feat: sub /new out from starting slash commands

- Since it doesn't work elegantly at the moment.

.github/actions/codex/bun.lock

@@ -11,7 +11,7 @@
         "@types/bun": "^1.2.19",
         "@types/node": "^24.1.0",
         "prettier": "^3.6.2",
-        "typescript": "^5.8.3",
+        "typescript": "^5.9.2",
       },
     },
   },
@@ -68,7 +68,7 @@
 
     "tunnel": ["tunnel@0.0.6", "", {}, "sha512-1h/Lnq9yajKY2PEbBadPXj3VxsDDu844OnaAo52UVmIzIvwwtBPIuNvkjuzBlTWpfJyUbG3ez0KSBibQkj4ojg=="],
 
-    "typescript": ["typescript@5.8.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-p1diW6TqL9L07nNxvRMM7hMMw4c5XOo/1ibL4aAIGmSAt9slTE1Xgw5KWuof2uTOvCg9BY7ZRi+GaF+7sfgPeQ=="],
+    "typescript": ["typescript@5.9.2", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-CWBzXQrc/qOkhidw1OzBTQuYRbfyxDXJMVJ1XNwUHGROVmuaeiEm3OslpZ1RV96d7SKKjZKrSJu3+t/xlw3R9A=="],
 
     "undici": ["undici@5.29.0", "", { "dependencies": { "@fastify/busboy": "^2.0.0" } }, "sha512-raqeBD6NQK4SkWhQzeYKd1KmIG6dllBOTt55Rmkt4HtI9mwdWtJljnrXjAFUBLTSN67HWrOIZ3EPF4kjUw80Bg=="],
 

.github/actions/codex/package.json

@@ -16,6 +16,6 @@
         "@types/bun": "^1.2.19",
         "@types/node": "^24.1.0",
         "prettier": "^3.6.2",
-        "typescript": "^5.8.3"
+        "typescript": "^5.9.2"
     }
 }

.github/actions/codex/src/process-label.ts

@@ -91,7 +91,38 @@ async function processLabel(
   labelConfig: LabelConfig,
 ): Promise<void> {
   const template = labelConfig.getPromptTemplate();
-  const populatedTemplate = await renderPromptTemplate(template, ctx);
+
+  // If this is a review label, prepend explicit PR-diff scoping guidance to
+  // reduce out-of-scope feedback. Do this before rendering so placeholders in
+  // the guidance (e.g., {CODEX_ACTION_GITHUB_EVENT_PATH}) are substituted.
+  const isReview = label.toLowerCase().includes("review");
+  const reviewScopeGuidance = `
+PR Diff Scope
+- Only review changes between the PR's merge-base and head; do not comment on commits or files outside this range.
+- Derive the base/head SHAs from the event JSON at {CODEX_ACTION_GITHUB_EVENT_PATH}, then compute and use the PR diff for all analysis and comments.
+
+Commands to determine scope
+- Resolve SHAs:
+  - BASE_SHA=$(jq -r '.pull_request.base.sha // .pull_request.base.ref' "{CODEX_ACTION_GITHUB_EVENT_PATH}")
+  - HEAD_SHA=$(jq -r '.pull_request.head.sha // .pull_request.head.ref' "{CODEX_ACTION_GITHUB_EVENT_PATH}")
+  - BASE_SHA=$(git rev-parse "$BASE_SHA")
+  - HEAD_SHA=$(git rev-parse "$HEAD_SHA")
+- Prefer triple-dot (merge-base) semantics for PR diffs:
+  - Changed commits: git log --oneline "$BASE_SHA...$HEAD_SHA"
+  - Changed files: git diff --name-status "$BASE_SHA...$HEAD_SHA"
+  - Review hunks: git diff -U0 "$BASE_SHA...$HEAD_SHA"
+
+Review rules
+- Anchor every comment to a file and hunk present in git diff "$BASE_SHA...$HEAD_SHA".
+- If you mention context outside the diff, label it as "Follow-up (outside this PR scope)" and keep it brief (<=2 bullets).
+- Do not critique commits or files not reachable in the PR range (merge-base(base, head) → head).
+`.trim();
+
+  const effectiveTemplate = isReview
+    ? `${reviewScopeGuidance}\n\n${template}`
+    : template;
+
+  const populatedTemplate = await renderPromptTemplate(effectiveTemplate, ctx);
 
   // Always run Codex and post the resulting message as a comment.
   let commentBody = await runCodex(populatedTemplate, ctx);

.github/codex/labels/codex-rust-review.md

@@ -0,0 +1,139 @@
+Review this PR and respond with a very concise final message, formatted in Markdown.
+
+There should be a summary of the changes (1-2 sentences) and a few bullet points if necessary.
+
+Then provide the **review** (1-2 sentences plus bullet points, friendly tone).
+
+Things to look out for when doing the review:
+
+## General Principles
+
+- **Make sure the pull request body explains the motivation behind the change.** If the author has failed to do this, call it out, and if you think you can deduce the motivation behind the change, propose copy.
+- Ideally, the PR body also contains a small summary of the change. For small changes, the PR title may be sufficient.
+- Each PR should ideally do one conceptual thing. For example, if a PR does a refactoring as well as introducing a new feature, push back and suggest the refactoring be done in a separate PR. This makes things easier for the reviewer, as refactoring changes can often be far-reaching, yet quick to review.
+- When introducing new code, be on the lookout for code that duplicates existing code. When found, propose a way to refactor the existing code such that it should be reused.
+
+## Code Organization
+
+- Each create in the Cargo workspace in `codex-rs` has a specific purpose: make a note if you believe new code is not introduced in the correct crate.
+- When possible, try to keep the `core` crate as small as possible. Non-core but shared logic is often a good candidate for `codex-rs/common`.
+- Be wary of large files and offer suggestions for how to break things into more reasonably-sized files.
+- Rust files should generally be organized such that the public parts of the API appear near the top of the file and helper functions go below. This is analagous to the "inverted pyramid" structure that is favored in journalism.
+
+## Assertions in Tests
+
+Assert the equality of the entire objects instead of doing "piecemeal comparisons," performing `assert_eq!()` on individual fields.
+
+Note that unit tests also function as "executable documentation." As shown in the following example, "piecemeal comparisons" are often more verbose, provide less coverage, and are not as useful as executable documentation.
+
+For example, suppose you have the following enum:
+
+```rust
+#[derive(Debug, PartialEq)]
+enum Message {
+    Request {
+        id: String,
+        method: String,
+        params: Option<serde_json::Value>,
+    },
+    Notification {
+        method: String,
+        params: Option<serde_json::Value>,
+    },
+}
+```
+
+This is an example of a _piecemeal_ comparison:
+
+```rust
+// BAD: Piecemeal Comparison
+
+#[test]
+fn test_get_latest_messages() {
+    let messages = get_latest_messages();
+    assert_eq!(messages.len(), 2);
+
+    let m0 = &messages[0];
+    match m0 {
+        Message::Request { id, method, params } => {
+            assert_eq!(id, "123");
+            assert_eq!(method, "subscribe");
+            assert_eq!(
+                *params,
+                Some(json!({
+                    "conversation_id": "x42z86"
+                }))
+            )
+        }
+        Message::Notification { .. } => {
+            panic!("expected Request");
+        }
+    }
+
+    let m1 = &messages[1];
+    match m1 {
+        Message::Request { .. } => {
+            panic!("expected Notification");
+        }
+        Message::Notification { method, params } => {
+            assert_eq!(method, "log");
+            assert_eq!(
+                *params,
+                Some(json!({
+                    "level": "info",
+                    "message": "subscribed"
+                }))
+            )
+        }
+    }
+}
+```
+
+This is a _deep_ comparison:
+
+```rust
+// GOOD: Verify the entire structure with a single assert_eq!().
+
+use pretty_assertions::assert_eq;
+
+#[test]
+fn test_get_latest_messages() {
+    let messages = get_latest_messages();
+
+    assert_eq!(
+        vec![
+            Message::Request {
+                id: "123".to_string(),
+                method: "subscribe".to_string(),
+                params: Some(json!({
+                    "conversation_id": "x42z86"
+                })),
+            },
+            Message::Notification {
+                method: "log".to_string(),
+                params: Some(json!({
+                    "level": "info",
+                    "message": "subscribed"
+                })),
+            },
+        ],
+        messages,
+    );
+}
+```
+
+## More Tactical Rust Things To Look Out For
+
+- Do not use `unsafe` (unless you have a really, really good reason like using an operating system API directly and no safe wrapper exists). For example, there are cases where it is tempting to use `unsafe` in order to use `std::env::set_var()`, but this indeed `unsafe` and has led to race conditions on multiple occasions. (When this happens, find a mechanism other than environment variables to use for configuration.)
+- Encourage the use of small enums or the newtype pattern in Rust if it helps readability without adding significant cognitive load or lines of code.
+- If you see opportunities for the changes in a diff to use more idiomatic Rust, please make specific recommendations. For example, favor the use of expressions over `return`.
+- When modifying a `Cargo.toml` file, make sure that dependency lists stay alphabetically sorted. Also consider whether a new dependency is added to the appropriate place (e.g., `[dependencies]` versus `[dev-dependencies]`)
+
+## Pull Request Body
+
+- If the nature of the change seems to have a visual component (which is often the case for changes to `codex-rs/tui`), recommend including a screenshot or video to demonstrate the change, if appropriate.
+- References to existing GitHub issues and PRs are encouraged, where appropriate, though you likely do not have network access, so may not be able to help here.
+
+# PR Information
+
+{CODEX_ACTION_GITHUB_EVENT_PATH} contains the JSON that triggered this GitHub workflow. It contains the `base` and `head` refs that define this PR. Both refs are available locally.

.github/workflows/codex.yml

@@ -20,7 +20,7 @@ jobs:
       (github.event_name == 'issues' && (
         (github.event.action == 'labeled' && (github.event.label.name == 'codex-attempt' || github.event.label.name == 'codex-triage'))
       )) ||
-      (github.event_name == 'pull_request' && github.event.action == 'labeled' && github.event.label.name == 'codex-review')
+      (github.event_name == 'pull_request' && github.event.action == 'labeled' && (github.event.label.name == 'codex-review' || github.event.label.name == 'codex-rust-review'))
     runs-on: ubuntu-latest
     permissions:
       contents: write # can push or create branches

.github/workflows/rust-release.yml

@@ -181,9 +181,9 @@ jobs:
           name: ${{ steps.release_name.outputs.name }}
           tag_name: ${{ github.ref_name }}
           files: dist/**
-          # For now, tag releases as "prerelease" because we are not claiming
-          # the Rust CLI is stable yet.
-          prerelease: true
+          # Mark as prerelease only when the version has a suffix after x.y.z
+          # (e.g. -alpha, -beta). Otherwise publish a normal release.
+          prerelease: ${{ contains(steps.release_name.outputs.name, '-') }}
 
       - uses: facebook/dotslash-publish-release@v2
         env:

.vscode/extensions.json

@@ -0,0 +1,5 @@
+{
+    "recommendations": [
+        "tamasfe.even-better-toml",
+    ]
+}

.vscode/settings.json

@@ -6,5 +6,13 @@
     "[rust]": {
         "editor.defaultFormatter": "rust-lang.rust-analyzer",
         "editor.formatOnSave": true,
-    }
+    },
+    "[toml]": {
+        "editor.defaultFormatter": "tamasfe.even-better-toml",
+        "editor.formatOnSave": true,
+    },
+    // Array order for options in ~/.codex/config.toml such as `notify` and the
+    // `args` for an MCP server is significant, so we disable reordering.
+    "evenBetterToml.formatter.reorderArrays": false,
+    "evenBetterToml.formatter.reorderKeys": true,
 }

AGENTS.md

@@ -2,7 +2,9 @@
 
 In the codex-rs folder where the rust code lives:
 
-- Never add or modify any code related to `CODEX_SANDBOX_NETWORK_DISABLED_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.
+- 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.
 
 Before creating a pull request with changes to `codex-rs`, run `just fmt` (in `codex-rs` directory) to format the code and `just fix` (in `codex-rs` directory) to fix any linter issues in the code, ensure the test suite passes by running `cargo test --all-features` in the `codex-rs` directory.
 

NOTICE

@@ -1,2 +1,6 @@
 OpenAI Codex
 Copyright 2025 OpenAI
+
+This project includes code derived from [Ratatui](https://github.com/ratatui/ratatui), licensed under the MIT license.
+Copyright (c) 2016-2022 Florian Dehau
+Copyright (c) 2023-2025 The Ratatui Developers

README.md

@@ -18,6 +18,7 @@ This is the home of the **Codex CLI**, which is a coding agent from OpenAI that
 - [Quickstart](#quickstart)
   - [OpenAI API Users](#openai-api-users)
   - [OpenAI Plus/Pro Users](#openai-pluspro-users)
+- [Using Open Source Models](#using-open-source-models)
 - [Why Codex?](#why-codex)
 - [Security model & permissions](#security-model--permissions)
   - [Platform sandboxing details](#platform-sandboxing-details)
@@ -95,6 +96,12 @@ codex login
 
 If you complete the process successfully, you should have a `~/.codex/auth.json` file that contains the credentials that Codex will use.
 
+To verify whether you are currently logged in, run:
+
+```
+codex login status
+```
+
 If you encounter problems with the login flow, please comment on <https://github.com/openai/codex/issues/1243>.
 
 <details>
@@ -180,6 +187,41 @@ they'll be committed to your working directory.
 
 ---
 
+## Using Open Source Models
+
+Codex can run fully locally against an OpenAI-compatible OSS host (like Ollama) using the `--oss` flag:
+
+- Interactive UI:
+  - codex --oss
+- Non-interactive (programmatic) mode:
+  - echo "Refactor utils" | codex exec --oss
+
+Model selection when using `--oss`:
+
+- If you omit `-m/--model`, Codex defaults to -m gpt-oss:20b and will verify it exists locally (downloading if needed).
+- To pick a different size, pass one of:
+  - -m "gpt-oss:20b"
+  - -m "gpt-oss:120b"
+
+Point Codex at your own OSS host:
+
+- By default, `--oss` talks to http://localhost:11434/v1.
+- To use a different host, set one of these environment variables before running Codex:
+  - CODEX_OSS_BASE_URL, for example:
+    - CODEX_OSS_BASE_URL="http://my-ollama.example.com:11434/v1" codex --oss -m gpt-oss:20b
+  - or CODEX_OSS_PORT (when the host is localhost):
+    - CODEX_OSS_PORT=11434 codex --oss
+
+Advanced: you can persist this in your config instead of environment variables by overriding the built-in `oss` provider in `~/.codex/config.toml`:
+
+```toml
+[model_providers.oss]
+name = "Open Source"
+base_url = "http://my-ollama.example.com:11434/v1"
+```
+
+---
+
 ## Why Codex?
 
 Codex CLI is built for developers who already **live in the terminal** and want

SUMMARY.md

@@ -0,0 +1,21 @@
+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-cli/bin/codex.js

@@ -83,6 +83,7 @@ if (wantsNative && process.platform !== 'win32') {
 
   const child = spawn(binaryPath, process.argv.slice(2), {
     stdio: "inherit",
+    env: { ...process.env, CODEX_MANAGED_BY_NPM: "1" },
   });
 
   child.on("error", (err) => {

codex-cli/src/utils/agent/sandbox/macos-seatbelt.ts

@@ -147,4 +147,8 @@ const READ_ONLY_SEATBELT_POLICY = `
   (sysctl-name "kern.version")
   (sysctl-name "sysctl.proc_cputype")
   (sysctl-name-prefix "hw.perflevel")
-)`.trim();
+)
+
+; Added on top of Chrome profile
+; Needed for python multiprocessing on MacOS for the SemLock
+(allow ipc-posix-sem)`.trim();

codex-rs/Cargo.lock

@@ -661,7 +661,7 @@ dependencies = [
  "clap",
  "codex-core",
  "serde",
- "toml 0.9.2",
+ "toml 0.9.4",
 ]
 
 [[package]]
@@ -673,7 +673,9 @@ dependencies = [
  "async-channel",
  "base64 0.22.1",
  "bytes",
+ "chrono",
  "codex-apply-patch",
+ "codex-login",
  "codex-mcp-client",
  "core_test_support",
  "dirs",
@@ -693,9 +695,11 @@ dependencies = [
  "reqwest",
  "seccompiler",
  "serde",
+ "serde_bytes",
  "serde_json",
  "sha1",
  "shlex",
+ "similar",
  "strum_macros 0.27.2",
  "tempfile",
  "thiserror 2.0.12",
@@ -703,7 +707,7 @@ dependencies = [
  "tokio",
  "tokio-test",
  "tokio-util",
- "toml 0.9.2",
+ "toml 0.9.4",
  "tracing",
  "tree-sitter",
  "tree-sitter-bash",
@@ -725,6 +729,7 @@ dependencies = [
  "codex-arg0",
  "codex-common",
  "codex-core",
+ "codex-ollama",
  "owo-colors",
  "predicates",
  "serde_json",
@@ -791,6 +796,7 @@ dependencies = [
  "reqwest",
  "serde",
  "serde_json",
+ "tempfile",
  "tokio",
 ]
 
@@ -822,22 +828,41 @@ dependencies = [
  "serde",
  "serde_json",
  "shlex",
+ "strum_macros 0.27.2",
  "tempfile",
  "tokio",
  "tokio-test",
- "toml 0.9.2",
+ "toml 0.9.4",
  "tracing",
  "tracing-subscriber",
  "uuid",
  "wiremock",
 ]
 
+[[package]]
+name = "codex-ollama"
+version = "0.0.0"
+dependencies = [
+ "async-stream",
+ "bytes",
+ "codex-core",
+ "futures",
+ "reqwest",
+ "serde_json",
+ "tempfile",
+ "tokio",
+ "toml 0.9.4",
+ "tracing",
+ "wiremock",
+]
+
 [[package]]
 name = "codex-tui"
 version = "0.0.0"
 dependencies = [
  "anyhow",
  "base64 0.22.1",
+ "chrono",
  "clap",
  "codex-ansi-escape",
  "codex-arg0",
@@ -845,6 +870,7 @@ dependencies = [
  "codex-core",
  "codex-file-search",
  "codex-login",
+ "codex-ollama",
  "color-eyre",
  "crossterm",
  "image",
@@ -853,23 +879,28 @@ dependencies = [
  "mcp-types",
  "path-clean",
  "pretty_assertions",
+ "rand 0.8.5",
  "ratatui",
  "ratatui-image",
  "regex-lite",
+ "reqwest",
+ "serde",
  "serde_json",
  "shlex",
  "strum 0.27.2",
  "strum_macros 0.27.2",
+ "supports-color",
+ "textwrap 0.16.2",
  "tokio",
  "tracing",
  "tracing-appender",
  "tracing-subscriber",
  "tui-input",
  "tui-markdown",
- "tui-textarea",
  "unicode-segmentation",
  "unicode-width 0.1.14",
  "uuid",
+ "vt100",
 ]
 
 [[package]]
@@ -1462,7 +1493,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "778e2ac28f6c47af28e4907f13ffd1e1ddbd400980a9abd7c8df189bf578a5ad"
 dependencies = [
  "libc",
- "windows-sys 0.60.2",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -1542,7 +1573,7 @@ checksum = "0ce92ff622d6dadf7349484f42c93271a0d49b7cc4d466a936405bacbe10aa78"
 dependencies = [
  "cfg-if",
  "rustix 1.0.8",
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -1745,7 +1776,7 @@ version = "0.2.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cba6ae63eb948698e300f645f87c70f76630d505f23b8907cf1e193ee85048c1"
 dependencies = [
- "unicode-width 0.2.0",
+ "unicode-width 0.2.1",
 ]
 
 [[package]]
@@ -2006,7 +2037,7 @@ dependencies = [
  "libc",
  "percent-encoding",
  "pin-project-lite",
- "socket2 0.6.0",
+ "socket2",
  "system-configuration",
  "tokio",
  "tower-service",
@@ -2325,9 +2356,15 @@ checksum = "e04d7f318608d35d4b61ddd75cbdaee86b023ebe2bd5a66ee0915f0bf93095a9"
 dependencies = [
  "hermit-abi",
  "libc",
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
+[[package]]
+name = "is_ci"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7655c9839580ee829dfacba1d1278c2b7883e50a277ff7541299489d6bdfdc45"
+
 [[package]]
 name = "is_terminal_polyfill"
 version = "1.70.1"
@@ -2639,6 +2676,7 @@ version = "0.0.0"
 dependencies = [
  "anyhow",
  "assert_cmd",
+ "codex-core",
  "codex-mcp-server",
  "mcp-types",
  "pretty_assertions",
@@ -2646,6 +2684,7 @@ dependencies = [
  "shlex",
  "tempfile",
  "tokio",
+ "uuid",
  "wiremock",
 ]
 
@@ -3373,7 +3412,7 @@ dependencies = [
 [[package]]
 name = "ratatui"
 version = "0.29.0"
-source = "git+https://github.com/nornagon/ratatui?branch=nornagon-v0.29.0-patch#bca287ddc5d38fe088c79e2eda22422b96226f2e"
+source = "git+https://github.com/nornagon/ratatui?branch=nornagon-v0.29.0-patch#9b2ad1298408c45918ee9f8241a6f95498cdbed2"
 dependencies = [
  "bitflags 2.9.1",
  "cassowary",
@@ -3387,7 +3426,7 @@ dependencies = [
  "strum 0.26.3",
  "unicode-segmentation",
  "unicode-truncate",
- "unicode-width 0.2.0",
+ "unicode-width 0.2.1",
 ]
 
 [[package]]
@@ -3701,7 +3740,7 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys 0.4.15",
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -3714,7 +3753,7 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys 0.9.4",
- "windows-sys 0.60.2",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -3945,6 +3984,15 @@ dependencies = [
  "serde_derive",
 ]
 
+[[package]]
+name = "serde_bytes"
+version = "0.11.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8437fd221bde2d4ca316d61b90e337e9e702b3820b87d63caa9ba6c02bd06d96"
+dependencies = [
+ "serde",
+]
+
 [[package]]
 name = "serde_derive"
 version = "1.0.219"
@@ -3969,9 +4017,9 @@ dependencies = [
 
 [[package]]
 name = "serde_json"
-version = "1.0.141"
+version = "1.0.142"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "30b9eff21ebe718216c6ec64e1d9ac57087aad11efc64e32002bce4a0d4c03d3"
+checksum = "030fedb782600dcbd6f02d479bf0d817ac3bb40d644745b769d6a96bc3afc5a7"
 dependencies = [
  "indexmap 2.10.0",
  "itoa",
@@ -4155,14 +4203,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "67b1b7a3b5fe4f1376887184045fcf45c69e92af734b7aaddc05fb777b6fbd03"
 
 [[package]]
-name = "socket2"
-version = "0.5.10"
+name = "smawk"
+version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e22376abed350d73dd1cd119b57ffccad95b4e585a7cda43e286245ce23c0678"
-dependencies = [
- "libc",
- "windows-sys 0.52.0",
-]
+checksum = "b7c388c1b5e93756d0c740965c41e8822f866621d41acbdf6336a6a168f8840c"
 
 [[package]]
 name = "socket2"
@@ -4216,7 +4260,7 @@ dependencies = [
  "starlark_syntax",
  "static_assertions",
  "strsim 0.10.0",
- "textwrap",
+ "textwrap 0.11.0",
  "thiserror 1.0.69",
 ]
 
@@ -4352,6 +4396,15 @@ version = "2.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
 
+[[package]]
+name = "supports-color"
+version = "3.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c64fc7232dd8d2e4ac5ce4ef302b1d81e0b80d055b9d77c7c4f51f6aa4c867d6"
+dependencies = [
+ "is_ci",
+]
+
 [[package]]
 name = "syn"
 version = "1.0.109"
@@ -4466,7 +4519,7 @@ dependencies = [
  "getrandom 0.3.3",
  "once_cell",
  "rustix 1.0.8",
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -4505,6 +4558,17 @@ dependencies = [
  "unicode-width 0.1.14",
 ]
 
+[[package]]
+name = "textwrap"
+version = "0.16.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c13547615a44dc9c452a8a534638acdf07120d4b6847c8178705da06306a3057"
+dependencies = [
+ "smawk",
+ "unicode-linebreak",
+ "unicode-width 0.2.1",
+]
+
 [[package]]
 name = "thiserror"
 version = "1.0.69"
@@ -4619,9 +4683,9 @@ dependencies = [
 
 [[package]]
 name = "tokio"
-version = "1.46.1"
+version = "1.47.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0cc3a2344dafbe23a245241fe8b09735b521110d30fcefbbd5feb1797ca35d17"
+checksum = "89e49afdadebb872d3145a5638b59eb0691ea23e46ca484037cfab3b76b95038"
 dependencies = [
  "backtrace",
  "bytes",
@@ -4632,9 +4696,9 @@ dependencies = [
  "pin-project-lite",
  "signal-hook-registry",
  "slab",
- "socket2 0.5.10",
+ "socket2",
  "tokio-macros",
- "windows-sys 0.52.0",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -4719,9 +4783,9 @@ dependencies = [
 
 [[package]]
 name = "toml"
-version = "0.9.2"
+version = "0.9.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ed0aee96c12fa71097902e0bb061a5e1ebd766a6636bb605ba401c45c1650eac"
+checksum = "41ae868b5a0f67631c14589f7e250c1ea2c574ee5ba21c6c8dd4b1485705a5a1"
 dependencies = [
  "indexmap 2.10.0",
  "serde",
@@ -4950,7 +5014,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "911e93158bf80bbc94bad533b2b16e3d711e1132d69a6a6980c3920a63422c19"
 dependencies = [
  "ratatui",
- "unicode-width 0.2.0",
+ "unicode-width 0.2.1",
 ]
 
 [[package]]
@@ -4969,17 +5033,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "tui-textarea"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0a5318dd619ed73c52a9417ad19046724effc1287fb75cdcc4eca1d6ac1acbae"
-dependencies = [
- "crossterm",
- "ratatui",
- "unicode-width 0.2.0",
-]
-
 [[package]]
 name = "typenum"
 version = "1.18.0"
@@ -4998,6 +5051,12 @@ version = "1.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5a5f39404a5da50712a4c1eecf25e90dd62b613502b7e925fd4e4d19b5c96512"
 
+[[package]]
+name = "unicode-linebreak"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b09c83c3c29d37506a3e260c08c03743a6bb66a9cd432c6934ab501a190571f"
+
 [[package]]
 name = "unicode-segmentation"
 version = "1.12.0"
@@ -5023,9 +5082,9 @@ checksum = "7dd6e30e90baa6f72411720665d41d89b9a3d039dc45b8faea1ddd07f617f6af"
 
 [[package]]
 name = "unicode-width"
-version = "0.2.0"
+version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1fc81956842c57dac11422a97c3b8195a1ff727f06e85c84ed2e8aa277c9a0fd"
+checksum = "4a1a07cc7db3810833284e8d372ccdc6da29741639ecc70c9ec107df0fa6154c"
 
 [[package]]
 name = "unicode-xid"
@@ -5110,6 +5169,27 @@ version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
 
+[[package]]
+name = "vt100"
+version = "0.16.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "054ff75fb8fa83e609e685106df4faeffdf3a735d3c74ebce97ec557d5d36fd9"
+dependencies = [
+ "itoa",
+ "unicode-width 0.2.1",
+ "vte",
+]
+
+[[package]]
+name = "vte"
+version = "0.15.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a5924018406ce0063cd67f8e008104968b74b563ee1b85dde3ed1f7cb87d3dbd"
+dependencies = [
+ "arrayvec",
+ "memchr",
+]
+
 [[package]]
 name = "wait-timeout"
 version = "0.2.1"
@@ -5298,7 +5378,7 @@ version = "0.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cf221c93e13a30d793f7645a0e7762c55d169dbb0a49671918a2319d289b10bb"
 dependencies = [
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]

codex-rs/Cargo.toml

@@ -1,5 +1,4 @@
 [workspace]
-resolver = "2"
 members = [
     "ansi-escape",
     "apply-patch",
@@ -15,8 +14,10 @@ members = [
     "mcp-client",
     "mcp-server",
     "mcp-types",
+    "ollama",
     "tui",
 ]
+resolver = "2"
 
 [workspace.package]
 version = "0.0.0"
@@ -45,4 +46,3 @@ codegen-units = 1
 [patch.crates-io]
 # ratatui = { path = "../../ratatui" }
 ratatui = { git = "https://github.com/nornagon/ratatui", branch = "nornagon-v0.29.0-patch" }
-

codex-rs/ansi-escape/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-ansi-escape"
 version = { workspace = true }
-edition = "2024"
 
 [lib]
 name = "codex_ansi_escape"
@@ -10,7 +10,7 @@ path = "src/lib.rs"
 [dependencies]
 ansi-to-tui = "7.0.0"
 ratatui = { version = "0.29.0", features = [
-    "unstable-widget-ref",
     "unstable-rendered-line-info",
+    "unstable-widget-ref",
 ] }
 tracing = { version = "0.1.41", features = ["log"] }

codex-rs/apply-patch/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-apply-patch"
 version = { workspace = true }
-edition = "2024"
 
 [lib]
 name = "codex_apply_patch"

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

@@ -58,16 +58,24 @@ impl PartialEq for IoError {
 
 #[derive(Debug, PartialEq)]
 pub enum MaybeApplyPatch {
-    Body(Vec<Hunk>),
+    Body(ApplyPatchArgs),
     ShellParseError(ExtractHeredocError),
     PatchParseError(ParseError),
     NotApplyPatch,
 }
 
+/// Both the raw PATCH argument to `apply_patch` as well as the PATCH argument
+/// parsed into hunks.
+#[derive(Debug, PartialEq)]
+pub struct ApplyPatchArgs {
+    pub patch: String,
+    pub hunks: Vec<Hunk>,
+}
+
 pub fn maybe_parse_apply_patch(argv: &[String]) -> MaybeApplyPatch {
     match argv {
         [cmd, body] if cmd == "apply_patch" => match parse_patch(body) {
-            Ok(hunks) => MaybeApplyPatch::Body(hunks),
+            Ok(source) => MaybeApplyPatch::Body(source),
             Err(e) => MaybeApplyPatch::PatchParseError(e),
         },
         [bash, flag, script]
@@ -77,7 +85,7 @@ pub fn maybe_parse_apply_patch(argv: &[String]) -> MaybeApplyPatch {
         {
             match extract_heredoc_body_from_apply_patch_command(script) {
                 Ok(body) => match parse_patch(&body) {
-                    Ok(hunks) => MaybeApplyPatch::Body(hunks),
+                    Ok(source) => MaybeApplyPatch::Body(source),
                     Err(e) => MaybeApplyPatch::PatchParseError(e),
                 },
                 Err(e) => MaybeApplyPatch::ShellParseError(e),
@@ -116,11 +124,19 @@ pub enum MaybeApplyPatchVerified {
     NotApplyPatch,
 }
 
-#[derive(Debug, PartialEq)]
 /// ApplyPatchAction is the result of parsing an `apply_patch` command. By
 /// construction, all paths should be absolute paths.
+#[derive(Debug, PartialEq)]
 pub struct ApplyPatchAction {
     changes: HashMap<PathBuf, ApplyPatchFileChange>,
+
+    /// The raw patch argument that can be used with `apply_patch` as an exec
+    /// call. i.e., if the original arg was parsed in "lenient" mode with a
+    /// heredoc, this should be the value without the heredoc wrapper.
+    pub patch: String,
+
+    /// The working directory that was used to resolve relative paths in the patch.
+    pub cwd: PathBuf,
 }
 
 impl ApplyPatchAction {
@@ -140,8 +156,28 @@ impl ApplyPatchAction {
             panic!("path must be absolute");
         }
 
+        #[allow(clippy::expect_used)]
+        let filename = path
+            .file_name()
+            .expect("path should not be empty")
+            .to_string_lossy();
+        let patch = format!(
+            r#"*** Begin Patch
+*** Update File: {filename}
+@@
++ {content}
+*** End Patch"#,
+        );
         let changes = HashMap::from([(path.to_path_buf(), ApplyPatchFileChange::Add { content })]);
-        Self { changes }
+        #[allow(clippy::expect_used)]
+        Self {
+            changes,
+            cwd: path
+                .parent()
+                .expect("path should have parent")
+                .to_path_buf(),
+            patch,
+        }
     }
 }
 
@@ -149,7 +185,7 @@ impl ApplyPatchAction {
 /// patch.
 pub fn maybe_parse_apply_patch_verified(argv: &[String], cwd: &Path) -> MaybeApplyPatchVerified {
     match maybe_parse_apply_patch(argv) {
-        MaybeApplyPatch::Body(hunks) => {
+        MaybeApplyPatch::Body(ApplyPatchArgs { patch, hunks }) => {
             let mut changes = HashMap::new();
             for hunk in hunks {
                 let path = hunk.resolve_path(cwd);
@@ -183,7 +219,11 @@ pub fn maybe_parse_apply_patch_verified(argv: &[String], cwd: &Path) -> MaybeApp
                     }
                 }
             }
-            MaybeApplyPatchVerified::Body(ApplyPatchAction { changes })
+            MaybeApplyPatchVerified::Body(ApplyPatchAction {
+                changes,
+                patch,
+                cwd: cwd.to_path_buf(),
+            })
         }
         MaybeApplyPatch::ShellParseError(e) => MaybeApplyPatchVerified::ShellParseError(e),
         MaybeApplyPatch::PatchParseError(e) => MaybeApplyPatchVerified::CorrectnessError(e.into()),
@@ -264,7 +304,7 @@ pub fn apply_patch(
     stderr: &mut impl std::io::Write,
 ) -> Result<(), ApplyPatchError> {
     let hunks = match parse_patch(patch) {
-        Ok(hunks) => hunks,
+        Ok(source) => source.hunks,
         Err(e) => {
             match &e {
                 InvalidPatchError(message) => {
@@ -652,7 +692,7 @@ mod tests {
         ]);
 
         match maybe_parse_apply_patch(&args) {
-            MaybeApplyPatch::Body(hunks) => {
+            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, patch: _ }) => {
                 assert_eq!(
                     hunks,
                     vec![Hunk::AddFile {
@@ -679,7 +719,7 @@ PATCH"#,
         ]);
 
         match maybe_parse_apply_patch(&args) {
-            MaybeApplyPatch::Body(hunks) => {
+            MaybeApplyPatch::Body(ApplyPatchArgs { hunks, patch: _ }) => {
                 assert_eq!(
                     hunks,
                     vec![Hunk::AddFile {
@@ -954,7 +994,7 @@ PATCH"#,
         ));
         let patch = parse_patch(&patch).unwrap();
 
-        let update_file_chunks = match patch.as_slice() {
+        let update_file_chunks = match patch.hunks.as_slice() {
             [Hunk::UpdateFile { chunks, .. }] => chunks,
             _ => panic!("Expected a single UpdateFile hunk"),
         };
@@ -992,7 +1032,7 @@ PATCH"#,
         ));
 
         let patch = parse_patch(&patch).unwrap();
-        let chunks = match patch.as_slice() {
+        let chunks = match patch.hunks.as_slice() {
             [Hunk::UpdateFile { chunks, .. }] => chunks,
             _ => panic!("Expected a single UpdateFile hunk"),
         };
@@ -1029,7 +1069,7 @@ PATCH"#,
         ));
 
         let patch = parse_patch(&patch).unwrap();
-        let chunks = match patch.as_slice() {
+        let chunks = match patch.hunks.as_slice() {
             [Hunk::UpdateFile { chunks, .. }] => chunks,
             _ => panic!("Expected a single UpdateFile hunk"),
         };
@@ -1064,7 +1104,7 @@ PATCH"#,
         ));
 
         let patch = parse_patch(&patch).unwrap();
-        let chunks = match patch.as_slice() {
+        let chunks = match patch.hunks.as_slice() {
             [Hunk::UpdateFile { chunks, .. }] => chunks,
             _ => panic!("Expected a single UpdateFile hunk"),
         };
@@ -1110,7 +1150,7 @@ PATCH"#,
 
         // Extract chunks then build the unified diff.
         let parsed = parse_patch(&patch).unwrap();
-        let chunks = match parsed.as_slice() {
+        let chunks = match parsed.hunks.as_slice() {
             [Hunk::UpdateFile { chunks, .. }] => chunks,
             _ => panic!("Expected a single UpdateFile hunk"),
         };
@@ -1193,6 +1233,8 @@ g
                         new_content: "updated session directory content\n".to_string(),
                     },
                 )]),
+                patch: argv[1].clone(),
+                cwd: session_dir.path().to_path_buf(),
             })
         );
     }

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

@@ -22,6 +22,7 @@
 //!
 //! The parser below is a little more lenient than the explicit spec and allows for
 //! leading/trailing whitespace around patch markers.
+use crate::ApplyPatchArgs;
 use std::path::Path;
 use std::path::PathBuf;
 
@@ -102,7 +103,7 @@ pub struct UpdateFileChunk {
     pub is_end_of_file: bool,
 }
 
-pub fn parse_patch(patch: &str) -> Result<Vec<Hunk>, ParseError> {
+pub fn parse_patch(patch: &str) -> Result<ApplyPatchArgs, ParseError> {
     let mode = if PARSE_IN_STRICT_MODE {
         ParseMode::Strict
     } else {
@@ -150,7 +151,7 @@ enum ParseMode {
     Lenient,
 }
 
-fn parse_patch_text(patch: &str, mode: ParseMode) -> Result<Vec<Hunk>, ParseError> {
+fn parse_patch_text(patch: &str, mode: ParseMode) -> Result<ApplyPatchArgs, ParseError> {
     let lines: Vec<&str> = patch.trim().lines().collect();
     let lines: &[&str] = match check_patch_boundaries_strict(&lines) {
         Ok(()) => &lines,
@@ -173,7 +174,8 @@ fn parse_patch_text(patch: &str, mode: ParseMode) -> Result<Vec<Hunk>, ParseErro
         line_number += hunk_lines;
         remaining_lines = &remaining_lines[hunk_lines..]
     }
-    Ok(hunks)
+    let patch = lines.join("\n");
+    Ok(ApplyPatchArgs { hunks, patch })
 }
 
 /// Checks the start and end lines of the patch text for `apply_patch`,
@@ -425,6 +427,7 @@ fn parse_update_file_chunk(
 }
 
 #[test]
+#[allow(clippy::unwrap_used)]
 fn test_parse_patch() {
     assert_eq!(
         parse_patch_text("bad", ParseMode::Strict),
@@ -455,8 +458,10 @@ fn test_parse_patch() {
             "*** Begin Patch\n\
              *** End Patch",
             ParseMode::Strict
-        ),
-        Ok(Vec::new())
+        )
+        .unwrap()
+        .hunks,
+        Vec::new()
     );
     assert_eq!(
         parse_patch_text(
@@ -472,8 +477,10 @@ fn test_parse_patch() {
              +    return 123\n\
              *** End Patch",
             ParseMode::Strict
-        ),
-        Ok(vec![
+        )
+        .unwrap()
+        .hunks,
+        vec![
             AddFile {
                 path: PathBuf::from("path/add.py"),
                 contents: "abc\ndef\n".to_string()
@@ -491,7 +498,7 @@ fn test_parse_patch() {
                     is_end_of_file: false
                 }]
             }
-        ])
+        ]
     );
     // Update hunk followed by another hunk (Add File).
     assert_eq!(
@@ -504,8 +511,10 @@ fn test_parse_patch() {
              +content\n\
              *** End Patch",
             ParseMode::Strict
-        ),
-        Ok(vec![
+        )
+        .unwrap()
+        .hunks,
+        vec![
             UpdateFile {
                 path: PathBuf::from("file.py"),
                 move_path: None,
@@ -520,7 +529,7 @@ fn test_parse_patch() {
                 path: PathBuf::from("other.py"),
                 contents: "content\n".to_string()
             }
-        ])
+        ]
     );
 
     // Update hunk without an explicit @@ header for the first chunk should parse.
@@ -533,8 +542,10 @@ fn test_parse_patch() {
 +bar
 *** End Patch"#,
             ParseMode::Strict
-        ),
-        Ok(vec![UpdateFile {
+        )
+        .unwrap()
+        .hunks,
+        vec![UpdateFile {
             path: PathBuf::from("file2.py"),
             move_path: None,
             chunks: vec![UpdateFileChunk {
@@ -543,7 +554,7 @@ fn test_parse_patch() {
                 new_lines: vec!["import foo".to_string(), "bar".to_string()],
                 is_end_of_file: false,
             }],
-        }])
+        }]
     );
 }
 
@@ -574,7 +585,10 @@ fn test_parse_patch_lenient() {
     );
     assert_eq!(
         parse_patch_text(&patch_text_in_heredoc, ParseMode::Lenient),
-        Ok(expected_patch.clone())
+        Ok(ApplyPatchArgs {
+            hunks: expected_patch.clone(),
+            patch: patch_text.to_string()
+        })
     );
 
     let patch_text_in_single_quoted_heredoc = format!("<<'EOF'\n{patch_text}\nEOF\n");
@@ -584,7 +598,10 @@ fn test_parse_patch_lenient() {
     );
     assert_eq!(
         parse_patch_text(&patch_text_in_single_quoted_heredoc, ParseMode::Lenient),
-        Ok(expected_patch.clone())
+        Ok(ApplyPatchArgs {
+            hunks: expected_patch.clone(),
+            patch: patch_text.to_string()
+        })
     );
 
     let patch_text_in_double_quoted_heredoc = format!("<<\"EOF\"\n{patch_text}\nEOF\n");
@@ -594,7 +611,10 @@ fn test_parse_patch_lenient() {
     );
     assert_eq!(
         parse_patch_text(&patch_text_in_double_quoted_heredoc, ParseMode::Lenient),
-        Ok(expected_patch.clone())
+        Ok(ApplyPatchArgs {
+            hunks: expected_patch.clone(),
+            patch: patch_text.to_string()
+        })
     );
 
     let patch_text_in_mismatched_quotes_heredoc = format!("<<\"EOF'\n{patch_text}\nEOF\n");

codex-rs/arg0/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-arg0"
 version = { workspace = true }
-edition = "2024"
 
 [lib]
 name = "codex_arg0"

codex-rs/arg0/src/lib.rs

@@ -2,6 +2,8 @@ use std::future::Future;
 use std::path::Path;
 use std::path::PathBuf;
 
+use codex_core::CODEX_APPLY_PATCH_ARG1;
+
 /// While we want to deploy the Codex CLI as a single executable for simplicity,
 /// we also want to expose some of its functionality as distinct CLIs, so we use
 /// the "arg0 trick" to determine which CLI to dispatch. This effectively allows
@@ -43,7 +45,7 @@ where
     }
 
     let argv1 = args.next().unwrap_or_default();
-    if argv1 == "--codex-run-as-apply-patch" {
+    if argv1 == CODEX_APPLY_PATCH_ARG1 {
         let patch_arg = args.next().and_then(|s| s.to_str().map(|s| s.to_owned()));
         let exit_code = match patch_arg {
             Some(patch_arg) => {
@@ -55,7 +57,7 @@ where
                 }
             }
             None => {
-                eprintln!("Error: --codex-run-as-apply-patch requires a UTF-8 PATCH argument.");
+                eprintln!("Error: {CODEX_APPLY_PATCH_ARG1} requires a UTF-8 PATCH argument.");
                 1
             }
         };

codex-rs/chatgpt/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-chatgpt"
 version = { workspace = true }
-edition = "2024"
 
 [lints]
 workspace = true
@@ -9,12 +9,12 @@ workspace = true
 [dependencies]
 anyhow = "1"
 clap = { version = "4", features = ["derive"] }
-serde = { version = "1", features = ["derive"] }
-serde_json = "1"
 codex-common = { path = "../common", features = ["cli"] }
 codex-core = { path = "../core" }
 codex-login = { path = "../login" }
 reqwest = { version = "0.12", features = ["json", "stream"] }
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
 tokio = { version = "1", features = ["full"] }
 
 [dev-dependencies]

codex-rs/chatgpt/src/chatgpt_client.rs

@@ -21,10 +21,14 @@ pub(crate) async fn chatgpt_get_request<T: DeserializeOwned>(
     let token =
         get_chatgpt_token_data().ok_or_else(|| anyhow::anyhow!("ChatGPT token not available"))?;
 
+    let account_id = token.account_id.ok_or_else(|| {
+        anyhow::anyhow!("ChatGPT account ID not available, please re-run `codex login`")
+    });
+
     let response = client
         .get(&url)
         .bearer_auth(&token.access_token)
-        .header("chatgpt-account-id", &token.account_id)
+        .header("chatgpt-account-id", account_id?)
         .header("Content-Type", "application/json")
         .header("User-Agent", "codex-cli")
         .send()

codex-rs/chatgpt/src/chatgpt_token.rs

@@ -18,7 +18,10 @@ pub fn set_chatgpt_token_data(value: TokenData) {
 
 /// Initialize the ChatGPT token from auth.json file
 pub async fn init_chatgpt_token_from_auth(codex_home: &Path) -> std::io::Result<()> {
-    let auth_json = codex_login::try_read_auth_json(codex_home).await?;
-    set_chatgpt_token_data(auth_json.tokens.clone());
+    let auth = codex_login::load_auth(codex_home, true)?;
+    if let Some(auth) = auth {
+        let token_data = auth.get_token_data().await?;
+        set_chatgpt_token_data(token_data);
+    }
     Ok(())
 }

codex-rs/chatgpt/tests/apply_command_e2e.rs

@@ -10,8 +10,13 @@ use tokio::process::Command;
 async fn create_temp_git_repo() -> anyhow::Result<TempDir> {
     let temp_dir = TempDir::new()?;
     let repo_path = temp_dir.path();
+    let envs = vec![
+        ("GIT_CONFIG_GLOBAL", "/dev/null"),
+        ("GIT_CONFIG_NOSYSTEM", "1"),
+    ];
 
     let output = Command::new("git")
+        .envs(envs.clone())
         .args(["init"])
         .current_dir(repo_path)
         .output()
@@ -25,12 +30,14 @@ async fn create_temp_git_repo() -> anyhow::Result<TempDir> {
     }
 
     Command::new("git")
+        .envs(envs.clone())
         .args(["config", "user.email", "test@example.com"])
         .current_dir(repo_path)
         .output()
         .await?;
 
     Command::new("git")
+        .envs(envs.clone())
         .args(["config", "user.name", "Test User"])
         .current_dir(repo_path)
         .output()
@@ -39,12 +46,14 @@ async fn create_temp_git_repo() -> anyhow::Result<TempDir> {
     std::fs::write(repo_path.join("README.md"), "# Test Repo\n")?;
 
     Command::new("git")
+        .envs(envs.clone())
         .args(["add", "README.md"])
         .current_dir(repo_path)
         .output()
         .await?;
 
     let output = Command::new("git")
+        .envs(envs.clone())
         .args(["commit", "-m", "Initial commit"])
         .current_dir(repo_path)
         .output()

codex-rs/cli/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-cli"
 version = { workspace = true }
-edition = "2024"
 
 [[bin]]
 name = "codex"
@@ -20,8 +20,8 @@ clap = { version = "4", features = ["derive"] }
 clap_complete = "4"
 codex-arg0 = { path = "../arg0" }
 codex-chatgpt = { path = "../chatgpt" }
-codex-core = { path = "../core" }
 codex-common = { path = "../common", features = ["cli"] }
+codex-core = { path = "../core" }
 codex-exec = { path = "../exec" }
 codex-login = { path = "../login" }
 codex-mcp-server = { path = "../mcp-server" }

codex-rs/cli/src/debug_sandbox.rs

@@ -4,10 +4,10 @@ use codex_common::CliConfigOverrides;
 use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
 use codex_core::config_types::SandboxMode;
-use codex_core::exec::StdioPolicy;
 use codex_core::exec::spawn_command_under_linux_sandbox;
-use codex_core::exec::spawn_command_under_seatbelt;
 use codex_core::exec_env::create_env;
+use codex_core::seatbelt::spawn_command_under_seatbelt;
+use codex_core::spawn::StdioPolicy;
 
 use crate::LandlockCommand;
 use crate::SeatbeltCommand;

codex-rs/cli/src/login.rs

@@ -1,25 +1,16 @@
+use std::env;
+
 use codex_common::CliConfigOverrides;
 use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
+use codex_login::AuthMode;
+use codex_login::OPENAI_API_KEY_ENV_VAR;
+use codex_login::load_auth;
+use codex_login::login_with_api_key;
 use codex_login::login_with_chatgpt;
 
 pub async fn run_login_with_chatgpt(cli_config_overrides: CliConfigOverrides) -> ! {
-    let cli_overrides = match cli_config_overrides.parse_overrides() {
-        Ok(v) => v,
-        Err(e) => {
-            eprintln!("Error parsing -c overrides: {e}");
-            std::process::exit(1);
-        }
-    };
-
-    let config_overrides = ConfigOverrides::default();
-    let config = match Config::load_with_cli_overrides(cli_overrides, config_overrides) {
-        Ok(config) => config,
-        Err(e) => {
-            eprintln!("Error loading configuration: {e}");
-            std::process::exit(1);
-        }
-    };
+    let config = load_config_or_exit(cli_config_overrides);
 
     let capture_output = false;
     match login_with_chatgpt(&config.codex_home, capture_output).await {
@@ -33,3 +24,103 @@ pub async fn run_login_with_chatgpt(cli_config_overrides: CliConfigOverrides) ->
         }
     }
 }
+
+pub async fn run_login_with_api_key(
+    cli_config_overrides: CliConfigOverrides,
+    api_key: String,
+) -> ! {
+    let config = load_config_or_exit(cli_config_overrides);
+
+    match login_with_api_key(&config.codex_home, &api_key) {
+        Ok(_) => {
+            eprintln!("Successfully logged in");
+            std::process::exit(0);
+        }
+        Err(e) => {
+            eprintln!("Error logging in: {e}");
+            std::process::exit(1);
+        }
+    }
+}
+
+pub async fn run_login_status(cli_config_overrides: CliConfigOverrides) -> ! {
+    let config = load_config_or_exit(cli_config_overrides);
+
+    match load_auth(&config.codex_home, true) {
+        Ok(Some(auth)) => match auth.mode {
+            AuthMode::ApiKey => {
+                if let Some(api_key) = auth.api_key.as_deref() {
+                    eprintln!("Logged in using an API key - {}", safe_format_key(api_key));
+
+                    if let Ok(env_api_key) = env::var(OPENAI_API_KEY_ENV_VAR) {
+                        if env_api_key == api_key {
+                            eprintln!(
+                                "   API loaded from OPENAI_API_KEY environment variable or .env file"
+                            );
+                        }
+                    }
+                } else {
+                    eprintln!("Logged in using an API key");
+                }
+                std::process::exit(0);
+            }
+            AuthMode::ChatGPT => {
+                eprintln!("Logged in using ChatGPT");
+                std::process::exit(0);
+            }
+        },
+        Ok(None) => {
+            eprintln!("Not logged in");
+            std::process::exit(1);
+        }
+        Err(e) => {
+            eprintln!("Error checking login status: {e}");
+            std::process::exit(1);
+        }
+    }
+}
+
+fn load_config_or_exit(cli_config_overrides: CliConfigOverrides) -> Config {
+    let cli_overrides = match cli_config_overrides.parse_overrides() {
+        Ok(v) => v,
+        Err(e) => {
+            eprintln!("Error parsing -c overrides: {e}");
+            std::process::exit(1);
+        }
+    };
+
+    let config_overrides = ConfigOverrides::default();
+    match Config::load_with_cli_overrides(cli_overrides, config_overrides) {
+        Ok(config) => config,
+        Err(e) => {
+            eprintln!("Error loading configuration: {e}");
+            std::process::exit(1);
+        }
+    }
+}
+
+fn safe_format_key(key: &str) -> String {
+    if key.len() <= 13 {
+        return "***".to_string();
+    }
+    let prefix = &key[..8];
+    let suffix = &key[key.len() - 5..];
+    format!("{prefix}***{suffix}")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::safe_format_key;
+
+    #[test]
+    fn formats_long_key() {
+        let key = "sk-proj-1234567890ABCDE";
+        assert_eq!(safe_format_key(key), "sk-proj-***ABCDE");
+    }
+
+    #[test]
+    fn short_key_returns_stars() {
+        let key = "sk-proj-12345";
+        assert_eq!(safe_format_key(key), "***");
+    }
+}

codex-rs/cli/src/main.rs

@@ -7,6 +7,8 @@ use codex_chatgpt::apply_command::ApplyCommand;
 use codex_chatgpt::apply_command::run_apply_command;
 use codex_cli::LandlockCommand;
 use codex_cli::SeatbeltCommand;
+use codex_cli::login::run_login_status;
+use codex_cli::login::run_login_with_api_key;
 use codex_cli::login::run_login_with_chatgpt;
 use codex_cli::proto;
 use codex_common::CliConfigOverrides;
@@ -43,7 +45,7 @@ enum Subcommand {
     #[clap(visible_alias = "e")]
     Exec(ExecCli),
 
-    /// Login with ChatGPT.
+    /// Manage login.
     Login(LoginCommand),
 
     /// Experimental: run Codex as an MCP server.
@@ -90,6 +92,18 @@ enum DebugCommand {
 struct LoginCommand {
     #[clap(skip)]
     config_overrides: CliConfigOverrides,
+
+    #[arg(long = "api-key", value_name = "API_KEY")]
+    api_key: Option<String>,
+
+    #[command(subcommand)]
+    action: Option<LoginSubcommand>,
+}
+
+#[derive(Debug, clap::Subcommand)]
+enum LoginSubcommand {
+    /// Show login status.
+    Status,
 }
 
 fn main() -> anyhow::Result<()> {
@@ -106,7 +120,7 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
         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)?;
+            let usage = codex_tui::run_main(tui_cli, codex_linux_sandbox_exe).await?;
             println!("{}", codex_core::protocol::FinalOutput::from(usage));
         }
         Some(Subcommand::Exec(mut exec_cli)) => {
@@ -118,7 +132,18 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
         }
         Some(Subcommand::Login(mut login_cli)) => {
             prepend_config_flags(&mut login_cli.config_overrides, cli.config_overrides);
-            run_login_with_chatgpt(login_cli.config_overrides).await;
+            match login_cli.action {
+                Some(LoginSubcommand::Status) => {
+                    run_login_status(login_cli.config_overrides).await;
+                }
+                None => {
+                    if let Some(api_key) = login_cli.api_key {
+                        run_login_with_api_key(login_cli.config_overrides, api_key).await;
+                    } else {
+                        run_login_with_chatgpt(login_cli.config_overrides).await;
+                    }
+                }
+            }
         }
         Some(Subcommand::Proto(mut proto_cli)) => {
             prepend_config_flags(&mut proto_cli.config_overrides, cli.config_overrides);

codex-rs/cli/src/proto.rs

@@ -9,6 +9,7 @@ use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
 use codex_core::protocol::Submission;
 use codex_core::util::notify_on_sigint;
+use codex_login::load_auth;
 use tokio::io::AsyncBufReadExt;
 use tokio::io::BufReader;
 use tracing::error;
@@ -35,8 +36,9 @@ pub async fn run_main(opts: ProtoCli) -> anyhow::Result<()> {
         .map_err(anyhow::Error::msg)?;
 
     let config = Config::load_with_cli_overrides(overrides_vec, ConfigOverrides::default())?;
+    let auth = load_auth(&config.codex_home, true)?;
     let ctrl_c = notify_on_sigint();
-    let CodexSpawnOk { codex, .. } = Codex::spawn(config, ctrl_c.clone()).await?;
+    let CodexSpawnOk { codex, .. } = Codex::spawn(config, auth, ctrl_c.clone()).await?;
     let codex = Arc::new(codex);
 
     // Task that reads JSON lines from stdin and forwards to Submission Queue

codex-rs/common/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-common"
 version = { workspace = true }
-edition = "2024"
 
 [lints]
 workspace = true
@@ -9,11 +9,11 @@ workspace = true
 [dependencies]
 clap = { version = "4", features = ["derive", "wrap_help"], optional = true }
 codex-core = { path = "../core" }
-toml = { version = "0.9", optional = true }
 serde = { version = "1", optional = true }
+toml = { version = "0.9", optional = true }
 
 [features]
 # Separate feature so that `clap` is not a mandatory dependency.
-cli = ["clap", "toml", "serde"]
+cli = ["clap", "serde", "toml"]
 elapsed = []
 sandbox_summary = []

codex-rs/common/src/approval_mode_cli_arg.rs

@@ -18,6 +18,9 @@ pub enum ApprovalModeCliArg {
     /// will escalate to the user to ask for un-sandboxed execution.
     OnFailure,
 
+    /// The model decides when to ask the user for approval.
+    OnRequest,
+
     /// Never ask for user approval
     /// Execution failures are immediately returned to the model.
     Never,
@@ -28,6 +31,7 @@ impl From<ApprovalModeCliArg> for AskForApproval {
         match value {
             ApprovalModeCliArg::Untrusted => AskForApproval::UnlessTrusted,
             ApprovalModeCliArg::OnFailure => AskForApproval::OnFailure,
+            ApprovalModeCliArg::OnRequest => AskForApproval::OnRequest,
             ApprovalModeCliArg::Never => AskForApproval::Never,
         }
     }

codex-rs/common/src/config_summary.rs

@@ -0,0 +1,29 @@
+use codex_core::WireApi;
+use codex_core::config::Config;
+
+use crate::sandbox_summary::summarize_sandbox_policy;
+
+/// Build a list of key/value pairs summarizing the effective configuration.
+pub fn create_config_summary_entries(config: &Config) -> Vec<(&'static str, String)> {
+    let mut entries = vec![
+        ("workdir", config.cwd.display().to_string()),
+        ("model", config.model.clone()),
+        ("provider", config.model_provider_id.clone()),
+        ("approval", config.approval_policy.to_string()),
+        ("sandbox", summarize_sandbox_policy(&config.sandbox_policy)),
+    ];
+    if config.model_provider.wire_api == WireApi::Responses
+        && config.model_family.supports_reasoning_summaries
+    {
+        entries.push((
+            "reasoning effort",
+            config.model_reasoning_effort.to_string(),
+        ));
+        entries.push((
+            "reasoning summaries",
+            config.model_reasoning_summary.to_string(),
+        ));
+    }
+
+    entries
+}

codex-rs/common/src/lib.rs

@@ -23,3 +23,7 @@ mod sandbox_summary;
 
 #[cfg(feature = "sandbox_summary")]
 pub use sandbox_summary::summarize_sandbox_policy;
+
+mod config_summary;
+
+pub use config_summary::create_config_summary_entries;

codex-rs/common/src/sandbox_summary.rs

@@ -7,6 +7,7 @@ pub fn summarize_sandbox_policy(sandbox_policy: &SandboxPolicy) -> String {
         SandboxPolicy::WorkspaceWrite {
             writable_roots,
             network_access,
+            include_default_writable_roots,
         } => {
             let mut summary = "workspace-write".to_string();
             if !writable_roots.is_empty() {
@@ -19,6 +20,9 @@ pub fn summarize_sandbox_policy(sandbox_policy: &SandboxPolicy) -> String {
                         .join(", ")
                 ));
             }
+            if !*include_default_writable_roots {
+                summary.push_str(" (exact writable roots)");
+            }
             if *network_access {
                 summary.push_str(" (network access enabled)");
             }

codex-rs/config.md

@@ -110,12 +110,15 @@ stream_idle_timeout_ms = 300000    # 5m idle timeout
 ```
 
 #### request_max_retries
+
 How many times Codex will retry a failed HTTP request to the model provider. Defaults to `4`.
 
 #### stream_max_retries
+
 Number of times Codex will attempt to reconnect when a streaming response is interrupted. Defaults to `10`.
 
 #### stream_idle_timeout_ms
+
 How long Codex will wait for activity on a streaming response before treating the connection as lost. Defaults to `300_000` (5 minutes).
 
 ## model_provider
@@ -145,12 +148,20 @@ Determines when the user should be prompted to approve whether Codex can execute
 approval_policy = "untrusted"
 ```
 
+If you want to be notified whenever a command fails, use "on-failure":
 ```toml
 # If the command fails when run in the sandbox, Codex asks for permission to
 # retry the command outside the sandbox.
 approval_policy = "on-failure"
 ```
 
+If you want the model to run until it decides that it needs to ask you for escalated permissions, use "on-request":
+```toml
+# The model decides when to escalate
+approval_policy = "on-request"
+```
+
+Alternatively, you can have the model run until it is done, and never ask to run a command with escalated permissions:
 ```toml
 # User is never prompted: if the command fails, Codex will automatically try
 # something out. Note the `exec` subcommand always uses this mode.
@@ -256,6 +267,8 @@ disk, but attempts to write a file or access the network will be blocked.
 
 A more relaxed policy is `workspace-write`. When specified, the current working directory for the Codex task will be writable (as well as `$TMPDIR` on macOS). Note that the CLI defaults to using the directory where it was spawned as `cwd`, though this can be overridden using `--cwd/-C`.
 
+On macOS (and soon Linux), all writable roots (including `cwd`) that contain a `.git/` folder _as an immediate child_ will configure the `.git/` folder to be read-only while the rest of the Git repository will be writable. This means that commands like `git commit` will fail, by default (as it entails writing to `.git/`), and will require Codex to ask for permission.
+
 ```toml
 # same as `--sandbox workspace-write`
 sandbox_mode = "workspace-write"
@@ -478,6 +491,19 @@ Setting `hide_agent_reasoning` to `true` suppresses these events in **both** the
 hide_agent_reasoning = true   # defaults to false
 ```
 
+## show_raw_agent_reasoning
+
+Surfaces the model’s raw chain-of-thought ("raw reasoning content") when available.
+
+Notes:
+- Only takes effect if the selected model/provider actually emits raw reasoning content. Many models do not. When unsupported, this option has no visible effect.
+- Raw reasoning may include intermediate thoughts or sensitive context. Enable only if acceptable for your workflow.
+
+Example:
+```toml
+show_raw_agent_reasoning = true  # defaults to false
+```
+
 ## model_context_window
 
 The size of the context window for the model, in tokens.

codex-rs/core/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-core"
 version = { workspace = true }
-edition = "2024"
 
 [lib]
 name = "codex_core"
@@ -15,7 +15,9 @@ anyhow = "1"
 async-channel = "2.3.1"
 base64 = "0.22"
 bytes = "1.10.1"
+chrono = { version = "0.4", features = ["serde"] }
 codex-apply-patch = { path = "../apply-patch" }
+codex-login = { path = "../login" }
 codex-mcp-client = { path = "../mcp-client" }
 dirs = "6"
 env-flags = "0.1.1"
@@ -29,8 +31,10 @@ rand = "0.9"
 reqwest = { version = "0.12", features = ["json", "stream"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+serde_bytes = "0.11"
 sha1 = "0.10.6"
 shlex = "1.3.0"
+similar = "2.7.0"
 strum_macros = "0.27.2"
 thiserror = "2.0.12"
 time = { version = "0.3", features = ["formatting", "local-offset", "macros"] }
@@ -42,13 +46,13 @@ tokio = { version = "1", features = [
     "signal",
 ] }
 tokio-util = "0.7.14"
-toml = "0.9.2"
+toml = "0.9.4"
 tracing = { version = "0.1.41", features = ["log"] }
 tree-sitter = "0.25.8"
 tree-sitter-bash = "0.25.0"
 uuid = { version = "1", features = ["serde", "v4"] }
-wildmatch = "2.4.0"
 whoami = "1.6.0"
+wildmatch = "2.4.0"
 
 
 [target.'cfg(target_os = "linux")'.dependencies]

codex-rs/core/prompt.md

@@ -1,16 +1,31 @@
-Please resolve the user's task by editing and testing the code files in your current code execution session.
-You are a deployed coding agent.
-Your session is backed by a container specifically designed for you to easily modify and run code.
-The repo(s) are already cloned in your working directory, and you must fully solve the problem for your answer to be considered correct.
+You are operating as and within the Codex CLI, an open-source, terminal-based agentic coding assistant built by OpenAI. It wraps OpenAI models to enable natural language interaction with a local codebase. You are expected to be precise, safe, and helpful.
 
+Your capabilities:
+- Receive user prompts, project context, and files.
+- Stream responses and emit function calls (e.g., shell commands, code edits).
+- Run commands, like apply_patch, and manage user approvals based on policy.
+- Work inside a workspace with sandboxing instructions specified by the policy described in (## Sandbox environment and approval instructions)
+
+Within this context, Codex refers to the open-source agentic coding interface (not the old Codex language model built by OpenAI).
+
+## General guidelines
+As a deployed coding agent, please continue working on the user's task until their query is resolved, before ending your turn and yielding back to the user. Only terminate your turn when you are sure that the task is solved. If you are not sure about file content or codebase structure pertaining to the user's request, use your tools to read files and gather the relevant information. Do NOT guess or make up an answer.
+
+After a user sends their first message, you should immediately provide a brief message acknowledging their request to set the tone and expectation of future work to be done (no more than 8-10 words). This should be done before performing work like exploring the codebase, writing or reading files, or other tool calls needed to complete the task. Use a natural, collaborative tone similar to how a teammate would receive a task during a pair programming session.
+
+Please resolve the user's task by editing the code files in your current code execution session. Your session allows for you to modify and run code. The repo(s) are already cloned in your working directory, and you must fully solve the problem for your answer to be considered correct.
+
+### Task execution
 You MUST adhere to the following criteria when executing the task:
 
 - Working on the repo(s) in the current environment is allowed, even if they are proprietary.
 - Analyzing code for vulnerabilities is allowed.
 - Showing user code and tool call details is allowed.
 - User instructions may overwrite the _CODING GUIDELINES_ section in this developer message.
+- `user_instructions` are not part of the user's request, but guidance for how to complete the task.
+- Do not cite `user_instructions` back to the user unless a specific piece is relevant.
 - Do not use \`ls -R\`, \`find\`, or \`grep\` - these are slow in large repos. Use \`rg\` and \`rg --files\`.
-- Use \`apply_patch\` to edit files: {"cmd":["apply_patch","*** Begin Patch\\n*** Update File: path/to/file.py\\n@@ def example():\\n- pass\\n+ return 123\\n*** End Patch"]}
+- Use the \`apply_patch\` shell command to edit files: {"command":["apply_patch","*** Begin Patch\\n*** Update File: path/to/file.py\\n@@ def example():\\n- pass\\n+ return 123\\n*** End Patch"]}
 - If completing the user's task requires writing or modifying files:
   - Your code and final answer should follow these _CODING GUIDELINES_:
     - Fix the problem at the root cause rather than applying surface-level patches, when possible.
@@ -33,23 +48,22 @@ You MUST adhere to the following criteria when executing the task:
 - If completing the user's task DOES NOT require writing or modifying files (e.g., the user asks a question about the code base):
   - Respond in a friendly tune as a remote teammate, who is knowledgeable, capable and eager to help with coding.
 - When your task involves writing or modifying files:
-  - Do NOT tell the user to "save the file" or "copy the code into a file" if you already created or modified the file using \`apply_patch\`. Instead, reference the file as already saved.
+  - Do NOT tell the user to "save the file" or "copy the code into a file" if you already created or modified the file using the `apply_patch` shell command. Instead, reference the file as already saved.
   - Do NOT show the full contents of large files you have already written, unless the user explicitly asks for them.
 
-§ `apply-patch` Specification
+## Using the shell command `apply_patch` to edit files
+`apply_patch` is a shell command for editing files. Your patch language is a stripped‑down, file‑oriented diff format designed to be easy to parse and safe to apply. You can think of it as a high‑level envelope:
 
-Your patch language is a stripped‑down, file‑oriented diff format designed to be easy to parse and safe to apply. You can think of it as a high‑level envelope:
-
-**_ Begin Patch
+*** Begin Patch
 [ one or more file sections ]
-_** End Patch
+*** End Patch
 
 Within that envelope, you get a sequence of file operations.
 You MUST include a header to specify the action you are taking.
 Each operation starts with one of three headers:
 
-**_ Add File: <path> - create a new file. Every following line is a + line (the initial contents).
-_** Delete File: <path> - remove an existing file. Nothing follows.
+*** Add File: <path> - create a new file. Every following line is a + line (the initial contents).
+*** Delete File: <path> - remove an existing file. Nothing follows.
 \*\*\* Update File: <path> - patch an existing file in place (optionally with a rename).
 
 May be immediately followed by \*\*\* Move to: <new path> if you want to rename the file.
@@ -63,36 +77,60 @@ Within a hunk each line starts with:
   At the end of a truncated hunk you can emit \*\*\* End of File.
 
 Patch := Begin { FileOp } End
-Begin := "**_ Begin Patch" NEWLINE
-End := "_** End Patch" NEWLINE
+Begin := "*** Begin Patch" NEWLINE
+End := "*** End Patch" NEWLINE
 FileOp := AddFile | DeleteFile | UpdateFile
-AddFile := "**_ Add File: " path NEWLINE { "+" line NEWLINE }
-DeleteFile := "_** Delete File: " path NEWLINE
-UpdateFile := "**_ Update File: " path NEWLINE [ MoveTo ] { Hunk }
-MoveTo := "_** Move to: " newPath NEWLINE
+AddFile := "*** Add File: " path NEWLINE { "+" line NEWLINE }
+DeleteFile := "*** Delete File: " path NEWLINE
+UpdateFile := "*** Update File: " path NEWLINE [ MoveTo ] { Hunk }
+MoveTo := "*** Move to: " newPath NEWLINE
 Hunk := "@@" [ header ] NEWLINE { HunkLine } [ "*** End of File" NEWLINE ]
 HunkLine := (" " | "-" | "+") text NEWLINE
 
 A full patch can combine several operations:
 
-**_ Begin Patch
-_** Add File: hello.txt
+*** Begin Patch
+*** Add File: hello.txt
 +Hello world
-**_ Update File: src/app.py
-_** Move to: src/main.py
+*** Update File: src/app.py
+*** Move to: src/main.py
 @@ def greet():
 -print("Hi")
 +print("Hello, world!")
-**_ Delete File: obsolete.txt
-_** End Patch
+*** Delete File: obsolete.txt
+*** End Patch
 
 It is important to remember:
 
 - You must include a header with your intended action (Add/Delete/Update)
 - You must prefix new lines with `+` even when creating a new file
+- You must follow this schema exactly when providing a patch
 
-You can invoke apply_patch like:
+You can invoke apply_patch with the following shell command:
 
 ```
 shell {"command":["apply_patch","*** Begin Patch\n*** Add File: hello.txt\n+Hello, world!\n*** End Patch\n"]}
 ```
+
+## Sandbox environment and approval instructions
+
+You are running in a sandboxed workspace backed by version control. The sandbox might be configured by the user to restrict certain behaviors, like accessing the internet or writing to files outside the current directory.
+
+Commands that are blocked by sandbox settings will be automatically sent to the user for approval. The result of the request will be returned (i.e. the command result, or the request denial).
+The user also has an opportunity to approve the same command for the rest of the session.
+
+Guidance on running within the sandbox:
+- When running commands that will likely require approval, attempt to use simple, precise commands, to reduce frequency of approval requests.
+- When approval is denied or a command fails due to a permission error, do not retry the exact command in a different way. Move on and continue trying to address the user's request.
+
+
+## Tools available
+### Plan updates
+
+A tool named `update_plan` is available. Use it to keep an up‑to‑date, step‑by‑step plan for the task so you can follow your progress. When making your plans, keep in mind that you are a deployed coding agent - `update_plan` calls should not involve doing anything that you aren't capable of doing. For example, `update_plan` calls should NEVER contain tasks to merge your own pull requests. Only stop to ask the user if you genuinely need their feedback on a change.
+
+- At the start of any nontrivial task, call `update_plan` with an initial plan: a short list of 1‑sentence steps with a `status` for each step (`pending`, `in_progress`, or `completed`). There should always be exactly one `in_progress` step until everything is done.
+- Whenever you finish a step, call `update_plan` again, marking the finished step as `completed` and the next step as `in_progress`.
+- If your plan needs to change, call `update_plan` with the revised steps and include an `explanation` describing the change.
+- When all steps are complete, make a final `update_plan` call with all steps marked `completed`.
+

codex-rs/core/src/apply_patch.rs

@@ -1,277 +1,107 @@
 use crate::codex::Session;
 use crate::models::FunctionCallOutputPayload;
 use crate::models::ResponseInputItem;
-use crate::protocol::Event;
-use crate::protocol::EventMsg;
 use crate::protocol::FileChange;
-use crate::protocol::PatchApplyBeginEvent;
-use crate::protocol::PatchApplyEndEvent;
 use crate::protocol::ReviewDecision;
 use crate::safety::SafetyCheck;
 use crate::safety::assess_patch_safety;
-use anyhow::Context;
-use codex_apply_patch::AffectedPaths;
 use codex_apply_patch::ApplyPatchAction;
 use codex_apply_patch::ApplyPatchFileChange;
-use codex_apply_patch::print_summary;
 use std::collections::HashMap;
 use std::path::Path;
 use std::path::PathBuf;
 
+pub const CODEX_APPLY_PATCH_ARG1: &str = "--codex-run-as-apply-patch";
+
+pub(crate) enum InternalApplyPatchInvocation {
+    /// The `apply_patch` call was handled programmatically, without any sort
+    /// of sandbox, because the user explicitly approved it. This is the
+    /// result to use with the `shell` function call that contained `apply_patch`.
+    Output(ResponseInputItem),
+
+    /// The `apply_patch` call was approved, either automatically because it
+    /// appears that it should be allowed based on the user's sandbox policy
+    /// *or* because the user explicitly approved it. In either case, we use
+    /// exec with [`CODEX_APPLY_PATCH_ARG1`] to realize the `apply_patch` call,
+    /// but [`ApplyPatchExec::auto_approved`] is used to determine the sandbox
+    /// used with the `exec()`.
+    DelegateToExec(ApplyPatchExec),
+}
+
+pub(crate) struct ApplyPatchExec {
+    pub(crate) action: ApplyPatchAction,
+    pub(crate) user_explicitly_approved_this_action: bool,
+}
+
+impl From<ResponseInputItem> for InternalApplyPatchInvocation {
+    fn from(item: ResponseInputItem) -> Self {
+        InternalApplyPatchInvocation::Output(item)
+    }
+}
+
 pub(crate) async fn apply_patch(
     sess: &Session,
-    sub_id: String,
-    call_id: String,
+    sub_id: &str,
+    call_id: &str,
     action: ApplyPatchAction,
-) -> ResponseInputItem {
+) -> InternalApplyPatchInvocation {
     let writable_roots_snapshot = {
         #[allow(clippy::unwrap_used)]
         let guard = sess.writable_roots.lock().unwrap();
         guard.clone()
     };
 
-    let auto_approved = match assess_patch_safety(
+    match assess_patch_safety(
         &action,
         sess.approval_policy,
         &writable_roots_snapshot,
         &sess.cwd,
     ) {
-        SafetyCheck::AutoApprove { .. } => true,
+        SafetyCheck::AutoApprove { .. } => {
+            InternalApplyPatchInvocation::DelegateToExec(ApplyPatchExec {
+                action,
+                user_explicitly_approved_this_action: false,
+            })
+        }
         SafetyCheck::AskUser => {
             // Compute a readable summary of path changes to include in the
             // approval request so the user can make an informed decision.
+            //
+            // Note that it might be worth expanding this approval request to
+            // give the user the option to expand the set of writable roots so
+            // that similar patches can be auto-approved in the future during
+            // this session.
             let rx_approve = sess
-                .request_patch_approval(sub_id.clone(), call_id.clone(), &action, None, None)
+                .request_patch_approval(sub_id.to_owned(), call_id.to_owned(), &action, None, None)
                 .await;
             match rx_approve.await.unwrap_or_default() {
-                ReviewDecision::Approved | ReviewDecision::ApprovedForSession => false,
+                ReviewDecision::Approved | ReviewDecision::ApprovedForSession => {
+                    InternalApplyPatchInvocation::DelegateToExec(ApplyPatchExec {
+                        action,
+                        user_explicitly_approved_this_action: true,
+                    })
+                }
                 ReviewDecision::Denied | ReviewDecision::Abort => {
-                    return ResponseInputItem::FunctionCallOutput {
-                        call_id,
+                    ResponseInputItem::FunctionCallOutput {
+                        call_id: call_id.to_owned(),
                         output: FunctionCallOutputPayload {
                             content: "patch rejected by user".to_string(),
                             success: Some(false),
                         },
-                    };
+                    }
+                    .into()
                 }
             }
         }
-        SafetyCheck::Reject { reason } => {
-            return ResponseInputItem::FunctionCallOutput {
-                call_id,
-                output: FunctionCallOutputPayload {
-                    content: format!("patch rejected: {reason}"),
-                    success: Some(false),
-                },
-            };
-        }
-    };
-
-    // Verify write permissions before touching the filesystem.
-    let writable_snapshot = {
-        #[allow(clippy::unwrap_used)]
-        sess.writable_roots.lock().unwrap().clone()
-    };
-
-    if let Some(offending) = first_offending_path(&action, &writable_snapshot, &sess.cwd) {
-        let root = offending.parent().unwrap_or(&offending).to_path_buf();
-
-        let reason = Some(format!(
-            "grant write access to {} for this session",
-            root.display()
-        ));
-
-        let rx = sess
-            .request_patch_approval(
-                sub_id.clone(),
-                call_id.clone(),
-                &action,
-                reason.clone(),
-                Some(root.clone()),
-            )
-            .await;
-
-        if !matches!(
-            rx.await.unwrap_or_default(),
-            ReviewDecision::Approved | ReviewDecision::ApprovedForSession
-        ) {
-            return ResponseInputItem::FunctionCallOutput {
-                call_id,
-                output: FunctionCallOutputPayload {
-                    content: "patch rejected by user".to_string(),
-                    success: Some(false),
-                },
-            };
-        }
-
-        // user approved, extend writable roots for this session
-        #[allow(clippy::unwrap_used)]
-        sess.writable_roots.lock().unwrap().push(root);
-    }
-
-    let _ = sess
-        .tx_event
-        .send(Event {
-            id: sub_id.clone(),
-            msg: EventMsg::PatchApplyBegin(PatchApplyBeginEvent {
-                call_id: call_id.clone(),
-                auto_approved,
-                changes: convert_apply_patch_to_protocol(&action),
-            }),
-        })
-        .await;
-
-    let mut stdout = Vec::new();
-    let mut stderr = Vec::new();
-    // Enforce writable roots. If a write is blocked, collect offending root
-    // and prompt the user to extend permissions.
-    let mut result = apply_changes_from_apply_patch_and_report(&action, &mut stdout, &mut stderr);
-
-    if let Err(err) = &result {
-        if err.kind() == std::io::ErrorKind::PermissionDenied {
-            // Determine first offending path.
-            let offending_opt = action
-                .changes()
-                .iter()
-                .flat_map(|(path, change)| match change {
-                    ApplyPatchFileChange::Add { .. } => vec![path.as_ref()],
-                    ApplyPatchFileChange::Delete => vec![path.as_ref()],
-                    ApplyPatchFileChange::Update {
-                        move_path: Some(move_path),
-                        ..
-                    } => {
-                        vec![path.as_ref(), move_path.as_ref()]
-                    }
-                    ApplyPatchFileChange::Update {
-                        move_path: None, ..
-                    } => vec![path.as_ref()],
-                })
-                .find_map(|path: &Path| {
-                    // ApplyPatchAction promises to guarantee absolute paths.
-                    if !path.is_absolute() {
-                        panic!("apply_patch invariant failed: path is not absolute: {path:?}");
-                    }
-
-                    let writable = {
-                        #[allow(clippy::unwrap_used)]
-                        let roots = sess.writable_roots.lock().unwrap();
-                        roots.iter().any(|root| path.starts_with(root))
-                    };
-                    if writable {
-                        None
-                    } else {
-                        Some(path.to_path_buf())
-                    }
-                });
-
-            if let Some(offending) = offending_opt {
-                let root = offending.parent().unwrap_or(&offending).to_path_buf();
-
-                let reason = Some(format!(
-                    "grant write access to {} for this session",
-                    root.display()
-                ));
-                let rx = sess
-                    .request_patch_approval(
-                        sub_id.clone(),
-                        call_id.clone(),
-                        &action,
-                        reason.clone(),
-                        Some(root.clone()),
-                    )
-                    .await;
-                if matches!(
-                    rx.await.unwrap_or_default(),
-                    ReviewDecision::Approved | ReviewDecision::ApprovedForSession
-                ) {
-                    // Extend writable roots.
-                    #[allow(clippy::unwrap_used)]
-                    sess.writable_roots.lock().unwrap().push(root);
-                    stdout.clear();
-                    stderr.clear();
-                    result = apply_changes_from_apply_patch_and_report(
-                        &action,
-                        &mut stdout,
-                        &mut stderr,
-                    );
-                }
-            }
-        }
-    }
-
-    // Emit PatchApplyEnd event.
-    let success_flag = result.is_ok();
-    let _ = sess
-        .tx_event
-        .send(Event {
-            id: sub_id.clone(),
-            msg: EventMsg::PatchApplyEnd(PatchApplyEndEvent {
-                call_id: call_id.clone(),
-                stdout: String::from_utf8_lossy(&stdout).to_string(),
-                stderr: String::from_utf8_lossy(&stderr).to_string(),
-                success: success_flag,
-            }),
-        })
-        .await;
-
-    match result {
-        Ok(_) => ResponseInputItem::FunctionCallOutput {
-            call_id,
+        SafetyCheck::Reject { reason } => ResponseInputItem::FunctionCallOutput {
+            call_id: call_id.to_owned(),
             output: FunctionCallOutputPayload {
-                content: String::from_utf8_lossy(&stdout).to_string(),
-                success: None,
-            },
-        },
-        Err(e) => ResponseInputItem::FunctionCallOutput {
-            call_id,
-            output: FunctionCallOutputPayload {
-                content: format!("error: {e:#}, stderr: {}", String::from_utf8_lossy(&stderr)),
+                content: format!("patch rejected: {reason}"),
                 success: Some(false),
             },
-        },
-    }
-}
-
-/// Return the first path in `hunks` that is NOT under any of the
-/// `writable_roots` (after normalising). If all paths are acceptable,
-/// returns None.
-fn first_offending_path(
-    action: &ApplyPatchAction,
-    writable_roots: &[PathBuf],
-    cwd: &Path,
-) -> Option<PathBuf> {
-    let changes = action.changes();
-    for (path, change) in changes {
-        let candidate = match change {
-            ApplyPatchFileChange::Add { .. } => path,
-            ApplyPatchFileChange::Delete => path,
-            ApplyPatchFileChange::Update { move_path, .. } => move_path.as_ref().unwrap_or(path),
-        };
-
-        let abs = if candidate.is_absolute() {
-            candidate.clone()
-        } else {
-            cwd.join(candidate)
-        };
-
-        let mut allowed = false;
-        for root in writable_roots {
-            let root_abs = if root.is_absolute() {
-                root.clone()
-            } else {
-                cwd.join(root)
-            };
-            if abs.starts_with(&root_abs) {
-                allowed = true;
-                break;
-            }
-        }
-
-        if !allowed {
-            return Some(candidate.clone());
         }
+        .into(),
     }
-    None
 }
 
 pub(crate) fn convert_apply_patch_to_protocol(
@@ -299,85 +129,6 @@ pub(crate) fn convert_apply_patch_to_protocol(
     result
 }
 
-fn apply_changes_from_apply_patch_and_report(
-    action: &ApplyPatchAction,
-    stdout: &mut impl std::io::Write,
-    stderr: &mut impl std::io::Write,
-) -> std::io::Result<()> {
-    match apply_changes_from_apply_patch(action) {
-        Ok(affected_paths) => {
-            print_summary(&affected_paths, stdout)?;
-        }
-        Err(err) => {
-            writeln!(stderr, "{err:?}")?;
-        }
-    }
-
-    Ok(())
-}
-
-fn apply_changes_from_apply_patch(action: &ApplyPatchAction) -> anyhow::Result<AffectedPaths> {
-    let mut added: Vec<PathBuf> = Vec::new();
-    let mut modified: Vec<PathBuf> = Vec::new();
-    let mut deleted: Vec<PathBuf> = Vec::new();
-
-    let changes = action.changes();
-    for (path, change) in changes {
-        match change {
-            ApplyPatchFileChange::Add { content } => {
-                if let Some(parent) = path.parent() {
-                    if !parent.as_os_str().is_empty() {
-                        std::fs::create_dir_all(parent).with_context(|| {
-                            format!("Failed to create parent directories for {}", path.display())
-                        })?;
-                    }
-                }
-                std::fs::write(path, content)
-                    .with_context(|| format!("Failed to write file {}", path.display()))?;
-                added.push(path.clone());
-            }
-            ApplyPatchFileChange::Delete => {
-                std::fs::remove_file(path)
-                    .with_context(|| format!("Failed to delete file {}", path.display()))?;
-                deleted.push(path.clone());
-            }
-            ApplyPatchFileChange::Update {
-                unified_diff: _unified_diff,
-                move_path,
-                new_content,
-            } => {
-                if let Some(move_path) = move_path {
-                    if let Some(parent) = move_path.parent() {
-                        if !parent.as_os_str().is_empty() {
-                            std::fs::create_dir_all(parent).with_context(|| {
-                                format!(
-                                    "Failed to create parent directories for {}",
-                                    move_path.display()
-                                )
-                            })?;
-                        }
-                    }
-
-                    std::fs::rename(path, move_path)
-                        .with_context(|| format!("Failed to rename file {}", path.display()))?;
-                    std::fs::write(move_path, new_content)?;
-                    modified.push(move_path.clone());
-                    deleted.push(path.clone());
-                } else {
-                    std::fs::write(path, new_content)?;
-                    modified.push(path.clone());
-                }
-            }
-        }
-    }
-
-    Ok(AffectedPaths {
-        added,
-        modified,
-        deleted,
-    })
-}
-
 pub(crate) fn get_writable_roots(cwd: &Path) -> Vec<PathBuf> {
     let mut writable_roots = Vec::new();
     if cfg!(target_os = "macos") {

codex-rs/core/src/chat_completions.rs

@@ -21,7 +21,9 @@ 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::models::ContentItem;
+use crate::models::ReasoningItemContent;
 use crate::models::ResponseItem;
 use crate::openai_tools::create_tools_json_for_chat_completions_api;
 use crate::util::backoff;
@@ -29,17 +31,19 @@ use crate::util::backoff;
 /// Implementation for the classic Chat Completions API.
 pub(crate) async fn stream_chat_completions(
     prompt: &Prompt,
-    model: &str,
+    model_family: &ModelFamily,
     client: &reqwest::Client,
     provider: &ModelProviderInfo,
 ) -> Result<ResponseStream> {
     // Build messages array
     let mut messages = Vec::<serde_json::Value>::new();
 
-    let full_instructions = prompt.get_full_instructions(model);
+    let full_instructions = prompt.get_full_instructions(model_family);
     messages.push(json!({"role": "system", "content": full_instructions}));
 
-    for item in &prompt.input {
+    let input = prompt.get_formatted_input();
+
+    for item in &input {
         match item {
             ResponseItem::Message { role, content, .. } => {
                 let mut text = String::new();
@@ -105,9 +109,9 @@ pub(crate) async fn stream_chat_completions(
         }
     }
 
-    let tools_json = create_tools_json_for_chat_completions_api(prompt, model)?;
+    let tools_json = create_tools_json_for_chat_completions_api(&prompt.tools)?;
     let payload = json!({
-        "model": model,
+        "model": model_family.slug,
         "messages": messages,
         "stream": true,
         "tools": tools_json,
@@ -115,7 +119,7 @@ pub(crate) async fn stream_chat_completions(
 
     debug!(
         "POST to {}: {}",
-        provider.get_full_url(),
+        provider.get_full_url(&None),
         serde_json::to_string_pretty(&payload).unwrap_or_default()
     );
 
@@ -124,7 +128,7 @@ pub(crate) async fn stream_chat_completions(
     loop {
         attempt += 1;
 
-        let req_builder = provider.create_request_builder(client)?;
+        let req_builder = provider.create_request_builder(client, &None).await?;
 
         let res = req_builder
             .header(reqwest::header::ACCEPT, "text/event-stream")
@@ -202,6 +206,8 @@ async fn process_chat_sse<S>(
     }
 
     let mut fn_call_state = FunctionCallState::default();
+    let mut assistant_text = String::new();
+    let mut reasoning_text = String::new();
 
     loop {
         let sse = match timeout(idle_timeout, stream.next()).await {
@@ -230,6 +236,31 @@ async fn process_chat_sse<S>(
 
         // OpenAI Chat streaming sends a literal string "[DONE]" when finished.
         if sse.data.trim() == "[DONE]" {
+            // Emit any finalized items before closing so downstream consumers receive
+            // terminal events for both assistant content and raw reasoning.
+            if !assistant_text.is_empty() {
+                let item = ResponseItem::Message {
+                    role: "assistant".to_string(),
+                    content: vec![ContentItem::OutputText {
+                        text: std::mem::take(&mut assistant_text),
+                    }],
+                    id: None,
+                };
+                let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+            }
+
+            if !reasoning_text.is_empty() {
+                let item = ResponseItem::Reasoning {
+                    id: String::new(),
+                    summary: Vec::new(),
+                    content: Some(vec![ReasoningItemContent::ReasoningText {
+                        text: std::mem::take(&mut reasoning_text),
+                    }]),
+                    encrypted_content: None,
+                };
+                let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+            }
+
             let _ = tx_event
                 .send(Ok(ResponseEvent::Completed {
                     response_id: String::new(),
@@ -249,21 +280,47 @@ async fn process_chat_sse<S>(
         let choice_opt = chunk.get("choices").and_then(|c| c.get(0));
 
         if let Some(choice) = choice_opt {
-            // Handle assistant content tokens.
+            // Handle assistant content tokens as streaming deltas.
             if let Some(content) = choice
                 .get("delta")
                 .and_then(|d| d.get("content"))
                 .and_then(|c| c.as_str())
             {
-                let item = ResponseItem::Message {
-                    role: "assistant".to_string(),
-                    content: vec![ContentItem::OutputText {
-                        text: content.to_string(),
-                    }],
-                    id: None,
-                };
+                if !content.is_empty() {
+                    assistant_text.push_str(content);
+                    let _ = tx_event
+                        .send(Ok(ResponseEvent::OutputTextDelta(content.to_string())))
+                        .await;
+                }
+            }
 
-                let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+            // Forward any reasoning/thinking deltas if present.
+            // Some providers stream `reasoning` as a plain string while others
+            // nest the text under an object (e.g. `{ "reasoning": { "text": "…" } }`).
+            if let Some(reasoning_val) = choice.get("delta").and_then(|d| d.get("reasoning")) {
+                let mut maybe_text = reasoning_val.as_str().map(|s| s.to_string());
+
+                if maybe_text.is_none() && reasoning_val.is_object() {
+                    if let Some(s) = reasoning_val
+                        .get("text")
+                        .and_then(|t| t.as_str())
+                        .filter(|s| !s.is_empty())
+                    {
+                        maybe_text = Some(s.to_string());
+                    } else if let Some(s) = reasoning_val
+                        .get("content")
+                        .and_then(|t| t.as_str())
+                        .filter(|s| !s.is_empty())
+                    {
+                        maybe_text = Some(s.to_string());
+                    }
+                }
+
+                if let Some(reasoning) = maybe_text {
+                    let _ = tx_event
+                        .send(Ok(ResponseEvent::ReasoningContentDelta(reasoning)))
+                        .await;
+                }
             }
 
             // Handle streaming function / tool calls.
@@ -300,7 +357,21 @@ async fn process_chat_sse<S>(
             if let Some(finish_reason) = choice.get("finish_reason").and_then(|v| v.as_str()) {
                 match finish_reason {
                     "tool_calls" if fn_call_state.active => {
-                        // Build the FunctionCall response item.
+                        // First, flush the terminal raw reasoning so UIs can finalize
+                        // the reasoning stream before any exec/tool events begin.
+                        if !reasoning_text.is_empty() {
+                            let item = ResponseItem::Reasoning {
+                                id: String::new(),
+                                summary: Vec::new(),
+                                content: Some(vec![ReasoningItemContent::ReasoningText {
+                                    text: std::mem::take(&mut reasoning_text),
+                                }]),
+                                encrypted_content: None,
+                            };
+                            let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+                        }
+
+                        // Then emit the FunctionCall response item.
                         let item = ResponseItem::FunctionCall {
                             id: None,
                             name: fn_call_state.name.clone().unwrap_or_else(|| "".to_string()),
@@ -308,11 +379,33 @@ async fn process_chat_sse<S>(
                             call_id: fn_call_state.call_id.clone().unwrap_or_else(String::new),
                         };
 
-                        // Emit it downstream.
                         let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
                     }
                     "stop" => {
-                        // Regular turn without tool-call.
+                        // Regular turn without tool-call. Emit the final assistant message
+                        // as a single OutputItemDone so non-delta consumers see the result.
+                        if !assistant_text.is_empty() {
+                            let item = ResponseItem::Message {
+                                role: "assistant".to_string(),
+                                content: vec![ContentItem::OutputText {
+                                    text: std::mem::take(&mut assistant_text),
+                                }],
+                                id: None,
+                            };
+                            let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+                        }
+                        // Also emit a terminal Reasoning item so UIs can finalize raw reasoning.
+                        if !reasoning_text.is_empty() {
+                            let item = ResponseItem::Reasoning {
+                                id: String::new(),
+                                summary: Vec::new(),
+                                content: Some(vec![ReasoningItemContent::ReasoningText {
+                                    text: std::mem::take(&mut reasoning_text),
+                                }]),
+                                encrypted_content: None,
+                            };
+                            let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+                        }
                     }
                     _ => {}
                 }
@@ -350,10 +443,17 @@ async fn process_chat_sse<S>(
 /// The adapter is intentionally *lossless*: callers who do **not** opt in via
 /// [`AggregateStreamExt::aggregate()`] keep receiving the original unmodified
 /// events.
+#[derive(Copy, Clone, Eq, PartialEq)]
+enum AggregateMode {
+    AggregatedOnly,
+    Streaming,
+}
 pub(crate) struct AggregatedChatStream<S> {
     inner: S,
     cumulative: String,
-    pending_completed: Option<ResponseEvent>,
+    cumulative_reasoning: String,
+    pending: std::collections::VecDeque<ResponseEvent>,
+    mode: AggregateMode,
 }
 
 impl<S> Stream for AggregatedChatStream<S>
@@ -365,8 +465,8 @@ where
     fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
         let this = self.get_mut();
 
-        // First, flush any buffered Completed event from the previous call.
-        if let Some(ev) = this.pending_completed.take() {
+        // First, flush any buffered events from the previous call.
+        if let Some(ev) = this.pending.pop_front() {
             return Poll::Ready(Some(Ok(ev)));
         }
 
@@ -383,16 +483,21 @@ where
                     let is_assistant_delta = matches!(&item, crate::models::ResponseItem::Message { role, .. } if role == "assistant");
 
                     if is_assistant_delta {
-                        if let crate::models::ResponseItem::Message { content, .. } = &item {
-                            if let Some(text) = content.iter().find_map(|c| match c {
-                                crate::models::ContentItem::OutputText { text } => Some(text),
-                                _ => None,
-                            }) {
-                                this.cumulative.push_str(text);
+                        // Only use the final assistant message if we have not
+                        // seen any deltas; otherwise, deltas already built the
+                        // cumulative text and this would duplicate it.
+                        if this.cumulative.is_empty() {
+                            if let crate::models::ResponseItem::Message { content, .. } = &item {
+                                if let Some(text) = content.iter().find_map(|c| match c {
+                                    crate::models::ContentItem::OutputText { text } => Some(text),
+                                    _ => None,
+                                }) {
+                                    this.cumulative.push_str(text);
+                                }
                             }
                         }
 
-                        // Swallow partial assistant chunk; keep polling.
+                        // Swallow assistant message here; emit on Completed.
                         continue;
                     }
 
@@ -403,24 +508,50 @@ where
                     response_id,
                     token_usage,
                 }))) => {
+                    // Build any aggregated items in the correct order: Reasoning first, then Message.
+                    let mut emitted_any = false;
+
+                    if !this.cumulative_reasoning.is_empty()
+                        && matches!(this.mode, AggregateMode::AggregatedOnly)
+                    {
+                        let aggregated_reasoning = crate::models::ResponseItem::Reasoning {
+                            id: String::new(),
+                            summary: Vec::new(),
+                            content: Some(vec![
+                                crate::models::ReasoningItemContent::ReasoningText {
+                                    text: std::mem::take(&mut this.cumulative_reasoning),
+                                },
+                            ]),
+                            encrypted_content: None,
+                        };
+                        this.pending
+                            .push_back(ResponseEvent::OutputItemDone(aggregated_reasoning));
+                        emitted_any = true;
+                    }
+
                     if !this.cumulative.is_empty() {
-                        let aggregated_item = crate::models::ResponseItem::Message {
+                        let aggregated_message = crate::models::ResponseItem::Message {
                             id: None,
                             role: "assistant".to_string(),
                             content: vec![crate::models::ContentItem::OutputText {
                                 text: std::mem::take(&mut this.cumulative),
                             }],
                         };
+                        this.pending
+                            .push_back(ResponseEvent::OutputItemDone(aggregated_message));
+                        emitted_any = true;
+                    }
 
-                        // Buffer Completed so it is returned *after* the aggregated message.
-                        this.pending_completed = Some(ResponseEvent::Completed {
-                            response_id,
-                            token_usage,
+                    // Always emit Completed last when anything was aggregated.
+                    if emitted_any {
+                        this.pending.push_back(ResponseEvent::Completed {
+                            response_id: response_id.clone(),
+                            token_usage: token_usage.clone(),
                         });
-
-                        return Poll::Ready(Some(Ok(ResponseEvent::OutputItemDone(
-                            aggregated_item,
-                        ))));
+                        // Return the first pending event now.
+                        if let Some(ev) = this.pending.pop_front() {
+                            return Poll::Ready(Some(Ok(ev)));
+                        }
                     }
 
                     // Nothing aggregated – forward Completed directly.
@@ -434,10 +565,27 @@ where
                     // will never appear in a Chat Completions stream.
                     continue;
                 }
-                Poll::Ready(Some(Ok(ResponseEvent::OutputTextDelta(_))))
-                | Poll::Ready(Some(Ok(ResponseEvent::ReasoningSummaryDelta(_)))) => {
-                    // Deltas are ignored here since aggregation waits for the
-                    // final OutputItemDone.
+                Poll::Ready(Some(Ok(ResponseEvent::OutputTextDelta(delta)))) => {
+                    // Always accumulate deltas so we can emit a final OutputItemDone at Completed.
+                    this.cumulative.push_str(&delta);
+                    if matches!(this.mode, AggregateMode::Streaming) {
+                        // In streaming mode, also forward the delta immediately.
+                        return Poll::Ready(Some(Ok(ResponseEvent::OutputTextDelta(delta))));
+                    } else {
+                        continue;
+                    }
+                }
+                Poll::Ready(Some(Ok(ResponseEvent::ReasoningContentDelta(delta)))) => {
+                    // Always accumulate reasoning deltas so we can emit a final Reasoning item at Completed.
+                    this.cumulative_reasoning.push_str(&delta);
+                    if matches!(this.mode, AggregateMode::Streaming) {
+                        // In streaming mode, also forward the delta immediately.
+                        return Poll::Ready(Some(Ok(ResponseEvent::ReasoningContentDelta(delta))));
+                    } else {
+                        continue;
+                    }
+                }
+                Poll::Ready(Some(Ok(ResponseEvent::ReasoningSummaryDelta(_)))) => {
                     continue;
                 }
             }
@@ -467,12 +615,24 @@ pub(crate) trait AggregateStreamExt: Stream<Item = Result<ResponseEvent>> + Size
     /// }
     /// ```
     fn aggregate(self) -> AggregatedChatStream<Self> {
-        AggregatedChatStream {
-            inner: self,
-            cumulative: String::new(),
-            pending_completed: None,
-        }
+        AggregatedChatStream::new(self, AggregateMode::AggregatedOnly)
     }
 }
 
 impl<T> AggregateStreamExt for T where T: Stream<Item = Result<ResponseEvent>> + Sized {}
+
+impl<S> AggregatedChatStream<S> {
+    fn new(inner: S, mode: AggregateMode) -> Self {
+        AggregatedChatStream {
+            inner,
+            cumulative: String::new(),
+            cumulative_reasoning: String::new(),
+            pending: std::collections::VecDeque::new(),
+            mode,
+        }
+    }
+
+    pub(crate) fn streaming_mode(inner: S) -> Self {
+        Self::new(inner, AggregateMode::Streaming)
+    }
+}

codex-rs/core/src/client.rs

@@ -3,6 +3,8 @@ use std::path::Path;
 use std::time::Duration;
 
 use bytes::Bytes;
+use codex_login::AuthMode;
+use codex_login::CodexAuth;
 use eventsource_stream::Eventsource;
 use futures::prelude::*;
 use reqwest::StatusCode;
@@ -41,6 +43,7 @@ use std::sync::Arc;
 #[derive(Clone)]
 pub struct ModelClient {
     config: Arc<Config>,
+    auth: Option<CodexAuth>,
     client: reqwest::Client,
     provider: ModelProviderInfo,
     session_id: Uuid,
@@ -51,6 +54,7 @@ pub struct ModelClient {
 impl ModelClient {
     pub fn new(
         config: Arc<Config>,
+        auth: Option<CodexAuth>,
         provider: ModelProviderInfo,
         effort: ReasoningEffortConfig,
         summary: ReasoningSummaryConfig,
@@ -58,6 +62,7 @@ impl ModelClient {
     ) -> Self {
         Self {
             config,
+            auth,
             client: reqwest::Client::new(),
             provider,
             session_id,
@@ -76,7 +81,7 @@ impl ModelClient {
                 // Create the raw streaming connection first.
                 let response_stream = stream_chat_completions(
                     prompt,
-                    &self.config.model,
+                    &self.config.model_family,
                     &self.client,
                     &self.provider,
                 )
@@ -85,7 +90,11 @@ impl ModelClient {
                 // Wrap it with the aggregation adapter so callers see *only*
                 // the final assistant message per turn (matching the
                 // behaviour of the Responses API).
-                let mut aggregated = response_stream.aggregate();
+                let mut aggregated = if self.config.show_raw_agent_reasoning {
+                    crate::chat_completions::AggregatedChatStream::streaming_mode(response_stream)
+                } else {
+                    response_stream.aggregate()
+                };
 
                 // Bridge the aggregated stream back into a standard
                 // `ResponseStream` by forwarding events through a channel.
@@ -114,51 +123,80 @@ impl ModelClient {
             return stream_from_fixture(path, self.provider.clone()).await;
         }
 
-        let full_instructions = prompt.get_full_instructions(&self.config.model);
-        let tools_json = create_tools_json_for_responses_api(prompt, &self.config.model)?;
-        let reasoning = create_reasoning_param_for_request(&self.config, self.effort, self.summary);
+        let auth = self.auth.clone();
+
+        let auth_mode = auth.as_ref().map(|a| a.mode);
+
+        let store = prompt.store && auth_mode != Some(AuthMode::ChatGPT);
+
+        let full_instructions = prompt.get_full_instructions(&self.config.model_family);
+        let tools_json = create_tools_json_for_responses_api(&prompt.tools)?;
+        let reasoning = create_reasoning_param_for_request(
+            &self.config.model_family,
+            self.effort,
+            self.summary,
+        );
 
         // Request encrypted COT if we are not storing responses,
         // otherwise reasoning items will be referenced by ID
-        let include = if !prompt.store && reasoning.is_some() {
+        let include: Vec<String> = if !store && reasoning.is_some() {
             vec!["reasoning.encrypted_content".to_string()]
         } else {
             vec![]
         };
 
+        let input_with_instructions = prompt.get_formatted_input();
+
         let payload = ResponsesApiRequest {
             model: &self.config.model,
             instructions: &full_instructions,
-            input: &prompt.input,
+            input: &input_with_instructions,
             tools: &tools_json,
             tool_choice: "auto",
             parallel_tool_calls: false,
             reasoning,
-            store: prompt.store,
-            // TODO: make this configurable
+            store,
             stream: true,
             include,
         };
 
+        let mut attempt = 0;
+        let max_retries = self.provider.request_max_retries();
+
         trace!(
             "POST to {}: {}",
-            self.provider.get_full_url(),
+            self.provider.get_full_url(&auth),
             serde_json::to_string(&payload)?
         );
 
-        let mut attempt = 0;
-        let max_retries = self.provider.request_max_retries();
         loop {
             attempt += 1;
 
-            let req_builder = self
+            let mut req_builder = self
                 .provider
-                .create_request_builder(&self.client)?
+                .create_request_builder(&self.client, &auth)
+                .await?;
+
+            req_builder = req_builder
                 .header("OpenAI-Beta", "responses=experimental")
                 .header("session_id", self.session_id.to_string())
                 .header(reqwest::header::ACCEPT, "text/event-stream")
                 .json(&payload);
 
+            if let Some(auth) = auth.as_ref()
+                && auth.mode == AuthMode::ChatGPT
+                && let Some(account_id) = auth.get_account_id().await
+            {
+                req_builder = req_builder.header("chatgpt-account-id", account_id);
+            }
+
+            let originator = self
+                .config
+                .internal_originator
+                .as_deref()
+                .unwrap_or("codex_cli_rs");
+            req_builder = req_builder.header("originator", originator);
+
             let res = req_builder.send().await;
             if let Ok(resp) = &res {
                 trace!(
@@ -389,6 +427,14 @@ async fn process_sse<S>(
                     }
                 }
             }
+            "response.reasoning_text.delta" => {
+                if let Some(delta) = event.delta {
+                    let event = ResponseEvent::ReasoningContentDelta(delta);
+                    if tx_event.send(Ok(event)).await.is_err() {
+                        return;
+                    }
+                }
+            }
             "response.created" => {
                 if event.response.is_some() {
                     let _ = tx_event.send(Ok(ResponseEvent::Created {})).await;
@@ -567,7 +613,7 @@ mod tests {
 
         let provider = ModelProviderInfo {
             name: "test".to_string(),
-            base_url: "https://test.com".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,
@@ -577,6 +623,7 @@ mod tests {
             request_max_retries: Some(0),
             stream_max_retries: Some(0),
             stream_idle_timeout_ms: Some(1000),
+            requires_openai_auth: false,
         };
 
         let events = collect_events(
@@ -626,7 +673,7 @@ mod tests {
         let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
         let provider = ModelProviderInfo {
             name: "test".to_string(),
-            base_url: "https://test.com".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,
@@ -636,6 +683,7 @@ mod tests {
             request_max_retries: Some(0),
             stream_max_retries: Some(0),
             stream_idle_timeout_ms: Some(1000),
+            requires_openai_auth: false,
         };
 
         let events = collect_events(&[sse1.as_bytes()], provider).await;
@@ -728,7 +776,7 @@ mod tests {
 
             let provider = ModelProviderInfo {
                 name: "test".to_string(),
-                base_url: "https://test.com".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,
@@ -738,6 +786,7 @@ mod tests {
                 request_max_retries: Some(0),
                 stream_max_retries: Some(0),
                 stream_idle_timeout_ms: Some(1000),
+                requires_openai_auth: false,
             };
 
             let out = run_sse(evs, provider).await;

codex-rs/core/src/client_common.rs

@@ -1,13 +1,19 @@
 use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
 use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::error::Result;
+use crate::model_family::ModelFamily;
+use crate::models::ContentItem;
 use crate::models::ResponseItem;
+use crate::openai_tools::OpenAiTool;
+use crate::protocol::AskForApproval;
+use crate::protocol::SandboxPolicy;
 use crate::protocol::TokenUsage;
 use codex_apply_patch::APPLY_PATCH_TOOL_INSTRUCTIONS;
 use futures::Stream;
 use serde::Serialize;
 use std::borrow::Cow;
-use std::collections::HashMap;
+use std::fmt::Display;
+use std::path::PathBuf;
 use std::pin::Pin;
 use std::task::Context;
 use std::task::Poll;
@@ -17,6 +23,47 @@ use tokio::sync::mpsc;
 /// with this content.
 const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
 
+/// wraps environment context message in a tag for the model to parse more easily.
+const ENVIRONMENT_CONTEXT_START: &str = "<environment_context>\n\n";
+const ENVIRONMENT_CONTEXT_END: &str = "\n\n</environment_context>";
+
+/// wraps user instructions message in a tag for the model to parse more easily.
+const USER_INSTRUCTIONS_START: &str = "<user_instructions>\n\n";
+const USER_INSTRUCTIONS_END: &str = "\n\n</user_instructions>";
+
+#[derive(Debug, Clone)]
+pub(crate) struct EnvironmentContext {
+    pub cwd: PathBuf,
+    pub approval_policy: AskForApproval,
+    pub sandbox_policy: SandboxPolicy,
+}
+
+impl Display for EnvironmentContext {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        writeln!(
+            f,
+            "Current working directory: {}",
+            self.cwd.to_string_lossy()
+        )?;
+        writeln!(f, "Approval policy: {}", self.approval_policy)?;
+        writeln!(f, "Sandbox policy: {}", self.sandbox_policy)?;
+
+        let network_access = match self.sandbox_policy.clone() {
+            SandboxPolicy::DangerFullAccess => "enabled",
+            SandboxPolicy::ReadOnly => "restricted",
+            SandboxPolicy::WorkspaceWrite { network_access, .. } => {
+                if network_access {
+                    "enabled"
+                } else {
+                    "restricted"
+                }
+            }
+        };
+        writeln!(f, "Network access: {network_access}")?;
+        Ok(())
+    }
+}
+
 /// API request payload for a single model turn.
 #[derive(Default, Debug, Clone)]
 pub struct Prompt {
@@ -28,30 +75,62 @@ pub struct Prompt {
     /// Whether to store response on server side (disable_response_storage = !store).
     pub store: bool,
 
-    /// Additional tools sourced from external MCP servers. Note each key is
-    /// the "fully qualified" tool name (i.e., prefixed with the server name),
-    /// which should be reported to the model in place of Tool::name.
-    pub extra_tools: HashMap<String, mcp_types::Tool>,
+    /// A list of key-value pairs that will be added as a developer message
+    /// for the model to use
+    pub environment_context: Option<EnvironmentContext>,
+
+    /// Tools available to the model, including additional tools sourced from
+    /// external MCP servers.
+    pub tools: Vec<OpenAiTool>,
 
     /// Optional override for the built-in BASE_INSTRUCTIONS.
     pub base_instructions_override: Option<String>,
 }
 
 impl Prompt {
-    pub(crate) fn get_full_instructions(&self, model: &str) -> Cow<'_, str> {
+    pub(crate) fn get_full_instructions(&self, model: &ModelFamily) -> Cow<'_, str> {
         let base = self
             .base_instructions_override
             .as_deref()
             .unwrap_or(BASE_INSTRUCTIONS);
         let mut sections: Vec<&str> = vec![base];
-        if let Some(ref user) = self.user_instructions {
-            sections.push(user);
-        }
-        if model.starts_with("gpt-4.1") {
+        if model.needs_special_apply_patch_instructions {
             sections.push(APPLY_PATCH_TOOL_INSTRUCTIONS);
         }
         Cow::Owned(sections.join("\n"))
     }
+
+    fn get_formatted_user_instructions(&self) -> Option<String> {
+        self.user_instructions
+            .as_ref()
+            .map(|ui| format!("{USER_INSTRUCTIONS_START}{ui}{USER_INSTRUCTIONS_END}"))
+    }
+
+    fn get_formatted_environment_context(&self) -> Option<String> {
+        self.environment_context
+            .as_ref()
+            .map(|ec| format!("{ENVIRONMENT_CONTEXT_START}{ec}{ENVIRONMENT_CONTEXT_END}"))
+    }
+
+    pub(crate) fn get_formatted_input(&self) -> Vec<ResponseItem> {
+        let mut input_with_instructions = Vec::with_capacity(self.input.len() + 2);
+        if let Some(ec) = self.get_formatted_environment_context() {
+            input_with_instructions.push(ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText { text: ec }],
+            });
+        }
+        if let Some(ui) = self.get_formatted_user_instructions() {
+            input_with_instructions.push(ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText { text: ui }],
+            });
+        }
+        input_with_instructions.extend(self.input.clone());
+        input_with_instructions
+    }
 }
 
 #[derive(Debug)]
@@ -64,6 +143,7 @@ pub enum ResponseEvent {
     },
     OutputTextDelta(String),
     ReasoningSummaryDelta(String),
+    ReasoningContentDelta(String),
 }
 
 #[derive(Debug, Serialize)]
@@ -137,14 +217,12 @@ pub(crate) struct ResponsesApiRequest<'a> {
     pub(crate) include: Vec<String>,
 }
 
-use crate::config::Config;
-
 pub(crate) fn create_reasoning_param_for_request(
-    config: &Config,
+    model_family: &ModelFamily,
     effort: ReasoningEffortConfig,
     summary: ReasoningSummaryConfig,
 ) -> Option<Reasoning> {
-    if model_supports_reasoning_summaries(config) {
+    if model_family.supports_reasoning_summaries {
         let effort: Option<OpenAiReasoningEffort> = effort.into();
         let effort = effort?;
         Some(Reasoning {
@@ -156,27 +234,6 @@ pub(crate) fn create_reasoning_param_for_request(
     }
 }
 
-pub fn model_supports_reasoning_summaries(config: &Config) -> bool {
-    // Currently, we hardcode this rule to decide whether to enable reasoning.
-    // We expect reasoning to apply only to OpenAI models, but we do not want
-    // users to have to mess with their config to disable reasoning for models
-    // that do not support it, such as `gpt-4.1`.
-    //
-    // Though if a user is using Codex with non-OpenAI models that, say, happen
-    // to start with "o", then they can set `model_reasoning_effort = "none"` in
-    // config.toml to disable reasoning.
-    //
-    // Converseley, if a user has a non-OpenAI provider that supports reasoning,
-    // they can set the top-level `model_supports_reasoning_summaries = true`
-    // config option to enable reasoning.
-    if config.model_supports_reasoning_summaries {
-        return true;
-    }
-
-    let model = &config.model;
-    model.starts_with("o") || model.starts_with("codex")
-}
-
 pub(crate) struct ResponseStream {
     pub(crate) rx_event: mpsc::Receiver<Result<ResponseEvent>>,
 }
@@ -188,3 +245,23 @@ impl Stream for ResponseStream {
         self.rx_event.poll_recv(cx)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::expect_used)]
+    use crate::model_family::find_family_for_model;
+
+    use super::*;
+
+    #[test]
+    fn get_full_instructions_no_user_content() {
+        let prompt = Prompt {
+            user_instructions: Some("custom instruction".to_string()),
+            ..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);
+    }
+}

codex-rs/core/src/codex_wrapper.rs

@@ -6,6 +6,7 @@ use crate::config::Config;
 use crate::protocol::Event;
 use crate::protocol::EventMsg;
 use crate::util::notify_on_sigint;
+use codex_login::load_auth;
 use tokio::sync::Notify;
 use uuid::Uuid;
 
@@ -25,11 +26,12 @@ pub struct CodexConversation {
 /// that callers can surface the information to the UI.
 pub async fn init_codex(config: Config) -> anyhow::Result<CodexConversation> {
     let ctrl_c = notify_on_sigint();
+    let auth = load_auth(&config.codex_home, true)?;
     let CodexSpawnOk {
         codex,
         init_id,
         session_id,
-    } = Codex::spawn(config, ctrl_c.clone()).await?;
+    } = Codex::spawn(config, auth, ctrl_c.clone()).await?;
 
     // The first event must be `SessionInitialized`. Validate and forward it to
     // the caller so that they can display it in the conversation history.

codex-rs/core/src/config.rs

@@ -10,6 +10,8 @@ use crate::config_types::ShellEnvironmentPolicyToml;
 use crate::config_types::Tui;
 use crate::config_types::UriBasedFileOpener;
 use crate::flags::OPENAI_DEFAULT_MODEL;
+use crate::model_family::ModelFamily;
+use crate::model_family::find_family_for_model;
 use crate::model_provider_info::ModelProviderInfo;
 use crate::model_provider_info::built_in_model_providers;
 use crate::openai_model_info::get_model_info;
@@ -33,6 +35,8 @@ pub struct Config {
     /// Optional override of model selection.
     pub model: String,
 
+    pub model_family: ModelFamily,
+
     /// Size of the context window for the model, in tokens.
     pub model_context_window: Option<u64>,
 
@@ -57,12 +61,16 @@ pub struct Config {
     /// users are only interested in the final agent responses.
     pub hide_agent_reasoning: bool,
 
+    /// When set to `true`, `AgentReasoningRawContentEvent` events will be shown in the UI/output.
+    /// Defaults to `false`.
+    pub show_raw_agent_reasoning: bool,
+
     /// Disable server-side response storage (sends the full conversation
     /// context with every request). Currently necessary for OpenAI customers
     /// who have opted into Zero Data Retention (ZDR).
     pub disable_response_storage: bool,
 
-    /// User-provided instructions from instructions.md.
+    /// User-provided instructions from AGENTS.md.
     pub user_instructions: Option<String>,
 
     /// Base instructions override.
@@ -134,15 +142,17 @@ pub struct Config {
     /// request using the Responses API.
     pub model_reasoning_summary: ReasoningSummary,
 
-    /// When set to `true`, overrides the default heuristic and forces
-    /// `model_supports_reasoning_summaries()` to return `true`.
-    pub model_supports_reasoning_summaries: bool,
-
     /// 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,
+
+    /// The value for the `originator` header included with Responses API requests.
+    pub internal_originator: Option<String>,
 }
 
 impl Config {
@@ -319,6 +329,10 @@ pub struct ConfigToml {
     /// UI/output. Defaults to `false`.
     pub hide_agent_reasoning: Option<bool>,
 
+    /// When set to `true`, `AgentReasoningRawContentEvent` events will be shown in the UI/output.
+    /// Defaults to `false`.
+    pub show_raw_agent_reasoning: Option<bool>,
+
     pub model_reasoning_effort: Option<ReasoningEffort>,
     pub model_reasoning_summary: Option<ReasoningSummary>,
 
@@ -333,6 +347,9 @@ pub struct ConfigToml {
 
     /// Experimental path to a file whose contents replace the built-in BASE_INSTRUCTIONS.
     pub experimental_instructions_file: Option<PathBuf>,
+
+    /// The value for the `originator` header included with Responses API requests.
+    pub internal_originator: Option<String>,
 }
 
 impl ConfigToml {
@@ -347,6 +364,7 @@ impl ConfigToml {
                 Some(s) => SandboxPolicy::WorkspaceWrite {
                     writable_roots: s.writable_roots.clone(),
                     network_access: s.network_access,
+                    include_default_writable_roots: true,
                 },
                 None => SandboxPolicy::new_workspace_write_policy(),
             },
@@ -366,6 +384,9 @@ pub struct ConfigOverrides {
     pub config_profile: Option<String>,
     pub codex_linux_sandbox_exe: Option<PathBuf>,
     pub base_instructions: Option<String>,
+    pub include_plan_tool: Option<bool>,
+    pub disable_response_storage: Option<bool>,
+    pub show_raw_agent_reasoning: Option<bool>,
 }
 
 impl Config {
@@ -388,6 +409,9 @@ impl Config {
             config_profile: config_profile_key,
             codex_linux_sandbox_exe,
             base_instructions,
+            include_plan_tool,
+            disable_response_storage,
+            show_raw_agent_reasoning,
         } = overrides;
 
         let config_profile = match config_profile_key.as_ref().or(cfg.profile.as_ref()) {
@@ -453,7 +477,19 @@ impl Config {
             .or(config_profile.model)
             .or(cfg.model)
             .unwrap_or_else(default_model);
-        let openai_model_info = get_model_info(&model);
+        let model_family = find_family_for_model(&model).unwrap_or_else(|| {
+            let supports_reasoning_summaries =
+                cfg.model_supports_reasoning_summaries.unwrap_or(false);
+            ModelFamily {
+                slug: model.clone(),
+                family: model.clone(),
+                needs_special_apply_patch_instructions: false,
+                supports_reasoning_summaries,
+                uses_local_shell_tool: false,
+            }
+        });
+
+        let openai_model_info = get_model_info(&model_family);
         let model_context_window = cfg
             .model_context_window
             .or_else(|| openai_model_info.as_ref().map(|info| info.context_window));
@@ -465,12 +501,20 @@ impl Config {
 
         let experimental_resume = cfg.experimental_resume;
 
-        let base_instructions = base_instructions.or(Self::get_base_instructions(
-            cfg.experimental_instructions_file.as_ref(),
-        ));
+        // Load base instructions override from a file if specified. If the
+        // path is relative, resolve it against the effective cwd so the
+        // behaviour matches other path-like config values.
+        let experimental_instructions_path = config_profile
+            .experimental_instructions_file
+            .as_ref()
+            .or(cfg.experimental_instructions_file.as_ref());
+        let file_base_instructions =
+            Self::get_base_instructions(experimental_instructions_path, &resolved_cwd)?;
+        let base_instructions = base_instructions.or(file_base_instructions);
 
         let config = Self {
             model,
+            model_family,
             model_context_window,
             model_max_output_tokens,
             model_provider_id,
@@ -485,6 +529,7 @@ impl Config {
             disable_response_storage: config_profile
                 .disable_response_storage
                 .or(cfg.disable_response_storage)
+                .or(disable_response_storage)
                 .unwrap_or(false),
             notify: cfg.notify,
             user_instructions,
@@ -499,6 +544,10 @@ impl Config {
             codex_linux_sandbox_exe,
 
             hide_agent_reasoning: cfg.hide_agent_reasoning.unwrap_or(false),
+            show_raw_agent_reasoning: cfg
+                .show_raw_agent_reasoning
+                .or(show_raw_agent_reasoning)
+                .unwrap_or(false),
             model_reasoning_effort: config_profile
                 .model_reasoning_effort
                 .or(cfg.model_reasoning_effort)
@@ -508,16 +557,14 @@ impl Config {
                 .or(cfg.model_reasoning_summary)
                 .unwrap_or_default(),
 
-            model_supports_reasoning_summaries: cfg
-                .model_supports_reasoning_summaries
-                .unwrap_or(false),
-
             chatgpt_base_url: config_profile
                 .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),
+            internal_originator: cfg.internal_originator,
         };
         Ok(config)
     }
@@ -528,7 +575,7 @@ impl Config {
             None => return None,
         };
 
-        p.push("instructions.md");
+        p.push("AGENTS.md");
         std::fs::read_to_string(&p).ok().and_then(|s| {
             let s = s.trim();
             if s.is_empty() {
@@ -539,13 +586,46 @@ impl Config {
         })
     }
 
-    fn get_base_instructions(path: Option<&PathBuf>) -> Option<String> {
-        let path = path.as_ref()?;
+    fn get_base_instructions(
+        path: Option<&PathBuf>,
+        cwd: &Path,
+    ) -> std::io::Result<Option<String>> {
+        let p = match path.as_ref() {
+            None => return Ok(None),
+            Some(p) => p,
+        };
 
-        std::fs::read_to_string(path)
-            .ok()
-            .map(|s| s.trim().to_string())
-            .filter(|s| !s.is_empty())
+        // Resolve relative paths against the provided cwd to make CLI
+        // overrides consistent regardless of where the process was launched
+        // from.
+        let full_path = if p.is_relative() {
+            cwd.join(p)
+        } else {
+            p.to_path_buf()
+        };
+
+        let contents = std::fs::read_to_string(&full_path).map_err(|e| {
+            std::io::Error::new(
+                e.kind(),
+                format!(
+                    "failed to read experimental instructions file {}: {e}",
+                    full_path.display()
+                ),
+            )
+        })?;
+
+        let s = contents.trim().to_string();
+        if s.is_empty() {
+            Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                format!(
+                    "experimental instructions file is empty: {}",
+                    full_path.display()
+                ),
+            ))
+        } else {
+            Ok(Some(s))
+        }
     }
 }
 
@@ -676,6 +756,7 @@ writable_roots = [
             SandboxPolicy::WorkspaceWrite {
                 writable_roots: vec![PathBuf::from("/tmp")],
                 network_access: false,
+                include_default_writable_roots: true,
             },
             sandbox_workspace_write_cfg.derive_sandbox_policy(sandbox_mode_override)
         );
@@ -751,7 +832,7 @@ disable_response_storage = true
 
         let openai_chat_completions_provider = ModelProviderInfo {
             name: "OpenAI using Chat Completions".to_string(),
-            base_url: "https://api.openai.com/v1".to_string(),
+            base_url: Some("https://api.openai.com/v1".to_string()),
             env_key: Some("OPENAI_API_KEY".to_string()),
             wire_api: crate::WireApi::Chat,
             env_key_instructions: None,
@@ -761,6 +842,7 @@ disable_response_storage = true
             request_max_retries: Some(4),
             stream_max_retries: Some(10),
             stream_idle_timeout_ms: Some(300_000),
+            requires_openai_auth: false,
         };
         let model_provider_map = {
             let mut model_provider_map = built_in_model_providers();
@@ -791,7 +873,7 @@ disable_response_storage = true
     ///
     /// 1. custom command-line argument, e.g. `--model o3`
     /// 2. as part of a profile, where the `--profile` is specified via a CLI
-    ///    (or in the config file itelf)
+    ///    (or in the config file itself)
     /// 3. as an entry in `config.toml`, e.g. `model = "o3"`
     /// 4. the default value for a required field defined in code, e.g.,
     ///    `crate::flags::OPENAI_DEFAULT_MODEL`
@@ -815,6 +897,7 @@ disable_response_storage = true
         assert_eq!(
             Config {
                 model: "o3".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_provider_id: "openai".to_string(),
@@ -835,12 +918,14 @@ disable_response_storage = true
                 tui: Tui::default(),
                 codex_linux_sandbox_exe: None,
                 hide_agent_reasoning: false,
+                show_raw_agent_reasoning: false,
                 model_reasoning_effort: ReasoningEffort::High,
                 model_reasoning_summary: ReasoningSummary::Detailed,
-                model_supports_reasoning_summaries: false,
                 chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
                 experimental_resume: None,
                 base_instructions: None,
+                include_plan_tool: false,
+                internal_originator: None,
             },
             o3_profile_config
         );
@@ -863,6 +948,7 @@ disable_response_storage = true
         )?;
         let expected_gpt3_profile_config = Config {
             model: "gpt-3.5-turbo".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_provider_id: "openai-chat-completions".to_string(),
@@ -883,12 +969,14 @@ disable_response_storage = true
             tui: Tui::default(),
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
+            show_raw_agent_reasoning: false,
             model_reasoning_effort: ReasoningEffort::default(),
             model_reasoning_summary: ReasoningSummary::default(),
-            model_supports_reasoning_summaries: false,
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
             experimental_resume: None,
             base_instructions: None,
+            include_plan_tool: false,
+            internal_originator: None,
         };
 
         assert_eq!(expected_gpt3_profile_config, gpt3_profile_config);
@@ -926,6 +1014,7 @@ disable_response_storage = true
         )?;
         let expected_zdr_profile_config = Config {
             model: "o3".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_provider_id: "openai".to_string(),
@@ -946,12 +1035,14 @@ disable_response_storage = true
             tui: Tui::default(),
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
+            show_raw_agent_reasoning: false,
             model_reasoning_effort: ReasoningEffort::default(),
             model_reasoning_summary: ReasoningSummary::default(),
-            model_supports_reasoning_summaries: false,
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
             experimental_resume: None,
             base_instructions: None,
+            include_plan_tool: false,
+            internal_originator: None,
         };
 
         assert_eq!(expected_zdr_profile_config, zdr_profile_config);

codex-rs/core/src/config_profile.rs

@@ -1,4 +1,5 @@
 use serde::Deserialize;
+use std::path::PathBuf;
 
 use crate::config_types::ReasoningEffort;
 use crate::config_types::ReasoningSummary;
@@ -17,4 +18,5 @@ pub struct ConfigProfile {
     pub model_reasoning_effort: Option<ReasoningEffort>,
     pub model_reasoning_summary: Option<ReasoningSummary>,
     pub chatgpt_base_url: Option<String>,
+    pub experimental_instructions_file: Option<PathBuf>,
 }

codex-rs/core/src/config_types.rs

@@ -78,7 +78,7 @@ pub enum HistoryPersistence {
 #[derive(Deserialize, Debug, Clone, PartialEq, Default)]
 pub struct Tui {}
 
-#[derive(Deserialize, Debug, Clone, Copy, PartialEq, Default)]
+#[derive(Deserialize, Debug, Clone, Copy, PartialEq, Default, Serialize)]
 #[serde(rename_all = "kebab-case")]
 pub enum SandboxMode {
     #[serde(rename = "read-only")]

codex-rs/core/src/conversation_history.rs

@@ -24,12 +24,83 @@ impl ConversationHistory {
         I::Item: std::ops::Deref<Target = ResponseItem>,
     {
         for item in items {
-            if is_api_message(&item) {
-                // Note agent-loop.ts also does filtering on some of the fields.
-                self.items.push(item.clone());
+            if !is_api_message(&item) {
+                continue;
+            }
+
+            // Merge adjacent assistant messages into a single history entry.
+            // This prevents duplicates when a partial assistant message was
+            // streamed into history earlier in the turn and the final full
+            // message is recorded at turn end.
+            match (&*item, self.items.last_mut()) {
+                (
+                    ResponseItem::Message {
+                        role: new_role,
+                        content: new_content,
+                        ..
+                    },
+                    Some(ResponseItem::Message {
+                        role: last_role,
+                        content: last_content,
+                        ..
+                    }),
+                ) if new_role == "assistant" && last_role == "assistant" => {
+                    append_text_content(last_content, new_content);
+                }
+                _ => {
+                    self.items.push(item.clone());
+                }
             }
         }
     }
+
+    /// Append a text `delta` to the latest assistant message, creating a new
+    /// assistant entry if none exists yet (e.g. first delta for this turn).
+    pub(crate) fn append_assistant_text(&mut self, delta: &str) {
+        match self.items.last_mut() {
+            Some(ResponseItem::Message { role, content, .. }) if role == "assistant" => {
+                append_text_delta(content, delta);
+            }
+            _ => {
+                // Start a new assistant message with the delta.
+                self.items.push(ResponseItem::Message {
+                    id: None,
+                    role: "assistant".to_string(),
+                    content: vec![crate::models::ContentItem::OutputText {
+                        text: delta.to_string(),
+                    }],
+                });
+            }
+        }
+    }
+
+    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;
+    }
 }
 
 /// Anything that is not a system message or "reasoning" message is considered
@@ -44,3 +115,140 @@ fn is_api_message(message: &ResponseItem) -> bool {
         ResponseItem::Other => false,
     }
 }
+
+/// Helper to append the textual content from `src` into `dst` in place.
+fn append_text_content(
+    dst: &mut Vec<crate::models::ContentItem>,
+    src: &Vec<crate::models::ContentItem>,
+) {
+    for c in src {
+        if let crate::models::ContentItem::OutputText { text } = c {
+            append_text_delta(dst, text);
+        }
+    }
+}
+
+/// Append a single text delta to the last OutputText item in `content`, or
+/// push a new OutputText item if none exists.
+fn append_text_delta(content: &mut Vec<crate::models::ContentItem>, delta: &str) {
+    if let Some(crate::models::ContentItem::OutputText { text }) = content
+        .iter_mut()
+        .rev()
+        .find(|c| matches!(c, crate::models::ContentItem::OutputText { .. }))
+    {
+        text.push_str(delta);
+    } else {
+        content.push(crate::models::ContentItem::OutputText {
+            text: delta.to_string(),
+        });
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::models::ContentItem;
+
+    fn assistant_msg(text: &str) -> ResponseItem {
+        ResponseItem::Message {
+            id: None,
+            role: "assistant".to_string(),
+            content: vec![ContentItem::OutputText {
+                text: text.to_string(),
+            }],
+        }
+    }
+
+    fn user_msg(text: &str) -> ResponseItem {
+        ResponseItem::Message {
+            id: None,
+            role: "user".to_string(),
+            content: vec![ContentItem::OutputText {
+                text: text.to_string(),
+            }],
+        }
+    }
+
+    #[test]
+    fn merges_adjacent_assistant_messages() {
+        let mut h = ConversationHistory::default();
+        let a1 = assistant_msg("Hello");
+        let a2 = assistant_msg(", world!");
+        h.record_items([&a1, &a2]);
+
+        let items = h.contents();
+        assert_eq!(
+            items,
+            vec![ResponseItem::Message {
+                id: None,
+                role: "assistant".to_string(),
+                content: vec![ContentItem::OutputText {
+                    text: "Hello, world!".to_string()
+                }]
+            }]
+        );
+    }
+
+    #[test]
+    fn append_assistant_text_creates_and_appends() {
+        let mut h = ConversationHistory::default();
+        h.append_assistant_text("Hello");
+        h.append_assistant_text(", world");
+
+        // Now record a final full assistant message and verify it merges.
+        let final_msg = assistant_msg("!");
+        h.record_items([&final_msg]);
+
+        let items = h.contents();
+        assert_eq!(
+            items,
+            vec![ResponseItem::Message {
+                id: None,
+                role: "assistant".to_string(),
+                content: vec![ContentItem::OutputText {
+                    text: "Hello, world!".to_string()
+                }]
+            }]
+        );
+    }
+
+    #[test]
+    fn filters_non_api_messages() {
+        let mut h = ConversationHistory::default();
+        // System message is not an API message; Other is ignored.
+        let system = ResponseItem::Message {
+            id: None,
+            role: "system".to_string(),
+            content: vec![ContentItem::OutputText {
+                text: "ignored".to_string(),
+            }],
+        };
+        h.record_items([&system, &ResponseItem::Other]);
+
+        // User and assistant should be retained.
+        let u = user_msg("hi");
+        let a = assistant_msg("hello");
+        h.record_items([&u, &a]);
+
+        let items = h.contents();
+        assert_eq!(
+            items,
+            vec![
+                ResponseItem::Message {
+                    id: None,
+                    role: "user".to_string(),
+                    content: vec![ContentItem::OutputText {
+                        text: "hi".to_string()
+                    }]
+                },
+                ResponseItem::Message {
+                    id: None,
+                    role: "assistant".to_string(),
+                    content: vec![ContentItem::OutputText {
+                        text: "hello".to_string()
+                    }]
+                }
+            ]
+        );
+    }
+}

codex-rs/core/src/exec.rs

@@ -6,23 +6,29 @@ use std::io;
 use std::path::Path;
 use std::path::PathBuf;
 use std::process::ExitStatus;
-use std::process::Stdio;
 use std::sync::Arc;
 use std::time::Duration;
 use std::time::Instant;
 
+use async_channel::Sender;
 use tokio::io::AsyncRead;
 use tokio::io::AsyncReadExt;
 use tokio::io::BufReader;
 use tokio::process::Child;
-use tokio::process::Command;
 use tokio::sync::Notify;
-use tracing::trace;
 
 use crate::error::CodexErr;
 use crate::error::Result;
 use crate::error::SandboxErr;
+use crate::protocol::Event;
+use crate::protocol::EventMsg;
+use crate::protocol::ExecCommandOutputDeltaEvent;
+use crate::protocol::ExecOutputStream;
 use crate::protocol::SandboxPolicy;
+use crate::seatbelt::spawn_command_under_seatbelt;
+use crate::spawn::StdioPolicy;
+use crate::spawn::spawn_child_async;
+use serde_bytes::ByteBuf;
 
 // Maximum we send for each stream, which is either:
 // - 10KiB OR
@@ -37,30 +43,20 @@ const DEFAULT_TIMEOUT_MS: u64 = 10_000;
 const SIGKILL_CODE: i32 = 9;
 const TIMEOUT_CODE: i32 = 64;
 
-const MACOS_SEATBELT_BASE_POLICY: &str = include_str!("seatbelt_base_policy.sbpl");
-
-/// When working with `sandbox-exec`, only consider `sandbox-exec` in `/usr/bin`
-/// to defend against an attacker trying to inject a malicious version on the
-/// PATH. If /usr/bin/sandbox-exec has been tampered with, then the attacker
-/// already has root access.
-const MACOS_PATH_TO_SEATBELT_EXECUTABLE: &str = "/usr/bin/sandbox-exec";
-
-/// Experimental environment variable that will be set to some non-empty value
-/// if both of the following are true:
-///
-/// 1. The process was spawned by Codex as part of a shell tool call.
-/// 2. SandboxPolicy.has_full_network_access() was false for the tool call.
-///
-/// We may try to have just one environment variable for all sandboxing
-/// attributes, so this may change in the future.
-pub const CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR: &str = "CODEX_SANDBOX_NETWORK_DISABLED";
-
 #[derive(Debug, Clone)]
 pub struct ExecParams {
     pub command: Vec<String>,
     pub cwd: PathBuf,
     pub timeout_ms: Option<u64>,
     pub env: HashMap<String, String>,
+    pub with_escalated_permissions: Option<bool>,
+    pub justification: Option<String>,
+}
+
+impl ExecParams {
+    pub fn timeout_duration(&self) -> Duration {
+        Duration::from_millis(self.timeout_ms.unwrap_or(DEFAULT_TIMEOUT_MS))
+    }
 }
 
 #[derive(Clone, Copy, Debug, PartialEq)]
@@ -74,24 +70,30 @@ pub enum SandboxType {
     LinuxSeccomp,
 }
 
+#[derive(Clone)]
+pub struct StdoutStream {
+    pub sub_id: String,
+    pub call_id: String,
+    pub tx_event: Sender<Event>,
+}
+
 pub async fn process_exec_tool_call(
     params: ExecParams,
     sandbox_type: SandboxType,
     ctrl_c: Arc<Notify>,
     sandbox_policy: &SandboxPolicy,
     codex_linux_sandbox_exe: &Option<PathBuf>,
+    stdout_stream: Option<StdoutStream>,
 ) -> Result<ExecToolCallOutput> {
     let start = Instant::now();
 
     let raw_output_result: std::result::Result<RawExecToolCallOutput, CodexErr> = match sandbox_type
     {
-        SandboxType::None => exec(params, sandbox_policy, ctrl_c).await,
+        SandboxType::None => exec(params, sandbox_policy, ctrl_c, stdout_stream.clone()).await,
         SandboxType::MacosSeatbelt => {
+            let timeout = params.timeout_duration();
             let ExecParams {
-                command,
-                cwd,
-                timeout_ms,
-                env,
+                command, cwd, env, ..
             } = params;
             let child = spawn_command_under_seatbelt(
                 command,
@@ -101,14 +103,12 @@ pub async fn process_exec_tool_call(
                 env,
             )
             .await?;
-            consume_truncated_output(child, ctrl_c, timeout_ms).await
+            consume_truncated_output(child, ctrl_c, timeout, stdout_stream.clone()).await
         }
         SandboxType::LinuxSeccomp => {
+            let timeout = params.timeout_duration();
             let ExecParams {
-                command,
-                cwd,
-                timeout_ms,
-                env,
+                command, cwd, env, ..
             } = params;
 
             let codex_linux_sandbox_exe = codex_linux_sandbox_exe
@@ -124,7 +124,7 @@ pub async fn process_exec_tool_call(
             )
             .await?;
 
-            consume_truncated_output(child, ctrl_c, timeout_ms).await
+            consume_truncated_output(child, ctrl_c, timeout, stdout_stream).await
         }
     };
     let duration = start.elapsed();
@@ -144,11 +144,7 @@ pub async fn process_exec_tool_call(
 
             let exit_code = raw_output.exit_status.code().unwrap_or(-1);
 
-            // NOTE(ragona): This is much less restrictive than the previous check. If we exec
-            // a command, and it returns anything other than success, we assume that it may have
-            // been a sandboxing error and allow the user to retry. (The user of course may choose
-            // not to retry, or in a non-interactive mode, would automatically reject the approval.)
-            if exit_code != 0 && sandbox_type != SandboxType::None {
+            if exit_code != 0 && is_likely_sandbox_denied(sandbox_type, exit_code) {
                 return Err(CodexErr::Sandbox(SandboxErr::Denied(
                     exit_code, stdout, stderr,
                 )));
@@ -168,27 +164,6 @@ pub async fn process_exec_tool_call(
     }
 }
 
-pub async fn spawn_command_under_seatbelt(
-    command: Vec<String>,
-    sandbox_policy: &SandboxPolicy,
-    cwd: PathBuf,
-    stdio_policy: StdioPolicy,
-    env: HashMap<String, String>,
-) -> std::io::Result<Child> {
-    let args = create_seatbelt_command_args(command, sandbox_policy, &cwd);
-    let arg0 = None;
-    spawn_child_async(
-        PathBuf::from(MACOS_PATH_TO_SEATBELT_EXECUTABLE),
-        args,
-        arg0,
-        cwd,
-        sandbox_policy,
-        stdio_policy,
-        env,
-    )
-    .await
-}
-
 /// Spawn a shell tool command under the Linux Landlock+seccomp sandbox helper
 /// (codex-linux-sandbox).
 ///
@@ -248,63 +223,24 @@ fn create_linux_sandbox_command_args(
     linux_cmd
 }
 
-fn create_seatbelt_command_args(
-    command: Vec<String>,
-    sandbox_policy: &SandboxPolicy,
-    cwd: &Path,
-) -> Vec<String> {
-    let (file_write_policy, extra_cli_args) = {
-        if sandbox_policy.has_full_disk_write_access() {
-            // Allegedly, this is more permissive than `(allow file-write*)`.
-            (
-                r#"(allow file-write* (regex #"^/"))"#.to_string(),
-                Vec::<String>::new(),
-            )
-        } else {
-            let writable_roots = sandbox_policy.get_writable_roots_with_cwd(cwd);
-            let (writable_folder_policies, cli_args): (Vec<String>, Vec<String>) = writable_roots
-                .iter()
-                .enumerate()
-                .map(|(index, root)| {
-                    let param_name = format!("WRITABLE_ROOT_{index}");
-                    let policy: String = format!("(subpath (param \"{param_name}\"))");
-                    let cli_arg = format!("-D{param_name}={}", root.to_string_lossy());
-                    (policy, cli_arg)
-                })
-                .unzip();
-            if writable_folder_policies.is_empty() {
-                ("".to_string(), Vec::<String>::new())
-            } else {
-                let file_write_policy = format!(
-                    "(allow file-write*\n{}\n)",
-                    writable_folder_policies.join(" ")
-                );
-                (file_write_policy, cli_args)
-            }
-        }
-    };
+/// We don't have a fully deterministic way to tell if our command failed
+/// because of the sandbox - a command in the user's zshrc file might hit an
+/// error, but the command itself might fail or succeed for other reasons.
+/// For now, we conservatively check for 'command not found' (exit code 127),
+/// and can add additional cases as necessary.
+fn is_likely_sandbox_denied(sandbox_type: SandboxType, exit_code: i32) -> bool {
+    if sandbox_type == SandboxType::None {
+        return false;
+    }
 
-    let file_read_policy = if sandbox_policy.has_full_disk_read_access() {
-        "; allow read-only file operations\n(allow file-read*)"
-    } else {
-        ""
-    };
+    // Quick rejects: well-known non-sandbox shell exit codes
+    // 127: command not found, 2: misuse of shell builtins
+    if exit_code == 127 {
+        return false;
+    }
 
-    // TODO(mbolin): apply_patch calls must also honor the SandboxPolicy.
-    let network_policy = if sandbox_policy.has_full_network_access() {
-        "(allow network-outbound)\n(allow network-inbound)\n(allow system-socket)"
-    } else {
-        ""
-    };
-
-    let full_policy = format!(
-        "{MACOS_SEATBELT_BASE_POLICY}\n{file_read_policy}\n{file_write_policy}\n{network_policy}"
-    );
-    let mut seatbelt_args: Vec<String> = vec!["-p".to_string(), full_policy];
-    seatbelt_args.extend(extra_cli_args);
-    seatbelt_args.push("--".to_string());
-    seatbelt_args.extend(command);
-    seatbelt_args
+    // For all other cases, we assume the sandbox is the cause
+    true
 }
 
 #[derive(Debug)]
@@ -323,15 +259,16 @@ pub struct ExecToolCallOutput {
 }
 
 async fn exec(
-    ExecParams {
-        command,
-        cwd,
-        timeout_ms,
-        env,
-    }: ExecParams,
+    params: ExecParams,
     sandbox_policy: &SandboxPolicy,
     ctrl_c: Arc<Notify>,
+    stdout_stream: Option<StdoutStream>,
 ) -> Result<RawExecToolCallOutput> {
+    let timeout = params.timeout_duration();
+    let ExecParams {
+        command, cwd, env, ..
+    } = params;
+
     let (program, args) = command.split_first().ok_or_else(|| {
         CodexErr::Io(io::Error::new(
             io::ErrorKind::InvalidInput,
@@ -349,91 +286,7 @@ async fn exec(
         env,
     )
     .await?;
-    consume_truncated_output(child, ctrl_c, timeout_ms).await
-}
-
-#[derive(Debug, Clone, Copy)]
-pub enum StdioPolicy {
-    RedirectForShellTool,
-    Inherit,
-}
-
-/// Spawns the appropriate child process for the ExecParams and SandboxPolicy,
-/// ensuring the args and environment variables used to create the `Command`
-/// (and `Child`) honor the configuration.
-///
-/// For now, we take `SandboxPolicy` as a parameter to spawn_child() because
-/// we need to determine whether to set the
-/// `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` environment variable.
-async fn spawn_child_async(
-    program: PathBuf,
-    args: Vec<String>,
-    #[cfg_attr(not(unix), allow(unused_variables))] arg0: Option<&str>,
-    cwd: PathBuf,
-    sandbox_policy: &SandboxPolicy,
-    stdio_policy: StdioPolicy,
-    env: HashMap<String, String>,
-) -> std::io::Result<Child> {
-    trace!(
-        "spawn_child_async: {program:?} {args:?} {arg0:?} {cwd:?} {sandbox_policy:?} {stdio_policy:?} {env:?}"
-    );
-
-    let mut cmd = Command::new(&program);
-    #[cfg(unix)]
-    cmd.arg0(arg0.map_or_else(|| program.to_string_lossy().to_string(), String::from));
-    cmd.args(args);
-    cmd.current_dir(cwd);
-    cmd.env_clear();
-    cmd.envs(env);
-
-    if !sandbox_policy.has_full_network_access() {
-        cmd.env(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR, "1");
-    }
-
-    // If this Codex process dies (including being killed via SIGKILL), we want
-    // any child processes that were spawned as part of a `"shell"` tool call
-    // to also be terminated.
-
-    // This relies on prctl(2), so it only works on Linux.
-    #[cfg(target_os = "linux")]
-    unsafe {
-        cmd.pre_exec(|| {
-            // This prctl call effectively requests, "deliver SIGTERM when my
-            // current parent dies."
-            if libc::prctl(libc::PR_SET_PDEATHSIG, libc::SIGTERM) == -1 {
-                return Err(io::Error::last_os_error());
-            }
-
-            // Though if there was a race condition and this pre_exec() block is
-            // run _after_ the parent (i.e., the Codex process) has already
-            // exited, then the parent is the _init_ process (which will never
-            // die), so we should just terminate the child process now.
-            if libc::getppid() == 1 {
-                libc::raise(libc::SIGTERM);
-            }
-            Ok(())
-        });
-    }
-
-    match stdio_policy {
-        StdioPolicy::RedirectForShellTool => {
-            // Do not create a file descriptor for stdin because otherwise some
-            // commands may hang forever waiting for input. For example, ripgrep has
-            // a heuristic where it may try to read from stdin as explained here:
-            // https://github.com/BurntSushi/ripgrep/blob/e2362d4d5185d02fa857bf381e7bd52e66fafc73/crates/core/flags/hiargs.rs#L1101-L1103
-            cmd.stdin(Stdio::null());
-
-            cmd.stdout(Stdio::piped()).stderr(Stdio::piped());
-        }
-        StdioPolicy::Inherit => {
-            // Inherit stdin, stdout, and stderr from the parent process.
-            cmd.stdin(Stdio::inherit())
-                .stdout(Stdio::inherit())
-                .stderr(Stdio::inherit());
-        }
-    }
-
-    cmd.kill_on_drop(true).spawn()
+    consume_truncated_output(child, ctrl_c, timeout, stdout_stream).await
 }
 
 /// Consumes the output of a child process, truncating it so it is suitable for
@@ -441,7 +294,8 @@ async fn spawn_child_async(
 pub(crate) async fn consume_truncated_output(
     mut child: Child,
     ctrl_c: Arc<Notify>,
-    timeout_ms: Option<u64>,
+    timeout: Duration,
+    stdout_stream: Option<StdoutStream>,
 ) -> Result<RawExecToolCallOutput> {
     // Both stdout and stderr were configured with `Stdio::piped()`
     // above, therefore `take()` should normally return `Some`.  If it doesn't
@@ -462,15 +316,18 @@ pub(crate) async fn consume_truncated_output(
         BufReader::new(stdout_reader),
         MAX_STREAM_OUTPUT,
         MAX_STREAM_OUTPUT_LINES,
+        stdout_stream.clone(),
+        false,
     ));
     let stderr_handle = tokio::spawn(read_capped(
         BufReader::new(stderr_reader),
         MAX_STREAM_OUTPUT,
         MAX_STREAM_OUTPUT_LINES,
+        stdout_stream.clone(),
+        true,
     ));
 
     let interrupted = ctrl_c.notified();
-    let timeout = Duration::from_millis(timeout_ms.unwrap_or(DEFAULT_TIMEOUT_MS));
     let exit_status = tokio::select! {
         result = tokio::time::timeout(timeout, child.wait()) => {
             match result {
@@ -500,10 +357,12 @@ pub(crate) async fn consume_truncated_output(
     })
 }
 
-async fn read_capped<R: AsyncRead + Unpin>(
+async fn read_capped<R: AsyncRead + Unpin + Send + 'static>(
     mut reader: R,
     max_output: usize,
     max_lines: usize,
+    stream: Option<StdoutStream>,
+    is_stderr: bool,
 ) -> io::Result<Vec<u8>> {
     let mut buf = Vec::with_capacity(max_output.min(8 * 1024));
     let mut tmp = [0u8; 8192];
@@ -517,6 +376,25 @@ async fn read_capped<R: AsyncRead + Unpin>(
             break;
         }
 
+        if let Some(stream) = &stream {
+            let chunk = tmp[..n].to_vec();
+            let msg = EventMsg::ExecCommandOutputDelta(ExecCommandOutputDeltaEvent {
+                call_id: stream.call_id.clone(),
+                stream: if is_stderr {
+                    ExecOutputStream::Stderr
+                } else {
+                    ExecOutputStream::Stdout
+                },
+                chunk: ByteBuf::from(chunk),
+            });
+            let event = Event {
+                id: stream.sub_id.clone(),
+                msg,
+            };
+            #[allow(clippy::let_unit_value)]
+            let _ = stream.tx_event.send(event).await;
+        }
+
         // Copy into the buffer only while we still have byte and line budget.
         if remaining_bytes > 0 && remaining_lines > 0 {
             let mut copy_len = 0;

codex-rs/core/src/git_info.rs

@@ -9,7 +9,7 @@ use tokio::time::timeout;
 /// Timeout for git commands to prevent freezing on large repositories
 const GIT_COMMAND_TIMEOUT: TokioDuration = TokioDuration::from_secs(5);
 
-#[derive(Serialize, Deserialize, Clone)]
+#[derive(Serialize, Deserialize, Clone, Debug)]
 pub struct GitInfo {
     /// Current commit hash (SHA)
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -111,9 +111,14 @@ mod tests {
     // Helper function to create a test git repository
     async fn create_test_git_repo(temp_dir: &TempDir) -> PathBuf {
         let repo_path = temp_dir.path().to_path_buf();
+        let envs = vec![
+            ("GIT_CONFIG_GLOBAL", "/dev/null"),
+            ("GIT_CONFIG_NOSYSTEM", "1"),
+        ];
 
         // Initialize git repo
         Command::new("git")
+            .envs(envs.clone())
             .args(["init"])
             .current_dir(&repo_path)
             .output()
@@ -122,6 +127,7 @@ mod tests {
 
         // Configure git user (required for commits)
         Command::new("git")
+            .envs(envs.clone())
             .args(["config", "user.name", "Test User"])
             .current_dir(&repo_path)
             .output()
@@ -129,6 +135,7 @@ mod tests {
             .expect("Failed to set git user name");
 
         Command::new("git")
+            .envs(envs.clone())
             .args(["config", "user.email", "test@example.com"])
             .current_dir(&repo_path)
             .output()
@@ -140,6 +147,7 @@ mod tests {
         fs::write(&test_file, "test content").expect("Failed to write test file");
 
         Command::new("git")
+            .envs(envs.clone())
             .args(["add", "."])
             .current_dir(&repo_path)
             .output()
@@ -147,6 +155,7 @@ mod tests {
             .expect("Failed to add files");
 
         Command::new("git")
+            .envs(envs.clone())
             .args(["commit", "-m", "Initial commit"])
             .current_dir(&repo_path)
             .output()

codex-rs/core/src/lib.rs

@@ -28,18 +28,25 @@ mod mcp_connection_manager;
 mod mcp_tool_call;
 mod message_history;
 mod model_provider_info;
+pub use model_provider_info::BUILT_IN_OSS_MODEL_PROVIDER_ID;
 pub use model_provider_info::ModelProviderInfo;
 pub use model_provider_info::WireApi;
+pub use model_provider_info::built_in_model_providers;
+pub use model_provider_info::create_oss_provider_with_base_url;
+pub mod model_family;
 mod models;
-pub mod openai_api_key;
 mod openai_model_info;
 mod openai_tools;
+pub mod plan_tool;
 mod project_doc;
 pub mod protocol;
 mod rollout;
-mod safety;
+pub(crate) mod safety;
+pub mod seatbelt;
 pub mod shell;
+pub mod spawn;
+pub mod turn_diff_tracker;
 mod user_notification;
 pub mod util;
-
-pub use client_common::model_supports_reasoning_summaries;
+pub use apply_patch::CODEX_APPLY_PATCH_ARG1;
+pub use safety::get_platform_sandbox;

codex-rs/core/src/mcp_tool_call.rs

@@ -1,4 +1,5 @@
 use std::time::Duration;
+use std::time::Instant;
 
 use tracing::error;
 
@@ -7,6 +8,7 @@ use crate::models::FunctionCallOutputPayload;
 use crate::models::ResponseInputItem;
 use crate::protocol::Event;
 use crate::protocol::EventMsg;
+use crate::protocol::McpInvocation;
 use crate::protocol::McpToolCallBeginEvent;
 use crate::protocol::McpToolCallEndEvent;
 
@@ -41,21 +43,28 @@ pub(crate) async fn handle_mcp_tool_call(
         }
     };
 
-    let tool_call_begin_event = EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
-        call_id: call_id.clone(),
+    let invocation = McpInvocation {
         server: server.clone(),
         tool: tool_name.clone(),
         arguments: arguments_value.clone(),
+    };
+
+    let tool_call_begin_event = EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
+        call_id: call_id.clone(),
+        invocation: invocation.clone(),
     });
     notify_mcp_tool_call_event(sess, sub_id, tool_call_begin_event).await;
 
+    let start = Instant::now();
     // Perform the tool call.
     let result = sess
-        .call_tool(&server, &tool_name, arguments_value, timeout)
+        .call_tool(&server, &tool_name, arguments_value.clone(), timeout)
         .await
         .map_err(|e| format!("tool call error: {e}"));
     let tool_call_end_event = EventMsg::McpToolCallEnd(McpToolCallEndEvent {
         call_id: call_id.clone(),
+        invocation,
+        duration: start.elapsed(),
         result: result.clone(),
     });
 

codex-rs/core/src/model_family.rs

@@ -0,0 +1,95 @@
+/// A model family is a group of models that share certain characteristics.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub struct ModelFamily {
+    /// The full model slug used to derive this model family, e.g.
+    /// "gpt-4.1-2025-04-14".
+    pub slug: String,
+
+    /// The model family name, e.g. "gpt-4.1". Note this should able to be used
+    /// with [`crate::openai_model_info::get_model_info`].
+    pub family: String,
+
+    /// True if the model needs additional instructions on how to use the
+    /// "virtual" `apply_patch` CLI.
+    pub needs_special_apply_patch_instructions: bool,
+
+    // Whether the `reasoning` field can be set when making a request to this
+    // model family. Note it has `effort` and `summary` subfields (though
+    // `summary` is optional).
+    pub supports_reasoning_summaries: bool,
+
+    // This should be set to true when the model expects a tool named
+    // "local_shell" to be provided. Its contract must be understood natively by
+    // the model such that its description can be omitted.
+    // See https://platform.openai.com/docs/guides/tools-local-shell
+    pub uses_local_shell_tool: bool,
+}
+
+macro_rules! model_family {
+    (
+        $slug:expr, $family:expr $(, $key:ident : $value:expr )* $(,)?
+    ) => {{
+        // defaults
+        let mut mf = ModelFamily {
+            slug: $slug.to_string(),
+            family: $family.to_string(),
+            needs_special_apply_patch_instructions: false,
+            supports_reasoning_summaries: false,
+            uses_local_shell_tool: false,
+        };
+        // apply overrides
+        $(
+            mf.$key = $value;
+        )*
+        Some(mf)
+    }};
+}
+
+macro_rules! simple_model_family {
+    (
+        $slug:expr, $family:expr
+    ) => {{
+        Some(ModelFamily {
+            slug: $slug.to_string(),
+            family: $family.to_string(),
+            needs_special_apply_patch_instructions: false,
+            supports_reasoning_summaries: false,
+            uses_local_shell_tool: false,
+        })
+    }};
+}
+
+/// 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> {
+    if slug.starts_with("o3") {
+        model_family!(
+            slug, "o3",
+            supports_reasoning_summaries: true,
+        )
+    } else if slug.starts_with("o4-mini") {
+        model_family!(
+            slug, "o4-mini",
+            supports_reasoning_summaries: 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("gpt-4.1") {
+        model_family!(
+            slug, "gpt-4.1",
+            needs_special_apply_patch_instructions: true,
+        )
+    } else if slug.starts_with("gpt-4o") {
+        simple_model_family!(slug, "gpt-4o")
+    } else if slug.starts_with("gpt-oss") {
+        simple_model_family!(slug, "gpt-oss")
+    } else if slug.starts_with("gpt-3.5") {
+        simple_model_family!(slug, "gpt-3.5")
+    } else {
+        None
+    }
+}

codex-rs/core/src/model_provider_info.rs

@@ -5,6 +5,8 @@
 //!   2. User-defined entries inside `~/.codex/config.toml` under the `model_providers`
 //!      key. These override or extend the defaults at runtime.
 
+use codex_login::AuthMode;
+use codex_login::CodexAuth;
 use serde::Deserialize;
 use serde::Serialize;
 use std::collections::HashMap;
@@ -12,11 +14,6 @@ use std::env::VarError;
 use std::time::Duration;
 
 use crate::error::EnvVarError;
-use crate::openai_api_key::get_openai_api_key;
-
-/// Value for the `OpenAI-Originator` header that is sent with requests to
-/// OpenAI.
-const OPENAI_ORIGINATOR_HEADER: &str = "codex_cli_rs";
 const DEFAULT_STREAM_IDLE_TIMEOUT_MS: u64 = 300_000;
 const DEFAULT_STREAM_MAX_RETRIES: u64 = 10;
 const DEFAULT_REQUEST_MAX_RETRIES: u64 = 4;
@@ -30,7 +27,7 @@ const DEFAULT_REQUEST_MAX_RETRIES: u64 = 4;
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub enum WireApi {
-    /// The experimental "Responses" API exposed by OpenAI at `/v1/responses`.
+    /// The Responses API exposed by OpenAI at `/v1/responses`.
     Responses,
 
     /// Regular Chat Completions compatible with `/v1/chat/completions`.
@@ -44,7 +41,7 @@ pub struct ModelProviderInfo {
     /// Friendly display name.
     pub name: String,
     /// Base URL for the provider's OpenAI-compatible API.
-    pub base_url: String,
+    pub base_url: Option<String>,
     /// Environment variable that stores the user's API key for this provider.
     pub env_key: Option<String>,
 
@@ -78,6 +75,10 @@ pub struct ModelProviderInfo {
     /// Idle timeout (in milliseconds) to wait for activity on a streaming response before treating
     /// the connection as lost.
     pub stream_idle_timeout_ms: Option<u64>,
+
+    /// Whether this provider requires some form of standard authentication (API key, ChatGPT token).
+    #[serde(default)]
+    pub requires_openai_auth: bool,
 }
 
 impl ModelProviderInfo {
@@ -85,29 +86,40 @@ impl ModelProviderInfo {
     /// reqwest Client applying:
     ///   • provider-specific headers (static + env based)
     ///   • Bearer auth header when an API key is available.
+    ///   • Auth token for OAuth.
     ///
-    /// When `require_api_key` is true and the provider declares an `env_key`
-    /// but the variable is missing/empty, returns an [`Err`] identical to the
+    /// If the provider declares an `env_key` but the variable is missing/empty, returns an [`Err`] identical to the
     /// one produced by [`ModelProviderInfo::api_key`].
-    pub fn create_request_builder<'a>(
+    pub async fn create_request_builder<'a>(
         &'a self,
         client: &'a reqwest::Client,
+        auth: &Option<CodexAuth>,
     ) -> crate::error::Result<reqwest::RequestBuilder> {
-        let api_key = self.api_key()?;
+        let effective_auth = match self.api_key() {
+            Ok(Some(key)) => Some(CodexAuth::from_api_key(key)),
+            Ok(None) => auth.clone(),
+            Err(err) => {
+                if auth.is_some() {
+                    auth.clone()
+                } else {
+                    return Err(err);
+                }
+            }
+        };
 
-        let url = self.get_full_url();
+        let url = self.get_full_url(&effective_auth);
 
         let mut builder = client.post(url);
-        if let Some(key) = api_key {
-            builder = builder.bearer_auth(key);
+
+        if let Some(auth) = effective_auth.as_ref() {
+            builder = builder.bearer_auth(auth.get_token().await?);
         }
 
         Ok(self.apply_http_headers(builder))
     }
 
-    pub(crate) fn get_full_url(&self) -> String {
-        let query_string = self
-            .query_params
+    fn get_query_string(&self) -> String {
+        self.query_params
             .as_ref()
             .map_or_else(String::new, |params| {
                 let full_params = params
@@ -116,8 +128,27 @@ impl ModelProviderInfo {
                     .collect::<Vec<_>>()
                     .join("&");
                 format!("?{full_params}")
-            });
-        let base_url = &self.base_url;
+            })
+    }
+
+    pub(crate) fn get_full_url(&self, auth: &Option<CodexAuth>) -> String {
+        let default_base_url = if matches!(
+            auth,
+            Some(CodexAuth {
+                mode: AuthMode::ChatGPT,
+                ..
+            })
+        ) {
+            "https://chatgpt.com/backend-api/codex"
+        } else {
+            "https://api.openai.com/v1"
+        };
+        let query_string = self.get_query_string();
+        let base_url = self
+            .base_url
+            .clone()
+            .unwrap_or(default_base_url.to_string());
+
         match self.wire_api {
             WireApi::Responses => format!("{base_url}/responses{query_string}"),
             WireApi::Chat => format!("{base_url}/chat/completions{query_string}"),
@@ -149,14 +180,10 @@ impl ModelProviderInfo {
     /// If `env_key` is Some, returns the API key for this provider if present
     /// (and non-empty) in the environment. If `env_key` is required but
     /// cannot be found, returns an error.
-    fn api_key(&self) -> crate::error::Result<Option<String>> {
+    pub fn api_key(&self) -> crate::error::Result<Option<String>> {
         match &self.env_key {
             Some(env_key) => {
-                let env_value = if env_key == crate::openai_api_key::OPENAI_API_KEY_ENV_VAR {
-                    get_openai_api_key().map_or_else(|| Err(VarError::NotPresent), Ok)
-                } else {
-                    std::env::var(env_key)
-                };
+                let env_value = std::env::var(env_key);
                 env_value
                     .and_then(|v| {
                         if v.trim().is_empty() {
@@ -196,14 +223,18 @@ impl ModelProviderInfo {
     }
 }
 
+const DEFAULT_OLLAMA_PORT: u32 = 11434;
+
+pub const BUILT_IN_OSS_MODEL_PROVIDER_ID: &str = "oss";
+
 /// Built-in default provider list.
 pub fn built_in_model_providers() -> HashMap<String, ModelProviderInfo> {
     use ModelProviderInfo as P;
 
     // We do not want to be in the business of adjucating which third-party
-    // providers are bundled with Codex CLI, so we only include the OpenAI
-    // provider by default. Users are encouraged to add to `model_providers`
-    // in config.toml to add their own providers.
+    // providers are bundled with Codex CLI, so we only include the OpenAI and
+    // open source ("oss") providers by default. Users are encouraged to add to
+    // `model_providers` in config.toml to add their own providers.
     [
         (
             "openai",
@@ -216,40 +247,79 @@ pub fn built_in_model_providers() -> HashMap<String, ModelProviderInfo> {
                 // OpenAI provider.
                 base_url: std::env::var("OPENAI_BASE_URL")
                     .ok()
-                    .filter(|v| !v.trim().is_empty())
-                    .unwrap_or_else(|| "https://api.openai.com/v1".to_string()),
-                env_key: Some("OPENAI_API_KEY".into()),
-                env_key_instructions: Some("Create an API key (https://platform.openai.com) and export it as an environment variable.".into()),
+                    .filter(|v| !v.trim().is_empty()),
+                env_key: None,
+                env_key_instructions: None,
                 wire_api: WireApi::Responses,
                 query_params: None,
                 http_headers: Some(
-                    [
-                        ("originator".to_string(), OPENAI_ORIGINATOR_HEADER.to_string()),
-                        ("version".to_string(), env!("CARGO_PKG_VERSION").to_string()),
-                    ]
+                    [("version".to_string(), env!("CARGO_PKG_VERSION").to_string())]
                         .into_iter()
                         .collect(),
                 ),
                 env_http_headers: Some(
                     [
-                        ("OpenAI-Organization".to_string(), "OPENAI_ORGANIZATION".to_string()),
+                        (
+                            "OpenAI-Organization".to_string(),
+                            "OPENAI_ORGANIZATION".to_string(),
+                        ),
                         ("OpenAI-Project".to_string(), "OPENAI_PROJECT".to_string()),
                     ]
-                        .into_iter()
-                        .collect(),
+                    .into_iter()
+                    .collect(),
                 ),
                 // Use global defaults for retry/timeout unless overridden in config.toml.
                 request_max_retries: None,
                 stream_max_retries: None,
                 stream_idle_timeout_ms: None,
+                requires_openai_auth: true,
             },
         ),
+        (BUILT_IN_OSS_MODEL_PROVIDER_ID, create_oss_provider()),
     ]
     .into_iter()
     .map(|(k, v)| (k.to_string(), v))
     .collect()
 }
 
+pub fn create_oss_provider() -> ModelProviderInfo {
+    // These CODEX_OSS_ environment variables are experimental: we may
+    // switch to reading values from config.toml instead.
+    let codex_oss_base_url = match std::env::var("CODEX_OSS_BASE_URL")
+        .ok()
+        .filter(|v| !v.trim().is_empty())
+    {
+        Some(url) => url,
+        None => format!(
+            "http://localhost:{port}/v1",
+            port = std::env::var("CODEX_OSS_PORT")
+                .ok()
+                .filter(|v| !v.trim().is_empty())
+                .and_then(|v| v.parse::<u32>().ok())
+                .unwrap_or(DEFAULT_OLLAMA_PORT)
+        ),
+    };
+
+    create_oss_provider_with_base_url(&codex_oss_base_url)
+}
+
+pub fn create_oss_provider_with_base_url(base_url: &str) -> ModelProviderInfo {
+    ModelProviderInfo {
+        name: "gpt-oss".into(),
+        base_url: Some(base_url.into()),
+        env_key: None,
+        env_key_instructions: None,
+        wire_api: WireApi::Chat,
+        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,
+    }
+}
+
 #[cfg(test)]
 mod tests {
     #![allow(clippy::unwrap_used)]
@@ -264,7 +334,7 @@ base_url = "http://localhost:11434/v1"
         "#;
         let expected_provider = ModelProviderInfo {
             name: "Ollama".into(),
-            base_url: "http://localhost:11434/v1".into(),
+            base_url: Some("http://localhost:11434/v1".into()),
             env_key: None,
             env_key_instructions: None,
             wire_api: WireApi::Chat,
@@ -274,6 +344,7 @@ base_url = "http://localhost:11434/v1"
             request_max_retries: None,
             stream_max_retries: None,
             stream_idle_timeout_ms: None,
+            requires_openai_auth: false,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
@@ -290,7 +361,7 @@ query_params = { api-version = "2025-04-01-preview" }
         "#;
         let expected_provider = ModelProviderInfo {
             name: "Azure".into(),
-            base_url: "https://xxxxx.openai.azure.com/openai".into(),
+            base_url: Some("https://xxxxx.openai.azure.com/openai".into()),
             env_key: Some("AZURE_OPENAI_API_KEY".into()),
             env_key_instructions: None,
             wire_api: WireApi::Chat,
@@ -302,6 +373,7 @@ query_params = { api-version = "2025-04-01-preview" }
             request_max_retries: None,
             stream_max_retries: None,
             stream_idle_timeout_ms: None,
+            requires_openai_auth: false,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
@@ -319,7 +391,7 @@ env_http_headers = { "X-Example-Env-Header" = "EXAMPLE_ENV_VAR" }
         "#;
         let expected_provider = ModelProviderInfo {
             name: "Example".into(),
-            base_url: "https://example.com".into(),
+            base_url: Some("https://example.com".into()),
             env_key: Some("API_KEY".into()),
             env_key_instructions: None,
             wire_api: WireApi::Chat,
@@ -333,6 +405,7 @@ env_http_headers = { "X-Example-Env-Header" = "EXAMPLE_ENV_VAR" }
             request_max_retries: None,
             stream_max_retries: None,
             stream_idle_timeout_ms: None,
+            requires_openai_auth: false,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();

codex-rs/core/src/models.rs

@@ -9,7 +9,7 @@ use serde::ser::Serializer;
 
 use crate::protocol::InputItem;
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ResponseInputItem {
     Message {
@@ -26,7 +26,7 @@ pub enum ResponseInputItem {
     },
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ContentItem {
     InputText { text: String },
@@ -34,7 +34,7 @@ pub enum ContentItem {
     OutputText { text: String },
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ResponseItem {
     Message {
@@ -45,6 +45,8 @@ pub enum ResponseItem {
     Reasoning {
         id: String,
         summary: Vec<ReasoningItemReasoningSummary>,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        content: Option<Vec<ReasoningItemContent>>,
         encrypted_content: Option<String>,
     },
     LocalShellCall {
@@ -107,7 +109,7 @@ impl From<ResponseInputItem> for ResponseItem {
     }
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(rename_all = "snake_case")]
 pub enum LocalShellStatus {
     Completed,
@@ -115,13 +117,13 @@ pub enum LocalShellStatus {
     Incomplete,
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum LocalShellAction {
     Exec(LocalShellExecAction),
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 pub struct LocalShellExecAction {
     pub command: Vec<String>,
     pub timeout_ms: Option<u64>,
@@ -130,12 +132,18 @@ pub struct LocalShellExecAction {
     pub user: Option<String>,
 }
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ReasoningItemReasoningSummary {
     SummaryText { text: String },
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum ReasoningItemContent {
+    ReasoningText { text: String },
+}
+
 impl From<Vec<InputItem>> for ResponseInputItem {
     fn from(items: Vec<InputItem>) -> Self {
         Self::Message {
@@ -183,12 +191,15 @@ pub struct ShellToolCallParams {
     // The wire format uses `timeout`, which has ambiguous units, so we use
     // `timeout_ms` as the field name so it is clear in code.
     pub timeout_ms: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub with_escalated_permissions: Option<bool>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub justification: Option<String>,
 }
 
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, PartialEq)]
 pub struct FunctionCallOutputPayload {
     pub content: String,
-    #[expect(dead_code)]
     pub success: Option<bool>,
 }
 
@@ -295,6 +306,8 @@ mod tests {
                 command: vec!["ls".to_string(), "-l".to_string()],
                 workdir: Some("/tmp".to_string()),
                 timeout_ms: Some(1000),
+                with_escalated_permissions: None,
+                justification: None,
             },
             params
         );

codex-rs/core/src/openai_api_key.rs

@@ -1,24 +0,0 @@
-use std::env;
-use std::sync::LazyLock;
-use std::sync::RwLock;
-
-pub const OPENAI_API_KEY_ENV_VAR: &str = "OPENAI_API_KEY";
-
-static OPENAI_API_KEY: LazyLock<RwLock<Option<String>>> = LazyLock::new(|| {
-    let val = env::var(OPENAI_API_KEY_ENV_VAR)
-        .ok()
-        .and_then(|s| if s.is_empty() { None } else { Some(s) });
-    RwLock::new(val)
-});
-
-pub fn get_openai_api_key() -> Option<String> {
-    #![allow(clippy::unwrap_used)]
-    OPENAI_API_KEY.read().unwrap().clone()
-}
-
-pub fn set_openai_api_key(value: String) {
-    #![allow(clippy::unwrap_used)]
-    if !value.is_empty() {
-        *OPENAI_API_KEY.write().unwrap() = Some(value);
-    }
-}

codex-rs/core/src/openai_model_info.rs

@@ -1,3 +1,5 @@
+use crate::model_family::ModelFamily;
+
 /// Metadata about a model, particularly OpenAI models.
 /// We may want to consider including details like the pricing for
 /// input tokens, output tokens, etc., though users will need to be able to
@@ -12,10 +14,19 @@ pub(crate) struct ModelInfo {
     pub(crate) max_output_tokens: u64,
 }
 
-/// Note details such as what a model like gpt-4o is aliased to may be out of
-/// date.
-pub(crate) fn get_model_info(name: &str) -> Option<ModelInfo> {
-    match name {
+pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
+    match model_family.slug.as_str() {
+        // 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,
+        }),
         // https://platform.openai.com/docs/models/o3
         "o3" => Some(ModelInfo {
             context_window: 200_000,

codex-rs/core/src/openai_tools.rs

@@ -1,21 +1,28 @@
+use serde::Deserialize;
 use serde::Serialize;
 use serde_json::json;
 use std::collections::BTreeMap;
-use std::sync::LazyLock;
+use std::collections::HashMap;
 
-use crate::client_common::Prompt;
+use crate::model_family::ModelFamily;
+use crate::plan_tool::PLAN_TOOL;
+use crate::protocol::AskForApproval;
+use crate::protocol::SandboxPolicy;
 
-#[derive(Debug, Clone, Serialize)]
-pub(crate) struct ResponsesApiTool {
-    name: &'static str,
-    description: &'static str,
-    strict: bool,
-    parameters: JsonSchema,
+#[derive(Debug, Clone, Serialize, PartialEq)]
+pub struct ResponsesApiTool {
+    pub(crate) name: String,
+    pub(crate) description: String,
+    /// TODO: Validation. When strict is set to true, the JSON schema,
+    /// `required` and `additional_properties` must be present. All fields in
+    /// `properties` must be present in `required`.
+    pub(crate) strict: bool,
+    pub(crate) parameters: JsonSchema,
 }
 
 /// When serialized as JSON, this produces a valid "Tool" in the OpenAI
 /// Responses API.
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone, Serialize, PartialEq)]
 #[serde(tag = "type")]
 pub(crate) enum OpenAiTool {
     #[serde(rename = "function")]
@@ -24,74 +31,218 @@ pub(crate) enum OpenAiTool {
     LocalShell {},
 }
 
+#[derive(Debug, Clone)]
+pub enum ConfigShellToolType {
+    DefaultShell,
+    ShellWithRequest { sandbox_policy: SandboxPolicy },
+    LocalShell,
+}
+
+#[derive(Debug, Clone)]
+pub struct ToolsConfig {
+    pub shell_type: ConfigShellToolType,
+    pub plan_tool: bool,
+}
+
+impl ToolsConfig {
+    pub fn new(
+        model_family: &ModelFamily,
+        approval_policy: AskForApproval,
+        sandbox_policy: SandboxPolicy,
+        include_plan_tool: bool,
+    ) -> Self {
+        let mut shell_type = if model_family.uses_local_shell_tool {
+            ConfigShellToolType::LocalShell
+        } else {
+            ConfigShellToolType::DefaultShell
+        };
+        if matches!(approval_policy, AskForApproval::OnRequest) {
+            shell_type = ConfigShellToolType::ShellWithRequest {
+                sandbox_policy: sandbox_policy.clone(),
+            }
+        }
+
+        Self {
+            shell_type,
+            plan_tool: include_plan_tool,
+        }
+    }
+}
+
 /// Generic JSON‑Schema subset needed for our tool definitions
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
 #[serde(tag = "type", rename_all = "lowercase")]
 pub(crate) enum JsonSchema {
-    String,
-    Number,
+    Boolean {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        description: Option<String>,
+    },
+    String {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        description: Option<String>,
+    },
+    Number {
+        #[serde(skip_serializing_if = "Option::is_none")]
+        description: Option<String>,
+    },
     Array {
         items: Box<JsonSchema>,
+
+        #[serde(skip_serializing_if = "Option::is_none")]
+        description: Option<String>,
     },
     Object {
         properties: BTreeMap<String, JsonSchema>,
-        required: &'static [&'static str],
-        #[serde(rename = "additionalProperties")]
-        additional_properties: bool,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        required: Option<Vec<String>>,
+        #[serde(
+            rename = "additionalProperties",
+            skip_serializing_if = "Option::is_none"
+        )]
+        additional_properties: Option<bool>,
     },
 }
 
-/// Tool usage specification
-static DEFAULT_TOOLS: LazyLock<Vec<OpenAiTool>> = LazyLock::new(|| {
+fn create_shell_tool() -> OpenAiTool {
     let mut properties = BTreeMap::new();
     properties.insert(
         "command".to_string(),
         JsonSchema::Array {
-            items: Box::new(JsonSchema::String),
+            items: Box::new(JsonSchema::String { description: None }),
+            description: None,
         },
     );
-    properties.insert("workdir".to_string(), JsonSchema::String);
-    properties.insert("timeout".to_string(), JsonSchema::Number);
+    properties.insert(
+        "workdir".to_string(),
+        JsonSchema::String { description: None },
+    );
+    properties.insert(
+        "timeout".to_string(),
+        JsonSchema::Number { description: None },
+    );
 
-    vec![OpenAiTool::Function(ResponsesApiTool {
-        name: "shell",
-        description: "Runs a shell command, and returns its output.",
+    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: &["command"],
-            additional_properties: false,
+            required: Some(vec!["command".to_string()]),
+            additional_properties: Some(false),
         },
-    })]
-});
+    })
+}
 
-static DEFAULT_CODEX_MODEL_TOOLS: LazyLock<Vec<OpenAiTool>> =
-    LazyLock::new(|| vec![OpenAiTool::LocalShell {}]);
+fn create_shell_tool_for_sandbox(sandbox_policy: &SandboxPolicy) -> 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".to_string(),
+        JsonSchema::Number {
+            description: Some("The timeout for the command in milliseconds".to_string()),
+        },
+    );
+
+    if matches!(sandbox_policy, SandboxPolicy::WorkspaceWrite { .. }) {
+        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(
+        "justification".to_string(),
+        JsonSchema::String {
+            description: Some("Only set if ask_for_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,
+            ..
+        } => {
+            format!(
+                r#"
+The shell tool is used to execute shell commands.
+- When invoking the shell tool, your call will be running in a landlock sandbox, and some shell commands will require escalated privileges:
+  - Types of actions that require escalated privileges:
+    - Reading files outside the current directory
+    - Writing files outside the current directory, and protected folders like .git or .env{}
+  - Examples of commands that require escalated privileges:
+    - git commit
+    - npm install or pnpm install
+    - cargo build
+    - cargo test
+- When invoking a command that will require escalated privileges:
+  - Provide the with_escalated_permissions parameter with the boolean value true
+  - Include a short, 1 sentence explanation for why we need to run with_escalated_permissions in the justification parameter."#,
+                if !network_access {
+                    "\n  - Commands that require network access\n"
+                } else {
+                    ""
+                }
+            )
+        }
+        SandboxPolicy::DangerFullAccess => {
+            "Runs a shell command and returns its output.".to_string()
+        }
+        SandboxPolicy::ReadOnly => {
+            r#"
+The shell tool is used to execute shell commands.
+- When invoking the shell tool, your call will be running in a landlock sandbox, and some shell commands (including apply_patch) will require escalated permissions:
+  - Types of actions that require escalated privileges:
+    - Reading files outside the current directory
+    - Writing files
+    - Applying patches
+  - Examples of commands that require escalated privileges:
+    - apply_patch
+    - git commit
+    - npm install or pnpm install
+    - cargo build
+    - cargo test
+- When invoking a command that will require escalated privileges:
+  - Provide the with_escalated_permissions parameter with the boolean value true
+  - Include a short, 1 sentence explanation for why we need to run with_escalated_permissions in the justification parameter"#.to_string()
+        }
+    };
+
+    OpenAiTool::Function(ResponsesApiTool {
+        name: "shell".to_string(),
+        description,
+        strict: false,
+        parameters: JsonSchema::Object {
+            properties,
+            required: Some(vec!["command".to_string()]),
+            additional_properties: Some(false),
+        },
+    })
+}
 
 /// Returns JSON values that are compatible with Function Calling in the
 /// Responses API:
 /// https://platform.openai.com/docs/guides/function-calling?api-mode=responses
 pub(crate) fn create_tools_json_for_responses_api(
-    prompt: &Prompt,
-    model: &str,
+    tools: &Vec<OpenAiTool>,
 ) -> crate::error::Result<Vec<serde_json::Value>> {
-    // Assemble tool list: built-in tools + any extra tools from the prompt.
-    let default_tools = if model.starts_with("codex") {
-        &DEFAULT_CODEX_MODEL_TOOLS
-    } else {
-        &DEFAULT_TOOLS
-    };
-    let mut tools_json = Vec::with_capacity(default_tools.len() + prompt.extra_tools.len());
-    for t in default_tools.iter() {
-        tools_json.push(serde_json::to_value(t)?);
+    let mut tools_json = Vec::new();
+
+    for tool in tools {
+        tools_json.push(serde_json::to_value(tool)?);
     }
-    tools_json.extend(
-        prompt
-            .extra_tools
-            .clone()
-            .into_iter()
-            .map(|(name, tool)| mcp_tool_to_openai_tool(name, tool)),
-    );
 
     Ok(tools_json)
 }
@@ -100,12 +251,11 @@ pub(crate) 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(
-    prompt: &Prompt,
-    model: &str,
+    tools: &Vec<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.
-    let responses_api_tools_json = create_tools_json_for_responses_api(prompt, model)?;
+    let responses_api_tools_json = create_tools_json_for_responses_api(tools)?;
     let tools_json = responses_api_tools_json
         .into_iter()
         .filter_map(|mut tool| {
@@ -128,10 +278,10 @@ pub(crate) fn create_tools_json_for_chat_completions_api(
     Ok(tools_json)
 }
 
-fn mcp_tool_to_openai_tool(
+pub(crate) fn mcp_tool_to_openai_tool(
     fully_qualified_name: String,
     tool: mcp_types::Tool,
-) -> serde_json::Value {
+) -> Result<ResponsesApiTool, serde_json::Error> {
     let mcp_types::Tool {
         description,
         mut input_schema,
@@ -146,12 +296,205 @@ fn mcp_tool_to_openai_tool(
         input_schema.properties = Some(serde_json::Value::Object(serde_json::Map::new()));
     }
 
-    // TODO(mbolin): Change the contract of this function to return
-    // ResponsesApiTool.
-    json!({
-        "name": fully_qualified_name,
-        "description": description,
-        "parameters": input_schema,
-        "type": "function",
+    let serialized_input_schema = serde_json::to_value(input_schema)?;
+    let input_schema = serde_json::from_value::<JsonSchema>(serialized_input_schema)?;
+
+    Ok(ResponsesApiTool {
+        name: fully_qualified_name,
+        description: description.unwrap_or_default(),
+        strict: false,
+        parameters: input_schema,
     })
 }
+
+/// Returns a list of OpenAiTools based on the provided config and MCP tools.
+/// Note that the keys of mcp_tools should be fully qualified names. See
+/// [`McpConnectionManager`] for more details.
+pub(crate) fn get_openai_tools(
+    config: &ToolsConfig,
+    mcp_tools: Option<HashMap<String, mcp_types::Tool>>,
+) -> Vec<OpenAiTool> {
+    let mut tools: Vec<OpenAiTool> = Vec::new();
+
+    match &config.shell_type {
+        ConfigShellToolType::DefaultShell => {
+            tools.push(create_shell_tool());
+        }
+        ConfigShellToolType::ShellWithRequest { sandbox_policy } => {
+            tools.push(create_shell_tool_for_sandbox(sandbox_policy));
+        }
+        ConfigShellToolType::LocalShell => {
+            tools.push(OpenAiTool::LocalShell {});
+        }
+    }
+
+    if config.plan_tool {
+        tools.push(PLAN_TOOL.clone());
+    }
+
+    if let Some(mcp_tools) = mcp_tools {
+        for (name, tool) in mcp_tools {
+            match mcp_tool_to_openai_tool(name.clone(), tool.clone()) {
+                Ok(converted_tool) => tools.push(OpenAiTool::Function(converted_tool)),
+                Err(e) => {
+                    tracing::error!("Failed to convert {name:?} MCP tool to OpenAI tool: {e:?}");
+                }
+            }
+        }
+    }
+
+    tools
+}
+
+#[cfg(test)]
+#[allow(clippy::expect_used)]
+mod tests {
+    use crate::model_family::find_family_for_model;
+    use mcp_types::ToolInputSchema;
+
+    use super::*;
+
+    fn assert_eq_tool_names(tools: &[OpenAiTool], expected_names: &[&str]) {
+        let tool_names = tools
+            .iter()
+            .map(|tool| match tool {
+                OpenAiTool::Function(ResponsesApiTool { name, .. }) => name,
+                OpenAiTool::LocalShell {} => "local_shell",
+            })
+            .collect::<Vec<_>>();
+
+        assert_eq!(
+            tool_names.len(),
+            expected_names.len(),
+            "tool_name mismatch, {tool_names:?}, {expected_names:?}",
+        );
+        for (name, expected_name) in tool_names.iter().zip(expected_names.iter()) {
+            assert_eq!(
+                name, expected_name,
+                "tool_name mismatch, {name:?}, {expected_name:?}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_get_openai_tools() {
+        let model_family = find_family_for_model("codex-mini-latest")
+            .expect("codex-mini-latest should be a valid model family");
+        let config = ToolsConfig::new(
+            &model_family,
+            AskForApproval::Never,
+            SandboxPolicy::ReadOnly,
+            true,
+        );
+        let tools = get_openai_tools(&config, Some(HashMap::new()));
+
+        assert_eq_tool_names(&tools, &["local_shell", "update_plan"]);
+    }
+
+    #[test]
+    fn test_get_openai_tools_default_shell() {
+        let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
+        let config = ToolsConfig::new(
+            &model_family,
+            AskForApproval::Never,
+            SandboxPolicy::ReadOnly,
+            true,
+        );
+        let tools = get_openai_tools(&config, Some(HashMap::new()));
+
+        assert_eq_tool_names(&tools, &["shell", "update_plan"]);
+    }
+
+    #[test]
+    fn test_get_openai_tools_mcp_tools() {
+        let model_family = find_family_for_model("o3").expect("o3 should be a valid model family");
+        let config = ToolsConfig::new(
+            &model_family,
+            AskForApproval::Never,
+            SandboxPolicy::ReadOnly,
+            false,
+        );
+        let tools = get_openai_tools(
+            &config,
+            Some(HashMap::from([(
+                "test_server/do_something_cool".to_string(),
+                mcp_types::Tool {
+                    name: "do_something_cool".to_string(),
+                    input_schema: ToolInputSchema {
+                        properties: Some(serde_json::json!({
+                            "string_argument": {
+                                "type": "string",
+                            },
+                            "number_argument": {
+                                "type": "number",
+                            },
+                            "object_argument": {
+                                "type": "object",
+                                "properties": {
+                                    "string_property": { "type": "string" },
+                                    "number_property": { "type": "number" },
+                                },
+                                "required": [
+                                    "string_property",
+                                    "number_property"
+                                ],
+                                "additionalProperties": Some(false),
+                            },
+                        })),
+                        required: None,
+                        r#type: "object".to_string(),
+                    },
+                    output_schema: None,
+                    title: None,
+                    annotations: None,
+                    description: Some("Do something cool".to_string()),
+                },
+            )])),
+        );
+
+        assert_eq_tool_names(&tools, &["shell", "test_server/do_something_cool"]);
+
+        assert_eq!(
+            tools[1],
+            OpenAiTool::Function(ResponsesApiTool {
+                name: "test_server/do_something_cool".to_string(),
+                parameters: JsonSchema::Object {
+                    properties: BTreeMap::from([
+                        (
+                            "string_argument".to_string(),
+                            JsonSchema::String { description: None }
+                        ),
+                        (
+                            "number_argument".to_string(),
+                            JsonSchema::Number { description: None }
+                        ),
+                        (
+                            "object_argument".to_string(),
+                            JsonSchema::Object {
+                                properties: BTreeMap::from([
+                                    (
+                                        "string_property".to_string(),
+                                        JsonSchema::String { description: None }
+                                    ),
+                                    (
+                                        "number_property".to_string(),
+                                        JsonSchema::Number { description: None }
+                                    ),
+                                ]),
+                                required: Some(vec![
+                                    "string_property".to_string(),
+                                    "number_property".to_string(),
+                                ]),
+                                additional_properties: Some(false),
+                            },
+                        ),
+                    ]),
+                    required: None,
+                    additional_properties: None,
+                },
+                description: "Do something cool".to_string(),
+                strict: false,
+            })
+        );
+    }
+}

codex-rs/core/src/plan_tool.rs

@@ -0,0 +1,133 @@
+use std::collections::BTreeMap;
+use std::sync::LazyLock;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::codex::Session;
+use crate::models::FunctionCallOutputPayload;
+use crate::models::ResponseInputItem;
+use crate::openai_tools::JsonSchema;
+use crate::openai_tools::OpenAiTool;
+use crate::openai_tools::ResponsesApiTool;
+use crate::protocol::Event;
+use crate::protocol::EventMsg;
+
+// Types for the TODO tool arguments matching codex-vscode/todo-mcp/src/main.rs
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum StepStatus {
+    Pending,
+    InProgress,
+    Completed,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct PlanItemArg {
+    pub step: String,
+    pub status: StepStatus,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UpdatePlanArgs {
+    #[serde(default)]
+    pub explanation: Option<String>,
+    pub plan: Vec<PlanItemArg>,
+}
+
+pub(crate) static PLAN_TOOL: LazyLock<OpenAiTool> = LazyLock::new(|| {
+    let mut plan_item_props = BTreeMap::new();
+    plan_item_props.insert("step".to_string(), JsonSchema::String { description: None });
+    plan_item_props.insert(
+        "status".to_string(),
+        JsonSchema::String { description: None },
+    );
+
+    let plan_items_schema = JsonSchema::Array {
+        description: Some("The list of steps".to_string()),
+        items: Box::new(JsonSchema::Object {
+            properties: plan_item_props,
+            required: Some(vec!["step".to_string(), "status".to_string()]),
+            additional_properties: Some(false),
+        }),
+    };
+
+    let mut properties = BTreeMap::new();
+    properties.insert(
+        "explanation".to_string(),
+        JsonSchema::String { description: None },
+    );
+    properties.insert("plan".to_string(), plan_items_schema);
+
+    OpenAiTool::Function(ResponsesApiTool {
+        name: "update_plan".to_string(),
+        description: r#"Use the update_plan tool to keep the user updated on the current plan for the task.
+After understanding the user's task, call the update_plan tool with an initial plan. An example of a plan:
+1. Explore the codebase to find relevant files (status: in_progress)
+2. Implement the feature in the XYZ component (status: pending)
+3. Commit changes and make a pull request (status: pending)
+Each step should be a short, 1-sentence description.
+Until all the steps are finished, there should always be exactly one in_progress step in the plan.
+Call the update_plan tool whenever you finish a step, marking the completed step as `completed` and marking the next step as `in_progress`.
+Before running a command, consider whether or not you have completed the previous step, and make sure to mark it as completed before moving on to the next step.
+Sometimes, you may need to change plans in the middle of a task: call `update_plan` with the updated plan and make sure to provide an `explanation` of the rationale when doing so.
+When all steps are completed, call update_plan one last time with all steps marked as `completed`."#.to_string(),
+        strict: false,
+        parameters: JsonSchema::Object {
+            properties,
+            required: Some(vec!["plan".to_string()]),
+            additional_properties: Some(false),
+        },
+    })
+});
+
+/// This function doesn't do anything useful. However, it gives the model a structured way to record its plan that clients can read and render.
+/// So it's the _inputs_ to this function that are useful to clients, not the outputs and neither are actually useful for the model other
+/// than forcing it to come up and document a plan (TBD how that affects performance).
+pub(crate) async fn handle_update_plan(
+    session: &Session,
+    arguments: String,
+    sub_id: String,
+    call_id: String,
+) -> ResponseInputItem {
+    match parse_update_plan_arguments(arguments, &call_id) {
+        Ok(args) => {
+            let output = ResponseInputItem::FunctionCallOutput {
+                call_id,
+                output: FunctionCallOutputPayload {
+                    content: "Plan updated".to_string(),
+                    success: Some(true),
+                },
+            };
+            session
+                .send_event(Event {
+                    id: sub_id.to_string(),
+                    msg: EventMsg::PlanUpdate(args),
+                })
+                .await;
+            output
+        }
+        Err(output) => *output,
+    }
+}
+
+fn parse_update_plan_arguments(
+    arguments: String,
+    call_id: &str,
+) -> Result<UpdatePlanArgs, Box<ResponseInputItem>> {
+    match serde_json::from_str::<UpdatePlanArgs>(&arguments) {
+        Ok(args) => Ok(args),
+        Err(e) => {
+            let output = ResponseInputItem::FunctionCallOutput {
+                call_id: call_id.to_string(),
+                output: FunctionCallOutputPayload {
+                    content: format!("failed to parse function arguments: {e}"),
+                    success: None,
+                },
+            };
+            Err(Box::new(output))
+        }
+    }
+}

codex-rs/core/src/protocol.rs

@@ -7,11 +7,13 @@ use std::collections::HashMap;
 use std::fmt;
 use std::path::Path;
 use std::path::PathBuf;
-use std::str::FromStr; // Added for FinalOutput Display implementation
+use std::str::FromStr;
+use std::time::Duration;
 
 use mcp_types::CallToolResult;
 use serde::Deserialize;
 use serde::Serialize;
+use serde_bytes::ByteBuf;
 use strum_macros::Display;
 use uuid::Uuid;
 
@@ -19,6 +21,7 @@ use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
 use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::message_history::HistoryEntry;
 use crate::model_provider_info::ModelProviderInfo;
+use crate::plan_tool::UpdatePlanArgs;
 
 /// Submission Queue Entry - requests from user
 #[derive(Debug, Clone, Deserialize, Serialize)]
@@ -119,6 +122,10 @@ pub enum Op {
     /// Request a single history entry identified by `log_id` + `offset`.
     GetHistoryEntryRequest { offset: usize, log_id: u64 },
 
+    /// Request the agent to summarize the current conversation context.
+    /// The agent will use its existing context (either conversation history or previous response id)
+    /// to generate a summary which will be returned as an AgentMessage event.
+    Compact,
     /// Request to shut down codex instance.
     Shutdown,
 }
@@ -143,13 +150,17 @@ pub enum AskForApproval {
     /// the user to approve execution without a sandbox.
     OnFailure,
 
+    /// The model decides when to ask the user for approval.
+    OnRequest,
+
     /// Never ask the user to approve commands. Failures are immediately returned
     /// to the model, and never escalated to the user for approval.
     Never,
 }
 
 /// Determines execution restrictions for model shell commands.
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Display)]
+#[strum(serialize_all = "kebab-case")]
 #[serde(tag = "mode", rename_all = "kebab-case")]
 pub enum SandboxPolicy {
     /// No restrictions whatsoever. Use with caution.
@@ -173,9 +184,29 @@ pub enum SandboxPolicy {
         /// default.
         #[serde(default)]
         network_access: bool,
+
+        /// When set to `true`, will include defaults like the current working
+        /// directory and TMPDIR (on macOS). When `false`, only `writable_roots`
+        /// are used. (Mainly used for testing.)
+        #[serde(default = "default_true")]
+        include_default_writable_roots: bool,
     },
 }
 
+/// A writable root path accompanied by a list of subpaths that should remain
+/// read‑only even when the root is writable. This is primarily used to ensure
+/// top‑level VCS metadata directories (e.g. `.git`) under a writable root are
+/// not modified by the agent.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct WritableRoot {
+    pub root: PathBuf,
+    pub read_only_subpaths: Vec<PathBuf>,
+}
+
+fn default_true() -> bool {
+    true
+}
+
 impl FromStr for SandboxPolicy {
     type Err = serde_json::Error;
 
@@ -197,6 +228,7 @@ impl SandboxPolicy {
         SandboxPolicy::WorkspaceWrite {
             writable_roots: vec![],
             network_access: false,
+            include_default_writable_roots: true,
         }
     }
 
@@ -222,27 +254,51 @@ impl SandboxPolicy {
         }
     }
 
-    /// Returns the list of writable roots that should be passed down to the
-    /// Landlock rules installer, tailored to the current working directory.
-    pub fn get_writable_roots_with_cwd(&self, cwd: &Path) -> Vec<PathBuf> {
+    /// Returns the list of writable roots (tailored to the current working
+    /// directory) together with subpaths that should remain read‑only under
+    /// each writable root.
+    pub fn get_writable_roots_with_cwd(&self, cwd: &Path) -> Vec<WritableRoot> {
         match self {
             SandboxPolicy::DangerFullAccess => Vec::new(),
             SandboxPolicy::ReadOnly => Vec::new(),
-            SandboxPolicy::WorkspaceWrite { writable_roots, .. } => {
-                let mut roots = writable_roots.clone();
-                roots.push(cwd.to_path_buf());
+            SandboxPolicy::WorkspaceWrite {
+                writable_roots,
+                include_default_writable_roots,
+                ..
+            } => {
+                // Start from explicitly configured writable roots.
+                let mut roots: Vec<PathBuf> = writable_roots.clone();
 
-                // Also include the per-user tmp dir on macOS.
-                // Note this is added dynamically rather than storing it in
-                // writable_roots because writable_roots contains only static
-                // values deserialized from the config file.
-                if cfg!(target_os = "macos") {
-                    if let Some(tmpdir) = std::env::var_os("TMPDIR") {
-                        roots.push(PathBuf::from(tmpdir));
+                // Optionally include defaults (cwd and TMPDIR on macOS).
+                if *include_default_writable_roots {
+                    roots.push(cwd.to_path_buf());
+
+                    // Also include the per-user tmp dir on macOS.
+                    // Note this is added dynamically rather than storing it in
+                    // `writable_roots` because `writable_roots` contains only static
+                    // values deserialized from the config file.
+                    if cfg!(target_os = "macos") {
+                        if let Some(tmpdir) = std::env::var_os("TMPDIR") {
+                            roots.push(PathBuf::from(tmpdir));
+                        }
                     }
                 }
 
+                // For each root, compute subpaths that should remain read-only.
                 roots
+                    .into_iter()
+                    .map(|writable_root| {
+                        let mut subpaths = Vec::new();
+                        let top_level_git = writable_root.join(".git");
+                        if top_level_git.is_dir() {
+                            subpaths.push(top_level_git);
+                        }
+                        WritableRoot {
+                            root: writable_root,
+                            read_only_subpaths: subpaths,
+                        }
+                    })
+                    .collect()
             }
         }
     }
@@ -307,6 +363,12 @@ pub enum EventMsg {
     /// Agent reasoning delta event from agent.
     AgentReasoningDelta(AgentReasoningDeltaEvent),
 
+    /// Raw chain-of-thought from agent.
+    AgentReasoningRawContent(AgentReasoningRawContentEvent),
+
+    /// Agent reasoning content delta event from agent.
+    AgentReasoningRawContentDelta(AgentReasoningRawContentDeltaEvent),
+
     /// Ack the client's configure message.
     SessionConfigured(SessionConfiguredEvent),
 
@@ -317,6 +379,9 @@ pub enum EventMsg {
     /// Notification that the server is about to execute a command.
     ExecCommandBegin(ExecCommandBeginEvent),
 
+    /// Incremental chunk of output from a running command.
+    ExecCommandOutputDelta(ExecCommandOutputDeltaEvent),
+
     ExecCommandEnd(ExecCommandEndEvent),
 
     ExecApprovalRequest(ExecApprovalRequestEvent),
@@ -332,9 +397,13 @@ pub enum EventMsg {
     /// Notification that a patch application has finished.
     PatchApplyEnd(PatchApplyEndEvent),
 
+    TurnDiff(TurnDiffEvent),
+
     /// Response to GetHistoryEntryRequest.
     GetHistoryEntryResponse(GetHistoryEntryResponseEvent),
 
+    PlanUpdate(UpdatePlanArgs),
+
     /// Notification that the agent is shutting down.
     ShutdownComplete,
 }
@@ -405,15 +474,23 @@ pub struct AgentReasoningEvent {
     pub text: String,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct AgentReasoningRawContentEvent {
+    pub text: String,
+}
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct AgentReasoningRawContentDeltaEvent {
+    pub delta: String,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct AgentReasoningDeltaEvent {
     pub delta: String,
 }
 
 #[derive(Debug, Clone, Deserialize, Serialize)]
-pub struct McpToolCallBeginEvent {
-    /// Identifier so this can be paired with the McpToolCallEnd event.
-    pub call_id: String,
+pub struct McpInvocation {
     /// Name of the MCP server as defined in the config.
     pub server: String,
     /// Name of the tool as given by the MCP server.
@@ -422,10 +499,19 @@ pub struct McpToolCallBeginEvent {
     pub arguments: Option<serde_json::Value>,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct McpToolCallBeginEvent {
+    /// Identifier so this can be paired with the McpToolCallEnd event.
+    pub call_id: String,
+    pub invocation: McpInvocation,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct McpToolCallEndEvent {
     /// Identifier for the corresponding McpToolCallBegin that finished.
     pub call_id: String,
+    pub invocation: McpInvocation,
+    pub duration: Duration,
     /// Result of the tool call. Note this could be an error.
     pub result: Result<CallToolResult, String>,
 }
@@ -459,6 +545,26 @@ pub struct ExecCommandEndEvent {
     pub stderr: String,
     /// The command's exit code.
     pub exit_code: i32,
+    /// The duration of the command execution.
+    pub duration: Duration,
+}
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum ExecOutputStream {
+    Stdout,
+    Stderr,
+}
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct ExecCommandOutputDeltaEvent {
+    /// Identifier for the ExecCommandBegin that produced this chunk.
+    pub call_id: String,
+    /// Which stream produced this chunk.
+    pub stream: ExecOutputStream,
+    /// Raw bytes from the stream (may not be valid UTF-8).
+    #[serde(with = "serde_bytes")]
+    pub chunk: ByteBuf,
 }
 
 #[derive(Debug, Clone, Deserialize, Serialize)]
@@ -514,6 +620,11 @@ pub struct PatchApplyEndEvent {
     pub success: bool,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct TurnDiffEvent {
+    pub unified_diff: String,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct GetHistoryEntryResponseEvent {
     pub offset: usize,

codex-rs/core/src/safety.rs

@@ -11,7 +11,7 @@ use crate::is_safe_command::is_known_safe_command;
 use crate::protocol::AskForApproval;
 use crate::protocol::SandboxPolicy;
 
-#[derive(Debug)]
+#[derive(Debug, PartialEq)]
 pub enum SafetyCheck {
     AutoApprove { sandbox_type: SandboxType },
     AskUser,
@@ -31,7 +31,7 @@ pub fn assess_patch_safety(
     }
 
     match policy {
-        AskForApproval::OnFailure | AskForApproval::Never => {
+        AskForApproval::OnFailure | AskForApproval::Never | AskForApproval::OnRequest => {
             // Continue to see if this can be auto-approved.
         }
         // TODO(ragona): I'm not sure this is actually correct? I believe in this case
@@ -41,11 +41,13 @@ pub fn assess_patch_safety(
         }
     }
 
-    if is_write_patch_constrained_to_writable_paths(action, writable_roots, cwd) {
-        SafetyCheck::AutoApprove {
-            sandbox_type: SandboxType::None,
-        }
-    } else if policy == AskForApproval::OnFailure {
+    // Even though the patch *appears* to be constrained to writable paths, it
+    // is possible that paths in the patch are hard links to files outside the
+    // writable roots, so we should still run `apply_patch` in a sandbox in that
+    // case.
+    if is_write_patch_constrained_to_writable_paths(action, writable_roots, cwd)
+        || policy == AskForApproval::OnFailure
+    {
         // Only auto‑approve when we can actually enforce a sandbox. Otherwise
         // fall back to asking the user because the patch may touch arbitrary
         // paths outside the project.
@@ -74,10 +76,8 @@ pub fn assess_command_safety(
     approval_policy: AskForApproval,
     sandbox_policy: &SandboxPolicy,
     approved: &HashSet<Vec<String>>,
+    with_escalated_permissions: bool,
 ) -> SafetyCheck {
-    use AskForApproval::*;
-    use SandboxPolicy::*;
-
     // A command is "trusted" because either:
     // - it belongs to a set of commands we consider "safe" by default, or
     // - the user has explicitly approved the command for this session
@@ -97,6 +97,17 @@ pub fn assess_command_safety(
         };
     }
 
+    assess_safety_for_untrusted_command(approval_policy, sandbox_policy, with_escalated_permissions)
+}
+
+pub(crate) fn assess_safety_for_untrusted_command(
+    approval_policy: AskForApproval,
+    sandbox_policy: &SandboxPolicy,
+    with_escalated_permissions: bool,
+) -> SafetyCheck {
+    use AskForApproval::*;
+    use SandboxPolicy::*;
+
     match (approval_policy, sandbox_policy) {
         (UnlessTrusted, _) => {
             // Even though the user may have opted into DangerFullAccess,
@@ -104,9 +115,23 @@ pub fn assess_command_safety(
             // commands.
             SafetyCheck::AskUser
         }
-        (OnFailure, DangerFullAccess) | (Never, DangerFullAccess) => SafetyCheck::AutoApprove {
+        (OnFailure, DangerFullAccess)
+        | (Never, DangerFullAccess)
+        | (OnRequest, DangerFullAccess) => SafetyCheck::AutoApprove {
             sandbox_type: SandboxType::None,
         },
+        (OnRequest, ReadOnly) | (OnRequest, WorkspaceWrite { .. }) => {
+            if with_escalated_permissions {
+                SafetyCheck::AskUser
+            } else {
+                match get_platform_sandbox() {
+                    Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+                    // Fall back to asking since the command is untrusted and
+                    // we do not have a sandbox available
+                    None => SafetyCheck::AskUser,
+                }
+            }
+        }
         (Never, ReadOnly)
         | (Never, WorkspaceWrite { .. })
         | (OnFailure, ReadOnly)
@@ -255,4 +280,47 @@ mod tests {
             &cwd,
         ))
     }
+
+    #[test]
+    fn test_request_escalated_privileges() {
+        // Should not be a trusted command
+        let command = vec!["git commit".to_string()];
+        let approval_policy = AskForApproval::OnRequest;
+        let sandbox_policy = SandboxPolicy::ReadOnly;
+        let approved: HashSet<Vec<String>> = HashSet::new();
+        let request_escalated_privileges = true;
+
+        let safety_check = assess_command_safety(
+            &command,
+            approval_policy,
+            &sandbox_policy,
+            &approved,
+            request_escalated_privileges,
+        );
+
+        assert_eq!(safety_check, SafetyCheck::AskUser);
+    }
+
+    #[test]
+    fn test_request_escalated_privileges_no_sandbox_fallback() {
+        let command = vec!["git".to_string(), "commit".to_string()];
+        let approval_policy = AskForApproval::OnRequest;
+        let sandbox_policy = SandboxPolicy::ReadOnly;
+        let approved: HashSet<Vec<String>> = HashSet::new();
+        let request_escalated_privileges = false;
+
+        let safety_check = assess_command_safety(
+            &command,
+            approval_policy,
+            &sandbox_policy,
+            &approved,
+            request_escalated_privileges,
+        );
+
+        let expected = match get_platform_sandbox() {
+            Some(sandbox_type) => SafetyCheck::AutoApprove { sandbox_type },
+            None => SafetyCheck::AskUser,
+        };
+        assert_eq!(safety_check, expected);
+    }
 }

codex-rs/core/src/seatbelt.rs

@@ -0,0 +1,312 @@
+use std::collections::HashMap;
+use std::path::Path;
+use std::path::PathBuf;
+use tokio::process::Child;
+
+use crate::protocol::SandboxPolicy;
+use crate::spawn::CODEX_SANDBOX_ENV_VAR;
+use crate::spawn::StdioPolicy;
+use crate::spawn::spawn_child_async;
+
+const MACOS_SEATBELT_BASE_POLICY: &str = include_str!("seatbelt_base_policy.sbpl");
+
+/// When working with `sandbox-exec`, only consider `sandbox-exec` in `/usr/bin`
+/// to defend against an attacker trying to inject a malicious version on the
+/// PATH. If /usr/bin/sandbox-exec has been tampered with, then the attacker
+/// already has root access.
+const MACOS_PATH_TO_SEATBELT_EXECUTABLE: &str = "/usr/bin/sandbox-exec";
+
+pub async fn spawn_command_under_seatbelt(
+    command: Vec<String>,
+    sandbox_policy: &SandboxPolicy,
+    cwd: PathBuf,
+    stdio_policy: StdioPolicy,
+    mut env: HashMap<String, String>,
+) -> std::io::Result<Child> {
+    let args = create_seatbelt_command_args(command, 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,
+        sandbox_policy,
+        stdio_policy,
+        env,
+    )
+    .await
+}
+
+fn create_seatbelt_command_args(
+    command: Vec<String>,
+    sandbox_policy: &SandboxPolicy,
+    cwd: &Path,
+) -> Vec<String> {
+    let (file_write_policy, extra_cli_args) = {
+        if sandbox_policy.has_full_disk_write_access() {
+            // Allegedly, this is more permissive than `(allow file-write*)`.
+            (
+                r#"(allow file-write* (regex #"^/"))"#.to_string(),
+                Vec::<String>::new(),
+            )
+        } else {
+            let writable_roots = sandbox_policy.get_writable_roots_with_cwd(cwd);
+
+            let mut writable_folder_policies: Vec<String> = Vec::new();
+            let mut cli_args: Vec<String> = Vec::new();
+
+            for (index, wr) in writable_roots.iter().enumerate() {
+                // Canonicalize to avoid mismatches like /var vs /private/var on macOS.
+                let canonical_root = wr.root.canonicalize().unwrap_or_else(|_| wr.root.clone());
+                let root_param = format!("WRITABLE_ROOT_{index}");
+                cli_args.push(format!(
+                    "-D{root_param}={}",
+                    canonical_root.to_string_lossy()
+                ));
+
+                if wr.read_only_subpaths.is_empty() {
+                    writable_folder_policies.push(format!("(subpath (param \"{root_param}\"))"));
+                } else {
+                    // Add parameters for each read-only subpath and generate
+                    // the `(require-not ...)` clauses.
+                    let mut require_parts: Vec<String> = Vec::new();
+                    require_parts.push(format!("(subpath (param \"{root_param}\"))"));
+                    for (subpath_index, ro) in wr.read_only_subpaths.iter().enumerate() {
+                        let canonical_ro = ro.canonicalize().unwrap_or_else(|_| ro.clone());
+                        let ro_param = format!("WRITABLE_ROOT_{index}_RO_{subpath_index}");
+                        cli_args.push(format!("-D{ro_param}={}", canonical_ro.to_string_lossy()));
+                        require_parts
+                            .push(format!("(require-not (subpath (param \"{ro_param}\")))"));
+                    }
+                    let policy_component = format!("(require-all {} )", require_parts.join(" "));
+                    writable_folder_policies.push(policy_component);
+                }
+            }
+
+            if writable_folder_policies.is_empty() {
+                ("".to_string(), Vec::<String>::new())
+            } else {
+                let file_write_policy = format!(
+                    "(allow file-write*\n{}\n)",
+                    writable_folder_policies.join(" ")
+                );
+                (file_write_policy, cli_args)
+            }
+        }
+    };
+
+    let file_read_policy = if sandbox_policy.has_full_disk_read_access() {
+        "; allow read-only file operations\n(allow file-read*)"
+    } else {
+        ""
+    };
+
+    // TODO(mbolin): apply_patch calls must also honor the SandboxPolicy.
+    let network_policy = if sandbox_policy.has_full_network_access() {
+        "(allow network-outbound)\n(allow network-inbound)\n(allow system-socket)"
+    } else {
+        ""
+    };
+
+    let full_policy = format!(
+        "{MACOS_SEATBELT_BASE_POLICY}\n{file_read_policy}\n{file_write_policy}\n{network_policy}"
+    );
+
+    let mut seatbelt_args: Vec<String> = vec!["-p".to_string(), full_policy];
+    seatbelt_args.extend(extra_cli_args);
+    seatbelt_args.push("--".to_string());
+    seatbelt_args.extend(command);
+    seatbelt_args
+}
+
+#[cfg(test)]
+mod tests {
+    #![expect(clippy::expect_used)]
+    use super::MACOS_SEATBELT_BASE_POLICY;
+    use super::create_seatbelt_command_args;
+    use crate::protocol::SandboxPolicy;
+    use pretty_assertions::assert_eq;
+    use std::fs;
+    use std::path::Path;
+    use std::path::PathBuf;
+    use tempfile::TempDir;
+
+    #[test]
+    fn create_seatbelt_args_with_read_only_git_subpath() {
+        // Create a temporary workspace with two writable roots: one containing
+        // a top-level .git directory and one without it.
+        let tmp = TempDir::new().expect("tempdir");
+        let PopulatedTmp {
+            root_with_git,
+            root_without_git,
+            root_with_git_canon,
+            root_with_git_git_canon,
+            root_without_git_canon,
+        } = populate_tmpdir(tmp.path());
+
+        // Build a policy that only includes the two test roots as writable and
+        // does not automatically include defaults like cwd or TMPDIR.
+        let policy = SandboxPolicy::WorkspaceWrite {
+            writable_roots: vec![root_with_git.clone(), root_without_git.clone()],
+            network_access: false,
+            include_default_writable_roots: false,
+        };
+
+        let args = create_seatbelt_command_args(
+            vec!["/bin/echo".to_string(), "hello".to_string()],
+            &policy,
+            tmp.path(),
+        );
+
+        // Build the expected policy text using a raw string for readability.
+        // Note that the policy includes:
+        // - the base policy,
+        // - read-only access to the filesystem,
+        // - write access to WRITABLE_ROOT_0 (but not its .git) and WRITABLE_ROOT_1.
+        let expected_policy = format!(
+            r#"{MACOS_SEATBELT_BASE_POLICY}
+; allow read-only file operations
+(allow file-read*)
+(allow file-write*
+(require-all (subpath (param "WRITABLE_ROOT_0")) (require-not (subpath (param "WRITABLE_ROOT_0_RO_0"))) ) (subpath (param "WRITABLE_ROOT_1"))
+)
+"#,
+        );
+
+        let expected_args = vec![
+            "-p".to_string(),
+            expected_policy,
+            format!(
+                "-DWRITABLE_ROOT_0={}",
+                root_with_git_canon.to_string_lossy()
+            ),
+            format!(
+                "-DWRITABLE_ROOT_0_RO_0={}",
+                root_with_git_git_canon.to_string_lossy()
+            ),
+            format!(
+                "-DWRITABLE_ROOT_1={}",
+                root_without_git_canon.to_string_lossy()
+            ),
+            "--".to_string(),
+            "/bin/echo".to_string(),
+            "hello".to_string(),
+        ];
+
+        assert_eq!(args, expected_args);
+    }
+
+    #[test]
+    fn create_seatbelt_args_for_cwd_as_git_repo() {
+        // Create a temporary workspace with two writable roots: one containing
+        // a top-level .git directory and one without it.
+        let tmp = TempDir::new().expect("tempdir");
+        let PopulatedTmp {
+            root_with_git,
+            root_with_git_canon,
+            root_with_git_git_canon,
+            ..
+        } = populate_tmpdir(tmp.path());
+
+        // Build a policy that does not specify any writable_roots, but does
+        // use the default ones (cwd and TMPDIR) and verifies the `.git` check
+        // is done properly for cwd.
+        let policy = SandboxPolicy::WorkspaceWrite {
+            writable_roots: vec![],
+            network_access: false,
+            include_default_writable_roots: true,
+        };
+
+        let args = create_seatbelt_command_args(
+            vec!["/bin/echo".to_string(), "hello".to_string()],
+            &policy,
+            root_with_git.as_path(),
+        );
+
+        let tmpdir_env_var = if cfg!(target_os = "macos") {
+            std::env::var("TMPDIR")
+                .ok()
+                .map(PathBuf::from)
+                .and_then(|p| p.canonicalize().ok())
+                .map(|p| p.to_string_lossy().to_string())
+        } else {
+            None
+        };
+        let tempdir_policy_entry = if tmpdir_env_var.is_some() {
+            " (subpath (param \"WRITABLE_ROOT_1\"))"
+        } else {
+            ""
+        };
+
+        // Build the expected policy text using a raw string for readability.
+        // Note that the policy includes:
+        // - the base policy,
+        // - read-only access to the filesystem,
+        // - write access to WRITABLE_ROOT_0 (but not its .git) and WRITABLE_ROOT_1.
+        let expected_policy = format!(
+            r#"{MACOS_SEATBELT_BASE_POLICY}
+; allow read-only file operations
+(allow file-read*)
+(allow file-write*
+(require-all (subpath (param "WRITABLE_ROOT_0")) (require-not (subpath (param "WRITABLE_ROOT_0_RO_0"))) ){tempdir_policy_entry}
+)
+"#,
+        );
+
+        let mut expected_args = vec![
+            "-p".to_string(),
+            expected_policy,
+            format!(
+                "-DWRITABLE_ROOT_0={}",
+                root_with_git_canon.to_string_lossy()
+            ),
+            format!(
+                "-DWRITABLE_ROOT_0_RO_0={}",
+                root_with_git_git_canon.to_string_lossy()
+            ),
+        ];
+
+        if let Some(p) = tmpdir_env_var {
+            expected_args.push(format!("-DWRITABLE_ROOT_1={p}"));
+        }
+
+        expected_args.extend(vec![
+            "--".to_string(),
+            "/bin/echo".to_string(),
+            "hello".to_string(),
+        ]);
+
+        assert_eq!(args, expected_args);
+    }
+
+    struct PopulatedTmp {
+        root_with_git: PathBuf,
+        root_without_git: PathBuf,
+        root_with_git_canon: PathBuf,
+        root_with_git_git_canon: PathBuf,
+        root_without_git_canon: PathBuf,
+    }
+
+    fn populate_tmpdir(tmp: &Path) -> PopulatedTmp {
+        let root_with_git = tmp.join("with_git");
+        let root_without_git = tmp.join("no_git");
+        fs::create_dir_all(&root_with_git).expect("create with_git");
+        fs::create_dir_all(&root_without_git).expect("create no_git");
+        fs::create_dir_all(root_with_git.join(".git")).expect("create .git");
+
+        // Ensure we have canonical paths for -D parameter matching.
+        let root_with_git_canon = root_with_git.canonicalize().expect("canonicalize with_git");
+        let root_with_git_git_canon = root_with_git_canon.join(".git");
+        let root_without_git_canon = root_without_git
+            .canonicalize()
+            .expect("canonicalize no_git");
+        PopulatedTmp {
+            root_with_git,
+            root_without_git,
+            root_with_git_canon,
+            root_with_git_git_canon,
+            root_without_git_canon,
+        }
+    }
+}

codex-rs/core/src/seatbelt_base_policy.sbpl

@@ -65,3 +65,7 @@
   (sysctl-name "sysctl.proc_cputype")
   (sysctl-name-prefix "hw.perflevel")
 )
+
+; Added on top of Chrome profile
+; Needed for python multiprocessing on MacOS for the SemLock
+(allow ipc-posix-sem)

codex-rs/core/src/shell.rs

@@ -20,8 +20,13 @@ impl Shell {
                     return None;
                 }
 
-                let mut result = vec![zsh.shell_path.clone(), "-c".to_string()];
-                if let Ok(joined) = shlex::try_join(command.iter().map(|s| s.as_str())) {
+                let mut result = vec![zsh.shell_path.clone()];
+                result.push("-lc".to_string());
+
+                let joined = strip_bash_lc(&command)
+                    .or_else(|| shlex::try_join(command.iter().map(|s| s.as_str())).ok());
+
+                if let Some(joined) = joined {
                     result.push(format!("source {} && ({joined})", zsh.zshrc_path));
                 } else {
                     return None;
@@ -33,6 +38,19 @@ impl Shell {
     }
 }
 
+fn strip_bash_lc(command: &Vec<String>) -> Option<String> {
+    match command.as_slice() {
+        // exactly three items
+        [first, second, third]
+            // first two must be "bash", "-lc"
+            if first == "bash" && second == "-lc" =>
+        {
+            Some(third.clone())
+        }
+        _ => None,
+    }
+}
+
 #[cfg(target_os = "macos")]
 pub async fn default_user_shell() -> Shell {
     use tokio::process::Command;
@@ -119,15 +137,29 @@ mod tests {
         let cases = vec![
             (
                 vec!["myecho"],
-                vec![shell_path, "-c", "source ZSHRC_PATH && (myecho)"],
+                vec![shell_path, "-lc", "source ZSHRC_PATH && (myecho)"],
                 Some("It works!\n"),
             ),
+            (
+                vec!["myecho"],
+                vec![shell_path, "-lc", "source ZSHRC_PATH && (myecho)"],
+                Some("It works!\n"),
+            ),
+            (
+                vec!["bash", "-c", "echo 'single' \"double\""],
+                vec![
+                    shell_path,
+                    "-lc",
+                    "source ZSHRC_PATH && (bash -c \"echo 'single' \\\"double\\\"\")",
+                ],
+                Some("single double\n"),
+            ),
             (
                 vec!["bash", "-lc", "echo 'single' \"double\""],
                 vec![
                     shell_path,
-                    "-c",
-                    "source ZSHRC_PATH && (bash -lc \"echo 'single' \\\"double\\\"\")",
+                    "-lc",
+                    "source ZSHRC_PATH && (echo 'single' \"double\")",
                 ],
                 Some("single double\n"),
             ),
@@ -183,11 +215,14 @@ mod tests {
                         "HOME".to_string(),
                         temp_home.path().to_str().unwrap().to_string(),
                     )]),
+                    with_escalated_permissions: None,
+                    justification: None,
                 },
                 SandboxType::None,
                 Arc::new(Notify::new()),
                 &SandboxPolicy::DangerFullAccess,
                 &None,
+                None,
             )
             .await
             .unwrap();

codex-rs/core/src/spawn.rs

@@ -0,0 +1,107 @@
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::process::Stdio;
+use tokio::process::Child;
+use tokio::process::Command;
+use tracing::trace;
+
+use crate::protocol::SandboxPolicy;
+
+/// Experimental environment variable that will be set to some non-empty value
+/// if both of the following are true:
+///
+/// 1. The process was spawned by Codex as part of a shell tool call.
+/// 2. SandboxPolicy.has_full_network_access() was false for the tool call.
+///
+/// We may try to have just one environment variable for all sandboxing
+/// attributes, so this may change in the future.
+pub const CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR: &str = "CODEX_SANDBOX_NETWORK_DISABLED";
+
+/// Should be set when the process is spawned under a sandbox. Currently, the
+/// value is "seatbelt" for macOS, but it may change in the future to
+/// accommodate sandboxing configuration and other sandboxing mechanisms.
+pub const CODEX_SANDBOX_ENV_VAR: &str = "CODEX_SANDBOX";
+
+#[derive(Debug, Clone, Copy)]
+pub enum StdioPolicy {
+    RedirectForShellTool,
+    Inherit,
+}
+
+/// Spawns the appropriate child process for the ExecParams and SandboxPolicy,
+/// ensuring the args and environment variables used to create the `Command`
+/// (and `Child`) honor the configuration.
+///
+/// For now, we take `SandboxPolicy` as a parameter to spawn_child() because
+/// we need to determine whether to set the
+/// `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` environment variable.
+pub(crate) async fn spawn_child_async(
+    program: PathBuf,
+    args: Vec<String>,
+    #[cfg_attr(not(unix), allow(unused_variables))] arg0: Option<&str>,
+    cwd: PathBuf,
+    sandbox_policy: &SandboxPolicy,
+    stdio_policy: StdioPolicy,
+    env: HashMap<String, String>,
+) -> std::io::Result<Child> {
+    trace!(
+        "spawn_child_async: {program:?} {args:?} {arg0:?} {cwd:?} {sandbox_policy:?} {stdio_policy:?} {env:?}"
+    );
+
+    let mut cmd = Command::new(&program);
+    #[cfg(unix)]
+    cmd.arg0(arg0.map_or_else(|| program.to_string_lossy().to_string(), String::from));
+    cmd.args(args);
+    cmd.current_dir(cwd);
+    cmd.env_clear();
+    cmd.envs(env);
+
+    if !sandbox_policy.has_full_network_access() {
+        cmd.env(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR, "1");
+    }
+
+    // If this Codex process dies (including being killed via SIGKILL), we want
+    // any child processes that were spawned as part of a `"shell"` tool call
+    // to also be terminated.
+
+    // This relies on prctl(2), so it only works on Linux.
+    #[cfg(target_os = "linux")]
+    unsafe {
+        cmd.pre_exec(|| {
+            // This prctl call effectively requests, "deliver SIGTERM when my
+            // current parent dies."
+            if libc::prctl(libc::PR_SET_PDEATHSIG, libc::SIGTERM) == -1 {
+                return Err(std::io::Error::last_os_error());
+            }
+
+            // Though if there was a race condition and this pre_exec() block is
+            // run _after_ the parent (i.e., the Codex process) has already
+            // exited, then the parent is the _init_ process (which will never
+            // die), so we should just terminate the child process now.
+            if libc::getppid() == 1 {
+                libc::raise(libc::SIGTERM);
+            }
+            Ok(())
+        });
+    }
+
+    match stdio_policy {
+        StdioPolicy::RedirectForShellTool => {
+            // Do not create a file descriptor for stdin because otherwise some
+            // commands may hang forever waiting for input. For example, ripgrep has
+            // a heuristic where it may try to read from stdin as explained here:
+            // https://github.com/BurntSushi/ripgrep/blob/e2362d4d5185d02fa857bf381e7bd52e66fafc73/crates/core/flags/hiargs.rs#L1101-L1103
+            cmd.stdin(Stdio::null());
+
+            cmd.stdout(Stdio::piped()).stderr(Stdio::piped());
+        }
+        StdioPolicy::Inherit => {
+            // Inherit stdin, stdout, and stderr from the parent process.
+            cmd.stdin(Stdio::inherit())
+                .stdout(Stdio::inherit())
+                .stderr(Stdio::inherit());
+        }
+    }
+
+    cmd.kill_on_drop(true).spawn()
+}

codex-rs/core/src/turn_diff_tracker.rs

@@ -0,0 +1,887 @@
+use std::collections::HashMap;
+use std::fs;
+use std::path::Path;
+use std::path::PathBuf;
+use std::process::Command;
+
+use anyhow::Context;
+use anyhow::Result;
+use anyhow::anyhow;
+use sha1::digest::Output;
+use uuid::Uuid;
+
+use crate::protocol::FileChange;
+
+const ZERO_OID: &str = "0000000000000000000000000000000000000000";
+const DEV_NULL: &str = "/dev/null";
+
+struct BaselineFileInfo {
+    path: PathBuf,
+    content: Vec<u8>,
+    mode: FileMode,
+    oid: String,
+}
+
+/// Tracks sets of changes to files and exposes the overall unified diff.
+/// Internally, the way this works is now:
+/// 1. Maintain an in-memory baseline snapshot of files when they are first seen.
+///    For new additions, do not create a baseline so that diffs are shown as proper additions (using /dev/null).
+/// 2. Keep a stable internal filename (uuid) per external path for rename tracking.
+/// 3. To compute the aggregated unified diff, compare each baseline snapshot to the current file on disk entirely in-memory
+///    using the `similar` crate and emit unified diffs with rewritten external paths.
+#[derive(Default)]
+pub struct TurnDiffTracker {
+    /// Map external path -> internal filename (uuid).
+    external_to_temp_name: HashMap<PathBuf, String>,
+    /// Internal filename -> baseline file info.
+    baseline_file_info: HashMap<String, BaselineFileInfo>,
+    /// Internal filename -> external path as of current accumulated state (after applying all changes).
+    /// This is where renames are tracked.
+    temp_name_to_current_path: HashMap<String, PathBuf>,
+    /// Cache of known git worktree roots to avoid repeated filesystem walks.
+    git_root_cache: Vec<PathBuf>,
+}
+
+impl TurnDiffTracker {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Front-run apply patch calls to track the starting contents of any modified files.
+    /// - Creates an in-memory baseline snapshot for files that already exist on disk when first seen.
+    /// - For additions, we intentionally do not create a baseline snapshot so that diffs are proper additions.
+    /// - Also updates internal mappings for move/rename events.
+    pub fn on_patch_begin(&mut self, changes: &HashMap<PathBuf, FileChange>) {
+        for (path, change) in changes.iter() {
+            // Ensure a stable internal filename exists for this external path.
+            if !self.external_to_temp_name.contains_key(path) {
+                let internal = Uuid::new_v4().to_string();
+                self.external_to_temp_name
+                    .insert(path.clone(), internal.clone());
+                self.temp_name_to_current_path
+                    .insert(internal.clone(), path.clone());
+
+                // If the file exists on disk now, snapshot as baseline; else leave missing to represent /dev/null.
+                let baseline_file_info = if path.exists() {
+                    let mode = file_mode_for_path(path);
+                    let mode_val = mode.unwrap_or(FileMode::Regular);
+                    let content = blob_bytes(path, &mode_val).unwrap_or_default();
+                    let oid = if mode == Some(FileMode::Symlink) {
+                        format!("{:x}", git_blob_sha1_hex_bytes(&content))
+                    } else {
+                        self.git_blob_oid_for_path(path)
+                            .unwrap_or_else(|| format!("{:x}", git_blob_sha1_hex_bytes(&content)))
+                    };
+                    Some(BaselineFileInfo {
+                        path: path.clone(),
+                        content,
+                        mode: mode_val,
+                        oid,
+                    })
+                } else {
+                    Some(BaselineFileInfo {
+                        path: path.clone(),
+                        content: vec![],
+                        mode: FileMode::Regular,
+                        oid: ZERO_OID.to_string(),
+                    })
+                };
+
+                if let Some(baseline_file_info) = baseline_file_info {
+                    self.baseline_file_info
+                        .insert(internal.clone(), baseline_file_info);
+                }
+            }
+
+            // Track rename/move in current mapping if provided in an Update.
+            if let FileChange::Update {
+                move_path: Some(dest),
+                ..
+            } = change
+            {
+                let uuid_filename = match self.external_to_temp_name.get(path) {
+                    Some(i) => i.clone(),
+                    None => {
+                        // This should be rare, but if we haven't mapped the source, create it with no baseline.
+                        let i = Uuid::new_v4().to_string();
+                        self.baseline_file_info.insert(
+                            i.clone(),
+                            BaselineFileInfo {
+                                path: path.clone(),
+                                content: vec![],
+                                mode: FileMode::Regular,
+                                oid: ZERO_OID.to_string(),
+                            },
+                        );
+                        i
+                    }
+                };
+                // Update current external mapping for temp file name.
+                self.temp_name_to_current_path
+                    .insert(uuid_filename.clone(), dest.clone());
+                // Update forward file_mapping: external current -> internal name.
+                self.external_to_temp_name.remove(path);
+                self.external_to_temp_name
+                    .insert(dest.clone(), uuid_filename);
+            };
+        }
+    }
+
+    fn get_path_for_internal(&self, internal: &str) -> Option<PathBuf> {
+        self.temp_name_to_current_path
+            .get(internal)
+            .cloned()
+            .or_else(|| {
+                self.baseline_file_info
+                    .get(internal)
+                    .map(|info| info.path.clone())
+            })
+    }
+
+    /// Find the git worktree root for a file/directory by walking up to the first ancestor containing a `.git` entry.
+    /// Uses a simple cache of known roots and avoids negative-result caching for simplicity.
+    fn find_git_root_cached(&mut self, start: &Path) -> Option<PathBuf> {
+        let dir = if start.is_dir() {
+            start
+        } else {
+            start.parent()?
+        };
+
+        // Fast path: if any cached root is an ancestor of this path, use it.
+        if let Some(root) = self
+            .git_root_cache
+            .iter()
+            .find(|r| dir.starts_with(r))
+            .cloned()
+        {
+            return Some(root);
+        }
+
+        // Walk up to find a `.git` marker.
+        let mut cur = dir.to_path_buf();
+        loop {
+            let git_marker = cur.join(".git");
+            if git_marker.is_dir() || git_marker.is_file() {
+                if !self.git_root_cache.iter().any(|r| r == &cur) {
+                    self.git_root_cache.push(cur.clone());
+                }
+                return Some(cur);
+            }
+
+            // On Windows, avoid walking above the drive or UNC share root.
+            #[cfg(windows)]
+            {
+                if is_windows_drive_or_unc_root(&cur) {
+                    return None;
+                }
+            }
+
+            if let Some(parent) = cur.parent() {
+                cur = parent.to_path_buf();
+            } else {
+                return None;
+            }
+        }
+    }
+
+    /// Return a display string for `path` relative to its git root if found, else absolute.
+    fn relative_to_git_root_str(&mut self, path: &Path) -> String {
+        let s = if let Some(root) = self.find_git_root_cached(path) {
+            if let Ok(rel) = path.strip_prefix(&root) {
+                rel.display().to_string()
+            } else {
+                path.display().to_string()
+            }
+        } else {
+            path.display().to_string()
+        };
+        s.replace('\\', "/")
+    }
+
+    /// Ask git to compute the blob SHA-1 for the file at `path` within its repository.
+    /// Returns None if no repository is found or git invocation fails.
+    fn git_blob_oid_for_path(&mut self, path: &Path) -> Option<String> {
+        let root = self.find_git_root_cached(path)?;
+        // Compute a path relative to the repo root for better portability across platforms.
+        let rel = path.strip_prefix(&root).unwrap_or(path);
+        let output = Command::new("git")
+            .arg("-C")
+            .arg(&root)
+            .arg("hash-object")
+            .arg("--")
+            .arg(rel)
+            .output()
+            .ok()?;
+        if !output.status.success() {
+            return None;
+        }
+        let s = String::from_utf8_lossy(&output.stdout).trim().to_string();
+        if s.len() == 40 { Some(s) } else { None }
+    }
+
+    /// Recompute the aggregated unified diff by comparing all of the in-memory snapshots that were
+    /// collected before the first time they were touched by apply_patch during this turn with
+    /// the current repo state.
+    pub fn get_unified_diff(&mut self) -> Result<Option<String>> {
+        let mut aggregated = String::new();
+
+        // Compute diffs per tracked internal file in a stable order by external path.
+        let mut baseline_file_names: Vec<String> =
+            self.baseline_file_info.keys().cloned().collect();
+        // Sort lexicographically by full repo-relative path to match git behavior.
+        baseline_file_names.sort_by_key(|internal| {
+            self.get_path_for_internal(internal)
+                .map(|p| self.relative_to_git_root_str(&p))
+                .unwrap_or_default()
+        });
+
+        for internal in baseline_file_names {
+            aggregated.push_str(self.get_file_diff(&internal).as_str());
+            if !aggregated.ends_with('\n') {
+                aggregated.push('\n');
+            }
+        }
+
+        if aggregated.trim().is_empty() {
+            Ok(None)
+        } else {
+            Ok(Some(aggregated))
+        }
+    }
+
+    fn get_file_diff(&mut self, internal_file_name: &str) -> String {
+        let mut aggregated = String::new();
+
+        // Snapshot lightweight fields only.
+        let (baseline_external_path, baseline_mode, left_oid) = {
+            if let Some(info) = self.baseline_file_info.get(internal_file_name) {
+                (info.path.clone(), info.mode, info.oid.clone())
+            } else {
+                (PathBuf::new(), FileMode::Regular, ZERO_OID.to_string())
+            }
+        };
+        let current_external_path = match self.get_path_for_internal(internal_file_name) {
+            Some(p) => p,
+            None => return aggregated,
+        };
+
+        let current_mode = file_mode_for_path(&current_external_path).unwrap_or(FileMode::Regular);
+        let right_bytes = blob_bytes(&current_external_path, &current_mode);
+
+        // Compute displays with &mut self before borrowing any baseline content.
+        let left_display = self.relative_to_git_root_str(&baseline_external_path);
+        let right_display = self.relative_to_git_root_str(&current_external_path);
+
+        // Compute right oid before borrowing baseline content.
+        let right_oid = if let Some(b) = right_bytes.as_ref() {
+            if current_mode == FileMode::Symlink {
+                format!("{:x}", git_blob_sha1_hex_bytes(b))
+            } else {
+                self.git_blob_oid_for_path(&current_external_path)
+                    .unwrap_or_else(|| format!("{:x}", git_blob_sha1_hex_bytes(b)))
+            }
+        } else {
+            ZERO_OID.to_string()
+        };
+
+        // Borrow baseline content only after all &mut self uses are done.
+        let left_present = left_oid.as_str() != ZERO_OID;
+        let left_bytes: Option<&[u8]> = if left_present {
+            self.baseline_file_info
+                .get(internal_file_name)
+                .map(|i| i.content.as_slice())
+        } else {
+            None
+        };
+
+        // Fast path: identical bytes or both missing.
+        if left_bytes == right_bytes.as_deref() {
+            return aggregated;
+        }
+
+        aggregated.push_str(&format!("diff --git a/{left_display} b/{right_display}\n"));
+
+        let is_add = !left_present && right_bytes.is_some();
+        let is_delete = left_present && right_bytes.is_none();
+
+        if is_add {
+            aggregated.push_str(&format!("new file mode {current_mode}\n"));
+        } else if is_delete {
+            aggregated.push_str(&format!("deleted file mode {baseline_mode}\n"));
+        } else if baseline_mode != current_mode {
+            aggregated.push_str(&format!("old mode {baseline_mode}\n"));
+            aggregated.push_str(&format!("new mode {current_mode}\n"));
+        }
+
+        let left_text = left_bytes.and_then(|b| std::str::from_utf8(b).ok());
+        let right_text = right_bytes
+            .as_deref()
+            .and_then(|b| std::str::from_utf8(b).ok());
+
+        let can_text_diff = matches!(
+            (left_text, right_text, is_add, is_delete),
+            (Some(_), Some(_), _, _) | (_, Some(_), true, _) | (Some(_), _, _, true)
+        );
+
+        if can_text_diff {
+            let l = left_text.unwrap_or("");
+            let r = right_text.unwrap_or("");
+
+            aggregated.push_str(&format!("index {left_oid}..{right_oid}\n"));
+
+            let old_header = if left_present {
+                format!("a/{left_display}")
+            } else {
+                DEV_NULL.to_string()
+            };
+            let new_header = if right_bytes.is_some() {
+                format!("b/{right_display}")
+            } else {
+                DEV_NULL.to_string()
+            };
+
+            let diff = similar::TextDiff::from_lines(l, r);
+            let unified = diff
+                .unified_diff()
+                .context_radius(3)
+                .header(&old_header, &new_header)
+                .to_string();
+
+            aggregated.push_str(&unified);
+        } else {
+            aggregated.push_str(&format!("index {left_oid}..{right_oid}\n"));
+            let old_header = if left_present {
+                format!("a/{left_display}")
+            } else {
+                DEV_NULL.to_string()
+            };
+            let new_header = if right_bytes.is_some() {
+                format!("b/{right_display}")
+            } else {
+                DEV_NULL.to_string()
+            };
+            aggregated.push_str(&format!("--- {old_header}\n"));
+            aggregated.push_str(&format!("+++ {new_header}\n"));
+            aggregated.push_str("Binary files differ\n");
+        }
+        aggregated
+    }
+}
+
+/// Compute the Git SHA-1 blob object ID for the given content (bytes).
+fn git_blob_sha1_hex_bytes(data: &[u8]) -> Output<sha1::Sha1> {
+    // Git blob hash is sha1 of: "blob <len>\0<data>"
+    let header = format!("blob {}\0", data.len());
+    use sha1::Digest;
+    let mut hasher = sha1::Sha1::new();
+    hasher.update(header.as_bytes());
+    hasher.update(data);
+    hasher.finalize()
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum FileMode {
+    Regular,
+    #[cfg(unix)]
+    Executable,
+    Symlink,
+}
+
+impl FileMode {
+    fn as_str(&self) -> &'static str {
+        match self {
+            FileMode::Regular => "100644",
+            #[cfg(unix)]
+            FileMode::Executable => "100755",
+            FileMode::Symlink => "120000",
+        }
+    }
+}
+
+impl std::fmt::Display for FileMode {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.as_str())
+    }
+}
+
+#[cfg(unix)]
+fn file_mode_for_path(path: &Path) -> Option<FileMode> {
+    use std::os::unix::fs::PermissionsExt;
+    let meta = fs::symlink_metadata(path).ok()?;
+    let ft = meta.file_type();
+    if ft.is_symlink() {
+        return Some(FileMode::Symlink);
+    }
+    let mode = meta.permissions().mode();
+    let is_exec = (mode & 0o111) != 0;
+    Some(if is_exec {
+        FileMode::Executable
+    } else {
+        FileMode::Regular
+    })
+}
+
+#[cfg(not(unix))]
+fn file_mode_for_path(_path: &Path) -> Option<FileMode> {
+    // Default to non-executable on non-unix.
+    Some(FileMode::Regular)
+}
+
+fn blob_bytes(path: &Path, mode: &FileMode) -> Option<Vec<u8>> {
+    if path.exists() {
+        let contents = if *mode == FileMode::Symlink {
+            symlink_blob_bytes(path)
+                .ok_or_else(|| anyhow!("failed to read symlink target for {}", path.display()))
+        } else {
+            fs::read(path)
+                .with_context(|| format!("failed to read current file for diff {}", path.display()))
+        };
+        contents.ok()
+    } else {
+        None
+    }
+}
+
+#[cfg(unix)]
+fn symlink_blob_bytes(path: &Path) -> Option<Vec<u8>> {
+    use std::os::unix::ffi::OsStrExt;
+    let target = std::fs::read_link(path).ok()?;
+    Some(target.as_os_str().as_bytes().to_vec())
+}
+
+#[cfg(not(unix))]
+fn symlink_blob_bytes(_path: &Path) -> Option<Vec<u8>> {
+    None
+}
+
+#[cfg(windows)]
+fn is_windows_drive_or_unc_root(p: &std::path::Path) -> bool {
+    use std::path::Component;
+    let mut comps = p.components();
+    matches!(
+        (comps.next(), comps.next(), comps.next()),
+        (Some(Component::Prefix(_)), Some(Component::RootDir), None)
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used)]
+    use super::*;
+    use pretty_assertions::assert_eq;
+    use tempfile::tempdir;
+
+    /// Compute the Git SHA-1 blob object ID for the given content (string).
+    /// This delegates to the bytes version to avoid UTF-8 lossy conversions here.
+    fn git_blob_sha1_hex(data: &str) -> String {
+        format!("{:x}", git_blob_sha1_hex_bytes(data.as_bytes()))
+    }
+
+    fn normalize_diff_for_test(input: &str, root: &Path) -> String {
+        let root_str = root.display().to_string().replace('\\', "/");
+        let replaced = input.replace(&root_str, "<TMP>");
+        // Split into blocks on lines starting with "diff --git ", sort blocks for determinism, and rejoin
+        let mut blocks: Vec<String> = Vec::new();
+        let mut current = String::new();
+        for line in replaced.lines() {
+            if line.starts_with("diff --git ") && !current.is_empty() {
+                blocks.push(current);
+                current = String::new();
+            }
+            if !current.is_empty() {
+                current.push('\n');
+            }
+            current.push_str(line);
+        }
+        if !current.is_empty() {
+            blocks.push(current);
+        }
+        blocks.sort();
+        let mut out = blocks.join("\n");
+        if !out.ends_with('\n') {
+            out.push('\n');
+        }
+        out
+    }
+
+    #[test]
+    fn accumulates_add_and_update() {
+        let mut acc = TurnDiffTracker::new();
+
+        let dir = tempdir().unwrap();
+        let file = dir.path().join("a.txt");
+
+        // First patch: add file (baseline should be /dev/null).
+        let add_changes = HashMap::from([(
+            file.clone(),
+            FileChange::Add {
+                content: "foo\n".to_string(),
+            },
+        )]);
+        acc.on_patch_begin(&add_changes);
+
+        // Simulate apply: create the file on disk.
+        fs::write(&file, "foo\n").unwrap();
+        let first = acc.get_unified_diff().unwrap().unwrap();
+        let first = normalize_diff_for_test(&first, dir.path());
+        let expected_first = {
+            let mode = file_mode_for_path(&file).unwrap_or(FileMode::Regular);
+            let right_oid = git_blob_sha1_hex("foo\n");
+            format!(
+                r#"diff --git a/<TMP>/a.txt b/<TMP>/a.txt
+new file mode {mode}
+index {ZERO_OID}..{right_oid}
+--- {DEV_NULL}
++++ b/<TMP>/a.txt
+@@ -0,0 +1 @@
++foo
+"#,
+            )
+        };
+        assert_eq!(first, expected_first);
+
+        // Second patch: update the file on disk.
+        let update_changes = HashMap::from([(
+            file.clone(),
+            FileChange::Update {
+                unified_diff: "".to_owned(),
+                move_path: None,
+            },
+        )]);
+        acc.on_patch_begin(&update_changes);
+
+        // Simulate apply: append a new line.
+        fs::write(&file, "foo\nbar\n").unwrap();
+        let combined = acc.get_unified_diff().unwrap().unwrap();
+        let combined = normalize_diff_for_test(&combined, dir.path());
+        let expected_combined = {
+            let mode = file_mode_for_path(&file).unwrap_or(FileMode::Regular);
+            let right_oid = git_blob_sha1_hex("foo\nbar\n");
+            format!(
+                r#"diff --git a/<TMP>/a.txt b/<TMP>/a.txt
+new file mode {mode}
+index {ZERO_OID}..{right_oid}
+--- {DEV_NULL}
++++ b/<TMP>/a.txt
+@@ -0,0 +1,2 @@
++foo
++bar
+"#,
+            )
+        };
+        assert_eq!(combined, expected_combined);
+    }
+
+    #[test]
+    fn accumulates_delete() {
+        let dir = tempdir().unwrap();
+        let file = dir.path().join("b.txt");
+        fs::write(&file, "x\n").unwrap();
+
+        let mut acc = TurnDiffTracker::new();
+        let del_changes = HashMap::from([(file.clone(), FileChange::Delete)]);
+        acc.on_patch_begin(&del_changes);
+
+        // Simulate apply: delete the file from disk.
+        let baseline_mode = file_mode_for_path(&file).unwrap_or(FileMode::Regular);
+        fs::remove_file(&file).unwrap();
+        let diff = acc.get_unified_diff().unwrap().unwrap();
+        let diff = normalize_diff_for_test(&diff, dir.path());
+        let expected = {
+            let left_oid = git_blob_sha1_hex("x\n");
+            format!(
+                r#"diff --git a/<TMP>/b.txt b/<TMP>/b.txt
+deleted file mode {baseline_mode}
+index {left_oid}..{ZERO_OID}
+--- a/<TMP>/b.txt
++++ {DEV_NULL}
+@@ -1 +0,0 @@
+-x
+"#,
+            )
+        };
+        assert_eq!(diff, expected);
+    }
+
+    #[test]
+    fn accumulates_move_and_update() {
+        let dir = tempdir().unwrap();
+        let src = dir.path().join("src.txt");
+        let dest = dir.path().join("dst.txt");
+        fs::write(&src, "line\n").unwrap();
+
+        let mut acc = TurnDiffTracker::new();
+        let mv_changes = HashMap::from([(
+            src.clone(),
+            FileChange::Update {
+                unified_diff: "".to_owned(),
+                move_path: Some(dest.clone()),
+            },
+        )]);
+        acc.on_patch_begin(&mv_changes);
+
+        // Simulate apply: move and update content.
+        fs::rename(&src, &dest).unwrap();
+        fs::write(&dest, "line2\n").unwrap();
+
+        let out = acc.get_unified_diff().unwrap().unwrap();
+        let out = normalize_diff_for_test(&out, dir.path());
+        let expected = {
+            let left_oid = git_blob_sha1_hex("line\n");
+            let right_oid = git_blob_sha1_hex("line2\n");
+            format!(
+                r#"diff --git a/<TMP>/src.txt b/<TMP>/dst.txt
+index {left_oid}..{right_oid}
+--- a/<TMP>/src.txt
++++ b/<TMP>/dst.txt
+@@ -1 +1 @@
+-line
++line2
+"#
+            )
+        };
+        assert_eq!(out, expected);
+    }
+
+    #[test]
+    fn move_without_1change_yields_no_diff() {
+        let dir = tempdir().unwrap();
+        let src = dir.path().join("moved.txt");
+        let dest = dir.path().join("renamed.txt");
+        fs::write(&src, "same\n").unwrap();
+
+        let mut acc = TurnDiffTracker::new();
+        let mv_changes = HashMap::from([(
+            src.clone(),
+            FileChange::Update {
+                unified_diff: "".to_owned(),
+                move_path: Some(dest.clone()),
+            },
+        )]);
+        acc.on_patch_begin(&mv_changes);
+
+        // Simulate apply: move only, no content change.
+        fs::rename(&src, &dest).unwrap();
+
+        let diff = acc.get_unified_diff().unwrap();
+        assert_eq!(diff, None);
+    }
+
+    #[test]
+    fn move_declared_but_file_only_appears_at_dest_is_add() {
+        let dir = tempdir().unwrap();
+        let src = dir.path().join("src.txt");
+        let dest = dir.path().join("dest.txt");
+        let mut acc = TurnDiffTracker::new();
+        let mv = HashMap::from([(
+            src.clone(),
+            FileChange::Update {
+                unified_diff: "".into(),
+                move_path: Some(dest.clone()),
+            },
+        )]);
+        acc.on_patch_begin(&mv);
+        // No file existed initially; create only dest
+        fs::write(&dest, "hello\n").unwrap();
+        let diff = acc.get_unified_diff().unwrap().unwrap();
+        let diff = normalize_diff_for_test(&diff, dir.path());
+        let expected = {
+            let mode = file_mode_for_path(&dest).unwrap_or(FileMode::Regular);
+            let right_oid = git_blob_sha1_hex("hello\n");
+            format!(
+                r#"diff --git a/<TMP>/src.txt b/<TMP>/dest.txt
+new file mode {mode}
+index {ZERO_OID}..{right_oid}
+--- {DEV_NULL}
++++ b/<TMP>/dest.txt
+@@ -0,0 +1 @@
++hello
+"#,
+            )
+        };
+        assert_eq!(diff, expected);
+    }
+
+    #[test]
+    fn update_persists_across_new_baseline_for_new_file() {
+        let dir = tempdir().unwrap();
+        let a = dir.path().join("a.txt");
+        let b = dir.path().join("b.txt");
+        fs::write(&a, "foo\n").unwrap();
+        fs::write(&b, "z\n").unwrap();
+
+        let mut acc = TurnDiffTracker::new();
+
+        // First: update existing a.txt (baseline snapshot is created for a).
+        let update_a = HashMap::from([(
+            a.clone(),
+            FileChange::Update {
+                unified_diff: "".to_owned(),
+                move_path: None,
+            },
+        )]);
+        acc.on_patch_begin(&update_a);
+        // Simulate apply: modify a.txt on disk.
+        fs::write(&a, "foo\nbar\n").unwrap();
+        let first = acc.get_unified_diff().unwrap().unwrap();
+        let first = normalize_diff_for_test(&first, dir.path());
+        let expected_first = {
+            let left_oid = git_blob_sha1_hex("foo\n");
+            let right_oid = git_blob_sha1_hex("foo\nbar\n");
+            format!(
+                r#"diff --git a/<TMP>/a.txt b/<TMP>/a.txt
+index {left_oid}..{right_oid}
+--- a/<TMP>/a.txt
++++ b/<TMP>/a.txt
+@@ -1 +1,2 @@
+ foo
++bar
+"#
+            )
+        };
+        assert_eq!(first, expected_first);
+
+        // Next: introduce a brand-new path b.txt into baseline snapshots via a delete change.
+        let del_b = HashMap::from([(b.clone(), FileChange::Delete)]);
+        acc.on_patch_begin(&del_b);
+        // Simulate apply: delete b.txt.
+        let baseline_mode = file_mode_for_path(&b).unwrap_or(FileMode::Regular);
+        fs::remove_file(&b).unwrap();
+
+        let combined = acc.get_unified_diff().unwrap().unwrap();
+        let combined = normalize_diff_for_test(&combined, dir.path());
+        let expected = {
+            let left_oid_a = git_blob_sha1_hex("foo\n");
+            let right_oid_a = git_blob_sha1_hex("foo\nbar\n");
+            let left_oid_b = git_blob_sha1_hex("z\n");
+            format!(
+                r#"diff --git a/<TMP>/a.txt b/<TMP>/a.txt
+index {left_oid_a}..{right_oid_a}
+--- a/<TMP>/a.txt
++++ b/<TMP>/a.txt
+@@ -1 +1,2 @@
+ foo
++bar
+diff --git a/<TMP>/b.txt b/<TMP>/b.txt
+deleted file mode {baseline_mode}
+index {left_oid_b}..{ZERO_OID}
+--- a/<TMP>/b.txt
++++ {DEV_NULL}
+@@ -1 +0,0 @@
+-z
+"#,
+            )
+        };
+        assert_eq!(combined, expected);
+    }
+
+    #[test]
+    fn binary_files_differ_update() {
+        let dir = tempdir().unwrap();
+        let file = dir.path().join("bin.dat");
+
+        // Initial non-UTF8 bytes
+        let left_bytes: Vec<u8> = vec![0xff, 0xfe, 0xfd, 0x00];
+        // Updated non-UTF8 bytes
+        let right_bytes: Vec<u8> = vec![0x01, 0x02, 0x03, 0x00];
+
+        fs::write(&file, &left_bytes).unwrap();
+
+        let mut acc = TurnDiffTracker::new();
+        let update_changes = HashMap::from([(
+            file.clone(),
+            FileChange::Update {
+                unified_diff: "".to_owned(),
+                move_path: None,
+            },
+        )]);
+        acc.on_patch_begin(&update_changes);
+
+        // Apply update on disk
+        fs::write(&file, &right_bytes).unwrap();
+
+        let diff = acc.get_unified_diff().unwrap().unwrap();
+        let diff = normalize_diff_for_test(&diff, dir.path());
+        let expected = {
+            let left_oid = format!("{:x}", git_blob_sha1_hex_bytes(&left_bytes));
+            let right_oid = format!("{:x}", git_blob_sha1_hex_bytes(&right_bytes));
+            format!(
+                r#"diff --git a/<TMP>/bin.dat b/<TMP>/bin.dat
+index {left_oid}..{right_oid}
+--- a/<TMP>/bin.dat
++++ b/<TMP>/bin.dat
+Binary files differ
+"#
+            )
+        };
+        assert_eq!(diff, expected);
+    }
+
+    #[test]
+    fn filenames_with_spaces_add_and_update() {
+        let mut acc = TurnDiffTracker::new();
+
+        let dir = tempdir().unwrap();
+        let file = dir.path().join("name with spaces.txt");
+
+        // First patch: add file (baseline should be /dev/null).
+        let add_changes = HashMap::from([(
+            file.clone(),
+            FileChange::Add {
+                content: "foo\n".to_string(),
+            },
+        )]);
+        acc.on_patch_begin(&add_changes);
+
+        // Simulate apply: create the file on disk.
+        fs::write(&file, "foo\n").unwrap();
+        let first = acc.get_unified_diff().unwrap().unwrap();
+        let first = normalize_diff_for_test(&first, dir.path());
+        let expected_first = {
+            let mode = file_mode_for_path(&file).unwrap_or(FileMode::Regular);
+            let right_oid = git_blob_sha1_hex("foo\n");
+            format!(
+                r#"diff --git a/<TMP>/name with spaces.txt b/<TMP>/name with spaces.txt
+new file mode {mode}
+index {ZERO_OID}..{right_oid}
+--- {DEV_NULL}
++++ b/<TMP>/name with spaces.txt
+@@ -0,0 +1 @@
++foo
+"#,
+            )
+        };
+        assert_eq!(first, expected_first);
+
+        // Second patch: update the file on disk.
+        let update_changes = HashMap::from([(
+            file.clone(),
+            FileChange::Update {
+                unified_diff: "".to_owned(),
+                move_path: None,
+            },
+        )]);
+        acc.on_patch_begin(&update_changes);
+
+        // Simulate apply: append a new line with a space.
+        fs::write(&file, "foo\nbar baz\n").unwrap();
+        let combined = acc.get_unified_diff().unwrap().unwrap();
+        let combined = normalize_diff_for_test(&combined, dir.path());
+        let expected_combined = {
+            let mode = file_mode_for_path(&file).unwrap_or(FileMode::Regular);
+            let right_oid = git_blob_sha1_hex("foo\nbar baz\n");
+            format!(
+                r#"diff --git a/<TMP>/name with spaces.txt b/<TMP>/name with spaces.txt
+new file mode {mode}
+index {ZERO_OID}..{right_oid}
+--- {DEV_NULL}
++++ b/<TMP>/name with spaces.txt
+@@ -0,0 +1,2 @@
++foo
++bar baz
+"#,
+            )
+        };
+        assert_eq!(combined, expected_combined);
+    }
+}

codex-rs/core/tests/cli_stream.rs

@@ -1,7 +1,7 @@
 #![expect(clippy::unwrap_used)]
 
 use assert_cmd::Command as AssertCommand;
-use codex_core::exec::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
 use std::time::Duration;
 use std::time::Instant;
 use tempfile::TempDir;
@@ -81,6 +81,96 @@ async fn chat_mode_stream_cli() {
     server.verify().await;
 }
 
+/// Verify that passing `-c experimental_instructions_file=...` to the CLI
+/// overrides the built-in base instructions by inspecting the request body
+/// received by a mock OpenAI Responses endpoint.
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn exec_cli_applies_experimental_instructions_file() {
+    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 server which will capture the request and return a minimal
+    // SSE stream for a single turn.
+    let server = MockServer::start().await;
+    let sse = concat!(
+        "data: {\"type\":\"response.created\",\"response\":{}}\n\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"r1\"}}\n\n"
+    );
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(
+            ResponseTemplate::new(200)
+                .insert_header("content-type", "text/event-stream")
+                .set_body_raw(sse, "text/event-stream"),
+        )
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    // Create a temporary instructions file with a unique marker we can assert
+    // appears in the outbound request payload.
+    let custom = TempDir::new().unwrap();
+    let marker = "cli-experimental-instructions-marker";
+    let custom_path = custom.path().join("instr.md");
+    std::fs::write(&custom_path, marker).unwrap();
+    let custom_path_str = custom_path.to_string_lossy().replace('\\', "/");
+
+    // Build a provider override that points at the mock server and instructs
+    // Codex to use the Responses API with the dummy env var.
+    let provider_override = format!(
+        "model_providers.mock={{ name = \"mock\", base_url = \"{}/v1\", env_key = \"PATH\", wire_api = \"responses\" }}",
+        server.uri()
+    );
+
+    let home = TempDir::new().unwrap();
+    let mut cmd = AssertCommand::new("cargo");
+    cmd.arg("run")
+        .arg("-p")
+        .arg("codex-cli")
+        .arg("--quiet")
+        .arg("--")
+        .arg("exec")
+        .arg("--skip-git-repo-check")
+        .arg("-c")
+        .arg(&provider_override)
+        .arg("-c")
+        .arg("model_provider=\"mock\"")
+        .arg("-c")
+        .arg(format!(
+            "experimental_instructions_file=\"{custom_path_str}\""
+        ))
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg("hello?\n");
+    cmd.env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("OPENAI_BASE_URL", format!("{}/v1", server.uri()));
+
+    let output = cmd.output().unwrap();
+    println!("Status: {}", output.status);
+    println!("Stdout:\n{}", String::from_utf8_lossy(&output.stdout));
+    println!("Stderr:\n{}", String::from_utf8_lossy(&output.stderr));
+    assert!(output.status.success());
+
+    // Inspect the captured request and verify our custom base instructions were
+    // included in the `instructions` field.
+    let request = &server.received_requests().await.unwrap()[0];
+    let body = request.body_json::<serde_json::Value>().unwrap();
+    let instructions = body
+        .get("instructions")
+        .and_then(|v| v.as_str())
+        .unwrap_or_default()
+        .to_string();
+    assert!(
+        instructions.contains(marker),
+        "instructions did not contain custom marker; got: {instructions}"
+    );
+}
+
 /// Tests streaming responses through the CLI using a local SSE fixture file.
 /// This test:
 /// 1. Uses a pre-recorded SSE response fixture instead of a live server
@@ -370,9 +460,14 @@ async fn integration_git_info_unit_test() {
     // 1. Create temp directory for git repo
     let temp_dir = TempDir::new().unwrap();
     let git_repo = temp_dir.path().to_path_buf();
+    let envs = vec![
+        ("GIT_CONFIG_GLOBAL", "/dev/null"),
+        ("GIT_CONFIG_NOSYSTEM", "1"),
+    ];
 
     // 2. Initialize a git repository with some content
     let init_output = std::process::Command::new("git")
+        .envs(envs.clone())
         .args(["init"])
         .current_dir(&git_repo)
         .output()
@@ -381,12 +476,14 @@ async fn integration_git_info_unit_test() {
 
     // Configure git user (required for commits)
     std::process::Command::new("git")
+        .envs(envs.clone())
         .args(["config", "user.name", "Integration Test"])
         .current_dir(&git_repo)
         .output()
         .unwrap();
 
     std::process::Command::new("git")
+        .envs(envs.clone())
         .args(["config", "user.email", "test@example.com"])
         .current_dir(&git_repo)
         .output()
@@ -397,12 +494,14 @@ async fn integration_git_info_unit_test() {
     std::fs::write(&test_file, "integration test content").unwrap();
 
     std::process::Command::new("git")
+        .envs(envs.clone())
         .args(["add", "."])
         .current_dir(&git_repo)
         .output()
         .unwrap();
 
     let commit_output = std::process::Command::new("git")
+        .envs(envs.clone())
         .args(["commit", "-m", "Integration test commit"])
         .current_dir(&git_repo)
         .output()
@@ -411,6 +510,7 @@ async fn integration_git_info_unit_test() {
 
     // Create a branch to test branch detection
     std::process::Command::new("git")
+        .envs(envs.clone())
         .args(["checkout", "-b", "integration-test-branch"])
         .current_dir(&git_repo)
         .output()
@@ -418,6 +518,7 @@ async fn integration_git_info_unit_test() {
 
     // Add a remote to test repository URL detection
     std::process::Command::new("git")
+        .envs(envs.clone())
         .args([
             "remote",
             "add",

codex-rs/core/tests/client.rs

@@ -1,11 +1,22 @@
+#![allow(clippy::expect_used)]
+#![allow(clippy::unwrap_used)]
+use std::path::PathBuf;
+
+use chrono::Utc;
 use codex_core::Codex;
 use codex_core::CodexSpawnOk;
 use codex_core::ModelProviderInfo;
-use codex_core::exec::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+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::protocol::SessionConfiguredEvent;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_login::AuthDotJson;
+use codex_login::AuthMode;
+use codex_login::CodexAuth;
+use codex_login::TokenData;
 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;
@@ -13,14 +24,42 @@ use tempfile::TempDir;
 use wiremock::Mock;
 use wiremock::MockServer;
 use wiremock::ResponseTemplate;
+use wiremock::matchers::header_regex;
 use wiremock::matchers::method;
 use wiremock::matchers::path;
+use wiremock::matchers::query_param;
 
 /// 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)
 }
 
+fn assert_message_role(request_body: &serde_json::Value, role: &str) {
+    assert_eq!(request_body["role"].as_str().unwrap(), role);
+}
+
+fn assert_message_starts_with(request_body: &serde_json::Value, text: &str) {
+    let content = request_body["content"][0]["text"]
+        .as_str()
+        .expect("invalid message content");
+
+    assert!(
+        content.starts_with(text),
+        "expected message content '{content}' to start with '{text}'"
+    );
+}
+
+fn assert_message_ends_with(request_body: &serde_json::Value, text: &str) {
+    let content = request_body["content"][0]["text"]
+        .as_str()
+        .expect("invalid message content");
+
+    assert!(
+        content.ends_with(text),
+        "expected message content '{content}' to end with '{text}'"
+    );
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn includes_session_id_and_model_headers_in_request() {
     #![allow(clippy::unwrap_used)]
@@ -48,32 +87,23 @@ async fn includes_session_id_and_model_headers_in_request() {
         .await;
 
     let model_provider = ModelProviderInfo {
-        name: "openai".into(),
-        base_url: format!("{}/v1", server.uri()),
-        // Environment variable that should exist in the test environment.
-        // ModelClient will return an error if the environment variable for the
-        // provider is not set.
-        env_key: Some("PATH".into()),
-        env_key_instructions: None,
-        wire_api: codex_core::WireApi::Responses,
-        query_params: None,
-        http_headers: Some(
-            [("originator".to_string(), "codex_cli_rs".to_string())]
-                .into_iter()
-                .collect(),
-        ),
-        env_http_headers: None,
-        request_max_retries: Some(0),
-        stream_max_retries: Some(0),
-        stream_idle_timeout_ms: None,
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
     };
 
     // Init session
     let codex_home = TempDir::new().unwrap();
     let mut config = load_default_config_for_test(&codex_home);
     config.model_provider = model_provider;
+
     let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
-    let CodexSpawnOk { codex, .. } = Codex::spawn(config, ctrl_c.clone()).await.unwrap();
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(CodexAuth::from_api_key("Test API Key".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
 
     codex
         .submit(Op::UserInput {
@@ -95,15 +125,20 @@ async fn includes_session_id_and_model_headers_in_request() {
 
     // get request from the server
     let request = &server.received_requests().await.unwrap()[0];
-    let request_body = request.headers.get("session_id").unwrap();
-    let originator = request.headers.get("originator").unwrap();
+    let request_session_id = request.headers.get("session_id").unwrap();
+    let request_authorization = request.headers.get("authorization").unwrap();
+    let request_originator = request.headers.get("originator").unwrap();
 
     assert!(current_session_id.is_some());
     assert_eq!(
-        request_body.to_str().unwrap(),
+        request_session_id.to_str().unwrap(),
         current_session_id.as_ref().unwrap()
     );
-    assert_eq!(originator.to_str().unwrap(), "codex_cli_rs");
+    assert_eq!(request_originator.to_str().unwrap(), "codex_cli_rs");
+    assert_eq!(
+        request_authorization.to_str().unwrap(),
+        "Bearer Test API Key"
+    );
 }
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
@@ -126,22 +161,9 @@ async fn includes_base_instructions_override_in_request() {
         .await;
 
     let model_provider = ModelProviderInfo {
-        name: "openai".into(),
-        base_url: format!("{}/v1", server.uri()),
-        // Environment variable that should exist in the test environment.
-        // ModelClient will return an error if the environment variable for the
-        // provider is not set.
-        env_key: Some("PATH".into()),
-        env_key_instructions: None,
-        wire_api: codex_core::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: None,
+        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);
 
@@ -149,7 +171,13 @@ async fn includes_base_instructions_override_in_request() {
     config.model_provider = model_provider;
 
     let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
-    let CodexSpawnOk { codex, .. } = Codex::spawn(config, ctrl_c.clone()).await.unwrap();
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(CodexAuth::from_api_key("Test API Key".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
 
     codex
         .submit(Op::UserInput {
@@ -172,3 +200,381 @@ async fn includes_base_instructions_override_in_request() {
             .contains("test instructions")
     );
 }
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn originator_config_override_is_used() {
+    #![allow(clippy::unwrap_used)]
+
+    // Mock server
+    let server = MockServer::start().await;
+
+    let first = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse_completed("resp1"), "text/event-stream");
+
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(first)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/v1", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider = model_provider;
+    config.internal_originator = Some("my_override".to_string());
+
+    let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(CodexAuth::from_api_key("Test API Key".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello".into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let request = &server.received_requests().await.unwrap()[0];
+    let request_originator = request.headers.get("originator").unwrap();
+    assert_eq!(request_originator.to_str().unwrap(), "my_override");
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn chatgpt_auth_sends_correct_request() {
+    #![allow(clippy::unwrap_used)]
+
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // Mock server
+    let server = MockServer::start().await;
+
+    // First request – must NOT include `previous_response_id`.
+    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("/api/codex/responses"))
+        .respond_with(first)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let model_provider = ModelProviderInfo {
+        base_url: Some(format!("{}/api/codex", server.uri())),
+        ..built_in_model_providers()["openai"].clone()
+    };
+
+    // Init session
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider = model_provider;
+    let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(auth_from_token("Access Token".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello".into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    let EventMsg::SessionConfigured(SessionConfiguredEvent { session_id, .. }) =
+        wait_for_event(&codex, |ev| matches!(ev, EventMsg::SessionConfigured(_))).await
+    else {
+        unreachable!()
+    };
+
+    let current_session_id = Some(session_id.to_string());
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // get request from the server
+    let request = &server.received_requests().await.unwrap()[0];
+    let request_session_id = request.headers.get("session_id").unwrap();
+    let request_authorization = request.headers.get("authorization").unwrap();
+    let request_originator = request.headers.get("originator").unwrap();
+    let request_chatgpt_account_id = request.headers.get("chatgpt-account-id").unwrap();
+    let request_body = request.body_json::<serde_json::Value>().unwrap();
+
+    assert!(current_session_id.is_some());
+    assert_eq!(
+        request_session_id.to_str().unwrap(),
+        current_session_id.as_ref().unwrap()
+    );
+    assert_eq!(request_originator.to_str().unwrap(), "codex_cli_rs");
+    assert_eq!(
+        request_authorization.to_str().unwrap(),
+        "Bearer Access Token"
+    );
+    assert_eq!(request_chatgpt_account_id.to_str().unwrap(), "account_id");
+    assert!(!request_body["store"].as_bool().unwrap());
+    assert!(request_body["stream"].as_bool().unwrap());
+    assert_eq!(
+        request_body["include"][0].as_str().unwrap(),
+        "reasoning.encrypted_content"
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn includes_user_instructions_message_in_request() {
+    #![allow(clippy::unwrap_used)]
+
+    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 ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(CodexAuth::from_api_key("Test API Key".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello".into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let request = &server.received_requests().await.unwrap()[0];
+    let request_body = request.body_json::<serde_json::Value>().unwrap();
+
+    assert!(
+        !request_body["instructions"]
+            .as_str()
+            .unwrap()
+            .contains("be nice")
+    );
+    assert_message_role(&request_body["input"][0], "user");
+    assert_message_starts_with(&request_body["input"][0], "<environment_context>\n\n");
+    assert_message_ends_with(&request_body["input"][0], "</environment_context>");
+    assert_message_role(&request_body["input"][1], "user");
+    assert_message_starts_with(&request_body["input"][1], "<user_instructions>\n\n");
+    assert_message_ends_with(&request_body["input"][1], "</user_instructions>");
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn azure_overrides_assign_properties_used_for_responses_url() {
+    #![allow(clippy::unwrap_used)]
+
+    let existing_env_var_with_random_value = if cfg!(windows) { "USERNAME" } else { "USER" };
+
+    // Mock server
+    let server = MockServer::start().await;
+
+    // First request – must NOT include `previous_response_id`.
+    let first = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse_completed("resp1"), "text/event-stream");
+
+    // Expect POST to /openai/responses with api-version query param
+    Mock::given(method("POST"))
+        .and(path("/openai/responses"))
+        .and(query_param("api-version", "2025-04-01-preview"))
+        .and(header_regex("Custom-Header", "Value"))
+        .and(header_regex(
+            "Authorization",
+            format!(
+                "Bearer {}",
+                std::env::var(existing_env_var_with_random_value).unwrap()
+            )
+            .as_str(),
+        ))
+        .respond_with(first)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let provider = ModelProviderInfo {
+        name: "custom".to_string(),
+        base_url: Some(format!("{}/openai", server.uri())),
+        // Reuse the existing environment variable to avoid using unsafe code
+        env_key: Some(existing_env_var_with_random_value.to_string()),
+        query_params: Some(std::collections::HashMap::from([(
+            "api-version".to_string(),
+            "2025-04-01-preview".to_string(),
+        )])),
+        env_key_instructions: None,
+        wire_api: WireApi::Responses,
+        http_headers: Some(std::collections::HashMap::from([(
+            "Custom-Header".to_string(),
+            "Value".to_string(),
+        )])),
+        env_http_headers: None,
+        request_max_retries: None,
+        stream_max_retries: None,
+        stream_idle_timeout_ms: None,
+        requires_openai_auth: false,
+    };
+
+    // Init session
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider = provider;
+
+    let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
+    let CodexSpawnOk { codex, .. } = Codex::spawn(config, None, ctrl_c.clone()).await.unwrap();
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello".into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn env_var_overrides_loaded_auth() {
+    #![allow(clippy::unwrap_used)]
+
+    let existing_env_var_with_random_value = if cfg!(windows) { "USERNAME" } else { "USER" };
+
+    // Mock server
+    let server = MockServer::start().await;
+
+    // First request – must NOT include `previous_response_id`.
+    let first = ResponseTemplate::new(200)
+        .insert_header("content-type", "text/event-stream")
+        .set_body_raw(sse_completed("resp1"), "text/event-stream");
+
+    // Expect POST to /openai/responses with api-version query param
+    Mock::given(method("POST"))
+        .and(path("/openai/responses"))
+        .and(query_param("api-version", "2025-04-01-preview"))
+        .and(header_regex("Custom-Header", "Value"))
+        .and(header_regex(
+            "Authorization",
+            format!(
+                "Bearer {}",
+                std::env::var(existing_env_var_with_random_value).unwrap()
+            )
+            .as_str(),
+        ))
+        .respond_with(first)
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let provider = ModelProviderInfo {
+        name: "custom".to_string(),
+        base_url: Some(format!("{}/openai", server.uri())),
+        // Reuse the existing environment variable to avoid using unsafe code
+        env_key: Some(existing_env_var_with_random_value.to_string()),
+        query_params: Some(std::collections::HashMap::from([(
+            "api-version".to_string(),
+            "2025-04-01-preview".to_string(),
+        )])),
+        env_key_instructions: None,
+        wire_api: WireApi::Responses,
+        http_headers: Some(std::collections::HashMap::from([(
+            "Custom-Header".to_string(),
+            "Value".to_string(),
+        )])),
+        env_http_headers: None,
+        request_max_retries: None,
+        stream_max_retries: None,
+        stream_idle_timeout_ms: None,
+        requires_openai_auth: false,
+    };
+
+    // Init session
+    let codex_home = TempDir::new().unwrap();
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider = provider;
+
+    let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(auth_from_token("Default Access Token".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
+
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello".into(),
+            }],
+        })
+        .await
+        .unwrap();
+
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+}
+
+fn auth_from_token(id_token: String) -> CodexAuth {
+    CodexAuth::new(
+        None,
+        AuthMode::ChatGPT,
+        PathBuf::new(),
+        Some(AuthDotJson {
+            openai_api_key: None,
+            tokens: Some(TokenData {
+                id_token,
+                access_token: "Access Token".to_string(),
+                refresh_token: "test".to_string(),
+                account_id: Some("account_id".to_string()),
+            }),
+            last_refresh: Some(Utc::now()),
+        }),
+    )
+}

codex-rs/core/tests/compact.rs

@@ -0,0 +1,254 @@
+#![expect(clippy::unwrap_used)]
+
+use codex_core::Codex;
+use codex_core::CodexSpawnOk;
+use codex_core::ModelProviderInfo;
+use codex_core::built_in_model_providers;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::InputItem;
+use codex_core::protocol::Op;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_login::CodexAuth;
+use core_test_support::load_default_config_for_test;
+use core_test_support::wait_for_event;
+use serde_json::Value;
+use tempfile::TempDir;
+use wiremock::Mock;
+use wiremock::MockServer;
+use wiremock::ResponseTemplate;
+use wiremock::matchers::method;
+use wiremock::matchers::path;
+
+use pretty_assertions::assert_eq;
+
+// --- 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";
+const THIRD_USER_MSG: &str = "next turn";
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn summarize_context_three_requests_and_instructions() {
+    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;
+    }
+
+    // Set up a mock server that we can inspect after the run.
+    let server = MockServer::start().await;
+
+    // SSE 1: assistant replies normally so it is recorded in history.
+    let sse1 = sse(vec![
+        ev_assistant_message("m1", FIRST_REPLY),
+        ev_completed("r1"),
+    ]);
+
+    // SSE 2: summarizer returns a summary message.
+    let sse2 = sse(vec![
+        ev_assistant_message("m2", SUMMARY_TEXT),
+        ev_completed("r2"),
+    ]);
+
+    // SSE 3: minimal completed; we only need to capture the request body.
+    let sse3 = sse(vec![ev_completed("r3")]);
+
+    // Mount three expectations, one per request, matched by body content.
+    let first_matcher = |req: &wiremock::Request| {
+        let body = std::str::from_utf8(&req.body).unwrap_or("");
+        body.contains("\"text\":\"hello world\"")
+            && !body.contains(&format!("\"text\":\"{SUMMARIZE_TRIGGER}\""))
+    };
+    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;
+
+    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;
+
+    // Build config pointing to the mock server and spawn Codex.
+    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 ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(CodexAuth::from_api_key("dummy".to_string())),
+        ctrl_c.clone(),
+    )
+    .await
+    .unwrap();
+
+    // 1) Normal user input – should hit server once.
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: "hello world".into(),
+            }],
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // 2) Summarize – second hit with summarization instructions.
+    codex.submit(Op::Compact).await.unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // 3) Next user input – third hit; history should include only the summary.
+    codex
+        .submit(Op::UserInput {
+            items: vec![InputItem::Text {
+                text: THIRD_USER_MSG.into(),
+            }],
+        })
+        .await
+        .unwrap();
+    wait_for_event(&codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // Inspect the three captured requests.
+    let requests = server.received_requests().await.unwrap();
+    assert_eq!(requests.len(), 3, "expected exactly three requests");
+
+    let req1 = &requests[0];
+    let req2 = &requests[1];
+    let req3 = &requests[2];
+
+    let body1 = req1.body_json::<serde_json::Value>().unwrap();
+    let body2 = req2.body_json::<serde_json::Value>().unwrap();
+    let body3 = req3.body_json::<serde_json::Value>().unwrap();
+
+    // System instructions should change for the summarization turn.
+    let instr1 = body1.get("instructions").and_then(|v| v.as_str()).unwrap();
+    let instr2 = body2.get("instructions").and_then(|v| v.as_str()).unwrap();
+    assert_ne!(
+        instr1, instr2,
+        "summarization should override base instructions"
+    );
+    assert!(
+        instr2.contains("You are a summarization assistant"),
+        "summarization instructions not applied"
+    );
+
+    // The summarization request should include the injected user input marker.
+    let input2 = body2.get("input").and_then(|v| v.as_array()).unwrap();
+    // The last item is the user message created from the injected input.
+    let last2 = input2.last().unwrap();
+    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));
+
+    // Third request must contain only the summary from step 2 as prior history plus 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"
+    );
+
+    // Collect all (role, text) message tuples.
+    let mut messages: Vec<(String, String)> = Vec::new();
+    for item in input3 {
+        if item["type"].as_str() == Some("message") {
+            let role = item["role"].as_str().unwrap_or_default().to_string();
+            let text = item["content"][0]["text"]
+                .as_str()
+                .unwrap_or_default()
+                .to_string();
+            messages.push((role, text));
+        }
+    }
+
+    // Exactly one assistant message should remain after compaction 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!(
+        messages
+            .iter()
+            .any(|(r, t)| r == "user" && t == THIRD_USER_MSG),
+        "third request should include the new user message"
+    );
+    assert!(
+        !messages.iter().any(|(_, t)| t.contains("hello world")),
+        "third request should not include the original user input"
+    );
+    assert!(
+        !messages.iter().any(|(_, t)| t.contains(SUMMARIZE_TRIGGER)),
+        "third request should not include the summarize trigger"
+    );
+}

codex-rs/core/tests/exec.rs

@@ -0,0 +1,71 @@
+#![cfg(target_os = "macos")]
+#![expect(clippy::expect_used)]
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use codex_core::exec::ExecParams;
+use codex_core::exec::SandboxType;
+use codex_core::exec::process_exec_tool_call;
+use codex_core::protocol::SandboxPolicy;
+use codex_core::spawn::CODEX_SANDBOX_ENV_VAR;
+use tempfile::TempDir;
+use tokio::sync::Notify;
+
+use codex_core::get_platform_sandbox;
+
+async fn run_test_cmd(tmp: TempDir, cmd: Vec<&str>, should_be_ok: bool) {
+    if std::env::var(CODEX_SANDBOX_ENV_VAR) == Ok("seatbelt".to_string()) {
+        eprintln!("{CODEX_SANDBOX_ENV_VAR} is set to 'seatbelt', skipping test.");
+        return;
+    }
+
+    let sandbox_type = get_platform_sandbox().expect("should be able to get sandbox type");
+    assert_eq!(sandbox_type, SandboxType::MacosSeatbelt);
+
+    let params = ExecParams {
+        command: cmd.iter().map(|s| s.to_string()).collect(),
+        cwd: tmp.path().to_path_buf(),
+        timeout_ms: Some(1000),
+        env: HashMap::new(),
+        with_escalated_permissions: None,
+        justification: None,
+    };
+
+    let ctrl_c = Arc::new(Notify::new());
+    let policy = SandboxPolicy::new_read_only_policy();
+
+    let result = process_exec_tool_call(params, sandbox_type, ctrl_c, &policy, &None, None).await;
+
+    assert!(result.is_ok() == should_be_ok);
+}
+
+/// Command succeeds with exit code 0 normally
+#[tokio::test]
+async fn exit_code_0_succeeds() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let cmd = vec!["echo", "hello"];
+
+    run_test_cmd(tmp, cmd, true).await
+}
+
+/// Command not found returns exit code 127, this is not considered a sandbox error
+#[tokio::test]
+async fn exit_command_not_found_is_ok() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let cmd = vec!["/bin/bash", "-c", "nonexistent_command_12345"];
+    run_test_cmd(tmp, cmd, true).await
+}
+
+/// Writing a file fails and should be considered a sandbox error
+#[tokio::test]
+async fn write_file_fails_as_sandbox_error() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let path = tmp.path().join("test.txt");
+    let cmd = vec![
+        "/user/bin/touch",
+        path.to_str().expect("should be able to get path"),
+    ];
+
+    run_test_cmd(tmp, cmd, false).await;
+}

codex-rs/core/tests/exec_stream_events.rs

@@ -0,0 +1,147 @@
+#![cfg(unix)]
+
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use async_channel::Receiver;
+use codex_core::exec::ExecParams;
+use codex_core::exec::SandboxType;
+use codex_core::exec::StdoutStream;
+use codex_core::exec::process_exec_tool_call;
+use codex_core::protocol::Event;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::ExecCommandOutputDeltaEvent;
+use codex_core::protocol::ExecOutputStream;
+use codex_core::protocol::SandboxPolicy;
+use tokio::sync::Notify;
+
+fn collect_stdout_events(rx: Receiver<Event>) -> Vec<u8> {
+    let mut out = Vec::new();
+    while let Ok(ev) = rx.try_recv() {
+        if let EventMsg::ExecCommandOutputDelta(ExecCommandOutputDeltaEvent {
+            stream: ExecOutputStream::Stdout,
+            chunk,
+            ..
+        }) = ev.msg
+        {
+            out.extend_from_slice(&chunk);
+        }
+    }
+    out
+}
+
+#[tokio::test]
+async fn test_exec_stdout_stream_events_echo() {
+    let (tx, rx) = async_channel::unbounded::<Event>();
+
+    let stdout_stream = StdoutStream {
+        sub_id: "test-sub".to_string(),
+        call_id: "call-1".to_string(),
+        tx_event: tx,
+    };
+
+    let cmd = vec![
+        "/bin/sh".to_string(),
+        "-c".to_string(),
+        // Use printf for predictable behavior across shells
+        "printf 'hello-world\n'".to_string(),
+    ];
+
+    let params = ExecParams {
+        command: cmd,
+        cwd: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
+        timeout_ms: Some(5_000),
+        env: HashMap::new(),
+        with_escalated_permissions: None,
+        justification: None,
+    };
+
+    let ctrl_c = Arc::new(Notify::new());
+    let policy = SandboxPolicy::new_read_only_policy();
+
+    let result = process_exec_tool_call(
+        params,
+        SandboxType::None,
+        ctrl_c,
+        &policy,
+        &None,
+        Some(stdout_stream),
+    )
+    .await;
+
+    let result = match result {
+        Ok(r) => r,
+        Err(e) => panic!("process_exec_tool_call failed: {e}"),
+    };
+
+    assert_eq!(result.exit_code, 0);
+    assert_eq!(result.stdout, "hello-world\n");
+
+    let streamed = collect_stdout_events(rx);
+    // We should have received at least the same contents (possibly in one chunk)
+    assert_eq!(String::from_utf8_lossy(&streamed), "hello-world\n");
+}
+
+#[tokio::test]
+async fn test_exec_stderr_stream_events_echo() {
+    let (tx, rx) = async_channel::unbounded::<Event>();
+
+    let stdout_stream = StdoutStream {
+        sub_id: "test-sub".to_string(),
+        call_id: "call-2".to_string(),
+        tx_event: tx,
+    };
+
+    let cmd = vec![
+        "/bin/sh".to_string(),
+        "-c".to_string(),
+        // Write to stderr explicitly
+        "printf 'oops\n' 1>&2".to_string(),
+    ];
+
+    let params = ExecParams {
+        command: cmd,
+        cwd: std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")),
+        timeout_ms: Some(5_000),
+        env: HashMap::new(),
+        with_escalated_permissions: None,
+        justification: None,
+    };
+
+    let ctrl_c = Arc::new(Notify::new());
+    let policy = SandboxPolicy::new_read_only_policy();
+
+    let result = process_exec_tool_call(
+        params,
+        SandboxType::None,
+        ctrl_c,
+        &policy,
+        &None,
+        Some(stdout_stream),
+    )
+    .await;
+
+    let result = match result {
+        Ok(r) => r,
+        Err(e) => panic!("process_exec_tool_call failed: {e}"),
+    };
+
+    assert_eq!(result.exit_code, 0);
+    assert_eq!(result.stdout, "");
+    assert_eq!(result.stderr, "oops\n");
+
+    // Collect only stderr delta events
+    let mut err = Vec::new();
+    while let Ok(ev) = rx.try_recv() {
+        if let EventMsg::ExecCommandOutputDelta(ExecCommandOutputDeltaEvent {
+            stream: ExecOutputStream::Stderr,
+            chunk,
+            ..
+        }) = ev.msg
+        {
+            err.extend_from_slice(&chunk);
+        }
+    }
+    assert_eq!(String::from_utf8_lossy(&err), "oops\n");
+}

codex-rs/core/tests/live_agent.rs

@@ -50,7 +50,7 @@ async fn spawn_codex() -> Result<Codex, CodexErr> {
     config.model_provider.request_max_retries = Some(2);
     config.model_provider.stream_max_retries = Some(2);
     let CodexSpawnOk { codex: agent, .. } =
-        Codex::spawn(config, std::sync::Arc::new(Notify::new())).await?;
+        Codex::spawn(config, None, std::sync::Arc::new(Notify::new())).await?;
 
     Ok(agent)
 }
@@ -177,8 +177,7 @@ async fn live_shell_function_call() {
         match ev.msg {
             EventMsg::ExecCommandBegin(codex_core::protocol::ExecCommandBeginEvent {
                 command,
-                call_id: _,
-                cwd: _,
+                ..
             }) => {
                 assert_eq!(command, vec!["echo", MARKER]);
                 saw_begin = true;
@@ -186,8 +185,7 @@ async fn live_shell_function_call() {
             EventMsg::ExecCommandEnd(codex_core::protocol::ExecCommandEndEvent {
                 stdout,
                 exit_code,
-                call_id: _,
-                stderr: _,
+                ..
             }) => {
                 assert_eq!(exit_code, 0, "echo returned non‑zero exit code");
                 assert!(stdout.contains(MARKER));

codex-rs/core/tests/sandbox.rs

@@ -0,0 +1,195 @@
+#![cfg(target_os = "macos")]
+#![expect(clippy::expect_used)]
+
+use std::collections::HashMap;
+use std::path::Path;
+use std::path::PathBuf;
+
+use codex_core::protocol::SandboxPolicy;
+use codex_core::seatbelt::spawn_command_under_seatbelt;
+use codex_core::spawn::CODEX_SANDBOX_ENV_VAR;
+use codex_core::spawn::StdioPolicy;
+use tempfile::TempDir;
+
+struct TestScenario {
+    repo_parent: PathBuf,
+    file_outside_repo: PathBuf,
+    repo_root: PathBuf,
+    file_in_repo_root: PathBuf,
+    file_in_dot_git_dir: PathBuf,
+}
+
+struct TestExpectations {
+    file_outside_repo_is_writable: bool,
+    file_in_repo_root_is_writable: bool,
+    file_in_dot_git_dir_is_writable: bool,
+}
+
+impl TestScenario {
+    async fn run_test(&self, policy: &SandboxPolicy, expectations: TestExpectations) {
+        if std::env::var(CODEX_SANDBOX_ENV_VAR) == Ok("seatbelt".to_string()) {
+            eprintln!("{CODEX_SANDBOX_ENV_VAR} is set to 'seatbelt', skipping test.");
+            return;
+        }
+
+        assert_eq!(
+            touch(&self.file_outside_repo, policy).await,
+            expectations.file_outside_repo_is_writable
+        );
+        assert_eq!(
+            self.file_outside_repo.exists(),
+            expectations.file_outside_repo_is_writable
+        );
+
+        assert_eq!(
+            touch(&self.file_in_repo_root, policy).await,
+            expectations.file_in_repo_root_is_writable
+        );
+        assert_eq!(
+            self.file_in_repo_root.exists(),
+            expectations.file_in_repo_root_is_writable
+        );
+
+        assert_eq!(
+            touch(&self.file_in_dot_git_dir, policy).await,
+            expectations.file_in_dot_git_dir_is_writable
+        );
+        assert_eq!(
+            self.file_in_dot_git_dir.exists(),
+            expectations.file_in_dot_git_dir_is_writable
+        );
+    }
+}
+
+/// If the user has added a workspace root that is not a Git repo root, then
+/// the user has to specify `--skip-git-repo-check` or go through some
+/// interstitial that indicates they are taking on some risk because Git
+/// cannot be used to backup their work before the agent begins.
+///
+/// Because the user has agreed to this risk, we do not try find all .git
+/// folders in the workspace and block them (though we could change our
+/// position on this in the future).
+#[tokio::test]
+async fn if_parent_of_repo_is_writable_then_dot_git_folder_is_writable() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let test_scenario = create_test_scenario(&tmp);
+    let policy = SandboxPolicy::WorkspaceWrite {
+        writable_roots: vec![test_scenario.repo_parent.clone()],
+        network_access: false,
+        include_default_writable_roots: false,
+    };
+
+    test_scenario
+        .run_test(
+            &policy,
+            TestExpectations {
+                file_outside_repo_is_writable: true,
+                file_in_repo_root_is_writable: true,
+                file_in_dot_git_dir_is_writable: true,
+            },
+        )
+        .await;
+}
+
+/// When the writable root is the root of a Git repository (as evidenced by the
+/// presence of a .git folder), then the .git folder should be read-only if
+/// the policy is `WorkspaceWrite`.
+#[tokio::test]
+async fn if_git_repo_is_writable_root_then_dot_git_folder_is_read_only() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let test_scenario = create_test_scenario(&tmp);
+    let policy = SandboxPolicy::WorkspaceWrite {
+        writable_roots: vec![test_scenario.repo_root.clone()],
+        network_access: false,
+        include_default_writable_roots: false,
+    };
+
+    test_scenario
+        .run_test(
+            &policy,
+            TestExpectations {
+                file_outside_repo_is_writable: false,
+                file_in_repo_root_is_writable: true,
+                file_in_dot_git_dir_is_writable: false,
+            },
+        )
+        .await;
+}
+
+/// Under DangerFullAccess, all writes should be permitted anywhere on disk,
+/// including inside the .git folder.
+#[tokio::test]
+async fn danger_full_access_allows_all_writes() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let test_scenario = create_test_scenario(&tmp);
+    let policy = SandboxPolicy::DangerFullAccess;
+
+    test_scenario
+        .run_test(
+            &policy,
+            TestExpectations {
+                file_outside_repo_is_writable: true,
+                file_in_repo_root_is_writable: true,
+                file_in_dot_git_dir_is_writable: true,
+            },
+        )
+        .await;
+}
+
+/// Under ReadOnly, writes should not be permitted anywhere on disk.
+#[tokio::test]
+async fn read_only_forbids_all_writes() {
+    let tmp = TempDir::new().expect("should be able to create temp dir");
+    let test_scenario = create_test_scenario(&tmp);
+    let policy = SandboxPolicy::ReadOnly;
+
+    test_scenario
+        .run_test(
+            &policy,
+            TestExpectations {
+                file_outside_repo_is_writable: false,
+                file_in_repo_root_is_writable: false,
+                file_in_dot_git_dir_is_writable: false,
+            },
+        )
+        .await;
+}
+
+fn create_test_scenario(tmp: &TempDir) -> TestScenario {
+    let repo_parent = tmp.path().to_path_buf();
+    let repo_root = repo_parent.join("repo");
+    let dot_git_dir = repo_root.join(".git");
+
+    std::fs::create_dir(&repo_root).expect("should be able to create repo root");
+    std::fs::create_dir(&dot_git_dir).expect("should be able to create .git dir");
+
+    TestScenario {
+        file_outside_repo: repo_parent.join("outside.txt"),
+        repo_parent,
+        file_in_repo_root: repo_root.join("repo_file.txt"),
+        repo_root,
+        file_in_dot_git_dir: dot_git_dir.join("dot_git_file.txt"),
+    }
+}
+
+/// Note that `path` must be absolute.
+async fn touch(path: &Path, policy: &SandboxPolicy) -> bool {
+    assert!(path.is_absolute(), "Path must be absolute: {path:?}");
+    let mut child = spawn_command_under_seatbelt(
+        vec![
+            "/usr/bin/touch".to_string(),
+            path.to_string_lossy().to_string(),
+        ],
+        policy,
+        std::env::current_dir().expect("should be able to get current dir"),
+        StdioPolicy::RedirectForShellTool,
+        HashMap::new(),
+    )
+    .await
+    .expect("should be able to spawn command under seatbelt");
+    child
+        .wait()
+        .await
+        .expect("should be able to wait for child process")
+        .success()
+}

codex-rs/core/tests/stream_no_completed.rs

@@ -6,10 +6,11 @@ use std::time::Duration;
 use codex_core::Codex;
 use codex_core::CodexSpawnOk;
 use codex_core::ModelProviderInfo;
-use codex_core::exec::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::InputItem;
 use codex_core::protocol::Op;
+use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use codex_login::CodexAuth;
 use core_test_support::load_default_config_for_test;
 use core_test_support::load_sse_fixture;
 use core_test_support::load_sse_fixture_with_id;
@@ -75,7 +76,7 @@ async fn retries_on_early_close() {
 
     let model_provider = ModelProviderInfo {
         name: "openai".into(),
-        base_url: format!("{}/v1", server.uri()),
+        base_url: Some(format!("{}/v1", server.uri())),
         // Environment variable that should exist in the test environment.
         // ModelClient will return an error if the environment variable for the
         // provider is not set.
@@ -89,13 +90,20 @@ async fn retries_on_early_close() {
         request_max_retries: Some(0),
         stream_max_retries: Some(1),
         stream_idle_timeout_ms: Some(2000),
+        requires_openai_auth: false,
     };
 
     let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());
     let codex_home = TempDir::new().unwrap();
     let mut config = load_default_config_for_test(&codex_home);
     config.model_provider = model_provider;
-    let CodexSpawnOk { codex, .. } = Codex::spawn(config, ctrl_c).await.unwrap();
+    let CodexSpawnOk { codex, .. } = Codex::spawn(
+        config,
+        Some(CodexAuth::from_api_key("Test API Key".to_string())),
+        ctrl_c,
+    )
+    .await
+    .unwrap();
 
     codex
         .submit(Op::UserInput {

codex-rs/exec/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-exec"
 version = { workspace = true }
-edition = "2024"
 
 [[bin]]
 name = "codex-exec"
@@ -19,12 +19,13 @@ anyhow = "1"
 chrono = "0.4.40"
 clap = { version = "4", features = ["derive"] }
 codex-arg0 = { path = "../arg0" }
-codex-core = { path = "../core" }
 codex-common = { path = "../common", features = [
     "cli",
     "elapsed",
     "sandbox_summary",
 ] }
+codex-core = { path = "../core" }
+codex-ollama = { path = "../ollama" }
 owo-colors = "4.2.0"
 serde_json = "1"
 shlex = "1.3.0"

codex-rs/exec/src/cli.rs

@@ -14,6 +14,9 @@ pub struct Cli {
     #[arg(long, short = 'm')]
     pub model: Option<String>,
 
+    #[arg(long = "oss", default_value_t = false)]
+    pub oss: bool,
+
     /// Select the sandbox policy to use when executing model-generated shell
     /// commands.
     #[arg(long = "sandbox", short = 's')]

codex-rs/exec/src/event_processor.rs

@@ -1,9 +1,6 @@
 use std::path::Path;
 
-use codex_common::summarize_sandbox_policy;
-use codex_core::WireApi;
 use codex_core::config::Config;
-use codex_core::model_supports_reasoning_summaries;
 use codex_core::protocol::Event;
 
 pub(crate) enum CodexStatus {
@@ -20,44 +17,14 @@ pub(crate) trait EventProcessor {
     fn process_event(&mut self, event: Event) -> CodexStatus;
 }
 
-pub(crate) fn create_config_summary_entries(config: &Config) -> Vec<(&'static str, String)> {
-    let mut entries = vec![
-        ("workdir", config.cwd.display().to_string()),
-        ("model", config.model.clone()),
-        ("provider", config.model_provider_id.clone()),
-        ("approval", config.approval_policy.to_string()),
-        ("sandbox", summarize_sandbox_policy(&config.sandbox_policy)),
-    ];
-    if config.model_provider.wire_api == WireApi::Responses
-        && model_supports_reasoning_summaries(config)
-    {
-        entries.push((
-            "reasoning effort",
-            config.model_reasoning_effort.to_string(),
-        ));
-        entries.push((
-            "reasoning summaries",
-            config.model_reasoning_summary.to_string(),
-        ));
-    }
-
-    entries
-}
-
-pub(crate) fn handle_last_message(
-    last_agent_message: Option<&str>,
-    last_message_path: Option<&Path>,
-) {
-    match (last_message_path, last_agent_message) {
-        (Some(path), Some(msg)) => write_last_message_file(msg, Some(path)),
-        (Some(path), None) => {
-            write_last_message_file("", Some(path));
-            eprintln!(
-                "Warning: no last agent message; wrote empty content to {}",
-                path.display()
-            );
-        }
-        (None, _) => eprintln!("Warning: no file to write last message to."),
+pub(crate) fn handle_last_message(last_agent_message: Option<&str>, output_file: &Path) {
+    let message = last_agent_message.unwrap_or_default();
+    write_last_message_file(message, Some(output_file));
+    if last_agent_message.is_none() {
+        eprintln!(
+            "Warning: no last agent message; wrote empty content to {}",
+            output_file.display()
+        );
     }
 }
 

codex-rs/exec/src/event_processor_with_human_output.rs

@@ -1,8 +1,12 @@
+use codex_common::elapsed::format_duration;
 use codex_common::elapsed::format_elapsed;
 use codex_core::config::Config;
+use codex_core::plan_tool::UpdatePlanArgs;
 use codex_core::protocol::AgentMessageDeltaEvent;
 use codex_core::protocol::AgentMessageEvent;
 use codex_core::protocol::AgentReasoningDeltaEvent;
+use codex_core::protocol::AgentReasoningRawContentDeltaEvent;
+use codex_core::protocol::AgentReasoningRawContentEvent;
 use codex_core::protocol::BackgroundEventEvent;
 use codex_core::protocol::ErrorEvent;
 use codex_core::protocol::Event;
@@ -10,6 +14,7 @@ use codex_core::protocol::EventMsg;
 use codex_core::protocol::ExecCommandBeginEvent;
 use codex_core::protocol::ExecCommandEndEvent;
 use codex_core::protocol::FileChange;
+use codex_core::protocol::McpInvocation;
 use codex_core::protocol::McpToolCallBeginEvent;
 use codex_core::protocol::McpToolCallEndEvent;
 use codex_core::protocol::PatchApplyBeginEvent;
@@ -17,6 +22,7 @@ use codex_core::protocol::PatchApplyEndEvent;
 use codex_core::protocol::SessionConfiguredEvent;
 use codex_core::protocol::TaskCompleteEvent;
 use codex_core::protocol::TokenUsage;
+use codex_core::protocol::TurnDiffEvent;
 use owo_colors::OwoColorize;
 use owo_colors::Style;
 use shlex::try_join;
@@ -27,8 +33,8 @@ use std::time::Instant;
 
 use crate::event_processor::CodexStatus;
 use crate::event_processor::EventProcessor;
-use crate::event_processor::create_config_summary_entries;
 use crate::event_processor::handle_last_message;
+use codex_common::create_config_summary_entries;
 
 /// This should be configurable. When used in CI, users may not want to impose
 /// a limit so they can see the full transcript.
@@ -37,11 +43,6 @@ pub(crate) struct EventProcessorWithHumanOutput {
     call_id_to_command: HashMap<String, ExecCommandBegin>,
     call_id_to_patch: HashMap<String, PatchApplyBegin>,
 
-    /// Tracks in-flight MCP tool calls so we can calculate duration and print
-    /// a concise summary when the corresponding `McpToolCallEnd` event is
-    /// received.
-    call_id_to_tool_call: HashMap<String, McpToolCallBegin>,
-
     // To ensure that --color=never is respected, ANSI escapes _must_ be added
     // using .style() with one of these fields. If you need a new style, add a
     // new field here.
@@ -56,8 +57,10 @@ pub(crate) struct EventProcessorWithHumanOutput {
 
     /// Whether to include `AgentReasoning` events in the output.
     show_agent_reasoning: bool,
+    show_raw_agent_reasoning: bool,
     answer_started: bool,
     reasoning_started: bool,
+    raw_reasoning_started: bool,
     last_message_path: Option<PathBuf>,
 }
 
@@ -69,7 +72,6 @@ impl EventProcessorWithHumanOutput {
     ) -> Self {
         let call_id_to_command = HashMap::new();
         let call_id_to_patch = HashMap::new();
-        let call_id_to_tool_call = HashMap::new();
 
         if with_ansi {
             Self {
@@ -82,10 +84,11 @@ impl EventProcessorWithHumanOutput {
                 red: Style::new().red(),
                 green: Style::new().green(),
                 cyan: Style::new().cyan(),
-                call_id_to_tool_call,
                 show_agent_reasoning: !config.hide_agent_reasoning,
+                show_raw_agent_reasoning: config.show_raw_agent_reasoning,
                 answer_started: false,
                 reasoning_started: false,
+                raw_reasoning_started: false,
                 last_message_path,
             }
         } else {
@@ -99,10 +102,11 @@ impl EventProcessorWithHumanOutput {
                 red: Style::new(),
                 green: Style::new(),
                 cyan: Style::new(),
-                call_id_to_tool_call,
                 show_agent_reasoning: !config.hide_agent_reasoning,
+                show_raw_agent_reasoning: config.show_raw_agent_reasoning,
                 answer_started: false,
                 reasoning_started: false,
+                raw_reasoning_started: false,
                 last_message_path,
             }
         }
@@ -111,15 +115,6 @@ impl EventProcessorWithHumanOutput {
 
 struct ExecCommandBegin {
     command: Vec<String>,
-    start_time: Instant,
-}
-
-/// Metadata captured when an `McpToolCallBegin` event is received.
-struct McpToolCallBegin {
-    /// Formatted invocation string, e.g. `server.tool({"city":"sf"})`.
-    invocation: String,
-    /// Timestamp when the call started so we can compute duration later.
-    start_time: Instant,
 }
 
 struct PatchApplyBegin {
@@ -183,10 +178,9 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 // Ignore.
             }
             EventMsg::TaskComplete(TaskCompleteEvent { last_agent_message }) => {
-                handle_last_message(
-                    last_agent_message.as_deref(),
-                    self.last_message_path.as_deref(),
-                );
+                if let Some(output_file) = self.last_message_path.as_deref() {
+                    handle_last_message(last_agent_message.as_deref(), output_file);
+                }
                 return CodexStatus::InitiateShutdown;
             }
             EventMsg::TokenCount(TokenUsage { total_tokens, .. }) => {
@@ -217,6 +211,32 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 #[allow(clippy::expect_used)]
                 std::io::stdout().flush().expect("could not flush stdout");
             }
+            EventMsg::AgentReasoningRawContent(AgentReasoningRawContentEvent { text }) => {
+                if !self.show_raw_agent_reasoning {
+                    return CodexStatus::Running;
+                }
+                if !self.raw_reasoning_started {
+                    print!("{text}");
+                    #[allow(clippy::expect_used)]
+                    std::io::stdout().flush().expect("could not flush stdout");
+                } else {
+                    println!();
+                    self.raw_reasoning_started = false;
+                }
+            }
+            EventMsg::AgentReasoningRawContentDelta(AgentReasoningRawContentDeltaEvent {
+                delta,
+            }) => {
+                if !self.show_raw_agent_reasoning {
+                    return CodexStatus::Running;
+                }
+                if !self.raw_reasoning_started {
+                    self.raw_reasoning_started = true;
+                }
+                print!("{delta}");
+                #[allow(clippy::expect_used)]
+                std::io::stdout().flush().expect("could not flush stdout");
+            }
             EventMsg::AgentMessage(AgentMessageEvent { message }) => {
                 // if answer_started is false, this means we haven't received any
                 // delta. Thus, we need to print the message as a new answer.
@@ -241,7 +261,6 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                     call_id.clone(),
                     ExecCommandBegin {
                         command: command.clone(),
-                        start_time: Instant::now(),
                     },
                 );
                 ts_println!(
@@ -252,20 +271,19 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                     cwd.to_string_lossy(),
                 );
             }
+            EventMsg::ExecCommandOutputDelta(_) => {}
             EventMsg::ExecCommandEnd(ExecCommandEndEvent {
                 call_id,
                 stdout,
                 stderr,
+                duration,
                 exit_code,
             }) => {
                 let exec_command = self.call_id_to_command.remove(&call_id);
-                let (duration, call) = if let Some(ExecCommandBegin {
-                    command,
-                    start_time,
-                }) = exec_command
+                let (duration, call) = if let Some(ExecCommandBegin { command, .. }) = exec_command
                 {
                     (
-                        format!(" in {}", format_elapsed(start_time)),
+                        format!(" in {}", format_duration(duration)),
                         format!("{}", escape_command(&command).style(self.bold)),
                     )
                 } else {
@@ -291,63 +309,33 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 println!("{}", truncated_output.style(self.dimmed));
             }
             EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
-                call_id,
-                server,
-                tool,
-                arguments,
+                call_id: _,
+                invocation,
             }) => {
-                // Build fully-qualified tool name: server.tool
-                let fq_tool_name = format!("{server}.{tool}");
-
-                // Format arguments as compact JSON so they fit on one line.
-                let args_str = arguments
-                    .as_ref()
-                    .map(|v: &serde_json::Value| {
-                        serde_json::to_string(v).unwrap_or_else(|_| v.to_string())
-                    })
-                    .unwrap_or_default();
-
-                let invocation = if args_str.is_empty() {
-                    format!("{fq_tool_name}()")
-                } else {
-                    format!("{fq_tool_name}({args_str})")
-                };
-
-                self.call_id_to_tool_call.insert(
-                    call_id.clone(),
-                    McpToolCallBegin {
-                        invocation: invocation.clone(),
-                        start_time: Instant::now(),
-                    },
-                );
-
                 ts_println!(
                     self,
                     "{} {}",
                     "tool".style(self.magenta),
-                    invocation.style(self.bold),
+                    format_mcp_invocation(&invocation).style(self.bold),
                 );
             }
             EventMsg::McpToolCallEnd(tool_call_end_event) => {
                 let is_success = tool_call_end_event.is_success();
-                let McpToolCallEndEvent { call_id, result } = tool_call_end_event;
-                // Retrieve start time and invocation for duration calculation and labeling.
-                let info = self.call_id_to_tool_call.remove(&call_id);
-
-                let (duration, invocation) = if let Some(McpToolCallBegin {
+                let McpToolCallEndEvent {
+                    call_id: _,
+                    result,
                     invocation,
-                    start_time,
-                    ..
-                }) = info
-                {
-                    (format!(" in {}", format_elapsed(start_time)), invocation)
-                } else {
-                    (String::new(), format!("tool('{call_id}')"))
-                };
+                    duration,
+                } = tool_call_end_event;
+
+                let duration = format!(" in {}", format_duration(duration));
 
                 let status_str = if is_success { "success" } else { "failed" };
                 let title_style = if is_success { self.green } else { self.red };
-                let title = format!("{invocation} {status_str}{duration}:");
+                let title = format!(
+                    "{} {status_str}{duration}:",
+                    format_mcp_invocation(&invocation)
+                );
 
                 ts_println!(self, "{}", title.style(title_style));
 
@@ -445,6 +433,7 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 stdout,
                 stderr,
                 success,
+                ..
             }) => {
                 let patch_begin = self.call_id_to_patch.remove(&call_id);
 
@@ -474,6 +463,10 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                     println!("{}", line.style(self.dimmed));
                 }
             }
+            EventMsg::TurnDiff(TurnDiffEvent { unified_diff }) => {
+                ts_println!(self, "{}", "turn diff:".style(self.magenta));
+                println!("{unified_diff}");
+            }
             EventMsg::ExecApprovalRequest(_) => {
                 // Should we exit?
             }
@@ -513,6 +506,11 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 ts_println!(self, "model: {}", model);
                 println!();
             }
+            EventMsg::PlanUpdate(plan_update_event) => {
+                let UpdatePlanArgs { explanation, plan } = plan_update_event;
+                ts_println!(self, "explanation: {explanation:?}");
+                ts_println!(self, "plan: {plan:?}");
+            }
             EventMsg::GetHistoryEntryResponse(_) => {
                 // Currently ignored in exec output.
             }
@@ -538,3 +536,21 @@ fn format_file_change(change: &FileChange) -> &'static str {
         } => "M",
     }
 }
+
+fn format_mcp_invocation(invocation: &McpInvocation) -> String {
+    // Build fully-qualified tool name: server.tool
+    let fq_tool_name = format!("{}.{}", invocation.server, invocation.tool);
+
+    // Format arguments as compact JSON so they fit on one line.
+    let args_str = invocation
+        .arguments
+        .as_ref()
+        .map(|v: &serde_json::Value| serde_json::to_string(v).unwrap_or_else(|_| v.to_string()))
+        .unwrap_or_default();
+
+    if args_str.is_empty() {
+        format!("{fq_tool_name}()")
+    } else {
+        format!("{fq_tool_name}({args_str})")
+    }
+}

codex-rs/exec/src/event_processor_with_json_output.rs

@@ -9,8 +9,8 @@ use serde_json::json;
 
 use crate::event_processor::CodexStatus;
 use crate::event_processor::EventProcessor;
-use crate::event_processor::create_config_summary_entries;
 use crate::event_processor::handle_last_message;
+use codex_common::create_config_summary_entries;
 
 pub(crate) struct EventProcessorWithJsonOutput {
     last_message_path: Option<PathBuf>,
@@ -46,10 +46,9 @@ impl EventProcessor for EventProcessorWithJsonOutput {
                 CodexStatus::Running
             }
             EventMsg::TaskComplete(TaskCompleteEvent { last_agent_message }) => {
-                handle_last_message(
-                    last_agent_message.as_deref(),
-                    self.last_message_path.as_deref(),
-                );
+                if let Some(output_file) = self.last_message_path.as_deref() {
+                    handle_last_message(last_agent_message.as_deref(), output_file);
+                }
                 CodexStatus::InitiateShutdown
             }
             EventMsg::ShutdownComplete => CodexStatus::Shutdown,

codex-rs/exec/src/lib.rs

@@ -9,6 +9,7 @@ use std::path::PathBuf;
 use std::sync::Arc;
 
 pub use cli::Cli;
+use codex_core::BUILT_IN_OSS_MODEL_PROVIDER_ID;
 use codex_core::codex_wrapper::CodexConversation;
 use codex_core::codex_wrapper::{self};
 use codex_core::config::Config;
@@ -21,6 +22,7 @@ use codex_core::protocol::InputItem;
 use codex_core::protocol::Op;
 use codex_core::protocol::TaskCompleteEvent;
 use codex_core::util::is_inside_git_repo;
+use codex_ollama::DEFAULT_OSS_MODEL;
 use event_processor_with_human_output::EventProcessorWithHumanOutput;
 use event_processor_with_json_output::EventProcessorWithJsonOutput;
 use tracing::debug;
@@ -34,7 +36,8 @@ use crate::event_processor::EventProcessor;
 pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()> {
     let Cli {
         images,
-        model,
+        model: model_cli_arg,
+        oss,
         config_profile,
         full_auto,
         dangerously_bypass_approvals_and_sandbox,
@@ -92,6 +95,20 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         ),
     };
 
+    // TODO(mbolin): Take a more thoughtful approach to logging.
+    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)),
+        )
+        .with_ansi(stderr_with_ansi)
+        .with_writer(std::io::stderr)
+        .try_init();
+
     let sandbox_mode = if full_auto {
         Some(SandboxMode::WorkspaceWrite)
     } else if dangerously_bypass_approvals_and_sandbox {
@@ -100,6 +117,23 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         sandbox_mode_cli_arg.map(Into::<SandboxMode>::into)
     };
 
+    // When using `--oss`, let the bootstrapper pick the model (defaulting to
+    // gpt-oss:20b) and ensure it is present locally. Also, force the built‑in
+    // `oss` model provider.
+    let model = if let Some(model) = model_cli_arg {
+        Some(model)
+    } else if oss {
+        Some(DEFAULT_OSS_MODEL.to_owned())
+    } else {
+        None // No model specified, will use the default.
+    };
+
+    let model_provider = if oss {
+        Some(BUILT_IN_OSS_MODEL_PROVIDER_ID.to_string())
+    } else {
+        None // No specific model provider override.
+    };
+
     // Load configuration and determine approval policy
     let overrides = ConfigOverrides {
         model,
@@ -109,9 +143,12 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         approval_policy: Some(AskForApproval::Never),
         sandbox_mode,
         cwd: cwd.map(|p| p.canonicalize().unwrap_or(p)),
-        model_provider: None,
+        model_provider,
         codex_linux_sandbox_exe,
         base_instructions: None,
+        include_plan_tool: None,
+        disable_response_storage: oss.then_some(true),
+        show_raw_agent_reasoning: oss.then_some(true),
     };
     // Parse `-c` overrides.
     let cli_kv_overrides = match config_overrides.parse_overrides() {
@@ -133,6 +170,12 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         ))
     };
 
+    if oss {
+        codex_ollama::ensure_oss_ready(&config)
+            .await
+            .map_err(|e| anyhow::anyhow!("OSS setup failed: {e}"))?;
+    }
+
     // Print the effective configuration and prompt so users can see what Codex
     // is using.
     event_processor.print_config_summary(&config, &prompt);
@@ -142,20 +185,6 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         std::process::exit(1);
     }
 
-    // TODO(mbolin): Take a more thoughtful approach to logging.
-    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)),
-        )
-        .with_ansi(stderr_with_ansi)
-        .with_writer(std::io::stderr)
-        .try_init();
-
     let CodexConversation {
         codex: codex_wrapper,
         session_configured,
@@ -187,10 +216,16 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
                     res = codex.next_event() => match res {
                         Ok(event) => {
                             debug!("Received event: {event:?}");
+
+                            let is_shutdown_complete = matches!(event.msg, EventMsg::ShutdownComplete);
                             if let Err(e) = tx.send(event) {
                                 error!("Error sending event: {e:?}");
                                 break;
                             }
+                            if is_shutdown_complete {
+                                info!("Received shutdown event, exiting event loop.");
+                                break;
+                            }
                         },
                         Err(e) => {
                             error!("Error receiving event: {e:?}");

codex-rs/exec/tests/apply_patch.rs

@@ -1,5 +1,6 @@
 use anyhow::Context;
 use assert_cmd::prelude::*;
+use codex_core::CODEX_APPLY_PATCH_ARG1;
 use std::fs;
 use std::process::Command;
 use tempfile::tempdir;
@@ -16,7 +17,7 @@ fn test_standalone_exec_cli_can_use_apply_patch() -> anyhow::Result<()> {
 
     Command::cargo_bin("codex-exec")
         .context("should find binary for codex-exec")?
-        .arg("--codex-run-as-apply-patch")
+        .arg(CODEX_APPLY_PATCH_ARG1)
         .arg(
             r#"*** Begin Patch
 *** Update File: source.txt

codex-rs/execpolicy/Cargo.toml

@@ -26,7 +26,7 @@ multimap = "0.10.0"
 path-absolutize = "3.1.1"
 regex-lite = "0.1"
 serde = { version = "1.0.194", features = ["derive"] }
-serde_json = "1.0.110"
+serde_json = "1.0.142"
 serde_with = { version = "3", features = ["macros"] }
 
 [dev-dependencies]

codex-rs/file-search/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-file-search"
 version = { workspace = true }
-edition = "2024"
 
 [[bin]]
 name = "codex-file-search"
@@ -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 = "1.0.110"
+serde_json = "1.0.142"
 tokio = { version = "1", features = ["full"] }

codex-rs/linux-sandbox/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-linux-sandbox"
 version = { workspace = true }
-edition = "2024"
 
 [[bin]]
 name = "codex-linux-sandbox"
@@ -19,8 +19,8 @@ anyhow = "1"
 clap = { version = "4", features = ["derive"] }
 codex-common = { path = "../common", features = ["cli"] }
 codex-core = { path = "../core" }
-libc = "0.2.172"
 landlock = "0.4.1"
+libc = "0.2.172"
 seccompiler = "0.5.0"
 
 [target.'cfg(target_os = "linux")'.dev-dependencies]

codex-rs/linux-sandbox/src/landlock.rs

@@ -36,7 +36,11 @@ pub(crate) fn apply_sandbox_policy_to_current_thread(
     }
 
     if !sandbox_policy.has_full_disk_write_access() {
-        let writable_roots = sandbox_policy.get_writable_roots_with_cwd(cwd);
+        let writable_roots = sandbox_policy
+            .get_writable_roots_with_cwd(cwd)
+            .into_iter()
+            .map(|writable_root| writable_root.root)
+            .collect();
         install_filesystem_landlock_rules_on_current_thread(writable_roots)?;
     }
 

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

@@ -44,11 +44,14 @@ async fn run_cmd(cmd: &[&str], writable_roots: &[PathBuf], timeout_ms: u64) {
         cwd: std::env::current_dir().expect("cwd should exist"),
         timeout_ms: Some(timeout_ms),
         env: create_env_from_core_vars(),
+        with_escalated_permissions: None,
+        justification: None,
     };
 
     let sandbox_policy = SandboxPolicy::WorkspaceWrite {
         writable_roots: writable_roots.to_vec(),
         network_access: false,
+        include_default_writable_roots: true,
     };
     let sandbox_program = env!("CARGO_BIN_EXE_codex-linux-sandbox");
     let codex_linux_sandbox_exe = Some(PathBuf::from(sandbox_program));
@@ -59,6 +62,7 @@ async fn run_cmd(cmd: &[&str], writable_roots: &[PathBuf], timeout_ms: u64) {
         ctrl_c,
         &sandbox_policy,
         &codex_linux_sandbox_exe,
+        None,
     )
     .await
     .unwrap();
@@ -137,6 +141,8 @@ async fn assert_network_blocked(cmd: &[&str]) {
         // do not stall the suite.
         timeout_ms: Some(NETWORK_TIMEOUT_MS),
         env: create_env_from_core_vars(),
+        with_escalated_permissions: None,
+        justification: None,
     };
 
     let sandbox_policy = SandboxPolicy::new_read_only_policy();
@@ -149,6 +155,7 @@ async fn assert_network_blocked(cmd: &[&str]) {
         ctrl_c,
         &sandbox_policy,
         &codex_linux_sandbox_exe,
+        None,
     )
     .await;
 

codex-rs/login/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-login"
 version = { workspace = true }
-edition = "2024"
 
 [lints]
 workspace = true
@@ -18,3 +18,6 @@ tokio = { version = "1", features = [
     "rt-multi-thread",
     "signal",
 ] }
+
+[dev-dependencies]
+tempfile = "3"

codex-rs/login/src/lib.rs

@@ -1,19 +1,187 @@
 use chrono::DateTime;
+
 use chrono::Utc;
 use serde::Deserialize;
 use serde::Serialize;
+use std::env;
 use std::fs::OpenOptions;
 use std::io::Read;
 use std::io::Write;
 #[cfg(unix)]
 use std::os::unix::fs::OpenOptionsExt;
 use std::path::Path;
+use std::path::PathBuf;
 use std::process::Stdio;
+use std::sync::Arc;
+use std::sync::Mutex;
+use std::time::Duration;
 use tokio::process::Command;
 
 const SOURCE_FOR_PYTHON_SERVER: &str = include_str!("./login_with_chatgpt.py");
 
 const CLIENT_ID: &str = "app_EMoamEEZ73f0CkXaXp7hrann";
+pub const OPENAI_API_KEY_ENV_VAR: &str = "OPENAI_API_KEY";
+
+#[derive(Clone, Debug, PartialEq, Copy)]
+pub enum AuthMode {
+    ApiKey,
+    ChatGPT,
+}
+
+#[derive(Debug, Clone)]
+pub struct CodexAuth {
+    pub api_key: Option<String>,
+    pub mode: AuthMode,
+    auth_dot_json: Arc<Mutex<Option<AuthDotJson>>>,
+    auth_file: PathBuf,
+}
+
+impl PartialEq for CodexAuth {
+    fn eq(&self, other: &Self) -> bool {
+        self.mode == other.mode
+    }
+}
+
+impl CodexAuth {
+    pub fn new(
+        api_key: Option<String>,
+        mode: AuthMode,
+        auth_file: PathBuf,
+        auth_dot_json: Option<AuthDotJson>,
+    ) -> Self {
+        let auth_dot_json = Arc::new(Mutex::new(auth_dot_json));
+        Self {
+            api_key,
+            mode,
+            auth_file,
+            auth_dot_json,
+        }
+    }
+
+    pub fn from_api_key(api_key: String) -> Self {
+        Self {
+            api_key: Some(api_key),
+            mode: AuthMode::ApiKey,
+            auth_file: PathBuf::new(),
+            auth_dot_json: Arc::new(Mutex::new(None)),
+        }
+    }
+
+    pub async fn get_token_data(&self) -> Result<TokenData, std::io::Error> {
+        #[expect(clippy::unwrap_used)]
+        let auth_dot_json = self.auth_dot_json.lock().unwrap().clone();
+        match auth_dot_json {
+            Some(AuthDotJson {
+                tokens: Some(mut tokens),
+                last_refresh: Some(last_refresh),
+                ..
+            }) => {
+                if last_refresh < Utc::now() - chrono::Duration::days(28) {
+                    let refresh_response = tokio::time::timeout(
+                        Duration::from_secs(60),
+                        try_refresh_token(tokens.refresh_token.clone()),
+                    )
+                    .await
+                    .map_err(|_| {
+                        std::io::Error::other("timed out while refreshing OpenAI API key")
+                    })?
+                    .map_err(std::io::Error::other)?;
+
+                    let updated_auth_dot_json = update_tokens(
+                        &self.auth_file,
+                        refresh_response.id_token,
+                        refresh_response.access_token,
+                        refresh_response.refresh_token,
+                    )
+                    .await?;
+
+                    tokens = updated_auth_dot_json
+                        .tokens
+                        .clone()
+                        .ok_or(std::io::Error::other(
+                            "Token data is not available after refresh.",
+                        ))?;
+
+                    #[expect(clippy::unwrap_used)]
+                    let mut auth_lock = self.auth_dot_json.lock().unwrap();
+                    *auth_lock = Some(updated_auth_dot_json);
+                }
+
+                Ok(tokens)
+            }
+            _ => Err(std::io::Error::other("Token data is not available.")),
+        }
+    }
+
+    pub async fn get_token(&self) -> Result<String, std::io::Error> {
+        match self.mode {
+            AuthMode::ApiKey => Ok(self.api_key.clone().unwrap_or_default()),
+            AuthMode::ChatGPT => {
+                let id_token = self.get_token_data().await?.access_token;
+
+                Ok(id_token)
+            }
+        }
+    }
+
+    pub async fn get_account_id(&self) -> Option<String> {
+        match self.mode {
+            AuthMode::ApiKey => None,
+            AuthMode::ChatGPT => {
+                let token_data = self.get_token_data().await.ok()?;
+
+                token_data.account_id.clone()
+            }
+        }
+    }
+}
+
+// Loads the available auth information from the auth.json or OPENAI_API_KEY environment variable.
+pub fn load_auth(codex_home: &Path, include_env_var: bool) -> std::io::Result<Option<CodexAuth>> {
+    let auth_file = get_auth_file(codex_home);
+
+    let auth_dot_json = try_read_auth_json(&auth_file).ok();
+
+    let auth_json_api_key = auth_dot_json
+        .as_ref()
+        .and_then(|a| a.openai_api_key.clone())
+        .filter(|s| !s.is_empty());
+
+    let openai_api_key = if include_env_var {
+        env::var(OPENAI_API_KEY_ENV_VAR)
+            .ok()
+            .filter(|s| !s.is_empty())
+            .or(auth_json_api_key)
+    } else {
+        auth_json_api_key
+    };
+
+    let has_tokens = auth_dot_json
+        .as_ref()
+        .and_then(|a| a.tokens.as_ref())
+        .is_some();
+
+    if openai_api_key.is_none() && !has_tokens {
+        return Ok(None);
+    }
+
+    let mode = if openai_api_key.is_some() {
+        AuthMode::ApiKey
+    } else {
+        AuthMode::ChatGPT
+    };
+
+    Ok(Some(CodexAuth {
+        api_key: openai_api_key,
+        mode,
+        auth_file,
+        auth_dot_json: Arc::new(Mutex::new(auth_dot_json)),
+    }))
+}
+
+fn get_auth_file(codex_home: &Path) -> PathBuf {
+    codex_home.join("auth.json")
+}
 
 /// Run `python3 -c {{SOURCE_FOR_PYTHON_SERVER}}` with the CODEX_HOME
 /// environment variable set to the provided `codex_home` path. If the
@@ -24,14 +192,12 @@ const CLIENT_ID: &str = "app_EMoamEEZ73f0CkXaXp7hrann";
 /// If `capture_output` is true, the subprocess's output will be captured and
 /// recorded in memory. Otherwise, the subprocess's output will be sent to the
 /// current process's stdout/stderr.
-pub async fn login_with_chatgpt(
-    codex_home: &Path,
-    capture_output: bool,
-) -> std::io::Result<String> {
+pub async fn login_with_chatgpt(codex_home: &Path, capture_output: bool) -> std::io::Result<()> {
     let child = Command::new("python3")
         .arg("-c")
         .arg(SOURCE_FOR_PYTHON_SERVER)
         .env("CODEX_HOME", codex_home)
+        .env("CODEX_CLIENT_ID", CLIENT_ID)
         .stdin(Stdio::null())
         .stdout(if capture_output {
             Stdio::piped()
@@ -47,7 +213,7 @@ pub async fn login_with_chatgpt(
 
     let output = child.wait_with_output().await?;
     if output.status.success() {
-        try_read_openai_api_key(codex_home).await
+        Ok(())
     } else {
         let stderr = String::from_utf8_lossy(&output.stderr);
         Err(std::io::Error::other(format!(
@@ -56,61 +222,66 @@ pub async fn login_with_chatgpt(
     }
 }
 
-/// Attempt to read the `OPENAI_API_KEY` from the `auth.json` file in the given
-/// `CODEX_HOME` directory, refreshing it, if necessary.
-pub async fn try_read_openai_api_key(codex_home: &Path) -> std::io::Result<String> {
-    let auth_dot_json = try_read_auth_json(codex_home).await?;
-    Ok(auth_dot_json.openai_api_key)
+pub fn login_with_api_key(codex_home: &Path, api_key: &str) -> std::io::Result<()> {
+    let auth_dot_json = AuthDotJson {
+        openai_api_key: Some(api_key.to_string()),
+        tokens: None,
+        last_refresh: None,
+    };
+    write_auth_json(&get_auth_file(codex_home), &auth_dot_json)
 }
 
 /// Attempt to read and refresh the `auth.json` file in the given `CODEX_HOME` directory.
 /// Returns the full AuthDotJson structure after refreshing if necessary.
-pub async fn try_read_auth_json(codex_home: &Path) -> std::io::Result<AuthDotJson> {
-    let auth_path = codex_home.join("auth.json");
-    let mut file = std::fs::File::open(&auth_path)?;
+pub fn try_read_auth_json(auth_file: &Path) -> std::io::Result<AuthDotJson> {
+    let mut file = std::fs::File::open(auth_file)?;
     let mut contents = String::new();
     file.read_to_string(&mut contents)?;
     let auth_dot_json: AuthDotJson = serde_json::from_str(&contents)?;
 
-    if is_expired(&auth_dot_json) {
-        let refresh_response = try_refresh_token(&auth_dot_json).await?;
-        let mut auth_dot_json = auth_dot_json;
-        auth_dot_json.tokens.id_token = refresh_response.id_token;
-        if let Some(refresh_token) = refresh_response.refresh_token {
-            auth_dot_json.tokens.refresh_token = refresh_token;
-        }
-        auth_dot_json.last_refresh = Utc::now();
+    Ok(auth_dot_json)
+}
 
-        let mut options = OpenOptions::new();
-        options.truncate(true).write(true).create(true);
-        #[cfg(unix)]
-        {
-            options.mode(0o600);
-        }
-
-        let json_data = serde_json::to_string(&auth_dot_json)?;
-        {
-            let mut file = options.open(&auth_path)?;
-            file.write_all(json_data.as_bytes())?;
-            file.flush()?;
-        }
-
-        Ok(auth_dot_json)
-    } else {
-        Ok(auth_dot_json)
+fn write_auth_json(auth_file: &Path, auth_dot_json: &AuthDotJson) -> std::io::Result<()> {
+    let json_data = serde_json::to_string_pretty(auth_dot_json)?;
+    let mut options = OpenOptions::new();
+    options.truncate(true).write(true).create(true);
+    #[cfg(unix)]
+    {
+        options.mode(0o600);
     }
+    let mut file = options.open(auth_file)?;
+    file.write_all(json_data.as_bytes())?;
+    file.flush()?;
+    Ok(())
 }
 
-fn is_expired(auth_dot_json: &AuthDotJson) -> bool {
-    let last_refresh = auth_dot_json.last_refresh;
-    last_refresh < Utc::now() - chrono::Duration::days(28)
+async fn update_tokens(
+    auth_file: &Path,
+    id_token: String,
+    access_token: Option<String>,
+    refresh_token: Option<String>,
+) -> std::io::Result<AuthDotJson> {
+    let mut auth_dot_json = try_read_auth_json(auth_file)?;
+
+    let tokens = auth_dot_json.tokens.get_or_insert_with(TokenData::default);
+    tokens.id_token = id_token.to_string();
+    if let Some(access_token) = access_token {
+        tokens.access_token = access_token.to_string();
+    }
+    if let Some(refresh_token) = refresh_token {
+        tokens.refresh_token = refresh_token.to_string();
+    }
+    auth_dot_json.last_refresh = Some(Utc::now());
+    write_auth_json(auth_file, &auth_dot_json)?;
+    Ok(auth_dot_json)
 }
 
-async fn try_refresh_token(auth_dot_json: &AuthDotJson) -> std::io::Result<RefreshResponse> {
+async fn try_refresh_token(refresh_token: String) -> std::io::Result<RefreshResponse> {
     let refresh_request = RefreshRequest {
         client_id: CLIENT_ID,
         grant_type: "refresh_token",
-        refresh_token: auth_dot_json.tokens.refresh_token.clone(),
+        refresh_token,
         scope: "openid profile email",
     };
 
@@ -145,24 +316,27 @@ struct RefreshRequest {
     scope: &'static str,
 }
 
-#[derive(Deserialize)]
+#[derive(Deserialize, Clone)]
 struct RefreshResponse {
     id_token: String,
+    access_token: Option<String>,
     refresh_token: Option<String>,
 }
 
 /// Expected structure for $CODEX_HOME/auth.json.
-#[derive(Deserialize, Serialize)]
+#[derive(Deserialize, Serialize, Clone, Debug, PartialEq)]
 pub struct AuthDotJson {
     #[serde(rename = "OPENAI_API_KEY")]
-    pub openai_api_key: String,
+    pub openai_api_key: Option<String>,
 
-    pub tokens: TokenData,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tokens: Option<TokenData>,
 
-    pub last_refresh: DateTime<Utc>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub last_refresh: Option<DateTime<Utc>>,
 }
 
-#[derive(Deserialize, Serialize, Clone)]
+#[derive(Deserialize, Serialize, Clone, Debug, PartialEq, Default)]
 pub struct TokenData {
     /// This is a JWT.
     pub id_token: String,
@@ -172,5 +346,97 @@ pub struct TokenData {
 
     pub refresh_token: String,
 
-    pub account_id: String,
+    pub account_id: Option<String>,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use tempfile::tempdir;
+
+    #[test]
+    #[expect(clippy::unwrap_used)]
+    fn writes_api_key_and_loads_auth() {
+        let dir = tempdir().unwrap();
+        login_with_api_key(dir.path(), "sk-test-key").unwrap();
+        let auth = load_auth(dir.path(), false).unwrap().unwrap();
+        assert_eq!(auth.mode, AuthMode::ApiKey);
+        assert_eq!(auth.api_key.as_deref(), Some("sk-test-key"));
+    }
+
+    #[test]
+    #[expect(clippy::unwrap_used)]
+    fn loads_from_env_var_if_env_var_exists() {
+        let dir = tempdir().unwrap();
+
+        let env_var = std::env::var(OPENAI_API_KEY_ENV_VAR);
+
+        if let Ok(env_var) = env_var {
+            let auth = load_auth(dir.path(), true).unwrap().unwrap();
+            assert_eq!(auth.mode, AuthMode::ApiKey);
+            assert_eq!(auth.api_key, Some(env_var));
+        }
+    }
+
+    #[tokio::test]
+    #[expect(clippy::unwrap_used)]
+    async fn loads_token_data_from_auth_json() {
+        let dir = tempdir().unwrap();
+        let auth_file = dir.path().join("auth.json");
+        std::fs::write(
+            auth_file,
+            format!(
+                r#"
+        {{
+            "OPENAI_API_KEY": null,
+            "tokens": {{
+                "id_token": "test-id-token",
+                "access_token": "test-access-token",
+                "refresh_token": "test-refresh-token"
+            }},
+            "last_refresh": "{}"
+        }}
+        "#,
+                Utc::now().to_rfc3339()
+            ),
+        )
+        .unwrap();
+
+        let auth = load_auth(dir.path(), false).unwrap().unwrap();
+        assert_eq!(auth.mode, AuthMode::ChatGPT);
+        assert_eq!(auth.api_key, None);
+        assert_eq!(
+            auth.get_token_data().await.unwrap(),
+            TokenData {
+                id_token: "test-id-token".to_string(),
+                access_token: "test-access-token".to_string(),
+                refresh_token: "test-refresh-token".to_string(),
+                account_id: None,
+            }
+        );
+    }
+
+    #[tokio::test]
+    #[expect(clippy::unwrap_used)]
+    async fn loads_api_key_from_auth_json() {
+        let dir = tempdir().unwrap();
+        let auth_file = dir.path().join("auth.json");
+        std::fs::write(
+            auth_file,
+            r#"
+        {
+            "OPENAI_API_KEY": "sk-test-key",
+            "tokens": null,
+            "last_refresh": null
+        }
+        "#,
+        )
+        .unwrap();
+
+        let auth = load_auth(dir.path(), false).unwrap().unwrap();
+        assert_eq!(auth.mode, AuthMode::ApiKey);
+        assert_eq!(auth.api_key, Some("sk-test-key".to_string()));
+
+        assert!(auth.get_token_data().await.is_err());
+    }
 }

codex-rs/login/src/login_with_chatgpt.py

@@ -41,7 +41,6 @@ from typing import Any, Dict  # for type hints
 REQUIRED_PORT = 1455
 URL_BASE = f"http://localhost:{REQUIRED_PORT}"
 DEFAULT_ISSUER = "https://auth.openai.com"
-DEFAULT_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
 
 EXIT_CODE_WHEN_ADDRESS_ALREADY_IN_USE = 13
 
@@ -58,7 +57,7 @@ class TokenData:
 class AuthBundle:
     """Aggregates authentication data produced after successful OAuth flow."""
 
-    api_key: str
+    api_key: str | None
     token_data: TokenData
     last_refresh: str
 
@@ -78,12 +77,18 @@ def main() -> None:
         eprint("ERROR: CODEX_HOME environment variable is not set")
         sys.exit(1)
 
+    client_id = os.getenv("CODEX_CLIENT_ID")
+    if not client_id:
+        eprint("ERROR: CODEX_CLIENT_ID environment variable is not set")
+        sys.exit(1)
+
     # Spawn server.
     try:
         httpd = _ApiKeyHTTPServer(
             ("127.0.0.1", REQUIRED_PORT),
             _ApiKeyHTTPHandler,
             codex_home=codex_home,
+            client_id=client_id,
             verbose=args.verbose,
         )
     except OSError as e:
@@ -157,7 +162,7 @@ class _ApiKeyHTTPHandler(http.server.BaseHTTPRequestHandler):
                 return
 
             try:
-                auth_bundle, success_url = self._exchange_code_for_api_key(code)
+                auth_bundle, success_url = self._exchange_code(code)
             except Exception as exc:  # noqa: BLE001 – propagate to client
                 self.send_error(500, f"Token exchange failed: {exc}")
                 return
@@ -211,68 +216,22 @@ class _ApiKeyHTTPHandler(http.server.BaseHTTPRequestHandler):
         if getattr(self.server, "verbose", False):  # type: ignore[attr-defined]
             super().log_message(fmt, *args)
 
-    def _exchange_code_for_api_key(self, code: str) -> tuple[AuthBundle, str]:
-        """Perform token + token-exchange to obtain an OpenAI API key.
+    def _obtain_api_key(
+        self,
+        token_claims: Dict[str, Any],
+        access_claims: Dict[str, Any],
+        token_data: TokenData,
+    ) -> tuple[str | None, str | None]:
+        """Obtain an API key from the auth service.
 
-        Returns (AuthBundle, success_url).
+        Returns (api_key, success_url) if successful, None otherwise.
         """
 
-        token_endpoint = f"{self.server.issuer}/oauth/token"
-
-        # 1. Authorization-code -> (id_token, access_token, refresh_token)
-        data = urllib.parse.urlencode(
-            {
-                "grant_type": "authorization_code",
-                "code": code,
-                "redirect_uri": self.server.redirect_uri,
-                "client_id": self.server.client_id,
-                "code_verifier": self.server.pkce.code_verifier,
-            }
-        ).encode()
-
-        token_data: TokenData
-
-        with urllib.request.urlopen(
-            urllib.request.Request(
-                token_endpoint,
-                data=data,
-                method="POST",
-                headers={"Content-Type": "application/x-www-form-urlencoded"},
-            )
-        ) as resp:
-            payload = json.loads(resp.read().decode())
-            
-            # Extract chatgpt_account_id from id_token
-            id_token_parts = payload["id_token"].split(".")
-            if len(id_token_parts) != 3:
-                raise ValueError("Invalid ID token")
-            id_token_claims = _decode_jwt_segment(id_token_parts[1])
-            auth_claims = id_token_claims.get("https://api.openai.com/auth", {})
-            chatgpt_account_id = auth_claims.get("chatgpt_account_id", "")
-            
-            token_data = TokenData(
-                id_token=payload["id_token"],
-                access_token=payload["access_token"],
-                refresh_token=payload["refresh_token"],
-                account_id=chatgpt_account_id,
-            )
-
-        access_token_parts = token_data.access_token.split(".")
-        if len(access_token_parts) != 3:
-            raise ValueError("Invalid access token")
-
-        access_token_claims = _decode_jwt_segment(access_token_parts[1])
-
-        token_claims = id_token_claims.get("https://api.openai.com/auth", {})
-        access_claims = access_token_claims.get("https://api.openai.com/auth", {})
-
         org_id = token_claims.get("organization_id")
-        if not org_id:
-            raise ValueError("Missing organization in id_token claims")
-
         project_id = token_claims.get("project_id")
-        if not project_id:
-            raise ValueError("Missing project in id_token claims")
+
+        if not org_id or not project_id:
+            return (None, None)
 
         random_id = secrets.token_hex(6)
 
@@ -292,7 +251,7 @@ class _ApiKeyHTTPHandler(http.server.BaseHTTPRequestHandler):
         exchanged_access_token: str
         with urllib.request.urlopen(
             urllib.request.Request(
-                token_endpoint,
+                self.server.token_endpoint,
                 data=exchange_data,
                 method="POST",
                 headers={"Content-Type": "application/x-www-form-urlencoded"},
@@ -340,6 +299,65 @@ class _ApiKeyHTTPHandler(http.server.BaseHTTPRequestHandler):
         except Exception as exc:  # pragma: no cover – best-effort only
             eprint(f"Unable to redeem ChatGPT subscriber API credits: {exc}")
 
+        return (exchanged_access_token, success_url)
+
+    def _exchange_code(self, code: str) -> tuple[AuthBundle, str]:
+        """Perform token + token-exchange to obtain an OpenAI API key.
+
+        Returns (AuthBundle, success_url).
+        """
+
+        # 1. Authorization-code -> (id_token, access_token, refresh_token)
+        data = urllib.parse.urlencode(
+            {
+                "grant_type": "authorization_code",
+                "code": code,
+                "redirect_uri": self.server.redirect_uri,
+                "client_id": self.server.client_id,
+                "code_verifier": self.server.pkce.code_verifier,
+            }
+        ).encode()
+
+        token_data: TokenData
+
+        with urllib.request.urlopen(
+            urllib.request.Request(
+                self.server.token_endpoint,
+                data=data,
+                method="POST",
+                headers={"Content-Type": "application/x-www-form-urlencoded"},
+            )
+        ) as resp:
+            payload = json.loads(resp.read().decode())
+
+            # Extract chatgpt_account_id from id_token
+            id_token_parts = payload["id_token"].split(".")
+            if len(id_token_parts) != 3:
+                raise ValueError("Invalid ID token")
+            id_token_claims = _decode_jwt_segment(id_token_parts[1])
+            auth_claims = id_token_claims.get("https://api.openai.com/auth", {})
+            chatgpt_account_id = auth_claims.get("chatgpt_account_id", "")
+
+            token_data = TokenData(
+                id_token=payload["id_token"],
+                access_token=payload["access_token"],
+                refresh_token=payload["refresh_token"],
+                account_id=chatgpt_account_id,
+            )
+
+        access_token_parts = token_data.access_token.split(".")
+        if len(access_token_parts) != 3:
+            raise ValueError("Invalid access token")
+
+        access_token_claims = _decode_jwt_segment(access_token_parts[1])
+
+        token_claims = id_token_claims.get("https://api.openai.com/auth", {})
+        access_claims = access_token_claims.get("https://api.openai.com/auth", {})
+
+        exchanged_access_token, success_url = self._obtain_api_key(
+            token_claims, access_claims, token_data
+        )
+
         # Persist refresh_token/id_token for future use (redeem credits etc.)
         last_refresh_str = (
             datetime.datetime.now(datetime.timezone.utc)
@@ -353,7 +371,7 @@ class _ApiKeyHTTPHandler(http.server.BaseHTTPRequestHandler):
             last_refresh=last_refresh_str,
         )
 
-        return (auth_bundle, success_url)
+        return (auth_bundle, success_url or f"{URL_BASE}/success")
 
     def request_shutdown(self) -> None:
         # shutdown() must be invoked from another thread to avoid
@@ -413,6 +431,7 @@ class _ApiKeyHTTPServer(http.server.HTTPServer):
         request_handler_class: type[http.server.BaseHTTPRequestHandler],
         *,
         codex_home: str,
+        client_id: str,
         verbose: bool = False,
     ) -> None:
         super().__init__(server_address, request_handler_class, bind_and_activate=True)
@@ -422,7 +441,8 @@ class _ApiKeyHTTPServer(http.server.HTTPServer):
         self.verbose: bool = verbose
 
         self.issuer: str = DEFAULT_ISSUER
-        self.client_id: str = DEFAULT_CLIENT_ID
+        self.token_endpoint: str = f"{self.issuer}/oauth/token"
+        self.client_id: str = client_id
         port = server_address[1]
         self.redirect_uri: str = f"http://localhost:{port}/auth/callback"
         self.pkce: PkceCodes = _generate_pkce()
@@ -438,6 +458,7 @@ class _ApiKeyHTTPServer(http.server.HTTPServer):
             "code_challenge": self.pkce.code_challenge,
             "code_challenge_method": "S256",
             "id_token_add_organizations": "true",
+            "codex_cli_simplified_flow": "true",
             "state": self.state,
         }
         return f"{self.issuer}/oauth/authorize?" + urllib.parse.urlencode(params)
@@ -581,8 +602,8 @@ def maybe_redeem_credits(
         granted = redeem_data.get("granted_chatgpt_subscriber_api_credits", 0)
         if granted and granted > 0:
             eprint(
-                f"""Thanks for being a ChatGPT {'Plus' if plan_type=='plus' else 'Pro'} subscriber!
-If you haven't already redeemed, you should receive {'$5' if plan_type=='plus' else '$50'} in API credits.
+                f"""Thanks for being a ChatGPT {"Plus" if plan_type == "plus" else "Pro"} subscriber!
+If you haven't already redeemed, you should receive {"$5" if plan_type == "plus" else "$50"} in API credits.
 
 Credits: https://platform.openai.com/settings/organization/billing/credit-grants
 More info: https://help.openai.com/en/articles/11381614""",
@@ -666,6 +687,7 @@ LOGIN_SUCCESS_HTML = """<!DOCTYPE html>
         justify-content: center;
         position: relative;
         background: white;
+
         font-family: system-ui, -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, Cantarell, 'Open Sans', 'Helvetica Neue', sans-serif;
       }
       .inner-container {
@@ -683,6 +705,7 @@ LOGIN_SUCCESS_HTML = """<!DOCTYPE html>
         align-items: center;
         gap: 20px;
         display: flex;
+        margin-top: 15vh;
       }
       .svg-wrapper {
         position: relative;
@@ -690,9 +713,9 @@ LOGIN_SUCCESS_HTML = """<!DOCTYPE html>
       .title {
         text-align: center;
         color: var(--text-primary, #0D0D0D);
-        font-size: 28px;
+        font-size: 32px;
         font-weight: 400;
-        line-height: 36.40px;
+        line-height: 40px;
         word-wrap: break-word;
       }
       .setup-box {
@@ -765,16 +788,26 @@ LOGIN_SUCCESS_HTML = """<!DOCTYPE html>
         word-wrap: break-word;
         text-decoration: none;
       }
+      .logo {
+        display: flex;
+        align-items: center;
+        justify-content: center;
+        width: 4rem;
+        height: 4rem;
+        border-radius: 16px;
+        border: .5px solid rgba(0, 0, 0, 0.1);
+        box-shadow: rgba(0, 0, 0, 0.1) 0px 4px 16px 0px;
+        box-sizing: border-box;
+        background-color: rgb(255, 255, 255);
+      }
     </style>
   </head>
   <body>
     <div class="container">
       <div class="inner-container">
         <div class="content">
-          <div data-svg-wrapper class="svg-wrapper">
-            <svg width="56" height="56" viewBox="0 0 56 56" fill="none" xmlns="http://www.w3.org/2000/svg">
-              <path d="M4.6665 28.0003C4.6665 15.1137 15.1132 4.66699 27.9998 4.66699C40.8865 4.66699 51.3332 15.1137 51.3332 28.0003C51.3332 40.887 40.8865 51.3337 27.9998 51.3337C15.1132 51.3337 4.6665 40.887 4.6665 28.0003ZM37.5093 18.5088C36.4554 17.7672 34.9999 18.0203 34.2583 19.0742L24.8508 32.4427L20.9764 28.1808C20.1095 27.2272 18.6338 27.1569 17.6803 28.0238C16.7267 28.8906 16.6565 30.3664 17.5233 31.3199L23.3566 37.7366C23.833 38.2606 24.5216 38.5399 25.2284 38.4958C25.9353 38.4517 26.5838 38.089 26.9914 37.5098L38.0747 21.7598C38.8163 20.7059 38.5632 19.2504 37.5093 18.5088Z" fill="var(--green-400, #04B84C)"/>
-            </svg>
+          <div class="logo">
+            <svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" fill="none" viewBox="0 0 32 32"><path stroke="#000" stroke-linecap="round" stroke-width="2.484" d="M22.356 19.797H17.17M9.662 12.29l1.979 3.576a.511.511 0 0 1-.005.504l-1.974 3.409M30.758 16c0 8.15-6.607 14.758-14.758 14.758-8.15 0-14.758-6.607-14.758-14.758C1.242 7.85 7.85 1.242 16 1.242c8.15 0 14.758 6.608 14.758 14.758Z"></path></svg>
           </div>
           <div class="title">Signed in to Codex CLI</div>
         </div>

codex-rs/mcp-server/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
+edition = "2024"
 name = "codex-mcp-server"
 version = { workspace = true }
-edition = "2024"
 
 [[bin]]
 name = "codex-mcp-server"
@@ -23,9 +23,7 @@ schemars = "0.8.22"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 shlex = "1.3.0"
-toml = "0.9"
-tracing = { version = "0.1.41", features = ["log"] }
-tracing-subscriber = { version = "0.3", features = ["fmt", "env-filter"] }
+strum_macros = "0.27.2"
 tokio = { version = "1", features = [
     "io-std",
     "macros",
@@ -33,6 +31,9 @@ tokio = { version = "1", features = [
     "rt-multi-thread",
     "signal",
 ] }
+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"] }
 
 [dev-dependencies]

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

@@ -50,6 +50,10 @@ pub struct CodexToolCallParam {
     /// The set of instructions to use instead of the default ones.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub base_instructions: Option<String>,
+
+    /// Whether to include the plan tool in the conversation.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub include_plan_tool: Option<bool>,
 }
 
 /// Custom enum mirroring [`AskForApproval`], but has an extra dependency on
@@ -140,9 +144,10 @@ impl CodexToolCallParam {
             sandbox,
             config: cli_overrides,
             base_instructions,
+            include_plan_tool,
         } = self;
 
-        // Build the `ConfigOverrides` recognised by codex-core.
+        // Build the `ConfigOverrides` recognized by codex-core.
         let overrides = codex_core::config::ConfigOverrides {
             model,
             config_profile: profile,
@@ -152,6 +157,9 @@ impl CodexToolCallParam {
             model_provider: None,
             codex_linux_sandbox_exe,
             base_instructions,
+            include_plan_tool,
+            disable_response_storage: None,
+            show_raw_agent_reasoning: None,
         };
 
         let cli_overrides = cli_overrides
@@ -262,6 +270,10 @@ mod tests {
                 "description": "Working directory for the session. If relative, it is resolved against the server process's current working directory.",
                 "type": "string"
               },
+              "include-plan-tool": {
+                "description": "Whether to include the plan tool in the conversation.",
+                "type": "boolean"
+              },
               "model": {
                 "description": "Optional override for the model name (e.g. \"o3\", \"o4-mini\").",
                 "type": "string"

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

@@ -252,17 +252,22 @@ async fn run_codex_tool_session_inner(
                     EventMsg::AgentMessage(AgentMessageEvent { .. }) => {
                         // TODO: think how we want to support this in the MCP
                     }
-                    EventMsg::TaskStarted
+                    EventMsg::AgentReasoningRawContent(_)
+                    | EventMsg::AgentReasoningRawContentDelta(_)
+                    | EventMsg::TaskStarted
                     | EventMsg::TokenCount(_)
                     | EventMsg::AgentReasoning(_)
                     | EventMsg::McpToolCallBegin(_)
                     | EventMsg::McpToolCallEnd(_)
                     | EventMsg::ExecCommandBegin(_)
+                    | EventMsg::ExecCommandOutputDelta(_)
                     | EventMsg::ExecCommandEnd(_)
                     | EventMsg::BackgroundEvent(_)
                     | EventMsg::PatchApplyBegin(_)
                     | EventMsg::PatchApplyEnd(_)
+                    | EventMsg::TurnDiff(_)
                     | EventMsg::GetHistoryEntryResponse(_)
+                    | EventMsg::PlanUpdate(_)
                     | EventMsg::ShutdownComplete => {
                         // For now, we do not do anything extra for these
                         // events. Note that

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

@@ -0,0 +1,124 @@
+use std::sync::Arc;
+
+use crate::exec_approval::handle_exec_approval_request;
+use crate::outgoing_message::OutgoingMessageSender;
+use crate::outgoing_message::OutgoingNotificationMeta;
+use crate::patch_approval::handle_patch_approval_request;
+use codex_core::Codex;
+use codex_core::protocol::AgentMessageEvent;
+use codex_core::protocol::ApplyPatchApprovalRequestEvent;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::ExecApprovalRequestEvent;
+use mcp_types::RequestId;
+use tracing::error;
+
+pub async fn run_conversation_loop(
+    codex: Arc<Codex>,
+    outgoing: Arc<OutgoingMessageSender>,
+    request_id: RequestId,
+) {
+    let request_id_str = match &request_id {
+        RequestId::String(s) => s.clone(),
+        RequestId::Integer(n) => n.to_string(),
+    };
+
+    // Stream events until the task needs to pause for user interaction or
+    // completes.
+    loop {
+        match codex.next_event().await {
+            Ok(event) => {
+                outgoing
+                    .send_event_as_notification(
+                        &event,
+                        Some(OutgoingNotificationMeta::new(Some(request_id.clone()))),
+                    )
+                    .await;
+
+                match event.msg {
+                    EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
+                        command,
+                        cwd,
+                        call_id,
+                        reason: _,
+                    }) => {
+                        handle_exec_approval_request(
+                            command,
+                            cwd,
+                            outgoing.clone(),
+                            codex.clone(),
+                            request_id.clone(),
+                            request_id_str.clone(),
+                            event.id.clone(),
+                            call_id,
+                        )
+                        .await;
+                        continue;
+                    }
+                    EventMsg::Error(_) => {
+                        error!("Codex runtime error");
+                    }
+                    EventMsg::ApplyPatchApprovalRequest(ApplyPatchApprovalRequestEvent {
+                        call_id,
+                        reason,
+                        grant_root,
+                        changes,
+                    }) => {
+                        handle_patch_approval_request(
+                            call_id,
+                            reason,
+                            grant_root,
+                            changes,
+                            outgoing.clone(),
+                            codex.clone(),
+                            request_id.clone(),
+                            request_id_str.clone(),
+                            event.id.clone(),
+                        )
+                        .await;
+                        continue;
+                    }
+                    EventMsg::TaskComplete(_) => {}
+                    EventMsg::SessionConfigured(_) => {
+                        tracing::error!("unexpected SessionConfigured event");
+                    }
+                    EventMsg::AgentMessageDelta(_) => {
+                        // TODO: think how we want to support this in the MCP
+                    }
+                    EventMsg::AgentReasoningDelta(_) => {
+                        // TODO: think how we want to support this in the MCP
+                    }
+                    EventMsg::AgentMessage(AgentMessageEvent { .. }) => {
+                        // TODO: think how we want to support this in the MCP
+                    }
+                    EventMsg::AgentReasoningRawContent(_)
+                    | EventMsg::AgentReasoningRawContentDelta(_)
+                    | EventMsg::TaskStarted
+                    | EventMsg::TokenCount(_)
+                    | EventMsg::AgentReasoning(_)
+                    | EventMsg::McpToolCallBegin(_)
+                    | EventMsg::McpToolCallEnd(_)
+                    | EventMsg::ExecCommandBegin(_)
+                    | EventMsg::ExecCommandEnd(_)
+                    | EventMsg::TurnDiff(_)
+                    | EventMsg::BackgroundEvent(_)
+                    | EventMsg::ExecCommandOutputDelta(_)
+                    | EventMsg::PatchApplyBegin(_)
+                    | EventMsg::PatchApplyEnd(_)
+                    | EventMsg::GetHistoryEntryResponse(_)
+                    | EventMsg::PlanUpdate(_)
+                    | EventMsg::ShutdownComplete => {
+                        // For now, we do not do anything extra for these
+                        // events. Note that
+                        // send(codex_event_to_notification(&event)) above has
+                        // already dispatched these events as notifications,
+                        // though we may want to do give different treatment to
+                        // individual events in the future.
+                    }
+                }
+            }
+            Err(e) => {
+                error!("Codex runtime error: {e}");
+            }
+        }
+    }
+}

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

@@ -18,7 +18,7 @@ use crate::codex_tool_runner::INVALID_PARAMS_ERROR_CODE;
 
 /// Conforms to [`mcp_types::ElicitRequestParams`] so that it can be used as the
 /// `params` field of an [`ElicitRequest`].
-#[derive(Debug, Serialize)]
+#[derive(Debug, Deserialize, Serialize)]
 pub struct ExecApprovalElicitRequestParams {
     // These fields are required so that `params`
     // conforms to ElicitRequestParams.

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

@@ -17,11 +17,14 @@ use tracing_subscriber::EnvFilter;
 
 mod codex_tool_config;
 mod codex_tool_runner;
+mod conversation_loop;
 mod exec_approval;
 mod json_to_toml;
-mod message_processor;
+pub mod mcp_protocol;
+pub(crate) mod message_processor;
 mod outgoing_message;
 mod patch_approval;
+pub(crate) mod tool_handlers;
 
 use crate::message_processor::MessageProcessor;
 use crate::outgoing_message::OutgoingMessage;

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

@@ -1,4 +1,5 @@
 use std::collections::HashMap;
+use std::collections::HashSet;
 use std::path::PathBuf;
 use std::sync::Arc;
 
@@ -6,11 +7,17 @@ use crate::codex_tool_config::CodexToolCallParam;
 use crate::codex_tool_config::CodexToolCallReplyParam;
 use crate::codex_tool_config::create_tool_for_codex_tool_call_param;
 use crate::codex_tool_config::create_tool_for_codex_tool_call_reply_param;
+use crate::mcp_protocol::ToolCallRequestParams;
+use crate::mcp_protocol::ToolCallResponse;
+use crate::mcp_protocol::ToolCallResponseResult;
 use crate::outgoing_message::OutgoingMessageSender;
+use crate::tool_handlers::create_conversation::handle_create_conversation;
+use crate::tool_handlers::send_message::handle_send_message;
 
 use codex_core::Codex;
 use codex_core::config::Config as CodexConfig;
 use codex_core::protocol::Submission;
+use mcp_types::CallToolRequest;
 use mcp_types::CallToolRequestParams;
 use mcp_types::CallToolResult;
 use mcp_types::ClientRequest;
@@ -37,6 +44,7 @@ pub(crate) struct MessageProcessor {
     codex_linux_sandbox_exe: Option<PathBuf>,
     session_map: Arc<Mutex<HashMap<Uuid, Arc<Codex>>>>,
     running_requests_id_to_codex_uuid: Arc<Mutex<HashMap<RequestId, Uuid>>>,
+    running_session_ids: Arc<Mutex<HashSet<Uuid>>>,
 }
 
 impl MessageProcessor {
@@ -52,9 +60,22 @@ impl MessageProcessor {
             codex_linux_sandbox_exe,
             session_map: Arc::new(Mutex::new(HashMap::new())),
             running_requests_id_to_codex_uuid: Arc::new(Mutex::new(HashMap::new())),
+            running_session_ids: Arc::new(Mutex::new(HashSet::new())),
         }
     }
 
+    pub(crate) fn session_map(&self) -> Arc<Mutex<HashMap<Uuid, Arc<Codex>>>> {
+        self.session_map.clone()
+    }
+
+    pub(crate) fn outgoing(&self) -> Arc<OutgoingMessageSender> {
+        self.outgoing.clone()
+    }
+
+    pub(crate) fn running_session_ids(&self) -> Arc<Mutex<HashSet<Uuid>>> {
+        self.running_session_ids.clone()
+    }
+
     pub(crate) async fn process_request(&mut self, request: JSONRPCRequest) {
         // Hold on to the ID so we can respond.
         let request_id = request.id.clone();
@@ -300,6 +321,14 @@ impl MessageProcessor {
         params: <mcp_types::CallToolRequest as mcp_types::ModelContextProtocolRequest>::Params,
     ) {
         tracing::info!("tools/call -> params: {:?}", params);
+        // Serialize params into JSON and try to parse as new type
+        if let Ok(new_params) =
+            serde_json::to_value(&params).and_then(serde_json::from_value::<ToolCallRequestParams>)
+        {
+            // New tool call matched → forward
+            self.handle_new_tool_calls(id, new_params).await;
+            return;
+        }
         let CallToolRequestParams { name, arguments } = params;
 
         match name.as_str() {
@@ -323,6 +352,29 @@ impl MessageProcessor {
             }
         }
     }
+    async fn handle_new_tool_calls(&self, request_id: RequestId, params: ToolCallRequestParams) {
+        match params {
+            ToolCallRequestParams::ConversationCreate(args) => {
+                handle_create_conversation(self, request_id, args).await;
+            }
+            ToolCallRequestParams::ConversationSendMessage(args) => {
+                handle_send_message(self, request_id, args).await;
+            }
+            _ => {
+                let result = CallToolResult {
+                    content: vec![ContentBlock::TextContent(TextContent {
+                        r#type: "text".to_string(),
+                        text: "Unknown tool".to_string(),
+                        annotations: None,
+                    })],
+                    is_error: Some(true),
+                    structured_content: None,
+                };
+                self.send_response::<CallToolRequest>(request_id, result)
+                    .await;
+            }
+        }
+    }
 
     async fn handle_tool_call_codex(&self, id: RequestId, arguments: Option<serde_json::Value>) {
         let (initial_prompt, config): (String, CodexConfig) = match arguments {
@@ -631,4 +683,20 @@ impl MessageProcessor {
     ) {
         tracing::info!("notifications/message -> params: {:?}", params);
     }
+
+    pub(crate) async fn send_response_with_optional_error(
+        &self,
+        id: RequestId,
+        message: Option<ToolCallResponseResult>,
+        error: Option<bool>,
+    ) {
+        let response = ToolCallResponse {
+            request_id: id.clone(),
+            is_error: error,
+            result: message,
+        };
+        let result: CallToolResult = response.into();
+        self.send_response::<mcp_types::CallToolRequest>(id.clone(), result)
+            .await;
+    }
 }

codex-rs/mcp-server/src/tool_handlers/create_conversation.rs

@@ -0,0 +1,162 @@
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use codex_core::Codex;
+use codex_core::codex_wrapper::init_codex;
+use codex_core::config::Config as CodexConfig;
+use codex_core::config::ConfigOverrides;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::SessionConfiguredEvent;
+use mcp_types::RequestId;
+use tokio::sync::Mutex;
+use uuid::Uuid;
+
+use crate::conversation_loop::run_conversation_loop;
+use crate::json_to_toml::json_to_toml;
+use crate::mcp_protocol::ConversationCreateArgs;
+use crate::mcp_protocol::ConversationCreateResult;
+use crate::mcp_protocol::ConversationId;
+use crate::mcp_protocol::ToolCallResponseResult;
+use crate::message_processor::MessageProcessor;
+
+pub(crate) async fn handle_create_conversation(
+    message_processor: &MessageProcessor,
+    id: RequestId,
+    args: ConversationCreateArgs,
+) {
+    // Build ConfigOverrides from args
+    let ConversationCreateArgs {
+        prompt: _, // not used here; creation only establishes the session
+        model,
+        cwd,
+        approval_policy,
+        sandbox,
+        config,
+        profile,
+        base_instructions,
+    } = args;
+
+    // Convert config overrides JSON into CLI-style TOML overrides
+    let cli_overrides: Vec<(String, toml::Value)> = match config {
+        Some(v) => match v.as_object() {
+            Some(map) => map
+                .into_iter()
+                .map(|(k, v)| (k.clone(), json_to_toml(v.clone())))
+                .collect(),
+            None => Vec::new(),
+        },
+        None => Vec::new(),
+    };
+
+    let overrides = ConfigOverrides {
+        model: Some(model.clone()),
+        cwd: Some(PathBuf::from(cwd)),
+        approval_policy,
+        sandbox_mode: sandbox,
+        model_provider: None,
+        config_profile: profile,
+        codex_linux_sandbox_exe: None,
+        base_instructions,
+        include_plan_tool: None,
+        disable_response_storage: None,
+        show_raw_agent_reasoning: None,
+    };
+
+    let cfg: CodexConfig = match CodexConfig::load_with_cli_overrides(cli_overrides, overrides) {
+        Ok(cfg) => cfg,
+        Err(e) => {
+            message_processor
+                .send_response_with_optional_error(
+                    id,
+                    Some(ToolCallResponseResult::ConversationCreate(
+                        ConversationCreateResult::Error {
+                            message: format!("Failed to load config: {e}"),
+                        },
+                    )),
+                    Some(true),
+                )
+                .await;
+            return;
+        }
+    };
+
+    // Initialize Codex session
+    let codex_conversation = match init_codex(cfg).await {
+        Ok(conv) => conv,
+        Err(e) => {
+            message_processor
+                .send_response_with_optional_error(
+                    id,
+                    Some(ToolCallResponseResult::ConversationCreate(
+                        ConversationCreateResult::Error {
+                            message: format!("Failed to initialize session: {e}"),
+                        },
+                    )),
+                    Some(true),
+                )
+                .await;
+            return;
+        }
+    };
+
+    // Expect SessionConfigured; if not, return error.
+    let EventMsg::SessionConfigured(SessionConfiguredEvent { model, .. }) =
+        &codex_conversation.session_configured.msg
+    else {
+        message_processor
+            .send_response_with_optional_error(
+                id,
+                Some(ToolCallResponseResult::ConversationCreate(
+                    ConversationCreateResult::Error {
+                        message: "Expected SessionConfigured event".to_string(),
+                    },
+                )),
+                Some(true),
+            )
+            .await;
+        return;
+    };
+
+    let effective_model = model.clone();
+
+    let session_id = codex_conversation.session_id;
+    let codex_arc = Arc::new(codex_conversation.codex);
+
+    // Store session for future calls
+    insert_session(
+        session_id,
+        codex_arc.clone(),
+        message_processor.session_map(),
+    )
+    .await;
+    // Run the conversation loop in the background so this request can return immediately.
+    let outgoing = message_processor.outgoing();
+    let spawn_id = id.clone();
+    tokio::spawn(async move {
+        run_conversation_loop(codex_arc.clone(), outgoing, spawn_id).await;
+    });
+
+    // Reply with the new conversation id and effective model
+    message_processor
+        .send_response_with_optional_error(
+            id,
+            Some(ToolCallResponseResult::ConversationCreate(
+                ConversationCreateResult::Ok {
+                    conversation_id: ConversationId(session_id),
+                    model: effective_model,
+                },
+            )),
+            Some(false),
+        )
+        .await;
+}
+
+async fn insert_session(
+    session_id: Uuid,
+    codex: Arc<Codex>,
+    session_map: Arc<Mutex<HashMap<Uuid, Arc<Codex>>>>,
+) {
+    let mut guard = session_map.lock().await;
+    guard.insert(session_id, codex);
+}

codex-rs/mcp-server/src/tool_handlers/mod.rs

@@ -0,0 +1,2 @@
+pub(crate) mod create_conversation;
+pub(crate) mod send_message;