Wrap-Up — Session End

Execute all steps. Present findings to the user before committing anything.

Step 1: Review Deliverables

Scan what happened this session:

1. Run git status and git diff --stat to see all changes
2. List every file created or modified, grouped by location:
   • context/ — context files updated
   • .claude/skills/ — skills created or modified
   • Source code — implementation work
   • Docs repo — documentation created/updated (if workflow-enabled)
   • Other locations — flag for review
3. File placement check: Verify outputs follow naming conventions:
   • Skills use {category}-{function} naming
   • Daily memory uses YYYY-MM-DD.md format (main sessions) or YYYY-MM-DD_{session-name}.md (sandbox sessions)
   • If anything is in the wrong place, flag it

Step 2: Detect Sandbox Environment

4. Check if you're running inside an agentsandbox:
   • Check if the AGENTSANDBOX_SESSION env var is set
   • OR check if /workspace/CLAUDE.md exists (container indicator)
   • OR check if the current branch starts with asb/ (git symbolic-ref --short HEAD)
   • If any is true, you are in a sandbox session
5. If in a sandbox session, determine the session name:
   • From $AGENTSANDBOX_SESSION env var (preferred)
   • Or from hostname: strip asb- prefix from $HOSTNAME
   • Or from branch name: strip asb/{source}- prefix

Step 3: Update Daily Memory

6. Choose the correct filename:
   • Main session: context/memory/{YYYY-MM-DD}.md
   • Sandbox session: context/memory/{YYYY-MM-DD}_{session-name}.md
     • Example: context/memory/2026-03-21_ux-study-test.md
   • This keeps sandbox memory separate from main — no merge conflicts
7. For full history context in a sandbox, read from .context-main/memory/ (read-only mount of main's context)
8. Update the session section with:
   • Goal: What the user wanted to accomplish
   • Work Done: Bullet points of what was accomplished
   • Decisions: Key decisions made and why
   • Learnings: Things discovered about the codebase, tools, or patterns
   • Open Items: Unfinished work, blockers, things to pick up next

Step 4: Documentation Check (if workflow-enabled)

9. Read workflow.json to check if this is a workflow-enabled project
10. If yes, review the session for documentation gaps:

Decisions → ADRs:

• Were any architectural or technical decisions made this session?
• If yes, does an ADR already exist for this decision?
• If no ADR exists, suggest: "Decision about [topic] — want me to create an ADR?"

New features → PRDs/Design Docs:

• Was a new feature discussed or started?
• Does it have a PRD or design doc?
• If not, suggest: "New feature [name] — want a PRD or design doc?"

Architecture changes → C4 Diagrams:

• Were new components, services, or integrations added?
• Do the C4 diagrams in the docs repo reflect these changes?
• If not, suggest: "Architecture changed — C4 diagrams may need updating"

Incidents → Post-Mortems:

• Was a production incident investigated or resolved?
• If yes, suggest a post-mortem

Research → Spike Reports:

• Was research or option comparison done?
• If yes, suggest capturing it as a spike report

Present documentation suggestions as a checklist — the user picks what to do:

### Documentation Suggestions
- [ ] Create ADR for [decision about X]
- [ ] Update C4 container diagram (new service added)
- [ ] Create spike report for [research topic]

Step 5: Distill to Long-Term Memory

10. Review today's session notes. Ask yourself:
    • Any lessons learned that future-me should know?
    • Any decisions that affect how we work going forward?
    • Any user preferences discovered?
11. If yes, update context/MEMORY.md with the distilled insight
12. If any context files (SOUL.md, USER.md) should be updated based on what you learned, suggest the changes to the user

Step 6: Collect Feedback (Optional)

13. Ask the user briefly:
    • "Anything I should do differently next time?"
    • "Any preferences to remember?"
14. If they provide feedback, update the relevant context file

Step 7: Commit (If Requested)

15. Only if the user asks to commit:
    • Stage relevant files
    • Draft a clean commit message (no AI mentions, no Co-Authored-By)
    • Show the message for approval before committing
    • If changes span multiple repos (project + docs), handle each repo separately

Step 8: Cost Summary

16. Check if ~/.claude/metrics/costs.jsonl exists. If it does, calculate today's total spend:
    bashCopy

    jq -s '[.[] | select(.timestamp | startswith("'"$(date -u +%Y-%m-%d)"'"))] | {total_eur: (map(.estimated_cost_eur) | add), input_tokens: (map(.input_tokens) | add), output_tokens: (map(.output_tokens) | add), responses: length}' ~/.claude/metrics/costs.jsonl
    

17. Present the cost summary to the user:
    Copy

    ### Today's Usage
    - Responses: N
    - Tokens: Xk input / Yk output
    - Estimated cost: €X.XX
    

Step 9: Summary

18. Present a brief session summary:
    • What was accomplished
    • What's left for next time
    • Any context files that were updated
    • Any documentation suggestions that were deferred (so we remember next session)
    • Today's cost (from step 16)