Skip to content

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

Rust

share system token counting

Raine Virtaee93378b4d87
diff --git a/src/providers/codex/count-tokens.ts b/src/providers/codex/count-tokens.ts
index 3a1253f..b45f867 100644
--- a/src/providers/codex/count-tokens.ts
+++ b/src/providers/codex/count-tokens.ts
@@ -1,17 +1,13 @@
import { encode } from "gpt-tokenizer/model/gpt-4o";
import type { AnthropicRequest } from "../../anthropic/schema.ts";
import type { ResponsesRequest } from "./translate/request.ts";
-import { flattenSystemText } from "../translate/anthropic-content.ts";
-import { countAnthropicRequestTokens } from "../shared/count-tokens.ts";
+import { countAnthropicRequestTokensWithSystem } from "../shared/count-tokens.ts";
import { countToolSchemaTokens } from "../shared/tool-schema.ts";
const IMAGE_TOKEN_ESTIMATE = 2000;
export function countTokens(req: AnthropicRequest): number {
- let total = 0;
- const instructions = flattenSystemText(req.system);
- if (instructions) total += encode(instructions).length;
- total += countAnthropicRequestTokens({
+ return countAnthropicRequestTokensWithSystem({
req,
countToken: (value) => encode(value).length,
tools: req.tools,
@@ -19,7 +15,6 @@ export function countTokens(req: AnthropicRequest): number {
readToolDescription: (tool) => tool.description,
readToolSchema: (tool) => tool.input_schema,
});
- return total;
}
export function countTranslatedTokens(
diff --git a/src/providers/kimi/count-tokens.ts b/src/providers/kimi/count-tokens.ts
index 2ef710f..4d3b188 100644
--- a/src/providers/kimi/count-tokens.ts
+++ b/src/providers/kimi/count-tokens.ts
@@ -1,8 +1,7 @@
import { encode } from "gpt-tokenizer/model/gpt-4o";
import type { AnthropicRequest } from "../../anthropic/schema.ts";
import type { KimiChatRequest } from "./translate/request.ts";
-import { flattenSystemText } from "../translate/anthropic-content.ts";
-import { countAnthropicRequestTokens } from "../shared/count-tokens.ts";
+import { countAnthropicRequestTokensWithSystem } from "../shared/count-tokens.ts";
import { countToolSchemaTokens } from "../shared/tool-schema.ts";
const IMAGE_TOKEN_ESTIMATE = 2000;
@@ -10,10 +9,7 @@ const IMAGE_TOKEN_ESTIMATE = 2000;
// Approximate: Kimi's tokenizer isn't gpt-tokenizer, but Claude Code's
// compaction logic only needs a monotonic estimate, not an exact count.
export function countTokens(req: AnthropicRequest): number {
- let total = 0;
- const system = flattenSystemText(req.system);
- if (system) total += encode(system).length;
- total += countAnthropicRequestTokens({
+ return countAnthropicRequestTokensWithSystem({
req,
countToken: (value) => encode(value).length,
tools: req.tools,
@@ -22,7 +18,6 @@ export function countTokens(req: AnthropicRequest): number {
readToolSchema: (tool) => tool.input_schema,
includeThinking: true,
});
- return total;
}
export function countTranslatedTokens(req: KimiChatRequest): number {
diff --git a/src/providers/shared/count-tokens.ts b/src/providers/shared/count-tokens.ts
index 4dcf30e..e6538e9 100644
--- a/src/providers/shared/count-tokens.ts
+++ b/src/providers/shared/count-tokens.ts
@@ -1,5 +1,9 @@
import { countToolSchemaTokens } from "./tool-schema.ts";
-import { normalizeContent, toolResultToString } from "../translate/anthropic-content.ts";
+import {
+ flattenSystemText,
+ normalizeContent,
+ toolResultToString,
+} from "../translate/anthropic-content.ts";
import type { AnthropicRequest } from "../../anthropic/schema.ts";
const IMAGE_TOKEN_ESTIMATE = 2000;
@@ -70,3 +74,27 @@ export function countAnthropicRequestTokens<TTool>(
total += req.messages.length * 4;
return total;
}
+
+export function countAnthropicRequestTokensWithSystem<TTool>(
+ options: CountAnthropicRequestTokenOptions<TTool> & {
+ includeThinking?: false;
+ },
+): number;
+
+export function countAnthropicRequestTokensWithSystem<TTool>(
+ options: CountAnthropicRequestTokenOptions<TTool> & {
+ includeThinking: true;
+ },
+): number;
+
+export function countAnthropicRequestTokensWithSystem<TTool>(
+ options: CountAnthropicRequestTokenOptions<TTool> & { includeThinking?: boolean },
+): number {
+ let total = 0;
+ const system = flattenSystemText(options.req.system);
+ if (system) total += options.countToken(system);
+ total += options.includeThinking
+ ? countAnthropicRequestTokens({ ...options, includeThinking: true })
+ : countAnthropicRequestTokens({ ...options, includeThinking: false });
+ return total;
+}