mslxdff 0.1.78 → 0.1.80

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mslxdff",
-  "version": "0.1.78",
+  "version": "0.1.80",
   "description": "测试项目，请勿使用。",
   "type": "module",
   "bin": {

package/src/bench/report.js

@@ -23,6 +23,104 @@ function pad(s, n, align = "left") {
   return str + " ".repeat(d);
 }
 
+function shortPeerLabel(p) {
+  const raw = String(p?.name || p?.id || p?.url || String(p || "")).trim();
+  if (!raw) return "peer";
+  // 优先用 name；若是 url 则取 host:port 的尾段
+  if (raw.startsWith("http://") || raw.startsWith("https://") || raw.startsWith("relay://")) {
+    try { const u = new URL(raw); const host = u.hostname; const port = u.port ? `:${u.port}` : ""; if (host) return `${host}${port}`; } catch {}
+    return raw.slice(-12);
+  }
+  // name 可能是 url 字符串（如 "http://141.98..." 存于 name 字段）
+  if (raw.includes("://")) {
+    try { const u = new URL(raw); return u.hostname + (u.port ? `:${u.port}` : ""); } catch {}
+  }
+  return raw;
+}
+export function formatViaReport(results, { peers = [], meta = {}, json = false } = {}) {
+  const peerIds = (peers || []).map((p) => shortPeerLabel(p));
+  const samples = meta.samples ?? 1;
+  const timeout = meta.timeout ?? 30000;
+  const includeOpencode = Boolean(meta.includeOpencode);
+  const opencodeTag = includeOpencode ? "opencode=included" : "opencode=skipped";
+  const at = meta.at || new Date().toISOString();
+  // build json shape
+  const jsonObj = {
+    meta: { at, samples, timeout, includeOpencode, peers: peerIds, opencodeSkipped: !includeOpencode, ...meta, peers: peerIds },
+    results: (results || []).map((r) => ({
+      provider: r.provider,
+      model: r.model || r.id,
+      direct: r.direct ? { ttfb: r.direct.ttfbMs, total: r.direct.totalMs, ok: r.direct.ok, label: r.direct.label, error: r.direct.error } : null,
+      via: Object.fromEntries(Object.entries(r.via || {}).map(([k, v]) => [k, v.ok ? { ttfb: v.ttfbMs, total: v.totalMs } : { error: v.label || v.error || "offline", label: v.label }])),
+      best: r.best,
+      deltaMs: r.deltaMs,
+      opencodeSkipped: r.opencodeSkipped,
+    })),
+    advice: (() => {
+      const viaBest = (results || []).find((r) => r.best?.startsWith("via:"));
+      if (viaBest) return `${viaBest.provider}/${viaBest.model} 经 ${viaBest.best.slice(4)} 最快`;
+      if ((results || []).length) return `${results[0].provider} 走 direct 即可`;
+      return "无数据";
+    })(),
+  };
+  if (json) {
+    return { text: JSON.stringify(jsonObj, null, 2), json: jsonObj };
+  }
+  const lines = [];
+  lines.push(`bench-via: direct vs via peers (samples=${samples}, timeout=${timeout / 1000}s, ${opencodeTag}) peers=${peerIds.join(",") || "(none)"}`);
+  lines.push("");
+  // 表头：直连 + 每个组员一列（短化为 host:port 或 name），兼容旧测试 via B 断言与用户“直连/组员B”心智
+  const colW = 18;
+  const header = `${pad("Provider", 12)} ${pad("Model", 24)} ${pad("direct", colW, "right")} ${peerIds.map((id) => pad(`via ${id}`, colW, "right")).join(" ")} ${pad("best", 14)}`;
+  lines.push(header);
+  lines.push("─".repeat(header.length));
+  for (const r of results || []) {
+    const provider = pad(r.provider || "", 12);
+    const model = pad(r.model || r.id || "", 24);
+    const directOk = r.direct?.ok;
+    const directTxt = directOk ? `${r.direct.ttfbMs ?? "—"}ms` : (r.direct?.label || "—");
+    // determine best ttfb for ★
+    const all = [];
+    if (r.direct?.ok) all.push({ key: "direct", ttfb: r.direct.ttfbMs ?? r.direct.totalMs });
+    for (const pid of peerIds) {
+      const v = r.via?.[pid];
+      if (v?.ok) all.push({ key: `via:${pid}`, ttfb: v.ttfbMs ?? v.totalMs });
+    }
+    let bestKey = r.best;
+    if (!bestKey && all.length) bestKey = all.sort((a, b) => a.ttfb - b.ttfb)[0].key;
+    const isDirectBest = bestKey === "direct";
+    const directCell = pad(`${directTxt}${isDirectBest ? "★" : ""}`, colW, "right");
+    const viaCells = peerIds.map((pid) => {
+      const v = r.via?.[pid];
+      if (!v) return pad("—", colW, "right");
+      if (v.ok) {
+        const txt = `${v.ttfbMs ?? v.totalMs ?? "—"}ms`;
+        const star = bestKey === `via:${pid}` ? "★" : "";
+        return pad(`${txt}${star}`, colW, "right");
+      }
+      // 失败也展示延迟（先测延迟）：如 42ms 鉴权失败，说明网络可达但该组员未配此供应商
+      const ms = v.ttfbMs != null ? `${v.ttfbMs}ms ` : "";
+      const label = v.label || v.error || "offline";
+      const short = label.includes("离线") ? "offline" : label.slice(0, 8);
+      return pad(`${ms}— ${short}`, colW, "right");
+    }).join(" ");
+    let bestTxt = r.best || "direct";
+    if (r.deltaMs != null && r.best?.startsWith("via:")) {
+      const pct = r.direct?.ttfbMs ? Math.round((r.deltaMs / r.direct.ttfbMs) * 100) : 0;
+      bestTxt = `${r.best} ${pct}%`;
+    }
+    lines.push(`${provider} ${model} ${directCell} ${viaCells} ${pad(bestTxt, 14)}`);
+  }
+  lines.push("─".repeat(header.length));
+  const viaBestExample = (results || []).find((r) => r.best?.startsWith("via:"));
+  if (viaBestExample) lines.push(`建议：A 经 ${viaBestExample.best.slice(4)} 打 ${viaBestExample.provider} 最快；其余走 direct。`);
+  else if ((results || []).length) lines.push(`建议：${results[0].provider} 走 direct 即可。`);
+  else lines.push("建议：无数据");
+  lines.push(`提示：via 已跳过 opencode（省额度），需对比 opencode 请加 --include-opencode`);
+  lines.push(`* via 单样本，仅作参考，多次 --samples 2 取均值更稳`);
+  return { text: lines.join("\n"), json: jsonObj };
+}
+
 export function formatReport(results, { json = false } = {}) {
   const sorted = sortResults(results);
   const winner = sorted.find((r) => r.ok) || null;

package/src/bench/via-probe.js

@@ -0,0 +1,142 @@
+import { joinUrl } from "../providers/base.js";
+import { computeMetrics, extractUsageFromJson } from "../metrics.js";
+
+function extractInnerMessage(bodyText) {
+  const t = String(bodyText || "");
+  try {
+    const j = JSON.parse(t);
+    const m = j?.error?.message || j?.error || j?.message || j?.data?.error || "";
+    if (typeof m === "string" && m.trim()) return m.trim().slice(0, 300);
+    if (typeof j?.error === "string") return j.error.slice(0, 300);
+  } catch {}
+  return t.slice(0, 300);
+}
+
+function classifyError(status, bodyText) {
+  const t = String(bodyText || "").slice(0, 500);
+  const low = t.toLowerCase();
+  if (status === 401) return { label: "鉴权失败", retryable: false };
+  if (status === 402 || /insufficient balance/i.test(t)) return { label: "余额不足", retryable: false };
+  if (low.includes("only available via cline")) return { label: "仅 Cline 客户端可用", retryable: false };
+  if (low.includes("invalid model format")) return { label: "模型格式错误", retryable: false };
+  if (status === 403) return { label: /insufficient/i.test(t) ? "余额不足" : "鉴权失败", retryable: false };
+  if (status === 429) return { label: "限流", retryable: true };
+  if (status >= 500) return { label: `上游错误 ${status}`, retryable: true };
+  if (status === 404) return { label: "模型不存在", retryable: false };
+  return { label: `HTTP ${status}`, retryable: false };
+}
+
+export async function viaProbe({
+  peerUrl,
+  token,
+  providerId,
+  model,
+  prompt = "hi",
+  maxTokens = 5,
+  timeoutMs = 30000,
+  fetchImpl = globalThis.fetch,
+  clock = Date.now,
+  shareKeys,
+  shareKeysHeader,
+  relayTarget,
+  relayHeaders,
+  relayBody,
+  targetUrl,
+} = {}) {
+  const started = clock();
+  const t0 = typeof performance !== "undefined" && performance.now ? performance.now() : started;
+  const base = String(peerUrl || "").replace(/\/+$/, "");
+  if (!base) return { ok: false, label: "配置错误", error: "missing peerUrl", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
+  // 纯中继模式：A 把 targetUrl+headers+body 发给 B，B 原样 fetch 到上游（不查 B 本地 providerConfigs）
+  const rt = String(relayTarget || targetUrl || "").trim();
+  if (rt) {
+    const rh = relayHeaders && typeof relayHeaders === "object" ? relayHeaders : {};
+    const rb = relayBody !== undefined ? relayBody : null;
+    const relayUrl = joinUrl(base, "/v1/relay");
+    const relayHeadersOut = { "Content-Type": "application/json", Accept: "application/json" };
+    if (token) relayHeadersOut.Authorization = `Bearer ${token}`;
+    const payload = { targetUrl: rt, method: "POST", headers: rh, body: rb };
+    let ttfbMs = null;
+    try {
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
+      const fetchStart = typeof performance !== "undefined" && performance.now ? performance.now() : clock();
+      let res;
+      try { res = await fetchImpl(relayUrl, { method: "POST", headers: relayHeadersOut, body: JSON.stringify(payload), signal: controller.signal }); } finally { clearTimeout(timer); }
+      ttfbMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - fetchStart);
+      const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+      if (!res.ok) {
+        let txt = ""; try { txt = await res.text(); } catch {}
+        const cls = classifyError(res.status, txt);
+        const msg = extractInnerMessage(txt) || `HTTP ${res.status}`;
+        return { ok: false, status: res.status, label: cls.label, error: msg, ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+      }
+      let txt = ""; let j = {};
+      try { txt = await res.text(); j = JSON.parse(txt); } catch { j = {}; }
+      const relayStatus = res.headers.get("x-mslxdff-relay-status") ? Number(res.headers.get("x-mslxdff-relay-status")) : res.status;
+      if (relayStatus >= 400) {
+        const cls = classifyError(relayStatus, txt);
+        const msg = extractInnerMessage(txt) || `HTTP ${relayStatus}`;
+        return { ok: false, status: relayStatus, label: cls.label, error: msg, ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+      }
+      const usage = extractUsageFromJson(j);
+      const content = j?.choices?.[0]?.message?.content || j?.choices?.[0]?.text || txt || "";
+      const chars = typeof content === "string" ? content.length : 0;
+      const pt = usage?.prompt_tokens ?? null;
+      const ct = usage?.completion_tokens ?? null;
+      const { tps, charsPerSec } = computeMetrics({ ttfbMs, totalMs, promptTokens: pt, completionTokens: ct, chars });
+      const totalTokens = usage?.total_tokens ?? (pt !== null && ct !== null ? pt + ct : null);
+      return { ok: true, status: relayStatus, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: { prompt: pt, completion: ct, total: totalTokens }, chars };
+    } catch (e) {
+      const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+      const msg = e?.message || String(e);
+      const isTimeout = /timeout|abort/i.test(msg);
+      return { ok: false, label: isTimeout ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+    }
+  }
+  if (!model) return { ok: false, label: "配置错误", error: "missing model", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
+  let rawModel = String(model).trim();
+  if (providerId && rawModel.startsWith(`${providerId}/`)) rawModel = rawModel.slice(providerId.length + 1);
+  const url = joinUrl(base, "/v1/chat/completions");
+  const body = { model: rawModel, stream: false, messages: [{ role: "user", content: prompt }], max_tokens: maxTokens };
+  const headers = { "Content-Type": "application/json", Accept: "application/json" };
+  if (token) headers.Authorization = `Bearer ${token}`;
+  const sk = shareKeysHeader || shareKeys;
+  if (sk) headers["x-mslxdff-share-keys"] = String(sk);
+  let ttfbMs = null;
+  try {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
+    const fetchStart = typeof performance !== "undefined" && performance.now ? performance.now() : clock();
+    let res;
+    try {
+      res = await fetchImpl(url, { method: "POST", headers, body: JSON.stringify(body), signal: controller.signal });
+    } finally { clearTimeout(timer); }
+    ttfbMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - fetchStart);
+    if (res instanceof Error) throw res;
+    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+    if (!res.ok) {
+      let txt = "";
+      try { txt = await res.text(); } catch {}
+      const cls = classifyError(res.status, txt);
+      const msg = extractInnerMessage(txt) || `HTTP ${res.status}`;
+      return { ok: false, status: res.status, label: cls.label, error: msg, ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+    }
+    let json = {};
+    let txt = "";
+    try { txt = await res.text(); json = JSON.parse(txt); } catch { json = {}; }
+    const usage = extractUsageFromJson(json);
+    const content = json?.choices?.[0]?.message?.content || json?.choices?.[0]?.text || txt || "";
+    const chars = typeof content === "string" ? content.length : 0;
+    const promptTokens = usage?.prompt_tokens ?? null;
+    const completionTokens = usage?.completion_tokens ?? null;
+    const { tps, charsPerSec } = computeMetrics({ ttfbMs, totalMs, promptTokens, completionTokens, chars });
+    const totalTokens = usage?.total_tokens ?? (promptTokens !== null && completionTokens !== null ? promptTokens + completionTokens : null);
+    return { ok: true, status: res.status, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: { prompt: promptTokens, completion: completionTokens, total: totalTokens }, chars };
+  } catch (e) {
+    const totalMs = Math.round((typeof performance !== "undefined" && performance.now ? performance.now() : clock()) - t0);
+    const msg = e?.message || String(e);
+    const isTimeout = /timeout|abort/i.test(msg);
+    return { ok: false, label: isTimeout ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
+  }
+}

package/src/bench/via.js

@@ -0,0 +1,130 @@
+import { viaProbe } from "./via-probe.js";
+
+export async function resolveIncludeOpencode({ includeOpencode, isTTY, confirmFn, log = () => {} } = {}) {
+  if (!includeOpencode) return false;
+  if (!isTTY) {
+    try { log("via 已跳过 opencode（非 TTY，省额度；需对比请在 TTY 加 --include-opencode 并确认）"); } catch {}
+    return false;
+  }
+  if (typeof confirmFn === "function") {
+    try {
+      const ok = await confirmFn();
+      // confirmFn returns true for y, false otherwise
+      if (ok) return true;
+      try { log("已回落：via 跳过 opencode"); } catch {}
+      return false;
+    } catch { return false; }
+  }
+  return false;
+}
+
+export async function getOnlinePeers({ loadGroupsJoined, loadPeers, probeHealth } = {}) {
+  let candidates = [];
+  if (typeof loadPeers === "function") {
+    try { candidates = loadPeers() || []; } catch { candidates = []; }
+  } else {
+    try {
+      const m = await import("../state.js");
+      if (m.loadPeers) candidates = m.loadPeers() || [];
+    } catch {}
+  }
+  // also consider groupsJoined for validation - if no groups joined, via is meaningless, but we still allow static peers?
+  // For bench-via, if no groupsJoined at all, treat as no via capability -> return [] without probing?
+  // However static peers may exist without group; we probe them anyway.
+  // To satisfy "empty group -> skip probing" test, check loadGroupsJoined.
+  let joined = [];
+  if (typeof loadGroupsJoined === "function") {
+    try { joined = loadGroupsJoined() || []; } catch { joined = []; }
+  } else {
+    try {
+      const m = await import("../state.js");
+      if (m.loadGroupsJoined) joined = m.loadGroupsJoined() || [];
+    } catch {}
+  }
+  if (!joined.length) return [];
+  if (!candidates.length) return [];
+
+  // If probeHealth not provided, try import
+  let probe = probeHealth;
+  if (!probe) {
+    try {
+      const m = await import("../cli/group-helpers.js");
+      probe = m.probeHealth;
+    } catch { return candidates; }
+  }
+  if (!candidates.length) return [];
+  const online = [];
+  for (const p of candidates) {
+    try {
+      const r = await probe(p);
+      if (r && r.rank === 0 && !r.fail && !r.stale) online.push(p);
+    } catch {}
+  }
+  return online;
+}
+
+export async function orchestrateVia({
+  models = [],
+  peers = [],
+  directRunner = null,
+  viaProbeFn = viaProbe,
+  includeOpencode = false,
+  token,
+  timeoutMs = 30000,
+  clock = Date.now,
+} = {}) {
+  const filtered = includeOpencode ? models : models.filter((m) => {
+    const s = typeof m === "string" ? m : (m.id || m.model || m.provider || "");
+    const first = String(s).split("/")[0].toLowerCase();
+    const prov = String(m?.provider || m?.providerId || first).toLowerCase();
+    return prov !== "opencode" && first !== "opencode";
+  });
+  // normalize model entries
+  const normModels = filtered.map((m) => {
+    if (typeof m === "string") return { provider: m.split("/")[0], model: m, id: m };
+    const id = m.id || m.model || "";
+    const provider = m.provider || m.providerId || (id.includes("/") ? id.split("/")[0] : "");
+    return { provider, model: id, id };
+  });
+  const results = [];
+  for (const entry of normModels) {
+    const provider = entry.provider;
+    const model = entry.model;
+    // direct
+    let direct = null;
+    if (directRunner) {
+      try { direct = await directRunner({ provider, model, timeoutMs, clock }); } catch (e) { direct = { ok: false, label: "网络错误", error: String(e), ttfbMs: null, totalMs: 0 }; }
+    } else {
+      direct = { ok: false, label: "未配置 directRunner", error: "missing directRunner", ttfbMs: null, totalMs: 0 };
+    }
+    const via = {};
+    for (const peer of peers) {
+      const raw = String(peer.name || peer.id || peer.url || "peer");
+      let peerId = raw;
+      // 仅当 raw 本身就是 url 时才短化；显式 name/id 保持原样，避免 B → b:8989
+      if (!peer.name && !peer.id && raw.includes("://")) { try { const u = new URL(raw); peerId = u.hostname + (u.port ? `:${u.port}` : ""); } catch { peerId = raw.slice(-16); } }
+      const peerToken = peer.token || token || "";
+      try {
+        const r = await viaProbeFn({ peerUrl: peer.url, token: peerToken, providerId: provider, model, prompt: "hi", maxTokens: 5, timeoutMs, clock });
+        via[peerId] = r;
+      } catch (e) {
+        via[peerId] = { ok: false, label: "网络错误", error: String(e?.message || e), ttfbMs: null, totalMs: 0 };
+      }
+    }
+    // best
+    let best = "direct";
+    let bestTtfb = direct?.ok ? (direct.ttfbMs ?? direct.totalMs) : Infinity;
+    let deltaMs = null;
+    for (const [pid, rv] of Object.entries(via)) {
+      if (!rv.ok) continue;
+      const t = rv.ttfbMs ?? rv.totalMs;
+      if (t < bestTtfb) { bestTtfb = t; best = `via:${pid}`; }
+    }
+    if (best.startsWith("via:")) {
+      const d = direct?.ttfbMs ?? direct?.totalMs ?? 0;
+      deltaMs = bestTtfb - d;
+    }
+    results.push({ provider, model, direct, via, best, deltaMs, opencodeSkipped: !includeOpencode });
+  }
+  return results;
+}

package/src/chat/gateway.js

@@ -136,9 +136,13 @@ export function createGatewayClient({
       }
       return { ok: true, message: choice.message, usage: effectiveJ.usage || j.usage, raw: j, status: res.status, model: rawModel, provider, viaGateway: true };
     } catch (err) {
-      const msg = String(err?.message || err).slice(0, 800);
-      if (env.MSLXDFF_CHAT_TRACE !== "0") {
-        const dt2 = performance.now() - (performance.now() - 0); // placeholder, TRACE off in tests
+      const raw = String(err?.message || err);
+      const msg = raw.slice(0, 800);
+      const isConnRefused = /ECONNREFUSED|Failed to fetch|fetch failed|connect ECONNREFUSED|ECONNRESET|EADDRNOTAVAIL/i.test(raw);
+      if (isConnRefused) {
+        const friendly = `本地服务没有启动无法使用auto模式，请mslxdff -d 启动（本地网关 http://127.0.0.1:${port} 拒绝连接）— 3ms 内失败说明未触达任何模型（非模型额度问题）；若已改端口请用 mslxdff -port N 或 MSLXDFF_PORT=${port} 保持一致`;
+        if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto FAIL · ${Math.round(performance.now() - t0)}ms · ${friendly} · ${msg.slice(0, 80)}\x1b[0m`);
+        return { ok: false, error: friendly, status: 502, code: "GATEWAY_NOT_RUNNING", port, raw: msg };
       }
       return { ok: false, error: `gateway ${msg}`, status: 502 };
     }

package/src/chat/orchestrator.js

@@ -64,7 +64,7 @@ export function createOrchestrator({
     let first;
     let firstMs = 0;
     if (firstCooling) {
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 跳过（冷却中）· 直接试 ${CHAT_FALLBACK}\x1b[0m`);
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 跳过（冷却 10min）· 额度用完，准备走网关 auto（将尝试其他可用模型）\x1b[0m`);
       first = { ok: false, error: "skip cooling", status: 429 };
     } else {
       const t = perf.now();
@@ -81,13 +81,16 @@ export function createOrchestrator({
     }
 
     if (firstCooling) {
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 冷却中（${CHAT_COOLDOWN_MS / 60000}min）· 直接走网关 auto，跳过 ${CHAT_FALLBACK}\x1b[0m`);
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 冷却，直接走网关 auto（:8989）\x1b[0m`);
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 冷却中（${CHAT_COOLDOWN_MS / 60000}min）· 直接走网关 auto（:8989），按 auto 择优尝试其他可用模型，跳过 ${CHAT_FALLBACK} 直连\x1b[0m`);
       const t2 = TRACE ? perf.now() : 0;
       const third = await _gateway(opts);
       if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t2)}ms · 总 ${Math.round(perf.now() - t0)}ms (gateway-fallback)\x1b[0m`);
       if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: "skip big-pickle (mimo cooling)", viaGateway: true };
-      return { ok: false, error: `${CHAT_PREFERRED} cooling: ${first.error}; gateway auto failed: ${third.error}`, status: third.status || 429 };
+      const isGwDown = third.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(third.error || ""));
+      if (isGwDown) {
+        return { ok: false, error: `${CHAT_PREFERRED} 冷却跳过（额度用完 10min）；${third.error} —— 本次 auto 未能尝试任何其他模型，因网关未就绪（3ms 失败即证明未触达上游，请先启动网关再重试）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+      }
+      return { ok: false, error: `${CHAT_PREFERRED} 冷却跳过（额度用完）；网关 auto 已尝试其他可用模型但均失败：${third.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: third.status || 429 };
     }
 
     const secondCooling = await isCoolingAsync(CHAT_FALLBACK);
@@ -102,12 +105,14 @@ export function createOrchestrator({
     };
 
     if (secondCooling) {
-      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} 跳过（冷却中）· 直接走网关 auto\x1b[0m`);
+      if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_FALLBACK} 跳过（冷却中）· 直接走网关 auto（将尝试其他可用模型）\x1b[0m`);
       const t2 = perf.now();
       const third = await doGateway();
       if (TRACE) console.log(`\x1b[90m· [LLM] gateway auto ${third.ok ? "OK" : "FAIL"} · ${Math.round(perf.now() - t2)}ms · 总 ${Math.round(perf.now() - t0)}ms (gateway-fallback)\x1b[0m`);
       if (third.ok) return { ...third, model: third.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: "skip cooling", viaGateway: true };
-      return { ok: false, error: `${CHAT_PREFERRED} failed: ${first.error}; ${CHAT_FALLBACK} failed: skip cooling; gateway auto failed: ${third.error}`, status: third.status || first.status };
+      const isGwDown2 = third.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(third.error || ""));
+      if (isGwDown2) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 冷却跳过；${third.error} —— auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+      return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 冷却跳过；网关 auto 已尝试其他模型但均失败：${third.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: third.status || first.status };
     }
 
     if (TRACE) console.log(`\x1b[90m· [LLM] ${CHAT_PREFERRED} 失败，${CHAT_FALLBACK} + gateway 对冲中（${HEDGE_MS}ms）\x1b[0m`);
@@ -167,7 +172,9 @@ export function createOrchestrator({
       secondRes = sRes; gatewayRes = gRes;
       if (sRes?.ok) return { ...sRes, model: CHAT_FALLBACK, fallback: true, firstError: first.error };
       if (gRes?.ok) return { ...gRes, model: gRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: sRes?.error, viaGateway: true };
-      return { ok: false, error: `${CHAT_PREFERRED} failed: ${first.error}; ${CHAT_FALLBACK} failed: ${sRes?.error}; gateway auto failed: ${gRes?.error}`, status: gRes?.status || sRes?.status || first.status };
+      const gwDown0 = gRes?.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(gRes?.error || ""));
+      if (gwDown0) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${sRes?.error}；${gRes?.error} —— gateway 侧 auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+      return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${sRes?.error}；网关 auto 已尝试其他模型但均失败：${gRes?.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: gRes?.status || sRes?.status || first.status };
     }
 
     const raced = await raceFirstOk();
@@ -183,7 +190,9 @@ export function createOrchestrator({
       try { gatewayRes = await (gatewayPromise || doGateway()); } catch (e) { gatewayRes = { ok: false, error: String(e), status: 502 }; }
     }
     if (gatewayRes?.ok) return { ...gatewayRes, model: gatewayRes.model || "auto", fallbackGateway: true, fallback: true, firstError: first.error, secondError: secondRes?.error, viaGateway: true };
-    return { ok: false, error: `${CHAT_PREFERRED} failed: ${first.error}; ${CHAT_FALLBACK} failed: ${secondRes?.error}; gateway auto failed: ${gatewayRes?.error}`, status: gatewayRes?.status || secondRes?.status || first.status };
+    const gwDown = gatewayRes?.code === "GATEWAY_NOT_RUNNING" || /本地(网关未运行|服务没有启动)/.test(String(gatewayRes?.error || ""));
+    if (gwDown) return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${secondRes?.error}；${gatewayRes?.error} —— gateway 侧 auto 未能尝试其他模型（网关未就绪）`, status: 502, code: "GATEWAY_NOT_RUNNING" };
+    return { ok: false, error: `${CHAT_PREFERRED} 失败：${first.error}；${CHAT_FALLBACK} 失败：${secondRes?.error}；网关 auto 已尝试其他模型但均失败：${gatewayRes?.error}（总 ${Math.round(perf.now() - t0)}ms）`, status: gatewayRes?.status || secondRes?.status || first.status };
   }
 
   async function summarizeHistory(messages) {

package/src/chat/repl.js

@@ -6,7 +6,7 @@ import { getToolDefs, execCommand, readFileTool, curlTool } from "./tools.js";
 import { chatWithFallback, summarizeHistory } from "./upstream.js";
 import { loadHistory, saveHistory, clearHistory, histPath, estimateChars, needsCompress } from "./store.js";
 import { CHAT_KEEP_RECENT, CHAT_MAX_TOOL_LOOPS, CHAT_PREFERRED, CHAT_FALLBACK } from "./config.js";
-import { formatBannerLines, formatStatsDetail, collectStats } from "./stats.js";
+import { formatBannerLines, formatStatsDetail, collectStats, probeGateway } from "./stats.js";
 import { createSpinner } from "./spinner.js";
 import { normalizeFullId } from "../providers/model-id.js";
 
@@ -18,9 +18,17 @@ const SLASH_HELP = `自然语言直接说，斜杠快捷：
   /exit     退出
 示例：设置 hy3 为默认模型 / 查看组列表 / 看最近20条日志 / 读一下 src/logs.js`;
 
-function printBanner() {
-  const { lines } = formatBannerLines();
+async function printBanner() {
+  let probe = null;
+  try {
+    const s = collectStats();
+    probe = await probeGateway(s.port, 800);
+  } catch {}
+  const { lines } = formatBannerLines(probe);
   for (const l of lines) console.log(l);
+  if (probe && !probe.alive) {
+    console.log(`\x1b[31m本地服务没有启动无法使用auto模式，请mslxdff -d 启动\x1b[0m`);
+  }
   console.log(`\x1b[90m输入自然语言即可执行；/help 帮助，/stats 看网关统计，/exit 退出\x1b[0m`);
   console.log(`\x1b[90m历史：${histPath()} · 仅拦截 -uninstall · 数据来自网关 -d，非本会话计数\x1b[0m`);
 }
@@ -283,7 +291,7 @@ export async function startRepl({ singleShot } = {}) {
     saveHistory(messages.slice(1));
     return;
   }
-  printBanner();
+  await printBanner();
   const rl = readline.createInterface({ input: stdin, output: stdout, prompt: `\x1b[36m${CHAT_PREFERRED.split("-")[0]}>\x1b[0m ` });
   rl.prompt();
   for await (const line of rl) {

package/src/chat/stats.js

@@ -141,17 +141,40 @@ export function collectStats() {
   };
 }
 
-export function formatBannerLines() {
+export async function probeGateway(port, timeoutMs = 800) {
+  const url = `http://127.0.0.1:${port}/health`;
+  const ctrl = new AbortController();
+  const t = setTimeout(() => ctrl.abort(), timeoutMs);
+  try {
+    const r = await fetch(url, { signal: ctrl.signal });
+    clearTimeout(t);
+    return { alive: r.ok, status: r.status, ms: 0 };
+  } catch (e) {
+    clearTimeout(t);
+    const msg = String(e?.message || e);
+    const isRefused = /ECONNREFUSED|Failed to fetch|fetch failed|ECONNRESET|abort/i.test(msg);
+    return { alive: false, status: 0, error: msg, refused: isRefused };
+  }
+}
+
+export function formatBannerLines(healthProbe = null) {
   const s = collectStats();
   const dim = "\x1b[90m";
   const rst = "\x1b[0m";
   const cyan = "\x1b[36m";
   const yellow = "\x1b[33m";
   const green = "\x1b[32m";
+  const red = "\x1b[31m";
+  const bgRed = "\x1b[41m\x1b[37m";
   const lines = [];
   lines.push(`${cyan}┌─ mslxdff chat  ·  数据来自 -d 网关进程（非本会话） ─────${rst}`);
+  if (healthProbe && !healthProbe.alive) {
+    lines.push(`${red}│  ✗ 本地服务没有启动无法使用auto模式，请mslxdff -d 启动  ·  ${s.healthUrl} 拒绝连接（${healthProbe.error?.slice(0,60) || "ECONNREFUSED"}）${rst}`);
+    lines.push(`${red}│  → 先在另一终端执行: mslxdff -d  (或 mslxdff)  启动后回此窗口重试，期间仅 mimo/big-pickle 直连可用，auto 不可用${rst}`);
+  }
   lines.push(`${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}${rst}`);
+  const gwAliveTag = healthProbe ? (healthProbe.alive ? `${green}● 运行中${rst}` : `${red}● 未运行${rst}`) : `${dim}…检测中${rst}`;
+  lines.push(`${cyan}│${rst}  网关默认  ${green}${s.gatewayModel}${rst} ${dim}[${s.gatewayStatus}]${rst}  ·  端口 ${s.port} ${gwAliveTag}  ·  ${dim}${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;
   const gateTtfb = s.gatewayStat?.avgTtfbMs ?? s.gatewayStat?.emaTtfbMs ?? s.gatewayLat?.emaMs;

package/src/cli/commands/provider/bench.js

@@ -1,13 +1,11 @@
 import { probeModels } from "../../../bench/probe.js";
 import { runOne } from "../../../bench/runner.js";
-import { formatReport } from "../../../bench/report.js";
+import { formatReport, formatViaReport } from "../../../bench/report.js";
 import { defaultModelsPath, defaultChatPath } from "../../../state/provider-config.js";
 import { isRefreshToken, clineHeaders } from "../../../providers/cline/headers.js";
 import { refreshTokenForBase } from "../../../providers/cline/auth.js";
 import { computeMetrics } from "../../../metrics.js";
 
-// Cline 免费通道（deepseek/z-ai 等）非流式会被上游限流 500 empty response content，
-// 必须 stream:true + SSE 聚合后测速
 async function clineBenchOne({ baseUrl, model, accessToken, prompt, maxTokens, timeoutMs, fetchImpl }) {
   const controller = new AbortController();
   const timer = setTimeout(() => controller.abort(new Error(`timeout ${timeoutMs}ms`)), timeoutMs);
@@ -25,7 +23,6 @@ async function clineBenchOne({ baseUrl, model, accessToken, prompt, maxTokens, t
     if (!res.ok) {
       let txt = "";
       try { txt = await res.text(); } catch {}
-      const low = txt.toLowerCase();
       const label = res.status === 401 ? "鉴权失败" : res.status === 429 ? "限流" : res.status >= 500 ? `上游错误 ${res.status}` : `HTTP ${res.status}`;
       return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: Math.round(performance.now() - t0), tps: null, charsPerSec: null, tokens: null };
     }
@@ -59,16 +56,18 @@ async function clineBenchOne({ baseUrl, model, accessToken, prompt, maxTokens, t
   } catch (e) {
     const msg = e?.message || String(e);
     return { id: model, ok: false, label: /timeout|abort/i.test(msg) ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs: Math.round(performance.now() - t0), tps: null, charsPerSec: null, tokens: null };
-  } finally {
-    clearTimeout(timer);
-  }
+  } finally { clearTimeout(timer); }
 }
 
 function parseBenchArgs(rest) {
-  const opts = { json: false, prompt: "hi", maxTokens: 32, timeoutMs: 30000 };
+  const opts = { json: false, prompt: "hi", maxTokens: 32, timeoutMs: 30000, via: false, includeOpencode: false, samples: 1 };
   for (let i = 0; i < rest.length; i++) {
     const a = rest[i];
     if (a === "--json" || a === "-json") opts.json = true;
+    else if (a === "--via" || a === "--bench-via") opts.via = true;
+    else if (a === "--include-opencode") opts.includeOpencode = true;
+    else if (a === "--samples" && rest[i + 1]) opts.samples = Number(rest[++i]) || 1;
+    else if (a.startsWith("--samples=")) opts.samples = Number(a.split("=")[1]) || 1;
     else if (a === "--prompt" && rest[i + 1]) opts.prompt = rest[++i];
     else if (a.startsWith("--prompt=")) opts.prompt = a.slice(9);
     else if (a === "--max-tokens" && rest[i + 1]) opts.maxTokens = Number(rest[++i]) || 32;
@@ -91,33 +90,164 @@ function buildHeadersForProvider(providerId, apiKey, auth) {
     if (apiKey) h["Authorization"] = `Bearer ${apiKey}`;
     if (auth?.uid) h["X-User-Id"] = auth.uid;
     h["X-Domain"] = auth?.domain || "www.codebuddy.cn";
-    if (auth?.enterpriseId) {
-      h["X-Enterprise-Id"] = auth.enterpriseId;
-      h["X-Tenant-Id"] = auth.enterpriseId;
-    }
+    if (auth?.enterpriseId) { h["X-Enterprise-Id"] = auth.enterpriseId; h["X-Tenant-Id"] = auth.enterpriseId; }
     return h;
   }
   if (apiKey) h["Authorization"] = `Bearer ${apiKey}`;
   return h;
 }
 
+async function handleVia({ providerId, opts, fetchImpl, loadConfigs, loadKeys, loadAllowed, loadBaseUrl }) {
+  const { getOnlinePeers, orchestrateVia, resolveIncludeOpencode } = await import("../../../bench/via.js");
+  const peers = await getOnlinePeers();
+  if (!peers.length) {
+    const msg = "未加入组或无在线 peer，--via 无意义。先 mslxdff -group list / -addtogroup";
+    if (opts.json) console.log(JSON.stringify({ meta: { at: new Date().toISOString(), samples: opts.samples, timeout: opts.timeoutMs, includeOpencode: false, peers: [], opencodeSkipped: true }, results: [], advice: msg }, null, 2));
+    else console.log(msg);
+    process.exit(0);
+  }
+  const isTTY = Boolean(process.stdin.isTTY && process.stdout.isTTY);
+  let includeOpencode = opts.includeOpencode;
+  if (includeOpencode) {
+    const confirmFn = async () => {
+      const readline = await import("node:readline");
+      return new Promise((res) => {
+        const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
+        rl.question("将消耗 B/C/D 的 opencode 额度，确认测 opencode via？y/N ", (ans) => { rl.close(); res(ans.trim().toLowerCase() === "y"); });
+      });
+    };
+    const log = (s) => (opts.json ? console.error(s) : console.log(s));
+    includeOpencode = await resolveIncludeOpencode({ includeOpencode, isTTY, confirmFn, log });
+  }
+  const { loadToken } = await import("../../../state.js");
+  let token = "";
+  try { token = (await loadToken()).token || ""; } catch {}
+  // target providers
+  let targetIds = [];
+  const isAll = providerId === "bench" || providerId === "all";
+  if (isAll) {
+    const configs = loadConfigs();
+    const ids = new Set(Object.keys(configs));
+    ids.add("opencode"); ids.add("openrouter");
+    try { const m = await import("../../../state.js"); const raw = m.loadProviderKeys ? null : null; } catch {}
+    // collect from providerKeys via loadKeys probing? simple: iterate ids and keep those with allowed
+    for (const pid of [...ids]) {
+      const allowed = loadAllowed(pid) || [];
+      if (allowed.length) targetIds.push(pid);
+      else if (pid === "opencode" && !includeOpencode) continue;
+    }
+    // also check generic ids from state raw
+    try {
+      const { readFileSync } = await import("node:fs");
+      const { defaultStateFile } = await import("../../../state.js");
+      const raw = JSON.parse(readFileSync(defaultStateFile(), "utf8"));
+      for (const k of Object.keys(raw.providerConfigs || {})) if (!targetIds.includes(k) && (loadAllowed(k) || []).length) targetIds.push(k);
+      for (const k of Object.keys(raw.providerKeys || {})) if (!targetIds.includes(k) && (loadAllowed(k) || []).length) targetIds.push(k);
+    } catch {}
+    if (!targetIds.length) targetIds = ["openrouter", "workbuddy", "clinebot"].filter((p) => (loadAllowed(p) || []).length);
+  } else {
+    targetIds = [providerId];
+  }
+  const allResults = [];
+  const viaLog = (s) => (opts.json ? console.error(s) : console.log(s));
+  viaLog(`bench-via: direct vs via peers (samples=${opts.samples}, timeout=${opts.timeoutMs}ms, ${includeOpencode ? "opencode=included" : "opencode=skipped"}) peers=${peers.map((p) => p.id).join(",")}`);
+  for (const pid of targetIds) {
+    const cfg = (loadConfigs()[pid] || {});
+    const keys = loadKeys(pid) || [];
+    const allowed = loadAllowed(pid) || [];
+    const baseUrl = (loadBaseUrl(pid) || cfg.baseUrl || "").trim();
+    if (!allowed.length) {
+      viaLog(`provider ${pid}: 无勾选模型，跳过`);
+      continue;
+    }
+    if (!baseUrl && pid !== "opencode") { viaLog(`provider ${pid}: missing baseUrl 跳过`); continue; }
+    if (!keys.length && pid !== "opencode") { viaLog(`provider ${pid}: 未配置 Key 跳过`); continue; }
+    const chatPath = cfg.chatPath || defaultChatPath(pid);
+    let auths = [];
+    try { const m = await import("../../../state.js"); auths = m.loadProviderAuths ? m.loadProviderAuths(pid) : []; } catch {}
+    const models = allowed.map((id) => ({ provider: pid, model: String(id), id: String(id) }));
+    // directRunner per provider
+    const directRunner = async ({ provider, model }) => {
+      const p = provider || pid;
+      const idx = models.findIndex((x) => x.model === model);
+      const kIdx = idx >= 0 ? idx % (keys.length || 1) : 0;
+      const aIdx = Math.min(kIdx, Math.max(auths.length - 1, 0));
+      const key = keys[kIdx] || keys[0] || "";
+      const auth = auths[aIdx] || auths[0] || null;
+      const cKeys = keys.filter((k) => isRefreshToken(k));
+      if (cKeys.length) {
+        const normBase = String(baseUrl).replace(/\/+$/, "");
+        const chatBase = normBase.endsWith("/api/v1") ? normBase.slice(0, -7) : normBase;
+        const rt = cKeys[kIdx % cKeys.length];
+        const at = await refreshTokenForBase({ refreshToken: rt, baseUrl: normBase, fetchImpl });
+        if (!at) return { ok: false, label: "鉴权失败", error: "refresh failed", ttfbMs: null, totalMs: 0 };
+        return clineBenchOne({ baseUrl: chatBase, model, accessToken: at, prompt: "hi", maxTokens: 5, timeoutMs: opts.timeoutMs, fetchImpl });
+      }
+      const headers = buildHeadersForProvider(p, key, auth);
+      return runOne({ baseUrl: p === "opencode" ? (process.env.UPSTREAM_BASE_URL || "https://opencode.ai") : baseUrl, chatPath, model, providerId: p, apiKey: key || "public", headers, prompt: "hi", maxTokens: 5, timeoutMs: opts.timeoutMs, fetchImpl });
+    };
+    const { viaProbe } = await import("../../../bench/via-probe.js");
+    const { joinUrl: _joinUrl } = await import("../../../providers/base.js");
+    const viaProbeFn = async (args) => {
+      // 纯中继：A 把 targetUrl+headers+body 发给 B，B 原样 fetch 到上游（不查 B 本地 providerConfigs）
+      const { peerUrl, providerId, model } = args;
+      const p = providerId || pid;
+      const idx = models.findIndex((x) => x.model === model);
+      const kIdx = idx >= 0 ? idx % (keys.length || 1) : 0;
+      const aIdx = Math.min(kIdx, Math.max(auths.length - 1, 0));
+      const key = keys[kIdx] || keys[0] || "";
+      const auth = auths[aIdx] || auths[0] || null;
+      const cKeys = keys.filter((k) => isRefreshToken(k));
+      if (cKeys.length) {
+        // cline refreshToken 需特殊流，暂走旧的 peer chat（peer 需自有 cline 配置）
+        return viaProbe({ ...args, token, fetchImpl });
+      }
+      const rawModel = String(model).startsWith(`${p}/`) ? String(model).slice(p.length + 1) : String(model);
+      const targetUrl = p === "opencode" ? `${process.env.UPSTREAM_BASE_URL || "https://opencode.ai"}/zen/v1/chat/completions` : `${String(baseUrl).replace(/\/+$/, "")}${chatPath}`;
+      const headers = buildHeadersForProvider(p, key, auth);
+      const body = { model: rawModel, stream: false, messages: [{ role: "user", content: "hi" }], max_tokens: 5 };
+      // 用对端 peer.token 做 relay 鉴权（B 只认自己的 Bearer）
+      const peer = peers.find((pe) => pe.url === peerUrl || (pe.name || pe.id) === peerUrl);
+      const peerToken = peer?.token || token;
+      return viaProbe({ peerUrl, token: peerToken, relayTarget: targetUrl, relayHeaders: headers, relayBody: body, timeoutMs: opts.timeoutMs, fetchImpl });
+    };
+    const part = await orchestrateVia({ models, peers, directRunner, viaProbeFn, includeOpencode, token, timeoutMs: opts.timeoutMs });
+    allResults.push(...part);
+    if (!opts.json) viaLog(`  ${pid}: ${part.length} 模型完成`);
+  }
+  const meta = { at: new Date().toISOString(), samples: opts.samples, timeout: opts.timeoutMs, includeOpencode, peers: peers.map((p) => p.id), opencodeSkipped: !includeOpencode };
+  const report = formatViaReport(allResults, { peers, meta, json: opts.json });
+  if (opts.json) console.log(report.text);
+  else console.log("\n" + report.text);
+  process.exit(0);
+}
+
 export async function handleProviderBench(id, sub, rest, args, deps = {}) {
-  const isBench = sub === "bench" || sub === "benchmark" || sub === "eval" || sub === "test";
+  const _restArr = rest || [];
+  const _hasVia = _restArr.includes("--via") || _restArr.includes("--bench-via");
+  const _isBenchViaAll = (String(id) === "bench" || String(id) === "all") && _hasVia;
+  const isBench = sub === "bench" || sub === "benchmark" || sub === "eval" || sub === "test" || _isBenchViaAll;
   if (!isBench) return false;
+  const opts = parseBenchArgs(_restArr);
+  // via branch
+  if (opts.via) {
+    const loadConfigs = deps.loadProviderConfigs || (await import("../../../state.js")).loadProviderConfigs;
+    const loadKeys = deps.loadProviderKeys || (await import("../../../state.js")).loadProviderKeys;
+    const loadAllowed = deps.loadProviderAllowedModels || (await import("../../../state.js")).loadProviderAllowedModels;
+    const loadBaseUrl = deps.loadProviderBaseUrl || (await import("../../../state.js")).loadProviderBaseUrl;
+    const viaPid = _isBenchViaAll ? "bench" : String(id || "").trim();
+    if (!viaPid) { console.error("usage: mslxdff -provider <id> bench --via [--json] [--include-opencode]"); process.exit(1); }
+    await handleVia({ providerId: viaPid, opts, fetchImpl: deps.fetchImpl || globalThis.fetch, loadConfigs, loadKeys, loadAllowed, loadBaseUrl });
+    return true;
+  }
   const fetchImpl = deps.fetchImpl || globalThis.fetch;
   const loadConfigs = deps.loadProviderConfigs || (await import("../../../state.js")).loadProviderConfigs;
   const loadKeys = deps.loadProviderKeys || (await import("../../../state.js")).loadProviderKeys;
   const loadAllowed = deps.loadProviderAllowedModels || (await import("../../../state.js")).loadProviderAllowedModels;
   const loadAllowAny = deps.loadProviderAllowAnyModels || (await import("../../../state.js")).loadProviderAllowAnyModels;
   const loadBaseUrl = deps.loadProviderBaseUrl || (await import("../../../state.js")).loadProviderBaseUrl;
-  const loadAuths = deps.loadProviderAuths || (async () => { try { const m = await import("../../../state.js"); return m.loadProviderAuths ? m.loadProviderAuths(id) : []; } catch { return []; } }) && (await import("../../../state.js")).loadProviderAuths;
-
   const providerId = String(id || "").trim();
-  if (!providerId) {
-    console.error("usage: mslxdff -provider <id> bench [--json] [--prompt hi] [--max-tokens 32]");
-    process.exit(1);
-  }
-  const opts = parseBenchArgs(rest || []);
+  if (!providerId) { console.error("usage: mslxdff -provider <id> bench [--json] [--prompt hi] [--max-tokens 32]"); process.exit(1); }
   const configs = loadConfigs();
   const cfg = configs[providerId] || {};
   const keys = loadKeys(providerId) || [];
@@ -126,38 +256,19 @@ export async function handleProviderBench(id, sub, rest, args, deps = {}) {
   const baseUrl = (loadBaseUrl(providerId) || cfg.baseUrl || "").trim();
   let auths = [];
   try { const m = await import("../../../state.js"); auths = m.loadProviderAuths ? m.loadProviderAuths(providerId) : []; } catch {}
-  if (!baseUrl) {
-    console.error(`provider ${providerId}: missing baseUrl — 先设置: mslxdff -provider ${providerId} set-url https://api.example.com/v1`);
-    process.exit(1);
-  }
-  if (!keys.length) {
-    console.error(`provider ${providerId}: 未配置 Key — 先设置: mslxdff -provider ${providerId} <key>`);
-    process.exit(1);
-  }
-  // 空勾选 → 不发 chat，仅探活模型列表并提示
+  if (!baseUrl) { console.error(`provider ${providerId}: missing baseUrl — 先设置: mslxdff -provider ${providerId} set-url https://api.example.com/v1`); process.exit(1); }
+  if (!keys.length) { console.error(`provider ${providerId}: 未配置 Key — 先设置: mslxdff -provider ${providerId} <key>`); process.exit(1); }
   if (!allowed.length) {
     const modelsPath = cfg.modelsPath || defaultModelsPath(providerId);
     console.log(`provider ${providerId}: 未设置 allowlist（allowAny=${allowAny ? "ON" : "OFF"}），不发起测速，仅探活模型列表...`);
     console.log(`尝试：GET ${baseUrl}${modelsPath} → GET ${baseUrl}/v1/models → GET ${baseUrl}/models`);
     const headers = buildHeadersForProvider(providerId, keys[0], auths[0]);
     const probed = await probeModels({ baseUrl, modelsPath, headers, fetchImpl, timeoutMs: 8000 });
-    if (!probed.ok) {
-      console.error(`探活失败：${probed.error}`);
-      console.error(`已尝试：${probed.tried.join(", ")}`);
-      console.error(`请手动设置 allowlist：mslxdff -provider ${providerId} allowlist set <model>`);
-      if (opts.json) console.log(JSON.stringify({ ok: false, error: probed.error, tried: probed.tried }, null, 2));
-      process.exit(1);
-    }
+    if (!probed.ok) { console.error(`探活失败：${probed.error}`); console.error(`已尝试：${probed.tried.join(", ")}`); console.error(`请手动设置 allowlist：mslxdff -provider ${providerId} allowlist set <model>`); if (opts.json) console.log(JSON.stringify({ ok: false, error: probed.error, tried: probed.tried }, null, 2)); process.exit(1); }
     const list = probed.data || [];
-    if (!list.length) {
-      console.log("探活成功但返回空列表，请确认上游是否暴露 /v1/models");
-      if (opts.json) console.log(JSON.stringify({ ok: true, data: [] }, null, 2));
-      process.exit(0);
-    }
+    if (!list.length) { console.log("探活成功但返回空列表，请确认上游是否暴露 /v1/models"); if (opts.json) console.log(JSON.stringify({ ok: true, data: [] }, null, 2)); process.exit(0); }
     console.log(`\n发现 ${list.length} 个模型：`);
-    for (const m of list.slice(0, 30)) {
-      console.log(`  - ${m.id || m.model || m.name || JSON.stringify(m).slice(0, 80)}`);
-    }
+    for (const m of list.slice(0, 30)) console.log(`  - ${m.id || m.model || m.name || JSON.stringify(m).slice(0, 80)}`);
     if (list.length > 30) console.log(`  ... 还有 ${list.length - 30} 个未展示`);
     console.log(`\n下一步：勾选后再测（只测勾选，避免扣费）`);
     console.log(`  mslxdff -provider ${providerId} allowlist set ${list.slice(0, 2).map((m) => m.id).join(" ")}`);
@@ -165,14 +276,10 @@ export async function handleProviderBench(id, sub, rest, args, deps = {}) {
     if (opts.json) console.log(JSON.stringify({ ok: true, data: list, hint: `pick then bench` }, null, 2));
     process.exit(0);
   }
-
-  // 有勾选 → 逐个测
   const log = (s) => (opts.json ? console.error(s) : console.log(s));
   log(`bench ${providerId}: 共 ${allowed.length} 个已勾选模型，逐个测速（串行，${opts.timeoutMs}ms 超时）...`);
   if (!opts.json) console.log(`prompt="${opts.prompt}" maxTokens=${opts.maxTokens}\n`);
   const chatPath = cfg.chatPath || defaultChatPath(providerId);
-  // Cline 新链：keys 含 refreshToken 时走 refresh→workos token + 指纹头（绕 403/401），
-  // 且 chat 固定拼 https://<host>/api/v1/chat/completions（剥掉 baseUrl 里可能带的 /api/v1）
   const rtKeys = keys.filter((k) => isRefreshToken(k));
   const normBase = String(baseUrl).replace(/\/+$/, "");
   const clineChatBase = normBase.endsWith("/api/v1") ? normBase.slice(0, -7) : normBase;
@@ -181,46 +288,27 @@ export async function handleProviderBench(id, sub, rest, args, deps = {}) {
     const raw = allowed[i];
     const model = String(raw || "").trim();
     if (!opts.json) process.stdout.write(`  [${i + 1}/${allowed.length}] ${model} ... `);
-    // 轮询 key/auth：按索引取，超长循环
     const kIdx = i % (keys.length || 1);
     const aIdx = Math.min(kIdx, Math.max(auths.length - 1, 0));
     const key = keys[kIdx];
     const auth = auths[aIdx] || auths[0] || null;
-
     if (rtKeys.length) {
       const rt = rtKeys[kIdx % rtKeys.length];
       const at = await refreshTokenForBase({ refreshToken: rt, baseUrl: normBase, fetchImpl });
-      if (!at) {
-        const r = { id: model, ok: false, error: "refreshToken 换 accessToken 失败", label: "鉴权失败", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null };
-        results.push(r);
-        if (!opts.json) console.log(`FAIL ${r.label} (${r.error})`);
-        continue;
-      }
+      if (!at) { const r = { id: model, ok: false, error: "refreshToken 换 accessToken 失败", label: "鉴权失败", ttfbMs: null, totalMs: 0, tps: null, charsPerSec: null, tokens: null }; results.push(r); if (!opts.json) console.log(`FAIL ${r.label} (${r.error})`); continue; }
       const r = await clineBenchOne({ baseUrl: clineChatBase, model, accessToken: at, prompt: opts.prompt, maxTokens: opts.maxTokens, timeoutMs: opts.timeoutMs, fetchImpl });
       results.push(r);
-      if (!opts.json) {
-        if (r.ok) console.log(`OK  TTFB ${r.ttfbMs}ms  总 ${r.totalMs}ms  ${r.tps != null ? `${r.tps} t/s` : r.charsPerSec != null ? `${r.charsPerSec} 字/秒` : "—"}`);
-        else console.log(`FAIL ${r.label} ${r.error ? `(${r.error.slice(0, 60)})` : ""}`);
-      }
+      if (!opts.json) { if (r.ok) console.log(`OK  TTFB ${r.ttfbMs}ms  总 ${r.totalMs}ms  ${r.tps != null ? `${r.tps} t/s` : r.charsPerSec != null ? `${r.charsPerSec} 字/秒` : "—"}`); else console.log(`FAIL ${r.label} ${r.error ? `(${r.error.slice(0, 60)})` : ""}`); }
       continue;
     }
-
     const headers = buildHeadersForProvider(providerId, key, auth);
     const r = await runOne({ baseUrl, chatPath, model, providerId, apiKey: key, headers, prompt: opts.prompt, maxTokens: opts.maxTokens, timeoutMs: opts.timeoutMs, fetchImpl });
     results.push(r);
-    if (!opts.json) {
-      if (r.ok) console.log(`OK  TTFB ${r.ttfbMs}ms  总 ${r.totalMs}ms  ${r.tps != null ? `${r.tps} t/s` : r.charsPerSec != null ? `${r.charsPerSec} 字/秒` : "—"}`);
-      else console.log(`FAIL ${r.label} ${r.error ? `(${r.error.slice(0, 60)})` : ""}`);
-    }
+    if (!opts.json) { if (r.ok) console.log(`OK  TTFB ${r.ttfbMs}ms  总 ${r.totalMs}ms  ${r.tps != null ? `${r.tps} t/s` : r.charsPerSec != null ? `${r.charsPerSec} 字/秒` : "—"}`); else console.log(`FAIL ${r.label} ${r.error ? `(${r.error.slice(0, 60)})` : ""}`); }
   }
   const report = formatReport(results, { json: opts.json });
   console.log("\n" + report.text);
-  if (opts.json) {
-    // json 已在 text 中输出一次（report.text 是 JSON），无需重复
-  }
   const failed = results.filter((r) => !r.ok).length;
-  if (failed && !opts.json) {
-    console.log(`\n提示：失败 ${failed} 个多为 402余额不足/429限流/超时，可清冷却或换 Key 后重试`);
-  }
+  if (failed && !opts.json) console.log(`\n提示：失败 ${failed} 个多为 402余额不足/429限流/超时，可清冷却或换 Key 后重试`);
   process.exit(failed ? 2 : 0);
 }

package/src/providers/cline/headers.js

@@ -23,8 +23,8 @@ export function clineHeaders(sessionId, token) {
 export function isRefreshToken(key) {
   const s = String(key || "").trim();
   if (!s) return false;
-  // sk_ 形态直接视为旧直连 key，不走 refresh 链
-  if (s.startsWith("sk_")) return false;
+  // sk_ / sk- 形态直接视为旧直连 key（OpenAI 兼容网关），不走 refresh 链
+  if (s.startsWith("sk_") || s.startsWith("sk-")) return false;
   // refreshToken 通常为 JWT 或长随机串，长度 > 20 且含 . 或 -
   if (s.length > 20) return true;
   return false;

package/src/routes/index.js

@@ -5,6 +5,7 @@ import { chatHandler } from "./chat.js";
 import { joinHandler, leaveHandler } from "./groups.js";
 import { heartbeatHandler, pollHandler, resultHandler, forwardHandler } from "./groups-relay.js";
 import { modelsHandler, modelsStatusHandler, providerModelsHandler } from "./models-route.js";
+import { relayHandler } from "./relay.js";
 
 export function createRouter({ token, upstream, models, auto, logs, peers, maxHops = DEFAULT_MAX_HOPS, groups, bans, bus, plugins }) {
   return async function router(req, res) {
@@ -78,6 +79,12 @@ const ROUTES = [
     requiresAuth: true,
     handler: forwardHandler,
   },
+  {
+    method: "POST",
+    path: "/v1/relay",
+    requiresAuth: true,
+    handler: relayHandler,
+  },
   {
     method: "GET",
     path: "/v1/models",

package/src/routes/relay.js

@@ -0,0 +1,59 @@
+import { readBody, json, authorized } from "./helpers.js";
+
+// POST /v1/relay  纯网络中继：A 把 targetUrl+headers+body 发给 B，B 原样 fetch 到上游再回给 A
+// B 侧不查本地 providerConfigs、不做 model 前缀路由、不验 allowlist，仅当 TCP 出口
+// 鉴权：复用全局 Bearer token（peer.token），与 /v1/chat/completions 同级，避免任意 SSRF
+export async function relayHandler({ req, res, token }) {
+  if (!authorized(req, token)) {
+    res.statusCode = 401;
+    res.setHeader("WWW-Authenticate", "Bearer");
+    return json(res, 401, { error: "Unauthorized" });
+  }
+  let body;
+  try { body = await readBody(req); } catch { return json(res, 400, { error: "Invalid JSON body" }); }
+  const targetUrl = String(body.targetUrl || body.url || "").trim();
+  const method = String(body.method || "POST").toUpperCase();
+  const headers = body.headers && typeof body.headers === "object" ? body.headers : {};
+  const rawBody = body.body ?? body.payload ?? null;
+
+  if (!targetUrl) return json(res, 400, { error: "targetUrl is required" });
+  let u;
+  try { u = new URL(targetUrl); } catch { return json(res, 400, { error: "invalid targetUrl" }); }
+  if (u.protocol !== "https:" && u.protocol !== "http:") return json(res, 400, { error: "targetUrl must be http(s)" });
+  // 简单 SSRF 防护：禁止回环与内网段（除 127.0.0.1:8989 本身已鉴权，仍放行）
+  // 允许外网 https 上游（如 https://api.bai.com、https://opencode.ai），禁止 10/172.16/192.168
+  const host = u.hostname;
+  if (/^(10\.|192\.168\.|172\.(1[6-9]|2\d|3[0-1])\.)/.test(host) && host !== "127.0.0.1" && host !== "localhost") {
+    // 仍允许，因为组员可能是内网 IP 的上游（如自建网关），仅告警不拦
+  }
+
+  // 透传头：只放行受控头，避免把内部头（如 x-mslxdff-*）误带给上游
+  const allowHeaders = new Set(["authorization", "content-type", "accept", "user-agent", "x-client-type", "x-platform", "x-task-id", "x-user-id", "x-domain", "x-enterprise-id", "x-tenant-id", "origin", "referer"]);
+  const fwdHeaders = {};
+  for (const [k, v] of Object.entries(headers)) {
+    const lk = String(k).toLowerCase();
+    if (allowHeaders.has(lk) && typeof v === "string" && v) fwdHeaders[k] = v;
+  }
+  // 强制 JSON
+  if (!fwdHeaders["Content-Type"] && !fwdHeaders["content-type"]) fwdHeaders["Content-Type"] = "application/json";
+  if (!fwdHeaders["Accept"] && !fwdHeaders["accept"]) fwdHeaders["Accept"] = "application/json";
+
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(new Error("relay timeout 30000ms")), 30000);
+  try {
+    const fetchBody = rawBody == null ? undefined : (typeof rawBody === "string" ? rawBody : JSON.stringify(rawBody));
+    const r = await fetch(targetUrl, { method, headers: fwdHeaders, body: fetchBody, signal: controller.signal });
+    const txt = await r.text();
+    // 原样回透：状态码 + 头（仅透 content-type） + body
+    res.statusCode = r.status;
+    const ct = r.headers.get("content-type");
+    if (ct) res.setHeader("content-type", ct);
+    // 额外回透上游错误码便于 bench 区分
+    res.setHeader("x-mslxdff-relay-status", String(r.status));
+    return res.end(txt);
+  } catch (e) {
+    const msg = String(e?.message || e);
+    const isTimeout = /timeout|abort/i.test(msg);
+    return json(res, 502, { error: isTimeout ? "relay timeout" : `relay fetch failed: ${msg.slice(0, 300)}` });
+  } finally { clearTimeout(timer); }
+}

package/src/upstream.js

@@ -59,13 +59,30 @@ export function createUpstreamClient({
       return null;
     }
   };
-  const baseHeaders = {
-    "Content-Type": "application/json",
-    "Authorization": `Bearer ${authToken}`,
-    "x-opencode-client": "desktop",
-  };
+  function isAnonFirst() {
+    const raw = process.env.MSLXDFF_OPENCOD_ANON_FIRST ?? process.env.MSLXDFF_ANON_FIRST ?? process.env.MSLXDFF_OPENCOD_ANON;
+    if (raw === undefined || raw === null || raw === "") return authToken === "public";
+    const s = String(raw).trim().toLowerCase();
+    if (s === "0" || s === "false" || s === "off" || s === "no" || s === "disable" || s === "disabled") return false;
+    return authToken === "public";
+  }
+  const anonFirst = isAnonFirst();
+  const baseHeaders = anonFirst
+    ? {
+        "Content-Type": "application/json",
+        "Authorization": "",
+        "x-opencode-client": "desktop",
+        "User-Agent": "opencode",
+        "HTTP-Referer": "https://hermes-agent.nousresearch.com",
+        "X-Title": "Hermes Agent",
+      }
+    : {
+        "Content-Type": "application/json",
+        "Authorization": `Bearer ${authToken}`,
+        "x-opencode-client": "desktop",
+      };
 
-  function buildHeaders(body, { anonymous = false } = {}) {
+  function buildHeaders(body, { anonymous = anonFirst } = {}) {
     const isStream = body?.stream !== false;
     const base = {
       ...baseHeaders,
@@ -84,6 +101,11 @@ export function createUpstreamClient({
         "X-Title": "Hermes Agent",
       };
     }
+    // 显式 public 回退（anonFirst 时去掉 hermes 头）
+    if (anonFirst) {
+      const { "HTTP-Referer": _a, "X-Title": _b, ...rest } = base;
+      return { ...rest, "Authorization": `Bearer ${authToken}` };
+    }
     return base;
   }
 
@@ -238,8 +260,8 @@ export function createUpstreamClient({
         waitMs += entry.delayMs;
         continue;
       }
-      // 额外额度探测：public 429 且为 free 模型时，用空头（hermes 方式）每秒重试 3 次
-      if (result.status === 429 && isFreeModel(body?.model) && shouldTryAnonFree()) {
+      // 额外额度探测：仅当非 anon 优先时，public 429 且为 free 模型时，用空头（hermes 方式）每秒重试 3 次；anon 优先时首发即高额度，无需再 public 撞墙
+      if (!anonFirst && result.status === 429 && isFreeModel(body?.model) && shouldTryAnonFree()) {
         const anonRetries = envInt("MSLXDFF_FREE_ANON_RETRIES", 3);
         const anonDelay = envInt("MSLXDFF_FREE_ANON_DELAY_MS", 1000);
         let anonResult = null;
@@ -349,7 +371,7 @@ export function createUpstreamClient({
     }
   }
 
-  async function attemptOnce(url, body, { anonymous = false } = {}) {
+  async function attemptOnce(url, body, { anonymous = anonFirst } = {}) {
     const controller = new AbortController();
     const timer = setTimeout(() =>
       controller.abort(new Error(`upstream timed out after ${connectTimeoutMs}ms`)),