Python: [Bug]: run_skill_script approval handshake leaves function call unanswered in local Inspector (Responses API) #7267

Description

Bug: run_skill_script approval handshake leaves function call unanswered in local Inspector

Environment

Description

When using a SkillsProvider with a script_runner and the default approval mode, approving the run_skill_script tool call in the Agent Inspector does not complete correctly. The tool block shows a red ✗ and the function call result is never submitted back to the Responses API. On the next turn the API rejects the request with a 400 error.

Steps to Reproduce

1. Create an agent with a file-based skill and a script_runner (you can use the 07-skills example provided by Microsoft):

   skills_provider = SkillsProvider.from_paths(
       skill_paths=Path(__file__).parent / "skills",
       script_runner=run_local_skill_script,
       # default: approvals enabled for all tools
   )
   agent = Agent(client=client, context_providers=[skills_provider], ...)
   server = ResponsesHostServer(agent)
   server.run()

2. Start the agent locally with azd ai agent run or F5 to start the service and open the Agent Inspector

3. Open the Agent Inspector and send a message that triggers run_skill_script (e.g. "Greet me" with a greet-user skill).

4. The Inspector shows an approval dialog for load_skill → approve it. ✅

5. The Inspector shows run_skill_script but stays there forever

Actual Behavior

• After run_skill_script, the tool block shows a red ✗ and the agent hangs.

• The script runner function is never called (confirmed by adding logging before and inside the function — the entry log never appears).

• On the next conversation turn, the Responses API returns:

  server_error: (
    "<class 'agent_framework_foundry._chat_client.FoundryChatClient'> service failed to complete the prompt:
     Error code: 400 - {'error': {'message': 'No tool output found for function call call_<id>.',
     'type': 'invalid_request_error', 'param': 'input', 'code': None}}",
    BadRequestError(...)
  )
  

Expected Behavior

After the approval, the framework should:

1. Call the registered script_runner.
2. Submit the returned string as the tool output to the Responses API.
3. Allow the model to continue and produce a final reply.

Root Cause Hypothesis

The approval round-trip for run_skill_script uses the Responses API's function call / tool output mechanism. After the user clicks Approve in the Inspector, the hosting layer (ResponsesHostServer) fails to deliver the tool output back to the API in the follow-up request, leaving the function call call_<id> dangling. The API then rejects the next request because the conversation history contains an unresolved function call with no corresponding tool output.

The script_runner itself is never invoked — the failure occurs inside the framework before the runner is reached.

Workaround

Pass disable_run_skill_script_approval=True (and optionally disable_load_skill_approval=True) to SkillsProvider.from_paths(). This is safe for trusted local scripts and bypasses the broken approval round-trip entirely:

skills_provider = SkillsProvider.from_paths(
    skill_paths=Path(__file__).parent / "skills",
    script_runner=run_local_skill_script,
    disable_load_skill_approval=True,
    disable_run_skill_script_approval=True,
)

### Code Sample

```markdown

Error Messages / Stack Traces

: (
      "<class 'agent_framework_foundry._chat_client.FoundryChatClient'> service failed to complete the prompt:
       Error code: 400 - {'error': {'message': 'No tool output found for function call call_<id>.',
       'type': 'invalid_request_error', 'param': 'input', 'code': None}}",
      BadRequestError(...)
    )

Package Versions

agent-framework-foundry-hosting: 1.0.0b260721, agent-framework-foundry: 1.10.2

Python Version

Python 3.13

Additional Context

With default approval values:
Screenshot 1: Waiting for approval

Screenshot 2: After approval was accepted

Screenshot 3: In second round (entered the same prompt to trigger the skill, 'Greet me')

After the workaround:

Screenshot 4: After disabling the approvals