## Why
Dispatched invocations, model-visible Responses calls, and Code Mode's
per-cell runtime calls use IDs with different roles. A dedicated Rust type
makes the dispatch identity explicit in APIs that carry it.
## What changed
Add `codex-tool-call-id` and use its `ToolCallId` across core tool dispatch,
nested Code Mode calls, and rollout tracing. The type wraps a string and
continues to serialize as a string.
GitOrigin-RevId: 119de57a9af840c4de72895adc9acdd0ed15b4fd
## Why
Accepted tool outputs may never appear in a later inference request, and request preparation can strip their history IDs. Replay needs to retain these outputs and match them to later snapshots, including trimmed results and outputs with identical wire bodies.
## What changed
- Capture accepted function and custom tool outputs, preserving their acceptance order, identity, and tool provenance during replay.
- Record stripped tool output IDs alongside inference requests so replay can restore identity without changing the captured wire payload. Reuse accepted items and adopt their first model-visible snapshot bodies.
- Record compaction checkpoints for local context-window resets using the selected input and frozen installed history, including goal edits retained during installation.
- Exclude `configuration_update` items from normalized conversation history.
## Testing
Add replay coverage for outputs accepted without subsequent inference, trimmed results, identical outputs with distinct IDs, and acceptance before or after a request snapshot. Extend checkpoint coverage to verify that retained goal edits are captured and later edits do not alter the checkpoint.
GitOrigin-RevId: 70f737782f5000812f389a4e01b9f9690e39549f
## Why
Interrupting a turn while a tool awaits approval can leave its rollout trace execution marked as running without an end timestamp.
## What changed
Record an `Aborted` end event when a tool dispatch trace is dropped before producing a result. Make completion and failure methods consume the trace handle so each dispatch is closed once.
## Testing
Add coverage for dropping a pending dispatch and for interrupting a turn with a pending command approval, verifying that replay reports an aborted execution with an end timestamp.
GitOrigin-RevId: 63cf68312fcbff8c6b85a01a7713cf4460c9d795
## What changed
- Derive sampling diagnostic and rollout trace labels directly from permission profiles instead of constructing legacy sandbox policies. Rename the trace metadata field from `sandbox_policy` to `permissions_profile`.
- Extract `with_workspace_write_cwd` to share default workspace metadata protections between legacy policy conversion and the TUI's sandbox-mode compatibility check. Build the check's workspace-write profile from filesystem and network permissions directly.
## Testing
Add a regression test confirming that the sandbox-mode compatibility check accepts standard workspace metadata protections and rejects a profile with an additional writable `.git` root.
GitOrigin-RevId: f20d23afd80dff6347865df3d6a94d6f5e90e1e0
## Why
Recording inference starts synchronously serializes prompts and writes trace payloads before model work can proceed. Defer that work so inference dispatch can continue while the trace is recorded.
## What changed
- Use one worker per trace bundle to write events in submission order, with up to eight pending jobs.
- Share prepared prompts and encoded requests with the worker, and build remote compaction snapshots there. Write payloads before publishing events that reference them.
- Add `TraceWriter::flush` and `ThreadTraceContext::flush` for live reads. Drain accepted jobs and seal the bundle on shutdown; suppress the closing marker after a recording failure.
## Testing
Add tests for deferred prompt serialization, ordered start and completion events, shutdown draining, and serialization failures that preserve earlier evidence without publishing a closing marker. Verify compaction snapshots are built on the worker and skipped when tracing is disabled.
GitOrigin-RevId: 32fd787bb1d281f1814df2f08c73b667628f8add
## What changed
- Replace individual JSON payload files with numbered Zstandard shards in trace schema v3. Add `PayloadReader` to resolve payload references, cache decoded shards, and validate payloads during replay.
- Capture exact outgoing HTTP bodies and WebSocket messages alongside prepared prompts. Preserve full formatted context for continuations from unrecorded WebSocket warmups while retaining the actual incremental request.
- Flush payload records before publishing references and seal the final shard before recording completion. Replace `recording_complete` with `recording_status` to distinguish unclosed recordings, closed recordings with partial payloads, and complete recordings.
Reduce trace bundles with the binary that recorded them.
## Testing
Add coverage for shard rotation, oversized records, live reads, cache eviction, and malformed or corrupt payloads. Extend HTTP and WebSocket integration assertions to verify captured requests and prepared prompts, including warmup continuations.
GitOrigin-RevId: 40ce3c11383f48e08f52e7120b706c61a449df9c
## What changed
- Build and assert `PermissionProfile` values directly in memory consolidation tests using filesystem and network permissions.
- Remove the `SandboxPolicy` dependency from rollout trace test fixtures while preserving their `DangerFullAccess` value.
- Update sandbox documentation to describe legacy `sandbox_mode` conversion to `PermissionProfile` and nested filesystem rules, and reference `SandboxMode` in the CLI documentation.
GitOrigin-RevId: ac1be62625b1bfd98dde5f762b379eb840f9186b
## Why
A completion marker missing its final newline can make an incomplete trace appear complete. Opaque JSON payload validation can also miss invalid UTF-8 inside strings.
## What changed
Require a newline-terminated `RecordingCompleted` record before setting `recording_complete`. Read raw payloads as UTF-8 text before validating their JSON structure.
## Testing
Extend the lifecycle replay test to cover a completion marker missing its final newline and an opaque payload containing invalid UTF-8.
GitOrigin-RevId: a0007ff449cdefeb44065ed9165ef60bf720dad9
## Why
Enabled rollout tracing could silently fail to initialize, and a terminal task status did not establish whether all recording had finished. Trace consumers need to distinguish a complete recording from a partial bundle, including when a task is interrupted.
## What changed
- Fail thread startup or resume if an enabled recorder cannot initialize, while cleaning up session resources.
- Stop recording after the first write or serialization failure and preserve that failure across shared producers.
- Append a closing record when the last writer handle is released without recording errors. Expose `recording_complete` in `state.json` independently of task success, and mark submission-channel closure as an aborted rollout.
- Reject replay logs with missing or reordered events, mismatched rollout IDs, events after completion, or missing or invalid JSON payloads. Valid prefixes remain replayable with `recording_complete` set to `false`.
## Testing
Add coverage for startup failure and recovery, shared-writer failures, finalization after child handles close, incomplete logs, and missing or malformed payloads. Extend session and CLI tests to check recording completion after shutdown, interruption, stream failure, and submission-channel closure.
GitOrigin-RevId: 30d62fa95912fb0442ad1199776d110884066c30
## Why
Tool interception and MCP aliases can route a requested callable to a different handler. Rollout traces need to reflect the handler that actually executes, including when an `exec_command` invocation is intercepted as `apply_patch`.
## What changed
Carry an explicit `ToolCallKind` in runtime-start events and use it to update the tool call's execution classification. This replaces MCP-specific payload inspection while preserving the original callable summary and invocation for intercepted patch and aliased MCP calls.
## Testing
Add a regression test verifying that an intercepted `exec_command` is recorded as `ApplyPatch`, retains its original summary and invocation, completes successfully, and creates no terminal operations or sessions.
GitOrigin-RevId: 0e72f3ef55cda2796e779efd16241e4a0dbf0c5b
## Why
Remote compaction can process large histories. Repeatedly estimating and replacing the full history, and cloning it when tracing is disabled, adds avoidable CPU and memory overhead.
## What changed
- Estimate each history item's token count once, update the total as trailing tool outputs are rewritten, and replace history only after all rewrites are selected.
- Preserve unclamped token totals while calculating removed tokens so saturated estimates do not hide overflow.
- Snapshot compaction input history only when rollout tracing is enabled.
- Reuse the v2 request input instead of cloning it before adding the compaction trigger.
## Testing
- Cover enabled and disabled compaction trace contexts through `is_enabled()` assertions.
GitOrigin-RevId: 8de9c9704ba29532fba05430eb204bbae2e8bf83
## Why
Pre-sampling compaction intentionally uses the previous turn's model
when the compaction compatibility hash changes or when switching to a
model with a smaller context window. This keeps compaction aligned with
the settings that produced the history, but it can block the next turn
when a resumed ChatGPT thread still references a model slug that has
since been retired. The Codex backend rejects that compaction request
before the user's currently selected model gets a chance to sample.
This PR lets those threads recover without changing previous-model
compaction behavior for API-key authentication or custom providers. It
is stacked on #31316, which is a behavior-preserving extraction of the
individual remote compaction attempts; this PR contains the fallback
behavior.
## What changed
- For automatic previous-model compaction, capture the selected model's
request context when using ChatGPT authentication with the OpenAI
provider and the selected model differs from the previous model.
- If the previous-model attempt returns an `InvalidRequest`, retry
compaction once with the selected model for both `/responses/compact`
and Responses Compaction V2.
- Complete history processing, lifecycle events, and token accounting
with the context of the model that successfully compacted the thread.
- If the fallback also fails, return the original previous-model error
so the retry does not change the user-visible failure.
- Record fallback attempts with reason, implementation, and outcome
telemetry.
- Leave API-key authentication, custom providers, same-model turns, and
non-`InvalidRequest` failures on their existing paths.
## Testing
- `just test -p codex-core -E 'test(pre_sampling_compact) |
test(model_unavailable_error)'` (10 tests)
- Added integration coverage for a resumed thread whose model was
renamed, a model downshift using Responses Compaction V2, and API-key
authentication with a custom provider.
## Why
Rollout traces need an identifier that can be used to correlate a Codex
inference with upstream Responses API, proxy, and engine logs. The
reduced trace model already exposed `upstream_request_id`, but it was
being populated from the Responses API `response.id`. That value is
useful for `previous_response_id` chaining, but it is not the transport
request id that upstream systems key on.
This PR separates those concepts so trace consumers can reliably answer
both questions:
- which Responses API response did this inference produce?
- which upstream request handled it?
## Structure
The change keeps the upstream request id at the same lifecycle level as
the provider stream:
- `codex-api` captures the `x-request-id` HTTP response header when the
SSE stream is created and exposes it on `ResponseStream`. Fixture and
websocket streams set the field to `None` because they do not have that
HTTP response header.
- `codex-core` carries that stream-level id into `InferenceTraceAttempt`
when recording terminal stream outcomes. Completed, failed, cancelled,
dropped-stream, and pre-response error paths all record the id when it
is available.
- `rollout-trace` now records both identifiers in raw terminal inference
events and response payloads: `response_id` for the Responses API
`response.id`, and `upstream_request_id` for `x-request-id`.
- The reducer stores both fields on `InferenceCall`. It also uses
`response_id` for `previous_response_id` conversation linking, which
removes the old accidental dependency on the misnamed
`upstream_request_id` field.
- Terminal inference reduction now consumes the full terminal payload
(`InferenceCompleted`, `InferenceFailed`, or `InferenceCancelled`) in
one place. That keeps status, partial payloads, response ids, and
upstream request ids consistent across success, failure, cancellation,
and late stream-mapper events.
## Why This Shape
`x-request-id` is a property of the HTTP/provider response envelope, not
an SSE event. Capturing it once in `codex-api` and plumbing it through
terminal trace recording avoids trying to infer the value from stream
contents, and it preserves the id even when the stream fails or is
cancelled after only partial output.
Keeping `response_id` separate from `upstream_request_id` also makes the
reduced trace model less surprising: `response_id` remains the
conversation-continuation id, while `upstream_request_id` is the
operational correlation id for upstream debugging.
## Validation
The PR updates trace and reducer coverage for:
- reading `x-request-id` from SSE response headers;
- storing the true upstream request id on completed inference calls;
- preserving upstream request ids for cancelled and late-cancelled
inference streams;
- keeping `previous_response_id` reconstruction tied to `response_id`
rather than transport request ids.
Records cancelled inference streams when Codex stops consuming a
provider response before `response.completed`, preserving complete
output items observed before cancellation.
Also closes still-running inference calls when the owning turn ends, so
reduced rollout traces do not leave stale `Running` inference nodes.
Covered by focused reducer coverage and a core stream-drop test for
partial output preservation.
## Summary
Adds the debug CLI entry point for reducing recorded rollout traces.
This gives developers a direct way to inspect whether the emitted trace
stream reduces into the expected conversation/runtime model.
## Stack
This is PR 5/5 in the rollout trace stack.
- [#18876](https://github.com/openai/codex/pull/18876): Add rollout
trace crate
- [#18877](https://github.com/openai/codex/pull/18877): Record core
session rollout traces
- [#18878](https://github.com/openai/codex/pull/18878): Trace tool and
code-mode boundaries
- [#18879](https://github.com/openai/codex/pull/18879): Trace sessions
and multi-agent edges
- [#18880](https://github.com/openai/codex/pull/18880): Add debug trace
reduction command
## Review Notes
This PR is intentionally last: it depends on the trace crate, core
recorder, runtime/tool events, and session/agent edge data all existing.
The command should remain a debug/developer tool and avoid adding new
runtime behavior.
The useful review question is whether the CLI exposes the reducer in the
smallest practical way for local inspection without turning the debug
command into a supported user-facing workflow.