openclaw - 💡(How to fix) Fix Local model token usage not passed through to session logs [1 comments, 2 participants]

Fix Plan

To resolve the issue, we need to update the gateway to correctly map the OpenAI-format usage fields from local model responses into the session log usage object.

Steps:

• Update the gateway's request proxying logic to parse the usage object from the local model response.
• Map the prompt_tokens, completion_tokens, and total_tokens fields to the corresponding fields in the session log usage object.
• Ensure the gateway correctly handles cases where the local model response may not include all usage fields.

Example Code:

# Assuming the gateway uses Python and a similar request/response handling framework
def proxy_request(request):
    # ... (existing code)

    # Parse the local model response
    response_json = response.json()
    usage = response_json.get('usage', {})

    # Map the usage fields
    session_usage = {
        'input': usage.get('prompt_tokens', 0),
        'output': usage.get('completion_tokens', 0),
        'totalTokens': usage.get('total_tokens', 0)
    }

    # Update the session log usage object
    session_log['usage'] = session_usage

    # ... (existing code)

Verification

To verify the fix, test the gateway with a local model request and check the session JSONL logs for correct usage data.

Extra Tips

• Ensure the gateway handles cases where the local model response may not include all usage fields to avoid errors.
• Consider adding logging or monitoring to track any issues with usage data mapping or parsing.