sh-fix

name	sh-fix
description	Fix QA issues: reads qa-issues.json written by superharness qa, fixes each issue using TDD, then re-runs QA to verify.

Fix QA Issues

Read qa-issues.json and fix pending issues using TDD, then re-run QA to verify fixes.

Prerequisites

superharness qa (CLI or /superharness:sh-qa skill) must have already run and written qa-issues.json to the current task directory.

Process

Step 1: Load Issues

cat .superharness/tasks/{task}/qa-issues.json

Filter for status: "pending" issues. Sort by severity: critical → major → minor → suggestion.

If no pending issues, report "No pending QA issues" and stop.

Step 2: Fix Each Issue (TDD)

For each pending issue, in severity order:

1. Read the issue: file, line, message, fix_hint, category
2. Write a failing test that reproduces the issue (RED phase)
3. Fix the code to make the test pass (GREEN phase)
4. Verify: run the test suite to confirm fix doesn't break anything
5. Update issue status in qa-issues.json to "fixed" and increment fix_round
6. Commit the fix with a descriptive message

Log trace events:

echo '{"ts":"...","phase":"fix","event":"fix_start","detail":"修复中: {issue.id}"}' >> .superharness/tasks/{task}/trace.jsonl
echo '{"ts":"...","phase":"fix","event":"fix_complete","detail":"已修复: {issue.id}"}' >> .superharness/tasks/{task}/trace.jsonl

Step 3: Re-run QA

After fixing all critical and major issues, re-run QA:

superharness qa --task .superharness/tasks/{task}

This performs incremental verification:

• Re-runs all deterministic checks
• Only re-evaluates rubric dimensions related to fixed issues

Step 4: Check Convergence

If new issues appear or regressions are detected:

• Regression issues (previously fixed, now failing again) get severity bumped one level
• If fix_round exceeds max_fix_rounds (default 3), mark remaining issues as "escalated"
• Report escalated issues to user for manual intervention

Anti-Oscillation Rules

• Max 3 fix rounds per issue (configurable in config.yaml)
• Regression detection: same issue id reappearing after fix → severity bumped
• Two consecutive regressions → auto-escalate to human
• Never fix suggestion-level issues automatically (only if user requests)

Red Flags

• Never skip the failing test step — even for "obvious" fixes
• Never mark an issue as fixed without running the test suite
• Never exceed max_fix_rounds without escalating
• Never fix issues in a different order than severity (critical first)