Codex runtime mapping

This published workflow is explicit-only through agents/openai.yaml; the proven branch is a bounded local design Arena. Before starting, require at least two authorized, isolated candidates and an available independent judge; retain the same rubric and verification contract. Outside that proven scope, report the workflow as unavailable unless a separate task-specific proof supports it. Use native Codex subagents and isolated paths for parallel candidates, not Cursor Task, Work cloud, or ~/.cursor/rules. In the observed native path, candidates used collaboration.spawn_agent with agent_type=general-worker and the cross-judge used agent_type=reviewer; the exact upstream Cursor roles and automatic trigger are not proven. Keep the same rubric, full-candidate read, cross-judge, base selection, named graft, redesign-on-failure, and verification phases. Follow the standing model router and task authority; record invocation metadata separately from worker self-report.

Arena

Open a todolist with one entry per phase (Frame, Fan out, Cross-judge, Pick, Graft, Verify) before launching anything.

Phase A: Frame

Every candidate gets the same prompt, so the prompt is the contract.

1. State the artifact each candidate is producing.
2. Derive the rubric. State what success looks like for this task, then turn it into 3-6 concrete gradeable criteria. The rubric is the picker's tool in Phase D. Candidates only see the task. Choose available native Codex runners under the standing model-capability-router policy and the operator’s explicit selection. Use structurally distinct prompts or models when comparison needs diversity; record the actual runners.
3. Assign output paths. Each candidate writes to its own location (a git worktree where possible, otherwise /tmp/arena-<slug>/candidate-<n>/), per cursor-principle-separate-before-serializing-shared-state.

Phase B: Fan out