mslxdff 0.1.160 → 0.1.161

package/src/chat/orchestrator.js

@@ -1,234 +1,234 @@
-import { performance as nodePerf } from "node:perf_hooks";
-
-const ORIG_PREFERRED = "mimo-v2.5-free";
-const ORIG_FALLBACK = "big-pickle";
-
-/**
- * 编排深模块：mimo → pickle → gateway 三级降级 + 800ms 对冲
- * 注入化：chatOnce / chatViaGateway / cooling / config / env / performance
- */
-export function createOrchestrator({
-  chatOnce,
-  chatViaGateway,
-  cooling,
-  config = {},
-  env = process.env,
-  performance: perf = nodePerf,
-  chatWithFallbackImpl,
-} = {}) {
-  const CHAT_PREFERRED = config.CHAT_PREFERRED || ORIG_PREFERRED;
-  const CHAT_FALLBACK = config.CHAT_FALLBACK || ORIG_FALLBACK;
-  const CHAT_GATEWAY_TIMEOUT_MS = config.CHAT_GATEWAY_TIMEOUT_MS || 25000;
-  const CHAT_COOLDOWN_MS = 10 * 60 * 1000;
-  const CHAT_SLOW_COOLDOWN_MS = 10 * 60 * 1000;
-
-  const _chatOnce = chatOnce || (async () => ({ ok: false, error: "no chatOnce", status: 500 }));
-  const _gateway = chatViaGateway || (async () => ({ ok: false, error: "no gateway", status: 500 }));
-  const _cooling = cooling || {
-    isCooling: async () => false,
-    recordError: async () => {},
-    recordOk: async () => {},
-  };
-
-  async function isCoolingAsync(id) {
-    try { return await _cooling.isCooling(id); } catch { return false; }
-  }
-  async function recordChatError(id, status, opts) {
-    try { await _cooling.recordError(id, status, opts); } catch {}
-  }
-  async function recordChatOk(id, latencyMs) {
-    try { await _cooling.recordOk(id, latencyMs); } catch {}
-  }
-
-  async function safeChatOnce(opts, model) {
-    try {
-      const r = await _chatOnce({ ...opts, model });
-      return r;
-    } catch (err) {
-      const msg = String(err?.message || err).slice(0, 800);
-      const status = err?._t ? 502 : 502;
-      return { ok: false, error: msg, status, _thrown: err };
-    }
-  }
-
-  async function chatWithFallback(opts) {
-    if (chatWithFallbackImpl) return chatWithFallbackImpl(opts);
-    // 严格模式：显式指定了具体模型（非 auto/空）→ 只用该模型，失败即报错，绝不降级到其他模型
-    const pinned = String(opts?.model || "").trim();
-    if (pinned && pinned.toLowerCase() !== "auto") {
-      const r = await safeChatOnce(opts, pinned);
-      if (r.ok) return { ...r, model: pinned };
-      return { ok: false, error: `指定模型 ${pinned} 失败：${r.error}（已锁定不自动换模型；如需自动择优请输入 /model auto）`, status: r.status || 502, pinnedModel: pinned };
-    }
-    const TRACE = env.MSLXDFF_CHAT_TRACE !== "0";
-    const HEDGE_MS = (() => {
-      const v = Number(env.MSLXDFF_HEDGE_DELAY_MS);
-      return Number.isInteger(v) && v >= 0 ? v : 800;
-    })();
-    const t0 = TRACE ? perf.now() : 0;
-
-    const firstCooling = await isCoolingAsync(CHAT_PREFERRED);
-    let first;
-    let firstMs = 0;
-    if (firstCooling) {
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 跳过（冷却 10min）· 额度用完，准备走网关 auto（将尝试其他可用模型）\x1b[0m`);
-      first = { ok: false, error: "skip cooling", status: 429 };
-    } else {
-      const t = perf.now();
-      first = await safeChatOnce(opts, CHAT_PREFERRED);
-      firstMs = Math.round(perf.now() - t);
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} ${first.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t0)}ms${first.ok ? "" : ` · ${String(first.error).slice(0, 80)}`}\x1b[0m`);
-      if (first.ok) {
-        await recordChatOk(CHAT_PREFERRED, firstMs);
-        return { ...first, model: CHAT_PREFERRED };
-      } else {
-        const slow = firstMs > 20000;
-        await recordChatError(CHAT_PREFERRED, first.status, { slow, latencyMs: firstMs });
-      }
-    }
-
-    if (firstCooling) {
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 冷却中（${CHAT_COOLDOWN_MS / 60000}min）· 直接走网关 auto（:8989），按 auto 择优尝试其他可用模型，跳过 ${CHAT_FALLBACK} 直连\x1b[0m`);
-      const t2 = TRACE ? perf.now() : 0;
-      const third = await _gateway(opts);
-      if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t2)}ms · 总 ${Math.round(perf.now() - t0)}ms (gateway-fallback)\x1b[0m`);
-      if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: "skip big-pickle (mimo cooling)", viaGateway: true };
-      const isGwDown = third.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(third.error || ""));
-      if (isGwDown) {
-        return { ok: false, error: `${CHAT_PREFERRED} 冷却跳过（额度用完 10min）；${third.error} —— 本次 auto 未能尝试任何其他模型，因网关未就绪（3ms 失败即证明未触达上游，请先启动网关再重试）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
-      }
-      return { ok: false, error: `${CHAT_PREFERRED} 冷却跳过（额度用完）；网关 auto 已尝试其他可用模型但均失败：${third.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: third.status || 429 };
-    }
-
-    const secondCooling = await isCoolingAsync(CHAT_FALLBACK);
-    const doGateway = () => _gateway(opts);
-    const doSecond = async () => {
-      const t = perf.now();
-      const r = await safeChatOnce(opts, CHAT_FALLBACK);
-      const ms = Math.round(perf.now() - t);
-      if (r.ok) await recordChatOk(CHAT_FALLBACK, ms);
-      else await recordChatError(CHAT_FALLBACK, r.status, { slow: ms > 20000, latencyMs: ms });
-      return { res: r, ms };
-    };
-
-    if (secondCooling) {
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} 跳过（冷却中）· 直接走网关 auto（将尝试其他可用模型）\x1b[0m`);
-      const t2 = perf.now();
-      const third = await doGateway();
-      if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t2)}ms · 总 ${Math.round(perf.now() - t0)}ms (gateway-fallback)\x1b[0m`);
-      if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: "skip cooling", viaGateway: true };
-      const isGwDown2 = third.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(third.error || ""));
-      if (isGwDown2) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 冷却跳过；${third.error} —— auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
-      return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 冷却跳过；网关 auto 已尝试其他模型但均失败：${third.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: third.status || first.status };
-    }
-
-    if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 失败，${CHAT_FALLBACK} + gateway 对冲中（${HEDGE_MS}ms）\x1b[0m`);
-    const t1 = TRACE ? perf.now() : 0;
-    let secondRes = null;
-    let secondMs = 0;
-    let gatewayRes = null;
-
-    const secondPromise = doSecond().then(({ res, ms }) => {
-      secondRes = res; secondMs = ms;
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} ${res.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t1)}ms · 总 ${Math.round(perf.now() - t0)}ms (hedge)\x1b[0m`);
-      return res;
-    });
-
-    let gatewayPromise = null;
-    const gatewayDelay = HEDGE_MS > 0 ? HEDGE_MS : 0;
-    if (gatewayDelay > 0) {
-      gatewayPromise = new Promise((resolve) => {
-        setTimeout(async () => {
-          const r = await doGateway();
-          gatewayRes = r;
-          resolve(r);
-        }, gatewayDelay);
-      });
-    } else {
-      gatewayPromise = doGateway().then((r) => { gatewayRes = r; return r; });
-    }
-
-    const raceFirstOk = async () => {
-      const secondOrTimeout = await Promise.race([
-        secondPromise.then((r) => ({ kind: "second", r })),
-        new Promise((resolve) => setTimeout(() => resolve({ kind: "timeout" }), gatewayDelay)),
-      ]);
-      if (secondOrTimeout.kind === "second" && secondOrTimeout.r?.ok) {
-        return { ok: true, res: secondOrTimeout.r, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
-      }
-      if (!gatewayPromise || secondOrTimeout.kind === "timeout") {
-        const immediate = doGateway().then((r) => { gatewayRes = r; return r; });
-        if (gatewayPromise) {
-          gatewayRes = await Promise.race([gatewayPromise, immediate]);
-        } else {
-          gatewayRes = await immediate;
-        }
-      } else {
-        gatewayRes = await gatewayPromise;
-      }
-      if (gatewayRes?.ok) return { ok: true, res: gatewayRes, model: gatewayRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: secondRes?.error, viaGateway: true };
-      if (secondRes?.ok) return { ok: true, res: secondRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
-      return { ok: false, gatewayRes, secondRes };
-    };
-
-    if (gatewayDelay === 0) {
-      const [sRes, gRes] = await Promise.all([
-        secondPromise.catch((e) => ({ ok: false, error: String(e), status: 502 })),
-        doGateway().catch((e) => ({ ok: false, error: String(e), status: 502 })),
-      ]);
-      secondRes = sRes; gatewayRes = gRes;
-      if (sRes?.ok) return { ...sRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
-      if (gRes?.ok) return { ...gRes, model: gRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: sRes?.error, viaGateway: true };
-      const gwDown0 = gRes?.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(gRes?.error || ""));
-      if (gwDown0) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${sRes?.error}；${gRes?.error} —— gateway 侧 auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
-      return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${sRes?.error}；网关 auto 已尝试其他模型但均失败：${gRes?.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: gRes?.status || sRes?.status || first.status };
-    }
-
-    const raced = await raceFirstOk();
-    if (raced.ok) {
-      const r = raced.res;
-      return { ...r, model: raced.model, fallback: raced.fallback, fallbackGateway: raced.fallbackGateway, firstError: raced.firstError, secondError: raced.secondError, viaGateway: raced.viaGateway };
-    }
-    if (!secondRes) {
-      try { secondRes = await secondPromise; } catch (e) { secondRes = { ok: false, error: String(e), status: 502 }; }
-    }
-    if (secondRes?.ok) return { ...secondRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
-    if (!gatewayRes) {
-      try { gatewayRes = await (gatewayPromise || doGateway()); } catch (e) { gatewayRes = { ok: false, error: String(e), status: 502 }; }
-    }
-    if (gatewayRes?.ok) return { ...gatewayRes, model: gatewayRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: secondRes?.error, viaGateway: true };
-    const gwDown = gatewayRes?.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(gatewayRes?.error || ""));
-    if (gwDown) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${secondRes?.error}；${gatewayRes?.error} —— gateway 侧 auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
-    return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${secondRes?.error}；网关 auto 已尝试其他模型但均失败：${gatewayRes?.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: gatewayRes?.status || secondRes?.status || first.status };
-  }
-
-  async function summarizeHistory(messages) {
-    const prompt = [
-      { role: "system", content: "你是对话压缩助手，把以下历史对话压缩成 800 字以内的中文摘要，保留关键操作与结果、用户的偏好与待办、模型设置与群组操作及时间线，不要遗漏重要细节。" },
-      { role: "user", content: messages.map((m) => `${m.role}: ${m.content || JSON.stringify(m.tool_calls || "")}`).join("\n").slice(0, 90000) },
-    ];
-    const r = await chatWithFallback({ messages: prompt });
-    if (!r.ok) return null;
-    const txt = String(r.message?.content || "").trim();
-    return txt ? `【历史摘要】${txt}` : null;
-  }
-
-  // 兼容测试的额外覆盖参数
-  if (chatWithFallbackImpl) {
-    const orig = chatWithFallback;
-    // 允许测试覆盖 summarize 内部的 chatWithFallback
-    const wrapped = async (opts) => chatWithFallbackImpl(opts);
-    return { chatWithFallback: wrapped, summarizeHistory: async (msgs) => {
-      const prompt = [
-        { role: "system", content: "你是对话压缩助手，把以下历史对话压缩成 800 字以内的中文摘要，保留关键操作与结果、用户的偏好与待办、模型设置与群组操作及时间线，不要遗漏重要细节。" },
-        { role: "user", content: msgs.map((m) => `${m.role}: ${m.content || JSON.stringify(m.tool_calls || "")}`).join("\n").slice(0, 90000) },
-      ];
-      const r = await wrapped({ messages: prompt });
-      if (!r.ok) return null;
-      const txt = String(r.message?.content || "").trim();
-      return txt ? `【历史摘要】${txt}` : null;
-    } };
-  }
-
-  return { chatWithFallback, summarizeHistory };
-}
+import { performance as nodePerf } from "node:perf_hooks";
+
+const ORIG_PREFERRED = "mimo-v2.5-free";
+const ORIG_FALLBACK = "big-pickle";
+
+/**
+ * 编排深模块：mimo → pickle → gateway 三级降级 + 800ms 对冲
+ * 注入化：chatOnce / chatViaGateway / cooling / config / env / performance
+ */
+export function createOrchestrator({
+  chatOnce,
+  chatViaGateway,
+  cooling,
+  config = {},
+  env = process.env,
+  performance: perf = nodePerf,
+  chatWithFallbackImpl,
+} = {}) {
+  const CHAT_PREFERRED = config.CHAT_PREFERRED || ORIG_PREFERRED;
+  const CHAT_FALLBACK = config.CHAT_FALLBACK || ORIG_FALLBACK;
+  const CHAT_GATEWAY_TIMEOUT_MS = config.CHAT_GATEWAY_TIMEOUT_MS || 25000;
+  const CHAT_COOLDOWN_MS = 10 * 60 * 1000;
+  const CHAT_SLOW_COOLDOWN_MS = 10 * 60 * 1000;
+
+  const _chatOnce = chatOnce || (async () => ({ ok: false, error: "no chatOnce", status: 500 }));
+  const _gateway = chatViaGateway || (async () => ({ ok: false, error: "no gateway", status: 500 }));
+  const _cooling = cooling || {
+    isCooling: async () => false,
+    recordError: async () => {},
+    recordOk: async () => {},
+  };
+
+  async function isCoolingAsync(id) {
+    try { return await _cooling.isCooling(id); } catch { return false; }
+  }
+  async function recordChatError(id, status, opts) {
+    try { await _cooling.recordError(id, status, opts); } catch {}
+  }
+  async function recordChatOk(id, latencyMs) {
+    try { await _cooling.recordOk(id, latencyMs); } catch {}
+  }
+
+  async function safeChatOnce(opts, model) {
+    try {
+      const r = await _chatOnce({ ...opts, model });
+      return r;
+    } catch (err) {
+      const msg = String(err?.message || err).slice(0, 800);
+      const status = err?._t ? 502 : 502;
+      return { ok: false, error: msg, status, _thrown: err };
+    }
+  }
+
+  async function chatWithFallback(opts) {
+    if (chatWithFallbackImpl) return chatWithFallbackImpl(opts);
+    // 严格模式：显式指定了具体模型（非 auto/空）→ 只用该模型，失败即报错，绝不降级到其他模型
+    const pinned = String(opts?.model || "").trim();
+    if (pinned && pinned.toLowerCase() !== "auto") {
+      const r = await safeChatOnce(opts, pinned);
+      if (r.ok) return { ...r, model: pinned };
+      return { ok: false, error: `指定模型 ${pinned} 失败：${r.error}（已锁定不自动换模型；如需自动择优请输入 /model auto）`, status: r.status || 502, pinnedModel: pinned };
+    }
+    const TRACE = env.MSLXDFF_CHAT_TRACE !== "0";
+    const HEDGE_MS = (() => {
+      const v = Number(env.MSLXDFF_HEDGE_DELAY_MS);
+      return Number.isInteger(v) && v >= 0 ? v : 800;
+    })();
+    const t0 = TRACE ? perf.now() : 0;
+
+    const firstCooling = await isCoolingAsync(CHAT_PREFERRED);
+    let first;
+    let firstMs = 0;
+    if (firstCooling) {
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 跳过（冷却 10min）· 额度用完，准备走网关 auto（将尝试其他可用模型）\x1b[0m`);
+      first = { ok: false, error: "skip cooling", status: 429 };
+    } else {
+      const t = perf.now();
+      first = await safeChatOnce(opts, CHAT_PREFERRED);
+      firstMs = Math.round(perf.now() - t);
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} ${first.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t0)}ms${first.ok ? "" : ` · ${String(first.error).slice(0, 80)}`}\x1b[0m`);
+      if (first.ok) {
+        await recordChatOk(CHAT_PREFERRED, firstMs);
+        return { ...first, model: CHAT_PREFERRED };
+      } else {
+        const slow = firstMs > 20000;
+        await recordChatError(CHAT_PREFERRED, first.status, { slow, latencyMs: firstMs });
+      }
+    }
+
+    if (firstCooling) {
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 冷却中（${CHAT_COOLDOWN_MS / 60000}min）· 直接走网关 auto（:8989），按 auto 择优尝试其他可用模型，跳过 ${CHAT_FALLBACK} 直连\x1b[0m`);
+      const t2 = TRACE ? perf.now() : 0;
+      const third = await _gateway(opts);
+      if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t2)}ms · 总 ${Math.round(perf.now() - t0)}ms (gateway-fallback)\x1b[0m`);
+      if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: "skip big-pickle (mimo cooling)", viaGateway: true };
+      const isGwDown = third.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(third.error || ""));
+      if (isGwDown) {
+        return { ok: false, error: `${CHAT_PREFERRED} 冷却跳过（额度用完 10min）；${third.error} —— 本次 auto 未能尝试任何其他模型，因网关未就绪（3ms 失败即证明未触达上游，请先启动网关再重试）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+      }
+      return { ok: false, error: `${CHAT_PREFERRED} 冷却跳过（额度用完）；网关 auto 已尝试其他可用模型但均失败：${third.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: third.status || 429 };
+    }
+
+    const secondCooling = await isCoolingAsync(CHAT_FALLBACK);
+    const doGateway = () => _gateway(opts);
+    const doSecond = async () => {
+      const t = perf.now();
+      const r = await safeChatOnce(opts, CHAT_FALLBACK);
+      const ms = Math.round(perf.now() - t);
+      if (r.ok) await recordChatOk(CHAT_FALLBACK, ms);
+      else await recordChatError(CHAT_FALLBACK, r.status, { slow: ms > 20000, latencyMs: ms });
+      return { res: r, ms };
+    };
+
+    if (secondCooling) {
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} 跳过（冷却中）· 直接走网关 auto（将尝试其他可用模型）\x1b[0m`);
+      const t2 = perf.now();
+      const third = await doGateway();
+      if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t2)}ms · 总 ${Math.round(perf.now() - t0)}ms (gateway-fallback)\x1b[0m`);
+      if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: "skip cooling", viaGateway: true };
+      const isGwDown2 = third.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(third.error || ""));
+      if (isGwDown2) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 冷却跳过；${third.error} —— auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+      return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 冷却跳过；网关 auto 已尝试其他模型但均失败：${third.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: third.status || first.status };
+    }
+
+    if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 失败，${CHAT_FALLBACK} + gateway 对冲中（${HEDGE_MS}ms）\x1b[0m`);
+    const t1 = TRACE ? perf.now() : 0;
+    let secondRes = null;
+    let secondMs = 0;
+    let gatewayRes = null;
+
+    const secondPromise = doSecond().then(({ res, ms }) => {
+      secondRes = res; secondMs = ms;
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} ${res.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t1)}ms · 总 ${Math.round(perf.now() - t0)}ms (hedge)\x1b[0m`);
+      return res;
+    });
+
+    let gatewayPromise = null;
+    const gatewayDelay = HEDGE_MS > 0 ? HEDGE_MS : 0;
+    if (gatewayDelay > 0) {
+      gatewayPromise = new Promise((resolve) => {
+        setTimeout(async () => {
+          const r = await doGateway();
+          gatewayRes = r;
+          resolve(r);
+        }, gatewayDelay);
+      });
+    } else {
+      gatewayPromise = doGateway().then((r) => { gatewayRes = r; return r; });
+    }
+
+    const raceFirstOk = async () => {
+      const secondOrTimeout = await Promise.race([
+        secondPromise.then((r) => ({ kind: "second", r })),
+        new Promise((resolve) => setTimeout(() => resolve({ kind: "timeout" }), gatewayDelay)),
+      ]);
+      if (secondOrTimeout.kind === "second" && secondOrTimeout.r?.ok) {
+        return { ok: true, res: secondOrTimeout.r, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
+      }
+      if (!gatewayPromise || secondOrTimeout.kind === "timeout") {
+        const immediate = doGateway().then((r) => { gatewayRes = r; return r; });
+        if (gatewayPromise) {
+          gatewayRes = await Promise.race([gatewayPromise, immediate]);
+        } else {
+          gatewayRes = await immediate;
+        }
+      } else {
+        gatewayRes = await gatewayPromise;
+      }
+      if (gatewayRes?.ok) return { ok: true, res: gatewayRes, model: gatewayRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: secondRes?.error, viaGateway: true };
+      if (secondRes?.ok) return { ok: true, res: secondRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
+      return { ok: false, gatewayRes, secondRes };
+    };
+
+    if (gatewayDelay === 0) {
+      const [sRes, gRes] = await Promise.all([
+        secondPromise.catch((e) => ({ ok: false, error: String(e), status: 502 })),
+        doGateway().catch((e) => ({ ok: false, error: String(e), status: 502 })),
+      ]);
+      secondRes = sRes; gatewayRes = gRes;
+      if (sRes?.ok) return { ...sRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
+      if (gRes?.ok) return { ...gRes, model: gRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: sRes?.error, viaGateway: true };
+      const gwDown0 = gRes?.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(gRes?.error || ""));
+      if (gwDown0) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${sRes?.error}；${gRes?.error} —— gateway 侧 auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+      return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${sRes?.error}；网关 auto 已尝试其他模型但均失败：${gRes?.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: gRes?.status || sRes?.status || first.status };
+    }
+
+    const raced = await raceFirstOk();
+    if (raced.ok) {
+      const r = raced.res;
+      return { ...r, model: raced.model, fallback: raced.fallback, fallbackGateway: raced.fallbackGateway, firstError: raced.firstError, secondError: raced.secondError, viaGateway: raced.viaGateway };
+    }
+    if (!secondRes) {
+      try { secondRes = await secondPromise; } catch (e) { secondRes = { ok: false, error: String(e), status: 502 }; }
+    }
+    if (secondRes?.ok) return { ...secondRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
+    if (!gatewayRes) {
+      try { gatewayRes = await (gatewayPromise || doGateway()); } catch (e) { gatewayRes = { ok: false, error: String(e), status: 502 }; }
+    }
+    if (gatewayRes?.ok) return { ...gatewayRes, model: gatewayRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: secondRes?.error, viaGateway: true };
+    const gwDown = gatewayRes?.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(gatewayRes?.error || ""));
+    if (gwDown) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${secondRes?.error}；${gatewayRes?.error} —— gateway 侧 auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+    return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${secondRes?.error}；网关 auto 已尝试其他模型但均失败：${gatewayRes?.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: gatewayRes?.status || secondRes?.status || first.status };
+  }
+
+  async function summarizeHistory(messages) {
+    const prompt = [
+      { role: "system", content: "你是对话压缩助手，把以下历史对话压缩成 800 字以内的中文摘要，保留关键操作与结果、用户的偏好与待办、模型设置与群组操作及时间线，不要遗漏重要细节。" },
+      { role: "user", content: messages.map((m) => `${m.role}: ${m.content || JSON.stringify(m.tool_calls || "")}`).join("\n").slice(0, 90000) },
+    ];
+    const r = await chatWithFallback({ messages: prompt });
+    if (!r.ok) return null;
+    const txt = String(r.message?.content || "").trim();
+    return txt ? `【历史摘要】${txt}` : null;
+  }
+
+  // 兼容测试的额外覆盖参数
+  if (chatWithFallbackImpl) {
+    const orig = chatWithFallback;
+    // 允许测试覆盖 summarize 内部的 chatWithFallback
+    const wrapped = async (opts) => chatWithFallbackImpl(opts);
+    return { chatWithFallback: wrapped, summarizeHistory: async (msgs) => {
+      const prompt = [
+        { role: "system", content: "你是对话压缩助手，把以下历史对话压缩成 800 字以内的中文摘要，保留关键操作与结果、用户的偏好与待办、模型设置与群组操作及时间线，不要遗漏重要细节。" },
+        { role: "user", content: msgs.map((m) => `${m.role}: ${m.content || JSON.stringify(m.tool_calls || "")}`).join("\n").slice(0, 90000) },
+      ];
+      const r = await wrapped({ messages: prompt });
+      if (!r.ok) return null;
+      const txt = String(r.message?.content || "").trim();
+      return txt ? `【历史摘要】${txt}` : null;
+    } };
+  }
+
+  return { chatWithFallback, summarizeHistory };
+}

package/src/chat/prompt.js

@@ -1,70 +1,70 @@
-import { readFileSync, existsSync } from "node:fs";
-import { join, dirname } from "node:path";
-import { fileURLToPath } from "node:url";
-import { logDir } from "../logs.js";
-import { nowShanghaiYMDHM } from "../time.js";
-
-const pkgRoot = join(dirname(fileURLToPath(import.meta.url)), "../..");
-
-function readMini() {
-  try {
-    const p = join(pkgRoot, "docs/cli_help_mini.md");
-    return readFileSync(p, "utf8");
-  } catch { return ""; }
-}
-
-function readModels() {
-  try {
-    const cache = join(logDir(), "models.json");
-    if (existsSync(cache)) {
-      const j = JSON.parse(readFileSync(cache, "utf8"));
-      const ids = (j.data || []).map((m) => m.id).filter(Boolean);
-      if (ids.length) return ids;
-    }
-  } catch {}
-  // 兜底：硬编码常见 free 模型，供离线时参考（2026-09 实测 /zen/v1/models free 池）
-  return ["big-pickle", "mimo-v2.5-free", "ling-3.0-flash-fin-free", "deepseek-v4-flash-free", "nemotron-3.5-lightning-free", "nemotron-3-ultra-free", "muse-spark-1.3-contributor-free", "muse-spark-1.2-contributor-free"];
-}
-
-export function buildSystemPrompt({ modelsOverride } = {}) {
-  const mini = readMini();
-  const models = modelsOverride || readModels();
-  const now = nowShanghaiYMDHM();
-  // 按供应商分组，便于“bai有哪些模型”这类问题直接回答
-  const byProv = {};
-  for (const id of models) {
-    const slash = id.indexOf("/");
-    const prov = slash > 0 ? id.slice(0, slash) : "opencode";
-    if (!byProv[prov]) byProv[prov] = [];
-    byProv[prov].push(id);
-  }
-  const provSummary = Object.entries(byProv).map(([p, arr]) => `${p}(${arr.length}): ${arr.slice(0, 12).join(", ")}${arr.length > 12 ? " …" : ""}`).join(" | ");
-  return `你是 mslxdff 的终端助手，运行在用户本机，帮用户把自然语言翻译成精确的 mslxdff CLI 命令并执行。
-
-当前时间：${now}
-可用模型（你必须从中精确选择，禁止自创，共 ${models.length} 个）：${models.join(", ")}
-按供应商：${provSummary}
-
-${mini}
-
-语言（最高优先级）：
-- 全部面向用户的自然语言回复**必须使用简体中文**（无论用户用英文/日文/拼音提问，都用中文回答）。
-- 仅代码、命令、模型 id、路径、JSON 等技术标识保持原文，不做翻译。
-- 禁止输出英文长段解释；中英文混排时中文为主。
-
-规则：
-- 用户说简称你必须自行查“可用模型”找到全称，例如 hy3→hy3-free，mimo→mimo-v2.5-free，bigpickle→big-pickle。
-- 永远输出精确的命令与模型 id，大小写敏感。
-- 需要执行命令时调用 run_command，需要看文件时调用 read_file，需要检查网络/服务可用性时调用 curl。
-- curl 简写：upstream(=上游 https://opencode.ai/zen/v1/models)、local/health(=本机 /health)、local/models(=本机 /v1/models)，也支持完整 http(s) URL；会自动补上游头、本机 token 与已配置供应商 key（直连 https://api.b.ai/v1/models 会自动带 bai 的 key，无需手动加头）。
-- 查“某供应商有哪些模型”**优先用 CLI 直查**：若上方“可用模型”已能回答，直接前缀过滤回答（如 workbuddy/ 即 workbuddy）；需实时拉取时调用 run_command: "-provider <id> models"（如 -provider workbuddy models）或 "-model list --provider <id>"，按 allowlist 过滤，--json 供脚本。**禁止**调 -provider <id> list（这是查配置，不是查模型！）。**错误示例**：workbuddy有哪些模型 → 调 -provider workbuddy list → 错。**正确**：-provider workbuddy models。禁止为此调用 -showtoken。
-- 严禁幻觉命令：mslxdff "hi" --model X / mslxdff --model X "hi" / mslxdff -chat --model X 都不存在，输出只会是 status 页。探活任意模型（含 cline/*、workbuddy/*、bai/*）必须用 curl POST http://localhost:8989/v1/chat/completions，body 为 {"model":"<前缀/模型>","messages":[{"role":"user","content":"hi"}],"stream":false}，成功 200 + x-mslxdff-via:local 即通；401 代表本机 token 陈旧需提示 mslxdff -stop && mslxdff；403 + x-mslxdff-allowlist:1 代表白名单未放行需 allowlist add。
-- **禁止重复调用（最高优先级）**：同一 run_command/curl/read_file 在本轮只执行一次，重复会被工具侧 SKIPPED_DUP 拦截；查询类（-showtoken/-status/-provider list/-providers list/-model list/-group list/-log 等）**调用一次即答案**，拿到 OK 结果后必须**立即用中文直接回答用户**，禁止再发起任何工具调用。收到 SKIPPED_DUP 或“请直接回答/禁止再调用”提示时，必须 0 工具直接回答。
-- 禁止调用 -uninstall，包含即拒绝；-showtoken 仅在用户明确要求查看/调试本机 token 时才用，查模型/查供应商严禁调用。
-- 回复风格：简洁友好，执行前后用中文说明你在做什么。
-- 若用户只是闲聊/提问且可用模型列表已能回答，不调工具，直接用中文回答。`;
-}
-
-export function getModelsForPrompt() {
-  return readModels();
-}
+import { readFileSync, existsSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { logDir } from "../logs.js";
+import { nowShanghaiYMDHM } from "../time.js";
+
+const pkgRoot = join(dirname(fileURLToPath(import.meta.url)), "../..");
+
+function readMini() {
+  try {
+    const p = join(pkgRoot, "docs/cli_help_mini.md");
+    return readFileSync(p, "utf8");
+  } catch { return ""; }
+}
+
+function readModels() {
+  try {
+    const cache = join(logDir(), "models.json");
+    if (existsSync(cache)) {
+      const j = JSON.parse(readFileSync(cache, "utf8"));
+      const ids = (j.data || []).map((m) => m.id).filter(Boolean);
+      if (ids.length) return ids;
+    }
+  } catch {}
+  // 兜底：硬编码常见 free 模型，供离线时参考（2026-09 实测 /zen/v1/models free 池）
+  return ["big-pickle", "mimo-v2.5-free", "ling-3.0-flash-fin-free", "deepseek-v4-flash-free", "nemotron-3.5-lightning-free", "nemotron-3-ultra-free", "muse-spark-1.3-contributor-free", "muse-spark-1.2-contributor-free"];
+}
+
+export function buildSystemPrompt({ modelsOverride } = {}) {
+  const mini = readMini();
+  const models = modelsOverride || readModels();
+  const now = nowShanghaiYMDHM();
+  // 按供应商分组，便于“bai有哪些模型”这类问题直接回答
+  const byProv = {};
+  for (const id of models) {
+    const slash = id.indexOf("/");
+    const prov = slash > 0 ? id.slice(0, slash) : "opencode";
+    if (!byProv[prov]) byProv[prov] = [];
+    byProv[prov].push(id);
+  }
+  const provSummary = Object.entries(byProv).map(([p, arr]) => `${p}(${arr.length}): ${arr.slice(0, 12).join(", ")}${arr.length > 12 ? " …" : ""}`).join(" | ");
+  return `你是 mslxdff 的终端助手，运行在用户本机，帮用户把自然语言翻译成精确的 mslxdff CLI 命令并执行。
+
+当前时间：${now}
+可用模型（你必须从中精确选择，禁止自创，共 ${models.length} 个）：${models.join(", ")}
+按供应商：${provSummary}
+
+${mini}
+
+语言（最高优先级）：
+- 全部面向用户的自然语言回复**必须使用简体中文**（无论用户用英文/日文/拼音提问，都用中文回答）。
+- 仅代码、命令、模型 id、路径、JSON 等技术标识保持原文，不做翻译。
+- 禁止输出英文长段解释；中英文混排时中文为主。
+
+规则：
+- 用户说简称你必须自行查“可用模型”找到全称，例如 hy3→hy3-free，mimo→mimo-v2.5-free，bigpickle→big-pickle。
+- 永远输出精确的命令与模型 id，大小写敏感。
+- 需要执行命令时调用 run_command，需要看文件时调用 read_file，需要检查网络/服务可用性时调用 curl。
+- curl 简写：upstream(=上游 https://opencode.ai/zen/v1/models)、local/health(=本机 /health)、local/models(=本机 /v1/models)，也支持完整 http(s) URL；会自动补上游头、本机 token 与已配置供应商 key（直连 https://api.b.ai/v1/models 会自动带 bai 的 key，无需手动加头）。
+- 查“某供应商有哪些模型”**优先用 CLI 直查**：若上方“可用模型”已能回答，直接前缀过滤回答（如 workbuddy/ 即 workbuddy）；需实时拉取时调用 run_command: "-provider <id> models"（如 -provider workbuddy models）或 "-model list --provider <id>"，按 allowlist 过滤，--json 供脚本。**禁止**调 -provider <id> list（这是查配置，不是查模型！）。**错误示例**：workbuddy有哪些模型 → 调 -provider workbuddy list → 错。**正确**：-provider workbuddy models。禁止为此调用 -showtoken。
+- 严禁幻觉命令：mslxdff "hi" --model X / mslxdff --model X "hi" / mslxdff -chat --model X 都不存在，输出只会是 status 页。探活任意模型（含 cline/*、workbuddy/*、bai/*）必须用 curl POST http://localhost:8989/v1/chat/completions，body 为 {"model":"<前缀/模型>","messages":[{"role":"user","content":"hi"}],"stream":false}，成功 200 + x-mslxdff-via:local 即通；401 代表本机 token 陈旧需提示 mslxdff -stop && mslxdff；403 + x-mslxdff-allowlist:1 代表白名单未放行需 allowlist add。
+- **禁止重复调用（最高优先级）**：同一 run_command/curl/read_file 在本轮只执行一次，重复会被工具侧 SKIPPED_DUP 拦截；查询类（-showtoken/-status/-provider list/-providers list/-model list/-group list/-log 等）**调用一次即答案**，拿到 OK 结果后必须**立即用中文直接回答用户**，禁止再发起任何工具调用。收到 SKIPPED_DUP 或“请直接回答/禁止再调用”提示时，必须 0 工具直接回答。
+- 禁止调用 -uninstall，包含即拒绝；-showtoken 仅在用户明确要求查看/调试本机 token 时才用，查模型/查供应商严禁调用。
+- 回复风格：简洁友好，执行前后用中文说明你在做什么。
+- 若用户只是闲聊/提问且可用模型列表已能回答，不调工具，直接用中文回答。`;
+}
+
+export function getModelsForPrompt() {
+  return readModels();
+}