@aipack-ai/multi-agent 0.0.1 → 1.0.0

package/dist/index.js

@@ -345,6 +345,7 @@ var ExtensionManager = class {
       beforeRun: new AsyncSeriesWaterfallHook("beforeRun"),
       beforeTransform: new AsyncSeriesWaterfallHook("beforeTransform"),
       afterTransform: new AsyncSeriesWaterfallHook("afterTransform"),
+      beforeModelCall: new AsyncSeriesWaterfallHook("beforeModelCall"),
       beforeEmit: new AsyncSeriesHook("beforeEmit"),
       afterEmit: new AsyncSeriesHook("afterEmit"),
       done: new AsyncSeriesHook("done"),
@@ -768,6 +769,24 @@ function isContextOverflow(message, contextWindow) {
       return true;
     }
   }
+  if (message.stopReason === "length" && message.content) {
+    const blocks = Array.isArray(message.content) ? message.content : null;
+    if (blocks) {
+      const hasMeaningfulText = blocks.some(
+        (b) => b.type === "text" && typeof b.text === "string" && b.text.trim().length > 0
+      );
+      const hasToolCall = blocks.some((b) => b.type === "toolCall");
+      const hasThinking = blocks.some((b) => b.type === "thinking");
+      if (!hasMeaningfulText && !hasToolCall && hasThinking) {
+        return true;
+      }
+    }
+  }
+  if (message.maxTokens && message.maxTokens > 0 && message.stopReason === "length" && message.usage) {
+    if (message.usage.output >= Math.floor(message.maxTokens * 0.95)) {
+      return true;
+    }
+  }
   return false;
 }
 
@@ -1036,7 +1055,8 @@ var AgentRuntime = class _AgentRuntime {
       config: runtime._config,
       workspace: options.workspace ?? process.cwd(),
       sessionKey: runtime._sessionKey,
-      shared: /* @__PURE__ */ new Map()
+      shared: /* @__PURE__ */ new Map(),
+      runtime
     };
     runtime._extensionContext = ctx;
     runtime._extensions.applyAll(ctx);
@@ -1863,12 +1883,14 @@ ${text}`,
   /**
    * 单回合模型调用 + 上下文溢出自动恢复闭环。
    *
-   * 检测（isContextOverflow，统一传入 model.contextWindow，覆盖显式错误 /
-   * 静默溢出 / 截断溢出三模式）→ 丢弃失败的 assistant 消息 → 截断会话历史
-   * → 同回合重试（不消耗回合数，上限 OVERFLOW_RECOVERY_LIMIT）：
+   * 检测（isContextOverflow，统一传入 model.contextWindow / content / maxTokens，
+   * 覆盖显式错误 / 静默溢出 / 输入截断溢出 / 输出 thinking 耗尽 / 输出打满 五模式）
+   * → 丢弃失败的 assistant 消息 → 截断会话历史 → 同回合重试（不消耗回合数，
+   * 上限 OVERFLOW_RECOVERY_LIMIT）：
    *
-   * - 显式错误 / 零产出截断溢出：丢弃错误消息后重试；流式路径吞掉可恢复的
-   *   error chunk（消费者看不到瞬态错误），不可恢复时补发。
+   * - 显式错误 / 零产出截断溢出 / thinking 耗尽溢出 / 输出打满溢出：丢弃失败
+   *   消息后重试；流式路径吞掉可恢复的 error chunk（消费者看不到瞬态错误），
+   *   不可恢复时补发。
    * - 静默溢出（stop + 有完整产出）：保留回复，仅压缩旧上下文供后续轮次。
    * - 恢复耗尽或单请求超窗（无可丢弃）：返回最后一次错误消息，维持旧行为。
    *
@@ -1877,14 +1899,22 @@ ${text}`,
    */
   async *modelTurnWithRecovery(compilation, signal, sessionKey, stream) {
     const contextWindow = this._model.contextWindow;
+    const maxTokens = this._model.maxTokens;
     let recoveries = 0;
+    const toProbe = (m) => ({
+      stopReason: m.stopReason,
+      errorMessage: m.errorMessage,
+      usage: m.usage,
+      content: m.content,
+      maxTokens
+    });
     while (true) {
       let assistant = null;
       let suppressed = null;
       for await (const event of this.streamModelEvents(compilation, signal, stream)) {
         if (event.type === "error") {
           const msg = event.message;
-          if (recoveries < OVERFLOW_RECOVERY_LIMIT && isContextOverflow(msg, contextWindow)) {
+          if (recoveries < OVERFLOW_RECOVERY_LIMIT && isContextOverflow(toProbe(msg), contextWindow)) {
             suppressed = msg;
             continue;
           }
@@ -1899,13 +1929,22 @@ ${text}`,
       }
       const final = assistant ?? suppressed;
       if (!final) return null;
-      if (isContextOverflow(final, contextWindow)) {
-        const failed = final.stopReason === "error" || (final.usage?.output ?? 0) === 0;
+      if (isContextOverflow(toProbe(final), contextWindow)) {
+        const output = final.usage?.output ?? 0;
+        const blocks = Array.isArray(final.content) ? final.content : null;
+        const hasMeaningfulText = blocks?.some(
+          (b) => b.type === "text" && "text" in b && typeof b.text === "string" && b.text.trim().length > 0
+        );
+        const hasToolCall = blocks?.some((b) => b.type === "toolCall");
+        const thinkingOnly = blocks ? !hasMeaningfulText && !hasToolCall && blocks.some((b) => b.type === "thinking") : false;
+        const outputFull = maxTokens > 0 && output >= Math.floor(maxTokens * 0.95);
+        const failed = final.stopReason === "error" || output === 0 || thinkingOnly || outputFull;
         if (failed && recoveries < OVERFLOW_RECOVERY_LIMIT) {
           recoveries += 1;
           if (await this.recoverFromOverflow(compilation, recoveries, sessionKey, signal)) {
+            const reason = thinkingOnly ? "thinking \u8017\u5C3D" : outputFull ? "output \u6253\u6EE1" : final.stopReason;
             console.warn(
-              `[Runtime] \u4E0A\u4E0B\u6587\u6EA2\u51FA\uFF0C\u5DF2\u538B\u7F29\u5386\u53F2\uFF08\u6458\u8981\u6216\u622A\u65AD\uFF09\u5E76\u540C\u56DE\u5408\u91CD\u8BD5\uFF08${recoveries}/${OVERFLOW_RECOVERY_LIMIT}\uFF09`
+              `[Runtime] \u4E0A\u4E0B\u6587\u6EA2\u51FA\uFF08${reason}\uFF09\uFF0C\u5DF2\u538B\u7F29\u5386\u53F2\u5E76\u540C\u56DE\u5408\u91CD\u8BD5\uFF08${recoveries}/${OVERFLOW_RECOVERY_LIMIT}\uFF09`
             );
             await this.emitTelemetry("onRetry", {
               traceId: compilation.traceId,
@@ -2215,7 +2254,10 @@ ${text}`,
       });
     };
     try {
-      for await (const event of this._streamFn(this._model, this.buildContext(compilation.messages), options)) {
+      const context = await this._hooks.beforeModelCall.promise(
+        this.buildContext(compilation.messages)
+      );
+      for await (const event of this._streamFn(this._model, context, options)) {
         if (stream && event.type === "text_delta" && ttftAt === void 0) {
           ttftAt = Date.now();
         }