DurableAgent surfaces fatal stream errors as [object Object] and does not clearly signal a terminal failure to stream consumers #1545

Summary

When DurableAgent hits a fatal mid-turn error, stream consumers can receive an opaque "[object Object]" error instead of a real message.

In the same failure path, the run/stream state can remain ambiguous enough that consumers may still attempt reconnect/resume logic even though the turn is effectively dead.

This appears to be a separate bug from the underlying reasoning/tool-loop issue in:

• DurableAgent should preserve reasoning content in conversation history across tool loop steps #1393
• Preserve reasoning content in DurableAgent conversation history #1444
• Bug: DurableAgent + OpenAI Responses API fails on tool calls due to missing required reasoning item #880
• fix(ai): strip openai itemID from tool call metadata  #889

Environment

• workflow: 4.2.0-beta.71
• @workflow/ai: 4.1.0-beta.57
• ai: 6.0.116

Problem

In a fatal DurableAgent stream failure, a structured error object can be collapsed to:

[object Object]

I also saw the library log:

Error processing UI message chunks: Error: [object Object]

This makes the real failure cause invisible to downstream consumers.

Separately, after this kind of fatal error, the failure is not always clearly terminal from the perspective of a stream consumer. A consumer may still treat the run as reconnectable/resumable even though the turn has already failed.

Expected behavior

• Fatal stream errors should surface a useful message, not "[object Object]".
• Non-recoverable stream failures should be clearly terminal so consumers do not keep trying to resume a dead turn.

Suggested fix

1. Replace String(error) for unknown errors with a safer formatter:

   • prefer error.message
   • otherwise serialize structured objects with JSON.stringify(...) when possible
   • fall back to a generic message only as a last resort

2. Ensure fatal stream-processing errors are exposed to consumers as terminal failures, so reconnect/resume logic can stop immediately.

3. Keep Preserve reasoning content in DurableAgent conversation history #1444 moving, since reasoning preservation appears to address one upstream trigger for this failure path.