Coordination

Coordination Skill

Multi-agent coordination for parallel task execution using the two-track architecture:

• plan tool: Create specs from prose/PRDs (interview → scout → elaborate → structure)
• coordinate tool: Execute TASK-XX format specs (validate → dispatch → execute → review)

YOU HAVE THE coordinate AND plan TOOLS AVAILABLE.

Two-Track Architecture

┌─────────────────────────────────────────────────────────────────┐
│  PLAN TOOL (create specs)          COORDINATE TOOL (execute)    │
│  ─────────────────────────         ───────────────────────────  │
│  prose/PRD → interview             TASK-XX spec → validate      │
│           → scout                              → dispatch       │
│           → elaborate                          → workers        │
│           → structure                          → review         │
│           → spec.md ─────────────────────────→                  │
└─────────────────────────────────────────────────────────────────┘

When to Use Each Tool

Use plan when:

• User has prose, PRD, or requirements to convert into a spec
• User says "plan", "create a spec", "help me design"
• Input is NOT already in TASK-XX format

Use coordinate when:

• User has a valid TASK-XX format spec
• User says "implement", "execute", "run", "coordinate"
• The file already contains ## TASK-XX: sections

Coordinate Tool

IMPORTANT: The coordinate tool now requires a valid TASK-XX format spec. It will NOT auto-convert prose or PRDs.

// Execute a valid spec
coordinate({ plan: "./spec.md" })

// If validation fails, the error will tell you to use the plan tool first

Parameters

coordinate({
  // Required
  plan: string,                    // Path to TASK-XX format spec file

  // Common options (all optional)
  agents: number | string[],       // Worker count or array (default: 4)
  reviewCycles: number | false,    // Worker self-review cycles (default: 5, false to disable)
  supervisor: boolean | object,    // Monitor stuck workers (default: true)
  costLimit: number,               // End gracefully at limit (default: $40)
  
  // Async mode
  async: boolean,                  // Run in background (default: false)
  
  // Advanced
  maxFixCycles: number,            // Review/fix iterations (default: 3)
  validate: boolean,               // Run validation after (default: false)
  checkTests: boolean,             // Reviewer checks tests (default: true)
  
  // Model overrides (string sets model, object for full config)
  coordinator: string | { model: string },
  worker: string | { model: string },
  reviewer: string | { model: string },
})

Examples

// Basic - 4 workers
coordinate({ plan: "./spec.md" })

// More workers
coordinate({ plan: "./spec.md", agents: 8 })

// Async mode - returns immediately, use /jobs to monitor
coordinate({ plan: "./spec.md", async: true })

// Disable self-review for speed
coordinate({ plan: "./spec.md", reviewCycles: false })

// Custom models
coordinate({ 
  plan: "./spec.md",
  worker: "claude-sonnet-4-20250514"
})

Pipeline Phases (coordinate)

1. Validate — Checks spec is valid TASK-XX format
2. Dispatch — Assigns tasks to workers respecting dependencies
3. Workers — Execute tasks in parallel with self-review
4. Review — Code reviewer checks all changes
5. Fixes — Workers fix any issues found
6. Complete — Final summary

Plan Tool

Creates TASK-XX specs from prose, PRDs, or requirements through an interactive flow.

// Create spec from a file
plan({ input: "./requirements.md" })

// Create spec from inline text
plan({ input: "Add user authentication with JWT tokens" })

// Refine an existing spec
plan({ continue: "./spec.md" })

// Skip interview (use defaults)
plan({ input: "./prd.md", skipInterview: true })

Parameters

plan({
  // For NEW plans
  input: string,                   // File path or inline text
  
  // For REFINING existing specs
  continue: string,                // Path to existing spec to refine
  
  // Options
  skipInterview: boolean,          // Skip interactive interview (default: false)
  skipScout: boolean,              // Skip codebase analysis (default: false)
  maxInterviewRounds: number,      // Limit interview rounds (default: 5 new, 3 refine)
  output: string,                  // Where to save spec (default: auto-named in specs/)
  format: "markdown" | "json",     // Output format (default: markdown)
  
  // Model overrides
  model: string,                   // Model for elaboration (default: frontier)
  scoutModel: string,              // Model for scout (default: fast)
})

Pipeline Phases (plan)

1. Interview — Gather requirements interactively (can skip with Ctrl+D or timeout)
2. Scout — Analyze codebase structure and patterns
3. Elaborate — Expand requirements with codebase context
4. Structure — Convert to TASK-XX format spec
5. Output — Save spec to file

Required Spec Format (TASK-XX)

The coordinate tool requires this format:

# Project Title

Description of what we're building.

## TASK-01: Create auth types
Priority: P1
Files: src/types.ts (create)
Depends on: none
Acceptance: AuthCredentials and User interfaces exported

## TASK-02: Implement login endpoint
Priority: P1
Files: src/routes/auth.ts (create)
Depends on: TASK-01
Acceptance: POST /login returns JWT, tests pass

Required fields per task:

• ## TASK-XX: Title — Task ID and title
• Priority: P0|P1|P2|P3 — Execution priority
• Files: — Files to create/modify
• Depends on: — Dependencies (or "none")
• Acceptance: — Testable completion criteria

Typical Workflow

// 1. User has prose requirements
plan({ input: "./requirements.md" })
// → Creates specs/requirements-spec.md

// 2. Execute the generated spec
coordinate({ plan: "./specs/requirements-spec.md" })

Or if user already has a valid spec:

// Execute directly
coordinate({ plan: "./valid-spec.md" })

Monitoring

During execution: TUI shows live progress with pipeline status, workers, events.

Async mode: Use /jobs command to open full dashboard:

• Pipeline status
• Task queue with dependencies
• Worker status with cost/duration
• Event stream

Output Retrieval

When worker output is truncated, use coord_output:

coord_output({ ids: ["worker-04ea"] })
coord_output({ ids: ["scout", "planner", "review"] })
coord_output({ ids: ["worker-04ea"], format: "stripped" })  // No ANSI codes

DO NOT

• Don't pass prose to coordinate — Use plan tool first to create a spec
• Don't manually write TASK-XX specs — Let plan tool generate them
• Don't skip validation errors — Fix the spec or use plan to create a valid one

Error Handling

If coordinate returns a validation error like:

Invalid spec format. The coordinate tool requires a valid TASK-XX format spec.

Errors:
- No valid TASK-XX format tasks found

Use the plan tool first:

plan({ input: "./your-file.md" })
// Then coordinate the output
coordinate({ plan: "./specs/your-file-spec.md" })