fix: reject premature Chat Completions stream EOF

[jbeckwith-oai]

Summary

This pull request fixes premature Chat Completions stream termination being reported as a successful, truncated response. For the official OpenAI endpoint, streams now require a provider finish reason for choice zero before completing; otherwise they raise ModelBehaviorError before tool execution. Interrupted requests remain counted in generation spans even when no usage trailer arrives.

Tool-call buffering preserves real terminal choices instead of synthesizing completion at EOF. Third-party adapter behavior, nonempty length completions, content filtering, and usage trailers remain supported.

Test plan

• Added offline HTTP SSE tests through the production AsyncOpenAI parser and Runner.run_streamed, covering premature text/tool/empty EOF, normal text and tool completion, buffering, usage, and third-party compatibility.
• Confirmed the text EOF regressions fail on the unchanged baseline.
• Focused checks: 206 adapter tests, 29 reasoning/SSE tests, and 125 EOF/tracing tests passed.
• Full verification: formatting, lint, Mypy, Pyright, parallel tests, and serial tests passed. Native macOS sandbox tests are skipped by the mandated nested-sandbox setting and remain enabled in CI.
• Two independent reviews of the final complete diff found no actionable issues.

Issue number

No linked issue.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• I've completed the repository-required independent review before submitting this PR

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-28T01:17:27.635122Z	af0ea10	New commits
🔒 Security Review	✅ Completed 2026-09-28T01:17:26.603732Z	af0ea10	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 93233dac21

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[markstuart-oai]

Reviewed af0ea10395d6961656ea4b796c20e22f5a5121e9. No actionable findings. Official OpenAI streams now need a real terminal choice before final completion; buffering preserves that evidence instead of fabricating it. The EOF path retains request usage for tracing and closes the stream, while third-party EOF behavior and valid length/tool completions remain supported. The real-parser Runner regressions cover failure before tool execution and successful terminal paths.

All 22 hosted checks passed on this commit. Validation here was source inspection plus hosted CI; I did not run repository tests.