fix some errors

packages/cli/src/utils/codeCommand.ts

@@ -62,21 +62,20 @@ export async function executeCodeCommand(
     };
   }
 
-  args.push('--settings', getSettingsPath(`${JSON.stringify(settingsFlag)}`));
-  console.log(args)
-
   // Non-interactive mode for automation environments
   if (config.NON_INTERACTIVE_MODE) {
-    env.CI = "true";
-    env.FORCE_COLOR = "0";
-    env.NODE_NO_READLINE = "1";
-    env.TERM = "dumb";
+    settingsFlag.env = {
+      ...settingsFlag.env,
+      CI: "true",
+      FORCE_COLOR: "0",
+      NODE_NO_READLINE: "1",
+      TERM: "dumb"
+    }
   }
 
-  // Set ANTHROPIC_SMALL_FAST_MODEL if it exists in config
-  if (config?.ANTHROPIC_SMALL_FAST_MODEL) {
-    env.ANTHROPIC_SMALL_FAST_MODEL = config.ANTHROPIC_SMALL_FAST_MODEL;
-  }
+  const settingsFile = await getSettingsPath(`${JSON.stringify(settingsFlag)}`)
+
+  args.push('--settings', settingsFile);
 
   // Increment reference count when command starts
   incrementReferenceCount();

packages/cli/src/utils/index.ts

@@ -3,6 +3,7 @@ import readline from "node:readline";
 import JSON5 from "json5";
 import path from "node:path";
 import { createHash } from "node:crypto";
+import os from "node:os";
 import {
   CONFIG_FILE,
   HOME_DIR, PID_FILE,
@@ -12,7 +13,7 @@ import {
   readPresetFile,
 } from "@CCR/shared";
 import { getServer } from "@CCR/server";
-import { writeFileSync, existsSync, readFileSync } from "fs";
+import { writeFileSync, existsSync, readFileSync, mkdirSync } from "fs";
 import { checkForUpdates, performUpdate } from "./update";
 import { version } from "../../package.json";
 import { spawn } from "child_process";
@@ -254,25 +255,35 @@ export const restartService = async () => {
 
 
 /**
- * 获取设置文件的临时路径
- * 对内容进行 hash，如果临时目录中已存在该 hash 的文件则直接返回，否则创建新文件
- * @param content 设置内容字符串
- * @returns 临时文件的完整路径
+ * Get a temporary path for the settings file
+ * Hash the content and return the file path if it already exists in temp directory,
+ * otherwise create a new file with the content
+ * @param content Settings content string
+ * @returns Full path to the temporary file
  */
-export const getSettingsPath = (content: string): string => {
-  // 对内容进行 hash（使用 SHA256 算法）
+export const getSettingsPath = async (content: string): Promise<string> => {
+  // Hash the content using SHA256 algorithm
   const hash = createHash('sha256').update(content, 'utf-8').digest('hex');
 
+  // Create claude-code-router directory in system temp folder
+  const tempDir = path.join(os.tmpdir(), 'claude-code-router');
   const fileName = `ccr-settings-${hash}.json`;
-  const tempFilePath = path.join(HOME_DIR, fileName);
+  const tempFilePath = path.join(tempDir, fileName);
 
-  // 检查文件是否已存在
-  if (existsSync(tempFilePath)) {
-    return tempFilePath;
+  // Ensure the directory exists
+  try {
+    await fs.access(tempDir);
+  } catch {
+    await fs.mkdir(tempDir, { recursive: true });
   }
 
-  // 文件不存在，创建并写入内容
-  writeFileSync(tempFilePath, content, 'utf-8');
-
+  // Check if the file already exists
+  try {
+    await fs.access(tempFilePath);
+    return tempFilePath;
+  } catch {
+    // File doesn't exist, create and write content
+    await fs.writeFile(tempFilePath, content, 'utf-8');
     return tempFilePath;
   }
+}

packages/core/src/plugins/token-speed.ts

@@ -1,8 +1,8 @@
 import fp from 'fastify-plugin';
 import { CCRPlugin, CCRPluginOptions } from './types';
 import { SSEParserTransform } from '../utils/sse';
-import { Tiktoken } from 'tiktoken';
 import { OutputHandlerConfig, OutputOptions, outputManager } from './output';
+import { ITokenizer, TokenizerConfig } from '../types/tokenizer';
 
 /**
  * Token statistics interface
@@ -45,6 +45,9 @@ interface TokenSpeedOptions extends CCRPluginOptions {
 // Store request-level statistics
 const requestStats = new Map<string, TokenStats>();
 
+// Cache tokenizers by provider and model to avoid repeated initialization
+const tokenizerCache = new Map<string, ITokenizer>();
+
 // Cross-request statistics
 const globalStats = {
   totalRequests: 0,
@@ -94,15 +97,53 @@ export const tokenSpeedPlugin: CCRPlugin = {
       outputManager.setDefaultOptions(opts.outputOptions);
     }
 
-    // Initialize tiktoken encoder
-    let encoding: Tiktoken | null = null;
-    try {
-      const { get_encoding } = await import('tiktoken');
-      encoding = get_encoding('cl100k_base');
-    } catch (error) {
-      fastify.log?.warn('Failed to load tiktoken, falling back to estimation');
+    /**
+     * Get or create tokenizer for a specific provider and model
+     */
+    const getTokenizerForRequest = async (request: any): Promise<ITokenizer | null> => {
+      const tokenizerService = (fastify as any).tokenizerService;
+      if (!tokenizerService) {
+        fastify.log?.warn('TokenizerService not available');
+        return null;
       }
 
+      // Extract provider and model from request
+      // Format: "provider,model" or just "model"
+      if (!request.provider || !request.model) {
+        return null;
+      }
+      const providerName = request.provider;
+      const modelName = request.model;
+
+      // Create cache key
+      const cacheKey = `${providerName}:${modelName}`;
+
+      // Check cache first
+      if (tokenizerCache.has(cacheKey)) {
+        return tokenizerCache.get(cacheKey)!;
+      }
+
+      // Get tokenizer config for this model
+      const tokenizerConfig = tokenizerService.getTokenizerConfigForModel(providerName, modelName);
+
+      if (!tokenizerConfig) {
+        // No specific config, use fallback
+        fastify.log?.debug(`No tokenizer config for ${providerName}:${modelName}, using fallback`);
+        return null;
+      }
+
+      try {
+        // Create and cache tokenizer
+        const tokenizer = await tokenizerService.getTokenizer(tokenizerConfig);
+        tokenizerCache.set(cacheKey, tokenizer);
+        fastify.log?.info(`Created tokenizer for ${providerName}:${modelName} - ${tokenizer.name}`);
+        return tokenizer;
+      } catch (error: any) {
+        fastify.log?.warn(`Failed to create tokenizer for ${providerName}:${modelName}: ${error.message}`);
+        return null;
+      }
+    };
+
     // Add onSend hook to intercept streaming responses
     fastify.addHook('onSend', async (request, reply, payload) => {
       // Only handle streaming responses
@@ -122,6 +163,10 @@ export const tokenSpeedPlugin: CCRPlugin = {
         tokensPerSecond: 0,
         contentBlocks: []
       });
+
+      // Get tokenizer for this specific request
+      const tokenizer = await getTokenizerForRequest(request);
+
       // Tee the stream: one for stats, one for the client
       const [originalStream, statsStream] = payload.tee();
 
@@ -156,8 +201,8 @@ export const tokenSpeedPlugin: CCRPlugin = {
             // Detect content_block_delta event (incremental tokens)
             if (data.event === 'content_block_delta' && data.data?.delta?.type === 'text_delta') {
               const text = data.data.delta.text;
-              const tokenCount = encoding
-                ? encoding.encode(text).length
+              const tokenCount = tokenizer
+                ? (tokenizer.encodeText ? tokenizer.encodeText(text).length : estimateTokens(text))
                 : estimateTokens(text);
 
               stats.tokenCount += tokenCount;

packages/core/src/server.ts

@@ -38,6 +38,7 @@ import { sessionUsageCache } from "./utils/cache";
 declare module "fastify" {
   interface FastifyRequest {
     provider?: string;
+    model?: string;
   }
   interface FastifyInstance {
     _server?: Server;
@@ -226,6 +227,7 @@ class Server {
               const [provider, ...model] = body.model.split(",");
               body.model = model.join(",");
               req.provider = provider;
+              req.model = model;
               return;
             } catch (err) {
               req.log.error({error: err}, "Error in modelProviderMiddleware:");

packages/core/src/tokenizer/huggingface-tokenizer.ts

@@ -164,6 +164,16 @@ export class HuggingFaceTokenizer implements ITokenizer {
     return this.tokenizer !== null;
   }
 
+  /**
+   * Encode text to tokens (for simple text tokenization)
+   */
+  encodeText(text: string): number[] {
+    if (!this.tokenizer) {
+      throw new Error("Tokenizer not initialized");
+    }
+    return this.tokenizer.encode(text).ids;
+  }
+
   dispose(): void {
     this.tokenizer = null;
   }

packages/core/src/tokenizer/tiktoken-tokenizer.ts

@@ -99,6 +99,17 @@ export class TiktokenTokenizer implements ITokenizer {
     return this.encoding !== undefined;
   }
 
+  /**
+   * Encode text to tokens (for simple text tokenization)
+   */
+  encodeText(text: string): number[] {
+    const encoding = this.encoding;
+    if (!encoding) {
+      throw new Error("Encoding not initialized");
+    }
+    return Array.from(encoding.encode(text));
+  }
+
   dispose(): void {
     if (this.encoding) {
       this.encoding.free();

packages/core/src/types/tokenizer.d.ts

@@ -112,6 +112,9 @@ export interface ITokenizer {
   /** Count tokens for a given request */
   countTokens(request: TokenizeRequest): Promise<number>;
 
+  /** Encode text to tokens (for simple text tokenization) */
+  encodeText?(text: string): number[];
+
   /** Check if tokenizer is initialized */
   isInitialized(): boolean;