wrap

Brain OS : Session Wrap

Close the session cleanly. Update entity state. Capture decisions. Detect momentum shifts.

Not just "what changed" : "what does the system need to remember?"

REQUIRED FIRST READ

Before any tool call, read ~/.claude/brain-os/PROTOCOL.md. The wrap writes state back via MCP mutating tools; it must read current state via MCP tools first.

Step 0: Check for Compact Checkpoints

Before asking wrap questions, check for unconfirmed checkpoint files:

.brain/sessions/checkpoints/*-*.json

Look for any file where "confirmed": false and "status": "captured".

If found:

1. Read the checkpoint(s).
2. Show the user a human summary of what was captured:

   Before context was compacted, I saved a checkpoint:
   - You were working on: [last_user_goals summary]
   - Files touched: [count] files in [project]
   - Open threads: [open_questions summary]
   - Last direction: [last_assistant_summary, one line]
   
   Does this look right? I'll fold the confirmed parts into today's wrap.
   

3. User confirms, edits, or discards.
4. Confirmed content feeds into the normal wrap questions below (pre-filling suggestions).
5. After wrap completes, mark the checkpoint file: "status": "confirmed" or "status": "discarded".

Auto-wraps from a previous session

Also check for unconfirmed auto-wrap records written by wrap_auto:

.brain/sessions/autowrap-*.json

Each has "confirmed": false, the low-risk fields already applied (applied), and high-risk proposals staged under pending_review. For each one:

1. Tell the user what was already auto-saved (the applied list — next move, open questions, evidence).
2. For each pending_review item (status, mode, blocked, decisions), ask one short yes/no: "Last session I held this back for you to confirm: [item]. Apply it?"
3. Apply confirmed items via the normal tools (entity_update; decision_check then decision_log for decisions). Discard the rest.
4. Mark the record "confirmed": true.