[codex] cover emited events by otel with tests

Hopefully fixes incorrectly showing ^J instead of Shift+Enter in the key
hints occasionally.

.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-release.yml

@@ -167,6 +167,12 @@ jobs:
     needs: build
     name: release
     runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      actions: read
+    outputs:
+      version: ${{ steps.release_name.outputs.name }}
+      tag: ${{ github.ref_name }}
 
     steps:
       - name: Checkout repository
@@ -220,6 +226,48 @@ jobs:
           tag: ${{ github.ref_name }}
           config: .github/dotslash-config.json
 
+  # Publish to npm using OIDC authentication.
+  # July 31, 2025: https://github.blog/changelog/2025-07-31-npm-trusted-publishing-with-oidc-is-generally-available/
+  # npm docs: https://docs.npmjs.com/trusted-publishers
+  publish-npm:
+    # Skip this step for pre-releases (alpha/beta).
+    if: ${{ !contains(needs.release.outputs.version, '-') }}
+    name: publish-npm
+    needs: release
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write # Required for OIDC
+      contents: read
+
+    steps:
+      - name: Setup Node.js
+        uses: actions/setup-node@v5
+        with:
+          node-version: 22
+          registry-url: "https://registry.npmjs.org"
+          scope: "@openai"
+
+      # Trusted publishing requires npm CLI version 11.5.1 or later.
+      - name: Update npm
+        run: npm install -g npm@latest
+
+      - name: Download npm tarball from release
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -euo pipefail
+          version="${{ needs.release.outputs.version }}"
+          tag="${{ needs.release.outputs.tag }}"
+          mkdir -p dist/npm
+          gh release download "$tag" \
+            --repo "${GITHUB_REPOSITORY}" \
+            --pattern "codex-npm-${version}.tgz" \
+            --dir dist/npm
+
+      # No NODE_AUTH_TOKEN needed because we use OIDC.
+      - name: Publish to npm
+        run: npm publish "${GITHUB_WORKSPACE}/dist/npm/codex-npm-${{ needs.release.outputs.version }}.tgz"
+
   update-branch:
     name: Update latest-alpha-cli branch
     permissions:

AGENTS.md

@@ -4,6 +4,7 @@ In the codex-rs folder where the rust code lives:
 
 - Crate names are prefixed with `codex-`. For example, the `core` folder's crate is named `codex-core`
 - When using format! and you can inline variables into {}, always do that.
+- Install any commands the repo relies on (for example `just`, `rg`, or `cargo-insta`) if they aren't already available before running instructions here.
 - Never add or modify any code related to `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` or `CODEX_SANDBOX_ENV_VAR`.
   - 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.

codex-cli/package.json

@@ -15,7 +15,8 @@
   ],
   "repository": {
     "type": "git",
-    "url": "git+https://github.com/openai/codex.git"
+    "url": "git+https://github.com/openai/codex.git",
+    "directory": "codex-cli"
   },
   "dependencies": {
     "@vscode/ripgrep": "^1.15.14"

codex-rs/Cargo.toml

@@ -17,6 +17,7 @@ members = [
     "ollama",
     "protocol",
     "protocol-ts",
+    "otel",
     "tui",
 ]
 resolver = "2"

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

@@ -40,6 +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 {
@@ -93,10 +98,12 @@ pub struct ApplyPatchArgs {
 
 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 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) {
@@ -207,6 +214,26 @@ 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,
@@ -875,6 +902,28 @@ mod tests {
         ));
     }
 
+    #[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(&[

codex-rs/cli/Cargo.toml

@@ -36,5 +36,11 @@ tokio = { version = "1", features = [
     "signal",
 ] }
 tracing = "0.1.41"
-tracing-subscriber = "0.3.19"
+tracing-subscriber = "0.3.20"
 codex-protocol-ts = { path = "../protocol-ts" }
+
+[dev-dependencies]
+assert_cmd = "2"
+predicates = "3"
+pretty_assertions = "1"
+tempfile = "3"

codex-rs/cli/src/debug_sandbox.rs

@@ -64,7 +64,6 @@ async fn run_command_under_sandbox(
     sandbox_type: SandboxType,
 ) -> anyhow::Result<()> {
     let sandbox_mode = create_sandbox_mode(full_auto);
-    let cwd = std::env::current_dir()?;
     let config = Config::load_with_cli_overrides(
         config_overrides
             .parse_overrides()
@@ -75,13 +74,29 @@ async fn run_command_under_sandbox(
             ..Default::default()
         },
     )?;
+
+    // In practice, this should be `std::env::current_dir()` because this CLI
+    // does not support `--cwd`, but let's use the config value for consistency.
+    let cwd = config.cwd.clone();
+    // For now, we always use the same cwd for both the command and the
+    // sandbox policy. In the future, we could add a CLI option to set them
+    // separately.
+    let sandbox_policy_cwd = cwd.clone();
+
     let stdio_policy = StdioPolicy::Inherit;
     let env = create_env(&config.shell_environment_policy);
 
     let mut child = match sandbox_type {
         SandboxType::Seatbelt => {
-            spawn_command_under_seatbelt(command, &config.sandbox_policy, cwd, stdio_policy, env)
-                .await?
+            spawn_command_under_seatbelt(
+                command,
+                cwd,
+                &config.sandbox_policy,
+                sandbox_policy_cwd.as_path(),
+                stdio_policy,
+                env,
+            )
+            .await?
         }
         SandboxType::Landlock => {
             #[expect(clippy::expect_used)]
@@ -91,8 +106,9 @@ async fn run_command_under_sandbox(
             spawn_command_under_linux_sandbox(
                 codex_linux_sandbox_exe,
                 command,
-                &config.sandbox_policy,
                 cwd,
+                &config.sandbox_policy,
+                sandbox_policy_cwd.as_path(),
                 stdio_policy,
                 env,
             )

codex-rs/cli/src/main.rs

@@ -17,6 +17,9 @@ use codex_exec::Cli as ExecCli;
 use codex_tui::Cli as TuiCli;
 use std::path::PathBuf;
 
+mod mcp_cmd;
+
+use crate::mcp_cmd::McpCli;
 use crate::proto::ProtoCli;
 
 /// Codex CLI
@@ -56,8 +59,8 @@ enum Subcommand {
     /// Remove stored authentication credentials.
     Logout(LogoutCommand),
 
-    /// Experimental: run Codex as an MCP server.
-    Mcp,
+    /// [experimental] Run Codex as an MCP server and manage MCP servers.
+    Mcp(McpCli),
 
     /// Run the Protocol stream via stdin/stdout
     #[clap(visible_alias = "p")]
@@ -73,6 +76,9 @@ 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 +91,21 @@ 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,
+
+    #[clap(flatten)]
+    config_overrides: TuiCli,
+}
+
 #[derive(Debug, Parser)]
 struct DebugArgs {
     #[command(subcommand)]
@@ -143,26 +164,57 @@ 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?;
-            if !usage.is_zero() {
-                println!("{}", codex_core::protocol::FinalOutput::from(usage));
+            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.token_usage.is_zero() {
+                println!(
+                    "{}",
+                    codex_core::protocol::FinalOutput::from(usage.token_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?;
+        Some(Subcommand::Mcp(mut mcp_cli)) => {
+            // Propagate any root-level config overrides (e.g. `-c key=value`).
+            prepend_config_flags(&mut mcp_cli.config_overrides, root_config_overrides.clone());
+            mcp_cli.run(codex_linux_sandbox_exe).await?;
+        }
+        Some(Subcommand::Resume(ResumeCommand {
+            session_id,
+            last,
+            config_overrides,
+        })) => {
+            interactive = finalize_resume_interactive(
+                interactive,
+                root_config_overrides.clone(),
+                session_id,
+                last,
+                config_overrides,
+            );
+            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 +229,17 @@ 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 +247,10 @@ 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 +258,10 @@ 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 +270,10 @@ 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)) => {
@@ -228,8 +295,208 @@ fn prepend_config_flags(
         .splice(0..0, cli_config_overrides.raw_overrides);
 }
 
+/// Build the final `TuiCli` for a `codex resume` invocation.
+fn finalize_resume_interactive(
+    mut interactive: TuiCli,
+    root_config_overrides: CliConfigOverrides,
+    session_id: Option<String>,
+    last: bool,
+    resume_cli: TuiCli,
+) -> TuiCli {
+    // 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;
+
+    // Merge resume-scoped flags and overrides with highest precedence.
+    merge_resume_cli_flags(&mut interactive, resume_cli);
+
+    // Propagate any root-level config overrides (e.g. `-c key=value`).
+    prepend_config_flags(&mut interactive.config_overrides, root_config_overrides);
+
+    interactive
+}
+
+/// Merge flags provided to `codex resume` so they take precedence over any
+/// root-level flags. Only overrides fields explicitly set on the resume-scoped
+/// CLI. Also appends `-c key=value` overrides with highest precedence.
+fn merge_resume_cli_flags(interactive: &mut TuiCli, resume_cli: TuiCli) {
+    if let Some(model) = resume_cli.model {
+        interactive.model = Some(model);
+    }
+    if resume_cli.oss {
+        interactive.oss = true;
+    }
+    if let Some(profile) = resume_cli.config_profile {
+        interactive.config_profile = Some(profile);
+    }
+    if let Some(sandbox) = resume_cli.sandbox_mode {
+        interactive.sandbox_mode = Some(sandbox);
+    }
+    if let Some(approval) = resume_cli.approval_policy {
+        interactive.approval_policy = Some(approval);
+    }
+    if resume_cli.full_auto {
+        interactive.full_auto = true;
+    }
+    if resume_cli.dangerously_bypass_approvals_and_sandbox {
+        interactive.dangerously_bypass_approvals_and_sandbox = true;
+    }
+    if let Some(cwd) = resume_cli.cwd {
+        interactive.cwd = Some(cwd);
+    }
+    if resume_cli.web_search {
+        interactive.web_search = true;
+    }
+    if !resume_cli.images.is_empty() {
+        interactive.images = resume_cli.images;
+    }
+    if let Some(prompt) = resume_cli.prompt {
+        interactive.prompt = Some(prompt);
+    }
+
+    interactive
+        .config_overrides
+        .raw_overrides
+        .extend(resume_cli.config_overrides.raw_overrides);
+}
+
 fn print_completion(cmd: CompletionCommand) {
     let mut app = MultitoolCli::command();
     let name = "codex";
     generate(cmd.shell, &mut app, name, &mut std::io::stdout());
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn finalize_from_args(args: &[&str]) -> TuiCli {
+        let cli = MultitoolCli::try_parse_from(args).expect("parse");
+        let MultitoolCli {
+            interactive,
+            config_overrides: root_overrides,
+            subcommand,
+        } = cli;
+
+        let Subcommand::Resume(ResumeCommand {
+            session_id,
+            last,
+            config_overrides: resume_cli,
+        }) = subcommand.expect("resume present")
+        else {
+            unreachable!()
+        };
+
+        finalize_resume_interactive(interactive, root_overrides, session_id, last, resume_cli)
+    }
+
+    #[test]
+    fn resume_model_flag_applies_when_no_root_flags() {
+        let interactive = finalize_from_args(["codex", "resume", "-m", "gpt-5-test"].as_ref());
+
+        assert_eq!(interactive.model.as_deref(), Some("gpt-5-test"));
+        assert!(interactive.resume_picker);
+        assert!(!interactive.resume_last);
+        assert_eq!(interactive.resume_session_id, None);
+    }
+
+    #[test]
+    fn resume_picker_logic_none_and_not_last() {
+        let interactive = finalize_from_args(["codex", "resume"].as_ref());
+        assert!(interactive.resume_picker);
+        assert!(!interactive.resume_last);
+        assert_eq!(interactive.resume_session_id, None);
+    }
+
+    #[test]
+    fn resume_picker_logic_last() {
+        let interactive = finalize_from_args(["codex", "resume", "--last"].as_ref());
+        assert!(!interactive.resume_picker);
+        assert!(interactive.resume_last);
+        assert_eq!(interactive.resume_session_id, None);
+    }
+
+    #[test]
+    fn resume_picker_logic_with_session_id() {
+        let interactive = finalize_from_args(["codex", "resume", "1234"].as_ref());
+        assert!(!interactive.resume_picker);
+        assert!(!interactive.resume_last);
+        assert_eq!(interactive.resume_session_id.as_deref(), Some("1234"));
+    }
+
+    #[test]
+    fn resume_merges_option_flags_and_full_auto() {
+        let interactive = finalize_from_args(
+            [
+                "codex",
+                "resume",
+                "sid",
+                "--oss",
+                "--full-auto",
+                "--search",
+                "--sandbox",
+                "workspace-write",
+                "--ask-for-approval",
+                "on-request",
+                "-m",
+                "gpt-5-test",
+                "-p",
+                "my-profile",
+                "-C",
+                "/tmp",
+                "-i",
+                "/tmp/a.png,/tmp/b.png",
+            ]
+            .as_ref(),
+        );
+
+        assert_eq!(interactive.model.as_deref(), Some("gpt-5-test"));
+        assert!(interactive.oss);
+        assert_eq!(interactive.config_profile.as_deref(), Some("my-profile"));
+        assert!(matches!(
+            interactive.sandbox_mode,
+            Some(codex_common::SandboxModeCliArg::WorkspaceWrite)
+        ));
+        assert!(matches!(
+            interactive.approval_policy,
+            Some(codex_common::ApprovalModeCliArg::OnRequest)
+        ));
+        assert!(interactive.full_auto);
+        assert_eq!(
+            interactive.cwd.as_deref(),
+            Some(std::path::Path::new("/tmp"))
+        );
+        assert!(interactive.web_search);
+        let has_a = interactive
+            .images
+            .iter()
+            .any(|p| p == std::path::Path::new("/tmp/a.png"));
+        let has_b = interactive
+            .images
+            .iter()
+            .any(|p| p == std::path::Path::new("/tmp/b.png"));
+        assert!(has_a && has_b);
+        assert!(!interactive.resume_picker);
+        assert!(!interactive.resume_last);
+        assert_eq!(interactive.resume_session_id.as_deref(), Some("sid"));
+    }
+
+    #[test]
+    fn resume_merges_dangerously_bypass_flag() {
+        let interactive = finalize_from_args(
+            [
+                "codex",
+                "resume",
+                "--dangerously-bypass-approvals-and-sandbox",
+            ]
+            .as_ref(),
+        );
+        assert!(interactive.dangerously_bypass_approvals_and_sandbox);
+        assert!(interactive.resume_picker);
+        assert!(!interactive.resume_last);
+        assert_eq!(interactive.resume_session_id, None);
+    }
+}

codex-rs/cli/src/mcp_cmd.rs

@@ -0,0 +1,370 @@
+use std::collections::BTreeMap;
+use std::collections::HashMap;
+use std::path::PathBuf;
+
+use anyhow::Context;
+use anyhow::Result;
+use anyhow::anyhow;
+use anyhow::bail;
+use codex_common::CliConfigOverrides;
+use codex_core::config::Config;
+use codex_core::config::ConfigOverrides;
+use codex_core::config::find_codex_home;
+use codex_core::config::load_global_mcp_servers;
+use codex_core::config::write_global_mcp_servers;
+use codex_core::config_types::McpServerConfig;
+
+/// [experimental] Launch Codex as an MCP server or manage configured MCP servers.
+///
+/// Subcommands:
+/// - `serve`  — run the MCP server on stdio
+/// - `list`   — list configured servers (with `--json`)
+/// - `get`    — show a single server (with `--json`)
+/// - `add`    — add a server launcher entry to `~/.codex/config.toml`
+/// - `remove` — delete a server entry
+#[derive(Debug, clap::Parser)]
+pub struct McpCli {
+    #[clap(flatten)]
+    pub config_overrides: CliConfigOverrides,
+
+    #[command(subcommand)]
+    pub cmd: Option<McpSubcommand>,
+}
+
+#[derive(Debug, clap::Subcommand)]
+pub enum McpSubcommand {
+    /// [experimental] Run the Codex MCP server (stdio transport).
+    Serve,
+
+    /// [experimental] List configured MCP servers.
+    List(ListArgs),
+
+    /// [experimental] Show details for a configured MCP server.
+    Get(GetArgs),
+
+    /// [experimental] Add a global MCP server entry.
+    Add(AddArgs),
+
+    /// [experimental] Remove a global MCP server entry.
+    Remove(RemoveArgs),
+}
+
+#[derive(Debug, clap::Parser)]
+pub struct ListArgs {
+    /// Output the configured servers as JSON.
+    #[arg(long)]
+    pub json: bool,
+}
+
+#[derive(Debug, clap::Parser)]
+pub struct GetArgs {
+    /// Name of the MCP server to display.
+    pub name: String,
+
+    /// Output the server configuration as JSON.
+    #[arg(long)]
+    pub json: bool,
+}
+
+#[derive(Debug, clap::Parser)]
+pub struct AddArgs {
+    /// Name for the MCP server configuration.
+    pub name: String,
+
+    /// Environment variables to set when launching the server.
+    #[arg(long, value_parser = parse_env_pair, value_name = "KEY=VALUE")]
+    pub env: Vec<(String, String)>,
+
+    /// Command to launch the MCP server.
+    #[arg(trailing_var_arg = true, num_args = 1..)]
+    pub command: Vec<String>,
+}
+
+#[derive(Debug, clap::Parser)]
+pub struct RemoveArgs {
+    /// Name of the MCP server configuration to remove.
+    pub name: String,
+}
+
+impl McpCli {
+    pub async fn run(self, codex_linux_sandbox_exe: Option<PathBuf>) -> Result<()> {
+        let McpCli {
+            config_overrides,
+            cmd,
+        } = self;
+        let subcommand = cmd.unwrap_or(McpSubcommand::Serve);
+
+        match subcommand {
+            McpSubcommand::Serve => {
+                codex_mcp_server::run_main(codex_linux_sandbox_exe, config_overrides).await?;
+            }
+            McpSubcommand::List(args) => {
+                run_list(&config_overrides, args)?;
+            }
+            McpSubcommand::Get(args) => {
+                run_get(&config_overrides, args)?;
+            }
+            McpSubcommand::Add(args) => {
+                run_add(&config_overrides, args)?;
+            }
+            McpSubcommand::Remove(args) => {
+                run_remove(&config_overrides, args)?;
+            }
+        }
+
+        Ok(())
+    }
+}
+
+fn run_add(config_overrides: &CliConfigOverrides, add_args: AddArgs) -> Result<()> {
+    // Validate any provided overrides even though they are not currently applied.
+    config_overrides.parse_overrides().map_err(|e| anyhow!(e))?;
+
+    let AddArgs { name, env, command } = add_args;
+
+    validate_server_name(&name)?;
+
+    let mut command_parts = command.into_iter();
+    let command_bin = command_parts
+        .next()
+        .ok_or_else(|| anyhow!("command is required"))?;
+    let command_args: Vec<String> = command_parts.collect();
+
+    let env_map = if env.is_empty() {
+        None
+    } else {
+        let mut map = HashMap::new();
+        for (key, value) in env {
+            map.insert(key, value);
+        }
+        Some(map)
+    };
+
+    let codex_home = find_codex_home().context("failed to resolve CODEX_HOME")?;
+    let mut servers = load_global_mcp_servers(&codex_home)
+        .with_context(|| format!("failed to load MCP servers from {}", codex_home.display()))?;
+
+    let new_entry = McpServerConfig {
+        command: command_bin,
+        args: command_args,
+        env: env_map,
+        startup_timeout_ms: None,
+    };
+
+    servers.insert(name.clone(), new_entry);
+
+    write_global_mcp_servers(&codex_home, &servers)
+        .with_context(|| format!("failed to write MCP servers to {}", codex_home.display()))?;
+
+    println!("Added global MCP server '{name}'.");
+
+    Ok(())
+}
+
+fn run_remove(config_overrides: &CliConfigOverrides, remove_args: RemoveArgs) -> Result<()> {
+    config_overrides.parse_overrides().map_err(|e| anyhow!(e))?;
+
+    let RemoveArgs { name } = remove_args;
+
+    validate_server_name(&name)?;
+
+    let codex_home = find_codex_home().context("failed to resolve CODEX_HOME")?;
+    let mut servers = load_global_mcp_servers(&codex_home)
+        .with_context(|| format!("failed to load MCP servers from {}", codex_home.display()))?;
+
+    let removed = servers.remove(&name).is_some();
+
+    if removed {
+        write_global_mcp_servers(&codex_home, &servers)
+            .with_context(|| format!("failed to write MCP servers to {}", codex_home.display()))?;
+    }
+
+    if removed {
+        println!("Removed global MCP server '{name}'.");
+    } else {
+        println!("No MCP server named '{name}' found.");
+    }
+
+    Ok(())
+}
+
+fn run_list(config_overrides: &CliConfigOverrides, list_args: ListArgs) -> Result<()> {
+    let overrides = config_overrides.parse_overrides().map_err(|e| anyhow!(e))?;
+    let config = Config::load_with_cli_overrides(overrides, ConfigOverrides::default())
+        .context("failed to load configuration")?;
+
+    let mut entries: Vec<_> = config.mcp_servers.iter().collect();
+    entries.sort_by(|(a, _), (b, _)| a.cmp(b));
+
+    if list_args.json {
+        let json_entries: Vec<_> = entries
+            .into_iter()
+            .map(|(name, cfg)| {
+                let env = cfg.env.as_ref().map(|env| {
+                    env.iter()
+                        .map(|(k, v)| (k.clone(), v.clone()))
+                        .collect::<BTreeMap<_, _>>()
+                });
+                serde_json::json!({
+                    "name": name,
+                    "command": cfg.command,
+                    "args": cfg.args,
+                    "env": env,
+                    "startup_timeout_ms": cfg.startup_timeout_ms,
+                })
+            })
+            .collect();
+        let output = serde_json::to_string_pretty(&json_entries)?;
+        println!("{output}");
+        return Ok(());
+    }
+
+    if entries.is_empty() {
+        println!("No MCP servers configured yet. Try `codex mcp add my-tool -- my-command`.");
+        return Ok(());
+    }
+
+    let mut rows: Vec<[String; 4]> = Vec::new();
+    for (name, cfg) in entries {
+        let args = if cfg.args.is_empty() {
+            "-".to_string()
+        } else {
+            cfg.args.join(" ")
+        };
+
+        let env = match cfg.env.as_ref() {
+            None => "-".to_string(),
+            Some(map) if map.is_empty() => "-".to_string(),
+            Some(map) => {
+                let mut pairs: Vec<_> = map.iter().collect();
+                pairs.sort_by(|(a, _), (b, _)| a.cmp(b));
+                pairs
+                    .into_iter()
+                    .map(|(k, v)| format!("{k}={v}"))
+                    .collect::<Vec<_>>()
+                    .join(", ")
+            }
+        };
+
+        rows.push([name.clone(), cfg.command.clone(), args, env]);
+    }
+
+    let mut widths = ["Name".len(), "Command".len(), "Args".len(), "Env".len()];
+    for row in &rows {
+        for (i, cell) in row.iter().enumerate() {
+            widths[i] = widths[i].max(cell.len());
+        }
+    }
+
+    println!(
+        "{:<name_w$}  {:<cmd_w$}  {:<args_w$}  {:<env_w$}",
+        "Name",
+        "Command",
+        "Args",
+        "Env",
+        name_w = widths[0],
+        cmd_w = widths[1],
+        args_w = widths[2],
+        env_w = widths[3],
+    );
+
+    for row in rows {
+        println!(
+            "{:<name_w$}  {:<cmd_w$}  {:<args_w$}  {:<env_w$}",
+            row[0],
+            row[1],
+            row[2],
+            row[3],
+            name_w = widths[0],
+            cmd_w = widths[1],
+            args_w = widths[2],
+            env_w = widths[3],
+        );
+    }
+
+    Ok(())
+}
+
+fn run_get(config_overrides: &CliConfigOverrides, get_args: GetArgs) -> Result<()> {
+    let overrides = config_overrides.parse_overrides().map_err(|e| anyhow!(e))?;
+    let config = Config::load_with_cli_overrides(overrides, ConfigOverrides::default())
+        .context("failed to load configuration")?;
+
+    let Some(server) = config.mcp_servers.get(&get_args.name) else {
+        bail!("No MCP server named '{name}' found.", name = get_args.name);
+    };
+
+    if get_args.json {
+        let env = server.env.as_ref().map(|env| {
+            env.iter()
+                .map(|(k, v)| (k.clone(), v.clone()))
+                .collect::<BTreeMap<_, _>>()
+        });
+        let output = serde_json::to_string_pretty(&serde_json::json!({
+            "name": get_args.name,
+            "command": server.command,
+            "args": server.args,
+            "env": env,
+            "startup_timeout_ms": server.startup_timeout_ms,
+        }))?;
+        println!("{output}");
+        return Ok(());
+    }
+
+    println!("{}", get_args.name);
+    println!("  command: {}", server.command);
+    let args = if server.args.is_empty() {
+        "-".to_string()
+    } else {
+        server.args.join(" ")
+    };
+    println!("  args: {args}");
+    let env_display = match server.env.as_ref() {
+        None => "-".to_string(),
+        Some(map) if map.is_empty() => "-".to_string(),
+        Some(map) => {
+            let mut pairs: Vec<_> = map.iter().collect();
+            pairs.sort_by(|(a, _), (b, _)| a.cmp(b));
+            pairs
+                .into_iter()
+                .map(|(k, v)| format!("{k}={v}"))
+                .collect::<Vec<_>>()
+                .join(", ")
+        }
+    };
+    println!("  env: {env_display}");
+    if let Some(timeout) = server.startup_timeout_ms {
+        println!("  startup_timeout_ms: {timeout}");
+    }
+    println!("  remove: codex mcp remove {}", get_args.name);
+
+    Ok(())
+}
+
+fn parse_env_pair(raw: &str) -> Result<(String, String), String> {
+    let mut parts = raw.splitn(2, '=');
+    let key = parts
+        .next()
+        .map(str::trim)
+        .filter(|s| !s.is_empty())
+        .ok_or_else(|| "environment entries must be in KEY=VALUE form".to_string())?;
+    let value = parts
+        .next()
+        .map(str::to_string)
+        .ok_or_else(|| "environment entries must be in KEY=VALUE form".to_string())?;
+
+    Ok((key.to_string(), value))
+}
+
+fn validate_server_name(name: &str) -> Result<()> {
+    let is_valid = !name.is_empty()
+        && name
+            .chars()
+            .all(|c| c.is_ascii_alphanumeric() || c == '-' || c == '_');
+
+    if is_valid {
+        Ok(())
+    } else {
+        bail!("invalid server name '{name}' (use letters, numbers, '-', '_')");
+    }
+}

codex-rs/cli/tests/mcp_add_remove.rs

@@ -0,0 +1,86 @@
+use std::path::Path;
+
+use anyhow::Result;
+use codex_core::config::load_global_mcp_servers;
+use predicates::str::contains;
+use pretty_assertions::assert_eq;
+use tempfile::TempDir;
+
+fn codex_command(codex_home: &Path) -> Result<assert_cmd::Command> {
+    let mut cmd = assert_cmd::Command::cargo_bin("codex")?;
+    cmd.env("CODEX_HOME", codex_home);
+    Ok(cmd)
+}
+
+#[test]
+fn add_and_remove_server_updates_global_config() -> Result<()> {
+    let codex_home = TempDir::new()?;
+
+    let mut add_cmd = codex_command(codex_home.path())?;
+    add_cmd
+        .args(["mcp", "add", "docs", "--", "echo", "hello"])
+        .assert()
+        .success()
+        .stdout(contains("Added global MCP server 'docs'."));
+
+    let servers = load_global_mcp_servers(codex_home.path())?;
+    assert_eq!(servers.len(), 1);
+    let docs = servers.get("docs").expect("server should exist");
+    assert_eq!(docs.command, "echo");
+    assert_eq!(docs.args, vec!["hello".to_string()]);
+    assert!(docs.env.is_none());
+
+    let mut remove_cmd = codex_command(codex_home.path())?;
+    remove_cmd
+        .args(["mcp", "remove", "docs"])
+        .assert()
+        .success()
+        .stdout(contains("Removed global MCP server 'docs'."));
+
+    let servers = load_global_mcp_servers(codex_home.path())?;
+    assert!(servers.is_empty());
+
+    let mut remove_again_cmd = codex_command(codex_home.path())?;
+    remove_again_cmd
+        .args(["mcp", "remove", "docs"])
+        .assert()
+        .success()
+        .stdout(contains("No MCP server named 'docs' found."));
+
+    let servers = load_global_mcp_servers(codex_home.path())?;
+    assert!(servers.is_empty());
+
+    Ok(())
+}
+
+#[test]
+fn add_with_env_preserves_key_order_and_values() -> Result<()> {
+    let codex_home = TempDir::new()?;
+
+    let mut add_cmd = codex_command(codex_home.path())?;
+    add_cmd
+        .args([
+            "mcp",
+            "add",
+            "envy",
+            "--env",
+            "FOO=bar",
+            "--env",
+            "ALPHA=beta",
+            "--",
+            "python",
+            "server.py",
+        ])
+        .assert()
+        .success();
+
+    let servers = load_global_mcp_servers(codex_home.path())?;
+    let envy = servers.get("envy").expect("server should exist");
+    let env = envy.env.as_ref().expect("env should be present");
+
+    assert_eq!(env.len(), 2);
+    assert_eq!(env.get("FOO"), Some(&"bar".to_string()));
+    assert_eq!(env.get("ALPHA"), Some(&"beta".to_string()));
+
+    Ok(())
+}

codex-rs/cli/tests/mcp_list.rs

@@ -0,0 +1,106 @@
+use std::path::Path;
+
+use anyhow::Result;
+use predicates::str::contains;
+use pretty_assertions::assert_eq;
+use serde_json::Value as JsonValue;
+use tempfile::TempDir;
+
+fn codex_command(codex_home: &Path) -> Result<assert_cmd::Command> {
+    let mut cmd = assert_cmd::Command::cargo_bin("codex")?;
+    cmd.env("CODEX_HOME", codex_home);
+    Ok(cmd)
+}
+
+#[test]
+fn list_shows_empty_state() -> Result<()> {
+    let codex_home = TempDir::new()?;
+
+    let mut cmd = codex_command(codex_home.path())?;
+    let output = cmd.args(["mcp", "list"]).output()?;
+    assert!(output.status.success());
+    let stdout = String::from_utf8(output.stdout)?;
+    assert!(stdout.contains("No MCP servers configured yet."));
+
+    Ok(())
+}
+
+#[test]
+fn list_and_get_render_expected_output() -> Result<()> {
+    let codex_home = TempDir::new()?;
+
+    let mut add = codex_command(codex_home.path())?;
+    add.args([
+        "mcp",
+        "add",
+        "docs",
+        "--env",
+        "TOKEN=secret",
+        "--",
+        "docs-server",
+        "--port",
+        "4000",
+    ])
+    .assert()
+    .success();
+
+    let mut list_cmd = codex_command(codex_home.path())?;
+    let list_output = list_cmd.args(["mcp", "list"]).output()?;
+    assert!(list_output.status.success());
+    let stdout = String::from_utf8(list_output.stdout)?;
+    assert!(stdout.contains("Name"));
+    assert!(stdout.contains("docs"));
+    assert!(stdout.contains("docs-server"));
+    assert!(stdout.contains("TOKEN=secret"));
+
+    let mut list_json_cmd = codex_command(codex_home.path())?;
+    let json_output = list_json_cmd.args(["mcp", "list", "--json"]).output()?;
+    assert!(json_output.status.success());
+    let stdout = String::from_utf8(json_output.stdout)?;
+    let parsed: JsonValue = serde_json::from_str(&stdout)?;
+    let array = parsed.as_array().expect("expected array");
+    assert_eq!(array.len(), 1);
+    let entry = &array[0];
+    assert_eq!(entry.get("name"), Some(&JsonValue::String("docs".into())));
+    assert_eq!(
+        entry.get("command"),
+        Some(&JsonValue::String("docs-server".into()))
+    );
+
+    let args = entry
+        .get("args")
+        .and_then(|v| v.as_array())
+        .expect("args array");
+    assert_eq!(
+        args,
+        &vec![
+            JsonValue::String("--port".into()),
+            JsonValue::String("4000".into())
+        ]
+    );
+
+    let env = entry
+        .get("env")
+        .and_then(|v| v.as_object())
+        .expect("env map");
+    assert_eq!(env.get("TOKEN"), Some(&JsonValue::String("secret".into())));
+
+    let mut get_cmd = codex_command(codex_home.path())?;
+    let get_output = get_cmd.args(["mcp", "get", "docs"]).output()?;
+    assert!(get_output.status.success());
+    let stdout = String::from_utf8(get_output.stdout)?;
+    assert!(stdout.contains("docs"));
+    assert!(stdout.contains("command: docs-server"));
+    assert!(stdout.contains("args: --port 4000"));
+    assert!(stdout.contains("env: TOKEN=secret"));
+    assert!(stdout.contains("remove: codex mcp remove docs"));
+
+    let mut get_json_cmd = codex_command(codex_home.path())?;
+    get_json_cmd
+        .args(["mcp", "get", "docs", "--json"])
+        .assert()
+        .success()
+        .stdout(contains("\"name\": \"docs\""));
+
+    Ok(())
+}

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/model_presets.rs

@@ -1,4 +1,6 @@
+use codex_core::config::GPT_5_CODEX_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)]
@@ -12,50 +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,
-        },
-        ModelPreset {
-            id: "gpt-5-high-new",
-            label: "gpt-5 high new",
-            description: "— our latest release tuned to rely on the model's built-in reasoning defaults",
-            model: "gpt-5-high-new",
-            effort: ReasoningEffort::Medium,
-        },
-    ];
-    PRESETS
+const PRESETS: &[ModelPreset] = &[
+    ModelPreset {
+        id: "gpt-5-codex-low",
+        label: "gpt-5-codex low",
+        description: "",
+        model: "gpt-5-codex",
+        effort: Some(ReasoningEffort::Low),
+    },
+    ModelPreset {
+        id: "gpt-5-codex-medium",
+        label: "gpt-5-codex medium",
+        description: "",
+        model: "gpt-5-codex",
+        effort: None,
+    },
+    ModelPreset {
+        id: "gpt-5-codex-high",
+        label: "gpt-5-codex high",
+        description: "",
+        model: "gpt-5-codex",
+        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 != GPT_5_CODEX_MEDIUM_MODEL)
+            .collect(),
+        _ => PRESETS.to_vec(),
+    }
 }

codex-rs/core/Cargo.toml

@@ -4,22 +4,25 @@ name = "codex-core"
 version = { workspace = true }
 
 [lib]
+doctest = false
 name = "codex_core"
 path = "src/lib.rs"
-doctest = false
 
 [lints]
 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-file-search = { path = "../file-search" }
 codex-mcp-client = { path = "../mcp-client" }
 codex-protocol = { path = "../protocol" }
+codex-otel = { path = "../otel", features = ["otel"] }
 dirs = "6"
 env-flags = "0.1.1"
 eventsource-stream = "0.2.3"
@@ -39,7 +42,12 @@ similar = "2.7.0"
 strum_macros = "0.27.2"
 tempfile = "3"
 thiserror = "2.0.16"
-time = { version = "0.3", features = ["formatting", "parsing", "local-offset", "macros"] }
+time = { version = "0.3", features = [
+    "formatting",
+    "parsing",
+    "local-offset",
+    "macros",
+] }
 tokio = { version = "1", features = [
     "io-std",
     "macros",
@@ -55,7 +63,7 @@ tree-sitter = "0.25.9"
 tree-sitter-bash = "0.25.0"
 uuid = { version = "1", features = ["serde", "v4"] }
 which = "6"
-wildmatch = "2.4.0"
+wildmatch = "2.5.0"
 
 
 [target.'cfg(target_os = "linux")'.dependencies]
@@ -80,6 +88,7 @@ tempfile = "3"
 tokio-test = "0.4"
 walkdir = "2.5.0"
 wiremock = "0.6"
+tracing-test = {  version = "0.2.5", features = ["no-env-filter"] }
 
 [package.metadata.cargo-shear]
 ignored = ["openssl-sys"]

codex-rs/core/gpt_5_codex_prompt.md

@@ -0,0 +1,104 @@
+You are Codex, based on GPT-5. You are running as a coding agent in the Codex CLI on a user's computer.
+
+## General
+
+- The arguments to `shell` will be passed to execvp(). Most terminal commands should be prefixed with ["bash", "-lc"].
+- Always set the `workdir` param when using the shell function. 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.)
+
+## 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.
+- 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.
+- While you are working, you might notice unexpected changes that you didn't make. If this happens, STOP IMMEDIATELY and ask the user how they would like to proceed.
+
+## 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.
+
+## Codex CLI harness, sandboxing, and approvals
+
+The Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.
+
+Filesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:
+- **read-only**: The sandbox only permits reading files.
+- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.
+- **danger-full-access**: No filesystem sandboxing - all commands are permitted.
+
+Network sandboxing defines whether network can be accessed without approval. Options for `network_access` are:
+- **restricted**: Requires approval
+- **enabled**: No approval needed
+
+Approvals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` 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 `approval_policy == 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 /var)
+- 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. ALWAYS proceed to use the `with_escalated_permissions` and `justification` parameters - do not message the user before requesting approval for the command.
+- 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 `sandbox_mode` 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.
+
+Although they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, 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.
+
+When requesting approval to execute 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 you need to enable `with_escalated_permissions` in the justification parameter
+
+## Special user requests
+
+- 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.
+- If 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.
+
+## 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.
+- File References: When referencing files in your response, make sure to include the relevant start line and always follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Line/column (1‑based, optional): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5

codex-rs/core/prompt.md

@@ -251,6 +251,16 @@ You are producing plain text that will later be styled by the CLI. Follow these
 - Apply to inline examples and to bullet keywords if the keyword itself is a literal file/command.
 - Never mix monospace and bold markers; choose one based on whether it’s a keyword (`**`) or inline code/path (`` ` ``).
 
+**File References**
+When referencing files in your response, make sure to include the relevant start line and always follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Line/column (1‑based, optional): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5
+
 **Structure**
 
 - Place related bullets together; don’t mix unrelated concepts in the same section.

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/apply_patch.rs

@@ -52,12 +52,13 @@ pub(crate) async fn apply_patch(
         &turn_context.sandbox_policy,
         &turn_context.cwd,
     ) {
-        SafetyCheck::AutoApprove { .. } => {
-            InternalApplyPatchInvocation::DelegateToExec(ApplyPatchExec {
-                action,
-                user_explicitly_approved_this_action: false,
-            })
-        }
+        SafetyCheck::AutoApprove {
+            user_explicitly_approved,
+            ..
+        } => InternalApplyPatchInvocation::DelegateToExec(ApplyPatchExec {
+            action,
+            user_explicitly_approved_this_action: user_explicitly_approved,
+        }),
         SafetyCheck::AskUser => {
             // Compute a readable summary of path changes to include in the
             // approval request so the user can make an informed decision.

codex-rs/core/src/auth.rs

@@ -408,6 +408,32 @@ mod tests {
         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();

codex-rs/core/src/bash.rs

@@ -73,6 +73,9 @@ pub fn try_parse_word_only_commands_sequence(tree: &Tree, src: &str) -> Option<V
         }
     }
 
+    // Walk uses a stack (LIFO), so re-sort by position to restore source order.
+    command_nodes.sort_by_key(|node| node.start_byte());
+
     let mut commands = Vec::new();
     for node in command_nodes {
         if let Some(words) = parse_plain_command_from_node(node, src) {
@@ -150,10 +153,10 @@ mod tests {
         let src = "ls && pwd; echo 'hi there' | wc -l";
         let cmds = parse_seq(src).unwrap();
         let expected: Vec<Vec<String>> = vec![
-            vec!["wc".to_string(), "-l".to_string()],
-            vec!["echo".to_string(), "hi there".to_string()],
-            vec!["pwd".to_string()],
             vec!["ls".to_string()],
+            vec!["pwd".to_string()],
+            vec!["echo".to_string(), "hi there".to_string()],
+            vec!["wc".to_string(), "-l".to_string()],
         ];
         assert_eq!(cmds, expected);
     }

codex-rs/core/src/chat_completions.rs

@@ -1,6 +1,20 @@
 use std::time::Duration;
+use std::time::Instant;
 
+use crate::ModelProviderInfo;
+use crate::client_common::Prompt;
+use crate::client_common::ResponseEvent;
+use crate::client_common::ResponseStream;
+use crate::error::CodexErr;
+use crate::error::Result;
+use crate::model_family::ModelFamily;
+use crate::openai_tools::create_tools_json_for_chat_completions_api;
+use crate::util::backoff;
 use bytes::Bytes;
+use codex_otel::otel_event_manager::OtelEventManager;
+use codex_protocol::models::ContentItem;
+use codex_protocol::models::ReasoningItemContent;
+use codex_protocol::models::ResponseItem;
 use eventsource_stream::Eventsource;
 use futures::Stream;
 use futures::StreamExt;
@@ -15,25 +29,13 @@ use tokio::time::timeout;
 use tracing::debug;
 use tracing::trace;
 
-use crate::ModelProviderInfo;
-use crate::client_common::Prompt;
-use crate::client_common::ResponseEvent;
-use crate::client_common::ResponseStream;
-use crate::error::CodexErr;
-use crate::error::Result;
-use crate::model_family::ModelFamily;
-use crate::openai_tools::create_tools_json_for_chat_completions_api;
-use crate::util::backoff;
-use codex_protocol::models::ContentItem;
-use codex_protocol::models::ReasoningItemContent;
-use codex_protocol::models::ResponseItem;
-
 /// Implementation for the classic Chat Completions API.
 pub(crate) async fn stream_chat_completions(
     prompt: &Prompt,
     model_family: &ModelFamily,
     client: &reqwest::Client,
     provider: &ModelProviderInfo,
+    otel_event_manager: &OtelEventManager,
 ) -> Result<ResponseStream> {
     // Build messages array
     let mut messages = Vec::<serde_json::Value>::new();
@@ -288,12 +290,30 @@ pub(crate) async fn stream_chat_completions(
 
         let req_builder = provider.create_request_builder(client, &None).await?;
 
+        let start = Instant::now();
+
         let res = req_builder
             .header(reqwest::header::ACCEPT, "text/event-stream")
             .json(&payload)
             .send()
             .await;
 
+        let cf_ray = if let Ok(resp) = &res {
+            let cf_ray = resp
+                .headers()
+                .get("cf-ray")
+                .map(|v| v.to_str().unwrap_or_default())
+                .unwrap_or_default()
+                .to_string();
+
+            trace!("Response status: {}, cf-ray: {}", resp.status(), cf_ray);
+            Some(cf_ray)
+        } else {
+            None
+        };
+
+        otel_event_manager.request(cf_ray, attempt, start.elapsed(), &res);
+
         match res {
             Ok(resp) if resp.status().is_success() => {
                 let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
@@ -302,6 +322,7 @@ pub(crate) async fn stream_chat_completions(
                     stream,
                     tx_event,
                     provider.stream_idle_timeout(),
+                    otel_event_manager.clone(),
                 ));
                 return Ok(ResponseStream { rx_event });
             }
@@ -345,6 +366,7 @@ async fn process_chat_sse<S>(
     stream: S,
     tx_event: mpsc::Sender<Result<ResponseEvent>>,
     idle_timeout: Duration,
+    otel_event_manager: OtelEventManager,
 ) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
@@ -368,12 +390,13 @@ async fn process_chat_sse<S>(
     let mut reasoning_text = String::new();
 
     loop {
+        let start = Instant::now();
         let sse = match timeout(idle_timeout, stream.next()).await {
             Ok(Some(Ok(ev))) => ev,
             Ok(Some(Err(e))) => {
-                let _ = tx_event
-                    .send(Err(CodexErr::Stream(e.to_string(), None)))
-                    .await;
+                let error = e.to_string();
+                otel_event_manager.sse_event_failed(None, start.elapsed(), error.as_str());
+                let _ = tx_event.send(Err(CodexErr::Stream(error, None))).await;
                 return;
             }
             Ok(None) => {
@@ -387,11 +410,10 @@ async fn process_chat_sse<S>(
                 return;
             }
             Err(_) => {
+                let error = "idle timeout waiting for SSE";
+                otel_event_manager.sse_event_failed(None, start.elapsed(), error);
                 let _ = tx_event
-                    .send(Err(CodexErr::Stream(
-                        "idle timeout waiting for SSE".into(),
-                        None,
-                    )))
+                    .send(Err(CodexErr::Stream(error.into(), None)))
                     .await;
                 return;
             }
@@ -399,6 +421,7 @@ async fn process_chat_sse<S>(
 
         // OpenAI Chat streaming sends a literal string "[DONE]" when finished.
         if sse.data.trim() == "[DONE]" {
+            otel_event_manager.sse_event(&sse.event, start.elapsed());
             // Emit any finalized items before closing so downstream consumers receive
             // terminal events for both assistant content and raw reasoning.
             if !assistant_text.is_empty() {
@@ -436,9 +459,14 @@ async fn process_chat_sse<S>(
         // Parse JSON chunk
         let chunk: serde_json::Value = match serde_json::from_str(&sse.data) {
             Ok(v) => v,
-            Err(_) => continue,
+            Err(e) => {
+                let error = format!("Failed to parse SSE event: {e}, data: {}", &sse.data);
+                otel_event_manager.sse_event_failed(None, start.elapsed(), error.as_str());
+                continue;
+            }
         };
         trace!("chat_completions received SSE chunk: {chunk:?}");
+        otel_event_manager.sse_event(&sse.event, start.elapsed());
 
         let choice_opt = chunk.get("choices").and_then(|c| c.get(0));
 

codex-rs/core/src/client.rs

@@ -2,6 +2,7 @@ use std::io::BufRead;
 use std::path::Path;
 use std::sync::OnceLock;
 use std::time::Duration;
+use std::time::Instant;
 
 use crate::AuthManager;
 use bytes::Bytes;
@@ -43,6 +44,7 @@ use crate::openai_tools::create_tools_json_for_responses_api;
 use crate::protocol::TokenUsage;
 use crate::token_data::PlanType;
 use crate::util::backoff;
+use codex_otel::otel_event_manager::OtelEventManager;
 use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
 use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use codex_protocol::models::ResponseItem;
@@ -69,10 +71,11 @@ struct Error {
 pub struct ModelClient {
     config: Arc<Config>,
     auth_manager: Option<Arc<AuthManager>>,
+    otel_event_manager: OtelEventManager,
     client: reqwest::Client,
     provider: ModelProviderInfo,
     conversation_id: ConversationId,
-    effort: ReasoningEffortConfig,
+    effort: Option<ReasoningEffortConfig>,
     summary: ReasoningSummaryConfig,
 }
 
@@ -80,8 +83,9 @@ impl ModelClient {
     pub fn new(
         config: Arc<Config>,
         auth_manager: Option<Arc<AuthManager>>,
+        otel_event_manager: OtelEventManager,
         provider: ModelProviderInfo,
-        effort: ReasoningEffortConfig,
+        effort: Option<ReasoningEffortConfig>,
         summary: ReasoningSummaryConfig,
         conversation_id: ConversationId,
     ) -> Self {
@@ -90,6 +94,7 @@ impl ModelClient {
         Self {
             config,
             auth_manager,
+            otel_event_manager,
             client,
             provider,
             conversation_id,
@@ -104,6 +109,12 @@ 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.
@@ -117,6 +128,7 @@ impl ModelClient {
                     &self.config.model_family,
                     &self.client,
                     &self.provider,
+                    &self.otel_event_manager,
                 )
                 .await?;
 
@@ -153,7 +165,12 @@ impl ModelClient {
         if let Some(path) = &*CODEX_RS_SSE_FIXTURE {
             // short circuit for tests
             warn!(path, "Streaming from fixture");
-            return stream_from_fixture(path, self.provider.clone()).await;
+            return stream_from_fixture(
+                path,
+                self.provider.clone(),
+                self.otel_event_manager.clone(),
+            )
+            .await;
         }
 
         let auth_manager = self.auth_manager.clone();
@@ -187,6 +204,15 @@ 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,
@@ -195,13 +221,19 @@ impl ModelClient {
             tool_choice: "auto",
             parallel_tool_calls: false,
             reasoning,
-            store: false,
+            store: azure_workaround,
             stream: true,
             include,
             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();
 
@@ -214,7 +246,7 @@ 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
@@ -228,7 +260,7 @@ impl ModelClient {
                 .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
@@ -237,17 +269,25 @@ impl ModelClient {
                 req_builder = req_builder.header("chatgpt-account-id", account_id);
             }
 
+            let start = Instant::now();
             let res = req_builder.send().await;
-            if let Ok(resp) = &res {
-                trace!(
-                    "Response status: {}, cf-ray: {}",
-                    resp.status(),
-                    resp.headers()
-                        .get("cf-ray")
-                        .map(|v| v.to_str().unwrap_or_default())
-                        .unwrap_or_default()
-                );
-            }
+
+            let cf_ray = if let Ok(resp) = &res {
+                let cf_ray = resp
+                    .headers()
+                    .get("cf-ray")
+                    .map(|v| v.to_str().unwrap_or_default())
+                    .unwrap_or_default()
+                    .to_string();
+
+                trace!("Response status: {}, cf-ray: {}", resp.status(), cf_ray);
+                Some(cf_ray)
+            } else {
+                None
+            };
+
+            self.otel_event_manager
+                .request(cf_ray, attempt, start.elapsed(), &res);
 
             match res {
                 Ok(resp) if resp.status().is_success() => {
@@ -259,6 +299,7 @@ impl ModelClient {
                         stream,
                         tx_event,
                         self.provider.stream_idle_timeout(),
+                        self.otel_event_manager.clone(),
                     ));
 
                     return Ok(ResponseStream { rx_event });
@@ -345,6 +386,10 @@ impl ModelClient {
         self.provider.clone()
     }
 
+    pub fn get_otel_event_manager(&self) -> OtelEventManager {
+        self.otel_event_manager.clone()
+    }
+
     /// Returns the currently configured model slug.
     pub fn get_model(&self) -> String {
         self.config.model.clone()
@@ -356,7 +401,7 @@ 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
     }
 
@@ -368,6 +413,10 @@ impl ModelClient {
     pub fn get_auth_manager(&self) -> Option<Arc<AuthManager>> {
         self.auth_manager.clone()
     }
+
+    pub fn get_conversation_id(&self) -> ConversationId {
+        self.conversation_id
+    }
 }
 
 #[derive(Debug, Deserialize, Serialize)]
@@ -425,10 +474,38 @@ 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>>,
     idle_timeout: Duration,
+    otel_event_manager: OtelEventManager,
 ) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
@@ -440,11 +517,14 @@ async fn process_sse<S>(
     let mut response_error: Option<CodexErr> = None;
 
     loop {
+        let start = Instant::now();
         let sse = match timeout(idle_timeout, stream.next()).await {
             Ok(Some(Ok(sse))) => sse,
             Ok(Some(Err(e))) => {
                 debug!("SSE Error: {e:#}");
-                let event = CodexErr::Stream(e.to_string(), None);
+                let error = e.to_string();
+                otel_event_manager.sse_event_failed(None, start.elapsed(), error.as_str());
+                let event = CodexErr::Stream(error, None);
                 let _ = tx_event.send(Err(event)).await;
                 return;
             }
@@ -454,6 +534,22 @@ async fn process_sse<S>(
                         id: response_id,
                         usage,
                     }) => {
+                        if let Some(token_usage) = &usage {
+                            otel_event_manager.sse_event_completed(
+                                start.elapsed(),
+                                token_usage.input_tokens,
+                                token_usage.output_tokens,
+                                token_usage
+                                    .input_tokens_details
+                                    .as_ref()
+                                    .map(|d| d.cached_tokens),
+                                token_usage
+                                    .output_tokens_details
+                                    .as_ref()
+                                    .map(|d| d.reasoning_tokens),
+                                token_usage.total_tokens,
+                            );
+                        }
                         let event = ResponseEvent::Completed {
                             response_id,
                             token_usage: usage.map(Into::into),
@@ -461,22 +557,31 @@ async fn process_sse<S>(
                         let _ = tx_event.send(Ok(event)).await;
                     }
                     None => {
-                        let _ = tx_event
-                            .send(Err(response_error.unwrap_or(CodexErr::Stream(
-                                "stream closed before response.completed".into(),
+                        let error = response_error.unwrap_or(CodexErr::Stream(
+                            "stream closed before response.completed".into(),
+                            None,
+                        ));
+
+                        if let CodexErr::Stream(message, _) = &error {
+                            otel_event_manager.sse_event_failed(
                                 None,
-                            ))))
-                            .await;
+                                start.elapsed(),
+                                message.as_str(),
+                            );
+                        }
+
+                        let _ = tx_event.send(Err(error)).await;
                     }
                 }
                 return;
             }
             Err(_) => {
+                let error = "idle timeout waiting for SSE";
+
+                otel_event_manager.sse_event_failed(None, start.elapsed(), error);
+
                 let _ = tx_event
-                    .send(Err(CodexErr::Stream(
-                        "idle timeout waiting for SSE".into(),
-                        None,
-                    )))
+                    .send(Err(CodexErr::Stream(error.into(), None)))
                     .await;
                 return;
             }
@@ -488,11 +593,15 @@ async fn process_sse<S>(
         let event: SseEvent = match serde_json::from_str(&sse.data) {
             Ok(event) => event,
             Err(e) => {
-                debug!("Failed to parse SSE event: {e}, data: {}", &sse.data);
+                let error = format!("Failed to parse SSE event: {e}, data: {}", &sse.data);
+                otel_event_manager.sse_event_failed(None, start.elapsed(), error.as_str());
+                debug!(error);
                 continue;
             }
         };
 
+        otel_event_manager.sse_event(event.kind.as_str(), start.elapsed());
+
         match event.kind.as_str() {
             // Individual output item finalised. Forward immediately so the
             // rest of the agent can stream assistant text/functions *live*
@@ -513,9 +622,13 @@ async fn process_sse<S>(
             // The fix is to forward the incremental events *as they come* and
             // drop the duplicated list inside `response.completed`.
             "response.output_item.done" => {
-                let Some(item_val) = event.item else { continue };
+                let Some(item_val) = event.item else {
+                    continue
+                };
                 let Ok(item) = serde_json::from_value::<ResponseItem>(item_val) else {
-                    debug!("failed to parse ResponseItem from output_item.done");
+                    let error = "failed to parse ResponseItem from output_item.done";
+                    debug!(error);
+                    otel_event_manager.sse_event_failed(Some(event.kind), start.elapsed(), error);
                     continue;
                 };
 
@@ -555,8 +668,9 @@ async fn process_sse<S>(
             }
             "response.failed" => {
                 if let Some(resp_val) = event.response {
+                    let error_message = "response.failed event received";
                     response_error = Some(CodexErr::Stream(
-                        "response.failed event received".to_string(),
+                        error_message.to_string(),
                         None,
                     ));
 
@@ -567,12 +681,17 @@ async fn process_sse<S>(
                             Ok(error) => {
                                 let delay = try_parse_retry_after(&error);
                                 let message = error.message.unwrap_or_default();
+                                otel_event_manager.sse_event_failed(Some(event.kind), start.elapsed(), message.as_str());
                                 response_error = Some(CodexErr::Stream(message, delay));
                             }
                             Err(e) => {
-                                debug!("failed to parse ErrorResponse: {e}");
+                                let error_message = format!("failed to parse ErrorResponse: {e}");
+                                debug!(error_message);
+                                otel_event_manager.sse_event_failed(Some(event.kind), start.elapsed(), error_message.as_str());
                             }
                         }
+                    } else {
+                        otel_event_manager.sse_event_failed(Some(event.kind), start.elapsed(), error_message);
                     }
                 }
             }
@@ -584,7 +703,9 @@ async fn process_sse<S>(
                             response_completed = Some(r);
                         }
                         Err(e) => {
-                            debug!("failed to parse ResponseCompleted: {e}");
+                            let error_message = format!("failed to parse ResponseCompleted: {e}");
+                            debug!(error_message);
+                            otel_event_manager.sse_event_failed(Some(event.kind), start.elapsed(), error_message.as_str());
                             continue;
                         }
                     };
@@ -621,8 +742,10 @@ async fn process_sse<S>(
                     return;
                 }
             }
-            "response.reasoning_summary_text.done" => {}
-            _ => {}
+            "response.reasoning_summary_text.done" => {
+            }
+            _ => {
+            }
         }
     }
 }
@@ -631,6 +754,7 @@ async fn process_sse<S>(
 async fn stream_from_fixture(
     path: impl AsRef<Path>,
     provider: ModelProviderInfo,
+    otel_event_manager: OtelEventManager,
 ) -> Result<ResponseStream> {
     let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
     let f = std::fs::File::open(path.as_ref())?;
@@ -649,6 +773,7 @@ async fn stream_from_fixture(
         stream,
         tx_event,
         provider.stream_idle_timeout(),
+        otel_event_manager,
     ));
     Ok(ResponseStream { rx_event })
 }
@@ -694,7 +819,7 @@ mod tests {
     use tokio::sync::mpsc;
     use tokio_test::io::Builder as IoBuilder;
     use tokio_util::io::ReaderStream;
-
+    use tracing_test::traced_test;
     // ────────────────────────────
     // Helpers
     // ────────────────────────────
@@ -704,6 +829,7 @@ mod tests {
     async fn collect_events(
         chunks: &[&[u8]],
         provider: ModelProviderInfo,
+        otel_event_manager: OtelEventManager,
     ) -> Vec<Result<ResponseEvent>> {
         let mut builder = IoBuilder::new();
         for chunk in chunks {
@@ -713,7 +839,12 @@ mod tests {
         let reader = builder.build();
         let stream = ReaderStream::new(reader).map_err(CodexErr::Io);
         let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(16);
-        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));
+        tokio::spawn(process_sse(
+            stream,
+            tx,
+            provider.stream_idle_timeout(),
+            otel_event_manager,
+        ));
 
         let mut events = Vec::new();
         while let Some(ev) = rx.recv().await {
@@ -722,11 +853,35 @@ mod tests {
         events
     }
 
+    async fn run_sse_allow_errors(
+        events: Vec<serde_json::Value>,
+        provider: ModelProviderInfo,
+        otel_event_manager: OtelEventManager,
+    ) {
+        let chunks: Vec<Vec<u8>> = events
+            .into_iter()
+            .map(|event| {
+                let kind = event
+                    .get("type")
+                    .and_then(|v| v.as_str())
+                    .expect("fixture event missing type");
+                if event.as_object().map(|o| o.len() == 1).unwrap_or(false) {
+                    format!("event: {kind}\n\n").into_bytes()
+                } else {
+                    format!("event: {kind}\ndata: {event}\n\n").into_bytes()
+                }
+            })
+            .collect();
+        let chunk_refs: Vec<&[u8]> = chunks.iter().map(Vec::as_slice).collect();
+        let _ = collect_events(&chunk_refs, provider, otel_event_manager).await;
+    }
+
     /// Builds an in-memory SSE stream from JSON fixtures and returns only the
     /// successfully parsed events (panics on internal channel errors).
     async fn run_sse(
         events: Vec<serde_json::Value>,
         provider: ModelProviderInfo,
+        otel_event_manager: OtelEventManager,
     ) -> Vec<ResponseEvent> {
         let mut body = String::new();
         for e in events {
@@ -743,7 +898,12 @@ mod tests {
 
         let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(8);
         let stream = ReaderStream::new(std::io::Cursor::new(body)).map_err(CodexErr::Io);
-        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));
+        tokio::spawn(process_sse(
+            stream,
+            tx,
+            provider.stream_idle_timeout(),
+            otel_event_manager,
+        ));
 
         let mut out = Vec::new();
         while let Some(ev) = rx.recv().await {
@@ -752,6 +912,378 @@ mod tests {
         out
     }
 
+    fn otel_event_manager() -> OtelEventManager {
+        OtelEventManager::new(
+            ConversationId::new(),
+            "test-model",
+            "test-slug",
+            None,
+            Some(AuthMode::ChatGPT),
+            true,
+            "test-terminal".to_string(),
+        )
+    }
+
+    fn model_provider() -> ModelProviderInfo {
+        ModelProviderInfo {
+            name: "test".to_string(),
+            base_url: Some("https://example.com".to_string()),
+            env_key: Some("TEST_KEY".to_string()),
+            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(1_000),
+            requires_openai_auth: false,
+        }
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_emits_tracing_for_output_item() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+
+        let _ = run_sse(
+            vec![
+                json!({
+                    "type": "response.output_item.done",
+                    "item": {
+                        "type": "message",
+                        "role": "assistant",
+                        "content": [{ "type": "output_text", "text": "hi" }]
+                    }
+                }),
+                json!({
+                    "type": "response.completed",
+                    "response": { "id": "resp1" }
+                }),
+            ],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.output_item.done")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing response.output_item.done event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_emits_failed_event_on_parse_error() {
+        let chunk = "data: not-json\n\n";
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+
+        let _ = collect_events(&[chunk.as_bytes()], provider, otel_event_manager).await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("error.message")
+                        && line.contains("Failed to parse SSE event")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_stream_error_records_failed_event() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(1);
+
+        let stream = stream::iter(vec![Err::<Bytes, CodexErr>(
+            std::io::Error::new(std::io::ErrorKind::Other, "boom").into(),
+        )]);
+
+        let handle = tokio::spawn(process_sse(
+            stream,
+            tx,
+            provider.stream_idle_timeout(),
+            otel_event_manager,
+        ));
+
+        while rx.recv().await.is_some() {}
+        handle.await.expect("process_sse task");
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("error.message")
+                        && line.contains("boom")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_idle_timeout_records_failed_event() {
+        let otel_event_manager = otel_event_manager();
+        let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(1);
+        let stream = stream::pending::<std::result::Result<Bytes, CodexErr>>();
+        let idle_timeout = std::time::Duration::from_millis(1);
+
+        let handle = tokio::spawn(process_sse(stream, tx, idle_timeout, otel_event_manager));
+        let _ = rx.recv().await;
+        handle.await.expect("process_sse task");
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("error.message")
+                        && line.contains("idle timeout")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_records_failed_event_when_stream_closes_without_completed() {
+        let item = json!({
+            "type": "response.output_item.done",
+            "item": {
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "hi"}]
+            }
+        })
+        .to_string();
+        let sse = format!("event: response.output_item.done\ndata: {item}\n\n");
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+
+        let _ = collect_events(&[sse.as_bytes()], provider, otel_event_manager).await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("error.message")
+                        && line.contains("stream closed before response.completed")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_output_item_parse_failure_logs_event() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let _ = run_sse_allow_errors(
+            vec![json!({
+                "type": "response.output_item.done",
+                "item": "unexpected"
+            })],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.output_item.done")
+                        && line.contains("error.message")
+                        && line.contains("failed to parse ResponseItem")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_failed_event_records_response_error_message() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let _ = run_sse_allow_errors(
+            vec![json!({
+                "type": "response.failed",
+                "response": {
+                    "error": {
+                        "message": "boom",
+                        "code": "bad"
+                    }
+                }
+            })],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.failed")
+                        && line.contains("error.message")
+                        && line.contains("boom")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_failed_event_logs_parse_error() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let _ = run_sse_allow_errors(
+            vec![json!({
+                "type": "response.failed",
+                "response": {
+                    "error": "not-an-object"
+                }
+            })],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.failed")
+                        && line.contains("error.message")
+                        && line.contains("failed to parse ErrorResponse")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_failed_event_logs_missing_error() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let _ = run_sse_allow_errors(
+            vec![json!({
+                "type": "response.failed",
+                "response": {}
+            })],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.failed")
+                        && line.contains("error.message")
+                        && line.contains("response.failed event received")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_failed_event_logs_response_completed_parse_error() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let _ = run_sse_allow_errors(
+            vec![json!({
+                "type": "response.completed",
+                "response": {}
+            })],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.completed")
+                        && line.contains("error.message")
+                        && line.contains("failed to parse ResponseCompleted")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing codex.sse_event".to_string()))
+        });
+    }
+
+    #[tokio::test]
+    #[traced_test]
+    async fn process_sse_emits_completed_telemetry() {
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
+        let _ = run_sse_allow_errors(
+            vec![json!({
+                "type": "response.completed",
+                "response": {
+                    "id": "resp1",
+                    "usage": {
+                        "input_tokens": 3,
+                        "input_tokens_details": { "cached_tokens": 1 },
+                        "output_tokens": 5,
+                        "output_tokens_details": { "reasoning_tokens": 2 },
+                        "total_tokens": 9
+                    }
+                }
+            })],
+            provider,
+            otel_event_manager,
+        )
+        .await;
+
+        logs_assert(|lines: &[&str]| {
+            lines
+                .iter()
+                .find(|line| {
+                    line.contains("codex.sse_event")
+                        && line.contains("event.kind=response.completed")
+                        && line.contains("input_token_count=3")
+                        && line.contains("output_token_count=5")
+                        && line.contains("cached_token_count=1")
+                        && line.contains("reasoning_token_count=2")
+                        && line.contains("tool_token_count=9")
+                })
+                .map(|_| Ok(()))
+                .unwrap_or(Err("missing response.completed telemetry".to_string()))
+        });
+    }
+
     // ────────────────────────────
     // Tests from `implement-test-for-responses-api-sse-parser`
     // ────────────────────────────
@@ -788,24 +1320,13 @@ mod tests {
         let sse2 = format!("event: response.output_item.done\ndata: {item2}\n\n");
         let sse3 = format!("event: response.completed\ndata: {completed}\n\n");
 
-        let provider = ModelProviderInfo {
-            name: "test".to_string(),
-            base_url: Some("https://test.com".to_string()),
-            env_key: Some("TEST_API_KEY".to_string()),
-            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(1000),
-            requires_openai_auth: false,
-        };
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
 
         let events = collect_events(
             &[sse1.as_bytes(), sse2.as_bytes(), sse3.as_bytes()],
             provider,
+            otel_event_manager,
         )
         .await;
 
@@ -848,22 +1369,10 @@ mod tests {
         .to_string();
 
         let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
-        let provider = ModelProviderInfo {
-            name: "test".to_string(),
-            base_url: Some("https://test.com".to_string()),
-            env_key: Some("TEST_API_KEY".to_string()),
-            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(1000),
-            requires_openai_auth: false,
-        };
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
 
-        let events = collect_events(&[sse1.as_bytes()], provider).await;
+        let events = collect_events(&[sse1.as_bytes()], provider, otel_event_manager).await;
 
         assert_eq!(events.len(), 2);
 
@@ -882,22 +1391,10 @@ mod tests {
         let raw_error = r#"{"type":"response.failed","sequence_number":3,"response":{"id":"resp_689bcf18d7f08194bf3440ba62fe05d803fee0cdac429894","object":"response","created_at":1755041560,"status":"failed","background":false,"error":{"code":"rate_limit_exceeded","message":"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."}, "usage":null,"user":null,"metadata":{}}}"#;
 
         let sse1 = format!("event: response.failed\ndata: {raw_error}\n\n");
-        let provider = ModelProviderInfo {
-            name: "test".to_string(),
-            base_url: Some("https://test.com".to_string()),
-            env_key: Some("TEST_API_KEY".to_string()),
-            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(1000),
-            requires_openai_auth: false,
-        };
+        let provider = model_provider();
+        let otel_event_manager = otel_event_manager();
 
-        let events = collect_events(&[sse1.as_bytes()], provider).await;
+        let events = collect_events(&[sse1.as_bytes()], provider, otel_event_manager).await;
 
         assert_eq!(events.len(), 1);
 
@@ -987,22 +1484,10 @@ mod tests {
             let mut evs = vec![case.event];
             evs.push(completed.clone());
 
-            let provider = ModelProviderInfo {
-                name: "test".to_string(),
-                base_url: Some("https://test.com".to_string()),
-                env_key: Some("TEST_API_KEY".to_string()),
-                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(1000),
-                requires_openai_auth: false,
-            };
+            let provider = model_provider();
+            let otel_event_manager = otel_event_manager();
 
-            let out = run_sse(evs, provider).await;
+            let out = run_sse(evs, provider, otel_event_manager).await;
             assert_eq!(out.len(), case.expected_len, "case {}", case.name);
             assert!(
                 (case.expect_first)(&out[0]),

codex-rs/core/src/client_common.rs

@@ -10,14 +10,14 @@ 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");
+/// 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)]
@@ -34,15 +34,14 @@ pub struct Prompt {
 }
 
 impl Prompt {
-    pub(crate) fn get_full_instructions(&self, model: &ModelFamily) -> Cow<'_, str> {
+    pub(crate) fn get_full_instructions<'a>(&'a self, model: &'a ModelFamily) -> Cow<'a, str> {
         let base = self
             .base_instructions_override
             .as_deref()
-            .unwrap_or(BASE_INSTRUCTIONS);
-        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
+            .unwrap_or(model.base_instructions.deref());
+        // 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",
@@ -50,11 +49,13 @@ 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);
+            Cow::Owned(format!("{base}\n{APPLY_PATCH_TOOL_INSTRUCTIONS}"))
+        } else {
+            Cow::Borrowed(base)
         }
-        Cow::Owned(sections.join("\n"))
     }
 
     pub(crate) fn get_formatted_input(&self) -> Vec<ResponseItem> {
@@ -81,8 +82,10 @@ 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.
@@ -136,14 +139,17 @@ 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(
@@ -169,18 +175,64 @@ 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: "gpt-5-codex",
+                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]

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

@@ -0,0 +1,399 @@
+use std::sync::Arc;
+
+use super::AgentTask;
+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) async 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).await;
+}
+
+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,
+) {
+    let start_event = Event {
+        id: sub_id.clone(),
+        msg: EventMsg::TaskStarted(TaskStartedEvent {
+            model_context_window: turn_context.client.get_model_context_window(),
+        }),
+    };
+    sess.send_event(start_event).await;
+    run_compact_task_inner(
+        sess.clone(),
+        turn_context,
+        sub_id.clone(),
+        input,
+        compact_instructions,
+        true,
+    )
+    .await;
+    let event = Event {
+        id: sub_id,
+        msg: EventMsg::TaskComplete(TaskCompleteEvent {
+            last_agent_message: None,
+        }),
+    };
+    sess.send_event(event).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 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()])
+        .await;
+
+    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).await;
+    }
+    let history_snapshot = {
+        let state = sess.state.lock().await;
+        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().await;
+        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;
+}
+
+pub 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()
+}
+
+pub 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().await;
+                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.rs

@@ -1,6 +1,7 @@
 use crate::config_profile::ConfigProfile;
 use crate::config_types::History;
 use crate::config_types::McpServerConfig;
+use crate::config_types::Notifications;
 use crate::config_types::ReasoningSummaryFormat;
 use crate::config_types::SandboxWorkspaceWrite;
 use crate::config_types::ShellEnvironmentPolicy;
@@ -9,6 +10,7 @@ use crate::config_types::Tui;
 use crate::config_types::UriBasedFileOpener;
 use crate::git_info::resolve_root_git_project_for_trust;
 use crate::model_family::ModelFamily;
+use crate::model_family::derive_default_model_family;
 use crate::model_family::find_family_for_model;
 use crate::model_provider_info::ModelProviderInfo;
 use crate::model_provider_info::built_in_model_providers;
@@ -24,15 +26,20 @@ use codex_protocol::mcp_protocol::Tools;
 use codex_protocol::mcp_protocol::UserSavedConfig;
 use dirs::home_dir;
 use serde::Deserialize;
+use std::collections::BTreeMap;
 use std::collections::HashMap;
 use std::path::Path;
 use std::path::PathBuf;
 use tempfile::NamedTempFile;
 use toml::Value as TomlValue;
+use toml_edit::Array as TomlArray;
 use toml_edit::DocumentMut;
+use toml_edit::Item as TomlItem;
+use toml_edit::Table as TomlTable;
 
 const OPENAI_DEFAULT_MODEL: &str = "gpt-5";
-pub const GPT5_HIGH_MODEL: &str = "gpt-5-high";
+const OPENAI_DEFAULT_REVIEW_MODEL: &str = "gpt-5-codex";
+pub const GPT_5_CODEX_MEDIUM_MODEL: &str = "gpt-5-codex";
 
 /// Maximum number of bytes of the documentation that will be embedded. Larger
 /// files are *silently truncated* to this size so we do not take up too much of
@@ -41,12 +48,17 @@ pub(crate) const PROJECT_DOC_MAX_BYTES: usize = 32 * 1024; // 32 KiB
 
 pub(crate) const CONFIG_TOML_FILE: &str = "config.toml";
 
+const DEFAULT_OTEL_ENVIRONMENT: &str = "dev";
+
 /// Application configuration loaded from disk and merged with overrides.
 #[derive(Debug, Clone, PartialEq)]
 pub struct Config {
     /// Optional override of model selection.
     pub model: String,
 
+    /// Model used specifically for review sessions. Defaults to "gpt-5".
+    pub review_model: String,
+
     pub model_family: ModelFamily,
 
     /// Size of the context window for the model, in tokens.
@@ -55,6 +67,9 @@ pub struct Config {
     /// Maximum number of output tokens.
     pub model_max_output_tokens: Option<u64>,
 
+    /// Token usage threshold triggering auto-compaction of conversation history.
+    pub model_auto_compact_token_limit: Option<i64>,
+
     /// Key into the model_providers map that specifies which provider to use.
     pub model_provider_id: String,
 
@@ -105,6 +120,10 @@ pub struct Config {
     /// If unset the feature is disabled.
     pub notify: Option<Vec<String>>,
 
+    /// TUI notifications preference. When set, the TUI will send OSC 9 notifications on approvals
+    /// and turn completions when not focused.
+    pub tui_notifications: Notifications,
+
     /// The directory that should be treated as the current working directory
     /// for the session. All relative paths inside the business-logic layer are
     /// resolved against this path.
@@ -140,7 +159,7 @@ pub struct Config {
 
     /// Value to use for `reasoning.effort` when making a request using the
     /// Responses API.
-    pub model_reasoning_effort: ReasoningEffort,
+    pub model_reasoning_effort: Option<ReasoningEffort>,
 
     /// If not "none", the value to use for `reasoning.summary` when making a
     /// request using the Responses API.
@@ -152,9 +171,6 @@ pub struct Config {
     /// Base URL for requests to ChatGPT (as opposed to the OpenAI API).
     pub chatgpt_base_url: String,
 
-    /// Experimental rollout resume path (absolute path to .jsonl; undocumented).
-    pub experimental_resume: Option<PathBuf>,
-
     /// Include an experimental plan tool that the model can use to update its current plan and status of each step.
     pub include_plan_tool: bool,
 
@@ -180,6 +196,9 @@ pub struct Config {
     /// All characters are inserted as they are received, and no buffering
     /// or placeholder replacement will occur for fast keypress bursts.
     pub disable_paste_burst: bool,
+
+    /// OTEL configuration (exporter type, endpoint, headers, etc.).
+    pub otel: crate::config_types::OtelConfig,
 }
 
 impl Config {
@@ -259,6 +278,88 @@ pub fn load_config_as_toml(codex_home: &Path) -> std::io::Result<TomlValue> {
     }
 }
 
+pub fn load_global_mcp_servers(
+    codex_home: &Path,
+) -> std::io::Result<BTreeMap<String, McpServerConfig>> {
+    let root_value = load_config_as_toml(codex_home)?;
+    let Some(servers_value) = root_value.get("mcp_servers") else {
+        return Ok(BTreeMap::new());
+    };
+
+    servers_value
+        .clone()
+        .try_into()
+        .map_err(|e| std::io::Error::new(std::io::ErrorKind::InvalidData, e))
+}
+
+pub fn write_global_mcp_servers(
+    codex_home: &Path,
+    servers: &BTreeMap<String, McpServerConfig>,
+) -> std::io::Result<()> {
+    let config_path = codex_home.join(CONFIG_TOML_FILE);
+    let mut doc = match std::fs::read_to_string(&config_path) {
+        Ok(contents) => contents
+            .parse::<DocumentMut>()
+            .map_err(|e| std::io::Error::new(std::io::ErrorKind::InvalidData, e))?,
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => DocumentMut::new(),
+        Err(e) => return Err(e),
+    };
+
+    doc.as_table_mut().remove("mcp_servers");
+
+    if !servers.is_empty() {
+        let mut table = TomlTable::new();
+        table.set_implicit(true);
+        doc["mcp_servers"] = TomlItem::Table(table);
+
+        for (name, config) in servers {
+            let mut entry = TomlTable::new();
+            entry.set_implicit(false);
+            entry["command"] = toml_edit::value(config.command.clone());
+
+            if !config.args.is_empty() {
+                let mut args = TomlArray::new();
+                for arg in &config.args {
+                    args.push(arg.clone());
+                }
+                entry["args"] = TomlItem::Value(args.into());
+            }
+
+            if let Some(env) = &config.env
+                && !env.is_empty()
+            {
+                let mut env_table = TomlTable::new();
+                env_table.set_implicit(false);
+                let mut pairs: Vec<_> = env.iter().collect();
+                pairs.sort_by(|(a, _), (b, _)| a.cmp(b));
+                for (key, value) in pairs {
+                    env_table.insert(key, toml_edit::value(value.clone()));
+                }
+                entry["env"] = TomlItem::Table(env_table);
+            }
+
+            if let Some(timeout) = config.startup_timeout_ms {
+                let timeout = i64::try_from(timeout).map_err(|_| {
+                    std::io::Error::new(
+                        std::io::ErrorKind::InvalidData,
+                        "startup_timeout_ms exceeds supported range",
+                    )
+                })?;
+                entry["startup_timeout_ms"] = toml_edit::value(timeout);
+            }
+
+            doc["mcp_servers"][name.as_str()] = TomlItem::Table(entry);
+        }
+    }
+
+    std::fs::create_dir_all(codex_home)?;
+    let tmp_file = NamedTempFile::new_in(codex_home)?;
+    std::fs::write(tmp_file.path(), doc.to_string())?;
+    tmp_file.persist(config_path).map_err(|err| err.error)?;
+
+    Ok(())
+}
+
 fn set_project_trusted_inner(doc: &mut DocumentMut, project_path: &Path) -> anyhow::Result<()> {
     // Ensure we render a human-friendly structure:
     //
@@ -423,14 +524,24 @@ pub async fn persist_model_selection(
     if let Some(profile_name) = active_profile {
         let profile_table = ensure_profile_table(&mut doc, profile_name)?;
         profile_table["model"] = toml_edit::value(model);
-        if let Some(effort) = effort {
-            profile_table["model_reasoning_effort"] = toml_edit::value(effort.to_string());
+        match effort {
+            Some(effort) => {
+                profile_table["model_reasoning_effort"] = toml_edit::value(effort.to_string());
+            }
+            None => {
+                profile_table.remove("model_reasoning_effort");
+            }
         }
     } else {
         let table = doc.as_table_mut();
         table["model"] = toml_edit::value(model);
-        if let Some(effort) = effort {
-            table["model_reasoning_effort"] = toml_edit::value(effort.to_string());
+        match effort {
+            Some(effort) => {
+                table["model_reasoning_effort"] = toml_edit::value(effort.to_string());
+            }
+            None => {
+                table.remove("model_reasoning_effort");
+            }
         }
     }
 
@@ -499,6 +610,8 @@ fn apply_toml_override(root: &mut TomlValue, path: &str, value: TomlValue) {
 pub struct ConfigToml {
     /// Optional override of model selection.
     pub model: Option<String>,
+    /// Review model override used by the `/review` feature.
+    pub review_model: Option<String>,
 
     /// Provider to use from the model_providers map.
     pub model_provider: Option<String>,
@@ -509,6 +622,9 @@ pub struct ConfigToml {
     /// Maximum number of output tokens.
     pub model_max_output_tokens: Option<u64>,
 
+    /// Token usage threshold triggering auto-compaction of conversation history.
+    pub model_auto_compact_token_limit: Option<i64>,
+
     /// Default approval policy for executing commands.
     pub approval_policy: Option<AskForApproval>,
 
@@ -579,9 +695,6 @@ pub struct ConfigToml {
     /// Base URL for requests to ChatGPT (as opposed to the OpenAI API).
     pub chatgpt_base_url: Option<String>,
 
-    /// Experimental rollout resume path (absolute path to .jsonl; undocumented).
-    pub experimental_resume: Option<PathBuf>,
-
     /// Experimental path to a file whose contents replace the built-in BASE_INSTRUCTIONS.
     pub experimental_instructions_file: Option<PathBuf>,
 
@@ -597,6 +710,9 @@ pub struct ConfigToml {
     /// All characters are inserted as they are received, and no buffering
     /// or placeholder replacement will occur for fast keypress bursts.
     pub disable_paste_burst: Option<bool>,
+
+    /// OTEL configuration.
+    pub otel: Option<crate::config_types::OtelConfigToml>,
 }
 
 impl From<ConfigToml> for UserSavedConfig {
@@ -724,6 +840,7 @@ impl ConfigToml {
 #[derive(Default, Debug, Clone)]
 pub struct ConfigOverrides {
     pub model: Option<String>,
+    pub review_model: Option<String>,
     pub cwd: Option<PathBuf>,
     pub approval_policy: Option<AskForApproval>,
     pub sandbox_mode: Option<SandboxMode>,
@@ -751,6 +868,7 @@ impl Config {
         // Destructure ConfigOverrides fully to ensure all overrides are applied.
         let ConfigOverrides {
             model,
+            review_model: override_review_model,
             cwd,
             approval_policy,
             sandbox_mode,
@@ -841,15 +959,8 @@ impl Config {
             .or(cfg.model)
             .unwrap_or_else(default_model);
 
-        let mut model_family = find_family_for_model(&model).unwrap_or_else(|| ModelFamily {
-            slug: model.clone(),
-            family: model.clone(),
-            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,
-        });
+        let mut model_family =
+            find_family_for_model(&model).unwrap_or_else(|| derive_default_model_family(&model));
 
         if let Some(supports_reasoning_summaries) = cfg.model_supports_reasoning_summaries {
             model_family.supports_reasoning_summaries = supports_reasoning_summaries;
@@ -867,8 +978,11 @@ impl Config {
                 .as_ref()
                 .map(|info| info.max_output_tokens)
         });
-
-        let experimental_resume = cfg.experimental_resume;
+        let model_auto_compact_token_limit = cfg.model_auto_compact_token_limit.or_else(|| {
+            openai_model_info
+                .as_ref()
+                .and_then(|info| info.auto_compact_token_limit)
+        });
 
         // Load base instructions override from a file if specified. If the
         // path is relative, resolve it against the effective cwd so the
@@ -881,11 +995,18 @@ impl Config {
             Self::get_base_instructions(experimental_instructions_path, &resolved_cwd)?;
         let base_instructions = base_instructions.or(file_base_instructions);
 
+        // Default review model when not set in config; allow CLI override to take precedence.
+        let review_model = override_review_model
+            .or(cfg.review_model)
+            .unwrap_or_else(default_review_model);
+
         let config = Self {
             model,
+            review_model,
             model_family,
             model_context_window,
             model_max_output_tokens,
+            model_auto_compact_token_limit,
             model_provider_id,
             model_provider,
             cwd: resolved_cwd,
@@ -913,8 +1034,7 @@ impl Config {
                 .unwrap_or(false),
             model_reasoning_effort: config_profile
                 .model_reasoning_effort
-                .or(cfg.model_reasoning_effort)
-                .unwrap_or_default(),
+                .or(cfg.model_reasoning_effort),
             model_reasoning_summary: config_profile
                 .model_reasoning_summary
                 .or(cfg.model_reasoning_summary)
@@ -924,8 +1044,6 @@ impl Config {
                 .chatgpt_base_url
                 .or(cfg.chatgpt_base_url)
                 .unwrap_or("https://chatgpt.com/backend-api/".to_string()),
-
-            experimental_resume,
             include_plan_tool: include_plan_tool.unwrap_or(false),
             include_apply_patch_tool: include_apply_patch_tool.unwrap_or(false),
             tools_web_search_request,
@@ -938,6 +1056,27 @@ impl Config {
             include_view_image_tool,
             active_profile: active_profile_name,
             disable_paste_burst: cfg.disable_paste_burst.unwrap_or(false),
+            tui_notifications: cfg
+                .tui
+                .as_ref()
+                .map(|t| t.notifications.clone())
+                .unwrap_or_default(),
+            otel: {
+                use crate::config_types::OtelConfig;
+                use crate::config_types::OtelConfigToml;
+                use crate::config_types::OtelExporterKind;
+                let t: OtelConfigToml = cfg.otel.unwrap_or_default();
+                let log_user_prompt = t.log_user_prompt.unwrap_or(false);
+                let environment = t
+                    .environment
+                    .unwrap_or(DEFAULT_OTEL_ENVIRONMENT.to_string());
+                let exporter = t.exporter.unwrap_or(OtelExporterKind::None);
+                OtelConfig {
+                    log_user_prompt,
+                    environment,
+                    exporter,
+                }
+            },
         };
         Ok(config)
     }
@@ -1006,6 +1145,10 @@ fn default_model() -> String {
     OPENAI_DEFAULT_MODEL.to_string()
 }
 
+fn default_review_model() -> String {
+    OPENAI_DEFAULT_REVIEW_MODEL.to_string()
+}
+
 /// Returns the path to the Codex configuration directory, which can be
 /// specified by the `CODEX_HOME` environment variable. If not set, defaults to
 /// `~/.codex`.
@@ -1044,6 +1187,7 @@ pub fn log_dir(cfg: &Config) -> std::io::Result<PathBuf> {
 #[cfg(test)]
 mod tests {
     use crate::config_types::HistoryPersistence;
+    use crate::config_types::Notifications;
 
     use super::*;
     use pretty_assertions::assert_eq;
@@ -1082,6 +1226,19 @@ persistence = "none"
         );
     }
 
+    #[test]
+    fn tui_config_missing_notifications_field_defaults_to_disabled() {
+        let cfg = r#"
+[tui]
+"#;
+
+        let parsed = toml::from_str::<ConfigToml>(cfg)
+            .expect("TUI config without notifications should succeed");
+        let tui = parsed.tui.expect("config should include tui section");
+
+        assert_eq!(tui.notifications, Notifications::Enabled(false));
+    }
+
     #[test]
     fn test_sandbox_config_parsing() {
         let sandbox_full_access = r#"
@@ -1138,6 +1295,47 @@ exclude_slash_tmp = true
         );
     }
 
+    #[test]
+    fn load_global_mcp_servers_returns_empty_if_missing() -> anyhow::Result<()> {
+        let codex_home = TempDir::new()?;
+
+        let servers = load_global_mcp_servers(codex_home.path())?;
+        assert!(servers.is_empty());
+
+        Ok(())
+    }
+
+    #[test]
+    fn write_global_mcp_servers_round_trips_entries() -> anyhow::Result<()> {
+        let codex_home = TempDir::new()?;
+
+        let mut servers = BTreeMap::new();
+        servers.insert(
+            "docs".to_string(),
+            McpServerConfig {
+                command: "echo".to_string(),
+                args: vec!["hello".to_string()],
+                env: None,
+                startup_timeout_ms: None,
+            },
+        );
+
+        write_global_mcp_servers(codex_home.path(), &servers)?;
+
+        let loaded = load_global_mcp_servers(codex_home.path())?;
+        assert_eq!(loaded.len(), 1);
+        let docs = loaded.get("docs").expect("docs entry");
+        assert_eq!(docs.command, "echo");
+        assert_eq!(docs.args, vec!["hello".to_string()]);
+
+        let empty = BTreeMap::new();
+        write_global_mcp_servers(codex_home.path(), &empty)?;
+        let loaded = load_global_mcp_servers(codex_home.path())?;
+        assert!(loaded.is_empty());
+
+        Ok(())
+    }
+
     #[tokio::test]
     async fn persist_model_selection_updates_defaults() -> anyhow::Result<()> {
         let codex_home = TempDir::new()?;
@@ -1145,7 +1343,7 @@ exclude_slash_tmp = true
         persist_model_selection(
             codex_home.path(),
             None,
-            "gpt-5-high-new",
+            "gpt-5-codex",
             Some(ReasoningEffort::High),
         )
         .await?;
@@ -1154,7 +1352,7 @@ exclude_slash_tmp = true
             tokio::fs::read_to_string(codex_home.path().join(CONFIG_TOML_FILE)).await?;
         let parsed: ConfigToml = toml::from_str(&serialized)?;
 
-        assert_eq!(parsed.model.as_deref(), Some("gpt-5-high-new"));
+        assert_eq!(parsed.model.as_deref(), Some("gpt-5-codex"));
         assert_eq!(parsed.model_reasoning_effort, Some(ReasoningEffort::High));
 
         Ok(())
@@ -1208,8 +1406,8 @@ model = "gpt-4.1"
         persist_model_selection(
             codex_home.path(),
             Some("dev"),
-            "gpt-5-high-new",
-            Some(ReasoningEffort::Low),
+            "gpt-5-codex",
+            Some(ReasoningEffort::Medium),
         )
         .await?;
 
@@ -1221,8 +1419,11 @@ model = "gpt-4.1"
             .get("dev")
             .expect("profile should be created");
 
-        assert_eq!(profile.model.as_deref(), Some("gpt-5-high-new"));
-        assert_eq!(profile.model_reasoning_effort, Some(ReasoningEffort::Low));
+        assert_eq!(profile.model.as_deref(), Some("gpt-5-codex"));
+        assert_eq!(
+            profile.model_reasoning_effort,
+            Some(ReasoningEffort::Medium)
+        );
 
         Ok(())
     }
@@ -1418,9 +1619,11 @@ model_verbosity = "high"
         assert_eq!(
             Config {
                 model: "o3".to_string(),
+                review_model: OPENAI_DEFAULT_REVIEW_MODEL.to_string(),
                 model_family: find_family_for_model("o3").expect("known model slug"),
                 model_context_window: Some(200_000),
                 model_max_output_tokens: Some(100_000),
+                model_auto_compact_token_limit: None,
                 model_provider_id: "openai".to_string(),
                 model_provider: fixture.openai_provider.clone(),
                 approval_policy: AskForApproval::Never,
@@ -1438,11 +1641,10 @@ model_verbosity = "high"
                 codex_linux_sandbox_exe: None,
                 hide_agent_reasoning: false,
                 show_raw_agent_reasoning: false,
-                model_reasoning_effort: ReasoningEffort::High,
+                model_reasoning_effort: Some(ReasoningEffort::High),
                 model_reasoning_summary: ReasoningSummary::Detailed,
                 model_verbosity: None,
                 chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
-                experimental_resume: None,
                 base_instructions: None,
                 include_plan_tool: false,
                 include_apply_patch_tool: false,
@@ -1452,6 +1654,12 @@ model_verbosity = "high"
                 include_view_image_tool: true,
                 active_profile: Some("o3".to_string()),
                 disable_paste_burst: false,
+                tui_notifications: Default::default(),
+                otel: crate::config_types::OtelConfig {
+                    log_user_prompt: false,
+                    environment: DEFAULT_OTEL_ENVIRONMENT.to_string(),
+                    exporter: crate::config_types::OtelExporterKind::None,
+                },
             },
             o3_profile_config
         );
@@ -1474,9 +1682,11 @@ model_verbosity = "high"
         )?;
         let expected_gpt3_profile_config = Config {
             model: "gpt-3.5-turbo".to_string(),
+            review_model: OPENAI_DEFAULT_REVIEW_MODEL.to_string(),
             model_family: find_family_for_model("gpt-3.5-turbo").expect("known model slug"),
             model_context_window: Some(16_385),
             model_max_output_tokens: Some(4_096),
+            model_auto_compact_token_limit: None,
             model_provider_id: "openai-chat-completions".to_string(),
             model_provider: fixture.openai_chat_completions_provider.clone(),
             approval_policy: AskForApproval::UnlessTrusted,
@@ -1494,11 +1704,10 @@ model_verbosity = "high"
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
             show_raw_agent_reasoning: false,
-            model_reasoning_effort: ReasoningEffort::default(),
+            model_reasoning_effort: None,
             model_reasoning_summary: ReasoningSummary::default(),
             model_verbosity: None,
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
-            experimental_resume: None,
             base_instructions: None,
             include_plan_tool: false,
             include_apply_patch_tool: false,
@@ -1508,6 +1717,12 @@ model_verbosity = "high"
             include_view_image_tool: true,
             active_profile: Some("gpt3".to_string()),
             disable_paste_burst: false,
+            tui_notifications: Default::default(),
+            otel: crate::config_types::OtelConfig {
+                log_user_prompt: false,
+                environment: DEFAULT_OTEL_ENVIRONMENT.to_string(),
+                exporter: crate::config_types::OtelExporterKind::None,
+            },
         };
 
         assert_eq!(expected_gpt3_profile_config, gpt3_profile_config);
@@ -1545,9 +1760,11 @@ model_verbosity = "high"
         )?;
         let expected_zdr_profile_config = Config {
             model: "o3".to_string(),
+            review_model: OPENAI_DEFAULT_REVIEW_MODEL.to_string(),
             model_family: find_family_for_model("o3").expect("known model slug"),
             model_context_window: Some(200_000),
             model_max_output_tokens: Some(100_000),
+            model_auto_compact_token_limit: None,
             model_provider_id: "openai".to_string(),
             model_provider: fixture.openai_provider.clone(),
             approval_policy: AskForApproval::OnFailure,
@@ -1565,11 +1782,10 @@ model_verbosity = "high"
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
             show_raw_agent_reasoning: false,
-            model_reasoning_effort: ReasoningEffort::default(),
+            model_reasoning_effort: None,
             model_reasoning_summary: ReasoningSummary::default(),
             model_verbosity: None,
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
-            experimental_resume: None,
             base_instructions: None,
             include_plan_tool: false,
             include_apply_patch_tool: false,
@@ -1579,6 +1795,12 @@ model_verbosity = "high"
             include_view_image_tool: true,
             active_profile: Some("zdr".to_string()),
             disable_paste_burst: false,
+            tui_notifications: Default::default(),
+            otel: crate::config_types::OtelConfig {
+                log_user_prompt: false,
+                environment: DEFAULT_OTEL_ENVIRONMENT.to_string(),
+                exporter: crate::config_types::OtelExporterKind::None,
+            },
         };
 
         assert_eq!(expected_zdr_profile_config, zdr_profile_config);
@@ -1602,9 +1824,11 @@ model_verbosity = "high"
         )?;
         let expected_gpt5_profile_config = Config {
             model: "gpt-5".to_string(),
+            review_model: OPENAI_DEFAULT_REVIEW_MODEL.to_string(),
             model_family: find_family_for_model("gpt-5").expect("known model slug"),
             model_context_window: Some(272_000),
             model_max_output_tokens: Some(128_000),
+            model_auto_compact_token_limit: None,
             model_provider_id: "openai".to_string(),
             model_provider: fixture.openai_provider.clone(),
             approval_policy: AskForApproval::OnFailure,
@@ -1622,11 +1846,10 @@ model_verbosity = "high"
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
             show_raw_agent_reasoning: false,
-            model_reasoning_effort: ReasoningEffort::High,
+            model_reasoning_effort: Some(ReasoningEffort::High),
             model_reasoning_summary: ReasoningSummary::Detailed,
             model_verbosity: Some(Verbosity::High),
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
-            experimental_resume: None,
             base_instructions: None,
             include_plan_tool: false,
             include_apply_patch_tool: false,
@@ -1636,6 +1859,12 @@ model_verbosity = "high"
             include_view_image_tool: true,
             active_profile: Some("gpt5".to_string()),
             disable_paste_burst: false,
+            tui_notifications: Default::default(),
+            otel: crate::config_types::OtelConfig {
+                log_user_prompt: false,
+                environment: DEFAULT_OTEL_ENVIRONMENT.to_string(),
+                exporter: crate::config_types::OtelExporterKind::None,
+            },
         };
 
         assert_eq!(expected_gpt5_profile_config, gpt5_profile_config);
@@ -1739,3 +1968,46 @@ trust_level = "trusted"
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod notifications_tests {
+    use crate::config_types::Notifications;
+    use serde::Deserialize;
+
+    #[derive(Deserialize, Debug, PartialEq)]
+    struct TuiTomlTest {
+        notifications: Notifications,
+    }
+
+    #[derive(Deserialize, Debug, PartialEq)]
+    struct RootTomlTest {
+        tui: TuiTomlTest,
+    }
+
+    #[test]
+    fn test_tui_notifications_true() {
+        let toml = r#"
+            [tui]
+            notifications = true
+        "#;
+        let parsed: RootTomlTest = toml::from_str(toml).expect("deserialize notifications=true");
+        assert!(matches!(
+            parsed.tui.notifications,
+            Notifications::Enabled(true)
+        ));
+    }
+
+    #[test]
+    fn test_tui_notifications_custom_array() {
+        let toml = r#"
+            [tui]
+            notifications = ["foo"]
+        "#;
+        let parsed: RootTomlTest =
+            toml::from_str(toml).expect("deserialize notifications=[\"foo\"]");
+        assert!(matches!(
+            parsed.tui.notifications,
+            Notifications::Custom(ref v) if v == &vec!["foo".to_string()]
+        ));
+    }
+}

codex-rs/core/src/config_edit.rs

@@ -7,6 +7,12 @@ 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(
@@ -14,47 +20,12 @@ pub async fn persist_overrides(
     profile: Option<&str>,
     overrides: &[(&[&str], &str)],
 ) -> Result<()> {
-    let config_path = codex_home.join(CONFIG_TOML_FILE);
+    let with_options: Vec<(&[&str], Option<&str>)> = overrides
+        .iter()
+        .map(|(segments, value)| (*segments, Some(*value)))
+        .collect();
 
-    let mut doc = match tokio::fs::read_to_string(&config_path).await {
-        Ok(s) => s.parse::<DocumentMut>()?,
-        Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
-            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())
-    };
-
-    for (segments, val) in overrides.iter().copied() {
-        let value = toml_edit::value(val);
-        if let Some(ref name) = effective_profile {
-            if segments.first().copied() == Some("profiles") {
-                apply_toml_edit_override_segments(&mut doc, segments, value);
-            } else {
-                let mut seg_buf: Vec<&str> = Vec::with_capacity(2 + segments.len());
-                seg_buf.push("profiles");
-                seg_buf.push(name.as_str());
-                seg_buf.extend_from_slice(segments);
-                apply_toml_edit_override_segments(&mut doc, &seg_buf, value);
-            }
-        } else {
-            apply_toml_edit_override_segments(&mut doc, segments, value);
-        }
-    }
-
-    let tmp_file = NamedTempFile::new_in(codex_home)?;
-    tokio::fs::write(tmp_file.path(), doc.to_string()).await?;
-    tmp_file.persist(config_path)?;
-
-    Ok(())
+    persist_overrides_with_behavior(codex_home, profile, &with_options, NoneBehavior::Skip).await
 }
 
 /// Persist overrides where values may be optional. Any entries with `None`
@@ -65,16 +36,17 @@ pub async fn persist_non_null_overrides(
     profile: Option<&str>,
     overrides: &[(&[&str], Option<&str>)],
 ) -> Result<()> {
-    let filtered: Vec<(&[&str], &str)> = overrides
-        .iter()
-        .filter_map(|(k, v)| v.map(|vv| (*k, vv)))
-        .collect();
+    persist_overrides_with_behavior(codex_home, profile, overrides, NoneBehavior::Skip).await
+}
 
-    if filtered.is_empty() {
-        return Ok(());
-    }
-
-    persist_overrides(codex_home, profile, &filtered).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
@@ -121,6 +93,125 @@ fn apply_toml_edit_override_segments(
     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::*;
@@ -574,6 +665,81 @@ 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);

codex-rs/core/src/config_types.rs

@@ -76,9 +76,75 @@ pub enum HistoryPersistence {
     None,
 }
 
+// ===== OTEL configuration =====
+
+#[derive(Deserialize, Debug, Clone, PartialEq)]
+#[serde(rename_all = "kebab-case")]
+pub enum OtelHttpProtocol {
+    /// Binary payload
+    Binary,
+    /// JSON payload
+    Json,
+}
+
+/// Which OTEL exporter to use.
+#[derive(Deserialize, Debug, Clone, PartialEq)]
+#[serde(rename_all = "kebab-case")]
+pub enum OtelExporterKind {
+    None,
+    OtlpHttp {
+        endpoint: String,
+        headers: HashMap<String, String>,
+        protocol: OtelHttpProtocol,
+    },
+    OtlpGrpc {
+        endpoint: String,
+        headers: HashMap<String, String>,
+    },
+}
+
+/// OTEL settings loaded from config.toml. Fields are optional so we can apply defaults.
+#[derive(Deserialize, Debug, Clone, PartialEq, Default)]
+pub struct OtelConfigToml {
+    /// Log user prompt in traces
+    pub log_user_prompt: Option<bool>,
+
+    /// Mark traces with environment (dev, staging, prod, test). Defaults to dev.
+    pub environment: Option<String>,
+
+    /// Exporter to use. Defaults to `otlp-file`.
+    pub exporter: Option<OtelExporterKind>,
+}
+
+/// Effective OTEL settings after defaults are applied.
+#[derive(Debug, Clone, PartialEq)]
+pub struct OtelConfig {
+    pub log_user_prompt: bool,
+    pub environment: String,
+    pub exporter: OtelExporterKind,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Deserialize)]
+#[serde(untagged)]
+pub enum Notifications {
+    Enabled(bool),
+    Custom(Vec<String>),
+}
+
+impl Default for Notifications {
+    fn default() -> Self {
+        Self::Enabled(false)
+    }
+}
+
 /// Collection of settings that are specific to the TUI.
 #[derive(Deserialize, Debug, Clone, PartialEq, Default)]
-pub struct Tui {}
+pub struct Tui {
+    /// Enable desktop notifications from the TUI when the terminal is unfocused.
+    /// Defaults to `false`.
+    #[serde(default)]
+    pub notifications: Notifications,
+}
 
 #[derive(Deserialize, Debug, Clone, PartialEq, Default)]
 pub struct SandboxWorkspaceWrite {

codex-rs/core/src/conversation_history.rs

@@ -1,4 +1,3 @@
-use codex_protocol::models::ContentItem;
 use codex_protocol::models::ResponseItem;
 
 /// Transcript of conversation history
@@ -33,52 +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 last_agent_message(&self) -> String {
-        for item in self.items.iter().rev() {
-            if let ResponseItem::Message { role, content, .. } = item
-                && role == "assistant"
-            {
-                return content
-                    .iter()
-                    .find_map(|ci| {
-                        if let ContentItem::OutputText { text } = ci {
-                            Some(text.clone())
-                        } else {
-                            None
-                        }
-                    })
-                    .unwrap_or_default();
-            }
-        }
-        String::new()
+    pub(crate) fn replace(&mut self, items: Vec<ResponseItem>) {
+        self.items = items;
     }
 }
 
@@ -92,8 +47,9 @@ fn is_api_message(message: &ResponseItem) -> bool {
         | ResponseItem::CustomToolCall { .. }
         | ResponseItem::CustomToolCallOutput { .. }
         | ResponseItem::LocalShellCall { .. }
-        | ResponseItem::Reasoning { .. } => true,
-        ResponseItem::WebSearchCall { .. } | ResponseItem::Other => false,
+        | ResponseItem::Reasoning { .. }
+        | ResponseItem::WebSearchCall { .. } => true,
+        ResponseItem::Other => false,
     }
 }
 

codex-rs/core/src/conversation_manager.rs

@@ -3,6 +3,8 @@ use crate::CodexAuth;
 use crate::codex::Codex;
 use crate::codex::CodexSpawnOk;
 use crate::codex::INITIAL_SUBMIT_ID;
+use crate::codex::compact::content_items_to_text;
+use crate::codex::compact::is_session_prefix_message;
 use crate::codex_conversation::CodexConversation;
 use crate::config::Config;
 use crate::error::CodexErr;
@@ -59,21 +61,11 @@ impl ConversationManager {
         config: Config,
         auth_manager: Arc<AuthManager>,
     ) -> CodexResult<NewConversation> {
-        // TO BE REFACTORED: use the config experimental_resume field until we have a mainstream way.
-        if let Some(resume_path) = config.experimental_resume.as_ref() {
-            let initial_history = RolloutRecorder::get_rollout_history(resume_path).await?;
-            let CodexSpawnOk {
-                codex,
-                conversation_id,
-            } = Codex::spawn(config, auth_manager, initial_history).await?;
-            self.finalize_spawn(codex, conversation_id).await
-        } else {
-            let CodexSpawnOk {
-                codex,
-                conversation_id,
-            } = Codex::spawn(config, auth_manager, InitialHistory::New).await?;
-            self.finalize_spawn(codex, conversation_id).await
-        }
+        let CodexSpawnOk {
+            codex,
+            conversation_id,
+        } = Codex::spawn(config, auth_manager, InitialHistory::New).await?;
+        self.finalize_spawn(codex, conversation_id).await
     }
 
     async fn finalize_spawn(
@@ -144,19 +136,19 @@ impl ConversationManager {
         self.conversations.write().await.remove(conversation_id)
     }
 
-    /// Fork an existing conversation by dropping the last `drop_last_messages`
-    /// user/assistant messages from its transcript and starting a new
+    /// Fork an existing conversation by taking messages up to the given position
+    /// (not including the message at the given position) and starting a new
     /// conversation with identical configuration (unless overridden by the
     /// caller's `config`). The new conversation will have a fresh id.
     pub async fn fork_conversation(
         &self,
-        num_messages_to_drop: usize,
+        nth_user_message: usize,
         config: Config,
         path: PathBuf,
     ) -> CodexResult<NewConversation> {
         // Compute the prefix up to the cut point.
         let history = RolloutRecorder::get_rollout_history(&path).await?;
-        let history = truncate_after_dropping_last_messages(history, num_messages_to_drop);
+        let history = truncate_before_nth_user_message(history, nth_user_message);
 
         // Spawn a new conversation with the computed initial history.
         let auth_manager = self.auth_manager.clone();
@@ -169,33 +161,30 @@ 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(history: InitialHistory, n: usize) -> InitialHistory {
-    if n == 0 {
-        return InitialHistory::Forked(history.get_rollout_items());
-    }
-
-    // Work directly on rollout items, and cut the vector at the nth-from-last user message input.
+/// Return a prefix of `items` obtained by cutting strictly before the nth user message
+/// (0-based) and all items that follow it.
+fn truncate_before_nth_user_message(history: InitialHistory, n: usize) -> InitialHistory {
+    // Work directly on rollout items, and cut the vector at the nth 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
+        if let RolloutItem::ResponseItem(ResponseItem::Message { role, content, .. }) = item
             && role == "user"
+            && content_items_to_text(content).is_some_and(|text| !is_session_prefix_message(&text))
         {
             user_positions.push(idx);
         }
     }
 
-    // If fewer than n user messages exist, treat as empty.
-    if user_positions.len() < n {
+    // If fewer than or equal to n user messages exist, treat as empty (out of range).
+    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];
+    // Cut strictly before the nth user message (do not keep the nth itself).
+    let cut_idx = user_positions[n];
     let rolled: Vec<RolloutItem> = items.into_iter().take(cut_idx).collect();
 
     if rolled.is_empty() {
@@ -208,9 +197,11 @@ fn truncate_after_dropping_last_messages(history: InitialHistory, n: usize) -> I
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::codex::make_session_and_context;
     use codex_protocol::models::ContentItem;
     use codex_protocol::models::ReasoningItemReasoningSummary;
     use codex_protocol::models::ResponseItem;
+    use pretty_assertions::assert_eq;
 
     fn user_msg(text: &str) -> ResponseItem {
         ResponseItem::Message {
@@ -262,7 +253,7 @@ mod tests {
             .cloned()
             .map(RolloutItem::ResponseItem)
             .collect();
-        let truncated = truncate_after_dropping_last_messages(InitialHistory::Forked(initial), 1);
+        let truncated = truncate_before_nth_user_message(InitialHistory::Forked(initial), 1);
         let got_items = truncated.get_rollout_items();
         let expected_items = vec![
             RolloutItem::ResponseItem(items[0].clone()),
@@ -279,7 +270,37 @@ mod tests {
             .cloned()
             .map(RolloutItem::ResponseItem)
             .collect();
-        let truncated2 = truncate_after_dropping_last_messages(InitialHistory::Forked(initial2), 2);
+        let truncated2 = truncate_before_nth_user_message(InitialHistory::Forked(initial2), 2);
         assert!(matches!(truncated2, InitialHistory::New));
     }
+
+    #[test]
+    fn ignores_session_prefix_messages_when_truncating() {
+        let (session, turn_context) = make_session_and_context();
+        let mut items = session.build_initial_context(&turn_context);
+        items.push(user_msg("feature request"));
+        items.push(assistant_msg("ack"));
+        items.push(user_msg("second question"));
+        items.push(assistant_msg("answer"));
+
+        let rollout_items: Vec<RolloutItem> = items
+            .iter()
+            .cloned()
+            .map(RolloutItem::ResponseItem)
+            .collect();
+
+        let truncated = truncate_before_nth_user_message(InitialHistory::Forked(rollout_items), 1);
+        let got_items = truncated.get_rollout_items();
+
+        let expected: Vec<RolloutItem> = vec![
+            RolloutItem::ResponseItem(items[0].clone()),
+            RolloutItem::ResponseItem(items[1].clone()),
+            RolloutItem::ResponseItem(items[2].clone()),
+        ];
+
+        assert_eq!(
+            serde_json::to_value(&got_items).unwrap(),
+            serde_json::to_value(&expected).unwrap()
+        );
+    }
 }

codex-rs/core/src/environment_context.rs

@@ -2,6 +2,7 @@ use serde::Deserialize;
 use serde::Serialize;
 use strum_macros::Display as DeriveDisplay;
 
+use crate::codex::TurnContext;
 use crate::protocol::AskForApproval;
 use crate::protocol::SandboxPolicy;
 use crate::shell::Shell;
@@ -71,6 +72,39 @@ impl EnvironmentContext {
             shell,
         }
     }
+
+    /// Compares two environment contexts, ignoring the shell. Useful when
+    /// comparing turn to turn, since the initial environment_context will
+    /// include the shell, and then it is not configurable from turn to turn.
+    pub fn equals_except_shell(&self, other: &EnvironmentContext) -> bool {
+        let EnvironmentContext {
+            cwd,
+            approval_policy,
+            sandbox_mode,
+            network_access,
+            writable_roots,
+            // should compare all fields except shell
+            shell: _,
+        } = other;
+
+        self.cwd == *cwd
+            && self.approval_policy == *approval_policy
+            && self.sandbox_mode == *sandbox_mode
+            && self.network_access == *network_access
+            && self.writable_roots == *writable_roots
+    }
+}
+
+impl From<&TurnContext> for EnvironmentContext {
+    fn from(turn_context: &TurnContext) -> Self {
+        Self::new(
+            Some(turn_context.cwd.clone()),
+            Some(turn_context.approval_policy),
+            Some(turn_context.sandbox_policy.clone()),
+            // Shell is not configurable from turn to turn
+            None,
+        )
+    }
 }
 
 impl EnvironmentContext {
@@ -140,6 +174,9 @@ impl From<EnvironmentContext> for ResponseItem {
 
 #[cfg(test)]
 mod tests {
+    use crate::shell::BashShell;
+    use crate::shell::ZshShell;
+
     use super::*;
     use pretty_assertions::assert_eq;
 
@@ -210,4 +247,82 @@ mod tests {
 
         assert_eq!(context.serialize_to_xml(), expected);
     }
+
+    #[test]
+    fn equals_except_shell_compares_approval_policy() {
+        // Approval policy
+        let context1 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(workspace_write_policy(vec!["/repo"], false)),
+            None,
+        );
+        let context2 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::Never),
+            Some(workspace_write_policy(vec!["/repo"], true)),
+            None,
+        );
+        assert!(!context1.equals_except_shell(&context2));
+    }
+
+    #[test]
+    fn equals_except_shell_compares_sandbox_policy() {
+        let context1 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(SandboxPolicy::new_read_only_policy()),
+            None,
+        );
+        let context2 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(SandboxPolicy::new_workspace_write_policy()),
+            None,
+        );
+
+        assert!(!context1.equals_except_shell(&context2));
+    }
+
+    #[test]
+    fn equals_except_shell_compares_workspace_write_policy() {
+        let context1 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(workspace_write_policy(vec!["/repo", "/tmp", "/var"], false)),
+            None,
+        );
+        let context2 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(workspace_write_policy(vec!["/repo", "/tmp"], true)),
+            None,
+        );
+
+        assert!(!context1.equals_except_shell(&context2));
+    }
+
+    #[test]
+    fn equals_except_shell_ignores_shell() {
+        let context1 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(workspace_write_policy(vec!["/repo"], false)),
+            Some(Shell::Bash(BashShell {
+                shell_path: "/bin/bash".into(),
+                bashrc_path: "/home/user/.bashrc".into(),
+            })),
+        );
+        let context2 = EnvironmentContext::new(
+            Some(PathBuf::from("/repo")),
+            Some(AskForApproval::OnRequest),
+            Some(workspace_write_policy(vec!["/repo"], false)),
+            Some(Shell::Zsh(ZshShell {
+                shell_path: "/bin/zsh".into(),
+                zshrc_path: "/home/user/.zshrc".into(),
+            })),
+        );
+
+        assert!(context1.equals_except_shell(&context2));
+    }
 }

codex-rs/core/src/error.rs

@@ -1,3 +1,4 @@
+use crate::exec::ExecToolCallOutput;
 use crate::token_data::KnownPlan;
 use crate::token_data::PlanType;
 use codex_protocol::mcp_protocol::ConversationId;
@@ -13,8 +14,11 @@ 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")]
@@ -28,7 +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")]
@@ -245,9 +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(),
     }
 }

codex-rs/core/src/exec.rs

@@ -3,6 +3,7 @@ use std::os::unix::process::ExitStatusExt;
 
 use std::collections::HashMap;
 use std::io;
+use std::path::Path;
 use std::path::PathBuf;
 use std::process::ExitStatus;
 use std::time::Duration;
@@ -34,6 +35,7 @@ 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
@@ -43,7 +45,7 @@ const AGGREGATE_BUFFER_INITIAL_CAPACITY: usize = 8 * 1024; // 8 KiB
 /// Aggregation still collects full output; only the live event stream is capped.
 pub(crate) const MAX_EXEC_OUTPUT_DELTAS_PER_CALL: usize = 10_000;
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug)]
 pub struct ExecParams {
     pub command: Vec<String>,
     pub cwd: PathBuf,
@@ -81,33 +83,41 @@ pub async fn process_exec_tool_call(
     params: ExecParams,
     sandbox_type: SandboxType,
     sandbox_policy: &SandboxPolicy,
+    sandbox_cwd: &Path,
     codex_linux_sandbox_exe: &Option<PathBuf>,
     stdout_stream: Option<StdoutStream>,
 ) -> 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, ..
+                command,
+                cwd: command_cwd,
+                env,
+                ..
             } = params;
             let child = spawn_command_under_seatbelt(
                 command,
+                command_cwd,
                 sandbox_policy,
-                cwd,
+                sandbox_cwd,
                 StdioPolicy::RedirectForShellTool,
                 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, ..
+                command,
+                cwd: command_cwd,
+                env,
+                ..
             } = params;
 
             let codex_linux_sandbox_exe = codex_linux_sandbox_exe
@@ -116,48 +126,64 @@ pub async fn process_exec_tool_call(
             let child = spawn_command_under_linux_sandbox(
                 codex_linux_sandbox_exe,
                 command,
+                command_cwd,
                 sandbox_policy,
-                cwd,
+                sandbox_cwd,
                 StdioPolicy::RedirectForShellTool,
                 env,
             )
             .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}");
@@ -197,6 +223,7 @@ 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> {
@@ -229,6 +256,7 @@ pub struct ExecToolCallOutput {
     pub stderr: StreamOutput<String>,
     pub aggregated_output: StreamOutput<String>,
     pub duration: Duration,
+    pub timed_out: bool,
 }
 
 async fn exec(
@@ -298,22 +326,24 @@ 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)
         }
     };
 
@@ -336,6 +366,7 @@ async fn consume_truncated_output(
         stdout,
         stderr,
         aggregated_output,
+        timed_out,
     })
 }
 

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

@@ -38,16 +38,20 @@ impl ExecCommandSession {
         writer_handle: JoinHandle<()>,
         wait_handle: JoinHandle<()>,
         exit_status: std::sync::Arc<std::sync::atomic::AtomicBool>,
-    ) -> Self {
-        Self {
-            writer_tx,
-            output_tx,
-            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,
-        }
+    ) -> (Self, broadcast::Receiver<Vec<u8>>) {
+        let initial_output_rx = output_tx.subscribe();
+        (
+            Self {
+                writer_tx,
+                output_tx,
+                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,
+            },
+            initial_output_rx,
+        )
     }
 
     pub(crate) fn writer_sender(&self) -> mpsc::Sender<Vec<u8>> {

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

@@ -93,18 +93,16 @@ impl SessionManager {
                 .fetch_add(1, std::sync::atomic::Ordering::SeqCst),
         );
 
-        let (session, mut exit_rx) =
-            create_exec_command_session(params.clone())
-                .await
-                .map_err(|err| {
-                    format!(
-                        "failed to create exec command session for session id {}: {err}",
-                        session_id.0
-                    )
-                })?;
+        let (session, mut output_rx, mut exit_rx) = create_exec_command_session(params.clone())
+            .await
+            .map_err(|err| {
+                format!(
+                    "failed to create exec command session for session id {}: {err}",
+                    session_id.0
+                )
+            })?;
 
         // Insert into session map.
-        let mut output_rx = session.output_receiver();
         self.sessions.lock().await.insert(session_id, session);
 
         // Collect output until either timeout expires or process exits.
@@ -245,7 +243,11 @@ impl SessionManager {
 /// Spawn PTY and child process per spawn_exec_command_session logic.
 async fn create_exec_command_session(
     params: ExecCommandParams,
-) -> anyhow::Result<(ExecCommandSession, oneshot::Receiver<i32>)> {
+) -> anyhow::Result<(
+    ExecCommandSession,
+    tokio::sync::broadcast::Receiver<Vec<u8>>,
+    oneshot::Receiver<i32>,
+)> {
     let ExecCommandParams {
         cmd,
         yield_time_ms: _,
@@ -279,7 +281,6 @@ 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);
-
     // Reader task: drain PTY and forward chunks to output channel.
     let mut reader = pair.master.try_clone_reader()?;
     let output_tx_clone = output_tx.clone();
@@ -341,7 +342,7 @@ async fn create_exec_command_session(
     });
 
     // Create and store the session with channels.
-    let session = ExecCommandSession::new(
+    let (session, initial_output_rx) = ExecCommandSession::new(
         writer_tx,
         output_tx,
         killer,
@@ -350,7 +351,7 @@ async fn create_exec_command_session(
         wait_handle,
         exit_status,
     );
-    Ok((session, exit_rx))
+    Ok((session, initial_output_rx, exit_rx))
 }
 
 #[cfg(test)]

codex-rs/core/src/internal_storage.rs

@@ -1,6 +1,7 @@
 use anyhow::Context;
 use serde::Deserialize;
 use serde::Serialize;
+use std::io::ErrorKind;
 use std::path::Path;
 use std::path::PathBuf;
 
@@ -11,7 +12,7 @@ pub struct InternalStorage {
     #[serde(skip)]
     storage_path: PathBuf,
     #[serde(default)]
-    pub gpt_5_high_model_prompt_seen: bool,
+    pub gpt_5_codex_model_prompt_seen: bool,
 }
 
 // TODO(jif) generalise all the file writers and build proper async channel inserters.
@@ -31,7 +32,14 @@ impl InternalStorage {
                 }
             },
             Err(error) => {
-                tracing::warn!("failed to read internal storage: {error:?}");
+                if error.kind() == ErrorKind::NotFound {
+                    tracing::debug!(
+                        "internal storage not found at {}; initializing defaults",
+                        storage_path.display()
+                    );
+                } else {
+                    tracing::warn!("failed to read internal storage: {error:?}");
+                }
                 Self::empty(storage_path)
             }
         }

codex-rs/core/src/landlock.rs

@@ -16,21 +16,22 @@ use tokio::process::Child;
 pub async fn spawn_command_under_linux_sandbox<P>(
     codex_linux_sandbox_exe: P,
     command: Vec<String>,
+    command_cwd: PathBuf,
     sandbox_policy: &SandboxPolicy,
-    cwd: PathBuf,
+    sandbox_policy_cwd: &Path,
     stdio_policy: StdioPolicy,
     env: HashMap<String, String>,
 ) -> std::io::Result<Child>
 where
     P: AsRef<Path>,
 {
-    let args = create_linux_sandbox_command_args(command, sandbox_policy, &cwd);
+    let args = create_linux_sandbox_command_args(command, sandbox_policy, sandbox_policy_cwd);
     let arg0 = Some("codex-linux-sandbox");
     spawn_child_async(
         codex_linux_sandbox_exe.as_ref().to_path_buf(),
         args,
         arg0,
-        cwd,
+        command_cwd,
         sandbox_policy,
         stdio_policy,
         env,
@@ -42,10 +43,13 @@ where
 fn create_linux_sandbox_command_args(
     command: Vec<String>,
     sandbox_policy: &SandboxPolicy,
-    cwd: &Path,
+    sandbox_policy_cwd: &Path,
 ) -> Vec<String> {
     #[expect(clippy::expect_used)]
-    let sandbox_policy_cwd = cwd.to_str().expect("cwd must be valid UTF-8").to_string();
+    let sandbox_policy_cwd = sandbox_policy_cwd
+        .to_str()
+        .expect("cwd must be valid UTF-8")
+        .to_string();
 
     #[expect(clippy::expect_used)]
     let sandbox_policy_json =

codex-rs/core/src/lib.rs

@@ -46,6 +46,7 @@ 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 mod review_format;
 pub use codex_protocol::protocol::InitialHistory;
 pub use conversation_manager::ConversationManager;
 pub use conversation_manager::NewConversation;
@@ -70,6 +71,7 @@ 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;
@@ -87,11 +89,16 @@ pub use codex_protocol::config_types as protocol_config_types;
 
 pub use client::ModelClient;
 pub use client_common::Prompt;
+pub use client_common::REVIEW_PROMPT;
 pub use client_common::ResponseEvent;
 pub use client_common::ResponseStream;
+pub use codex::compact::content_items_to_text;
+pub use codex::compact::is_session_prefix_message;
 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;
+
+pub mod otel_init;

codex-rs/core/src/model_family.rs

@@ -1,6 +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 GPT_5_CODEX_INSTRUCTIONS: &str = include_str!("../gpt_5_codex_prompt.md");
+
 /// A model family is a group of models that share certain characteristics.
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
 pub struct ModelFamily {
@@ -33,6 +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 {
@@ -48,6 +56,7 @@ macro_rules! model_family {
             reasoning_summary_format: ReasoningSummaryFormat::None,
             uses_local_shell_tool: false,
             apply_patch_tool_type: None,
+            base_instructions: BASE_INSTRUCTIONS.to_string(),
         };
         // apply overrides
         $(
@@ -57,22 +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,
-            reasoning_summary_format: ReasoningSummaryFormat::None,
-            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> {
@@ -80,23 +73,20 @@ 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,
-            reasoning_summary_format: ReasoningSummaryFormat::Experimental,
+            needs_special_apply_patch_instructions: true,
         )
     } else if slug.starts_with("gpt-4.1") {
         model_family!(
@@ -106,15 +96,36 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
     } 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("gpt-5-codex") {
+        model_family!(
+            slug, slug,
+            supports_reasoning_summaries: true,
+            reasoning_summary_format: ReasoningSummaryFormat::Experimental,
+            base_instructions: GPT_5_CODEX_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

@@ -162,6 +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.
@@ -329,6 +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::*;
@@ -419,4 +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/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)),
 
-        _ if slug.starts_with("gpt-5") => Some(ModelInfo {
-            context_window: 272_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: 272_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

@@ -7,8 +7,6 @@ use std::collections::HashMap;
 
 use crate::model_family::ModelFamily;
 use crate::plan_tool::PLAN_TOOL;
-use crate::protocol::AskForApproval;
-use crate::protocol::SandboxPolicy;
 use crate::tool_apply_patch::ApplyPatchToolType;
 use crate::tool_apply_patch::create_apply_patch_freeform_tool;
 use crate::tool_apply_patch::create_apply_patch_json_tool;
@@ -57,10 +55,9 @@ pub(crate) enum OpenAiTool {
 
 #[derive(Debug, Clone)]
 pub enum ConfigShellToolType {
-    DefaultShell,
-    ShellWithRequest { sandbox_policy: SandboxPolicy },
-    LocalShell,
-    StreamableShell,
+    Default,
+    Local,
+    Streamable,
 }
 
 #[derive(Debug, Clone)]
@@ -75,8 +72,6 @@ pub(crate) struct ToolsConfig {
 
 pub(crate) struct ToolsConfigParams<'a> {
     pub(crate) model_family: &'a ModelFamily,
-    pub(crate) approval_policy: AskForApproval,
-    pub(crate) sandbox_policy: SandboxPolicy,
     pub(crate) include_plan_tool: bool,
     pub(crate) include_apply_patch_tool: bool,
     pub(crate) include_web_search_request: bool,
@@ -89,8 +84,6 @@ impl ToolsConfig {
     pub fn new(params: &ToolsConfigParams) -> Self {
         let ToolsConfigParams {
             model_family,
-            approval_policy,
-            sandbox_policy,
             include_plan_tool,
             include_apply_patch_tool,
             include_web_search_request,
@@ -98,18 +91,13 @@ impl ToolsConfig {
             include_view_image_tool,
             experimental_unified_exec_tool,
         } = params;
-        let mut shell_type = if *use_streamable_shell_tool {
-            ConfigShellToolType::StreamableShell
+        let shell_type = if *use_streamable_shell_tool {
+            ConfigShellToolType::Streamable
         } else if model_family.uses_local_shell_tool {
-            ConfigShellToolType::LocalShell
+            ConfigShellToolType::Local
         } else {
-            ConfigShellToolType::DefaultShell
+            ConfigShellToolType::Default
         };
-        if matches!(approval_policy, AskForApproval::OnRequest) && !use_streamable_shell_tool {
-            shell_type = ConfigShellToolType::ShellWithRequest {
-                sandbox_policy: sandbox_policy.clone(),
-            }
-        }
 
         let apply_patch_tool_type = match model_family.apply_patch_tool_type {
             Some(ApplyPatchToolType::Freeform) => Some(ApplyPatchToolType::Freeform),
@@ -170,40 +158,6 @@ pub(crate) enum JsonSchema {
     },
 }
 
-fn create_shell_tool() -> OpenAiTool {
-    let mut properties = BTreeMap::new();
-    properties.insert(
-        "command".to_string(),
-        JsonSchema::Array {
-            items: Box::new(JsonSchema::String { description: None }),
-            description: Some("The command to execute".to_string()),
-        },
-    );
-    properties.insert(
-        "workdir".to_string(),
-        JsonSchema::String {
-            description: Some("The working directory to execute the command in".to_string()),
-        },
-    );
-    properties.insert(
-        "timeout_ms".to_string(),
-        JsonSchema::Number {
-            description: Some("The timeout for the command in milliseconds".to_string()),
-        },
-    );
-
-    OpenAiTool::Function(ResponsesApiTool {
-        name: "shell".to_string(),
-        description: "Runs a shell command and returns its output".to_string(),
-        strict: false,
-        parameters: JsonSchema::Object {
-            properties,
-            required: Some(vec!["command".to_string()]),
-            additional_properties: Some(false),
-        },
-    })
-}
-
 fn create_unified_exec_tool() -> OpenAiTool {
     let mut properties = BTreeMap::new();
     properties.insert(
@@ -251,7 +205,7 @@ fn create_unified_exec_tool() -> OpenAiTool {
     })
 }
 
-fn create_shell_tool_for_sandbox(sandbox_policy: &SandboxPolicy) -> OpenAiTool {
+fn create_shell_tool() -> OpenAiTool {
     let mut properties = BTreeMap::new();
     properties.insert(
         "command".to_string(),
@@ -273,73 +227,22 @@ fn create_shell_tool_for_sandbox(sandbox_policy: &SandboxPolicy) -> OpenAiTool {
         },
     );
 
-    if matches!(sandbox_policy, SandboxPolicy::WorkspaceWrite { .. }) {
-        properties.insert(
+    properties.insert(
         "with_escalated_permissions".to_string(),
         JsonSchema::Boolean {
             description: Some("Whether to request escalated permissions. Set to true if command needs to be run without sandbox restrictions".to_string()),
         },
     );
-        properties.insert(
+    properties.insert(
         "justification".to_string(),
         JsonSchema::String {
             description: Some("Only set if with_escalated_permissions is true. 1-sentence explanation of why we want to run this command.".to_string()),
         },
     );
-    }
-
-    let description = match sandbox_policy {
-        SandboxPolicy::WorkspaceWrite {
-            network_access,
-            ..
-        } => {
-            let network_line = if !network_access {
-                "\n    - Commands that require network access"
-            } else {
-                ""
-            };
-
-            format!(
-                r#"
-The shell tool is used to execute shell commands.
-- When invoking the shell tool, your call will be running in a sandbox, and some shell commands will require escalated privileges:
-  - Types of actions that require escalated privileges:
-    - Writing files other than those in the writable roots (see the environment context for the allowed directories){network_line}
-  - 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."#,
-            )
-        }
-        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 sandbox, and some shell commands (including apply_patch) will require escalated permissions:
-  - Types of actions that require escalated privileges:
-    - 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,
@@ -382,7 +285,7 @@ pub(crate) struct ApplyPatchToolArgs {
 /// Responses API:
 /// https://platform.openai.com/docs/guides/function-calling?api-mode=responses
 pub fn create_tools_json_for_responses_api(
-    tools: &Vec<OpenAiTool>,
+    tools: &[OpenAiTool],
 ) -> crate::error::Result<Vec<serde_json::Value>> {
     let mut tools_json = Vec::new();
 
@@ -397,7 +300,7 @@ pub fn create_tools_json_for_responses_api(
 /// Chat Completions API:
 /// https://platform.openai.com/docs/guides/function-calling?api-mode=chat
 pub(crate) fn create_tools_json_for_chat_completions_api(
-    tools: &Vec<OpenAiTool>,
+    tools: &[OpenAiTool],
 ) -> crate::error::Result<Vec<serde_json::Value>> {
     // We start with the JSON for the Responses API and than rewrite it to match
     // the chat completions tool call format.
@@ -586,16 +489,13 @@ pub(crate) fn get_openai_tools(
         tools.push(create_unified_exec_tool());
     } else {
         match &config.shell_type {
-            ConfigShellToolType::DefaultShell => {
+            ConfigShellToolType::Default => {
                 tools.push(create_shell_tool());
             }
-            ConfigShellToolType::ShellWithRequest { sandbox_policy } => {
-                tools.push(create_shell_tool_for_sandbox(sandbox_policy));
-            }
-            ConfigShellToolType::LocalShell => {
+            ConfigShellToolType::Local => {
                 tools.push(OpenAiTool::LocalShell {});
             }
-            ConfigShellToolType::StreamableShell => {
+            ConfigShellToolType::Streamable => {
                 tools.push(OpenAiTool::Function(
                     crate::exec_command::create_exec_command_tool_for_responses_api(),
                 ));
@@ -685,8 +585,6 @@ mod tests {
             .expect("codex-mini-latest should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: true,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -707,8 +605,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: true,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -729,8 +625,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: false,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -835,8 +729,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: false,
             include_apply_patch_tool: false,
             include_web_search_request: false,
@@ -913,8 +805,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: false,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -976,8 +866,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: false,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -1034,8 +922,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: false,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -1095,8 +981,6 @@ mod tests {
         let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
         let config = ToolsConfig::new(&ToolsConfigParams {
             model_family: &model_family,
-            approval_policy: AskForApproval::Never,
-            sandbox_policy: SandboxPolicy::ReadOnly,
             include_plan_tool: false,
             include_apply_patch_tool: false,
             include_web_search_request: true,
@@ -1149,14 +1033,8 @@ 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);
+    fn test_shell_tool() {
+        let tool = super::create_shell_tool();
         let OpenAiTool::Function(ResponsesApiTool {
             description, name, ..
         }) = &tool
@@ -1165,63 +1043,7 @@ mod tests {
         };
         assert_eq!(name, "shell");
 
-        let expected = r#"
-The shell tool is used to execute shell commands.
-- When invoking the shell tool, your call will be running in a sandbox, and some shell commands will require escalated privileges:
-  - Types of actions that require escalated privileges:
-    - Writing files other than those in the writable roots (see the environment context for the allowed directories)
-    - Commands that require network access
-  - 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."#;
+        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 = r#"
-The shell tool is used to execute shell commands.
-- When invoking the shell tool, your call will be running in a sandbox, and some shell commands (including apply_patch) will require escalated permissions:
-  - Types of actions that require escalated privileges:
-    - 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"#;
-        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/otel_init.rs

@@ -0,0 +1,61 @@
+use crate::config::Config;
+use crate::config_types::OtelExporterKind as Kind;
+use crate::config_types::OtelHttpProtocol as Protocol;
+use crate::default_client::ORIGINATOR;
+use codex_otel::config::OtelExporter;
+use codex_otel::config::OtelHttpProtocol;
+use codex_otel::config::OtelSettings;
+use codex_otel::otel_provider::OtelProvider;
+use std::error::Error;
+
+/// Build an OpenTelemetry provider from the app Config.
+///
+/// Returns `None` when OTEL export is disabled.
+pub fn build_provider(
+    config: &Config,
+    service_version: &str,
+) -> Result<Option<OtelProvider>, Box<dyn Error>> {
+    let exporter = match &config.otel.exporter {
+        Kind::None => OtelExporter::None,
+        Kind::OtlpHttp {
+            endpoint,
+            headers,
+            protocol,
+        } => {
+            let protocol = match protocol {
+                Protocol::Json => OtelHttpProtocol::Json,
+                Protocol::Binary => OtelHttpProtocol::Binary,
+            };
+
+            OtelExporter::OtlpHttp {
+                endpoint: endpoint.clone(),
+                headers: headers
+                    .iter()
+                    .map(|(k, v)| (k.clone(), v.clone()))
+                    .collect(),
+                protocol,
+            }
+        }
+        Kind::OtlpGrpc { endpoint, headers } => OtelExporter::OtlpGrpc {
+            endpoint: endpoint.clone(),
+            headers: headers
+                .iter()
+                .map(|(k, v)| (k.clone(), v.clone()))
+                .collect(),
+        },
+    };
+
+    OtelProvider::from(&OtelSettings {
+        service_name: ORIGINATOR.value.to_owned(),
+        service_version: service_version.to_string(),
+        codex_home: config.codex_home.clone(),
+        environment: config.otel.environment.to_string(),
+        exporter,
+    })
+}
+
+/// Filter predicate for exporting only Codex-owned events via OTEL.
+/// Keeps events that originated from codex_otel module
+pub fn codex_export_filter(meta: &tracing::Metadata<'_>) -> bool {
+    meta.target().starts_with("codex_otel")
+}

codex-rs/core/src/parse_command.rs

@@ -1156,10 +1156,8 @@ fn parse_bash_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
         // bias toward the primary command when pipelines are present.
         // First, drop obvious small formatting helpers (e.g., wc/awk/etc).
         let had_multiple_commands = all_commands.len() > 1;
-        // The bash AST walker yields commands in right-to-left order for
-        // connector/pipeline sequences. Reverse to reflect actual execution order.
-        let mut filtered_commands = drop_small_formatting_commands(all_commands);
-        filtered_commands.reverse();
+        // Commands arrive in source order; drop formatting helpers while preserving it.
+        let filtered_commands = drop_small_formatting_commands(all_commands);
         if filtered_commands.is_empty() {
             return Some(vec![ParsedCommand::Unknown {
                 cmd: script.clone(),

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/review_format.rs

@@ -0,0 +1,55 @@
+use crate::protocol::ReviewFinding;
+
+// Note: We keep this module UI-agnostic. It returns plain strings that
+// higher layers (e.g., TUI) may style as needed.
+
+fn format_location(item: &ReviewFinding) -> String {
+    let path = item.code_location.absolute_file_path.display();
+    let start = item.code_location.line_range.start;
+    let end = item.code_location.line_range.end;
+    format!("{path}:{start}-{end}")
+}
+
+/// Format a full review findings block as plain text lines.
+///
+/// - When `selection` is `Some`, each item line includes a checkbox marker:
+///   "[x]" for selected items and "[ ]" for unselected. Missing indices
+///   default to selected.
+/// - When `selection` is `None`, the marker is omitted and a simple bullet is
+///   rendered ("- Title — path:start-end").
+pub fn format_review_findings_block(
+    findings: &[ReviewFinding],
+    selection: Option<&[bool]>,
+) -> String {
+    let mut lines: Vec<String> = Vec::new();
+    lines.push(String::new());
+
+    // Header
+    if findings.len() > 1 {
+        lines.push("Full review comments:".to_string());
+    } else {
+        lines.push("Review comment:".to_string());
+    }
+
+    for (idx, item) in findings.iter().enumerate() {
+        lines.push(String::new());
+
+        let title = &item.title;
+        let location = format_location(item);
+
+        if let Some(flags) = selection {
+            // Default to selected if index is out of bounds.
+            let checked = flags.get(idx).copied().unwrap_or(true);
+            let marker = if checked { "[x]" } else { "[ ]" };
+            lines.push(format!("- {marker} {title} — {location}"));
+        } else {
+            lines.push(format!("- {title} — {location}"));
+        }
+
+        for body_line in item.body.lines() {
+            lines.push(format!("  {body_line}"));
+        }
+    }
+
+    lines.join("\n")
+}

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

@@ -3,6 +3,10 @@ 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;
@@ -334,3 +338,48 @@ async fn read_head_and_flags(
 
     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

@@ -8,6 +8,7 @@ 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;
 

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

@@ -25,8 +25,9 @@ pub(crate) fn should_persist_response_item(item: &ResponseItem) -> bool {
         | ResponseItem::FunctionCall { .. }
         | ResponseItem::FunctionCallOutput { .. }
         | ResponseItem::CustomToolCall { .. }
-        | ResponseItem::CustomToolCallOutput { .. } => true,
-        ResponseItem::WebSearchCall { .. } | ResponseItem::Other => false,
+        | ResponseItem::CustomToolCallOutput { .. }
+        | ResponseItem::WebSearchCall { .. } => true,
+        ResponseItem::Other => false,
     }
 }
 
@@ -38,7 +39,10 @@ pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
         | EventMsg::AgentMessage(_)
         | EventMsg::AgentReasoning(_)
         | EventMsg::AgentReasoningRawContent(_)
-        | EventMsg::TokenCount(_) => true,
+        | EventMsg::TokenCount(_)
+        | EventMsg::EnteredReviewMode(_)
+        | EventMsg::ExitedReviewMode(_)
+        | EventMsg::TurnAborted(_) => true,
         EventMsg::Error(_)
         | EventMsg::TaskStarted(_)
         | EventMsg::TaskComplete(_)
@@ -65,7 +69,6 @@ pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
         | EventMsg::McpListToolsResponse(_)
         | EventMsg::ListCustomPromptsResponse(_)
         | EventMsg::PlanUpdate(_)
-        | EventMsg::TurnAborted(_)
         | EventMsg::ShutdownComplete
         | EventMsg::ConversationPath(_) => false,
     }

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

@@ -204,7 +204,6 @@ impl RolloutRecorder {
 
     pub(crate) async fn get_rollout_history(path: &Path) -> std::io::Result<InitialHistory> {
         info!("Resuming rollout from {path:?}");
-        tracing::error!("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"));
@@ -254,7 +253,7 @@ impl RolloutRecorder {
             }
         }
 
-        tracing::error!(
+        info!(
             "Resumed rollout with {} items, conversation ID: {:?}",
             items.len(),
             conversation_id

codex-rs/core/src/safety.rs

@@ -13,9 +13,14 @@ use crate::protocol::SandboxPolicy;
 
 #[derive(Debug, PartialEq)]
 pub enum SafetyCheck {
-    AutoApprove { sandbox_type: SandboxType },
+    AutoApprove {
+        sandbox_type: SandboxType,
+        user_explicitly_approved: bool,
+    },
     AskUser,
-    Reject { reason: String },
+    Reject {
+        reason: String,
+    },
 }
 
 pub fn assess_patch_safety(
@@ -52,12 +57,16 @@ pub fn assess_patch_safety(
         // fall back to asking the user because the patch may touch arbitrary
         // paths outside the project.
         match get_platform_sandbox() {
-            Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+            Some(sandbox_type) => SafetyCheck::AutoApprove {
+                sandbox_type,
+                user_explicitly_approved: false,
+            },
             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,
+                    user_explicitly_approved: false,
                 }
             }
             None => SafetyCheck::AskUser,
@@ -101,6 +110,7 @@ pub fn assess_command_safety(
     if is_known_safe_command(command) || approved.contains(command) {
         return SafetyCheck::AutoApprove {
             sandbox_type: SandboxType::None,
+            user_explicitly_approved: false,
         };
     }
 
@@ -126,13 +136,17 @@ pub(crate) fn assess_safety_for_untrusted_command(
         | (Never, DangerFullAccess)
         | (OnRequest, DangerFullAccess) => SafetyCheck::AutoApprove {
             sandbox_type: SandboxType::None,
+            user_explicitly_approved: false,
         },
         (OnRequest, ReadOnly) | (OnRequest, WorkspaceWrite { .. }) => {
             if with_escalated_permissions {
                 SafetyCheck::AskUser
             } else {
                 match get_platform_sandbox() {
-                    Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+                    Some(sandbox_type) => SafetyCheck::AutoApprove {
+                        sandbox_type,
+                        user_explicitly_approved: false,
+                    },
                     // Fall back to asking since the command is untrusted and
                     // we do not have a sandbox available
                     None => SafetyCheck::AskUser,
@@ -144,7 +158,10 @@ pub(crate) fn assess_safety_for_untrusted_command(
         | (OnFailure, ReadOnly)
         | (OnFailure, WorkspaceWrite { .. }) => {
             match get_platform_sandbox() {
-                Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+                Some(sandbox_type) => SafetyCheck::AutoApprove {
+                    sandbox_type,
+                    user_explicitly_approved: false,
+                },
                 None => {
                     if matches!(approval_policy, OnFailure) {
                         // Since the command is not trusted, even though the
@@ -342,7 +359,10 @@ mod tests {
         );
 
         let expected = match get_platform_sandbox() {
-            Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+            Some(sandbox_type) => SafetyCheck::AutoApprove {
+                sandbox_type,
+                user_explicitly_approved: false,
+            },
             None => SafetyCheck::AskUser,
         };
         assert_eq!(safety_check, expected);

codex-rs/core/src/seatbelt.rs

@@ -18,19 +18,20 @@ const MACOS_PATH_TO_SEATBELT_EXECUTABLE: &str = "/usr/bin/sandbox-exec";
 
 pub async fn spawn_command_under_seatbelt(
     command: Vec<String>,
+    command_cwd: PathBuf,
     sandbox_policy: &SandboxPolicy,
-    cwd: PathBuf,
+    sandbox_policy_cwd: &Path,
     stdio_policy: StdioPolicy,
     mut env: HashMap<String, String>,
 ) -> std::io::Result<Child> {
-    let args = create_seatbelt_command_args(command, sandbox_policy, &cwd);
+    let args = create_seatbelt_command_args(command, sandbox_policy, sandbox_policy_cwd);
     let arg0 = None;
     env.insert(CODEX_SANDBOX_ENV_VAR.to_string(), "seatbelt".to_string());
     spawn_child_async(
         PathBuf::from(MACOS_PATH_TO_SEATBELT_EXECUTABLE),
         args,
         arg0,
-        cwd,
+        command_cwd,
         sandbox_policy,
         stdio_policy,
         env,
@@ -41,7 +42,7 @@ pub async fn spawn_command_under_seatbelt(
 fn create_seatbelt_command_args(
     command: Vec<String>,
     sandbox_policy: &SandboxPolicy,
-    cwd: &Path,
+    sandbox_policy_cwd: &Path,
 ) -> Vec<String> {
     let (file_write_policy, extra_cli_args) = {
         if sandbox_policy.has_full_disk_write_access() {
@@ -51,7 +52,7 @@ fn create_seatbelt_command_args(
                 Vec::<String>::new(),
             )
         } else {
-            let writable_roots = sandbox_policy.get_writable_roots_with_cwd(cwd);
+            let writable_roots = sandbox_policy.get_writable_roots_with_cwd(sandbox_policy_cwd);
 
             let mut writable_folder_policies: Vec<String> = Vec::new();
             let mut cli_args: Vec<String> = Vec::new();

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")
@@ -63,14 +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)
 
-; needed to look up user info, see https://crbug.com/792228
 (allow mach-lookup
-  (global-name "com.apple.system.opendirectoryd.libinfo")
+  (global-name "com.apple.PowerManagement.control")
 )

codex-rs/core/src/shell.rs

@@ -5,20 +5,20 @@ use std::path::PathBuf;
 
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct ZshShell {
-    shell_path: String,
-    zshrc_path: String,
+    pub(crate) shell_path: String,
+    pub(crate) zshrc_path: String,
 }
 
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct BashShell {
-    shell_path: String,
-    bashrc_path: String,
+    pub(crate) shell_path: String,
+    pub(crate) 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".
-    bash_exe_fallback: Option<PathBuf>, // In case the model generates a bash command.
+    pub(crate) exe: String, // Executable name or path, e.g. "pwsh" or "powershell.exe".
+    pub(crate) bash_exe_fallback: Option<PathBuf>, // In case the model generates a bash command.
 }
 
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
@@ -32,15 +32,19 @@ pub enum Shell {
 impl Shell {
     pub fn format_default_shell_invocation(&self, command: Vec<String>) -> Option<Vec<String>> {
         match self {
-            Shell::Zsh(zsh) => {
-                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::Zsh(zsh) => format_shell_invocation_with_rc(
+                command.as_slice(),
+                &zsh.shell_path,
+                &zsh.zshrc_path,
+            ),
+            Shell::Bash(bash) => format_shell_invocation_with_rc(
+                command.as_slice(),
+                &bash.shell_path,
+                &bash.bashrc_path,
+            ),
             Shell::PowerShell(ps) => {
                 // If model generated a bash command, prefer a detected bash fallback
-                if let Some(script) = strip_bash_lc(&command) {
+                if let Some(script) = strip_bash_lc(command.as_slice()) {
                     return match &ps.bash_exe_fallback {
                         Some(bash) => Some(vec![
                             bash.to_string_lossy().to_string(),
@@ -102,7 +106,7 @@ impl Shell {
 }
 
 fn format_shell_invocation_with_rc(
-    command: &Vec<String>,
+    command: &[String],
     shell_path: &str,
     rc_path: &str,
 ) -> Option<Vec<String>> {
@@ -118,8 +122,8 @@ fn format_shell_invocation_with_rc(
     Some(vec![shell_path.to_string(), "-lc".to_string(), rc_command])
 }
 
-fn strip_bash_lc(command: &Vec<String>) -> Option<String> {
-    match command.as_slice() {
+fn strip_bash_lc(command: &[String]) -> Option<String> {
+    match command {
         // exactly three items
         [first, second, third]
             // first two must be "bash", "-lc"
@@ -345,6 +349,7 @@ mod tests {
                 },
                 SandboxType::None,
                 &SandboxPolicy::DangerFullAccess,
+                temp_home.path(),
                 &None,
                 None,
             )
@@ -451,6 +456,7 @@ mod macos_tests {
                 },
                 SandboxType::None,
                 &SandboxPolicy::DangerFullAccess,
+                temp_home.path(),
                 &None,
                 None,
             )

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

@@ -100,10 +100,13 @@ type OutputBuffer = Arc<Mutex<OutputBufferState>>;
 type OutputHandles = (OutputBuffer, Arc<Notify>);
 
 impl ManagedUnifiedExecSession {
-    fn new(session: ExecCommandSession) -> Self {
+    fn new(
+        session: ExecCommandSession,
+        initial_output_rx: tokio::sync::broadcast::Receiver<Vec<u8>>,
+    ) -> Self {
         let output_buffer = Arc::new(Mutex::new(OutputBufferState::default()));
         let output_notify = Arc::new(Notify::new());
-        let mut receiver = session.output_receiver();
+        let mut receiver = initial_output_rx;
         let buffer_clone = Arc::clone(&output_buffer);
         let notify_clone = Arc::clone(&output_notify);
         let output_task = tokio::spawn(async move {
@@ -193,8 +196,8 @@ impl UnifiedExecSessionManager {
         } 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 (session, initial_output_rx) = create_unified_exec_session(&command).await?;
+            let managed_session = ManagedUnifiedExecSession::new(session, initial_output_rx);
             let (buffer, notify) = managed_session.output_handles();
             writer_tx = managed_session.writer_sender();
             output_buffer = buffer;
@@ -297,7 +300,13 @@ impl UnifiedExecSessionManager {
 
 async fn create_unified_exec_session(
     command: &[String],
-) -> Result<ExecCommandSession, UnifiedExecError> {
+) -> Result<
+    (
+        ExecCommandSession,
+        tokio::sync::broadcast::Receiver<Vec<u8>>,
+    ),
+    UnifiedExecError,
+> {
     if command.is_empty() {
         return Err(UnifiedExecError::MissingCommandLine);
     }
@@ -380,7 +389,7 @@ async fn create_unified_exec_session(
         wait_exit_status.store(true, Ordering::SeqCst);
     });
 
-    Ok(ExecCommandSession::new(
+    let (session, initial_output_rx) = ExecCommandSession::new(
         writer_tx,
         output_tx,
         killer,
@@ -388,7 +397,8 @@ async fn create_unified_exec_session(
         writer_handle,
         wait_handle,
         exit_status,
-    ))
+    );
+    Ok((session, initial_output_rx))
 }
 
 #[cfg(test)]
@@ -421,7 +431,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: None,
                 input_chunks: &["bash".to_string(), "-i".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         let session_id = open_shell.session_id.expect("expected session_id");
@@ -441,7 +451,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: Some(session_id),
                 input_chunks: &["echo $CODEX_INTERACTIVE_SHELL_VAR\n".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         assert!(out_2.output.contains("codex"));
@@ -458,7 +468,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: None,
                 input_chunks: &["/bin/bash".to_string(), "-i".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         let session_a = shell_a.session_id.expect("expected session id");
@@ -467,7 +477,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: Some(session_a),
                 input_chunks: &["export CODEX_INTERACTIVE_SHELL_VAR=codex\n".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
 
@@ -478,7 +488,7 @@ mod tests {
                     "echo".to_string(),
                     "$CODEX_INTERACTIVE_SHELL_VAR\n".to_string(),
                 ],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         assert!(!out_2.output.contains("codex"));
@@ -487,7 +497,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: Some(session_a),
                 input_chunks: &["echo $CODEX_INTERACTIVE_SHELL_VAR\n".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         assert!(out_3.output.contains("codex"));
@@ -504,7 +514,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: None,
                 input_chunks: &["bash".to_string(), "-i".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         let session_id = open_shell.session_id.expect("expected session id");
@@ -516,7 +526,7 @@ mod tests {
                     "export".to_string(),
                     "CODEX_INTERACTIVE_SHELL_VAR=codex\n".to_string(),
                 ],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
 
@@ -547,6 +557,7 @@ mod tests {
 
     #[cfg(unix)]
     #[tokio::test]
+    #[ignore] // Ignored while we have a better way to test this.
     async fn requests_with_large_timeout_are_capped() -> Result<(), UnifiedExecError> {
         let manager = UnifiedExecSessionManager::default();
 
@@ -568,13 +579,14 @@ mod tests {
 
     #[cfg(unix)]
     #[tokio::test]
+    #[ignore] // Ignored while we have a better way to test this.
     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(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
 
@@ -595,7 +607,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: None,
                 input_chunks: &["/bin/bash".to_string(), "-i".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
         let session_id = open_shell.session_id.expect("expected session id");
@@ -604,7 +616,7 @@ mod tests {
             .handle_request(UnifiedExecRequest {
                 session_id: Some(session_id),
                 input_chunks: &["exit\n".to_string()],
-                timeout_ms: Some(1_500),
+                timeout_ms: Some(2_500),
             })
             .await?;
 

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

@@ -11,6 +11,8 @@ use codex_core::ReasoningItemContent;
 use codex_core::ResponseItem;
 use codex_core::WireApi;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_otel::otel_event_manager::OtelEventManager;
+use codex_protocol::mcp_protocol::AuthMode;
 use codex_protocol::mcp_protocol::ConversationId;
 use core_test_support::load_default_config_for_test;
 use futures::StreamExt;
@@ -70,13 +72,26 @@ async fn run_request(input: Vec<ResponseItem>) -> Value {
     let summary = config.model_reasoning_summary;
     let config = Arc::new(config);
 
+    let conversation_id = ConversationId::new();
+
+    let otel_event_manager = OtelEventManager::new(
+        conversation_id,
+        config.model.as_str(),
+        config.model_family.slug.as_str(),
+        None,
+        Some(AuthMode::ChatGPT),
+        false,
+        "test".to_string(),
+    );
+
     let client = ModelClient::new(
         Arc::clone(&config),
         None,
+        otel_event_manager,
         provider,
         effort,
         summary,
-        ConversationId::new(),
+        conversation_id,
     );
 
     let mut prompt = Prompt::default();

codex-rs/core/tests/chat_completions_sse.rs

@@ -1,4 +1,5 @@
 use std::sync::Arc;
+use tracing_test::traced_test;
 
 use codex_core::ContentItem;
 use codex_core::ModelClient;
@@ -8,6 +9,8 @@ use codex_core::ResponseEvent;
 use codex_core::ResponseItem;
 use codex_core::WireApi;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_otel::otel_event_manager::OtelEventManager;
+use codex_protocol::mcp_protocol::AuthMode;
 use codex_protocol::mcp_protocol::ConversationId;
 use core_test_support::load_default_config_for_test;
 use futures::StreamExt;
@@ -23,11 +26,15 @@ fn network_disabled() -> bool {
 }
 
 async fn run_stream(sse_body: &str) -> Vec<ResponseEvent> {
+    run_stream_with_bytes(sse_body.as_bytes()).await
+}
+
+async fn run_stream_with_bytes(sse_body: &[u8]) -> 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");
+        .set_body_bytes(sse_body.to_vec());
 
     Mock::given(method("POST"))
         .and(path("/v1/chat/completions"))
@@ -63,13 +70,26 @@ async fn run_stream(sse_body: &str) -> Vec<ResponseEvent> {
     let summary = config.model_reasoning_summary;
     let config = Arc::new(config);
 
+    let conversation_id = ConversationId::new();
+
+    let otel_event_manager = OtelEventManager::new(
+        conversation_id,
+        config.model.as_str(),
+        config.model_family.slug.as_str(),
+        None,
+        Some(AuthMode::ChatGPT),
+        false,
+        "test".to_string(),
+    );
+
     let client = ModelClient::new(
         Arc::clone(&config),
         None,
+        otel_event_manager,
         provider,
         effort,
         summary,
-        ConversationId::new(),
+        conversation_id,
     );
 
     let mut prompt = Prompt::default();
@@ -89,7 +109,8 @@ async fn run_stream(sse_body: &str) -> Vec<ResponseEvent> {
     while let Some(event) = stream.next().await {
         match event {
             Ok(ev) => events.push(ev),
-            Err(e) => panic!("stream event error: {e}"),
+            // We still collect the error to exercise telemetry and complete the task.
+            Err(_e) => break,
         }
     }
     events
@@ -318,3 +339,88 @@ async fn streams_reasoning_before_tool_call() {
 
     assert!(matches!(events[3], ResponseEvent::Completed { .. }));
 }
+
+#[tokio::test]
+#[traced_test]
+async fn chat_sse_emits_failed_on_parse_error() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse_body = concat!("data: not-json\n\n", "data: [DONE]\n\n");
+
+    let _ = run_stream(sse_body).await;
+
+    logs_assert(|lines: &[&str]| {
+        lines
+            .iter()
+            .find(|line| {
+                line.contains("codex.api_request") && line.contains("http.response.status_code=200")
+            })
+            .map(|_| Ok(()))
+            .unwrap_or(Err("cannot find codex.api_request event".to_string()))
+    });
+
+    logs_assert(|lines: &[&str]| {
+        lines
+            .iter()
+            .find(|line| {
+                line.contains("codex.sse_event")
+                    && line.contains("error.message")
+                    && line.contains("Failed to parse SSE event")
+            })
+            .map(|_| Ok(()))
+            .unwrap_or(Err("cannot find SSE event".to_string()))
+    });
+}
+
+#[tokio::test]
+#[traced_test]
+async fn chat_sse_done_chunk_emits_event() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse_body = concat!("data: [DONE]\n\n",);
+
+    let _ = run_stream(sse_body).await;
+
+    logs_assert(|lines: &[&str]| {
+        lines
+            .iter()
+            .find(|line| line.contains("codex.sse_event") && line.contains("event.kind=message"))
+            .map(|_| Ok(()))
+            .unwrap_or(Err("cannot find SSE event".to_string()))
+    });
+}
+
+#[tokio::test]
+#[traced_test]
+async fn chat_sse_emits_error_on_invalid_utf8() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let _ = run_stream_with_bytes(b"data: \x80\x80\n\n").await;
+
+    logs_assert(|lines: &[&str]| {
+        lines
+            .iter()
+            .find(|line| {
+                line.contains("codex.sse_event")
+                    && line.contains("error.message")
+                    && line.contains("UTF8 error: invalid utf-8 sequence of 1 bytes from index 0")
+            })
+            .map(|_| Ok(()))
+            .unwrap_or(Err("cannot find SSE event".to_string()))
+    });
+}

codex-rs/core/tests/common/Cargo.toml

@@ -11,3 +11,4 @@ codex-core = { path = "../.." }
 serde_json = "1"
 tempfile = "3"
 tokio = { version = "1", features = ["time"] }
+wiremock = "0.6"

codex-rs/core/tests/common/lib.rs

@@ -7,6 +7,8 @@ use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
 use codex_core::config::ConfigToml;
 
+pub mod responses;
+
 /// Returns a default `Config` whose on-disk state is confined to the provided
 /// temporary directory. Using a per-test directory keeps tests hermetic and
 /// avoids clobbering a developer’s real `~/.codex`.

codex-rs/core/tests/common/responses.rs

@@ -0,0 +1,133 @@
+use serde_json::Value;
+use wiremock::BodyPrintLimit;
+use wiremock::Mock;
+use wiremock::MockServer;
+use wiremock::ResponseTemplate;
+use wiremock::matchers::method;
+use wiremock::matchers::path;
+
+/// Build an SSE stream body from a list of JSON events.
+pub fn sse(events: Vec<Value>) -> String {
+    use std::fmt::Write as _;
+    let mut out = String::new();
+    for ev in events {
+        let kind = ev.get("type").and_then(|v| v.as_str()).unwrap();
+        writeln!(&mut out, "event: {kind}").unwrap();
+        if !ev.as_object().map(|o| o.len() == 1).unwrap_or(false) {
+            write!(&mut out, "data: {ev}\n\n").unwrap();
+        } else {
+            out.push('\n');
+        }
+    }
+    out
+}
+
+/// Convenience: SSE event for a completed response with a specific id.
+pub fn ev_completed(id: &str) -> Value {
+    serde_json::json!({
+        "type": "response.completed",
+        "response": {
+            "id": id,
+            "usage": {"input_tokens":0,"input_tokens_details":null,"output_tokens":0,"output_tokens_details":null,"total_tokens":0}
+        }
+    })
+}
+
+pub 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.
+pub fn ev_assistant_message(id: &str, text: &str) -> Value {
+    serde_json::json!({
+        "type": "response.output_item.done",
+        "item": {
+            "type": "message",
+            "role": "assistant",
+            "id": id,
+            "content": [{"type": "output_text", "text": text}]
+        }
+    })
+}
+
+pub 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
+        }
+    })
+}
+
+/// Convenience: SSE event for an `apply_patch` custom tool call with raw patch
+/// text. This mirrors the payload produced by the Responses API when the model
+/// invokes `apply_patch` directly (before we convert it to a function call).
+pub fn ev_apply_patch_custom_tool_call(call_id: &str, patch: &str) -> Value {
+    serde_json::json!({
+        "type": "response.output_item.done",
+        "item": {
+            "type": "custom_tool_call",
+            "name": "apply_patch",
+            "input": patch,
+            "call_id": call_id
+        }
+    })
+}
+
+/// Convenience: SSE event for an `apply_patch` function call. The Responses API
+/// wraps the patch content in a JSON string under the `input` key; we recreate
+/// the same structure so downstream code exercises the full parsing path.
+pub fn ev_apply_patch_function_call(call_id: &str, patch: &str) -> Value {
+    let arguments = serde_json::json!({ "input": patch });
+    let arguments = serde_json::to_string(&arguments).expect("serialize apply_patch arguments");
+
+    serde_json::json!({
+        "type": "response.output_item.done",
+        "item": {
+            "type": "function_call",
+            "name": "apply_patch",
+            "arguments": arguments,
+            "call_id": call_id
+        }
+    })
+}
+
+pub fn sse_response(body: String) -> ResponseTemplate {
+    ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(body, "text/event-stream")
+}
+
+pub async fn mount_sse_once<M>(server: &MockServer, matcher: M, body: String)
+where
+    M: wiremock::Match + Send + Sync + 'static,
+{
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .and(matcher)
+        .respond_with(sse_response(body))
+        .mount(server)
+        .await;
+}
+
+pub async fn start_mock_server() -> MockServer {
+    MockServer::builder()
+        .body_print_limit(BodyPrintLimit::Limited(80_000))
+        .start()
+        .await
+}

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

@@ -420,12 +420,6 @@ async fn integration_creates_and_checks_session_file() {
     // 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")
@@ -434,11 +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)

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

@@ -1,19 +1,36 @@
 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_otel::otel_event_manager::OtelEventManager;
+use codex_protocol::mcp_protocol::AuthMode;
+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 tracing_test::traced_test;
 use uuid::Uuid;
 use wiremock::Mock;
 use wiremock::MockServer;
@@ -222,20 +239,21 @@ async fn resume_includes_initial_messages_and_sends_prior_items() {
     let codex_home = TempDir::new().unwrap();
     let mut config = load_default_config_for_test(&codex_home);
     config.model_provider = model_provider;
-    config.experimental_resume = Some(session_path.clone());
     // 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
-        .new_conversation(config)
+        .resume_conversation_from_rollout(config, session_path.clone(), auth_manager)
         .await
-        .expect("create new conversation");
+        .expect("resume conversation");
 
     // 1) Assert initial_messages only includes existing EventMsg entries; response items are not converted
     let initial_msgs = session_configured
@@ -620,6 +638,160 @@ 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 conversation_id = ConversationId::new();
+
+    let otel_event_manager = OtelEventManager::new(
+        conversation_id,
+        config.model.as_str(),
+        config.model_family.slug.as_str(),
+        None,
+        Some(AuthMode::ChatGPT),
+        false,
+        "test".to_string(),
+    );
+
+    let client = ModelClient::new(
+        Arc::clone(&config),
+        None,
+        otel_event_manager,
+        provider,
+        effort,
+        summary,
+        conversation_id,
+    );
+
+    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" };
@@ -918,3 +1090,57 @@ async fn history_dedupes_streamed_and_final_messages_across_turns() {
         "request 3 tail mismatch",
     );
 }
+
+#[tokio::test]
+#[traced_test]
+async fn responses_api_emits_api_request_event() {
+    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.user_instructions = Some("be nice".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 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;
+
+    logs_assert(|lines: &[&str]| {
+        lines
+            .iter()
+            .find(|line| line.contains("codex.api_request"))
+            .map(|_| Ok(()))
+            .unwrap_or_else(|| Err("expected codex.api_request event".to_string()))
+    });
+}

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

@@ -1,10 +1,9 @@
-#![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;
@@ -12,82 +11,42 @@ use codex_core::protocol::RolloutItem;
 use codex_core::protocol::RolloutLine;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
 use core_test_support::load_default_config_for_test;
+use core_test_support::responses;
 use core_test_support::wait_for_event;
-use serde_json::Value;
 use tempfile::TempDir;
 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 responses::ev_assistant_message;
+use responses::ev_completed;
+use responses::sse;
+use responses::start_mock_server;
+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 {
-    use std::fmt::Write as _;
-    let mut out = String::new();
-    for ev in events {
-        let kind = ev.get("type").and_then(|v| v.as_str()).unwrap();
-        writeln!(&mut out, "event: {kind}").unwrap();
-        if !ev.as_object().map(|o| o.len() == 1).unwrap_or(false) {
-            write!(&mut out, "data: {ev}\n\n").unwrap();
-        } else {
-            out.push('\n');
-        }
-    }
-    out
-}
-
-/// Convenience: SSE event for a completed response with a specific id.
-fn ev_completed(id: &str) -> Value {
-    serde_json::json!({
-        "type": "response.completed",
-        "response": {
-            "id": id,
-            "usage": {"input_tokens":0,"input_tokens_details":null,"output_tokens":0,"output_tokens_details":null,"total_tokens":0}
-        }
-    })
-}
-
-/// Convenience: SSE event for a single assistant message output item.
-fn ev_assistant_message(id: &str, text: &str) -> Value {
-    serde_json::json!({
-        "type": "response.output_item.done",
-        "item": {
-            "type": "message",
-            "role": "assistant",
-            "id": id,
-            "content": [{"type": "output_text", "text": text}]
-        }
-    })
-}
-
-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)
-where
-    M: wiremock::Match + Send + Sync + 'static,
-{
-    Mock::given(method("POST"))
-        .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";
+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() {
@@ -99,7 +58,7 @@ 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![
@@ -122,19 +81,19 @@ async fn summarize_context_three_requests_and_instructions() {
         body.contains("\"text\":\"hello world\"")
             && !body.contains(&format!("\"text\":\"{SUMMARIZE_TRIGGER}\""))
     };
-    mount_sse_once(&server, first_matcher, sse1).await;
+    responses::mount_sse_once(&server, first_matcher, sse1).await;
 
     let second_matcher = |req: &wiremock::Request| {
         let body = std::str::from_utf8(&req.body).unwrap_or("");
         body.contains(&format!("\"text\":\"{SUMMARIZE_TRIGGER}\""))
     };
-    mount_sse_once(&server, second_matcher, sse2).await;
+    responses::mount_sse_once(&server, second_matcher, sse2).await;
 
     let third_matcher = |req: &wiremock::Request| {
         let body = std::str::from_utf8(&req.body).unwrap_or("");
         body.contains(&format!("\"text\":\"{THIRD_USER_MSG}\""))
     };
-    mount_sse_once(&server, third_matcher, sse3).await;
+    responses::mount_sse_once(&server, third_matcher, sse3).await;
 
     // Build config pointing to the mock server and spawn Codex.
     let model_provider = ModelProviderInfo {
@@ -144,6 +103,7 @@ 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 NewConversation {
         conversation: codex,
@@ -198,7 +158,7 @@ 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"
     );
 
@@ -209,14 +169,17 @@ 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.
@@ -232,24 +195,35 @@ 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"
     );
 
@@ -258,6 +232,7 @@ async fn summarize_context_three_requests_and_instructions() {
     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}",
@@ -296,3 +271,550 @@ async fn summarize_context_three_requests_and_instructions() {
         "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),
+        responses::ev_completed_with_tokens("r1", 70_000),
+    ]);
+
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", "SECOND_REPLY"),
+        responses::ev_completed_with_tokens("r2", 330_000),
+    ]);
+
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", AUTO_SUMMARY_TEXT),
+        responses::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(responses::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(responses::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(responses::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),
+        responses::ev_completed_with_tokens("r1", 70_000),
+    ]);
+
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", "SECOND_REPLY"),
+        responses::ev_completed_with_tokens("r2", 330_000),
+    ]);
+
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", AUTO_SUMMARY_TEXT),
+        responses::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(responses::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(responses::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(responses::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),
+        responses::ev_completed_with_tokens("r1", 500),
+    ]);
+
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", SUMMARY_TEXT),
+        responses::ev_completed_with_tokens("r2", 50),
+    ]);
+
+    let sse3 = sse(vec![
+        ev_assistant_message("m3", STILL_TOO_BIG_REPLY),
+        responses::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(responses::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(responses::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(responses::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),
+        responses::ev_completed_with_tokens("r1", 500),
+    ]);
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", FIRST_AUTO_SUMMARY),
+        responses::ev_completed_with_tokens("r2", 50),
+    ]);
+    let sse3 = sse(vec![
+        responses::ev_function_call(DUMMY_CALL_ID, DUMMY_FUNCTION_NAME, "{}"),
+        responses::ev_completed_with_tokens("r3", 150),
+    ]);
+    let sse4 = sse(vec![
+        ev_assistant_message("m4", SECOND_LARGE_REPLY),
+        responses::ev_completed_with_tokens("r4", 450),
+    ]);
+    let sse5 = sse(vec![
+        ev_assistant_message("m5", SECOND_AUTO_SUMMARY),
+        responses::ev_completed_with_tokens("r5", 60),
+    ]);
+    let sse6 = sse(vec![
+        ev_assistant_message("m6", FINAL_REPLY),
+        responses::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();
+
+    let mut auto_compact_lifecycle_events = Vec::new();
+    loop {
+        let event = codex.next_event().await.unwrap();
+        if event.id.starts_with("auto-compact-")
+            && matches!(
+                event.msg,
+                EventMsg::TaskStarted(_) | EventMsg::TaskComplete(_)
+            )
+        {
+            auto_compact_lifecycle_events.push(event);
+            continue;
+        }
+        if let EventMsg::TaskComplete(_) = &event.msg
+            && !event.id.starts_with("auto-compact-")
+        {
+            break;
+        }
+    }
+
+    assert!(
+        auto_compact_lifecycle_events.is_empty(),
+        "auto compact should not emit task lifecycle events"
+    );
+
+    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 +1,837 @@
+#![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 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::responses::ev_assistant_message;
+use core_test_support::responses::ev_completed;
+use core_test_support::responses::mount_sse_once;
+use core_test_support::responses::sse;
+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, 2).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()]);
+
+    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()[..compact_arr.len()],
+        &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
+}
+
+#[cfg(test)]
+async fn fork_conversation(
+    manager: &ConversationManager,
+    config: &Config,
+    path: std::path::PathBuf,
+    nth_user_message: usize,
+) -> Arc<CodexConversation> {
+    let NewConversation { conversation, .. } = manager
+        .fork_conversation(nth_user_message, config.clone(), path)
+        .await
+        .expect("fork conversation");
+    conversation
+}

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

@@ -39,7 +39,7 @@ async fn run_test_cmd(tmp: TempDir, cmd: Vec<&str>) -> Result<ExecToolCallOutput
 
     let policy = SandboxPolicy::new_read_only_policy();
 
-    process_exec_tool_call(params, sandbox_type, &policy, &None, None).await
+    process_exec_tool_call(params, sandbox_type, &policy, tmp.path(), &None, None).await
 }
 
 /// Command succeeds with exit code 0 normally

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;
@@ -46,9 +49,10 @@ async fn test_exec_stdout_stream_events_echo() {
         "printf 'hello-world\n'".to_string(),
     ];
 
+    let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
     let params = ExecParams {
         command: cmd,
-        cwd: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
+        cwd: cwd.clone(),
         timeout_ms: Some(5_000),
         env: HashMap::new(),
         with_escalated_permissions: None,
@@ -61,6 +65,7 @@ async fn test_exec_stdout_stream_events_echo() {
         params,
         SandboxType::None,
         &policy,
+        cwd.as_path(),
         &None,
         Some(stdout_stream),
     )
@@ -96,9 +101,10 @@ async fn test_exec_stderr_stream_events_echo() {
         "printf 'oops\n' 1>&2".to_string(),
     ];
 
+    let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
     let params = ExecParams {
         command: cmd,
-        cwd: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
+        cwd: cwd.clone(),
         timeout_ms: Some(5_000),
         env: HashMap::new(),
         with_escalated_permissions: None,
@@ -111,6 +117,7 @@ async fn test_exec_stderr_stream_events_echo() {
         params,
         SandboxType::None,
         &policy,
+        cwd.as_path(),
         &None,
         Some(stdout_stream),
     )
@@ -149,9 +156,10 @@ async fn test_aggregated_output_interleaves_in_order() {
         "printf 'O1\\n'; sleep 0.01; printf 'E1\\n' 1>&2; sleep 0.01; printf 'O2\\n'; sleep 0.01; printf 'E2\\n' 1>&2".to_string(),
     ];
 
+    let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
     let params = ExecParams {
         command: cmd,
-        cwd: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
+        cwd: cwd.clone(),
         timeout_ms: Some(5_000),
         env: HashMap::new(),
         with_escalated_permissions: None,
@@ -160,9 +168,16 @@ async fn test_aggregated_output_interleaves_in_order() {
 
     let policy = SandboxPolicy::new_read_only_policy();
 
-    let result = process_exec_tool_call(params, SandboxType::None, &policy, &None, None)
-        .await
-        .expect("process_exec_tool_call");
+    let result = process_exec_tool_call(
+        params,
+        SandboxType::None,
+        &policy,
+        cwd.as_path(),
+        &None,
+        None,
+    )
+    .await
+    .expect("process_exec_tool_call");
 
     assert_eq!(result.exit_code, 0);
     assert_eq!(result.stdout.text, "O1\nO2\n");
@@ -170,3 +185,45 @@ 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 cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
+    let params = ExecParams {
+        command: cmd,
+        cwd: cwd.clone(),
+        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,
+        cwd.as_path(),
+        &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

@@ -5,6 +5,8 @@ use codex_core::ModelProviderInfo;
 use codex_core::NewConversation;
 use codex_core::ResponseItem;
 use codex_core::built_in_model_providers;
+use codex_core::content_items_to_text;
+use codex_core::is_session_prefix_message;
 use codex_core::protocol::ConversationPathResponseEvent;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::InputItem;
@@ -104,13 +106,16 @@ async fn fork_conversation_twice_drops_to_first_message() {
         items
     };
 
-    // Compute expected prefixes after each fork by truncating base rollout at nth-from-last user input.
+    // Compute expected prefixes after each fork by truncating base rollout
+    // strictly before the nth user input (0-based).
     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"
+                && content_items_to_text(content)
+                    .is_some_and(|text| !is_session_prefix_message(&text))
             {
                 // 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.
@@ -126,11 +131,8 @@ async fn fork_conversation_twice_drops_to_first_message() {
     };
     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);
+    // After cutting at nth user input (n=1 → second user message), cut strictly before that input.
+    let cut1 = user_inputs.get(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.
@@ -161,12 +163,12 @@ async fn fork_conversation_twice_drops_to_first_message() {
         serde_json::to_value(&expected_after_first).unwrap()
     );
 
-    // Fork again with n=1 → drops the (new) last user message, leaving only the first.
+    // Fork again with n=0 → 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())
+        .fork_conversation(0, config_for_fork.clone(), fork1_path.clone())
         .await
         .expect("fork 2");
 

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

@@ -3,12 +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

@@ -36,7 +36,7 @@ async fn override_turn_context_does_not_persist_when_config_exists() {
             approval_policy: None,
             sandbox_policy: None,
             model: Some("o3".to_string()),
-            effort: Some(ReasoningEffort::High),
+            effort: Some(Some(ReasoningEffort::High)),
             summary: None,
         })
         .await
@@ -76,7 +76,7 @@ async fn override_turn_context_does_not_create_config_file() {
             approval_policy: None,
             sandbox_policy: None,
             model: Some("o3".to_string()),
-            effort: Some(ReasoningEffort::Medium),
+            effort: Some(Some(ReasoningEffort::Medium)),
             summary: None,
         })
         .await

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

@@ -12,6 +12,7 @@ use codex_core::protocol::Op;
 use codex_core::protocol::SandboxPolicy;
 use codex_core::protocol_config_types::ReasoningEffort;
 use codex_core::protocol_config_types::ReasoningSummary;
+use codex_core::shell::Shell;
 use codex_core::shell::default_user_shell;
 use core_test_support::load_default_config_for_test;
 use core_test_support::load_sse_fixture_with_id;
@@ -23,6 +24,30 @@ use wiremock::ResponseTemplate;
 use wiremock::matchers::method;
 use wiremock::matchers::path;
 
+fn text_user_input(text: String) -> serde_json::Value {
+    serde_json::json!({
+        "type": "message",
+        "role": "user",
+        "content": [ { "type": "input_text", "text": text } ]
+    })
+}
+
+fn default_env_context_str(cwd: &str, shell: &Shell) -> String {
+    format!(
+        r#"<environment_context>
+  <cwd>{}</cwd>
+  <approval_policy>on-request</approval_policy>
+  <sandbox_mode>read-only</sandbox_mode>
+  <network_access>restricted</network_access>
+{}</environment_context>"#,
+        cwd,
+        match shell.name() {
+            Some(name) => format!("  <shell>{name}</shell>\n"),
+            None => String::new(),
+        }
+    )
+}
+
 /// Build minimal SSE stream with completed marker using the JSON fixture.
 fn sse_completed(id: &str) -> String {
     load_sse_fixture_with_id("tests/fixtures/completed_template.json", id)
@@ -387,7 +412,7 @@ 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
@@ -519,7 +544,7 @@ 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
@@ -546,12 +571,262 @@ async fn per_turn_overrides_keep_cached_prefix_and_key_constant() {
         "role": "user",
         "content": [ { "type": "input_text", "text": "hello 2" } ]
     });
+    let expected_env_text_2 = format!(
+        r#"<environment_context>
+  <cwd>{}</cwd>
+  <approval_policy>never</approval_policy>
+  <sandbox_mode>workspace-write</sandbox_mode>
+  <network_access>enabled</network_access>
+  <writable_roots>
+    <root>{}</root>
+  </writable_roots>
+</environment_context>"#,
+        new_cwd.path().to_string_lossy(),
+        writable.path().to_string_lossy(),
+    );
+    let expected_env_msg_2 = serde_json::json!({
+        "type": "message",
+        "role": "user",
+        "content": [ { "type": "input_text", "text": expected_env_text_2 } ]
+    });
     let expected_body2 = serde_json::json!(
         [
             body1["input"].as_array().unwrap().as_slice(),
-            [expected_user_message_2].as_slice(),
+            [expected_env_msg_2, expected_user_message_2].as_slice(),
         ]
         .concat()
     );
     assert_eq!(body2["input"], expected_body2);
 }
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn send_user_turn_with_no_changes_does_not_send_environment_context() {
+    use pretty_assertions::assert_eq;
+
+    let server = MockServer::start().await;
+
+    let sse = sse_completed("resp");
+    let template = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse, "text/event-stream");
+
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(template)
+        .expect(2)
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let cwd = TempDir::new().unwrap();
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.cwd = cwd.path().to_path_buf();
+    config.model_provider = model_provider;
+    config.user_instructions = Some("be consistent and helpful".to_string());
+
+    let default_cwd = config.cwd.clone();
+    let default_approval_policy = config.approval_policy;
+    let default_sandbox_policy = config.sandbox_policy.clone();
+    let default_model = config.model.clone();
+    let default_effort = config.model_reasoning_effort;
+    let default_summary = config.model_reasoning_summary;
+
+    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::UserTurn {
+            items: vec![InputItem::Text {
+                text: "hello 1".into(),
+            }],
+            cwd: default_cwd.clone(),
+            approval_policy: default_approval_policy,
+            sandbox_policy: default_sandbox_policy.clone(),
+            model: default_model.clone(),
+            effort: default_effort,
+            summary: default_summary,
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    codex
+        .submit(Op::UserTurn {
+            items: vec![InputItem::Text {
+                text: "hello 2".into(),
+            }],
+            cwd: default_cwd.clone(),
+            approval_policy: default_approval_policy,
+            sandbox_policy: default_sandbox_policy.clone(),
+            model: default_model.clone(),
+            effort: default_effort,
+            summary: default_summary,
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let requests = server.received_requests().await.unwrap();
+    assert_eq!(requests.len(), 2, "expected two POST requests");
+
+    let body1 = requests[0].body_json::<serde_json::Value>().unwrap();
+    let body2 = requests[1].body_json::<serde_json::Value>().unwrap();
+
+    let shell = default_user_shell().await;
+    let expected_ui_text =
+        "<user_instructions>\n\nbe consistent and helpful\n\n</user_instructions>";
+    let expected_ui_msg = text_user_input(expected_ui_text.to_string());
+
+    let expected_env_msg_1 = text_user_input(default_env_context_str(
+        &cwd.path().to_string_lossy(),
+        &shell,
+    ));
+    let expected_user_message_1 = text_user_input("hello 1".to_string());
+
+    let expected_input_1 = serde_json::Value::Array(vec![
+        expected_ui_msg.clone(),
+        expected_env_msg_1.clone(),
+        expected_user_message_1.clone(),
+    ]);
+    assert_eq!(body1["input"], expected_input_1);
+
+    let expected_user_message_2 = text_user_input("hello 2".to_string());
+    let expected_input_2 = serde_json::Value::Array(vec![
+        expected_ui_msg,
+        expected_env_msg_1,
+        expected_user_message_1,
+        expected_user_message_2,
+    ]);
+    assert_eq!(body2["input"], expected_input_2);
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn send_user_turn_with_changes_sends_environment_context() {
+    use pretty_assertions::assert_eq;
+
+    let server = MockServer::start().await;
+
+    let sse = sse_completed("resp");
+    let template = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse, "text/event-stream");
+
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(template)
+        .expect(2)
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let cwd = TempDir::new().unwrap();
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.cwd = cwd.path().to_path_buf();
+    config.model_provider = model_provider;
+    config.user_instructions = Some("be consistent and helpful".to_string());
+
+    let default_cwd = config.cwd.clone();
+    let default_approval_policy = config.approval_policy;
+    let default_sandbox_policy = config.sandbox_policy.clone();
+    let default_model = config.model.clone();
+    let default_effort = config.model_reasoning_effort;
+    let default_summary = config.model_reasoning_summary;
+
+    let conversation_manager =
+        ConversationManager::with_auth(CodexAuth::from_api_key("Test API Key"));
+    let codex = conversation_manager
+        .new_conversation(config.clone())
+        .await
+        .expect("create new conversation")
+        .conversation;
+
+    codex
+        .submit(Op::UserTurn {
+            items: vec![InputItem::Text {
+                text: "hello 1".into(),
+            }],
+            cwd: default_cwd.clone(),
+            approval_policy: default_approval_policy,
+            sandbox_policy: default_sandbox_policy.clone(),
+            model: default_model,
+            effort: default_effort,
+            summary: default_summary,
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    codex
+        .submit(Op::UserTurn {
+            items: vec![InputItem::Text {
+                text: "hello 2".into(),
+            }],
+            cwd: default_cwd.clone(),
+            approval_policy: AskForApproval::Never,
+            sandbox_policy: SandboxPolicy::DangerFullAccess,
+            model: "o3".to_string(),
+            effort: Some(ReasoningEffort::High),
+            summary: ReasoningSummary::Detailed,
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let requests = server.received_requests().await.unwrap();
+    assert_eq!(requests.len(), 2, "expected two POST requests");
+
+    let body1 = requests[0].body_json::<serde_json::Value>().unwrap();
+    let body2 = requests[1].body_json::<serde_json::Value>().unwrap();
+
+    let shell = default_user_shell().await;
+    let expected_ui_text =
+        "<user_instructions>\n\nbe consistent and helpful\n\n</user_instructions>";
+    let expected_ui_msg = serde_json::json!({
+        "type": "message",
+        "role": "user",
+        "content": [ { "type": "input_text", "text": expected_ui_text } ]
+    });
+    let expected_env_text_1 = default_env_context_str(&default_cwd.to_string_lossy(), &shell);
+    let expected_env_msg_1 = text_user_input(expected_env_text_1);
+    let expected_user_message_1 = text_user_input("hello 1".to_string());
+    let expected_input_1 = serde_json::Value::Array(vec![
+        expected_ui_msg.clone(),
+        expected_env_msg_1.clone(),
+        expected_user_message_1.clone(),
+    ]);
+    assert_eq!(body1["input"], expected_input_1);
+
+    let expected_env_msg_2 = text_user_input(format!(
+        r#"<environment_context>
+  <cwd>{}</cwd>
+  <approval_policy>never</approval_policy>
+  <sandbox_mode>danger-full-access</sandbox_mode>
+  <network_access>enabled</network_access>
+</environment_context>"#,
+        default_cwd.to_string_lossy()
+    ));
+    let expected_user_message_2 = text_user_input("hello 2".to_string());
+    let expected_input_2 = serde_json::Value::Array(vec![
+        expected_ui_msg,
+        expected_env_msg_1,
+        expected_user_message_1,
+        expected_env_msg_2,
+        expected_user_message_2,
+    ]);
+    assert_eq!(body2["input"], expected_input_2);
+}

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

@@ -0,0 +1,701 @@
+use codex_core::CodexAuth;
+use codex_core::CodexConversation;
+use codex_core::ContentItem;
+use codex_core::ConversationManager;
+use codex_core::ModelProviderInfo;
+use codex_core::REVIEW_PROMPT;
+use codex_core::ResponseItem;
+use codex_core::built_in_model_providers;
+use codex_core::config::Config;
+use codex_core::protocol::ConversationPathResponseEvent;
+use codex_core::protocol::ENVIRONMENT_CONTEXT_OPEN_TAG;
+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::protocol::RolloutItem;
+use codex_core::protocol::RolloutLine;
+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;
+
+    // Also verify that a user message with the header and a formatted finding
+    // was recorded back in the parent session's rollout.
+    codex.submit(Op::GetPath).await.unwrap();
+    let history_event =
+        wait_for_event(&codex, |ev| matches!(ev, EventMsg::ConversationPath(_))).await;
+    let path = match history_event {
+        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path,
+        other => panic!("expected ConversationPath event, got {other:?}"),
+    };
+    let text = std::fs::read_to_string(&path).expect("read rollout file");
+
+    let mut saw_header = false;
+    let mut saw_finding_line = false;
+    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");
+        if let RolloutItem::ResponseItem(ResponseItem::Message { role, content, .. }) = rl.item
+            && role == "user"
+        {
+            for c in content {
+                if let ContentItem::InputText { text } = c {
+                    if text.contains("full review output from reviewer model") {
+                        saw_header = true;
+                    }
+                    if text.contains("- Prefer Stylize helpers — /tmp/file.rs:10-20") {
+                        saw_finding_line = true;
+                    }
+                }
+            }
+        }
+    }
+    assert!(saw_header, "user header missing from rollout");
+    assert!(
+        saw_finding_line,
+        "formatted finding line missing from rollout"
+    );
+
+    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 the environment context followed by the review prompt.
+    let request = &server.received_requests().await.unwrap()[0];
+    let body = request.body_json::<serde_json::Value>().unwrap();
+    let input = body["input"].as_array().expect("input array");
+    assert_eq!(
+        input.len(),
+        2,
+        "expected environment context and review prompt"
+    );
+
+    let env_msg = &input[0];
+    assert_eq!(env_msg["type"].as_str().unwrap(), "message");
+    assert_eq!(env_msg["role"].as_str().unwrap(), "user");
+    let env_text = env_msg["content"][0]["text"].as_str().expect("env text");
+    assert!(
+        env_text.starts_with(ENVIRONMENT_CONTEXT_OPEN_TAG),
+        "environment context must be the first item"
+    );
+    assert!(
+        env_text.contains("<cwd>"),
+        "environment context should include cwd"
+    );
+
+    let review_msg = &input[1];
+    assert_eq!(review_msg["type"].as_str().unwrap(), "message");
+    assert_eq!(review_msg["role"].as_str().unwrap(), "user");
+    assert_eq!(
+        review_msg["content"][0]["text"].as_str().unwrap(),
+        format!("{REVIEW_PROMPT}\n\n---\n\nNow, here's your task: Please review only this",)
+    );
+
+    // Also verify that a user interruption note was recorded in the rollout.
+    codex.submit(Op::GetPath).await.unwrap();
+    let history_event =
+        wait_for_event(&codex, |ev| matches!(ev, EventMsg::ConversationPath(_))).await;
+    let path = match history_event {
+        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path,
+        other => panic!("expected ConversationPath event, got {other:?}"),
+    };
+    let text = std::fs::read_to_string(&path).expect("read rollout file");
+    let mut saw_interruption_message = false;
+    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");
+        if let RolloutItem::ResponseItem(ResponseItem::Message { role, content, .. }) = rl.item
+            && role == "user"
+        {
+            for c in content {
+                if let ContentItem::InputText { text } = c
+                    && text.contains("User initiated a review task, but was interrupted.")
+                {
+                    saw_interruption_message = true;
+                    break;
+                }
+            }
+        }
+        if saw_interruption_message {
+            break;
+        }
+    }
+    assert!(
+        saw_interruption_message,
+        "expected user interruption message in rollout"
+    );
+
+    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

@@ -171,6 +171,8 @@ async fn python_getpwuid_works_under_seatbelt() {
 
     // ReadOnly is sufficient here since we are only exercising user lookup.
     let policy = SandboxPolicy::ReadOnly;
+    let command_cwd = std::env::current_dir().expect("getcwd");
+    let sandbox_cwd = command_cwd.clone();
 
     let mut child = spawn_command_under_seatbelt(
         vec![
@@ -179,8 +181,9 @@ async fn python_getpwuid_works_under_seatbelt() {
             // Print the passwd struct; success implies lookup worked.
             "import pwd, os; print(pwd.getpwuid(os.getuid()))".to_string(),
         ],
+        command_cwd,
         &policy,
-        std::env::current_dir().expect("should be able to get current dir"),
+        sandbox_cwd.as_path(),
         StdioPolicy::RedirectForShellTool,
         HashMap::new(),
     )
@@ -216,13 +219,16 @@ fn create_test_scenario(tmp: &TempDir) -> TestScenario {
 /// Note that `path` must be absolute.
 async fn touch(path: &Path, policy: &SandboxPolicy) -> bool {
     assert!(path.is_absolute(), "Path must be absolute: {path:?}");
+    let command_cwd = std::env::current_dir().expect("getcwd");
+    let sandbox_cwd = command_cwd.clone();
     let mut child = spawn_command_under_seatbelt(
         vec![
             "/usr/bin/touch".to_string(),
             path.to_string_lossy().to_string(),
         ],
+        command_cwd,
         policy,
-        std::env::current_dir().expect("should be able to get current dir"),
+        sandbox_cwd.as_path(),
         StdioPolicy::RedirectForShellTool,
         HashMap::new(),
     )

codex-rs/docs/codex_mcp_interface.md

@@ -0,0 +1,123 @@
+# Codex MCP Interface [experimental]
+
+This document describes Codex’s experimental MCP interface: a JSON‑RPC API that runs over the Model Context Protocol (MCP) transport to control a local Codex engine.
+
+- Status: experimental and subject to change without notice
+- Server binary: `codex mcp` (or `codex-mcp-server`)
+- Transport: standard MCP over stdio (JSON‑RPC 2.0, line‑delimited)
+
+## Overview
+
+Codex exposes a small set of MCP‑compatible methods to create and manage conversations, send user input, receive live events, and handle approval prompts. The types are defined in `protocol/src/mcp_protocol.rs` and re‑used by the MCP server implementation in `mcp-server/`.
+
+At a glance:
+
+- Conversations
+  - `newConversation` → start a Codex session
+  - `sendUserMessage` / `sendUserTurn` → send user input into a conversation
+  - `interruptConversation` → stop the current turn
+  - `listConversations`, `resumeConversation`, `archiveConversation`
+- Configuration and info
+  - `getUserSavedConfig`, `setDefaultModel`, `getUserAgent`, `userInfo`
+- Auth
+  - `loginApiKey`, `loginChatGpt`, `cancelLoginChatGpt`, `logoutChatGpt`, `getAuthStatus`
+- Utilities
+  - `gitDiffToRemote`, `execOneOffCommand`
+- Approvals (server → client requests)
+  - `applyPatchApproval`, `execCommandApproval`
+- Notifications (server → client)
+  - `loginChatGptComplete`, `authStatusChange`
+  - `codex/event` stream with agent events
+
+See code for full type definitions and exact shapes: `protocol/src/mcp_protocol.rs`.
+
+## Starting the server
+
+Run Codex as an MCP server and connect an MCP client:
+
+```bash
+codex mcp | your_mcp_client
+```
+
+For a simple inspection UI, you can also try:
+
+```bash
+npx @modelcontextprotocol/inspector codex mcp
+```
+
+## Conversations
+
+Start a new session with optional overrides:
+
+Request `newConversation` params (subset):
+
+- `model`: string model id (e.g. "o3", "gpt-5")
+- `profile`: optional named profile
+- `cwd`: optional working directory
+- `approvalPolicy`: `untrusted` | `on-request` | `on-failure` | `never`
+- `sandbox`: `read-only` | `workspace-write` | `danger-full-access`
+- `config`: map of additional config overrides
+- `baseInstructions`: optional instruction override
+- `includePlanTool` / `includeApplyPatchTool`: booleans
+
+Response: `{ conversationId, model, reasoningEffort?, rolloutPath }`
+
+Send input to the active turn:
+
+- `sendUserMessage` → enqueue items to the conversation
+- `sendUserTurn` → structured turn with explicit `cwd`, `approvalPolicy`, `sandboxPolicy`, `model`, optional `effort`, and `summary`
+
+Interrupt a running turn: `interruptConversation`.
+
+List/resume/archive: `listConversations`, `resumeConversation`, `archiveConversation`.
+
+## Event stream
+
+While a conversation runs, the server sends notifications:
+
+- `codex/event` with the serialized Codex event payload. The shape matches `core/src/protocol.rs`’s `Event` and `EventMsg` types. Some notifications include a `_meta.requestId` to correlate with the originating request.
+- Auth notifications via method names `loginChatGptComplete` and `authStatusChange`.
+
+Clients should render events and, when present, surface approval requests (see next section).
+
+## Approvals (server → client)
+
+When Codex needs approval to apply changes or run commands, the server issues JSON‑RPC requests to the client:
+
+- `applyPatchApproval { conversationId, callId, fileChanges, reason?, grantRoot? }`
+- `execCommandApproval { conversationId, callId, command, cwd, reason? }`
+
+The client must reply with `{ decision: "allow" | "deny" }` for each request.
+
+## Auth helpers
+
+For ChatGPT or API‑key based auth flows, the server exposes helpers:
+
+- `loginApiKey { apiKey }`
+- `loginChatGpt` → returns `{ loginId, authUrl }`; browser completes flow; then `loginChatGptComplete` notification follows
+- `cancelLoginChatGpt { loginId }`, `logoutChatGpt`, `getAuthStatus { includeToken?, refreshToken? }`
+
+## Example: start and send a message
+
+```json
+{ "jsonrpc": "2.0", "id": 1, "method": "newConversation", "params": { "model": "gpt-5", "approvalPolicy": "on-request" } }
+```
+
+Server responds:
+
+```json
+{ "jsonrpc": "2.0", "id": 1, "result": { "conversationId": "c7b0…", "model": "gpt-5", "rolloutPath": "/path/to/rollout.jsonl" } }
+```
+
+Then send input:
+
+```json
+{ "jsonrpc": "2.0", "id": 2, "method": "sendUserMessage", "params": { "conversationId": "c7b0…", "items": [{ "type": "text", "text": "Hello Codex" }] } }
+```
+
+While processing, the server emits `codex/event` notifications containing agent output, approvals, and status updates.
+
+## Compatibility and stability
+
+This interface is experimental. Method names, fields, and event shapes may evolve. For the authoritative schema, consult `protocol/src/mcp_protocol.rs` and the corresponding server wiring in `mcp-server/`.
+

codex-rs/exec/Cargo.toml

@@ -38,7 +38,8 @@ tokio = { version = "1", features = [
     "signal",
 ] }
 tracing = { version = "0.1.41", features = ["log"] }
-tracing-subscriber = { version = "0.3.19", features = ["env-filter"] }
+tracing-subscriber = { version = "0.3.20", features = ["env-filter"] }
+opentelemetry-appender-tracing = "0.30.1"
 
 [dev-dependencies]
 assert_cmd = "2"
@@ -46,4 +47,6 @@ core_test_support = { path = "../core/tests/common" }
 libc = "0.2"
 predicates = "3"
 tempfile = "3.13.0"
+uuid = "1"
+walkdir = "2"
 wiremock = "0.6"

codex-rs/exec/src/cli.rs

@@ -6,6 +6,10 @@ use std::path::PathBuf;
 #[derive(Parser, Debug)]
 #[command(version)]
 pub struct Cli {
+    /// Action to perform. If omitted, runs a new non-interactive session.
+    #[command(subcommand)]
+    pub command: Option<Command>,
+
     /// Optional image(s) to attach to the initial prompt.
     #[arg(long = "image", short = 'i', value_name = "FILE", value_delimiter = ',', num_args = 1..)]
     pub images: Vec<PathBuf>,
@@ -69,6 +73,28 @@ pub struct Cli {
     pub prompt: Option<String>,
 }
 
+#[derive(Debug, clap::Subcommand)]
+pub enum Command {
+    /// Resume a previous session by id or pick the most recent with --last.
+    Resume(ResumeArgs),
+}
+
+#[derive(Parser, Debug)]
+pub struct ResumeArgs {
+    /// 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")]
+    pub session_id: Option<String>,
+
+    /// Resume the most recent recorded session (newest) without specifying an id.
+    #[arg(long = "last", default_value_t = false, conflicts_with = "session_id")]
+    pub last: bool,
+
+    /// Prompt to send after resuming the session. If `-` is used, read from stdin.
+    #[arg(value_name = "PROMPT")]
+    pub prompt: Option<String>,
+}
+
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, ValueEnum)]
 #[value(rename_all = "kebab-case")]
 pub enum Color {

codex-rs/exec/src/event_processor_with_human_output.rs

@@ -558,10 +558,15 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 TurnAbortReason::Replaced => {
                     ts_println!(self, "task aborted: replaced by a new task");
                 }
+                TurnAbortReason::ReviewEnded => {
+                    ts_println!(self, "task aborted: review ended");
+                }
             },
             EventMsg::ShutdownComplete => return CodexStatus::Shutdown,
             EventMsg::ConversationPath(_) => {}
             EventMsg::UserMessage(_) => {}
+            EventMsg::EnteredReviewMode(_) => {}
+            EventMsg::ExitedReviewMode(_) => {}
         }
         CodexStatus::Running
     }

codex-rs/exec/src/lib.rs

@@ -3,10 +3,6 @@ mod event_processor;
 mod event_processor_with_human_output;
 mod event_processor_with_json_output;
 
-use std::io::IsTerminal;
-use std::io::Read;
-use std::path::PathBuf;
-
 pub use cli::Cli;
 use codex_core::AuthManager;
 use codex_core::BUILT_IN_OSS_MODEL_PROVIDER_ID;
@@ -25,16 +21,24 @@ use codex_ollama::DEFAULT_OSS_MODEL;
 use codex_protocol::config_types::SandboxMode;
 use event_processor_with_human_output::EventProcessorWithHumanOutput;
 use event_processor_with_json_output::EventProcessorWithJsonOutput;
+use opentelemetry_appender_tracing::layer::OpenTelemetryTracingBridge;
+use std::io::IsTerminal;
+use std::io::Read;
+use std::path::PathBuf;
 use tracing::debug;
 use tracing::error;
 use tracing::info;
 use tracing_subscriber::EnvFilter;
+use tracing_subscriber::prelude::*;
 
+use crate::cli::Command as ExecCommand;
 use crate::event_processor::CodexStatus;
 use crate::event_processor::EventProcessor;
+use codex_core::find_conversation_path_by_id_str;
 
 pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()> {
     let Cli {
+        command,
         images,
         model: model_cli_arg,
         oss,
@@ -51,8 +55,15 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         config_overrides,
     } = cli;
 
-    // Determine the prompt based on CLI arg and/or stdin.
-    let prompt = match prompt {
+    // Determine the prompt source (parent or subcommand) and read from stdin if needed.
+    let prompt_arg = match &command {
+        // Allow prompt before the subcommand by falling back to the parent-level prompt
+        // when the Resume subcommand did not provide its own prompt.
+        Some(ExecCommand::Resume(args)) => args.prompt.clone().or(prompt),
+        None => prompt,
+    };
+
+    let prompt = match prompt_arg {
         Some(p) if p != "-" => p,
         // Either `-` was passed or no positional arg.
         maybe_dash => {
@@ -95,19 +106,18 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         ),
     };
 
-    // TODO(mbolin): Take a more thoughtful approach to logging.
+    // Build fmt layer (existing logging) to compose with OTEL layer.
     let default_level = "error";
-    let _ = tracing_subscriber::fmt()
-        // Fallback to the `default_level` log filter if the environment
-        // variable is not set _or_ contains an invalid value
-        .with_env_filter(
-            EnvFilter::try_from_default_env()
-                .or_else(|_| EnvFilter::try_new(default_level))
-                .unwrap_or_else(|_| EnvFilter::new(default_level)),
-        )
+
+    // Build env_filter separately and attach via with_filter.
+    let env_filter = EnvFilter::try_from_default_env()
+        .or_else(|_| EnvFilter::try_new(default_level))
+        .unwrap_or_else(|_| EnvFilter::new(default_level));
+
+    let fmt_layer = tracing_subscriber::fmt::layer()
         .with_ansi(stderr_with_ansi)
         .with_writer(std::io::stderr)
-        .try_init();
+        .with_filter(env_filter);
 
     let sandbox_mode = if full_auto {
         Some(SandboxMode::WorkspaceWrite)
@@ -137,6 +147,7 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
     // Load configuration and determine approval policy
     let overrides = ConfigOverrides {
         model,
+        review_model: None,
         config_profile,
         // This CLI is intended to be headless and has no affordances for asking
         // the user for approval.
@@ -162,6 +173,31 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
     };
 
     let config = Config::load_with_cli_overrides(cli_kv_overrides, overrides)?;
+
+    let otel = codex_core::otel_init::build_provider(&config, env!("CARGO_PKG_VERSION"));
+
+    #[allow(clippy::print_stderr)]
+    let otel = match otel {
+        Ok(otel) => otel,
+        Err(e) => {
+            eprintln!("Could not create otel exporter: {e}");
+            std::process::exit(1);
+        }
+    };
+
+    if let Some(provider) = otel {
+        let otel_layer = OpenTelemetryTracingBridge::new(&provider.logger).with_filter(
+            tracing_subscriber::filter::filter_fn(codex_core::otel_init::codex_export_filter),
+        );
+
+        let _ = tracing_subscriber::registry()
+            .with(fmt_layer)
+            .with(otel_layer)
+            .try_init();
+    } else {
+        let _ = tracing_subscriber::registry().with(fmt_layer).try_init();
+    }
+
     let mut event_processor: Box<dyn EventProcessor> = if json_mode {
         Box::new(EventProcessorWithJsonOutput::new(last_message_file.clone()))
     } else {
@@ -189,11 +225,29 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
 
     let conversation_manager =
         ConversationManager::new(AuthManager::shared(config.codex_home.clone()));
+
+    // Handle resume subcommand by resolving a rollout path and using explicit resume API.
     let NewConversation {
         conversation_id: _,
         conversation,
         session_configured,
-    } = conversation_manager.new_conversation(config).await?;
+    } = if let Some(ExecCommand::Resume(args)) = command {
+        let resume_path = resolve_resume_path(&config, &args).await?;
+
+        if let Some(path) = resume_path {
+            conversation_manager
+                .resume_conversation_from_rollout(
+                    config.clone(),
+                    path,
+                    AuthManager::shared(config.codex_home.clone()),
+                )
+                .await?
+        } else {
+            conversation_manager.new_conversation(config).await?
+        }
+    } else {
+        conversation_manager.new_conversation(config).await?
+    };
     info!("Codex initialized with event: {session_configured:?}");
 
     let (tx, mut rx) = tokio::sync::mpsc::unbounded_channel::<Event>();
@@ -278,3 +332,23 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
 
     Ok(())
 }
+
+async fn resolve_resume_path(
+    config: &Config,
+    args: &crate::cli::ResumeArgs,
+) -> anyhow::Result<Option<PathBuf>> {
+    if args.last {
+        match codex_core::RolloutRecorder::list_conversations(&config.codex_home, 1, None).await {
+            Ok(page) => Ok(page.items.first().map(|it| it.path.clone())),
+            Err(e) => {
+                error!("Error listing conversations: {e}");
+                Ok(None)
+            }
+        }
+    } else if let Some(id_str) = args.session_id.as_deref() {
+        let path = find_conversation_path_by_id_str(&config.codex_home, id_str).await?;
+        Ok(path)
+    } else {
+        Ok(None)
+    }
+}

codex-rs/exec/tests/fixtures/cli_responses_fixture.sse

@@ -0,0 +1,10 @@
+event: response.created
+data: {"type":"response.created","response":{"id":"resp1"}}
+
+event: response.output_item.done
+data: {"type":"response.output_item.done","item":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"fixture hello"}]}}
+
+event: response.completed
+data: {"type":"response.completed","response":{"id":"resp1","output":[]}}
+
+

codex-rs/exec/tests/fixtures/sse_apply_patch_add.json

@@ -1,25 +0,0 @@
-[
-  {
-    "type": "response.output_item.done",
-    "item": {
-      "type": "custom_tool_call",
-      "name": "apply_patch",
-      "input": "*** Begin Patch\n*** Add File: test.md\n+Hello world\n*** End Patch",
-      "call_id": "__ID__"
-    }
-  },
-  {
-    "type": "response.completed",
-    "response": {
-      "id": "__ID__",
-      "usage": {
-        "input_tokens": 0,
-        "input_tokens_details": null,
-        "output_tokens": 0,
-        "output_tokens_details": null,
-        "total_tokens": 0
-      },
-      "output": []
-    }
-  }
-]

codex-rs/exec/tests/fixtures/sse_apply_patch_freeform_add.json

@@ -1,25 +0,0 @@
-[
-  {
-    "type": "response.output_item.done",
-    "item": {
-      "type": "custom_tool_call",
-      "name": "apply_patch",
-      "input": "*** Begin Patch\n*** Add File: app.py\n+class BaseClass:\n+  def method():\n+    return False\n*** End Patch",
-      "call_id": "__ID__"
-    }
-  },
-  {
-    "type": "response.completed",
-    "response": {
-      "id": "__ID__",
-      "usage": {
-        "input_tokens": 0,
-        "input_tokens_details": null,
-        "output_tokens": 0,
-        "output_tokens_details": null,
-        "total_tokens": 0
-      },
-      "output": []
-    }
-  }
-]

codex-rs/exec/tests/fixtures/sse_apply_patch_freeform_update.json

@@ -1,25 +0,0 @@
-[
-  {
-    "type": "response.output_item.done",
-    "item": {
-      "type": "custom_tool_call",
-      "name": "apply_patch",
-      "input": "*** Begin Patch\n*** Update File: app.py\n@@  def method():\n-    return False\n+\n+    return True\n*** End Patch",
-      "call_id": "__ID__"
-    }
-  },
-  {
-    "type": "response.completed",
-    "response": {
-      "id": "__ID__",
-      "usage": {
-        "input_tokens": 0,
-        "input_tokens_details": null,
-        "output_tokens": 0,
-        "output_tokens_details": null,
-        "total_tokens": 0
-      },
-      "output": []
-    }
-  }
-]

codex-rs/exec/tests/fixtures/sse_apply_patch_update.json

@@ -1,25 +0,0 @@
-[
-  {
-    "type": "response.output_item.done",
-    "item": {
-      "type": "function_call",
-      "name": "apply_patch",
-      "arguments": "{\n  \"input\": \"*** Begin Patch\\n*** Update File: test.md\\n@@\\n-Hello world\\n+Final text\\n*** End Patch\"\n}",
-      "call_id": "__ID__"
-    }
-  },
-  {
-    "type": "response.completed",
-    "response": {
-      "id": "__ID__",
-      "usage": {
-        "input_tokens": 0,
-        "input_tokens_details": null,
-        "output_tokens": 0,
-        "output_tokens_details": null,
-        "total_tokens": 0
-      },
-      "output": []
-    }
-  }
-]

codex-rs/exec/tests/fixtures/sse_response_completed.json

@@ -1,16 +0,0 @@
-[
-  {
-    "type": "response.completed",
-    "response": {
-      "id": "__ID__",
-      "usage": {
-        "input_tokens": 0,
-        "input_tokens_details": null,
-        "output_tokens": 0,
-        "output_tokens_details": null,
-        "total_tokens": 0
-      },
-      "output": []
-    }
-  }
-]

codex-rs/exec/tests/suite/apply_patch.rs

@@ -1,8 +1,12 @@
-#![allow(clippy::expect_used, clippy::unwrap_used)]
+#![allow(clippy::expect_used, clippy::unwrap_used, unused_imports)]
 
 use anyhow::Context;
 use assert_cmd::prelude::*;
 use codex_core::CODEX_APPLY_PATCH_ARG1;
+use core_test_support::responses::ev_apply_patch_custom_tool_call;
+use core_test_support::responses::ev_apply_patch_function_call;
+use core_test_support::responses::ev_completed;
+use core_test_support::responses::sse;
 use std::fs;
 use std::process::Command;
 use tempfile::tempdir;
@@ -55,15 +59,28 @@ async fn test_apply_patch_tool() -> anyhow::Result<()> {
 
     let tmp_cwd = tempdir().expect("failed to create temp dir");
     let tmp_path = tmp_cwd.path().to_path_buf();
-    run_e2e_exec_test(
-        tmp_cwd.path(),
-        vec![
-            include_str!("../fixtures/sse_apply_patch_add.json").to_string(),
-            include_str!("../fixtures/sse_apply_patch_update.json").to_string(),
-            include_str!("../fixtures/sse_response_completed.json").to_string(),
-        ],
-    )
-    .await;
+    let add_patch = r#"*** Begin Patch
+*** Add File: test.md
++Hello world
+*** End Patch"#;
+    let update_patch = r#"*** Begin Patch
+*** Update File: test.md
+@@
+-Hello world
++Final text
+*** End Patch"#;
+    let response_streams = vec![
+        sse(vec![
+            ev_apply_patch_custom_tool_call("request_0", add_patch),
+            ev_completed("request_0"),
+        ]),
+        sse(vec![
+            ev_apply_patch_function_call("request_1", update_patch),
+            ev_completed("request_1"),
+        ]),
+        sse(vec![ev_completed("request_2")]),
+    ];
+    run_e2e_exec_test(tmp_cwd.path(), response_streams).await;
 
     let final_path = tmp_path.join("test.md");
     let contents = std::fs::read_to_string(&final_path)
@@ -86,15 +103,31 @@ async fn test_apply_patch_freeform_tool() -> anyhow::Result<()> {
     }
 
     let tmp_cwd = tempdir().expect("failed to create temp dir");
-    run_e2e_exec_test(
-        tmp_cwd.path(),
-        vec![
-            include_str!("../fixtures/sse_apply_patch_freeform_add.json").to_string(),
-            include_str!("../fixtures/sse_apply_patch_freeform_update.json").to_string(),
-            include_str!("../fixtures/sse_response_completed.json").to_string(),
-        ],
-    )
-    .await;
+    let freeform_add_patch = r#"*** Begin Patch
+*** Add File: app.py
++class BaseClass:
++  def method():
++    return False
+*** End Patch"#;
+    let freeform_update_patch = r#"*** Begin Patch
+*** Update File: app.py
+@@  def method():
+-    return False
++
++    return True
+*** End Patch"#;
+    let response_streams = vec![
+        sse(vec![
+            ev_apply_patch_custom_tool_call("request_0", freeform_add_patch),
+            ev_completed("request_0"),
+        ]),
+        sse(vec![
+            ev_apply_patch_custom_tool_call("request_1", freeform_update_patch),
+            ev_completed("request_1"),
+        ]),
+        sse(vec![ev_completed("request_2")]),
+    ];
+    run_e2e_exec_test(tmp_cwd.path(), response_streams).await;
 
     // Verify final file contents
     let final_path = tmp_cwd.path().join("app.py");

codex-rs/exec/tests/suite/common.rs

@@ -4,7 +4,6 @@
 
 use anyhow::Context;
 use assert_cmd::prelude::*;
-use core_test_support::load_sse_fixture_with_id_from_str;
 use std::path::Path;
 use std::process::Command;
 use std::sync::atomic::AtomicUsize;
@@ -27,10 +26,7 @@ impl Respond for SeqResponder {
         match self.responses.get(call_num) {
             Some(body) => wiremock::ResponseTemplate::new(200)
                 .insert_header("content-type", "text/event-stream")
-                .set_body_raw(
-                    load_sse_fixture_with_id_from_str(body, &format!("request_{call_num}")),
-                    "text/event-stream",
-                ),
+                .set_body_string(body.clone()),
             None => panic!("no response for {call_num}"),
         }
     }

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

@@ -1,4 +1,5 @@
 // Aggregates all former standalone integration tests as modules.
 mod apply_patch;
 mod common;
+mod resume;
 mod sandbox;

codex-rs/exec/tests/suite/resume.rs

@@ -0,0 +1,267 @@
+#![allow(clippy::unwrap_used, clippy::expect_used)]
+use anyhow::Context;
+use assert_cmd::prelude::*;
+use serde_json::Value;
+use std::process::Command;
+use tempfile::TempDir;
+use uuid::Uuid;
+use walkdir::WalkDir;
+
+/// Utility: scan the sessions dir for a rollout file that contains `marker`
+/// in any response_item.message.content entry. Returns the absolute path.
+fn find_session_file_containing_marker(
+    sessions_dir: &std::path::Path,
+    marker: &str,
+) -> Option<std::path::PathBuf> {
+    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 path = entry.path();
+        let Ok(content) = std::fs::read_to_string(path) else {
+            continue;
+        };
+        // Skip the first meta line and scan remaining JSONL entries.
+        let mut lines = content.lines();
+        if lines.next().is_none() {
+            continue;
+        }
+        for line in lines {
+            if line.trim().is_empty() {
+                continue;
+            }
+            let Ok(item): Result<Value, _> = serde_json::from_str(line) else {
+                continue;
+            };
+            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")
+                && payload
+                    .get("content")
+                    .map(|c| c.to_string())
+                    .unwrap_or_default()
+                    .contains(marker)
+            {
+                return Some(path.to_path_buf());
+            }
+        }
+    }
+    None
+}
+
+/// Extract the conversation UUID from the first SessionMeta line in the rollout file.
+fn extract_conversation_id(path: &std::path::Path) -> String {
+    let content = std::fs::read_to_string(path).unwrap();
+    let mut lines = content.lines();
+    let meta_line = lines.next().expect("missing meta line");
+    let meta: Value = serde_json::from_str(meta_line).expect("invalid meta json");
+    meta.get("payload")
+        .and_then(|p| p.get("id"))
+        .and_then(|v| v.as_str())
+        .unwrap_or_default()
+        .to_string()
+}
+
+#[test]
+fn exec_resume_last_appends_to_existing_file() -> anyhow::Result<()> {
+    let home = TempDir::new()?;
+    let fixture = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("tests/fixtures/cli_responses_fixture.sse");
+
+    // 1) First run: create a session with a unique marker in the content.
+    let marker = format!("resume-last-{}", Uuid::new_v4());
+    let prompt = format!("echo {marker}");
+
+    Command::cargo_bin("codex-exec")
+        .context("should find binary for codex-exec")?
+        .env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        .env("OPENAI_BASE_URL", "http://unused.local")
+        .arg("--skip-git-repo-check")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt)
+        .assert()
+        .success();
+
+    // Find the created session file containing the marker.
+    let sessions_dir = home.path().join("sessions");
+    let path = find_session_file_containing_marker(&sessions_dir, &marker)
+        .expect("no session file found after first run");
+
+    // 2) Second run: resume the most recent file with a new marker.
+    let marker2 = format!("resume-last-2-{}", Uuid::new_v4());
+    let prompt2 = format!("echo {marker2}");
+
+    let mut binding = assert_cmd::Command::cargo_bin("codex-exec")
+        .context("should find binary for codex-exec")?;
+    let cmd = binding
+        .env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        .env("OPENAI_BASE_URL", "http://unused.local")
+        .arg("--skip-git-repo-check")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt2)
+        .arg("resume")
+        .arg("--last");
+    cmd.assert().success();
+
+    // Ensure the same file was updated and contains both markers.
+    let resumed_path = find_session_file_containing_marker(&sessions_dir, &marker2)
+        .expect("no resumed session file containing marker2");
+    assert_eq!(
+        resumed_path, path,
+        "resume --last should append to existing file"
+    );
+    let content = std::fs::read_to_string(&resumed_path)?;
+    assert!(content.contains(&marker));
+    assert!(content.contains(&marker2));
+    Ok(())
+}
+
+#[test]
+fn exec_resume_by_id_appends_to_existing_file() -> anyhow::Result<()> {
+    let home = TempDir::new()?;
+    let fixture = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("tests/fixtures/cli_responses_fixture.sse");
+
+    // 1) First run: create a session
+    let marker = format!("resume-by-id-{}", Uuid::new_v4());
+    let prompt = format!("echo {marker}");
+
+    Command::cargo_bin("codex-exec")
+        .context("should find binary for codex-exec")?
+        .env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        .env("OPENAI_BASE_URL", "http://unused.local")
+        .arg("--skip-git-repo-check")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt)
+        .assert()
+        .success();
+
+    let sessions_dir = home.path().join("sessions");
+    let path = find_session_file_containing_marker(&sessions_dir, &marker)
+        .expect("no session file found after first run");
+    let session_id = extract_conversation_id(&path);
+    assert!(
+        !session_id.is_empty(),
+        "missing conversation id in meta line"
+    );
+
+    // 2) Resume by id
+    let marker2 = format!("resume-by-id-2-{}", Uuid::new_v4());
+    let prompt2 = format!("echo {marker2}");
+
+    let mut binding = assert_cmd::Command::cargo_bin("codex-exec")
+        .context("should find binary for codex-exec")?;
+    let cmd = binding
+        .env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        .env("OPENAI_BASE_URL", "http://unused.local")
+        .arg("--skip-git-repo-check")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt2)
+        .arg("resume")
+        .arg(&session_id);
+    cmd.assert().success();
+
+    let resumed_path = find_session_file_containing_marker(&sessions_dir, &marker2)
+        .expect("no resumed session file containing marker2");
+    assert_eq!(
+        resumed_path, path,
+        "resume by id should append to existing file"
+    );
+    let content = std::fs::read_to_string(&resumed_path)?;
+    assert!(content.contains(&marker));
+    assert!(content.contains(&marker2));
+    Ok(())
+}
+
+#[test]
+fn exec_resume_preserves_cli_configuration_overrides() -> anyhow::Result<()> {
+    let home = TempDir::new()?;
+    let fixture = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("tests/fixtures/cli_responses_fixture.sse");
+
+    let marker = format!("resume-config-{}", Uuid::new_v4());
+    let prompt = format!("echo {marker}");
+
+    Command::cargo_bin("codex-exec")
+        .context("should find binary for codex-exec")?
+        .env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        .env("OPENAI_BASE_URL", "http://unused.local")
+        .arg("--skip-git-repo-check")
+        .arg("--sandbox")
+        .arg("workspace-write")
+        .arg("--model")
+        .arg("gpt-5")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt)
+        .assert()
+        .success();
+
+    let sessions_dir = home.path().join("sessions");
+    let path = find_session_file_containing_marker(&sessions_dir, &marker)
+        .expect("no session file found after first run");
+
+    let marker2 = format!("resume-config-2-{}", Uuid::new_v4());
+    let prompt2 = format!("echo {marker2}");
+
+    let output = Command::cargo_bin("codex-exec")
+        .context("should find binary for codex-exec")?
+        .env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        .env("OPENAI_BASE_URL", "http://unused.local")
+        .arg("--skip-git-repo-check")
+        .arg("--sandbox")
+        .arg("workspace-write")
+        .arg("--model")
+        .arg("gpt-5-high")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt2)
+        .arg("resume")
+        .arg("--last")
+        .output()
+        .context("resume run should succeed")?;
+
+    assert!(output.status.success(), "resume run failed: {output:?}");
+
+    let stdout = String::from_utf8(output.stdout)?;
+    assert!(
+        stdout.contains("model: gpt-5-high"),
+        "stdout missing model override: {stdout}"
+    );
+    assert!(
+        stdout.contains("sandbox: workspace-write"),
+        "stdout missing sandbox override: {stdout}"
+    );
+
+    let resumed_path = find_session_file_containing_marker(&sessions_dir, &marker2)
+        .expect("no resumed session file containing marker2");
+    assert_eq!(resumed_path, path, "resume should append to same file");
+
+    let content = std::fs::read_to_string(&resumed_path)?;
+    assert!(content.contains(&marker));
+    assert!(content.contains(&marker2));
+    Ok(())
+}

codex-rs/exec/tests/suite/sandbox.rs

@@ -4,27 +4,39 @@ use codex_core::spawn::StdioPolicy;
 use std::collections::HashMap;
 use std::future::Future;
 use std::io;
+use std::path::Path;
 use std::path::PathBuf;
 use std::process::ExitStatus;
+use tokio::fs::create_dir_all;
 use tokio::process::Child;
 
 #[cfg(target_os = "macos")]
 async fn spawn_command_under_sandbox(
     command: Vec<String>,
+    command_cwd: PathBuf,
     sandbox_policy: &SandboxPolicy,
-    cwd: PathBuf,
+    sandbox_cwd: &Path,
     stdio_policy: StdioPolicy,
     env: HashMap<String, String>,
 ) -> std::io::Result<Child> {
     use codex_core::seatbelt::spawn_command_under_seatbelt;
-    spawn_command_under_seatbelt(command, sandbox_policy, cwd, stdio_policy, env).await
+    spawn_command_under_seatbelt(
+        command,
+        command_cwd,
+        sandbox_policy,
+        sandbox_cwd,
+        stdio_policy,
+        env,
+    )
+    .await
 }
 
 #[cfg(target_os = "linux")]
 async fn spawn_command_under_sandbox(
     command: Vec<String>,
+    command_cwd: PathBuf,
     sandbox_policy: &SandboxPolicy,
-    cwd: PathBuf,
+    sandbox_cwd: &Path,
     stdio_policy: StdioPolicy,
     env: HashMap<String, String>,
 ) -> std::io::Result<Child> {
@@ -33,8 +45,9 @@ async fn spawn_command_under_sandbox(
     spawn_command_under_linux_sandbox(
         codex_linux_sandbox_exe,
         command,
+        command_cwd,
         sandbox_policy,
-        cwd,
+        sandbox_cwd,
         stdio_policy,
         env,
     )
@@ -74,14 +87,17 @@ if __name__ == '__main__':
     p.join()
 "#;
 
+    let command_cwd = std::env::current_dir().expect("should be able to get current dir");
+    let sandbox_cwd = command_cwd.clone();
     let mut child = spawn_command_under_sandbox(
         vec![
             "python3".to_string(),
             "-c".to_string(),
             python_code.to_string(),
         ],
+        command_cwd,
         &policy,
-        std::env::current_dir().expect("should be able to get current dir"),
+        sandbox_cwd.as_path(),
         StdioPolicy::Inherit,
         HashMap::new(),
     )
@@ -92,6 +108,88 @@ if __name__ == '__main__':
     assert!(status.success(), "python exited with {status:?}");
 }
 
+#[tokio::test]
+async fn sandbox_distinguishes_command_and_policy_cwds() {
+    let temp = tempfile::tempdir().expect("should be able to create temp dir");
+    let sandbox_root = temp.path().join("sandbox");
+    let command_root = temp.path().join("command");
+    create_dir_all(&sandbox_root).await.expect("mkdir");
+    create_dir_all(&command_root).await.expect("mkdir");
+    let canonical_sandbox_root = tokio::fs::canonicalize(&sandbox_root)
+        .await
+        .expect("canonicalize sandbox root");
+    let canonical_allowed_path = canonical_sandbox_root.join("allowed.txt");
+
+    let disallowed_path = command_root.join("forbidden.txt");
+
+    // Note writable_roots is empty: verify that `canonical_allowed_path` is
+    // writable only because it is under the sandbox policy cwd, not because it
+    // is under a writable root.
+    let policy = SandboxPolicy::WorkspaceWrite {
+        writable_roots: vec![],
+        network_access: false,
+        exclude_tmpdir_env_var: true,
+        exclude_slash_tmp: true,
+    };
+
+    // Attempt to write inside the command cwd, which is outside of the sandbox policy cwd.
+    let mut child = spawn_command_under_sandbox(
+        vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "echo forbidden > forbidden.txt".to_string(),
+        ],
+        command_root.clone(),
+        &policy,
+        canonical_sandbox_root.as_path(),
+        StdioPolicy::Inherit,
+        HashMap::new(),
+    )
+    .await
+    .expect("should spawn command writing to forbidden path");
+
+    let status = child
+        .wait()
+        .await
+        .expect("should wait for forbidden command");
+    assert!(
+        !status.success(),
+        "sandbox unexpectedly allowed writing to command cwd: {status:?}"
+    );
+    let forbidden_exists = tokio::fs::try_exists(&disallowed_path)
+        .await
+        .expect("try_exists failed");
+    assert!(
+        !forbidden_exists,
+        "forbidden path should not have been created"
+    );
+
+    // Writing to the sandbox policy cwd after changing directories into it should succeed.
+    let mut child = spawn_command_under_sandbox(
+        vec![
+            "/usr/bin/touch".to_string(),
+            canonical_allowed_path.to_string_lossy().into_owned(),
+        ],
+        command_root,
+        &policy,
+        canonical_sandbox_root.as_path(),
+        StdioPolicy::Inherit,
+        HashMap::new(),
+    )
+    .await
+    .expect("should spawn command writing to sandbox root");
+
+    let status = child.wait().await.expect("should wait for allowed command");
+    assert!(
+        status.success(),
+        "sandbox blocked allowed write: {status:?}"
+    );
+    let allowed_exists = tokio::fs::try_exists(&canonical_allowed_path)
+        .await
+        .expect("try_exists allowed failed");
+    assert!(allowed_exists, "allowed path should exist");
+}
+
 fn unix_sock_body() {
     unsafe {
         let mut fds = [0i32; 2];
@@ -200,10 +298,13 @@ where
         cmds.push(test_selector.into());
 
         // Your existing launcher:
+        let command_cwd = std::env::current_dir().expect("should be able to get current dir");
+        let sandbox_cwd = command_cwd.clone();
         let mut child = spawn_command_under_sandbox(
             cmds,
+            command_cwd,
             policy,
-            std::env::current_dir().expect("should be able to get current dir"),
+            sandbox_cwd.as_path(),
             stdio_policy,
             HashMap::from([("IN_SANDBOX".into(), "1".into())]),
         )

codex-rs/execpolicy/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-execpolicy"
 version = { workspace = true }
-edition = "2024"
 
 [[bin]]
 name = "codex-execpolicy"
@@ -15,9 +15,8 @@ path = "src/lib.rs"
 workspace = true
 
 [dependencies]
-anyhow = "1"
-starlark = "0.13.0"
 allocative = "0.3.3"
+anyhow = "1"
 clap = { version = "4", features = ["derive"] }
 derive_more = { version = "2", features = ["display"] }
 env_logger = "0.11.5"
@@ -25,9 +24,10 @@ log = "0.4"
 multimap = "0.10.0"
 path-absolutize = "3.1.1"
 regex-lite = "0.1"
-serde = { version = "1.0.194", features = ["derive"] }
-serde_json = "1.0.143"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
 serde_with = { version = "3", features = ["macros"] }
+starlark = "0.13.0"
 
 [dev-dependencies]
 tempfile = "3.13.0"

codex-rs/file-search/Cargo.toml

@@ -17,5 +17,5 @@ clap = { version = "4", features = ["derive"] }
 ignore = "0.4.23"
 nucleo-matcher = "0.3.1"
 serde = { version = "1", features = ["derive"] }
-serde_json = "1.0.143"
+serde_json = "1"
 tokio = { version = "1", features = ["full"] }

codex-rs/linux-sandbox/tests/suite/landlock.rs

@@ -35,9 +35,11 @@ fn create_env_from_core_vars() -> HashMap<String, String> {
 
 #[expect(clippy::print_stdout, clippy::expect_used, clippy::unwrap_used)]
 async fn run_cmd(cmd: &[&str], writable_roots: &[PathBuf], timeout_ms: u64) {
+    let cwd = std::env::current_dir().expect("cwd should exist");
+    let sandbox_cwd = cwd.clone();
     let params = ExecParams {
         command: cmd.iter().map(|elm| elm.to_string()).collect(),
-        cwd: std::env::current_dir().expect("cwd should exist"),
+        cwd,
         timeout_ms: Some(timeout_ms),
         env: create_env_from_core_vars(),
         with_escalated_permissions: None,
@@ -59,6 +61,7 @@ async fn run_cmd(cmd: &[&str], writable_roots: &[PathBuf], timeout_ms: u64) {
         params,
         SandboxType::LinuxSeccomp,
         &sandbox_policy,
+        sandbox_cwd.as_path(),
         &codex_linux_sandbox_exe,
         None,
     )
@@ -121,7 +124,7 @@ async fn test_writable_root() {
 }
 
 #[tokio::test]
-#[should_panic(expected = "Sandbox(Timeout)")]
+#[should_panic(expected = "Sandbox(Timeout")]
 async fn test_timeout() {
     run_cmd(&["sleep", "2"], &[], 50).await;
 }
@@ -133,6 +136,7 @@ async fn test_timeout() {
 #[expect(clippy::expect_used)]
 async fn assert_network_blocked(cmd: &[&str]) {
     let cwd = std::env::current_dir().expect("cwd should exist");
+    let sandbox_cwd = cwd.clone();
     let params = ExecParams {
         command: cmd.iter().map(|s| s.to_string()).collect(),
         cwd,
@@ -151,31 +155,33 @@ async fn assert_network_blocked(cmd: &[&str]) {
         params,
         SandboxType::LinuxSeccomp,
         &sandbox_policy,
+        sandbox_cwd.as_path(),
         &codex_linux_sandbox_exe,
         None,
     )
     .await;
 
-    let (exit_code, stdout, stderr) = match result {
-        Ok(output) => (output.exit_code, output.stdout.text, output.stderr.text),
-        Err(CodexErr::Sandbox(SandboxErr::Denied(exit_code, stdout, stderr))) => {
-            (exit_code, stdout, stderr)
-        }
+    let output = match result {
+        Ok(output) => output,
+        Err(CodexErr::Sandbox(SandboxErr::Denied { output })) => *output,
         _ => {
             panic!("expected sandbox denied error, got: {result:?}");
         }
     };
 
-    dbg!(&stderr);
-    dbg!(&stdout);
-    dbg!(&exit_code);
+    dbg!(&output.stderr.text);
+    dbg!(&output.stdout.text);
+    dbg!(&output.exit_code);
 
     // A completely missing binary exits with 127.  Anything else should also
     // be non‑zero (EPERM from seccomp will usually bubble up as 1, 2, 13…)
     // If—*and only if*—the command exits 0 we consider the sandbox breached.
 
-    if exit_code == 0 {
-        panic!("Network sandbox FAILED - {cmd:?} exited 0\nstdout:\n{stdout}\nstderr:\n{stderr}",);
+    if output.exit_code == 0 {
+        panic!(
+            "Network sandbox FAILED - {cmd:?} exited 0\nstdout:\n{}\nstderr:\n{}",
+            output.stdout.text, output.stderr.text
+        );
     }
 }
 

codex-rs/login/src/server.rs

@@ -238,8 +238,8 @@ async fn process_request(
                         &opts.codex_home,
                         api_key.clone(),
                         tokens.id_token.clone(),
-                        Some(tokens.access_token.clone()),
-                        Some(tokens.refresh_token.clone()),
+                        tokens.access_token.clone(),
+                        tokens.refresh_token.clone(),
                     )
                     .await
                     {
@@ -446,8 +446,8 @@ async fn persist_tokens_async(
     codex_home: &Path,
     api_key: Option<String>,
     id_token: String,
-    access_token: Option<String>,
-    refresh_token: Option<String>,
+    access_token: String,
+    refresh_token: String,
 ) -> io::Result<()> {
     // Reuse existing synchronous logic but run it off the async runtime.
     let codex_home = codex_home.to_path_buf();
@@ -459,43 +459,29 @@ async fn persist_tokens_async(
             std::fs::create_dir_all(parent).map_err(io::Error::other)?;
         }
 
-        let mut auth = read_or_default(&auth_file);
-        if let Some(key) = api_key {
-            auth.openai_api_key = Some(key);
-        }
-        let tokens = auth.tokens.get_or_insert_with(TokenData::default);
-        tokens.id_token = parse_id_token(&id_token).map_err(io::Error::other)?;
-        // Persist chatgpt_account_id if present in claims
+        let mut tokens = TokenData {
+            id_token: parse_id_token(&id_token).map_err(io::Error::other)?,
+            access_token,
+            refresh_token,
+            account_id: None,
+        };
         if let Some(acc) = jwt_auth_claims(&id_token)
             .get("chatgpt_account_id")
             .and_then(|v| v.as_str())
         {
             tokens.account_id = Some(acc.to_string());
         }
-        if let Some(at) = access_token {
-            tokens.access_token = at;
-        }
-        if let Some(rt) = refresh_token {
-            tokens.refresh_token = rt;
-        }
-        auth.last_refresh = Some(Utc::now());
+        let auth = AuthDotJson {
+            openai_api_key: api_key,
+            tokens: Some(tokens),
+            last_refresh: Some(Utc::now()),
+        };
         codex_core::auth::write_auth_json(&auth_file, &auth)
     })
     .await
     .map_err(|e| io::Error::other(format!("persist task failed: {e}")))?
 }
 
-fn read_or_default(path: &Path) -> AuthDotJson {
-    match codex_core::auth::try_read_auth_json(path) {
-        Ok(auth) => auth,
-        Err(_) => AuthDotJson {
-            openai_api_key: None,
-            tokens: None,
-            last_refresh: None,
-        },
-    }
-}
-
 fn compose_success_url(port: u16, issuer: &str, id_token: &str, access_token: &str) -> String {
     let token_claims = jwt_auth_claims(id_token);
     let access_claims = jwt_auth_claims(access_token);

codex-rs/login/tests/suite/login_server_e2e.rs

@@ -90,6 +90,22 @@ async fn end_to_end_login_flow_persists_auth_json() {
     let tmp = tempdir().unwrap();
     let codex_home = tmp.path().to_path_buf();
 
+    // Seed auth.json with stale API key + tokens that should be overwritten.
+    let stale_auth = serde_json::json!({
+        "OPENAI_API_KEY": "sk-stale",
+        "tokens": {
+            "id_token": "stale.header.payload",
+            "access_token": "stale-access",
+            "refresh_token": "stale-refresh",
+            "account_id": "stale-acc"
+        }
+    });
+    std::fs::write(
+        codex_home.join("auth.json"),
+        serde_json::to_string_pretty(&stale_auth).unwrap(),
+    )
+    .unwrap();
+
     let state = "test_state_123".to_string();
 
     // Run server in background
@@ -122,10 +138,10 @@ async fn end_to_end_login_flow_persists_auth_json() {
     let auth_path = codex_home.join("auth.json");
     let data = std::fs::read_to_string(&auth_path).unwrap();
     let json: serde_json::Value = serde_json::from_str(&data).unwrap();
-    assert!(
-        !json["OPENAI_API_KEY"].is_null(),
-        "OPENAI_API_KEY should be set"
-    );
+    // The following assert is here because of the old oauth flow that exchanges tokens for an
+    // API key. See obtain_api_key in server.rs for details. Once we remove this old mechanism
+    // from the code, this test should be updated to expect that the API key is no longer present.
+    assert_eq!(json["OPENAI_API_KEY"], "access-123");
     assert_eq!(json["tokens"]["access_token"], "access-123");
     assert_eq!(json["tokens"]["refresh_token"], "refresh-123");
     assert_eq!(json["tokens"]["account_id"], "acc-123");

codex-rs/mcp-server/Cargo.toml

@@ -36,7 +36,7 @@ tokio = { version = "1", features = [
 toml = "0.9"
 tracing = { version = "0.1.41", features = ["log"] }
 tracing-subscriber = { version = "0.3", features = ["env-filter", "fmt"] }
-uuid = { version = "1", features = ["serde", "v4"] }
+uuid = { version = "1", features = ["serde", "v7"] }
 
 [dev-dependencies]
 assert_cmd = "2"

codex-rs/mcp-server/src/codex_message_processor.rs

@@ -20,7 +20,7 @@ use codex_core::config::ConfigToml;
 use codex_core::config::load_config_as_toml;
 use codex_core::config_edit::CONFIG_KEY_EFFORT;
 use codex_core::config_edit::CONFIG_KEY_MODEL;
-use codex_core::config_edit::persist_non_null_overrides;
+use codex_core::config_edit::persist_overrides_and_clear_if_none;
 use codex_core::default_client::get_codex_user_agent;
 use codex_core::exec::ExecParams;
 use codex_core::exec_env::create_env;
@@ -423,32 +423,41 @@ impl CodexMessageProcessor {
         // Determine whether auth is required based on the active model provider.
         // If a custom provider is configured with `requires_openai_auth == false`,
         // then no auth step is required; otherwise, default to requiring auth.
-        let requires_openai_auth = Some(self.config.model_provider.requires_openai_auth);
+        let requires_openai_auth = self.config.model_provider.requires_openai_auth;
 
-        let response = match self.auth_manager.auth() {
-            Some(auth) => {
-                let (reported_auth_method, token_opt) = match auth.get_token().await {
-                    Ok(token) if !token.is_empty() => {
-                        let tok = if include_token { Some(token) } else { None };
-                        (Some(auth.mode), tok)
-                    }
-                    Ok(_) => (None, None),
-                    Err(err) => {
-                        tracing::warn!("failed to get token for auth status: {err}");
-                        (None, None)
-                    }
-                };
-                codex_protocol::mcp_protocol::GetAuthStatusResponse {
-                    auth_method: reported_auth_method,
-                    auth_token: token_opt,
-                    requires_openai_auth,
-                }
-            }
-            None => codex_protocol::mcp_protocol::GetAuthStatusResponse {
+        let response = if !requires_openai_auth {
+            codex_protocol::mcp_protocol::GetAuthStatusResponse {
                 auth_method: None,
                 auth_token: None,
-                requires_openai_auth,
-            },
+                requires_openai_auth: Some(false),
+            }
+        } else {
+            match self.auth_manager.auth() {
+                Some(auth) => {
+                    let auth_mode = auth.mode;
+                    let (reported_auth_method, token_opt) = match auth.get_token().await {
+                        Ok(token) if !token.is_empty() => {
+                            let tok = if include_token { Some(token) } else { None };
+                            (Some(auth_mode), tok)
+                        }
+                        Ok(_) => (None, None),
+                        Err(err) => {
+                            tracing::warn!("failed to get token for auth status: {err}");
+                            (None, None)
+                        }
+                    };
+                    codex_protocol::mcp_protocol::GetAuthStatusResponse {
+                        auth_method: reported_auth_method,
+                        auth_token: token_opt,
+                        requires_openai_auth: Some(true),
+                    }
+                }
+                None => codex_protocol::mcp_protocol::GetAuthStatusResponse {
+                    auth_method: None,
+                    auth_token: None,
+                    requires_openai_auth: Some(true),
+                },
+            }
         };
 
         self.outgoing.send_response(request_id, response).await;
@@ -519,7 +528,7 @@ impl CodexMessageProcessor {
             (&[CONFIG_KEY_EFFORT], effort_str.as_deref()),
         ];
 
-        match persist_non_null_overrides(
+        match persist_overrides_and_clear_if_none(
             &self.config.codex_home,
             self.config.active_profile.as_deref(),
             &overrides,
@@ -580,12 +589,14 @@ impl CodexMessageProcessor {
         let codex_linux_sandbox_exe = self.config.codex_linux_sandbox_exe.clone();
         let outgoing = self.outgoing.clone();
         let req_id = request_id;
+        let sandbox_cwd = self.config.cwd.clone();
 
         tokio::spawn(async move {
             match codex_core::exec::process_exec_tool_call(
                 exec_params,
                 sandbox_type,
                 &effective_policy,
+                sandbox_cwd.as_path(),
                 &codex_linux_sandbox_exe,
                 None,
             )
@@ -805,7 +816,7 @@ impl CodexMessageProcessor {
             return;
         };
 
-        let required_suffix = format!("{}.jsonl", conversation_id.0);
+        let required_suffix = format!("{conversation_id}.jsonl");
         let Some(file_name) = canonical_rollout_path.file_name().map(OsStr::to_owned) else {
             let error = JSONRPCErrorError {
                 code: INVALID_REQUEST_ERROR_CODE,
@@ -1248,6 +1259,7 @@ fn derive_config_from_params(
     } = params;
     let overrides = ConfigOverrides {
         model,
+        review_model: None,
         config_profile: profile,
         cwd: cwd.map(PathBuf::from),
         approval_policy,
@@ -1404,13 +1416,13 @@ mod tests {
     #[test]
     fn extract_conversation_summary_prefers_plain_user_messages() {
         let conversation_id =
-            ConversationId(Uuid::parse_str("3f941c35-29b3-493b-b0a4-e25800d9aeb0").unwrap());
+            ConversationId::from_string("3f941c35-29b3-493b-b0a4-e25800d9aeb0").unwrap();
         let timestamp = Some("2025-09-05T16:53:11.850Z".to_string());
         let path = PathBuf::from("rollout.jsonl");
 
         let head = vec![
             json!({
-                "id": conversation_id.0,
+                "id": conversation_id.to_string(),
                 "timestamp": timestamp,
                 "cwd": "/",
                 "originator": "codex",