Team Skill

When to use

Use only for an explicit $team request or omx team ... launch that benefits from durable tmux workers, shared task state, mailbox coordination, or worktree isolation. Read AGENTS.md#durable-runtime-invariants-canonical-ssot; do not copy its durable rules into this card.

Inputs and preconditions

• Launch shape: omx team [N:agent-type] "<task>" (for example omx team 3:executor "implement X").
• Require tmux -V, a leader session with $TMUX, and the intended omx executable. In Codex App/plain non-tmux sessions, explain the runtime boundary instead of pretending Team is available.
• Before launch, ground the task in a recent .omx/context/{slug}-*.md; create a concise snapshot when none exists. Include target, evidence, constraints, unknowns, and likely touchpoints.
• Do not launch nested Team runs. Pick worker roles deliberately; use OMX_TEAM_WORKER_CLI=codex|claude|auto or OMX_TEAM_WORKER_CLI_MAP=... only when needed.
• Per-agent reasoning is set by agentReasoning; accepted values are low, medium, high, xhigh, and max. max is passed unchanged and remains capability-dependent. ultra is unsupported and is not an alias for max.
• Invalid configured values use the built-in role-default fallback. An explicit raw -c model_reasoning_effort=... is opaque and wins. When both sources are present, explicit raw reasoning wins over inherited Team reasoning and environment reasoning. Do not downgrade or retry max as xhigh; built-in role defaults remain unchanged.

Operational steps

1. Start the runtime and capture startup evidence: Team started: <name>, tmux target, worker panes, and the leader ACK mailbox.
2. Current Runtime Behavior: runtime-owned task files and APIs are the operational interface; the durable ownership and safety rules remain in templates/AGENTS.md.
3. Runtime creates .omx/state/team/<name>/config.json, manifest.v2.json, tasks/task-<id>.json, worker identities/inboxes, mailbox files, and a dispatch queue. Workers receive OMX_TEAM_WORKER, OMX_TEAM_STATE_ROOT, and OMX_TEAM_LEADER_CWD.
4. Deliver assignments through durable inbox/task state and the CLI API. The worker card defines ACK, claim, transition, mailbox, and idle-status commands.
5. Prefer these machine-readable operations:

   omx team api send-message --input '{"team_name":"<name>","from_worker":"leader-fixed","to_worker":"worker-1","body":"<short trigger>"}' --json
   omx team api read-task --input '{"team_name":"<name>","task_id":"<id>"}' --json
   omx team api transition-task-status --input '{"team_name":"<name>","task_id":"<id>","from":"in_progress","to":"completed","claim_token":"<token>"}' --json

6. Monitor with omx team status <name> --json or omx team await <name> --timeout-ms 30000 --json; inspect mailbox/state files when a worker is blocked or stale.
7. Keep the team running until pending=0, in_progress=0, and failed=0 (or an explicitly acknowledged failure path). Run omx team shutdown <name> only then, unless the user explicitly aborts.
8. Verify shutdown evidence and state cleanup. Do not claim completion while workers are still writing.

Team and Ultragoal

When a leader-owned .omx/ultragoal/goals.json exists, workers return task evidence only. The leader checkpoints with a fresh Codex get_goal snapshot:

omx ultragoal checkpoint --goal-id <id> --status complete --evidence "<Team evidence>" --codex-goal-json <fresh-get-goal-json-or-path>

Team launch remains an explicit separate action; it does not create hidden Codex goals.

Dispatch and recovery

• If dispatch reports worker_notify_failed:<worker>, inspect tmux list-panes, capture the pane, then send one concise trigger and re-check mailbox/state.
• For Claude panes, do not spam Enter while work is active. Confirm runtime status first.
• If a worker reports omx team api ... ENOENT, check whether shutdown or state deletion happened too early; preserve state until all transitions finish.
• For a clean retry, kill only known stale worker panes, remove only the exact stale Team root, and relaunch; never kill the leader/HUD pane accidentally.

Exit and evidence

Report team name, launch command, pane/ACK evidence, task counts, worker verification, failures or blockers, shutdown result, and cleaned state paths. Include any worktree or CLI-map choices and keep final integration/verification in the leader lane.