GPT-5.6 Relay

Turn the user's task into a relay of persistent Codex threads: a Sol coordinator plans, then creates Terra or Luna child threads for phases they can complete reliably. Each child receives a concrete artifact and returns evidence to its parent. The model names describe capability tiers, not fixed job titles.

Invocation of this skill explicitly authorizes the user-visible child threads required by the route, including their model and effort settings. The Sol coordinator owns the whole outcome, keeps the canonical task state, verifies every handoff, and reports the route actually used.

Preflight

1. Restate the requested outcome, acceptance criteria, constraints, allowed mutations, and deployment authority. Infer ordinary implementation details, but do not infer permission for destructive or external actions.
2. Confirm the Codex app exposes project listing, thread creation, thread reading, and background thread messaging with model and effort controls. If it does not, show the proposed route and stop; do not replace persistent threads with hidden subprocesses.
3. For repository work, list projects and resolve the current workspace's project ID before creating a child. For a general task, use a projectless target.
4. Record the starting state. In a Git repository, capture the branch, revision, and dirty files so the relay can distinguish its work from pre-existing changes.
5. Classify the work by uncertainty, not apparent size:
   • Clear: the solution and finish line are already known.
   • Judgment-heavy: the implementation is tractable, but tradeoffs or failure modes need thought.
   • Open-ended: the problem, architecture, or safe path must be discovered.

Preflight is complete when the finish line is checkable and every proposed external side effect is authorized.