| name | anti-hall-model-policy |
| description | Codex model routing policy for anti-hall work. Use when selecting models for planning, implementation, review, debate, or mechanical execution in Codex. |
anti-hall Codex model policy
| Task shape | Codex model | Effort |
|---|
| Planning, architecture, validation, sensitive review, debate | gpt-5.6-sol | xhigh |
| Implementation from a settled plan | gpt-5.6-terra | medium |
| Mechanical command runner, cheap lookup, repetitive execution | gpt-5.4-mini (default) — gpt-5.6-luna available when 5.6-era capability/cutoff matters (not the default) — gpt-5.3-codex-spark is a distinct, faster/less-capable model, ChatGPT Pro only, not an effort setting of the flagship | low |
Rules:
- Debate roles always use
gpt-5.6-sol.
- Do not use mini models for security-sensitive validation or launch go/no-go decisions.
- Do not run every subtask on
gpt-5.6-sol; distribute by task shape.
- If a model is unavailable or rate-limited, record the limitation and choose the nearest safe fallback. Do not silently downgrade debate below
gpt-5.6-sol.
Claude-side mapping (when coordinating with Claude agents)
When this skill coordinates with anti-hall deadly-loop or ship-it, the Claude-side tier token sonnet resolves to Sonnet 5 (claude-sonnet-5). Claude-side role assignments:
- Reviewer (deadly-loop/ship-it):
model:"sonnet" = Sonnet 5, effort xhigh
- Planning secondary / medium scope:
model:"sonnet" = Sonnet 5, effort xhigh
- Implementation failover (when Codex unavailable):
model:"sonnet" = Sonnet 5, effort high
- Main coordinator, planning top-level, deep debug:
model:"opus" = latest Claude Opus
Cross-model rule: if Codex implements, Claude (Sonnet 5 or Opus) reviews. If Sonnet 5 implements, Codex or Opus reviews. No agent reviews its own implementation.
