Bug Health Check (Inline Orchestration)

You ARE the orchestrator. Execute this workflow directly without spawning a separate orchestrator agent.

Workflow Overview

Detection → Validate → Fix by Priority → Verify → Repeat if needed

Max iterations: 3 Priorities: critical → high → medium → low

---

Phase 1: Pre-flight

1. Setup directories:

   mkdir -p .tmp/current/{plans,changes,backups}
   

2. Validate environment:

   • Check

     package.json

     exists
   • Check

     type-check

     and

     build

     scripts exist

3. Initialize TodoWrite:

   [
     { "content": "Bug detection", "status": "in_progress", "activeForm": "Detecting bugs" },
     { "content": "Fix critical bugs", "status": "pending", "activeForm": "Fixing critical bugs" },
     { "content": "Fix high priority bugs", "status": "pending", "activeForm": "Fixing high bugs" },
     {
       "content": "Fix medium priority bugs",
       "status": "pending",
       "activeForm": "Fixing medium bugs"
     },
     { "content": "Fix low priority bugs", "status": "pending", "activeForm": "Fixing low bugs" },
     { "content": "Verification scan", "status": "pending", "activeForm": "Verifying fixes" }
   ]
   

---

Phase 2: Detection

Invoke bug-hunter via Task tool:

subagent_type: "bug-hunter"
description: "Detect all bugs"
prompt: |
  Scan the entire codebase for bugs:
  - Run type-check and build
  - Check for security vulnerabilities
  - Find dead code and debug statements
  - Categorize by priority (critical/high/medium/low)

  Generate: bug-hunting-report.md

  Return summary with bug counts per priority.

After bug-hunter returns:

1. Read

   bug-hunting-report.md

2. Parse bug counts by priority
3. If zero bugs → skip to Final Summary
4. Update TodoWrite: mark detection complete

---

Phase 3: Quality Gate (Detection)

Run inline validation:

pnpm type-check
pnpm build

• If both pass → proceed to fixing
• If fail → report to user, exit

---

Phase 4: Fixing Loop

For each priority (critical → high → medium → low):

1. Check if bugs exist for this priority

   • If zero → skip to next priority

2. Update TodoWrite: mark current priority in_progress

3. Invoke bug-fixer via Task tool:

   subagent_type: "bug-fixer"
   description: "Fix {priority} bugs"
   prompt: |
     Read bug-hunting-report.md and fix all {priority} priority bugs.
   
     For each bug:
     1. Backup file before editing
     2. Implement fix
     3. Log change to .tmp/current/changes/bug-changes.json
   
     Generate/update: bug-fixes-implemented.md
   
     Return: count of fixed bugs, count of failed fixes.
   

4. Quality Gate (inline):

   pnpm type-check
   pnpm build
   

   • If FAIL → report error, suggest rollback, exit
   • If PASS → continue

5. Update TodoWrite: mark priority complete

6. Repeat for next priority

---

Phase 5: Verification

After all priorities fixed:

1. Update TodoWrite: mark verification in_progress

2. Invoke bug-hunter (verification mode):

   subagent_type: "bug-hunter"
   description: "Verification scan"
   prompt: |
     Re-scan codebase after fixes.
     Compare with previous bug-hunting-report.md.
   
     Report:
     - Bugs fixed (count)
     - Bugs remaining (count)
     - New bugs introduced (count)
   

3. Decision:

   • If bugs_remaining == 0 → Final Summary
   • If iteration < 3 AND bugs_remaining > 0 → Go to Phase 2
   • If iteration >= 3 → Final Summary with remaining bugs

---

Phase 6: Final Summary

Generate summary for user:

## Bug Health Check Complete

**Iterations**: {count}/3
**Status**: {SUCCESS/PARTIAL}

### Results

- Found: {total} bugs
- Fixed: {fixed} ({percentage}%)
- Remaining: {remaining}

### By Priority

- Critical: {fixed}/{total}
- High: {fixed}/{total}
- Medium: {fixed}/{total}
- Low: {fixed}/{total}

### Validation

- Type Check: {status}
- Build: {status}

### Artifacts

- Detection: `bug-hunting-report.md`
- Fixes: `bug-fixes-implemented.md`

---

Error Handling

If quality gate fails:

Rollback available: .tmp/current/changes/bug-changes.json

To rollback:
1. Read changes log
2. Restore files from .tmp/current/backups/
3. Re-run workflow

If worker fails:

• Report error to user
• Suggest manual intervention
• Exit workflow

---

Key Differences from Old Approach

Old (Orchestrator Agent)	New (Inline Skill)
9+ orchestrator calls	0 orchestrator calls
~1400 lines (cmd + agent)	~150 lines
Context reload each call	Single session context
Plan files for each phase	Direct execution
~10,000+ tokens overhead	~500 tokens

---

Worker Prompts

See

references/worker-prompts.md

for detailed prompts.