Repository navigation
fix(fusion): stop grounded agents on incomplete tests - #1030
Conversation
|
🦞👀 Pull request received. I will update this pull request when review starts. ClawSweeper review completeClawSweeper finished reviewing this revision. The review result is being finalized. |
|
Codex review: needs maintainer review before merge. What this changesStop grounded and checked Fusion workers when visible tests return an incomplete result. Example: An edit introduces an infinite loop and the visible-test watchdog terminates execution.
Review scores
ProductKind: Bug fix · Worth it: Yes · Fix scope: Complete Merge readiness✅ Ready for maintainer review Keep open: this maintainer-authored PR fixes a grounded-worker gap still present on main; no actionable introduced defect was found. Priority: P2 Before mergeNone. FindingsNone. Agent review detailsHow this fits togetherThe experimental local-repair runner sends model actions to a sandboxed grounded worker, which edits source and runs visible checks before returning candidates for independent grading. flowchart TD
A[Local benchmark runner] --> B[Grounded worker]
B --> C[Model action]
C --> D[Sandbox visible checks]
D --> E[Complete result continues]
D --> F[Incomplete result throws]
E --> G[Candidate grading]
F --> H[Sandbox cleanup]
Technical reviewBest possible solution: Land the shared visible-check guard in the grounded worker while preserving repairable complete failures and existing cleanup. Do we have a high-confidence way to reproduce the issue? Main consumes passed without checking valid at both grounded visible-test call sites; the supplied watchdog regression exercises the resulting incomplete-run path, but no target code was executed during this read-only review. Is this the best way to solve the issue? The two-call-site helper is the smallest complete fix: it uses the checker's existing validity verdict and bypasses recoverable-action handling before existing cleanup. AGENTS.md: found and applied where relevant. Codex review notes: model internal, reasoning medium; reviewed against 1e4104fb47f7. Provenance checked
TestingProof path: in-process harness. SecurityNone. EvidenceWhat I checked:
Preflight sourceSource unknown. Source comparison at the observed PR snapshot only. Completion is self-reported; body, policy, tests, execution, and report bytes are not verified. LabelsLabel changes:
Label justifications:
Rating scale6/6 🦀 challenger crab · 5/6 🦞 diamond lobster · 4/6 🐚 platinum hermit · 3/6 🦐 gold shrimp · 2/6 🦪 silver shellfish · 1/6 🧂 unranked krab. Overall follows the weaker of proof and patch quality; ✨ marks media proof (a screenshot, video, or linked artifact) that directly shows the changed behavior. WorkflowClawSweeper edits this one comment on every review. Comment Reviewed November 13, 2026, 8:34 AM ET / 13:34 UTC. |
Summary
Stop grounded and checked repair workers when visible tests do not produce a complete result. Previously an incomplete check after an edit or submission could enter another model turn or final grading. Both paths now raise the existing finite incomplete_test_run error and close the sandbox.
Complete assertion failures remain repairable.
Verification
Deployment and risk
No deployment or default promotion. Incomplete execution now stops the run instead of being treated as model feedback; no retry or model fallback is added.