mslxdff 0.1.203 → 0.1.208

package/src/providers/raccoon/sse.js

@@ -1,135 +1,135 @@
-// raccoon 流处理：上游就是 OpenAI 形（`/api/web/llm/v2/chat/completions`），
-// 因此**不做协议转换**——本文件只做三件事：形状判定、错误信封分类、非流式聚合。
-// 流式路径直接把上游 body 透传（逐字节），避免任何整形引入的失真。
-import { createSseParser } from "../../transport/sse.js";
-import { isRaccoonAuthFailure, parseRaccoonEnvelope, raccoonEnvelopeText } from "./envelope.js";
-
-/**
- * 额度类措辞（业务码未取证，见 ADR-0050 决策 8）：**宁可漏判也不误报**。
- * 刻意不收裸 `insufficient`/`balance`/`points` —— 那些词会命中「权限不足」这类非额度错，
- * 一旦误判成 quota 就会长冷 1h 并回 429 `quota_exhausted`，违反 ADR-0044 的「非额度原因不得谎报」。
- */
-const QUOTA_HINTS = /(积分|额度|quota|欠费)/i;
-
-export function isRaccoonSseContentType(value) {
-  return String(value || "").includes("text/event-stream");
-}
-
-function isQuotaLike(envelope) {
-  const text = `${envelope?.message ?? ""} ${envelope?.details ?? ""}`;
-  if (QUOTA_HINTS.test(text)) return true;
-  return envelope?.status === 402;
-}
-
-/**
- * 失败分类 → `{ kind, code, status, message }`。
- * kind 取值：auth（重登）/ quota（长冷却）/ rate_limit（短冷却）/ server（上游 5xx，短冷却）/ unknown（4xx 如实透出）。
- */
-export function classifyRaccoonFailure(status, payload) {
-  const envelope = parseRaccoonEnvelope(payload, status);
-  let kind = "unknown";
-  if (isRaccoonAuthFailure(envelope, status)) kind = "auth";
-  else if (isQuotaLike(envelope)) kind = "quota";
-  else if (status === 429) kind = "rate_limit";
-  else if (status >= 500) kind = "server";
-  return { kind, code: envelope.code, status, message: raccoonEnvelopeText(envelope, `HTTP ${status}`) };
-}
-
-const ERROR_MAP = {
-  auth: { status: 401, type: "auth_error", text: (f) => `raccoon: 登录态已失效${f.code ? `（code ${f.code}）` : ""}。请运行 mslxdff -provider raccoon login 重新登录` },
-  quota: { status: 429, type: "quota_exhausted", text: (f) => `raccoon: 积分不足${f.message ? `（${f.message}）` : ""}。请运行 mslxdff -provider raccoon checkin 领取每日积分，或更换账号` },
-  rate_limit: { status: 429, type: "rate_limited", text: () => "raccoon: 上游限流（账号冷却中），请稍后重试" },
-  server: { status: 502, type: "upstream_error", text: (f) => `raccoon: 上游服务异常${f.code ? `（code ${f.code}）` : ""}${f.message ? `：${f.message}` : ""}` },
-  // unknown 只覆盖 4xx（5xx 已被上面分流成 server）：按请求缺陷如实回 400，不谎报成网关故障
-  unknown: { status: 400, type: "invalid_request", text: (f) => `raccoon: 上游拒绝请求（HTTP ${f.status || "?"}${f.code ? ` code ${f.code}` : ""}）${f.message ? `：${f.message}` : ""}` },
-};
-
-/** 失败 → Response（带 `x-mslxdff-raccoon-kind`，供工厂决定冷却档位）。 */
-export function raccoonErrorResponse(fail = {}) {
-  const kind = ERROR_MAP[fail.kind] ? fail.kind : "unknown";
-  const entry = ERROR_MAP[kind];
-  const headers = { "Content-Type": "application/json" };
-  if (fail.kind) headers["x-mslxdff-raccoon-kind"] = fail.kind;
-  return new Response(JSON.stringify({ error: { message: entry.text(fail), type: entry.type, code: fail.code || 0 } }), {
-    status: entry.status,
-    headers,
-  });
-}
-
-/** 工具调用增量合并（对齐 `qwenwork/sse.js` 的 mergeToolCallDelta：id/type 取一次，name/arguments 追加）。 */
-function mergeRaccoonToolCallDelta(merged, delta) {
-  for (const key of ["id", "type"]) {
-    if (merged[key] === undefined && typeof delta[key] === "string" && delta[key] !== "") merged[key] = delta[key];
-  }
-  const dfn = delta.function;
-  if (!dfn || typeof dfn !== "object") return;
-  if (!merged.function || typeof merged.function !== "object") merged.function = {};
-  if (typeof dfn.name === "string" && dfn.name !== "") merged.function.name = (merged.function.name || "") + dfn.name;
-  if (typeof dfn.arguments === "string" && dfn.arguments !== "") {
-    merged.function.arguments = (merged.function.arguments || "") + dfn.arguments;
-  }
-}
-
-/** 把 SSE 文本聚合成一条 OpenAI chat completion（非流式请求、以及「header 说是 JSON 但体是 SSE」时走这里）。 */
-export async function aggregateRaccoonSse(text, { model } = {}) {
-  const parser = createSseParser();
-  // 聚合拿到的是**完整 body**（不是分片）：补一个定界符，否则末尾缺 `\n\n` 的最后一帧会被当半包丢掉
-  const events = parser.push(`${String(text || "")}\n\n`);
-  const content = [];
-  const reasoning = [];
-  const toolCalls = new Map();
-  const toolOrder = [];
-  let finishReason = "stop";
-  let usage;
-  let id = "";
-  let created = 0;
-  for (const raw of events) {
-    if (raw === "[DONE]" || !raw) continue;
-    let payload;
-    try {
-      payload = JSON.parse(raw);
-    } catch {
-      continue;
-    }
-    // 流内错误信封：优先按失败分类返回，不要静默当成空回复
-    if (payload?.code !== undefined && Number(payload.code) !== 0) {
-      return { error: classifyRaccoonFailure(200, payload) };
-    }
-    if (payload?.error) {
-      return { error: classifyRaccoonFailure(200, payload) };
-    }
-    if (!id && typeof payload?.id === "string") id = payload.id;
-    if (!created && Number.isFinite(payload?.created)) created = Number(payload.created);
-    if (payload?.usage) usage = payload.usage;
-    const choice = Array.isArray(payload?.choices) ? payload.choices[0] : undefined;
-    const delta = choice?.delta ?? choice?.message;
-    if (typeof delta?.content === "string" && delta.content.length > 0) content.push(delta.content);
-    if (typeof delta?.reasoning_content === "string" && delta.reasoning_content.length > 0) reasoning.push(delta.reasoning_content);
-    for (const call of Array.isArray(delta?.tool_calls) ? delta.tool_calls : []) {
-      if (!call || typeof call !== "object") continue;
-      const idx = typeof call.index === "number" ? call.index : 0;
-      if (!toolCalls.has(idx)) {
-        toolCalls.set(idx, { index: idx });
-        toolOrder.push(idx);
-      }
-      mergeRaccoonToolCallDelta(toolCalls.get(idx), call);
-    }
-    if (typeof choice?.finish_reason === "string" && choice.finish_reason.length > 0) finishReason = choice.finish_reason;
-  }
-  const message = { role: "assistant", content: content.join("") || null };
-  if (reasoning.length > 0) message.reasoning_content = reasoning.join("");
-  if (toolOrder.length > 0) {
-    toolOrder.sort((a, b) => a - b);
-    message.tool_calls = toolOrder.map((idx) => toolCalls.get(idx));
-  }
-  return {
-    openAi: {
-      id: id || `chatcmpl-raccoon-${Date.now()}`,
-      object: "chat.completion",
-      created: created || Math.floor(Date.now() / 1000),
-      model: model || "",
-      choices: [{ index: 0, message, finish_reason: finishReason }],
-      ...(usage ? { usage } : {}),
-    },
-  };
-}
+// raccoon 流处理：上游就是 OpenAI 形（`/api/web/llm/v2/chat/completions`），
+// 因此**不做协议转换**——本文件只做三件事：形状判定、错误信封分类、非流式聚合。
+// 流式路径直接把上游 body 透传（逐字节），避免任何整形引入的失真。
+import { createSseParser } from "../../transport/sse.js";
+import { isRaccoonAuthFailure, parseRaccoonEnvelope, raccoonEnvelopeText } from "./envelope.js";
+
+/**
+ * 额度类措辞（业务码未取证，见 ADR-0050 决策 8）：**宁可漏判也不误报**。
+ * 刻意不收裸 `insufficient`/`balance`/`points` —— 那些词会命中「权限不足」这类非额度错，
+ * 一旦误判成 quota 就会长冷 1h 并回 429 `quota_exhausted`，违反 ADR-0044 的「非额度原因不得谎报」。
+ */
+const QUOTA_HINTS = /(积分|额度|quota|欠费)/i;
+
+export function isRaccoonSseContentType(value) {
+  return String(value || "").includes("text/event-stream");
+}
+
+function isQuotaLike(envelope) {
+  const text = `${envelope?.message ?? ""} ${envelope?.details ?? ""}`;
+  if (QUOTA_HINTS.test(text)) return true;
+  return envelope?.status === 402;
+}
+
+/**
+ * 失败分类 → `{ kind, code, status, message }`。
+ * kind 取值：auth（重登）/ quota（长冷却）/ rate_limit（短冷却）/ server（上游 5xx，短冷却）/ unknown（4xx 如实透出）。
+ */
+export function classifyRaccoonFailure(status, payload) {
+  const envelope = parseRaccoonEnvelope(payload, status);
+  let kind = "unknown";
+  if (isRaccoonAuthFailure(envelope, status)) kind = "auth";
+  else if (isQuotaLike(envelope)) kind = "quota";
+  else if (status === 429) kind = "rate_limit";
+  else if (status >= 500) kind = "server";
+  return { kind, code: envelope.code, status, message: raccoonEnvelopeText(envelope, `HTTP ${status}`) };
+}
+
+const ERROR_MAP = {
+  auth: { status: 401, type: "auth_error", text: (f) => `raccoon: 登录态已失效${f.code ? `（code ${f.code}）` : ""}。请运行 mslxdff -provider raccoon login 重新登录` },
+  quota: { status: 429, type: "quota_exhausted", text: (f) => `raccoon: 积分不足${f.message ? `（${f.message}）` : ""}。请运行 mslxdff -provider raccoon checkin 领取每日积分，或更换账号` },
+  rate_limit: { status: 429, type: "rate_limited", text: () => "raccoon: 上游限流（账号冷却中），请稍后重试" },
+  server: { status: 502, type: "upstream_error", text: (f) => `raccoon: 上游服务异常${f.code ? `（code ${f.code}）` : ""}${f.message ? `：${f.message}` : ""}` },
+  // unknown 只覆盖 4xx（5xx 已被上面分流成 server）：按请求缺陷如实回 400，不谎报成网关故障
+  unknown: { status: 400, type: "invalid_request", text: (f) => `raccoon: 上游拒绝请求（HTTP ${f.status || "?"}${f.code ? ` code ${f.code}` : ""}）${f.message ? `：${f.message}` : ""}` },
+};
+
+/** 失败 → Response（带 `x-mslxdff-raccoon-kind`，供工厂决定冷却档位）。 */
+export function raccoonErrorResponse(fail = {}) {
+  const kind = ERROR_MAP[fail.kind] ? fail.kind : "unknown";
+  const entry = ERROR_MAP[kind];
+  const headers = { "Content-Type": "application/json" };
+  if (fail.kind) headers["x-mslxdff-raccoon-kind"] = fail.kind;
+  return new Response(JSON.stringify({ error: { message: entry.text(fail), type: entry.type, code: fail.code || 0 } }), {
+    status: entry.status,
+    headers,
+  });
+}
+
+/** 工具调用增量合并（对齐 `qwenwork/sse.js` 的 mergeToolCallDelta：id/type 取一次，name/arguments 追加）。 */
+function mergeRaccoonToolCallDelta(merged, delta) {
+  for (const key of ["id", "type"]) {
+    if (merged[key] === undefined && typeof delta[key] === "string" && delta[key] !== "") merged[key] = delta[key];
+  }
+  const dfn = delta.function;
+  if (!dfn || typeof dfn !== "object") return;
+  if (!merged.function || typeof merged.function !== "object") merged.function = {};
+  if (typeof dfn.name === "string" && dfn.name !== "") merged.function.name = (merged.function.name || "") + dfn.name;
+  if (typeof dfn.arguments === "string" && dfn.arguments !== "") {
+    merged.function.arguments = (merged.function.arguments || "") + dfn.arguments;
+  }
+}
+
+/** 把 SSE 文本聚合成一条 OpenAI chat completion（非流式请求、以及「header 说是 JSON 但体是 SSE」时走这里）。 */
+export async function aggregateRaccoonSse(text, { model } = {}) {
+  const parser = createSseParser();
+  // 聚合拿到的是**完整 body**（不是分片）：补一个定界符，否则末尾缺 `\n\n` 的最后一帧会被当半包丢掉
+  const events = parser.push(`${String(text || "")}\n\n`);
+  const content = [];
+  const reasoning = [];
+  const toolCalls = new Map();
+  const toolOrder = [];
+  let finishReason = "stop";
+  let usage;
+  let id = "";
+  let created = 0;
+  for (const raw of events) {
+    if (raw === "[DONE]" || !raw) continue;
+    let payload;
+    try {
+      payload = JSON.parse(raw);
+    } catch {
+      continue;
+    }
+    // 流内错误信封：优先按失败分类返回，不要静默当成空回复
+    if (payload?.code !== undefined && Number(payload.code) !== 0) {
+      return { error: classifyRaccoonFailure(200, payload) };
+    }
+    if (payload?.error) {
+      return { error: classifyRaccoonFailure(200, payload) };
+    }
+    if (!id && typeof payload?.id === "string") id = payload.id;
+    if (!created && Number.isFinite(payload?.created)) created = Number(payload.created);
+    if (payload?.usage) usage = payload.usage;
+    const choice = Array.isArray(payload?.choices) ? payload.choices[0] : undefined;
+    const delta = choice?.delta ?? choice?.message;
+    if (typeof delta?.content === "string" && delta.content.length > 0) content.push(delta.content);
+    if (typeof delta?.reasoning_content === "string" && delta.reasoning_content.length > 0) reasoning.push(delta.reasoning_content);
+    for (const call of Array.isArray(delta?.tool_calls) ? delta.tool_calls : []) {
+      if (!call || typeof call !== "object") continue;
+      const idx = typeof call.index === "number" ? call.index : 0;
+      if (!toolCalls.has(idx)) {
+        toolCalls.set(idx, { index: idx });
+        toolOrder.push(idx);
+      }
+      mergeRaccoonToolCallDelta(toolCalls.get(idx), call);
+    }
+    if (typeof choice?.finish_reason === "string" && choice.finish_reason.length > 0) finishReason = choice.finish_reason;
+  }
+  const message = { role: "assistant", content: content.join("") || null };
+  if (reasoning.length > 0) message.reasoning_content = reasoning.join("");
+  if (toolOrder.length > 0) {
+    toolOrder.sort((a, b) => a - b);
+    message.tool_calls = toolOrder.map((idx) => toolCalls.get(idx));
+  }
+  return {
+    openAi: {
+      id: id || `chatcmpl-raccoon-${Date.now()}`,
+      object: "chat.completion",
+      created: created || Math.floor(Date.now() / 1000),
+      model: model || "",
+      choices: [{ index: 0, message, finish_reason: finishReason }],
+      ...(usage ? { usage } : {}),
+    },
+  };
+}

package/src/providers/responses-channel.js

@@ -1,120 +1,120 @@
-// 通用（带 key）供应商的 responses 通道：muse-spark* 这类模型上游只挂 /responses，
-// 打同 host 的 /chat/completions 会 503 "Endpoint is unavailable"（ocgo 实测 2026-09-22）。
-// 判定单一来源 = isResponsesModel（models.dev 模型级 provider.npm + muse-spark 前缀兜底），
-// 与 GET /v1/models 的 capabilities.upstreamApi 同源，新模型无需改码。
-//
-// 契约：run() 的出参恒为 **chat 形状**（SSE 或 JSON），调用方无需再判分支。
-//   · 流式优先 @ai-sdk/openai 的 responses 适配器（含加密思考往返，ADR-0017 复用）；
-//   · SDK 不可用 / 非流式 → 原生 fetch + chatToResponsesBody 正转换 + 反向整形。
-// key 轮换、退避重试、_t 计时与 createChatRunner（base.js）同语义，避免两条通道行为漂移。
-import { chatToResponsesBody, toChatResponse, reshapeResponsesSse } from "../upstream-responses.js";
-import { attemptOnceResponsesSdk } from "../upstream-engine/sdk/responses.js";
-import { ENGINE_MARKER } from "../upstream-engine/sdk/chat.js";
-import { sleep } from "./base.js";
-
-/** 供应商 id → env 片段（`my-api` → `MY_API`），与 providerKeyEnv 同规则 */
-export function envSlug(id) {
-  return String(id || "").toUpperCase().replace(/[^A-Z0-9]/g, "_");
-}
-
-/**
- * responses 端点路径：显式入参 > env `MSLXDFF_<ID>_RESPONSES_PATH` > 缺省 `/responses`。
- * 归一为以 `/` 开头（joinUrl 会再拼 baseUrl）。
- */
-export function resolveResponsesPath(id, responsesPath) {
-  const raw = responsesPath || process.env[`MSLXDFF_${envSlug(id)}_RESPONSES_PATH`] || "";
-  const s = String(raw).trim();
-  if (!s) return "/responses";
-  return s.startsWith("/") ? s : `/${s}`;
-}
-
-/**
- * 建 responses 通道。
- * @param {object} o
- * @param {string} o.id 供应商 id（用于报错文案与 providerName）
- * @param {string} o.url 已拼好的 responses 绝对 URL
- * @param {number} o.connectTimeoutMs 单次尝试超时
- * @param {object} o.retry 重试表（与 createChatRunner 同形状：network/429/50x → {attempts, delayMs}）
- * @param {number} o.cooldownMs 全 key 冷却时的人话报错文案用
- * @param {boolean} o.sdkEnabled 是否允许走 AI SDK responses 适配器（流式路径）
- * @param {Function} o.buildHeaders (body, key, opts) → headers
- * @param {Function} o.fetchImpl 注入的 fetch（连接池/测试接缝）
- * @param {object} [o.dispatcher] undici Agent
- */
-export function createResponsesChannel({
-  id,
-  url,
-  connectTimeoutMs = 30_000,
-  retry = {},
-  cooldownMs = 30_000,
-  sdkEnabled = true,
-  buildHeaders,
-  fetchImpl,
-  dispatcher = null,
-} = {}) {
-  async function attemptOnce(body, key, opts) {
-    const headers = buildHeaders ? buildHeaders(body, key, opts) : {};
-    const wantsStream = body?.stream !== false;
-    if (sdkEnabled && wantsStream) {
-      try {
-        const r = await attemptOnceResponsesSdk({
-          url, body, headers, providerName: id, marker: ENGINE_MARKER, fetchImpl,
-        });
-        if (r) return r; // 适配器已产出 chat SSE
-      } catch (e) {
-        if (!e || (!e._sdkLoadFailed && !e._sdkUnsupported)) throw e;
-        // SDK 装载失败 / URL 异形 → 落原生兜底（保底不变差）
-      }
-    }
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(new Error(`${id} timed out after ${connectTimeoutMs}ms`)), connectTimeoutMs);
-    try {
-      const fetchOpts = { method: "POST", headers, body: JSON.stringify(chatToResponsesBody(body)), signal: controller.signal };
-      if (dispatcher) fetchOpts.dispatcher = dispatcher;
-      const res = await fetchImpl(url, fetchOpts);
-      if (!res.ok) return res;
-      if (wantsStream) return reshapeResponsesSse(res, body?.model);
-      const json = await res.json().catch(() => null);
-      return json ? toChatResponse(res, json) : res;
-    } finally {
-      clearTimeout(timer);
-    }
-  }
-
-  /** key 轮换 + 退避重试 + _t 计时；语义对齐 base.js 的 createChatRunner.runChat */
-  async function run(body, activeRing, sourceKey, opts) {
-    const t0 = performance.now();
-    const attempts = [];
-    let waitMs = 0;
-    const key = activeRing.next();
-    if (!key && activeRing.size > 0) {
-      const err = new Error(`${id}: all API keys are in cooldown (last error < ${cooldownMs}ms ago) — provider temporarily unavailable`);
-      err._t = { attempts: [], waitMs: 0, totalMs: Math.round(performance.now() - t0), cooldownMs };
-      throw err;
-    }
-    for (let attempt = 0; ; attempt++) {
-      const t = performance.now();
-      let result;
-      try {
-        result = await attemptOnce(body, key, opts);
-      } catch (err) {
-        result = err;
-      }
-      attempts.push({ attempt, type: result instanceof Error ? "network" : `http${result?.status}`, ms: Math.round(performance.now() - t) });
-      if (result instanceof Error) {
-        const entry = retry?.network;
-        if (entry && attempt < entry.attempts) { await sleep(entry.delayMs); waitMs += entry.delayMs; continue; }
-        activeRing.onError(key);
-        result._t = { attempts, waitMs, totalMs: Math.round(performance.now() - t0) };
-        throw result;
-      }
-      const entry = retry?.[result.status];
-      if (entry && attempt < entry.attempts) { await sleep(entry.delayMs); waitMs += entry.delayMs; continue; }
-      if (result.status === 401 || result.status === 403 || result.status === 429 || result.status >= 500) activeRing.onError(key);
-      try { result._t = { attempts, waitMs, totalMs: Math.round(performance.now() - t0) }; } catch {}
-      return result;
-    }
-  }
-
-  return { url, attemptOnce, run };
-}
+// 通用（带 key）供应商的 responses 通道：muse-spark* 这类模型上游只挂 /responses，
+// 打同 host 的 /chat/completions 会 503 "Endpoint is unavailable"（ocgo 实测 2026-09-22）。
+// 判定单一来源 = isResponsesModel（models.dev 模型级 provider.npm + muse-spark 前缀兜底），
+// 与 GET /v1/models 的 capabilities.upstreamApi 同源，新模型无需改码。
+//
+// 契约：run() 的出参恒为 **chat 形状**（SSE 或 JSON），调用方无需再判分支。
+//   · 流式优先 @ai-sdk/openai 的 responses 适配器（含加密思考往返，ADR-0017 复用）；
+//   · SDK 不可用 / 非流式 → 原生 fetch + chatToResponsesBody 正转换 + 反向整形。
+// key 轮换、退避重试、_t 计时与 createChatRunner（base.js）同语义，避免两条通道行为漂移。
+import { chatToResponsesBody, toChatResponse, reshapeResponsesSse } from "../upstream-responses.js";
+import { attemptOnceResponsesSdk } from "../upstream-engine/sdk/responses.js";
+import { ENGINE_MARKER } from "../upstream-engine/sdk/chat.js";
+import { sleep } from "./base.js";
+
+/** 供应商 id → env 片段（`my-api` → `MY_API`），与 providerKeyEnv 同规则 */
+export function envSlug(id) {
+  return String(id || "").toUpperCase().replace(/[^A-Z0-9]/g, "_");
+}
+
+/**
+ * responses 端点路径：显式入参 > env `MSLXDFF_<ID>_RESPONSES_PATH` > 缺省 `/responses`。
+ * 归一为以 `/` 开头（joinUrl 会再拼 baseUrl）。
+ */
+export function resolveResponsesPath(id, responsesPath) {
+  const raw = responsesPath || process.env[`MSLXDFF_${envSlug(id)}_RESPONSES_PATH`] || "";
+  const s = String(raw).trim();
+  if (!s) return "/responses";
+  return s.startsWith("/") ? s : `/${s}`;
+}
+
+/**
+ * 建 responses 通道。
+ * @param {object} o
+ * @param {string} o.id 供应商 id（用于报错文案与 providerName）
+ * @param {string} o.url 已拼好的 responses 绝对 URL
+ * @param {number} o.connectTimeoutMs 单次尝试超时
+ * @param {object} o.retry 重试表（与 createChatRunner 同形状：network/429/50x → {attempts, delayMs}）
+ * @param {number} o.cooldownMs 全 key 冷却时的人话报错文案用
+ * @param {boolean} o.sdkEnabled 是否允许走 AI SDK responses 适配器（流式路径）
+ * @param {Function} o.buildHeaders (body, key, opts) → headers
+ * @param {Function} o.fetchImpl 注入的 fetch（连接池/测试接缝）
+ * @param {object} [o.dispatcher] undici Agent
+ */
+export function createResponsesChannel({
+  id,
+  url,
+  connectTimeoutMs = 30_000,
+  retry = {},
+  cooldownMs = 30_000,
+  sdkEnabled = true,
+  buildHeaders,
+  fetchImpl,
+  dispatcher = null,
+} = {}) {
+  async function attemptOnce(body, key, opts) {
+    const headers = buildHeaders ? buildHeaders(body, key, opts) : {};
+    const wantsStream = body?.stream !== false;
+    if (sdkEnabled && wantsStream) {
+      try {
+        const r = await attemptOnceResponsesSdk({
+          url, body, headers, providerName: id, marker: ENGINE_MARKER, fetchImpl,
+        });
+        if (r) return r; // 适配器已产出 chat SSE
+      } catch (e) {
+        if (!e || (!e._sdkLoadFailed && !e._sdkUnsupported)) throw e;
+        // SDK 装载失败 / URL 异形 → 落原生兜底（保底不变差）
+      }
+    }
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(new Error(`${id} timed out after ${connectTimeoutMs}ms`)), connectTimeoutMs);
+    try {
+      const fetchOpts = { method: "POST", headers, body: JSON.stringify(chatToResponsesBody(body)), signal: controller.signal };
+      if (dispatcher) fetchOpts.dispatcher = dispatcher;
+      const res = await fetchImpl(url, fetchOpts);
+      if (!res.ok) return res;
+      if (wantsStream) return reshapeResponsesSse(res, body?.model);
+      const json = await res.json().catch(() => null);
+      return json ? toChatResponse(res, json) : res;
+    } finally {
+      clearTimeout(timer);
+    }
+  }
+
+  /** key 轮换 + 退避重试 + _t 计时；语义对齐 base.js 的 createChatRunner.runChat */
+  async function run(body, activeRing, sourceKey, opts) {
+    const t0 = performance.now();
+    const attempts = [];
+    let waitMs = 0;
+    const key = activeRing.next();
+    if (!key && activeRing.size > 0) {
+      const err = new Error(`${id}: all API keys are in cooldown (last error < ${cooldownMs}ms ago) — provider temporarily unavailable`);
+      err._t = { attempts: [], waitMs: 0, totalMs: Math.round(performance.now() - t0), cooldownMs };
+      throw err;
+    }
+    for (let attempt = 0; ; attempt++) {
+      const t = performance.now();
+      let result;
+      try {
+        result = await attemptOnce(body, key, opts);
+      } catch (err) {
+        result = err;
+      }
+      attempts.push({ attempt, type: result instanceof Error ? "network" : `http${result?.status}`, ms: Math.round(performance.now() - t) });
+      if (result instanceof Error) {
+        const entry = retry?.network;
+        if (entry && attempt < entry.attempts) { await sleep(entry.delayMs); waitMs += entry.delayMs; continue; }
+        activeRing.onError(key);
+        result._t = { attempts, waitMs, totalMs: Math.round(performance.now() - t0) };
+        throw result;
+      }
+      const entry = retry?.[result.status];
+      if (entry && attempt < entry.attempts) { await sleep(entry.delayMs); waitMs += entry.delayMs; continue; }
+      if (result.status === 401 || result.status === 403 || result.status === 429 || result.status >= 500) activeRing.onError(key);
+      try { result._t = { attempts, waitMs, totalMs: Math.round(performance.now() - t0) }; } catch {}
+      return result;
+    }
+  }
+
+  return { url, attemptOnce, run };
+}

package/src/providers/share-keys.js

@@ -1,54 +1,54 @@
-// ADR-0019：借道 = 用你的 key —— 转发时自动借出，无开关、无白名单（组内互信是前提）。
-// header: `x-mslxdff-share-keys: <providerId>=<key1>,<key2>;<providerId2>=<key3>`
-import { splitModelId, DEFAULT_PROVIDER } from "./model-id.js";
-import { classifyProvider } from "./classify.js";
-import { loadProviderKeys, listProviderIdsWithKeys } from "../state.js";
-import { isRefreshToken } from "./cline/headers.js";
-
-export const SHARE_KEYS_HEADER = "x-mslxdff-share-keys";
-
-// 硬排除：refresh-token 型凭据（对端刷新会轮换 token，与本机互相踢下线）——正确性问题。
-// 二层保护：① 按 id 硬排除（cline 恒 local-only）；② 按凭据形状兜底（见 shareableProviderIds）。
-// traework/workbuddy/zcode 同为本机账号绑定型（auths/*.json + uid；zcode 还带 per-account deviceMid），恒不借出。
-const NEVER_SHARE_IDS = new Set(["cline", "codearts", "traework", "qoder", "zcode", "raccoon"]);
-// 本节点应 cast key 到出站转发的供应商 id 集合 = 所有「本机有 key」的供应商，
-// 减去：opencode（无 key 且恒排除）、local-only（workbuddy/cline系，本就不走组员）、刷新型凭据。
-export function shareableProviderIds({ file } = {}) {
-  const ids = [];
-  for (const id of listProviderIdsWithKeys({ file })) {
-    if (id === DEFAULT_PROVIDER) continue;
-    if (NEVER_SHARE_IDS.has(id)) continue;
-    try { if (classifyProvider(id) === "local-only") continue; } catch {}
-    // 形状兜底：id 白名单可能有残留（迁移未跑/回滚/手改 config），只要该供应商现有 key 里
-    // 有一个是 refreshToken 形态，就一律不外借——refreshToken 借出去会被对端刷新轮换，双向踢下线。
-    try { if (loadProviderKeys(id, { file }).some((k) => isRefreshToken(k, id))) continue; } catch {}
-    ids.push(id);
-  }
-  return ids;
-}
-
-// 组装 share header 值（只含 shareable 且当前请求命中前缀的供应商）。
-// splitModelId 用 shareableProviderIds() 作为已知集合：命中 → 该供应商在此节点 shareable；
-// 未命中或裸 id（默认供应商 opencode）→ 不附带。返回 null 表示本请求不需要附带。
-export function buildShareKeysHeader(model, { file } = {}) {
-  const split = splitModelId(model, shareableProviderIds({ file }));
-  if (!split.provider || split.provider === DEFAULT_PROVIDER) return null;
-  const keys = loadProviderKeys(split.provider, { file });
-  if (!keys.length) return null;
-  return `${split.provider}=${keys.join(",")}`;
-}
-
-// 组员侧：解析 share header 为 { providerId -> keys[] }
-// 防御：默认供应商（opencode）即使出现在 header 一律忽略——它不该有共享 key。
-export function parseShareKeysHeader(value) {
-  const out = {};
-  if (!value) return out;
-  for (const seg of String(value).split(";")) {
-    const [provider, keysPart] = seg.split("=");
-    const id = String(provider || "").trim();
-    if (!id || id === DEFAULT_PROVIDER || !keysPart) continue;
-    const keys = String(keysPart).split(",").map((k) => k.trim()).filter(Boolean);
-    if (keys.length) out[id] = keys;
-  }
-  return out;
+// ADR-0019：借道 = 用你的 key —— 转发时自动借出，无开关、无白名单（组内互信是前提）。
+// header: `x-mslxdff-share-keys: <providerId>=<key1>,<key2>;<providerId2>=<key3>`
+import { splitModelId, DEFAULT_PROVIDER } from "./model-id.js";
+import { classifyProvider } from "./classify.js";
+import { loadProviderKeys, listProviderIdsWithKeys } from "../state.js";
+import { isRefreshToken } from "./cline/headers.js";
+
+export const SHARE_KEYS_HEADER = "x-mslxdff-share-keys";
+
+// 硬排除：refresh-token 型凭据（对端刷新会轮换 token，与本机互相踢下线）——正确性问题。
+// 二层保护：① 按 id 硬排除（cline 恒 local-only）；② 按凭据形状兜底（见 shareableProviderIds）。
+// traework/workbuddy/zcode 同为本机账号绑定型（auths/*.json + uid；zcode 还带 per-account deviceMid），恒不借出。
+const NEVER_SHARE_IDS = new Set(["cline", "codearts", "traework", "qoder", "zcode", "raccoon"]);
+// 本节点应 cast key 到出站转发的供应商 id 集合 = 所有「本机有 key」的供应商，
+// 减去：opencode（无 key 且恒排除）、local-only（workbuddy/cline系，本就不走组员）、刷新型凭据。
+export function shareableProviderIds({ file } = {}) {
+  const ids = [];
+  for (const id of listProviderIdsWithKeys({ file })) {
+    if (id === DEFAULT_PROVIDER) continue;
+    if (NEVER_SHARE_IDS.has(id)) continue;
+    try { if (classifyProvider(id) === "local-only") continue; } catch {}
+    // 形状兜底：id 白名单可能有残留（迁移未跑/回滚/手改 config），只要该供应商现有 key 里
+    // 有一个是 refreshToken 形态，就一律不外借——refreshToken 借出去会被对端刷新轮换，双向踢下线。
+    try { if (loadProviderKeys(id, { file }).some((k) => isRefreshToken(k, id))) continue; } catch {}
+    ids.push(id);
+  }
+  return ids;
+}
+
+// 组装 share header 值（只含 shareable 且当前请求命中前缀的供应商）。
+// splitModelId 用 shareableProviderIds() 作为已知集合：命中 → 该供应商在此节点 shareable；
+// 未命中或裸 id（默认供应商 opencode）→ 不附带。返回 null 表示本请求不需要附带。
+export function buildShareKeysHeader(model, { file } = {}) {
+  const split = splitModelId(model, shareableProviderIds({ file }));
+  if (!split.provider || split.provider === DEFAULT_PROVIDER) return null;
+  const keys = loadProviderKeys(split.provider, { file });
+  if (!keys.length) return null;
+  return `${split.provider}=${keys.join(",")}`;
+}
+
+// 组员侧：解析 share header 为 { providerId -> keys[] }
+// 防御：默认供应商（opencode）即使出现在 header 一律忽略——它不该有共享 key。
+export function parseShareKeysHeader(value) {
+  const out = {};
+  if (!value) return out;
+  for (const seg of String(value).split(";")) {
+    const [provider, keysPart] = seg.split("=");
+    const id = String(provider || "").trim();
+    if (!id || id === DEFAULT_PROVIDER || !keysPart) continue;
+    const keys = String(keysPart).split(",").map((k) => k.trim()).filter(Boolean);
+    if (keys.length) out[id] = keys;
+  }
+  return out;
 }