Streaming responses return zero token counts

[eugenio]

Bug Description

When streaming is enabled (stream: true), the final SSE chunk always returns usage: {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0} instead of actual token counts. Non-streaming responses work correctly.

This breaks token/cost tracking in downstream clients like OpenCode.

Root Cause

Two issues in server.py (_stream_litellm function):

1. Missing stream_options parameter (line ~1448)

The LiteLLM acompletion() call sets stream: True but never passes stream_options: {"include_usage": True}. Without this, the upstream provider does not include token counts in streaming chunks.

Before:

call_kwargs: Dict[str, Any] = {"model": litellm_model, "messages": messages, "stream": True}

Fix:

call_kwargs: Dict[str, Any] = {
    "model": litellm_model,
    "messages": messages,
    "stream": True,
    "stream_options": {"include_usage": True},
}

2. Usage-only final chunks are silently dropped (lines ~1485-1488)

Some providers send a final chunk with usage data but empty choices. The current code skips these with if choice is None: continue, discarding the usage data.

Before:

async for chunk in response:
    choice = chunk.choices[0] if chunk.choices else None
    if choice is None:
        continue
    # ... usage extraction happens AFTER this skip

Fix: Extract usage BEFORE the choice is None check, and yield usage-only chunks:

async for chunk in response:
    usage = None
    if hasattr(chunk, "usage") and chunk.usage:
        usage = {
            "prompt_tokens": chunk.usage.prompt_tokens or 0,
            "completion_tokens": chunk.usage.completion_tokens or 0,
        }

    choice = chunk.choices[0] if chunk.choices else None
    if choice is None:
        if usage:
            yield {}, usage, None
        continue
    # ... rest unchanged

Verification

After applying both fixes, streaming responses correctly return token counts:

Before: "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
After:  "usage": {"prompt_tokens": 40, "completion_tokens": 10, "total_tokens": 50}

Impact

• OpenCode TUI context meter shows "0 tokens, 0% used"
• Cost tracking shows $0.00 for all sessions
• Any OpenAI-compatible client relying on streaming usage data is affected

Affected Version

NadirClaw v0.13.0