Arena

Arena

Use independent candidates when one early shape would constrain the answer. The user or an active workflow must authorize subagent use.

1. Define one artifact and three to six task-specific criteria. Give every candidate the same task and keep the rubric with the parent.
2. Read $CODEX_HOME/pstack-models.toml when present. Otherwise use deep, general, and fast from the Codex runtime mapping. Two candidates are enough for a bounded design; use three when model-tier diversity materially helps.
3. Give writers separate worktrees or output directories. Launch the candidates together through Codex collaboration tools.
4. Wait for all viable candidates. Record dropouts without retry loops that do not change the conditions.
5. Read every candidate. Score each criterion, choose the most maintainable coherent base, and record the decision.
6. Inspect losing candidates for one or two compatible improvements. Integrate them deliberately; do not average incompatible mental models.
7. Verify the synthesized artifact against the same success criteria.

Use a separate deep judge only when the candidates are close, the stakes justify another pass, or the user explicitly asks for multi-model judgment. The parent still owns the final choice.

Return one artifact plus a short synthesis note naming candidates, scores, selected base, grafts, rejected ideas, dropouts, and verification.