fix(agent): propagate LLM stop reason through lifecycle events by visionik · Pull Request #24867 · openclaw/openclaw

visionik · 2026-02-23T23:12:42Z

Problem

The LLM stop reason (end_turn, max_tokens, etc.) is lost between the agent runner and the gateway event stream. Downstream consumers like the ACP bridge, TUI, and WebSocket clients cannot distinguish a complete response from one truncated by max_tokens.

The stop reason is dropped at two points:

pi-embedded-runner/run.ts only sets meta.stopReason for tool_calls, discarding the actual LLM stop reason from lastAssistant.stopReason
commands/agent.ts emits the lifecycle phase: "end" event without including stopReason

Fix

run.ts: Propagate lastAssistant.stopReason into result.meta.stopReason when there are no client tool calls
agent.ts: Include stopReason in the lifecycle end event data and log non-end_turn stop reasons to stderr

Context

Related to #24856 (flush throttled delta before final). Together these two PRs improve streaming reliability for gateway clients:

fix(gateway): flush throttled delta before emitChatFinal #24856 ensures the complete text is delivered via deltas before the final event
This PR ensures consumers can detect truncated responses via stopReason

Changes

2 files, 10 lines added, 1 removed.

Greptile Summary

Propagates LLM stop reason (end_turn, max_tokens, etc.) from the agent runner to lifecycle events and downstream consumers.

Changes:

run.ts:1138-1140 - propagates lastAssistant.stopReason when no client tool calls present
agent.ts:600-612 - includes stopReason in lifecycle end event and logs non-end_turn reasons to stderr

This allows downstream consumers (ACP bridge, TUI, WebSocket clients) to distinguish complete responses from ones truncated by token limits.

Confidence Score: 5/5

This PR is safe to merge with minimal risk
The changes are minimal, well-scoped, and backward compatible. The new stopReason field is optional and doesn't break existing consumers. The implementation correctly propagates the stop reason from the LLM through the execution flow. Type safety is maintained with appropriate casts, and the change aligns with existing patterns in the codebase.
No files require special attention

_{Last reviewed commit: e8f66e8}

_{(4/5) You can add custom instructions or style guidelines for the agent here!}

kevinWangSheng

Review Summary

Good change in concept - propagating stopReason helps downstream consumers detect truncated responses.

Issue found

⚠️ CI check failed - check job failed: https://github.com/openclaw/openclaw/actions/runs/22329117075/job/64607637997

Please fix the failing check and I'll be happy to approve.

Code review notes

The implementation is clean and follows existing patterns
Adding stopReason to lifecycle events is backward compatible
Good logging for non-end_turn cases

Status: Request changes until CI passes.

Complete the stop reason propagation chain so ACP clients can distinguish end_turn from max_tokens: - server-chat.ts: emitChatFinal accepts optional stopReason param, includes it in the final payload, reads it from lifecycle event data - translator.ts: read stopReason from the final payload instead of hardcoding end_turn Chain: LLM API → run.ts (meta.stopReason) → agent.ts (lifecycle event) → server-chat.ts (final payload) → ACP translator (PromptResponse)

visionik · 2026-03-03T05:42:15Z

@kevinWangSheng thank you for the comment. Passed all tests and I just merged. Happy to have your input on any of my future PR's!

…openclaw#24867) Complete the stop reason propagation chain so ACP clients can distinguish end_turn from max_tokens: - server-chat.ts: emitChatFinal accepts optional stopReason param, includes it in the final payload, reads it from lifecycle event data - translator.ts: read stopReason from the final payload instead of hardcoding end_turn Chain: LLM API → run.ts (meta.stopReason) → agent.ts (lifecycle event) → server-chat.ts (final payload) → ACP translator (PromptResponse)

@Takhoffman

* main: (134 commits) fix(telegram): warn when accounts.default is missing in multi-account setup (openclaw#32544) agents: propagate config for embedded skill loading Gateway: fix stale self version in status output (openclaw#32655) feat(mattermost): add native slash command support (refresh) (openclaw#32467) Diffs: Migrate tool usage guidance from before_prompt_build to a plugin skill (openclaw#32630) bug: Workaround for QMD upstream bug (openclaw#27028) fix: improve compaction summary instructions to preserve active work (openclaw#8903) chore: Updated Brave documentation (openclaw#26860) security(line): synthesize strict LINE auth boundary hardening test(e2e): isolate module mocks across harnesses fix(telegram): debounce forwarded media-only bursts test(live): harden gateway model profile probes fix(ci): handle disabled systemd units in docker doctor flow fix(test): stabilize appcast version assertion fix(line): synthesize media/auth/routing webhook regressions (openclaw#32546) thanks @Takhoffman fix(gateway+acp): thread stopReason through final event to ACP bridge (openclaw#24867) docs(changelog): reattribute duplicated PR credits fix: scope extension runtime deps to plugin manifests ci: enable stale workflow chore(release): bump to 2026.3.3 and seed changelog ...

…openclaw#24867) Complete the stop reason propagation chain so ACP clients can distinguish end_turn from max_tokens: - server-chat.ts: emitChatFinal accepts optional stopReason param, includes it in the final payload, reads it from lifecycle event data - translator.ts: read stopReason from the final payload instead of hardcoding end_turn Chain: LLM API → run.ts (meta.stopReason) → agent.ts (lifecycle event) → server-chat.ts (final payload) → ACP translator (PromptResponse)

openclaw-barnacle bot added gateway Gateway runtime commands Command implementations agents Agent runtime and tooling size: XS maintainer Maintainer-authored PR labels Feb 23, 2026

kevinWangSheng reviewed Feb 23, 2026

View reviewed changes

visionik force-pushed the fix/propagate-stop-reason branch from 83b30ee to 333c160 Compare March 3, 2026 05:16

openclaw-barnacle bot removed commands Command implementations agents Agent runtime and tooling labels Mar 3, 2026

visionik merged commit 0b3bbfe into openclaw:main Mar 3, 2026
29 checks passed

github-actions bot mentioned this pull request Mar 3, 2026

📡 Upstream Digest — 2026-03-03 06:40 UTC curtismercier/openclaw-mods#170

Open

MillionthOdin16 added a commit to MillionthOdin16/openclaw that referenced this pull request Mar 3, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(agent): propagate LLM stop reason through lifecycle events#24867

fix(agent): propagate LLM stop reason through lifecycle events#24867
visionik merged 1 commit intoopenclaw:mainfrom
visionik:fix/propagate-stop-reason

visionik commented Feb 23, 2026 •

edited by greptile-apps bot

Loading

Uh oh!

kevinWangSheng left a comment

Uh oh!

Uh oh!

visionik commented Mar 3, 2026

Uh oh!

Uh oh!

Conversation

visionik commented Feb 23, 2026 • edited by greptile-apps bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Context

Changes

Greptile Summary

Confidence Score: 5/5

Uh oh!

kevinWangSheng left a comment

Choose a reason for hiding this comment

Review Summary

Issue found

Code review notes

Uh oh!

Uh oh!

visionik commented Mar 3, 2026

Uh oh!

visionik commented Feb 23, 2026 •

edited by greptile-apps bot

Loading