@jeffreycao/copilot-api 1.4.2 → 1.4.3

package/README.md

@@ -22,7 +22,22 @@
 ---
 
 > [!NOTE]
-> [opencode](https://github.com/sst/opencode) already ships with a built-in GitHub Copilot provider, so you may not need this project for basic usage. This proxy is still useful if you want OpenCode to talk to Copilot through `@ai-sdk/anthropic`, preserve Anthropic Messages semantics for tool use, prefer the native Messages API over plain Chat Completions for Claude-family models, use `gpt-5.4` phase-aware commentary, or fine-tune premium-request usage with small-model fallbacks.
+> [opencode](https://github.com/sst/opencode) already ships with a built-in GitHub Copilot provider, so you may not need this project for basic usage. This proxy is still useful if you want OpenCode to talk to Copilot through `@ai-sdk/anthropic`, preserve Anthropic Messages semantics for tool use, prefer the native Messages API over Chat Completions API for Claude-family models, use gpt phase-aware commentary, or optimize premium requests.
+
+---
+## Important Notes
+
+> [!IMPORTANT]
+> **Before using, please be aware of the following:**
+>
+> 1. **Claude Code model ID configuration:** When using with Claude Code, please configure the model ID as `claude-opus-4-6` or `claude-opus-4.6` (without the `[1m]` suffix, exceeding GitHub Copilot's context window limit too much may lead to being banned).
+>
+> 2. **Recommend for Opencode:** When using with opencode, we recommend starting with the opencode OAuth app. This approach behaves identically to opencode's built-in GitHub Copilot provider with no Terms of Service risk:
+>    ```sh
+>    npx @jeffreycao/copilot-api@latest --oauth-app=opencode start
+>    ```
+>
+> 3. **Disable multi agent when using codex:** If you're using codex via GitHub Copilot, it's recommended to disable the multi agent feature. Currently, GitHub Copilot charges based on the last message being a user role when using codex, and the billing logic has not been adjusted.
 
 ---
 
@@ -36,7 +51,7 @@ Compared with routing everything through plain Chat Completions compatibility, t
 
 - **OpenAI & Anthropic Compatibility**: Exposes GitHub Copilot as an OpenAI-compatible (`/v1/responses`, `/v1/chat/completions`, `/v1/models`, `/v1/embeddings`) and Anthropic-compatible (`/v1/messages`) API.
 - **Anthropic-First Routing for Claude Models**: When a model supports Copilot's native `/v1/messages` endpoint, the proxy prefers it over `/responses` or `/chat/completions`, preserving Anthropic-style `tool_use` / `tool_result` flows and more Claude-native behavior.
-- **Fewer Unnecessary Premium Requests**: Reduces wasted premium usage by routing warmup and compact/background requests to `smallModel`, merging `tool_result` follow-ups back into the tool flow, and treating resumed tool turns as continuation traffic instead of fresh premium interactions.
+- **Fewer Unnecessary Premium Requests**: Reduces wasted premium usage by routing warmup requests to `smallModel`, merging `tool_result` follow-ups back into the tool flow, and treating resumed tool turns as continuation traffic instead of fresh premium interactions.
 - **Phase-Aware `gpt-5.4` and `gpt-5.3-codex`**: These models can emit user-friendly commentary before deeper reasoning or tool use, so long-running coding actions are easier to understand instead of appearing as a sudden tool burst.
 - **Claude Native Beta Support**: On the Messages API path, supports Anthropic-native capabilities such as `interleaved-thinking`, `advanced-tool-use`, and `context-management`, which are difficult or unavailable through plain Chat Completions compatibility.
 - **Subagent Marker Integration**: Optional Claude Code and opencode plugins can inject `__SUBAGENT_MARKER__...` and propagate `x-session-id` so subagent traffic keeps the correct root session and agent/user semantics.
@@ -52,6 +67,7 @@ Compared with routing everything through plain Chat Completions compatibility, t
 - **GitHub Enterprise Support**: Connect to GHE.com by setting `COPILOT_API_ENTERPRISE_URL` environment variable (e.g., `company.ghe.com`) or using `--enterprise-url=company.ghe.com` command line option.
 - **Custom Data Directory**: Change the default data directory (where tokens and config are stored) by setting `COPILOT_API_HOME` environment variable or using `--api-home=/path/to/dir` command line option.
 - **Multi-Provider Anthropic Proxy Routes**: Add global provider configs and call external Anthropic-compatible APIs via `/:provider/v1/messages` and `/:provider/v1/models`.
+- **Accurate Claude Token Counting**: Optionally forward `/v1/messages/count_tokens` requests for Claude models to Anthropic's free token counting endpoint for exact counts instead of GPT tokenizer estimation.
 
 ## Better Agent Semantics
 
@@ -72,7 +88,6 @@ Supported `anthropic-beta` values are filtered and forwarded on the native Messa
 The proxy includes request-accounting safeguards designed for tool-heavy coding workflows:
 
 - tool-less warmup or probe requests can be forced onto `smallModel` so background checks do not spend premium usage;
-- compact/background requests can be downgraded to `smallModel` automatically;
 - mixed `tool_result` + reminder text blocks are merged back into the `tool_result` flow instead of being counted like fresh user turns;
 - `x-initiator` is derived from the latest message or item, not stale assistant history.
 
@@ -92,6 +107,23 @@ The marker flow uses `__SUBAGENT_MARKER__...` inside a `<system-reminder>` block
 
 Optional marker producers are included for both Claude Code and opencode; see [Subagent Marker Integration](#subagent-marker-integration-optional) below for setup details.
 
+### Accurate Claude token counting
+
+By default, `/v1/messages/count_tokens` estimates Claude token counts using the GPT `o200k_base` tokenizer with a 1.15x multiplier. This consistently underestimates actual Claude token usage, which can cause tools like Claude Code to compact too late and hit "prompt token count exceeds limit" errors.
+
+When an Anthropic API key is configured, the proxy forwards Claude model token counting requests to [Anthropic's real `/v1/messages/count_tokens` endpoint](https://docs.anthropic.com/en/docs/build-with-claude/token-counting) instead. This returns exact counts and eliminates the estimation mismatch. Non-Claude models and failures fall back to the GPT tokenizer estimation automatically.
+
+**Setup:**
+
+1. Create an Anthropic API account at [console.anthropic.com](https://console.anthropic.com) and add a minimum $5 credit balance (required to activate the API key, but the token counting endpoint itself is free)
+2. Create an API key from Settings > API Keys
+3. Configure the key via **one** of:
+   - `config.json`: set `"anthropicApiKey": "sk-ant-..."`
+   - Environment variable: `ANTHROPIC_API_KEY=sk-ant-...`
+
+> [!NOTE]
+> Anthropic's `/v1/messages/count_tokens` endpoint is **free** (no per-token cost). It is rate-limited to 100 RPM at Tier 1. The $5 credit purchase is only needed to activate API access — the token counting calls themselves cost nothing.
+
 ## Demo
 
 https://github.com/user-attachments/assets/7654b383-669d-4eb9-b23c-06d7aefee8c5
@@ -276,7 +308,8 @@ The following command line options are available for the `start` command:
       "gpt-5.4": "xhigh"
     },
     "useFunctionApplyPatch": true,
-    "useMessagesApi": true
+    "useMessagesApi": true,
+    "anthropicApiKey": ""
   }
   ```
 - **auth.apiKeys:** API keys used for request authentication. Supports multiple keys for rotation. Requests can authenticate with either `x-api-key: <key>` or `Authorization: Bearer <key>`. If empty or omitted, authentication is disabled.
@@ -295,6 +328,7 @@ The following command line options are available for the `start` command:
 - **modelReasoningEfforts:** Per-model `reasoning.effort` sent to the Copilot Responses API. Allowed values are `none`, `minimal`, `low`, `medium`, `high`, and `xhigh`. If a model isn’t listed, `high` is used by default.
 - **useFunctionApplyPatch:** When `true`, the server will convert any custom tool named `apply_patch` in Responses payloads into an OpenAI-style function tool (`type: "function"`) with a parameter schema so assistants can call it using function-calling semantics to edit files. Set to `false` to leave tools unchanged. Defaults to `true`.
 - **useMessagesApi:** When `true`, Claude-family models that support Copilot's native `/v1/messages` endpoint will use the Messages API; otherwise they fall back to `/chat/completions`. Set to `false` to disable Messages API routing and always use `/chat/completions`. Defaults to `true`.
+- **anthropicApiKey:** Anthropic API key used for accurate Claude token counting (see [Accurate Claude Token Counting](#accurate-claude-token-counting) below). Can also be set via the `ANTHROPIC_API_KEY` environment variable. If not set, token counting falls back to GPT tokenizer estimation.
 
 Edit this file to customize prompts or swap in your own fast model. Restart the server (or rerun the command) after changes so the cached config is refreshed.
 

package/dist/{config-DIqcOsnZ.js → config-bVq-BhC7.js}

@@ -166,7 +166,10 @@ function getProviderConfig(name) {
 function isMessagesApiEnabled() {
 	return getConfig().useMessagesApi ?? true;
 }
+function getAnthropicApiKey() {
+	return getConfig().anthropicApiKey ?? process.env.ANTHROPIC_API_KEY ?? void 0;
+}
 
 //#endregion
-export { getConfig, getExtraPromptForModel, getProviderConfig, getReasoningEffortForModel, getSmallModel, isMessagesApiEnabled, isResponsesApiContextManagementModel, mergeConfigWithDefaults };
-//# sourceMappingURL=config-DIqcOsnZ.js.map
+export { getAnthropicApiKey, getConfig, getExtraPromptForModel, getProviderConfig, getReasoningEffortForModel, getSmallModel, isMessagesApiEnabled, isResponsesApiContextManagementModel, mergeConfigWithDefaults };
+//# sourceMappingURL=config-bVq-BhC7.js.map

package/dist/config-bVq-BhC7.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"config-bVq-BhC7.js","names":["defaultConfig: AppConfig","cachedConfig: AppConfig | null"],"sources":["../src/lib/config.ts"],"sourcesContent":["import consola from \"consola\"\nimport fs from \"node:fs\"\n\nimport { PATHS } from \"./paths\"\n\nexport interface AppConfig {\n  auth?: {\n    apiKeys?: Array<string>\n  }\n  providers?: Record<string, ProviderConfig>\n  extraPrompts?: Record<string, string>\n  smallModel?: string\n  responsesApiContextManagementModels?: Array<string>\n  modelReasoningEfforts?: Record<\n    string,\n    \"none\" | \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\"\n  >\n  useFunctionApplyPatch?: boolean\n  useMessagesApi?: boolean\n  anthropicApiKey?: string\n}\n\nexport interface ModelConfig {\n  temperature?: number\n  topP?: number\n  topK?: number\n}\n\nexport interface ProviderConfig {\n  type?: string\n  enabled?: boolean\n  baseUrl?: string\n  apiKey?: string\n  models?: Record<string, ModelConfig>\n  adjustInputTokens?: boolean\n}\n\nexport interface ResolvedProviderConfig {\n  name: string\n  type: \"anthropic\"\n  baseUrl: string\n  apiKey: string\n  models?: Record<string, ModelConfig>\n  adjustInputTokens?: boolean\n}\n\nconst gpt5ExplorationPrompt = `## Exploration and reading files\n- **Think first.** Before any tool call, decide ALL files/resources you will need.\n- **Batch everything.** If you need multiple files (even from different places), read them together.\n- **multi_tool_use.parallel** Use multi_tool_use.parallel to parallelize tool calls and only this.\n- **Only make sequential calls if you truly cannot know the next file without seeing a result first.**\n- **Workflow:** (a) plan all needed reads → (b) issue one parallel batch → (c) analyze results → (d) repeat if new, unpredictable reads arise.`\n\nconst gpt5CommentaryPrompt = `# Working with the user\n\nYou interact with the user through a terminal. You have 2 ways of communicating with the users:  \n- Share intermediary updates in \\`commentary\\` channel.  \n- After you have completed all your work, send a message to the \\`final\\` channel.  \n\n## Intermediary updates\n\n- Intermediary updates go to the \\`commentary\\` channel.\n- User updates are short updates while you are working, they are NOT final answers.\n- You use 1-2 sentence user updates to communicate progress and new information to the user as you are doing work.\n- Do not begin responses with conversational interjections or meta commentary. Avoid openers such as acknowledgements (“Done —”, “Got it”, “Great question, ”) or framing phrases.\n- You provide user updates frequently, every 20s.\n- Before exploring or doing substantial work, you start with a user update acknowledging the request and explaining your first step. You should include your understanding of the user request and explain what you will do. Avoid commenting on the request or using starters such as \"Got it -\" or \"Understood -\" etc.\n- When exploring, e.g. searching, reading files, you provide user updates as you go, every 20s, explaining what context you are gathering and what you've learned. Vary your sentence structure when providing these updates to avoid sounding repetitive - in particular, don't start each sentence the same way.\n- After you have sufficient context, and the work is substantial, you provide a longer plan (this is the only user update that may be longer than 2 sentences and can contain formatting).\n- Before performing file edits of any kind, you provide updates explaining what edits you are making.\n- As you are thinking, you very frequently provide updates even if not taking any actions, informing the user of your progress. You interrupt your thinking and send multiple updates in a row if thinking for more than 100 words.\n- Tone of your updates MUST match your personality.`\n\nconst defaultConfig: AppConfig = {\n  auth: {\n    apiKeys: [],\n  },\n  providers: {},\n  extraPrompts: {\n    \"gpt-5-mini\": gpt5ExplorationPrompt,\n    \"gpt-5.3-codex\": gpt5CommentaryPrompt,\n    \"gpt-5.4-mini\": gpt5CommentaryPrompt,\n    \"gpt-5.4\": gpt5CommentaryPrompt,\n  },\n  smallModel: \"gpt-5-mini\",\n  responsesApiContextManagementModels: [],\n  modelReasoningEfforts: {\n    \"gpt-5-mini\": \"low\",\n    \"gpt-5.3-codex\": \"xhigh\",\n    \"gpt-5.4-mini\": \"xhigh\",\n    \"gpt-5.4\": \"xhigh\",\n  },\n  useFunctionApplyPatch: true,\n  useMessagesApi: true,\n}\n\nlet cachedConfig: AppConfig | null = null\n\nfunction ensureConfigFile(): void {\n  try {\n    fs.accessSync(PATHS.CONFIG_PATH, fs.constants.R_OK | fs.constants.W_OK)\n  } catch {\n    fs.mkdirSync(PATHS.APP_DIR, { recursive: true })\n    fs.writeFileSync(\n      PATHS.CONFIG_PATH,\n      `${JSON.stringify(defaultConfig, null, 2)}\\n`,\n      \"utf8\",\n    )\n    try {\n      fs.chmodSync(PATHS.CONFIG_PATH, 0o600)\n    } catch {\n      return\n    }\n  }\n}\n\nfunction readConfigFromDisk(): AppConfig {\n  ensureConfigFile()\n  try {\n    const raw = fs.readFileSync(PATHS.CONFIG_PATH, \"utf8\")\n    if (!raw.trim()) {\n      fs.writeFileSync(\n        PATHS.CONFIG_PATH,\n        `${JSON.stringify(defaultConfig, null, 2)}\\n`,\n        \"utf8\",\n      )\n      return defaultConfig\n    }\n    return JSON.parse(raw) as AppConfig\n  } catch (error) {\n    consola.error(\"Failed to read config file, using default config\", error)\n    return defaultConfig\n  }\n}\n\nfunction mergeDefaultConfig(config: AppConfig): {\n  mergedConfig: AppConfig\n  changed: boolean\n} {\n  const extraPrompts = config.extraPrompts ?? {}\n  const defaultExtraPrompts = defaultConfig.extraPrompts ?? {}\n  const modelReasoningEfforts = config.modelReasoningEfforts ?? {}\n  const defaultModelReasoningEfforts = defaultConfig.modelReasoningEfforts ?? {}\n\n  const missingExtraPromptModels = Object.keys(defaultExtraPrompts).filter(\n    (model) => !Object.hasOwn(extraPrompts, model),\n  )\n\n  const missingReasoningEffortModels = Object.keys(\n    defaultModelReasoningEfforts,\n  ).filter((model) => !Object.hasOwn(modelReasoningEfforts, model))\n\n  const hasExtraPromptChanges = missingExtraPromptModels.length > 0\n  const hasReasoningEffortChanges = missingReasoningEffortModels.length > 0\n\n  if (!hasExtraPromptChanges && !hasReasoningEffortChanges) {\n    return { mergedConfig: config, changed: false }\n  }\n\n  return {\n    mergedConfig: {\n      ...config,\n      extraPrompts: {\n        ...defaultExtraPrompts,\n        ...extraPrompts,\n      },\n      modelReasoningEfforts: {\n        ...defaultModelReasoningEfforts,\n        ...modelReasoningEfforts,\n      },\n    },\n    changed: true,\n  }\n}\n\nexport function mergeConfigWithDefaults(): AppConfig {\n  const config = readConfigFromDisk()\n  const { mergedConfig, changed } = mergeDefaultConfig(config)\n\n  if (changed) {\n    try {\n      fs.writeFileSync(\n        PATHS.CONFIG_PATH,\n        `${JSON.stringify(mergedConfig, null, 2)}\\n`,\n        \"utf8\",\n      )\n    } catch (writeError) {\n      consola.warn(\n        \"Failed to write merged extraPrompts to config file\",\n        writeError,\n      )\n    }\n  }\n\n  cachedConfig = mergedConfig\n  return mergedConfig\n}\n\nexport function getConfig(): AppConfig {\n  cachedConfig ??= readConfigFromDisk()\n  return cachedConfig\n}\n\nexport function getExtraPromptForModel(model: string): string {\n  const config = getConfig()\n  return config.extraPrompts?.[model] ?? \"\"\n}\n\nexport function getSmallModel(): string {\n  const config = getConfig()\n  return config.smallModel ?? \"gpt-5-mini\"\n}\n\nexport function getResponsesApiContextManagementModels(): Array<string> {\n  const config = getConfig()\n  return (\n    config.responsesApiContextManagementModels\n    ?? defaultConfig.responsesApiContextManagementModels\n    ?? []\n  )\n}\n\nexport function isResponsesApiContextManagementModel(model: string): boolean {\n  return getResponsesApiContextManagementModels().includes(model)\n}\n\nexport function getReasoningEffortForModel(\n  model: string,\n): \"none\" | \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" {\n  const config = getConfig()\n  return config.modelReasoningEfforts?.[model] ?? \"high\"\n}\n\nexport function normalizeProviderBaseUrl(url: string): string {\n  return url.trim().replace(/\\/+$/u, \"\")\n}\n\nexport function getProviderConfig(name: string): ResolvedProviderConfig | null {\n  const providerName = name.trim()\n  if (!providerName) {\n    return null\n  }\n\n  const config = getConfig()\n  const provider = config.providers?.[providerName]\n  if (!provider) {\n    return null\n  }\n\n  if (provider.enabled === false) {\n    return null\n  }\n\n  const type = provider.type ?? \"anthropic\"\n  if (type !== \"anthropic\") {\n    consola.warn(\n      `Provider ${providerName} is ignored because only anthropic type is supported`,\n    )\n    return null\n  }\n\n  const baseUrl = normalizeProviderBaseUrl(provider.baseUrl ?? \"\")\n  const apiKey = (provider.apiKey ?? \"\").trim()\n  if (!baseUrl || !apiKey) {\n    consola.warn(\n      `Provider ${providerName} is enabled but missing baseUrl or apiKey`,\n    )\n    return null\n  }\n\n  return {\n    name: providerName,\n    type,\n    baseUrl,\n    apiKey,\n    models: provider.models,\n    adjustInputTokens: provider.adjustInputTokens,\n  }\n}\n\nexport function listEnabledProviders(): Array<string> {\n  const config = getConfig()\n  const providerNames = Object.keys(config.providers ?? {})\n  return providerNames.filter((name) => getProviderConfig(name) !== null)\n}\n\nexport function isMessagesApiEnabled(): boolean {\n  const config = getConfig()\n  return config.useMessagesApi ?? true\n}\n\nexport function getAnthropicApiKey(): string | undefined {\n  const config = getConfig()\n  return config.anthropicApiKey ?? process.env.ANTHROPIC_API_KEY ?? undefined\n}\n"],"mappings":";;;;;AA8CA,MAAM,wBAAwB;;;;;;AAO9B,MAAM,uBAAuB;;;;;;;;;;;;;;;;;;;AAoB7B,MAAMA,gBAA2B;CAC/B,MAAM,EACJ,SAAS,EAAE,EACZ;CACD,WAAW,EAAE;CACb,cAAc;EACZ,cAAc;EACd,iBAAiB;EACjB,gBAAgB;EAChB,WAAW;EACZ;CACD,YAAY;CACZ,qCAAqC,EAAE;CACvC,uBAAuB;EACrB,cAAc;EACd,iBAAiB;EACjB,gBAAgB;EAChB,WAAW;EACZ;CACD,uBAAuB;CACvB,gBAAgB;CACjB;AAED,IAAIC,eAAiC;AAErC,SAAS,mBAAyB;AAChC,KAAI;AACF,KAAG,WAAW,MAAM,aAAa,GAAG,UAAU,OAAO,GAAG,UAAU,KAAK;SACjE;AACN,KAAG,UAAU,MAAM,SAAS,EAAE,WAAW,MAAM,CAAC;AAChD,KAAG,cACD,MAAM,aACN,GAAG,KAAK,UAAU,eAAe,MAAM,EAAE,CAAC,KAC1C,OACD;AACD,MAAI;AACF,MAAG,UAAU,MAAM,aAAa,IAAM;UAChC;AACN;;;;AAKN,SAAS,qBAAgC;AACvC,mBAAkB;AAClB,KAAI;EACF,MAAM,MAAM,GAAG,aAAa,MAAM,aAAa,OAAO;AACtD,MAAI,CAAC,IAAI,MAAM,EAAE;AACf,MAAG,cACD,MAAM,aACN,GAAG,KAAK,UAAU,eAAe,MAAM,EAAE,CAAC,KAC1C,OACD;AACD,UAAO;;AAET,SAAO,KAAK,MAAM,IAAI;UACf,OAAO;AACd,UAAQ,MAAM,oDAAoD,MAAM;AACxE,SAAO;;;AAIX,SAAS,mBAAmB,QAG1B;CACA,MAAM,eAAe,OAAO,gBAAgB,EAAE;CAC9C,MAAM,sBAAsB,cAAc,gBAAgB,EAAE;CAC5D,MAAM,wBAAwB,OAAO,yBAAyB,EAAE;CAChE,MAAM,+BAA+B,cAAc,yBAAyB,EAAE;CAE9E,MAAM,2BAA2B,OAAO,KAAK,oBAAoB,CAAC,QAC/D,UAAU,CAAC,OAAO,OAAO,cAAc,MAAM,CAC/C;CAED,MAAM,+BAA+B,OAAO,KAC1C,6BACD,CAAC,QAAQ,UAAU,CAAC,OAAO,OAAO,uBAAuB,MAAM,CAAC;CAEjE,MAAM,wBAAwB,yBAAyB,SAAS;CAChE,MAAM,4BAA4B,6BAA6B,SAAS;AAExE,KAAI,CAAC,yBAAyB,CAAC,0BAC7B,QAAO;EAAE,cAAc;EAAQ,SAAS;EAAO;AAGjD,QAAO;EACL,cAAc;GACZ,GAAG;GACH,cAAc;IACZ,GAAG;IACH,GAAG;IACJ;GACD,uBAAuB;IACrB,GAAG;IACH,GAAG;IACJ;GACF;EACD,SAAS;EACV;;AAGH,SAAgB,0BAAqC;CACnD,MAAM,SAAS,oBAAoB;CACnC,MAAM,EAAE,cAAc,YAAY,mBAAmB,OAAO;AAE5D,KAAI,QACF,KAAI;AACF,KAAG,cACD,MAAM,aACN,GAAG,KAAK,UAAU,cAAc,MAAM,EAAE,CAAC,KACzC,OACD;UACM,YAAY;AACnB,UAAQ,KACN,sDACA,WACD;;AAIL,gBAAe;AACf,QAAO;;AAGT,SAAgB,YAAuB;AACrC,kBAAiB,oBAAoB;AACrC,QAAO;;AAGT,SAAgB,uBAAuB,OAAuB;AAE5D,QADe,WAAW,CACZ,eAAe,UAAU;;AAGzC,SAAgB,gBAAwB;AAEtC,QADe,WAAW,CACZ,cAAc;;AAG9B,SAAgB,yCAAwD;AAEtE,QADe,WAAW,CAEjB,uCACJ,cAAc,uCACd,EAAE;;AAIT,SAAgB,qCAAqC,OAAwB;AAC3E,QAAO,wCAAwC,CAAC,SAAS,MAAM;;AAGjE,SAAgB,2BACd,OAC0D;AAE1D,QADe,WAAW,CACZ,wBAAwB,UAAU;;AAGlD,SAAgB,yBAAyB,KAAqB;AAC5D,QAAO,IAAI,MAAM,CAAC,QAAQ,SAAS,GAAG;;AAGxC,SAAgB,kBAAkB,MAA6C;CAC7E,MAAM,eAAe,KAAK,MAAM;AAChC,KAAI,CAAC,aACH,QAAO;CAIT,MAAM,WADS,WAAW,CACF,YAAY;AACpC,KAAI,CAAC,SACH,QAAO;AAGT,KAAI,SAAS,YAAY,MACvB,QAAO;CAGT,MAAM,OAAO,SAAS,QAAQ;AAC9B,KAAI,SAAS,aAAa;AACxB,UAAQ,KACN,YAAY,aAAa,sDAC1B;AACD,SAAO;;CAGT,MAAM,UAAU,yBAAyB,SAAS,WAAW,GAAG;CAChE,MAAM,UAAU,SAAS,UAAU,IAAI,MAAM;AAC7C,KAAI,CAAC,WAAW,CAAC,QAAQ;AACvB,UAAQ,KACN,YAAY,aAAa,2CAC1B;AACD,SAAO;;AAGT,QAAO;EACL,MAAM;EACN;EACA;EACA;EACA,QAAQ,SAAS;EACjB,mBAAmB,SAAS;EAC7B;;AASH,SAAgB,uBAAgC;AAE9C,QADe,WAAW,CACZ,kBAAkB;;AAGlC,SAAgB,qBAAyC;AAEvD,QADe,WAAW,CACZ,mBAAmB,QAAQ,IAAI,qBAAqB"}

package/dist/main.js

@@ -23,7 +23,7 @@ if (typeof args["enterprise-url"] === "string") process.env.COPILOT_API_ENTERPRI
 const { auth } = await import("./auth-BCT7hxQA.js");
 const { checkUsage } = await import("./check-usage-YCS0L_nI.js");
 const { debug } = await import("./debug-Dx1S6uWG.js");
-const { start } = await import("./start-BECWyGAL.js");
+const { start } = await import("./start-jLhEk66O.js");
 const main = defineCommand({
 	meta: {
 		name: "copilot-api",

package/dist/{server-CSVu0O6G.js → server-DPTfZSVR.js}

@@ -1,7 +1,7 @@
 import { PATHS } from "./paths-Cla6y5eD.js";
 import { HTTPError, cacheModels, copilotBaseUrl, copilotHeaders, forwardError, generateRequestIdFromPayload, getRootSessionId, getUUID, isNullish, parseUserIdMetadata, prepareForCompact, prepareInteractionHeaders, sleep, state } from "./utils-BzmtBkw4.js";
 import { getCopilotUsage } from "./get-copilot-usage-CnL_6H-N.js";
-import { getConfig, getExtraPromptForModel, getProviderConfig, getReasoningEffortForModel, getSmallModel, isMessagesApiEnabled, isResponsesApiContextManagementModel } from "./config-DIqcOsnZ.js";
+import { getAnthropicApiKey, getConfig, getExtraPromptForModel, getProviderConfig, getReasoningEffortForModel, getSmallModel, isMessagesApiEnabled, isResponsesApiContextManagementModel } from "./config-bVq-BhC7.js";
 import consola from "consola";
 import path from "node:path";
 import { Hono } from "hono";
@@ -851,12 +851,48 @@ function getAnthropicToolUseBlocks(toolCalls) {
 //#endregion
 //#region src/routes/messages/count-tokens-handler.ts
 /**
-* Handles token counting for Anthropic messages
+* Forwards token counting to Anthropic's real /v1/messages/count_tokens endpoint.
+* Returns the result on success, or null to fall through to estimation.
+*/
+async function countTokensViaAnthropic(c, payload) {
+	if (!payload.model.startsWith("claude")) return null;
+	const apiKey = getAnthropicApiKey();
+	if (!apiKey) return null;
+	const model = payload.model.replaceAll(".", "-");
+	const res = await fetch("https://api.anthropic.com/v1/messages/count_tokens", {
+		method: "POST",
+		headers: {
+			"content-type": "application/json",
+			"x-api-key": apiKey,
+			"anthropic-version": "2023-06-01",
+			"anthropic-beta": "token-counting-2024-11-01"
+		},
+		body: JSON.stringify({
+			...payload,
+			model
+		})
+	});
+	if (!res.ok) {
+		consola.warn("Anthropic count_tokens failed:", res.status, await res.text().catch(() => ""), "- falling back to estimation");
+		return null;
+	}
+	const result = await res.json();
+	consola.info("Token count (Anthropic API):", result.input_tokens);
+	return c.json(result);
+}
+/**
+* Handles token counting for Anthropic messages.
+*
+* When an Anthropic API key is available (via config or ANTHROPIC_API_KEY env var)
+* and the model is a Claude model, forwards to Anthropic's free /v1/messages/count_tokens
+* endpoint for accurate counts. Otherwise falls back to GPT tokenizer estimation.
 */
 async function handleCountTokens(c) {
 	try {
-		const anthropicBeta = c.req.header("anthropic-beta");
 		const anthropicPayload = await c.req.json();
+		const anthropicResult = await countTokensViaAnthropic(c, anthropicPayload);
+		if (anthropicResult) return anthropicResult;
+		const anthropicBeta = c.req.header("anthropic-beta");
 		const openAIPayload = translateToOpenAI(anthropicPayload);
 		const selectedModel = findEndpointModel(anthropicPayload.model);
 		anthropicPayload.model = selectedModel?.id ?? anthropicPayload.model;
@@ -2877,4 +2913,4 @@ server.route("/:provider/v1/models", providerModelRoutes);
 
 //#endregion
 export { server };
-//# sourceMappingURL=server-CSVu0O6G.js.map
+//# sourceMappingURL=server-DPTfZSVR.js.map