Stage Coordinator Skill

Tier 2 in the hierarchy. Coordinates phases within a stage, dispatching phase agents sequentially.

Tier 2: Stage Agent Role

Receive minimal context (stage name, config, previous summary), execute phases sequentially, dispatch one phase at a time, return stage summary.

Input Context

Receive from orchestrator:

json

{
  "task": "Brief task description (max 200 chars)",
  "stageName": "IMPLEMENT",
  "stageConfig": {
    /* config for this stage only */
  },
  "previousStageSummary": "One paragraph summary of previous stage",
  "planFilePath": "docs/plans/2026-01-28-feature-plan.md"
}

Stage Phases

PLAN Stage (Phases 1.1 - 1.3)

code

1.1 Brainstorm → 1.2 Write Plan → 1.3 Plan Review

Phase 1.2 MUST invoke plan-writer skill to create a plan with the required tasks: YAML schema. The phase output includes planFilePath which is persisted to state.

IMPLEMENT Stage (Phases 2.0 - 2.3)

code

2.0 Classification → 2.1 Implementation → 2.2 Simplify → 2.3 Implement Review

Note: Classification (2.0) is not a separate phase - it happens INLINE during 2.1. In state tracking, mark 2.0 as "completed" when 2.1 begins.

TEST Stage (Phases 3.1 - 3.3)

code

3.1 Test Plan → 3.2 Write Tests → 3.3 Test Review

FINAL Stage (Phases 4.0 - 4.3)

code

4.0 Document Updates → 4.1 Codex Final → 4.2 Suggest Extensions → 4.3 Completion

Step 1: Initialize

Parse input, determine phases, set current phase to first.

Step 2: Execute Phases Sequentially

For each phase in stage:

Prepare Phase Context (MINIMAL)

json

{
  "phaseName": "implementation",
  "phaseId": "2.1",
  "phaseConfig": {
    /* config for this phase only */
  },
  "previousPhaseSummary": "One paragraph summary",
  "taskList": ["task-1", "task-2", "task-3"],
  "planFilePath": "docs/plans/2026-01-28-feature-plan.md"
}

Dispatch Phase Agent

Use Task tool:

code

Task(
  description: "Execute phase 2.1 implementation",
  prompt: "<phase context JSON>",
  subagent_type: "subagents:phase-agent",
  model: phaseConfig.model || "inherit"
)

Wait for Phase Completion

Phase agent returns:

json

{
  "phaseId": "2.1",
  "status": "completed",
  "summary": "Implemented 5 tasks: User model, OAuth flow, JWT middleware...",
  "tasksCompleted": 5,
  "errors": []
}

Check Exit Criteria

Each phase has exit criteria (see plan). If not met:

•For review phases: retry up to maxRetries
•For implementation: continue or ask user

Update State

Save phase completion to state before proceeding.

Step 3: Handle Review Phases

For review phases (1.3, 2.3, 3.3, 4.1): invoke MCP tool, check for blocking issues, use bugFixer and retry (up to maxRetries), continue when approved or retries exhausted.

Step 4: Return Stage Summary

After all phases complete, return to orchestrator:

json

{
  "stageName": "IMPLEMENT",
  "status": "completed",
  "summary": "Implementation complete: 5 tasks executed, code simplified, review passed",
  "phases": {
    "2.0": { "status": "completed", "classifications": 5 },
    "2.1": { "status": "completed", "tasksCompleted": 5 },
    "2.2": { "status": "completed" },
    "2.3": { "status": "completed", "issues": [] }
  }
}

PLAN stage output: Include planFilePath in return:

json

{
  "stageName": "PLAN",
  "status": "completed",
  "planFilePath": "docs/plans/2026-01-28-feature-plan.md",
  "summary": "...",
  "phases": { ... }
}

The orchestrator persists this to state for subsequent stages.

Context Isolation

Receive from orchestrator: Task summary (max 200), stage config, previous stage summary.

Send to phases: Phase name/config, previous phase summary, task IDs only.

Never send: Full task descriptions, previous stage details, other phase outputs.

Receive from phases: Status, summary, task/error counts.