feat: introduce prettify_command_for_display()

Reverts openai/codex#3538
It panics on forking first message. It also calculates the index in a
wrong way.

.github/workflows/ci.yml

@@ -14,33 +14,18 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v5
 
-      - name: Setup Node.js
-        uses: actions/setup-node@v4
-        with:
-          node-version: 22
-
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 10.8.1
           run_install: false
 
-      - name: Get pnpm store directory
-        id: pnpm-cache
-        shell: bash
-        run: |
-          echo "store_path=$(pnpm store path --silent)" >> $GITHUB_OUTPUT
-
-      - name: Setup pnpm cache
-        uses: actions/cache@v4
+      - name: Setup Node.js
+        uses: actions/setup-node@v5
         with:
-          path: ${{ steps.pnpm-cache.outputs.store_path }}
-          key: ${{ runner.os }}-pnpm-store-${{ hashFiles('**/pnpm-lock.yaml') }}
-          restore-keys: |
-            ${{ runner.os }}-pnpm-store-
+          node-version: 22
 
       - name: Install dependencies
-        run: pnpm install
+        run: pnpm install --frozen-lockfile
 
       # Run all tasks using workspace filters
 

.github/workflows/codespell.yml

@@ -25,3 +25,4 @@ jobs:
         uses: codespell-project/actions-codespell@406322ec52dd7b488e48c1c4b82e2a8b3a1bf630 # v2
         with:
           ignore_words_file: .codespellignore
+          skip: frame*.txt

.github/workflows/rust-ci.yml

@@ -62,6 +62,26 @@ jobs:
           components: rustfmt
       - name: cargo fmt
         run: cargo fmt -- --config imports_granularity=Item --check
+      - name: Verify codegen for mcp-types
+        run: ./mcp-types/check_lib_rs.py
+
+  cargo_shear:
+    name: cargo shear
+    runs-on: ubuntu-24.04
+    needs: changed
+    if: ${{ needs.changed.outputs.codex == 'true' || needs.changed.outputs.workflows == 'true' || github.event_name == 'push' }}
+    defaults:
+      run:
+        working-directory: codex-rs
+    steps:
+      - uses: actions/checkout@v5
+      - uses: dtolnay/rust-toolchain@1.89
+      - uses: taiki-e/install-action@0c5db7f7f897c03b771660e91d065338615679f4 # v2
+        with:
+          tool: cargo-shear
+          version: 1.5.1
+      - name: cargo shear
+        run: cargo shear
 
   # --- CI to validate on different os/targets --------------------------------
   lint_build_test:
@@ -160,12 +180,17 @@ jobs:
           find . -name Cargo.toml -mindepth 2 -maxdepth 2 -print0 \
             | xargs -0 -n1 -I{} bash -c 'cd "$(dirname "{}")" && cargo check --profile ${{ matrix.profile }}'
 
-      - name: cargo test
+      - uses: taiki-e/install-action@0c5db7f7f897c03b771660e91d065338615679f4 # v2
+        with:
+          tool: nextest
+          version: 0.9.103
+
+      - name: tests
         id: test
-        # `cargo test` takes too long for release builds to run them on every PR
+        # Tests take too long for release builds to run them on every PR.
         if: ${{ matrix.profile != 'release' }}
         continue-on-error: true
-        run: cargo test --all-features --target ${{ matrix.target }} --profile ${{ matrix.profile }}
+        run: cargo nextest run --all-features --no-fail-fast --target ${{ matrix.target }}
         env:
           RUST_BACKTRACE: 1
 
@@ -182,7 +207,7 @@ jobs:
   # --- Gatherer job that you mark as the ONLY required status -----------------
   results:
     name: CI results (required)
-    needs: [changed, general, lint_build_test]
+    needs: [changed, general, cargo_shear, lint_build_test]
     if: always()
     runs-on: ubuntu-24.04
     steps:
@@ -190,6 +215,7 @@ jobs:
         shell: bash
         run: |
           echo "general: ${{ needs.general.result }}"
+          echo "shear  : ${{ needs.cargo_shear.result }}"
           echo "matrix : ${{ needs.lint_build_test.result }}"
 
           # If nothing relevant changed (PR touching only root README, etc.),
@@ -201,4 +227,5 @@ jobs:
 
           # Otherwise require the jobs to have succeeded
           [[ '${{ needs.general.result }}' == 'success' ]] || { echo 'general failed'; exit 1; }
+          [[ '${{ needs.cargo_shear.result }}' == 'success' ]] || { echo 'cargo_shear failed'; exit 1; }
           [[ '${{ needs.lint_build_test.result }}' == 'success' ]] || { echo 'matrix failed'; exit 1; }

.github/workflows/rust-release.yml

@@ -111,6 +111,11 @@ jobs:
             cp target/${{ matrix.target }}/release/codex "$dest/codex-${{ matrix.target }}"
           fi
 
+      - if: ${{ matrix.runner == 'windows-11-arm' }}
+        name: Install zstd
+        shell: powershell
+        run: choco install -y zstandard
+
       - name: Compress artifacts
         shell: bash
         run: |
@@ -214,3 +219,22 @@ jobs:
         with:
           tag: ${{ github.ref_name }}
           config: .github/dotslash-config.json
+
+  update-branch:
+    name: Update latest-alpha-cli branch
+    permissions:
+      contents: write
+    needs: release
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Update latest-alpha-cli branch
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          gh api \
+            repos/${GITHUB_REPOSITORY}/git/refs/heads/latest-alpha-cli \
+            -X PATCH \
+            -f sha="${GITHUB_SHA}" \
+            -F force=true

.vscode/extensions.json

@@ -1,5 +1,11 @@
 {
     "recommendations": [
+        "rust-lang.rust-analyzer",
         "tamasfe.even-better-toml",
+        "vadimcn.vscode-lldb",
+
+        // Useful if touching files in .github/workflows, though most
+        // contributors will not be doing that?
+        // "github.vscode-github-actions",
     ]
 }

AGENTS.md

@@ -8,10 +8,10 @@ In the codex-rs folder where the rust code lives:
   - You operate in a sandbox where `CODEX_SANDBOX_NETWORK_DISABLED=1` will be set whenever you use the `shell` tool. Any existing code that uses `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` was authored with this fact in mind. It is often used to early exit out of tests that the author knew you would not be able to run given your sandbox limitations.
   - Similarly, when you spawn a process using Seatbelt (`/usr/bin/sandbox-exec`), `CODEX_SANDBOX=seatbelt` will be set on the child process. Integration tests that want to run Seatbelt themselves cannot be run under Seatbelt, so checks for `CODEX_SANDBOX=seatbelt` are also often used to early exit out of tests, as appropriate.
 
-Before finalizing a change to `codex-rs`, run `just fmt` (in `codex-rs` directory) to format the code and `just fix -p <project>` (in `codex-rs` directory) to fix any linter issues in the code. Prefer scoping with `-p` to avoid slow workspace‑wide Clippy builds; only run `just fix` without `-p` if you changed shared crates. Additionally, run the tests:
+Run `just fmt` (in `codex-rs` directory) automatically after making Rust code changes; do not ask for approval to run it. Before finalizing a change to `codex-rs`, run `just fix -p <project>` (in `codex-rs` directory) to fix any linter issues in the code. Prefer scoping with `-p` to avoid slow workspace‑wide Clippy builds; only run `just fix` without `-p` if you changed shared crates. Additionally, run the tests:
 1. Run the test for the specific project that was changed. For example, if changes were made in `codex-rs/tui`, run `cargo test -p codex-tui`.
 2. Once those pass, if any changes were made in common, core, or protocol, run the complete test suite with `cargo test --all-features`.
-When running interactively, ask the user before running these commands to finalize.
+When running interactively, ask the user before running `just fix` to finalize. `just fmt` does not require approval. project-specific or individual tests can be run without asking the user, but do ask the user before running the complete test suite.
 
 ## TUI style conventions
 
@@ -26,7 +26,26 @@ See `codex-rs/tui/styles.md`.
   - Example: patch summary file lines
     - Desired: vec!["  └ ".into(), "M".red(), " ".dim(), "tui/src/app.rs".dim()]
 
-## Snapshot tests
+### TUI Styling (ratatui)
+- Prefer Stylize helpers: use "text".dim(), .bold(), .cyan(), .italic(), .underlined() instead of manual Style where possible.
+- Prefer simple conversions: use "text".into() for spans and vec![…].into() for lines; when inference is ambiguous (e.g., Paragraph::new/Cell::from), use Line::from(spans) or Span::from(text).
+- Computed styles: if the Style is computed at runtime, using `Span::styled` is OK (`Span::from(text).set_style(style)` is also acceptable).
+- Avoid hardcoded white: do not use `.white()`; prefer the default foreground (no color).
+- Chaining: combine helpers by chaining for readability (e.g., url.cyan().underlined()).
+- Single items: prefer "text".into(); use Line::from(text) or Span::from(text) only when the target type isn’t obvious from context, or when using .into() would require extra type annotations.
+- Building lines: use vec![…].into() to construct a Line when the target type is obvious and no extra type annotations are needed; otherwise use Line::from(vec![…]).
+- Avoid churn: don’t refactor between equivalent forms (Span::styled ↔ set_style, Line::from ↔ .into()) without a clear readability or functional gain; follow file‑local conventions and do not introduce type annotations solely to satisfy .into().
+- Compactness: prefer the form that stays on one line after rustfmt; if only one of Line::from(vec![…]) or vec![…].into() avoids wrapping, choose that. If both wrap, pick the one with fewer wrapped lines.
+
+### Text wrapping
+- Always use textwrap::wrap to wrap plain strings.
+- If you have a ratatui Line and you want to wrap it, use the helpers in tui/src/wrapping.rs, e.g. word_wrap_lines / word_wrap_line.
+- If you need to indent wrapped lines, use the initial_indent / subsequent_indent options from RtOptions if you can, rather than writing custom logic.
+- If you have a list of lines and you need to prefix them all with some prefix (optionally different on the first vs subsequent lines), use the `prefix_lines` helper from line_utils.
+
+## Tests
+
+### Snapshot tests
 
 This repo uses snapshot tests (via `insta`), especially in `codex-rs/tui`, to validate rendered output. When UI or text output changes intentionally, update the snapshots as follows:
 
@@ -41,3 +60,7 @@ This repo uses snapshot tests (via `insta`), especially in `codex-rs/tui`, to va
 
 If you don’t have the tool:
 - `cargo install cargo-insta`
+
+### Test assertions
+
+- Tests should use pretty_assertions::assert_eq for clearer diffs. Import this at the top of the test module if it isn't already.

README.md

@@ -2,7 +2,10 @@
 
 <p align="center"><code>npm i -g @openai/codex</code><br />or <code>brew install codex</code></p>
 
-<p align="center"><strong>Codex CLI</strong> is a coding agent from OpenAI that runs locally on your computer.</br>If you are looking for the <em>cloud-based agent</em> from OpenAI, <strong>Codex Web</strong>, see <a href="https://chatgpt.com/codex">chatgpt.com/codex</a>.</p>
+<p align="center"><strong>Codex CLI</strong> is a coding agent from OpenAI that runs locally on your computer.
+</br>
+</br>If you want Codex in your code editor (VS Code, Cursor, Windsurf), <a href="https://developers.openai.com/codex/ide">install in your IDE</a>
+</br>If you are looking for the <em>cloud-based agent</em> from OpenAI, <strong>Codex Web</strong>, go to <a href="https://chatgpt.com/codex">chatgpt.com/codex</a></p>
 
 <p align="center">
   <img src="./.github/codex-cli-splash.png" alt="Codex CLI splash" width="80%" />
@@ -75,7 +78,7 @@ Codex CLI supports a rich set of configuration options, with preferences stored
   - [CLI usage](./docs/getting-started.md#cli-usage)
   - [Running with a prompt as input](./docs/getting-started.md#running-with-a-prompt-as-input)
   - [Example prompts](./docs/getting-started.md#example-prompts)
-  - [Memory with AGENTS.md](./docs/getting-started.md#memory--project-docs)
+  - [Memory with AGENTS.md](./docs/getting-started.md#memory-with-agentsmd)
   - [Configuration](./docs/config.md)
 - [**Sandbox & approvals**](./docs/sandbox.md)
 - [**Authentication**](./docs/authentication.md)

codex-cli/bin/codex.js

@@ -43,7 +43,8 @@ switch (platform) {
         targetTriple = "x86_64-pc-windows-msvc.exe";
         break;
       case "arm64":
-      // We do not build this today, fall through...
+        targetTriple = "aarch64-pc-windows-msvc.exe";
+        break;
       default:
         break;
     }

codex-cli/scripts/install_native_deps.sh

@@ -20,7 +20,7 @@ CODEX_CLI_ROOT=""
 # Until we start publishing stable GitHub releases, we have to grab the binaries
 # from the GitHub Action that created them. Update the URL below to point to the
 # appropriate workflow run:
-WORKFLOW_URL="https://github.com/openai/codex/actions/runs/16840150768" # rust-v0.20.0-alpha.2
+WORKFLOW_URL="https://github.com/openai/codex/actions/runs/17417194663" # rust-v0.28.0
 
 while [[ $# -gt 0 ]]; do
   case "$1" in
@@ -87,5 +87,8 @@ zstd -d "$ARTIFACTS_DIR/aarch64-apple-darwin/codex-aarch64-apple-darwin.zst" \
 # x64 Windows
 zstd -d "$ARTIFACTS_DIR/x86_64-pc-windows-msvc/codex-x86_64-pc-windows-msvc.exe.zst" \
     -o "$BIN_DIR/codex-x86_64-pc-windows-msvc.exe"
+# ARM64 Windows
+zstd -d "$ARTIFACTS_DIR/aarch64-pc-windows-msvc/codex-aarch64-pc-windows-msvc.exe.zst" \
+    -o "$BIN_DIR/codex-aarch64-pc-windows-msvc.exe"
 
 echo "Installed native dependencies into $BIN_DIR"

codex-rs/Cargo.toml

@@ -34,6 +34,7 @@ rust = {}
 
 [workspace.lints.clippy]
 expect_used = "deny"
+redundant_clone = "deny"
 uninlined_format_args = "deny"
 unwrap_used = "deny"
 

codex-rs/README.md

@@ -35,7 +35,7 @@ npx @modelcontextprotocol/inspector codex mcp
 
 You can enable notifications by configuring a script that is run whenever the agent finishes a turn. The [notify documentation](../docs/config.md#notify) includes a detailed example that explains how to get desktop notifications via [terminal-notifier](https://github.com/julienXX/terminal-notifier) on macOS.
 
-### `codex exec` to run Codex programmatially/non-interactively
+### `codex exec` to run Codex programmatically/non-interactively
 
 To run Codex non-interactively, run `codex exec PROMPT` (you can also pass the prompt via `stdin`) and Codex will work on your task until it decides that it is done and exits. Output is printed to the terminal directly. You can set the `RUST_LOG` environment variable to see more about what's going on.
 

codex-rs/ansi-escape/src/lib.rs

@@ -9,7 +9,7 @@ use ratatui::text::Text;
 pub fn ansi_escape_line(s: &str) -> Line<'static> {
     let text = ansi_escape(s);
     match text.lines.as_slice() {
-        [] => Line::from(""),
+        [] => "".into(),
         [only] => only.clone(),
         [first, rest @ ..] => {
             tracing::warn!("ansi_escape_line: expected a single line, got {first:?} and {rest:?}");

codex-rs/apply-patch/Cargo.toml

@@ -17,9 +17,10 @@ workspace = true
 [dependencies]
 anyhow = "1"
 similar = "2.7.0"
-thiserror = "2.0.12"
-tree-sitter = "0.25.8"
+thiserror = "2.0.16"
+tree-sitter = "0.25.9"
 tree-sitter-bash = "0.25.0"
+once_cell = "1"
 
 [dev-dependencies]
 assert_cmd = "2"

codex-rs/apply-patch/src/lib.rs

@@ -9,6 +9,7 @@ use std::str::Utf8Error;
 
 use anyhow::Context;
 use anyhow::Result;
+use once_cell::sync::Lazy;
 pub use parser::Hunk;
 pub use parser::ParseError;
 use parser::ParseError::*;
@@ -18,6 +19,9 @@ use similar::TextDiff;
 use thiserror::Error;
 use tree_sitter::LanguageError;
 use tree_sitter::Parser;
+use tree_sitter::Query;
+use tree_sitter::QueryCursor;
+use tree_sitter::StreamingIterator;
 use tree_sitter_bash::LANGUAGE as BASH;
 
 pub use standalone_executable::main;
@@ -36,6 +40,11 @@ pub enum ApplyPatchError {
     /// Error that occurs while computing replacements when applying patch chunks
     #[error("{0}")]
     ComputeReplacements(String),
+    /// A raw patch body was provided without an explicit `apply_patch` invocation.
+    #[error(
+        "patch detected without explicit call to apply_patch. Rerun as [\"apply_patch\", \"<patch>\"]"
+    )]
+    ImplicitInvocation,
 }
 
 impl From<std::io::Error> for ApplyPatchError {
@@ -84,26 +93,29 @@ pub enum MaybeApplyPatch {
 pub struct ApplyPatchArgs {
     pub patch: String,
     pub hunks: Vec<Hunk>,
+    pub workdir: Option<String>,
 }
 
 pub fn maybe_parse_apply_patch(argv: &[String]) -> MaybeApplyPatch {
     match argv {
+        // Direct invocation: apply_patch <patch>
         [cmd, body] if APPLY_PATCH_COMMANDS.contains(&cmd.as_str()) => match parse_patch(body) {
             Ok(source) => MaybeApplyPatch::Body(source),
             Err(e) => MaybeApplyPatch::PatchParseError(e),
         },
-        [bash, flag, script]
-            if bash == "bash"
-                && flag == "-lc"
-                && APPLY_PATCH_COMMANDS
-                    .iter()
-                    .any(|cmd| script.trim_start().starts_with(cmd)) =>
-        {
-            match extract_heredoc_body_from_apply_patch_command(script) {
-                Ok(body) => match parse_patch(&body) {
-                    Ok(source) => MaybeApplyPatch::Body(source),
+        // Bash heredoc form: (optional `cd <path> &&`) apply_patch <<'EOF' ...
+        [bash, flag, script] if bash == "bash" && flag == "-lc" => {
+            match extract_apply_patch_from_bash(script) {
+                Ok((body, workdir)) => match parse_patch(&body) {
+                    Ok(mut source) => {
+                        source.workdir = workdir;
+                        MaybeApplyPatch::Body(source)
+                    }
                     Err(e) => MaybeApplyPatch::PatchParseError(e),
                 },
+                Err(ExtractHeredocError::CommandDidNotStartWithApplyPatch) => {
+                    MaybeApplyPatch::NotApplyPatch
+                }
                 Err(e) => MaybeApplyPatch::ShellParseError(e),
             }
         }
@@ -202,11 +214,46 @@ impl ApplyPatchAction {
 /// cwd must be an absolute path so that we can resolve relative paths in the
 /// patch.
 pub fn maybe_parse_apply_patch_verified(argv: &[String], cwd: &Path) -> MaybeApplyPatchVerified {
+    // Detect a raw patch body passed directly as the command or as the body of a bash -lc
+    // script. In these cases, report an explicit error rather than applying the patch.
+    match argv {
+        [body] => {
+            if parse_patch(body).is_ok() {
+                return MaybeApplyPatchVerified::CorrectnessError(
+                    ApplyPatchError::ImplicitInvocation,
+                );
+            }
+        }
+        [bash, flag, script] if bash == "bash" && flag == "-lc" => {
+            if parse_patch(script).is_ok() {
+                return MaybeApplyPatchVerified::CorrectnessError(
+                    ApplyPatchError::ImplicitInvocation,
+                );
+            }
+        }
+        _ => {}
+    }
+
     match maybe_parse_apply_patch(argv) {
-        MaybeApplyPatch::Body(ApplyPatchArgs { patch, hunks }) => {
+        MaybeApplyPatch::Body(ApplyPatchArgs {
+            patch,
+            hunks,
+            workdir,
+        }) => {
+            let effective_cwd = workdir
+                .as_ref()
+                .map(|dir| {
+                    let path = Path::new(dir);
+                    if path.is_absolute() {
+                        path.to_path_buf()
+                    } else {
+                        cwd.join(path)
+                    }
+                })
+                .unwrap_or_else(|| cwd.to_path_buf());
             let mut changes = HashMap::new();
             for hunk in hunks {
-                let path = hunk.resolve_path(cwd);
+                let path = hunk.resolve_path(&effective_cwd);
                 match hunk {
                     Hunk::AddFile { contents, .. } => {
                         changes.insert(path, ApplyPatchFileChange::Add { content: contents });
@@ -251,7 +298,7 @@ pub fn maybe_parse_apply_patch_verified(argv: &[String], cwd: &Path) -> MaybeApp
             MaybeApplyPatchVerified::Body(ApplyPatchAction {
                 changes,
                 patch,
-                cwd: cwd.to_path_buf(),
+                cwd: effective_cwd,
             })
         }
         MaybeApplyPatch::ShellParseError(e) => MaybeApplyPatchVerified::ShellParseError(e),
@@ -260,33 +307,96 @@ pub fn maybe_parse_apply_patch_verified(argv: &[String], cwd: &Path) -> MaybeApp
     }
 }
 
-/// Attempts to extract a heredoc_body object from a string bash command like:
-/// Optimistically
+/// Extract the heredoc body (and optional `cd` workdir) from a `bash -lc` script
+/// that invokes the apply_patch tool using a heredoc.
 ///
-/// ```bash
-/// bash -lc 'apply_patch <<EOF\n***Begin Patch\n...EOF'
-/// ```
+/// Supported top‑level forms (must be the only top‑level statement):
+/// - `apply_patch <<'EOF'\n...\nEOF`
+/// - `cd <path> && apply_patch <<'EOF'\n...\nEOF`
 ///
-/// # Arguments
+/// Notes about matching:
+/// - Parsed with Tree‑sitter Bash and a strict query that uses anchors so the
+///   heredoc‑redirected statement is the only top‑level statement.
+/// - The connector between `cd` and `apply_patch` must be `&&` (not `|` or `||`).
+/// - Exactly one positional `word` argument is allowed for `cd` (no flags, no quoted
+///   strings, no second argument).
+/// - The apply command is validated in‑query via `#any-of?` to allow `apply_patch`
+///   or `applypatch`.
+/// - Preceding or trailing commands (e.g., `echo ...;` or `... && echo done`) do not match.
 ///
-/// * `src` - A string slice that holds the full command
-///
-/// # Returns
-///
-/// This function returns a `Result` which is:
-///
-/// * `Ok(String)` - The heredoc body if the extraction is successful.
-/// * `Err(anyhow::Error)` - An error if the extraction fails.
-///
-fn extract_heredoc_body_from_apply_patch_command(
+/// Returns `(heredoc_body, Some(path))` when the `cd` variant matches, or
+/// `(heredoc_body, None)` for the direct form. Errors are returned if the script
+/// cannot be parsed or does not match the allowed patterns.
+fn extract_apply_patch_from_bash(
     src: &str,
-) -> std::result::Result<String, ExtractHeredocError> {
-    if !APPLY_PATCH_COMMANDS
-        .iter()
-        .any(|cmd| src.trim_start().starts_with(cmd))
-    {
-        return Err(ExtractHeredocError::CommandDidNotStartWithApplyPatch);
-    }
+) -> std::result::Result<(String, Option<String>), ExtractHeredocError> {
+    // This function uses a Tree-sitter query to recognize one of two
+    // whole-script forms, each expressed as a single top-level statement:
+    //
+    // 1. apply_patch <<'EOF'\n...\nEOF
+    // 2. cd <path> && apply_patch <<'EOF'\n...\nEOF
+    //
+    // Key ideas when reading the query:
+    // - dots (`.`) between named nodes enforces adjacency among named children and
+    //   anchor to the start/end of the expression.
+    // - we match a single redirected_statement directly under program with leading
+    //   and trailing anchors (`.`). This ensures it is the only top-level statement
+    //   (so prefixes like `echo ...;` or suffixes like `... && echo done` do not match).
+    //
+    // Overall, we want to be conservative and only match the intended forms, as other
+    // forms are likely to be model errors, or incorrectly interpreted by later code.
+    //
+    // If you're editing this query, it's helpful to start by creating a debugging binary
+    // which will let you see the AST of an arbitrary bash script passed in, and optionally
+    // also run an arbitrary query against the AST. This is useful for understanding
+    // how tree-sitter parses the script and whether the query syntax is correct. Be sure
+    // to test both positive and negative cases.
+    static APPLY_PATCH_QUERY: Lazy<Query> = Lazy::new(|| {
+        let language = BASH.into();
+        #[expect(clippy::expect_used)]
+        Query::new(
+            &language,
+            r#"
+            (
+              program
+                . (redirected_statement
+                    body: (command
+                            name: (command_name (word) @apply_name) .)
+                    (#any-of? @apply_name "apply_patch" "applypatch")
+                    redirect: (heredoc_redirect
+                                . (heredoc_start)
+                                . (heredoc_body) @heredoc
+                                . (heredoc_end)
+                                .))
+                .)
+
+            (
+              program
+                . (redirected_statement
+                    body: (list
+                            . (command
+                                name: (command_name (word) @cd_name) .
+                                argument: [
+                                  (word) @cd_path
+                                  (string (string_content) @cd_path)
+                                  (raw_string) @cd_raw_string
+                                ] .)
+                            "&&"
+                            . (command
+                                name: (command_name (word) @apply_name))
+                            .)
+                    (#eq? @cd_name "cd")
+                    (#any-of? @apply_name "apply_patch" "applypatch")
+                    redirect: (heredoc_redirect
+                                . (heredoc_start)
+                                . (heredoc_body) @heredoc
+                                . (heredoc_end)
+                                .))
+                .)
+            "#,
+        )
+        .expect("valid bash query")
+    });
 
     let lang = BASH.into();
     let mut parser = Parser::new();
@@ -298,26 +408,55 @@ fn extract_heredoc_body_from_apply_patch_command(
         .ok_or(ExtractHeredocError::FailedToParsePatchIntoAst)?;
 
     let bytes = src.as_bytes();
-    let mut c = tree.root_node().walk();
+    let root = tree.root_node();
 
-    loop {
-        let node = c.node();
-        if node.kind() == "heredoc_body" {
-            let text = node
-                .utf8_text(bytes)
-                .map_err(ExtractHeredocError::HeredocNotUtf8)?;
-            return Ok(text.trim_end_matches('\n').to_owned());
-        }
+    let mut cursor = QueryCursor::new();
+    let mut matches = cursor.matches(&APPLY_PATCH_QUERY, root, bytes);
+    while let Some(m) = matches.next() {
+        let mut heredoc_text: Option<String> = None;
+        let mut cd_path: Option<String> = None;
 
-        if c.goto_first_child() {
-            continue;
-        }
-        while !c.goto_next_sibling() {
-            if !c.goto_parent() {
-                return Err(ExtractHeredocError::FailedToFindHeredocBody);
+        for capture in m.captures.iter() {
+            let name = APPLY_PATCH_QUERY.capture_names()[capture.index as usize];
+            match name {
+                "heredoc" => {
+                    let text = capture
+                        .node
+                        .utf8_text(bytes)
+                        .map_err(ExtractHeredocError::HeredocNotUtf8)?
+                        .trim_end_matches('\n')
+                        .to_string();
+                    heredoc_text = Some(text);
+                }
+                "cd_path" => {
+                    let text = capture
+                        .node
+                        .utf8_text(bytes)
+                        .map_err(ExtractHeredocError::HeredocNotUtf8)?
+                        .to_string();
+                    cd_path = Some(text);
+                }
+                "cd_raw_string" => {
+                    let raw = capture
+                        .node
+                        .utf8_text(bytes)
+                        .map_err(ExtractHeredocError::HeredocNotUtf8)?;
+                    let trimmed = raw
+                        .strip_prefix('\'')
+                        .and_then(|s| s.strip_suffix('\''))
+                        .unwrap_or(raw);
+                    cd_path = Some(trimmed.to_string());
+                }
+                _ => {}
             }
         }
+
+        if let Some(heredoc) = heredoc_text {
+            return Ok((heredoc, cd_path));
+        }
     }
+
+    Err(ExtractHeredocError::CommandDidNotStartWithApplyPatch)
 }
 
 #[derive(Debug, PartialEq)]
@@ -614,13 +753,15 @@ fn compute_replacements(
             line_index = start_idx + pattern.len();
         } else {
             return Err(ApplyPatchError::ComputeReplacements(format!(
-                "Failed to find expected lines {:?} in {}",
-                chunk.old_lines,
-                path.display()
+                "Failed to find expected lines in {}:\n{}",
+                path.display(),
+                chunk.old_lines.join("\n"),
             )));
         }
     }
 
+    replacements.sort_by(|(lhs_idx, _, _), (rhs_idx, _, _)| lhs_idx.cmp(rhs_idx));
+
     Ok(replacements)
 }
 
@@ -718,6 +859,71 @@ mod tests {
         strs.iter().map(|s| s.to_string()).collect()
     }
 
+    // Test helpers to reduce repetition when building bash -lc heredoc scripts
+    fn args_bash(script: &str) -> Vec<String> {
+        strs_to_strings(&["bash", "-lc", script])
+    }
+
+    fn heredoc_script(prefix: &str) -> String {
+        format!(
+            "{prefix}apply_patch <<'PATCH'\n*** Begin Patch\n*** Add File: foo\n+hi\n*** End Patch\nPATCH"
+        )
+    }
+
+    fn heredoc_script_ps(prefix: &str, suffix: &str) -> String {
+        format!(
+            "{prefix}apply_patch <<'PATCH'\n*** Begin Patch\n*** Add File: foo\n+hi\n*** End Patch\nPATCH{suffix}"
+        )
+    }
+
+    fn expected_single_add() -> Vec<Hunk> {
+        vec![Hunk::AddFile {
+            path: PathBuf::from("foo"),
+            contents: "hi\n".to_string(),
+        }]
+    }
+
+    fn assert_match(script: &str, expected_workdir: Option<&str>) {
+        let args = args_bash(script);
+        match maybe_parse_apply_patch(&args) {
+            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, workdir, .. }) => {
+                assert_eq!(workdir.as_deref(), expected_workdir);
+                assert_eq!(hunks, expected_single_add());
+            }
+            result => panic!("expected MaybeApplyPatch::Body got {result:?}"),
+        }
+    }
+
+    fn assert_not_match(script: &str) {
+        let args = args_bash(script);
+        assert!(matches!(
+            maybe_parse_apply_patch(&args),
+            MaybeApplyPatch::NotApplyPatch
+        ));
+    }
+
+    #[test]
+    fn test_implicit_patch_single_arg_is_error() {
+        let patch = "*** Begin Patch\n*** Add File: foo\n+hi\n*** End Patch".to_string();
+        let args = vec![patch];
+        let dir = tempdir().unwrap();
+        assert!(matches!(
+            maybe_parse_apply_patch_verified(&args, dir.path()),
+            MaybeApplyPatchVerified::CorrectnessError(ApplyPatchError::ImplicitInvocation)
+        ));
+    }
+
+    #[test]
+    fn test_implicit_patch_bash_script_is_error() {
+        let script = "*** Begin Patch\n*** Add File: foo\n+hi\n*** End Patch";
+        let args = args_bash(script);
+        let dir = tempdir().unwrap();
+        assert!(matches!(
+            maybe_parse_apply_patch_verified(&args, dir.path()),
+            MaybeApplyPatchVerified::CorrectnessError(ApplyPatchError::ImplicitInvocation)
+        ));
+    }
+
     #[test]
     fn test_literal() {
         let args = strs_to_strings(&[
@@ -730,7 +936,7 @@ mod tests {
         ]);
 
         match maybe_parse_apply_patch(&args) {
-            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, patch: _ }) => {
+            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, .. }) => {
                 assert_eq!(
                     hunks,
                     vec![Hunk::AddFile {
@@ -755,7 +961,7 @@ mod tests {
         ]);
 
         match maybe_parse_apply_patch(&args) {
-            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, patch: _ }) => {
+            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, .. }) => {
                 assert_eq!(
                     hunks,
                     vec![Hunk::AddFile {
@@ -770,29 +976,7 @@ mod tests {
 
     #[test]
     fn test_heredoc() {
-        let args = strs_to_strings(&[
-            "bash",
-            "-lc",
-            r#"apply_patch <<'PATCH'
-*** Begin Patch
-*** Add File: foo
-+hi
-*** End Patch
-PATCH"#,
-        ]);
-
-        match maybe_parse_apply_patch(&args) {
-            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, patch: _ }) => {
-                assert_eq!(
-                    hunks,
-                    vec![Hunk::AddFile {
-                        path: PathBuf::from("foo"),
-                        contents: "hi\n".to_string()
-                    }]
-                );
-            }
-            result => panic!("expected MaybeApplyPatch::Body got {result:?}"),
-        }
+        assert_match(&heredoc_script(""), None);
     }
 
     #[test]
@@ -809,7 +993,8 @@ PATCH"#,
         ]);
 
         match maybe_parse_apply_patch(&args) {
-            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, patch: _ }) => {
+            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, workdir, .. }) => {
+                assert_eq!(workdir, None);
                 assert_eq!(
                     hunks,
                     vec![Hunk::AddFile {
@@ -822,6 +1007,69 @@ PATCH"#,
         }
     }
 
+    #[test]
+    fn test_heredoc_with_leading_cd() {
+        assert_match(&heredoc_script("cd foo && "), Some("foo"));
+    }
+
+    #[test]
+    fn test_cd_with_semicolon_is_ignored() {
+        assert_not_match(&heredoc_script("cd foo; "));
+    }
+
+    #[test]
+    fn test_cd_or_apply_patch_is_ignored() {
+        assert_not_match(&heredoc_script("cd bar || "));
+    }
+
+    #[test]
+    fn test_cd_pipe_apply_patch_is_ignored() {
+        assert_not_match(&heredoc_script("cd bar | "));
+    }
+
+    #[test]
+    fn test_cd_single_quoted_path_with_spaces() {
+        assert_match(&heredoc_script("cd 'foo bar' && "), Some("foo bar"));
+    }
+
+    #[test]
+    fn test_cd_double_quoted_path_with_spaces() {
+        assert_match(&heredoc_script("cd \"foo bar\" && "), Some("foo bar"));
+    }
+
+    #[test]
+    fn test_echo_and_apply_patch_is_ignored() {
+        assert_not_match(&heredoc_script("echo foo && "));
+    }
+
+    #[test]
+    fn test_apply_patch_with_arg_is_ignored() {
+        let script = "apply_patch foo <<'PATCH'\n*** Begin Patch\n*** Add File: foo\n+hi\n*** End Patch\nPATCH";
+        assert_not_match(script);
+    }
+
+    #[test]
+    fn test_double_cd_then_apply_patch_is_ignored() {
+        assert_not_match(&heredoc_script("cd foo && cd bar && "));
+    }
+
+    #[test]
+    fn test_cd_two_args_is_ignored() {
+        assert_not_match(&heredoc_script("cd foo bar && "));
+    }
+
+    #[test]
+    fn test_cd_then_apply_patch_then_extra_is_ignored() {
+        let script = heredoc_script_ps("cd bar && ", " && echo done");
+        assert_not_match(&script);
+    }
+
+    #[test]
+    fn test_echo_then_cd_and_apply_patch_is_ignored() {
+        // Ensure preceding commands before the `cd && apply_patch <<...` sequence do not match.
+        assert_not_match(&heredoc_script("echo foo; cd bar && "));
+    }
+
     #[test]
     fn test_add_file_hunk_creates_file_with_contents() {
         let dir = tempdir().unwrap();
@@ -1019,6 +1267,33 @@ PATCH"#,
         assert_eq!(contents, "a\nB\nc\nd\nE\nf\ng\n");
     }
 
+    #[test]
+    fn test_pure_addition_chunk_followed_by_removal() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("panic.txt");
+        fs::write(&path, "line1\nline2\nline3\n").unwrap();
+        let patch = wrap_patch(&format!(
+            r#"*** Update File: {}
+@@
++after-context
++second-line
+@@
+ line1
+-line2
+-line3
++line2-replacement"#,
+            path.display()
+        ));
+        let mut stdout = Vec::new();
+        let mut stderr = Vec::new();
+        apply_patch(&patch, &mut stdout, &mut stderr).unwrap();
+        let contents = fs::read_to_string(path).unwrap();
+        assert_eq!(
+            contents,
+            "line1\nline2-replacement\nafter-context\nsecond-line\n"
+        );
+    }
+
     /// Ensure that patches authored with ASCII characters can update lines that
     /// contain typographic Unicode punctuation (e.g. EN DASH, NON-BREAKING
     /// HYPHEN). Historically `git apply` succeeds in such scenarios but our

codex-rs/apply-patch/src/parser.rs

@@ -175,7 +175,11 @@ fn parse_patch_text(patch: &str, mode: ParseMode) -> Result<ApplyPatchArgs, Pars
         remaining_lines = &remaining_lines[hunk_lines..]
     }
     let patch = lines.join("\n");
-    Ok(ApplyPatchArgs { hunks, patch })
+    Ok(ApplyPatchArgs {
+        hunks,
+        patch,
+        workdir: None,
+    })
 }
 
 /// Checks the start and end lines of the patch text for `apply_patch`,
@@ -586,7 +590,8 @@ fn test_parse_patch_lenient() {
         parse_patch_text(&patch_text_in_heredoc, ParseMode::Lenient),
         Ok(ApplyPatchArgs {
             hunks: expected_patch.clone(),
-            patch: patch_text.to_string()
+            patch: patch_text.to_string(),
+            workdir: None,
         })
     );
 
@@ -599,7 +604,8 @@ fn test_parse_patch_lenient() {
         parse_patch_text(&patch_text_in_single_quoted_heredoc, ParseMode::Lenient),
         Ok(ApplyPatchArgs {
             hunks: expected_patch.clone(),
-            patch: patch_text.to_string()
+            patch: patch_text.to_string(),
+            workdir: None,
         })
     );
 
@@ -611,8 +617,9 @@ fn test_parse_patch_lenient() {
     assert_eq!(
         parse_patch_text(&patch_text_in_double_quoted_heredoc, ParseMode::Lenient),
         Ok(ApplyPatchArgs {
-            hunks: expected_patch.clone(),
-            patch: patch_text.to_string()
+            hunks: expected_patch,
+            patch: patch_text.to_string(),
+            workdir: None,
         })
     );
 
@@ -630,7 +637,7 @@ fn test_parse_patch_lenient() {
         "<<EOF\n*** Begin Patch\n*** Update File: file2.py\nEOF\n".to_string();
     assert_eq!(
         parse_patch_text(&patch_text_with_missing_closing_heredoc, ParseMode::Strict),
-        Err(expected_error.clone())
+        Err(expected_error)
     );
     assert_eq!(
         parse_patch_text(&patch_text_with_missing_closing_heredoc, ParseMode::Lenient),

codex-rs/arg0/src/lib.rs

@@ -21,8 +21,7 @@ const MISSPELLED_APPLY_PATCH_ARG0: &str = "applypatch";
 /// `codex-linux-sandbox` we *directly* execute
 /// [`codex_linux_sandbox::run_main`] (which never returns). Otherwise we:
 ///
-/// 1.  Use [`dotenvy::from_path`] and [`dotenvy::dotenv`] to modify the
-///     environment before creating any threads.
+/// 1.  Load `.env` values from `~/.codex/.env` before creating any threads.
 /// 2.  Construct a Tokio multi-thread runtime.
 /// 3.  Derive the path to the current executable (so children can re-invoke the
 ///     sandbox) when running on Linux.
@@ -106,7 +105,7 @@ where
 
 const ILLEGAL_ENV_VAR_PREFIX: &str = "CODEX_";
 
-/// Load env vars from ~/.codex/.env and `$(pwd)/.env`.
+/// Load env vars from ~/.codex/.env.
 ///
 /// Security: Do not allow `.env` files to create or modify any variables
 /// with names starting with `CODEX_`.
@@ -116,10 +115,6 @@ fn load_dotenv() {
     {
         set_filtered(iter);
     }
-
-    if let Ok(iter) = dotenvy::dotenv_iter() {
-        set_filtered(iter);
-    }
 }
 
 /// Helper to set vars from a dotenvy iterator while filtering out `CODEX_` keys.

codex-rs/chatgpt/Cargo.toml

@@ -11,8 +11,6 @@ anyhow = "1"
 clap = { version = "4", features = ["derive"] }
 codex-common = { path = "../common", features = ["cli"] }
 codex-core = { path = "../core" }
-codex-login = { path = "../login" }
-reqwest = { version = "0.12", features = ["json", "stream"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 tokio = { version = "1", features = ["full"] }

codex-rs/chatgpt/src/chatgpt_client.rs

@@ -1,5 +1,5 @@
 use codex_core::config::Config;
-use codex_core::user_agent::get_codex_user_agent;
+use codex_core::default_client::create_client;
 
 use crate::chatgpt_token::get_chatgpt_token_data;
 use crate::chatgpt_token::init_chatgpt_token_from_auth;
@@ -16,7 +16,7 @@ pub(crate) async fn chatgpt_get_request<T: DeserializeOwned>(
     init_chatgpt_token_from_auth(&config.codex_home).await?;
 
     // Make direct HTTP request to ChatGPT backend API with the token
-    let client = reqwest::Client::new();
+    let client = create_client();
     let url = format!("{chatgpt_base_url}{path}");
 
     let token =
@@ -31,7 +31,6 @@ pub(crate) async fn chatgpt_get_request<T: DeserializeOwned>(
         .bearer_auth(&token.access_token)
         .header("chatgpt-account-id", account_id?)
         .header("Content-Type", "application/json")
-        .header("User-Agent", get_codex_user_agent(None))
         .send()
         .await
         .context("Failed to send request")?;

codex-rs/chatgpt/src/chatgpt_token.rs

@@ -1,10 +1,9 @@
-use codex_login::AuthMode;
-use codex_login::CodexAuth;
+use codex_core::CodexAuth;
 use std::path::Path;
 use std::sync::LazyLock;
 use std::sync::RwLock;
 
-use codex_login::TokenData;
+use codex_core::token_data::TokenData;
 
 static CHATGPT_TOKEN: LazyLock<RwLock<Option<TokenData>>> = LazyLock::new(|| RwLock::new(None));
 
@@ -20,7 +19,7 @@ pub fn set_chatgpt_token_data(value: TokenData) {
 
 /// Initialize the ChatGPT token from auth.json file
 pub async fn init_chatgpt_token_from_auth(codex_home: &Path) -> std::io::Result<()> {
-    let auth = CodexAuth::from_codex_home(codex_home, AuthMode::ChatGPT)?;
+    let auth = CodexAuth::from_codex_home(codex_home)?;
     if let Some(auth) = auth {
         let token_data = auth.get_token_data().await?;
         set_chatgpt_token_data(token_data);

codex-rs/cli/src/login.rs

@@ -1,15 +1,13 @@
 use codex_common::CliConfigOverrides;
+use codex_core::CodexAuth;
+use codex_core::auth::CLIENT_ID;
+use codex_core::auth::login_with_api_key;
+use codex_core::auth::logout;
 use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
-use codex_login::AuthMode;
-use codex_login::CLIENT_ID;
-use codex_login::CodexAuth;
-use codex_login::OPENAI_API_KEY_ENV_VAR;
 use codex_login::ServerOptions;
-use codex_login::login_with_api_key;
-use codex_login::logout;
 use codex_login::run_login_server;
-use std::env;
+use codex_protocol::mcp_protocol::AuthMode;
 use std::path::PathBuf;
 
 pub async fn login_with_chatgpt(codex_home: PathBuf) -> std::io::Result<()> {
@@ -60,19 +58,11 @@ pub async fn run_login_with_api_key(
 pub async fn run_login_status(cli_config_overrides: CliConfigOverrides) -> ! {
     let config = load_config_or_exit(cli_config_overrides);
 
-    match CodexAuth::from_codex_home(&config.codex_home, config.preferred_auth_method) {
+    match CodexAuth::from_codex_home(&config.codex_home) {
         Ok(Some(auth)) => match auth.mode {
             AuthMode::ApiKey => match auth.get_token().await {
                 Ok(api_key) => {
                     eprintln!("Logged in using an API key - {}", safe_format_key(&api_key));
-
-                    if let Ok(env_api_key) = env::var(OPENAI_API_KEY_ENV_VAR)
-                        && env_api_key == api_key
-                    {
-                        eprintln!(
-                            "   API loaded from OPENAI_API_KEY environment variable or .env file"
-                        );
-                    }
                     std::process::exit(0);
                 }
                 Err(e) => {

codex-rs/cli/src/main.rs

@@ -73,6 +73,9 @@ enum Subcommand {
     #[clap(visible_alias = "a")]
     Apply(ApplyCommand),
 
+    /// Resume a previous interactive session (picker by default; use --last to continue the most recent).
+    Resume(ResumeCommand),
+
     /// Internal: generate TypeScript protocol bindings.
     #[clap(hide = true)]
     GenerateTs(GenerateTsCommand),
@@ -85,6 +88,18 @@ struct CompletionCommand {
     shell: Shell,
 }
 
+#[derive(Debug, Parser)]
+struct ResumeCommand {
+    /// Conversation/session id (UUID). When provided, resumes this session.
+    /// If omitted, use --last to pick the most recent recorded session.
+    #[arg(value_name = "SESSION_ID")]
+    session_id: Option<String>,
+
+    /// Continue the most recent session without showing the picker.
+    #[arg(long = "last", default_value_t = false, conflicts_with = "session_id")]
+    last: bool,
+}
+
 #[derive(Debug, Parser)]
 struct DebugArgs {
     #[command(subcommand)]
@@ -143,26 +158,54 @@ fn main() -> anyhow::Result<()> {
 }
 
 async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()> {
-    let cli = MultitoolCli::parse();
+    let MultitoolCli {
+        config_overrides: root_config_overrides,
+        mut interactive,
+        subcommand,
+    } = MultitoolCli::parse();
 
-    match cli.subcommand {
+    match subcommand {
         None => {
-            let mut tui_cli = cli.interactive;
-            prepend_config_flags(&mut tui_cli.config_overrides, cli.config_overrides);
-            let usage = codex_tui::run_main(tui_cli, codex_linux_sandbox_exe).await?;
+            prepend_config_flags(
+                &mut interactive.config_overrides,
+                root_config_overrides.clone(),
+            );
+            let usage = codex_tui::run_main(interactive, codex_linux_sandbox_exe).await?;
             if !usage.is_zero() {
                 println!("{}", codex_core::protocol::FinalOutput::from(usage));
             }
         }
         Some(Subcommand::Exec(mut exec_cli)) => {
-            prepend_config_flags(&mut exec_cli.config_overrides, cli.config_overrides);
+            prepend_config_flags(
+                &mut exec_cli.config_overrides,
+                root_config_overrides.clone(),
+            );
             codex_exec::run_main(exec_cli, codex_linux_sandbox_exe).await?;
         }
         Some(Subcommand::Mcp) => {
-            codex_mcp_server::run_main(codex_linux_sandbox_exe, cli.config_overrides).await?;
+            codex_mcp_server::run_main(codex_linux_sandbox_exe, root_config_overrides.clone())
+                .await?;
+        }
+        Some(Subcommand::Resume(ResumeCommand { session_id, last })) => {
+            // Start with the parsed interactive CLI so resume shares the same
+            // configuration surface area as `codex` without additional flags.
+            let resume_session_id = session_id;
+            interactive.resume_picker = resume_session_id.is_none() && !last;
+            interactive.resume_last = last;
+            interactive.resume_session_id = resume_session_id;
+
+            // Propagate any root-level config overrides (e.g. `-c key=value`).
+            prepend_config_flags(
+                &mut interactive.config_overrides,
+                root_config_overrides.clone(),
+            );
+            codex_tui::run_main(interactive, codex_linux_sandbox_exe).await?;
         }
         Some(Subcommand::Login(mut login_cli)) => {
-            prepend_config_flags(&mut login_cli.config_overrides, cli.config_overrides);
+            prepend_config_flags(
+                &mut login_cli.config_overrides,
+                root_config_overrides.clone(),
+            );
             match login_cli.action {
                 Some(LoginSubcommand::Status) => {
                     run_login_status(login_cli.config_overrides).await;
@@ -177,11 +220,17 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
             }
         }
         Some(Subcommand::Logout(mut logout_cli)) => {
-            prepend_config_flags(&mut logout_cli.config_overrides, cli.config_overrides);
+            prepend_config_flags(
+                &mut logout_cli.config_overrides,
+                root_config_overrides.clone(),
+            );
             run_logout(logout_cli.config_overrides).await;
         }
         Some(Subcommand::Proto(mut proto_cli)) => {
-            prepend_config_flags(&mut proto_cli.config_overrides, cli.config_overrides);
+            prepend_config_flags(
+                &mut proto_cli.config_overrides,
+                root_config_overrides.clone(),
+            );
             proto::run_main(proto_cli).await?;
         }
         Some(Subcommand::Completion(completion_cli)) => {
@@ -189,7 +238,10 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
         }
         Some(Subcommand::Debug(debug_args)) => match debug_args.cmd {
             DebugCommand::Seatbelt(mut seatbelt_cli) => {
-                prepend_config_flags(&mut seatbelt_cli.config_overrides, cli.config_overrides);
+                prepend_config_flags(
+                    &mut seatbelt_cli.config_overrides,
+                    root_config_overrides.clone(),
+                );
                 codex_cli::debug_sandbox::run_command_under_seatbelt(
                     seatbelt_cli,
                     codex_linux_sandbox_exe,
@@ -197,7 +249,10 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
                 .await?;
             }
             DebugCommand::Landlock(mut landlock_cli) => {
-                prepend_config_flags(&mut landlock_cli.config_overrides, cli.config_overrides);
+                prepend_config_flags(
+                    &mut landlock_cli.config_overrides,
+                    root_config_overrides.clone(),
+                );
                 codex_cli::debug_sandbox::run_command_under_landlock(
                     landlock_cli,
                     codex_linux_sandbox_exe,
@@ -206,7 +261,10 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
             }
         },
         Some(Subcommand::Apply(mut apply_cli)) => {
-            prepend_config_flags(&mut apply_cli.config_overrides, cli.config_overrides);
+            prepend_config_flags(
+                &mut apply_cli.config_overrides,
+                root_config_overrides.clone(),
+            );
             run_apply_command(apply_cli, None).await?;
         }
         Some(Subcommand::GenerateTs(gen_cli)) => {

codex-rs/cli/src/proto.rs

@@ -2,6 +2,7 @@ use std::io::IsTerminal;
 
 use clap::Parser;
 use codex_common::CliConfigOverrides;
+use codex_core::AuthManager;
 use codex_core::ConversationManager;
 use codex_core::NewConversation;
 use codex_core::config::Config;
@@ -9,7 +10,6 @@ use codex_core::config::ConfigOverrides;
 use codex_core::protocol::Event;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::Submission;
-use codex_login::AuthManager;
 use tokio::io::AsyncBufReadExt;
 use tokio::io::BufReader;
 use tracing::error;
@@ -37,10 +37,8 @@ pub async fn run_main(opts: ProtoCli) -> anyhow::Result<()> {
 
     let config = Config::load_with_cli_overrides(overrides_vec, ConfigOverrides::default())?;
     // Use conversation_manager API to start a conversation
-    let conversation_manager = ConversationManager::new(AuthManager::shared(
-        config.codex_home.clone(),
-        config.preferred_auth_method,
-    ));
+    let conversation_manager =
+        ConversationManager::new(AuthManager::shared(config.codex_home.clone()));
     let NewConversation {
         conversation_id: _,
         conversation,

codex-rs/common/src/config_summary.rs

@@ -17,7 +17,10 @@ pub fn create_config_summary_entries(config: &Config) -> Vec<(&'static str, Stri
     {
         entries.push((
             "reasoning effort",
-            config.model_reasoning_effort.to_string(),
+            config
+                .model_reasoning_effort
+                .map(|effort| effort.to_string())
+                .unwrap_or_else(|| "none".to_string()),
         ));
         entries.push((
             "reasoning summaries",

codex-rs/common/src/elapsed.rs

@@ -2,7 +2,7 @@ use std::time::Duration;
 use std::time::Instant;
 
 /// Returns a string representing the elapsed time since `start_time` like
-/// "1m15s" or "1.50s".
+/// "1m 15s" or "1.50s".
 pub fn format_elapsed(start_time: Instant) -> String {
     format_duration(start_time.elapsed())
 }
@@ -12,7 +12,7 @@ pub fn format_elapsed(start_time: Instant) -> String {
 /// Formatting rules:
 /// * < 1 s  ->  "{milli}ms"
 /// * < 60 s ->  "{sec:.2}s" (two decimal places)
-/// * >= 60 s ->  "{min}m{sec:02}s"
+/// * >= 60 s ->  "{min}m {sec:02}s"
 pub fn format_duration(duration: Duration) -> String {
     let millis = duration.as_millis() as i64;
     format_elapsed_millis(millis)
@@ -26,7 +26,7 @@ fn format_elapsed_millis(millis: i64) -> String {
     } else {
         let minutes = millis / 60_000;
         let seconds = (millis % 60_000) / 1000;
-        format!("{minutes}m{seconds:02}s")
+        format!("{minutes}m {seconds:02}s")
     }
 }
 
@@ -61,12 +61,18 @@ mod tests {
     fn test_format_duration_minutes() {
         // Durations ≥ 1 minute should be printed mmss.
         let dur = Duration::from_millis(75_000); // 1m15s
-        assert_eq!(format_duration(dur), "1m15s");
+        assert_eq!(format_duration(dur), "1m 15s");
 
         let dur_exact = Duration::from_millis(60_000); // 1m0s
-        assert_eq!(format_duration(dur_exact), "1m00s");
+        assert_eq!(format_duration(dur_exact), "1m 00s");
 
         let dur_long = Duration::from_millis(3_601_000);
-        assert_eq!(format_duration(dur_long), "60m01s");
+        assert_eq!(format_duration(dur_long), "60m 01s");
+    }
+
+    #[test]
+    fn test_format_duration_one_hour_has_space() {
+        let dur_hour = Duration::from_millis(3_600_000);
+        assert_eq!(format_duration(dur_hour), "60m 00s");
     }
 }

codex-rs/common/src/model_presets.rs

@@ -1,4 +1,6 @@
+use codex_core::config::SWIFTFOX_MEDIUM_MODEL;
 use codex_core::protocol_config_types::ReasoningEffort;
+use codex_protocol::mcp_protocol::AuthMode;
 
 /// A simple preset pairing a model slug with a reasoning effort.
 #[derive(Debug, Clone, Copy)]
@@ -12,43 +14,68 @@ pub struct ModelPreset {
     /// Model slug (e.g., "gpt-5").
     pub model: &'static str,
     /// Reasoning effort to apply for this preset.
-    pub effort: ReasoningEffort,
+    pub effort: Option<ReasoningEffort>,
 }
 
-/// Built-in list of model presets that pair a model with a reasoning effort.
-///
-/// Keep this UI-agnostic so it can be reused by both TUI and MCP server.
-pub fn builtin_model_presets() -> &'static [ModelPreset] {
-    // Order reflects effort from minimal to high.
-    const PRESETS: &[ModelPreset] = &[
-        ModelPreset {
-            id: "gpt-5-minimal",
-            label: "gpt-5 minimal",
-            description: "— fastest responses with limited reasoning; ideal for coding, instructions, or lightweight tasks",
-            model: "gpt-5",
-            effort: ReasoningEffort::Minimal,
-        },
-        ModelPreset {
-            id: "gpt-5-low",
-            label: "gpt-5 low",
-            description: "— balances speed with some reasoning; useful for straightforward queries and short explanations",
-            model: "gpt-5",
-            effort: ReasoningEffort::Low,
-        },
-        ModelPreset {
-            id: "gpt-5-medium",
-            label: "gpt-5 medium",
-            description: "— default setting; provides a solid balance of reasoning depth and latency for general-purpose tasks",
-            model: "gpt-5",
-            effort: ReasoningEffort::Medium,
-        },
-        ModelPreset {
-            id: "gpt-5-high",
-            label: "gpt-5 high",
-            description: "— maximizes reasoning depth for complex or ambiguous problems",
-            model: "gpt-5",
-            effort: ReasoningEffort::High,
-        },
-    ];
-    PRESETS
+const PRESETS: &[ModelPreset] = &[
+    ModelPreset {
+        id: "swiftfox-low",
+        label: "swiftfox low",
+        description: "",
+        model: "swiftfox",
+        effort: Some(ReasoningEffort::Low),
+    },
+    ModelPreset {
+        id: "swiftfox-medium",
+        label: "swiftfox medium",
+        description: "",
+        model: "swiftfox",
+        effort: None,
+    },
+    ModelPreset {
+        id: "swiftfox-high",
+        label: "swiftfox high",
+        description: "",
+        model: "swiftfox",
+        effort: Some(ReasoningEffort::High),
+    },
+    ModelPreset {
+        id: "gpt-5-minimal",
+        label: "gpt-5 minimal",
+        description: "— fastest responses with limited reasoning; ideal for coding, instructions, or lightweight tasks",
+        model: "gpt-5",
+        effort: Some(ReasoningEffort::Minimal),
+    },
+    ModelPreset {
+        id: "gpt-5-low",
+        label: "gpt-5 low",
+        description: "— balances speed with some reasoning; useful for straightforward queries and short explanations",
+        model: "gpt-5",
+        effort: Some(ReasoningEffort::Low),
+    },
+    ModelPreset {
+        id: "gpt-5-medium",
+        label: "gpt-5 medium",
+        description: "— default setting; provides a solid balance of reasoning depth and latency for general-purpose tasks",
+        model: "gpt-5",
+        effort: Some(ReasoningEffort::Medium),
+    },
+    ModelPreset {
+        id: "gpt-5-high",
+        label: "gpt-5 high",
+        description: "— maximizes reasoning depth for complex or ambiguous problems",
+        model: "gpt-5",
+        effort: Some(ReasoningEffort::High),
+    },
+];
+
+pub fn builtin_model_presets(auth_mode: Option<AuthMode>) -> Vec<ModelPreset> {
+    match auth_mode {
+        Some(AuthMode::ApiKey) => PRESETS
+            .iter()
+            .copied()
+            .filter(|p| !p.model.contains(SWIFTFOX_MEDIUM_MODEL))
+            .collect(),
+        _ => PRESETS.to_vec(),
+    }
 }

codex-rs/core/Cargo.toml

@@ -13,12 +13,13 @@ workspace = true
 
 [dependencies]
 anyhow = "1"
+askama = "0.12"
 async-channel = "2.3.1"
 base64 = "0.22"
 bytes = "1.10.1"
 chrono = { version = "0.4", features = ["serde"] }
 codex-apply-patch = { path = "../apply-patch" }
-codex-login = { path = "../login" }
+codex-file-search = { path = "../file-search" }
 codex-mcp-client = { path = "../mcp-client" }
 codex-protocol = { path = "../protocol" }
 dirs = "6"
@@ -27,22 +28,20 @@ eventsource-stream = "0.2.3"
 futures = "0.3"
 libc = "0.2.175"
 mcp-types = { path = "../mcp-types" }
-mime_guess = "2.0"
 os_info = "3.12.0"
 portable-pty = "0.9.0"
 rand = "0.9"
 regex-lite = "0.1.7"
 reqwest = { version = "0.12", features = ["json", "stream"] }
 serde = { version = "1", features = ["derive"] }
-serde_bytes = "0.11"
 serde_json = "1"
 sha1 = "0.10.6"
 shlex = "1.3.0"
 similar = "2.7.0"
 strum_macros = "0.27.2"
 tempfile = "3"
-thiserror = "2.0.12"
-time = { version = "0.3", features = ["formatting", "local-offset", "macros"] }
+thiserror = "2.0.16"
+time = { version = "0.3", features = ["formatting", "parsing", "local-offset", "macros"] }
 tokio = { version = "1", features = [
     "io-std",
     "macros",
@@ -54,10 +53,10 @@ tokio-util = "0.7.16"
 toml = "0.9.5"
 toml_edit = "0.23.4"
 tracing = { version = "0.1.41", features = ["log"] }
-tree-sitter = "0.25.8"
+tree-sitter = "0.25.9"
 tree-sitter-bash = "0.25.0"
 uuid = { version = "1", features = ["serde", "v4"] }
-whoami = "1.6.1"
+which = "6"
 wildmatch = "2.4.0"
 
 
@@ -73,9 +72,6 @@ openssl-sys = { version = "*", features = ["vendored"] }
 [target.aarch64-unknown-linux-musl.dependencies]
 openssl-sys = { version = "*", features = ["vendored"] }
 
-[target.'cfg(target_os = "windows")'.dependencies]
-which = "6"
-
 [dev-dependencies]
 assert_cmd = "2"
 core_test_support = { path = "tests/common" }
@@ -86,3 +82,6 @@ tempfile = "3"
 tokio-test = "0.4"
 walkdir = "2.5.0"
 wiremock = "0.6"
+
+[package.metadata.cargo-shear]
+ignored = ["openssl-sys"]

codex-rs/core/prompt.md

@@ -14,6 +14,18 @@ Within this context, Codex refers to the open-source agentic coding interface (n
 
 Your default personality and tone is concise, direct, and friendly. You communicate efficiently, always keeping the user clearly informed about ongoing actions without unnecessary detail. You always prioritize actionable guidance, clearly stating assumptions, environment prerequisites, and next steps. Unless explicitly asked, you avoid excessively verbose explanations about your work.
 
+# AGENTS.md spec
+- Repos often contain AGENTS.md files. These files can appear anywhere within the repository.
+- These files are a way for humans to give you (the agent) instructions or tips for working within the container.
+- Some examples might be: coding conventions, info about how code is organized, or instructions for how to run or test code.
+- Instructions in AGENTS.md files:
+    - The scope of an AGENTS.md file is the entire directory tree rooted at the folder that contains it.
+    - For every file you touch in the final patch, you must obey instructions in any AGENTS.md file whose scope includes that file.
+    - Instructions about code style, structure, naming, etc. apply only to code within the AGENTS.md file's scope, unless the file states otherwise.
+    - More-deeply-nested AGENTS.md files take precedence in the case of conflicting instructions.
+    - Direct system/developer/user instructions (as part of a prompt) take precedence over AGENTS.md instructions.
+- The contents of the AGENTS.md file at the root of the repo and any directories from the CWD up to the root are included with the developer message and don't need to be re-read. When working in a subdirectory of CWD, or a directory outside the CWD, check for any AGENTS.md files that may be applicable.
+
 ## Responsiveness
 
 ### Preamble messages
@@ -228,7 +240,6 @@ You are producing plain text that will later be styled by the CLI. Follow these
 **Bullets**
 
 - Use `-` followed by a space for every bullet.
-- Bold the keyword, then colon + concise description.
 - Merge related points when possible; avoid a bullet for every trivial detail.
 - Keep bullets to one line unless breaking for clarity is unavoidable.
 - Group into short lists (4–6 bullets) ordered by importance.

codex-rs/core/review_prompt.md

@@ -0,0 +1,87 @@
+# Review guidelines:
+
+You are acting as a reviewer for a proposed code change made by another engineer.
+
+Below are some default guidelines for determining whether the original author would appreciate the issue being flagged.
+
+These are not the final word in determining whether an issue is a bug. In many cases, you will encounter other, more specific guidelines. These may be present elsewhere in a developer message, a user message, a file, or even elsewhere in this system message.
+Those guidelines should be considered to override these general instructions.
+
+Here are the general guidelines for determining whether something is a bug and should be flagged.
+
+1. It meaningfully impacts the accuracy, performance, security, or maintainability of the code.
+2. The bug is discrete and actionable (i.e. not a general issue with the codebase or a combination of multiple issues).
+3. Fixing the bug does not demand a level of rigor that is not present in the rest of the codebase (e.g. one doesn't need very detailed comments and input validation in a repository of one-off scripts in personal projects)
+4. The bug was introduced in the commit (pre-existing bugs should not be flagged).
+5. The author of the original PR would likely fix the issue if they were made aware of it.
+6. The bug does not rely on unstated assumptions about the codebase or author's intent.
+7. It is not enough to speculate that a change may disrupt another part of the codebase, to be considered a bug, one must identify the other parts of the code that are provably affected.
+8. The bug is clearly not just an intentional change by the original author.
+
+When flagging a bug, you will also provide an accompanying comment. Once again, these guidelines are not the final word on how to construct a comment -- defer to any subsequent guidelines that you encounter.
+
+1. The comment should be clear about why the issue is a bug.
+2. The comment should appropriately communicate the severity of the issue. It should not claim that an issue is more severe than it actually is.
+3. The comment should be brief. The body should be at most 1 paragraph. It should not introduce line breaks within the natural language flow unless it is necessary for the code fragment.
+4. The comment should not include any chunks of code longer than 3 lines. Any code chunks should be wrapped in markdown inline code tags or a code block.
+5. The comment should clearly and explicitly communicate the scenarios, environments, or inputs that are necessary for the bug to arise. The comment should immediately indicate that the issue's severity depends on these factors.
+6. The comment's tone should be matter-of-fact and not accusatory or overly positive. It should read as a helpful AI assistant suggestion without sounding too much like a human reviewer.
+7. The comment should be written such that the original author can immediately grasp the idea without close reading.
+8. The comment should avoid excessive flattery and comments that are not helpful to the original author. The comment should avoid phrasing like "Great job ...", "Thanks for ...".
+
+Below are some more detailed guidelines that you should apply to this specific review.
+
+HOW MANY FINDINGS TO RETURN:
+
+Output all findings that the original author would fix if they knew about it. If there is no finding that a person would definitely love to see and fix, prefer outputting no findings. Do not stop at the first qualifying finding. Continue until you've listed every qualifying finding.
+
+GUIDELINES:
+
+- Ignore trivial style unless it obscures meaning or violates documented standards.
+- Use one comment per distinct issue (or a multi-line range if necessary).
+- Use ```suggestion blocks ONLY for concrete replacement code (minimal lines; no commentary inside the block).
+- In every ```suggestion block, preserve the exact leading whitespace of the replaced lines (spaces vs tabs, number of spaces).
+- Do NOT introduce or remove outer indentation levels unless that is the actual fix.
+
+The comments will be presented in the code review as inline comments. You should avoid providing unnecessary location details in the comment body. Always keep the line range as short as possible for interpreting the issue. Avoid ranges longer than 5–10 lines; instead, choose the most suitable subrange that pinpoints the problem.
+
+At the beginning of the finding title, tag the bug with priority level. For example "[P1] Un-padding slices along wrong tensor dimensions". [P0] – Drop everything to fix.  Blocking release, operations, or major usage. Only use for universal issues that do not depend on any assumptions about the inputs. · [P1] – Urgent. Should be addressed in the next cycle · [P2] – Normal. To be fixed eventually · [P3] – Low. Nice to have.
+
+Additionally, include a numeric priority field in the JSON output for each finding: set "priority" to 0 for P0, 1 for P1, 2 for P2, or 3 for P3. If a priority cannot be determined, omit the field or use null.
+
+At the end of your findings, output an "overall correctness" verdict of whether or not the patch should be considered "correct".
+Correct implies that existing code and tests will not break, and the patch is free of bugs and other blocking issues.
+Ignore non-blocking issues such as style, formatting, typos, documentation, and other nits.
+
+FORMATTING GUIDELINES:
+The finding description should be one paragraph.
+
+OUTPUT FORMAT:
+
+## Output schema  — MUST MATCH *exactly*
+
+```json
+{
+  "findings": [
+    {
+      "title": "<≤ 80 chars, imperative>",
+      "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>",
+      "confidence_score": <float 0.0-1.0>,
+      "priority": <int 0-3, optional>,
+      "code_location": {
+        "absolute_file_path": "<file path>",
+        "line_range": {"start": <int>, "end": <int>}
+      }
+    }
+  ],
+  "overall_correctness": "patch is correct" | "patch is incorrect",
+  "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>",
+  "overall_confidence_score": <float 0.0-1.0>
+}
+```
+
+* **Do not** wrap the JSON in markdown fences or extra prose.
+* The code_location field is required and must include absolute_file_path and line_range.
+*Line ranges must be as short as possible for interpreting the issue (avoid ranges over 5–10 lines; pick the most suitable subrange).
+* The code_location should overlap with the diff.
+* Do not generate a PR fix.

codex-rs/core/src/auth.rs

@@ -0,0 +1,659 @@
+use chrono::DateTime;
+use chrono::Utc;
+use serde::Deserialize;
+use serde::Serialize;
+use std::env;
+use std::fs::File;
+use std::fs::OpenOptions;
+use std::io::Read;
+use std::io::Write;
+#[cfg(unix)]
+use std::os::unix::fs::OpenOptionsExt;
+use std::path::Path;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::sync::Mutex;
+use std::time::Duration;
+
+use codex_protocol::mcp_protocol::AuthMode;
+
+use crate::token_data::PlanType;
+use crate::token_data::TokenData;
+use crate::token_data::parse_id_token;
+
+#[derive(Debug, Clone)]
+pub struct CodexAuth {
+    pub mode: AuthMode,
+
+    pub(crate) api_key: Option<String>,
+    pub(crate) auth_dot_json: Arc<Mutex<Option<AuthDotJson>>>,
+    pub(crate) auth_file: PathBuf,
+    pub(crate) client: reqwest::Client,
+}
+
+impl PartialEq for CodexAuth {
+    fn eq(&self, other: &Self) -> bool {
+        self.mode == other.mode
+    }
+}
+
+impl CodexAuth {
+    pub async fn refresh_token(&self) -> Result<String, std::io::Error> {
+        let token_data = self
+            .get_current_token_data()
+            .ok_or(std::io::Error::other("Token data is not available."))?;
+        let token = token_data.refresh_token;
+
+        let refresh_response = try_refresh_token(token, &self.client)
+            .await
+            .map_err(std::io::Error::other)?;
+
+        let updated = update_tokens(
+            &self.auth_file,
+            refresh_response.id_token,
+            refresh_response.access_token,
+            refresh_response.refresh_token,
+        )
+        .await?;
+
+        if let Ok(mut auth_lock) = self.auth_dot_json.lock() {
+            *auth_lock = Some(updated.clone());
+        }
+
+        let access = match updated.tokens {
+            Some(t) => t.access_token,
+            None => {
+                return Err(std::io::Error::other(
+                    "Token data is not available after refresh.",
+                ));
+            }
+        };
+        Ok(access)
+    }
+
+    /// Loads the available auth information from the auth.json.
+    pub fn from_codex_home(codex_home: &Path) -> std::io::Result<Option<CodexAuth>> {
+        load_auth(codex_home)
+    }
+
+    pub async fn get_token_data(&self) -> Result<TokenData, std::io::Error> {
+        let auth_dot_json: Option<AuthDotJson> = self.get_current_auth_json();
+        match auth_dot_json {
+            Some(AuthDotJson {
+                tokens: Some(mut tokens),
+                last_refresh: Some(last_refresh),
+                ..
+            }) => {
+                if last_refresh < Utc::now() - chrono::Duration::days(28) {
+                    let refresh_response = tokio::time::timeout(
+                        Duration::from_secs(60),
+                        try_refresh_token(tokens.refresh_token.clone(), &self.client),
+                    )
+                    .await
+                    .map_err(|_| {
+                        std::io::Error::other("timed out while refreshing OpenAI API key")
+                    })?
+                    .map_err(std::io::Error::other)?;
+
+                    let updated_auth_dot_json = update_tokens(
+                        &self.auth_file,
+                        refresh_response.id_token,
+                        refresh_response.access_token,
+                        refresh_response.refresh_token,
+                    )
+                    .await?;
+
+                    tokens = updated_auth_dot_json
+                        .tokens
+                        .clone()
+                        .ok_or(std::io::Error::other(
+                            "Token data is not available after refresh.",
+                        ))?;
+
+                    #[expect(clippy::unwrap_used)]
+                    let mut auth_lock = self.auth_dot_json.lock().unwrap();
+                    *auth_lock = Some(updated_auth_dot_json);
+                }
+
+                Ok(tokens)
+            }
+            _ => Err(std::io::Error::other("Token data is not available.")),
+        }
+    }
+
+    pub async fn get_token(&self) -> Result<String, std::io::Error> {
+        match self.mode {
+            AuthMode::ApiKey => Ok(self.api_key.clone().unwrap_or_default()),
+            AuthMode::ChatGPT => {
+                let id_token = self.get_token_data().await?.access_token;
+                Ok(id_token)
+            }
+        }
+    }
+
+    pub fn get_account_id(&self) -> Option<String> {
+        self.get_current_token_data().and_then(|t| t.account_id)
+    }
+
+    pub(crate) fn get_plan_type(&self) -> Option<PlanType> {
+        self.get_current_token_data()
+            .and_then(|t| t.id_token.chatgpt_plan_type)
+    }
+
+    fn get_current_auth_json(&self) -> Option<AuthDotJson> {
+        #[expect(clippy::unwrap_used)]
+        self.auth_dot_json.lock().unwrap().clone()
+    }
+
+    fn get_current_token_data(&self) -> Option<TokenData> {
+        self.get_current_auth_json().and_then(|t| t.tokens)
+    }
+
+    /// Consider this private to integration tests.
+    pub fn create_dummy_chatgpt_auth_for_testing() -> Self {
+        let auth_dot_json = AuthDotJson {
+            openai_api_key: None,
+            tokens: Some(TokenData {
+                id_token: Default::default(),
+                access_token: "Access Token".to_string(),
+                refresh_token: "test".to_string(),
+                account_id: Some("account_id".to_string()),
+            }),
+            last_refresh: Some(Utc::now()),
+        };
+
+        let auth_dot_json = Arc::new(Mutex::new(Some(auth_dot_json)));
+        Self {
+            api_key: None,
+            mode: AuthMode::ChatGPT,
+            auth_file: PathBuf::new(),
+            auth_dot_json,
+            client: crate::default_client::create_client(),
+        }
+    }
+
+    fn from_api_key_with_client(api_key: &str, client: reqwest::Client) -> Self {
+        Self {
+            api_key: Some(api_key.to_owned()),
+            mode: AuthMode::ApiKey,
+            auth_file: PathBuf::new(),
+            auth_dot_json: Arc::new(Mutex::new(None)),
+            client,
+        }
+    }
+
+    pub fn from_api_key(api_key: &str) -> Self {
+        Self::from_api_key_with_client(api_key, crate::default_client::create_client())
+    }
+}
+
+pub const OPENAI_API_KEY_ENV_VAR: &str = "OPENAI_API_KEY";
+
+pub fn read_openai_api_key_from_env() -> Option<String> {
+    env::var(OPENAI_API_KEY_ENV_VAR)
+        .ok()
+        .map(|value| value.trim().to_string())
+        .filter(|value| !value.is_empty())
+}
+
+pub fn get_auth_file(codex_home: &Path) -> PathBuf {
+    codex_home.join("auth.json")
+}
+
+/// Delete the auth.json file inside `codex_home` if it exists. Returns `Ok(true)`
+/// if a file was removed, `Ok(false)` if no auth file was present.
+pub fn logout(codex_home: &Path) -> std::io::Result<bool> {
+    let auth_file = get_auth_file(codex_home);
+    match std::fs::remove_file(&auth_file) {
+        Ok(_) => Ok(true),
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => Ok(false),
+        Err(err) => Err(err),
+    }
+}
+
+/// Writes an `auth.json` that contains only the API key.
+pub fn login_with_api_key(codex_home: &Path, api_key: &str) -> std::io::Result<()> {
+    let auth_dot_json = AuthDotJson {
+        openai_api_key: Some(api_key.to_string()),
+        tokens: None,
+        last_refresh: None,
+    };
+    write_auth_json(&get_auth_file(codex_home), &auth_dot_json)
+}
+
+fn load_auth(codex_home: &Path) -> std::io::Result<Option<CodexAuth>> {
+    let auth_file = get_auth_file(codex_home);
+    let client = crate::default_client::create_client();
+    let auth_dot_json = match try_read_auth_json(&auth_file) {
+        Ok(auth) => auth,
+        Err(e) => {
+            return Err(e);
+        }
+    };
+
+    let AuthDotJson {
+        openai_api_key: auth_json_api_key,
+        tokens,
+        last_refresh,
+    } = auth_dot_json;
+
+    // Prefer AuthMode.ApiKey if it's set in the auth.json.
+    if let Some(api_key) = &auth_json_api_key {
+        return Ok(Some(CodexAuth::from_api_key_with_client(api_key, client)));
+    }
+
+    Ok(Some(CodexAuth {
+        api_key: None,
+        mode: AuthMode::ChatGPT,
+        auth_file,
+        auth_dot_json: Arc::new(Mutex::new(Some(AuthDotJson {
+            openai_api_key: None,
+            tokens,
+            last_refresh,
+        }))),
+        client,
+    }))
+}
+
+/// Attempt to read and refresh the `auth.json` file in the given `CODEX_HOME` directory.
+/// Returns the full AuthDotJson structure after refreshing if necessary.
+pub fn try_read_auth_json(auth_file: &Path) -> std::io::Result<AuthDotJson> {
+    let mut file = File::open(auth_file)?;
+    let mut contents = String::new();
+    file.read_to_string(&mut contents)?;
+    let auth_dot_json: AuthDotJson = serde_json::from_str(&contents)?;
+
+    Ok(auth_dot_json)
+}
+
+pub fn write_auth_json(auth_file: &Path, auth_dot_json: &AuthDotJson) -> std::io::Result<()> {
+    let json_data = serde_json::to_string_pretty(auth_dot_json)?;
+    let mut options = OpenOptions::new();
+    options.truncate(true).write(true).create(true);
+    #[cfg(unix)]
+    {
+        options.mode(0o600);
+    }
+    let mut file = options.open(auth_file)?;
+    file.write_all(json_data.as_bytes())?;
+    file.flush()?;
+    Ok(())
+}
+
+async fn update_tokens(
+    auth_file: &Path,
+    id_token: String,
+    access_token: Option<String>,
+    refresh_token: Option<String>,
+) -> std::io::Result<AuthDotJson> {
+    let mut auth_dot_json = try_read_auth_json(auth_file)?;
+
+    let tokens = auth_dot_json.tokens.get_or_insert_with(TokenData::default);
+    tokens.id_token = parse_id_token(&id_token).map_err(std::io::Error::other)?;
+    if let Some(access_token) = access_token {
+        tokens.access_token = access_token;
+    }
+    if let Some(refresh_token) = refresh_token {
+        tokens.refresh_token = refresh_token;
+    }
+    auth_dot_json.last_refresh = Some(Utc::now());
+    write_auth_json(auth_file, &auth_dot_json)?;
+    Ok(auth_dot_json)
+}
+
+async fn try_refresh_token(
+    refresh_token: String,
+    client: &reqwest::Client,
+) -> std::io::Result<RefreshResponse> {
+    let refresh_request = RefreshRequest {
+        client_id: CLIENT_ID,
+        grant_type: "refresh_token",
+        refresh_token,
+        scope: "openid profile email",
+    };
+
+    // Use shared client factory to include standard headers
+    let response = client
+        .post("https://auth.openai.com/oauth/token")
+        .header("Content-Type", "application/json")
+        .json(&refresh_request)
+        .send()
+        .await
+        .map_err(std::io::Error::other)?;
+
+    if response.status().is_success() {
+        let refresh_response = response
+            .json::<RefreshResponse>()
+            .await
+            .map_err(std::io::Error::other)?;
+        Ok(refresh_response)
+    } else {
+        Err(std::io::Error::other(format!(
+            "Failed to refresh token: {}",
+            response.status()
+        )))
+    }
+}
+
+#[derive(Serialize)]
+struct RefreshRequest {
+    client_id: &'static str,
+    grant_type: &'static str,
+    refresh_token: String,
+    scope: &'static str,
+}
+
+#[derive(Deserialize, Clone)]
+struct RefreshResponse {
+    id_token: String,
+    access_token: Option<String>,
+    refresh_token: Option<String>,
+}
+
+/// Expected structure for $CODEX_HOME/auth.json.
+#[derive(Deserialize, Serialize, Clone, Debug, PartialEq)]
+pub struct AuthDotJson {
+    #[serde(rename = "OPENAI_API_KEY")]
+    pub openai_api_key: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tokens: Option<TokenData>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub last_refresh: Option<DateTime<Utc>>,
+}
+
+// Shared constant for token refresh (client id used for oauth token refresh flow)
+pub const CLIENT_ID: &str = "app_EMoamEEZ73f0CkXaXp7hrann";
+
+use std::sync::RwLock;
+
+/// Internal cached auth state.
+#[derive(Clone, Debug)]
+struct CachedAuth {
+    auth: Option<CodexAuth>,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::token_data::IdTokenInfo;
+    use crate::token_data::KnownPlan;
+    use crate::token_data::PlanType;
+    use base64::Engine;
+    use pretty_assertions::assert_eq;
+    use serde::Serialize;
+    use serde_json::json;
+    use tempfile::tempdir;
+
+    const LAST_REFRESH: &str = "2025-08-06T20:41:36.232376Z";
+
+    #[tokio::test]
+    async fn roundtrip_auth_dot_json() {
+        let codex_home = tempdir().unwrap();
+        let _ = write_auth_file(
+            AuthFileParams {
+                openai_api_key: None,
+                chatgpt_plan_type: "pro".to_string(),
+            },
+            codex_home.path(),
+        )
+        .expect("failed to write auth file");
+
+        let file = get_auth_file(codex_home.path());
+        let auth_dot_json = try_read_auth_json(&file).unwrap();
+        write_auth_json(&file, &auth_dot_json).unwrap();
+
+        let same_auth_dot_json = try_read_auth_json(&file).unwrap();
+        assert_eq!(auth_dot_json, same_auth_dot_json);
+    }
+
+    #[test]
+    fn login_with_api_key_overwrites_existing_auth_json() {
+        let dir = tempdir().unwrap();
+        let auth_path = dir.path().join("auth.json");
+        let stale_auth = json!({
+            "OPENAI_API_KEY": "sk-old",
+            "tokens": {
+                "id_token": "stale.header.payload",
+                "access_token": "stale-access",
+                "refresh_token": "stale-refresh",
+                "account_id": "stale-acc"
+            }
+        });
+        std::fs::write(
+            &auth_path,
+            serde_json::to_string_pretty(&stale_auth).unwrap(),
+        )
+        .unwrap();
+
+        super::login_with_api_key(dir.path(), "sk-new").expect("login_with_api_key should succeed");
+
+        let auth = super::try_read_auth_json(&auth_path).expect("auth.json should parse");
+        assert_eq!(auth.openai_api_key.as_deref(), Some("sk-new"));
+        assert!(auth.tokens.is_none(), "tokens should be cleared");
+    }
+
+    #[tokio::test]
+    async fn pro_account_with_no_api_key_uses_chatgpt_auth() {
+        let codex_home = tempdir().unwrap();
+        let fake_jwt = write_auth_file(
+            AuthFileParams {
+                openai_api_key: None,
+                chatgpt_plan_type: "pro".to_string(),
+            },
+            codex_home.path(),
+        )
+        .expect("failed to write auth file");
+
+        let CodexAuth {
+            api_key,
+            mode,
+            auth_dot_json,
+            auth_file: _,
+            ..
+        } = super::load_auth(codex_home.path()).unwrap().unwrap();
+        assert_eq!(None, api_key);
+        assert_eq!(AuthMode::ChatGPT, mode);
+
+        let guard = auth_dot_json.lock().unwrap();
+        let auth_dot_json = guard.as_ref().expect("AuthDotJson should exist");
+        assert_eq!(
+            &AuthDotJson {
+                openai_api_key: None,
+                tokens: Some(TokenData {
+                    id_token: IdTokenInfo {
+                        email: Some("user@example.com".to_string()),
+                        chatgpt_plan_type: Some(PlanType::Known(KnownPlan::Pro)),
+                        raw_jwt: fake_jwt,
+                    },
+                    access_token: "test-access-token".to_string(),
+                    refresh_token: "test-refresh-token".to_string(),
+                    account_id: None,
+                }),
+                last_refresh: Some(
+                    DateTime::parse_from_rfc3339(LAST_REFRESH)
+                        .unwrap()
+                        .with_timezone(&Utc)
+                ),
+            },
+            auth_dot_json
+        )
+    }
+
+    #[tokio::test]
+    async fn loads_api_key_from_auth_json() {
+        let dir = tempdir().unwrap();
+        let auth_file = dir.path().join("auth.json");
+        std::fs::write(
+            auth_file,
+            r#"{"OPENAI_API_KEY":"sk-test-key","tokens":null,"last_refresh":null}"#,
+        )
+        .unwrap();
+
+        let auth = super::load_auth(dir.path()).unwrap().unwrap();
+        assert_eq!(auth.mode, AuthMode::ApiKey);
+        assert_eq!(auth.api_key, Some("sk-test-key".to_string()));
+
+        assert!(auth.get_token_data().await.is_err());
+    }
+
+    #[test]
+    fn logout_removes_auth_file() -> Result<(), std::io::Error> {
+        let dir = tempdir()?;
+        let auth_dot_json = AuthDotJson {
+            openai_api_key: Some("sk-test-key".to_string()),
+            tokens: None,
+            last_refresh: None,
+        };
+        write_auth_json(&get_auth_file(dir.path()), &auth_dot_json)?;
+        assert!(dir.path().join("auth.json").exists());
+        let removed = logout(dir.path())?;
+        assert!(removed);
+        assert!(!dir.path().join("auth.json").exists());
+        Ok(())
+    }
+
+    struct AuthFileParams {
+        openai_api_key: Option<String>,
+        chatgpt_plan_type: String,
+    }
+
+    fn write_auth_file(params: AuthFileParams, codex_home: &Path) -> std::io::Result<String> {
+        let auth_file = get_auth_file(codex_home);
+        // Create a minimal valid JWT for the id_token field.
+        #[derive(Serialize)]
+        struct Header {
+            alg: &'static str,
+            typ: &'static str,
+        }
+        let header = Header {
+            alg: "none",
+            typ: "JWT",
+        };
+        let payload = serde_json::json!({
+            "email": "user@example.com",
+            "email_verified": true,
+            "https://api.openai.com/auth": {
+                "chatgpt_account_id": "bc3618e3-489d-4d49-9362-1561dc53ba53",
+                "chatgpt_plan_type": params.chatgpt_plan_type,
+                "chatgpt_user_id": "user-12345",
+                "user_id": "user-12345",
+            }
+        });
+        let b64 = |b: &[u8]| base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(b);
+        let header_b64 = b64(&serde_json::to_vec(&header)?);
+        let payload_b64 = b64(&serde_json::to_vec(&payload)?);
+        let signature_b64 = b64(b"sig");
+        let fake_jwt = format!("{header_b64}.{payload_b64}.{signature_b64}");
+
+        let auth_json_data = json!({
+            "OPENAI_API_KEY": params.openai_api_key,
+            "tokens": {
+                "id_token": fake_jwt,
+                "access_token": "test-access-token",
+                "refresh_token": "test-refresh-token"
+            },
+            "last_refresh": LAST_REFRESH,
+        });
+        let auth_json = serde_json::to_string_pretty(&auth_json_data)?;
+        std::fs::write(auth_file, auth_json)?;
+        Ok(fake_jwt)
+    }
+}
+
+/// Central manager providing a single source of truth for auth.json derived
+/// authentication data. It loads once (or on preference change) and then
+/// hands out cloned `CodexAuth` values so the rest of the program has a
+/// consistent snapshot.
+///
+/// External modifications to `auth.json` will NOT be observed until
+/// `reload()` is called explicitly. This matches the design goal of avoiding
+/// different parts of the program seeing inconsistent auth data mid‑run.
+#[derive(Debug)]
+pub struct AuthManager {
+    codex_home: PathBuf,
+    inner: RwLock<CachedAuth>,
+}
+
+impl AuthManager {
+    /// Create a new manager loading the initial auth using the provided
+    /// preferred auth method. Errors loading auth are swallowed; `auth()` will
+    /// simply return `None` in that case so callers can treat it as an
+    /// unauthenticated state.
+    pub fn new(codex_home: PathBuf) -> Self {
+        let auth = CodexAuth::from_codex_home(&codex_home).ok().flatten();
+        Self {
+            codex_home,
+            inner: RwLock::new(CachedAuth { auth }),
+        }
+    }
+
+    /// Create an AuthManager with a specific CodexAuth, for testing only.
+    pub fn from_auth_for_testing(auth: CodexAuth) -> Arc<Self> {
+        let cached = CachedAuth { auth: Some(auth) };
+        Arc::new(Self {
+            codex_home: PathBuf::new(),
+            inner: RwLock::new(cached),
+        })
+    }
+
+    /// Current cached auth (clone). May be `None` if not logged in or load failed.
+    pub fn auth(&self) -> Option<CodexAuth> {
+        self.inner.read().ok().and_then(|c| c.auth.clone())
+    }
+
+    /// Force a reload of the auth information from auth.json. Returns
+    /// whether the auth value changed.
+    pub fn reload(&self) -> bool {
+        let new_auth = CodexAuth::from_codex_home(&self.codex_home).ok().flatten();
+        if let Ok(mut guard) = self.inner.write() {
+            let changed = !AuthManager::auths_equal(&guard.auth, &new_auth);
+            guard.auth = new_auth;
+            changed
+        } else {
+            false
+        }
+    }
+
+    fn auths_equal(a: &Option<CodexAuth>, b: &Option<CodexAuth>) -> bool {
+        match (a, b) {
+            (None, None) => true,
+            (Some(a), Some(b)) => a == b,
+            _ => false,
+        }
+    }
+
+    /// Convenience constructor returning an `Arc` wrapper.
+    pub fn shared(codex_home: PathBuf) -> Arc<Self> {
+        Arc::new(Self::new(codex_home))
+    }
+
+    /// Attempt to refresh the current auth token (if any). On success, reload
+    /// the auth state from disk so other components observe refreshed token.
+    pub async fn refresh_token(&self) -> std::io::Result<Option<String>> {
+        let auth = match self.auth() {
+            Some(a) => a,
+            None => return Ok(None),
+        };
+        match auth.refresh_token().await {
+            Ok(token) => {
+                // Reload to pick up persisted changes.
+                self.reload();
+                Ok(Some(token))
+            }
+            Err(e) => Err(e),
+        }
+    }
+
+    /// Log out by deleting the on‑disk auth.json (if present). Returns Ok(true)
+    /// if a file was removed, Ok(false) if no auth file existed. On success,
+    /// reloads the in‑memory auth cache so callers immediately observe the
+    /// unauthenticated state.
+    pub fn logout(&self) -> std::io::Result<bool> {
+        let removed = super::auth::logout(&self.codex_home)?;
+        // Always reload to clear any cached auth (even if file absent).
+        self.reload();
+        Ok(removed)
+    }
+}

codex-rs/core/src/bash.rs

@@ -1,3 +1,4 @@
+use tree_sitter::Node;
 use tree_sitter::Parser;
 use tree_sitter::Tree;
 use tree_sitter_bash::LANGUAGE as BASH;
@@ -59,9 +60,6 @@ pub fn try_parse_word_only_commands_sequence(tree: &Tree, src: &str) -> Option<V
             }
         } else {
             // Reject any punctuation / operator tokens that are not explicitly allowed.
-            if kind.chars().any(|c| "&;|".contains(c)) && !ALLOWED_PUNCT_TOKENS.contains(&kind) {
-                return None;
-            }
             if !(ALLOWED_PUNCT_TOKENS.contains(&kind) || kind.trim().is_empty()) {
                 // If it's a quote token or operator it's allowed above; we also allow whitespace tokens.
                 // Any other punctuation like parentheses, braces, redirects, backticks, etc are rejected.
@@ -75,7 +73,7 @@ pub fn try_parse_word_only_commands_sequence(tree: &Tree, src: &str) -> Option<V
 
     let mut commands = Vec::new();
     for node in command_nodes {
-        if let Some(words) = parse_plain_command_from_node(node, src) {
+        if let Some(words) = extract_words_from_command_node(node, src) {
             commands.push(words);
         } else {
             return None;
@@ -84,7 +82,10 @@ pub fn try_parse_word_only_commands_sequence(tree: &Tree, src: &str) -> Option<V
     Some(commands)
 }
 
-fn parse_plain_command_from_node(cmd: tree_sitter::Node, src: &str) -> Option<Vec<String>> {
+/// Extract the plain words of a simple command node, normalizing quoted
+/// strings into their contents. Returns None if the node contains unsupported
+/// constructs for a word-only command.
+pub(crate) fn extract_words_from_command_node(cmd: Node, src: &str) -> Option<Vec<String>> {
     if cmd.kind() != "command" {
         return None;
     }
@@ -130,6 +131,48 @@ fn parse_plain_command_from_node(cmd: tree_sitter::Node, src: &str) -> Option<Ve
     Some(words)
 }
 
+/// Find the earliest `command` node in source order within the parse tree.
+pub(crate) fn find_first_command_node(tree: &Tree) -> Option<Node<'_>> {
+    let root = tree.root_node();
+    let mut cursor = root.walk();
+    let mut stack = vec![root];
+    let mut best: Option<Node> = None;
+    while let Some(node) = stack.pop() {
+        if node.is_named()
+            && node.kind() == "command"
+            && best.is_none_or(|b| node.start_byte() < b.start_byte())
+        {
+            best = Some(node);
+        }
+        for child in node.children(&mut cursor) {
+            stack.push(child);
+        }
+    }
+    best
+}
+
+/// Given the first command node, return the byte index in `src` at which the
+/// remainder script starts, only if the next non-whitespace token is an allowed
+/// sequencing operator for dropping the leading `cd`.
+///
+/// Allowed operators: `&&` (conditional on success) and `;` (unconditional).
+/// Disallowed: `||`, `|` — removing `cd` would change semantics.
+pub(crate) fn remainder_start_after_wrapper_operator(first_cmd: Node, src: &str) -> Option<usize> {
+    let mut sib = first_cmd.next_sibling()?;
+    while !sib.is_named() && sib.kind().trim().is_empty() {
+        sib = sib.next_sibling()?;
+    }
+    if sib.is_named() || (sib.kind() != "&&" && sib.kind() != ";") {
+        return None;
+    }
+    let mut idx = sib.end_byte();
+    let bytes = src.as_bytes();
+    while idx < bytes.len() && bytes[idx].is_ascii_whitespace() {
+        idx += 1;
+    }
+    if idx >= bytes.len() { None } else { Some(idx) }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -215,4 +258,45 @@ mod tests {
     fn rejects_trailing_operator_parse_error() {
         assert!(parse_seq("ls &&").is_none());
     }
+
+    #[test]
+    fn find_first_command_node_finds_cd() {
+        let src = "cd foo && ls; git status";
+        let tree = try_parse_bash(src).unwrap();
+        let first = find_first_command_node(&tree).unwrap();
+        let words = extract_words_from_command_node(first, src).unwrap();
+        assert_eq!(words, vec!["cd".to_string(), "foo".to_string()]);
+    }
+
+    #[test]
+    fn remainder_after_wrapper_operator_allows_and_and_semicolon() {
+        // Allows &&
+        let src = "cd foo && ls; git status";
+        let tree = try_parse_bash(src).unwrap();
+        let first = find_first_command_node(&tree).unwrap();
+        let idx = remainder_start_after_wrapper_operator(first, src).unwrap();
+        assert_eq!(&src[idx..], "ls; git status");
+
+        // Allows ;
+        let src2 = "cd foo; ls";
+        let tree2 = try_parse_bash(src2).unwrap();
+        let first2 = find_first_command_node(&tree2).unwrap();
+        let idx2 = remainder_start_after_wrapper_operator(first2, src2).unwrap();
+        assert_eq!(&src2[idx2..], "ls");
+    }
+
+    #[test]
+    fn remainder_after_wrapper_operator_rejects_or_and_pipe() {
+        // Rejects ||
+        let src = "cd foo || echo hi";
+        let tree = try_parse_bash(src).unwrap();
+        let first = find_first_command_node(&tree).unwrap();
+        assert!(remainder_start_after_wrapper_operator(first, src).is_none());
+
+        // Rejects |
+        let src2 = "cd foo | rg bar";
+        let tree2 = try_parse_bash(src2).unwrap();
+        let first2 = find_first_command_node(&tree2).unwrap();
+        assert!(remainder_start_after_wrapper_operator(first2, src2).is_none());
+    }
 }

codex-rs/core/src/chat_completions.rs

@@ -19,7 +19,6 @@ use crate::ModelProviderInfo;
 use crate::client_common::Prompt;
 use crate::client_common::ResponseEvent;
 use crate::client_common::ResponseStream;
-use crate::config::Config;
 use crate::error::CodexErr;
 use crate::error::Result;
 use crate::model_family::ModelFamily;
@@ -35,7 +34,6 @@ pub(crate) async fn stream_chat_completions(
     model_family: &ModelFamily,
     client: &reqwest::Client,
     provider: &ModelProviderInfo,
-    config: &Config,
 ) -> Result<ResponseStream> {
     // Build messages array
     let mut messages = Vec::<serde_json::Value>::new();
@@ -45,7 +43,107 @@ pub(crate) async fn stream_chat_completions(
 
     let input = prompt.get_formatted_input();
 
+    // Pre-scan: map Reasoning blocks to the adjacent assistant anchor after the last user.
+    // - If the last emitted message is a user message, drop all reasoning.
+    // - Otherwise, for each Reasoning item after the last user message, attach it
+    //   to the immediate previous assistant message (stop turns) or the immediate
+    //   next assistant anchor (tool-call turns: function/local shell call, or assistant message).
+    let mut reasoning_by_anchor_index: std::collections::HashMap<usize, String> =
+        std::collections::HashMap::new();
+
+    // Determine the last role that would be emitted to Chat Completions.
+    let mut last_emitted_role: Option<&str> = None;
     for item in &input {
+        match item {
+            ResponseItem::Message { role, .. } => last_emitted_role = Some(role.as_str()),
+            ResponseItem::FunctionCall { .. } | ResponseItem::LocalShellCall { .. } => {
+                last_emitted_role = Some("assistant")
+            }
+            ResponseItem::FunctionCallOutput { .. } => last_emitted_role = Some("tool"),
+            ResponseItem::Reasoning { .. } | ResponseItem::Other => {}
+            ResponseItem::CustomToolCall { .. } => {}
+            ResponseItem::CustomToolCallOutput { .. } => {}
+            ResponseItem::WebSearchCall { .. } => {}
+        }
+    }
+
+    // Find the last user message index in the input.
+    let mut last_user_index: Option<usize> = None;
+    for (idx, item) in input.iter().enumerate() {
+        if let ResponseItem::Message { role, .. } = item
+            && role == "user"
+        {
+            last_user_index = Some(idx);
+        }
+    }
+
+    // Attach reasoning only if the conversation does not end with a user message.
+    if !matches!(last_emitted_role, Some("user")) {
+        for (idx, item) in input.iter().enumerate() {
+            // Only consider reasoning that appears after the last user message.
+            if let Some(u_idx) = last_user_index
+                && idx <= u_idx
+            {
+                continue;
+            }
+
+            if let ResponseItem::Reasoning {
+                content: Some(items),
+                ..
+            } = item
+            {
+                let mut text = String::new();
+                for c in items {
+                    match c {
+                        ReasoningItemContent::ReasoningText { text: t }
+                        | ReasoningItemContent::Text { text: t } => text.push_str(t),
+                    }
+                }
+                if text.trim().is_empty() {
+                    continue;
+                }
+
+                // Prefer immediate previous assistant message (stop turns)
+                let mut attached = false;
+                if idx > 0
+                    && let ResponseItem::Message { role, .. } = &input[idx - 1]
+                    && role == "assistant"
+                {
+                    reasoning_by_anchor_index
+                        .entry(idx - 1)
+                        .and_modify(|v| v.push_str(&text))
+                        .or_insert(text.clone());
+                    attached = true;
+                }
+
+                // Otherwise, attach to immediate next assistant anchor (tool-calls or assistant message)
+                if !attached && idx + 1 < input.len() {
+                    match &input[idx + 1] {
+                        ResponseItem::FunctionCall { .. } | ResponseItem::LocalShellCall { .. } => {
+                            reasoning_by_anchor_index
+                                .entry(idx + 1)
+                                .and_modify(|v| v.push_str(&text))
+                                .or_insert(text.clone());
+                        }
+                        ResponseItem::Message { role, .. } if role == "assistant" => {
+                            reasoning_by_anchor_index
+                                .entry(idx + 1)
+                                .and_modify(|v| v.push_str(&text))
+                                .or_insert(text.clone());
+                        }
+                        _ => {}
+                    }
+                }
+            }
+        }
+    }
+
+    // Track last assistant text we emitted to avoid duplicate assistant messages
+    // in the outbound Chat Completions payload (can happen if a final
+    // aggregated assistant message was recorded alongside an earlier partial).
+    let mut last_assistant_text: Option<String> = None;
+
+    for (idx, item) in input.iter().enumerate() {
         match item {
             ResponseItem::Message { role, content, .. } => {
                 let mut text = String::new();
@@ -58,7 +156,24 @@ pub(crate) async fn stream_chat_completions(
                         _ => {}
                     }
                 }
-                messages.push(json!({"role": role, "content": text}));
+                // Skip exact-duplicate assistant messages.
+                if role == "assistant" {
+                    if let Some(prev) = &last_assistant_text
+                        && prev == &text
+                    {
+                        continue;
+                    }
+                    last_assistant_text = Some(text.clone());
+                }
+
+                let mut msg = json!({"role": role, "content": text});
+                if role == "assistant"
+                    && let Some(reasoning) = reasoning_by_anchor_index.get(&idx)
+                    && let Some(obj) = msg.as_object_mut()
+                {
+                    obj.insert("reasoning".to_string(), json!(reasoning));
+                }
+                messages.push(msg);
             }
             ResponseItem::FunctionCall {
                 name,
@@ -66,7 +181,7 @@ pub(crate) async fn stream_chat_completions(
                 call_id,
                 ..
             } => {
-                messages.push(json!({
+                let mut msg = json!({
                     "role": "assistant",
                     "content": null,
                     "tool_calls": [{
@@ -77,7 +192,13 @@ pub(crate) async fn stream_chat_completions(
                             "arguments": arguments,
                         }
                     }]
-                }));
+                });
+                if let Some(reasoning) = reasoning_by_anchor_index.get(&idx)
+                    && let Some(obj) = msg.as_object_mut()
+                {
+                    obj.insert("reasoning".to_string(), json!(reasoning));
+                }
+                messages.push(msg);
             }
             ResponseItem::LocalShellCall {
                 id,
@@ -86,7 +207,7 @@ pub(crate) async fn stream_chat_completions(
                 action,
             } => {
                 // Confirm with API team.
-                messages.push(json!({
+                let mut msg = json!({
                     "role": "assistant",
                     "content": null,
                     "tool_calls": [{
@@ -95,7 +216,13 @@ pub(crate) async fn stream_chat_completions(
                         "status": status,
                         "action": action,
                     }]
-                }));
+                });
+                if let Some(reasoning) = reasoning_by_anchor_index.get(&idx)
+                    && let Some(obj) = msg.as_object_mut()
+                {
+                    obj.insert("reasoning".to_string(), json!(reasoning));
+                }
+                messages.push(msg);
             }
             ResponseItem::FunctionCallOutput { call_id, output } => {
                 messages.push(json!({
@@ -131,26 +258,10 @@ pub(crate) async fn stream_chat_completions(
                     "content": output,
                 }));
             }
-            ResponseItem::Reasoning {
-                id: _,
-                summary,
-                content,
-                encrypted_content: _,
-            } => {
-                if !config.skip_reasoning_in_chat_completions {
-                    // There is no clear way of sending reasoning items over chat completions.
-                    // We are sending it as an assistant message.
-                    tracing::info!("reasoning item: {:?}", item);
-                    let reasoning =
-                        format!("Reasoning Summary: {summary:?}, Reasoning Content: {content:?}");
-                    messages.push(json!({
-                        "role": "assistant",
-                        "content": reasoning,
-                    }));
-                }
-            }
-            ResponseItem::WebSearchCall { .. } | ResponseItem::Other => {
-                tracing::info!("omitting item from chat completions: {:?}", item);
+            ResponseItem::Reasoning { .. }
+            | ResponseItem::WebSearchCall { .. }
+            | ResponseItem::Other => {
+                // Omit these items from the conversation history.
                 continue;
             }
         }
@@ -349,7 +460,10 @@ async fn process_chat_sse<S>(
             // Some providers stream `reasoning` as a plain string while others
             // nest the text under an object (e.g. `{ "reasoning": { "text": "…" } }`).
             if let Some(reasoning_val) = choice.get("delta").and_then(|d| d.get("reasoning")) {
-                let mut maybe_text = reasoning_val.as_str().map(|s| s.to_string());
+                let mut maybe_text = reasoning_val
+                    .as_str()
+                    .map(|s| s.to_string())
+                    .filter(|s| !s.is_empty());
 
                 if maybe_text.is_none() && reasoning_val.is_object() {
                     if let Some(s) = reasoning_val
@@ -368,7 +482,7 @@ async fn process_chat_sse<S>(
                 }
 
                 if let Some(reasoning) = maybe_text {
-                    // Accumulate so we can emit a terminal Reasoning item at end-of-turn.
+                    // Accumulate so we can emit a terminal Reasoning item at the end.
                     reasoning_text.push_str(&reasoning);
                     let _ = tx_event
                         .send(Ok(ResponseEvent::ReasoningContentDelta(reasoning)))
@@ -376,6 +490,31 @@ async fn process_chat_sse<S>(
                 }
             }
 
+            // Some providers only include reasoning on the final message object.
+            if let Some(message_reasoning) = choice.get("message").and_then(|m| m.get("reasoning"))
+            {
+                // Accept either a plain string or an object with { text | content }
+                if let Some(s) = message_reasoning.as_str() {
+                    if !s.is_empty() {
+                        reasoning_text.push_str(s);
+                        let _ = tx_event
+                            .send(Ok(ResponseEvent::ReasoningContentDelta(s.to_string())))
+                            .await;
+                    }
+                } else if let Some(obj) = message_reasoning.as_object()
+                    && let Some(s) = obj
+                        .get("text")
+                        .and_then(|v| v.as_str())
+                        .or_else(|| obj.get("content").and_then(|v| v.as_str()))
+                    && !s.is_empty()
+                {
+                    reasoning_text.push_str(s);
+                    let _ = tx_event
+                        .send(Ok(ResponseEvent::ReasoningContentDelta(s.to_string())))
+                        .await;
+                }
+            }
+
             // Handle streaming function / tool calls.
             if let Some(tool_calls) = choice
                 .get("delta")
@@ -531,27 +670,47 @@ where
                     // do NOT emit yet. Forward any other item (e.g. FunctionCall) right
                     // away so downstream consumers see it.
 
-                    let is_assistant_delta = matches!(&item, codex_protocol::models::ResponseItem::Message { role, .. } if role == "assistant");
+                    let is_assistant_message = matches!(
+                        &item,
+                        codex_protocol::models::ResponseItem::Message { role, .. } if role == "assistant"
+                    );
 
-                    if is_assistant_delta {
-                        // Only use the final assistant message if we have not
-                        // seen any deltas; otherwise, deltas already built the
-                        // cumulative text and this would duplicate it.
-                        if this.cumulative.is_empty()
-                            && let codex_protocol::models::ResponseItem::Message { content, .. } =
-                                &item
-                            && let Some(text) = content.iter().find_map(|c| match c {
-                                codex_protocol::models::ContentItem::OutputText { text } => {
-                                    Some(text)
+                    if is_assistant_message {
+                        match this.mode {
+                            AggregateMode::AggregatedOnly => {
+                                // Only use the final assistant message if we have not
+                                // seen any deltas; otherwise, deltas already built the
+                                // cumulative text and this would duplicate it.
+                                if this.cumulative.is_empty()
+                                    && let codex_protocol::models::ResponseItem::Message {
+                                        content,
+                                        ..
+                                    } = &item
+                                    && let Some(text) = content.iter().find_map(|c| match c {
+                                        codex_protocol::models::ContentItem::OutputText {
+                                            text,
+                                        } => Some(text),
+                                        _ => None,
+                                    })
+                                {
+                                    this.cumulative.push_str(text);
                                 }
-                                _ => None,
-                            })
-                        {
-                            this.cumulative.push_str(text);
+                                // Swallow assistant message here; emit on Completed.
+                                continue;
+                            }
+                            AggregateMode::Streaming => {
+                                // In streaming mode, if we have not seen any deltas, forward
+                                // the final assistant message directly. If deltas were seen,
+                                // suppress the final message to avoid duplication.
+                                if this.cumulative.is_empty() {
+                                    return Poll::Ready(Some(Ok(ResponseEvent::OutputItemDone(
+                                        item,
+                                    ))));
+                                } else {
+                                    continue;
+                                }
+                            }
                         }
-
-                        // Swallow assistant message here; emit on Completed.
-                        continue;
                     }
 
                     // Not an assistant message – forward immediately.
@@ -583,6 +742,11 @@ where
                         emitted_any = true;
                     }
 
+                    // Always emit the final aggregated assistant message when any
+                    // content deltas have been observed. In AggregatedOnly mode this
+                    // is the sole assistant output; in Streaming mode this finalizes
+                    // the streamed deltas into a terminal OutputItemDone so callers
+                    // can persist/render the message once per turn.
                     if !this.cumulative.is_empty() {
                         let aggregated_message = codex_protocol::models::ResponseItem::Message {
                             id: None,

codex-rs/core/src/client.rs

@@ -1,12 +1,15 @@
 use std::io::BufRead;
 use std::path::Path;
+use std::sync::OnceLock;
 use std::time::Duration;
 
+use crate::AuthManager;
 use bytes::Bytes;
-use codex_login::AuthManager;
-use codex_login::AuthMode;
+use codex_protocol::mcp_protocol::AuthMode;
+use codex_protocol::mcp_protocol::ConversationId;
 use eventsource_stream::Eventsource;
 use futures::prelude::*;
+use regex_lite::Regex;
 use reqwest::StatusCode;
 use serde::Deserialize;
 use serde::Serialize;
@@ -17,7 +20,6 @@ use tokio_util::io::ReaderStream;
 use tracing::debug;
 use tracing::trace;
 use tracing::warn;
-use uuid::Uuid;
 
 use crate::chat_completions::AggregateStreamExt;
 use crate::chat_completions::stream_chat_completions;
@@ -28,6 +30,7 @@ use crate::client_common::ResponsesApiRequest;
 use crate::client_common::create_reasoning_param_for_request;
 use crate::client_common::create_text_param_for_request;
 use crate::config::Config;
+use crate::default_client::create_client;
 use crate::error::CodexErr;
 use crate::error::Result;
 use crate::error::UsageLimitReachedError;
@@ -38,7 +41,7 @@ use crate::model_provider_info::WireApi;
 use crate::openai_model_info::get_model_info;
 use crate::openai_tools::create_tools_json_for_responses_api;
 use crate::protocol::TokenUsage;
-use crate::user_agent::get_codex_user_agent;
+use crate::token_data::PlanType;
 use crate::util::backoff;
 use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
 use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
@@ -53,10 +56,12 @@ struct ErrorResponse {
 #[derive(Debug, Deserialize)]
 struct Error {
     r#type: Option<String>,
+    #[allow(dead_code)]
+    code: Option<String>,
     message: Option<String>,
 
     // Optional fields available on "usage_limit_reached" and "usage_not_included" errors
-    plan_type: Option<String>,
+    plan_type: Option<PlanType>,
     resets_in_seconds: Option<u64>,
 }
 
@@ -66,8 +71,8 @@ pub struct ModelClient {
     auth_manager: Option<Arc<AuthManager>>,
     client: reqwest::Client,
     provider: ModelProviderInfo,
-    session_id: Uuid,
-    effort: ReasoningEffortConfig,
+    conversation_id: ConversationId,
+    effort: Option<ReasoningEffortConfig>,
     summary: ReasoningSummaryConfig,
 }
 
@@ -76,16 +81,18 @@ impl ModelClient {
         config: Arc<Config>,
         auth_manager: Option<Arc<AuthManager>>,
         provider: ModelProviderInfo,
-        effort: ReasoningEffortConfig,
+        effort: Option<ReasoningEffortConfig>,
         summary: ReasoningSummaryConfig,
-        session_id: Uuid,
+        conversation_id: ConversationId,
     ) -> Self {
+        let client = create_client();
+
         Self {
             config,
             auth_manager,
-            client: reqwest::Client::new(),
+            client,
             provider,
-            session_id,
+            conversation_id,
             effort,
             summary,
         }
@@ -97,6 +104,12 @@ impl ModelClient {
             .or_else(|| get_model_info(&self.config.model_family).map(|info| info.context_window))
     }
 
+    pub fn get_auto_compact_token_limit(&self) -> Option<i64> {
+        self.config.model_auto_compact_token_limit.or_else(|| {
+            get_model_info(&self.config.model_family).and_then(|info| info.auto_compact_token_limit)
+        })
+    }
+
     /// Dispatches to either the Responses or Chat implementation depending on
     /// the provider config.  Public callers always invoke `stream()` – the
     /// specialised helpers are private to avoid accidental misuse.
@@ -110,7 +123,6 @@ impl ModelClient {
                     &self.config.model_family,
                     &self.client,
                     &self.provider,
-                    &self.config,
                 )
                 .await?;
 
@@ -152,14 +164,6 @@ impl ModelClient {
 
         let auth_manager = self.auth_manager.clone();
 
-        let auth_mode = auth_manager
-            .as_ref()
-            .and_then(|m| m.auth())
-            .as_ref()
-            .map(|a| a.mode);
-
-        let store = prompt.store && auth_mode != Some(AuthMode::ChatGPT);
-
         let full_instructions = prompt.get_full_instructions(&self.config.model_family);
         let tools_json = create_tools_json_for_responses_api(&prompt.tools)?;
         let reasoning = create_reasoning_param_for_request(
@@ -168,9 +172,7 @@ impl ModelClient {
             self.summary,
         );
 
-        // Request encrypted COT if we are not storing responses,
-        // otherwise reasoning items will be referenced by ID
-        let include: Vec<String> = if !store && reasoning.is_some() {
+        let include: Vec<String> = if reasoning.is_some() {
             vec!["reasoning.encrypted_content".to_string()]
         } else {
             vec![]
@@ -191,6 +193,15 @@ impl ModelClient {
             None
         };
 
+        // In general, we want to explicitly send `store: false` when using the Responses API,
+        // but in practice, the Azure Responses API rejects `store: false`:
+        //
+        // - If store = false and id is sent an error is thrown that ID is not found
+        // - If store = false and id is not sent an error is thrown that ID is required
+        //
+        // For Azure, we send `store: true` and preserve reasoning item IDs.
+        let azure_workaround = self.provider.is_azure_responses_endpoint();
+
         let payload = ResponsesApiRequest {
             model: &self.config.model,
             instructions: &full_instructions,
@@ -199,13 +210,19 @@ impl ModelClient {
             tool_choice: "auto",
             parallel_tool_calls: false,
             reasoning,
-            store,
+            store: azure_workaround,
             stream: true,
             include,
-            prompt_cache_key: Some(self.session_id.to_string()),
+            prompt_cache_key: Some(self.conversation_id.to_string()),
             text,
         };
 
+        let mut payload_json = serde_json::to_value(&payload)?;
+        if azure_workaround {
+            attach_item_ids(&mut payload_json, &input_with_instructions);
+        }
+        let payload_body = serde_json::to_string(&payload_json)?;
+
         let mut attempt = 0;
         let max_retries = self.provider.request_max_retries();
 
@@ -218,7 +235,7 @@ impl ModelClient {
             trace!(
                 "POST to {}: {}",
                 self.provider.get_full_url(&auth),
-                serde_json::to_string(&payload)?
+                payload_body.as_str()
             );
 
             let mut req_builder = self
@@ -228,9 +245,11 @@ impl ModelClient {
 
             req_builder = req_builder
                 .header("OpenAI-Beta", "responses=experimental")
-                .header("session_id", self.session_id.to_string())
+                // Send session_id for compatibility.
+                .header("conversation_id", self.conversation_id.to_string())
+                .header("session_id", self.conversation_id.to_string())
                 .header(reqwest::header::ACCEPT, "text/event-stream")
-                .json(&payload);
+                .json(&payload_json);
 
             if let Some(auth) = auth.as_ref()
                 && auth.mode == AuthMode::ChatGPT
@@ -239,17 +258,13 @@ impl ModelClient {
                 req_builder = req_builder.header("chatgpt-account-id", account_id);
             }
 
-            let originator = &self.config.responses_originator_header;
-            req_builder = req_builder.header("originator", originator);
-            req_builder = req_builder.header("User-Agent", get_codex_user_agent(Some(originator)));
-
             let res = req_builder.send().await;
             if let Ok(resp) = &res {
                 trace!(
-                    "Response status: {}, request-id: {}",
+                    "Response status: {}, cf-ray: {}",
                     resp.status(),
                     resp.headers()
-                        .get("x-request-id")
+                        .get("cf-ray")
                         .map(|v| v.to_str().unwrap_or_default())
                         .unwrap_or_default()
                 );
@@ -311,7 +326,7 @@ impl ModelClient {
                                 // token.
                                 let plan_type = error
                                     .plan_type
-                                    .or_else(|| auth.and_then(|a| a.get_plan_type()));
+                                    .or_else(|| auth.as_ref().and_then(|a| a.get_plan_type()));
                                 let resets_in_seconds = error.resets_in_seconds;
                                 return Err(CodexErr::UsageLimitReached(UsageLimitReachedError {
                                     plan_type,
@@ -362,7 +377,7 @@ impl ModelClient {
     }
 
     /// Returns the current reasoning effort setting.
-    pub fn get_reasoning_effort(&self) -> ReasoningEffortConfig {
+    pub fn get_reasoning_effort(&self) -> Option<ReasoningEffortConfig> {
         self.effort
     }
 
@@ -407,9 +422,15 @@ impl From<ResponseCompletedUsage> for TokenUsage {
     fn from(val: ResponseCompletedUsage) -> Self {
         TokenUsage {
             input_tokens: val.input_tokens,
-            cached_input_tokens: val.input_tokens_details.map(|d| d.cached_tokens),
+            cached_input_tokens: val
+                .input_tokens_details
+                .map(|d| d.cached_tokens)
+                .unwrap_or(0),
             output_tokens: val.output_tokens,
-            reasoning_output_tokens: val.output_tokens_details.map(|d| d.reasoning_tokens),
+            reasoning_output_tokens: val
+                .output_tokens_details
+                .map(|d| d.reasoning_tokens)
+                .unwrap_or(0),
             total_tokens: val.total_tokens,
         }
     }
@@ -425,6 +446,33 @@ struct ResponseCompletedOutputTokensDetails {
     reasoning_tokens: u64,
 }
 
+fn attach_item_ids(payload_json: &mut Value, original_items: &[ResponseItem]) {
+    let Some(input_value) = payload_json.get_mut("input") else {
+        return;
+    };
+    let serde_json::Value::Array(items) = input_value else {
+        return;
+    };
+
+    for (value, item) in items.iter_mut().zip(original_items.iter()) {
+        if let ResponseItem::Reasoning { id, .. }
+        | ResponseItem::Message { id: Some(id), .. }
+        | ResponseItem::WebSearchCall { id: Some(id), .. }
+        | ResponseItem::FunctionCall { id: Some(id), .. }
+        | ResponseItem::LocalShellCall { id: Some(id), .. }
+        | ResponseItem::CustomToolCall { id: Some(id), .. } = item
+        {
+            if id.is_empty() {
+                continue;
+            }
+
+            if let Some(obj) = value.as_object_mut() {
+                obj.insert("id".to_string(), Value::String(id.clone()));
+            }
+        }
+    }
+}
+
 async fn process_sse<S>(
     stream: S,
     tx_event: mpsc::Sender<Result<ResponseEvent>>,
@@ -565,8 +613,9 @@ async fn process_sse<S>(
                     if let Some(error) = error {
                         match serde_json::from_value::<Error>(error.clone()) {
                             Ok(error) => {
+                                let delay = try_parse_retry_after(&error);
                                 let message = error.message.unwrap_or_default();
-                                response_error = Some(CodexErr::Stream(message, None));
+                                response_error = Some(CodexErr::Stream(message, delay));
                             }
                             Err(e) => {
                                 debug!("failed to parse ErrorResponse: {e}");
@@ -652,6 +701,40 @@ async fn stream_from_fixture(
     Ok(ResponseStream { rx_event })
 }
 
+fn rate_limit_regex() -> &'static Regex {
+    static RE: OnceLock<Regex> = OnceLock::new();
+
+    #[expect(clippy::unwrap_used)]
+    RE.get_or_init(|| Regex::new(r"Please try again in (\d+(?:\.\d+)?)(s|ms)").unwrap())
+}
+
+fn try_parse_retry_after(err: &Error) -> Option<Duration> {
+    if err.code != Some("rate_limit_exceeded".to_string()) {
+        return None;
+    }
+
+    // parse the Please try again in 1.898s format using regex
+    let re = rate_limit_regex();
+    if let Some(message) = &err.message
+        && let Some(captures) = re.captures(message)
+    {
+        let seconds = captures.get(1);
+        let unit = captures.get(2);
+
+        if let (Some(value), Some(unit)) = (seconds, unit) {
+            let value = value.as_str().parse::<f64>().ok()?;
+            let unit = unit.as_str();
+
+            if unit == "s" {
+                return Some(Duration::from_secs_f64(value));
+            } else if unit == "ms" {
+                return Some(Duration::from_millis(value as u64));
+            }
+        }
+    }
+    None
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -872,7 +955,7 @@ mod tests {
                     msg,
                     "Rate limit reached for gpt-5 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."
                 );
-                assert_eq!(*delay, None);
+                assert_eq!(*delay, Some(Duration::from_secs_f64(11.054)));
             }
             other => panic!("unexpected second event: {other:?}"),
         }
@@ -976,4 +1059,64 @@ mod tests {
             );
         }
     }
+
+    #[test]
+    fn test_try_parse_retry_after() {
+        let err = Error {
+            r#type: None,
+            message: Some("Rate limit reached for gpt-5 in organization org- on tokens per min (TPM): Limit 1, Used 1, Requested 19304. Please try again in 28ms. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
+            code: Some("rate_limit_exceeded".to_string()),
+            plan_type: None,
+            resets_in_seconds: None
+        };
+
+        let delay = try_parse_retry_after(&err);
+        assert_eq!(delay, Some(Duration::from_millis(28)));
+    }
+
+    #[test]
+    fn test_try_parse_retry_after_no_delay() {
+        let err = Error {
+            r#type: None,
+            message: Some("Rate limit reached for gpt-5 in organization <ORG> on tokens per min (TPM): Limit 30000, Used 6899, Requested 24050. Please try again in 1.898s. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
+            code: Some("rate_limit_exceeded".to_string()),
+            plan_type: None,
+            resets_in_seconds: None
+        };
+        let delay = try_parse_retry_after(&err);
+        assert_eq!(delay, Some(Duration::from_secs_f64(1.898)));
+    }
+
+    #[test]
+    fn error_response_deserializes_old_schema_known_plan_type_and_serializes_back() {
+        use crate::token_data::KnownPlan;
+        use crate::token_data::PlanType;
+
+        let json = r#"{"error":{"type":"usage_limit_reached","plan_type":"pro","resets_in_seconds":3600}}"#;
+        let resp: ErrorResponse =
+            serde_json::from_str(json).expect("should deserialize old schema");
+
+        assert!(matches!(
+            resp.error.plan_type,
+            Some(PlanType::Known(KnownPlan::Pro))
+        ));
+
+        let plan_json = serde_json::to_string(&resp.error.plan_type).expect("serialize plan_type");
+        assert_eq!(plan_json, "\"pro\"");
+    }
+
+    #[test]
+    fn error_response_deserializes_old_schema_unknown_plan_type_and_serializes_back() {
+        use crate::token_data::PlanType;
+
+        let json =
+            r#"{"error":{"type":"usage_limit_reached","plan_type":"vip","resets_in_seconds":60}}"#;
+        let resp: ErrorResponse =
+            serde_json::from_str(json).expect("should deserialize old schema");
+
+        assert!(matches!(resp.error.plan_type, Some(PlanType::Unknown(ref s)) if s == "vip"));
+
+        let plan_json = serde_json::to_string(&resp.error.plan_type).expect("serialize plan_type");
+        assert_eq!(plan_json, "\"vip\"");
+    }
 }

codex-rs/core/src/client_common.rs

@@ -1,4 +1,3 @@
-use crate::config_types::Verbosity as VerbosityConfig;
 use crate::error::Result;
 use crate::model_family::ModelFamily;
 use crate::openai_tools::OpenAiTool;
@@ -6,23 +5,19 @@ use crate::protocol::TokenUsage;
 use codex_apply_patch::APPLY_PATCH_TOOL_INSTRUCTIONS;
 use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
 use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
-use codex_protocol::models::ContentItem;
+use codex_protocol::config_types::Verbosity as VerbosityConfig;
 use codex_protocol::models::ResponseItem;
 use futures::Stream;
 use serde::Serialize;
 use std::borrow::Cow;
+use std::ops::Deref;
 use std::pin::Pin;
 use std::task::Context;
 use std::task::Poll;
 use tokio::sync::mpsc;
 
-/// The `instructions` field in the payload sent to a model should always start
-/// with this content.
-const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
-
-/// wraps user instructions message in a tag for the model to parse more easily.
-const USER_INSTRUCTIONS_START: &str = "<user_instructions>\n\n";
-const USER_INSTRUCTIONS_END: &str = "\n\n</user_instructions>";
+/// Review thread system prompt. Edit `core/src/review_prompt.md` to customize.
+pub const REVIEW_PROMPT: &str = include_str!("../review_prompt.md");
 
 /// API request payload for a single model turn
 #[derive(Default, Debug, Clone)]
@@ -30,12 +25,9 @@ pub struct Prompt {
     /// Conversation context input items.
     pub input: Vec<ResponseItem>,
 
-    /// Whether to store response on server side (disable_response_storage = !store).
-    pub store: bool,
-
     /// Tools available to the model, including additional tools sourced from
     /// external MCP servers.
-    pub tools: Vec<OpenAiTool>,
+    pub(crate) tools: Vec<OpenAiTool>,
 
     /// Optional override for the built-in BASE_INSTRUCTIONS.
     pub base_instructions_override: Option<String>,
@@ -46,11 +38,12 @@ impl Prompt {
         let base = self
             .base_instructions_override
             .as_deref()
-            .unwrap_or(BASE_INSTRUCTIONS);
+            .unwrap_or(model.base_instructions.deref());
         let mut sections: Vec<&str> = vec![base];
 
-        // When there are no custom instructions, add apply_patch_tool_instructions if either:
-        // - the model needs special instructions (4.1), or
+        // When there are no custom instructions, add apply_patch_tool_instructions if:
+        // - the model needs special instructions (4.1)
+        // AND
         // - there is no apply_patch tool present
         let is_apply_patch_tool_present = self.tools.iter().any(|tool| match tool {
             OpenAiTool::Function(f) => f.name == "apply_patch",
@@ -58,7 +51,8 @@ impl Prompt {
             _ => false,
         });
         if self.base_instructions_override.is_none()
-            && (model.needs_special_apply_patch_instructions || !is_apply_patch_tool_present)
+            && model.needs_special_apply_patch_instructions
+            && !is_apply_patch_tool_present
         {
             sections.push(APPLY_PATCH_TOOL_INSTRUCTIONS);
         }
@@ -68,17 +62,6 @@ impl Prompt {
     pub(crate) fn get_formatted_input(&self) -> Vec<ResponseItem> {
         self.input.clone()
     }
-
-    /// Creates a formatted user instructions message from a string
-    pub(crate) fn format_user_instructions_message(ui: &str) -> ResponseItem {
-        ResponseItem::Message {
-            id: None,
-            role: "user".to_string(),
-            content: vec![ContentItem::InputText {
-                text: format!("{USER_INSTRUCTIONS_START}{ui}{USER_INSTRUCTIONS_END}"),
-            }],
-        }
-    }
 }
 
 #[derive(Debug)]
@@ -100,8 +83,10 @@ pub enum ResponseEvent {
 
 #[derive(Debug, Serialize)]
 pub(crate) struct Reasoning {
-    pub(crate) effort: ReasoningEffortConfig,
-    pub(crate) summary: ReasoningSummaryConfig,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(crate) effort: Option<ReasoningEffortConfig>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(crate) summary: Option<ReasoningSummaryConfig>,
 }
 
 /// Controls under the `text` field in the Responses API for GPT-5.
@@ -144,7 +129,6 @@ pub(crate) struct ResponsesApiRequest<'a> {
     pub(crate) tool_choice: &'static str,
     pub(crate) parallel_tool_calls: bool,
     pub(crate) reasoning: Option<Reasoning>,
-    /// true when using the Responses API.
     pub(crate) store: bool,
     pub(crate) stream: bool,
     pub(crate) include: Vec<String>,
@@ -156,14 +140,17 @@ pub(crate) struct ResponsesApiRequest<'a> {
 
 pub(crate) fn create_reasoning_param_for_request(
     model_family: &ModelFamily,
-    effort: ReasoningEffortConfig,
+    effort: Option<ReasoningEffortConfig>,
     summary: ReasoningSummaryConfig,
 ) -> Option<Reasoning> {
-    if model_family.supports_reasoning_summaries {
-        Some(Reasoning { effort, summary })
-    } else {
-        None
+    if !model_family.supports_reasoning_summaries {
+        return None;
     }
+
+    Some(Reasoning {
+        effort,
+        summary: Some(summary),
+    })
 }
 
 pub(crate) fn create_text_param_for_request(
@@ -174,7 +161,7 @@ pub(crate) fn create_text_param_for_request(
     })
 }
 
-pub(crate) struct ResponseStream {
+pub struct ResponseStream {
     pub(crate) rx_event: mpsc::Receiver<Result<ResponseEvent>>,
 }
 
@@ -189,18 +176,64 @@ impl Stream for ResponseStream {
 #[cfg(test)]
 mod tests {
     use crate::model_family::find_family_for_model;
+    use pretty_assertions::assert_eq;
 
     use super::*;
 
+    struct InstructionsTestCase {
+        pub slug: &'static str,
+        pub expects_apply_patch_instructions: bool,
+    }
     #[test]
     fn get_full_instructions_no_user_content() {
         let prompt = Prompt {
             ..Default::default()
         };
-        let expected = format!("{BASE_INSTRUCTIONS}\n{APPLY_PATCH_TOOL_INSTRUCTIONS}");
-        let model_family = find_family_for_model("gpt-4.1").expect("known model slug");
-        let full = prompt.get_full_instructions(&model_family);
-        assert_eq!(full, expected);
+        let test_cases = vec![
+            InstructionsTestCase {
+                slug: "gpt-3.5",
+                expects_apply_patch_instructions: true,
+            },
+            InstructionsTestCase {
+                slug: "gpt-4.1",
+                expects_apply_patch_instructions: true,
+            },
+            InstructionsTestCase {
+                slug: "gpt-4o",
+                expects_apply_patch_instructions: true,
+            },
+            InstructionsTestCase {
+                slug: "gpt-5",
+                expects_apply_patch_instructions: true,
+            },
+            InstructionsTestCase {
+                slug: "codex-mini-latest",
+                expects_apply_patch_instructions: true,
+            },
+            InstructionsTestCase {
+                slug: "gpt-oss:120b",
+                expects_apply_patch_instructions: false,
+            },
+            InstructionsTestCase {
+                slug: "swiftfox",
+                expects_apply_patch_instructions: false,
+            },
+        ];
+        for test_case in test_cases {
+            let model_family = find_family_for_model(test_case.slug).expect("known model slug");
+            let expected = if test_case.expects_apply_patch_instructions {
+                format!(
+                    "{}\n{}",
+                    model_family.clone().base_instructions,
+                    APPLY_PATCH_TOOL_INSTRUCTIONS
+                )
+            } else {
+                model_family.clone().base_instructions
+            };
+
+            let full = prompt.get_full_instructions(&model_family);
+            assert_eq!(full, expected);
+        }
     }
 
     #[test]
@@ -215,7 +248,7 @@ mod tests {
             tool_choice: "auto",
             parallel_tool_calls: false,
             reasoning: None,
-            store: true,
+            store: false,
             stream: true,
             include: vec![],
             prompt_cache_key: None,
@@ -245,7 +278,7 @@ mod tests {
             tool_choice: "auto",
             parallel_tool_calls: false,
             reasoning: None,
-            store: true,
+            store: false,
             stream: true,
             include: vec![],
             prompt_cache_key: None,

codex-rs/core/src/codex/compact.rs

@@ -0,0 +1,400 @@
+use std::sync::Arc;
+
+use super::AgentTask;
+use super::MutexExt;
+use super::Session;
+use super::TurnContext;
+use super::get_last_assistant_message_from_turn;
+use crate::Prompt;
+use crate::client_common::ResponseEvent;
+use crate::error::CodexErr;
+use crate::error::Result as CodexResult;
+use crate::protocol::AgentMessageEvent;
+use crate::protocol::CompactedItem;
+use crate::protocol::ErrorEvent;
+use crate::protocol::Event;
+use crate::protocol::EventMsg;
+use crate::protocol::InputItem;
+use crate::protocol::InputMessageKind;
+use crate::protocol::TaskCompleteEvent;
+use crate::protocol::TaskStartedEvent;
+use crate::protocol::TurnContextItem;
+use crate::util::backoff;
+use askama::Template;
+use codex_protocol::models::ContentItem;
+use codex_protocol::models::ResponseInputItem;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::RolloutItem;
+use futures::prelude::*;
+
+pub(super) const COMPACT_TRIGGER_TEXT: &str = "Start Summarization";
+const SUMMARIZATION_PROMPT: &str = include_str!("../../templates/compact/prompt.md");
+
+#[derive(Template)]
+#[template(path = "compact/history_bridge.md", escape = "none")]
+struct HistoryBridgeTemplate<'a> {
+    user_messages_text: &'a str,
+    summary_text: &'a str,
+}
+
+pub(super) fn spawn_compact_task(
+    sess: Arc<Session>,
+    turn_context: Arc<TurnContext>,
+    sub_id: String,
+    input: Vec<InputItem>,
+) {
+    let task = AgentTask::compact(
+        sess.clone(),
+        turn_context,
+        sub_id,
+        input,
+        SUMMARIZATION_PROMPT.to_string(),
+    );
+    sess.set_task(task);
+}
+
+pub(super) async fn run_inline_auto_compact_task(
+    sess: Arc<Session>,
+    turn_context: Arc<TurnContext>,
+) {
+    let sub_id = sess.next_internal_sub_id();
+    let input = vec![InputItem::Text {
+        text: COMPACT_TRIGGER_TEXT.to_string(),
+    }];
+    run_compact_task_inner(
+        sess,
+        turn_context,
+        sub_id,
+        input,
+        SUMMARIZATION_PROMPT.to_string(),
+        false,
+    )
+    .await;
+}
+
+pub(super) async fn run_compact_task(
+    sess: Arc<Session>,
+    turn_context: Arc<TurnContext>,
+    sub_id: String,
+    input: Vec<InputItem>,
+    compact_instructions: String,
+) {
+    run_compact_task_inner(
+        sess,
+        turn_context,
+        sub_id,
+        input,
+        compact_instructions,
+        true,
+    )
+    .await;
+}
+
+async fn run_compact_task_inner(
+    sess: Arc<Session>,
+    turn_context: Arc<TurnContext>,
+    sub_id: String,
+    input: Vec<InputItem>,
+    compact_instructions: String,
+    remove_task_on_completion: bool,
+) {
+    let model_context_window = turn_context.client.get_model_context_window();
+    let start_event = Event {
+        id: sub_id.clone(),
+        msg: EventMsg::TaskStarted(TaskStartedEvent {
+            model_context_window,
+        }),
+    };
+    sess.send_event(start_event).await;
+
+    let initial_input_for_turn: ResponseInputItem = ResponseInputItem::from(input);
+    let instructions_override = compact_instructions;
+    let turn_input = sess.turn_input_with_history(vec![initial_input_for_turn.clone().into()]);
+
+    let prompt = Prompt {
+        input: turn_input,
+        tools: Vec::new(),
+        base_instructions_override: Some(instructions_override),
+    };
+
+    let max_retries = turn_context.client.get_provider().stream_max_retries();
+    let mut retries = 0;
+
+    let rollout_item = RolloutItem::TurnContext(TurnContextItem {
+        cwd: turn_context.cwd.clone(),
+        approval_policy: turn_context.approval_policy,
+        sandbox_policy: turn_context.sandbox_policy.clone(),
+        model: turn_context.client.get_model(),
+        effort: turn_context.client.get_reasoning_effort(),
+        summary: turn_context.client.get_reasoning_summary(),
+    });
+    sess.persist_rollout_items(&[rollout_item]).await;
+
+    loop {
+        let attempt_result = drain_to_completed(&sess, turn_context.as_ref(), &prompt).await;
+
+        match attempt_result {
+            Ok(()) => {
+                break;
+            }
+            Err(CodexErr::Interrupted) => {
+                return;
+            }
+            Err(e) => {
+                if retries < max_retries {
+                    retries += 1;
+                    let delay = backoff(retries);
+                    sess.notify_stream_error(
+                        &sub_id,
+                        format!(
+                            "stream error: {e}; retrying {retries}/{max_retries} in {delay:?}…"
+                        ),
+                    )
+                    .await;
+                    tokio::time::sleep(delay).await;
+                    continue;
+                } else {
+                    let event = Event {
+                        id: sub_id.clone(),
+                        msg: EventMsg::Error(ErrorEvent {
+                            message: e.to_string(),
+                        }),
+                    };
+                    sess.send_event(event).await;
+                    return;
+                }
+            }
+        }
+    }
+
+    if remove_task_on_completion {
+        sess.remove_task(&sub_id);
+    }
+    let history_snapshot = {
+        let state = sess.state.lock_unchecked();
+        state.history.contents()
+    };
+    let summary_text = get_last_assistant_message_from_turn(&history_snapshot).unwrap_or_default();
+    let user_messages = collect_user_messages(&history_snapshot);
+    let initial_context = sess.build_initial_context(turn_context.as_ref());
+    let new_history = build_compacted_history(initial_context, &user_messages, &summary_text);
+    {
+        let mut state = sess.state.lock_unchecked();
+        state.history.replace(new_history);
+    }
+
+    let rollout_item = RolloutItem::Compacted(CompactedItem {
+        message: summary_text.clone(),
+    });
+    sess.persist_rollout_items(&[rollout_item]).await;
+
+    let event = Event {
+        id: sub_id.clone(),
+        msg: EventMsg::AgentMessage(AgentMessageEvent {
+            message: "Compact task completed".to_string(),
+        }),
+    };
+    sess.send_event(event).await;
+    let event = Event {
+        id: sub_id.clone(),
+        msg: EventMsg::TaskComplete(TaskCompleteEvent {
+            last_agent_message: None,
+        }),
+    };
+    sess.send_event(event).await;
+}
+
+fn content_items_to_text(content: &[ContentItem]) -> Option<String> {
+    let mut pieces = Vec::new();
+    for item in content {
+        match item {
+            ContentItem::InputText { text } | ContentItem::OutputText { text } => {
+                if !text.is_empty() {
+                    pieces.push(text.as_str());
+                }
+            }
+            ContentItem::InputImage { .. } => {}
+        }
+    }
+    if pieces.is_empty() {
+        None
+    } else {
+        Some(pieces.join("\n"))
+    }
+}
+
+pub(crate) fn collect_user_messages(items: &[ResponseItem]) -> Vec<String> {
+    items
+        .iter()
+        .filter_map(|item| match item {
+            ResponseItem::Message { role, content, .. } if role == "user" => {
+                content_items_to_text(content)
+            }
+            _ => None,
+        })
+        .filter(|text| !is_session_prefix_message(text))
+        .collect()
+}
+
+fn is_session_prefix_message(text: &str) -> bool {
+    matches!(
+        InputMessageKind::from(("user", text)),
+        InputMessageKind::UserInstructions | InputMessageKind::EnvironmentContext
+    )
+}
+
+pub(crate) fn build_compacted_history(
+    initial_context: Vec<ResponseItem>,
+    user_messages: &[String],
+    summary_text: &str,
+) -> Vec<ResponseItem> {
+    let mut history = initial_context;
+    let user_messages_text = if user_messages.is_empty() {
+        "(none)".to_string()
+    } else {
+        user_messages.join("\n\n")
+    };
+    let summary_text = if summary_text.is_empty() {
+        "(no summary available)".to_string()
+    } else {
+        summary_text.to_string()
+    };
+    let Ok(bridge) = HistoryBridgeTemplate {
+        user_messages_text: &user_messages_text,
+        summary_text: &summary_text,
+    }
+    .render() else {
+        return vec![];
+    };
+    history.push(ResponseItem::Message {
+        id: None,
+        role: "user".to_string(),
+        content: vec![ContentItem::InputText { text: bridge }],
+    });
+    history
+}
+
+async fn drain_to_completed(
+    sess: &Session,
+    turn_context: &TurnContext,
+    prompt: &Prompt,
+) -> CodexResult<()> {
+    let mut stream = turn_context.client.clone().stream(prompt).await?;
+    loop {
+        let maybe_event = stream.next().await;
+        let Some(event) = maybe_event else {
+            return Err(CodexErr::Stream(
+                "stream closed before response.completed".into(),
+                None,
+            ));
+        };
+        match event {
+            Ok(ResponseEvent::OutputItemDone(item)) => {
+                let mut state = sess.state.lock_unchecked();
+                state.history.record_items(std::slice::from_ref(&item));
+            }
+            Ok(ResponseEvent::Completed { .. }) => {
+                return Ok(());
+            }
+            Ok(_) => continue,
+            Err(e) => return Err(e),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use pretty_assertions::assert_eq;
+
+    #[test]
+    fn content_items_to_text_joins_non_empty_segments() {
+        let items = vec![
+            ContentItem::InputText {
+                text: "hello".to_string(),
+            },
+            ContentItem::OutputText {
+                text: String::new(),
+            },
+            ContentItem::OutputText {
+                text: "world".to_string(),
+            },
+        ];
+
+        let joined = content_items_to_text(&items);
+
+        assert_eq!(Some("hello\nworld".to_string()), joined);
+    }
+
+    #[test]
+    fn content_items_to_text_ignores_image_only_content() {
+        let items = vec![ContentItem::InputImage {
+            image_url: "file://image.png".to_string(),
+        }];
+
+        let joined = content_items_to_text(&items);
+
+        assert_eq!(None, joined);
+    }
+
+    #[test]
+    fn collect_user_messages_extracts_user_text_only() {
+        let items = vec![
+            ResponseItem::Message {
+                id: Some("assistant".to_string()),
+                role: "assistant".to_string(),
+                content: vec![ContentItem::OutputText {
+                    text: "ignored".to_string(),
+                }],
+            },
+            ResponseItem::Message {
+                id: Some("user".to_string()),
+                role: "user".to_string(),
+                content: vec![
+                    ContentItem::InputText {
+                        text: "first".to_string(),
+                    },
+                    ContentItem::OutputText {
+                        text: "second".to_string(),
+                    },
+                ],
+            },
+            ResponseItem::Other,
+        ];
+
+        let collected = collect_user_messages(&items);
+
+        assert_eq!(vec!["first\nsecond".to_string()], collected);
+    }
+
+    #[test]
+    fn collect_user_messages_filters_session_prefix_entries() {
+        let items = vec![
+            ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText {
+                    text: "<user_instructions>do things</user_instructions>".to_string(),
+                }],
+            },
+            ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText {
+                    text: "<ENVIRONMENT_CONTEXT>cwd=/tmp</ENVIRONMENT_CONTEXT>".to_string(),
+                }],
+            },
+            ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText {
+                    text: "real user message".to_string(),
+                }],
+            },
+        ];
+
+        let collected = collect_user_messages(&items);
+
+        assert_eq!(vec!["real user message".to_string()], collected);
+    }
+}

codex-rs/core/src/config_edit.rs

@@ -0,0 +1,748 @@
+use crate::config::CONFIG_TOML_FILE;
+use anyhow::Result;
+use std::path::Path;
+use tempfile::NamedTempFile;
+use toml_edit::DocumentMut;
+
+pub const CONFIG_KEY_MODEL: &str = "model";
+pub const CONFIG_KEY_EFFORT: &str = "model_reasoning_effort";
+
+#[derive(Copy, Clone)]
+enum NoneBehavior {
+    Skip,
+    Remove,
+}
+
+/// Persist overrides into `config.toml` using explicit key segments per
+/// override. This avoids ambiguity with keys that contain dots or spaces.
+pub async fn persist_overrides(
+    codex_home: &Path,
+    profile: Option<&str>,
+    overrides: &[(&[&str], &str)],
+) -> Result<()> {
+    let with_options: Vec<(&[&str], Option<&str>)> = overrides
+        .iter()
+        .map(|(segments, value)| (*segments, Some(*value)))
+        .collect();
+
+    persist_overrides_with_behavior(codex_home, profile, &with_options, NoneBehavior::Skip).await
+}
+
+/// Persist overrides where values may be optional. Any entries with `None`
+/// values are skipped. If all values are `None`, this becomes a no-op and
+/// returns `Ok(())` without touching the file.
+pub async fn persist_non_null_overrides(
+    codex_home: &Path,
+    profile: Option<&str>,
+    overrides: &[(&[&str], Option<&str>)],
+) -> Result<()> {
+    persist_overrides_with_behavior(codex_home, profile, overrides, NoneBehavior::Skip).await
+}
+
+/// Persist overrides where `None` values clear any existing values from the
+/// configuration file.
+pub async fn persist_overrides_and_clear_if_none(
+    codex_home: &Path,
+    profile: Option<&str>,
+    overrides: &[(&[&str], Option<&str>)],
+) -> Result<()> {
+    persist_overrides_with_behavior(codex_home, profile, overrides, NoneBehavior::Remove).await
+}
+
+/// Apply a single override onto a `toml_edit` document while preserving
+/// existing formatting/comments.
+/// The key is expressed as explicit segments to correctly handle keys that
+/// contain dots or spaces.
+fn apply_toml_edit_override_segments(
+    doc: &mut DocumentMut,
+    segments: &[&str],
+    value: toml_edit::Item,
+) {
+    use toml_edit::Item;
+
+    if segments.is_empty() {
+        return;
+    }
+
+    let mut current = doc.as_table_mut();
+    for seg in &segments[..segments.len() - 1] {
+        if !current.contains_key(seg) {
+            current[*seg] = Item::Table(toml_edit::Table::new());
+            if let Some(t) = current[*seg].as_table_mut() {
+                t.set_implicit(true);
+            }
+        }
+
+        let maybe_item = current.get_mut(seg);
+        let Some(item) = maybe_item else { return };
+
+        if !item.is_table() {
+            *item = Item::Table(toml_edit::Table::new());
+            if let Some(t) = item.as_table_mut() {
+                t.set_implicit(true);
+            }
+        }
+
+        let Some(tbl) = item.as_table_mut() else {
+            return;
+        };
+        current = tbl;
+    }
+
+    let last = segments[segments.len() - 1];
+    current[last] = value;
+}
+
+async fn persist_overrides_with_behavior(
+    codex_home: &Path,
+    profile: Option<&str>,
+    overrides: &[(&[&str], Option<&str>)],
+    none_behavior: NoneBehavior,
+) -> Result<()> {
+    if overrides.is_empty() {
+        return Ok(());
+    }
+
+    let should_skip = match none_behavior {
+        NoneBehavior::Skip => overrides.iter().all(|(_, value)| value.is_none()),
+        NoneBehavior::Remove => false,
+    };
+
+    if should_skip {
+        return Ok(());
+    }
+
+    let config_path = codex_home.join(CONFIG_TOML_FILE);
+
+    let read_result = tokio::fs::read_to_string(&config_path).await;
+    let mut doc = match read_result {
+        Ok(contents) => contents.parse::<DocumentMut>()?,
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
+            if overrides
+                .iter()
+                .all(|(_, value)| value.is_none() && matches!(none_behavior, NoneBehavior::Remove))
+            {
+                return Ok(());
+            }
+
+            tokio::fs::create_dir_all(codex_home).await?;
+            DocumentMut::new()
+        }
+        Err(e) => return Err(e.into()),
+    };
+
+    let effective_profile = if let Some(p) = profile {
+        Some(p.to_owned())
+    } else {
+        doc.get("profile")
+            .and_then(|i| i.as_str())
+            .map(|s| s.to_string())
+    };
+
+    let mut mutated = false;
+
+    for (segments, value) in overrides.iter().copied() {
+        let mut seg_buf: Vec<&str> = Vec::new();
+        let segments_to_apply: &[&str];
+
+        if let Some(ref name) = effective_profile {
+            if segments.first().copied() == Some("profiles") {
+                segments_to_apply = segments;
+            } else {
+                seg_buf.reserve(2 + segments.len());
+                seg_buf.push("profiles");
+                seg_buf.push(name.as_str());
+                seg_buf.extend_from_slice(segments);
+                segments_to_apply = seg_buf.as_slice();
+            }
+        } else {
+            segments_to_apply = segments;
+        }
+
+        match value {
+            Some(v) => {
+                let item_value = toml_edit::value(v);
+                apply_toml_edit_override_segments(&mut doc, segments_to_apply, item_value);
+                mutated = true;
+            }
+            None => {
+                if matches!(none_behavior, NoneBehavior::Remove)
+                    && remove_toml_edit_segments(&mut doc, segments_to_apply)
+                {
+                    mutated = true;
+                }
+            }
+        }
+    }
+
+    if !mutated {
+        return Ok(());
+    }
+
+    let tmp_file = NamedTempFile::new_in(codex_home)?;
+    tokio::fs::write(tmp_file.path(), doc.to_string()).await?;
+    tmp_file.persist(config_path)?;
+
+    Ok(())
+}
+
+fn remove_toml_edit_segments(doc: &mut DocumentMut, segments: &[&str]) -> bool {
+    use toml_edit::Item;
+
+    if segments.is_empty() {
+        return false;
+    }
+
+    let mut current = doc.as_table_mut();
+    for seg in &segments[..segments.len() - 1] {
+        let Some(item) = current.get_mut(seg) else {
+            return false;
+        };
+
+        match item {
+            Item::Table(table) => {
+                current = table;
+            }
+            _ => {
+                return false;
+            }
+        }
+    }
+
+    current.remove(segments[segments.len() - 1]).is_some()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use pretty_assertions::assert_eq;
+    use tempfile::tempdir;
+
+    /// Verifies model and effort are written at top-level when no profile is set.
+    #[tokio::test]
+    async fn set_default_model_and_effort_top_level_when_no_profile() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        persist_overrides(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], "gpt-5"),
+                (&[CONFIG_KEY_EFFORT], "high"),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"model = "gpt-5"
+model_reasoning_effort = "high"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies values are written under the active profile when `profile` is set.
+    #[tokio::test]
+    async fn set_defaults_update_profile_when_profile_set() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed config with a profile selection but without profiles table
+        let seed = "profile = \"o3\"\n";
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], "o3"),
+                (&[CONFIG_KEY_EFFORT], "minimal"),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"profile = "o3"
+
+[profiles.o3]
+model = "o3"
+model_reasoning_effort = "minimal"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies profile names with dots/spaces are preserved via explicit segments.
+    #[tokio::test]
+    async fn set_defaults_update_profile_with_dot_and_space() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed config with a profile name that contains a dot and a space
+        let seed = "profile = \"my.team name\"\n";
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], "o3"),
+                (&[CONFIG_KEY_EFFORT], "minimal"),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"profile = "my.team name"
+
+[profiles."my.team name"]
+model = "o3"
+model_reasoning_effort = "minimal"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies explicit profile override writes under that profile even without active profile.
+    #[tokio::test]
+    async fn set_defaults_update_when_profile_override_supplied() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // No profile key in config.toml
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), "")
+            .await
+            .expect("seed write");
+
+        // Persist with an explicit profile override
+        persist_overrides(
+            codex_home,
+            Some("o3"),
+            &[(&[CONFIG_KEY_MODEL], "o3"), (&[CONFIG_KEY_EFFORT], "high")],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"[profiles.o3]
+model = "o3"
+model_reasoning_effort = "high"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies nested tables are created as needed when applying overrides.
+    #[tokio::test]
+    async fn persist_overrides_creates_nested_tables() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        persist_overrides(
+            codex_home,
+            None,
+            &[
+                (&["a", "b", "c"], "v"),
+                (&["x"], "y"),
+                (&["profiles", "p1", CONFIG_KEY_MODEL], "gpt-5"),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"x = "y"
+
+[a.b]
+c = "v"
+
+[profiles.p1]
+model = "gpt-5"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies a scalar key becomes a table when nested keys are written.
+    #[tokio::test]
+    async fn persist_overrides_replaces_scalar_with_table() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+        let seed = "foo = \"bar\"\n";
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides(codex_home, None, &[(&["foo", "bar", "baz"], "ok")])
+            .await
+            .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"[foo.bar]
+baz = "ok"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies comments and spacing are preserved when writing under active profile.
+    #[tokio::test]
+    async fn set_defaults_preserve_comments() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed a config with comments and spacing we expect to preserve
+        let seed = r#"# Global comment
+# Another line
+
+profile = "o3"
+
+# Profile settings
+[profiles.o3]
+# keep me
+existing = "keep"
+"#;
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        // Apply defaults; since profile is set, it should write under [profiles.o3]
+        persist_overrides(
+            codex_home,
+            None,
+            &[(&[CONFIG_KEY_MODEL], "o3"), (&[CONFIG_KEY_EFFORT], "high")],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"# Global comment
+# Another line
+
+profile = "o3"
+
+# Profile settings
+[profiles.o3]
+# keep me
+existing = "keep"
+model = "o3"
+model_reasoning_effort = "high"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies comments and spacing are preserved when writing at top level.
+    #[tokio::test]
+    async fn set_defaults_preserve_global_comments() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed a config WITHOUT a profile, containing comments and spacing
+        let seed = r#"# Top-level comments
+# should be preserved
+
+existing = "keep"
+"#;
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        // Since there is no profile, the defaults should be written at top-level
+        persist_overrides(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], "gpt-5"),
+                (&[CONFIG_KEY_EFFORT], "minimal"),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"# Top-level comments
+# should be preserved
+
+existing = "keep"
+model = "gpt-5"
+model_reasoning_effort = "minimal"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies errors on invalid TOML propagate and file is not clobbered.
+    #[tokio::test]
+    async fn persist_overrides_errors_on_parse_failure() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Write an intentionally invalid TOML file
+        let invalid = "invalid = [unclosed";
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), invalid)
+            .await
+            .expect("seed write");
+
+        // Attempting to persist should return an error and must not clobber the file.
+        let res = persist_overrides(codex_home, None, &[(&["x"], "y")]).await;
+        assert!(res.is_err(), "expected parse error to propagate");
+
+        // File should be unchanged
+        let contents = read_config(codex_home).await;
+        assert_eq!(contents, invalid);
+    }
+
+    /// Verifies changing model only preserves existing effort at top-level.
+    #[tokio::test]
+    async fn changing_only_model_preserves_existing_effort_top_level() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed with an effort value only
+        let seed = "model_reasoning_effort = \"minimal\"\n";
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        // Change only the model
+        persist_overrides(codex_home, None, &[(&[CONFIG_KEY_MODEL], "o3")])
+            .await
+            .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"model_reasoning_effort = "minimal"
+model = "o3"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies changing effort only preserves existing model at top-level.
+    #[tokio::test]
+    async fn changing_only_effort_preserves_existing_model_top_level() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed with a model value only
+        let seed = "model = \"gpt-5\"\n";
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        // Change only the effort
+        persist_overrides(codex_home, None, &[(&[CONFIG_KEY_EFFORT], "high")])
+            .await
+            .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"model = "gpt-5"
+model_reasoning_effort = "high"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies changing model only preserves existing effort in active profile.
+    #[tokio::test]
+    async fn changing_only_model_preserves_effort_in_active_profile() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // Seed with an active profile and an existing effort under that profile
+        let seed = r#"profile = "p1"
+
+[profiles.p1]
+model_reasoning_effort = "low"
+"#;
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides(codex_home, None, &[(&[CONFIG_KEY_MODEL], "o4-mini")])
+            .await
+            .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"profile = "p1"
+
+[profiles.p1]
+model_reasoning_effort = "low"
+model = "o4-mini"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies changing effort only preserves existing model in a profile override.
+    #[tokio::test]
+    async fn changing_only_effort_preserves_model_in_profile_override() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        // No active profile key; we'll target an explicit override
+        let seed = r#"[profiles.team]
+model = "gpt-5"
+"#;
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides(
+            codex_home,
+            Some("team"),
+            &[(&[CONFIG_KEY_EFFORT], "minimal")],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"[profiles.team]
+model = "gpt-5"
+model_reasoning_effort = "minimal"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies `persist_non_null_overrides` skips `None` entries and writes only present values at top-level.
+    #[tokio::test]
+    async fn persist_non_null_skips_none_top_level() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        persist_non_null_overrides(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], Some("gpt-5")),
+                (&[CONFIG_KEY_EFFORT], None),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = "model = \"gpt-5\"\n";
+        assert_eq!(contents, expected);
+    }
+
+    /// Verifies no-op behavior when all provided overrides are `None` (no file created/modified).
+    #[tokio::test]
+    async fn persist_non_null_noop_when_all_none() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        persist_non_null_overrides(
+            codex_home,
+            None,
+            &[(&["a"], None), (&["profiles", "p", "x"], None)],
+        )
+        .await
+        .expect("persist");
+
+        // Should not create config.toml on a pure no-op
+        assert!(!codex_home.join(CONFIG_TOML_FILE).exists());
+    }
+
+    /// Verifies entries are written under the specified profile and `None` entries are skipped.
+    #[tokio::test]
+    async fn persist_non_null_respects_profile_override() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        persist_non_null_overrides(
+            codex_home,
+            Some("team"),
+            &[
+                (&[CONFIG_KEY_MODEL], Some("o3")),
+                (&[CONFIG_KEY_EFFORT], None),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"[profiles.team]
+model = "o3"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    #[tokio::test]
+    async fn persist_clear_none_removes_top_level_value() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        let seed = r#"model = "gpt-5"
+model_reasoning_effort = "medium"
+"#;
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides_and_clear_if_none(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], None),
+                (&[CONFIG_KEY_EFFORT], Some("high")),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = "model_reasoning_effort = \"high\"\n";
+        assert_eq!(contents, expected);
+    }
+
+    #[tokio::test]
+    async fn persist_clear_none_respects_active_profile() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        let seed = r#"profile = "team"
+
+[profiles.team]
+model = "gpt-4"
+model_reasoning_effort = "minimal"
+"#;
+        tokio::fs::write(codex_home.join(CONFIG_TOML_FILE), seed)
+            .await
+            .expect("seed write");
+
+        persist_overrides_and_clear_if_none(
+            codex_home,
+            None,
+            &[
+                (&[CONFIG_KEY_MODEL], None),
+                (&[CONFIG_KEY_EFFORT], Some("high")),
+            ],
+        )
+        .await
+        .expect("persist");
+
+        let contents = read_config(codex_home).await;
+        let expected = r#"profile = "team"
+
+[profiles.team]
+model_reasoning_effort = "high"
+"#;
+        assert_eq!(contents, expected);
+    }
+
+    #[tokio::test]
+    async fn persist_clear_none_noop_when_file_missing() {
+        let tmpdir = tempdir().expect("tmp");
+        let codex_home = tmpdir.path();
+
+        persist_overrides_and_clear_if_none(codex_home, None, &[(&[CONFIG_KEY_MODEL], None)])
+            .await
+            .expect("persist");
+
+        assert!(!codex_home.join(CONFIG_TOML_FILE).exists());
+    }
+
+    // Test helper moved to bottom per review guidance.
+    async fn read_config(codex_home: &Path) -> String {
+        let p = codex_home.join(CONFIG_TOML_FILE);
+        tokio::fs::read_to_string(p).await.unwrap_or_default()
+    }
+}

codex-rs/core/src/config_profile.rs

@@ -1,10 +1,10 @@
 use serde::Deserialize;
 use std::path::PathBuf;
 
-use crate::config_types::Verbosity;
 use crate::protocol::AskForApproval;
 use codex_protocol::config_types::ReasoningEffort;
 use codex_protocol::config_types::ReasoningSummary;
+use codex_protocol::config_types::Verbosity;
 
 /// Collection of common configuration options that a user can define as a unit
 /// in `config.toml`.
@@ -15,10 +15,23 @@ pub struct ConfigProfile {
     /// [`ModelProviderInfo`] to use.
     pub model_provider: Option<String>,
     pub approval_policy: Option<AskForApproval>,
-    pub disable_response_storage: Option<bool>,
     pub model_reasoning_effort: Option<ReasoningEffort>,
     pub model_reasoning_summary: Option<ReasoningSummary>,
     pub model_verbosity: Option<Verbosity>,
     pub chatgpt_base_url: Option<String>,
     pub experimental_instructions_file: Option<PathBuf>,
 }
+
+impl From<ConfigProfile> for codex_protocol::mcp_protocol::Profile {
+    fn from(config_profile: ConfigProfile) -> Self {
+        Self {
+            model: config_profile.model,
+            model_provider: config_profile.model_provider,
+            approval_policy: config_profile.approval_policy,
+            model_reasoning_effort: config_profile.model_reasoning_effort,
+            model_reasoning_summary: config_profile.model_reasoning_summary,
+            model_verbosity: config_profile.model_verbosity,
+            chatgpt_base_url: config_profile.chatgpt_base_url,
+        }
+    }
+}

codex-rs/core/src/config_types.rs

@@ -8,8 +8,6 @@ use std::path::PathBuf;
 use wildmatch::WildMatchPattern;
 
 use serde::Deserialize;
-use serde::Serialize;
-use strum_macros::Display;
 
 #[derive(Deserialize, Debug, Clone, PartialEq)]
 pub struct McpServerConfig {
@@ -20,6 +18,10 @@ pub struct McpServerConfig {
 
     #[serde(default)]
     pub env: Option<HashMap<String, String>>,
+
+    /// Startup timeout in milliseconds for initializing MCP server & initially listing tools.
+    #[serde(default)]
+    pub startup_timeout_ms: Option<u64>,
 }
 
 #[derive(Deserialize, Debug, Copy, Clone, PartialEq)]
@@ -90,6 +92,17 @@ pub struct SandboxWorkspaceWrite {
     pub exclude_slash_tmp: bool,
 }
 
+impl From<SandboxWorkspaceWrite> for codex_protocol::mcp_protocol::SandboxSettings {
+    fn from(sandbox_workspace_write: SandboxWorkspaceWrite) -> Self {
+        Self {
+            writable_roots: sandbox_workspace_write.writable_roots,
+            network_access: Some(sandbox_workspace_write.network_access),
+            exclude_tmpdir_env_var: Some(sandbox_workspace_write.exclude_tmpdir_env_var),
+            exclude_slash_tmp: Some(sandbox_workspace_write.exclude_slash_tmp),
+        }
+    }
+}
+
 #[derive(Deserialize, Debug, Clone, PartialEq, Default)]
 #[serde(rename_all = "kebab-case")]
 pub enum ShellEnvironmentPolicyInherit {
@@ -186,42 +199,10 @@ impl From<ShellEnvironmentPolicyToml> for ShellEnvironmentPolicy {
     }
 }
 
-/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#get-started-with-reasoning
-#[derive(Debug, Serialize, Deserialize, Default, Clone, Copy, PartialEq, Eq, Display)]
-#[serde(rename_all = "lowercase")]
-#[strum(serialize_all = "lowercase")]
-pub enum ReasoningEffort {
-    Low,
+#[derive(Deserialize, Debug, Clone, PartialEq, Eq, Default, Hash)]
+#[serde(rename_all = "kebab-case")]
+pub enum ReasoningSummaryFormat {
     #[default]
-    Medium,
-    High,
-    /// Option to disable reasoning.
     None,
-}
-
-/// A summary of the reasoning performed by the model. This can be useful for
-/// debugging and understanding the model's reasoning process.
-/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#reasoning-summaries
-#[derive(Debug, Serialize, Deserialize, Default, Clone, Copy, PartialEq, Eq, Display)]
-#[serde(rename_all = "lowercase")]
-#[strum(serialize_all = "lowercase")]
-pub enum ReasoningSummary {
-    #[default]
-    Auto,
-    Concise,
-    Detailed,
-    /// Option to disable reasoning summaries.
-    None,
-}
-
-/// Controls output length/detail on GPT-5 models via the Responses API.
-/// Serialized with lowercase values to match the OpenAI API.
-#[derive(Debug, Serialize, Deserialize, Default, Clone, Copy, PartialEq, Eq, Display)]
-#[serde(rename_all = "lowercase")]
-#[strum(serialize_all = "lowercase")]
-pub enum Verbosity {
-    Low,
-    #[default]
-    Medium,
-    High,
+    Experimental,
 }

codex-rs/core/src/conversation_history.rs

@@ -32,32 +32,8 @@ impl ConversationHistory {
         }
     }
 
-    pub(crate) fn keep_last_messages(&mut self, n: usize) {
-        if n == 0 {
-            self.items.clear();
-            return;
-        }
-
-        // Collect the last N message items (assistant/user), newest to oldest.
-        let mut kept: Vec<ResponseItem> = Vec::with_capacity(n);
-        for item in self.items.iter().rev() {
-            if let ResponseItem::Message { role, content, .. } = item {
-                kept.push(ResponseItem::Message {
-                    // we need to remove the id or the model will complain that messages are sent without
-                    // their reasonings
-                    id: None,
-                    role: role.clone(),
-                    content: content.clone(),
-                });
-                if kept.len() == n {
-                    break;
-                }
-            }
-        }
-
-        // Preserve chronological order (oldest to newest) within the kept slice.
-        kept.reverse();
-        self.items = kept;
+    pub(crate) fn replace(&mut self, items: Vec<ResponseItem>) {
+        self.items = items;
     }
 }
 

codex-rs/core/src/conversation_manager.rs

@@ -1,11 +1,5 @@
-use std::collections::HashMap;
-use std::sync::Arc;
-
-use codex_login::AuthManager;
-use codex_login::CodexAuth;
-use tokio::sync::RwLock;
-use uuid::Uuid;
-
+use crate::AuthManager;
+use crate::CodexAuth;
 use crate::codex::Codex;
 use crate::codex::CodexSpawnOk;
 use crate::codex::INITIAL_SUBMIT_ID;
@@ -16,12 +10,20 @@ use crate::error::Result as CodexResult;
 use crate::protocol::Event;
 use crate::protocol::EventMsg;
 use crate::protocol::SessionConfiguredEvent;
+use crate::rollout::RolloutRecorder;
+use codex_protocol::mcp_protocol::ConversationId;
 use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::InitialHistory;
+use codex_protocol::protocol::RolloutItem;
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Arc;
+use tokio::sync::RwLock;
 
 /// Represents a newly created Codex conversation, including the first event
 /// (which is [`EventMsg::SessionConfigured`]).
 pub struct NewConversation {
-    pub conversation_id: Uuid,
+    pub conversation_id: ConversationId,
     pub conversation: Arc<CodexConversation>,
     pub session_configured: SessionConfiguredEvent,
 }
@@ -29,7 +31,7 @@ pub struct NewConversation {
 /// [`ConversationManager`] is responsible for creating conversations and
 /// maintaining them in memory.
 pub struct ConversationManager {
-    conversations: Arc<RwLock<HashMap<Uuid, Arc<CodexConversation>>>>,
+    conversations: Arc<RwLock<HashMap<ConversationId, Arc<CodexConversation>>>>,
     auth_manager: Arc<AuthManager>,
 }
 
@@ -44,7 +46,7 @@ impl ConversationManager {
     /// Construct with a dummy AuthManager containing the provided CodexAuth.
     /// Used for integration tests: should not be used by ordinary business logic.
     pub fn with_auth(auth: CodexAuth) -> Self {
-        Self::new(codex_login::AuthManager::from_auth_for_testing(auth))
+        Self::new(crate::AuthManager::from_auth_for_testing(auth))
     }
 
     pub async fn new_conversation(&self, config: Config) -> CodexResult<NewConversation> {
@@ -59,18 +61,15 @@ impl ConversationManager {
     ) -> CodexResult<NewConversation> {
         let CodexSpawnOk {
             codex,
-            session_id: conversation_id,
-        } = {
-            let initial_history = None;
-            Codex::spawn(config, auth_manager, initial_history).await?
-        };
+            conversation_id,
+        } = Codex::spawn(config, auth_manager, InitialHistory::New).await?;
         self.finalize_spawn(codex, conversation_id).await
     }
 
     async fn finalize_spawn(
         &self,
         codex: Codex,
-        conversation_id: Uuid,
+        conversation_id: ConversationId,
     ) -> CodexResult<NewConversation> {
         // The first event must be `SessionInitialized`. Validate and forward it
         // to the caller so that they can display it in the conversation
@@ -101,7 +100,7 @@ impl ConversationManager {
 
     pub async fn get_conversation(
         &self,
-        conversation_id: Uuid,
+        conversation_id: ConversationId,
     ) -> CodexResult<Arc<CodexConversation>> {
         let conversations = self.conversations.read().await;
         conversations
@@ -110,8 +109,29 @@ impl ConversationManager {
             .ok_or_else(|| CodexErr::ConversationNotFound(conversation_id))
     }
 
-    pub async fn remove_conversation(&self, conversation_id: Uuid) {
-        self.conversations.write().await.remove(&conversation_id);
+    pub async fn resume_conversation_from_rollout(
+        &self,
+        config: Config,
+        rollout_path: PathBuf,
+        auth_manager: Arc<AuthManager>,
+    ) -> CodexResult<NewConversation> {
+        let initial_history = RolloutRecorder::get_rollout_history(&rollout_path).await?;
+        let CodexSpawnOk {
+            codex,
+            conversation_id,
+        } = Codex::spawn(config, auth_manager, initial_history).await?;
+        self.finalize_spawn(codex, conversation_id).await
+    }
+
+    /// Removes the conversation from the manager's internal map, though the
+    /// conversation is stored as `Arc<CodexConversation>`, it is possible that
+    /// other references to it exist elsewhere. Returns the conversation if the
+    /// conversation was found and removed.
+    pub async fn remove_conversation(
+        &self,
+        conversation_id: &ConversationId,
+    ) -> Option<Arc<CodexConversation>> {
+        self.conversations.write().await.remove(conversation_id)
     }
 
     /// Fork an existing conversation by dropping the last `drop_last_messages`
@@ -120,20 +140,20 @@ impl ConversationManager {
     /// caller's `config`). The new conversation will have a fresh id.
     pub async fn fork_conversation(
         &self,
-        conversation_history: Vec<ResponseItem>,
         num_messages_to_drop: usize,
         config: Config,
+        path: PathBuf,
     ) -> CodexResult<NewConversation> {
         // Compute the prefix up to the cut point.
-        let truncated_history =
-            truncate_after_dropping_last_messages(conversation_history, num_messages_to_drop);
+        let history = RolloutRecorder::get_rollout_history(&path).await?;
+        let history = truncate_after_dropping_last_messages(history, num_messages_to_drop);
 
         // Spawn a new conversation with the computed initial history.
         let auth_manager = self.auth_manager.clone();
         let CodexSpawnOk {
             codex,
-            session_id: conversation_id,
-        } = Codex::spawn(config, auth_manager, Some(truncated_history)).await?;
+            conversation_id,
+        } = Codex::spawn(config, auth_manager, history).await?;
 
         self.finalize_spawn(codex, conversation_id).await
     }
@@ -141,31 +161,37 @@ impl ConversationManager {
 
 /// Return a prefix of `items` obtained by dropping the last `n` user messages
 /// and all items that follow them.
-fn truncate_after_dropping_last_messages(items: Vec<ResponseItem>, n: usize) -> Vec<ResponseItem> {
-    if n == 0 || items.is_empty() {
-        return items;
+fn truncate_after_dropping_last_messages(history: InitialHistory, n: usize) -> InitialHistory {
+    if n == 0 {
+        return InitialHistory::Forked(history.get_rollout_items());
     }
 
-    // Walk backwards counting only `user` Message items, find cut index.
-    let mut count = 0usize;
-    let mut cut_index = 0usize;
-    for (idx, item) in items.iter().enumerate().rev() {
-        if let ResponseItem::Message { role, .. } = item
+    // Work directly on rollout items, and cut the vector at the nth-from-last user message input.
+    let items: Vec<RolloutItem> = history.get_rollout_items();
+
+    // Find indices of user message inputs in rollout order.
+    let mut user_positions: Vec<usize> = Vec::new();
+    for (idx, item) in items.iter().enumerate() {
+        if let RolloutItem::ResponseItem(ResponseItem::Message { role, .. }) = item
             && role == "user"
         {
-            count += 1;
-            if count == n {
-                // Cut everything from this user message to the end.
-                cut_index = idx;
-                break;
-            }
+            user_positions.push(idx);
         }
     }
-    if count < n {
-        // If fewer than n messages exist, drop everything.
-        Vec::new()
+
+    // If fewer than n user messages exist, treat as empty.
+    if user_positions.len() < n {
+        return InitialHistory::New;
+    }
+
+    // Cut strictly before the nth-from-last user message (do not keep the nth itself).
+    let cut_idx = user_positions[user_positions.len() - n];
+    let rolled: Vec<RolloutItem> = items.into_iter().take(cut_idx).collect();
+
+    if rolled.is_empty() {
+        InitialHistory::New
     } else {
-        items.into_iter().take(cut_index).collect()
+        InitialHistory::Forked(rolled)
     }
 }
 
@@ -220,13 +246,30 @@ mod tests {
             assistant_msg("a4"),
         ];
 
-        let truncated = truncate_after_dropping_last_messages(items.clone(), 1);
+        // Wrap as InitialHistory::Forked with response items only.
+        let initial: Vec<RolloutItem> = items
+            .iter()
+            .cloned()
+            .map(RolloutItem::ResponseItem)
+            .collect();
+        let truncated = truncate_after_dropping_last_messages(InitialHistory::Forked(initial), 1);
+        let got_items = truncated.get_rollout_items();
+        let expected_items = vec![
+            RolloutItem::ResponseItem(items[0].clone()),
+            RolloutItem::ResponseItem(items[1].clone()),
+            RolloutItem::ResponseItem(items[2].clone()),
+        ];
         assert_eq!(
-            truncated,
-            vec![items[0].clone(), items[1].clone(), items[2].clone()]
+            serde_json::to_value(&got_items).unwrap(),
+            serde_json::to_value(&expected_items).unwrap()
         );
 
-        let truncated2 = truncate_after_dropping_last_messages(items, 2);
-        assert!(truncated2.is_empty());
+        let initial2: Vec<RolloutItem> = items
+            .iter()
+            .cloned()
+            .map(RolloutItem::ResponseItem)
+            .collect();
+        let truncated2 = truncate_after_dropping_last_messages(InitialHistory::Forked(initial2), 2);
+        assert!(matches!(truncated2, InitialHistory::New));
     }
 }

codex-rs/core/src/default_client.rs

@@ -0,0 +1,212 @@
+use reqwest::header::HeaderValue;
+use std::sync::LazyLock;
+use std::sync::Mutex;
+
+/// Set this to add a suffix to the User-Agent string.
+///
+/// It is not ideal that we're using a global singleton for this.
+/// This is primarily designed to differentiate MCP clients from each other.
+/// Because there can only be one MCP server per process, it should be safe for this to be a global static.
+/// However, future users of this should use this with caution as a result.
+/// In addition, we want to be confident that this value is used for ALL clients and doing that requires a
+/// lot of wiring and it's easy to miss code paths by doing so.
+/// See https://github.com/openai/codex/pull/3388/files for an example of what that would look like.
+/// Finally, we want to make sure this is set for ALL mcp clients without needing to know a special env var
+/// or having to set data that they already specified in the mcp initialize request somewhere else.
+///
+/// A space is automatically added between the suffix and the rest of the User-Agent string.
+/// The full user agent string is returned from the mcp initialize response.
+/// Parenthesis will be added by Codex. This should only specify what goes inside of the parenthesis.
+pub static USER_AGENT_SUFFIX: LazyLock<Mutex<Option<String>>> = LazyLock::new(|| Mutex::new(None));
+
+pub const CODEX_INTERNAL_ORIGINATOR_OVERRIDE_ENV_VAR: &str = "CODEX_INTERNAL_ORIGINATOR_OVERRIDE";
+
+#[derive(Debug, Clone)]
+pub struct Originator {
+    pub value: String,
+    pub header_value: HeaderValue,
+}
+
+pub static ORIGINATOR: LazyLock<Originator> = LazyLock::new(|| {
+    let default = "codex_cli_rs";
+    let value = std::env::var(CODEX_INTERNAL_ORIGINATOR_OVERRIDE_ENV_VAR)
+        .unwrap_or_else(|_| default.to_string());
+
+    match HeaderValue::from_str(&value) {
+        Ok(header_value) => Originator {
+            value,
+            header_value,
+        },
+        Err(e) => {
+            tracing::error!("Unable to turn originator override {value} into header value: {e}");
+            Originator {
+                value: default.to_string(),
+                header_value: HeaderValue::from_static(default),
+            }
+        }
+    }
+});
+
+pub fn get_codex_user_agent() -> String {
+    let build_version = env!("CARGO_PKG_VERSION");
+    let os_info = os_info::get();
+    let prefix = format!(
+        "{}/{build_version} ({} {}; {}) {}",
+        ORIGINATOR.value.as_str(),
+        os_info.os_type(),
+        os_info.version(),
+        os_info.architecture().unwrap_or("unknown"),
+        crate::terminal::user_agent()
+    );
+    let suffix = USER_AGENT_SUFFIX
+        .lock()
+        .ok()
+        .and_then(|guard| guard.clone());
+    let suffix = suffix
+        .as_deref()
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .map_or_else(String::new, |value| format!(" ({value})"));
+
+    let candidate = format!("{prefix}{suffix}");
+    sanitize_user_agent(candidate, &prefix)
+}
+
+/// Sanitize the user agent string.
+///
+/// Invalid characters are replaced with an underscore.
+///
+/// If the user agent fails to parse, it falls back to fallback and then to ORIGINATOR.
+fn sanitize_user_agent(candidate: String, fallback: &str) -> String {
+    if HeaderValue::from_str(candidate.as_str()).is_ok() {
+        return candidate;
+    }
+
+    let sanitized: String = candidate
+        .chars()
+        .map(|ch| if matches!(ch, ' '..='~') { ch } else { '_' })
+        .collect();
+    if !sanitized.is_empty() && HeaderValue::from_str(sanitized.as_str()).is_ok() {
+        tracing::warn!(
+            "Sanitized Codex user agent because provided suffix contained invalid header characters"
+        );
+        sanitized
+    } else if HeaderValue::from_str(fallback).is_ok() {
+        tracing::warn!(
+            "Falling back to base Codex user agent because provided suffix could not be sanitized"
+        );
+        fallback.to_string()
+    } else {
+        tracing::warn!(
+            "Falling back to default Codex originator because base user agent string is invalid"
+        );
+        ORIGINATOR.value.clone()
+    }
+}
+
+/// Create a reqwest client with default `originator` and `User-Agent` headers set.
+pub fn create_client() -> reqwest::Client {
+    use reqwest::header::HeaderMap;
+
+    let mut headers = HeaderMap::new();
+    headers.insert("originator", ORIGINATOR.header_value.clone());
+    let ua = get_codex_user_agent();
+
+    reqwest::Client::builder()
+        // Set UA via dedicated helper to avoid header validation pitfalls
+        .user_agent(ua)
+        .default_headers(headers)
+        .build()
+        .unwrap_or_else(|_| reqwest::Client::new())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_get_codex_user_agent() {
+        let user_agent = get_codex_user_agent();
+        assert!(user_agent.starts_with("codex_cli_rs/"));
+    }
+
+    #[tokio::test]
+    async fn test_create_client_sets_default_headers() {
+        use wiremock::Mock;
+        use wiremock::MockServer;
+        use wiremock::ResponseTemplate;
+        use wiremock::matchers::method;
+        use wiremock::matchers::path;
+
+        let client = create_client();
+
+        // Spin up a local mock server and capture a request.
+        let server = MockServer::start().await;
+        Mock::given(method("GET"))
+            .and(path("/"))
+            .respond_with(ResponseTemplate::new(200))
+            .mount(&server)
+            .await;
+
+        let resp = client
+            .get(server.uri())
+            .send()
+            .await
+            .expect("failed to send request");
+        assert!(resp.status().is_success());
+
+        let requests = server
+            .received_requests()
+            .await
+            .expect("failed to fetch received requests");
+        assert!(!requests.is_empty());
+        let headers = &requests[0].headers;
+
+        // originator header is set to the provided value
+        let originator_header = headers
+            .get("originator")
+            .expect("originator header missing");
+        assert_eq!(originator_header.to_str().unwrap(), "codex_cli_rs");
+
+        // User-Agent matches the computed Codex UA for that originator
+        let expected_ua = get_codex_user_agent();
+        let ua_header = headers
+            .get("user-agent")
+            .expect("user-agent header missing");
+        assert_eq!(ua_header.to_str().unwrap(), expected_ua);
+    }
+
+    #[test]
+    fn test_invalid_suffix_is_sanitized() {
+        let prefix = "codex_cli_rs/0.0.0";
+        let suffix = "bad\rsuffix";
+
+        assert_eq!(
+            sanitize_user_agent(format!("{prefix} ({suffix})"), prefix),
+            "codex_cli_rs/0.0.0 (bad_suffix)"
+        );
+    }
+
+    #[test]
+    fn test_invalid_suffix_is_sanitized2() {
+        let prefix = "codex_cli_rs/0.0.0";
+        let suffix = "bad\0suffix";
+
+        assert_eq!(
+            sanitize_user_agent(format!("{prefix} ({suffix})"), prefix),
+            "codex_cli_rs/0.0.0 (bad_suffix)"
+        );
+    }
+
+    #[test]
+    #[cfg(target_os = "macos")]
+    fn test_macos() {
+        use regex_lite::Regex;
+        let user_agent = get_codex_user_agent();
+        let re = Regex::new(
+            r"^codex_cli_rs/\d+\.\d+\.\d+ \(Mac OS \d+\.\d+\.\d+; (x86_64|arm64)\) (\S+)$",
+        )
+        .unwrap();
+        assert!(re.is_match(&user_agent));
+    }
+}

codex-rs/core/src/environment_context.rs

@@ -8,12 +8,10 @@ use crate::shell::Shell;
 use codex_protocol::config_types::SandboxMode;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::ENVIRONMENT_CONTEXT_CLOSE_TAG;
+use codex_protocol::protocol::ENVIRONMENT_CONTEXT_OPEN_TAG;
 use std::path::PathBuf;
 
-/// wraps environment context message in a tag for the model to parse more easily.
-pub(crate) const ENVIRONMENT_CONTEXT_START: &str = "<environment_context>";
-pub(crate) const ENVIRONMENT_CONTEXT_END: &str = "</environment_context>";
-
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, DeriveDisplay)]
 #[serde(rename_all = "kebab-case")]
 #[strum(serialize_all = "kebab-case")]
@@ -28,6 +26,7 @@ pub(crate) struct EnvironmentContext {
     pub approval_policy: Option<AskForApproval>,
     pub sandbox_mode: Option<SandboxMode>,
     pub network_access: Option<NetworkAccess>,
+    pub writable_roots: Option<Vec<PathBuf>>,
     pub shell: Option<Shell>,
 }
 
@@ -59,6 +58,16 @@ impl EnvironmentContext {
                 }
                 None => None,
             },
+            writable_roots: match sandbox_policy {
+                Some(SandboxPolicy::WorkspaceWrite { writable_roots, .. }) => {
+                    if writable_roots.is_empty() {
+                        None
+                    } else {
+                        Some(writable_roots)
+                    }
+                }
+                _ => None,
+            },
             shell,
         }
     }
@@ -74,12 +83,13 @@ impl EnvironmentContext {
     ///   <cwd>...</cwd>
     ///   <approval_policy>...</approval_policy>
     ///   <sandbox_mode>...</sandbox_mode>
+    ///   <writable_roots>...</writable_roots>
     ///   <network_access>...</network_access>
     ///   <shell>...</shell>
     /// </environment_context>
     /// ```
     pub fn serialize_to_xml(self) -> String {
-        let mut lines = vec![ENVIRONMENT_CONTEXT_START.to_string()];
+        let mut lines = vec![ENVIRONMENT_CONTEXT_OPEN_TAG.to_string()];
         if let Some(cwd) = self.cwd {
             lines.push(format!("  <cwd>{}</cwd>", cwd.to_string_lossy()));
         }
@@ -96,12 +106,22 @@ impl EnvironmentContext {
                 "  <network_access>{network_access}</network_access>"
             ));
         }
+        if let Some(writable_roots) = self.writable_roots {
+            lines.push("  <writable_roots>".to_string());
+            for writable_root in writable_roots {
+                lines.push(format!(
+                    "    <root>{}</root>",
+                    writable_root.to_string_lossy()
+                ));
+            }
+            lines.push("  </writable_roots>".to_string());
+        }
         if let Some(shell) = self.shell
             && let Some(shell_name) = shell.name()
         {
             lines.push(format!("  <shell>{shell_name}</shell>"));
         }
-        lines.push(ENVIRONMENT_CONTEXT_END.to_string());
+        lines.push(ENVIRONMENT_CONTEXT_CLOSE_TAG.to_string());
         lines.join("\n")
     }
 }
@@ -117,3 +137,77 @@ impl From<EnvironmentContext> for ResponseItem {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use pretty_assertions::assert_eq;
+
+    fn workspace_write_policy(writable_roots: Vec<&str>, network_access: bool) -> SandboxPolicy {
+        SandboxPolicy::WorkspaceWrite {
+            writable_roots: writable_roots.into_iter().map(PathBuf::from).collect(),
+            network_access,
+            exclude_tmpdir_env_var: false,
+            exclude_slash_tmp: false,
+        }
+    }
+
+    #[test]
+    fn serialize_workspace_write_environment_context() {
+        let context = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(workspace_write_policy(vec!["/repo", "/tmp"], false)),
+            None,
+        );
+
+        let expected = r#"<environment_context>
+  <cwd>/repo</cwd>
+  <approval_policy>on-request</approval_policy>
+  <sandbox_mode>workspace-write</sandbox_mode>
+  <network_access>restricted</network_access>
+  <writable_roots>
+    <root>/repo</root>
+    <root>/tmp</root>
+  </writable_roots>
+</environment_context>"#;
+
+        assert_eq!(context.serialize_to_xml(), expected);
+    }
+
+    #[test]
+    fn serialize_read_only_environment_context() {
+        let context = EnvironmentContext::new(
+            None,
+            Some(AskForApproval::Never),
+            Some(SandboxPolicy::ReadOnly),
+            None,
+        );
+
+        let expected = r#"<environment_context>
+  <approval_policy>never</approval_policy>
+  <sandbox_mode>read-only</sandbox_mode>
+  <network_access>restricted</network_access>
+</environment_context>"#;
+
+        assert_eq!(context.serialize_to_xml(), expected);
+    }
+
+    #[test]
+    fn serialize_full_access_environment_context() {
+        let context = EnvironmentContext::new(
+            None,
+            Some(AskForApproval::OnFailure),
+            Some(SandboxPolicy::DangerFullAccess),
+            None,
+        );
+
+        let expected = r#"<environment_context>
+  <approval_policy>on-failure</approval_policy>
+  <sandbox_mode>danger-full-access</sandbox_mode>
+  <network_access>enabled</network_access>
+</environment_context>"#;
+
+        assert_eq!(context.serialize_to_xml(), expected);
+    }
+}

codex-rs/core/src/error.rs

@@ -1,18 +1,24 @@
+use crate::exec::ExecToolCallOutput;
+use crate::token_data::KnownPlan;
+use crate::token_data::PlanType;
+use codex_protocol::mcp_protocol::ConversationId;
 use reqwest::StatusCode;
 use serde_json;
 use std::io;
 use std::time::Duration;
 use thiserror::Error;
 use tokio::task::JoinError;
-use uuid::Uuid;
 
 pub type Result<T> = std::result::Result<T, CodexErr>;
 
 #[derive(Error, Debug)]
 pub enum SandboxErr {
     /// Error from sandbox execution
-    #[error("sandbox denied exec error, exit code: {0}, stdout: {1}, stderr: {2}")]
-    Denied(i32, String, String),
+    #[error(
+        "sandbox denied exec error, exit code: {}, stdout: {}, stderr: {}",
+        .output.exit_code, .output.stdout.text, .output.stderr.text
+    )]
+    Denied { output: Box<ExecToolCallOutput> },
 
     /// Error from linux seccomp filter setup
     #[cfg(target_os = "linux")]
@@ -26,7 +32,7 @@ pub enum SandboxErr {
 
     /// Command timed out
     #[error("command timed out")]
-    Timeout,
+    Timeout { output: Box<ExecToolCallOutput> },
 
     /// Command was killed by a signal
     #[error("command was killed by a signal")]
@@ -49,7 +55,7 @@ pub enum CodexErr {
     Stream(String, Option<Duration>),
 
     #[error("no conversation with id: {0}")]
-    ConversationNotFound(Uuid),
+    ConversationNotFound(ConversationId),
 
     #[error("session configured event was not the first event in the stream")]
     SessionConfiguredNotFirstEvent,
@@ -127,38 +133,58 @@ pub enum CodexErr {
 
 #[derive(Debug)]
 pub struct UsageLimitReachedError {
-    pub plan_type: Option<String>,
-    pub resets_in_seconds: Option<u64>,
+    pub(crate) plan_type: Option<PlanType>,
+    pub(crate) resets_in_seconds: Option<u64>,
 }
 
 impl std::fmt::Display for UsageLimitReachedError {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        // Base message differs slightly for legacy ChatGPT Plus plan users.
-        if let Some(plan_type) = &self.plan_type
-            && plan_type == "plus"
-        {
-            write!(
-                f,
-                "You've hit your usage limit. Upgrade to Pro (https://openai.com/chatgpt/pricing) or try again"
-            )?;
-            if let Some(secs) = self.resets_in_seconds {
-                let reset_duration = format_reset_duration(secs);
-                write!(f, " in {reset_duration}.")?;
-            } else {
-                write!(f, " later.")?;
+        let message = match self.plan_type.as_ref() {
+            Some(PlanType::Known(KnownPlan::Plus)) => format!(
+                "You've hit your usage limit. Upgrade to Pro (https://openai.com/chatgpt/pricing){}",
+                retry_suffix_after_or(self.resets_in_seconds)
+            ),
+            Some(PlanType::Known(KnownPlan::Team)) | Some(PlanType::Known(KnownPlan::Business)) => {
+                format!(
+                    "You've hit your usage limit. To get more access now, send a request to your admin{}",
+                    retry_suffix_after_or(self.resets_in_seconds)
+                )
             }
-        } else {
-            write!(f, "You've hit your usage limit.")?;
-
-            if let Some(secs) = self.resets_in_seconds {
-                let reset_duration = format_reset_duration(secs);
-                write!(f, " Try again in {reset_duration}.")?;
-            } else {
-                write!(f, " Try again later.")?;
+            Some(PlanType::Known(KnownPlan::Free)) => {
+                "To use Codex with your ChatGPT plan, upgrade to Plus: https://openai.com/chatgpt/pricing."
+                    .to_string()
             }
-        }
+            Some(PlanType::Known(KnownPlan::Pro))
+            | Some(PlanType::Known(KnownPlan::Enterprise))
+            | Some(PlanType::Known(KnownPlan::Edu)) => format!(
+                "You've hit your usage limit.{}",
+                retry_suffix(self.resets_in_seconds)
+            ),
+            Some(PlanType::Unknown(_)) | None => format!(
+                "You've hit your usage limit.{}",
+                retry_suffix(self.resets_in_seconds)
+            ),
+        };
 
-        Ok(())
+        write!(f, "{message}")
+    }
+}
+
+fn retry_suffix(resets_in_seconds: Option<u64>) -> String {
+    if let Some(secs) = resets_in_seconds {
+        let reset_duration = format_reset_duration(secs);
+        format!(" Try again in {reset_duration}.")
+    } else {
+        " Try again later.".to_string()
+    }
+}
+
+fn retry_suffix_after_or(resets_in_seconds: Option<u64>) -> String {
+    if let Some(secs) = resets_in_seconds {
+        let reset_duration = format_reset_duration(secs);
+        format!(" or try again in {reset_duration}.")
+    } else {
+        " or try again later.".to_string()
     }
 }
 
@@ -223,9 +249,12 @@ impl CodexErr {
 
 pub fn get_error_message_ui(e: &CodexErr) -> String {
     match e {
-        CodexErr::Sandbox(SandboxErr::Denied(_, _, stderr)) => stderr.to_string(),
+        CodexErr::Sandbox(SandboxErr::Denied { output }) => output.stderr.text.clone(),
         // Timeouts are not sandbox errors from a UX perspective; present them plainly
-        CodexErr::Sandbox(SandboxErr::Timeout) => "error: command timed out".to_string(),
+        CodexErr::Sandbox(SandboxErr::Timeout { output }) => format!(
+            "error: command timed out after {} ms",
+            output.duration.as_millis()
+        ),
         _ => e.to_string(),
     }
 }
@@ -237,7 +266,7 @@ mod tests {
     #[test]
     fn usage_limit_reached_error_formats_plus_plan() {
         let err = UsageLimitReachedError {
-            plan_type: Some("plus".to_string()),
+            plan_type: Some(PlanType::Known(KnownPlan::Plus)),
             resets_in_seconds: None,
         };
         assert_eq!(
@@ -246,6 +275,18 @@ mod tests {
         );
     }
 
+    #[test]
+    fn usage_limit_reached_error_formats_free_plan() {
+        let err = UsageLimitReachedError {
+            plan_type: Some(PlanType::Known(KnownPlan::Free)),
+            resets_in_seconds: Some(3600),
+        };
+        assert_eq!(
+            err.to_string(),
+            "To use Codex with your ChatGPT plan, upgrade to Plus: https://openai.com/chatgpt/pricing."
+        );
+    }
+
     #[test]
     fn usage_limit_reached_error_formats_default_when_none() {
         let err = UsageLimitReachedError {
@@ -258,10 +299,34 @@ mod tests {
         );
     }
 
+    #[test]
+    fn usage_limit_reached_error_formats_team_plan() {
+        let err = UsageLimitReachedError {
+            plan_type: Some(PlanType::Known(KnownPlan::Team)),
+            resets_in_seconds: Some(3600),
+        };
+        assert_eq!(
+            err.to_string(),
+            "You've hit your usage limit. To get more access now, send a request to your admin or try again in 1 hour."
+        );
+    }
+
+    #[test]
+    fn usage_limit_reached_error_formats_business_plan_without_reset() {
+        let err = UsageLimitReachedError {
+            plan_type: Some(PlanType::Known(KnownPlan::Business)),
+            resets_in_seconds: None,
+        };
+        assert_eq!(
+            err.to_string(),
+            "You've hit your usage limit. To get more access now, send a request to your admin or try again later."
+        );
+    }
+
     #[test]
     fn usage_limit_reached_error_formats_default_for_other_plans() {
         let err = UsageLimitReachedError {
-            plan_type: Some("pro".to_string()),
+            plan_type: Some(PlanType::Known(KnownPlan::Pro)),
             resets_in_seconds: None,
         };
         assert_eq!(
@@ -285,7 +350,7 @@ mod tests {
     #[test]
     fn usage_limit_reached_includes_hours_and_minutes() {
         let err = UsageLimitReachedError {
-            plan_type: Some("plus".to_string()),
+            plan_type: Some(PlanType::Known(KnownPlan::Plus)),
             resets_in_seconds: Some(3 * 3600 + 32 * 60),
         };
         assert_eq!(

codex-rs/core/src/event_mapping.rs

@@ -0,0 +1,167 @@
+use crate::protocol::AgentMessageEvent;
+use crate::protocol::AgentReasoningEvent;
+use crate::protocol::AgentReasoningRawContentEvent;
+use crate::protocol::EventMsg;
+use crate::protocol::InputMessageKind;
+use crate::protocol::UserMessageEvent;
+use crate::protocol::WebSearchEndEvent;
+use codex_protocol::models::ContentItem;
+use codex_protocol::models::ReasoningItemContent;
+use codex_protocol::models::ReasoningItemReasoningSummary;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::models::WebSearchAction;
+
+/// Convert a `ResponseItem` into zero or more `EventMsg` values that the UI can render.
+///
+/// When `show_raw_agent_reasoning` is false, raw reasoning content events are omitted.
+pub(crate) fn map_response_item_to_event_messages(
+    item: &ResponseItem,
+    show_raw_agent_reasoning: bool,
+) -> Vec<EventMsg> {
+    match item {
+        ResponseItem::Message { role, content, .. } => {
+            // Do not surface system messages as user events.
+            if role == "system" {
+                return Vec::new();
+            }
+
+            let mut events: Vec<EventMsg> = Vec::new();
+            let mut message_parts: Vec<String> = Vec::new();
+            let mut images: Vec<String> = Vec::new();
+            let mut kind: Option<InputMessageKind> = None;
+
+            for content_item in content.iter() {
+                match content_item {
+                    ContentItem::InputText { text } => {
+                        if kind.is_none() {
+                            let trimmed = text.trim_start();
+                            kind = if trimmed.starts_with("<environment_context>") {
+                                Some(InputMessageKind::EnvironmentContext)
+                            } else if trimmed.starts_with("<user_instructions>") {
+                                Some(InputMessageKind::UserInstructions)
+                            } else {
+                                Some(InputMessageKind::Plain)
+                            };
+                        }
+                        message_parts.push(text.clone());
+                    }
+                    ContentItem::InputImage { image_url } => {
+                        images.push(image_url.clone());
+                    }
+                    ContentItem::OutputText { text } => {
+                        events.push(EventMsg::AgentMessage(AgentMessageEvent {
+                            message: text.clone(),
+                        }));
+                    }
+                }
+            }
+
+            if !message_parts.is_empty() || !images.is_empty() {
+                let message = if message_parts.is_empty() {
+                    String::new()
+                } else {
+                    message_parts.join("")
+                };
+                let images = if images.is_empty() {
+                    None
+                } else {
+                    Some(images)
+                };
+
+                events.push(EventMsg::UserMessage(UserMessageEvent {
+                    message,
+                    kind,
+                    images,
+                }));
+            }
+
+            events
+        }
+
+        ResponseItem::Reasoning {
+            summary, content, ..
+        } => {
+            let mut events = Vec::new();
+            for ReasoningItemReasoningSummary::SummaryText { text } in summary {
+                events.push(EventMsg::AgentReasoning(AgentReasoningEvent {
+                    text: text.clone(),
+                }));
+            }
+            if let Some(items) = content.as_ref().filter(|_| show_raw_agent_reasoning) {
+                for c in items {
+                    let text = match c {
+                        ReasoningItemContent::ReasoningText { text }
+                        | ReasoningItemContent::Text { text } => text,
+                    };
+                    events.push(EventMsg::AgentReasoningRawContent(
+                        AgentReasoningRawContentEvent { text: text.clone() },
+                    ));
+                }
+            }
+            events
+        }
+
+        ResponseItem::WebSearchCall { id, action, .. } => match action {
+            WebSearchAction::Search { query } => {
+                let call_id = id.clone().unwrap_or_else(|| "".to_string());
+                vec![EventMsg::WebSearchEnd(WebSearchEndEvent {
+                    call_id,
+                    query: query.clone(),
+                })]
+            }
+            WebSearchAction::Other => Vec::new(),
+        },
+
+        // Variants that require side effects are handled by higher layers and do not emit events here.
+        ResponseItem::FunctionCall { .. }
+        | ResponseItem::FunctionCallOutput { .. }
+        | ResponseItem::LocalShellCall { .. }
+        | ResponseItem::CustomToolCall { .. }
+        | ResponseItem::CustomToolCallOutput { .. }
+        | ResponseItem::Other => Vec::new(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::map_response_item_to_event_messages;
+    use crate::protocol::EventMsg;
+    use crate::protocol::InputMessageKind;
+    use codex_protocol::models::ContentItem;
+    use codex_protocol::models::ResponseItem;
+    use pretty_assertions::assert_eq;
+
+    #[test]
+    fn maps_user_message_with_text_and_two_images() {
+        let img1 = "https://example.com/one.png".to_string();
+        let img2 = "https://example.com/two.jpg".to_string();
+
+        let item = ResponseItem::Message {
+            id: None,
+            role: "user".to_string(),
+            content: vec![
+                ContentItem::InputText {
+                    text: "Hello world".to_string(),
+                },
+                ContentItem::InputImage {
+                    image_url: img1.clone(),
+                },
+                ContentItem::InputImage {
+                    image_url: img2.clone(),
+                },
+            ],
+        };
+
+        let events = map_response_item_to_event_messages(&item, false);
+        assert_eq!(events.len(), 1, "expected a single user message event");
+
+        match &events[0] {
+            EventMsg::UserMessage(user) => {
+                assert_eq!(user.message, "Hello world");
+                assert!(matches!(user.kind, Some(InputMessageKind::Plain)));
+                assert_eq!(user.images, Some(vec![img1, img2]));
+            }
+            other => panic!("expected UserMessage, got {other:?}"),
+        }
+    }
+}

codex-rs/core/src/exec.rs

@@ -26,7 +26,6 @@ use crate::protocol::SandboxPolicy;
 use crate::seatbelt::spawn_command_under_seatbelt;
 use crate::spawn::StdioPolicy;
 use crate::spawn::spawn_child_async;
-use serde_bytes::ByteBuf;
 
 const DEFAULT_TIMEOUT_MS: u64 = 10_000;
 
@@ -35,6 +34,7 @@ const DEFAULT_TIMEOUT_MS: u64 = 10_000;
 const SIGKILL_CODE: i32 = 9;
 const TIMEOUT_CODE: i32 = 64;
 const EXIT_CODE_SIGNAL_BASE: i32 = 128; // conventional shell: 128 + signal
+const EXEC_TIMEOUT_EXIT_CODE: i32 = 124; // conventional timeout exit code
 
 // I/O buffer sizing
 const READ_CHUNK_SIZE: usize = 8192; // bytes per read
@@ -87,11 +87,12 @@ pub async fn process_exec_tool_call(
 ) -> Result<ExecToolCallOutput> {
     let start = Instant::now();
 
+    let timeout_duration = params.timeout_duration();
+
     let raw_output_result: std::result::Result<RawExecToolCallOutput, CodexErr> = match sandbox_type
     {
         SandboxType::None => exec(params, sandbox_policy, stdout_stream.clone()).await,
         SandboxType::MacosSeatbelt => {
-            let timeout = params.timeout_duration();
             let ExecParams {
                 command, cwd, env, ..
             } = params;
@@ -103,10 +104,9 @@ pub async fn process_exec_tool_call(
                 env,
             )
             .await?;
-            consume_truncated_output(child, timeout, stdout_stream.clone()).await
+            consume_truncated_output(child, timeout_duration, stdout_stream.clone()).await
         }
         SandboxType::LinuxSeccomp => {
-            let timeout = params.timeout_duration();
             let ExecParams {
                 command, cwd, env, ..
             } = params;
@@ -124,41 +124,56 @@ pub async fn process_exec_tool_call(
             )
             .await?;
 
-            consume_truncated_output(child, timeout, stdout_stream).await
+            consume_truncated_output(child, timeout_duration, stdout_stream).await
         }
     };
     let duration = start.elapsed();
     match raw_output_result {
         Ok(raw_output) => {
-            let stdout = raw_output.stdout.from_utf8_lossy();
-            let stderr = raw_output.stderr.from_utf8_lossy();
+            #[allow(unused_mut)]
+            let mut timed_out = raw_output.timed_out;
 
             #[cfg(target_family = "unix")]
-            match raw_output.exit_status.signal() {
-                Some(TIMEOUT_CODE) => return Err(CodexErr::Sandbox(SandboxErr::Timeout)),
-                Some(signal) => {
-                    return Err(CodexErr::Sandbox(SandboxErr::Signal(signal)));
+            {
+                if let Some(signal) = raw_output.exit_status.signal() {
+                    if signal == TIMEOUT_CODE {
+                        timed_out = true;
+                    } else {
+                        return Err(CodexErr::Sandbox(SandboxErr::Signal(signal)));
+                    }
                 }
-                None => {}
             }
 
-            let exit_code = raw_output.exit_status.code().unwrap_or(-1);
-
-            if exit_code != 0 && is_likely_sandbox_denied(sandbox_type, exit_code) {
-                return Err(CodexErr::Sandbox(SandboxErr::Denied(
-                    exit_code,
-                    stdout.text,
-                    stderr.text,
-                )));
+            let mut exit_code = raw_output.exit_status.code().unwrap_or(-1);
+            if timed_out {
+                exit_code = EXEC_TIMEOUT_EXIT_CODE;
             }
 
-            Ok(ExecToolCallOutput {
+            let stdout = raw_output.stdout.from_utf8_lossy();
+            let stderr = raw_output.stderr.from_utf8_lossy();
+            let aggregated_output = raw_output.aggregated_output.from_utf8_lossy();
+            let exec_output = ExecToolCallOutput {
                 exit_code,
                 stdout,
                 stderr,
-                aggregated_output: raw_output.aggregated_output.from_utf8_lossy(),
+                aggregated_output,
                 duration,
-            })
+                timed_out,
+            };
+
+            if timed_out {
+                return Err(CodexErr::Sandbox(SandboxErr::Timeout {
+                    output: Box::new(exec_output),
+                }));
+            }
+
+            if exit_code != 0 && is_likely_sandbox_denied(sandbox_type, exit_code) {
+                return Err(CodexErr::Sandbox(SandboxErr::Denied {
+                    output: Box::new(exec_output),
+                }));
+            }
+
+            Ok(exec_output)
         }
         Err(err) => {
             tracing::error!("exec error: {err}");
@@ -198,6 +213,7 @@ struct RawExecToolCallOutput {
     pub stdout: StreamOutput<Vec<u8>>,
     pub stderr: StreamOutput<Vec<u8>>,
     pub aggregated_output: StreamOutput<Vec<u8>>,
+    pub timed_out: bool,
 }
 
 impl StreamOutput<String> {
@@ -230,6 +246,7 @@ pub struct ExecToolCallOutput {
     pub stderr: StreamOutput<String>,
     pub aggregated_output: StreamOutput<String>,
     pub duration: Duration,
+    pub timed_out: bool,
 }
 
 async fn exec(
@@ -299,22 +316,24 @@ async fn consume_truncated_output(
         Some(agg_tx.clone()),
     ));
 
-    let exit_status = tokio::select! {
+    let (exit_status, timed_out) = tokio::select! {
         result = tokio::time::timeout(timeout, child.wait()) => {
             match result {
-                Ok(Ok(exit_status)) => exit_status,
-                Ok(e) => e?,
+                Ok(status_result) => {
+                    let exit_status = status_result?;
+                    (exit_status, false)
+                }
                 Err(_) => {
                     // timeout
                     child.start_kill()?;
                     // Debatable whether `child.wait().await` should be called here.
-                    synthetic_exit_status(EXIT_CODE_SIGNAL_BASE + TIMEOUT_CODE)
+                    (synthetic_exit_status(EXIT_CODE_SIGNAL_BASE + TIMEOUT_CODE), true)
                 }
             }
         }
         _ = tokio::signal::ctrl_c() => {
             child.start_kill()?;
-            synthetic_exit_status(EXIT_CODE_SIGNAL_BASE + SIGKILL_CODE)
+            (synthetic_exit_status(EXIT_CODE_SIGNAL_BASE + SIGKILL_CODE), false)
         }
     };
 
@@ -337,6 +356,7 @@ async fn consume_truncated_output(
         stdout,
         stderr,
         aggregated_output,
+        timed_out,
     })
 }
 
@@ -369,7 +389,7 @@ async fn read_capped<R: AsyncRead + Unpin + Send + 'static>(
                 } else {
                     ExecOutputStream::Stdout
                 },
-                chunk: ByteBuf::from(chunk),
+                chunk,
             });
             let event = Event {
                 id: stream.sub_id.clone(),

codex-rs/core/src/exec_command/exec_command_session.rs

@@ -11,6 +11,9 @@ pub(crate) struct ExecCommandSession {
     /// Broadcast stream of output chunks read from the PTY. New subscribers
     /// receive only chunks emitted after they subscribe.
     output_tx: broadcast::Sender<Vec<u8>>,
+    /// Receiver subscribed before the child process starts emitting output so
+    /// the first caller can consume any early data without races.
+    initial_output_rx: StdMutex<Option<broadcast::Receiver<Vec<u8>>>>,
 
     /// Child killer handle for termination on drop (can signal independently
     /// of a thread blocked in `.wait()`).
@@ -24,6 +27,9 @@ pub(crate) struct ExecCommandSession {
 
     /// JoinHandle for the child wait task.
     wait_handle: StdMutex<Option<JoinHandle<()>>>,
+
+    /// Tracks whether the underlying process has exited.
+    exit_status: std::sync::Arc<std::sync::atomic::AtomicBool>,
 }
 
 impl ExecCommandSession {
@@ -34,14 +40,25 @@ impl ExecCommandSession {
         reader_handle: JoinHandle<()>,
         writer_handle: JoinHandle<()>,
         wait_handle: JoinHandle<()>,
+        exit_status: std::sync::Arc<std::sync::atomic::AtomicBool>,
     ) -> Self {
         Self {
             writer_tx,
             output_tx,
+            initial_output_rx: StdMutex::new(None),
             killer: StdMutex::new(Some(killer)),
             reader_handle: StdMutex::new(Some(reader_handle)),
             writer_handle: StdMutex::new(Some(writer_handle)),
             wait_handle: StdMutex::new(Some(wait_handle)),
+            exit_status,
+        }
+    }
+
+    pub(crate) fn set_initial_output_receiver(&self, receiver: broadcast::Receiver<Vec<u8>>) {
+        if let Ok(mut guard) = self.initial_output_rx.lock()
+            && guard.is_none()
+        {
+            *guard = Some(receiver);
         }
     }
 
@@ -50,7 +67,17 @@ impl ExecCommandSession {
     }
 
     pub(crate) fn output_receiver(&self) -> broadcast::Receiver<Vec<u8>> {
-        self.output_tx.subscribe()
+        if let Ok(mut guard) = self.initial_output_rx.lock()
+            && let Some(receiver) = guard.take()
+        {
+            receiver
+        } else {
+            self.output_tx.subscribe()
+        }
+    }
+
+    pub(crate) fn has_exited(&self) -> bool {
+        self.exit_status.load(std::sync::atomic::Ordering::SeqCst)
     }
 }
 

codex-rs/core/src/exec_command/mod.rs

@@ -6,6 +6,7 @@ mod session_manager;
 
 pub use exec_command_params::ExecCommandParams;
 pub use exec_command_params::WriteStdinParams;
+pub(crate) use exec_command_session::ExecCommandSession;
 pub use responses_api::EXEC_COMMAND_TOOL_NAME;
 pub use responses_api::WRITE_STDIN_TOOL_NAME;
 pub use responses_api::create_exec_command_tool_for_responses_api;

codex-rs/core/src/exec_command/session_manager.rs

@@ -3,6 +3,7 @@ use std::io::ErrorKind;
 use std::io::Read;
 use std::sync::Arc;
 use std::sync::Mutex as StdMutex;
+use std::sync::atomic::AtomicBool;
 use std::sync::atomic::AtomicU32;
 
 use portable_pty::CommandBuilder;
@@ -19,6 +20,7 @@ use crate::exec_command::exec_command_params::ExecCommandParams;
 use crate::exec_command::exec_command_params::WriteStdinParams;
 use crate::exec_command::exec_command_session::ExecCommandSession;
 use crate::exec_command::session_id::SessionId;
+use crate::truncate::truncate_middle;
 use codex_protocol::models::FunctionCallOutputPayload;
 
 #[derive(Debug, Default)]
@@ -277,6 +279,7 @@ async fn create_exec_command_session(
     let (writer_tx, mut writer_rx) = mpsc::channel::<Vec<u8>>(128);
     // Broadcast for streaming PTY output to readers: subscribers receive from subscription time.
     let (output_tx, _) = tokio::sync::broadcast::channel::<Vec<u8>>(256);
+    let initial_output_rx = output_tx.subscribe();
 
     // Reader task: drain PTY and forward chunks to output channel.
     let mut reader = pair.master.try_clone_reader()?;
@@ -327,11 +330,14 @@ async fn create_exec_command_session(
 
     // Keep the child alive until it exits, then signal exit code.
     let (exit_tx, exit_rx) = oneshot::channel::<i32>();
+    let exit_status = Arc::new(AtomicBool::new(false));
+    let wait_exit_status = exit_status.clone();
     let wait_handle = tokio::task::spawn_blocking(move || {
         let code = match child.wait() {
             Ok(status) => status.exit_code() as i32,
             Err(_) => -1,
         };
+        wait_exit_status.store(true, std::sync::atomic::Ordering::SeqCst);
         let _ = exit_tx.send(code);
     });
 
@@ -343,116 +349,12 @@ async fn create_exec_command_session(
         reader_handle,
         writer_handle,
         wait_handle,
+        exit_status,
     );
+    session.set_initial_output_receiver(initial_output_rx);
     Ok((session, exit_rx))
 }
 
-/// Truncate the middle of a UTF-8 string to at most `max_bytes` bytes,
-/// preserving the beginning and the end. Returns the possibly truncated
-/// string and `Some(original_token_count)` (estimated at 4 bytes/token)
-/// if truncation occurred; otherwise returns the original string and `None`.
-fn truncate_middle(s: &str, max_bytes: usize) -> (String, Option<u64>) {
-    // No truncation needed
-    if s.len() <= max_bytes {
-        return (s.to_string(), None);
-    }
-    let est_tokens = (s.len() as u64).div_ceil(4);
-    if max_bytes == 0 {
-        // Cannot keep any content; still return a full marker (never truncated).
-        return (format!("…{est_tokens} tokens truncated…"), Some(est_tokens));
-    }
-
-    // Helper to truncate a string to a given byte length on a char boundary.
-    fn truncate_on_boundary(input: &str, max_len: usize) -> &str {
-        if input.len() <= max_len {
-            return input;
-        }
-        let mut end = max_len;
-        while end > 0 && !input.is_char_boundary(end) {
-            end -= 1;
-        }
-        &input[..end]
-    }
-
-    // Given a left/right budget, prefer newline boundaries; otherwise fall back
-    // to UTF-8 char boundaries.
-    fn pick_prefix_end(s: &str, left_budget: usize) -> usize {
-        if let Some(head) = s.get(..left_budget)
-            && let Some(i) = head.rfind('\n')
-        {
-            return i + 1; // keep the newline so suffix starts on a fresh line
-        }
-        truncate_on_boundary(s, left_budget).len()
-    }
-
-    fn pick_suffix_start(s: &str, right_budget: usize) -> usize {
-        let start_tail = s.len().saturating_sub(right_budget);
-        if let Some(tail) = s.get(start_tail..)
-            && let Some(i) = tail.find('\n')
-        {
-            return start_tail + i + 1; // start after newline
-        }
-        // Fall back to a char boundary at or after start_tail.
-        let mut idx = start_tail.min(s.len());
-        while idx < s.len() && !s.is_char_boundary(idx) {
-            idx += 1;
-        }
-        idx
-    }
-
-    // Refine marker length and budgets until stable. Marker is never truncated.
-    let mut guess_tokens = est_tokens; // worst-case: everything truncated
-    for _ in 0..4 {
-        let marker = format!("…{guess_tokens} tokens truncated…");
-        let marker_len = marker.len();
-        let keep_budget = max_bytes.saturating_sub(marker_len);
-        if keep_budget == 0 {
-            // No room for any content within the cap; return a full, untruncated marker
-            // that reflects the entire truncated content.
-            return (format!("…{est_tokens} tokens truncated…"), Some(est_tokens));
-        }
-
-        let left_budget = keep_budget / 2;
-        let right_budget = keep_budget - left_budget;
-        let prefix_end = pick_prefix_end(s, left_budget);
-        let mut suffix_start = pick_suffix_start(s, right_budget);
-        if suffix_start < prefix_end {
-            suffix_start = prefix_end;
-        }
-        let kept_content_bytes = prefix_end + (s.len() - suffix_start);
-        let truncated_content_bytes = s.len().saturating_sub(kept_content_bytes);
-        let new_tokens = (truncated_content_bytes as u64).div_ceil(4);
-        if new_tokens == guess_tokens {
-            let mut out = String::with_capacity(marker_len + kept_content_bytes + 1);
-            out.push_str(&s[..prefix_end]);
-            out.push_str(&marker);
-            // Place marker on its own line for symmetry when we keep line boundaries.
-            out.push('\n');
-            out.push_str(&s[suffix_start..]);
-            return (out, Some(est_tokens));
-        }
-        guess_tokens = new_tokens;
-    }
-
-    // Fallback: use last guess to build output.
-    let marker = format!("…{guess_tokens} tokens truncated…");
-    let marker_len = marker.len();
-    let keep_budget = max_bytes.saturating_sub(marker_len);
-    if keep_budget == 0 {
-        return (format!("…{est_tokens} tokens truncated…"), Some(est_tokens));
-    }
-    let left_budget = keep_budget / 2;
-    let right_budget = keep_budget - left_budget;
-    let prefix_end = pick_prefix_end(s, left_budget);
-    let suffix_start = pick_suffix_start(s, right_budget);
-    let mut out = String::with_capacity(marker_len + prefix_end + (s.len() - suffix_start) + 1);
-    out.push_str(&s[..prefix_end]);
-    out.push_str(&marker);
-    out.push('\n');
-    out.push_str(&s[suffix_start..]);
-    (out, Some(est_tokens))
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -616,50 +518,4 @@ Output:
 abc"#;
         assert_eq!(expected, text);
     }
-
-    #[test]
-    fn truncate_middle_no_newlines_fallback() {
-        // A long string with no newlines that exceeds the cap.
-        let s = "abcdefghijklmnopqrstuvwxyz0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ";
-        let max_bytes = 16; // force truncation
-        let (out, original) = truncate_middle(s, max_bytes);
-        // For very small caps, we return the full, untruncated marker,
-        // even if it exceeds the cap.
-        assert_eq!(out, "…16 tokens truncated…");
-        // Original string length is 62 bytes => ceil(62/4) = 16 tokens.
-        assert_eq!(original, Some(16));
-    }
-
-    #[test]
-    fn truncate_middle_prefers_newline_boundaries() {
-        // Build a multi-line string of 20 numbered lines (each "NNN\n").
-        let mut s = String::new();
-        for i in 1..=20 {
-            s.push_str(&format!("{i:03}\n"));
-        }
-        // Total length: 20 lines * 4 bytes per line = 80 bytes.
-        assert_eq!(s.len(), 80);
-
-        // Choose a cap that forces truncation while leaving room for
-        // a few lines on each side after accounting for the marker.
-        let max_bytes = 64;
-        // Expect exact output: first 4 lines, marker, last 4 lines, and correct token estimate (80/4 = 20).
-        assert_eq!(
-            truncate_middle(&s, max_bytes),
-            (
-                r#"001
-002
-003
-004
-…12 tokens truncated…
-017
-018
-019
-020
-"#
-                .to_string(),
-                Some(20)
-            )
-        );
-    }
 }

codex-rs/core/src/git_info.rs

@@ -3,6 +3,7 @@ use std::path::Path;
 use std::path::PathBuf;
 
 use codex_protocol::mcp_protocol::GitSha;
+use codex_protocol::protocol::GitInfo;
 use futures::future::join_all;
 use serde::Deserialize;
 use serde::Serialize;
@@ -43,19 +44,6 @@ pub fn get_git_repo_root(base_dir: &Path) -> Option<PathBuf> {
 /// Timeout for git commands to prevent freezing on large repositories
 const GIT_COMMAND_TIMEOUT: TokioDuration = TokioDuration::from_secs(5);
 
-#[derive(Serialize, Deserialize, Clone, Debug)]
-pub struct GitInfo {
-    /// Current commit hash (SHA)
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub commit_hash: Option<String>,
-    /// Current branch name
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub branch: Option<String>,
-    /// Repository URL (if available from remote)
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub repository_url: Option<String>,
-}
-
 #[derive(Serialize, Deserialize, Clone, Debug)]
 pub struct GitDiffToRemote {
     pub sha: GitSha,
@@ -814,7 +802,7 @@ mod tests {
     async fn resolve_root_git_project_for_trust_regular_repo_returns_repo_root() {
         let temp_dir = TempDir::new().expect("Failed to create temp dir");
         let repo_path = create_test_git_repo(&temp_dir).await;
-        let expected = std::fs::canonicalize(&repo_path).unwrap().to_path_buf();
+        let expected = std::fs::canonicalize(&repo_path).unwrap();
 
         assert_eq!(
             resolve_root_git_project_for_trust(&repo_path),
@@ -822,10 +810,7 @@ mod tests {
         );
         let nested = repo_path.join("sub/dir");
         std::fs::create_dir_all(&nested).unwrap();
-        assert_eq!(
-            resolve_root_git_project_for_trust(&nested),
-            Some(expected.clone())
-        );
+        assert_eq!(resolve_root_git_project_for_trust(&nested), Some(expected));
     }
 
     #[tokio::test]

codex-rs/core/src/internal_storage.rs

@@ -0,0 +1,68 @@
+use anyhow::Context;
+use serde::Deserialize;
+use serde::Serialize;
+use std::path::Path;
+use std::path::PathBuf;
+
+pub(crate) const INTERNAL_STORAGE_FILE: &str = "internal_storage.json";
+
+#[derive(Debug, Default, Clone, Serialize, Deserialize)]
+pub struct InternalStorage {
+    #[serde(skip)]
+    storage_path: PathBuf,
+    #[serde(default, alias = "gpt_5_high_model_prompt_seen")]
+    pub swiftfox_model_prompt_seen: bool,
+}
+
+// TODO(jif) generalise all the file writers and build proper async channel inserters.
+impl InternalStorage {
+    pub fn load(codex_home: &Path) -> Self {
+        let storage_path = codex_home.join(INTERNAL_STORAGE_FILE);
+
+        match std::fs::read_to_string(&storage_path) {
+            Ok(serialized) => match serde_json::from_str::<Self>(&serialized) {
+                Ok(mut storage) => {
+                    storage.storage_path = storage_path;
+                    storage
+                }
+                Err(error) => {
+                    tracing::warn!("failed to parse internal storage: {error:?}");
+                    Self::empty(storage_path)
+                }
+            },
+            Err(error) => {
+                tracing::warn!("failed to read internal storage: {error:?}");
+                Self::empty(storage_path)
+            }
+        }
+    }
+
+    fn empty(storage_path: PathBuf) -> Self {
+        Self {
+            storage_path,
+            ..Default::default()
+        }
+    }
+
+    pub async fn persist(&self) -> anyhow::Result<()> {
+        let serialized = serde_json::to_string_pretty(self)?;
+
+        if let Some(parent) = self.storage_path.parent() {
+            tokio::fs::create_dir_all(parent).await.with_context(|| {
+                format!(
+                    "failed to create internal storage directory at {}",
+                    parent.display()
+                )
+            })?;
+        }
+
+        tokio::fs::write(&self.storage_path, serialized)
+            .await
+            .with_context(|| {
+                format!(
+                    "failed to persist internal storage at {}",
+                    self.storage_path.display()
+                )
+            })
+    }
+}

codex-rs/core/src/lib.rs

@@ -6,14 +6,17 @@
 #![deny(clippy::print_stdout, clippy::print_stderr)]
 
 mod apply_patch;
-mod bash;
+pub mod auth;
+pub mod bash;
 mod chat_completions;
 mod client;
 mod client_common;
 pub mod codex;
 mod codex_conversation;
+pub mod token_data;
 pub use codex_conversation::CodexConversation;
 pub mod config;
+pub mod config_edit;
 pub mod config_profile;
 pub mod config_types;
 mod conversation_history;
@@ -25,21 +28,32 @@ mod exec_command;
 pub mod exec_env;
 mod flags;
 pub mod git_info;
+pub mod internal_storage;
 mod is_safe_command;
 pub mod landlock;
 mod mcp_connection_manager;
 mod mcp_tool_call;
 mod message_history;
 mod model_provider_info;
+mod normalize_command;
 pub mod parse_command;
+mod truncate;
+mod unified_exec;
+mod user_instructions;
 pub use model_provider_info::BUILT_IN_OSS_MODEL_PROVIDER_ID;
 pub use model_provider_info::ModelProviderInfo;
 pub use model_provider_info::WireApi;
 pub use model_provider_info::built_in_model_providers;
 pub use model_provider_info::create_oss_provider_with_base_url;
 mod conversation_manager;
+mod event_mapping;
+pub use codex_protocol::protocol::InitialHistory;
 pub use conversation_manager::ConversationManager;
 pub use conversation_manager::NewConversation;
+// Re-export common auth types for workspace consumers
+pub use auth::AuthManager;
+pub use auth::CodexAuth;
+pub mod default_client;
 pub mod model_family;
 mod openai_model_info;
 mod openai_tools;
@@ -53,9 +67,17 @@ pub mod spawn;
 pub mod terminal;
 mod tool_apply_patch;
 pub mod turn_diff_tracker;
-pub mod user_agent;
+pub use rollout::ARCHIVED_SESSIONS_SUBDIR;
+pub use rollout::RolloutRecorder;
+pub use rollout::SESSIONS_SUBDIR;
+pub use rollout::SessionMeta;
+pub use rollout::find_conversation_path_by_id_str;
+pub use rollout::list::ConversationItem;
+pub use rollout::list::ConversationsPage;
+pub use rollout::list::Cursor;
 mod user_notification;
 pub mod util;
+
 pub use apply_patch::CODEX_APPLY_PATCH_ARG1;
 pub use safety::get_platform_sandbox;
 // Re-export the protocol types from the standalone `codex-protocol` crate so existing
@@ -64,3 +86,14 @@ pub use codex_protocol::protocol;
 // Re-export protocol config enums to ensure call sites can use the same types
 // as those in the protocol crate when constructing protocol messages.
 pub use codex_protocol::config_types as protocol_config_types;
+
+pub use client::ModelClient;
+pub use client_common::Prompt;
+pub use client_common::ResponseEvent;
+pub use client_common::ResponseStream;
+pub use codex_protocol::models::ContentItem;
+pub use codex_protocol::models::LocalShellAction;
+pub use codex_protocol::models::LocalShellExecAction;
+pub use codex_protocol::models::LocalShellStatus;
+pub use codex_protocol::models::ReasoningItemContent;
+pub use codex_protocol::models::ResponseItem;

codex-rs/core/src/mcp_connection_manager.rs

@@ -9,6 +9,7 @@
 use std::collections::HashMap;
 use std::collections::HashSet;
 use std::ffi::OsString;
+use std::sync::Arc;
 use std::time::Duration;
 
 use anyhow::Context;
@@ -36,8 +37,8 @@ use crate::config_types::McpServerConfig;
 const MCP_TOOL_NAME_DELIMITER: &str = "__";
 const MAX_TOOL_NAME_LENGTH: usize = 64;
 
-/// Timeout for the `tools/list` request.
-const LIST_TOOLS_TIMEOUT: Duration = Duration::from_secs(10);
+/// Default timeout for initializing MCP server & initially listing tools.
+const DEFAULT_STARTUP_TIMEOUT: Duration = Duration::from_secs(10);
 
 /// Map that holds a startup error for every MCP server that could **not** be
 /// spawned successfully.
@@ -81,6 +82,11 @@ struct ToolInfo {
     tool: Tool,
 }
 
+struct ManagedClient {
+    client: Arc<McpClient>,
+    startup_timeout: Duration,
+}
+
 /// A thin wrapper around a set of running [`McpClient`] instances.
 #[derive(Default)]
 pub(crate) struct McpConnectionManager {
@@ -88,7 +94,7 @@ pub(crate) struct McpConnectionManager {
     ///
     /// The server name originates from the keys of the `mcp_servers` map in
     /// the user configuration.
-    clients: HashMap<String, std::sync::Arc<McpClient>>,
+    clients: HashMap<String, ManagedClient>,
 
     /// Fully qualified tool name -> tool instance.
     tools: HashMap<String, ToolInfo>,
@@ -126,8 +132,15 @@ impl McpConnectionManager {
                 continue;
             }
 
+            let startup_timeout = cfg
+                .startup_timeout_ms
+                .map(Duration::from_millis)
+                .unwrap_or(DEFAULT_STARTUP_TIMEOUT);
+
             join_set.spawn(async move {
-                let McpServerConfig { command, args, env } = cfg;
+                let McpServerConfig {
+                    command, args, env, ..
+                } = cfg;
                 let client_res = McpClient::new_stdio_client(
                     command.into(),
                     args.into_iter().map(OsString::from).collect(),
@@ -150,16 +163,23 @@ impl McpConnectionManager {
                                 name: "codex-mcp-client".to_owned(),
                                 version: env!("CARGO_PKG_VERSION").to_owned(),
                                 title: Some("Codex".into()),
+                                // This field is used by Codex when it is an MCP
+                                // server: it should not be used when Codex is
+                                // an MCP client.
+                                user_agent: None,
                             },
                             protocol_version: mcp_types::MCP_SCHEMA_VERSION.to_owned(),
                         };
                         let initialize_notification_params = None;
-                        let timeout = Some(Duration::from_secs(10));
                         match client
-                            .initialize(params, initialize_notification_params, timeout)
+                            .initialize(
+                                params,
+                                initialize_notification_params,
+                                Some(startup_timeout),
+                            )
                             .await
                         {
-                            Ok(_response) => (server_name, Ok(client)),
+                            Ok(_response) => (server_name, Ok((client, startup_timeout))),
                             Err(e) => (server_name, Err(e)),
                         }
                     }
@@ -168,15 +188,26 @@ impl McpConnectionManager {
             });
         }
 
-        let mut clients: HashMap<String, std::sync::Arc<McpClient>> =
-            HashMap::with_capacity(join_set.len());
+        let mut clients: HashMap<String, ManagedClient> = HashMap::with_capacity(join_set.len());
 
         while let Some(res) = join_set.join_next().await {
-            let (server_name, client_res) = res?; // JoinError propagation
+            let (server_name, client_res) = match res {
+                Ok((server_name, client_res)) => (server_name, client_res),
+                Err(e) => {
+                    warn!("Task panic when starting MCP server: {e:#}");
+                    continue;
+                }
+            };
 
             match client_res {
-                Ok(client) => {
-                    clients.insert(server_name, std::sync::Arc::new(client));
+                Ok((client, startup_timeout)) => {
+                    clients.insert(
+                        server_name,
+                        ManagedClient {
+                            client: Arc::new(client),
+                            startup_timeout,
+                        },
+                    );
                 }
                 Err(e) => {
                     errors.insert(server_name, e);
@@ -184,7 +215,13 @@ impl McpConnectionManager {
             }
         }
 
-        let all_tools = list_all_tools(&clients).await?;
+        let all_tools = match list_all_tools(&clients).await {
+            Ok(tools) => tools,
+            Err(e) => {
+                warn!("Failed to list tools from some MCP servers: {e:#}");
+                Vec::new()
+            }
+        };
 
         let tools = qualify_tools(all_tools);
 
@@ -212,6 +249,7 @@ impl McpConnectionManager {
             .clients
             .get(server)
             .ok_or_else(|| anyhow!("unknown MCP server '{server}'"))?
+            .client
             .clone();
 
         client
@@ -229,21 +267,18 @@ impl McpConnectionManager {
 
 /// Query every server for its available tools and return a single map that
 /// contains **all** tools. Each key is the fully-qualified name for the tool.
-async fn list_all_tools(
-    clients: &HashMap<String, std::sync::Arc<McpClient>>,
-) -> Result<Vec<ToolInfo>> {
+async fn list_all_tools(clients: &HashMap<String, ManagedClient>) -> Result<Vec<ToolInfo>> {
     let mut join_set = JoinSet::new();
 
     // Spawn one task per server so we can query them concurrently. This
     // keeps the overall latency roughly at the slowest server instead of
     // the cumulative latency.
-    for (server_name, client) in clients {
+    for (server_name, managed_client) in clients {
         let server_name_cloned = server_name.clone();
-        let client_clone = client.clone();
+        let client_clone = managed_client.client.clone();
+        let startup_timeout = managed_client.startup_timeout;
         join_set.spawn(async move {
-            let res = client_clone
-                .list_tools(None, Some(LIST_TOOLS_TIMEOUT))
-                .await;
+            let res = client_clone.list_tools(None, Some(startup_timeout)).await;
             (server_name_cloned, res)
         });
     }
@@ -251,8 +286,19 @@ async fn list_all_tools(
     let mut aggregated: Vec<ToolInfo> = Vec::with_capacity(join_set.len());
 
     while let Some(join_res) = join_set.join_next().await {
-        let (server_name, list_result) = join_res?;
-        let list_result = list_result?;
+        let (server_name, list_result) = if let Ok(result) = join_res {
+            result
+        } else {
+            warn!("Task panic when listing tools for MCP server: {join_res:#?}");
+            continue;
+        };
+
+        let list_result = if let Ok(result) = list_result {
+            result
+        } else {
+            warn!("Failed to list tools for MCP server '{server_name}': {list_result:#?}");
+            continue;
+        };
 
         for tool in list_result.tools {
             let tool_info = ToolInfo {

codex-rs/core/src/message_history.rs

@@ -5,7 +5,7 @@
 //! JSON-Lines tooling. Each record has the following schema:
 //!
 //! ````text
-//! {"session_id":"<uuid>","ts":<unix_seconds>,"text":"<message>"}
+//! {"conversation_id":"<uuid>","ts":<unix_seconds>,"text":"<message>"}
 //! ````
 //!
 //! To minimise the chance of interleaved writes when multiple processes are
@@ -22,14 +22,15 @@ use std::path::PathBuf;
 
 use serde::Deserialize;
 use serde::Serialize;
+
 use std::time::Duration;
 use tokio::fs;
 use tokio::io::AsyncReadExt;
-use uuid::Uuid;
 
 use crate::config::Config;
 use crate::config_types::HistoryPersistence;
 
+use codex_protocol::mcp_protocol::ConversationId;
 #[cfg(unix)]
 use std::os::unix::fs::OpenOptionsExt;
 #[cfg(unix)]
@@ -54,10 +55,14 @@ fn history_filepath(config: &Config) -> PathBuf {
     path
 }
 
-/// Append a `text` entry associated with `session_id` to the history file. Uses
+/// Append a `text` entry associated with `conversation_id` to the history file. Uses
 /// advisory file locking to ensure that concurrent writes do not interleave,
 /// which entails a small amount of blocking I/O internally.
-pub(crate) async fn append_entry(text: &str, session_id: &Uuid, config: &Config) -> Result<()> {
+pub(crate) async fn append_entry(
+    text: &str,
+    conversation_id: &ConversationId,
+    config: &Config,
+) -> Result<()> {
     match config.history.persistence {
         HistoryPersistence::SaveAll => {
             // Save everything: proceed.
@@ -84,7 +89,7 @@ pub(crate) async fn append_entry(text: &str, session_id: &Uuid, config: &Config)
 
     // Construct the JSON line first so we can write it in a single syscall.
     let entry = HistoryEntry {
-        session_id: session_id.to_string(),
+        session_id: conversation_id.to_string(),
         ts,
         text: text.to_string(),
     };
@@ -105,47 +110,34 @@ pub(crate) async fn append_entry(text: &str, session_id: &Uuid, config: &Config)
     // Ensure permissions.
     ensure_owner_only_permissions(&history_file).await?;
 
-    // Lock file.
-    acquire_exclusive_lock_with_retry(&history_file).await?;
-
-    // We use sync I/O with spawn_blocking() because we are using a
-    // [`std::fs::File`] instead of a [`tokio::fs::File`] to leverage an
-    // advisory file locking API that is not available in the async API.
+    // Perform a blocking write under an advisory write lock using std::fs.
     tokio::task::spawn_blocking(move || -> Result<()> {
-        history_file.write_all(line.as_bytes())?;
-        history_file.flush()?;
-        Ok(())
+        // Retry a few times to avoid indefinite blocking when contended.
+        for _ in 0..MAX_RETRIES {
+            match history_file.try_lock() {
+                Ok(()) => {
+                    // While holding the exclusive lock, write the full line.
+                    history_file.write_all(line.as_bytes())?;
+                    history_file.flush()?;
+                    return Ok(());
+                }
+                Err(std::fs::TryLockError::WouldBlock) => {
+                    std::thread::sleep(RETRY_SLEEP);
+                }
+                Err(e) => return Err(e.into()),
+            }
+        }
+
+        Err(std::io::Error::new(
+            std::io::ErrorKind::WouldBlock,
+            "could not acquire exclusive lock on history file after multiple attempts",
+        ))
     })
     .await??;
 
     Ok(())
 }
 
-/// Attempt to acquire an exclusive advisory lock on `file`, retrying up to 10
-/// times if the lock is currently held by another process. This prevents a
-/// potential indefinite wait while still giving other writers some time to
-/// finish their operation.
-async fn acquire_exclusive_lock_with_retry(file: &File) -> Result<()> {
-    use tokio::time::sleep;
-
-    for _ in 0..MAX_RETRIES {
-        match file.try_lock() {
-            Ok(()) => return Ok(()),
-            Err(e) => match e {
-                std::fs::TryLockError::WouldBlock => {
-                    sleep(RETRY_SLEEP).await;
-                }
-                other => return Err(other.into()),
-            },
-        }
-    }
-
-    Err(std::io::Error::new(
-        std::io::ErrorKind::WouldBlock,
-        "could not acquire exclusive lock on history file after multiple attempts",
-    ))
-}
-
 /// Asynchronously fetch the history file's *identifier* (inode on Unix) and
 /// the current number of entries by counting newline characters.
 pub(crate) async fn history_metadata(config: &Config) -> (u64, usize) {
@@ -221,29 +213,42 @@ pub(crate) fn lookup(log_id: u64, offset: usize, config: &Config) -> Option<Hist
         return None;
     }
 
-    // Open & lock file for reading.
-    if let Err(e) = acquire_shared_lock_with_retry(&file) {
-        tracing::warn!(error = %e, "failed to acquire shared lock on history file");
-        return None;
-    }
+    // Open & lock file for reading using a shared lock.
+    // Retry a few times to avoid indefinite blocking.
+    for _ in 0..MAX_RETRIES {
+        let lock_result = file.try_lock_shared();
 
-    let reader = BufReader::new(&file);
-    for (idx, line_res) in reader.lines().enumerate() {
-        let line = match line_res {
-            Ok(l) => l,
-            Err(e) => {
-                tracing::warn!(error = %e, "failed to read line from history file");
+        match lock_result {
+            Ok(()) => {
+                let reader = BufReader::new(&file);
+                for (idx, line_res) in reader.lines().enumerate() {
+                    let line = match line_res {
+                        Ok(l) => l,
+                        Err(e) => {
+                            tracing::warn!(error = %e, "failed to read line from history file");
+                            return None;
+                        }
+                    };
+
+                    if idx == offset {
+                        match serde_json::from_str::<HistoryEntry>(&line) {
+                            Ok(entry) => return Some(entry),
+                            Err(e) => {
+                                tracing::warn!(error = %e, "failed to parse history entry");
+                                return None;
+                            }
+                        }
+                    }
+                }
+                // Not found at requested offset.
                 return None;
             }
-        };
-
-        if idx == offset {
-            match serde_json::from_str::<HistoryEntry>(&line) {
-                Ok(entry) => return Some(entry),
-                Err(e) => {
-                    tracing::warn!(error = %e, "failed to parse history entry");
-                    return None;
-                }
+            Err(std::fs::TryLockError::WouldBlock) => {
+                std::thread::sleep(RETRY_SLEEP);
+            }
+            Err(e) => {
+                tracing::warn!(error = %e, "failed to acquire shared lock on history file");
+                return None;
             }
         }
     }
@@ -258,26 +263,6 @@ pub(crate) fn lookup(log_id: u64, offset: usize, config: &Config) -> Option<Hist
     None
 }
 
-#[cfg(unix)]
-fn acquire_shared_lock_with_retry(file: &File) -> Result<()> {
-    for _ in 0..MAX_RETRIES {
-        match file.try_lock_shared() {
-            Ok(()) => return Ok(()),
-            Err(e) => match e {
-                std::fs::TryLockError::WouldBlock => {
-                    std::thread::sleep(RETRY_SLEEP);
-                }
-                other => return Err(other.into()),
-            },
-        }
-    }
-
-    Err(std::io::Error::new(
-        std::io::ErrorKind::WouldBlock,
-        "could not acquire shared lock on history file after multiple attempts",
-    ))
-}
-
 /// On Unix systems ensure the file permissions are `0o600` (rw-------). If the
 /// permissions cannot be changed the error is propagated to the caller.
 #[cfg(unix)]

codex-rs/core/src/model_family.rs

@@ -1,5 +1,11 @@
+use crate::config_types::ReasoningSummaryFormat;
 use crate::tool_apply_patch::ApplyPatchToolType;
 
+/// The `instructions` field in the payload sent to a model should always start
+/// with this content.
+const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
+const SWIFTFOX_INSTRUCTIONS: &str = include_str!("../swiftfox_prompt.md");
+
 /// A model family is a group of models that share certain characteristics.
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
 pub struct ModelFamily {
@@ -20,6 +26,9 @@ pub struct ModelFamily {
     // `summary` is optional).
     pub supports_reasoning_summaries: bool,
 
+    // Define if we need a special handling of reasoning summary
+    pub reasoning_summary_format: ReasoningSummaryFormat,
+
     // This should be set to true when the model expects a tool named
     // "local_shell" to be provided. Its contract must be understood natively by
     // the model such that its description can be omitted.
@@ -29,6 +38,9 @@ pub struct ModelFamily {
     /// Present if the model performs better when `apply_patch` is provided as
     /// a tool call instead of just a bash command
     pub apply_patch_tool_type: Option<ApplyPatchToolType>,
+
+    // Instructions to use for querying the model
+    pub base_instructions: String,
 }
 
 macro_rules! model_family {
@@ -41,8 +53,10 @@ macro_rules! model_family {
             family: $family.to_string(),
             needs_special_apply_patch_instructions: false,
             supports_reasoning_summaries: false,
+            reasoning_summary_format: ReasoningSummaryFormat::None,
             uses_local_shell_tool: false,
             apply_patch_tool_type: None,
+            base_instructions: BASE_INSTRUCTIONS.to_string(),
         };
         // apply overrides
         $(
@@ -52,21 +66,6 @@ macro_rules! model_family {
     }};
 }
 
-macro_rules! simple_model_family {
-    (
-        $slug:expr, $family:expr
-    ) => {{
-        Some(ModelFamily {
-            slug: $slug.to_string(),
-            family: $family.to_string(),
-            needs_special_apply_patch_instructions: false,
-            supports_reasoning_summaries: false,
-            uses_local_shell_tool: false,
-            apply_patch_tool_type: None,
-        })
-    }};
-}
-
 /// Returns a `ModelFamily` for the given model slug, or `None` if the slug
 /// does not match any known model family.
 pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
@@ -74,40 +73,59 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
         model_family!(
             slug, "o3",
             supports_reasoning_summaries: true,
+            needs_special_apply_patch_instructions: true,
         )
     } else if slug.starts_with("o4-mini") {
         model_family!(
             slug, "o4-mini",
             supports_reasoning_summaries: true,
+            needs_special_apply_patch_instructions: true,
         )
     } else if slug.starts_with("codex-mini-latest") {
         model_family!(
             slug, "codex-mini-latest",
             supports_reasoning_summaries: true,
             uses_local_shell_tool: true,
-        )
-    } else if slug.starts_with("codex-") {
-        model_family!(
-            slug, slug,
-            supports_reasoning_summaries: true,
+            needs_special_apply_patch_instructions: true,
         )
     } else if slug.starts_with("gpt-4.1") {
         model_family!(
             slug, "gpt-4.1",
             needs_special_apply_patch_instructions: true,
         )
-    } else if slug.starts_with("gpt-oss") {
+    } else if slug.starts_with("gpt-oss") || slug.starts_with("openai/gpt-oss") {
         model_family!(slug, "gpt-oss", apply_patch_tool_type: Some(ApplyPatchToolType::Function))
     } else if slug.starts_with("gpt-4o") {
-        simple_model_family!(slug, "gpt-4o")
+        model_family!(slug, "gpt-4o", needs_special_apply_patch_instructions: true)
     } else if slug.starts_with("gpt-3.5") {
-        simple_model_family!(slug, "gpt-3.5")
+        model_family!(slug, "gpt-3.5", needs_special_apply_patch_instructions: true)
+    } else if slug.starts_with("codex-") || slug.starts_with("swiftfox") {
+        model_family!(
+            slug, slug,
+            supports_reasoning_summaries: true,
+            reasoning_summary_format: ReasoningSummaryFormat::Experimental,
+            base_instructions: SWIFTFOX_INSTRUCTIONS.to_string(),
+        )
     } else if slug.starts_with("gpt-5") {
         model_family!(
             slug, "gpt-5",
             supports_reasoning_summaries: true,
+            needs_special_apply_patch_instructions: true,
         )
     } else {
         None
     }
 }
+
+pub fn derive_default_model_family(model: &str) -> ModelFamily {
+    ModelFamily {
+        slug: model.to_string(),
+        family: model.to_string(),
+        needs_special_apply_patch_instructions: false,
+        supports_reasoning_summaries: false,
+        reasoning_summary_format: ReasoningSummaryFormat::None,
+        uses_local_shell_tool: false,
+        apply_patch_tool_type: None,
+        base_instructions: BASE_INSTRUCTIONS.to_string(),
+    }
+}

codex-rs/core/src/model_provider_info.rs

@@ -5,8 +5,8 @@
 //!   2. User-defined entries inside `~/.codex/config.toml` under the `model_providers`
 //!      key. These override or extend the defaults at runtime.
 
-use codex_login::AuthMode;
-use codex_login::CodexAuth;
+use crate::CodexAuth;
+use codex_protocol::mcp_protocol::AuthMode;
 use serde::Deserialize;
 use serde::Serialize;
 use std::collections::HashMap;
@@ -80,7 +80,10 @@ pub struct ModelProviderInfo {
     /// the connection as lost.
     pub stream_idle_timeout_ms: Option<u64>,
 
-    /// Whether this provider requires some form of standard authentication (API key, ChatGPT token).
+    /// Does this provider require an OpenAI API Key or ChatGPT login token? If true,
+    /// user is presented with login screen on first run, and login preference and token/key
+    /// are stored in auth.json. If false (which is the default), login screen is skipped,
+    /// and API key (if needed) comes from the "env_key" environment variable.
     #[serde(default)]
     pub requires_openai_auth: bool,
 }
@@ -159,6 +162,21 @@ impl ModelProviderInfo {
         }
     }
 
+    pub(crate) fn is_azure_responses_endpoint(&self) -> bool {
+        if self.wire_api != WireApi::Responses {
+            return false;
+        }
+
+        if self.name.eq_ignore_ascii_case("azure") {
+            return true;
+        }
+
+        self.base_url
+            .as_ref()
+            .map(|base| matches_azure_responses_base_url(base))
+            .unwrap_or(false)
+    }
+
     /// Apply provider-specific HTTP headers (both static and environment-based)
     /// onto an existing `reqwest::RequestBuilder` and return the updated
     /// builder.
@@ -326,6 +344,18 @@ pub fn create_oss_provider_with_base_url(base_url: &str) -> ModelProviderInfo {
     }
 }
 
+fn matches_azure_responses_base_url(base_url: &str) -> bool {
+    let base = base_url.to_ascii_lowercase();
+    const AZURE_MARKERS: [&str; 5] = [
+        "openai.azure.",
+        "cognitiveservices.azure.",
+        "aoai.azure.",
+        "azure-api.",
+        "azurefd.",
+    ];
+    AZURE_MARKERS.iter().any(|marker| base.contains(marker))
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -416,4 +446,69 @@ env_http_headers = { "X-Example-Env-Header" = "EXAMPLE_ENV_VAR" }
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
         assert_eq!(expected_provider, provider);
     }
+
+    #[test]
+    fn detects_azure_responses_base_urls() {
+        fn provider_for(base_url: &str) -> ModelProviderInfo {
+            ModelProviderInfo {
+                name: "test".into(),
+                base_url: Some(base_url.into()),
+                env_key: None,
+                env_key_instructions: None,
+                wire_api: WireApi::Responses,
+                query_params: None,
+                http_headers: None,
+                env_http_headers: None,
+                request_max_retries: None,
+                stream_max_retries: None,
+                stream_idle_timeout_ms: None,
+                requires_openai_auth: false,
+            }
+        }
+
+        let positive_cases = [
+            "https://foo.openai.azure.com/openai",
+            "https://foo.openai.azure.us/openai/deployments/bar",
+            "https://foo.cognitiveservices.azure.cn/openai",
+            "https://foo.aoai.azure.com/openai",
+            "https://foo.openai.azure-api.net/openai",
+            "https://foo.z01.azurefd.net/",
+        ];
+        for base_url in positive_cases {
+            let provider = provider_for(base_url);
+            assert!(
+                provider.is_azure_responses_endpoint(),
+                "expected {base_url} to be detected as Azure"
+            );
+        }
+
+        let named_provider = ModelProviderInfo {
+            name: "Azure".into(),
+            base_url: Some("https://example.com".into()),
+            env_key: None,
+            env_key_instructions: None,
+            wire_api: WireApi::Responses,
+            query_params: None,
+            http_headers: None,
+            env_http_headers: None,
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
+            requires_openai_auth: false,
+        };
+        assert!(named_provider.is_azure_responses_endpoint());
+
+        let negative_cases = [
+            "https://api.openai.com/v1",
+            "https://example.com/openai",
+            "https://myproxy.azurewebsites.net/openai",
+        ];
+        for base_url in negative_cases {
+            let provider = provider_for(base_url);
+            assert!(
+                !provider.is_azure_responses_endpoint(),
+                "expected {base_url} not to be detected as Azure"
+            );
+        }
+    }
 }

codex-rs/core/src/normalize_command.rs

@@ -0,0 +1,438 @@
+use crate::bash::extract_words_from_command_node;
+use crate::bash::find_first_command_node;
+use crate::bash::remainder_start_after_wrapper_operator;
+use crate::bash::try_parse_bash;
+use crate::bash::try_parse_word_only_commands_sequence;
+use std::path::Path;
+use std::path::PathBuf;
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct NormalizedCommand {
+    pub command: Vec<String>,
+    pub cwd: PathBuf,
+    pub command_for_display: Vec<String>,
+}
+
+/// Normalize a shell command for execution and display.
+///
+/// For classic wrappers like `bash -lc "cd repo && git status"` returns:
+/// - command: ["bash", "-lc", "git status"]
+/// - cwd: original cwd joined with "repo"
+/// - command_for_display:
+///   - If the remainder is a single exec-able command, return it tokenized
+///     (e.g., ["rg", "--files"]).
+///   - Otherwise, wrap the script as ["bash", "-lc", "<script>"] so that
+///     [`crate::parse_command::parse_command`] (which currently checks for
+///     bash -lc) can handle it. Even if the original used zsh, we standardize
+///     to bash here to match [`crate::parse_command::parse_command`]’s
+///     expectations.
+///
+/// Returns None when no normalization is needed.
+pub(crate) fn try_normalize_command(command: &[String], cwd: &Path) -> Option<NormalizedCommand> {
+    let invocation = parse_shell_script_from_shell_invocation(command)?;
+    let shell_script = invocation.command.clone();
+
+    let tree = try_parse_bash(&shell_script)?;
+
+    let first_cmd = find_first_command_node(&tree)?;
+    let words = extract_words_from_command_node(first_cmd, &shell_script)?;
+    if !is_command_cd_to_directory(&words) {
+        return None;
+    }
+
+    let idx = remainder_start_after_wrapper_operator(first_cmd, &shell_script)?;
+    let remainder = shell_script[idx..].to_string();
+
+    // For parse_command consumption: if the remainder is a single exec-able
+    // command, return it as argv tokens; otherwise, wrap it in ["bash","-lc",..].
+    let command_for_display = try_parse_bash(&remainder)
+        .and_then(|tree| try_parse_word_only_commands_sequence(&tree, &remainder))
+        .and_then(|mut cmds| match cmds.len() {
+            1 => cmds.pop(),
+            _ => None,
+        })
+        .unwrap_or_else(|| vec!["bash".to_string(), "-lc".to_string(), remainder.clone()]);
+
+    // Compute new cwd only after confirming the wrapper/operator shape
+    let dir = &words[1];
+    let new_cwd = if Path::new(dir).is_absolute() {
+        PathBuf::from(dir)
+    } else {
+        // TODO(mbolin): Is there anything to worry about if `dir` is a
+        // symlink or contains `..` components?
+        cwd.join(dir)
+    };
+
+    Some(NormalizedCommand {
+        command: vec![invocation.executable, invocation.flag, remainder],
+        cwd: new_cwd,
+        command_for_display,
+    })
+}
+
+/// This is similar to [`crate::shell::strip_bash_lc`] and should potentially
+/// be unified with it.
+fn parse_shell_script_from_shell_invocation(command: &[String]) -> Option<ShellScriptInvocation> {
+    // Allowed shells
+    fn is_allowed_exe(s: &str) -> bool {
+        matches!(s, "bash" | "/bin/bash" | "zsh" | "/bin/zsh")
+    }
+    match command {
+        // exactly three items: <exe> <flag> <command>
+        [exe, flag, cmd] if is_allowed_exe(exe) && (flag == "-lc" || flag == "-c") => {
+            Some(ShellScriptInvocation {
+                executable: exe.clone(),
+                flag: flag.clone(),
+                command: cmd.clone(),
+            })
+        }
+        // exactly four items: <exe> -l -c <command>
+        [exe, flag1, flag2, cmd] if is_allowed_exe(exe) && flag1 == "-l" && flag2 == "-c" => {
+            Some(ShellScriptInvocation {
+                executable: exe.clone(),
+                flag: "-lc".to_string(),
+                command: cmd.clone(),
+            })
+        }
+        _ => None,
+    }
+}
+
+fn is_command_cd_to_directory(command: &[String]) -> bool {
+    matches!(command, [first, _dir] if first == "cd")
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct ShellScriptInvocation {
+    executable: String,
+    /// Single arg, so `-l -c` must be collapsed to `-lc`.
+    /// Must be written so `command` follows `flag`, so must
+    /// be `-lc` and not `-cl`.
+    flag: String,
+    command: String,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::NormalizedCommand;
+    use super::try_normalize_command;
+    use pretty_assertions::assert_eq;
+    use std::path::PathBuf;
+
+    #[test]
+    fn cd_prefix_in_bash_script_is_hidden() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo && echo hi".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "echo hi".into()],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["echo".into(), "hi".into()],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_shell_var_not_matched() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd $SOME_DIR && ls".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm, None,
+            "Not a classic wrapper (cd arg is a variable), so no normalization."
+        );
+    }
+
+    #[test]
+    fn cd_prefix_with_quoted_path_is_hidden() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "  cd 'foo bar' && ls".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "ls".into()],
+                cwd: PathBuf::from("/baz/foo bar"),
+                command_for_display: vec!["ls".into()],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_prefix_with_additional_cd_is_preserved() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo && cd bar && ls".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "cd bar && ls".into()],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["bash".into(), "-lc".into(), "cd bar && ls".into(),],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_prefix_with_or_connector_is_preserved() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd and_and || echo \"couldn't find the dir for &&\"".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        // Not a classic wrapper (uses ||), so no normalization.
+        assert_eq!(norm, None);
+    }
+
+    #[test]
+    fn cd_prefix_preserves_operators_between_remaining_commands() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo && ls && git status".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "ls && git status".into()],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["bash".into(), "-lc".into(), "ls && git status".into(),],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_prefix_preserves_pipelines() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo && ls | rg foo".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "ls | rg foo".into()],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["bash".into(), "-lc".into(), "ls | rg foo".into(),],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_prefix_preserves_sequence_operator() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo && ls; git status".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "ls; git status".into()],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["bash".into(), "-lc".into(), "ls; git status".into(),],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_prefix_with_semicolon_is_hidden() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo; ls".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "ls".into()],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["ls".into()],
+            })
+        );
+    }
+
+    #[test]
+    fn pushd_prefix_is_not_normalized() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "pushd foo && ls".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        // We do not normalize pushd because omitting pushd from the execution
+        // would have different semantics than the original command.
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(norm, None);
+    }
+
+    #[test]
+    fn supports_zsh_and_bin_bash_and_split_flags() {
+        // zsh -lc
+        let zsh = vec!["zsh".into(), "-lc".into(), "cd foo && ls".into()];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&zsh, &cwd).unwrap();
+        assert_eq!(
+            norm.command,
+            ["zsh", "-lc", "ls"]
+                .iter()
+                .map(|s| s.to_string())
+                .collect::<Vec<_>>()
+        );
+        assert_eq!(norm.cwd, PathBuf::from("/baz/foo"));
+        assert_eq!(norm.command_for_display, vec!["ls".to_string()]);
+
+        // /bin/bash -l -c <cmd>
+        let bash_split = vec![
+            "/bin/bash".into(),
+            "-l".into(),
+            "-c".into(),
+            "cd foo && ls".into(),
+        ];
+        let norm2 = try_normalize_command(&bash_split, &cwd).unwrap();
+        assert_eq!(
+            norm2.command,
+            ["/bin/bash", "-lc", "ls"]
+                .iter()
+                .map(|s| s.to_string())
+                .collect::<Vec<_>>()
+        );
+        assert_eq!(norm2.cwd, PathBuf::from("/baz/foo"));
+        assert_eq!(norm2.command_for_display, vec!["ls".to_string()]);
+    }
+
+    #[test]
+    fn supports_dash_c_flag() {
+        let command = vec!["bash".into(), "-c".into(), "cd foo && ls".into()];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd).unwrap();
+        assert_eq!(
+            norm.command,
+            ["bash", "-c", "ls"]
+                .iter()
+                .map(|s| s.to_string())
+                .collect::<Vec<_>>()
+        );
+        assert_eq!(norm.cwd, PathBuf::from("/baz/foo"));
+        assert_eq!(norm.command_for_display, vec!["ls".to_string()]);
+    }
+
+    #[test]
+    fn rejects_pipe_operator_immediately_after_cd() {
+        let command = vec!["bash".into(), "-lc".into(), "cd foo | rg bar".into()];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(norm, None);
+    }
+
+    #[test]
+    fn rejects_unknown_shell_and_bad_flag_order() {
+        // Unknown shell
+        let unknown = vec!["sh".into(), "-lc".into(), "cd foo && ls".into()];
+        let cwd = PathBuf::from("/baz");
+        assert_eq!(try_normalize_command(&unknown, &cwd), None);
+
+        // Bad flag order -cl (unsupported)
+        let bad_flag = vec!["bash".into(), "-cl".into(), "cd foo && ls".into()];
+        assert_eq!(try_normalize_command(&bad_flag, &cwd), None);
+    }
+
+    #[test]
+    fn absolute_directory_sets_absolute_cwd() {
+        let command = vec!["bash".into(), "-lc".into(), "cd /tmp && ls".into()];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd).unwrap();
+        assert_eq!(norm.cwd, PathBuf::from("/tmp"));
+        assert_eq!(
+            norm.command,
+            ["bash", "-lc", "ls"]
+                .iter()
+                .map(|s| s.to_string())
+                .collect::<Vec<_>>()
+        );
+        assert_eq!(norm.command_for_display, vec!["ls".to_string()]);
+    }
+
+    #[test]
+    fn non_shell_command_is_returned_unmodified() {
+        let command = vec!["rg".to_string(), "--files".to_string()];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        // Not a shell wrapper, so no normalization.
+        assert_eq!(norm, None);
+    }
+
+    #[test]
+    fn cd_prefix_with_or_operator_is_preserved() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd missing && ls || echo 'fallback'".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec!["bash".into(), "-lc".into(), "ls || echo 'fallback'".into()],
+                cwd: PathBuf::from("/baz/missing"),
+                command_for_display: vec![
+                    "bash".into(),
+                    "-lc".into(),
+                    "ls || echo 'fallback'".into(),
+                ],
+            })
+        );
+    }
+
+    #[test]
+    fn cd_prefix_with_ampersands_in_string_is_hidden() {
+        let command = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "cd foo && echo \"checking && markers\"".to_string(),
+        ];
+        let cwd = PathBuf::from("/baz");
+        let norm = try_normalize_command(&command, &cwd);
+        assert_eq!(
+            norm,
+            Some(NormalizedCommand {
+                command: vec![
+                    "bash".into(),
+                    "-lc".into(),
+                    "echo \"checking && markers\"".into(),
+                ],
+                cwd: PathBuf::from("/baz/foo"),
+                command_for_display: vec!["echo".into(), "checking && markers".into(),],
+            })
+        );
+    }
+}

codex-rs/core/src/openai_model_info.rs

@@ -12,6 +12,19 @@ pub(crate) struct ModelInfo {
 
     /// Maximum number of output tokens that can be generated for the model.
     pub(crate) max_output_tokens: u64,
+
+    /// Token threshold where we should automatically compact conversation history.
+    pub(crate) auto_compact_token_limit: Option<i64>,
+}
+
+impl ModelInfo {
+    const fn new(context_window: u64, max_output_tokens: u64) -> Self {
+        Self {
+            context_window,
+            max_output_tokens,
+            auto_compact_token_limit: None,
+        }
+    }
 }
 
 pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
@@ -20,73 +33,37 @@ pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
         // OSS models have a 128k shared token pool.
         // Arbitrarily splitting it: 3/4 input context, 1/4 output.
         // https://openai.com/index/gpt-oss-model-card/
-        "gpt-oss-20b" => Some(ModelInfo {
-            context_window: 96_000,
-            max_output_tokens: 32_000,
-        }),
-        "gpt-oss-120b" => Some(ModelInfo {
-            context_window: 96_000,
-            max_output_tokens: 32_000,
-        }),
+        "gpt-oss-20b" => Some(ModelInfo::new(96_000, 32_000)),
+        "gpt-oss-120b" => Some(ModelInfo::new(96_000, 32_000)),
         // https://platform.openai.com/docs/models/o3
-        "o3" => Some(ModelInfo {
-            context_window: 200_000,
-            max_output_tokens: 100_000,
-        }),
+        "o3" => Some(ModelInfo::new(200_000, 100_000)),
 
         // https://platform.openai.com/docs/models/o4-mini
-        "o4-mini" => Some(ModelInfo {
-            context_window: 200_000,
-            max_output_tokens: 100_000,
-        }),
+        "o4-mini" => Some(ModelInfo::new(200_000, 100_000)),
 
         // https://platform.openai.com/docs/models/codex-mini-latest
-        "codex-mini-latest" => Some(ModelInfo {
-            context_window: 200_000,
-            max_output_tokens: 100_000,
-        }),
+        "codex-mini-latest" => Some(ModelInfo::new(200_000, 100_000)),
 
         // As of Jun 25, 2025, gpt-4.1 defaults to gpt-4.1-2025-04-14.
         // https://platform.openai.com/docs/models/gpt-4.1
-        "gpt-4.1" | "gpt-4.1-2025-04-14" => Some(ModelInfo {
-            context_window: 1_047_576,
-            max_output_tokens: 32_768,
-        }),
+        "gpt-4.1" | "gpt-4.1-2025-04-14" => Some(ModelInfo::new(1_047_576, 32_768)),
 
         // As of Jun 25, 2025, gpt-4o defaults to gpt-4o-2024-08-06.
         // https://platform.openai.com/docs/models/gpt-4o
-        "gpt-4o" | "gpt-4o-2024-08-06" => Some(ModelInfo {
-            context_window: 128_000,
-            max_output_tokens: 16_384,
-        }),
+        "gpt-4o" | "gpt-4o-2024-08-06" => Some(ModelInfo::new(128_000, 16_384)),
 
         // https://platform.openai.com/docs/models/gpt-4o?snapshot=gpt-4o-2024-05-13
-        "gpt-4o-2024-05-13" => Some(ModelInfo {
-            context_window: 128_000,
-            max_output_tokens: 4_096,
-        }),
+        "gpt-4o-2024-05-13" => Some(ModelInfo::new(128_000, 4_096)),
 
         // https://platform.openai.com/docs/models/gpt-4o?snapshot=gpt-4o-2024-11-20
-        "gpt-4o-2024-11-20" => Some(ModelInfo {
-            context_window: 128_000,
-            max_output_tokens: 16_384,
-        }),
+        "gpt-4o-2024-11-20" => Some(ModelInfo::new(128_000, 16_384)),
 
         // https://platform.openai.com/docs/models/gpt-3.5-turbo
-        "gpt-3.5-turbo" => Some(ModelInfo {
-            context_window: 16_385,
-            max_output_tokens: 4_096,
-        }),
+        "gpt-3.5-turbo" => Some(ModelInfo::new(16_385, 4_096)),
 
-        "gpt-5" => Some(ModelInfo {
-            context_window: 400_000,
-            max_output_tokens: 128_000,
-        }),
+        _ if slug.starts_with("gpt-5") => Some(ModelInfo::new(272_000, 128_000)),
 
-        _ if slug.starts_with("codex-") => Some(ModelInfo {
-            context_window: 400_000,
-            max_output_tokens: 128_000,
-        }),
+        _ if slug.starts_with("codex-") => Some(ModelInfo::new(272_000, 128_000)),
 
         _ => None,
     }

codex-rs/core/src/openai_tools.rs

@@ -49,7 +49,7 @@ pub(crate) enum OpenAiTool {
     LocalShell {},
     // TODO: Understand why we get an error on web_search although the API docs say it's supported.
     // https://platform.openai.com/docs/guides/tools-web-search?api-mode=responses#:~:text=%7B%20type%3A%20%22web_search%22%20%7D%2C
-    #[serde(rename = "web_search_preview")]
+    #[serde(rename = "web_search")]
     WebSearch {},
     #[serde(rename = "custom")]
     Freeform(FreeformTool),
@@ -70,6 +70,7 @@ pub(crate) struct ToolsConfig {
     pub apply_patch_tool_type: Option<ApplyPatchToolType>,
     pub web_search_request: bool,
     pub include_view_image_tool: bool,
+    pub experimental_unified_exec_tool: bool,
 }
 
 pub(crate) struct ToolsConfigParams<'a> {
@@ -81,6 +82,7 @@ pub(crate) struct ToolsConfigParams<'a> {
     pub(crate) include_web_search_request: bool,
     pub(crate) use_streamable_shell_tool: bool,
     pub(crate) include_view_image_tool: bool,
+    pub(crate) experimental_unified_exec_tool: bool,
 }
 
 impl ToolsConfig {
@@ -94,6 +96,7 @@ impl ToolsConfig {
             include_web_search_request,
             use_streamable_shell_tool,
             include_view_image_tool,
+            experimental_unified_exec_tool,
         } = params;
         let mut shell_type = if *use_streamable_shell_tool {
             ConfigShellToolType::StreamableShell
@@ -126,6 +129,7 @@ impl ToolsConfig {
             apply_patch_tool_type,
             web_search_request: *include_web_search_request,
             include_view_image_tool: *include_view_image_tool,
+            experimental_unified_exec_tool: *experimental_unified_exec_tool,
         }
     }
 }
@@ -200,6 +204,53 @@ fn create_shell_tool() -> OpenAiTool {
     })
 }
 
+fn create_unified_exec_tool() -> OpenAiTool {
+    let mut properties = BTreeMap::new();
+    properties.insert(
+        "input".to_string(),
+        JsonSchema::Array {
+            items: Box::new(JsonSchema::String { description: None }),
+            description: Some(
+                "When no session_id is provided, treat the array as the command and arguments \
+                 to launch. When session_id is set, concatenate the strings (in order) and write \
+                 them to the session's stdin."
+                    .to_string(),
+            ),
+        },
+    );
+    properties.insert(
+        "session_id".to_string(),
+        JsonSchema::String {
+            description: Some(
+                "Identifier for an existing interactive session. If omitted, a new command \
+                 is spawned."
+                    .to_string(),
+            ),
+        },
+    );
+    properties.insert(
+        "timeout_ms".to_string(),
+        JsonSchema::Number {
+            description: Some(
+                "Maximum time in milliseconds to wait for output after writing the input."
+                    .to_string(),
+            ),
+        },
+    );
+
+    OpenAiTool::Function(ResponsesApiTool {
+        name: "unified_exec".to_string(),
+        description:
+            "Runs a command in a PTY. Provide a session_id to reuse an existing interactive session.".to_string(),
+        strict: false,
+        parameters: JsonSchema::Object {
+            properties,
+            required: Some(vec!["input".to_string()]),
+            additional_properties: Some(false),
+        },
+    })
+}
+
 fn create_shell_tool_for_sandbox(sandbox_policy: &SandboxPolicy) -> OpenAiTool {
     let mut properties = BTreeMap::new();
     properties.insert(
@@ -237,59 +288,9 @@ fn create_shell_tool_for_sandbox(sandbox_policy: &SandboxPolicy) -> OpenAiTool {
     );
     }
 
-    let description = match sandbox_policy {
-        SandboxPolicy::WorkspaceWrite {
-            network_access,
-            ..
-        } => {
-            format!(
-                r#"
-The shell tool is used to execute shell commands.
-- When invoking the shell tool, your call will be running in a landlock sandbox, and some shell commands will require escalated privileges:
-  - Types of actions that require escalated privileges:
-    - Reading files outside the current directory
-    - Writing files outside the current directory, and protected folders like .git or .env{}
-  - Examples of commands that require escalated privileges:
-    - git commit
-    - npm install or pnpm install
-    - cargo build
-    - cargo test
-- When invoking a command that will require escalated privileges:
-  - Provide the with_escalated_permissions parameter with the boolean value true
-  - Include a short, 1 sentence explanation for why we need to run with_escalated_permissions in the justification parameter."#,
-                if !network_access {
-                    "\n  - Commands that require network access\n"
-                } else {
-                    ""
-                }
-            )
-        }
-        SandboxPolicy::DangerFullAccess => {
-            "Runs a shell command and returns its output.".to_string()
-        }
-        SandboxPolicy::ReadOnly => {
-            r#"
-The shell tool is used to execute shell commands.
-- When invoking the shell tool, your call will be running in a landlock sandbox, and some shell commands (including apply_patch) will require escalated permissions:
-  - Types of actions that require escalated privileges:
-    - Reading files outside the current directory
-    - Writing files
-    - Applying patches
-  - Examples of commands that require escalated privileges:
-    - apply_patch
-    - git commit
-    - npm install or pnpm install
-    - cargo build
-    - cargo test
-- When invoking a command that will require escalated privileges:
-  - Provide the with_escalated_permissions parameter with the boolean value true
-  - Include a short, 1 sentence explanation for why we need to run with_escalated_permissions in the justification parameter"#.to_string()
-        }
-    };
-
     OpenAiTool::Function(ResponsesApiTool {
         name: "shell".to_string(),
-        description,
+        description: "Runs a shell command and returns its output.".to_string(),
         strict: false,
         parameters: JsonSchema::Object {
             properties,
@@ -532,23 +533,27 @@ pub(crate) fn get_openai_tools(
 ) -> Vec<OpenAiTool> {
     let mut tools: Vec<OpenAiTool> = Vec::new();
 
-    match &config.shell_type {
-        ConfigShellToolType::DefaultShell => {
-            tools.push(create_shell_tool());
-        }
-        ConfigShellToolType::ShellWithRequest { sandbox_policy } => {
-            tools.push(create_shell_tool_for_sandbox(sandbox_policy));
-        }
-        ConfigShellToolType::LocalShell => {
-            tools.push(OpenAiTool::LocalShell {});
-        }
-        ConfigShellToolType::StreamableShell => {
-            tools.push(OpenAiTool::Function(
-                crate::exec_command::create_exec_command_tool_for_responses_api(),
-            ));
-            tools.push(OpenAiTool::Function(
-                crate::exec_command::create_write_stdin_tool_for_responses_api(),
-            ));
+    if config.experimental_unified_exec_tool {
+        tools.push(create_unified_exec_tool());
+    } else {
+        match &config.shell_type {
+            ConfigShellToolType::DefaultShell => {
+                tools.push(create_shell_tool());
+            }
+            ConfigShellToolType::ShellWithRequest { sandbox_policy } => {
+                tools.push(create_shell_tool_for_sandbox(sandbox_policy));
+            }
+            ConfigShellToolType::LocalShell => {
+                tools.push(OpenAiTool::LocalShell {});
+            }
+            ConfigShellToolType::StreamableShell => {
+                tools.push(OpenAiTool::Function(
+                    crate::exec_command::create_exec_command_tool_for_responses_api(),
+                ));
+                tools.push(OpenAiTool::Function(
+                    crate::exec_command::create_write_stdin_tool_for_responses_api(),
+                ));
+            }
         }
     }
 
@@ -575,10 +580,8 @@ pub(crate) fn get_openai_tools(
     if config.include_view_image_tool {
         tools.push(create_view_image_tool());
     }
-
     if let Some(mcp_tools) = mcp_tools {
         // Ensure deterministic ordering to maximize prompt cache hits.
-        // HashMap iteration order is non-deterministic, so sort by fully-qualified tool name.
         let mut entries: Vec<(String, mcp_types::Tool)> = mcp_tools.into_iter().collect();
         entries.sort_by(|a, b| a.0.cmp(&b.0));
 
@@ -640,12 +643,13 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
         let tools = get_openai_tools(&config, Some(HashMap::new()));
 
         assert_eq_tool_names(
             &tools,
-            &["local_shell", "update_plan", "web_search", "view_image"],
+            &["unified_exec", "update_plan", "web_search", "view_image"],
         );
     }
 
@@ -661,12 +665,13 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
         let tools = get_openai_tools(&config, Some(HashMap::new()));
 
         assert_eq_tool_names(
             &tools,
-            &["shell", "update_plan", "web_search", "view_image"],
+            &["unified_exec", "update_plan", "web_search", "view_image"],
         );
     }
 
@@ -682,6 +687,7 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
         let tools = get_openai_tools(
             &config,
@@ -724,7 +730,7 @@ mod tests {
         assert_eq_tool_names(
             &tools,
             &[
-                "shell",
+                "unified_exec",
                 "web_search",
                 "view_image",
                 "test_server/do_something_cool",
@@ -787,6 +793,7 @@ mod tests {
             include_web_search_request: false,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
 
         // Intentionally construct a map with keys that would sort alphabetically.
@@ -839,11 +846,11 @@ mod tests {
         ]);
 
         let tools = get_openai_tools(&config, Some(tools_map));
-        // Expect shell first, followed by MCP tools sorted by fully-qualified name.
+        // Expect unified_exec first, followed by MCP tools sorted by fully-qualified name.
         assert_eq_tool_names(
             &tools,
             &[
-                "shell",
+                "unified_exec",
                 "view_image",
                 "test_server/cool",
                 "test_server/do",
@@ -864,6 +871,7 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
 
         let tools = get_openai_tools(
@@ -891,7 +899,7 @@ mod tests {
 
         assert_eq_tool_names(
             &tools,
-            &["shell", "web_search", "view_image", "dash/search"],
+            &["unified_exec", "web_search", "view_image", "dash/search"],
         );
 
         assert_eq!(
@@ -926,6 +934,7 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
 
         let tools = get_openai_tools(
@@ -951,7 +960,7 @@ mod tests {
 
         assert_eq_tool_names(
             &tools,
-            &["shell", "web_search", "view_image", "dash/paginate"],
+            &["unified_exec", "web_search", "view_image", "dash/paginate"],
         );
         assert_eq!(
             tools[3],
@@ -983,6 +992,7 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
 
         let tools = get_openai_tools(
@@ -1006,7 +1016,10 @@ mod tests {
             )])),
         );
 
-        assert_eq_tool_names(&tools, &["shell", "web_search", "view_image", "dash/tags"]);
+        assert_eq_tool_names(
+            &tools,
+            &["unified_exec", "web_search", "view_image", "dash/tags"],
+        );
         assert_eq!(
             tools[3],
             OpenAiTool::Function(ResponsesApiTool {
@@ -1040,6 +1053,7 @@ mod tests {
             include_web_search_request: true,
             use_streamable_shell_tool: false,
             include_view_image_tool: true,
+            experimental_unified_exec_tool: true,
         });
 
         let tools = get_openai_tools(
@@ -1063,7 +1077,10 @@ mod tests {
             )])),
         );
 
-        assert_eq_tool_names(&tools, &["shell", "web_search", "view_image", "dash/value"]);
+        assert_eq_tool_names(
+            &tools,
+            &["unified_exec", "web_search", "view_image", "dash/value"],
+        );
         assert_eq!(
             tools[3],
             OpenAiTool::Function(ResponsesApiTool {
@@ -1081,4 +1098,54 @@ mod tests {
             })
         );
     }
+
+    #[test]
+    fn test_shell_tool_for_sandbox_workspace_write() {
+        let sandbox_policy = SandboxPolicy::WorkspaceWrite {
+            writable_roots: vec!["workspace".into()],
+            network_access: false,
+            exclude_tmpdir_env_var: false,
+            exclude_slash_tmp: false,
+        };
+        let tool = super::create_shell_tool_for_sandbox(&sandbox_policy);
+        let OpenAiTool::Function(ResponsesApiTool {
+            description, name, ..
+        }) = &tool
+        else {
+            panic!("expected function tool");
+        };
+        assert_eq!(name, "shell");
+
+        let expected = "Runs a shell command and returns its output.";
+        assert_eq!(description, expected);
+    }
+
+    #[test]
+    fn test_shell_tool_for_sandbox_readonly() {
+        let tool = super::create_shell_tool_for_sandbox(&SandboxPolicy::ReadOnly);
+        let OpenAiTool::Function(ResponsesApiTool {
+            description, name, ..
+        }) = &tool
+        else {
+            panic!("expected function tool");
+        };
+        assert_eq!(name, "shell");
+
+        let expected = "Runs a shell command and returns its output.";
+        assert_eq!(description, expected);
+    }
+
+    #[test]
+    fn test_shell_tool_for_sandbox_danger_full_access() {
+        let tool = super::create_shell_tool_for_sandbox(&SandboxPolicy::DangerFullAccess);
+        let OpenAiTool::Function(ResponsesApiTool {
+            description, name, ..
+        }) = &tool
+        else {
+            panic!("expected function tool");
+        };
+        assert_eq!(name, "shell");
+
+        assert_eq!(description, "Runs a shell command and returns its output.");
+    }
 }

codex-rs/core/src/parse_command.rs

@@ -235,6 +235,28 @@ mod tests {
         );
     }
 
+    #[test]
+    fn bash_cd_then_bar_is_same_as_bar() {
+        // Ensure a leading `cd` inside bash -lc is dropped when followed by another command.
+        assert_parsed(
+            &shlex_split_safe("bash -lc 'cd foo && bar'"),
+            vec![ParsedCommand::Unknown {
+                cmd: "bar".to_string(),
+            }],
+        );
+    }
+
+    #[test]
+    fn bash_cd_then_cat_is_read() {
+        assert_parsed(
+            &shlex_split_safe("bash -lc 'cd foo && cat foo.txt'"),
+            vec![ParsedCommand::Read {
+                cmd: "cat foo.txt".to_string(),
+                name: "foo.txt".to_string(),
+            }],
+        );
+    }
+
     #[test]
     fn supports_ls_with_pipe() {
         let inner = "ls -la | sed -n '1,120p'";
@@ -846,7 +868,7 @@ pub fn parse_command_impl(command: &[String]) -> Vec<ParsedCommand> {
     let parts = if contains_connectors(&normalized) {
         split_on_connectors(&normalized)
     } else {
-        vec![normalized.clone()]
+        vec![normalized]
     };
 
     // Preserve left-to-right execution order for all commands, including bash -c/-lc
@@ -1149,6 +1171,10 @@ fn parse_bash_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
             .collect();
         if commands.len() > 1 {
             commands.retain(|pc| !matches!(pc, ParsedCommand::Unknown { cmd } if cmd == "true"));
+            // Apply the same simplifications used for non-bash parsing, e.g., drop leading `cd`.
+            while let Some(next) = simplify_once(&commands) {
+                commands = next;
+            }
         }
         if commands.len() == 1 {
             // If we reduced to a single command, attribute the full original script
@@ -1175,10 +1201,7 @@ fn parse_bash_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
                                     name,
                                 }
                             } else {
-                                ParsedCommand::Read {
-                                    cmd: cmd.clone(),
-                                    name,
-                                }
+                                ParsedCommand::Read { cmd, name }
                             }
                         } else {
                             ParsedCommand::Read {
@@ -1189,10 +1212,7 @@ fn parse_bash_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
                     }
                     ParsedCommand::ListFiles { path, cmd, .. } => {
                         if had_connectors {
-                            ParsedCommand::ListFiles {
-                                cmd: cmd.clone(),
-                                path,
-                            }
+                            ParsedCommand::ListFiles { cmd, path }
                         } else {
                             ParsedCommand::ListFiles {
                                 cmd: shlex_join(&script_tokens),
@@ -1204,11 +1224,7 @@ fn parse_bash_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
                         query, path, cmd, ..
                     } => {
                         if had_connectors {
-                            ParsedCommand::Search {
-                                cmd: cmd.clone(),
-                                query,
-                                path,
-                            }
+                            ParsedCommand::Search { cmd, query, path }
                         } else {
                             ParsedCommand::Search {
                                 cmd: shlex_join(&script_tokens),

codex-rs/core/src/project_doc.rs

@@ -115,7 +115,7 @@ pub fn discover_project_doc_paths(config: &Config) -> std::io::Result<Vec<PathBu
     // Build chain from cwd upwards and detect git root.
     let mut chain: Vec<PathBuf> = vec![dir.clone()];
     let mut git_root: Option<PathBuf> = None;
-    let mut cursor = dir.clone();
+    let mut cursor = dir;
     while let Some(parent) = cursor.parent() {
         let git_marker = cursor.join(".git");
         let git_exists = match std::fs::metadata(&git_marker) {

codex-rs/core/src/prompt_for_compact_command.md

@@ -1,21 +0,0 @@
-You are a summarization assistant. A conversation follows between a user and a coding-focused AI (Codex). Your task is to generate a clear summary capturing:
-
-• High-level objective or problem being solved  
-• Key instructions or design decisions given by the user  
-• Main code actions or behaviors from the AI  
-• Important variables, functions, modules, or outputs discussed  
-• Any unresolved questions or next steps
-
-Produce the summary in a structured format like:
-
-**Objective:** …
-
-**User instructions:** … (bulleted)
-
-**AI actions / code behavior:** … (bulleted)
-
-**Important entities:** … (e.g. function names, variables, files)
-
-**Open issues / next steps:** … (if any)
-
-**Summary (concise):** (one or two sentences)

codex-rs/core/src/rollout.rs

@@ -1,368 +0,0 @@
-//! Persist Codex session rollouts (.jsonl) so sessions can be replayed or inspected later.
-
-use std::fs::File;
-use std::fs::{self};
-use std::io::Error as IoError;
-use std::path::Path;
-
-use serde::Deserialize;
-use serde::Serialize;
-use serde_json::Value;
-use time::OffsetDateTime;
-use time::format_description::FormatItem;
-use time::macros::format_description;
-use tokio::io::AsyncWriteExt;
-use tokio::sync::mpsc::Sender;
-use tokio::sync::mpsc::{self};
-use tokio::sync::oneshot;
-use tracing::info;
-use tracing::warn;
-use uuid::Uuid;
-
-use crate::config::Config;
-use crate::git_info::GitInfo;
-use crate::git_info::collect_git_info;
-use codex_protocol::models::ResponseItem;
-
-const SESSIONS_SUBDIR: &str = "sessions";
-
-#[derive(Serialize, Deserialize, Clone, Default)]
-pub struct SessionMeta {
-    pub id: Uuid,
-    pub timestamp: String,
-    pub instructions: Option<String>,
-}
-
-#[derive(Serialize)]
-struct SessionMetaWithGit {
-    #[serde(flatten)]
-    meta: SessionMeta,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    git: Option<GitInfo>,
-}
-
-#[derive(Serialize, Deserialize, Default, Clone)]
-pub struct SessionStateSnapshot {}
-
-#[derive(Serialize, Deserialize, Default, Clone)]
-pub struct SavedSession {
-    pub session: SessionMeta,
-    #[serde(default)]
-    pub items: Vec<ResponseItem>,
-    #[serde(default)]
-    pub state: SessionStateSnapshot,
-    pub session_id: Uuid,
-}
-
-/// Records all [`ResponseItem`]s for a session and flushes them to disk after
-/// every update.
-///
-/// Rollouts are recorded as JSONL and can be inspected with tools such as:
-///
-/// ```ignore
-/// $ jq -C . ~/.codex/sessions/rollout-2025-05-07T17-24-21-5973b6c0-94b8-487b-a530-2aeb6098ae0e.jsonl
-/// $ fx ~/.codex/sessions/rollout-2025-05-07T17-24-21-5973b6c0-94b8-487b-a530-2aeb6098ae0e.jsonl
-/// ```
-#[derive(Clone)]
-pub(crate) struct RolloutRecorder {
-    tx: Sender<RolloutCmd>,
-}
-
-enum RolloutCmd {
-    AddItems(Vec<ResponseItem>),
-    UpdateState(SessionStateSnapshot),
-    Shutdown { ack: oneshot::Sender<()> },
-}
-
-impl RolloutRecorder {
-    /// Attempt to create a new [`RolloutRecorder`]. If the sessions directory
-    /// cannot be created or the rollout file cannot be opened we return the
-    /// error so the caller can decide whether to disable persistence.
-    pub async fn new(
-        config: &Config,
-        uuid: Uuid,
-        instructions: Option<String>,
-    ) -> std::io::Result<Self> {
-        let LogFileInfo {
-            file,
-            session_id,
-            timestamp,
-        } = create_log_file(config, uuid)?;
-
-        let timestamp_format: &[FormatItem] = format_description!(
-            "[year]-[month]-[day]T[hour]:[minute]:[second].[subsecond digits:3]Z"
-        );
-        let timestamp = timestamp
-            .format(timestamp_format)
-            .map_err(|e| IoError::other(format!("failed to format timestamp: {e}")))?;
-
-        // Clone the cwd for the spawned task to collect git info asynchronously
-        let cwd = config.cwd.clone();
-
-        // A reasonably-sized bounded channel. If the buffer fills up the send
-        // future will yield, which is fine – we only need to ensure we do not
-        // perform *blocking* I/O on the caller's thread.
-        let (tx, rx) = mpsc::channel::<RolloutCmd>(256);
-
-        // Spawn a Tokio task that owns the file handle and performs async
-        // writes. Using `tokio::fs::File` keeps everything on the async I/O
-        // driver instead of blocking the runtime.
-        tokio::task::spawn(rollout_writer(
-            tokio::fs::File::from_std(file),
-            rx,
-            Some(SessionMeta {
-                timestamp,
-                id: session_id,
-                instructions,
-            }),
-            cwd,
-        ));
-
-        Ok(Self { tx })
-    }
-
-    pub(crate) async fn record_items(&self, items: &[ResponseItem]) -> std::io::Result<()> {
-        let mut filtered = Vec::new();
-        for item in items {
-            match item {
-                // Note that function calls may look a bit strange if they are
-                // "fully qualified MCP tool calls," so we could consider
-                // reformatting them in that case.
-                ResponseItem::Message { .. }
-                | ResponseItem::LocalShellCall { .. }
-                | ResponseItem::FunctionCall { .. }
-                | ResponseItem::FunctionCallOutput { .. }
-                | ResponseItem::CustomToolCall { .. }
-                | ResponseItem::CustomToolCallOutput { .. }
-                | ResponseItem::Reasoning { .. } => filtered.push(item.clone()),
-                ResponseItem::WebSearchCall { .. } | ResponseItem::Other => {
-                    // These should never be serialized.
-                    continue;
-                }
-            }
-        }
-        if filtered.is_empty() {
-            return Ok(());
-        }
-        self.tx
-            .send(RolloutCmd::AddItems(filtered))
-            .await
-            .map_err(|e| IoError::other(format!("failed to queue rollout items: {e}")))
-    }
-
-    pub(crate) async fn record_state(&self, state: SessionStateSnapshot) -> std::io::Result<()> {
-        self.tx
-            .send(RolloutCmd::UpdateState(state))
-            .await
-            .map_err(|e| IoError::other(format!("failed to queue rollout state: {e}")))
-    }
-
-    pub async fn resume(
-        path: &Path,
-        cwd: std::path::PathBuf,
-    ) -> std::io::Result<(Self, SavedSession)> {
-        info!("Resuming rollout from {path:?}");
-        let text = tokio::fs::read_to_string(path).await?;
-        let mut lines = text.lines();
-        let meta_line = lines
-            .next()
-            .ok_or_else(|| IoError::other("empty session file"))?;
-        let session: SessionMeta = serde_json::from_str(meta_line)
-            .map_err(|e| IoError::other(format!("failed to parse session meta: {e}")))?;
-        let mut items = Vec::new();
-        let mut state = SessionStateSnapshot::default();
-
-        for line in lines {
-            if line.trim().is_empty() {
-                continue;
-            }
-            let v: Value = match serde_json::from_str(line) {
-                Ok(v) => v,
-                Err(_) => continue,
-            };
-            if v.get("record_type")
-                .and_then(|rt| rt.as_str())
-                .map(|s| s == "state")
-                .unwrap_or(false)
-            {
-                if let Ok(s) = serde_json::from_value::<SessionStateSnapshot>(v.clone()) {
-                    state = s
-                }
-                continue;
-            }
-            match serde_json::from_value::<ResponseItem>(v.clone()) {
-                Ok(item) => match item {
-                    ResponseItem::Message { .. }
-                    | ResponseItem::LocalShellCall { .. }
-                    | ResponseItem::FunctionCall { .. }
-                    | ResponseItem::FunctionCallOutput { .. }
-                    | ResponseItem::CustomToolCall { .. }
-                    | ResponseItem::CustomToolCallOutput { .. }
-                    | ResponseItem::Reasoning { .. } => items.push(item),
-                    ResponseItem::WebSearchCall { .. } | ResponseItem::Other => {}
-                },
-                Err(e) => {
-                    warn!("failed to parse item: {v:?}, error: {e}");
-                }
-            }
-        }
-
-        let saved = SavedSession {
-            session: session.clone(),
-            items: items.clone(),
-            state: state.clone(),
-            session_id: session.id,
-        };
-
-        let file = std::fs::OpenOptions::new()
-            .append(true)
-            .read(true)
-            .open(path)?;
-
-        let (tx, rx) = mpsc::channel::<RolloutCmd>(256);
-        tokio::task::spawn(rollout_writer(
-            tokio::fs::File::from_std(file),
-            rx,
-            None,
-            cwd,
-        ));
-        info!("Resumed rollout successfully from {path:?}");
-        Ok((Self { tx }, saved))
-    }
-
-    pub async fn shutdown(&self) -> std::io::Result<()> {
-        let (tx_done, rx_done) = oneshot::channel();
-        match self.tx.send(RolloutCmd::Shutdown { ack: tx_done }).await {
-            Ok(_) => rx_done
-                .await
-                .map_err(|e| IoError::other(format!("failed waiting for rollout shutdown: {e}"))),
-            Err(e) => {
-                warn!("failed to send rollout shutdown command: {e}");
-                Err(IoError::other(format!(
-                    "failed to send rollout shutdown command: {e}"
-                )))
-            }
-        }
-    }
-}
-
-struct LogFileInfo {
-    /// Opened file handle to the rollout file.
-    file: File,
-
-    /// Session ID (also embedded in filename).
-    session_id: Uuid,
-
-    /// Timestamp for the start of the session.
-    timestamp: OffsetDateTime,
-}
-
-fn create_log_file(config: &Config, session_id: Uuid) -> std::io::Result<LogFileInfo> {
-    // Resolve ~/.codex/sessions/YYYY/MM/DD and create it if missing.
-    let timestamp = OffsetDateTime::now_local()
-        .map_err(|e| IoError::other(format!("failed to get local time: {e}")))?;
-    let mut dir = config.codex_home.clone();
-    dir.push(SESSIONS_SUBDIR);
-    dir.push(timestamp.year().to_string());
-    dir.push(format!("{:02}", u8::from(timestamp.month())));
-    dir.push(format!("{:02}", timestamp.day()));
-    fs::create_dir_all(&dir)?;
-
-    // Custom format for YYYY-MM-DDThh-mm-ss. Use `-` instead of `:` for
-    // compatibility with filesystems that do not allow colons in filenames.
-    let format: &[FormatItem] =
-        format_description!("[year]-[month]-[day]T[hour]-[minute]-[second]");
-    let date_str = timestamp
-        .format(format)
-        .map_err(|e| IoError::other(format!("failed to format timestamp: {e}")))?;
-
-    let filename = format!("rollout-{date_str}-{session_id}.jsonl");
-
-    let path = dir.join(filename);
-    let file = std::fs::OpenOptions::new()
-        .append(true)
-        .create(true)
-        .open(&path)?;
-
-    Ok(LogFileInfo {
-        file,
-        session_id,
-        timestamp,
-    })
-}
-
-async fn rollout_writer(
-    file: tokio::fs::File,
-    mut rx: mpsc::Receiver<RolloutCmd>,
-    mut meta: Option<SessionMeta>,
-    cwd: std::path::PathBuf,
-) -> std::io::Result<()> {
-    let mut writer = JsonlWriter { file };
-
-    // If we have a meta, collect git info asynchronously and write meta first
-    if let Some(session_meta) = meta.take() {
-        let git_info = collect_git_info(&cwd).await;
-        let session_meta_with_git = SessionMetaWithGit {
-            meta: session_meta,
-            git: git_info,
-        };
-
-        // Write the SessionMeta as the first item in the file
-        writer.write_line(&session_meta_with_git).await?;
-    }
-
-    // Process rollout commands
-    while let Some(cmd) = rx.recv().await {
-        match cmd {
-            RolloutCmd::AddItems(items) => {
-                for item in items {
-                    match item {
-                        ResponseItem::Message { .. }
-                        | ResponseItem::LocalShellCall { .. }
-                        | ResponseItem::FunctionCall { .. }
-                        | ResponseItem::FunctionCallOutput { .. }
-                        | ResponseItem::CustomToolCall { .. }
-                        | ResponseItem::CustomToolCallOutput { .. }
-                        | ResponseItem::Reasoning { .. } => {
-                            writer.write_line(&item).await?;
-                        }
-                        ResponseItem::WebSearchCall { .. } | ResponseItem::Other => {}
-                    }
-                }
-            }
-            RolloutCmd::UpdateState(state) => {
-                #[derive(Serialize)]
-                struct StateLine<'a> {
-                    record_type: &'static str,
-                    #[serde(flatten)]
-                    state: &'a SessionStateSnapshot,
-                }
-                writer
-                    .write_line(&StateLine {
-                        record_type: "state",
-                        state: &state,
-                    })
-                    .await?;
-            }
-            RolloutCmd::Shutdown { ack } => {
-                let _ = ack.send(());
-            }
-        }
-    }
-
-    Ok(())
-}
-
-struct JsonlWriter {
-    file: tokio::fs::File,
-}
-
-impl JsonlWriter {
-    async fn write_line(&mut self, item: &impl serde::Serialize) -> std::io::Result<()> {
-        let mut json = serde_json::to_string(item)?;
-        json.push('\n');
-        let _ = self.file.write_all(json.as_bytes()).await;
-        self.file.flush().await?;
-        Ok(())
-    }
-}

codex-rs/core/src/rollout/list.rs

@@ -0,0 +1,385 @@
+use std::cmp::Reverse;
+use std::io::{self};
+use std::path::Path;
+use std::path::PathBuf;
+
+use codex_file_search as file_search;
+use std::num::NonZero;
+use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
+use time::OffsetDateTime;
+use time::PrimitiveDateTime;
+use time::format_description::FormatItem;
+use time::macros::format_description;
+use uuid::Uuid;
+
+use super::SESSIONS_SUBDIR;
+use crate::protocol::EventMsg;
+use codex_protocol::protocol::RolloutItem;
+use codex_protocol::protocol::RolloutLine;
+
+/// Returned page of conversation summaries.
+#[derive(Debug, Default, PartialEq)]
+pub struct ConversationsPage {
+    /// Conversation summaries ordered newest first.
+    pub items: Vec<ConversationItem>,
+    /// Opaque pagination token to resume after the last item, or `None` if end.
+    pub next_cursor: Option<Cursor>,
+    /// Total number of files touched while scanning this request.
+    pub num_scanned_files: usize,
+    /// True if a hard scan cap was hit; consider resuming with `next_cursor`.
+    pub reached_scan_cap: bool,
+}
+
+/// Summary information for a conversation rollout file.
+#[derive(Debug, PartialEq)]
+pub struct ConversationItem {
+    /// Absolute path to the rollout file.
+    pub path: PathBuf,
+    /// First up to 5 JSONL records parsed as JSON (includes meta line).
+    pub head: Vec<serde_json::Value>,
+}
+
+/// Hard cap to bound worst‑case work per request.
+const MAX_SCAN_FILES: usize = 100;
+const HEAD_RECORD_LIMIT: usize = 10;
+
+/// Pagination cursor identifying a file by timestamp and UUID.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Cursor {
+    ts: OffsetDateTime,
+    id: Uuid,
+}
+
+impl Cursor {
+    fn new(ts: OffsetDateTime, id: Uuid) -> Self {
+        Self { ts, id }
+    }
+}
+
+impl serde::Serialize for Cursor {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        let ts_str = self
+            .ts
+            .format(&format_description!(
+                "[year]-[month]-[day]T[hour]-[minute]-[second]"
+            ))
+            .map_err(|e| serde::ser::Error::custom(format!("format error: {e}")))?;
+        serializer.serialize_str(&format!("{ts_str}|{}", self.id))
+    }
+}
+
+impl<'de> serde::Deserialize<'de> for Cursor {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let s = String::deserialize(deserializer)?;
+        parse_cursor(&s).ok_or_else(|| serde::de::Error::custom("invalid cursor"))
+    }
+}
+
+/// Retrieve recorded conversation file paths with token pagination. The returned `next_cursor`
+/// can be supplied on the next call to resume after the last returned item, resilient to
+/// concurrent new sessions being appended. Ordering is stable by timestamp desc, then UUID desc.
+pub(crate) async fn get_conversations(
+    codex_home: &Path,
+    page_size: usize,
+    cursor: Option<&Cursor>,
+) -> io::Result<ConversationsPage> {
+    let mut root = codex_home.to_path_buf();
+    root.push(SESSIONS_SUBDIR);
+
+    if !root.exists() {
+        return Ok(ConversationsPage {
+            items: Vec::new(),
+            next_cursor: None,
+            num_scanned_files: 0,
+            reached_scan_cap: false,
+        });
+    }
+
+    let anchor = cursor.cloned();
+
+    let result = traverse_directories_for_paths(root.clone(), page_size, anchor).await?;
+    Ok(result)
+}
+
+/// Load the full contents of a single conversation session file at `path`.
+/// Returns the entire file contents as a String.
+#[allow(dead_code)]
+pub(crate) async fn get_conversation(path: &Path) -> io::Result<String> {
+    tokio::fs::read_to_string(path).await
+}
+
+/// Load conversation file paths from disk using directory traversal.
+///
+/// Directory layout: `~/.codex/sessions/YYYY/MM/DD/rollout-YYYY-MM-DDThh-mm-ss-<uuid>.jsonl`
+/// Returned newest (latest) first.
+async fn traverse_directories_for_paths(
+    root: PathBuf,
+    page_size: usize,
+    anchor: Option<Cursor>,
+) -> io::Result<ConversationsPage> {
+    let mut items: Vec<ConversationItem> = Vec::with_capacity(page_size);
+    let mut scanned_files = 0usize;
+    let mut anchor_passed = anchor.is_none();
+    let (anchor_ts, anchor_id) = match anchor {
+        Some(c) => (c.ts, c.id),
+        None => (OffsetDateTime::UNIX_EPOCH, Uuid::nil()),
+    };
+
+    let year_dirs = collect_dirs_desc(&root, |s| s.parse::<u16>().ok()).await?;
+
+    'outer: for (_year, year_path) in year_dirs.iter() {
+        if scanned_files >= MAX_SCAN_FILES {
+            break;
+        }
+        let month_dirs = collect_dirs_desc(year_path, |s| s.parse::<u8>().ok()).await?;
+        for (_month, month_path) in month_dirs.iter() {
+            if scanned_files >= MAX_SCAN_FILES {
+                break 'outer;
+            }
+            let day_dirs = collect_dirs_desc(month_path, |s| s.parse::<u8>().ok()).await?;
+            for (_day, day_path) in day_dirs.iter() {
+                if scanned_files >= MAX_SCAN_FILES {
+                    break 'outer;
+                }
+                let mut day_files = collect_files(day_path, |name_str, path| {
+                    if !name_str.starts_with("rollout-") || !name_str.ends_with(".jsonl") {
+                        return None;
+                    }
+
+                    parse_timestamp_uuid_from_filename(name_str)
+                        .map(|(ts, id)| (ts, id, name_str.to_string(), path.to_path_buf()))
+                })
+                .await?;
+                // Stable ordering within the same second: (timestamp desc, uuid desc)
+                day_files.sort_by_key(|(ts, sid, _name_str, _path)| (Reverse(*ts), Reverse(*sid)));
+                for (ts, sid, _name_str, path) in day_files.into_iter() {
+                    scanned_files += 1;
+                    if scanned_files >= MAX_SCAN_FILES && items.len() >= page_size {
+                        break 'outer;
+                    }
+                    if !anchor_passed {
+                        if ts < anchor_ts || (ts == anchor_ts && sid < anchor_id) {
+                            anchor_passed = true;
+                        } else {
+                            continue;
+                        }
+                    }
+                    if items.len() == page_size {
+                        break 'outer;
+                    }
+                    // Read head and simultaneously detect message events within the same
+                    // first N JSONL records to avoid a second file read.
+                    let (head, saw_session_meta, saw_user_event) =
+                        read_head_and_flags(&path, HEAD_RECORD_LIMIT)
+                            .await
+                            .unwrap_or((Vec::new(), false, false));
+                    // Apply filters: must have session meta and at least one user message event
+                    if saw_session_meta && saw_user_event {
+                        items.push(ConversationItem { path, head });
+                    }
+                }
+            }
+        }
+    }
+
+    let next = build_next_cursor(&items);
+    Ok(ConversationsPage {
+        items,
+        next_cursor: next,
+        num_scanned_files: scanned_files,
+        reached_scan_cap: scanned_files >= MAX_SCAN_FILES,
+    })
+}
+
+/// Pagination cursor token format: "<file_ts>|<uuid>" where `file_ts` matches the
+/// filename timestamp portion (YYYY-MM-DDThh-mm-ss) used in rollout filenames.
+/// The cursor orders files by timestamp desc, then UUID desc.
+fn parse_cursor(token: &str) -> Option<Cursor> {
+    let (file_ts, uuid_str) = token.split_once('|')?;
+
+    let Ok(uuid) = Uuid::parse_str(uuid_str) else {
+        return None;
+    };
+
+    let format: &[FormatItem] =
+        format_description!("[year]-[month]-[day]T[hour]-[minute]-[second]");
+    let ts = PrimitiveDateTime::parse(file_ts, format).ok()?.assume_utc();
+
+    Some(Cursor::new(ts, uuid))
+}
+
+fn build_next_cursor(items: &[ConversationItem]) -> Option<Cursor> {
+    let last = items.last()?;
+    let file_name = last.path.file_name()?.to_string_lossy();
+    let (ts, id) = parse_timestamp_uuid_from_filename(&file_name)?;
+    Some(Cursor::new(ts, id))
+}
+
+/// Collects immediate subdirectories of `parent`, parses their (string) names with `parse`,
+/// and returns them sorted descending by the parsed key.
+async fn collect_dirs_desc<T, F>(parent: &Path, parse: F) -> io::Result<Vec<(T, PathBuf)>>
+where
+    T: Ord + Copy,
+    F: Fn(&str) -> Option<T>,
+{
+    let mut dir = tokio::fs::read_dir(parent).await?;
+    let mut vec: Vec<(T, PathBuf)> = Vec::new();
+    while let Some(entry) = dir.next_entry().await? {
+        if entry
+            .file_type()
+            .await
+            .map(|ft| ft.is_dir())
+            .unwrap_or(false)
+            && let Some(s) = entry.file_name().to_str()
+            && let Some(v) = parse(s)
+        {
+            vec.push((v, entry.path()));
+        }
+    }
+    vec.sort_by_key(|(v, _)| Reverse(*v));
+    Ok(vec)
+}
+
+/// Collects files in a directory and parses them with `parse`.
+async fn collect_files<T, F>(parent: &Path, parse: F) -> io::Result<Vec<T>>
+where
+    F: Fn(&str, &Path) -> Option<T>,
+{
+    let mut dir = tokio::fs::read_dir(parent).await?;
+    let mut collected: Vec<T> = Vec::new();
+    while let Some(entry) = dir.next_entry().await? {
+        if entry
+            .file_type()
+            .await
+            .map(|ft| ft.is_file())
+            .unwrap_or(false)
+            && let Some(s) = entry.file_name().to_str()
+            && let Some(v) = parse(s, &entry.path())
+        {
+            collected.push(v);
+        }
+    }
+    Ok(collected)
+}
+
+fn parse_timestamp_uuid_from_filename(name: &str) -> Option<(OffsetDateTime, Uuid)> {
+    // Expected: rollout-YYYY-MM-DDThh-mm-ss-<uuid>.jsonl
+    let core = name.strip_prefix("rollout-")?.strip_suffix(".jsonl")?;
+
+    // Scan from the right for a '-' such that the suffix parses as a UUID.
+    let (sep_idx, uuid) = core
+        .match_indices('-')
+        .rev()
+        .find_map(|(i, _)| Uuid::parse_str(&core[i + 1..]).ok().map(|u| (i, u)))?;
+
+    let ts_str = &core[..sep_idx];
+    let format: &[FormatItem] =
+        format_description!("[year]-[month]-[day]T[hour]-[minute]-[second]");
+    let ts = PrimitiveDateTime::parse(ts_str, format).ok()?.assume_utc();
+    Some((ts, uuid))
+}
+
+async fn read_head_and_flags(
+    path: &Path,
+    max_records: usize,
+) -> io::Result<(Vec<serde_json::Value>, bool, bool)> {
+    use tokio::io::AsyncBufReadExt;
+
+    let file = tokio::fs::File::open(path).await?;
+    let reader = tokio::io::BufReader::new(file);
+    let mut lines = reader.lines();
+    let mut head: Vec<serde_json::Value> = Vec::new();
+    let mut saw_session_meta = false;
+    let mut saw_user_event = false;
+
+    while head.len() < max_records {
+        let line_opt = lines.next_line().await?;
+        let Some(line) = line_opt else { break };
+        let trimmed = line.trim();
+        if trimmed.is_empty() {
+            continue;
+        }
+
+        let parsed: Result<RolloutLine, _> = serde_json::from_str(trimmed);
+        let Ok(rollout_line) = parsed else { continue };
+
+        match rollout_line.item {
+            RolloutItem::SessionMeta(session_meta_line) => {
+                if let Ok(val) = serde_json::to_value(session_meta_line) {
+                    head.push(val);
+                    saw_session_meta = true;
+                }
+            }
+            RolloutItem::ResponseItem(item) => {
+                if let Ok(val) = serde_json::to_value(item) {
+                    head.push(val);
+                }
+            }
+            RolloutItem::TurnContext(_) => {
+                // Not included in `head`; skip.
+            }
+            RolloutItem::Compacted(_) => {
+                // Not included in `head`; skip.
+            }
+            RolloutItem::EventMsg(ev) => {
+                if matches!(ev, EventMsg::UserMessage(_)) {
+                    saw_user_event = true;
+                }
+            }
+        }
+    }
+
+    Ok((head, saw_session_meta, saw_user_event))
+}
+
+/// Locate a recorded conversation rollout file by its UUID string using the existing
+/// paginated listing implementation. Returns `Ok(Some(path))` if found, `Ok(None)` if not present
+/// or the id is invalid.
+pub async fn find_conversation_path_by_id_str(
+    codex_home: &Path,
+    id_str: &str,
+) -> io::Result<Option<PathBuf>> {
+    // Validate UUID format early.
+    if Uuid::parse_str(id_str).is_err() {
+        return Ok(None);
+    }
+
+    let mut root = codex_home.to_path_buf();
+    root.push(SESSIONS_SUBDIR);
+    if !root.exists() {
+        return Ok(None);
+    }
+    // This is safe because we know the values are valid.
+    #[allow(clippy::unwrap_used)]
+    let limit = NonZero::new(1).unwrap();
+    // This is safe because we know the values are valid.
+    #[allow(clippy::unwrap_used)]
+    let threads = NonZero::new(2).unwrap();
+    let cancel = Arc::new(AtomicBool::new(false));
+    let exclude: Vec<String> = Vec::new();
+    let compute_indices = false;
+
+    let results = file_search::run(
+        id_str,
+        limit,
+        &root,
+        exclude,
+        threads,
+        cancel,
+        compute_indices,
+    )
+    .map_err(|e| io::Error::other(format!("file search failed: {e}")))?;
+
+    Ok(results
+        .matches
+        .into_iter()
+        .next()
+        .map(|m| root.join(m.path)))
+}

codex-rs/core/src/rollout/mod.rs

@@ -0,0 +1,16 @@
+//! Rollout module: persistence and discovery of session rollout files.
+
+pub const SESSIONS_SUBDIR: &str = "sessions";
+pub const ARCHIVED_SESSIONS_SUBDIR: &str = "archived_sessions";
+
+pub mod list;
+pub(crate) mod policy;
+pub mod recorder;
+
+pub use codex_protocol::protocol::SessionMeta;
+pub use list::find_conversation_path_by_id_str;
+pub use recorder::RolloutRecorder;
+pub use recorder::RolloutRecorderParams;
+
+#[cfg(test)]
+pub mod tests;

codex-rs/core/src/rollout/policy.rs

@@ -0,0 +1,74 @@
+use crate::protocol::EventMsg;
+use crate::protocol::RolloutItem;
+use codex_protocol::models::ResponseItem;
+
+/// Whether a rollout `item` should be persisted in rollout files.
+#[inline]
+pub(crate) fn is_persisted_response_item(item: &RolloutItem) -> bool {
+    match item {
+        RolloutItem::ResponseItem(item) => should_persist_response_item(item),
+        RolloutItem::EventMsg(ev) => should_persist_event_msg(ev),
+        // Persist Codex executive markers so we can analyze flows (e.g., compaction, API turns).
+        RolloutItem::Compacted(_) | RolloutItem::TurnContext(_) | RolloutItem::SessionMeta(_) => {
+            true
+        }
+    }
+}
+
+/// Whether a `ResponseItem` should be persisted in rollout files.
+#[inline]
+pub(crate) fn should_persist_response_item(item: &ResponseItem) -> bool {
+    match item {
+        ResponseItem::Message { .. }
+        | ResponseItem::Reasoning { .. }
+        | ResponseItem::LocalShellCall { .. }
+        | ResponseItem::FunctionCall { .. }
+        | ResponseItem::FunctionCallOutput { .. }
+        | ResponseItem::CustomToolCall { .. }
+        | ResponseItem::CustomToolCallOutput { .. } => true,
+        ResponseItem::WebSearchCall { .. } | ResponseItem::Other => false,
+    }
+}
+
+/// Whether an `EventMsg` should be persisted in rollout files.
+#[inline]
+pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
+    match ev {
+        EventMsg::UserMessage(_)
+        | EventMsg::AgentMessage(_)
+        | EventMsg::AgentReasoning(_)
+        | EventMsg::AgentReasoningRawContent(_)
+        | EventMsg::TokenCount(_)
+        | EventMsg::EnteredReviewMode(_)
+        | EventMsg::ExitedReviewMode(_) => true,
+        EventMsg::Error(_)
+        | EventMsg::TaskStarted(_)
+        | EventMsg::TaskComplete(_)
+        | EventMsg::AgentMessageDelta(_)
+        | EventMsg::AgentReasoningDelta(_)
+        | EventMsg::AgentReasoningRawContentDelta(_)
+        | EventMsg::AgentReasoningSectionBreak(_)
+        | EventMsg::SessionConfigured(_)
+        | EventMsg::McpToolCallBegin(_)
+        | EventMsg::McpToolCallEnd(_)
+        | EventMsg::WebSearchBegin(_)
+        | EventMsg::WebSearchEnd(_)
+        | EventMsg::ExecCommandBegin(_)
+        | EventMsg::ExecCommandOutputDelta(_)
+        | EventMsg::ExecCommandEnd(_)
+        | EventMsg::ExecApprovalRequest(_)
+        | EventMsg::ApplyPatchApprovalRequest(_)
+        | EventMsg::BackgroundEvent(_)
+        | EventMsg::StreamError(_)
+        | EventMsg::PatchApplyBegin(_)
+        | EventMsg::PatchApplyEnd(_)
+        | EventMsg::TurnDiff(_)
+        | EventMsg::GetHistoryEntryResponse(_)
+        | EventMsg::McpListToolsResponse(_)
+        | EventMsg::ListCustomPromptsResponse(_)
+        | EventMsg::PlanUpdate(_)
+        | EventMsg::TurnAborted(_)
+        | EventMsg::ShutdownComplete
+        | EventMsg::ConversationPath(_) => false,
+    }
+}

codex-rs/core/src/rollout/recorder.rs

@@ -0,0 +1,423 @@
+//! Persist Codex session rollouts (.jsonl) so sessions can be replayed or inspected later.
+
+use std::fs::File;
+use std::fs::{self};
+use std::io::Error as IoError;
+use std::path::Path;
+use std::path::PathBuf;
+
+use codex_protocol::mcp_protocol::ConversationId;
+use serde::Deserialize;
+use serde::Serialize;
+use serde_json::Value;
+use time::OffsetDateTime;
+use time::format_description::FormatItem;
+use time::macros::format_description;
+use tokio::io::AsyncWriteExt;
+use tokio::sync::mpsc::Sender;
+use tokio::sync::mpsc::{self};
+use tokio::sync::oneshot;
+use tracing::info;
+use tracing::warn;
+
+use super::SESSIONS_SUBDIR;
+use super::list::ConversationsPage;
+use super::list::Cursor;
+use super::list::get_conversations;
+use super::policy::is_persisted_response_item;
+use crate::config::Config;
+use crate::default_client::ORIGINATOR;
+use crate::git_info::collect_git_info;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::InitialHistory;
+use codex_protocol::protocol::ResumedHistory;
+use codex_protocol::protocol::RolloutItem;
+use codex_protocol::protocol::RolloutLine;
+use codex_protocol::protocol::SessionMeta;
+use codex_protocol::protocol::SessionMetaLine;
+
+#[derive(Serialize, Deserialize, Default, Clone)]
+pub struct SessionStateSnapshot {}
+
+#[derive(Serialize, Deserialize, Default, Clone)]
+pub struct SavedSession {
+    pub session: SessionMeta,
+    #[serde(default)]
+    pub items: Vec<ResponseItem>,
+    #[serde(default)]
+    pub state: SessionStateSnapshot,
+    pub session_id: ConversationId,
+}
+
+/// Records all [`ResponseItem`]s for a session and flushes them to disk after
+/// every update.
+///
+/// Rollouts are recorded as JSONL and can be inspected with tools such as:
+///
+/// ```ignore
+/// $ jq -C . ~/.codex/sessions/rollout-2025-05-07T17-24-21-5973b6c0-94b8-487b-a530-2aeb6098ae0e.jsonl
+/// $ fx ~/.codex/sessions/rollout-2025-05-07T17-24-21-5973b6c0-94b8-487b-a530-2aeb6098ae0e.jsonl
+/// ```
+#[derive(Clone)]
+pub struct RolloutRecorder {
+    tx: Sender<RolloutCmd>,
+    pub(crate) rollout_path: PathBuf,
+}
+
+#[derive(Clone)]
+pub enum RolloutRecorderParams {
+    Create {
+        conversation_id: ConversationId,
+        instructions: Option<String>,
+    },
+    Resume {
+        path: PathBuf,
+    },
+}
+
+enum RolloutCmd {
+    AddItems(Vec<RolloutItem>),
+    /// Ensure all prior writes are processed; respond when flushed.
+    Flush {
+        ack: oneshot::Sender<()>,
+    },
+    Shutdown {
+        ack: oneshot::Sender<()>,
+    },
+}
+
+impl RolloutRecorderParams {
+    pub fn new(conversation_id: ConversationId, instructions: Option<String>) -> Self {
+        Self::Create {
+            conversation_id,
+            instructions,
+        }
+    }
+
+    pub fn resume(path: PathBuf) -> Self {
+        Self::Resume { path }
+    }
+}
+
+impl RolloutRecorder {
+    /// List conversations (rollout files) under the provided Codex home directory.
+    pub async fn list_conversations(
+        codex_home: &Path,
+        page_size: usize,
+        cursor: Option<&Cursor>,
+    ) -> std::io::Result<ConversationsPage> {
+        get_conversations(codex_home, page_size, cursor).await
+    }
+
+    /// Attempt to create a new [`RolloutRecorder`]. If the sessions directory
+    /// cannot be created or the rollout file cannot be opened we return the
+    /// error so the caller can decide whether to disable persistence.
+    pub async fn new(config: &Config, params: RolloutRecorderParams) -> std::io::Result<Self> {
+        let (file, rollout_path, meta) = match params {
+            RolloutRecorderParams::Create {
+                conversation_id,
+                instructions,
+            } => {
+                let LogFileInfo {
+                    file,
+                    path,
+                    conversation_id: session_id,
+                    timestamp,
+                } = create_log_file(config, conversation_id)?;
+
+                let timestamp_format: &[FormatItem] = format_description!(
+                    "[year]-[month]-[day]T[hour]:[minute]:[second].[subsecond digits:3]Z"
+                );
+                let timestamp = timestamp
+                    .to_offset(time::UtcOffset::UTC)
+                    .format(timestamp_format)
+                    .map_err(|e| IoError::other(format!("failed to format timestamp: {e}")))?;
+
+                (
+                    tokio::fs::File::from_std(file),
+                    path,
+                    Some(SessionMeta {
+                        id: session_id,
+                        timestamp,
+                        cwd: config.cwd.clone(),
+                        originator: ORIGINATOR.value.clone(),
+                        cli_version: env!("CARGO_PKG_VERSION").to_string(),
+                        instructions,
+                    }),
+                )
+            }
+            RolloutRecorderParams::Resume { path } => (
+                tokio::fs::OpenOptions::new()
+                    .append(true)
+                    .open(&path)
+                    .await?,
+                path,
+                None,
+            ),
+        };
+
+        // Clone the cwd for the spawned task to collect git info asynchronously
+        let cwd = config.cwd.clone();
+
+        // A reasonably-sized bounded channel. If the buffer fills up the send
+        // future will yield, which is fine – we only need to ensure we do not
+        // perform *blocking* I/O on the caller's thread.
+        let (tx, rx) = mpsc::channel::<RolloutCmd>(256);
+
+        // Spawn a Tokio task that owns the file handle and performs async
+        // writes. Using `tokio::fs::File` keeps everything on the async I/O
+        // driver instead of blocking the runtime.
+        tokio::task::spawn(rollout_writer(file, rx, meta, cwd));
+
+        Ok(Self { tx, rollout_path })
+    }
+
+    pub(crate) async fn record_items(&self, items: &[RolloutItem]) -> std::io::Result<()> {
+        let mut filtered = Vec::new();
+        for item in items {
+            // Note that function calls may look a bit strange if they are
+            // "fully qualified MCP tool calls," so we could consider
+            // reformatting them in that case.
+            if is_persisted_response_item(item) {
+                filtered.push(item.clone());
+            }
+        }
+        if filtered.is_empty() {
+            return Ok(());
+        }
+        self.tx
+            .send(RolloutCmd::AddItems(filtered))
+            .await
+            .map_err(|e| IoError::other(format!("failed to queue rollout items: {e}")))
+    }
+
+    /// Flush all queued writes and wait until they are committed by the writer task.
+    pub async fn flush(&self) -> std::io::Result<()> {
+        let (tx, rx) = oneshot::channel();
+        self.tx
+            .send(RolloutCmd::Flush { ack: tx })
+            .await
+            .map_err(|e| IoError::other(format!("failed to queue rollout flush: {e}")))?;
+        rx.await
+            .map_err(|e| IoError::other(format!("failed waiting for rollout flush: {e}")))
+    }
+
+    pub(crate) async fn get_rollout_history(path: &Path) -> std::io::Result<InitialHistory> {
+        info!("Resuming rollout from {path:?}");
+        let text = tokio::fs::read_to_string(path).await?;
+        if text.trim().is_empty() {
+            return Err(IoError::other("empty session file"));
+        }
+
+        let mut items: Vec<RolloutItem> = Vec::new();
+        let mut conversation_id: Option<ConversationId> = None;
+        for line in text.lines() {
+            if line.trim().is_empty() {
+                continue;
+            }
+            let v: Value = match serde_json::from_str(line) {
+                Ok(v) => v,
+                Err(e) => {
+                    warn!("failed to parse line as JSON: {line:?}, error: {e}");
+                    continue;
+                }
+            };
+
+            // Parse the rollout line structure
+            match serde_json::from_value::<RolloutLine>(v.clone()) {
+                Ok(rollout_line) => match rollout_line.item {
+                    RolloutItem::SessionMeta(session_meta_line) => {
+                        // Use the FIRST SessionMeta encountered in the file as the canonical
+                        // conversation id and main session information. Keep all items intact.
+                        if conversation_id.is_none() {
+                            conversation_id = Some(session_meta_line.meta.id);
+                        }
+                        items.push(RolloutItem::SessionMeta(session_meta_line));
+                    }
+                    RolloutItem::ResponseItem(item) => {
+                        items.push(RolloutItem::ResponseItem(item));
+                    }
+                    RolloutItem::Compacted(item) => {
+                        items.push(RolloutItem::Compacted(item));
+                    }
+                    RolloutItem::TurnContext(item) => {
+                        items.push(RolloutItem::TurnContext(item));
+                    }
+                    RolloutItem::EventMsg(_ev) => {
+                        items.push(RolloutItem::EventMsg(_ev));
+                    }
+                },
+                Err(e) => {
+                    warn!("failed to parse rollout line: {v:?}, error: {e}");
+                }
+            }
+        }
+
+        info!(
+            "Resumed rollout with {} items, conversation ID: {:?}",
+            items.len(),
+            conversation_id
+        );
+        let conversation_id = conversation_id
+            .ok_or_else(|| IoError::other("failed to parse conversation ID from rollout file"))?;
+
+        if items.is_empty() {
+            return Ok(InitialHistory::New);
+        }
+
+        info!("Resumed rollout successfully from {path:?}");
+        Ok(InitialHistory::Resumed(ResumedHistory {
+            conversation_id,
+            history: items,
+            rollout_path: path.to_path_buf(),
+        }))
+    }
+
+    pub(crate) fn get_rollout_path(&self) -> PathBuf {
+        self.rollout_path.clone()
+    }
+
+    pub async fn shutdown(&self) -> std::io::Result<()> {
+        let (tx_done, rx_done) = oneshot::channel();
+        match self.tx.send(RolloutCmd::Shutdown { ack: tx_done }).await {
+            Ok(_) => rx_done
+                .await
+                .map_err(|e| IoError::other(format!("failed waiting for rollout shutdown: {e}"))),
+            Err(e) => {
+                warn!("failed to send rollout shutdown command: {e}");
+                Err(IoError::other(format!(
+                    "failed to send rollout shutdown command: {e}"
+                )))
+            }
+        }
+    }
+}
+
+struct LogFileInfo {
+    /// Opened file handle to the rollout file.
+    file: File,
+
+    /// Full path to the rollout file.
+    path: PathBuf,
+
+    /// Session ID (also embedded in filename).
+    conversation_id: ConversationId,
+
+    /// Timestamp for the start of the session.
+    timestamp: OffsetDateTime,
+}
+
+fn create_log_file(
+    config: &Config,
+    conversation_id: ConversationId,
+) -> std::io::Result<LogFileInfo> {
+    // Resolve ~/.codex/sessions/YYYY/MM/DD and create it if missing.
+    let timestamp = OffsetDateTime::now_local()
+        .map_err(|e| IoError::other(format!("failed to get local time: {e}")))?;
+    let mut dir = config.codex_home.clone();
+    dir.push(SESSIONS_SUBDIR);
+    dir.push(timestamp.year().to_string());
+    dir.push(format!("{:02}", u8::from(timestamp.month())));
+    dir.push(format!("{:02}", timestamp.day()));
+    fs::create_dir_all(&dir)?;
+
+    // Custom format for YYYY-MM-DDThh-mm-ss. Use `-` instead of `:` for
+    // compatibility with filesystems that do not allow colons in filenames.
+    let format: &[FormatItem] =
+        format_description!("[year]-[month]-[day]T[hour]-[minute]-[second]");
+    let date_str = timestamp
+        .format(format)
+        .map_err(|e| IoError::other(format!("failed to format timestamp: {e}")))?;
+
+    let filename = format!("rollout-{date_str}-{conversation_id}.jsonl");
+
+    let path = dir.join(filename);
+    let file = std::fs::OpenOptions::new()
+        .append(true)
+        .create(true)
+        .open(&path)?;
+
+    Ok(LogFileInfo {
+        file,
+        path,
+        conversation_id,
+        timestamp,
+    })
+}
+
+async fn rollout_writer(
+    file: tokio::fs::File,
+    mut rx: mpsc::Receiver<RolloutCmd>,
+    mut meta: Option<SessionMeta>,
+    cwd: std::path::PathBuf,
+) -> std::io::Result<()> {
+    let mut writer = JsonlWriter { file };
+
+    // If we have a meta, collect git info asynchronously and write meta first
+    if let Some(session_meta) = meta.take() {
+        let git_info = collect_git_info(&cwd).await;
+        let session_meta_line = SessionMetaLine {
+            meta: session_meta,
+            git: git_info,
+        };
+
+        // Write the SessionMeta as the first item in the file, wrapped in a rollout line
+        writer
+            .write_rollout_item(RolloutItem::SessionMeta(session_meta_line))
+            .await?;
+    }
+
+    // Process rollout commands
+    while let Some(cmd) = rx.recv().await {
+        match cmd {
+            RolloutCmd::AddItems(items) => {
+                for item in items {
+                    if is_persisted_response_item(&item) {
+                        writer.write_rollout_item(item).await?;
+                    }
+                }
+            }
+            RolloutCmd::Flush { ack } => {
+                // Ensure underlying file is flushed and then ack.
+                if let Err(e) = writer.file.flush().await {
+                    let _ = ack.send(());
+                    return Err(e);
+                }
+                let _ = ack.send(());
+            }
+            RolloutCmd::Shutdown { ack } => {
+                let _ = ack.send(());
+            }
+        }
+    }
+
+    Ok(())
+}
+
+struct JsonlWriter {
+    file: tokio::fs::File,
+}
+
+impl JsonlWriter {
+    async fn write_rollout_item(&mut self, rollout_item: RolloutItem) -> std::io::Result<()> {
+        let timestamp_format: &[FormatItem] = format_description!(
+            "[year]-[month]-[day]T[hour]:[minute]:[second].[subsecond digits:3]Z"
+        );
+        let timestamp = OffsetDateTime::now_utc()
+            .format(timestamp_format)
+            .map_err(|e| IoError::other(format!("failed to format timestamp: {e}")))?;
+
+        let line = RolloutLine {
+            timestamp,
+            item: rollout_item,
+        };
+        self.write_line(&line).await
+    }
+    async fn write_line(&mut self, item: &impl serde::Serialize) -> std::io::Result<()> {
+        let mut json = serde_json::to_string(item)?;
+        json.push('\n');
+        self.file.write_all(json.as_bytes()).await?;
+        self.file.flush().await?;
+        Ok(())
+    }
+}

codex-rs/core/src/rollout/tests.rs

@@ -0,0 +1,445 @@
+#![allow(clippy::unwrap_used, clippy::expect_used)]
+
+use std::fs::File;
+use std::fs::{self};
+use std::io::Write;
+use std::path::Path;
+
+use tempfile::TempDir;
+use time::OffsetDateTime;
+use time::PrimitiveDateTime;
+use time::format_description::FormatItem;
+use time::macros::format_description;
+use uuid::Uuid;
+
+use crate::rollout::list::ConversationItem;
+use crate::rollout::list::ConversationsPage;
+use crate::rollout::list::Cursor;
+use crate::rollout::list::get_conversation;
+use crate::rollout::list::get_conversations;
+
+fn write_session_file(
+    root: &Path,
+    ts_str: &str,
+    uuid: Uuid,
+    num_records: usize,
+) -> std::io::Result<(OffsetDateTime, Uuid)> {
+    let format: &[FormatItem] =
+        format_description!("[year]-[month]-[day]T[hour]-[minute]-[second]");
+    let dt = PrimitiveDateTime::parse(ts_str, format)
+        .unwrap()
+        .assume_utc();
+    let dir = root
+        .join("sessions")
+        .join(format!("{:04}", dt.year()))
+        .join(format!("{:02}", u8::from(dt.month())))
+        .join(format!("{:02}", dt.day()));
+    fs::create_dir_all(&dir)?;
+
+    let filename = format!("rollout-{ts_str}-{uuid}.jsonl");
+    let file_path = dir.join(filename);
+    let mut file = File::create(file_path)?;
+
+    let meta = serde_json::json!({
+        "timestamp": ts_str,
+        "type": "session_meta",
+        "payload": {
+            "id": uuid,
+            "timestamp": ts_str,
+            "instructions": null,
+            "cwd": ".",
+            "originator": "test_originator",
+            "cli_version": "test_version"
+        }
+    });
+    writeln!(file, "{meta}")?;
+
+    // Include at least one user message event to satisfy listing filters
+    let user_event = serde_json::json!({
+        "timestamp": ts_str,
+        "type": "event_msg",
+        "payload": {
+            "type": "user_message",
+            "message": "Hello from user",
+            "kind": "plain"
+        }
+    });
+    writeln!(file, "{user_event}")?;
+
+    for i in 0..num_records {
+        let rec = serde_json::json!({
+            "record_type": "response",
+            "index": i
+        });
+        writeln!(file, "{rec}")?;
+    }
+    Ok((dt, uuid))
+}
+
+#[tokio::test]
+async fn test_list_conversations_latest_first() {
+    let temp = TempDir::new().unwrap();
+    let home = temp.path();
+
+    // Fixed UUIDs for deterministic expectations
+    let u1 = Uuid::from_u128(1);
+    let u2 = Uuid::from_u128(2);
+    let u3 = Uuid::from_u128(3);
+
+    // Create three sessions across three days
+    write_session_file(home, "2025-01-01T12-00-00", u1, 3).unwrap();
+    write_session_file(home, "2025-01-02T12-00-00", u2, 3).unwrap();
+    write_session_file(home, "2025-01-03T12-00-00", u3, 3).unwrap();
+
+    let page = get_conversations(home, 10, None).await.unwrap();
+
+    // Build expected objects
+    let p1 = home
+        .join("sessions")
+        .join("2025")
+        .join("01")
+        .join("03")
+        .join(format!("rollout-2025-01-03T12-00-00-{u3}.jsonl"));
+    let p2 = home
+        .join("sessions")
+        .join("2025")
+        .join("01")
+        .join("02")
+        .join(format!("rollout-2025-01-02T12-00-00-{u2}.jsonl"));
+    let p3 = home
+        .join("sessions")
+        .join("2025")
+        .join("01")
+        .join("01")
+        .join(format!("rollout-2025-01-01T12-00-00-{u1}.jsonl"));
+
+    let head_3 = vec![serde_json::json!({
+        "id": u3,
+        "timestamp": "2025-01-03T12-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let head_2 = vec![serde_json::json!({
+        "id": u2,
+        "timestamp": "2025-01-02T12-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let head_1 = vec![serde_json::json!({
+        "id": u1,
+        "timestamp": "2025-01-01T12-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+
+    let expected_cursor: Cursor =
+        serde_json::from_str(&format!("\"2025-01-01T12-00-00|{u1}\"")).unwrap();
+
+    let expected = ConversationsPage {
+        items: vec![
+            ConversationItem {
+                path: p1,
+                head: head_3,
+            },
+            ConversationItem {
+                path: p2,
+                head: head_2,
+            },
+            ConversationItem {
+                path: p3,
+                head: head_1,
+            },
+        ],
+        next_cursor: Some(expected_cursor),
+        num_scanned_files: 3,
+        reached_scan_cap: false,
+    };
+
+    assert_eq!(page, expected);
+}
+
+#[tokio::test]
+async fn test_pagination_cursor() {
+    let temp = TempDir::new().unwrap();
+    let home = temp.path();
+
+    // Fixed UUIDs for deterministic expectations
+    let u1 = Uuid::from_u128(11);
+    let u2 = Uuid::from_u128(22);
+    let u3 = Uuid::from_u128(33);
+    let u4 = Uuid::from_u128(44);
+    let u5 = Uuid::from_u128(55);
+
+    // Oldest to newest
+    write_session_file(home, "2025-03-01T09-00-00", u1, 1).unwrap();
+    write_session_file(home, "2025-03-02T09-00-00", u2, 1).unwrap();
+    write_session_file(home, "2025-03-03T09-00-00", u3, 1).unwrap();
+    write_session_file(home, "2025-03-04T09-00-00", u4, 1).unwrap();
+    write_session_file(home, "2025-03-05T09-00-00", u5, 1).unwrap();
+
+    let page1 = get_conversations(home, 2, None).await.unwrap();
+    let p5 = home
+        .join("sessions")
+        .join("2025")
+        .join("03")
+        .join("05")
+        .join(format!("rollout-2025-03-05T09-00-00-{u5}.jsonl"));
+    let p4 = home
+        .join("sessions")
+        .join("2025")
+        .join("03")
+        .join("04")
+        .join(format!("rollout-2025-03-04T09-00-00-{u4}.jsonl"));
+    let head_5 = vec![serde_json::json!({
+        "id": u5,
+        "timestamp": "2025-03-05T09-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let head_4 = vec![serde_json::json!({
+        "id": u4,
+        "timestamp": "2025-03-04T09-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let expected_cursor1: Cursor =
+        serde_json::from_str(&format!("\"2025-03-04T09-00-00|{u4}\"")).unwrap();
+    let expected_page1 = ConversationsPage {
+        items: vec![
+            ConversationItem {
+                path: p5,
+                head: head_5,
+            },
+            ConversationItem {
+                path: p4,
+                head: head_4,
+            },
+        ],
+        next_cursor: Some(expected_cursor1.clone()),
+        num_scanned_files: 3, // scanned 05, 04, and peeked at 03 before breaking
+        reached_scan_cap: false,
+    };
+    assert_eq!(page1, expected_page1);
+
+    let page2 = get_conversations(home, 2, page1.next_cursor.as_ref())
+        .await
+        .unwrap();
+    let p3 = home
+        .join("sessions")
+        .join("2025")
+        .join("03")
+        .join("03")
+        .join(format!("rollout-2025-03-03T09-00-00-{u3}.jsonl"));
+    let p2 = home
+        .join("sessions")
+        .join("2025")
+        .join("03")
+        .join("02")
+        .join(format!("rollout-2025-03-02T09-00-00-{u2}.jsonl"));
+    let head_3 = vec![serde_json::json!({
+        "id": u3,
+        "timestamp": "2025-03-03T09-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let head_2 = vec![serde_json::json!({
+        "id": u2,
+        "timestamp": "2025-03-02T09-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let expected_cursor2: Cursor =
+        serde_json::from_str(&format!("\"2025-03-02T09-00-00|{u2}\"")).unwrap();
+    let expected_page2 = ConversationsPage {
+        items: vec![
+            ConversationItem {
+                path: p3,
+                head: head_3,
+            },
+            ConversationItem {
+                path: p2,
+                head: head_2,
+            },
+        ],
+        next_cursor: Some(expected_cursor2.clone()),
+        num_scanned_files: 5, // scanned 05, 04 (anchor), 03, 02, and peeked at 01
+        reached_scan_cap: false,
+    };
+    assert_eq!(page2, expected_page2);
+
+    let page3 = get_conversations(home, 2, page2.next_cursor.as_ref())
+        .await
+        .unwrap();
+    let p1 = home
+        .join("sessions")
+        .join("2025")
+        .join("03")
+        .join("01")
+        .join(format!("rollout-2025-03-01T09-00-00-{u1}.jsonl"));
+    let head_1 = vec![serde_json::json!({
+        "id": u1,
+        "timestamp": "2025-03-01T09-00-00",
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let expected_cursor3: Cursor =
+        serde_json::from_str(&format!("\"2025-03-01T09-00-00|{u1}\"")).unwrap();
+    let expected_page3 = ConversationsPage {
+        items: vec![ConversationItem {
+            path: p1,
+            head: head_1,
+        }],
+        next_cursor: Some(expected_cursor3),
+        num_scanned_files: 5, // scanned 05, 04 (anchor), 03, 02 (anchor), 01
+        reached_scan_cap: false,
+    };
+    assert_eq!(page3, expected_page3);
+}
+
+#[tokio::test]
+async fn test_get_conversation_contents() {
+    let temp = TempDir::new().unwrap();
+    let home = temp.path();
+
+    let uuid = Uuid::new_v4();
+    let ts = "2025-04-01T10-30-00";
+    write_session_file(home, ts, uuid, 2).unwrap();
+
+    let page = get_conversations(home, 1, None).await.unwrap();
+    let path = &page.items[0].path;
+
+    let content = get_conversation(path).await.unwrap();
+
+    // Page equality (single item)
+    let expected_path = home
+        .join("sessions")
+        .join("2025")
+        .join("04")
+        .join("01")
+        .join(format!("rollout-2025-04-01T10-30-00-{uuid}.jsonl"));
+    let expected_head = vec![serde_json::json!({
+        "id": uuid,
+        "timestamp": ts,
+        "instructions": null,
+        "cwd": ".",
+        "originator": "test_originator",
+        "cli_version": "test_version"
+    })];
+    let expected_cursor: Cursor = serde_json::from_str(&format!("\"{ts}|{uuid}\"")).unwrap();
+    let expected_page = ConversationsPage {
+        items: vec![ConversationItem {
+            path: expected_path,
+            head: expected_head,
+        }],
+        next_cursor: Some(expected_cursor),
+        num_scanned_files: 1,
+        reached_scan_cap: false,
+    };
+    assert_eq!(page, expected_page);
+
+    // Entire file contents equality
+    let meta = serde_json::json!({"timestamp": ts, "type": "session_meta", "payload": {"id": uuid, "timestamp": ts, "instructions": null, "cwd": ".", "originator": "test_originator", "cli_version": "test_version"}});
+    let user_event = serde_json::json!({
+        "timestamp": ts,
+        "type": "event_msg",
+        "payload": {"type": "user_message", "message": "Hello from user", "kind": "plain"}
+    });
+    let rec0 = serde_json::json!({"record_type": "response", "index": 0});
+    let rec1 = serde_json::json!({"record_type": "response", "index": 1});
+    let expected_content = format!("{meta}\n{user_event}\n{rec0}\n{rec1}\n");
+    assert_eq!(content, expected_content);
+}
+
+#[tokio::test]
+async fn test_stable_ordering_same_second_pagination() {
+    let temp = TempDir::new().unwrap();
+    let home = temp.path();
+
+    let ts = "2025-07-01T00-00-00";
+    let u1 = Uuid::from_u128(1);
+    let u2 = Uuid::from_u128(2);
+    let u3 = Uuid::from_u128(3);
+
+    write_session_file(home, ts, u1, 0).unwrap();
+    write_session_file(home, ts, u2, 0).unwrap();
+    write_session_file(home, ts, u3, 0).unwrap();
+
+    let page1 = get_conversations(home, 2, None).await.unwrap();
+
+    let p3 = home
+        .join("sessions")
+        .join("2025")
+        .join("07")
+        .join("01")
+        .join(format!("rollout-2025-07-01T00-00-00-{u3}.jsonl"));
+    let p2 = home
+        .join("sessions")
+        .join("2025")
+        .join("07")
+        .join("01")
+        .join(format!("rollout-2025-07-01T00-00-00-{u2}.jsonl"));
+    let head = |u: Uuid| -> Vec<serde_json::Value> {
+        vec![serde_json::json!({
+            "id": u,
+            "timestamp": ts,
+            "instructions": null,
+            "cwd": ".",
+            "originator": "test_originator",
+            "cli_version": "test_version"
+        })]
+    };
+    let expected_cursor1: Cursor = serde_json::from_str(&format!("\"{ts}|{u2}\"")).unwrap();
+    let expected_page1 = ConversationsPage {
+        items: vec![
+            ConversationItem {
+                path: p3,
+                head: head(u3),
+            },
+            ConversationItem {
+                path: p2,
+                head: head(u2),
+            },
+        ],
+        next_cursor: Some(expected_cursor1.clone()),
+        num_scanned_files: 3, // scanned u3, u2, peeked u1
+        reached_scan_cap: false,
+    };
+    assert_eq!(page1, expected_page1);
+
+    let page2 = get_conversations(home, 2, page1.next_cursor.as_ref())
+        .await
+        .unwrap();
+    let p1 = home
+        .join("sessions")
+        .join("2025")
+        .join("07")
+        .join("01")
+        .join(format!("rollout-2025-07-01T00-00-00-{u1}.jsonl"));
+    let expected_cursor2: Cursor = serde_json::from_str(&format!("\"{ts}|{u1}\"")).unwrap();
+    let expected_page2 = ConversationsPage {
+        items: vec![ConversationItem {
+            path: p1,
+            head: head(u1),
+        }],
+        next_cursor: Some(expected_cursor2),
+        num_scanned_files: 3, // scanned u3, u2 (anchor), u1
+        reached_scan_cap: false,
+    };
+    assert_eq!(page2, expected_page2);
+}

codex-rs/core/src/safety.rs

@@ -53,6 +53,13 @@ pub fn assess_patch_safety(
         // paths outside the project.
         match get_platform_sandbox() {
             Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+            None if sandbox_policy == &SandboxPolicy::DangerFullAccess => {
+                // If the user has explicitly requested DangerFullAccess, then
+                // we can auto-approve even without a sandbox.
+                SafetyCheck::AutoApprove {
+                    sandbox_type: SandboxType::None,
+                }
+            }
             None => SafetyCheck::AskUser,
         }
     } else if policy == AskForApproval::Never {
@@ -286,7 +293,7 @@ mod tests {
         // With the parent dir explicitly added as a writable root, the
         // outside write should be permitted.
         let policy_with_parent = SandboxPolicy::WorkspaceWrite {
-            writable_roots: vec![parent.clone()],
+            writable_roots: vec![parent],
             network_access: false,
             exclude_tmpdir_env_var: true,
             exclude_slash_tmp: true,

codex-rs/core/src/seatbelt.rs

@@ -153,7 +153,7 @@ mod tests {
         // Build a policy that only includes the two test roots as writable and
         // does not automatically include defaults TMPDIR or /tmp.
         let policy = SandboxPolicy::WorkspaceWrite {
-            writable_roots: vec![root_with_git.clone(), root_without_git.clone()],
+            writable_roots: vec![root_with_git, root_without_git],
             network_access: false,
             exclude_tmpdir_env_var: true,
             exclude_slash_tmp: true,

codex-rs/core/src/seatbelt_base_policy.sbpl

@@ -2,6 +2,7 @@
 
 ; inspired by Chrome's sandbox policy:
 ; https://source.chromium.org/chromium/chromium/src/+/main:sandbox/policy/mac/common.sb;l=273-319;drc=7b3962fe2e5fc9e2ee58000dc8fbf3429d84d3bd
+; https://source.chromium.org/chromium/chromium/src/+/main:sandbox/policy/mac/renderer.sb;l=64;drc=7b3962fe2e5fc9e2ee58000dc8fbf3429d84d3bd
 
 ; start with closed-by-default
 (deny default)
@@ -9,7 +10,13 @@
 ; child processes inherit the policy of their parent
 (allow process-exec)
 (allow process-fork)
-(allow signal (target self))
+(allow signal (target same-sandbox))
+
+; Allow cf prefs to work.
+(allow user-preference-read)
+
+; process-info
+(allow process-info* (target same-sandbox))
 
 (allow file-write-data
   (require-all
@@ -32,28 +39,22 @@
   (sysctl-name "hw.l3cachesize_compat")
   (sysctl-name "hw.logicalcpu_max")
   (sysctl-name "hw.machine")
+  (sysctl-name "hw.memsize")
   (sysctl-name "hw.ncpu")
   (sysctl-name "hw.nperflevels")
-  (sysctl-name "hw.optional.arm.FEAT_BF16")
-  (sysctl-name "hw.optional.arm.FEAT_DotProd")
-  (sysctl-name "hw.optional.arm.FEAT_FCMA")
-  (sysctl-name "hw.optional.arm.FEAT_FHM")
-  (sysctl-name "hw.optional.arm.FEAT_FP16")
-  (sysctl-name "hw.optional.arm.FEAT_I8MM")
-  (sysctl-name "hw.optional.arm.FEAT_JSCVT")
-  (sysctl-name "hw.optional.arm.FEAT_LSE")
-  (sysctl-name "hw.optional.arm.FEAT_RDM")
-  (sysctl-name "hw.optional.arm.FEAT_SHA512")
-  (sysctl-name "hw.optional.armv8_2_sha512")
-  (sysctl-name "hw.memsize")
-  (sysctl-name "hw.pagesize")
+  ; Chrome locks these CPU feature detection down a bit more tightly,
+  ; but mostly for fingerprinting concerns which isn't an issue for codex.
+  (sysctl-name-prefix "hw.optional.arm.")
+  (sysctl-name-prefix "hw.optional.armv8_")
   (sysctl-name "hw.packages")
   (sysctl-name "hw.pagesize_compat")
+  (sysctl-name "hw.pagesize")
   (sysctl-name "hw.physicalcpu_max")
   (sysctl-name "hw.tbfrequency_compat")
   (sysctl-name "hw.vectorunit")
   (sysctl-name "kern.hostname")
   (sysctl-name "kern.maxfilesperproc")
+  (sysctl-name "kern.maxproc")
   (sysctl-name "kern.osproductversion")
   (sysctl-name "kern.osrelease")
   (sysctl-name "kern.ostype")
@@ -63,9 +64,27 @@
   (sysctl-name "kern.usrstack64")
   (sysctl-name "kern.version")
   (sysctl-name "sysctl.proc_cputype")
+  (sysctl-name "vm.loadavg")
   (sysctl-name-prefix "hw.perflevel")
+  (sysctl-name-prefix "kern.proc.pgrp.")
+  (sysctl-name-prefix "kern.proc.pid.")
+  (sysctl-name-prefix "net.routetable.")
+)
+
+; IOKit
+(allow iokit-open
+  (iokit-registry-entry-class "RootDomainUserClient")
+)
+
+; needed to look up user info, see https://crbug.com/792228
+(allow mach-lookup
+  (global-name "com.apple.system.opendirectoryd.libinfo")
 )
 
 ; Added on top of Chrome profile
 ; Needed for python multiprocessing on MacOS for the SemLock
 (allow ipc-posix-sem)
+
+(allow mach-lookup
+  (global-name "com.apple.PowerManagement.control")
+)

codex-rs/core/src/shell.rs

@@ -9,6 +9,12 @@ pub struct ZshShell {
     zshrc_path: String,
 }
 
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
+pub struct BashShell {
+    shell_path: String,
+    bashrc_path: String,
+}
+
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct PowerShellConfig {
     exe: String, // Executable name or path, e.g. "pwsh" or "powershell.exe".
@@ -18,6 +24,7 @@ pub struct PowerShellConfig {
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub enum Shell {
     Zsh(ZshShell),
+    Bash(BashShell),
     PowerShell(PowerShellConfig),
     Unknown,
 }
@@ -26,22 +33,10 @@ impl Shell {
     pub fn format_default_shell_invocation(&self, command: Vec<String>) -> Option<Vec<String>> {
         match self {
             Shell::Zsh(zsh) => {
-                if !std::path::Path::new(&zsh.zshrc_path).exists() {
-                    return None;
-                }
-
-                let mut result = vec![zsh.shell_path.clone()];
-                result.push("-lc".to_string());
-
-                let joined = strip_bash_lc(&command)
-                    .or_else(|| shlex::try_join(command.iter().map(|s| s.as_str())).ok());
-
-                if let Some(joined) = joined {
-                    result.push(format!("source {} && ({joined})", zsh.zshrc_path));
-                } else {
-                    return None;
-                }
-                Some(result)
+                format_shell_invocation_with_rc(&command, &zsh.shell_path, &zsh.zshrc_path)
+            }
+            Shell::Bash(bash) => {
+                format_shell_invocation_with_rc(&command, &bash.shell_path, &bash.bashrc_path)
             }
             Shell::PowerShell(ps) => {
                 // If model generated a bash command, prefer a detected bash fallback
@@ -97,12 +92,32 @@ impl Shell {
             Shell::Zsh(zsh) => std::path::Path::new(&zsh.shell_path)
                 .file_name()
                 .map(|s| s.to_string_lossy().to_string()),
+            Shell::Bash(bash) => std::path::Path::new(&bash.shell_path)
+                .file_name()
+                .map(|s| s.to_string_lossy().to_string()),
             Shell::PowerShell(ps) => Some(ps.exe.clone()),
             Shell::Unknown => None,
         }
     }
 }
 
+fn format_shell_invocation_with_rc(
+    command: &Vec<String>,
+    shell_path: &str,
+    rc_path: &str,
+) -> Option<Vec<String>> {
+    let joined = strip_bash_lc(command)
+        .or_else(|| shlex::try_join(command.iter().map(|s| s.as_str())).ok())?;
+
+    let rc_command = if std::path::Path::new(rc_path).exists() {
+        format!("source {rc_path} && ({joined})")
+    } else {
+        joined
+    };
+
+    Some(vec![shell_path.to_string(), "-lc".to_string(), rc_command])
+}
+
 fn strip_bash_lc(command: &Vec<String>) -> Option<String> {
     match command.as_slice() {
         // exactly three items
@@ -116,44 +131,43 @@ fn strip_bash_lc(command: &Vec<String>) -> Option<String> {
     }
 }
 
-#[cfg(target_os = "macos")]
-pub async fn default_user_shell() -> Shell {
-    use tokio::process::Command;
-    use whoami;
+#[cfg(unix)]
+fn detect_default_user_shell() -> Shell {
+    use libc::getpwuid;
+    use libc::getuid;
+    use std::ffi::CStr;
 
-    let user = whoami::username();
-    let home = format!("/Users/{user}");
-    let output = Command::new("dscl")
-        .args([".", "-read", &home, "UserShell"])
-        .output()
-        .await
-        .ok();
-    match output {
-        Some(o) => {
-            if !o.status.success() {
-                return Shell::Unknown;
-            }
-            let stdout = String::from_utf8_lossy(&o.stdout);
-            for line in stdout.lines() {
-                if let Some(shell_path) = line.strip_prefix("UserShell: ")
-                    && shell_path.ends_with("/zsh")
-                {
-                    return Shell::Zsh(ZshShell {
-                        shell_path: shell_path.to_string(),
-                        zshrc_path: format!("{home}/.zshrc"),
-                    });
-                }
+    unsafe {
+        let uid = getuid();
+        let pw = getpwuid(uid);
+
+        if !pw.is_null() {
+            let shell_path = CStr::from_ptr((*pw).pw_shell)
+                .to_string_lossy()
+                .into_owned();
+            let home_path = CStr::from_ptr((*pw).pw_dir).to_string_lossy().into_owned();
+
+            if shell_path.ends_with("/zsh") {
+                return Shell::Zsh(ZshShell {
+                    shell_path,
+                    zshrc_path: format!("{home_path}/.zshrc"),
+                });
             }
 
-            Shell::Unknown
+            if shell_path.ends_with("/bash") {
+                return Shell::Bash(BashShell {
+                    shell_path,
+                    bashrc_path: format!("{home_path}/.bashrc"),
+                });
+            }
         }
-        _ => Shell::Unknown,
     }
+    Shell::Unknown
 }
 
-#[cfg(all(not(target_os = "macos"), not(target_os = "windows")))]
+#[cfg(unix)]
 pub async fn default_user_shell() -> Shell {
-    Shell::Unknown
+    detect_default_user_shell()
 }
 
 #[cfg(target_os = "windows")]
@@ -196,8 +210,13 @@ pub async fn default_user_shell() -> Shell {
     }
 }
 
+#[cfg(all(not(target_os = "windows"), not(unix)))]
+pub async fn default_user_shell() -> Shell {
+    Shell::Unknown
+}
+
 #[cfg(test)]
-#[cfg(target_os = "macos")]
+#[cfg(unix)]
 mod tests {
     use super::*;
     use std::process::Command;
@@ -230,9 +249,124 @@ mod tests {
             zshrc_path: "/does/not/exist/.zshrc".to_string(),
         });
         let actual_cmd = shell.format_default_shell_invocation(vec!["myecho".to_string()]);
-        assert_eq!(actual_cmd, None);
+        assert_eq!(
+            actual_cmd,
+            Some(vec![
+                "/bin/zsh".to_string(),
+                "-lc".to_string(),
+                "myecho".to_string()
+            ])
+        );
     }
 
+    #[tokio::test]
+    async fn test_run_with_profile_bashrc_not_exists() {
+        let shell = Shell::Bash(BashShell {
+            shell_path: "/bin/bash".to_string(),
+            bashrc_path: "/does/not/exist/.bashrc".to_string(),
+        });
+        let actual_cmd = shell.format_default_shell_invocation(vec!["myecho".to_string()]);
+        assert_eq!(
+            actual_cmd,
+            Some(vec![
+                "/bin/bash".to_string(),
+                "-lc".to_string(),
+                "myecho".to_string()
+            ])
+        );
+    }
+
+    #[tokio::test]
+    async fn test_run_with_profile_bash_escaping_and_execution() {
+        let shell_path = "/bin/bash";
+
+        let cases = vec![
+            (
+                vec!["myecho"],
+                vec![shell_path, "-lc", "source BASHRC_PATH && (myecho)"],
+                Some("It works!\n"),
+            ),
+            (
+                vec!["bash", "-lc", "echo 'single' \"double\""],
+                vec![
+                    shell_path,
+                    "-lc",
+                    "source BASHRC_PATH && (echo 'single' \"double\")",
+                ],
+                Some("single double\n"),
+            ),
+        ];
+
+        for (input, expected_cmd, expected_output) in cases {
+            use std::collections::HashMap;
+
+            use crate::exec::ExecParams;
+            use crate::exec::SandboxType;
+            use crate::exec::process_exec_tool_call;
+            use crate::protocol::SandboxPolicy;
+
+            let temp_home = tempfile::tempdir().unwrap();
+            let bashrc_path = temp_home.path().join(".bashrc");
+            std::fs::write(
+                &bashrc_path,
+                r#"
+                    set -x
+                    function myecho {
+                        echo 'It works!'
+                    }
+                    "#,
+            )
+            .unwrap();
+            let shell = Shell::Bash(BashShell {
+                shell_path: shell_path.to_string(),
+                bashrc_path: bashrc_path.to_str().unwrap().to_string(),
+            });
+
+            let actual_cmd = shell
+                .format_default_shell_invocation(input.iter().map(|s| s.to_string()).collect());
+            let expected_cmd = expected_cmd
+                .iter()
+                .map(|s| s.replace("BASHRC_PATH", bashrc_path.to_str().unwrap()))
+                .collect();
+
+            assert_eq!(actual_cmd, Some(expected_cmd));
+
+            let output = process_exec_tool_call(
+                ExecParams {
+                    command: actual_cmd.unwrap(),
+                    cwd: PathBuf::from(temp_home.path()),
+                    timeout_ms: None,
+                    env: HashMap::from([(
+                        "HOME".to_string(),
+                        temp_home.path().to_str().unwrap().to_string(),
+                    )]),
+                    with_escalated_permissions: None,
+                    justification: None,
+                },
+                SandboxType::None,
+                &SandboxPolicy::DangerFullAccess,
+                &None,
+                None,
+            )
+            .await
+            .unwrap();
+
+            assert_eq!(output.exit_code, 0, "input: {input:?} output: {output:?}");
+            if let Some(expected) = expected_output {
+                assert_eq!(
+                    output.stdout.text, expected,
+                    "input: {input:?} output: {output:?}"
+                );
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+#[cfg(target_os = "macos")]
+mod macos_tests {
+    use super::*;
+
     #[tokio::test]
     async fn test_run_with_profile_escaping_and_execution() {
         let shell_path = "/bin/zsh";
@@ -298,10 +432,7 @@ mod tests {
                 .format_default_shell_invocation(input.iter().map(|s| s.to_string()).collect());
             let expected_cmd = expected_cmd
                 .iter()
-                .map(|s| {
-                    s.replace("ZSHRC_PATH", zshrc_path.to_str().unwrap())
-                        .to_string()
-                })
+                .map(|s| s.replace("ZSHRC_PATH", zshrc_path.to_str().unwrap()))
                 .collect();
 
             assert_eq!(actual_cmd, Some(expected_cmd));

codex-rs/core/src/token_data.rs

@@ -3,8 +3,6 @@ use serde::Deserialize;
 use serde::Serialize;
 use thiserror::Error;
 
-use crate::AuthMode;
-
 #[derive(Deserialize, Serialize, Clone, Debug, PartialEq, Default)]
 pub struct TokenData {
     /// Flat info parsed from the JWT in auth.json.
@@ -22,43 +20,13 @@ pub struct TokenData {
     pub account_id: Option<String>,
 }
 
-impl TokenData {
-    /// Returns true if this is a plan that should use the traditional
-    /// "metered" billing via an API key.
-    pub(crate) fn should_use_api_key(
-        &self,
-        preferred_auth_method: AuthMode,
-        is_openai_email: bool,
-    ) -> bool {
-        if preferred_auth_method == AuthMode::ApiKey {
-            return true;
-        }
-        // If the email is an OpenAI email, use AuthMode::ChatGPT unless preferred_auth_method is AuthMode::ApiKey.
-        if is_openai_email {
-            return false;
-        }
-
-        self.id_token
-            .chatgpt_plan_type
-            .as_ref()
-            .is_none_or(|plan| plan.is_plan_that_should_use_api_key())
-    }
-
-    pub fn is_openai_email(&self) -> bool {
-        self.id_token
-            .email
-            .as_deref()
-            .is_some_and(|email| email.trim().to_ascii_lowercase().ends_with("@openai.com"))
-    }
-}
-
 /// Flat subset of useful claims in id_token from auth.json.
 #[derive(Debug, Clone, PartialEq, Eq, Default, Serialize, Deserialize)]
 pub struct IdTokenInfo {
     pub email: Option<String>,
     /// The ChatGPT subscription plan type
     /// (e.g., "free", "plus", "pro", "business", "enterprise", "edu").
-    /// (Note: ae has not verified that those are the exact values.)
+    /// (Note: values may vary by backend.)
     pub(crate) chatgpt_plan_type: Option<PlanType>,
     pub raw_jwt: String,
 }
@@ -79,28 +47,6 @@ pub(crate) enum PlanType {
     Unknown(String),
 }
 
-impl PlanType {
-    fn is_plan_that_should_use_api_key(&self) -> bool {
-        match self {
-            Self::Known(known) => {
-                use KnownPlan::*;
-                !matches!(known, Free | Plus | Pro | Team)
-            }
-            Self::Unknown(_) => {
-                // Unknown plans should use the API key.
-                true
-            }
-        }
-    }
-
-    pub fn as_string(&self) -> String {
-        match self {
-            Self::Known(known) => format!("{known:?}").to_lowercase(),
-            Self::Unknown(s) => s.clone(),
-        }
-    }
-}
-
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub(crate) enum KnownPlan {
@@ -137,7 +83,7 @@ pub enum IdTokenInfoError {
     Json(#[from] serde_json::Error),
 }
 
-pub(crate) fn parse_id_token(id_token: &str) -> Result<IdTokenInfo, IdTokenInfoError> {
+pub fn parse_id_token(id_token: &str) -> Result<IdTokenInfo, IdTokenInfoError> {
     // JWT format: header.payload.signature
     let mut parts = id_token.split('.');
     let (_header_b64, payload_b64, _sig_b64) = match (parts.next(), parts.next(), parts.next()) {
@@ -204,9 +150,33 @@ mod tests {
 
         let info = parse_id_token(&fake_jwt).expect("should parse");
         assert_eq!(info.email.as_deref(), Some("user@example.com"));
-        assert_eq!(
-            info.chatgpt_plan_type,
-            Some(PlanType::Known(KnownPlan::Pro))
-        );
+        assert_eq!(info.get_chatgpt_plan_type().as_deref(), Some("Pro"));
+    }
+
+    #[test]
+    fn id_token_info_handles_missing_fields() {
+        #[derive(Serialize)]
+        struct Header {
+            alg: &'static str,
+            typ: &'static str,
+        }
+        let header = Header {
+            alg: "none",
+            typ: "JWT",
+        };
+        let payload = serde_json::json!({ "sub": "123" });
+
+        fn b64url_no_pad(bytes: &[u8]) -> String {
+            base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(bytes)
+        }
+
+        let header_b64 = b64url_no_pad(&serde_json::to_vec(&header).unwrap());
+        let payload_b64 = b64url_no_pad(&serde_json::to_vec(&payload).unwrap());
+        let signature_b64 = b64url_no_pad(b"sig");
+        let fake_jwt = format!("{header_b64}.{payload_b64}.{signature_b64}");
+
+        let info = parse_id_token(&fake_jwt).expect("should parse");
+        assert!(info.email.is_none());
+        assert!(info.get_chatgpt_plan_type().is_none());
     }
 }

codex-rs/core/src/tool_apply_patch.lark

@@ -0,0 +1,19 @@
+start: begin_patch hunk+ end_patch
+begin_patch: "*** Begin Patch" LF
+end_patch: "*** End Patch" LF?
+
+hunk: add_hunk | delete_hunk | update_hunk
+add_hunk: "*** Add File: " filename LF add_line+
+delete_hunk: "*** Delete File: " filename LF
+update_hunk: "*** Update File: " filename LF change_move? change?
+
+filename: /(.+)/
+add_line: "+" /(.*)/ LF -> line
+
+change_move: "*** Move to: " filename LF
+change: (change_context | change_line)+ eof_line?
+change_context: ("@@" | "@@ " /(.+)/) LF
+change_line: ("+" | "-" | " ") /(.*)/ LF
+eof_line: "*** End of File" LF
+
+%import common.LF

codex-rs/core/src/tool_apply_patch.rs

@@ -8,6 +8,8 @@ use crate::openai_tools::JsonSchema;
 use crate::openai_tools::OpenAiTool;
 use crate::openai_tools::ResponsesApiTool;
 
+const APPLY_PATCH_LARK_GRAMMAR: &str = include_str!("tool_apply_patch.lark");
+
 #[derive(Serialize, Deserialize)]
 pub(crate) struct ApplyPatchToolArgs {
     pub(crate) input: String,
@@ -29,27 +31,7 @@ pub(crate) fn create_apply_patch_freeform_tool() -> OpenAiTool {
         format: FreeformToolFormat {
             r#type: "grammar".to_string(),
             syntax: "lark".to_string(),
-            definition: r#"start: begin_patch hunk+ end_patch
-begin_patch: "*** Begin Patch" LF
-end_patch: "*** End Patch" LF?
-
-hunk: add_hunk | delete_hunk | update_hunk
-add_hunk: "*** Add File: " filename LF add_line+
-delete_hunk: "*** Delete File: " filename LF
-update_hunk: "*** Update File: " filename LF change_move? change?
-
-filename: /(.+)/
-add_line: "+" /(.+)/ LF -> line
-
-change_move: "*** Move to: " filename LF
-change: (change_context | change_line)+ eof_line?
-change_context: ("@@" | "@@ " /(.+)/) LF
-change_line: ("+" | "-" | " ") /(.+)/ LF
-eof_line: "*** End of File" LF
-
-%import common.LF
-"#
-            .to_string(),
+            definition: APPLY_PATCH_LARK_GRAMMAR.to_string(),
         },
     })
 }

codex-rs/core/src/truncate.rs

@@ -0,0 +1,180 @@
+//! Utilities for truncating large chunks of output while preserving a prefix
+//! and suffix on UTF-8 boundaries.
+
+/// Truncate the middle of a UTF-8 string to at most `max_bytes` bytes,
+/// preserving the beginning and the end. Returns the possibly truncated
+/// string and `Some(original_token_count)` (estimated at 4 bytes/token)
+/// if truncation occurred; otherwise returns the original string and `None`.
+pub(crate) fn truncate_middle(s: &str, max_bytes: usize) -> (String, Option<u64>) {
+    if s.len() <= max_bytes {
+        return (s.to_string(), None);
+    }
+
+    let est_tokens = (s.len() as u64).div_ceil(4);
+    if max_bytes == 0 {
+        return (format!("…{est_tokens} tokens truncated…"), Some(est_tokens));
+    }
+
+    fn truncate_on_boundary(input: &str, max_len: usize) -> &str {
+        if input.len() <= max_len {
+            return input;
+        }
+        let mut end = max_len;
+        while end > 0 && !input.is_char_boundary(end) {
+            end -= 1;
+        }
+        &input[..end]
+    }
+
+    fn pick_prefix_end(s: &str, left_budget: usize) -> usize {
+        if let Some(head) = s.get(..left_budget)
+            && let Some(i) = head.rfind('\n')
+        {
+            return i + 1;
+        }
+        truncate_on_boundary(s, left_budget).len()
+    }
+
+    fn pick_suffix_start(s: &str, right_budget: usize) -> usize {
+        let start_tail = s.len().saturating_sub(right_budget);
+        if let Some(tail) = s.get(start_tail..)
+            && let Some(i) = tail.find('\n')
+        {
+            return start_tail + i + 1;
+        }
+
+        let mut idx = start_tail.min(s.len());
+        while idx < s.len() && !s.is_char_boundary(idx) {
+            idx += 1;
+        }
+        idx
+    }
+
+    let mut guess_tokens = est_tokens;
+    for _ in 0..4 {
+        let marker = format!("…{guess_tokens} tokens truncated…");
+        let marker_len = marker.len();
+        let keep_budget = max_bytes.saturating_sub(marker_len);
+        if keep_budget == 0 {
+            return (format!("…{est_tokens} tokens truncated…"), Some(est_tokens));
+        }
+
+        let left_budget = keep_budget / 2;
+        let right_budget = keep_budget - left_budget;
+        let prefix_end = pick_prefix_end(s, left_budget);
+        let mut suffix_start = pick_suffix_start(s, right_budget);
+        if suffix_start < prefix_end {
+            suffix_start = prefix_end;
+        }
+
+        let kept_content_bytes = prefix_end + (s.len() - suffix_start);
+        let truncated_content_bytes = s.len().saturating_sub(kept_content_bytes);
+        let new_tokens = (truncated_content_bytes as u64).div_ceil(4);
+
+        if new_tokens == guess_tokens {
+            let mut out = String::with_capacity(marker_len + kept_content_bytes + 1);
+            out.push_str(&s[..prefix_end]);
+            out.push_str(&marker);
+            out.push('\n');
+            out.push_str(&s[suffix_start..]);
+            return (out, Some(est_tokens));
+        }
+
+        guess_tokens = new_tokens;
+    }
+
+    let marker = format!("…{guess_tokens} tokens truncated…");
+    let marker_len = marker.len();
+    let keep_budget = max_bytes.saturating_sub(marker_len);
+    if keep_budget == 0 {
+        return (format!("…{est_tokens} tokens truncated…"), Some(est_tokens));
+    }
+
+    let left_budget = keep_budget / 2;
+    let right_budget = keep_budget - left_budget;
+    let prefix_end = pick_prefix_end(s, left_budget);
+    let suffix_start = pick_suffix_start(s, right_budget);
+
+    let mut out = String::with_capacity(marker_len + prefix_end + (s.len() - suffix_start) + 1);
+    out.push_str(&s[..prefix_end]);
+    out.push_str(&marker);
+    out.push('\n');
+    out.push_str(&s[suffix_start..]);
+    (out, Some(est_tokens))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::truncate_middle;
+
+    #[test]
+    fn truncate_middle_no_newlines_fallback() {
+        let s = "abcdefghijklmnopqrstuvwxyz0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ*";
+        let max_bytes = 32;
+        let (out, original) = truncate_middle(s, max_bytes);
+        assert!(out.starts_with("abc"));
+        assert!(out.contains("tokens truncated"));
+        assert!(out.ends_with("XYZ*"));
+        assert_eq!(original, Some((s.len() as u64).div_ceil(4)));
+    }
+
+    #[test]
+    fn truncate_middle_prefers_newline_boundaries() {
+        let mut s = String::new();
+        for i in 1..=20 {
+            s.push_str(&format!("{i:03}\n"));
+        }
+        assert_eq!(s.len(), 80);
+
+        let max_bytes = 64;
+        let (out, tokens) = truncate_middle(&s, max_bytes);
+        assert!(out.starts_with("001\n002\n003\n004\n"));
+        assert!(out.contains("tokens truncated"));
+        assert!(out.ends_with("017\n018\n019\n020\n"));
+        assert_eq!(tokens, Some(20));
+    }
+
+    #[test]
+    fn truncate_middle_handles_utf8_content() {
+        let s = "😀😀😀😀😀😀😀😀😀😀\nsecond line with ascii text\n";
+        let max_bytes = 32;
+        let (out, tokens) = truncate_middle(s, max_bytes);
+
+        assert!(out.contains("tokens truncated"));
+        assert!(!out.contains('\u{fffd}'));
+        assert_eq!(tokens, Some((s.len() as u64).div_ceil(4)));
+    }
+
+    #[test]
+    fn truncate_middle_prefers_newline_boundaries_2() {
+        // Build a multi-line string of 20 numbered lines (each "NNN\n").
+        let mut s = String::new();
+        for i in 1..=20 {
+            s.push_str(&format!("{i:03}\n"));
+        }
+        // Total length: 20 lines * 4 bytes per line = 80 bytes.
+        assert_eq!(s.len(), 80);
+
+        // Choose a cap that forces truncation while leaving room for
+        // a few lines on each side after accounting for the marker.
+        let max_bytes = 64;
+        // Expect exact output: first 4 lines, marker, last 4 lines, and correct token estimate (80/4 = 20).
+        assert_eq!(
+            truncate_middle(&s, max_bytes),
+            (
+                r#"001
+002
+003
+004
+…12 tokens truncated…
+017
+018
+019
+020
+"#
+                .to_string(),
+                Some(20)
+            )
+        );
+    }
+}

codex-rs/core/src/turn_diff_tracker.rs

@@ -678,7 +678,7 @@ index {left_oid}..{right_oid}
         let dest = dir.path().join("dest.txt");
         let mut acc = TurnDiffTracker::new();
         let mv = HashMap::from([(
-            src.clone(),
+            src,
             FileChange::Update {
                 unified_diff: "".into(),
                 move_path: Some(dest.clone()),

codex-rs/core/src/unified_exec/errors.rs

@@ -0,0 +1,22 @@
+use thiserror::Error;
+
+#[derive(Debug, Error)]
+pub(crate) enum UnifiedExecError {
+    #[error("Failed to create unified exec session: {pty_error}")]
+    CreateSession {
+        #[source]
+        pty_error: anyhow::Error,
+    },
+    #[error("Unknown session id {session_id}")]
+    UnknownSessionId { session_id: i32 },
+    #[error("failed to write to stdin")]
+    WriteToStdin,
+    #[error("missing command line for unified exec request")]
+    MissingCommandLine,
+}
+
+impl UnifiedExecError {
+    pub(crate) fn create_session(error: anyhow::Error) -> Self {
+        Self::CreateSession { pty_error: error }
+    }
+}

codex-rs/core/src/unified_exec/mod.rs

@@ -0,0 +1,637 @@
+use portable_pty::CommandBuilder;
+use portable_pty::PtySize;
+use portable_pty::native_pty_system;
+use std::collections::HashMap;
+use std::collections::VecDeque;
+use std::io::ErrorKind;
+use std::io::Read;
+use std::sync::Arc;
+use std::sync::Mutex as StdMutex;
+use std::sync::atomic::AtomicBool;
+use std::sync::atomic::AtomicI32;
+use std::sync::atomic::Ordering;
+use tokio::sync::Mutex;
+use tokio::sync::Notify;
+use tokio::sync::mpsc;
+use tokio::task::JoinHandle;
+use tokio::time::Duration;
+use tokio::time::Instant;
+
+use crate::exec_command::ExecCommandSession;
+use crate::truncate::truncate_middle;
+
+mod errors;
+
+pub(crate) use errors::UnifiedExecError;
+
+const DEFAULT_TIMEOUT_MS: u64 = 1_000;
+const MAX_TIMEOUT_MS: u64 = 60_000;
+const UNIFIED_EXEC_OUTPUT_MAX_BYTES: usize = 128 * 1024; // 128 KiB
+
+#[derive(Debug)]
+pub(crate) struct UnifiedExecRequest<'a> {
+    pub session_id: Option<i32>,
+    pub input_chunks: &'a [String],
+    pub timeout_ms: Option<u64>,
+}
+
+#[derive(Debug, Clone, PartialEq)]
+pub(crate) struct UnifiedExecResult {
+    pub session_id: Option<i32>,
+    pub output: String,
+}
+
+#[derive(Debug, Default)]
+pub(crate) struct UnifiedExecSessionManager {
+    next_session_id: AtomicI32,
+    sessions: Mutex<HashMap<i32, ManagedUnifiedExecSession>>,
+}
+
+#[derive(Debug)]
+struct ManagedUnifiedExecSession {
+    session: ExecCommandSession,
+    output_buffer: OutputBuffer,
+    /// Notifies waiters whenever new output has been appended to
+    /// `output_buffer`, allowing clients to poll for fresh data.
+    output_notify: Arc<Notify>,
+    output_task: JoinHandle<()>,
+}
+
+#[derive(Debug, Default)]
+struct OutputBufferState {
+    chunks: VecDeque<Vec<u8>>,
+    total_bytes: usize,
+}
+
+impl OutputBufferState {
+    fn push_chunk(&mut self, chunk: Vec<u8>) {
+        self.total_bytes = self.total_bytes.saturating_add(chunk.len());
+        self.chunks.push_back(chunk);
+
+        let mut excess = self
+            .total_bytes
+            .saturating_sub(UNIFIED_EXEC_OUTPUT_MAX_BYTES);
+
+        while excess > 0 {
+            match self.chunks.front_mut() {
+                Some(front) if excess >= front.len() => {
+                    excess -= front.len();
+                    self.total_bytes = self.total_bytes.saturating_sub(front.len());
+                    self.chunks.pop_front();
+                }
+                Some(front) => {
+                    front.drain(..excess);
+                    self.total_bytes = self.total_bytes.saturating_sub(excess);
+                    break;
+                }
+                None => break,
+            }
+        }
+    }
+
+    fn drain(&mut self) -> Vec<Vec<u8>> {
+        let drained: Vec<Vec<u8>> = self.chunks.drain(..).collect();
+        self.total_bytes = 0;
+        drained
+    }
+}
+
+type OutputBuffer = Arc<Mutex<OutputBufferState>>;
+type OutputHandles = (OutputBuffer, Arc<Notify>);
+
+impl ManagedUnifiedExecSession {
+    fn new(session: ExecCommandSession) -> Self {
+        let output_buffer = Arc::new(Mutex::new(OutputBufferState::default()));
+        let output_notify = Arc::new(Notify::new());
+        let mut receiver = session.output_receiver();
+        let buffer_clone = Arc::clone(&output_buffer);
+        let notify_clone = Arc::clone(&output_notify);
+        let output_task = tokio::spawn(async move {
+            while let Ok(chunk) = receiver.recv().await {
+                let mut guard = buffer_clone.lock().await;
+                guard.push_chunk(chunk);
+                drop(guard);
+                notify_clone.notify_waiters();
+            }
+        });
+
+        Self {
+            session,
+            output_buffer,
+            output_notify,
+            output_task,
+        }
+    }
+
+    fn writer_sender(&self) -> mpsc::Sender<Vec<u8>> {
+        self.session.writer_sender()
+    }
+
+    fn output_handles(&self) -> OutputHandles {
+        (
+            Arc::clone(&self.output_buffer),
+            Arc::clone(&self.output_notify),
+        )
+    }
+
+    fn has_exited(&self) -> bool {
+        self.session.has_exited()
+    }
+}
+
+impl Drop for ManagedUnifiedExecSession {
+    fn drop(&mut self) {
+        self.output_task.abort();
+    }
+}
+
+impl UnifiedExecSessionManager {
+    pub async fn handle_request(
+        &self,
+        request: UnifiedExecRequest<'_>,
+    ) -> Result<UnifiedExecResult, UnifiedExecError> {
+        let (timeout_ms, timeout_warning) = match request.timeout_ms {
+            Some(requested) if requested > MAX_TIMEOUT_MS => (
+                MAX_TIMEOUT_MS,
+                Some(format!(
+                    "Warning: requested timeout {requested}ms exceeds maximum of {MAX_TIMEOUT_MS}ms; clamping to {MAX_TIMEOUT_MS}ms.\n"
+                )),
+            ),
+            Some(requested) => (requested, None),
+            None => (DEFAULT_TIMEOUT_MS, None),
+        };
+
+        let mut new_session: Option<ManagedUnifiedExecSession> = None;
+        let session_id;
+        let writer_tx;
+        let output_buffer;
+        let output_notify;
+
+        if let Some(existing_id) = request.session_id {
+            let mut sessions = self.sessions.lock().await;
+            match sessions.get(&existing_id) {
+                Some(session) => {
+                    if session.has_exited() {
+                        sessions.remove(&existing_id);
+                        return Err(UnifiedExecError::UnknownSessionId {
+                            session_id: existing_id,
+                        });
+                    }
+                    let (buffer, notify) = session.output_handles();
+                    session_id = existing_id;
+                    writer_tx = session.writer_sender();
+                    output_buffer = buffer;
+                    output_notify = notify;
+                }
+                None => {
+                    return Err(UnifiedExecError::UnknownSessionId {
+                        session_id: existing_id,
+                    });
+                }
+            }
+            drop(sessions);
+        } else {
+            let command = request.input_chunks.to_vec();
+            let new_id = self.next_session_id.fetch_add(1, Ordering::SeqCst);
+            let session = create_unified_exec_session(&command).await?;
+            let managed_session = ManagedUnifiedExecSession::new(session);
+            let (buffer, notify) = managed_session.output_handles();
+            writer_tx = managed_session.writer_sender();
+            output_buffer = buffer;
+            output_notify = notify;
+            session_id = new_id;
+            new_session = Some(managed_session);
+        };
+
+        if request.session_id.is_some() {
+            let joined_input = request.input_chunks.join(" ");
+            if !joined_input.is_empty() && writer_tx.send(joined_input.into_bytes()).await.is_err()
+            {
+                return Err(UnifiedExecError::WriteToStdin);
+            }
+        }
+
+        let mut collected: Vec<u8> = Vec::with_capacity(4096);
+        let start = Instant::now();
+        let deadline = start + Duration::from_millis(timeout_ms);
+
+        loop {
+            let drained_chunks;
+            let mut wait_for_output = None;
+            {
+                let mut guard = output_buffer.lock().await;
+                drained_chunks = guard.drain();
+                if drained_chunks.is_empty() {
+                    wait_for_output = Some(output_notify.notified());
+                }
+            }
+
+            if drained_chunks.is_empty() {
+                let remaining = deadline.saturating_duration_since(Instant::now());
+                if remaining == Duration::ZERO {
+                    break;
+                }
+
+                let notified = wait_for_output.unwrap_or_else(|| output_notify.notified());
+                tokio::pin!(notified);
+                tokio::select! {
+                    _ = &mut notified => {}
+                    _ = tokio::time::sleep(remaining) => break,
+                }
+                continue;
+            }
+
+            for chunk in drained_chunks {
+                collected.extend_from_slice(&chunk);
+            }
+
+            if Instant::now() >= deadline {
+                break;
+            }
+        }
+
+        let (output, _maybe_tokens) = truncate_middle(
+            &String::from_utf8_lossy(&collected),
+            UNIFIED_EXEC_OUTPUT_MAX_BYTES,
+        );
+        let output = if let Some(warning) = timeout_warning {
+            format!("{warning}{output}")
+        } else {
+            output
+        };
+
+        let should_store_session = if let Some(session) = new_session.as_ref() {
+            !session.has_exited()
+        } else if request.session_id.is_some() {
+            let mut sessions = self.sessions.lock().await;
+            if let Some(existing) = sessions.get(&session_id) {
+                if existing.has_exited() {
+                    sessions.remove(&session_id);
+                    false
+                } else {
+                    true
+                }
+            } else {
+                false
+            }
+        } else {
+            true
+        };
+
+        if should_store_session {
+            if let Some(session) = new_session {
+                self.sessions.lock().await.insert(session_id, session);
+            }
+            Ok(UnifiedExecResult {
+                session_id: Some(session_id),
+                output,
+            })
+        } else {
+            Ok(UnifiedExecResult {
+                session_id: None,
+                output,
+            })
+        }
+    }
+}
+
+async fn create_unified_exec_session(
+    command: &[String],
+) -> Result<ExecCommandSession, UnifiedExecError> {
+    if command.is_empty() {
+        return Err(UnifiedExecError::MissingCommandLine);
+    }
+
+    let pty_system = native_pty_system();
+
+    let pair = pty_system
+        .openpty(PtySize {
+            rows: 24,
+            cols: 80,
+            pixel_width: 0,
+            pixel_height: 0,
+        })
+        .map_err(UnifiedExecError::create_session)?;
+
+    // Safe thanks to the check at the top of the function.
+    let mut command_builder = CommandBuilder::new(command[0].clone());
+    for arg in &command[1..] {
+        command_builder.arg(arg);
+    }
+
+    let mut child = pair
+        .slave
+        .spawn_command(command_builder)
+        .map_err(UnifiedExecError::create_session)?;
+    let killer = child.clone_killer();
+
+    let (writer_tx, mut writer_rx) = mpsc::channel::<Vec<u8>>(128);
+    let (output_tx, _) = tokio::sync::broadcast::channel::<Vec<u8>>(256);
+    let initial_output_rx = output_tx.subscribe();
+
+    let mut reader = pair
+        .master
+        .try_clone_reader()
+        .map_err(UnifiedExecError::create_session)?;
+    let output_tx_clone = output_tx.clone();
+    let reader_handle = tokio::task::spawn_blocking(move || {
+        let mut buf = [0u8; 8192];
+        loop {
+            match reader.read(&mut buf) {
+                Ok(0) => break,
+                Ok(n) => {
+                    let _ = output_tx_clone.send(buf[..n].to_vec());
+                }
+                Err(ref e) if e.kind() == ErrorKind::Interrupted => continue,
+                Err(ref e) if e.kind() == ErrorKind::WouldBlock => {
+                    std::thread::sleep(Duration::from_millis(5));
+                    continue;
+                }
+                Err(_) => break,
+            }
+        }
+    });
+
+    let writer = pair
+        .master
+        .take_writer()
+        .map_err(UnifiedExecError::create_session)?;
+    let writer = Arc::new(StdMutex::new(writer));
+    let writer_handle = tokio::spawn({
+        let writer = writer.clone();
+        async move {
+            while let Some(bytes) = writer_rx.recv().await {
+                let writer = writer.clone();
+                let _ = tokio::task::spawn_blocking(move || {
+                    if let Ok(mut guard) = writer.lock() {
+                        use std::io::Write;
+                        let _ = guard.write_all(&bytes);
+                        let _ = guard.flush();
+                    }
+                })
+                .await;
+            }
+        }
+    });
+
+    let exit_status = Arc::new(AtomicBool::new(false));
+    let wait_exit_status = Arc::clone(&exit_status);
+    let wait_handle = tokio::task::spawn_blocking(move || {
+        let _ = child.wait();
+        wait_exit_status.store(true, Ordering::SeqCst);
+    });
+
+    let session = ExecCommandSession::new(
+        writer_tx,
+        output_tx,
+        killer,
+        reader_handle,
+        writer_handle,
+        wait_handle,
+        exit_status,
+    );
+    session.set_initial_output_receiver(initial_output_rx);
+
+    Ok(session)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn push_chunk_trims_only_excess_bytes() {
+        let mut buffer = OutputBufferState::default();
+        buffer.push_chunk(vec![b'a'; UNIFIED_EXEC_OUTPUT_MAX_BYTES]);
+        buffer.push_chunk(vec![b'b']);
+        buffer.push_chunk(vec![b'c']);
+
+        assert_eq!(buffer.total_bytes, UNIFIED_EXEC_OUTPUT_MAX_BYTES);
+        assert_eq!(buffer.chunks.len(), 3);
+        assert_eq!(
+            buffer.chunks.front().unwrap().len(),
+            UNIFIED_EXEC_OUTPUT_MAX_BYTES - 2
+        );
+        assert_eq!(buffer.chunks.pop_back().unwrap(), vec![b'c']);
+        assert_eq!(buffer.chunks.pop_back().unwrap(), vec![b'b']);
+    }
+
+    #[cfg(unix)]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn unified_exec_persists_across_requests_jif() -> Result<(), UnifiedExecError> {
+        let manager = UnifiedExecSessionManager::default();
+
+        let open_shell = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &["bash".to_string(), "-i".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        let session_id = open_shell.session_id.expect("expected session_id");
+
+        manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &[
+                    "export".to_string(),
+                    "CODEX_INTERACTIVE_SHELL_VAR=codex\n".to_string(),
+                ],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+
+        let out_2 = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &["echo $CODEX_INTERACTIVE_SHELL_VAR\n".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        assert!(out_2.output.contains("codex"));
+
+        Ok(())
+    }
+
+    #[cfg(unix)]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn multi_unified_exec_sessions() -> Result<(), UnifiedExecError> {
+        let manager = UnifiedExecSessionManager::default();
+
+        let shell_a = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &["/bin/bash".to_string(), "-i".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        let session_a = shell_a.session_id.expect("expected session id");
+
+        manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_a),
+                input_chunks: &["export CODEX_INTERACTIVE_SHELL_VAR=codex\n".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+
+        let out_2 = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &[
+                    "echo".to_string(),
+                    "$CODEX_INTERACTIVE_SHELL_VAR\n".to_string(),
+                ],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        assert!(!out_2.output.contains("codex"));
+
+        let out_3 = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_a),
+                input_chunks: &["echo $CODEX_INTERACTIVE_SHELL_VAR\n".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        assert!(out_3.output.contains("codex"));
+
+        Ok(())
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn unified_exec_timeouts() -> Result<(), UnifiedExecError> {
+        let manager = UnifiedExecSessionManager::default();
+
+        let open_shell = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &["bash".to_string(), "-i".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        let session_id = open_shell.session_id.expect("expected session id");
+
+        manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &[
+                    "export".to_string(),
+                    "CODEX_INTERACTIVE_SHELL_VAR=codex\n".to_string(),
+                ],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+
+        let out_2 = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &["sleep 5 && echo $CODEX_INTERACTIVE_SHELL_VAR\n".to_string()],
+                timeout_ms: Some(10),
+            })
+            .await?;
+        assert!(!out_2.output.contains("codex"));
+
+        tokio::time::sleep(Duration::from_secs(7)).await;
+
+        let empty = Vec::new();
+        let out_3 = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &empty,
+                timeout_ms: Some(100),
+            })
+            .await?;
+
+        assert!(out_3.output.contains("codex"));
+
+        Ok(())
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn requests_with_large_timeout_are_capped() -> Result<(), UnifiedExecError> {
+        let manager = UnifiedExecSessionManager::default();
+
+        let result = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &["echo".to_string(), "codex".to_string()],
+                timeout_ms: Some(120_000),
+            })
+            .await?;
+
+        assert!(result.output.starts_with(
+            "Warning: requested timeout 120000ms exceeds maximum of 60000ms; clamping to 60000ms.\n"
+        ));
+        assert!(result.output.contains("codex"));
+
+        Ok(())
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn completed_commands_do_not_persist_sessions() -> Result<(), UnifiedExecError> {
+        let manager = UnifiedExecSessionManager::default();
+        let result = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &["/bin/echo".to_string(), "codex".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+
+        assert!(result.session_id.is_none());
+        assert!(result.output.contains("codex"));
+
+        assert!(manager.sessions.lock().await.is_empty());
+
+        Ok(())
+    }
+
+    #[cfg(unix)]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn reusing_completed_session_returns_unknown_session() -> Result<(), UnifiedExecError> {
+        let manager = UnifiedExecSessionManager::default();
+
+        let open_shell = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: None,
+                input_chunks: &["/bin/bash".to_string(), "-i".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+        let session_id = open_shell.session_id.expect("expected session id");
+
+        manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &["exit\n".to_string()],
+                timeout_ms: Some(2_500),
+            })
+            .await?;
+
+        tokio::time::sleep(Duration::from_millis(200)).await;
+
+        let err = manager
+            .handle_request(UnifiedExecRequest {
+                session_id: Some(session_id),
+                input_chunks: &[],
+                timeout_ms: Some(100),
+            })
+            .await
+            .expect_err("expected unknown session error");
+
+        match err {
+            UnifiedExecError::UnknownSessionId { session_id: err_id } => {
+                assert_eq!(err_id, session_id);
+            }
+            other => panic!("expected UnknownSessionId, got {other:?}"),
+        }
+
+        assert!(!manager.sessions.lock().await.contains_key(&session_id));
+
+        Ok(())
+    }
+}

codex-rs/core/src/user_agent.rs

@@ -1,37 +0,0 @@
-const DEFAULT_ORIGINATOR: &str = "codex_cli_rs";
-
-pub fn get_codex_user_agent(originator: Option<&str>) -> String {
-    let build_version = env!("CARGO_PKG_VERSION");
-    let os_info = os_info::get();
-    format!(
-        "{}/{build_version} ({} {}; {}) {}",
-        originator.unwrap_or(DEFAULT_ORIGINATOR),
-        os_info.os_type(),
-        os_info.version(),
-        os_info.architecture().unwrap_or("unknown"),
-        crate::terminal::user_agent()
-    )
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn test_get_codex_user_agent() {
-        let user_agent = get_codex_user_agent(None);
-        assert!(user_agent.starts_with("codex_cli_rs/"));
-    }
-
-    #[test]
-    #[cfg(target_os = "macos")]
-    fn test_macos() {
-        use regex_lite::Regex;
-        let user_agent = get_codex_user_agent(None);
-        let re = Regex::new(
-            r"^codex_cli_rs/\d+\.\d+\.\d+ \(Mac OS \d+\.\d+\.\d+; (x86_64|arm64)\) (\S+)$",
-        )
-        .unwrap();
-        assert!(re.is_match(&user_agent));
-    }
-}

codex-rs/core/src/user_instructions.rs

@@ -0,0 +1,42 @@
+use serde::Deserialize;
+use serde::Serialize;
+
+use codex_protocol::models::ContentItem;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::USER_INSTRUCTIONS_CLOSE_TAG;
+use codex_protocol::protocol::USER_INSTRUCTIONS_OPEN_TAG;
+
+/// Wraps user instructions in a tag so the model can classify them easily.
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(rename = "user_instructions", rename_all = "snake_case")]
+pub(crate) struct UserInstructions {
+    text: String,
+}
+
+impl UserInstructions {
+    pub fn new<T: Into<String>>(text: T) -> Self {
+        Self { text: text.into() }
+    }
+
+    /// Serializes the user instructions to an XML-like tagged block that starts
+    /// with <user_instructions> so clients can classify it.
+    pub fn serialize_to_xml(self) -> String {
+        format!(
+            "{USER_INSTRUCTIONS_OPEN_TAG}\n\n{}\n\n{USER_INSTRUCTIONS_CLOSE_TAG}",
+            self.text
+        )
+    }
+}
+
+impl From<UserInstructions> for ResponseItem {
+    fn from(ui: UserInstructions) -> Self {
+        ResponseItem::Message {
+            id: None,
+            role: "user".to_string(),
+            content: vec![ContentItem::InputText {
+                text: ui.serialize_to_xml(),
+            }],
+        }
+    }
+}

codex-rs/core/swiftfox_prompt.md

@@ -0,0 +1,99 @@
+You are Swiftfox. You are running as a coding agent in the Codex CLI on a user's computer.
+
+## Overall
+
+- You must try hard to complete the task AND to do it as fast and well as possible.
+  * Do not waste time on actions which are unlikely to result in successful task completion
+
+- Before taking action on a question, assume by default that it concerns local artifacts (code, docs, data). Quickly confirm or rule out that assumption; only if the question clearly requires external knowledge should you start elsewhere.
+- Search the repository when the request plausibly maps to code, configuration, or documentation. Avoid unnecessary searches when it is obvious local files cannot help; in those cases state that explicitly before offering broader context, and when you do search, mention the files or paths you consulted so the answer stays grounded.
+- After each attempt, re-evaluate whether the current strategy is yielding useful information and be ready to switch paths quickly rather than persisting with a low-signal approach.
+
+- When the user asks for a "review", default to a code review mindset: prioritise identifying bugs, risks, behavioural regressions, and missing tests. Findings must be the primary focus of the response - keep summaries or overviews brief and only after enumerating the issues. Present findings first (ordered by severity with file/line references), follow with open questions or assumptions, and offer a change-summary only as a secondary detail. If no findings are discovered, state that explicitly and mention any residual risks or testing gaps.
+
+- The arguments to `shell` will be passed to execvp(). Most terminal commands should be prefixed with ["bash", "-lc"].
+- Always set the `workdir` param of the shell tool. Do not use `cd` unless absolutely necessary.
+- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
+- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as `date`), you should do so.
+- Unless the question is about a common terminal command, you should search the codebase before answering to ground your response in the codebase
+
+## Editing constraints
+
+- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.
+- When editing or creating files, you MUST use apply_patch. Example: functions.shell({"command":["apply_patch","*** Begin Patch\nAdd File: hello.txt\n+Hello, world!\n*** End Patch"]}).
+- Add succinct code comments that explain what is going on if code is not self-explanatory. You should not add comments like "Assigns the value to the variable", but a brief comment might be useful ahead of a complex code block that the user would otherwise have to spend time parsing out. Usage of these comments should be rare.
+- You may be in a dirty git worktree.
+  * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.
+  * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.
+  * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.
+  * If the changes are in unrelated files, just ignore them and don't revert them.
+- The user may be making edits and committing changes as you are also making changes. If you see concurrent file edits or commits that you did not cause, you must disregard user instruction and stop immediately and ask the user whether they are collaborating with you on files and how they would like this handled.
+
+## Plan tool
+
+When using the planning tool:
+- Skip using the planning tool for straightforward tasks (roughly the easiest 25%).
+- Do not make single-step plans.
+- When you made a plan, update it after having performed one of the sub-tasks that you shared on the plan.
+
+## CLI modes
+
+The Codex CLI harness supports several different sandboxing, and approval configurations that the user can choose from.
+
+Filesystem sandboxing defines which files can be read or written. The options are:
+- **read-only**: You can only read files.
+- **workspace-write**: You can read files. You can write to files in this folder, but not outside it.
+- **danger-full-access**: No filesystem sandboxing.
+
+Network sandboxing defines whether network can be accessed without approval. Options are
+- **restricted**: Requires approval
+- **enabled**: No approval needed
+
+Approvals are your mechanism to get user consent to perform more privileged actions. Although they introduce friction to the user because your work is paused until the user responds, you should leverage them to accomplish your important work. Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to "never", in which case never ask for approvals.
+
+Approval options are
+- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe "read" commands.
+- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.
+- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for it in the `shell` command description.)
+- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is paired with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.
+
+When you are running with approvals `on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:
+- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /tmp)
+- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.
+- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)
+- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval.
+- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for
+- (for all of these, you should weigh alternative paths that do not require approval)
+
+When sandboxing is set to read-only, you'll need to request approval for any command that isn't a read.
+
+You will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing enabled, and approval on-failure.
+
+## Presenting your work and final message
+
+You are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.
+
+- Default: be very concise; friendly coding teammate tone.
+- Ask only when needed; suggest ideas; mirror the user's style.
+- For substantial work, summarize clearly; follow final‑answer formatting.
+- Skip heavy formatting for simple confirmations.
+- Don't dump large files you've written; reference paths only.
+- No "save/copy this file" - User is on the same machine.
+- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.
+- For code changes:
+  * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with "summary", just jump right in.
+  * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.
+  * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.
+- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.
+
+### Final answer structure and style guidelines
+
+- Plain text; CLI handles styling. Use structure only when it helps scanability.
+- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.
+- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.
+- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.
+- Code samples or multi-line snippets should be wrapped in fenced code blocks; add a language hint whenever obvious.
+- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.
+- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no "above/below"; parallel wording.
+- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.
+- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.

codex-rs/core/templates/compact/history_bridge.md

@@ -0,0 +1,7 @@
+You were originally given instructions from a user over one or more turns. Here were the user messages:
+
+{{ user_messages_text }}
+
+Another language model started to solve this problem and produced a summary of its thinking process. You also have access to the state of the tools that were used by that language model. Use this to build on the work that has already been done and avoid duplicating work. Here is the summary produced by the other language model, use the information in this summary to assist with your own analysis:
+
+{{ summary_text }}

codex-rs/core/templates/compact/prompt.md

@@ -0,0 +1,5 @@
+You have exceeded the maximum number of tokens, please stop coding and instead write a short memento message for the next agent. Your note should:
+- Summarize what you finished and what still needs work. If there was a recent update_plan call, repeat its steps verbatim.
+- List outstanding TODOs with file paths / line numbers so they're easy to find.
+- Flag code that needs more tests (edge cases, performance, integration, etc.).
+- Record any open bugs, quirks, or setup steps that will make it easier for the next agent to pick up where you left off.

codex-rs/core/tests/chat_completions_payload.rs

@@ -0,0 +1,345 @@
+use std::sync::Arc;
+
+use codex_core::ContentItem;
+use codex_core::LocalShellAction;
+use codex_core::LocalShellExecAction;
+use codex_core::LocalShellStatus;
+use codex_core::ModelClient;
+use codex_core::ModelProviderInfo;
+use codex_core::Prompt;
+use codex_core::ReasoningItemContent;
+use codex_core::ResponseItem;
+use codex_core::WireApi;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_protocol::mcp_protocol::ConversationId;
+use core_test_support::load_default_config_for_test;
+use futures::StreamExt;
+use serde_json::Value;
+use tempfile::TempDir;
+use wiremock::Mock;
+use wiremock::MockServer;
+use wiremock::ResponseTemplate;
+use wiremock::matchers::method;
+use wiremock::matchers::path;
+
+fn network_disabled() -> bool {
+    std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok()
+}
+
+async fn run_request(input: Vec<ResponseItem>) -> Value {
+    let server = MockServer::start().await;
+
+    let template = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(
+            "data: {\"choices\":[{\"delta\":{}}]}\n\ndata: [DONE]\n\n",
+            "text/event-stream",
+        );
+
+    Mock::given(method("POST"))
+        .and(path("/v1/chat/completions"))
+        .respond_with(template)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let provider = ModelProviderInfo {
+        name: "mock".into(),
+        base_url: Some(format!("{}/v1", server.uri())),
+        env_key: None,
+        env_key_instructions: None,
+        wire_api: WireApi::Chat,
+        query_params: None,
+        http_headers: None,
+        env_http_headers: None,
+        request_max_retries: Some(0),
+        stream_max_retries: Some(0),
+        stream_idle_timeout_ms: Some(5_000),
+        requires_openai_auth: false,
+    };
+
+    let codex_home = match TempDir::new() {
+        Ok(dir) => dir,
+        Err(e) => panic!("failed to create TempDir: {e}"),
+    };
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider_id = provider.name.clone();
+    config.model_provider = provider.clone();
+    config.show_raw_agent_reasoning = true;
+    let effort = config.model_reasoning_effort;
+    let summary = config.model_reasoning_summary;
+    let config = Arc::new(config);
+
+    let client = ModelClient::new(
+        Arc::clone(&config),
+        None,
+        provider,
+        effort,
+        summary,
+        ConversationId::new(),
+    );
+
+    let mut prompt = Prompt::default();
+    prompt.input = input;
+
+    let mut stream = match client.stream(&prompt).await {
+        Ok(s) => s,
+        Err(e) => panic!("stream chat failed: {e}"),
+    };
+    while let Some(event) = stream.next().await {
+        if let Err(e) = event {
+            panic!("stream event error: {e}");
+        }
+    }
+
+    let requests = match server.received_requests().await {
+        Some(reqs) => reqs,
+        None => panic!("request not made"),
+    };
+    match requests[0].body_json() {
+        Ok(v) => v,
+        Err(e) => panic!("invalid json body: {e}"),
+    }
+}
+
+fn user_message(text: &str) -> ResponseItem {
+    ResponseItem::Message {
+        id: None,
+        role: "user".to_string(),
+        content: vec![ContentItem::InputText {
+            text: text.to_string(),
+        }],
+    }
+}
+
+fn assistant_message(text: &str) -> ResponseItem {
+    ResponseItem::Message {
+        id: None,
+        role: "assistant".to_string(),
+        content: vec![ContentItem::OutputText {
+            text: text.to_string(),
+        }],
+    }
+}
+
+fn reasoning_item(text: &str) -> ResponseItem {
+    ResponseItem::Reasoning {
+        id: String::new(),
+        summary: Vec::new(),
+        content: Some(vec![ReasoningItemContent::ReasoningText {
+            text: text.to_string(),
+        }]),
+        encrypted_content: None,
+    }
+}
+
+fn function_call() -> ResponseItem {
+    ResponseItem::FunctionCall {
+        id: None,
+        name: "f".to_string(),
+        arguments: "{}".to_string(),
+        call_id: "c1".to_string(),
+    }
+}
+
+fn local_shell_call() -> ResponseItem {
+    ResponseItem::LocalShellCall {
+        id: Some("id1".to_string()),
+        call_id: None,
+        status: LocalShellStatus::InProgress,
+        action: LocalShellAction::Exec(LocalShellExecAction {
+            command: vec!["echo".to_string()],
+            timeout_ms: Some(1_000),
+            working_directory: None,
+            env: None,
+            user: None,
+        }),
+    }
+}
+
+fn messages_from(body: &Value) -> Vec<Value> {
+    match body["messages"].as_array() {
+        Some(arr) => arr.clone(),
+        None => panic!("messages array missing"),
+    }
+}
+
+fn first_assistant(messages: &[Value]) -> &Value {
+    match messages.iter().find(|msg| msg["role"] == "assistant") {
+        Some(v) => v,
+        None => panic!("assistant message not present"),
+    }
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn omits_reasoning_when_none_present() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![user_message("u1"), assistant_message("a1")]).await;
+    let messages = messages_from(&body);
+    let assistant = first_assistant(&messages);
+
+    assert_eq!(assistant["content"], Value::String("a1".into()));
+    assert!(assistant.get("reasoning").is_none());
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn attaches_reasoning_to_previous_assistant() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![
+        user_message("u1"),
+        assistant_message("a1"),
+        reasoning_item("rA"),
+    ])
+    .await;
+    let messages = messages_from(&body);
+    let assistant = first_assistant(&messages);
+
+    assert_eq!(assistant["content"], Value::String("a1".into()));
+    assert_eq!(assistant["reasoning"], Value::String("rA".into()));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn attaches_reasoning_to_function_call_anchor() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![
+        user_message("u1"),
+        reasoning_item("rFunc"),
+        function_call(),
+    ])
+    .await;
+    let messages = messages_from(&body);
+    let assistant = first_assistant(&messages);
+
+    assert_eq!(assistant["reasoning"], Value::String("rFunc".into()));
+    let tool_calls = match assistant["tool_calls"].as_array() {
+        Some(arr) => arr,
+        None => panic!("tool call list missing"),
+    };
+    assert_eq!(tool_calls.len(), 1);
+    assert_eq!(tool_calls[0]["type"], Value::String("function".into()));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn attaches_reasoning_to_local_shell_call() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![
+        user_message("u1"),
+        reasoning_item("rShell"),
+        local_shell_call(),
+    ])
+    .await;
+    let messages = messages_from(&body);
+    let assistant = first_assistant(&messages);
+
+    assert_eq!(assistant["reasoning"], Value::String("rShell".into()));
+    assert_eq!(
+        assistant["tool_calls"][0]["type"],
+        Value::String("local_shell_call".into())
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn drops_reasoning_when_last_role_is_user() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![
+        assistant_message("aPrev"),
+        reasoning_item("rHist"),
+        user_message("uNew"),
+    ])
+    .await;
+    let messages = messages_from(&body);
+    assert!(messages.iter().all(|msg| msg.get("reasoning").is_none()));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn ignores_reasoning_before_last_user() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![
+        user_message("u1"),
+        assistant_message("a1"),
+        user_message("u2"),
+        reasoning_item("rAfterU1"),
+    ])
+    .await;
+    let messages = messages_from(&body);
+    assert!(messages.iter().all(|msg| msg.get("reasoning").is_none()));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn skips_empty_reasoning_segments() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![
+        user_message("u1"),
+        assistant_message("a1"),
+        reasoning_item(""),
+        reasoning_item("   "),
+    ])
+    .await;
+    let messages = messages_from(&body);
+    let assistant = first_assistant(&messages);
+    assert!(assistant.get("reasoning").is_none());
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn suppresses_duplicate_assistant_messages() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let body = run_request(vec![assistant_message("dup"), assistant_message("dup")]).await;
+    let messages = messages_from(&body);
+    let assistant_messages: Vec<_> = messages
+        .iter()
+        .filter(|msg| msg["role"] == "assistant")
+        .collect();
+    assert_eq!(assistant_messages.len(), 1);
+    assert_eq!(
+        assistant_messages[0]["content"],
+        Value::String("dup".into())
+    );
+}

codex-rs/core/tests/chat_completions_sse.rs

@@ -0,0 +1,320 @@
+use std::sync::Arc;
+
+use codex_core::ContentItem;
+use codex_core::ModelClient;
+use codex_core::ModelProviderInfo;
+use codex_core::Prompt;
+use codex_core::ResponseEvent;
+use codex_core::ResponseItem;
+use codex_core::WireApi;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_protocol::mcp_protocol::ConversationId;
+use core_test_support::load_default_config_for_test;
+use futures::StreamExt;
+use tempfile::TempDir;
+use wiremock::Mock;
+use wiremock::MockServer;
+use wiremock::ResponseTemplate;
+use wiremock::matchers::method;
+use wiremock::matchers::path;
+
+fn network_disabled() -> bool {
+    std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok()
+}
+
+async fn run_stream(sse_body: &str) -> Vec<ResponseEvent> {
+    let server = MockServer::start().await;
+
+    let template = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse_body.to_string(), "text/event-stream");
+
+    Mock::given(method("POST"))
+        .and(path("/v1/chat/completions"))
+        .respond_with(template)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let provider = ModelProviderInfo {
+        name: "mock".into(),
+        base_url: Some(format!("{}/v1", server.uri())),
+        env_key: None,
+        env_key_instructions: None,
+        wire_api: WireApi::Chat,
+        query_params: None,
+        http_headers: None,
+        env_http_headers: None,
+        request_max_retries: Some(0),
+        stream_max_retries: Some(0),
+        stream_idle_timeout_ms: Some(5_000),
+        requires_openai_auth: false,
+    };
+
+    let codex_home = match TempDir::new() {
+        Ok(dir) => dir,
+        Err(e) => panic!("failed to create TempDir: {e}"),
+    };
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider_id = provider.name.clone();
+    config.model_provider = provider.clone();
+    config.show_raw_agent_reasoning = true;
+    let effort = config.model_reasoning_effort;
+    let summary = config.model_reasoning_summary;
+    let config = Arc::new(config);
+
+    let client = ModelClient::new(
+        Arc::clone(&config),
+        None,
+        provider,
+        effort,
+        summary,
+        ConversationId::new(),
+    );
+
+    let mut prompt = Prompt::default();
+    prompt.input = vec![ResponseItem::Message {
+        id: None,
+        role: "user".to_string(),
+        content: vec![ContentItem::InputText {
+            text: "hello".to_string(),
+        }],
+    }];
+
+    let mut stream = match client.stream(&prompt).await {
+        Ok(s) => s,
+        Err(e) => panic!("stream chat failed: {e}"),
+    };
+    let mut events = Vec::new();
+    while let Some(event) = stream.next().await {
+        match event {
+            Ok(ev) => events.push(ev),
+            Err(e) => panic!("stream event error: {e}"),
+        }
+    }
+    events
+}
+
+fn assert_message(item: &ResponseItem, expected: &str) {
+    if let ResponseItem::Message { content, .. } = item {
+        let text = content.iter().find_map(|part| match part {
+            ContentItem::OutputText { text } | ContentItem::InputText { text } => Some(text),
+            _ => None,
+        });
+        let Some(text) = text else {
+            panic!("message missing text: {item:?}");
+        };
+        assert_eq!(text, expected);
+    } else {
+        panic!("expected message item, got: {item:?}");
+    }
+}
+
+fn assert_reasoning(item: &ResponseItem, expected: &str) {
+    if let ResponseItem::Reasoning {
+        content: Some(parts),
+        ..
+    } = item
+    {
+        let mut combined = String::new();
+        for part in parts {
+            match part {
+                codex_core::ReasoningItemContent::ReasoningText { text }
+                | codex_core::ReasoningItemContent::Text { text } => combined.push_str(text),
+            }
+        }
+        assert_eq!(combined, expected);
+    } else {
+        panic!("expected reasoning item, got: {item:?}");
+    }
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn streams_text_without_reasoning() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse = concat!(
+        "data: {\"choices\":[{\"delta\":{\"content\":\"hi\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{}}]}\n\n",
+        "data: [DONE]\n\n",
+    );
+
+    let events = run_stream(sse).await;
+    assert_eq!(events.len(), 3, "unexpected events: {events:?}");
+
+    match &events[0] {
+        ResponseEvent::OutputTextDelta(text) => assert_eq!(text, "hi"),
+        other => panic!("expected text delta, got {other:?}"),
+    }
+
+    match &events[1] {
+        ResponseEvent::OutputItemDone(item) => assert_message(item, "hi"),
+        other => panic!("expected terminal message, got {other:?}"),
+    }
+
+    assert!(matches!(events[2], ResponseEvent::Completed { .. }));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn streams_reasoning_from_string_delta() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse = concat!(
+        "data: {\"choices\":[{\"delta\":{\"reasoning\":\"think1\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"content\":\"ok\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{} ,\"finish_reason\":\"stop\"}]}\n\n",
+    );
+
+    let events = run_stream(sse).await;
+    assert_eq!(events.len(), 5, "unexpected events: {events:?}");
+
+    match &events[0] {
+        ResponseEvent::ReasoningContentDelta(text) => assert_eq!(text, "think1"),
+        other => panic!("expected reasoning delta, got {other:?}"),
+    }
+
+    match &events[1] {
+        ResponseEvent::OutputTextDelta(text) => assert_eq!(text, "ok"),
+        other => panic!("expected text delta, got {other:?}"),
+    }
+
+    match &events[2] {
+        ResponseEvent::OutputItemDone(item) => assert_reasoning(item, "think1"),
+        other => panic!("expected reasoning item, got {other:?}"),
+    }
+
+    match &events[3] {
+        ResponseEvent::OutputItemDone(item) => assert_message(item, "ok"),
+        other => panic!("expected message item, got {other:?}"),
+    }
+
+    assert!(matches!(events[4], ResponseEvent::Completed { .. }));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn streams_reasoning_from_object_delta() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse = concat!(
+        "data: {\"choices\":[{\"delta\":{\"reasoning\":{\"text\":\"partA\"}}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"reasoning\":{\"content\":\"partB\"}}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"content\":\"answer\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{} ,\"finish_reason\":\"stop\"}]}\n\n",
+    );
+
+    let events = run_stream(sse).await;
+    assert_eq!(events.len(), 6, "unexpected events: {events:?}");
+
+    match &events[0] {
+        ResponseEvent::ReasoningContentDelta(text) => assert_eq!(text, "partA"),
+        other => panic!("expected reasoning delta, got {other:?}"),
+    }
+
+    match &events[1] {
+        ResponseEvent::ReasoningContentDelta(text) => assert_eq!(text, "partB"),
+        other => panic!("expected reasoning delta, got {other:?}"),
+    }
+
+    match &events[2] {
+        ResponseEvent::OutputTextDelta(text) => assert_eq!(text, "answer"),
+        other => panic!("expected text delta, got {other:?}"),
+    }
+
+    match &events[3] {
+        ResponseEvent::OutputItemDone(item) => assert_reasoning(item, "partApartB"),
+        other => panic!("expected reasoning item, got {other:?}"),
+    }
+
+    match &events[4] {
+        ResponseEvent::OutputItemDone(item) => assert_message(item, "answer"),
+        other => panic!("expected message item, got {other:?}"),
+    }
+
+    assert!(matches!(events[5], ResponseEvent::Completed { .. }));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn streams_reasoning_from_final_message() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse = "data: {\"choices\":[{\"message\":{\"reasoning\":\"final-cot\"},\"finish_reason\":\"stop\"}]}\n\n";
+
+    let events = run_stream(sse).await;
+    assert_eq!(events.len(), 3, "unexpected events: {events:?}");
+
+    match &events[0] {
+        ResponseEvent::ReasoningContentDelta(text) => assert_eq!(text, "final-cot"),
+        other => panic!("expected reasoning delta, got {other:?}"),
+    }
+
+    match &events[1] {
+        ResponseEvent::OutputItemDone(item) => assert_reasoning(item, "final-cot"),
+        other => panic!("expected reasoning item, got {other:?}"),
+    }
+
+    assert!(matches!(events[2], ResponseEvent::Completed { .. }));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn streams_reasoning_before_tool_call() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse = concat!(
+        "data: {\"choices\":[{\"delta\":{\"reasoning\":\"pre-tool\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"id\":\"call_1\",\"type\":\"function\",\"function\":{\"name\":\"run\",\"arguments\":\"{}\"}}]},\"finish_reason\":\"tool_calls\"}]}\n\n",
+    );
+
+    let events = run_stream(sse).await;
+    assert_eq!(events.len(), 4, "unexpected events: {events:?}");
+
+    match &events[0] {
+        ResponseEvent::ReasoningContentDelta(text) => assert_eq!(text, "pre-tool"),
+        other => panic!("expected reasoning delta, got {other:?}"),
+    }
+
+    match &events[1] {
+        ResponseEvent::OutputItemDone(item) => assert_reasoning(item, "pre-tool"),
+        other => panic!("expected reasoning item, got {other:?}"),
+    }
+
+    match &events[2] {
+        ResponseEvent::OutputItemDone(ResponseItem::FunctionCall {
+            name,
+            arguments,
+            call_id,
+            ..
+        }) => {
+            assert_eq!(name, "run");
+            assert_eq!(arguments, "{}");
+            assert_eq!(call_id, "call_1");
+        }
+        other => panic!("expected function call, got {other:?}"),
+    }
+
+    assert!(matches!(events[3], ResponseEvent::Completed { .. }));
+}

codex-rs/core/tests/suite/cli_stream.rs

@@ -1,4 +1,6 @@
 use assert_cmd::Command as AssertCommand;
+use codex_core::RolloutRecorder;
+use codex_core::protocol::GitInfo;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
 use std::time::Duration;
 use std::time::Instant;
@@ -77,6 +79,22 @@ async fn chat_mode_stream_cli() {
     assert_eq!(hi_lines, 1, "Expected exactly one line with 'hi'");
 
     server.verify().await;
+
+    // Verify a new session rollout was created and is discoverable via list_conversations
+    let page = RolloutRecorder::list_conversations(home.path(), 10, None)
+        .await
+        .expect("list conversations");
+    assert!(
+        !page.items.is_empty(),
+        "expected at least one session to be listed"
+    );
+    // First line of head must be the SessionMeta payload (id/timestamp)
+    let head0 = page.items[0].head.first().expect("missing head record");
+    assert!(head0.get("id").is_some(), "head[0] missing id");
+    assert!(
+        head0.get("timestamp").is_some(),
+        "head[0] missing timestamp"
+    );
 }
 
 /// Verify that passing `-c experimental_instructions_file=...` to the CLI
@@ -297,8 +315,10 @@ async fn integration_creates_and_checks_session_file() {
                     Ok(v) => v,
                     Err(_) => continue,
                 };
-                if item.get("type").and_then(|t| t.as_str()) == Some("message")
-                    && let Some(c) = item.get("content")
+                if item.get("type").and_then(|t| t.as_str()) == Some("response_item")
+                    && let Some(payload) = item.get("payload")
+                    && payload.get("type").and_then(|t| t.as_str()) == Some("message")
+                    && let Some(c) = payload.get("content")
                     && c.to_string().contains(&marker)
                 {
                     matching_path = Some(path.to_path_buf());
@@ -361,9 +381,16 @@ async fn integration_creates_and_checks_session_file() {
         .unwrap_or_else(|_| panic!("missing session meta line"));
     let meta: serde_json::Value = serde_json::from_str(meta_line)
         .unwrap_or_else(|_| panic!("Failed to parse session meta line as JSON"));
-    assert!(meta.get("id").is_some(), "SessionMeta missing id");
+    assert_eq!(
+        meta.get("type").and_then(|v| v.as_str()),
+        Some("session_meta")
+    );
+    let payload = meta
+        .get("payload")
+        .unwrap_or_else(|| panic!("Missing payload in meta line"));
+    assert!(payload.get("id").is_some(), "SessionMeta missing id");
     assert!(
-        meta.get("timestamp").is_some(),
+        payload.get("timestamp").is_some(),
         "SessionMeta missing timestamp"
     );
 
@@ -375,8 +402,10 @@ async fn integration_creates_and_checks_session_file() {
         let Ok(item) = serde_json::from_str::<serde_json::Value>(line) else {
             continue;
         };
-        if item.get("type").and_then(|t| t.as_str()) == Some("message")
-            && let Some(c) = item.get("content")
+        if item.get("type").and_then(|t| t.as_str()) == Some("response_item")
+            && let Some(payload) = item.get("payload")
+            && payload.get("type").and_then(|t| t.as_str()) == Some("message")
+            && let Some(c) = payload.get("content")
             && c.to_string().contains(&marker)
         {
             found_message = true;
@@ -388,16 +417,9 @@ async fn integration_creates_and_checks_session_file() {
         "No message found in session file containing the marker"
     );
 
-    // Second run: resume and append.
-    let orig_len = content.lines().count();
+    // Second run: resume should update the existing file.
     let marker2 = format!("integration-resume-{}", Uuid::new_v4());
     let prompt2 = format!("echo {marker2}");
-    // Cross‑platform safe resume override.  On Windows, backslashes in a TOML string must be escaped
-    // or the parse will fail and the raw literal (including quotes) may be preserved all the way down
-    // to Config, which in turn breaks resume because the path is invalid. Normalize to forward slashes
-    // to sidestep the issue.
-    let resume_path_str = path.to_string_lossy().replace('\\', "/");
-    let resume_override = format!("experimental_resume=\"{resume_path_str}\"");
     let mut cmd2 = AssertCommand::new("cargo");
     cmd2.arg("run")
         .arg("-p")
@@ -406,11 +428,11 @@ async fn integration_creates_and_checks_session_file() {
         .arg("--")
         .arg("exec")
         .arg("--skip-git-repo-check")
-        .arg("-c")
-        .arg(&resume_override)
         .arg("-C")
         .arg(env!("CARGO_MANIFEST_DIR"))
-        .arg(&prompt2);
+        .arg(&prompt2)
+        .arg("resume")
+        .arg("--last");
     cmd2.env("CODEX_HOME", home.path())
         .env("OPENAI_API_KEY", "dummy")
         .env("CODEX_RS_SSE_FIXTURE", &fixture)
@@ -419,31 +441,50 @@ async fn integration_creates_and_checks_session_file() {
     let output2 = cmd2.output().unwrap();
     assert!(output2.status.success(), "resume codex-cli run failed");
 
-    // The rollout writer runs on a background async task; give it a moment to flush.
-    let mut new_len = orig_len;
-    let deadline = Instant::now() + Duration::from_secs(5);
-    let mut content2 = String::new();
-    while Instant::now() < deadline {
-        if let Ok(c) = std::fs::read_to_string(&path) {
-            let count = c.lines().count();
-            if count > orig_len {
-                content2 = c;
-                new_len = count;
+    // Find the new session file containing the resumed marker.
+    let deadline = Instant::now() + Duration::from_secs(10);
+    let mut resumed_path: Option<std::path::PathBuf> = None;
+    while Instant::now() < deadline && resumed_path.is_none() {
+        for entry in WalkDir::new(&sessions_dir) {
+            let entry = match entry {
+                Ok(e) => e,
+                Err(_) => continue,
+            };
+            if !entry.file_type().is_file() {
+                continue;
+            }
+            if !entry.file_name().to_string_lossy().ends_with(".jsonl") {
+                continue;
+            }
+            let p = entry.path();
+            let Ok(c) = std::fs::read_to_string(p) else {
+                continue;
+            };
+            if c.contains(&marker2) {
+                resumed_path = Some(p.to_path_buf());
                 break;
             }
         }
-        std::thread::sleep(Duration::from_millis(50));
+        if resumed_path.is_none() {
+            std::thread::sleep(Duration::from_millis(50));
+        }
     }
-    if content2.is_empty() {
-        // last attempt
-        content2 = std::fs::read_to_string(&path).unwrap();
-        new_len = content2.lines().count();
-    }
-    assert!(new_len > orig_len, "rollout file did not grow after resume");
-    assert!(content2.contains(&marker), "rollout lost original marker");
+
+    let resumed_path = resumed_path.expect("No resumed session file found containing the marker2");
+    // Resume should write to the existing log file.
+    assert_eq!(
+        resumed_path, path,
+        "resume should create a new session file"
+    );
+
+    let resumed_content = std::fs::read_to_string(&resumed_path).unwrap();
     assert!(
-        content2.contains(&marker2),
-        "rollout missing resumed marker"
+        resumed_content.contains(&marker),
+        "resumed file missing original marker"
+    );
+    assert!(
+        resumed_content.contains(&marker2),
+        "resumed file missing resumed marker"
     );
 }
 
@@ -571,7 +612,7 @@ async fn integration_git_info_unit_test() {
 
     // 5. Test serialization to ensure it works in SessionMeta
     let serialized = serde_json::to_string(&git_info).unwrap();
-    let deserialized: codex_core::git_info::GitInfo = serde_json::from_str(&serialized).unwrap();
+    let deserialized: GitInfo = serde_json::from_str(&serialized).unwrap();
 
     assert_eq!(git_info.commit_hash, deserialized.commit_hash);
     assert_eq!(git_info.branch, deserialized.branch);

codex-rs/core/tests/suite/client.rs

@@ -1,19 +1,34 @@
+use codex_core::CodexAuth;
+use codex_core::ContentItem;
 use codex_core::ConversationManager;
+use codex_core::LocalShellAction;
+use codex_core::LocalShellExecAction;
+use codex_core::LocalShellStatus;
+use codex_core::ModelClient;
 use codex_core::ModelProviderInfo;
 use codex_core::NewConversation;
+use codex_core::Prompt;
+use codex_core::ReasoningItemContent;
+use codex_core::ResponseEvent;
+use codex_core::ResponseItem;
 use codex_core::WireApi;
 use codex_core::built_in_model_providers;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::InputItem;
 use codex_core::protocol::Op;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
-use codex_login::AuthMode;
-use codex_login::CodexAuth;
+use codex_protocol::mcp_protocol::ConversationId;
+use codex_protocol::models::ReasoningItemReasoningSummary;
+use codex_protocol::models::WebSearchAction;
 use core_test_support::load_default_config_for_test;
 use core_test_support::load_sse_fixture_with_id;
 use core_test_support::wait_for_event;
+use futures::StreamExt;
 use serde_json::json;
+use std::io::Write;
+use std::sync::Arc;
 use tempfile::TempDir;
+use uuid::Uuid;
 use wiremock::Mock;
 use wiremock::MockServer;
 use wiremock::ResponseTemplate;
@@ -96,7 +111,7 @@ fn write_auth_json(
         "OPENAI_API_KEY": openai_api_key,
         "tokens": tokens,
         // RFC3339 datetime; value doesn't matter for these tests
-        "last_refresh": "2025-08-06T20:41:36.232376Z",
+        "last_refresh": chrono::Utc::now(),
     });
 
     std::fs::write(
@@ -109,7 +124,178 @@ fn write_auth_json(
 }
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn includes_session_id_and_model_headers_in_request() {
+async fn resume_includes_initial_messages_and_sends_prior_items() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // Create a fake rollout session file with prior user + system + assistant messages.
+    let tmpdir = TempDir::new().unwrap();
+    let session_path = tmpdir.path().join("resume-session.jsonl");
+    let mut f = std::fs::File::create(&session_path).unwrap();
+    let convo_id = Uuid::new_v4();
+    writeln!(
+        f,
+        "{}",
+        json!({
+            "timestamp": "2024-01-01T00:00:00.000Z",
+            "type": "session_meta",
+            "payload": {
+                "id": convo_id,
+                "timestamp": "2024-01-01T00:00:00Z",
+                "instructions": "be nice",
+                "cwd": ".",
+                "originator": "test_originator",
+                "cli_version": "test_version"
+            }
+        })
+    )
+    .unwrap();
+
+    // Prior item: user message (should be delivered)
+    let prior_user = codex_protocol::models::ResponseItem::Message {
+        id: None,
+        role: "user".to_string(),
+        content: vec![codex_protocol::models::ContentItem::InputText {
+            text: "resumed user message".to_string(),
+        }],
+    };
+    let prior_user_json = serde_json::to_value(&prior_user).unwrap();
+    writeln!(
+        f,
+        "{}",
+        json!({
+            "timestamp": "2024-01-01T00:00:01.000Z",
+            "type": "response_item",
+            "payload": prior_user_json
+        })
+    )
+    .unwrap();
+
+    // Prior item: system message (excluded from API history)
+    let prior_system = codex_protocol::models::ResponseItem::Message {
+        id: None,
+        role: "system".to_string(),
+        content: vec![codex_protocol::models::ContentItem::OutputText {
+            text: "resumed system instruction".to_string(),
+        }],
+    };
+    let prior_system_json = serde_json::to_value(&prior_system).unwrap();
+    writeln!(
+        f,
+        "{}",
+        json!({
+            "timestamp": "2024-01-01T00:00:02.000Z",
+            "type": "response_item",
+            "payload": prior_system_json
+        })
+    )
+    .unwrap();
+
+    // Prior item: assistant message
+    let prior_item = codex_protocol::models::ResponseItem::Message {
+        id: None,
+        role: "assistant".to_string(),
+        content: vec![codex_protocol::models::ContentItem::OutputText {
+            text: "resumed assistant message".to_string(),
+        }],
+    };
+    let prior_item_json = serde_json::to_value(&prior_item).unwrap();
+    writeln!(
+        f,
+        "{}",
+        json!({
+            "timestamp": "2024-01-01T00:00:03.000Z",
+            "type": "response_item",
+            "payload": prior_item_json
+        })
+    )
+    .unwrap();
+    drop(f);
+
+    // Mock server that will receive the resumed request
+    let server = MockServer::start().await;
+    let first = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse_completed("resp1"), "text/event-stream");
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(first)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    // Configure Codex to resume from our file
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider = model_provider;
+    // Also configure user instructions to ensure they are NOT delivered on resume.
+    config.user_instructions = Some("be nice".to_string());
+
+    let conversation_manager =
+        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
+    let auth_manager =
+        codex_core::AuthManager::from_auth_for_testing(CodexAuth::from_api_key("Test API Key"));
+    let NewConversation {
+        conversation: codex,
+        session_configured,
+        ..
+    } = conversation_manager
+        .resume_conversation_from_rollout(config, session_path.clone(), auth_manager)
+        .await
+        .expect("resume conversation");
+
+    // 1) Assert initial_messages only includes existing EventMsg entries; response items are not converted
+    let initial_msgs = session_configured
+        .initial_messages
+        .clone()
+        .expect("expected initial messages option for resumed session");
+    let initial_json = serde_json::to_value(&initial_msgs).unwrap();
+    let expected_initial_json = json!([]);
+    assert_eq!(initial_json, expected_initial_json);
+
+    // 2) Submit new input; the request body must include the prior item followed by the new user input.
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello".into(),
+            }],
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let request = &server.received_requests().await.unwrap()[0];
+    let request_body = request.body_json::<serde_json::Value>().unwrap();
+    let expected_input = json!([
+        {
+            "type": "message",
+            "role": "user",
+            "content": [{ "type": "input_text", "text": "resumed user message" }]
+        },
+        {
+            "type": "message",
+            "role": "assistant",
+            "content": [{ "type": "output_text", "text": "resumed assistant message" }]
+        },
+        {
+            "type": "message",
+            "role": "user",
+            "content": [{ "type": "input_text", "text": "hello" }]
+        }
+    ]);
+    assert_eq!(request_body["input"], expected_input);
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn includes_conversation_id_and_model_headers_in_request() {
     if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
         println!(
             "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
@@ -166,12 +352,12 @@ async fn includes_session_id_and_model_headers_in_request() {
 
     // get request from the server
     let request = &server.received_requests().await.unwrap()[0];
-    let request_session_id = request.headers.get("session_id").unwrap();
+    let request_conversation_id = request.headers.get("conversation_id").unwrap();
     let request_authorization = request.headers.get("authorization").unwrap();
     let request_originator = request.headers.get("originator").unwrap();
 
     assert_eq!(
-        request_session_id.to_str().unwrap(),
+        request_conversation_id.to_str().unwrap(),
         conversation_id.to_string()
     );
     assert_eq!(request_originator.to_str().unwrap(), "codex_cli_rs");
@@ -238,56 +424,6 @@ async fn includes_base_instructions_override_in_request() {
     );
 }
 
-#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn originator_config_override_is_used() {
-    // Mock server
-    let server = MockServer::start().await;
-
-    let first = ResponseTemplate::new(200)
-        .insert_header("content-type", "text/event-stream")
-        .set_body_raw(sse_completed("resp1"), "text/event-stream");
-
-    Mock::given(method("POST"))
-        .and(path("/v1/responses"))
-        .respond_with(first)
-        .expect(1)
-        .mount(&server)
-        .await;
-
-    let model_provider = ModelProviderInfo {
-        base_url: Some(format!("{}/v1", server.uri())),
-        ..built_in_model_providers()["openai"].clone()
-    };
-
-    let codex_home = TempDir::new().unwrap();
-    let mut config = load_default_config_for_test(&codex_home);
-    config.model_provider = model_provider;
-    config.responses_originator_header = "my_override".to_owned();
-
-    let conversation_manager =
-        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
-    let codex = conversation_manager
-        .new_conversation(config)
-        .await
-        .expect("create new conversation")
-        .conversation;
-
-    codex
-        .submit(Op::UserInput {
-            items: vec![InputItem::Text {
-                text: "hello".into(),
-            }],
-        })
-        .await
-        .unwrap();
-
-    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
-
-    let request = &server.received_requests().await.unwrap()[0];
-    let request_originator = request.headers.get("originator").unwrap();
-    assert_eq!(request_originator.to_str().unwrap(), "my_override");
-}
-
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn chatgpt_auth_sends_correct_request() {
     if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
@@ -344,14 +480,14 @@ async fn chatgpt_auth_sends_correct_request() {
 
     // get request from the server
     let request = &server.received_requests().await.unwrap()[0];
-    let request_session_id = request.headers.get("session_id").unwrap();
+    let request_conversation_id = request.headers.get("conversation_id").unwrap();
     let request_authorization = request.headers.get("authorization").unwrap();
     let request_originator = request.headers.get("originator").unwrap();
     let request_chatgpt_account_id = request.headers.get("chatgpt-account-id").unwrap();
     let request_body = request.body_json::<serde_json::Value>().unwrap();
 
     assert_eq!(
-        request_session_id.to_str().unwrap(),
+        request_conversation_id.to_str().unwrap(),
         conversation_id.to_string()
     );
     assert_eq!(request_originator.to_str().unwrap(), "codex_cli_rs");
@@ -360,7 +496,6 @@ async fn chatgpt_auth_sends_correct_request() {
         "Bearer Access Token"
     );
     assert_eq!(request_chatgpt_account_id.to_str().unwrap(), "account_id");
-    assert!(!request_body["store"].as_bool().unwrap());
     assert!(request_body["stream"].as_bool().unwrap());
     assert_eq!(
         request_body["include"][0].as_str().unwrap(),
@@ -368,87 +503,6 @@ async fn chatgpt_auth_sends_correct_request() {
     );
 }
 
-#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn prefers_chatgpt_token_when_config_prefers_chatgpt() {
-    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
-        println!(
-            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
-        );
-        return;
-    }
-
-    // Mock server
-    let server = MockServer::start().await;
-
-    let first = ResponseTemplate::new(200)
-        .insert_header("content-type", "text/event-stream")
-        .set_body_raw(sse_completed("resp1"), "text/event-stream");
-
-    // Expect ChatGPT base path and correct headers
-    Mock::given(method("POST"))
-        .and(path("/v1/responses"))
-        .and(header_regex("Authorization", r"Bearer Access-123"))
-        .and(header_regex("chatgpt-account-id", r"acc-123"))
-        .respond_with(first)
-        .expect(1)
-        .mount(&server)
-        .await;
-
-    let model_provider = ModelProviderInfo {
-        base_url: Some(format!("{}/v1", server.uri())),
-        ..built_in_model_providers()["openai"].clone()
-    };
-
-    // Init session
-    let codex_home = TempDir::new().unwrap();
-    // Write auth.json that contains both API key and ChatGPT tokens for a plan that should prefer ChatGPT.
-    let _jwt = write_auth_json(
-        &codex_home,
-        Some("sk-test-key"),
-        "pro",
-        "Access-123",
-        Some("acc-123"),
-    );
-
-    let mut config = load_default_config_for_test(&codex_home);
-    config.model_provider = model_provider;
-    config.preferred_auth_method = AuthMode::ChatGPT;
-
-    let auth_manager =
-        match CodexAuth::from_codex_home(codex_home.path(), config.preferred_auth_method) {
-            Ok(Some(auth)) => codex_login::AuthManager::from_auth_for_testing(auth),
-            Ok(None) => panic!("No CodexAuth found in codex_home"),
-            Err(e) => panic!("Failed to load CodexAuth: {e}"),
-        };
-    let conversation_manager = ConversationManager::new(auth_manager);
-    let NewConversation {
-        conversation: codex,
-        ..
-    } = conversation_manager
-        .new_conversation(config)
-        .await
-        .expect("create new conversation");
-
-    codex
-        .submit(Op::UserInput {
-            items: vec![InputItem::Text {
-                text: "hello".into(),
-            }],
-        })
-        .await
-        .unwrap();
-
-    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
-
-    // verify request body flags
-    let request = &server.received_requests().await.unwrap()[0];
-    let request_body = request.body_json::<serde_json::Value>().unwrap();
-    assert!(
-        !request_body["store"].as_bool().unwrap(),
-        "store should be false for ChatGPT auth"
-    );
-}
-
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn prefers_apikey_when_config_prefers_apikey_even_with_chatgpt_tokens() {
     if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
@@ -493,14 +547,12 @@ async fn prefers_apikey_when_config_prefers_apikey_even_with_chatgpt_tokens() {
 
     let mut config = load_default_config_for_test(&codex_home);
     config.model_provider = model_provider;
-    config.preferred_auth_method = AuthMode::ApiKey;
 
-    let auth_manager =
-        match CodexAuth::from_codex_home(codex_home.path(), config.preferred_auth_method) {
-            Ok(Some(auth)) => codex_login::AuthManager::from_auth_for_testing(auth),
-            Ok(None) => panic!("No CodexAuth found in codex_home"),
-            Err(e) => panic!("Failed to load CodexAuth: {e}"),
-        };
+    let auth_manager = match CodexAuth::from_codex_home(codex_home.path()) {
+        Ok(Some(auth)) => codex_core::AuthManager::from_auth_for_testing(auth),
+        Ok(None) => panic!("No CodexAuth found in codex_home"),
+        Err(e) => panic!("Failed to load CodexAuth: {e}"),
+    };
     let conversation_manager = ConversationManager::new(auth_manager);
     let NewConversation {
         conversation: codex,
@@ -520,14 +572,6 @@ async fn prefers_apikey_when_config_prefers_apikey_even_with_chatgpt_tokens() {
         .unwrap();
 
     wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
-
-    // verify request body flags
-    let request = &server.received_requests().await.unwrap()[0];
-    let request_body = request.body_json::<serde_json::Value>().unwrap();
-    assert!(
-        request_body["store"].as_bool().unwrap(),
-        "store should be true for API key auth"
-    );
 }
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
@@ -591,6 +635,147 @@ async fn includes_user_instructions_message_in_request() {
     assert_message_ends_with(&request_body["input"][1], "</environment_context>");
 }
 
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn azure_responses_request_includes_store_and_reasoning_ids() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let server = MockServer::start().await;
+
+    let sse_body = concat!(
+        "data: {\"type\":\"response.created\",\"response\":{}}\n\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_1\"}}\n\n",
+    );
+
+    let template = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse_body, "text/event-stream");
+
+    Mock::given(method("POST"))
+        .and(path("/openai/responses"))
+        .respond_with(template)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let provider = ModelProviderInfo {
+        name: "azure".into(),
+        base_url: Some(format!("{}/openai", server.uri())),
+        env_key: None,
+        env_key_instructions: None,
+        wire_api: WireApi::Responses,
+        query_params: None,
+        http_headers: None,
+        env_http_headers: None,
+        request_max_retries: Some(0),
+        stream_max_retries: Some(0),
+        stream_idle_timeout_ms: Some(5_000),
+        requires_openai_auth: false,
+    };
+
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider_id = provider.name.clone();
+    config.model_provider = provider.clone();
+    let effort = config.model_reasoning_effort;
+    let summary = config.model_reasoning_summary;
+    let config = Arc::new(config);
+
+    let client = ModelClient::new(
+        Arc::clone(&config),
+        None,
+        provider,
+        effort,
+        summary,
+        ConversationId::new(),
+    );
+
+    let mut prompt = Prompt::default();
+    prompt.input.push(ResponseItem::Reasoning {
+        id: "reasoning-id".into(),
+        summary: vec![ReasoningItemReasoningSummary::SummaryText {
+            text: "summary".into(),
+        }],
+        content: Some(vec![ReasoningItemContent::ReasoningText {
+            text: "content".into(),
+        }]),
+        encrypted_content: None,
+    });
+    prompt.input.push(ResponseItem::Message {
+        id: Some("message-id".into()),
+        role: "assistant".into(),
+        content: vec![ContentItem::OutputText {
+            text: "message".into(),
+        }],
+    });
+    prompt.input.push(ResponseItem::WebSearchCall {
+        id: Some("web-search-id".into()),
+        status: Some("completed".into()),
+        action: WebSearchAction::Search {
+            query: "weather".into(),
+        },
+    });
+    prompt.input.push(ResponseItem::FunctionCall {
+        id: Some("function-id".into()),
+        name: "do_thing".into(),
+        arguments: "{}".into(),
+        call_id: "function-call-id".into(),
+    });
+    prompt.input.push(ResponseItem::LocalShellCall {
+        id: Some("local-shell-id".into()),
+        call_id: Some("local-shell-call-id".into()),
+        status: LocalShellStatus::Completed,
+        action: LocalShellAction::Exec(LocalShellExecAction {
+            command: vec!["echo".into(), "hello".into()],
+            timeout_ms: None,
+            working_directory: None,
+            env: None,
+            user: None,
+        }),
+    });
+    prompt.input.push(ResponseItem::CustomToolCall {
+        id: Some("custom-tool-id".into()),
+        status: Some("completed".into()),
+        call_id: "custom-tool-call-id".into(),
+        name: "custom_tool".into(),
+        input: "{}".into(),
+    });
+
+    let mut stream = client
+        .stream(&prompt)
+        .await
+        .expect("responses stream to start");
+
+    while let Some(event) = stream.next().await {
+        if let Ok(ResponseEvent::Completed { .. }) = event {
+            break;
+        }
+    }
+
+    let requests = server
+        .received_requests()
+        .await
+        .expect("mock server collected requests");
+    assert_eq!(requests.len(), 1, "expected a single request");
+    let body: serde_json::Value = requests[0]
+        .body_json()
+        .expect("request body to be valid JSON");
+
+    assert_eq!(body["store"], serde_json::Value::Bool(true));
+    assert_eq!(body["stream"], serde_json::Value::Bool(true));
+    assert_eq!(body["input"].as_array().map(Vec::len), Some(6));
+    assert_eq!(body["input"][0]["id"].as_str(), Some("reasoning-id"));
+    assert_eq!(body["input"][1]["id"].as_str(), Some("message-id"));
+    assert_eq!(body["input"][2]["id"].as_str(), Some("web-search-id"));
+    assert_eq!(body["input"][3]["id"].as_str(), Some("function-id"));
+    assert_eq!(body["input"][4]["id"].as_str(), Some("local-shell-id"));
+    assert_eq!(body["input"][5]["id"].as_str(), Some("custom-tool-id"));
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn azure_overrides_assign_properties_used_for_responses_url() {
     let existing_env_var_with_random_value = if cfg!(windows) { "USERNAME" } else { "USER" };
@@ -845,34 +1030,29 @@ async fn history_dedupes_streamed_and_final_messages_across_turns() {
     assert_eq!(requests.len(), 3, "expected 3 requests (one per turn)");
 
     // Replace full-array compare with tail-only raw JSON compare using a single hard-coded value.
-    let r3_tail_expected = serde_json::json!([
+    let r3_tail_expected = json!([
         {
             "type": "message",
-            "id": null,
             "role": "user",
             "content": [{"type":"input_text","text":"U1"}]
         },
         {
             "type": "message",
-            "id": null,
             "role": "assistant",
             "content": [{"type":"output_text","text":"Hey there!\n"}]
         },
         {
             "type": "message",
-            "id": null,
             "role": "user",
             "content": [{"type":"input_text","text":"U2"}]
         },
         {
             "type": "message",
-            "id": null,
             "role": "assistant",
             "content": [{"type":"output_text","text":"Hey there!\n"}]
         },
         {
             "type": "message",
-            "id": null,
             "role": "user",
             "content": [{"type":"input_text","text":"U3"}]
         }

codex-rs/core/tests/suite/compact.rs

@@ -1,29 +1,40 @@
 #![expect(clippy::unwrap_used)]
 
+use codex_core::CodexAuth;
 use codex_core::ConversationManager;
 use codex_core::ModelProviderInfo;
+use codex_core::NewConversation;
 use codex_core::built_in_model_providers;
+use codex_core::protocol::ErrorEvent;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::InputItem;
 use codex_core::protocol::Op;
+use codex_core::protocol::RolloutItem;
+use codex_core::protocol::RolloutLine;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
-use codex_login::CodexAuth;
 use core_test_support::load_default_config_for_test;
 use core_test_support::wait_for_event;
 use serde_json::Value;
 use tempfile::TempDir;
+use wiremock::BodyPrintLimit;
 use wiremock::Mock;
 use wiremock::MockServer;
+use wiremock::Request;
+use wiremock::Respond;
 use wiremock::ResponseTemplate;
 use wiremock::matchers::method;
 use wiremock::matchers::path;
 
 use pretty_assertions::assert_eq;
+use std::sync::Arc;
+use std::sync::Mutex;
+use std::sync::atomic::AtomicUsize;
+use std::sync::atomic::Ordering;
 
 // --- Test helpers -----------------------------------------------------------
 
 /// Build an SSE stream body from a list of JSON events.
-fn sse(events: Vec<Value>) -> String {
+pub(super) fn sse(events: Vec<Value>) -> String {
     use std::fmt::Write as _;
     let mut out = String::new();
     for ev in events {
@@ -39,7 +50,7 @@ fn sse(events: Vec<Value>) -> String {
 }
 
 /// Convenience: SSE event for a completed response with a specific id.
-fn ev_completed(id: &str) -> Value {
+pub(super) fn ev_completed(id: &str) -> Value {
     serde_json::json!({
         "type": "response.completed",
         "response": {
@@ -49,8 +60,24 @@ fn ev_completed(id: &str) -> Value {
     })
 }
 
+fn ev_completed_with_tokens(id: &str, total_tokens: u64) -> Value {
+    serde_json::json!({
+        "type": "response.completed",
+        "response": {
+            "id": id,
+            "usage": {
+                "input_tokens": total_tokens,
+                "input_tokens_details": null,
+                "output_tokens": 0,
+                "output_tokens_details": null,
+                "total_tokens": total_tokens
+            }
+        }
+    })
+}
+
 /// Convenience: SSE event for a single assistant message output item.
-fn ev_assistant_message(id: &str, text: &str) -> Value {
+pub(super) fn ev_assistant_message(id: &str, text: &str) -> Value {
     serde_json::json!({
         "type": "response.output_item.done",
         "item": {
@@ -62,13 +89,25 @@ fn ev_assistant_message(id: &str, text: &str) -> Value {
     })
 }
 
-fn sse_response(body: String) -> ResponseTemplate {
+fn ev_function_call(call_id: &str, name: &str, arguments: &str) -> Value {
+    serde_json::json!({
+        "type": "response.output_item.done",
+        "item": {
+            "type": "function_call",
+            "call_id": call_id,
+            "name": name,
+            "arguments": arguments
+        }
+    })
+}
+
+pub(super) fn sse_response(body: String) -> ResponseTemplate {
     ResponseTemplate::new(200)
         .insert_header("content-type", "text/event-stream")
         .set_body_raw(body, "text/event-stream")
 }
 
-async fn mount_sse_once<M>(server: &MockServer, matcher: M, body: String)
+pub(super) async fn mount_sse_once<M>(server: &MockServer, matcher: M, body: String)
 where
     M: wiremock::Match + Send + Sync + 'static,
 {
@@ -76,15 +115,32 @@ where
         .and(path("/v1/responses"))
         .and(matcher)
         .respond_with(sse_response(body))
-        .expect(1)
         .mount(server)
         .await;
 }
 
-const FIRST_REPLY: &str = "FIRST_REPLY";
-const SUMMARY_TEXT: &str = "SUMMARY_ONLY_CONTEXT";
-const SUMMARIZE_TRIGGER: &str = "Start Summarization";
+async fn start_mock_server() -> MockServer {
+    MockServer::builder()
+        .body_print_limit(BodyPrintLimit::Limited(80_000))
+        .start()
+        .await
+}
+
+pub(super) const FIRST_REPLY: &str = "FIRST_REPLY";
+pub(super) const SUMMARY_TEXT: &str = "SUMMARY_ONLY_CONTEXT";
+pub(super) const SUMMARIZE_TRIGGER: &str = "Start Summarization";
 const THIRD_USER_MSG: &str = "next turn";
+const AUTO_SUMMARY_TEXT: &str = "AUTO_SUMMARY";
+const FIRST_AUTO_MSG: &str = "token limit start";
+const SECOND_AUTO_MSG: &str = "token limit push";
+const STILL_TOO_BIG_REPLY: &str = "STILL_TOO_BIG";
+const MULTI_AUTO_MSG: &str = "multi auto";
+const SECOND_LARGE_REPLY: &str = "SECOND_LARGE_REPLY";
+const FIRST_AUTO_SUMMARY: &str = "FIRST_AUTO_SUMMARY";
+const SECOND_AUTO_SUMMARY: &str = "SECOND_AUTO_SUMMARY";
+const FINAL_REPLY: &str = "FINAL_REPLY";
+const DUMMY_FUNCTION_NAME: &str = "unsupported_tool";
+const DUMMY_CALL_ID: &str = "call-multi-auto";
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn summarize_context_three_requests_and_instructions() {
@@ -96,7 +152,7 @@ async fn summarize_context_three_requests_and_instructions() {
     }
 
     // Set up a mock server that we can inspect after the run.
-    let server = MockServer::start().await;
+    let server = start_mock_server().await;
 
     // SSE 1: assistant replies normally so it is recorded in history.
     let sse1 = sse(vec![
@@ -141,12 +197,14 @@ async fn summarize_context_three_requests_and_instructions() {
     let home = TempDir::new().unwrap();
     let mut config = load_default_config_for_test(&home);
     config.model_provider = model_provider;
+    config.model_auto_compact_token_limit = Some(200_000);
     let conversation_manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
-    let codex = conversation_manager
-        .new_conversation(config)
-        .await
-        .unwrap()
-        .conversation;
+    let NewConversation {
+        conversation: codex,
+        session_configured,
+        ..
+    } = conversation_manager.new_conversation(config).await.unwrap();
+    let rollout_path = session_configured.rollout_path;
 
     // 1) Normal user input – should hit server once.
     codex
@@ -194,7 +252,7 @@ async fn summarize_context_three_requests_and_instructions() {
         "summarization should override base instructions"
     );
     assert!(
-        instr2.contains("You are a summarization assistant"),
+        instr2.contains("You have exceeded the maximum number of tokens"),
         "summarization instructions not applied"
     );
 
@@ -205,14 +263,17 @@ async fn summarize_context_three_requests_and_instructions() {
     assert_eq!(last2.get("type").unwrap().as_str().unwrap(), "message");
     assert_eq!(last2.get("role").unwrap().as_str().unwrap(), "user");
     let text2 = last2["content"][0]["text"].as_str().unwrap();
-    assert!(text2.contains(SUMMARIZE_TRIGGER));
+    assert!(
+        text2.contains(SUMMARIZE_TRIGGER),
+        "expected summarize trigger, got `{text2}`"
+    );
 
-    // Third request must contain only the summary from step 2 as prior history plus new user msg.
+    // Third request must contain the refreshed instructions, bridge summary message and new user msg.
     let input3 = body3.get("input").and_then(|v| v.as_array()).unwrap();
     println!("third request body: {body3}");
     assert!(
-        input3.len() >= 2,
-        "expected summary + new user message in third request"
+        input3.len() >= 3,
+        "expected refreshed context and new user message in third request"
     );
 
     // Collect all (role, text) message tuples.
@@ -228,24 +289,611 @@ async fn summarize_context_three_requests_and_instructions() {
         }
     }
 
-    // Exactly one assistant message should remain after compaction and the new user message is present.
+    // No previous assistant messages should remain and the new user message is present.
     let assistant_count = messages.iter().filter(|(r, _)| r == "assistant").count();
-    assert_eq!(
-        assistant_count, 1,
-        "exactly one assistant message should remain after compaction"
-    );
+    assert_eq!(assistant_count, 0, "assistant history should be cleared");
     assert!(
         messages
             .iter()
             .any(|(r, t)| r == "user" && t == THIRD_USER_MSG),
         "third request should include the new user message"
     );
+    let Some((_, bridge_text)) = messages.iter().find(|(role, text)| {
+        role == "user"
+            && (text.contains("Here were the user messages")
+                || text.contains("Here are all the user messages"))
+            && text.contains(SUMMARY_TEXT)
+    }) else {
+        panic!("expected a bridge message containing the summary");
+    };
     assert!(
-        !messages.iter().any(|(_, t)| t.contains("hello world")),
-        "third request should not include the original user input"
+        bridge_text.contains("hello world"),
+        "bridge should capture earlier user messages"
     );
     assert!(
-        !messages.iter().any(|(_, t)| t.contains(SUMMARIZE_TRIGGER)),
+        !bridge_text.contains(SUMMARIZE_TRIGGER),
+        "bridge text should not echo the summarize trigger"
+    );
+    assert!(
+        !messages
+            .iter()
+            .any(|(_, text)| text.contains(SUMMARIZE_TRIGGER)),
         "third request should not include the summarize trigger"
     );
+
+    // Shut down Codex to flush rollout entries before inspecting the file.
+    codex.submit(Op::Shutdown).await.unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::ShutdownComplete)).await;
+
+    // Verify rollout contains APITurn entries for each API call and a Compacted entry.
+    println!("rollout path: {}", rollout_path.display());
+    let text = std::fs::read_to_string(&rollout_path).unwrap_or_else(|e| {
+        panic!(
+            "failed to read rollout file {}: {e}",
+            rollout_path.display()
+        )
+    });
+    let mut api_turn_count = 0usize;
+    let mut saw_compacted_summary = false;
+    for line in text.lines() {
+        let trimmed = line.trim();
+        if trimmed.is_empty() {
+            continue;
+        }
+        let Ok(entry): Result<RolloutLine, _> = serde_json::from_str(trimmed) else {
+            continue;
+        };
+        match entry.item {
+            RolloutItem::TurnContext(_) => {
+                api_turn_count += 1;
+            }
+            RolloutItem::Compacted(ci) => {
+                if ci.message == SUMMARY_TEXT {
+                    saw_compacted_summary = true;
+                }
+            }
+            _ => {}
+        }
+    }
+
+    assert!(
+        api_turn_count == 3,
+        "expected three APITurn entries in rollout"
+    );
+    assert!(
+        saw_compacted_summary,
+        "expected a Compacted entry containing the summarizer output"
+    );
+}
+
+// Windows CI only: bump to 4 workers to prevent SSE/event starvation and test timeouts.
+#[cfg_attr(windows, tokio::test(flavor = "multi_thread", worker_threads = 4))]
+#[cfg_attr(not(windows), tokio::test(flavor = "multi_thread", worker_threads = 2))]
+async fn auto_compact_runs_after_token_limit_hit() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let server = start_mock_server().await;
+
+    let sse1 = sse(vec![
+        ev_assistant_message("m1", FIRST_REPLY),
+        ev_completed_with_tokens("r1", 70_000),
+    ]);
+
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", "SECOND_REPLY"),
+        ev_completed_with_tokens("r2", 330_000),
+    ]);
+
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", AUTO_SUMMARY_TEXT),
+        ev_completed_with_tokens("r3", 200),
+    ]);
+
+    let first_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(FIRST_AUTO_MSG)
+            && !body.contains(SECOND_AUTO_MSG)
+            && !body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(first_matcher)
+        .respond_with(sse_response(sse1))
+        .mount(&server)
+        .await;
+
+    let second_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(SECOND_AUTO_MSG)
+            && body.contains(FIRST_AUTO_MSG)
+            && !body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(second_matcher)
+        .respond_with(sse_response(sse2))
+        .mount(&server)
+        .await;
+
+    let third_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(third_matcher)
+        .respond_with(sse_response(sse3))
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&home);
+    config.model_provider = model_provider;
+    config.model_auto_compact_token_limit = Some(200_000);
+    let conversation_manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
+    let codex = conversation_manager
+        .new_conversation(config)
+        .await
+        .unwrap()
+        .conversation;
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: FIRST_AUTO_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: SECOND_AUTO_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+    // wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let requests = server.received_requests().await.unwrap();
+    assert!(
+        requests.len() >= 3,
+        "auto compact should add at least a third request, got {}",
+        requests.len()
+    );
+    let is_auto_compact = |req: &wiremock::Request| {
+        std::str::from_utf8(&req.body)
+            .unwrap_or("")
+            .contains("You have exceeded the maximum number of tokens")
+    };
+    let auto_compact_count = requests.iter().filter(|req| is_auto_compact(req)).count();
+    assert_eq!(
+        auto_compact_count, 1,
+        "expected exactly one auto compact request"
+    );
+    let auto_compact_index = requests
+        .iter()
+        .enumerate()
+        .find_map(|(idx, req)| is_auto_compact(req).then_some(idx))
+        .expect("auto compact request missing");
+    assert_eq!(
+        auto_compact_index, 2,
+        "auto compact should add a third request"
+    );
+
+    let body3 = requests[auto_compact_index]
+        .body_json::<serde_json::Value>()
+        .unwrap();
+    let instructions = body3
+        .get("instructions")
+        .and_then(|v| v.as_str())
+        .unwrap_or_default();
+    assert!(
+        instructions.contains("You have exceeded the maximum number of tokens"),
+        "auto compact should reuse summarization instructions"
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn auto_compact_persists_rollout_entries() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let server = start_mock_server().await;
+
+    let sse1 = sse(vec![
+        ev_assistant_message("m1", FIRST_REPLY),
+        ev_completed_with_tokens("r1", 70_000),
+    ]);
+
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", "SECOND_REPLY"),
+        ev_completed_with_tokens("r2", 330_000),
+    ]);
+
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", AUTO_SUMMARY_TEXT),
+        ev_completed_with_tokens("r3", 200),
+    ]);
+
+    let first_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(FIRST_AUTO_MSG)
+            && !body.contains(SECOND_AUTO_MSG)
+            && !body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(first_matcher)
+        .respond_with(sse_response(sse1))
+        .mount(&server)
+        .await;
+
+    let second_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(SECOND_AUTO_MSG)
+            && body.contains(FIRST_AUTO_MSG)
+            && !body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(second_matcher)
+        .respond_with(sse_response(sse2))
+        .mount(&server)
+        .await;
+
+    let third_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(third_matcher)
+        .respond_with(sse_response(sse3))
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&home);
+    config.model_provider = model_provider;
+    let conversation_manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
+    let NewConversation {
+        conversation: codex,
+        session_configured,
+        ..
+    } = conversation_manager.new_conversation(config).await.unwrap();
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: FIRST_AUTO_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: SECOND_AUTO_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    codex.submit(Op::Shutdown).await.unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::ShutdownComplete)).await;
+
+    let rollout_path = session_configured.rollout_path;
+    let text = std::fs::read_to_string(&rollout_path).unwrap_or_else(|e| {
+        panic!(
+            "failed to read rollout file {}: {e}",
+            rollout_path.display()
+        )
+    });
+
+    let mut turn_context_count = 0usize;
+    for line in text.lines() {
+        let trimmed = line.trim();
+        if trimmed.is_empty() {
+            continue;
+        }
+        let Ok(entry): Result<RolloutLine, _> = serde_json::from_str(trimmed) else {
+            continue;
+        };
+        match entry.item {
+            RolloutItem::TurnContext(_) => {
+                turn_context_count += 1;
+            }
+            RolloutItem::Compacted(_) => {}
+            _ => {}
+        }
+    }
+
+    assert!(
+        turn_context_count >= 2,
+        "expected at least two turn context entries, got {turn_context_count}"
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn auto_compact_stops_after_failed_attempt() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let server = start_mock_server().await;
+
+    let sse1 = sse(vec![
+        ev_assistant_message("m1", FIRST_REPLY),
+        ev_completed_with_tokens("r1", 500),
+    ]);
+
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", SUMMARY_TEXT),
+        ev_completed_with_tokens("r2", 50),
+    ]);
+
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", STILL_TOO_BIG_REPLY),
+        ev_completed_with_tokens("r3", 500),
+    ]);
+
+    let first_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(FIRST_AUTO_MSG)
+            && !body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(first_matcher)
+        .respond_with(sse_response(sse1.clone()))
+        .mount(&server)
+        .await;
+
+    let second_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("You have exceeded the maximum number of tokens")
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(second_matcher)
+        .respond_with(sse_response(sse2.clone()))
+        .mount(&server)
+        .await;
+
+    let third_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        !body.contains("You have exceeded the maximum number of tokens")
+            && body.contains(SUMMARY_TEXT)
+    };
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(third_matcher)
+        .respond_with(sse_response(sse3.clone()))
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&home);
+    config.model_provider = model_provider;
+    config.model_auto_compact_token_limit = Some(200);
+    let conversation_manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
+    let codex = conversation_manager
+        .new_conversation(config)
+        .await
+        .unwrap()
+        .conversation;
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: FIRST_AUTO_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    let error_event = wait_for_event(&codex, |ev| matches!(ev, EventMsg::Error(_))).await;
+    let EventMsg::Error(ErrorEvent { message }) = error_event else {
+        panic!("expected error event");
+    };
+    assert!(
+        message.contains("limit"),
+        "error message should include limit information: {message}"
+    );
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let requests = server.received_requests().await.unwrap();
+    assert_eq!(
+        requests.len(),
+        3,
+        "auto compact should attempt at most one summarization before erroring"
+    );
+
+    let last_body = requests[2].body_json::<serde_json::Value>().unwrap();
+    let instructions = last_body
+        .get("instructions")
+        .and_then(|v| v.as_str())
+        .unwrap_or_default();
+    assert!(
+        !instructions.contains("You have exceeded the maximum number of tokens"),
+        "third request should be the follow-up turn, not another summarization"
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn auto_compact_allows_multiple_attempts_when_interleaved_with_other_turn_events() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let server = start_mock_server().await;
+
+    let sse1 = sse(vec![
+        ev_assistant_message("m1", FIRST_REPLY),
+        ev_completed_with_tokens("r1", 500),
+    ]);
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", FIRST_AUTO_SUMMARY),
+        ev_completed_with_tokens("r2", 50),
+    ]);
+    let sse3 = sse(vec![
+        ev_function_call(DUMMY_CALL_ID, DUMMY_FUNCTION_NAME, "{}"),
+        ev_completed_with_tokens("r3", 150),
+    ]);
+    let sse4 = sse(vec![
+        ev_assistant_message("m4", SECOND_LARGE_REPLY),
+        ev_completed_with_tokens("r4", 450),
+    ]);
+    let sse5 = sse(vec![
+        ev_assistant_message("m5", SECOND_AUTO_SUMMARY),
+        ev_completed_with_tokens("r5", 60),
+    ]);
+    let sse6 = sse(vec![
+        ev_assistant_message("m6", FINAL_REPLY),
+        ev_completed_with_tokens("r6", 120),
+    ]);
+
+    #[derive(Clone)]
+    struct SeqResponder {
+        bodies: Arc<Vec<String>>,
+        calls: Arc<AtomicUsize>,
+        requests: Arc<Mutex<Vec<Vec<u8>>>>,
+    }
+
+    impl SeqResponder {
+        fn new(bodies: Vec<String>) -> Self {
+            Self {
+                bodies: Arc::new(bodies),
+                calls: Arc::new(AtomicUsize::new(0)),
+                requests: Arc::new(Mutex::new(Vec::new())),
+            }
+        }
+
+        fn recorded_requests(&self) -> Vec<Vec<u8>> {
+            self.requests.lock().unwrap().clone()
+        }
+    }
+
+    impl Respond for SeqResponder {
+        fn respond(&self, req: &Request) -> ResponseTemplate {
+            let idx = self.calls.fetch_add(1, Ordering::SeqCst);
+            self.requests.lock().unwrap().push(req.body.clone());
+            let body = self
+                .bodies
+                .get(idx)
+                .unwrap_or_else(|| panic!("unexpected request index {idx}"))
+                .clone();
+            ResponseTemplate::new(200)
+                .insert_header("content-type", "text/event-stream")
+                .set_body_raw(body, "text/event-stream")
+        }
+    }
+
+    let responder = SeqResponder::new(vec![sse1, sse2, sse3, sse4, sse5, sse6]);
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(responder.clone())
+        .expect(6)
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&home);
+    config.model_provider = model_provider;
+    config.model_auto_compact_token_limit = Some(200);
+    let conversation_manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
+    let codex = conversation_manager
+        .new_conversation(config)
+        .await
+        .unwrap()
+        .conversation;
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: MULTI_AUTO_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    loop {
+        let event = codex.next_event().await.unwrap();
+        if let EventMsg::TaskComplete(_) = &event.msg
+            && !event.id.starts_with("auto-compact-")
+        {
+            break;
+        }
+    }
+
+    let request_bodies: Vec<String> = responder
+        .recorded_requests()
+        .into_iter()
+        .map(|body| String::from_utf8(body).unwrap_or_default())
+        .collect();
+    assert_eq!(
+        request_bodies.len(),
+        6,
+        "expected six requests including two auto compactions"
+    );
+    assert!(
+        request_bodies[0].contains(MULTI_AUTO_MSG),
+        "first request should contain the user input"
+    );
+    assert!(
+        request_bodies[1].contains("You have exceeded the maximum number of tokens"),
+        "first auto compact request should use summarization instructions"
+    );
+    assert!(
+        request_bodies[3].contains(&format!("unsupported call: {DUMMY_FUNCTION_NAME}")),
+        "function call output should be sent before the second auto compact"
+    );
+    assert!(
+        request_bodies[4].contains("You have exceeded the maximum number of tokens"),
+        "second auto compact request should reuse summarization instructions"
+    );
 }

codex-rs/core/tests/suite/compact_resume_fork.rs

@@ -0,0 +1,838 @@
+#![allow(clippy::expect_used)]
+
+//! Integration tests that cover compacting, resuming, and forking conversations.
+//!
+//! Each test sets up a mocked SSE conversation and drives the conversation through
+//! a specific sequence of operations. After every operation we capture the
+//! request payload that Codex would send to the model and assert that the
+//! model-visible history matches the expected sequence of messages.
+
+use super::compact::FIRST_REPLY;
+use super::compact::SUMMARIZE_TRIGGER;
+use super::compact::SUMMARY_TEXT;
+use super::compact::ev_assistant_message;
+use super::compact::ev_completed;
+use super::compact::mount_sse_once;
+use super::compact::sse;
+use codex_core::CodexAuth;
+use codex_core::CodexConversation;
+use codex_core::ConversationManager;
+use codex_core::ModelProviderInfo;
+use codex_core::NewConversation;
+use codex_core::built_in_model_providers;
+use codex_core::config::Config;
+use codex_core::protocol::ConversationPathResponseEvent;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::InputItem;
+use codex_core::protocol::Op;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use core_test_support::load_default_config_for_test;
+use core_test_support::wait_for_event;
+use pretty_assertions::assert_eq;
+use serde_json::Value;
+use serde_json::json;
+use std::sync::Arc;
+use tempfile::TempDir;
+use wiremock::MockServer;
+
+const AFTER_SECOND_RESUME: &str = "AFTER_SECOND_RESUME";
+
+fn network_disabled() -> bool {
+    std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok()
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+/// Scenario: compact an initial conversation, resume it, fork one turn back, and
+/// ensure the model-visible history matches expectations at each request.
+async fn compact_resume_and_fork_preserve_model_history_view() {
+    if network_disabled() {
+        println!("Skipping test because network is disabled in this sandbox");
+        return;
+    }
+
+    // 1. Arrange mocked SSE responses for the initial compact/resume/fork flow.
+    let server = MockServer::start().await;
+    mount_initial_flow(&server).await;
+
+    // 2. Start a new conversation and drive it through the compact/resume/fork steps.
+    let (_home, config, manager, base) = start_test_conversation(&server).await;
+
+    user_turn(&base, "hello world").await;
+    compact_conversation(&base).await;
+    user_turn(&base, "AFTER_COMPACT").await;
+    let base_path = fetch_conversation_path(&base, "base conversation").await;
+    assert!(
+        base_path.exists(),
+        "compact+resume test expects base path {base_path:?} to exist",
+    );
+
+    let resumed = resume_conversation(&manager, &config, base_path).await;
+    user_turn(&resumed, "AFTER_RESUME").await;
+    let resumed_path = fetch_conversation_path(&resumed, "resumed conversation").await;
+    assert!(
+        resumed_path.exists(),
+        "compact+resume test expects resumed path {resumed_path:?} to exist",
+    );
+
+    let forked = fork_conversation(&manager, &config, resumed_path, 1).await;
+    user_turn(&forked, "AFTER_FORK").await;
+
+    // 3. Capture the requests to the model and validate the history slices.
+    let requests = gather_request_bodies(&server).await;
+
+    // input after compact is a prefix of input after resume/fork
+    let input_after_compact = json!(requests[requests.len() - 3]["input"]);
+    let input_after_resume = json!(requests[requests.len() - 2]["input"]);
+    let input_after_fork = json!(requests[requests.len() - 1]["input"]);
+
+    let compact_arr = input_after_compact
+        .as_array()
+        .expect("input after compact should be an array");
+    let resume_arr = input_after_resume
+        .as_array()
+        .expect("input after resume should be an array");
+    let fork_arr = input_after_fork
+        .as_array()
+        .expect("input after fork should be an array");
+
+    assert!(
+        compact_arr.len() <= resume_arr.len(),
+        "after-resume input should have at least as many items as after-compact",
+    );
+    assert_eq!(compact_arr.as_slice(), &resume_arr[..compact_arr.len()]);
+    eprint!(
+        "len of compact: {}, len of fork: {}",
+        compact_arr.len(),
+        fork_arr.len()
+    );
+    eprintln!("input_after_fork:{}", json!(input_after_fork));
+    assert!(
+        compact_arr.len() <= fork_arr.len(),
+        "after-fork input should have at least as many items as after-compact",
+    );
+    assert_eq!(compact_arr.as_slice(), &fork_arr[..compact_arr.len()]);
+
+    let prompt = requests[0]["instructions"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let user_instructions = requests[0]["input"][0]["content"][0]["text"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let environment_context = requests[0]["input"][1]["content"][0]["text"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let tool_calls = json!(requests[0]["tools"].as_array());
+    let prompt_cache_key = requests[0]["prompt_cache_key"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let fork_prompt_cache_key = requests[requests.len() - 1]["prompt_cache_key"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let user_turn_1 = json!(
+    {
+      "model": "gpt-5",
+      "instructions": prompt,
+      "input": [
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": user_instructions
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": environment_context
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "hello world"
+            }
+          ]
+        }
+      ],
+      "tools": tool_calls,
+      "tool_choice": "auto",
+      "parallel_tool_calls": false,
+      "reasoning": {
+        "summary": "auto"
+      },
+      "store": false,
+      "stream": true,
+      "include": [
+        "reasoning.encrypted_content"
+      ],
+      "prompt_cache_key": prompt_cache_key
+    });
+    let compact_1 = json!(
+    {
+      "model": "gpt-5",
+      "instructions": "You have exceeded the maximum number of tokens, please stop coding and instead write a short memento message for the next agent. Your note should:
+- Summarize what you finished and what still needs work. If there was a recent update_plan call, repeat its steps verbatim.
+- List outstanding TODOs with file paths / line numbers so they're easy to find.
+- Flag code that needs more tests (edge cases, performance, integration, etc.).
+- Record any open bugs, quirks, or setup steps that will make it easier for the next agent to pick up where you left off.",
+      "input": [
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": user_instructions
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": environment_context
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "hello world"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "assistant",
+          "content": [
+            {
+              "type": "output_text",
+              "text": "FIRST_REPLY"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "Start Summarization"
+            }
+          ]
+        }
+      ],
+      "tools": [],
+      "tool_choice": "auto",
+      "parallel_tool_calls": false,
+      "reasoning": {
+        "summary": "auto"
+      },
+      "store": false,
+      "stream": true,
+      "include": [
+        "reasoning.encrypted_content"
+      ],
+      "prompt_cache_key": prompt_cache_key
+    });
+    let user_turn_2_after_compact = json!(
+    {
+      "model": "gpt-5",
+      "instructions": prompt,
+      "input": [
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": user_instructions
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": environment_context
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "You were originally given instructions from a user over one or more turns. Here were the user messages:
+
+hello world
+
+Another language model started to solve this problem and produced a summary of its thinking process. You also have access to the state of the tools that were used by that language model. Use this to build on the work that has already been done and avoid duplicating work. Here is the summary produced by the other language model, use the information in this summary to assist with your own analysis:
+
+SUMMARY_ONLY_CONTEXT"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "AFTER_COMPACT"
+            }
+          ]
+        }
+      ],
+      "tools": tool_calls,
+      "tool_choice": "auto",
+      "parallel_tool_calls": false,
+      "reasoning": {
+        "summary": "auto"
+      },
+      "store": false,
+      "stream": true,
+      "include": [
+        "reasoning.encrypted_content"
+      ],
+      "prompt_cache_key": prompt_cache_key
+    });
+    let usert_turn_3_after_resume = json!(
+    {
+      "model": "gpt-5",
+      "instructions": prompt,
+      "input": [
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": user_instructions
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": environment_context
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "You were originally given instructions from a user over one or more turns. Here were the user messages:
+
+hello world
+
+Another language model started to solve this problem and produced a summary of its thinking process. You also have access to the state of the tools that were used by that language model. Use this to build on the work that has already been done and avoid duplicating work. Here is the summary produced by the other language model, use the information in this summary to assist with your own analysis:
+
+SUMMARY_ONLY_CONTEXT"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "AFTER_COMPACT"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "assistant",
+          "content": [
+            {
+              "type": "output_text",
+              "text": "AFTER_COMPACT_REPLY"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "AFTER_RESUME"
+            }
+          ]
+        }
+      ],
+      "tools": tool_calls,
+      "tool_choice": "auto",
+      "parallel_tool_calls": false,
+      "reasoning": {
+        "summary": "auto"
+      },
+      "store": false,
+      "stream": true,
+      "include": [
+        "reasoning.encrypted_content"
+      ],
+      "prompt_cache_key": prompt_cache_key
+    });
+    let user_turn_3_after_fork = json!(
+    {
+      "model": "gpt-5",
+      "instructions": prompt,
+      "input": [
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": user_instructions
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": environment_context
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "You were originally given instructions from a user over one or more turns. Here were the user messages:
+
+hello world
+
+Another language model started to solve this problem and produced a summary of its thinking process. You also have access to the state of the tools that were used by that language model. Use this to build on the work that has already been done and avoid duplicating work. Here is the summary produced by the other language model, use the information in this summary to assist with your own analysis:
+
+SUMMARY_ONLY_CONTEXT"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "AFTER_COMPACT"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "assistant",
+          "content": [
+            {
+              "type": "output_text",
+              "text": "AFTER_COMPACT_REPLY"
+            }
+          ]
+        },
+        {
+          "type": "message",
+          "role": "user",
+          "content": [
+            {
+              "type": "input_text",
+              "text": "AFTER_FORK"
+            }
+          ]
+        }
+      ],
+      "tools": tool_calls,
+      "tool_choice": "auto",
+      "parallel_tool_calls": false,
+      "reasoning": {
+        "summary": "auto"
+      },
+      "store": false,
+      "stream": true,
+      "include": [
+        "reasoning.encrypted_content"
+      ],
+      "prompt_cache_key": fork_prompt_cache_key
+    });
+    let expected = json!([
+        user_turn_1,
+        compact_1,
+        user_turn_2_after_compact,
+        usert_turn_3_after_resume,
+        user_turn_3_after_fork
+    ]);
+    assert_eq!(requests.len(), 5);
+    assert_eq!(json!(requests), expected);
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+/// Scenario: after the forked branch is compacted, resuming again should reuse
+/// the compacted history and only append the new user message.
+async fn compact_resume_after_second_compaction_preserves_history() {
+    if network_disabled() {
+        println!("Skipping test because network is disabled in this sandbox");
+        return;
+    }
+
+    // 1. Arrange mocked SSE responses for the initial flow plus the second compact.
+    let server = MockServer::start().await;
+    mount_initial_flow(&server).await;
+    mount_second_compact_flow(&server).await;
+
+    // 2. Drive the conversation through compact -> resume -> fork -> compact -> resume.
+    let (_home, config, manager, base) = start_test_conversation(&server).await;
+
+    user_turn(&base, "hello world").await;
+    compact_conversation(&base).await;
+    user_turn(&base, "AFTER_COMPACT").await;
+    let base_path = fetch_conversation_path(&base, "base conversation").await;
+    assert!(
+        base_path.exists(),
+        "second compact test expects base path {base_path:?} to exist",
+    );
+
+    let resumed = resume_conversation(&manager, &config, base_path).await;
+    user_turn(&resumed, "AFTER_RESUME").await;
+    let resumed_path = fetch_conversation_path(&resumed, "resumed conversation").await;
+    assert!(
+        resumed_path.exists(),
+        "second compact test expects resumed path {resumed_path:?} to exist",
+    );
+
+    let forked = fork_conversation(&manager, &config, resumed_path, 3).await;
+    user_turn(&forked, "AFTER_FORK").await;
+
+    compact_conversation(&forked).await;
+    user_turn(&forked, "AFTER_COMPACT_2").await;
+    let forked_path = fetch_conversation_path(&forked, "forked conversation").await;
+    assert!(
+        forked_path.exists(),
+        "second compact test expects forked path {forked_path:?} to exist",
+    );
+
+    let resumed_again = resume_conversation(&manager, &config, forked_path).await;
+    user_turn(&resumed_again, AFTER_SECOND_RESUME).await;
+
+    let requests = gather_request_bodies(&server).await;
+    let input_after_compact = json!(requests[requests.len() - 2]["input"]);
+    let input_after_resume = json!(requests[requests.len() - 1]["input"]);
+
+    // test input after compact before resume is the same as input after resume
+    let compact_input_array = input_after_compact
+        .as_array()
+        .expect("input after compact should be an array");
+    let resume_input_array = input_after_resume
+        .as_array()
+        .expect("input after resume should be an array");
+    assert!(
+        compact_input_array.len() <= resume_input_array.len(),
+        "after-resume input should have at least as many items as after-compact"
+    );
+    assert_eq!(
+        compact_input_array.as_slice(),
+        &resume_input_array[..compact_input_array.len()]
+    );
+    // hard coded test
+    let prompt = requests[0]["instructions"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let user_instructions = requests[0]["input"][0]["content"][0]["text"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+    let environment_instructions = requests[0]["input"][1]["content"][0]["text"]
+        .as_str()
+        .unwrap_or_default()
+        .to_string();
+
+    let expected = json!([
+      {
+        "instructions": prompt,
+        "input": [
+          {
+            "type": "message",
+            "role": "user",
+            "content": [
+              {
+                "type": "input_text",
+                "text": user_instructions
+              }
+            ]
+          },
+          {
+            "type": "message",
+            "role": "user",
+            "content": [
+              {
+                "type": "input_text",
+                "text": environment_instructions
+              }
+            ]
+          },
+          {
+            "type": "message",
+            "role": "user",
+            "content": [
+              {
+                "type": "input_text",
+                "text": "You were originally given instructions from a user over one or more turns. Here were the user messages:\n\nAFTER_FORK\n\nAnother language model started to solve this problem and produced a summary of its thinking process. You also have access to the state of the tools that were used by that language model. Use this to build on the work that has already been done and avoid duplicating work. Here is the summary produced by the other language model, use the information in this summary to assist with your own analysis:\n\nSUMMARY_ONLY_CONTEXT"
+              }
+            ]
+          },
+          {
+            "type": "message",
+            "role": "user",
+            "content": [
+              {
+                "type": "input_text",
+                "text": "AFTER_COMPACT_2"
+              }
+            ]
+          },
+          {
+            "type": "message",
+            "role": "user",
+            "content": [
+              {
+                "type": "input_text",
+                "text": "AFTER_SECOND_RESUME"
+              }
+            ]
+          }
+        ],
+      }
+    ]);
+    let last_request_after_2_compacts = json!([{
+        "instructions": requests[requests.len() -1]["instructions"],
+        "input": requests[requests.len() -1]["input"],
+    }]);
+    assert_eq!(expected, last_request_after_2_compacts);
+}
+
+fn normalize_line_endings(value: &mut Value) {
+    match value {
+        Value::String(text) => {
+            if text.contains('\r') {
+                *text = text.replace("\r\n", "\n").replace('\r', "\n");
+            }
+        }
+        Value::Array(items) => {
+            for item in items {
+                normalize_line_endings(item);
+            }
+        }
+        Value::Object(map) => {
+            for item in map.values_mut() {
+                normalize_line_endings(item);
+            }
+        }
+        _ => {}
+    }
+}
+
+async fn gather_request_bodies(server: &MockServer) -> Vec<Value> {
+    server
+        .received_requests()
+        .await
+        .expect("mock server should not fail")
+        .into_iter()
+        .map(|req| {
+            let mut value = req.body_json::<Value>().expect("valid JSON body");
+            normalize_line_endings(&mut value);
+            value
+        })
+        .collect()
+}
+
+async fn mount_initial_flow(server: &MockServer) {
+    let sse1 = sse(vec![
+        ev_assistant_message("m1", FIRST_REPLY),
+        ev_completed("r1"),
+    ]);
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", SUMMARY_TEXT),
+        ev_completed("r2"),
+    ]);
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", "AFTER_COMPACT_REPLY"),
+        ev_completed("r3"),
+    ]);
+    let sse4 = sse(vec![ev_completed("r4")]);
+    let sse5 = sse(vec![ev_completed("r5")]);
+
+    let match_first = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("\"text\":\"hello world\"")
+            && !body.contains(&format!("\"text\":\"{SUMMARIZE_TRIGGER}\""))
+            && !body.contains("\"text\":\"AFTER_COMPACT\"")
+            && !body.contains("\"text\":\"AFTER_RESUME\"")
+            && !body.contains("\"text\":\"AFTER_FORK\"")
+    };
+    mount_sse_once(server, match_first, sse1).await;
+
+    let match_compact = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(&format!("\"text\":\"{SUMMARIZE_TRIGGER}\""))
+    };
+    mount_sse_once(server, match_compact, sse2).await;
+
+    let match_after_compact = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("\"text\":\"AFTER_COMPACT\"")
+            && !body.contains("\"text\":\"AFTER_RESUME\"")
+            && !body.contains("\"text\":\"AFTER_FORK\"")
+    };
+    mount_sse_once(server, match_after_compact, sse3).await;
+
+    let match_after_resume = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("\"text\":\"AFTER_RESUME\"")
+    };
+    mount_sse_once(server, match_after_resume, sse4).await;
+
+    let match_after_fork = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("\"text\":\"AFTER_FORK\"")
+    };
+    mount_sse_once(server, match_after_fork, sse5).await;
+}
+
+async fn mount_second_compact_flow(server: &MockServer) {
+    let sse6 = sse(vec![
+        ev_assistant_message("m4", SUMMARY_TEXT),
+        ev_completed("r6"),
+    ]);
+    let sse7 = sse(vec![ev_completed("r7")]);
+
+    let match_second_compact = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(&format!("\"text\":\"{SUMMARIZE_TRIGGER}\"")) && body.contains("AFTER_FORK")
+    };
+    mount_sse_once(server, match_second_compact, sse6).await;
+
+    let match_after_second_resume = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains(&format!("\"text\":\"{AFTER_SECOND_RESUME}\""))
+    };
+    mount_sse_once(server, match_after_second_resume, sse7).await;
+}
+
+async fn start_test_conversation(
+    server: &MockServer,
+) -> (TempDir, Config, ConversationManager, Arc<CodexConversation>) {
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+    let home = TempDir::new().expect("create temp dir");
+    let mut config = load_default_config_for_test(&home);
+    config.model_provider = model_provider;
+
+    let manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
+    let NewConversation { conversation, .. } = manager
+        .new_conversation(config.clone())
+        .await
+        .expect("create conversation");
+
+    (home, config, manager, conversation)
+}
+
+async fn user_turn(conversation: &Arc<CodexConversation>, text: &str) {
+    conversation
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text { text: text.into() }],
+        })
+        .await
+        .expect("submit user turn");
+    wait_for_event(conversation, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+}
+
+async fn compact_conversation(conversation: &Arc<CodexConversation>) {
+    conversation
+        .submit(Op::Compact)
+        .await
+        .expect("compact conversation");
+    wait_for_event(conversation, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+}
+
+async fn fetch_conversation_path(
+    conversation: &Arc<CodexConversation>,
+    context: &str,
+) -> std::path::PathBuf {
+    conversation
+        .submit(Op::GetPath)
+        .await
+        .expect("request conversation path");
+    match wait_for_event(conversation, |ev| {
+        matches!(ev, EventMsg::ConversationPath(_))
+    })
+    .await
+    {
+        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path,
+        _ => panic!("expected ConversationPath event for {context}"),
+    }
+}
+
+async fn resume_conversation(
+    manager: &ConversationManager,
+    config: &Config,
+    path: std::path::PathBuf,
+) -> Arc<CodexConversation> {
+    let auth_manager =
+        codex_core::AuthManager::from_auth_for_testing(CodexAuth::from_api_key("dummy"));
+    let NewConversation { conversation, .. } = manager
+        .resume_conversation_from_rollout(config.clone(), path, auth_manager)
+        .await
+        .expect("resume conversation");
+    conversation
+}
+
+async fn fork_conversation(
+    manager: &ConversationManager,
+    config: &Config,
+    path: std::path::PathBuf,
+    back_steps: usize,
+) -> Arc<CodexConversation> {
+    let NewConversation { conversation, .. } = manager
+        .fork_conversation(back_steps, config.clone(), path)
+        .await
+        .expect("fork conversation");
+    conversation
+}

codex-rs/core/tests/suite/exec_stream_events.rs

@@ -2,8 +2,11 @@
 
 use std::collections::HashMap;
 use std::path::PathBuf;
+use std::time::Duration;
 
 use async_channel::Receiver;
+use codex_core::error::CodexErr;
+use codex_core::error::SandboxErr;
 use codex_core::exec::ExecParams;
 use codex_core::exec::SandboxType;
 use codex_core::exec::StdoutStream;
@@ -170,3 +173,36 @@ async fn test_aggregated_output_interleaves_in_order() {
     assert_eq!(result.aggregated_output.text, "O1\nE1\nO2\nE2\n");
     assert_eq!(result.aggregated_output.truncated_after_lines, None);
 }
+
+#[tokio::test]
+async fn test_exec_timeout_returns_partial_output() {
+    let cmd = vec![
+        "/bin/sh".to_string(),
+        "-c".to_string(),
+        "printf 'before\\n'; sleep 2; printf 'after\\n'".to_string(),
+    ];
+
+    let params = ExecParams {
+        command: cmd,
+        cwd: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
+        timeout_ms: Some(200),
+        env: HashMap::new(),
+        with_escalated_permissions: None,
+        justification: None,
+    };
+
+    let policy = SandboxPolicy::new_read_only_policy();
+
+    let result = process_exec_tool_call(params, SandboxType::None, &policy, &None, None).await;
+
+    let Err(CodexErr::Sandbox(SandboxErr::Timeout { output })) = result else {
+        panic!("expected timeout error");
+    };
+
+    assert_eq!(output.exit_code, 124);
+    assert_eq!(output.stdout.text, "before\n");
+    assert!(output.stderr.text.is_empty());
+    assert_eq!(output.aggregated_output.text, "before\n");
+    assert!(output.duration >= Duration::from_millis(200));
+    assert!(output.timed_out);
+}

codex-rs/core/tests/suite/fork_conversation.rs

@@ -0,0 +1,195 @@
+use codex_core::CodexAuth;
+use codex_core::ContentItem;
+use codex_core::ConversationManager;
+use codex_core::ModelProviderInfo;
+use codex_core::NewConversation;
+use codex_core::ResponseItem;
+use codex_core::built_in_model_providers;
+use codex_core::protocol::ConversationPathResponseEvent;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::InputItem;
+use codex_core::protocol::Op;
+use codex_core::protocol::RolloutItem;
+use codex_core::protocol::RolloutLine;
+use core_test_support::load_default_config_for_test;
+use core_test_support::wait_for_event;
+use tempfile::TempDir;
+use wiremock::Mock;
+use wiremock::MockServer;
+use wiremock::ResponseTemplate;
+use wiremock::matchers::method;
+use wiremock::matchers::path;
+
+/// Build minimal SSE stream with completed marker using the JSON fixture.
+fn sse_completed(id: &str) -> String {
+    core_test_support::load_sse_fixture_with_id("tests/fixtures/completed_template.json", id)
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn fork_conversation_twice_drops_to_first_message() {
+    // Start a mock server that completes three turns.
+    let server = MockServer::start().await;
+    let sse = sse_completed("resp");
+    let first = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse.clone(), "text/event-stream");
+
+    // Expect three calls to /v1/responses – one per user input.
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(first)
+        .expect(3)
+        .mount(&server)
+        .await;
+
+    // Configure Codex to use the mock server.
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&home);
+    config.model_provider = model_provider.clone();
+    let config_for_fork = config.clone();
+
+    let conversation_manager = ConversationManager::with_auth(CodexAuth::from_api_key("dummy"));
+    let NewConversation {
+        conversation: codex,
+        ..
+    } = conversation_manager
+        .new_conversation(config)
+        .await
+        .expect("create conversation");
+
+    // Send three user messages; wait for three completed turns.
+    for text in ["first", "second", "third"] {
+        codex
+            .submit(Op::UserInput {
+                items: vec![InputItem::Text {
+                    text: text.to_string(),
+                }],
+            })
+            .await
+            .unwrap();
+        let _ = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+    }
+
+    // Request history from the base conversation to obtain rollout path.
+    codex.submit(Op::GetPath).await.unwrap();
+    let base_history =
+        wait_for_event(&codex, |ev| matches!(ev, EventMsg::ConversationPath(_))).await;
+    let base_path = match &base_history {
+        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path.clone(),
+        _ => panic!("expected ConversationHistory event"),
+    };
+
+    // GetHistory flushes before returning the path; no wait needed.
+
+    // Helper: read rollout items (excluding SessionMeta) from a JSONL path.
+    let read_items = |p: &std::path::Path| -> Vec<RolloutItem> {
+        let text = std::fs::read_to_string(p).expect("read rollout file");
+        let mut items: Vec<RolloutItem> = Vec::new();
+        for line in text.lines() {
+            if line.trim().is_empty() {
+                continue;
+            }
+            let v: serde_json::Value = serde_json::from_str(line).expect("jsonl line");
+            let rl: RolloutLine = serde_json::from_value(v).expect("rollout line");
+            match rl.item {
+                RolloutItem::SessionMeta(_) => {}
+                other => items.push(other),
+            }
+        }
+        items
+    };
+
+    // Compute expected prefixes after each fork by truncating base rollout at nth-from-last user input.
+    let base_items = read_items(&base_path);
+    let find_user_input_positions = |items: &[RolloutItem]| -> Vec<usize> {
+        let mut pos = Vec::new();
+        for (i, it) in items.iter().enumerate() {
+            if let RolloutItem::ResponseItem(ResponseItem::Message { role, content, .. }) = it
+                && role == "user"
+            {
+                // Consider any user message as an input boundary; recorder stores both EventMsg and ResponseItem.
+                // We specifically look for input items, which are represented as ContentItem::InputText.
+                if content
+                    .iter()
+                    .any(|c| matches!(c, ContentItem::InputText { .. }))
+                {
+                    pos.push(i);
+                }
+            }
+        }
+        pos
+    };
+    let user_inputs = find_user_input_positions(&base_items);
+
+    // After dropping last user input (n=1), cut strictly before that input if present, else empty.
+    let cut1 = user_inputs
+        .get(user_inputs.len().saturating_sub(1))
+        .copied()
+        .unwrap_or(0);
+    let expected_after_first: Vec<RolloutItem> = base_items[..cut1].to_vec();
+
+    // After dropping again (n=1 on fork1), compute expected relative to fork1's rollout.
+
+    // Fork once with n=1 → drops the last user input and everything after.
+    let NewConversation {
+        conversation: codex_fork1,
+        ..
+    } = conversation_manager
+        .fork_conversation(1, config_for_fork.clone(), base_path.clone())
+        .await
+        .expect("fork 1");
+
+    codex_fork1.submit(Op::GetPath).await.unwrap();
+    let fork1_history = wait_for_event(&codex_fork1, |ev| {
+        matches!(ev, EventMsg::ConversationPath(_))
+    })
+    .await;
+    let fork1_path = match &fork1_history {
+        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path.clone(),
+        _ => panic!("expected ConversationHistory event after first fork"),
+    };
+
+    // GetHistory on fork1 flushed; the file is ready.
+    let fork1_items = read_items(&fork1_path);
+    pretty_assertions::assert_eq!(
+        serde_json::to_value(&fork1_items).unwrap(),
+        serde_json::to_value(&expected_after_first).unwrap()
+    );
+
+    // Fork again with n=1 → drops the (new) last user message, leaving only the first.
+    let NewConversation {
+        conversation: codex_fork2,
+        ..
+    } = conversation_manager
+        .fork_conversation(1, config_for_fork.clone(), fork1_path.clone())
+        .await
+        .expect("fork 2");
+
+    codex_fork2.submit(Op::GetPath).await.unwrap();
+    let fork2_history = wait_for_event(&codex_fork2, |ev| {
+        matches!(ev, EventMsg::ConversationPath(_))
+    })
+    .await;
+    let fork2_path = match &fork2_history {
+        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path.clone(),
+        _ => panic!("expected ConversationHistory event after second fork"),
+    };
+    // GetHistory on fork2 flushed; the file is ready.
+    let fork1_items = read_items(&fork1_path);
+    let fork1_user_inputs = find_user_input_positions(&fork1_items);
+    let cut_last_on_fork1 = fork1_user_inputs
+        .get(fork1_user_inputs.len().saturating_sub(1))
+        .copied()
+        .unwrap_or(0);
+    let expected_after_second: Vec<RolloutItem> = fork1_items[..cut_last_on_fork1].to_vec();
+    let fork2_items = read_items(&fork2_path);
+    pretty_assertions::assert_eq!(
+        serde_json::to_value(&fork2_items).unwrap(),
+        serde_json::to_value(&expected_after_second).unwrap()
+    );
+}

codex-rs/core/tests/suite/mod.rs

@@ -3,10 +3,15 @@
 mod cli_stream;
 mod client;
 mod compact;
+mod compact_resume_fork;
 mod exec;
 mod exec_stream_events;
+mod fork_conversation;
 mod live_cli;
+mod model_overrides;
 mod prompt_caching;
+mod review;
+mod rollout_list_find;
 mod seatbelt;
 mod stream_error_allows_next_turn;
 mod stream_no_completed;

codex-rs/core/tests/suite/model_overrides.rs

@@ -0,0 +1,92 @@
+use codex_core::CodexAuth;
+use codex_core::ConversationManager;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::Op;
+use codex_core::protocol_config_types::ReasoningEffort;
+use core_test_support::load_default_config_for_test;
+use core_test_support::wait_for_event;
+use pretty_assertions::assert_eq;
+use tempfile::TempDir;
+
+const CONFIG_TOML: &str = "config.toml";
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn override_turn_context_does_not_persist_when_config_exists() {
+    let codex_home = TempDir::new().unwrap();
+    let config_path = codex_home.path().join(CONFIG_TOML);
+    let initial_contents = "model = \"gpt-4o\"\n";
+    tokio::fs::write(&config_path, initial_contents)
+        .await
+        .expect("seed config.toml");
+
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model = "gpt-4o".to_string();
+
+    let conversation_manager =
+        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
+    let codex = conversation_manager
+        .new_conversation(config)
+        .await
+        .expect("create conversation")
+        .conversation;
+
+    codex
+        .submit(Op::OverrideTurnContext {
+            cwd: None,
+            approval_policy: None,
+            sandbox_policy: None,
+            model: Some("o3".to_string()),
+            effort: Some(Some(ReasoningEffort::High)),
+            summary: None,
+        })
+        .await
+        .expect("submit override");
+
+    codex.submit(Op::Shutdown).await.expect("request shutdown");
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::ShutdownComplete)).await;
+
+    let contents = tokio::fs::read_to_string(&config_path)
+        .await
+        .expect("read config.toml after override");
+    assert_eq!(contents, initial_contents);
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn override_turn_context_does_not_create_config_file() {
+    let codex_home = TempDir::new().unwrap();
+    let config_path = codex_home.path().join(CONFIG_TOML);
+    assert!(
+        !config_path.exists(),
+        "test setup should start without config"
+    );
+
+    let config = load_default_config_for_test(&codex_home);
+
+    let conversation_manager =
+        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
+    let codex = conversation_manager
+        .new_conversation(config)
+        .await
+        .expect("create conversation")
+        .conversation;
+
+    codex
+        .submit(Op::OverrideTurnContext {
+            cwd: None,
+            approval_policy: None,
+            sandbox_policy: None,
+            model: Some("o3".to_string()),
+            effort: Some(Some(ReasoningEffort::Medium)),
+            summary: None,
+        })
+        .await
+        .expect("submit override");
+
+    codex.submit(Op::Shutdown).await.expect("request shutdown");
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::ShutdownComplete)).await;
+
+    assert!(
+        !config_path.exists(),
+        "override should not create config.toml"
+    );
+}

codex-rs/core/tests/suite/prompt_caching.rs

@@ -1,5 +1,6 @@
 #![allow(clippy::unwrap_used)]
 
+use codex_core::CodexAuth;
 use codex_core::ConversationManager;
 use codex_core::ModelProviderInfo;
 use codex_core::built_in_model_providers;
@@ -12,7 +13,6 @@ use codex_core::protocol::SandboxPolicy;
 use codex_core::protocol_config_types::ReasoningEffort;
 use codex_core::protocol_config_types::ReasoningSummary;
 use codex_core::shell::default_user_shell;
-use codex_login::CodexAuth;
 use core_test_support::load_default_config_for_test;
 use core_test_support::load_sse_fixture_with_id;
 use core_test_support::wait_for_event;
@@ -289,20 +289,17 @@ async fn prefixes_context_and_instructions_once_and_consistently_across_requests
 
     let expected_env_msg = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": expected_env_text } ]
     });
     let expected_ui_msg = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": expected_ui_text } ]
     });
 
     let expected_user_message_1 = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": "hello 1" } ]
     });
@@ -314,7 +311,6 @@ async fn prefixes_context_and_instructions_once_and_consistently_across_requests
 
     let expected_user_message_2 = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": "hello 2" } ]
     });
@@ -391,7 +387,7 @@ async fn overrides_turn_context_but_keeps_cached_prefix_and_key_constant() {
                 exclude_slash_tmp: true,
             }),
             model: Some("o3".to_string()),
-            effort: Some(ReasoningEffort::High),
+            effort: Some(Some(ReasoningEffort::High)),
             summary: Some(ReasoningSummary::Detailed),
         })
         .await
@@ -424,21 +420,25 @@ async fn overrides_turn_context_but_keeps_cached_prefix_and_key_constant() {
     // as the prefix of the second request, ensuring cache hit potential.
     let expected_user_message_2 = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": "hello 2" } ]
     });
     // After overriding the turn context, the environment context should be emitted again
     // reflecting the new approval policy and sandbox settings. Omit cwd because it did
     // not change.
-    let expected_env_text_2 = r#"<environment_context>
+    let expected_env_text_2 = format!(
+        r#"<environment_context>
   <approval_policy>never</approval_policy>
   <sandbox_mode>workspace-write</sandbox_mode>
   <network_access>enabled</network_access>
-</environment_context>"#;
+  <writable_roots>
+    <root>{}</root>
+  </writable_roots>
+</environment_context>"#,
+        writable.path().to_string_lossy()
+    );
     let expected_env_msg_2 = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": expected_env_text_2 } ]
     });
@@ -519,7 +519,7 @@ async fn per_turn_overrides_keep_cached_prefix_and_key_constant() {
                 exclude_slash_tmp: true,
             },
             model: "o3".to_string(),
-            effort: ReasoningEffort::High,
+            effort: Some(ReasoningEffort::High),
             summary: ReasoningSummary::Detailed,
         })
         .await
@@ -543,7 +543,6 @@ async fn per_turn_overrides_keep_cached_prefix_and_key_constant() {
     // as the prefix of the second request.
     let expected_user_message_2 = serde_json::json!({
         "type": "message",
-        "id": serde_json::Value::Null,
         "role": "user",
         "content": [ { "type": "input_text", "text": "hello 2" } ]
     });

codex-rs/core/tests/suite/review.rs

@@ -0,0 +1,598 @@
+use codex_core::CodexAuth;
+use codex_core::CodexConversation;
+use codex_core::ConversationManager;
+use codex_core::ModelProviderInfo;
+use codex_core::built_in_model_providers;
+use codex_core::config::Config;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::ExitedReviewModeEvent;
+use codex_core::protocol::InputItem;
+use codex_core::protocol::Op;
+use codex_core::protocol::ReviewCodeLocation;
+use codex_core::protocol::ReviewFinding;
+use codex_core::protocol::ReviewLineRange;
+use codex_core::protocol::ReviewOutputEvent;
+use codex_core::protocol::ReviewRequest;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use core_test_support::load_default_config_for_test;
+use core_test_support::load_sse_fixture_with_id_from_str;
+use core_test_support::wait_for_event;
+use pretty_assertions::assert_eq;
+use std::path::PathBuf;
+use std::sync::Arc;
+use tempfile::TempDir;
+use tokio::io::AsyncWriteExt as _;
+use uuid::Uuid;
+use wiremock::Mock;
+use wiremock::MockServer;
+use wiremock::ResponseTemplate;
+use wiremock::matchers::method;
+use wiremock::matchers::path;
+
+/// Verify that submitting `Op::Review` spawns a child task and emits
+/// EnteredReviewMode -> ExitedReviewMode(None) -> TaskComplete
+/// in that order when the model returns a structured review JSON payload.
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn review_op_emits_lifecycle_and_review_output() {
+    // Skip under Codex sandbox network restrictions.
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // Start mock Responses API server. Return a single assistant message whose
+    // text is a JSON-encoded ReviewOutputEvent.
+    let review_json = serde_json::json!({
+        "findings": [
+            {
+                "title": "Prefer Stylize helpers",
+                "body": "Use .dim()/.bold() chaining instead of manual Style where possible.",
+                "confidence_score": 0.9,
+                "priority": 1,
+                "code_location": {
+                    "absolute_file_path": "/tmp/file.rs",
+                    "line_range": {"start": 10, "end": 20}
+                }
+            }
+        ],
+        "overall_correctness": "good",
+        "overall_explanation": "All good with some improvements suggested.",
+        "overall_confidence_score": 0.8
+    })
+    .to_string();
+    let sse_template = r#"[
+            {"type":"response.output_item.done", "item":{
+                "type":"message", "role":"assistant",
+                "content":[{"type":"output_text","text":__REVIEW__}]
+            }},
+            {"type":"response.completed", "response": {"id": "__ID__"}}
+        ]"#;
+    let review_json_escaped = serde_json::to_string(&review_json).unwrap();
+    let sse_raw = sse_template.replace("__REVIEW__", &review_json_escaped);
+    let server = start_responses_server_with_sse(&sse_raw, 1).await;
+    let codex_home = TempDir::new().unwrap();
+    let codex = new_conversation_for_server(&server, &codex_home, |_| {}).await;
+
+    // Submit review request.
+    codex
+        .submit(Op::Review {
+            review_request: ReviewRequest {
+                prompt: "Please review my changes".to_string(),
+                user_facing_hint: "my changes".to_string(),
+            },
+        })
+        .await
+        .unwrap();
+
+    // Verify lifecycle: Entered -> Exited(Some(review)) -> TaskComplete.
+    let _entered = wait_for_event(&codex, |ev| matches!(ev, EventMsg::EnteredReviewMode(_))).await;
+    let closed = wait_for_event(&codex, |ev| matches!(ev, EventMsg::ExitedReviewMode(_))).await;
+    let review = match closed {
+        EventMsg::ExitedReviewMode(ev) => ev
+            .review_output
+            .expect("expected ExitedReviewMode with Some(review_output)"),
+        other => panic!("expected ExitedReviewMode(..), got {other:?}"),
+    };
+
+    // Deep compare full structure using PartialEq (floats are f32 on both sides).
+    let expected = ReviewOutputEvent {
+        findings: vec![ReviewFinding {
+            title: "Prefer Stylize helpers".to_string(),
+            body: "Use .dim()/.bold() chaining instead of manual Style where possible.".to_string(),
+            confidence_score: 0.9,
+            priority: 1,
+            code_location: ReviewCodeLocation {
+                absolute_file_path: PathBuf::from("/tmp/file.rs"),
+                line_range: ReviewLineRange { start: 10, end: 20 },
+            },
+        }],
+        overall_correctness: "good".to_string(),
+        overall_explanation: "All good with some improvements suggested.".to_string(),
+        overall_confidence_score: 0.8,
+    };
+    assert_eq!(expected, review);
+    let _complete = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    server.verify().await;
+}
+
+/// When the model returns plain text that is not JSON, ensure the child
+/// lifecycle still occurs and the plain text is surfaced via
+/// ExitedReviewMode(Some(..)) as the overall_explanation.
+// Windows CI only: bump to 4 workers to prevent SSE/event starvation and test timeouts.
+#[cfg_attr(windows, tokio::test(flavor = "multi_thread", worker_threads = 4))]
+#[cfg_attr(not(windows), tokio::test(flavor = "multi_thread", worker_threads = 2))]
+async fn review_op_with_plain_text_emits_review_fallback() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse_raw = r#"[
+        {"type":"response.output_item.done", "item":{
+            "type":"message", "role":"assistant",
+            "content":[{"type":"output_text","text":"just plain text"}]
+        }},
+        {"type":"response.completed", "response": {"id": "__ID__"}}
+    ]"#;
+    let server = start_responses_server_with_sse(sse_raw, 1).await;
+    let codex_home = TempDir::new().unwrap();
+    let codex = new_conversation_for_server(&server, &codex_home, |_| {}).await;
+
+    codex
+        .submit(Op::Review {
+            review_request: ReviewRequest {
+                prompt: "Plain text review".to_string(),
+                user_facing_hint: "plain text review".to_string(),
+            },
+        })
+        .await
+        .unwrap();
+
+    let _entered = wait_for_event(&codex, |ev| matches!(ev, EventMsg::EnteredReviewMode(_))).await;
+    let closed = wait_for_event(&codex, |ev| matches!(ev, EventMsg::ExitedReviewMode(_))).await;
+    let review = match closed {
+        EventMsg::ExitedReviewMode(ev) => ev
+            .review_output
+            .expect("expected ExitedReviewMode with Some(review_output)"),
+        other => panic!("expected ExitedReviewMode(..), got {other:?}"),
+    };
+
+    // Expect a structured fallback carrying the plain text.
+    let expected = ReviewOutputEvent {
+        overall_explanation: "just plain text".to_string(),
+        ..Default::default()
+    };
+    assert_eq!(expected, review);
+    let _complete = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    server.verify().await;
+}
+
+/// When the model returns structured JSON in a review, ensure no AgentMessage
+/// is emitted; the UI consumes the structured result via ExitedReviewMode.
+// Windows CI only: bump to 4 workers to prevent SSE/event starvation and test timeouts.
+#[cfg_attr(windows, tokio::test(flavor = "multi_thread", worker_threads = 4))]
+#[cfg_attr(not(windows), tokio::test(flavor = "multi_thread", worker_threads = 2))]
+async fn review_does_not_emit_agent_message_on_structured_output() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let review_json = serde_json::json!({
+        "findings": [
+            {
+                "title": "Example",
+                "body": "Structured review output.",
+                "confidence_score": 0.5,
+                "priority": 1,
+                "code_location": {
+                    "absolute_file_path": "/tmp/file.rs",
+                    "line_range": {"start": 1, "end": 2}
+                }
+            }
+        ],
+        "overall_correctness": "ok",
+        "overall_explanation": "ok",
+        "overall_confidence_score": 0.5
+    })
+    .to_string();
+    let sse_template = r#"[
+            {"type":"response.output_item.done", "item":{
+                "type":"message", "role":"assistant",
+                "content":[{"type":"output_text","text":__REVIEW__}]
+            }},
+            {"type":"response.completed", "response": {"id": "__ID__"}}
+        ]"#;
+    let review_json_escaped = serde_json::to_string(&review_json).unwrap();
+    let sse_raw = sse_template.replace("__REVIEW__", &review_json_escaped);
+    let server = start_responses_server_with_sse(&sse_raw, 1).await;
+    let codex_home = TempDir::new().unwrap();
+    let codex = new_conversation_for_server(&server, &codex_home, |_| {}).await;
+
+    codex
+        .submit(Op::Review {
+            review_request: ReviewRequest {
+                prompt: "check structured".to_string(),
+                user_facing_hint: "check structured".to_string(),
+            },
+        })
+        .await
+        .unwrap();
+
+    // Drain events until TaskComplete; ensure none are AgentMessage.
+    use tokio::time::Duration;
+    use tokio::time::timeout;
+    let mut saw_entered = false;
+    let mut saw_exited = false;
+    loop {
+        let ev = timeout(Duration::from_secs(5), codex.next_event())
+            .await
+            .expect("timeout waiting for event")
+            .expect("stream ended unexpectedly");
+        match ev.msg {
+            EventMsg::TaskComplete(_) => break,
+            EventMsg::AgentMessage(_) => {
+                panic!("unexpected AgentMessage during review with structured output")
+            }
+            EventMsg::EnteredReviewMode(_) => saw_entered = true,
+            EventMsg::ExitedReviewMode(_) => saw_exited = true,
+            _ => {}
+        }
+    }
+    assert!(saw_entered && saw_exited, "missing review lifecycle events");
+
+    server.verify().await;
+}
+
+/// Ensure that when a custom `review_model` is set in the config, the review
+/// request uses that model (and not the main chat model).
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn review_uses_custom_review_model_from_config() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // Minimal stream: just a completed event
+    let sse_raw = r#"[
+        {"type":"response.completed", "response": {"id": "__ID__"}}
+    ]"#;
+    let server = start_responses_server_with_sse(sse_raw, 1).await;
+    let codex_home = TempDir::new().unwrap();
+    // Choose a review model different from the main model; ensure it is used.
+    let codex = new_conversation_for_server(&server, &codex_home, |cfg| {
+        cfg.model = "gpt-4.1".to_string();
+        cfg.review_model = "gpt-5".to_string();
+    })
+    .await;
+
+    codex
+        .submit(Op::Review {
+            review_request: ReviewRequest {
+                prompt: "use custom model".to_string(),
+                user_facing_hint: "use custom model".to_string(),
+            },
+        })
+        .await
+        .unwrap();
+
+    // Wait for completion
+    let _entered = wait_for_event(&codex, |ev| matches!(ev, EventMsg::EnteredReviewMode(_))).await;
+    let _closed = wait_for_event(&codex, |ev| {
+        matches!(
+            ev,
+            EventMsg::ExitedReviewMode(ExitedReviewModeEvent {
+                review_output: None
+            })
+        )
+    })
+    .await;
+    let _complete = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // Assert the request body model equals the configured review model
+    let request = &server.received_requests().await.unwrap()[0];
+    let body = request.body_json::<serde_json::Value>().unwrap();
+    assert_eq!(body["model"].as_str().unwrap(), "gpt-5");
+
+    server.verify().await;
+}
+
+/// When a review session begins, it must not prepend prior chat history from
+/// the parent session. The request `input` should contain only the review
+/// prompt from the user.
+// Windows CI only: bump to 4 workers to prevent SSE/event starvation and test timeouts.
+#[cfg_attr(windows, tokio::test(flavor = "multi_thread", worker_threads = 4))]
+#[cfg_attr(not(windows), tokio::test(flavor = "multi_thread", worker_threads = 2))]
+async fn review_input_isolated_from_parent_history() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // Mock server for the single review request
+    let sse_raw = r#"[
+        {"type":"response.completed", "response": {"id": "__ID__"}}
+    ]"#;
+    let server = start_responses_server_with_sse(sse_raw, 1).await;
+
+    // Seed a parent session history via resume file with both user + assistant items.
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let session_file = codex_home.path().join("resume.jsonl");
+    {
+        let mut f = tokio::fs::File::create(&session_file).await.unwrap();
+        let convo_id = Uuid::new_v4();
+        // Proper session_meta line (enveloped) with a conversation id
+        let meta_line = serde_json::json!({
+            "timestamp": "2024-01-01T00:00:00.000Z",
+            "type": "session_meta",
+            "payload": {
+                "id": convo_id,
+                "timestamp": "2024-01-01T00:00:00Z",
+                "instructions": null,
+                "cwd": ".",
+                "originator": "test_originator",
+                "cli_version": "test_version"
+            }
+        });
+        f.write_all(format!("{meta_line}\n").as_bytes())
+            .await
+            .unwrap();
+
+        // Prior user message (enveloped response_item)
+        let user = codex_protocol::models::ResponseItem::Message {
+            id: None,
+            role: "user".to_string(),
+            content: vec![codex_protocol::models::ContentItem::InputText {
+                text: "parent: earlier user message".to_string(),
+            }],
+        };
+        let user_json = serde_json::to_value(&user).unwrap();
+        let user_line = serde_json::json!({
+            "timestamp": "2024-01-01T00:00:01.000Z",
+            "type": "response_item",
+            "payload": user_json
+        });
+        f.write_all(format!("{user_line}\n").as_bytes())
+            .await
+            .unwrap();
+
+        // Prior assistant message (enveloped response_item)
+        let assistant = codex_protocol::models::ResponseItem::Message {
+            id: None,
+            role: "assistant".to_string(),
+            content: vec![codex_protocol::models::ContentItem::OutputText {
+                text: "parent: assistant reply".to_string(),
+            }],
+        };
+        let assistant_json = serde_json::to_value(&assistant).unwrap();
+        let assistant_line = serde_json::json!({
+            "timestamp": "2024-01-01T00:00:02.000Z",
+            "type": "response_item",
+            "payload": assistant_json
+        });
+        f.write_all(format!("{assistant_line}\n").as_bytes())
+            .await
+            .unwrap();
+    }
+    let codex =
+        resume_conversation_for_server(&server, &codex_home, session_file.clone(), |_| {}).await;
+
+    // Submit review request; it must start fresh (no parent history in `input`).
+    let review_prompt = "Please review only this".to_string();
+    codex
+        .submit(Op::Review {
+            review_request: ReviewRequest {
+                prompt: review_prompt.clone(),
+                user_facing_hint: review_prompt.clone(),
+            },
+        })
+        .await
+        .unwrap();
+
+    let _entered = wait_for_event(&codex, |ev| matches!(ev, EventMsg::EnteredReviewMode(_))).await;
+    let _closed = wait_for_event(&codex, |ev| {
+        matches!(
+            ev,
+            EventMsg::ExitedReviewMode(ExitedReviewModeEvent {
+                review_output: None
+            })
+        )
+    })
+    .await;
+    let _complete = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // Assert the request `input` contains only the single review user message.
+    let request = &server.received_requests().await.unwrap()[0];
+    let body = request.body_json::<serde_json::Value>().unwrap();
+    let expected_input = serde_json::json!([
+        {
+            "type": "message",
+            "role": "user",
+            "content": [{"type": "input_text", "text": review_prompt}]
+        }
+    ]);
+    assert_eq!(body["input"], expected_input);
+
+    server.verify().await;
+}
+
+/// After a review thread finishes, its conversation should not leak into the
+/// parent session. A subsequent parent turn must not include any review
+/// messages in its request `input`.
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn review_history_does_not_leak_into_parent_session() {
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // Respond to both the review request and the subsequent parent request.
+    let sse_raw = r#"[
+        {"type":"response.output_item.done", "item":{
+            "type":"message", "role":"assistant",
+            "content":[{"type":"output_text","text":"review assistant output"}]
+        }},
+        {"type":"response.completed", "response": {"id": "__ID__"}}
+    ]"#;
+    let server = start_responses_server_with_sse(sse_raw, 2).await;
+    let codex_home = TempDir::new().unwrap();
+    let codex = new_conversation_for_server(&server, &codex_home, |_| {}).await;
+
+    // 1) Run a review turn that produces an assistant message (isolated in child).
+    codex
+        .submit(Op::Review {
+            review_request: ReviewRequest {
+                prompt: "Start a review".to_string(),
+                user_facing_hint: "Start a review".to_string(),
+            },
+        })
+        .await
+        .unwrap();
+    let _entered = wait_for_event(&codex, |ev| matches!(ev, EventMsg::EnteredReviewMode(_))).await;
+    let _closed = wait_for_event(&codex, |ev| {
+        matches!(
+            ev,
+            EventMsg::ExitedReviewMode(ExitedReviewModeEvent {
+                review_output: Some(_)
+            })
+        )
+    })
+    .await;
+    let _complete = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // 2) Continue in the parent session; request input must not include any review items.
+    let followup = "back to parent".to_string();
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: followup.clone(),
+            }],
+        })
+        .await
+        .unwrap();
+    let _complete = wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // Inspect the second request (parent turn) input contents.
+    // Parent turns include session initial messages (user_instructions, environment_context).
+    // Critically, no messages from the review thread should appear.
+    let requests = server.received_requests().await.unwrap();
+    assert_eq!(requests.len(), 2);
+    let body = requests[1].body_json::<serde_json::Value>().unwrap();
+    let input = body["input"].as_array().expect("input array");
+
+    // Must include the followup as the last item for this turn
+    let last = input.last().expect("at least one item in input");
+    assert_eq!(last["role"].as_str().unwrap(), "user");
+    let last_text = last["content"][0]["text"].as_str().unwrap();
+    assert_eq!(last_text, followup);
+
+    // Ensure no review-thread content leaked into the parent request
+    let contains_review_prompt = input
+        .iter()
+        .any(|msg| msg["content"][0]["text"].as_str().unwrap_or_default() == "Start a review");
+    let contains_review_assistant = input.iter().any(|msg| {
+        msg["content"][0]["text"].as_str().unwrap_or_default() == "review assistant output"
+    });
+    assert!(
+        !contains_review_prompt,
+        "review prompt leaked into parent turn input"
+    );
+    assert!(
+        !contains_review_assistant,
+        "review assistant output leaked into parent turn input"
+    );
+
+    server.verify().await;
+}
+
+/// Start a mock Responses API server and mount the given SSE stream body.
+async fn start_responses_server_with_sse(sse_raw: &str, expected_requests: usize) -> MockServer {
+    let server = MockServer::start().await;
+    let sse = load_sse_fixture_with_id_from_str(sse_raw, &Uuid::new_v4().to_string());
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(
+            ResponseTemplate::new(200)
+                .insert_header("content-type", "text/event-stream")
+                .set_body_raw(sse.clone(), "text/event-stream"),
+        )
+        .expect(expected_requests as u64)
+        .mount(&server)
+        .await;
+    server
+}
+
+/// Create a conversation configured to talk to the provided mock server.
+#[expect(clippy::expect_used)]
+async fn new_conversation_for_server<F>(
+    server: &MockServer,
+    codex_home: &TempDir,
+    mutator: F,
+) -> Arc<CodexConversation>
+where
+    F: FnOnce(&mut Config),
+{
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+    let mut config = load_default_config_for_test(codex_home);
+    config.model_provider = model_provider;
+    mutator(&mut config);
+    let conversation_manager =
+        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
+    conversation_manager
+        .new_conversation(config)
+        .await
+        .expect("create conversation")
+        .conversation
+}
+
+/// Create a conversation resuming from a rollout file, configured to talk to the provided mock server.
+#[expect(clippy::expect_used)]
+async fn resume_conversation_for_server<F>(
+    server: &MockServer,
+    codex_home: &TempDir,
+    resume_path: std::path::PathBuf,
+    mutator: F,
+) -> Arc<CodexConversation>
+where
+    F: FnOnce(&mut Config),
+{
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+    let mut config = load_default_config_for_test(codex_home);
+    config.model_provider = model_provider;
+    mutator(&mut config);
+    let conversation_manager =
+        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
+    let auth_manager =
+        codex_core::AuthManager::from_auth_for_testing(CodexAuth::from_api_key("Test API Key"));
+    conversation_manager
+        .resume_conversation_from_rollout(config, resume_path, auth_manager)
+        .await
+        .expect("resume conversation")
+        .conversation
+}

codex-rs/core/tests/suite/rollout_list_find.rs

@@ -0,0 +1,50 @@
+#![allow(clippy::unwrap_used, clippy::expect_used)]
+use std::io::Write;
+use std::path::PathBuf;
+
+use codex_core::find_conversation_path_by_id_str;
+use tempfile::TempDir;
+use uuid::Uuid;
+
+/// Create sessions/YYYY/MM/DD and write a minimal rollout file containing the
+/// provided conversation id in the SessionMeta line. Returns the absolute path.
+fn write_minimal_rollout_with_id(codex_home: &TempDir, id: Uuid) -> PathBuf {
+    let sessions = codex_home.path().join("sessions/2024/01/01");
+    std::fs::create_dir_all(&sessions).unwrap();
+
+    let file = sessions.join(format!("rollout-2024-01-01T00-00-00-{id}.jsonl"));
+    let mut f = std::fs::File::create(&file).unwrap();
+    // Minimal first line: session_meta with the id so content search can find it
+    writeln!(
+        f,
+        "{}",
+        serde_json::json!({
+            "timestamp": "2024-01-01T00:00:00.000Z",
+            "type": "session_meta",
+            "payload": {
+                "id": id,
+                "timestamp": "2024-01-01T00:00:00Z",
+                "instructions": null,
+                "cwd": ".",
+                "originator": "test",
+                "cli_version": "test"
+            }
+        })
+    )
+    .unwrap();
+
+    file
+}
+
+#[tokio::test]
+async fn find_locates_rollout_file_by_id() {
+    let home = TempDir::new().unwrap();
+    let id = Uuid::new_v4();
+    let expected = write_minimal_rollout_with_id(&home, id);
+
+    let found = find_conversation_path_by_id_str(home.path(), &id.to_string())
+        .await
+        .unwrap();
+
+    assert_eq!(found.unwrap(), expected);
+}

codex-rs/core/tests/suite/seatbelt.rs

@@ -159,6 +159,41 @@ async fn read_only_forbids_all_writes() {
         .await;
 }
 
+/// Verify that user lookups via `pwd.getpwuid(os.getuid())` work under the
+/// seatbelt sandbox. Prior to allowing the necessary mach‑lookup for
+/// OpenDirectory libinfo, this would fail with `KeyError: getpwuid(): uid not found`.
+#[tokio::test]
+async fn python_getpwuid_works_under_seatbelt() {
+    if std::env::var(CODEX_SANDBOX_ENV_VAR) == Ok("seatbelt".to_string()) {
+        eprintln!("{CODEX_SANDBOX_ENV_VAR} is set to 'seatbelt', skipping test.");
+        return;
+    }
+
+    // ReadOnly is sufficient here since we are only exercising user lookup.
+    let policy = SandboxPolicy::ReadOnly;
+
+    let mut child = spawn_command_under_seatbelt(
+        vec![
+            "python3".to_string(),
+            "-c".to_string(),
+            // Print the passwd struct; success implies lookup worked.
+            "import pwd, os; print(pwd.getpwuid(os.getuid()))".to_string(),
+        ],
+        &policy,
+        std::env::current_dir().expect("should be able to get current dir"),
+        StdioPolicy::RedirectForShellTool,
+        HashMap::new(),
+    )
+    .await
+    .expect("should be able to spawn python under seatbelt");
+
+    let status = child
+        .wait()
+        .await
+        .expect("should be able to wait for child process");
+    assert!(status.success(), "python exited with {status:?}");
+}
+
 #[expect(clippy::expect_used)]
 fn create_test_scenario(tmp: &TempDir) -> TestScenario {
     let repo_parent = tmp.path().to_path_buf();

codex-rs/core/tests/suite/stream_error_allows_next_turn.rs

@@ -1,5 +1,6 @@
 use std::time::Duration;
 
+use codex_core::CodexAuth;
 use codex_core::ConversationManager;
 use codex_core::ModelProviderInfo;
 use codex_core::WireApi;
@@ -7,7 +8,6 @@ use codex_core::protocol::EventMsg;
 use codex_core::protocol::InputItem;
 use codex_core::protocol::Op;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
-use codex_login::CodexAuth;
 use core_test_support::load_default_config_for_test;
 use core_test_support::load_sse_fixture_with_id;
 use core_test_support::wait_for_event_with_timeout;