Maestro

Orchestration stance for a main agent that oversees delegated work instead of doing it. Three skills:

• conduct — the boss-not-doer stance for an implement → review → address pipeline.
• patient-waiting — the zero-token discipline for waiting on anything.
• your-cue — a read-only briefing of every live agent session, one cue per workstream.

Install

claude plugin marketplace add jtsternberg/claude-plugins
claude plugin install maestro@jtsternberg

Invoke as /maestro:conduct, /maestro:patient-waiting, and /maestro:your-cue in Claude Code, or $maestro:conduct, $maestro:patient-waiting, and $maestro:your-cue in Codex. Bare names are prose identifiers only.

Skills

conduct

You are the boss, not the doer: your cycles go to choosing what happens next, writing work orders, verifying results, and talking to the human — edits, searches, test runs, and reviews all happen in delegated agents. Each pipeline phase gets a fresh agent in a fresh session, because an implementer that reviews its own work and a reviewer that addresses its own review both lose the skeptical context that made the phase worth running. Fresh means defeating the session-routing cache too — dispatch the next phase with --fresh — or the “reviewer” is silently the implementer resumed.

Work orders carry what a fresh context can’t cheaply rediscover: verified facts (including known inaccuracies in the source material), hard constraints as behaviors rather than vibes, every environment trap a predecessor already hit, tripwires armed with the evidence to self-settle them, and any charter text embedded verbatim. Reports come back as claims, not conclusions — check the diff, PR state, and CI yourself before telling the human a phase succeeded. When the human has an action, end with Next for you: <the single action>, never a menu. Next for you: nothing means work is still active — implementation, review, or waiting on an armed callback that will report — and the agent will relay when the human is needed. Completed handoffs state the outcome, any scheduled pending job, and that the agent is idle and ready for the next instruction. A future timer alone does not mean current work continues; completed responses need not end with nothing.

Decisions the human owns get a plain-language section in /tmp/maestro/decisions-<session-id>.md, written before the message that asks the question — what happened, why it matters, the options as the exact tokens to type, each with a one-letter alias (`ok` (k)) so one keystroke answers, the recommendation, what happens next either way, with every GitHub ref a link labeled owner/repo#N. A Next for you: that asks for a decision carries the answer tokens, a clause naming the subject, and the ledger path with its anchor as a separate token (… .md - #decide-92) — a path fused to an anchor opens nothing — because still owe "fold" or "defer" is unanswerable to a human who left four hours ago. A decision bundling numbered recommendations also accepts N: <change>, overriding item N alone. An answered decision leaves Open and collapses to a one-line Decided entry with the date, its decide-<slug>, and the answer — the list that is both the audit trail and the decisions half of a handoff.

It binds to whatever is installed: hotline to dispatch into another workspace so that repo’s own CLAUDE.md and skills load, cmux-cli to give each phase a visible, resumable surface the human can step into. With neither, it falls back to in-process subagents for quick mechanical work and headless sessions launched in the owning repo for substantial phases.

Triggers when the session’s job is overseeing delegated agents — orchestrating a multi-agent workstream, running boss mode, delegating a pipeline — rather than doing the work.

patient-waiting

A waiting ladder you never skip down: a background bash until loop first, the Monitor tool second but only inside the deadline it reports back — Claude Code 2.1.272 accepts persistent: true and still caps the task at 30 minutes, so accepted is not honored — and if both keep dying, stop and hand the loop to the human rather than falling through to model-in-the-loop polling. The rule underneath it all: never machine-poll a human. If the event is human-triggered — a review submitted, a doc approved, “when I’m ready” — they can close the loop by speaking, and a scheduled model wake adds cost and nothing else.

ScheduleWakeup loops are reserved for machine-paced state neither watcher can see, and even then come with three backstops: max three quiet iterations, no off-hours polling, and a quiet-poll count in every reschedule reason so drift stays visible. The skill exists because a killed watcher once got “recovered” into an hourly self-reschedule that ran for a week — roughly 140 full-context premium-model turns spent confirming nothing had changed.

Triggers before setting up any poll, watch loop, recurring check-in, or ScheduleWakeup loop — on “check in on X”, “watch for Y”, “tell me when Z”, “poll”, “monitor this” — and whenever a background watcher was killed and you’re about to work around it.

your-cue

A read-only briefing across every live agent session, answering the question a list of running panes never does: what does each workstream need from you now. Catch-up is the input; the cue is the product. Every workstream ends with Your cue: — a concrete verb (answer, approve, review, review and close, resume, clarify, close, bury) or Your cue: nothing — <reason>, so “nothing needed” is a conclusion stated rather than a line omitted — and the whole briefing ends with a single prioritized Next for you:, never a menu. A cue to answer or clarify carries the subject of the decision, and conduct’s decisions-ledger path plus its separate anchor token when one exists — a briefing read hours later cannot be acted on from answer tokens alone.

It takes no arguments and sweeps whatever is reachable: graveyard candidates --json --no-verdict as the unified fast path, or herdr agent list plus cmux tree/sidebar state as the portable one, enriched only where a row is missing a visible name or status. When just one host answers it stops and asks rather than briefing half the fleet as if it were the fleet, and a partial run the user approved carries a Coverage section. Hotline callees nest beneath their caller, filtered to sessions actually in the live inventory, because the call registry is append-only and unpruned; most rows are stale.

Every item names a place the user can find with their eyes and keys — fable overseer watch PRs — the "agentic-dev" window › ⌘1 agentic-dev › left pane, tab 3/3 — built by the bundled scripts/locate.sh from one cmux tree snapshot. cmux windows have no title, so a window is named by the workspace it is showing. Sessions join to tabs on the exact surface, then on a unique title match; anything else is reported unlocated rather than guessed. Items sharing a workspace group under one sub-header.

Every row the host doesn’t report as working gets a bounded transcript summary — --window 8 --max-chars 8000 — about five sessions per cheap subagent, all batches dispatched at once (it asks first above about 60 sessions). Each subagent routes to the model-invocable sessions-catch-up and reports three sentences plus a STATE (complete, partial, blocked, waiting on you), so the briefing itself never reads a transcript. Tabs titled hotline: count as callees even when the registry has no matching row, and a callee the user has typed into since its work order is flagged human-driven hotline callee with their last typed prompt, so an orchestrator asks before sending it more work. Everything it runs is a read: no focus stealing, no keystrokes into a live REPL, no cleared notifications or scrollback. Burial is a cue, requiring both a finished transcript and Graveyard’s buryable; execution waits for the user.

Triggers when the user wants to know what their agents are doing, what is waiting on them, or what finished while they were away — a morning briefing, a status sweep of delegated work, a catch-up across workspaces.