[Bug] GPT-6 Astra is missing from the Codex OAuth catalog

[endcycles]

What happened?

A local models.json entry gets gpt-6-astra working through ChatGPT OAuth in Prime 0.9.1. We verified a normal CLI response and a real ipython file-read round-trip at xhigh, not just whether the model appeared in the picker.

Reproduction

With ChatGPT OAuth configured and no custom Astra entry:

$ prime-agent model list gpt-6-astra
No models matching "gpt-6-astra"

Expected: an account with Astra access can select it and resolve it for RLM subagents.

The remaining gap

The explicit openai-codex list omits Astra. Discovery also sends client_version=0.147.0. Our read-only check with the same account returned HTTP 200 in both cases:

• 0.147.0: Astra absent.
• 0.153.3: Astra present.

PR #2057, checked at 1b1c4d1ef2e8, adds Astra for other providers but leaves both OAuth details unchanged.

I'd like to contribute a small follow-up: the Codex entry and effort map, a compatible discovery version, focused regression tests, and the required changelog fragments. Account filtering would stay intact. I'd rather fix the picker and subagent paths together than leave Astra working in one place and missing in the other.

Direct chat and the file-read tool call passed locally. RLM execution has not been tested; 0.153.3 is a tested discovery version, not a claimed minimum.

Would you welcome a PR with that scope?

Environment

• Affected areas: AI providers and models; Coding agent and CLI.
• Prime Agent: 0.9.1.
• Installation: npm-layout package; original install command not recorded.
• OS: macOS 26.6.2 (25G83), arm64. Node v25.9.0.
• Shell/terminal: reproduced through non-interactive subprocesses; no TUI-specific failure claimed.

[api-moose]

Thanks for merging PR #2057. I hit the same missing Codex/ChatGPT OAuth option in Prime Agent beta after it merged. Astra appears through other providers, but not through the subscriber path.

Checking the merged main at 3b51ce33054f271b496dce351d97f167b081d595 confirms:

• codexModels in packages/ai/scripts/generate-models.ts still omits gpt-6-astra.
• The GPT-6 mandatory-reasoning rule excludes openai-codex-responses; the new Codex entry also needs to suppress off and minimal while exposing the supported efforts.
• Codex discovery still sends client_version=0.147.0. Your comparison with 0.153.3 indicates that the entry and discovery version should be fixed together so RLM/model discovery can resolve Astra too.

I checked open catalog PRs #2042 and #1928; neither adds the Astra Codex entry. A focused follow-up from current main, referencing #2057 and this Discussion, seems appropriate. It should include generated catalog output, account-filtering and reasoning regressions, and changelog fragments, while keeping unrelated catalog work separate.

Would maintainers welcome that follow-up, or is it already planned elsewhere? Happy to coordinate here to avoid duplicate work.

[api-moose]

CC @xeophon

[snimu]

Solved in #2074
Thanks you for the report!