crew

Crew

A crew is resident agents in Herdr panes: each keeps its context across turns, and the user and other windows reach it by name. This session is the coordinator. It designs the crew for the task in front of it and reshapes it as the work changes; this skill holds the rules every crew keeps and how to seat each model, not a workflow.

Read the user's crew preferences (~/.codex/agent-notes/crew.md, when present) before designing the crew: they set the coordinator's role, the flow of each unit and the review policy, and win over this skill where they differ.

The user decides: whether to open a crew; which models join, when the user names them; the points the user settles personally; anything destructive, outward-facing, or costly beyond what the card allows. Invoking this skill authorizes opening the panes the card lists.

Rules for any crew

• Earn the seat. A resident agent joins when it brings what the coordinator lacks: a context to hold for later, a different model's eye, or work that runs in parallel on separate files. Work that needs none of these stays in this session or goes to a subagent or Workflow. Start from the smallest crew that covers the need; the coordinator can hold a seat itself, as the builder beside an outside reviewer or as one voice at a table. A model the user names gets the job it fits best; when none fits, say so before seating it.
• One coordinator. It holds the goal and finish line, assigns work, accepts results and merges. It keeps the card at .work/crew/<run>/card.md, with .work/ in .git/info/exclude unless the repository already ignores it: goal and finish line, its own pane, each agent's name, model, pane and worktree, who may commit, merge or push, the budget and round cap when they matter, and what goes to the user. Show the card before starting unless the user already said go, and update it whenever the crew changes.
• Stop rules with a floor. A lane's stop rule names a minimum sample and both a count and a rate (more than 3 failures and more than 3% after 30 calls), and the lane classifies each failure before counting it: a provider refusal, a malformed answer and a bug in its own checker call for different rulings. Rules without a floor stopped one lane three times, twice on a single failure or a bug in its own checker.
• Independent lanes. Parallel lanes touch separate files; an agent that edits works in its own worktree and branch. The coordinator orders dependent work and says which lane waits. Worktrees still share the machine: the card gives each lane its own ports or a turn at a shared service or heavy test run, and an agent kills only processes it started, by PID.
• Families for judgement. Seat a different model family where a shared blind spot costs most: a design, the review of a unit whose wrong accept is expensive, a check of the whole run against its goal. Throughput work goes to the model the project's rules name.
• Evidence over reports. A result counts when the coordinator can see it: a commit, a file, a finished check's log. Another family's findings are claims: check each against the code or source and record the ruling beside it before another agent acts on it. When the claims are many, have each carry path:line, check the quotes and numbers against those lines with a script, and read what it flags.
• Paths, not pastes. Every model works on the repository or worktree itself; a copy of the material is only for a model that cannot read it, since copies go stale and hide the rest. Messages carry file paths. An answer longer than a few lines goes to a file the reply names. Residents write at full length; a length limit belongs only to an Agent-tool subagent, whose return lands in your own context.
• Arm a wait. A resident never messages back. After each prompt whose answer you need, arm a wait that notifies you: the prompt with --wait in the background, or a background loop until the expected file is new or rewritten since the prompt and herdr agent get shows idle or done, since a status file a lane reuses across stages already exists (scripts/crew-watch.sh checks both). Size the loop's limit to the work, as one synthesis ran 70 minutes; a loop that reaches it says so, and you re-arm it. Track each wait by the PID you recorded when you armed it, since a pgrep -f pattern also matches the loop's own command line. Tell the user what you are waiting on.
• Address by roster. Reach an agent by a name or pane that herdr agent list or the card shows; when they leave the target open, ask the user.
• Fresh context when it pays. Compact or reset an agent whose context has grown large or whose next work is unrelated. A coordinator whose own context is filling hands over through the card: it records there what is in flight and which waits are armed, stops its own watchers, and the fresh session starts from the card and re-arms them.
• The user at decisions. Bring each decision when it arises, with the options and a recommendation; report progress at milestones, quietly otherwise, naming each running seat and armed wait by what it does or waits for.
• Close what you opened once its work is merged or recorded. Side agents stay until the user retires them. When the user stops a crew, cancel the seats that just started, let a seat near its end finish without new model calls, start nothing new, and record on the card, per seat, what finished, what was cancelled and what is left undone.