Repair Loop

Signal, not vibes: each stage narrows the suspect set with evidence, and selection is by execution, not by asking the model which patch looks right.

The loop

failing tests → extract signals → localize → sample K patches
             → select by running the real suite → blast-radius gate → commit

1. Signals — traceback paths, exception types, failing-test identity, plus a flake-tolerance pass (a test that also fails on the unpatched baseline is evidence, not a verdict).
2. Localization — evidence-ranked suspects: traceback frames → symbol resolution → call-graph blast radius (CIE-Integration when CIE is present; the built-in ToolBackends otherwise). Long functions get a second pass at statement granularity — Statement-Level-Graph.
3. K sampled attempts — K independent patch attempts run in parallel (adaptive, rate-limit-aware worker pool).
4. Selection by execution — each candidate is applied, the real suite is run, the tree is restored. The winner is empirically better, not the model's favorite. (sandbox.py::_purge_pycache exists because a write→retest cycle could otherwise evaluate a stale .pyc and silently corrupt this selection — root-caused and fixed; see Parallel-Execution.)
5. Blast-radius gate — a static check that rejects a winning patch if it changes or removes a function/method signature while a caller outside the patched file still depends on it — the rejection is fed back into the next round's prompt rather than dropped.
6. Commit — each accepted edit is auto-committed with a descriptive message (Auto-Commit-Messages); any round that makes things worse is reverted.

Verdicts, not booleans

Every run's outcome is one of 7 checkpointed verdicts: passed / failed / partial / timeout / lint_error / crashed / skipped — plus the full phase-by-phase history (Checkpointing-and-Resumability).

Where it runs

• Local — direct subprocess execution with per-command timeouts.
• Docker — per-project persistent container (CLI-CI-Native); forge degrades gracefully with FORGE_DISABLE_DOCKER_TESTS=1.
• Bootstrap-Gate — before any of this runs against a cold clone, the environment is proven capable of running at least one test.

Research grounding

Execution-guided selection and critique-gated repairs trace to SWE-agent's agent-computer interface work (arXiv:2405.15793), Agentless's localize-repair decompose (arXiv:2408.03310), and self-refine loops (arXiv:2303.17651) — the full survey with what forge adopts and what it deliberately does not is in Execution-Guided-Repair and Critic-Verification-Gate.