You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
.NET: Fix reasoning remnants and handoff bounce-back - #9174
When an assistant response contains both reasoning and a handoff function call, handoff filtering removes the function call and its result but retains a reasoning-only message. The receiving agent then sees that message after the user's request. The final report in #7384 identifies this deterministic filtering defect as a contributor to silent turns when providers return reasoning summaries.
The same issue also confirms a separate same-turn bounce-back defect: when control returns to an agent, its pending handoff result is delivered without a fresh request to act on. With service-managed history this can reduce the entire request delta to one tool result. This PR now addresses both defects.
Description & Review Guide
What are the major changes? Drop reasoning-only remnants of filtered tool messages. Track each participant's pending handoff call across checkpoints and, on return without new user input, reissue the latest user request with a fresh message identifier. Add model-free filtering and bounce-back regressions covering local/service-managed history, reasoning, streaming/non-streaming, checkpoint restore, and ordinary tool continuations. Document the behavior. The net472 CI correction also keeps the preexisting stale-waiter test's blocking lock off the thread pool.
What is the impact of these changes? Returning agents receive an actionable request while retaining their existing session, ConversationId, and required handoff call/result pairing. New user input is not replayed again; participant-local replays never enter the shared workflow transcript. Filtering still preserves standalone reasoning, unfiltered tools, visible text, and original messages. There are no public API changes, and production background-agent code is unchanged.
What do you want reviewers to focus on? Matching reactivation to the participant's own pending handoff result, distinguishing genuine user input from approval responses, and preserving session state and transcript boundaries. The checkpoint controls rebuild the workflow and verify both replay and no-replay paths. This fixes the confirmed handoff payload defects rather than retrying or inventing text for arbitrary empty provider completions.
Local validation, run from dotnet\ without a live model:
dotnet format 'tests\Microsoft.Agents.AI.UnitTests\Microsoft.Agents.AI.UnitTests.csproj' --no-restore --include 'tests\Microsoft.Agents.AI.UnitTests\Harness\BackgroundAgents\BackgroundAgentsProviderTests.cs' --verbosity minimal
Passed.
Before the corresponding fixes, both reasoning-enabled filtering cases failed, and all eight local/service-managed, reasoning-on/off, streaming/non-streaming bounce-back cases reproduced an empty completion. All now pass. The full solution unit-test suite and live-provider scenarios were not run.
This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.
Drop reasoning-only remnants only when tool content was filtered, preserve unfiltered content, and cover the model-free filter and mocked workflow paths.
Co-authored-by: Copilot <[email protected]>
Copilot-Session: c2b8751b-fffc-4069-ad5b-ebb5d3d4335d
The reason will be displayed to describe this comment to others. Learn more.
MAF Automated Review — Iteration 1
Result: No findings Scope: full PR (1 commit(s)): 890da44564a0 Model:gpt-5.6-sol
Overview
The PR narrowly removes reasoning-only remnants when handoff or tool content was actually filtered, while preserving standalone reasoning, visible text, unfiltered tools, and the None behavior. Per-invocation state, clone-before-mutation behavior, and unit plus end-to-end coverage constrain the change to handoff reception. No publishable Critical, High, or Medium issue was established.
Reviewed the supplied pull-request change set across correctness, security/reliability, architecture, and failure behavior.
No publishable findings remained after source verification for this scope.
Keep the blocking runtime-lock holder off the thread pool and await continuation startup after releasing the lock, preserving the stale-waiter race assertions and existing timeout bounds.
Co-authored-by: Copilot <[email protected]>
Copilot-Session: c2b8751b-fffc-4069-ad5b-ebb5d3d4335d
Vincent Biret (baywet)
changed the title
.NET: Fix reasoning-only handoff message filtering
.NET: Fix reasoning remnants and handoff bounce-back
Oct 7, 2026
This new string comparison conflicts with the explicit repository rule in dotnet/AGENTS.md:43, which requires string.Equals with a StringComparison. Use an ordinal comparison for the call ID.
This issue also appears in the following locations of the same file:
The change tracks each participant's pending handoff result and uses call-ID matching, fresh-input detection, participant-local state, and broad streaming/checkpoint tests to constrain when bounce-back recovery runs. Those guards prevent unrelated handoffs and ordinary tool continuations from triggering the recovery path. The residual risk is that recovery represents continuation by submitting the original user command as a new request, which can cause a returning agent to repeat side effects it already performed.
Reviewed the supplied incremental change set across correctness, security/reliability, architecture, and failure behavior.
1 verified finding remained after source verification (1 medium) across 1 file. Details are attached to the affected lines below.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
documentationUsage: [Issues, PRs], Target: documentation in the code base and learn docs.NETUsage: [Issues, PRs], Target: .NetworkflowsUsage: [Issues, PRs], Target: Workflows
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
GitHub Actions
Inactive
agent-framework-automation
Bot
added
.NET
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Motivation & Context
When an assistant response contains both reasoning and a handoff function call, handoff filtering removes the function call and its result but retains a reasoning-only message. The receiving agent then sees that message after the user's request. The final report in #7384 identifies this deterministic filtering defect as a contributor to silent turns when providers return reasoning summaries.
The same issue also confirms a separate same-turn bounce-back defect: when control returns to an agent, its pending handoff result is delivered without a fresh request to act on. With service-managed history this can reduce the entire request delta to one tool result. This PR now addresses both defects.
Description & Review Guide
Local validation, run from
dotnet\without a live model:dotnet build 'src\Microsoft.Agents.AI.Workflows\Microsoft.Agents.AI.Workflows.csproj' --no-restore --tl:offdotnet test --project 'tests\Microsoft.Agents.AI.Workflows.UnitTests\Microsoft.Agents.AI.Workflows.UnitTests.csproj' -f net10.0 --no-restore --filter-query '/*/*/Handoff*/*'dotnet test --project 'tests\Microsoft.Agents.AI.Workflows.UnitTests\Microsoft.Agents.AI.Workflows.UnitTests.csproj' -f net472 -c Release --no-restore --filter-query '/*/*/Handoff*/*'dotnet test --project 'tests\Microsoft.Agents.AI.Workflows.UnitTests\Microsoft.Agents.AI.Workflows.UnitTests.csproj' -f net10.0 --no-build --no-restoredotnet test --project 'tests\Microsoft.Agents.AI.Workflows.UnitTests\Microsoft.Agents.AI.Workflows.UnitTests.csproj' -f net472 -c Release --no-build --no-restoredotnet format 'src\Microsoft.Agents.AI.Workflows\Microsoft.Agents.AI.Workflows.csproj' --no-restore --include 'src\Microsoft.Agents.AI.Workflows\Specialized\HandoffAgentExecutor.cs' 'src\Microsoft.Agents.AI.Workflows\HandoffWorkflowBuilder.cs' --verbosity minimaldotnet format 'tests\Microsoft.Agents.AI.Workflows.UnitTests\Microsoft.Agents.AI.Workflows.UnitTests.csproj' --no-restore --include 'tests\Microsoft.Agents.AI.Workflows.UnitTests\HandoffOrchestrationTests.cs' --verbosity minimaldotnet test --project 'tests\Microsoft.Agents.AI.UnitTests\Microsoft.Agents.AI.UnitTests.csproj' -f net472 -c Release --no-restoredotnet test --project 'tests\Microsoft.Agents.AI.UnitTests\Microsoft.Agents.AI.UnitTests.csproj' -f net10.0 -c Release --no-restore --filter-query '/*/*/BackgroundAgentsProviderTests/*'dotnet format 'tests\Microsoft.Agents.AI.UnitTests\Microsoft.Agents.AI.UnitTests.csproj' --no-restore --include 'tests\Microsoft.Agents.AI.UnitTests\Harness\BackgroundAgents\BackgroundAgentsProviderTests.cs' --verbosity minimalBefore the corresponding fixes, both reasoning-enabled filtering cases failed, and all eight local/service-managed, reasoning-on/off, streaming/non-streaming bounce-back cases reproduced an empty completion. All now pass. The full solution unit-test suite and live-provider scenarios were not run.
Related Issue
Fixes #7384.
No other open PR referencing #7384 was found.
Contribution Checklist
breaking changelabel (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.