coordination-playbook

Project Coordinator Engineer: Coordination Playbook

Output format, death loop prevention, error handling, preferred patterns, anti-rationalizations, blocker criteria.

---

Output Format — Implementation Schema

**Phase 1: ANALYZE** — Review scope, identify agents, map dependencies, find parallel opportunities. **Phase 2: PLAN** — TodoWrite with assignments, execution sequence, communication file plan. **Phase 3: COORDINATE** — Spawn agents, monitor progress, update STATUS.md, detect death loops. **Phase 4: INTEGRATE** — Coordinate handoffs, validate cross-agent consistency, update PROGRESS.md.

**Final Output**:

═══════════════════════════════════════════════════════════════
 PROJECT COORDINATION COMPLETE
 Agents Coordinated: {count}
 Tasks Completed: {count}
 Parallel Execution: {count} concurrent streams
 Documentation: STATUS.md, PROGRESS.md, BLOCKERS.md ({count})
 Death Loop Prevention: 3 max enforced, compilation-first ✓, context {%}%
═══════════════════════════════════════════════════════════════

Death Loop Prevention

3-Attempt Maximum (Hard Limit)

After 3 failures per agent per task: STOP, document in BLOCKERS.md, analyze root cause, change strategy or escalate.

Compilation-First Protocol

For code-modifying agents: (1) modify code, (2) verify build, (3) verify tests, (4) ONLY then lint/format. Prevents build-lint-break cycles.

Context Window Monitoring

At 70%: summarize to PROGRESS.md, archive logs, clear non-essential history, continue with summary.

Identical Error Detection

Same error 3+ times → STOP all agents, document in BLOCKERS.md, require root cause before retry.

See [death-loop-prevention.md](death-loop-prevention.md) for comprehensive patterns.

Error Handling

| Error | Fix | |-------|-----| | Death loop (3+ same failures) | STOP, BLOCKERS.md, root cause before retry | | File domain conflict | Serialize or partition file domains | | Context overflow | PROGRESS.md at 70%, archive, clear |

See [error-catalog.md](error-catalog.md) for full catalog.

Preferred Patterns

| Signal | Fix | |--------|-----| | Agent fails repeatedly, coordinator re-spawns | Enforce 3-attempt max, then reassess | | Code → lint → compilation breaks | Verify compilation BEFORE linting | | Two agents modifying same file | Serialize same-file modifications |

See [preferred-patterns.md](preferred-patterns.md) for full catalog.

Anti-Rationalization

| Rationalization | Required Action | |----------------|-----------------| | "4th attempt might work" | STOP at 3, analyze root cause | | "Linting is quick, run it first" | Verify compilation first | | "Agents can coordinate file changes" | Enforce non-overlapping domains | | "Context still has space" | Summarize at 70% | | "Same error but different line number" | Treat as identical for loop detection |

Blocker Criteria

Resolve routine issues from repository evidence and existing authorization. When you are sure, decide, record the assumption in your report, and continue; ask Jev when genuinely unsure. Ask only when these situations leave a material decision uncovered:

| Situation | Ask This | |-----------|----------| | Agent fails 3 times | "Change strategy or escalate?" | | Circular dependencies | "How to break cycle?" | | All agents blocked | "Which dependency first?" | | Context approaching 90% | "Compact and continue?" |

Decide Explicitly Before Acting On

• Strategy after 3 failed attempts
• Breaking circular dependencies
• File modification conflict resolution
• Continuing at 90% context