claude-sdk-rs: Command for Claude Code

Wrap Up — Log work and commit at the end of a session.

Runs /log-work then /commit in sequence. Use this after finishing a piece of work to sync status.md, append a log entry, and commit everything cleanly. For a full end-of-context hand-off to a fresh agent, use /handoff instead.

Variables

$ARGUMENTS — optional free-text note about what was done (passed straight through to /log-work as its narrative). May be brief ("shipped D17 --from flag") or detailed. If omitted, /log-work derives the narrative from git history.

Instructions

1. Drain any durable caveat first. If this session surfaced something the next agent must not lose, append it to planning/state.json carryover[] (field shape in docs/state/state-schema.md) with one of these kind values:

   kind	for
   defect	a real unticketed bug with a fix — not yet filed as its own block
   deferred	a real follow-on you haven't ticketed yet
   drift	a doc, comment, block title or generated surface that has fallen out of step with the code or the graph
   env	a transient environmental caveat ("installed binary is stale, rebuild first")

   constraint and known_issue are retired (HQ D72) — okf-core preserves them only through its Unknown(String) fallback so legacy entries still round-trip. Do not mint new entries with either.

   /wrap-up writes no handoff file, so carryover[] is the only place this kind of note survives. Skip if the session produced none or the repo has no state.json.

   Route at write time — three destinations, not two. Ask both questions before appending:

   1. Can only a human do this? A decision only the operator can make, a credential only they hold, a judgement call, a thing they must look at — that is not a carryover[] entry. File it as a {"type":"operator", slug, exit, start, what?} edge on the block it gates, per the operator-work rule below. Why it matters here and not only there: a carryover entry gates nothing, so operator work parked in it is never forced; an operator edge blocks the work standing behind it, which is what gets it done. Measured 2026-08-19 — 30 of the fleet's 202 carryover[] entries are operator work misfiled this way, filed as defect or deferred because the table above offers no row meaning "not an agent's to do."
   2. Is it permanently true? A gotcha still true next month, a deliberate non-fix nobody intends to reverse, a load-bearing measured number someone will need again — that belongs in reference[]. A fact with no clears_when because nothing will ever make it stop being true is the signal. See docs/state/reference-container-schema.md for its field table and kind vocabulary.