Python: Workflow evaluation fails with reasoning models but succeeds with non-reasoning models #4047

Summary

The workflow evaluation sample fails when the workflow uses a reasoning model (for example gpt-5.2), but completes when switched to a non-reasoning model (for example gpt-4.1-nano).

Repro

1. Use python/samples/05-end-to-end/workflow_evaluation/run_evaluation.py (or create_workflow.py) and configure the workflow model to a reasoning deployment (for example gpt-5.2).
2. Run the workflow step.
3. Observe failure during tool-calling/workflow execution.
4. Switch the workflow model to a non-reasoning deployment (for example gpt-4.1-nano) and run again.
5. Workflow completes.

Observed error(s)

A common error is:

• No tool output found for function call call_...

(Also seen earlier in the same investigation: function_call/reasoning pairing validation failures.)

Expected behavior

Workflow execution should be model-agnostic for reasoning vs non-reasoning deployments and should not fail in workflow/tool-call continuation paths.

Investigation notes

• Focused/isolated responses tests and diagnostics (streaming vs non-streaming, store=True/False/None) did not reliably reproduce the failure.
• The failure appears specific to workflow orchestration/chaining behavior and how response history is handled across workflow agent boundaries when reasoning responses are involved.
• Likely area: interaction between workflow message handoff/history replay and Responses function-call continuation semantics.

Suspected areas

• python/packages/core/agent_framework/_workflows/_agent_executor.py
• python/packages/core/agent_framework/_tools.py
• python/packages/core/agent_framework/openai/_responses_client.py
• python/samples/05-end-to-end/workflow_evaluation/create_workflow.py