feat: Codex model fallback — probe and cache working models (v0.12.6.0)

[amoeri]

Summary

If your ChatGPT subscription doesn't include the model Codex defaults to, every
/codex invocation silently failed and fell back to Claude reviewing itself. The
whole point of Codex is an independent second opinion from a different AI. That was broken.

Probe-and-cache script (bin/gstack-codex-model):

• Probes Codex CLI for a working model, tries user's config default first
• Falls through gpt-5.3-codex → gpt-5.2-codex on "not supported" errors
• Caches result for 24h (~6ms reads, ~6s probe on cache miss)
• Configurable via gstack-config set codex_fallback_models

11 invocation sites updated across 4 source files + 10 regenerated skills:

• codex/SKILL.md.tmpl — Step 0.5 model probe + error handling
• scripts/resolvers/review.ts — 4 sites (OH voice, adversarial, structured review, plan review)
• scripts/resolvers/design.ts — 3 sites (design lite, sketch, outside voice)
• autoplan/SKILL.md.tmpl — 1 probe + 3 phase commands

Uses -c model="$CODEX_MODEL" which works for both codex exec and codex review
(discovered during review that codex review doesn't accept -m).

Commits

1. feat: add gstack-codex-model probe-and-cache script — new bin/ script
2. feat: wire model probe into all codex invocation sites — templates + resolvers
3. chore: regenerate SKILL.md files from updated templates — 10 generated files
4. chore: bump version and changelog (v0.12.6.0)

Pre-Landing Review

Reviewed via /autoplan (CEO + Eng phases). 12 decisions made, 10 auto-decided,
2 taste decisions confirmed by user (inline two-liners over helper abstraction,
manual verification over automated bash tests).

Key findings addressed:

• Use -c model= not -m (codex review doesn't support -m)
• Don't parse ~/.codex/config.toml — probe the default directly
• Atomic cache writes (temp + mv)
• Parse specific "not supported" errors, don't cache transient failures

Test Coverage

• 378/378 skill validation tests pass
• Probe script tested manually: broken model → fallback works, cache hit ~6ms, --probe bypasses cache

Test plan

• Skill validation tests pass (378/378)
• Skill parser tests pass
• SKILL.md regeneration clean (29 files)
• Manual: broken model config → probe finds gpt-5.3-codex
• Manual: cache read ~6ms, probe ~6.6s
• Manual: --probe bypasses cache

🤖 Generated with Claude Code

[time-attack]

@16francej — recommend closing this PR as superseded by #1628. The branch hard-codes outdated model choices and may confuse authentication or network failures with unsupported models. Preserve contributor credit for the probing idea, but carry forward only bounded capability checks with explicit failure classification.

[capy-ai]

Thanks for this, @amoeri. This is now covered on main: Codex model resolution now goes request, then GSTACK_CODEX_MODEL, then config.toml, then the default, with a probe for unusable models (shipped in v1.91.16.0, #3014). Closing as superseded; if you still see the problem on the latest gstack, please open an issue with ./setup --status output.