Framework-agent failures return `None` instead of the failed task's reason

[ambiorix2099]

In short: when a framework agent fails, the SDK prints None instead of the reason — even though the server recorded a perfectly good one. Every such failure costs a manual walk of the server API to diagnose.

File: src/conductor/ai/agents/runtime/runtime.py

Symptom

$ CONDUCTOR_AGENT_LLM_MODEL=anthropic/claude-sonnet-4-6 python examples/agents/93_openai_runner_hello_world.py
Framework agent 'Assistant' execution FAILED
None

No reason, no task name, nothing actionable. Diagnosing it means walking the server API by hand:

curl -s localhost:8080/api/agent/executions?size=20
curl -s "localhost:8080/api/workflow/<executionId>?includeTasks=true"   # find FAILED tasks, read reasonForIncompletion

Cause

_run_framework() reports:

error=status.reason if raw_status in ("FAILED", "TERMINATED") else None

For these failures status.reason is empty, so the result is None.

The reason was available all along, at both levels of the execution:

• the workflow's own reasonForIncompletion:

  Task 41830611-7c25-4d51-a36f-50e324e59239 failed with status: FAILED and reason:
  'Task execution failed: OpenAI Responses API call failed: Responses API failed with status 401 ...'
  

• the failed task's reasonForIncompletion (Assistant_llm), carrying the same text.

There is also an existing helper that already does this. _extract_failed_task_reason(wf) reads the first FAILED task's reasonForIncompletion and returns Task '<ref>' failed: <reason>. It is called by run() (line ~2534) and _run_by_name() (line ~2623) — which is why native-agent failures report properly, e.g. 54_software_bug_assistant prints:

ERROR: Task 'software_assistant_54_list_mcp_0' failed: Failed to list MCP tools ... HTTP 400

_run_framework() never calls it.

Fix

When status.reason is empty, fall back to the workflow's reasonForIncompletion, or call _extract_failed_task_reason as run() does. Either is sufficient.

Verify

Start the server without OPENAI_API_KEY, run 93 — it should name the failing task and its reason instead of printing None.

Note — possible second gap, unconfirmed

59_coding_agent goes through run() yet also reported no reason, while its server-side SUB_WORKFLOW task did carry one (Anthropic Messages API failed with status 404). _extract_failed_task_reason only inspects the top-level workflow's tasks, so failures inside a sub-workflow may need the same treatment. Worth checking while fixing this.