Python: [Bug]: Responses stream parser silently drops unhandled platform/hosted tool call and result item types, causing AG-UI tool calls to go missing or show "tool calls missing results"

[djw-bsn]

Description

When a Foundry hosted agent uses a Foundry Toolbox to dispatch a knowledge-base tool call, the Responses API emits the result as a response.output_item.added item with type == "function_call_output". agent_framework_openai has no handler for this item type, logs it as unparsed, and discards it. No Content.from_function_result(...) reaches agent_framework_ag_ui.

Downstream, this means:

• agent_framework_ag_ui emits TOOL_CALL_END but no matching TOOL_CALL_RESULT; its fallback treats the call as declaration-only.

The model still receives and uses the real output, but the structured result never reaches the AG-UI client.

Caveat: the model sees a generic call_tool wrapper; the Toolbox executes the inner MCP tool locally and ResponsesHostServer intentionally serializes its function_result as the standard function_call_output Responses item. The same KB tool called directly is a normal mcp_call and works, confirming this is a client-parser gap rather than a KB-tool failure.

Reproduction

1. Configure a Foundry hosted agent with a Foundry Toolbox connected to a knowledge-base retrieval tool. The agent must call the KB tool through the toolbox's generic call_tool function; a direct call does not reproduce this.
2. Drive it via agent_framework_ag_ui's AgentFrameworkAgent.run() (AG-UI protocol) with DEBUG logging enabled for agent_framework.openai.
3. Ask a question that triggers the hosted tool.
4. Observe, with DEBUG logging on agent_framework.openai:

   DEBUG agent_framework.openai - Unparsed event of type: response.output_item.added:
   ResponseOutputItemAddedEvent(item=ResponseFunctionToolCallOutputItem(...,
   call_id='call_XXXX', output='<actual tool result text>',
   type='function_call_output', ...))
   

5. The stream emits TOOL_CALL_END but no matching TOOL_CALL_RESULT. agent_framework_ag_ui._agent_run logs:

   INFO agent_framework_ag_ui._agent_run - Found 1 pending tool calls without end event
   INFO agent_framework_ag_ui._agent_run - Emitting ToolCallEndEvent for declaration-only tool 'call_XXXX'
   

   No function_result content is produced, despite the successful tool output in step 4.

Controls

• The same FoundryIQ KB tool called directly, without a Toolbox, is exposed as mcp_call and produces the complete TOOL_CALL_START → TOOL_CALL_END → TOOL_CALL_RESULT lifecycle. This shows the issue is specific to the local Toolbox execution path, not the KB tool.
• A Databricks Genie tool routed through mcp_call also works correctly.
• A SharePoint grounding preview emits sharepoint_grounding_preview_call and sharepoint_grounding_preview_call_output, which are both logged as unparsed and produce no AG-UI events. This is secondary evidence because the feature is preview.

Expected behavior

Parse function_call_output into Content.from_function_result(call_id=..., result=event_item.output, ...) so agent_framework_ag_ui emits the matching TOOL_CALL_RESULT. Other supported Responses item types should likewise be handled instead of silently dropped.

Suggested fix location

Add function_call_output handling to agent_framework_openai/_chat_client.py. The SharePoint preview types may need equivalent handling.

Code Sample

Error Messages / Stack Traces

Package Versions

agent-framework-openai==1.14.2, agent-framework-core==1.17.0, agent-framework-ag-ui==1.2.2, agent-framework-foundry==1.12.0

Python Version

Python 3.12

Additional Context

No response

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

python/packages/openai/agent_framework_openai/_chat_client.py::_parse_chunk_from_openai drops function_call_output items in the response.output_item.added branch around lines 3205-3335. The trigger is a Responses stream item containing type="function_call_output", a call_id, and hosted tool output. Minimal repro: pass that event shape directly to _parse_chunk_from_openai and assert it emits one correlated function_result; the current result is an empty content list.

• Failing test: python/packages/openai/tests/openai/test_openai_chat_client.py::test_parse_chunk_from_openai_with_function_call_output_added
• Files examined: python/packages/openai/agent_framework_openai/_chat_client.py, python/packages/openai/tests/openai/test_openai_chat_client.py, python/packages/ag-ui/agent_framework_ag_ui/_agent_run.py, python/packages/openai/pyproject.toml, docs/specs/004-python-function-calling-loop.md
• Tests run: test_parse_chunk_from_openai_with_web_search_call_added, test_parse_chunk_from_openai_with_mcp_call_added_defers_result, test_parse_chunk_from_openai_with_function_call_output_added
• Reported version: 1.14.2
• Current version: 1.14.2