[u7-u7]

Change Summary

Closing an SDK agent stream after consuming a chunk currently closes only the outer async generator. Python's async for does not automatically close the iterator it consumes. As a result, the nested retry/step/model generators remain suspended: the HTTP model stream stays open and run_loop's existing interrupted-round persistence is deferred to eventual async-generator finalization.

Use contextlib.aclosing at the three ownership boundaries (run(stream=True) → run_loop → retried step) so explicit consumer shutdown reaches the model stream before aclose() returns. This does not change retries, payloads, or normal streaming output.

Validation

• New regression exercises step, run_loop, and the public run(stream=True) against a real OpenAI-compatible SDK stream using httpx.MockTransport. All three failed on the base (model stream outlived its consumer), and pass after the fix; run-loop/public-run cases also verify interrupted-round persistence is invoked.
• A focused retry-wrapper regression verifies that closing runs the active attempt's finalizer before returning and does not start another retry.
• pytest -q tests/utils/test_retry_classification.py tests/agent/test_llm_io.py tests/agent/test_partial_round.py: 59 passed, 15 skipped.
• Flake8 for changed production files, YAPF for changed sections, and git diff --check: passed.

Not tested: live providers, a full initialized tool/MCP session, the entire repository suite, or other Python versions. Optional Anthropic cases were skipped because its dependency is not installed. Tests use the real stream/agent loop with unrelated setup mocked; these are local results, not upstream CI.

AI assistance

Implementation and automated verification were performed with OpenAI Codex. No human-review attestation is made.