fix(acp): report token usage on PromptResponse#1858
Conversation
The ACP adapter left PromptResponse.usage unset, so ACP clients driving kimi (Multica, OpenCode, Claude Code) had zero token/cost visibility even though the SDK tracks usage internally. Populate PromptResponse.usage from session.getUsage() on turn end, mapping the SDK's cumulative total to the ACP Usage shape. Reading usage is best-effort: any failure resolves the prompt with stopReason alone rather than hanging or rejecting.
🦋 Changeset detectedLatest commit: 6f621b9 The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d8a973ef3e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| Promise.resolve() | ||
| .then(() => this.session.getUsage()) | ||
| .then((usage) => { | ||
| resolve({ stopReason, usage: sessionUsageToAcpUsage(usage) }); |
There was a problem hiding this comment.
Because the usage read is only added in runTurnBody, ACP prompts intercepted by runBuiltInCommand still return the old { stopReason: 'end_turn' } path without usage. In particular, /compact calls runCompactCommand/session.compact, and compaction records model token usage, so ACP clients still see no PromptResponse.usage for a token-consuming compaction prompt; apply the same best-effort usage wrapping to that built-in return path as well.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Addressed in 6f621b9. Extracted a shared promptResponseWithUsage helper and wired it into both runTurnBody and the runBuiltInCommand return path, so /compact (and every other built-in) now reports PromptResponse.usage. Added a test in session-slash.test.ts covering the built-in return path. acp-adapter: 309 tests passed, typecheck clean, oxlint 0 errors.
Codex review (P2): the usage read was only wired into runTurnBody, so ACP prompts handled by runBuiltInCommand (notably /compact, which runs a token-consuming compaction) still returned end_turn with no usage. Extract a shared promptResponseWithUsage helper and use it from both the model-turn path and the built-in return path so every terminal prompt reports usage.
1 participant
Related Issue
Resolve #1855
Problem
The ACP adapter leaves
PromptResponse.usageunset and never emits usage, so ACP clients driving kimi over ACP (Multica, OpenCode, Claude Code using kimi as a backend) have zero visibility into token consumption and cost — even though the SDK already tracks usage internally (session.getUsage(), serialized towire.jsonl). On Multica specifically,runtime usagereturns zero rows for the kimi runtime while other providers on the same daemon return full per-model breakdowns.What changed
PromptResponse.usageis now populated on turn end fromsession.getUsage(), mapping the SDK's cumulativetotal(TokenUsage) to the ACPUsageshape:inputOther→inputTokensoutput→outputTokensinputCacheRead→cachedReadTokensinputCacheCreation→cachedWriteTokenstotalTokensgetUsage()the adapter already calls for the/usagebuilt-in, so no new RPC or event accumulation is introduced.getUsage) resolves the prompt withstopReasonalone rather than hanging or rejecting. ThePromise.resolve().then(() => this.session.getUsage())wrapper also converts a synchronous throw into a caught rejection so a completed turn can never be left pending.sessionUsageToAcpUsageinevents-map.tsand two tests insession-prompt.test.ts(usage reported; usage-read failure degrades gracefully).Verification:
pnpm --filter @moonshot-ai/acp-adapter test(308 passed),typecheckclean,oxlint --type-aware0 errors on changed files.Checklist
gen-changesetsskill, or this PR needs no changeset.gen-docsskill, or this PR needs no doc update.