jankneumann/fix-scrub

Fix Scrub

Consume the bug-scrub report and apply fixes with clean separation from the diagnostic phase. Classifies findings into three tiers (auto/agent/manual), applies fixes, verifies quality, and commits.

Arguments

$ARGUMENTS - Optional flags:

• --report <path> (default: docs/bug-scrub/bug-scrub-report.json)
• --tier <list> (comma-separated; default: auto,agent; values: auto, agent, manual)
• --severity <level> (minimum severity; default: medium)
• --dry-run (plan fixes without applying — skips branch creation)
• --max-agent-fixes <N> (default: 10)
• --worktree (create an isolated git worktree for the fix-scrub branch; auto-enabled when running inside an existing worktree)

Script Location

Scripts live in <agent-skills-dir>/fix-scrub/scripts/. Each agent runtime substitutes <agent-skills-dir> with its config directory:

• Claude: .claude/skills
• Codex: .codex/skills
• Gemini: .gemini/skills

If scripts are missing, run skills/install.sh to sync them from the canonical skills/ source.

Prerequisites

• Bug-scrub report must exist (run /bug-scrub first)
• Python 3.11+
• ruff for auto-fixes

Steps

0. Branch Setup

Skip this step if --dry-run is active.

Create an isolated branch for fix-scrub changes. All fixes go through PR review before reaching main.

# Pull latest main
git checkout main
git pull origin main

# Determine branch name (date-based, with collision suffix)
BRANCH_DATE=$(date +%Y-%m-%d)
BRANCH_NAME="fix-scrub/${BRANCH_DATE}"
SUFFIX=1
while git show-ref --verify --quiet "refs/heads/${BRANCH_NAME}"; do
  SUFFIX=$((SUFFIX + 1))
  BRANCH_NAME="fix-scrub/${BRANCH_DATE}-${SUFFIX}"
done

# Create and switch to the fix-scrub branch
git checkout -b "$BRANCH_NAME"
echo "Created branch: $BRANCH_NAME"

Optional: Worktree Isolation

Create a worktree when --worktree is passed or when already inside a git worktree (auto-detection). This prevents fix-scrub from modifying an active implementation worktree.

# Detect if worktree isolation is needed
eval "$(python3 "<skill-base-dir>/../worktree/scripts/worktree.py" detect)"

if [[ "$IN_WORKTREE" == "true" ]] || [[ "<--worktree flag passed>" == "true" ]]; then
  # Agent-id for worktree disambiguation
  FIX_AGENT_ID="${AGENT_ID:-fix-$(date +%s)}"

  # Setup fix-scrub worktree (creates .git-worktrees/fix-scrub/<agent-id>/)
  eval "$(python3 "<skill-base-dir>/../worktree/scripts/worktree.py" setup "${BRANCH_DATE}" --branch "${BRANCH_NAME}" --prefix fix-scrub --agent-id "${FIX_AGENT_ID}")"
  cd "$WORKTREE_PATH"
  echo "Working directory: $(pwd)"
fi

After the fix-scrub is complete and changes are pushed, tear down the worktree:

# Teardown fix-scrub worktree
python3 "<skill-base-dir>/../worktree/scripts/worktree.py" teardown "${BRANCH_DATE}" --prefix fix-scrub --agent-id "${FIX_AGENT_ID}"

1. Load Report and Classify

python3 <agent-skills-dir>/fix-scrub/scripts/main.py \
  --report <report-path> \
  --tier <tiers> \
  --severity <level> \
  --dry-run

Review the dry-run output before applying.

2. Apply Fixes

python3 <agent-skills-dir>/fix-scrub/scripts/main.py \
  --report <report-path> \
  --tier <tiers> \
  --severity <level> \
  --max-agent-fixes <N>

This will:

1. Apply auto-fixes (ruff --fix)
2. Generate agent-fix prompts to docs/bug-scrub/agent-fix-prompts.json
3. Track OpenSpec task completions

3. Dispatch Agent Fixes (if applicable)

If agent-fix prompts were generated, resolve the implementer archetype and dispatch as parallel Task() agents:

import json
from src.agents_config import load_archetypes_config, resolve_model

archetypes = load_archetypes_config()
implementer = archetypes.get("implementer")
resolved_model = resolve_model(implementer, {}) if implementer else "sonnet"

with open("docs/bug-scrub/agent-fix-prompts.json") as f:
    prompts = json.load(f)

# Launch parallel agents — one per file group
for entry in prompts:
    Task(
        subagent_type="general-purpose",
        model=resolved_model,  # archetype: implementer (sonnet, or opus if escalated)
        description=f"Fix issues in {entry['file']}",
        prompt=entry["prompt"],
        run_in_background=True,
    )

Wait for all agents to complete, then verify.

4. Quality Verification

After all fixes (auto + agent) are applied, the orchestrator runs pytest, mypy, ruff, and openspec validate. If regressions are detected, review before committing.

5. Commit

git add .
git commit -m "$(cat <<'EOF'
fix(scrub): apply <N> fixes from bug-scrub report

Auto-fixes: <count> (ruff)
Agent-fixes: <count> (mypy, markers, deferred)
Manual-only: <count> (reported, not fixed)

Source report: <report-path>

Co-Authored-By: Claude <[email protected]>
EOF
)"

6. Push and Create PR

Skip if no fixes were applied (dry-run or all findings were manual-only).

# Push the fix-scrub branch
git push -u origin "$BRANCH_NAME"

# Create PR with fix-scrub report as body
gh pr create \
  --title "fix(scrub): apply fixes from bug-scrub report $(date +%Y-%m-%d)" \
  --body "$(cat <<'EOF'
## Summary

Automated fix-scrub remediation from bug-scrub report.

### Fix Summary
<!-- Paste from docs/bug-scrub/fix-scrub-report.md -->
- Auto-fixes: <count> (ruff)
- Agent-fixes: <count> (mypy, markers, deferred)
- Manual-only: <count> (reported, not fixed)

### Quality Checks
- [ ] pytest passing
- [ ] mypy passing
- [ ] ruff passing
- [ ] openspec validate passing

Source report: `docs/bug-scrub/bug-scrub-report.json`

---
🤖 Generated with Claude Code
EOF
)"

7. Review Summary

Check docs/bug-scrub/fix-scrub-report.md for the full summary including:

• Fixes applied by tier
• Files changed
• OpenSpec tasks marked as completed
• Quality check results
• Manual action items requiring human attention

Fixability Tiers

| Tier | Criteria | Action | |------|----------|--------| | auto | ruff with fixable rules | ruff check --fix | | agent | mypy type errors, markers with 10+ chars context, deferred items with proposed fix | Task() agent with file scope | | manual | architecture, security, deferred without fix, markers with insufficient context | Reported only |

Quality Checks

python3 -m pytest <agent-skills-dir>/fix-scrub/tests -q