fix(core): preserve OpenAI OAuth model context limits - #53080

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

The following comment was made by an LLM, it may be inaccurate:

The needs:issue label appears to be a false positive: this PR has contained Fixes #47646 since opening. The standards run checks only closingIssuesReferences, which is empty for this PR targeting v2 rather than the default dev branch; the template-compliance check passes.

@rekram1-node, could you review this OpenAI provider fix when you have a chance, and help link #47646/remove the incorrect label? I chose you based on your recent changes in this provider. The production change removes the two-line blanket limit override; regression coverage checks long-context models, a fast alias, and a smaller-context model. The regression failed before the fix and passes afterward; the focused suite, core typecheck, and full bun run check pass locally.

Additional runtime verification on PR head dc65500: built the Linux x64 CLI with Bun 1.4.2 and ran its private server with a ChatGPT OAuth connection. After catalog initialization, /api/model reports {context: 1050000, input: 922000, output: 128000} for gpt-6.1-sol, gpt-6-astra, gpt-6-astra-fast, gpt-5.6-sol, and gpt-5.5, all using https://chatgpt.com/backend-api/codex. The five provider tests also pass on this head.

This verifies the effective metadata in the compiled server; it does not claim a million-token inference test.

Could a maintainer triage this for an OpenAI provider review, clear the false-positive needs:issue label, and enable any required upstream CI? This targets v2 and references #47646; local verification is listed above.

The key review question is whether ChatGPT OAuth should preserve model catalog limits instead of imposing the Codex CLI budget. #44821 reports a controlled GPT-5.6 Sol request accepting 921,858 input tokens through OAuth—evidence that the blanket 272K input cap is too low for that account/route, though not proof of identical capacity for every account. This PR preserves each model’s existing limits rather than introducing another universal limit.

If maintainers require account-specific limit discovery or an opt-in approach instead, guidance on that decision would help move the fix forward.