[Bug]: Ollama native usage parser ignores prompt_eval_cached_count (cache shows 0) #8101

Client or integration

Kilo Code through OpenCodex

Area

Provider adapter / usage reporting

Summary

While routing Kilo Code requests to Ollama Cloud's DeepSeek V4.1 Flash through OpenCodex, the Usage dashboard consistently reports 0 cached input tokens. A concrete source-level cause is that the ollama-native adapter's usageFromNative() currently maps prompt_eval_count and eval_count but drops the native prompt_eval_cached_count response field entirely. If upstream returns nonzero cached tokens, those counts cannot reach OpenCodex usage accounting through this function.

This issue concerns missing accounting for a native response field, not a claim that Ollama Cloud necessarily served cached tokens in the observed session. We have not captured a raw Ollama Cloud response proving a nonzero count.

Reproduction

Deterministic source-level reproduction (no account or live API call required):

1. Open src/adapters/ollama-native.ts on the current default branch and locate usageFromNative().

2. Trace its handling of a syntactically valid Ollama native final response usage payload with these fields:

   {
     "prompt_eval_count": 2000,
     "prompt_eval_cached_count": 1500,
     "eval_count": 100
   }

3. The function reads prompt_eval_count and eval_count, then returns only { inputTokens: 2000, outputTokens: 100 }. There is no cached-token field in the returned OcxUsage object.

4. Thus the native cached-token count of 1500, if supplied upstream, is lost before downstream usage recording and dashboard accounting.

Actual code:

function usageFromNative(value: JsonRecord | undefined): OcxUsage | undefined {
  if (!value) return undefined;
  const input = isFiniteNonNegativeInteger(value.prompt_eval_count) ? value.prompt_eval_count : undefined;
  const output = isFiniteNonNegativeInteger(value.eval_count) ? value.eval_count : undefined;
  if (input === undefined && output === undefined) return undefined;
  return { inputTokens: input ?? 0, outputTokens: output ?? 0 };
}

Observed UI workflow:

1. Configure ollama-cloud in OCX using the ollama-native adapter.
2. Select ollama-cloud/deepseek-v4.1-flash in Kilo Code through OCX.
3. Make repeated coding turns within the same session.
4. Inspect OCX's Usage dashboard; cached input tokens remain 0.

The input above is a synthetic parser testcase, not a captured upstream response. A regression test could feed this payload through the native adapter's streamed or nonstreamed final-response processing and assert that the resulting usage event carries the cache-read token count.

Expected behavior

• Parse prompt_eval_cached_count when present in the Ollama native response.
• Map it to OCX's canonical cache-read usage field, retaining correct total input accounting.
• Surface cache-read tokens in the ledger/dashboard (and distinguish unavailable cache information from an explicit zero when possible).

Actual behavior

The parser ignores prompt_eval_cached_count entirely, and a user sees 0 cached tokens for Ollama Cloud sessions.

Version

OpenCodex v2.96.0 (latest stable release as of 2026-11-02, per user confirmation and official release). Source-level reproduction is verified against the default main branch corresponding to this release.

Operating system

Cross-platform source-level defect (pure TypeScript mapping; not dependent on host OS). The host operating system and version for the observed Kilo Code session were not captured.

Provider and model

Ollama Cloud, native /api/chat adapter, DeepSeek V4.1 Flash.

Additional context

Related but potentially distinct upstream Ollama issue: ollama/ollama#15758 (Ollama Cloud sometimes reports zero cached tokens). Even if the upstream service omits this metric for some requests, OCX should not discard it when supplied.

Checks

• Searched existing issues for Ollama cached-token accounting; no exact duplicate found.
• Reproduced the field omission by source inspection (synthetic input; not a live API assertion).
• No credentials, tokens, or private user data included.