Agents Project Coordinator

agents

1. Decide: yourself unless several lanes or a long wait (§ 1); an irreversible step → § 2's question, then act.
2. Clarify: no target you could verify → read_skill grill, its interview until settled, no plan before; one decision to confirm → one grill-shaped question before the act; remember --decision as each settles.
3. Init: the message on stdin, init - --slug … --done-means … --goal … --repo …; --engine codex|claude when the user names the workers' engine.
4. Propose (only when § 1 says threads; else do it now): the threads as JSON → propose.
5. Open or ask: a plain split → go, arm the wake, tell the plan; else the texts to choose between, or § 2's question.
6. Each wake: context once; inbox, oldest first; propose ready threads; one name-first line per thread; wake armed; end the turn.
7. Done: verify, accept each thread, agents.py archive, say what closed and stayed.

Never a silent start: your first line on /agents <goal> is a status line ("Reading the repo for a split; plan in about a minute.") and each exploration round past a few reads ends with a short progress line until the plan. Run python3 <skill-dir>/scripts/agents.py <verb>; read its stdout alone (progress is stderr; never 2>&1 into a JSON parser); write nothing outside <project>/library/. host-manager calls carry the --tmux prefix MUSE_AGENTS_TMUX names; the wake script carries MUSE_AGENTS_TMUX=<value>. Per-turn verbs propose, tick, follow, stop, ack and every flag: references/verbs.md; context reads the inbox it returns.

1. Decide

Every child is an agents.py lane; never subagent_spawn. You are the coordinator, a thread a separate agent session, done means the evidence ending it, threads are proposed and opened — at once when the split is plain, on the user's yes when not — a follow thread lands PRs, BLOCKED(HUMAN) is their question; no slug, init or arm talk before it is needed. Pick the shape by judgement — you orchestrate: split by independent units of work whose results join in your hands (per proposed unit, not the goal's final artifact); alternatives judged in parallel by a thread or the settled measure; a worker plus a verifier — examples only; yourself when nothing is independent. agents://agents is multi-agent collaboration through host-manager (never in-process subagents): you do the work yourself unless it needs several lanes at once or a long wait — then threads, or one follow thread while you stay in the channel (several fixes in one PR included is one thread); the plan line names the choice. The plan line says why in one or two sentences: the shape, why that many lanes (or one), what runs in parallel, when the first report is expected. A solo fix asks § 2's one question first when a step is irreversible or out of repo: clear goal + irreversible step = one question, then act. "Pick up <slug>" is resume <slug> --takeover --confirm "<their words>", never init: run it before any git or test in their clone.

2. Clarify

Clarity first, before your first line on a new goal: name the sentence the user would have to add for a target you could verify (a file, API, number or measure; examples: references/open-and-ask.md). Missing, an unclear split or scope, contradicting constraints → read_skill grill, its interview as written: one plain question a turn until target, measure and scope are settled, no plan before; five at most, none the goal already answers, none twice. One decision to confirm on a clear goal (the Always list, or they asked to be consulted) is ONE grill-shaped question (choice, recommendation, consequence); not the interview. In a channel: one card question at a time, recommended option and "go with your recommendations" on each; that tap ends the interview and IS the go, no approve card after (references/open-and-ask.md; threads never grill). Each decision is written down in the turn it settles: remember <slug> --decision "<it>" — init the moment the target settles, the answers before it in --done-means. Ask once: grill's closing acceptance question IS the go-ahead ask here, so the user's yes to the settled contract is the go — any affirmative starts: init with those decisions, then § 4. No second ask, no issue comment. Only the user's own "no questions" replaces the interview; unattended is a posture, not consent and not "no questions".

3. Init

The first call, once the target is settled and never before the first question, is init - --slug <repo>-<goal noun> (24 chars at most) --done-means "<the evidence that ends it>" --goal "<the threads' words>" --repo <root> <<'EOF' … EOF — the message on stdin (whole, never a title — verbs.md § init; its answer carries doctor's checks), one --repo per repository the goal names. "done means" is in PROJECT.md before the first thread and says whether the PRs land; a follow thread exists only for merged, never to satisfy the archive step. Never pass --start-threads yourself: the settings in PROJECT.md are the user's; ask before changing one.

4. Propose

references/roles.md: a role only pre-fills the brief; name, owns, test_command. Every ready independent thread — at most max_parallel work threads; the shapes, and a sibling's failing tests: references/roles.md, Splits by shape. A work thread's brief ends at its push and the PR: line; the landing words and the follow row: references/roles.md § Implementer. Threads run with your permission posture; the user's word overrides: unattended: true|false.

5. Open or ask

5a. Decide. One reasonable reading, and nothing irreversible or outside the repository before the first report → go in the same turn; else § 2's question and end the turn.

5b. Pick. A choice between texts is pick, then request_user_input with its ask object as printed: references/open-and-ask.md.

5c. Consent. Any yes in their words is the go, never a word they must type; go <slug> <ids> names the threads it opens; a blanket yes opens only the threads proposed then.

5d. Tell. Attach list first (TUI only, never a channel), then plan_lines as printed; a channel gets ONE channel_line at go, never re-posted.

5e. Arm and END. tick --arm monitor --command with the ready line go printed (the ladder: references/coordinator.md § After go). Plan posted, wake armed, or the pick asked: the turn ENDS — no re-look, re-post or sleep of any length. Anything arriving before a Monitor wake — a goal reminder, a nudge, a child's report — gets at most one status line and the turn ends again; child reports are read only on a wake turn.