mslxdff 0.1.63 → 0.1.64

package/bin/mslxdff.js

@@ -295,11 +295,54 @@ if (args.includes("-model") || args.includes("-models")) {
     // provider filter: bare ids are opencode, prefixed are <provider>/...
     if (modelListProvider) {
       const prov = String(modelListProvider).toLowerCase();
-      ids = ids.filter((id) => {
+      const filtered = ids.filter((id) => {
         const slash = String(id).indexOf("/");
         const p = slash > 0 ? String(id).slice(0, slash).toLowerCase() : "opencode";
         return p === prov;
       });
+      // 非 opencode 供应商的模型不在 opencode 缓存中，改从 allowlist 展示（原名+别名）
+      if (prov !== "opencode" && filtered.length === 0) {
+        try {
+          const { loadProviderAllowedModels, loadProviderAllowAnyModels, loadProviderBaseUrl } = await import("../src/state.js");
+          const { loadModelAliases, getAliasForModel } = await import("../src/providers/model-id.js");
+          try { loadModelAliases(); } catch {}
+          const allowed = loadProviderAllowedModels(prov);
+          const allowAny = loadProviderAllowAnyModels(prov);
+          const baseUrl = loadProviderBaseUrl(prov);
+          if (modelListJson) {
+            const data = allowed.length
+              ? allowed.map((raw) => ({ id: `${prov}/${raw}`, object: "model" }))
+              : [];
+            console.log(JSON.stringify({ object: "list", data }, null, 2));
+            process.exit(0);
+          }
+          if (!allowed.length) {
+            if (allowAny) {
+              console.log(`provider "${prov}" allowAny ON (allowlist 空=放行全部)${baseUrl ? `  baseUrl=${baseUrl}` : ""}`);
+              console.log(`  (未设 allowlist，全部模型放行)  查看 live 列表: mslxdff -provider ${prov} models`);
+            } else {
+              console.log(`no models for provider "${prov}" — allowlist 空 + allowAny OFF = 阻塞`);
+              console.log(`  设白名单: mslxdff -provider ${prov} allowlist set <model1> <model2>  或  mslxdff -provider ${prov} allowAny on`);
+              console.log(`  live 查看: mslxdff -provider ${prov} models`);
+            }
+            process.exit(0);
+          }
+          const at2 = cachedAt ? ` (cached ${fmtShanghaiYMDHM(cachedAt)})` : "";
+          console.log(`${allowed.length} model(s) for ${prov}${at2} (allowlist，原名 + 别名):`);
+          const pickedIds2 = loadModelPicks();
+          for (const raw of allowed) {
+            const canonical = `${prov}/${raw}`;
+            let alias = null;
+            try { alias = getAliasForModel(canonical); } catch {}
+            if (!alias && String(canonical).includes("/")) alias = String(canonical).replace(/\//g, "-");
+            const aliasStr = alias && alias !== canonical ? `  (别名: ${alias})` : "";
+            const mark2 = pickedIds2.includes(canonical) || (alias && pickedIds2.includes(alias)) ? "*" : " ";
+            console.log(`  ${mark2} ${canonical}${aliasStr}`);
+          }
+          process.exit(0);
+        } catch {}
+      }
+      ids = filtered;
       if (modelListJson) {
         console.log(JSON.stringify({ object: "list", data: ids.map((id) => ({ id, object: "model" })) }, null, 2));
         process.exit(0);
@@ -316,12 +359,42 @@ if (args.includes("-model") || args.includes("-models")) {
       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 && !modelListProvider) {
+    // TTY 交互式：`mslxdff -models`（无 list）走交互；` -model list` 默认列表，但交互入口统一为 -models
+    // 为支持“opencode + 其他供应商 allowlist 原名/别名”一起勾选，交互池 = opencode 免费池 + 各供应商 allowlist + 已勾选的遗留 picks
+    if (sub === undefined && process.stdin.isTTY && process.stdout.isTTY) {
       const statuses = loadModelErrors();
       const current = getPreferredModel();
       const pickedIds = loadModelPicks();
-      const items = ids.map((id) => {
+      // 基础：opencode 免费池
+      const combinedIds = [...ids];
+      const seen = new Set(combinedIds);
+      try {
+        const { loadProviderConfigs, loadProviderAllowedModels } = await import("../src/state.js");
+        const configs = loadProviderConfigs();
+        for (const pid of Object.keys(configs).filter((k) => String(k).toLowerCase() !== "opencode")) {
+          const allowed = loadProviderAllowedModels(pid);
+          for (const raw of allowed) {
+            const canonical = `${pid}/${raw}`;
+            if (!seen.has(canonical)) {
+              seen.add(canonical);
+              combinedIds.push(canonical);
+            }
+            // 别名（dash 版）若已被 picks 选中，也确保可见（便于取消）
+            const aliasDash = String(canonical).replace(/\//g, "-");
+            if (aliasDash !== canonical && pickedIds.includes(aliasDash) && !seen.has(aliasDash)) {
+              // 不直接加入 alias 作为独立选项，保留 canonical 即可（canonical 与 alias 视为同一模型，勾选 canonical）
+            }
+          }
+        }
+      } catch {}
+      // 已勾选但不在上述池中的（例如 workbuddy/* 当 allowAny ON 时无 allowlist，或历史 picks），补齐以便取消
+      for (const pid of pickedIds) {
+        if (!seen.has(pid)) {
+          seen.add(pid);
+          combinedIds.push(pid);
+        }
+      }
+      const items = combinedIds.map((id) => {
         const e = statuses[id];
         return {
           id,
@@ -342,38 +415,158 @@ if (args.includes("-model") || args.includes("-models")) {
     const at = cachedAt ? ` (cached ${fmtShanghaiYMDHM(cachedAt)})` : "";
     const pickedIds = loadModelPicks();
     const mark = (id) => (pickedIds.includes(id) ? "*" : " ");
-    console.log(`${ids.length} free model(s)${at} (${pickedIds.length} picked, * = picked):`);
-    // 加载别名映射，显示原始名+别名
-    let aliasMap = {};
-    let fullAliases = {};
-    try {
-      const { loadModelAliases, getAliasForModel } = await import("../src/providers/model-id.js");
-      loadModelAliases();
-      for (const id of ids) {
-        const alias = getAliasForModel(id);
-        if (alias) aliasMap[id] = alias;
+    // 分组展示：按供应商前缀分组，opencode 裸 id 归为 opencode
+    const groups = {};
+    for (const id of ids) {
+      const prov = String(id).includes("/") ? String(id).split("/")[0] : "opencode";
+      if (!groups[prov]) groups[prov] = [];
+      groups[prov].push(id);
+    }
+    const order = ["opencode", "workbuddy", "clinebot", "openrouter"];
+    const sortedProvs = Object.keys(groups).sort((a, b) => {
+      const ia = order.indexOf(a), ib = order.indexOf(b);
+      if (ia !== -1 || ib !== -1) {
+        if (ia === -1) return 1;
+        if (ib === -1) return -1;
+        return ia - ib;
       }
-      // 读取完整 alias 表（用于展示本地别名，如 clinebot/*）
+      return a.localeCompare(b);
+    });
+    if (modelListProvider) {
+      console.log(`${ids.length} model(s) for ${modelListProvider}${at} (${pickedIds.length} picked, * = picked):`);
+      // 加载别名映射，显示原始名+别名
+      let aliasMap = {};
       try {
-        const aliasesFile = join(homedir(), ".config", "mslxdff", "model-aliases.json");
-        const raw = JSON.parse(readFileSync(aliasesFile, "utf8"));
-        if (raw && typeof raw === "object") fullAliases = raw;
+        const { loadModelAliases, getAliasForModel } = await import("../src/providers/model-id.js");
+        loadModelAliases();
+        for (const id of ids) {
+          const alias = getAliasForModel(id);
+          if (alias) aliasMap[id] = alias;
+          else if (String(id).includes("/")) {
+            const dashAlias = String(id).replace(/\//g, "-");
+            if (dashAlias !== id) aliasMap[id] = dashAlias;
+          }
+        }
       } catch {}
-    } catch {}
-    for (const id of ids) {
-      const alias = aliasMap[id];
-      const aliasStr = alias ? `  (alias: ${alias})` : "";
-      console.log(`  ${mark(id)} ${id}${aliasStr}`);
-    }
-    // 额外展示本地别名（不在上游列表中的，如 clinebot/*）
-    const aliasEntries = Object.entries(fullAliases).filter(([alias, canonical]) => !ids.includes(canonical));
-    if (aliasEntries.length) {
-      console.log(`\nlocal aliases (${aliasEntries.length}):`);
-      for (const [alias, canonical] of aliasEntries) {
-        console.log(`  ${canonical}  =>  ${alias}`);
+      for (const prov of sortedProvs) {
+        const list = groups[prov];
+        console.log(`\n  ── ${prov} (${list.length}) ──`);
+        for (const id of list) {
+          const alias = aliasMap[id];
+          const aliasStr = alias ? `  (别名: ${alias})` : "";
+          console.log(`  ${mark(id)} ${id}${aliasStr}`);
+        }
       }
+      console.log(`\npicked only constrains auto; manage with: mslxdff -models (TTY) | mslxdff -model pick <id> | mslxdff -model unpick <id> | mslxdff -model pick clear`);
+    } else {
+      console.log(`${ids.length} free model(s)${at} (${pickedIds.length} picked, * = picked):`);
+      // 加载别名映射，显示原始名+别名
+      let aliasMap = {};
+      let fullAliases = {};
+      try {
+        const { loadModelAliases, getAliasForModel } = await import("../src/providers/model-id.js");
+        loadModelAliases();
+        for (const id of ids) {
+          const alias = getAliasForModel(id);
+          if (alias) aliasMap[id] = alias;
+          else if (String(id).includes("/")) {
+            const dashAlias = String(id).replace(/\//g, "-");
+            if (dashAlias !== id) aliasMap[id] = dashAlias;
+          }
+        }
+        try {
+          const aliasesFile = join(homedir(), ".config", "mslxdff", "model-aliases.json");
+          const raw = JSON.parse(readFileSync(aliasesFile, "utf8"));
+          if (raw && typeof raw === "object") fullAliases = raw;
+        } catch {}
+      } catch {}
+      for (const prov of sortedProvs) {
+        const list = groups[prov];
+        console.log(`\n  ── ${prov} (${list.length}) ──`);
+        for (const id of list) {
+          const alias = aliasMap[id];
+          const aliasStr = alias ? `  (别名: ${alias})` : "";
+          console.log(`  ${mark(id)} ${id}${aliasStr}`);
+        }
+      }
+      // ── 分隔 + 其他供应商 allowlist（原名 + 别名） ──
+      // 其余供应商的可用模型由 allowlist 决定（空 allowlist + allowAny OFF = 阻塞），此处直接展示 allowlist 条目与别名，不做网络拉取，速度快
+      try {
+        const { loadProviderConfigs, loadProviderAllowedModels, loadProviderAllowAnyModels, loadProviderBaseUrl } = await import("../src/state.js");
+        const { loadModelAliases: _la2, getAliasForModel: _gaf } = await import("../src/providers/model-id.js");
+        try { _la2(); } catch {}
+        const configs = loadProviderConfigs();
+        const otherIds = Object.keys(configs).filter((k) => String(k).toLowerCase() !== "opencode");
+        // 仅展示已配置过的供应商（有 baseUrl/keys/allowlist 任一），按固定顺序 + 字典序
+        const order2 = ["workbuddy", "clinebot", "openrouter", "bai"];
+        otherIds.sort((a, b) => {
+          const ia = order2.indexOf(a), ib = order2.indexOf(b);
+          if (ia !== -1 || ib !== -1) {
+            if (ia === -1) return 1;
+            if (ib === -1) return -1;
+            return ia - ib;
+          }
+          return a.localeCompare(b);
+        });
+        if (otherIds.length) {
+          console.log(`\n────────────────────────────────────────`);
+          console.log(`其他供应商 (allowlist，原名 + 别名) (${otherIds.length} providers):`);
+          for (const pid of otherIds) {
+            const allowed = loadProviderAllowedModels(pid);
+            const allowAny = loadProviderAllowAnyModels(pid);
+            const baseUrl = loadProviderBaseUrl(pid) || configs[pid]?.baseUrl || "";
+            const header = allowAny
+              ? (allowed.length ? `allowlist ${allowed.length} (allowAny ON)` : `allowAny ON (allowlist 空=放行全部)`)
+              : (allowed.length ? `allowlist ${allowed.length} (allowAny OFF)` : `allowlist 空 + allowAny OFF = 阻塞`);
+            console.log(`\n  ── ${pid} (${header})${baseUrl ? `  baseUrl=${baseUrl}` : ""} ──`);
+            if (!allowed.length) {
+              if (allowAny) {
+                console.log(`     (未设 allowlist，全部模型放行)  查看 live 列表: mslxdff -provider ${pid} models`);
+                console.log(`     限制可用模型: mslxdff -provider ${pid} allowlist set <model1> <model2>`);
+              } else {
+                console.log(`     阻塞中：无可用模型 — 设白名单: mslxdff -provider ${pid} allowlist set <model1> <model2>`);
+                console.log(`     或放行全部: mslxdff -provider ${pid} allowAny on`);
+              }
+            } else {
+              for (const raw of allowed) {
+                const canonical = `${pid}/${raw}`;
+                let alias = null;
+                try { alias = _gaf(canonical); } catch {}
+                if (!alias && String(canonical).includes("/")) alias = String(canonical).replace(/\//g, "-");
+                const aliasStr = alias && alias !== canonical ? `  (别名: ${alias})` : "";
+                const pickedMark = pickedIds.includes(canonical) || pickedIds.includes(alias || "") ? "*" : " ";
+                console.log(`     ${pickedMark} ${canonical}${aliasStr}`);
+              }
+              console.log(`     管理: mslxdff -provider ${pid} allowlist [list|add|remove|clear] | allowAny on|off`);
+            }
+          }
+        } else {
+          console.log(`\n────────────────────────────────────────`);
+          console.log(`其他供应商 (allowlist，原名 + 别名): (none — 尚未配置)`);
+          console.log(`  添加示例: mslxdff -provider add myapi https://api.example.com/v1 sk-xxx --models-path /v1/models`);
+        }
+        // 额外兜底：本地别名中不在任何 allowlist 里的，单独提示
+        const aliasEntries = Object.entries(fullAliases).filter(([alias, canonical]) => {
+          if (ids.includes(canonical)) return false;
+          // 已在 allowlist 中展示过的别名不重复
+          for (const pid of otherIds) {
+            const allowed = loadProviderAllowedModels(pid);
+            for (const raw of allowed) {
+              const can = `${pid}/${raw}`;
+              if (can === canonical) return false;
+            }
+          }
+          return true;
+        });
+        if (aliasEntries.length) {
+          console.log(`\n  本地别名 (不在 allowlist 里的遗留映射 ${aliasEntries.length}):`);
+          for (const [alias, canonical] of aliasEntries) {
+            console.log(`     ${canonical}  =>  ${alias}`);
+          }
+        }
+      } catch {}
+      console.log(`\npicked only constrains auto; manage with: mslxdff -models (TTY) | mslxdff -model pick <id> | mslxdff -model unpick <id> | mslxdff -model pick clear`);
     }
-    console.log(`\npicked only constrains auto; manage with: mslxdff -models (TTY) | mslxdff -model pick <id> | mslxdff -model unpick <id> | mslxdff -model pick clear`);
   } catch (err) {
     console.error(`could not fetch models: ${String(err?.message || err)}`);
     process.exit(1);
@@ -764,17 +957,21 @@ if (args.includes("-provider") || args.includes("--provider")) {
   const sub = args[idx + 2];
   const rest = args.slice(idx + 2);
   if (!id) {
-    console.error("usage: mslxdff -provider <id> [key...|add|remove|list|clear|share|set-url|allowlist|allowAny]");
+    console.error("usage: mslxdff -provider <id> [key...|add|remove|list|models|clear|share|set-url|allowlist|allowAny]");
     console.error("       e.g. mslxdff -provider openrouter sk-1 sk-2 sk-3      set multiple keys (replaces all)");
     console.error("            mslxdff -provider openrouter add sk-4             append one key");
     console.error("            mslxdff -provider openrouter remove sk-1          remove a key by value");
     console.error("            mslxdff -provider openrouter list                 list all keys (masked)");
+    console.error("            mslxdff -provider openrouter models [--json]      list provider models (allowlist filtered)");
     console.error("            mslxdff -provider openrouter share on|off         share keys with peers on outgoing forward (ADR-0008)");
     console.error("            mslxdff -provider openrouter set-url https://api.example.com/v1");
+    console.error("            mslxdff -provider openrouter set-models-path /v1/models");
+    console.error("            mslxdff -provider openrouter set-chat-path /v1/chat/completions");
     console.error("            mslxdff -provider openrouter allowlist set gpt-4 gpt-3.5  manage allowed models (empty=block unless allowAny ON)");
     console.error("            mslxdff -provider openrouter allowAny on|off     empty allowlist = allow all or block all (default OFF, secure)");
     console.error("            mslxdff -provider add myapi https://api.example.com/v1 sk-xxx   add generic OpenAI-compatible provider");
     console.error("            mslxdff -provider add myapi https://api.example.com/v1 sk-xxx gpt-4  add with allowlist");
+    console.error("            mslxdff -provider add myapi https://api.example.com/v1 sk-xxx --models-path /v1/models --chat-path /v1/chat/completions");
     console.error("            mslxdff -provider openrouter                      interactive hidden input (append)");
     console.error("            mslxdff -provider openrouter clear                remove all keys");
     process.exit(1);
@@ -832,9 +1029,10 @@ if (args.includes("-provider") || args.includes("--provider")) {
     const gBase = rest[1];
     const gKey = rest[2];
     if (!gid || !gBase || !gKey) {
-      console.error("usage: mslxdff -provider add <id> <baseUrl> <key> [allowedModel...]");
+      console.error("usage: mslxdff -provider add <id> <baseUrl> <key> [allowedModel...] [--models-path <path>] [--chat-path <path>]");
       console.error("       e.g. mslxdff -provider add myapi https://api.example.com/v1 sk-xxx");
       console.error("            mslxdff -provider add myapi https://api.example.com/v1 sk-xxx gpt-4 gpt-3.5");
+      console.error("            mslxdff -provider add myapi https://api.example.com/v1 sk-xxx --models-path /v1/models --chat-path /v1/chat/completions");
       process.exit(1);
     }
     if (gid === "opencode" || gid === "oc" || gid === "openrouter") {
@@ -1317,7 +1515,8 @@ async function pickInteractive(items, startCursor = 0) {
   let cursor = Math.min(Math.max(startCursor, 0), items.length - 1);
   const draw = () => {
     const lines = [...renderChooser(items, cursor), ...renderChooserHelp()];
-    process.stdout.write("\x1b[G\x1b[J" + lines.join("\n"));
+    // PowerShell 兼容：\x1b[2J 清屏 + \x1b[H 回到左上，比 \x1b[G\x1b[J 更可靠
+    process.stdout.write("\x1b[2J\x1b[H" + lines.join("\n"));
   };
   draw();
   return new Promise((resolve) => {
@@ -1359,7 +1558,7 @@ async function pickInteractiveMulti(items, initialPicked = new Set(), startCurso
   const draw = () => {
     const rows = items.map((it, i) => ({ ...it, picked: picked.has(it.id) }));
     const lines = [...renderChooser(rows, cursor, { multi: true }), ...renderChooserHelp(true)];
-    process.stdout.write("\x1b[G\x1b[J" + lines.join("\n"));
+    process.stdout.write("\x1b[2J\x1b[H" + lines.join("\n"));
   };
   draw();
   return new Promise((resolve) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mslxdff",
-  "version": "0.1.63",
+  "version": "0.1.64",
   "description": "测试项目，请勿使用。",
   "type": "module",
   "bin": {

package/src/auto.js

@@ -85,6 +85,7 @@ function effectiveCooldown(entry, slowCooldownMs, cooldownMs) {
 function inCooldown(id, errors, now, cooldownMs, slowCooldownMs) {
   const e = normEntry(errors[id]);
   if (!e || !(e.at > 0)) return false;
+  if (e.status === MODEL_STATUS.NORMAL) return false;
   const cd = effectiveCooldown(e, slowCooldownMs, cooldownMs);
   return cd > 0 && now - e.at < cd;
 }
@@ -98,6 +99,16 @@ function normLatency(e) {
 
 export function rankModels(ids, errors = {}, { now = Date.now(), cooldownMs = 0, slowCooldownMs = 0, latencies = {}, preferred } = {}) {
   const pref = preferred ?? getPreferredModel();
+  // 最近一次成功的模型（NORMAL 且非慢且 at 最大），用于“上次成功优先”——慢模型即使刚成功也不应钉死
+  let lastSuccessId = null;
+  let lastSuccessAt = 0;
+  for (const [id, e] of Object.entries(errors)) {
+    const ne = normEntry(e);
+    if (ne && ne.status === MODEL_STATUS.NORMAL && !ne.slow && ne.at > lastSuccessAt) {
+      lastSuccessAt = ne.at;
+      lastSuccessId = id;
+    }
+  }
   return [...new Set(ids)]
     .filter(Boolean)
     .map((id) => ({
@@ -105,12 +116,14 @@ export function rankModels(ids, errors = {}, { now = Date.now(), cooldownMs = 0,
       e: normEntry(errors[id]),
       err: normEntry(errors[id])?.at ?? 0,
       isPreferred: id === pref,
+      isLastSuccess: id === lastSuccessId,
       cooling: inCooldown(id, errors, now, cooldownMs, slowCooldownMs),
       latency: normLatency(latencies[id]) ?? Number.MAX_SAFE_INTEGER,
     }))
     .sort(
       (a, b) =>
         (a.cooling ? 1 : 0) - (b.cooling ? 1 : 0) ||
+        (b.isLastSuccess ? 1 : 0) - (a.isLastSuccess ? 1 : 0) ||
         (b.isPreferred ? 1 : 0) - (a.isPreferred ? 1 : 0) ||
         a.latency - b.latency ||
         a.err - b.err

package/src/chat/config.js

@@ -6,5 +6,6 @@ export const CHAT_HISTORY_MAX_CHARS = 400000;
 export const CHAT_KEEP_RECENT = 40;
 export const CHAT_SUMMARY_TRIGGER = 400000;
 export const CHAT_MAX_TOOL_LOOPS = 6;
-export const CHAT_TIMEOUT_MS = 30000;
+export const CHAT_TIMEOUT_MS = 15000;
+export const CHAT_GATEWAY_TIMEOUT_MS = 25000;
 export const FORBIDDEN = ["-uninstall", "--uninstall"];

package/src/chat/repl.js

@@ -30,35 +30,7 @@ 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];
@@ -85,23 +57,11 @@ 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;
   let lastFallback = false;
+  let lastFallbackGateway = false;
   let lastLatency = 0;
   const t0 = performance.now();
   const turnStart = performance.now();
@@ -147,6 +107,7 @@ async function runAgentTurn(userText, messages) {
     lastModel = res.model;
     lastUsage = res.usage || null;
     lastFallback = !!res.fallback;
+    lastFallbackGateway = !!res.fallbackGateway || !!res.viaGateway;
     const msg = res.message;
     const toolCalls = msg.tool_calls || [];
     let fallbackCmd = null;
@@ -158,9 +119,11 @@ async function runAgentTurn(userText, messages) {
       const text = String(msg.content || "").trim() || "(空回复)";
       messages.push({ role: "assistant", content: text });
       const totalMs = Math.round(performance.now() - t0);
-      const note = lastFallback ? "\n\x1b[90m[注：mimo 不可用，已用 big-pickle]\x1b[0m" : "";
-      trace(`[turn] 完成 总计 ${totalMs}ms · LLM ${lastLatency}ms · 0 工具`);
-      return { text: text + note, model: lastModel, latency: lastLatency, usage: lastUsage, fallback: lastFallback, ok: true, totalMs };
+      let note = "";
+      if (lastFallbackGateway) note = "\n\x1b[90m[注：mimo/big-pickle 均不可用，已自动切本地网关 auto（:8989）]\x1b[0m";
+      else if (lastFallback) note = "\n\x1b[90m[注：mimo 不可用，已用 big-pickle]\x1b[0m";
+      trace(`[turn] 完成 总计 ${totalMs}ms · LLM ${lastLatency}ms · 0 工具${lastFallbackGateway ? " · gateway-fallback" : ""}`);
+      return { text: text + note, model: lastModel, 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 })) });
@@ -262,19 +225,19 @@ async function runAgentTurn(userText, messages) {
     trace(`[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, ok: false };
+  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 };
 }
 
-function printFooter({ model, latency, usage, totalMs, fallback }) {
+function printFooter({ model, latency, usage, totalMs, fallback, fallbackGateway, viaGateway }) {
   const dim = "\x1b[90m";
   const rst = "\x1b[0m";
   let gw = null;
   try { gw = collectStats(); } catch {}
-  const modelLabel = model || "—";
+  const modelLabel = model ? (fallbackGateway || viaGateway ? `${model} (gateway auto)` : model) : "—";
   const latLabel = latency ? `${latency}ms` : "—";
   const totalLabel = totalMs ? ` · 总耗时 ${totalMs}ms` : "";
   const tokLabel = usage ? ` · tokens ${usage.prompt_tokens ?? "?"}→${usage.completion_tokens ?? "?"}` : "";
-  const fbLabel = fallback ? " · fallback" : "";
+  const fbLabel = fallbackGateway || viaGateway ? " · gateway-fallback" : fallback ? " · fallback" : "";
   let gwLabel = "";
   let extra = "";
   if (gw) {

package/src/chat/stats.js

@@ -150,7 +150,7 @@ export function formatBannerLines() {
   const green = "\x1b[32m";
   const lines = [];
   lines.push(`${cyan}┌─ mslxdff chat  ·  数据来自 -d 网关进程（非本会话） ─────${rst}`);
-  lines.push(`${cyan}│${rst}  对话模型  ${yellow}${s.chatPref}${rst} ${dim}→ ${s.chatFall}（自动降级）${rst}  ${dim}[${s.chatPrefStatus}/${s.chatFallStatus}]${rst}`);
+  lines.push(`${cyan}│${rst}  对话模型  ${yellow}${s.chatPref}${rst} ${dim}→ ${s.chatFall} → gateway auto:8989${rst}  ${dim}[${s.chatPrefStatus}/${s.chatFallStatus}]${rst}  ${dim}三级兜底${rst}`);
   lines.push(`${cyan}│${rst}  网关默认  ${green}${s.gatewayModel}${rst} ${dim}[${s.gatewayStatus}]${rst}  ·  端口 ${s.port}  ·  ${dim}${s.endpointUrl}${rst}`);
   const prefTtfb = s.chatPrefStat?.avgTtfbMs ?? s.chatPrefStat?.emaTtfbMs ?? s.chatPrefLat?.emaMs;
   const fallTtfb = s.chatFallStat?.avgTtfbMs ?? s.chatFallStat?.emaTtfbMs ?? s.chatFallLat?.emaMs;

package/src/chat/upstream.js

@@ -1,6 +1,7 @@
 import { performance } from "node:perf_hooks";
-import { CHAT_PREFERRED, CHAT_FALLBACK, CHAT_TIMEOUT_MS } from "./config.js";
+import { CHAT_PREFERRED, CHAT_FALLBACK, CHAT_TIMEOUT_MS, CHAT_GATEWAY_TIMEOUT_MS } from "./config.js";
 import { createUpstreamClient } from "../upstream.js";
+import { DEFAULT_PORT } from "../state.js";
 
 function modelForAttempt(attempt) {
   return attempt === 0 ? CHAT_PREFERRED : CHAT_FALLBACK;
@@ -56,25 +57,202 @@ async function chatOnceNoTools({ messages, model }) {
   } finally { try { await client.close(); } catch {} }
 }
 
-// 带自动降级：仅 mimo-v2.5-free → big-pickle，不引入其他模型
+// 带自动降级：mimo-v2.5-free → big-pickle → 本地 8989 auto（网关 auto，含多供应商择优/hedge/peer）
+async function chatViaGateway({ messages, tools }) {
+  const TRACE = process.env.MSLXDFF_CHAT_TRACE !== "0";
+  const t0 = TRACE ? performance.now() : 0;
+  let port = DEFAULT_PORT;
+  let token = "";
+  try {
+    const state = await import("../state.js");
+    const loaded = await state.loadToken();
+    token = String(loaded?.token || "").trim();
+    const p = state.getPort();
+    if (Number.isInteger(p) && p > 0) port = p;
+    else if (Number.isInteger(Number(process.env.MSLXDFF_PORT)) && Number(process.env.MSLXDFF_PORT) > 0) port = Number(process.env.MSLXDFF_PORT);
+  } catch {}
+  // token 为空则尝试直接读 state 文件兜底（避免 loadToken 异常时无 token）
+  if (!token) {
+    try {
+      const { readFileSync, existsSync } = await import("node:fs");
+      const { join } = await import("node:path");
+      const { homedir } = await import("node:os");
+      const sf = process.env.MSLXDFF_STATE_FILE || join(homedir(), ".config", "mslxdff", "state.json");
+      if (existsSync(sf)) {
+        const j = JSON.parse(readFileSync(sf, "utf8"));
+        if (typeof j.token === "string" && j.token.trim()) token = j.token.trim();
+      }
+    } catch {}
+  }
+  const url = `http://127.0.0.1:${port}/v1/chat/completions`;
+  const body = { model: "auto", messages, stream: false };
+  if (tools?.length) {
+    body.tools = tools;
+    body.tool_choice = "auto";
+  }
+  try {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), CHAT_GATEWAY_TIMEOUT_MS);
+    const res = await fetch(url, {
+      method: "POST",
+      headers: { "Content-Type": "application/json", ...(token ? { Authorization: `Bearer ${token}` } : {}) },
+      body: JSON.stringify(body),
+      signal: controller.signal,
+    });
+    clearTimeout(timer);
+    const txt = await res.text();
+    let j;
+    try { j = JSON.parse(txt); } catch {
+      // 网关可能因 workbuddy 强制 stream:true 而返回 SSE（text/event-stream），需兼容
+      if (txt.includes("data:")) {
+        try {
+          const lines = txt.split(/\r?\n/);
+          let content = "";
+          let model = "auto";
+          let usage = null;
+          let sseOk = false;
+          for (const line of lines) {
+            const t = String(line).trim();
+            if (!t.startsWith("data:")) continue;
+            const payload = t.slice(5).trim();
+            if (!payload || payload === "[DONE]") continue;
+            try {
+              const obj = JSON.parse(payload);
+              sseOk = true;
+              const ch = obj.choices?.[0];
+              // 兼容 thinking 模型的 reasoning_content（delta 阶段 content 为空，实际在 reasoning_content）
+              if (ch?.delta?.content) content += ch.delta.content;
+              else if (ch?.delta?.reasoning_content) content += ch.delta.reasoning_content;
+              else if (ch?.message?.content) content += ch.message.content;
+              else if (ch?.message?.reasoning_content) content += ch.message.reasoning_content;
+              else if (typeof ch?.text === "string") content += ch.text;
+              else if (typeof obj.content === "string") content += obj.content;
+              if (obj.model) model = obj.model;
+              if (obj.usage) usage = obj.usage;
+              // 有些 SSE 直接是完整 chat.completion
+              if (obj.choices?.[0]?.message?.content && !content) content = obj.choices[0].message.content;
+              if (obj.choices?.[0]?.message?.reasoning_content && !content) content = obj.choices[0].message.reasoning_content;
+            } catch {}
+          }
+          if (sseOk && content) {
+            j = { id: `sse-${Date.now()}`, object: "chat.completion", model, choices: [{ index: 0, finish_reason: "stop", message: { role: "assistant", content } }], usage };
+          } else if (sseOk) {
+            // SSE 但无 content，按失败处理
+            return { ok: false, error: `gateway SSE no content: ${txt.slice(0, 800)}`, status: res.status };
+          } else {
+            return { ok: false, error: `gateway non-json: ${txt.slice(0, 800)}`, status: res.status };
+          }
+        } catch {
+          return { ok: false, error: `gateway non-json: ${txt.slice(0, 800)}`, status: res.status };
+        }
+      } else {
+        return { ok: false, error: `gateway non-json: ${txt.slice(0, 800)}`, status: res.status };
+      }
+    }
+    if (!res.ok) {
+      const msg = j?.error?.message || j?.error || j?.data?.error?.message || j?.data?.error || txt.slice(0, 800);
+      if (TRACE) {
+        const dt = Math.round(performance.now() - t0);
+        console.log(`\x1b[90m· [LLM] gateway auto FAIL · ${dt}ms · HTTP ${res.status} ${String(msg).slice(0, 80)}\x1b[0m`);
+      }
+      return { ok: false, error: msg, status: res.status };
+    }
+    // 兼容网关返回的两种形状：标准 {"choices":...} 与 workbuddy 聚合后的 {"data":{"choices":...}}
+    const choice = j.choices?.[0] || j.data?.choices?.[0];
+    const effectiveJ = j.choices ? j : (j.data?.choices ? j.data : j);
+    if (!choice) {
+      if (TRACE) console.log(`\x1b[90m· [gateway debug] no choice, txt=${txt.slice(0, 800)} · j=${JSON.stringify(j).slice(0, 800)}\x1b[0m`);
+      return { ok: false, error: `gateway no choice: ${txt.slice(0, 800)}`, status: res.status };
+    }
+    if (TRACE) {
+      const dt = Math.round(performance.now() - t0);
+      const m = effectiveJ.model || j.model || choice.message?.model || "auto";
+      console.log(`\x1b[90m· [LLM] gateway auto OK · ${dt}ms · 模型 ${m} · 总 ${Math.round(performance.now() - t0)}ms (gateway-fallback)\x1b[0m`);
+    }
+    // 透传 usage/raw，并标记 gateway（兼容 data 包装）
+    return { ok: true, message: choice.message, usage: effectiveJ.usage || j.usage, raw: j, status: res.status, model: effectiveJ.model || j.model || "auto", viaGateway: true };
+  } catch (err) {
+    const msg = String(err?.message || err).slice(0, 800);
+    if (TRACE) {
+      const dt = Math.round(performance.now() - t0);
+      console.log(`\x1b[90m· [LLM] gateway auto FAIL · ${dt}ms · ${msg.slice(0, 80)}\x1b[0m`);
+    }
+    return { ok: false, error: `gateway ${msg}`, status: 502 };
+  }
+}
+
+// 带自动降级：mimo-v2.5-free → big-pickle → 本地 8989 auto（网关 auto，含多供应商择优/hedge/peer）
+async function safeChatOnce(opts, model) {
+  try {
+    const r = await chatOnce({ ...opts, model });
+    return r;
+  } catch (err) {
+    const msg = String(err?.message || err).slice(0, 800);
+    const status = err?._t ? 502 : (err?.cause?.code ? 502 : 502);
+    return { ok: false, error: msg, status, _thrown: err };
+  }
+}
+
+async function isCoolingAsync(id) {
+  try {
+    const state = await import("../state.js");
+    const errors = state.loadModelErrors();
+    const e = errors[id];
+    if (!e || typeof e !== "object") return false;
+    const at = Number(e.at || 0);
+    if (!at) return false;
+    const isSlow = !!e.slow;
+    const cd = isSlow ? 5 * 60 * 1000 : 60 * 1000;
+    return Date.now() - at < cd && (e.status === "limit" || e.status === "error");
+  } catch { return false; }
+}
+
 export async function chatWithFallback(opts) {
   const TRACE = process.env.MSLXDFF_CHAT_TRACE !== "0";
   const t0 = TRACE ? performance.now() : 0;
-  const first = await chatOnce({ ...opts, model: CHAT_PREFERRED });
+  // 若上次已确认冷却（429/limit），直接跳过，避免 7+7 秒白等，第二次直接走“上次成功”的网关
+  const firstCooling = await isCoolingAsync(CHAT_PREFERRED);
+  let first;
+  if (firstCooling) {
+    if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 跳过（冷却中）· 直接试 ${CHAT_FALLBACK}\x1b[0m`);
+    first = { ok: false, error: "skip cooling", status: 429 };
+  } else {
+    first = await safeChatOnce(opts, CHAT_PREFERRED);
+  }
   if (TRACE) {
     const dt = Math.round(performance.now() - t0);
     console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} ${first.ok ? "OK" : "FAIL"} · ${dt}ms${first.ok ? "" : ` · ${String(first.error).slice(0, 80)}`}\x1b[0m`);
   }
   if (first.ok) return { ...first, model: CHAT_PREFERRED };
   const t1 = TRACE ? performance.now() : 0;
-  const second = await chatOnce({ ...opts, model: CHAT_FALLBACK });
+  const secondCooling = await isCoolingAsync(CHAT_FALLBACK);
+  let second;
+  if (secondCooling && firstCooling) {
+    if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} 跳过（冷却中）· 直接走网关 auto\x1b[0m`);
+    second = { ok: false, error: "skip cooling", status: 429 };
+  } else if (secondCooling) {
+    if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} 跳过（冷却中）· 直接走网关 auto\x1b[0m`);
+    second = { ok: false, error: "skip cooling", status: 429 };
+  } else {
+    second = await safeChatOnce(opts, CHAT_FALLBACK);
+  }
   if (TRACE) {
     const dt = Math.round(performance.now() - t1);
     const total = Math.round(performance.now() - t0);
     console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} ${second.ok ? "OK" : "FAIL"} · ${dt}ms · 总 ${total}ms (fallback)\x1b[0m`);
   }
   if (second.ok) return { ...second, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
-  return { ok: false, error: `${CHAT_PREFERRED} failed: ${first.error}; ${CHAT_FALLBACK} failed: ${second.error}`, status: second.status || first.status };
+  // 两者皆失败 → 兜底到本地网关 auto（会走 auto 择优、hedge、peer 等完整链路）
+  const t2 = TRACE ? performance.now() : 0;
+  if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} + ${CHAT_FALLBACK} 均失败，尝试本地网关 auto（:8989）\x1b[0m`);
+  const third = await chatViaGateway(opts);
+  if (TRACE) {
+    const dt = Math.round(performance.now() - t2);
+    const total = Math.round(performance.now() - t0);
+    console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${dt}ms · 总 ${total}ms (gateway-fallback)\x1b[0m`);
+  }
+  if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: second.error, viaGateway: true };
+  return { ok: false, error: `${CHAT_PREFERRED} failed: ${first.error}; ${CHAT_FALLBACK} failed: ${second.error}; gateway auto failed: ${third.error}`, status: third.status || second.status || first.status };
 }
 
 // 压缩用：简短摘要请求（不带 tools），128k 上下文下仅 95% 触发，需完整摘要