Bug: Usage data discrepancy between CLI /usage command and web dashboard

首先，按照 issue 中的重現步驟，將 CLI /usage 的彙總結果與 Web 儀表板的模型明細進行比較。檢查 #440 中提到的背景或內部模型呼叫是否包含在 CLI 中但被儀表板排除；當兩個介面回報一致的模型數量和成本時，即視為完成。

Bug: Usage data discrepancy between CLI /usage command and web dashboard

Summary

There is a significant discrepancy between the usage data shown by the CLI /usage command and the web dashboard (https://commandcode.ai/settings/usage). The CLI shows one model as the most used, while the web dashboard shows a completely different model.

Environment

• Command Code version: 0.30.1
• OS: Windows
• Shell: PowerShell

Steps to Reproduce

1. Use Command Code with multiple models over several sessions
2. Run /usage in the CLI to see usage summary
3. Check the web dashboard at https://commandcode.ai/settings/usage
4. Compare the model usage breakdown

Expected Behavior

The usage data should be consistent between the CLI and web dashboard. Both should show the same model as the most used.

Actual Behavior

CLI /usage command shows:

• moonshotai/Kimi-K2.5: 154 requests, $0.4753
• xiaomi/mimo-v2.5-pro: 316 requests, $0.4663
• deepseek/deepseek-v4-pro: 81 requests, $0.4430
• MiniMaxAI/MiniMax-M2.5: 258 requests, $0.0805

Web dashboard shows:

• xiaomi/mimo-v2.5-pro: Most used model (expected)
• moonshotai/Kimi-K2.5: Only 3 requests (discrepancy)

The CLI shows Kimi-K2.5 as the most used model (154 requests), but the web dashboard shows it was only used 3 times. The web dashboard correctly shows MiMo 2.5 Pro as the most used model.

Analysis

This appears to be a data synchronization or attribution issue. The CLI may be:

1. Aggregating data differently than the web dashboard
2. Including background/internal model calls (as mentioned in #440) in the CLI summary but not in the web dashboard
3. Having a caching or sync issue between local and server-side data

Related Issues

• #440 — "Background model MiniMax-M2.5 is expensive and uncontrollable" (mentions Kimi-K2.5 background usage)
• #326 — "Unexpected usage of multiple models during a single session" (closed as non-issue, but same underlying problem)
• This discrepancy may be related to the background model usage discussed in #440

Additional Notes

• The web dashboard appears to show the correct user-selected model usage
• The CLI may be including background/internal model calls in its summary
• This discrepancy makes it difficult to accurately track actual model usage and costs