Skip to content

Use Claude Code with your ChatGPT, Kimi, Cursor or Grok subscription via a local Anthropic-compatible proxy

Rust

preserve output budget and log usage drift

Raine Virta0e540551dd3c
diff --git a/src/server.ts b/src/server.ts
index c0312ec..b0ff6f6 100644
--- a/src/server.ts
+++ b/src/server.ts
@@ -4,6 +4,7 @@ import { assertAllowedModel, ModelNotAllowedError, resolveModel } from "./transl
import { translateRequest } from "./translate/request.ts"
import { translateStream } from "./translate/stream.ts"
import { accumulateResponse, UpstreamStreamError } from "./translate/accumulate.ts"
+import { mapUsageToAnthropic } from "./translate/reducer.ts"
import { CodexError, postCodex } from "./codex/client.ts"
import { countTokens, countTranslatedTokens } from "./count-tokens.ts"
@@ -31,6 +32,20 @@ interface SessionMessageSnapshot {
translatedInputTokens?: number
}
+function usageWindowTokens(usage: {
+ input_tokens: number
+ output_tokens: number
+ cache_creation_input_tokens: number
+ cache_read_input_tokens: number
+}): number {
+ return (
+ usage.input_tokens +
+ usage.output_tokens +
+ usage.cache_creation_input_tokens +
+ usage.cache_read_input_tokens
+ )
+}
+
interface SessionTimelineState {
seq: number
lastCount?: SessionCountSnapshot
@@ -160,6 +175,8 @@ async function handleMessages(req: Request, reqId: string): Promise<Response> {
toolCount,
stream: wantStream,
sessionId,
+ requestedMaxTokens: body.max_tokens,
+ hasContextManagement: body.context_management !== undefined,
hasJsonSchemaFormat: body.output_config?.format?.type === "json_schema",
})
if (VERBOSE) log.debug("anthropic request body", { reqId, body })
@@ -199,6 +216,9 @@ async function handleMessages(req: Request, reqId: string): Promise<Response> {
inputItems: translated.input.length,
tools: translated.tools?.length ?? 0,
hasInstructions: !!translated.instructions,
+ requestedMaxTokens: body.max_tokens,
+ translatedMaxOutputTokens: translated.max_output_tokens,
+ hasContextManagement: body.context_management !== undefined,
promptCacheKey: translated.prompt_cache_key,
})
if (VERBOSE) log.debug("translated request body", { reqId, body: translated })
@@ -217,6 +237,9 @@ async function handleMessages(req: Request, reqId: string): Promise<Response> {
inputItems: translated.input.length,
translatedToolCount: translated.tools?.length ?? 0,
hasInstructions: !!translated.instructions,
+ requestedMaxTokens: body.max_tokens,
+ translatedMaxOutputTokens: translated.max_output_tokens,
+ hasContextManagement: body.context_management !== undefined,
previousCountReqId: state?.lastCount?.reqId,
previousCountModel: state?.lastCount?.model,
previousCountTokens: state?.lastCount?.tokens,
@@ -257,6 +280,7 @@ async function handleMessages(req: Request, reqId: string): Promise<Response> {
sessionId,
onFinish: VERBOSE
? (finish) => {
+ const mappedUsage = finish.usage ? mapUsageToAnthropic(finish.usage) : undefined
log.info("compaction telemetry", {
reqId,
phase: "upstream_finish",
@@ -269,9 +293,18 @@ async function handleMessages(req: Request, reqId: string): Promise<Response> {
toolCount,
localInputTokens,
translatedInputTokens,
+ requestedMaxTokens: body.max_tokens,
+ translatedMaxOutputTokens: translated.max_output_tokens,
+ hasContextManagement: body.context_management !== undefined,
upstreamInputTokens: finish.usage?.input_tokens ?? 0,
upstreamOutputTokens: finish.usage?.output_tokens ?? 0,
upstreamCachedInputTokens: finish.usage?.input_tokens_details?.cached_tokens ?? 0,
+ upstreamReasoningTokens:
+ finish.usage?.output_tokens_details?.reasoning_tokens ?? 0,
+ mappedInputTokens: mappedUsage?.input_tokens ?? 0,
+ mappedOutputTokens: mappedUsage?.output_tokens ?? 0,
+ mappedCachedInputTokens: mappedUsage?.cache_read_input_tokens ?? 0,
+ mappedContextWindowTokens: mappedUsage ? usageWindowTokens(mappedUsage) : 0,
stopReason: finish.stopReason,
})
}
@@ -302,13 +335,21 @@ async function handleMessages(req: Request, reqId: string): Promise<Response> {
toolCount,
localInputTokens,
translatedInputTokens,
- upstreamInputTokens: result.usage.input_tokens,
- upstreamOutputTokens: result.usage.output_tokens,
- upstreamCachedInputTokens: result.usage.cache_read_input_tokens,
- stopReason: result.stop_reason,
+ requestedMaxTokens: body.max_tokens,
+ translatedMaxOutputTokens: translated.max_output_tokens,
+ hasContextManagement: body.context_management !== undefined,
+ upstreamInputTokens: result.rawUsage?.input_tokens ?? 0,
+ upstreamOutputTokens: result.rawUsage?.output_tokens ?? 0,
+ upstreamCachedInputTokens: result.rawUsage?.input_tokens_details?.cached_tokens ?? 0,
+ upstreamReasoningTokens: result.rawUsage?.output_tokens_details?.reasoning_tokens ?? 0,
+ mappedInputTokens: result.response.usage.input_tokens,
+ mappedOutputTokens: result.response.usage.output_tokens,
+ mappedCachedInputTokens: result.response.usage.cache_read_input_tokens,
+ mappedContextWindowTokens: usageWindowTokens(result.response.usage),
+ stopReason: result.response.stop_reason,
})
}
- return new Response(JSON.stringify(result), {
+ return new Response(JSON.stringify(result.response), {
headers: { "content-type": "application/json" },
})
} catch (err) {
diff --git a/src/translate/accumulate.ts b/src/translate/accumulate.ts
index 9e1ae95..14fb047 100644
--- a/src/translate/accumulate.ts
+++ b/src/translate/accumulate.ts
@@ -1,4 +1,5 @@
import { mapUsageToAnthropic, reduceUpstream } from "./reducer.ts"
+import type { CodexUsage } from "./reducer.ts"
export { UpstreamStreamError } from "./reducer.ts"
@@ -21,6 +22,11 @@ export interface AnthropicNonStreamResponse {
}
}
+export interface AccumulatedResponse {
+ response: AnthropicNonStreamResponse
+ rawUsage?: CodexUsage
+}
+
/**
* Drive the Codex SSE stream to completion through the shared reducer
* and fold the ReducerEvents into a single Anthropic non-streaming
@@ -30,7 +36,7 @@ export interface AnthropicNonStreamResponse {
export async function accumulateResponse(
upstream: ReadableStream<Uint8Array>,
opts: { messageId: string; model: string },
-): Promise<AnthropicNonStreamResponse> {
+): Promise<AccumulatedResponse> {
type Block =
| { kind: "text"; text: string }
| { kind: "tool"; id: string; name: string; args: string }
@@ -39,6 +45,7 @@ export async function accumulateResponse(
const blocks = new Map<number, Block>()
let stopReason: AnthropicNonStreamResponse["stop_reason"] = null
let usage: ReturnType<typeof mapUsageToAnthropic> | undefined
+ let rawUsage: CodexUsage | undefined
for await (const e of reduceUpstream(upstream)) {
switch (e.kind) {
@@ -65,6 +72,7 @@ export async function accumulateResponse(
break
case "finish":
stopReason = e.stopReason
+ rawUsage = e.usage
usage = mapUsageToAnthropic(e.usage)
break
}
@@ -88,18 +96,21 @@ export async function accumulateResponse(
}
return {
- id: opts.messageId,
- type: "message",
- role: "assistant",
- model: opts.model,
- content,
- stop_reason: stopReason,
- stop_sequence: null,
- usage: usage ?? {
- input_tokens: 0,
- output_tokens: 0,
- cache_creation_input_tokens: 0,
- cache_read_input_tokens: 0,
+ rawUsage,
+ response: {
+ id: opts.messageId,
+ type: "message",
+ role: "assistant",
+ model: opts.model,
+ content,
+ stop_reason: stopReason,
+ stop_sequence: null,
+ usage: usage ?? {
+ input_tokens: 0,
+ output_tokens: 0,
+ cache_creation_input_tokens: 0,
+ cache_read_input_tokens: 0,
+ },
},
}
}
diff --git a/src/translate/request.ts b/src/translate/request.ts
index ec5a393..8d2b50f 100644
--- a/src/translate/request.ts
+++ b/src/translate/request.ts
@@ -12,6 +12,7 @@ export interface ResponsesRequest {
instructions?: string
input: ResponsesInputItem[]
tools?: ResponsesTool[]
+ max_output_tokens?: number
tool_choice?:
| "auto"
| "none"
@@ -95,6 +96,7 @@ export function translateRequest(req: AnthropicRequest, opts: TranslateOptions =
}
if (instructions) out.instructions = instructions
if (tools && tools.length) out.tools = tools
+ if (req.max_tokens) out.max_output_tokens = req.max_tokens
if (opts.sessionId) out.prompt_cache_key = opts.sessionId
const effort = req.output_config?.effort
if (effort) out.reasoning = { effort }