tests

### Summary
After #7022, we no longer need this warning. We should also clean up the
schema for the notification, but this is a quick fix to just stop the
behavior in the VSCE

## Testing
- [x] Ran locally

README.md

@@ -69,7 +69,38 @@ Codex can access MCP servers. To configure them, refer to the [config docs](./do
 
 Codex CLI supports a rich set of configuration options, with preferences stored in `~/.codex/config.toml`. For full configuration options, see [Configuration](./docs/config.md).
 
----
+### Execpolicy Quickstart
+
+Codex can enforce your own rules-based execution policy before it runs shell commands.
+
+1. Create a policy directory: `mkdir -p ~/.codex/policy`.
+2. Create one or more `.codexpolicy` files in that folder. Codex automatically loads every `.codexpolicy` file in there on startup.
+3. Write `prefix_rule` entries to describe the commands you want to allow, prompt, or block:
+
+```starlark
+prefix_rule(
+    pattern = ["git", ["push", "fetch"]],
+    decision = "prompt",  # allow | prompt | forbidden
+    match = [["git", "push", "origin", "main"]],  # examples that must match
+    not_match = [["git", "status"]],              # examples that must not match
+)
+```
+
+- `pattern` is a list of shell tokens, evaluated from left to right; wrap tokens in a nested list to express alternatives (e.g., match both `push` and `fetch`).
+- `decision` sets the severity; Codex picks the strictest decision when multiple rules match (forbidden > prompt > allow).
+- `match` and `not_match` act as (optional) unit tests. Codex validates them when it loads your policy, so you get feedback if an example has unexpected behavior.
+
+In this example rule, if Codex wants to run commands with the prefix `git push` or `git fetch`, it will first ask for user approval.
+
+Use the `codex execpolicy check` subcommand to preview decisions before you save a rule (see the [`codex-execpolicy` README](./codex-rs/execpolicy/README.md) for syntax details):
+
+```shell
+codex execpolicy check --policy ~/.codex/policy/default.codexpolicy git push origin main
+```
+
+Pass multiple `--policy` flags to test how several files combine, and use `--pretty` for formatted JSON output. See the [`codex-rs/execpolicy` README](./codex-rs/execpolicy/README.md) for a more detailed walkthrough of the available syntax.
+
+## Note: `execpolicy` commands are still in preview. The API may have breaking changes in the future.
 
 ### Docs & FAQ
 

codex-rs/.config/nextest.toml

@@ -7,3 +7,7 @@ slow-timeout = { period = "15s", terminate-after = 2 }
 # Do not add new tests here
 filter = 'test(rmcp_client) | test(humanlike_typing_1000_chars_appears_live_no_placeholder)'
 slow-timeout = { period = "1m", terminate-after = 4 }
+
+[[profile.default.overrides]]
+filter = 'test(approval_matrix_covers_all_modes)'
+slow-timeout = { period = "30s", terminate-after = 2 }

codex-rs/Cargo.lock

@@ -260,7 +260,7 @@ dependencies = [
  "memchr",
  "proc-macro2",
  "quote",
- "rustc-hash 2.1.1",
+ "rustc-hash",
  "serde",
  "serde_derive",
  "syn 2.0.104",
@@ -726,6 +726,17 @@ version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 
+[[package]]
+name = "chardetng"
+version = "0.1.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "14b8f0b65b7b08ae3c8187e8d77174de20cb6777864c6b832d8ad365999cf1ea"
+dependencies = [
+ "cfg-if",
+ "encoding_rs",
+ "memchr",
+]
+
 [[package]]
 name = "chrono"
 version = "0.4.42"
@@ -879,6 +890,7 @@ dependencies = [
  "serde",
  "serde_json",
  "strum_macros 0.27.2",
+ "thiserror 2.0.17",
  "ts-rs",
  "uuid",
 ]
@@ -989,6 +1001,7 @@ dependencies = [
  "codex-common",
  "codex-core",
  "codex-exec",
+ "codex-execpolicy",
  "codex-login",
  "codex-mcp-server",
  "codex-process-hardening",
@@ -1080,11 +1093,13 @@ dependencies = [
  "async-trait",
  "base64",
  "bytes",
+ "chardetng",
  "chrono",
  "codex-app-server-protocol",
  "codex-apply-patch",
  "codex-arg0",
  "codex-async-utils",
+ "codex-execpolicy",
  "codex-file-search",
  "codex-git",
  "codex-keyring-store",
@@ -1094,13 +1109,13 @@ dependencies = [
  "codex-utils-pty",
  "codex-utils-readiness",
  "codex-utils-string",
- "codex-utils-tokenizer",
  "codex-windows-sandbox",
  "core-foundation 0.9.4",
  "core_test_support",
  "ctor 0.5.0",
  "dirs",
  "dunce",
+ "encoding_rs",
  "env-flags",
  "escargot",
  "eventsource-stream",
@@ -1182,9 +1197,47 @@ dependencies = [
  "wiremock",
 ]
 
+[[package]]
+name = "codex-exec-server"
+version = "0.0.0"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "clap",
+ "codex-core",
+ "libc",
+ "path-absolutize",
+ "pretty_assertions",
+ "rmcp",
+ "serde",
+ "serde_json",
+ "shlex",
+ "socket2 0.6.0",
+ "tempfile",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "tracing-subscriber",
+]
+
 [[package]]
 name = "codex-execpolicy"
 version = "0.0.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "multimap",
+ "pretty_assertions",
+ "serde",
+ "serde_json",
+ "shlex",
+ "starlark",
+ "thiserror 2.0.17",
+]
+
+[[package]]
+name = "codex-execpolicy-legacy"
+version = "0.0.0"
 dependencies = [
  "allocative",
  "anyhow",
@@ -1202,21 +1255,6 @@ 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"
@@ -1366,6 +1404,7 @@ dependencies = [
  "codex-app-server-protocol",
  "codex-protocol",
  "eventsource-stream",
+ "http",
  "opentelemetry",
  "opentelemetry-otlp",
  "opentelemetry-semantic-conventions",
@@ -1399,6 +1438,7 @@ dependencies = [
  "icu_provider",
  "mcp-types",
  "mime_guess",
+ "pretty_assertions",
  "schemars 0.8.22",
  "serde",
  "serde_json",
@@ -1589,23 +1629,12 @@ dependencies = [
 name = "codex-utils-string"
 version = "0.0.0"
 
-[[package]]
-name = "codex-utils-tokenizer"
-version = "0.0.0"
-dependencies = [
- "anyhow",
- "codex-utils-cache",
- "pretty_assertions",
- "thiserror 2.0.17",
- "tiktoken-rs",
- "tokio",
-]
-
 [[package]]
 name = "codex-windows-sandbox"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "codex-protocol",
  "dirs-next",
  "dunce",
  "rand 0.8.5",
@@ -2421,17 +2450,6 @@ dependencies = [
  "once_cell",
 ]
 
-[[package]]
-name = "fancy-regex"
-version = "0.13.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "531e46835a22af56d1e3b66f04844bed63158bc094a628bec1d321d9b4c44bf2"
-dependencies = [
- "bit-set",
- "regex-automata",
- "regex-syntax 0.8.5",
-]
-
 [[package]]
 name = "fastrand"
 version = "2.3.0"
@@ -4756,7 +4774,7 @@ dependencies = [
  "pin-project-lite",
  "quinn-proto",
  "quinn-udp",
- "rustc-hash 2.1.1",
+ "rustc-hash",
  "rustls",
  "socket2 0.6.0",
  "thiserror 2.0.17",
@@ -4776,7 +4794,7 @@ dependencies = [
  "lru-slab",
  "rand 0.9.2",
  "ring",
- "rustc-hash 2.1.1",
+ "rustc-hash",
  "rustls",
  "rustls-pki-types",
  "slab",
@@ -5121,12 +5139,6 @@ version = "0.1.25"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "989e6739f80c4ad5b13e0fd7fe89531180375b18520cc8c82080e4dc4035b84f"
 
-[[package]]
-name = "rustc-hash"
-version = "1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2"
-
 [[package]]
 name = "rustc-hash"
 version = "2.1.1"
@@ -5174,6 +5186,7 @@ version = "0.23.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2491382039b29b9b11ff08b76ff6c97cf287671dbb74f0be44bda389fffe9bd1"
 dependencies = [
+ "log",
  "once_cell",
  "ring",
  "rustls-pki-types",
@@ -6346,21 +6359,6 @@ dependencies = [
  "zune-jpeg",
 ]
 
-[[package]]
-name = "tiktoken-rs"
-version = "0.9.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3a19830747d9034cd9da43a60eaa8e552dfda7712424aebf187b7a60126bae0d"
-dependencies = [
- "anyhow",
- "base64",
- "bstr",
- "fancy-regex",
- "lazy_static",
- "regex",
- "rustc-hash 1.1.0",
-]
-
 [[package]]
 name = "time"
 version = "0.3.44"
@@ -6601,8 +6599,10 @@ 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

@@ -16,8 +16,9 @@ members = [
     "common",
     "core",
     "exec",
+    "exec-server",
     "execpolicy",
-    "execpolicy2",
+    "execpolicy-legacy",
     "keyring-store",
     "file-search",
     "linux-sandbox",
@@ -40,7 +41,6 @@ members = [
     "utils/pty",
     "utils/readiness",
     "utils/string",
-    "utils/tokenizer",
 ]
 resolver = "2"
 
@@ -66,6 +66,7 @@ codex-chatgpt = { path = "chatgpt" }
 codex-common = { path = "common" }
 codex-core = { path = "core" }
 codex-exec = { path = "exec" }
+codex-execpolicy = { path = "execpolicy" }
 codex-feedback = { path = "feedback" }
 codex-file-search = { path = "file-search" }
 codex-git = { path = "utils/git" }
@@ -88,7 +89,6 @@ codex-utils-json-to-toml = { path = "utils/json-to-toml" }
 codex-utils-pty = { path = "utils/pty" }
 codex-utils-readiness = { path = "utils/readiness" }
 codex-utils-string = { path = "utils/string" }
-codex-utils-tokenizer = { path = "utils/tokenizer" }
 codex-windows-sandbox = { path = "windows-sandbox-rs" }
 core_test_support = { path = "core/tests/common" }
 mcp-types = { path = "mcp-types" }
@@ -109,6 +109,7 @@ axum = { version = "0.8", default-features = false }
 base64 = "0.22.1"
 bytes = "1.10.1"
 chrono = "0.4.42"
+chardetng = "0.1.17"
 clap = "4"
 clap_complete = "4"
 color-eyre = "0.6.3"
@@ -121,6 +122,7 @@ dotenvy = "0.15.7"
 dunce = "1.0.4"
 env-flags = "0.1.1"
 env_logger = "0.11.5"
+encoding_rs = "0.8.35"
 escargot = "0.5"
 eventsource-stream = "0.2.3"
 futures = { version = "0.3", default-features = false }
@@ -167,7 +169,6 @@ reqwest = "0.12"
 rmcp = { version = "0.8.5", default-features = false }
 schemars = "0.8.22"
 seccompiler = "0.5.0"
-sentry = "0.34.0"
 serde = "1"
 serde_json = "1"
 serde_with = "3.14"
@@ -176,6 +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"
@@ -185,7 +187,6 @@ tempfile = "3.23.0"
 test-log = "0.2.18"
 textwrap = "0.16.2"
 thiserror = "2.0.17"
-tiktoken-rs = "0.9"
 time = "0.3"
 tiny_http = "0.12"
 tokio = "1"
@@ -263,7 +264,6 @@ ignored = [
     "icu_provider",
     "openssl-sys",
     "codex-utils-readiness",
-    "codex-utils-tokenizer",
 ]
 
 [profile.release]

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

@@ -19,6 +19,7 @@ schemars = { workspace = true }
 serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }
 strum_macros = { workspace = true }
+thiserror = { workspace = true }
 ts-rs = { workspace = true }
 uuid = { workspace = true, features = ["serde", "v7"] }
 

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

@@ -129,6 +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,
@@ -374,7 +378,7 @@ macro_rules! server_notification_definitions {
         impl TryFrom<JSONRPCNotification> for ServerNotification {
             type Error = serde_json::Error;
 
-            fn try_from(value: JSONRPCNotification) -> Result<Self, Self::Error> {
+            fn try_from(value: JSONRPCNotification) -> Result<Self, serde_json::Error> {
                 serde_json::from_value(serde_json::to_value(value)?)
             }
         }
@@ -434,6 +438,13 @@ server_request_definitions! {
         response: v2::CommandExecutionRequestApprovalResponse,
     },
 
+    /// Sent when approval is requested for a specific file change.
+    /// This request is used for Turns started via turn/start.
+    FileChangeRequestApproval => "item/fileChange/requestApproval" {
+        params: v2::FileChangeRequestApprovalParams,
+        response: v2::FileChangeRequestApprovalResponse,
+    },
+
     /// DEPRECATED APIs below
     /// Request to approve a patch.
     /// This request is used for Turns started via the legacy APIs (i.e. SendUserTurn, SendUserMessage).
@@ -476,6 +487,7 @@ pub struct FuzzyFileSearchResponse {
 
 server_notification_definitions! {
     /// NEW NOTIFICATIONS
+    Error => "error" (v2::ErrorNotification),
     ThreadStarted => "thread/started" (v2::ThreadStartedNotification),
     TurnStarted => "turn/started" (v2::TurnStartedNotification),
     TurnCompleted => "turn/completed" (v2::TurnCompletedNotification),
@@ -490,6 +502,9 @@ 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")]
@@ -524,7 +539,7 @@ mod tests {
         let request = ClientRequest::NewConversation {
             request_id: RequestId::Integer(42),
             params: v1::NewConversationParams {
-                model: Some("gpt-5.1-codex".to_string()),
+                model: Some("gpt-5.1-codex-max".to_string()),
                 model_provider: None,
                 profile: None,
                 cwd: None,
@@ -542,7 +557,7 @@ mod tests {
                 "method": "newConversation",
                 "id": 42,
                 "params": {
-                    "model": "gpt-5.1-codex",
+                    "model": "gpt-5.1-codex-max",
                     "modelProvider": null,
                     "profile": null,
                     "cwd": null,

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/v2.rs

@@ -11,6 +11,8 @@ 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::CodexErrorInfo as CoreCodexErrorInfo;
+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;
@@ -19,6 +21,7 @@ use schemars::JsonSchema;
 use serde::Deserialize;
 use serde::Serialize;
 use serde_json::Value as JsonValue;
+use thiserror::Error;
 use ts_rs::TS;
 
 // Macro to declare a camelCased API v2 enum mirroring a core enum which
@@ -46,6 +49,72 @@ macro_rules! v2_enum_from_core {
     };
 }
 
+/// This translation layer make sure that we expose codex error code in camel case.
+///
+/// When an upstream HTTP status is available (for example, from the Responses API or a provider),
+/// it is forwarded in `httpStatusCode` on the relevant `codexErrorInfo` variant.
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Eq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub enum CodexErrorInfo {
+    ContextWindowExceeded,
+    UsageLimitExceeded,
+    HttpConnectionFailed {
+        #[serde(rename = "httpStatusCode")]
+        #[ts(rename = "httpStatusCode")]
+        http_status_code: Option<u16>,
+    },
+    /// Failed to connect to the response SSE stream.
+    ResponseStreamConnectionFailed {
+        #[serde(rename = "httpStatusCode")]
+        #[ts(rename = "httpStatusCode")]
+        http_status_code: Option<u16>,
+    },
+    InternalServerError,
+    Unauthorized,
+    BadRequest,
+    SandboxError,
+    /// The response SSE stream disconnected in the middle of a turn before completion.
+    ResponseStreamDisconnected {
+        #[serde(rename = "httpStatusCode")]
+        #[ts(rename = "httpStatusCode")]
+        http_status_code: Option<u16>,
+    },
+    /// Reached the retry limit for responses.
+    ResponseTooManyFailedAttempts {
+        #[serde(rename = "httpStatusCode")]
+        #[ts(rename = "httpStatusCode")]
+        http_status_code: Option<u16>,
+    },
+    Other,
+}
+
+impl From<CoreCodexErrorInfo> for CodexErrorInfo {
+    fn from(value: CoreCodexErrorInfo) -> Self {
+        match value {
+            CoreCodexErrorInfo::ContextWindowExceeded => CodexErrorInfo::ContextWindowExceeded,
+            CoreCodexErrorInfo::UsageLimitExceeded => CodexErrorInfo::UsageLimitExceeded,
+            CoreCodexErrorInfo::HttpConnectionFailed { http_status_code } => {
+                CodexErrorInfo::HttpConnectionFailed { http_status_code }
+            }
+            CoreCodexErrorInfo::ResponseStreamConnectionFailed { http_status_code } => {
+                CodexErrorInfo::ResponseStreamConnectionFailed { http_status_code }
+            }
+            CoreCodexErrorInfo::InternalServerError => CodexErrorInfo::InternalServerError,
+            CoreCodexErrorInfo::Unauthorized => CodexErrorInfo::Unauthorized,
+            CoreCodexErrorInfo::BadRequest => CodexErrorInfo::BadRequest,
+            CoreCodexErrorInfo::SandboxError => CodexErrorInfo::SandboxError,
+            CoreCodexErrorInfo::ResponseStreamDisconnected { http_status_code } => {
+                CodexErrorInfo::ResponseStreamDisconnected { http_status_code }
+            }
+            CoreCodexErrorInfo::ResponseTooManyFailedAttempts { http_status_code } => {
+                CodexErrorInfo::ResponseTooManyFailedAttempts { http_status_code }
+            }
+            CoreCodexErrorInfo::Other => CodexErrorInfo::Other,
+        }
+    }
+}
+
 v2_enum_from_core!(
     pub enum AskForApproval from codex_protocol::protocol::AskForApproval {
         UnlessTrusted, OnFailure, OnRequest, Never
@@ -402,6 +471,12 @@ 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)]
@@ -444,6 +519,12 @@ 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)]
@@ -505,6 +586,10 @@ 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)]
@@ -519,25 +604,37 @@ 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)]
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS, Error)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]
+#[error("{message}")]
 pub struct TurnError {
     pub message: String,
+    pub codex_error_info: Option<CodexErrorInfo>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]
+pub struct ErrorNotification {
+    pub error: TurnError,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(tag = "status", rename_all = "camelCase")]
+#[ts(tag = "status", export_to = "v2/")]
 pub enum TurnStatus {
     Completed,
     Interrupted,
-    Failed,
+    Failed { error: TurnError },
     InProgress,
 }
 
@@ -562,6 +659,45 @@ 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/")]
@@ -735,20 +871,23 @@ pub struct FileUpdateChange {
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-#[serde(rename_all = "camelCase")]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[ts(tag = "type")]
 #[ts(export_to = "v2/")]
 pub enum PatchChangeKind {
     Add,
     Delete,
-    Update,
+    Update { move_path: Option<PathBuf> },
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]
 pub enum PatchApplyStatus {
+    InProgress,
     Completed,
     Failed,
+    Declined,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -814,8 +953,6 @@ 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)]
@@ -883,6 +1020,15 @@ 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/")]
@@ -916,6 +1062,26 @@ pub struct CommandExecutionRequestApprovalResponse {
     pub accept_settings: Option<CommandExecutionRequestAcceptSettings>,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+#[ts(export_to = "v2/")]
+pub struct FileChangeRequestApprovalParams {
+    pub thread_id: String,
+    pub turn_id: String,
+    pub item_id: String,
+    /// Optional explanatory reason (e.g. request for extra write access).
+    pub reason: Option<String>,
+    /// [UNSTABLE] 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)]
+#[ts(export_to = "v2/")]
+pub struct FileChangeRequestApprovalResponse {
+    pub decision: ApprovalDecision,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 #[ts(export_to = "v2/")]
@@ -929,6 +1095,7 @@ pub struct AccountRateLimitsUpdatedNotification {
 pub struct RateLimitSnapshot {
     pub primary: Option<RateLimitWindow>,
     pub secondary: Option<RateLimitWindow>,
+    pub credits: Option<CreditsSnapshot>,
 }
 
 impl From<CoreRateLimitSnapshot> for RateLimitSnapshot {
@@ -936,6 +1103,7 @@ impl From<CoreRateLimitSnapshot> for RateLimitSnapshot {
         Self {
             primary: value.primary.map(RateLimitWindow::from),
             secondary: value.secondary.map(RateLimitWindow::from),
+            credits: value.credits.map(CreditsSnapshot::from),
         }
     }
 }
@@ -959,6 +1127,25 @@ 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/")]
@@ -981,6 +1168,7 @@ mod tests {
     use codex_protocol::items::WebSearchItem;
     use codex_protocol::user_input::UserInput as CoreUserInput;
     use pretty_assertions::assert_eq;
+    use serde_json::json;
     use std::path::PathBuf;
 
     #[test]
@@ -1066,4 +1254,20 @@ mod tests {
             }
         );
     }
+
+    #[test]
+    fn codex_error_info_serializes_http_status_code_in_camel_case() {
+        let value = CodexErrorInfo::ResponseTooManyFailedAttempts {
+            http_status_code: Some(401),
+        };
+
+        assert_eq!(
+            serde_json::to_value(value).unwrap(),
+            json!({
+                "responseTooManyFailedAttempts": {
+                    "httpStatusCode": 401
+                }
+            })
+        );
+    }
 }

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

@@ -17,15 +17,22 @@ 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::FileChangeRequestApprovalParams;
+use codex_app_server_protocol::FileChangeRequestApprovalResponse;
 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 +43,17 @@ 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 +512,9 @@ 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 +612,8 @@ 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 +633,8 @@ 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 +670,115 @@ 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)?;
+            }
+            ServerRequest::FileChangeRequestApproval { request_id, params } => {
+                self.approve_file_change_request(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 approve_file_change_request(
+        &mut self,
+        request_id: RequestId,
+        params: FileChangeRequestApprovalParams,
+    ) -> Result<()> {
+        let FileChangeRequestApprovalParams {
+            thread_id,
+            turn_id,
+            item_id,
+            reason,
+            grant_root,
+        } = params;
+
+        println!(
+            "\n< fileChange 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(grant_root) = grant_root.as_deref() {
+            println!("< grant root: {}", grant_root.display());
+        }
+
+        let response = FileChangeRequestApprovalResponse {
+            decision: ApprovalDecision::Accept,
+        };
+        self.send_server_request_response(request_id, &response)?;
+        println!("< approved fileChange 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/README.md

@@ -65,6 +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
 
@@ -181,6 +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.
@@ -286,6 +339,29 @@ Event notifications are the server-initiated event stream for thread lifecycles,
 
 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`.
 
+- `turn/started` — `{ turn }` with the turn id, empty `items`, and `status: "inProgress"`.
+- `turn/completed` — `{ turn }` where `turn.status` is `completed`, `interrupted`, or `failed`; failures carry `{ error: { message, codexErrorInfo? } }`.
+
+Today both notifications carry an empty `items` array even when item events were streamed; rely on `item/*` notifications for the canonical item list until this is fixed.
+
+#### Errors
+`error` event is emitted whenever the server hits an error mid-turn (for example, upstream model errors or quota limits). Carries the same `{ error: { message, codexErrorInfo? } }` payload as `turn.status: "failed"` and may precede that terminal notification.
+
+  `codexErrorInfo` maps to the `CodexErrorInfo` enum. Common values:
+  - `ContextWindowExceeded`
+  - `UsageLimitExceeded`
+  - `HttpConnectionFailed { httpStatusCode? }`: upstream HTTP failures including 4xx/5xx
+  - `ResponseStreamConnectionFailed { httpStatusCode? }`: failure to connect to the response SSE stream
+  - `ResponseStreamDisconnected { httpStatusCode? }`: disconnect of the response SSE stream in the middle of a turn before completion
+  - `ResponseTooManyFailedAttempts { httpStatusCode? }`
+  - `BadRequest`
+  - `Unauthorized`
+  - `SandboxError`
+  - `InternalServerError`
+  - `Other`: all unclassified errors
+
+When an upstream HTTP status is available (for example, from the Responses API or a provider), it is forwarded in `httpStatusCode` on the relevant `codexErrorInfo` variant.
+
 #### Thread items
 
 `ThreadItem` is the tagged union carried in turn responses and `item/*` notifications. Currently we support events for the following items:

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

@@ -1,24 +1,33 @@
 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::CodexErrorInfo as V2CodexErrorInfo;
 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::ErrorNotification;
 use codex_app_server_protocol::ExecCommandApprovalParams;
 use codex_app_server_protocol::ExecCommandApprovalResponse;
+use codex_app_server_protocol::FileChangeRequestApprovalParams;
+use codex_app_server_protocol::FileChangeRequestApprovalResponse;
+use codex_app_server_protocol::FileUpdateChange;
 use codex_app_server_protocol::InterruptConversationResponse;
 use codex_app_server_protocol::ItemCompletedNotification;
 use codex_app_server_protocol::ItemStartedNotification;
 use codex_app_server_protocol::McpToolCallError;
 use codex_app_server_protocol::McpToolCallResult;
 use codex_app_server_protocol::McpToolCallStatus;
+use codex_app_server_protocol::PatchApplyStatus;
+use codex_app_server_protocol::PatchChangeKind as V2PatchChangeKind;
 use codex_app_server_protocol::ReasoningSummaryPartAddedNotification;
 use codex_app_server_protocol::ReasoningSummaryTextDeltaNotification;
 use codex_app_server_protocol::ReasoningTextDeltaNotification;
@@ -26,7 +35,11 @@ use codex_app_server_protocol::SandboxCommandAssessment as V2SandboxCommandAsses
 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;
@@ -34,12 +47,17 @@ use codex_core::protocol::Event;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::ExecApprovalRequestEvent;
 use codex_core::protocol::ExecCommandEndEvent;
+use codex_core::protocol::FileChange as CoreFileChange;
 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::collections::HashMap;
 use std::convert::TryFrom;
+use std::path::PathBuf;
 use std::sync::Arc;
 use tokio::sync::oneshot;
 use tracing::error;
@@ -52,30 +70,84 @@ 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,
+            turn_id,
             changes,
             reason,
             grant_root,
-        }) => {
-            let params = ApplyPatchApprovalParams {
-                conversation_id,
-                call_id,
-                file_changes: changes,
-                reason,
-                grant_root,
-            };
-            let rx = outgoing
-                .send_request(ServerRequestPayload::ApplyPatchApproval(params))
-                .await;
-            tokio::spawn(async move {
-                on_patch_approval_response(event_id, rx, conversation).await;
-            });
-        }
+        }) => match api_version {
+            ApiVersion::V1 => {
+                let params = ApplyPatchApprovalParams {
+                    conversation_id,
+                    call_id,
+                    file_changes: changes.clone(),
+                    reason,
+                    grant_root,
+                };
+                let rx = outgoing
+                    .send_request(ServerRequestPayload::ApplyPatchApproval(params))
+                    .await;
+                tokio::spawn(async move {
+                    on_patch_approval_response(event_id, rx, conversation).await;
+                });
+            }
+            ApiVersion::V2 => {
+                // Until we migrate the core to be aware of a first class FileChangeItem
+                // and emit the corresponding EventMsg, we repurpose the call_id as the item_id.
+                let item_id = call_id.clone();
+                let patch_changes = convert_patch_changes(&changes);
+
+                let first_start = {
+                    let mut map = turn_summary_store.lock().await;
+                    let summary = map.entry(conversation_id).or_default();
+                    summary.file_change_started.insert(item_id.clone())
+                };
+                if first_start {
+                    let item = ThreadItem::FileChange {
+                        id: item_id.clone(),
+                        changes: patch_changes.clone(),
+                        status: PatchApplyStatus::InProgress,
+                    };
+                    let notification = ItemStartedNotification { item };
+                    outgoing
+                        .send_server_notification(ServerNotification::ItemStarted(notification))
+                        .await;
+                }
+
+                let params = FileChangeRequestApprovalParams {
+                    thread_id: conversation_id.to_string(),
+                    turn_id: turn_id.clone(),
+                    item_id: item_id.clone(),
+                    reason,
+                    grant_root,
+                };
+                let rx = outgoing
+                    .send_request(ServerRequestPayload::FileChangeRequestApproval(params))
+                    .await;
+                tokio::spawn(async move {
+                    on_file_change_request_approval_response(
+                        event_id,
+                        conversation_id,
+                        item_id,
+                        patch_changes,
+                        rx,
+                        conversation,
+                        outgoing,
+                        turn_summary_store,
+                    )
+                    .await;
+                });
+            }
+        },
         EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
             call_id,
             turn_id,
@@ -189,6 +261,42 @@ pub(crate) async fn apply_bespoke_event_handling(
                     .await;
             }
         }
+        EventMsg::Error(ev) => {
+            let turn_error = TurnError {
+                message: ev.message,
+                codex_error_info: ev.codex_error_info.map(V2CodexErrorInfo::from),
+            };
+            handle_error(conversation_id, turn_error.clone(), &turn_summary_store).await;
+            outgoing
+                .send_server_notification(ServerNotification::Error(ErrorNotification {
+                    error: turn_error,
+                }))
+                .await;
+        }
+        EventMsg::StreamError(ev) => {
+            // We don't need to update the turn summary store for stream errors as they are intermediate error states for retries,
+            // but we notify the client.
+            let turn_error = TurnError {
+                message: ev.message,
+                codex_error_info: ev.codex_error_info.map(V2CodexErrorInfo::from),
+            };
+            outgoing
+                .send_server_notification(ServerNotification::Error(ErrorNotification {
+                    error: turn_error,
+                }))
+                .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 };
@@ -203,6 +311,64 @@ 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::PatchApplyBegin(patch_begin_event) => {
+            // Until we migrate the core to be aware of a first class FileChangeItem
+            // and emit the corresponding EventMsg, we repurpose the call_id as the item_id.
+            let item_id = patch_begin_event.call_id.clone();
+
+            let first_start = {
+                let mut map = turn_summary_store.lock().await;
+                let summary = map.entry(conversation_id).or_default();
+                summary.file_change_started.insert(item_id.clone())
+            };
+            if first_start {
+                let item = ThreadItem::FileChange {
+                    id: item_id.clone(),
+                    changes: convert_patch_changes(&patch_begin_event.changes),
+                    status: PatchApplyStatus::InProgress,
+                };
+                let notification = ItemStartedNotification { item };
+                outgoing
+                    .send_server_notification(ServerNotification::ItemStarted(notification))
+                    .await;
+            }
+        }
+        EventMsg::PatchApplyEnd(patch_end_event) => {
+            // Until we migrate the core to be aware of a first class FileChangeItem
+            // and emit the corresponding EventMsg, we repurpose the call_id as the item_id.
+            let item_id = patch_end_event.call_id.clone();
+
+            let status = if patch_end_event.success {
+                PatchApplyStatus::Completed
+            } else {
+                PatchApplyStatus::Failed
+            };
+            let changes = convert_patch_changes(&patch_end_event.changes);
+            complete_file_change_item(
+                conversation_id,
+                item_id,
+                changes,
+                status,
+                outgoing.as_ref(),
+                &turn_summary_store,
+            )
+            .await;
+        }
         EventMsg::ExecCommandBegin(exec_command_begin_event) => {
             let item = ThreadItem::CommandExecution {
                 id: exec_command_begin_event.call_id.clone(),
@@ -298,12 +464,103 @@ 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 complete_file_change_item(
+    conversation_id: ConversationId,
+    item_id: String,
+    changes: Vec<FileUpdateChange>,
+    status: PatchApplyStatus,
+    outgoing: &OutgoingMessageSender,
+    turn_summary_store: &TurnSummaryStore,
+) {
+    {
+        let mut map = turn_summary_store.lock().await;
+        if let Some(summary) = map.get_mut(&conversation_id) {
+            summary.file_change_started.remove(&item_id);
+        }
+    }
+
+    let item = ThreadItem::FileChange {
+        id: item_id,
+        changes,
+        status,
+    };
+    let notification = ItemCompletedNotification { item };
+    outgoing
+        .send_server_notification(ServerNotification::ItemCompleted(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(error) = turn_summary.last_error {
+        TurnStatus::Failed { error }
+    } 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,
+    error: TurnError,
+    turn_summary_store: &TurnSummaryStore,
+) {
+    let mut map = turn_summary_store.lock().await;
+    map.entry(conversation_id).or_default().last_error = Some(error);
+}
+
 async fn on_patch_approval_response(
     event_id: String,
     receiver: oneshot::Receiver<JsonValue>,
@@ -382,6 +639,132 @@ 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")
+    }
+}
+
+fn convert_patch_changes(changes: &HashMap<PathBuf, CoreFileChange>) -> Vec<FileUpdateChange> {
+    let mut converted: Vec<FileUpdateChange> = changes
+        .iter()
+        .map(|(path, change)| FileUpdateChange {
+            path: path.to_string_lossy().into_owned(),
+            kind: map_patch_change_kind(change),
+            diff: format_file_change_diff(change),
+        })
+        .collect();
+    converted.sort_by(|a, b| a.path.cmp(&b.path));
+    converted
+}
+
+fn map_patch_change_kind(change: &CoreFileChange) -> V2PatchChangeKind {
+    match change {
+        CoreFileChange::Add { .. } => V2PatchChangeKind::Add,
+        CoreFileChange::Delete { .. } => V2PatchChangeKind::Delete,
+        CoreFileChange::Update { move_path, .. } => V2PatchChangeKind::Update {
+            move_path: move_path.clone(),
+        },
+    }
+}
+
+fn format_file_change_diff(change: &CoreFileChange) -> String {
+    match change {
+        CoreFileChange::Add { content } => content.clone(),
+        CoreFileChange::Delete { content } => content.clone(),
+        CoreFileChange::Update {
+            unified_diff,
+            move_path,
+        } => {
+            if let Some(path) = move_path {
+                format!("{unified_diff}\n\nMoved to: {}", path.display())
+            } else {
+                unified_diff.clone()
+            }
+        }
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn on_file_change_request_approval_response(
+    event_id: String,
+    conversation_id: ConversationId,
+    item_id: String,
+    changes: Vec<FileUpdateChange>,
+    receiver: oneshot::Receiver<JsonValue>,
+    codex: Arc<CodexConversation>,
+    outgoing: Arc<OutgoingMessageSender>,
+    turn_summary_store: TurnSummaryStore,
+) {
+    let response = receiver.await;
+    let (decision, completion_status) = match response {
+        Ok(value) => {
+            let response = serde_json::from_value::<FileChangeRequestApprovalResponse>(value)
+                .unwrap_or_else(|err| {
+                    error!("failed to deserialize FileChangeRequestApprovalResponse: {err}");
+                    FileChangeRequestApprovalResponse {
+                        decision: ApprovalDecision::Decline,
+                    }
+                });
+
+            let (decision, completion_status) = match response.decision {
+                ApprovalDecision::Accept => (ReviewDecision::Approved, None),
+                ApprovalDecision::Decline => {
+                    (ReviewDecision::Denied, Some(PatchApplyStatus::Declined))
+                }
+                ApprovalDecision::Cancel => {
+                    (ReviewDecision::Abort, Some(PatchApplyStatus::Declined))
+                }
+            };
+            // Allow EventMsg::PatchApplyEnd to emit ItemCompleted for accepted patches.
+            // Only short-circuit on declines/cancels/failures.
+            (decision, completion_status)
+        }
+        Err(err) => {
+            error!("request failed: {err:?}");
+            (ReviewDecision::Denied, Some(PatchApplyStatus::Failed))
+        }
+    };
+
+    if let Some(status) = completion_status {
+        complete_file_change_item(
+            conversation_id,
+            item_id,
+            changes,
+            status,
+            outgoing.as_ref(),
+            &turn_summary_store,
+        )
+        .await;
+    }
+
+    if let Err(err) = codex
+        .submit(Op::PatchApproval {
+            id: event_id,
+            decision,
+        })
+        .await
+    {
+        error!("failed to submit PatchApproval: {err}");
+    }
+}
+
 async fn on_command_execution_request_approval_response(
     event_id: String,
     receiver: oneshot::Receiver<JsonValue>,
@@ -486,13 +869,171 @@ 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,
+            TurnError {
+                message: "boom".to_string(),
+                codex_error_info: Some(V2CodexErrorInfo::InternalServerError),
+            },
+            &turn_summary_store,
+        )
+        .await;
+
+        let turn_summary = find_and_remove_turn_summary(conversation_id, &turn_summary_store).await;
+        assert_eq!(
+            turn_summary.last_error,
+            Some(TurnError {
+                message: "boom".to_string(),
+                codex_error_info: Some(V2CodexErrorInfo::InternalServerError),
+            })
+        );
+        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,
+            TurnError {
+                message: "oops".to_string(),
+                codex_error_info: None,
+            },
+            &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,
+            TurnError {
+                message: "bad".to_string(),
+                codex_error_info: Some(V2CodexErrorInfo::Other),
+            },
+            &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(),
+                            codex_error_info: Some(V2CodexErrorInfo::Other),
+                        }
+                    }
+                );
+            }
+            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() {
@@ -522,6 +1063,123 @@ 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,
+            TurnError {
+                message: "a1".to_string(),
+                codex_error_info: Some(V2CodexErrorInfo::BadRequest),
+            },
+            &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,
+            TurnError {
+                message: "b1".to_string(),
+                codex_error_info: None,
+            },
+            &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(),
+                            codex_error_info: Some(V2CodexErrorInfo::BadRequest),
+                        }
+                    }
+                );
+            }
+            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(),
+                            codex_error_info: None,
+                        }
+                    }
+                );
+            }
+            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;
@@ -81,6 +83,7 @@ use codex_app_server_protocol::ThreadStartParams;
 use codex_app_server_protocol::ThreadStartResponse;
 use codex_app_server_protocol::ThreadStartedNotification;
 use codex_app_server_protocol::Turn;
+use codex_app_server_protocol::TurnError;
 use codex_app_server_protocol::TurnInterruptParams;
 use codex_app_server_protocol::TurnStartParams;
 use codex_app_server_protocol::TurnStartResponse;
@@ -89,6 +92,7 @@ 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::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;
@@ -132,6 +139,7 @@ use codex_protocol::protocol::USER_MESSAGE_BEGIN;
 use codex_protocol::user_input::UserInput as CoreInputItem;
 use codex_utils_json_to_toml::json_to_toml;
 use std::collections::HashMap;
+use std::collections::HashSet;
 use std::ffi::OsStr;
 use std::io::Error as IoError;
 use std::path::Path;
@@ -151,6 +159,15 @@ 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) file_change_started: HashSet<String>,
+    pub(crate) last_error: Option<TurnError>,
+}
+
+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 +175,8 @@ struct ActiveLogin {
     login_id: Uuid,
 }
 
-impl ActiveLogin {
-    fn drop(&self) {
+impl Drop for ActiveLogin {
+    fn drop(&mut self) {
         self.shutdown_handle.shutdown();
     }
 }
@@ -175,6 +192,7 @@ 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 +245,97 @@ 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 +367,9 @@ 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 +524,7 @@ impl CodexMessageProcessor {
         {
             let mut guard = self.active_login.lock().await;
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
         }
 
@@ -525,7 +632,7 @@ 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 +722,7 @@ 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 +811,7 @@ 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 +865,7 @@ impl CodexMessageProcessor {
         {
             let mut guard = self.active_login.lock().await;
             if let Some(active) = guard.take() {
-                active.drop();
+                drop(active);
             }
         }
 
@@ -1063,11 +1170,13 @@ impl CodexMessageProcessor {
         let exec_params = ExecParams {
             command: params.command,
             cwd,
-            timeout_ms,
+            expiration: timeout_ms.into(),
             env,
             with_escalated_permissions: None,
             justification: None,
             arg0: None,
+            max_output_tokens: None,
+            max_output_chars: None,
         };
 
         let effective_policy = params
@@ -1135,7 +1244,7 @@ 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 +1256,17 @@ 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 {
@@ -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
@@ -1238,8 +1361,22 @@ 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.
@@ -1521,6 +1658,11 @@ 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, ApiVersion::V2)
@@ -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()),
         };
@@ -2272,9 +2436,6 @@ 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 +2475,8 @@ 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 +2499,64 @@ 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;
 
@@ -2441,6 +2656,7 @@ 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 {
@@ -2497,6 +2713,7 @@ impl CodexMessageProcessor {
                             conversation.clone(),
                             outgoing_for_task.clone(),
                             pending_interrupts.clone(),
+                            turn_summary_store.clone(),
                             api_version_for_task,
                         )
                         .await;
@@ -2791,6 +3008,7 @@ 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-5.1-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-5.1-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-5.1-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-5.1-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-5.1-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-5.1-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

@@ -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

@@ -1,22 +1,30 @@
 use anyhow::Result;
 use app_test_support::McpProcess;
+use app_test_support::create_apply_patch_sse_response;
 use app_test_support::create_final_assistant_message_sse_response;
 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::ApprovalDecision;
 use codex_app_server_protocol::CommandExecutionStatus;
+use codex_app_server_protocol::FileChangeRequestApprovalResponse;
+use codex_app_server_protocol::ItemCompletedNotification;
 use codex_app_server_protocol::ItemStartedNotification;
 use codex_app_server_protocol::JSONRPCNotification;
 use codex_app_server_protocol::JSONRPCResponse;
+use codex_app_server_protocol::PatchApplyStatus;
+use codex_app_server_protocol::PatchChangeKind;
 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;
@@ -57,7 +65,7 @@ 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 +126,17 @@ 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 +169,7 @@ 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.
@@ -233,7 +245,7 @@ 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 CommandExecutionRequestApproval request from server
     let first_turn_id = mcp
@@ -274,6 +286,11 @@ 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
@@ -302,6 +319,11 @@ 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??;
 
     Ok(())
 }
@@ -362,7 +384,7 @@ 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
@@ -455,6 +477,300 @@ async fn turn_start_updates_sandbox_and_cwd_between_turns_v2() -> Result<()> {
     Ok(())
 }
 
+#[tokio::test]
+async fn turn_start_file_change_approval_v2() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let tmp = TempDir::new()?;
+    let codex_home = tmp.path().join("codex_home");
+    std::fs::create_dir(&codex_home)?;
+    let workspace = tmp.path().join("workspace");
+    std::fs::create_dir(&workspace)?;
+
+    let patch = r#"*** Begin Patch
+*** Add File: README.md
++new line
+*** End Patch
+"#;
+    let responses = vec![
+        create_apply_patch_sse_response(patch, "patch-call")?,
+        create_final_assistant_message_sse_response("patch applied")?,
+    ];
+    let server = create_mock_chat_completions_server(responses).await;
+    create_config_toml(&codex_home, &server.uri(), "untrusted")?;
+
+    let mut mcp = McpProcess::new(&codex_home).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+
+    let start_req = mcp
+        .send_thread_start_request(ThreadStartParams {
+            model: Some("mock-model".to_string()),
+            cwd: Some(workspace.to_string_lossy().into_owned()),
+            ..Default::default()
+        })
+        .await?;
+    let start_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(start_req)),
+    )
+    .await??;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
+
+    let turn_req = mcp
+        .send_turn_start_request(TurnStartParams {
+            thread_id: thread.id.clone(),
+            input: vec![V2UserInput::Text {
+                text: "apply patch".into(),
+            }],
+            cwd: Some(workspace.clone()),
+            ..Default::default()
+        })
+        .await?;
+    let turn_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(turn_req)),
+    )
+    .await??;
+    let TurnStartResponse { turn } = to_response::<TurnStartResponse>(turn_resp)?;
+
+    let started_file_change = timeout(DEFAULT_READ_TIMEOUT, async {
+        loop {
+            let started_notif = mcp
+                .read_stream_until_notification_message("item/started")
+                .await?;
+            let started: ItemStartedNotification =
+                serde_json::from_value(started_notif.params.clone().expect("item/started params"))?;
+            if let ThreadItem::FileChange { .. } = started.item {
+                return Ok::<ThreadItem, anyhow::Error>(started.item);
+            }
+        }
+    })
+    .await??;
+    let ThreadItem::FileChange {
+        ref id,
+        status,
+        ref changes,
+    } = started_file_change
+    else {
+        unreachable!("loop ensures we break on file change items");
+    };
+    assert_eq!(id, "patch-call");
+    assert_eq!(status, PatchApplyStatus::InProgress);
+    let started_changes = changes.clone();
+
+    let server_req = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_request_message(),
+    )
+    .await??;
+    let ServerRequest::FileChangeRequestApproval { request_id, params } = server_req else {
+        panic!("expected FileChangeRequestApproval request")
+    };
+    assert_eq!(params.item_id, "patch-call");
+    assert_eq!(params.thread_id, thread.id);
+    assert_eq!(params.turn_id, turn.id);
+    let expected_readme_path = workspace.join("README.md");
+    let expected_readme_path = expected_readme_path.to_string_lossy().into_owned();
+    pretty_assertions::assert_eq!(
+        started_changes,
+        vec![codex_app_server_protocol::FileUpdateChange {
+            path: expected_readme_path.clone(),
+            kind: PatchChangeKind::Add,
+            diff: "new line\n".to_string(),
+        }]
+    );
+
+    mcp.send_response(
+        request_id,
+        serde_json::to_value(FileChangeRequestApprovalResponse {
+            decision: ApprovalDecision::Accept,
+        })?,
+    )
+    .await?;
+
+    let completed_file_change = timeout(DEFAULT_READ_TIMEOUT, async {
+        loop {
+            let completed_notif = mcp
+                .read_stream_until_notification_message("item/completed")
+                .await?;
+            let completed: ItemCompletedNotification = serde_json::from_value(
+                completed_notif
+                    .params
+                    .clone()
+                    .expect("item/completed params"),
+            )?;
+            if let ThreadItem::FileChange { .. } = completed.item {
+                return Ok::<ThreadItem, anyhow::Error>(completed.item);
+            }
+        }
+    })
+    .await??;
+    let ThreadItem::FileChange { ref id, status, .. } = completed_file_change else {
+        unreachable!("loop ensures we break on file change items");
+    };
+    assert_eq!(id, "patch-call");
+    assert_eq!(status, PatchApplyStatus::Completed);
+
+    timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("codex/event/task_complete"),
+    )
+    .await??;
+
+    let readme_contents = std::fs::read_to_string(expected_readme_path)?;
+    assert_eq!(readme_contents, "new line\n");
+
+    Ok(())
+}
+
+#[tokio::test]
+async fn turn_start_file_change_approval_decline_v2() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let tmp = TempDir::new()?;
+    let codex_home = tmp.path().join("codex_home");
+    std::fs::create_dir(&codex_home)?;
+    let workspace = tmp.path().join("workspace");
+    std::fs::create_dir(&workspace)?;
+
+    let patch = r#"*** Begin Patch
+*** Add File: README.md
++new line
+*** End Patch
+"#;
+    let responses = vec![
+        create_apply_patch_sse_response(patch, "patch-call")?,
+        create_final_assistant_message_sse_response("patch declined")?,
+    ];
+    let server = create_mock_chat_completions_server(responses).await;
+    create_config_toml(&codex_home, &server.uri(), "untrusted")?;
+
+    let mut mcp = McpProcess::new(&codex_home).await?;
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
+
+    let start_req = mcp
+        .send_thread_start_request(ThreadStartParams {
+            model: Some("mock-model".to_string()),
+            cwd: Some(workspace.to_string_lossy().into_owned()),
+            ..Default::default()
+        })
+        .await?;
+    let start_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(start_req)),
+    )
+    .await??;
+    let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
+
+    let turn_req = mcp
+        .send_turn_start_request(TurnStartParams {
+            thread_id: thread.id.clone(),
+            input: vec![V2UserInput::Text {
+                text: "apply patch".into(),
+            }],
+            cwd: Some(workspace.clone()),
+            ..Default::default()
+        })
+        .await?;
+    let turn_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(turn_req)),
+    )
+    .await??;
+    let TurnStartResponse { turn } = to_response::<TurnStartResponse>(turn_resp)?;
+
+    let started_file_change = timeout(DEFAULT_READ_TIMEOUT, async {
+        loop {
+            let started_notif = mcp
+                .read_stream_until_notification_message("item/started")
+                .await?;
+            let started: ItemStartedNotification =
+                serde_json::from_value(started_notif.params.clone().expect("item/started params"))?;
+            if let ThreadItem::FileChange { .. } = started.item {
+                return Ok::<ThreadItem, anyhow::Error>(started.item);
+            }
+        }
+    })
+    .await??;
+    let ThreadItem::FileChange {
+        ref id,
+        status,
+        ref changes,
+    } = started_file_change
+    else {
+        unreachable!("loop ensures we break on file change items");
+    };
+    assert_eq!(id, "patch-call");
+    assert_eq!(status, PatchApplyStatus::InProgress);
+    let started_changes = changes.clone();
+
+    let server_req = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_request_message(),
+    )
+    .await??;
+    let ServerRequest::FileChangeRequestApproval { request_id, params } = server_req else {
+        panic!("expected FileChangeRequestApproval request")
+    };
+    assert_eq!(params.item_id, "patch-call");
+    assert_eq!(params.thread_id, thread.id);
+    assert_eq!(params.turn_id, turn.id);
+    let expected_readme_path = workspace.join("README.md");
+    let expected_readme_path_str = expected_readme_path.to_string_lossy().into_owned();
+    pretty_assertions::assert_eq!(
+        started_changes,
+        vec![codex_app_server_protocol::FileUpdateChange {
+            path: expected_readme_path_str.clone(),
+            kind: PatchChangeKind::Add,
+            diff: "new line\n".to_string(),
+        }]
+    );
+
+    mcp.send_response(
+        request_id,
+        serde_json::to_value(FileChangeRequestApprovalResponse {
+            decision: ApprovalDecision::Decline,
+        })?,
+    )
+    .await?;
+
+    let completed_file_change = timeout(DEFAULT_READ_TIMEOUT, async {
+        loop {
+            let completed_notif = mcp
+                .read_stream_until_notification_message("item/completed")
+                .await?;
+            let completed: ItemCompletedNotification = serde_json::from_value(
+                completed_notif
+                    .params
+                    .clone()
+                    .expect("item/completed params"),
+            )?;
+            if let ThreadItem::FileChange { .. } = completed.item {
+                return Ok::<ThreadItem, anyhow::Error>(completed.item);
+            }
+        }
+    })
+    .await??;
+    let ThreadItem::FileChange { ref id, status, .. } = completed_file_change else {
+        unreachable!("loop ensures we break on file change items");
+    };
+    assert_eq!(id, "patch-call");
+    assert_eq!(status, PatchApplyStatus::Declined);
+
+    timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("codex/event/task_complete"),
+    )
+    .await??;
+
+    assert!(
+        !expected_readme_path.exists(),
+        "declined patch should not be applied"
+    );
+
+    Ok(())
+}
+
 // Helper to create a config.toml pointing at the mock model server.
 fn create_config_toml(
     codex_home: &Path,

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/Cargo.toml

@@ -26,6 +26,7 @@ codex-cloud-tasks = { path = "../cloud-tasks" }
 codex-common = { workspace = true, features = ["cli"] }
 codex-core = { workspace = true }
 codex-exec = { workspace = true }
+codex-execpolicy = { workspace = true }
 codex-login = { workspace = true }
 codex-mcp-server = { workspace = true }
 codex-process-hardening = { workspace = true }

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();
@@ -153,7 +149,7 @@ 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,

codex-rs/cli/src/main.rs

@@ -18,6 +18,7 @@ use codex_cli::login::run_logout;
 use codex_cloud_tasks::Cli as CloudTasksCli;
 use codex_common::CliConfigOverrides;
 use codex_exec::Cli as ExecCli;
+use codex_execpolicy::ExecPolicyCheckCommand;
 use codex_responses_api_proxy::Args as ResponsesApiProxyArgs;
 use codex_tui::AppExitInfo;
 use codex_tui::Cli as TuiCli;
@@ -93,6 +94,10 @@ enum Subcommand {
     #[clap(visible_alias = "debug")]
     Sandbox(SandboxArgs),
 
+    /// Execpolicy tooling.
+    #[clap(hide = true)]
+    Execpolicy(ExecpolicyCommand),
+
     /// Apply the latest diff produced by Codex agent as a `git apply` to your local working tree.
     #[clap(visible_alias = "a")]
     Apply(ApplyCommand),
@@ -134,6 +139,10 @@ 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,
 }
@@ -158,6 +167,19 @@ enum SandboxCommand {
     Windows(WindowsCommand),
 }
 
+#[derive(Debug, Parser)]
+struct ExecpolicyCommand {
+    #[command(subcommand)]
+    sub: ExecpolicySubcommand,
+}
+
+#[derive(Debug, clap::Subcommand)]
+enum ExecpolicySubcommand {
+    /// Check execpolicy files against a command.
+    #[clap(name = "check")]
+    Check(ExecPolicyCheckCommand),
+}
+
 #[derive(Debug, Parser)]
 struct LoginCommand {
     #[clap(skip)]
@@ -323,6 +345,10 @@ fn run_update_action(action: UpdateAction) -> anyhow::Result<()> {
     Ok(())
 }
 
+fn run_execpolicycheck(cmd: ExecPolicyCheckCommand) -> anyhow::Result<()> {
+    cmd.run()
+}
+
 #[derive(Debug, Default, Parser, Clone)]
 struct FeatureToggles {
     /// Enable a feature (repeatable). Equivalent to `-c features.<name>=true`.
@@ -448,6 +474,7 @@ 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 +482,7 @@ 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?;
@@ -543,6 +571,9 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
                 .await?;
             }
         },
+        Some(Subcommand::Execpolicy(ExecpolicyCommand { sub })) => match sub {
+            ExecpolicySubcommand::Check(cmd) => run_execpolicycheck(cmd)?,
+        },
         Some(Subcommand::Apply(mut apply_cli)) => {
             prepend_config_flags(
                 &mut apply_cli.config_overrides,
@@ -611,6 +642,7 @@ 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 +651,7 @@ 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 +735,21 @@ 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 {
@@ -775,6 +816,7 @@ 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 +825,7 @@ 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 +834,14 @@ 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]

codex-rs/cli/src/mcp_cmd.rs

@@ -79,6 +79,7 @@ pub struct GetArgs {
 }
 
 #[derive(Debug, clap::Parser)]
+#[command(override_usage = "codex mcp add [OPTIONS] <NAME> (--url <URL> | -- <COMMAND>...)")]
 pub struct AddArgs {
     /// Name for the MCP server configuration.
     pub name: String,

codex-rs/cli/tests/execpolicy.rs

@@ -0,0 +1,58 @@
+use std::fs;
+
+use assert_cmd::Command;
+use pretty_assertions::assert_eq;
+use serde_json::json;
+use tempfile::TempDir;
+
+#[test]
+fn execpolicy_check_matches_expected_json() -> Result<(), Box<dyn std::error::Error>> {
+    let codex_home = TempDir::new()?;
+    let policy_path = codex_home.path().join("policy.codexpolicy");
+    fs::write(
+        &policy_path,
+        r#"
+prefix_rule(
+    pattern = ["git", "push"],
+    decision = "forbidden",
+)
+"#,
+    )?;
+
+    let output = Command::cargo_bin("codex")?
+        .env("CODEX_HOME", codex_home.path())
+        .args([
+            "execpolicy",
+            "check",
+            "--policy",
+            policy_path
+                .to_str()
+                .expect("policy path should be valid UTF-8"),
+            "git",
+            "push",
+            "origin",
+            "main",
+        ])
+        .output()?;
+
+    assert!(output.status.success());
+    let result: serde_json::Value = serde_json::from_slice(&output.stdout)?;
+    assert_eq!(
+        result,
+        json!({
+            "match": {
+                "decision": "forbidden",
+                "matchedRules": [
+                    {
+                        "prefixRuleMatch": {
+                            "matchedPrefix": ["git", "push"],
+                            "decision": "forbidden"
+                        }
+                    }
+                ]
+            }
+        })
+    );
+
+    Ok(())
+}

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/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/config_summary.rs

@@ -15,13 +15,12 @@ pub fn create_config_summary_entries(config: &Config) -> Vec<(&'static str, Stri
     if config.model_provider.wire_api == WireApi::Responses
         && config.model_family.supports_reasoning_summaries
     {
-        entries.push((
-            "reasoning effort",
-            config
-                .model_reasoning_effort
-                .map(|effort| effort.to_string())
-                .unwrap_or_else(|| "none".to_string()),
-        ));
+        let reasoning_effort = config
+            .model_reasoning_effort
+            .or(config.model_family.default_reasoning_effort)
+            .map(|effort| effort.to_string())
+            .unwrap_or_else(|| "none".to_string());
+        entries.push(("reasoning effort", reasoning_effort));
         entries.push((
             "reasoning summaries",
             config.model_reasoning_summary.to_string(),

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/core/Cargo.toml

@@ -19,9 +19,11 @@ async-trait = { workspace = true }
 base64 = { workspace = true }
 bytes = { workspace = true }
 chrono = { workspace = true, features = ["serde"] }
+chardetng = { workspace = true }
 codex-app-server-protocol = { workspace = true }
 codex-apply-patch = { workspace = true }
 codex-async-utils = { workspace = true }
+codex-execpolicy = { workspace = true }
 codex-file-search = { workspace = true }
 codex-git = { workspace = true }
 codex-keyring-store = { workspace = true }
@@ -31,11 +33,11 @@ codex-rmcp-client = { workspace = true }
 codex-utils-pty = { workspace = true }
 codex-utils-readiness = { workspace = true }
 codex-utils-string = { workspace = true }
-codex-utils-tokenizer = { workspace = true }
 codex-windows-sandbox = { package = "codex-windows-sandbox", path = "../windows-sandbox-rs" }
 dirs = { workspace = true }
 dunce = { workspace = true }
 env-flags = { workspace = true }
+encoding_rs = { workspace = true }
 eventsource-stream = { workspace = true }
 futures = { workspace = true }
 http = { 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/src/bash.rs

@@ -100,7 +100,7 @@ pub fn extract_bash_command(command: &[String]) -> Option<(&str, &str)> {
     if !matches!(flag.as_str(), "-lc" | "-c")
         || !matches!(
             detect_shell_type(&PathBuf::from(shell)),
-            Some(ShellType::Zsh) | Some(ShellType::Bash)
+            Some(ShellType::Zsh) | Some(ShellType::Bash) | Some(ShellType::Sh)
         )
     {
         return None;

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;
             }

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()
 }

codex-rs/core/src/client_common.rs

@@ -136,7 +136,7 @@ fn reserialize_shell_outputs(items: &mut [ResponseItem]) {
 }
 
 fn is_shell_tool_name(name: &str) -> bool {
-    matches!(name, "shell" | "container.exec" | "shell_command")
+    matches!(name, "shell" | "container.exec")
 }
 
 #[derive(Deserialize)]
@@ -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)]
@@ -431,7 +422,7 @@ mod tests {
                 expects_apply_patch_instructions: false,
             },
             InstructionsTestCase {
-                slug: "gpt-5.1-codex",
+                slug: "gpt-5.1-codex-max",
                 expects_apply_patch_instructions: false,
             },
         ];

codex-rs/core/src/codex.rs

@@ -7,12 +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::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;
@@ -55,6 +59,7 @@ 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::ShellEnvironmentPolicy;
 use crate::context_manager::ContextManager;
@@ -63,10 +68,6 @@ use crate::error::CodexErr;
 use crate::error::Result as CodexResult;
 #[cfg(test)]
 use crate::exec::StreamOutput;
-// Removed: legacy executor wiring replaced by ToolOrchestrator flows.
-// legacy normalize_exec_result no longer used after orchestrator migration
-use crate::compact::build_compacted_history;
-use crate::compact::collect_user_messages;
 use crate::mcp::auth::compute_auth_statuses;
 use crate::mcp_connection_manager::McpConnectionManager;
 use crate::model_family::find_family_for_model;
@@ -78,7 +79,6 @@ 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;
@@ -120,6 +120,7 @@ use crate::user_instructions::UserInstructions;
 use crate::user_notification::UserNotification;
 use crate::util::backoff;
 use codex_async_utils::OrCancelExt;
+use codex_execpolicy::Policy as ExecPolicy;
 use codex_otel::otel_event_manager::OtelEventManager;
 use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
 use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
@@ -127,11 +128,11 @@ use codex_protocol::models::ContentItem;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseInputItem;
 use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::CodexErrorInfo;
 use codex_protocol::protocol::InitialHistory;
 use codex_protocol::user_input::UserInput;
 use codex_utils_readiness::Readiness;
 use codex_utils_readiness::ReadinessFlag;
-use codex_utils_tokenizer::warm_model_cache;
 
 /// The high-level interface to the Codex system.
 /// It operates as a queue pair where you send submissions and receive events.
@@ -165,6 +166,10 @@ impl Codex {
 
         let user_instructions = get_user_instructions(&config).await;
 
+        let exec_policy = crate::exec_policy::exec_policy_for(&config.features, &config.codex_home)
+            .await
+            .map_err(|err| CodexErr::Fatal(format!("failed to load execpolicy: {err}")))?;
+
         let config = Arc::new(config);
 
         let session_configuration = SessionConfiguration {
@@ -181,6 +186,7 @@ impl Codex {
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
             features: config.features.clone(),
+            exec_policy,
             session_source,
         };
 
@@ -278,6 +284,8 @@ pub(crate) struct TurnContext {
     pub(crate) final_output_json_schema: Option<Value>,
     pub(crate) codex_linux_sandbox_exe: Option<PathBuf>,
     pub(crate) tool_call_gate: Arc<ReadinessFlag>,
+    pub(crate) exec_policy: Arc<ExecPolicy>,
+    pub(crate) truncation_policy: TruncationPolicy,
 }
 
 impl TurnContext {
@@ -294,7 +302,6 @@ impl TurnContext {
     }
 }
 
-#[allow(dead_code)]
 #[derive(Clone)]
 pub(crate) struct SessionConfiguration {
     /// Provider identifier ("openai", "openrouter", ...).
@@ -334,6 +341,8 @@ pub(crate) struct SessionConfiguration {
 
     /// Set of feature flags for this session
     features: Features,
+    /// Execpolicy policy, applied only when enabled by feature flag.
+    exec_policy: Arc<ExecPolicy>,
 
     // TODO(pakrym): Remove config from here
     original_config_do_not_use: Arc<Config>,
@@ -404,7 +413,7 @@ impl Session {
         );
 
         let client = ModelClient::new(
-            Arc::new(per_turn_config),
+            Arc::new(per_turn_config.clone()),
             auth_manager,
             otel_event_manager,
             provider,
@@ -434,6 +443,8 @@ impl Session {
             final_output_json_schema: None,
             codex_linux_sandbox_exe: config.codex_linux_sandbox_exe.clone(),
             tool_call_gate: Arc::new(ReadinessFlag::new()),
+            exec_policy: session_configuration.exec_policy.clone(),
+            truncation_policy: TruncationPolicy::new(&per_turn_config),
         }
     }
 
@@ -481,7 +492,7 @@ impl Session {
         // - load history metadata
         let rollout_fut = RolloutRecorder::new(&config, rollout_params);
 
-        let default_shell_fut = shell::default_user_shell();
+        let default_shell = shell::default_user_shell();
         let history_meta_fut = crate::message_history::history_metadata(&config);
         let auth_statuses_fut = compute_auth_statuses(
             config.mcp_servers.iter(),
@@ -489,12 +500,8 @@ impl Session {
         );
 
         // Join all independent futures.
-        let (rollout_recorder, default_shell, (history_log_id, history_entry_count), auth_statuses) = tokio::join!(
-            rollout_fut,
-            default_shell_fut,
-            history_meta_fut,
-            auth_statuses_fut
-        );
+        let (rollout_recorder, (history_log_id, history_entry_count), auth_statuses) =
+            tokio::join!(rollout_fut, history_meta_fut, auth_statuses_fut);
 
         let rollout_recorder = rollout_recorder.map_err(|e| {
             error!("failed to initialize rollout recorder: {e:#}");
@@ -547,9 +554,6 @@ impl Session {
         // Create the mutable state for the Session.
         let state = SessionState::new(session_configuration.clone());
 
-        // Warm the tokenizer cache for the session model without blocking startup.
-        warm_model_cache(&session_configuration.model);
-
         let services = SessionServices {
             mcp_connection_manager: Arc::new(RwLock::new(McpConnectionManager::default())),
             mcp_startup_cancellation_token: CancellationToken::new(),
@@ -581,6 +585,10 @@ 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,
@@ -681,7 +689,8 @@ 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)
@@ -902,6 +911,7 @@ impl Session {
 
         let event = EventMsg::ApplyPatchApprovalRequest(ApplyPatchApprovalRequestEvent {
             call_id,
+            turn_id: turn_context.sub_id.clone(),
             changes,
             reason,
             grant_root,
@@ -938,7 +948,7 @@ 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;
     }
@@ -952,17 +962,25 @@ 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);
+                    }
                 }
                 _ => {}
             }
@@ -971,9 +989,13 @@ 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>) {
@@ -990,6 +1012,15 @@ 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(
@@ -1018,7 +1049,7 @@ impl Session {
             Some(turn_context.cwd.clone()),
             Some(turn_context.approval_policy),
             Some(turn_context.sandbox_policy.clone()),
-            Some(self.user_shell().clone()),
+            self.user_shell().clone(),
         )));
         items
     }
@@ -1057,11 +1088,14 @@ 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 {
@@ -1155,9 +1189,14 @@ impl Session {
         &self,
         turn_context: &TurnContext,
         message: impl Into<String>,
+        codex_error: CodexErr,
     ) {
+        let codex_error_info = CodexErrorInfo::ResponseStreamDisconnected {
+            http_status_code: codex_error.http_status_code_value(),
+        };
         let event = EventMsg::StreamError(StreamErrorEvent {
             message: message.into(),
+            codex_error_info: Some(codex_error_info),
         });
         self.send_event(turn_context, event).await;
     }
@@ -1167,14 +1206,7 @@ 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 {
@@ -1311,7 +1343,10 @@ impl Session {
 }
 
 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");
@@ -1407,6 +1442,7 @@ mod handlers {
     use crate::tasks::UndoTask;
     use crate::tasks::UserShellCommandTask;
     use codex_protocol::custom_prompts::CustomPrompt;
+    use codex_protocol::protocol::CodexErrorInfo;
     use codex_protocol::protocol::ErrorEvent;
     use codex_protocol::protocol::Event;
     use codex_protocol::protocol::EventMsg;
@@ -1415,6 +1451,7 @@ 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;
@@ -1623,16 +1660,15 @@ 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 {
@@ -1653,6 +1689,7 @@ mod handlers {
                 id: sub_id.clone(),
                 msg: EventMsg::Error(ErrorEvent {
                     message: "Failed to shutdown rollout recorder".to_string(),
+                    codex_error_info: Some(CodexErrorInfo::Other),
                 }),
             };
             sess.send_event_raw(event).await;
@@ -1758,6 +1795,8 @@ async fn spawn_review_thread(
         final_output_json_schema: None,
         codex_linux_sandbox_exe: parent_turn_context.codex_linux_sandbox_exe.clone(),
         tool_call_gate: Arc::new(ReadinessFlag::new()),
+        exec_policy: parent_turn_context.exec_policy.clone(),
+        truncation_policy: TruncationPolicy::new(&per_turn_config),
     };
 
     // Seed the child task with the review prompt as the initial user message.
@@ -1765,7 +1804,12 @@ 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))
@@ -1866,7 +1910,12 @@ pub(crate) async fn run_task(
 
                 // 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 {
-                    compact::run_inline_auto_compact_task(sess.clone(), turn_context.clone()).await;
+                    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;
+                    }
                     continue;
                 }
 
@@ -1895,9 +1944,7 @@ pub(crate) async fn run_task(
             }
             Err(e) => {
                 info!("Turn error: {e:#}");
-                let event = EventMsg::Error(ErrorEvent {
-                    message: e.to_string(),
-                });
+                let event = EventMsg::Error(e.to_error_event(None));
                 sess.send_event(&turn_context, event).await;
                 // let the user continue the conversation
                 break;
@@ -1937,12 +1984,32 @@ async fn run_turn(
         .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(),
     };
 
@@ -2002,6 +2069,7 @@ async fn run_turn(
                     sess.notify_stream_error(
                         &turn_context,
                         format!("Reconnecting... {retries}/{max_retries}"),
+                        e,
                     )
                     .await;
 
@@ -2320,6 +2388,7 @@ mod tests {
     use crate::config::ConfigOverrides;
     use crate::config::ConfigToml;
     use crate::exec::ExecToolCallOutput;
+    use crate::shell::default_user_shell;
     use crate::tools::format_exec_output_str;
 
     use crate::protocol::CompactedItem;
@@ -2429,8 +2498,9 @@ 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,
@@ -2546,6 +2616,7 @@ mod tests {
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
             features: Features::default(),
+            exec_policy: Arc::new(ExecPolicy::empty()),
             session_source: SessionSource::Exec,
         };
 
@@ -2557,7 +2628,7 @@ mod tests {
             unified_exec_manager: UnifiedExecSessionManager::default(),
             notifier: UserNotifier::new(None),
             rollout: Mutex::new(None),
-            user_shell: shell::Shell::Unknown,
+            user_shell: default_user_shell(),
             show_raw_agent_reasoning: config.show_raw_agent_reasoning,
             auth_manager: Arc::clone(&auth_manager),
             otel_event_manager: otel_event_manager.clone(),
@@ -2623,6 +2694,7 @@ mod tests {
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
             features: Features::default(),
+            exec_policy: Arc::new(ExecPolicy::empty()),
             session_source: SessionSource::Exec,
         };
 
@@ -2634,7 +2706,7 @@ mod tests {
             unified_exec_manager: UnifiedExecSessionManager::default(),
             notifier: UserNotifier::new(None),
             rollout: Mutex::new(None),
-            user_shell: shell::Shell::Unknown,
+            user_shell: default_user_shell(),
             show_raw_agent_reasoning: config.show_raw_agent_reasoning,
             auth_manager: Arc::clone(&auth_manager),
             otel_event_manager: otel_event_manager.clone(),
@@ -2753,7 +2825,8 @@ 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;
 
@@ -2871,7 +2944,7 @@ 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,
@@ -2880,7 +2953,7 @@ 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 {
@@ -2890,13 +2963,13 @@ 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,
@@ -2904,6 +2977,7 @@ mod tests {
         live_history.replace(rebuilt1);
         rollout_items.push(RolloutItem::Compacted(CompactedItem {
             message: summary1.to_string(),
+            replacement_history: None,
         }));
 
         let user2 = ResponseItem::Message {
@@ -2913,7 +2987,7 @@ 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 {
@@ -2923,13 +2997,13 @@ 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,
@@ -2937,6 +3011,7 @@ mod tests {
         live_history.replace(rebuilt2);
         rollout_items.push(RolloutItem::Compacted(CompactedItem {
             message: summary2.to_string(),
+            replacement_history: None,
         }));
 
         let user3 = ResponseItem::Message {
@@ -2946,7 +3021,7 @@ 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 {
@@ -2956,7 +3031,7 @@ 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())
@@ -2976,6 +3051,7 @@ mod tests {
         let session = Arc::new(session);
         let mut turn_context = Arc::new(turn_context_raw);
 
+        let timeout_ms = 1000;
         let params = ExecParams {
             command: if cfg!(windows) {
                 vec![
@@ -2991,16 +3067,25 @@ mod tests {
                 ]
             },
             cwd: turn_context.cwd.clone(),
-            timeout_ms: Some(1000),
+            expiration: timeout_ms.into(),
             env: HashMap::new(),
             with_escalated_permissions: Some(true),
             justification: Some("test".to_string()),
             arg0: None,
+            max_output_tokens: None,
+            max_output_chars: None,
         };
 
         let params2 = ExecParams {
             with_escalated_permissions: Some(false),
-            ..params.clone()
+            command: params.command.clone(),
+            cwd: params.cwd.clone(),
+            expiration: timeout_ms.into(),
+            env: HashMap::new(),
+            justification: params.justification.clone(),
+            arg0: None,
+            max_output_tokens: None,
+            max_output_chars: None,
         };
 
         let turn_diff_tracker = Arc::new(tokio::sync::Mutex::new(TurnDiffTracker::new()));
@@ -3020,7 +3105,7 @@ mod tests {
                     arguments: serde_json::json!({
                         "command": params.command.clone(),
                         "workdir": Some(turn_context.cwd.to_string_lossy().to_string()),
-                        "timeout_ms": params.timeout_ms,
+                        "timeout_ms": params.expiration.timeout_ms(),
                         "with_escalated_permissions": params.with_escalated_permissions,
                         "justification": params.justification.clone(),
                     })
@@ -3057,7 +3142,7 @@ mod tests {
                     arguments: serde_json::json!({
                         "command": params2.command.clone(),
                         "workdir": Some(turn_context.cwd.to_string_lossy().to_string()),
-                        "timeout_ms": params2.timeout_ms,
+                        "timeout_ms": params2.expiration.timeout_ms(),
                         "with_escalated_permissions": params2.with_escalated_permissions,
                         "justification": params2.justification.clone(),
                     })

codex-rs/core/src/command_safety/is_dangerous_command.rs

@@ -1,6 +1,8 @@
 use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::SandboxPolicy;
 
+use crate::sandboxing::SandboxPermissions;
+
 use crate::bash::parse_shell_lc_plain_commands;
 use crate::is_safe_command::is_known_safe_command;
 
@@ -8,7 +10,7 @@ pub fn requires_initial_appoval(
     policy: AskForApproval,
     sandbox_policy: &SandboxPolicy,
     command: &[String],
-    with_escalated_permissions: bool,
+    sandbox_permissions: SandboxPermissions,
 ) -> bool {
     if is_known_safe_command(command) {
         return false;
@@ -24,8 +26,7 @@ pub fn requires_initial_appoval(
             // In restricted sandboxes (ReadOnly/WorkspaceWrite), do not prompt for
             // non‑escalated, non‑dangerous commands — let the sandbox enforce
             // restrictions (e.g., block network/write) without a user prompt.
-            let wants_escalation: bool = with_escalated_permissions;
-            if wants_escalation {
+            if sandbox_permissions.requires_escalated_permissions() {
                 return true;
             }
             command_might_be_dangerous(command)

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;
@@ -29,12 +32,22 @@ pub const SUMMARIZATION_PROMPT: &str = include_str!("../templates/compact/prompt
 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;
 }
 
@@ -42,13 +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(
@@ -59,7 +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;
 
@@ -112,9 +127,7 @@ 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(),
-                });
+                let event = EventMsg::Error(e.to_error_event(None));
                 sess.send_event(&turn_context, event).await;
                 return;
             }
@@ -125,14 +138,13 @@ async fn run_compact_task_inner(
                     sess.notify_stream_error(
                         turn_context.as_ref(),
                         format!("Reconnecting... {retries}/{max_retries}"),
+                        e,
                     )
                     .await;
                     tokio::time::sleep(delay).await;
                     continue;
                 } else {
-                    let event = EventMsg::Error(ErrorEvent {
-                        message: e.to_string(),
-                    });
+                    let event = EventMsg::Error(e.to_error_event(None));
                     sess.send_event(&turn_context, event).await;
                     return;
                 }
@@ -155,18 +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;
 
@@ -229,7 +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,
     )
 }
 
@@ -237,20 +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;
             }
@@ -299,7 +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;
@@ -317,6 +324,7 @@ async fn drain_to_completed(
 
 #[cfg(test)]
 mod tests {
+
     use super::*;
     use pretty_assertions::assert_eq;
 
@@ -408,16 +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);
 
@@ -450,7 +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 +1,83 @@
+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 = EventMsg::Error(
+            err.to_error_event(Some("Error running remote compact task".to_string())),
+        );
+        sess.send_event(turn_context, 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

@@ -4,7 +4,6 @@ use crate::config::types::Notice;
 use anyhow::Context;
 use codex_protocol::config_types::ReasoningEffort;
 use codex_protocol::config_types::TrustLevel;
-use codex_utils_tokenizer::warm_model_cache;
 use std::collections::BTreeMap;
 use std::path::Path;
 use std::path::PathBuf;
@@ -231,9 +230,6 @@ impl ConfigDocument {
     fn apply(&mut self, edit: &ConfigEdit) -> anyhow::Result<bool> {
         match edit {
             ConfigEdit::SetModel { model, effort } => Ok({
-                if let Some(model) = &model {
-                    warm_model_cache(model)
-                }
                 let mut mutated = false;
                 mutated |= self.write_profile_value(
                     &["model"],
@@ -550,6 +546,15 @@ 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)
@@ -836,6 +841,36 @@ 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");

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

@@ -61,9 +61,6 @@ pub mod edit;
 pub mod profile;
 pub mod types;
 
-#[cfg(target_os = "windows")]
-pub const OPENAI_DEFAULT_MODEL: &str = "gpt-5.1";
-#[cfg(not(target_os = "windows"))]
 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";
@@ -81,7 +78,7 @@ pub struct Config {
     /// Optional override of model selection.
     pub model: String,
 
-    /// Model used specifically for review sessions. Defaults to "gpt-5.1-codex".
+    /// Model used specifically for review sessions. Defaults to "gpt-5.1-codex-max".
     pub review_model: String,
 
     pub model_family: ModelFamily,
@@ -163,6 +160,9 @@ pub struct Config {
     /// and turn completions when not focused.
     pub tui_notifications: Notifications,
 
+    /// Enable ASCII animations and shimmer effects in the TUI.
+    pub animations: bool,
+
     /// The directory that should be treated as the current working directory
     /// for the session. All relative paths inside the business-logic layer are
     /// resolved against this path.
@@ -195,6 +195,9 @@ pub struct Config {
     /// Additional filenames to try when looking for project-level docs.
     pub project_doc_fallback_filenames: Vec<String>,
 
+    /// Token budget applied when storing tool/function outputs in the context manager.
+    pub tool_output_token_limit: Option<usize>,
+
     /// Directory containing all Codex state (defaults to `~/.codex` but can be
     /// overridden by the `CODEX_HOME` environment variable).
     pub codex_home: PathBuf,
@@ -636,6 +639,9 @@ pub struct ConfigToml {
     /// Ordered list of fallback filenames to look for when AGENTS.md is missing.
     pub project_doc_fallback_filenames: Option<Vec<String>>,
 
+    /// Token budget applied when storing tool/function outputs in the context manager.
+    pub tool_output_token_limit: Option<usize>,
+
     /// Profile to use from the `profiles` map.
     pub profile: Option<String>,
 
@@ -1209,6 +1215,7 @@ impl Config {
                     }
                 })
                 .collect(),
+            tool_output_token_limit: cfg.tool_output_token_limit,
             codex_home,
             history,
             file_opener: cfg.file_opener.unwrap_or(UriBasedFileOpener::VsCode),
@@ -1249,6 +1256,7 @@ impl Config {
                 .as_ref()
                 .map(|t| t.notifications.clone())
                 .unwrap_or_default(),
+            animations: cfg.tui.as_ref().map(|t| t.animations).unwrap_or(true),
             otel: {
                 let t: OtelConfigToml = cfg.otel.unwrap_or_default();
                 let log_user_prompt = t.log_user_prompt.unwrap_or(false);
@@ -1313,6 +1321,16 @@ 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 {
@@ -2961,6 +2979,7 @@ 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,
@@ -2988,6 +3007,7 @@ model_verbosity = "high"
                 notices: Default::default(),
                 disable_paste_burst: false,
                 tui_notifications: Default::default(),
+                animations: true,
                 otel: OtelConfig::default(),
             },
             o3_profile_config
@@ -3032,6 +3052,7 @@ 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,
@@ -3059,6 +3080,7 @@ model_verbosity = "high"
             notices: Default::default(),
             disable_paste_burst: false,
             tui_notifications: Default::default(),
+            animations: true,
             otel: OtelConfig::default(),
         };
 
@@ -3118,6 +3140,7 @@ 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,
@@ -3145,6 +3168,7 @@ model_verbosity = "high"
             notices: Default::default(),
             disable_paste_burst: false,
             tui_notifications: Default::default(),
+            animations: true,
             otel: OtelConfig::default(),
         };
 
@@ -3190,6 +3214,7 @@ 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,
@@ -3217,6 +3242,7 @@ model_verbosity = "high"
             notices: Default::default(),
             disable_paste_burst: false,
             tui_notifications: Default::default(),
+            animations: true,
             otel: OtelConfig::default(),
         };
 

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>,
     },
 }
 
@@ -349,6 +363,15 @@ pub struct Tui {
     /// Defaults to `true`.
     #[serde(default)]
     pub notifications: Notifications,
+
+    /// Enable animations (welcome screen, shimmer effects, spinners).
+    /// Defaults to `true`.
+    #[serde(default = "default_true")]
+    pub animations: bool,
+}
+
+const fn default_true() -> bool {
+    true
 }
 
 /// Settings for notices we display to users via the tui and app-server clients
@@ -364,6 +387,9 @@ 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
+    #[serde(rename = "hide_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

@@ -1,21 +1,16 @@
 use crate::codex::TurnContext;
 use crate::context_manager::normalize;
-use crate::truncate;
-use crate::truncate::format_output_for_model_body;
-use crate::truncate::globally_truncate_function_output_items;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::approx_token_count;
+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::models::ShellToolCallParams;
 use codex_protocol::protocol::TokenUsage;
 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 +45,7 @@ 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 +57,7 @@ impl ContextManager {
                 continue;
             }
 
-            let processed = Self::process_item(&item);
+            let processed = self.process_item(item_ref, policy);
             self.items.push(processed);
         }
     }
@@ -80,26 +75,21 @@ impl ContextManager {
         history
     }
 
-    // Estimate the number of tokens in the history. Return None if no tokenizer
-    // is available. This does not consider the reasoning traces.
-    // /!\ The value is a lower bound estimate and does not represent the exact
-    // context length.
+    // Estimate token usage using byte-based heuristics from the truncation helpers.
+    // This is a coarse lower bound, not a tokenizer-accurate count.
     pub(crate) fn estimate_token_count(&self, turn_context: &TurnContext) -> Option<i64> {
-        let model = turn_context.client.get_model();
-        let tokenizer = Tokenizer::for_model(model.as_str()).ok()?;
         let model_family = turn_context.client.get_model_family();
+        let base_tokens =
+            i64::try_from(approx_token_count(model_family.base_instructions.as_str()))
+                .unwrap_or(i64::MAX);
 
-        Some(
-            self.items
-                .iter()
-                .map(|item| {
-                    serde_json::to_string(&item)
-                        .map(|item| tokenizer.count(&item))
-                        .unwrap_or_default()
-                })
-                .sum::<i64>()
-                + tokenizer.count(model_family.base_instructions.as_str()),
-        )
+        let items_tokens = self.items.iter().fold(0i64, |acc, item| {
+            let serialized = serde_json::to_string(item).unwrap_or_default();
+            let item_tokens = i64::try_from(approx_token_count(&serialized)).unwrap_or(i64::MAX);
+            acc.saturating_add(item_tokens)
+        });
+
+        Some(base_tokens.saturating_add(items_tokens))
     }
 
     pub(crate) fn remove_first_item(&mut self) {
@@ -141,6 +131,47 @@ impl ContextManager {
         normalize::remove_orphan_outputs(&mut self.items);
     }
 
+    fn get_shell_truncation_policy(&self, call_id: &str) -> Option<TruncationPolicy> {
+        let call = self.get_call_for_call_id(call_id)?;
+        match call {
+            ResponseItem::FunctionCall { arguments, .. } => {
+                let shell_tool_call_params =
+                    serde_json::from_str::<ShellToolCallParams>(&arguments).ok()?;
+                Self::create_truncation_policy(
+                    shell_tool_call_params.max_output_tokens,
+                    shell_tool_call_params.max_output_chars,
+                )
+            }
+            _ => None,
+        }
+    }
+
+    fn create_truncation_policy(
+        max_output_tokens: Option<usize>,
+        max_output_chars: Option<usize>,
+    ) -> Option<TruncationPolicy> {
+        if let Some(max_output_tokens) = max_output_tokens {
+            Some(TruncationPolicy::Tokens(max_output_tokens))
+        } else {
+            max_output_chars.map(TruncationPolicy::Bytes)
+        }
+    }
+
+    fn get_call_for_call_id(&self, call_id: &str) -> Option<ResponseItem> {
+        self.items.iter().find_map(|item| match item {
+            ResponseItem::FunctionCall {
+                call_id: existing, ..
+            } => {
+                if existing == call_id {
+                    Some(item.clone())
+                } else {
+                    None
+                }
+            }
+            _ => None,
+        })
+    }
+
     /// Returns a clone of the contents in the transcript.
     fn contents(&self) -> Vec<ResponseItem> {
         self.items.clone()
@@ -150,18 +181,17 @@ 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 truncation_policy_override = self.get_shell_truncation_policy(call_id);
+                let truncation_policy =
+                    truncation_policy_override.unwrap_or(policy_with_serialization_budget);
+                let truncated = truncate_text(output.content.as_str(), truncation_policy);
+                let truncated_items = output.content_items.as_ref().map(|items| {
+                    truncate_function_output_items_with_policy(items, truncation_policy)
+                });
                 ResponseItem::FunctionCallOutput {
                     call_id: call_id.clone(),
                     output: FunctionCallOutputPayload {
@@ -172,11 +202,7 @@ 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 +214,7 @@ impl ContextManager {
             | ResponseItem::FunctionCall { .. }
             | ResponseItem::WebSearchCall { .. }
             | ResponseItem::CustomToolCall { .. }
+            | ResponseItem::CompactionSummary { .. }
             | ResponseItem::GhostSnapshot { .. }
             | ResponseItem::Other => item.clone(),
         }
@@ -205,7 +232,8 @@ 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::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

@@ -1,7 +1,4 @@
 mod history;
 mod normalize;
 
-pub(crate) use crate::truncate::MODEL_FORMAT_MAX_BYTES;
-pub(crate) use crate::truncate::MODEL_FORMAT_MAX_LINES;
-pub(crate) use crate::truncate::format_output_for_model_body;
 pub(crate) use history::ContextManager;

codex-rs/core/src/environment_context.rs

@@ -6,6 +6,7 @@ use crate::codex::TurnContext;
 use crate::protocol::AskForApproval;
 use crate::protocol::SandboxPolicy;
 use crate::shell::Shell;
+use crate::shell::default_user_shell;
 use codex_protocol::config_types::SandboxMode;
 use codex_protocol::models::ContentItem;
 use codex_protocol::models::ResponseItem;
@@ -28,7 +29,7 @@ pub(crate) struct EnvironmentContext {
     pub sandbox_mode: Option<SandboxMode>,
     pub network_access: Option<NetworkAccess>,
     pub writable_roots: Option<Vec<PathBuf>>,
-    pub shell: Option<Shell>,
+    pub shell: Shell,
 }
 
 impl EnvironmentContext {
@@ -36,7 +37,7 @@ impl EnvironmentContext {
         cwd: Option<PathBuf>,
         approval_policy: Option<AskForApproval>,
         sandbox_policy: Option<SandboxPolicy>,
-        shell: Option<Shell>,
+        shell: Shell,
     ) -> Self {
         Self {
             cwd,
@@ -110,7 +111,7 @@ impl EnvironmentContext {
         } else {
             None
         };
-        EnvironmentContext::new(cwd, approval_policy, sandbox_policy, None)
+        EnvironmentContext::new(cwd, approval_policy, sandbox_policy, default_user_shell())
     }
 }
 
@@ -121,7 +122,7 @@ impl From<&TurnContext> for EnvironmentContext {
             Some(turn_context.approval_policy),
             Some(turn_context.sandbox_policy.clone()),
             // Shell is not configurable from turn to turn
-            None,
+            default_user_shell(),
         )
     }
 }
@@ -169,11 +170,9 @@ impl EnvironmentContext {
             }
             lines.push("  </writable_roots>".to_string());
         }
-        if let Some(shell) = self.shell
-            && let Some(shell_name) = shell.name()
-        {
-            lines.push(format!("  <shell>{shell_name}</shell>"));
-        }
+
+        let shell_name = self.shell.name();
+        lines.push(format!("  <shell>{shell_name}</shell>"));
         lines.push(ENVIRONMENT_CONTEXT_CLOSE_TAG.to_string());
         lines.join("\n")
     }
@@ -193,12 +192,18 @@ impl From<EnvironmentContext> for ResponseItem {
 
 #[cfg(test)]
 mod tests {
-    use crate::shell::BashShell;
-    use crate::shell::ZshShell;
+    use crate::shell::ShellType;
 
     use super::*;
     use pretty_assertions::assert_eq;
 
+    fn fake_shell() -> Shell {
+        Shell {
+            shell_type: ShellType::Bash,
+            shell_path: PathBuf::from("/bin/bash"),
+        }
+    }
+
     fn workspace_write_policy(writable_roots: Vec<&str>, network_access: bool) -> SandboxPolicy {
         SandboxPolicy::WorkspaceWrite {
             writable_roots: writable_roots.into_iter().map(PathBuf::from).collect(),
@@ -214,7 +219,7 @@ mod tests {
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(workspace_write_policy(vec!["/repo", "/tmp"], false)),
-            None,
+            fake_shell(),
         );
 
         let expected = r#"<environment_context>
@@ -226,6 +231,7 @@ mod tests {
     <root>/repo</root>
     <root>/tmp</root>
   </writable_roots>
+  <shell>bash</shell>
 </environment_context>"#;
 
         assert_eq!(context.serialize_to_xml(), expected);
@@ -237,13 +243,14 @@ mod tests {
             None,
             Some(AskForApproval::Never),
             Some(SandboxPolicy::ReadOnly),
-            None,
+            fake_shell(),
         );
 
         let expected = r#"<environment_context>
   <approval_policy>never</approval_policy>
   <sandbox_mode>read-only</sandbox_mode>
   <network_access>restricted</network_access>
+  <shell>bash</shell>
 </environment_context>"#;
 
         assert_eq!(context.serialize_to_xml(), expected);
@@ -255,13 +262,14 @@ mod tests {
             None,
             Some(AskForApproval::OnFailure),
             Some(SandboxPolicy::DangerFullAccess),
-            None,
+            fake_shell(),
         );
 
         let expected = r#"<environment_context>
   <approval_policy>on-failure</approval_policy>
   <sandbox_mode>danger-full-access</sandbox_mode>
   <network_access>enabled</network_access>
+  <shell>bash</shell>
 </environment_context>"#;
 
         assert_eq!(context.serialize_to_xml(), expected);
@@ -274,13 +282,13 @@ mod tests {
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(workspace_write_policy(vec!["/repo"], false)),
-            None,
+            fake_shell(),
         );
         let context2 = EnvironmentContext::new(
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::Never),
             Some(workspace_write_policy(vec!["/repo"], true)),
-            None,
+            fake_shell(),
         );
         assert!(!context1.equals_except_shell(&context2));
     }
@@ -291,13 +299,13 @@ mod tests {
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(SandboxPolicy::new_read_only_policy()),
-            None,
+            fake_shell(),
         );
         let context2 = EnvironmentContext::new(
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(SandboxPolicy::new_workspace_write_policy()),
-            None,
+            fake_shell(),
         );
 
         assert!(!context1.equals_except_shell(&context2));
@@ -309,13 +317,13 @@ mod tests {
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(workspace_write_policy(vec!["/repo", "/tmp", "/var"], false)),
-            None,
+            fake_shell(),
         );
         let context2 = EnvironmentContext::new(
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(workspace_write_policy(vec!["/repo", "/tmp"], true)),
-            None,
+            fake_shell(),
         );
 
         assert!(!context1.equals_except_shell(&context2));
@@ -327,17 +335,19 @@ mod tests {
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(workspace_write_policy(vec!["/repo"], false)),
-            Some(Shell::Bash(BashShell {
+            Shell {
+                shell_type: ShellType::Bash,
                 shell_path: "/bin/bash".into(),
-            })),
+            },
         );
         let context2 = EnvironmentContext::new(
             Some(PathBuf::from("/repo")),
             Some(AskForApproval::OnRequest),
             Some(workspace_write_policy(vec!["/repo"], false)),
-            Some(Shell::Zsh(ZshShell {
+            Shell {
+                shell_type: ShellType::Zsh,
                 shell_path: "/bin/zsh".into(),
-            })),
+            },
         );
 
         assert!(context1.equals_except_shell(&context2));

codex-rs/core/src/error.rs

@@ -2,13 +2,16 @@ 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::CodexErrorInfo;
+use codex_protocol::protocol::ErrorEvent;
 use codex_protocol::protocol::RateLimitSnapshot;
 use reqwest::StatusCode;
 use serde_json;
@@ -429,6 +432,57 @@ impl CodexErr {
     pub fn downcast_ref<T: std::any::Any>(&self) -> Option<&T> {
         (self as &dyn std::any::Any).downcast_ref::<T>()
     }
+
+    /// Translate core error to client-facing protocol error.
+    pub fn to_codex_protocol_error(&self) -> CodexErrorInfo {
+        match self {
+            CodexErr::ContextWindowExceeded => CodexErrorInfo::ContextWindowExceeded,
+            CodexErr::UsageLimitReached(_)
+            | CodexErr::QuotaExceeded
+            | CodexErr::UsageNotIncluded => CodexErrorInfo::UsageLimitExceeded,
+            CodexErr::RetryLimit(_) => CodexErrorInfo::ResponseTooManyFailedAttempts {
+                http_status_code: self.http_status_code_value(),
+            },
+            CodexErr::ConnectionFailed(_) => CodexErrorInfo::HttpConnectionFailed {
+                http_status_code: self.http_status_code_value(),
+            },
+            CodexErr::ResponseStreamFailed(_) => CodexErrorInfo::ResponseStreamConnectionFailed {
+                http_status_code: self.http_status_code_value(),
+            },
+            CodexErr::RefreshTokenFailed(_) => CodexErrorInfo::Unauthorized,
+            CodexErr::SessionConfiguredNotFirstEvent
+            | CodexErr::InternalServerError
+            | CodexErr::InternalAgentDied => CodexErrorInfo::InternalServerError,
+            CodexErr::UnsupportedOperation(_) | CodexErr::ConversationNotFound(_) => {
+                CodexErrorInfo::BadRequest
+            }
+            CodexErr::Sandbox(_) => CodexErrorInfo::SandboxError,
+            _ => CodexErrorInfo::Other,
+        }
+    }
+
+    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,
+            codex_error_info: Some(self.to_codex_protocol_error()),
+        }
+    }
+
+    pub fn http_status_code_value(&self) -> Option<u16> {
+        let http_status_code = match self {
+            CodexErr::RetryLimit(err) => Some(err.status),
+            CodexErr::UnexpectedStatus(err) => Some(err.status),
+            CodexErr::ConnectionFailed(err) => err.source.status(),
+            CodexErr::ResponseStreamFailed(err) => err.source.status(),
+            _ => None,
+        };
+        http_status_code.as_ref().map(StatusCode::as_u16)
+    }
 }
 
 pub fn get_error_message_ui(e: &CodexErr) -> String {
@@ -461,7 +515,10 @@ 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)]
@@ -474,6 +531,10 @@ mod tests {
     use chrono::Utc;
     use codex_protocol::protocol::RateLimitWindow;
     use pretty_assertions::assert_eq;
+    use reqwest::Response;
+    use reqwest::ResponseBuilderExt;
+    use reqwest::StatusCode;
+    use reqwest::Url;
 
     fn rate_limit_snapshot() -> RateLimitSnapshot {
         let primary_reset_at = Utc
@@ -495,6 +556,7 @@ mod tests {
                 window_minutes: Some(120),
                 resets_at: Some(secondary_reset_at),
             }),
+            credits: None,
         }
     }
 
@@ -568,6 +630,33 @@ mod tests {
         assert_eq!(get_error_message_ui(&err), "stdout only");
     }
 
+    #[test]
+    fn to_error_event_handles_response_stream_failed() {
+        let response = http::Response::builder()
+            .status(StatusCode::TOO_MANY_REQUESTS)
+            .url(Url::parse("http://example.com").unwrap())
+            .body("")
+            .unwrap();
+        let source = Response::from(response).error_for_status_ref().unwrap_err();
+        let err = CodexErr::ResponseStreamFailed(ResponseStreamFailed {
+            source,
+            request_id: Some("req-123".to_string()),
+        });
+
+        let event = err.to_error_event(Some("prefix".to_string()));
+
+        assert_eq!(
+            event.message,
+            "prefix: Error while reading the server response: HTTP status client error (429 Too Many Requests) for url (http://example.com/), request id: req-123"
+        );
+        assert_eq!(
+            event.codex_error_info,
+            Some(CodexErrorInfo::ResponseStreamConnectionFailed {
+                http_status_code: Some(429)
+            })
+        );
+    }
+
     #[test]
     fn sandbox_denied_reports_exit_code_when_no_output_available() {
         let output = ExecToolCallOutput {

codex-rs/core/src/exec.rs

@@ -14,6 +14,7 @@ use tokio::io::AsyncRead;
 use tokio::io::AsyncReadExt;
 use tokio::io::BufReader;
 use tokio::process::Child;
+use tokio_util::sync::CancellationToken;
 
 use crate::error::CodexErr;
 use crate::error::Result;
@@ -28,8 +29,9 @@ use crate::sandboxing::ExecEnv;
 use crate::sandboxing::SandboxManager;
 use crate::spawn::StdioPolicy;
 use crate::spawn::spawn_child_async;
+use crate::text_encoding::bytes_to_string_smart;
 
-const DEFAULT_TIMEOUT_MS: u64 = 10_000;
+pub const DEFAULT_EXEC_COMMAND_TIMEOUT_MS: u64 = 10_000;
 
 // Hardcode these since it does not seem worth including the libc crate just
 // for these.
@@ -46,20 +48,61 @@ const AGGREGATE_BUFFER_INITIAL_CAPACITY: usize = 8 * 1024; // 8 KiB
 /// Aggregation still collects full output; only the live event stream is capped.
 pub(crate) const MAX_EXEC_OUTPUT_DELTAS_PER_CALL: usize = 10_000;
 
-#[derive(Clone, Debug)]
+#[derive(Debug)]
 pub struct ExecParams {
     pub command: Vec<String>,
     pub cwd: PathBuf,
-    pub timeout_ms: Option<u64>,
+    pub expiration: ExecExpiration,
     pub env: HashMap<String, String>,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
     pub arg0: Option<String>,
+    pub max_output_tokens: Option<usize>,
+    pub max_output_chars: Option<usize>,
 }
 
-impl ExecParams {
-    pub fn timeout_duration(&self) -> Duration {
-        Duration::from_millis(self.timeout_ms.unwrap_or(DEFAULT_TIMEOUT_MS))
+/// Mechanism to terminate an exec invocation before it finishes naturally.
+#[derive(Debug)]
+pub enum ExecExpiration {
+    Timeout(Duration),
+    DefaultTimeout,
+    Cancellation(CancellationToken),
+}
+
+impl From<Option<u64>> for ExecExpiration {
+    fn from(timeout_ms: Option<u64>) -> Self {
+        timeout_ms.map_or(ExecExpiration::DefaultTimeout, |timeout_ms| {
+            ExecExpiration::Timeout(Duration::from_millis(timeout_ms))
+        })
+    }
+}
+
+impl From<u64> for ExecExpiration {
+    fn from(timeout_ms: u64) -> Self {
+        ExecExpiration::Timeout(Duration::from_millis(timeout_ms))
+    }
+}
+
+impl ExecExpiration {
+    async fn wait(self) {
+        match self {
+            ExecExpiration::Timeout(duration) => tokio::time::sleep(duration).await,
+            ExecExpiration::DefaultTimeout => {
+                tokio::time::sleep(Duration::from_millis(DEFAULT_EXEC_COMMAND_TIMEOUT_MS)).await
+            }
+            ExecExpiration::Cancellation(cancel) => {
+                cancel.cancelled().await;
+            }
+        }
+    }
+
+    /// If ExecExpiration is a timeout, returns the timeout in milliseconds.
+    pub(crate) fn timeout_ms(&self) -> Option<u64> {
+        match self {
+            ExecExpiration::Timeout(duration) => Some(duration.as_millis() as u64),
+            ExecExpiration::DefaultTimeout => Some(DEFAULT_EXEC_COMMAND_TIMEOUT_MS),
+            ExecExpiration::Cancellation(_) => None,
+        }
     }
 }
 
@@ -95,11 +138,13 @@ pub async fn process_exec_tool_call(
     let ExecParams {
         command,
         cwd,
-        timeout_ms,
+        expiration,
         env,
         with_escalated_permissions,
         justification,
         arg0: _,
+        max_output_tokens,
+        max_output_chars,
     } = params;
 
     let (program, args) = command.split_first().ok_or_else(|| {
@@ -114,15 +159,17 @@ pub async fn process_exec_tool_call(
         args: args.to_vec(),
         cwd,
         env,
-        timeout_ms,
+        expiration,
         with_escalated_permissions,
         justification,
+        max_output_tokens,
+        max_output_chars,
     };
 
     let manager = SandboxManager::new();
     let exec_env = manager
         .transform(
-            &spec,
+            spec,
             sandbox_policy,
             sandbox_type,
             sandbox_cwd,
@@ -131,7 +178,7 @@ pub async fn process_exec_tool_call(
         .map_err(CodexErr::from)?;
 
     // Route through the sandboxing module for a single, unified execution path.
-    crate::sandboxing::execute_env(&exec_env, sandbox_policy, stdout_stream).await
+    crate::sandboxing::execute_env(exec_env, sandbox_policy, stdout_stream).await
 }
 
 pub(crate) async fn execute_exec_env(
@@ -143,21 +190,25 @@ pub(crate) async fn execute_exec_env(
         command,
         cwd,
         env,
-        timeout_ms,
+        expiration,
         sandbox,
         with_escalated_permissions,
         justification,
         arg0,
+        max_output_tokens,
+        max_output_chars,
     } = env;
 
     let params = ExecParams {
         command,
         cwd,
-        timeout_ms,
+        expiration,
         env,
         with_escalated_permissions,
         justification,
         arg0,
+        max_output_tokens,
+        max_output_chars,
     };
 
     let start = Instant::now();
@@ -178,16 +229,18 @@ async fn exec_windows_sandbox(
         command,
         cwd,
         env,
-        timeout_ms,
+        expiration,
         ..
     } = params;
+    // TODO(iceweasel-oai): run_windows_sandbox_capture should support all
+    // variants of ExecExpiration, not just timeout.
+    let timeout_ms = expiration.timeout_ms();
 
-    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 codex_home = find_codex_home().map_err(|err| {
         CodexErr::Io(io::Error::other(format!(
@@ -196,7 +249,7 @@ async fn exec_windows_sandbox(
     })?;
     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,
@@ -415,7 +468,7 @@ impl StreamOutput<String> {
 impl StreamOutput<Vec<u8>> {
     pub fn from_utf8_lossy(&self) -> StreamOutput<String> {
         StreamOutput {
-            text: String::from_utf8_lossy(&self.text).to_string(),
+            text: bytes_to_string_smart(&self.text),
             truncated_after_lines: self.truncated_after_lines,
         }
     }
@@ -444,15 +497,17 @@ 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();
     let ExecParams {
         command,
         cwd,
         env,
         arg0,
+        expiration,
         ..
     } = params;
 
@@ -473,14 +528,14 @@ async fn exec(
         env,
     )
     .await?;
-    consume_truncated_output(child, timeout, stdout_stream).await
+    consume_truncated_output(child, expiration, stdout_stream).await
 }
 
 /// Consumes the output of a child process, truncating it so it is suitable for
 /// use as the output of a `shell` tool call. Also enforces specified timeout.
 async fn consume_truncated_output(
     mut child: Child,
-    timeout: Duration,
+    expiration: ExecExpiration,
     stdout_stream: Option<StdoutStream>,
 ) -> Result<RawExecToolCallOutput> {
     // Both stdout and stderr were configured with `Stdio::piped()`
@@ -514,20 +569,14 @@ async fn consume_truncated_output(
     ));
 
     let (exit_status, timed_out) = tokio::select! {
-        result = tokio::time::timeout(timeout, child.wait()) => {
-            match result {
-                Ok(status_result) => {
-                    let exit_status = status_result?;
-                    (exit_status, false)
-                }
-                Err(_) => {
-                    // timeout
-                    kill_child_process_group(&mut child)?;
-                    child.start_kill()?;
-                    // Debatable whether `child.wait().await` should be called here.
-                    (synthetic_exit_status(EXIT_CODE_SIGNAL_BASE + TIMEOUT_CODE), true)
-                }
-            }
+        status_result = child.wait() => {
+            let exit_status = status_result?;
+            (exit_status, false)
+        }
+        _ = expiration.wait() => {
+            kill_child_process_group(&mut child)?;
+            child.start_kill()?;
+            (synthetic_exit_status(EXIT_CODE_SIGNAL_BASE + TIMEOUT_CODE), true)
         }
         _ = tokio::signal::ctrl_c() => {
             kill_child_process_group(&mut child)?;
@@ -779,6 +828,15 @@ mod tests {
     #[cfg(unix)]
     #[tokio::test]
     async fn kill_child_process_group_kills_grandchildren_on_timeout() -> Result<()> {
+        // On Linux/macOS, /bin/bash is typically present; on FreeBSD/OpenBSD,
+        // prefer /bin/sh to avoid NotFound errors.
+        #[cfg(any(target_os = "freebsd", target_os = "openbsd"))]
+        let command = vec![
+            "/bin/sh".to_string(),
+            "-c".to_string(),
+            "sleep 60 & echo $!; sleep 60".to_string(),
+        ];
+        #[cfg(all(unix, not(any(target_os = "freebsd", target_os = "openbsd"))))]
         let command = vec![
             "/bin/bash".to_string(),
             "-c".to_string(),
@@ -788,11 +846,13 @@ mod tests {
         let params = ExecParams {
             command,
             cwd: std::env::current_dir()?,
-            timeout_ms: Some(500),
+            expiration: 500.into(),
             env,
             with_escalated_permissions: None,
             justification: None,
             arg0: None,
+            max_output_tokens: None,
+            max_output_chars: None,
         };
 
         let output = exec(params, SandboxType::None, &SandboxPolicy::ReadOnly, None).await?;
@@ -822,4 +882,64 @@ mod tests {
         assert!(killed, "grandchild process with pid {pid} is still alive");
         Ok(())
     }
+
+    #[tokio::test]
+    async fn process_exec_tool_call_respects_cancellation_token() -> Result<()> {
+        let command = long_running_command();
+        let cwd = std::env::current_dir()?;
+        let env: HashMap<String, String> = std::env::vars().collect();
+        let cancel_token = CancellationToken::new();
+        let cancel_tx = cancel_token.clone();
+        let params = ExecParams {
+            command,
+            cwd: cwd.clone(),
+            expiration: ExecExpiration::Cancellation(cancel_token),
+            env,
+            with_escalated_permissions: None,
+            justification: None,
+            arg0: None,
+            max_output_tokens: None,
+            max_output_chars: None,
+        };
+        tokio::spawn(async move {
+            tokio::time::sleep(Duration::from_millis(1_000)).await;
+            cancel_tx.cancel();
+        });
+        let result = process_exec_tool_call(
+            params,
+            SandboxType::None,
+            &SandboxPolicy::DangerFullAccess,
+            cwd.as_path(),
+            &None,
+            None,
+        )
+        .await;
+        let output = match result {
+            Err(CodexErr::Sandbox(SandboxErr::Timeout { output })) => output,
+            other => panic!("expected timeout error, got {other:?}"),
+        };
+        assert!(output.timed_out);
+        assert_eq!(output.exit_code, EXEC_TIMEOUT_EXIT_CODE);
+        Ok(())
+    }
+
+    #[cfg(unix)]
+    fn long_running_command() -> Vec<String> {
+        vec![
+            "/bin/sh".to_string(),
+            "-c".to_string(),
+            "sleep 30".to_string(),
+        ]
+    }
+
+    #[cfg(windows)]
+    fn long_running_command() -> Vec<String> {
+        vec![
+            "powershell.exe".to_string(),
+            "-NonInteractive".to_string(),
+            "-NoLogo".to_string(),
+            "-Command".to_string(),
+            "Start-Sleep -Seconds 30".to_string(),
+        ]
+    }
 }

codex-rs/core/src/exec_policy.rs

@@ -0,0 +1,365 @@
+use std::io::ErrorKind;
+use std::path::Path;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use crate::command_safety::is_dangerous_command::requires_initial_appoval;
+use codex_execpolicy::Decision;
+use codex_execpolicy::Evaluation;
+use codex_execpolicy::Policy;
+use codex_execpolicy::PolicyParser;
+use codex_protocol::protocol::AskForApproval;
+use codex_protocol::protocol::SandboxPolicy;
+use thiserror::Error;
+use tokio::fs;
+
+use crate::bash::parse_shell_lc_plain_commands;
+use crate::features::Feature;
+use crate::features::Features;
+use crate::sandboxing::SandboxPermissions;
+use crate::tools::sandboxing::ApprovalRequirement;
+
+const FORBIDDEN_REASON: &str = "execpolicy forbids this command";
+const PROMPT_REASON: &str = "execpolicy requires approval for this command";
+const POLICY_DIR_NAME: &str = "policy";
+const POLICY_EXTENSION: &str = "codexpolicy";
+
+#[derive(Debug, Error)]
+pub enum ExecPolicyError {
+    #[error("failed to read execpolicy files from {dir}: {source}")]
+    ReadDir {
+        dir: PathBuf,
+        source: std::io::Error,
+    },
+
+    #[error("failed to read execpolicy file {path}: {source}")]
+    ReadFile {
+        path: PathBuf,
+        source: std::io::Error,
+    },
+
+    #[error("failed to parse execpolicy file {path}: {source}")]
+    ParsePolicy {
+        path: String,
+        source: codex_execpolicy::Error,
+    },
+}
+
+pub(crate) async fn exec_policy_for(
+    features: &Features,
+    codex_home: &Path,
+) -> Result<Arc<Policy>, ExecPolicyError> {
+    if !features.enabled(Feature::ExecPolicy) {
+        return Ok(Arc::new(Policy::empty()));
+    }
+
+    let policy_dir = codex_home.join(POLICY_DIR_NAME);
+    let policy_paths = collect_policy_files(&policy_dir).await?;
+
+    let mut parser = PolicyParser::new();
+    for policy_path in &policy_paths {
+        let contents =
+            fs::read_to_string(policy_path)
+                .await
+                .map_err(|source| ExecPolicyError::ReadFile {
+                    path: policy_path.clone(),
+                    source,
+                })?;
+        let identifier = policy_path.to_string_lossy().to_string();
+        parser
+            .parse(&identifier, &contents)
+            .map_err(|source| ExecPolicyError::ParsePolicy {
+                path: identifier,
+                source,
+            })?;
+    }
+
+    let policy = Arc::new(parser.build());
+    tracing::debug!(
+        "loaded execpolicy from {} files in {}",
+        policy_paths.len(),
+        policy_dir.display()
+    );
+
+    Ok(policy)
+}
+
+fn evaluate_with_policy(
+    policy: &Policy,
+    command: &[String],
+    approval_policy: AskForApproval,
+) -> Option<ApprovalRequirement> {
+    let commands = parse_shell_lc_plain_commands(command).unwrap_or_else(|| vec![command.to_vec()]);
+    let evaluation = policy.check_multiple(commands.iter());
+
+    match evaluation {
+        Evaluation::Match { decision, .. } => match decision {
+            Decision::Forbidden => Some(ApprovalRequirement::Forbidden {
+                reason: FORBIDDEN_REASON.to_string(),
+            }),
+            Decision::Prompt => {
+                let reason = PROMPT_REASON.to_string();
+                if matches!(approval_policy, AskForApproval::Never) {
+                    Some(ApprovalRequirement::Forbidden { reason })
+                } else {
+                    Some(ApprovalRequirement::NeedsApproval {
+                        reason: Some(reason),
+                    })
+                }
+            }
+            Decision::Allow => Some(ApprovalRequirement::Skip),
+        },
+        Evaluation::NoMatch { .. } => None,
+    }
+}
+
+pub(crate) fn create_approval_requirement_for_command(
+    policy: &Policy,
+    command: &[String],
+    approval_policy: AskForApproval,
+    sandbox_policy: &SandboxPolicy,
+    sandbox_permissions: SandboxPermissions,
+) -> ApprovalRequirement {
+    if let Some(requirement) = evaluate_with_policy(policy, command, approval_policy) {
+        return requirement;
+    }
+
+    if requires_initial_appoval(
+        approval_policy,
+        sandbox_policy,
+        command,
+        sandbox_permissions,
+    ) {
+        ApprovalRequirement::NeedsApproval { reason: None }
+    } else {
+        ApprovalRequirement::Skip
+    }
+}
+
+async fn collect_policy_files(dir: &Path) -> Result<Vec<PathBuf>, ExecPolicyError> {
+    let mut read_dir = match fs::read_dir(dir).await {
+        Ok(read_dir) => read_dir,
+        Err(err) if err.kind() == ErrorKind::NotFound => return Ok(Vec::new()),
+        Err(source) => {
+            return Err(ExecPolicyError::ReadDir {
+                dir: dir.to_path_buf(),
+                source,
+            });
+        }
+    };
+
+    let mut policy_paths = Vec::new();
+    while let Some(entry) =
+        read_dir
+            .next_entry()
+            .await
+            .map_err(|source| ExecPolicyError::ReadDir {
+                dir: dir.to_path_buf(),
+                source,
+            })?
+    {
+        let path = entry.path();
+        let file_type = entry
+            .file_type()
+            .await
+            .map_err(|source| ExecPolicyError::ReadDir {
+                dir: dir.to_path_buf(),
+                source,
+            })?;
+
+        if path
+            .extension()
+            .and_then(|ext| ext.to_str())
+            .is_some_and(|ext| ext == POLICY_EXTENSION)
+            && file_type.is_file()
+        {
+            policy_paths.push(path);
+        }
+    }
+
+    policy_paths.sort();
+
+    Ok(policy_paths)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::features::Feature;
+    use crate::features::Features;
+    use codex_protocol::protocol::AskForApproval;
+    use codex_protocol::protocol::SandboxPolicy;
+    use pretty_assertions::assert_eq;
+    use std::fs;
+    use tempfile::tempdir;
+
+    #[tokio::test]
+    async fn returns_empty_policy_when_feature_disabled() {
+        let mut features = Features::with_defaults();
+        features.disable(Feature::ExecPolicy);
+        let temp_dir = tempdir().expect("create temp dir");
+
+        let policy = exec_policy_for(&features, temp_dir.path())
+            .await
+            .expect("policy result");
+
+        let commands = [vec!["rm".to_string()]];
+        assert!(matches!(
+            policy.check_multiple(commands.iter()),
+            Evaluation::NoMatch { .. }
+        ));
+        assert!(!temp_dir.path().join(POLICY_DIR_NAME).exists());
+    }
+
+    #[tokio::test]
+    async fn collect_policy_files_returns_empty_when_dir_missing() {
+        let temp_dir = tempdir().expect("create temp dir");
+
+        let policy_dir = temp_dir.path().join(POLICY_DIR_NAME);
+        let files = collect_policy_files(&policy_dir)
+            .await
+            .expect("collect policy files");
+
+        assert!(files.is_empty());
+    }
+
+    #[tokio::test]
+    async fn loads_policies_from_policy_subdirectory() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let policy_dir = temp_dir.path().join(POLICY_DIR_NAME);
+        fs::create_dir_all(&policy_dir).expect("create policy dir");
+        fs::write(
+            policy_dir.join("deny.codexpolicy"),
+            r#"prefix_rule(pattern=["rm"], decision="forbidden")"#,
+        )
+        .expect("write policy file");
+
+        let policy = exec_policy_for(&Features::with_defaults(), temp_dir.path())
+            .await
+            .expect("policy result");
+        let command = [vec!["rm".to_string()]];
+        assert!(matches!(
+            policy.check_multiple(command.iter()),
+            Evaluation::Match { .. }
+        ));
+    }
+
+    #[tokio::test]
+    async fn ignores_policies_outside_policy_dir() {
+        let temp_dir = tempdir().expect("create temp dir");
+        fs::write(
+            temp_dir.path().join("root.codexpolicy"),
+            r#"prefix_rule(pattern=["ls"], decision="prompt")"#,
+        )
+        .expect("write policy file");
+
+        let policy = exec_policy_for(&Features::with_defaults(), temp_dir.path())
+            .await
+            .expect("policy result");
+        let command = [vec!["ls".to_string()]];
+        assert!(matches!(
+            policy.check_multiple(command.iter()),
+            Evaluation::NoMatch { .. }
+        ));
+    }
+
+    #[test]
+    fn evaluates_bash_lc_inner_commands() {
+        let policy_src = r#"
+prefix_rule(pattern=["rm"], decision="forbidden")
+"#;
+        let mut parser = PolicyParser::new();
+        parser
+            .parse("test.codexpolicy", policy_src)
+            .expect("parse policy");
+        let policy = parser.build();
+
+        let forbidden_script = vec![
+            "bash".to_string(),
+            "-lc".to_string(),
+            "rm -rf /tmp".to_string(),
+        ];
+
+        let requirement =
+            evaluate_with_policy(&policy, &forbidden_script, AskForApproval::OnRequest)
+                .expect("expected match for forbidden command");
+
+        assert_eq!(
+            requirement,
+            ApprovalRequirement::Forbidden {
+                reason: FORBIDDEN_REASON.to_string()
+            }
+        );
+    }
+
+    #[test]
+    fn approval_requirement_prefers_execpolicy_match() {
+        let policy_src = r#"prefix_rule(pattern=["rm"], decision="prompt")"#;
+        let mut parser = PolicyParser::new();
+        parser
+            .parse("test.codexpolicy", policy_src)
+            .expect("parse policy");
+        let policy = parser.build();
+        let command = vec!["rm".to_string()];
+
+        let requirement = create_approval_requirement_for_command(
+            &policy,
+            &command,
+            AskForApproval::OnRequest,
+            &SandboxPolicy::DangerFullAccess,
+            SandboxPermissions::UseDefault,
+        );
+
+        assert_eq!(
+            requirement,
+            ApprovalRequirement::NeedsApproval {
+                reason: Some(PROMPT_REASON.to_string())
+            }
+        );
+    }
+
+    #[test]
+    fn approval_requirement_respects_approval_policy() {
+        let policy_src = r#"prefix_rule(pattern=["rm"], decision="prompt")"#;
+        let mut parser = PolicyParser::new();
+        parser
+            .parse("test.codexpolicy", policy_src)
+            .expect("parse policy");
+        let policy = parser.build();
+        let command = vec!["rm".to_string()];
+
+        let requirement = create_approval_requirement_for_command(
+            &policy,
+            &command,
+            AskForApproval::Never,
+            &SandboxPolicy::DangerFullAccess,
+            SandboxPermissions::UseDefault,
+        );
+
+        assert_eq!(
+            requirement,
+            ApprovalRequirement::Forbidden {
+                reason: PROMPT_REASON.to_string()
+            }
+        );
+    }
+
+    #[test]
+    fn approval_requirement_falls_back_to_heuristics() {
+        let command = vec!["python".to_string()];
+
+        let empty_policy = Policy::empty();
+        let requirement = create_approval_requirement_for_command(
+            &empty_policy,
+            &command,
+            AskForApproval::UnlessTrusted,
+            &SandboxPolicy::ReadOnly,
+            SandboxPermissions::UseDefault,
+        );
+
+        assert_eq!(
+            requirement,
+            ApprovalRequirement::NeedsApproval { reason: None }
+        );
+    }
+}

codex-rs/core/src/features.rs

@@ -27,11 +27,10 @@ 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
-    /// shell instead of an array of args passed to `execvp(3)`.
-    ShellCommandTool,
     /// Enable experimental RMCP features such as OAuth login.
     RmcpClient,
     /// Include the freeform apply_patch tool.
@@ -40,14 +39,18 @@ pub enum Feature {
     ViewImageTool,
     /// Allow the model to request web searches.
     WebSearchRequest,
+    /// Gate the execpolicy enforcement for shell/unified exec.
+    ExecPolicy,
     /// Enable the model-based risk assessments for sandboxed commands.
     SandboxCommandAssessment,
-    /// Create a ghost commit at each turn.
-    GhostCommit,
     /// Enable Windows sandbox (restricted token) on Windows.
     WindowsSandbox,
+    /// Remote compaction enabled (only for ChatGPT auth)
+    RemoteCompaction,
     /// Enable the default shell tool.
     ShellTool,
+    /// Allow model to call multiple tools in parallel (only for models supporting it).
+    ParallelToolCalls,
 }
 
 impl Feature {
@@ -249,18 +252,26 @@ 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",
         stage: Stage::Experimental,
         default_enabled: false,
     },
-    FeatureSpec {
-        id: Feature::ShellCommandTool,
-        key: "shell_command_tool",
-        stage: Stage::Experimental,
-        default_enabled: false,
-    },
     FeatureSpec {
         id: Feature::RmcpClient,
         key: "rmcp_client",
@@ -273,18 +284,18 @@ 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",
         stage: Stage::Stable,
         default_enabled: false,
     },
+    FeatureSpec {
+        id: Feature::ExecPolicy,
+        key: "exec_policy",
+        stage: Stage::Experimental,
+        default_enabled: true,
+    },
     FeatureSpec {
         id: Feature::SandboxCommandAssessment,
         key: "experimental_sandbox_command_assessment",
@@ -292,14 +303,20 @@ pub const FEATURES: &[FeatureSpec] = &[
         default_enabled: false,
     },
     FeatureSpec {
-        id: Feature::GhostCommit,
-        key: "ghost_commit",
+        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::WindowsSandbox,
-        key: "enable_experimental_windows_sandbox",
+        id: Feature::ParallelToolCalls,
+        key: "parallel",
         stage: Stage::Experimental,
         default_enabled: false,
     },

codex-rs/core/src/lib.rs

@@ -13,6 +13,7 @@ mod client;
 mod client_common;
 pub mod codex;
 mod codex_conversation;
+mod compact_remote;
 pub use codex_conversation::CodexConversation;
 mod codex_delegate;
 mod command_safety;
@@ -24,6 +25,7 @@ mod environment_context;
 pub mod error;
 pub mod exec;
 pub mod exec_env;
+mod exec_policy;
 pub mod features;
 mod flags;
 pub mod git_info;
@@ -34,8 +36,10 @@ mod mcp_tool_call;
 mod message_history;
 mod model_provider_info;
 pub mod parse_command;
+pub mod powershell;
 mod response_processing;
 pub mod sandboxing;
+mod text_encoding;
 pub mod token_data;
 mod truncate;
 mod unified_exec;

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)]
@@ -66,12 +68,15 @@ 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 {
     (
         $slug:expr, $family:expr $(, $key:ident : $value:expr )* $(,)?
     ) => {{
+        let truncation_policy = TruncationPolicy::Bytes(10_000);
         // defaults
         #[allow(unused_mut)]
         let mut mf = ModelFamily {
@@ -86,9 +91,10 @@ macro_rules! model_family {
             experimental_supported_tools: Vec::new(),
             effective_context_window_percent: 95,
             support_verbosity: false,
-            shell_type: ConfigShellToolType::Default,
+            shell_type: ConfigShellToolType::Default(truncation_policy),
             default_verbosity: None,
             default_reasoning_effort: None,
+            truncation_policy,
         };
 
         // apply overrides
@@ -133,6 +139,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
     } 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") {
+        let truncation_policy = TruncationPolicy::Tokens(10_000);
         model_family!(
             slug, slug,
             supports_reasoning_summaries: true,
@@ -145,11 +152,13 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
                 "test_sync_tool".to_string(),
             ],
             supports_parallel_tool_calls: true,
+            shell_type: ConfigShellToolType::ShellCommand(truncation_policy),
             support_verbosity: true,
         )
 
     // Internal models.
     } else if slug.starts_with("codex-exp-") {
+        let truncation_policy = TruncationPolicy::Tokens(10_000);
         model_family!(
             slug, slug,
             supports_reasoning_summaries: true,
@@ -161,24 +170,44 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
                 "list_dir".to_string(),
                 "read_file".to_string(),
             ],
+            shell_type: ConfigShellToolType::ShellCommand(truncation_policy),
             supports_parallel_tool_calls: true,
             support_verbosity: true,
+            truncation_policy: truncation_policy,
         )
 
     // Production models.
+    } else if slug.starts_with("gpt-5.1-codex-max") {
+        let truncation_policy = TruncationPolicy::Tokens(10_000);
+        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(truncation_policy),
+            supports_parallel_tool_calls: true,
+            support_verbosity: false,
+            truncation_policy: truncation_policy,
+        )
     } else if slug.starts_with("gpt-5-codex")
         || slug.starts_with("gpt-5.1-codex")
         || slug.starts_with("codex-")
     {
+        let truncation_policy = TruncationPolicy::Tokens(10_000);
         model_family!(
             slug, slug,
             supports_reasoning_summaries: true,
             reasoning_summary_format: ReasoningSummaryFormat::Experimental,
             base_instructions: GPT_5_CODEX_INSTRUCTIONS.to_string(),
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
+            shell_type: ConfigShellToolType::ShellCommand(truncation_policy),
+            supports_parallel_tool_calls: true,
             support_verbosity: false,
+            truncation_policy: truncation_policy,
         )
     } else if slug.starts_with("gpt-5.1") {
+        let truncation_policy = TruncationPolicy::Tokens(10_000);
         model_family!(
             slug, "gpt-5.1",
             supports_reasoning_summaries: true,
@@ -187,13 +216,18 @@ 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(truncation_policy),
+            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(TruncationPolicy::Bytes(10_000)),
             support_verbosity: true,
+            truncation_policy: TruncationPolicy::Bytes(10_000),
         )
     } else {
         None
@@ -201,6 +235,7 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
 }
 
 pub fn derive_default_model_family(model: &str) -> ModelFamily {
+    let truncation_policy = TruncationPolicy::Bytes(10_000);
     ModelFamily {
         slug: model.to_string(),
         family: model.to_string(),
@@ -213,8 +248,9 @@ pub fn derive_default_model_family(model: &str) -> ModelFamily {
         experimental_supported_tools: Vec::new(),
         effective_context_window_percent: 95,
         support_verbosity: false,
-        shell_type: ConfigShellToolType::Default,
+        shell_type: ConfigShellToolType::Default(truncation_policy),
         default_verbosity: None,
         default_reasoning_effort: None,
+        truncation_policy,
     }
 }

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;

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

@@ -1,6 +1,7 @@
 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;
@@ -11,6 +12,11 @@ pub fn shlex_join(tokens: &[String]) -> String {
         .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,
     }
 }

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/sandboxing/mod.rs

@@ -8,6 +8,7 @@ ready‑to‑spawn environment.
 
 pub mod assessment;
 
+use crate::exec::ExecExpiration;
 use crate::exec::ExecToolCallOutput;
 use crate::exec::SandboxType;
 use crate::exec::StdoutStream;
@@ -26,27 +27,53 @@ use std::collections::HashMap;
 use std::path::Path;
 use std::path::PathBuf;
 
-#[derive(Clone, Debug)]
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum SandboxPermissions {
+    UseDefault,
+    RequireEscalated,
+}
+
+impl SandboxPermissions {
+    pub fn requires_escalated_permissions(self) -> bool {
+        matches!(self, SandboxPermissions::RequireEscalated)
+    }
+}
+
+impl From<bool> for SandboxPermissions {
+    fn from(with_escalated_permissions: bool) -> Self {
+        if with_escalated_permissions {
+            SandboxPermissions::RequireEscalated
+        } else {
+            SandboxPermissions::UseDefault
+        }
+    }
+}
+
+#[derive(Debug)]
 pub struct CommandSpec {
     pub program: String,
     pub args: Vec<String>,
     pub cwd: PathBuf,
     pub env: HashMap<String, String>,
-    pub timeout_ms: Option<u64>,
+    pub expiration: ExecExpiration,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
+    pub max_output_tokens: Option<usize>,
+    pub max_output_chars: Option<usize>,
 }
 
-#[derive(Clone, Debug)]
+#[derive(Debug)]
 pub struct ExecEnv {
     pub command: Vec<String>,
     pub cwd: PathBuf,
     pub env: HashMap<String, String>,
-    pub timeout_ms: Option<u64>,
+    pub expiration: ExecExpiration,
     pub sandbox: SandboxType,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
     pub arg0: Option<String>,
+    pub max_output_tokens: Option<usize>,
+    pub max_output_chars: Option<usize>,
 }
 
 pub enum SandboxPreference {
@@ -93,13 +120,13 @@ impl SandboxManager {
 
     pub(crate) fn transform(
         &self,
-        spec: &CommandSpec,
+        mut spec: CommandSpec,
         policy: &SandboxPolicy,
         sandbox: SandboxType,
         sandbox_policy_cwd: &Path,
         codex_linux_sandbox_exe: Option<&PathBuf>,
     ) -> Result<ExecEnv, SandboxTransformError> {
-        let mut env = spec.env.clone();
+        let mut env = spec.env;
         if !policy.has_full_network_access() {
             env.insert(
                 CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR.to_string(),
@@ -108,8 +135,8 @@ impl SandboxManager {
         }
 
         let mut command = Vec::with_capacity(1 + spec.args.len());
-        command.push(spec.program.clone());
-        command.extend(spec.args.iter().cloned());
+        command.push(spec.program);
+        command.append(&mut spec.args);
 
         let (command, sandbox_env, arg0_override) = match sandbox {
             SandboxType::None => (command, HashMap::new(), None),
@@ -154,13 +181,15 @@ impl SandboxManager {
 
         Ok(ExecEnv {
             command,
-            cwd: spec.cwd.clone(),
+            cwd: spec.cwd,
             env,
-            timeout_ms: spec.timeout_ms,
+            expiration: spec.expiration,
             sandbox,
             with_escalated_permissions: spec.with_escalated_permissions,
-            justification: spec.justification.clone(),
+            justification: spec.justification,
             arg0: arg0_override,
+            max_output_tokens: spec.max_output_tokens,
+            max_output_chars: spec.max_output_chars,
         })
     }
 
@@ -170,9 +199,9 @@ impl SandboxManager {
 }
 
 pub async fn execute_env(
-    env: &ExecEnv,
+    env: ExecEnv,
     policy: &SandboxPolicy,
     stdout_stream: Option<StdoutStream>,
 ) -> crate::error::Result<ExecToolCallOutput> {
-    execute_exec_env(env.clone(), policy, stdout_stream).await
+    execute_exec_env(env, policy, stdout_stream).await
 }

codex-rs/core/src/shell.rs

@@ -7,64 +7,41 @@ pub enum ShellType {
     Zsh,
     Bash,
     PowerShell,
+    Sh,
+    Cmd,
 }
 
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
-pub struct ZshShell {
+pub struct Shell {
+    pub(crate) shell_type: ShellType,
     pub(crate) shell_path: PathBuf,
 }
 
-#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
-pub struct BashShell {
-    pub(crate) shell_path: PathBuf,
-}
-
-#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
-pub struct PowerShellConfig {
-    pub(crate) shell_path: PathBuf, // Executable name or path, e.g. "pwsh" or "powershell.exe".
-}
-
-#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
-pub enum Shell {
-    Zsh(ZshShell),
-    Bash(BashShell),
-    PowerShell(PowerShellConfig),
-    Unknown,
-}
-
 impl Shell {
-    pub fn name(&self) -> Option<String> {
-        match self {
-            Shell::Zsh(ZshShell { shell_path, .. }) | Shell::Bash(BashShell { shell_path, .. }) => {
-                std::path::Path::new(shell_path)
-                    .file_name()
-                    .map(|s| s.to_string_lossy().to_string())
-            }
-            Shell::PowerShell(ps) => ps
-                .shell_path
-                .file_stem()
-                .map(|s| s.to_string_lossy().to_string()),
-            Shell::Unknown => None,
+    pub fn name(&self) -> &'static str {
+        match self.shell_type {
+            ShellType::Zsh => "zsh",
+            ShellType::Bash => "bash",
+            ShellType::PowerShell => "powershell",
+            ShellType::Sh => "sh",
+            ShellType::Cmd => "cmd",
         }
     }
 
     /// Takes a string of shell and returns the full list of command args to
     /// use with `exec()` to run the shell command.
     pub fn derive_exec_args(&self, command: &str, use_login_shell: bool) -> Vec<String> {
-        match self {
-            Shell::Zsh(ZshShell { shell_path, .. }) | Shell::Bash(BashShell { shell_path, .. }) => {
+        match self.shell_type {
+            ShellType::Zsh | ShellType::Bash | ShellType::Sh => {
                 let arg = if use_login_shell { "-lc" } else { "-c" };
                 vec![
-                    shell_path.to_string_lossy().to_string(),
+                    self.shell_path.to_string_lossy().to_string(),
                     arg.to_string(),
                     command.to_string(),
                 ]
             }
-            Shell::PowerShell(ps) => {
-                let mut args = vec![
-                    ps.shell_path.to_string_lossy().to_string(),
-                    "-NoLogo".to_string(),
-                ];
+            ShellType::PowerShell => {
+                let mut args = vec![self.shell_path.to_string_lossy().to_string()];
                 if !use_login_shell {
                     args.push("-NoProfile".to_string());
                 }
@@ -73,7 +50,12 @@ impl Shell {
                 args.push(command.to_string());
                 args
             }
-            Shell::Unknown => shlex::split(command).unwrap_or_else(|| vec![command.to_string()]),
+            ShellType::Cmd => {
+                let mut args = vec![self.shell_path.to_string_lossy().to_string()];
+                args.push("/c".to_string());
+                args.push(command.to_string());
+                args
+            }
         }
     }
 }
@@ -146,19 +128,34 @@ fn get_shell_path(
     None
 }
 
-fn get_zsh_shell(path: Option<&PathBuf>) -> Option<ZshShell> {
+fn get_zsh_shell(path: Option<&PathBuf>) -> Option<Shell> {
     let shell_path = get_shell_path(ShellType::Zsh, path, "zsh", vec!["/bin/zsh"]);
 
-    shell_path.map(|shell_path| ZshShell { shell_path })
+    shell_path.map(|shell_path| Shell {
+        shell_type: ShellType::Zsh,
+        shell_path,
+    })
 }
 
-fn get_bash_shell(path: Option<&PathBuf>) -> Option<BashShell> {
+fn get_bash_shell(path: Option<&PathBuf>) -> Option<Shell> {
     let shell_path = get_shell_path(ShellType::Bash, path, "bash", vec!["/bin/bash"]);
 
-    shell_path.map(|shell_path| BashShell { shell_path })
+    shell_path.map(|shell_path| Shell {
+        shell_type: ShellType::Bash,
+        shell_path,
+    })
 }
 
-fn get_powershell_shell(path: Option<&PathBuf>) -> Option<PowerShellConfig> {
+fn get_sh_shell(path: Option<&PathBuf>) -> Option<Shell> {
+    let shell_path = get_shell_path(ShellType::Sh, path, "sh", vec!["/bin/sh"]);
+
+    shell_path.map(|shell_path| Shell {
+        shell_type: ShellType::Sh,
+        shell_path,
+    })
+}
+
+fn get_powershell_shell(path: Option<&PathBuf>) -> Option<Shell> {
     let shell_path = get_shell_path(
         ShellType::PowerShell,
         path,
@@ -167,32 +164,61 @@ fn get_powershell_shell(path: Option<&PathBuf>) -> Option<PowerShellConfig> {
     )
     .or_else(|| get_shell_path(ShellType::PowerShell, path, "powershell", vec![]));
 
-    shell_path.map(|shell_path| PowerShellConfig { shell_path })
+    shell_path.map(|shell_path| Shell {
+        shell_type: ShellType::PowerShell,
+        shell_path,
+    })
+}
+
+fn get_cmd_shell(path: Option<&PathBuf>) -> Option<Shell> {
+    let shell_path = get_shell_path(ShellType::Cmd, path, "cmd", vec![]);
+
+    shell_path.map(|shell_path| Shell {
+        shell_type: ShellType::Cmd,
+        shell_path,
+    })
+}
+
+fn ultimate_fallback_shell() -> Shell {
+    if cfg!(windows) {
+        Shell {
+            shell_type: ShellType::Cmd,
+            shell_path: PathBuf::from("cmd.exe"),
+        }
+    } else {
+        Shell {
+            shell_type: ShellType::Sh,
+            shell_path: PathBuf::from("/bin/sh"),
+        }
+    }
 }
 
 pub fn get_shell_by_model_provided_path(shell_path: &PathBuf) -> Shell {
     detect_shell_type(shell_path)
         .and_then(|shell_type| get_shell(shell_type, Some(shell_path)))
-        .unwrap_or(Shell::Unknown)
+        .unwrap_or(ultimate_fallback_shell())
 }
 
 pub fn get_shell(shell_type: ShellType, path: Option<&PathBuf>) -> Option<Shell> {
     match shell_type {
-        ShellType::Zsh => get_zsh_shell(path).map(Shell::Zsh),
-        ShellType::Bash => get_bash_shell(path).map(Shell::Bash),
-        ShellType::PowerShell => get_powershell_shell(path).map(Shell::PowerShell),
+        ShellType::Zsh => get_zsh_shell(path),
+        ShellType::Bash => get_bash_shell(path),
+        ShellType::PowerShell => get_powershell_shell(path),
+        ShellType::Sh => get_sh_shell(path),
+        ShellType::Cmd => get_cmd_shell(path),
     }
 }
 
 pub fn detect_shell_type(shell_path: &PathBuf) -> Option<ShellType> {
     match shell_path.as_os_str().to_str() {
         Some("zsh") => Some(ShellType::Zsh),
+        Some("sh") => Some(ShellType::Sh),
+        Some("cmd") => Some(ShellType::Cmd),
         Some("bash") => Some(ShellType::Bash),
         Some("pwsh") => Some(ShellType::PowerShell),
         Some("powershell") => Some(ShellType::PowerShell),
         _ => {
             let shell_name = shell_path.file_stem();
-
             if let Some(shell_name) = shell_name
                 && shell_name != shell_path
             {
@@ -204,14 +230,29 @@ pub fn detect_shell_type(shell_path: &PathBuf) -> Option<ShellType> {
     }
 }
 
-pub async fn default_user_shell() -> Shell {
+pub fn default_user_shell() -> Shell {
+    default_user_shell_from_path(get_user_shell_path())
+}
+
+fn default_user_shell_from_path(user_shell_path: Option<PathBuf>) -> Shell {
     if cfg!(windows) {
-        get_shell(ShellType::PowerShell, None).unwrap_or(Shell::Unknown)
+        get_shell(ShellType::PowerShell, None).unwrap_or(ultimate_fallback_shell())
     } else {
-        get_user_shell_path()
+        let user_default_shell = user_shell_path
             .and_then(|shell| detect_shell_type(&shell))
-            .and_then(|shell_type| get_shell(shell_type, None))
-            .unwrap_or(Shell::Unknown)
+            .and_then(|shell_type| get_shell(shell_type, None));
+
+        let shell_with_fallback = if cfg!(target_os = "macos") {
+            user_default_shell
+                .or_else(|| get_shell(ShellType::Zsh, None))
+                .or_else(|| get_shell(ShellType::Bash, None))
+        } else {
+            user_default_shell
+                .or_else(|| get_shell(ShellType::Bash, None))
+                .or_else(|| get_shell(ShellType::Zsh, None))
+        };
+
+        shell_with_fallback.unwrap_or(ultimate_fallback_shell())
     }
 }
 
@@ -251,6 +292,14 @@ 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)
@@ -259,6 +308,19 @@ mod detect_shell_type_tests {
             detect_shell_type(&PathBuf::from("/usr/local/bin/pwsh")),
             Some(ShellType::PowerShell)
         );
+        assert_eq!(
+            detect_shell_type(&PathBuf::from("/bin/sh")),
+            Some(ShellType::Sh)
+        );
+        assert_eq!(detect_shell_type(&PathBuf::from("sh")), Some(ShellType::Sh));
+        assert_eq!(
+            detect_shell_type(&PathBuf::from("cmd")),
+            Some(ShellType::Cmd)
+        );
+        assert_eq!(
+            detect_shell_type(&PathBuf::from("cmd.exe")),
+            Some(ShellType::Cmd)
+        );
     }
 }
 
@@ -274,10 +336,17 @@ mod tests {
     fn detects_zsh() {
         let zsh_shell = get_shell(ShellType::Zsh, None).unwrap();
 
-        let ZshShell { shell_path } = match zsh_shell {
-            Shell::Zsh(zsh_shell) => zsh_shell,
-            _ => panic!("expected zsh shell"),
-        };
+        let shell_path = zsh_shell.shell_path;
+
+        assert_eq!(shell_path, PathBuf::from("/bin/zsh"));
+    }
+
+    #[test]
+    #[cfg(target_os = "macos")]
+    fn fish_fallback_to_zsh() {
+        let zsh_shell = default_user_shell_from_path(Some(PathBuf::from("/bin/fish")));
+
+        let shell_path = zsh_shell.shell_path;
 
         assert_eq!(shell_path, PathBuf::from("/bin/zsh"));
     }
@@ -285,18 +354,60 @@ mod tests {
     #[test]
     fn detects_bash() {
         let bash_shell = get_shell(ShellType::Bash, None).unwrap();
-        let BashShell { shell_path } = match bash_shell {
-            Shell::Bash(bash_shell) => bash_shell,
-            _ => panic!("expected bash shell"),
-        };
+        let shell_path = bash_shell.shell_path;
 
         assert!(
             shell_path == PathBuf::from("/bin/bash")
-                || shell_path == PathBuf::from("/usr/bin/bash"),
+                || shell_path == PathBuf::from("/usr/bin/bash")
+                || shell_path == PathBuf::from("/usr/local/bin/bash"),
             "shell path: {shell_path:?}",
         );
     }
 
+    #[test]
+    fn detects_sh() {
+        let sh_shell = get_shell(ShellType::Sh, None).unwrap();
+        let shell_path = sh_shell.shell_path;
+        assert!(
+            shell_path == PathBuf::from("/bin/sh") || shell_path == PathBuf::from("/usr/bin/sh"),
+            "shell path: {shell_path:?}",
+        );
+    }
+
+    #[test]
+    fn can_run_on_shell_test() {
+        let cmd = "echo \"Works\"";
+        if cfg!(windows) {
+            assert!(shell_works(
+                get_shell(ShellType::PowerShell, None),
+                "Out-String 'Works'",
+                true,
+            ));
+            assert!(shell_works(get_shell(ShellType::Cmd, None), cmd, true,));
+            assert!(shell_works(Some(ultimate_fallback_shell()), cmd, true));
+        } else {
+            assert!(shell_works(Some(ultimate_fallback_shell()), cmd, true));
+            assert!(shell_works(get_shell(ShellType::Zsh, None), cmd, false));
+            assert!(shell_works(get_shell(ShellType::Bash, None), cmd, true));
+            assert!(shell_works(get_shell(ShellType::Sh, None), cmd, true));
+        }
+    }
+
+    fn shell_works(shell: Option<Shell>, command: &str, required: bool) -> bool {
+        if let Some(shell) = shell {
+            let args = shell.derive_exec_args(command, false);
+            let output = Command::new(args[0].clone())
+                .args(&args[1..])
+                .output()
+                .unwrap();
+            assert!(output.status.success());
+            assert!(String::from_utf8_lossy(&output.stdout).contains("Works"));
+            true
+        } else {
+            !required
+        }
+    }
+
     #[tokio::test]
     async fn test_current_shell_detects_zsh() {
         let shell = Command::new("sh")
@@ -308,10 +419,11 @@ mod tests {
         let shell_path = String::from_utf8_lossy(&shell.stdout).trim().to_string();
         if shell_path.ends_with("/zsh") {
             assert_eq!(
-                default_user_shell().await,
-                Shell::Zsh(ZshShell {
+                default_user_shell(),
+                Shell {
+                    shell_type: ShellType::Zsh,
                     shell_path: PathBuf::from(shell_path),
-                })
+                }
             );
         }
     }
@@ -322,11 +434,8 @@ mod tests {
             return;
         }
 
-        let powershell_shell = default_user_shell().await;
-        let PowerShellConfig { shell_path } = match powershell_shell {
-            Shell::PowerShell(powershell_shell) => powershell_shell,
-            _ => panic!("expected powershell shell"),
-        };
+        let powershell_shell = default_user_shell();
+        let shell_path = powershell_shell.shell_path;
 
         assert!(shell_path.ends_with("pwsh.exe") || shell_path.ends_with("powershell.exe"));
     }
@@ -338,10 +447,7 @@ mod tests {
         }
 
         let powershell_shell = get_shell(ShellType::PowerShell, None).unwrap();
-        let PowerShellConfig { shell_path } = match powershell_shell {
-            Shell::PowerShell(powershell_shell) => powershell_shell,
-            _ => panic!("expected powershell shell"),
-        };
+        let shell_path = powershell_shell.shell_path;
 
         assert!(shell_path.ends_with("pwsh.exe") || shell_path.ends_with("powershell.exe"));
     }

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

@@ -31,6 +31,8 @@ use crate::user_shell_command::user_shell_command_record_item;
 use super::SessionTask;
 use super::SessionTaskContext;
 
+const USER_SHELL_TIMEOUT_MS: u64 = 60 * 60 * 1000; // 1 hour
+
 #[derive(Clone)]
 pub(crate) struct UserShellCommandTask {
     command: String,
@@ -93,11 +95,15 @@ impl SessionTask for UserShellCommandTask {
             command: command.clone(),
             cwd: cwd.clone(),
             env: create_env(&turn_context.shell_environment_policy),
-            timeout_ms: None,
+            // TODO(zhao-oai): Now that we have ExecExpiration::Cancellation, we
+            // should use that instead of an "arbitrarily large" timeout here.
+            expiration: USER_SHELL_TIMEOUT_MS.into(),
             sandbox: SandboxType::None,
             with_escalated_permissions: None,
             justification: None,
             arg0: None,
+            max_output_tokens: None,
+            max_output_chars: None,
         };
 
         let stdout_stream = Some(StdoutStream {
@@ -122,7 +128,11 @@ 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;
@@ -164,12 +174,19 @@ impl SessionTask for UserShellCommandTask {
                             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;
@@ -201,11 +218,18 @@ impl SessionTask for UserShellCommandTask {
                             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/text_encoding.rs

@@ -0,0 +1,461 @@
+//! Text encoding detection and conversion utilities for shell output.
+//!
+//! Windows users frequently run into code pages such as CP1251 or CP866 when invoking commands
+//! through VS Code. Those bytes show up as invalid UTF-8 and used to be replaced with the standard
+//! Unicode replacement character. We now lean on `chardetng` and `encoding_rs` so we can
+//! automatically detect and decode the vast majority of legacy encodings before falling back to
+//! lossy UTF-8 decoding.
+
+use chardetng::EncodingDetector;
+use encoding_rs::Encoding;
+use encoding_rs::IBM866;
+use encoding_rs::WINDOWS_1252;
+
+/// Attempts to convert arbitrary bytes to UTF-8 with best-effort encoding detection.
+pub fn bytes_to_string_smart(bytes: &[u8]) -> String {
+    if bytes.is_empty() {
+        return String::new();
+    }
+
+    if let Ok(utf8_str) = std::str::from_utf8(bytes) {
+        return utf8_str.to_owned();
+    }
+
+    let encoding = detect_encoding(bytes);
+    decode_bytes(bytes, encoding)
+}
+
+// Windows-1252 reassigns a handful of 0x80-0x9F slots to smart punctuation (curly quotes, dashes,
+// ™). CP866 uses those *same byte values* for uppercase Cyrillic letters. When chardetng sees shell
+// snippets that mix these bytes with ASCII it sometimes guesses IBM866, so “smart quotes” render as
+// Cyrillic garbage (“УФЦ”) in VS Code. However, CP866 uppercase tokens are perfectly valid output
+// (e.g., `ПРИ test`) so we cannot flip every 0x80-0x9F byte to Windows-1252 either. The compromise
+// is to only coerce IBM866 to Windows-1252 when (a) the high bytes are exclusively the punctuation
+// values listed below and (b) we spot adjacent ASCII. This targets the real failure case without
+// clobbering legitimate Cyrillic text. If another code page has a similar collision, introduce a
+// dedicated allowlist (like this one) plus unit tests that capture the actual shell output we want
+// to preserve. Windows-1252 byte values for smart punctuation.
+const WINDOWS_1252_PUNCT_BYTES: [u8; 8] = [
+    0x91, // ‘ (left single quotation mark)
+    0x92, // ’ (right single quotation mark)
+    0x93, // “ (left double quotation mark)
+    0x94, // ” (right double quotation mark)
+    0x95, // • (bullet)
+    0x96, // – (en dash)
+    0x97, // — (em dash)
+    0x99, // ™ (trade mark sign)
+];
+
+fn detect_encoding(bytes: &[u8]) -> &'static Encoding {
+    let mut detector = EncodingDetector::new();
+    detector.feed(bytes, true);
+    let (encoding, _is_confident) = detector.guess_assess(None, true);
+
+    // chardetng occasionally reports IBM866 for short strings that only contain Windows-1252 “smart
+    // punctuation” bytes (0x80-0x9F) because that range maps to Cyrillic letters in IBM866. When
+    // those bytes show up alongside an ASCII word (typical shell output: `"“`test), we know the
+    // intent was likely CP1252 quotes/dashes. Prefer WINDOWS_1252 in that specific situation so we
+    // render the characters users expect instead of Cyrillic junk. References:
+    // - Windows-1252 reserving 0x80-0x9F for curly quotes/dashes:
+    //   https://en.wikipedia.org/wiki/Windows-1252
+    // - CP866 mapping 0x93/0x94/0x96 to Cyrillic letters, so the same bytes show up as “УФЦ” when
+    //   mis-decoded: https://www.unicode.org/Public/MAPPINGS/VENDORS/MICSFT/PC/CP866.TXT
+    if encoding == IBM866 && looks_like_windows_1252_punctuation(bytes) {
+        return WINDOWS_1252;
+    }
+
+    encoding
+}
+
+fn decode_bytes(bytes: &[u8], encoding: &'static Encoding) -> String {
+    let (decoded, _, had_errors) = encoding.decode(bytes);
+
+    if had_errors {
+        return String::from_utf8_lossy(bytes).into_owned();
+    }
+
+    decoded.into_owned()
+}
+
+/// Detect whether the byte stream looks like Windows-1252 “smart punctuation” wrapped around
+/// otherwise-ASCII text.
+///
+/// Context: IBM866 and Windows-1252 share the 0x80-0x9F slot range. In IBM866 these bytes decode to
+/// Cyrillic letters, whereas Windows-1252 maps them to curly quotes and dashes. chardetng can guess
+/// IBM866 for short snippets that only contain those bytes, which turns shell output such as
+/// `“test”` into unreadable Cyrillic. To avoid that, we treat inputs comprising a handful of bytes
+/// from the problematic range plus ASCII letters as CP1252 punctuation. We deliberately do *not*
+/// cap how many of those punctuation bytes we accept: VS Code frequently prints several quoted
+/// phrases (e.g., `"foo" – "bar"`), and truncating the count would once again mis-decode those as
+/// Cyrillic. If we discover additional encodings with overlapping byte ranges, prefer adding
+/// encoding-specific byte allowlists like `WINDOWS_1252_PUNCT` and tests that exercise real-world
+/// shell snippets.
+fn looks_like_windows_1252_punctuation(bytes: &[u8]) -> bool {
+    let mut saw_extended_punctuation = false;
+    let mut saw_ascii_word = false;
+
+    for &byte in bytes {
+        if byte >= 0xA0 {
+            return false;
+        }
+        if (0x80..=0x9F).contains(&byte) {
+            if !is_windows_1252_punct(byte) {
+                return false;
+            }
+            saw_extended_punctuation = true;
+        }
+        if byte.is_ascii_alphabetic() {
+            saw_ascii_word = true;
+        }
+    }
+
+    saw_extended_punctuation && saw_ascii_word
+}
+
+fn is_windows_1252_punct(byte: u8) -> bool {
+    WINDOWS_1252_PUNCT_BYTES.contains(&byte)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use encoding_rs::BIG5;
+    use encoding_rs::EUC_KR;
+    use encoding_rs::GBK;
+    use encoding_rs::ISO_8859_2;
+    use encoding_rs::ISO_8859_3;
+    use encoding_rs::ISO_8859_4;
+    use encoding_rs::ISO_8859_5;
+    use encoding_rs::ISO_8859_6;
+    use encoding_rs::ISO_8859_7;
+    use encoding_rs::ISO_8859_8;
+    use encoding_rs::ISO_8859_10;
+    use encoding_rs::ISO_8859_13;
+    use encoding_rs::SHIFT_JIS;
+    use encoding_rs::WINDOWS_874;
+    use encoding_rs::WINDOWS_1250;
+    use encoding_rs::WINDOWS_1251;
+    use encoding_rs::WINDOWS_1253;
+    use encoding_rs::WINDOWS_1254;
+    use encoding_rs::WINDOWS_1255;
+    use encoding_rs::WINDOWS_1256;
+    use encoding_rs::WINDOWS_1257;
+    use encoding_rs::WINDOWS_1258;
+    use pretty_assertions::assert_eq;
+
+    #[test]
+    fn test_utf8_passthrough() {
+        // Fast path: when UTF-8 is valid we should avoid copies and return as-is.
+        let utf8_text = "Hello, мир! 世界";
+        let bytes = utf8_text.as_bytes();
+        assert_eq!(bytes_to_string_smart(bytes), utf8_text);
+    }
+
+    #[test]
+    fn test_cp1251_russian_text() {
+        // Cyrillic text emitted by PowerShell/WSL in CP1251 should decode cleanly.
+        let bytes = b"\xEF\xF0\xE8\xEC\xE5\xF0"; // "пример" encoded with Windows-1251
+        assert_eq!(bytes_to_string_smart(bytes), "пример");
+    }
+
+    #[test]
+    fn test_cp1251_privet_word() {
+        // Regression: CP1251 words like "Привет" must not be mis-identified as Windows-1252.
+        let bytes = b"\xCF\xF0\xE8\xE2\xE5\xF2"; // "Привет" encoded with Windows-1251
+        assert_eq!(bytes_to_string_smart(bytes), "Привет");
+    }
+
+    #[test]
+    fn test_koi8_r_privet_word() {
+        // KOI8-R output should decode to the original Cyrillic as well.
+        let bytes = b"\xF0\xD2\xC9\xD7\xC5\xD4"; // "Привет" encoded with KOI8-R
+        assert_eq!(bytes_to_string_smart(bytes), "Привет");
+    }
+
+    #[test]
+    fn test_cp866_russian_text() {
+        // Legacy consoles (cmd.exe) commonly emit CP866 bytes for Cyrillic content.
+        let bytes = b"\xAF\xE0\xA8\xAC\xA5\xE0"; // "пример" encoded with CP866
+        assert_eq!(bytes_to_string_smart(bytes), "пример");
+    }
+
+    #[test]
+    fn test_cp866_uppercase_text() {
+        // Ensure the IBM866 heuristic still returns IBM866 for uppercase-only words.
+        let bytes = b"\x8F\x90\x88"; // "ПРИ" encoded with CP866 uppercase letters
+        assert_eq!(bytes_to_string_smart(bytes), "ПРИ");
+    }
+
+    #[test]
+    fn test_cp866_uppercase_followed_by_ascii() {
+        // Regression test: uppercase CP866 tokens next to ASCII text should not be treated as
+        // CP1252.
+        let bytes = b"\x8F\x90\x88 test"; // "ПРИ test" encoded with CP866 uppercase letters followed by ASCII
+        assert_eq!(bytes_to_string_smart(bytes), "ПРИ test");
+    }
+
+    #[test]
+    fn test_windows_1252_quotes() {
+        // Smart detection should map Windows-1252 punctuation into proper Unicode.
+        let bytes = b"\x93\x94test";
+        assert_eq!(bytes_to_string_smart(bytes), "\u{201C}\u{201D}test");
+    }
+
+    #[test]
+    fn test_windows_1252_multiple_quotes() {
+        // Longer snippets of punctuation (e.g., “foo” – “bar”) should still flip to CP1252.
+        let bytes = b"\x93foo\x94 \x96 \x93bar\x94";
+        assert_eq!(
+            bytes_to_string_smart(bytes),
+            "\u{201C}foo\u{201D} \u{2013} \u{201C}bar\u{201D}"
+        );
+    }
+
+    #[test]
+    fn test_windows_1252_privet_gibberish_is_preserved() {
+        // Windows-1252 cannot encode Cyrillic; if the input literally contains "ПÑ..." we should not "fix" it.
+        let bytes = "Привет".as_bytes();
+        assert_eq!(bytes_to_string_smart(bytes), "Привет");
+    }
+
+    #[test]
+    fn test_iso8859_1_latin_text() {
+        // ISO-8859-1 (code page 28591) is the Latin segment used by LatArCyrHeb.
+        // encoding_rs unifies ISO-8859-1 with Windows-1252, so reuse that constant here.
+        let (encoded, _, had_errors) = WINDOWS_1252.encode("Hello");
+        assert!(!had_errors, "failed to encode Latin sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Hello");
+    }
+
+    #[test]
+    fn test_iso8859_2_central_european_text() {
+        // ISO-8859-2 (code page 28592) covers additional Central European glyphs.
+        let (encoded, _, had_errors) = ISO_8859_2.encode("Příliš žluťoučký kůň");
+        assert!(!had_errors, "failed to encode ISO-8859-2 sample");
+        assert_eq!(
+            bytes_to_string_smart(encoded.as_ref()),
+            "Příliš žluťoučký kůň"
+        );
+    }
+
+    #[test]
+    fn test_iso8859_3_south_europe_text() {
+        // ISO-8859-3 (code page 28593) adds support for Maltese/Esperanto letters.
+        // chardetng rarely distinguishes ISO-8859-3 from neighboring Latin code pages, so we rely on
+        // an ASCII-only sample to ensure round-tripping still succeeds.
+        let (encoded, _, had_errors) = ISO_8859_3.encode("Esperanto and Maltese");
+        assert!(!had_errors, "failed to encode ISO-8859-3 sample");
+        assert_eq!(
+            bytes_to_string_smart(encoded.as_ref()),
+            "Esperanto and Maltese"
+        );
+    }
+
+    #[test]
+    fn test_iso8859_4_baltic_text() {
+        // ISO-8859-4 (code page 28594) targets the Baltic/Nordic repertoire.
+        let sample = "Šis ir rakstzīmju kodēšanas tests. Dažās valodās, kurās tiek \
+                      izmantotas latīņu valodas burti, lēmuma pieņemšanai mums ir nepieciešams \
+                      vairāk ieguldījuma.";
+        let (encoded, _, had_errors) = ISO_8859_4.encode(sample);
+        assert!(!had_errors, "failed to encode ISO-8859-4 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), sample);
+    }
+
+    #[test]
+    fn test_iso8859_5_cyrillic_text() {
+        // ISO-8859-5 (code page 28595) covers the Cyrillic portion.
+        let (encoded, _, had_errors) = ISO_8859_5.encode("Привет");
+        assert!(!had_errors, "failed to encode Cyrillic sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Привет");
+    }
+
+    #[test]
+    fn test_iso8859_6_arabic_text() {
+        // ISO-8859-6 (code page 28596) covers the Arabic glyphs.
+        let (encoded, _, had_errors) = ISO_8859_6.encode("مرحبا");
+        assert!(!had_errors, "failed to encode Arabic sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "مرحبا");
+    }
+
+    #[test]
+    fn test_iso8859_7_greek_text() {
+        // ISO-8859-7 (code page 28597) is used for Greek locales.
+        let (encoded, _, had_errors) = ISO_8859_7.encode("Καλημέρα");
+        assert!(!had_errors, "failed to encode ISO-8859-7 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Καλημέρα");
+    }
+
+    #[test]
+    fn test_iso8859_8_hebrew_text() {
+        // ISO-8859-8 (code page 28598) covers the Hebrew glyphs.
+        let (encoded, _, had_errors) = ISO_8859_8.encode("שלום");
+        assert!(!had_errors, "failed to encode Hebrew sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "שלום");
+    }
+
+    #[test]
+    fn test_iso8859_9_turkish_text() {
+        // ISO-8859-9 (code page 28599) mirrors Latin-1 but inserts Turkish letters.
+        // encoding_rs exposes the equivalent Windows-1254 mapping.
+        let (encoded, _, had_errors) = WINDOWS_1254.encode("İstanbul");
+        assert!(!had_errors, "failed to encode ISO-8859-9 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "İstanbul");
+    }
+
+    #[test]
+    fn test_iso8859_10_nordic_text() {
+        // ISO-8859-10 (code page 28600) adds additional Nordic letters.
+        let sample = "Þetta er prófun fyrir Ægir og Øystein.";
+        let (encoded, _, had_errors) = ISO_8859_10.encode(sample);
+        assert!(!had_errors, "failed to encode ISO-8859-10 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), sample);
+    }
+
+    #[test]
+    fn test_iso8859_11_thai_text() {
+        // ISO-8859-11 (code page 28601) mirrors TIS-620 / Windows-874 for Thai.
+        let sample = "ภาษาไทยสำหรับการทดสอบ ISO-8859-11";
+        // encoding_rs exposes the equivalent Windows-874 encoding, so use that constant.
+        let (encoded, _, had_errors) = WINDOWS_874.encode(sample);
+        assert!(!had_errors, "failed to encode ISO-8859-11 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), sample);
+    }
+
+    // ISO-8859-12 was never standardized, and encodings 14–16 cannot be distinguished reliably
+    // without the heuristics we removed (chardetng generally reports neighboring Latin pages), so
+    // we intentionally omit coverage for those slots until the detector can identify them.
+
+    #[test]
+    fn test_iso8859_13_baltic_text() {
+        // ISO-8859-13 (code page 28603) is common across Baltic languages.
+        let (encoded, _, had_errors) = ISO_8859_13.encode("Sveiki");
+        assert!(!had_errors, "failed to encode ISO-8859-13 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Sveiki");
+    }
+
+    #[test]
+    fn test_windows_1250_central_european_text() {
+        let (encoded, _, had_errors) = WINDOWS_1250.encode("Příliš žluťoučký kůň");
+        assert!(!had_errors, "failed to encode Central European sample");
+        assert_eq!(
+            bytes_to_string_smart(encoded.as_ref()),
+            "Příliš žluťoučký kůň"
+        );
+    }
+
+    #[test]
+    fn test_windows_1251_encoded_text() {
+        let (encoded, _, had_errors) = WINDOWS_1251.encode("Привет из Windows-1251");
+        assert!(!had_errors, "failed to encode Windows-1251 sample");
+        assert_eq!(
+            bytes_to_string_smart(encoded.as_ref()),
+            "Привет из Windows-1251"
+        );
+    }
+
+    #[test]
+    fn test_windows_1253_greek_text() {
+        let (encoded, _, had_errors) = WINDOWS_1253.encode("Γειά σου");
+        assert!(!had_errors, "failed to encode Greek sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Γειά σου");
+    }
+
+    #[test]
+    fn test_windows_1254_turkish_text() {
+        let (encoded, _, had_errors) = WINDOWS_1254.encode("İstanbul");
+        assert!(!had_errors, "failed to encode Turkish sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "İstanbul");
+    }
+
+    #[test]
+    fn test_windows_1255_hebrew_text() {
+        let (encoded, _, had_errors) = WINDOWS_1255.encode("שלום");
+        assert!(!had_errors, "failed to encode Windows-1255 Hebrew sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "שלום");
+    }
+
+    #[test]
+    fn test_windows_1256_arabic_text() {
+        let (encoded, _, had_errors) = WINDOWS_1256.encode("مرحبا");
+        assert!(!had_errors, "failed to encode Windows-1256 Arabic sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "مرحبا");
+    }
+
+    #[test]
+    fn test_windows_1257_baltic_text() {
+        let (encoded, _, had_errors) = WINDOWS_1257.encode("Pērkons");
+        assert!(!had_errors, "failed to encode Baltic sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Pērkons");
+    }
+
+    #[test]
+    fn test_windows_1258_vietnamese_text() {
+        let (encoded, _, had_errors) = WINDOWS_1258.encode("Xin chào");
+        assert!(!had_errors, "failed to encode Vietnamese sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "Xin chào");
+    }
+
+    #[test]
+    fn test_windows_874_thai_text() {
+        let (encoded, _, had_errors) = WINDOWS_874.encode("สวัสดีครับ นี่คือการทดสอบภาษาไทย");
+        assert!(!had_errors, "failed to encode Thai sample");
+        assert_eq!(
+            bytes_to_string_smart(encoded.as_ref()),
+            "สวัสดีครับ นี่คือการทดสอบภาษาไทย"
+        );
+    }
+
+    #[test]
+    fn test_windows_932_shift_jis_text() {
+        let (encoded, _, had_errors) = SHIFT_JIS.encode("こんにちは");
+        assert!(!had_errors, "failed to encode Shift-JIS sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "こんにちは");
+    }
+
+    #[test]
+    fn test_windows_936_gbk_text() {
+        let (encoded, _, had_errors) = GBK.encode("你好，世界，这是一个测试");
+        assert!(!had_errors, "failed to encode GBK sample");
+        assert_eq!(
+            bytes_to_string_smart(encoded.as_ref()),
+            "你好，世界，这是一个测试"
+        );
+    }
+
+    #[test]
+    fn test_windows_949_korean_text() {
+        let (encoded, _, had_errors) = EUC_KR.encode("안녕하세요");
+        assert!(!had_errors, "failed to encode Korean sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "안녕하세요");
+    }
+
+    #[test]
+    fn test_windows_950_big5_text() {
+        let (encoded, _, had_errors) = BIG5.encode("繁體");
+        assert!(!had_errors, "failed to encode Big5 sample");
+        assert_eq!(bytes_to_string_smart(encoded.as_ref()), "繁體");
+    }
+
+    #[test]
+    fn test_latin1_cafe() {
+        // Latin-1 bytes remain common in Western-European locales; decode them directly.
+        let bytes = b"caf\xE9"; // codespell:ignore caf
+        assert_eq!(bytes_to_string_smart(bytes), "café");
+    }
+
+    #[test]
+    fn test_preserves_ansi_sequences() {
+        // ANSI escape sequences should survive regardless of the detected encoding.
+        let bytes = b"\x1b[31mred\x1b[0m";
+        assert_eq!(bytes_to_string_smart(bytes), "\x1b[31mred\x1b[0m");
+    }
+
+    #[test]
+    fn test_fallback_to_lossy() {
+        // Completely invalid sequences fall back to the old lossy behavior.
+        let invalid_bytes = [0xFF, 0xFE, 0xFD];
+        let result = bytes_to_string_smart(&invalid_bytes);
+        assert_eq!(result, String::from_utf8_lossy(&invalid_bytes));
+    }
+}

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

@@ -15,6 +15,8 @@ use crate::protocol::PatchApplyEndEvent;
 use crate::protocol::TurnDiffEvent;
 use crate::tools::context::SharedTurnDiffTracker;
 use crate::tools::sandboxing::ToolError;
+use crate::truncate::TruncationPolicy;
+use crate::truncate::formatted_truncate_text;
 use codex_protocol::parse_command::ParsedCommand;
 use std::collections::HashMap;
 use std::path::Path;
@@ -29,6 +31,7 @@ pub(crate) struct ToolEventCtx<'a> {
     pub turn: &'a TurnContext,
     pub call_id: &'a str,
     pub turn_diff_tracker: Option<&'a SharedTurnDiffTracker>,
+    pub override_truncation_policy: Option<&'a TruncationPolicy>,
 }
 
 impl<'a> ToolEventCtx<'a> {
@@ -37,12 +40,14 @@ impl<'a> ToolEventCtx<'a> {
         turn: &'a TurnContext,
         call_id: &'a str,
         turn_diff_tracker: Option<&'a SharedTurnDiffTracker>,
+        override_truncation_policy: Option<&'a TruncationPolicy>,
     ) -> Self {
         Self {
             session,
             turn,
             call_id,
             turn_diff_tracker,
+            override_truncation_policy,
         }
     }
 }
@@ -88,6 +93,7 @@ pub(crate) enum ToolEmitter {
         cwd: PathBuf,
         source: ExecCommandSource,
         parsed_cmd: Vec<ParsedCommand>,
+        freeform: bool,
     },
     ApplyPatch {
         changes: HashMap<PathBuf, FileChange>,
@@ -103,13 +109,19 @@ pub(crate) enum ToolEmitter {
 }
 
 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,
         }
     }
 
@@ -144,6 +156,7 @@ impl ToolEmitter {
                     cwd,
                     source,
                     parsed_cmd,
+                    ..
                 },
                 stage,
             ) => {
@@ -171,15 +184,17 @@ impl ToolEmitter {
                         ctx.turn,
                         EventMsg::PatchApplyBegin(PatchApplyBeginEvent {
                             call_id: ctx.call_id.to_string(),
+                            turn_id: ctx.turn.sub_id.clone(),
                             auto_approved: *auto_approved,
                             changes: changes.clone(),
                         }),
                     )
                     .await;
             }
-            (Self::ApplyPatch { .. }, ToolEventStage::Success(output)) => {
+            (Self::ApplyPatch { changes, .. }, ToolEventStage::Success(output)) => {
                 emit_patch_end(
                     ctx,
+                    changes.clone(),
                     output.stdout.text.clone(),
                     output.stderr.text.clone(),
                     output.exit_code == 0,
@@ -187,11 +202,12 @@ impl ToolEmitter {
                 .await;
             }
             (
-                Self::ApplyPatch { .. },
+                Self::ApplyPatch { changes, .. },
                 ToolEventStage::Failure(ToolEventFailure::Output(output)),
             ) => {
                 emit_patch_end(
                     ctx,
+                    changes.clone(),
                     output.stdout.text.clone(),
                     output.stderr.text.clone(),
                     output.exit_code == 0,
@@ -199,10 +215,17 @@ impl ToolEmitter {
                 .await;
             }
             (
-                Self::ApplyPatch { .. },
+                Self::ApplyPatch { changes, .. },
                 ToolEventStage::Failure(ToolEventFailure::Message(message)),
             ) => {
-                emit_patch_end(ctx, String::new(), (*message).to_string(), false).await;
+                emit_patch_end(
+                    ctx,
+                    changes.clone(),
+                    String::new(),
+                    (*message).to_string(),
+                    false,
+                )
+                .await;
             }
             (
                 Self::UnifiedExec {
@@ -234,14 +257,30 @@ impl ToolEmitter {
         self.emit(ctx, ToolEventStage::Begin).await;
     }
 
+    fn format_exec_output_for_model(
+        &self,
+        output: &ExecToolCallOutput,
+        truncation_policy: &TruncationPolicy,
+    ) -> String {
+        match self {
+            Self::Shell { freeform: true, .. } => {
+                super::format_exec_output_for_model_freeform(output, *truncation_policy)
+            }
+            _ => super::format_exec_output_for_model_structured(output, *truncation_policy),
+        }
+    }
+
     pub async fn finish(
         &self,
         ctx: ToolEventCtx<'_>,
         out: Result<ExecToolCallOutput, ToolError>,
     ) -> Result<String, FunctionCallError> {
+        let truncation_policy = ctx
+            .override_truncation_policy
+            .unwrap_or(&ctx.turn.truncation_policy);
         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, truncation_policy);
                 let exit_code = output.exit_code;
                 let event = ToolEventStage::Success(output);
                 let result = if exit_code == 0 {
@@ -253,24 +292,26 @@ 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, truncation_policy);
                 let event = ToolEventStage::Failure(ToolEventFailure::Output(*output));
                 let result = Err(FunctionCallError::RespondToModel(response));
                 (event, result)
             }
             Err(ToolError::Codex(err)) => {
-                let message = format!("execution error: {err:?}");
-                let event = ToolEventStage::Failure(ToolEventFailure::Message(message.clone()));
-                let result = Err(FunctionCallError::RespondToModel(message));
+                let formatted_error = formatted_truncate_text(&err.to_string(), *truncation_policy);
+                let message = format!("execution error: {formatted_error}");
+                let event = ToolEventStage::Failure(ToolEventFailure::Message(message));
+                let result = Err(FunctionCallError::RespondToModel(formatted_error));
                 (event, result)
             }
             Err(ToolError::Rejected(msg)) => {
+                let formatted_msg = formatted_truncate_text(&msg, *truncation_policy);
                 // Normalize common rejection messages for exec tools so tests and
                 // users see a clear, consistent phrase.
-                let normalized = if msg == "rejected by user" {
+                let normalized = if formatted_msg == "rejected by user" {
                     "exec command rejected by user".to_string()
                 } else {
-                    msg
+                    formatted_msg
                 };
                 let event = ToolEventStage::Failure(ToolEventFailure::Message(normalized.clone()));
                 let result = Err(FunctionCallError::RespondToModel(normalized));
@@ -342,7 +383,7 @@ async fn emit_exec_stage(
                 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, ctx.turn.truncation_policy),
             };
             emit_exec_end(ctx, exec_input, exec_result).await;
         }
@@ -388,15 +429,23 @@ async fn emit_exec_end(
         .await;
 }
 
-async fn emit_patch_end(ctx: ToolEventCtx<'_>, stdout: String, stderr: String, success: bool) {
+async fn emit_patch_end(
+    ctx: ToolEventCtx<'_>,
+    changes: HashMap<PathBuf, FileChange>,
+    stdout: String,
+    stderr: String,
+    success: bool,
+) {
     ctx.session
         .send_event(
             ctx.turn,
             EventMsg::PatchApplyEnd(PatchApplyEndEvent {
                 call_id: ctx.call_id.to_string(),
+                turn_id: ctx.turn.sub_id.clone(),
                 stdout,
                 stderr,
                 success,
+                changes,
             }),
         )
         .await;

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

@@ -100,6 +100,7 @@ impl ToolHandler for ApplyPatchHandler {
                             turn.as_ref(),
                             &call_id,
                             Some(&tracker),
+                            None,
                         );
                         emitter.begin(event_ctx).await;
 
@@ -127,6 +128,7 @@ impl ToolHandler for ApplyPatchHandler {
                             turn.as_ref(),
                             &call_id,
                             Some(&tracker),
+                            None,
                         );
                         let content = emitter.finish(event_ctx, out).await?;
                         Ok(ToolOutput::Function {

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

@@ -9,9 +9,11 @@ use crate::apply_patch::convert_apply_patch_to_protocol;
 use crate::codex::TurnContext;
 use crate::exec::ExecParams;
 use crate::exec_env::create_env;
+use crate::exec_policy::create_approval_requirement_for_command;
 use crate::function_tool::FunctionCallError;
 use crate::is_safe_command::is_known_safe_command;
 use crate::protocol::ExecCommandSource;
+use crate::sandboxing::SandboxPermissions;
 use crate::tools::context::ToolInvocation;
 use crate::tools::context::ToolOutput;
 use crate::tools::context::ToolPayload;
@@ -25,6 +27,7 @@ use crate::tools::runtimes::apply_patch::ApplyPatchRuntime;
 use crate::tools::runtimes::shell::ShellRequest;
 use crate::tools::runtimes::shell::ShellRuntime;
 use crate::tools::sandboxing::ToolCtx;
+use crate::truncate::TruncationPolicy;
 
 pub struct ShellHandler;
 
@@ -35,11 +38,13 @@ impl ShellHandler {
         ExecParams {
             command: params.command,
             cwd: turn_context.resolve_path(params.workdir.clone()),
-            timeout_ms: params.timeout_ms,
+            expiration: params.timeout_ms.into(),
             env: create_env(&turn_context.shell_environment_policy),
             with_escalated_permissions: params.with_escalated_permissions,
             justification: params.justification,
             arg0: None,
+            max_output_tokens: params.max_output_tokens,
+            max_output_chars: params.max_output_chars,
         }
     }
 }
@@ -57,11 +62,13 @@ impl ShellCommandHandler {
         ExecParams {
             command,
             cwd: turn_context.resolve_path(params.workdir.clone()),
-            timeout_ms: params.timeout_ms,
+            expiration: params.timeout_ms.into(),
             env: create_env(&turn_context.shell_environment_policy),
             with_escalated_permissions: params.with_escalated_permissions,
             justification: params.justification,
             arg0: None,
+            max_output_tokens: params.max_output_tokens,
+            max_output_chars: params.max_output_chars,
         }
     }
 }
@@ -130,7 +137,7 @@ impl ToolHandler for ShellHandler {
                     turn,
                     tracker,
                     call_id,
-                    true,
+                    false,
                 )
                 .await
             }
@@ -178,7 +185,7 @@ impl ToolHandler for ShellCommandHandler {
             turn,
             tracker,
             call_id,
-            false,
+            true,
         )
         .await
     }
@@ -192,7 +199,7 @@ 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)
@@ -207,6 +214,9 @@ impl ShellHandler {
             )));
         }
 
+        let override_truncation_policy =
+            create_truncation_policy(exec_params.max_output_tokens, exec_params.max_output_chars);
+
         // Intercept apply_patch if present.
         match codex_apply_patch::maybe_parse_apply_patch_verified(
             &exec_params.command,
@@ -235,13 +245,14 @@ impl ShellHandler {
                             turn.as_ref(),
                             &call_id,
                             Some(&tracker),
+                            override_truncation_policy.as_ref(),
                         );
                         emitter.begin(event_ctx).await;
 
                         let req = ApplyPatchRequest {
                             patch: apply.action.patch.clone(),
                             cwd: apply.action.cwd.clone(),
-                            timeout_ms: exec_params.timeout_ms,
+                            timeout_ms: exec_params.expiration.timeout_ms(),
                             user_explicitly_approved: apply.user_explicitly_approved_this_action,
                             codex_exe: turn.codex_linux_sandbox_exe.clone(),
                         };
@@ -261,6 +272,7 @@ impl ShellHandler {
                             turn.as_ref(),
                             &call_id,
                             Some(&tracker),
+                            override_truncation_policy.as_ref(),
                         );
                         let content = emitter.finish(event_ctx, out).await?;
                         return Ok(ToolOutput::Function {
@@ -285,24 +297,38 @@ 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 event_ctx = ToolEventCtx::new(session.as_ref(), turn.as_ref(), &call_id, None);
+        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,
+            override_truncation_policy.as_ref(),
+        );
         emitter.begin(event_ctx).await;
 
         let req = ShellRequest {
             command: exec_params.command.clone(),
             cwd: exec_params.cwd.clone(),
-            timeout_ms: exec_params.timeout_ms,
+            timeout_ms: exec_params.expiration.timeout_ms(),
             env: exec_params.env.clone(),
             with_escalated_permissions: exec_params.with_escalated_permissions,
             justification: exec_params.justification.clone(),
+            max_output_tokens: exec_params.max_output_tokens,
+            max_output_chars: exec_params.max_output_chars,
+            approval_requirement: create_approval_requirement_for_command(
+                &turn.exec_policy,
+                &exec_params.command,
+                turn.approval_policy,
+                &turn.sandbox_policy,
+                SandboxPermissions::from(exec_params.with_escalated_permissions.unwrap_or(false)),
+            ),
         };
         let mut orchestrator = ToolOrchestrator::new();
         let mut runtime = ShellRuntime::new();
@@ -315,7 +341,13 @@ impl ShellHandler {
         let out = orchestrator
             .run(&mut runtime, &req, &tool_ctx, &turn, turn.approval_policy)
             .await;
-        let event_ctx = ToolEventCtx::new(session.as_ref(), turn.as_ref(), &call_id, None);
+        let event_ctx = ToolEventCtx::new(
+            session.as_ref(),
+            turn.as_ref(),
+            &call_id,
+            None,
+            override_truncation_policy.as_ref(),
+        );
         let content = emitter.finish(event_ctx, out).await?;
         Ok(ToolOutput::Function {
             content,
@@ -325,34 +357,45 @@ impl ShellHandler {
     }
 }
 
+fn create_truncation_policy(
+    max_output_tokens: Option<usize>,
+    max_output_chars: Option<usize>,
+) -> Option<TruncationPolicy> {
+    if let Some(max_output_tokens) = max_output_tokens {
+        Some(TruncationPolicy::Tokens(max_output_tokens))
+    } else {
+        max_output_chars.map(TruncationPolicy::Bytes)
+    }
+}
 #[cfg(test)]
 mod tests {
     use std::path::PathBuf;
 
     use crate::is_safe_command::is_known_safe_command;
-    use crate::shell::BashShell;
-    use crate::shell::PowerShellConfig;
     use crate::shell::Shell;
-    use crate::shell::ZshShell;
+    use crate::shell::ShellType;
 
     /// The logic for is_known_safe_command() has heuristics for known shells,
     /// so we must ensure the commands generated by [ShellCommandHandler] can be
     /// recognized as safe if the `command` is safe.
     #[test]
     fn commands_generated_by_shell_command_handler_can_be_matched_by_is_known_safe_command() {
-        let bash_shell = Shell::Bash(BashShell {
+        let bash_shell = Shell {
+            shell_type: ShellType::Bash,
             shell_path: PathBuf::from("/bin/bash"),
-        });
+        };
         assert_safe(&bash_shell, "ls -la");
 
-        let zsh_shell = Shell::Zsh(ZshShell {
+        let zsh_shell = Shell {
+            shell_type: ShellType::Zsh,
             shell_path: PathBuf::from("/bin/zsh"),
-        });
+        };
         assert_safe(&zsh_shell, "ls -la");
 
-        let powershell = Shell::PowerShell(PowerShellConfig {
+        let powershell = Shell {
+            shell_type: ShellType::PowerShell,
             shell_path: PathBuf::from("pwsh.exe"),
-        });
+        };
         assert_safe(&powershell, "ls -Name");
     }
 

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

@@ -162,6 +162,7 @@ impl ToolHandler for UnifiedExecHandler {
                     context.turn.as_ref(),
                     &context.call_id,
                     None,
+                    None,
                 );
                 let emitter = ToolEmitter::unified_exec(
                     &command,

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/orchestrator.rs

@@ -11,11 +11,13 @@ use crate::error::get_error_message_ui;
 use crate::exec::ExecToolCallOutput;
 use crate::sandboxing::SandboxManager;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ApprovalRequirement;
 use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::ToolCtx;
 use crate::tools::sandboxing::ToolError;
 use crate::tools::sandboxing::ToolRuntime;
+use crate::tools::sandboxing::default_approval_requirement;
 use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::ReviewDecision;
 
@@ -49,40 +51,52 @@ impl ToolOrchestrator {
         let otel_cfg = codex_otel::otel_event_manager::ToolDecisionSource::Config;
 
         // 1) Approval
-        let needs_initial_approval =
-            tool.wants_initial_approval(req, approval_policy, &turn_ctx.sandbox_policy);
         let mut already_approved = false;
 
-        if needs_initial_approval {
-            let mut risk = None;
-
-            if let Some(metadata) = req.sandbox_retry_data() {
-                risk = tool_ctx
-                    .session
-                    .assess_sandbox_command(turn_ctx, &tool_ctx.call_id, &metadata.command, None)
-                    .await;
+        let requirement = tool.approval_requirement(req).unwrap_or_else(|| {
+            default_approval_requirement(approval_policy, &turn_ctx.sandbox_policy)
+        });
+        match requirement {
+            ApprovalRequirement::Skip => {
+                otel.tool_decision(otel_tn, otel_ci, ReviewDecision::Approved, otel_cfg);
             }
+            ApprovalRequirement::Forbidden { reason } => {
+                return Err(ToolError::Rejected(reason));
+            }
+            ApprovalRequirement::NeedsApproval { reason } => {
+                let mut risk = None;
 
-            let approval_ctx = ApprovalCtx {
-                session: tool_ctx.session,
-                turn: turn_ctx,
-                call_id: &tool_ctx.call_id,
-                retry_reason: None,
-                risk,
-            };
-            let decision = tool.start_approval_async(req, approval_ctx).await;
-
-            otel.tool_decision(otel_tn, otel_ci, decision, otel_user.clone());
-
-            match decision {
-                ReviewDecision::Denied | ReviewDecision::Abort => {
-                    return Err(ToolError::Rejected("rejected by user".to_string()));
+                if let Some(metadata) = req.sandbox_retry_data() {
+                    risk = tool_ctx
+                        .session
+                        .assess_sandbox_command(
+                            turn_ctx,
+                            &tool_ctx.call_id,
+                            &metadata.command,
+                            None,
+                        )
+                        .await;
                 }
-                ReviewDecision::Approved | ReviewDecision::ApprovedForSession => {}
+
+                let approval_ctx = ApprovalCtx {
+                    session: tool_ctx.session,
+                    turn: turn_ctx,
+                    call_id: &tool_ctx.call_id,
+                    retry_reason: reason,
+                    risk,
+                };
+                let decision = tool.start_approval_async(req, approval_ctx).await;
+
+                otel.tool_decision(otel_tn, otel_ci, decision, otel_user.clone());
+
+                match decision {
+                    ReviewDecision::Denied | ReviewDecision::Abort => {
+                        return Err(ToolError::Rejected("rejected by user".to_string()));
+                    }
+                    ReviewDecision::Approved | ReviewDecision::ApprovedForSession => {}
+                }
+                already_approved = true;
             }
-            already_approved = true;
-        } else {
-            otel.tool_decision(otel_tn, otel_ci, ReviewDecision::Approved, otel_cfg);
         }
 
         // 2) First attempt under the selected sandbox.

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

@@ -116,6 +116,8 @@ impl ToolRouter {
                             timeout_ms: exec.timeout_ms,
                             with_escalated_permissions: None,
                             justification: None,
+                            max_output_tokens: None,
+                            max_output_chars: None,
                         };
                         Ok(Some(ToolCall {
                             tool_name: "local_shell".to_string(),

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

@@ -67,11 +67,13 @@ impl ApplyPatchRuntime {
             program,
             args: vec![CODEX_APPLY_PATCH_ARG1.to_string(), req.patch.clone()],
             cwd: req.cwd.clone(),
-            timeout_ms: req.timeout_ms,
+            expiration: req.timeout_ms.into(),
             // Run apply_patch with a minimal environment for determinism and to avoid leaks.
             env: HashMap::new(),
             with_escalated_permissions: None,
             justification: None,
+            max_output_tokens: None,
+            max_output_chars: None,
         })
     }
 
@@ -153,9 +155,9 @@ impl ToolRuntime<ApplyPatchRequest, ExecToolCallOutput> for ApplyPatchRuntime {
     ) -> Result<ExecToolCallOutput, ToolError> {
         let spec = Self::build_command_spec(req)?;
         let env = attempt
-            .env_for(&spec)
+            .env_for(spec)
             .map_err(|err| ToolError::Codex(err.into()))?;
-        let out = execute_env(&env, attempt.policy, Self::stdout_stream(ctx))
+        let out = execute_env(env, attempt.policy, Self::stdout_stream(ctx))
             .await
             .map_err(ToolError::Codex)?;
         Ok(out)

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

@@ -4,6 +4,7 @@ Module: runtimes
 Concrete ToolRuntime implementations for specific tools. Each runtime stays
 small and focused and reuses the orchestrator for approvals + sandbox + retry.
 */
+use crate::exec::ExecExpiration;
 use crate::sandboxing::CommandSpec;
 use crate::tools::sandboxing::ToolError;
 use std::collections::HashMap;
@@ -15,13 +16,16 @@ pub mod unified_exec;
 
 /// Shared helper to construct a CommandSpec from a tokenized command line.
 /// Validates that at least a program is present.
+#[allow(clippy::too_many_arguments)]
 pub(crate) fn build_command_spec(
     command: &[String],
     cwd: &Path,
     env: &HashMap<String, String>,
-    timeout_ms: Option<u64>,
+    expiration: ExecExpiration,
     with_escalated_permissions: Option<bool>,
     justification: Option<String>,
+    max_output_tokens: Option<usize>,
+    max_output_chars: Option<usize>,
 ) -> Result<CommandSpec, ToolError> {
     let (program, args) = command
         .split_first()
@@ -31,8 +35,10 @@ pub(crate) fn build_command_spec(
         args: args.to_vec(),
         cwd: cwd.to_path_buf(),
         env: env.clone(),
-        timeout_ms,
+        expiration,
         with_escalated_permissions,
         justification,
+        max_output_tokens,
+        max_output_chars,
     })
 }

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

@@ -4,13 +4,12 @@ Runtime: shell
 Executes shell requests under the orchestrator: asks for approval when needed,
 builds a CommandSpec, and runs it under the current SandboxAttempt.
 */
-use crate::command_safety::is_dangerous_command::requires_initial_appoval;
 use crate::exec::ExecToolCallOutput;
-use crate::protocol::SandboxPolicy;
 use crate::sandboxing::execute_env;
 use crate::tools::runtimes::build_command_spec;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ApprovalRequirement;
 use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::SandboxRetryData;
@@ -20,7 +19,6 @@ use crate::tools::sandboxing::ToolCtx;
 use crate::tools::sandboxing::ToolError;
 use crate::tools::sandboxing::ToolRuntime;
 use crate::tools::sandboxing::with_cached_approval;
-use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::ReviewDecision;
 use futures::future::BoxFuture;
 use std::path::PathBuf;
@@ -33,6 +31,9 @@ pub struct ShellRequest {
     pub env: std::collections::HashMap<String, String>,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
+    pub max_output_tokens: Option<usize>,
+    pub max_output_chars: Option<usize>,
+    pub approval_requirement: ApprovalRequirement,
 }
 
 impl ProvidesSandboxRetryData for ShellRequest {
@@ -114,18 +115,8 @@ impl Approvable<ShellRequest> for ShellRuntime {
         })
     }
 
-    fn wants_initial_approval(
-        &self,
-        req: &ShellRequest,
-        policy: AskForApproval,
-        sandbox_policy: &SandboxPolicy,
-    ) -> bool {
-        requires_initial_appoval(
-            policy,
-            sandbox_policy,
-            &req.command,
-            req.with_escalated_permissions.unwrap_or(false),
-        )
+    fn approval_requirement(&self, req: &ShellRequest) -> Option<ApprovalRequirement> {
+        Some(req.approval_requirement.clone())
     }
 
     fn wants_escalated_first_attempt(&self, req: &ShellRequest) -> bool {
@@ -144,14 +135,16 @@ impl ToolRuntime<ShellRequest, ExecToolCallOutput> for ShellRuntime {
             &req.command,
             &req.cwd,
             &req.env,
-            req.timeout_ms,
+            req.timeout_ms.into(),
             req.with_escalated_permissions,
             req.justification.clone(),
+            req.max_output_tokens,
+            req.max_output_chars,
         )?;
         let env = attempt
-            .env_for(&spec)
+            .env_for(spec)
             .map_err(|err| ToolError::Codex(err.into()))?;
-        let out = execute_env(&env, attempt.policy, Self::stdout_stream(ctx))
+        let out = execute_env(env, attempt.policy, Self::stdout_stream(ctx))
             .await
             .map_err(ToolError::Codex)?;
         Ok(out)

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

@@ -1,4 +1,3 @@
-use crate::command_safety::is_dangerous_command::requires_initial_appoval;
 /*
 Runtime: unified exec
 
@@ -7,9 +6,11 @@ the session manager to spawn PTYs once an ExecEnv is prepared.
 */
 use crate::error::CodexErr;
 use crate::error::SandboxErr;
+use crate::exec::ExecExpiration;
 use crate::tools::runtimes::build_command_spec;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ApprovalRequirement;
 use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::SandboxRetryData;
@@ -22,9 +23,7 @@ use crate::tools::sandboxing::with_cached_approval;
 use crate::unified_exec::UnifiedExecError;
 use crate::unified_exec::UnifiedExecSession;
 use crate::unified_exec::UnifiedExecSessionManager;
-use codex_protocol::protocol::AskForApproval;
 use codex_protocol::protocol::ReviewDecision;
-use codex_protocol::protocol::SandboxPolicy;
 use futures::future::BoxFuture;
 use std::collections::HashMap;
 use std::path::PathBuf;
@@ -36,6 +35,9 @@ pub struct UnifiedExecRequest {
     pub env: HashMap<String, String>,
     pub with_escalated_permissions: Option<bool>,
     pub justification: Option<String>,
+    pub max_output_tokens: Option<usize>,
+    pub max_output_chars: Option<usize>,
+    pub approval_requirement: ApprovalRequirement,
 }
 
 impl ProvidesSandboxRetryData for UnifiedExecRequest {
@@ -65,6 +67,7 @@ impl UnifiedExecRequest {
         env: HashMap<String, String>,
         with_escalated_permissions: Option<bool>,
         justification: Option<String>,
+        approval_requirement: ApprovalRequirement,
     ) -> Self {
         Self {
             command,
@@ -72,6 +75,9 @@ impl UnifiedExecRequest {
             env,
             with_escalated_permissions,
             justification,
+            max_output_tokens: None,
+            max_output_chars: None,
+            approval_requirement,
         }
     }
 }
@@ -129,18 +135,8 @@ impl Approvable<UnifiedExecRequest> for UnifiedExecRuntime<'_> {
         })
     }
 
-    fn wants_initial_approval(
-        &self,
-        req: &UnifiedExecRequest,
-        policy: AskForApproval,
-        sandbox_policy: &SandboxPolicy,
-    ) -> bool {
-        requires_initial_appoval(
-            policy,
-            sandbox_policy,
-            &req.command,
-            req.with_escalated_permissions.unwrap_or(false),
-        )
+    fn approval_requirement(&self, req: &UnifiedExecRequest) -> Option<ApprovalRequirement> {
+        Some(req.approval_requirement.clone())
     }
 
     fn wants_escalated_first_attempt(&self, req: &UnifiedExecRequest) -> bool {
@@ -159,13 +155,15 @@ impl<'a> ToolRuntime<UnifiedExecRequest, UnifiedExecSession> for UnifiedExecRunt
             &req.command,
             &req.cwd,
             &req.env,
-            None,
+            ExecExpiration::DefaultTimeout,
             req.with_escalated_permissions,
             req.justification.clone(),
+            req.max_output_tokens,
+            req.max_output_chars,
         )
         .map_err(|_| ToolError::Rejected("missing command line for PTY".to_string()))?;
         let exec_env = attempt
-            .env_for(&spec)
+            .env_for(spec)
             .map_err(|err| ToolError::Codex(err.into()))?;
         self.manager
             .open_session_with_exec_env(&exec_env)

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

@@ -86,6 +86,37 @@ pub(crate) struct ApprovalCtx<'a> {
     pub risk: Option<SandboxCommandAssessment>,
 }
 
+// Specifies what tool orchestrator should do with a given tool call.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(crate) enum ApprovalRequirement {
+    /// No approval required for this tool call
+    Skip,
+    /// Approval required for this tool call
+    NeedsApproval { reason: Option<String> },
+    /// Execution forbidden for this tool call
+    Forbidden { reason: String },
+}
+
+/// - Never, OnFailure: do not ask
+/// - OnRequest: ask unless sandbox policy is DangerFullAccess
+/// - UnlessTrusted: always ask
+pub(crate) fn default_approval_requirement(
+    policy: AskForApproval,
+    sandbox_policy: &SandboxPolicy,
+) -> ApprovalRequirement {
+    let needs_approval = match policy {
+        AskForApproval::Never | AskForApproval::OnFailure => false,
+        AskForApproval::OnRequest => !matches!(sandbox_policy, SandboxPolicy::DangerFullAccess),
+        AskForApproval::UnlessTrusted => true,
+    };
+
+    if needs_approval {
+        ApprovalRequirement::NeedsApproval { reason: None }
+    } else {
+        ApprovalRequirement::Skip
+    }
+}
+
 pub(crate) trait Approvable<Req> {
     type ApprovalKey: Hash + Eq + Clone + Debug + Serialize;
 
@@ -106,22 +137,11 @@ pub(crate) trait Approvable<Req> {
         matches!(policy, AskForApproval::Never)
     }
 
-    /// Decide whether an initial user approval should be requested before the
-    /// first attempt. Defaults to the orchestrator's behavior (pre‑refactor):
-    /// - Never, OnFailure: do not ask
-    /// - OnRequest: ask unless sandbox policy is DangerFullAccess
-    /// - UnlessTrusted: always ask
-    fn wants_initial_approval(
-        &self,
-        _req: &Req,
-        policy: AskForApproval,
-        sandbox_policy: &SandboxPolicy,
-    ) -> bool {
-        match policy {
-            AskForApproval::Never | AskForApproval::OnFailure => false,
-            AskForApproval::OnRequest => !matches!(sandbox_policy, SandboxPolicy::DangerFullAccess),
-            AskForApproval::UnlessTrusted => true,
-        }
+    /// Override the default approval requirement. Return `Some(_)` to specify
+    /// a custom requirement, or `None` to fall back to
+    /// policy-based default.
+    fn approval_requirement(&self, _req: &Req) -> Option<ApprovalRequirement> {
+        None
     }
 
     /// Decide we can request an approval for no-sandbox execution.
@@ -196,7 +216,7 @@ pub(crate) struct SandboxAttempt<'a> {
 impl<'a> SandboxAttempt<'a> {
     pub fn env_for(
         &self,
-        spec: &CommandSpec,
+        spec: CommandSpec,
     ) -> Result<crate::sandboxing::ExecEnv, SandboxTransformError> {
         self.manager.transform(
             spec,

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

@@ -8,6 +8,7 @@ use crate::tools::handlers::apply_patch::ApplyPatchToolType;
 use crate::tools::handlers::apply_patch::create_apply_patch_freeform_tool;
 use crate::tools::handlers::apply_patch::create_apply_patch_json_tool;
 use crate::tools::registry::ToolRegistryBuilder;
+use crate::truncate::TruncationPolicy;
 use serde::Deserialize;
 use serde::Serialize;
 use serde_json::Value as JsonValue;
@@ -17,7 +18,7 @@ use std::collections::HashMap;
 
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
 pub enum ConfigShellToolType {
-    Default,
+    Default(TruncationPolicy),
     Local,
     UnifiedExec,
     /// Do not include a shell tool by default. Useful when using Codex
@@ -26,7 +27,7 @@ pub enum ConfigShellToolType {
     /// to customize agent behavior.
     Disabled,
     /// Takes a command as a single string to be run in the user's default shell.
-    ShellCommand,
+    ShellCommand(TruncationPolicy),
 }
 
 #[derive(Debug, Clone)]
@@ -57,8 +58,6 @@ impl ToolsConfig {
             ConfigShellToolType::Disabled
         } else if features.enabled(Feature::UnifiedExec) {
             ConfigShellToolType::UnifiedExec
-        } else if features.enabled(Feature::ShellCommandTool) {
-            ConfigShellToolType::ShellCommand
         } else {
             model_family.shell_type.clone()
         };
@@ -266,7 +265,7 @@ fn create_write_stdin_tool() -> ToolSpec {
     })
 }
 
-fn create_shell_tool() -> ToolSpec {
+fn create_shell_tool(truncation_policy: TruncationPolicy) -> ToolSpec {
     let mut properties = BTreeMap::new();
     properties.insert(
         "command".to_string(),
@@ -300,6 +299,24 @@ fn create_shell_tool() -> ToolSpec {
             description: Some("Only set if with_escalated_permissions is true. 1-sentence explanation of why we want to run this command.".to_string()),
         },
     );
+    match truncation_policy {
+        TruncationPolicy::Tokens(_) => {
+            properties.insert(
+                "max_output_tokens".to_string(),
+                JsonSchema::Number {
+                    description: Some("Maximum number of tokens to return from stdout/stderr. Excess tokens will be truncated".to_string()),
+                },
+            );
+        }
+        TruncationPolicy::Bytes(_) => {
+            properties.insert(
+                "max_output_chars".to_string(),
+                JsonSchema::Number {
+                    description: Some("Maximum number of characters to return from stdout/stderr. Excess characters will be truncated".to_string()),
+                },
+            );
+        }
+    }
 
     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"].
@@ -330,7 +347,7 @@ Examples of valid command strings:
     })
 }
 
-fn create_shell_command_tool() -> ToolSpec {
+fn create_shell_command_tool(truncation_policy: TruncationPolicy) -> ToolSpec {
     let mut properties = BTreeMap::new();
     properties.insert(
         "command".to_string(),
@@ -364,6 +381,30 @@ fn create_shell_command_tool() -> ToolSpec {
             description: Some("Only set if with_escalated_permissions is true. 1-sentence explanation of why we want to run this command.".to_string()),
         },
     );
+    match truncation_policy {
+        TruncationPolicy::Tokens(_) => {
+            properties.insert(
+                "max_output_tokens".to_string(),
+                JsonSchema::Number {
+                    description: Some(
+                        "Maximum number of tokens to return. Excess output will be truncated."
+                            .to_string(),
+                    ),
+                },
+            );
+        }
+        TruncationPolicy::Bytes(_) => {
+            properties.insert(
+                "max_output_chars".to_string(),
+                JsonSchema::Number {
+                    description: Some(
+                        "Maximum number of tokens to return. Excess output will be truncated."
+                            .to_string(),
+                    ),
+                },
+            );
+        }
+    }
 
     let description = if cfg!(windows) {
         r#"Runs a Powershell command (Windows) and returns its output.
@@ -1001,8 +1042,8 @@ pub(crate) fn build_specs(
     let shell_command_handler = Arc::new(ShellCommandHandler);
 
     match &config.shell_type {
-        ConfigShellToolType::Default => {
-            builder.push_spec(create_shell_tool());
+        ConfigShellToolType::Default(truncation_policy) => {
+            builder.push_spec(create_shell_tool(*truncation_policy));
         }
         ConfigShellToolType::Local => {
             builder.push_spec(ToolSpec::LocalShell {});
@@ -1016,8 +1057,8 @@ pub(crate) fn build_specs(
         ConfigShellToolType::Disabled => {
             // Do nothing.
         }
-        ConfigShellToolType::ShellCommand => {
-            builder.push_spec(create_shell_command_tool());
+        ConfigShellToolType::ShellCommand(truncation_policy) => {
+            builder.push_spec(create_shell_command_tool(*truncation_policy));
         }
     }
 
@@ -1160,11 +1201,11 @@ mod tests {
 
     fn shell_tool_name(config: &ToolsConfig) -> Option<&'static str> {
         match config.shell_type {
-            ConfigShellToolType::Default => Some("shell"),
+            ConfigShellToolType::Default(_) => Some("shell"),
             ConfigShellToolType::Local => Some("local_shell"),
             ConfigShellToolType::UnifiedExec => None,
             ConfigShellToolType::Disabled => None,
-            ConfigShellToolType::ShellCommand => Some("shell_command"),
+            ConfigShellToolType::ShellCommand(_) => Some("shell_command"),
         }
     }
 
@@ -1292,7 +1333,7 @@ mod tests {
             "gpt-5-codex",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1309,7 +1350,7 @@ mod tests {
             "gpt-5.1-codex",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1384,7 +1425,7 @@ mod tests {
             "gpt-5.1-codex-mini",
             &Features::with_defaults(),
             &[
-                "shell",
+                "shell_command",
                 "list_mcp_resources",
                 "list_mcp_resource_templates",
                 "read_mcp_resource",
@@ -1395,13 +1436,29 @@ 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",
@@ -1452,22 +1509,6 @@ mod tests {
         assert_contains_tool_names(&tools, &subset);
     }
 
-    #[test]
-    fn test_build_specs_shell_command_present() {
-        assert_model_tools(
-            "codex-mini-latest",
-            Features::with_defaults().enable(Feature::ShellCommandTool),
-            &[
-                "shell_command",
-                "list_mcp_resources",
-                "list_mcp_resource_templates",
-                "read_mcp_resource",
-                "update_plan",
-                "view_image",
-            ],
-        );
-    }
-
     #[test]
     #[ignore]
     fn test_parallel_support_flags() {
@@ -1910,7 +1951,7 @@ mod tests {
 
     #[test]
     fn test_shell_tool() {
-        let tool = super::create_shell_tool();
+        let tool = super::create_shell_tool(TruncationPolicy::Bytes(10_000));
         let ToolSpec::Function(ResponsesApiTool {
             description, name, ..
         }) = &tool
@@ -1940,7 +1981,7 @@ Examples of valid command strings:
 
     #[test]
     fn test_shell_command_tool() {
-        let tool = super::create_shell_command_tool();
+        let tool = super::create_shell_command_tool(TruncationPolicy::Tokens(10_000));
         let ToolSpec::Function(ResponsesApiTool {
             description, name, ..
         }) = &tool

codex-rs/core/src/truncate.rs

@@ -2,48 +2,140 @@
 //! and suffix on UTF-8 boundaries, and helpers for line/token‑based truncation
 //! used across the core crate.
 
+use crate::config::Config;
 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 codex_utils_tokenizer::Tokenizer;
 
-/// 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
+const APPROX_BYTES_PER_TOKEN: usize = 4;
 
-/// Globally truncate function output items to fit within `MODEL_FORMAT_MAX_BYTES`
-/// by preserving as many text/image items as possible and appending a summary
-/// for any omitted text items.
-pub(crate) fn globally_truncate_function_output_items(
+#[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 = MODEL_FORMAT_MAX_BYTES;
+    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 == 0 {
+                if remaining_budget == 0 {
                     omitted_text_items += 1;
                     continue;
                 }
 
-                let len = text.len();
-                if len <= remaining {
+                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 -= len;
+                    remaining_budget = remaining_budget.saturating_sub(cost);
                 } else {
-                    let slice = take_bytes_at_char_boundary(text, remaining);
-                    if !slice.is_empty() {
-                        out.push(FunctionCallOutputContentItem::InputText {
-                            text: slice.to_string(),
-                        });
+                    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 = 0;
+                    remaining_budget = 0;
                 }
             }
-            // todo(aibrahim): handle input images; resize
             FunctionCallOutputContentItem::InputImage { image_url } => {
                 out.push(FunctionCallOutputContentItem::InputImage {
                     image_url: image_url.clone(),
@@ -61,513 +153,332 @@ pub(crate) fn globally_truncate_function_output_items(
     out
 }
 
-/// Format a block of exec/tool output for model consumption, truncating by
-/// lines and bytes while preserving head and tail segments.
-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();
+/// 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 output = truncate_formatted_exec_output(content, total_lines, limit_bytes, limit_lines);
-    format!("Total output lines: {total_lines}\n\n{output}")
-}
+    let max_tokens = policy.token_budget();
 
-fn truncate_formatted_exec_output(
-    content: &str,
-    total_lines: usize,
-    limit_bytes: usize,
-    limit_lines: usize,
-) -> String {
-    error_on_double_truncation(content);
-    let head_lines: usize = limit_lines / 2;
-    let tail_lines: usize = limit_lines - head_lines; // 128
-    let head_bytes: usize = limit_bytes / 2;
-    let segments: Vec<&str> = content.split_inclusive('\n').collect();
-    let head_take = head_lines.min(segments.len());
-    let tail_take = tail_lines.min(segments.len().saturating_sub(head_take));
-    let omitted = segments.len().saturating_sub(head_take + tail_take);
-
-    let head_slice_end: usize = segments
-        .iter()
-        .take(head_take)
-        .map(|segment| segment.len())
-        .sum();
-    let tail_slice_start: usize = if tail_take == 0 {
-        content.len()
-    } else {
-        content.len()
-            - segments
-                .iter()
-                .rev()
-                .take(tail_take)
-                .map(|segment| segment.len())
-                .sum::<usize>()
-    };
-    let head_slice = &content[..head_slice_end];
-    let tail_slice = &content[tail_slice_start..];
-    let truncated_by_bytes = content.len() > limit_bytes;
-    // this is a bit wrong. We are counting metadata lines and not just shell output lines.
-    let marker = if omitted > 0 {
-        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 error_on_double_truncation(content: &str) {
-    if content.contains("Total output lines:") && content.contains("omitted") {
-        tracing::error!(
-            "FunctionCallOutput content was already truncated before ContextManager::record_items; this would cause double truncation {content}"
-        );
-    }
-}
-
-/// Truncate an output string to a maximum number of “tokens”, where tokens are
-/// approximated as individual `char`s. Preserves a prefix and suffix with an
-/// elision marker describing how many tokens were omitted.
-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))
-}
-
-/// 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 {
+    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();
+
+    let (left_budget, right_budget) = split_budget(max_bytes);
+
+    let (removed_chars, left, right) = split_string(s, left_budget, right_budget);
+
+    let marker = format_truncation_marker(
+        policy,
+        removed_units_for_source(policy, total_bytes.saturating_sub(max_bytes), removed_chars),
+    );
+
+    assemble_truncated_output(left, right, &marker)
+}
+
+fn split_string(s: &str, beginning_bytes: usize, end_bytes: usize) -> (usize, &str, &str) {
+    if s.is_empty() {
+        return (0, "", "");
     }
 
-    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 len = s.len();
+    let tail_start_target = len.saturating_sub(end_bytes);
+    let mut prefix_end = 0usize;
+    let mut suffix_start = len;
+    let mut removed_chars = 0usize;
+    let mut suffix_started = false;
+
+    for (idx, ch) in s.char_indices() {
+        let char_end = idx + ch.len_utf8();
+        if char_end <= beginning_bytes {
+            prefix_end = char_end;
+            continue;
         }
 
-        let mut idx = start_tail.min(s.len());
-        while idx < s.len() && !s.is_char_boundary(idx) {
-            idx += 1;
+        if idx >= tail_start_target {
+            if !suffix_started {
+                suffix_start = idx;
+                suffix_started = true;
+            }
+            continue;
         }
-        idx
+
+        removed_chars = removed_chars.saturating_add(1);
     }
 
-    // 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 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 before = &s[..prefix_end];
+    let after = &s[suffix_start..];
+
+    (removed_chars, before, after)
+}
+
+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)
 }
 
 #[cfg(test)]
 mod tests {
-    use super::MODEL_FORMAT_MAX_BYTES;
-    use super::MODEL_FORMAT_MAX_LINES;
-    use super::format_output_for_model_body;
-    use super::globally_truncate_function_output_items;
-    use super::truncate_middle;
-    use super::truncate_output_to_tokens;
+
+    use super::TruncationPolicy;
+    use super::approx_token_count;
+    use super::formatted_truncate_text;
+    use super::split_string;
+    use super::truncate_function_output_items_with_policy;
+    use super::truncate_text;
+    use super::truncate_with_token_budget;
     use codex_protocol::models::FunctionCallOutputContentItem;
-    use codex_utils_tokenizer::Tokenizer;
     use pretty_assertions::assert_eq;
-    use regex_lite::Regex;
 
-    fn truncated_message_pattern(line: &str, total_lines: usize) -> String {
-        let head_lines = MODEL_FORMAT_MAX_LINES / 2;
-        let tail_lines = MODEL_FORMAT_MAX_LINES - head_lines;
-        let head_take = head_lines.min(total_lines);
-        let tail_take = tail_lines.min(total_lines.saturating_sub(head_take));
-        let omitted = total_lines.saturating_sub(head_take + tail_take);
-        let escaped_line = regex_lite::escape(line);
-        if omitted == 0 {
-            return format!(
-                r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes \.{{3}}]\n\n.*)$",
-            );
-        }
-        format!(
-            r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} omitted {omitted} of {total_lines} lines \.{{3}}]\n\n.*)$",
-        )
+    #[test]
+    fn split_string_works() {
+        assert_eq!(split_string("hello world", 5, 5), (1, "hello", "world"));
+        assert_eq!(split_string("abc", 0, 0), (3, "", ""));
     }
 
     #[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 split_string_handles_empty_string() {
+        assert_eq!(split_string("", 4, 4), (0, "", ""));
     }
 
     #[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);
-
-        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));
+    fn split_string_only_keeps_prefix_when_tail_budget_is_zero() {
+        assert_eq!(split_string("abcdef", 3, 0), (3, "abc", ""));
     }
 
     #[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);
-
-        assert!(out.contains("tokens truncated"));
-        assert!(!out.contains('\u{fffd}'));
-        assert_eq!(tokens, Some(tok.count(s) as u64));
+    fn split_string_only_keeps_suffix_when_prefix_budget_is_zero() {
+        assert_eq!(split_string("abcdef", 0, 3), (3, "", "def"));
     }
 
     #[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);
-
-        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));
+    fn split_string_handles_overlapping_budgets_without_removal() {
+        assert_eq!(split_string("abcdef", 4, 4), (0, "abcd", "ef"));
     }
 
     #[test]
-    fn truncate_output_to_tokens_returns_original_when_under_limit() {
+    fn split_string_respects_utf8_boundaries() {
+        assert_eq!(split_string("😀abc😀", 5, 5), (1, "😀a", "c😀"));
+
+        assert_eq!(split_string("😀😀😀😀😀", 1, 1), (5, "", ""));
+        assert_eq!(split_string("😀😀😀😀😀", 7, 7), (3, "😀", "😀"));
+        assert_eq!(split_string("😀😀😀😀😀", 8, 8), (1, "😀😀", "😀😀"));
+    }
+
+    #[test]
+    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_tokens_less_than_placeholder_returns_placeholder() {
+        let content = "example output";
+
+        assert_eq!(
+            "Total output lines: 1\n\nex…3 tokens truncated…ut",
+            formatted_truncate_text(content, TruncationPolicy::Tokens(1)),
+        );
+    }
+
+    #[test]
+    fn truncate_tokens_under_limit_returns_original() {
+        let content = "example output";
+
+        assert_eq!(
+            content,
+            formatted_truncate_text(content, TruncationPolicy::Tokens(10)),
+        );
+    }
+
+    #[test]
+    fn truncate_bytes_under_limit_returns_original() {
+        let content = "example output";
+
+        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 (truncated, original) = truncate_output_to_tokens(s, 100);
-        assert_eq!(truncated, s);
+        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_output_to_tokens_reports_truncation_at_zero_limit() {
+    fn truncate_with_token_budget_reports_truncation_at_zero_limit() {
         let s = "abcdef";
-        let (truncated, original) = truncate_output_to_tokens(s, 0);
-        assert!(truncated.contains("tokens truncated"));
-        assert_eq!(original, Some(s.chars().count()));
+        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_output_to_tokens_preserves_prefix_and_suffix() {
-        let s = "abcdefghijklmnopqrstuvwxyz";
-        let max_tokens = 10;
-        let (truncated, original) = truncate_output_to_tokens(s, max_tokens);
-        assert!(truncated.starts_with("abcde"));
-        assert!(truncated.ends_with("vwxyz"));
-        assert_eq!(original, Some(s.chars().count()));
+    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 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 = format_output_for_model_body(
-            &large_error,
-            MODEL_FORMAT_MAX_BYTES,
-            MODEL_FORMAT_MAX_LINES,
-        );
-
-        let total_lines = large_error.lines().count();
-        let pattern = truncated_message_pattern(line, total_lines);
-        let regex = Regex::new(&pattern).unwrap_or_else(|err| {
-            panic!("failed to compile regex {pattern}: {err}");
-        });
-        let captures = regex
-            .captures(&truncated)
-            .unwrap_or_else(|| panic!("message failed to match pattern {pattern}: {truncated}"));
-        let body = captures
-            .name("body")
-            .expect("missing body capture")
-            .as_str();
-        assert!(
-            body.len() <= MODEL_FORMAT_MAX_BYTES,
-            "body exceeds byte limit: {} bytes",
-            body.len()
-        );
-        assert_ne!(truncated, large_error);
-    }
-
-    #[test]
-    fn format_exec_output_marks_byte_truncation_without_omitted_lines() {
-        let long_line = "a".repeat(MODEL_FORMAT_MAX_BYTES + 50);
-        let truncated = format_output_for_model_body(
-            &long_line,
-            MODEL_FORMAT_MAX_BYTES,
-            MODEL_FORMAT_MAX_LINES,
-        );
-
-        assert_ne!(truncated, long_line);
-        let marker_line =
-            format!("[... output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes ...]");
-        assert!(
-            truncated.contains(&marker_line),
-            "missing byte truncation marker: {truncated}"
-        );
-        assert!(
-            !truncated.contains("omitted"),
-            "line omission marker should not appear when no lines were dropped: {truncated}"
-        );
-    }
-
-    #[test]
-    fn format_exec_output_returns_original_when_within_limits() {
-        let content = "example output\n".repeat(10);
-
-        assert_eq!(
-            format_output_for_model_body(&content, MODEL_FORMAT_MAX_BYTES, MODEL_FORMAT_MAX_LINES),
-            content
-        );
-    }
-
-    #[test]
-    fn format_exec_output_reports_omitted_lines_and_keeps_head_and_tail() {
-        let total_lines = MODEL_FORMAT_MAX_LINES + 100;
-        let content: String = (0..total_lines)
-            .map(|idx| format!("line-{idx}\n"))
-            .collect();
-
-        let truncated =
-            format_output_for_model_body(&content, MODEL_FORMAT_MAX_BYTES, MODEL_FORMAT_MAX_LINES);
-
-        let omitted = total_lines - MODEL_FORMAT_MAX_LINES;
-        let expected_marker = format!("[... omitted {omitted} of {total_lines} lines ...]");
-
-        assert!(
-            truncated.contains(&expected_marker),
-            "missing omitted marker: {truncated}"
-        );
-        assert!(
-            truncated.contains("line-0\n"),
-            "expected head line to remain: {truncated}"
-        );
-
-        let last_line = format!("line-{}\n", total_lines - 1);
-        assert!(
-            truncated.contains(&last_line),
-            "expected tail line to remain: {truncated}"
-        );
-    }
-
-    #[test]
-    fn format_exec_output_prefers_line_marker_when_both_limits_exceeded() {
-        let total_lines = MODEL_FORMAT_MAX_LINES + 42;
-        let long_line = "x".repeat(256);
-        let content: String = (0..total_lines)
-            .map(|idx| format!("line-{idx}-{long_line}\n"))
-            .collect();
-
-        let truncated =
-            format_output_for_model_body(&content, MODEL_FORMAT_MAX_BYTES, MODEL_FORMAT_MAX_LINES);
-
-        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}"
-        );
+    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() {
-        // Arrange: several text items, none exceeding per-item limit, but total exceeds budget.
-        let budget = 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 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 },
-            FunctionCallOutputContentItem::InputText { text: t2 },
+            FunctionCallOutputContentItem::InputText { text: t1.clone() },
+            FunctionCallOutputContentItem::InputText { text: t2.clone() },
             FunctionCallOutputContentItem::InputImage {
                 image_url: "img:mid".to_string(),
             },
@@ -576,7 +487,8 @@ mod tests {
             FunctionCallOutputContentItem::InputText { text: t5 },
         ];
 
-        let output = globally_truncate_function_output_items(&items);
+        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);
@@ -585,13 +497,13 @@ mod tests {
             FunctionCallOutputContentItem::InputText { text } => text,
             other => panic!("unexpected first item: {other:?}"),
         };
-        assert_eq!(first_text.len(), t1_len);
+        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.len(), t2_len);
+        assert_eq!(second_text, &t2);
 
         assert_eq!(
             output[2],
@@ -604,7 +516,10 @@ mod tests {
             FunctionCallOutputContentItem::InputText { text } => text,
             other => panic!("unexpected fourth item: {other:?}"),
         };
-        assert_eq!(fourth_text.len(), remaining_after_t1_t2);
+        assert!(
+            fourth_text.contains("tokens truncated"),
+            "expected marker in truncated snippet: {fourth_text}"
+        );
 
         let summary_text = match &output[4] {
             FunctionCallOutputContentItem::InputText { text } => text,

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>,

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

@@ -2,23 +2,25 @@
 
 use std::collections::VecDeque;
 use std::sync::Arc;
-
 use tokio::sync::Mutex;
 use tokio::sync::Notify;
 use tokio::sync::mpsc;
 use tokio::sync::oneshot::error::TryRecvError;
 use tokio::task::JoinHandle;
 use tokio::time::Duration;
+use tokio_util::sync::CancellationToken;
 
 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)]
@@ -65,13 +67,18 @@ impl OutputBufferState {
 }
 
 pub(crate) type OutputBuffer = Arc<Mutex<OutputBufferState>>;
-pub(crate) type OutputHandles = (OutputBuffer, Arc<Notify>);
+pub(crate) struct OutputHandles {
+    pub(crate) output_buffer: OutputBuffer,
+    pub(crate) output_notify: Arc<Notify>,
+    pub(crate) cancellation_token: CancellationToken,
+}
 
 #[derive(Debug)]
 pub(crate) struct UnifiedExecSession {
     session: ExecCommandSession,
     output_buffer: OutputBuffer,
     output_notify: Arc<Notify>,
+    cancellation_token: CancellationToken,
     output_task: JoinHandle<()>,
     sandbox_type: SandboxType,
 }
@@ -84,9 +91,11 @@ impl UnifiedExecSession {
     ) -> Self {
         let output_buffer = Arc::new(Mutex::new(OutputBufferState::default()));
         let output_notify = Arc::new(Notify::new());
+        let cancellation_token = CancellationToken::new();
         let mut receiver = initial_output_rx;
         let buffer_clone = Arc::clone(&output_buffer);
         let notify_clone = Arc::clone(&output_notify);
+        let cancellation_token_clone = cancellation_token.clone();
         let output_task = tokio::spawn(async move {
             loop {
                 match receiver.recv().await {
@@ -97,7 +106,10 @@ impl UnifiedExecSession {
                         notify_clone.notify_waiters();
                     }
                     Err(tokio::sync::broadcast::error::RecvError::Lagged(_)) => continue,
-                    Err(tokio::sync::broadcast::error::RecvError::Closed) => break,
+                    Err(tokio::sync::broadcast::error::RecvError::Closed) => {
+                        cancellation_token_clone.cancel();
+                        break;
+                    }
                 }
             }
         });
@@ -106,6 +118,7 @@ impl UnifiedExecSession {
             session,
             output_buffer,
             output_notify,
+            cancellation_token,
             output_task,
             sandbox_type,
         }
@@ -116,10 +129,11 @@ impl UnifiedExecSession {
     }
 
     pub(super) fn output_handles(&self) -> OutputHandles {
-        (
-            Arc::clone(&self.output_buffer),
-            Arc::clone(&self.output_notify),
-        )
+        OutputHandles {
+            output_buffer: Arc::clone(&self.output_buffer),
+            output_notify: Arc::clone(&self.output_notify),
+            cancellation_token: self.cancellation_token.clone(),
+        }
     }
 
     pub(super) fn has_exited(&self) -> bool {
@@ -165,7 +179,10 @@ 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 {
@@ -194,20 +211,34 @@ impl UnifiedExecSession {
         };
 
         if exit_ready {
+            managed.signal_exit();
             managed.check_for_sandbox_denial().await?;
             return Ok(managed);
         }
 
-        tokio::pin!(exit_rx);
         if tokio::time::timeout(Duration::from_millis(50), &mut exit_rx)
             .await
             .is_ok()
         {
+            managed.signal_exit();
             managed.check_for_sandbox_denial().await?;
+            return Ok(managed);
         }
 
+        tokio::spawn({
+            let cancellation_token = managed.cancellation_token.clone();
+            async move {
+                let _ = exit_rx.await;
+                cancellation_token.cancel();
+            }
+        });
+
         Ok(managed)
     }
+
+    fn signal_exit(&self) {
+        self.cancellation_token.cancel();
+    }
 }
 
 impl Drop for UnifiedExecSession {

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

@@ -5,16 +5,19 @@ use tokio::sync::Notify;
 use tokio::sync::mpsc;
 use tokio::time::Duration;
 use tokio::time::Instant;
+use tokio_util::sync::CancellationToken;
 
 use crate::codex::Session;
 use crate::codex::TurnContext;
 use crate::exec::ExecToolCallOutput;
 use crate::exec::StreamOutput;
 use crate::exec_env::create_env;
+use crate::exec_policy::create_approval_requirement_for_command;
 use crate::protocol::BackgroundEventEvent;
 use crate::protocol::EventMsg;
 use crate::protocol::ExecCommandSource;
 use crate::sandboxing::ExecEnv;
+use crate::sandboxing::SandboxPermissions;
 use crate::tools::events::ToolEmitter;
 use crate::tools::events::ToolEventCtx;
 use crate::tools::events::ToolEventFailure;
@@ -23,6 +26,9 @@ 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;
@@ -35,8 +41,19 @@ use super::clamp_yield_time;
 use super::generate_chunk_id;
 use super::resolve_max_tokens;
 use super::session::OutputBuffer;
+use super::session::OutputHandles;
 use super::session::UnifiedExecSession;
-use crate::truncate::truncate_output_to_tokens;
+
+struct PreparedSessionHandles {
+    writer_tx: mpsc::Sender<Vec<u8>>,
+    output_buffer: OutputBuffer,
+    output_notify: Arc<Notify>,
+    cancellation_token: CancellationToken,
+    session_ref: Arc<Session>,
+    turn_ref: Arc<TurnContext>,
+    command: Vec<String>,
+    cwd: PathBuf,
+}
 
 impl UnifiedExecSessionManager {
     pub(crate) async fn exec_command(
@@ -63,14 +80,23 @@ impl UnifiedExecSessionManager {
         let yield_time_ms = clamp_yield_time(request.yield_time_ms);
 
         let start = Instant::now();
-        let (output_buffer, output_notify) = session.output_handles();
+        let OutputHandles {
+            output_buffer,
+            output_notify,
+            cancellation_token,
+        } = session.output_handles();
         let deadline = start + Duration::from_millis(yield_time_ms);
-        let collected =
-            Self::collect_output_until_deadline(&output_buffer, &output_notify, deadline).await;
+        let collected = Self::collect_output_until_deadline(
+            &output_buffer,
+            &output_notify,
+            &cancellation_token,
+            deadline,
+        )
+        .await;
         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 +111,8 @@ 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 +120,7 @@ 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()),
         };
 
@@ -123,15 +151,16 @@ impl UnifiedExecSessionManager {
     ) -> Result<UnifiedExecResponse, UnifiedExecError> {
         let session_id = request.session_id;
 
-        let (
+        let PreparedSessionHandles {
             writer_tx,
             output_buffer,
             output_notify,
+            cancellation_token,
             session_ref,
             turn_ref,
-            session_command,
-            session_cwd,
-        ) = self.prepare_session_handles(session_id).await?;
+            command: session_command,
+            cwd: session_cwd,
+        } = self.prepare_session_handles(session_id).await?;
 
         let interaction_emitter = ToolEmitter::unified_exec(
             &session_command,
@@ -145,6 +174,7 @@ impl UnifiedExecSessionManager {
                 turn_ref.as_ref(),
                 request.call_id,
                 None,
+                None,
             )
         };
         interaction_emitter
@@ -170,12 +200,18 @@ impl UnifiedExecSessionManager {
         let yield_time_ms = clamp_yield_time(request.yield_time_ms);
         let start = Instant::now();
         let deadline = start + Duration::from_millis(yield_time_ms);
-        let collected =
-            Self::collect_output_until_deadline(&output_buffer, &output_notify, deadline).await;
+        let collected = Self::collect_output_until_deadline(
+            &output_buffer,
+            &output_notify,
+            &cancellation_token,
+            deadline,
+        )
+        .await;
         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 +235,7 @@ impl UnifiedExecSessionManager {
             output,
             session_id,
             exit_code,
-            original_token_count,
+            original_token_count: Some(original_token_count),
             session_command: Some(session_command.clone()),
         };
 
@@ -258,44 +294,27 @@ impl UnifiedExecSessionManager {
     async fn prepare_session_handles(
         &self,
         session_id: i32,
-    ) -> Result<
-        (
-            mpsc::Sender<Vec<u8>>,
-            OutputBuffer,
-            Arc<Notify>,
-            Arc<Session>,
-            Arc<TurnContext>,
-            Vec<String>,
-            PathBuf,
-        ),
-        UnifiedExecError,
-    > {
+    ) -> Result<PreparedSessionHandles, UnifiedExecError> {
         let sessions = self.sessions.lock().await;
-        let (output_buffer, output_notify, writer_tx, session, turn, command, cwd) =
-            if let Some(entry) = sessions.get(&session_id) {
-                let (buffer, notify) = entry.session.output_handles();
-                (
-                    buffer,
-                    notify,
-                    entry.session.writer_sender(),
-                    Arc::clone(&entry.session_ref),
-                    Arc::clone(&entry.turn_ref),
-                    entry.command.clone(),
-                    entry.cwd.clone(),
-                )
-            } else {
-                return Err(UnifiedExecError::UnknownSessionId { session_id });
-            };
-
-        Ok((
-            writer_tx,
+        let entry = sessions
+            .get(&session_id)
+            .ok_or(UnifiedExecError::UnknownSessionId { session_id })?;
+        let OutputHandles {
             output_buffer,
             output_notify,
-            session,
-            turn,
-            command,
-            cwd,
-        ))
+            cancellation_token,
+        } = entry.session.output_handles();
+
+        Ok(PreparedSessionHandles {
+            writer_tx: entry.session.writer_sender(),
+            output_buffer,
+            output_notify,
+            cancellation_token,
+            session_ref: Arc::clone(&entry.session_ref),
+            turn_ref: Arc::clone(&entry.turn_ref),
+            command: entry.command.clone(),
+            cwd: entry.cwd.clone(),
+        })
     }
 
     async fn send_input(
@@ -351,6 +370,7 @@ impl UnifiedExecSessionManager {
             entry.turn_ref.as_ref(),
             &entry.call_id,
             None,
+            None,
         );
         let emitter = ToolEmitter::unified_exec(
             &entry.command,
@@ -384,6 +404,7 @@ impl UnifiedExecSessionManager {
             context.turn.as_ref(),
             &context.call_id,
             None,
+            None,
         );
         let emitter =
             ToolEmitter::unified_exec(command, cwd, ExecCommandSource::UnifiedExecStartup, None);
@@ -444,6 +465,13 @@ impl UnifiedExecSessionManager {
             create_env(&context.turn.shell_environment_policy),
             with_escalated_permissions,
             justification,
+            create_approval_requirement_for_command(
+                &context.turn.exec_policy,
+                command,
+                context.turn.approval_policy,
+                &context.turn.sandbox_policy,
+                SandboxPermissions::from(with_escalated_permissions.unwrap_or(false)),
+            ),
         );
         let tool_ctx = ToolCtx {
             session: context.session.as_ref(),
@@ -466,9 +494,13 @@ impl UnifiedExecSessionManager {
     pub(super) async fn collect_output_until_deadline(
         output_buffer: &OutputBuffer,
         output_notify: &Arc<Notify>,
+        cancellation_token: &CancellationToken,
         deadline: Instant,
     ) -> Vec<u8> {
+        const POST_EXIT_OUTPUT_GRACE: Duration = Duration::from_millis(25);
+
         let mut collected: Vec<u8> = Vec::with_capacity(4096);
+        let mut exit_signal_received = cancellation_token.is_cancelled();
         loop {
             let drained_chunks;
             let mut wait_for_output = None;
@@ -481,15 +513,27 @@ impl UnifiedExecSessionManager {
             }
 
             if drained_chunks.is_empty() {
+                exit_signal_received |= cancellation_token.is_cancelled();
                 let remaining = deadline.saturating_duration_since(Instant::now());
                 if remaining == Duration::ZERO {
                     break;
                 }
 
                 let notified = wait_for_output.unwrap_or_else(|| output_notify.notified());
+                if exit_signal_received {
+                    let grace = remaining.min(POST_EXIT_OUTPUT_GRACE);
+                    if tokio::time::timeout(grace, notified).await.is_err() {
+                        break;
+                    }
+                    continue;
+                }
+
                 tokio::pin!(notified);
+                let exit_notified = cancellation_token.cancelled();
+                tokio::pin!(exit_notified);
                 tokio::select! {
                     _ = &mut notified => {}
+                    _ = &mut exit_notified => exit_signal_received = true,
                     _ = tokio::time::sleep(remaining) => break,
                 }
                 continue;
@@ -499,6 +543,7 @@ impl UnifiedExecSessionManager {
                 collected.extend_from_slice(&chunk);
             }
 
+            exit_signal_received |= cancellation_token.is_cancelled();
             if Instant::now() >= deadline {
                 break;
             }