Evolve Task

Step 2: TRIAGE (Fill Information Gaps)

• Map First: Must read architecture diagrams under codemaps/ or docs/SYSTEM_PANORAMA.md to accurately assess modification risk. Output:
• Goal (one sentence)
• Problem (reproducible description)
• Evidence (files/commands/logs)
• Risk level (low/medium/high)

Step 3: Delegate Explorer (Evidence Collection)

• Let Explorer sub-agent output: Evidence list / Repro / Hypotheses(<=3)
• Performance Evaluation: After task completion, evaluate Explorer performance and write to .state/.verdict.json. Format must strictly follow @.principles/schemas/agent_verdict_schema.json.

Step 4: Delegate Diagnostician (Root Cause)

• Let Diagnostician output: Proximal cause / Root cause / 5 Whys / Category
• Performance Evaluation: After task completion, write to .state/.verdict.json. Format follows @.principles/schemas/agent_verdict_schema.json.

Step 5: Delegate Auditor (Deductive Audit)

• Let Auditor output: Axiom/System/Via negativa / RESULT: PASS/FAIL
• Write audit result to AUDIT.md (RESULT line must exist).
• Performance Evaluation: After task completion, write to .state/.verdict.json. Format follows @.principles/schemas/agent_verdict_schema.json.

Branch Handling (Must Follow)

• If RESULT = FAIL:
  1. Write Must-fix list to AUDIT.md
  2. Go back to Step 4 to re-delegate Diagnostician, request supplemental root cause
  3. Max 2 retries; if still FAIL, write to memory/DECISIONS.md and request user intervention
• If RESULT = PASS: Continue to Step 6

Step 6: Delegate Planner (Movie Script Plan)

• Planner outputs Plan (steps/commands/metrics/rollback).
• Write plan to PLAN.md (STATUS line must exist).
• Task Sync:
  • If CLAUDE_CODE_TASK_LIST_ID is set, you must convert the Plan's core steps to Native Tasks (via natural language command "Add task..." or related tools).
  • If not set and in interactive mode, prompt user: "Recommend running export CLAUDE_CODE_TASK_LIST_ID=task-$(date +%s) to enable persistent task tracking."
  • If in background/headless mode, skip prompt.
• Performance Evaluation: After task completion, write to .state/.verdict.json. Format follows @.principles/schemas/agent_verdict_schema.json.

Step 7: Delegate Implementer (Execution)

• Implementer can only execute according to PLAN. Any deviation must first update PLAN.
• Performance Evaluation: After task completion, write to .state/.verdict.json based on verification results. Format follows @.principles/schemas/agent_verdict_schema.json.

Step 8: Delegate Reviewer (Review)

• Reviewer outputs: Critical/Warning/Suggestion.
• Performance Evaluation: After task completion, write to .state/.verdict.json. Format follows @.principles/schemas/agent_verdict_schema.json.

Branch Handling

• If Critical exists: Go back to Step 6 to revise plan, max 2 retries
• If no Critical: Continue to Step 9

Step 9: Reflection & Persistence

1. System Evolution: Append Pain/Root cause/New principle candidates/Gate suggestions to memory/ISSUE_LOG.md, and update memory/DECISIONS.md.
2. User Profile Update (Mandatory):
   • Review user's performance in this task (instruction quality, domain knowledge, preferences) and system highlights.
   • Must write to .state/.user_verdict.json (incremental update). Format must strictly follow @.principles/schemas/user_verdict_schema.json.
   • Note: If user shows no obvious characteristics or preferences, corresponding fields can be empty.

Step 10: Final Briefing

• Action: Delegate pd_spawn_agent(reporter) for closing statement.
• Requirement: Pass Implementer and Reviewer's final outputs as input. Let it decide report depth and style based on USER_CONTEXT.md.
• Goal: Ensure the boss (user) clearly understands task outcomes and potential risks without cognitive load.