Guided Closeout

The operator runs many sessions and clones in parallel and does not read long final messages. A stopping point is handed back as a short recap plus one structured prompt whose options are the concrete next steps. The law is the "Guided Closeout" section of AGENTS.md; this skill is the shape.

When holding the orchestrator role, update the private handoff under the orchestrate skill before the ledger row and prompt. Record each unit's T3 thread link (or its environmentId plus thread id on a linked host) and any send whose reply is still unread. A t3_thread_send receipt confirms delivery, not completion; role takeover requires the successor's explicit acknowledgment and orchestrator ledger row. Do not resume an existing Desktop task merely to deliver a closeout message.

At session start

Read bun run beep session open (the Claude SessionStart hook prints the same rows). Each live row is a resume candidate: offer them in the first prompt alongside whatever the operator asked for, newest first, and never silently take over a row another live session owns.

When

Every stopping point, not only the end of a task:

• the requested work is done and verified;
• the work is blocked on something only the operator can decide or do;
• a PR reached merge-ready: yes: it left draft at content-final, its checks are green, its threads are answered and its 20-minute review window has elapsed (hand-off, never a merge unless authorized);
• a merge landed and the lane was retired;
• any moment you would otherwise end the turn with a question.

If the harness has no question tool (headless -p runs, a detached job), write the same options as a numbered list with the recommended item first and say which tool the operator should use to answer.

Shape

0. Write the ledger row first. bun run beep session note --state <open|blocked|done> --next "<one line>" [--summary "<one line>"] [--pr <n>] from the checkout you worked in. This is what the next session (yours or a sibling's) sees at start through bun run beep session open; the prompt below is for the operator, the row is for the fleet. yeet sweep and yeet sweep --retire write the done row themselves.
1. Recap, three lines at most, above the prompt. What landed (PR numbers, commits), what is verified, what is not. Numbers in a short table if more than two.
2. One prompt call. Claude Code: AskUserQuestion. Codex: request_user_input. At most four questions per call; most closeouts need one.
3. The question carries the state. One line: clone, lane or branch, PR, where it stopped. The operator must be able to answer without scrolling up.
4. Two to four options, each a real next step. Recommended option first, labelled "(Recommended)", with the trade-off in its description. Always include a "stop here" or "park it" option so a no is one click.
5. Multi-select only when the options are independent (several lanes to retire, several packets to close). Decisions between alternatives stay single-select.
6. Act on the answer in the same turn when it is reversible and in scope; otherwise open the next prompt.

Option vocabulary

Prefer these, so prompts read the same across sessions:

• Continue with <next step> (Recommended) — the work that follows naturally.
• Open the follow-up PR for <x> — when the next step is a separate branch.
• Run /grill-with-docs on <decision> — when direction, not execution, is next.
• Retire the lane / Sweep the clone — post-merge closeout.
• Adopt the <slug> closeout (Recommended) — after a goal's final PR merges: goals completion refresh, then goals set-status in a packet-state PR. A goal is not done, and /goal must not stop, until that PR merges.
• Archive the <slug> packet (Recommended) — after the adoption PR merges: bun run beep goals archive --kind goals --slug <slug> --plan, then --write --expect <planId>, push the archive checkout, then land the tombstone in a packet-state PR; or set retentionReason in the manifest when the evidence must stay on main. Explorations use --kind explorations once graduated or killed.
• Park it: record state and stop — writes the state to memory or the packet and ends the turn.

What it is not

• Not a substitute for /grill-with-docs: direction-setting decisions are grilled first; the closeout prompt carries the remaining frontier.
• Not an exploration exit: an /explore packet that needs the operator switches into /grill-with-docs and keeps going; closeout comes only once the packet is graduated, parked, or killed.
• Not a permission laundering path: an action the harness denied is reported as denied, never offered as an option for the operator to "approve" on your behalf.
• Not a reason to skip the on-disk handoff: packet state, memory, and the yeet inbox are still written before the prompt.