fix: close model providers created by Runner

[Excelius-Wang]

Summary

This pull request fixes a resource leak when Runner creates the default model provider for a
single run. Direct Responses WebSocket runs could finish successfully while the provider retained
its cached connection because the existing provider aclose() chain was never called.

Runner now closes only providers it creates implicitly after async, sync, or streamed execution
settles. A caller-supplied RunConfig or model_provider remains caller-owned and reusable, and
cleanup preserves the primary result, exception, and cancellation behavior.

Test plan

• Added lifecycle coverage for implicit and explicit provider ownership, sync and async runs,
  streamed completion, cancellation, repeated cancellation, cancel-before-start, cleanup failure,
  sandbox wrapping, and a real loopback Responses WebSocket connection.
• Ran .agents/skills/code-change-verification/scripts/run.sh: format, lint, mypy, pyright, and the
  full test suite passed (9301 passed, 29 skipped).

Issue number

N/A

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
• If using Codex, I've run /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 2026-08-31T03:24:21.727741Z	99853c0	New commits
🔒 Security Review	✅ Completed 2026-08-31T03:25:41.414262Z	99853c0	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: 55c0e3874f

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".

[chatgpt-codex-connector]

Move provider cleanup policy into run_internal

Move the new ownership normalization and cancellation-resistant provider cleanup helpers into src/agents/run_internal/ and leave only their orchestration calls here. These functions introduce substantive resource-lifecycle policy directly into the public runtime entrypoint, contrary to the repository requirement that new runtime logic live under run_internal; keeping them here further concentrates lifecycle behavior in the already-large run.py.

AGENTS.md reference: AGENTS.md:L190-L191

Useful? React with 👍 / 👎.

[Excelius-Wang]

Thanks — moved the run-config ownership normalization and cancellation-resistant provider cleanup helpers into src/agents/run_internal/model_provider_lifecycle.py. src/agents/run.py now retains only the orchestration imports and call sites, with behavior and coverage unchanged.

[seratch]

Please move the ownership normalization and cancellation-resistant provider cleanup policy out of run.py and into run_internal/, leaving only orchestration calls in the public runtime entrypoint.

[sylvesterkaczmarek]

I reviewed the current head after the lifecycle policy was moved under run_internal. I traced implicit provider ownership through AgentRunner.run() and run_streamed(), then through streamed completion/cancellation in RunResultStreaming: non-streaming cleanup is protected by finally, streamed cleanup is registered against run-loop completion and awaited through the same one-shot cleanup task, and caller-supplied RunConfig / model_provider instances remain caller-owned. _close_runner_owned_model_provider() also preserves caller cancellation while preventing provider-close failures from replacing the primary run result/error. The focused coverage exercises normal async/sync completion, explicit ownership, cancellation/repeated cancellation, cancel-before-start, cleanup failure, and a real Responses WebSocket close. I don't see a blocking lifecycle issue in the current revision.