bug: Chat Completions usage drops cache-write token counts
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Newbie friendliness
- 85/100
- Issue type
- Bug
- Clarity
- Clearly specified
- Activity status
- Active
- Tech stack
- node.js, typescript
- Domain
- backend
Research direction
The converters extractUsage (packages/kosong/src/providers/openai-common.ts) and parseOpenAIUsage (packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts) drop cache-write tokens. Read both functions, add cache_write_tokens to inputCacheCreation and subtract from inputOther. Done when normalized output matches expected inputCacheCreation: 256 and inputOther: 117.
Written by the indexing model from the issue text.
Description
What version of Kimi Code is running?
Source checkout at 7ccac2b45b8a3481c76f771bb99c043b562de2c5, including @moonshot-ai/kosong 0.5.6. This is an offline source-function reproduction, not a running CLI session.
Which open platform/subscription were you using?
Moonshot Open Platform Chat Completions response schema. No subscription login or hosted API request was needed for this reproduction.
Which model were you using?
The documented kimi-k3.1-pro usage example.
What platform is your computer?
macOS arm64, Node.js v26.5.0.
What issue are you seeing?
Both Chat Completions usage converters ignore prompt_tokens_details.cache_write_tokens. They return inputCacheCreation: 0 and include those tokens in inputOther instead. The aggregate input count remains correct, but the normalized cache-write breakdown is lost.
The affected conversions are:
What steps can reproduce the bug?
Pass this usage object from the official Chat API example to either converter:
{
"prompt_tokens": 373,
"completion_tokens": 112,
"prompt_tokens_details": {
"cache_write_tokens": 256
}
}
Both currently return the following normalized counters:
{
"inputOther": 373,
"output": 112,
"inputCacheRead": 0,
"inputCacheCreation": 0
}
The v2 converter additionally preserves the original object in raw.
What is the expected behavior?
{
"inputOther": 117,
"output": 112,
"inputCacheRead": 0,
"inputCacheCreation": 256
}
The API documents ordinary input, cache reads and cache writes as mutually exclusive parts of prompt_tokens. Kosong's TokenUsage contract likewise defines inputOther as input that is neither cache-read nor cache-created.
A response with 1,000 input tokens, 600 cached tokens and 256 cache-write tokens should therefore produce 144 ordinary input tokens, 600 cache-read tokens and 256 cache-created tokens. Both converters currently produce 400, 600 and 0.
Additional information
I executed the exact two source functions after stripping TypeScript types. The write-only and read/write cases fail in both implementations; ordinary usage and the legacy top-level cached_tokens cases pass. This was not a package test run, CLI end-to-end test or billing experiment.
This is separate from #2160, which handles partial usage objects without aggregate input counts, and #930, which exposes existing cache counters in the UI. It does not require adding a cache-TTL setting or changing the API's aggregate totals.
The proposed fix is to preserve the reported cache-write count in both existing converters and subtract it from ordinary input, while keeping the existing top-level cache-read precedence and behavior when the new field is absent.
Contribution
- I am willing to submit a PR for this bug fix myself and will wait for a maintainer's
/approvecomment before proceeding with the contribution.
- Dominant language
- TypeScript
- Stars
- 7.9k
- Forks
- 1.3k
- Avg merge
- 1d 7h
- Merged PRs (30d)
- 96
Getting set up
- No Dockerfile or Docker Compose file
- Has a pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from MoonshotAI/kimi-code
-
Difficulty 2/5 1-3 hours Newbie friendliness 85/100
MoonshotAI/kimi-code#4398 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 85/100
MoonshotAI/kimi-code#4395 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
MoonshotAI/kimi-code#4392 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
MoonshotAI/kimi-code#4381 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 85/100
MoonshotAI/kimi-code#4369 · 1 comment ·
Maintainers usually reply within 1 day
All issues in MoonshotAI/kimi-code
Similar issues
-
type: bug
Difficulty 2/5 1-3 hours Newbie friendliness 79/100
Maintainers usually reply within 1 day
-
0️⃣ phase/new
Difficulty 2/5 1-3 hours Newbie friendliness 82/100
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
-
Difficulty 1/5 Under an hour Newbie friendliness 70/100
microsoft/playwright#43640 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 70/100