[pratik wayase (PratikWayase)]

Motivation & Context

When using SequentialBuilder with multiple agent participants, the second (and later) participant is invoked with the prior agent's full_conversation, which produces a history ending on an assistant message (e.g., [user, assistant(writer output)]). Some chat-completions providers (e.g., qwen3.8-27b@4bit behind an OpenAI-compatible chat-completions API) return essentially empty text (finish_reason=stop, contents empty, usage.output_token_count ≈ 1) when the prompt ends on an assistant message. The second participant (e.g., a "reviewer") then "runs successfully" but produces no assistant text, breaking downstream AG-UI workflow streaming (TEXT_MESSAGE_* events are never emitted for that turn).

The bug reporter themselves confirmed that adding any trailing user message resolves the issue:

agent.run([user, assistant("W-OK"), user("continue")], stream=True)  # → normal text

This PR provides an opt-in mechanism to append a synthetic user continuation message before invoking an agent when the cache ends on an assistant turn.

Description & Review Guide

• What are the major changes?

  • Added ensure_trailing_user_turn: bool = False parameter to AgentExecutor.__init__. When enabled, a new private method _ensure_cache_not_assistant_ending() is invoked at the start of _handle_request. If the cache is non-empty and ends on an assistant message, a synthetic Message(role="user", contents=["please continue."]) is appended before the agent runs.
  • Threaded the same parameter through AgentApprovalExecutor.__init__ → its inner AgentExecutor, so request-info-wrapped agents also honor the flag.
  • Threaded the same parameter through SequentialBuilder.__init__ → all participant executors (both the AgentApprovalExecutor branch and the plain AgentExecutor branch).
  • Added docstrings for the new parameter on all three classes.
  • Added 13 new tests across three test files covering: default-off behavior, flag-disabled no-op, cache-already-ending-on-user no-op, empty cache no-op, synthetic turn appended when enabled, end-to-end two-agent chaining with and without the flag, chain_only_agent_responses=True interaction, and propagation through AgentApprovalExecutor → inner AgentExecutor.

• What is the impact of these changes?

  • No breaking changes. The default value is False everywhere, preserving all existing behavior.
  • Users who hit the empty-completion issue on assistant-ending prompts can now opt in by passing ensure_trailing_user_turn=True to SequentialBuilder.
  • The fix is applied at the AgentExecutor level, so it also works for non-Sequential workflows that chain AgentExecutor instances manually (verified by test_from_response_ensure_trailing_user_turn_end_to_end_full_mode).

• What do you want reviewers to focus on?

  • Whether the synthetic message text "please continue." is acceptable as a hardcoded default, or whether it should be configurable (e.g., ensure_trailing_user_turn: bool | str = False). For non-continuation agents like a "reviewer," "please continue." is semantically odd but functionally sufficient.
  • Whether the synthetic user turn should be excluded from the agent's full_conversation propagation so that downstream agents (3rd, 4th in a chain) don't accumulate synthetic turns in their context. Currently the synthetic turn is appended to self._cache only, but it may still appear in the agent's persisted history depending on the agent implementation. This PR does not test the 3-agent accumulation case.
  • Whether the flag should eventually default to True. We chose False for backward compatibility, but maintainers may want to discuss whether the empty-completion behavior should be considered a bug to fix by default rather than a workaround to opt into.

Related Issue

Fixes #8240

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• 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.

[Copilot]

Copilot review overview

🟢 Approval recommended

The implementation and coverage are sound; remaining comments concern minor documentation and logging corrections.

Review effort: Balanced
Findings: 2

Open (2)

• Add spaces between adjacent debug message literals · New
• Clarify builder flag scope for prebuilt executors · New

What changed in this PR

Adds an opt-in synthetic trailing user turn to prevent empty completions from providers that reject assistant-ending prompts.

Changes:

• Extends agent and sequential executors with ensure_trailing_user_turn.
• Propagates the option through approval-wrapped agents.
• Adds unit and end-to-end coverage for chaining scenarios.

File	Description
python/​packages/​core/​agent_framework/​_workflows/​_agent_executor.py	Implements synthetic user-turn insertion.
python/​packages/​core/​tests/​workflow/​test_agent_executor.py	Tests core executor behavior.
python/​packages/​orchestrations/​agent_framework_orchestrations/​_sequential.py	Exposes and propagates the builder option.
python/​packages/​orchestrations/​agent_framework_orchestrations/​_orchestration_request_info.py	Propagates the option through approval workflows.
python/​packages/​orchestrations/​tests/​test_sequential.py	Tests sequential chaining configurations.
python/​packages/​orchestrations/​tests/​test_orchestration_request_info.py	Tests approval-executor propagation.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[Evan Mattson (moonbox3)]

Can you have a look at the file conflict please

[pratik wayase (PratikWayase)]

The single failing test test_cancellation_signal_preempts_stuck_agent_call in packages/foundry_hosting/tests/test_responses.py is unrelated to this PR. It mocks agent.run directly and asserts a 1-second cancellation-preemption timeout — it flakes on Windows / Python 3.14 under CI load. None of the files in this PR (_agent_executor.py, _orchestration_request_info.py, _sequential.py and their tests) are on the test's call path.

[pratik wayase (PratikWayase)]

Hi Eduard van Valkenburg (@eavanvalkenburg), Evan Mattson (@moonbox3), and westey (@westey-m), It should be ready to merge now, but let me know if you need any further changes. Thanks!