| name | session-wrapup |
| description | Session closer. Writes chat summary, reflection note, ACCESS entries, and flags pending system maintenance. Produces deferred actions on read-only platforms. Activate when the user ends the session or context is running low. |
| compatibility | Requires agent-memory MCP server with memory_record_session |
| trigger | {"event":"session-end","priority":50} |
| source | user-stated |
| origin_session | manual |
| created | "2026-03-16T00:00:00.000Z" |
| last_verified | "2026-03-20T00:00:00.000Z" |
| trust | high |
Session Wrap-Up
Load this skill on your first bootstrap or when uncertain about the wrap-up protocol. Load core/governance/session-checklists.md only when you want the shorter session-end runbook there during normal sessions; it is an on-demand reference, not the live router.
When to use this skill
Activate when:
- The user says "wrap up", "end session", "that's all", or similar.
- The session is clearly concluding (final thanks, sign-off language).
- Context window is running low and the session should be archived before context is lost.
Steps
1. Record the session atomically
- Prefer
memory_record_session when the local agent-memory MCP surface is available. It should write the session SUMMARY.md, optional reflection.md, update core/memory/activity/SUMMARY.md, and append ACCESS entries in one commit.
- If the composite tool is unavailable, fall back to the individual governed writes or direct file writes as needed.
Create the session's chat folder if it doesn't exist: core/memory/activity/YYYY/MM/DD/chat-NNN/.
Write SUMMARY.md following the compression hierarchy in README.md § "Summaries":
- Key topics discussed.
- Decisions made and their reasoning.
- Action items (for the user or for future sessions).
- Notable context that a future agent should know.
If writing a reflection, follow the canonical format in README.md § "Session reflection".
When ACCESS entries are available, include them in the same composite call so session_id is injected automatically.
2. Update summaries if warranted
If this session produced significant new knowledge, identity changes, or skill refinements:
- Update the relevant folder's
SUMMARY.md to reflect the new content.
- For identity or governance changes, ensure they were proposed and approved per
core/governance/update-guidelines.md.
3. Check for system maintenance
- If any ACCESS.jsonl has hit the aggregation trigger (see
core/INIT.md), load core/governance/curation-algorithms.md and run aggregation now, or flag it for the next session start.
- If periodic review is overdue, add a reminder to
core/governance/review-queue.md.
4. Produce deferred actions (if read-only)
If write access is unavailable, produce a deferred-action summary listing:
- All ACCESS entries that should be appended.
- All file writes (summaries, reflections, knowledge updates) that should be applied.
- All review-queue items.
Present this using the format in core/governance/update-guidelines.md § "How to communicate deferred actions". If this is your first read-only session, also review the worked example in core/governance/update-guidelines.md § "Worked example" for the output format. (HUMANS/tooling/scripts/onboard-export.sh is for first-session onboarding only; it does not apply here.)
5. Sign off
Brief, warm sign-off. Reference something specific from the session to demonstrate continuity — not a generic "have a great day."
Quality criteria
- The chat summary should be useful to a future agent reading only SUMMARY.md files (no transcript access needed for basic context).
- The reflection note should be honest — low helpfulness scores and gap observations are more valuable than optimistic self-assessment.
- ACCESS entries are complete — every content file read is logged, including misses.
- If read-only, the deferred-action summary is comprehensive enough that a user can apply all changes without needing to re-read the session.
Anti-patterns
- Don't skip the reflection. It's tempting to write only the summary. The reflection is what makes the system self-improving.
- Don't inflate helpfulness scores. A file that was opened but didn't influence the response is a 0.2–0.4, not a 0.7.
- Don't write a novel. The summary should be 10–30 lines, not a transcript rehash.
- Don't forget deferred actions on read-only platforms. This is the user's only way to persist the session's value.