菜单
Lightweight notification skill — atmux supervisor pings teammates via `/atmux:heads-up <message>` to nudge them about new tasks, cascade unblocks, or inbox updates. Skill just acknowledges; teammate folds the nudge into next idle turn.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | heads-up |
| description | Lightweight notification skill — atmux supervisor pings teammates via `/atmux:heads-up <message>` to nudge them about new tasks, cascade unblocks, or inbox updates. Skill just acknowledges; teammate folds the nudge into next idle turn. |
| argument-hint | <event description> |
Invoked by atmux when notifying a teammate's pane between turns. Examples:
/atmux:heads-up new task t-abc1234 dispatched/atmux:heads-up cascade unblock — claim --next/atmux:heads-up message from reviewer/atmux:heads-up new tell-lead from driver/atmux:heads-up decisions-add/atmux:heads-up flag-addWithout it, every supervisor injection produces Unknown command: /atmux:heads-up errors in the pane (cosmetic but noisy — log spam buries real signal). The skill silently absorbs the nudge so atmux's notification mechanism is clean.
The skill itself does not act. Per the team-member brief in atmux templates, atmux's between-turn injections are "always safe to consume" — pane-state preflight runs upstream (post-ADR-202 Honker substrate, this is event-driven; pre-substrate, it's the legacy supervisor's keystroke gate). Treat the heads-up as a normal nudge: read the args (what changed), fold into your loop on the next idle turn, no special handling.
Read the args ($ARGUMENTS). The args summarise what changed:
new task <id> dispatched → next idle turn, claim or note the taskcascade unblock — claim --next → run atmux claim --next when readynew tell-lead from driver → re-read .atmux/driver-inbox.md ## Open section (the lead's incoming inbox; an earlier proposed rename was REVERTED — canonical filename is driver-inbox.md; driver→lead direction per ADR-215 §D3)new flag <id> / flag-add / flag-resolve → atmux flags list --status opennew reply from <member> → re-read .atmux/lead-outbox.mddecisions-add → re-read .atmux/decisions.md<etype> event → check kanban + inbox for the changed surfaceDo not interrupt in-flight work. If currently mid-task, finish the turn first. The supervisor's pane-state preflight already gated the keystroke — receipt is a hint, not a preempt.
No output required. The skill is a no-op acknowledgement. Don't echo "I received the heads-up" or similar — Claude Code's silent consumption keeps the pane clean.
State files are the source of truth. State + notification are transactional (verb writes state.db row → publishes event → atmux injects), so a missed keystroke can't desync you from kanban truth — re-read state via atmux inbox <member> + atmux task list when in doubt. If you missed multiple heads-ups while busy, just re-query on next idle turn; you'll catch up. The kanban itself lives in .atmux/state.db (SQLite per ADR-126) — never grep the .db directly; always use the verbs.
atmux claim --next, separate verb).It exists purely so /atmux:heads-up <args> is a valid skill invocation rather than an Unknown command error.
Unlike /atmux:whip §8.0, /atmux:sweep §9.5 (formerly the medic operator surface; per ADR-212 the role retired but the marker scheme stayed), /atmux:bau header, /atmux:bruh §7, /atmux:session global, and /atmux:team global — which all mandate an ✅/⚠/🔴 + 👁 attention-marker header on every operator-facing report — /atmux:heads-up deliberately produces no output, attention markers included. Pane silence IS the success state: the injection is a hint folded into the next idle loop, not a turn the operator reads. Echoing any verdict line — even ✅ — defeats the noise-suppression purpose.
If something genuinely warrants the operator's attention as a result of a heads-up (e.g. parsing the args revealed a 🔴-class condition), the receiving agent surfaces it via its own normal channel (lead-queue entry, whip status line, BAU report) — NOT via this skill's output. Keep this skill silent.
.atmux/state.db SQLite store (replaces legacy .atmux/kanban.json + .atmux/inboxes/<m>.json).driver-inbox.md + lead-outbox.md SQLite-canonical with rendered markdown view; clarifies surface vocab used above.Unified whip skill. Verbs — run (autonomous-work nudge loop, default), cadence (tune re-arm interval), watchdog (liveness one-shot). Use /atmux:whip <verb> or `/loop /atmux:whip` for the main loop.
Business-as-usual status check + velocity-fix escalation. Reports commit cadence, rate-limits, kanban, churn per team across the cockpit (or single team from inside a cage). Auto-escalates Dormant teams to lead. `/atmux:bau [hours] [--no-fix]`. Default window 24h.
Bruh-mode unblocker — sweeps pending decisions, blockers, flags, and worktrees and pushes everything forward in one pass. Unblock all, approve all, flip all, merge all (all worktrees we worked on).
Unified session-continuity skill. Verbs — cont, preclear, handoff, stop. Replaces cont, preclear, handoff, full-stop. Use /atmux:session <verb> [args].
Fleet-wide diagnose + complain sweep — runs `atmux doctor` and `atmux status --json` across every enabled team, files complaints on anomalies, and takes structural fixes (rotate lead, clear member, push branch fix). Persisted host-pressure playbook (per ADR-198) is one trigger. Per ADR-077 substrate; manually invoked (the auto-spawned cockpit role was retired per ADR-212).
Send a driver→lead ask via `atmux tell-lead` — durable to .atmux/driver-inbox.md plus best-effort wake-up to the lead's pane. Explains the two listener-absent warnings as expected output, not a failure.