mslxdff 0.1.62 → 0.1.64

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

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

@@ -1,6 +1,6 @@
 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";
@@ -14,6 +14,13 @@ 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;
@@ -120,6 +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 = {
@@ -134,6 +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
@@ -201,7 +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}`,
@@ -268,7 +279,7 @@ export function createWorkbuddyProvider({
   }
 
   async function runChat(body, activeRing, opts = {}) {
-    const url = `${resolvedBase}/v2/chat/completions`;
+    const url = joinUrl(resolvedBase, resolvedChatPath);
     const t0 = performance.now();
     const preferredUid = opts?.workbuddyUid ? String(opts.workbuddyUid).trim() : "";
     const modelForLog = body?.model || "";
@@ -488,7 +499,7 @@ export function createWorkbuddyProvider({
   async function listModels() {
     const now = Date.now();
     if (cache && now - fetchedAt < CACHE_TTL_MS) return cache;
-    const url = `${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 {
@@ -523,7 +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) });
+  }
+}