Relay

A model-agnostic coding-environment skill for discovering available models and routing one task through frontier planning, workhorse execution, and independent frontier review.

The relay first inventories the models that the active environment can actually delegate to, classifies them from current evidence, and then assigns each phase by capability and cost:

frontier planner
    -> faster, cheaper executor
    -> separate frontier reviewer

For example: Fable plans, GPT-5.5 executes, and GPT-5.6 Sol reviews. Those names are configuration, not requirements. The route can use any models exposed by the host, and deployment is always a separate authorized phase.

Install and invoke

Environment	Project installation	Invocation
Codex	.agents/skills/relay/	$relay
Cursor	.agents/skills/relay/	/relay
Claude Code	.agents/skills/relay/ plus .claude/skills/relay/	/relay

Keep the repository layout intact when installing. Codex and Cursor read the canonical skill directly; the Claude Code entrypoint loads that same implementation with host-compatible explicit-invocation metadata. Restart an environment if it does not detect the newly installed skill.

The skill uses the coding environment's native visible delegation controls with explicit model and reasoning settings. It does not use hidden subprocesses or direct provider calls to manufacture unavailable capabilities.

Before creating any child, Relay announces the role, selected model, tier, effort, and reason for the assignment. Parallel batches are shown as a model-to-role table, and any verified model substitution is disclosed immediately.

Saved model preferences

Users can select models for a single run or save defaults for planning, execution, and review. Project preferences live in .relay/models.yaml; optional user-wide preferences live under the user's Relay config directory. Precedence is one-run choice, project preference, user preference, then automatic suggestion. Relay revalidates every preference against the current environment and can either suggest a valid replacement or stop when a preferred model is unavailable.

version: 1
phases:
  planning: { model: "fable", effort: "high" }
  execution: { model: "gpt-5.5", effort: "medium" }
  review: { model: "gpt-5.6-sol", effort: "high", on_unavailable: "stop" }

Discovery

relay inspects native child-task capabilities, distinguishes parent availability from child selectability, and verifies the actual model assigned to each child. Public provider catalogs and remembered names never prove current availability. See ENVIRONMENTS.md for adapters and MODEL_SELECTION.md for tier evidence and routing gates.

Requirements

The host must expose native controls for selecting and verifying child models. Verification can be returned model metadata or a native contract that guarantees an accepted exact ID is used and rejects unsupported values. Aliases and fallback-capable selectors require resolved-model readback. When the host cannot verify the actual model, Relay stops instead of pretending that model routing happened.

License

MIT. See LICENSE.