Session handover

Use this skill when pausing work, changing the lead model, mining session history, or preparing continuity for a fresh model that will wake up without the outgoing model's private context.

Operating model

Treat the system as a working kitchen staffed in shifts.

The incoming shift does not need the outgoing shift's entire internal monologue. It needs an accurate station state:

• what is done;
• what is still outstanding;
• what is currently in progress and who owns it;
• what failed and must not be repeated blindly;
• what is running low or close to a limit;
• what is hazardous, uncertain, or requires verification;
• what is reserved or explicitly DO NOT TOUCH;
• the exact next safe action.

The purpose is operational continuity, not narrative completeness.

Goals

• Preserve signal: user constraints, decisions, stable workflows, current state, ownership, risks, and receipts.
• Remove noise: retries, verbose tool output, duplicated pings, abandoned speculation, and private chain-of-thought.
• Prevent a fresh model from duplicating work, acting on stale state, exhausting resources, or crossing protected boundaries.
• Leave the station usable and the next action executable.

Inputs

• Session directory: ~/.openclaw/agents/<agentId>/sessions/
• Optional index: sessions.json
• Memory files in the workspace (memory/*.md; MEMORY.md only where allowed)
• Current plans, blackboards, receipts, service state, git state, and task artefacts when relevant

Recommended output

Write or update the appropriate durable handover note, normally SESSION_NOTES.md, a daily memory entry, or the active typed blackboard.

Use this compact shift sheet:

1. Outcome / station state — what is now true.
2. Completed — finished work with receipts.
3. In progress / owner — active work and who owns it; say none if empty.
4. Outstanding — bounded remaining jobs.
5. Failures / do not repeat — failed approaches and exact boundaries.
6. Running low / limits — tokens, time, storage, quotas, hop budgets, locks, deadlines, or other capacity.
7. Hazards / verify — uncertain claims and checks required before action.
8. Reserved / DO NOT TOUCH — protected files, data, services, credentials, user decisions, and exact target scope.
9. Next action — one exact safe first step for the incoming shift.
10. Receipts / rollback — paths, revisions, hashes, backups, commands, and status.
11. Security redactions — confirm secrets and irrelevant private material were not copied.

Omit empty prose, but never silently omit the protected-target, hazard, resource-limit, or ownership state. Use none known where ambiguity would be dangerous.

Method

1. Inventory the station

   • List relevant sessions and artefacts by date/size.
   • Read current plans, blackboards, locks, service state, and receipts when the task used them.
   • Resolve which work is complete, active, abandoned, or blocked.

2. Extract user intent and constraints first

   • Read authenticated user messages before assistant/tool output.
   • Preserve the newest applicable decision.
   • Render exact protected targets as DO_NOT_TOUCH: <target>.

3. Separate state from story

   • Record observed facts, decisions, actions, failures, evidence, and open questions distinctly.
   • Do not pass guesses as facts.
   • Do not copy private reasoning or verbose transcripts when the operational result suffices.

4. Establish ownership

   • Name the active owner for in-progress work.
   • If ownership cannot be established, mark the item unclaimed; do not imply it is safe to duplicate.
   • Clear stale ownership only with evidence that the prior worker is no longer active.

5. Check resources and hazards

   • Record remaining budgets, quotas, storage, deadlines, lock/revision state, and recovery boundaries where material.
   • Flag anything that could run out, collide, become destructive, or leave the system unrecoverable.
   • State what the incoming shift must verify before trusting volatile or high-impact state.

6. Attach receipts

   • Prefer exact paths, revisions, hashes, timestamps, test outcomes, backups, and rollback instructions.
   • A claim such as “fixed” requires proportionate proof.

7. Write the shortest complete handover

   • Make the exact next action executable without reconstructing the whole session.
   • Expire or remove stale notes rather than letting old state masquerade as current state.

8. Read it as the incoming shift

   • Confirm a fresh model can continue safely without asking what happened.
   • Confirm it cannot mistake reserved work for free stock or completed work for outstanding work.

Useful command snippets

ls -lt ~/.openclaw/agents/main/sessions/*.jsonl
jq -r 'select(.message.role=="user") | .message.content[]? | select(.type=="text") | .text' <session>.jsonl
jq -r '.message.content[]? | select(.type=="toolCall") | .name' <session>.jsonl | sort | uniq -c | sort -rn

Quality gate

A handover passes only if a fresh model can answer:

• What is true now?
• What was completed, and where is the proof?
• What remains, and who owns it?
• What failed or must not be repeated?
• What resource or limit is close to exhaustion?
• What must be verified before acting?
• What is reserved or forbidden to touch?
• What exact action should happen next?
• How is rollback performed if that action fails?

If any answer that matters is unclear, the outgoing shift has not finished closing down.