[Kevin Dong (KevinAndrewDong)]

Motivation and Context

Dropping a message or raising from an intervention handler can leave InProcessRuntime.stop_when_idle() waiting indefinitely. The runtime has removed the envelope from the queue, but its unfinished-task count is never decremented on those early-return paths.

Description

Call task_done() before returning from the six send, publish, and response interception branches that drop a message or handle an exception. Normal dispatch retains its existing completion accounting.

Add 12 parameterized regression cases covering those three stages with a drop marker class, a drop marker instance, a raised exception, and normal pass-through.

Validation

Baseline: 483ffb487a33a1644886992d56599feae3b69b76.

• Before the fix, runtime/serialization plus the new regression: 9 failed, 23 passed on both Python 3.12.14 and 3.13.15. Each failure reaches the idle-stop deadline.
• After the fix, runtime, orchestration, and group-chat suites: 175 passed, 1 skipped on each interpreter. The skip is an existing upstream timing-related skip.
• These tests use normal complete-package imports and upstream tests/conftest.py. No namespace bootstrap, model API, or external service is involved.
• Ruff 0.13.1 lint/format, Python AST, whitespace checks, and clean patch application pass.
• The entire repository suite, type checking, and pinned pre-commit hooks have not run. Local command-line network access is blocked, so the repository's Ruff 0.9.6 hook and other missing hook dependencies could not be installed.

Command, from python/:
python -m pytest tests/unit/agents/runtime tests/unit/agents/orchestration tests/unit/agents/test_group_chat -rs

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the complete pre-submission checks pass
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

AI assistance

OpenAI Codex assisted with source inspection, the regression tests, the six-line fix, and local verification.

[Copilot]

🟢 Approval recommended

The six targeted completion calls preserve normal dispatch accounting, and the regression cases cover every affected stage and outcome with no identified blocking issues.

No open issues 🎉

What changed in this PR

Fixes queue completion accounting in the Python in-process agent runtime so intercepted messages do not leave stop_when_idle() waiting indefinitely.

Changes:

• Marks dequeued messages complete on all six drop and exception paths across send, publish, and response interception.
• Adds 12 parameterized regression cases covering both drop marker forms, exceptions, and normal pass-through.

File	Description
python/​tests/​unit/​agents/​runtime/​test_intervention_queue.py	Adds regression coverage for idle shutdown after interception.
python/​semantic_kernel/​agents/​runtime/​in_process/​in_process_runtime.py	Completes queue accounting on terminal interception paths.

_{Review effort: Balanced}

[Kevin Dong (KevinAndrewDong)]

@microsoft-github-policy-service agree