feat: add OpenAI Codex provider metadata

[opencode-agent]

Summary

• add an openai-codex provider for ChatGPT subscription models served through the Codex backend
• inherit provider-agnostic OpenAI model metadata while publishing Codex-specific 400K context / 272K input / 128K output limits
• allow OAuth-only providers to declare an empty environment-variable list instead of inventing an API-key credential

OpenAI corrected the GPT-5.6 Codex window to 272K in Codex 0.144.6. Keeping this metadata separate preserves the direct OpenAI API models' 1.05M context window and gives harnesses a canonical route-specific source.

Sources:

• https://github.com/openai/codex/releases/tag/rust-v0.144.6
• https://github.com/openai/codex/blob/rust-v0.145.0/codex-rs/models-manager/models.json

@rekram1-node, this is the model-catalog half of the GPT-5.6 Sol compaction failure Kit reproduced through ChatGPT OAuth.

Verification

• bun validate
• bun test packages/core/test/schema.test.ts
• catalog generation tests covering metadata inheritance and schema projection pass; the full file retains a pre-existing open-weight metadata failure unrelated to this change

Fixes #3183

Requested by: @kitlangton (Kit via Slack)

[github-actions]

Action items

• [medium] [violation] providers/openai-codex/models/gpt-5.5.toml:4 - Check: Provider models must author cost (AGENTS.md required fields; cost is provider-specific and not inherited via base_model). Why: All four new Codex models omit [cost]. Subscription/coding-plan peers (e.g. kimi-for-coding, zai-coding-plan, minimax-coding-plan) publish explicit zero pricing; omitting cost leaves catalog consumers without a provider-authored price. Action: Add a provider-local [cost] on each model—typically input/output/cache_* = 0 for ChatGPT subscription usage, or verified non-zero rates if Codex bills differently—and cite what the values represent.
• [medium] [possible mistake] providers/openai-codex/models/gpt-5.5.toml:12 - Check: experimental.modes must reflect controls this inference route actually accepts. Why: Every model copies OpenAI API service_tier = "priority" under [experimental.modes.fast]. That is a paid OpenAI API body field; this provider is the ChatGPT/Codex OAuth backend (https://chatgpt.com/backend-api/codex). Closest subscription-style peers (e.g. GitHub Copilot GPT entries) do not advertise this mode. Publishing it can make harnesses send an unsupported body on this route. Action: Verify Codex accepts service_tier: "priority" on this API; if not, drop the fast mode (or replace it with a Codex-documented control).

[Cretezy]

Currently, the following models are offered (on Pro plan): gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark.

I believe we may need to introduce a way to gate models per-plan, as spark models are Pro-only.

5.6 Sol, Terra, and Luna offer low, medium, high, xhigh, max, and ultra. Also also likely none, even if not exposed through Codex. I believe ultra is also Pro-only.

For 5.5/5.4/5.4-mini/5.3-codex-spark, low, medium, high, and xhigh are offered. None also probably.

I think keeping this to 5.6 models makes sense though, as those are what most users will want to use.

[github-actions]

Closing this pull request as stale because it has not been updated in 30 days. Feel free to reopen it or submit a new pull request if the work is resumed.