What
When running as an ACP server (gemini --acp), the token usage reported for a prompt turn only includes input and output tokens. The cached and thought/reasoning token counts are dropped — even though Gemini's usageMetadata provides them and the CLI already consumes them elsewhere (telemetry, UI, chat recording).
In packages/cli/src/acp/acpSession.ts, the GeminiEventType.Finished handler reads only promptTokenCount and candidatesTokenCount:
const usage = event.value.usageMetadata;
if (usage) {
turnInputTokens = usage.promptTokenCount ?? turnInputTokens;
turnOutputTokens = usage.candidatesTokenCount ?? turnOutputTokens;
}cachedContentTokenCount and thoughtsTokenCount are never read, and the emitted _meta.quota.token_count contains only input_tokens / output_tokens. By contrast, these fields are read in packages/core/src/agent/event-translator.ts, packages/core/src/telemetry/types.ts, and packages/core/src/services/chatRecordingService.ts — so the data is available; only the ACP path discards it.
Why it matters
ACP clients that estimate cost from token counts treat all input tokens as uncached, because the cached count is absent/0. For long agentic sessions where most of the prompt is cached turn-to-turn, this overstates the estimated cost by ~3× relative to the real (cached) spend.
We hit this analyzing OpenHands eval runs through the Gemini CLI ACP harness: every instance reports cache_read_tokens = 0, so the token-based cost estimate is ~3× the actual proxy bill (the underlying spend already benefits from caching — only the reporting is wrong).
Relationship to #24280
This is related to but distinct from #24280 (“populate the standard PromptResponse.usage field”). That issue's proposed payload lists only input_tokens / output_tokens / total_tokens, so implementing it as written would not close the cache-token gap. The standard ACP Usage type already has cachedReadTokens and thoughtTokens fields — they just need to be populated.
Proposed fix
In acpSession.ts, also capture usageMetadata.cachedContentTokenCount and usageMetadata.thoughtsTokenCount, and populate the standard ACP PromptResponse.usage field (inputTokens, outputTokens, cachedReadTokens, thoughtTokens, totalTokens). This aligns Gemini CLI with Claude Agent ACP / Codex ACP and lets any ACP client read accurate usage without Gemini-specific _meta.quota parsing. The existing _meta.quota payload can be kept for backward compatibility.
I have a small PR ready for this.
What
When running as an ACP server (
gemini --acp), the token usage reported for a prompt turn only includes input and output tokens. The cached and thought/reasoning token counts are dropped — even though Gemini'susageMetadataprovides them and the CLI already consumes them elsewhere (telemetry, UI, chat recording).In
packages/cli/src/acp/acpSession.ts, theGeminiEventType.Finishedhandler reads onlypromptTokenCountandcandidatesTokenCount:cachedContentTokenCountandthoughtsTokenCountare never read, and the emitted_meta.quota.token_countcontains onlyinput_tokens/output_tokens. By contrast, these fields are read inpackages/core/src/agent/event-translator.ts,packages/core/src/telemetry/types.ts, andpackages/core/src/services/chatRecordingService.ts— so the data is available; only the ACP path discards it.Why it matters
ACP clients that estimate cost from token counts treat all input tokens as uncached, because the cached count is absent/
0. For long agentic sessions where most of the prompt is cached turn-to-turn, this overstates the estimated cost by ~3× relative to the real (cached) spend.We hit this analyzing OpenHands eval runs through the Gemini CLI ACP harness: every instance reports
cache_read_tokens = 0, so the token-based cost estimate is ~3× the actual proxy bill (the underlying spend already benefits from caching — only the reporting is wrong).Relationship to #24280
This is related to but distinct from #24280 (“populate the standard
PromptResponse.usagefield”). That issue's proposed payload lists onlyinput_tokens/output_tokens/total_tokens, so implementing it as written would not close the cache-token gap. The standard ACPUsagetype already hascachedReadTokensandthoughtTokensfields — they just need to be populated.Proposed fix
In
acpSession.ts, also captureusageMetadata.cachedContentTokenCountandusageMetadata.thoughtsTokenCount, and populate the standard ACPPromptResponse.usagefield (inputTokens,outputTokens,cachedReadTokens,thoughtTokens,totalTokens). This aligns Gemini CLI with Claude Agent ACP / Codex ACP and lets any ACP client read accurate usage without Gemini-specific_meta.quotaparsing. The existing_meta.quotapayload can be kept for backward compatibility.I have a small PR ready for this.