Agent安装

1. Initialize

# Load context
Read AI-GUIDE.md → minimal context strategy
Read TRACKER.md → current state
Read ROADMAP.md → phases list

Display startup banner:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 VIEPILOT ► AUTONOMOUS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

 Project: {project_name}
 Phase: {current_phase}/{total_phases}
 Progress: [████░░░░░░] {percent}%

2. Select Phase

• Find first incomplete phase
• Or resume in_progress phase
• Check dependencies met

3. Execute Phase Loop

For each task in phase:

3a. Load Task Context

Read:
- AI-GUIDE.md (always)
- TRACKER.md (always)
- PHASE-STATE.md (current phase)
- tasks/{task}.md (current task)
- context_required files (from task)

Stack preflight:
- Identify stacks used by task
- Load stack cache summary first
- Apply stack do/don't rules during implementation

Doc-first gate + checkpoint order

1. Validate task contract (workflows/autonomous.md).
2. Record plan in task file; set task in_progress in PHASE-STATE.md.
3. Stack preflight.
4. Create git tag {projectPrefix}-vp-p{phase}-t{task} only after steps 1–3 pass.

3b. Execute Task

• Implement according to objective
• Follow acceptance criteria
• Split large tasks into sub-tasks (30-90 minutes)
• Record plan + files + best-practice checklist before each sub-task
• Write tests if required
• Atomic commits per sub-task
• Log notes in task file

3c. Verify Task

automated:
  - Run defined commands
  - Check expected outputs
manual:
  - Present if defined
  - Ask user if required
quality_gate:
  - All acceptance criteria checked
  - Automated tests pass
  - No lint errors

3d. Handle Result

• PASS → Mark task done, create tag -done, next task
• PARTIAL → Retry with fix
• FAIL → Control point (retry/skip/rollback/stop)

4. Update State

After each PASS task and PASS sub-task:

• Update PHASE-STATE.md
• Update TRACKER.md
• Update HANDOFF.json
• Update ROADMAP.md if progress/status changed
• Update CHANGELOG.md (if feature/fix)

Rule: state-first then continue. Do not batch updates only at end of phase.

5. Phase Complete

When all tasks done:

• Run phase verification
• Check phase quality gate
• Write SUMMARY.md
• Create git tag: {projectPrefix}-vp-p{phase}-complete
• Increment version if needed

6. Iterate

• More phases? → Loop to step 2
• All complete? → Suggest /vp-docs

7. Handle Blockers

At any control point, offer:

1. Fix and retry
2. Skip (with reason)
3. Rollback
4. Stop autonomous mode

Display progress summary on stop.

Adapter Compatibility

AskUserQuestion Tool (ENH-059)

Control-point prompts use AskUserQuestion on Claude Code (terminal).

Claude Code (terminal) — AUQ preload required (ENH-059): Before the first interactive prompt, call ToolSearch with query: "select:AskUserQuestion" to load the deferred tool schema. Only after ToolSearch succeeds can AskUserQuestion be invoked. If ToolSearch returns an error, fall back to plain-text numbered list for that session.

Prompts using AskUserQuestion in this skill:

• Control point decisions (retry / skip / rollback / stop — Step 7)

Skill Context Execution (FEAT-020)

At session start, vp-auto reads PROJECT-CONTEXT.md ## Skills (locked by /vp-crystallize Step 1E) and builds a SKILL_CONTEXT_MAP.

During each task:

• Required skills whose phases match the current task's phase are silently injected into the execution context
• Best practices from matched skills appear as a silent checklist before implementation
• Task output records skills_applied: [id@version, ...]

No prompts are shown — all skill decisions are locked at crystallize time. /vp-auto is execution-only for skills.

Install skills: vp-tools install-skill <source> Lock decisions: /vp-crystallize Step 1E Docs: docs/user/features/skill-registry.md