🐞 Bug Report

Describe the bug

When Honcho is configured to use a custom OpenAI-compatible provider backed by
DashScope/Qwen (e.g. qwen3.5-plus), any background task that uses JSON mode
(response_format={"type":"json_object"}) fails immediately with a 400 error:

Error code: 400 - {'error': {'message': "<400> InternalError.Algo.InvalidParameter:
'messages' must contain the word 'json' in some form, to use 'response_format' of
type 'json_object'.", 'type': 'invalid_request_error', ...}}

The deriver starts up successfully but its first structured-output LLM call is
rejected before generation even begins. The worker retries 3 times then stays in a
broken loop.

To Reproduce

1. Self-host Honcho and set the following in config.toml:

   [llm]
   OPENAI_COMPATIBLE_API_KEY = "your-dashscope-key"
   OPENAI_COMPATIBLE_BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
   
   [deriver]
   PROVIDER = "custom"
   MODEL = "qwen3.5-plus"

2. Start Honcho with docker compose up.
3. Watch the deriver logs: docker compose logs --tail=100 -f deriver
4. See the repeated 400 invalid_parameter_error on every structured-output attempt.

Expected behaviour

Honcho should be able to use DashScope/Qwen through the custom OpenAI-compatible
provider path for JSON-mode requests without request rejection. Qwen does support
response_format={"type":"json_object"}; it just additionally requires the request
messages to explicitly contain the word json.

Media prove

Deriver log output:

2026-03-26 14:48:07,442 - src.utils.clients - WARNING - Error on attempt 1/3 with
custom/qwen3.5-plus: Error code: 400 - {'error': {'message': "<400>
InternalError.Algo.InvalidParameter: 'messages' must contain the word 'json' in
some form, to use 'response_format' of type 'json_object'.", ...}}

Your environment

• OS: Ubuntu (Docker self-hosted)
• Honcho Server Version: latest main
• Custom Provider: Alibaba Cloud DashScope — https://dashscope-intl.aliyuncs.com/compatible-mode/v1
• Model: qwen3.5-plus

Additional context

Root cause: src/utils/clients.py sets response_format={"type":"json_object"}
in the shared AsyncOpenAI branch for JSON mode, but does not guarantee the outgoing
messages explicitly mention the word json. DashScope/Qwen validates this server-side
and rejects the request if the word is absent.

This is documented Qwen behaviour:
https://www.alibabacloud.com/help/en/model-studio/qwen-structured-output

Affected components: The bug affects any Honcho component that uses JSON mode
through the custom OpenAI-compatible path, not just the deriver — summary and dream
likely hit the same issue.

Suggested fix: In the AsyncOpenAI client path in src/utils/clients.py, when
json_mode=True and the provider is custom and the model name contains qwen,
prepend a system message containing a JSON instruction if none of the existing messages
already mention json:

if json_mode and provider != "vllm":
    if provider == "custom" and "qwen" in model.lower():
        has_json_hint = any(
            isinstance(m.get("content"), str)
            and "json" in m["content"].lower()
            for m in openai_params["messages"]
        )
        if not has_json_hint:
            openai_params["messages"] = [
                {
                    "role": "system",
                    "content": "Return valid JSON only. The output must be a JSON object.",
                },
                *openai_params["messages"],
            ]
    openai_params["response_format"] = {"type": "json_object"}

This needs to be applied to both the non-streaming and streaming AsyncOpenAI
branches. The same pattern applies for the elif json_mode: branch inside the
streaming path.

I have tested this fix locally and it resolves the deriver 400 failures.