Python: [Bug]: SequentialBuilder second agent receives assistant-ending history and some models return empty text

Description

When using SequentialBuilder with multiple agent participants (default settings), the second (and later) participant is invoked with the prior agent's full_conversation.

Observed shape for a two-agent sequential run:

[user, assistant(writer output)]

That matches the documented chaining behavior (AgentExecutor.from_response with default context_mode="full" → prior.full_conversation).

With at least one OpenAI-compatible chat model we tested (qwen3.8-27b@4bit behind a chat-completions API), a completion whose input ends with an assistant message returns essentially empty text:

• finish_reason=stop
• contents empty (or only a usage chunk)
• usage.output_token_count ≈ 1

So the second participant (e.g. “reviewer”) appears to “run successfully”, but produces no assistant text. Downstream AG-UI workflow streaming then has no TEXT_MESSAGE_* events for that turn (only an empty AgentResponseUpdate surfaced as a custom/workflow_output-style payload).

Expected: We are looking for guidance from maintainers on how Sequential / multi-agent conversation chaining is intended to work with providers that do not emit text when the prompt ends on assistant.

Repro outline

1. Build SequentialBuilder(participants=[writer, reviewer]).build() with two chat agents on the model above.
2. Run with a simple user prompt (streaming or non-streaming).
3. Writer returns normal assistant text.
4. Reviewer is called with [user, assistant(writer)] and returns empty assistant content.

Same empty result when calling the second agent directly:

agent.run([
    Message(role="user", contents=["go"]),
    Message(role="assistant", contents=["W-OK"]),
], stream=True)
# → stop with empty text

While:

agent.run([Message(role="user", contents=["go"])], stream=True)
# → normal text

and

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

No exception is raised; the workflow completes with executor_completed / RUN_FINISHED.

Code Sample

from agent_framework.orchestrations import SequentialBuilder

workflow = SequentialBuilder(participants=[writer, reviewer]).build()

# stream=True or stream=False both show the same empty second-agent text
async for event in workflow.run("go", stream=True):
    ...
# reviewer AgentResponse / AgentResponseUpdate has empty text contents

Error Messages / Stack Traces

No exception. Empty model completion only.

Example observed update for the second agent:

AgentResponseUpdate(contents=[], role=assistant, finish_reason=stop, ...)
# followed by a usage-only update

Package Versions

agent-framework (core): 1.15.0
agent-framework-orchestrations: 1.1.1
agent-framework-ag-ui: (used when observing AG-UI TEXT_MESSAGE absence; core issue reproduces without AG-UI via workflow.run / agent.run)

Python Version

Python 3.14

Additional Context

• Default Sequential settings (chain_only_agent_responses=False). With chain_only_agent_responses=True, the second agent still receives an assistant-ending message list ([assistant(writer)]), and we observed the same empty completion pattern.
• First participant is fine; only subsequent participants that inherit assistant-ending history are affected on this model.
• Happy to provide more traces (executor_invoked payloads / raw chat-completions request bodies) if useful.