Optimize

Optimize Skill

Diagnose → Propose → Confirm → Apply → Verify loop for continuous improvement.

Architecture

User: "/optimize [target]" or "优化 [目标]"
       │
       ▼
1. DISCOVER — identify what to optimize based on context or explicit target
       │
       ▼
2. DIAGNOSE — analyze current state, find issues and improvement opportunities
       │
       ▼
3. BOOTSTRAP — check state/bootstrap.json for prior successful patterns
       │
       ▼
4. PROPOSE — generate one proposal per actionable finding, write to state/proposals/
       │
       ▼
5. CONFIRM — present proposals one at a time, wait for accept/reject/skip/edit
       │
       ▼
6. APPLY — execute accepted proposals (edit files, run commands, adjust config)
       │
       ▼
7. VERIFY — check results, compare before/after
       │
       ▼
8. RECORD — if improvement confirmed, record to state/bootstrap.json

Step 1: Discover

Determine what to optimize:

• If user provides explicit target: use it
• If user just ran observability: optimize the findings
• If user has state/user_dna.json with low confidence: optimize the cognitive model
• If .pi/permissions.json recently changed: optimize permission rules
• Otherwise: ask user what they want to improve

Step 2: Diagnose

Analyze the target and present findings:

## 诊断: [target]

Issues found:
  1. [severity] issue description — impact
  2. [severity] issue description — impact
  ...

Severity levels:

• HIGH: Blocking or causing significant problems
• MEDIUM: Degrading quality or adding friction
• LOW: Minor improvement opportunity

For each issue, note whether it's fixable by configuration change vs. requires structural change.

Step 3: Bootstrap

Check state/bootstrap.json for prior successful patterns related to the current target. If found:

"这个领域之前有 N 次成功的优化记录。参考：[summary of past successful approaches]."

If no bootstrap data: "这是第一次优化这个领域——每次采纳的方案会积累成经验。"

Step 4: Propose

For each actionable diagnostic, generate ONE proposal:

{
  "id": "prop_YYYYMMDD_HHMMSS",
  "created_at": "ISO8601",
  "target": "what's being optimized",
  "issue": "description of the problem",
  "severity": "high|medium|low",
  "diagnosis": "root cause analysis",
  "proposed_change": {
    "action": "what to do",
    "details": {}
  },
  "expected_impact": {
    "metric": "what should improve",
    "current": "before state",
    "expected": "after state",
    "basis": "why we think this will work"
  },
  "risk": "low|medium|high",
  "status": "pending"
}

Risk levels:

• low: Non-destructive change, easy to revert
• medium: Moderate blast radius, may affect related areas
• high: Significant change, hard to roll back

Write proposals to state/proposals/prop_YYYYMMDD_HHMMSS.json.

Step 5: Confirm

Present proposals one at a time:

## Proposal 1/3: [severity] title

**Problem:** [what's wrong, with evidence]
**Root cause:** [why it's happening]
**Fix:** [what will change]
**Expected impact:** [before → after]
**Risk:** [level — what could go wrong]

**Accept?** (a)ccept / (r)eject / (s)kip / (e)dit

Wait for user response after each proposal.

If "edit", let the user modify the proposal before applying. If "reject", mark the proposal rejected and move on. If "skip", mark skipped — can revisit later.

Step 6: Apply

Execute accepted proposals. For each:

1. Make the change (file edit, config update, command execution)
2. Track what was changed (old value, new value, reason)
3. If the change is reversible, note how to revert

If multiple proposals touch the same target, apply in dependency order.

Step 7: Verify

Check the result:

1. Run relevant verification (tests, smoke test, before/after comparison)
2. Present results:

## 验证结果

Proposal 1: [status] — [evidence of improvement or regression]
Proposal 2: [status] — [evidence]

If improvement: mark applied. If regression: mark failed, suggest rollback. If no change: mark applied (still a valid attempt).

Step 8: Record

For proposals marked applied:

1. Update proposal JSON status to applied
2. Record to state/bootstrap.json:

{
  "domain": "optimization target",
  "approach": "what worked",
  "result": "outcome",
  "timestamp": "ISO8601",
  "quality": "high|medium|low"
}

Rollback

If regression detected:

1. Revert the change using tracked old/new values
2. Mark proposal failed
3. Suggest alternative approach
4. Record the failure to bootstrap to avoid repeating

Summary

After all proposals processed:

## Optimization Summary

Target: [what was optimized]
Accepted: X/Y proposals
Applied:  A/X (B failed, C no change)

Changes:
  [list of applied changes with results]

Next: [suggestion for what to check or optimize next]

Edge Cases

Key Files