swebench-verified

Verification phase for SWE-bench instances. Runs the full test suite on the craft patch, classifies regressions against the captured fail-on-base baseline, and emits a verdict plus a re-entry route for the outer loop.

Implementation phase for SWE-bench instances. You generate the patch from the recon handoff; a codex subagent challenges it; the gate arbitrates. Loops until FAIL_TO_PASS tests pass. On audit re-entry, narrows a fix that regressed. Stops when gate is green or the attempt budget is exhausted.

Read-only diagnosis phase for SWE-bench instances. Walks the codebase, reproduces the failure, and produces a structured hypothesis handoff for /craft. Single diagnostician; correction comes from the audit→recon outer loop, not a parallel blind. No edits, no external access.