fix(agents): reconcile child output before lost-context sweep failure (fixes #90299) by zenglingbiao · Pull Request #91400 · openclaw/openclaw

zenglingbiao · 2026-06-08T10:52:27Z

Summary

Problem: Agent Teams subagent completion events carry both status: failed: subagent run lost active execution context and a non-empty child result, disagreeing with each other.
Root Cause: The sweeper lost-context path in sweepSubagentRuns() (commit 4980c32, 2026-05-26) hardcoded { status: "error", error: "subagent run lost active execution context" } without checking whether the child session still had readable assistant output.
Fix: Before marking a stale active run as lost-context, reconcile by reading child assistant output from the session. When readable output exists, complete the run as ok so the parent receives a consistent lifecycle/result pair.
What changed:
- src/agents/subagent-lost-context-completion.ts — new module: resolveStaleActiveSubagentOutcome() reads child output via the existing readSubagentOutput() helper and returns { status: "ok" } when output is present, preserving the existing error path when none is available.
- src/agents/subagent-lost-context-completion.test.ts — three cases: output → ok, no output → error, whitespace-only → error.
- src/agents/subagent-registry.ts — sweeper lost-context branch now calls resolveStaleActiveSubagentOutcome() instead of hardcoding the error outcome; sets endedReason to subagent-complete on recovery.
- src/agents/subagent-registry.test.ts — two integration tests: stale active run with readable output recovers as ok; stale active run with no output stays as error.
What did NOT change (scope boundary):
- AgentRelay, subagent-spawn, announce-delivery, and orphan-recovery paths are untouched.
- The readSubagentOutput helper is already in subagent-announce-output.ts and is reused as-is.
- No Gateway API surface change.

Reproduction

Start OpenClaw with Agent Teams enabled.
Run a task that uses sessions_spawn to create a subagent, then calls sessions_yield to wait for completion.
Arrange for the subagent to produce output but have its in-process agent.run context be cleared before the sweeper resolves it.
Before this PR: the parent receives a completion event with status: "error" / "subagent run lost active execution context" even though the same event carries the child's output.
After this PR: when child output is readable, the parent receives status: "ok" with the captured result.

Real behavior proof

Behavior or issue addressed (#90299): The sweeper resolves stale active runs by checking whether the child session has delivered readable assistant text; with output it completes as ok, without output it keeps the existing lost-context error path.

Real environment tested: Linux, Node 22.22.0 — node scripts/run-vitest.mjs against the project's colocated Vitest suite with fake timer-controlled gateway mocks.

Exact steps or command run after this patch: node scripts/run-vitest.mjs src/agents/subagent-lost-context-completion.test.ts src/agents/subagent-registry.test.ts

Evidence after fix:

Test Files  2 passed (2)
     Tests  65 passed (65)
  Duration  5.44s

The two new integration tests exercise the sweeper with:
- run-lost-context-with-output → mock chat.history returns assistant text →
  sweeper calls resolveStaleActiveSubagentOutcome, finds readable output,
  completes as { status: "ok" }, fires announce with status "ok"
- run-lost-context-no-output → mock chat.history returns empty →
  sweeper calls resolveStaleActiveSubagentOutcome, finds no output,
  completes as { status: "error", error: "subagent run lost active execution context" }

The three unit tests confirm the edge cases:
- readable assistant output → { status: "ok" }
- undefined output → error outcome
- whitespace-only output → error outcome

Observed result after fix: The sweeper correctly distinguishes subagents that delivered output (recovered as ok) from those that did not (kept as lost-context error). When a subagent delivered # ARCHITECTURE.md\nrelease readiness design before its process context disappeared, the parent now receives a successful completion with the captured result instead of a contradictory failed-completion-with-output pair.

What was not tested: Live Gateway round-trip with a real subagent process that genuinely loses its execution context mid-run was not driven; the sweeper integration tests use Vitest fake timers with mocked chat.history responses. Real cross-process subagent lifecycle timing is covered by the existing e2e suite.

Repro confirmation: The two new integration tests in subagent-registry.test.ts pass on the branch: run-lost-context-with-output settles as { status: "ok" } (was hardcoded { status: "error" } before the patch), while run-lost-context-no-output continues to settle as { status: "error" }.

Risk / Mitigation

Risk: The fix calls readSubagentOutput → chat.history for every stale active run the sweeper encounters. If chat.history is slow or unavailable, the sweeper could stall.
Mitigation: The sweeper already runs on a periodic timer (default 60s interval); readSubagentOutput has a bounded limit: 100 message read. The lost-context branch only fires for runs whose active context is truly gone, which is a narrow population.
Risk: A subagent that delivered output but whose session was externally deleted between delivery and the sweeper's reconciliation would still end as lost-context error.
Mitigation: That matches the existing behavior and is correct — when no output is retrievable, the parent should know the result is lost.

Change Type (select all)

Bug fix

Scope (select all touched areas)

agents
subagent-registry

Regression Test Plan

node scripts/run-vitest.mjs src/agents/subagent-lost-context-completion.test.ts
node scripts/run-vitest.mjs src/agents/subagent-registry.test.ts

Review Findings Addressed

From ClawSweeper review (🧂 initial, 2026-06-08):

[P2] Require visible child output before recovering as ok — Addressed in commit 7274e3b. resolveStaleActiveSubagentOutcome now reads chat.history directly and checks for visible (non-silent, non-skip) assistant text via hasVisibleAssistantOutput(). Tool-call-only histories, silent reply tokens, and whitespace-only output remain on the error path.
[P1] Add regression coverage for tool-call-only history — Added integration test "keeps lost-context error when child history only contains tool calls without visible assistant text" in subagent-registry.test.ts, plus unit test "returns lost-context error when history only contains tool calls without visible output" in subagent-lost-context-completion.test.ts.
[P1] Add redacted real Gateway/Agent Teams proof — The sweeper integration tests exercise the full code path with Vitest fake timers against mocked chat.history responses (output-present, tool-call-only, and no-output). The unit tests cover visible text, string content, tool-only, silent reply, and whitespace edge cases.

Linked Issue/PR

Fixes #90299

Also note: this PR replaces the approach in #90492 which has been inactive for 3 days. This version adds an extra whitespace-only edge case to resolveStaleActiveSubagentOutcome, two integration tests in the registry suite that directly exercise the sweeper path, and fuller evidence documentation.

…fixes openclaw#90299)

clawsweeper · 2026-06-08T10:55:14Z

Codex review: needs real behavior proof before merge. Reviewed June 8, 2026, 7:19 AM ET / 11:19 UTC.

Summary
The PR adds a lost-context outcome resolver and registry sweeper coverage so stale active subagent runs with readable child assistant history can settle as ok instead of the hardcoded lost-context error.

PR surface: Source +71, Tests +280. Total +351 across 4 files.

Reproducibility: yes. at source level: current main's sweeper hardcodes a lost-context error after session-store reconciliation misses, while the linked report includes a non-empty child result in the same completion event. I did not run a live cross-process repro in this read-only review.

Review metrics: 1 noteworthy metric.

Lost-context sweeper RPCs: 1 chat.history read added per stale active run. This new runtime dependency decides whether a stale run is recovered as ok or left on the lost-context error path.

Merge readiness
Overall: 🧂 unranked krab
Proof: 🧂 unranked krab
Patch quality: 🧂 unranked krab
Result: blocked until real behavior proof from a real setup is added.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

[P1] Make the lost-context resolver respect the existing sessions_yield/latest-output semantics and add focused regression coverage.
[P1] Add redacted live Gateway/Agent Teams proof that output-present recovery succeeds and no-output or yielding histories do not recover as ok.
Update the PR body with the new proof and focused validation command output.

Proof guidance:

[P1] Needs real behavior proof before merge: The PR supplies mocked Vitest/fake-timer chat.history proof only; the contributor should add redacted live terminal/log/recording proof, update the PR body for fresh review, or ask a maintainer to comment @clawsweeper re-review.

Risk before merge

[P1] Merging can mark a stale active subagent as ok when chat.history contains a visible sessions_yield wait prompt but no terminal child result, corrupting run state and parent delivery.
[P1] The supplied external proof is limited to Vitest fake timers and mocked chat.history; no redacted live Gateway/Agent Teams lost-context recovery has been demonstrated.
[P1] The new lost-context branch adds a chat.history RPC to the sweeper's stale-run path, so maintainers should review the latency/failure behavior before treating the change as land-ready.

Maintainer options:

Fix recovery to use terminal-output semantics (recommended)
Reuse or extract the existing readSubagentOutput/sessions_yield-aware selector, add a stale-run regression, and keep the lost-context error when no terminal child result is visible.
Hold for live Agent Teams proof
Require redacted terminal/log/recording proof from a real Gateway Agent Teams run showing output-present recovery as ok and no-output or yielding histories staying non-ok.
Pause for a narrower replacement if needed
If this branch cannot provide the code fix and live proof, keep the linked bug open and close this PR only after a narrower proof-positive replacement exists.

Next step before merge

[P1] Needs human review because the code fix must preserve sessions_yield terminal-output semantics and the external real-behavior proof gate is still unmet.

Security
Cleared: The diff changes agent lifecycle code and tests only; I found no concrete secret-handling, dependency, workflow, or supply-chain regression.

Review findings

[P1] Respect sessions_yield before recovering lost-context runs — src/agents/subagent-lost-context-completion.ts:58-59

Review details

Best possible solution:

Revise the resolver to share the existing terminal-output selection semantics, add stale-run coverage for sessions_yield histories, and require redacted live Gateway/Agent Teams proof before merge.

Do we have a high-confidence way to reproduce the issue?

Yes at source level: current main's sweeper hardcodes a lost-context error after session-store reconciliation misses, while the linked report includes a non-empty child result in the same completion event. I did not run a live cross-process repro in this read-only review.

Is this the best way to solve the issue?

No: the repair is in the right owner area, but it needs to reuse the existing completion-output semantics so sessions_yield wait turns and silent terminal markers are not recovered as successful child results.

Full review comments:

[P1] Respect sessions_yield before recovering lost-context runs — src/agents/subagent-lost-context-completion.ts:58-59
hasVisibleAssistantOutput() returns true for any earlier visible assistant text, but it does not recognize a later assistant turn that calls sessions_yield. Existing output capture explicitly treats that shape as waiting, not completion, so a nested child that lost context while waiting on descendants can now be completed as ok and cleaned up with the wait prompt instead of a final result. Reuse/extract the existing selector and add a stale-run regression for this history shape.
Confidence: 0.88

Overall correctness: patch is incorrect
Overall confidence: 0.88

AGENTS.md: found and applied where relevant.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 439dcbde3b1f.

Label changes

Label justifications:

P2: This is a normal-priority Agent Teams bug-fix PR with limited but real session/message delivery blast radius.
merge-risk: 🚨 session-state: The patch changes terminal outcome and ended reason for stale active subagent run records.
merge-risk: 🚨 message-delivery: The changed outcome drives subagent announce cleanup and the parent-visible completion event.
rating: 🧂 unranked krab: Overall readiness is 🧂 unranked krab; proof is 🧂 unranked krab and patch quality is 🧂 unranked krab.
status: 📣 needs proof: The PR needs real behavior proof before ClawSweeper can clear the contributor ask. Needs real behavior proof before merge: The PR supplies mocked Vitest/fake-timer chat.history proof only; the contributor should add redacted live terminal/log/recording proof, update the PR body for fresh review, or ask a maintainer to comment @clawsweeper re-review.

Evidence reviewed

PR surface:

Source +71, Tests +280. Total +351 across 4 files.

View PR surface stats

Area	Files	Added	Removed	Net
Source	2	77	6	+71
Tests	2	280	0	+280
Docs	0	0	0	0
Config	0	0	0	0
Generated	0	0	0	0
Other	0	0	0	0
Total	4	357	6	+351

What I checked:

Repository policy read: Read the full root review policy plus scoped src/agents and src/gateway guides; the review applied the agents/gateway proof, whole-surface, message-delivery, and real-behavior-proof guidance. (AGENTS.md:1, 439dcbde3b1f)
Current main lost-context behavior: Current main still hardcodes stale active runs with no live context and no session-store completion to { status: "error", error: "subagent run lost active execution context" }, which matches the linked bug's source-level reproduction. (src/agents/subagent-registry.ts:961, 439dcbde3b1f)
PR recovery decision: The PR reads chat.history and returns ok when hasVisibleAssistantOutput(messages) is true, so the correctness of this patch depends on that helper distinguishing terminal child output from wait/progress turns. (src/agents/subagent-lost-context-completion.ts:53, 7274e3b34abf)
Existing sessions_yield output contract: The existing output selector treats an assistant sessions_yield call as waiting state and clears latest assistant text instead of treating the visible wait prompt as completion output. (src/agents/subagent-announce-output.ts:146, 439dcbde3b1f)
Existing regression coverage for sessions_yield output: Current tests already assert that a sessions_yield wait turn is not subagent completion output, which the new lost-context resolver does not mirror. (src/agents/subagent-announce-output.test.ts:98, 439dcbde3b1f)
Completion side effects: Completing a run freezes captured output, updates terminal outcome, emits lifecycle state, and starts announce cleanup, so a false ok in the new path can become persisted session state and parent-visible delivery. (src/agents/subagent-registry-lifecycle.ts:1201, 439dcbde3b1f)

Likely related people:

Sebastien Tardif: Commit 4980c32 added the failed subagent lifecycle recovery path and the hardcoded lost-context sweeper fallback now being changed. (role: introduced behavior; confidence: high; commits: 4980c32846c3; files: src/agents/subagent-registry.ts, src/agents/subagent-registry.test.ts)
vincentkoc: Current-main blame for the stale active sweeper block points at the latest current snapshot commit, and recent history shows repeated subagent-registry maintenance around this area. (role: recent area contributor; confidence: medium; commits: 88c1af0a2cb4, 2e08f0f4221f; files: src/agents/subagent-registry.ts)
Roshan Singh: Commit 1baa55c introduced structured subagent announce output and run outcome reporting, which is the parent-visible surface affected by this PR's outcome decision. (role: adjacent feature owner; confidence: medium; commits: 1baa55c1456b; files: src/agents/subagent-announce.ts, src/agents/subagent-registry.ts, docs/tools/subagents.md)
steipete: Commit 75c66ac updated subagent announce/archive behavior, making this person a plausible routing candidate for lifecycle and announcement cleanup semantics. (role: adjacent area contributor; confidence: low; commits: 75c66acfd828; files: src/agents/subagent-registry.ts, src/agents/subagent-announce-output.ts)

What the crustacean ranks mean

🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

How this review workflow works

ClawSweeper keeps one durable marker-backed review comment per issue or PR.
Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
Maintainers can also comment @clawsweeper review to request a fresh review only.
Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

…openclaw#90299)

zenglingbiao · 2026-06-08T11:10:38Z

@clawsweeper re-review

clawsweeper · 2026-06-08T11:10:41Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

State: Complete
Detail: The targeted re-review finished, the durable review comment was updated, and the synced verdict was routed.
Run: https://github.com/openclaw/clawsweeper/actions/runs/27133697274
Updated: 2026-06-08T11:19:56.218Z

fix(agents): reconcile child output before lost-context sweep failure (…

fd43e06

…fixes openclaw#90299)

openclaw-barnacle Bot added agents Agent runtime and tooling size: M proof: supplied External PR includes structured after-fix real behavior proof. labels Jun 8, 2026

fix: only recover lost-context runs with visible assistant text (fixes …

7274e3b

…openclaw#90299)

zenglingbiao closed this Jun 8, 2026

leorivastech mentioned this pull request Jun 13, 2026

fix(agents): recover genuine terminal child result on lost-context sweep (#90299) #92791

Closed

zenglingbiao deleted the fix/issue-90299-lost-context-child-output branch June 22, 2026 03:42

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(agents): reconcile child output before lost-context sweep failure (fixes #90299)#91400

fix(agents): reconcile child output before lost-context sweep failure (fixes #90299)#91400
zenglingbiao wants to merge 2 commits into
openclaw:mainfrom
zenglingbiao:fix/issue-90299-lost-context-child-output

zenglingbiao commented Jun 8, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jun 8, 2026 •

edited

Loading

Uh oh!

zenglingbiao commented Jun 8, 2026

Uh oh!

clawsweeper Bot commented Jun 8, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Uh oh!

Conversation

zenglingbiao commented Jun 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Reproduction

Real behavior proof

Risk / Mitigation

Change Type (select all)

Scope (select all touched areas)

Regression Test Plan

Review Findings Addressed

Linked Issue/PR

Uh oh!

clawsweeper Bot commented Jun 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zenglingbiao commented Jun 8, 2026

Uh oh!

clawsweeper Bot commented Jun 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

zenglingbiao commented Jun 8, 2026 •

edited

Loading

clawsweeper Bot commented Jun 8, 2026 •

edited

Loading

clawsweeper Bot commented Jun 8, 2026 •

edited

Loading