Execute

Leading words: current first, same-family ok, slug shield, auto-pick stays distinct, this chat does the work.

Workers research. This chat does the work. There is no agent-count flag — pick 2 or 3 from the job (default 3; 2 if the request is small, local, or mechanical).

Operating contract

• Do not auto-start this skill. The user must ask for it.
• Current-for-all is the default. Same family may occupy more than one slot when they chose that.
• Auto-pick is the only path that still requires distinct families.
• Do not invent a model name. Use the live list this harness exposes.
• Never write a spawn identifier in a question, option, chat roster, worker prompt, or take label. Map family → live-list name only in the spawn tool's model argument. Current → inherit (or omit model).
• Workers are read-only. A worker that edits files has failed; discard that take.
• This chat writes after the takes. Do not send writes back to workers.
• If this host cannot spawn a child at all, stop. Auto-pick that cannot gather two families: offer Current-for-all or deep-deliberation.

Runtime state

EXECUTE_STATE
host: cursor | claude-code | opencode | copilot | codex | other
agents: 2 | 3
roster: unset | Current,OpenAI,…
takes: 0/N
next_action: one action only

Pipeline

Copy these steps into the todo list verbatim. Skip with skip: <reason>.

1. Detect the host. Read references/hosts.md. Build the live spawnable model list. Current-for-all does not need two families.
2. Set agents to 3, or 2 if the request is small, local, or mechanical. If the user named models (or current), use that list (2–5) instead of auto-pick. One named model fills every slot.
3. Resolve roster: models the user named, or Current for all, or auto-pick, or assign each worker. Announce with display names before spawning.
4. Spawn one read-only worker per roster slot. Same prompt. Parallel. Read references/worker-prompt.md.
5. Render labeled takes ([Current], [OpenAI], …). Do not write yet.
6. This chat merges with [display name] attribution, does the work, proves the real path.

Roster

If the user already named 2–5 models that exist on this host, use them. Same family is allowed. If they named one model, or said current / inherit, fill every slot with that choice.

If the user said auto-pick (or chose it), run auto-pick for agents families from references/hosts.md, announce with display names, then spawn.

Otherwise ask:

prompt: Which models should research?
options:
- Current model for all workers (Recommended)
- Auto-pick distinct families
- I'll assign each worker
- Say this in plain English
allow_multiple: false

Option ids: current, auto, assign, simplify. Never live-list names.

Say this in plain English is a meta question: explain Current vs auto-pick with display-name examples (never spawn ids), then re-ask. Do not spawn yet.

If they assign each worker, ask one question per slot (not a multi-select). Each slot's first option is Current model (Recommended), then one option per remaining family using display names and fam-* ids from references/hosts.md. Same family may be chosen on more than one slot.

No structured question tool → numbered text fallback.

Write

• Resolve consensus and divergence with [display name] attribution.
• Keep valuable minority observations. Reject weak claims explicitly.
• Do the work here. Do not delegate writes back to workers.
• Prove on the real path (or closest executable check). skip: if the surface cannot be driven.

Reply contract

Playbook: execute
roster: Current, OpenAI, Grok

Then the labeled takes, what shipped, and how you verified.

Delegate failure

1. Retry one transient spawn failure.
2. Never invent a missing take.
3. Fewer than two successful takes → stop, ask whether to retry or abort.