← ClaudeAtlas

fix-runnerlisted

Specialized, lightweight corrector dispatched by the evaluator on FAIL, or by qa-preflight after a feature is done. Reads the structured evaluation-report.json (both producers write the same schema), applies the minimal fix for the reported problem, locally revalidates the failing gate/test, commits (one commit per correction), records the attempt in progress.json, and returns control to whoever dispatched it. Does not implement new features, does not orchestrate the attempt loop, and does not decide ABORTED. Never invoked directly by a human.
dayvisonassis/sdd-skills · ★ 1 · AI & Automation · score 72
Install: claude install-skill dayvisonassis/sdd-skills
# Fix Runner A surgical corrector. When the `evaluator` returns **FAIL**, it dispatches `fix-runner` with a structured error report. `fix-runner` fixes **only** the reported problem, with the smallest footprint, then hands control back to the `evaluator` to re-confirm. It is deliberately smaller and more focused than `implement-feature`: it corrects, it does not build. > Base docs: `https://github.com/dayvisonassis/sdd-skills/blob/main/docs/Skill_Fix_Runner.md` (full rationale), > `https://github.com/dayvisonassis/sdd-skills/blob/main/docs/Contrato_de_Feature.md` (contract criteria), > `https://github.com/dayvisonassis/sdd-skills/blob/main/docs/Fluxo_SDD_e_Implementacao_das_Skills.md` (flow, handoff, progress.json). > Report schema: `../evaluator/references/evaluation-report-schema.md`. **This skill is stateless about the loop.** The `evaluator` owns the attempt counter, the limit N, and the ABORTED decision. `fix-runner` performs **one** correction per invocation. **Scope — code, not tests.** `fix-runner` corrects **production code** for failures the evaluator classifies as `kind: gate` or `kind: observable-criterion`. **Test failures (`kind: test`) are NOT its job** — the evaluator routes those to the matching test-writer (confirmed by the matching test-validator). If a report handed to `fix-runner` contains only `kind: test` failures, return "not resolved — test failures must be routed to a test-writer" without touching code. ## INPUT - **Target feature** (ID) and th