fix(windows) parse powershell commands

This adds the following fields to `ThreadStartResponse` and
`ThreadResumeResponse`:

```rust
    pub model: String,
    pub model_provider: String,
    pub cwd: PathBuf,
    pub approval_policy: AskForApproval,
    pub sandbox: SandboxPolicy,
    pub reasoning_effort: Option<ReasoningEffort>,
```

This is important because these fields are optional in
`ThreadStartParams` and `ThreadResumeParams`, so the caller needs to be
able to determine what values were ultimately used to start/resume the
conversation. (Though note that any of these could be changed later
between turns in the conversation.)

Though to get this information reliably, it must be read from the
internal `SessionConfiguredEvent` that is created in response to the
start of a conversation. Because `SessionConfiguredEvent` (as defined in
`codex-rs/protocol/src/protocol.rs`) did not have all of these fields, a
number of them had to be added as part of this PR.

Because `SessionConfiguredEvent` is referenced in many tests, test
instances of `SessionConfiguredEvent` had to be updated, as well, which
is why this PR touches so many files.

.github/codex/home/config.toml

@@ -1,3 +1,3 @@
-model = "gpt-5"
+model = "gpt-5.1"
 
 # Consider setting [mcp_servers] here!

.github/workflows/cla.yml

@@ -13,6 +13,9 @@ permissions:
 
 jobs:
   cla:
+    # Only run the CLA assistant for the canonical openai repo so forks are not blocked
+    # and contributors who signed previously do not receive duplicate CLA notifications.
+    if: ${{ github.repository_owner == 'openai' }}
     runs-on: ubuntu-latest
     steps:
       - uses: contributor-assistant/github-action@v2.6.1
@@ -43,4 +46,6 @@ jobs:
           path-to-document: https://github.com/openai/codex/blob/main/docs/CLA.md
           path-to-signatures: signatures/cla.json
           branch: cla-signatures
-          allowlist: dependabot[bot]
+          allowlist: |
+            codex
+            dependabot[bot]

.github/workflows/close-stale-contributor-prs.yml

@@ -15,7 +15,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Close inactive PRs from contributors
-        uses: actions/github-script@v7
+        uses: actions/github-script@v8
         with:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           script: |

.github/workflows/issue-deduplicator.yml

@@ -46,7 +46,7 @@ jobs:
         with:
           openai-api-key: ${{ secrets.CODEX_OPENAI_API_KEY }}
           allow-users: "*"
-          model: gpt-5
+          model: gpt-5.1
           prompt: |
             You are an assistant that triages new GitHub issues by identifying potential duplicates.
 

.github/workflows/rust-ci.yml

@@ -95,8 +95,8 @@ jobs:
       run:
         working-directory: codex-rs
     env:
-      # Speed up repeated builds across CI runs by caching compiled objects.
-      RUSTC_WRAPPER: sccache
+      # Speed up repeated builds across CI runs by caching compiled objects (non-Windows).
+      USE_SCCACHE: ${{ startsWith(matrix.runner, 'windows') && 'false' || 'true' }}
       CARGO_INCREMENTAL: "0"
       SCCACHE_CACHE_SIZE: 10G
 
@@ -170,12 +170,14 @@ jobs:
 
       # Install and restore sccache cache
       - name: Install sccache
+        if: ${{ env.USE_SCCACHE == 'true' }}
         uses: taiki-e/install-action@44c6d64aa62cd779e873306675c7a58e86d6d532 # v2
         with:
           tool: sccache
           version: 0.7.5
 
       - name: Configure sccache backend
+        if: ${{ env.USE_SCCACHE == 'true' }}
         shell: bash
         run: |
           set -euo pipefail
@@ -188,8 +190,13 @@ jobs:
             echo "Using sccache local disk + actions/cache fallback"
           fi
 
+      - name: Enable sccache wrapper
+        if: ${{ env.USE_SCCACHE == 'true' }}
+        shell: bash
+        run: echo "RUSTC_WRAPPER=sccache" >> "$GITHUB_ENV"
+
       - name: Restore sccache cache (fallback)
-        if: ${{ env.SCCACHE_GHA_ENABLED != 'true' }}
+        if: ${{ env.USE_SCCACHE == 'true' && env.SCCACHE_GHA_ENABLED != 'true' }}
         id: cache_sccache_restore
         uses: actions/cache/restore@v4
         with:
@@ -274,7 +281,7 @@ jobs:
           key: cargo-home-${{ matrix.runner }}-${{ matrix.target }}-${{ matrix.profile }}-${{ hashFiles('**/Cargo.lock') }}-${{ hashFiles('codex-rs/rust-toolchain.toml') }}
 
       - name: Save sccache cache (fallback)
-        if: always() && !cancelled() && env.SCCACHE_GHA_ENABLED != 'true'
+        if: always() && !cancelled() && env.USE_SCCACHE == 'true' && env.SCCACHE_GHA_ENABLED != 'true'
         continue-on-error: true
         uses: actions/cache/save@v4
         with:
@@ -282,12 +289,12 @@ jobs:
           key: sccache-${{ matrix.runner }}-${{ matrix.target }}-${{ matrix.profile }}-${{ hashFiles('**/Cargo.lock') }}-${{ github.run_id }}
 
       - name: sccache stats
-        if: always()
+        if: always() && env.USE_SCCACHE == 'true'
         continue-on-error: true
         run: sccache --show-stats || true
 
       - name: sccache summary
-        if: always()
+        if: always() && env.USE_SCCACHE == 'true'
         shell: bash
         run: |
           {
@@ -326,7 +333,8 @@ jobs:
       run:
         working-directory: codex-rs
     env:
-      RUSTC_WRAPPER: sccache
+      # Speed up repeated builds across CI runs by caching compiled objects (non-Windows).
+      USE_SCCACHE: ${{ startsWith(matrix.runner, 'windows') && 'false' || 'true' }}
       CARGO_INCREMENTAL: "0"
       SCCACHE_CACHE_SIZE: 10G
 
@@ -370,12 +378,14 @@ jobs:
             cargo-home-${{ matrix.runner }}-${{ matrix.target }}-${{ matrix.profile }}-
 
       - name: Install sccache
+        if: ${{ env.USE_SCCACHE == 'true' }}
         uses: taiki-e/install-action@44c6d64aa62cd779e873306675c7a58e86d6d532 # v2
         with:
           tool: sccache
           version: 0.7.5
 
       - name: Configure sccache backend
+        if: ${{ env.USE_SCCACHE == 'true' }}
         shell: bash
         run: |
           set -euo pipefail
@@ -388,8 +398,13 @@ jobs:
             echo "Using sccache local disk + actions/cache fallback"
           fi
 
+      - name: Enable sccache wrapper
+        if: ${{ env.USE_SCCACHE == 'true' }}
+        shell: bash
+        run: echo "RUSTC_WRAPPER=sccache" >> "$GITHUB_ENV"
+
       - name: Restore sccache cache (fallback)
-        if: ${{ env.SCCACHE_GHA_ENABLED != 'true' }}
+        if: ${{ env.USE_SCCACHE == 'true' && env.SCCACHE_GHA_ENABLED != 'true' }}
         id: cache_sccache_restore
         uses: actions/cache/restore@v4
         with:
@@ -424,7 +439,7 @@ jobs:
           key: cargo-home-${{ matrix.runner }}-${{ matrix.target }}-${{ matrix.profile }}-${{ hashFiles('**/Cargo.lock') }}-${{ hashFiles('codex-rs/rust-toolchain.toml') }}
 
       - name: Save sccache cache (fallback)
-        if: always() && !cancelled() && env.SCCACHE_GHA_ENABLED != 'true'
+        if: always() && !cancelled() && env.USE_SCCACHE == 'true' && env.SCCACHE_GHA_ENABLED != 'true'
         continue-on-error: true
         uses: actions/cache/save@v4
         with:
@@ -432,12 +447,12 @@ jobs:
           key: sccache-${{ matrix.runner }}-${{ matrix.target }}-${{ matrix.profile }}-${{ hashFiles('**/Cargo.lock') }}-${{ github.run_id }}
 
       - name: sccache stats
-        if: always()
+        if: always() && env.USE_SCCACHE == 'true'
         continue-on-error: true
         run: sccache --show-stats || true
 
       - name: sccache summary
-        if: always()
+        if: always() && env.USE_SCCACHE == 'true'
         shell: bash
         run: |
           {

.gitignore

@@ -64,6 +64,9 @@ apply_patch/
 # coverage
 coverage/
 
+# personal files
+personal/
+
 # os
 .DS_Store
 Thumbs.db

codex-rs/.codex/policy.codexpolicy

@@ -1,13 +0,0 @@
-# Execpolicy v2 smoke-test policy
-prefix_rule(
-    pattern = ["echo", ["a", "b"]],
-    decision = "allow",
-    match = ["echo c"],
-    not_match = ["echo-policy-allowed"],
-)
-
-prefix_rule(
-    pattern = ["echo", "c"],
-    decision = "forbidden",
-    match = ["echo c"],
-)

codex-rs/.config/nextest.toml

@@ -0,0 +1,9 @@
+[profile.default]
+# Do not increase, fix your test instead
+slow-timeout = { period = "15s", terminate-after = 2 }
+
+
+[[profile.default.overrides]]
+# Do not add new tests here
+filter = 'test(rmcp_client) | test(humanlike_typing_1000_chars_appears_live_no_placeholder)'
+slow-timeout = { period = "1m", terminate-after = 4 }

codex-rs/Cargo.lock

@@ -850,6 +850,7 @@ dependencies = [
  "codex-protocol",
  "codex-utils-json-to-toml",
  "core_test_support",
+ "mcp-types",
  "opentelemetry-appender-tracing",
  "os_info",
  "pretty_assertions",
@@ -873,7 +874,6 @@ dependencies = [
  "clap",
  "codex-protocol",
  "mcp-types",
- "paste",
  "pretty_assertions",
  "schemars 0.8.22",
  "serde",
@@ -883,6 +883,19 @@ dependencies = [
  "uuid",
 ]
 
+[[package]]
+name = "codex-app-server-test-client"
+version = "0.0.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "codex-app-server-protocol",
+ "codex-protocol",
+ "serde",
+ "serde_json",
+ "uuid",
+]
+
 [[package]]
 name = "codex-apply-patch"
 version = "0.0.0"
@@ -1047,6 +1060,8 @@ dependencies = [
  "clap",
  "codex-app-server-protocol",
  "codex-core",
+ "codex-lmstudio",
+ "codex-ollama",
  "codex-protocol",
  "once_cell",
  "serde",
@@ -1070,7 +1085,6 @@ dependencies = [
  "codex-apply-patch",
  "codex-arg0",
  "codex-async-utils",
- "codex-execpolicy2",
  "codex-file-search",
  "codex-git",
  "codex-keyring-store",
@@ -1146,7 +1160,6 @@ dependencies = [
  "codex-arg0",
  "codex-common",
  "codex-core",
- "codex-ollama",
  "codex-protocol",
  "core_test_support",
  "libc",
@@ -1169,6 +1182,26 @@ dependencies = [
  "wiremock",
 ]
 
+[[package]]
+name = "codex-exec-server"
+version = "0.0.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "codex-core",
+ "libc",
+ "path-absolutize",
+ "pretty_assertions",
+ "rmcp",
+ "serde",
+ "serde_json",
+ "socket2 0.6.0",
+ "tempfile",
+ "tokio",
+ "tracing",
+ "tracing-subscriber",
+]
+
 [[package]]
 name = "codex-execpolicy"
 version = "0.0.0"
@@ -1265,6 +1298,19 @@ dependencies = [
  "tokio",
 ]
 
+[[package]]
+name = "codex-lmstudio"
+version = "0.0.0"
+dependencies = [
+ "codex-core",
+ "reqwest",
+ "serde_json",
+ "tokio",
+ "tracing",
+ "which",
+ "wiremock",
+]
+
 [[package]]
 name = "codex-login"
 version = "0.0.0"
@@ -1340,6 +1386,7 @@ dependencies = [
  "codex-app-server-protocol",
  "codex-protocol",
  "eventsource-stream",
+ "http",
  "opentelemetry",
  "opentelemetry-otlp",
  "opentelemetry-semantic-conventions",
@@ -1373,6 +1420,7 @@ dependencies = [
  "icu_provider",
  "mcp-types",
  "mime_guess",
+ "pretty_assertions",
  "schemars 0.8.22",
  "serde",
  "serde_json",
@@ -1429,6 +1477,7 @@ dependencies = [
  "tracing",
  "urlencoding",
  "webbrowser",
+ "which",
 ]
 
 [[package]]
@@ -1456,12 +1505,12 @@ dependencies = [
  "codex-ansi-escape",
  "codex-app-server-protocol",
  "codex-arg0",
+ "codex-backend-client",
  "codex-common",
  "codex-core",
  "codex-feedback",
  "codex-file-search",
  "codex-login",
- "codex-ollama",
  "codex-protocol",
  "codex-windows-sandbox",
  "color-eyre",
@@ -1484,6 +1533,7 @@ dependencies = [
  "ratatui",
  "ratatui-macros",
  "regex-lite",
+ "reqwest",
  "serde",
  "serde_json",
  "serial_test",
@@ -1566,9 +1616,11 @@ name = "codex-utils-tokenizer"
 version = "0.0.0"
 dependencies = [
  "anyhow",
+ "codex-utils-cache",
  "pretty_assertions",
  "thiserror 2.0.17",
  "tiktoken-rs",
+ "tokio",
 ]
 
 [[package]]
@@ -1576,6 +1628,7 @@ name = "codex-windows-sandbox"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "codex-protocol",
  "dirs-next",
  "dunce",
  "rand 0.8.5",
@@ -5144,6 +5197,7 @@ version = "0.23.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2491382039b29b9b11ff08b76ff6c97cf287671dbb74f0be44bda389fffe9bd1"
 dependencies = [
+ "log",
  "once_cell",
  "ring",
  "rustls-pki-types",
@@ -6318,9 +6372,9 @@ dependencies = [
 
 [[package]]
 name = "tiktoken-rs"
-version = "0.7.0"
+version = "0.9.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "25563eeba904d770acf527e8b370fe9a5547bacd20ff84a0b6c3bc41288e5625"
+checksum = "3a19830747d9034cd9da43a60eaa8e552dfda7712424aebf187b7a60126bae0d"
 dependencies = [
  "anyhow",
  "base64",
@@ -6571,8 +6625,10 @@ dependencies = [
  "percent-encoding",
  "pin-project",
  "prost",
+ "rustls-native-certs",
  "socket2 0.5.10",
  "tokio",
+ "tokio-rustls",
  "tokio-stream",
  "tower",
  "tower-layer",

codex-rs/Cargo.toml

@@ -5,6 +5,7 @@ members = [
     "async-utils",
     "app-server",
     "app-server-protocol",
+    "app-server-test-client",
     "apply-patch",
     "arg0",
     "feedback",
@@ -15,11 +16,13 @@ members = [
     "common",
     "core",
     "exec",
+    "exec-server",
     "execpolicy",
     "execpolicy2",
     "keyring-store",
     "file-search",
     "linux-sandbox",
+    "lmstudio",
     "login",
     "mcp-server",
     "mcp-types",
@@ -64,12 +67,12 @@ codex-chatgpt = { path = "chatgpt" }
 codex-common = { path = "common" }
 codex-core = { path = "core" }
 codex-exec = { path = "exec" }
-codex-execpolicy2 = { path = "execpolicy2" }
 codex-feedback = { path = "feedback" }
 codex-file-search = { path = "file-search" }
 codex-git = { path = "utils/git" }
 codex-keyring-store = { path = "keyring-store" }
 codex-linux-sandbox = { path = "linux-sandbox" }
+codex-lmstudio = { path = "lmstudio" }
 codex-login = { path = "login" }
 codex-mcp-server = { path = "mcp-server" }
 codex-ollama = { path = "ollama" }
@@ -131,7 +134,7 @@ image = { version = "^0.25.8", default-features = false }
 indexmap = "2.12.0"
 insta = "1.43.2"
 itertools = "0.14.0"
-keyring = "3.6"
+keyring = { version = "3.6", default-features = false }
 landlock = "0.4.1"
 lazy_static = "1"
 libc = "0.2.175"
@@ -151,7 +154,6 @@ opentelemetry-semantic-conventions = "0.30.0"
 opentelemetry_sdk = "0.30.0"
 os_info = "3.12.0"
 owo-colors = "4.2.0"
-paste = "1.0.15"
 path-absolutize = "3.1.1"
 pathdiff = "0.2"
 portable-pty = "0.9.0"
@@ -175,6 +177,7 @@ sha1 = "0.10.6"
 sha2 = "0.10"
 shlex = "1.3.0"
 similar = "2.7.0"
+socket2 = "0.6.0"
 starlark = "0.13.0"
 strum = "0.27.2"
 strum_macros = "0.27.2"
@@ -184,6 +187,7 @@ tempfile = "3.23.0"
 test-log = "0.2.18"
 textwrap = "0.16.2"
 thiserror = "2.0.17"
+tiktoken-rs = "0.9"
 time = "0.3"
 tiny_http = "0.12"
 tokio = "1"

codex-rs/app-server-protocol/Cargo.toml

@@ -15,7 +15,6 @@ anyhow = { workspace = true }
 clap = { workspace = true, features = ["derive"] }
 codex-protocol = { workspace = true }
 mcp-types = { workspace = true }
-paste = { workspace = true }
 schemars = { workspace = true }
 serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }

codex-rs/app-server-protocol/src/export.rs

@@ -61,7 +61,32 @@ pub fn generate_types(out_dir: &Path, prettier: Option<&Path>) -> Result<()> {
     Ok(())
 }
 
+#[derive(Clone, Copy, Debug)]
+pub struct GenerateTsOptions {
+    pub generate_indices: bool,
+    pub ensure_headers: bool,
+    pub run_prettier: bool,
+}
+
+impl Default for GenerateTsOptions {
+    fn default() -> Self {
+        Self {
+            generate_indices: true,
+            ensure_headers: true,
+            run_prettier: true,
+        }
+    }
+}
+
 pub fn generate_ts(out_dir: &Path, prettier: Option<&Path>) -> Result<()> {
+    generate_ts_with_options(out_dir, prettier, GenerateTsOptions::default())
+}
+
+pub fn generate_ts_with_options(
+    out_dir: &Path,
+    prettier: Option<&Path>,
+    options: GenerateTsOptions,
+) -> Result<()> {
     let v2_out_dir = out_dir.join("v2");
     ensure_dir(out_dir)?;
     ensure_dir(&v2_out_dir)?;
@@ -74,17 +99,28 @@ pub fn generate_ts(out_dir: &Path, prettier: Option<&Path>) -> Result<()> {
     export_server_responses(out_dir)?;
     ServerNotification::export_all_to(out_dir)?;
 
-    generate_index_ts(out_dir)?;
-    generate_index_ts(&v2_out_dir)?;
+    if options.generate_indices {
+        generate_index_ts(out_dir)?;
+        generate_index_ts(&v2_out_dir)?;
+    }
 
     // Ensure our header is present on all TS files (root + subdirs like v2/).
-    let ts_files = ts_files_in_recursive(out_dir)?;
-    for file in &ts_files {
-        prepend_header_if_missing(file)?;
+    let mut ts_files = Vec::new();
+    let should_collect_ts_files =
+        options.ensure_headers || (options.run_prettier && prettier.is_some());
+    if should_collect_ts_files {
+        ts_files = ts_files_in_recursive(out_dir)?;
+    }
+
+    if options.ensure_headers {
+        for file in &ts_files {
+            prepend_header_if_missing(file)?;
+        }
     }
 
     // Optionally run Prettier on all generated TS files.
-    if let Some(prettier_bin) = prettier
+    if options.run_prettier
+        && let Some(prettier_bin) = prettier
         && !ts_files.is_empty()
     {
         let status = Command::new(prettier_bin)
@@ -723,7 +759,13 @@ mod tests {
 
         let _guard = TempDirGuard(output_dir.clone());
 
-        generate_ts(&output_dir, None)?;
+        // Avoid doing more work than necessary to keep the test from timing out.
+        let options = GenerateTsOptions {
+            generate_indices: false,
+            ensure_headers: false,
+            run_prettier: false,
+        };
+        generate_ts_with_options(&output_dir, None, options)?;
 
         let mut undefined_offenders = Vec::new();
         let mut optional_nullable_offenders = BTreeSet::new();

codex-rs/app-server-protocol/src/protocol/common.rs

@@ -1,6 +1,4 @@
-use std::collections::HashMap;
 use std::path::Path;
-use std::path::PathBuf;
 
 use crate::JSONRPCNotification;
 use crate::JSONRPCRequest;
@@ -9,12 +7,6 @@ use crate::export::GeneratedSchema;
 use crate::export::write_json_schema;
 use crate::protocol::v1;
 use crate::protocol::v2;
-use codex_protocol::ConversationId;
-use codex_protocol::parse_command::ParsedCommand;
-use codex_protocol::protocol::FileChange;
-use codex_protocol::protocol::ReviewDecision;
-use codex_protocol::protocol::SandboxCommandAssessment;
-use paste::paste;
 use schemars::JsonSchema;
 use serde::Deserialize;
 use serde::Serialize;
@@ -137,6 +129,10 @@ client_request_definitions! {
         params: v2::TurnInterruptParams,
         response: v2::TurnInterruptResponse,
     },
+    ReviewStart => "review/start" {
+        params: v2::ReviewStartParams,
+        response: v2::TurnStartResponse,
+    },
 
     ModelList => "model/list" {
         params: v2::ModelListParams,
@@ -277,34 +273,36 @@ macro_rules! server_request_definitions {
     (
         $(
             $(#[$variant_meta:meta])*
-            $variant:ident
+            $variant:ident $(=> $wire:literal)? {
+                params: $params:ty,
+                response: $response:ty,
+            }
         ),* $(,)?
     ) => {
-        paste! {
-            /// Request initiated from the server and sent to the client.
-            #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-            #[serde(tag = "method", rename_all = "camelCase")]
-            pub enum ServerRequest {
-                $(
-                    $(#[$variant_meta])*
-                    $variant {
-                        #[serde(rename = "id")]
-                        request_id: RequestId,
-                        params: [<$variant Params>],
-                    },
-                )*
-            }
+        /// Request initiated from the server and sent to the client.
+        #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+        #[serde(tag = "method", rename_all = "camelCase")]
+        pub enum ServerRequest {
+            $(
+                $(#[$variant_meta])*
+                $(#[serde(rename = $wire)] #[ts(rename = $wire)])?
+                $variant {
+                    #[serde(rename = "id")]
+                    request_id: RequestId,
+                    params: $params,
+                },
+            )*
+        }
 
-            #[derive(Debug, Clone, PartialEq, JsonSchema)]
-            pub enum ServerRequestPayload {
-                $( $variant([<$variant Params>]), )*
-            }
+        #[derive(Debug, Clone, PartialEq, JsonSchema)]
+        pub enum ServerRequestPayload {
+            $( $variant($params), )*
+        }
 
-            impl ServerRequestPayload {
-                pub fn request_with_id(self, request_id: RequestId) -> ServerRequest {
-                    match self {
-                        $(Self::$variant(params) => ServerRequest::$variant { request_id, params },)*
-                    }
+        impl ServerRequestPayload {
+            pub fn request_with_id(self, request_id: RequestId) -> ServerRequest {
+                match self {
+                    $(Self::$variant(params) => ServerRequest::$variant { request_id, params },)*
                 }
             }
         }
@@ -312,9 +310,9 @@ macro_rules! server_request_definitions {
         pub fn export_server_responses(
             out_dir: &::std::path::Path,
         ) -> ::std::result::Result<(), ::ts_rs::ExportError> {
-            paste! {
-                $(<[<$variant Response>] as ::ts_rs::TS>::export_all_to(out_dir)?;)*
-            }
+            $(
+                <$response as ::ts_rs::TS>::export_all_to(out_dir)?;
+            )*
             Ok(())
         }
 
@@ -323,9 +321,12 @@ macro_rules! server_request_definitions {
             out_dir: &Path,
         ) -> ::anyhow::Result<Vec<GeneratedSchema>> {
             let mut schemas = Vec::new();
-            paste! {
-                $(schemas.push(crate::export::write_json_schema::<[<$variant Response>]>(out_dir, stringify!([<$variant Response>]))?);)*
-            }
+            $(
+                schemas.push(crate::export::write_json_schema::<$response>(
+                    out_dir,
+                    concat!(stringify!($variant), "Response"),
+                )?);
+            )*
             Ok(schemas)
         }
 
@@ -334,9 +335,12 @@ macro_rules! server_request_definitions {
             out_dir: &Path,
         ) -> ::anyhow::Result<Vec<GeneratedSchema>> {
             let mut schemas = Vec::new();
-            paste! {
-                $(schemas.push(crate::export::write_json_schema::<[<$variant Params>]>(out_dir, stringify!([<$variant Params>]))?);)*
-            }
+            $(
+                schemas.push(crate::export::write_json_schema::<$params>(
+                    out_dir,
+                    concat!(stringify!($variant), "Params"),
+                )?);
+            )*
             Ok(schemas)
         }
     };
@@ -426,49 +430,27 @@ impl TryFrom<JSONRPCRequest> for ServerRequest {
 }
 
 server_request_definitions! {
+    /// NEW APIs
+    /// Sent when approval is requested for a specific command execution.
+    /// This request is used for Turns started via turn/start.
+    CommandExecutionRequestApproval => "item/commandExecution/requestApproval" {
+        params: v2::CommandExecutionRequestApprovalParams,
+        response: v2::CommandExecutionRequestApprovalResponse,
+    },
+
+    /// DEPRECATED APIs below
     /// Request to approve a patch.
-    ApplyPatchApproval,
+    /// This request is used for Turns started via the legacy APIs (i.e. SendUserTurn, SendUserMessage).
+    ApplyPatchApproval {
+        params: v1::ApplyPatchApprovalParams,
+        response: v1::ApplyPatchApprovalResponse,
+    },
     /// Request to exec a command.
-    ExecCommandApproval,
-}
-
-#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-#[serde(rename_all = "camelCase")]
-pub struct ApplyPatchApprovalParams {
-    pub conversation_id: ConversationId,
-    /// Use to correlate this with [codex_core::protocol::PatchApplyBeginEvent]
-    /// and [codex_core::protocol::PatchApplyEndEvent].
-    pub call_id: String,
-    pub file_changes: HashMap<PathBuf, FileChange>,
-    /// Optional explanatory reason (e.g. request for extra write access).
-    pub reason: Option<String>,
-    /// When set, the agent is asking the user to allow writes under this root
-    /// for the remainder of the session (unclear if this is honored today).
-    pub grant_root: Option<PathBuf>,
-}
-
-#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-#[serde(rename_all = "camelCase")]
-pub struct ExecCommandApprovalParams {
-    pub conversation_id: ConversationId,
-    /// Use to correlate this with [codex_core::protocol::ExecCommandBeginEvent]
-    /// and [codex_core::protocol::ExecCommandEndEvent].
-    pub call_id: String,
-    pub command: Vec<String>,
-    pub cwd: PathBuf,
-    pub reason: Option<String>,
-    pub risk: Option<SandboxCommandAssessment>,
-    pub parsed_cmd: Vec<ParsedCommand>,
-}
-
-#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-pub struct ExecCommandApprovalResponse {
-    pub decision: ReviewDecision,
-}
-
-#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-pub struct ApplyPatchApprovalResponse {
-    pub decision: ReviewDecision,
+    /// This request is used for Turns started via the legacy APIs (i.e. SendUserTurn, SendUserMessage).
+    ExecCommandApproval {
+        params: v1::ExecCommandApprovalParams,
+        response: v1::ExecCommandApprovalResponse,
+    },
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -533,17 +515,20 @@ client_notification_definitions! {
 mod tests {
     use super::*;
     use anyhow::Result;
+    use codex_protocol::ConversationId;
     use codex_protocol::account::PlanType;
+    use codex_protocol::parse_command::ParsedCommand;
     use codex_protocol::protocol::AskForApproval;
     use pretty_assertions::assert_eq;
     use serde_json::json;
+    use std::path::PathBuf;
 
     #[test]
     fn serialize_new_conversation() -> Result<()> {
         let request = ClientRequest::NewConversation {
             request_id: RequestId::Integer(42),
             params: v1::NewConversationParams {
-                model: Some("gpt-5-codex".to_string()),
+                model: Some("gpt-5.1-codex".to_string()),
                 model_provider: None,
                 profile: None,
                 cwd: None,
@@ -561,7 +546,7 @@ mod tests {
                 "method": "newConversation",
                 "id": 42,
                 "params": {
-                    "model": "gpt-5-codex",
+                    "model": "gpt-5.1-codex",
                     "modelProvider": null,
                     "profile": null,
                     "cwd": null,
@@ -616,7 +601,7 @@ mod tests {
     #[test]
     fn serialize_server_request() -> Result<()> {
         let conversation_id = ConversationId::from_string("67e55044-10b1-426f-9247-bb680e5fe0c8")?;
-        let params = ExecCommandApprovalParams {
+        let params = v1::ExecCommandApprovalParams {
             conversation_id,
             call_id: "call-42".to_string(),
             command: vec!["echo".to_string(), "hello".to_string()],

codex-rs/app-server-protocol/src/protocol/v1.rs

@@ -8,8 +8,12 @@ use codex_protocol::config_types::ReasoningSummary;
 use codex_protocol::config_types::SandboxMode;
 use codex_protocol::config_types::Verbosity;
 use codex_protocol::models::ResponseItem;
+use codex_protocol::parse_command::ParsedCommand;
 use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::EventMsg;
+use codex_protocol::protocol::FileChange;
+use codex_protocol::protocol::ReviewDecision;
+use codex_protocol::protocol::SandboxCommandAssessment;
 use codex_protocol::protocol::SandboxPolicy;
 use codex_protocol::protocol::SessionSource;
 use codex_protocol::protocol::TurnAbortReason;
@@ -191,6 +195,46 @@ pub struct GitDiffToRemoteResponse {
     pub diff: String,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct ApplyPatchApprovalParams {
+    pub conversation_id: ConversationId,
+    /// Use to correlate this with [codex_core::protocol::PatchApplyBeginEvent]
+    /// and [codex_core::protocol::PatchApplyEndEvent].
+    pub call_id: String,
+    pub file_changes: HashMap<PathBuf, FileChange>,
+    /// Optional explanatory reason (e.g. request for extra write access).
+    pub reason: Option<String>,
+    /// When set, the agent is asking the user to allow writes under this root
+    /// for the remainder of the session (unclear if this is honored today).
+    pub grant_root: Option<PathBuf>,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct ApplyPatchApprovalResponse {
+    pub decision: ReviewDecision,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct ExecCommandApprovalParams {
+    pub conversation_id: ConversationId,
+    /// Use to correlate this with [codex_core::protocol::ExecCommandBeginEvent]
+    /// and [codex_core::protocol::ExecCommandEndEvent].
+    pub call_id: String,
+    pub command: Vec<String>,
+    pub cwd: PathBuf,
+    pub reason: Option<String>,
+    pub risk: Option<SandboxCommandAssessment>,
+    pub parsed_cmd: Vec<ParsedCommand>,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+pub struct ExecCommandApprovalResponse {
+    pub decision: ReviewDecision,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 pub struct CancelLoginChatGptParams {

codex-rs/app-server-protocol/src/protocol/v2.rs

@@ -4,11 +4,13 @@ use std::path::PathBuf;
 use crate::protocol::common::AuthMode;
 use codex_protocol::ConversationId;
 use codex_protocol::account::PlanType;
+use codex_protocol::approvals::SandboxCommandAssessment as CoreSandboxCommandAssessment;
 use codex_protocol::config_types::ReasoningEffort;
 use codex_protocol::config_types::ReasoningSummary;
 use codex_protocol::items::AgentMessageContent as CoreAgentMessageContent;
 use codex_protocol::items::TurnItem as CoreTurnItem;
 use codex_protocol::models::ResponseItem;
+use codex_protocol::parse_command::ParsedCommand as CoreParsedCommand;
 use codex_protocol::protocol::RateLimitSnapshot as CoreRateLimitSnapshot;
 use codex_protocol::protocol::RateLimitWindow as CoreRateLimitWindow;
 use codex_protocol::user_input::UserInput as CoreUserInput;
@@ -20,7 +22,7 @@ use serde_json::Value as JsonValue;
 use ts_rs::TS;
 
 // Macro to declare a camelCased API v2 enum mirroring a core enum which
-// tends to use kebab-case.
+// tends to use either snake_case or kebab-case.
 macro_rules! v2_enum_from_core {
     (
         pub enum $Name:ident from $Src:path { $( $Variant:ident ),+ $(,)? }
@@ -56,6 +58,23 @@ v2_enum_from_core!(
     }
 );
 
+v2_enum_from_core!(
+    pub enum CommandRiskLevel from codex_protocol::approvals::SandboxRiskLevel {
+        Low,
+        Medium,
+        High
+    }
+);
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub enum ApprovalDecision {
+    Accept,
+    Decline,
+    Cancel,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq, JsonSchema, TS)]
 #[serde(tag = "type", rename_all = "camelCase")]
 #[ts(tag = "type")]
@@ -63,6 +82,8 @@ v2_enum_from_core!(
 pub enum SandboxPolicy {
     DangerFullAccess,
     ReadOnly,
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
     WorkspaceWrite {
         #[serde(default)]
         writable_roots: Vec<PathBuf>,
@@ -119,6 +140,98 @@ impl From<codex_protocol::protocol::SandboxPolicy> for SandboxPolicy {
     }
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct SandboxCommandAssessment {
+    pub description: String,
+    pub risk_level: CommandRiskLevel,
+}
+
+impl SandboxCommandAssessment {
+    pub fn into_core(self) -> CoreSandboxCommandAssessment {
+        CoreSandboxCommandAssessment {
+            description: self.description,
+            risk_level: self.risk_level.to_core(),
+        }
+    }
+}
+
+impl From<CoreSandboxCommandAssessment> for SandboxCommandAssessment {
+    fn from(value: CoreSandboxCommandAssessment) -> Self {
+        Self {
+            description: value.description,
+            risk_level: CommandRiskLevel::from(value.risk_level),
+        }
+    }
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[ts(tag = "type")]
+#[ts(export_to = "v2/")]
+pub enum CommandAction {
+    Read {
+        command: String,
+        name: String,
+        path: PathBuf,
+    },
+    ListFiles {
+        command: String,
+        path: Option<String>,
+    },
+    Search {
+        command: String,
+        query: Option<String>,
+        path: Option<String>,
+    },
+    Unknown {
+        command: String,
+    },
+}
+
+impl CommandAction {
+    pub fn into_core(self) -> CoreParsedCommand {
+        match self {
+            CommandAction::Read {
+                command: cmd,
+                name,
+                path,
+            } => CoreParsedCommand::Read { cmd, name, path },
+            CommandAction::ListFiles { command: cmd, path } => {
+                CoreParsedCommand::ListFiles { cmd, path }
+            }
+            CommandAction::Search {
+                command: cmd,
+                query,
+                path,
+            } => CoreParsedCommand::Search { cmd, query, path },
+            CommandAction::Unknown { command: cmd } => CoreParsedCommand::Unknown { cmd },
+        }
+    }
+}
+
+impl From<CoreParsedCommand> for CommandAction {
+    fn from(value: CoreParsedCommand) -> Self {
+        match value {
+            CoreParsedCommand::Read { cmd, name, path } => CommandAction::Read {
+                command: cmd,
+                name,
+                path,
+            },
+            CoreParsedCommand::ListFiles { cmd, path } => {
+                CommandAction::ListFiles { command: cmd, path }
+            }
+            CoreParsedCommand::Search { cmd, query, path } => CommandAction::Search {
+                command: cmd,
+                query,
+                path,
+            },
+            CoreParsedCommand::Unknown { cmd } => CommandAction::Unknown { command: cmd },
+        }
+    }
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(tag = "type", rename_all = "camelCase")]
 #[ts(tag = "type")]
@@ -279,7 +392,7 @@ pub struct ThreadStartParams {
     pub cwd: Option<String>,
     pub approval_policy: Option<AskForApproval>,
     pub sandbox: Option<SandboxMode>,
-    pub config: Option<HashMap<String, serde_json::Value>>,
+    pub config: Option<HashMap<String, JsonValue>>,
     pub base_instructions: Option<String>,
     pub developer_instructions: Option<String>,
 }
@@ -289,6 +402,12 @@ pub struct ThreadStartParams {
 #[ts(export_to = "v2/")]
 pub struct ThreadStartResponse {
     pub thread: Thread,
+    pub model: String,
+    pub model_provider: String,
+    pub cwd: PathBuf,
+    pub approval_policy: AskForApproval,
+    pub sandbox: SandboxPolicy,
+    pub reasoning_effort: Option<ReasoningEffort>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Default, Clone, PartialEq, JsonSchema, TS)]
@@ -331,6 +450,12 @@ pub struct ThreadResumeParams {
 #[ts(export_to = "v2/")]
 pub struct ThreadResumeResponse {
     pub thread: Thread,
+    pub model: String,
+    pub model_provider: String,
+    pub cwd: PathBuf,
+    pub approval_policy: AskForApproval,
+    pub sandbox: SandboxPolicy,
+    pub reasoning_effort: Option<ReasoningEffort>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -406,9 +531,11 @@ pub struct AccountUpdatedNotification {
 #[ts(export_to = "v2/")]
 pub struct Turn {
     pub id: String,
+    /// This is currently only populated for resumed threads.
+    /// TODO: properly populate items for all turns.
     pub items: Vec<ThreadItem>,
+    #[serde(flatten)]
     pub status: TurnStatus,
-    pub error: Option<TurnError>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -419,12 +546,12 @@ pub struct TurnError {
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-#[serde(rename_all = "camelCase")]
-#[ts(export_to = "v2/")]
+#[serde(tag = "status", rename_all = "camelCase")]
+#[ts(tag = "status", export_to = "v2/")]
 pub enum TurnStatus {
     Completed,
     Interrupted,
-    Failed,
+    Failed { error: TurnError },
     InProgress,
 }
 
@@ -449,6 +576,45 @@ pub struct TurnStartParams {
     pub summary: Option<ReasoningSummary>,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct ReviewStartParams {
+    pub thread_id: String,
+    pub target: ReviewTarget,
+
+    /// When true, also append the final review message to the original thread.
+    #[serde(default)]
+    pub append_to_original_thread: bool,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[ts(tag = "type", export_to = "v2/")]
+pub enum ReviewTarget {
+    /// Review the working tree: staged, unstaged, and untracked files.
+    UncommittedChanges,
+
+    /// Review changes between the current branch and the given base branch.
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    BaseBranch { branch: String },
+
+    /// Review the changes introduced by a specific commit.
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    Commit {
+        sha: String,
+        /// Optional human-readable label (e.g., commit subject) for UIs.
+        title: Option<String>,
+    },
+
+    /// Arbitrary instructions, equivalent to the old free-form prompt.
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    Custom { instructions: String },
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]
@@ -506,14 +672,14 @@ impl From<CoreUserInput> for UserInput {
 #[ts(tag = "type")]
 #[ts(export_to = "v2/")]
 pub enum ThreadItem {
-    UserMessage {
-        id: String,
-        content: Vec<UserInput>,
-    },
-    AgentMessage {
-        id: String,
-        text: String,
-    },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    UserMessage { id: String, content: Vec<UserInput> },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    AgentMessage { id: String, text: String },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
     Reasoning {
         id: String,
         #[serde(default)]
@@ -521,19 +687,35 @@ pub enum ThreadItem {
         #[serde(default)]
         content: Vec<String>,
     },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
     CommandExecution {
         id: String,
+        /// The command to be executed.
         command: String,
-        aggregated_output: String,
-        exit_code: Option<i32>,
+        /// The command's working directory.
+        cwd: PathBuf,
         status: CommandExecutionStatus,
+        /// A best-effort parsing of the command to understand the action(s) it will perform.
+        /// This returns a list of CommandAction objects because a single shell command may
+        /// be composed of many commands piped together.
+        command_actions: Vec<CommandAction>,
+        /// The command's output, aggregated from stdout and stderr.
+        aggregated_output: Option<String>,
+        /// The command's exit code.
+        exit_code: Option<i32>,
+        /// The duration of the command execution in milliseconds.
         duration_ms: Option<i64>,
     },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
     FileChange {
         id: String,
         changes: Vec<FileUpdateChange>,
         status: PatchApplyStatus,
     },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
     McpToolCall {
         id: String,
         server: String,
@@ -543,22 +725,18 @@ pub enum ThreadItem {
         result: Option<McpToolCallResult>,
         error: Option<McpToolCallError>,
     },
-    WebSearch {
-        id: String,
-        query: String,
-    },
-    TodoList {
-        id: String,
-        items: Vec<TodoItem>,
-    },
-    ImageView {
-        id: String,
-        path: String,
-    },
-    CodeReview {
-        id: String,
-        review: String,
-    },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    WebSearch { id: String, query: String },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    TodoList { id: String, items: Vec<TodoItem> },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    ImageView { id: String, path: String },
+    #[serde(rename_all = "camelCase")]
+    #[ts(rename_all = "camelCase")]
+    CodeReview { id: String, review: String },
 }
 
 impl From<CoreTurnItem> for ThreadItem {
@@ -640,7 +818,7 @@ pub enum McpToolCallStatus {
 #[ts(export_to = "v2/")]
 pub struct McpToolCallResult {
     pub content: Vec<McpContentBlock>,
-    pub structured_content: JsonValue,
+    pub structured_content: Option<JsonValue>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -689,8 +867,6 @@ pub struct Usage {
 #[ts(export_to = "v2/")]
 pub struct TurnCompletedNotification {
     pub turn: Turn,
-    // TODO: should usage be stored on the Turn object, and we return that instead?
-    pub usage: Usage,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -758,6 +934,39 @@ pub struct McpToolCallProgressNotification {
     pub message: String,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct CommandExecutionRequestApprovalParams {
+    pub thread_id: String,
+    pub turn_id: String,
+    pub item_id: String,
+    /// Optional explanatory reason (e.g. request for network access).
+    pub reason: Option<String>,
+    /// Optional model-provided risk assessment describing the blocked command.
+    pub risk: Option<SandboxCommandAssessment>,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct CommandExecutionRequestAcceptSettings {
+    /// If true, automatically approve this command for the duration of the session.
+    #[serde(default)]
+    pub for_session: bool,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct CommandExecutionRequestApprovalResponse {
+    pub decision: ApprovalDecision,
+    /// Optional approval settings for when the decision is `accept`.
+    /// Ignored if the decision is `decline` or `cancel`.
+    #[serde(default)]
+    pub accept_settings: Option<CommandExecutionRequestAcceptSettings>,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]

codex-rs/app-server-test-client/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "codex-app-server-test-client"
+version = { workspace = true }
+edition = "2024"
+
+[lints]
+workspace = true
+
+[dependencies]
+anyhow = { workspace = true }
+clap = { workspace = true, features = ["derive", "env"] }
+codex-app-server-protocol = { workspace = true }
+codex-protocol = { workspace = true }
+serde = { workspace = true, features = ["derive"] }
+serde_json = { workspace = true }
+uuid = { workspace = true, features = ["v4"] }

codex-rs/app-server-test-client/README.md

@@ -0,0 +1,2 @@
+# App Server Test Client
+Exercises simple `codex app-server` flows end-to-end, logging JSON-RPC messages sent between client and server to stdout.

codex-rs/app-server-test-client/src/main.rs

@@ -0,0 +1,773 @@
+use std::collections::VecDeque;
+use std::io::BufRead;
+use std::io::BufReader;
+use std::io::Write;
+use std::process::Child;
+use std::process::ChildStdin;
+use std::process::ChildStdout;
+use std::process::Command;
+use std::process::Stdio;
+use std::thread;
+use std::time::Duration;
+
+use anyhow::Context;
+use anyhow::Result;
+use anyhow::bail;
+use clap::Parser;
+use clap::Subcommand;
+use codex_app_server_protocol::AddConversationListenerParams;
+use codex_app_server_protocol::AddConversationSubscriptionResponse;
+use codex_app_server_protocol::ApprovalDecision;
+use codex_app_server_protocol::AskForApproval;
+use codex_app_server_protocol::ClientInfo;
+use codex_app_server_protocol::ClientRequest;
+use codex_app_server_protocol::CommandExecutionRequestAcceptSettings;
+use codex_app_server_protocol::CommandExecutionRequestApprovalParams;
+use codex_app_server_protocol::CommandExecutionRequestApprovalResponse;
+use codex_app_server_protocol::GetAccountRateLimitsResponse;
+use codex_app_server_protocol::InitializeParams;
+use codex_app_server_protocol::InitializeResponse;
+use codex_app_server_protocol::InputItem;
+use codex_app_server_protocol::JSONRPCMessage;
+use codex_app_server_protocol::JSONRPCNotification;
+use codex_app_server_protocol::JSONRPCRequest;
+use codex_app_server_protocol::JSONRPCResponse;
+use codex_app_server_protocol::LoginChatGptCompleteNotification;
+use codex_app_server_protocol::LoginChatGptResponse;
+use codex_app_server_protocol::NewConversationParams;
+use codex_app_server_protocol::NewConversationResponse;
+use codex_app_server_protocol::RequestId;
+use codex_app_server_protocol::SandboxPolicy;
+use codex_app_server_protocol::SendUserMessageParams;
+use codex_app_server_protocol::SendUserMessageResponse;
+use codex_app_server_protocol::ServerNotification;
+use codex_app_server_protocol::ServerRequest;
+use codex_app_server_protocol::ThreadStartParams;
+use codex_app_server_protocol::ThreadStartResponse;
+use codex_app_server_protocol::TurnStartParams;
+use codex_app_server_protocol::TurnStartResponse;
+use codex_app_server_protocol::TurnStatus;
+use codex_app_server_protocol::UserInput as V2UserInput;
+use codex_protocol::ConversationId;
+use codex_protocol::protocol::Event;
+use codex_protocol::protocol::EventMsg;
+use serde::Serialize;
+use serde::de::DeserializeOwned;
+use serde_json::Value;
+use uuid::Uuid;
+
+/// Minimal launcher that initializes the Codex app-server and logs the handshake.
+#[derive(Parser)]
+#[command(author = "Codex", version, about = "Bootstrap Codex app-server", long_about = None)]
+struct Cli {
+    /// Path to the `codex` CLI binary.
+    #[arg(long, env = "CODEX_BIN", default_value = "codex")]
+    codex_bin: String,
+
+    #[command(subcommand)]
+    command: CliCommand,
+}
+
+#[derive(Subcommand)]
+enum CliCommand {
+    /// Send a user message through the Codex app-server.
+    SendMessage {
+        /// User message to send to Codex.
+        #[arg()]
+        user_message: String,
+    },
+    /// Send a user message through the app-server V2 thread/turn APIs.
+    SendMessageV2 {
+        /// User message to send to Codex.
+        #[arg()]
+        user_message: String,
+    },
+    /// Start a V2 turn that elicits an ExecCommand approval.
+    #[command(name = "trigger-cmd-approval")]
+    TriggerCmdApproval {
+        /// Optional prompt; defaults to a simple python command.
+        #[arg()]
+        user_message: Option<String>,
+    },
+    /// Start a V2 turn that elicits an ApplyPatch approval.
+    #[command(name = "trigger-patch-approval")]
+    TriggerPatchApproval {
+        /// Optional prompt; defaults to creating a file via apply_patch.
+        #[arg()]
+        user_message: Option<String>,
+    },
+    /// Start a V2 turn that should not elicit an ExecCommand approval.
+    #[command(name = "no-trigger-cmd-approval")]
+    NoTriggerCmdApproval,
+    /// Trigger the ChatGPT login flow and wait for completion.
+    TestLogin,
+    /// Fetch the current account rate limits from the Codex app-server.
+    GetAccountRateLimits,
+}
+
+fn main() -> Result<()> {
+    let Cli { codex_bin, command } = Cli::parse();
+
+    match command {
+        CliCommand::SendMessage { user_message } => send_message(codex_bin, user_message),
+        CliCommand::SendMessageV2 { user_message } => send_message_v2(codex_bin, user_message),
+        CliCommand::TriggerCmdApproval { user_message } => {
+            trigger_cmd_approval(codex_bin, user_message)
+        }
+        CliCommand::TriggerPatchApproval { user_message } => {
+            trigger_patch_approval(codex_bin, user_message)
+        }
+        CliCommand::NoTriggerCmdApproval => no_trigger_cmd_approval(codex_bin),
+        CliCommand::TestLogin => test_login(codex_bin),
+        CliCommand::GetAccountRateLimits => get_account_rate_limits(codex_bin),
+    }
+}
+
+fn send_message(codex_bin: String, user_message: String) -> Result<()> {
+    let mut client = CodexClient::spawn(codex_bin)?;
+
+    let initialize = client.initialize()?;
+    println!("< initialize response: {initialize:?}");
+
+    let conversation = client.new_conversation()?;
+    println!("< newConversation response: {conversation:?}");
+
+    let subscription = client.add_conversation_listener(&conversation.conversation_id)?;
+    println!("< addConversationListener response: {subscription:?}");
+
+    let send_response = client.send_user_message(&conversation.conversation_id, &user_message)?;
+    println!("< sendUserMessage response: {send_response:?}");
+
+    client.stream_conversation(&conversation.conversation_id)?;
+
+    client.remove_conversation_listener(subscription.subscription_id)?;
+
+    Ok(())
+}
+
+fn send_message_v2(codex_bin: String, user_message: String) -> Result<()> {
+    send_message_v2_with_policies(codex_bin, user_message, None, None)
+}
+
+fn trigger_cmd_approval(codex_bin: String, user_message: Option<String>) -> Result<()> {
+    let default_prompt =
+        "Run `touch /tmp/should-trigger-approval` so I can confirm the file exists.";
+    let message = user_message.unwrap_or_else(|| default_prompt.to_string());
+    send_message_v2_with_policies(
+        codex_bin,
+        message,
+        Some(AskForApproval::OnRequest),
+        Some(SandboxPolicy::ReadOnly),
+    )
+}
+
+fn trigger_patch_approval(codex_bin: String, user_message: Option<String>) -> Result<()> {
+    let default_prompt =
+        "Create a file named APPROVAL_DEMO.txt containing a short hello message using apply_patch.";
+    let message = user_message.unwrap_or_else(|| default_prompt.to_string());
+    send_message_v2_with_policies(
+        codex_bin,
+        message,
+        Some(AskForApproval::OnRequest),
+        Some(SandboxPolicy::ReadOnly),
+    )
+}
+
+fn no_trigger_cmd_approval(codex_bin: String) -> Result<()> {
+    let prompt = "Run `touch should_not_trigger_approval.txt`";
+    send_message_v2_with_policies(codex_bin, prompt.to_string(), None, None)
+}
+
+fn send_message_v2_with_policies(
+    codex_bin: String,
+    user_message: String,
+    approval_policy: Option<AskForApproval>,
+    sandbox_policy: Option<SandboxPolicy>,
+) -> Result<()> {
+    let mut client = CodexClient::spawn(codex_bin)?;
+
+    let initialize = client.initialize()?;
+    println!("< initialize response: {initialize:?}");
+
+    let thread_response = client.thread_start(ThreadStartParams::default())?;
+    println!("< thread/start response: {thread_response:?}");
+    let mut turn_params = TurnStartParams {
+        thread_id: thread_response.thread.id.clone(),
+        input: vec![V2UserInput::Text { text: user_message }],
+        ..Default::default()
+    };
+    turn_params.approval_policy = approval_policy;
+    turn_params.sandbox_policy = sandbox_policy;
+
+    let turn_response = client.turn_start(turn_params)?;
+    println!("< turn/start response: {turn_response:?}");
+
+    client.stream_turn(&thread_response.thread.id, &turn_response.turn.id)?;
+
+    Ok(())
+}
+
+fn test_login(codex_bin: String) -> Result<()> {
+    let mut client = CodexClient::spawn(codex_bin)?;
+
+    let initialize = client.initialize()?;
+    println!("< initialize response: {initialize:?}");
+
+    let login_response = client.login_chat_gpt()?;
+    println!("< loginChatGpt response: {login_response:?}");
+    println!(
+        "Open the following URL in your browser to continue:\n{}",
+        login_response.auth_url
+    );
+
+    let completion = client.wait_for_login_completion(&login_response.login_id)?;
+    println!("< loginChatGptComplete notification: {completion:?}");
+
+    if completion.success {
+        println!("Login succeeded.");
+        Ok(())
+    } else {
+        bail!(
+            "login failed: {}",
+            completion
+                .error
+                .as_deref()
+                .unwrap_or("unknown error from loginChatGptComplete")
+        );
+    }
+}
+
+fn get_account_rate_limits(codex_bin: String) -> Result<()> {
+    let mut client = CodexClient::spawn(codex_bin)?;
+
+    let initialize = client.initialize()?;
+    println!("< initialize response: {initialize:?}");
+
+    let response = client.get_account_rate_limits()?;
+    println!("< account/rateLimits/read response: {response:?}");
+
+    Ok(())
+}
+
+struct CodexClient {
+    child: Child,
+    stdin: Option<ChildStdin>,
+    stdout: BufReader<ChildStdout>,
+    pending_notifications: VecDeque<JSONRPCNotification>,
+}
+
+impl CodexClient {
+    fn spawn(codex_bin: String) -> Result<Self> {
+        let mut codex_app_server = Command::new(&codex_bin)
+            .arg("app-server")
+            .stdin(Stdio::piped())
+            .stdout(Stdio::piped())
+            .stderr(Stdio::inherit())
+            .spawn()
+            .with_context(|| format!("failed to start `{codex_bin}` app-server"))?;
+
+        let stdin = codex_app_server
+            .stdin
+            .take()
+            .context("codex app-server stdin unavailable")?;
+        let stdout = codex_app_server
+            .stdout
+            .take()
+            .context("codex app-server stdout unavailable")?;
+
+        Ok(Self {
+            child: codex_app_server,
+            stdin: Some(stdin),
+            stdout: BufReader::new(stdout),
+            pending_notifications: VecDeque::new(),
+        })
+    }
+
+    fn initialize(&mut self) -> Result<InitializeResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::Initialize {
+            request_id: request_id.clone(),
+            params: InitializeParams {
+                client_info: ClientInfo {
+                    name: "codex-toy-app-server".to_string(),
+                    title: Some("Codex Toy App Server".to_string()),
+                    version: env!("CARGO_PKG_VERSION").to_string(),
+                },
+            },
+        };
+
+        self.send_request(request, request_id, "initialize")
+    }
+
+    fn new_conversation(&mut self) -> Result<NewConversationResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::NewConversation {
+            request_id: request_id.clone(),
+            params: NewConversationParams::default(),
+        };
+
+        self.send_request(request, request_id, "newConversation")
+    }
+
+    fn add_conversation_listener(
+        &mut self,
+        conversation_id: &ConversationId,
+    ) -> Result<AddConversationSubscriptionResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::AddConversationListener {
+            request_id: request_id.clone(),
+            params: AddConversationListenerParams {
+                conversation_id: *conversation_id,
+                experimental_raw_events: false,
+            },
+        };
+
+        self.send_request(request, request_id, "addConversationListener")
+    }
+
+    fn remove_conversation_listener(&mut self, subscription_id: Uuid) -> Result<()> {
+        let request_id = self.request_id();
+        let request = ClientRequest::RemoveConversationListener {
+            request_id: request_id.clone(),
+            params: codex_app_server_protocol::RemoveConversationListenerParams { subscription_id },
+        };
+
+        self.send_request::<codex_app_server_protocol::RemoveConversationSubscriptionResponse>(
+            request,
+            request_id,
+            "removeConversationListener",
+        )?;
+
+        Ok(())
+    }
+
+    fn send_user_message(
+        &mut self,
+        conversation_id: &ConversationId,
+        message: &str,
+    ) -> Result<SendUserMessageResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::SendUserMessage {
+            request_id: request_id.clone(),
+            params: SendUserMessageParams {
+                conversation_id: *conversation_id,
+                items: vec![InputItem::Text {
+                    text: message.to_string(),
+                }],
+            },
+        };
+
+        self.send_request(request, request_id, "sendUserMessage")
+    }
+
+    fn thread_start(&mut self, params: ThreadStartParams) -> Result<ThreadStartResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::ThreadStart {
+            request_id: request_id.clone(),
+            params,
+        };
+
+        self.send_request(request, request_id, "thread/start")
+    }
+
+    fn turn_start(&mut self, params: TurnStartParams) -> Result<TurnStartResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::TurnStart {
+            request_id: request_id.clone(),
+            params,
+        };
+
+        self.send_request(request, request_id, "turn/start")
+    }
+
+    fn login_chat_gpt(&mut self) -> Result<LoginChatGptResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::LoginChatGpt {
+            request_id: request_id.clone(),
+            params: None,
+        };
+
+        self.send_request(request, request_id, "loginChatGpt")
+    }
+
+    fn get_account_rate_limits(&mut self) -> Result<GetAccountRateLimitsResponse> {
+        let request_id = self.request_id();
+        let request = ClientRequest::GetAccountRateLimits {
+            request_id: request_id.clone(),
+            params: None,
+        };
+
+        self.send_request(request, request_id, "account/rateLimits/read")
+    }
+
+    fn stream_conversation(&mut self, conversation_id: &ConversationId) -> Result<()> {
+        loop {
+            let notification = self.next_notification()?;
+
+            if !notification.method.starts_with("codex/event/") {
+                continue;
+            }
+
+            if let Some(event) = self.extract_event(notification, conversation_id)? {
+                match &event.msg {
+                    EventMsg::AgentMessage(event) => {
+                        println!("{}", event.message);
+                    }
+                    EventMsg::AgentMessageDelta(event) => {
+                        print!("{}", event.delta);
+                        std::io::stdout().flush().ok();
+                    }
+                    EventMsg::TaskComplete(event) => {
+                        println!("\n[task complete: {event:?}]");
+                        break;
+                    }
+                    EventMsg::TurnAborted(event) => {
+                        println!("\n[turn aborted: {:?}]", event.reason);
+                        break;
+                    }
+                    EventMsg::Error(event) => {
+                        println!("[error] {event:?}");
+                    }
+                    _ => {
+                        println!("[UNKNOWN EVENT] {:?}", event.msg);
+                    }
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    fn wait_for_login_completion(
+        &mut self,
+        expected_login_id: &Uuid,
+    ) -> Result<LoginChatGptCompleteNotification> {
+        loop {
+            let notification = self.next_notification()?;
+
+            if let Ok(server_notification) = ServerNotification::try_from(notification) {
+                match server_notification {
+                    ServerNotification::LoginChatGptComplete(completion) => {
+                        if &completion.login_id == expected_login_id {
+                            return Ok(completion);
+                        }
+
+                        println!(
+                            "[ignoring loginChatGptComplete for unexpected login_id: {}]",
+                            completion.login_id
+                        );
+                    }
+                    ServerNotification::AuthStatusChange(status) => {
+                        println!("< authStatusChange notification: {status:?}");
+                    }
+                    ServerNotification::AccountRateLimitsUpdated(snapshot) => {
+                        println!("< accountRateLimitsUpdated notification: {snapshot:?}");
+                    }
+                    ServerNotification::SessionConfigured(_) => {
+                        // SessionConfigured notifications are unrelated to login; skip.
+                    }
+                    _ => {}
+                }
+            }
+
+            // Not a server notification (likely a conversation event); keep waiting.
+        }
+    }
+
+    fn stream_turn(&mut self, thread_id: &str, turn_id: &str) -> Result<()> {
+        loop {
+            let notification = self.next_notification()?;
+
+            let Ok(server_notification) = ServerNotification::try_from(notification) else {
+                continue;
+            };
+
+            match server_notification {
+                ServerNotification::ThreadStarted(payload) => {
+                    if payload.thread.id == thread_id {
+                        println!("< thread/started notification: {:?}", payload.thread);
+                    }
+                }
+                ServerNotification::TurnStarted(payload) => {
+                    if payload.turn.id == turn_id {
+                        println!("< turn/started notification: {:?}", payload.turn.status);
+                    }
+                }
+                ServerNotification::AgentMessageDelta(delta) => {
+                    print!("{}", delta.delta);
+                    std::io::stdout().flush().ok();
+                }
+                ServerNotification::CommandExecutionOutputDelta(delta) => {
+                    print!("{}", delta.delta);
+                    std::io::stdout().flush().ok();
+                }
+                ServerNotification::ItemStarted(payload) => {
+                    println!("\n< item started: {:?}", payload.item);
+                }
+                ServerNotification::ItemCompleted(payload) => {
+                    println!("< item completed: {:?}", payload.item);
+                }
+                ServerNotification::TurnCompleted(payload) => {
+                    if payload.turn.id == turn_id {
+                        println!("\n< turn/completed notification: {:?}", payload.turn.status);
+                        if let TurnStatus::Failed { error } = &payload.turn.status {
+                            println!("[turn error] {}", error.message);
+                        }
+                        break;
+                    }
+                }
+                ServerNotification::McpToolCallProgress(payload) => {
+                    println!("< MCP tool progress: {}", payload.message);
+                }
+                _ => {
+                    println!("[UNKNOWN SERVER NOTIFICATION] {server_notification:?}");
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    fn extract_event(
+        &self,
+        notification: JSONRPCNotification,
+        conversation_id: &ConversationId,
+    ) -> Result<Option<Event>> {
+        let params = notification
+            .params
+            .context("event notification missing params")?;
+
+        let mut map = match params {
+            Value::Object(map) => map,
+            other => bail!("unexpected params shape: {other:?}"),
+        };
+
+        let conversation_value = map
+            .remove("conversationId")
+            .context("event missing conversationId")?;
+        let notification_conversation: ConversationId = serde_json::from_value(conversation_value)
+            .context("conversationId was not a valid UUID")?;
+
+        if &notification_conversation != conversation_id {
+            return Ok(None);
+        }
+
+        let event_value = Value::Object(map);
+        let event: Event =
+            serde_json::from_value(event_value).context("failed to decode event payload")?;
+        Ok(Some(event))
+    }
+
+    fn send_request<T>(
+        &mut self,
+        request: ClientRequest,
+        request_id: RequestId,
+        method: &str,
+    ) -> Result<T>
+    where
+        T: DeserializeOwned,
+    {
+        self.write_request(&request)?;
+        self.wait_for_response(request_id, method)
+    }
+
+    fn write_request(&mut self, request: &ClientRequest) -> Result<()> {
+        let request_json = serde_json::to_string(request)?;
+        let request_pretty = serde_json::to_string_pretty(request)?;
+        print_multiline_with_prefix("> ", &request_pretty);
+
+        if let Some(stdin) = self.stdin.as_mut() {
+            writeln!(stdin, "{request_json}")?;
+            stdin
+                .flush()
+                .context("failed to flush request to codex app-server")?;
+        } else {
+            bail!("codex app-server stdin closed");
+        }
+
+        Ok(())
+    }
+
+    fn wait_for_response<T>(&mut self, request_id: RequestId, method: &str) -> Result<T>
+    where
+        T: DeserializeOwned,
+    {
+        loop {
+            let message = self.read_jsonrpc_message()?;
+
+            match message {
+                JSONRPCMessage::Response(JSONRPCResponse { id, result }) => {
+                    if id == request_id {
+                        return serde_json::from_value(result)
+                            .with_context(|| format!("{method} response missing payload"));
+                    }
+                }
+                JSONRPCMessage::Error(err) => {
+                    if err.id == request_id {
+                        bail!("{method} failed: {err:?}");
+                    }
+                }
+                JSONRPCMessage::Notification(notification) => {
+                    self.pending_notifications.push_back(notification);
+                }
+                JSONRPCMessage::Request(request) => {
+                    self.handle_server_request(request)?;
+                }
+            }
+        }
+    }
+
+    fn next_notification(&mut self) -> Result<JSONRPCNotification> {
+        if let Some(notification) = self.pending_notifications.pop_front() {
+            return Ok(notification);
+        }
+
+        loop {
+            let message = self.read_jsonrpc_message()?;
+
+            match message {
+                JSONRPCMessage::Notification(notification) => return Ok(notification),
+                JSONRPCMessage::Response(_) | JSONRPCMessage::Error(_) => {
+                    // No outstanding requests, so ignore stray responses/errors for now.
+                    continue;
+                }
+                JSONRPCMessage::Request(request) => {
+                    self.handle_server_request(request)?;
+                }
+            }
+        }
+    }
+
+    fn read_jsonrpc_message(&mut self) -> Result<JSONRPCMessage> {
+        loop {
+            let mut response_line = String::new();
+            let bytes = self
+                .stdout
+                .read_line(&mut response_line)
+                .context("failed to read from codex app-server")?;
+
+            if bytes == 0 {
+                bail!("codex app-server closed stdout");
+            }
+
+            let trimmed = response_line.trim();
+            if trimmed.is_empty() {
+                continue;
+            }
+
+            let parsed: Value =
+                serde_json::from_str(trimmed).context("response was not valid JSON-RPC")?;
+            let pretty = serde_json::to_string_pretty(&parsed)?;
+            print_multiline_with_prefix("< ", &pretty);
+            let message: JSONRPCMessage = serde_json::from_value(parsed)
+                .context("response was not a valid JSON-RPC message")?;
+            return Ok(message);
+        }
+    }
+
+    fn request_id(&self) -> RequestId {
+        RequestId::String(Uuid::new_v4().to_string())
+    }
+
+    fn handle_server_request(&mut self, request: JSONRPCRequest) -> Result<()> {
+        let server_request = ServerRequest::try_from(request)
+            .context("failed to deserialize ServerRequest from JSONRPCRequest")?;
+
+        match server_request {
+            ServerRequest::CommandExecutionRequestApproval { request_id, params } => {
+                self.handle_command_execution_request_approval(request_id, params)?;
+            }
+            other => {
+                bail!("received unsupported server request: {other:?}");
+            }
+        }
+
+        Ok(())
+    }
+
+    fn handle_command_execution_request_approval(
+        &mut self,
+        request_id: RequestId,
+        params: CommandExecutionRequestApprovalParams,
+    ) -> Result<()> {
+        let CommandExecutionRequestApprovalParams {
+            thread_id,
+            turn_id,
+            item_id,
+            reason,
+            risk,
+        } = params;
+
+        println!(
+            "\n< commandExecution approval requested for thread {thread_id}, turn {turn_id}, item {item_id}"
+        );
+        if let Some(reason) = reason.as_deref() {
+            println!("< reason: {reason}");
+        }
+        if let Some(risk) = risk.as_ref() {
+            println!("< risk assessment: {risk:?}");
+        }
+
+        let response = CommandExecutionRequestApprovalResponse {
+            decision: ApprovalDecision::Accept,
+            accept_settings: Some(CommandExecutionRequestAcceptSettings { for_session: false }),
+        };
+        self.send_server_request_response(request_id, &response)?;
+        println!("< approved commandExecution request for item {item_id}");
+        Ok(())
+    }
+
+    fn send_server_request_response<T>(&mut self, request_id: RequestId, response: &T) -> Result<()>
+    where
+        T: Serialize,
+    {
+        let message = JSONRPCMessage::Response(JSONRPCResponse {
+            id: request_id,
+            result: serde_json::to_value(response)?,
+        });
+        self.write_jsonrpc_message(message)
+    }
+
+    fn write_jsonrpc_message(&mut self, message: JSONRPCMessage) -> Result<()> {
+        let payload = serde_json::to_string(&message)?;
+        let pretty = serde_json::to_string_pretty(&message)?;
+        print_multiline_with_prefix("> ", &pretty);
+
+        if let Some(stdin) = self.stdin.as_mut() {
+            writeln!(stdin, "{payload}")?;
+            stdin
+                .flush()
+                .context("failed to flush response to codex app-server")?;
+            return Ok(());
+        }
+
+        bail!("codex app-server stdin closed")
+    }
+}
+
+fn print_multiline_with_prefix(prefix: &str, payload: &str) {
+    for line in payload.lines() {
+        println!("{prefix}{line}");
+    }
+}
+
+impl Drop for CodexClient {
+    fn drop(&mut self) {
+        let _ = self.stdin.take();
+
+        if let Ok(Some(status)) = self.child.try_wait() {
+            println!("[codex app-server exited: {status}]");
+            return;
+        }
+
+        thread::sleep(Duration::from_millis(100));
+
+        if let Ok(Some(status)) = self.child.try_wait() {
+            println!("[codex app-server exited: {status}]");
+            return;
+        }
+
+        let _ = self.child.kill();
+        let _ = self.child.wait();
+    }
+}

codex-rs/app-server/Cargo.toml

@@ -46,6 +46,7 @@ app_test_support = { workspace = true }
 assert_cmd = { workspace = true }
 base64 = { workspace = true }
 core_test_support = { workspace = true }
+mcp-types = { workspace = true }
 os_info = { workspace = true }
 pretty_assertions = { workspace = true }
 serial_test = { workspace = true }

codex-rs/app-server/README.md

@@ -2,6 +2,16 @@
 
 `codex app-server` is the interface Codex uses to power rich interfaces such as the [Codex VS Code extension](https://marketplace.visualstudio.com/items?itemName=openai.chatgpt). The message schema is currently unstable, but those who wish to build experimental UIs on top of Codex may find it valuable.
 
+## Table of Contents
+- [Protocol](#protocol)
+- [Message Schema](#message-schema)
+- [Lifecycle Overview](#lifecycle-overview)
+- [Initialization](#initialization)
+- [Core primitives](#core-primitives)
+- [Thread & turn endpoints](#thread--turn-endpoints)
+- [Auth endpoints](#auth-endpoints)
+- [Events (work-in-progress)](#v2-streaming-events-work-in-progress)
+
 ## Protocol
 
 Similar to [MCP](https://modelcontextprotocol.io/), `codex app-server` supports bidirectional communication, streaming JSONL over stdio. The protocol is JSON-RPC 2.0, though the `"jsonrpc":"2.0"` header is omitted.
@@ -15,6 +25,14 @@ codex app-server generate-ts --out DIR
 codex app-server generate-json-schema --out DIR
 ```
 
+## Lifecycle Overview
+
+- Initialize once: Immediately after launching the codex app-server process, send an `initialize` request with your client metadata, then emit an `initialized` notification. Any other request before this handshake gets rejected.
+- Start (or resume) a thread: Call `thread/start` to open a fresh conversation. The response returns the thread object and you’ll also get a `thread/started` notification. If you’re continuing an existing conversation, call `thread/resume` with its ID instead.
+- Begin a turn: To send user input, call `turn/start` with the target `threadId` and the user's input. Optional fields let you override model, cwd, sandbox policy, etc. This immediately returns the new turn object and triggers a `turn/started` notification.
+- Stream events: After `turn/start`, keep reading JSON-RPC notifications on stdout. You’ll see `item/started`, `item/completed`, deltas like `item/agentMessage/delta`, tool progress, etc. These represent streaming model output plus any side effects (commands, tool calls, reasoning notes).
+- Finish the turn: When the model is done (or the turn is interrupted via making the `turn/interrupt` call), the server sends `turn/completed` with the final turn state and token usage.
+
 ## Initialization
 
 Clients must send a single `initialize` request before invoking any other method, then acknowledge with an `initialized` notification. The server returns the user agent string it will present to upstream services; subsequent requests issued before initialization receive a `"Not initialized"` error, and repeated `initialize` calls receive an `"Already initialized"` error.
@@ -47,6 +65,7 @@ The JSON-RPC API exposes dedicated methods for managing Codex conversations. Thr
 - `thread/archive` — move a thread’s rollout file into the archived directory; returns `{}` on success.
 - `turn/start` — add user input to a thread and begin Codex generation; responds with the initial `turn` object and streams `turn/started`, `item/*`, and `turn/completed` notifications.
 - `turn/interrupt` — request cancellation of an in-flight turn by `(thread_id, turn_id)`; success is an empty `{}` response and the turn finishes with `status: "interrupted"`.
+- `review/start` — kick off Codex’s automated reviewer for a thread; responds like `turn/start` and emits a `item/completed` notification with a `codeReview` item when results are ready.
 
 ### 1) Start or resume a thread
 
@@ -56,7 +75,7 @@ Start a fresh thread when you need a new Codex conversation.
 { "method": "thread/start", "id": 10, "params": {
     // Optionally set config settings. If not specified, will use the user's
     // current config settings.
-    "model": "gpt-5-codex",
+    "model": "gpt-5.1-codex",
     "cwd": "/Users/me/project",
     "approvalPolicy": "never",
     "sandbox": "workspaceWrite",
@@ -137,7 +156,7 @@ You can optionally specify config overrides on the new turn. If specified, these
         "writableRoots": ["/Users/me/project"],
         "networkAccess": true
     },
-    "model": "gpt-5-codex",
+    "model": "gpt-5.1-codex",
     "effort": "medium",
     "summary": "concise"
 } }
@@ -163,6 +182,58 @@ You can cancel a running Turn with `turn/interrupt`.
 
 The server requests cancellations for running subprocesses, then emits a `turn/completed` event with `status: "interrupted"`. Rely on the `turn/completed` to know when Codex-side cleanup is done.
 
+### 6) Request a code review
+
+Use `review/start` to run Codex’s reviewer on the currently checked-out project. The request takes the thread id plus a `target` describing what should be reviewed:
+
+- `{"type":"uncommittedChanges"}` — staged, unstaged, and untracked files.
+- `{"type":"baseBranch","branch":"main"}` — diff against the provided branch’s upstream (see prompt for the exact `git merge-base`/`git diff` instructions Codex will run).
+- `{"type":"commit","sha":"abc1234","title":"Optional subject"}` — review a specific commit.
+- `{"type":"custom","instructions":"Free-form reviewer instructions"}` — fallback prompt equivalent to the legacy manual review request.
+- `appendToOriginalThread` (bool, default `false`) — when `true`, Codex also records a final assistant-style message with the review summary in the original thread. When `false`, only the `codeReview` item is emitted for the review run and no extra message is added to the original thread.
+
+Example request/response:
+
+```json
+{ "method": "review/start", "id": 40, "params": {
+    "threadId": "thr_123",
+    "appendToOriginalThread": true,
+    "target": { "type": "commit", "sha": "1234567deadbeef", "title": "Polish tui colors" }
+} }
+{ "id": 40, "result": { "turn": {
+    "id": "turn_900",
+    "status": "inProgress",
+    "items": [
+        { "type": "userMessage", "id": "turn_900", "content": [ { "type": "text", "text": "Review commit 1234567: Polish tui colors" } ] }
+    ],
+    "error": null
+} } }
+```
+
+Codex streams the usual `turn/started` notification followed by an `item/started`
+with the same `codeReview` item id so clients can show progress:
+
+```json
+{ "method": "item/started", "params": { "item": {
+    "type": "codeReview",
+    "id": "turn_900",
+    "review": "current changes"
+} } }
+```
+
+When the reviewer finishes, the server emits `item/completed` containing the same
+`codeReview` item with the final review text:
+
+```json
+{ "method": "item/completed", "params": { "item": {
+    "type": "codeReview",
+    "id": "turn_900",
+    "review": "Looks solid overall...\n\n- Prefer Stylize helpers — app.rs:10-20\n  ..."
+} } }
+```
+
+The `review` string is plain text that already bundles the overall explanation plus a bullet list for each structured finding (matching `ThreadItem::CodeReview` in the generated schema). Use this notification to render the reviewer output in your client.
+
 ## Auth endpoints
 
 The JSON-RPC auth/account surface exposes request/response methods plus server-initiated notifications (no `id`). Use these to determine auth state, start or cancel logins, logout, and inspect ChatGPT rate limits.
@@ -258,3 +329,33 @@ Field notes:
 - `codex app-server generate-ts --out <dir>` emits v2 types under `v2/`.
 - `codex app-server generate-json-schema --out <dir>` outputs `codex_app_server_protocol.schemas.json`.
 - See [“Authentication and authorization” in the config docs](../../docs/config.md#authentication-and-authorization) for configuration knobs.
+
+
+## Events (work-in-progress)
+
+Event notifications are the server-initiated event stream for thread lifecycles, turn lifecycles, and the items within them. After you start or resume a thread, keep reading stdout for `thread/started`, `turn/*`, and `item/*` notifications.
+
+### Turn events
+
+The app-server streams JSON-RPC notifications while a turn is running. Each turn starts with `turn/started` (initial `turn`) and ends with `turn/completed` (final `turn` plus token `usage`), and clients subscribe to the events they care about, rendering each item incrementally as updates arrive. The per-item lifecycle is always: `item/started` → zero or more item-specific deltas → `item/completed`.
+
+#### Thread items
+
+`ThreadItem` is the tagged union carried in turn responses and `item/*` notifications. Currently we support events for the following items:
+- `userMessage` — `{id, content}` where `content` is a list of user inputs (`text`, `image`, or `localImage`).
+- `agentMessage` — `{id, text}` containing the accumulated agent reply.
+- `reasoning` — `{id, summary, content}` where `summary` holds streamed reasoning summaries (applicable for most OpenAI models) and `content` holds raw reasoning blocks (applicable for e.g. open source models).
+- `mcpToolCall` — `{id, server, tool, status, arguments, result?, error?}` describing MCP calls; `status` is `inProgress`, `completed`, or `failed`.
+- `webSearch` — `{id, query}` for a web search request issued by the agent.
+
+All items emit two shared lifecycle events:
+- `item/started` — emits the full `item` when a new unit of work begins so the UI can render it immediately; the `item.id` in this payload matches the `itemId` used by deltas.
+- `item/completed` — sends the final `item` once that work finishes (e.g., after a tool call or message completes); treat this as the authoritative state.
+
+There are additional item-specific events:
+#### agentMessage
+- `item/agentMessage/delta` — appends streamed text for the agent message; concatenate `delta` values for the same `itemId` in order to reconstruct the full reply.
+#### reasoning
+- `item/reasoning/summaryTextDelta` — streams readable reasoning summaries; `summaryIndex` increments when a new summary section opens.
+- `item/reasoning/summaryPartAdded` — marks the boundary between reasoning summary sections for an `itemId`; subsequent `summaryTextDelta` entries share the same `summaryIndex`.
+- `item/reasoning/textDelta` — streams raw reasoning text (only applicable for e.g. open source models); use `contentIndex` to group deltas that belong together before showing them in the UI.

codex-rs/app-server/src/bespoke_event_handling.rs

@@ -0,0 +1,984 @@
+use crate::codex_message_processor::ApiVersion;
+use crate::codex_message_processor::PendingInterrupts;
+use crate::codex_message_processor::TurnSummary;
+use crate::codex_message_processor::TurnSummaryStore;
+use crate::outgoing_message::OutgoingMessageSender;
+use codex_app_server_protocol::AccountRateLimitsUpdatedNotification;
+use codex_app_server_protocol::AgentMessageDeltaNotification;
+use codex_app_server_protocol::ApplyPatchApprovalParams;
+use codex_app_server_protocol::ApplyPatchApprovalResponse;
+use codex_app_server_protocol::ApprovalDecision;
+use codex_app_server_protocol::CommandAction as V2ParsedCommand;
+use codex_app_server_protocol::CommandExecutionOutputDeltaNotification;
+use codex_app_server_protocol::CommandExecutionRequestApprovalParams;
+use codex_app_server_protocol::CommandExecutionRequestApprovalResponse;
+use codex_app_server_protocol::CommandExecutionStatus;
+use codex_app_server_protocol::ExecCommandApprovalParams;
+use codex_app_server_protocol::ExecCommandApprovalResponse;
+use codex_app_server_protocol::InterruptConversationResponse;
+use codex_app_server_protocol::ItemCompletedNotification;
+use codex_app_server_protocol::ItemStartedNotification;
+use codex_app_server_protocol::McpToolCallError;
+use codex_app_server_protocol::McpToolCallResult;
+use codex_app_server_protocol::McpToolCallStatus;
+use codex_app_server_protocol::ReasoningSummaryPartAddedNotification;
+use codex_app_server_protocol::ReasoningSummaryTextDeltaNotification;
+use codex_app_server_protocol::ReasoningTextDeltaNotification;
+use codex_app_server_protocol::SandboxCommandAssessment as V2SandboxCommandAssessment;
+use codex_app_server_protocol::ServerNotification;
+use codex_app_server_protocol::ServerRequestPayload;
+use codex_app_server_protocol::ThreadItem;
+use codex_app_server_protocol::Turn;
+use codex_app_server_protocol::TurnCompletedNotification;
+use codex_app_server_protocol::TurnError;
+use codex_app_server_protocol::TurnInterruptResponse;
+use codex_app_server_protocol::TurnStatus;
+use codex_core::CodexConversation;
+use codex_core::parse_command::shlex_join;
+use codex_core::protocol::ApplyPatchApprovalRequestEvent;
+use codex_core::protocol::Event;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::ExecApprovalRequestEvent;
+use codex_core::protocol::ExecCommandEndEvent;
+use codex_core::protocol::McpToolCallBeginEvent;
+use codex_core::protocol::McpToolCallEndEvent;
+use codex_core::protocol::Op;
+use codex_core::protocol::ReviewDecision;
+use codex_core::review_format::format_review_findings_block;
+use codex_protocol::ConversationId;
+use codex_protocol::protocol::ReviewOutputEvent;
+use std::convert::TryFrom;
+use std::sync::Arc;
+use tokio::sync::oneshot;
+use tracing::error;
+
+type JsonValue = serde_json::Value;
+
+pub(crate) async fn apply_bespoke_event_handling(
+    event: Event,
+    conversation_id: ConversationId,
+    conversation: Arc<CodexConversation>,
+    outgoing: Arc<OutgoingMessageSender>,
+    pending_interrupts: PendingInterrupts,
+    turn_summary_store: TurnSummaryStore,
+    api_version: ApiVersion,
+) {
+    let Event { id: event_id, msg } = event;
+    match msg {
+        EventMsg::TaskComplete(_ev) => {
+            handle_turn_complete(conversation_id, event_id, &outgoing, &turn_summary_store).await;
+        }
+        EventMsg::ApplyPatchApprovalRequest(ApplyPatchApprovalRequestEvent {
+            call_id,
+            changes,
+            reason,
+            grant_root,
+        }) => {
+            let params = ApplyPatchApprovalParams {
+                conversation_id,
+                call_id,
+                file_changes: changes,
+                reason,
+                grant_root,
+            };
+            let rx = outgoing
+                .send_request(ServerRequestPayload::ApplyPatchApproval(params))
+                .await;
+            tokio::spawn(async move {
+                on_patch_approval_response(event_id, rx, conversation).await;
+            });
+        }
+        EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
+            call_id,
+            turn_id,
+            command,
+            cwd,
+            reason,
+            risk,
+            parsed_cmd,
+        }) => match api_version {
+            ApiVersion::V1 => {
+                let params = ExecCommandApprovalParams {
+                    conversation_id,
+                    call_id,
+                    command,
+                    cwd,
+                    reason,
+                    risk,
+                    parsed_cmd,
+                };
+                let rx = outgoing
+                    .send_request(ServerRequestPayload::ExecCommandApproval(params))
+                    .await;
+                tokio::spawn(async move {
+                    on_exec_approval_response(event_id, rx, conversation).await;
+                });
+            }
+            ApiVersion::V2 => {
+                let params = CommandExecutionRequestApprovalParams {
+                    thread_id: conversation_id.to_string(),
+                    turn_id: turn_id.clone(),
+                    // Until we migrate the core to be aware of a first class CommandExecutionItem
+                    // and emit the corresponding EventMsg, we repurpose the call_id as the item_id.
+                    item_id: call_id.clone(),
+                    reason,
+                    risk: risk.map(V2SandboxCommandAssessment::from),
+                };
+                let rx = outgoing
+                    .send_request(ServerRequestPayload::CommandExecutionRequestApproval(
+                        params,
+                    ))
+                    .await;
+                tokio::spawn(async move {
+                    on_command_execution_request_approval_response(event_id, rx, conversation)
+                        .await;
+                });
+            }
+        },
+        // TODO(celia): properly construct McpToolCall TurnItem in core.
+        EventMsg::McpToolCallBegin(begin_event) => {
+            let notification = construct_mcp_tool_call_notification(begin_event).await;
+            outgoing
+                .send_server_notification(ServerNotification::ItemStarted(notification))
+                .await;
+        }
+        EventMsg::McpToolCallEnd(end_event) => {
+            let notification = construct_mcp_tool_call_end_notification(end_event).await;
+            outgoing
+                .send_server_notification(ServerNotification::ItemCompleted(notification))
+                .await;
+        }
+        EventMsg::AgentMessageContentDelta(event) => {
+            let notification = AgentMessageDeltaNotification {
+                item_id: event.item_id,
+                delta: event.delta,
+            };
+            outgoing
+                .send_server_notification(ServerNotification::AgentMessageDelta(notification))
+                .await;
+        }
+        EventMsg::ReasoningContentDelta(event) => {
+            let notification = ReasoningSummaryTextDeltaNotification {
+                item_id: event.item_id,
+                delta: event.delta,
+                summary_index: event.summary_index,
+            };
+            outgoing
+                .send_server_notification(ServerNotification::ReasoningSummaryTextDelta(
+                    notification,
+                ))
+                .await;
+        }
+        EventMsg::ReasoningRawContentDelta(event) => {
+            let notification = ReasoningTextDeltaNotification {
+                item_id: event.item_id,
+                delta: event.delta,
+                content_index: event.content_index,
+            };
+            outgoing
+                .send_server_notification(ServerNotification::ReasoningTextDelta(notification))
+                .await;
+        }
+        EventMsg::AgentReasoningSectionBreak(event) => {
+            let notification = ReasoningSummaryPartAddedNotification {
+                item_id: event.item_id,
+                summary_index: event.summary_index,
+            };
+            outgoing
+                .send_server_notification(ServerNotification::ReasoningSummaryPartAdded(
+                    notification,
+                ))
+                .await;
+        }
+        EventMsg::TokenCount(token_count_event) => {
+            if let Some(rate_limits) = token_count_event.rate_limits {
+                outgoing
+                    .send_server_notification(ServerNotification::AccountRateLimitsUpdated(
+                        AccountRateLimitsUpdatedNotification {
+                            rate_limits: rate_limits.into(),
+                        },
+                    ))
+                    .await;
+            }
+        }
+        EventMsg::Error(ev) => {
+            handle_error(conversation_id, ev.message, &turn_summary_store).await;
+        }
+        EventMsg::EnteredReviewMode(review_request) => {
+            let notification = ItemStartedNotification {
+                item: ThreadItem::CodeReview {
+                    id: event_id.clone(),
+                    review: review_request.user_facing_hint,
+                },
+            };
+            outgoing
+                .send_server_notification(ServerNotification::ItemStarted(notification))
+                .await;
+        }
+        EventMsg::ItemStarted(item_started_event) => {
+            let item: ThreadItem = item_started_event.item.clone().into();
+            let notification = ItemStartedNotification { item };
+            outgoing
+                .send_server_notification(ServerNotification::ItemStarted(notification))
+                .await;
+        }
+        EventMsg::ItemCompleted(item_completed_event) => {
+            let item: ThreadItem = item_completed_event.item.clone().into();
+            let notification = ItemCompletedNotification { item };
+            outgoing
+                .send_server_notification(ServerNotification::ItemCompleted(notification))
+                .await;
+        }
+        EventMsg::ExitedReviewMode(review_event) => {
+            let review_text = match review_event.review_output {
+                Some(output) => render_review_output_text(&output),
+                None => REVIEW_FALLBACK_MESSAGE.to_string(),
+            };
+            let notification = ItemCompletedNotification {
+                item: ThreadItem::CodeReview {
+                    id: event_id,
+                    review: review_text,
+                },
+            };
+            outgoing
+                .send_server_notification(ServerNotification::ItemCompleted(notification))
+                .await;
+        }
+        EventMsg::ExecCommandBegin(exec_command_begin_event) => {
+            let item = ThreadItem::CommandExecution {
+                id: exec_command_begin_event.call_id.clone(),
+                command: shlex_join(&exec_command_begin_event.command),
+                cwd: exec_command_begin_event.cwd,
+                status: CommandExecutionStatus::InProgress,
+                command_actions: exec_command_begin_event
+                    .parsed_cmd
+                    .into_iter()
+                    .map(V2ParsedCommand::from)
+                    .collect(),
+                aggregated_output: None,
+                exit_code: None,
+                duration_ms: None,
+            };
+            let notification = ItemStartedNotification { item };
+            outgoing
+                .send_server_notification(ServerNotification::ItemStarted(notification))
+                .await;
+        }
+        EventMsg::ExecCommandOutputDelta(exec_command_output_delta_event) => {
+            let notification = CommandExecutionOutputDeltaNotification {
+                item_id: exec_command_output_delta_event.call_id.clone(),
+                delta: String::from_utf8_lossy(&exec_command_output_delta_event.chunk).to_string(),
+            };
+            outgoing
+                .send_server_notification(ServerNotification::CommandExecutionOutputDelta(
+                    notification,
+                ))
+                .await;
+        }
+        EventMsg::ExecCommandEnd(exec_command_end_event) => {
+            let ExecCommandEndEvent {
+                call_id,
+                command,
+                cwd,
+                parsed_cmd,
+                aggregated_output,
+                exit_code,
+                duration,
+                ..
+            } = exec_command_end_event;
+
+            let status = if exit_code == 0 {
+                CommandExecutionStatus::Completed
+            } else {
+                CommandExecutionStatus::Failed
+            };
+
+            let aggregated_output = if aggregated_output.is_empty() {
+                None
+            } else {
+                Some(aggregated_output)
+            };
+
+            let duration_ms = i64::try_from(duration.as_millis()).unwrap_or(i64::MAX);
+
+            let item = ThreadItem::CommandExecution {
+                id: call_id,
+                command: shlex_join(&command),
+                cwd,
+                status,
+                command_actions: parsed_cmd.into_iter().map(V2ParsedCommand::from).collect(),
+                aggregated_output,
+                exit_code: Some(exit_code),
+                duration_ms: Some(duration_ms),
+            };
+
+            let notification = ItemCompletedNotification { item };
+            outgoing
+                .send_server_notification(ServerNotification::ItemCompleted(notification))
+                .await;
+        }
+        // If this is a TurnAborted, reply to any pending interrupt requests.
+        EventMsg::TurnAborted(turn_aborted_event) => {
+            let pending = {
+                let mut map = pending_interrupts.lock().await;
+                map.remove(&conversation_id).unwrap_or_default()
+            };
+            if !pending.is_empty() {
+                for (rid, ver) in pending {
+                    match ver {
+                        ApiVersion::V1 => {
+                            let response = InterruptConversationResponse {
+                                abort_reason: turn_aborted_event.reason.clone(),
+                            };
+                            outgoing.send_response(rid, response).await;
+                        }
+                        ApiVersion::V2 => {
+                            let response = TurnInterruptResponse {};
+                            outgoing.send_response(rid, response).await;
+                        }
+                    }
+                }
+            }
+
+            handle_turn_interrupted(conversation_id, event_id, &outgoing, &turn_summary_store)
+                .await;
+        }
+
+        _ => {}
+    }
+}
+
+async fn emit_turn_completed_with_status(
+    event_id: String,
+    status: TurnStatus,
+    outgoing: &OutgoingMessageSender,
+) {
+    let notification = TurnCompletedNotification {
+        turn: Turn {
+            id: event_id,
+            items: vec![],
+            status,
+        },
+    };
+    outgoing
+        .send_server_notification(ServerNotification::TurnCompleted(notification))
+        .await;
+}
+
+async fn find_and_remove_turn_summary(
+    conversation_id: ConversationId,
+    turn_summary_store: &TurnSummaryStore,
+) -> TurnSummary {
+    let mut map = turn_summary_store.lock().await;
+    map.remove(&conversation_id).unwrap_or_default()
+}
+
+async fn handle_turn_complete(
+    conversation_id: ConversationId,
+    event_id: String,
+    outgoing: &OutgoingMessageSender,
+    turn_summary_store: &TurnSummaryStore,
+) {
+    let turn_summary = find_and_remove_turn_summary(conversation_id, turn_summary_store).await;
+
+    let status = if let Some(message) = turn_summary.last_error_message {
+        TurnStatus::Failed {
+            error: TurnError { message },
+        }
+    } else {
+        TurnStatus::Completed
+    };
+
+    emit_turn_completed_with_status(event_id, status, outgoing).await;
+}
+
+async fn handle_turn_interrupted(
+    conversation_id: ConversationId,
+    event_id: String,
+    outgoing: &OutgoingMessageSender,
+    turn_summary_store: &TurnSummaryStore,
+) {
+    find_and_remove_turn_summary(conversation_id, turn_summary_store).await;
+
+    emit_turn_completed_with_status(event_id, TurnStatus::Interrupted, outgoing).await;
+}
+
+async fn handle_error(
+    conversation_id: ConversationId,
+    message: String,
+    turn_summary_store: &TurnSummaryStore,
+) {
+    let mut map = turn_summary_store.lock().await;
+    map.entry(conversation_id).or_default().last_error_message = Some(message);
+}
+
+async fn on_patch_approval_response(
+    event_id: String,
+    receiver: oneshot::Receiver<JsonValue>,
+    codex: Arc<CodexConversation>,
+) {
+    let response = receiver.await;
+    let value = match response {
+        Ok(value) => value,
+        Err(err) => {
+            error!("request failed: {err:?}");
+            if let Err(submit_err) = codex
+                .submit(Op::PatchApproval {
+                    id: event_id.clone(),
+                    decision: ReviewDecision::Denied,
+                })
+                .await
+            {
+                error!("failed to submit denied PatchApproval after request failure: {submit_err}");
+            }
+            return;
+        }
+    };
+
+    let response =
+        serde_json::from_value::<ApplyPatchApprovalResponse>(value).unwrap_or_else(|err| {
+            error!("failed to deserialize ApplyPatchApprovalResponse: {err}");
+            ApplyPatchApprovalResponse {
+                decision: ReviewDecision::Denied,
+            }
+        });
+
+    if let Err(err) = codex
+        .submit(Op::PatchApproval {
+            id: event_id,
+            decision: response.decision,
+        })
+        .await
+    {
+        error!("failed to submit PatchApproval: {err}");
+    }
+}
+
+async fn on_exec_approval_response(
+    event_id: String,
+    receiver: oneshot::Receiver<JsonValue>,
+    conversation: Arc<CodexConversation>,
+) {
+    let response = receiver.await;
+    let value = match response {
+        Ok(value) => value,
+        Err(err) => {
+            error!("request failed: {err:?}");
+            return;
+        }
+    };
+
+    // Try to deserialize `value` and then make the appropriate call to `codex`.
+    let response =
+        serde_json::from_value::<ExecCommandApprovalResponse>(value).unwrap_or_else(|err| {
+            error!("failed to deserialize ExecCommandApprovalResponse: {err}");
+            // If we cannot deserialize the response, we deny the request to be
+            // conservative.
+            ExecCommandApprovalResponse {
+                decision: ReviewDecision::Denied,
+            }
+        });
+
+    if let Err(err) = conversation
+        .submit(Op::ExecApproval {
+            id: event_id,
+            decision: response.decision,
+        })
+        .await
+    {
+        error!("failed to submit ExecApproval: {err}");
+    }
+}
+
+const REVIEW_FALLBACK_MESSAGE: &str = "Reviewer failed to output a response.";
+
+fn render_review_output_text(output: &ReviewOutputEvent) -> String {
+    let mut sections = Vec::new();
+    let explanation = output.overall_explanation.trim();
+    if !explanation.is_empty() {
+        sections.push(explanation.to_string());
+    }
+    if !output.findings.is_empty() {
+        let findings = format_review_findings_block(&output.findings, None);
+        let trimmed = findings.trim();
+        if !trimmed.is_empty() {
+            sections.push(trimmed.to_string());
+        }
+    }
+    if sections.is_empty() {
+        REVIEW_FALLBACK_MESSAGE.to_string()
+    } else {
+        sections.join("\n\n")
+    }
+}
+
+async fn on_command_execution_request_approval_response(
+    event_id: String,
+    receiver: oneshot::Receiver<JsonValue>,
+    conversation: Arc<CodexConversation>,
+) {
+    let response = receiver.await;
+    let value = match response {
+        Ok(value) => value,
+        Err(err) => {
+            error!("request failed: {err:?}");
+            return;
+        }
+    };
+
+    let response = serde_json::from_value::<CommandExecutionRequestApprovalResponse>(value)
+        .unwrap_or_else(|err| {
+            error!("failed to deserialize CommandExecutionRequestApprovalResponse: {err}");
+            CommandExecutionRequestApprovalResponse {
+                decision: ApprovalDecision::Decline,
+                accept_settings: None,
+            }
+        });
+
+    let CommandExecutionRequestApprovalResponse {
+        decision,
+        accept_settings,
+    } = response;
+
+    let decision = match (decision, accept_settings) {
+        (ApprovalDecision::Accept, Some(settings)) if settings.for_session => {
+            ReviewDecision::ApprovedForSession
+        }
+        (ApprovalDecision::Accept, _) => ReviewDecision::Approved,
+        (ApprovalDecision::Decline, _) => ReviewDecision::Denied,
+        (ApprovalDecision::Cancel, _) => ReviewDecision::Abort,
+    };
+    if let Err(err) = conversation
+        .submit(Op::ExecApproval {
+            id: event_id,
+            decision,
+        })
+        .await
+    {
+        error!("failed to submit ExecApproval: {err}");
+    }
+}
+
+/// similar to handle_mcp_tool_call_begin in exec
+async fn construct_mcp_tool_call_notification(
+    begin_event: McpToolCallBeginEvent,
+) -> ItemStartedNotification {
+    let item = ThreadItem::McpToolCall {
+        id: begin_event.call_id,
+        server: begin_event.invocation.server,
+        tool: begin_event.invocation.tool,
+        status: McpToolCallStatus::InProgress,
+        arguments: begin_event.invocation.arguments.unwrap_or(JsonValue::Null),
+        result: None,
+        error: None,
+    };
+    ItemStartedNotification { item }
+}
+
+/// simiilar to handle_mcp_tool_call_end in exec
+async fn construct_mcp_tool_call_end_notification(
+    end_event: McpToolCallEndEvent,
+) -> ItemCompletedNotification {
+    let status = if end_event.is_success() {
+        McpToolCallStatus::Completed
+    } else {
+        McpToolCallStatus::Failed
+    };
+
+    let (result, error) = match &end_event.result {
+        Ok(value) => (
+            Some(McpToolCallResult {
+                content: value.content.clone(),
+                structured_content: value.structured_content.clone(),
+            }),
+            None,
+        ),
+        Err(message) => (
+            None,
+            Some(McpToolCallError {
+                message: message.clone(),
+            }),
+        ),
+    };
+
+    let item = ThreadItem::McpToolCall {
+        id: end_event.call_id,
+        server: end_event.invocation.server,
+        tool: end_event.invocation.tool,
+        status,
+        arguments: end_event.invocation.arguments.unwrap_or(JsonValue::Null),
+        result,
+        error,
+    };
+    ItemCompletedNotification { item }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::CHANNEL_CAPACITY;
+    use crate::outgoing_message::OutgoingMessage;
+    use crate::outgoing_message::OutgoingMessageSender;
+    use anyhow::Result;
+    use anyhow::anyhow;
+    use anyhow::bail;
+    use codex_core::protocol::McpInvocation;
+    use mcp_types::CallToolResult;
+    use mcp_types::ContentBlock;
+    use mcp_types::TextContent;
+    use pretty_assertions::assert_eq;
+    use serde_json::Value as JsonValue;
+    use std::collections::HashMap;
+    use std::time::Duration;
+    use tokio::sync::Mutex;
+    use tokio::sync::mpsc;
+
+    fn new_turn_summary_store() -> TurnSummaryStore {
+        Arc::new(Mutex::new(HashMap::new()))
+    }
+
+    #[tokio::test]
+    async fn test_handle_error_records_message() -> Result<()> {
+        let conversation_id = ConversationId::new();
+        let turn_summary_store = new_turn_summary_store();
+
+        handle_error(conversation_id, "boom".to_string(), &turn_summary_store).await;
+
+        let turn_summary = find_and_remove_turn_summary(conversation_id, &turn_summary_store).await;
+        assert_eq!(turn_summary.last_error_message, Some("boom".to_string()));
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_handle_turn_complete_emits_completed_without_error() -> Result<()> {
+        let conversation_id = ConversationId::new();
+        let event_id = "complete1".to_string();
+        let (tx, mut rx) = mpsc::channel(CHANNEL_CAPACITY);
+        let outgoing = Arc::new(OutgoingMessageSender::new(tx));
+        let turn_summary_store = new_turn_summary_store();
+
+        handle_turn_complete(
+            conversation_id,
+            event_id.clone(),
+            &outgoing,
+            &turn_summary_store,
+        )
+        .await;
+
+        let msg = rx
+            .recv()
+            .await
+            .ok_or_else(|| anyhow!("should send one notification"))?;
+        match msg {
+            OutgoingMessage::AppServerNotification(ServerNotification::TurnCompleted(n)) => {
+                assert_eq!(n.turn.id, event_id);
+                assert_eq!(n.turn.status, TurnStatus::Completed);
+            }
+            other => bail!("unexpected message: {other:?}"),
+        }
+        assert!(rx.try_recv().is_err(), "no extra messages expected");
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_handle_turn_interrupted_emits_interrupted_with_error() -> Result<()> {
+        let conversation_id = ConversationId::new();
+        let event_id = "interrupt1".to_string();
+        let turn_summary_store = new_turn_summary_store();
+        handle_error(conversation_id, "oops".to_string(), &turn_summary_store).await;
+        let (tx, mut rx) = mpsc::channel(CHANNEL_CAPACITY);
+        let outgoing = Arc::new(OutgoingMessageSender::new(tx));
+
+        handle_turn_interrupted(
+            conversation_id,
+            event_id.clone(),
+            &outgoing,
+            &turn_summary_store,
+        )
+        .await;
+
+        let msg = rx
+            .recv()
+            .await
+            .ok_or_else(|| anyhow!("should send one notification"))?;
+        match msg {
+            OutgoingMessage::AppServerNotification(ServerNotification::TurnCompleted(n)) => {
+                assert_eq!(n.turn.id, event_id);
+                assert_eq!(n.turn.status, TurnStatus::Interrupted);
+            }
+            other => bail!("unexpected message: {other:?}"),
+        }
+        assert!(rx.try_recv().is_err(), "no extra messages expected");
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_handle_turn_complete_emits_failed_with_error() -> Result<()> {
+        let conversation_id = ConversationId::new();
+        let event_id = "complete_err1".to_string();
+        let turn_summary_store = new_turn_summary_store();
+        handle_error(conversation_id, "bad".to_string(), &turn_summary_store).await;
+        let (tx, mut rx) = mpsc::channel(CHANNEL_CAPACITY);
+        let outgoing = Arc::new(OutgoingMessageSender::new(tx));
+
+        handle_turn_complete(
+            conversation_id,
+            event_id.clone(),
+            &outgoing,
+            &turn_summary_store,
+        )
+        .await;
+
+        let msg = rx
+            .recv()
+            .await
+            .ok_or_else(|| anyhow!("should send one notification"))?;
+        match msg {
+            OutgoingMessage::AppServerNotification(ServerNotification::TurnCompleted(n)) => {
+                assert_eq!(n.turn.id, event_id);
+                assert_eq!(
+                    n.turn.status,
+                    TurnStatus::Failed {
+                        error: TurnError {
+                            message: "bad".to_string(),
+                        }
+                    }
+                );
+            }
+            other => bail!("unexpected message: {other:?}"),
+        }
+        assert!(rx.try_recv().is_err(), "no extra messages expected");
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_construct_mcp_tool_call_begin_notification_with_args() {
+        let begin_event = McpToolCallBeginEvent {
+            call_id: "call_123".to_string(),
+            invocation: McpInvocation {
+                server: "codex".to_string(),
+                tool: "list_mcp_resources".to_string(),
+                arguments: Some(serde_json::json!({"server": ""})),
+            },
+        };
+
+        let notification = construct_mcp_tool_call_notification(begin_event.clone()).await;
+
+        let expected = ItemStartedNotification {
+            item: ThreadItem::McpToolCall {
+                id: begin_event.call_id,
+                server: begin_event.invocation.server,
+                tool: begin_event.invocation.tool,
+                status: McpToolCallStatus::InProgress,
+                arguments: serde_json::json!({"server": ""}),
+                result: None,
+                error: None,
+            },
+        };
+
+        assert_eq!(notification, expected);
+    }
+
+    #[tokio::test]
+    async fn test_handle_turn_complete_emits_error_multiple_turns() -> Result<()> {
+        // Conversation A will have two turns; Conversation B will have one turn.
+        let conversation_a = ConversationId::new();
+        let conversation_b = ConversationId::new();
+        let turn_summary_store = new_turn_summary_store();
+
+        let (tx, mut rx) = mpsc::channel(CHANNEL_CAPACITY);
+        let outgoing = Arc::new(OutgoingMessageSender::new(tx));
+
+        // Turn 1 on conversation A
+        let a_turn1 = "a_turn1".to_string();
+        handle_error(conversation_a, "a1".to_string(), &turn_summary_store).await;
+        handle_turn_complete(
+            conversation_a,
+            a_turn1.clone(),
+            &outgoing,
+            &turn_summary_store,
+        )
+        .await;
+
+        // Turn 1 on conversation B
+        let b_turn1 = "b_turn1".to_string();
+        handle_error(conversation_b, "b1".to_string(), &turn_summary_store).await;
+        handle_turn_complete(
+            conversation_b,
+            b_turn1.clone(),
+            &outgoing,
+            &turn_summary_store,
+        )
+        .await;
+
+        // Turn 2 on conversation A
+        let a_turn2 = "a_turn2".to_string();
+        handle_turn_complete(
+            conversation_a,
+            a_turn2.clone(),
+            &outgoing,
+            &turn_summary_store,
+        )
+        .await;
+
+        // Verify: A turn 1
+        let msg = rx
+            .recv()
+            .await
+            .ok_or_else(|| anyhow!("should send first notification"))?;
+        match msg {
+            OutgoingMessage::AppServerNotification(ServerNotification::TurnCompleted(n)) => {
+                assert_eq!(n.turn.id, a_turn1);
+                assert_eq!(
+                    n.turn.status,
+                    TurnStatus::Failed {
+                        error: TurnError {
+                            message: "a1".to_string(),
+                        }
+                    }
+                );
+            }
+            other => bail!("unexpected message: {other:?}"),
+        }
+
+        // Verify: B turn 1
+        let msg = rx
+            .recv()
+            .await
+            .ok_or_else(|| anyhow!("should send second notification"))?;
+        match msg {
+            OutgoingMessage::AppServerNotification(ServerNotification::TurnCompleted(n)) => {
+                assert_eq!(n.turn.id, b_turn1);
+                assert_eq!(
+                    n.turn.status,
+                    TurnStatus::Failed {
+                        error: TurnError {
+                            message: "b1".to_string(),
+                        }
+                    }
+                );
+            }
+            other => bail!("unexpected message: {other:?}"),
+        }
+
+        // Verify: A turn 2
+        let msg = rx
+            .recv()
+            .await
+            .ok_or_else(|| anyhow!("should send third notification"))?;
+        match msg {
+            OutgoingMessage::AppServerNotification(ServerNotification::TurnCompleted(n)) => {
+                assert_eq!(n.turn.id, a_turn2);
+                assert_eq!(n.turn.status, TurnStatus::Completed);
+            }
+            other => bail!("unexpected message: {other:?}"),
+        }
+
+        assert!(rx.try_recv().is_err(), "no extra messages expected");
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_construct_mcp_tool_call_begin_notification_without_args() {
+        let begin_event = McpToolCallBeginEvent {
+            call_id: "call_456".to_string(),
+            invocation: McpInvocation {
+                server: "codex".to_string(),
+                tool: "list_mcp_resources".to_string(),
+                arguments: None,
+            },
+        };
+
+        let notification = construct_mcp_tool_call_notification(begin_event.clone()).await;
+
+        let expected = ItemStartedNotification {
+            item: ThreadItem::McpToolCall {
+                id: begin_event.call_id,
+                server: begin_event.invocation.server,
+                tool: begin_event.invocation.tool,
+                status: McpToolCallStatus::InProgress,
+                arguments: JsonValue::Null,
+                result: None,
+                error: None,
+            },
+        };
+
+        assert_eq!(notification, expected);
+    }
+
+    #[tokio::test]
+    async fn test_construct_mcp_tool_call_end_notification_success() {
+        let content = vec![ContentBlock::TextContent(TextContent {
+            annotations: None,
+            text: "{\"resources\":[]}".to_string(),
+            r#type: "text".to_string(),
+        })];
+        let result = CallToolResult {
+            content: content.clone(),
+            is_error: Some(false),
+            structured_content: None,
+        };
+
+        let end_event = McpToolCallEndEvent {
+            call_id: "call_789".to_string(),
+            invocation: McpInvocation {
+                server: "codex".to_string(),
+                tool: "list_mcp_resources".to_string(),
+                arguments: Some(serde_json::json!({"server": ""})),
+            },
+            duration: Duration::from_nanos(92708),
+            result: Ok(result),
+        };
+
+        let notification = construct_mcp_tool_call_end_notification(end_event.clone()).await;
+
+        let expected = ItemCompletedNotification {
+            item: ThreadItem::McpToolCall {
+                id: end_event.call_id,
+                server: end_event.invocation.server,
+                tool: end_event.invocation.tool,
+                status: McpToolCallStatus::Completed,
+                arguments: serde_json::json!({"server": ""}),
+                result: Some(McpToolCallResult {
+                    content,
+                    structured_content: None,
+                }),
+                error: None,
+            },
+        };
+
+        assert_eq!(notification, expected);
+    }
+
+    #[tokio::test]
+    async fn test_construct_mcp_tool_call_end_notification_error() {
+        let end_event = McpToolCallEndEvent {
+            call_id: "call_err".to_string(),
+            invocation: McpInvocation {
+                server: "codex".to_string(),
+                tool: "list_mcp_resources".to_string(),
+                arguments: None,
+            },
+            duration: Duration::from_millis(1),
+            result: Err("boom".to_string()),
+        };
+
+        let notification = construct_mcp_tool_call_end_notification(end_event.clone()).await;
+
+        let expected = ItemCompletedNotification {
+            item: ThreadItem::McpToolCall {
+                id: end_event.call_id,
+                server: end_event.invocation.server,
+                tool: end_event.invocation.tool,
+                status: McpToolCallStatus::Failed,
+                arguments: JsonValue::Null,
+                result: None,
+                error: Some(McpToolCallError {
+                    message: "boom".to_string(),
+                }),
+            },
+        };
+
+        assert_eq!(notification, expected);
+    }
+}

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

@@ -1,3 +1,4 @@
+use crate::bespoke_event_handling::apply_bespoke_event_handling;
 use crate::error_code::INTERNAL_ERROR_CODE;
 use crate::error_code::INVALID_REQUEST_ERROR_CODE;
 use crate::fuzzy_file_search::run_fuzzy_file_search;
@@ -8,13 +9,9 @@ use chrono::DateTime;
 use chrono::Utc;
 use codex_app_server_protocol::Account;
 use codex_app_server_protocol::AccountLoginCompletedNotification;
-use codex_app_server_protocol::AccountRateLimitsUpdatedNotification;
 use codex_app_server_protocol::AccountUpdatedNotification;
 use codex_app_server_protocol::AddConversationListenerParams;
 use codex_app_server_protocol::AddConversationSubscriptionResponse;
-use codex_app_server_protocol::AgentMessageDeltaNotification;
-use codex_app_server_protocol::ApplyPatchApprovalParams;
-use codex_app_server_protocol::ApplyPatchApprovalResponse;
 use codex_app_server_protocol::ArchiveConversationParams;
 use codex_app_server_protocol::ArchiveConversationResponse;
 use codex_app_server_protocol::AskForApproval;
@@ -26,8 +23,6 @@ use codex_app_server_protocol::CancelLoginChatGptResponse;
 use codex_app_server_protocol::ClientRequest;
 use codex_app_server_protocol::ConversationGitInfo;
 use codex_app_server_protocol::ConversationSummary;
-use codex_app_server_protocol::ExecCommandApprovalParams;
-use codex_app_server_protocol::ExecCommandApprovalResponse;
 use codex_app_server_protocol::ExecOneOffCommandParams;
 use codex_app_server_protocol::ExecOneOffCommandResponse;
 use codex_app_server_protocol::FeedbackUploadParams;
@@ -46,9 +41,6 @@ use codex_app_server_protocol::GetUserSavedConfigResponse;
 use codex_app_server_protocol::GitDiffToRemoteResponse;
 use codex_app_server_protocol::InputItem as WireInputItem;
 use codex_app_server_protocol::InterruptConversationParams;
-use codex_app_server_protocol::InterruptConversationResponse;
-use codex_app_server_protocol::ItemCompletedNotification;
-use codex_app_server_protocol::ItemStartedNotification;
 use codex_app_server_protocol::JSONRPCErrorError;
 use codex_app_server_protocol::ListConversationsParams;
 use codex_app_server_protocol::ListConversationsResponse;
@@ -63,22 +55,19 @@ use codex_app_server_protocol::ModelListParams;
 use codex_app_server_protocol::ModelListResponse;
 use codex_app_server_protocol::NewConversationParams;
 use codex_app_server_protocol::NewConversationResponse;
-use codex_app_server_protocol::ReasoningSummaryPartAddedNotification;
-use codex_app_server_protocol::ReasoningSummaryTextDeltaNotification;
-use codex_app_server_protocol::ReasoningTextDeltaNotification;
 use codex_app_server_protocol::RemoveConversationListenerParams;
 use codex_app_server_protocol::RemoveConversationSubscriptionResponse;
 use codex_app_server_protocol::RequestId;
-use codex_app_server_protocol::Result as JsonRpcResult;
 use codex_app_server_protocol::ResumeConversationParams;
 use codex_app_server_protocol::ResumeConversationResponse;
+use codex_app_server_protocol::ReviewStartParams;
+use codex_app_server_protocol::ReviewTarget;
 use codex_app_server_protocol::SandboxMode;
 use codex_app_server_protocol::SendUserMessageParams;
 use codex_app_server_protocol::SendUserMessageResponse;
 use codex_app_server_protocol::SendUserTurnParams;
 use codex_app_server_protocol::SendUserTurnResponse;
 use codex_app_server_protocol::ServerNotification;
-use codex_app_server_protocol::ServerRequestPayload;
 use codex_app_server_protocol::SessionConfiguredNotification;
 use codex_app_server_protocol::SetDefaultModelParams;
 use codex_app_server_protocol::SetDefaultModelResponse;
@@ -95,7 +84,6 @@ use codex_app_server_protocol::ThreadStartResponse;
 use codex_app_server_protocol::ThreadStartedNotification;
 use codex_app_server_protocol::Turn;
 use codex_app_server_protocol::TurnInterruptParams;
-use codex_app_server_protocol::TurnInterruptResponse;
 use codex_app_server_protocol::TurnStartParams;
 use codex_app_server_protocol::TurnStartResponse;
 use codex_app_server_protocol::TurnStartedNotification;
@@ -127,12 +115,10 @@ use codex_core::find_conversation_path_by_id_str;
 use codex_core::get_platform_sandbox;
 use codex_core::git_info::git_diff_to_remote;
 use codex_core::parse_cursor;
-use codex_core::protocol::ApplyPatchApprovalRequestEvent;
-use codex_core::protocol::Event;
 use codex_core::protocol::EventMsg;
-use codex_core::protocol::ExecApprovalRequestEvent;
 use codex_core::protocol::Op;
-use codex_core::protocol::ReviewDecision;
+use codex_core::protocol::ReviewRequest;
+use codex_core::protocol::SessionConfiguredEvent;
 use codex_core::read_head_for_summary;
 use codex_feedback::CodexFeedback;
 use codex_login::ServerOptions as LoginServerOptions;
@@ -167,7 +153,15 @@ use tracing::warn;
 use uuid::Uuid;
 
 type PendingInterruptQueue = Vec<(RequestId, ApiVersion)>;
-type PendingInterrupts = Arc<Mutex<HashMap<ConversationId, PendingInterruptQueue>>>;
+pub(crate) type PendingInterrupts = Arc<Mutex<HashMap<ConversationId, PendingInterruptQueue>>>;
+
+/// Per-conversation accumulation of the latest states e.g. error message while a turn runs.
+#[derive(Default, Clone)]
+pub(crate) struct TurnSummary {
+    pub(crate) last_error_message: Option<String>,
+}
+
+pub(crate) type TurnSummaryStore = Arc<Mutex<HashMap<ConversationId, TurnSummary>>>;
 
 // Duration before a ChatGPT login attempt is abandoned.
 const LOGIN_CHATGPT_TIMEOUT: Duration = Duration::from_secs(10 * 60);
@@ -176,8 +170,8 @@ struct ActiveLogin {
     login_id: Uuid,
 }
 
-impl ActiveLogin {
-    fn drop(&self) {
+impl Drop for ActiveLogin {
+    fn drop(&mut self) {
         self.shutdown_handle.shutdown();
     }
 }
@@ -193,12 +187,13 @@ pub(crate) struct CodexMessageProcessor {
     active_login: Arc<Mutex<Option<ActiveLogin>>>,
     // Queue of pending interrupt requests per conversation. We reply when TurnAborted arrives.
     pending_interrupts: PendingInterrupts,
+    turn_summary_store: TurnSummaryStore,
     pending_fuzzy_searches: Arc<Mutex<HashMap<String, Arc<AtomicBool>>>>,
     feedback: CodexFeedback,
 }
 
 #[derive(Clone, Copy, Debug)]
-enum ApiVersion {
+pub(crate) enum ApiVersion {
     V1,
     V2,
 }
@@ -245,11 +240,97 @@ impl CodexMessageProcessor {
             conversation_listeners: HashMap::new(),
             active_login: Arc::new(Mutex::new(None)),
             pending_interrupts: Arc::new(Mutex::new(HashMap::new())),
+            turn_summary_store: Arc::new(Mutex::new(HashMap::new())),
             pending_fuzzy_searches: Arc::new(Mutex::new(HashMap::new())),
             feedback,
         }
     }
 
+    fn review_request_from_target(
+        target: ReviewTarget,
+        append_to_original_thread: bool,
+    ) -> Result<(ReviewRequest, String), JSONRPCErrorError> {
+        fn invalid_request(message: String) -> JSONRPCErrorError {
+            JSONRPCErrorError {
+                code: INVALID_REQUEST_ERROR_CODE,
+                message,
+                data: None,
+            }
+        }
+
+        match target {
+            // TODO(jif) those messages will be extracted in a follow-up PR.
+            ReviewTarget::UncommittedChanges => Ok((
+                ReviewRequest {
+                    prompt: "Review the current code changes (staged, unstaged, and untracked files) and provide prioritized findings.".to_string(),
+                    user_facing_hint: "current changes".to_string(),
+                    append_to_original_thread,
+                },
+                "Review uncommitted changes".to_string(),
+            )),
+            ReviewTarget::BaseBranch { branch } => {
+                let branch = branch.trim().to_string();
+                if branch.is_empty() {
+                    return Err(invalid_request("branch must not be empty".to_string()));
+                }
+                let prompt = format!("Review the code changes against the base branch '{branch}'. Start by finding the merge diff between the current branch and {branch}'s upstream e.g. (`git merge-base HEAD \"$(git rev-parse --abbrev-ref \"{branch}@{{upstream}}\")\"`), then run `git diff` against that SHA to see what changes we would merge into the {branch} branch. Provide prioritized, actionable findings.");
+                let hint = format!("changes against '{branch}'");
+                let display = format!("Review changes against base branch '{branch}'");
+                Ok((
+                    ReviewRequest {
+                        prompt,
+                        user_facing_hint: hint,
+                        append_to_original_thread,
+                    },
+                    display,
+                ))
+            }
+            ReviewTarget::Commit { sha, title } => {
+                let sha = sha.trim().to_string();
+                if sha.is_empty() {
+                    return Err(invalid_request("sha must not be empty".to_string()));
+                }
+                let brief_title = title
+                    .map(|t| t.trim().to_string())
+                    .filter(|t| !t.is_empty());
+                let prompt = if let Some(title) = brief_title.clone() {
+                    format!("Review the code changes introduced by commit {sha} (\"{title}\"). Provide prioritized, actionable findings.")
+                } else {
+                    format!("Review the code changes introduced by commit {sha}. Provide prioritized, actionable findings.")
+                };
+                let short_sha = sha.chars().take(7).collect::<String>();
+                let hint = format!("commit {short_sha}");
+                let display = if let Some(title) = brief_title {
+                    format!("Review commit {short_sha}: {title}")
+                } else {
+                    format!("Review commit {short_sha}")
+                };
+                Ok((
+                    ReviewRequest {
+                        prompt,
+                        user_facing_hint: hint,
+                        append_to_original_thread,
+                    },
+                    display,
+                ))
+            }
+            ReviewTarget::Custom { instructions } => {
+                let trimmed = instructions.trim().to_string();
+                if trimmed.is_empty() {
+                    return Err(invalid_request("instructions must not be empty".to_string()));
+                }
+                Ok((
+                    ReviewRequest {
+                        prompt: trimmed.clone(),
+                        user_facing_hint: trimmed.clone(),
+                        append_to_original_thread,
+                    },
+                    trimmed,
+                ))
+            }
+        }
+    }
+
     pub async fn process_request(&mut self, request: ClientRequest) {
         match request {
             ClientRequest::Initialize { .. } => {
@@ -281,6 +362,9 @@ impl CodexMessageProcessor {
             ClientRequest::TurnInterrupt { request_id, params } => {
                 self.turn_interrupt(request_id, params).await;
             }
+            ClientRequest::ReviewStart { request_id, params } => {
+                self.review_start(request_id, params).await;
+            }
             ClientRequest::NewConversation { request_id, params } => {
                 // Do not tokio::spawn() to process new_conversation()
                 // asynchronously because we need to ensure the conversation is
@@ -435,7 +519,7 @@ impl CodexMessageProcessor {
         {
             let mut guard = self.active_login.lock().await;
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
         }
 
@@ -543,7 +627,7 @@ impl CodexMessageProcessor {
                     {
                         let mut guard = self.active_login.lock().await;
                         if let Some(existing) = guard.take() {
-                            existing.drop();
+                            drop(existing);
                         }
                         *guard = Some(ActiveLogin {
                             shutdown_handle: shutdown_handle.clone(),
@@ -633,7 +717,7 @@ impl CodexMessageProcessor {
                     {
                         let mut guard = self.active_login.lock().await;
                         if let Some(existing) = guard.take() {
-                            existing.drop();
+                            drop(existing);
                         }
                         *guard = Some(ActiveLogin {
                             shutdown_handle: shutdown_handle.clone(),
@@ -722,7 +806,7 @@ impl CodexMessageProcessor {
         let mut guard = self.active_login.lock().await;
         if guard.as_ref().map(|l| l.login_id) == Some(login_id) {
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
             Ok(())
         } else {
@@ -776,7 +860,7 @@ impl CodexMessageProcessor {
         {
             let mut guard = self.active_login.lock().await;
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
         }
 
@@ -1230,8 +1314,12 @@ impl CodexMessageProcessor {
 
         match self.conversation_manager.new_conversation(config).await {
             Ok(new_conv) => {
-                let conversation_id = new_conv.conversation_id;
-                let rollout_path = new_conv.session_configured.rollout_path.clone();
+                let NewConversation {
+                    conversation_id,
+                    session_configured,
+                    ..
+                } = new_conv;
+                let rollout_path = session_configured.rollout_path.clone();
                 let fallback_provider = self.config.model_provider_id.as_str();
 
                 // A bit hacky, but the summary contains a lot of useful information for the thread
@@ -1256,14 +1344,28 @@ impl CodexMessageProcessor {
                     }
                 };
 
+                let SessionConfiguredEvent {
+                    model,
+                    model_provider_id,
+                    cwd,
+                    approval_policy,
+                    sandbox_policy,
+                    ..
+                } = session_configured;
                 let response = ThreadStartResponse {
                     thread: thread.clone(),
+                    model,
+                    model_provider: model_provider_id,
+                    cwd,
+                    approval_policy: approval_policy.into(),
+                    sandbox: sandbox_policy.into(),
+                    reasoning_effort: session_configured.reasoning_effort,
                 };
 
                 // Auto-attach a conversation listener when starting a thread.
                 // Use the same behavior as the v1 API with experimental_raw_events=false.
                 if let Err(err) = self
-                    .attach_conversation_listener(conversation_id, false)
+                    .attach_conversation_listener(conversation_id, false, ApiVersion::V2)
                     .await
                 {
                     tracing::warn!(
@@ -1541,7 +1643,7 @@ impl CodexMessageProcessor {
             }) => {
                 // Auto-attach a conversation listener when resuming a thread.
                 if let Err(err) = self
-                    .attach_conversation_listener(conversation_id, false)
+                    .attach_conversation_listener(conversation_id, false, ApiVersion::V2)
                     .await
                 {
                     tracing::warn!(
@@ -1570,7 +1672,15 @@ impl CodexMessageProcessor {
                         return;
                     }
                 };
-                let response = ThreadResumeResponse { thread };
+                let response = ThreadResumeResponse {
+                    thread,
+                    model: session_configured.model,
+                    model_provider: session_configured.model_provider_id,
+                    cwd: session_configured.cwd,
+                    approval_policy: session_configured.approval_policy.into(),
+                    sandbox: session_configured.sandbox_policy.into(),
+                    reasoning_effort: session_configured.reasoning_effort,
+                };
                 self.outgoing.send_response(request_id, response).await;
             }
             Err(err) => {
@@ -2290,9 +2400,6 @@ impl CodexMessageProcessor {
             }
         };
 
-        // Keep a copy of v2 inputs for the notification payload.
-        let v2_inputs_for_notif = params.input.clone();
-
         // Map v2 input items to core input items.
         let mapped_items: Vec<CoreInputItem> = params
             .input
@@ -2332,12 +2439,8 @@ impl CodexMessageProcessor {
             Ok(turn_id) => {
                 let turn = Turn {
                     id: turn_id.clone(),
-                    items: vec![ThreadItem::UserMessage {
-                        id: turn_id,
-                        content: v2_inputs_for_notif,
-                    }],
+                    items: vec![],
                     status: TurnStatus::InProgress,
-                    error: None,
                 };
 
                 let response = TurnStartResponse { turn: turn.clone() };
@@ -2360,6 +2463,64 @@ impl CodexMessageProcessor {
         }
     }
 
+    async fn review_start(&self, request_id: RequestId, params: ReviewStartParams) {
+        let ReviewStartParams {
+            thread_id,
+            target,
+            append_to_original_thread,
+        } = params;
+        let (_, conversation) = match self.conversation_from_thread_id(&thread_id).await {
+            Ok(v) => v,
+            Err(error) => {
+                self.outgoing.send_error(request_id, error).await;
+                return;
+            }
+        };
+
+        let (review_request, display_text) =
+            match Self::review_request_from_target(target, append_to_original_thread) {
+                Ok(value) => value,
+                Err(err) => {
+                    self.outgoing.send_error(request_id, err).await;
+                    return;
+                }
+            };
+
+        let turn_id = conversation.submit(Op::Review { review_request }).await;
+
+        match turn_id {
+            Ok(turn_id) => {
+                let mut items = Vec::new();
+                if !display_text.is_empty() {
+                    items.push(ThreadItem::UserMessage {
+                        id: turn_id.clone(),
+                        content: vec![V2UserInput::Text { text: display_text }],
+                    });
+                }
+                let turn = Turn {
+                    id: turn_id.clone(),
+                    items,
+                    status: TurnStatus::InProgress,
+                };
+                let response = TurnStartResponse { turn: turn.clone() };
+                self.outgoing.send_response(request_id, response).await;
+
+                let notif = TurnStartedNotification { turn };
+                self.outgoing
+                    .send_server_notification(ServerNotification::TurnStarted(notif))
+                    .await;
+            }
+            Err(err) => {
+                let error = JSONRPCErrorError {
+                    code: INTERNAL_ERROR_CODE,
+                    message: format!("failed to start review: {err}"),
+                    data: None,
+                };
+                self.outgoing.send_error(request_id, error).await;
+            }
+        }
+    }
+
     async fn turn_interrupt(&mut self, request_id: RequestId, params: TurnInterruptParams) {
         let TurnInterruptParams { thread_id, .. } = params;
 
@@ -2394,7 +2555,7 @@ impl CodexMessageProcessor {
             experimental_raw_events,
         } = params;
         match self
-            .attach_conversation_listener(conversation_id, experimental_raw_events)
+            .attach_conversation_listener(conversation_id, experimental_raw_events, ApiVersion::V1)
             .await
         {
             Ok(subscription_id) => {
@@ -2435,6 +2596,7 @@ impl CodexMessageProcessor {
         &mut self,
         conversation_id: ConversationId,
         experimental_raw_events: bool,
+        api_version: ApiVersion,
     ) -> Result<Uuid, JSONRPCErrorError> {
         let conversation = match self
             .conversation_manager
@@ -2458,6 +2620,8 @@ impl CodexMessageProcessor {
 
         let outgoing_for_task = self.outgoing.clone();
         let pending_interrupts = self.pending_interrupts.clone();
+        let turn_summary_store = self.turn_summary_store.clone();
+        let api_version_for_task = api_version;
         tokio::spawn(async move {
             loop {
                 tokio::select! {
@@ -2513,6 +2677,8 @@ impl CodexMessageProcessor {
                             conversation.clone(),
                             outgoing_for_task.clone(),
                             pending_interrupts.clone(),
+                            turn_summary_store.clone(),
+                            api_version_for_task,
                         )
                         .await;
                     }
@@ -2655,157 +2821,6 @@ impl CodexMessageProcessor {
     }
 }
 
-async fn apply_bespoke_event_handling(
-    event: Event,
-    conversation_id: ConversationId,
-    conversation: Arc<CodexConversation>,
-    outgoing: Arc<OutgoingMessageSender>,
-    pending_interrupts: PendingInterrupts,
-) {
-    let Event { id: event_id, msg } = event;
-    match msg {
-        EventMsg::ApplyPatchApprovalRequest(ApplyPatchApprovalRequestEvent {
-            call_id,
-            changes,
-            reason,
-            grant_root,
-        }) => {
-            let params = ApplyPatchApprovalParams {
-                conversation_id,
-                call_id,
-                file_changes: changes,
-                reason,
-                grant_root,
-            };
-            let rx = outgoing
-                .send_request(ServerRequestPayload::ApplyPatchApproval(params))
-                .await;
-            // TODO(mbolin): Enforce a timeout so this task does not live indefinitely?
-            tokio::spawn(async move {
-                on_patch_approval_response(event_id, rx, conversation).await;
-            });
-        }
-        EventMsg::AgentMessageContentDelta(event) => {
-            let notification = AgentMessageDeltaNotification {
-                item_id: event.item_id,
-                delta: event.delta,
-            };
-            outgoing
-                .send_server_notification(ServerNotification::AgentMessageDelta(notification))
-                .await;
-        }
-        EventMsg::ReasoningContentDelta(event) => {
-            let notification = ReasoningSummaryTextDeltaNotification {
-                item_id: event.item_id,
-                delta: event.delta,
-                summary_index: event.summary_index,
-            };
-            outgoing
-                .send_server_notification(ServerNotification::ReasoningSummaryTextDelta(
-                    notification,
-                ))
-                .await;
-        }
-        EventMsg::ReasoningRawContentDelta(event) => {
-            let notification = ReasoningTextDeltaNotification {
-                item_id: event.item_id,
-                delta: event.delta,
-                content_index: event.content_index,
-            };
-            outgoing
-                .send_server_notification(ServerNotification::ReasoningTextDelta(notification))
-                .await;
-        }
-        EventMsg::AgentReasoningSectionBreak(event) => {
-            let notification = ReasoningSummaryPartAddedNotification {
-                item_id: event.item_id,
-                summary_index: event.summary_index,
-            };
-            outgoing
-                .send_server_notification(ServerNotification::ReasoningSummaryPartAdded(
-                    notification,
-                ))
-                .await;
-        }
-        EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
-            call_id,
-            command,
-            cwd,
-            reason,
-            risk,
-            parsed_cmd,
-        }) => {
-            let params = ExecCommandApprovalParams {
-                conversation_id,
-                call_id,
-                command,
-                cwd,
-                reason,
-                risk,
-                parsed_cmd,
-            };
-            let rx = outgoing
-                .send_request(ServerRequestPayload::ExecCommandApproval(params))
-                .await;
-
-            // TODO(mbolin): Enforce a timeout so this task does not live indefinitely?
-            tokio::spawn(async move {
-                on_exec_approval_response(event_id, rx, conversation).await;
-            });
-        }
-        EventMsg::TokenCount(token_count_event) => {
-            if let Some(rate_limits) = token_count_event.rate_limits {
-                outgoing
-                    .send_server_notification(ServerNotification::AccountRateLimitsUpdated(
-                        AccountRateLimitsUpdatedNotification {
-                            rate_limits: rate_limits.into(),
-                        },
-                    ))
-                    .await;
-            }
-        }
-        EventMsg::ItemStarted(item_started_event) => {
-            let item: ThreadItem = item_started_event.item.clone().into();
-            let notification = ItemStartedNotification { item };
-            outgoing
-                .send_server_notification(ServerNotification::ItemStarted(notification))
-                .await;
-        }
-        EventMsg::ItemCompleted(item_completed_event) => {
-            let item: ThreadItem = item_completed_event.item.clone().into();
-            let notification = ItemCompletedNotification { item };
-            outgoing
-                .send_server_notification(ServerNotification::ItemCompleted(notification))
-                .await;
-        }
-        // If this is a TurnAborted, reply to any pending interrupt requests.
-        EventMsg::TurnAborted(turn_aborted_event) => {
-            let pending = {
-                let mut map = pending_interrupts.lock().await;
-                map.remove(&conversation_id).unwrap_or_default()
-            };
-            if !pending.is_empty() {
-                for (rid, ver) in pending {
-                    match ver {
-                        ApiVersion::V1 => {
-                            let response = InterruptConversationResponse {
-                                abort_reason: turn_aborted_event.reason.clone(),
-                            };
-                            outgoing.send_response(rid, response).await;
-                        }
-                        ApiVersion::V2 => {
-                            let response = TurnInterruptResponse {};
-                            outgoing.send_response(rid, response).await;
-                        }
-                    }
-                }
-            }
-        }
-
-        _ => {}
-    }
-}
-
 async fn derive_config_from_params(
     overrides: ConfigOverrides,
     cli_overrides: Option<std::collections::HashMap<String, serde_json::Value>>,
@@ -2819,84 +2834,6 @@ async fn derive_config_from_params(
     Config::load_with_cli_overrides(cli_overrides, overrides).await
 }
 
-async fn on_patch_approval_response(
-    event_id: String,
-    receiver: oneshot::Receiver<JsonRpcResult>,
-    codex: Arc<CodexConversation>,
-) {
-    let response = receiver.await;
-    let value = match response {
-        Ok(value) => value,
-        Err(err) => {
-            error!("request failed: {err:?}");
-            if let Err(submit_err) = codex
-                .submit(Op::PatchApproval {
-                    id: event_id.clone(),
-                    decision: ReviewDecision::Denied,
-                })
-                .await
-            {
-                error!("failed to submit denied PatchApproval after request failure: {submit_err}");
-            }
-            return;
-        }
-    };
-
-    let response =
-        serde_json::from_value::<ApplyPatchApprovalResponse>(value).unwrap_or_else(|err| {
-            error!("failed to deserialize ApplyPatchApprovalResponse: {err}");
-            ApplyPatchApprovalResponse {
-                decision: ReviewDecision::Denied,
-            }
-        });
-
-    if let Err(err) = codex
-        .submit(Op::PatchApproval {
-            id: event_id,
-            decision: response.decision,
-        })
-        .await
-    {
-        error!("failed to submit PatchApproval: {err}");
-    }
-}
-
-async fn on_exec_approval_response(
-    event_id: String,
-    receiver: oneshot::Receiver<JsonRpcResult>,
-    conversation: Arc<CodexConversation>,
-) {
-    let response = receiver.await;
-    let value = match response {
-        Ok(value) => value,
-        Err(err) => {
-            error!("request failed: {err:?}");
-            return;
-        }
-    };
-
-    // Try to deserialize `value` and then make the appropriate call to `codex`.
-    let response =
-        serde_json::from_value::<ExecCommandApprovalResponse>(value).unwrap_or_else(|err| {
-            error!("failed to deserialize ExecCommandApprovalResponse: {err}");
-            // If we cannot deserialize the response, we deny the request to be
-            // conservative.
-            ExecCommandApprovalResponse {
-                decision: ReviewDecision::Denied,
-            }
-        });
-
-    if let Err(err) = conversation
-        .submit(Op::ExecApproval {
-            id: event_id,
-            decision: response.decision,
-        })
-        .await
-    {
-        error!("failed to submit ExecApproval: {err}");
-    }
-}
-
 async fn read_summary_from_rollout(
     path: &Path,
     fallback_provider: &str,

codex-rs/app-server/src/fuzzy_file_search.rs

@@ -19,6 +19,10 @@ pub(crate) async fn run_fuzzy_file_search(
     roots: Vec<String>,
     cancellation_flag: Arc<AtomicBool>,
 ) -> Vec<FuzzyFileSearchResult> {
+    if roots.is_empty() {
+        return Vec::new();
+    }
+
     #[expect(clippy::expect_used)]
     let limit_per_root =
         NonZero::new(LIMIT_PER_ROOT).expect("LIMIT_PER_ROOT should be a valid non-zero usize");

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

@@ -28,6 +28,7 @@ use tracing_subscriber::filter::Targets;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::util::SubscriberInitExt;
 
+mod bespoke_event_handling;
 mod codex_message_processor;
 mod error_code;
 mod fuzzy_file_search;
@@ -46,7 +47,7 @@ pub async fn run_main(
 ) -> IoResult<()> {
     // Set up channels.
     let (incoming_tx, mut incoming_rx) = mpsc::channel::<JSONRPCMessage>(CHANNEL_CAPACITY);
-    let (outgoing_tx, mut outgoing_rx) = mpsc::unbounded_channel::<OutgoingMessage>();
+    let (outgoing_tx, mut outgoing_rx) = mpsc::channel::<OutgoingMessage>(CHANNEL_CAPACITY);
 
     // Task: read from stdin, push to `incoming_tx`.
     let stdin_reader_handle = tokio::spawn({

codex-rs/app-server/src/outgoing_message.rs

@@ -19,12 +19,12 @@ use crate::error_code::INTERNAL_ERROR_CODE;
 /// Sends messages to the client and manages request callbacks.
 pub(crate) struct OutgoingMessageSender {
     next_request_id: AtomicI64,
-    sender: mpsc::UnboundedSender<OutgoingMessage>,
+    sender: mpsc::Sender<OutgoingMessage>,
     request_id_to_callback: Mutex<HashMap<RequestId, oneshot::Sender<Result>>>,
 }
 
 impl OutgoingMessageSender {
-    pub(crate) fn new(sender: mpsc::UnboundedSender<OutgoingMessage>) -> Self {
+    pub(crate) fn new(sender: mpsc::Sender<OutgoingMessage>) -> Self {
         Self {
             next_request_id: AtomicI64::new(0),
             sender,
@@ -45,8 +45,12 @@ impl OutgoingMessageSender {
         }
 
         let outgoing_message =
-            OutgoingMessage::Request(request.request_with_id(outgoing_message_id));
-        let _ = self.sender.send(outgoing_message);
+            OutgoingMessage::Request(request.request_with_id(outgoing_message_id.clone()));
+        if let Err(err) = self.sender.send(outgoing_message).await {
+            warn!("failed to send request {outgoing_message_id:?} to client: {err:?}");
+            let mut request_id_to_callback = self.request_id_to_callback.lock().await;
+            request_id_to_callback.remove(&outgoing_message_id);
+        }
         rx_approve
     }
 
@@ -72,7 +76,9 @@ impl OutgoingMessageSender {
         match serde_json::to_value(response) {
             Ok(result) => {
                 let outgoing_message = OutgoingMessage::Response(OutgoingResponse { id, result });
-                let _ = self.sender.send(outgoing_message);
+                if let Err(err) = self.sender.send(outgoing_message).await {
+                    warn!("failed to send response to client: {err:?}");
+                }
             }
             Err(err) => {
                 self.send_error(
@@ -89,21 +95,29 @@ impl OutgoingMessageSender {
     }
 
     pub(crate) async fn send_server_notification(&self, notification: ServerNotification) {
-        let _ = self
+        if let Err(err) = self
             .sender
-            .send(OutgoingMessage::AppServerNotification(notification));
+            .send(OutgoingMessage::AppServerNotification(notification))
+            .await
+        {
+            warn!("failed to send server notification to client: {err:?}");
+        }
     }
 
     /// All notifications should be migrated to [`ServerNotification`] and
     /// [`OutgoingMessage::Notification`] should be removed.
     pub(crate) async fn send_notification(&self, notification: OutgoingNotification) {
         let outgoing_message = OutgoingMessage::Notification(notification);
-        let _ = self.sender.send(outgoing_message);
+        if let Err(err) = self.sender.send(outgoing_message).await {
+            warn!("failed to send notification to client: {err:?}");
+        }
     }
 
     pub(crate) async fn send_error(&self, id: RequestId, error: JSONRPCErrorError) {
         let outgoing_message = OutgoingMessage::Error(OutgoingError { id, error });
-        let _ = self.sender.send(outgoing_message);
+        if let Err(err) = self.sender.send(outgoing_message).await {
+            warn!("failed to send error to client: {err:?}");
+        }
     }
 }
 

codex-rs/app-server/tests/common/mcp_process.rs

@@ -35,6 +35,7 @@ use codex_app_server_protocol::NewConversationParams;
 use codex_app_server_protocol::RemoveConversationListenerParams;
 use codex_app_server_protocol::RequestId;
 use codex_app_server_protocol::ResumeConversationParams;
+use codex_app_server_protocol::ReviewStartParams;
 use codex_app_server_protocol::SendUserMessageParams;
 use codex_app_server_protocol::SendUserTurnParams;
 use codex_app_server_protocol::ServerRequest;
@@ -377,6 +378,15 @@ impl McpProcess {
         self.send_request("turn/interrupt", params).await
     }
 
+    /// Send a `review/start` JSON-RPC request (v2).
+    pub async fn send_review_start_request(
+        &mut self,
+        params: ReviewStartParams,
+    ) -> anyhow::Result<i64> {
+        let params = Some(serde_json::to_value(params)?);
+        self.send_request("review/start", params).await
+    }
+
     /// Send a `cancelLoginChatGpt` JSON-RPC request.
     pub async fn send_cancel_login_chat_gpt_request(
         &mut self,

codex-rs/app-server/tests/suite/config.rs

@@ -27,7 +27,7 @@ fn create_config_toml(codex_home: &Path) -> std::io::Result<()> {
     std::fs::write(
         config_toml,
         r#"
-model = "gpt-5-codex"
+model = "gpt-5.1-codex"
 approval_policy = "on-request"
 sandbox_mode = "workspace-write"
 model_reasoning_summary = "detailed"
@@ -87,7 +87,7 @@ async fn get_config_toml_parses_all_fields() -> Result<()> {
             }),
             forced_chatgpt_workspace_id: Some("12345678-0000-0000-0000-000000000000".into()),
             forced_login_method: Some(ForcedLoginMethod::Chatgpt),
-            model: Some("gpt-5-codex".into()),
+            model: Some("gpt-5.1-codex".into()),
             model_reasoning_effort: Some(ReasoningEffort::High),
             model_reasoning_summary: Some(ReasoningSummary::Detailed),
             model_verbosity: Some(Verbosity::Medium),

codex-rs/app-server/tests/suite/set_default_model.rs

@@ -57,7 +57,7 @@ fn create_config_toml(codex_home: &Path) -> std::io::Result<()> {
     std::fs::write(
         config_toml,
         r#"
-model = "gpt-5-codex"
+model = "gpt-5.1-codex"
 model_reasoning_effort = "medium"
 "#,
     )

codex-rs/app-server/tests/suite/v2/mod.rs

@@ -1,6 +1,7 @@
 mod account;
 mod model_list;
 mod rate_limits;
+mod review;
 mod thread_archive;
 mod thread_list;
 mod thread_resume;

codex-rs/app-server/tests/suite/v2/review.rs

@@ -0,0 +1,279 @@
+use anyhow::Result;
+use app_test_support::McpProcess;
+use app_test_support::create_final_assistant_message_sse_response;
+use app_test_support::create_mock_chat_completions_server_unchecked;
+use app_test_support::to_response;
+use codex_app_server_protocol::ItemCompletedNotification;
+use codex_app_server_protocol::ItemStartedNotification;
+use codex_app_server_protocol::JSONRPCError;
+use codex_app_server_protocol::JSONRPCNotification;
+use codex_app_server_protocol::JSONRPCResponse;
+use codex_app_server_protocol::RequestId;
+use codex_app_server_protocol::ReviewStartParams;
+use codex_app_server_protocol::ReviewTarget;
+use codex_app_server_protocol::ThreadItem;
+use codex_app_server_protocol::ThreadStartParams;
+use codex_app_server_protocol::ThreadStartResponse;
+use codex_app_server_protocol::TurnStartResponse;
+use codex_app_server_protocol::TurnStatus;
+use serde_json::json;
+use tempfile::TempDir;
+use tokio::time::timeout;
+
+const DEFAULT_READ_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(10);
+const INVALID_REQUEST_ERROR_CODE: i64 = -32600;
+
+#[tokio::test]
+async fn review_start_runs_review_turn_and_emits_code_review_item() -> Result<()> {
+    let review_payload = json!({
+        "findings": [
+            {
+                "title": "Prefer Stylize helpers",
+                "body": "Use .dim()/.bold() chaining instead of manual Style.",
+                "confidence_score": 0.9,
+                "priority": 1,
+                "code_location": {
+                    "absolute_file_path": "/tmp/file.rs",
+                    "line_range": {"start": 10, "end": 20}
+                }
+            }
+        ],
+        "overall_correctness": "good",
+        "overall_explanation": "Looks solid overall with minor polish suggested.",
+        "overall_confidence_score": 0.75
+    })
+    .to_string();
+    let responses = vec![create_final_assistant_message_sse_response(
+        &review_payload,
+    )?];
+    let server = create_mock_chat_completions_server_unchecked(responses).await;
+
+    let codex_home = TempDir::new()?;
+    create_config_toml(codex_home.path(), &server.uri())?;
+
+    let mut mcp = McpProcess::new(codex_home.path()).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+
+    let thread_id = start_default_thread(&mut mcp).await?;
+
+    let review_req = mcp
+        .send_review_start_request(ReviewStartParams {
+            thread_id: thread_id.clone(),
+            append_to_original_thread: true,
+            target: ReviewTarget::Commit {
+                sha: "1234567deadbeef".to_string(),
+                title: Some("Tidy UI colors".to_string()),
+            },
+        })
+        .await?;
+    let review_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(review_req)),
+    )
+    .await??;
+    let TurnStartResponse { turn } = to_response::<TurnStartResponse>(review_resp)?;
+    let turn_id = turn.id.clone();
+    assert_eq!(turn.status, TurnStatus::InProgress);
+    assert_eq!(turn.items.len(), 1);
+    match &turn.items[0] {
+        ThreadItem::UserMessage { content, .. } => {
+            assert_eq!(content.len(), 1);
+            assert!(matches!(
+                &content[0],
+                codex_app_server_protocol::UserInput::Text { .. }
+            ));
+        }
+        other => panic!("expected user message, got {other:?}"),
+    }
+
+    let _started: JSONRPCNotification = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("turn/started"),
+    )
+    .await??;
+    let item_started: JSONRPCNotification = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("item/started"),
+    )
+    .await??;
+    let started: ItemStartedNotification =
+        serde_json::from_value(item_started.params.expect("params must be present"))?;
+    match started.item {
+        ThreadItem::CodeReview { id, review } => {
+            assert_eq!(id, turn_id);
+            assert_eq!(review, "commit 1234567");
+        }
+        other => panic!("expected code review item, got {other:?}"),
+    }
+
+    let mut review_body: Option<String> = None;
+    for _ in 0..5 {
+        let review_notif: JSONRPCNotification = timeout(
+            DEFAULT_READ_TIMEOUT,
+            mcp.read_stream_until_notification_message("item/completed"),
+        )
+        .await??;
+        let completed: ItemCompletedNotification =
+            serde_json::from_value(review_notif.params.expect("params must be present"))?;
+        match completed.item {
+            ThreadItem::CodeReview { id, review } => {
+                assert_eq!(id, turn_id);
+                review_body = Some(review);
+                break;
+            }
+            ThreadItem::UserMessage { .. } => continue,
+            other => panic!("unexpected item/completed payload: {other:?}"),
+        }
+    }
+
+    let review = review_body.expect("did not observe a code review item");
+    assert!(review.contains("Prefer Stylize helpers"));
+    assert!(review.contains("/tmp/file.rs:10-20"));
+
+    Ok(())
+}
+
+#[tokio::test]
+async fn review_start_rejects_empty_base_branch() -> Result<()> {
+    let server = create_mock_chat_completions_server_unchecked(vec![]).await;
+    let codex_home = TempDir::new()?;
+    create_config_toml(codex_home.path(), &server.uri())?;
+
+    let mut mcp = McpProcess::new(codex_home.path()).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+    let thread_id = start_default_thread(&mut mcp).await?;
+
+    let request_id = mcp
+        .send_review_start_request(ReviewStartParams {
+            thread_id,
+            append_to_original_thread: true,
+            target: ReviewTarget::BaseBranch {
+                branch: "   ".to_string(),
+            },
+        })
+        .await?;
+    let error: JSONRPCError = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_error_message(RequestId::Integer(request_id)),
+    )
+    .await??;
+    assert_eq!(error.error.code, INVALID_REQUEST_ERROR_CODE);
+    assert!(
+        error.error.message.contains("branch must not be empty"),
+        "unexpected message: {}",
+        error.error.message
+    );
+
+    Ok(())
+}
+
+#[tokio::test]
+async fn review_start_rejects_empty_commit_sha() -> Result<()> {
+    let server = create_mock_chat_completions_server_unchecked(vec![]).await;
+    let codex_home = TempDir::new()?;
+    create_config_toml(codex_home.path(), &server.uri())?;
+
+    let mut mcp = McpProcess::new(codex_home.path()).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+    let thread_id = start_default_thread(&mut mcp).await?;
+
+    let request_id = mcp
+        .send_review_start_request(ReviewStartParams {
+            thread_id,
+            append_to_original_thread: true,
+            target: ReviewTarget::Commit {
+                sha: "\t".to_string(),
+                title: None,
+            },
+        })
+        .await?;
+    let error: JSONRPCError = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_error_message(RequestId::Integer(request_id)),
+    )
+    .await??;
+    assert_eq!(error.error.code, INVALID_REQUEST_ERROR_CODE);
+    assert!(
+        error.error.message.contains("sha must not be empty"),
+        "unexpected message: {}",
+        error.error.message
+    );
+
+    Ok(())
+}
+
+#[tokio::test]
+async fn review_start_rejects_empty_custom_instructions() -> Result<()> {
+    let server = create_mock_chat_completions_server_unchecked(vec![]).await;
+    let codex_home = TempDir::new()?;
+    create_config_toml(codex_home.path(), &server.uri())?;
+
+    let mut mcp = McpProcess::new(codex_home.path()).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+    let thread_id = start_default_thread(&mut mcp).await?;
+
+    let request_id = mcp
+        .send_review_start_request(ReviewStartParams {
+            thread_id,
+            append_to_original_thread: true,
+            target: ReviewTarget::Custom {
+                instructions: "\n\n".to_string(),
+            },
+        })
+        .await?;
+    let error: JSONRPCError = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_error_message(RequestId::Integer(request_id)),
+    )
+    .await??;
+    assert_eq!(error.error.code, INVALID_REQUEST_ERROR_CODE);
+    assert!(
+        error
+            .error
+            .message
+            .contains("instructions must not be empty"),
+        "unexpected message: {}",
+        error.error.message
+    );
+
+    Ok(())
+}
+
+async fn start_default_thread(mcp: &mut McpProcess) -> Result<String> {
+    let thread_req = mcp
+        .send_thread_start_request(ThreadStartParams {
+            model: Some("mock-model".to_string()),
+            ..Default::default()
+        })
+        .await?;
+    let thread_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(thread_req)),
+    )
+    .await??;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(thread_resp)?;
+    Ok(thread.id)
+}
+
+fn create_config_toml(codex_home: &std::path::Path, server_uri: &str) -> std::io::Result<()> {
+    let config_toml = codex_home.join("config.toml");
+    std::fs::write(
+        config_toml,
+        format!(
+            r#"
+model = "mock-model"
+approval_policy = "never"
+sandbox_mode = "read-only"
+
+model_provider = "mock_provider"
+
+[model_providers.mock_provider]
+name = "Mock provider"
+base_url = "{server_uri}/v1"
+wire_api = "chat"
+request_max_retries = 0
+stream_max_retries = 0
+"#
+        ),
+    )
+}

codex-rs/app-server/tests/suite/v2/thread_archive.rs

@@ -35,7 +35,7 @@ async fn thread_archive_moves_rollout_into_archived_directory() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(start_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
     assert!(!thread.id.is_empty());
 
     // Locate the rollout path recorded for this thread id.

codex-rs/app-server/tests/suite/v2/thread_resume.rs

@@ -27,7 +27,7 @@ async fn thread_resume_returns_original_thread() -> Result<()> {
     // Start a thread.
     let start_id = mcp
         .send_thread_start_request(ThreadStartParams {
-            model: Some("gpt-5-codex".to_string()),
+            model: Some("gpt-5.1-codex".to_string()),
             ..Default::default()
         })
         .await?;
@@ -36,7 +36,7 @@ async fn thread_resume_returns_original_thread() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(start_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
 
     // Resume it via v2 API.
     let resume_id = mcp
@@ -50,8 +50,9 @@ async fn thread_resume_returns_original_thread() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(resume_id)),
     )
     .await??;
-    let ThreadResumeResponse { thread: resumed } =
-        to_response::<ThreadResumeResponse>(resume_resp)?;
+    let ThreadResumeResponse {
+        thread: resumed, ..
+    } = to_response::<ThreadResumeResponse>(resume_resp)?;
     assert_eq!(resumed, thread);
 
     Ok(())
@@ -68,7 +69,7 @@ async fn thread_resume_prefers_path_over_thread_id() -> Result<()> {
 
     let start_id = mcp
         .send_thread_start_request(ThreadStartParams {
-            model: Some("gpt-5-codex".to_string()),
+            model: Some("gpt-5.1-codex".to_string()),
             ..Default::default()
         })
         .await?;
@@ -77,7 +78,7 @@ async fn thread_resume_prefers_path_over_thread_id() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(start_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
 
     let thread_path = thread.path.clone();
     let resume_id = mcp
@@ -93,8 +94,9 @@ async fn thread_resume_prefers_path_over_thread_id() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(resume_id)),
     )
     .await??;
-    let ThreadResumeResponse { thread: resumed } =
-        to_response::<ThreadResumeResponse>(resume_resp)?;
+    let ThreadResumeResponse {
+        thread: resumed, ..
+    } = to_response::<ThreadResumeResponse>(resume_resp)?;
     assert_eq!(resumed, thread);
 
     Ok(())
@@ -112,7 +114,7 @@ async fn thread_resume_supports_history_and_overrides() -> Result<()> {
     // Start a thread.
     let start_id = mcp
         .send_thread_start_request(ThreadStartParams {
-            model: Some("gpt-5-codex".to_string()),
+            model: Some("gpt-5.1-codex".to_string()),
             ..Default::default()
         })
         .await?;
@@ -121,7 +123,7 @@ async fn thread_resume_supports_history_and_overrides() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(start_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
 
     let history_text = "Hello from history";
     let history = vec![ResponseItem::Message {
@@ -147,10 +149,13 @@ async fn thread_resume_supports_history_and_overrides() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(resume_id)),
     )
     .await??;
-    let ThreadResumeResponse { thread: resumed } =
-        to_response::<ThreadResumeResponse>(resume_resp)?;
+    let ThreadResumeResponse {
+        thread: resumed,
+        model_provider,
+        ..
+    } = to_response::<ThreadResumeResponse>(resume_resp)?;
     assert!(!resumed.id.is_empty());
-    assert_eq!(resumed.model_provider, "mock_provider");
+    assert_eq!(model_provider, "mock_provider");
     assert_eq!(resumed.preview, history_text);
 
     Ok(())

codex-rs/app-server/tests/suite/v2/thread_start.rs

@@ -29,7 +29,7 @@ async fn thread_start_creates_thread_and_emits_started() -> Result<()> {
     // Start a v2 thread with an explicit model override.
     let req_id = mcp
         .send_thread_start_request(ThreadStartParams {
-            model: Some("gpt-5".to_string()),
+            model: Some("gpt-5.1".to_string()),
             ..Default::default()
         })
         .await?;
@@ -40,13 +40,17 @@ async fn thread_start_creates_thread_and_emits_started() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(req_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(resp)?;
+    let ThreadStartResponse {
+        thread,
+        model_provider,
+        ..
+    } = to_response::<ThreadStartResponse>(resp)?;
     assert!(!thread.id.is_empty(), "thread id should not be empty");
     assert!(
         thread.preview.is_empty(),
         "new threads should start with an empty preview"
     );
-    assert_eq!(thread.model_provider, "mock_provider");
+    assert_eq!(model_provider, "mock_provider");
     assert!(
         thread.created_at > 0,
         "created_at should be a positive UNIX timestamp"

codex-rs/app-server/tests/suite/v2/turn_interrupt.rs

@@ -5,14 +5,17 @@ use app_test_support::McpProcess;
 use app_test_support::create_mock_chat_completions_server;
 use app_test_support::create_shell_sse_response;
 use app_test_support::to_response;
+use codex_app_server_protocol::JSONRPCNotification;
 use codex_app_server_protocol::JSONRPCResponse;
 use codex_app_server_protocol::RequestId;
 use codex_app_server_protocol::ThreadStartParams;
 use codex_app_server_protocol::ThreadStartResponse;
+use codex_app_server_protocol::TurnCompletedNotification;
 use codex_app_server_protocol::TurnInterruptParams;
 use codex_app_server_protocol::TurnInterruptResponse;
 use codex_app_server_protocol::TurnStartParams;
 use codex_app_server_protocol::TurnStartResponse;
+use codex_app_server_protocol::TurnStatus;
 use codex_app_server_protocol::UserInput as V2UserInput;
 use tempfile::TempDir;
 use tokio::time::timeout;
@@ -62,7 +65,7 @@ async fn turn_interrupt_aborts_running_turn() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(thread_req)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(thread_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(thread_resp)?;
 
     // Start a turn that triggers a long-running command.
     let turn_req = mcp
@@ -99,7 +102,18 @@ async fn turn_interrupt_aborts_running_turn() -> Result<()> {
     .await??;
     let _resp: TurnInterruptResponse = to_response::<TurnInterruptResponse>(interrupt_resp)?;
 
-    // No fields to assert on; successful deserialization confirms proper response shape.
+    let completed_notif: JSONRPCNotification = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("turn/completed"),
+    )
+    .await??;
+    let completed: TurnCompletedNotification = serde_json::from_value(
+        completed_notif
+            .params
+            .expect("turn/completed params must be present"),
+    )?;
+    assert_eq!(completed.turn.status, TurnStatus::Interrupted);
+
     Ok(())
 }
 

codex-rs/app-server/tests/suite/v2/turn_start.rs

@@ -5,21 +5,23 @@ use app_test_support::create_mock_chat_completions_server;
 use app_test_support::create_mock_chat_completions_server_unchecked;
 use app_test_support::create_shell_sse_response;
 use app_test_support::to_response;
+use codex_app_server_protocol::CommandExecutionStatus;
+use codex_app_server_protocol::ItemStartedNotification;
 use codex_app_server_protocol::JSONRPCNotification;
 use codex_app_server_protocol::JSONRPCResponse;
 use codex_app_server_protocol::RequestId;
 use codex_app_server_protocol::ServerRequest;
+use codex_app_server_protocol::ThreadItem;
 use codex_app_server_protocol::ThreadStartParams;
 use codex_app_server_protocol::ThreadStartResponse;
+use codex_app_server_protocol::TurnCompletedNotification;
 use codex_app_server_protocol::TurnStartParams;
 use codex_app_server_protocol::TurnStartResponse;
 use codex_app_server_protocol::TurnStartedNotification;
+use codex_app_server_protocol::TurnStatus;
 use codex_app_server_protocol::UserInput as V2UserInput;
 use codex_core::protocol_config_types::ReasoningEffort;
 use codex_core::protocol_config_types::ReasoningSummary;
-use codex_protocol::parse_command::ParsedCommand;
-use codex_protocol::protocol::Event;
-use codex_protocol::protocol::EventMsg;
 use core_test_support::skip_if_no_network;
 use pretty_assertions::assert_eq;
 use std::path::Path;
@@ -57,7 +59,7 @@ async fn turn_start_emits_notifications_and_accepts_model_override() -> Result<(
         mcp.read_stream_until_response_message(RequestId::Integer(thread_req)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(thread_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(thread_resp)?;
 
     // Start a turn with only input and thread_id set (no overrides).
     let turn_req = mcp
@@ -118,13 +120,17 @@ async fn turn_start_emits_notifications_and_accepts_model_override() -> Result<(
     )
     .await??;
 
-    // And we should ultimately get a task_complete without having to add a
-    // legacy conversation listener explicitly (auto-attached by thread/start).
-    let _task_complete: JSONRPCNotification = timeout(
+    let completed_notif: JSONRPCNotification = timeout(
         DEFAULT_READ_TIMEOUT,
-        mcp.read_stream_until_notification_message("codex/event/task_complete"),
+        mcp.read_stream_until_notification_message("turn/completed"),
     )
     .await??;
+    let completed: TurnCompletedNotification = serde_json::from_value(
+        completed_notif
+            .params
+            .expect("turn/completed params must be present"),
+    )?;
+    assert_eq!(completed.turn.status, TurnStatus::Completed);
 
     Ok(())
 }
@@ -157,7 +163,7 @@ async fn turn_start_accepts_local_image_input() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(thread_req)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(thread_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(thread_resp)?;
 
     let image_path = codex_home.path().join("image.png");
     // No need to actually write the file; we just exercise the input path.
@@ -233,9 +239,9 @@ async fn turn_start_exec_approval_toggle_v2() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(start_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
 
-    // turn/start — expect ExecCommandApproval request from server
+    // turn/start — expect CommandExecutionRequestApproval request from server
     let first_turn_id = mcp
         .send_turn_start_request(TurnStartParams {
             thread_id: thread.id.clone(),
@@ -258,16 +264,10 @@ async fn turn_start_exec_approval_toggle_v2() -> Result<()> {
         mcp.read_stream_until_request_message(),
     )
     .await??;
-    let ServerRequest::ExecCommandApproval { request_id, params } = server_req else {
-        panic!("expected ExecCommandApproval request");
+    let ServerRequest::CommandExecutionRequestApproval { request_id, params } = server_req else {
+        panic!("expected CommandExecutionRequestApproval request");
     };
-    assert_eq!(params.call_id, "call1");
-    assert_eq!(
-        params.parsed_cmd,
-        vec![ParsedCommand::Unknown {
-            cmd: "python3 -c 'print(42)'".to_string()
-        }]
-    );
+    assert_eq!(params.item_id, "call1");
 
     // Approve and wait for task completion
     mcp.send_response(
@@ -280,6 +280,11 @@ async fn turn_start_exec_approval_toggle_v2() -> Result<()> {
         mcp.read_stream_until_notification_message("codex/event/task_complete"),
     )
     .await??;
+    timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("turn/completed"),
+    )
+    .await??;
 
     // Second turn with approval_policy=never should not elicit approval
     let second_turn_id = mcp
@@ -302,20 +307,23 @@ async fn turn_start_exec_approval_toggle_v2() -> Result<()> {
     )
     .await??;
 
-    // Ensure we do NOT receive an ExecCommandApproval request before task completes
+    // Ensure we do NOT receive a CommandExecutionRequestApproval request before task completes
     timeout(
         DEFAULT_READ_TIMEOUT,
         mcp.read_stream_until_notification_message("codex/event/task_complete"),
     )
     .await??;
+    timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("turn/completed"),
+    )
+    .await??;
 
     Ok(())
 }
 
 #[tokio::test]
 async fn turn_start_updates_sandbox_and_cwd_between_turns_v2() -> Result<()> {
-    // When returning Result from a test, pass an Ok(()) to the skip macro
-    // so the early return type matches. The no-arg form returns unit.
     skip_if_no_network!(Ok(()));
 
     let tmp = TempDir::new()?;
@@ -370,7 +378,7 @@ async fn turn_start_updates_sandbox_and_cwd_between_turns_v2() -> Result<()> {
         mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
     )
     .await??;
-    let ThreadStartResponse { thread } = to_response::<ThreadStartResponse>(start_resp)?;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
 
     // first turn with workspace-write sandbox and first_cwd
     let first_turn = mcp
@@ -424,29 +432,35 @@ async fn turn_start_updates_sandbox_and_cwd_between_turns_v2() -> Result<()> {
     )
     .await??;
 
-    let exec_begin_notification = timeout(
-        DEFAULT_READ_TIMEOUT,
-        mcp.read_stream_until_notification_message("codex/event/exec_command_begin"),
-    )
+    let command_exec_item = timeout(DEFAULT_READ_TIMEOUT, async {
+        loop {
+            let item_started_notification = mcp
+                .read_stream_until_notification_message("item/started")
+                .await?;
+            let params = item_started_notification
+                .params
+                .clone()
+                .expect("item/started params");
+            let item_started: ItemStartedNotification =
+                serde_json::from_value(params).expect("deserialize item/started notification");
+            if matches!(item_started.item, ThreadItem::CommandExecution { .. }) {
+                return Ok::<ThreadItem, anyhow::Error>(item_started.item);
+            }
+        }
+    })
     .await??;
-    let params = exec_begin_notification
-        .params
-        .clone()
-        .expect("exec_command_begin params");
-    let event: Event = serde_json::from_value(params).expect("deserialize exec begin event");
-    let exec_begin = match event.msg {
-        EventMsg::ExecCommandBegin(exec_begin) => exec_begin,
-        other => panic!("expected ExecCommandBegin event, got {other:?}"),
+    let ThreadItem::CommandExecution {
+        cwd,
+        command,
+        status,
+        ..
+    } = command_exec_item
+    else {
+        unreachable!("loop ensures we break on command execution items");
     };
-    assert_eq!(exec_begin.cwd, second_cwd);
-    assert_eq!(
-        exec_begin.command,
-        vec![
-            "bash".to_string(),
-            "-lc".to_string(),
-            "echo second turn".to_string()
-        ]
-    );
+    assert_eq!(cwd, second_cwd);
+    assert_eq!(command, "bash -lc 'echo second turn'");
+    assert_eq!(status, CommandExecutionStatus::InProgress);
 
     timeout(
         DEFAULT_READ_TIMEOUT,

codex-rs/cli/src/debug_sandbox.rs

@@ -138,11 +138,7 @@ async fn run_command_under_sandbox(
         {
             use codex_windows_sandbox::run_windows_sandbox_capture;
 
-            let policy_str = match &config.sandbox_policy {
-                codex_core::protocol::SandboxPolicy::DangerFullAccess => "workspace-write",
-                codex_core::protocol::SandboxPolicy::ReadOnly => "read-only",
-                codex_core::protocol::SandboxPolicy::WorkspaceWrite { .. } => "workspace-write",
-            };
+            let policy_str = serde_json::to_string(&config.sandbox_policy)?;
 
             let sandbox_cwd = sandbox_policy_cwd.clone();
             let cwd_clone = cwd.clone();
@@ -153,13 +149,13 @@ async fn run_command_under_sandbox(
             // Preflight audit is invoked elsewhere at the appropriate times.
             let res = tokio::task::spawn_blocking(move || {
                 run_windows_sandbox_capture(
-                    policy_str,
+                    policy_str.as_str(),
                     &sandbox_cwd,
+                    base_dir.as_path(),
                     command_vec,
                     &cwd_clone,
                     env_map,
                     None,
-                    Some(base_dir.as_path()),
                 )
             })
             .await;

codex-rs/cli/src/main.rs

@@ -134,6 +134,10 @@ struct ResumeCommand {
     #[arg(long = "last", default_value_t = false, conflicts_with = "session_id")]
     last: bool,
 
+    /// Show all sessions (disables cwd filtering and shows CWD column).
+    #[arg(long = "all", default_value_t = false)]
+    all: bool,
+
     #[clap(flatten)]
     config_overrides: TuiCli,
 }
@@ -448,6 +452,7 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
         Some(Subcommand::Resume(ResumeCommand {
             session_id,
             last,
+            all,
             config_overrides,
         })) => {
             interactive = finalize_resume_interactive(
@@ -455,6 +460,7 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
                 root_config_overrides.clone(),
                 session_id,
                 last,
+                all,
                 config_overrides,
             );
             let exit_info = codex_tui::run_main(interactive, codex_linux_sandbox_exe).await?;
@@ -611,6 +617,7 @@ fn finalize_resume_interactive(
     root_config_overrides: CliConfigOverrides,
     session_id: Option<String>,
     last: bool,
+    show_all: bool,
     resume_cli: TuiCli,
 ) -> TuiCli {
     // Start with the parsed interactive CLI so resume shares the same
@@ -619,6 +626,7 @@ fn finalize_resume_interactive(
     interactive.resume_picker = resume_session_id.is_none() && !last;
     interactive.resume_last = last;
     interactive.resume_session_id = resume_session_id;
+    interactive.resume_show_all = show_all;
 
     // Merge resume-scoped flags and overrides with highest precedence.
     merge_resume_cli_flags(&mut interactive, resume_cli);
@@ -702,13 +710,21 @@ mod tests {
         let Subcommand::Resume(ResumeCommand {
             session_id,
             last,
+            all,
             config_overrides: resume_cli,
         }) = subcommand.expect("resume present")
         else {
             unreachable!()
         };
 
-        finalize_resume_interactive(interactive, root_overrides, session_id, last, resume_cli)
+        finalize_resume_interactive(
+            interactive,
+            root_overrides,
+            session_id,
+            last,
+            all,
+            resume_cli,
+        )
     }
 
     fn sample_exit_info(conversation: Option<&str>) -> AppExitInfo {
@@ -761,9 +777,9 @@ mod tests {
 
     #[test]
     fn resume_model_flag_applies_when_no_root_flags() {
-        let interactive = finalize_from_args(["codex", "resume", "-m", "gpt-5-test"].as_ref());
+        let interactive = finalize_from_args(["codex", "resume", "-m", "gpt-5.1-test"].as_ref());
 
-        assert_eq!(interactive.model.as_deref(), Some("gpt-5-test"));
+        assert_eq!(interactive.model.as_deref(), Some("gpt-5.1-test"));
         assert!(interactive.resume_picker);
         assert!(!interactive.resume_last);
         assert_eq!(interactive.resume_session_id, None);
@@ -775,6 +791,7 @@ mod tests {
         assert!(interactive.resume_picker);
         assert!(!interactive.resume_last);
         assert_eq!(interactive.resume_session_id, None);
+        assert!(!interactive.resume_show_all);
     }
 
     #[test]
@@ -783,6 +800,7 @@ mod tests {
         assert!(!interactive.resume_picker);
         assert!(interactive.resume_last);
         assert_eq!(interactive.resume_session_id, None);
+        assert!(!interactive.resume_show_all);
     }
 
     #[test]
@@ -791,6 +809,14 @@ mod tests {
         assert!(!interactive.resume_picker);
         assert!(!interactive.resume_last);
         assert_eq!(interactive.resume_session_id.as_deref(), Some("1234"));
+        assert!(!interactive.resume_show_all);
+    }
+
+    #[test]
+    fn resume_all_flag_sets_show_all() {
+        let interactive = finalize_from_args(["codex", "resume", "--all"].as_ref());
+        assert!(interactive.resume_picker);
+        assert!(interactive.resume_show_all);
     }
 
     #[test]
@@ -808,7 +834,7 @@ mod tests {
                 "--ask-for-approval",
                 "on-request",
                 "-m",
-                "gpt-5-test",
+                "gpt-5.1-test",
                 "-p",
                 "my-profile",
                 "-C",
@@ -819,7 +845,7 @@ mod tests {
             .as_ref(),
         );
 
-        assert_eq!(interactive.model.as_deref(), Some("gpt-5-test"));
+        assert_eq!(interactive.model.as_deref(), Some("gpt-5.1-test"));
         assert!(interactive.oss);
         assert_eq!(interactive.config_profile.as_deref(), Some("my-profile"));
         assert_matches!(

codex-rs/cloud-tasks/src/lib.rs

@@ -1721,6 +1721,7 @@ mod tests {
     use ratatui::layout::Rect;
 
     #[test]
+    #[ignore = "very slow"]
     fn composer_input_renders_typed_characters() {
         let mut composer = ComposerInput::new();
         let key = KeyEvent::new(KeyCode::Char('a'), KeyModifiers::NONE);

codex-rs/codex-backend-openapi-models/src/models/credit_status_details.rs

@@ -0,0 +1,52 @@
+/*
+ * codex-backend
+ *
+ * codex-backend
+ *
+ * The version of the OpenAPI document: 0.0.1
+ *
+ * Generated by: https://openapi-generator.tech
+ */
+use serde::Deserialize;
+use serde::Serialize;
+
+#[derive(Clone, Default, Debug, PartialEq, Serialize, Deserialize)]
+pub struct CreditStatusDetails {
+    #[serde(rename = "has_credits")]
+    pub has_credits: bool,
+    #[serde(rename = "unlimited")]
+    pub unlimited: bool,
+    #[serde(
+        rename = "balance",
+        default,
+        with = "::serde_with::rust::double_option",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub balance: Option<Option<String>>,
+    #[serde(
+        rename = "approx_local_messages",
+        default,
+        with = "::serde_with::rust::double_option",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub approx_local_messages: Option<Option<Vec<serde_json::Value>>>,
+    #[serde(
+        rename = "approx_cloud_messages",
+        default,
+        with = "::serde_with::rust::double_option",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub approx_cloud_messages: Option<Option<Vec<serde_json::Value>>>,
+}
+
+impl CreditStatusDetails {
+    pub fn new(has_credits: bool, unlimited: bool) -> CreditStatusDetails {
+        CreditStatusDetails {
+            has_credits,
+            unlimited,
+            balance: None,
+            approx_local_messages: None,
+            approx_cloud_messages: None,
+        }
+    }
+}

codex-rs/codex-backend-openapi-models/src/models/mod.rs

@@ -32,3 +32,6 @@ pub use self::rate_limit_status_details::RateLimitStatusDetails;
 
 pub mod rate_limit_window_snapshot;
 pub use self::rate_limit_window_snapshot::RateLimitWindowSnapshot;
+
+pub mod credit_status_details;
+pub use self::credit_status_details::CreditStatusDetails;

codex-rs/codex-backend-openapi-models/src/models/rate_limit_status_payload.rs

@@ -23,6 +23,13 @@ pub struct RateLimitStatusPayload {
         skip_serializing_if = "Option::is_none"
     )]
     pub rate_limit: Option<Option<Box<models::RateLimitStatusDetails>>>,
+    #[serde(
+        rename = "credits",
+        default,
+        with = "::serde_with::rust::double_option",
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub credits: Option<Option<Box<models::CreditStatusDetails>>>,
 }
 
 impl RateLimitStatusPayload {
@@ -30,12 +37,15 @@ impl RateLimitStatusPayload {
         RateLimitStatusPayload {
             plan_type,
             rate_limit: None,
+            credits: None,
         }
     }
 }
 
 #[derive(Clone, Copy, Debug, Eq, PartialEq, Ord, PartialOrd, Hash, Serialize, Deserialize)]
 pub enum PlanType {
+    #[serde(rename = "guest")]
+    Guest,
     #[serde(rename = "free")]
     Free,
     #[serde(rename = "go")]
@@ -44,6 +54,8 @@ pub enum PlanType {
     Plus,
     #[serde(rename = "pro")]
     Pro,
+    #[serde(rename = "free_workspace")]
+    FreeWorkspace,
     #[serde(rename = "team")]
     Team,
     #[serde(rename = "business")]
@@ -52,6 +64,8 @@ pub enum PlanType {
     Education,
     #[serde(rename = "quorum")]
     Quorum,
+    #[serde(rename = "k12")]
+    K12,
     #[serde(rename = "enterprise")]
     Enterprise,
     #[serde(rename = "edu")]
@@ -60,6 +74,6 @@ pub enum PlanType {
 
 impl Default for PlanType {
     fn default() -> PlanType {
-        Self::Free
+        Self::Guest
     }
 }

codex-rs/codex-backend-openapi-models/src/models/rate_limit_window_snapshot.rs

@@ -7,7 +7,6 @@
  *
  * Generated by: https://openapi-generator.tech
  */
-
 use serde::Deserialize;
 use serde::Serialize;
 

codex-rs/common/Cargo.toml

@@ -10,6 +10,8 @@ workspace = true
 clap = { workspace = true, features = ["derive", "wrap_help"], optional = true }
 codex-app-server-protocol = { workspace = true }
 codex-core = { workspace = true }
+codex-lmstudio = { workspace = true }
+codex-ollama = { workspace = true }
 codex-protocol = { workspace = true }
 once_cell = { workspace = true }
 serde = { workspace = true, optional = true }

codex-rs/common/src/approval_presets.rs

@@ -24,21 +24,21 @@ pub fn builtin_approval_presets() -> Vec<ApprovalPreset> {
         ApprovalPreset {
             id: "read-only",
             label: "Read Only",
-            description: "Codex can read files and answer questions. Codex requires approval to make edits, run commands, or access network.",
+            description: "Requires approval to edit files and run commands.",
             approval: AskForApproval::OnRequest,
             sandbox: SandboxPolicy::ReadOnly,
         },
         ApprovalPreset {
             id: "auto",
-            label: "Auto",
-            description: "Codex can read files, make edits, and run commands in the workspace. Codex requires approval to work outside the workspace or access network.",
+            label: "Agent",
+            description: "Read and edit files, and run commands.",
             approval: AskForApproval::OnRequest,
             sandbox: SandboxPolicy::new_workspace_write_policy(),
         },
         ApprovalPreset {
             id: "full-access",
-            label: "Full Access",
-            description: "Codex can read files, make edits, and run commands with network access, without approval. Exercise caution.",
+            label: "Agent (full access)",
+            description: "Codex can edit files outside this workspace and run commands with network access. Exercise caution when using.",
             approval: AskForApproval::Never,
             sandbox: SandboxPolicy::DangerFullAccess,
         },

codex-rs/common/src/lib.rs

@@ -37,3 +37,5 @@ pub mod model_presets;
 // Shared approval presets (AskForApproval + Sandbox) used by TUI and MCP server
 // Not to be confused with AskForApproval, which we should probably rename to EscalationPolicy.
 pub mod approval_presets;
+// Shared OSS provider utilities used by TUI and exec
+pub mod oss;

codex-rs/common/src/oss.rs

@@ -0,0 +1,60 @@
+//! OSS provider utilities shared between TUI and exec.
+
+use codex_core::LMSTUDIO_OSS_PROVIDER_ID;
+use codex_core::OLLAMA_OSS_PROVIDER_ID;
+use codex_core::config::Config;
+
+/// Returns the default model for a given OSS provider.
+pub fn get_default_model_for_oss_provider(provider_id: &str) -> Option<&'static str> {
+    match provider_id {
+        LMSTUDIO_OSS_PROVIDER_ID => Some(codex_lmstudio::DEFAULT_OSS_MODEL),
+        OLLAMA_OSS_PROVIDER_ID => Some(codex_ollama::DEFAULT_OSS_MODEL),
+        _ => None,
+    }
+}
+
+/// Ensures the specified OSS provider is ready (models downloaded, service reachable).
+pub async fn ensure_oss_provider_ready(
+    provider_id: &str,
+    config: &Config,
+) -> Result<(), std::io::Error> {
+    match provider_id {
+        LMSTUDIO_OSS_PROVIDER_ID => {
+            codex_lmstudio::ensure_oss_ready(config)
+                .await
+                .map_err(|e| std::io::Error::other(format!("OSS setup failed: {e}")))?;
+        }
+        OLLAMA_OSS_PROVIDER_ID => {
+            codex_ollama::ensure_oss_ready(config)
+                .await
+                .map_err(|e| std::io::Error::other(format!("OSS setup failed: {e}")))?;
+        }
+        _ => {
+            // Unknown provider, skip setup
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_get_default_model_for_provider_lmstudio() {
+        let result = get_default_model_for_oss_provider(LMSTUDIO_OSS_PROVIDER_ID);
+        assert_eq!(result, Some(codex_lmstudio::DEFAULT_OSS_MODEL));
+    }
+
+    #[test]
+    fn test_get_default_model_for_provider_ollama() {
+        let result = get_default_model_for_oss_provider(OLLAMA_OSS_PROVIDER_ID);
+        assert_eq!(result, Some(codex_ollama::DEFAULT_OSS_MODEL));
+    }
+
+    #[test]
+    fn test_get_default_model_for_provider_unknown() {
+        let result = get_default_model_for_oss_provider("unknown-provider");
+        assert_eq!(result, None);
+    }
+}

codex-rs/core/Cargo.toml

@@ -25,7 +25,6 @@ codex-async-utils = { workspace = true }
 codex-file-search = { workspace = true }
 codex-git = { workspace = true }
 codex-keyring-store = { workspace = true }
-codex-execpolicy2 = { workspace = true }
 codex-otel = { workspace = true, features = ["otel"] }
 codex-protocol = { workspace = true }
 codex-rmcp-client = { workspace = true }
@@ -41,12 +40,7 @@ eventsource-stream = { workspace = true }
 futures = { workspace = true }
 http = { workspace = true }
 indexmap = { workspace = true }
-keyring = { workspace = true, features = [
-    "apple-native",
-    "crypto-rust",
-    "linux-native-async-persistent",
-    "windows-native",
-] }
+keyring = { workspace = true, features = ["crypto-rust"] }
 libc = { workspace = true }
 mcp-types = { workspace = true }
 os_info = { workspace = true }
@@ -91,9 +85,11 @@ wildmatch = { workspace = true }
 [target.'cfg(target_os = "linux")'.dependencies]
 landlock = { workspace = true }
 seccompiler = { workspace = true }
+keyring = { workspace = true, features = ["linux-native-async-persistent"] }
 
 [target.'cfg(target_os = "macos")'.dependencies]
 core-foundation = "0.9"
+keyring = { workspace = true, features = ["apple-native"] }
 
 # Build OpenSSL from source for musl builds.
 [target.x86_64-unknown-linux-musl.dependencies]
@@ -103,6 +99,12 @@ openssl-sys = { workspace = true, features = ["vendored"] }
 [target.aarch64-unknown-linux-musl.dependencies]
 openssl-sys = { workspace = true, features = ["vendored"] }
 
+[target.'cfg(target_os = "windows")'.dependencies]
+keyring = { workspace = true, features = ["windows-native"] }
+
+[target.'cfg(any(target_os = "freebsd", target_os = "openbsd"))'.dependencies]
+keyring = { workspace = true, features = ["sync-secret-service"] }
+
 [dev-dependencies]
 assert_cmd = { workspace = true }
 assert_matches = { workspace = true }

codex-rs/core/gpt_5_1_prompt.md

@@ -318,8 +318,6 @@ For casual greetings, acknowledgements, or other one-off conversational messages
 
 When using the shell, you must adhere to the following guidelines:
 
-- The arguments to `shell` will be passed to execvp().
-- Always set the `workdir` param when using the shell function. Do not use `cd` unless absolutely necessary.
 - When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
 - Read files in chunks with a max chunk size of 250 lines. Do not use python scripts to attempt to output larger chunks of a file. Command line output will be truncated after 10 kilobytes or 256 lines of output, regardless of the command used.
 

codex-rs/core/gpt_5_codex_prompt.md

@@ -2,8 +2,6 @@ You are Codex, based on GPT-5. You are running as a coding agent in the Codex CL
 
 ## General
 
-- The arguments to `shell` will be passed to execvp(). Most terminal commands should be prefixed with ["bash", "-lc"].
-- Always set the `workdir` param when using the shell function. Do not use `cd` unless absolutely necessary.
 - When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
 
 ## Editing constraints

codex-rs/core/src/chat_completions.rs

@@ -81,6 +81,7 @@ pub(crate) async fn stream_chat_completions(
             ResponseItem::CustomToolCallOutput { .. } => {}
             ResponseItem::WebSearchCall { .. } => {}
             ResponseItem::GhostSnapshot { .. } => {}
+            ResponseItem::CompactionSummary { .. } => {}
         }
     }
 
@@ -320,7 +321,8 @@ pub(crate) async fn stream_chat_completions(
             }
             ResponseItem::Reasoning { .. }
             | ResponseItem::WebSearchCall { .. }
-            | ResponseItem::Other => {
+            | ResponseItem::Other
+            | ResponseItem::CompactionSummary { .. } => {
                 // Omit these items from the conversation history.
                 continue;
             }
@@ -673,7 +675,9 @@ async fn process_chat_sse<S>(
             }
 
             // Emit end-of-turn when finish_reason signals completion.
-            if let Some(finish_reason) = choice.get("finish_reason").and_then(|v| v.as_str()) {
+            if let Some(finish_reason) = choice.get("finish_reason").and_then(|v| v.as_str())
+                && !finish_reason.is_empty()
+            {
                 match finish_reason {
                     "tool_calls" if fn_call_state.active => {
                         // First, flush the terminal raw reasoning so UIs can finalize

codex-rs/core/src/client.rs

@@ -26,6 +26,7 @@ use tokio::sync::mpsc;
 use tokio::time::timeout;
 use tokio_util::io::ReaderStream;
 use tracing::debug;
+use tracing::enabled;
 use tracing::trace;
 use tracing::warn;
 
@@ -78,6 +79,18 @@ struct Error {
     resets_at: Option<i64>,
 }
 
+#[derive(Debug, Serialize)]
+struct CompactHistoryRequest<'a> {
+    model: &'a str,
+    input: &'a [ResponseItem],
+    instructions: &'a str,
+}
+
+#[derive(Debug, Deserialize)]
+struct CompactHistoryResponse {
+    output: Vec<ResponseItem>,
+}
+
 #[derive(Debug, Clone)]
 pub struct ModelClient {
     config: Arc<Config>,
@@ -507,6 +520,70 @@ impl ModelClient {
     pub fn get_auth_manager(&self) -> Option<Arc<AuthManager>> {
         self.auth_manager.clone()
     }
+
+    pub async fn compact_conversation_history(&self, prompt: &Prompt) -> Result<Vec<ResponseItem>> {
+        if prompt.input.is_empty() {
+            return Ok(Vec::new());
+        }
+        let auth_manager = self.auth_manager.clone();
+        let auth = auth_manager.as_ref().and_then(|m| m.auth());
+        let mut req_builder = self
+            .provider
+            .create_compact_request_builder(&self.client, &auth)
+            .await?;
+        if let SessionSource::SubAgent(sub) = &self.session_source {
+            let subagent = if let crate::protocol::SubAgentSource::Other(label) = sub {
+                label.clone()
+            } else {
+                serde_json::to_value(sub)
+                    .ok()
+                    .and_then(|v| v.as_str().map(std::string::ToString::to_string))
+                    .unwrap_or_else(|| "other".to_string())
+            };
+            req_builder = req_builder.header("x-openai-subagent", subagent);
+        }
+        if let Some(auth) = auth.as_ref()
+            && auth.mode == AuthMode::ChatGPT
+            && let Some(account_id) = auth.get_account_id()
+        {
+            req_builder = req_builder.header("chatgpt-account-id", account_id);
+        }
+        let payload = CompactHistoryRequest {
+            model: &self.config.model,
+            input: &prompt.input,
+            instructions: &prompt.get_full_instructions(&self.config.model_family),
+        };
+
+        if enabled!(tracing::Level::TRACE) {
+            trace!(
+                "POST to {}: {}",
+                self.provider
+                    .get_compact_url(&auth)
+                    .unwrap_or("<none>".to_string()),
+                serde_json::to_value(&payload).unwrap_or_default()
+            );
+        }
+
+        let response = req_builder
+            .json(&payload)
+            .send()
+            .await
+            .map_err(|source| CodexErr::ConnectionFailed(ConnectionFailedError { source }))?;
+        let status = response.status();
+        let body = response
+            .text()
+            .await
+            .map_err(|source| CodexErr::ConnectionFailed(ConnectionFailedError { source }))?;
+        if !status.is_success() {
+            return Err(CodexErr::UnexpectedStatus(UnexpectedResponseError {
+                status,
+                body,
+                request_id: None,
+            }));
+        }
+        let CompactHistoryResponse { output } = serde_json::from_str(&body)?;
+        Ok(output)
+    }
 }
 
 enum StreamAttemptError {
@@ -1225,7 +1302,7 @@ mod tests {
 
     #[tokio::test]
     async fn error_when_error_event() {
-        let raw_error = r#"{"type":"response.failed","sequence_number":3,"response":{"id":"resp_689bcf18d7f08194bf3440ba62fe05d803fee0cdac429894","object":"response","created_at":1755041560,"status":"failed","background":false,"error":{"code":"rate_limit_exceeded","message":"Rate limit reached for gpt-5 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."}, "usage":null,"user":null,"metadata":{}}}"#;
+        let raw_error = r#"{"type":"response.failed","sequence_number":3,"response":{"id":"resp_689bcf18d7f08194bf3440ba62fe05d803fee0cdac429894","object":"response","created_at":1755041560,"status":"failed","background":false,"error":{"code":"rate_limit_exceeded","message":"Rate limit reached for gpt-5.1 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."}, "usage":null,"user":null,"metadata":{}}}"#;
 
         let sse1 = format!("event: response.failed\ndata: {raw_error}\n\n");
         let provider = ModelProviderInfo {
@@ -1254,7 +1331,7 @@ mod tests {
             Err(CodexErr::Stream(msg, delay)) => {
                 assert_eq!(
                     msg,
-                    "Rate limit reached for gpt-5 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."
+                    "Rate limit reached for gpt-5.1 in organization org-AAA on tokens per min (TPM): Limit 30000, Used 22999, Requested 12528. Please try again in 11.054s. Visit https://platform.openai.com/account/rate-limits to learn more."
                 );
                 assert_eq!(*delay, Some(Duration::from_secs_f64(11.054)));
             }
@@ -1473,7 +1550,7 @@ mod tests {
     fn test_try_parse_retry_after() {
         let err = Error {
             r#type: None,
-            message: Some("Rate limit reached for gpt-5 in organization org- on tokens per min (TPM): Limit 1, Used 1, Requested 19304. Please try again in 28ms. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
+            message: Some("Rate limit reached for gpt-5.1 in organization org- on tokens per min (TPM): Limit 1, Used 1, Requested 19304. Please try again in 28ms. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
             code: Some("rate_limit_exceeded".to_string()),
             plan_type: None,
             resets_at: None
@@ -1487,7 +1564,7 @@ mod tests {
     fn test_try_parse_retry_after_no_delay() {
         let err = Error {
             r#type: None,
-            message: Some("Rate limit reached for gpt-5 in organization <ORG> on tokens per min (TPM): Limit 30000, Used 6899, Requested 24050. Please try again in 1.898s. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
+            message: Some("Rate limit reached for gpt-5.1 in organization <ORG> on tokens per min (TPM): Limit 30000, Used 6899, Requested 24050. Please try again in 1.898s. Visit https://platform.openai.com/account/rate-limits to learn more.".to_string()),
             code: Some("rate_limit_exceeded".to_string()),
             plan_type: None,
             resets_at: None

codex-rs/core/src/client_common.rs

@@ -136,7 +136,7 @@ fn reserialize_shell_outputs(items: &mut [ResponseItem]) {
 }
 
 fn is_shell_tool_name(name: &str) -> bool {
-    matches!(name, "shell" | "container.exec")
+    matches!(name, "shell" | "container.exec" | "shell_command")
 }
 
 #[derive(Deserialize)]
@@ -427,7 +427,7 @@ mod tests {
                 expects_apply_patch_instructions: false,
             },
             InstructionsTestCase {
-                slug: "gpt-5-codex",
+                slug: "gpt-5.1-codex",
                 expects_apply_patch_instructions: false,
             },
             InstructionsTestCase {
@@ -457,7 +457,7 @@ mod tests {
         let input: Vec<ResponseItem> = vec![];
         let tools: Vec<serde_json::Value> = vec![];
         let req = ResponsesApiRequest {
-            model: "gpt-5",
+            model: "gpt-5.1",
             instructions: "i",
             input: &input,
             tools: &tools,
@@ -498,7 +498,7 @@ mod tests {
             create_text_param_for_request(None, &Some(schema.clone())).expect("text controls");
 
         let req = ResponsesApiRequest {
-            model: "gpt-5",
+            model: "gpt-5.1",
             instructions: "i",
             input: &input,
             tools: &tools,
@@ -534,7 +534,7 @@ mod tests {
         let input: Vec<ResponseItem> = vec![];
         let tools: Vec<serde_json::Value> = vec![];
         let req = ResponsesApiRequest {
-            model: "gpt-5",
+            model: "gpt-5.1",
             instructions: "i",
             input: &input,
             tools: &tools,

codex-rs/core/src/codex.rs

@@ -9,12 +9,11 @@ use crate::client_common::REVIEW_PROMPT;
 use crate::compact;
 use crate::features::Feature;
 use crate::function_tool::FunctionCallError;
-use crate::mcp::auth::McpAuthStatusEntry;
-use crate::mcp_connection_manager::DEFAULT_STARTUP_TIMEOUT;
 use crate::parse_command::parse_command;
 use crate::parse_turn_item;
 use crate::response_processing::process_items;
 use crate::terminal;
+use crate::truncate::TruncationPolicy;
 use crate::user_notification::UserNotifier;
 use crate::util::error_or_panic;
 use async_channel::Receiver;
@@ -45,6 +44,7 @@ use mcp_types::ReadResourceResult;
 use serde_json;
 use serde_json::Value;
 use tokio::sync::Mutex;
+use tokio::sync::RwLock;
 use tokio::sync::oneshot;
 use tokio_util::sync::CancellationToken;
 use tracing::debug;
@@ -56,8 +56,8 @@ use crate::ModelProviderInfo;
 use crate::client::ModelClient;
 use crate::client_common::Prompt;
 use crate::client_common::ResponseEvent;
+use crate::compact::collect_user_messages;
 use crate::config::Config;
-use crate::config::types::McpServerTransportConfig;
 use crate::config::types::ShellEnvironmentPolicy;
 use crate::context_manager::ContextManager;
 use crate::environment_context::EnvironmentContext;
@@ -65,10 +65,6 @@ use crate::error::CodexErr;
 use crate::error::Result as CodexResult;
 #[cfg(test)]
 use crate::exec::StreamOutput;
-// Removed: legacy executor wiring replaced by ToolOrchestrator flows.
-// legacy normalize_exec_result no longer used after orchestrator migration
-use crate::compact::build_compacted_history;
-use crate::compact::collect_user_messages;
 use crate::mcp::auth::compute_auth_statuses;
 use crate::mcp_connection_manager::McpConnectionManager;
 use crate::model_family::find_family_for_model;
@@ -122,7 +118,6 @@ use crate::user_instructions::UserInstructions;
 use crate::user_notification::UserNotification;
 use crate::util::backoff;
 use codex_async_utils::OrCancelExt;
-use codex_execpolicy2::Policy as ExecPolicyV2;
 use codex_otel::otel_event_manager::OtelEventManager;
 use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
 use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
@@ -134,6 +129,7 @@ use codex_protocol::protocol::InitialHistory;
 use codex_protocol::user_input::UserInput;
 use codex_utils_readiness::Readiness;
 use codex_utils_readiness::ReadinessFlag;
+use codex_utils_tokenizer::warm_model_cache;
 
 /// The high-level interface to the Codex system.
 /// It operates as a queue pair where you send submissions and receive events.
@@ -167,9 +163,6 @@ impl Codex {
 
         let user_instructions = get_user_instructions(&config).await;
 
-        let exec_policy_v2 = crate::exec_policy::exec_policy_for(&config.features, &config.cwd)
-            .map_err(|err| CodexErr::Fatal(format!("failed to load execpolicy2: {err}")))?;
-
         let config = Arc::new(config);
 
         let session_configuration = SessionConfiguration {
@@ -186,7 +179,6 @@ impl Codex {
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
             features: config.features.clone(),
-            exec_policy_v2,
             session_source,
         };
 
@@ -284,7 +276,7 @@ pub(crate) struct TurnContext {
     pub(crate) final_output_json_schema: Option<Value>,
     pub(crate) codex_linux_sandbox_exe: Option<PathBuf>,
     pub(crate) tool_call_gate: Arc<ReadinessFlag>,
-    pub(crate) exec_policy_v2: Option<Arc<ExecPolicyV2>>,
+    pub(crate) truncation_policy: TruncationPolicy,
 }
 
 impl TurnContext {
@@ -301,7 +293,6 @@ impl TurnContext {
     }
 }
 
-#[allow(dead_code)]
 #[derive(Clone)]
 pub(crate) struct SessionConfiguration {
     /// Provider identifier ("openai", "openrouter", ...).
@@ -341,8 +332,6 @@ pub(crate) struct SessionConfiguration {
 
     /// Set of feature flags for this session
     features: Features,
-    /// Optional execpolicy2 policy, applied only when enabled by feature flag.
-    exec_policy_v2: Option<Arc<ExecPolicyV2>>,
 
     // TODO(pakrym): Remove config from here
     original_config_do_not_use: Arc<Config>,
@@ -413,7 +402,7 @@ impl Session {
         );
 
         let client = ModelClient::new(
-            Arc::new(per_turn_config),
+            Arc::new(per_turn_config.clone()),
             auth_manager,
             otel_event_manager,
             provider,
@@ -443,7 +432,7 @@ impl Session {
             final_output_json_schema: None,
             codex_linux_sandbox_exe: config.codex_linux_sandbox_exe.clone(),
             tool_call_gate: Arc::new(ReadinessFlag::new()),
-            exec_policy_v2: session_configuration.exec_policy_v2.clone(),
+            truncation_policy: TruncationPolicy::new(&per_turn_config),
         }
     }
 
@@ -484,21 +473,13 @@ impl Session {
             ),
         };
 
-        // Error messages to dispatch after SessionConfigured is sent.
-        let mut post_session_configured_events = Vec::<Event>::new();
-
         // Kick off independent async setup tasks in parallel to reduce startup latency.
         //
         // - initialize RolloutRecorder with new or resumed session info
-        // - spin up MCP connection manager
         // - perform default shell discovery
         // - load history metadata
         let rollout_fut = RolloutRecorder::new(&config, rollout_params);
 
-        let mcp_fut = McpConnectionManager::new(
-            config.mcp_servers.clone(),
-            config.mcp_oauth_credentials_store_mode,
-        );
         let default_shell_fut = shell::default_user_shell();
         let history_meta_fut = crate::message_history::history_metadata(&config);
         let auth_statuses_fut = compute_auth_statuses(
@@ -507,15 +488,8 @@ impl Session {
         );
 
         // Join all independent futures.
-        let (
-            rollout_recorder,
-            mcp_res,
-            default_shell,
-            (history_log_id, history_entry_count),
-            auth_statuses,
-        ) = tokio::join!(
+        let (rollout_recorder, default_shell, (history_log_id, history_entry_count), auth_statuses) = tokio::join!(
             rollout_fut,
-            mcp_fut,
             default_shell_fut,
             history_meta_fut,
             auth_statuses_fut
@@ -527,34 +501,7 @@ impl Session {
         })?;
         let rollout_path = rollout_recorder.rollout_path.clone();
 
-        // Handle MCP manager result and record any startup failures.
-        let (mcp_connection_manager, failed_clients) = match mcp_res {
-            Ok((mgr, failures)) => (mgr, failures),
-            Err(e) => {
-                let message = format!("Failed to create MCP connection manager: {e:#}");
-                error!("{message}");
-                post_session_configured_events.push(Event {
-                    id: INITIAL_SUBMIT_ID.to_owned(),
-                    msg: EventMsg::Error(ErrorEvent { message }),
-                });
-                (McpConnectionManager::default(), Default::default())
-            }
-        };
-
-        // Surface individual client start-up failures to the user.
-        if !failed_clients.is_empty() {
-            for (server_name, err) in failed_clients {
-                let auth_entry = auth_statuses.get(&server_name);
-                let display_message = mcp_init_error_display(&server_name, auth_entry, &err);
-                warn!("MCP client for `{server_name}` failed to start: {err:#}");
-                post_session_configured_events.push(Event {
-                    id: INITIAL_SUBMIT_ID.to_owned(),
-                    msg: EventMsg::Error(ErrorEvent {
-                        message: display_message,
-                    }),
-                });
-            }
-        }
+        let mut post_session_configured_events = Vec::<Event>::new();
 
         for (alias, feature) in session_configuration.features.legacy_feature_usages() {
             let canonical = feature.key();
@@ -599,8 +546,12 @@ impl Session {
         // Create the mutable state for the Session.
         let state = SessionState::new(session_configuration.clone());
 
+        // Warm the tokenizer cache for the session model without blocking startup.
+        warm_model_cache(&session_configuration.model);
+
         let services = SessionServices {
-            mcp_connection_manager,
+            mcp_connection_manager: Arc::new(RwLock::new(McpConnectionManager::default())),
+            mcp_startup_cancellation_token: CancellationToken::new(),
             unified_exec_manager: UnifiedExecSessionManager::default(),
             notifier: UserNotifier::new(config.notify.clone()),
             rollout: Mutex::new(Some(rollout_recorder)),
@@ -629,6 +580,10 @@ impl Session {
             msg: EventMsg::SessionConfigured(SessionConfiguredEvent {
                 session_id: conversation_id,
                 model: session_configuration.model.clone(),
+                model_provider_id: config.model_provider_id.clone(),
+                approval_policy: session_configuration.approval_policy,
+                sandbox_policy: session_configuration.sandbox_policy.clone(),
+                cwd: session_configuration.cwd.clone(),
                 reasoning_effort: session_configuration.model_reasoning_effort,
                 history_log_id,
                 history_entry_count,
@@ -640,6 +595,18 @@ impl Session {
         for event in events {
             sess.send_event_raw(event).await;
         }
+        sess.services
+            .mcp_connection_manager
+            .write()
+            .await
+            .initialize(
+                config.mcp_servers.clone(),
+                config.mcp_oauth_credentials_store_mode,
+                auth_statuses.clone(),
+                tx_event.clone(),
+                sess.services.mcp_startup_cancellation_token.clone(),
+            )
+            .await;
 
         // record_initial_history can emit events. We record only after the SessionConfiguredEvent is emitted.
         sess.record_initial_history(initial_history).await;
@@ -717,7 +684,8 @@ impl Session {
                 let reconstructed_history =
                     self.reconstruct_history_from_rollout(&turn_context, &rollout_items);
                 if !reconstructed_history.is_empty() {
-                    self.record_into_history(&reconstructed_history).await;
+                    self.record_into_history(&reconstructed_history, &turn_context)
+                        .await;
                 }
 
                 // If persisting, persist all rollout items as-is (recorder filters)
@@ -899,6 +867,7 @@ impl Session {
         let parsed_cmd = parse_command(&command);
         let event = EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
             call_id,
+            turn_id: turn_context.sub_id.clone(),
             command,
             cwd,
             reason,
@@ -973,7 +942,7 @@ impl Session {
         turn_context: &TurnContext,
         items: &[ResponseItem],
     ) {
-        self.record_into_history(items).await;
+        self.record_into_history(items, turn_context).await;
         self.persist_rollout_response_items(items).await;
         self.send_raw_response_items(turn_context, items).await;
     }
@@ -987,17 +956,25 @@ impl Session {
         for item in rollout_items {
             match item {
                 RolloutItem::ResponseItem(response_item) => {
-                    history.record_items(std::iter::once(response_item));
+                    history.record_items(
+                        std::iter::once(response_item),
+                        turn_context.truncation_policy,
+                    );
                 }
                 RolloutItem::Compacted(compacted) => {
                     let snapshot = history.get_history();
-                    let user_messages = collect_user_messages(&snapshot);
-                    let rebuilt = build_compacted_history(
-                        self.build_initial_context(turn_context),
-                        &user_messages,
-                        &compacted.message,
-                    );
-                    history.replace(rebuilt);
+                    // TODO(jif) clean
+                    if let Some(replacement) = &compacted.replacement_history {
+                        history.replace(replacement.clone());
+                    } else {
+                        let user_messages = collect_user_messages(&snapshot);
+                        let rebuilt = compact::build_compacted_history(
+                            self.build_initial_context(turn_context),
+                            &user_messages,
+                            &compacted.message,
+                        );
+                        history.replace(rebuilt);
+                    }
                 }
                 _ => {}
             }
@@ -1006,9 +983,13 @@ impl Session {
     }
 
     /// Append ResponseItems to the in-memory conversation history only.
-    pub(crate) async fn record_into_history(&self, items: &[ResponseItem]) {
+    pub(crate) async fn record_into_history(
+        &self,
+        items: &[ResponseItem],
+        turn_context: &TurnContext,
+    ) {
         let mut state = self.state.lock().await;
-        state.record_items(items.iter());
+        state.record_items(items.iter(), turn_context.truncation_policy);
     }
 
     pub(crate) async fn replace_history(&self, items: Vec<ResponseItem>) {
@@ -1025,6 +1006,15 @@ impl Session {
         self.persist_rollout_items(&rollout_items).await;
     }
 
+    pub async fn enabled(&self, feature: Feature) -> bool {
+        self.state
+            .lock()
+            .await
+            .session_configuration
+            .features
+            .enabled(feature)
+    }
+
     async fn send_raw_response_items(&self, turn_context: &TurnContext, items: &[ResponseItem]) {
         for item in items {
             self.send_event(
@@ -1202,14 +1192,7 @@ impl Session {
         turn_context: Arc<TurnContext>,
         cancellation_token: CancellationToken,
     ) {
-        if !self
-            .state
-            .lock()
-            .await
-            .session_configuration
-            .features
-            .enabled(Feature::GhostCommit)
-        {
+        if !self.enabled(Feature::GhostCommit).await {
             return;
         }
         let token = match turn_context.tool_call_gate.subscribe().await {
@@ -1263,6 +1246,8 @@ impl Session {
     ) -> anyhow::Result<ListResourcesResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .list_resources(server, params)
             .await
     }
@@ -1274,6 +1259,8 @@ impl Session {
     ) -> anyhow::Result<ListResourceTemplatesResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .list_resource_templates(server, params)
             .await
     }
@@ -1285,6 +1272,8 @@ impl Session {
     ) -> anyhow::Result<ReadResourceResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .read_resource(server, params)
             .await
     }
@@ -1297,19 +1286,29 @@ impl Session {
     ) -> anyhow::Result<CallToolResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .call_tool(server, tool, arguments)
             .await
     }
 
-    pub(crate) fn parse_mcp_tool_name(&self, tool_name: &str) -> Option<(String, String)> {
+    pub(crate) async fn parse_mcp_tool_name(&self, tool_name: &str) -> Option<(String, String)> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .parse_tool_name(tool_name)
+            .await
     }
 
     pub async fn interrupt_task(self: &Arc<Self>) {
         info!("interrupt received: abort current task, if any");
-        self.abort_all_tasks(TurnAbortReason::Interrupted).await;
+        let has_active_turn = { self.active_turn.lock().await.is_some() };
+        if has_active_turn {
+            self.abort_all_tasks(TurnAbortReason::Interrupted).await;
+        } else {
+            self.cancel_mcp_startup().await;
+        }
     }
 
     pub(crate) fn notifier(&self) -> &UserNotifier {
@@ -1323,6 +1322,10 @@ impl Session {
     fn show_raw_agent_reasoning(&self) -> bool {
         self.services.show_raw_agent_reasoning
     }
+
+    async fn cancel_mcp_startup(&self) {
+        self.services.mcp_startup_cancellation_token.cancel();
+    }
 }
 
 async fn submission_loop(sess: Arc<Session>, config: Arc<Config>, rx_sub: Receiver<Submission>) {
@@ -1430,6 +1433,7 @@ mod handlers {
     use codex_protocol::protocol::ReviewDecision;
     use codex_protocol::protocol::ReviewRequest;
     use codex_protocol::protocol::TurnAbortReason;
+
     use codex_protocol::user_input::UserInput;
     use std::sync::Arc;
     use tracing::info;
@@ -1580,17 +1584,15 @@ mod handlers {
     }
 
     pub async fn list_mcp_tools(sess: &Session, config: &Arc<Config>, sub_id: String) {
-        // This is a cheap lookup from the connection manager's cache.
-        let tools = sess.services.mcp_connection_manager.list_all_tools();
-        let (auth_status_entries, resources, resource_templates) = tokio::join!(
+        let mcp_connection_manager = sess.services.mcp_connection_manager.read().await;
+        let (tools, auth_status_entries, resources, resource_templates) = tokio::join!(
+            mcp_connection_manager.list_all_tools(),
             compute_auth_statuses(
                 config.mcp_servers.iter(),
                 config.mcp_oauth_credentials_store_mode,
             ),
-            sess.services.mcp_connection_manager.list_all_resources(),
-            sess.services
-                .mcp_connection_manager
-                .list_all_resource_templates()
+            mcp_connection_manager.list_all_resources(),
+            mcp_connection_manager.list_all_resource_templates(),
         );
         let auth_statuses = auth_status_entries
             .iter()
@@ -1599,7 +1601,10 @@ mod handlers {
         let event = Event {
             id: sub_id,
             msg: EventMsg::McpListToolsResponse(crate::protocol::McpListToolsResponseEvent {
-                tools,
+                tools: tools
+                    .into_iter()
+                    .map(|(name, tool)| (name, tool.tool))
+                    .collect(),
                 resources,
                 resource_templates,
                 auth_statuses,
@@ -1637,16 +1642,15 @@ mod handlers {
         let turn_context = sess
             .new_turn_with_sub_id(sub_id, SessionSettingsUpdate::default())
             .await;
-        // Attempt to inject input into current task
-        if let Err(items) = sess
-            .inject_input(vec![UserInput::Text {
+
+        sess.spawn_task(
+            Arc::clone(&turn_context),
+            vec![UserInput::Text {
                 text: turn_context.compact_prompt().to_string(),
-            }])
-            .await
-        {
-            sess.spawn_task(Arc::clone(&turn_context), items, CompactTask)
-                .await;
-        }
+            }],
+            CompactTask,
+        )
+        .await;
     }
 
     pub async fn shutdown(sess: &Arc<Session>, sub_id: String) -> bool {
@@ -1772,7 +1776,7 @@ async fn spawn_review_thread(
         final_output_json_schema: None,
         codex_linux_sandbox_exe: parent_turn_context.codex_linux_sandbox_exe.clone(),
         tool_call_gate: Arc::new(ReadinessFlag::new()),
-        exec_policy_v2: parent_turn_context.exec_policy_v2.clone(),
+        truncation_policy: TruncationPolicy::new(&per_turn_config),
     };
 
     // Seed the child task with the review prompt as the initial user message.
@@ -1780,7 +1784,12 @@ async fn spawn_review_thread(
         text: review_prompt,
     }];
     let tc = Arc::new(review_turn_context);
-    sess.spawn_task(tc.clone(), input, ReviewTask).await;
+    sess.spawn_task(
+        tc.clone(),
+        input,
+        ReviewTask::new(review_request.append_to_original_thread),
+    )
+    .await;
 
     // Announce entering review mode so UIs can switch modes.
     sess.send_event(&tc, EventMsg::EnteredReviewMode(review_request))
@@ -1825,7 +1834,6 @@ pub(crate) async fn run_task(
     // Although from the perspective of codex.rs, TurnDiffTracker has the lifecycle of a Task which contains
     // many turns, from the perspective of the user, it is a single turn.
     let turn_diff_tracker = Arc::new(tokio::sync::Mutex::new(TurnDiffTracker::new()));
-    let mut auto_compact_recently_attempted = false;
 
     loop {
         // Note that pending_input would be something like a message the user
@@ -1880,27 +1888,12 @@ pub(crate) async fn run_task(
                 let (responses, items_to_record_in_conversation_history) =
                     process_items(processed_items, &sess, &turn_context).await;
 
+                // as long as compaction works well in getting us way below the token limit, we shouldn't worry about being in an infinite loop.
                 if token_limit_reached {
-                    if auto_compact_recently_attempted {
-                        let limit_str = limit.to_string();
-                        let current_tokens = total_usage_tokens
-                            .map(|tokens| tokens.to_string())
-                            .unwrap_or_else(|| "unknown".to_string());
-                        let event = EventMsg::Error(ErrorEvent {
-                            message: format!(
-                                "Conversation is still above the token limit after automatic summarization (limit {limit_str}, current {current_tokens}). Please start a new session or trim your input."
-                            ),
-                        });
-                        sess.send_event(&turn_context, event).await;
-                        break;
-                    }
-                    auto_compact_recently_attempted = true;
                     compact::run_inline_auto_compact_task(sess.clone(), turn_context.clone()).await;
                     continue;
                 }
 
-                auto_compact_recently_attempted = false;
-
                 if responses.is_empty() {
                     last_agent_message = get_last_assistant_message_from_turn(
                         &items_to_record_in_conversation_history,
@@ -1946,22 +1939,52 @@ async fn run_turn(
     input: Vec<ResponseItem>,
     cancellation_token: CancellationToken,
 ) -> CodexResult<TurnRunResult> {
-    let mcp_tools = sess.services.mcp_connection_manager.list_all_tools();
+    let mcp_tools = sess
+        .services
+        .mcp_connection_manager
+        .read()
+        .await
+        .list_all_tools()
+        .or_cancel(&cancellation_token)
+        .await?;
     let router = Arc::new(ToolRouter::from_config(
         &turn_context.tools_config,
-        Some(mcp_tools),
+        Some(
+            mcp_tools
+                .into_iter()
+                .map(|(name, tool)| (name, tool.tool))
+                .collect(),
+        ),
     ));
 
     let model_supports_parallel = turn_context
         .client
         .get_model_family()
         .supports_parallel_tool_calls;
-    let parallel_tool_calls = model_supports_parallel;
+
+    // TODO(jif) revert once testing phase is done.
+    let parallel_tool_calls = model_supports_parallel
+        && sess
+            .state
+            .lock()
+            .await
+            .session_configuration
+            .features
+            .enabled(Feature::ParallelToolCalls);
+    let mut base_instructions = turn_context.base_instructions.clone();
+    if parallel_tool_calls {
+        static INSTRUCTIONS: &str = include_str!("../templates/parallel/instructions.md");
+        static INSERTION_SPOT: &str = "## Editing constraints";
+        base_instructions
+            .as_mut()
+            .map(|base| base.replace(INSERTION_SPOT, INSTRUCTIONS));
+    }
+
     let prompt = Prompt {
         input,
         tools: router.specs(),
         parallel_tool_calls,
-        base_instructions_override: turn_context.base_instructions.clone(),
+        base_instructions_override: base_instructions,
         output_schema: turn_context.final_output_json_schema.clone(),
     };
 
@@ -2118,7 +2141,7 @@ async fn try_run_turn(
             ResponseEvent::Created => {}
             ResponseEvent::OutputItemDone(item) => {
                 let previously_active_item = active_item.take();
-                match ToolRouter::build_tool_call(sess.as_ref(), item.clone()) {
+                match ToolRouter::build_tool_call(sess.as_ref(), item.clone()).await {
                     Ok(Some(call)) => {
                         let payload_preview = call.payload.log_payload().into_owned();
                         tracing::info!("ToolCall: {} {}", call.tool_name, payload_preview);
@@ -2329,59 +2352,6 @@ pub(super) fn get_last_assistant_message_from_turn(responses: &[ResponseItem]) -
     })
 }
 
-fn mcp_init_error_display(
-    server_name: &str,
-    entry: Option<&McpAuthStatusEntry>,
-    err: &anyhow::Error,
-) -> String {
-    if let Some(McpServerTransportConfig::StreamableHttp {
-        url,
-        bearer_token_env_var,
-        http_headers,
-        ..
-    }) = &entry.map(|entry| &entry.config.transport)
-        && url == "https://api.githubcopilot.com/mcp/"
-        && bearer_token_env_var.is_none()
-        && http_headers.as_ref().map(HashMap::is_empty).unwrap_or(true)
-    {
-        // GitHub only supports OAUth for first party MCP clients.
-        // That means that the user has to specify a personal access token either via bearer_token_env_var or http_headers.
-        // https://github.com/github/github-mcp-server/issues/921#issuecomment-3221026448
-        format!(
-            "GitHub MCP does not support OAuth. Log in by adding a personal access token (https://github.com/settings/personal-access-tokens) to your environment and config.toml:\n[mcp_servers.{server_name}]\nbearer_token_env_var = CODEX_GITHUB_PERSONAL_ACCESS_TOKEN"
-        )
-    } else if is_mcp_client_auth_required_error(err) {
-        format!(
-            "The {server_name} MCP server is not logged in. Run `codex mcp login {server_name}`."
-        )
-    } else if is_mcp_client_startup_timeout_error(err) {
-        let startup_timeout_secs = match entry {
-            Some(entry) => match entry.config.startup_timeout_sec {
-                Some(timeout) => timeout,
-                None => DEFAULT_STARTUP_TIMEOUT,
-            },
-            None => DEFAULT_STARTUP_TIMEOUT,
-        }
-        .as_secs();
-        format!(
-            "MCP client for `{server_name}` timed out after {startup_timeout_secs} seconds. Add or adjust `startup_timeout_sec` in your config.toml:\n[mcp_servers.{server_name}]\nstartup_timeout_sec = XX"
-        )
-    } else {
-        format!("MCP client for `{server_name}` failed to start: {err:#}")
-    }
-}
-
-fn is_mcp_client_auth_required_error(error: &anyhow::Error) -> bool {
-    // StreamableHttpError::AuthRequired from the MCP SDK.
-    error.to_string().contains("Auth required")
-}
-
-fn is_mcp_client_startup_timeout_error(error: &anyhow::Error) -> bool {
-    let error_message = error.to_string();
-    error_message.contains("request timed out")
-        || error_message.contains("timed out handshaking with MCP server")
-}
-
 use crate::features::Features;
 #[cfg(test)]
 pub(crate) use tests::make_session_and_context;
@@ -2391,10 +2361,7 @@ mod tests {
     use super::*;
     use crate::config::ConfigOverrides;
     use crate::config::ConfigToml;
-    use crate::config::types::McpServerConfig;
-    use crate::config::types::McpServerTransportConfig;
     use crate::exec::ExecToolCallOutput;
-    use crate::mcp::auth::McpAuthStatusEntry;
     use crate::tools::format_exec_output_str;
 
     use crate::protocol::CompactedItem;
@@ -2414,7 +2381,6 @@ mod tests {
     use codex_app_server_protocol::AuthMode;
     use codex_protocol::models::ContentItem;
     use codex_protocol::models::ResponseItem;
-    use codex_protocol::protocol::McpAuthStatus;
     use std::time::Duration;
     use tokio::time::sleep;
 
@@ -2622,14 +2588,14 @@ mod tests {
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
             features: Features::default(),
-            exec_policy_v2: None,
             session_source: SessionSource::Exec,
         };
 
         let state = SessionState::new(session_configuration.clone());
 
         let services = SessionServices {
-            mcp_connection_manager: McpConnectionManager::default(),
+            mcp_connection_manager: Arc::new(RwLock::new(McpConnectionManager::default())),
+            mcp_startup_cancellation_token: CancellationToken::new(),
             unified_exec_manager: UnifiedExecSessionManager::default(),
             notifier: UserNotifier::new(None),
             rollout: Mutex::new(None),
@@ -2699,14 +2665,14 @@ mod tests {
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
             features: Features::default(),
-            exec_policy_v2: None,
             session_source: SessionSource::Exec,
         };
 
         let state = SessionState::new(session_configuration.clone());
 
         let services = SessionServices {
-            mcp_connection_manager: McpConnectionManager::default(),
+            mcp_connection_manager: Arc::new(RwLock::new(McpConnectionManager::default())),
+            mcp_startup_cancellation_token: CancellationToken::new(),
             unified_exec_manager: UnifiedExecSessionManager::default(),
             notifier: UserNotifier::new(None),
             rollout: Mutex::new(None),
@@ -2829,7 +2795,8 @@ mod tests {
         let input = vec![UserInput::Text {
             text: "start review".to_string(),
         }];
-        sess.spawn_task(Arc::clone(&tc), input, ReviewTask).await;
+        sess.spawn_task(Arc::clone(&tc), input, ReviewTask::new(true))
+            .await;
 
         sess.abort_all_tasks(TurnAbortReason::Interrupted).await;
 
@@ -2887,9 +2854,23 @@ mod tests {
     #[tokio::test]
     async fn fatal_tool_error_stops_turn_and_reports_error() {
         let (session, turn_context, _rx) = make_session_and_context_with_rx();
+        let tools = {
+            session
+                .services
+                .mcp_connection_manager
+                .read()
+                .await
+                .list_all_tools()
+                .await
+        };
         let router = ToolRouter::from_config(
             &turn_context.tools_config,
-            Some(session.services.mcp_connection_manager.list_all_tools()),
+            Some(
+                tools
+                    .into_iter()
+                    .map(|(name, tool)| (name, tool.tool))
+                    .collect(),
+            ),
         );
         let item = ResponseItem::CustomToolCall {
             id: None,
@@ -2900,6 +2881,7 @@ mod tests {
         };
 
         let call = ToolRouter::build_tool_call(session.as_ref(), item.clone())
+            .await
             .expect("build tool call")
             .expect("tool call present");
         let tracker = Arc::new(tokio::sync::Mutex::new(TurnDiffTracker::new()));
@@ -2932,7 +2914,7 @@ mod tests {
         for item in &initial_context {
             rollout_items.push(RolloutItem::ResponseItem(item.clone()));
         }
-        live_history.record_items(initial_context.iter());
+        live_history.record_items(initial_context.iter(), turn_context.truncation_policy);
 
         let user1 = ResponseItem::Message {
             id: None,
@@ -2941,7 +2923,7 @@ mod tests {
                 text: "first user".to_string(),
             }],
         };
-        live_history.record_items(std::iter::once(&user1));
+        live_history.record_items(std::iter::once(&user1), turn_context.truncation_policy);
         rollout_items.push(RolloutItem::ResponseItem(user1.clone()));
 
         let assistant1 = ResponseItem::Message {
@@ -2951,13 +2933,13 @@ mod tests {
                 text: "assistant reply one".to_string(),
             }],
         };
-        live_history.record_items(std::iter::once(&assistant1));
+        live_history.record_items(std::iter::once(&assistant1), turn_context.truncation_policy);
         rollout_items.push(RolloutItem::ResponseItem(assistant1.clone()));
 
         let summary1 = "summary one";
         let snapshot1 = live_history.get_history();
         let user_messages1 = collect_user_messages(&snapshot1);
-        let rebuilt1 = build_compacted_history(
+        let rebuilt1 = compact::build_compacted_history(
             session.build_initial_context(turn_context),
             &user_messages1,
             summary1,
@@ -2965,6 +2947,7 @@ mod tests {
         live_history.replace(rebuilt1);
         rollout_items.push(RolloutItem::Compacted(CompactedItem {
             message: summary1.to_string(),
+            replacement_history: None,
         }));
 
         let user2 = ResponseItem::Message {
@@ -2974,7 +2957,7 @@ mod tests {
                 text: "second user".to_string(),
             }],
         };
-        live_history.record_items(std::iter::once(&user2));
+        live_history.record_items(std::iter::once(&user2), turn_context.truncation_policy);
         rollout_items.push(RolloutItem::ResponseItem(user2.clone()));
 
         let assistant2 = ResponseItem::Message {
@@ -2984,13 +2967,13 @@ mod tests {
                 text: "assistant reply two".to_string(),
             }],
         };
-        live_history.record_items(std::iter::once(&assistant2));
+        live_history.record_items(std::iter::once(&assistant2), turn_context.truncation_policy);
         rollout_items.push(RolloutItem::ResponseItem(assistant2.clone()));
 
         let summary2 = "summary two";
         let snapshot2 = live_history.get_history();
         let user_messages2 = collect_user_messages(&snapshot2);
-        let rebuilt2 = build_compacted_history(
+        let rebuilt2 = compact::build_compacted_history(
             session.build_initial_context(turn_context),
             &user_messages2,
             summary2,
@@ -2998,6 +2981,7 @@ mod tests {
         live_history.replace(rebuilt2);
         rollout_items.push(RolloutItem::Compacted(CompactedItem {
             message: summary2.to_string(),
+            replacement_history: None,
         }));
 
         let user3 = ResponseItem::Message {
@@ -3007,7 +2991,7 @@ mod tests {
                 text: "third user".to_string(),
             }],
         };
-        live_history.record_items(std::iter::once(&user3));
+        live_history.record_items(std::iter::once(&user3), turn_context.truncation_policy);
         rollout_items.push(RolloutItem::ResponseItem(user3.clone()));
 
         let assistant3 = ResponseItem::Message {
@@ -3017,7 +3001,7 @@ mod tests {
                 text: "assistant reply three".to_string(),
             }],
         };
-        live_history.record_items(std::iter::once(&assistant3));
+        live_history.record_items(std::iter::once(&assistant3), turn_context.truncation_policy);
         rollout_items.push(RolloutItem::ResponseItem(assistant3.clone()));
 
         (rollout_items, live_history.get_history())
@@ -3149,7 +3133,6 @@ mod tests {
         pretty_assertions::assert_eq!(exec_output.metadata, ResponseExecMetadata { exit_code: 0 });
         assert!(exec_output.output.contains("hi"));
     }
-
     #[tokio::test]
     async fn unified_exec_rejects_escalated_permissions_when_policy_not_on_request() {
         use crate::protocol::AskForApproval;
@@ -3191,89 +3174,4 @@ mod tests {
 
         pretty_assertions::assert_eq!(output, expected);
     }
-
-    #[test]
-    fn mcp_init_error_display_prompts_for_github_pat() {
-        let server_name = "github";
-        let entry = McpAuthStatusEntry {
-            config: McpServerConfig {
-                transport: McpServerTransportConfig::StreamableHttp {
-                    url: "https://api.githubcopilot.com/mcp/".to_string(),
-                    bearer_token_env_var: None,
-                    http_headers: None,
-                    env_http_headers: None,
-                },
-                enabled: true,
-                startup_timeout_sec: None,
-                tool_timeout_sec: None,
-                enabled_tools: None,
-                disabled_tools: None,
-            },
-            auth_status: McpAuthStatus::Unsupported,
-        };
-        let err = anyhow::anyhow!("OAuth is unsupported");
-
-        let display = mcp_init_error_display(server_name, Some(&entry), &err);
-
-        let expected = format!(
-            "GitHub MCP does not support OAuth. Log in by adding a personal access token (https://github.com/settings/personal-access-tokens) to your environment and config.toml:\n[mcp_servers.{server_name}]\nbearer_token_env_var = CODEX_GITHUB_PERSONAL_ACCESS_TOKEN"
-        );
-
-        assert_eq!(expected, display);
-    }
-
-    #[test]
-    fn mcp_init_error_display_prompts_for_login_when_auth_required() {
-        let server_name = "example";
-        let err = anyhow::anyhow!("Auth required for server");
-
-        let display = mcp_init_error_display(server_name, None, &err);
-
-        let expected = format!(
-            "The {server_name} MCP server is not logged in. Run `codex mcp login {server_name}`."
-        );
-
-        assert_eq!(expected, display);
-    }
-
-    #[test]
-    fn mcp_init_error_display_reports_generic_errors() {
-        let server_name = "custom";
-        let entry = McpAuthStatusEntry {
-            config: McpServerConfig {
-                transport: McpServerTransportConfig::StreamableHttp {
-                    url: "https://example.com".to_string(),
-                    bearer_token_env_var: Some("TOKEN".to_string()),
-                    http_headers: None,
-                    env_http_headers: None,
-                },
-                enabled: true,
-                startup_timeout_sec: None,
-                tool_timeout_sec: None,
-                enabled_tools: None,
-                disabled_tools: None,
-            },
-            auth_status: McpAuthStatus::Unsupported,
-        };
-        let err = anyhow::anyhow!("boom");
-
-        let display = mcp_init_error_display(server_name, Some(&entry), &err);
-
-        let expected = format!("MCP client for `{server_name}` failed to start: {err:#}");
-
-        assert_eq!(expected, display);
-    }
-
-    #[test]
-    fn mcp_init_error_display_includes_startup_timeout_hint() {
-        let server_name = "slow";
-        let err = anyhow::anyhow!("request timed out");
-
-        let display = mcp_init_error_display(server_name, None, &err);
-
-        assert_eq!(
-            "MCP client for `slow` timed out after 10 seconds. Add or adjust `startup_timeout_sec` in your config.toml:\n[mcp_servers.slow]\nstartup_timeout_sec = XX",
-            display
-        );
-    }
 }

codex-rs/core/src/compact.rs

@@ -14,7 +14,9 @@ use crate::protocol::EventMsg;
 use crate::protocol::TaskStartedEvent;
 use crate::protocol::TurnContextItem;
 use crate::protocol::WarningEvent;
-use crate::truncate::truncate_middle;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::approx_token_count;
+use crate::truncate::truncate_text;
 use crate::util::backoff;
 use codex_protocol::items::TurnItem;
 use codex_protocol::models::ContentItem;
@@ -26,6 +28,7 @@ use futures::prelude::*;
 use tracing::error;
 
 pub const SUMMARIZATION_PROMPT: &str = include_str!("../templates/compact/prompt.md");
+pub const SUMMARY_PREFIX: &str = include_str!("../templates/compact/summary_prefix.md");
 const COMPACT_USER_MESSAGE_MAX_TOKENS: usize = 20_000;
 
 pub(crate) async fn run_inline_auto_compact_task(
@@ -58,7 +61,10 @@ async fn run_compact_task_inner(
     let initial_input_for_turn: ResponseInputItem = ResponseInputItem::from(input);
 
     let mut history = sess.clone_history().await;
-    history.record_items(&[initial_input_for_turn.into()]);
+    history.record_items(
+        &[initial_input_for_turn.into()],
+        turn_context.truncation_policy,
+    );
 
     let mut truncated_count = 0usize;
 
@@ -140,7 +146,9 @@ async fn run_compact_task_inner(
     }
 
     let history_snapshot = sess.clone_history().await.get_history();
-    let summary_text = get_last_assistant_message_from_turn(&history_snapshot).unwrap_or_default();
+    let summary_suffix =
+        get_last_assistant_message_from_turn(&history_snapshot).unwrap_or_default();
+    let summary_text = format!("{SUMMARY_PREFIX}\n{summary_suffix}");
     let user_messages = collect_user_messages(&history_snapshot);
 
     let initial_context = sess.build_initial_context(turn_context.as_ref());
@@ -164,6 +172,7 @@ async fn run_compact_task_inner(
 
     let rollout_item = RolloutItem::Compacted(CompactedItem {
         message: summary_text.clone(),
+        replacement_history: None,
     });
     sess.persist_rollout_items(&[rollout_item]).await;
 
@@ -201,12 +210,22 @@ pub(crate) fn collect_user_messages(items: &[ResponseItem]) -> Vec<String> {
     items
         .iter()
         .filter_map(|item| match crate::event_mapping::parse_turn_item(item) {
-            Some(TurnItem::UserMessage(user)) => Some(user.message()),
+            Some(TurnItem::UserMessage(user)) => {
+                if is_summary_message(&user.message()) {
+                    None
+                } else {
+                    Some(user.message())
+                }
+            }
             _ => None,
         })
         .collect()
 }
 
+pub(crate) fn is_summary_message(message: &str) -> bool {
+    message.starts_with(format!("{SUMMARY_PREFIX}\n").as_str())
+}
+
 pub(crate) fn build_compacted_history(
     initial_context: Vec<ResponseItem>,
     user_messages: &[String],
@@ -216,7 +235,7 @@ pub(crate) fn build_compacted_history(
         initial_context,
         user_messages,
         summary_text,
-        COMPACT_USER_MESSAGE_MAX_TOKENS * 4,
+        COMPACT_USER_MESSAGE_MAX_TOKENS,
     )
 }
 
@@ -224,20 +243,21 @@ fn build_compacted_history_with_limit(
     mut history: Vec<ResponseItem>,
     user_messages: &[String],
     summary_text: &str,
-    max_bytes: usize,
+    max_tokens: usize,
 ) -> Vec<ResponseItem> {
     let mut selected_messages: Vec<String> = Vec::new();
-    if max_bytes > 0 {
-        let mut remaining = max_bytes;
+    if max_tokens > 0 {
+        let mut remaining = max_tokens;
         for message in user_messages.iter().rev() {
             if remaining == 0 {
                 break;
             }
-            if message.len() <= remaining {
+            let tokens = approx_token_count(message);
+            if tokens <= remaining {
                 selected_messages.push(message.clone());
-                remaining = remaining.saturating_sub(message.len());
+                remaining = remaining.saturating_sub(tokens);
             } else {
-                let (truncated, _) = truncate_middle(message, remaining);
+                let truncated = truncate_text(message, TruncationPolicy::Tokens(remaining));
                 selected_messages.push(truncated);
                 break;
             }
@@ -286,7 +306,8 @@ async fn drain_to_completed(
         };
         match event {
             Ok(ResponseEvent::OutputItemDone(item)) => {
-                sess.record_into_history(std::slice::from_ref(&item)).await;
+                sess.record_into_history(std::slice::from_ref(&item), turn_context)
+                    .await;
             }
             Ok(ResponseEvent::RateLimits(snapshot)) => {
                 sess.update_rate_limits(turn_context, snapshot).await;
@@ -304,6 +325,7 @@ async fn drain_to_completed(
 
 #[cfg(test)]
 mod tests {
+
     use super::*;
     use pretty_assertions::assert_eq;
 
@@ -395,16 +417,16 @@ mod tests {
     }
 
     #[test]
-    fn build_compacted_history_truncates_overlong_user_messages() {
+    fn build_token_limited_compacted_history_truncates_overlong_user_messages() {
         // Use a small truncation limit so the test remains fast while still validating
         // that oversized user content is truncated.
-        let max_bytes = 128;
-        let big = "X".repeat(max_bytes + 50);
+        let max_tokens = 16;
+        let big = "word ".repeat(200);
         let history = super::build_compacted_history_with_limit(
             Vec::new(),
             std::slice::from_ref(&big),
             "SUMMARY",
-            max_bytes,
+            max_tokens,
         );
         assert_eq!(history.len(), 2);
 
@@ -437,7 +459,7 @@ mod tests {
     }
 
     #[test]
-    fn build_compacted_history_appends_summary_message() {
+    fn build_token_limited_compacted_history_appends_summary_message() {
         let initial_context: Vec<ResponseItem> = Vec::new();
         let user_messages = vec!["first user message".to_string()];
         let summary_text = "summary text";

codex-rs/core/src/compact_remote.rs

@@ -0,0 +1,88 @@
+use std::sync::Arc;
+
+use crate::Prompt;
+use crate::codex::Session;
+use crate::codex::TurnContext;
+use crate::error::Result as CodexResult;
+use crate::protocol::AgentMessageEvent;
+use crate::protocol::CompactedItem;
+use crate::protocol::ErrorEvent;
+use crate::protocol::EventMsg;
+use crate::protocol::RolloutItem;
+use crate::protocol::TaskStartedEvent;
+use codex_protocol::models::ResponseItem;
+
+pub(crate) async fn run_remote_compact_task(
+    sess: Arc<Session>,
+    turn_context: Arc<TurnContext>,
+) -> Option<String> {
+    let start_event = EventMsg::TaskStarted(TaskStartedEvent {
+        model_context_window: turn_context.client.get_model_context_window(),
+    });
+    sess.send_event(&turn_context, start_event).await;
+
+    match run_remote_compact_task_inner(&sess, &turn_context).await {
+        Ok(()) => {
+            let event = EventMsg::AgentMessage(AgentMessageEvent {
+                message: "Compact task completed".to_string(),
+            });
+            sess.send_event(&turn_context, event).await;
+        }
+        Err(err) => {
+            let event = EventMsg::Error(ErrorEvent {
+                message: err.to_string(),
+            });
+            sess.send_event(&turn_context, event).await;
+        }
+    }
+
+    None
+}
+
+async fn run_remote_compact_task_inner(
+    sess: &Arc<Session>,
+    turn_context: &Arc<TurnContext>,
+) -> CodexResult<()> {
+    let mut history = sess.clone_history().await;
+    let prompt = Prompt {
+        input: history.get_history_for_prompt(),
+        tools: vec![],
+        parallel_tool_calls: false,
+        base_instructions_override: turn_context.base_instructions.clone(),
+        output_schema: None,
+    };
+
+    let mut new_history = turn_context
+        .client
+        .compact_conversation_history(&prompt)
+        .await?;
+    // Required to keep `/undo` available after compaction
+    let ghost_snapshots: Vec<ResponseItem> = history
+        .get_history()
+        .iter()
+        .filter(|item| matches!(item, ResponseItem::GhostSnapshot { .. }))
+        .cloned()
+        .collect();
+
+    if !ghost_snapshots.is_empty() {
+        new_history.extend(ghost_snapshots);
+    }
+    sess.replace_history(new_history.clone()).await;
+
+    if let Some(estimated_tokens) = sess
+        .clone_history()
+        .await
+        .estimate_token_count(turn_context.as_ref())
+    {
+        sess.override_last_token_usage_estimate(turn_context.as_ref(), estimated_tokens)
+            .await;
+    }
+
+    let compacted_item = CompactedItem {
+        message: String::new(),
+        replacement_history: Some(new_history),
+    };
+    sess.persist_rollout_items(&[RolloutItem::Compacted(compacted_item)])
+        .await;
+    Ok(())
+}

codex-rs/core/src/config/edit.rs

@@ -3,6 +3,8 @@ use crate::config::types::McpServerConfig;
 use crate::config::types::Notice;
 use anyhow::Context;
 use codex_protocol::config_types::ReasoningEffort;
+use codex_protocol::config_types::TrustLevel;
+use codex_utils_tokenizer::warm_model_cache;
 use std::collections::BTreeMap;
 use std::path::Path;
 use std::path::PathBuf;
@@ -33,9 +35,9 @@ pub enum ConfigEdit {
     SetNoticeHideModelMigrationPrompt(String, bool),
     /// Replace the entire `[mcp_servers]` table.
     ReplaceMcpServers(BTreeMap<String, McpServerConfig>),
-    /// Set trust_level = "trusted" under `[projects."<path>"]`,
+    /// Set trust_level under `[projects."<path>"]`,
     /// migrating inline tables to explicit tables.
-    SetProjectTrusted(PathBuf),
+    SetProjectTrustLevel { path: PathBuf, level: TrustLevel },
     /// Set the value stored at the exact dotted path.
     SetPath {
         segments: Vec<String>,
@@ -229,6 +231,9 @@ impl ConfigDocument {
     fn apply(&mut self, edit: &ConfigEdit) -> anyhow::Result<bool> {
         match edit {
             ConfigEdit::SetModel { model, effort } => Ok({
+                if let Some(model) = &model {
+                    warm_model_cache(model)
+                }
                 let mut mutated = false;
                 mutated |= self.write_profile_value(
                     &["model"],
@@ -270,10 +275,14 @@ impl ConfigDocument {
             ConfigEdit::ReplaceMcpServers(servers) => Ok(self.replace_mcp_servers(servers)),
             ConfigEdit::SetPath { segments, value } => Ok(self.insert(segments, value.clone())),
             ConfigEdit::ClearPath { segments } => Ok(self.clear_owned(segments)),
-            ConfigEdit::SetProjectTrusted(project_path) => {
+            ConfigEdit::SetProjectTrustLevel { path, level } => {
                 // Delegate to the existing, tested logic in config.rs to
                 // ensure tables are explicit and migration is preserved.
-                crate::config::set_project_trusted_inner(&mut self.doc, project_path.as_path())?;
+                crate::config::set_project_trust_level_inner(
+                    &mut self.doc,
+                    path.as_path(),
+                    *level,
+                )?;
                 Ok(true)
             }
         }
@@ -529,9 +538,24 @@ impl ConfigEditsBuilder {
         self
     }
 
-    pub fn set_project_trusted<P: Into<PathBuf>>(mut self, project_path: P) -> Self {
-        self.edits
-            .push(ConfigEdit::SetProjectTrusted(project_path.into()));
+    pub fn set_project_trust_level<P: Into<PathBuf>>(
+        mut self,
+        project_path: P,
+        trust_level: TrustLevel,
+    ) -> Self {
+        self.edits.push(ConfigEdit::SetProjectTrustLevel {
+            path: project_path.into(),
+            level: trust_level,
+        });
+        self
+    }
+
+    /// Enable or disable a feature flag by key under the `[features]` table.
+    pub fn set_feature_enabled(mut self, key: &str, enabled: bool) -> Self {
+        self.edits.push(ConfigEdit::SetPath {
+            segments: vec!["features".to_string(), key.to_string()],
+            value: value(enabled),
+        });
         self
     }
 
@@ -569,7 +593,7 @@ mod tests {
             codex_home,
             None,
             &[ConfigEdit::SetModel {
-                model: Some("gpt-5-codex".to_string()),
+                model: Some("gpt-5.1-codex".to_string()),
                 effort: Some(ReasoningEffort::High),
             }],
         )
@@ -577,7 +601,7 @@ mod tests {
 
         let contents =
             std::fs::read_to_string(codex_home.join(CONFIG_TOML_FILE)).expect("read config");
-        let expected = r#"model = "gpt-5-codex"
+        let expected = r#"model = "gpt-5.1-codex"
 model_reasoning_effort = "high"
 "#;
         assert_eq!(contents, expected);
@@ -707,7 +731,7 @@ model = "o5-preview"
         std::fs::write(
             codex_home.join(CONFIG_TOML_FILE),
             r#"[profiles."team a"]
-model = "gpt-5-codex"
+model = "gpt-5.1-codex"
 "#,
         )
         .expect("seed");
@@ -957,14 +981,14 @@ B = \"2\"
         let codex_home = tmp.path().to_path_buf();
 
         ConfigEditsBuilder::new(&codex_home)
-            .set_model(Some("gpt-5-codex"), Some(ReasoningEffort::High))
+            .set_model(Some("gpt-5.1-codex"), Some(ReasoningEffort::High))
             .apply()
             .await
             .expect("persist");
 
         let contents =
             std::fs::read_to_string(codex_home.join(CONFIG_TOML_FILE)).expect("read config");
-        let expected = r#"model = "gpt-5-codex"
+        let expected = r#"model = "gpt-5.1-codex"
 model_reasoning_effort = "high"
 "#;
         assert_eq!(contents, expected);
@@ -986,11 +1010,11 @@ model_reasoning_effort = "low"
             std::fs::read_to_string(codex_home.join(CONFIG_TOML_FILE)).expect("read config");
         assert_eq!(contents, initial_expected);
 
-        let updated_expected = r#"model = "gpt-5-codex"
+        let updated_expected = r#"model = "gpt-5.1-codex"
 model_reasoning_effort = "high"
 "#;
         ConfigEditsBuilder::new(codex_home)
-            .set_model(Some("gpt-5-codex"), Some(ReasoningEffort::High))
+            .set_model(Some("gpt-5.1-codex"), Some(ReasoningEffort::High))
             .apply_blocking()
             .expect("persist update");
         contents = std::fs::read_to_string(codex_home.join(CONFIG_TOML_FILE)).expect("read config");

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

@@ -25,7 +25,9 @@ use crate::git_info::resolve_root_git_project_for_trust;
 use crate::model_family::ModelFamily;
 use crate::model_family::derive_default_model_family;
 use crate::model_family::find_family_for_model;
+use crate::model_provider_info::LMSTUDIO_OSS_PROVIDER_ID;
 use crate::model_provider_info::ModelProviderInfo;
+use crate::model_provider_info::OLLAMA_OSS_PROVIDER_ID;
 use crate::model_provider_info::built_in_model_providers;
 use crate::openai_model_info::get_model_info;
 use crate::project_doc::DEFAULT_PROJECT_DOC_FILENAME;
@@ -38,6 +40,7 @@ use codex_protocol::config_types::ForcedLoginMethod;
 use codex_protocol::config_types::ReasoningEffort;
 use codex_protocol::config_types::ReasoningSummary;
 use codex_protocol::config_types::SandboxMode;
+use codex_protocol::config_types::TrustLevel;
 use codex_protocol::config_types::Verbosity;
 use codex_rmcp_client::OAuthCredentialsStoreMode;
 use dirs::home_dir;
@@ -59,11 +62,11 @@ pub mod profile;
 pub mod types;
 
 #[cfg(target_os = "windows")]
-pub const OPENAI_DEFAULT_MODEL: &str = "gpt-5";
+pub const OPENAI_DEFAULT_MODEL: &str = "gpt-5.1";
 #[cfg(not(target_os = "windows"))]
-pub const OPENAI_DEFAULT_MODEL: &str = "gpt-5-codex";
-const OPENAI_DEFAULT_REVIEW_MODEL: &str = "gpt-5-codex";
-pub const GPT_5_CODEX_MEDIUM_MODEL: &str = "gpt-5-codex";
+pub const OPENAI_DEFAULT_MODEL: &str = "gpt-5.1-codex";
+const OPENAI_DEFAULT_REVIEW_MODEL: &str = "gpt-5.1-codex";
+pub const GPT_5_CODEX_MEDIUM_MODEL: &str = "gpt-5.1-codex";
 
 /// Maximum number of bytes of the documentation that will be embedded. Larger
 /// files are *silently truncated* to this size so we do not take up too much of
@@ -78,7 +81,7 @@ pub struct Config {
     /// Optional override of model selection.
     pub model: String,
 
-    /// Model used specifically for review sessions. Defaults to "gpt-5-codex".
+    /// Model used specifically for review sessions. Defaults to "gpt-5.1-codex".
     pub review_model: String,
 
     pub model_family: ModelFamily,
@@ -192,6 +195,9 @@ pub struct Config {
     /// Additional filenames to try when looking for project-level docs.
     pub project_doc_fallback_filenames: Vec<String>,
 
+    /// Token budget applied when storing tool/function outputs in the context manager.
+    pub tool_output_token_limit: Option<usize>,
+
     /// Directory containing all Codex state (defaults to `~/.codex` but can be
     /// overridden by the `CODEX_HOME` environment variable).
     pub codex_home: PathBuf,
@@ -382,15 +388,16 @@ fn ensure_no_inline_bearer_tokens(value: &TomlValue) -> std::io::Result<()> {
     Ok(())
 }
 
-pub(crate) fn set_project_trusted_inner(
+pub(crate) fn set_project_trust_level_inner(
     doc: &mut DocumentMut,
     project_path: &Path,
+    trust_level: TrustLevel,
 ) -> anyhow::Result<()> {
     // Ensure we render a human-friendly structure:
     //
     // [projects]
     // [projects."/path/to/project"]
-    // trust_level = "trusted"
+    // trust_level = "trusted" or "untrusted"
     //
     // rather than inline tables like:
     //
@@ -446,20 +453,66 @@ pub(crate) fn set_project_trusted_inner(
         return Err(anyhow::anyhow!("project table missing for {project_key}"));
     };
     proj_tbl.set_implicit(false);
-    proj_tbl["trust_level"] = toml_edit::value("trusted");
+    proj_tbl["trust_level"] = toml_edit::value(trust_level.to_string());
     Ok(())
 }
 
-/// Patch `CODEX_HOME/config.toml` project state.
+/// Patch `CODEX_HOME/config.toml` project state to set trust level.
 /// Use with caution.
-pub fn set_project_trusted(codex_home: &Path, project_path: &Path) -> anyhow::Result<()> {
+pub fn set_project_trust_level(
+    codex_home: &Path,
+    project_path: &Path,
+    trust_level: TrustLevel,
+) -> anyhow::Result<()> {
     use crate::config::edit::ConfigEditsBuilder;
 
     ConfigEditsBuilder::new(codex_home)
-        .set_project_trusted(project_path)
+        .set_project_trust_level(project_path, trust_level)
         .apply_blocking()
 }
 
+/// Save the default OSS provider preference to config.toml
+pub fn set_default_oss_provider(codex_home: &Path, provider: &str) -> std::io::Result<()> {
+    // Validate that the provider is one of the known OSS providers
+    match provider {
+        LMSTUDIO_OSS_PROVIDER_ID | OLLAMA_OSS_PROVIDER_ID => {
+            // Valid provider, continue
+        }
+        _ => {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidInput,
+                format!(
+                    "Invalid OSS provider '{provider}'. Must be one of: {LMSTUDIO_OSS_PROVIDER_ID}, {OLLAMA_OSS_PROVIDER_ID}"
+                ),
+            ));
+        }
+    }
+    let config_path = codex_home.join(CONFIG_TOML_FILE);
+
+    // Read existing config or create empty string if file doesn't exist
+    let content = match std::fs::read_to_string(&config_path) {
+        Ok(content) => content,
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => String::new(),
+        Err(e) => return Err(e),
+    };
+
+    // Parse as DocumentMut for editing while preserving structure
+    let mut doc = content.parse::<DocumentMut>().map_err(|e| {
+        std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            format!("failed to parse config.toml: {e}"),
+        )
+    })?;
+
+    // Set the default_oss_provider at root level
+    use toml_edit::value;
+    doc["oss_provider"] = value(provider);
+
+    // Write the modified document back
+    std::fs::write(&config_path, doc.to_string())?;
+    Ok(())
+}
+
 /// Apply a single dotted-path override onto a TOML value.
 fn apply_toml_override(root: &mut TomlValue, path: &str, value: TomlValue) {
     use toml::value::Table;
@@ -586,6 +639,9 @@ pub struct ConfigToml {
     /// Ordered list of fallback filenames to look for when AGENTS.md is missing.
     pub project_doc_fallback_filenames: Option<Vec<String>>,
 
+    /// Token budget applied when storing tool/function outputs in the context manager.
+    pub tool_output_token_limit: Option<usize>,
+
     /// Profile to use from the `profiles` map.
     pub profile: Option<String>,
 
@@ -657,6 +713,8 @@ pub struct ConfigToml {
     pub experimental_use_rmcp_client: Option<bool>,
     pub experimental_use_freeform_apply_patch: Option<bool>,
     pub experimental_sandbox_command_assessment: Option<bool>,
+    /// Preferred OSS provider for local models, e.g. "lmstudio" or "ollama".
+    pub oss_provider: Option<String>,
 }
 
 impl From<ConfigToml> for UserSavedConfig {
@@ -686,15 +744,16 @@ impl From<ConfigToml> for UserSavedConfig {
 
 #[derive(Deserialize, Debug, Clone, PartialEq, Eq)]
 pub struct ProjectConfig {
-    pub trust_level: Option<String>,
+    pub trust_level: Option<TrustLevel>,
 }
 
 impl ProjectConfig {
     pub fn is_trusted(&self) -> bool {
-        match &self.trust_level {
-            Some(trust_level) => trust_level == "trusted",
-            None => false,
-        }
+        matches!(self.trust_level, Some(TrustLevel::Trusted))
+    }
+
+    pub fn is_untrusted(&self) -> bool {
+        matches!(self.trust_level, Some(TrustLevel::Untrusted))
     }
 }
 
@@ -735,9 +794,9 @@ impl ConfigToml {
             .or(profile_sandbox_mode)
             .or(self.sandbox_mode)
             .or_else(|| {
-                // if no sandbox_mode is set, but user has marked directory as trusted, use WorkspaceWrite
+                // if no sandbox_mode is set, but user has marked directory as trusted or untrusted, use WorkspaceWrite
                 self.get_active_project(resolved_cwd).and_then(|p| {
-                    if p.is_trusted() {
+                    if p.is_trusted() || p.is_untrusted() {
                         Some(SandboxMode::WorkspaceWrite)
                     } else {
                         None
@@ -844,6 +903,34 @@ pub struct ConfigOverrides {
     pub additional_writable_roots: Vec<PathBuf>,
 }
 
+/// Resolves the OSS provider from CLI override, profile config, or global config.
+/// Returns `None` if no provider is configured at any level.
+pub fn resolve_oss_provider(
+    explicit_provider: Option<&str>,
+    config_toml: &ConfigToml,
+    config_profile: Option<String>,
+) -> Option<String> {
+    if let Some(provider) = explicit_provider {
+        // Explicit provider specified (e.g., via --local-provider)
+        Some(provider.to_string())
+    } else {
+        // Check profile config first, then global config
+        let profile = config_toml.get_config_profile(config_profile).ok();
+        if let Some(profile) = &profile {
+            // Check if profile has an oss provider
+            if let Some(profile_oss_provider) = &profile.oss_provider {
+                Some(profile_oss_provider.clone())
+            }
+            // If not then check if the toml has an oss provider
+            else {
+                config_toml.oss_provider.clone()
+            }
+        } else {
+            config_toml.oss_provider.clone()
+        }
+    }
+}
+
 impl Config {
     /// Meant to be used exclusively for tests: `load_with_overrides()` should
     /// be used in all other cases.
@@ -958,6 +1045,9 @@ impl Config {
                 if active_project.is_trusted() {
                     // If no explicit approval policy is set, but we trust cwd, default to OnRequest
                     AskForApproval::OnRequest
+                } else if active_project.is_untrusted() {
+                    // If project is explicitly marked untrusted, require approval for non-safe commands
+                    AskForApproval::UnlessTrusted
                 } else {
                     AskForApproval::default()
                 }
@@ -1125,6 +1215,7 @@ impl Config {
                     }
                 })
                 .collect(),
+            tool_output_token_limit: cfg.tool_output_token_limit,
             codex_home,
             history,
             file_opener: cfg.file_opener.unwrap_or(UriBasedFileOpener::VsCode),
@@ -1229,6 +1320,16 @@ impl Config {
             Ok(Some(s))
         }
     }
+
+    pub fn set_windows_sandbox_globally(&mut self, value: bool) {
+        crate::safety::set_windows_sandbox_enabled(value);
+        if value {
+            self.features.enable(Feature::WindowsSandbox);
+        } else {
+            self.features.disable(Feature::WindowsSandbox);
+        }
+        self.forced_auto_mode_downgraded_on_windows = !value;
+    }
 }
 
 fn default_model() -> String {
@@ -1323,7 +1424,7 @@ persistence = "none"
     }
 
     #[test]
-    fn tui_config_missing_notifications_field_defaults_to_disabled() {
+    fn tui_config_missing_notifications_field_defaults_to_enabled() {
         let cfg = r#"
 [tui]
 "#;
@@ -1332,7 +1433,7 @@ persistence = "none"
             .expect("TUI config without notifications should succeed");
         let tui = parsed.tui.expect("config should include tui section");
 
-        assert_eq!(tui.notifications, Notifications::Enabled(false));
+        assert_eq!(tui.notifications, Notifications::Enabled(true));
     }
 
     #[test]
@@ -2532,7 +2633,7 @@ url = "https://example.com/mcp"
         let codex_home = TempDir::new()?;
 
         ConfigEditsBuilder::new(codex_home.path())
-            .set_model(Some("gpt-5-codex"), Some(ReasoningEffort::High))
+            .set_model(Some("gpt-5.1-codex"), Some(ReasoningEffort::High))
             .apply()
             .await?;
 
@@ -2540,7 +2641,7 @@ url = "https://example.com/mcp"
             tokio::fs::read_to_string(codex_home.path().join(CONFIG_TOML_FILE)).await?;
         let parsed: ConfigToml = toml::from_str(&serialized)?;
 
-        assert_eq!(parsed.model.as_deref(), Some("gpt-5-codex"));
+        assert_eq!(parsed.model.as_deref(), Some("gpt-5.1-codex"));
         assert_eq!(parsed.model_reasoning_effort, Some(ReasoningEffort::High));
 
         Ok(())
@@ -2554,7 +2655,7 @@ url = "https://example.com/mcp"
         tokio::fs::write(
             &config_path,
             r#"
-model = "gpt-5-codex"
+model = "gpt-5.1-codex"
 model_reasoning_effort = "medium"
 
 [profiles.dev]
@@ -2590,7 +2691,7 @@ model = "gpt-4.1"
 
         ConfigEditsBuilder::new(codex_home.path())
             .with_profile(Some("dev"))
-            .set_model(Some("gpt-5-codex"), Some(ReasoningEffort::Medium))
+            .set_model(Some("gpt-5.1-codex"), Some(ReasoningEffort::Medium))
             .apply()
             .await?;
 
@@ -2602,7 +2703,7 @@ model = "gpt-4.1"
             .get("dev")
             .expect("profile should be created");
 
-        assert_eq!(profile.model.as_deref(), Some("gpt-5-codex"));
+        assert_eq!(profile.model.as_deref(), Some("gpt-5.1-codex"));
         assert_eq!(
             profile.model_reasoning_effort,
             Some(ReasoningEffort::Medium)
@@ -2624,7 +2725,7 @@ model = "gpt-4"
 model_reasoning_effort = "medium"
 
 [profiles.prod]
-model = "gpt-5-codex"
+model = "gpt-5.1-codex"
 "#,
         )
         .await?;
@@ -2653,7 +2754,7 @@ model = "gpt-5-codex"
                 .profiles
                 .get("prod")
                 .and_then(|profile| profile.model.as_deref()),
-            Some("gpt-5-codex"),
+            Some("gpt-5.1-codex"),
         );
 
         Ok(())
@@ -2768,7 +2869,7 @@ model_provider = "openai"
 approval_policy = "on-failure"
 
 [profiles.gpt5]
-model = "gpt-5"
+model = "gpt-5.1"
 model_provider = "openai"
 approval_policy = "on-failure"
 model_reasoning_effort = "high"
@@ -2877,6 +2978,7 @@ model_verbosity = "high"
                 model_providers: fixture.model_provider_map.clone(),
                 project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
                 project_doc_fallback_filenames: Vec::new(),
+                tool_output_token_limit: None,
                 codex_home: fixture.codex_home(),
                 history: History::default(),
                 file_opener: UriBasedFileOpener::VsCode,
@@ -2948,6 +3050,7 @@ model_verbosity = "high"
             model_providers: fixture.model_provider_map.clone(),
             project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
             project_doc_fallback_filenames: Vec::new(),
+            tool_output_token_limit: None,
             codex_home: fixture.codex_home(),
             history: History::default(),
             file_opener: UriBasedFileOpener::VsCode,
@@ -3034,6 +3137,7 @@ model_verbosity = "high"
             model_providers: fixture.model_provider_map.clone(),
             project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
             project_doc_fallback_filenames: Vec::new(),
+            tool_output_token_limit: None,
             codex_home: fixture.codex_home(),
             history: History::default(),
             file_opener: UriBasedFileOpener::VsCode,
@@ -3084,9 +3188,9 @@ model_verbosity = "high"
             fixture.codex_home(),
         )?;
         let expected_gpt5_profile_config = Config {
-            model: "gpt-5".to_string(),
+            model: "gpt-5.1".to_string(),
             review_model: OPENAI_DEFAULT_REVIEW_MODEL.to_string(),
-            model_family: find_family_for_model("gpt-5").expect("known model slug"),
+            model_family: find_family_for_model("gpt-5.1").expect("known model slug"),
             model_context_window: Some(272_000),
             model_max_output_tokens: Some(128_000),
             model_auto_compact_token_limit: Some(244_800),
@@ -3106,6 +3210,7 @@ model_verbosity = "high"
             model_providers: fixture.model_provider_map.clone(),
             project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
             project_doc_fallback_filenames: Vec::new(),
+            tool_output_token_limit: None,
             codex_home: fixture.codex_home(),
             history: History::default(),
             file_opener: UriBasedFileOpener::VsCode,
@@ -3164,7 +3269,7 @@ model_verbosity = "high"
         let project_dir = Path::new("/some/path");
         let mut doc = DocumentMut::new();
 
-        set_project_trusted_inner(&mut doc, project_dir)?;
+        set_project_trust_level_inner(&mut doc, project_dir, TrustLevel::Trusted)?;
 
         let contents = doc.to_string();
 
@@ -3204,7 +3309,7 @@ trust_level = "trusted"
         let mut doc = initial.parse::<DocumentMut>()?;
 
         // Run the function; it should convert to explicit tables and set trusted
-        set_project_trusted_inner(&mut doc, project_dir)?;
+        set_project_trust_level_inner(&mut doc, project_dir, TrustLevel::Trusted)?;
 
         let contents = doc.to_string();
 
@@ -3231,7 +3336,7 @@ model = "foo""#;
 
         // Approve a new directory
         let new_project = Path::new("/Users/mbolin/code/codex2");
-        set_project_trusted_inner(&mut doc, new_project)?;
+        set_project_trust_level_inner(&mut doc, new_project, TrustLevel::Trusted)?;
 
         let contents = doc.to_string();
 
@@ -3254,6 +3359,201 @@ trust_level = "trusted"
 
         Ok(())
     }
+
+    #[test]
+    fn test_set_default_oss_provider() -> std::io::Result<()> {
+        let temp_dir = TempDir::new()?;
+        let codex_home = temp_dir.path();
+        let config_path = codex_home.join(CONFIG_TOML_FILE);
+
+        // Test setting valid provider on empty config
+        set_default_oss_provider(codex_home, OLLAMA_OSS_PROVIDER_ID)?;
+        let content = std::fs::read_to_string(&config_path)?;
+        assert!(content.contains("oss_provider = \"ollama\""));
+
+        // Test updating existing config
+        std::fs::write(&config_path, "model = \"gpt-4\"\n")?;
+        set_default_oss_provider(codex_home, LMSTUDIO_OSS_PROVIDER_ID)?;
+        let content = std::fs::read_to_string(&config_path)?;
+        assert!(content.contains("oss_provider = \"lmstudio\""));
+        assert!(content.contains("model = \"gpt-4\""));
+
+        // Test overwriting existing oss_provider
+        set_default_oss_provider(codex_home, OLLAMA_OSS_PROVIDER_ID)?;
+        let content = std::fs::read_to_string(&config_path)?;
+        assert!(content.contains("oss_provider = \"ollama\""));
+        assert!(!content.contains("oss_provider = \"lmstudio\""));
+
+        // Test invalid provider
+        let result = set_default_oss_provider(codex_home, "invalid_provider");
+        assert!(result.is_err());
+        let error = result.unwrap_err();
+        assert_eq!(error.kind(), std::io::ErrorKind::InvalidInput);
+        assert!(error.to_string().contains("Invalid OSS provider"));
+        assert!(error.to_string().contains("invalid_provider"));
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_untrusted_project_gets_workspace_write_sandbox() -> anyhow::Result<()> {
+        let config_with_untrusted = r#"
+[projects."/tmp/test"]
+trust_level = "untrusted"
+"#;
+
+        let cfg = toml::from_str::<ConfigToml>(config_with_untrusted)
+            .expect("TOML deserialization should succeed");
+
+        let resolution = cfg.derive_sandbox_policy(None, None, &PathBuf::from("/tmp/test"));
+
+        // Verify that untrusted projects get WorkspaceWrite (or ReadOnly on Windows due to downgrade)
+        if cfg!(target_os = "windows") {
+            assert!(
+                matches!(resolution.policy, SandboxPolicy::ReadOnly),
+                "Expected ReadOnly on Windows, got {:?}",
+                resolution.policy
+            );
+        } else {
+            assert!(
+                matches!(resolution.policy, SandboxPolicy::WorkspaceWrite { .. }),
+                "Expected WorkspaceWrite for untrusted project, got {:?}",
+                resolution.policy
+            );
+        }
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_resolve_oss_provider_explicit_override() {
+        let config_toml = ConfigToml::default();
+        let result = resolve_oss_provider(Some("custom-provider"), &config_toml, None);
+        assert_eq!(result, Some("custom-provider".to_string()));
+    }
+
+    #[test]
+    fn test_resolve_oss_provider_from_profile() {
+        let mut profiles = std::collections::HashMap::new();
+        let profile = ConfigProfile {
+            oss_provider: Some("profile-provider".to_string()),
+            ..Default::default()
+        };
+        profiles.insert("test-profile".to_string(), profile);
+        let config_toml = ConfigToml {
+            profiles,
+            ..Default::default()
+        };
+
+        let result = resolve_oss_provider(None, &config_toml, Some("test-profile".to_string()));
+        assert_eq!(result, Some("profile-provider".to_string()));
+    }
+
+    #[test]
+    fn test_resolve_oss_provider_from_global_config() {
+        let config_toml = ConfigToml {
+            oss_provider: Some("global-provider".to_string()),
+            ..Default::default()
+        };
+
+        let result = resolve_oss_provider(None, &config_toml, None);
+        assert_eq!(result, Some("global-provider".to_string()));
+    }
+
+    #[test]
+    fn test_resolve_oss_provider_profile_fallback_to_global() {
+        let mut profiles = std::collections::HashMap::new();
+        let profile = ConfigProfile::default(); // No oss_provider set
+        profiles.insert("test-profile".to_string(), profile);
+        let config_toml = ConfigToml {
+            oss_provider: Some("global-provider".to_string()),
+            profiles,
+            ..Default::default()
+        };
+
+        let result = resolve_oss_provider(None, &config_toml, Some("test-profile".to_string()));
+        assert_eq!(result, Some("global-provider".to_string()));
+    }
+
+    #[test]
+    fn test_resolve_oss_provider_none_when_not_configured() {
+        let config_toml = ConfigToml::default();
+        let result = resolve_oss_provider(None, &config_toml, None);
+        assert_eq!(result, None);
+    }
+
+    #[test]
+    fn test_resolve_oss_provider_explicit_overrides_all() {
+        let mut profiles = std::collections::HashMap::new();
+        let profile = ConfigProfile {
+            oss_provider: Some("profile-provider".to_string()),
+            ..Default::default()
+        };
+        profiles.insert("test-profile".to_string(), profile);
+        let config_toml = ConfigToml {
+            oss_provider: Some("global-provider".to_string()),
+            profiles,
+            ..Default::default()
+        };
+
+        let result = resolve_oss_provider(
+            Some("explicit-provider"),
+            &config_toml,
+            Some("test-profile".to_string()),
+        );
+        assert_eq!(result, Some("explicit-provider".to_string()));
+    }
+
+    #[test]
+    fn test_untrusted_project_gets_unless_trusted_approval_policy() -> std::io::Result<()> {
+        let codex_home = TempDir::new()?;
+        let test_project_dir = TempDir::new()?;
+        let test_path = test_project_dir.path();
+
+        let mut projects = std::collections::HashMap::new();
+        projects.insert(
+            test_path.to_string_lossy().to_string(),
+            ProjectConfig {
+                trust_level: Some(TrustLevel::Untrusted),
+            },
+        );
+
+        let cfg = ConfigToml {
+            projects: Some(projects),
+            ..Default::default()
+        };
+
+        let config = Config::load_from_base_config_with_overrides(
+            cfg,
+            ConfigOverrides {
+                cwd: Some(test_path.to_path_buf()),
+                ..Default::default()
+            },
+            codex_home.path().to_path_buf(),
+        )?;
+
+        // Verify that untrusted projects get UnlessTrusted approval policy
+        assert_eq!(
+            config.approval_policy,
+            AskForApproval::UnlessTrusted,
+            "Expected UnlessTrusted approval policy for untrusted project"
+        );
+
+        // Verify that untrusted projects still get WorkspaceWrite sandbox (or ReadOnly on Windows)
+        if cfg!(target_os = "windows") {
+            assert!(
+                matches!(config.sandbox_policy, SandboxPolicy::ReadOnly),
+                "Expected ReadOnly on Windows"
+            );
+        } else {
+            assert!(
+                matches!(config.sandbox_policy, SandboxPolicy::WorkspaceWrite { .. }),
+                "Expected WorkspaceWrite sandbox for untrusted project"
+            );
+        }
+
+        Ok(())
+    }
 }
 
 #[cfg(test)]

codex-rs/core/src/config/profile.rs

@@ -33,6 +33,7 @@ pub struct ConfigProfile {
     /// Optional feature toggles scoped to this profile.
     #[serde(default)]
     pub features: Option<crate::features::FeaturesToml>,
+    pub oss_provider: Option<String>,
 }
 
 impl From<ConfigProfile> for codex_app_server_protocol::Profile {

codex-rs/core/src/config/types.rs

@@ -282,6 +282,14 @@ pub enum OtelHttpProtocol {
     Json,
 }
 
+#[derive(Deserialize, Debug, Clone, PartialEq, Default)]
+#[serde(rename_all = "kebab-case")]
+pub struct OtelTlsConfig {
+    pub ca_certificate: Option<PathBuf>,
+    pub client_certificate: Option<PathBuf>,
+    pub client_private_key: Option<PathBuf>,
+}
+
 /// Which OTEL exporter to use.
 #[derive(Deserialize, Debug, Clone, PartialEq)]
 #[serde(rename_all = "kebab-case")]
@@ -289,12 +297,18 @@ pub enum OtelExporterKind {
     None,
     OtlpHttp {
         endpoint: String,
+        #[serde(default)]
         headers: HashMap<String, String>,
         protocol: OtelHttpProtocol,
+        #[serde(default)]
+        tls: Option<OtelTlsConfig>,
     },
     OtlpGrpc {
         endpoint: String,
+        #[serde(default)]
         headers: HashMap<String, String>,
+        #[serde(default)]
+        tls: Option<OtelTlsConfig>,
     },
 }
 
@@ -338,7 +352,7 @@ pub enum Notifications {
 
 impl Default for Notifications {
     fn default() -> Self {
-        Self::Enabled(false)
+        Self::Enabled(true)
     }
 }
 
@@ -346,7 +360,7 @@ impl Default for Notifications {
 #[derive(Deserialize, Debug, Clone, PartialEq, Default)]
 pub struct Tui {
     /// Enable desktop notifications from the TUI when the terminal is unfocused.
-    /// Defaults to `false`.
+    /// Defaults to `true`.
     #[serde(default)]
     pub notifications: Notifications,
 }

codex-rs/core/src/context_manager/history.rs

@@ -1,8 +1,8 @@
 use crate::codex::TurnContext;
 use crate::context_manager::normalize;
-use crate::context_manager::truncate;
-use crate::context_manager::truncate::format_output_for_model_body;
-use crate::context_manager::truncate::globally_truncate_function_output_items;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::truncate_function_output_items_with_policy;
+use crate::truncate::truncate_text;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseItem;
 use codex_protocol::protocol::TokenUsage;
@@ -10,12 +10,6 @@ use codex_protocol::protocol::TokenUsageInfo;
 use codex_utils_tokenizer::Tokenizer;
 use std::ops::Deref;
 
-const CONTEXT_WINDOW_HARD_LIMIT_FACTOR: f64 = 1.1;
-const CONTEXT_WINDOW_HARD_LIMIT_BYTES: usize =
-    (truncate::MODEL_FORMAT_MAX_BYTES as f64 * CONTEXT_WINDOW_HARD_LIMIT_FACTOR) as usize;
-const CONTEXT_WINDOW_HARD_LIMIT_LINES: usize =
-    (truncate::MODEL_FORMAT_MAX_LINES as f64 * CONTEXT_WINDOW_HARD_LIMIT_FACTOR) as usize;
-
 /// Transcript of conversation history
 #[derive(Debug, Clone, Default)]
 pub(crate) struct ContextManager {
@@ -50,7 +44,7 @@ impl ContextManager {
     }
 
     /// `items` is ordered from oldest to newest.
-    pub(crate) fn record_items<I>(&mut self, items: I)
+    pub(crate) fn record_items<I>(&mut self, items: I, policy: TruncationPolicy)
     where
         I: IntoIterator,
         I::Item: std::ops::Deref<Target = ResponseItem>,
@@ -62,7 +56,7 @@ impl ContextManager {
                 continue;
             }
 
-            let processed = Self::process_item(&item);
+            let processed = self.process_item(item_ref, policy);
             self.items.push(processed);
         }
     }
@@ -150,18 +144,14 @@ impl ContextManager {
         items.retain(|item| !matches!(item, ResponseItem::GhostSnapshot { .. }));
     }
 
-    fn process_item(item: &ResponseItem) -> ResponseItem {
+    fn process_item(&self, item: &ResponseItem, policy: TruncationPolicy) -> ResponseItem {
         match item {
             ResponseItem::FunctionCallOutput { call_id, output } => {
-                let truncated = format_output_for_model_body(
-                    output.content.as_str(),
-                    CONTEXT_WINDOW_HARD_LIMIT_BYTES,
-                    CONTEXT_WINDOW_HARD_LIMIT_LINES,
-                );
+                let truncated = truncate_text(output.content.as_str(), policy);
                 let truncated_items = output
                     .content_items
                     .as_ref()
-                    .map(|items| globally_truncate_function_output_items(items));
+                    .map(|items| truncate_function_output_items_with_policy(items, policy));
                 ResponseItem::FunctionCallOutput {
                     call_id: call_id.clone(),
                     output: FunctionCallOutputPayload {
@@ -172,11 +162,7 @@ impl ContextManager {
                 }
             }
             ResponseItem::CustomToolCallOutput { call_id, output } => {
-                let truncated = format_output_for_model_body(
-                    output,
-                    CONTEXT_WINDOW_HARD_LIMIT_BYTES,
-                    CONTEXT_WINDOW_HARD_LIMIT_LINES,
-                );
+                let truncated = truncate_text(output, policy);
                 ResponseItem::CustomToolCallOutput {
                     call_id: call_id.clone(),
                     output: truncated,
@@ -188,6 +174,7 @@ impl ContextManager {
             | ResponseItem::FunctionCall { .. }
             | ResponseItem::WebSearchCall { .. }
             | ResponseItem::CustomToolCall { .. }
+            | ResponseItem::CompactionSummary { .. }
             | ResponseItem::GhostSnapshot { .. }
             | ResponseItem::Other => item.clone(),
         }
@@ -205,7 +192,8 @@ fn is_api_message(message: &ResponseItem) -> bool {
         | ResponseItem::CustomToolCallOutput { .. }
         | ResponseItem::LocalShellCall { .. }
         | ResponseItem::Reasoning { .. }
-        | ResponseItem::WebSearchCall { .. } => true,
+        | ResponseItem::WebSearchCall { .. }
+        | ResponseItem::CompactionSummary { .. } => true,
         ResponseItem::GhostSnapshot { .. } => false,
         ResponseItem::Other => false,
     }

codex-rs/core/src/context_manager/history_tests.rs

@@ -1,9 +1,8 @@
 use super::*;
-use crate::context_manager::MODEL_FORMAT_MAX_LINES;
-use crate::context_manager::truncate;
+use crate::truncate;
+use crate::truncate::TruncationPolicy;
 use codex_git::GhostCommit;
 use codex_protocol::models::ContentItem;
-use codex_protocol::models::FunctionCallOutputContentItem;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::LocalShellAction;
 use codex_protocol::models::LocalShellExecAction;
@@ -13,6 +12,9 @@ use codex_protocol::models::ReasoningItemReasoningSummary;
 use pretty_assertions::assert_eq;
 use regex_lite::Regex;
 
+const EXEC_FORMAT_MAX_LINES: usize = 256;
+const EXEC_FORMAT_MAX_BYTES: usize = 10_000;
+
 fn assistant_msg(text: &str) -> ResponseItem {
     ResponseItem::Message {
         id: None,
@@ -25,7 +27,9 @@ fn assistant_msg(text: &str) -> ResponseItem {
 
 fn create_history_with_items(items: Vec<ResponseItem>) -> ContextManager {
     let mut h = ContextManager::new();
-    h.record_items(items.iter());
+    // Use a generous but fixed token budget; tests only rely on truncation
+    // behavior, not on a specific model's token limit.
+    h.record_items(items.iter(), TruncationPolicy::Tokens(10_000));
     h
 }
 
@@ -55,6 +59,7 @@ fn reasoning_msg(text: &str) -> ResponseItem {
 #[test]
 fn filters_non_api_messages() {
     let mut h = ContextManager::default();
+    let policy = TruncationPolicy::Tokens(10_000);
     // System message is not API messages; Other is ignored.
     let system = ResponseItem::Message {
         id: None,
@@ -64,12 +69,12 @@ fn filters_non_api_messages() {
         }],
     };
     let reasoning = reasoning_msg("thinking...");
-    h.record_items([&system, &reasoning, &ResponseItem::Other]);
+    h.record_items([&system, &reasoning, &ResponseItem::Other], policy);
 
     // User and assistant should be retained.
     let u = user_msg("hi");
     let a = assistant_msg("hello");
-    h.record_items([&u, &a]);
+    h.record_items([&u, &a], policy);
 
     let items = h.contents();
     assert_eq!(
@@ -237,6 +242,9 @@ fn normalization_retains_local_shell_outputs() {
 #[test]
 fn record_items_truncates_function_call_output_content() {
     let mut history = ContextManager::new();
+    // Any reasonably small token budget works; the test only cares that
+    // truncation happens and the marker is present.
+    let policy = TruncationPolicy::Tokens(1_000);
     let long_line = "a very long line to trigger truncation\n";
     let long_output = long_line.repeat(2_500);
     let item = ResponseItem::FunctionCallOutput {
@@ -248,15 +256,20 @@ fn record_items_truncates_function_call_output_content() {
         },
     };
 
-    history.record_items([&item]);
+    history.record_items([&item], policy);
 
     assert_eq!(history.items.len(), 1);
     match &history.items[0] {
         ResponseItem::FunctionCallOutput { output, .. } => {
             assert_ne!(output.content, long_output);
             assert!(
-                output.content.starts_with("Total output lines:"),
-                "expected truncated summary, got {}",
+                output.content.contains("tokens truncated"),
+                "expected token-based truncation marker, got {}",
+                output.content
+            );
+            assert!(
+                output.content.contains("tokens truncated"),
+                "expected truncation marker, got {}",
                 output.content
             );
         }
@@ -267,6 +280,7 @@ fn record_items_truncates_function_call_output_content() {
 #[test]
 fn record_items_truncates_custom_tool_call_output_content() {
     let mut history = ContextManager::new();
+    let policy = TruncationPolicy::Tokens(1_000);
     let line = "custom output that is very long\n";
     let long_output = line.repeat(2_500);
     let item = ResponseItem::CustomToolCallOutput {
@@ -274,21 +288,48 @@ fn record_items_truncates_custom_tool_call_output_content() {
         output: long_output.clone(),
     };
 
-    history.record_items([&item]);
+    history.record_items([&item], policy);
 
     assert_eq!(history.items.len(), 1);
     match &history.items[0] {
         ResponseItem::CustomToolCallOutput { output, .. } => {
             assert_ne!(output, &long_output);
             assert!(
-                output.starts_with("Total output lines:"),
-                "expected truncated summary, got {output}"
+                output.contains("tokens truncated"),
+                "expected token-based truncation marker, got {output}"
+            );
+            assert!(
+                output.contains("tokens truncated") || output.contains("bytes truncated"),
+                "expected truncation marker, got {output}"
             );
         }
         other => panic!("unexpected history item: {other:?}"),
     }
 }
 
+#[test]
+fn record_items_respects_custom_token_limit() {
+    let mut history = ContextManager::new();
+    let policy = TruncationPolicy::Tokens(10);
+    let long_output = "tokenized content repeated many times ".repeat(200);
+    let item = ResponseItem::FunctionCallOutput {
+        call_id: "call-custom-limit".to_string(),
+        output: FunctionCallOutputPayload {
+            content: long_output,
+            success: Some(true),
+            ..Default::default()
+        },
+    };
+
+    history.record_items([&item], policy);
+
+    let stored = match &history.items[0] {
+        ResponseItem::FunctionCallOutput { output, .. } => output,
+        other => panic!("unexpected history item: {other:?}"),
+    };
+    assert!(stored.content.contains("tokens truncated"));
+}
+
 fn assert_truncated_message_matches(message: &str, line: &str, total_lines: usize) {
     let pattern = truncated_message_pattern(line, total_lines);
     let regex = Regex::new(&pattern).unwrap_or_else(|err| {
@@ -302,23 +343,22 @@ fn assert_truncated_message_matches(message: &str, line: &str, total_lines: usiz
         .expect("missing body capture")
         .as_str();
     assert!(
-        body.len() <= truncate::MODEL_FORMAT_MAX_BYTES,
+        body.len() <= EXEC_FORMAT_MAX_BYTES,
         "body exceeds byte limit: {} bytes",
         body.len()
     );
 }
 
 fn truncated_message_pattern(line: &str, total_lines: usize) -> String {
-    let head_lines = MODEL_FORMAT_MAX_LINES / 2;
-    let tail_lines = MODEL_FORMAT_MAX_LINES - head_lines;
+    let head_lines = EXEC_FORMAT_MAX_LINES / 2;
+    let tail_lines = EXEC_FORMAT_MAX_LINES - head_lines;
     let head_take = head_lines.min(total_lines);
     let tail_take = tail_lines.min(total_lines.saturating_sub(head_take));
     let omitted = total_lines.saturating_sub(head_take + tail_take);
     let escaped_line = regex_lite::escape(line);
     if omitted == 0 {
         return format!(
-            r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} output truncated to fit {max_bytes} bytes \.{{3}}]\n\n.*)$",
-            max_bytes = truncate::MODEL_FORMAT_MAX_BYTES,
+            r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} removed \d+ bytes to fit {EXEC_FORMAT_MAX_BYTES} byte limit \.{{3}}]\n\n.*)$",
         );
     }
     format!(
@@ -331,11 +371,7 @@ fn format_exec_output_truncates_large_error() {
     let line = "very long execution error line that should trigger truncation\n";
     let large_error = line.repeat(2_500); // way beyond both byte and line limits
 
-    let truncated = truncate::format_output_for_model_body(
-        &large_error,
-        truncate::MODEL_FORMAT_MAX_BYTES,
-        truncate::MODEL_FORMAT_MAX_LINES,
-    );
+    let truncated = truncate::truncate_with_line_bytes_budget(&large_error, EXEC_FORMAT_MAX_BYTES);
 
     let total_lines = large_error.lines().count();
     assert_truncated_message_matches(&truncated, line, total_lines);
@@ -344,17 +380,13 @@ fn format_exec_output_truncates_large_error() {
 
 #[test]
 fn format_exec_output_marks_byte_truncation_without_omitted_lines() {
-    let long_line = "a".repeat(truncate::MODEL_FORMAT_MAX_BYTES + 50);
-    let truncated = truncate::format_output_for_model_body(
-        &long_line,
-        truncate::MODEL_FORMAT_MAX_BYTES,
-        truncate::MODEL_FORMAT_MAX_LINES,
-    );
+    let long_line = "a".repeat(EXEC_FORMAT_MAX_BYTES + 50);
+    let truncated = truncate::truncate_with_line_bytes_budget(&long_line, EXEC_FORMAT_MAX_BYTES);
 
     assert_ne!(truncated, long_line);
+    let removed_bytes = long_line.len().saturating_sub(EXEC_FORMAT_MAX_BYTES);
     let marker_line = format!(
-        "[... output truncated to fit {} bytes ...]",
-        truncate::MODEL_FORMAT_MAX_BYTES
+        "[... removed {removed_bytes} bytes to fit {EXEC_FORMAT_MAX_BYTES} byte limit ...]"
     );
     assert!(
         truncated.contains(&marker_line),
@@ -371,28 +403,20 @@ fn format_exec_output_returns_original_when_within_limits() {
     let content = "example output\n".repeat(10);
 
     assert_eq!(
-        truncate::format_output_for_model_body(
-            &content,
-            truncate::MODEL_FORMAT_MAX_BYTES,
-            truncate::MODEL_FORMAT_MAX_LINES
-        ),
+        truncate::truncate_with_line_bytes_budget(&content, EXEC_FORMAT_MAX_BYTES),
         content
     );
 }
 
 #[test]
 fn format_exec_output_reports_omitted_lines_and_keeps_head_and_tail() {
-    let total_lines = truncate::MODEL_FORMAT_MAX_LINES + 100;
+    let total_lines = EXEC_FORMAT_MAX_LINES + 100;
     let content: String = (0..total_lines)
         .map(|idx| format!("line-{idx}\n"))
         .collect();
 
-    let truncated = truncate::format_output_for_model_body(
-        &content,
-        truncate::MODEL_FORMAT_MAX_BYTES,
-        truncate::MODEL_FORMAT_MAX_LINES,
-    );
-    let omitted = total_lines - truncate::MODEL_FORMAT_MAX_LINES;
+    let truncated = truncate::truncate_with_line_bytes_budget(&content, EXEC_FORMAT_MAX_BYTES);
+    let omitted = total_lines - EXEC_FORMAT_MAX_LINES;
     let expected_marker = format!("[... omitted {omitted} of {total_lines} lines ...]");
 
     assert!(
@@ -413,103 +437,24 @@ fn format_exec_output_reports_omitted_lines_and_keeps_head_and_tail() {
 
 #[test]
 fn format_exec_output_prefers_line_marker_when_both_limits_exceeded() {
-    let total_lines = truncate::MODEL_FORMAT_MAX_LINES + 42;
+    let total_lines = EXEC_FORMAT_MAX_LINES + 42;
     let long_line = "x".repeat(256);
     let content: String = (0..total_lines)
         .map(|idx| format!("line-{idx}-{long_line}\n"))
         .collect();
 
-    let truncated = truncate::format_output_for_model_body(
-        &content,
-        truncate::MODEL_FORMAT_MAX_BYTES,
-        truncate::MODEL_FORMAT_MAX_LINES,
-    );
+    let truncated = truncate::truncate_with_line_bytes_budget(&content, EXEC_FORMAT_MAX_BYTES);
 
     assert!(
         truncated.contains("[... omitted 42 of 298 lines ...]"),
         "expected omitted marker when line count exceeds limit: {truncated}"
     );
     assert!(
-        !truncated.contains("output truncated to fit"),
+        !truncated.contains("byte limit"),
         "line omission marker should take precedence over byte marker: {truncated}"
     );
 }
 
-#[test]
-fn truncates_across_multiple_under_limit_texts_and_reports_omitted() {
-    // Arrange: several text items, none exceeding per-item limit, but total exceeds budget.
-    let budget = truncate::MODEL_FORMAT_MAX_BYTES;
-    let t1_len = (budget / 2).saturating_sub(10);
-    let t2_len = (budget / 2).saturating_sub(10);
-    let remaining_after_t1_t2 = budget.saturating_sub(t1_len + t2_len);
-    let t3_len = 50; // gets truncated to remaining_after_t1_t2
-    let t4_len = 5; // omitted
-    let t5_len = 7; // omitted
-
-    let t1 = "a".repeat(t1_len);
-    let t2 = "b".repeat(t2_len);
-    let t3 = "c".repeat(t3_len);
-    let t4 = "d".repeat(t4_len);
-    let t5 = "e".repeat(t5_len);
-
-    let item = ResponseItem::FunctionCallOutput {
-        call_id: "call-omit".to_string(),
-        output: FunctionCallOutputPayload {
-            content: "irrelevant".to_string(),
-            content_items: Some(vec![
-                FunctionCallOutputContentItem::InputText { text: t1 },
-                FunctionCallOutputContentItem::InputText { text: t2 },
-                FunctionCallOutputContentItem::InputImage {
-                    image_url: "img:mid".to_string(),
-                },
-                FunctionCallOutputContentItem::InputText { text: t3 },
-                FunctionCallOutputContentItem::InputText { text: t4 },
-                FunctionCallOutputContentItem::InputText { text: t5 },
-            ]),
-            success: Some(true),
-        },
-    };
-
-    let mut history = ContextManager::new();
-    history.record_items([&item]);
-    assert_eq!(history.items.len(), 1);
-    let json = serde_json::to_value(&history.items[0]).expect("serialize to json");
-
-    let output = json
-        .get("output")
-        .expect("output field")
-        .as_array()
-        .expect("array output");
-
-    // Expect: t1 (full), t2 (full), image, t3 (truncated), summary mentioning 2 omitted.
-    assert_eq!(output.len(), 5);
-
-    let first = output[0].as_object().expect("first obj");
-    assert_eq!(first.get("type").unwrap(), "input_text");
-    let first_text = first.get("text").unwrap().as_str().unwrap();
-    assert_eq!(first_text.len(), t1_len);
-
-    let second = output[1].as_object().expect("second obj");
-    assert_eq!(second.get("type").unwrap(), "input_text");
-    let second_text = second.get("text").unwrap().as_str().unwrap();
-    assert_eq!(second_text.len(), t2_len);
-
-    assert_eq!(
-        output[2],
-        serde_json::json!({"type": "input_image", "image_url": "img:mid"})
-    );
-
-    let fourth = output[3].as_object().expect("fourth obj");
-    assert_eq!(fourth.get("type").unwrap(), "input_text");
-    let fourth_text = fourth.get("text").unwrap().as_str().unwrap();
-    assert_eq!(fourth_text.len(), remaining_after_t1_t2);
-
-    let summary = output[4].as_object().expect("summary obj");
-    assert_eq!(summary.get("type").unwrap(), "input_text");
-    let summary_text = summary.get("text").unwrap().as_str().unwrap();
-    assert!(summary_text.contains("omitted 2 text items"));
-}
-
 //TODO(aibrahim): run CI in release mode.
 #[cfg(not(debug_assertions))]
 #[test]

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

@@ -1,8 +1,5 @@
 mod history;
 mod normalize;
-mod truncate;
 
+pub(crate) use crate::truncate::truncate_with_line_bytes_budget;
 pub(crate) use history::ContextManager;
-pub(crate) use truncate::MODEL_FORMAT_MAX_BYTES;
-pub(crate) use truncate::MODEL_FORMAT_MAX_LINES;
-pub(crate) use truncate::format_output_for_model_body;

codex-rs/core/src/context_manager/truncate.rs

@@ -1,148 +0,0 @@
-use codex_protocol::models::FunctionCallOutputContentItem;
-use codex_utils_string::take_bytes_at_char_boundary;
-use codex_utils_string::take_last_bytes_at_char_boundary;
-
-use crate::util::error_or_panic;
-
-// Model-formatting limits: clients get full streams; only content sent to the model is truncated.
-pub const MODEL_FORMAT_MAX_BYTES: usize = 10 * 1024; // 10 KiB
-pub const MODEL_FORMAT_MAX_LINES: usize = 256; // lines
-
-pub(crate) fn globally_truncate_function_output_items(
-    items: &[FunctionCallOutputContentItem],
-) -> Vec<FunctionCallOutputContentItem> {
-    let mut out: Vec<FunctionCallOutputContentItem> = Vec::with_capacity(items.len());
-    let mut remaining = MODEL_FORMAT_MAX_BYTES;
-    let mut omitted_text_items = 0usize;
-
-    for it in items {
-        match it {
-            FunctionCallOutputContentItem::InputText { text } => {
-                if remaining == 0 {
-                    omitted_text_items += 1;
-                    continue;
-                }
-
-                let len = text.len();
-                if len <= remaining {
-                    out.push(FunctionCallOutputContentItem::InputText { text: text.clone() });
-                    remaining -= len;
-                } else {
-                    let slice = take_bytes_at_char_boundary(text, remaining);
-                    if !slice.is_empty() {
-                        out.push(FunctionCallOutputContentItem::InputText {
-                            text: slice.to_string(),
-                        });
-                    }
-                    remaining = 0;
-                }
-            }
-            // todo(aibrahim): handle input images; resize
-            FunctionCallOutputContentItem::InputImage { image_url } => {
-                out.push(FunctionCallOutputContentItem::InputImage {
-                    image_url: image_url.clone(),
-                });
-            }
-        }
-    }
-
-    if omitted_text_items > 0 {
-        out.push(FunctionCallOutputContentItem::InputText {
-            text: format!("[omitted {omitted_text_items} text items ...]"),
-        });
-    }
-
-    out
-}
-
-pub(crate) fn format_output_for_model_body(
-    content: &str,
-    limit_bytes: usize,
-    limit_lines: usize,
-) -> String {
-    // Head+tail truncation for the model: show the beginning and end with an elision.
-    // Clients still receive full streams; only this formatted summary is capped.
-    let total_lines = content.lines().count();
-    if content.len() <= limit_bytes && total_lines <= limit_lines {
-        return content.to_string();
-    }
-    let output = truncate_formatted_exec_output(content, total_lines, limit_bytes, limit_lines);
-    format!("Total output lines: {total_lines}\n\n{output}")
-}
-
-fn truncate_formatted_exec_output(
-    content: &str,
-    total_lines: usize,
-    limit_bytes: usize,
-    limit_lines: usize,
-) -> String {
-    debug_panic_on_double_truncation(content);
-    let head_lines: usize = limit_lines / 2;
-    let tail_lines: usize = limit_lines - head_lines; // 128
-    let head_bytes: usize = limit_bytes / 2;
-    let segments: Vec<&str> = content.split_inclusive('\n').collect();
-    let head_take = head_lines.min(segments.len());
-    let tail_take = tail_lines.min(segments.len().saturating_sub(head_take));
-    let omitted = segments.len().saturating_sub(head_take + tail_take);
-
-    let head_slice_end: usize = segments
-        .iter()
-        .take(head_take)
-        .map(|segment| segment.len())
-        .sum();
-    let tail_slice_start: usize = if tail_take == 0 {
-        content.len()
-    } else {
-        content.len()
-            - segments
-                .iter()
-                .rev()
-                .take(tail_take)
-                .map(|segment| segment.len())
-                .sum::<usize>()
-    };
-    let head_slice = &content[..head_slice_end];
-    let tail_slice = &content[tail_slice_start..];
-    let truncated_by_bytes = content.len() > limit_bytes;
-    // this is a bit wrong. We are counting metadata lines and not just shell output lines.
-    let marker = if omitted > 0 {
-        Some(format!(
-            "\n[... omitted {omitted} of {total_lines} lines ...]\n\n"
-        ))
-    } else if truncated_by_bytes {
-        Some(format!(
-            "\n[... output truncated to fit {limit_bytes} bytes ...]\n\n"
-        ))
-    } else {
-        None
-    };
-
-    let marker_len = marker.as_ref().map_or(0, String::len);
-    let base_head_budget = head_bytes.min(limit_bytes);
-    let head_budget = base_head_budget.min(limit_bytes.saturating_sub(marker_len));
-    let head_part = take_bytes_at_char_boundary(head_slice, head_budget);
-    let mut result = String::with_capacity(limit_bytes.min(content.len()));
-
-    result.push_str(head_part);
-    if let Some(marker_text) = marker.as_ref() {
-        result.push_str(marker_text);
-    }
-
-    let remaining = limit_bytes.saturating_sub(result.len());
-    if remaining == 0 {
-        return result;
-    }
-
-    let tail_part = take_last_bytes_at_char_boundary(tail_slice, remaining);
-    result.push_str(tail_part);
-
-    result
-}
-
-fn debug_panic_on_double_truncation(content: &str) {
-    if content.contains("Total output lines:") && content.contains("omitted") {
-        error_or_panic(format!(
-            "FunctionCallOutput content was already truncated before ContextManager::record_items; this would cause double truncation {content}"
-        ));
-    }
-}

codex-rs/core/src/error.rs

@@ -2,7 +2,8 @@ use crate::codex::ProcessedResponseItem;
 use crate::exec::ExecToolCallOutput;
 use crate::token_data::KnownPlan;
 use crate::token_data::PlanType;
-use crate::truncate::truncate_middle;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::truncate_text;
 use chrono::DateTime;
 use chrono::Datelike;
 use chrono::Local;
@@ -461,7 +462,10 @@ pub fn get_error_message_ui(e: &CodexErr) -> String {
         _ => e.to_string(),
     };
 
-    truncate_middle(&message, ERROR_MESSAGE_UI_MAX_BYTES).0
+    truncate_text(
+        &message,
+        TruncationPolicy::Bytes(ERROR_MESSAGE_UI_MAX_BYTES),
+    )
 }
 
 #[cfg(test)]

codex-rs/core/src/exec.rs

@@ -182,23 +182,26 @@ async fn exec_windows_sandbox(
         ..
     } = params;
 
-    let policy_str = match sandbox_policy {
-        SandboxPolicy::DangerFullAccess => "workspace-write",
-        SandboxPolicy::ReadOnly => "read-only",
-        SandboxPolicy::WorkspaceWrite { .. } => "workspace-write",
-    };
-
+    let policy_str = serde_json::to_string(sandbox_policy).map_err(|err| {
+        CodexErr::Io(io::Error::other(format!(
+            "failed to serialize Windows sandbox policy: {err}"
+        )))
+    })?;
     let sandbox_cwd = cwd.clone();
-    let logs_base_dir = find_codex_home().ok();
+    let codex_home = find_codex_home().map_err(|err| {
+        CodexErr::Io(io::Error::other(format!(
+            "windows sandbox: failed to resolve codex_home: {err}"
+        )))
+    })?;
     let spawn_res = tokio::task::spawn_blocking(move || {
         run_windows_sandbox_capture(
-            policy_str,
+            policy_str.as_str(),
             &sandbox_cwd,
+            codex_home.as_ref(),
             command,
             &cwd,
             env,
             timeout_ms,
-            logs_base_dir.as_deref(),
         )
     })
     .await;
@@ -440,7 +443,9 @@ async fn exec(
     stdout_stream: Option<StdoutStream>,
 ) -> Result<RawExecToolCallOutput> {
     #[cfg(target_os = "windows")]
-    if sandbox == SandboxType::WindowsRestrictedToken {
+    if sandbox == SandboxType::WindowsRestrictedToken
+        && !matches!(sandbox_policy, SandboxPolicy::DangerFullAccess)
+    {
         return exec_windows_sandbox(params, sandbox_policy).await;
     }
     let timeout = params.timeout_duration();

codex-rs/core/src/exec_policy.rs

@@ -1,173 +0,0 @@
-use std::fs;
-use std::path::Path;
-use std::path::PathBuf;
-use std::sync::Arc;
-
-use codex_execpolicy2::Decision;
-use codex_execpolicy2::Evaluation;
-use codex_execpolicy2::Policy;
-use codex_execpolicy2::PolicyParser;
-use codex_protocol::protocol::AskForApproval;
-use thiserror::Error;
-
-use crate::bash::parse_shell_lc_plain_commands;
-use crate::features::Feature;
-use crate::features::Features;
-use crate::tools::sandboxing::ApprovalRequirement;
-
-const FORBIDDEN_REASON: &str = "execpolicy forbids this command";
-const PROMPT_REASON: &str = "execpolicy requires approval for this command";
-
-#[derive(Debug, Error)]
-pub enum ExecPolicyError {
-    #[error("failed to read execpolicy files from {dir}: {source}")]
-    ReadDir {
-        dir: PathBuf,
-        source: std::io::Error,
-    },
-
-    #[error("failed to read execpolicy file {path}: {source}")]
-    ReadFile {
-        path: PathBuf,
-        source: std::io::Error,
-    },
-
-    #[error("failed to parse execpolicy file {path}: {source}")]
-    ParsePolicy {
-        path: String,
-        source: codex_execpolicy2::Error,
-    },
-}
-
-pub(crate) fn exec_policy_for(
-    features: &Features,
-    cwd: &Path,
-) -> Result<Option<Arc<Policy>>, ExecPolicyError> {
-    if !features.enabled(Feature::ExecPolicyV2) {
-        return Ok(None);
-    }
-
-    load_policy(cwd).map(Some)
-}
-
-pub(crate) fn evaluate_with_policy(
-    policy: &Policy,
-    command: &[String],
-    approval_policy: AskForApproval,
-) -> Option<ApprovalRequirement> {
-    let commands = parse_shell_lc_plain_commands(command).unwrap_or_else(|| vec![command.to_vec()]);
-    let evaluation = policy.check_multiple(commands.iter());
-
-    match evaluation {
-        Evaluation::Match { decision, .. } => match decision {
-            Decision::Forbidden => Some(ApprovalRequirement::Forbidden {
-                reason: FORBIDDEN_REASON.to_string(),
-            }),
-            Decision::Prompt => {
-                let reason = PROMPT_REASON.to_string();
-                if matches!(approval_policy, AskForApproval::Never) {
-                    Some(ApprovalRequirement::Forbidden { reason })
-                } else {
-                    Some(ApprovalRequirement::NeedsApproval {
-                        reason: Some(reason),
-                    })
-                }
-            }
-            Decision::Allow => Some(ApprovalRequirement::Skip),
-        },
-        Evaluation::NoMatch => None,
-    }
-}
-
-fn load_policy(cwd: &Path) -> Result<Arc<Policy>, ExecPolicyError> {
-    let codex_dir = cwd.join(".codex");
-    let entries = match fs::read_dir(&codex_dir) {
-        Ok(entries) => entries,
-        Err(source) => {
-            return Err(ExecPolicyError::ReadDir {
-                dir: codex_dir,
-                source,
-            });
-        }
-    };
-
-    let mut policy_paths: Vec<PathBuf> = Vec::new();
-    for entry in entries {
-        let entry = entry.map_err(|source| ExecPolicyError::ReadDir {
-            dir: codex_dir.clone(),
-            source,
-        })?;
-        let path = entry.path();
-        if path
-            .extension()
-            .and_then(|ext| ext.to_str())
-            .is_some_and(|ext| ext == "codexpolicy")
-            && path.is_file()
-        {
-            policy_paths.push(path);
-        }
-    }
-
-    policy_paths.sort();
-
-    let mut parser = PolicyParser::new();
-    for policy_path in &policy_paths {
-        let contents =
-            fs::read_to_string(policy_path).map_err(|source| ExecPolicyError::ReadFile {
-                path: policy_path.clone(),
-                source,
-            })?;
-        let identifier = policy_path.to_string_lossy().to_string();
-        parser
-            .parse(&identifier, &contents)
-            .map_err(|source| ExecPolicyError::ParsePolicy {
-                path: identifier,
-                source,
-            })?;
-    }
-
-    let policy = Arc::new(parser.build());
-    tracing::debug!(
-        file_count = policy_paths.len(),
-        "loaded execpolicy2 from {}",
-        codex_dir.display()
-    );
-
-    Ok(policy)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use codex_protocol::protocol::AskForApproval;
-    use pretty_assertions::assert_eq;
-
-    #[test]
-    fn evaluates_bash_lc_inner_commands() {
-        let policy_src = r#"
-prefix_rule(pattern=["rm"], decision="forbidden")
-"#;
-        let mut parser = PolicyParser::new();
-        parser
-            .parse("test.codexpolicy", policy_src)
-            .expect("parse policy");
-        let policy = parser.build();
-
-        let forbidden_script = vec![
-            "bash".to_string(),
-            "-lc".to_string(),
-            "rm -rf /tmp".to_string(),
-        ];
-
-        let requirement =
-            evaluate_with_policy(&policy, &forbidden_script, AskForApproval::OnRequest)
-                .expect("expected match for forbidden command");
-
-        assert_eq!(
-            requirement,
-            ApprovalRequirement::Forbidden {
-                reason: FORBIDDEN_REASON.to_string()
-            }
-        );
-    }
-}

codex-rs/core/src/features.rs

@@ -27,6 +27,8 @@ pub enum Stage {
 /// Unique features toggled via configuration.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
 pub enum Feature {
+    /// Create a ghost commit at each turn.
+    GhostCommit,
     /// Use the single unified PTY-backed exec tool.
     UnifiedExec,
     /// Use the shell command tool that takes `command` as a single string of
@@ -40,14 +42,16 @@ pub enum Feature {
     ViewImageTool,
     /// Allow the model to request web searches.
     WebSearchRequest,
-    /// Gate the execpolicy2 enforcement for shell/unified exec.
-    ExecPolicyV2,
     /// Enable the model-based risk assessments for sandboxed commands.
     SandboxCommandAssessment,
-    /// Create a ghost commit at each turn.
-    GhostCommit,
     /// Enable Windows sandbox (restricted token) on Windows.
     WindowsSandbox,
+    /// Remote compaction enabled (only for ChatGPT auth)
+    RemoteCompaction,
+    /// Enable the default shell tool.
+    ShellTool,
+    /// Allow model to call multiple tools in parallel (only for models supporting it).
+    ParallelToolCalls,
 }
 
 impl Feature {
@@ -249,6 +253,14 @@ pub struct FeatureSpec {
 }
 
 pub const FEATURES: &[FeatureSpec] = &[
+    // Stable features.
+    FeatureSpec {
+        id: Feature::GhostCommit,
+        key: "undo",
+        stage: Stage::Stable,
+        default_enabled: true,
+    },
+    // Unstable features.
     FeatureSpec {
         id: Feature::UnifiedExec,
         key: "unified_exec",
@@ -285,28 +297,34 @@ pub const FEATURES: &[FeatureSpec] = &[
         stage: Stage::Stable,
         default_enabled: false,
     },
-    FeatureSpec {
-        id: Feature::ExecPolicyV2,
-        key: "exec_policy_v2",
-        stage: Stage::Experimental,
-        default_enabled: false,
-    },
     FeatureSpec {
         id: Feature::SandboxCommandAssessment,
         key: "experimental_sandbox_command_assessment",
         stage: Stage::Experimental,
         default_enabled: false,
     },
-    FeatureSpec {
-        id: Feature::GhostCommit,
-        key: "ghost_commit",
-        stage: Stage::Experimental,
-        default_enabled: true,
-    },
     FeatureSpec {
         id: Feature::WindowsSandbox,
         key: "enable_experimental_windows_sandbox",
         stage: Stage::Experimental,
         default_enabled: false,
     },
+    FeatureSpec {
+        id: Feature::RemoteCompaction,
+        key: "remote_compaction",
+        stage: Stage::Experimental,
+        default_enabled: true,
+    },
+    FeatureSpec {
+        id: Feature::ParallelToolCalls,
+        key: "parallel",
+        stage: Stage::Experimental,
+        default_enabled: false,
+    },
+    FeatureSpec {
+        id: Feature::ShellTool,
+        key: "shell_tool",
+        stage: Stage::Stable,
+        default_enabled: true,
+    },
 ];

codex-rs/core/src/lib.rs

@@ -13,6 +13,7 @@ mod client;
 mod client_common;
 pub mod codex;
 mod codex_conversation;
+mod compact_remote;
 pub use codex_conversation::CodexConversation;
 mod codex_delegate;
 mod command_safety;
@@ -24,7 +25,6 @@ mod environment_context;
 pub mod error;
 pub mod exec;
 pub mod exec_env;
-mod exec_policy;
 pub mod features;
 mod flags;
 pub mod git_info;
@@ -35,14 +35,18 @@ mod mcp_tool_call;
 mod message_history;
 mod model_provider_info;
 pub mod parse_command;
+pub mod powershell;
 mod response_processing;
 pub mod sandboxing;
 pub mod token_data;
 mod truncate;
 mod unified_exec;
 mod user_instructions;
-pub use model_provider_info::BUILT_IN_OSS_MODEL_PROVIDER_ID;
+pub use model_provider_info::DEFAULT_LMSTUDIO_PORT;
+pub use model_provider_info::DEFAULT_OLLAMA_PORT;
+pub use model_provider_info::LMSTUDIO_OSS_PROVIDER_ID;
 pub use model_provider_info::ModelProviderInfo;
+pub use model_provider_info::OLLAMA_OSS_PROVIDER_ID;
 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;

codex-rs/core/src/model_family.rs

@@ -4,6 +4,7 @@ use codex_protocol::config_types::Verbosity;
 use crate::config::types::ReasoningSummaryFormat;
 use crate::tools::handlers::apply_patch::ApplyPatchToolType;
 use crate::tools::spec::ConfigShellToolType;
+use crate::truncate::TruncationPolicy;
 
 /// The `instructions` field in the payload sent to a model should always start
 /// with this content.
@@ -66,6 +67,8 @@ pub struct ModelFamily {
 
     /// Preferred shell tool type for this model family when features do not override it.
     pub shell_type: ConfigShellToolType,
+
+    pub truncation_policy: TruncationPolicy,
 }
 
 macro_rules! model_family {
@@ -89,6 +92,7 @@ macro_rules! model_family {
             shell_type: ConfigShellToolType::Default,
             default_verbosity: None,
             default_reasoning_effort: None,
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         };
 
         // apply overrides
@@ -132,7 +136,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
         model_family!(slug, "gpt-4o", needs_special_apply_patch_instructions: true)
     } else if slug.starts_with("gpt-3.5") {
         model_family!(slug, "gpt-3.5", needs_special_apply_patch_instructions: true)
-    } else if slug.starts_with("test-gpt-5-codex") {
+    } else if slug.starts_with("test-gpt-5") {
         model_family!(
             slug, slug,
             supports_reasoning_summaries: true,
@@ -145,7 +149,9 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
                 "test_sync_tool".to_string(),
             ],
             supports_parallel_tool_calls: true,
+            shell_type: ConfigShellToolType::ShellCommand,
             support_verbosity: true,
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
 
     // Internal models.
@@ -161,8 +167,10 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
                 "list_dir".to_string(),
                 "read_file".to_string(),
             ],
+            shell_type: ConfigShellToolType::ShellCommand,
             supports_parallel_tool_calls: true,
             support_verbosity: true,
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
 
     // Production models.
@@ -176,7 +184,10 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             reasoning_summary_format: ReasoningSummaryFormat::Experimental,
             base_instructions: GPT_5_CODEX_INSTRUCTIONS.to_string(),
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
+            shell_type: ConfigShellToolType::ShellCommand,
+            supports_parallel_tool_calls: true,
             support_verbosity: false,
+            truncation_policy: TruncationPolicy::Tokens(10_000),
         )
     } else if slug.starts_with("gpt-5.1") {
         model_family!(
@@ -187,13 +198,18 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
             default_verbosity: Some(Verbosity::Low),
             base_instructions: GPT_5_1_INSTRUCTIONS.to_string(),
             default_reasoning_effort: Some(ReasoningEffort::Medium),
+            truncation_policy: TruncationPolicy::Bytes(10_000),
+            shell_type: ConfigShellToolType::ShellCommand,
+            supports_parallel_tool_calls: true,
         )
     } else if slug.starts_with("gpt-5") {
         model_family!(
             slug, "gpt-5",
             supports_reasoning_summaries: true,
             needs_special_apply_patch_instructions: true,
+            shell_type: ConfigShellToolType::ShellCommand,
             support_verbosity: true,
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else {
         None
@@ -216,5 +232,6 @@ pub fn derive_default_model_family(model: &str) -> ModelFamily {
         shell_type: ConfigShellToolType::Default,
         default_verbosity: None,
         default_reasoning_effort: None,
+        truncation_policy: TruncationPolicy::Bytes(10_000),
     }
 }

codex-rs/core/src/model_provider_info.rs

@@ -8,6 +8,7 @@
 use crate::CodexAuth;
 use crate::default_client::CodexHttpClient;
 use crate::default_client::CodexRequestBuilder;
+use crate::error::CodexErr;
 use codex_app_server_protocol::AuthMode;
 use serde::Deserialize;
 use serde::Serialize;
@@ -109,21 +110,7 @@ impl ModelProviderInfo {
         client: &'a CodexHttpClient,
         auth: &Option<CodexAuth>,
     ) -> crate::error::Result<CodexRequestBuilder> {
-        let effective_auth = if let Some(secret_key) = &self.experimental_bearer_token {
-            Some(CodexAuth::from_api_key(secret_key))
-        } else {
-            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 effective_auth = self.effective_auth(auth)?;
 
         let url = self.get_full_url(&effective_auth);
 
@@ -136,6 +123,51 @@ impl ModelProviderInfo {
         Ok(self.apply_http_headers(builder))
     }
 
+    pub async fn create_compact_request_builder<'a>(
+        &'a self,
+        client: &'a CodexHttpClient,
+        auth: &Option<CodexAuth>,
+    ) -> crate::error::Result<CodexRequestBuilder> {
+        if self.wire_api != WireApi::Responses {
+            return Err(CodexErr::UnsupportedOperation(
+                "Compaction endpoint requires Responses API providers".to_string(),
+            ));
+        }
+        let effective_auth = self.effective_auth(auth)?;
+
+        let url = self.get_compact_url(&effective_auth).ok_or_else(|| {
+            CodexErr::UnsupportedOperation(
+                "Compaction endpoint requires Responses API providers".to_string(),
+            )
+        })?;
+
+        let mut builder = client.post(url);
+
+        if let Some(auth) = effective_auth.as_ref() {
+            builder = builder.bearer_auth(auth.get_token().await?);
+        }
+
+        Ok(self.apply_http_headers(builder))
+    }
+
+    fn effective_auth(&self, auth: &Option<CodexAuth>) -> crate::error::Result<Option<CodexAuth>> {
+        if let Some(secret_key) = &self.experimental_bearer_token {
+            return Ok(Some(CodexAuth::from_api_key(secret_key)));
+        }
+
+        match self.api_key() {
+            Ok(Some(key)) => Ok(Some(CodexAuth::from_api_key(&key))),
+            Ok(None) => Ok(auth.clone()),
+            Err(err) => {
+                if auth.is_some() {
+                    Ok(auth.clone())
+                } else {
+                    Err(err)
+                }
+            }
+        }
+    }
+
     fn get_query_string(&self) -> String {
         self.query_params
             .as_ref()
@@ -173,6 +205,18 @@ impl ModelProviderInfo {
         }
     }
 
+    pub(crate) fn get_compact_url(&self, auth: &Option<CodexAuth>) -> Option<String> {
+        if self.wire_api != WireApi::Responses {
+            return None;
+        }
+        let full = self.get_full_url(auth);
+        if let Some((path, query)) = full.split_once('?') {
+            Some(format!("{path}/compact?{query}"))
+        } else {
+            Some(format!("{full}/compact"))
+        }
+    }
+
     pub(crate) fn is_azure_responses_endpoint(&self) -> bool {
         if self.wire_api != WireApi::Responses {
             return false;
@@ -258,9 +302,11 @@ impl ModelProviderInfo {
     }
 }
 
-const DEFAULT_OLLAMA_PORT: u32 = 11434;
+pub const DEFAULT_LMSTUDIO_PORT: u16 = 1234;
+pub const DEFAULT_OLLAMA_PORT: u16 = 11434;
 
-pub const BUILT_IN_OSS_MODEL_PROVIDER_ID: &str = "oss";
+pub const LMSTUDIO_OSS_PROVIDER_ID: &str = "lmstudio";
+pub const OLLAMA_OSS_PROVIDER_ID: &str = "ollama";
 
 /// Built-in default provider list.
 pub fn built_in_model_providers() -> HashMap<String, ModelProviderInfo> {
@@ -311,14 +357,21 @@ pub fn built_in_model_providers() -> HashMap<String, ModelProviderInfo> {
                 requires_openai_auth: true,
             },
         ),
-        (BUILT_IN_OSS_MODEL_PROVIDER_ID, create_oss_provider()),
+        (
+            OLLAMA_OSS_PROVIDER_ID,
+            create_oss_provider(DEFAULT_OLLAMA_PORT, WireApi::Chat),
+        ),
+        (
+            LMSTUDIO_OSS_PROVIDER_ID,
+            create_oss_provider(DEFAULT_LMSTUDIO_PORT, WireApi::Responses),
+        ),
     ]
     .into_iter()
     .map(|(k, v)| (k.to_string(), v))
     .collect()
 }
 
-pub fn create_oss_provider() -> ModelProviderInfo {
+pub fn create_oss_provider(default_provider_port: u16, wire_api: WireApi) -> 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")
@@ -331,22 +384,21 @@ pub fn create_oss_provider() -> ModelProviderInfo {
             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)
+                .and_then(|v| v.parse::<u16>().ok())
+                .unwrap_or(default_provider_port)
         ),
     };
-
-    create_oss_provider_with_base_url(&codex_oss_base_url)
+    create_oss_provider_with_base_url(&codex_oss_base_url, wire_api)
 }
 
-pub fn create_oss_provider_with_base_url(base_url: &str) -> ModelProviderInfo {
+pub fn create_oss_provider_with_base_url(base_url: &str, wire_api: WireApi) -> ModelProviderInfo {
     ModelProviderInfo {
         name: "gpt-oss".into(),
         base_url: Some(base_url.into()),
         env_key: None,
         env_key_instructions: None,
         experimental_bearer_token: None,
-        wire_api: WireApi::Chat,
+        wire_api,
         query_params: None,
         http_headers: None,
         env_http_headers: None,

codex-rs/core/src/otel_init.rs

@@ -5,6 +5,7 @@ use crate::default_client::originator;
 use codex_otel::config::OtelExporter;
 use codex_otel::config::OtelHttpProtocol;
 use codex_otel::config::OtelSettings;
+use codex_otel::config::OtelTlsConfig as OtelTlsSettings;
 use codex_otel::otel_provider::OtelProvider;
 use std::error::Error;
 
@@ -21,6 +22,7 @@ pub fn build_provider(
             endpoint,
             headers,
             protocol,
+            tls,
         } => {
             let protocol = match protocol {
                 Protocol::Json => OtelHttpProtocol::Json,
@@ -34,14 +36,28 @@ pub fn build_provider(
                     .map(|(k, v)| (k.clone(), v.clone()))
                     .collect(),
                 protocol,
+                tls: tls.as_ref().map(|config| OtelTlsSettings {
+                    ca_certificate: config.ca_certificate.clone(),
+                    client_certificate: config.client_certificate.clone(),
+                    client_private_key: config.client_private_key.clone(),
+                }),
             }
         }
-        Kind::OtlpGrpc { endpoint, headers } => OtelExporter::OtlpGrpc {
+        Kind::OtlpGrpc {
+            endpoint,
+            headers,
+            tls,
+        } => OtelExporter::OtlpGrpc {
             endpoint: endpoint.clone(),
             headers: headers
                 .iter()
                 .map(|(k, v)| (k.clone(), v.clone()))
                 .collect(),
+            tls: tls.as_ref().map(|config| OtelTlsSettings {
+                ca_certificate: config.ca_certificate.clone(),
+                client_certificate: config.client_certificate.clone(),
+                client_private_key: config.client_private_key.clone(),
+            }),
         },
     };
 

codex-rs/core/src/parse_command.rs

@@ -1,16 +1,22 @@
 use crate::bash::extract_bash_command;
 use crate::bash::try_parse_shell;
 use crate::bash::try_parse_word_only_commands_sequence;
+use crate::powershell::extract_powershell_command;
 use codex_protocol::parse_command::ParsedCommand;
 use shlex::split as shlex_split;
 use shlex::try_join as shlex_try_join;
 use std::path::PathBuf;
 
-fn shlex_join(tokens: &[String]) -> String {
+pub fn shlex_join(tokens: &[String]) -> String {
     shlex_try_join(tokens.iter().map(String::as_str))
         .unwrap_or_else(|_| "<command included NUL byte>".to_string())
 }
 
+/// Extracts the shell and script from a command, regardless of platform
+pub fn extract_shell_command(command: &[String]) -> Option<(&str, &str)> {
+    extract_bash_command(command).or_else(|| extract_powershell_command(command))
+}
+
 /// DO NOT REVIEW THIS CODE BY HAND
 /// This parsing code is quite complex and not easy to hand-modify.
 /// The easiest way to iterate is to add unit tests and have Codex fix the implementation.
@@ -877,6 +883,87 @@ mod tests {
             }],
         );
     }
+
+    #[test]
+    fn powershell_command_is_stripped() {
+        assert_parsed(
+            &vec_str(&["powershell", "-Command", "Get-ChildItem"]),
+            vec![ParsedCommand::Unknown {
+                cmd: "Get-ChildItem".to_string(),
+            }],
+        );
+    }
+
+    #[test]
+    fn pwsh_with_noprofile_and_c_alias_is_stripped() {
+        assert_parsed(
+            &vec_str(&["pwsh", "-NoProfile", "-c", "Write-Host hi"]),
+            vec![ParsedCommand::Unknown {
+                cmd: "Write-Host hi".to_string(),
+            }],
+        );
+    }
+
+    #[test]
+    fn powershell_with_path_is_stripped() {
+        let command = if cfg!(windows) {
+            "C:\\windows\\System32\\WindowsPowerShell\\v1.0\\powershell.exe"
+        } else {
+            "/usr/local/bin/powershell.exe"
+        };
+
+        assert_parsed(
+            &vec_str(&[command, "-NoProfile", "-c", "Write-Host hi"]),
+            vec![ParsedCommand::Unknown {
+                cmd: "Write-Host hi".to_string(),
+            }],
+        );
+    }
+
+    #[test]
+    fn powershell_cd_is_stripped() {
+        assert_parsed(
+            &vec_str(&["powershell", "-Command", "cd foo; Get-ChildItem"]),
+            vec![ParsedCommand::Unknown {
+                cmd: "Get-ChildItem".to_string(),
+            }],
+        );
+    }
+    #[test]
+    fn powershell_pipes_are_preserved() {
+        assert_parsed(
+            &vec_str(&[
+                "powershell",
+                "-Command",
+                "Get-Content core/src/parse_command.rs | Select-Object -First 980 | Select-Object -Last 140",
+            ]),
+            vec![
+                ParsedCommand::Unknown {
+                    cmd: "Get-Content core/src/parse_command.rs".to_string(),
+                },
+                ParsedCommand::Unknown {
+                    cmd: "Select-Object -First 980".to_string(),
+                },
+                ParsedCommand::Unknown {
+                    cmd: "Select-Object -Last 140".to_string(),
+                },
+            ],
+        );
+    }
+
+    #[test]
+    fn powershell_at_escaping_is_preserved() {
+        assert_parsed(
+            &vec_str(&[
+                "powershell",
+                "-Command",
+                "@'\\nprint('Hello, world!')\\n'@ | python -",
+            ]),
+            vec![ParsedCommand::Unknown {
+                cmd: "@'\\nprint('Hello, world!')\\n'@ | python -".to_string(),
+            }],
+        );
+    }
 }
 
 pub fn parse_command_impl(command: &[String]) -> Vec<ParsedCommand> {
@@ -1190,7 +1277,8 @@ fn parse_find_query_and_path(tail: &[String]) -> (Option<String>, Option<String>
 }
 
 fn parse_shell_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
-    let (_, script) = extract_bash_command(original)?;
+    let (_, script) =
+        extract_bash_command(original).or_else(|| extract_powershell_command(original))?;
 
     if let Some(tree) = try_parse_shell(script)
         && let Some(all_commands) = try_parse_word_only_commands_sequence(&tree, script)

codex-rs/core/src/powershell.rs

@@ -0,0 +1,93 @@
+use std::path::PathBuf;
+
+use crate::shell::ShellType;
+use crate::shell::detect_shell_type;
+
+const POWERSHELL_FLAGS: &[&str] = &["-nologo", "-noprofile", "-command", "-c"];
+
+/// Extract the PowerShell script body from an invocation such as:
+///
+/// - ["pwsh", "-NoProfile", "-Command", "Get-ChildItem -Recurse | Select-String foo"]
+/// - ["powershell.exe", "-Command", "Write-Host hi"]
+/// - ["powershell", "-NoLogo", "-NoProfile", "-Command", "...script..."]
+///
+/// Returns (`shell`, `script`) when the first arg is a PowerShell executable and a
+/// `-Command` (or `-c`) flag is present followed by a script string.
+pub fn extract_powershell_command(command: &[String]) -> Option<(&str, &str)> {
+    if command.len() < 3 {
+        return None;
+    }
+
+    let shell = &command[0];
+    if detect_shell_type(&PathBuf::from(shell)) != Some(ShellType::PowerShell) {
+        return None;
+    }
+
+    // Find the first occurrence of -Command (accept common short alias -c as well)
+    let mut i = 1usize;
+    while i + 1 < command.len() {
+        let flag = &command[i];
+        // Reject unknown flags
+        if !POWERSHELL_FLAGS.contains(&flag.to_ascii_lowercase().as_str()) {
+            return None;
+        }
+        if flag.eq_ignore_ascii_case("-Command") || flag.eq_ignore_ascii_case("-c") {
+            let script = &command[i + 1];
+            return Some((shell, script.as_str()));
+        }
+        i += 1;
+    }
+    None
+}
+
+#[cfg(test)]
+mod tests {
+    use super::extract_powershell_command;
+
+    #[test]
+    fn extracts_basic_powershell_command() {
+        let cmd = vec![
+            "powershell".to_string(),
+            "-Command".to_string(),
+            "Write-Host hi".to_string(),
+        ];
+        let (_shell, script) = extract_powershell_command(&cmd).expect("extract");
+        assert_eq!(script, "Write-Host hi");
+    }
+
+    #[test]
+    fn extracts_lowercase_flags() {
+        let cmd = vec![
+            "powershell".to_string(),
+            "-nologo".to_string(),
+            "-command".to_string(),
+            "Write-Host hi".to_string(),
+        ];
+        let (_shell, script) = extract_powershell_command(&cmd).expect("extract");
+        assert_eq!(script, "Write-Host hi");
+    }
+
+    #[test]
+    fn extracts_full_path_powershell_command() {
+        let command = if cfg!(windows) {
+            "C:\\windows\\System32\\WindowsPowerShell\\v1.0\\powershell.exe".to_string()
+        } else {
+            "/usr/local/bin/powershell.exe".to_string()
+        };
+        let cmd = vec![command, "-Command".to_string(), "Write-Host hi".to_string()];
+        let (_shell, script) = extract_powershell_command(&cmd).expect("extract");
+        assert_eq!(script, "Write-Host hi");
+    }
+
+    #[test]
+    fn extracts_with_noprofile_and_alias() {
+        let cmd = vec![
+            "pwsh".to_string(),
+            "-NoProfile".to_string(),
+            "-c".to_string(),
+            "Get-ChildItem | Select-String foo".to_string(),
+        ];
+        let (_shell, script) = extract_powershell_command(&cmd).expect("extract");
+        assert_eq!(script, "Get-ChildItem | Select-String foo");
+    }
+}

codex-rs/core/src/response_processing.rs

@@ -13,92 +13,58 @@ pub(crate) async fn process_items(
     sess: &Session,
     turn_context: &TurnContext,
 ) -> (Vec<ResponseInputItem>, Vec<ResponseItem>) {
-    let mut items_to_record_in_conversation_history = Vec::<ResponseItem>::new();
+    let mut outputs_to_record = Vec::<ResponseItem>::new();
+    let mut new_inputs_to_record = Vec::<ResponseItem>::new();
     let mut responses = Vec::<ResponseInputItem>::new();
     for processed_response_item in processed_items {
         let crate::codex::ProcessedResponseItem { item, response } = processed_response_item;
-        match (&item, &response) {
-            (ResponseItem::Message { role, .. }, None) if role == "assistant" => {
-                // If the model returned a message, we need to record it.
-                items_to_record_in_conversation_history.push(item);
-            }
-            (
-                ResponseItem::LocalShellCall { .. },
-                Some(ResponseInputItem::FunctionCallOutput { call_id, output }),
-            ) => {
-                items_to_record_in_conversation_history.push(item);
-                items_to_record_in_conversation_history.push(ResponseItem::FunctionCallOutput {
+
+        if let Some(response) = &response {
+            responses.push(response.clone());
+        }
+
+        match response {
+            Some(ResponseInputItem::FunctionCallOutput { call_id, output }) => {
+                new_inputs_to_record.push(ResponseItem::FunctionCallOutput {
                     call_id: call_id.clone(),
                     output: output.clone(),
                 });
             }
-            (
-                ResponseItem::FunctionCall { .. },
-                Some(ResponseInputItem::FunctionCallOutput { call_id, output }),
-            ) => {
-                items_to_record_in_conversation_history.push(item);
-                items_to_record_in_conversation_history.push(ResponseItem::FunctionCallOutput {
+
+            Some(ResponseInputItem::CustomToolCallOutput { call_id, output }) => {
+                new_inputs_to_record.push(ResponseItem::CustomToolCallOutput {
                     call_id: call_id.clone(),
                     output: output.clone(),
                 });
             }
-            (
-                ResponseItem::CustomToolCall { .. },
-                Some(ResponseInputItem::CustomToolCallOutput { call_id, output }),
-            ) => {
-                items_to_record_in_conversation_history.push(item);
-                items_to_record_in_conversation_history.push(ResponseItem::CustomToolCallOutput {
-                    call_id: call_id.clone(),
-                    output: output.clone(),
-                });
-            }
-            (
-                ResponseItem::FunctionCall { .. },
-                Some(ResponseInputItem::McpToolCallOutput { call_id, result }),
-            ) => {
-                items_to_record_in_conversation_history.push(item);
+            Some(ResponseInputItem::McpToolCallOutput { call_id, result }) => {
                 let output = match result {
-                    Ok(call_tool_result) => FunctionCallOutputPayload::from(call_tool_result),
+                    Ok(call_tool_result) => FunctionCallOutputPayload::from(&call_tool_result),
                     Err(err) => FunctionCallOutputPayload {
                         content: err.clone(),
                         success: Some(false),
                         ..Default::default()
                     },
                 };
-                items_to_record_in_conversation_history.push(ResponseItem::FunctionCallOutput {
+                new_inputs_to_record.push(ResponseItem::FunctionCallOutput {
                     call_id: call_id.clone(),
                     output,
                 });
             }
-            (
-                ResponseItem::Reasoning {
-                    id,
-                    summary,
-                    content,
-                    encrypted_content,
-                },
-                None,
-            ) => {
-                items_to_record_in_conversation_history.push(ResponseItem::Reasoning {
-                    id: id.clone(),
-                    summary: summary.clone(),
-                    content: content.clone(),
-                    encrypted_content: encrypted_content.clone(),
-                });
-            }
+            None => {}
             _ => {
                 warn!("Unexpected response item: {item:?} with response: {response:?}");
             }
         };
-        if let Some(response) = response {
-            responses.push(response);
-        }
+
+        outputs_to_record.push(item);
     }
 
+    let all_items_to_record = [outputs_to_record, new_inputs_to_record].concat();
     // Only attempt to take the lock if there is something to record.
-    if !items_to_record_in_conversation_history.is_empty() {
-        sess.record_conversation_items(turn_context, &items_to_record_in_conversation_history)
+    if !all_items_to_record.is_empty() {
+        sess.record_conversation_items(turn_context, &all_items_to_record)
             .await;
     }
-    (responses, items_to_record_in_conversation_history)
+    (responses, all_items_to_record)
 }

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

@@ -27,7 +27,8 @@ pub(crate) fn should_persist_response_item(item: &ResponseItem) -> bool {
         | ResponseItem::CustomToolCall { .. }
         | ResponseItem::CustomToolCallOutput { .. }
         | ResponseItem::WebSearchCall { .. }
-        | ResponseItem::GhostSnapshot { .. } => true,
+        | ResponseItem::GhostSnapshot { .. }
+        | ResponseItem::CompactionSummary { .. } => true,
         ResponseItem::Other => false,
     }
 }
@@ -72,6 +73,8 @@ pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
         | EventMsg::GetHistoryEntryResponse(_)
         | EventMsg::UndoStarted(_)
         | EventMsg::McpListToolsResponse(_)
+        | EventMsg::McpStartupUpdate(_)
+        | EventMsg::McpStartupComplete(_)
         | EventMsg::ListCustomPromptsResponse(_)
         | EventMsg::PlanUpdate(_)
         | EventMsg::ShutdownComplete

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

@@ -814,6 +814,7 @@ async fn test_tail_skips_trailing_non_responses() -> Result<()> {
         timestamp: format!("{ts}-compacted"),
         item: RolloutItem::Compacted(CompactedItem {
             message: "compacted".into(),
+            replacement_history: None,
         }),
     };
     writeln!(file, "{}", serde_json::to_string(&compacted_line)?)?;

codex-rs/core/src/shell.rs

@@ -61,10 +61,7 @@ impl Shell {
                 ]
             }
             Shell::PowerShell(ps) => {
-                let mut args = vec![
-                    ps.shell_path.to_string_lossy().to_string(),
-                    "-NoLogo".to_string(),
-                ];
+                let mut args = vec![ps.shell_path.to_string_lossy().to_string()];
                 if !use_login_shell {
                     args.push("-NoProfile".to_string());
                 }
@@ -192,7 +189,6 @@ pub fn detect_shell_type(shell_path: &PathBuf) -> Option<ShellType> {
         Some("powershell") => Some(ShellType::PowerShell),
         _ => {
             let shell_name = shell_path.file_stem();
-
             if let Some(shell_name) = shell_name
                 && shell_name != shell_path
             {
@@ -251,6 +247,14 @@ mod detect_shell_type_tests {
             detect_shell_type(&PathBuf::from("powershell.exe")),
             Some(ShellType::PowerShell)
         );
+        assert_eq!(
+            detect_shell_type(&PathBuf::from(if cfg!(windows) {
+                "C:\\windows\\System32\\WindowsPowerShell\\v1.0\\powershell.exe"
+            } else {
+                "/usr/local/bin/pwsh"
+            })),
+            Some(ShellType::PowerShell)
+        );
         assert_eq!(
             detect_shell_type(&PathBuf::from("pwsh.exe")),
             Some(ShellType::PowerShell)

codex-rs/core/src/state/service.rs

@@ -8,9 +8,12 @@ use crate::unified_exec::UnifiedExecSessionManager;
 use crate::user_notification::UserNotifier;
 use codex_otel::otel_event_manager::OtelEventManager;
 use tokio::sync::Mutex;
+use tokio::sync::RwLock;
+use tokio_util::sync::CancellationToken;
 
 pub(crate) struct SessionServices {
-    pub(crate) mcp_connection_manager: McpConnectionManager,
+    pub(crate) mcp_connection_manager: Arc<RwLock<McpConnectionManager>>,
+    pub(crate) mcp_startup_cancellation_token: CancellationToken,
     pub(crate) unified_exec_manager: UnifiedExecSessionManager,
     pub(crate) notifier: UserNotifier,
     pub(crate) rollout: Mutex<Option<RolloutRecorder>>,

codex-rs/core/src/state/session.rs

@@ -7,6 +7,7 @@ use crate::context_manager::ContextManager;
 use crate::protocol::RateLimitSnapshot;
 use crate::protocol::TokenUsage;
 use crate::protocol::TokenUsageInfo;
+use crate::truncate::TruncationPolicy;
 
 /// Persistent, session-scoped state previously stored directly on `Session`.
 pub(crate) struct SessionState {
@@ -18,20 +19,21 @@ pub(crate) struct SessionState {
 impl SessionState {
     /// Create a new session state mirroring previous `State::default()` semantics.
     pub(crate) fn new(session_configuration: SessionConfiguration) -> Self {
+        let history = ContextManager::new();
         Self {
             session_configuration,
-            history: ContextManager::new(),
+            history,
             latest_rate_limits: None,
         }
     }
 
     // History helpers
-    pub(crate) fn record_items<I>(&mut self, items: I)
+    pub(crate) fn record_items<I>(&mut self, items: I, policy: TruncationPolicy)
     where
         I: IntoIterator,
         I::Item: std::ops::Deref<Target = ResponseItem>,
     {
-        self.history.record_items(items)
+        self.history.record_items(items, policy);
     }
 
     pub(crate) fn clone_history(&self) -> ContextManager {

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

@@ -1,15 +1,14 @@
 use std::sync::Arc;
 
-use async_trait::async_trait;
-use tokio_util::sync::CancellationToken;
-
-use crate::codex::TurnContext;
-use crate::compact;
-use crate::state::TaskKind;
-use codex_protocol::user_input::UserInput;
-
 use super::SessionTask;
 use super::SessionTaskContext;
+use crate::codex::TurnContext;
+use crate::features::Feature;
+use crate::state::TaskKind;
+use async_trait::async_trait;
+use codex_app_server_protocol::AuthMode;
+use codex_protocol::user_input::UserInput;
+use tokio_util::sync::CancellationToken;
 
 #[derive(Clone, Copy, Default)]
 pub(crate) struct CompactTask;
@@ -27,6 +26,17 @@ impl SessionTask for CompactTask {
         input: Vec<UserInput>,
         _cancellation_token: CancellationToken,
     ) -> Option<String> {
-        compact::run_compact_task(session.clone_session(), ctx, input).await
+        let session = session.clone_session();
+        if session
+            .services
+            .auth_manager
+            .auth()
+            .is_some_and(|auth| auth.mode == AuthMode::ChatGPT)
+            && session.enabled(Feature::RemoteCompaction).await
+        {
+            crate::compact_remote::run_remote_compact_task(session, ctx).await
+        } else {
+            crate::compact::run_compact_task(session, ctx, input).await
+        }
     }
 }

codex-rs/core/src/tasks/review.rs

@@ -23,8 +23,18 @@ use codex_protocol::user_input::UserInput;
 use super::SessionTask;
 use super::SessionTaskContext;
 
-#[derive(Clone, Copy, Default)]
-pub(crate) struct ReviewTask;
+#[derive(Clone, Copy)]
+pub(crate) struct ReviewTask {
+    append_to_original_thread: bool,
+}
+
+impl ReviewTask {
+    pub(crate) fn new(append_to_original_thread: bool) -> Self {
+        Self {
+            append_to_original_thread,
+        }
+    }
+}
 
 #[async_trait]
 impl SessionTask for ReviewTask {
@@ -52,13 +62,25 @@ impl SessionTask for ReviewTask {
             None => None,
         };
         if !cancellation_token.is_cancelled() {
-            exit_review_mode(session.clone_session(), output.clone(), ctx.clone()).await;
+            exit_review_mode(
+                session.clone_session(),
+                output.clone(),
+                ctx.clone(),
+                self.append_to_original_thread,
+            )
+            .await;
         }
         None
     }
 
     async fn abort(&self, session: Arc<SessionTaskContext>, ctx: Arc<TurnContext>) {
-        exit_review_mode(session.clone_session(), None, ctx).await;
+        exit_review_mode(
+            session.clone_session(),
+            None,
+            ctx,
+            self.append_to_original_thread,
+        )
+        .await;
     }
 }
 
@@ -175,32 +197,35 @@ pub(crate) async fn exit_review_mode(
     session: Arc<Session>,
     review_output: Option<ReviewOutputEvent>,
     ctx: Arc<TurnContext>,
+    append_to_original_thread: bool,
 ) {
-    let user_message = if let Some(out) = review_output.clone() {
-        let mut findings_str = String::new();
-        let text = out.overall_explanation.trim();
-        if !text.is_empty() {
-            findings_str.push_str(text);
-        }
-        if !out.findings.is_empty() {
-            let block = format_review_findings_block(&out.findings, None);
-            findings_str.push_str(&format!("\n{block}"));
-        }
-        crate::client_common::REVIEW_EXIT_SUCCESS_TMPL.replace("{results}", &findings_str)
-    } else {
-        crate::client_common::REVIEW_EXIT_INTERRUPTED_TMPL.to_string()
-    };
+    if append_to_original_thread {
+        let user_message = if let Some(out) = review_output.clone() {
+            let mut findings_str = String::new();
+            let text = out.overall_explanation.trim();
+            if !text.is_empty() {
+                findings_str.push_str(text);
+            }
+            if !out.findings.is_empty() {
+                let block = format_review_findings_block(&out.findings, None);
+                findings_str.push_str(&format!("\n{block}"));
+            }
+            crate::client_common::REVIEW_EXIT_SUCCESS_TMPL.replace("{results}", &findings_str)
+        } else {
+            crate::client_common::REVIEW_EXIT_INTERRUPTED_TMPL.to_string()
+        };
 
-    session
-        .record_conversation_items(
-            &ctx,
-            &[ResponseItem::Message {
-                id: None,
-                role: "user".to_string(),
-                content: vec![ContentItem::InputText { text: user_message }],
-            }],
-        )
-        .await;
+        session
+            .record_conversation_items(
+                &ctx,
+                &[ResponseItem::Message {
+                    id: None,
+                    role: "user".to_string(),
+                    content: vec![ContentItem::InputText { text: user_message }],
+                }],
+            )
+            .await;
+    }
     session
         .send_event(
             ctx.as_ref(),

codex-rs/core/src/tasks/user_shell.rs

@@ -20,6 +20,7 @@ use crate::parse_command::parse_command;
 use crate::protocol::EventMsg;
 use crate::protocol::ExecCommandBeginEvent;
 use crate::protocol::ExecCommandEndEvent;
+use crate::protocol::ExecCommandSource;
 use crate::protocol::SandboxPolicy;
 use crate::protocol::TaskStartedEvent;
 use crate::sandboxing::ExecEnv;
@@ -64,30 +65,33 @@ impl SessionTask for UserShellCommandTask {
         // allows commands that use shell features (pipes, &&, redirects, etc.).
         // We do not source rc files or otherwise reformat the script.
         let use_login_shell = true;
-        let shell_invocation = session
+        let command = session
             .user_shell()
             .derive_exec_args(&self.command, use_login_shell);
 
         let call_id = Uuid::new_v4().to_string();
         let raw_command = self.command.clone();
+        let cwd = turn_context.cwd.clone();
 
-        let parsed_cmd = parse_command(&shell_invocation);
+        let parsed_cmd = parse_command(&command);
         session
             .send_event(
                 turn_context.as_ref(),
                 EventMsg::ExecCommandBegin(ExecCommandBeginEvent {
                     call_id: call_id.clone(),
-                    command: shell_invocation.clone(),
-                    cwd: turn_context.cwd.clone(),
-                    parsed_cmd,
-                    is_user_shell_command: true,
+                    turn_id: turn_context.sub_id.clone(),
+                    command: command.clone(),
+                    cwd: cwd.clone(),
+                    parsed_cmd: parsed_cmd.clone(),
+                    source: ExecCommandSource::UserShell,
+                    interaction_input: None,
                 }),
             )
             .await;
 
         let exec_env = ExecEnv {
-            command: shell_invocation,
-            cwd: turn_context.cwd.clone(),
+            command: command.clone(),
+            cwd: cwd.clone(),
             env: create_env(&turn_context.shell_environment_policy),
             timeout_ms: None,
             sandbox: SandboxType::None,
@@ -127,6 +131,12 @@ impl SessionTask for UserShellCommandTask {
                         turn_context.as_ref(),
                         EventMsg::ExecCommandEnd(ExecCommandEndEvent {
                             call_id,
+                            turn_id: turn_context.sub_id.clone(),
+                            command: command.clone(),
+                            cwd: cwd.clone(),
+                            parsed_cmd: parsed_cmd.clone(),
+                            source: ExecCommandSource::UserShell,
+                            interaction_input: None,
                             stdout: String::new(),
                             stderr: aborted_message.clone(),
                             aggregated_output: aborted_message.clone(),
@@ -143,6 +153,12 @@ impl SessionTask for UserShellCommandTask {
                         turn_context.as_ref(),
                         EventMsg::ExecCommandEnd(ExecCommandEndEvent {
                             call_id: call_id.clone(),
+                            turn_id: turn_context.sub_id.clone(),
+                            command: command.clone(),
+                            cwd: cwd.clone(),
+                            parsed_cmd: parsed_cmd.clone(),
+                            source: ExecCommandSource::UserShell,
+                            interaction_input: None,
                             stdout: output.stdout.text.clone(),
                             stderr: output.stderr.text.clone(),
                             aggregated_output: output.aggregated_output.text.clone(),
@@ -174,6 +190,12 @@ impl SessionTask for UserShellCommandTask {
                         turn_context.as_ref(),
                         EventMsg::ExecCommandEnd(ExecCommandEndEvent {
                             call_id,
+                            turn_id: turn_context.sub_id.clone(),
+                            command,
+                            cwd,
+                            parsed_cmd,
+                            source: ExecCommandSource::UserShell,
+                            interaction_input: None,
                             stdout: exec_output.stdout.text.clone(),
                             stderr: exec_output.stderr.text.clone(),
                             aggregated_output: exec_output.aggregated_output.text.clone(),

codex-rs/core/src/tools/context.rs

@@ -4,17 +4,13 @@ use crate::tools::TELEMETRY_PREVIEW_MAX_BYTES;
 use crate::tools::TELEMETRY_PREVIEW_MAX_LINES;
 use crate::tools::TELEMETRY_PREVIEW_TRUNCATION_NOTICE;
 use crate::turn_diff_tracker::TurnDiffTracker;
-use codex_otel::otel_event_manager::OtelEventManager;
 use codex_protocol::models::FunctionCallOutputContentItem;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseInputItem;
 use codex_protocol::models::ShellToolCallParams;
-use codex_protocol::protocol::FileChange;
 use codex_utils_string::take_bytes_at_char_boundary;
 use mcp_types::CallToolResult;
 use std::borrow::Cow;
-use std::collections::HashMap;
-use std::path::PathBuf;
 use std::sync::Arc;
 use tokio::sync::Mutex;
 
@@ -244,25 +240,3 @@ mod tests {
         assert_eq!(lines.last(), Some(&TELEMETRY_PREVIEW_TRUNCATION_NOTICE));
     }
 }
-
-#[derive(Clone, Debug)]
-#[allow(dead_code)]
-pub(crate) struct ExecCommandContext {
-    pub(crate) turn: Arc<TurnContext>,
-    pub(crate) call_id: String,
-    pub(crate) command_for_display: Vec<String>,
-    pub(crate) cwd: PathBuf,
-    pub(crate) apply_patch: Option<ApplyPatchCommandContext>,
-    pub(crate) tool_name: String,
-    pub(crate) otel_event_manager: OtelEventManager,
-    // TODO(abhisek-oai): Find a better way to track this.
-    // https://github.com/openai/codex/pull/2471/files#r2470352242
-    pub(crate) is_user_shell_command: bool,
-}
-
-#[derive(Clone, Debug)]
-#[allow(dead_code)]
-pub(crate) struct ApplyPatchCommandContext {
-    pub(crate) user_explicitly_approved_this_action: bool,
-    pub(crate) changes: HashMap<PathBuf, FileChange>,
-}

codex-rs/core/src/tools/events.rs

@@ -8,12 +8,14 @@ use crate::parse_command::parse_command;
 use crate::protocol::EventMsg;
 use crate::protocol::ExecCommandBeginEvent;
 use crate::protocol::ExecCommandEndEvent;
+use crate::protocol::ExecCommandSource;
 use crate::protocol::FileChange;
 use crate::protocol::PatchApplyBeginEvent;
 use crate::protocol::PatchApplyEndEvent;
 use crate::protocol::TurnDiffEvent;
 use crate::tools::context::SharedTurnDiffTracker;
 use crate::tools::sandboxing::ToolError;
+use codex_protocol::parse_command::ParsedCommand;
 use std::collections::HashMap;
 use std::path::Path;
 use std::path::PathBuf;
@@ -60,17 +62,21 @@ pub(crate) async fn emit_exec_command_begin(
     ctx: ToolEventCtx<'_>,
     command: &[String],
     cwd: &Path,
-    is_user_shell_command: bool,
+    parsed_cmd: &[ParsedCommand],
+    source: ExecCommandSource,
+    interaction_input: Option<String>,
 ) {
     ctx.session
         .send_event(
             ctx.turn,
             EventMsg::ExecCommandBegin(ExecCommandBeginEvent {
                 call_id: ctx.call_id.to_string(),
+                turn_id: ctx.turn.sub_id.clone(),
                 command: command.to_vec(),
                 cwd: cwd.to_path_buf(),
-                parsed_cmd: parse_command(command),
-                is_user_shell_command,
+                parsed_cmd: parsed_cmd.to_vec(),
+                source,
+                interaction_input,
             }),
         )
         .await;
@@ -80,7 +86,8 @@ pub(crate) enum ToolEmitter {
     Shell {
         command: Vec<String>,
         cwd: PathBuf,
-        is_user_shell_command: bool,
+        source: ExecCommandSource,
+        parsed_cmd: Vec<ParsedCommand>,
     },
     ApplyPatch {
         changes: HashMap<PathBuf, FileChange>,
@@ -89,18 +96,20 @@ pub(crate) enum ToolEmitter {
     UnifiedExec {
         command: Vec<String>,
         cwd: PathBuf,
-        // True for `exec_command` and false for `write_stdin`.
-        #[allow(dead_code)]
-        is_startup_command: bool,
+        source: ExecCommandSource,
+        interaction_input: Option<String>,
+        parsed_cmd: Vec<ParsedCommand>,
     },
 }
 
 impl ToolEmitter {
-    pub fn shell(command: Vec<String>, cwd: PathBuf, is_user_shell_command: bool) -> Self {
+    pub fn shell(command: Vec<String>, cwd: PathBuf, source: ExecCommandSource) -> Self {
+        let parsed_cmd = parse_command(&command);
         Self::Shell {
             command,
             cwd,
-            is_user_shell_command,
+            source,
+            parsed_cmd,
         }
     }
 
@@ -111,11 +120,19 @@ impl ToolEmitter {
         }
     }
 
-    pub fn unified_exec(command: &[String], cwd: PathBuf, is_startup_command: bool) -> Self {
+    pub fn unified_exec(
+        command: &[String],
+        cwd: PathBuf,
+        source: ExecCommandSource,
+        interaction_input: Option<String>,
+    ) -> Self {
+        let parsed_cmd = parse_command(command);
         Self::UnifiedExec {
             command: command.to_vec(),
             cwd,
-            is_startup_command,
+            source,
+            interaction_input,
+            parsed_cmd,
         }
     }
 
@@ -125,45 +142,15 @@ impl ToolEmitter {
                 Self::Shell {
                     command,
                     cwd,
-                    is_user_shell_command,
+                    source,
+                    parsed_cmd,
                 },
-                ToolEventStage::Begin,
+                stage,
             ) => {
-                emit_exec_command_begin(ctx, command, cwd.as_path(), *is_user_shell_command).await;
-            }
-            (Self::Shell { .. }, ToolEventStage::Success(output)) => {
-                emit_exec_end(
+                emit_exec_stage(
                     ctx,
-                    output.stdout.text.clone(),
-                    output.stderr.text.clone(),
-                    output.aggregated_output.text.clone(),
-                    output.exit_code,
-                    output.duration,
-                    format_exec_output_str(&output),
-                )
-                .await;
-            }
-            (Self::Shell { .. }, ToolEventStage::Failure(ToolEventFailure::Output(output))) => {
-                emit_exec_end(
-                    ctx,
-                    output.stdout.text.clone(),
-                    output.stderr.text.clone(),
-                    output.aggregated_output.text.clone(),
-                    output.exit_code,
-                    output.duration,
-                    format_exec_output_str(&output),
-                )
-                .await;
-            }
-            (Self::Shell { .. }, ToolEventStage::Failure(ToolEventFailure::Message(message))) => {
-                emit_exec_end(
-                    ctx,
-                    String::new(),
-                    (*message).to_string(),
-                    (*message).to_string(),
-                    -1,
-                    Duration::ZERO,
-                    message.clone(),
+                    ExecCommandInput::new(command, cwd.as_path(), parsed_cmd, *source, None),
+                    stage,
                 )
                 .await;
             }
@@ -217,48 +204,26 @@ impl ToolEmitter {
             ) => {
                 emit_patch_end(ctx, String::new(), (*message).to_string(), false).await;
             }
-            (Self::UnifiedExec { command, cwd, .. }, ToolEventStage::Begin) => {
-                emit_exec_command_begin(ctx, command, cwd.as_path(), false).await;
-            }
-            (Self::UnifiedExec { .. }, ToolEventStage::Success(output)) => {
-                emit_exec_end(
-                    ctx,
-                    output.stdout.text.clone(),
-                    output.stderr.text.clone(),
-                    output.aggregated_output.text.clone(),
-                    output.exit_code,
-                    output.duration,
-                    format_exec_output_str(&output),
-                )
-                .await;
-            }
             (
-                Self::UnifiedExec { .. },
-                ToolEventStage::Failure(ToolEventFailure::Output(output)),
+                Self::UnifiedExec {
+                    command,
+                    cwd,
+                    source,
+                    interaction_input,
+                    parsed_cmd,
+                },
+                stage,
             ) => {
-                emit_exec_end(
+                emit_exec_stage(
                     ctx,
-                    output.stdout.text.clone(),
-                    output.stderr.text.clone(),
-                    output.aggregated_output.text.clone(),
-                    output.exit_code,
-                    output.duration,
-                    format_exec_output_str(&output),
-                )
-                .await;
-            }
-            (
-                Self::UnifiedExec { .. },
-                ToolEventStage::Failure(ToolEventFailure::Message(message)),
-            ) => {
-                emit_exec_end(
-                    ctx,
-                    String::new(),
-                    (*message).to_string(),
-                    (*message).to_string(),
-                    -1,
-                    Duration::ZERO,
-                    message.clone(),
+                    ExecCommandInput::new(
+                        command,
+                        cwd.as_path(),
+                        parsed_cmd,
+                        *source,
+                        interaction_input.as_deref(),
+                    ),
+                    stage,
                 )
                 .await;
             }
@@ -317,26 +282,107 @@ impl ToolEmitter {
     }
 }
 
-async fn emit_exec_end(
-    ctx: ToolEventCtx<'_>,
+struct ExecCommandInput<'a> {
+    command: &'a [String],
+    cwd: &'a Path,
+    parsed_cmd: &'a [ParsedCommand],
+    source: ExecCommandSource,
+    interaction_input: Option<&'a str>,
+}
+
+impl<'a> ExecCommandInput<'a> {
+    fn new(
+        command: &'a [String],
+        cwd: &'a Path,
+        parsed_cmd: &'a [ParsedCommand],
+        source: ExecCommandSource,
+        interaction_input: Option<&'a str>,
+    ) -> Self {
+        Self {
+            command,
+            cwd,
+            parsed_cmd,
+            source,
+            interaction_input,
+        }
+    }
+}
+
+struct ExecCommandResult {
     stdout: String,
     stderr: String,
     aggregated_output: String,
     exit_code: i32,
     duration: Duration,
     formatted_output: String,
+}
+
+async fn emit_exec_stage(
+    ctx: ToolEventCtx<'_>,
+    exec_input: ExecCommandInput<'_>,
+    stage: ToolEventStage,
+) {
+    match stage {
+        ToolEventStage::Begin => {
+            emit_exec_command_begin(
+                ctx,
+                exec_input.command,
+                exec_input.cwd,
+                exec_input.parsed_cmd,
+                exec_input.source,
+                exec_input.interaction_input.map(str::to_owned),
+            )
+            .await;
+        }
+        ToolEventStage::Success(output)
+        | ToolEventStage::Failure(ToolEventFailure::Output(output)) => {
+            let exec_result = ExecCommandResult {
+                stdout: output.stdout.text.clone(),
+                stderr: output.stderr.text.clone(),
+                aggregated_output: output.aggregated_output.text.clone(),
+                exit_code: output.exit_code,
+                duration: output.duration,
+                formatted_output: format_exec_output_str(&output),
+            };
+            emit_exec_end(ctx, exec_input, exec_result).await;
+        }
+        ToolEventStage::Failure(ToolEventFailure::Message(message)) => {
+            let text = message.to_string();
+            let exec_result = ExecCommandResult {
+                stdout: String::new(),
+                stderr: text.clone(),
+                aggregated_output: text.clone(),
+                exit_code: -1,
+                duration: Duration::ZERO,
+                formatted_output: text,
+            };
+            emit_exec_end(ctx, exec_input, exec_result).await;
+        }
+    }
+}
+
+async fn emit_exec_end(
+    ctx: ToolEventCtx<'_>,
+    exec_input: ExecCommandInput<'_>,
+    exec_result: ExecCommandResult,
 ) {
     ctx.session
         .send_event(
             ctx.turn,
             EventMsg::ExecCommandEnd(ExecCommandEndEvent {
                 call_id: ctx.call_id.to_string(),
-                stdout,
-                stderr,
-                aggregated_output,
-                exit_code,
-                duration,
-                formatted_output,
+                turn_id: ctx.turn.sub_id.clone(),
+                command: exec_input.command.to_vec(),
+                cwd: exec_input.cwd.to_path_buf(),
+                parsed_cmd: exec_input.parsed_cmd.to_vec(),
+                source: exec_input.source,
+                interaction_input: exec_input.interaction_input.map(str::to_owned),
+                stdout: exec_result.stdout,
+                stderr: exec_result.stderr,
+                aggregated_output: exec_result.aggregated_output,
+                exit_code: exec_result.exit_code,
+                duration: exec_result.duration,
+                formatted_output: exec_result.formatted_output,
             }),
         )
         .await;

codex-rs/core/src/tools/handlers/mcp_resource.rs

@@ -287,6 +287,8 @@ async fn handle_list_resources(
             let resources = session
                 .services
                 .mcp_connection_manager
+                .read()
+                .await
                 .list_all_resources()
                 .await;
             Ok(ListResourcesPayload::from_all_servers(resources))
@@ -396,6 +398,8 @@ async fn handle_list_resource_templates(
             let templates = session
                 .services
                 .mcp_connection_manager
+                .read()
+                .await
                 .list_all_resource_templates()
                 .await;
             Ok(ListResourceTemplatesPayload::from_all_servers(templates))

codex-rs/core/src/tools/handlers/shell.rs

@@ -11,6 +11,7 @@ use crate::exec::ExecParams;
 use crate::exec_env::create_env;
 use crate::function_tool::FunctionCallError;
 use crate::is_safe_command::is_known_safe_command;
+use crate::protocol::ExecCommandSource;
 use crate::tools::context::ToolInvocation;
 use crate::tools::context::ToolOutput;
 use crate::tools::context::ToolPayload;
@@ -116,7 +117,6 @@ impl ToolHandler for ShellHandler {
                     turn,
                     tracker,
                     call_id,
-                    false,
                 )
                 .await
             }
@@ -129,7 +129,6 @@ impl ToolHandler for ShellHandler {
                     turn,
                     tracker,
                     call_id,
-                    true,
                 )
                 .await
             }
@@ -177,7 +176,6 @@ impl ToolHandler for ShellCommandHandler {
             turn,
             tracker,
             call_id,
-            false,
         )
         .await
     }
@@ -191,7 +189,6 @@ impl ShellHandler {
         turn: Arc<TurnContext>,
         tracker: crate::tools::context::SharedTurnDiffTracker,
         call_id: String,
-        is_user_shell_command: bool,
     ) -> Result<ToolOutput, FunctionCallError> {
         // Approval policy guard for explicit escalation in non-OnRequest modes.
         if exec_params.with_escalated_permissions.unwrap_or(false)
@@ -284,12 +281,9 @@ impl ShellHandler {
             }
         }
 
-        // Regular shell execution path.
-        let emitter = ToolEmitter::shell(
-            exec_params.command.clone(),
-            exec_params.cwd.clone(),
-            is_user_shell_command,
-        );
+        let source = ExecCommandSource::Agent;
+        let emitter =
+            ToolEmitter::shell(exec_params.command.clone(), exec_params.cwd.clone(), source);
         let event_ctx = ToolEventCtx::new(session.as_ref(), turn.as_ref(), &call_id, None);
         emitter.begin(event_ctx).await;
 
@@ -300,11 +294,6 @@ impl ShellHandler {
             env: exec_params.env.clone(),
             with_escalated_permissions: exec_params.with_escalated_permissions,
             justification: exec_params.justification.clone(),
-            exec_policy: if is_user_shell_command {
-                None
-            } else {
-                turn.exec_policy_v2.clone()
-            },
         };
         let mut orchestrator = ToolOrchestrator::new();
         let mut runtime = ShellRuntime::new();

codex-rs/core/src/tools/handlers/unified_exec.rs

@@ -4,6 +4,7 @@ use crate::function_tool::FunctionCallError;
 use crate::is_safe_command::is_known_safe_command;
 use crate::protocol::EventMsg;
 use crate::protocol::ExecCommandOutputDeltaEvent;
+use crate::protocol::ExecCommandSource;
 use crate::protocol::ExecOutputStream;
 use crate::shell::get_shell_by_model_provided_path;
 use crate::tools::context::ToolInvocation;
@@ -162,8 +163,12 @@ impl ToolHandler for UnifiedExecHandler {
                     &context.call_id,
                     None,
                 );
-
-                let emitter = ToolEmitter::unified_exec(&command, cwd.clone(), true);
+                let emitter = ToolEmitter::unified_exec(
+                    &command,
+                    cwd.clone(),
+                    ExecCommandSource::UnifiedExecStartup,
+                    None,
+                );
                 emitter.emit(event_ctx, ToolEventStage::Begin).await;
 
                 manager
@@ -191,6 +196,7 @@ impl ToolHandler for UnifiedExecHandler {
                 })?;
                 manager
                     .write_stdin(WriteStdinRequest {
+                        call_id: &call_id,
                         session_id: args.session_id,
                         input: &args.chars,
                         yield_time_ms: args.yield_time_ms,

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

@@ -9,9 +9,7 @@ pub mod runtimes;
 pub mod sandboxing;
 pub mod spec;
 
-use crate::context_manager::MODEL_FORMAT_MAX_BYTES;
-use crate::context_manager::MODEL_FORMAT_MAX_LINES;
-use crate::context_manager::format_output_for_model_body;
+use crate::context_manager::truncate_with_line_bytes_budget;
 use crate::exec::ExecToolCallOutput;
 pub use router::ToolRouter;
 use serde::Serialize;
@@ -22,6 +20,9 @@ pub(crate) const TELEMETRY_PREVIEW_MAX_LINES: usize = 64; // lines
 pub(crate) const TELEMETRY_PREVIEW_TRUNCATION_NOTICE: &str =
     "[... telemetry preview truncated ...]";
 
+// TODO(aibrahim): migrate shell tool to use truncate text and respect config value
+const SHELL_OUTPUT_MAX_BYTES: usize = 10_000;
+
 /// Format the combined exec output for sending back to the model.
 /// Includes exit code and duration metadata; truncates large bodies safely.
 pub fn format_exec_output_for_model(exec_output: &ExecToolCallOutput) -> String {
@@ -77,5 +78,5 @@ pub fn format_exec_output_str(exec_output: &ExecToolCallOutput) -> String {
     };
 
     // Truncate for model consumption before serialization.
-    format_output_for_model_body(&body, MODEL_FORMAT_MAX_BYTES, MODEL_FORMAT_MAX_LINES)
+    truncate_with_line_bytes_budget(&body, SHELL_OUTPUT_MAX_BYTES)
 }

codex-rs/core/src/tools/orchestrator.rs

@@ -11,7 +11,6 @@ use crate::error::get_error_message_ui;
 use crate::exec::ExecToolCallOutput;
 use crate::sandboxing::SandboxManager;
 use crate::tools::sandboxing::ApprovalCtx;
-use crate::tools::sandboxing::ApprovalRequirement;
 use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::ToolCtx;
@@ -50,49 +49,40 @@ impl ToolOrchestrator {
         let otel_cfg = codex_otel::otel_event_manager::ToolDecisionSource::Config;
 
         // 1) Approval
+        let needs_initial_approval =
+            tool.wants_initial_approval(req, approval_policy, &turn_ctx.sandbox_policy);
         let mut already_approved = false;
 
-        match tool.approval_requirement(req, approval_policy, &turn_ctx.sandbox_policy) {
-            ApprovalRequirement::Skip => {
-                otel.tool_decision(otel_tn, otel_ci, ReviewDecision::Approved, otel_cfg);
-            }
-            ApprovalRequirement::Forbidden { reason } => {
-                return Err(ToolError::Rejected(reason));
-            }
-            ApprovalRequirement::NeedsApproval { reason } => {
-                let mut risk = None;
+        if needs_initial_approval {
+            let mut risk = None;
 
-                if let Some(metadata) = req.sandbox_retry_data() {
-                    risk = tool_ctx
-                        .session
-                        .assess_sandbox_command(
-                            turn_ctx,
-                            &tool_ctx.call_id,
-                            &metadata.command,
-                            None,
-                        )
-                        .await;
+            if let Some(metadata) = req.sandbox_retry_data() {
+                risk = tool_ctx
+                    .session
+                    .assess_sandbox_command(turn_ctx, &tool_ctx.call_id, &metadata.command, None)
+                    .await;
+            }
+
+            let approval_ctx = ApprovalCtx {
+                session: tool_ctx.session,
+                turn: turn_ctx,
+                call_id: &tool_ctx.call_id,
+                retry_reason: None,
+                risk,
+            };
+            let decision = tool.start_approval_async(req, approval_ctx).await;
+
+            otel.tool_decision(otel_tn, otel_ci, decision, otel_user.clone());
+
+            match decision {
+                ReviewDecision::Denied | ReviewDecision::Abort => {
+                    return Err(ToolError::Rejected("rejected by user".to_string()));
                 }
-
-                let approval_ctx = ApprovalCtx {
-                    session: tool_ctx.session,
-                    turn: turn_ctx,
-                    call_id: &tool_ctx.call_id,
-                    retry_reason: reason,
-                    risk,
-                };
-                let decision = tool.start_approval_async(req, approval_ctx).await;
-
-                otel.tool_decision(otel_tn, otel_ci, decision, otel_user.clone());
-
-                match decision {
-                    ReviewDecision::Denied | ReviewDecision::Abort => {
-                        return Err(ToolError::Rejected("rejected by user".to_string()));
-                    }
-                    ReviewDecision::Approved | ReviewDecision::ApprovedForSession => {}
-                }
-                already_approved = true;
+                ReviewDecision::Approved | ReviewDecision::ApprovedForSession => {}
             }
+            already_approved = true;
+        } else {
+            otel.tool_decision(otel_tn, otel_ci, ReviewDecision::Approved, otel_cfg);
         }
 
         // 2) First attempt under the selected sandbox.

codex-rs/core/src/tools/parallel.rs

@@ -112,7 +112,7 @@ impl ToolCallRuntime {
 
     fn abort_message(call: &ToolCall, secs: f32) -> String {
         match call.tool_name.as_str() {
-            "shell" | "container.exec" | "local_shell" | "unified_exec" => {
+            "shell" | "container.exec" | "local_shell" | "shell_command" | "unified_exec" => {
                 format!("Wall time: {secs:.1} seconds\naborted by user")
             }
             _ => format!("aborted by user after {secs:.1}s"),

codex-rs/core/src/tools/router.rs

@@ -54,7 +54,7 @@ impl ToolRouter {
             .any(|config| config.spec.name() == tool_name)
     }
 
-    pub fn build_tool_call(
+    pub async fn build_tool_call(
         session: &Session,
         item: ResponseItem,
     ) -> Result<Option<ToolCall>, FunctionCallError> {
@@ -65,7 +65,7 @@ impl ToolRouter {
                 call_id,
                 ..
             } => {
-                if let Some((server, tool)) = session.parse_mcp_tool_name(&name) {
+                if let Some((server, tool)) = session.parse_mcp_tool_name(&name).await {
                     Ok(Some(ToolCall {
                         tool_name: name,
                         call_id,

codex-rs/core/src/tools/runtimes/shell.rs

@@ -6,13 +6,11 @@ builds a CommandSpec, and runs it under the current SandboxAttempt.
 */
 use crate::command_safety::is_dangerous_command::requires_initial_appoval;
 use crate::exec::ExecToolCallOutput;
-use crate::exec_policy::evaluate_with_policy;
 use crate::protocol::SandboxPolicy;
 use crate::sandboxing::execute_env;
 use crate::tools::runtimes::build_command_spec;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
-use crate::tools::sandboxing::ApprovalRequirement;
 use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::SandboxRetryData;
@@ -22,12 +20,10 @@ use crate::tools::sandboxing::ToolCtx;
 use crate::tools::sandboxing::ToolError;
 use crate::tools::sandboxing::ToolRuntime;
 use crate::tools::sandboxing::with_cached_approval;
-use codex_execpolicy2::Policy as ExecPolicyV2;
 use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::ReviewDecision;
 use futures::future::BoxFuture;
 use std::path::PathBuf;
-use std::sync::Arc;
 
 #[derive(Clone, Debug)]
 pub struct ShellRequest {
@@ -37,7 +33,6 @@ pub struct ShellRequest {
     pub env: std::collections::HashMap<String, String>,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
-    pub exec_policy: Option<Arc<ExecPolicyV2>>,
 }
 
 impl ProvidesSandboxRetryData for ShellRequest {
@@ -119,26 +114,18 @@ impl Approvable<ShellRequest> for ShellRuntime {
         })
     }
 
-    fn approval_requirement(
+    fn wants_initial_approval(
         &self,
         req: &ShellRequest,
         policy: AskForApproval,
         sandbox_policy: &SandboxPolicy,
-    ) -> ApprovalRequirement {
-        if let Some(exec_policy) = &req.exec_policy
-            && let Some(requirement) = evaluate_with_policy(exec_policy, &req.command, policy)
-        {
-            requirement
-        } else if requires_initial_appoval(
+    ) -> bool {
+        requires_initial_appoval(
             policy,
             sandbox_policy,
             &req.command,
             req.with_escalated_permissions.unwrap_or(false),
-        ) {
-            ApprovalRequirement::NeedsApproval { reason: None }
-        } else {
-            ApprovalRequirement::Skip
-        }
+        )
     }
 
     fn wants_escalated_first_attempt(&self, req: &ShellRequest) -> bool {
@@ -170,85 +157,3 @@ impl ToolRuntime<ShellRequest, ExecToolCallOutput> for ShellRuntime {
         Ok(out)
     }
 }
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use codex_execpolicy2::PolicyParser;
-    use pretty_assertions::assert_eq;
-    use std::collections::HashMap;
-
-    fn parse_policy(src: &str) -> Arc<ExecPolicyV2> {
-        let mut parser = PolicyParser::new();
-        parser
-            .parse("test.codexpolicy", src)
-            .expect("parse execpolicy2 file");
-        Arc::new(parser.build())
-    }
-
-    fn shell_request(command: &[&str], exec_policy: Option<Arc<ExecPolicyV2>>) -> ShellRequest {
-        ShellRequest {
-            command: command.iter().map(ToString::to_string).collect(),
-            cwd: PathBuf::from("."),
-            timeout_ms: None,
-            env: HashMap::new(),
-            with_escalated_permissions: None,
-            justification: None,
-            exec_policy,
-        }
-    }
-
-    #[test]
-    fn prompt_decision_requires_approval() {
-        let policy = parse_policy(r#"prefix_rule(pattern=["echo"], decision="prompt")"#);
-        let req = shell_request(&["echo", "hi"], Some(policy));
-        let runtime = ShellRuntime::new();
-
-        let requirement = runtime.approval_requirement(
-            &req,
-            AskForApproval::OnRequest,
-            &SandboxPolicy::DangerFullAccess,
-        );
-
-        assert_eq!(
-            requirement,
-            ApprovalRequirement::NeedsApproval {
-                reason: Some("execpolicy requires approval for this command".to_string())
-            }
-        );
-    }
-
-    #[test]
-    fn prompt_blocked_when_approval_disabled() {
-        let policy = parse_policy(r#"prefix_rule(pattern=["echo"], decision="prompt")"#);
-        let req = shell_request(&["echo", "hi"], Some(policy));
-        let runtime = ShellRuntime::new();
-
-        let requirement = runtime.approval_requirement(
-            &req,
-            AskForApproval::Never,
-            &SandboxPolicy::DangerFullAccess,
-        );
-
-        assert_eq!(
-            requirement,
-            ApprovalRequirement::Forbidden {
-                reason: "execpolicy requires approval for this command".to_string()
-            }
-        );
-    }
-
-    #[test]
-    fn user_shell_commands_skip_execpolicy() {
-        let req = shell_request(&["echo", "hi"], None);
-        let runtime = ShellRuntime::new();
-
-        let requirement = runtime.approval_requirement(
-            &req,
-            AskForApproval::OnRequest,
-            &SandboxPolicy::DangerFullAccess,
-        );
-
-        assert_eq!(requirement, ApprovalRequirement::Skip);
-    }
-}

codex-rs/core/src/tools/runtimes/unified_exec.rs

@@ -1,5 +1,4 @@
 use crate::command_safety::is_dangerous_command::requires_initial_appoval;
-use crate::exec_policy::evaluate_with_policy;
 /*
 Runtime: unified exec
 
@@ -11,7 +10,6 @@ use crate::error::SandboxErr;
 use crate::tools::runtimes::build_command_spec;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
-use crate::tools::sandboxing::ApprovalRequirement;
 use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::SandboxRetryData;
@@ -24,21 +22,18 @@ use crate::tools::sandboxing::with_cached_approval;
 use crate::unified_exec::UnifiedExecError;
 use crate::unified_exec::UnifiedExecSession;
 use crate::unified_exec::UnifiedExecSessionManager;
-use codex_execpolicy2::Policy as ExecPolicyV2;
 use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::ReviewDecision;
 use codex_protocol::protocol::SandboxPolicy;
 use futures::future::BoxFuture;
 use std::collections::HashMap;
 use std::path::PathBuf;
-use std::sync::Arc;
 
 #[derive(Clone, Debug)]
 pub struct UnifiedExecRequest {
     pub command: Vec<String>,
     pub cwd: PathBuf,
     pub env: HashMap<String, String>,
-    pub exec_policy: Option<Arc<ExecPolicyV2>>,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
 }
@@ -68,7 +63,6 @@ impl UnifiedExecRequest {
         command: Vec<String>,
         cwd: PathBuf,
         env: HashMap<String, String>,
-        exec_policy: Option<Arc<ExecPolicyV2>>,
         with_escalated_permissions: Option<bool>,
         justification: Option<String>,
     ) -> Self {
@@ -76,7 +70,6 @@ impl UnifiedExecRequest {
             command,
             cwd,
             env,
-            exec_policy,
             with_escalated_permissions,
             justification,
         }
@@ -136,26 +129,18 @@ impl Approvable<UnifiedExecRequest> for UnifiedExecRuntime<'_> {
         })
     }
 
-    fn approval_requirement(
+    fn wants_initial_approval(
         &self,
         req: &UnifiedExecRequest,
         policy: AskForApproval,
         sandbox_policy: &SandboxPolicy,
-    ) -> ApprovalRequirement {
-        if let Some(exec_policy) = &req.exec_policy
-            && let Some(requirement) = evaluate_with_policy(exec_policy, &req.command, policy)
-        {
-            requirement
-        } else if requires_initial_appoval(
+    ) -> bool {
+        requires_initial_appoval(
             policy,
             sandbox_policy,
             &req.command,
             req.with_escalated_permissions.unwrap_or(false),
-        ) {
-            ApprovalRequirement::NeedsApproval { reason: None }
-        } else {
-            ApprovalRequirement::Skip
-        }
+        )
     }
 
     fn wants_escalated_first_attempt(&self, req: &UnifiedExecRequest) -> bool {

codex-rs/core/src/tools/sandboxing.rs

@@ -86,13 +86,6 @@ pub(crate) struct ApprovalCtx<'a> {
     pub risk: Option<SandboxCommandAssessment>,
 }
 
-#[derive(Clone, Debug, PartialEq, Eq)]
-pub(crate) enum ApprovalRequirement {
-    Skip,
-    NeedsApproval { reason: Option<String> },
-    Forbidden { reason: String },
-}
-
 pub(crate) trait Approvable<Req> {
     type ApprovalKey: Hash + Eq + Clone + Debug + Serialize;
 
@@ -113,27 +106,21 @@ pub(crate) trait Approvable<Req> {
         matches!(policy, AskForApproval::Never)
     }
 
-    /// Decide whether an initial user approval should be requested or forbidden before the
+    /// Decide whether an initial user approval should be requested before the
     /// first attempt. Defaults to the orchestrator's behavior (pre‑refactor):
     /// - Never, OnFailure: do not ask
     /// - OnRequest: ask unless sandbox policy is DangerFullAccess
     /// - UnlessTrusted: always ask
-    fn approval_requirement(
+    fn wants_initial_approval(
         &self,
         _req: &Req,
         policy: AskForApproval,
         sandbox_policy: &SandboxPolicy,
-    ) -> ApprovalRequirement {
-        let needs_approval = match policy {
+    ) -> bool {
+        match policy {
             AskForApproval::Never | AskForApproval::OnFailure => false,
             AskForApproval::OnRequest => !matches!(sandbox_policy, SandboxPolicy::DangerFullAccess),
             AskForApproval::UnlessTrusted => true,
-        };
-
-        if needs_approval {
-            ApprovalRequirement::NeedsApproval { reason: None }
-        } else {
-            ApprovalRequirement::Skip
         }
     }
 

codex-rs/core/src/tools/spec.rs

@@ -20,6 +20,11 @@ pub enum ConfigShellToolType {
     Default,
     Local,
     UnifiedExec,
+    /// Do not include a shell tool by default. Useful when using Codex
+    /// with tools provided exclusively provided by MCP servers. Often used
+    /// with `--config base_instructions=CUSTOM_INSTRUCTIONS`
+    /// to customize agent behavior.
+    Disabled,
     /// Takes a command as a single string to be run in the user's default shell.
     ShellCommand,
 }
@@ -48,7 +53,9 @@ impl ToolsConfig {
         let include_web_search_request = features.enabled(Feature::WebSearchRequest);
         let include_view_image_tool = features.enabled(Feature::ViewImageTool);
 
-        let shell_type = if features.enabled(Feature::UnifiedExec) {
+        let shell_type = if !features.enabled(Feature::ShellTool) {
+            ConfigShellToolType::Disabled
+        } else if features.enabled(Feature::UnifiedExec) {
             ConfigShellToolType::UnifiedExec
         } else if features.enabled(Feature::ShellCommandTool) {
             ConfigShellToolType::ShellCommand
@@ -294,9 +301,26 @@ fn create_shell_tool() -> ToolSpec {
         },
     );
 
+    let description  = if cfg!(windows) {
+        r#"Runs a Powershell command (Windows) and returns its output. Arguments to `shell` will be passed to CreateProcessW(). Most commands should be prefixed with ["powershell.exe", "-Command"].
+        
+Examples of valid command strings:
+
+- ls -a (show hidden): ["powershell.exe", "-Command", "Get-ChildItem -Force"]
+- recursive find by name: ["powershell.exe", "-Command", "Get-ChildItem -Recurse -Filter *.py"]
+- recursive grep: ["powershell.exe", "-Command", "Get-ChildItem -Path C:\\myrepo -Recurse | Select-String -Pattern 'TODO' -CaseSensitive"]
+- ps aux | grep python: ["powershell.exe", "-Command", "Get-Process | Where-Object { $_.ProcessName -like '*python*' }"]
+- setting an env var: ["powershell.exe", "-Command", "$env:FOO='bar'; echo $env:FOO"]
+- running an inline Python script: ["powershell.exe", "-Command", "@'\\nprint('Hello, world!')\\n'@ | python -"]"#
+    } else {
+        r#"Runs a shell command and returns its output.
+- The arguments to `shell` will be passed to execvp(). Most terminal commands should be prefixed with ["bash", "-lc"].
+- Always set the `workdir` param when using the shell function. Do not use `cd` unless absolutely necessary."#
+    }.to_string();
+
     ToolSpec::Function(ResponsesApiTool {
         name: "shell".to_string(),
-        description: "Runs a shell command and returns its output.".to_string(),
+        description,
         strict: false,
         parameters: JsonSchema::Object {
             properties,
@@ -341,9 +365,25 @@ fn create_shell_command_tool() -> ToolSpec {
         },
     );
 
+    let description = if cfg!(windows) {
+        r#"Runs a Powershell command (Windows) and returns its output.
+        
+Examples of valid command strings:
+
+- ls -a (show hidden): "Get-ChildItem -Force"
+- recursive find by name: "Get-ChildItem -Recurse -Filter *.py"
+- recursive grep: "Get-ChildItem -Path C:\\myrepo -Recurse | Select-String -Pattern 'TODO' -CaseSensitive"
+- ps aux | grep python: "Get-Process | Where-Object { $_.ProcessName -like '*python*' }"
+- setting an env var: "$env:FOO='bar'; echo $env:FOO"
+- running an inline Python script: "@'\\nprint('Hello, world!')\\n'@ | python -"#
+    } else {
+        r#"Runs a shell command and returns its output.
+- Always set the `workdir` param when using the shell_command function. Do not use `cd` unless absolutely necessary."#
+    }.to_string();
+
     ToolSpec::Function(ResponsesApiTool {
         name: "shell_command".to_string(),
-        description: "Runs a shell command string and returns its output.".to_string(),
+        description,
         strict: false,
         parameters: JsonSchema::Object {
             properties,
@@ -973,16 +1013,21 @@ pub(crate) fn build_specs(
             builder.register_handler("exec_command", unified_exec_handler.clone());
             builder.register_handler("write_stdin", unified_exec_handler);
         }
+        ConfigShellToolType::Disabled => {
+            // Do nothing.
+        }
         ConfigShellToolType::ShellCommand => {
             builder.push_spec(create_shell_command_tool());
         }
     }
 
-    // Always register shell aliases so older prompts remain compatible.
-    builder.register_handler("shell", shell_handler.clone());
-    builder.register_handler("container.exec", shell_handler.clone());
-    builder.register_handler("local_shell", shell_handler);
-    builder.register_handler("shell_command", shell_command_handler);
+    if config.shell_type != ConfigShellToolType::Disabled {
+        // Always register shell aliases so older prompts remain compatible.
+        builder.register_handler("shell", shell_handler.clone());
+        builder.register_handler("container.exec", shell_handler.clone());
+        builder.register_handler("local_shell", shell_handler);
+        builder.register_handler("shell_command", shell_command_handler);
+    }
 
     builder.push_spec_with_parallel_support(create_list_mcp_resources_tool(), true);
     builder.push_spec_with_parallel_support(create_list_mcp_resource_templates_tool(), true);
@@ -1118,6 +1163,7 @@ mod tests {
             ConfigShellToolType::Default => Some("shell"),
             ConfigShellToolType::Local => Some("local_shell"),
             ConfigShellToolType::UnifiedExec => None,
+            ConfigShellToolType::Disabled => None,
             ConfigShellToolType::ShellCommand => Some("shell_command"),
         }
     }
@@ -1246,7 +1292,7 @@ mod tests {
             "gpt-5-codex",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1263,7 +1309,7 @@ mod tests {
             "gpt-5.1-codex",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1338,7 +1384,7 @@ mod tests {
             "gpt-5.1-codex-mini",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1355,7 +1401,7 @@ mod tests {
             "gpt-5.1",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1873,8 +1919,23 @@ mod tests {
         };
         assert_eq!(name, "shell");
 
-        let expected = "Runs a shell command and returns its output.";
-        assert_eq!(description, expected);
+        let expected = if cfg!(windows) {
+            r#"Runs a Powershell command (Windows) and returns its output. Arguments to `shell` will be passed to CreateProcessW(). Most commands should be prefixed with ["powershell.exe", "-Command"].
+        
+Examples of valid command strings:
+
+- ls -a (show hidden): ["powershell.exe", "-Command", "Get-ChildItem -Force"]
+- recursive find by name: ["powershell.exe", "-Command", "Get-ChildItem -Recurse -Filter *.py"]
+- recursive grep: ["powershell.exe", "-Command", "Get-ChildItem -Path C:\\myrepo -Recurse | Select-String -Pattern 'TODO' -CaseSensitive"]
+- ps aux | grep python: ["powershell.exe", "-Command", "Get-Process | Where-Object { $_.ProcessName -like '*python*' }"]
+- setting an env var: ["powershell.exe", "-Command", "$env:FOO='bar'; echo $env:FOO"]
+- running an inline Python script: ["powershell.exe", "-Command", "@'\\nprint('Hello, world!')\\n'@ | python -"]"#
+        } else {
+            r#"Runs a shell command and returns its output.
+- The arguments to `shell` will be passed to execvp(). Most terminal commands should be prefixed with ["bash", "-lc"].
+- Always set the `workdir` param when using the shell function. Do not use `cd` unless absolutely necessary."#
+        }.to_string();
+        assert_eq!(description, &expected);
     }
 
     #[test]
@@ -1888,8 +1949,22 @@ mod tests {
         };
         assert_eq!(name, "shell_command");
 
-        let expected = "Runs a shell command string and returns its output.";
-        assert_eq!(description, expected);
+        let expected = if cfg!(windows) {
+            r#"Runs a Powershell command (Windows) and returns its output.
+        
+Examples of valid command strings:
+
+- ls -a (show hidden): "Get-ChildItem -Force"
+- recursive find by name: "Get-ChildItem -Recurse -Filter *.py"
+- recursive grep: "Get-ChildItem -Path C:\\myrepo -Recurse | Select-String -Pattern 'TODO' -CaseSensitive"
+- ps aux | grep python: "Get-Process | Where-Object { $_.ProcessName -like '*python*' }"
+- setting an env var: "$env:FOO='bar'; echo $env:FOO"
+- running an inline Python script: "@'\\nprint('Hello, world!')\\n'@ | python -"#.to_string()
+        } else {
+            r#"Runs a shell command and returns its output.
+- Always set the `workdir` param when using the shell_command function. Do not use `cd` unless absolutely necessary."#.to_string()
+        };
+        assert_eq!(description, &expected);
     }
 
     #[test]

codex-rs/core/src/truncate.rs

@@ -1,200 +1,670 @@
 //! Utilities for truncating large chunks of output while preserving a prefix
-//! and suffix on UTF-8 boundaries.
+//! and suffix on UTF-8 boundaries, and helpers for line/token‑based truncation
+//! used across the core crate.
 
-use codex_utils_tokenizer::Tokenizer;
+use codex_protocol::models::FunctionCallOutputContentItem;
+use codex_utils_string::take_bytes_at_char_boundary;
+use codex_utils_string::take_last_bytes_at_char_boundary;
 
-/// Truncate the middle of a UTF-8 string to at most `max_bytes` bytes,
-/// preserving the beginning and the end. Returns the possibly truncated
-/// string and `Some(original_token_count)` (counted with the local tokenizer;
-/// falls back to a 4-bytes-per-token estimate if the tokenizer cannot load)
-/// if truncation occurred; otherwise returns the original string and `None`.
-pub(crate) fn truncate_middle(s: &str, max_bytes: usize) -> (String, Option<u64>) {
-    if s.len() <= max_bytes {
-        return (s.to_string(), None);
+use crate::config::Config;
+
+const APPROX_BYTES_PER_TOKEN: usize = 4;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum TruncationPolicy {
+    Bytes(usize),
+    Tokens(usize),
+}
+
+impl TruncationPolicy {
+    pub fn new(config: &Config) -> Self {
+        let config_token_limit = config.tool_output_token_limit;
+
+        match config.model_family.truncation_policy {
+            TruncationPolicy::Bytes(family_bytes) => {
+                if let Some(token_limit) = config_token_limit {
+                    Self::Bytes(approx_bytes_for_tokens(token_limit))
+                } else {
+                    Self::Bytes(family_bytes)
+                }
+            }
+            TruncationPolicy::Tokens(family_tokens) => {
+                if let Some(token_limit) = config_token_limit {
+                    Self::Tokens(token_limit)
+                } else {
+                    Self::Tokens(family_tokens)
+                }
+            }
+        }
     }
 
-    // Build a tokenizer for counting (default to o200k_base; fall back to cl100k_base).
-    // If both fail, fall back to a 4-bytes-per-token estimate.
-    let tok = Tokenizer::try_default().ok();
-    let token_count = |text: &str| -> u64 {
-        if let Some(ref t) = tok {
-            t.count(text) as u64
-        } else {
-            (text.len() as u64).div_ceil(4)
+    /// Returns a token budget derived from this policy.
+    ///
+    /// - For `Tokens`, this is the explicit token limit.
+    /// - For `Bytes`, this is an approximate token budget using the global
+    ///   bytes-per-token heuristic.
+    pub fn token_budget(&self) -> usize {
+        match self {
+            TruncationPolicy::Bytes(bytes) => {
+                usize::try_from(approx_tokens_from_byte_count(*bytes)).unwrap_or(usize::MAX)
+            }
+            TruncationPolicy::Tokens(tokens) => *tokens,
         }
-    };
-
-    let total_tokens = token_count(s);
-    if max_bytes == 0 {
-        return (
-            format!("…{total_tokens} tokens truncated…"),
-            Some(total_tokens),
-        );
     }
 
-    fn truncate_on_boundary(input: &str, max_len: usize) -> &str {
-        if input.len() <= max_len {
-            return input;
+    /// Returns a byte budget derived from this policy.
+    ///
+    /// - For `Bytes`, this is the explicit byte limit.
+    /// - For `Tokens`, this is an approximate byte budget using the global
+    ///   bytes-per-token heuristic.
+    pub fn byte_budget(&self) -> usize {
+        match self {
+            TruncationPolicy::Bytes(bytes) => *bytes,
+            TruncationPolicy::Tokens(tokens) => approx_bytes_for_tokens(*tokens),
         }
-        let mut end = max_len;
-        while end > 0 && !input.is_char_boundary(end) {
-            end -= 1;
-        }
-        &input[..end]
     }
+}
 
-    fn pick_prefix_end(s: &str, left_budget: usize) -> usize {
-        if let Some(head) = s.get(..left_budget)
-            && let Some(i) = head.rfind('\n')
-        {
-            return i + 1;
-        }
-        truncate_on_boundary(s, left_budget).len()
+/// Format a block of exec/tool output for model consumption, truncating by
+/// lines and bytes while preserving head and tail segments.
+pub(crate) fn truncate_with_line_bytes_budget(content: &str, bytes_budget: usize) -> String {
+    // TODO(aibrahim): to be removed
+    let lines_budget = 256;
+    // Head+tail truncation for the model: show the beginning and end with an elision.
+    // Clients still receive full streams; only this formatted summary is capped.
+    let total_lines = content.lines().count();
+    if content.len() <= bytes_budget && total_lines <= lines_budget {
+        return content.to_string();
     }
+    let output = truncate_formatted_exec_output(content, total_lines, bytes_budget, lines_budget);
+    format!("Total output lines: {total_lines}\n\n{output}")
+}
 
-    fn pick_suffix_start(s: &str, right_budget: usize) -> usize {
-        let start_tail = s.len().saturating_sub(right_budget);
-        if let Some(tail) = s.get(start_tail..)
-            && let Some(i) = tail.find('\n')
-        {
-            return start_tail + i + 1;
-        }
-
-        let mut idx = start_tail.min(s.len());
-        while idx < s.len() && !s.is_char_boundary(idx) {
-            idx += 1;
-        }
-        idx
-    }
-
-    // Iterate to stabilize marker length → keep budget → boundaries.
-    let mut guess_tokens: u64 = 1;
-    for _ in 0..4 {
-        let marker = format!("…{guess_tokens} tokens truncated…");
-        let marker_len = marker.len();
-        let keep_budget = max_bytes.saturating_sub(marker_len);
-        if keep_budget == 0 {
-            return (
-                format!("…{total_tokens} tokens truncated…"),
-                Some(total_tokens),
+pub(crate) fn truncate_text(content: &str, policy: TruncationPolicy) -> String {
+    match policy {
+        TruncationPolicy::Bytes(bytes) => truncate_with_byte_estimate(
+            content,
+            bytes,
+            TruncationSource::Policy(TruncationPolicy::Bytes(bytes)),
+        ),
+        TruncationPolicy::Tokens(tokens) => {
+            let (truncated, _) = truncate_with_token_budget(
+                content,
+                tokens,
+                TruncationSource::Policy(TruncationPolicy::Tokens(tokens)),
             );
+            truncated
         }
+    }
+}
+/// Globally truncate function output items to fit within the given
+/// truncation policy's budget, preserving as many text/image items as
+/// possible and appending a summary for any omitted text items.
+pub(crate) fn truncate_function_output_items_with_policy(
+    items: &[FunctionCallOutputContentItem],
+    policy: TruncationPolicy,
+) -> Vec<FunctionCallOutputContentItem> {
+    let mut out: Vec<FunctionCallOutputContentItem> = Vec::with_capacity(items.len());
+    let mut remaining_budget = match policy {
+        TruncationPolicy::Bytes(_) => policy.byte_budget(),
+        TruncationPolicy::Tokens(_) => policy.token_budget(),
+    };
+    let mut omitted_text_items = 0usize;
 
-        let left_budget = keep_budget / 2;
-        let right_budget = keep_budget - left_budget;
-        let prefix_end = pick_prefix_end(s, left_budget);
-        let mut suffix_start = pick_suffix_start(s, right_budget);
-        if suffix_start < prefix_end {
-            suffix_start = prefix_end;
+    for it in items {
+        match it {
+            FunctionCallOutputContentItem::InputText { text } => {
+                if remaining_budget == 0 {
+                    omitted_text_items += 1;
+                    continue;
+                }
+
+                let cost = match policy {
+                    TruncationPolicy::Bytes(_) => text.len(),
+                    TruncationPolicy::Tokens(_) => approx_token_count(text),
+                };
+
+                if cost <= remaining_budget {
+                    out.push(FunctionCallOutputContentItem::InputText { text: text.clone() });
+                    remaining_budget = remaining_budget.saturating_sub(cost);
+                } else {
+                    let snippet_policy = match policy {
+                        TruncationPolicy::Bytes(_) => TruncationPolicy::Bytes(remaining_budget),
+                        TruncationPolicy::Tokens(_) => TruncationPolicy::Tokens(remaining_budget),
+                    };
+                    let snippet = truncate_text(text, snippet_policy);
+                    if snippet.is_empty() {
+                        omitted_text_items += 1;
+                    } else {
+                        out.push(FunctionCallOutputContentItem::InputText { text: snippet });
+                    }
+                    remaining_budget = 0;
+                }
+            }
+            FunctionCallOutputContentItem::InputImage { image_url } => {
+                out.push(FunctionCallOutputContentItem::InputImage {
+                    image_url: image_url.clone(),
+                });
+            }
         }
-
-        // Tokens actually removed (middle slice) using the real tokenizer.
-        let removed_tokens = token_count(&s[prefix_end..suffix_start]);
-
-        // If the number of digits in the token count does not change the marker length,
-        // we can finalize output.
-        let final_marker = format!("…{removed_tokens} tokens truncated…");
-        if final_marker.len() == marker_len {
-            let kept_content_bytes = prefix_end + (s.len() - suffix_start);
-            let mut out = String::with_capacity(final_marker.len() + kept_content_bytes + 1);
-            out.push_str(&s[..prefix_end]);
-            out.push_str(&final_marker);
-            out.push('\n');
-            out.push_str(&s[suffix_start..]);
-            return (out, Some(total_tokens));
-        }
-
-        guess_tokens = removed_tokens;
     }
 
-    // Fallback build after iterations: compute with the last guess.
-    let marker = format!("…{guess_tokens} tokens truncated…");
+    if omitted_text_items > 0 {
+        out.push(FunctionCallOutputContentItem::InputText {
+            text: format!("[omitted {omitted_text_items} text items ...]"),
+        });
+    }
+
+    out
+}
+
+/// Truncate the middle of a UTF-8 string to at most `max_tokens` tokens,
+/// preserving the beginning and the end. Returns the possibly truncated string
+/// and `Some(original_token_count)` if truncation occurred; otherwise returns
+/// the original string and `None`.
+fn truncate_with_token_budget(
+    s: &str,
+    max_tokens: usize,
+    source: TruncationSource,
+) -> (String, Option<u64>) {
+    if s.is_empty() {
+        return (String::new(), None);
+    }
+
+    let byte_len = s.len();
+    if max_tokens > 0 {
+        let small_threshold = approx_bytes_for_tokens(max_tokens / 4);
+        if small_threshold > 0 && byte_len <= small_threshold {
+            return (s.to_string(), None);
+        }
+    }
+
+    let truncated = truncate_with_byte_estimate(s, approx_bytes_for_tokens(max_tokens), source);
+    let approx_total_usize = approx_token_count(s);
+    let approx_total = u64::try_from(approx_total_usize).unwrap_or(u64::MAX);
+    if truncated == s {
+        (truncated, None)
+    } else {
+        (truncated, Some(approx_total))
+    }
+}
+
+/// Truncate a string using a byte budget derived from the token budget, without
+/// performing any real tokenization. This keeps the logic purely byte-based and
+/// uses a bytes placeholder in the truncated output.
+fn truncate_with_byte_estimate(s: &str, max_bytes: usize, source: TruncationSource) -> String {
+    if s.is_empty() {
+        return String::new();
+    }
+
+    if max_bytes == 0 {
+        // No budget to show content; just report that everything was truncated.
+        let marker = format_truncation_marker(source, removed_units_for_source(source, s.len()));
+        return marker;
+    }
+
+    if s.len() <= max_bytes {
+        return s.to_string();
+    }
+
+    let total_bytes = s.len();
+    let removed_bytes = total_bytes.saturating_sub(max_bytes);
+    let marker = format_truncation_marker(source, removed_units_for_source(source, removed_bytes));
     let marker_len = marker.len();
-    let keep_budget = max_bytes.saturating_sub(marker_len);
-    if keep_budget == 0 {
-        return (
-            format!("…{total_tokens} tokens truncated…"),
-            Some(total_tokens),
-        );
+
+    if marker_len >= max_bytes {
+        let truncated_marker = truncate_on_boundary(&marker, max_bytes);
+        return truncated_marker.to_string();
     }
 
-    let left_budget = keep_budget / 2;
-    let right_budget = keep_budget - left_budget;
+    let keep_budget = max_bytes - marker_len;
+    let (left_budget, right_budget) = split_budget(keep_budget);
     let prefix_end = pick_prefix_end(s, left_budget);
     let mut suffix_start = pick_suffix_start(s, right_budget);
     if suffix_start < prefix_end {
         suffix_start = prefix_end;
     }
 
-    let mut out = String::with_capacity(marker_len + prefix_end + (s.len() - suffix_start) + 1);
-    out.push_str(&s[..prefix_end]);
-    out.push_str(&marker);
+    let mut out = assemble_truncated_output(&s[..prefix_end], &s[suffix_start..], &marker);
+
+    if out.len() > max_bytes {
+        let boundary = truncate_on_boundary(&out, max_bytes);
+        out.truncate(boundary.len());
+    }
+
+    out
+}
+
+fn truncate_formatted_exec_output(
+    content: &str,
+    total_lines: usize,
+    limit_bytes: usize,
+    limit_lines: usize,
+) -> String {
+    error_on_double_truncation(content);
+    let head_lines: usize = limit_lines / 2;
+    let tail_lines: usize = limit_lines - head_lines; // 128
+    let head_bytes: usize = limit_bytes / 2;
+    let segments: Vec<&str> = content.split_inclusive('\n').collect();
+    let head_take = head_lines.min(segments.len());
+    let tail_take = tail_lines.min(segments.len().saturating_sub(head_take));
+    let omitted = segments.len().saturating_sub(head_take + tail_take);
+
+    let head_slice_end: usize = segments
+        .iter()
+        .take(head_take)
+        .map(|segment| segment.len())
+        .sum();
+    let tail_slice_start: usize = if tail_take == 0 {
+        content.len()
+    } else {
+        content.len()
+            - segments
+                .iter()
+                .rev()
+                .take(tail_take)
+                .map(|segment| segment.len())
+                .sum::<usize>()
+    };
+    let head_slice = &content[..head_slice_end];
+    let tail_slice = &content[tail_slice_start..];
+    let truncated_by_bytes = content.len() > limit_bytes;
+    // this is a bit wrong. We are counting metadata lines and not just shell output lines.
+    let marker = if omitted > 0 {
+        let marker_text = format_truncation_marker(
+            TruncationSource::LineOmission { total_lines },
+            u64::try_from(omitted).unwrap_or(u64::MAX),
+        );
+        Some(format!("\n{marker_text}\n\n"))
+    } else if truncated_by_bytes {
+        let removed_bytes =
+            u64::try_from(content.len().saturating_sub(limit_bytes)).unwrap_or(u64::MAX);
+        let marker_text =
+            format_truncation_marker(TruncationSource::ByteLimit { limit_bytes }, removed_bytes);
+        Some(format!("\n{marker_text}\n\n"))
+    } else {
+        None
+    };
+
+    let marker_len = marker.as_ref().map_or(0, String::len);
+    let base_head_budget = head_bytes.min(limit_bytes);
+    let head_budget = base_head_budget.min(limit_bytes.saturating_sub(marker_len));
+    let head_part = take_bytes_at_char_boundary(head_slice, head_budget);
+    let mut result = String::with_capacity(limit_bytes.min(content.len()));
+
+    result.push_str(head_part);
+    if let Some(marker_text) = marker.as_ref() {
+        result.push_str(marker_text);
+    }
+
+    let remaining = limit_bytes.saturating_sub(result.len());
+    if remaining == 0 {
+        return result;
+    }
+
+    let tail_part = take_last_bytes_at_char_boundary(tail_slice, remaining);
+    result.push_str(tail_part);
+
+    result
+}
+
+#[derive(Clone, Copy)]
+pub enum TruncationSource {
+    Policy(TruncationPolicy),
+    LineOmission { total_lines: usize },
+    ByteLimit { limit_bytes: usize },
+}
+
+fn format_truncation_marker(source: TruncationSource, removed_count: u64) -> String {
+    match source {
+        TruncationSource::Policy(TruncationPolicy::Tokens(_)) => {
+            format!("[…{removed_count} tokens truncated…]")
+        }
+        TruncationSource::Policy(TruncationPolicy::Bytes(_)) => {
+            format!("[…{removed_count} bytes truncated…]")
+        }
+        TruncationSource::LineOmission { total_lines } => {
+            format!("[... omitted {removed_count} of {total_lines} lines ...]")
+        }
+        TruncationSource::ByteLimit { limit_bytes } => {
+            format!("[... removed {removed_count} bytes to fit {limit_bytes} byte limit ...]")
+        }
+    }
+}
+
+fn split_budget(budget: usize) -> (usize, usize) {
+    let left = budget / 2;
+    (left, budget - left)
+}
+
+fn removed_units_for_source(source: TruncationSource, removed_bytes: usize) -> u64 {
+    match source {
+        TruncationSource::Policy(TruncationPolicy::Tokens(_)) => {
+            approx_tokens_from_byte_count(removed_bytes)
+        }
+        _ => u64::try_from(removed_bytes).unwrap_or(u64::MAX),
+    }
+}
+
+fn assemble_truncated_output(prefix: &str, suffix: &str, marker: &str) -> String {
+    let mut out = String::with_capacity(prefix.len() + marker.len() + suffix.len() + 1);
+    out.push_str(prefix);
+    out.push_str(marker);
     out.push('\n');
-    out.push_str(&s[suffix_start..]);
-    (out, Some(total_tokens))
+    out.push_str(suffix);
+    out
+}
+
+pub(crate) fn approx_token_count(text: &str) -> usize {
+    let len = text.len();
+    len.saturating_add(APPROX_BYTES_PER_TOKEN.saturating_sub(1)) / APPROX_BYTES_PER_TOKEN
+}
+
+fn approx_bytes_for_tokens(tokens: usize) -> usize {
+    tokens.saturating_mul(APPROX_BYTES_PER_TOKEN)
+}
+
+fn approx_tokens_from_byte_count(bytes: usize) -> u64 {
+    let bytes_u64 = bytes as u64;
+    bytes_u64.saturating_add((APPROX_BYTES_PER_TOKEN as u64).saturating_sub(1))
+        / (APPROX_BYTES_PER_TOKEN as u64)
+}
+
+fn truncate_on_boundary(input: &str, max_len: usize) -> &str {
+    if input.len() <= max_len {
+        return input;
+    }
+    let mut end = max_len;
+    while end > 0 && !input.is_char_boundary(end) {
+        end -= 1;
+    }
+    &input[..end]
+}
+
+fn pick_prefix_end(s: &str, left_budget: usize) -> usize {
+    if let Some(head) = s.get(..left_budget)
+        && let Some(i) = head.rfind('\n')
+    {
+        return i + 1;
+    }
+    truncate_on_boundary(s, left_budget).len()
+}
+
+fn pick_suffix_start(s: &str, right_budget: usize) -> usize {
+    let start_tail = s.len().saturating_sub(right_budget);
+    if let Some(tail) = s.get(start_tail..)
+        && let Some(i) = tail.find('\n')
+    {
+        return start_tail + i + 1;
+    }
+
+    let mut idx = start_tail.min(s.len());
+    while idx < s.len() && !s.is_char_boundary(idx) {
+        idx += 1;
+    }
+    idx
+}
+
+fn error_on_double_truncation(content: &str) {
+    if content.contains("Total output lines:") && content.contains("omitted") {
+        tracing::error!(
+            "FunctionCallOutput content was already truncated before ContextManager::record_items; this would cause double truncation {content}"
+        );
+    }
 }
 
 #[cfg(test)]
 mod tests {
-    use super::truncate_middle;
-    use codex_utils_tokenizer::Tokenizer;
+    use crate::config::OPENAI_DEFAULT_MODEL;
+    use crate::model_family::derive_default_model_family;
+    use crate::model_family::find_family_for_model;
 
-    #[test]
-    fn truncate_middle_no_newlines_fallback() {
-        let tok = Tokenizer::try_default().expect("load tokenizer");
-        let s = "abcdefghijklmnopqrstuvwxyz0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ*";
-        let max_bytes = 32;
-        let (out, original) = truncate_middle(s, max_bytes);
-        assert!(out.starts_with("abc"));
-        assert!(out.contains("tokens truncated"));
-        assert!(out.ends_with("XYZ*"));
-        assert_eq!(original, Some(tok.count(s) as u64));
+    use super::TruncationPolicy;
+    use super::TruncationSource;
+    use super::approx_token_count;
+    use super::truncate_function_output_items_with_policy;
+    use super::truncate_with_line_bytes_budget;
+    use super::truncate_with_token_budget;
+    use codex_protocol::models::FunctionCallOutputContentItem;
+    use pretty_assertions::assert_eq;
+    use regex_lite::Regex;
+
+    const MODEL_FORMAT_MAX_LINES: usize = 256;
+
+    fn model_format_max_bytes() -> usize {
+        find_family_for_model(OPENAI_DEFAULT_MODEL)
+            .unwrap_or_else(|| derive_default_model_family(OPENAI_DEFAULT_MODEL))
+            .truncation_policy
+            .byte_budget()
+    }
+
+    fn truncated_message_pattern(line: &str, total_lines: usize) -> String {
+        let head_lines = MODEL_FORMAT_MAX_LINES / 2;
+        let tail_lines = MODEL_FORMAT_MAX_LINES - head_lines;
+        let head_take = head_lines.min(total_lines);
+        let tail_take = tail_lines.min(total_lines.saturating_sub(head_take));
+        let omitted = total_lines.saturating_sub(head_take + tail_take);
+        let escaped_line = regex_lite::escape(line);
+        if omitted == 0 {
+            return format!(
+                r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} removed \d+ bytes to fit {max_bytes} byte limit \.{{3}}]\n\n.*)$",
+                max_bytes = model_format_max_bytes(),
+            );
+        }
+        format!(
+            r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} omitted {omitted} of {total_lines} lines \.{{3}}]\n\n.*)$",
+        )
     }
 
     #[test]
-    fn truncate_middle_prefers_newline_boundaries() {
-        let tok = Tokenizer::try_default().expect("load tokenizer");
-        let mut s = String::new();
-        for i in 1..=20 {
-            s.push_str(&format!("{i:03}\n"));
-        }
-        assert_eq!(s.len(), 80);
+    fn truncate_middle_returns_original_when_under_limit() {
+        let s = "short output";
+        let limit = 100;
+        let source = TruncationSource::Policy(TruncationPolicy::Tokens(limit));
+        let (out, original) = truncate_with_token_budget(s, limit, source);
+        assert_eq!(out, s);
+        assert_eq!(original, None);
+    }
 
-        let max_bytes = 64;
-        let (out, tokens) = truncate_middle(&s, max_bytes);
-        assert!(out.starts_with("001\n002\n003\n004\n"));
+    #[test]
+    fn truncate_middle_reports_truncation_at_zero_limit() {
+        let s = "abcdef";
+        let source = TruncationSource::Policy(TruncationPolicy::Tokens(0));
+        let (out, original) = truncate_with_token_budget(s, 0, source);
+        assert_eq!(out, "[…2 tokens truncated…]");
+        assert_eq!(original, Some(approx_token_count(s) as u64));
+    }
+
+    #[test]
+    fn truncate_middle_enforces_token_budget() {
+        let s = "alpha beta gamma delta epsilon zeta eta theta iota kappa";
+        let max_tokens = 12;
+        let source = TruncationSource::Policy(TruncationPolicy::Tokens(max_tokens));
+        let (out, original) = truncate_with_token_budget(s, max_tokens, source);
         assert!(out.contains("tokens truncated"));
-        assert!(out.ends_with("017\n018\n019\n020\n"));
-        assert_eq!(tokens, Some(tok.count(&s) as u64));
+        assert_eq!(original, Some(approx_token_count(s) as u64));
+        assert!(out.len() < s.len(), "truncated output should be shorter");
     }
 
     #[test]
     fn truncate_middle_handles_utf8_content() {
-        let tok = Tokenizer::try_default().expect("load tokenizer");
-        let s = "😀😀😀😀😀😀😀😀😀😀\nsecond line with ascii text\n";
-        let max_bytes = 32;
-        let (out, tokens) = truncate_middle(s, max_bytes);
+        let s = "😀😀😀😀😀😀😀😀😀😀\nsecond line with text\n";
+        let max_tokens = 8;
+        let source = TruncationSource::Policy(TruncationPolicy::Tokens(max_tokens));
+        let (out, tokens) = truncate_with_token_budget(s, max_tokens, source);
 
         assert!(out.contains("tokens truncated"));
         assert!(!out.contains('\u{fffd}'));
-        assert_eq!(tokens, Some(tok.count(s) as u64));
+        assert_eq!(tokens, Some(approx_token_count(s) as u64));
+        assert!(out.len() < s.len(), "UTF-8 content should be shortened");
     }
 
     #[test]
-    fn truncate_middle_prefers_newline_boundaries_2() {
-        let tok = Tokenizer::try_default().expect("load tokenizer");
-        // Build a multi-line string of 20 numbered lines (each "NNN\n").
-        let mut s = String::new();
-        for i in 1..=20 {
-            s.push_str(&format!("{i:03}\n"));
-        }
-        assert_eq!(s.len(), 80);
+    fn format_exec_output_truncates_large_error() {
+        let line = "very long execution error line that should trigger truncation\n";
+        let large_error = line.repeat(2_500); // way beyond both byte and line limits
 
-        let max_bytes = 64;
-        let (out, total) = truncate_middle(&s, max_bytes);
-        assert!(out.starts_with("001\n002\n003\n004\n"));
-        assert!(out.contains("tokens truncated"));
-        assert!(out.ends_with("017\n018\n019\n020\n"));
-        assert_eq!(total, Some(tok.count(&s) as u64));
+        let truncated = truncate_with_line_bytes_budget(&large_error, model_format_max_bytes());
+
+        let total_lines = large_error.lines().count();
+        let pattern = truncated_message_pattern(line, total_lines);
+        let regex = Regex::new(&pattern).unwrap_or_else(|err| {
+            panic!("failed to compile regex {pattern}: {err}");
+        });
+        let captures = regex
+            .captures(&truncated)
+            .unwrap_or_else(|| panic!("message failed to match pattern {pattern}: {truncated}"));
+        let body = captures
+            .name("body")
+            .expect("missing body capture")
+            .as_str();
+        assert!(
+            body.len() <= model_format_max_bytes(),
+            "body exceeds byte limit: {} bytes",
+            body.len()
+        );
+        assert_ne!(truncated, large_error);
+    }
+
+    #[test]
+    fn format_exec_output_marks_byte_truncation_without_omitted_lines() {
+        let max_bytes = model_format_max_bytes();
+        let long_line = "a".repeat(max_bytes + 50);
+        let truncated = truncate_with_line_bytes_budget(&long_line, max_bytes);
+
+        assert_ne!(truncated, long_line);
+        let removed_bytes = long_line.len().saturating_sub(max_bytes);
+        let marker_line =
+            format!("[... removed {removed_bytes} bytes to fit {max_bytes} byte limit ...]");
+        assert!(
+            truncated.contains(&marker_line),
+            "missing byte truncation marker: {truncated}"
+        );
+        assert!(
+            !truncated.contains("omitted"),
+            "line omission marker should not appear when no lines were dropped: {truncated}"
+        );
+    }
+
+    #[test]
+    fn format_exec_output_returns_original_when_within_limits() {
+        let content = "example output\n".repeat(10);
+
+        assert_eq!(
+            truncate_with_line_bytes_budget(&content, model_format_max_bytes()),
+            content
+        );
+    }
+
+    #[test]
+    fn format_exec_output_reports_omitted_lines_and_keeps_head_and_tail() {
+        let total_lines = MODEL_FORMAT_MAX_LINES + 100;
+        let content: String = (0..total_lines)
+            .map(|idx| format!("line-{idx}\n"))
+            .collect();
+
+        let truncated = truncate_with_line_bytes_budget(&content, model_format_max_bytes());
+
+        let omitted = total_lines - MODEL_FORMAT_MAX_LINES;
+        let expected_marker = format!("[... omitted {omitted} of {total_lines} lines ...]");
+
+        assert!(
+            truncated.contains(&expected_marker),
+            "missing omitted marker: {truncated}"
+        );
+        assert!(
+            truncated.contains("line-0\n"),
+            "expected head line to remain: {truncated}"
+        );
+
+        let last_line = format!("line-{}\n", total_lines - 1);
+        assert!(
+            truncated.contains(&last_line),
+            "expected tail line to remain: {truncated}"
+        );
+    }
+
+    #[test]
+    fn format_exec_output_prefers_line_marker_when_both_limits_exceeded() {
+        let total_lines = MODEL_FORMAT_MAX_LINES + 42;
+        let long_line = "x".repeat(256);
+        let content: String = (0..total_lines)
+            .map(|idx| format!("line-{idx}-{long_line}\n"))
+            .collect();
+
+        let truncated = truncate_with_line_bytes_budget(&content, model_format_max_bytes());
+
+        assert!(
+            truncated.contains("[... omitted 42 of 298 lines ...]"),
+            "expected omitted marker when line count exceeds limit: {truncated}"
+        );
+        assert!(
+            !truncated.contains("byte limit"),
+            "line omission marker should take precedence over byte marker: {truncated}"
+        );
+    }
+
+    #[test]
+    fn truncates_across_multiple_under_limit_texts_and_reports_omitted() {
+        let chunk = "alpha beta gamma delta epsilon zeta eta theta iota kappa lambda mu nu xi omicron pi rho sigma tau upsilon phi chi psi omega.\n";
+        let chunk_tokens = approx_token_count(chunk);
+        assert!(chunk_tokens > 0, "chunk must consume tokens");
+        let limit = chunk_tokens * 3;
+        let t1 = chunk.to_string();
+        let t2 = chunk.to_string();
+        let t3 = chunk.repeat(10);
+        let t4 = chunk.to_string();
+        let t5 = chunk.to_string();
+
+        let items = vec![
+            FunctionCallOutputContentItem::InputText { text: t1.clone() },
+            FunctionCallOutputContentItem::InputText { text: t2.clone() },
+            FunctionCallOutputContentItem::InputImage {
+                image_url: "img:mid".to_string(),
+            },
+            FunctionCallOutputContentItem::InputText { text: t3 },
+            FunctionCallOutputContentItem::InputText { text: t4 },
+            FunctionCallOutputContentItem::InputText { text: t5 },
+        ];
+
+        let output =
+            truncate_function_output_items_with_policy(&items, TruncationPolicy::Tokens(limit));
+
+        // Expect: t1 (full), t2 (full), image, t3 (truncated), summary mentioning 2 omitted.
+        assert_eq!(output.len(), 5);
+
+        let first_text = match &output[0] {
+            FunctionCallOutputContentItem::InputText { text } => text,
+            other => panic!("unexpected first item: {other:?}"),
+        };
+        assert_eq!(first_text, &t1);
+
+        let second_text = match &output[1] {
+            FunctionCallOutputContentItem::InputText { text } => text,
+            other => panic!("unexpected second item: {other:?}"),
+        };
+        assert_eq!(second_text, &t2);
+
+        assert_eq!(
+            output[2],
+            FunctionCallOutputContentItem::InputImage {
+                image_url: "img:mid".to_string()
+            }
+        );
+
+        let fourth_text = match &output[3] {
+            FunctionCallOutputContentItem::InputText { text } => text,
+            other => panic!("unexpected fourth item: {other:?}"),
+        };
+        assert!(
+            fourth_text.contains("tokens truncated"),
+            "expected marker in truncated snippet: {fourth_text}"
+        );
+
+        let summary_text = match &output[4] {
+            FunctionCallOutputContentItem::InputText { text } => text,
+            other => panic!("unexpected summary item: {other:?}"),
+        };
+        assert!(summary_text.contains("omitted 2 text items"));
     }
 }

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

@@ -45,6 +45,7 @@ pub(crate) const MIN_YIELD_TIME_MS: u64 = 250;
 pub(crate) const MAX_YIELD_TIME_MS: u64 = 30_000;
 pub(crate) const DEFAULT_MAX_OUTPUT_TOKENS: usize = 10_000;
 pub(crate) const UNIFIED_EXEC_OUTPUT_MAX_BYTES: usize = 1024 * 1024; // 1 MiB
+pub(crate) const UNIFIED_EXEC_OUTPUT_MAX_TOKENS: usize = UNIFIED_EXEC_OUTPUT_MAX_BYTES / 4;
 
 pub(crate) struct UnifiedExecContext {
     pub session: Arc<Session>,
@@ -74,6 +75,7 @@ pub(crate) struct ExecCommandRequest {
 
 #[derive(Debug)]
 pub(crate) struct WriteStdinRequest<'a> {
+    pub call_id: &'a str,
     pub session_id: i32,
     pub input: &'a str,
     pub yield_time_ms: u64,
@@ -89,6 +91,7 @@ pub(crate) struct UnifiedExecResponse {
     pub session_id: Option<i32>,
     pub exit_code: Option<i32>,
     pub original_token_count: Option<usize>,
+    pub session_command: Option<Vec<String>>,
 }
 
 #[derive(Default)]
@@ -122,37 +125,6 @@ pub(crate) fn generate_chunk_id() -> String {
         .collect()
 }
 
-pub(crate) fn truncate_output_to_tokens(
-    output: &str,
-    max_tokens: usize,
-) -> (String, Option<usize>) {
-    if max_tokens == 0 {
-        let total_tokens = output.chars().count();
-        let message = format!("…{total_tokens} tokens truncated…");
-        return (message, Some(total_tokens));
-    }
-
-    let tokens: Vec<char> = output.chars().collect();
-    let total_tokens = tokens.len();
-    if total_tokens <= max_tokens {
-        return (output.to_string(), None);
-    }
-
-    let half = max_tokens / 2;
-    if half == 0 {
-        let truncated = total_tokens.saturating_sub(max_tokens);
-        let message = format!("…{truncated} tokens truncated…");
-        return (message, Some(total_tokens));
-    }
-
-    let truncated = total_tokens.saturating_sub(half * 2);
-    let mut truncated_output = String::new();
-    truncated_output.extend(&tokens[..half]);
-    truncated_output.push_str(&format!("…{truncated} tokens truncated…"));
-    truncated_output.extend(&tokens[total_tokens - half..]);
-    (truncated_output, Some(total_tokens))
-}
-
 #[cfg(test)]
 #[cfg(unix)]
 mod tests {
@@ -213,6 +185,7 @@ mod tests {
             .services
             .unified_exec_manager
             .write_stdin(WriteStdinRequest {
+                call_id: "write-stdin",
                 session_id,
                 input,
                 yield_time_ms,

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

@@ -14,11 +14,13 @@ use crate::exec::ExecToolCallOutput;
 use crate::exec::SandboxType;
 use crate::exec::StreamOutput;
 use crate::exec::is_likely_sandbox_denied;
-use crate::truncate::truncate_middle;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::truncate_text;
 use codex_utils_pty::ExecCommandSession;
 use codex_utils_pty::SpawnedPty;
 
 use super::UNIFIED_EXEC_OUTPUT_MAX_BYTES;
+use super::UNIFIED_EXEC_OUTPUT_MAX_TOKENS;
 use super::UnifiedExecError;
 
 #[derive(Debug, Default)]
@@ -165,7 +167,10 @@ impl UnifiedExecSession {
         };
 
         if is_likely_sandbox_denied(self.sandbox_type(), &exec_output) {
-            let (snippet, _) = truncate_middle(&aggregated_text, UNIFIED_EXEC_OUTPUT_MAX_BYTES);
+            let snippet = truncate_text(
+                &aggregated_text,
+                TruncationPolicy::Tokens(UNIFIED_EXEC_OUTPUT_MAX_TOKENS),
+            );
             let message = if snippet.is_empty() {
                 format!("exit code {exit_code}")
             } else {