CONTRIBUTING.md: GPT-5.4-only AI policy is no longer current SOTA

[flyinghanger]

Context

CONTRIBUTING.md (L12-20) currently states:

AI-Generated Code Policy (Vibe Coding)

If you are using AI tools to generate code ("Vibe Coding"), please note:

We only accept code generated by the GPT-5.4 model.

Pull requests containing code generated by other AI models (including other GPT versions, Claude, Gemini, or any other LLM) will be rejected.

Issue

GPT-5.4 is no longer current SOTA — OpenAI has shipped successors, and other vendors (Anthropic, Google, etc.) ship competitive frontier models in the same tier. As written, the policy:

1. Excludes the current SOTA, including OpenAI's own newer releases — the pin is to a specific past version, not "the best available".
2. Isn't reliably detectable. There is no robust way for a reviewer to verify which model produced a given diff, so the policy primarily filters honesty rather than quality.
3. Penalizes good-faith disclosure. A contributor who states the model used (e.g. fix(claude): propagate real contextWindow from SDK to web status bar #720, where I disclosed Claude in the PR description) is auto-rejected, while undisclosed AI-assisted code goes through unflagged. This inverts the incentive the policy presumably wants to create.

Suggestions

Three options, lightest to heaviest, in case any of them is useful:

1. Drop the model whitelist. Lean on the existing quality bar already in CONTRIBUTING (focused PRs, tests, no mega-PRs). Quality of output is what matters, and that can be judged from the diff directly.
2. Switch to disclosure-only. Authors disclose the model in the PR description; reviewers judge the code on merit. Keeps transparency without the enforceability problem.
3. Rolling allowlist. Maintain a short list of "currently acceptable" SOTA models, refreshed periodically (e.g. quarterly), so the policy stays calibrated to the actual frontier.

Happy to send a docs PR for whichever direction lands well, or to close this if the restriction is intentional and I'm missing the rationale (e.g. a specific past incident with non-GPT models that motivated the policy).

[github-actions]

Thanks for raising this. Verified: CONTRIBUTING.md currently hard-pins AI-generated code to GPT-5.4 and says PRs generated by other LLMs will be rejected.

I did not find a repository workflow that enforces the contributor-side model whitelist automatically. The visible PR checks run typecheck/tests, and the PR review workflow runs Codex via openai/codex-action with a configurable model/default gpt-5.2-codex. So from the codebase/config I can verify this is currently documented policy text, not an enforcement path found in the checked workflows.

This looks like a maintainer policy decision rather than a code bug. The concrete docs change would be in CONTRIBUTING.md if maintainers choose one of the proposed directions.

Relevant code:

• CONTRIBUTING.md:12 - AI-generated code policy section starts here.
• CONTRIBUTING.md:16 - GPT-5.4-only requirement.
• CONTRIBUTING.md:18 - states PRs from other LLMs will be rejected.
• .github/workflows/test.yml:15 - PR test workflow installs dependencies.
• .github/workflows/test.yml:16 - PR test workflow runs typecheck.
• .github/workflows/test.yml:26 - PR test workflow runs tests.
• .github/workflows/codex-pr-review.yml:105 - PR review uses openai/codex-action.
• .github/workflows/codex-pr-review.yml:116 - review model is configurable, defaulting to gpt-5.2-codex.

---

HAPI Bot