mslxdff 0.1.203 → 0.1.207

package/src/routes/responses-route.js

@@ -1,213 +1,213 @@
-import { json, readBody } from "./helpers.js";
-import { createChatPipeline } from "../chat-pipeline/index.js";
-import { responsesToChatBody, chatJsonToResponse, createChunkTranslator, chunkToString } from "../responses/translate.js";
-
-/**
- * POST /v1/responses — 给 Codex/ChatGPT 桌面端用的 Responses API 薄壳。
- * 复用 ChatPipeline 全链路（auto/hedge/failover/tool_calls），只做形状翻译。
- * 非流式：收集 chat JSON → 转 Response 对象；流式：逐块实时翻成 responses SSE。
- */
-
-export function withBody(req, body) {
-  return Object.assign(Object.create(Object.getPrototypeOf(req)), req, { body });
-}
-
-function errorShape(text, status) {
-  let msg = String(text || "").slice(0, 500);
-  try {
-    const j = JSON.parse(String(text || ""));
-    msg = String(j?.error?.message || j?.error || msg);
-  } catch {}
-  return { error: { message: msg || "upstream error", type: "server_error", code: status } };
-}
-
-// 最小事件发射器：pipeline 靠 res.on("close") 感知下游断开，垫片必须有
-export function createEmitter() {
-  const map = new Map();
-  const self = {
-    on(ev, fn) { if (typeof fn === "function") { if (!map.has(ev)) map.set(ev, []); map.get(ev).push(fn); } return self; },
-    once(ev, fn) { const w = (...a) => { self.off(ev, w); fn(...a); }; return self.on(ev, w); },
-    off(ev, fn) { const l = map.get(ev); if (l) { const i = l.indexOf(fn); if (i >= 0) l.splice(i, 1); } return self; },
-    removeListener(ev, fn) { return self.off(ev, fn); },
-    emit(ev, ...a) { for (const fn of [...(map.get(ev) || [])]) { try { fn(...a); } catch {} } return true; },
-  };
-  return self;
-}
-
-// 非流式：收集捕获
-export function createCollector() {
-  let status = 200;
-  let text = "";
-  const res = {
-    ...createEmitter(),
-    set statusCode(v) { status = v; },
-    get statusCode() { return status; },
-    headersSent: false,
-    setHeader() {},
-    write(c) { text += chunkToString(c); return true; },
-    end(c) { if (c != null) text += chunkToString(c); res.headersSent = true; },
-  };
-  return { res, get: () => ({ status, text }) };
-}
-
-/**
- * 纯错误帧直通判定（垫片 write 的前置拦截）：整段只含 `data:` 行、且每帧都是
- * `{error:{…}}` 形状（无 `choices`）时，返回原样 SSE 帧数组；否则返回 null（交翻译器）。
- * 为什么必须拦：翻译器只认 chat 帧（choices.delta），错误帧会被它静默 `continue` 掉。
- */
-function passthroughErrorFrames(text) {
-  const src = String(text ?? "");
-  if (!src.includes("data:")) return null;
-  const out = [];
-  let sawError = false;
-  for (const rawLine of src.split("\n")) {
-    const line = rawLine.trim();
-    if (!line) continue;
-    if (line.startsWith(":")) { out.push(`${line}\n\n`); continue; } // 注释帧照原样（翻译器也是跳过）
-    if (!line.startsWith("data:")) return null;
-    const payload = line.slice(5).trim();
-    if (!payload) continue;
-    if (payload === "[DONE]") { out.push("data: [DONE]\n\n"); continue; }
-    let j;
-    try { j = JSON.parse(payload); } catch { return null; } // 解析失败 = 不是我们的错误帧，交翻译器（它自会记 jsonFails）
-    if (!j || typeof j !== "object" || Array.isArray(j) || j.choices) return null; // 有 choices = chat 帧，绝不拦
-    if (!j.error) return null; // 无 choices 又无 error：形状未知，交翻译器
-    sawError = true;
-    out.push(`data: ${JSON.stringify(j)}\n\n`);
-  }
-  return sawError ? out : null;
-}
-
-// 流式：逐 write 实时翻译并转发（head 延迟到首块，避免错误状态码已发送）
-export function createLiveForwarder(realRes, translator) {
-  let status = 200;
-  let headSent = false;
-  let ended = false;
-  const sendHead = () => {
-    if (headSent) return;
-    headSent = true;
-    realRes.writeHead(200, { "Content-Type": "text/event-stream", "Cache-Control": "no-cache", Connection: "keep-alive" });
-    for (const e of translator.begin()) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
-  };
-  // headersSent 必须是**计算值**：垫片的 write() 会调 sendHead() 真发头，若 headersSent 仍是普通属性
-  // （只在 end() 里置真），则「头已发但流未结束」这段窗口里它为假 ⇒ helpers.json() 会走 `!headersSent`
-  // 分支去 setHeader+end（而本垫片 setHeader 是 no-op、且真头已发）⇒ 错误体静默丢失、只剩裸关流。
-  // messages-route 的 createAnthropicForwarder 早已用 getter 实现并点名了同一个坑；本处是同契约的遗漏。
-  // 口径与 messages-route 的 createAnthropicForwarder 逐字对齐（哪里算「头已发」只此一处判）。
-  const sent = () => headSent || realRes.headersSent === true;
-  const res = {
-    ...createEmitter(),
-    set statusCode(v) { status = v; },
-    get statusCode() { return status; },
-    get headersSent() { return sent(); },
-    set headersSent(_v) { /* 兼容旧赋值点：真实状态由 headSent/ended 推导，不接受外部覆写 */ },
-    getHeader(k) { return String(k).toLowerCase() === "content-type" && sent() ? "text/event-stream" : undefined; },
-    setHeader() {},
-    write(c) {
-      if (ended) return true;
-      sendHead();
-      const text = chunkToString(c);
-      // 纯错误帧直写：helpers.json() 的 in-band 收场写的是 OpenAI 形状 `{error:{…}}`（无 choices），
-      // 若交给 chat→responses 翻译器，会被当作「非 chat 帧」静默忽略（translate.js 只认 choices.delta）——
-      // 错误在翻译层被吞掉，客户端只看到流被关。错误帧与 chat 帧形状互斥，故这里拦下直通。
-      const errOut = passthroughErrorFrames(text);
-      if (errOut != null) { for (const chunk of errOut) realRes.write(chunk); return true; }
-      for (const e of translator.push(text)) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
-      return true;
-    },
-    end(c) {
-      if (ended) return;
-      ended = true;
-      res.headersSent = true;
-      if (status >= 400 && !headSent) {
-        json(realRes, status, errorShape(c, status));
-        return;
-      }
-      sendHead();
-      if (c != null) for (const e of translator.push(chunkToString(c))) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
-      const fin = translator.getFinal();
-      for (const e of translator.end({ finish: fin.finish, usage: fin.usage })) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
-      realRes.write("data: [DONE]\n\n");
-      try { realRes.end(); } catch {}
-    },
-  };
-  return res;
-}
-
-const RDEBUG = process.env.MSLXDFF_RESPONSES_DEBUG === "1";
-function rlog(...a) {
-  if (RDEBUG) console.log("[responses]", ...a);
-}
-
-export async function responsesHandler(ctx) {
-  const { req, res } = ctx;
-  const t0 = Date.now();
-  let body;
-  try {
-    body = await readBody(req);
-  } catch {
-    return json(res, 400, { error: { message: "Invalid JSON body" } });
-  }
-  rlog("req", JSON.stringify({
-    model: body?.model, stream: body?.stream,
-    inputType: typeof body?.input, inputLen: Array.isArray(body?.input) ? body.input.length : String(body?.input || "").length,
-    inputKinds: Array.isArray(body?.input) ? [...new Set(body.input.map((i) => i?.type))] : null,
-    tools: Array.isArray(body?.tools) ? body.tools.length : 0,
-    instructionsLen: String(body?.instructions || "").length,
-    // 文本丢失定位：各 message item 的 content 形状（part 类型 + 文本长度 + 头部）
-    contentShapes: Array.isArray(body?.input)
-      ? body.input.filter((i) => i && (i.type === "message" || i.role)).map((m) => ({
-          type: m.type ?? "(无type)", role: m.role,
-          cType: Array.isArray(m.content) ? "array" : typeof m.content,
-          parts: Array.isArray(m.content) ? m.content.map((p) => `${p?.type ?? "?"}:${typeof p?.text === "string" ? p.text.length : "-"}`).slice(0, 6) : null,
-          textHead: typeof m.content === "string" ? m.content.slice(0, 50) : null,
-        })).slice(0, 10)
-      : null,
-  }));
-  let chatBody;
-  try {
-    chatBody = responsesToChatBody(body);
-  } catch (e) {
-    rlog("translate-req-400", String(e?.message || e));
-    return json(res, 400, { error: { message: String(e?.message || e) } });
-  }
-  rlog("chat", JSON.stringify({ model: chatBody.model, stream: chatBody.stream, msgs: chatBody.messages.map((m) => `${m.role}:${String(m.content || "").length}${m.tool_calls ? `+${m.tool_calls.length}tc` : ""}`) }));
-  const pipeline = createChatPipeline(ctx);
-  const fakeReq = withBody(req, chatBody);
-  // 真连接断开 → 透传给垫片，pipeline 的 abort 链路不断
-  const forwardClose = (shim) => { try { req.on?.("close", () => shim.emit("close")); } catch {} };
-  try {
-    if (chatBody.stream) {
-      const translator = createChunkTranslator(chatBody.model);
-      const live = createLiveForwarder(res, translator);
-      forwardClose(live);
-      await pipeline.execute({ req: fakeReq, res: live });
-      const st = translator.stats();
-      rlog("done-stream", JSON.stringify({ ms: Date.now() - t0, model: chatBody.model, ...st, finish: translator.getFinal().finish, usage: translator.getFinal().usage }));
-    } else {
-      const cap = createCollector();
-      forwardClose(cap.res);
-      await pipeline.execute({ req: fakeReq, res: cap.res });
-      const { status, text } = cap.get();
-      rlog("done-json", JSON.stringify({ ms: Date.now() - t0, model: chatBody.model, status, bytes: text.length, head: text.slice(0, 160) }));
-      if (status >= 400) return json(res, status, errorShape(text, status));
-      let chatJson = null;
-      try { chatJson = JSON.parse(text); } catch {}
-      if (!chatJson || chatJson.object === "error" || chatJson.error) {
-        rlog("translate-resp-502", text.slice(0, 300));
-        return json(res, status >= 400 ? status : 502, errorShape(text, status));
-      }
-      const respObj = chatJsonToResponse(chatJson, chatBody.model);
-      rlog("done-resp", JSON.stringify({ ms: Date.now() - t0, model: chatBody.model, status: respObj.status, items: respObj.output.length, textLen: JSON.stringify(respObj.output).length, usage: respObj.usage }));
-      // 非 JSON 透传（上游偶发）：包成纯文本 Response
-      if (!chatJson.choices) {
-        return json(res, 200, { id: chatJson.id || `resp_${Date.now()}`, object: "response", created_at: Math.floor(Date.now() / 1000), status: "completed", model: chatBody.model, output: [{ type: "message", role: "assistant", content: [{ type: "output_text", text: String(text).slice(0, 8000) }] }], usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 } });
-      }
-      return json(res, 200, respObj);
-    }
-  } catch (err) {
-    rlog("execute-throw", String(err?.message || err).slice(0, 300));
-    if (!res.headersSent) return json(res, 502, { error: { message: String(err?.message || err).slice(0, 500) } });
-    try { res.end(); } catch {}
-  }
-}
+import { json, readBody } from "./helpers.js";
+import { createChatPipeline } from "../chat-pipeline/index.js";
+import { responsesToChatBody, chatJsonToResponse, createChunkTranslator, chunkToString } from "../responses/translate.js";
+
+/**
+ * POST /v1/responses — 给 Codex/ChatGPT 桌面端用的 Responses API 薄壳。
+ * 复用 ChatPipeline 全链路（auto/hedge/failover/tool_calls），只做形状翻译。
+ * 非流式：收集 chat JSON → 转 Response 对象；流式：逐块实时翻成 responses SSE。
+ */
+
+export function withBody(req, body) {
+  return Object.assign(Object.create(Object.getPrototypeOf(req)), req, { body });
+}
+
+function errorShape(text, status) {
+  let msg = String(text || "").slice(0, 500);
+  try {
+    const j = JSON.parse(String(text || ""));
+    msg = String(j?.error?.message || j?.error || msg);
+  } catch {}
+  return { error: { message: msg || "upstream error", type: "server_error", code: status } };
+}
+
+// 最小事件发射器：pipeline 靠 res.on("close") 感知下游断开，垫片必须有
+export function createEmitter() {
+  const map = new Map();
+  const self = {
+    on(ev, fn) { if (typeof fn === "function") { if (!map.has(ev)) map.set(ev, []); map.get(ev).push(fn); } return self; },
+    once(ev, fn) { const w = (...a) => { self.off(ev, w); fn(...a); }; return self.on(ev, w); },
+    off(ev, fn) { const l = map.get(ev); if (l) { const i = l.indexOf(fn); if (i >= 0) l.splice(i, 1); } return self; },
+    removeListener(ev, fn) { return self.off(ev, fn); },
+    emit(ev, ...a) { for (const fn of [...(map.get(ev) || [])]) { try { fn(...a); } catch {} } return true; },
+  };
+  return self;
+}
+
+// 非流式：收集捕获
+export function createCollector() {
+  let status = 200;
+  let text = "";
+  const res = {
+    ...createEmitter(),
+    set statusCode(v) { status = v; },
+    get statusCode() { return status; },
+    headersSent: false,
+    setHeader() {},
+    write(c) { text += chunkToString(c); return true; },
+    end(c) { if (c != null) text += chunkToString(c); res.headersSent = true; },
+  };
+  return { res, get: () => ({ status, text }) };
+}
+
+/**
+ * 纯错误帧直通判定（垫片 write 的前置拦截）：整段只含 `data:` 行、且每帧都是
+ * `{error:{…}}` 形状（无 `choices`）时，返回原样 SSE 帧数组；否则返回 null（交翻译器）。
+ * 为什么必须拦：翻译器只认 chat 帧（choices.delta），错误帧会被它静默 `continue` 掉。
+ */
+function passthroughErrorFrames(text) {
+  const src = String(text ?? "");
+  if (!src.includes("data:")) return null;
+  const out = [];
+  let sawError = false;
+  for (const rawLine of src.split("\n")) {
+    const line = rawLine.trim();
+    if (!line) continue;
+    if (line.startsWith(":")) { out.push(`${line}\n\n`); continue; } // 注释帧照原样（翻译器也是跳过）
+    if (!line.startsWith("data:")) return null;
+    const payload = line.slice(5).trim();
+    if (!payload) continue;
+    if (payload === "[DONE]") { out.push("data: [DONE]\n\n"); continue; }
+    let j;
+    try { j = JSON.parse(payload); } catch { return null; } // 解析失败 = 不是我们的错误帧，交翻译器（它自会记 jsonFails）
+    if (!j || typeof j !== "object" || Array.isArray(j) || j.choices) return null; // 有 choices = chat 帧，绝不拦
+    if (!j.error) return null; // 无 choices 又无 error：形状未知，交翻译器
+    sawError = true;
+    out.push(`data: ${JSON.stringify(j)}\n\n`);
+  }
+  return sawError ? out : null;
+}
+
+// 流式：逐 write 实时翻译并转发（head 延迟到首块，避免错误状态码已发送）
+export function createLiveForwarder(realRes, translator) {
+  let status = 200;
+  let headSent = false;
+  let ended = false;
+  const sendHead = () => {
+    if (headSent) return;
+    headSent = true;
+    realRes.writeHead(200, { "Content-Type": "text/event-stream", "Cache-Control": "no-cache", Connection: "keep-alive" });
+    for (const e of translator.begin()) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
+  };
+  // headersSent 必须是**计算值**：垫片的 write() 会调 sendHead() 真发头，若 headersSent 仍是普通属性
+  // （只在 end() 里置真），则「头已发但流未结束」这段窗口里它为假 ⇒ helpers.json() 会走 `!headersSent`
+  // 分支去 setHeader+end（而本垫片 setHeader 是 no-op、且真头已发）⇒ 错误体静默丢失、只剩裸关流。
+  // messages-route 的 createAnthropicForwarder 早已用 getter 实现并点名了同一个坑；本处是同契约的遗漏。
+  // 口径与 messages-route 的 createAnthropicForwarder 逐字对齐（哪里算「头已发」只此一处判）。
+  const sent = () => headSent || realRes.headersSent === true;
+  const res = {
+    ...createEmitter(),
+    set statusCode(v) { status = v; },
+    get statusCode() { return status; },
+    get headersSent() { return sent(); },
+    set headersSent(_v) { /* 兼容旧赋值点：真实状态由 headSent/ended 推导，不接受外部覆写 */ },
+    getHeader(k) { return String(k).toLowerCase() === "content-type" && sent() ? "text/event-stream" : undefined; },
+    setHeader() {},
+    write(c) {
+      if (ended) return true;
+      sendHead();
+      const text = chunkToString(c);
+      // 纯错误帧直写：helpers.json() 的 in-band 收场写的是 OpenAI 形状 `{error:{…}}`（无 choices），
+      // 若交给 chat→responses 翻译器，会被当作「非 chat 帧」静默忽略（translate.js 只认 choices.delta）——
+      // 错误在翻译层被吞掉，客户端只看到流被关。错误帧与 chat 帧形状互斥，故这里拦下直通。
+      const errOut = passthroughErrorFrames(text);
+      if (errOut != null) { for (const chunk of errOut) realRes.write(chunk); return true; }
+      for (const e of translator.push(text)) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
+      return true;
+    },
+    end(c) {
+      if (ended) return;
+      ended = true;
+      res.headersSent = true;
+      if (status >= 400 && !headSent) {
+        json(realRes, status, errorShape(c, status));
+        return;
+      }
+      sendHead();
+      if (c != null) for (const e of translator.push(chunkToString(c))) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
+      const fin = translator.getFinal();
+      for (const e of translator.end({ finish: fin.finish, usage: fin.usage })) realRes.write(`data: ${JSON.stringify(e)}\n\n`);
+      realRes.write("data: [DONE]\n\n");
+      try { realRes.end(); } catch {}
+    },
+  };
+  return res;
+}
+
+const RDEBUG = process.env.MSLXDFF_RESPONSES_DEBUG === "1";
+function rlog(...a) {
+  if (RDEBUG) console.log("[responses]", ...a);
+}
+
+export async function responsesHandler(ctx) {
+  const { req, res } = ctx;
+  const t0 = Date.now();
+  let body;
+  try {
+    body = await readBody(req);
+  } catch {
+    return json(res, 400, { error: { message: "Invalid JSON body" } });
+  }
+  rlog("req", JSON.stringify({
+    model: body?.model, stream: body?.stream,
+    inputType: typeof body?.input, inputLen: Array.isArray(body?.input) ? body.input.length : String(body?.input || "").length,
+    inputKinds: Array.isArray(body?.input) ? [...new Set(body.input.map((i) => i?.type))] : null,
+    tools: Array.isArray(body?.tools) ? body.tools.length : 0,
+    instructionsLen: String(body?.instructions || "").length,
+    // 文本丢失定位：各 message item 的 content 形状（part 类型 + 文本长度 + 头部）
+    contentShapes: Array.isArray(body?.input)
+      ? body.input.filter((i) => i && (i.type === "message" || i.role)).map((m) => ({
+          type: m.type ?? "(无type)", role: m.role,
+          cType: Array.isArray(m.content) ? "array" : typeof m.content,
+          parts: Array.isArray(m.content) ? m.content.map((p) => `${p?.type ?? "?"}:${typeof p?.text === "string" ? p.text.length : "-"}`).slice(0, 6) : null,
+          textHead: typeof m.content === "string" ? m.content.slice(0, 50) : null,
+        })).slice(0, 10)
+      : null,
+  }));
+  let chatBody;
+  try {
+    chatBody = responsesToChatBody(body);
+  } catch (e) {
+    rlog("translate-req-400", String(e?.message || e));
+    return json(res, 400, { error: { message: String(e?.message || e) } });
+  }
+  rlog("chat", JSON.stringify({ model: chatBody.model, stream: chatBody.stream, msgs: chatBody.messages.map((m) => `${m.role}:${String(m.content || "").length}${m.tool_calls ? `+${m.tool_calls.length}tc` : ""}`) }));
+  const pipeline = createChatPipeline(ctx);
+  const fakeReq = withBody(req, chatBody);
+  // 真连接断开 → 透传给垫片，pipeline 的 abort 链路不断
+  const forwardClose = (shim) => { try { req.on?.("close", () => shim.emit("close")); } catch {} };
+  try {
+    if (chatBody.stream) {
+      const translator = createChunkTranslator(chatBody.model);
+      const live = createLiveForwarder(res, translator);
+      forwardClose(live);
+      await pipeline.execute({ req: fakeReq, res: live });
+      const st = translator.stats();
+      rlog("done-stream", JSON.stringify({ ms: Date.now() - t0, model: chatBody.model, ...st, finish: translator.getFinal().finish, usage: translator.getFinal().usage }));
+    } else {
+      const cap = createCollector();
+      forwardClose(cap.res);
+      await pipeline.execute({ req: fakeReq, res: cap.res });
+      const { status, text } = cap.get();
+      rlog("done-json", JSON.stringify({ ms: Date.now() - t0, model: chatBody.model, status, bytes: text.length, head: text.slice(0, 160) }));
+      if (status >= 400) return json(res, status, errorShape(text, status));
+      let chatJson = null;
+      try { chatJson = JSON.parse(text); } catch {}
+      if (!chatJson || chatJson.object === "error" || chatJson.error) {
+        rlog("translate-resp-502", text.slice(0, 300));
+        return json(res, status >= 400 ? status : 502, errorShape(text, status));
+      }
+      const respObj = chatJsonToResponse(chatJson, chatBody.model);
+      rlog("done-resp", JSON.stringify({ ms: Date.now() - t0, model: chatBody.model, status: respObj.status, items: respObj.output.length, textLen: JSON.stringify(respObj.output).length, usage: respObj.usage }));
+      // 非 JSON 透传（上游偶发）：包成纯文本 Response
+      if (!chatJson.choices) {
+        return json(res, 200, { id: chatJson.id || `resp_${Date.now()}`, object: "response", created_at: Math.floor(Date.now() / 1000), status: "completed", model: chatBody.model, output: [{ type: "message", role: "assistant", content: [{ type: "output_text", text: String(text).slice(0, 8000) }] }], usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 } });
+      }
+      return json(res, 200, respObj);
+    }
+  } catch (err) {
+    rlog("execute-throw", String(err?.message || err).slice(0, 300));
+    if (!res.headersSent) return json(res, 502, { error: { message: String(err?.message || err).slice(0, 500) } });
+    try { res.end(); } catch {}
+  }
+}

package/src/routes/stream-body.js

@@ -1,84 +1,84 @@
-// relay 的读取层工具（从 stream.js 拆出：HEAD 版已 19.8KB / 20KB 硬门，先拆后写）。
-// 全部是「不碰闸门、不碰 res、不碰计时」的纯 IO 适配：字节→文本、取消上游、注释帧判定、读取迭代。
-// 语义与拆出前逐字一致 —— 这里只搬家，不改行为。
-import { enrichSseChunkText, enrichNonStreamJson as enrichNonStreamJsonFn } from "./fallback.js";
-import { extractUsageFromSseText as extractUsageFn } from "../metrics.js";
-import { scanNonStreamBody as scanNonStreamBodyFn } from "./stream-scan.js";
-
-// SDK 通道（TextEncoder）产出 Uint8Array，legacy 通道为 Buffer；
-// 统一转文本，避免 [DONE]/finish_reason/usage/chars 统计在 SDK 路径下静默失效。
-export function chunkText(chunk) {
-  if (typeof chunk === "string") return chunk;
-  if (Buffer.isBuffer(chunk)) return chunk.toString("utf8");
-  if (chunk instanceof Uint8Array) return Buffer.from(chunk).toString("utf8");
-  return "";
-}
-
-// body.cancel() 可能返回非 Promise（自定义/AI SDK 流）——同步异常与 rejection 双路径都要吞掉
-export function cancelBody(body) {
-  try {
-    const p = typeof body?.cancel === "function" ? body.cancel() : null;
-    if (p && typeof p.catch === "function") p.catch(() => {});
-  } catch { /* ignore */ }
-}
-
-// 注释帧（": keepalive" 等）不是模型输出：不算首块、不解除闸门、不触发超时救回，但照常透传
-export function hasPayload(text) {
-  for (const line of String(text).split("\n")) {
-    const t = line.trim();
-    if (!t || t.startsWith(":")) continue;
-    return true;
-  }
-  return false;
-}
-
-/**
- * 非流式的「先判后写」：在 headers 还没 flush 之前解析上游体，决定能不能富化 / 是不是空轮。
- * 返回 { enriched, emptyTurn }：enriched=null 表示上游给的不是可解析 JSON（调用方走原文透传）；
- * emptyTurn=true 表示这一发是空轮、且留口是开的（调用方沿用留口语义，不写任何响应）。
- * 纯搬家：行为与拆出前逐字一致。
- */
-export function scanNonStreamReply(detail, text, { fallback, holdEnd, isEmptyTurn, captureFallback }) {
-  let enriched = null;
-  try {
-    const parsed = JSON.parse(text);
-    scanNonStreamBodyFn(detail, parsed);
-    enriched = enrichNonStreamJsonFn(parsed, fallback);
-  } catch { /* 上游给的不是可解析 JSON：调用方走原文透传 */ }
-  if (enriched === null) {
-    // 纯文本时按长度估 chars；若文本其实是 SSE（client 未要流但上游给流）仍尝试提 usage
-    try { detail.chars = text.length; } catch {}
-    try { const u = extractUsageFn(text); if (u) detail.usage = u; } catch {}
-    try { if (detail.talk) captureFallback(detail.talk, text); } catch {}
-    return { enriched: null, emptyTurn: false };
-  }
-  const emptyTurn = Boolean(holdEnd() && !detail.downstreamClosed && isEmptyTurn(detail));
-  return { enriched, emptyTurn };
-}
-
-// 上游回的是普通 reply（无流式 delta）时，客户端仍要看到增量：把整块按 fallback 富化成 SSE 帧。
-// 只在首块前、且本轮尚未写出 payload 时调用一次；解析不出就原样返回（绝不改字节）。
-export function enrichFirstReply(chunk, fallback) {
-  if (!fallback?.fallback) return chunk;
-  try {
-    const txt = chunkText(chunk);
-    if (!txt.includes("data:")) return chunk;
-    const enriched = enrichSseChunkText(txt, fallback);
-    return enriched !== txt ? Buffer.from(enriched, "utf8") : chunk;
-  } catch {
-    return chunk;
-  }
-}
-
-// 自持 reader 优先：for-await 会锁定 ReadableStream，使 body.cancel() 必 reject（真流上等于空操作），
-// 超时/断下游要真能掐上游必须走 reader.cancel()
-export async function* bodyChunks(body, reader) {
-  if (reader) {
-    for (;;) {
-      const { done, value } = await reader.read();
-      if (done) return;
-      yield value;
-    }
-  }
-  for await (const c of body) yield c;
-}
+// relay 的读取层工具（从 stream.js 拆出：HEAD 版已 19.8KB / 20KB 硬门，先拆后写）。
+// 全部是「不碰闸门、不碰 res、不碰计时」的纯 IO 适配：字节→文本、取消上游、注释帧判定、读取迭代。
+// 语义与拆出前逐字一致 —— 这里只搬家，不改行为。
+import { enrichSseChunkText, enrichNonStreamJson as enrichNonStreamJsonFn } from "./fallback.js";
+import { extractUsageFromSseText as extractUsageFn } from "../metrics.js";
+import { scanNonStreamBody as scanNonStreamBodyFn } from "./stream-scan.js";
+
+// SDK 通道（TextEncoder）产出 Uint8Array，legacy 通道为 Buffer；
+// 统一转文本，避免 [DONE]/finish_reason/usage/chars 统计在 SDK 路径下静默失效。
+export function chunkText(chunk) {
+  if (typeof chunk === "string") return chunk;
+  if (Buffer.isBuffer(chunk)) return chunk.toString("utf8");
+  if (chunk instanceof Uint8Array) return Buffer.from(chunk).toString("utf8");
+  return "";
+}
+
+// body.cancel() 可能返回非 Promise（自定义/AI SDK 流）——同步异常与 rejection 双路径都要吞掉
+export function cancelBody(body) {
+  try {
+    const p = typeof body?.cancel === "function" ? body.cancel() : null;
+    if (p && typeof p.catch === "function") p.catch(() => {});
+  } catch { /* ignore */ }
+}
+
+// 注释帧（": keepalive" 等）不是模型输出：不算首块、不解除闸门、不触发超时救回，但照常透传
+export function hasPayload(text) {
+  for (const line of String(text).split("\n")) {
+    const t = line.trim();
+    if (!t || t.startsWith(":")) continue;
+    return true;
+  }
+  return false;
+}
+
+/**
+ * 非流式的「先判后写」：在 headers 还没 flush 之前解析上游体，决定能不能富化 / 是不是空轮。
+ * 返回 { enriched, emptyTurn }：enriched=null 表示上游给的不是可解析 JSON（调用方走原文透传）；
+ * emptyTurn=true 表示这一发是空轮、且留口是开的（调用方沿用留口语义，不写任何响应）。
+ * 纯搬家：行为与拆出前逐字一致。
+ */
+export function scanNonStreamReply(detail, text, { fallback, holdEnd, isEmptyTurn, captureFallback }) {
+  let enriched = null;
+  try {
+    const parsed = JSON.parse(text);
+    scanNonStreamBodyFn(detail, parsed);
+    enriched = enrichNonStreamJsonFn(parsed, fallback);
+  } catch { /* 上游给的不是可解析 JSON：调用方走原文透传 */ }
+  if (enriched === null) {
+    // 纯文本时按长度估 chars；若文本其实是 SSE（client 未要流但上游给流）仍尝试提 usage
+    try { detail.chars = text.length; } catch {}
+    try { const u = extractUsageFn(text); if (u) detail.usage = u; } catch {}
+    try { if (detail.talk) captureFallback(detail.talk, text); } catch {}
+    return { enriched: null, emptyTurn: false };
+  }
+  const emptyTurn = Boolean(holdEnd() && !detail.downstreamClosed && isEmptyTurn(detail));
+  return { enriched, emptyTurn };
+}
+
+// 上游回的是普通 reply（无流式 delta）时，客户端仍要看到增量：把整块按 fallback 富化成 SSE 帧。
+// 只在首块前、且本轮尚未写出 payload 时调用一次；解析不出就原样返回（绝不改字节）。
+export function enrichFirstReply(chunk, fallback) {
+  if (!fallback?.fallback) return chunk;
+  try {
+    const txt = chunkText(chunk);
+    if (!txt.includes("data:")) return chunk;
+    const enriched = enrichSseChunkText(txt, fallback);
+    return enriched !== txt ? Buffer.from(enriched, "utf8") : chunk;
+  } catch {
+    return chunk;
+  }
+}
+
+// 自持 reader 优先：for-await 会锁定 ReadableStream，使 body.cancel() 必 reject（真流上等于空操作），
+// 超时/断下游要真能掐上游必须走 reader.cancel()
+export async function* bodyChunks(body, reader) {
+  if (reader) {
+    for (;;) {
+      const { done, value } = await reader.read();
+      if (done) return;
+      yield value;
+    }
+  }
+  for await (const c of body) yield c;
+}