Session Wrap-Up

Run three phases in order before ending the session. Auto-apply all changes without asking. Present a consolidated report at the end.

When to run: Before closing terminal, ending a work session, or when you're done for the day. Runs BEFORE the automatic Stop hooks (evaluate-session.sh) fire.

Phase 1: Ship It

1.1 Commit check:

• Run git status — if uncommitted changes exist:
  • Stage relevant files (NOT .env*, credentials, or generated files)
  • Commit with descriptive message following <type>: <description> convention
  • Push to remote

1.2 File placement check:

• If any files were created this session, verify they follow naming conventions from .claude/rules/conventions.md
• Auto-fix misplaced files (e.g., docs in src/, skills without SKILL.md)

1.3 Task cleanup:

• If beads tasks were worked on: bd update <id> --append-notes "session wrap-up: <status>"
• If task is complete: bd close <id>
• If task has remaining work: note what's left in the update

1.4 Worktree check:

• If the current worktree's feature branch is fully merged, remind user to clean up:
  Copy

  Worktree cleanup available: git worktree remove ~/worktrees/property-tracker/<name>
  

Phase 2: Remember It

Review what was learned during the session. Use this decision framework:

Actions:

1. Check if any debugging insights, user corrections, or conventions should persist
2. If something belongs in auto memory, update MEMORY.md (keep under 200 lines)
3. If a CLAUDE.md rule is needed, add it to the appropriate scoped file
4. Staleness check: Scan MEMORY.md for references to PRs, branches, files, or features that may be stale. Look for:
   • PR references — check if merged/closed with gh pr view <num> --json state
   • Branch references — check if branch still exists with git branch -a | grep <name>
   • "IN PROGRESS" items — verify if actually still in progress or completed
   • File paths — spot-check 2-3 referenced paths still exist
   • Mark stale entries for cleanup or update them
5. If nothing worth persisting: say "No new learnings to persist"

Phase 3: Review & Apply

Analyze the conversation for self-improvement findings. If the session was short or routine, say "Nothing to improve" and skip to the report.

Focus areas (in priority order):

1. Mistakes & near-misses — What went wrong? What took multiple attempts? What did the user have to correct? (Most valuable — don't skip this)
2. Friction — Repeated manual steps that should be automatic
3. Skill gaps — Things Claude struggled with or got wrong
4. Automation candidates — Repetitive patterns that could become skills or hooks

For each finding, auto-apply:

Conflict check: Before adding any new rule or instinct, search existing rules and instincts for contradictions. If "always do X" already exists and you're about to add "never do X in context Y", explicitly note the scoping difference.

Report format:

## Wrap-Up Report

### Ship It
- [x] Committed: <commit hash> "<message>"
- [x] Pushed to origin/<branch>
- [ ] No beads tasks active

### Remember It
- [x] Added "<insight>" to auto memory
- [ ] No CLAUDE.md updates needed

### Review & Apply
Applied:
1. ✅ Near-miss: <description> → [destination] Added <what>
2. ✅ Friction: <description> → [destination] Added <what>

No action needed:
3. Already documented in anti-patterns.md

### Session Stats
- Duration: ~Xh
- Files modified: N
- Commits: M

Important Notes

• This skill runs BEFORE the Stop hook. The Stop hook (evaluate-session.sh) will separately extract instincts from tool usage — that's complementary, not duplicative.
• Do NOT auto-commit to main/develop. Only commit to the current feature branch.
• If there are failing tests, do NOT wrap up — fix tests first or note them as blockers.
• Keep the report concise. The value is in the actions taken, not the report itself.