[bug] /v1/chat/completions returns 400 when stream=true, breaking Agent framework integration

[tangxinyao]

Prerequisites

• I have searched the issue tracker and this bug has not been reported yet.
• The bug is in AReno itself, not in my own code.
• I can reproduce it with an official install (not a heavily modified environment).
• I can consistently reproduce the bug across multiple trials or seeds.

Describe the bug

The /v1/chat/completions endpoint of areno serve does not support stream=true and directly returns a 400 error. As a result, any Agent framework that relies on the SSE streaming protocol (such as OpenAI SDK, LangChain, AutoGen, CrewAI, etc.) cannot use the areno service as a backend.

Currently, this limit is explicitly hardcoded in the codebase:

• areno/cli/serve.py (line 433):

if request.stream:
    raise HTTPException(status_code=400, detail="stream=true is not supported")

• areno/api/agentic.py (line 456):

if body.get("stream"):
    _write_json(handler, 400, {"error": {"message": "streaming chat completions are not supported yet"}})

Why it is needed:

1. Evaluating Agent Model Performance: Mainstream Agent frameworks drive multi-turn conversations and tool usage via client.chat.completions.create(stream=True). Lack of streaming support prevents downstream evaluation of trained models.
2. Agentic Training (Tool-Use RL): During the rollout phase, agentic training loops rely on SSE streaming to dynamically parse incremental tool_calls and delta payloads. Non-streaming responses freeze the whole execution loop.

Expected behavior

When receiving stream=true, the /v1/chat/completions endpoint of areno serve should return an SSE event stream (text/event-stream) compliant with the OpenAI streaming specification:

• Begins with delta.role = "assistant"
• Incrementally emits text or tool_calls deltas
• Concludes with finish_reason, usage, and data: [DONE]

Note: Streaming support for /v1/chat/completions is target for the serve layer in this PR; areno/api/agentic.py is currently out of scope.

Full logs

HTTP/1.1 400 Bad Request
Content-Type: application/json

{
  "detail": "stream=true is not supported"
}

Reproduction

import openai

client = openai.OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="empty",
)

# This call will fail with a 400 Bad Request error
response = client.chat.completions.create(
    model="my-model",
    messages=[{"role": "user", "content": "Hello"}],
    stream=True,
)

for chunk in response:
    print(chunk)

Exact command

areno serve --model-path ling_v3_tiny_f16 --tp-size 1 --world-size 1 --port 8000

Full traceback

Environment

• OS: *
• Python: *
• CUDA: *
• PyTorch: *
• AReno: *
• GPU model: *

Install mode

editable source / pip install -e .

Was ARENO_BUILD_EXT=0 used?

Not sure

areno env --json

areno check

Additional context

No response