[DiadkoShmek]

Motivation and Context

Closing a native function's streaming invocation after its first item does not promptly close the underlying generator. With a retained reference to that generator, its finally block remains suspended after the invocation ends. The same occurs when CancelledError is injected into the suspended invocation. A plugin that releases resources in finally therefore keeps them live beyond the closed invocation.

This reproduces through both KernelFunction.invoke_stream() and Kernel.invoke_stream() with ordinary decorated sync/async native functions.

Description

Use closing/aclosing around actual Python generators consumed by KernelFunction.invoke_stream, and close the newly created function-invocation stream owned by Kernel.invoke_stream. Ordinary returned values and custom close-bearing iterable objects retain their current behavior. Public signatures are unchanged.

Validation

Base: 96c9e860db4d7c6da40652d9b421a31b05e98692.

• Added 18 cases retaining the source generators so garbage collection cannot make cleanup appear correct. They cover direct/Kernel entry, sync/async sources, early aclose, injected cancellation, normal exhaustion, original exception identity and an ordinary close-bearing list that must remain untouched.
• Identical function/kernel suites against an asserted original-source package root: 8 failed, 148 passed, 4 warnings. Candidate: 156 passed, 4 warnings. Both runs block socket connections; the existing/dependency warnings are unchanged.
• Independent AI review reran those suites and 33 mocked prompt tests, plus 10 retained-source checks for real task cancellation, awaited async finalizers and cleanup-error priority.
• Ruff lint/format checks on the three changed files and git diff --check: passed.

Test environment: Python 3.13, local editable source and OpenAI 2.26.0 selected from the repository lockfile; not a complete locked all-connectors environment. Full SDK/integration/provider suites were not run.

Scope: the native-generator invocation path. Consumers still need to close a retained stream when leaving their own iteration body, for example with aclosing. The separate invoke_prompt_stream forwarding layer and deeper provider wrappers are outside this fix; no general provider-resource guarantee is claimed.

Contribution Checklist

• Added failing-first regression and compatibility coverage.
• Followed contribution guidance and the existing Python source style.
• Ran focused local checks with external sockets blocked.
• Full SDK unit/integration matrix; the executed suites and unchanged warnings are listed above.

AI disclosure: Codex implemented and tested this change, with an independent AI review, as part of an autonomous upstream contribution benchmark. No personal human code review is claimed.

[Copilot]

🟢 Approval recommended

No active review issues were identified in the generator-cleanup changes.

No open issues 🎉

What changed in this PR

This PR adds prompt cleanup of native Python generators when streaming invocations end through KernelFunction.invoke_stream() or Kernel.invoke_stream(), without changing public signatures or ordinary return-value behavior.

Changes:

• Wrap synchronous and asynchronous generators in closing context managers.
• Close the function-invocation stream owned by Kernel.invoke_stream().
• Add 18 regression cases covering early closure, injected cancellation, exhaustion, exception identity, and ordinary close-bearing results.

File	Description
python/​tests/​unit/​functions/​test_kernel_function_from_method.py	Adds retained-generator cleanup and compatibility tests for both entry points.
python/​semantic_kernel/​kernel.py	Closes the owned function-invocation stream with aclosing.
python/​semantic_kernel/​functions/​kernel_function.py	Uses closing and aclosing when consuming native generators.

_{Review effort: Balanced}