Bug_194_EVALUATE: CHAT-BOUNDARY-011 — vendor_id=None silently forwarded to orchestrator with no validation error

[steadhac]

Component: finbot/agents/chat.py → VendorChatAssistant._call_start_workflow

Root cause:

# chat.py lines 198-201
task_data: dict[str, Any] = {
    "description": description,
    "vendor_id": vendor_id,   # no None check
    "parent_workflow_id": self._workflow_id,
}

vendor_id is a required parameter (no default value in the function signature) but has
no runtime guard. Python type hints are not enforced at call time. When the LLM omits
the field or sends null, vendor_id=None is accepted and written directly into
task_data, which is then dispatched to the orchestrator.

Steps to reproduce:

1. Create a VendorChatAssistant with a mock background_tasks.
2. Call _call_start_workflow("do something", vendor_id=None).
3. Inspect task_data passed to background_tasks.add_task.

Expected: ValueError or {"error": ...} JSON — validation rejects None for a required field
Actual: status='started', task_data["vendor_id"] is None — dispatched to orchestrator

How to execute:

pytest tests/unit/agents/test_chat_assistant.py::TestBoundaryAndTypeValues::test_chat_boundary_011_vendor_id_none_flows_into_task_data -v

Proposed fix:

if vendor_id is None:
    return json.dumps({"error": "vendor_id is required"})

Impact: The orchestrator receives None as the vendor primary key. Any downstream
DB query using vendor_id=None will either raise a DataError (PostgreSQL) or silently
return no rows (SQLite), with no signal propagated back to the chat session. The
workflow appears to start successfully from the caller's perspective while the
orchestrator operates on a phantom vendor.

Acceptance criteria:

• test_chat_boundary_011_vendor_id_none_flows_into_task_data updated to assert an error response (once fix is applied)
• _call_start_workflow returns {"error": "vendor_id is required"} when vendor_id is None
• All other _call_start_workflow tests continue to pass

[Jean-Regis-M]

@saikishu @steadhac The orchestrator is silently passing vendor_id=None downstream, which shifts validation debt to the DB/LLM layers. I’ll add an explicit early-boundary check in _call_start_workflow (or equivalent) and attach a regression test covering None, empty strings, and valid UUIDs. Fix is <30 LOC. Ready to implement.