@wengine-ai/llms 2.3.2417 → 2.3.2419

package/README.md

@@ -93,7 +93,8 @@ npm install -g @wengine-ai/claude-code-router-next@latest && ccr restart
 
 | 版本 | 发布内容 |
 | --- | --- |
-| **v2.3.2417** | <ul><li>**空回复现在会重试并降级**：上游返回 200 且 `content` 为空时不再被记成成功，改为触发同模型重试与 fallback；判据是 content 集合为空（tool_use-only 等合法回合不受影响），探测有字节与时间上限，慢速流不丢字节；Codex 侧同样拦住空 `response.completed`。</li><li>**流式回合不再记录 `output_tokens: 0`**：普通流式请求此前没有 usage 读取器，健康回复被记成全零且无 TTFT；现从 tee 分支采集并按流序合并。</li><li>**不健康 provider 从模型列表降级剔除**：熔断模型对所有 provider 类型生效隐藏，恢复后自动放回；订阅类 provider 新增两级后台探活（先查账号状态，再发真实最小请求），不再依赖真实请求反复撞死账号。</li></ul> |
+| **v2.3.2419** | <ul><li>**codex 0 输出回合根因修复**：数据面此前绕过 `SUBSCRIPTION_PROXY_URL` 走直连路径，中间盒剥离响应头并破坏压缩体（插件改为走订阅代理 + identity 编码，ccr-provider-codex 1.0.7）；同时单次读取的格式探测在代理分块下漏掉格式标记，Responses 流被误送入 Chat 转换器输出「完整但空」的消息（现在累积读取判定格式，且原生 Responses 流仅对 codex 客户端放行，其他客户端走 Responses→Chat 转换）。</li><li>**gemini-oauth 数据面加 identity 编码**，登录失败记录 Google 原始错误与失败步骤（1.0.7）。</li></ul> |
+| **v2.3.2418** | <ul><li>**空回复现在会重试并降级**：上游返回 200 且 `content` 为空时不再被记成成功，改为触发同模型重试与 fallback；判据是 content 集合为空（tool_use-only 等合法回合不受影响），探测有字节与时间上限，慢速流不丢字节；Codex 侧同样拦住空 `response.completed`。</li><li>**流式回合不再记录 `output_tokens: 0`**：普通流式请求此前没有 usage 读取器，健康回复被记成全零且无 TTFT；现从 tee 分支采集并按流序合并。</li><li>**不健康 provider 从模型列表降级剔除**：熔断模型对所有 provider 类型生效隐藏，恢复后自动放回；订阅类 provider 新增两级后台探活（先查账号状态，再发真实最小请求），不再依赖真实请求反复撞死账号。</li></ul> |
 | **v2.3.2415** | <ul><li>**新增 Qoder 订阅 provider（qoder.com + qoder.cn）**: 同一插件同引擎支持国际站与国内站 PAT 订阅，双 managed provider + 卡片双区块（中英双语）；含整配置写串行化、401 单次刷新重放、每帧 1MiB 上限等健壮性修复。</li><li>**修复 WorkBuddy 6004 重置时间戳时区**: 测试改用 Asia/Shanghai 渲染，UTC CI 不再挂。</li><li>**修复 tool_result 图片按 base64 文本发送**: 图片改为视觉块发送，避免 200 万+ token 膨胀与上游拒绝。</li></ul> |
 | **v2.3.2414** | <ul><li>**CCR Desktop（可选 Tauri 客户端）与 Pi 配置生命周期修复**: 新增可选 Tauri 桌面壳（托盘快速切换全局默认模型、quota 摘要、单实例；生命周期伴侣化、退出不杀 gateway）；修复 `ccr restart` 的 disable→enable 与 legacy 项目 provider 改名迁移丢失 Pi 自定义模型字段（thinkingLevelMap/modelOverrides）。</li></ul> |
 | **v2.3.2413** | <ul><li>**修复 TTFT 永久为空**: #65（pi 会话粘连）与 #67（temp-file 重写）合并冲突解决时丢失了 #65 的 \`statsKey\` 文件名优先逻辑，handler 用会话 UUID 命名临时文件而 reader 用 requestId 查找 → 永久不匹配 → 所有流式请求的 TTFT 都是 null、tokens_per_second 退化为按完整时长重算。修复后 handler 优先用 \`metadata.statsKey\` 命名文件（对齐 reader），TTFT 与真实速度恢复。</li></ul> |

package/dist/api/empty-answer-guard.d.ts

@@ -0,0 +1,66 @@
+/**
+ * Detect an upstream answer that completed successfully (HTTP 200) but carried
+ * no content blocks at all.
+ *
+ * The client shows such a turn as "no response" — a message is sent and nothing
+ * happens. Because the wire response is an ordinary 200 with a normal
+ * message_start / message_delta / message_stop sequence, nothing in the request
+ * path throws, so neither the same-model retry nor the fallback chain runs, and
+ * the empty turn is recorded as a provider success (which also keeps the
+ * unhealthy model out of the fail pool).
+ *
+ * Judging the whole content-block set rather than the text keeps legitimate
+ * turns alive: a tool_use-only turn, a stop_sequence hit, and a refusal (which
+ * still emits a text block) all contain at least one content block.
+ *
+ * The probe is bounded by EMPTY_ANSWER_INSPECT_BYTES, so a healthy answer is
+ * released right after its first content block and is never buffered further.
+ */
+/** Upper bound on the bytes inspected before a stream is forwarded untouched. */
+export declare const EMPTY_ANSWER_INSPECT_BYTES: number;
+/**
+ * Upper bound on how long the probe may hold the response before forwarding.
+ *
+ * Long enough that a thinking model's first content block arrives well within it
+ * on a healthy stream, short enough that a stalled upstream cannot pin the
+ * request (and its connection) open.
+ */
+export declare const EMPTY_ANSWER_INSPECT_DEADLINE_MS = 10000;
+/** `content` = real output; `empty` = finished with no block; `pending` = undecided. */
+export type AnswerVerdict = "content" | "empty" | "pending";
+/**
+ * Judge the buffered prefix of an answer.
+ *
+ * Line-prefix scanning (rather than JSON parsing) is deliberate: Anthropic SSE
+ * names each event on both `event:` and `data:` lines, and payloads may be split
+ * across reads, so a marker scan stays correct and cheap on every chunk.
+ */
+export declare function inspectAnswerPrefix(text: string): AnswerVerdict;
+export interface AnswerProbeResult {
+    verdict: AnswerVerdict;
+    /** Chunks consumed while probing, in order, to be re-emitted verbatim. */
+    chunks: Uint8Array[];
+    /** Bytes consumed while probing. */
+    bytes: number;
+    /** Remaining upstream stream, or null when the answer ended / was cancelled. */
+    rest: ReadableStream<Uint8Array> | null;
+}
+/**
+ * Read enough of `body` to decide whether the answer carries any content.
+ *
+ * Consumed bytes are returned in `chunks` so a healthy stream can be rebuilt
+ * byte-for-byte. Reading stops at the first content block (fast, non-blocking
+ * for real answers), at the byte budget (forwarded untouched — an unknown shape
+ * is never second-guessed), at the terminal event (verdict `empty`), or when
+ * `deadlineMs` elapses.
+ *
+ * The deadline matters: before this guard, the first SSE event was forwarded as
+ * soon as it arrived. Waiting for a content block is open-ended by nature (a
+ * thinking model may emit nothing but heartbeats for a long time), so an
+ * upstream that stalls after `message_start` would hold the response — and the
+ * whole request — forever. On timeout the answer is treated as `pending` and
+ * forwarded as-is, which preserves the old behaviour for slow-but-alive streams.
+ */
+export declare function probeAnswer(body: ReadableStream<Uint8Array>, limitBytes?: number, deadlineMs?: number): Promise<AnswerProbeResult>;
+/** Re-emit probed chunks, then continue with `rest`. */
+export declare function rebuildAnswerStream(chunks: Uint8Array[], rest: ReadableStream<Uint8Array> | null): ReadableStream<Uint8Array>;

package/dist/ccr/models.d.ts

@@ -23,4 +23,12 @@ export interface AnthropicModelInfo {
         maxOutputTokens: number;
     };
 }
-export declare function listAnthropicCompatibleModels(config: any): AnthropicModelInfo[];
+/**
+ * Predicate used to hide a provider/model pair from the model list.
+ *
+ * Pairs it returns true for are still routed (fallback, an explicit request for
+ * that exact model) but the client never sees them, so an unhealthy model stops
+ * being picked session after session.
+ */
+export type ModelAvailabilityFilter = (provider: string, model: string) => boolean;
+export declare function listAnthropicCompatibleModels(config: any, isAvailable?: ModelAvailabilityFilter): AnthropicModelInfo[];

package/dist/ccr/plugin-install-routes.d.ts

@@ -31,6 +31,17 @@ export interface CatalogEntry {
     version: string;
     installed: boolean;
     installedVersion: string | null;
+    /**
+     * Present only when `installed` and the versions differ: the client shows an
+     * update affordance for exactly these entries. Compares the npm `version`
+     * against the installed manifest version, so a downgrade is also surfaced
+     * rather than silently hidden.
+     */
+    updateAvailable?: boolean;
+    /** True for a plugin installed locally that npm search did not return (unlisted,
+     *  renamed, or an offline/local-only install). No npm version is known, so the
+     *  row shows the installed version without an update affordance. */
+    localOnly?: boolean;
 }
 export interface CatalogResult {
     ok: true;