sweep

name	sweep
description	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).
argument-hint	[run | once | dry-run]

/atmux:sweep — fleet-wide diagnose + complain (per ADR-077 substrate)

Single entry point for the cockpit-tier diagnosis loop. The auto-spawned cockpit role was retired per ADR-212; the probe substrate persists and is now operator-invoked via this slash command.

• run (default, bare) — main loop turn. Reads ~/.atmux/cockpit.json, sweeps each enabled team, triages, acts.
• once — single turn, do NOT re-arm. For ad-hoc dispatch from the operator.
• dry-run — full sweep + report what it would do, but take no action and write no complaints. For testing.

Empty args → run. This preserves the /loop /atmux:sweep contract.

Verb dispatch

case "${1:-run}" in
  run)            ;;   # main loop turn — see §Run
  once)           ;;   # one-shot, no re-arm
  dry-run)        ;;   # full sweep, no side effects
  help|--help|-h) echo "Usage: /atmux:sweep [run | once | dry-run]"; exit 0 ;;
  *)              echo "Unknown verb '$1'. Usage: /atmux:sweep [run | once | dry-run]"; exit 1 ;;
esac

---

§Run — main loop turn (default)

Hourly diagnosis-and-prevention sweep. Fires every 3600s by default — anomalies that warrant a fleet-wide sweep are rare; tighter loops burn Opus tokens on idle sweeps. Operators wanting a different cadence can tune via the re-arm step below.

How to run

• Start: /loop /atmux:sweep — dynamic mode. The skill self-paces via ScheduleWakeup(3600) at the end of each turn.
• Stop: reply stop on the next turn, or exit the running pane.
• Restart after /clear: /loop /atmux:sweep again.

The expected runtime location is any operator-driven pane with cockpit-tier visibility (typically the cockpit superdriver REPL or a dedicated sweep-loop pane). Running it from inside a per-team cage is undefined — the role assumes it can see every enabled team's state from the host.

Instructions

Each time this skill fires in run mode, do the following. Keep responses terse; goal is action and durable complaints, not narration.

1. Load the brief. Read ${CLAUDE_PLUGIN_ROOT}/skills/sweep/sweep-prompt.md and follow it verbatim. It is the single source of truth for sweep behaviour.

   If ${CLAUDE_PLUGIN_ROOT} is unset (legacy install), fall back to ~/.claude/plugins/atmux/skills/sweep/sweep-prompt.md.

2. Verify environment. Confirm ~/.atmux/cockpit.json exists. If absent, exit cleanly — no cockpit, no role.

3. Execute the brief. Run the per-team sweep, triage, investigate, decide authority, act, log to complaint box.

   Eternal-improvement fallback: if the per-team sweep finds zero anomalies AND zero open complaints AND every team is shipping (commit-cadence green per BAU verdict), trigger one cycle of /atmux:bruh §0.7 in driver-scope per team in scope — file ONE [improve P3] task per the heuristics there (tech-debt grep / ADR §OQ / coverage gap / aged doctor warn / lint sweep / stale memory). Sweep's hourly cadence + low-frequency anomaly profile makes it the natural place to inject background improvements without burning whip-rate tokens. Skip if operator freeze in lead's driver-inbox last 24h. Skip if CPU/RAM throttle active. Log under terminal report as ♻️ eternal improvement: <team> · t-XXXXXXXX <title>.

4. Re-arm. At the end of the turn:

   ScheduleWakeup(delaySeconds: 3600, prompt: "<<autonomous-loop-dynamic>>", reason: "sweep re-arm")
   

   Default 3600s (1 hour). The cache-miss-vs-amortize tradeoff (see ScheduleWakeup docs) favours hourly here because anomaly investigation is genuinely sparse — the cache miss is amortized across an hour of "no work" instead of being paid every 5 minutes.

   once mode skips this step.

§Run notes

• Not a daemon. Sweep dies with the pane session. /clear or exit kills the loop. Re-arm with /loop /atmux:sweep.
• Action authority is real. Per ADR-077 §D3 sweep may rotate leads, clear members, cycle cages, push fixes to atmux on its own branch + open PR, and modify ~/.atmux/cockpit.json. Hard limits in §What it must NOT do (in the brief).
• Action-before-log audit trail. Every action is written to the complaint box BEFORE it executes. If the action turns out to be wrong, the audit trail survives.
• Discord uses [sweep] prefix. Operator-facing reports gate on the attention+verdict marker scheme; ban on unprefixed pings still applies.
• Operator-facing report format. All three modes (run, once, dry-run) emit reports using the attention+verdict scheme (👁 / ✅ / ⚠ / 🔴 / ℹ). The discipline + per-team derivation rules live in the brief at sweep-prompt.md §9.5; mirrors /atmux:whip §8.0, /atmux:bau header, /atmux:bruh §7, /atmux:session global, /atmux:team global, /atmux:tell-lead, and /atmux:budget.

---

§Once — single turn, no re-arm

Operator-triggered one-shot. Same logic as run but skips the ScheduleWakeup re-arm at step 4. Useful when the operator wants sweep to do a sweep right now (e.g. after manually adding a new team to the roster, or post-incident before re-enabling the loop). Also the canonical entry point for the ADR-198 host-pressure playbook — when the host is under memory/IO pressure, the operator runs /atmux:sweep once to fire the 5-step playbook surface as one turn.

/atmux:sweep once

---

§Dry-run — full sweep, no side effects

Executes the complete brief — read inbox, sweep teams, triage, investigate — but stops short of acting. No complaints written, no atmux send writes, no tmux send-keys, no commits or pushes. Reports findings to the calling pane only.

Used for operator-side spot-checking after a topology change, or for first-time setup when the complaint box (per ADR-077 §F2) hasn't shipped yet on a given install.

/atmux:sweep dry-run

---

Status of the deferred follow-ups

This skill exists. The ADR-077 §F1-F4 deferreds remain tracked in atmux's kanban:

• F2 (complaint box SQLite schema + atmux complaints verb) — until F2 ships, sweep logs complaints inline to its lead-queue + Discord. The brief's "log to complaint box" steps degrade to "log via Discord ping with [sweep] prefix".
• F3 (atmux send __sweep__) — when shipped, members can ping sweep with heads-up. Until then, members surface signals via team Discord pings; sweep reads Discord history out-of-band.
• F4 (P0 send-keys runbook) — folded into the brief's §Authority levels.
• F5 (status verb sweep surface) — operator-side; doesn't change sweep's own behaviour.

Cross-references

• ADR-077 — original cockpit self-healing role (this skill's substrate).
• ADR-133 — historical rename of the cockpit role superdoctor → medic (to avoid collision with the atmux doctor verb).
• ADR-212 — retirement of the medic cockpit-role auto-spawn; probe substrate + this slash command persist as on-demand surfaces.
• ADR-198 — host-pressure playbook (one trigger inside /atmux:sweep).
• ADR-217 §D2.1 — naming rationale (/atmux:sweep vs /atmux:superdoctor//atmux:medic).