/handover

Baton pass between two chats. A session is getting long, or you want to close it, but the work is not done: handover captures the state in a temporary file, and a fresh chat picks up from there with a clean context.

Not a debrief. It does not write to memory/, does not write to daily-log/, does not commit and does not push anything. If the work needs to be persisted (project status, session log), that is session-debrief. This is surgical and instant: a single throwaway file, outside the repo.

The file lives at a fixed path: ${TMPDIR:-/tmp}/brain-handover.md. Outside the repo (git never sees it), the OS cleans it up, and at most one exists at a time. The slot has no owner: the first "resume" from any chat consumes it, which is why Resume mode confirms the title before deleting anything.

No secrets in the file. The temp directory has no restrictive permissions and the file can outlive the chat that wrote it, so it must never be the only place a credential lives. Record the name of the variable, never its value; the value is re-entered through the environment in the new chat.

Detect the mode

Two modes in the same skill, distinguished by the operator's intent:

• Save - the operator wants to freeze the current chat (save triggers above).
• Resume - the operator wants to pick up from a saved handover (resume triggers above).

If the intent is ambiguous, ask which of the two before acting.

Mode ① - Save

1. Scan the conversation and reconstruct the state: goal, what is done, next step, decisions made, files touched, guardrails, open threads.

2. Write ${TMPDIR:-/tmp}/brain-handover.md with the template below. Adaptive template: include only the sections that actually have content - no empty sections with a placeholder dash. Every save overwrites the previous file (use the Write tool, never shell text-munging: apostrophes in natural language break single-quoted shell strings).

   # Handover - <session title> - <YYYY-MM-DD HH:MM>
   
   ## Goal
   What we are doing and why (1-3 lines).
   
   ## Current state
   Where things stand now, what is already done in this chat.
   
   ## Next step
   The next concrete action to pick up from.
   
   ## Decisions made
   Choices and constraints settled in chat, so the new chat does not reopen them.
   
   ## Relevant files / paths
   Files touched or useful, with path, so the new chat does not re-search from scratch.
   
   ## Do not
   Explicit guardrails: what to avoid, paths already ruled out.
   
   ## References
   Links to other documents, resources, people, useful numbers.
   
   ## Open / do not forget
   Open threads, gotchas, anything pending.

   Get today's date with date +%Y-%m-%d (and date +%H:%M for the time) before writing it into the title; never compute it by hand. Before writing, scan the draft for anything that looks like a token, a key, or a password: if one is there, replace it with the variable name.

3. Confirm in chat: the path plus a 2-line recap of what was captured. No approval needed, no preview before writing: it is temporary and zero risk. Optionally remind the operator of the trigger for the other chat: "in the new chat: resume the handover."

Do not update memory, do not write to the daily log, do not perform any git operation. That is the whole point of this skill.

Mode ② - Resume

1. Read ${TMPDIR:-/tmp}/brain-handover.md.

   • If the file does not exist, say so calmly ("no saved handover to resume") and never invent context. Then look for a hand-written pass note before giving up: untracked files in the tree whose name says handover, and the ## Open threads section of your own recent daily logs.

     cd "$(git rev-parse --show-toplevel)"
     git -c core.quotePath=false status --short --untracked-files=all | grep -i "handover\|handoff" || echo "no hand-written pass note in the tree"
     SLUG="$(scripts/lib/operator-registry.sh resolve)" && ls daily-log/*-"$SLUG".md 2>/dev/null | tail -2

     If something related turns up, propose it ("I found <path>, open it?"); never open it silently, and stop if nothing does. A missing file on a resume is a breakage of the previous chat's save: apply the Self-improvement footer to this run as well.

2. Declare the title (the first line of the file) and ask the operator to confirm it is the thread they expect. The slot is single and has no owner, so a "resume" from any chat consumes whichever handover is there. Only after a yes go on to step 3. If it is not the expected thread, leave the file where it is and stop.
3. Load the context: absorb the content as the starting state of the new chat. Recap in chat where things stood and what the next step is, so the operator confirms alignment.
4. Delete the file (throwaway): rm -f "${TMPDIR:-/tmp}/brain-handover.md". Delete only after a successful read and a confirmed title, never before.
5. Confirm: "Baton taken, file discarded. Picking up from: <next step>."

Rules

• Never write to memory, daily-log, or git. That is the clean split from session-debrief. If persistence is needed, point there instead. Resume may run a read-only git status to look for a hand-written pass note; it never writes.
• One file, fixed path, overwrite. ${TMPDIR:-/tmp}/brain-handover.md. No timestamped files piling up.
• Delete after reading and confirming, not before. In Resume mode, the rm is the last step, only after a successful read and a confirmed title.
• No credentials in the file. Variable names only, never values.
• Adaptive template. Only the sections with real content.
• Write tool for the file, not sed/perl/heredoc text-munging. Apostrophes in natural language break single-quoted shell strings.
• Absolute dates, from date +%Y-%m-%d / date +%H:%M.

Self-improvement

At the end of the run, check it against these friction triggers: avoidable round-trip · improvisation · breakage · token waste · operator correction (full definitions in skills/_improvements/capture.md). If at least one fired, ask the operator "there was friction on X, should I log it?"; on their ok, append an entry to skills/_improvements/friction-log.md in the format capture.md defines. Clean run: say nothing. Never self-edit this skill; improvements go through /skill-improve.