[RongJie G (CorgiBoyG)]

Motivation & Context

A streaming AgentLoopMiddleware run fires turn-scoped after_run providers (after_run_once_per_turn = True, e.g. CompactionProvider) from a finally block, so an iteration that raises — or an early stream close — still runs them with the stale final response from a prior iteration. The non-streaming path keeps that call outside its try/finally, and a normal non-loop run only fires result hooks on successful finalization. Streaming therefore diverged from both, silently advancing once-per-turn finalization (e.g. the compaction/summarization turn boundary) off a superseded response on a failed turn.

Description & Review Guide

• What are the major changes? In the streaming loop generator, fire the turn-scoped after-providers only on a clean loop exit. The three normal-stop returns become break, a completed_cleanly flag is set after the loop body, and the finally call is gated on that flag. An exception, or an early aclose()/GeneratorExit, leaves the flag unset and skips the call. Only the loop-iteration token cleanup remains unconditional in finally.
• What is the impact of these changes? Normal completion is unchanged (should_continue stop, max_iterations, approval escape hatch, MiddlewareTermination all still fire turn-scoped after_run exactly once). Exception and early-close paths now skip it, matching the non-streaming path and the non-loop run contract.
• What do you want reviewers to focus on? That every clean stop still reaches completed_cleanly = True before the generator is finalized, and that no exception/early-close path sets it.

Contribution Checklist

• I have read the CONTRIBUTING.md.
• The code builds and tests pass locally.
• I have added tests to cover my changes (deterministic red-on-main / green-on-fix: streaming exception, streaming early-close, non-streaming exception control; verified by reverting only the production file).
• All unit tests pass locally (tests/core/: 7056 passed, 19 skipped, 0 failed).

AI Assistance

This is an AI-assisted contribution. The defect was found by source audit; AI assisted with drafting the regression tests, the fix, and this description. The root-cause analysis, the empirical contract verification, and the red-on-main / green-on-fix validation were reviewed and confirmed by me.

Related Issue

Fixes #9817

Draft pending maintainer confirmation of the scoped fix direction.

[Copilot]

🟢 Approval recommended

No active review issues were identified for the change restricting streaming turn-scoped finalization to clean loop exits.

No open issues 🎉

What changed in this PR

Addresses #9817 by aligning streaming loop turn-scoped after_run behavior with non-streaming runs, preventing finalization with a stale response after an exception or early stream closure.

Changes:

• Gates turn-scoped providers on clean loop completion while preserving unconditional iteration-token cleanup.
• Converts normal-stop returns to breaks so successful exits still trigger finalization.
• Adds regression tests for streaming exceptions, early closure, and the non-streaming exception control.

File	Description
python/​packages/​core/​tests/​core/​test_harness_loop.py	Adds regression coverage for failed and early-closed loop runs.
python/​packages/​core/​agent_framework/​_harness/​_loop.py	Restricts streaming turn-scoped finalization to clean exits.

_{Review effort: Balanced}