mslxdff 0.1.88 → 0.1.90

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mslxdff",
-  "version": "0.1.88",
+  "version": "0.1.90",
   "description": "测试项目，请勿使用。",
   "type": "module",
   "bin": {

package/src/bench/cline-bench.js

@@ -1,54 +1,42 @@
 import { clineHeaders } from "../providers/cline/headers.js";
 import { computeMetrics } from "../metrics.js";
+import { createTransport } from "../transport/index.js";
+
+function sseContent(obj) {
+  const c = obj?.choices?.[0]?.delta?.content || obj?.choices?.[0]?.message?.content || "";
+  return typeof c === "string" ? c : "";
+}
 
 export async function clineBenchOne({ baseUrl, model, accessToken, prompt, maxTokens, timeoutMs, fetchImpl }) {
-  const controller = new AbortController();
-  const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
-  const t0 = performance.now();
+  const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
   let ttfbMs = null;
+  let totalMs = null;
   let content = "";
   try {
-    const res = await fetchImpl(`${baseUrl}/api/v1/chat/completions`, {
+    const sid = `sess_bench_${Date.now()}`;
+    const res = await tr.request({
+      url: `${baseUrl}/api/v1/chat/completions`,
       method: "POST",
-      headers: { ...clineHeaders(`sess_bench_${Date.now()}`, accessToken), Accept: "text/event-stream" },
-      body: JSON.stringify({ model, messages: [{ role: "user", content: prompt }], stream: true, max_tokens: maxTokens, session_id: `sess_bench_${Date.now()}`, reasoning_effort: "high" }),
-      signal: controller.signal,
+      headers: { ...clineHeaders(sid, accessToken), Accept: "text/event-stream" },
+      body: { model, messages: [{ role: "user", content: prompt }], stream: true, max_tokens: maxTokens, session_id: sid, reasoning_effort: "high" },
+      stream: true,
     });
-    if (res instanceof Error) throw res;
     if (!res.ok) {
       let txt = "";
       try { txt = await res.text(); } catch {}
       const label = res.status === 401 ? "鉴权失败" : res.status === 429 ? "限流" : res.status >= 500 ? `上游错误 ${res.status}` : `HTTP ${res.status}`;
-      return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: Math.round(performance.now() - t0), tps: null, charsPerSec: null, tokens: null };
+      return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: res.totalMs, tps: null, charsPerSec: null, tokens: null };
     }
-    const reader = res.body.getReader();
-    const decoder = new TextDecoder();
-    let buf = "";
-    for (;;) {
-      const { done, value } = await reader.read();
-      if (done) break;
-      if (ttfbMs === null) ttfbMs = Math.round(performance.now() - t0);
-      buf += decoder.decode(value, { stream: true });
-      let idx;
-      while ((idx = buf.indexOf("\n")) >= 0) {
-        const line = buf.slice(0, idx);
-        buf = buf.slice(idx + 1);
-        if (!line.startsWith("data:")) continue;
-        const payload = line.slice(5).trim();
-        if (!payload || payload === "[DONE]") continue;
-        try {
-          const j = JSON.parse(payload);
-          const c = j?.choices?.[0]?.delta?.content || j?.choices?.[0]?.message?.content || "";
-          if (typeof c === "string") content += c;
-        } catch {}
-      }
+    for await (const ev of res.stream()) {
+      try { content += sseContent(JSON.parse(ev)); } catch {}
     }
-    const totalMs = Math.round(performance.now() - t0);
+    ttfbMs = res.ttfbMs;
+    totalMs = res.totalMs;
     const chars = content.length;
     const { tps, charsPerSec } = computeMetrics({ ttfbMs, totalMs, promptTokens: null, completionTokens: null, chars });
     return { id: model, ok: true, status: 200, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: null, chars };
   } catch (e) {
     const msg = e?.message || String(e);
-    return { id: model, ok: false, label: /timeout|abort/i.test(msg) ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs: Math.round(performance.now() - t0), tps: null, charsPerSec: null, tokens: null };
-  } finally { clearTimeout(timer); }
+    return { id: model, ok: false, label: /timeout|abort/i.test(msg) ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+  }
 }

package/src/bench/probe.js

@@ -1,9 +1,9 @@
 import { joinUrl } from "../providers/base.js";
 import { getCustomNormalizer } from "../providers/registry.js";
+import { createTransport } from "../transport/index.js";
 
 function normalizeModelsPayload(json, baseUrl = "") {
   if (!json) return [];
-  // 可扩展：优先走注册表的定制化解析（如 cline.bot 仅 free）
   try {
     const custom = getCustomNormalizer(baseUrl);
     if (custom) {
@@ -46,27 +46,20 @@ export async function probeModels({
     if (!seen.has(norm)) { seen.add(norm); candidates.push(norm); }
   };
   if (modelsPath) push(modelsPath);
-  // 通用回退：/v1/models 与 /models
   push("/v1/models");
   push("/models");
-  // workbuddy 异形已在 defaultModelsPath 注入，若仍未命中则 candidates 已含
+  const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
   const tried = [];
   let lastError = "";
   for (const p of candidates) {
     const url = joinUrl(base, p);
     tried.push(url);
     try {
-      const controller = new AbortController();
-      const t = setTimeout(() => controller.abort(new Error(`probe timeout ${timeoutMs}ms`)), timeoutMs);
-      let res;
-      try {
-        res = await fetchImpl(url, { headers, signal: controller.signal });
-      } finally { clearTimeout(t); }
-      if (res instanceof Error) { lastError = res.message || String(res); continue; }
+      const res = await tr.request({ url, method: "GET", headers, stream: false });
       if (!res.ok) { lastError = `HTTP ${res.status}`; continue; }
       let json = {};
       try { json = await res.json(); } catch { json = {}; }
-      const raw = normalizeModelsPayload(json);
+      const raw = normalizeModelsPayload(json, baseUrl);
       const data = raw.map((m) => ({ id: toModelId(m), raw: m })).filter((x) => x.id).map((x) => ({ id: x.id, ...x.raw }));
       return { ok: true, data, tried, url, rawCount: raw.length };
     } catch (e) {

package/src/bench/runner.js

@@ -1,5 +1,6 @@
 import { joinUrl } from "../providers/base.js";
 import { computeMetrics, extractUsageFromJson } from "../metrics.js";
+import { createTransport } from "../transport/index.js";
 
 function extractInnerMessage(bodyText) {
   const t = String(bodyText || "");
@@ -37,43 +38,28 @@ export async function runOne({
   maxTokens = 32,
   timeoutMs = 30000,
   fetchImpl = globalThis.fetch,
-  clock = Date.now,
 } = {}) {
-  const started = clock();
-  const t0 = typeof performance !== "undefined" && performance.now ? performance.now() : started;
   if (!baseUrl) return { id: model, ok: false, error: "missing baseUrl", label: "配置错误", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
   if (!model) return { id: model, ok: false, error: "missing model", label: "配置错误", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
   if (!apiKey) return { id: model, ok: false, error: "missing apiKey", label: "未配置 Key", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
   const url = joinUrl(String(baseUrl).replace(/\/+$/, ""), chatPath);
   let rawModel = String(model || "").trim();
-  // allowlist 已是 raw（如 z-ai/glm-5.3-flash），仅当带供应商前缀时才剥
   if (providerId && rawModel.startsWith(`${providerId}/`)) rawModel = rawModel.slice(providerId.length + 1);
   const body = { model: rawModel, stream: false, messages: [{ role: "user", content: prompt }], max_tokens: maxTokens };
-  // workbuddy 需额外 workbuddy 透传头由调用方 headers 注入；此处直接透传
   const finalHeaders = { "Content-Type": "application/json", Accept: "application/json", ...headers };
   if (apiKey && !finalHeaders.Authorization) finalHeaders.Authorization = `Bearer ${apiKey}`;
-
-  let ttfbMs = null;
-  let res;
+  const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
   try {
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
-    const fetchStart = typeof performance !== "undefined" && performance.now ? performance.now() : clock();
-    try {
-      res = await fetchImpl(url, { method: "POST", headers: finalHeaders, body: JSON.stringify(body), signal: controller.signal });
-    } finally { clearTimeout(timer); }
-    ttfbMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - fetchStart);
-    if (res instanceof Error) throw res;
-    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+    const res = await tr.request({ url, method: "POST", headers: finalHeaders, body, stream: false });
+    const ttfbMs = res.ttfbMs;
+    const totalMs = res.totalMs;
     if (!res.ok) {
-      let txt = "";
-      try { txt = await res.text(); } catch {}
+      let txt = ""; try { txt = await res.text(); } catch {}
       const cls = classifyError(res.status, txt);
       const msg = extractInnerMessage(txt) || `HTTP ${res.status}`;
       return { id: model, providerId, ok: false, status: res.status, error: msg, label: cls.label, ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
     }
-    let json = {};
-    let txt = "";
+    let json = {}; let txt = "";
     try { txt = await res.text(); json = JSON.parse(txt); } catch { json = {}; }
     const usage = extractUsageFromJson(json);
     const content = json?.choices?.[0]?.message?.content || json?.choices?.[0]?.text || txt || "";
@@ -84,9 +70,8 @@ export async function runOne({
     const totalTokens = usage?.total_tokens ?? (promptTokens !== null && completionTokens !== null ? promptTokens + completionTokens : null);
     return { id: model, providerId, ok: true, status: res.status, ttfbMs, totalMs, tps, charsPerSec, tokens: { prompt: promptTokens, completion: completionTokens, total: totalTokens }, chars, label: "成功", raw: json };
   } catch (e) {
-    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
     const msg = e?.message || String(e);
     const isTimeout = /timeout|abort/i.test(msg);
-    return { id: model, providerId, ok: false, error: msg.slice(0, 300), label: isTimeout ? "超时" : "网络错误", ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+    return { id: model, providerId, ok: false, error: msg.slice(0, 300), label: isTimeout ? "超时" : "网络错误", ttfbMs: null, totalMs: null, tps: null, charsPerSec: null, tokens: null };
   }
 }

package/src/bench/via-probe.js

@@ -1,5 +1,6 @@
 import { joinUrl } from "../providers/base.js";
 import { computeMetrics, extractUsageFromJson } from "../metrics.js";
+import { createTransport } from "../transport/index.js";
 
 function extractInnerMessage(bodyText) {
   const t = String(bodyText || "");
@@ -43,11 +44,9 @@ export async function viaProbe({
   relayBody,
   targetUrl,
 } = {}) {
-  const started = clock();
-  const t0 = typeof performance !== "undefined" && performance.now ? performance.now() : started;
   const base = String(peerUrl || "").replace(/\/+$/, "");
   if (!base) return { ok: false, label: "配置错误", error: "missing peerUrl", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
-  // 纯中继模式：A 把 targetUrl+headers+body 发给 B，B 原样 fetch 到上游（不查 B 本地 providerConfigs）
+  const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
   const rt = String(relayTarget || targetUrl || "").trim();
   if (rt) {
     const rh = relayHeaders && typeof relayHeaders === "object" ? relayHeaders : {};
@@ -56,15 +55,10 @@ export async function viaProbe({
     const relayHeadersOut = { "Content-Type": "application/json", Accept: "application/json" };
     if (token) relayHeadersOut.Authorization = `Bearer ${token}`;
     const payload = { targetUrl: rt, method: "POST", headers: rh, body: rb };
-    let ttfbMs = null;
     try {
-      const controller = new AbortController();
-      const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
-      const fetchStart = typeof performance !== "undefined" && performance.now ? performance.now() : clock();
-      let res;
-      try { res = await fetchImpl(relayUrl, { method: "POST", headers: relayHeadersOut, body: JSON.stringify(payload), signal: controller.signal }); } finally { clearTimeout(timer); }
-      ttfbMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - fetchStart);
-      const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+      const res = await tr.request({ url: relayUrl, method: "POST", headers: relayHeadersOut, body: payload, stream: false });
+      const ttfbMs = res.ttfbMs;
+      const totalMs = res.totalMs;
       if (!res.ok) {
         let txt = ""; try { txt = await res.text(); } catch {}
         const cls = classifyError(res.status, txt);
@@ -88,10 +82,9 @@ export async function viaProbe({
       const totalTokens = usage?.total_tokens ?? (pt !== null && ct !== null ? pt + ct : null);
       return { ok: true, status: relayStatus, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: { prompt: pt, completion: ct, total: totalTokens }, chars };
     } catch (e) {
-      const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
       const msg = e?.message || String(e);
       const isTimeout = /timeout|abort/i.test(msg);
-      return { ok: false, label: isTimeout ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+      return { ok: false, label: isTimeout ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs: null, totalMs: null, tps: null, charsPerSec: null, tokens: null };
     }
   }
   if (!model) return { ok: false, label: "配置错误", error: "missing model", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
@@ -103,27 +96,17 @@ export async function viaProbe({
   if (token) headers.Authorization = `Bearer ${token}`;
   const sk = shareKeysHeader || shareKeys;
   if (sk) headers["x-mslxdff-share-keys"] = String(sk);
-  let ttfbMs = null;
   try {
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
-    const fetchStart = typeof performance !== "undefined" && performance.now ? performance.now() : clock();
-    let res;
-    try {
-      res = await fetchImpl(url, { method: "POST", headers, body: JSON.stringify(body), signal: controller.signal });
-    } finally { clearTimeout(timer); }
-    ttfbMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - fetchStart);
-    if (res instanceof Error) throw res;
-    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+    const res = await tr.request({ url, method: "POST", headers, body, stream: false });
+    const ttfbMs = res.ttfbMs;
+    const totalMs = res.totalMs;
     if (!res.ok) {
-      let txt = "";
-      try { txt = await res.text(); } catch {}
+      let txt = ""; try { txt = await res.text(); } catch {}
       const cls = classifyError(res.status, txt);
       const msg = extractInnerMessage(txt) || `HTTP ${res.status}`;
       return { ok: false, status: res.status, label: cls.label, error: msg, ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
     }
-    let json = {};
-    let txt = "";
+    let json = {}; let txt = "";
     try { txt = await res.text(); json = JSON.parse(txt); } catch { json = {}; }
     const usage = extractUsageFromJson(json);
     const content = json?.choices?.[0]?.message?.content || json?.choices?.[0]?.text || txt || "";
@@ -134,9 +117,8 @@ export async function viaProbe({
     const totalTokens = usage?.total_tokens ?? (promptTokens !== null && completionTokens !== null ? promptTokens + completionTokens : null);
     return { ok: true, status: res.status, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: { prompt: promptTokens, completion: completionTokens, total: totalTokens }, chars };
   } catch (e) {
-    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
     const msg = e?.message || String(e);
     const isTimeout = /timeout|abort/i.test(msg);
-    return { ok: false, label: isTimeout ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+    return { ok: false, label: isTimeout ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs: null, totalMs: null, tps: null, charsPerSec: null, tokens: null };
   }
 }

package/src/bench/workbuddy-bench.js

@@ -1,4 +1,5 @@
 import { computeMetrics } from "../metrics.js";
+import { createTransport } from "../transport/index.js";
 
 function buildWorkbuddyHeaders(apiKey, auth) {
   const h = {
@@ -16,55 +17,38 @@ function buildWorkbuddyHeaders(apiKey, auth) {
   return h;
 }
 
+function sseContent(obj) {
+  const c = obj?.choices?.[0]?.delta?.content || obj?.choices?.[0]?.message?.content || "";
+  return typeof c === "string" ? c : "";
+}
+
 export async function workbuddyBenchOne({ baseUrl, chatPath = "/v2/chat/completions", model, apiKey, auth, prompt = "hi", maxTokens = 5, timeoutMs = 30000, fetchImpl = globalThis.fetch }) {
-  const controller = new AbortController();
-  const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
-  const t0 = typeof performance !== "undefined" && performance.now ? performance.now() : Date.now();
+  const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
   let ttfbMs = null;
+  let totalMs = null;
   let content = "";
   try {
     const url = String(baseUrl).replace(/\/+$/, "") + String(chatPath || "/v2/chat/completions");
     const headers = buildWorkbuddyHeaders(apiKey, auth);
     const rawModel = String(model || "").trim();
     const body = { model: rawModel, stream: true, messages: [{ role: "user", content: prompt }], max_tokens: maxTokens };
-    const res = await fetchImpl(url, { method: "POST", headers, body: JSON.stringify(body), signal: controller.signal });
-    if (res instanceof Error) throw res;
+    const res = await tr.request({ url, method: "POST", headers, body, stream: true });
     if (!res.ok) {
       let txt = "";
       try { txt = await res.text(); } catch {}
       const label = res.status === 401 ? "鉴权失败" : res.status === 402 ? "余额不足" : res.status === 429 ? "限流" : res.status >= 500 ? `上游错误 ${res.status}` : `HTTP ${res.status}`;
-      return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : Date.now()) - t0), tps: null, charsPerSec: null, tokens: null };
+      return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: res.totalMs, tps: null, charsPerSec: null, tokens: null };
     }
-    // SSE stream parsing — 复用 cline 逻辑
-    const reader = res.body.getReader();
-    const decoder = new TextDecoder();
-    let buf = "";
-    for (;;) {
-      const { done, value } = await reader.read();
-      if (done) break;
-      if (ttfbMs === null) ttfbMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : Date.now()) - t0);
-      buf += decoder.decode(value, { stream: true });
-      let idx;
-      while ((idx = buf.indexOf("\n")) >= 0) {
-        const line = buf.slice(0, idx);
-        buf = buf.slice(idx + 1);
-        if (!line.startsWith("data:")) continue;
-        const payload = line.slice(5).trim();
-        if (!payload || payload === "[DONE]") continue;
-        try {
-          const j = JSON.parse(payload);
-          const c = j?.choices?.[0]?.delta?.content || j?.choices?.[0]?.message?.content || "";
-          if (typeof c === "string") content += c;
-        } catch {}
-      }
+    for await (const ev of res.stream()) {
+      try { content += sseContent(JSON.parse(ev)); } catch {}
     }
-    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : Date.now()) - t0);
+    ttfbMs = res.ttfbMs;
+    totalMs = res.totalMs;
     const chars = content.length;
     const { tps, charsPerSec } = computeMetrics({ ttfbMs, totalMs, promptTokens: null, completionTokens: null, chars });
     return { id: model, ok: true, status: 200, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: null, chars };
   } catch (e) {
     const msg = e?.message || String(e);
-    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : Date.now()) - t0);
     return { id: model, ok: false, label: /timeout|abort/i.test(msg) ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
-  } finally { clearTimeout(timer); }
+  }
 }

package/src/chat/engine.js

@@ -0,0 +1,160 @@
+import { performance } from "node:perf_hooks";
+import { estimateChars, needsCompress } from "./store.js";
+import { CHAT_KEEP_RECENT, CHAT_MAX_TOOL_LOOPS } from "./config.js";
+import { buildDedupKey, runTool } from "./tool-handlers.js";
+
+// 纯引擎：无 readline/ANSI/spinner，单一 runTurn 可测
+export function createEngine({
+  chatWithFallback,
+  summarizeHistory,
+  getToolDefs = () => [],
+  execCommand,
+  readFileTool,
+  curlTool,
+  onTrace = () => {},
+  config = {},
+} = {}) {
+  const CWF = chatWithFallback || (async () => ({ ok: false, error: "no chat" }));
+  const SUM = summarizeHistory || (async () => null);
+  const GTOOLS = getToolDefs;
+  const EXEC = execCommand || (async () => ({ ok: false, output: "no exec" }));
+  const READ = readFileTool || (async () => ({ ok: false, output: "no read" }));
+  const CURL = curlTool || (async () => ({ ok: false, output: "no curl" }));
+
+  async function maybeCompress(messages) {
+    if (!needsCompress(messages)) return [...messages];
+    const sys = messages[0];
+    const rest = messages.slice(1);
+    if (rest.length <= CHAT_KEEP_RECENT + 2) return [...messages];
+    const t0 = performance.now();
+    const toSummarize = rest.slice(0, -CHAT_KEEP_RECENT);
+    const keep = rest.slice(-CHAT_KEEP_RECENT);
+    const chars = estimateChars(messages);
+    onTrace(`[压缩] 触发 ${chars}字 > 400000阈值 · 待压 ${toSummarize.length}条 保留 ${keep.length}条`);
+    const summary = await SUM(toSummarize);
+    const dt = Math.round(performance.now() - t0);
+    if (!summary) {
+      onTrace(`[压缩] 失败/空 · ${dt}ms → 截断`);
+      return [sys, ...keep];
+    }
+    const summaryMsg = { role: "system", content: summary };
+    const next = [sys, summaryMsg, ...keep];
+    onTrace(`[压缩] 完成 ${toSummarize.length}条→${summary.length}字 · ${dt}ms · 新总量约 ${estimateChars(next)}字`);
+    return next;
+  }
+
+  async function runTurn(userText, messages) {
+    const tools = GTOOLS();
+    messages.push({ role: "user", content: userText });
+    let loops = 0;
+    let lastModel = null;
+    let lastProvider = null;
+    let lastUsage = null;
+    let lastFallback = false;
+    let lastFallbackGateway = false;
+    let lastLatency = 0;
+    const t0 = performance.now();
+    const turnStart = performance.now();
+    const seenCalls = new Map();
+    let duplicateStrikes = 0;
+    let forceNoTools = false;
+    onTrace(`[turn] 开始 "${userText.slice(0, 60)}${userText.length > 60 ? "…" : ""}" · 历史 ${messages.length}条 约 ${estimateChars(messages)}字`);
+    while (loops < CHAT_MAX_TOOL_LOOPS) {
+      const tLoop = performance.now();
+      const tComp = performance.now();
+      const cur = await maybeCompress(messages);
+      const compressMs = Math.round(performance.now() - tComp);
+      if (compressMs > 50) onTrace(`[loop ${loops}] 压缩耗时 ${compressMs}ms`);
+      messages.length = 0;
+      for (const m of cur) messages.push(m);
+      const tCall = performance.now();
+      let res;
+      const activeTools = forceNoTools ? [] : tools;
+      res = await CWF({ messages, tools: activeTools });
+      if (forceNoTools && res.ok && res.message?.tool_calls?.length) {
+        onTrace(`[guard] 禁工具模式下仍收到 tool_calls，已拦截`);
+        res.message.tool_calls = [];
+        if (!res.message.content) res.message.content = "（已拦截违规工具调用，请基于已有结果直接回答）";
+      }
+      const llmMs = Math.round(performance.now() - tCall);
+      onTrace(`[loop ${loops}] LLM ${llmMs}ms${compressMs > 50 ? ` (含压缩 ${compressMs}ms)` : ""} · ${estimateChars(messages)}字上下文`);
+      lastLatency = llmMs;
+      if (!res.ok) {
+        const err = `大模型暂不可用：${res.error}`;
+        messages.push({ role: "assistant", content: err });
+        return { text: err, model: null, latency: lastLatency, usage: null, fallback: false, ok: false };
+      }
+      lastModel = res.model;
+      lastProvider = res.provider || null;
+      lastUsage = res.usage || null;
+      lastFallback = !!res.fallback;
+      lastFallbackGateway = !!res.fallbackGateway || !!res.viaGateway;
+      const msg = res.message;
+      const toolCalls = msg.tool_calls || [];
+      let fallbackCmd = null;
+      if (!toolCalls.length && msg.content) {
+        const m = String(msg.content).match(/\{[^}]*"command"\s*:\s*"([^"]+)"[^}]*\}/);
+        if (m) fallbackCmd = m[1];
+      }
+      if (!toolCalls.length && !fallbackCmd) {
+        const text = String(msg.content || "").trim() || "(空回复)";
+        messages.push({ role: "assistant", content: text });
+        const totalMs = Math.round(performance.now() - t0);
+        let note = "";
+        if (lastFallbackGateway) note = "\n[注：mimo/big-pickle 均不可用，已自动切本地网关 auto（:8989）]";
+        else if (lastFallback) note = "\n[注：mimo 不可用，已用 big-pickle]";
+        onTrace(`[turn] 完成 总计 ${totalMs}ms · LLM ${lastLatency}ms · 0 工具${lastFallbackGateway ? " · gateway-fallback" : ""}`);
+        return { text: text + note, model: lastModel, provider: lastProvider, latency: lastLatency, usage: lastUsage, fallback: lastFallback, fallbackGateway: lastFallbackGateway, ok: true, totalMs };
+      }
+      const calls = toolCalls.length ? toolCalls : [{ id: "fallback-1", function: { name: "run_command", arguments: JSON.stringify({ command: fallbackCmd }) } }];
+      messages.push({ role: "assistant", content: msg.content || "", tool_calls: calls.map((c) => ({ id: c.id, type: "function", function: c.function })) });
+      onTrace(`[tools] 本轮 ${calls.length} 个调用 ${calls.map((c) => c.function?.name).join(",")} · 顺序执行`);
+      const tTools = performance.now();
+      const toolResults = [];
+      const toolDeps = { execCommand: EXEC, readFileTool: READ, curlTool: CURL, onTrace };
+      for (const c of calls) {
+        const name = c.function?.name;
+        let args = {};
+        try { args = JSON.parse(c.function?.arguments || "{}"); } catch {}
+        const t1 = performance.now();
+        const dedupKey = buildDedupKey(name, args);
+        const seen = seenCalls.get(dedupKey);
+        if (seen) {
+          const dt = Math.round(performance.now() - t1);
+          onTrace(`[tool] ${name} 重复调用已跳过 · ${dt}ms · 之前 ${seen.count} 次`);
+          toolResults.push({
+            id: c.id,
+            content: `SKIPPED_DUP: 此工具调用在本轮已执行过 ${seen.count} 次，结果相同请直接基于已有信息回答用户，不要再重复调用。\n--- 首次结果复用 ---\n${seen.firstResult.slice(0, 6000)}`,
+          });
+          continue;
+        }
+        const result = await runTool({ name, args, userText, ...toolDeps });
+        if (!seenCalls.has(dedupKey)) seenCalls.set(dedupKey, { count: 1, firstResult: result });
+        else seenCalls.get(dedupKey).count++;
+        toolResults.push({ id: c.id, content: result });
+      }
+      for (const tr of toolResults) messages.push({ role: "tool", tool_call_id: tr.id, content: tr.content });
+      if (toolResults.some((tr) => String(tr.content).startsWith("SKIPPED_DUP"))) {
+        duplicateStrikes++;
+        forceNoTools = true;
+        messages.push({ role: "system", content: "系统提示：你已重复调用相同工具，工具侧已复用首次结果并跳过执行。你已被禁止再调用任何工具，必须立即基于以上工具结果用中文直接回答用户，0 工具调用。" });
+        onTrace(`[dup] 检测到重复调用 ${duplicateStrikes} 次，已禁用后续工具调用`);
+        if (duplicateStrikes >= 2) {
+          const seen = [...seenCalls.values()].map((v) => v.firstResult).join("\n---\n").slice(0, 6000);
+          const synth = `检测到重复调用已达 ${duplicateStrikes} 次，为避免空转，直接基于已有结果回答：\n\n${seen}`;
+          messages.push({ role: "assistant", content: synth });
+          const totalMs = Math.round(performance.now() - t0);
+          onTrace(`[turn] 提前结束（重复阈值） 总计 ${totalMs}ms`);
+          return { text: synth, model: lastModel || "local", provider: lastProvider, latency: lastLatency, usage: lastUsage, fallback: lastFallback, ok: true, totalMs };
+        }
+      }
+      const toolsMs = Math.round(performance.now() - tTools);
+      const loopMs = Math.round(performance.now() - tLoop);
+      onTrace(`[loop ${loops}] 工具 ${toolsMs}ms · 本轮总 ${loopMs}ms · 累计 ${Math.round(performance.now() - turnStart)}ms`);
+      loops++;
+    }
+    return { text: "（工具调用次数已达上限，已停止）", model: lastModel, latency: lastLatency, usage: lastUsage, fallback: lastFallback, fallbackGateway: lastFallbackGateway, ok: false };
+  }
+
+  return { runTurn, maybeCompress };
+}