mslxdff 0.1.203 → 0.1.207

package/src/chat/tool-handlers.js

@@ -1,67 +1,67 @@
-/**
- * 工具执行层 — 从 engine.js 抽出的工具段。
- * buildDedupKey 纯函数（去重键归一）；runTool 执行单工具调用并返回结果文本。
- * SKIPPED_DUP/forceNoTools 状态机仍在 engine runTurn 侧。
- */
-export function buildDedupKey(name, args = {}) {
-  if (name === "run_command") {
-    const cmd = String(args.command || "").trim().toLowerCase().replace(/\s+/g, " ");
-    const norm = cmd.replace(/^-+providers\b/, "-provider").replace(/\s+/g, " ").trim();
-    return `run_command:${norm}`;
-  }
-  if (name === "curl") {
-    const u = String(args.url || "").trim().toLowerCase();
-    const m = String(args.method || "GET").toUpperCase();
-    return `curl:${m}:${u}:${String(args.body || "").slice(0, 200)}`;
-  }
-  if (name === "read_file") {
-    return `read_file:${String(args.path || "").trim().toLowerCase()}`;
-  }
-  return `${name}:${JSON.stringify(args)}`;
-}
-
-export async function runTool({
-  name,
-  args = {},
-  userText = "",
-  execCommand = async () => ({ ok: false, output: "no exec" }),
-  readFileTool = async () => ({ ok: false, output: "no read" }),
-  curlTool = async () => ({ ok: false, output: "no curl" }),
-  onTrace = () => {},
-} = {}) {
-  const t1 = performance.now();
-  if (name === "run_command") {
-    const cmd = String(args.command || "").trim();
-    const r = await execCommand(cmd);
-    let result = `${r.ok ? "OK" : "FAIL"}: ${r.output}`;
-    const lowCmd = cmd.toLowerCase().replace(/\s+/g, " ").trim();
-    const asksModel = String(userText || "").toLowerCase().includes("模型");
-    const isOnceAndDone =
-      /^-+(showtoken|status|s|providers?\b|model\b|group\b|log\b|workbuddy\b|free\b|autostart\b|plugins\b)/.test(lowCmd) ||
-      lowCmd === "-provider list" || lowCmd === "-providers list";
-    if (isOnceAndDone && r.ok) {
-      if (asksModel && lowCmd.includes("-provider")) {
-        result += `\n\n[提示：此命令仅显示供应商配置，不包含模型列表。用户问的是“有哪些模型”，请用系统提示中的“可用模型”按前缀过滤回答，或调 curl local/models，不要再调 provider list]`;
-      } else {
-        result += `\n\n[系统提示：此查询已完成，结果即答案，请直接用中文回答用户，禁止再调用相同或同类查询工具]`;
-      }
-    }
-    onTrace(`[tool] run_command "${cmd.slice(0, 40)}" · ${Math.round(performance.now() - t1)}ms · ${r.ok ? "OK" : "FAIL"} ${r.output.length}字`);
-    return result;
-  }
-  if (name === "read_file") {
-    const r = await readFileTool(args);
-    let result = `${r.ok ? "OK" : "FAIL"}: ${r.output.slice(0, 6000)}`;
-    if (r.ok) result += `\n\n[系统提示：文件已读取，请直接基于内容回答，禁止重复读取同一文件]`;
-    onTrace(`[tool] read_file ${args.path} · ${Math.round(performance.now() - t1)}ms · ${r.ok ? "OK" : "FAIL"} ${r.output.length}字`);
-    return result;
-  }
-  if (name === "curl") {
-    const u = String(args.url || "").trim();
-    const r = await curlTool(args);
-    const result = `${r.ok ? "OK" : "FAIL"}: ${r.output.slice(0, 6000)}`;
-    onTrace(`[tool] curl ${u} · ${Math.round(performance.now() - t1)}ms`);
-    return result;
-  }
-  return `unknown tool ${name}`;
-}
+/**
+ * 工具执行层 — 从 engine.js 抽出的工具段。
+ * buildDedupKey 纯函数（去重键归一）；runTool 执行单工具调用并返回结果文本。
+ * SKIPPED_DUP/forceNoTools 状态机仍在 engine runTurn 侧。
+ */
+export function buildDedupKey(name, args = {}) {
+  if (name === "run_command") {
+    const cmd = String(args.command || "").trim().toLowerCase().replace(/\s+/g, " ");
+    const norm = cmd.replace(/^-+providers\b/, "-provider").replace(/\s+/g, " ").trim();
+    return `run_command:${norm}`;
+  }
+  if (name === "curl") {
+    const u = String(args.url || "").trim().toLowerCase();
+    const m = String(args.method || "GET").toUpperCase();
+    return `curl:${m}:${u}:${String(args.body || "").slice(0, 200)}`;
+  }
+  if (name === "read_file") {
+    return `read_file:${String(args.path || "").trim().toLowerCase()}`;
+  }
+  return `${name}:${JSON.stringify(args)}`;
+}
+
+export async function runTool({
+  name,
+  args = {},
+  userText = "",
+  execCommand = async () => ({ ok: false, output: "no exec" }),
+  readFileTool = async () => ({ ok: false, output: "no read" }),
+  curlTool = async () => ({ ok: false, output: "no curl" }),
+  onTrace = () => {},
+} = {}) {
+  const t1 = performance.now();
+  if (name === "run_command") {
+    const cmd = String(args.command || "").trim();
+    const r = await execCommand(cmd);
+    let result = `${r.ok ? "OK" : "FAIL"}: ${r.output}`;
+    const lowCmd = cmd.toLowerCase().replace(/\s+/g, " ").trim();
+    const asksModel = String(userText || "").toLowerCase().includes("模型");
+    const isOnceAndDone =
+      /^-+(showtoken|status|s|providers?\b|model\b|group\b|log\b|workbuddy\b|free\b|autostart\b|plugins\b)/.test(lowCmd) ||
+      lowCmd === "-provider list" || lowCmd === "-providers list";
+    if (isOnceAndDone && r.ok) {
+      if (asksModel && lowCmd.includes("-provider")) {
+        result += `\n\n[提示：此命令仅显示供应商配置，不包含模型列表。用户问的是“有哪些模型”，请用系统提示中的“可用模型”按前缀过滤回答，或调 curl local/models，不要再调 provider list]`;
+      } else {
+        result += `\n\n[系统提示：此查询已完成，结果即答案，请直接用中文回答用户，禁止再调用相同或同类查询工具]`;
+      }
+    }
+    onTrace(`[tool] run_command "${cmd.slice(0, 40)}" · ${Math.round(performance.now() - t1)}ms · ${r.ok ? "OK" : "FAIL"} ${r.output.length}字`);
+    return result;
+  }
+  if (name === "read_file") {
+    const r = await readFileTool(args);
+    let result = `${r.ok ? "OK" : "FAIL"}: ${r.output.slice(0, 6000)}`;
+    if (r.ok) result += `\n\n[系统提示：文件已读取，请直接基于内容回答，禁止重复读取同一文件]`;
+    onTrace(`[tool] read_file ${args.path} · ${Math.round(performance.now() - t1)}ms · ${r.ok ? "OK" : "FAIL"} ${r.output.length}字`);
+    return result;
+  }
+  if (name === "curl") {
+    const u = String(args.url || "").trim();
+    const r = await curlTool(args);
+    const result = `${r.ok ? "OK" : "FAIL"}: ${r.output.slice(0, 6000)}`;
+    onTrace(`[tool] curl ${u} · ${Math.round(performance.now() - t1)}ms`);
+    return result;
+  }
+  return `unknown tool ${name}`;
+}

package/src/chat/tools.js

@@ -7,7 +7,7 @@ import { FORBIDDEN } from "./config.js";
 import { logDir } from "../logs.js";
 import { defaultStateFile, loadProviderKeys, loadProviderConfigs } from "../state.js";
 import { compatFetch } from "../compat.js";
-import { opencodeClientIdentity } from "../opencode-identity.js";
+import { opencodeClientIdentity } from "../opencode-identity.js";
 
 const pkgRoot = resolve(dirname(fileURLToPath(import.meta.url)), "../..");
 const allowedRoots = [

package/src/chat/upstream.js

@@ -1,59 +1,59 @@
-import { createDirectClient } from "./direct.js";
-import { createGatewayClient } from "./gateway.js";
-import { createCooling } from "./cooling.js";
-import { createOrchestrator } from "./orchestrator.js";
-import { createUpstreamClient } from "../upstream.js";
-import { CHAT_TIMEOUT_MS, CHAT_GATEWAY_TIMEOUT_MS, CHAT_PREFERRED, CHAT_FALLBACK } from "./config.js";
-import * as state from "../state.js";
-import { performance } from "node:perf_hooks";
-import { compatFetch } from "../compat.js";
-
-// 冷却深模块对接真实 state
-const cooling = createCooling({
-  loadModelErrors: () => {
-    try { return state.loadModelErrors(); } catch { return {}; }
-  },
-  saveModelErrors: (o) => { try { state.saveModelErrors(o); } catch {} },
-  loadModelLatencies: () => {
-    try { return state.loadModelLatencies(); } catch { return {}; }
-  },
-  saveModelLatencies: (o) => { try { state.saveModelLatencies(o); } catch {} },
-  flush: () => { try { state.flushStateSync(); } catch {} },
-  now: () => Date.now(),
-});
-
-const direct = createDirectClient({
-  createUpstreamClient,
-  chatTimeoutMs: CHAT_TIMEOUT_MS,
-  env: process.env,
-  fetchImpl: compatFetch,
-});
-
-const gateway = createGatewayClient({
-  fetchImpl: compatFetch,
-  loadToken: async () => {
-    try { const l = await state.loadToken(); return String(l?.token || "").trim(); } catch { return ""; }
-  },
-  getPort: () => {
-    try { const p = state.getPort(); if (Number.isInteger(p) && p > 0) return p; } catch {}
-    const v = Number(process.env.MSLXDFF_PORT);
-    if (Number.isInteger(v) && v > 0) return v;
-    return 8989;
-  },
-  gatewayTimeoutMs: CHAT_GATEWAY_TIMEOUT_MS,
-  env: process.env,
-});
-
-const orch = createOrchestrator({
-  chatOnce: direct.chatOnce,
-  chatViaGateway: gateway.chatViaGateway,
-  cooling,
-  config: { CHAT_PREFERRED, CHAT_FALLBACK, CHAT_GATEWAY_TIMEOUT_MS },
-  env: process.env,
-  performance,
-});
-
-export const chatOnce = direct.chatOnce;
-export const chatWithFallback = orch.chatWithFallback;
-export const summarizeHistory = orch.summarizeHistory;
-// keywords for legacy file-content check: chatViaGateway gateway auto safeChatOnce
+import { createDirectClient } from "./direct.js";
+import { createGatewayClient } from "./gateway.js";
+import { createCooling } from "./cooling.js";
+import { createOrchestrator } from "./orchestrator.js";
+import { createUpstreamClient } from "../upstream.js";
+import { CHAT_TIMEOUT_MS, CHAT_GATEWAY_TIMEOUT_MS, CHAT_PREFERRED, CHAT_FALLBACK } from "./config.js";
+import * as state from "../state.js";
+import { performance } from "node:perf_hooks";
+import { compatFetch } from "../compat.js";
+
+// 冷却深模块对接真实 state
+const cooling = createCooling({
+  loadModelErrors: () => {
+    try { return state.loadModelErrors(); } catch { return {}; }
+  },
+  saveModelErrors: (o) => { try { state.saveModelErrors(o); } catch {} },
+  loadModelLatencies: () => {
+    try { return state.loadModelLatencies(); } catch { return {}; }
+  },
+  saveModelLatencies: (o) => { try { state.saveModelLatencies(o); } catch {} },
+  flush: () => { try { state.flushStateSync(); } catch {} },
+  now: () => Date.now(),
+});
+
+const direct = createDirectClient({
+  createUpstreamClient,
+  chatTimeoutMs: CHAT_TIMEOUT_MS,
+  env: process.env,
+  fetchImpl: compatFetch,
+});
+
+const gateway = createGatewayClient({
+  fetchImpl: compatFetch,
+  loadToken: async () => {
+    try { const l = await state.loadToken(); return String(l?.token || "").trim(); } catch { return ""; }
+  },
+  getPort: () => {
+    try { const p = state.getPort(); if (Number.isInteger(p) && p > 0) return p; } catch {}
+    const v = Number(process.env.MSLXDFF_PORT);
+    if (Number.isInteger(v) && v > 0) return v;
+    return 8989;
+  },
+  gatewayTimeoutMs: CHAT_GATEWAY_TIMEOUT_MS,
+  env: process.env,
+});
+
+const orch = createOrchestrator({
+  chatOnce: direct.chatOnce,
+  chatViaGateway: gateway.chatViaGateway,
+  cooling,
+  config: { CHAT_PREFERRED, CHAT_FALLBACK, CHAT_GATEWAY_TIMEOUT_MS },
+  env: process.env,
+  performance,
+});
+
+export const chatOnce = direct.chatOnce;
+export const chatWithFallback = orch.chatWithFallback;
+export const summarizeHistory = orch.summarizeHistory;
+// keywords for legacy file-content check: chatViaGateway gateway auto safeChatOnce

package/src/chat-pipeline/auto-race.js

@@ -1,158 +1,158 @@
-import { performance } from "node:perf_hooks";
-import { injectReasoningContent } from "../reasoning.js";
-import { runHook } from "../plugins.js";
-import { errMsg } from "../routes/helpers.js";
-import { withOutputFloor } from "./empty-turn.js";
-import { classifyTerminalUpstream, readVerdictText, terminalShortcircuitOff } from "./terminal-verdict.js";
-import { handleLocalRelay } from "../routes/chat/local-handler.js";
-import { handleExhaustedAll } from "../routes/chat/exhausted-handler.js";
-
-/**
- * auto 首次并发择优 — 从 engine.js 抽出的第一段。
- * 无 prior-success 且多非冷却候选时并发竞速，胜者走 local relay。
- * 返回 { done:true } 表示已终结（handled/exhausted），调用方直接 return；
- * 返回 { done:false, order } 表示未终结，调用方用过滤后的 order 继续。
- */
-export async function runAutoRace(ctx, deps = {}) {
-  const { localRelay = handleLocalRelay, exhaustedAll = handleExhaustedAll } = deps;
-  const {
-    reqId, requested, body, hops, useAuto, lockModel, plugins,
-    auto, upstream, perf0, stages, mark, evt, logCall, logError,
-    startedAt, handlerCtx, res,
-  } = ctx;
-  let order = ctx.order;
-  const shareKeys = ctx.shareKeys ?? ctx.policy?.shareKeys ?? {};
-  const workbuddyUid = ctx.workbuddyUid ?? ctx.policy?.workbuddyUid ?? null;
-
-  if (!(useAuto && order.length > 1 && auto && !lockModel)) return { done: false, order };
-  const statuses = auto.statuses?.() ?? {};
-  const hasPriorSuccess = Object.values(statuses).some((e) => e && typeof e === "object" && e.status === "normal");
-  const nonCoolingOrder = order.filter((m) => { try { return !auto.isCooling(m); } catch { return true; } });
-  if (hasPriorSuccess || nonCoolingOrder.length <= 1) return { done: false, order };
-
-  const concLimit = (() => {
-    const v = Number(process.env.MSLXDFF_AUTO_CONCURRENT);
-    if (Number.isInteger(v) && v > 0) return Math.min(v, nonCoolingOrder.length);
-    return Math.min(nonCoolingOrder.length, 5);
-  })();
-  let raceModels = nonCoolingOrder.slice(0, concLimit);
-  if (plugins?.length) {
-    const k = [];
-    for (const m of raceModels) {
-      const b = await runHook(plugins, "model:beforeTry", { reqId, requested, model: m, hops });
-      if (b.value === false || b.value?.skip) continue;
-      k.push(m);
-    }
-    raceModels = k;
-  }
-  if (!raceModels.length) {
-    order = order.filter((m) => !new Set(nonCoolingOrder.slice(0, concLimit)).has(m));
-    if (!order.length) {
-      await exhaustedAll({ res, body, lastErr: { model: requested, status: 502, message: "all concurrent candidates skipped by plugin" }, order: nonCoolingOrder.slice(0, concLimit), requested, handlerCtx: { ...handlerCtx, reqId, startedAt }, evt, logCall, mark, perf0, stages });
-      return { done: true };
-    }
-    return { done: false, order };
-  }
-
-  evt("auto-concurrent-race", { reqId, models: raceModels, skippedFaulty: order.length - nonCoolingOrder.length, limit: concLimit });
-  const raceStart = performance.now();
-  const attempts = raceModels.map(async (m) => {
-    let f = { ...injectReasoningContent(m, body), model: m };
-    // 首发输出额度托底（同 serial-trial；auto 竞速路径独立组装，不能漏）
-    const _fl = withOutputFloor(f);
-    if (_fl !== f) {
-      const _fk = ["max_tokens", "max_completion_tokens"].find((k) => _fl[k] !== f[k]) || "max_tokens";
-      evt("output-floor", { reqId, model: m, key: _fk, raiseFrom: Number(f[_fk]), raiseTo: Number(_fl[_fk]) });
-      f = _fl;
-    }
-    if (plugins?.length) {
-      const u = await runHook(plugins, "upstream:request", { reqId, requested, model: m, payload: f, stream: Boolean(body.stream) });
-      if (u.changed && u.value?.payload) f = u.value.payload;
-    }
-    let r = null;
-    try {
-      const o = {};
-      if (Object.keys(shareKeys).length) o.shareKeys = shareKeys;
-      if (workbuddyUid) o.workbuddyUid = workbuddyUid;
-      if (handlerCtx?.sessionId) o.sessionId = handlerCtx.sessionId;
-      r = await upstream.chat(f, Object.keys(o).length ? o : undefined);
-    } catch (e) {
-      if (plugins?.length) runHook(plugins, "upstream:response", { reqId, requested, model: m, status: null, ok: false, error: errMsg(e), timing: e?._t ?? null }).catch(() => {});
-      return { model: m, ok: false, error: errMsg(e), status: 502, timing: e?._t ?? null };
-    }
-    if (plugins?.length) runHook(plugins, "upstream:response", { reqId, requested, model: m, status: r instanceof Error ? null : r?.status ?? null, ok: !(r instanceof Error) && r ? r.status < 400 : false, error: r instanceof Error ? errMsg(r) : null, timing: r?._t ?? null }).catch(() => {});
-    if (r && r.status >= 400) {
-      const a = r.status === 403 && r.headers?.get?.("x-mslxdff-allowlist") === "1";
-      if (a) return { model: m, ok: false, error: "allowlist", status: 403, allowlist: true };
-      return { model: m, ok: false, error: `upstream ${r.status}`, status: r.status, res: r, timing: r._t ?? null };
-    }
-    if (r instanceof Error) return { model: m, ok: false, error: errMsg(r), status: 502 };
-    return { model: m, ok: true, res: r, status: r.status, timing: r._t ?? null };
-  });
-  const results = await Promise.allSettled(attempts);
-  const okList = results.map((r, i) => ({ r, i, model: raceModels[i] }))
-    .filter(({ r }) => r.status === "fulfilled" && r.value?.ok)
-    .map(({ r, i, model }) => ({ model, idx: i, val: r.value, t: r.value.timing?.totalMs ?? r.value.timing?.ms ?? Number.MAX_SAFE_INTEGER }));
-  if (okList.length) {
-    okList.sort((a, b) => a.t - b.t);
-    const best = okList[0];
-    const winModel = best.model;
-    evt("auto-concurrent-win", { reqId, model: winModel, timing: best.val.timing, totalMs: Math.round(performance.now() - raceStart), tried: raceModels.length });
-    for (const { r, i } of results.map((r, i) => ({ r, i }))) {
-      const m = raceModels[i];
-      if (r.status === "fulfilled" && r.value?.ok) {
-        if (m === winModel) {
-          const latencyMs = r.value.timing?.totalMs ?? Math.round(performance.now() - raceStart);
-          await auto.recordOk(m, { latencyMs });
-          try { const { savePreferredModel } = await import("../state.js"); savePreferredModel(m); evt("auto-concurrent-preferred", { reqId, model: m }); } catch {}
-        }
-      } else if (r.status === "fulfilled" && !r.value?.ok && !r.value?.allowlist) await auto.recordError(m, { status: r.value.status || 502 });
-      else if (r.status === "rejected") await auto.recordError(m, { status: 502 });
-    }
-    handlerCtx.model = winModel;
-    const lr = await localRelay({ upRes: best.val.res, model: winModel, body, order: raceModels, idx: best.idx, lastErr: null, requested, useAuto, lockModel, auto, handlerCtx, evt, logCall, logError, mark, perf0, attemptStartMs: raceStart, stages, startedAt, plugins, res });
-    if (lr.handled) return { done: true };
-    if (!lr.lastErr) return { done: true };
-  } else {
-    evt("auto-concurrent-all-fail", { reqId, tried: raceModels.length, totalMs: Math.round(performance.now() - raceStart) });
-    for (const { r, i } of results.map((r, i) => ({ r, i }))) {
-      const m = raceModels[i];
-      if (r.status === "fulfilled" && !r.value?.ok && !r.value?.allowlist) await auto.recordError(m, { status: r.value.status || 502 });
-      else if (r.status === "rejected") await auto.recordError(m, { status: 502 });
-    }
-  }
-  const triedSet = new Set(raceModels);
-  order = order.filter((m) => !triedSet.has(m));
-  if (!order.length) {
-    // 并发竞速全挂时也要认终局判决（判据见 terminal-verdict.js）：否则「上游 6s 就把话说死」在这条出口
-    // 换成 502 "all concurrent candidates failed"，serial-trial 那条短路救不到它——同一失败模式的另一条命
-    // （`auto` 是 chat 门缺省路径，冷启动无 prior-success 且 >1 可用候选时必走竞速）。
-    // 走到这里 order 已空（没有下一候选可换）、组员也不在这条路上，所以本段只改「客户端看到什么」：
-    // 挑第一条能定终局的失败，把它的**原响应**交给 exhaustedAll 透传（原状态码 + 原字节 + 原头）。
-    let terminalLast = null;
-    if (!terminalShortcircuitOff()) {
-      for (const r of results) {
-        const v = r.status === "fulfilled" ? r.value : null;
-        if (!v || v.ok || v.allowlist || !v.res) continue;
-        const txt = await readVerdictText(v.res);
-        if (!txt.trim()) continue; // 读不到正文一律不判终局（保守回退「照跑收场」的旧行为）
-        const verdict = classifyTerminalUpstream(v.status, txt);
-        if (!verdict.terminal) continue;
-        evt("upstream-terminal", { reqId, model: v.model, status: v.status, kind: verdict.kind, marker: verdict.marker, peersSkipped: true, via: "auto-race" });
-        logError(v.model, v.status, `upstream ${v.status} body=${txt.replace(/\s+/g, " ").slice(0, 300)} · 终局判决 ${verdict.kind}（auto 竞速收场直发原文）`);
-        terminalLast = { model: v.model, upstream: v.res, status: v.status, message: txt.replace(/\s+/g, " ").slice(0, 400) || `upstream ${v.status}`, terminal: verdict.kind };
-        break;
-      }
-    }
-    const failedStatuses = results.map((r) => (r.status === "fulfilled" ? r.value?.status : null)).filter((s) => Number.isInteger(s));
-    const lastStatus = failedStatuses[failedStatuses.length - 1] || failedStatuses[0] || 502;
-    // 竞速出口的取证与串行一致：状态码、原响应、同名回显头（客户端只看得到这一个面）
-    if (terminalLast && !res.headersSent && typeof res?.setHeader === "function") {
-      try { res.setHeader("x-mslxdff-terminal", terminalLast.terminal); } catch {}
-    }
-    const last = terminalLast || { model: raceModels[0] || requested, status: lastStatus, message: "all concurrent candidates failed" };
-    await exhaustedAll({ res, body, lastErr: last, order: raceModels, requested, handlerCtx: { ...handlerCtx, reqId, startedAt }, evt, logCall, mark, perf0, stages });
-    return { done: true };
-  }
-  return { done: false, order };
-}
+import { performance } from "node:perf_hooks";
+import { injectReasoningContent } from "../reasoning.js";
+import { runHook } from "../plugins.js";
+import { errMsg } from "../routes/helpers.js";
+import { withOutputFloor } from "./empty-turn.js";
+import { classifyTerminalUpstream, readVerdictText, terminalShortcircuitOff } from "./terminal-verdict.js";
+import { handleLocalRelay } from "../routes/chat/local-handler.js";
+import { handleExhaustedAll } from "../routes/chat/exhausted-handler.js";
+
+/**
+ * auto 首次并发择优 — 从 engine.js 抽出的第一段。
+ * 无 prior-success 且多非冷却候选时并发竞速，胜者走 local relay。
+ * 返回 { done:true } 表示已终结（handled/exhausted），调用方直接 return；
+ * 返回 { done:false, order } 表示未终结，调用方用过滤后的 order 继续。
+ */
+export async function runAutoRace(ctx, deps = {}) {
+  const { localRelay = handleLocalRelay, exhaustedAll = handleExhaustedAll } = deps;
+  const {
+    reqId, requested, body, hops, useAuto, lockModel, plugins,
+    auto, upstream, perf0, stages, mark, evt, logCall, logError,
+    startedAt, handlerCtx, res,
+  } = ctx;
+  let order = ctx.order;
+  const shareKeys = ctx.shareKeys ?? ctx.policy?.shareKeys ?? {};
+  const workbuddyUid = ctx.workbuddyUid ?? ctx.policy?.workbuddyUid ?? null;
+
+  if (!(useAuto && order.length > 1 && auto && !lockModel)) return { done: false, order };
+  const statuses = auto.statuses?.() ?? {};
+  const hasPriorSuccess = Object.values(statuses).some((e) => e && typeof e === "object" && e.status === "normal");
+  const nonCoolingOrder = order.filter((m) => { try { return !auto.isCooling(m); } catch { return true; } });
+  if (hasPriorSuccess || nonCoolingOrder.length <= 1) return { done: false, order };
+
+  const concLimit = (() => {
+    const v = Number(process.env.MSLXDFF_AUTO_CONCURRENT);
+    if (Number.isInteger(v) && v > 0) return Math.min(v, nonCoolingOrder.length);
+    return Math.min(nonCoolingOrder.length, 5);
+  })();
+  let raceModels = nonCoolingOrder.slice(0, concLimit);
+  if (plugins?.length) {
+    const k = [];
+    for (const m of raceModels) {
+      const b = await runHook(plugins, "model:beforeTry", { reqId, requested, model: m, hops });
+      if (b.value === false || b.value?.skip) continue;
+      k.push(m);
+    }
+    raceModels = k;
+  }
+  if (!raceModels.length) {
+    order = order.filter((m) => !new Set(nonCoolingOrder.slice(0, concLimit)).has(m));
+    if (!order.length) {
+      await exhaustedAll({ res, body, lastErr: { model: requested, status: 502, message: "all concurrent candidates skipped by plugin" }, order: nonCoolingOrder.slice(0, concLimit), requested, handlerCtx: { ...handlerCtx, reqId, startedAt }, evt, logCall, mark, perf0, stages });
+      return { done: true };
+    }
+    return { done: false, order };
+  }
+
+  evt("auto-concurrent-race", { reqId, models: raceModels, skippedFaulty: order.length - nonCoolingOrder.length, limit: concLimit });
+  const raceStart = performance.now();
+  const attempts = raceModels.map(async (m) => {
+    let f = { ...injectReasoningContent(m, body), model: m };
+    // 首发输出额度托底（同 serial-trial；auto 竞速路径独立组装，不能漏）
+    const _fl = withOutputFloor(f);
+    if (_fl !== f) {
+      const _fk = ["max_tokens", "max_completion_tokens"].find((k) => _fl[k] !== f[k]) || "max_tokens";
+      evt("output-floor", { reqId, model: m, key: _fk, raiseFrom: Number(f[_fk]), raiseTo: Number(_fl[_fk]) });
+      f = _fl;
+    }
+    if (plugins?.length) {
+      const u = await runHook(plugins, "upstream:request", { reqId, requested, model: m, payload: f, stream: Boolean(body.stream) });
+      if (u.changed && u.value?.payload) f = u.value.payload;
+    }
+    let r = null;
+    try {
+      const o = {};
+      if (Object.keys(shareKeys).length) o.shareKeys = shareKeys;
+      if (workbuddyUid) o.workbuddyUid = workbuddyUid;
+      if (handlerCtx?.sessionId) o.sessionId = handlerCtx.sessionId;
+      r = await upstream.chat(f, Object.keys(o).length ? o : undefined);
+    } catch (e) {
+      if (plugins?.length) runHook(plugins, "upstream:response", { reqId, requested, model: m, status: null, ok: false, error: errMsg(e), timing: e?._t ?? null }).catch(() => {});
+      return { model: m, ok: false, error: errMsg(e), status: 502, timing: e?._t ?? null };
+    }
+    if (plugins?.length) runHook(plugins, "upstream:response", { reqId, requested, model: m, status: r instanceof Error ? null : r?.status ?? null, ok: !(r instanceof Error) && r ? r.status < 400 : false, error: r instanceof Error ? errMsg(r) : null, timing: r?._t ?? null }).catch(() => {});
+    if (r && r.status >= 400) {
+      const a = r.status === 403 && r.headers?.get?.("x-mslxdff-allowlist") === "1";
+      if (a) return { model: m, ok: false, error: "allowlist", status: 403, allowlist: true };
+      return { model: m, ok: false, error: `upstream ${r.status}`, status: r.status, res: r, timing: r._t ?? null };
+    }
+    if (r instanceof Error) return { model: m, ok: false, error: errMsg(r), status: 502 };
+    return { model: m, ok: true, res: r, status: r.status, timing: r._t ?? null };
+  });
+  const results = await Promise.allSettled(attempts);
+  const okList = results.map((r, i) => ({ r, i, model: raceModels[i] }))
+    .filter(({ r }) => r.status === "fulfilled" && r.value?.ok)
+    .map(({ r, i, model }) => ({ model, idx: i, val: r.value, t: r.value.timing?.totalMs ?? r.value.timing?.ms ?? Number.MAX_SAFE_INTEGER }));
+  if (okList.length) {
+    okList.sort((a, b) => a.t - b.t);
+    const best = okList[0];
+    const winModel = best.model;
+    evt("auto-concurrent-win", { reqId, model: winModel, timing: best.val.timing, totalMs: Math.round(performance.now() - raceStart), tried: raceModels.length });
+    for (const { r, i } of results.map((r, i) => ({ r, i }))) {
+      const m = raceModels[i];
+      if (r.status === "fulfilled" && r.value?.ok) {
+        if (m === winModel) {
+          const latencyMs = r.value.timing?.totalMs ?? Math.round(performance.now() - raceStart);
+          await auto.recordOk(m, { latencyMs });
+          try { const { savePreferredModel } = await import("../state.js"); savePreferredModel(m); evt("auto-concurrent-preferred", { reqId, model: m }); } catch {}
+        }
+      } else if (r.status === "fulfilled" && !r.value?.ok && !r.value?.allowlist) await auto.recordError(m, { status: r.value.status || 502 });
+      else if (r.status === "rejected") await auto.recordError(m, { status: 502 });
+    }
+    handlerCtx.model = winModel;
+    const lr = await localRelay({ upRes: best.val.res, model: winModel, body, order: raceModels, idx: best.idx, lastErr: null, requested, useAuto, lockModel, auto, handlerCtx, evt, logCall, logError, mark, perf0, attemptStartMs: raceStart, stages, startedAt, plugins, res });
+    if (lr.handled) return { done: true };
+    if (!lr.lastErr) return { done: true };
+  } else {
+    evt("auto-concurrent-all-fail", { reqId, tried: raceModels.length, totalMs: Math.round(performance.now() - raceStart) });
+    for (const { r, i } of results.map((r, i) => ({ r, i }))) {
+      const m = raceModels[i];
+      if (r.status === "fulfilled" && !r.value?.ok && !r.value?.allowlist) await auto.recordError(m, { status: r.value.status || 502 });
+      else if (r.status === "rejected") await auto.recordError(m, { status: 502 });
+    }
+  }
+  const triedSet = new Set(raceModels);
+  order = order.filter((m) => !triedSet.has(m));
+  if (!order.length) {
+    // 并发竞速全挂时也要认终局判决（判据见 terminal-verdict.js）：否则「上游 6s 就把话说死」在这条出口
+    // 换成 502 "all concurrent candidates failed"，serial-trial 那条短路救不到它——同一失败模式的另一条命
+    // （`auto` 是 chat 门缺省路径，冷启动无 prior-success 且 >1 可用候选时必走竞速）。
+    // 走到这里 order 已空（没有下一候选可换）、组员也不在这条路上，所以本段只改「客户端看到什么」：
+    // 挑第一条能定终局的失败，把它的**原响应**交给 exhaustedAll 透传（原状态码 + 原字节 + 原头）。
+    let terminalLast = null;
+    if (!terminalShortcircuitOff()) {
+      for (const r of results) {
+        const v = r.status === "fulfilled" ? r.value : null;
+        if (!v || v.ok || v.allowlist || !v.res) continue;
+        const txt = await readVerdictText(v.res);
+        if (!txt.trim()) continue; // 读不到正文一律不判终局（保守回退「照跑收场」的旧行为）
+        const verdict = classifyTerminalUpstream(v.status, txt);
+        if (!verdict.terminal) continue;
+        evt("upstream-terminal", { reqId, model: v.model, status: v.status, kind: verdict.kind, marker: verdict.marker, peersSkipped: true, via: "auto-race" });
+        logError(v.model, v.status, `upstream ${v.status} body=${txt.replace(/\s+/g, " ").slice(0, 300)} · 终局判决 ${verdict.kind}（auto 竞速收场直发原文）`);
+        terminalLast = { model: v.model, upstream: v.res, status: v.status, message: txt.replace(/\s+/g, " ").slice(0, 400) || `upstream ${v.status}`, terminal: verdict.kind };
+        break;
+      }
+    }
+    const failedStatuses = results.map((r) => (r.status === "fulfilled" ? r.value?.status : null)).filter((s) => Number.isInteger(s));
+    const lastStatus = failedStatuses[failedStatuses.length - 1] || failedStatuses[0] || 502;
+    // 竞速出口的取证与串行一致：状态码、原响应、同名回显头（客户端只看得到这一个面）
+    if (terminalLast && !res.headersSent && typeof res?.setHeader === "function") {
+      try { res.setHeader("x-mslxdff-terminal", terminalLast.terminal); } catch {}
+    }
+    const last = terminalLast || { model: raceModels[0] || requested, status: lastStatus, message: "all concurrent candidates failed" };
+    await exhaustedAll({ res, body, lastErr: last, order: raceModels, requested, handlerCtx: { ...handlerCtx, reqId, startedAt }, evt, logCall, mark, perf0, stages });
+    return { done: true };
+  }
+  return { done: false, order };
+}