mslxdff 0.1.229 → 0.1.230

package/src/providers/agentrouter.js

@@ -1,225 +1,231 @@
-// AgentRouter 定制供应商：OpenAI 兼容聚合网关，有三道硬门槛（2026-10-21 现网 curl 实测）：
-//   ① 客户端身份白名单 —— UA 为 `mslxdff`/`curl`/浏览器一律 401 `unauthorized client detected`，
-//      只放行 `opencode/*` 与 `claude-cli/<ver> (external…)`；
-//   ② 模型分端点 —— `deepseek-v4-flash` 吃 /v1/chat/completions，`gpt-6.1-sol`/`gpt-6-astra`
-//      对该端点回 400「请改用 /v1/responses」，只在 responses 通道可用；
-//   ③ 内容闸 —— 正文含 `x-anthropic` 字样稳定 500 `sensitive words detected`（连 user 消息也拦，
-//      与 workbuddy 只扫 system/assistant 不同），见 stripAgentRouterSensitive。
-// 故不复用 generic（其 UA 写死 `mslxdff`、responses 判定只认 models.dev npm 索引），这里自组装，
-// 底层仍复用 base.js 的 runner 与 responses-channel 的协议互转。
-import { createKeyRing } from "./keyring.js";
-import { loadProviderBaseUrl, loadProviderKeys } from "../state.js";
-import { createChatRunner, createAgent, envInt, getUndici, joinUrl, collectApiKeysGeneric } from "./base.js";
-import { createResponsesChannel, resolveResponsesPath } from "./responses-channel.js";
-import { compatFetch } from "../compat.js";
-import { opencodeUa } from "../opencode-identity.js";
-import { joinModelId } from "./model-id.js";
-
-const { UndiciFetch } = getUndici();
-
-export const DEFAULT_AGENTROUTER_BASE = "https://agentrouter.org";
-export const MODELS_PATH = "/v1/models";
-export const CHAT_PATH = "/v1/chat/completions";
-export const RESPONSES_PATH = "/v1/responses";
-
-// 只吃 responses 的模型族：现网实测 gpt-6* 打 chat 端点回 400 并显式要求改 /v1/responses。
-const RESPONSES_PREFIXES = ["gpt-6"];
-
-export function isAgentRouterResponsesModel(model) {
-  const m = String(model || "").trim().toLowerCase();
-  if (!m) return false;
-  return RESPONSES_PREFIXES.some((p) => m.startsWith(p));
-}
-
-// baseUrl 归一：缺省取 state/env，剥掉尾部 `/v1`（路径常量自带 `/v1`，避免 /v1/v1）
-export function resolveAgentRouterBase(baseUrl) {
-  const raw = baseUrl || loadProviderBaseUrl("agentrouter") || DEFAULT_AGENTROUTER_BASE;
-  return String(raw).trim().replace(/\/+$/, "").replace(/\/v1$/i, "");
-}
-
-// responses 通道内部两处 stream 判据不一致（响应通道用 `!== false` 决定读法，body 转换用
-// `=== true` 决定发什么）：客户端不显式给 stream 时，会去把未转换的上游 JSON 当 chat 形状读。
-// 这里先把意图钉成布尔，两条路径就同源了。
-export function normalizeResponsesBody(body) {
-  if (!body || typeof body !== "object") return body;
-  return { ...body, stream: body.stream === true };
-}
-
-// AgentRouter 内容闸：稳定拦截正文里含 `x-anthropic` 字样的请求（2026-10-21 现网实测：
-// `500 sensitive words detected` / `400 content-blocked`，连 `x-anthropic` 一个词也拦；
-// `x-google`/`x-meta`/`x-openai`/`anthropic` 均放行——这是对 Anthropic 内部计费头/代理指纹的反代理风控）。
-// 与 workbuddy 的 CC_SENTENCES 剥离同类，但**必须覆盖所有 role**（workbuddy 只扫 system/assistant，
-// 而 agentrouter 连 user/tool 消息里出现该串也拦）。用零宽空格断开子串匹配、不动其余内容。
-// 返回新对象（messages 浅拷贝），不污染调用方 body（dispatcher 可能复用同一 body 做兜底重试）。
-const AR_SENSITIVE = /x-anthropic/gi;
-
-export function stripAgentRouterSensitive(obj) {
-  if (!obj || typeof obj !== "object") return obj;
-  if (!Array.isArray(obj.messages)) return obj;
-  const mask = (text) => (typeof text === "string" && AR_SENSITIVE.test(text) ? text.replace(AR_SENSITIVE, "x​anthropic") : text);
-  const out = { ...obj, messages: obj.messages.map((m) => {
-    if (!m || typeof m !== "object") return m;
-    const nm = { ...m };
-    if (typeof m.content === "string") nm.content = mask(m.content);
-    else if (Array.isArray(m.content)) nm.content = m.content.map((p) => (p && typeof p === "object" && typeof p.text === "string" ? { ...p, text: mask(p.text) } : p));
-    if (typeof m.name === "string") nm.name = mask(m.name);
-    return nm;
-  }) };
-  return out;
-}
-
-export function createAgentRouterProvider({
-  id = "agentrouter",
-  baseUrl,
-  apiKeys,
-  apiKey,
-  connectTimeoutMs = envInt("MSLXDFF_AGENTROUTER_TIMEOUT_MS", 60_000),
-  cooldownMs = envInt("MSLXDFF_AGENTROUTER_COOLDOWN_MS", 30_000),
-  fetchImpl,
-  file,
-} = {}) {
-  if (id !== "agentrouter") throw new Error(`agentrouter provider id mismatch: ${id}`);
-  const resolvedBase = resolveAgentRouterBase(baseUrl);
-  if (!fetchImpl) fetchImpl = UndiciFetch || compatFetch;
-
-  const ring = createKeyRing(
-    collectApiKeysGeneric(id, apiKeys, apiKey, (pid) => loadProviderKeys(pid, file ? { file } : {})),
-    { cooldownMs },
-  );
-
-  let agent = null; let dispatcher = null;
-  {
-    const a = createAgent({
-      keepAliveTimeout: envInt("MSLXDFF_AGENTROUTER_KEEPALIVE_TIMEOUT", 30_000),
-      keepAliveMaxTimeout: envInt("MSLXDFF_AGENTROUTER_KEEPALIVE_MAX_TIMEOUT", 60_000),
-      connections: envInt("MSLXDFF_AGENTROUTER_KEEPALIVE_CONNECTIONS", 20),
-    });
-    agent = a.agent; dispatcher = a.dispatcher;
-  }
-
-  // 身份 UA：白名单只认 opencode / claude-cli，缺了恒 401。单独 env 可覆盖，不联动全局。
-  function identityUa() {
-    return process.env.MSLXDFF_AGENTROUTER_UA || opencodeUa();
-  }
-
-  function buildHeaders(body, key) {
-    const isStream = body?.stream !== false;
-    const h = {
-      "Content-Type": "application/json",
-      Accept: isStream ? "text/event-stream" : "*/*",
-      "User-Agent": identityUa(),
-    };
-    if (key) h.Authorization = `Bearer ${key}`;
-    return h;
-  }
-
-  function jsonHeaders(key) {
-    const h = { Accept: "application/json", "User-Agent": identityUa() };
-    if (key) h.Authorization = `Bearer ${key}`;
-    return h;
-  }
-
-  const { runChat } = createChatRunner({
-    id, ring, cooldownMs, fetchImpl, dispatcher, buildHeaders,
-    getUrl: () => joinUrl(resolvedBase, CHAT_PATH),
-    connectTimeoutMs,
-    // 非流式走原生（SDK 通道只服务 opencode zen 的形状），responses 另走专用通道
-    sdkDispatch: { enabled: false },
-  });
-
-  const responses = createResponsesChannel({
-    id,
-    url: joinUrl(resolvedBase, resolveResponsesPath(id, RESPONSES_PATH)),
-    connectTimeoutMs,
-    cooldownMs,
-    // 关 SDK：其缺省会强塞 include:["reasoning.encrypted_content"] / store:false，agentrouter 未验证支持
-    sdkEnabled: false,
-    buildHeaders,
-    fetchImpl,
-    dispatcher,
-  });
-
-  async function chat(body, opts) {
-    const sourceKey = `MSLXDFF_AGENTROUTER_KEY`;
-    const clean = stripAgentRouterSensitive(body);
-    if (isAgentRouterResponsesModel(clean?.model)) return responses.run(normalizeResponsesBody(clean), ring, sourceKey, opts);
-    return runChat(clean, ring, sourceKey);
-  }
-
-  async function chatWithKeys(body, keys, opts) {
-    const tmp = createKeyRing(keys, { cooldownMs });
-    const clean = stripAgentRouterSensitive(body);
-    if (isAgentRouterResponsesModel(clean?.model)) return responses.run(normalizeResponsesBody(clean), tmp, "shared provider keys", opts);
-    const scoped = createChatRunner({
-      id, ring: tmp, cooldownMs, fetchImpl, dispatcher, buildHeaders,
-      getUrl: () => joinUrl(resolvedBase, CHAT_PATH),
-      connectTimeoutMs,
-      sdkDispatch: { enabled: false },
-    });
-    return scoped.runChat(clean, tmp, "shared provider keys");
-  }
-
-  let modelsCache = { at: 0, list: [] };
-  const MODELS_TTL_MS = 10 * 60 * 1000;
-
-  async function fetchModels() {
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(new Error(`${id} models timed out`)), 15_000);
-    try {
-      const key = ring.next() || ring.keys[0];
-      const opts = { headers: jsonHeaders(key), signal: controller.signal };
-      if (dispatcher) opts.dispatcher = dispatcher;
-      const res = await fetchImpl(joinUrl(resolvedBase, MODELS_PATH), opts);
-      if (!res.ok) throw new Error(`${id} models: http ${res.status}`);
-      const json = await res.json().catch(() => ({}));
-      return Array.isArray(json?.data) ? json.data : Array.isArray(json) ? json : [];
-    } finally {
-      clearTimeout(timer);
-    }
-  }
-
-  async function listModels() {
-    if (modelsCache.list.length && Date.now() - modelsCache.at < MODELS_TTL_MS) return modelsCache.list;
-    let raw = [];
-    try {
-      raw = await fetchModels();
-    } catch {
-      return modelsCache.list; // 拉取失败给上次缓存（首次为空），不阻断启动
-    }
-    const list = raw
-      .filter((m) => m && typeof m.id === "string" && m.id.trim())
-      .map((m) => ({
-        id: joinModelId(id, m.id),
-        object: "model",
-        name: m.name || m.id,
-        owned_by: m.owned_by || "agentrouter",
-        ...(m.created ? { created: m.created } : {}),
-      }));
-    modelsCache = { at: Date.now(), list };
-    return list;
-  }
-
-  async function preheat() {
-    try {
-      const m = await listModels();
-      return { ok: m.length > 0, models: m.length };
-    } catch (err) {
-      return { ok: false, error: String(err?.message || err).slice(0, 120) };
-    }
-  }
-
-  async function close() { if (agent?.close) try { await agent.close(); } catch {} }
-
-  return {
-    id,
-    chat,
-    chatWithKeys,
-    listModels,
-    preheat,
-    close,
-    agent,
-    keyRing: ring,
-    baseUrl: resolvedBase,
-    responsesUrl: responses.url,
-  };
-}
+// AgentRouter 定制供应商：OpenAI 兼容聚合网关，有三道硬门槛（2026-10-21 现网 curl 实测）：
+//   ① 客户端身份白名单 —— UA 为 `mslxdff`/`curl`/浏览器一律 401 `unauthorized client detected`，
+//      只放行 `opencode/*` 与 `claude-cli/<ver> (external…)`；
+//   ② 模型分端点 —— `deepseek-v4-flash` 吃 /v1/chat/completions，`gpt-6.1-sol`/`gpt-6-astra`
+//      对该端点回 400「请改用 /v1/responses」，只在 responses 通道可用；
+//   ③ 内容闸 —— 正文含 `x-anthropic` 字样稳定 500 `sensitive words detected`（连 user 消息也拦，
+//      与 workbuddy 只扫 system/assistant 不同），见 stripAgentRouterSensitive。
+// 故不复用 generic（其 UA 写死 `mslxdff`、responses 判定只认 models.dev npm 索引），这里自组装，
+// 底层仍复用 base.js 的 runner 与 responses-channel 的协议互转。
+import { createKeyRing } from "./keyring.js";
+import { loadProviderBaseUrl, loadProviderKeys } from "../state.js";
+import { createChatRunner, createAgent, envInt, getUndici, joinUrl, collectApiKeysGeneric } from "./base.js";
+import { createResponsesChannel, resolveResponsesPath } from "./responses-channel.js";
+import { compatFetch } from "../compat.js";
+import { opencodeUa } from "../opencode-identity.js";
+import { joinModelId } from "./model-id.js";
+
+const { UndiciFetch } = getUndici();
+
+export const DEFAULT_AGENTROUTER_BASE = "https://agentrouter.org";
+export const MODELS_PATH = "/v1/models";
+export const CHAT_PATH = "/v1/chat/completions";
+export const RESPONSES_PATH = "/v1/responses";
+
+// 只吃 responses 的模型族：现网实测 gpt-6* 打 chat 端点回 400 并显式要求改 /v1/responses。
+const RESPONSES_PREFIXES = ["gpt-6"];
+
+export function isAgentRouterResponsesModel(model) {
+  const m = String(model || "").trim().toLowerCase();
+  if (!m) return false;
+  return RESPONSES_PREFIXES.some((p) => m.startsWith(p));
+}
+
+// baseUrl 归一：缺省取 state/env，剥掉尾部 `/v1`（路径常量自带 `/v1`，避免 /v1/v1）
+export function resolveAgentRouterBase(baseUrl) {
+  const raw = baseUrl || loadProviderBaseUrl("agentrouter") || DEFAULT_AGENTROUTER_BASE;
+  return String(raw).trim().replace(/\/+$/, "").replace(/\/v1$/i, "");
+}
+
+// responses 通道内部两处 stream 判据不一致（响应通道用 `!== false` 决定读法，body 转换用
+// `=== true` 决定发什么）：客户端不显式给 stream 时，会去把未转换的上游 JSON 当 chat 形状读。
+// 这里先把意图钉成布尔，两条路径就同源了。
+export function normalizeResponsesBody(body) {
+  if (!body || typeof body !== "object") return body;
+  return { ...body, stream: body.stream === true };
+}
+
+// AgentRouter 内容闸：稳定拦截正文里含 `x-anthropic` 字样的请求（2026-10-21 现网实测：
+// `500 sensitive words detected` / `400 content-blocked`，连 `x-anthropic` 一个词也拦；
+// `x-google`/`x-meta`/`x-openai`/`anthropic` 均放行——这是对 Anthropic 内部计费头/代理指纹的反代理风控）。
+// 与 workbuddy 的 CC_SENTENCES 剥离同类，但**必须覆盖所有 role**（workbuddy 只扫 system/assistant，
+// 而 agentrouter 连 user/tool 消息里出现该串也拦）。用零宽空格断开子串匹配、不动其余内容。
+// 返回新对象（messages 浅拷贝），不污染调用方 body（dispatcher 可能复用同一 body 做兜底重试）。
+const AR_SENSITIVE = /x-anthropic/gi;
+
+export function stripAgentRouterSensitive(obj) {
+  if (!obj || typeof obj !== "object") return obj;
+  if (!Array.isArray(obj.messages)) return obj;
+  const mask = (text) => (typeof text === "string" && AR_SENSITIVE.test(text) ? text.replace(AR_SENSITIVE, "x​anthropic") : text);
+  const out = { ...obj, messages: obj.messages.map((m) => {
+    if (!m || typeof m !== "object") return m;
+    const nm = { ...m };
+    if (typeof m.content === "string") nm.content = mask(m.content);
+    else if (Array.isArray(m.content)) nm.content = m.content.map((p) => (p && typeof p === "object" && typeof p.text === "string" ? { ...p, text: mask(p.text) } : p));
+    if (typeof m.name === "string") nm.name = mask(m.name);
+    return nm;
+  }) };
+  return out;
+}
+
+export function createAgentRouterProvider({
+  id = "agentrouter",
+  baseUrl,
+  apiKeys,
+  apiKey,
+  connectTimeoutMs = envInt("MSLXDFF_AGENTROUTER_TIMEOUT_MS", 60_000),
+  cooldownMs = envInt("MSLXDFF_AGENTROUTER_COOLDOWN_MS", 30_000),
+  fetchImpl,
+  file,
+} = {}) {
+  if (id !== "agentrouter") throw new Error(`agentrouter provider id mismatch: ${id}`);
+  const resolvedBase = resolveAgentRouterBase(baseUrl);
+  if (!fetchImpl) fetchImpl = UndiciFetch || compatFetch;
+
+  const ring = createKeyRing(
+    collectApiKeysGeneric(id, apiKeys, apiKey, (pid) => loadProviderKeys(pid, file ? { file } : {})),
+    { cooldownMs },
+  );
+
+  let agent = null; let dispatcher = null;
+  {
+    const a = createAgent({
+      keepAliveTimeout: envInt("MSLXDFF_AGENTROUTER_KEEPALIVE_TIMEOUT", 30_000),
+      keepAliveMaxTimeout: envInt("MSLXDFF_AGENTROUTER_KEEPALIVE_MAX_TIMEOUT", 60_000),
+      connections: envInt("MSLXDFF_AGENTROUTER_KEEPALIVE_CONNECTIONS", 20),
+    });
+    agent = a.agent; dispatcher = a.dispatcher;
+  }
+
+  // 身份 UA：白名单只认 opencode / claude-cli，缺了恒 401。单独 env 可覆盖，不联动全局。
+  function identityUa() {
+    return process.env.MSLXDFF_AGENTROUTER_UA || opencodeUa();
+  }
+
+  function buildHeaders(body, key) {
+    const isStream = body?.stream !== false;
+    const h = {
+      "Content-Type": "application/json",
+      Accept: isStream ? "text/event-stream" : "*/*",
+      "User-Agent": identityUa(),
+    };
+    if (key) h.Authorization = `Bearer ${key}`;
+    return h;
+  }
+
+  function jsonHeaders(key) {
+    const h = { Accept: "application/json", "User-Agent": identityUa() };
+    if (key) h.Authorization = `Bearer ${key}`;
+    return h;
+  }
+
+  // 网络重试：agentrouter 只有 1 把 key，任何一次瞬时连接抖动（现网实测 ECONNRESET 偶发，
+  // 上游高频时会触发）若直接 onError，= 立刻冷却 30s、把后续请求全挡成
+  // "all keys in cooldown"。补 attempts=2（对齐 generic/cline/workbuddy），让网络层重试先扛一下。
+  const chatRetry = { network: { attempts: 2, delayMs: 300 }, 429: { attempts: 1, delayMs: 100 }, 502: { attempts: 1, delayMs: 100 }, 503: { attempts: 1, delayMs: 100 }, 504: { attempts: 1, delayMs: 100 } };
+
+  const { runChat } = createChatRunner({
+    id, ring, cooldownMs, retry: chatRetry, fetchImpl, dispatcher, buildHeaders,
+    getUrl: () => joinUrl(resolvedBase, CHAT_PATH),
+    connectTimeoutMs,
+    // 非流式走原生（SDK 通道只服务 opencode zen 的形状），responses 另走专用通道
+    sdkDispatch: { enabled: false },
+  });
+
+  const responses = createResponsesChannel({
+    id,
+    url: joinUrl(resolvedBase, resolveResponsesPath(id, RESPONSES_PATH)),
+    connectTimeoutMs,
+    cooldownMs,
+    retry: chatRetry,
+    // 关 SDK：其缺省会强塞 include:["reasoning.encrypted_content"] / store:false，agentrouter 未验证支持
+    sdkEnabled: false,
+    buildHeaders,
+    fetchImpl,
+    dispatcher,
+  });
+
+  async function chat(body, opts) {
+    const sourceKey = `MSLXDFF_AGENTROUTER_KEY`;
+    const clean = stripAgentRouterSensitive(body);
+    if (isAgentRouterResponsesModel(clean?.model)) return responses.run(normalizeResponsesBody(clean), ring, sourceKey, opts);
+    return runChat(clean, ring, sourceKey);
+  }
+
+  async function chatWithKeys(body, keys, opts) {
+    const tmp = createKeyRing(keys, { cooldownMs });
+    const clean = stripAgentRouterSensitive(body);
+    if (isAgentRouterResponsesModel(clean?.model)) return responses.run(normalizeResponsesBody(clean), tmp, "shared provider keys", opts);
+    const scoped = createChatRunner({
+      id, ring: tmp, cooldownMs, retry: chatRetry, fetchImpl, dispatcher, buildHeaders,
+      getUrl: () => joinUrl(resolvedBase, CHAT_PATH),
+      connectTimeoutMs,
+      sdkDispatch: { enabled: false },
+    });
+    return scoped.runChat(clean, tmp, "shared provider keys");
+  }
+
+  let modelsCache = { at: 0, list: [] };
+  const MODELS_TTL_MS = 10 * 60 * 1000;
+
+  async function fetchModels() {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(new Error(`${id} models timed out`)), 15_000);
+    try {
+      const key = ring.next() || ring.keys[0];
+      const opts = { headers: jsonHeaders(key), signal: controller.signal };
+      if (dispatcher) opts.dispatcher = dispatcher;
+      const res = await fetchImpl(joinUrl(resolvedBase, MODELS_PATH), opts);
+      if (!res.ok) throw new Error(`${id} models: http ${res.status}`);
+      const json = await res.json().catch(() => ({}));
+      return Array.isArray(json?.data) ? json.data : Array.isArray(json) ? json : [];
+    } finally {
+      clearTimeout(timer);
+    }
+  }
+
+  async function listModels() {
+    if (modelsCache.list.length && Date.now() - modelsCache.at < MODELS_TTL_MS) return modelsCache.list;
+    let raw = [];
+    try {
+      raw = await fetchModels();
+    } catch {
+      return modelsCache.list; // 拉取失败给上次缓存（首次为空），不阻断启动
+    }
+    const list = raw
+      .filter((m) => m && typeof m.id === "string" && m.id.trim())
+      .map((m) => ({
+        id: joinModelId(id, m.id),
+        object: "model",
+        name: m.name || m.id,
+        owned_by: m.owned_by || "agentrouter",
+        ...(m.created ? { created: m.created } : {}),
+      }));
+    modelsCache = { at: Date.now(), list };
+    return list;
+  }
+
+  async function preheat() {
+    try {
+      const m = await listModels();
+      return { ok: m.length > 0, models: m.length };
+    } catch (err) {
+      return { ok: false, error: String(err?.message || err).slice(0, 120) };
+    }
+  }
+
+  async function close() { if (agent?.close) try { await agent.close(); } catch {} }
+
+  return {
+    id,
+    chat,
+    chatWithKeys,
+    listModels,
+    preheat,
+    close,
+    agent,
+    keyRing: ring,
+    baseUrl: resolvedBase,
+    responsesUrl: responses.url,
+  };
+}