supervisor

Plugin supervisor from marketplace claude-supervisor in bborbe/claude-supervisor. Commands read from the repo's manifest — 2 days ago.

Skills (4)

• asked-ledger
• hold
• open-items
• supervising-workers

Slash commands (94)

• /supervisor:attention-next

  What's next on the attention stack — show open items, answer one in chat, and deliver the answer to the session that asked

• /supervisor:close

  Close ONE finished session — a worker whose task is terminal or deferred, or a goal/topic manager whose page is handed over, ended or parked — idle, and holding no gate. Closes a tab by killing its pane and a headless worker by SIGTERM on its registry pid, running the same guards for both. Refuses i

• /supervisor:fleet-drive

  ONE round of the fleet loop — sweep, diff against the previous sweep, classify every peer (progressing / stalled / parked / done), dispatch the drive leg, act within a strict autonomy boundary, render and re-arm. The same round the fleet loop's wakeup runs, and runnable by hand. Repeats by default a

• /supervisor:fleet-loop

  Arm the stateful fleet loop — invoked ONCE, then every wakeup runs the round `/supervisor:fleet-drive`. Repeats by default at its `--interval` (default 15 min) via ScheduleWakeup; TTS on problems (voice-mode gated); routes to managers. Never re-invoked from a wakeup.

• /supervisor:fleet-reset

  Clear the fleet's remembered state — stop every session, empty the five stores, and (only behind a separate opt-in) strip the task-file links a keep-alive leg resumes. The strip unlinks only a session that is positively ABSENT. Destructive; refuses without a confirm token.

• /supervisor:fleet-status

  Read-only fleet roster — every peer Claude Code session, its live status, and its vault task mapping. Zero side effects. Ends with a blocked-by-you jump list (waiting sessions + clickable jump links).

• /supervisor:fleet-triage

  One pass over every live Claude Code session — classify each into CLOSED / OPERATOR / EXTERNAL / LIVE-IDLE HEADLESS / STUCK / WORKING, close the finished (headless workers included, via `close-worker.py`), drive everything that is not genuinely blocked on something outside the fleet, post a card for

• /supervisor:fleet-verify

  Verify the fleet layer's own contract — five gated checks, read-only, ending in a numbered fix list

• /supervisor:fleet-workers

  Show all supervisor workers and everything awaiting approval

• /supervisor:goal-answer

  Answer a pending worker permission prompt

• /supervisor:goal-drain

  Answer every pending permission by policy and escalate the rest

• /supervisor:goal-drive

  ONE round of the manager loop for ONE subject — reap the finished, nudge the stuck, spawn on the handover, classify, render and re-arm. The same round the manager loop's wakeup runs, and runnable by hand. Name via $1, or detected from the session when omitted; the branch (topic vs goal) is auto-dete

• /supervisor:goal-loop

  Run the manager loop for ONE topic OR ONE goal — watch its declared set + task files (default 15 min, `--interval`), detect problems, TTS the human (voice-mode gated), print a status table, recommend new sessions and session closes. Name via $1, or detected from the session when omitted; the branch

• /supervisor:goal-spawn

  Spawn a supervisor worker for a task

• /supervisor:goal-status

  Read-only goal-or-topic snapshot — ONE goal's declared task set (from `24 Goals/`) or ONE topic's declared goal set (from `23 Topics/`) *(operator's vault; not shipped with this plugin)*, its sessions, a status table with bucket icons, and a blocked-by-you jump list (clickable jump links). No vault

• /supervisor:goal-verify

  Verify one topic or goal and suggest the fixes — seven gated checks, read-only, ending in a numbered fix list

• /supervisor:inbox

  The operator's approval inbox — every task at phase: todo, and the three verbs that move one

• /supervisor:jump

  Jump to a Claude Code session's WezTerm tab. Bare = newest session needing attention; or by tab id, tab:<N>, pane:<N>, or a title substring. The executor for the `/supervisor:jump <pane-id>` lines that the fleet-status, fleet-loop, goal-status and goal-loop commands print but never run — those comma

• /supervisor:open-a2a

  Open a name through the a2a-worker-controller — resolves it with /supervisor:open's own gates, then creates every worker on the controller, the backend chosen by the task's `target:` frontmatter. Manager-tier, like /supervisor:open. Name via $ARGUMENTS; optional --task/--goal/--topic, --vault, --fir

• /supervisor:open

  Open the right thing for a name — resolves it to a task, goal or topic, then jumps to its live session, resumes its recorded session, or starts a new one. Batch mode opens every task flagged for today. Name via $ARGUMENTS; optional --task/--goal/--topic, --vault, --color, --flagged, --first-step.

• /supervisor:read-guides

  Read all supervisor guides for context before working — session tiers (which commands each session may run) plus the fleet-surface spec

• /supervisor:ready

  Ready ONE named task and open it — run the audit → repair → re-audit loop in a sub-agent until the audit returns zero hard-gate failures, then act on the Gate 2 verdict: open via `/supervisor:open`, or escalate naming the gap. Cross-tier — runnable by the fleet manager and by a goal/topic manager. I

• /supervisor:reset

  Re-discover this manager's state from disk rather than delete it. The reset contract is the vault's Manager Session runbook § Reset; this file owns the invocation and the report.

• /supervisor:restart

  Restart ONE named session — worker or manager — in a single verb: kill its registry pid, resume the same session id in a new tab in its own role colour, and tell it what changed (a manager is told to re-arm its loop). Takes a session id (or a pid). Refuses on a dirty worktree the target owns alone,

• /supervisor:steps:act
• /supervisor:steps:adopt
• /supervisor:steps:attention-watcher
• /supervisor:steps:auto-compaction
• /supervisor:steps:cause-verify
• /supervisor:steps:confirm-candidates
• /supervisor:steps:covered-subject-invariant
• /supervisor:steps:drive
• /supervisor:steps:escalate
• /supervisor:steps:flagged-open
• /supervisor:steps:fleet-round-tail
• /supervisor:steps:fleet-triage-act
• /supervisor:steps:fleet-triage-load-path
• /supervisor:steps:fleet-triage-render
• /supervisor:steps:fleet-triage-rules
• /supervisor:steps:fleet-triage-state
• /supervisor:steps:fleet-triage-sweep
• /supervisor:steps:fleet-triage-tier
• /supervisor:steps:fleet-triage-verify
• /supervisor:steps:goal-classify
• /supervisor:steps:goal-closure
• /supervisor:steps:goal-coverage
• /supervisor:steps:goal-drive
• /supervisor:steps:goal-gates
• /supervisor:steps:goal-handoff
• /supervisor:steps:goal-parked
• /supervisor:steps:goal-repair
• /supervisor:steps:goal-resolve
• /supervisor:steps:goal-spawn
• /supervisor:steps:goal-sweep
• /supervisor:steps:goal-under-target
• /supervisor:steps:goal-unowned
• /supervisor:steps:keepalive-decision-parks
• /supervisor:steps:keepalive-stood-down
• /supervisor:steps:keepalive-topicless-goals
• /supervisor:steps:keepalive
• /supervisor:steps:manager-reap
• /supervisor:steps:needs-input
• /supervisor:steps:open-items
• /supervisor:steps:open
• /supervisor:steps:output-shape
• /supervisor:steps:own-parks
• /supervisor:steps:persist
• /supervisor:steps:propose-manager
• /supervisor:steps:re-arm
• /supervisor:steps:readings
• /supervisor:steps:reap
• /supervisor:steps:round-liveness
• /supervisor:steps:round-quiet
• /supervisor:steps:round-readings
• /supervisor:steps:round-self-improvement
• /supervisor:steps:round-tail
• /supervisor:steps:rules
• /supervisor:steps:service-guard
• /supervisor:steps:sweep
• /supervisor:steps:topic-keepalive-outcome
• /supervisor:steps:topic-resolve
• /supervisor:steps:topic-sweep
• /supervisor:steps:voice-notify
• /supervisor:steps:worker-channels
• /supervisor:stop
• /supervisor:topic-drive
• /supervisor:topic-loop
• /supervisor:topic-spawn
• /supervisor:topic-status
• /supervisor:topic-verify
• /supervisor:who-needs-me
• /supervisor:who-needs-my-help
• /supervisor:worker-drive
• /supervisor:worker-restart

Subagents (17)

• fleet-drive

  Perform the fleet sweep's drive leg — read the digest render at the `payload=` path the caller passes, split its `parked` rows into revive (observed routine-continue closer, no verified blocker) and blocked (verified blocker), suppress re-nudges through a session-keyed ledger, close a nudged row sti

• fleet-flagged-open-actor

  Perform the fleet round's flagged-open leg — invoke `/supervisor:open --flagged` so every row the operator marked for today is opened, print every `REFUSED` line verbatim, open nothing a refusal names, and record the leg's `operator-flagged` count in the round's digest — then return the ≤ 1 KB verdi

• fleet-keepalive-actor

  Perform the fleet round's keep-alive leg — restart lapsed topic managers, resume due park watches, nudge watchers, resume dead workers; return the ≤ 1 KB hand-back. Router row 5b of /supervisor:fleet-drive.

• fleet-open-actor

  Perform the fleet round's open leg — run `open-queue.py --json` once, write its rendered block to this round's payload, run every `selected[].action` verbatim so each admitted row opens through `/supervisor:open`, and return the ≤ 1 KB verdict hand-back. Dispatched by router row 16 of /supervisor:fl

• fleet-sweep-reader

  Compute the fleet sweep's read half (Steps 0b–3 of /supervisor:fleet-loop) — read the four channels and the open-items ledger, diff against the previous snapshot, run the orphan reverse index, find collision and unmanaged-topic candidates, classify every session, persist the next snapshot through fl

• gate-relay-read

  Read a blocked tab worker's pane and return a compact per-pane summary of what it is waiting on, so the manager can ask the operator without paying the pane-dump cost in its own context. Use when the attention feed or a roster read shows a tab worker blocked and its live question is needed before su

• gate-relay-send

  Deliver an operator's already-given answer into a blocked tab worker's pane, verifying the gate is still open first and handing over a jump link instead whenever the pane holds a modal, a wizard, or a production-touching question. Use when the operator has answered a batched manager question in the

• goal-drive

  Perform the worker sweep's act leg for ONE subject — reap the finished, nudge stuck or error-marked workers, run the auto-resume gate on confirmed orphans. Reap runs BEFORE drive, always. Dispatched by `/supervisor:goal-drive` (operator, by hand) and by `/supervisor:goal-loop` (every tick, after its

• goal-spawn-reader

  Compute the spawn decisions for a goal manager's handed-over `To open` / `To resume` rows — read each row's task file, apply the approval, hold, role, dependency, target, shipping, mode and cap gates, resolve the vault and cwd, and return the ≤ 1 KB per-row call shape plus the payload path. Dispatch

• goal-sweep-reader

  Compute the worker sweep's read-and-render half — read a tracked task set, classify into the bucket set the vault's own runbook declares, extract full session-id sets with `python3 scripts/session-ids.py` (by field, never by line start, never by shape), flag orphan candidates and live collisions, re

• goal-verify

  Run the seven gated checks for ONE subject — a goal or a topic — and return the report plus a numbered fix list. Dispatched by `/supervisor:goal-verify` (operator, by hand). It owns the seven per-step forks and the report shape; it mutates nothing.

• goal-wrangler

  Run the routine approval loop over headless supervisor workers so the manager session is not occupied by approvals. Use when headless workers (spawned with interactive:false) are running and their permission prompts should be answered by policy rather than by the operator.

• round-readings-reader

  Compute a manager round's readings — the four verdict families (liveness, in-flight, attention feed, external wait) for every tracked row — by running scripts/round-readings.py over the caller's tracked set, recording the result on the round's receipt, and returning the ≤ 1 KB verdict hand-back plus

• topic-drive

  Perform the TOPIC sweep's act leg for ONE topic — keep each declared goal's manager alive (jump / resume / restart / leave / crash-loop), then rank and name the goals that should take the free N slots. Dispatched by /supervisor:topic-drive (operator, by hand) and by /supervisor:topic-loop (every tic

• topic-sweep-reader

  Compute the TOPIC sweep's read-and-render half — read the topic page's declared goal list, and for each goal read its status, phase, success-criterion tally and the state of the goal manager running on it, then render the goal table. Use when a topic command (/supervisor:topic-loop or /supervisor:to

• topic-verify

  Run the sixteen-row check for ONE TOPIC — eleven disk-backed page-consistency checks, one coverage read and four advisory rows — and return the report plus a numbered fix list. Dispatched by /supervisor:topic-verify (operator, by hand). It owns the per-row forks and the report shape; it mutates noth

• worker-drive-reader

  Compute the read half of one /supervisor:worker-drive round — the session diagnosis (the measured window, loop check, anchor re-verify, drift check) and the drive table (every open item, its evidence rating, blocker, lever and owner) — and return the ≤ 1 KB verdict hand-back. Use when /supervisor:wo

Hooks (7)

• PermissionRequest

  python3

• PreToolUse

  python3

• UserPromptSubmit

  python3

• Stop

  python3

• Stop

  python3

• Notification

  python3

• SessionEnd

  python3

Similar plugins