[core] add optional status_code to error events (#6865)

We want to better uncover error status code for clients. Add an optional
status_code to error events (thread error, error, stream error) so app
server could uncover the status code from the client side later.

in event log:
```
< {
<   "method": "codex/event/stream_error",
<   "params": {
<     "conversationId": "019a9a32-f576-7292-9711-8e57e8063536",
<     "id": "0",
<     "msg": {
<       "message": "Reconnecting... 5/5",
<       "status_code": 401,
<       "type": "stream_error"
<     }
<   }
< }
< {
<   "method": "codex/event/error",
<   "params": {
<     "conversationId": "019a9a32-f576-7292-9711-8e57e8063536",
<     "id": "0",
<     "msg": {
<       "message": "exceeded retry limit, last status: 401 Unauthorized, request id: 9a0cb03a485067f7-SJC",
<       "status_code": 401,
<       "type": "error"
<     }
<   }
< }
```

.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

@@ -46,4 +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

@@ -97,7 +97,6 @@ jobs:
     env:
       # Speed up repeated builds across CI runs by caching compiled objects (non-Windows).
       USE_SCCACHE: ${{ startsWith(matrix.runner, 'windows') && 'false' || 'true' }}
-      RUSTC_WRAPPER: ${{ startsWith(matrix.runner, 'windows') && '' || 'sccache' }}
       CARGO_INCREMENTAL: "0"
       SCCACHE_CACHE_SIZE: 10G
 
@@ -191,6 +190,11 @@ 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.USE_SCCACHE == 'true' && env.SCCACHE_GHA_ENABLED != 'true' }}
         id: cache_sccache_restore
@@ -331,7 +335,6 @@ jobs:
     env:
       # Speed up repeated builds across CI runs by caching compiled objects (non-Windows).
       USE_SCCACHE: ${{ startsWith(matrix.runner, 'windows') && 'false' || 'true' }}
-      RUSTC_WRAPPER: ${{ startsWith(matrix.runner, 'windows') && '' || 'sccache' }}
       CARGO_INCREMENTAL: "0"
       SCCACHE_CACHE_SIZE: 10G
 
@@ -395,6 +398,11 @@ 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.USE_SCCACHE == 'true' && env.SCCACHE_GHA_ENABLED != 'true' }}
         id: cache_sccache_restore

.gitignore

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

codex-rs/Cargo.lock

@@ -849,6 +849,7 @@ dependencies = [
  "codex-login",
  "codex-protocol",
  "codex-utils-json-to-toml",
+ "codex-windows-sandbox",
  "core_test_support",
  "mcp-types",
  "opentelemetry-appender-tracing",
@@ -874,7 +875,6 @@ dependencies = [
  "clap",
  "codex-protocol",
  "mcp-types",
- "paste",
  "pretty_assertions",
  "schemars 0.8.22",
  "serde",
@@ -1061,6 +1061,8 @@ dependencies = [
  "clap",
  "codex-app-server-protocol",
  "codex-core",
+ "codex-lmstudio",
+ "codex-ollama",
  "codex-protocol",
  "once_cell",
  "serde",
@@ -1159,7 +1161,6 @@ dependencies = [
  "codex-arg0",
  "codex-common",
  "codex-core",
- "codex-ollama",
  "codex-protocol",
  "core_test_support",
  "libc",
@@ -1182,6 +1183,26 @@ 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"
@@ -1202,6 +1223,21 @@ dependencies = [
  "tempfile",
 ]
 
+[[package]]
+name = "codex-execpolicy2"
+version = "0.0.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "multimap",
+ "pretty_assertions",
+ "serde",
+ "serde_json",
+ "shlex",
+ "starlark",
+ "thiserror 2.0.17",
+]
+
 [[package]]
 name = "codex-feedback"
 version = "0.0.0"
@@ -1263,6 +1299,19 @@ 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"
@@ -1338,6 +1387,7 @@ dependencies = [
  "codex-app-server-protocol",
  "codex-protocol",
  "eventsource-stream",
+ "http",
  "opentelemetry",
  "opentelemetry-otlp",
  "opentelemetry-semantic-conventions",
@@ -1371,6 +1421,7 @@ dependencies = [
  "icu_provider",
  "mcp-types",
  "mime_guess",
+ "pretty_assertions",
  "schemars 0.8.22",
  "serde",
  "serde_json",
@@ -1427,6 +1478,7 @@ dependencies = [
  "tracing",
  "urlencoding",
  "webbrowser",
+ "which",
 ]
 
 [[package]]
@@ -1454,12 +1506,12 @@ 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",
@@ -1482,6 +1534,7 @@ dependencies = [
  "ratatui",
  "ratatui-macros",
  "regex-lite",
+ "reqwest",
  "serde",
  "serde_json",
  "serial_test",
@@ -1576,6 +1629,7 @@ name = "codex-windows-sandbox"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "codex-protocol",
  "dirs-next",
  "dunce",
  "rand 0.8.5",
@@ -5144,6 +5198,7 @@ version = "0.23.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2491382039b29b9b11ff08b76ff6c97cf287671dbb74f0be44bda389fffe9bd1"
 dependencies = [
+ "log",
  "once_cell",
  "ring",
  "rustls-pki-types",
@@ -6571,8 +6626,10 @@ 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

@@ -16,10 +16,13 @@ members = [
     "common",
     "core",
     "exec",
+    "exec-server",
     "execpolicy",
+    "execpolicy2",
     "keyring-store",
     "file-search",
     "linux-sandbox",
+    "lmstudio",
     "login",
     "mcp-server",
     "mcp-types",
@@ -69,6 +72,7 @@ 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" }
@@ -130,7 +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"
@@ -150,7 +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"
@@ -174,6 +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"

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/lib.rs

@@ -7,5 +7,6 @@ pub use export::generate_ts;
 pub use export::generate_types;
 pub use jsonrpc_lite::*;
 pub use protocol::common::*;
+pub use protocol::thread_history::*;
 pub use protocol::v1::*;
 pub use protocol::v2::*;

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)]
@@ -512,6 +494,9 @@ server_notification_definitions! {
     ReasoningSummaryPartAdded => "item/reasoning/summaryPartAdded" (v2::ReasoningSummaryPartAddedNotification),
     ReasoningTextDelta => "item/reasoning/textDelta" (v2::ReasoningTextDeltaNotification),
 
+    /// Notifies the user of world-writable directories on Windows, which cannot be protected by the sandbox.
+    WindowsWorldWritableWarning => "windows/worldWritableWarning" (v2::WindowsWorldWritableWarningNotification),
+
     #[serde(rename = "account/login/completed")]
     #[ts(rename = "account/login/completed")]
     #[strum(serialize = "account/login/completed")]
@@ -533,17 +518,20 @@ 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-max".to_string()),
                 model_provider: None,
                 profile: None,
                 cwd: None,
@@ -561,7 +549,7 @@ mod tests {
                 "method": "newConversation",
                 "id": 42,
                 "params": {
-                    "model": "gpt-5-codex",
+                    "model": "gpt-5.1-codex-max",
                     "modelProvider": null,
                     "profile": null,
                     "cwd": null,
@@ -616,7 +604,7 @@ 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/mod.rs

@@ -2,5 +2,6 @@
 // Exposes protocol pieces used by `lib.rs` via `pub use protocol::common::*;`.
 
 pub mod common;
+pub mod thread_history;
 pub mod v1;
 pub mod v2;

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

@@ -0,0 +1,409 @@
+use crate::protocol::v2::ThreadItem;
+use crate::protocol::v2::Turn;
+use crate::protocol::v2::TurnStatus;
+use crate::protocol::v2::UserInput;
+use codex_protocol::protocol::AgentReasoningEvent;
+use codex_protocol::protocol::AgentReasoningRawContentEvent;
+use codex_protocol::protocol::EventMsg;
+use codex_protocol::protocol::TurnAbortedEvent;
+use codex_protocol::protocol::UserMessageEvent;
+
+/// Convert persisted [`EventMsg`] entries into a sequence of [`Turn`] values.
+///
+/// The purpose of this is to convert the EventMsgs persisted in a rollout file
+/// into a sequence of Turns and ThreadItems, which allows the client to render
+/// the historical messages when resuming a thread.
+pub fn build_turns_from_event_msgs(events: &[EventMsg]) -> Vec<Turn> {
+    let mut builder = ThreadHistoryBuilder::new();
+    for event in events {
+        builder.handle_event(event);
+    }
+    builder.finish()
+}
+
+struct ThreadHistoryBuilder {
+    turns: Vec<Turn>,
+    current_turn: Option<PendingTurn>,
+    next_turn_index: i64,
+    next_item_index: i64,
+}
+
+impl ThreadHistoryBuilder {
+    fn new() -> Self {
+        Self {
+            turns: Vec::new(),
+            current_turn: None,
+            next_turn_index: 1,
+            next_item_index: 1,
+        }
+    }
+
+    fn finish(mut self) -> Vec<Turn> {
+        self.finish_current_turn();
+        self.turns
+    }
+
+    /// This function should handle all EventMsg variants that can be persisted in a rollout file.
+    /// See `should_persist_event_msg` in `codex-rs/core/rollout/policy.rs`.
+    fn handle_event(&mut self, event: &EventMsg) {
+        match event {
+            EventMsg::UserMessage(payload) => self.handle_user_message(payload),
+            EventMsg::AgentMessage(payload) => self.handle_agent_message(payload.message.clone()),
+            EventMsg::AgentReasoning(payload) => self.handle_agent_reasoning(payload),
+            EventMsg::AgentReasoningRawContent(payload) => {
+                self.handle_agent_reasoning_raw_content(payload)
+            }
+            EventMsg::TokenCount(_) => {}
+            EventMsg::EnteredReviewMode(_) => {}
+            EventMsg::ExitedReviewMode(_) => {}
+            EventMsg::UndoCompleted(_) => {}
+            EventMsg::TurnAborted(payload) => self.handle_turn_aborted(payload),
+            _ => {}
+        }
+    }
+
+    fn handle_user_message(&mut self, payload: &UserMessageEvent) {
+        self.finish_current_turn();
+        let mut turn = self.new_turn();
+        let id = self.next_item_id();
+        let content = self.build_user_inputs(payload);
+        turn.items.push(ThreadItem::UserMessage { id, content });
+        self.current_turn = Some(turn);
+    }
+
+    fn handle_agent_message(&mut self, text: String) {
+        if text.is_empty() {
+            return;
+        }
+
+        let id = self.next_item_id();
+        self.ensure_turn()
+            .items
+            .push(ThreadItem::AgentMessage { id, text });
+    }
+
+    fn handle_agent_reasoning(&mut self, payload: &AgentReasoningEvent) {
+        if payload.text.is_empty() {
+            return;
+        }
+
+        // If the last item is a reasoning item, add the new text to the summary.
+        if let Some(ThreadItem::Reasoning { summary, .. }) = self.ensure_turn().items.last_mut() {
+            summary.push(payload.text.clone());
+            return;
+        }
+
+        // Otherwise, create a new reasoning item.
+        let id = self.next_item_id();
+        self.ensure_turn().items.push(ThreadItem::Reasoning {
+            id,
+            summary: vec![payload.text.clone()],
+            content: Vec::new(),
+        });
+    }
+
+    fn handle_agent_reasoning_raw_content(&mut self, payload: &AgentReasoningRawContentEvent) {
+        if payload.text.is_empty() {
+            return;
+        }
+
+        // If the last item is a reasoning item, add the new text to the content.
+        if let Some(ThreadItem::Reasoning { content, .. }) = self.ensure_turn().items.last_mut() {
+            content.push(payload.text.clone());
+            return;
+        }
+
+        // Otherwise, create a new reasoning item.
+        let id = self.next_item_id();
+        self.ensure_turn().items.push(ThreadItem::Reasoning {
+            id,
+            summary: Vec::new(),
+            content: vec![payload.text.clone()],
+        });
+    }
+
+    fn handle_turn_aborted(&mut self, _payload: &TurnAbortedEvent) {
+        let Some(turn) = self.current_turn.as_mut() else {
+            return;
+        };
+        turn.status = TurnStatus::Interrupted;
+    }
+
+    fn finish_current_turn(&mut self) {
+        if let Some(turn) = self.current_turn.take() {
+            if turn.items.is_empty() {
+                return;
+            }
+            self.turns.push(turn.into());
+        }
+    }
+
+    fn new_turn(&mut self) -> PendingTurn {
+        PendingTurn {
+            id: self.next_turn_id(),
+            items: Vec::new(),
+            status: TurnStatus::Completed,
+        }
+    }
+
+    fn ensure_turn(&mut self) -> &mut PendingTurn {
+        if self.current_turn.is_none() {
+            let turn = self.new_turn();
+            return self.current_turn.insert(turn);
+        }
+
+        if let Some(turn) = self.current_turn.as_mut() {
+            return turn;
+        }
+
+        unreachable!("current turn must exist after initialization");
+    }
+
+    fn next_turn_id(&mut self) -> String {
+        let id = format!("turn-{}", self.next_turn_index);
+        self.next_turn_index += 1;
+        id
+    }
+
+    fn next_item_id(&mut self) -> String {
+        let id = format!("item-{}", self.next_item_index);
+        self.next_item_index += 1;
+        id
+    }
+
+    fn build_user_inputs(&self, payload: &UserMessageEvent) -> Vec<UserInput> {
+        let mut content = Vec::new();
+        if !payload.message.trim().is_empty() {
+            content.push(UserInput::Text {
+                text: payload.message.clone(),
+            });
+        }
+        if let Some(images) = &payload.images {
+            for image in images {
+                content.push(UserInput::Image { url: image.clone() });
+            }
+        }
+        content
+    }
+}
+
+struct PendingTurn {
+    id: String,
+    items: Vec<ThreadItem>,
+    status: TurnStatus,
+}
+
+impl From<PendingTurn> for Turn {
+    fn from(value: PendingTurn) -> Self {
+        Self {
+            id: value.id,
+            items: value.items,
+            status: value.status,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use codex_protocol::protocol::AgentMessageEvent;
+    use codex_protocol::protocol::AgentReasoningEvent;
+    use codex_protocol::protocol::AgentReasoningRawContentEvent;
+    use codex_protocol::protocol::TurnAbortReason;
+    use codex_protocol::protocol::TurnAbortedEvent;
+    use codex_protocol::protocol::UserMessageEvent;
+    use pretty_assertions::assert_eq;
+
+    #[test]
+    fn builds_multiple_turns_with_reasoning_items() {
+        let events = vec![
+            EventMsg::UserMessage(UserMessageEvent {
+                message: "First turn".into(),
+                images: Some(vec!["https://example.com/one.png".into()]),
+            }),
+            EventMsg::AgentMessage(AgentMessageEvent {
+                message: "Hi there".into(),
+            }),
+            EventMsg::AgentReasoning(AgentReasoningEvent {
+                text: "thinking".into(),
+            }),
+            EventMsg::AgentReasoningRawContent(AgentReasoningRawContentEvent {
+                text: "full reasoning".into(),
+            }),
+            EventMsg::UserMessage(UserMessageEvent {
+                message: "Second turn".into(),
+                images: None,
+            }),
+            EventMsg::AgentMessage(AgentMessageEvent {
+                message: "Reply two".into(),
+            }),
+        ];
+
+        let turns = build_turns_from_event_msgs(&events);
+        assert_eq!(turns.len(), 2);
+
+        let first = &turns[0];
+        assert_eq!(first.id, "turn-1");
+        assert_eq!(first.status, TurnStatus::Completed);
+        assert_eq!(first.items.len(), 3);
+        assert_eq!(
+            first.items[0],
+            ThreadItem::UserMessage {
+                id: "item-1".into(),
+                content: vec![
+                    UserInput::Text {
+                        text: "First turn".into(),
+                    },
+                    UserInput::Image {
+                        url: "https://example.com/one.png".into(),
+                    }
+                ],
+            }
+        );
+        assert_eq!(
+            first.items[1],
+            ThreadItem::AgentMessage {
+                id: "item-2".into(),
+                text: "Hi there".into(),
+            }
+        );
+        assert_eq!(
+            first.items[2],
+            ThreadItem::Reasoning {
+                id: "item-3".into(),
+                summary: vec!["thinking".into()],
+                content: vec!["full reasoning".into()],
+            }
+        );
+
+        let second = &turns[1];
+        assert_eq!(second.id, "turn-2");
+        assert_eq!(second.items.len(), 2);
+        assert_eq!(
+            second.items[0],
+            ThreadItem::UserMessage {
+                id: "item-4".into(),
+                content: vec![UserInput::Text {
+                    text: "Second turn".into()
+                }],
+            }
+        );
+        assert_eq!(
+            second.items[1],
+            ThreadItem::AgentMessage {
+                id: "item-5".into(),
+                text: "Reply two".into(),
+            }
+        );
+    }
+
+    #[test]
+    fn splits_reasoning_when_interleaved() {
+        let events = vec![
+            EventMsg::UserMessage(UserMessageEvent {
+                message: "Turn start".into(),
+                images: None,
+            }),
+            EventMsg::AgentReasoning(AgentReasoningEvent {
+                text: "first summary".into(),
+            }),
+            EventMsg::AgentReasoningRawContent(AgentReasoningRawContentEvent {
+                text: "first content".into(),
+            }),
+            EventMsg::AgentMessage(AgentMessageEvent {
+                message: "interlude".into(),
+            }),
+            EventMsg::AgentReasoning(AgentReasoningEvent {
+                text: "second summary".into(),
+            }),
+        ];
+
+        let turns = build_turns_from_event_msgs(&events);
+        assert_eq!(turns.len(), 1);
+        let turn = &turns[0];
+        assert_eq!(turn.items.len(), 4);
+
+        assert_eq!(
+            turn.items[1],
+            ThreadItem::Reasoning {
+                id: "item-2".into(),
+                summary: vec!["first summary".into()],
+                content: vec!["first content".into()],
+            }
+        );
+        assert_eq!(
+            turn.items[3],
+            ThreadItem::Reasoning {
+                id: "item-4".into(),
+                summary: vec!["second summary".into()],
+                content: Vec::new(),
+            }
+        );
+    }
+
+    #[test]
+    fn marks_turn_as_interrupted_when_aborted() {
+        let events = vec![
+            EventMsg::UserMessage(UserMessageEvent {
+                message: "Please do the thing".into(),
+                images: None,
+            }),
+            EventMsg::AgentMessage(AgentMessageEvent {
+                message: "Working...".into(),
+            }),
+            EventMsg::TurnAborted(TurnAbortedEvent {
+                reason: TurnAbortReason::Replaced,
+            }),
+            EventMsg::UserMessage(UserMessageEvent {
+                message: "Let's try again".into(),
+                images: None,
+            }),
+            EventMsg::AgentMessage(AgentMessageEvent {
+                message: "Second attempt complete.".into(),
+            }),
+        ];
+
+        let turns = build_turns_from_event_msgs(&events);
+        assert_eq!(turns.len(), 2);
+
+        let first_turn = &turns[0];
+        assert_eq!(first_turn.status, TurnStatus::Interrupted);
+        assert_eq!(first_turn.items.len(), 2);
+        assert_eq!(
+            first_turn.items[0],
+            ThreadItem::UserMessage {
+                id: "item-1".into(),
+                content: vec![UserInput::Text {
+                    text: "Please do the thing".into()
+                }],
+            }
+        );
+        assert_eq!(
+            first_turn.items[1],
+            ThreadItem::AgentMessage {
+                id: "item-2".into(),
+                text: "Working...".into(),
+            }
+        );
+
+        let second_turn = &turns[1];
+        assert_eq!(second_turn.status, TurnStatus::Completed);
+        assert_eq!(second_turn.items.len(), 2);
+        assert_eq!(
+            second_turn.items[0],
+            ThreadItem::UserMessage {
+                id: "item-3".into(),
+                content: vec![UserInput::Text {
+                    text: "Let's try again".into()
+                }],
+            }
+        );
+        assert_eq!(
+            second_turn.items[1],
+            ThreadItem::AgentMessage {
+                id: "item-4".into(),
+                text: "Second attempt complete.".into(),
+            }
+        );
+    }
+}

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 +4,14 @@ 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::CreditsSnapshot as CoreCreditsSnapshot;
 use codex_protocol::protocol::RateLimitSnapshot as CoreRateLimitSnapshot;
 use codex_protocol::protocol::RateLimitWindow as CoreRateLimitWindow;
 use codex_protocol::user_input::UserInput as CoreUserInput;
@@ -20,7 +23,7 @@ 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 +59,23 @@ 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 +83,8 @@ 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 +141,98 @@ 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 +393,7 @@ 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 +403,12 @@ 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 +451,12 @@ 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)]
@@ -392,6 +518,10 @@ pub struct Thread {
     pub created_at: i64,
     /// [UNSTABLE] Path to the thread on disk.
     pub path: PathBuf,
+    /// Only populated on a `thread/resume` response.
+    /// For all other responses and notifications returning a Thread,
+    /// the turns field will be an empty list.
+    pub turns: Vec<Turn>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -406,9 +536,12 @@ pub struct AccountUpdatedNotification {
 #[ts(export_to = "v2/")]
 pub struct Turn {
     pub id: String,
+    /// Only populated on a `thread/resume` response.
+    /// For all other responses and notifications returning a Turn,
+    /// the items field will be an empty list.
     pub items: Vec<ThreadItem>,
+    #[serde(flatten)]
     pub status: TurnStatus,
-    pub error: Option<TurnError>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -419,12 +552,12 @@ 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 +582,45 @@ 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 +678,14 @@ 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 +693,35 @@ 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 +731,18 @@ 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 {
@@ -689,8 +873,6 @@ 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 +940,48 @@ pub struct McpToolCallProgressNotification {
     pub message: String,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct WindowsWorldWritableWarningNotification {
+    pub sample_paths: Vec<String>,
+    pub extra_count: usize,
+    pub failed_scan: bool,
+}
+
+#[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/")]
@@ -771,6 +995,7 @@ pub struct AccountRateLimitsUpdatedNotification {
 pub struct RateLimitSnapshot {
     pub primary: Option<RateLimitWindow>,
     pub secondary: Option<RateLimitWindow>,
+    pub credits: Option<CreditsSnapshot>,
 }
 
 impl From<CoreRateLimitSnapshot> for RateLimitSnapshot {
@@ -778,6 +1003,7 @@ impl From<CoreRateLimitSnapshot> for RateLimitSnapshot {
         Self {
             primary: value.primary.map(RateLimitWindow::from),
             secondary: value.secondary.map(RateLimitWindow::from),
+            credits: value.credits.map(CreditsSnapshot::from),
         }
     }
 }
@@ -801,6 +1027,25 @@ impl From<CoreRateLimitWindow> for RateLimitWindow {
     }
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct CreditsSnapshot {
+    pub has_credits: bool,
+    pub unlimited: bool,
+    pub balance: Option<String>,
+}
+
+impl From<CoreCreditsSnapshot> for CreditsSnapshot {
+    fn from(value: CoreCreditsSnapshot) -> Self {
+        Self {
+            has_credits: value.has_credits,
+            unlimited: value.unlimited,
+            balance: value.balance,
+        }
+    }
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]

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

@@ -17,15 +17,20 @@ 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;
@@ -36,14 +41,17 @@ 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;
@@ -502,10 +510,9 @@ impl CodexClient {
                 ServerNotification::TurnCompleted(payload) => {
                     if payload.turn.id == turn_id {
                         println!("\n< turn/completed notification: {:?}", payload.turn.status);
-                        if let Some(error) = payload.turn.error {
+                        if let TurnStatus::Failed { error } = &payload.turn.status {
                             println!("[turn error] {}", error.message);
                         }
-                        println!("< usage: {:?}", payload.usage);
                         break;
                     }
                 }
@@ -603,8 +610,8 @@ impl CodexClient {
                 JSONRPCMessage::Notification(notification) => {
                     self.pending_notifications.push_back(notification);
                 }
-                JSONRPCMessage::Request(_) => {
-                    bail!("unexpected request from codex app-server");
+                JSONRPCMessage::Request(request) => {
+                    self.handle_server_request(request)?;
                 }
             }
         }
@@ -624,8 +631,8 @@ impl CodexClient {
                     // No outstanding requests, so ignore stray responses/errors for now.
                     continue;
                 }
-                JSONRPCMessage::Request(_) => {
-                    bail!("unexpected request from codex app-server");
+                JSONRPCMessage::Request(request) => {
+                    self.handle_server_request(request)?;
                 }
             }
         }
@@ -661,6 +668,81 @@ impl CodexClient {
     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) {

codex-rs/app-server/Cargo.toml

@@ -40,6 +40,7 @@ tracing = { workspace = true, features = ["log"] }
 tracing-subscriber = { workspace = true, features = ["env-filter", "fmt"] }
 opentelemetry-appender-tracing = { workspace = true }
 uuid = { workspace = true, features = ["serde", "v7"] }
+codex-windows-sandbox.workspace = true
 
 [dev-dependencies]
 app_test_support = { 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

@@ -1,10 +1,18 @@
 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;
@@ -16,25 +24,35 @@ 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 JsonRpcResult = serde_json::Value;
+type JsonValue = serde_json::Value;
 
 pub(crate) async fn apply_bespoke_event_handling(
     event: Event,
@@ -42,9 +60,14 @@ pub(crate) async fn apply_bespoke_event_handling(
     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,
@@ -61,11 +84,57 @@ pub(crate) async fn apply_bespoke_event_handling(
             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::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;
@@ -121,32 +190,6 @@ pub(crate) async fn apply_bespoke_event_handling(
                 ))
                 .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
@@ -158,6 +201,20 @@ pub(crate) async fn apply_bespoke_event_handling(
                     .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 };
@@ -172,6 +229,94 @@ pub(crate) async fn apply_bespoke_event_handling(
                 .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 = {
@@ -194,15 +339,82 @@ pub(crate) async fn apply_bespoke_event_handling(
                     }
                 }
             }
+
+            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<JsonRpcResult>,
+    receiver: oneshot::Receiver<JsonValue>,
     codex: Arc<CodexConversation>,
 ) {
     let response = receiver.await;
@@ -244,7 +456,7 @@ async fn on_patch_approval_response(
 
 async fn on_exec_approval_response(
     event_id: String,
-    receiver: oneshot::Receiver<JsonRpcResult>,
+    receiver: oneshot::Receiver<JsonValue>,
     conversation: Arc<CodexConversation>,
 ) {
     let response = receiver.await;
@@ -278,6 +490,75 @@ async fn on_exec_approval_response(
     }
 }
 
+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,
@@ -287,10 +568,7 @@ async fn construct_mcp_tool_call_notification(
         server: begin_event.invocation.server,
         tool: begin_event.invocation.tool,
         status: McpToolCallStatus::InProgress,
-        arguments: begin_event
-            .invocation
-            .arguments
-            .unwrap_or(JsonRpcResult::Null),
+        arguments: begin_event.invocation.arguments.unwrap_or(JsonValue::Null),
         result: None,
         error: None,
     };
@@ -328,10 +606,7 @@ async fn construct_mcp_tool_call_end_notification(
         server: end_event.invocation.server,
         tool: end_event.invocation.tool,
         status,
-        arguments: end_event
-            .invocation
-            .arguments
-            .unwrap_or(JsonRpcResult::Null),
+        arguments: end_event.invocation.arguments.unwrap_or(JsonValue::Null),
         result,
         error,
     };
@@ -341,13 +616,140 @@ async fn construct_mcp_tool_call_end_notification(
 #[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() {
@@ -377,6 +779,105 @@ mod tests {
         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 {

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

@@ -60,6 +60,8 @@ use codex_app_server_protocol::RemoveConversationSubscriptionResponse;
 use codex_app_server_protocol::RequestId;
 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;
@@ -89,6 +91,8 @@ use codex_app_server_protocol::TurnStatus;
 use codex_app_server_protocol::UserInfoResponse;
 use codex_app_server_protocol::UserInput as V2UserInput;
 use codex_app_server_protocol::UserSavedConfig;
+use codex_app_server_protocol::WindowsWorldWritableWarningNotification;
+use codex_app_server_protocol::build_turns_from_event_msgs;
 use codex_backend_client::Client as BackendClient;
 use codex_core::AuthManager;
 use codex_core::CodexConversation;
@@ -109,12 +113,15 @@ use codex_core::config_loader::load_config_as_toml;
 use codex_core::default_client::get_codex_user_agent;
 use codex_core::exec::ExecParams;
 use codex_core::exec_env::create_env;
+use codex_core::features::Feature;
 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::EventMsg;
 use codex_core::protocol::Op;
+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;
@@ -151,6 +158,14 @@ use uuid::Uuid;
 type PendingInterruptQueue = Vec<(RequestId, ApiVersion)>;
 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);
 struct ActiveLogin {
@@ -158,8 +173,8 @@ struct ActiveLogin {
     login_id: Uuid,
 }
 
-impl ActiveLogin {
-    fn drop(&self) {
+impl Drop for ActiveLogin {
+    fn drop(&mut self) {
         self.shutdown_handle.shutdown();
     }
 }
@@ -175,6 +190,7 @@ 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,
 }
@@ -227,11 +243,97 @@ 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 { .. } => {
@@ -263,6 +365,9 @@ 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
@@ -417,7 +522,7 @@ impl CodexMessageProcessor {
         {
             let mut guard = self.active_login.lock().await;
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
         }
 
@@ -525,7 +630,7 @@ 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(),
@@ -615,7 +720,7 @@ 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(),
@@ -704,7 +809,7 @@ 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 {
@@ -758,7 +863,7 @@ impl CodexMessageProcessor {
         {
             let mut guard = self.active_login.lock().await;
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
         }
 
@@ -1135,7 +1240,7 @@ impl CodexMessageProcessor {
         let overrides = ConfigOverrides {
             model,
             config_profile: profile,
-            cwd: cwd.map(PathBuf::from),
+            cwd: cwd.clone().map(PathBuf::from),
             approval_policy,
             sandbox_mode,
             model_provider,
@@ -1147,7 +1252,17 @@ impl CodexMessageProcessor {
             ..Default::default()
         };
 
-        let config = match derive_config_from_params(overrides, cli_overrides).await {
+        // Persist windows sandbox feature.
+        // TODO: persist default config in general.
+        let mut cli_overrides = cli_overrides.unwrap_or_default();
+        if cfg!(windows) && self.config.features.enabled(Feature::WindowsSandbox) {
+            cli_overrides.insert(
+                "features.enable_experimental_windows_sandbox".to_string(),
+                serde_json::json!(true),
+            );
+        }
+
+        let config = match derive_config_from_params(overrides, Some(cli_overrides)).await {
             Ok(config) => config,
             Err(err) => {
                 let error = JSONRPCErrorError {
@@ -1159,6 +1274,10 @@ impl CodexMessageProcessor {
                 return;
             }
         };
+        if cfg!(windows) && config.features.enabled(Feature::WindowsSandbox) {
+            self.handle_windows_world_writable_warning(config.cwd.clone())
+                .await;
+        }
 
         match self.conversation_manager.new_conversation(config).await {
             Ok(conversation_id) => {
@@ -1212,8 +1331,12 @@ 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
@@ -1238,14 +1361,28 @@ 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!(
@@ -1521,9 +1658,14 @@ impl CodexMessageProcessor {
                 session_configured,
                 ..
             }) => {
+                let SessionConfiguredEvent {
+                    rollout_path,
+                    initial_messages,
+                    ..
+                } = session_configured;
                 // 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!(
@@ -1533,8 +1675,8 @@ impl CodexMessageProcessor {
                     );
                 }
 
-                let thread = match read_summary_from_rollout(
-                    session_configured.rollout_path.as_path(),
+                let mut thread = match read_summary_from_rollout(
+                    rollout_path.as_path(),
                     fallback_model_provider.as_str(),
                 )
                 .await
@@ -1545,14 +1687,27 @@ impl CodexMessageProcessor {
                             request_id,
                             format!(
                                 "failed to load rollout `{}` for conversation {conversation_id}: {err}",
-                                session_configured.rollout_path.display()
+                                rollout_path.display()
                             ),
                         )
                         .await;
                         return;
                     }
                 };
-                let response = ThreadResumeResponse { thread };
+                thread.turns = initial_messages
+                    .as_deref()
+                    .map_or_else(Vec::new, build_turns_from_event_msgs);
+
+                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) => {
@@ -1803,6 +1958,15 @@ impl CodexMessageProcessor {
                     include_apply_patch_tool,
                 } = overrides;
 
+                // Persist windows sandbox feature.
+                let mut cli_overrides = cli_overrides.unwrap_or_default();
+                if cfg!(windows) && self.config.features.enabled(Feature::WindowsSandbox) {
+                    cli_overrides.insert(
+                        "features.enable_experimental_windows_sandbox".to_string(),
+                        serde_json::json!(true),
+                    );
+                }
+
                 let overrides = ConfigOverrides {
                     model,
                     config_profile: profile,
@@ -1818,7 +1982,7 @@ impl CodexMessageProcessor {
                     ..Default::default()
                 };
 
-                derive_config_from_params(overrides, cli_overrides).await
+                derive_config_from_params(overrides, Some(cli_overrides)).await
             }
             None => Ok(self.config.as_ref().clone()),
         };
@@ -1833,6 +1997,10 @@ impl CodexMessageProcessor {
                 return;
             }
         };
+        if cfg!(windows) && config.features.enabled(Feature::WindowsSandbox) {
+            self.handle_windows_world_writable_warning(config.cwd.clone())
+                .await;
+        }
 
         let conversation_history = if let Some(path) = path {
             match RolloutRecorder::get_rollout_history(&path).await {
@@ -2272,9 +2440,6 @@ 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
@@ -2314,12 +2479,8 @@ 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() };
@@ -2342,6 +2503,64 @@ 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;
 
@@ -2376,7 +2595,7 @@ 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) => {
@@ -2417,6 +2636,7 @@ impl CodexMessageProcessor {
         &mut self,
         conversation_id: ConversationId,
         experimental_raw_events: bool,
+        api_version: ApiVersion,
     ) -> Result<Uuid, JSONRPCErrorError> {
         let conversation = match self
             .conversation_manager
@@ -2440,6 +2660,8 @@ 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! {
@@ -2495,6 +2717,8 @@ impl CodexMessageProcessor {
                             conversation.clone(),
                             outgoing_for_task.clone(),
                             pending_interrupts.clone(),
+                            turn_summary_store.clone(),
+                            api_version_for_task,
                         )
                         .await;
                     }
@@ -2635,6 +2859,53 @@ impl CodexMessageProcessor {
             Err(_) => None,
         }
     }
+
+    /// On Windows, when using the experimental sandbox, we need to warn the user about world-writable directories.
+    async fn handle_windows_world_writable_warning(&self, cwd: PathBuf) {
+        if !cfg!(windows) {
+            return;
+        }
+
+        if !self.config.features.enabled(Feature::WindowsSandbox) {
+            return;
+        }
+
+        if !matches!(
+            self.config.sandbox_policy,
+            codex_protocol::protocol::SandboxPolicy::WorkspaceWrite { .. }
+                | codex_protocol::protocol::SandboxPolicy::ReadOnly
+        ) {
+            return;
+        }
+
+        if self
+            .config
+            .notices
+            .hide_world_writable_warning
+            .unwrap_or(false)
+        {
+            return;
+        }
+
+        // This function is stubbed out to return None on non-Windows platforms
+        if let Some((sample_paths, extra_count, failed_scan)) =
+            codex_windows_sandbox::world_writable_warning_details(
+                self.config.codex_home.as_path(),
+                cwd,
+            )
+        {
+            tracing::warn!("world writable warning: {sample_paths:?} {extra_count} {failed_scan}");
+            self.outgoing
+                .send_server_notification(ServerNotification::WindowsWorldWritableWarning(
+                    WindowsWorldWritableWarningNotification {
+                        sample_paths,
+                        extra_count,
+                        failed_scan,
+                    },
+                ))
+                .await;
+        }
+    }
 }
 
 async fn derive_config_from_params(
@@ -2788,6 +3059,7 @@ fn summary_to_thread(summary: ConversationSummary) -> Thread {
         model_provider,
         created_at: created_at.map(|dt| dt.timestamp()).unwrap_or(0),
         path,
+        turns: Vec::new(),
     }
 }
 

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

@@ -47,7 +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/message_processor.rs

@@ -6,7 +6,6 @@ use crate::outgoing_message::OutgoingMessageSender;
 use codex_app_server_protocol::ClientInfo;
 use codex_app_server_protocol::ClientRequest;
 use codex_app_server_protocol::InitializeResponse;
-
 use codex_app_server_protocol::JSONRPCError;
 use codex_app_server_protocol::JSONRPCErrorError;
 use codex_app_server_protocol::JSONRPCNotification;
@@ -118,6 +117,7 @@ impl MessageProcessor {
                     self.outgoing.send_response(request_id, response).await;
 
                     self.initialized = true;
+
                     return;
                 }
             }

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:?}");
+        }
     }
 }
 
@@ -215,6 +229,7 @@ mod tests {
                         resets_at: Some(123),
                     }),
                     secondary: None,
+                    credits: None,
                 },
             });
 
@@ -229,7 +244,8 @@ mod tests {
                             "windowDurationMins": 15,
                             "resetsAt": 123
                         },
-                        "secondary": null
+                        "secondary": null,
+                        "credits": null
                     }
                 },
             }),

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-max"
 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-max".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-max"
 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/model_list.rs

@@ -45,6 +45,33 @@ async fn list_models_returns_all_models_with_large_limit() -> Result<()> {
     } = to_response::<ModelListResponse>(response)?;
 
     let expected_models = vec![
+        Model {
+            id: "gpt-5.1-codex-max".to_string(),
+            model: "gpt-5.1-codex-max".to_string(),
+            display_name: "gpt-5.1-codex-max".to_string(),
+            description: "Latest Codex-optimized flagship for deep and fast reasoning.".to_string(),
+            supported_reasoning_efforts: vec![
+                ReasoningEffortOption {
+                    reasoning_effort: ReasoningEffort::Low,
+                    description: "Fast responses with lighter reasoning".to_string(),
+                },
+                ReasoningEffortOption {
+                    reasoning_effort: ReasoningEffort::Medium,
+                    description: "Balances speed and reasoning depth for everyday tasks"
+                        .to_string(),
+                },
+                ReasoningEffortOption {
+                    reasoning_effort: ReasoningEffort::High,
+                    description: "Maximizes reasoning depth for complex problems".to_string(),
+                },
+                ReasoningEffortOption {
+                    reasoning_effort: ReasoningEffort::XHigh,
+                    description: "Extra high reasoning depth for complex problems".to_string(),
+                },
+            ],
+            default_reasoning_effort: ReasoningEffort::Medium,
+            is_default: true,
+        },
         Model {
             id: "gpt-5.1-codex".to_string(),
             model: "gpt-5.1-codex".to_string(),
@@ -66,7 +93,7 @@ async fn list_models_returns_all_models_with_large_limit() -> Result<()> {
                 },
             ],
             default_reasoning_effort: ReasoningEffort::Medium,
-            is_default: true,
+            is_default: false,
         },
         Model {
             id: "gpt-5.1-codex-mini".to_string(),
@@ -147,7 +174,7 @@ async fn list_models_pagination_works() -> Result<()> {
     } = to_response::<ModelListResponse>(first_response)?;
 
     assert_eq!(first_items.len(), 1);
-    assert_eq!(first_items[0].id, "gpt-5.1-codex");
+    assert_eq!(first_items[0].id, "gpt-5.1-codex-max");
     let next_cursor = first_cursor.ok_or_else(|| anyhow!("cursor for second page"))?;
 
     let second_request = mcp
@@ -169,7 +196,7 @@ async fn list_models_pagination_works() -> Result<()> {
     } = to_response::<ModelListResponse>(second_response)?;
 
     assert_eq!(second_items.len(), 1);
-    assert_eq!(second_items[0].id, "gpt-5.1-codex-mini");
+    assert_eq!(second_items[0].id, "gpt-5.1-codex");
     let third_cursor = second_cursor.ok_or_else(|| anyhow!("cursor for third page"))?;
 
     let third_request = mcp
@@ -191,8 +218,30 @@ async fn list_models_pagination_works() -> Result<()> {
     } = to_response::<ModelListResponse>(third_response)?;
 
     assert_eq!(third_items.len(), 1);
-    assert_eq!(third_items[0].id, "gpt-5.1");
-    assert!(third_cursor.is_none());
+    assert_eq!(third_items[0].id, "gpt-5.1-codex-mini");
+    let fourth_cursor = third_cursor.ok_or_else(|| anyhow!("cursor for fourth page"))?;
+
+    let fourth_request = mcp
+        .send_list_models_request(ModelListParams {
+            limit: Some(1),
+            cursor: Some(fourth_cursor.clone()),
+        })
+        .await?;
+
+    let fourth_response: JSONRPCResponse = timeout(
+        DEFAULT_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(fourth_request)),
+    )
+    .await??;
+
+    let ModelListResponse {
+        data: fourth_items,
+        next_cursor: fourth_cursor,
+    } = to_response::<ModelListResponse>(fourth_response)?;
+
+    assert_eq!(fourth_items.len(), 1);
+    assert_eq!(fourth_items[0].id, "gpt-5.1");
+    assert!(fourth_cursor.is_none());
     Ok(())
 }
 

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

@@ -152,6 +152,7 @@ async fn get_account_rate_limits_returns_snapshot() -> Result<()> {
                 window_duration_mins: Some(1440),
                 resets_at: Some(secondary_reset_timestamp),
             }),
+            credits: None,
         },
     };
     assert_eq!(received, expected);

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

@@ -1,13 +1,17 @@
 use anyhow::Result;
 use app_test_support::McpProcess;
+use app_test_support::create_fake_rollout;
 use app_test_support::create_mock_chat_completions_server;
 use app_test_support::to_response;
 use codex_app_server_protocol::JSONRPCResponse;
 use codex_app_server_protocol::RequestId;
+use codex_app_server_protocol::ThreadItem;
 use codex_app_server_protocol::ThreadResumeParams;
 use codex_app_server_protocol::ThreadResumeResponse;
 use codex_app_server_protocol::ThreadStartParams;
 use codex_app_server_protocol::ThreadStartResponse;
+use codex_app_server_protocol::TurnStatus;
+use codex_app_server_protocol::UserInput;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::ResponseItem;
 use tempfile::TempDir;
@@ -27,7 +31,7 @@ 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-max".to_string()),
             ..Default::default()
         })
         .await?;
@@ -36,7 +40,7 @@ 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
@@ -50,13 +54,73 @@ 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(())
 }
 
+#[tokio::test]
+async fn thread_resume_returns_rollout_history() -> Result<()> {
+    let server = create_mock_chat_completions_server(vec![]).await;
+    let codex_home = TempDir::new()?;
+    create_config_toml(codex_home.path(), &server.uri())?;
+
+    let preview = "Saved user message";
+    let conversation_id = create_fake_rollout(
+        codex_home.path(),
+        "2025-01-05T12-00-00",
+        "2025-01-05T12:00:00Z",
+        preview,
+        Some("mock_provider"),
+    )?;
+
+    let mut mcp = McpProcess::new(codex_home.path()).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+
+    let resume_id = mcp
+        .send_thread_resume_request(ThreadResumeParams {
+            thread_id: conversation_id.clone(),
+            ..Default::default()
+        })
+        .await?;
+    let resume_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(resume_id)),
+    )
+    .await??;
+    let ThreadResumeResponse { thread, .. } = to_response::<ThreadResumeResponse>(resume_resp)?;
+
+    assert_eq!(thread.id, conversation_id);
+    assert_eq!(thread.preview, preview);
+    assert_eq!(thread.model_provider, "mock_provider");
+    assert!(thread.path.is_absolute());
+
+    assert_eq!(
+        thread.turns.len(),
+        1,
+        "expected rollouts to include one turn"
+    );
+    let turn = &thread.turns[0];
+    assert_eq!(turn.status, TurnStatus::Completed);
+    assert_eq!(turn.items.len(), 1, "expected user message item");
+    match &turn.items[0] {
+        ThreadItem::UserMessage { content, .. } => {
+            assert_eq!(
+                content,
+                &vec![UserInput::Text {
+                    text: preview.to_string()
+                }]
+            );
+        }
+        other => panic!("expected user message item, got {other:?}"),
+    }
+
+    Ok(())
+}
+
 #[tokio::test]
 async fn thread_resume_prefers_path_over_thread_id() -> Result<()> {
     let server = create_mock_chat_completions_server(vec![]).await;
@@ -68,7 +132,7 @@ 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-max".to_string()),
             ..Default::default()
         })
         .await?;
@@ -77,7 +141,7 @@ 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 +157,9 @@ 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 +177,7 @@ 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-max".to_string()),
             ..Default::default()
         })
         .await?;
@@ -121,7 +186,7 @@ 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 +212,13 @@ 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/backend-client/src/client.rs

@@ -1,4 +1,5 @@
 use crate::types::CodeTaskDetailsResponse;
+use crate::types::CreditStatusDetails;
 use crate::types::PaginatedListTaskListItem;
 use crate::types::RateLimitStatusPayload;
 use crate::types::RateLimitWindowSnapshot;
@@ -6,6 +7,7 @@ use crate::types::TurnAttemptsSiblingTurnsResponse;
 use anyhow::Result;
 use codex_core::auth::CodexAuth;
 use codex_core::default_client::get_codex_user_agent;
+use codex_protocol::protocol::CreditsSnapshot;
 use codex_protocol::protocol::RateLimitSnapshot;
 use codex_protocol::protocol::RateLimitWindow;
 use reqwest::header::AUTHORIZATION;
@@ -272,19 +274,23 @@ impl Client {
 
     // rate limit helpers
     fn rate_limit_snapshot_from_payload(payload: RateLimitStatusPayload) -> RateLimitSnapshot {
-        let Some(details) = payload
+        let rate_limit_details = payload
             .rate_limit
-            .and_then(|inner| inner.map(|boxed| *boxed))
-        else {
-            return RateLimitSnapshot {
-                primary: None,
-                secondary: None,
-            };
+            .and_then(|inner| inner.map(|boxed| *boxed));
+
+        let (primary, secondary) = if let Some(details) = rate_limit_details {
+            (
+                Self::map_rate_limit_window(details.primary_window),
+                Self::map_rate_limit_window(details.secondary_window),
+            )
+        } else {
+            (None, None)
         };
 
         RateLimitSnapshot {
-            primary: Self::map_rate_limit_window(details.primary_window),
-            secondary: Self::map_rate_limit_window(details.secondary_window),
+            primary,
+            secondary,
+            credits: Self::map_credits(payload.credits),
         }
     }
 
@@ -306,6 +312,19 @@ impl Client {
         })
     }
 
+    fn map_credits(credits: Option<Option<Box<CreditStatusDetails>>>) -> Option<CreditsSnapshot> {
+        let details = match credits {
+            Some(Some(details)) => *details,
+            _ => return None,
+        };
+
+        Some(CreditsSnapshot {
+            has_credits: details.has_credits,
+            unlimited: details.unlimited,
+            balance: details.balance.and_then(|inner| inner),
+        })
+    }
+
     fn window_minutes_from_seconds(seconds: i32) -> Option<i64> {
         if seconds <= 0 {
             return None;

codex-rs/backend-client/src/types.rs

@@ -1,3 +1,4 @@
+pub use codex_backend_openapi_models::models::CreditStatusDetails;
 pub use codex_backend_openapi_models::models::PaginatedListTaskListItem;
 pub use codex_backend_openapi_models::models::PlanType;
 pub use codex_backend_openapi_models::models::RateLimitStatusDetails;

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/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/model_presets.rs

@@ -4,6 +4,10 @@ use codex_app_server_protocol::AuthMode;
 use codex_core::protocol_config_types::ReasoningEffort;
 use once_cell::sync::Lazy;
 
+pub const HIDE_GPT5_1_MIGRATION_PROMPT_CONFIG: &str = "hide_gpt5_1_migration_prompt";
+pub const HIDE_GPT_5_1_CODEX_MAX_MIGRATION_PROMPT_CONFIG: &str =
+    "hide_gpt-5.1-codex-max_migration_prompt";
+
 /// A reasoning effort option that can be surfaced for a model.
 #[derive(Debug, Clone, Copy)]
 pub struct ReasoningEffortPreset {
@@ -17,6 +21,7 @@ pub struct ReasoningEffortPreset {
 pub struct ModelUpgrade {
     pub id: &'static str,
     pub reasoning_effort_mapping: Option<HashMap<ReasoningEffort, ReasoningEffort>>,
+    pub migration_config_key: &'static str,
 }
 
 /// Metadata describing a Codex-supported model.
@@ -38,10 +43,40 @@ pub struct ModelPreset {
     pub is_default: bool,
     /// recommended upgrade model
     pub upgrade: Option<ModelUpgrade>,
+    /// Whether this preset should appear in the picker UI.
+    pub show_in_picker: bool,
 }
 
 static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
     vec![
+        ModelPreset {
+            id: "gpt-5.1-codex-max",
+            model: "gpt-5.1-codex-max",
+            display_name: "gpt-5.1-codex-max",
+            description: "Latest Codex-optimized flagship for deep and fast reasoning.",
+            default_reasoning_effort: ReasoningEffort::Medium,
+            supported_reasoning_efforts: &[
+                ReasoningEffortPreset {
+                    effort: ReasoningEffort::Low,
+                    description: "Fast responses with lighter reasoning",
+                },
+                ReasoningEffortPreset {
+                    effort: ReasoningEffort::Medium,
+                    description: "Balances speed and reasoning depth for everyday tasks",
+                },
+                ReasoningEffortPreset {
+                    effort: ReasoningEffort::High,
+                    description: "Maximizes reasoning depth for complex problems",
+                },
+                ReasoningEffortPreset {
+                    effort: ReasoningEffort::XHigh,
+                    description: "Extra high reasoning depth for complex problems",
+                },
+            ],
+            is_default: true,
+            upgrade: None,
+            show_in_picker: true,
+        },
         ModelPreset {
             id: "gpt-5.1-codex",
             model: "gpt-5.1-codex",
@@ -62,8 +97,13 @@ static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
                     description: "Maximizes reasoning depth for complex or ambiguous problems",
                 },
             ],
-            is_default: true,
-            upgrade: None,
+            is_default: false,
+            upgrade: Some(ModelUpgrade {
+                id: "gpt-5.1-codex-max",
+                reasoning_effort_mapping: None,
+                migration_config_key: HIDE_GPT_5_1_CODEX_MAX_MIGRATION_PROMPT_CONFIG,
+            }),
+            show_in_picker: true,
         },
         ModelPreset {
             id: "gpt-5.1-codex-mini",
@@ -82,7 +122,12 @@ static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
                 },
             ],
             is_default: false,
-            upgrade: None,
+            upgrade: Some(ModelUpgrade {
+                id: "gpt-5.1-codex-max",
+                reasoning_effort_mapping: None,
+                migration_config_key: HIDE_GPT_5_1_CODEX_MAX_MIGRATION_PROMPT_CONFIG,
+            }),
+            show_in_picker: true,
         },
         ModelPreset {
             id: "gpt-5.1",
@@ -105,7 +150,12 @@ static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
                 },
             ],
             is_default: false,
-            upgrade: None,
+            upgrade: Some(ModelUpgrade {
+                id: "gpt-5.1-codex-max",
+                reasoning_effort_mapping: None,
+                migration_config_key: HIDE_GPT_5_1_CODEX_MAX_MIGRATION_PROMPT_CONFIG,
+            }),
+            show_in_picker: true,
         },
         // Deprecated models.
         ModelPreset {
@@ -130,9 +180,11 @@ static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
             ],
             is_default: false,
             upgrade: Some(ModelUpgrade {
-                id: "gpt-5.1-codex",
+                id: "gpt-5.1-codex-max",
                 reasoning_effort_mapping: None,
+                migration_config_key: HIDE_GPT_5_1_CODEX_MAX_MIGRATION_PROMPT_CONFIG,
             }),
+            show_in_picker: false,
         },
         ModelPreset {
             id: "gpt-5-codex-mini",
@@ -154,7 +206,9 @@ static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
             upgrade: Some(ModelUpgrade {
                 id: "gpt-5.1-codex-mini",
                 reasoning_effort_mapping: None,
+                migration_config_key: HIDE_GPT5_1_MIGRATION_PROMPT_CONFIG,
             }),
+            show_in_picker: false,
         },
         ModelPreset {
             id: "gpt-5",
@@ -182,21 +236,22 @@ static PRESETS: Lazy<Vec<ModelPreset>> = Lazy::new(|| {
             ],
             is_default: false,
             upgrade: Some(ModelUpgrade {
-                id: "gpt-5.1",
-                reasoning_effort_mapping: Some(HashMap::from([(
-                    ReasoningEffort::Minimal,
-                    ReasoningEffort::Low,
-                )])),
+                id: "gpt-5.1-codex-max",
+                reasoning_effort_mapping: None,
+                migration_config_key: HIDE_GPT_5_1_CODEX_MAX_MIGRATION_PROMPT_CONFIG,
             }),
+            show_in_picker: false,
         },
     ]
 });
 
-pub fn builtin_model_presets(_auth_mode: Option<AuthMode>) -> Vec<ModelPreset> {
-    // leave auth mode for later use
+pub fn builtin_model_presets(auth_mode: Option<AuthMode>) -> Vec<ModelPreset> {
     PRESETS
         .iter()
-        .filter(|preset| preset.upgrade.is_none())
+        .filter(|preset| match auth_mode {
+            Some(AuthMode::ApiKey) => preset.show_in_picker && preset.id != "gpt-5.1-codex-max",
+            _ => preset.show_in_picker,
+        })
         .cloned()
         .collect()
 }
@@ -208,10 +263,21 @@ pub fn all_model_presets() -> &'static Vec<ModelPreset> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use codex_app_server_protocol::AuthMode;
 
     #[test]
     fn only_one_default_model_is_configured() {
         let default_models = PRESETS.iter().filter(|preset| preset.is_default).count();
         assert!(default_models == 1);
     }
+
+    #[test]
+    fn gpt_5_1_codex_max_hidden_for_api_key_auth() {
+        let presets = builtin_model_presets(Some(AuthMode::ApiKey));
+        assert!(
+            presets
+                .iter()
+                .all(|preset| preset.id != "gpt-5.1-codex-max")
+        );
+    }
 }

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

@@ -40,12 +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 }
@@ -90,9 +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]
@@ -102,6 +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-codex-max_prompt.md

@@ -0,0 +1,117 @@
+You are Codex, based on GPT-5. You are running as a coding agent in the Codex CLI on a user's computer.
+
+## General
+
+- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
+
+## Editing constraints
+
+- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.
+- Add succinct code comments that explain what is going on if code is not self-explanatory. You should not add comments like "Assigns the value to the variable", but a brief comment might be useful ahead of a complex code block that the user would otherwise have to spend time parsing out. Usage of these comments should be rare.
+- Try to use apply_patch for single file edits, but it is fine to explore other options to make the edit if it does not work well. Do not use apply_patch for changes that are auto-generated (i.e. generating package.json or running a lint or format command like gofmt) or when scripting is more efficient (such as search and replacing a string across a codebase).
+- You may be in a dirty git worktree.
+    * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.
+    * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.
+    * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.
+    * If the changes are in unrelated files, just ignore them and don't revert them.
+- Do not amend a commit unless explicitly requested to do so.
+- While you are working, you might notice unexpected changes that you didn't make. If this happens, STOP IMMEDIATELY and ask the user how they would like to proceed.
+- **NEVER** use destructive commands like `git reset --hard` or `git checkout --` unless specifically requested or approved by the user.
+
+## Plan tool
+
+When using the planning tool:
+- Skip using the planning tool for straightforward tasks (roughly the easiest 25%).
+- Do not make single-step plans.
+- When you made a plan, update it after having performed one of the sub-tasks that you shared on the plan.
+
+## Codex CLI harness, sandboxing, and approvals
+
+The Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.
+
+Filesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:
+- **read-only**: The sandbox only permits reading files.
+- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.
+- **danger-full-access**: No filesystem sandboxing - all commands are permitted.
+
+Network sandboxing defines whether network can be accessed without approval. Options for `network_access` are:
+- **restricted**: Requires approval
+- **enabled**: No approval needed
+
+Approvals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` are
+- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe "read" commands.
+- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.
+- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for it in the `shell` command description.)
+- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is paired with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.
+
+When you are running with `approval_policy == on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:
+- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /var)
+- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.
+- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)
+- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval. ALWAYS proceed to use the `with_escalated_permissions` and `justification` parameters - do not message the user before requesting approval for the command.
+- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for
+- (for all of these, you should weigh alternative paths that do not require approval)
+
+When `sandbox_mode` is set to read-only, you'll need to request approval for any command that isn't a read.
+
+You will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing enabled, and approval on-failure.
+
+Although they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to "never", in which case never ask for approvals.
+
+When requesting approval to execute a command that will require escalated privileges:
+  - Provide the `with_escalated_permissions` parameter with the boolean value true
+  - Include a short, 1 sentence explanation for why you need to enable `with_escalated_permissions` in the justification parameter
+
+## Special user requests
+
+- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as `date`), you should do so.
+- If the user asks for a "review", default to a code review mindset: prioritise identifying bugs, risks, behavioural regressions, and missing tests. Findings must be the primary focus of the response - keep summaries or overviews brief and only after enumerating the issues. Present findings first (ordered by severity with file/line references), follow with open questions or assumptions, and offer a change-summary only as a secondary detail. If no findings are discovered, state that explicitly and mention any residual risks or testing gaps.
+
+## Frontend tasks
+When doing frontend design tasks, avoid collapsing into "AI slop" or safe, average-looking layouts.
+Aim for interfaces that feel intentional, bold, and a bit surprising.
+- Typography: Use expressive, purposeful fonts and avoid default stacks (Inter, Roboto, Arial, system).
+- Color & Look: Choose a clear visual direction; define CSS variables; avoid purple-on-white defaults. No purple bias or dark mode bias.
+- Motion: Use a few meaningful animations (page-load, staggered reveals) instead of generic micro-motions.
+- Background: Don't rely on flat, single-color backgrounds; use gradients, shapes, or subtle patterns to build atmosphere.
+- Overall: Avoid boilerplate layouts and interchangeable UI patterns. Vary themes, type families, and visual languages across outputs.
+- Ensure the page loads properly on both desktop and mobile
+
+Exception: If working within an existing website or design system, preserve the established patterns, structure, and visual language.
+
+## Presenting your work and final message
+
+You are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.
+
+- Default: be very concise; friendly coding teammate tone.
+- Ask only when needed; suggest ideas; mirror the user's style.
+- For substantial work, summarize clearly; follow final‑answer formatting.
+- Skip heavy formatting for simple confirmations.
+- Don't dump large files you've written; reference paths only.
+- No "save/copy this file" - User is on the same machine.
+- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.
+- For code changes:
+  * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with "summary", just jump right in.
+  * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.
+  * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.
+- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.
+
+### Final answer structure and style guidelines
+
+- Plain text; CLI handles styling. Use structure only when it helps scanability.
+- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.
+- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.
+- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.
+- Code samples or multi-line snippets should be wrapped in fenced code blocks; include an info string as often as possible.
+- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.
+- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no "above/below"; parallel wording.
+- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.
+- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.
+- File References: When referencing files in your response follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Optionally include line/column (1‑based): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5

codex-rs/core/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;
 
@@ -55,6 +56,7 @@ use crate::model_family::ModelFamily;
 use crate::model_provider_info::ModelProviderInfo;
 use crate::model_provider_info::WireApi;
 use crate::openai_model_info::get_model_info;
+use crate::protocol::CreditsSnapshot;
 use crate::protocol::RateLimitSnapshot;
 use crate::protocol::RateLimitWindow;
 use crate::protocol::TokenUsage;
@@ -78,6 +80,18 @@ 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 +521,70 @@ 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 {
@@ -649,7 +727,13 @@ fn parse_rate_limit_snapshot(headers: &HeaderMap) -> Option<RateLimitSnapshot> {
         "x-codex-secondary-reset-at",
     );
 
-    Some(RateLimitSnapshot { primary, secondary })
+    let credits = parse_credits_snapshot(headers);
+
+    Some(RateLimitSnapshot {
+        primary,
+        secondary,
+        credits,
+    })
 }
 
 fn parse_rate_limit_window(
@@ -676,6 +760,20 @@ fn parse_rate_limit_window(
     })
 }
 
+fn parse_credits_snapshot(headers: &HeaderMap) -> Option<CreditsSnapshot> {
+    let has_credits = parse_header_bool(headers, "x-codex-credits-has-credits")?;
+    let unlimited = parse_header_bool(headers, "x-codex-credits-unlimited")?;
+    let balance = parse_header_str(headers, "x-codex-credits-balance")
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .map(std::string::ToString::to_string);
+    Some(CreditsSnapshot {
+        has_credits,
+        unlimited,
+        balance,
+    })
+}
+
 fn parse_header_f64(headers: &HeaderMap, name: &str) -> Option<f64> {
     parse_header_str(headers, name)?
         .parse::<f64>()
@@ -687,6 +785,17 @@ fn parse_header_i64(headers: &HeaderMap, name: &str) -> Option<i64> {
     parse_header_str(headers, name)?.parse::<i64>().ok()
 }
 
+fn parse_header_bool(headers: &HeaderMap, name: &str) -> Option<bool> {
+    let raw = parse_header_str(headers, name)?;
+    if raw.eq_ignore_ascii_case("true") || raw == "1" {
+        Some(true)
+    } else if raw.eq_ignore_ascii_case("false") || raw == "0" {
+        Some(false)
+    } else {
+        None
+    }
+}
+
 fn parse_header_str<'a>(headers: &'a HeaderMap, name: &str) -> Option<&'a str> {
     headers.get(name)?.to_str().ok()
 }
@@ -1225,7 +1334,7 @@ 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 +1363,7 @@ 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 +1582,7 @@ 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 +1596,7 @@ 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

@@ -165,11 +165,9 @@ fn build_structured_output(parsed: &ExecOutputJson) -> String {
     ));
 
     let mut output = parsed.output.clone();
-    if let Some(total_lines) = extract_total_output_lines(&parsed.output) {
+    if let Some((stripped, total_lines)) = strip_total_output_header(&parsed.output) {
         sections.push(format!("Total output lines: {total_lines}"));
-        if let Some(stripped) = strip_total_output_header(&output) {
-            output = stripped.to_string();
-        }
+        output = stripped.to_string();
     }
 
     sections.push("Output:".to_string());
@@ -178,19 +176,12 @@ fn build_structured_output(parsed: &ExecOutputJson) -> String {
     sections.join("\n")
 }
 
-fn extract_total_output_lines(output: &str) -> Option<u32> {
-    let marker_start = output.find("[... omitted ")?;
-    let marker = &output[marker_start..];
-    let (_, after_of) = marker.split_once(" of ")?;
-    let (total_segment, _) = after_of.split_once(' ')?;
-    total_segment.parse::<u32>().ok()
-}
-
-fn strip_total_output_header(output: &str) -> Option<&str> {
+fn strip_total_output_header(output: &str) -> Option<(&str, u32)> {
     let after_prefix = output.strip_prefix("Total output lines: ")?;
-    let (_, remainder) = after_prefix.split_once('\n')?;
+    let (total_segment, remainder) = after_prefix.split_once('\n')?;
+    let total_lines = total_segment.parse::<u32>().ok()?;
     let remainder = remainder.strip_prefix('\n').unwrap_or(remainder);
-    Some(remainder)
+    Some((remainder, total_lines))
 }
 
 #[derive(Debug)]
@@ -427,11 +418,11 @@ mod tests {
                 expects_apply_patch_instructions: false,
             },
             InstructionsTestCase {
-                slug: "gpt-5-codex",
+                slug: "gpt-5.1-codex",
                 expects_apply_patch_instructions: false,
             },
             InstructionsTestCase {
-                slug: "gpt-5.1-codex",
+                slug: "gpt-5.1-codex-max",
                 expects_apply_patch_instructions: false,
             },
         ];
@@ -457,7 +448,7 @@ 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 +489,7 @@ 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 +525,7 @@ 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

@@ -7,14 +7,16 @@ use std::sync::atomic::AtomicU64;
 use crate::AuthManager;
 use crate::client_common::REVIEW_PROMPT;
 use crate::compact;
+use crate::compact::run_inline_auto_compact_task;
+use crate::compact::should_use_remote_compact_task;
+use crate::compact_remote::run_inline_remote_auto_compact_task;
 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 +47,7 @@ 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;
@@ -56,19 +59,16 @@ 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;
 use crate::error::CodexErr;
 use crate::error::Result as CodexResult;
+use crate::error::http_status_code_value;
 #[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;
@@ -80,7 +80,6 @@ use crate::protocol::ApplyPatchApprovalRequestEvent;
 use crate::protocol::AskForApproval;
 use crate::protocol::BackgroundEventEvent;
 use crate::protocol::DeprecationNoticeEvent;
-use crate::protocol::ErrorEvent;
 use crate::protocol::Event;
 use crate::protocol::EventMsg;
 use crate::protocol::ExecApprovalRequestEvent;
@@ -134,6 +133,7 @@ use codex_protocol::user_input::UserInput;
 use codex_utils_readiness::Readiness;
 use codex_utils_readiness::ReadinessFlag;
 use codex_utils_tokenizer::warm_model_cache;
+use reqwest::StatusCode;
 
 /// The high-level interface to the Codex system.
 /// It operates as a queue pair where you send submissions and receive events.
@@ -280,6 +280,7 @@ 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) truncation_policy: TruncationPolicy,
 }
 
 impl TurnContext {
@@ -296,7 +297,6 @@ impl TurnContext {
     }
 }
 
-#[allow(dead_code)]
 #[derive(Clone)]
 pub(crate) struct SessionConfiguration {
     /// Provider identifier ("openai", "openrouter", ...).
@@ -406,7 +406,7 @@ impl Session {
         );
 
         let client = ModelClient::new(
-            Arc::new(per_turn_config),
+            Arc::new(per_turn_config.clone()),
             auth_manager,
             otel_event_manager,
             provider,
@@ -436,6 +436,7 @@ impl Session {
             final_output_json_schema: None,
             codex_linux_sandbox_exe: config.codex_linux_sandbox_exe.clone(),
             tool_call_gate: Arc::new(ReadinessFlag::new()),
+            truncation_policy: TruncationPolicy::new(&per_turn_config),
         }
     }
 
@@ -476,21 +477,13 @@ 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(
@@ -499,15 +492,8 @@ 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
@@ -519,34 +505,7 @@ 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();
@@ -595,7 +554,8 @@ impl Session {
         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)),
@@ -624,6 +584,10 @@ 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,
@@ -635,6 +599,18 @@ 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;
@@ -712,7 +688,8 @@ 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)
@@ -894,6 +871,7 @@ 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,
@@ -968,7 +946,7 @@ 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;
     }
@@ -982,17 +960,25 @@ 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);
+                    }
                 }
                 _ => {}
             }
@@ -1001,9 +987,13 @@ 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>) {
@@ -1020,6 +1010,15 @@ 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(
@@ -1087,11 +1086,14 @@ impl Session {
         self.send_token_count_event(turn_context).await;
     }
 
-    pub(crate) async fn override_last_token_usage_estimate(
-        &self,
-        turn_context: &TurnContext,
-        estimated_total_tokens: i64,
-    ) {
+    pub(crate) async fn recompute_token_usage(&self, turn_context: &TurnContext) {
+        let Some(estimated_total_tokens) = self
+            .clone_history()
+            .await
+            .estimate_token_count(turn_context)
+        else {
+            return;
+        };
         {
             let mut state = self.state.lock().await;
             let mut info = state.token_info().unwrap_or(TokenUsageInfo {
@@ -1185,9 +1187,11 @@ impl Session {
         &self,
         turn_context: &TurnContext,
         message: impl Into<String>,
+        http_status_code: Option<StatusCode>,
     ) {
         let event = EventMsg::StreamError(StreamErrorEvent {
             message: message.into(),
+            http_status_code: http_status_code_value(http_status_code),
         });
         self.send_event(turn_context, event).await;
     }
@@ -1197,14 +1201,7 @@ 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 {
@@ -1258,6 +1255,8 @@ impl Session {
     ) -> anyhow::Result<ListResourcesResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .list_resources(server, params)
             .await
     }
@@ -1269,6 +1268,8 @@ impl Session {
     ) -> anyhow::Result<ListResourceTemplatesResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .list_resource_templates(server, params)
             .await
     }
@@ -1280,6 +1281,8 @@ impl Session {
     ) -> anyhow::Result<ReadResourceResult> {
         self.services
             .mcp_connection_manager
+            .read()
+            .await
             .read_resource(server, params)
             .await
     }
@@ -1292,19 +1295,29 @@ 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 {
@@ -1318,10 +1331,17 @@ 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>) {
-    let mut previous_context: Option<Arc<TurnContext>> = None;
+    // Seed with context in case there is an OverrideTurnContext first.
+    let mut previous_context: Option<Arc<TurnContext>> =
+        Some(sess.new_turn(SessionSettingsUpdate::default()).await);
+
     // To break out of this loop, send Op::Shutdown.
     while let Ok(sub) = rx_sub.recv().await {
         debug!(?sub, "Submission");
@@ -1425,6 +1445,7 @@ 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;
@@ -1575,17 +1596,15 @@ 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()
@@ -1594,7 +1613,10 @@ 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,
@@ -1632,16 +1654,15 @@ 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 {
@@ -1662,6 +1683,7 @@ mod handlers {
                 id: sub_id.clone(),
                 msg: EventMsg::Error(ErrorEvent {
                     message: "Failed to shutdown rollout recorder".to_string(),
+                    http_status_code: None,
                 }),
             };
             sess.send_event_raw(event).await;
@@ -1767,6 +1789,7 @@ 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()),
+        truncation_policy: TruncationPolicy::new(&per_turn_config),
     };
 
     // Seed the child task with the review prompt as the initial user message.
@@ -1774,7 +1797,12 @@ 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))
@@ -1819,7 +1847,6 @@ 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
@@ -1874,27 +1901,17 @@ 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;
+                    if should_use_remote_compact_task(&sess).await {
+                        run_inline_remote_auto_compact_task(sess.clone(), turn_context.clone())
+                            .await;
+                    } else {
+                        run_inline_auto_compact_task(sess.clone(), turn_context.clone()).await;
                     }
-                    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,
@@ -1920,10 +1937,8 @@ pub(crate) async fn run_task(
             }
             Err(e) => {
                 info!("Turn error: {e:#}");
-                let event = EventMsg::Error(ErrorEvent {
-                    message: e.to_string(),
-                });
-                sess.send_event(&turn_context, event).await;
+                sess.send_event(&turn_context, EventMsg::Error(e.to_error_event(None)))
+                    .await;
                 // let the user continue the conversation
                 break;
             }
@@ -1940,22 +1955,54 @@ 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");
+        if let Some(family) =
+            find_family_for_model(&sess.state.lock().await.session_configuration.model)
+        {
+            let mut new_instructions = base_instructions.unwrap_or(family.base_instructions);
+            new_instructions.push_str(INSTRUCTIONS);
+            base_instructions = Some(new_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(),
     };
 
@@ -2015,6 +2062,7 @@ async fn run_turn(
                     sess.notify_stream_error(
                         &turn_context,
                         format!("Reconnecting... {retries}/{max_retries}"),
+                        e.http_status_code(),
                     )
                     .await;
 
@@ -2112,7 +2160,7 @@ 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);
@@ -2323,59 +2371,6 @@ 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;
@@ -2385,10 +2380,7 @@ 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;
@@ -2408,7 +2400,6 @@ 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;
 
@@ -2499,8 +2490,9 @@ mod tests {
             duration: StdDuration::from_secs(1),
             timed_out: true,
         };
+        let (_, turn_context) = make_session_and_context();
 
-        let out = format_exec_output_str(&exec);
+        let out = format_exec_output_str(&exec, turn_context.truncation_policy);
 
         assert_eq!(
             out,
@@ -2622,7 +2614,8 @@ mod tests {
         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),
@@ -2698,7 +2691,8 @@ mod tests {
         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),
@@ -2821,7 +2815,8 @@ 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;
 
@@ -2879,9 +2874,23 @@ 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,
@@ -2892,6 +2901,7 @@ 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()));
@@ -2924,7 +2934,7 @@ 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,
@@ -2933,7 +2943,7 @@ 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 {
@@ -2943,13 +2953,13 @@ 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,
@@ -2957,6 +2967,7 @@ mod tests {
         live_history.replace(rebuilt1);
         rollout_items.push(RolloutItem::Compacted(CompactedItem {
             message: summary1.to_string(),
+            replacement_history: None,
         }));
 
         let user2 = ResponseItem::Message {
@@ -2966,7 +2977,7 @@ 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 {
@@ -2976,13 +2987,13 @@ 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,
@@ -2990,6 +3001,7 @@ mod tests {
         live_history.replace(rebuilt2);
         rollout_items.push(RolloutItem::Compacted(CompactedItem {
             message: summary2.to_string(),
+            replacement_history: None,
         }));
 
         let user3 = ResponseItem::Message {
@@ -2999,7 +3011,7 @@ 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 {
@@ -3009,7 +3021,7 @@ 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())
@@ -3141,7 +3153,6 @@ 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;
@@ -3183,89 +3194,4 @@ 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

@@ -7,15 +7,18 @@ use crate::codex::TurnContext;
 use crate::codex::get_last_assistant_message_from_turn;
 use crate::error::CodexErr;
 use crate::error::Result as CodexResult;
+use crate::features::Feature;
 use crate::protocol::AgentMessageEvent;
 use crate::protocol::CompactedItem;
-use crate::protocol::ErrorEvent;
 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_app_server_protocol::AuthMode;
 use codex_protocol::items::TurnItem;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::ResponseInputItem;
@@ -26,14 +29,25 @@ 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 should_use_remote_compact_task(session: &Session) -> bool {
+    session
+        .services
+        .auth_manager
+        .auth()
+        .is_some_and(|auth| auth.mode == AuthMode::ChatGPT)
+        && session.enabled(Feature::RemoteCompaction).await
+}
+
 pub(crate) async fn run_inline_auto_compact_task(
     sess: Arc<Session>,
     turn_context: Arc<TurnContext>,
 ) {
     let prompt = turn_context.compact_prompt().to_string();
     let input = vec![UserInput::Text { text: prompt }];
+
     run_compact_task_inner(sess, turn_context, input).await;
 }
 
@@ -41,13 +55,12 @@ pub(crate) async fn run_compact_task(
     sess: Arc<Session>,
     turn_context: Arc<TurnContext>,
     input: Vec<UserInput>,
-) -> 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;
     run_compact_task_inner(sess.clone(), turn_context, input).await;
-    None
 }
 
 async fn run_compact_task_inner(
@@ -58,7 +71,10 @@ 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;
 
@@ -111,10 +127,8 @@ async fn run_compact_task_inner(
                     continue;
                 }
                 sess.set_total_tokens_full(turn_context.as_ref()).await;
-                let event = EventMsg::Error(ErrorEvent {
-                    message: e.to_string(),
-                });
-                sess.send_event(&turn_context, event).await;
+                sess.send_event(&turn_context, EventMsg::Error(e.to_error_event(None)))
+                    .await;
                 return;
             }
             Err(e) => {
@@ -124,15 +138,14 @@ async fn run_compact_task_inner(
                     sess.notify_stream_error(
                         turn_context.as_ref(),
                         format!("Reconnecting... {retries}/{max_retries}"),
+                        e.http_status_code(),
                     )
                     .await;
                     tokio::time::sleep(delay).await;
                     continue;
                 } else {
-                    let event = EventMsg::Error(ErrorEvent {
-                        message: e.to_string(),
-                    });
-                    sess.send_event(&turn_context, event).await;
+                    sess.send_event(&turn_context, EventMsg::Error(e.to_error_event(None)))
+                        .await;
                     return;
                 }
             }
@@ -140,7 +153,9 @@ 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());
@@ -152,18 +167,11 @@ async fn run_compact_task_inner(
         .collect();
     new_history.extend(ghost_snapshots);
     sess.replace_history(new_history).await;
-
-    if let Some(estimated_tokens) = sess
-        .clone_history()
-        .await
-        .estimate_token_count(&turn_context)
-    {
-        sess.override_last_token_usage_estimate(&turn_context, estimated_tokens)
-            .await;
-    }
+    sess.recompute_token_usage(&turn_context).await;
 
     let rollout_item = RolloutItem::Compacted(CompactedItem {
         message: summary_text.clone(),
+        replacement_history: None,
     });
     sess.persist_rollout_items(&[rollout_item]).await;
 
@@ -201,12 +209,22 @@ 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 +234,7 @@ 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 +242,21 @@ 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 +305,8 @@ 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 +324,7 @@ async fn drain_to_completed(
 
 #[cfg(test)]
 mod tests {
+
     use super::*;
     use pretty_assertions::assert_eq;
 
@@ -395,16 +416,16 @@ 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 +458,7 @@ 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 +1,81 @@
+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::EventMsg;
+use crate::protocol::RolloutItem;
+use crate::protocol::TaskStartedEvent;
+use codex_protocol::models::ResponseItem;
+
+pub(crate) async fn run_inline_remote_auto_compact_task(
+    sess: Arc<Session>,
+    turn_context: Arc<TurnContext>,
+) {
+    run_remote_compact_task_inner(&sess, &turn_context).await;
+}
+
+pub(crate) async fn run_remote_compact_task(sess: Arc<Session>, turn_context: Arc<TurnContext>) {
+    let start_event = EventMsg::TaskStarted(TaskStartedEvent {
+        model_context_window: turn_context.client.get_model_context_window(),
+    });
+    sess.send_event(&turn_context, start_event).await;
+
+    run_remote_compact_task_inner(&sess, &turn_context).await;
+}
+
+async fn run_remote_compact_task_inner(sess: &Arc<Session>, turn_context: &Arc<TurnContext>) {
+    if let Err(err) = run_remote_compact_task_inner_impl(sess, turn_context).await {
+        let event = err.to_error_event(Some("Error running remote compact task".to_string()));
+        sess.send_event(turn_context, EventMsg::Error(event)).await;
+    }
+}
+
+async fn run_remote_compact_task_inner_impl(
+    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;
+    sess.recompute_token_usage(turn_context).await;
+
+    let compacted_item = CompactedItem {
+        message: String::new(),
+        replacement_history: Some(new_history),
+    };
+    sess.persist_rollout_items(&[RolloutItem::Compacted(compacted_item)])
+        .await;
+
+    let event = EventMsg::AgentMessage(AgentMessageEvent {
+        message: "Compact task completed".to_string(),
+    });
+    sess.send_event(turn_context, event).await;
+
+    Ok(())
+}

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

@@ -550,6 +550,15 @@ impl ConfigEditsBuilder {
         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
+    }
+
     /// Apply edits on a blocking thread.
     pub fn apply_blocking(self) -> anyhow::Result<()> {
         apply_blocking(&self.codex_home, self.profile.as_deref(), &self.edits)
@@ -584,7 +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),
             }],
         )
@@ -592,7 +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);
@@ -722,7 +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");
@@ -836,6 +845,36 @@ hide_gpt5_1_migration_prompt = true
         assert_eq!(contents, expected);
     }
 
+    #[test]
+    fn blocking_set_hide_gpt_5_1_codex_max_migration_prompt_preserves_table() {
+        let tmp = tempdir().expect("tmpdir");
+        let codex_home = tmp.path();
+        std::fs::write(
+            codex_home.join(CONFIG_TOML_FILE),
+            r#"[notice]
+existing = "value"
+"#,
+        )
+        .expect("seed");
+        apply_blocking(
+            codex_home,
+            None,
+            &[ConfigEdit::SetNoticeHideModelMigrationPrompt(
+                "hide_gpt-5.1-codex-max_migration_prompt".to_string(),
+                true,
+            )],
+        )
+        .expect("persist");
+
+        let contents =
+            std::fs::read_to_string(codex_home.join(CONFIG_TOML_FILE)).expect("read config");
+        let expected = r#"[notice]
+existing = "value"
+"hide_gpt-5.1-codex-max_migration_prompt" = true
+"#;
+        assert_eq!(contents, expected);
+    }
+
     #[test]
     fn blocking_replace_mcp_servers_round_trips() {
         let tmp = tempdir().expect("tmpdir");
@@ -972,14 +1011,14 @@ 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);
@@ -1001,11 +1040,11 @@ 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;
@@ -59,12 +61,9 @@ pub mod edit;
 pub mod profile;
 pub mod types;
 
-#[cfg(target_os = "windows")]
-pub const OPENAI_DEFAULT_MODEL: &str = "gpt-5";
-#[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
@@ -79,7 +78,7 @@ 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-max".
     pub review_model: String,
 
     pub model_family: ModelFamily,
@@ -193,6 +192,9 @@ 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,
@@ -466,6 +468,48 @@ pub fn set_project_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;
@@ -592,6 +636,9 @@ 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>,
 
@@ -663,6 +710,8 @@ 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 {
@@ -851,6 +900,34 @@ 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.
@@ -1135,6 +1212,7 @@ impl Config {
                     }
                 })
                 .collect(),
+            tool_output_token_limit: cfg.tool_output_token_limit,
             codex_home,
             history,
             file_opener: cfg.file_opener.unwrap_or(UriBasedFileOpener::VsCode),
@@ -1239,6 +1317,16 @@ 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 {
@@ -2542,7 +2630,7 @@ 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?;
 
@@ -2550,7 +2638,7 @@ 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(())
@@ -2564,7 +2652,7 @@ 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]
@@ -2600,7 +2688,7 @@ 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?;
 
@@ -2612,7 +2700,7 @@ 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)
@@ -2634,7 +2722,7 @@ model = "gpt-4"
 model_reasoning_effort = "medium"
 
 [profiles.prod]
-model = "gpt-5-codex"
+model = "gpt-5.1-codex"
 "#,
         )
         .await?;
@@ -2663,7 +2751,7 @@ model = "gpt-5-codex"
                 .profiles
                 .get("prod")
                 .and_then(|profile| profile.model.as_deref()),
-            Some("gpt-5-codex"),
+            Some("gpt-5.1-codex"),
         );
 
         Ok(())
@@ -2778,7 +2866,7 @@ 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"
@@ -2887,6 +2975,7 @@ 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,
@@ -2958,6 +3047,7 @@ 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,
@@ -3044,6 +3134,7 @@ 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,
@@ -3094,9 +3185,9 @@ 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),
@@ -3116,6 +3207,7 @@ 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,
@@ -3265,6 +3357,41 @@ 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#"
@@ -3295,6 +3422,85 @@ trust_level = "untrusted"
         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()?;

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>,
     },
 }
 
@@ -364,6 +378,8 @@ pub struct Notice {
     pub hide_rate_limit_model_nudge: Option<bool>,
     /// Tracks whether the user has seen the model migration prompt
     pub hide_gpt5_1_migration_prompt: Option<bool>,
+    /// Tracks whether the user has seen the gpt-5.1-codex-max migration prompt
+    pub hide_gpt_5_1_codex_max_migration_prompt: Option<bool>,
 }
 
 impl Notice {

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 +144,18 @@ impl ContextManager {
         items.retain(|item| !matches!(item, ResponseItem::GhostSnapshot { .. }));
     }
 
-    fn process_item(item: &ResponseItem) -> ResponseItem {
+    fn process_item(&self, item: &ResponseItem, policy: TruncationPolicy) -> ResponseItem {
+        let policy_with_serialization_budget = policy.mul(1.2);
         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_items = output
-                    .content_items
-                    .as_ref()
-                    .map(|items| globally_truncate_function_output_items(items));
+                let truncated =
+                    truncate_text(output.content.as_str(), policy_with_serialization_budget);
+                let truncated_items = output.content_items.as_ref().map(|items| {
+                    truncate_function_output_items_with_policy(
+                        items,
+                        policy_with_serialization_budget,
+                    )
+                });
                 ResponseItem::FunctionCallOutput {
                     call_id: call_id.clone(),
                     output: FunctionCallOutputPayload {
@@ -172,11 +166,7 @@ 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_with_serialization_budget);
                 ResponseItem::CustomToolCallOutput {
                     call_id: call_id.clone(),
                     output: truncated,
@@ -188,6 +178,7 @@ impl ContextManager {
             | ResponseItem::FunctionCall { .. }
             | ResponseItem::WebSearchCall { .. }
             | ResponseItem::CustomToolCall { .. }
+            | ResponseItem::CompactionSummary { .. }
             | ResponseItem::GhostSnapshot { .. }
             | ResponseItem::Other => item.clone(),
         }
@@ -205,7 +196,8 @@ 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_BYTES: usize = 10_000;
+const EXEC_FORMAT_MAX_TOKENS: usize = 2_500;
+
 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
 }
 
@@ -52,9 +56,14 @@ fn reasoning_msg(text: &str) -> ResponseItem {
     }
 }
 
+fn truncate_exec_output(content: &str) -> String {
+    truncate::truncate_text(content, TruncationPolicy::Tokens(EXEC_FORMAT_MAX_TOKENS))
+}
+
 #[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 +73,12 @@ 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!(
@@ -223,7 +232,7 @@ fn normalization_retains_local_shell_outputs() {
         ResponseItem::FunctionCallOutput {
             call_id: "shell-1".to_string(),
             output: FunctionCallOutputPayload {
-                content: "ok".to_string(),
+                content: "Total output lines: 1\n\nok".to_string(),
                 ..Default::default()
             },
         },
@@ -237,6 +246,9 @@ 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 +260,20 @@ 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 +284,7 @@ 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 {
@@ -274,23 +292,50 @@ 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:?}"),
     }
 }
 
-fn assert_truncated_message_matches(message: &str, line: &str, total_lines: usize) {
-    let pattern = truncated_message_pattern(line, total_lines);
+#[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, expected_removed: usize) {
+    let pattern = truncated_message_pattern(line);
     let regex = Regex::new(&pattern).unwrap_or_else(|err| {
         panic!("failed to compile regex {pattern}: {err}");
     });
@@ -302,28 +347,22 @@ 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()
     );
+    let removed: usize = captures
+        .name("removed")
+        .expect("missing removed capture")
+        .as_str()
+        .parse()
+        .unwrap_or_else(|err| panic!("invalid removed tokens: {err}"));
+    assert_eq!(removed, expected_removed, "mismatched removed token count");
 }
 
-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);
+fn truncated_message_pattern(line: &str) -> String {
     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,
-        );
-    }
-    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.*)$",
-    )
+    format!(r"(?s)^(?P<body>{escaped_line}.*?)(?:\r?)?…(?P<removed>\d+) tokens truncated…(?:.*)?$")
 }
 
 #[test]
@@ -331,35 +370,18 @@ 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_exec_output(&large_error);
 
-    let total_lines = large_error.lines().count();
-    assert_truncated_message_matches(&truncated, line, total_lines);
+    assert_truncated_message_matches(&truncated, line, 36250);
     assert_ne!(truncated, 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 + 10000);
+    let truncated = truncate_exec_output(&long_line);
     assert_ne!(truncated, long_line);
-    let marker_line = format!(
-        "[... output truncated to fit {} bytes ...]",
-        truncate::MODEL_FORMAT_MAX_BYTES
-    );
-    assert!(
-        truncated.contains(&marker_line),
-        "missing byte truncation marker: {truncated}"
-    );
+    assert_truncated_message_matches(&truncated, "a", 2500);
     assert!(
         !truncated.contains("omitted"),
         "line omission marker should not appear when no lines were dropped: {truncated}"
@@ -369,42 +391,25 @@ fn format_exec_output_marks_byte_truncation_without_omitted_lines() {
 #[test]
 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
-        ),
-        content
-    );
+    assert_eq!(truncate_exec_output(&content), 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 = 2_000;
+    let filler = "x".repeat(64);
     let content: String = (0..total_lines)
-        .map(|idx| format!("line-{idx}\n"))
+        .map(|idx| format!("line-{idx}-{filler}\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 expected_marker = format!("[... omitted {omitted} of {total_lines} lines ...]");
-
+    let truncated = truncate_exec_output(&content);
+    assert_truncated_message_matches(&truncated, "line-0-", 34_723);
     assert!(
-        truncated.contains(&expected_marker),
-        "missing omitted marker: {truncated}"
-    );
-    assert!(
-        truncated.contains("line-0\n"),
+        truncated.contains("line-0-"),
         "expected head line to remain: {truncated}"
     );
 
-    let last_line = format!("line-{}\n", total_lines - 1);
+    let last_line = format!("line-{}-", total_lines - 1);
     assert!(
         truncated.contains(&last_line),
         "expected tail line to remain: {truncated}"
@@ -413,101 +418,15 @@ 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 = 300;
     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_exec_output(&content);
 
-    assert!(
-        truncated.contains("[... omitted 42 of 298 lines ...]"),
-        "expected omitted marker when line count exceeds limit: {truncated}"
-    );
-    assert!(
-        !truncated.contains("output truncated to fit"),
-        "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"));
+    assert_truncated_message_matches(&truncated, "line-0-", 17_423);
 }
 
 //TODO(aibrahim): run CI in release mode.

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

@@ -1,8 +1,4 @@
 mod history;
 mod normalize;
-mod truncate;
 
 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

@@ -2,13 +2,15 @@ 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;
 use chrono::Utc;
 use codex_async_utils::CancelErr;
 use codex_protocol::ConversationId;
+use codex_protocol::protocol::ErrorEvent;
 use codex_protocol::protocol::RateLimitSnapshot;
 use reqwest::StatusCode;
 use serde_json;
@@ -429,6 +431,37 @@ impl CodexErr {
     pub fn downcast_ref<T: std::any::Any>(&self) -> Option<&T> {
         (self as &dyn std::any::Any).downcast_ref::<T>()
     }
+
+    pub fn http_status_code(&self) -> Option<StatusCode> {
+        match self {
+            CodexErr::UnexpectedStatus(err) => Some(err.status),
+            CodexErr::RetryLimit(err) => Some(err.status),
+            CodexErr::UsageLimitReached(_) | CodexErr::UsageNotIncluded => {
+                Some(StatusCode::TOO_MANY_REQUESTS)
+            }
+            CodexErr::InternalServerError => Some(StatusCode::INTERNAL_SERVER_ERROR),
+            CodexErr::ResponseStreamFailed(err) => err.source.status(),
+            CodexErr::ConnectionFailed(err) => err.source.status(),
+            _ => None,
+        }
+    }
+
+    pub fn to_error_event(&self, message_prefix: Option<String>) -> ErrorEvent {
+        let error_message = self.to_string();
+        let message: String = match message_prefix {
+            Some(prefix) => format!("{prefix}: {error_message}"),
+            None => error_message,
+        };
+
+        ErrorEvent {
+            message,
+            http_status_code: http_status_code_value(self.http_status_code()),
+        }
+    }
+}
+
+pub fn http_status_code_value(http_status_code: Option<StatusCode>) -> Option<u16> {
+    http_status_code.as_ref().map(StatusCode::as_u16)
 }
 
 pub fn get_error_message_ui(e: &CodexErr) -> String {
@@ -461,7 +494,10 @@ 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)]
@@ -495,6 +531,7 @@ mod tests {
                 window_minutes: Some(120),
                 resets_at: Some(secondary_reset_at),
             }),
+            credits: None,
         }
     }
 
@@ -770,4 +807,43 @@ mod tests {
             assert_eq!(err.to_string(), expected);
         });
     }
+
+    #[test]
+    fn error_event_includes_http_status_code_when_available() {
+        let err = CodexErr::UnexpectedStatus(UnexpectedResponseError {
+            status: StatusCode::BAD_REQUEST,
+            body: "oops".to_string(),
+            request_id: Some("req-1".to_string()),
+        });
+        let event = err.to_error_event(None);
+
+        assert_eq!(
+            event.message,
+            "unexpected status 400 Bad Request: oops, request id: req-1"
+        );
+        assert_eq!(
+            event.http_status_code,
+            Some(StatusCode::BAD_REQUEST.as_u16())
+        );
+    }
+
+    #[test]
+    fn error_event_omits_http_status_code_when_unknown() {
+        let event = CodexErr::Fatal("boom".to_string()).to_error_event(None);
+
+        assert_eq!(event.message, "Fatal error: boom");
+        assert_eq!(event.http_status_code, None);
+    }
+
+    #[test]
+    fn error_event_applies_message_wrapper() {
+        let event = CodexErr::Fatal("boom".to_string())
+            .to_error_event(Some("Error running remote compact task".to_string()));
+
+        assert_eq!(
+            event.message,
+            "Error running remote compact task: Fatal error: boom"
+        );
+        assert_eq!(event.http_status_code, None);
+    }
 }

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/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
@@ -42,10 +44,14 @@ pub enum Feature {
     WebSearchRequest,
     /// 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 {
@@ -247,6 +253,20 @@ pub struct FeatureSpec {
 }
 
 pub const FEATURES: &[FeatureSpec] = &[
+    // Stable features.
+    FeatureSpec {
+        id: Feature::GhostCommit,
+        key: "undo",
+        stage: Stage::Stable,
+        default_enabled: true,
+    },
+    FeatureSpec {
+        id: Feature::ViewImageTool,
+        key: "view_image_tool",
+        stage: Stage::Stable,
+        default_enabled: true,
+    },
+    // Unstable features.
     FeatureSpec {
         id: Feature::UnifiedExec,
         key: "unified_exec",
@@ -271,12 +291,6 @@ pub const FEATURES: &[FeatureSpec] = &[
         stage: Stage::Beta,
         default_enabled: false,
     },
-    FeatureSpec {
-        id: Feature::ViewImageTool,
-        key: "view_image_tool",
-        stage: Stage::Stable,
-        default_enabled: true,
-    },
     FeatureSpec {
         id: Feature::WebSearchRequest,
         key: "web_search_request",
@@ -289,16 +303,28 @@ pub const FEATURES: &[FeatureSpec] = &[
         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;
@@ -34,14 +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.
@@ -11,6 +12,7 @@ const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
 
 const GPT_5_CODEX_INSTRUCTIONS: &str = include_str!("../gpt_5_codex_prompt.md");
 const GPT_5_1_INSTRUCTIONS: &str = include_str!("../gpt_5_1_prompt.md");
+const GPT_5_1_CODEX_MAX_INSTRUCTIONS: &str = include_str!("../gpt-5.1-codex-max_prompt.md");
 
 /// A model family is a group of models that share certain characteristics.
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
@@ -66,6 +68,8 @@ 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 +93,7 @@ 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 +137,7 @@ 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 +150,9 @@ 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.
@@ -161,11 +168,25 @@ 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.
+    } else if slug.starts_with("gpt-5.1-codex-max") {
+        model_family!(
+            slug, slug,
+            supports_reasoning_summaries: true,
+            reasoning_summary_format: ReasoningSummaryFormat::Experimental,
+            base_instructions: GPT_5_1_CODEX_MAX_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-codex")
         || slug.starts_with("gpt-5.1-codex")
         || slug.starts_with("codex-")
@@ -176,7 +197,10 @@ 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 +211,18 @@ 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::Default,
             support_verbosity: true,
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else {
         None
@@ -216,5 +245,6 @@ 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/openai_model_info.rs

@@ -70,7 +70,10 @@ pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
         // https://platform.openai.com/docs/models/gpt-3.5-turbo
         "gpt-3.5-turbo" => Some(ModelInfo::new(16_385, 4_096)),
 
-        _ if slug.starts_with("gpt-5-codex") || slug.starts_with("gpt-5.1-codex") => {
+        _ if slug.starts_with("gpt-5-codex")
+            || slug.starts_with("gpt-5.1-codex")
+            || slug.starts_with("gpt-5.1-codex-max") =>
+        {
             Some(ModelInfo::new(CONTEXT_WINDOW_272K, MAX_OUTPUT_TOKENS_128K))
         }
 

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 +883,42 @@ 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(),
+            }],
+        );
+    }
 }
 
 pub fn parse_command_impl(command: &[String]) -> Vec<ParsedCommand> {
@@ -884,6 +926,12 @@ pub fn parse_command_impl(command: &[String]) -> Vec<ParsedCommand> {
         return commands;
     }
 
+    if let Some((_, script)) = extract_powershell_command(command) {
+        return vec![ParsedCommand::Unknown {
+            cmd: script.to_string(),
+        }];
+    }
+
     let normalized = normalize_tokens(command);
 
     let parts = if contains_connectors(&normalized) {
@@ -1190,6 +1238,7 @@ fn parse_find_query_and_path(tail: &[String]) -> (Option<String>, Option<String>
 }
 
 fn parse_shell_lc_commands(original: &[String]) -> Option<Vec<ParsedCommand>> {
+    // Only handle bash/zsh here; PowerShell is stripped separately without bash parsing.
     let (_, script) = extract_bash_command(original)?;
 
     if let Some(tree) = try_parse_shell(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/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 +1,12 @@
 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::state::TaskKind;
+use async_trait::async_trait;
+use codex_protocol::user_input::UserInput;
+use tokio_util::sync::CancellationToken;
 
 #[derive(Clone, Copy, Default)]
 pub(crate) struct CompactTask;
@@ -27,6 +24,13 @@ 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 crate::compact::should_use_remote_compact_task(&session).await {
+            crate::compact_remote::run_remote_compact_task(session, ctx).await
+        } else {
+            crate::compact::run_compact_task(session, ctx, input).await
+        }
+
+        None
     }
 }

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

@@ -1,10 +1,14 @@
 use crate::codex::TurnContext;
+use crate::protocol::EventMsg;
+use crate::protocol::WarningEvent;
 use crate::state::TaskKind;
 use crate::tasks::SessionTask;
 use crate::tasks::SessionTaskContext;
 use async_trait::async_trait;
 use codex_git::CreateGhostCommitOptions;
+use codex_git::GhostSnapshotReport;
 use codex_git::GitToolingError;
+use codex_git::capture_ghost_snapshot_report;
 use codex_git::create_ghost_commit;
 use codex_protocol::models::ResponseItem;
 use codex_protocol::user_input::UserInput;
@@ -39,6 +43,27 @@ impl SessionTask for GhostSnapshotTask {
                 _ = cancellation_token.cancelled() => true,
                 _ = async {
                     let repo_path = ctx_for_task.cwd.clone();
+                    // First, compute a snapshot report so we can warn about
+                    // large untracked directories before running the heavier
+                    // snapshot logic.
+                    if let Ok(Ok(report)) = tokio::task::spawn_blocking({
+                        let repo_path = repo_path.clone();
+                        move || {
+                            let options = CreateGhostCommitOptions::new(&repo_path);
+                            capture_ghost_snapshot_report(&options)
+                        }
+                    })
+                    .await
+                        && let Some(message) = format_large_untracked_warning(&report) {
+                                session
+                                    .session
+                                    .send_event(
+                                        &ctx_for_task,
+                                        EventMsg::Warning(WarningEvent { message }),
+                                    )
+                                    .await;
+                            }
+
                     // Required to run in a dedicated blocking pool.
                     match tokio::task::spawn_blocking(move || {
                         let options = CreateGhostCommitOptions::new(&repo_path);
@@ -56,23 +81,18 @@ impl SessionTask for GhostSnapshotTask {
                                 .await;
                             info!("ghost commit captured: {}", ghost_commit.id());
                         }
-                        Ok(Err(err)) => {
-                            warn!(
+                        Ok(Err(err)) => match err {
+                            GitToolingError::NotAGitRepository { .. } => info!(
                                 sub_id = ctx_for_task.sub_id.as_str(),
-                                "failed to capture ghost snapshot: {err}"
-                            );
-                            let message = match err {
-                                GitToolingError::NotAGitRepository { .. } => {
-                                    "Snapshots disabled: current directory is not a Git repository."
-                                        .to_string()
-                                }
-                                _ => format!("Snapshots disabled after ghost snapshot error: {err}."),
-                            };
-                            session
-                                .session
-                                .notify_background_event(&ctx_for_task, message)
-                                .await;
-                        }
+                                "skipping ghost snapshot because current directory is not a Git repository"
+                            ),
+                            _ => {
+                                warn!(
+                                    sub_id = ctx_for_task.sub_id.as_str(),
+                                    "failed to capture ghost snapshot: {err}"
+                                );
+                            }
+                        },
                         Err(err) => {
                             warn!(
                                 sub_id = ctx_for_task.sub_id.as_str(),
@@ -108,3 +128,22 @@ impl GhostSnapshotTask {
         Self { token }
     }
 }
+
+fn format_large_untracked_warning(report: &GhostSnapshotReport) -> Option<String> {
+    if report.large_untracked_dirs.is_empty() {
+        return None;
+    }
+    const MAX_DIRS: usize = 3;
+    let mut parts: Vec<String> = Vec::new();
+    for dir in report.large_untracked_dirs.iter().take(MAX_DIRS) {
+        parts.push(format!("{} ({} files)", dir.path.display(), dir.file_count));
+    }
+    if report.large_untracked_dirs.len() > MAX_DIRS {
+        let remaining = report.large_untracked_dirs.len() - MAX_DIRS;
+        parts.push(format!("{remaining} more"));
+    }
+    Some(format!(
+        "Repository snapshot encountered large untracked directories: {}. This can slow Codex; consider adding these paths to .gitignore or disabling undo in your config.",
+        parts.join(", ")
+    ))
+}

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

@@ -65,22 +65,24 @@ 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,
+                    turn_id: turn_context.sub_id.clone(),
+                    command: command.clone(),
+                    cwd: cwd.clone(),
+                    parsed_cmd: parsed_cmd.clone(),
                     source: ExecCommandSource::UserShell,
                     interaction_input: None,
                 }),
@@ -88,8 +90,8 @@ impl SessionTask for UserShellCommandTask {
             .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,
@@ -120,7 +122,11 @@ impl SessionTask for UserShellCommandTask {
                     duration: Duration::ZERO,
                     timed_out: false,
                 };
-                let output_items = [user_shell_command_record_item(&raw_command, &exec_output)];
+                let output_items = [user_shell_command_record_item(
+                    &raw_command,
+                    &exec_output,
+                    &turn_context,
+                )];
                 session
                     .record_conversation_items(turn_context.as_ref(), &output_items)
                     .await;
@@ -129,6 +135,12 @@ 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(),
@@ -145,17 +157,30 @@ 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(),
                             exit_code: output.exit_code,
                             duration: output.duration,
-                            formatted_output: format_exec_output_str(&output),
+                            formatted_output: format_exec_output_str(
+                                &output,
+                                turn_context.truncation_policy,
+                            ),
                         }),
                     )
                     .await;
 
-                let output_items = [user_shell_command_record_item(&raw_command, &output)];
+                let output_items = [user_shell_command_record_item(
+                    &raw_command,
+                    &output,
+                    &turn_context,
+                )];
                 session
                     .record_conversation_items(turn_context.as_ref(), &output_items)
                     .await;
@@ -176,16 +201,29 @@ 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(),
                             exit_code: exec_output.exit_code,
                             duration: exec_output.duration,
-                            formatted_output: format_exec_output_str(&exec_output),
+                            formatted_output: format_exec_output_str(
+                                &exec_output,
+                                turn_context.truncation_policy,
+                            ),
                         }),
                     )
                     .await;
-                let output_items = [user_shell_command_record_item(&raw_command, &exec_output)];
+                let output_items = [user_shell_command_record_item(
+                    &raw_command,
+                    &exec_output,
+                    &turn_context,
+                )];
                 session
                     .record_conversation_items(turn_context.as_ref(), &output_items)
                     .await;

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

@@ -15,6 +15,7 @@ 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;
@@ -61,6 +62,7 @@ pub(crate) async fn emit_exec_command_begin(
     ctx: ToolEventCtx<'_>,
     command: &[String],
     cwd: &Path,
+    parsed_cmd: &[ParsedCommand],
     source: ExecCommandSource,
     interaction_input: Option<String>,
 ) {
@@ -69,9 +71,10 @@ pub(crate) async fn emit_exec_command_begin(
             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),
+                parsed_cmd: parsed_cmd.to_vec(),
                 source,
                 interaction_input,
             }),
@@ -84,6 +87,8 @@ pub(crate) enum ToolEmitter {
         command: Vec<String>,
         cwd: PathBuf,
         source: ExecCommandSource,
+        parsed_cmd: Vec<ParsedCommand>,
+        freeform: bool,
     },
     ApplyPatch {
         changes: HashMap<PathBuf, FileChange>,
@@ -94,15 +99,24 @@ pub(crate) enum ToolEmitter {
         cwd: PathBuf,
         source: ExecCommandSource,
         interaction_input: Option<String>,
+        parsed_cmd: Vec<ParsedCommand>,
     },
 }
 
 impl ToolEmitter {
-    pub fn shell(command: Vec<String>, cwd: PathBuf, source: ExecCommandSource) -> Self {
+    pub fn shell(
+        command: Vec<String>,
+        cwd: PathBuf,
+        source: ExecCommandSource,
+        freeform: bool,
+    ) -> Self {
+        let parsed_cmd = parse_command(&command);
         Self::Shell {
             command,
             cwd,
             source,
+            parsed_cmd,
+            freeform,
         }
     }
 
@@ -119,11 +133,13 @@ impl ToolEmitter {
         source: ExecCommandSource,
         interaction_input: Option<String>,
     ) -> Self {
+        let parsed_cmd = parse_command(command);
         Self::UnifiedExec {
             command: command.to_vec(),
             cwd,
             source,
             interaction_input,
+            parsed_cmd,
         }
     }
 
@@ -134,44 +150,15 @@ impl ToolEmitter {
                     command,
                     cwd,
                     source,
+                    parsed_cmd,
+                    ..
                 },
-                ToolEventStage::Begin,
+                stage,
             ) => {
-                emit_exec_command_begin(ctx, command, cwd.as_path(), *source, None).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;
             }
@@ -231,57 +218,20 @@ impl ToolEmitter {
                     cwd,
                     source,
                     interaction_input,
+                    parsed_cmd,
                 },
-                ToolEventStage::Begin,
+                stage,
             ) => {
-                emit_exec_command_begin(
+                emit_exec_stage(
                     ctx,
-                    command,
-                    cwd.as_path(),
-                    *source,
-                    interaction_input.clone(),
-                )
-                .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)),
-            ) => {
-                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::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;
             }
@@ -292,6 +242,19 @@ impl ToolEmitter {
         self.emit(ctx, ToolEventStage::Begin).await;
     }
 
+    fn format_exec_output_for_model(
+        &self,
+        output: &ExecToolCallOutput,
+        ctx: ToolEventCtx<'_>,
+    ) -> String {
+        match self {
+            Self::Shell { freeform: true, .. } => {
+                super::format_exec_output_for_model_freeform(output, ctx.turn.truncation_policy)
+            }
+            _ => super::format_exec_output_for_model_structured(output, ctx.turn.truncation_policy),
+        }
+    }
+
     pub async fn finish(
         &self,
         ctx: ToolEventCtx<'_>,
@@ -299,7 +262,7 @@ impl ToolEmitter {
     ) -> Result<String, FunctionCallError> {
         let (event, result) = match out {
             Ok(output) => {
-                let content = super::format_exec_output_for_model(&output);
+                let content = self.format_exec_output_for_model(&output, ctx);
                 let exit_code = output.exit_code;
                 let event = ToolEventStage::Success(output);
                 let result = if exit_code == 0 {
@@ -311,7 +274,7 @@ impl ToolEmitter {
             }
             Err(ToolError::Codex(CodexErr::Sandbox(SandboxErr::Timeout { output })))
             | Err(ToolError::Codex(CodexErr::Sandbox(SandboxErr::Denied { output }))) => {
-                let response = super::format_exec_output_for_model(&output);
+                let response = self.format_exec_output_for_model(&output, ctx);
                 let event = ToolEventStage::Failure(ToolEventFailure::Output(*output));
                 let result = Err(FunctionCallError::RespondToModel(response));
                 (event, result)
@@ -340,26 +303,107 @@ 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, ctx.turn.truncation_policy),
+            };
+            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

@@ -130,7 +130,7 @@ impl ToolHandler for ShellHandler {
                     turn,
                     tracker,
                     call_id,
-                    true,
+                    false,
                 )
                 .await
             }
@@ -178,7 +178,7 @@ impl ToolHandler for ShellCommandHandler {
             turn,
             tracker,
             call_id,
-            false,
+            true,
         )
         .await
     }
@@ -192,7 +192,7 @@ impl ShellHandler {
         turn: Arc<TurnContext>,
         tracker: crate::tools::context::SharedTurnDiffTracker,
         call_id: String,
-        is_user_shell_command: bool,
+        freeform: bool,
     ) -> Result<ToolOutput, FunctionCallError> {
         // Approval policy guard for explicit escalation in non-OnRequest modes.
         if exec_params.with_escalated_permissions.unwrap_or(false)
@@ -285,14 +285,13 @@ impl ShellHandler {
             }
         }
 
-        // Regular shell execution path.
-        let source = if is_user_shell_command {
-            ExecCommandSource::UserShell
-        } else {
-            ExecCommandSource::Agent
-        };
-        let emitter =
-            ToolEmitter::shell(exec_params.command.clone(), exec_params.cwd.clone(), source);
+        let source = ExecCommandSource::Agent;
+        let emitter = ToolEmitter::shell(
+            exec_params.command.clone(),
+            exec_params.cwd.clone(),
+            source,
+            freeform,
+        );
         let event_ctx = ToolEventCtx::new(session.as_ref(), turn.as_ref(), &call_id, None);
         emitter.begin(event_ctx).await;
 

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

@@ -9,10 +9,10 @@ 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::exec::ExecToolCallOutput;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::formatted_truncate_text;
+use crate::truncate::truncate_text;
 pub use router::ToolRouter;
 use serde::Serialize;
 
@@ -24,7 +24,10 @@ pub(crate) const TELEMETRY_PREVIEW_TRUNCATION_NOTICE: &str =
 
 /// 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 {
+pub fn format_exec_output_for_model_structured(
+    exec_output: &ExecToolCallOutput,
+    truncation_policy: TruncationPolicy,
+) -> String {
     let ExecToolCallOutput {
         exit_code,
         duration,
@@ -46,7 +49,7 @@ pub fn format_exec_output_for_model(exec_output: &ExecToolCallOutput) -> String
     // round to 1 decimal place
     let duration_seconds = ((duration.as_secs_f32()) * 10.0).round() / 10.0;
 
-    let formatted_output = format_exec_output_str(exec_output);
+    let formatted_output = format_exec_output_str(exec_output, truncation_policy);
 
     let payload = ExecOutput {
         output: &formatted_output,
@@ -60,7 +63,35 @@ pub fn format_exec_output_for_model(exec_output: &ExecToolCallOutput) -> String
     serde_json::to_string(&payload).expect("serialize ExecOutput")
 }
 
-pub fn format_exec_output_str(exec_output: &ExecToolCallOutput) -> String {
+pub fn format_exec_output_for_model_freeform(
+    exec_output: &ExecToolCallOutput,
+    truncation_policy: TruncationPolicy,
+) -> String {
+    // round to 1 decimal place
+    let duration_seconds = ((exec_output.duration.as_secs_f32()) * 10.0).round() / 10.0;
+
+    let total_lines = exec_output.aggregated_output.text.lines().count();
+
+    let formatted_output = truncate_text(&exec_output.aggregated_output.text, truncation_policy);
+
+    let mut sections = Vec::new();
+
+    sections.push(format!("Exit code: {}", exec_output.exit_code));
+    sections.push(format!("Wall time: {duration_seconds} seconds"));
+    if total_lines != formatted_output.lines().count() {
+        sections.push(format!("Total output lines: {total_lines}"));
+    }
+
+    sections.push("Output:".to_string());
+    sections.push(formatted_output);
+
+    sections.join("\n")
+}
+
+pub fn format_exec_output_str(
+    exec_output: &ExecToolCallOutput,
+    truncation_policy: TruncationPolicy,
+) -> String {
     let ExecToolCallOutput {
         aggregated_output, ..
     } = exec_output;
@@ -77,5 +108,5 @@ 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)
+    formatted_truncate_text(&body, truncation_policy)
 }

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/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",
@@ -1349,13 +1395,29 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_gpt_5_defaults() {
+        assert_model_tools(
+            "gpt-5",
+            &Features::with_defaults(),
+            &[
+                "shell",
+                "list_mcp_resources",
+                "list_mcp_resource_templates",
+                "read_mcp_resource",
+                "update_plan",
+                "view_image",
+            ],
+        );
+    }
+
     #[test]
     fn test_gpt_5_1_defaults() {
         assert_model_tools(
             "gpt-5.1",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1873,8 +1935,23 @@ 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 +1965,22 @@ 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 +1,486 @@
 //! 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 crate::config::Config;
+use codex_protocol::models::FunctionCallOutputContentItem;
 
-/// 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 {
+const APPROX_BYTES_PER_TOKEN: usize = 4;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum TruncationPolicy {
+    Bytes(usize),
+    Tokens(usize),
+}
+
+impl TruncationPolicy {
+    /// Scale the underlying budget by `multiplier`, rounding up to avoid under-budgeting.
+    pub fn mul(self, multiplier: f64) -> Self {
+        match self {
+            TruncationPolicy::Bytes(bytes) => {
+                TruncationPolicy::Bytes((bytes as f64 * multiplier).ceil() as usize)
+            }
+            TruncationPolicy::Tokens(tokens) => {
+                TruncationPolicy::Tokens((tokens as f64 * multiplier).ceil() as usize)
+            }
+        }
+    }
+
+    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)
+                }
+            }
+        }
+    }
+
+    /// 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,
+        }
+    }
+
+    /// 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),
+        }
+    }
+}
+
+pub(crate) fn formatted_truncate_text(content: &str, policy: TruncationPolicy) -> String {
+    if content.len() <= policy.byte_budget() {
+        return content.to_string();
+    }
+    let total_lines = content.lines().count();
+    let result = truncate_text(content, policy);
+    format!("Total output lines: {total_lines}\n\n{result}")
+}
+
+pub(crate) fn truncate_text(content: &str, policy: TruncationPolicy) -> String {
+    match policy {
+        TruncationPolicy::Bytes(_) => truncate_with_byte_estimate(content, policy),
+        TruncationPolicy::Tokens(_) => {
+            let (truncated, _) = truncate_with_token_budget(content, policy);
+            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;
+
+    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(),
+                });
+            }
+        }
+    }
+
+    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, policy: TruncationPolicy) -> (String, Option<u64>) {
+    if s.is_empty() {
+        return (String::new(), None);
+    }
+    let max_tokens = policy.token_budget();
+
+    let byte_len = s.len();
+    if max_tokens > 0 && byte_len <= approx_bytes_for_tokens(max_tokens) {
         return (s.to_string(), None);
     }
 
-    // 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)
-        }
-    };
+    let truncated = truncate_with_byte_estimate(s, policy);
+    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, policy: TruncationPolicy) -> String {
+    if s.is_empty() {
+        return String::new();
+    }
+    let total_chars = s.chars().count();
+    let max_bytes = policy.byte_budget();
 
-    let total_tokens = token_count(s);
     if max_bytes == 0 {
-        return (
-            format!("…{total_tokens} tokens truncated…"),
-            Some(total_tokens),
+        // No budget to show content; just report that everything was truncated.
+        let marker = format_truncation_marker(
+            policy,
+            removed_units_for_source(policy, s.len(), total_chars),
         );
+        return marker;
     }
 
-    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]
+    if s.len() <= max_bytes {
+        return s.to_string();
     }
 
-    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()
-    }
+    let total_bytes = s.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
-    }
-
-    // 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),
-            );
-        }
-
-        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;
-        }
-
-        // 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…");
-    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),
-        );
-    }
-
-    let left_budget = keep_budget / 2;
-    let right_budget = keep_budget - left_budget;
+    let (left_budget, right_budget) = split_budget(max_bytes);
     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);
-    out.push('\n');
-    out.push_str(&s[suffix_start..]);
-    (out, Some(total_tokens))
+    let left_chars = s[..prefix_end].chars().count();
+    let right_chars = s[suffix_start..].chars().count();
+    let removed_chars = total_chars
+        .saturating_sub(left_chars)
+        .saturating_sub(right_chars);
+
+    let marker = format_truncation_marker(
+        policy,
+        removed_units_for_source(policy, total_bytes.saturating_sub(max_bytes), removed_chars),
+    );
+
+    assemble_truncated_output(&s[..prefix_end], &s[suffix_start..], &marker)
+}
+
+fn format_truncation_marker(policy: TruncationPolicy, removed_count: u64) -> String {
+    match policy {
+        TruncationPolicy::Tokens(_) => format!("…{removed_count} tokens truncated…"),
+        TruncationPolicy::Bytes(_) => format!("…{removed_count} chars truncated…"),
+    }
+}
+
+fn split_budget(budget: usize) -> (usize, usize) {
+    let left = budget / 2;
+    (left, budget - left)
+}
+
+fn removed_units_for_source(
+    policy: TruncationPolicy,
+    removed_bytes: usize,
+    removed_chars: usize,
+) -> u64 {
+    match policy {
+        TruncationPolicy::Tokens(_) => approx_tokens_from_byte_count(removed_bytes),
+        TruncationPolicy::Bytes(_) => u64::try_from(removed_chars).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_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 {
+    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);
+    let mut idx = start_tail.min(s.len());
+    while idx < s.len() && !s.is_char_boundary(idx) {
+        idx += 1;
+    }
+    idx
 }
 
 #[cfg(test)]
 mod tests {
-    use super::truncate_middle;
-    use codex_utils_tokenizer::Tokenizer;
+
+    use super::TruncationPolicy;
+    use super::approx_token_count;
+    use super::formatted_truncate_text;
+    use super::truncate_function_output_items_with_policy;
+    use super::truncate_text;
+    use super::truncate_with_token_budget;
+    use codex_protocol::models::FunctionCallOutputContentItem;
+    use pretty_assertions::assert_eq;
 
     #[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));
+    fn truncate_bytes_less_than_placeholder_returns_placeholder() {
+        let content = "example output";
+
+        assert_eq!(
+            "Total output lines: 1\n\n…13 chars truncated…t",
+            formatted_truncate_text(content, TruncationPolicy::Bytes(1)),
+        );
     }
 
     #[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_tokens_less_than_placeholder_returns_placeholder() {
+        let content = "example output";
 
-        let max_bytes = 64;
-        let (out, tokens) = truncate_middle(&s, max_bytes);
-        assert!(out.starts_with("001\n002\n003\n004\n"));
-        assert!(out.contains("tokens truncated"));
-        assert!(out.ends_with("017\n018\n019\n020\n"));
-        assert_eq!(tokens, Some(tok.count(&s) as u64));
+        assert_eq!(
+            "Total output lines: 1\n\nex…3 tokens truncated…ut",
+            formatted_truncate_text(content, TruncationPolicy::Tokens(1)),
+        );
     }
 
     #[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);
+    fn truncate_tokens_under_limit_returns_original() {
+        let content = "example output";
 
-        assert!(out.contains("tokens truncated"));
-        assert!(!out.contains('\u{fffd}'));
-        assert_eq!(tokens, Some(tok.count(s) as u64));
+        assert_eq!(
+            content,
+            formatted_truncate_text(content, TruncationPolicy::Tokens(10)),
+        );
     }
 
     #[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 truncate_bytes_under_limit_returns_original() {
+        let content = "example output";
 
-        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));
+        assert_eq!(
+            content,
+            formatted_truncate_text(content, TruncationPolicy::Bytes(20)),
+        );
+    }
+
+    #[test]
+    fn truncate_tokens_over_limit_returns_truncated() {
+        let content = "this is an example of a long output that should be truncated";
+
+        assert_eq!(
+            "Total output lines: 1\n\nthis is an…10 tokens truncated… truncated",
+            formatted_truncate_text(content, TruncationPolicy::Tokens(5)),
+        );
+    }
+
+    #[test]
+    fn truncate_bytes_over_limit_returns_truncated() {
+        let content = "this is an example of a long output that should be truncated";
+
+        assert_eq!(
+            "Total output lines: 1\n\nthis is an exam…30 chars truncated…ld be truncated",
+            formatted_truncate_text(content, TruncationPolicy::Bytes(30)),
+        );
+    }
+
+    #[test]
+    fn truncate_bytes_reports_original_line_count_when_truncated() {
+        let content =
+            "this is an example of a long output that should be truncated\nalso some other line";
+
+        assert_eq!(
+            "Total output lines: 2\n\nthis is an exam…51 chars truncated…some other line",
+            formatted_truncate_text(content, TruncationPolicy::Bytes(30)),
+        );
+    }
+
+    #[test]
+    fn truncate_tokens_reports_original_line_count_when_truncated() {
+        let content =
+            "this is an example of a long output that should be truncated\nalso some other line";
+
+        assert_eq!(
+            "Total output lines: 2\n\nthis is an example o…11 tokens truncated…also some other line",
+            formatted_truncate_text(content, TruncationPolicy::Tokens(10)),
+        );
+    }
+
+    #[test]
+    fn truncate_with_token_budget_returns_original_when_under_limit() {
+        let s = "short output";
+        let limit = 100;
+        let (out, original) = truncate_with_token_budget(s, TruncationPolicy::Tokens(limit));
+        assert_eq!(out, s);
+        assert_eq!(original, None);
+    }
+
+    #[test]
+    fn truncate_with_token_budget_reports_truncation_at_zero_limit() {
+        let s = "abcdef";
+        let (out, original) = truncate_with_token_budget(s, TruncationPolicy::Tokens(0));
+        assert_eq!(out, "…2 tokens truncated…");
+        assert_eq!(original, Some(2));
+    }
+
+    #[test]
+    fn truncate_middle_tokens_handles_utf8_content() {
+        let s = "😀😀😀😀😀😀😀😀😀😀\nsecond line with text\n";
+        let (out, tokens) = truncate_with_token_budget(s, TruncationPolicy::Tokens(8));
+        assert_eq!(out, "😀😀😀😀…8 tokens truncated… line with text\n");
+        assert_eq!(tokens, Some(16));
+    }
+
+    #[test]
+    fn truncate_middle_bytes_handles_utf8_content() {
+        let s = "😀😀😀😀😀😀😀😀😀😀\nsecond line with text\n";
+        let out = truncate_text(s, TruncationPolicy::Bytes(20));
+        assert_eq!(out, "😀😀…21 chars truncated…with text\n");
+    }
+
+    #[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>,
@@ -124,37 +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 {

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::formatted_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 = formatted_truncate_text(
+                &aggregated_text,
+                TruncationPolicy::Tokens(UNIFIED_EXEC_OUTPUT_MAX_TOKENS),
+            );
             let message = if snippet.is_empty() {
                 format!("exit code {exit_code}")
             } else {

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

@@ -23,6 +23,9 @@ use crate::tools::orchestrator::ToolOrchestrator;
 use crate::tools::runtimes::unified_exec::UnifiedExecRequest as UnifiedExecToolRequest;
 use crate::tools::runtimes::unified_exec::UnifiedExecRuntime;
 use crate::tools::sandboxing::ToolCtx;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::approx_token_count;
+use crate::truncate::formatted_truncate_text;
 
 use super::ExecCommandRequest;
 use super::SessionEntry;
@@ -36,7 +39,6 @@ use super::generate_chunk_id;
 use super::resolve_max_tokens;
 use super::session::OutputBuffer;
 use super::session::UnifiedExecSession;
-use super::truncate_output_to_tokens;
 
 impl UnifiedExecSessionManager {
     pub(crate) async fn exec_command(
@@ -70,7 +72,7 @@ impl UnifiedExecSessionManager {
         let wall_time = Instant::now().saturating_duration_since(start);
 
         let text = String::from_utf8_lossy(&collected).to_string();
-        let (output, original_token_count) = truncate_output_to_tokens(&text, max_tokens);
+        let output = formatted_truncate_text(&text, TruncationPolicy::Tokens(max_tokens));
         let chunk_id = generate_chunk_id();
         let has_exited = session.has_exited();
         let stored_id = self
@@ -85,6 +87,8 @@ impl UnifiedExecSessionManager {
         // Only include a session_id in the response if the process is still alive.
         let session_id = if has_exited { None } else { Some(stored_id) };
 
+        let original_token_count = approx_token_count(&text);
+
         let response = UnifiedExecResponse {
             event_call_id: context.call_id.clone(),
             chunk_id,
@@ -92,7 +96,7 @@ impl UnifiedExecSessionManager {
             output,
             session_id,
             exit_code: exit_code.flatten(),
-            original_token_count,
+            original_token_count: Some(original_token_count),
             session_command: Some(request.command.clone()),
         };
 
@@ -175,7 +179,8 @@ impl UnifiedExecSessionManager {
         let wall_time = Instant::now().saturating_duration_since(start);
 
         let text = String::from_utf8_lossy(&collected).to_string();
-        let (output, original_token_count) = truncate_output_to_tokens(&text, max_tokens);
+        let output = formatted_truncate_text(&text, TruncationPolicy::Tokens(max_tokens));
+        let original_token_count = approx_token_count(&text);
         let chunk_id = generate_chunk_id();
 
         let status = self.refresh_session_state(session_id).await;
@@ -199,7 +204,7 @@ impl UnifiedExecSessionManager {
             output,
             session_id,
             exit_code,
-            original_token_count,
+            original_token_count: Some(original_token_count),
             session_command: Some(session_command.clone()),
         };
 

codex-rs/core/src/user_shell_command.rs

@@ -3,6 +3,7 @@ use std::time::Duration;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::ResponseItem;
 
+use crate::codex::TurnContext;
 use crate::exec::ExecToolCallOutput;
 use crate::tools::format_exec_output_str;
 
@@ -20,7 +21,11 @@ fn format_duration_line(duration: Duration) -> String {
     format!("Duration: {duration_seconds:.4} seconds")
 }
 
-fn format_user_shell_command_body(command: &str, exec_output: &ExecToolCallOutput) -> String {
+fn format_user_shell_command_body(
+    command: &str,
+    exec_output: &ExecToolCallOutput,
+    turn_context: &TurnContext,
+) -> String {
     let mut sections = Vec::new();
     sections.push("<command>".to_string());
     sections.push(command.to_string());
@@ -29,25 +34,33 @@ fn format_user_shell_command_body(command: &str, exec_output: &ExecToolCallOutpu
     sections.push(format!("Exit code: {}", exec_output.exit_code));
     sections.push(format_duration_line(exec_output.duration));
     sections.push("Output:".to_string());
-    sections.push(format_exec_output_str(exec_output));
+    sections.push(format_exec_output_str(
+        exec_output,
+        turn_context.truncation_policy,
+    ));
     sections.push("</result>".to_string());
     sections.join("\n")
 }
 
-pub fn format_user_shell_command_record(command: &str, exec_output: &ExecToolCallOutput) -> String {
-    let body = format_user_shell_command_body(command, exec_output);
+pub fn format_user_shell_command_record(
+    command: &str,
+    exec_output: &ExecToolCallOutput,
+    turn_context: &TurnContext,
+) -> String {
+    let body = format_user_shell_command_body(command, exec_output, turn_context);
     format!("{USER_SHELL_COMMAND_OPEN}\n{body}\n{USER_SHELL_COMMAND_CLOSE}")
 }
 
 pub fn user_shell_command_record_item(
     command: &str,
     exec_output: &ExecToolCallOutput,
+    turn_context: &TurnContext,
 ) -> ResponseItem {
     ResponseItem::Message {
         id: None,
         role: "user".to_string(),
         content: vec![ContentItem::InputText {
-            text: format_user_shell_command_record(command, exec_output),
+            text: format_user_shell_command_record(command, exec_output, turn_context),
         }],
     }
 }
@@ -55,6 +68,7 @@ pub fn user_shell_command_record_item(
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::codex::make_session_and_context;
     use crate::exec::StreamOutput;
     use pretty_assertions::assert_eq;
 
@@ -76,7 +90,8 @@ mod tests {
             duration: Duration::from_secs(1),
             timed_out: false,
         };
-        let item = user_shell_command_record_item("echo hi", &exec_output);
+        let (_, turn_context) = make_session_and_context();
+        let item = user_shell_command_record_item("echo hi", &exec_output, &turn_context);
         let ResponseItem::Message { content, .. } = item else {
             panic!("expected message");
         };
@@ -99,7 +114,8 @@ mod tests {
             duration: Duration::from_millis(120),
             timed_out: false,
         };
-        let record = format_user_shell_command_record("false", &exec_output);
+        let (_, turn_context) = make_session_and_context();
+        let record = format_user_shell_command_record("false", &exec_output, &turn_context);
         assert_eq!(
             record,
             "<user_shell_command>\n<command>\nfalse\n</command>\n<result>\nExit code: 42\nDuration: 0.1200 seconds\nOutput:\ncombined output wins\n</result>\n</user_shell_command>"