mslxdff 0.1.60 → 0.1.63

package/bin/mslxdff.js

@@ -232,8 +232,25 @@ if (args.includes("-model") || args.includes("-models")) {
     for (const id of picks) console.log(`  ${id}`);
     process.exit(0);
   }
+  // -model list supports optional provider filter: -model list --provider <id> | -model list <id> | --json
+  let modelListProvider = null;
+  let modelListJson = false;
+  if (sub === "list") {
+    const restArgs = args.slice(idx + 2);
+    for (let i = 0; i < restArgs.length; i++) {
+      const a = String(restArgs[i] || "");
+      if (a === "--json" || a === "-json") modelListJson = true;
+      else if (a === "--provider" || a === "-provider" || a === "--providerId") { modelListProvider = String(restArgs[i + 1] || "").trim() || null; i++; }
+      else if (!a.startsWith("-") && !modelListProvider) modelListProvider = a;
+    }
+    if (modelListProvider) {
+      const { normalizeProviderId } = await import("../src/providers/model-id.js");
+      const nid = normalizeProviderId(modelListProvider);
+      modelListProvider = nid || modelListProvider.toLowerCase();
+    }
+  }
   if (sub !== undefined && sub !== "list") {
-    console.error("usage: mslxdff -models (interactive multi-pick) | mslxdff -model list | mslxdff -model set <id> | mslxdff -model pick <id> | mslxdff -model unpick <id> | mslxdff -model pick clear | mslxdff -model picks | mslxdff -model status | mslxdff -model refresh");
+    console.error("usage: mslxdff -models (interactive multi-pick) | mslxdff -model list [--provider <id>] [--json] | mslxdff -model set <id> | mslxdff -model pick <id> | mslxdff -model unpick <id> | mslxdff -model pick clear | mslxdff -model picks | mslxdff -model status | mslxdff -model refresh");
     process.exit(1);
   }
   const cacheFile = join(logDir(), "models.json");
@@ -275,12 +292,32 @@ if (args.includes("-model") || args.includes("-models")) {
         throw new Error("no cached models and refresh failed");
       }
     }
+    // provider filter: bare ids are opencode, prefixed are <provider>/...
+    if (modelListProvider) {
+      const prov = String(modelListProvider).toLowerCase();
+      ids = ids.filter((id) => {
+        const slash = String(id).indexOf("/");
+        const p = slash > 0 ? String(id).slice(0, slash).toLowerCase() : "opencode";
+        return p === prov;
+      });
+      if (modelListJson) {
+        console.log(JSON.stringify({ object: "list", data: ids.map((id) => ({ id, object: "model" })) }, null, 2));
+        process.exit(0);
+      }
+      if (!ids.length) {
+        console.log(`no models for provider "${prov}" — try: mslxdff -provider ${prov} models  or  mslxdff -model refresh`);
+        process.exit(0);
+      }
+    } else if (modelListJson) {
+      console.log(JSON.stringify({ object: "list", data: ids.map((id) => ({ id, object: "model" })) }, null, 2));
+      process.exit(0);
+    }
     if (!ids.length) {
       console.log("no models available — try: mslxdff -model refresh");
       process.exit(0);
     }
     // TTY：交互式多选勾选常用模型（空格勾选，Enter 保存）；非 TTY（管道/脚本）：保持纯列表并标注勾选
-    if (process.stdin.isTTY && process.stdout.isTTY) {
+    if (process.stdin.isTTY && process.stdout.isTTY && !modelListProvider) {
       const statuses = loadModelErrors();
       const current = getPreferredModel();
       const pickedIds = loadModelPicks();
@@ -809,10 +846,24 @@ if (args.includes("-provider") || args.includes("--provider")) {
     const nid = normalizeProviderId(gid);
     if (!nid) { console.error(`invalid provider id: ${gid}`); process.exit(1); }
     if (!/^https?:\/\/.+/.test(String(gBase).trim())) { console.error(`invalid baseUrl: ${gBase} (must start with http:// or https://)`); process.exit(1); }
-    const cur = loadProviderConfig(nid) || { baseUrl: "", keys: [], allowedModels: [], auths: [] };
+    const cur = loadProviderConfig(nid) || { baseUrl: "", keys: [], allowedModels: [], auths: [], modelsPath: "", chatPath: "" };
     let keys, auths, baseUrl;
     baseUrl = String(gBase).trim();
-    const extraModels = rest.slice(3).filter((x) => x && !String(x).startsWith("-")).map((m) => String(m).trim()).filter(Boolean);
+    // parse --models-path / --chat-path from tail
+    let parsedModelsPath = null;
+    let parsedChatPath = null;
+    const extraTokens = [];
+    for (let _i = 3; _i < rest.length; _i++) {
+      const tok = String(rest[_i] || "");
+      if (tok === "--models-path" || tok === "--modelsPath" || tok === "--models_path") { parsedModelsPath = String(rest[_i + 1] || "").trim() || null; _i++; }
+      else if (tok.startsWith("--models-path=")) { parsedModelsPath = tok.slice("--models-path=".length).trim() || null; }
+      else if (tok === "--chat-path" || tok === "--chatPath" || tok === "--chat_path") { parsedChatPath = String(rest[_i + 1] || "").trim() || null; _i++; }
+      else if (tok.startsWith("--chat-path=")) { parsedChatPath = tok.slice("--chat-path=".length).trim() || null; }
+      else extraTokens.push(tok);
+    }
+    if (parsedModelsPath && !String(parsedModelsPath).startsWith("/")) { console.error(`invalid --models-path: ${parsedModelsPath} (must start with /)`); process.exit(1); }
+    if (parsedChatPath && !String(parsedChatPath).startsWith("/")) { console.error(`invalid --chat-path: ${parsedChatPath} (must start with /)`); process.exit(1); }
+    const extraModels = extraTokens.filter((x) => x && !String(x).startsWith("-")).map((m) => String(m).trim()).filter(Boolean);
     const allowedModels = extraModels.length ? [...new Set([...(cur.allowedModels || []), ...extraModels])] : (cur.allowedModels || []);
     if (nid === "workbuddy") {
       // workbuddy: keys/auths 一一对应，需解析 uid
@@ -838,7 +889,12 @@ if (args.includes("-provider") || args.includes("--provider")) {
         while (newKeys.length < newAuths.length) newKeys.push(token);
       }
       keys = newKeys; auths = newAuths;
-      saveProviderConfig(nid, { baseUrl, keys, auths, allowedModels });
+      const cfgToSave = { baseUrl, keys, auths, allowedModels };
+      if (parsedModelsPath) cfgToSave.modelsPath = parsedModelsPath;
+      else if (cur.modelsPath) cfgToSave.modelsPath = cur.modelsPath;
+      if (parsedChatPath) cfgToSave.chatPath = parsedChatPath;
+      else if (cur.chatPath) cfgToSave.chatPath = cur.chatPath;
+      saveProviderConfig(nid, cfgToSave);
       // 同步写 auths/workbuddy-<uid>.json 供 checkin 使用
       try {
         const { writeFileSync, mkdirSync } = await import("node:fs");
@@ -860,7 +916,12 @@ if (args.includes("-provider") || args.includes("--provider")) {
       }
       keys = [...new Set([...(cur.keys || []), trimmed].filter(Boolean))];
       auths = undefined;
-      saveProviderConfig(nid, { baseUrl, keys, allowedModels });
+      const cfgToSave2 = { baseUrl, keys, allowedModels };
+      if (parsedModelsPath) cfgToSave2.modelsPath = parsedModelsPath;
+      else if (cur.modelsPath) cfgToSave2.modelsPath = cur.modelsPath;
+      if (parsedChatPath) cfgToSave2.chatPath = parsedChatPath;
+      else if (cur.chatPath) cfgToSave2.chatPath = cur.chatPath;
+      saveProviderConfig(nid, cfgToSave2);
     }
     console.log(`added generic provider: ${nid}`);
     console.log(`  baseUrl: ${String(gBase).trim().replace(/\/+$/, "")}`);
@@ -880,6 +941,40 @@ if (args.includes("-provider") || args.includes("--provider")) {
     process.exit(0);
   }
   const { loadProviderKeys, saveProviderKeys, addProviderKey, removeProviderKeys, loadProviderShareKeys, saveProviderShareKeys, loadProviderConfig, saveProviderConfig, saveProviderBaseUrl, loadProviderConfigs } = await import("../src/state.js");
+  if (sub === "set-models-path" || sub === "setModelsPath" || sub === "models-path") {
+    const p = rest[1];
+    if (!p) {
+      console.error(`usage: mslxdff -provider ${id} set-models-path <path>  (e.g. /v1/models)`);
+      process.exit(1);
+    }
+    if (!String(p).trim().startsWith("/")) {
+      console.error(`invalid modelsPath: ${p} (must start with /)`);
+      process.exit(1);
+    }
+    const cur = loadProviderConfig(id) || { baseUrl: "", keys: [] };
+    const { normalizeProviderId } = await import("../src/providers/model-id.js");
+    const nid = normalizeProviderId(id);
+    saveProviderConfig(nid || id, { baseUrl: cur.baseUrl || "", keys: cur.keys || [], modelsPath: String(p).trim() });
+    console.log(`set ${nid || id} modelsPath: ${String(p).trim()} — restart daemon to activate`);
+    process.exit(0);
+  }
+  if (sub === "set-chat-path" || sub === "setChatPath" || sub === "chat-path") {
+    const p = rest[1];
+    if (!p) {
+      console.error(`usage: mslxdff -provider ${id} set-chat-path <path>  (e.g. /v1/chat/completions)`);
+      process.exit(1);
+    }
+    if (!String(p).trim().startsWith("/")) {
+      console.error(`invalid chatPath: ${p} (must start with /)`);
+      process.exit(1);
+    }
+    const cur = loadProviderConfig(id) || { baseUrl: "", keys: [] };
+    const { normalizeProviderId } = await import("../src/providers/model-id.js");
+    const nid = normalizeProviderId(id);
+    saveProviderConfig(nid || id, { baseUrl: cur.baseUrl || "", keys: cur.keys || [], chatPath: String(p).trim() });
+    console.log(`set ${nid || id} chatPath: ${String(p).trim()} — restart daemon to activate`);
+    process.exit(0);
+  }
   if (sub === "clear") {
     const configs = loadProviderConfigs();
     if (configs[id]) {
@@ -1020,6 +1115,69 @@ if (args.includes("-provider") || args.includes("--provider")) {
     console.error(`       mslxdff -provider ${id} allowAny on|off   (empty allowlist = block or allow all)`);
     process.exit(1);
   }
+  if (sub === "models" || sub === "list-models" || sub === "ls") {
+    const wantsJson = args.includes("--json") || args.includes("-json");
+    const cfg = loadProviderConfig(id);
+    // opencode: use aggregated cache + provider-specific? For opencode, show bare ids from cache
+    if (id === "opencode" || id === "oc") {
+      try {
+        const cacheFile = join(logDir(), "models.json");
+        const { readFileSync } = await import("node:fs");
+        const raw = JSON.parse(readFileSync(cacheFile, "utf8"));
+        const ids = (raw.data || []).map((m) => m.id).filter((x) => !String(x).includes("/"));
+        if (wantsJson) {
+          console.log(JSON.stringify({ object: "list", data: ids.map((id) => ({ id, object: "model" })) }, null, 2));
+        } else {
+          console.log(`opencode models (${ids.length}):`);
+          for (const mid of ids) console.log(`  ${mid}`);
+        }
+      } catch (e) {
+        console.error(`could not read models cache: ${String(e?.message || e)}`);
+        process.exit(1);
+      }
+      process.exit(0);
+    }
+    // generic/workbuddy: live fetch via provider listModels
+    try {
+      const { createGenericProvider } = await import("../src/providers/generic.js");
+      const { createWorkbuddyProvider } = await import("../src/providers/workbuddy.js");
+      const { isModelAllowed } = await import("../src/state.js");
+      const baseUrl = cfg?.baseUrl || (id === "workbuddy" ? "https://copilot.tencent.com" : "");
+      const keys = loadProviderKeys(id);
+      const auths = cfg?.auths || [];
+      let provider;
+      if (id === "workbuddy") {
+        provider = createWorkbuddyProvider({ baseUrl, apiKeys: keys, auths, file: defaultStateFile() });
+      } else {
+        if (!baseUrl) {
+          console.error(`provider ${id}: missing baseUrl — set via: mslxdff -provider ${id} set-url <baseUrl>`);
+          process.exit(1);
+        }
+        provider = createGenericProvider({ id, baseUrl, apiKeys: keys, file: defaultStateFile() });
+      }
+      const all = await provider.listModels();
+      // apply allowlist filtering mirroring dispatcher
+      const filtered = all.filter((m) => {
+        const raw = String(m.id || "").includes("/") ? String(m.id).split("/").slice(1).join("/") : String(m.id);
+        // for workbuddy, raw is like hy3, for generic it's raw id without prefix? joinModelId adds prefix, so need to extract raw
+        const checkRaw = m.id.startsWith(`${id}/`) ? m.id.slice(id.length + 1) : raw;
+        return isModelAllowed(id, checkRaw);
+      });
+      if (wantsJson) {
+        console.log(JSON.stringify({ object: "list", data: filtered }, null, 2));
+      } else {
+        console.log(`${id} models (${filtered.length}${filtered.length !== all.length ? `/${all.length}` : ""}):`);
+        for (const m of filtered) console.log(`  ${m.id}`);
+        if (!filtered.length && all.length) console.log(`  (all ${all.length} filtered by allowlist — use: mslxdff -provider ${id} allowlist list)`);
+        if (!all.length) console.log(`  (no models — check baseUrl/keys or try: curl ${baseUrl}/models)`);
+      }
+      try { await provider.close?.(); } catch {}
+    } catch (e) {
+      console.error(`could not list ${id} models: ${String(e?.message || e)}`);
+      process.exit(1);
+    }
+    process.exit(0);
+  }
   if (sub === "list" || sub === "status") {
     const keys = loadProviderKeys(id);
     const cfg = loadProviderConfig(id);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mslxdff",
-  "version": "0.1.60",
+  "version": "0.1.63",
   "description": "测试项目，请勿使用。",
   "type": "module",
   "bin": {

package/src/chat/prompt.js

@@ -52,7 +52,7 @@ ${mini}
 - 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”：优先直接用上方“可用模型”按前缀过滤回答（如 bai/ 开头的即 bai 供应商），无需调工具；如需实时刷新，调 curl local/models（GET，自动带本机 token）看网关聚合列表，或 curl https://api.b.ai/v1/models（自动带 key）看上游全量。禁止为此调用 -showtoken（本机 token 已自动注入）。
+- 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 clinebot/*、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。

package/src/chat/repl.js

@@ -30,6 +30,36 @@ function trace(line) {
   console.log(`\x1b[90m· ${line}\x1b[0m`);
 }
 
+function extractProvFromQuery(text) {
+  const low = String(text || "").toLowerCase();
+  const known = ["workbuddy", "clinebot", "opencode", "bai", "openrouter", "poolside", "z-ai", "deepseek"];
+  for (const k of known) if (low.includes(k)) return k;
+  return null;
+}
+function isModelListQuery(text) {
+  const low = String(text || "").toLowerCase();
+  if (!low.includes("模型")) return false;
+  return low.includes("哪些") || low.includes("可用") || low.includes("支持") || low.includes("列表") || low.includes("有啥") || low.includes("都有") || low.includes("可以") || low.includes("用");
+}
+function formatModelAnswer(prov, models) {
+  const byProv = {};
+  for (const id of models) {
+    const slash = id.indexOf("/");
+    const p = slash > 0 ? id.slice(0, slash) : "opencode";
+    if (!byProv[p]) byProv[p] = [];
+    byProv[p].push(id);
+  }
+  if (prov) {
+    const list = byProv[prov] || models.filter((m) => m.toLowerCase().startsWith(prov.toLowerCase() + "/"));
+    if (!list.length) return `**${prov}** 暂无可用模型（可能未配置或网关未聚合）。可用总量 ${models.length}，按供应商：${Object.entries(byProv).map(([k, v]) => `${k}(${v.length})`).join(" | ")}`;
+    // 更友好：直接列 id 和调用方式
+    return `**${prov}** 可用模型（共 ${list.length} 个，网关已聚合）：\n\n| 模型 id | 调用方式 |\n|:---|:---|\n${list.map((m) => `| \`${m}\` | \`${m}\` |`).join("\n")}\n\n> 提示：直接用 \`${prov}/<模型>\` 调用，例如 \`${list[0]}\``;
+  }
+  // 无指定供应商：按分组汇总
+  const summary = Object.entries(byProv).map(([p, arr]) => `**${p}**(${arr.length})：${arr.slice(0, 8).join(", ")}${arr.length > 8 ? " …" : ""}`).join("\n");
+  return `可用模型总计 ${models.length} 个，按供应商分组：\n\n${summary}\n\n> 查某供应商请说“workbuddy有哪些模型”`;
+}
+
 async function maybeCompress(messages) {
   if (!needsCompress(messages)) return [...messages];
   const sys = messages[0];
@@ -55,6 +85,19 @@ async function maybeCompress(messages) {
 async function runAgentTurn(userText, messages) {
   const tools = getToolDefs();
   messages.push({ role: "user", content: userText });
+  // Fast-path：模型列表类问题本地直答，不走 LLM，避免 “provider list” 幻觉和 6 轮重复
+  if (isModelListQuery(userText)) {
+    const prov = extractProvFromQuery(userText);
+    try {
+      const models = getModelsForPrompt();
+      const answer = formatModelAnswer(prov, models);
+      if (answer) {
+        messages.push({ role: "assistant", content: answer });
+        trace(`[fast] 模型列表直答 prov=${prov || "all"} 共 ${models.length} 个`);
+        return { text: answer, model: "local", latency: 0, usage: null, fallback: false, ok: true, totalMs: 0 };
+      }
+    } catch {}
+  }
   let loops = 0;
   let lastModel = null;
   let lastUsage = null;
@@ -64,6 +107,8 @@ async function runAgentTurn(userText, messages) {
   const turnStart = performance.now();
   // 同轮去重：同一工具+参数只真正执行一次，重复直接复用并提示 LLM
   const seenCalls = new Map(); // key -> { count, firstResult }
+  let duplicateStrikes = 0;
+  let forceNoTools = false;
   trace(`[turn] 开始 "${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();
@@ -74,12 +119,19 @@ async function runAgentTurn(userText, messages) {
     messages.length = 0;
     for (const m of cur) messages.push(m);
     const tCall = performance.now();
-    const spinnerLabel = loops === 0 ? "已发送给 AI，等待回复中" : "AI 正在整理回复中";
+    const spinnerLabel = loops === 0 ? "已发送给 AI，等待回复中" : forceNoTools ? "AI 整理回答中（已禁工具）" : "AI 正在整理回复中";
     const spinner = createSpinner(spinnerLabel);
     spinner.start();
     let res;
     try {
-      res = await chatWithFallback({ messages, tools });
+      const activeTools = forceNoTools ? [] : tools;
+      res = await chatWithFallback({ messages, tools: activeTools });
+      if (forceNoTools && res.ok && res.message?.tool_calls?.length) {
+        // LLM 在禁工具模式下仍尝试调工具，视为违规，直接转文本
+        trace(`[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）";
+      }
     } finally {
       const ms = Math.round(performance.now() - tCall);
       spinner.stop(`\x1b[90m✓ AI 已回复 · ${ms}ms\x1b[0m`);
@@ -149,13 +201,18 @@ async function runAgentTurn(userText, messages) {
         console.log(`\x1b[90m→ 执行: mslxdff ${cmd}\x1b[0m`);
         const r = await execCommand(cmd);
         result = `${r.ok ? "OK" : "FAIL"}: ${r.output}`;
-        // 查询类命令直接在结果里植入“立即回答”锚点，降低 LLM 再发一次的概率
+        // 查询类命令直接在结果里植入“立即回答”锚点，降低 LLM 再发一次的概率；若用户问模型，则 provider list 不算答案
         const lowCmd = cmd.toLowerCase().replace(/\s+/g, " ").trim();
+        const asksModel = String(userText || "").toLowerCase().includes("模型");
         const isOnceAndDone =
           /^-+(showtoken|status|s|providers?\b|model\b|group\b|log\b|workbuddy\b|free\b|autostart\b|plugins\b)/.test(lowCmd) ||
           lowCmd === "-provider list" || lowCmd === "-providers list";
         if (isOnceAndDone && r.ok) {
-          result += `\n\n[系统提示：此查询已完成，结果即答案，请直接用中文回答用户，禁止再调用相同或同类查询工具]`;
+          if (asksModel && lowCmd.includes("-provider")) {
+            result += `\n\n[提示：此命令仅显示供应商配置，不包含模型列表。用户问的是“有哪些模型”，请用系统提示中的“可用模型”按前缀过滤回答，或调 curl local/models，不要再调 provider list]`;
+          } else {
+            result += `\n\n[系统提示：此查询已完成，结果即答案，请直接用中文回答用户，禁止再调用相同或同类查询工具]`;
+          }
         }
         const dt = Math.round(performance.now() - t1);
         trace(`[tool] run_command "${cmd.slice(0, 40)}" · ${dt}ms · ${r.ok ? "OK" : "FAIL"} ${r.output.length}字`);
@@ -185,10 +242,20 @@ async function runAgentTurn(userText, messages) {
       return { id: c.id, content: result };
     }));
     for (const tr of toolResults) messages.push({ role: "tool", tool_call_id: tr.id, content: tr.content });
-    // 若本轮有 SKIPPED_DUP，额外追加一条系统提示，强制 LLM 基于已有结果回答而不是再调工具
+    // 若本轮有 SKIPPED_DUP，额外追加系统提示并禁用后续工具，强制直接回答
     if (toolResults.some((tr) => String(tr.content).startsWith("SKIPPED_DUP"))) {
-      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。请直接基于以上工具结果用中文回答用户，不要再发起任何工具调用。" });
-      trace(`[dup] 检测到重复调用，已注入系统提示要求直接回答`);
+      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。" });
+      trace(`[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);
+        trace(`[turn] 提前结束（重复阈值） 总计 ${totalMs}ms`);
+        return { text: synth, model: lastModel || "local", latency: lastLatency, usage: lastUsage, fallback: lastFallback, ok: true, totalMs };
+      }
     }
     const toolsMs = Math.round(performance.now() - tTools);
     const loopMs = Math.round(performance.now() - tLoop);

package/src/providers/generic.js

@@ -1,6 +1,14 @@
 import { joinModelId } from "./model-id.js";
 import { createKeyRing } from "./keyring.js";
-import { loadProviderKeys, loadProviderBaseUrl } from "../state.js";
+import { loadProviderKeys, loadProviderBaseUrl, loadProviderModelsPath, loadProviderChatPath } from "../state.js";
+
+function joinUrl(base, path) {
+  const b = String(base || "").trim().replace(/\/+$/, "");
+  const p = String(path || "").trim();
+  if (!p) return b;
+  const pp = p.startsWith("/") ? p : `/${p}`;
+  return `${b}${pp}`;
+}
 
 let UndiciAgent = null;
 let UndiciFetch = null;
@@ -36,6 +44,8 @@ export function createGenericProvider({
   baseUrl,
   apiKeys,
   apiKey,
+  modelsPath,
+  chatPath,
   connectTimeoutMs = Number(process.env.MSLXDFF_GENERIC_TIMEOUT_MS) || 30_000,
   cooldownMs = envInt("MSLXDFF_GENERIC_COOLDOWN_MS", 30_000),
   retry = {
@@ -48,11 +58,14 @@ export function createGenericProvider({
   fetchImpl,
   headers: extraHeaders,
   noAgent = false,
+  file,
 } = {}) {
   if (!id) throw new Error("generic provider requires id");
   const resolvedBase = resolveBaseUrl(id, baseUrl);
   if (!resolvedBase) throw new Error(`generic provider ${id}: missing baseUrl`);
   if (!fetchImpl) fetchImpl = UndiciFetch || fetch;
+  const resolvedModelsPath = modelsPath || loadProviderModelsPath(id, file ? { file } : {});
+  const resolvedChatPath = chatPath || loadProviderChatPath(id, file ? { file } : {});
 
   const ring = createKeyRing(collectApiKeys(id, apiKeys, apiKey), { cooldownMs });
 
@@ -98,7 +111,7 @@ export function createGenericProvider({
   }
 
   async function runChat(body, activeRing, sourceKey) {
-    const url = `${resolvedBase}/chat/completions`;
+    const url = joinUrl(resolvedBase, resolvedChatPath);
     const t0 = performance.now();
     const attempts = [];
     let waitMs = 0;
@@ -156,7 +169,7 @@ export function createGenericProvider({
   async function listModels() {
     const now = Date.now();
     if (cache && now - fetchedAt < CACHE_TTL_MS) return cache;
-    const url = `${resolvedBase}/models`;
+    const url = joinUrl(resolvedBase, resolvedModelsPath);
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(new Error(`${id} models timed out`)), 15_000);
     try {
@@ -181,7 +194,7 @@ export function createGenericProvider({
   }
 
   async function preheat() {
-    const url = `${resolvedBase}/models`;
+    const url = joinUrl(resolvedBase, resolvedModelsPath);
     const t0 = performance.now();
     try {
       const headers = { Accept: "application/json" };

package/src/providers/workbuddy.js

@@ -1,8 +1,9 @@
 import { joinModelId } from "./model-id.js";
 import { createKeyRing } from "./keyring.js";
-import { loadProviderKeys, loadProviderAuths, loadProviderBaseUrl, loadProviderShareKeys, saveProviderConfig, WORKBUDDY_DEFAULT_BASE_URL } from "../state.js";
+import { loadProviderKeys, loadProviderAuths, loadProviderBaseUrl, loadProviderShareKeys, saveProviderConfig, WORKBUDDY_DEFAULT_BASE_URL, loadProviderModelsPath, loadProviderChatPath } from "../state.js";
 import { writeFileSync, existsSync, mkdirSync, readdirSync, readFileSync, appendFileSync, statSync } from "node:fs";
 import { join, dirname } from "node:path";
+import { tmpdir } from "node:os";
 import { getCachedBalance, setCachedBalance } from "./workbuddy-balance.js";
 
 let UndiciAgent = null;
@@ -13,10 +14,25 @@ try {
   UndiciFetch = mod.fetch;
 } catch {}
 
+function joinUrl(base, path) {
+  const b = String(base || "").trim().replace(/\/+$/, "");
+  const p = String(path || "").trim();
+  if (!p) return b;
+  const pp = p.startsWith("/") ? p : `/${p}`;
+  return `${b}${pp}`;
+}
 function envInt(name, fallback) {
   const v = Number(process.env[name]);
   return Number.isInteger(v) && v > 0 ? v : fallback;
 }
+function isTestEnv() {
+  if (process.env.NODE_ENV === "test") return true;
+  if (process.env.MSLXDFF_STATE_FILE && String(process.env.MSLXDFF_STATE_FILE).includes("mslxdff-test")) return true;
+  if (process.argv.some((a) => String(a).includes("--test") || String(a).endsWith(".test.js"))) return true;
+  if (Array.isArray(process.execArgv) && process.execArgv.some((a) => String(a).includes("--test"))) return true;
+  if (process.env.NODE_TEST_CONTEXT) return true;
+  return false;
+}
 
 function resolveBaseUrl(baseUrl) {
   if (baseUrl) return String(baseUrl).trim().replace(/\/+$/, "");
@@ -111,6 +127,8 @@ export function createWorkbuddyProvider({
   apiKeys,
   apiKey,
   auths,
+  modelsPath,
+  chatPath,
   connectTimeoutMs = Number(process.env.MSLXDFF_WORKBUDDY_TIMEOUT_MS) || 30_000,
   cooldownMs = envInt("MSLXDFF_WORKBUDDY_COOLDOWN_MS", 30_000),
   retry = {
@@ -125,6 +143,8 @@ export function createWorkbuddyProvider({
 } = {}) {
   const id = "workbuddy";
   const resolvedBase = resolveBaseUrl(baseUrl);
+  const resolvedModelsPath = modelsPath || loadProviderModelsPath(id, file ? { file } : {});
+  const resolvedChatPath = chatPath || loadProviderChatPath(id, file ? { file } : {});
   if (!fetchImpl) fetchImpl = UndiciFetch || fetch;
 
   // keys 优先显式传入，其次 state
@@ -143,7 +163,7 @@ export function createWorkbuddyProvider({
   // fallback scan auths/workbuddy-*.json if still empty
   if (!authList.length && !keys.length) {
     try {
-      const authDir = process.env.WORKBUDDY_AUTH_DIR || (file && String(file).includes("mslxdff-") ? join(dirname(String(file)), "auths") : join(process.cwd(), "auths"));
+      const authDir = process.env.WORKBUDDY_AUTH_DIR || (isTestEnv() ? join(tmpdir(), "mslxdff-test-auths") : (file && String(file).includes("mslxdff-") ? join(dirname(String(file)), "auths") : join(process.cwd(), "auths")));
       if (existsSync(authDir)) {
         const files = readdirSync(authDir).filter((f) => f.startsWith("workbuddy-") && f.endsWith(".json"));
         for (const f of files) {
@@ -192,7 +212,7 @@ export function createWorkbuddyProvider({
     const rt = auth?.refreshToken;
     const uid = auth?.uid;
     if (!rt || !uid) return null;
-    const url = `${resolvedBase}/v2/plugin/auth/token/refresh`;
+    const url = joinUrl(resolvedBase, "/v2/plugin/auth/token/refresh");
     const headers = {
       "Content-Type": "application/json",
       Authorization: `Bearer ${key}`,
@@ -223,7 +243,7 @@ export function createWorkbuddyProvider({
             saveProviderConfig(id, { baseUrl: resolvedBase, keys: [...keys], auths: [...authList] }, file ? { file } : {});
           } catch {}
           try {
-            const authDir = process.env.WORKBUDDY_AUTH_DIR || (file && String(file).includes("mslxdff-") ? join(dirname(String(file)), "auths") : join(process.cwd(), "auths"));
+            const authDir = process.env.WORKBUDDY_AUTH_DIR || (isTestEnv() ? join(tmpdir(), "mslxdff-test-auths") : (file && String(file).includes("mslxdff-") ? join(dirname(String(file)), "auths") : join(process.cwd(), "auths")));
             mkdirSync(authDir, { recursive: true });
             const expAt = (() => { try { return JSON.parse(Buffer.from(newAt.split(".")[1], "base64").toString()).exp; } catch { return Math.floor(Date.now()/1000)+5184000; } })();
             const doc = { account: { uid, enterpriseId: auth.enterpriseId || "", nickname: "" }, auth: { accessToken: newAt, refreshToken: newRt, expiresAt: expAt, domain: auth.domain || "www.codebuddy.cn" } };
@@ -259,7 +279,7 @@ export function createWorkbuddyProvider({
   }
 
   async function runChat(body, activeRing, opts = {}) {
-    const url = `${resolvedBase}/v2/chat/completions`;
+    const url = joinUrl(resolvedBase, resolvedChatPath);
     const t0 = performance.now();
     const preferredUid = opts?.workbuddyUid ? String(opts.workbuddyUid).trim() : "";
     const modelForLog = body?.model || "";
@@ -479,7 +499,7 @@ export function createWorkbuddyProvider({
   async function listModels() {
     const now = Date.now();
     if (cache && now - fetchedAt < CACHE_TTL_MS) return cache;
-    const url = `${resolvedBase}/console/enterprises/personal/models`;
+    const url = joinUrl(resolvedBase, resolvedModelsPath);
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(new Error(`${id} models timed out`)), 15_000);
     try {
@@ -514,7 +534,7 @@ export function createWorkbuddyProvider({
   }
 
   async function preheat() {
-    const url = `${resolvedBase}/console/enterprises/personal/models`;
+    const url = joinUrl(resolvedBase, resolvedModelsPath);
     const t0 = performance.now();
     try {
       const key = ring.next() || keys[0] || "";

package/src/routes/index.js

@@ -4,12 +4,21 @@ import { json, notFound, authorized } from "./helpers.js";
 import { chatHandler } from "./chat.js";
 import { joinHandler, leaveHandler } from "./groups.js";
 import { heartbeatHandler, pollHandler, resultHandler, forwardHandler } from "./groups-relay.js";
-import { modelsHandler, modelsStatusHandler } from "./models-route.js";
+import { modelsHandler, modelsStatusHandler, providerModelsHandler } from "./models-route.js";
 
 export function createRouter({ token, upstream, models, auto, logs, peers, maxHops = DEFAULT_MAX_HOPS, groups, bans, bus, plugins }) {
   return async function router(req, res) {
     const method = req.method || "GET";
     const path = (req.url || "").split("?")[0];
+    // dynamic provider models route: GET /v1/providers/:id/models
+    if (method === "GET" && /^\/v1\/providers\/[^/]+\/models\/?$/.test(path)) {
+      if (!authorized(req, token)) {
+        res.statusCode = 401;
+        res.setHeader("WWW-Authenticate", "Bearer");
+        return json(res, 401, { error: "Unauthorized" });
+      }
+      return providerModelsHandler({ req, res, models, upstream });
+    }
     const route = ROUTES.find((r) => r.method === method && r.path === path);
     if (!route) return notFound(res);
     if (route.requiresAuth && !authorized(req, token)) {

package/src/routes/models-route.js

@@ -1,5 +1,6 @@
 import { json, errMsg } from "./helpers.js";
 import { runHook } from "../plugins.js";
+import { isModelAllowed } from "../state.js";
 
 export async function modelsHandler({ res, models, plugins }) {
   if (!models) return json(res, 501, { error: "Models service not configured" });
@@ -42,3 +43,30 @@ export async function modelsStatusHandler({ res, models, auto }) {
   }
   json(res, 200, { object: "list", data });
 }
+
+export async function providerModelsHandler({ req, res, models }) {
+  if (!models) return json(res, 501, { error: "Models service not configured" });
+  const url = req.url || "";
+  const path = url.split("?")[0] || "";
+  const m = path.match(/^\/v1\/providers\/([^/]+)\/models\/?$/);
+  const pid = m ? decodeURIComponent(m[1]).toLowerCase() : "";
+  if (!pid) return json(res, 400, { error: "missing provider id" });
+  try {
+    const data = await models.get();
+    const all = Array.isArray(data?.data) ? data.data : [];
+    const filtered = all.filter((entry) => {
+      const id = String(entry?.id || "");
+      const slash = id.indexOf("/");
+      const prov = slash > 0 ? id.slice(0, slash).toLowerCase() : "opencode";
+      if (prov !== pid) return false;
+      const raw = slash > 0 ? id.slice(slash + 1) : id;
+      // allowlist check: opencode always allowed via allowAny, others respect config
+      try {
+        return isModelAllowed(pid, raw);
+      } catch { return true; }
+    });
+    json(res, 200, { object: "list", data: filtered });
+  } catch (err) {
+    json(res, 502, { error: errMsg(err) });
+  }
+}

package/src/state.js

@@ -6,9 +6,19 @@ import os from "node:os";
 
 export const DEFAULT_PORT = 8989;
 
+function isTestEnv() {
+  if (process.env.NODE_ENV === "test") return true;
+  if (process.env.MSLXDFF_STATE_FILE && String(process.env.MSLXDFF_STATE_FILE).includes("mslxdff-test")) return true;
+  if (process.argv.some((a) => String(a).includes("--test") || String(a).endsWith(".test.js"))) return true;
+  if (Array.isArray(process.execArgv) && process.execArgv.some((a) => String(a).includes("--test"))) return true;
+  // Node test runner sets this
+  if (process.env.NODE_TEST_CONTEXT) return true;
+  return false;
+}
+
 export function defaultStateFile() {
   if (process.env.MSLXDFF_STATE_FILE) return process.env.MSLXDFF_STATE_FILE;
-  if (process.argv.includes("--test") || process.env.NODE_ENV === "test") {
+  if (isTestEnv()) {
     return join(os.tmpdir(), "mslxdff-test-state.json");
   }
   return join(os.homedir(), ".config", "mslxdff", "state.json");
@@ -19,7 +29,7 @@ export function tokenFile(file) {
   const sf = file || defaultStateFile();
   const realDefault = join(os.homedir(), ".config", "mslxdff", "state.json");
   if (sf !== realDefault) return join(dirname(sf), "token");
-  if (process.argv.includes("--test") || process.env.NODE_ENV === "test") {
+  if (isTestEnv()) {
     return join(os.tmpdir(), "mslxdff-test-token");
   }
   return join(os.homedir(), ".config", "mslxdff", "token");
@@ -385,6 +395,29 @@ function normalizeAllowedModel(model, providerId) {
   }
   return s;
 }
+function normalizeEndpointPath(v) {
+  const s = String(v || "").trim();
+  if (!s) return "";
+  return s.startsWith("/") ? s : `/${s}`;
+}
+function defaultModelsPath(id) {
+  if (String(id).toLowerCase() === "workbuddy") return "/console/enterprises/personal/models";
+  return "/models";
+}
+function defaultChatPath(id) {
+  if (String(id).toLowerCase() === "workbuddy") return "/v2/chat/completions";
+  return "/chat/completions";
+}
+export function loadProviderModelsPath(id, { file = defaultStateFile() } = {}) {
+  const cfg = loadProviderConfigs({ file })[id];
+  if (cfg && typeof cfg.modelsPath === "string" && cfg.modelsPath.trim()) return normalizeEndpointPath(cfg.modelsPath);
+  return defaultModelsPath(id);
+}
+export function loadProviderChatPath(id, { file = defaultStateFile() } = {}) {
+  const cfg = loadProviderConfigs({ file })[id];
+  if (cfg && typeof cfg.chatPath === "string" && cfg.chatPath.trim()) return normalizeEndpointPath(cfg.chatPath);
+  return defaultChatPath(id);
+}
 
 export function loadProviderConfigs({ file = defaultStateFile() } = {}) {
   const v = readState(file).providerConfigs;
@@ -401,7 +434,9 @@ export function loadProviderConfig(id, { file = defaultStateFile() } = {}) {
     const cfg = loadProviderConfigs({ file })[id];
     const allowedModels = cfg && Array.isArray(cfg.allowedModels) ? [...new Set(cfg.allowedModels.map((m) => normalizeAllowedModel(m, id)).filter(Boolean))] : [];
     const auths = cfg && Array.isArray(cfg.auths) ? normalizeAuths(cfg.auths) : [];
-    if (baseUrl || envKeys.length || allowedModels.length || auths.length) return { baseUrl: normalizeBaseUrl(baseUrl), keys: envKeys, auths, allowedModels };
+    const modelsPath = cfg && typeof cfg.modelsPath === "string" && cfg.modelsPath.trim() ? normalizeEndpointPath(cfg.modelsPath) : defaultModelsPath(id);
+    const chatPath = cfg && typeof cfg.chatPath === "string" && cfg.chatPath.trim() ? normalizeEndpointPath(cfg.chatPath) : defaultChatPath(id);
+    if (baseUrl || envKeys.length || allowedModels.length || auths.length) return { baseUrl: normalizeBaseUrl(baseUrl), keys: envKeys, auths, allowedModels, modelsPath, chatPath };
     return null;
   }
   const configs = loadProviderConfigs({ file });
@@ -412,11 +447,13 @@ export function loadProviderConfig(id, { file = defaultStateFile() } = {}) {
       keys: Array.isArray(cfg.keys) ? [...new Set(cfg.keys.filter((x) => typeof x === "string" && x.trim().length))] : [],
       auths: normalizeAuths(cfg.auths),
       allowedModels: Array.isArray(cfg.allowedModels) ? [...new Set(cfg.allowedModels.map((m) => normalizeAllowedModel(m, id)).filter(Boolean))] : [],
+      modelsPath: typeof cfg.modelsPath === "string" && cfg.modelsPath.trim() ? normalizeEndpointPath(cfg.modelsPath) : defaultModelsPath(id),
+      chatPath: typeof cfg.chatPath === "string" && cfg.chatPath.trim() ? normalizeEndpointPath(cfg.chatPath) : defaultChatPath(id),
     };
   }
   // 兼容旧 providerKeys 形态：有 key 但无 configs 时视为通用供应商（baseUrl 为空，需后补）
   const keys = loadProviderKeys(id, { file });
-  if (keys.length) return { baseUrl: String(id).toLowerCase() === "workbuddy" ? WORKBUDDY_DEFAULT_BASE_URL : "", keys, auths: [], allowedModels: [] };
+  if (keys.length) return { baseUrl: String(id).toLowerCase() === "workbuddy" ? WORKBUDDY_DEFAULT_BASE_URL : "", keys, auths: [], allowedModels: [], modelsPath: defaultModelsPath(id), chatPath: defaultChatPath(id) };
   return null;
 }
 
@@ -441,12 +478,16 @@ export function saveProviderAuths(id, list, { file = defaultStateFile() } = {})
   const baseUrl = normalizeBaseUrl(cur.baseUrl || loadProviderBaseUrl(id, { file }) || "");
   const keys = Array.isArray(cur.keys) ? [...new Set(cur.keys.filter((x) => typeof x === "string" && x.trim().length))] : loadProviderKeys(id, { file });
   const allowedModels = Array.isArray(cur.allowedModels) ? [...new Set(cur.allowedModels.map((m) => normalizeAllowedModel(m, id)).filter(Boolean))] : [];
-  if (!baseUrl && !keys.length && !clean.length && !allowedModels.length) {
+  const modelsPath = typeof cur.modelsPath === "string" ? normalizeEndpointPath(cur.modelsPath) : "";
+  const chatPath = typeof cur.chatPath === "string" ? normalizeEndpointPath(cur.chatPath) : "";
+  if (!baseUrl && !keys.length && !clean.length && !allowedModels.length && !modelsPath && !chatPath) {
     delete configs[id];
   } else {
     configs[id] = { baseUrl, keys };
     if (clean.length) configs[id].auths = clean;
     if (allowedModels.length) configs[id].allowedModels = allowedModels;
+    if (modelsPath) configs[id].modelsPath = modelsPath;
+    if (chatPath) configs[id].chatPath = chatPath;
   }
   writeStateImmediate(file, { providerConfigs: configs });
   return clean;
@@ -459,33 +500,43 @@ export function saveProviderBaseUrl(id, baseUrl, { file = defaultStateFile() } =
   const keys = Array.isArray(cur.keys) ? cur.keys : loadProviderKeys(id, { file });
   const auths = normalizeAuths(cur.auths);
   const allowedModels = Array.isArray(cur.allowedModels) ? [...new Set(cur.allowedModels.map((m) => normalizeAllowedModel(m, id)).filter(Boolean))] : [];
-  if (!clean && !keys.length && !auths.length && !allowedModels.length) {
+  const modelsPath = typeof cur.modelsPath === "string" ? normalizeEndpointPath(cur.modelsPath) : "";
+  const chatPath = typeof cur.chatPath === "string" ? normalizeEndpointPath(cur.chatPath) : "";
+  if (!clean && !keys.length && !auths.length && !allowedModels.length && !modelsPath && !chatPath) {
     delete configs[id];
   } else {
     configs[id] = { baseUrl: clean, keys };
     if (auths.length) configs[id].auths = auths;
     if (allowedModels.length) configs[id].allowedModels = allowedModels;
+    if (modelsPath) configs[id].modelsPath = modelsPath;
+    if (chatPath) configs[id].chatPath = chatPath;
   }
   writeStateImmediate(file, { providerConfigs: configs });
   return clean;
 }
 
-export function saveProviderConfig(id, { baseUrl, keys, auths, allowedModels }, { file = defaultStateFile() } = {}) {
+export function saveProviderConfig(id, { baseUrl, keys, auths, allowedModels, modelsPath, chatPath }, { file = defaultStateFile() } = {}) {
   const cleanUrl = normalizeBaseUrl(baseUrl);
   const cleanKeys = [...new Set((Array.isArray(keys) ? keys : []).map((k) => String(k || "").trim()).filter(Boolean))];
   const cleanAuths = auths === undefined ? undefined : normalizeAuths(auths);
   const cleanAllowed = [...new Set((Array.isArray(allowedModels) ? allowedModels : []).map((m) => normalizeAllowedModel(m, id)).filter(Boolean))];
+  const cleanModelsPath = modelsPath === undefined ? undefined : (String(modelsPath).trim() ? normalizeEndpointPath(modelsPath) : "");
+  const cleanChatPath = chatPath === undefined ? undefined : (String(chatPath).trim() ? normalizeEndpointPath(chatPath) : "");
   const configs = { ...loadProviderConfigs({ file }) };
   const cur = configs[id] && typeof configs[id] === "object" ? configs[id] : {};
-  // 保留已有的 allowedModels / auths 若本次未传入
+  // 保留已有的 allowedModels / auths / paths 若本次未传入
   const finalAllowed = allowedModels === undefined ? (Array.isArray(cur.allowedModels) ? [...new Set(cur.allowedModels.map((m) => normalizeAllowedModel(m, id)).filter(Boolean))] : []) : cleanAllowed;
   const finalAuths = cleanAuths === undefined ? normalizeAuths(cur.auths) : cleanAuths;
-  if (!cleanUrl && !cleanKeys.length && !finalAllowed.length && !finalAuths.length) {
+  const finalModelsPath = cleanModelsPath === undefined ? (typeof cur.modelsPath === "string" ? normalizeEndpointPath(cur.modelsPath) : "") : cleanModelsPath;
+  const finalChatPath = cleanChatPath === undefined ? (typeof cur.chatPath === "string" ? normalizeEndpointPath(cur.chatPath) : "") : cleanChatPath;
+  if (!cleanUrl && !cleanKeys.length && !finalAllowed.length && !finalAuths.length && !finalModelsPath && !finalChatPath) {
     delete configs[id];
   } else {
     configs[id] = { baseUrl: cleanUrl, keys: cleanKeys };
     if (finalAuths.length) configs[id].auths = finalAuths;
     if (finalAllowed.length) configs[id].allowedModels = finalAllowed;
+    if (finalModelsPath) configs[id].modelsPath = finalModelsPath;
+    if (finalChatPath) configs[id].chatPath = finalChatPath;
   }
   // 同步清理旧 providerKeys 中同 id 的残留，避免双写
   const oldKeys = readState(file).providerKeys;
@@ -493,10 +544,10 @@ export function saveProviderConfig(id, { baseUrl, keys, auths, allowedModels },
     const nextKeys = { ...oldKeys };
     delete nextKeys[id];
     writeStateImmediate(file, { providerKeys: nextKeys, providerConfigs: configs });
-    return { baseUrl: cleanUrl, keys: cleanKeys, auths: finalAuths, allowedModels: finalAllowed };
+    return { baseUrl: cleanUrl, keys: cleanKeys, auths: finalAuths, allowedModels: finalAllowed, modelsPath: finalModelsPath, chatPath: finalChatPath };
   }
   writeStateImmediate(file, { providerConfigs: configs });
-  return { baseUrl: cleanUrl, keys: cleanKeys, auths: finalAuths, allowedModels: finalAllowed };
+  return { baseUrl: cleanUrl, keys: cleanKeys, auths: finalAuths, allowedModels: finalAllowed, modelsPath: finalModelsPath, chatPath: finalChatPath };
 }
 
 // ---- 供应商模型白名单：providerConfigs.<id>.allowedModels（空 = 不限） ----
@@ -515,12 +566,16 @@ export function saveProviderAllowedModels(id, list, { file = defaultStateFile()
   const baseUrl = normalizeBaseUrl(cur.baseUrl || loadProviderBaseUrl(id, { file }) || "");
   const keys = Array.isArray(cur.keys) ? cur.keys : loadProviderKeys(id, { file });
   const auths = normalizeAuths(cur.auths);
-  if (!baseUrl && !keys.length && !auths.length && !clean.length) {
+  const modelsPath = typeof cur.modelsPath === "string" ? normalizeEndpointPath(cur.modelsPath) : "";
+  const chatPath = typeof cur.chatPath === "string" ? normalizeEndpointPath(cur.chatPath) : "";
+  if (!baseUrl && !keys.length && !auths.length && !clean.length && !modelsPath && !chatPath) {
     delete configs[id];
   } else {
     configs[id] = { baseUrl, keys };
     if (auths.length) configs[id].auths = auths;
     if (clean.length) configs[id].allowedModels = clean;
+    if (modelsPath) configs[id].modelsPath = modelsPath;
+    if (chatPath) configs[id].chatPath = chatPath;
   }
   writeStateImmediate(file, { providerConfigs: configs });
   return clean;