Execute Phase Skill

Execute a planned phase by spawning fresh executor subagents for each task group.

Usage

/execute-phase           Execute the current planned phase
/execute-phase 2         Execute a specific phase by number

Behavior

1. •Read .planning/PHASES.md to identify the target phase
2. •Verify the phase has status planned (has a PLAN.md)
3. •Read phases/NN-name/PLAN.md
4. •Emit phase_executing event:
   bash

   curl -s -X POST http://localhost:4011/event \
     -H "Content-Type: application/json" \
     -d '{"type":"phase_executing","questId":"<id>","phaseId":"<id>","phaseName":"<name>"}'
   

5. •For each task group, spawn phase-executor agent via Task tool with:
   • •The PLAN.md content
   • •Fresh 200k context (this is the key context rot prevention)
   • •Any diagnosis from previous failed verification (if retrying)
6. •Collect structured results from executor
7. •Update PHASES.md status to executed
8. •Update STATE.md with execution results (commits, files changed, issues)

Orchestrator Rules

The orchestrating skill stays lean (under 30% context):

• •Only read PLAN.md and executor results
• •Never read full source files directly
• •Route between executors, don't execute yourself
• •Log decisions to STATE.md

Deviation Handling

Executors follow deviation rules from the runbook:

• •Auto-fix: bugs, missing deps, blocking issues
• •Checkpoint: architectural changes, scope creep — executor stops, reports back

If an executor checkpoints, present the issue to the user and wait for direction before continuing.

Prerequisites

• •Active quest with planned phase
• •Phase status must be planned (or retrying for re-execution)

Quick Reference

• •Runbook: runbook/quest-workflow.md
• •Executor agent: claude-config/agents/phase-executor.md