tools-only SUPERVISOR MODE (Multi-Agent QA)
Architecture Principle
Context Isolation: Each phase operates with fresh focus on its specific task.
Token Multiplier: Multi-phase operations consume ~4x baseline. For complex tasks (>5 files), watch context budget.
Workflow
1. DEVELOPER PHASE
- Impact Analysis: Which files are affected?
- No-Touch Zones Check: Review your project's protected files
- Check CLAUDE.md for project-specific no-touch zones
- Common patterns: auth logic, core business logic, production config
- Generate Code with justification
2. REVIEWER PHASE
Run Quality Gates:
npx tsc --noEmit
npm run build
npm run test
- On failure: Return to Phase 1 (max 3 iterations)
- On success: Continue to Phase 3
Anti-Telephone-Game: When returning to Phase 1, pass the original error directly, don't paraphrase:
BAD: "The test somehow didn't work"
GOOD: "FAIL src/utils/parser.test.ts:42 - Expected 'string' but got 'undefined'"
3. SYNTHESIS PHASE
- Error Propagation Check: Can errors from Phase 1/2 affect downstream?
- Rollback Capability: Is git reset HEAD~1 sufficient?
- Summarize changes with file list
4. COMPLETION
- List affected files with line numbers
- Propose commit (only after confirmation!)
Failure Mode Mitigations
| Failure | Mitigation |
|---|---|
| Supervisor Bottleneck | Only distilled summaries between phases |
| Divergence | TTL: Max 3 iterations, then user decision |
| Error Propagation | Validate outputs before next phase starts |
Task:
$ARGUMENTS
IMPORTANT: This mode iterates until all Quality Gates pass or 3 attempts are exhausted.
Origin
Originally developed for fabrikIQ - AI-powered manufacturing data analysis.