Summary

Fixes #7073 - Kimi K2.5 cache stats showing cacheRead: 0

Root Cause

Kimi K2 models use automatic prefix caching (like Anthropic) and return cache stats in a nested field:

{
  "usage": {
    "prompt_tokens": 1113,
    "cached_tokens": 1024,
    "prompt_tokens_details": {
      "cached_tokens": 1024
    }
  }
}

OpenClaw wasn't parsing prompt_tokens_details.cached_tokens, so cache stats were lost.

Changes

• src/agents/usage.ts: Add prompt_tokens_details.cached_tokens to cacheRead parsing chain
• src/agents/usage.test.ts: Add test for K2 usage format

Testing

Verified with real Kimi K2.5 API calls - cached_tokens is correctly returned on repeated requests with the same prefix.

Note

This also adds cached_tokens (top-level) for moonshot-v1 explicit caching API, which will be used by a separate PR (#25104).

Greptile Summary

Adds support for Kimi K2 automatic prefix caching stats by parsing prompt_tokens_details.cached_tokens in the usage normalization chain. This fixes cache stats showing as 0 for K2 models that return cache information in a nested field structure (similar to Anthropic). The implementation includes proper test coverage for both the Moonshot v1 explicit caching format (cached_tokens) and the K2 automatic prefix caching format (prompt_tokens_details.cached_tokens).

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The change is straightforward and well-tested. It extends the existing fallback chain in normalizeUsage to handle two additional cache token formats without breaking existing behavior. The implementation follows the established pattern of checking multiple field variations using the nullish coalescing operator. Both new formats have dedicated unit tests that verify correct parsing.
• No files require special attention

_{Last reviewed commit: d2a1cf0}