Skill: wrap-up

When to Use

Run at the end of every coding session — whether work is complete or paused.
This ritual ensures nothing is lost, lessons are captured, and the next session
can start immediately without archaeology.

---

The 5-Step Wrap-Up Ritual

Step 1 — Changes Audit

Review everything that happened this session:

git status                      # uncommitted changes
git diff --stat HEAD            # what changed vs last commit
git log --oneline -10           # commits made this session

Check for:

• Uncommitted changes that should be committed (use smart-commit)
• Uncommitted changes that should be discarded
• TODOs left in code (grep -r "TODO\|FIXME\|HACK" --include="*.py" .)
• Any .claude/state/ files that need updating

---

Step 2 — Quality Check

pytest -x                       # tests must pass
python -m py_compile proxy.py   # syntax check main file

If tests fail: either fix now or document the failure in NEXT_ACTION.md as a blocker.

Do not leave the session with silently broken tests.

---

Step 3 — Learning Capture

For each meaningful correction, mistake, or discovery this session, append to
.claude/state/learnings.md:

## <date> — <short title>

**Context:** What was being worked on.
**Mistake / Discovery:** What went wrong or what was learned.
**Correction:** What the right approach is.
**Pattern:** One-line rule to remember (e.g., "Always read router/CLAUDE.md before touching model_router.py").

Create the file if it doesn't exist. Even 1 learning per session compounds over time.

---

Step 4 — Next Session Planning

Update .claude/state/NEXT_ACTION.md with a clear resume guide:

# Next Action

**Objective:** <what are we building?>
**Status:** <COMPLETE | IN PROGRESS | BLOCKED>
**Last completed:** <step or task just finished>

## If Resuming: Start Here

1. Read: <file to read first>
2. Run: `pytest -x` to confirm baseline
3. Then: <exact next action>

## Blockers

- <any known blockers or dependencies>

## Optional Next Steps

- <lower priority work to do after the main task>

Also update .claude/state/agent-state.json:

• Set next_step to the next action
• Set last_updated to current ISO8601 timestamp

---

Step 5 — One-Paragraph Summary

Write a single paragraph that any developer (or AI) could read to instantly
understand what happened this session. Output it to the user as the final
message of the session.

Include:

• What was the goal
• What was accomplished
• What was skipped or deferred
• Any important decisions made
• What comes next

Example:

Added Scout agent and 9 missing skills (pro-workflow, smart-commit, wrap-up, learn-rule,
parallel-worktrees, replay-learnings, session-handoff, insights, deslop) to the .claude/
system. Updated CLAUDE.md with the new skill table and model selection guide. All tests
pass. Next: push to claude/improve-repo-docs-9Q3eO and verify CI is green.

---

Acceptance Checks

• git status reviewed — no unintended uncommitted changes
• pytest -x green (or failure documented as blocker)
• At least one learning captured in .claude/state/learnings.md
• NEXT_ACTION.md updated with clear resume instructions
• agent-state.json next_step and last_updated updated
• One-paragraph summary written