Python: Fix handoff target invoked with no messages when response has no text

[Sadok Barbouche (cr-sbarbouche)]

Motivation & Context

HandoffAgentExecutor broadcasts only the cleaned agent response to the other agents, then sends the handoff target an AgentExecutorRequest. clean_conversation_for_handoff drops every message with no text content, so a response consisting solely of the handoff tool call (no text) cleans to an empty list. The handoff target's request then carries messages=[].

AgentExecutor tolerates an empty cache and only logs a warning, so most chat agents survive. Agents that require input do not — for example an A2AAgent target raises ValueError: At least one message is required when starting a new task (no continuation_token)., aborting the whole workflow. This is especially reachable in the handoff pattern because the handoff itself is a function call: an agent that hands off without also emitting text produces exactly the tool-only response that cleans away to nothing.

This is the same root cause as #7456, in a different orchestrator. PR #7549 fixed the group-chat orchestrators by supplying a continuation instruction when the selected speaker would otherwise receive no messages; the handoff path was deliberately left out of that PR to keep it scoped to the linked issue.

Description & Review Guide

• What are the major changes?
  HandoffAgentExecutor._run_agent_and_emit now sends the handoff target a continuation instruction message instead of messages=[] when the handing-off agent's cleaned response is empty. The instruction text lives in a private module constant, _HANDOFF_CONTINUATION_DEFAULT_INSTRUCTION, following the existing _AUTONOMOUS_MODE_DEFAULT_PROMPT pattern already in _handoff.py.

• What is the impact of these changes?
  The fix is scoped to the case where the handoff response cleans to no messages at all. When the response has any text content, nothing changes and no instruction is injected (covered by a dedicated test). It applies to every agent type rather than special-casing A2A, so any participant that rejects empty input benefits.

• What do you want reviewers to focus on?
  Whether the injected instruction is the right shape versus replaying other conversation context into the request, and whether cleaned_response being empty is the correct (and only) signal to check here in the handoff path.

Tests cover a handoff whose response has only the handoff tool call and no text (streaming and non-streaming), asserting the target's first invocation carries the continuation instruction, plus a companion test asserting the instruction is not injected when the handoff response already has text.

Related Issue

Fixes #7573

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.

[Copilot]

Pull request overview

Prevents handoff targets from receiving an empty request after a textless tool-only handoff response.

Changes:

• Adds a fallback continuation instruction for empty cleaned responses.
• Adds streaming, non-streaming, and text-preservation tests.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
_handoff.py	Injects the fallback handoff message.
test_handoff.py	Adds regression coverage for textless handoffs.

---

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