mslxdff 0.1.236 → 0.1.238

package/src/providers/codearts/chat.js

@@ -1,207 +1,207 @@
-// CodeArts 聊天服务：上游 body/头组装 + SDK-HMAC-SHA256 签名 + 401 刷新重试 + 环形切号。
-// 上游恒 stream:true（chatBodyV2 契约）；流式客户端转 OpenAI SSE，非流式聚合回 JSON。
-import crypto from "node:crypto";
-import { joinUrl } from "../base.js";
-import { createTransport } from "../../transport/index.js";
-import { SNAP_BASE, EP_CHAT_V2, chatBaseHeaders, MAAS_TYPE_HEADER, MAAS_BENEFIT, MAX_TOKENS_CEILING } from "./const.js";
-import { signRequest } from "./sign.js";
-import { newChatId, preflightResponse, sseToOpenAIResponse, aggregateToCompletion, UpstreamEventError } from "./stream.js";
-
-const MODEL_RULES = [
-  { match: (m) => String(m || "").toLowerCase().includes("deepseek"), scope: "all" },
-  { match: (m) => /(^kimi-|kimi)/i.test(String(m || "")), scope: "toolCalls" },
-];
-
-// DeepSeek 系思考模式要求 assistant 消息回传 reasoning_content，缺失补 " " 占位（ADR-0001 同款）。
-export function injectReasoningPlaceholder(messages, model) {
-  const rule = MODEL_RULES.find((r) => r.match(model));
-  if (!rule || !Array.isArray(messages)) return messages;
-  return messages.map((m) => {
-    if (m?.role !== "assistant") return m;
-    if (typeof m.reasoning_content === "string" && m.reasoning_content.length) return m;
-    if (rule.scope === "toolCalls" && !(Array.isArray(m.tool_calls) && m.tool_calls.length)) return m;
-    return { ...m, reasoning_content: " " };
-  });
-}
-
-const PASSTHROUGH_KEYS = ["temperature", "top_p", "max_tokens", "max_completion_tokens", "reasoning_effort", "tools", "tool_choice", "stop", "presence_penalty", "frequency_penalty", "response_format", "user", "n", "seed"];
-
-// 账号级「会话/并发饱和」：上游按账号限同时活跃会话数（TM.00001041 = 并发会话数已达上限(3个)）。
-// 这是账号此刻进不去，不是本请求参数错；槽位在上游保留分钟级才回收（与本地进程无关）。
-// 多号场景换另一个账号通常立即成功——故短冷却本账号让下一次 next() 换号并 continue 重试。
-const SESSION_SATURATED_RE = /TM\.00001041|并发会话数已达上限/;
-const SESSION_SATURATED_COOLDOWN_MS = Number.parseInt(process.env.MSLXDFF_CODEARTS_SESSION_COOLDOWN_MS ?? "", 10) || 30_000;
-
-export function buildUpstreamBody(body, { chatId, upstreamModel }) {
-  const out = {
-    model: upstreamModel,
-    stream: true, // v2 端点契约：恒流式，非流式由本层聚合
-    messages: injectReasoningPlaceholder(body?.messages || [], upstreamModel),
-    prompt_cache_key: chatId,
-    tool_stream: true,
-    chat_id: chatId,
-  };
-  for (const k of PASSTHROUGH_KEYS) {
-    if (k === "max_completion_tokens") continue; // v2 端点不认这个名字（透传 → InferHub.001001005.400，2026-10-16 现网 A/B 复现）
-    if (body?.[k] !== undefined && body?.[k] !== null) out[k] = body[k];
-  }
-  // Responses 风格客户端只发 max_completion_tokens：映射成 v2 认的 max_tokens（显式 max_tokens 优先）
-  if (out.max_tokens === undefined && body?.max_completion_tokens != null) out.max_tokens = body.max_completion_tokens;
-  // 超上限的 max_tokens 上游一律回 InferHub.001001005.400（参数非法），在这里夹住（env 可放宽）
-  const ceiling = Number.parseInt(process.env.MSLXDFF_CODEARTS_MAX_TOKENS_CEILING ?? "", 10) || MAX_TOKENS_CEILING;
-  if (typeof out.max_tokens === "number" && out.max_tokens > ceiling) out.max_tokens = ceiling;
-  return out;
-}
-
-// Session-Id：由 chatId 派生的稳定会话标识（prompt cache 亲和，对齐 Go sessionIDFor）。
-function sessionIdFor(chatId) {
-  return crypto.createHash("sha256").update(`codearts-session:${chatId}`, "utf8").digest("hex").slice(0, 32);
-}
-
-function stripOwnPrefix(model, providerId) {
-  const s = String(model || "");
-  const i = s.indexOf("/");
-  if (i <= 0) return s;
-  return s.slice(0, i).toLowerCase() === String(providerId || "").toLowerCase() ? s.slice(i + 1) : s;
-}
-
-// 从 4xx/5xx 正文里判「账号级会话饱和」：命中则可换号重试（不同于参数类 400，换号无意义）。
-export function isSessionSaturated(status, bodyText) {
-  if (status !== 400 && status !== 429 && status < 500) return false;
-  return SESSION_SATURATED_RE.test(String(bodyText || ""));
-}
-
-export function createChatService({
-  id = "codearts",
-  baseUrl = SNAP_BASE,
-  chatPath = EP_CHAT_V2,
-  fetchImpl,
-  dispatcher,
-  authPool,
-  catalog,
-  connectTimeoutMs = 30_000,
-  clock = Date.now,
-  } = {}) {
-  const resolvedBase = String(baseUrl).trim().replace(/\/+$/, "") || SNAP_BASE;
-  const transport = createTransport({ fetchImpl, dispatcher, keepAlive: !!dispatcher, timeoutMs: connectTimeoutMs, retry: {} });
-
-  async function sendOnce({ bodyStr, account, benefit, chatId, sessionId, traceId }) {
-    const url = joinUrl(resolvedBase, chatPath);
-    const h = chatBaseHeaders(account.securityToken, traceId);
-    if (benefit) h[MAAS_TYPE_HEADER] = MAAS_BENEFIT;
-    const { headers } = signRequest({ method: "POST", url, headers: h, body: bodyStr, cred: account });
-    // Chat-Id / Session-Id 在签名之后追加（对齐官方客户端：不进 SignedHeaders）。
-    headers["chat-id"] = chatId;
-    headers["session-id"] = sessionId;
-    return transport.request({ url, method: "POST", headers, body: bodyStr, stream: true, timeoutMs: connectTimeoutMs });
-  }
-
-  function errRes(status, msg, code) {
-    return new Response(JSON.stringify({ error: { message: msg, type: "upstream_error", ...(code ? { code } : {}) } }), {
-      status,
-      headers: { "Content-Type": "application/json" },
-    });
-  }
-
-  function withTiming(res, t0) {
-    try { res._t = { attempts: [], waitMs: 0, totalMs: Math.round(clock() - t0) }; } catch {}
-    return res;
-  }
-
-  /**
-   * 跑一次对话。body.model 已由 dispatcher 剥前缀（容忍自带前缀）。
-   * @returns {Promise<Response>} 流式 → OpenAI SSE；非流式 → chat.completion JSON
-   */
-  async function runChat(body, activeRing, opts = {}) {
-    const t0 = clock();
-    const requestedModel = String(body?.model || "");
-    const rawModel = stripOwnPrefix(requestedModel, id);
-    const upstreamModel = catalog?.canonical(rawModel) || rawModel;
-    const chatId = newChatId(body, opts);
-    const sessionId = sessionIdFor(chatId);
-    const upstreamBody = buildUpstreamBody(body, { chatId, upstreamModel });
-    const bodyStr = JSON.stringify(upstreamBody);
-    const isStream = body?.stream === true;
-    let lastErr = null;
-
-    const maxTries = Math.max(1, activeRing.size || 1);
-    for (let attempt = 0; attempt < maxTries; attempt++) {
-      const blob = activeRing.next();
-      if (!blob) break;
-      let account;
-      try {
-        account = await authPool.getCredential(blob);
-      } catch (err) {
-        try { activeRing.onError(blob); } catch {}
-        lastErr = err;
-        continue; // 死号/不在池：换下一个账号
-      }
-      const benefit = catalog?.isBenefit ? catalog.isBenefit(account.userId, upstreamModel) : false;
-      const traceId = crypto.randomBytes(8).toString("hex") + String(clock());
-      let res;
-      try {
-        res = await sendOnce({ bodyStr, account, benefit, chatId, sessionId, traceId });
-      } catch (err) {
-        try { activeRing.onError(blob); } catch {}
-        lastErr = err;
-        continue;
-      }
-      if (res.status === 401 || res.status === 403) {
-        // STS 临时凭证失效：强制刷新重试一次，仍 401/403 换号
-        authPool.invalidate(blob);
-        let account2 = null;
-        try { account2 = await authPool.getCredential(blob); } catch { /* 死号 → 换号 */ }
-        if (account2) {
-          const traceId2 = crypto.randomBytes(8).toString("hex") + String(clock());
-          try { res = await sendOnce({ bodyStr, account: account2, benefit, chatId, sessionId, traceId: traceId2 }); } catch (err2) { lastErr = err2; try { activeRing.onError(blob); } catch {} continue; }
-          if (res.status === 401 || res.status === 403) {
-            try { await res.text(); } catch {}
-            try { activeRing.onError(blob); } catch {}
-            lastErr = new Error(`codearts auth still failing after refresh (http ${res.status})`);
-            continue;
-          }
-        } else {
-          try { activeRing.onError(blob); } catch {}
-          continue;
-        }
-      }
-      if (res.status >= 400) {
-        const txt = await res.text().catch(() => "");
-        const msg = `codearts upstream http ${res.status}: ${String(txt).slice(0, 300)}`;
-        // 账号会话饱和：短冷却本号并换号重试（多号场景另一号通常立即成功）；
-        // 所有号都饱和才落到重试耗尽分支，交给上层按上游不可用输出。
-        if (isSessionSaturated(res.status, txt)) {
-          try { activeRing.onError(blob, SESSION_SATURATED_COOLDOWN_MS); } catch {}
-          lastErr = Object.assign(new Error(msg), { upstreamStatus: res.status, upstreamBody: txt });
-          continue;
-        }
-        if (res.status === 429 || res.status >= 500) try { activeRing.onError(blob); } catch {}
-        return withTiming(errRes(res.status, msg), t0);
-      }
-      try {
-        const pre = await preflightResponse(res); // HTTP 200 内嵌错误（排队/未注册/福利未领）→ 抛 UpstreamEventError
-        if (isStream) return withTiming(sseToOpenAIResponse(pre, { model: requestedModel || upstreamModel, id: `chatcmpl-${chatId.slice(0, 12)}` }), t0);
-        const data = await aggregateToCompletion(pre, { model: requestedModel || upstreamModel, id: `chatcmpl-${chatId.slice(0, 12)}` });
-        return withTiming(new Response(JSON.stringify(data), { status: 200, headers: { "Content-Type": "application/json" } }), t0);
-      } catch (err) {
-        if (err instanceof UpstreamEventError) {
-          if (err.status === 429 || err.status >= 500) try { activeRing.onError(blob); } catch {}
-          return withTiming(errRes(err.status, `codearts upstream: ${err.message}`, err.code), t0);
-        }
-        try { activeRing.onError(blob); } catch {}
-        lastErr = err;
-        continue;
-      }
-    }
-    // 全部号跑完仍无成功：若最后一个是账号会话饱和（非参数错），如实回传上游 400，
-    // 别退化成笼统的 "all accounts unavailable"——用户要看得到「并发会话数已达上限」。
-    if (lastErr?.upstreamStatus && lastErr?.upstreamBody != null && isSessionSaturated(lastErr.upstreamStatus, lastErr.upstreamBody)) {
-      return withTiming(errRes(lastErr.upstreamStatus, lastErr.message), t0);
-    }
-    throw Object.assign(lastErr || new Error("codearts: all accounts unavailable"), {
-      _t: { attempts: [], waitMs: 0, totalMs: Math.round(clock() - t0) },
-    });
-  }
-
-  return { runChat, buildUpstreamBody, injectReasoningPlaceholder, _sendOnce: sendOnce };
-}
+// CodeArts 聊天服务：上游 body/头组装 + SDK-HMAC-SHA256 签名 + 401 刷新重试 + 环形切号。
+// 上游恒 stream:true（chatBodyV2 契约）；流式客户端转 OpenAI SSE，非流式聚合回 JSON。
+import crypto from "node:crypto";
+import { joinUrl } from "../base.js";
+import { createTransport } from "../../transport/index.js";
+import { SNAP_BASE, EP_CHAT_V2, chatBaseHeaders, MAAS_TYPE_HEADER, MAAS_BENEFIT, MAX_TOKENS_CEILING } from "./const.js";
+import { signRequest } from "./sign.js";
+import { newChatId, preflightResponse, sseToOpenAIResponse, aggregateToCompletion, UpstreamEventError } from "./stream.js";
+
+const MODEL_RULES = [
+  { match: (m) => String(m || "").toLowerCase().includes("deepseek"), scope: "all" },
+  { match: (m) => /(^kimi-|kimi)/i.test(String(m || "")), scope: "toolCalls" },
+];
+
+// DeepSeek 系思考模式要求 assistant 消息回传 reasoning_content，缺失补 " " 占位（ADR-0001 同款）。
+export function injectReasoningPlaceholder(messages, model) {
+  const rule = MODEL_RULES.find((r) => r.match(model));
+  if (!rule || !Array.isArray(messages)) return messages;
+  return messages.map((m) => {
+    if (m?.role !== "assistant") return m;
+    if (typeof m.reasoning_content === "string" && m.reasoning_content.length) return m;
+    if (rule.scope === "toolCalls" && !(Array.isArray(m.tool_calls) && m.tool_calls.length)) return m;
+    return { ...m, reasoning_content: " " };
+  });
+}
+
+const PASSTHROUGH_KEYS = ["temperature", "top_p", "max_tokens", "max_completion_tokens", "reasoning_effort", "tools", "tool_choice", "stop", "presence_penalty", "frequency_penalty", "response_format", "user", "n", "seed"];
+
+// 账号级「会话/并发饱和」：上游按账号限同时活跃会话数（TM.00001041 = 并发会话数已达上限(3个)）。
+// 这是账号此刻进不去，不是本请求参数错；槽位在上游保留分钟级才回收（与本地进程无关）。
+// 多号场景换另一个账号通常立即成功——故短冷却本账号让下一次 next() 换号并 continue 重试。
+const SESSION_SATURATED_RE = /TM\.00001041|并发会话数已达上限/;
+const SESSION_SATURATED_COOLDOWN_MS = Number.parseInt(process.env.MSLXDFF_CODEARTS_SESSION_COOLDOWN_MS ?? "", 10) || 30_000;
+
+export function buildUpstreamBody(body, { chatId, upstreamModel }) {
+  const out = {
+    model: upstreamModel,
+    stream: true, // v2 端点契约：恒流式，非流式由本层聚合
+    messages: injectReasoningPlaceholder(body?.messages || [], upstreamModel),
+    prompt_cache_key: chatId,
+    tool_stream: true,
+    chat_id: chatId,
+  };
+  for (const k of PASSTHROUGH_KEYS) {
+    if (k === "max_completion_tokens") continue; // v2 端点不认这个名字（透传 → InferHub.001001005.400，2026-10-16 现网 A/B 复现）
+    if (body?.[k] !== undefined && body?.[k] !== null) out[k] = body[k];
+  }
+  // Responses 风格客户端只发 max_completion_tokens：映射成 v2 认的 max_tokens（显式 max_tokens 优先）
+  if (out.max_tokens === undefined && body?.max_completion_tokens != null) out.max_tokens = body.max_completion_tokens;
+  // 超上限的 max_tokens 上游一律回 InferHub.001001005.400（参数非法），在这里夹住（env 可放宽）
+  const ceiling = Number.parseInt(process.env.MSLXDFF_CODEARTS_MAX_TOKENS_CEILING ?? "", 10) || MAX_TOKENS_CEILING;
+  if (typeof out.max_tokens === "number" && out.max_tokens > ceiling) out.max_tokens = ceiling;
+  return out;
+}
+
+// Session-Id：由 chatId 派生的稳定会话标识（prompt cache 亲和，对齐 Go sessionIDFor）。
+function sessionIdFor(chatId) {
+  return crypto.createHash("sha256").update(`codearts-session:${chatId}`, "utf8").digest("hex").slice(0, 32);
+}
+
+function stripOwnPrefix(model, providerId) {
+  const s = String(model || "");
+  const i = s.indexOf("/");
+  if (i <= 0) return s;
+  return s.slice(0, i).toLowerCase() === String(providerId || "").toLowerCase() ? s.slice(i + 1) : s;
+}
+
+// 从 4xx/5xx 正文里判「账号级会话饱和」：命中则可换号重试（不同于参数类 400，换号无意义）。
+export function isSessionSaturated(status, bodyText) {
+  if (status !== 400 && status !== 429 && status < 500) return false;
+  return SESSION_SATURATED_RE.test(String(bodyText || ""));
+}
+
+export function createChatService({
+  id = "codearts",
+  baseUrl = SNAP_BASE,
+  chatPath = EP_CHAT_V2,
+  fetchImpl,
+  dispatcher,
+  authPool,
+  catalog,
+  connectTimeoutMs = 30_000,
+  clock = Date.now,
+  } = {}) {
+  const resolvedBase = String(baseUrl).trim().replace(/\/+$/, "") || SNAP_BASE;
+  const transport = createTransport({ fetchImpl, dispatcher, keepAlive: !!dispatcher, timeoutMs: connectTimeoutMs, retry: {} });
+
+  async function sendOnce({ bodyStr, account, benefit, chatId, sessionId, traceId }) {
+    const url = joinUrl(resolvedBase, chatPath);
+    const h = chatBaseHeaders(account.securityToken, traceId);
+    if (benefit) h[MAAS_TYPE_HEADER] = MAAS_BENEFIT;
+    const { headers } = signRequest({ method: "POST", url, headers: h, body: bodyStr, cred: account });
+    // Chat-Id / Session-Id 在签名之后追加（对齐官方客户端：不进 SignedHeaders）。
+    headers["chat-id"] = chatId;
+    headers["session-id"] = sessionId;
+    return transport.request({ url, method: "POST", headers, body: bodyStr, stream: true, timeoutMs: connectTimeoutMs });
+  }
+
+  function errRes(status, msg, code) {
+    return new Response(JSON.stringify({ error: { message: msg, type: "upstream_error", ...(code ? { code } : {}) } }), {
+      status,
+      headers: { "Content-Type": "application/json" },
+    });
+  }
+
+  function withTiming(res, t0) {
+    try { res._t = { attempts: [], waitMs: 0, totalMs: Math.round(clock() - t0) }; } catch {}
+    return res;
+  }
+
+  /**
+   * 跑一次对话。body.model 已由 dispatcher 剥前缀（容忍自带前缀）。
+   * @returns {Promise<Response>} 流式 → OpenAI SSE；非流式 → chat.completion JSON
+   */
+  async function runChat(body, activeRing, opts = {}) {
+    const t0 = clock();
+    const requestedModel = String(body?.model || "");
+    const rawModel = stripOwnPrefix(requestedModel, id);
+    const upstreamModel = catalog?.canonical(rawModel) || rawModel;
+    const chatId = newChatId(body, opts);
+    const sessionId = sessionIdFor(chatId);
+    const upstreamBody = buildUpstreamBody(body, { chatId, upstreamModel });
+    const bodyStr = JSON.stringify(upstreamBody);
+    const isStream = body?.stream === true;
+    let lastErr = null;
+
+    const maxTries = Math.max(1, activeRing.size || 1);
+    for (let attempt = 0; attempt < maxTries; attempt++) {
+      const blob = activeRing.next();
+      if (!blob) break;
+      let account;
+      try {
+        account = await authPool.getCredential(blob);
+      } catch (err) {
+        try { activeRing.onError(blob); } catch {}
+        lastErr = err;
+        continue; // 死号/不在池：换下一个账号
+      }
+      const benefit = catalog?.isBenefit ? catalog.isBenefit(account.userId, upstreamModel) : false;
+      const traceId = crypto.randomBytes(8).toString("hex") + String(clock());
+      let res;
+      try {
+        res = await sendOnce({ bodyStr, account, benefit, chatId, sessionId, traceId });
+      } catch (err) {
+        try { activeRing.onError(blob); } catch {}
+        lastErr = err;
+        continue;
+      }
+      if (res.status === 401 || res.status === 403) {
+        // STS 临时凭证失效：强制刷新重试一次，仍 401/403 换号
+        authPool.invalidate(blob);
+        let account2 = null;
+        try { account2 = await authPool.getCredential(blob); } catch { /* 死号 → 换号 */ }
+        if (account2) {
+          const traceId2 = crypto.randomBytes(8).toString("hex") + String(clock());
+          try { res = await sendOnce({ bodyStr, account: account2, benefit, chatId, sessionId, traceId: traceId2 }); } catch (err2) { lastErr = err2; try { activeRing.onError(blob); } catch {} continue; }
+          if (res.status === 401 || res.status === 403) {
+            try { await res.text(); } catch {}
+            try { activeRing.onError(blob); } catch {}
+            lastErr = new Error(`codearts auth still failing after refresh (http ${res.status})`);
+            continue;
+          }
+        } else {
+          try { activeRing.onError(blob); } catch {}
+          continue;
+        }
+      }
+      if (res.status >= 400) {
+        const txt = await res.text().catch(() => "");
+        const msg = `codearts upstream http ${res.status}: ${String(txt).slice(0, 300)}`;
+        // 账号会话饱和：短冷却本号并换号重试（多号场景另一号通常立即成功）；
+        // 所有号都饱和才落到重试耗尽分支，交给上层按上游不可用输出。
+        if (isSessionSaturated(res.status, txt)) {
+          try { activeRing.onError(blob, SESSION_SATURATED_COOLDOWN_MS); } catch {}
+          lastErr = Object.assign(new Error(msg), { upstreamStatus: res.status, upstreamBody: txt });
+          continue;
+        }
+        if (res.status === 429 || res.status >= 500) try { activeRing.onError(blob); } catch {}
+        return withTiming(errRes(res.status, msg), t0);
+      }
+      try {
+        const pre = await preflightResponse(res); // HTTP 200 内嵌错误（排队/未注册/福利未领）→ 抛 UpstreamEventError
+        if (isStream) return withTiming(sseToOpenAIResponse(pre, { model: requestedModel || upstreamModel, id: `chatcmpl-${chatId.slice(0, 12)}` }), t0);
+        const data = await aggregateToCompletion(pre, { model: requestedModel || upstreamModel, id: `chatcmpl-${chatId.slice(0, 12)}` });
+        return withTiming(new Response(JSON.stringify(data), { status: 200, headers: { "Content-Type": "application/json" } }), t0);
+      } catch (err) {
+        if (err instanceof UpstreamEventError) {
+          if (err.status === 429 || err.status >= 500) try { activeRing.onError(blob); } catch {}
+          return withTiming(errRes(err.status, `codearts upstream: ${err.message}`, err.code), t0);
+        }
+        try { activeRing.onError(blob); } catch {}
+        lastErr = err;
+        continue;
+      }
+    }
+    // 全部号跑完仍无成功：若最后一个是账号会话饱和（非参数错），如实回传上游 400，
+    // 别退化成笼统的 "all accounts unavailable"——用户要看得到「并发会话数已达上限」。
+    if (lastErr?.upstreamStatus && lastErr?.upstreamBody != null && isSessionSaturated(lastErr.upstreamStatus, lastErr.upstreamBody)) {
+      return withTiming(errRes(lastErr.upstreamStatus, lastErr.message), t0);
+    }
+    throw Object.assign(lastErr || new Error("codearts: all accounts unavailable"), {
+      _t: { attempts: [], waitMs: 0, totalMs: Math.round(clock() - t0) },
+    });
+  }
+
+  return { runChat, buildUpstreamBody, injectReasoningPlaceholder, _sendOnce: sendOnce };
+}

package/src/providers/codearts/const.js

@@ -44,12 +44,19 @@ export const MAAS_TYPE_HEADER = "maas_type";
 export const MAAS_BENEFIT = "benefit";
 
 // 冷启动种子：福利套餐轮换后新模型在首次发现前也按福利路由（宁多带不少带，少带即 404）。
-// 现网取证 2026-10-16（福利网关 /api/v1/gateway/config，error_code=0000）共 4 位：下列 4 个 id 逐字对齐上游。
-export const SEED_BENEFIT_MODELS = ["deepseek-v4-flash-0731", "deepseek-v4-pro-0813", "deepseek-v4.1-flash", "glm-5.3-flash"];
-// 内置模型种子（仅用于大小写归一，不标福利）。
+// 现网取证 2026-10-16（福利网关 /api/v1/gateway/config，error_code=0000）共 4 位：下列 id 逐字对齐上游。
+// 2026-10-24 再取证：`deepseek-v4.1-flash` 已转 builtin 正式模型（福利网关该号已空）。它留在福利种子里
+// 是有害的 —— preheat 未回来时带福利头 → InferHub.4004.200 benefit not found，不带才 200。故移出、转入正式表。
+export const SEED_BENEFIT_MODELS = ["deepseek-v4-flash-0731", "deepseek-v4-pro-0813", "glm-5.3-flash"];
+// 内置模型种子（仅用于大小写归一，不标福利）。id 一律写上游真正的 model_id 大小写：
+// `deepseek-v4.1-flash` 是全小写，`GLM-5.2` 就是混合大小写——猜错即 InferHub.002002009.404。
 export const SEED_KNOWN_MODELS = [
-  "GLM-5.2", "GLM-5.2-ArkTS-SPARK", "GLM-5.1", "GLM-4.7",
-  "OpenPangu-2.0-Pro", "OpenPangu-2.0-Flash",
+  // 现网取证 2026-10-24（builtin 路真 model_id 逐字）：这些就是全小写，展示名（OpenPangu-2.0-Pro /
+  // GLM-5.3 / Kimi-K3）出自 agent-center 的 model_name，当 id 发上去必 InferHub.002002009.404。
+  // 同名漂移记在 name 字段，不占用 id：GLM-5.2-ArkTS-SPARK 的真 id 是 glm-5.2-sft-harmony。
+  "deepseek-v4.1-flash", "glm-5.3", "kimi-k3", "openpangu-2.0-pro", "openpangu-2.0-flash",
+  "glm-5.2-sft-harmony", "kimi-k2.6-vl",
+  "GLM-5.2", "GLM-5.1", "GLM-4.7",
   "Qwen3-VL-235B", "Qwen3.5-397B-A17B-VL", "Qwen3.6-27B-VL",
 ];
 

package/src/providers/codearts/models.js

@@ -14,8 +14,19 @@ const lower = (s) => String(s || "").trim().toLowerCase();
 
 export function createCatalog() {
   const known = new Map(); // lower → 精确 ID（种子 + 已发现，大小写归一单点）
+  const knownReal = new Set(); // 该精确 ID 出自上游真正的 model_id（不是展示名兜底）
+
+  // id 权威性：真正的 model_id > 展示名（agent-center 无 model_id 时拿 model_name 顶包，
+  // 发上去必 InferHub.002002009.404）；同级才用历史规则「非小写的精确大小写覆盖小写」。
+  function noteKnown(key, id, byModelId) {
+    if (!known.has(key)) { known.set(key, id); if (byModelId) knownReal.add(key); return; }
+    if (byModelId && !knownReal.has(key)) { known.set(key, id); knownReal.add(key); return; }
+    if (!!byModelId === knownReal.has(key) && id !== key && known.get(key) === key) known.set(key, id);
+  }
+
   for (const id of [...SEED_KNOWN_MODELS, ...SEED_BENEFIT_MODELS]) {
-    if (!known.has(lower(id))) known.set(lower(id), id);
+    const key = lower(id);
+    if (key) noteKnown(key, id, false); // 种子按展示名记账：真 model_id 一发现就覆盖它
   }
   const accounts = new Map(); // userId → {models, benefit:Set, knows:Set, lastBenefit:Map, fetched}
 
@@ -28,7 +39,7 @@ export function createCatalog() {
       const key = lower(m.id);
       ac.knows.add(key);
       if (m.benefit) { ac.benefit.add(key); ac.lastBenefit.set(key, m); }
-      if (!known.has(key) || m.id !== key) known.set(key, m.id);
+      noteKnown(key, m.id, m.byModelId);
       ac.models.push(m);
     }
     if (keepBenefit) {
@@ -38,7 +49,7 @@ export function createCatalog() {
         ac.benefit.add(key);
         ac.knows.add(key);
         ac.models.push(mi);
-        if (!known.has(key) || mi.id !== key) known.set(key, mi.id);
+        noteKnown(key, mi.id, mi.byModelId);
       }
       ac.models.sort((a, b) => (lower(a.id) < lower(b.id) ? -1 : lower(a.id) > lower(b.id) ? 1 : (a.id < b.id ? -1 : 1)));
     }
@@ -72,7 +83,7 @@ export function createCatalog() {
   return { setAccount, isBenefit, canonical, accountModels, isStale };
 }
 
-/** 多路合并：按小写 ID 去重，福利标记取或，参数更全/精确大小写优先，按小写 ID 排序。 */
+/** 多路合并：按小写 ID 去重，福利标记取或，真 model_id 优先 > 参数更全 > 精确大小写，按小写 ID 排序。 */
 export function mergeModels(sets) {
   const merged = new Map();
   for (const set of sets || []) {
@@ -81,7 +92,11 @@ export function mergeModels(sets) {
       const key = lower(m.id);
       const cur = merged.get(key);
       if (!cur) { merged.set(key, { ...m }); continue; }
-      const keep = (m.contextWindow > 0 && !cur.contextWindow) || (m.id !== lower(m.id) && cur.id === lower(cur.id)) ? m : cur;
+      // 上游真正的 model_id 优先于展示名：agent-center 那份只有 model_name（DeepSeek-V4.1-Flash），
+      // builtin 那份才是可请求的 id（deepseek-v4.1-flash）；发错大小写 = InferHub.002002009.404。
+      const keep = !!m.byModelId !== !!cur.byModelId
+        ? (m.byModelId ? m : cur)
+        : ((m.contextWindow > 0 && !cur.contextWindow) || (m.id !== lower(m.id) && cur.id === lower(cur.id)) ? m : cur);
       const drop = keep === m ? cur : m;
       merged.set(key, {
         ...keep,
@@ -130,6 +145,9 @@ const toModelInfo = (m, { benefit = false } = {}) => {
   if (!id) return null;
   return {
     id: String(id),
+    // 出处记账：只有真正的 model_id 字段才敢当「发给上游的 id」；靠 model_name/modelAlias
+    // 兜底出来的是展示名，上游 InferHub 不认（现网 2026-10-24：DeepSeek-V4.1-Flash 必 404）。
+    byModelId: !!m?.model_id,
     name: String(m?.model_name || id),
     contextWindow: Number(m?.context_window) || 0,
     maxTokens: Number(m?.max_tokens) || 0,