Claude — all non-Fable Claude 5 / Opus 4.x:
- claude-opus-4-8 (new strong — the current top Opus)
- claude-opus-4-7 (previous strong Opus, still used by mrrlin bridge)
- claude-sonnet-5 (weak, quick)
- claude-haiku-4-5 → model=claude-haiku-4-5-20251001 (weak, fast — cheapest)
Fable-5 explicitly excluded per operator preference.

Codex/OpenAI — every current model surface incl. mini variants:
- codex (default — account picks whatever it exposes)
- codex-pro (gpt-5.4-codex — strongest reasoning, requires API-key plan)
- codex-gpt-5.4-mini (VERIFIED live on ChatGPT-plan accounts — 10k tok in
  <30s round-trip. Weak tier, cheap.)
- codex-gpt-5.3-codex (previous-gen strong reasoning)
- codex-gpt-5.6-terra (newest, may need Codex CLI upgrade)

Registry ids renamed to include model designator (claude-opus-4-8 vs
old `claude-opus`) so pick decisions are unambiguous — no more "which
opus did we get". Old `claude-opus` / `claude-sonnet` ids purged from
state.json in same commit; consumers should re-audit.

Total agents: 31 (was 28).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>