| name | fix-runner |
| description | Use after evaluator flags failures to run .specify/gates/run-gates.sh and apply minimal fixes, looping until gates are green. Does not invent new gates. |
fix-runner
When: evaluator reported a FAIL, or gates are suspected red, after implement-feature has touched code for the active feature.
Template: none — this skill edits existing source/test files to make gates pass, it does not produce a spec-style artifact.
Output: minimal code/test edits; a final gate run that is green (or an explicit report of what remains red and why).
Steps:
- Run
.specify/gates/run-gates.sh from the repo root.
- If it reports "no checks detected", there is nothing for this skill to fix at the gate level — fall back to whatever evaluator flagged directly (e.g. an unmet acceptance criterion) and fix that instead.
- For each failing gate, read its output and apply the MINIMAL fix that addresses the root cause — do not refactor unrelated code, do not add new lint rules, scripts, or gates beyond what
run-gates.sh already runs (it picks up npm run lint/test/build or make lint/test automatically; do not hand-roll alternatives).
- Re-run
.specify/gates/run-gates.sh after each fix attempt.
- Loop steps 3-4 until the script exits 0 ("all gates passed"), or until a fix attempt would require scope beyond the current task (e.g. a missing approved spec/plan) — in that case stop and report the blocker instead of working around it.
- Do not touch
spec.md, plan.md, or tasks.md frontmatter/approval state from this skill.
Model routing: when a gate fix is handed to a subagent (via dispatching-parallel-agents) instead of run inline, size the model to the fix's difficulty — do NOT inherit the session model by default. A single-file mechanical gate fix (lint, formatting, a missing import) is cheapest tier; a fix that needs cross-file reasoning is standard tier. See the routing-minimum-capable-model skill.
Next: evaluator, to re-confirm acceptance criteria now that gates are green; loop continues via implement-and-evaluate.
