@acosmi/sdk-ts 2.19.4 → 2.19.6

package/dist/node/index.cjs

@@ -150,6 +150,14 @@ var init_types = __esm({
 });
 
 // src/shared/errors.ts
+function finiteNonNegative(value) {
+  return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
+}
+function isStreamError(err) {
+  if (err == null || typeof err !== "object") return false;
+  const e = err;
+  return e.name === "StreamError" && typeof e.errorCode === "string";
+}
 function nullableID(value) {
   return value === null || typeof value === "string";
 }
@@ -359,8 +367,23 @@ var init_errors = __esm({
       userMessage;
       /** gateway 原始 error 字符串 */
       rawError;
-      /** 客户端是否值得重试 */
+      /** 客户端是否值得重试 (= 服务端 `retryable` **且** `requestDisposition === 'not_accepted'`) */
       retryable;
+      /**
+       * 服务端在帧上声明的 `retryable` **原值** (2.19.5 起) —— 不与 requestDisposition 合取,
+       * 缺席即 undefined。
+       *
+       * 为什么要并列存一份: 上面的 `retryable` 是两个判据的合取, 网关的流内错误帧并不总是
+       * 下发 `requestDisposition`, 于是"服务端明说可重试"会被合取成 false —— 两端各自正确,
+       * 合起来恒否。消费方要区分**「服务端明说不可重试」**(false, 一票否决) 与**「服务端没说」**
+       * (undefined, 回落到自己的本地分类表) 时, 只能读这个原值; 读 `retryable` 两种情形不可分。
+       */
+      serverRetryable;
+      /**
+       * 服务端建议的等待秒数 (2.19.5 起)。**有限非负数**才收, 其余 (null / 负数 /
+       * NaN / Infinity / 非数字) 一律 undefined —— 宁可不给, 绝不编造等待时间。
+       */
+      retryAfterSecs;
       errorContractVersion;
       faultDomain;
       requestDisposition;
@@ -383,6 +406,8 @@ var init_errors = __esm({
         this.userMessage = userMessage;
         this.rawError = rawError;
         this.retryable = retryable;
+        this.serverRetryable = opts.retryable;
+        this.retryAfterSecs = finiteNonNegative(opts.retryAfterSecs);
         this.errorContractVersion = opts.errorContractVersion;
         this.faultDomain = opts.faultDomain;
         this.requestDisposition = opts.requestDisposition;
@@ -2197,7 +2222,10 @@ var FileTokenStore = class {
   /**
    * 跨进程临界区. 用 sidecar `<path>.lock` 文件 + O_EXCL 创建语义实现互斥:
    *   - 创建成功 = 持有锁; 写入 pid+timestamp 便于诊断
-   *   - 创建失败 (EEXIST) = 别的进程持锁, backoff 重试
+   *   - 创建失败 (EEXIST) = 别的进程持锁, stale 检测后 backoff 重试
+   *   - 创建失败 (Windows 的 EPERM / EBUSY) = 前一持有者的 unlink 处于 delete-pending,
+   *     同属争用; 跳过 stale 检测直接 backoff 重试 (见 catch 内注释)。EACCES 不在其中 —
+   *     它是真权限错误, 立即抛
    *   - 锁文件 mtime > staleMs = 进程崩溃残留, unlink 后重试
    *   - acquireTimeoutMs 超时 = 抛错 (caller 应作为 transient error 处理, 上层 retry)
    *
@@ -2226,25 +2254,29 @@ ${Date.now()}
         release = async () => {
           try {
             await fs.unlink(lockPath);
-          } catch {
+          } catch (e) {
+            warnLockCleanupFailed("release", lockPath, e);
           }
         };
         break;
       } catch (e) {
-        if (!isAlreadyExistsError(e)) throw e;
-        let stale = false;
-        try {
-          const st = await fs.stat(lockPath);
-          if (Date.now() - st.mtimeMs > fileLockDefaults.staleMs) stale = true;
-        } catch {
-          continue;
-        }
-        if (stale) {
+        if (!isLockContendedError(e)) throw e;
+        if (!isWindowsLockBusyError(e)) {
+          let stale = false;
           try {
-            await fs.unlink(lockPath);
+            const st = await fs.stat(lockPath);
+            if (Date.now() - st.mtimeMs > fileLockDefaults.staleMs) stale = true;
           } catch {
+            continue;
+          }
+          if (stale) {
+            try {
+              await fs.unlink(lockPath);
+            } catch (e2) {
+              warnLockCleanupFailed("stale-break", lockPath, e2);
+            }
+            continue;
           }
-          continue;
         }
         if (Date.now() - startMs > fileLockDefaults.acquireTimeoutMs) {
           throw new Error(
@@ -2344,12 +2376,27 @@ ${Date.now()}
     });
   }
 };
+function warnLockCleanupFailed(stage, lockPath, e) {
+  if (isNotExistError(e)) return;
+  console.warn(
+    `[acosmi-sdk] warning: unlink token file lock (${stage}) failed: ${lockPath}: ${e instanceof Error ? e.message : String(e)}`
+  );
+}
 function isNotExistError(e) {
   if (typeof e === "object" && e !== null && "code" in e) {
     return e.code === "ENOENT";
   }
   return false;
 }
+function isWindowsLockBusyError(e) {
+  if (typeof process === "undefined" || process.platform !== "win32") return false;
+  if (typeof e !== "object" || e === null || !("code" in e)) return false;
+  const code = e.code;
+  return code === "EPERM" || code === "EBUSY";
+}
+function isLockContendedError(e) {
+  return isAlreadyExistsError(e) || isWindowsLockBusyError(e);
+}
 function isAlreadyExistsError(e) {
   if (typeof e === "object" && e !== null && "code" in e) {
     return e.code === "EEXIST";
@@ -2943,7 +2990,7 @@ function parseStreamError(data) {
   let code = payload.errorCode ?? "";
   const stage = payload.stage ?? "";
   let message = payload.message ?? "";
-  const retryable = payload.retryable ?? false;
+  const retryable = payload.retryable;
   let rawError = "";
   if (payload.error != null) {
     if (typeof payload.error === "string") {
@@ -2961,6 +3008,7 @@ function parseStreamError(data) {
     message,
     rawError,
     retryable,
+    retryAfterSecs: payload.retryAfterSecs,
     errorContractVersion: payload.errorContractVersion,
     faultDomain: payload.faultDomain,
     requestDisposition: payload.requestDisposition,
@@ -3171,6 +3219,25 @@ function notifyGatewayRequestID(cb, headers) {
   } catch {
   }
 }
+function notifyResponseHeaders(cb, headers) {
+  if (!cb) return;
+  try {
+    cb(headers);
+  } catch {
+  }
+}
+function armStreamRequestTimeout(opts, parent) {
+  const ms = opts?.requestTimeoutMs;
+  if (typeof ms !== "number" || !Number.isFinite(ms) || ms <= 0) return void 0;
+  return withRequestTimeout(ms, parent);
+}
+async function* disposeTimeoutWhenSettled(inner, ctl) {
+  try {
+    yield* inner;
+  } finally {
+    ctl.dispose();
+  }
+}
 var DEFAULT_API_TIMEOUT_MS = 6e4;
 function newDeferred() {
   let resolve;
@@ -4304,10 +4371,23 @@ var Client = class _Client {
    *
    * @param onUpstreamActivity 见 {@link UpstreamActivityCallback}
    * @param onGatewayRequestID 见 {@link GatewayRequestIDCallback}
+   * @param opts 见 {@link ChatStreamOptions} (2.19.5 起)
    */
-  chatStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID) {
+  chatStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID, opts) {
     return {
-      [Symbol.asyncIterator]: () => this.chatStreamGen(modelID, req, signal, false, onUpstreamActivity, onGatewayRequestID)
+      [Symbol.asyncIterator]: () => {
+        const ctl = armStreamRequestTimeout(opts, signal);
+        const inner = this.chatStreamGen(
+          modelID,
+          req,
+          ctl?.signal ?? signal,
+          false,
+          onUpstreamActivity,
+          onGatewayRequestID,
+          opts?.onResponseHeaders
+        );
+        return ctl ? disposeTimeoutWhenSettled(inner, ctl) : inner;
+      }
     };
   }
   /**
@@ -4317,13 +4397,26 @@ var Client = class _Client {
    *
    * @param onUpstreamActivity 见 {@link UpstreamActivityCallback}
    * @param onGatewayRequestID 见 {@link GatewayRequestIDCallback}
+   * @param opts 见 {@link ChatStreamOptions} (2.19.5 起)
    */
-  chatMessagesStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID) {
+  chatMessagesStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID, opts) {
     return {
-      [Symbol.asyncIterator]: () => this.chatMessagesStreamGen(modelID, req, signal, false, onUpstreamActivity, onGatewayRequestID)
+      [Symbol.asyncIterator]: () => {
+        const ctl = armStreamRequestTimeout(opts, signal);
+        const inner = this.chatMessagesStreamGen(
+          modelID,
+          req,
+          ctl?.signal ?? signal,
+          false,
+          onUpstreamActivity,
+          onGatewayRequestID,
+          opts?.onResponseHeaders
+        );
+        return ctl ? disposeTimeoutWhenSettled(inner, ctl) : inner;
+      }
     };
   }
-  async *chatStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID) {
+  async *chatStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID, onResponseHeaders) {
     const r = { ...req, stream: true };
     const { body, adapter } = await this.buildChatRequest(modelID, r, signal);
     const token = await this.ensureToken(signal);
@@ -4355,10 +4448,11 @@ var Client = class _Client {
           `stream: unauthorized and refresh failed: ${refreshErr instanceof Error ? refreshErr.message : String(refreshErr)}`
         );
       }
-      yield* this.chatStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID);
+      yield* this.chatStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID, onResponseHeaders);
       return;
     }
     notifyGatewayRequestID(onGatewayRequestID, resp.headers);
+    notifyResponseHeaders(onResponseHeaders, resp.headers);
     if (!resp.ok) {
       const bodyBytes = await readLimited(resp.body, maxErrorBodySize);
       throw parseHTTPErrorWithHeader(resp.status, bodyBytes, resp.headers);
@@ -4393,7 +4487,7 @@ var Client = class _Client {
       }
     }
   }
-  async *chatMessagesStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID) {
+  async *chatMessagesStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID, onResponseHeaders) {
     const r = { ...req, stream: true };
     const { body, adapter } = await this.buildChatRequest(modelID, r, signal);
     const token = await this.ensureToken(signal);
@@ -4425,10 +4519,11 @@ var Client = class _Client {
           `messages stream: unauthorized and refresh failed: ${refreshErr instanceof Error ? refreshErr.message : String(refreshErr)}`
         );
       }
-      yield* this.chatMessagesStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID);
+      yield* this.chatMessagesStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID, onResponseHeaders);
       return;
     }
     notifyGatewayRequestID(onGatewayRequestID, resp.headers);
+    notifyResponseHeaders(onResponseHeaders, resp.headers);
     if (!resp.ok) {
       const bodyBytes = await readLimited(resp.body, maxErrorBodySize);
       throw parseHTTPErrorWithHeader(resp.status, bodyBytes, resp.headers);
@@ -9114,6 +9209,7 @@ exports.isPlatform = isPlatform;
 exports.isRegion = isRegion;
 exports.isSSECommentLine = isSSECommentLine;
 exports.isSSLError = isSSLError;
+exports.isStreamError = isStreamError;
 exports.isTerminalRemoteEvent = isTerminalRemoteEvent;
 exports.isUserAccessTokenRejected = isUserAccessTokenRejected;
 exports.isValidTokenSet = isValidTokenSet;