You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
fix(codex): delegate_task children start as Codex subagents - #36525
The rework brief listed nothing to change; it asked me to stay on PR #36525. CI then failed 'Test Server 2' (and the aggregate 'Check') on 6bedaa6. The failure was real: the four OrchestratorMcpToolkit.integration.test.ts tests ('reports distinct/identical/emoji/long child results') replay the recorded delegated_task_status Codex transcript, and its child.thread/start frame had no threadSource. With the fix, T3 sends threadSource "subagent" for delegate_task children, so the replay harness rejected the frame. The job's later 'not enough space on the disk' error was runner noise, not the cause. I reproduced the four failures locally with the same test file, then added "threadSource":"subagent" to that one expected outbound frame in apps/server/src/orchestration-v2/testkit/fixtures/delegated_task_status/codex_transcript.ndjson. No other Codex fixture starts a delegated child. The fix is unchanged. On f8074ae all 8 required checks pass and Macroscope approved with no comments on any line. CodeRabbit's review of 6bedaa6 had no actionable comments. Its re-run on f8074ae was still in progress when I stopped watching, and it had reported 0 included reviews left that hour. No finding rated high or above is open.
On origin/main 56fd98b: ran T3CODE_DEV_SEED=0 CODEX_HOME=<scratch> vp run dev, where the scratch Codex home held the hub config plus a stdio MCP server metalog whose tool record_marker logs params._meta["x-codex-turn-metadata"] on every tools/call. Paired the web client in the preview browser and, in a new Codex thread (GPT-6-Astra, codex-cli 0.169.0), asked the agent to call record_marker("parent"), then delegate_task a Codex child that calls record_marker("child"). Both calls were logged. Each had session_id, thread_id, turn_id, model and codex_version, and neither had a thread_source key. The database showed the child's native Codex thread (01a1d8c9-b23f-…) owned by app thread thread:delegated-task:command%3Amcp%3A…, so it was a delegate_task child, not a native Codex subagent. Then added a failing adapter test (CodexAdapterV2.test.ts, 'starts delegate_task children with the subagent thread source'). It records threadSource on thread/start for a parent thread and for a thread whose id comes from delegatedTaskThread, and gets [undefined, undefined] where it expects [undefined, "subagent"].
Evidence
Run the real Codex parent → record_marker('parent') → T3 delegate_task → child record_marker('child') flow on fresh databases. GPT-6-Astra, codex-cli 0.169.0, Chromium, 1280×800, light theme. Reproduced twice per commit; screenshots show the second run. Received MCP logs independently confirm the displayed results and identify both children as T3-owned delegates.
Captures are cropped to the conversation for privacy. Recordings begin with delegation underway; the head recording ends with a three-second hold of the final still. Runtime exports remove only the private workspaces map. All captures were visually inspected before upload.
Tests
vp test run src/mcp/OrchestratorMcpToolkit.integration.test.ts on 6bedaa6: 4 failed (reproduces CI). Same file after the fixture fix: 4 passed. vp test run of RestartContinuation, ThreadStop, RemoteDelegatedTask, ProviderRuntimeRecoveryService, peer/RemoteDelegation and mcp/McpHttpServer tests: 196 passed. Pre-commit fmt and lint passed. CI on f8074ae: all 8 required checks passed.
Not in this change
thread/fork also accepts threadSource. delegate_task does not create children through forkThread today, so I left fork unchanged.
Other providers' adapters (Claude, Cursor, OpenCode, ...) were not checked for an equivalent sub-agent marker; the issue is Codex-specific.
thread/fork also accepts threadSource. delegate_task does not create children through forkThread today, so fork is unchanged.
Other providers' adapters were not checked for an equivalent sub-agent marker; the issue is Codex-specific.
Who did the work
Written by Claude Opus 5.5, in Claude Code
Reviewed by GPT-6.1 Sol, in Codex, and by Claude Opus 5.5, each on its own
Proof captured by GPT-6.1 Sol and checked by Claude Opus 5.5
…urce
Failing test for #36520: Codex thread/start for a delegate_task child
carries no threadSource, so the child's MCP tool calls look like a
user thread's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex delegate_task children were started with thread/start and no
threadSource, so their MCP tool calls carried no thread_source and MCP
servers could not tell them from threads the user opened. ensureThread
now sends threadSource "subagent" when the app thread id is a
delegate_task child.
Fixes#36520
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Macroscope's review found this PR approvable — Delegated Codex children now start with subagent classification so their MCP calls carry the appropriate source. Ordinary threads, identifier formats and runtime policies remain unchanged, keeping the change narrowly scoped.
Notes:
No code objects were reviewed. Approvability was decided on eligibility alone.
You can add or adjust custom eligibility rules. Learn more.
No actionable comments were generated in the recent review.
Walkthrough
The change identifies delegated-task thread IDs with a shared prefix. When Codex starts one of these threads, it sets threadSource to "subagent". A replay-based test checks the source values for the parent and child threads.
Codex sets threadSource to "subagent" when it starts a delegated-task thread. The replay-based test checks that the parent has no source and the child has the subagent source. The transcript fixture includes the child source.
Delegated-task children now start as Codex subagents, and user threads are unchanged. No merge-blocking risk was found.
Pre-merge checks: 4
Check name
Status
Explanation
Title check
The title "fix(codex): delegate_task children start as Codex subagents" names the main change. The PR adds threadSource "subagent" to thread/start params for delegate_task child threads in the Codex a…
Linked Issues check
[#36520] The issue requires that a Codex delegate_task child send thread_source "subagent" in its MCP turn metadata, and that threads opened by users stay unchanged. At the reviewed head, CodexAdapter…
Out of Scope Changes check
Every changed area connects to #36520. CodexAdapterV2.ts carries the behavior change. IdAllocator.ts adds the predicate and the shared prefix constant that the gate needs. It also changes derive.deleg…
Description check
The description has a clear problem statement, a change explanation, a reproduction on main, and detailed verification. It does not follow the template exactly. It has no Scope section that links an i…
Full details: Title check
Explanation
The title "fix(codex): delegate_task children start as Codex subagents" names the main change. The PR adds threadSource "subagent" to thread/start params for delegate_task child threads in the Codex adapter, so the title matches the changeset and follows the conventional commit form.
Full details: Linked Issues check
Explanation
[#36520] The issue requires that a Codex delegate_task child send thread_source "subagent" in its MCP turn metadata, and that threads opened by users stay unchanged. At the reviewed head, CodexAdapterV2.ensureThread adds threadSource: "subagent" to the thread/start params only when isDelegatedTaskThread(threadInput.threadId) is true. Other threads receive no threadSource. isDelegatedTaskThread in packages/provider-core/src/server/IdAllocator.ts is a startsWith check on DELEGATED_TASK_THREAD_PREFIX. derive.delegatedTaskThread builds its ID from the same constant, so the gate matches the IDs the app creates. The regression test 'starts delegate_task children with the subagent thread source' covers a parent thread (no threadSource) and a child thread (threadSource "subagent"). The PR reports that this test fails without the fix and that CodexAdapterV2.test.ts passes (168/168). The issue's app-server check shows that threadSource on thread/start reaches later tool-call metadata. The PR does not run a live codex-cli. Only coding requirements apply. The PR's pending evidence section is not a coding requirement. Incremental change since the previous review is limited to the same gate and the same test fixture, so the prior conclusion still holds.
Full details: Out of Scope Changes check
Explanation
Every changed area connects to #36520. CodexAdapterV2.ts carries the behavior change. IdAllocator.ts adds the predicate and the shared prefix constant that the gate needs. It also changes derive.delegatedTaskThread to use that constant. The test file adds the regression test. The replay fixture codex_transcript.ndjson adds threadSource "subagent" to the child's expected thread/start frame, which the strict replay harness needs. The PR leaves thread/fork and other providers unchanged. The issue does not require either change, so their absence is not out of scope. No unrelated change appears in the reviewed summary or diff hunks.
Full details: Description check
Explanation
The description has a clear problem statement, a change explanation, a reproduction on main, and detailed verification. It does not follow the template exactly. It has no Scope section that links an issue for the bug, but it does cite Fixes #36520. It also omits the required 'Problem' and 'Scope' headings and has no conventional commit title. The Verification section is thorough, with before/after evidence. It also has duplicated bullets in the 'Not in this change' section and refers to a 'rework brief' and 'factory' process that is internal context, not repository template content. Overall the description is mostly complete.
Finishing Touches
Generate unit tests (beta)
Commit to this branch
Create a new PR
Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 0 remain after this review.
The delegated_task_status Codex transcript recorded the child's
thread/start without threadSource; T3 now sends "subagent" for
delegate_task children.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
size:S10-29 changed lines (additions + deletions).vouch:trustedAuthor is trusted by repo permissions or the VOUCHED list.
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
github-actions
Bot
added
vouch:trusted
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
The rework brief listed nothing to change; it asked me to stay on PR #36525. CI then failed 'Test Server 2' (and the aggregate 'Check') on 6bedaa6. The failure was real: the four OrchestratorMcpToolkit.integration.test.ts tests ('reports distinct/identical/emoji/long child results') replay the recorded delegated_task_status Codex transcript, and its child.thread/start frame had no threadSource. With the fix, T3 sends threadSource "subagent" for delegate_task children, so the replay harness rejected the frame. The job's later 'not enough space on the disk' error was runner noise, not the cause. I reproduced the four failures locally with the same test file, then added "threadSource":"subagent" to that one expected outbound frame in apps/server/src/orchestration-v2/testkit/fixtures/delegated_task_status/codex_transcript.ndjson. No other Codex fixture starts a delegated child. The fix is unchanged. On f8074ae all 8 required checks pass and Macroscope approved with no comments on any line. CodeRabbit's review of 6bedaa6 had no actionable comments. Its re-run on f8074ae was still in progress when I stopped watching, and it had reported 0 included reviews left that hour. No finding rated high or above is open.
Fixes #36520
Reproduction on main
On origin/main 56fd98b: ran
T3CODE_DEV_SEED=0 CODEX_HOME=<scratch> vp run dev, where the scratch Codex home held the hub config plus a stdio MCP servermetalogwhose toolrecord_markerlogs params._meta["x-codex-turn-metadata"] on every tools/call. Paired the web client in the preview browser and, in a new Codex thread (GPT-6-Astra, codex-cli 0.169.0), asked the agent to call record_marker("parent"), then delegate_task a Codex child that calls record_marker("child"). Both calls were logged. Each had session_id, thread_id, turn_id, model and codex_version, and neither had a thread_source key. The database showed the child's native Codex thread (01a1d8c9-b23f-…) owned by app threadthread:delegated-task:command%3Amcp%3A…, so it was a delegate_task child, not a native Codex subagent. Then added a failing adapter test (CodexAdapterV2.test.ts, 'starts delegate_task children with the subagent thread source'). It records threadSource on thread/start for a parent thread and for a thread whose id comes from delegatedTaskThread, and gets [undefined, undefined] where it expects [undefined, "subagent"].Evidence
Run the real Codex parent → record_marker('parent') → T3 delegate_task → child record_marker('child') flow on fresh databases. GPT-6-Astra, codex-cli 0.169.0, Chromium, 1280×800, light theme. Reproduced twice per commit; screenshots show the second run. Received MCP logs independently confirm the displayed results and identify both children as T3-owned delegates.
Reproduction driver · MCP logger
Captures are cropped to the conversation for privacy. Recordings begin with delegation underway; the head recording ends with a three-second hold of the final still. Runtime exports remove only the private workspaces map. All captures were visually inspected before upload.
Tests
vp test run src/mcp/OrchestratorMcpToolkit.integration.test.ts on 6bedaa6: 4 failed (reproduces CI). Same file after the fixture fix: 4 passed. vp test run of RestartContinuation, ThreadStop, RemoteDelegatedTask, ProviderRuntimeRecoveryService, peer/RemoteDelegation and mcp/McpHttpServer tests: 196 passed. Pre-commit fmt and lint passed. CI on f8074ae: all 8 required checks passed.
Not in this change
Who did the work
Run by the t3code factory in T3 Code.