@wengine-ai/llms 2.3.23 → 2.3.232

package/README.md

@@ -93,6 +93,8 @@ npm install -g @wengine-ai/claude-code-router-next@latest && ccr restart
 
 | 版本 | 发布内容 |
 | --- | --- |
+| **v2.3.232** | <ul><li>**严格项目级路由成为权威边界**: 项目配置非空 `Router` 后，目标缺失、格式错误、被禁用、不健康或额度耗尽不再静默回落到全局，而是返回稳定错误码；项目内显式 fallback 仍可用，同时关闭全局 Router/custom router、全局 fallback、客户端与 subagent model override 等项目外逃逸路径。</li><li>**Provider 级代理策略**: 新增顶层 `PROXY_GLOBAL_ENABLED` 与 provider 级 `proxy_enabled`；全局关闭时仅显式标记的 provider 使用共享 `PROXY_URL`，其余直连，未配置或开启时保持全部走代理的旧行为。推理、fallback、健康探测、额度查询、wakeup、provider API tokenizer 等出站请求统一遵循该策略。</li><li>**代理安全校验与 UI 控制**: 保存配置前仅允许 `http://`、`https://`、空值或环境变量占位符，非法协议/畸形 URL 逐项报错且不覆盖原配置；设置页新增「全局生效」开关、provider 卡片新增独立代理开关，并提示仅使用可信代理。ProxyAgent 改为按 URL 缓存复用，关闭时释放，日志自动脱敏代理凭据。</li><li>**per-client adapters 与 pi 上下文路由**: runtime 下沉到 core 并用 adapter 统一 Claude Code、Codex、pi、qwen-code、opencode 的差异；pi 新增 `extendedContextRatio`（默认 80%），不再生成/消费 `[1m]` 后缀，longContext 改用绝对阈值链。</li><li>**修复严格路由、客户端识别与用量回归**: 畸形项目 target 不再逃逸全局；Codex Responses 请求不再被误判为 Claude Code；修复非流式 `upstreamModel` 丢失、失败请求串用上一请求 tokens、`response.completed` 将 usage merge base 归零、项目路由失败清除会话用量基线等问题。</li><li>**修复 runtime 生命周期与错误处理**: `createCcrServer({ port })` 正确采用传入端口，401/403 认证不再挂起，provider/transformer/tokenizer 初始化完成后才监听，preset 注册失败不再静默吞掉。</li></ul> |
+| **v2.3.231** | <ul><li>**项目级接管默认仅接管 Claude Code**: 此前打开项目「CCR 接管」默认接管全部客户端（Claude Code、pi、qwen-code、opencode），会把 `.pi/settings.json`、`.qwen/settings.json`、`opencode.json` 等配置写进项目目录（opencode 首次运行还会自行生成 `AGENTS.md`），污染项目根目录；现在所有默认路径（主开关、多选清空、legacy `enabled: true`）都只接管 Claude Code，其余客户端在多选中显式勾选才接管。</li><li>**修复「检查更新」永远提示已是最新（双重 bug）**: 后端 `checkForUpdates`/`performUpdate` 硬编码了已从 npm unpublish 的旧包名 `claude-code-router-next`，`npm view` 404 被 catch 静默吞掉、恒返回「无更新」，「立即更新」也因同样的错包名安装失败，现改为从 `package.json` 动态读取包名（`@wengine-ai/claude-code-router-next`），今后改包名不会再失效；另外前端更新弹窗条件要求 `changelog` 非空而后端该字段恒为空串，即使有新版本也不弹窗，现去掉该强制判断（弹窗内已有「暂无更新日志」兜底文案）。</li><li>**发布确认点闸门（release gate）**: `scripts/release.sh` 新增 `validate_release_docs`，发布前（含 dry-run）强制校验：6 个 `package.json` 版本一致、`CHANGELOG.md` 有该版本的非空段落、两份 README 表格有该版本行、版本号严格大于 npm 已发布 latest（逐段数字比较，registry 不可达时警告跳过）；任一不满足即中止发布。</li><li>**版本号策略：多位 patch 日常小迭代**: 自本版本起日常小迭代在 patch 段追加一位数字（`2.3.23` → `2.3.231` → `2.3.232`），避免每日发布推高主版本数字；patch 按数字比较，用过 `2.3.23x` 后下一个功能版本是 `2.3.240`（`2.3.24` 会被闸门当降级拦截）或直接升 `2.4.0`。</li></ul> |
 | **v2.3.23** | <ul><li>**修复状态栏 token 速率虚高（撞到 999 上限）**: `ccr statusline` 的 token 速率常显示几百、极端撞到 999 上限，与「用量统计」的几十 t/s 对不上。根因是 token-speed 插件流式过程中每秒上报的是「最近 1 秒到达的 token 数」滑动窗口值，而 SSE delta 常成批到达（同一时间戳）使瞬时计数虚高；现改为与用量统计同一套解码平均公式（`输出 token ÷ (总耗时 − TTFT)`）上报，速率随流式收敛到最终值、与用量统计一致，并移除失效的滑动窗口记账。</li><li>**状态栏默认改为无图标表格风格**: 默认主题不再带装饰图标，模块之间改用细竖线 `│` 分隔，呈简洁表格样式，默认模块顺序为「模型 │ 工作目录 │ git 分支 │ 上下文进度条 │ token 速率 │ 会话总 token」。动机是歧义宽度的 emoji 图标（如闪电 `⚡` U+26A1）会让 Claude Code 误算状态栏显示宽度、在双击等重绘时产生数字重影/位移；改用定宽字符或不带图标可避免。图标仍支持在 UI 中自定义。</li><li>**修复 UI 状态栏图标无法删除 + 旧圆圈模块迁移**: 切换所选模块时图标搜索框（`IconSearchInput`）不随 `value` 同步、与模块真实图标脱节导致看似删不掉，现已在 `value` 变化时同步；同时加载配置时把旧 `contextCircle` 模块迁移为 `contextBar`，配置弹窗显示长条进度条而非旧圆圈图标。</li><li>**`build:ui` 构建后同步产物到 CLI/根 dist**: `pnpm build:ui` 现会把 `index.html` 同步到 `packages/cli/dist` 与根 `dist`，使单独运行 `build:ui` 也能更新本地运行中的 ccr 实际读取的包。</li></ul> |
 | **v2.3.22** | <ul><li>**状态栏按用户配置的压缩阈值显示上下文上限**: `ccr statusline` 优先用 `CLAUDE_CODE_AUTO_COMPACT_WINDOW`（CCR 从顶层 `ContextWindow` 写入）作为上下文上限与百分比分母，未设才回退到 Claude Code 的窗口值，使状态栏与实际 auto-compact 触发点对齐，不再固定显示 1M/200k。</li><li>**接管保留用户手写的 auto-compact 自定义值**: CCR 接管 Claude Code 时不再无条件用 `ContextWindow` 覆盖 `CLAUDE_CODE_AUTO_COMPACT_WINDOW`，改用状态文件（`~/.claude-code-router/client-state.json` 全局、`<project-id>/ccr-state.json` 项目级）区分 CCR 写入值（可随 `ContextWindow` 更新）与用户手写值（保留）；卸载只清自己写的、保留用户的；状态缺失时保守保留，绝不误覆盖。</li><li>**UI 上下文窗口与扩展上下文配合提示**: 顶层 `ContextWindow` > 200000 但未启用对应模型家族「扩展上下文 (1M)」（模型名带 `[1m]`）时显示红色警告，并在两处配置补充配合说明——否则 Claude Code 会将其封顶至 200000。</li><li>**修复状态栏百分比按 1M 计算与偶发闪 0%**: statusline 子进程不一定继承项目级 settings 的环境变量，导致项目里设的 `CLAUDE_CODE_AUTO_COMPACT_WINDOW` 被漏读而按 1M 计算，现在依次从 `process.env`→项目 `settings.local.json`→全局 `settings.json` 读取；另 `current_usage` 在请求进行中/刚 auto-compact 后会暂态为空使百分比偶发闪 0%，现回退到 transcript 最近一条 assistant 消息的上下文用量保持稳定。</li><li>**修复项目接管 disable→enable 循环冻结窗口与 disable 残留**: 项目级 `ccr-state.json` 在 disable 后丢失、再 enable 时不重建，导致 CCR 无法识别自己写入的窗口——既会让 `ContextWindow` 变更无法刷新生效，也会在关闭接管时把 `CLAUDE_CODE_AUTO_COMPACT_WINDOW` 当用户值残留。现在状态缺失时按「值等于当前 `ContextWindow`」兜底识别 CCR-managed，enable 重建状态、disable 清残留，真正的用户手写值仍保留。</li></ul> |
 | **v2.3.21** | <ul><li>**新增 pi (earendil-works) 客户端接管**: 新增 pi 作为第三个接管目标（与 Claude Code、Codex 并列），pi 使用 Anthropic `/v1/messages` 协议直连无需 transformer；接管会在 `~/.pi/agent` 写入 `models.json`（注册指向 ccr 代理的 `ccr` provider，暴露 `ccr-opus`/`ccr-sonnet`/`ccr-haiku` 族别名）与 `settings.json`（默认 provider/model 指向 ccr），并备份原文件，仅做配置注入不含账号管理；用量统计也会把 pi 识别为独立客户端（pi 与 Claude Code 共用 `ccr-*` 别名，靠 pi system prompt 特征 + Anthropic SDK 请求头区分）。</li><li>**项目级接管支持多客户端多选**: 「项目级配置」页「CCR 接管」从只接管 Claude Code 改为客户端多选下拉（Claude Code + pi + qwen-code + opencode），不选默认接管全部；pi 项目级接管在项目目录写 `.pi/settings.json` 指向全局 ccr provider 并在 `trust.json` 信任该目录，接管状态从各客户端项目级配置文件实时推导、自动兼容既有项目（Codex 配置全局-only 不纳入）。</li><li>**新增 qwen-code (Alibaba) 客户端接管**: 新增 qwen-code（`@qwen-code/qwen-code`）作为第四个接管目标，走 Anthropic `/v1/messages` 直连；接管在 `~/.qwen/settings.json`（用户级）/ `<项目>/.qwen/settings.json`（项目级）写入指向 ccr 的 `modelProviders.anthropic`（暴露 ccr-opus/sonnet/haiku）、备份原文件，项目级还在 `trustedFolders.json` 信任该目录。用量统计区分 qwen-code：qwen 经代理时会伪装成 Claude Code 的 `claude-cli` UA，故改用 qwen system prompt 特征（`You are Qwen Code`）正向识别、Claude Code 用 `cc_version`/`metadata.user_id` 强信号识别。</li><li>**新增 opencode (opencode.ai) 客户端接管**: 新增 opencode 作为第五个接管目标，走 Anthropic `/v1/messages` 直连；接管在 `~/.config/opencode/opencode.json`（全局）/ `<项目>/opencode.json`（项目级）注入一个 `@ai-sdk/anthropic` 的 `provider.ccr`（baseURL=`…/v1`、apiKey 内联、暴露 ccr-opus/sonnet/haiku）并设 `model: ccr/ccr-opus`，备份原文件，无需 trust。用量统计区分 opencode：其请求带 `opencode/…` UA 且不伪装，直接按 UA 识别。</li></ul> |
@@ -101,8 +103,6 @@ npm install -g @wengine-ai/claude-code-router-next@latest && ccr restart
 | **v2.3.18** | <ul><li>**模型族长上下文阈值继承主路由配置**: `ccr-opus`/`ccr-sonnet`/`ccr-haiku` 的 family routing 未配置族级阈值时会继承 `Router.longContextThreshold`，避免主路由设为 100k 时约 70k token 请求仍误进 `<family>/longContext`。</li><li>**fallback 候选模型也执行 Double-Check 重试**: 每个 fallback 候选第一次 `fetch failed`/空 SSE/隐藏错误后会先对同一候选重试一次，第二次仍失败才切到下一个，降低智谱等接口短暂不稳定导致的过早切换；同时修正失败记录的 `originalModel` 显示。</li><li>**全局配置保存后同步项目级 CCR 接管字段**: 跟随全局 Router 且已开启 CCR Takeover 的项目，在保存全局配置或切回使用全局配置时自动刷新 `.claude/settings.local.json` 中的模型族别名、auto-compact、状态栏与代理字段。</li><li>**默认降低日志量并保留最近 7 天**: 服务器日志默认级别从 `debug` 调整为 `error`，只在显式配置 `LOG_LEVEL=info/debug/trace` 时输出详细日志，并自动清理超过 7 天的 `ccr-*.log`。</li></ul> |
 | **v2.3.17** | <ul><li>**修复 fallback 触发时所有备用模型报 Invalid URL**: 主模型限流（如智谱套餐）触发 fallback 后，此前所有备用模型都报 `Invalid URL`；根因是 fallback 用了原始 `ConfigProvider[]`（无 `baseUrl`）而非已注册的 `LLMProvider[]`，改为 `providerService.getProviders()`。</li><li>**修复 fireworks 托管上游用量统计全 0**: fireworks 流式把真实 usage 放在 `finish_reason` 之后的 `choices:[]` 空 chunk 里，此前 `finish_reason` 块用 null usage 覆盖了已 merge 的真实值、且 `break` 丢弃了后续 usage chunk，导致 `input` 有值但 `output`/缓存命中为 0；现改为 finish 后继续读取、只设 stop_reason 不碰 usage、`break` 改 `hasFinished=true`，并修复下游三层用量捕获与 `??` 对 0 不 fallback 的问题。</li><li>**忽略已删除模型的残留路由，防止健康池污染**: 从 provider 配置删除模型后，路由与 fallback 中残留的 `provider,model` 仍被当作有效候选反复失败污染健康池；现在对无法匹配的候选直接返回 `null` 并在请求前校验，`ProviderHealthStore` 统一拦截空 provider/model 防畸形 key 污染，同时抽取 `findProviderModel` 共用函数。</li></ul> |
 | **v2.3.16** | <ul><li>**修复 system 消息顺序兼容 DeepSeek/vLLM**: OpenAI 兼容提供商（DeepSeek V4、GLM、vLLM）要求 `[system, user, assistant]` 顺序，此前 system 排在 user/assistant 之后会导致输出乱码；现在将 system 消息前置并去重。</li><li>**接管时去除 Attribution 动态头以提升缓存命中**: CCR 接管 Claude Code 时默认写入 `CLAUDE_CODE_ATTRIBUTION_HEADER=0`，去掉系统提示词开头随请求变化的 attribution 头（客户端版本 + prompt fingerprint），稳定上游 prompt 缓存前缀；可在设置页或 `disableAttributionHeader: false` 关闭。</li><li>**设置页布局紧凑化**: 「日志级别」下拉从独占整行移入两列网格，与「API 密钥」并排显示，减少纵向留白。</li></ul> |
-| **v2.3.15** | <ul><li>**Fallback 前同模型重试一次**: 模型调用出现一次异常不再立即切换到备用模型，而是先对同一模型自动重试一次（Double-Check），避免因瞬时错误（网络抖动、偶发限流、空 SSE 响应等）导致不必要的模型切换；重试仍失败才走原有 fallback 流程。</li><li>**用量统计显示上游真实模型**: 部分上游网关会偷偷将请求路由/降级到另一个后端模型（如请求 glm-5 实际返回 MiniMax-M2.5）。用量统计的模型映射现在追加上游返回的真实模型，格式为 `originalModel → routedModel → upstreamModel`（如 `ccr-opus → glm-5 → minimax-m2.5`），上游未偷换时自动省略。</li></ul> |
-| **v2.3.14** | <ul><li>**修复状态栏显示 `<synthetic>` 模型名**：Claude Code 在 auto-compact 自动压缩、中断恢复时写入的 `<synthetic>` 合成消息（非真实 LLM 响应）此前被状态栏当作模型名显示（从 Claude 账号会话切换到 ccr 接管、或发生自动压缩后尤其常见）；现在过滤该合成标识，正确显示实际路由模型，同时不再把合成消息的 usage 计入 token 统计。</li></ul> |
 
 > 仅保留最近 10 个版本，更早版本的发布摘要见 [CHANGELOG-archive.md](./CHANGELOG-archive.md)，完整详细变更记录见 [CHANGELOG.md](./CHANGELOG.md)。
 
@@ -117,7 +117,10 @@ npm install -g @wengine-ai/claude-code-router-next@latest && ccr restart
 > ```
 
 `config.json` 文件有几个关键部分：
-- **`PROXY_URL`** (可选): 您可以为 API 请求设置代理，例如：`"PROXY_URL": "http://127.0.0.1:7890"`。
+- **`PROXY_URL`** (可选): 您可以为 API 请求设置代理，例如：`"PROXY_URL": "http://127.0.0.1:7890"`。CCR 进程自身通过此地址连接代理端口，无需开启系统代理、TUN 或代理软件的全局模式。
+- **`PROXY_GLOBAL_ENABLED`** (可选): 控制代理的作用范围。未配置或设为 `true`（默认）时，所有 provider 的出站请求均走代理，保持与旧配置的兼容性。设为 `false` 时，仅标记了 `proxy_enabled: true` 的 provider 走代理，其余 provider 直连。所有 provider 共用顶层 `PROXY_URL`，不支持为每个 provider 单独设置代理地址。如果 `PROXY_URL` 未设置（空地址），则所有代理开关均不生效，全部直连。provider 专属的出站请求（推理、fallback、健康探测、额度查询、wakeup 唤醒、provider API tokenizer 等）均遵循同一 provider 代理策略。
+  > [!WARNING]
+  > 代理可看到 API key 和请求内容，请仅配置可信代理。
 - **`LOG`** (可选): 您可以通过将其设置为 `true` 来启用日志记录。当设置为 `false` 时，将不会创建日志文件。默认值为 `true`。
 - **`LOG_LEVEL`** (可选): 设置日志级别。可用选项包括：`"fatal"`、`"error"`、`"warn"`、`"info"`、`"debug"`、`"trace"`。默认值为 `"error"`；仅在显式配置为 `"info"`、`"debug"` 或 `"trace"` 时输出详细日志。
 - **日志系统**: Claude Code Router 使用两个独立的日志系统：
@@ -136,6 +139,7 @@ npm install -g @wengine-ai/claude-code-router-next@latest && ccr restart
 {
   "APIKEY": "your-secret-key",
   "PROXY_URL": "http://127.0.0.1:7890",
+  "PROXY_GLOBAL_ENABLED": false,
   "LOG": true,
   "LOG_LEVEL": "error",
   "API_TIMEOUT_MS": 600000,
@@ -153,7 +157,8 @@ npm install -g @wengine-ai/claude-code-router-next@latest && ccr restart
       ],
       "transformer": {
         "use": ["openrouter"]
-      }
+      },
+      "proxy_enabled": true
     },
     {
       "name": "deepseek",

package/dist/api/middleware.d.ts

@@ -0,0 +1,9 @@
+import { FastifyRequest, FastifyReply } from "fastify";
+export interface ApiError extends Error {
+    statusCode?: number;
+    code?: string;
+    type?: string;
+    rawBody?: string;
+}
+export declare function createApiError(message: string, statusCode?: number, code?: string, type?: string): ApiError;
+export declare function errorHandler(error: ApiError, request: FastifyRequest, reply: FastifyReply): Promise<never>;

package/dist/api/routes.d.ts

@@ -0,0 +1,24 @@
+import { FastifyInstance } from "fastify";
+import { ConfigService } from "../services/config";
+import { ProviderService } from "../services/provider";
+import { TransformerService } from "../services/transformer";
+export declare const CCR_FAMILY_ALIAS: RegExp;
+/**
+ * Normalize a Responses API (Codex) request body into a unified chat request so
+ * downstream transformers (and the family router) can process it.
+ * Codex sends: { model, input: [...], instructions, tools, stream, ... }
+ * We convert to: { model, messages: [...], system, tools, stream, ... }
+ */
+export declare function normalizeResponsesBody(body: any): void;
+declare module "fastify" {
+    interface FastifyInstance {
+        configService: ConfigService;
+        providerService: ProviderService;
+        transformerService: TransformerService;
+        recordUsage?: (data: any) => void;
+    }
+    interface FastifyRequest {
+        provider?: string;
+    }
+}
+export declare const registerApiRoutes: (fastify: FastifyInstance) => Promise<void>;

package/dist/ccr/admin-routes.d.ts

@@ -0,0 +1,6 @@
+/**
+ * Register all CCR admin/management routes on the given Fastify app.
+ * This is the core of createServer() from the legacy server, minus the
+ * request-pipeline hooks (which are now in createCcrServer).
+ */
+export declare function registerAdminRoutes(server: any, config: any): Promise<any>;

package/dist/ccr/agents/image.agent.d.ts

@@ -0,0 +1,10 @@
+import { IAgent, ITool } from "./type";
+export declare class ImageAgent implements IAgent {
+    name: string;
+    tools: Map<string, ITool>;
+    constructor();
+    shouldHandle(req: any, config: any): boolean;
+    appendTools(): void;
+    reqHandler(req: any, config: any): void;
+}
+export declare const imageAgent: ImageAgent;

package/dist/ccr/agents/index.d.ts

@@ -0,0 +1,28 @@
+import { IAgent } from './type';
+export declare class AgentsManager {
+    private agents;
+    /**
+     * Register an agent
+     * @param agent The agent instance to register
+     * @param isDefault Whether to set as default agent
+     */
+    registerAgent(agent: IAgent): void;
+    /**
+     * Find agent by name
+     * @param name Agent name
+     * @returns Found agent instance, undefined if not found
+     */
+    getAgent(name: string): IAgent | undefined;
+    /**
+     * Get all registered agents
+     * @returns Array of all agent instances
+     */
+    getAllAgents(): IAgent[];
+    /**
+     * Get all agent tools
+     * @returns Array of tools
+     */
+    getAllTools(): any[];
+}
+declare const agentsManager: AgentsManager;
+export default agentsManager;

package/dist/ccr/agents/type.d.ts

@@ -0,0 +1,13 @@
+export interface ITool {
+    name: string;
+    description: string;
+    input_schema: any;
+    handler: (args: any, context: any) => Promise<string>;
+}
+export interface IAgent {
+    name: string;
+    tools: Map<string, ITool>;
+    shouldHandle: (req: any, config: any) => boolean;
+    reqHandler: (req: any, config: any) => void;
+    resHandler?: (payload: any, config: any) => void;
+}

package/dist/ccr/auth.d.ts

@@ -0,0 +1,8 @@
+/**
+ * API key authentication middleware for the CCR runtime.
+ *
+ * Migrated verbatim from packages/server/src/middleware/auth.ts.
+ * Local requests (127.0.0.1) are trusted and skip auth.
+ */
+import type { FastifyRequest, FastifyReply } from "fastify";
+export declare const apiKeyAuth: (config: any) => (req: FastifyRequest, reply: FastifyReply, done: () => void) => Promise<void>;

package/dist/ccr/claude-config-init.d.ts

@@ -0,0 +1 @@
+export declare function initializeClaudeConfig(): Promise<void>;

package/dist/ccr/codex-accounts.d.ts

@@ -0,0 +1,9 @@
+export declare function switchCodexAccountBeforeUsageLimit(): Promise<void>;
+export declare function switchCodexAccountAfterRateLimit(reason?: string): Promise<void>;
+/** Periodically refresh managed Codex account tokens, active or not, so refresh_tokens never go stale. */
+export declare function startCodexTokenRefreshScheduler(): void;
+export declare function isRateLimitMessage(statusCode: number, message?: string): boolean;
+export declare function getCurrentCodexAccountForRequest(): Promise<{
+    id?: string;
+    email?: string;
+}>;

package/dist/ccr/codex-usage-cache.d.ts

@@ -0,0 +1,2 @@
+import { type CodexAccountsResult } from "@wengine-ai/claude-code-router-shared";
+export declare function computeCodexAccountUsage(config: Record<string, any>): Promise<CodexAccountsResult>;

package/dist/ccr/config.d.ts

@@ -0,0 +1,6 @@
+export declare const initDir: () => Promise<void>;
+export declare const readConfigFileRaw: () => Promise<any>;
+export declare const readConfigFile: () => Promise<any>;
+export declare const backupConfigFile: () => Promise<string | null>;
+export declare const writeConfigFile: (config: any) => Promise<void>;
+export declare const initConfig: () => Promise<any>;

package/dist/ccr/create-server.d.ts

@@ -0,0 +1,6 @@
+import Server from "../server";
+export interface CcrRunOptions {
+    port?: number;
+    logger?: any;
+}
+export declare function createCcrServer(options?: CcrRunOptions): Promise<Server>;

package/dist/ccr/health-reconcile.d.ts

@@ -0,0 +1,22 @@
+/**
+ * Build the set of provider+model pairs the current config can actually route
+ * to: every model listed under a provider, plus every "provider,model" target
+ * referenced anywhere in Router / fallback. Some providers leave `models` empty
+ * and are only reachable via Router, so both sources must be considered to avoid
+ * pruning a still-routable model's health state.
+ */
+export declare function collectReachableModelKeys(config: any): Set<string>;
+/**
+ * Remove circuit-breaker health entries the current config can no longer route
+ * to (renamed/removed models, deleted providers). Returns the number pruned.
+ */
+export declare function reconcileHealthStore(config: any, logger?: any): number;
+/**
+ * Clear every circuit-breaker entry for a provider whose endpoint just probed
+ * successfully. The probe checks endpoint-level reachability (GET /v1/models),
+ * so a success means any lingering "open" breaker for that provider — including
+ * orphaned/renamed model names the per-model recover can't reach — should be
+ * cleared. Genuinely broken models will reopen on the next real request.
+ * Returns the number of entries cleared.
+ */
+export declare function clearProviderHealth(providerName: string, logger?: any): number;

package/dist/ccr/models.d.ts

@@ -0,0 +1,26 @@
+/** Anthropic-compatible model metadata exposed by /v1/models. */
+export interface AnthropicModelInfo {
+    type: "model";
+    id: string;
+    display_name: string;
+    created_at: string;
+    context_window: number;
+    contextWindow: number;
+    max_input_tokens: number;
+    maxInputTokens: number;
+    max_tokens: number;
+    maxTokens: number;
+    max_output_tokens: number;
+    maxOutputTokens: number;
+    capabilities: {
+        context_window: number;
+        contextWindow: number;
+        max_input_tokens: number;
+        maxInputTokens: number;
+        max_tokens: number;
+        maxTokens: number;
+        max_output_tokens: number;
+        maxOutputTokens: number;
+    };
+}
+export declare function listAnthropicCompatibleModels(config: any): AnthropicModelInfo[];

package/dist/ccr/pino-retention.d.ts

@@ -0,0 +1 @@
+export declare function startPinoLogRetention(): void;

package/dist/ccr/plugin-registration.d.ts

@@ -0,0 +1,6 @@
+/**
+ * Register plugins from configuration
+ * @param serverInstance Server instance
+ * @param config Application configuration
+ */
+export declare function registerPluginsFromConfig(serverInstance: any, config: any): Promise<void>;

package/dist/ccr/request-pipeline.d.ts

@@ -0,0 +1,37 @@
+/**
+ * CCR request pipeline — explicit, ordered hooks for the CCR runtime.
+ *
+ * Hook order (enforced by namespace phases + sequential global registration):
+ *   1. onRequest: request timing (requestStartTime)
+ *   2. namespace preHandler: request normalize
+ *   3. namespace preHandler: adapter
+ *   4. global preHandler: auth/Codex
+ *   5. global preHandler: agent mutation
+ *   6. namespace preHandler: router
+ *   7. namespace preHandler: provider model normalization
+ *   8. handler: routes.ts handleTransformerEndpoint
+ *   9. onSend: agent tool rewrite + usage/upstream-model capture
+ *  10. onResponse: TTFT/speed/health/final usage record
+ *
+ * Fastify runs ancestor preHandler hooks before child plugin hooks. To enforce
+ * adapter → auth/Codex → agent → router without relying on registration timing,
+ * the namespace registers a single ordered dispatcher (see Server.registerNamespace)
+ * whose phases invoke the global callbacks between adapter and router.
+ *
+ * The pipeline is registered by createCcrServer AFTER admin routes and BEFORE
+ * namespace registration + listen, so all hooks are in deterministic order.
+ */
+import { EventEmitter } from "node:events";
+export declare function getPipelineEventEmitter(): EventEmitter;
+/**
+ * Register all CCR request-pipeline hooks on the server instance.
+ * Called by createCcrServer after admin routes are registered.
+ */
+export interface CcrPreHandlerCallbacks {
+    authCodex(req: any, reply: any): Promise<void>;
+    agent(req: any, reply: any): Promise<void>;
+}
+declare function normalizeCcrRequest(req: any): void;
+export declare function createCcrPreHandlerCallbacks(config: any): CcrPreHandlerCallbacks;
+export { normalizeCcrRequest };
+export declare function registerRequestPipeline(serverInstance: any, config: any): void;

package/dist/ccr/usage-merge.d.ts

@@ -0,0 +1,32 @@
+/**
+ * Pure helpers for merging per-frame LLM usage into the per-request usage cache.
+ *
+ * Extracted from index.ts so the merge contract is unit-testable without
+ * standing up the whole Fastify server.
+ */
+export declare function getUsageCacheReadInputTokens(usage: any): number;
+export declare function getUsageCacheCreationInputTokens(usage: any): number;
+/**
+ * Normalize a raw usage object (from Anthropic, OpenAI Chat Completions, or
+ * OpenAI Responses APIs) into a common shape with input_tokens / output_tokens
+ * / cache_read_input_tokens / cache_creation_input_tokens fields.
+ */
+export declare function normalizeUsagePayload(usage: any): any;
+/**
+ * Merge an incoming usage frame into the previously captured usage for a
+ * request. Fields are merged so that an all-zero frame canNOT overwrite a real
+ * value captured earlier: each field is replaced only when the incoming value
+ * is non-zero.
+ *
+ * Why this matters: some OpenAI-compatible upstreams (e.g. fireworks-hosted
+ * GLM-5.2) report real usage on the first chunk (message_start /
+ * response.created) but emit an all-zero usage object on the trailing
+ * message_delta / response.completed. A plain `{...base, ...incoming}` spread
+ * let those zeros clobber the real input_tokens/output_tokens, leaving the
+ * whole request recorded as 0/0 in usage stats.
+ *
+ * `resetBase` should be true on the first frame of a new response
+ * (message_start / response.completed) so stale fields from a previous request
+ * in the same session don't leak in.
+ */
+export declare function mergeUsageCapture(existing: any, incoming: any, resetBase: boolean): any;

package/dist/ccr/usage-store.d.ts

@@ -0,0 +1,109 @@
+type UsageStatus = "success" | "error";
+export interface UsageRecord {
+    id: string;
+    timestamp: string;
+    sessionId: string;
+    provider: string;
+    originalModel: string;
+    model: string;
+    upstreamModel?: string;
+    modelFamily: string;
+    scenarioType: string;
+    clientType?: string;
+    codexAccountId?: string;
+    codexAccountEmail?: string;
+    stream: boolean;
+    inputTokens: number;
+    outputTokens: number;
+    cacheReadInputTokens: number;
+    cacheCreationInputTokens: number;
+    ttft: number | null;
+    tokensPerSecond: number | null;
+    durationMs: number;
+    status: UsageStatus;
+    errorMessage?: string;
+    responseBody?: string;
+}
+export interface UsageSummary {
+    totalRequests: number;
+    successCount: number;
+    errorCount: number;
+    totalInputTokens: number;
+    totalOutputTokens: number;
+    totalCacheReadInputTokens: number;
+    totalCacheCreationInputTokens: number;
+    avgTtft: number | null;
+    avgTokensPerSecond: number | null;
+    byModel: Record<string, {
+        count: number;
+        inputTokens: number;
+        outputTokens: number;
+        cacheReadInputTokens: number;
+        cacheCreationInputTokens: number;
+    }>;
+    byProvider: Record<string, {
+        count: number;
+        inputTokens: number;
+        outputTokens: number;
+        cacheReadInputTokens: number;
+        cacheCreationInputTokens: number;
+    }>;
+    byScenario: Record<string, {
+        count: number;
+        inputTokens: number;
+        outputTokens: number;
+        cacheReadInputTokens: number;
+        cacheCreationInputTokens: number;
+    }>;
+    byFamily: Record<string, {
+        count: number;
+        inputTokens: number;
+        outputTokens: number;
+        cacheReadInputTokens: number;
+        cacheCreationInputTokens: number;
+    }>;
+    byDay: Record<string, {
+        count: number;
+        inputTokens: number;
+        outputTokens: number;
+        cacheReadInputTokens: number;
+        cacheCreationInputTokens: number;
+    }>;
+    byClient: Record<string, {
+        count: number;
+        inputTokens: number;
+        outputTokens: number;
+        cacheReadInputTokens: number;
+        cacheCreationInputTokens: number;
+    }>;
+}
+export interface UsageQueryFilters {
+    startTime?: string;
+    endTime?: string;
+    model?: string;
+    provider?: string;
+    scenario?: string;
+    clientType?: string;
+    sessionId?: string;
+    status?: UsageStatus;
+    page?: number;
+    pageSize?: number;
+}
+export interface UsageQueryResult {
+    records: UsageRecord[];
+    summary: UsageSummary;
+    total: number;
+    page: number;
+    pageSize: number;
+}
+export declare function append(record: UsageRecord): void;
+export declare function query(filters: UsageQueryFilters): UsageQueryResult;
+export declare function querySummary(startTime?: string, endTime?: string, status?: UsageStatus): UsageSummary;
+export declare function clear(beforeDate?: string): void;
+export declare function close(): void;
+export declare function readTokenSpeedStats(sessionId: string): {
+    ttft: number | null;
+    tokensPerSecond: number | null;
+};
+export declare function getOldestRecordTimestamp(provider: string, startTime: string, endTime: string): string | undefined;
+export {};