Summary

We investigated repeated Ask failures in Sourcebot and captured the real upstream model traffic with a transparent proxy. We found two separate failure modes:

1. Streaming/network failure
   Some model requests return 200 OK with text/event-stream, start streaming, but never terminate correctly. Eventually the client fails with:

   • TypeError: terminated
   • TLSSocket.onHttpSocketClose
   • read ETIMEDOUT

2. Tool-call argument failure
   In OpenAI-compatible mode, the model sometimes emits invalid function arguments JSON for tool calls, which leads to:

   • invalid function arguments json string
   • upstream 400 Bad Request

These are not the same failure.

Environment

• Sourcebot: current Docker image as of 2026-04-10
• Runtime: Node 24 inside container
• Deployment: local Docker on WSL2
• Provider modes tested:
  • openai-compatible
  • anthropic

What we ruled out

We ruled out the following as the primary root cause:

• WSL general outbound connectivity
• Docker general outbound connectivity
• basic long-lived fetch/undici SSE streaming in both host and container
• host proxy env leakage into the Sourcebot container

Using the same sourcebot container and Node fetch/undici, an independent long streaming request to the same model endpoint completed successfully:

• host: 49 chunks, 23594 bytes, 38.4s
• container: 52 chunks, 25029 bytes, 43.6s

This narrowed the failures to Sourcebot's real Ask workflow rather than general WSL/Docker outbound connectivity.

Failure mode 1: hanging SSE stream

Symptom

Frontend remains stuck in generating, and backend eventually reports:

• TypeError: terminated
• TLSSocket.onHttpSocketClose

Captured behavior

For some Ask requests, upstream responds with:

• 200 OK
• content-type: text/event-stream

But the stream never ends normally. Our proxy later records:

• message: "terminated"
• cause_code: "ETIMEDOUT"
• cause_message: "read ETIMEDOUT"

This happened in both Anthropic-compatible and OpenAI-compatible runs.

Example evidence

OpenAI-compatible sample:

• request body size about 45 KB
• stream starts successfully
• then hangs until timeout

Anthropic-compatible sample:

• request to /anthropic/v1/messages
• stream=true
• thinking.enabled
• tool_choice={"type":"auto"}
• tools_count=8
• stream starts, then never cleanly finishes

See sanitized evidence: https://gist.github.com/leozhengliu-pixel/6252d9b8415ab65dd106d11e2bf59da0

Failure mode 2: invalid tool-call arguments

Symptom

Frontend error:

Failed after 2 attempts with non-retryable error: 'invalid params, invalid function arguments json string ...'

Captured behavior

In an OpenAI-compatible request, the assistant produced this tool call:

{
  "id": "call_function_j32rtdm8f446_1",
  "type": "function",
  "function": {
    "name": "read_file",
    "arguments": "\"{\""
  }
}

The arguments field is not a JSON object string like:

{"path":"...","repo":"..."}

It is just a malformed string literal.

The same request also contains the tool error returned into conversation state:

{
  "role": "tool",
  "tool_call_id": "call_function_j32rtdm8f446_1",
  "content": "Invalid input for tool read_file: JSON parsing failed: Text: {.\nError message: Expected property name or '}' in JSON at position 1 (line 1 column 2)"
}

That request ultimately receives 400 Bad Request.

Why I believe this is relevant to Sourcebot

Even if some instability may be provider-side, Sourcebot is the shared client here, and these failures happen specifically under Sourcebot Ask orchestration:

• tool-rich multi-step interaction
• title-generation subrequests
• conversation accumulation
• streaming consumer path

At minimum, it would help if Sourcebot:

• handled hanging SSE streams more defensively
• surfaced raw upstream request/response diagnostics more clearly
• isolated title-generation failure from main answer flow
• degraded more safely on malformed tool arguments

Requested guidance

Please help clarify:

1. Is this a known limitation of Ask with OpenAI-compatible / Anthropic-compatible providers?
2. Are there recommended settings to disable or simplify:
   • title generation
   • thinking
   • tool streaming
   • multi-step Ask behavior
3. Is there an existing safeguard for malformed tool_calls[].function.arguments?
4. Is there a supported way to capture Sourcebot's exact model payloads without patching Sourcebot?

I can provide more sanitized payloads if useful.