Agent诊断

What I Do

After ANY skill is used, this meta-skill triggers to:

1. Analyze what worked and what didn't
2. Identify new patterns or shortcuts discovered
3. Update the skill file with learnings
4. Prevent knowledge loss between sessions

Relationship with self-improve:

• This skill (skill-reinforcement) = WHEN to update (post-use triggers)
• The self-improve skill = HOW to update (templates, structures, decision trees)

When to Trigger

Invoke this skill automatically when:

• Any skill from .opencode/skill/*/SKILL.md completes
• A workflow succeeds or fails in a notable way
• New shortcuts or anti-patterns are discovered
• Token usage could be reduced with better patterns
• API behavior differs from documentation
• Commands fail and I find the fix
• User confirms something works
• I do the same task twice (should become a skill/tool)

Reinforcement Process

Step 1: Capture the Context

After using a skill, note:

- Skill used: [skill-name]
- Task: [what was being done]
- Outcome: [success/partial/failure]
- Token cost: [high/medium/low]
- Time taken: [fast/normal/slow]

Step 2: Identify Learnings

Ask these questions:

1. What took longer than expected? → Document the fix
2. What failed unexpectedly? → Add to "Common Issues"
3. What shortcut was discovered? → Add to "Token Saving Tips"
4. What assumption was wrong? → Correct in documentation
5. What worked better than documented? → Update the workflow

Step 3: Categorize the Learning

Step 4: Update the Skill File

# Read current skill
cat .opencode/skill/[skill-name]/SKILL.md

# Edit to add learning in appropriate section
# Use the Edit tool to append or modify

Step 5: Validate the Update

Ensure updates are:

• Actionable - Not vague observations
• Specific - Include exact commands/patterns
• Formatted - Match existing style
• Non-redundant - Don't duplicate existing content

Learning Templates

For New Shortcuts

### [Shortcut Name]

**Discovery**: [How it was found]
**Before**: [Old approach]
**After**: [New approach]
**Savings**: [Token/time reduction]

For Failure Modes

| Issue  | Symptom        | Fix              |
| ------ | -------------- | ---------------- |
| [Name] | [What you see] | [How to resolve] |

For Anti-Patterns

### Don't: [Bad Pattern Name]

```javascript
// BAD - [explanation]
[bad code]
```

Do: [Good Pattern Name]

// GOOD - [explanation]
[good code]

### For Workflow Improvements

```markdown
## Updated: [Section Name]

[New content that replaces or augments existing]

> **Note**: Updated [date] after discovering [context]

Real Examples

Example 1: Session Persistence Discovery

Context: Testing staging branch, went through full login flow Learning: Chrome MCP persists sessions - was already logged in Action: Added "Session Persistence is Your Friend" section with check-first pattern

Example 2: OTP Extraction Optimization

Context: Opened email, waited for load, extracted OTP Learning: OTP visible in Gmail list preview without opening email Action: Updated OTP section with faster "search + list preview" method

Example 3: Deployment URL Pattern

Context: Manually constructed URL, got it wrong Learning: Can extract URL directly from Vercel bot's PR comment Action: Added gh pr view command to get URL automatically

Skill File Structure Convention

Every skill should have these sections (add if missing):

## What I Do

[Core purpose]

## Prerequisites

[Requirements]

## Workflow

[Main steps]

## Common Issues

[Failure modes and fixes]

## Token Saving Tips

[Efficiency patterns]

## Anti-Patterns to Avoid

[What NOT to do]

## Real Examples

[Actual usage examples]

## Learnings Log

[Append-only log of discoveries - optional]

Integration with Other Skills

When reinforcing a skill, check if learnings apply to related skills:

Cross-pollinate learnings when applicable.

Deciding What to Create

If reinforcement reveals a need for new capability, use the self-improve skill's decision tree:

Need to extend capabilities?
│
├─ Just need docs/commands? → SKILL
├─ Need specialized AI persona? → AGENT
├─ Need event hooks? → PLUGIN
├─ Need callable function? → TOOL
└─ Need external integration? → MCP SERVER

Automation Hooks

Post-Skill Trigger

After completing any skill, automatically ask:

1. "Did anything unexpected happen?"
2. "Was there a faster way to do this?"
3. "What would I do differently next time?"

If answers exist, invoke skill-reinforcement.

Periodic Review

Every ~10 skill uses, review:

• Most frequently used skills (prioritize improvements)
• Skills with most "Common Issues" (need better documentation)
• Skills with outdated information (need refresh)

Meta: Reinforcing This Skill

This skill should also improve itself. Track:

• How often it triggers
• Quality of captured learnings
• Whether skills actually improve over time
• Time cost of reinforcement vs value gained

Quick Reinforcement Checklist

[ ] Skill completed
[ ] Outcome noted (success/fail/partial)
[ ] Any surprises? → Document
[ ] Any shortcuts found? → Add to tips
[ ] Any failures? → Add to issues
[ ] Could be faster? → Add anti-pattern
[ ] Update skill file
[ ] Validate formatting
[ ] Done

Immediate Update Rule

UPDATE IMMEDIATELY - Don't wait until end of conversation.

When any of these happen, stop and update the relevant skill:

1. API format is wrong (like -d vs -F for curl)
2. Commands fail and I find the fix
3. User confirms something works
4. I discover a better/faster way
5. Something is missing from docs

Example:

❌ "I'll note this for later"
✅ *immediately edits skill file*

Learnings Log

• 2026-01-12: Bulk remote branch cleanup can hit stale refs and timeouts; run git fetch --prune origin first, delete with a loop that tolerates missing refs, then prune again to verify only origin/main remains.
• 2026-01-13: When extending agent policy docs, renumber numbered headings after inserts and place testing tools + real-funds protocol near the Testability section for clarity.
• 2026-01-13: When refactoring MCP tool handlers into a registry, remove the legacy switch and align handler arg types with tool schemas to avoid type errors.
• 2026-01-13: When adding a new Next.js route handler export, ensure it sits outside existing handler functions to avoid "Modifiers cannot appear here" errors.