fix(agent-core): stop loop after aborted tool run by szsip239 · Pull Request #94412 · openclaw/openclaw

szsip239 · 2026-06-18T03:36:20Z

Summary

Stop the agent loop after an in-flight tool aborts the run, once the existing tool result and turn_end have been emitted.
Avoid requesting another model turn with an already-aborted signal.
Add a regression test that fails if a tool-triggered abort calls the model a second time.

Real behavior proof

Behavior addressed: When a tool aborts the active run during tool execution, current origin/main still starts another model turn with the already-aborted signal. That extra turn can keep the caller/session path open after the user-visible abort, which matches the observed session-lock symptom after switching away from an in-flight OpenClaw run. This PR ends the agent loop immediately after the aborted tool result and turn_end have been emitted.

Real environment tested: Local OpenClaw source checkout on macOS, Node v26.3.0, pnpm 11.2.2. I ran the same inline reproduction against current origin/main at 0cae5b3672 and this PR branch codex/fix-abort-tool-lock-release at 1096846d2f. The reproduction imports the real OpenClaw packages/agent-core/src/agent-loop.ts through tsx and exercises runAgentLoop with a model stream, an actual OpenClaw tool object, and an AbortController signal; no mocked unit-test runner is involved in this proof.

Exact steps or command run after this patch: On the PR branch, I ran an inline node --import tsx --input-type=module reproduction that creates a first assistant toolCall, executes an abort_tool whose execute calls controller.abort(new Error("user aborted")), then fails if streamFn is called more than once. I also ran node scripts/run-vitest.mjs packages/agent-core/src/agent-loop.test.ts -t "does not request another model turn after a tool aborts the run".

Evidence after fix: The same reproduction fails on origin/main and passes on this PR branch. On origin/main (0cae5b3672) it printed:

{
  "commit": "0cae5b3672",
  "streamCalls": 2,
  "aborted": true,
  "roles": ["user", "assistant", "toolResult", "assistant"],
  "finalEvent": "agent_end"
}

and then threw Error: BUG: model turn continued after abort. On this PR branch (1096846d2f) it printed:

{
  "commit": "1096846d2f",
  "streamCalls": 1,
  "aborted": true,
  "roles": ["user", "assistant", "toolResult"],
  "finalEvent": "agent_end"
}

The targeted regression test output was:

Test Files  1 passed (1)
Tests  1 passed | 13 skipped (14)
Duration  619ms

Observed result after fix: After the tool aborts the run, OpenClaw emits the tool result and agent_end without issuing a second model request. The returned messages stop at user, assistant, and toolResult; there is no extra assistant message from an aborted follow-up turn.

What was not tested: I did not rerun a full packaged gateway/CLI release build with a live external model provider for this PR-body refresh. The proof above isolates the real OpenClaw agent-loop path locally; full-suite and broader gateway checks remain supplemental CI/maintainer validation.

Verification

node scripts/run-vitest.mjs packages/agent-core/src/agent-loop.test.ts -t "does not request another model turn after a tool aborts the run"
node scripts/run-vitest.mjs packages/agent-core/src/agent-loop.test.ts
node scripts/run-vitest.mjs src/agents/embedded-agent-runner/run/attempt.subscription-cleanup.test.ts src/agents/embedded-agent-runner/run/attempt.session-lock.test.ts
pnpm tsgo:test:packages
pnpm tsgo:core

clawsweeper · 2026-06-18T04:59:16Z

Codex review: needs maintainer review before merge. Reviewed June 22, 2026, 5:00 AM ET / 09:00 UTC.

Summary
The PR adds a shared abort-stop helper in packages/agent-core/src/agent-loop.ts and regression tests for tool-triggered and next-turn-hook-triggered aborts.

PR surface: Source +43, Tests +148. Total +191 across 2 files.

Reproducibility: yes. Current source shows the missing post-tool abort check before prepareNextTurn, and the PR body supplies a before/after terminal reproduction showing the extra model call on main and one stream call on the PR branch.

Review metrics: 1 noteworthy metric.

Terminal abort outcome: 1 synthesized assistant terminal message added. Maintainers should notice that the latest patch intentionally changes the abort transcript from ending on a tool result to ending on an aborted assistant outcome.

Merge readiness
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
Patch quality: 🐚 platinum hermit
Result: ready for maintainer review.

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

Rank-up moves:

Refresh or investigate the failed checks-node-core-runtime-infra-process shard before landing.

Risk before merge

[P1] Current PR head still has one failing CI shard outside the changed files, so maintainers need a CI refresh or scoped failure review before landing.
[P1] The PR body's role-list proof predates the later terminal-aborted-message commits; the current tests and diff, not that older role list, reflect the latest terminal transcript behavior.

Maintainer options:

Decide the mitigation before merge
Land the focused shared agent-core abort boundary after current-head CI is clean or the remaining unrelated shard failure is explicitly scoped by maintainers.
Pause or close
Do not merge this PR until maintainers decide whether the risk is worth taking.

Next step before merge

[P2] The remaining action is maintainer review plus CI refresh/scoping, not a concrete ClawSweeper repair job on the patch.

Security
Cleared: The diff only changes agent-core loop control flow and colocated tests, with no dependency, workflow, secret, install, publishing, or supply-chain surface.

Review details

Best possible solution:

Land the focused shared agent-core abort boundary after current-head CI is clean or the remaining unrelated shard failure is explicitly scoped by maintainers.

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

Yes. Current source shows the missing post-tool abort check before prepareNextTurn, and the PR body supplies a before/after terminal reproduction showing the extra model call on main and one stream call on the PR branch.

Is this the best way to solve the issue?

Yes. The shared agent-core loop boundary is the right layer because both the plain Agent and harness paths consume runAgentLoop, and session consumers already key terminal handling off message and turn events.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against e20edd753bd7.

Label changes

Label justifications:

P1: The PR targets a broken agent abort/session workflow where a tool-triggered abort can continue into an extra model turn instead of releasing the run promptly.
rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🦞 diamond lobster and patch quality is 🐚 platinum hermit.
status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (terminal): The PR body includes terminal before/after output from a real source checkout showing current main makes a second stream call after abort while the PR branch stops at one call; current tests cover the later terminal assistant refinement.
proof: sufficient: Contributor real behavior proof is sufficient. The PR body includes terminal before/after output from a real source checkout showing current main makes a second stream call after abort while the PR branch stops at one call; current tests cover the later terminal assistant refinement.

Evidence reviewed

PR surface:

Source +43, Tests +148. Total +191 across 2 files.

View PR surface stats

Area	Files	Added	Net
Source	1	43	+43
Tests	1	148	+148
Docs	0	0	0
Config	0	0	0
Generated	0	0	0
Other	0	0	0
Total	2	191	+191

What I checked:

Current main abort gap: Current main emits turn_end after tool execution, then builds nextTurnContext and calls prepareNextTurn without checking whether the signal was aborted during tool execution. (packages/agent-core/src/agent-loop.ts:334, e20edd753bd7)
PR terminal abort behavior: The PR head appends a zero-usage aborted assistant message, opens a new terminal turn when needed, emits message and turn lifecycle events, and then emits agent_end. (packages/agent-core/src/agent-loop.ts:273, d25ab29d3a3b)
Regression coverage: The PR head covers both a tool that aborts during execution and an async prepareNextTurn hook that aborts, asserting only one model stream call and a terminal aborted assistant outcome. (packages/agent-core/src/agent-loop.test.ts:771, d25ab29d3a3b)
Caller lifecycle: Agent.runWithLifecycle keeps activeRun and isStreaming set until runAgentLoop returns and finishRun() clears the run state, which explains why an unnecessary post-abort model turn can delay session release. (packages/agent-core/src/agent.ts:510, e20edd753bd7)
Session consumer contract: Session handling persists normal assistant/tool messages on message_end, tracks the last assistant message for terminal handling, and forwards agent_end messages to extensions, so the PR's terminal assistant outcome is consumed by existing lifecycle surfaces. (src/agents/sessions/agent-session.ts:617, e20edd753bd7)
CI status: Current PR CI has one failed shard, checks-node-core-runtime-infra-process, failing src/config/plugin-auto-enable.core.test.ts:1223; the failure is outside the changed agent-core files. (d25ab29d3a3b)

Likely related people:

steipete: GitHub path history shows the agent-core package extraction and later terminal lifecycle work on agent-loop.ts, agent.ts, types.ts, and related tests. (role: package extraction and terminal lifecycle contributor; confidence: high; commits: bb46b79d3c14, aa0d6e1bca55, 0314819f918a; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/agent.ts, packages/agent-core/src/types.ts)
fuller-stack-dev: GitHub path history shows deferred tool hydration and co-authored terminal lifecycle work on the same agent-loop/type surfaces involved in tool continuation behavior. (role: recent agent-loop feature contributor; confidence: medium; commits: 965fa05df38b, 0314819f918a; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/agent.ts, packages/agent-core/src/types.ts)
vincentkoc: Local blame and GitHub history show recent agent-loop and harness maintenance, including text-delta loop behavior and harness API cleanup near this runtime surface. (role: recent adjacent loop contributor; confidence: medium; commits: 6fcc9457020e, 464adfe5e544, f05fd56d666a; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/harness/agent-harness.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.

vincentkoc · 2026-06-22T10:04:58Z

Merged via squash.

Prepared head SHA: e11d9718e3cd84e50f5659766b6a23f42b18a741
Merge commit: 7fe287b0d30d9ae3998dbf4da04f9498dd37f7a6

Thanks @szsip239!

@vincentkoc

Merged via squash. Prepared head SHA: e11d971 Co-authored-by: szsip239 <[email protected]> Co-authored-by: vincentkoc <[email protected]> Reviewed-by: @vincentkoc

openclaw-barnacle Bot added size: S triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup. labels Jun 18, 2026

openclaw-barnacle Bot added proof: supplied External PR includes structured after-fix real behavior proof. and removed triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup. labels Jun 18, 2026

vincentkoc force-pushed the codex/fix-abort-tool-lock-release branch 2 times, most recently from 0308d98 to 468234a Compare June 22, 2026 05:31

clawsweeper Bot added rating: 🐚 platinum hermit Good normal PR readiness with ordinary maintainer review expected. and removed rating: 🦞 diamond lobster Very strong PR readiness with only minor maintainer review expected. labels Jun 22, 2026

vincentkoc force-pushed the codex/fix-abort-tool-lock-release branch 9 times, most recently from fa08a2d to 530e68f Compare June 22, 2026 09:47

szsip239 and others added 2 commits June 22, 2026 17:53

fix(agent-core): stop loop after aborted tool run

25c5e39

fix(agent-core): recheck aborts before follow-up turns

6c2fbb0

vincentkoc added 4 commits June 22, 2026 17:53

fix(agent-core): preserve aborted terminal outcome

3da71a2

fix(agent-core): balance aborted turn lifecycle

647a5cd

fix(agent-core): balance aborted terminal turn events

16eb727

fix(agent-core): start aborted terminal turns before messages

e11d971

vincentkoc force-pushed the codex/fix-abort-tool-lock-release branch from 530e68f to e11d971 Compare June 22, 2026 09:53

vincentkoc merged commit 7fe287b into openclaw:main Jun 22, 2026
75 checks passed

clawsweeper Bot mentioned this pull request Jun 22, 2026

feat(session): add chat-channel session switching with LRU history retention #49914

Closed

20 tasks

This was referenced Jul 1, 2026

chore: bump openclaw/openclaw version from v2026.6.10 to v2026.6.11 Docker-Hub-sirmark/docker-openclaw#35

Merged

chore(openclaw): bump pinned version v2026.5.28 → v2026.6.11 Evans-Software-Solutions-Limited/axel-saas#121

Open

quick3210-gif mentioned this pull request Jul 8, 2026

[BUG] exec tool: aborted run wedges subsequent exec calls in same session (regression from PR #94412) #102006

Open

yetval mentioned this pull request Jul 8, 2026

agent-core: parallel batch runs a prepared paid tool after the run is aborted mid-batch #102252

Closed

This was referenced Jul 8, 2026

fix(agent-core): stop canceled parallel tools from starting #102276

Merged

fix(agent-core): skip prepared tool execution after abort #102283

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(agent-core): stop loop after aborted tool run#94412

fix(agent-core): stop loop after aborted tool run#94412
vincentkoc merged 6 commits into
openclaw:mainfrom
szsip239:codex/fix-abort-tool-lock-release

szsip239 commented Jun 18, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jun 18, 2026 •

edited

Loading

Uh oh!

Uh oh!

vincentkoc commented Jun 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

szsip239 commented Jun 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Real behavior proof

Verification

Uh oh!

clawsweeper Bot commented Jun 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

vincentkoc commented Jun 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

szsip239 commented Jun 18, 2026 •

edited

Loading

clawsweeper Bot commented Jun 18, 2026 •

edited

Loading