## Why
Codex app-server latency traces do not granularly cover turn task
startup and inter-request handoffs. These spans help attribute time
across task execution, startup prewarm, in-flight tool completion, and
rollout persistence.
## What changed
- Add `session_task.run` spans around task execution and
`session_task.flush_rollout` around flushing pending conversation
transcript writes to durable storage
- Add `regular_task.prepare_run_turn` around regular-turn startup (Send
the `TurnStarted` event, reset turn-specific reasoning state, and
resolve any startup prewarm)
- Add `startup_prewarm.resolve` around waiting for background session
prewarming to finish, fail, time out, or be cancelled
- Add a function-level trace span around draining in-flight tool calls
(Wait for tool calls to complete, record tool result in conversation
history, and other bookkeeping)
## Verification
Trigger Codex rollout and observe new spans are included
## 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`
## Why
Compaction window identity is part of session history, not model-client
transport state. Persisting it with the compacted rollout item lets
resumed threads continue from the reconstructed window without keeping
mutable window state on `ModelClient`.
## What changed
- Added `window_id` to `CompactedItem` and stamp it when
`replace_compacted_history` installs compacted history.
- Moved auto-compact window id ownership into `AutoCompactWindow` /
`SessionState`; `ModelClient` now receives the request window id from
callers instead of storing it.
- Returned `window_id` from rollout reconstruction for resume.
Reconstruction uses the newest surviving compacted item's stored
`window_id` when present, and falls back to the legacy compacted-item
count when it is absent.
- Kept fork startup at the fresh default window id and updated direct
model-client tests to pass explicit test window ids.
## Validation
- `cargo check -p codex-core --tests`
## Summary
We originally addressed startup prewarming holding the read side of
`RwLock<McpConnectionManager>` by snapshotting tool-list state. Review
feedback identified the broader ownership problem: the outer
synchronization should only publish or retrieve the current manager,
while MCP operations rely on the manager's internal synchronization. A
follow-up preserved operation retirement with a separate gate, but
further review questioned whether that synchronization was actually
required and whether we could support latest-wins replacement instead.
This PR now stores the current MCP manager in `ArcSwap`. Each operation
uses `load_full()` to obtain an owned `Arc<McpConnectionManager>`, then
performs MCP I/O without retaining the publication mechanism. Refresh
cancels obsolete startup work, constructs a replacement, and atomically
publishes it. New operations see the latest manager, while operations
that already loaded the previous manager retain a valid handle. Refresh
happens at a turn boundary, so there should be no active user tool calls
to drain.
Git history supports dropping the outer `RwLock`. It was introduced in
`03ffe4d595` on November 17, 2025 for non-blocking MCP startup: the
session published an empty manager, startup initialized that same object
while holding the write lock, and readers waited for initialization.
`7cd2e84026` on February 19, 2026 removed that two-phase initialization
in favor of constructing a fresh manager and swapping it in, explicitly
noting that `Option` or `OnceCell` could replace the placeholder design.
Hot reload later reused the existing lock to publish a replacement, but
I found no indication that the lock was introduced to guarantee
in-flight tool calls finish before refresh or shutdown.
Terminal shutdown remains separate from refresh: it aborts startup
prewarming and active tasks before shutting down the current manager, so
tool calls may be interrupted and no model WebSocket work continues
after shutdown. Focused regression coverage exercises pending tool-list
cancellation, deferred refresh, and startup-prewarm shutdown.
## Summary
- accept non-empty model-defined reasoning effort values while
preserving built-in effort behavior
- propagate the non-Copy effort type through core, app-server, TUI,
telemetry, and persistence call sites
- preserve string wire encoding and expose an open-string schema for
clients
- update model selection and shortcut behavior for model-advertised
effort values
## Root cause
`ReasoningEffort` gained a string-backed custom variant, so it could no
longer implement `Copy` or rely on derived closed-enum serialization.
Existing consumers still moved effort values from shared references and
assumed a fixed built-in value set.
## Validation
- `just fmt`
- Local tests and compilation were not run per request; relying on CI.
## Summary
- skip startup websocket prewarm setup when the model client has
Responses-over-WebSocket disabled
- avoid making HTTP-only sessions build prewarm prompt/tool state that
cannot produce a reusable websocket session
## Why
Recent macOS timing flakes were timing out while waiting for first-turn
events in HTTP-only core tests. Startup prewarm is only useful for
websocket-capable providers, but it was scheduled for every session. For
HTTP-only test providers this added unnecessary async startup work
before the regular turn could reach the mocked response flow.
## Testing
- bazel test //codex-rs/core:core-all-test
--test_filter=suite::auto_review::remote_model_override_uses_catalog_model_for_strict_auto_review
--test_output=errors
- bazel test //codex-rs/core:core-all-test
--test_filter=suite::request_permissions_tool::approved_folder_write_request_permissions_unblocks_later_apply_patch
--test_output=errors
Stack split from #25708. Original PR intentionally left open. This third
PR resolves the effective per-thread multi-agent runtime from persisted
metadata, inherited runtime, and current model selection.
## Summary
- Add `request_kind` values for foreground turn, startup prewarm,
compaction, and detached memory model requests.
- Attach compaction dispatch metadata to local Responses, legacy
`/v1/responses/compact`, and remote v2 compact requests.
- Add the existing logical context-window identifier as `window_id` on
turn-owned model request metadata.
- Keep identity fields optional for detached memory requests, while
still emitting `request_kind="memory"` in non-git/no-sandbox workspaces.
## Root Cause
`x-codex-turn-metadata` has more than one producer. Foreground turns and
compaction requests own a real turn and should carry that turn identity.
Detached memory stage-one requests do not own a foreground turn, so
absent identity fields are valid rather than missing data. Startup
websocket prewarm is also a model request, but it has `generate=false`
and must not be counted as a foreground turn.
`thread_source` or session source identifies where a thread came from
(for example review, guardian, or another subagent). `request_kind`
identifies what the current outbound model request is doing (`turn`,
`prewarm`, `compaction`, or `memory`). A review or guardian thread can
issue either a normal turn request or a compaction request, so source
cannot replace request kind.
## Behavior / Impact
- Ordinary foreground requests send `request_kind="turn"`, their real
identity fields, and `window_id="<thread_id>:<window_generation>"`.
- Startup websocket warmup requests send `request_kind="prewarm"` so
they are not counted as foreground turns.
- Compaction requests send `request_kind="compaction"`, their real
owning turn identity, the existing `window_id`, and
`compaction.{trigger,reason,implementation,phase,strategy}`.
- Detached memory stage-one requests send `request_kind="memory"`
without `session_id`, `thread_id`, `turn_id`, or `window_id`; when no
workspace metadata exists, the kind-only header is still emitted.
- `session_id`, `thread_id`, `turn_id`, and `window_id` remain optional
in the header schema because detached memory requests do not own a
foreground turn or context window.
- `window_id` is not a new ID system: it is copied from the already-sent
`x-codex-window-id` / WS client metadata value at model-request dispatch
time.
- Existing `x-codex-window-id` HTTP/WS emission, value format,
generation advancement, resume behavior, and fork reset behavior are
unchanged.
- `request_kind`, `window_id`, and upstream turn-owned identity fields
remain schema-owned; input `responsesapi_client_metadata` cannot replace
their canonical values.
- No table, DAG, export, app-server API, or MCP `_meta` schema changes
are included.
A compaction attempt stopped by a pre-compact hook issues no model
request and therefore has no request header; its outcome remains in
analytics events. Status, error, duration, and token deltas also remain
analytics fields rather than request-header fields.
Future detached-memory attribution using a real initiating turn ID as
`trigger_turn_id` is intentionally not part of this PR.
## Sync With Main
- Final pushed head `716342e79` is rebased onto `origin/main@0d37db4b2`.
- The metadata conflict came from upstream `#24160`, which added
`forked_from_thread_id` on the same `turn_metadata` surface. Resolution
preserves that field and its protection from client metadata override
alongside this PR's request-kind, compaction, and window-id fields.
- While resolving the overlapping commits, I removed an accidental
recursive model-request overlay and a duplicate detached-memory header
builder before completing the rebase.
## Latency / User Experience Boundary
- Foreground turns perform no new filesystem, git, or network work. New
fields are inserted into metadata already serialized for outgoing
requests.
- Compaction issues the same model/HTTP requests with the same prompt,
model, service tier, and sampling settings; only metadata bytes change.
- Startup prewarm already sent metadata; it is now correctly classified
as `prewarm`.
- Non-git detached memory now sends a small kind-only metadata header
rather than no header.
- This client diff adds no user-visible latency mechanism beyond
negligible serialization and header bytes on already-existing requests.
## Validation
On conflict-resolved head `1d35c2cfb` based on `origin/main@487521733`:
- `just fmt` (passed)
- `just fix -p codex-core` (passed)
- `git diff --check origin/main...HEAD` (passed)
- `just test -p codex-core -E 'test(turn_metadata) |
test(websocket_first_turn_uses_startup_prewarm_and_create) |
test(responses_stream_includes_turn_metadata_header_for_git_workspace_e2e)
|
test(responses_websocket_forwards_turn_metadata_on_initial_and_incremental_create)
| test(remote_compact_v2_retries_failures_with_stream_retry_budget) |
test(window_id_advances_after_compact_persists_on_resume_and_resets_on_fork)'`
(`23 passed`; `bench-smoke` passed)
- `just test -p codex-app-server -E
'test(turn_start_forwards_client_metadata_to_responses_request_v2) |
test(turn_start_forwards_client_metadata_to_responses_websocket_request_body_v2)
| test(auto_compaction_remote_emits_started_and_completed_items)'` (`3
passed`; `bench-smoke` passed)
- `just test -p codex-memories-write` (`29 passed`; `bench-smoke`
passed)
## Why
While investigating `codex exec hi` startup latency, the useful
questions were not "is startup slow?" but "which durable bucket is slow
in production?"
The path we observed has a few distinct stages:
1. `thread/start` creates the session
2. startup prewarm builds the turn context, tools, and prompt
3. startup prewarm warms the websocket
4. the first real turn resolves the prewarm
5. the model produces the first token
Before this PR, production telemetry had some of the raw measurements
already:
- aggregate startup-prewarm duration / age-at-first-turn metrics
- TTFT as a metric
- websocket request telemetry
But there was no coherent production event stream for the startup
breakdown itself, and TTFT was metric-only. That made it hard to answer
the same latency questions from OpenTelemetry-backed logs without adding
one-off local instrumentation.
## What changed
Add durable production telemetry on the existing `SessionTelemetry`
path:
- new `codex.startup_phase` OTel log/trace events plus
`codex.startup.phase.duration_ms`
- new `codex.turn_ttft` OTel log/trace events while preserving the
existing TTFT metric
The startup phase event is emitted for the coarse buckets we actually
observed while running `exec hi`:
- `thread_start_create_thread`
- `startup_prewarm_total`
- `startup_prewarm_create_turn_context`
- `startup_prewarm_build_tools`
- `startup_prewarm_build_prompt`
- `startup_prewarm_websocket_warmup`
- `startup_prewarm_resolve`
These phases are intentionally low-cardinality so they remain safe as
production telemetry tags.
## Why this shape
This keeps the instrumentation on the same production path as the rest
of the session telemetry instead of adding a local debug-only trace
mode. It also avoids changing startup behavior:
- prewarm still runs
- no control flow changes
- no extra remote calls
- no user-visible behavior changes
One boundary is intentional: very early process bootstrap that happens
before a session exists is not included here, because this PR uses
session-scoped production telemetry. The expensive buckets we were
trying to understand after `thread/start` are now covered durably.
## Verification
- `cargo test -p codex-otel`
- `cargo test -p codex-core turn_timing`
- `cargo test -p codex-core
regular_turn_emits_turn_started_without_waiting_for_startup_prewarm`
- `cargo test -p codex-core
interrupting_regular_turn_waiting_on_startup_prewarm_emits_turn_aborted`
- `cargo test -p codex-app-server thread_start`
- `just fix -p codex-otel -p codex-core -p codex-app-server`
I also ran `cargo test -p codex-core`; it built successfully and then
hit an existing unrelated stack overflow in
`tools::handlers::multi_agents::tests::tool_handlers_cascade_close_and_resume_and_keep_explicitly_closed_subtrees_closed`.
## Summary
- reduce public module visibility across Rust crates, preferring private
or crate-private modules with explicit crate-root public exports
- update external call sites and tests to use the intended public crate
APIs instead of reaching through module trees
- add the module visibility guideline to AGENTS.md
## Validation
- `cargo check --workspace --all-targets --message-format=short` passed
before the final fix/format pass
- `just fix` completed successfully
- `just fmt` completed successfully
- `git diff --check` passed
Stacked on #16508.
This removes the temporary `codex-core` / `codex-login` re-export shims
from the ownership split and rewrites callsites to import directly from
`codex-model-provider-info`, `codex-models-manager`, `codex-api`,
`codex-protocol`, `codex-feedback`, and `codex-response-debug-context`.
No behavior change intended; this is the mechanical import cleanup layer
split out from the ownership move.
---------
Co-authored-by: Codex <noreply@openai.com>
## 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
```
mchen-oaiandGitHub
Owen LinandGitHub
pakrym-oaiandGitHub
Charlie MarshandGitHub
Ahmed IbrahimandGitHub
jifandGitHub
ningyi-oaiandGitHub
sayan-oaiandGitHub
Matthew ZengandGitHub
Dylan HurdandGitHub