fix(agent-core): ignore truncated tool calls by steipete · Pull Request #97140 · openclaw/openclaw

steipete · 2026-06-27T01:47:56Z

What Problem This Solves

Closes #97091. A model response that reaches a length, error, or aborted terminal while still emitting tool calls can contain incomplete calls. Those calls must not execute, persist, or replay as valid tool use.

This supersedes #97092. The contributor fork cannot accept maintainer pushes, so this branch preserves @galiniliev's work and coauthor credit while carrying the review fixes.

Why This Change Was Made

Make toolUse the final agent-loop authorization gate for tool dispatch.
Remove tool-call blocks from non-executable terminal messages before events, persistence, return, and next-turn replay.
Preserve explicit length and error terminals across Google, Ollama, shared OpenAI-compatible, and DeepSeek DSML normalization instead of reclassifying them from parsed calls.
Require an observed provider terminal before recovered DSML calls become executable, so EOF cannot authorize a complete-looking partial call.
Emit provider failures through the shared error terminal path rather than a successful done event.
Keep completed tool-use behavior unchanged.

User Impact

Truncated or failed generations no longer execute partial tool calls, including unintended child-session spawns. Complete tool calls still execute normally.

Evidence

node scripts/run-vitest.mjs packages/agent-core/src/agent-loop.test.ts src/plugin-sdk/provider-stream-shared.test.ts src/llm/providers/google-shared.test.ts src/agents/openai-transport-stream.test.ts extensions/google/transport-stream.test.ts extensions/ollama/src/stream.test.ts — 451 tests passed across four shards.
oxfmt --check and oxlint --deny-warnings passed for all 12 changed files.
git diff --check origin/main...HEAD passed.
Fresh autoreview after all accepted fixes: no findings; patch correct at 0.96 confidence.

Co-authored-by: Galin Iliev <[email protected]>

clawsweeper · 2026-06-27T01:49:02Z

Codex review: needs real behavior proof before merge. Reviewed June 26, 2026, 10:13 PM ET / 02:13 UTC.

Summary
The branch gates agent-core tool dispatch on final toolUse turns, strips non-executable terminal tool calls from committed/replayed messages, and updates Google, Ollama, OpenAI-compatible, and provider-wrapper normalization tests.

PR surface: Source +26, Tests +382. Total +408 across 12 files.

Reproducibility: yes. Source inspection gives a high-confidence reproduction path: return a final assistant message with stopReason: "length" and a toolCall block, and current main routes it into executeToolCalls.

Review metrics: 1 noteworthy metric.

Tool execution contract: 1 core dispatch gate tightened; 5 provider/wrapper terminal paths covered. This is the boundary that decides whether provider-returned tool-call blocks can run side-effectful tools.

Root-cause cluster
Relationship: fixed_by_candidate
Canonical: #97091
Summary: This PR is the current candidate fix for the canonical truncated assistant tool-call execution bug and supersedes the earlier contributor branch.

Members:

canonical: [Bug]: truncated assistant tool-call turn executes unintended sessions_spawn #97091 - The PR body uses closing linkage for the issue, and both describe the same length/error/truncated assistant tool-call execution path.
superseded: fix(agent-core): ignore truncated assistant tool calls #97092 - The PR body says this replacement branch preserves the earlier contributor work while carrying review fixes because the original fork could not accept maintainer pushes.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge readiness
Overall: 🐚 platinum hermit
Proof: 🌊 off-meta tidepool
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:

[P2] Confirm maintainer acceptance of the toolUse-only execution contract after required checks finish.

Risk before merge

[P1] Merging intentionally changes compatibility for provider/plugin adapters: tool-call blocks with length, error, EOF, or other non-toolUse terminals will no longer execute, so adapters must emit toolUse for complete executable calls.
[P1] The original Telegram/GitHub Copilot incident was not rerun on the patched branch; the available proof is source-level, targeted runtime-seam tests, and loopback provider tests.

Maintainer options:

Accept the toolUse terminal contract (recommended)
Maintainers can accept that only assistant turns ending in toolUse may execute tools, with provider/plugin adapters responsible for preserving toolUse on complete executable tool-call turns.
Ask for adapter-contract proof first
If compatibility is uncertain, require targeted evidence for supported adapters showing complete tool-call turns still normalize to toolUse while length/error/EOF paths do not.
Keep the original branch superseded
If this replacement lands, close fix(agent-core): ignore truncated assistant tool calls #97092 as superseded so the older fork branch does not remain a competing landing path.

Next step before merge

[P1] Manual maintainer review is needed because the PR has a protected maintainer label and intentionally changes the provider/plugin tool execution contract.

Security
Cleared: No concrete security or supply-chain regression was found; the diff narrows side-effectful tool execution and does not add dependencies, scripts, workflow permissions, or secret handling.

Review details

Best possible solution:

Land this PR after maintainers accept the toolUse-only execution contract and required checks finish; then close the linked issue and superseded contributor PR.

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

Yes. Source inspection gives a high-confidence reproduction path: return a final assistant message with stopReason: "length" and a toolCall block, and current main routes it into executeToolCalls.

Is this the best way to solve the issue?

Yes. Agent-core is the correct final authorization boundary for tool execution and replay, and the PR also updates sibling provider normalization so the fix is not one-provider-only.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against 552ec2b49d4a.

Label changes

Label changes:

add P1: The PR fixes a real agent-core workflow bug where a truncated assistant turn can launch unintended side-effectful subagent work.
add merge-risk: 🚨 compatibility: The diff changes behavior for provider/plugin adapters that return tool-call blocks with non-toolUse terminal reasons.
add rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🌊 off-meta tidepool and patch quality is 🐚 platinum hermit.
add status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Not applicable: The contributor proof gate is not applied because this is a protected maintainer-labeled replacement PR; the PR body provides targeted test evidence rather than live-provider proof.
remove rating: 🦐 gold shrimp: Current PR rating is rating: 🐚 platinum hermit, so this older rating label is no longer current.
remove status: ⏳ waiting on author: Current PR status label is status: 👀 ready for maintainer look.

Label justifications:

P1: The PR fixes a real agent-core workflow bug where a truncated assistant turn can launch unintended side-effectful subagent work.
merge-risk: 🚨 compatibility: The diff changes behavior for provider/plugin adapters that return tool-call blocks with non-toolUse terminal reasons.
rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🌊 off-meta tidepool and patch quality is 🐚 platinum hermit.
status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Not applicable: The contributor proof gate is not applied because this is a protected maintainer-labeled replacement PR; the PR body provides targeted test evidence rather than live-provider proof.

Evidence reviewed

PR surface:

Source +26, Tests +382. Total +408 across 12 files.

View PR surface stats

Area	Files	Added	Removed	Net
Source	6	44	18	+26
Tests	6	396	14	+382
Docs	0	0	0	0
Config	0	0	0	0
Generated	0	0	0	0
Other	0	0	0	0
Total	12	440	32	+408

What I checked:

Current main failure path: Current main exits early only for error or aborted, then executes any toolCall blocks present in the final assistant message, so a length terminal with tool calls can dispatch tools. (packages/agent-core/src/agent-loop.ts:339, 552ec2b49d4a)
PR core guard: PR head adds removeNonExecutableToolCalls, gates dispatch on message.stopReason === "toolUse", and scrubs final messages before message_end and replay. (packages/agent-core/src/agent-loop.ts:83, e080b60311d3)
Regression coverage: The new agent-core test verifies a length assistant message containing sessions_spawn does not execute, does not emit tool_execution_start, and is scrubbed from returned messages, message_end, and next-turn replay context. (packages/agent-core/src/agent-loop.test.ts:168, e080b60311d3)
Sibling normalization coverage: PR head preserves non-executable terminals across DeepSeek DSML recovery, Google MAX_TOKENS, Ollama length, and shared plain-text tool-call wrappers instead of relying only on the core gate. (src/agents/openai-transport-stream.ts:3244, e080b60311d3)
Linked issue and superseded branch: Live GitHub data shows this PR closes the canonical issue and explicitly supersedes the earlier contributor PR, which remains open and unmerged.
Latest release still has old behavior: The latest release tag v2026.6.10 still executes collected tool calls without requiring toolUse, so the fix is not shipped yet and the PR remains necessary. (packages/agent-core/src/agent-loop.ts:306, aa69b12d0086)

Likely related people:

steipete: PR history shows steipete authored the agent-core extraction and later terminal-classifier work in the same loop/terminal area, so they are a strong routing candidate beyond this PR authorship. (role: area introducer and recent agent-runtime contributor; confidence: high; commits: bb46b79d3c14, 0314819f918a; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/agent-loop.test.ts)
vincentkoc: Recent merged work in the same agent-loop terminal lifecycle area includes several commits by vincentkoc, and they merged the aborted-tool-run fix PR. (role: recent adjacent owner/reviewer; confidence: high; commits: 7fe287b0d30d; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/agent-loop.test.ts)
szsip239: szsip239 authored the recent aborted-tool-run fix touching the same agent-core loop files, which is adjacent to this terminal-state execution change. (role: recent adjacent bug-fix contributor; confidence: medium; commits: 7fe287b0d30d; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/agent-loop.test.ts)
fuller-stack-dev: History shows coauthor metadata on terminal-classifier work and recent tool-search/deferred execution changes in the same agent-core test and dispatch area. (role: adjacent tool-execution contributor; confidence: medium; commits: 0314819f918a, 965fa05df38b; files: packages/agent-core/src/agent-loop.ts, packages/agent-core/src/agent-loop.test.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.

steipete · 2026-06-27T02:31:08Z

Land-ready proof for e080b60311d3f0edaa350bc9f67a14be90ee8840:

Local focused regression: node scripts/run-vitest.mjs packages/agent-core/src/agent-loop.test.ts src/plugin-sdk/provider-stream-shared.test.ts src/llm/providers/google-shared.test.ts src/agents/openai-transport-stream.test.ts extensions/google/transport-stream.test.ts extensions/ollama/src/stream.test.ts — 451 tests passed across four shards.
Local static proof: oxfmt --check, oxlint --deny-warnings, and git diff --check passed for all 12 changed files.
Blacksmith Testbox focused proof: provider blacksmith-testbox, lease tbx_01kw3d3p5rc0qffryfpaqqdktm, 451 tests passed.
Blacksmith Testbox full changed gate: provider blacksmith-testbox, lease tbx_01kw3d6f8fhm4wmr4pcjzb82mv; core/core-test/extension/extension-test typechecks, lint, import-cycle, architecture, and storage guards passed.
Hosted CI: run https://github.com/openclaw/openclaw/actions/runs/28275302226 passed on rerun attempt 2. The only first-attempt failure was an unrelated timeout in src/gateway/server-startup-session-migration.test.ts; exact-job rerun passed.
OpenGrep: run https://github.com/openclaw/openclaw/actions/runs/28275302204 passed on attempt 2 after cancelling an anomalous hung first attempt.
Fresh autoreview thread 019f06cd-56d3-7091-879e-a5b8f0e5278e: no findings; patch correct at 0.96 confidence.

Behavioral proof covers completed tool-use dispatch, length/error/abort suppression, provider stop-reason preservation, EOF without a finish reason, error versus done terminal events, message_end, persistence, returned messages, and next-turn replay.

Known proof gaps: none.

steipete · 2026-06-27T02:31:47Z

Merged via squash.

Prepared head SHA: e080b60311d3f0edaa350bc9f67a14be90ee8840
Landed commit: 56259606d15f66c90415baddca9e08b156ef6b33

* fix(agent-core): ignore truncated tool calls Co-authored-by: Galin Iliev <[email protected]> * fix(agent-core): require explicit tool-call terminals --------- Co-authored-by: Galin Iliev <[email protected]>

* fix(agent-core): ignore truncated tool calls Co-authored-by: Galin Iliev <[email protected]> * fix(agent-core): require explicit tool-call terminals --------- Co-authored-by: Galin Iliev <[email protected]> (cherry picked from commit 5625960)

fix(agent-core): ignore truncated tool calls

964ca3e

Co-authored-by: Galin Iliev <[email protected]>

openclaw-barnacle Bot added agents Agent runtime and tooling extensions: google extensions: ollama size: M maintainer Maintainer-authored PR labels Jun 27, 2026

fix(agent-core): require explicit tool-call terminals

e080b60

steipete merged commit 5625960 into main Jun 27, 2026
200 of 212 checks passed

steipete deleted the fix/truncated-tool-call-terminals-97092 branch June 27, 2026 02:31

steipete mentioned this pull request Jun 27, 2026

fix(agent-core): ignore truncated assistant tool calls #97092

Closed

clawsweeper Bot mentioned this pull request Jun 27, 2026

fix(llm): preserve structured tool results as text across providers #97183

Closed

github-actions Bot mentioned this pull request Jun 27, 2026

📡 Upstream Digest — 2026-06-27 06:46 UTC curtismercier/openclaw-mods#1158

Open

clawsweeper Bot mentioned this pull request Jun 30, 2026

fix(openai): preserve complete tool calls from clean streams without finish_reason #98124

Merged

23 tasks

clawsweeper Bot mentioned this pull request Jul 1, 2026

fix: preserve streaming tool_calls when provider omits finish_reason #98575

Closed

clawsweeper Bot mentioned this pull request Jul 8, 2026

fix(tool-call-repair): preserve visible suffix when XML tool text is incomplete #102240

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(agent-core): ignore truncated tool calls#97140

fix(agent-core): ignore truncated tool calls#97140
steipete merged 2 commits into
mainfrom
fix/truncated-tool-call-terminals-97092

steipete commented Jun 27, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jun 27, 2026 •

edited

Loading

Uh oh!

steipete commented Jun 27, 2026

Uh oh!

Uh oh!

steipete commented Jun 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Uh oh!

Conversation

steipete commented Jun 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What Problem This Solves

Why This Change Was Made

User Impact

Evidence

Uh oh!

clawsweeper Bot commented Jun 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

steipete commented Jun 27, 2026

Uh oh!

Uh oh!

steipete commented Jun 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

steipete commented Jun 27, 2026 •

edited

Loading

clawsweeper Bot commented Jun 27, 2026 •

edited

Loading