Self Reflect

Analyze the work done in this session and extract valuable learnings into the knowledge base.

Process

1. Gather Context

Review what happened in this session:

• Read recent git log to see what was committed
• Check git diff for uncommitted changes
• Review any test failures encountered during the session
• Note any corrections the user made to your approach

2. Identify Learnings

Extract insights across these categories:

3. Write Entries

For each learning, append a JSONL entry. Field order matches .agents/knowledge/README.md:

{
  "id": "<category-prefix>-<next-number>",
  "type": "<pattern|gotcha|decision|anti-pattern|codebase-fact|api-behavior>",
  "fact": "Core insight in one sentence",
  "recommendation": "How to apply this in future sessions",
  "confidence": "<high|medium|low>",
  "provenance": [{"source": "<human|agent|review|test-failure>", "reference": "PR #X or file path", "date": "<today>"}],
  "tags": ["relevant", "tags"],
  "affectedFiles": ["paths/that/this/applies/to"],
  "createdAt": "<ISO timestamp>",
  "updatedAt": "<ISO timestamp>",
  "usageCount": 0,
  "helpfulCount": 0,
  "outdatedReports": 0
}

Rules:

• Check for duplicates before writing — re-confirm existing entries (bump helpfulCount, set updatedAt) instead of creating new ones
• Use high confidence only for facts verified by tests or explicit user confirmation
• Keep fact under 120 characters — it's the headline
• recommendation should be actionable — what should an agent DO differently
• tags should match file paths or feature areas for filtering by /prime

4. Prune Stale Entries

If you notice existing entries that are outdated (e.g., a gotcha that was fixed, a decision that was reversed):

• Bump outdatedReports by 1 and set updatedAt to now if you're not sure — let /evolve decide
• Remove the entry only when the staleness is unambiguous (referenced file gone, contradicted by a newer entry, or outdatedReports already exceeds helpfulCount)
• Or update it with the new information and reset confidence

5. Report Summary

After reflecting, output a brief summary:

## Session Reflection

**Patterns captured:** N new, M updated
**Gotchas found:** N
**Decisions recorded:** N
**Anti-patterns noted:** N

### Key Insights
- [Most important learning]
- [Second most important]

When to Trigger

• After completing a feature or significant bug fix
• After a code review (human or council)
• After encountering surprising behavior
• When the user says "we're done" or "let's wrap up"
• Proactively suggest running /self-reflect at natural session endpoints