codex: 🐛 Resolve the small and medium tiers to gpt-5.6-luna

[yasyf]

Summary

The codex small and medium tiers resolved to gpt-5.4-mini. On a codex CLI signed in with a ChatGPT account, every call to that model fails at once with:

ERROR: {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'gpt-5.4-mini' model is not supported when using Codex with a ChatGPT account."}}

As a result, every captain-hook llm_nudge routed to codex's small tier fails, including cc-skills show_skills.wall_of_text and general.detours.

Fix

small → gpt-5.6-luna:low and medium → gpt-5.6-luna:medium, changed in the core tier table (rust/spawnllm-core/src/route.rs) and the Python backend's CodexCliBackend.models. large stays gpt-5.5:medium. The codex-agent and codex-effort-suffix plan vectors now use the same model name, so no vector names a model codex rejects.

Model matrix

I ran each candidate through spawnllm's own CodexCliBackend with codex 0.154.0 signed in with a ChatGPT account, using a one-word prompt and max_attempts=1:

model	ChatGPT sign-in	codex's fetched catalog	bundled catalog (API key)
gpt-5.4-mini:low / :medium	400, above (isolated and not)	absent	hidden, supported_in_api
gpt-5.6-luna:low / :medium	ok, 3–7s (isolated and not)	listed, supported_in_api	listed, supported_in_api
gpt-5.6-terra:low	ok, 3.0s	listed, supported_in_api	listed, supported_in_api
gpt-5.5:low	ok, 4.0s	listed, supported_in_api	listed, supported_in_api
gpt-5.3-codex-spark:low	ok, 3.7s	listed, not supported_in_api	absent
gpt-reserve:low	ok, 3.7s	hidden	absent

"Fetched catalog" is ~/.codex/models_cache.json, which codex downloads for the signed-in account. "Bundled catalog" is the model list compiled into the codex binary. gpt-5.6-luna is the one model both catalogs list and mark supported_in_api that codex describes as "Fast and affordable", so it is the only candidate fit for both small tiers. No API key is available on this machine, so API-key auth rests on the catalog metadata, not a live call.

Tests

tests/test_backends.py::TestModels::test_codex pins the new table. Against the old codex.py it fails, and with the change it passes. capabilities.json regenerates with the new tiers, and the Go and Rust hosts replay it.

https://claude.ai/code/session_015yr73szXYEmbjtNVLxnetK