Bug: modelstudio (DashScope) API returns correct usage but session stores all zeros

Summary

When using the modelstudio provider (Alibaba Cloud DashScope) with GLM-5 or other models, the API correctly returns usage statistics in the response, but OpenClaw's session storage records all usage values as 0.

Evidence

1. API Response (correct usage)

{
  "usage": {
    "prompt_tokens": 8,
    "completion_tokens": 352,
    "completion_tokens_details": {
      "reasoning_tokens": 348
    },
    "total_tokens": 360
  },
  "choices": [{
    "message": {
      "content": "梦。",
      "reasoning_content": "..."
    }
  }]
}

2. Session Storage (all zeros)

From ~/.openclaw/agents/main/sessions/*.jsonl:

{
  "type": "message",
  "message": {
    "usage": {
      "input": 0,
      "output": 0,
      "totalTokens": 0,
      "cost": {"total": 0}
    }
  }
}

3. OpenClaw Status Output

Tokens: unknown/203k (?%)

Analysis

Code Review

The normalizeUsage function in auth-profiles-B5ypC5S-.js should correctly parse the response:

const rawInput = asFiniteNumber(raw.input ?? raw.inputTokens ?? raw.input_tokens ?? raw.promptTokens ?? raw.prompt_tokens);
const output = asFiniteNumber(raw.output ?? raw.outputTokens ?? raw.output_tokens ?? raw.completionTokens ?? raw.completion_tokens);

This supports prompt_tokens and completion_tokens field names, which match the API response.

Potential Causes

1. GLM-5's reasoning_content field: GLM-5 returns an extra reasoning_content field along with content. This might interfere with usage extraction logic.

2. Usage data not passed to session storage: The usage data might be lost or overwritten with default values somewhere in the agent-runner pipeline.

3. Stream processing issue: If streaming is enabled, the final usage might not be captured correctly.

Configuration

{
  "models": {
    "providers": {
      "modelstudio": {
        "baseUrl": "https://coding.dashscope.aliyuncs.com/v1",
        "api": "openai-completions",
        "models": [{
          "id": "glm-5",
          "contextWindow": 202752,
          "cost": {
            "input": 0,
            "output": 0
          }
        }]
      }
    }
  }
}

Environment

• OpenClaw version: 2026.3.28
• Node: v24.14.1
• OS: macOS 26.3 (arm64)
• Provider: modelstudio (Alibaba Cloud DashScope)
• Model: glm-5 (also affects qwen3.5-plus, kimi-k2.5)

Expected Behavior

Session storage should record actual usage values from the API response, enabling accurate token tracking and cost estimation.

Workaround Attempts

Setting cost values to non-zero didn't help - the usage is still recorded as 0.

This bug affects all modelstudio models and prevents users from tracking their API usage and costs.