## Description
This PR adds an optional `metadata` field to `ResponseItem` for
Responses API calls. Only mechanical plumbing, no actual values
populated and sent yet. Turns out just adding a new field to
`ResponseItem` has quite a large blast radius already.
This change is backwards compatible because `metadata` is optional and
omitted when absent, so existing response items and rollout history
without it still deserialize and requests that do not set it keep the
same wire shape. For provider compatibility, we strip out `metadata`
before non-OpenAI Responses requests so Azure and AWS Bedrock never see
this field.
My followup PR here will actually make use of it to start storing and
passing along `turn_id`: https://github.com/openai/codex/pull/28360
## What changed
- Added `ResponseItemMetadata` with optional `turn_id`, plus optional
`metadata` on Responses API item variants and inter-agent communication.
- Preserved item metadata through response-item rewrites such as
truncation, missing tool-output synthesis, compaction history
rebuilding, visible-history conversion, rollout/resume, and generated
app-server schemas/types.
- Strip item metadata from non-OpenAI Responses requests while
preserving it for OpenAI-shaped requests.
- Updated the mechanical fixture/test construction churn required by the
new optional field.
## What
Introduce a `CodexResponsesMetadata` struct that defines all the core
metadata we send to Responses API. Example fields are `thread_id`,
`turn_id`, `window_id`, etc.
Going forward, `client_metadata["x-codex-turn-metadata"]` will be the
canonical way Codex sends metadata to Responses API across both HTTP and
websocket transports.
For now, we continue to emit the existing top-level HTTP headers and
top-level `client_metadata` fields from the same
`CodexResponsesMetadata` struct for compatibility reasons.
Also, app-server clients who specify additional
`responsesapi_client_metadata` via `turn/start` and `turn/steer` will
have those fields merged into
`client_metadata["x-codex-turn-metadata"]`, but cannot override the
reserved fields that core uses (i.e. the fields in
`CodexResponsesMetadata`).
## Why
Responses API request instrumentation is the source of truth for
downstream Codex analytics that join requests by Codex IDs such as
session, thread, turn, and context window. Before this change, those
values were assembled through several request-specific paths: HTTP
request bodies, websocket handshake headers, websocket `response.create`
payloads, compaction requests, and the rich `x-codex-turn-metadata`
envelope all had their own wiring.
That made metadata propagation easy to drift across API-key/direct
Responses API requests, ChatGPT-auth/proxied requests, websocket
requests, and compaction requests. It also made additions like
`window_id` error-prone because a field could be added to one transport
projection but missed in another.
## What changed
- Added `CodexResponsesMetadata` as the core-owned snapshot for Codex
metadata sent to ResponsesAPI.
- Render `client_metadata["x-codex-turn-metadata"]`, flat
`client_metadata` projections, and direct compatibility headers from
that same snapshot.
- Include the known Codex-owned fields in the turn metadata blob,
including installation/session/thread/turn/window IDs, request kind,
lineage, sandbox/workspace metadata, timing, and compaction details.
- Treat app-server `responsesapi_client_metadata` as enrichment for the
Codex turn metadata blob while preventing those extras from overriding
Codex-owned fields.
- Use the same metadata path for normal turns, websocket prewarm, local
compaction, remote v1 compaction, and remote v2 compaction.
- Keep websocket connection-only preconnect metadata separate so
handshakes carry compatibility identity headers without inventing a fake
turn metadata blob.
## Verification
- `cargo check -p codex-core`
- `just fix -p codex-core`
## Summary
Remove the dead experimental `persistExtendedHistory` app-server flag
and collapse rollout persistence to the single policy app-server already
used.
## What Changed
- Removed `persistExtendedHistory` from v2 thread start/resume/fork
params and deleted its deprecation notice path.
- Removed the persistence-mode enums and plumbing through core, rollout,
and thread-store.
- Made rollout filtering mode-free, keeping the existing limited
persisted-history behavior.
## Test Plan
- `just write-app-server-schema`
- `cargo nextest run --no-fail-fast -p codex-app-server-protocol
schema_fixtures`
- `cargo nextest run --no-fail-fast -p codex-app-server
thread_shell_command_history_responses_exclude_persisted_command_executions`
- `cargo nextest run --no-fail-fast -p codex-rollout -p
codex-thread-store`
- final `rg` for removed flag/type names
## Why
This PR
https://github.com/openai/codex/pull/24161#discussion_r3325692763
revealed a subagent data modeling issue, where we overloaded
`forked_from_id` to also mean `parent_thread_id`. That's incorrect since
guardian and review subagents can be a subagent and NOT fork the main
thread's history.
The solution here is to explicitly store a new `parent_thread_id` on
`SessionMeta`, alongside `forked_from_id` which already exists. While
we're at it, also expose it in the app-server protocol on the `Thread`
object.
A thread->subagent relationship and a fork of thread history are
orthogonal concepts.
## What Changed
- Added top-level `parent_thread_id` persistence on `SessionMeta` and
runtime/session plumbing through `SessionConfiguredEvent`,
`CodexSpawnArgs`, `SessionConfiguration`, `ThreadConfigSnapshot`,
`TurnContext`, and `ModelClient`.
- Made turn metadata, request headers, analytics, and subagent-start
events read the separate runtime/top-level parent field instead of
deriving general parent lineage from `SessionSource` or
`forked_from_thread_id`.
- Passed parent lineage separately at delegated subagent, review,
guardian, agent-job, and multi-agent spawn construction sites;
copied-history fork lineage remains derived only from `InitialHistory`.
- Persisted and exposed parent lineage through rollout/thread-store
projections and app-server v2 `Thread.parentThreadId`.
- Updated app-server README text and regenerated app-server schema
fixtures for the additive `parentThreadId` response field.
## Description
Bedrock currently only supports the implicit `default` service tier for
GPT models. This PR strips non-default service tier metadata from
Bedrock model catalogs so Codex does not advertise or send unsupported
tiers.
## What changed
- Normalize both built-in and configured Bedrock catalogs to
default-only service tier behavior.
- Add regression coverage for built-in and configured Bedrock catalogs.
## Validation
- `just fmt`
- `just test -p codex-model-provider`
## Why
We recently added `forked_from_thread_id` which lets us trace where a
thread's _context_ comes from, but we also want to understand subagent
lineage (e.g. which parent thread spawned this subagent? what kind of
subagent is it?) which is orthogonal.
This PR adds `parent_thread_id` and `subagent_kind` to the
`x-codex-turn-metadata` header sent to ResponsesAPI.
## What changed
- Adds `parent_thread_id` and `subagent_kind` to core-owned
`x-codex-turn-metadata`.
- Restores persisted `SessionSource` and `ThreadSource` from resumed
session metadata so cold-resumed subagent threads keep their lineage on
later Responses API requests.
- Centralizes parent-thread extraction on `SessionSource` /
`SubAgentSource` and reuses it in the Responses client, analytics, agent
control, and state parsing paths.
- Extends reserved-key, git-enrichment, thread-spawn, and app-server v2
metadata coverage for the new lineage fields.
## Verification
- Not run locally per request.
- Added focused coverage in `core/src/turn_metadata_tests.rs` and
`app-server/tests/suite/v2/client_metadata.rs`.
## Why
When Codex calls responsesapi, we currently send `session_id`,
`thread_id`, and `turn_id` among other things as
`client_metadata["x-codex-turn-metadata"]`. This PR adds
`forked_from_thread_id` which helps explain the "lineage" of a forked
thread.
## What's changed
- Track the immediate history source copied into a forked thread through
thread/session creation, including subagent and review turn metadata
paths.
- Include `forked_from_thread_id` in Codex turn metadata while
preventing turn-scoped Responses API client metadata from overwriting
Codex-owned lineage fields.
- Add coverage for fork lineage in turn metadata and the app-server
Responses API request path.
`#[serde(default)]` wasn't sufficient for our generated TS types to
reflect that clients didn't have to set them. We also need
`skip_serializing_if = "std::ops::Not::not"`. This is already a rule in
our agents.md file.
## Description
This PR makes `codex remote-control` behave like a foreground CLI
command by default. Running it now starts remote control, waits for
readiness, prints a clear status message with the machine name, and
stays alive until Ctrl-C.
Users who want daemon behavior can use `codex remote-control start`, and
`codex remote-control stop` now prints concise human-readable output.
`--json` remains available for scripts.
Implementation-wise, this now verifies the real app-server state instead
of just assuming startup worked. The CLI starts or connects to
app-server, probes its control socket, calls the `remoteControl/enable`
API, and waits for the remote-control status response/notification
before printing success.
For daemon mode, `codex remote-control start` also reports which managed
app-server binary was used, including its path and best-effort `codex
--version`, so failures are easier to diagnose.
## Examples
Example output:
```
> codex remote-control
Starting app-server with remote control enabled...
This machine is available for remote control as com-97826.
Press Ctrl-C to stop.
```
Error case using daemon (currently expected based on our publicly
released CLI version):
```
> ./target/debug/codex remote-control start
Starting app-server daemon with remote control enabled...
Error: app server did not become ready on /Users/owen/.codex/app-server-control/app-server-control.sock
Daemon used app-server:
path: /Users/owen/.codex/packages/standalone/current/codex
version: 0.130.0
Managed app-server stderr (/Users/owen/.codex/app-server-daemon/app-server.stderr.log):
error: unexpected argument '--remote-control' found
Usage: codex app-server [OPTIONS] [COMMAND]
For more information, try '--help'.
Caused by:
0: failed to connect to /Users/owen/.codex/app-server-control/app-server-control.sock
1: No such file or directory (os error 2)
```
## What changed
- `codex remote-control` now runs remote control in the foreground and
prints a Ctrl-C stop hint.
- `codex remote-control start` starts the daemon and waits for remote
control readiness before reporting success.
- `codex remote-control stop` reports stopped/not-running status in
plain language.
- Startup failures now include recent managed app-server stderr to make
daemon issues easier to diagnose.
- Added coverage for CLI output, readiness waiting, foreground shutdown,
and stderr log tailing.
## Why
`experimentalFeature/list` reports effective feature enablement, but
currently does not resolve it against a working directory where
project-local config.toml files can exist and toggle on/off features
when merged into the effective config after resolving the various config
layers. That means we effectively (and incorrectly) ignore features set
in project-local config.
To address that, this PR exposes an optional `thread_id` param which
allows us to load the thread's `cwd.
## Testing
- `cargo test -p codex-app-server-protocol`
- `cargo test -p codex-app-server experimental_feature_list`
## Why
To help improve `codex remote-control` CLI UX which I plan to do in a
followup, this PR adds `server-name` to the various remote control APIs:
- `remoteControl/enable`
- `remoteControl/disable`
- `remoteControl/status/changed`
Also, add a `remoteControl/status/read` API. This will be helpful in the
Codex App.
## Why
reapplies https://github.com/openai/codex/pull/22386 which was
previously reverted
Also, introduce `remoteControl/enable` and `remoteControl/disable`
app-server APIs to toggle on/off remote control at runtime for a given
running app-server instance.
## What Changed
- Adds experimental v2 RPCs:
- `remoteControl/enable`
- `remoteControl/disable`
- Adds `RemoteControlRequestProcessor` and routes the new RPCs through
it instead of `ConfigRequestProcessor`.
- Adds named `RemoteControlHandle::enable`, `disable`, and `status`
methods.
- Makes `remoteControl/enable` return an error when sqlite state DB is
unavailable, while keeping enrollment/websocket failures as async status
updates.
- Adds `AppServerRuntimeOptions.remote_control_enabled` and hidden
`--remote-control` flags for `codex app-server` and `codex-app-server`.
- Updates managed daemon startup to use `codex app-server
--remote-control --listen unix://`.
- Marks `Feature::RemoteControl` as removed and ignores
`[features].remote_control`.
- Updates app-server README entries for the new remote-control methods.
## Why
`codex remote-control` manages the app-server daemon with
`remote_control` enabled, but it previously only exposed an implicit
start path. Once started, there was no obvious top-level
`remote-control` command for stopping the daemon; users had to know
about the lower-level `codex app-server daemon stop` command.
The startup failure for missing managed installs was also ambiguous.
`codex remote-control` and daemon bootstrap require the standalone Codex
install under `CODEX_HOME/packages/standalone/current/codex`, but the
old error only said to install Codex first, which is unclear when
another `codex` binary is already on PATH. Now we add an explicit
instruction for how to get the standalone Codex install.
## What changed
- Converts `codex remote-control` into a command group while preserving
bare `codex remote-control` as the existing start behavior.
- Adds `codex remote-control start` as the explicit start path.
- Adds `codex remote-control stop`, which maps to app-server daemon
stop.
- Updates the shared daemon managed-install error to name the missing
standalone path, explain why that install is required, provide the
installer command, and tell users to rerun the command they just tried.
## Verification
- `cargo test -p codex-app-server-daemon`
- `cargo test -p codex-cli`
- `./target/debug/codex remote-control --help`
## Why
Remote control starts by letting `codex-backend` initialize against the
app-server as an infrastructure health/proxy client before the real
remote client connects. App-server initialization also sets the
process-wide `originator` from `client_info.name`, so `codex-backend`
could become the sticky originator for later model/API requests even
after the real client initialized.
## What changed
- Treat `codex-backend` as a non-originating initialize client,
alongside the existing `codex_app_server_daemon` probe client.
- Preserve normal per-connection initialize behavior, including session
metadata and initialize analytics.
- Add regression coverage that verifies `codex-backend` initialize does
not replace the default originator.
## Testing
- `cargo test -p codex-app-server --test all
initialize_codex_backend_does_not_override_originator`
## Why
`remote_control` can appear in `config.toml`, CLI feature overrides, and
the app-server config APIs. Before this PR, app-server startup treated
`config.features.enabled(Feature::RemoteControl)` as the signal to start
remote control ([base
code](https://github.com/openai/codex/blob/5e3ee5eddfa5333f2e0b011880abf0cbf92bd295/codex-rs/app-server/src/lib.rs#L678-L680)).
That meant a user with:
```toml
[features]
remote_control = true
```
would accidentally opt every app-server process into remote control.
Remote-control startup should instead be a per-process launch decision
made by CLI flags.
## What Changed
- Marks `Feature::RemoteControl` as `Stage::Removed`, keeping
`remote_control` as a known compatibility key while making it
config-inert.
- Adds a hidden `--remote-control` process flag to `codex app-server`
and standalone `codex-app-server`.
- Plumbs that flag through
`AppServerRuntimeOptions.remote_control_enabled` and makes app-server
startup use only that runtime option to decide whether to start remote
control.
- Removes the app-server config mutation hook that reloaded config and
toggled remote control at runtime.
- Updates managed daemon spawning to use `codex app-server
--remote-control --listen unix://` instead of `--enable remote_control`.
Config APIs can still list, read, write, and set `remote_control`; those
operations just no longer affect remote-control process enrollment.
## Why
Update `codex remote-control` to use the new app server daemon commands
instead.
- if the updater loop is not running, bootstrap the daemon with remote
control enabled (`codex app-server daemon bootstrap --remote-control`)
- otherwise, enable the persisted remote-control setting and start the
daemon normally
## Summary
Remote clients can still receive large `thread/resume` histories when
prior turns include MCP tool call payloads or image-generation results.
This adds a temporary response-only redaction path for the known remote
client names.
Longer term we will move towards fully paginated APIs backed by SQLite.
## Changes
- Redact MCP tool call payload-bearing fields in `thread/resume`
responses for `codex_chatgpt_android_remote` and
`codex_chatgpt_ios_remote`.
- Drop `imageGeneration` items from those `thread/resume` responses.
- Keep redaction out of persisted rollout files, `thread/read`,
`thread/turns/list`, live notifications, and token usage replay.
- Cover the behavior with app-server helper tests and a v2 resume
integration test that checks both remote clients plus a non-target
control client.
## Testing
- `cargo test -p codex-app-server thread_resume_redaction`
- `cargo test -p codex-app-server
thread_resume_redacts_payloads_for_chatgpt_remote_clients`
## Why
We'd like SQLite state to become required and load-bearing. As a first
step, let's remove the mechanism that allows us to blow away the SQLite
DB on a version bump, and instead rely on graceful migrations.
The original motivation
([PR](https://github.com/openai/codex/pull/10623)) behind this mechanism
was to care less about backwards compatibility while SQLite was being
landed, but I'd say it's quite important now to keep the data in it.
## What changed
- Make `STATE_DB_FILENAME` and `LOGS_DB_FILENAME` the full canonical
filenames: `state_5.sqlite` and `logs_2.sqlite`.
- Remove `STATE_DB_VERSION` / `LOGS_DB_VERSION` and the helper that
constructed filenames from versions.
- Stop `StateRuntime::init` from scanning for or deleting older SQLite
DB filenames at startup.
- Delete the tests that encoded legacy state/logs DB deletion behavior.
## Verification
- `cargo test -p codex-state`
## Why
The goal of this PR is to align on app-server and `ThreadStore` API
updates for paginating through large threads.
#### app-server
##### `thread/turns/list`
- Updates `thread/turns/list` to support `itemsView?: "notLoaded" |
"summary" | "full" | null`, defaulting to `summary`.
- Implements the current `thread/turns/list` behavior over the existing
persisted rollout-history fallback:
- `notLoaded` returns turn envelopes with empty `items`.
- `summary` returns the first user message and final assistant message
when available.
- `full` preserves the existing full item behavior.
Note that this method still uses the naive approach of loading the
entire rollout file, and returns just the filtered slice of the data.
Real pagination will come later by leveraging SQLite.
##### `thread/turns/items/list`
- Adds the experimental `thread/turns/items/list` protocol, schema,
dispatcher, and processor stub. The app-server currently returns
JSON-RPC `-32601` with `thread/turns/items/list is not supported yet`.
#### ThreadStore
- Adds the experimental `thread/turns/items/list` protocol, schema,
dispatcher, and processor stub. The app-server currently returns
JSON-RPC `-32601` with `thread/turns/items/list is not supported yet`.
- Adds `ThreadStore` contract types and stubbed methods for listing
thread turns and listing items within a turn.
- Adds a typed `StoredTurnStatus` and `StoredTurnError` to avoid baking
app-server API enums or lossy string status values into the store-facing
turn contract.
- Adds a typed `StoredTurnStatus` and `StoredTurnError` to avoid baking
app-server API enums or lossy string status values into the store-facing
turn contract.
This also sketches the storage abstraction we expect to need once turns
are indexed/stored. In particular, `notLoaded` is useful only if
ThreadStore can eventually list turn metadata without loading every
persisted item for each turn.
## Validation
- Added/updated protocol serialization coverage for the new request and
response shapes.
- Added app-server integration coverage for `thread/turns/list` default
summary behavior and all three `itemsView` modes.
- Added app-server integration coverage that `thread/turns/items/list`
returns the expected unsupported JSON-RPC error when experimental APIs
are enabled.
- Added thread-store coverage that the default trait methods return
`ThreadStoreError::Unsupported`.
No developers.openai.com documentation update is needed for this
internal experimental app-server API surface.
## Summary
`codex --enable remote_control app-server --listen off` is the current
way to start a headless, remote-controllable app-server, but it is hard
to remember and exposes implementation details.
This adds `codex remote-control` as a friendly top-level wrapper for
that flow. The command starts a foreground app-server with local
transports disabled and enables `remote_control` only for that
invocation.
## Changes
- Add a visible `codex remote-control` CLI subcommand.
- Launch app-server with `AppServerTransport::Off`.
- Append `features.remote_control=true` after root feature toggles so
the explicit command wins over `--disable remote_control`.
- Reject root `--remote` / `--remote-auth-token-env`, matching other
non-TUI subcommands.
- Add tests for parsing, launch defaults, override ordering, and remote
flag rejection.
## Verification
- `cargo test -p codex-cli`
- `just fix -p codex-cli`
## Why
`codex-rs/app-server-protocol/src/protocol/v2.rs` had grown into a
single ~12k-line definition file for the entire app-server v2 API.
This is purely a mechanical refactor to break up the monolithic `v2.rs`
file that contains all app-server API v2 types into more modular files,
grouped by resource (e.g. account, thread, turn, etc.).
`just write-app-server-schema` shows no real changes, so we can be sure
that this is purely an internal organizational change.
## What changed
- Replaced the monolithic `protocol/v2.rs` with a `protocol/v2/` module
tree and a small `mod.rs` that only declares and reexports modules.
- Grouped v2 API definitions by conceptual owner, including `account`,
`apps`, `collaboration_mode`, `command_exec`, `config`, `device_key`,
`experimental_feature`, `feedback`, `fs`, `hook`, `item`, `mcp`,
`model`, `notification`, `permissions`, `plugin`, `process`, `realtime`,
`review`, `thread`, `thread_data`, `turn`, and `windows_sandbox`.
- Moved v2 tests into `protocol/v2/tests.rs` so `mod.rs` stays small.
- Kept shared protocol helpers in `protocol/v2/shared.rs`, including the
enum mirroring macro and common cross-resource types.
- Co-located resource-specific notifications and server-request payloads
with the modules that own those resources.
- Regenerated app-server protocol schema fixtures. The schema diffs are
non-semantic newline-only changes after the refactor.
## Verification
- `cargo check -p codex-app-server-protocol`
- `cargo test -p codex-app-server-protocol`
- `just write-app-server-schema`
## Why
Taking a step to removing the `persistExtendedHistory` field. It's not
scalable to be persisting so much data in the rollout file and returning
it in the thread history.
When a client explicitly sends `true`, the server now tells that client
the parameter is deprecated and ignored so the caller has a clear
migration signal via the `deprecationNotice` notification.
## What changed
- Keep the `persist_extended_history` / `persistExtendedHistory` field
in the v2 protocol for compatibility, but document it as deprecated and
ignored.
- Ignore the parameter in app-server `thread/start`, `thread/resume`,
and `thread/fork`; those paths always use limited history persistence
now.
- Stop treating `persistExtendedHistory` as a running-thread resume
override mismatch.
- Emit a connection-scoped `deprecationNotice` when a request explicitly
sets `persist_extended_history: true`.
## Verification
- Added `thread_start_deprecates_persist_extended_history_true` to cover
the deprecation notice.
- `cargo test -p codex-app-server`
- `cargo test -p codex-app-server-protocol`
## Why
Whenever we return a thread's history (turns and items) over app-server,
always return the limited form as specified by the rollout policy
`EventPersistenceMode::Limited`, even if the thread was previously
started with `EventPersistenceMode::Extended`.
We're finding it is quite unscalable to be returning the extended
history, so let's apply the same filtering logic of the rollout policy
when we load and return the thread's history.
## What Changed
- Reuse the rollout persistence policy when reconstructing app-server
`ThreadItem` history so only `EventPersistenceMode::Limited` rollout
items are replayed into API turns.
- Route `thread/read`, `thread/resume`, `thread/fork`,
`thread/turns/list`, and rollback responses through the same filtered
app-server history projection.
- Keep live active turns intact when composing a response for a
currently running thread.
- Update command execution coverage so persisted extended command events
are excluded from returned history for `thread/read`, `thread/fork`, and
`thread/turns/list`.
## Test Plan
- `cargo test -p codex-app-server limited`
- `cargo test -p codex-app-server thread_shell_command`
- `cargo test -p codex-app-server thread_read`
- `cargo test -p codex-app-server thread_rollback`
- `cargo test -p codex-app-server thread_fork`
- `cargo test -p codex-app-server-protocol`
## Description
Ignore these top-level config keys when loading project-scoped
config.toml files:
```
"openai_base_url",
"chatgpt_base_url",
"model_provider",
"model_providers",
"profile",
"profiles",
"experimental_realtime_ws_base_url",
```
## What changed
- Add a project-local config denylist for credential-routing fields such
as `openai_base_url`, `chatgpt_base_url`, `model_provider`,
`model_providers`, `profile`, `profiles`, and
`experimental_realtime_ws_base_url`.
- Strip those fields from project config layers before they participate
in effective config merging, while leaving safe project-local settings
intact.
- Track ignored project-local keys on config layers and surface a
startup warning telling users to move those settings to user-level
`config.toml` if they intentionally need them.
- Update profile behavior coverage so project-local `profile` /
`profiles` entries are ignored instead of overriding user-level profile
selection.
## Verification
- `cargo test -p codex-config`
- `cargo test -p codex-core
project_layer_ignores_unsupported_config_keys`
- `cargo test -p codex-core project_profiles_are_ignored`
- `cargo test -p codex-core config::config_loader_tests`
The Codex App treats apply patch tool calls quite load-bearing in the UI
(always shown on a completed turn), so we'd like to persist
`EventMsg::ApplyPatchEnd` to guarantee that when a client reconnects to
app-server mid-turn, we always have the full diff to display at the end
of that turn.
## Why
Large MCP tool call outputs can make rollout JSONL files enormous. In
the session that motivated this change, the biggest JSONL records were:
- `event_msg/mcp_tool_call_end`
- `response_item/function_call_output`
both containing the same unbounded MCP payloads - just 3 MCP tool calls
that each were multi-hundred MBs 😱
This PR truncates both of those JSONL records.
## How
#### For `response_item/function_call_output`
Unified exec already bounds tool output before it is injected into
model-facing history, which also keeps the corresponding rollout
`response_item/function_call_output` records small.
MCP should follow the same pattern: truncate the model-facing tool
output at the tool-output boundary, while leaving code-mode/raw hook
consumers alone.
#### For `event_msg/mcp_tool_call_end`
`McpToolCallEnd` also needs its own bounded event copy because it is the
app-server/replay/UI event shape that backs `ThreadItem::McpToolCall`.
Unfortunately this is _not_ downstream of the `ToolOutput` trait.
## Model behavior
Model behavior is actually unchanged as a result of this PR.
Before this PR, MCP output was:
1. Converted to `FunctionCallOutput`.
2. Recorded into in-memory history.
3. Truncated by `ContextManager::record_items()` before later model
turns saw it.
After this branch, MCP output is truncated earlier, in
`McpToolOutput::response_payload()`, using the same helper. Then
`ContextManager::record_items()` sees an already-truncated output and
effectively has little/no additional work to do.
So the model should still see the same kind of truncated function-call
output. The practical difference is where truncation happens: earlier,
before rollout persistence/app-server emission can see the giant
payload.
## Verification
- `cargo test -p codex-core mcp_tool_output`
- `cargo test -p codex-core
mcp_tool_call::tests::truncate_mcp_tool_result_for_event`
- `cargo test -p codex-core
mcp_post_tool_use_payload_uses_model_tool_name_args_and_result`
- `just fmt`
- `just fix -p codex-core`
- `git diff --check`
## Description
This PR introduces `review_id` as the stable identifier for guardian
reviews and exposes it in app-server `item/autoApprovalReview/started`
and `item/autoApprovalReview/completed` events.
Internally, guardian rejection state is now keyed by `review_id` instead
of the reviewed tool item ID. `target_item_id` is still included when a
review maps to a concrete thread item, but it is no longer overloaded as
the review lifecycle identifier.
## Motivation
We'd like to give users the ability to preempt a guardian review while
it's running (approve or decline).
However, we can't implement the API that allows the user to override a
running guardian review because we didn't have a unique `review_id` per
guardian review. Using `target_item_id` is not correct since:
- with execve reviews, there can be multiple execve calls (and therefore
guardian reviews) per shell command
- with network policy reviews, there is no target item ID
The PR that actually implements user overrides will use `review_id` as
the stable identifier.
## Description
We reuse a guardian thread for a given user thread when we can. However,
we had always sent the full transcript history every time we made a
followup review request to an existing guardian thread.
This is especially bad for long guardian threads since we keep
re-appending old transcript entries instead of just what has changed.
The fix is to just send what's new.
**Caveat**: Whenever a thread is compacted or rolled back, we fall back
to sending the full transcript to guardian again since the thread's
history has been modified. However in the happy path we get a nice
optimization.
## Before
Initial guardian review sends the full parent transcript:
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please check the repo visibility and push the docs fix if needed.
[2] tool gh_repo_view call: {"repo":"openai/codex"}
[3] tool gh_repo_view result: repo visibility: public
[4] assistant: The repo is public; I now need approval to push the docs fix.
>>> TRANSCRIPT END
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
And a followup to the same guardian thread would send the full
transcript again (including items 1-4 we already sent):
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please check the repo visibility and push the docs fix if needed.
[2] tool gh_repo_view call: {"repo":"openai/codex"}
[3] tool gh_repo_view result: repo visibility: public
[4] assistant: The repo is public; I now need approval to push the docs fix.
[5] user: Please push the second docs fix too.
[6] assistant: I need approval for the second docs fix.
>>> TRANSCRIPT END
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
## After
Initial guardian review sends the full parent transcript (this is
unchanged):
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please check the repo visibility and push the docs fix if needed.
[2] tool gh_repo_view call: {"repo":"openai/codex"}
[3] tool gh_repo_view result: repo visibility: public
[4] assistant: The repo is public; I now need approval to push the docs fix.
>>> TRANSCRIPT END
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
But a followup now sends:
```
The following is the Codex agent history added since your last approval assessment. Continue the same review conversation...
>>> TRANSCRIPT DELTA START
[5] user: Please push the second docs fix too.
[6] assistant: I need approval for the second docs fix.
>>> TRANSCRIPT DELTA END
The Codex agent has requested the following next action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
## Description
This PR fixes `/debug-config` so it shows more of the active
requirements state, including reviewer requirements and managed feature
pins. This made it clear that legacy MDM config was setting
`approvals_reviewer = "guardian_subagent"` and that we were translating
that into a requirements constraint.
Also, translate `approvals_reviewer = "guardian_subagent"` (from legacy
managed_config.toml) to `allowed_approvals_reviewers: guardian_subagent,
user` instead of `allowed_approvals_reviewers: guardian_subagent`.
Example `/debug-config`:
```
Config layer stack (lowest precedence first):
1. system (/etc/codex/config.toml) (enabled)
2. user (/Users/owen/.codex/config.toml) (enabled)
3. project (/Users/owen/repos/codex/.codex/config.toml) (enabled)
4. legacy managed_config.toml (MDM) (enabled)
MDM value:
...
# Enable Guardian Mode
features.guardian_approval = true
approvals_reviewer = "guardian_subagent"
Requirements:
- allowed_approvals_reviewers: guardian_subagent, user (source: MDM managed_config.toml (legacy))
- features: apps=true, plugins=true (source: cloud requirements)
```
Before this PR, the `Requirements` section showed None.
## Description
This PR changes guardian transcript compaction so oversized
conversations no longer collapse into a nearly empty placeholder.
Before this change, if the retained user history alone exceeded the
message budget, guardian would replace the entire transcript with
`<transcript omitted to preserve budget for planned action>`!
That meant approvals, especially network approvals, could lose the
recent tool call and tool result that explained what guardian was
actually reviewing. Now we keep a compact but usable transcript instead
of dropping it all.
### Before
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
<transcript omitted to preserve budget for planned action>
>>> TRANSCRIPT END
Conversation transcript omitted due to size.
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
Retry reason:
Sandbox blocked outbound network access.
Assess the exact planned action below. Use read-only tool checks when local state matters.
Planned action JSON:
{
"tool": "network_access",
"target": "https://example.com:443",
"host": "example.com",
"protocol": "https",
"port": 443
}
>>> APPROVAL REQUEST END
```
### After
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please investigate why uploads to example.com are failing and retry if needed.
[8] user: If the request looks correct, go ahead and try again with network access.
[9] tool shell call: {"command":["curl","-X","POST","https://example.com/upload"],"cwd":"/repo"}
[10] tool shell result: sandbox blocked outbound network access
>>> TRANSCRIPT END
Some conversation entries were omitted.
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
Retry reason:
Sandbox blocked outbound network access.
Assess the exact planned action below. Use read-only tool checks when local state matters.
Planned action JSON:
{
"tool": "network_access",
"target": "https://example.com:443",
"host": "example.com",
"protocol": "https",
"port": 443
}
>>> APPROVAL REQUEST END
```
## Summary
This adds a stable Codex installation ID and includes it on Responses
API requests via `x-codex-installation-id` passed in via the
`client_metadata` field for analytics/debugging.
The main pieces are:
- persist a UUID in `$CODEX_HOME/installation_id`
- thread the installation ID into `ModelClient`
- send it in `client_metadata` on Responses requests so it works
consistently across HTTP and WebSocket transports
## Description
This PR makes the SQLite state runtime tolerate databases that have
already been migrated by a newer Codex binary.
Today, if an older CLI sees migration versions in `_sqlx_migrations`
that it doesn't know about, startup fails. This change relaxes that
check for the runtime migrators we use in `codex-state` so older
binaries can keep opening the DB in that case.
## Why
We can end up with mixed-version CLIs running against the same local
state DB. In that setup, treating "the database is ahead of me" as a
hard error is unnecessarily strict and breaks the older client even when
the migration history is otherwise fine.
## Follow-up
We still clean up versioned `state_*.sqlite` and `logs_*.sqlite` files
during init, so older binaries can treat newer DB files as legacy. That
should probably be tightened separately if we want mixed-version local
usage to be fully safe.
Guardian events were emitted a bit out of order for CommandExecution
items. This would make it hard for the frontend to render a guardian
auto-review, which has this payload:
```
pub struct ItemGuardianApprovalReviewStartedNotification {
pub thread_id: String,
pub turn_id: String,
pub target_item_id: String,
pub review: GuardianApprovalReview,
// FYI this is no longer a json blob
pub action: Option<JsonValue>,
}
```
There is a `target_item_id` the auto-approval review is referring to,
but the actual item had not been emitted yet.
Before this PR:
- `item/autoApprovalReview/started`
- `item/autoApprovalReview/completed`, and if approved...
- `item/started`
- `item/completed`
After this PR:
- `item/started`
- `item/autoApprovalReview/started`
- `item/autoApprovalReview/completed`
- `item/completed`
This lines up much better with existing patterns (i.e. human review in
`Default mode`, where app-server would send a server request to prompt
for user approval after `item/started`), and makes it easier for clients
to render what guardian is actually reviewing.
We do this following a similar pattern as `FileChange` (aka apply patch)
items, where we create a FileChange item and emit `item/started` if we
see the apply patch approval request, before the actual apply patch call
runs.
## Summary
This changes our V8 Bazel wiring so `simdutf` no longer comes from a
live `git_repository` fetch against Chromium's Googlesource host.
Instead, we pull `simdutf` from a pinned GitHub release archive and keep
the V8 `simdutf` target wired through the external repo. The
archive-backed target is set up to match the way V8 consumes `simdutf`
today, including the amalgamated `src/simdutf.cpp` entrypoint and the
internal files it includes.
## Why
CI was intermittently failing while Bazel tried to fetch:
`https://chromium.googlesource.com/chromium/src/third_party/simdutf/`
That fetch was returning HTTP 429s, which then fan out into failures in
the Bazel jobs, the SDK job, and the argument-comment lint jobs since
they all go through the same dependency resolution path.
## What changed
- replaced the `simdutf` `git_repository` in the patched V8 module deps
with a pinned `http_archive`
- pointed that archive at `simdutf` `v7.7.0` on GitHub
- added the archive hash so the fetch is deterministic
- kept the V8 BUILD patch pointing `:simdutf` at the external
`@simdutf//:simdutf` target
- configured the Bazel `cc_library` for the archive to use the
amalgamated `src/simdutf.cpp` source plus the internal headers / textual
includes it depends on
## Validation
- ran `bazel build @v8//:simdutf`
- confirmed the target builds successfully with the new archive-backed
wiring
## Description
Add requirements.toml support for `allowed_approvals_reviewers =
["user", "guardian_subagent"]`, so admins can now restrict the use of
guardian mode.
Note: If a user sets a reviewer that isn’t allowed by requirements.toml,
config loading falls back to the first allowed reviewer and emits a
startup warning.
The table below describes the possible admin controls.
| Admin intent | `requirements.toml` | User `config.toml` | End result |
|---|---|---|---|
| Leave Guardian optional | omit `allowed_approvals_reviewers` or set
`["user", "guardian_subagent"]` | user chooses `approvals_reviewer =
"user"` or `"guardian_subagent"` | Guardian off for `user`, on for
`guardian_subagent` + `approval_policy = "on-request"` |
| Force Guardian off | `allowed_approvals_reviewers = ["user"]` | any
user value | Effective reviewer is `user`; Guardian off |
| Force Guardian on | `allowed_approvals_reviewers =
["guardian_subagent"]` and usually `allowed_approval_policies =
["on-request"]` | any user reviewer value; user should also have
`approval_policy = "on-request"` unless policy is forced | Effective
reviewer is `guardian_subagent`; Guardian on when effective approval
policy is `on-request` |
| Allow both, but default to manual if user does nothing |
`allowed_approvals_reviewers = ["user", "guardian_subagent"]` | omit
`approvals_reviewer` | Effective reviewer is `user`; Guardian off |
| Allow both, and user explicitly opts into Guardian |
`allowed_approvals_reviewers = ["user", "guardian_subagent"]` |
`approvals_reviewer = "guardian_subagent"` and `approval_policy =
"on-request"` | Guardian on |
| Invalid admin config | `allowed_approvals_reviewers = []` | anything |
Config load error |
## Description
Previously the `action` field on `EventMsg::GuardianAssessment`, which
describes what Guardian is reviewing, was typed as an arbitrary JSON
blob. This PR cleans it up and defines a sum type representing all the
various actions that Guardian can review.
This is a breaking change (on purpose), which is fine because:
- the Codex app / VSCE does not actually use `action` at the moment
- the TUI code that consumes `action` is updated in this PR as well
- rollout files that serialized old `EventMsg::GuardianAssessment` will
just silently drop these guardian events
- the contract is defined as unstable, so other clients have a fair
warning :)
This will make things much easier for followup Guardian work.
## Why
The old guardian review payloads worked, but they pushed too much shape
knowledge into downstream consumers. The TUI had custom JSON parsing
logic for commands, patches, network requests, and MCP calls, and the
app-server protocol was effectively just passing through an opaque blob.
Typing this at the protocol boundary makes the contract clearer.
## Description
Dependent on:
- [responsesapi] https://github.com/openai/openai/pull/760991
- [codex-backend] https://github.com/openai/openai/pull/760985
`codex app-server -> codex-backend -> responsesapi` now reuses a
persistent websocket connection across many turns. This PR updates
tracing when using websockets so that each `response.create` websocket
request propagates the current tracing context, so we can get a holistic
end-to-end trace for each turn.
Tracing is propagated via special keys (`ws_request_header_traceparent`,
`ws_request_header_tracestate`) set in the `client_metadata` param in
Responses API.
Currently tracing on websockets is a bit broken because we only set
tracing context on ws connection time, so it's detached from a
`turn/start` request.
## Description
This PR fixes a bad first-turn failure mode in app-server when the
startup websocket prewarm hangs. Before this change, `initialize ->
thread/start -> turn/start` could sit behind the prewarm for up to five
minutes, so the client would not see `turn/started`, and even
`turn/interrupt` would block because the turn had not actually started
yet.
Now, we:
- set a (configurable) timeout of 15s for websocket startup time,
exposed as `websocket_startup_timeout_ms` in config.toml
- `turn/started` is sent immediately on `turn/start` even if the
websocket is still connecting
- `turn/interrupt` can be used to cancel a turn that is still waiting on
the websocket warmup
- the turn task will wait for the full 15s websocket warming timeout
before falling back
## Why
The old behavior made app-server feel stuck at exactly the moment the
client expects turn lifecycle events to start flowing. That was
especially painful for external clients, because from their point of
view the server had accepted the request but then went silent for
minutes.
## Configuring the websocket startup timeout
Can set it in config.toml like this:
```
[model_providers.openai]
supports_websockets = true
websocket_connect_timeout_ms = 15000
```
## Description
This PR expands tracing coverage across app-server thread startup, core
session initialization, and the Responses transport layer. It also gives
core dispatch spans stable operation-specific names so traces are easier
to follow than the old generic `submission_dispatch` spans.
Also use `fmt::Display` for types that we serialize in traces so we send
strings instead of rust types
This makes the test less flaky by checking the core invariant instead of
the full span chain.
Before, the test waited for several specific internal spans
(`submission_dispatch`, `session_task.turn`, `run_turn`) and asserted
their exact relationships. That was brittle because those spans are
exported asynchronously and are more of an implementation detail than
the thing we actually care about.
Now, the test only checks that:
- `turn/start` is on the expected remote trace with the expected remote
parent
- at least one representative core turn span on that same trace descends
from it
That keeps the sanity-check we want while making the test less sensitive
to timing and internal refactors.
## Summary
This PR keeps app-server RPC request trace context alive for the full
lifetime of the work that request kicks off (e.g. for `thread/start`,
this is `app-server rpc handler -> tokio background task -> core op
submissions`). Previously we lose trace lineage once the request handler
returns or hands work off to background tasks.
This approach is especially relevant for `thread/start` and other RPC
handlers that run in a non-blocking way. In the near future we'll most
likely want to make all app-server handlers run in a non-blocking way by
default, and only queue operations that must operate in order (e.g.
thread RPCs per thread?), so we want to make sure tracing in app-server
just generally works.
Depends on https://github.com/openai/codex/pull/14300
**Before**
<img width="155" height="207" alt="image"
src="https://github.com/user-attachments/assets/c9487459-36f1-436c-beb7-fafeb40737af"
/>
**After**
<img width="299" height="337" alt="image"
src="https://github.com/user-attachments/assets/727392b2-d072-4427-9dc4-0502d8652dea"
/>
## What changed
- Keep request-scoped trace context around until we send the final
response or error, or the connection closes.
- Thread that trace context through detached `thread/start` work so
background startup stays attached to the originating request.
- Pass request trace context through to downstream core operations,
including:
- thread creation
- resume/fork flows
- turn submission
- review
- interrupt
- realtime conversation operations
- Add tracing tests that verify:
- remote W3C trace context is preserved for `thread/start`
- remote W3C trace context is preserved for `turn/start`
- downstream core spans stay under the originating request span
- request-scoped tracing state is cleaned up correctly
- Clean up shutdown behavior so detached background tasks and spawned
threads are drained before process exit.
## Description
This PR trims `app-server-protocol`'s v1 surface down to the small set
of legacy types we still actually use.
Unfortunately, we can't delete all of them yet because:
- a few one-off v1 RPCs are still used by the Codex app
- a few of these app-server-protocol v1 types are actually imported by
core crates
This change deletes that unused RPC surface, keeps the remaining
compatibility types in place, and makes the crate root re-export only
the v1 structs that downstream crates still depend on.
## Why
The main goal here is to make the legacy protocol surface match reality.
Leaving a large pile of dead v1 structs in place makes it harder to tell
which compatibility paths are still intentional, and it keeps old
schema/types around even though nothing should be building against them
anymore.
This also gives us a cleaner boundary for future cleanup. Instead of
re-exporting all of `protocol::v1::*`, the crate now explicitly exposes
only the v1 types that are still live, which makes it much easier to see
what remains and delete more safely later.
## What changed
- Deleted the unused v1 RPC/request/response structs from
`app-server-protocol/src/protocol/v1.rs`.
- Kept the small set of v1 compatibility types that are still live,
including:
- `initialize`
- `getConversationSummary`
- `getAuthStatus`
- `gitDiffToRemote`
- legacy approval payloads
- config-related structs still used by downstream crates
- Replaced the blanket `pub use protocol::v1::*` export in
`app-server-protocol/src/lib.rs` with an explicit list of the remaining
supported v1 types.
- Regenerated the schema/type artifacts, which also updated the
`InitializeCapabilities` opt-out example to use `thread/started` instead
of the old `codex/event/session_configured` example.
## Validation
- `just write-app-server-schema`
- `cargo test -p codex-app-server-protocol`
## Follow-up
The next cleanup is to keep shrinking the remaining v1 compatibility
surface as callers migrate off it. Once the remaining consumers stop
importing these legacy types, we should be able to remove more of the v1
module and eventually stop exporting it from the crate root entirely.
## Description
This PR stops emitting legacy `codex/event/*` notifications from the
public app-server transports.
It's been a long time coming! app-server was still producing a raw
notification stream from core, alongside the typed app-server
notifications and server requests, for compatibility reasons. Now,
external clients should no longer be depending on those legacy
notifications, so this change removes them from the stdio and websocket
contract and updates the surrounding docs, examples, and tests to match.
### Caveat
I left the "in-process" version of app-server alone for now, since
`codex exec` was recently based on top of app-server via this in-process
form here: https://github.com/openai/codex/pull/14005
Seems like `codex exec` still consumes some legacy notifications
internally, so this branch only removes `codex/event/*` from app-server
over stdio and websockets.
## Follow-up
Once `codex exec` is fully migrated off `codex/event/*` notifications,
we'll be able to stop emitting them entirely entirely instead of just
filtering it at the external transport boundary.
Prevent binaries >500KB from being committed. And maintain an allowlist
if we need to bypass on a case-by-case basis.
I checked the currently tracked binary-like assets in the repo. There
are only 5 obvious committed binaries by extension/MIME type:
- `.github/codex-cli-splash.png`: `838,131` bytes, about `818 KiB`
- `codex-rs/vendor/bubblewrap/bubblewrap.jpg`: `40,239` bytes, about `39
KiB`
-
`codex-rs/skills/src/assets/samples/skill-creator/assets/skill-creator.png`:
`1,563` bytes
- `codex-rs/skills/src/assets/samples/openai-docs/assets/openai.png`:
`1,429` bytes
-
`codex-rs/skills/src/assets/samples/skill-installer/assets/skill-installer.png`:
`1,086` bytes
So `500 KB` looks like a good default for this repo. It would only trip
on one existing intentional asset, which keeps the allowlist small and
the policy easy to understand.
Here's a smoke-test from a throwaway branch that tries to commit a large
binary:
https://github.com/openai/codex/actions/runs/22971558828/job/66689330435?pr=14383
## Summary
This PR fixes OTLP HTTP trace export in runtimes where the previous
exporter setup was unreliable, especially around app-server usage. It
also removes the old `codex_otel::otel_provider` compatibility shim and
switches remaining call sites over to the crate-root
`codex_otel::OtelProvider` export.
## What changed
- Use a runtime-safe OTLP HTTP trace exporter path for Tokio runtimes.
- Add an async HTTP client path for trace export when we are already
inside a multi-thread Tokio runtime.
- Make provider shutdown flush traces before tearing down the tracer
provider.
- Add loopback coverage that verifies traces are actually sent to
`/v1/traces`:
- outside Tokio
- inside a multi-thread Tokio runtime
- inside a current-thread Tokio runtime
- Remove the `codex_otel::otel_provider` shim and update remaining
imports.
## Why
I hit cases where spans were being created correctly but never made it
to the collector. The issue turned out to be in exporter/runtime
behavior rather than the span plumbing itself. This PR narrows that gap
and gives us regression coverage for the actual export path.
This changes the web_search tool spec in codex-core to use dedicated
Responses-API payload structs instead of shared config types and custom
serializers.
Previously, `ToolSpec::WebSearch` stored `WebSearchFilters` and
`WebSearchUserLocation` directly and relied on hand-written serializers
to shape the outgoing JSON. This worked, but it mixed config/schema
types with the OpenAI Responses payload contract and created an easy
place for drift if those shared types changed later.
### Why
This keeps the boundary clearer:
- app-server/config/schema types stay focused on config
- Responses tool payload types stay focused on the OpenAI wire format
It also makes the serialization behavior obvious from the structs
themselves, instead of hiding it in custom serializer functions.
This cleans up a bunch of metric plumbing that had started to drift.
The main change is making `codex-otel` the canonical home for shared
metric definitions and metric tag helpers. I moved the `turn/thread`
metric names that were still duplicated into the OTEL metric registry,
added a shared `metrics::tags` module for common tag keys and session
tag construction, and updated `SessionTelemetry` to build its metadata
tags through that shared path.
On the codex-core side, TTFT/TTFM now use the shared metric-name
constants instead of local string definitions. I also switched the
obvious remaining turn/thread metric callsites over to the shared
constants, and added a small helper so TTFT/TTFM can attach an optional
sanitized client.name tag from TurnContext.
This should make follow-on telemetry work less ad hoc:
- one canonical place for metric names
- one canonical place for common metric tag keys/builders
- less duplication between `codex-core` and `codex-otel`
This fixes a schema export bug where two different `WebSearchAction`
types were getting merged under the same name in the app-server v2 JSON
schema bundle.
The problem was that v2 thread items use the app-server API's
`WebSearchAction` with camelCase variants like `openPage`, while
`ThreadResumeParams.history` and
`RawResponseItemCompletedNotification.item` pull in the upstream
`ResponseItem` graph, which uses the Responses API snake_case shape like
`open_page`. During bundle generation we were flattening nested
definitions into the v2 namespace by plain name, so the later definition
could silently overwrite the earlier one.
That meant clients generating code from the bundled schema could end up
with the wrong `WebSearchAction` definition for v2 thread history. In
practice this shows up on web search items reconstructed from rollout
files with persisted extended history.
This change does two things:
- Gives the upstream Responses API schema a distinct JSON schema name:
`ResponsesApiWebSearchAction`
- Makes namespace-level schema definition collisions fail loudly instead
of silently overwriting
## Summary
This is a purely mechanical refactor of `OtelManager` ->
`SessionTelemetry` to better convey what the struct is doing. No
behavior change.
## Why
`OtelManager` ended up sounding much broader than what this type
actually does. It doesn't manage OTEL globally; it's the session-scoped
telemetry surface for emitting log/trace events and recording metrics
with consistent session metadata (`app_version`, `model`, `slug`,
`originator`, etc.).
`SessionTelemetry` is a more accurate name, and updating the call sites
makes that boundary a lot easier to follow.
## Validation
- `just fmt`
- `cargo test -p codex-otel`
- `cargo test -p codex-core`
## Summary
This is a structural cleanup of `codex-otel` to make the ownership
boundaries a lot clearer.
For example, previously it was quite confusing that `OtelManager` which
emits log + trace event telemetry lived under
`codex-rs/otel/src/traces/`. Also, there were two places that defined
methods on OtelManager via `impl OtelManager` (`lib.rs` and
`otel_manager.rs`).
What changed:
- move the `OtelProvider` implementation into `src/provider.rs`
- move `OtelManager` and session-scoped event emission into
`src/events/otel_manager.rs`
- collapse the shared log/trace event helpers into
`src/events/shared.rs`
- pull target classification into `src/targets.rs`
- move `traceparent_context_from_env()` into `src/trace_context.rs`
- keep `src/otel_provider.rs` as a compatibility shim for existing
imports
- update the `codex-otel` README to reflect the new layout
## Why
`lib.rs` and `otel_provider.rs` were doing too many different jobs at
once: provider setup, export routing, trace-context helpers, and session
event emission all lived together.
This refactor separates those concerns without trying to change the
behavior of the crate. The goal is to make future OTEL work easier to
reason about and easier to review.
## Notes
- no intended behavior change
- `OtelManager` remains the session-scoped event emitter in this PR
- the `otel_provider` shim keeps downstream churn low while the
internals move around
## Validation
- `just fmt`
- `cargo test -p codex-otel`
- `just fix -p codex-otel`
### Summary
This adds turn-level latency metrics for the first model output and the
first completed agent message.
- `codex.turn.ttft.duration_ms` starts at turn start and records on the
first output signal we see from the model. That includes normal
assistant text, reasoning deltas, and non-text outputs like tool-call
items.
- `codex.turn.ttfm.duration_ms` also starts at turn start, but it
records when the first agent message finishes streaming rather than when
its first delta arrives.
### Implementation notes
The timing is tracked in codex-core, not app-server, so the definition
stays consistent across CLI, TUI, and app-server clients.
I reused the existing turn lifecycle boundary that already drives
`codex.turn.e2e_duration_ms`, stored the turn start timestamp in turn
state, and record each metric once per turn.
I also wired the new metric names into the OTEL runtime metrics summary
so they show up in the same in-memory/debug snapshot path as the
existing timing metrics.
This fixes a flaky `turn_start_shell_zsh_fork_executes_command_v2` test.
The interrupt path can race with the follow-up `/responses` request that
reports the aborted tool call, so the test now allows that extra no-op
response instead of assuming there will only ever be one request. The
assertions still stay focused on the behavior the test actually cares
about: starting the zsh-forked command correctly.
Testing:
- `just fmt`
- `cargo test -p codex-app-server --test all
suite::v2::turn_start_zsh_fork::turn_start_shell_zsh_fork_executes_command_v2
-- --exact --nocapture`
### Motivation
Today config.toml has three different OTEL knobs under `[otel]`:
- `exporter` controls where OTEL logs go
- `trace_exporter` controls where OTEL traces go
- `metrics_exporter` controls where metrics go
Those often (pretty much always?) serve different purposes.
For example, for OpenAI internal usage, the **log exporter** is already
being used for IT/security telemetry, and that use case is intentionally
content-rich: tool calls, arguments, outputs, MCP payloads, and in some
cases user content are all useful there. `log_user_prompt` is a good
example of that distinction. When it’s enabled, we include raw prompt
text in OTEL logs, which is acceptable for the security use case.
The **trace exporter** is a different story. The goal there is to give
OpenAI engineers visibility into latency and request behavior when they
run Codex locally, without sending sensitive prompt or tool data as
trace event data. In other words, traces should help answer “what was
slow?” or “where did time go?”, not “what did the user say?” or “what
did the tool return?”
The complication is that Rust’s `tracing` crate does not make a hard
distinction between “logs” and “trace events.” It gives us one
instrumentation API for logs and trace events (via `tracing::event!`),
and subscribers decide what gets treated as logs, trace events, or both.
Before this change, our OTEL trace layer was effectively attached to the
general tracing stream, which meant turning on `trace_exporter` could
pick up content-rich events that were originally written with logging
(and the `log_exporter`) in mind. That made it too easy for sensitive
data to end up in exported traces by accident.
### Concrete example
In `otel_manager.rs`, this `tracing::event!` call would be exported in
both logs AND traces (as a trace event).
```
pub fn user_prompt(&self, items: &[UserInput]) {
let prompt = items
.iter()
.flat_map(|item| match item {
UserInput::Text { text, .. } => Some(text.as_str()),
_ => None,
})
.collect::<String>();
let prompt_to_log = if self.metadata.log_user_prompts {
prompt.as_str()
} else {
"[REDACTED]"
};
tracing::event!(
tracing::Level::INFO,
event.name = "codex.user_prompt",
event.timestamp = %timestamp(),
// ...
prompt = %prompt_to_log,
);
}
```
Instead of `tracing::event!`, we should now be using `log_event!` and
`trace_event!` instead to more clearly indicate which sink (logs vs.
traces) that event should be exported to.
### What changed
This PR makes the log and trace export distinct instead of treating them
as two sinks for the same data.
On the provider side, OTEL logs and traces now have separate
routing/filtering policy. The log exporter keeps receiving the existing
`codex_otel` events, while trace export is limited to spans and trace
events.
On the event side, `OtelManager` now emits two flavors of telemetry
where needed:
- a log-only event with the current rich payloads
- a tracing-safe event with summaries only
It also has a convenience `log_and_trace_event!` macro for emitting to
both logs and traces when it's safe to do so, as well as log- and
trace-specific fields.
That means prompts, tool args, tool output, account email, MCP metadata,
and similar content stay in the log lane, while traces get the pieces
that are actually useful for performance work: durations, counts, sizes,
status, token counts, tool origin, and normalized error classes.
This preserves current IT/security logging behavior while making it safe
to turn on trace export for employees.
### Full list of things removed from trace export
- raw user prompt text from `codex.user_prompt`
- raw tool arguments and output from `codex.tool_result`
- MCP server metadata from `codex.tool_result` (mcp_server,
mcp_server_origin)
- account identity fields like `user.email` and `user.account_id` from
trace-safe OTEL events
- `host.name` from trace resources
- generic `codex.tool_decision` events from traces
- generic `codex.sse_event` events from traces
- the full ToolCall debug payload from the `handle_tool_call` span
What traces now keep instead is mostly:
- spans
- trace-safe OTEL events
- counts, lengths, durations, status, token counts, and tool origin
summaries
This PR adds a durable trace linkage for each turn by storing the active
trace ID on the rollout TurnContext record stored in session rollout
files.
Before this change, we propagated trace context at runtime but didn’t
persist a stable per-turn trace key in rollout history. That made
after-the-fact debugging harder (for example, mapping a historical turn
to the corresponding trace in datadog). This sets us up for much easier
debugging in the future.
### What changed
- Added an optional `trace_id` to TurnContextItem (rollout schema).
- Added a small OTEL helper to read the current span trace ID.
- Captured `trace_id` when creating `TurnContext` and included it in
`to_turn_context_item()`.
- Updated tests and fixtures that construct TurnContextItem so
older/no-trace cases still work.
### Why this approach
TurnContext is already the canonical durable per-turn metadata in
rollout. This keeps ownership clean: trace linkage lives with other
persisted turn metadata.
This adds a first-class server request for MCP server elicitations:
`mcpServer/elicitation/request`.
Until now, MCP elicitation requests only showed up as a raw
`codex/event/elicitation_request` event from core. That made it hard for
v2 clients to handle elicitations using the same request/response flow
as other server-driven interactions (like shell and `apply_patch`
tools).
This also updates the underlying MCP elicitation request handling in
core to pass through the full MCP request (including URL and form data)
so we can expose it properly in app-server.
### Why not `item/mcpToolCall/elicitationRequest`?
This is because MCP elicitations are related to MCP servers first, and
only optionally to a specific MCP tool call.
In the MCP protocol, elicitation is a server-to-client capability: the
server sends `elicitation/create`, and the client replies with an
elicitation result. RMCP models it that way as well.
In practice an elicitation is often triggered by an MCP tool call, but
not always.
### What changed
- add `mcpServer/elicitation/request` to the v2 app-server API
- translate core `codex/event/elicitation_request` events into the new
v2 server request
- map client responses back into `Op::ResolveElicitation` so the MCP
server can continue
- update app-server docs and generated protocol schema
- add an end-to-end app-server test that covers the full round trip
through a real RMCP elicitation flow
- The new test exercises a realistic case where an MCP tool call
triggers an elicitation, the app-server emits
mcpServer/elicitation/request, the client accepts it, and the tool call
resumes and completes successfully.
### app-server API flow
- Client starts a thread with `thread/start`.
- Client starts a turn with `turn/start`.
- App-server sends `item/started` for the `mcpToolCall`.
- While that tool call is in progress, app-server sends
`mcpServer/elicitation/request`.
- Client responds to that request with `{ action: "accept" | "decline" |
"cancel" }`.
- App-server sends `serverRequest/resolved`.
- App-server sends `item/completed` for the mcpToolCall.
- App-server sends `turn/completed`.
- If the turn is interrupted while the elicitation is pending,
app-server still sends `serverRequest/resolved` before the turn
finishes.
### Overview
This PR:
- Updates `app-server-test-client` to load OTEL settings from
`$CODEX_HOME/config.toml` and initializes its own OTEL provider.
- Add real client root spans to app-server test client traces.
This updates `codex-app-server-test-client` so its Datadog traces
reflect the full client-driven flow instead of a set of server spans
stitched together under a synthetic parent.
Before this change, the test client generated a fake `traceparent` once
and reused it for every JSON-RPC request. That kept the requests in one
trace, but there was no real client span at the top, so Datadog ended up
showing the sequence in a slightly misleading way, where all RPCs were
anchored under `initialize`.
Now the test client:
- loads OTEL settings from the normal Codex config path, including
`$CODEX_HOME/config.toml` and existing --config overrides
- initializes tracing the same way other Codex binaries do when trace
export is enabled
- creates a real client root span for each scripted command
- creates per-request client spans for JSON-RPC methods like
`initialize`, `thread/start`, and `turn/start`
- injects W3C trace context from the current client span into
request.trace instead of reusing a fabricated carrier
This gives us a cleaner trace shape in Datadog:
- one trace URL for the whole scripted flow
- a visible client root span
- proper client/server parent-child relationships for each app-server
request
This is PR 3 of the app-server tracing rollout.
PRs https://github.com/openai/codex/pull/13285 and
https://github.com/openai/codex/pull/13368 gave us inbound request spans
in app-server and propagated trace context through Submission. This
change finishes the next piece in core: when a request actually starts a
turn, we now create a core-owned long-lived span that stays open for the
real lifetime of the turn.
What changed:
- `Session::spawn_task` can now optionally create a long-lived turn span
and run the spawned task inside it
- `turn/start` uses that path, so normal turn execution stays under a
single core-owned span after the async handoff
- `review/start` uses the same pattern
- added a unit test that verifies the spawned turn task inherits the
submission dispatch trace ancestry
**Why**
The app-server request span is intentionally short-lived. Once work
crosses into core, we still want one span that covers the actual
execution window until completion or interruption. This keeps that
ownership where it belongs: in the layer that owns the runtime
lifecycle.
### Summary
Propagate trace context originating at app-server RPC method handlers ->
codex core submission loop (so this includes spans such as `run_turn`!).
This implements PR 2 of the app-server tracing rollout.
This also removes the old lower-level env-based reparenting in core so
explicit request/submission ancestry wins instead of being overridden by
ambient `TRACEPARENT` state.
### What changed
- Added `trace: Option<W3cTraceContext>` to codex_protocol::Submission
- Taught `Codex::submit()` / `submit_with_id()` to automatically capture
the current span context when constructing or forwarding a submission
- Wrapped the core submission loop in a submission_dispatch span
parented from Submission.trace
- Warn on invalid submission trace carriers and ignore them cleanly
- Removed the old env-based downstream reparenting path in core task
execution
- Stopped OTEL provider init from implicitly attaching env trace context
process-wide
- Updated mcp-server Submission call sites for the new field
Added focused unit tests for:
- capturing trace context into Submission
- preferring `Submission.trace` when building the core dispatch span
### Why
PR 1 gave us consistent inbound request spans in app-server, but that
only covered the transport boundary. For long-running work like turns
and reviews, the important missing piece was preserving ancestry after
the request handler returns and core continues work on a different async
path.
This change makes that handoff explicit and keeps the parentage rules
simple:
- app-server request span sets the current context
- `Submission.trace` snapshots that context
- core restores it once, at the submission boundary
- deeper core spans inherit naturally
That also lets us stop relying on env-based reparenting for this path,
which was too ambient and could override explicit ancestry.
This adds a first-class app-server v2 `skills/changed` notification for
the existing skills live-reload signal.
Before this change, clients only had the legacy raw
`codex/event/skills_update_available` event. With this PR, v2 clients
can listen for a typed JSON-RPC notification instead of depending on the
legacy `codex/event/*` stream, which we want to remove soon.
Realized EventMsg generated types were unintentionally removed as part
of this PR: https://github.com/openai/codex/pull/13375
Turns out our TypeScript export pipeline relied on transitively reaching
`EventMsg`. We should still export `EventMsg` explicitly since we're
still emitting `codex/event/*` events (for now, but getting dropped soon
as well).
## Summary
This removes the old app-server v1 methods and notifications we no
longer need, while keeping the small set the main codex app client still
depends on for now.
The remaining legacy surface is:
- `initialize`
- `getConversationSummary`
- `getAuthStatus`
- `gitDiffToRemote`
- `fuzzyFileSearch`
- `fuzzyFileSearch/sessionStart`
- `fuzzyFileSearch/sessionUpdate`
- `fuzzyFileSearch/sessionStop`
And the raw `codex/event/*` notifications emitted from core. These
notifications will be removed in a followup PR.
## What changed
- removed deprecated v1 request variants from the protocol and
app-server dispatcher
- removed deprecated typed notifications: `authStatusChange`,
`loginChatGptComplete`, and `sessionConfigured`
- updated the app-server test client to use v2 flows instead of deleted
v1 flows
- deleted legacy-only app-server test suites and added focused coverage
for `getConversationSummary`
- regenerated app-server schema fixtures and updated the MCP interface
docs to match the remaining compatibility surface
## Testing
- `just write-app-server-schema`
- `cargo test -p codex-app-server-protocol`
- `cargo test -p codex-app-server`
This is a follow-up for https://github.com/openai/codex/pull/13047
## Why
We had a race where `turn/started` could be observed before the thread
had actually transitioned to `Active`. This was because we eagerly
emitted `turn/started` in the request handler for `turn/start` (and
`review/start`).
That was showing up as flaky `thread/resume` tests, but the real issue
was broader: a client could see `turn/started` and still get back an
idle thread immediately afterward.
The first idea was to eagerly call
`thread_watch_manager.note_turn_started(...)` from the `turn/start`
request path. That turns out to be unsafe, because
`submit(Op::UserInput)` only queues work. If a turn starts and completes
quickly, request-path bookkeeping can race with the real lifecycle
events and leave stale running state behind.
**The real fix** is to move `turn/started` to emit only after the turn
_actually_ starts, so we do that by waiting for the
`EventMsg::TurnStarted` notification emitted by codex core. We do this
for both `turn/start` and `review/start`.
I also verified this change is safe for our first-party codex apps -
they don't have any assumptions that `turn/started` is emitted before
the RPC response to `turn/start` (which is correct anyway).
I also removed `single_client_mode` since it isn't really necessary now.
## Testing
- `cargo test -p codex-app-server thread_resume -- --nocapture`
- `cargo test -p codex-app-server
'suite::v2::turn_start::turn_start_emits_notifications_and_accepts_model_override'
-- --exact --nocapture`
- `cargo test -p codex-app-server`
### Overview
This PR adds the first piece of tracing for app-server JSON-RPC
requests.
There are two main changes:
- JSON-RPC requests can now take an optional W3C trace context at the
top level via a `trace` field (`traceparent` / `tracestate`).
- app-server now creates a dedicated request span for every inbound
JSON-RPC request in `MessageProcessor`, and uses the request-level trace
context as the parent when present.
For compatibility with existing flows, app-server still falls back to
the TRACEPARENT env var when there is no request-level traceparent.
This PR is intentionally scoped to the app-server boundary. In a
followup, we'll actually propagate trace context through the async
handoff into core execution spans like run_turn, which will make
app-server traces much more useful.
### Spans
A few details on the app-server span shape:
- each inbound request gets its own server span
- span/resource names are based on the JSON-RPC method (`initialize`,
`thread/start`, `turn/start`, etc.)
- spans record transport (stdio vs websocket), request id, connection
id, and client name/version when available
- `initialize` stores client metadata in session state so later requests
on the same connection can reuse it
Stop `thread/start` from blocking other app-server requests.
Before this change, `thread/start ran` inline on the request loop, so
slow startup paths like MCP auth checks could hold up unrelated requests
on the same connection, including `thread/loaded/list`. This moves
`thread/start` into a background task.
While doing so, it revealed an issue where we were doing nested locking
(and there were some race conditions possible that could introduce a
"phantom listener"). This PR also refactors the listener/subscription
bookkeeping - listener/subscription state is now centralized in
`ThreadStateManager` instead of being split across multiple lock
domains. That makes late auto-attach on `thread/start` race-safe and
avoids reintroducing disconnected clients as phantom subscribers.
Adds a new v2 app-server API for a client to be able to unsubscribe to a
thread:
- New RPC method: `thread/unsubscribe`
- New server notification: `thread/closed`
Today clients can start/resume/archive threads, but there wasn’t a way
to explicitly unload a live thread from memory without archiving it.
With `thread/unsubscribe`, a client can indicate it is no longer
actively working with a live Thread. If this is the only client
subscribed to that given thread, the thread will be automatically closed
by app-server, at which point the server will send `thread/closed` and
`thread/status/changed` with `status: notLoaded` notifications.
This gives clients a way to prevent long-running app-server processes
from accumulating too many thread (and related) objects in memory.
Closed threads will also be removed from `thread/loaded/list`.
Previously, clients would call `thread/start` with dynamic_tools set,
and when a model invokes a dynamic tool, it would just make the
server->client `item/tool/call` request and wait for the client's
response to complete the tool call. This works, but it doesn't have an
`item/started` or `item/completed` event.
Now we are doing this:
- [new] emit `item/started` with `DynamicToolCall` populated with the
call arguments
- send an `item/tool/call` server request
- [new] once the client responds, emit `item/completed` with
`DynamicToolCall` populated with the response.
Also, with `persistExtendedHistory: true`, dynamic tool calls are now
reconstructable in `thread/read` and `thread/resume` as
`ThreadItem::DynamicToolCall`.
zsh fork PR stack:
- https://github.com/openai/codex/pull/12051
- https://github.com/openai/codex/pull/12052👈
### Summary
This PR introduces a feature-gated native shell runtime path that routes
shell execution through a patched zsh exec bridge, removing MCP-specific
behavior from the shell hot path while preserving existing
CommandExecution lifecycle semantics.
When shell_zsh_fork is enabled, shell commands run via patched zsh with
per-`execve` interception through EXEC_WRAPPER. Core receives wrapper
IPC requests over a Unix socket, applies existing approval policy, and
returns allow/deny before the subcommand executes.
### What’s included
**1) New zsh exec bridge runtime in core**
- Wrapper-mode entrypoint (maybe_run_zsh_exec_wrapper_mode) for
EXEC_WRAPPER invocations.
- Per-execution Unix-socket IPC handling for wrapper requests/responses.
- Approval callback integration using existing core approval
orchestration.
- Streaming stdout/stderr deltas to existing command output event
pipeline.
- Error handling for malformed IPC, denial/abort, and execution
failures.
**2) Session lifecycle integration**
SessionServices now owns a `ZshExecBridge`.
Session startup initializes bridge state; shutdown tears it down
cleanly.
**3) Shell runtime routing (feature-gated)**
When `shell_zsh_fork` is enabled:
- Build execution env/spec as usual.
- Add wrapper socket env wiring.
- Execute via `zsh_exec_bridge.execute_shell_request(...)` instead of
the regular shell path.
- Non-zsh-fork behavior remains unchanged.
**4) Config + feature wiring**
- Added `Feature::ShellZshFork` (under development).
- Added config support for `zsh_path` (optional absolute path to patched
zsh):
- `Config`, `ConfigToml`, `ConfigProfile`, overrides, and schema.
- Session startup validates that `zsh_path` exists/usable when zsh-fork
is enabled.
- Added startup test for missing `zsh_path` failure mode.
**5) Seatbelt/sandbox updates for wrapper IPC**
- Extended seatbelt policy generation to optionally allow outbound
connection to explicitly permitted Unix sockets.
- Wired sandboxing path to pass wrapper socket path through to seatbelt
policy generation.
- Added/updated seatbelt tests for explicit socket allow rule and
argument emission.
**6) Runtime entrypoint hooks**
- This allows the same binary to act as the zsh wrapper subprocess when
invoked via `EXEC_WRAPPER`.
**7) Tool selection behavior**
- ToolsConfig now prefers ShellCommand type when shell_zsh_fork is
enabled.
- Added test coverage for precedence with unified-exec enabled.
zsh fork PR stack:
- https://github.com/openai/codex/pull/12051👈
- https://github.com/openai/codex/pull/12052
With upcoming support for a fork of zsh that allows us to intercept
`execve` and run execpolicy checks for each subcommand as part of a
`CommandExecution`, it will be possible for there to be multiple
approval requests for a shell command like `/path/to/zsh -lc 'git status
&& rg \"TODO\" src && make test'`.
To support that, this PR introduces a new `approval_id` field across
core, protocol, and app-server so that we can associate approvals
properly for subcommands.
Propagate client JSON-RPC errors for app-server request callbacks.
Previously a number of possible errors were collapsed to `channel
closed`. Now we should be able to see the underlying client error.
### Summary
This change stops masking client JSON-RPC error responses as generic
callback cancellation in app-server server->client request flows.
Previously, when the client responded with a JSON-RPC error, we removed
the callback entry but did not send anything to the waiting oneshot
receiver. Waiters then observed channel closure (for example, auth
refresh request canceled: channel closed), which hid the actual client
error.
Now, client JSON-RPC errors are forwarded through the callback channel
and handled explicitly by request consumers.
### User-visible behavior
- External auth refresh now surfaces real client JSON-RPC errors when
provided.
- True transport/callback-drop cases still report
canceled/channel-closed semantics.
### Example: client JSON-RPC error is now propagated (not masked as
"canceled")
When app-server asks the client to refresh ChatGPT auth tokens, it sends
a server->client JSON-RPC request like:
```json
{
"id": 42,
"method": "account/chatgptAuthTokens/refresh",
"params": {
"reason": "unauthorized",
"previousAccountId": "org-abc"
}
}
```
If the client cannot refresh and responds with a JSON-RPC error:
```
{
"id": 42,
"error": {
"code": -32000,
"message": "refresh failed",
"data": null
}
}
```
app-server now forwards that error through the callback path and
surfaces:
`auth refresh request failed: code=-32000 message=refresh failed`
Previously, this same case could be reported as:
`auth refresh request canceled: channel closed`
This PR adds an experimental `persist_extended_history` bool flag to
app-server thread APIs so rollout logs can retain a richer set of
EventMsgs for non-lossy Thread > Turn > ThreadItems reconstruction (i.e.
on `thread/resume`).
### Motivation
Today, our rollout recorder only persists a small subset (e.g. user
message, reasoning, assistant message) of `EventMsg` types, dropping a
good number (like command exec, file change, etc.) that are important
for reconstructing full item history for `thread/resume`, `thread/read`,
and `thread/fork`.
Some clients want to be able to resume a thread without lossiness. This
lossiness is primarily a UI thing, since what the model sees are
`ResponseItem` and not `EventMsg`.
### Approach
This change introduces an opt-in `persist_full_history` flag to preserve
those events when you start/resume/fork a thread (defaults to `false`).
This is done by adding an `EventPersistenceMode` to the rollout
recorder:
- `Limited` (existing behavior, default)
- `Extended` (new opt-in behavior)
In `Extended` mode, persist additional `EventMsg` variants needed for
non-lossy app-server `ThreadItem` reconstruction. We now store the
following ThreadItems that we didn't before:
- web search
- command execution
- patch/file changes
- MCP tool calls
- image view calls
- collab tool outcomes
- context compaction
- review mode enter/exit
For **command executions** in particular, we truncate the output using
the existing `truncate_text` from core to store an upper bound of 10,000
bytes, which is also the default value for truncating tool outputs shown
to the model. This keeps the size of the rollout file and command
execution items returned over the wire reasonable.
And we also persist `EventMsg::Error` which we can now map back to the
Turn's status and populates the Turn's error metadata.
#### Updates to EventMsgs
To truly make `thread/resume` non-lossy, we also needed to persist the
`status` on `EventMsg::CommandExecutionEndEvent` and
`EventMsg::PatchApplyEndEvent`. Previously it was not obvious whether a
command failed or was declined (similar for apply_patch). These
EventMsgs were never persisted before so I made it a required field.
…ount_id and chatgpt_plan_type
### Summary
Following up on external auth mode which was introduced here:
https://github.com/openai/codex/pull/10012
Turns out some clients have a differently shaped ID token and don't have
a chosen workspace (aka chatgpt_account_id) encoded in their ID token.
So, let's replace `id_token` param with `chatgpt_account_id` and
`chatgpt_plan_type` (optional) when initializing the external ChatGPT
auth mode (`account/login/start` with `chatgptAuthTokens`).
The client was able to test end-to-end with a Codex build from this
branch and verified it worked!
These fields had always been documented as experimental/unstable with
docstrings, but now let's actually use the `experimental` annotation to
be more explicit.
- thread/start.experimentalRawEvents
- thread/resume.history
- thread/resume.path
- thread/fork.path
- turn/start.collaborationMode
- account/login/start.chatgptAuthTokens
This PR adds a dedicated `turn/steer` API for appending user input to an
in-flight turn.
## Motivation
Currently, steering in the app is implemented by just calling
`turn/start` while a turn is running. This has some really weird quirks:
- Client gets back a new `turn.id`, even though streamed
events/approvals remained tied to the original active turn ID.
- All the various turn-level override params on `turn/start` do not
apply to the "steer", and would only apply to the next real turn.
- There can also be a race condition where the client thinks the turn is
active but the server has already completed it, so there might be bugs
if the client has baked in some client-specific behavior thinking it's a
steer when in fact the server kicked off a new turn. This is
particularly possible when running a client against a remote app-server.
Having a dedicated `turn/steer` API eliminates all those quirks.
`turn/steer` behavior:
- Requires an active turn on threadId. Returns a JSON-RPC error if there
is no active turn.
- If expectedTurnId is provided, it must match the active turn (more
useful when connecting to a remote app-server).
- Does not emit `turn/started`.
- Does not accept turn overrides (`cwd`, `model`, `sandbox`, etc.) or
`outputSchema` to accurately reflect that these are not applied when
steering.
So that the rest of the codebase (like TUI) don't need to be concerned
whether ChatGPT auth was handled by Codex itself or passed in via
app-server's external auth mode.
Took over the work that @aaronl-openai started here:
https://github.com/openai/codex/pull/10397
Now that app-server clients are able to set up custom tools (called
`dynamic_tools` in app-server), we should expose a way for clients to
pass in not just text, but also image outputs. This is something the
Responses API already supports for function call outputs, where you can
pass in either a string or an array of content outputs (text, image,
file):
https://platform.openai.com/docs/api-reference/responses/create#responses_create-input-input_item_list-item-function_tool_call_output-output-array-input_image
So let's just plumb it through in Codex (with the caveat that we only
support text and image for now). This is implemented end-to-end across
app-server v2 protocol types and core tool handling.
## Breaking API change
NOTE: This introduces a breaking change with dynamic tools, but I think
it's ok since this concept was only recently introduced
(https://github.com/openai/codex/pull/9539) and it's better to get the
API contract correct. I don't think there are any real consumers of this
yet (not even the Codex App).
Old shape:
`{ "output": "dynamic-ok", "success": true }`
New shape:
```
{
"contentItems": [
{ "type": "inputText", "text": "dynamic-ok" },
{ "type": "inputImage", "imageUrl": "data:image/png;base64,AAA" }
]
"success": true
}
```
One of our partners flagged that they were seeing the wrong order of
events when running `review/start` with command exec approvals:
```
{"method":"item/commandExecution/requestApproval","id":0,"params":{"threadId":"019c0b6b-6a42-7c02-99c4-98c80e88ac27","turnId":"0","itemId":"0","reason":"`/bin/zsh -lc 'git show b7a92b4eacf262c575f26b1e1ed621a357642e55 --stat'` requires approval: Xcode-required approval: Require explicit user confirmation for all commands.","proposedExecpolicyAmendment":null}}
{"method":"item/started","params":{"item":{"type":"commandExecution","id":"call_AEjlbHqLYNM7kbU3N6uw1CNi","command":"/bin/zsh -lc 'git show b7a92b4eacf262c575f26b1e1ed621a357642e55 --stat'","cwd":"/Users/devingreen/Desktop/SampleProject","processId":null,"status":"inProgress","commandActions":[{"type":"unknown","command":"git show b7a92b4eacf262c575f26b1e1ed621a357642e55 --stat"}],"aggregatedOutput":null,"exitCode":null,"durationMs":null},"threadId":"019c0b6b-6a42-7c02-99c4-98c80e88ac27","turnId":"0"}}
```
**Key fix**: In the review sub‑agent delegate we were forwarding exec
(and patch) approvals using the parent turn id (`parent_ctx.sub_id`) as
the approval call_id. That made
`item/commandExecution/requestApproval.itemId` differ from the actual
`item/started` id. We now forward the sub‑agent’s `call_id` from the
approval event instead, so the approval item id matches the
commandExecution item id in review flows.
Here’s the expected event order for an inline `review/start` that
triggers an exec approval after this fix:
1. Response to review/start (JSON‑RPC response)
- Includes `turn` (status inProgress) and `review_thread_id` (same as
parent thread for inline).
2. `turn/started` notification
- turnId is the review turn id (e.g., "0").
3. `item/started` → EnteredReviewMode
- item.id == turnId, marks entry into review mode.
4. `item/started` → commandExecution
- item.id == <call_id> (e.g., "review-call-1"), status: inProgress.
5. `item/commandExecution/requestApproval` request
- JSON‑RPC request (not a notification).
- params.itemId == <call_id> and params.turnId == turnId.
6. Client replies to approval request (Approved / Declined / etc).
7. If approved:
- Optional `item/commandExecution/outputDelta` notifications.
- `item/completed` → commandExecution with status and exitCode.
8. Review finishes:
- `item/started` → ExitedReviewMode
- `item/completed` → ExitedReviewMode
- (Agent message items may also appear, depending on review output.)
9. `turn/completed` notification
The key being #4 and #5 are now in the proper order with the correct
item id.
This updates our generated TypeScript types to be more correct with how
the server actually behaves, **specifically for JSON-RPC requests**.
Before this PR, we'd generate `field: T | null`. After this PR, we will
have `field?: T | null`. The latter matches how the server actually
works, in that if an optional field is omitted, the server will treat it
as null. This also makes it less annoying in theory for clients to
upgrade to newer versions of Codex, since adding a new optional field to
a JSON-RPC request should not require a client change.
NOTE: This only applies to JSON-RPC requests. All other payloads (i.e.
responses, notifications) will return `field: T | null` as usual.
This enables a new use case where `codex app-server` is embedded into a
parent application that will directly own the user's ChatGPT auth
lifecycle, which means it owns the user’s auth tokens and refreshes it
when necessary. The parent application would just want a way to pass in
the auth tokens for codex to use directly.
The idea is that we are introducing a new "auth mode" currently only
exposed via app server: **`chatgptAuthTokens`** which consist of the
`id_token` (stores account metadata) and `access_token` (the bearer
token used directly for backend API calls). These auth tokens are only
stored in-memory. This new mode is in addition to the existing `apiKey`
and `chatgpt` auth modes.
This PR reuses the shape of our existing app-server account APIs as much
as possible:
- Update `account/login/start` with a new `chatgptAuthTokens` variant,
which will allow the client to pass in the tokens and have codex
app-server use them directly. Upon success, the server emits
`account/login/completed` and `account/updated` notifications.
- A new server->client request called
`account/chatgptAuthTokens/refresh` which the server can use whenever
the access token previously passed in has expired and it needs a new one
from the parent application.
I leveraged the core 401 retry loop which typically triggers auth token
refreshes automatically, but made it pluggable:
- **chatgpt** mode refreshes internally, as usual.
- **chatgptAuthTokens** mode calls the client via
`account/chatgptAuthTokens/refresh`, the client responds with updated
tokens, codex updates its in-memory auth, then retries. This RPC has a
10s timeout and handles JSON-RPC errors from the client.
Also some additional things:
- chatgpt logins are blocked while external auth is active (have to log
out first. typically clients will pick one OR the other, not support
both)
- `account/logout` clears external auth in memory
- Ensures that if `forced_chatgpt_workspace_id` is set via the user's
config, we respect it in both:
- `account/login/start` with `chatgptAuthTokens` (returns a JSON-RPC
error back to the client)
- `account/chatgptAuthTokens/refresh` (fails the turn, and on next
request app-server will send another `account/chatgptAuthTokens/refresh`
request to the client).
### Overview
Currently calling `thread/resume` will always bump the thread's
`updated_at` timestamp. This PR makes it the `updated_at` timestamp
changes only if a turn is triggered.
### Additonal context
What we typically do on resuming a thread is **always** writing “initial
context” to the rollout file immediately. This initial context includes:
- Developer instructions derived from sandbox/approval policy + cwd
- Optional developer instructions (if provided)
- Optional collaboration-mode instructions
- Optional user instructions (if provided)
- Environment context (cwd, shell, etc.)
This PR defers writing the “initial context” to the rollout file until
the first `turn/start`, so we don't inadvertently bump the thread's
`updated_at` timestamp until a turn is actually triggered.
This works even though both `thread/resume` and `turn/start` accept
overrides (such as `model`, `cwd`, etc.) because the initial context is
seeded from the effective `TurnContext` in memory, computed at
`turn/start` time, after both sets of overrides have been applied.
**NOTE**: This is a very short-lived solution until we introduce sqlite.
Then we can remove this.
Remove `FileTimes::new().set_modified(SystemTime::now())` when resuming
a thread.
Context: It's awkward in UI built on top of app-server that resuming a
thread bumps the `updated_at` timestamp, even if no message is sent. So
if you open a thread (perhaps to just view its contents), it
automatically reorders it to the top which is almost certainly not what
you want.
Add support for returning threads by either `created_at` OR `updated_at`
descending. Previously core always returned threads ordered by
`created_at`.
This PR:
- updates core to be able to list threads by `updated_at` OR
`created_at` descending based on what the caller wants
- also update `thread/list` in app-server to expose this (default to
`created_at` if not specified)
All existing codepaths (app-server, TUI) still default to `created_at`,
so no behavior change is expected with this PR.
**Implementation**
To sort by `updated_at` is a bit nontrivial (whereas `created_at` is
easy due to the way we structure the folders and filenames on disk,
which are all based on `created_at`).
The most naive way to do this without introducing a cache file or sqlite
DB (which we have to implement/maintain) is to scan files in reverse
`created_at` order on disk, and look at the file's mtime (last modified
timestamp according to the filesystem) until we reach `MAX_SCAN_FILES`
(currently set to 10,000). Then, we can return the most recent N
threads.
Based on some quick and dirty benchmarking on my machine with ~1000
rollout files, calling `thread/list` with limit 50, the `updated_at`
path is slower as expected due to all the I/O:
- updated-at: average 103.10 ms
- created-at: average 41.10 ms
Those absolute numbers aren't a big deal IMO, but we can certainly
optimize this in a followup if needed by introducing more state stored
on disk.
**Caveat**
There's also a limitation in that any files older than `MAX_SCAN_FILES`
will be excluded, which means if a user continues a REALLY old thread,
it's possible to not be included. In practice that should not be too big
of an issue.
If a user makes...
- 1000 rollouts/day → threads older than 10 days won't show up
- 100 rollouts/day → ~100 days
If this becomes a problem for some reason, even more motivation to
implement an updated_at cache.
Owen Lin