@bolloon/bolloon-agent 0.1.35 → 0.1.36

package/src/agents/pi-sdk.ts

@@ -48,6 +48,8 @@ import { injectJudgmentGate, recordJudgmentUsage } from '../pi-ecosystem-judgmen
 import { monitorAfterReply } from '../pi-ecosystem-judgment/monitor-gate.js';
 // Bootstrap 生命周期 hook (SessionStart / Stop / PreToolUse)
 import { onSessionStart, onStop, onPreToolUse } from '../pi-ecosystem-judgment/human-value-pipeline.js';
+import { onPostToolUse, onJudgmentInjected, onMonitorViolation } from '../bootstrap/lifecycle-hooks.js';
+import { budgetReduce, snip, microcompact } from '../context-compaction/index.js';
 // React Harness: 8-gate + 4-guard (防越权 / 防 prompt 注入)
 import { ReactHarness } from '../security/react-harness.js';
 
@@ -480,10 +482,18 @@ export interface StreamCallback {
 }
 
 export interface StreamEvent {
-  type: 'status' | 'thinking' | 'tool' | 'token' | 'done' | 'error';
+  type: 'status' | 'thinking' | 'tool' | 'token' | 'done' | 'error'
+      | 'step_start' | 'step_done' | 'step_error';
   content: string;
   tool?: string;
   data?: unknown;
+  // step_* 专用: success / output / error
+  success?: boolean;
+  output?: string;
+  error?: string;
+  args?: Record<string, unknown>;
+  // step_* 可选: 步骤耗时 (server 端用来展示 in 状态条 + 性能分析)
+  durationMs?: number;
 }
 
 const TOOL_DEFINITIONS = `
@@ -572,9 +582,21 @@ class PiAgentSession implements AgentSession {
   private messageHistory: Message[] = [];
   private tools: Map<string, Tool> = new Map();
   private skillRegistry: SkillRegistry = new SkillRegistry();
-  private readonly MAX_REACT_ITERATIONS = 100;
+  // 2026-06-16 修: 父要求把 ReAct loop 上限放大到 "几乎无限", 靠自动压缩上下文 + fail-safe 兜底
+  // 默认 10000 — 正常任务永远跑不到, 但作为防 LLM 死循环 / 防 OOM 的最后一道闸
+  // 旧默认 100 写死导致中等复杂度任务 (10-50 个 tool call + 多步反思) 会被误杀
+  private readonly MAX_REACT_ITERATIONS = 10_000;
   private readonly MAX_REFINE_ATTEMPTS = 3;
   private readonly QUALITY_THRESHOLD = 0.6;
+  /** P1: 上下文溢出阈值 (单轮估算 token 数, 超过则强制终止防止 prompt-too-long) */
+  private readonly MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD = 60_000;  // 60K tokens 上限
+  /** 2026-06-16 新增: 累计错误总数兜底 (不管是否同工具, 累计 N 次就强制退出)
+   *  防 LLM 轮换工具名绕开 MAX_SAME_TOOL_FAILURES 的死循环攻击 */
+  private readonly MAX_TOTAL_ERRORS = 20;
+  /** 2026-06-16 新增: loop 内自动压缩触发阈值 (相对 60K 阈值的比例) */
+  private readonly LOOP_COMPACT_RATIO = 0.8;
+  /** P1: max output token 升级重试 (LLM 截断时重试, 最多 3 次) */
+  private readonly MAX_OUTPUT_TOKEN_ESCALATION_RETRIES = 3;
   private thinkingEngine = new DeepThinkingEngine(3);
   private coordinator = new AgentCoordinator(3);
   private harness: any = null;
@@ -583,6 +605,10 @@ class PiAgentSession implements AgentSession {
   private reactHarness: ReactHarness = new ReactHarness();
   private usePivotLoop: boolean = false;
   private pivotLoopConfig?: PivotLoopConfig;
+  /** P2: 当前会话的 permission mode (每次 promptStream 入口解析) */
+  private currentPermissionMode: import('./permission-mode.js').PermissionMode = 'default';
+  /** P1.2: Context Collapse 读时投影结果 (feature flag 开启时由 maybeAutoCompact 写入, buildContext 优先用) */
+  private projectedHistory: Message[] | null = null;
 
   /**
    * Judgment 注入门临时结果: 在 prompt / promptStream / promptWithPivotLoop 入口算一次, 拼到本轮 systemPrompt 末尾
@@ -620,11 +646,13 @@ class PiAgentSession implements AgentSession {
 
     safePhase('gate_compute', { detail: '正在检索相关判断力...' });
     try {
-      const gate = await injectJudgmentGate(input);
+      // P-Action 4 (2026-06-15) 路径 1 整合: 透传 maxChars=1500 (≈ 375 tokens 硬上限)
+      // 路径 2/3 检测由 injection-gate 内部 alreadyInjectedSources 处理 (目前 assembleSystemPrompt 还没注入 value-store 标记, 所以这里不传)
+      const gate = await injectJudgmentGate(input, {}, { maxChars: 1500 });
       this.judgmentGateAddition = gate.systemAddition;
       this.judgmentGateUsedIds = gate.usedIds;
       if (gate.usedIds.length > 0) {
-        safePhase('gate_done', { usedCount: gate.usedIds.length });
+        safePhase('gate_done', { usedCount: gate.usedIds.length, didInject: gate.didInject, skipReason: gate.skipReason });
       }
     } catch (err) {
       console.warn('[PiAgent] judgment gate failed (non-fatal):', err);
@@ -737,7 +765,7 @@ class PiAgentSession implements AgentSession {
     this.tools.set('read_document', {
       name: 'read_document',
       description: '读取文档内容，支持 .txt, .md, .pdf, .docx 格式',
-      parameters: { path: '文件路径' },
+      parameters: { path: 'string' },
       execute: async (args) => {
         try {
           const content = await documentReader.read(args.path);
@@ -754,7 +782,7 @@ class PiAgentSession implements AgentSession {
     this.tools.set('summarize_document', {
       name: 'summarize_document',
       description: '总结文档内容，分析并生成摘要',
-      parameters: { path: '文件路径', context: '可选上下文信息' },
+      parameters: { path: 'string', context: 'string' },
       execute: async (args) => {
         try {
           if (!this.minimaxAvailable) {
@@ -774,7 +802,7 @@ class PiAgentSession implements AgentSession {
     this.tools.set('improve_document', {
       name: 'improve_document',
       description: '根据要求改进文档内容',
-      parameters: { path: '文件路径', requirements: '改进要求' },
+      parameters: { path: 'string', requirements: 'string' },
       execute: async (args) => {
         try {
           if (!this.minimaxAvailable) {
@@ -1094,8 +1122,20 @@ class PiAgentSession implements AgentSession {
     this.currentSignal = options?.signal ?? null;
     this.currentOnStream = options?.onStream ?? null;
     await this.computeJudgmentGate(input);
+
+    // P2: 解析当前 permission mode
+    try {
+      const { resolvePermissionMode } = await import('./permission-mode.js');
+      this.currentPermissionMode = resolvePermissionMode();
+    } catch (err) {
+      console.warn('[PiAgent] resolvePermissionMode failed (non-fatal):', err);
+      this.currentPermissionMode = 'default';
+    }
+
     try {
-      return await this.runReActLoop(undefined, options?.signal);
+      // 2026-06-16: runReActLoop 现在返回 { reply, aiFailed, aiFailureReason } — 这里只需 reply 字符串
+      const loopResult = await this.runReActLoop(undefined, options?.signal);
+      return loopResult.reply;
     } finally {
       if (this.judgmentGateUsedIds.length > 0) {
         recordJudgmentUsage(this.judgmentGateUsedIds, { userInput: input }).catch((err) =>
@@ -1131,6 +1171,14 @@ class PiAgentSession implements AgentSession {
     this.currentSignal = signal ?? null;
     await this.computeJudgmentGate(input);
 
+    // P1.1: 异步跑 Auto-Compact (LLM 摘要, 仅在 budget 超限时触发, 失败静默)
+    // 复用 computeJudgmentGate 的 onStream 广播 phase, 跟 judgment 注入门风格一致
+    try {
+      await this.maybeAutoCompact(onStream, signal);
+    } catch (err) {
+      console.warn('[PiAgent] maybeAutoCompact failed (non-fatal):', err);
+    }
+
     // Bootstrap SessionStart: 收集项目 Context, 拼到 systemAddition 头部
     // (失败静默, 5s 限流防止循环)
     let bootstrapAddition = '';
@@ -1141,16 +1189,76 @@ class PiAgentSession implements AgentSession {
       console.warn('[PiAgent] onSessionStart failed (non-fatal):', err);
     }
     this.bootstrapAddition = bootstrapAddition;
-    this.promptStartTime = Date.now();
 
-    let result: string;
+    // P2: 解析当前 permission mode (BootstrapOptions > env BOLLOON_PERM_MODE > default)
     try {
-      result = await this.runReActLoop(onStream, signal);
-    } catch (err: any) {
-      // abort 失败: 视作"已中断", 抛错让上层用 partial 兜底
-      this.currentOnStream = null;
-      this.currentSignal = null;
-      throw err;
+      const { resolvePermissionMode } = await import('./permission-mode.js');
+      this.currentPermissionMode = resolvePermissionMode();
+    } catch (err) {
+      console.warn('[PiAgent] resolvePermissionMode failed (non-fatal, using default):', err);
+      this.currentPermissionMode = 'default';
+    }
+
+    this.promptStartTime = Date.now();
+
+    // 2026-06-16: loop 自动重试 — runReActLoop 内部遇到 [AI 服务调用失败] sentinel 时,
+    //   会设 aiFailed=true 并提前 break. 这里在外层重跑整个 loop (不是单次 LLM 调用),
+    //   临时网络抖动 / 配额瞬时超限可自愈. 最多 3 次, 指数退避 1s/2s/4s.
+    //   用户看到 status bar 显示 "自动重试中 X/N" — 不暴露按钮.
+    const MAX_LOOP_RETRIES = 3;
+    let attempt = 0;
+    let result: string = '';
+    let lastAiFailureReason = '';
+    while (attempt <= MAX_LOOP_RETRIES) {
+      try {
+        const loopResult = await this.runReActLoop(onStream, signal);
+        result = loopResult.reply;
+        if (!loopResult.aiFailed) break; // 正常完成, 退出 retry 循环
+        lastAiFailureReason = loopResult.aiFailureReason || 'AI 调用失败';
+      } catch (err: any) {
+        // abort 失败: 视作"已中断", 抛错让上层用 partial 兜底
+        this.currentOnStream = null;
+        this.currentSignal = null;
+        throw err;
+      }
+      attempt++;
+      if (attempt > MAX_LOOP_RETRIES) {
+        console.warn(`[PiAgent] loop 自动重试 ${MAX_LOOP_RETRIES} 次后仍失败, 终止`);
+        if (onStream) {
+          onStream({ type: 'status', content: `⛔ loop 自动重试 ${MAX_LOOP_RETRIES} 次后仍失败: ${lastAiFailureReason}`, tool: 'system' });
+        }
+        result = lastAiFailureReason || 'AI 服务调用失败, 自动重试后仍不可用';
+        break;
+      }
+      const backoffMs = 1000 * Math.pow(2, attempt - 1); // 1s, 2s, 4s
+      console.log(`[PiAgent] loop 自动重试 ${attempt}/${MAX_LOOP_RETRIES}, 等待 ${backoffMs}ms`);
+      if (onStream) {
+        onStream({ type: 'status', content: `↻ 自动重试 loop ${attempt}/${MAX_LOOP_RETRIES} (${(backoffMs / 1000).toFixed(0)}s 后)`, tool: 'system' });
+      }
+      // 中途 abort 也要响应
+      await new Promise<void>((resolve, reject) => {
+        const t = setTimeout(() => {
+          signal?.removeEventListener?.('abort', onAbort);
+          resolve();
+        }, backoffMs);
+        const onAbort = () => {
+          clearTimeout(t);
+          reject(new Error('aborted during retry backoff'));
+        };
+        if (signal?.aborted) {
+          clearTimeout(t);
+          reject(new Error('aborted during retry backoff'));
+          return;
+        }
+        signal?.addEventListener?.('abort', onAbort, { once: true });
+      });
+      // 重试时要把这条 user message 从 history 里移除 (避免下一次 runReActLoop 又重复加入),
+      // 因为 messageHistory.push({role:'user'}) 在 promptStream 顶部已经做过, 重跑 runReActLoop 不会重复 push,
+      // 但 assistant 失败那条也别留 (留了会污染下一轮 LLM context).
+      // 简化: 重试前 pop 一次 assistant (如果最后一条是 assistant)
+      if (this.messageHistory.length > 0 && this.messageHistory[this.messageHistory.length - 1].role === 'assistant') {
+        this.messageHistory.pop();
+      }
     }
     onStream({ type: 'done', content: '' });
 
@@ -1243,7 +1351,10 @@ ${this.getToolDefinitions()}
 - 当任务完成时，必须在回答末尾添加 <final gen> 标记表示结束
 - 如果需要更多信息，继续调用工具${this.judgmentGateAddition}`;
 
-    const result = await loop.execute(input, llm, systemPrompt);
+    // 2026-06-15: 把 currentOnStream 传给 loop, 让 step-timeline 在 pivot 循环里也能 emit step_start/done
+    //   之前 loop.execute() 不接 streamCallback, 导致 step-timeline 只能看到老 runReActLoop 路径
+    //   promptWithPivotLoop 路径 0 step events — UI 显示 timeline 但永远是空
+    const result = await loop.execute(input, llm, systemPrompt, this.currentOnStream ?? undefined);
 
     this.messageHistory.push({ role: 'user', content: input });
     if (result.response) {
@@ -1261,15 +1372,21 @@ ${this.getToolDefinitions()}
     return result;
   }
 
-  private async runReActLoop(onStream?: StreamCallback, signal?: AbortSignal): Promise<string> {
+  private async runReActLoop(onStream?: StreamCallback, signal?: AbortSignal): Promise<{ reply: string; aiFailed: boolean; aiFailureReason?: string }> {
     const llm = getMinimax();
     let iteration = 0;
     let finalResponse = '';
     let lastQualityScore = 0;
     let refineAttempts = 0;
     let consecutiveErrors = 0;
+    // 2026-06-16 新增: 累计错误数 (跨工具, 兜底防 LLM 轮换工具名死循环)
+    let totalErrors = 0;
     let lastFailedTool = ''; // 跟踪最近一次失败的 tool name
     let lastFailedToolCount = 0; // 最近失败工具的连续失败次数
+    // 2026-06-16: AI sentinel 标志 — runReActLoop 返回 aiFailed=true,
+    //   promptStream 据此自动重跑整个 loop 最多 N 次 (不是单次 LLM 重试)
+    let aiFailed = false;
+    let aiFailureReason = '';
     const MAX_CONSECUTIVE_ERRORS = 3;
     const MAX_SAME_TOOL_FAILURES = 3; // 同一工具连续失败 3 次, 强制让 LLM 给出最终答案
 
@@ -1289,6 +1406,55 @@ ${this.getToolDefinitions()}
     while (iteration < this.MAX_REACT_ITERATIONS) {
       iteration++;
 
+      // 停止条件 1: max turns (fail-safe 10000, 正常任务永远跑不到)
+      if (iteration >= this.MAX_REACT_ITERATIONS) {
+        console.warn(`[PiAgent] 达到最大循环数 ${this.MAX_REACT_ITERATIONS}, 强制终止 (fail-safe)`);
+        onStream?.({ type: 'error', content: `⏹️ 达到最大循环数 (${this.MAX_REACT_ITERATIONS}, fail-safe)`, tool: 'loop' });
+        finalResponse = finalResponse || '(本轮 ReAct 循环达到最大步数, 强制结束)';
+        break;
+      }
+
+      // 停止条件 2: signal.aborted (显式 abort / 用户中断)
+      if (signal?.aborted) {
+        console.warn('[PiAgent] runReActLoop aborted by signal');
+        onStream?.({ type: 'error', content: '⏹️ 用户中断', tool: 'loop' });
+        finalResponse = finalResponse || '(用户中断)';
+        break;
+      }
+
+      // 2026-06-16 新增: 累计错误兜底 — 跨工具, 防 LLM 轮换工具名绕过 MAX_SAME_TOOL_FAILURES
+      if (totalErrors >= this.MAX_TOTAL_ERRORS) {
+        console.warn(`[PiAgent] 累计错误 ${totalErrors} >= ${this.MAX_TOTAL_ERRORS}, 强制终止 (防死循环)`);
+        onStream?.({ type: 'error', content: `⛔ 累计 ${totalErrors} 次错误, 强制终止 (防止 LLM 死循环)`, tool: 'loop' });
+        finalResponse = finalResponse || `(本轮 ReAct 循环累计 ${totalErrors} 次错误, 强制结束。请换个思路或简化任务重试。)`;
+        break;
+      }
+
+      // 2026-06-16 新增: loop 内自动压缩 — token 超 80% 阈值时跑一次
+      // compact 失败走 C 路径: 不强行 break, 让现有 60K 阈值兜底 (后面有检查)
+      const compactThreshold = this.MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD * this.LOOP_COMPACT_RATIO;
+      const estimatedTokensBefore = this.estimateHistoryTokens();
+      if (estimatedTokensBefore > compactThreshold) {
+        const tokensBeforeCompact = estimatedTokensBefore;
+        console.log(`[PiAgent] loop 入口 token ${tokensBeforeCompact} > ${compactThreshold}, 触发自动压缩`);
+        onStream?.({ type: 'status', content: `🗜️ loop 自动压缩 (token ${tokensBeforeCompact} > ${compactThreshold})`, tool: 'compactor' });
+        try {
+          await this.maybeAutoCompact(onStream, signal);
+        } catch (compactErr) {
+          // C 路径: compact 失败不 break, 让 token 阈值检查兜底
+          console.warn(`[PiAgent] loop 内 maybeAutoCompact 失败 (non-fatal, 继续走 token 阈值):`, compactErr);
+        }
+      }
+
+      // 停止条件 3: context overflow (compact 后还超, 强制终止)
+      const estimatedTokens = this.estimateHistoryTokens();
+      if (estimatedTokens > this.MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD) {
+        console.warn(`[PiAgent] context overflow (${estimatedTokens} tokens > ${this.MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD})`);
+        onStream?.({ type: 'error', content: `⏹️ 上下文溢出 (${estimatedTokens} tokens, 阈值 ${this.MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD})`, tool: 'loop' });
+        finalResponse = finalResponse || '(本轮 ReAct 循环因上下文溢出终止)';
+        break;
+      }
+
       // 调试日志：显示每次循环开始
       console.log(`[PiAgent] 循环 ${iteration}/${this.MAX_REACT_ITERATIONS} 开始`);
       if (onStream) {
@@ -1335,8 +1501,27 @@ ${toolDefs}
 - 当任务完成时，必须在回答末尾添加 <final gen> 标记表示结束
 - 如果需要更多信息，继续调用工具${this.judgmentGateAddition}`;
 
-      const response = await llm.chat(context, systemPrompt, signal);
-      const reply = response.reply.trim();
+      // 3 个恢复机制 (Claude Code 论文 9-step pipeline 内部):
+      //   1. max output token 升级 (最多 3 次, 每次 maxOutputTokens 翻倍)
+      //   2. reactive compaction (prompt 估算超阈值, 跑压缩)
+      //   3. prompt-too-long (LLM 报错 4xxx token 错误, 跑 reactive compaction 再试 1 次)
+      // 失败静默: 全部重试失败 → 空 reply (上层用 no tool_use 终止)
+      const response = await this.callLlmWithRecovery(llm, context, systemPrompt, signal, onStream);
+      const reply = (response.reply || '').trim();
+
+      // 2026-06-16: 看到 [AI 服务调用失败] sentinel → 不再立即 break,
+      // 而是设 aiFailed=true, 让外层 promptStream 自动重跑整个 loop 最多 N 次
+      // (LLM API 401 / 网络错 / 配额满时, pi-ai 返回这个 prefix;
+      //  自动 retry 兜底: 临时网络抖动可自愈, 真挂 N 次后才报失败)
+      if (reply.startsWith('[AI 服务调用失败]')) {
+        console.log(`[PiAgent] 收到 AI 错误 sentinel, 标记 aiFailed, 外层会自动重试整个 loop`);
+        aiFailed = true;
+        aiFailureReason = reply.length > 200 ? reply.substring(0, 200) : reply;
+        if (onStream) {
+          onStream({ type: 'status', content: `⚠️ AI 调用失败, 将自动重试整个 loop`, tool: 'system' });
+        }
+        break;
+      }
 
       console.log(`[PiAgent] LLM 回复长度: ${reply.length}, 内容预览: "${reply.substring(0, 80)}..."`);
       console.log(`[PiAgent] LLM 完整回复:\n${reply}`);
@@ -1375,23 +1560,37 @@ ${toolDefs}
           if (toolCall.args && Object.keys(toolCall.args).length > 0) {
             onStream({ type: 'status', content: `📋 参数: ${JSON.stringify(toolCall.args)}`, tool: toolCall.name });
           }
+          // 2026-06-15: step-timeline 状态机 — 开新节点
+          onStream({
+            type: 'step_start',
+            content: `调用 ${toolCall.name}`,
+            tool: toolCall.name,
+            args: toolCall.args || {},
+          });
         }
 
         const tool = this.tools.get(toolCall.name);
         if (!tool) {
           consecutiveErrors++;
+          // 2026-06-16 新增: 未知工具也要累计 (LLM 幻觉高频场景)
+          totalErrors++;
           const errorResult: ToolResult = { success: false, error: `未知工具: ${toolCall.name}` };
           this.messageHistory.push({ role: 'tool', content: JSON.stringify(errorResult), toolResult: errorResult });
           this.logToHarness(toolCall.name, toolCall.args, errorResult);
-          console.warn(`[PiAgent] 未知工具: ${toolCall.name}，跳过并继续`);
+          console.warn(`[PiAgent] 未知工具: ${toolCall.name} (累计 ${totalErrors}/${this.MAX_TOTAL_ERRORS})，跳过并继续`);
           continue;
         }
 
         // Bootstrap PreToolUse hook: 调工具前校验 (危险命令拦截)
         // 失败静默 — hook 自身挂掉 = 放行
+        // P2: 透传 permissionMode (从 BootstrapOptions / env BOLLOON_PERM_MODE 解析)
         let toolToExecute = tool;
         try {
-          const pre = await onPreToolUse({ tool: toolCall.name, args: toolCall.args || {} });
+          const pre = await onPreToolUse({
+            tool: toolCall.name,
+            args: toolCall.args || {},
+            permissionMode: this.currentPermissionMode,
+          });
           if (!pre.allowed) {
             const deniedResult: ToolResult = {
               success: false,
@@ -1409,6 +1608,13 @@ ${toolDefs}
                 content: `🛡️ PreToolUse 拒绝 ${toolCall.name}: ${pre.reason || '安全校验失败'}`,
                 tool: toolCall.name,
               });
+              // 2026-06-15: step-timeline — 拦在 PreToolUse, 标 step_error
+              onStream({
+                type: 'step_error',
+                content: `PreToolUse 拒绝 ${toolCall.name}`,
+                tool: toolCall.name,
+                error: pre.reason || '安全校验失败',
+              });
             }
             console.warn(`[PiAgent] PreToolUse denied ${toolCall.name}: ${pre.reason}`);
             // 不调 tool.execute, 也不计 consecutiveErrors (这是用户级拒绝, 不是工具错)
@@ -1443,6 +1649,13 @@ ${toolDefs}
                 content: `🛡️ Harness ${pre.details.rejectedBy} 拒绝 ${toolCall.name}: ${pre.reason || '安全校验失败'}`,
                 tool: toolCall.name,
               });
+              // 2026-06-15: step-timeline — Harness gate 拒绝, 标 step_error
+              onStream({
+                type: 'step_error',
+                content: `Harness 拒绝 ${toolCall.name}`,
+                tool: toolCall.name,
+                error: pre.reason || '安全校验失败',
+              });
             }
             console.warn(`[PiAgent] Harness denied ${toolCall.name} (${pre.details.rejectedBy}): ${pre.reason}`);
             continue;
@@ -1452,8 +1665,26 @@ ${toolDefs}
         }
 
         try {
+          const toolStart = Date.now();
           let result = await tool.execute(toolCall.args);
-          console.log(`[PiAgent] 工具 ${toolCall.name} 执行完成: success=${result.success}`);
+          const toolDurationMs = Date.now() - toolStart;
+          console.log(`[PiAgent] 工具 ${toolCall.name} 执行完成: success=${result.success} (${toolDurationMs}ms)`);
+
+          // PostToolUse 审计 hook: 写 audit log, 默认 continue
+          try {
+            await onPostToolUse({
+              tool: toolCall.name,
+              args: toolCall.args || {},
+              result: {
+                success: result.success,
+                output: result.output?.substring(0, 500),
+                error: result.error,
+              },
+              durationMs: toolDurationMs,
+            });
+          } catch (postErr) {
+            console.warn('[PiAgent] onPostToolUse failed (non-fatal):', postErr);
+          }
 
           // Context router: 拿最近一次 preToolCall 算的 hint, 拼到 tool result messageHistory
           // (LLM 下次看到 tool result 时, 能"记得"这次调用的安全约束)
@@ -1506,8 +1737,23 @@ ${toolDefs}
                 const outputPreview = result.output.substring(0, 200);
                 onStream({ type: 'tool', content: `📤 结果: ${outputPreview}${result.output.length > 200 ? '...' : ''}`, tool: toolCall.name });
               }
+              // 2026-06-15: step-timeline 状态机 — 关闭当前节点 (成功)
+              onStream({
+                type: 'step_done',
+                content: `${toolCall.name} 执行成功`,
+                tool: toolCall.name,
+                success: true,
+                output: result.output,
+              });
             } else {
               onStream({ type: 'error', content: `❌ ${toolCall.name} 执行失败: ${result.error}`, tool: toolCall.name });
+              // 2026-06-15: step-timeline 状态机 — 关闭当前节点 (失败)
+              onStream({
+                type: 'step_error',
+                content: `${toolCall.name} 执行失败`,
+                tool: toolCall.name,
+                error: result.error,
+              });
             }
           }
 
@@ -1530,6 +1776,8 @@ ${toolDefs}
             // 不 break，继续下一次循环
           } else {
             consecutiveErrors++;
+            // 2026-06-16 新增: 累计错误 (跨工具, 兜底防 LLM 轮换工具名死循环)
+            totalErrors++;
             // 跟踪同一工具连续失败次数
             if (toolCall.name === lastFailedTool) {
               lastFailedToolCount++;
@@ -1537,7 +1785,7 @@ ${toolDefs}
               lastFailedTool = toolCall.name;
               lastFailedToolCount = 1;
             }
-            console.warn(`[PiAgent] 工具 ${toolCall.name} 执行失败 (${lastFailedToolCount}/${MAX_SAME_TOOL_FAILURES}): ${result.error}`);
+            console.warn(`[PiAgent] 工具 ${toolCall.name} 执行失败 (${lastFailedToolCount}/${MAX_SAME_TOOL_FAILURES}, 累计 ${totalErrors}/${this.MAX_TOTAL_ERRORS}): ${result.error}`);
 
             // 同一工具连续失败达到上限, 不再重试, 强制 LLM 给出最终答案
             if (lastFailedToolCount >= MAX_SAME_TOOL_FAILURES) {
@@ -1564,10 +1812,12 @@ ${toolDefs}
           }
         } catch (execError) {
           consecutiveErrors++;
+          // 2026-06-16 新增: 异常分支也要累计
+          totalErrors++;
           const errorResult: ToolResult = { success: false, error: String(execError) };
           this.messageHistory.push({ role: 'tool', content: JSON.stringify(errorResult), toolResult: errorResult });
           this.logToHarness(toolCall.name, toolCall.args, errorResult);
-          console.error(`[PiAgent] 工具执行异常: ${execError}`);
+          console.error(`[PiAgent] 工具执行异常 (累计 ${totalErrors}/${this.MAX_TOTAL_ERRORS}): ${execError}`);
         }
       } else {
         // LLM 返回的不是 tool call 格式
@@ -1586,7 +1836,10 @@ ${toolDefs}
         const containsToolCallIntent = reply.includes('调用工具') || reply.includes('tool(') ||
           reply.includes('使用工具') || reply.includes('需要获取') || reply.includes('需要查看') ||
           // 兼容 LLM 用对象字面量输出 tool call (上轮没解析成功时, 至少要继续)
-          reply.includes('tool =>') || reply.includes('[TOOL_CALL]');
+          reply.includes('tool =>') || reply.includes('[TOOL_CALL]') ||
+          // 2026-06-15 修: 兼容 LLM 用 XML 标签输出 tool call (<shell_exec>...</shell_exec>)
+          //   这时 parseToolCall 失败, 至少要让 loop 继续
+          /<\w+>[\s\S]*?<\/\w+>/.test(reply);
         const hasError = ['不存在', '找不到', '无法找到', 'not found', 'does not exist',
           '错误', 'error', '失败', 'failed'].some(k => reply.includes(k));
         const isTooShort = reply.length < 50 && reply.length > 0;
@@ -1647,7 +1900,8 @@ Workspace root folder: ${this.cwd}
       console.warn('[PiAgent] reactHarness.onSessionEnd failed (non-fatal):', err);
     }
 
-    return finalResponse;
+    // 2026-06-16: 暴露 aiFailed 标志 — promptStream 据此决定是否自动重试整个 loop
+    return { reply: finalResponse, aiFailed, aiFailureReason: aiFailureReason || undefined };
   }
 
   async deepThink(prompt: string): Promise<{ result: ThinkResult; response: string }> {
@@ -1703,7 +1957,13 @@ Workspace root folder: ${this.cwd}
   }
 
   private buildContext(): string {
-    const recentMessages = this.messageHistory.slice(-10);
+    // P1 接入: 同步跑前 3 层压缩 (Budget Reduction / Snip / Microcompact)
+    // 异步层 (Context Collapse / Auto-Compact) 在 promptStream 入口处单独跑 (用 LLM)
+    // 失败静默: 任何 stage 抛错 → 走老 slice(-10) 逻辑
+    //
+    // P1.2: 如果 maybeAutoCompact 算过 Context Collapse 投影, 用 this.projectedHistory (读时投影, 非破坏)
+    const source = this.projectedHistory ?? this.messageHistory;
+    const recentMessages = this.compressHistorySync(source).slice(-10);
     return recentMessages.map(m => {
       if (m.role === 'user') return `用户: ${m.content}`;
       if (m.role === 'assistant') return `助手: ${m.content}`;
@@ -1715,6 +1975,164 @@ Workspace root folder: ${this.cwd}
     }).join('\n');
   }
 
+  /**
+   * 估算 messageHistory 的 token 数 (4 字符 ≈ 1 token, 与 context-compaction 同步).
+   * 失败静默: 任何异常 → 0 (不阻塞)
+   */
+  private estimateHistoryTokens(): number {
+    try {
+      const { estimateTokens } = require('../context-compaction/index.js') as typeof import('../context-compaction/index.js');
+      return estimateTokens(this.messageHistory as any);
+    } catch {
+      return 0;
+    }
+  }
+
+  /**
+   * 3 个恢复机制合一:
+   *   1. max output token 升级: 最多 3 次, 每次 maxOutputTokens 翻倍 (如果 llm.chat 支持)
+   *   2. reactive compaction: 估算 > 80% 阈值, 跑 sync compressHistorySync + 必要时 maybeAutoCompact
+   *   3. prompt-too-long: LLM 报错 4xxx token 错误, 跑 reactive compaction 再试 1 次
+   *
+   * 失败静默: 全部失败 → 返回空 reply, 让上层 no-tool_use 终止
+   */
+  private async callLlmWithRecovery(
+    llm: any,
+    context: string,
+    systemPrompt: string,
+    signal: AbortSignal | undefined,
+    onStream?: (chunk: any) => void
+  ): Promise<{ reply: string }> {
+    // Reactive compaction 预检: 估算 token 超 80% 阈值, 跑一次
+    const estimated = this.estimateHistoryTokens();
+    if (estimated > this.MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD * 0.8) {
+      console.warn(`[PiAgent] reactive compaction pre-check (${estimated} tokens > 80% threshold)`);
+      onStream?.({ type: 'status', content: '⚠️ reactive compaction 预检触发', tool: 'recovery' });
+      try {
+        const compacted = this.compressHistorySync(this.messageHistory);
+        this.messageHistory = compacted;
+        if (this.estimateHistoryTokens() > this.MAX_OUTPUT_TOKEN_ESCALATION_THRESHOLD * 0.8) {
+          await this.maybeAutoCompact(onStream, signal);
+        }
+      } catch (err) {
+        console.warn('[PiAgent] reactive compaction pre-check failed:', err);
+      }
+    }
+
+    // 主调用 + 3 个恢复路径
+    let lastErr: any = null;
+    for (let attempt = 0; attempt <= this.MAX_OUTPUT_TOKEN_ESCALATION_RETRIES; attempt++) {
+      try {
+        const response = await llm.chat(context, systemPrompt, signal);
+        return { reply: response.reply || '' };
+      } catch (err: any) {
+        lastErr = err;
+        const errMsg = String(err?.message || err || '');
+        const isPromptTooLong = /token|too long|exceed|length|context|4000|413|429/i.test(errMsg);
+        if (isPromptTooLong) {
+          console.warn(`[PiAgent] prompt-too-long 触发 (attempt ${attempt + 1}), 跑 reactive compaction`);
+          onStream?.({ type: 'status', content: `⚠️ prompt-too-long 触发 (attempt ${attempt + 1}/${this.MAX_OUTPUT_TOKEN_ESCALATION_RETRIES + 1})`, tool: 'recovery' });
+          try {
+            await this.maybeAutoCompact(onStream, signal);
+          } catch (compactionErr) {
+            console.warn('[PiAgent] reactive compaction on prompt-too-long failed:', compactionErr);
+          }
+          // 重新生成 context (compressHistorySync + projected)
+          context = this.buildContext();
+        } else {
+          // 非 prompt-too-long 错误, 1 次重试就放弃
+          if (attempt === 0) {
+            console.warn(`[PiAgent] LLM 调用失败 (non-prompt-too-long), 1 次重试:`, err);
+            continue;
+          }
+          break;
+        }
+      }
+    }
+    console.warn('[PiAgent] callLlmWithRecovery 全失败 (silent):', lastErr);
+    return { reply: '' };
+  }
+
+  /**
+   * 同步压缩: 跑前 3 层 (Budget Reduction / Snip / Microcompact).
+   * Context Collapse / Auto-Compact 是 async, 不在 buildContext 同步链里跑.
+   * 失败静默: 任何 stage 抛错 → 返回原 history.
+   */
+  private compressHistorySync(history: Message[]): Message[] {
+    try {
+      // context-compaction 的 Message 与 pi-sdk 的 Message 字段兼容
+      // 这里用 any cast 跳过 structural type 严格校验 (避免双向 import)
+      let h: any = history;
+      const r1 = budgetReduce(h);
+      h = r1.history;
+      const r2 = snip(h);
+      h = r2.history;
+      const r3 = microcompact(h);
+      h = r3.history;
+      return h as Message[];
+    } catch (err) {
+      console.warn('[PiAgent] compressHistorySync failed (silent, using original):', err);
+      return history;
+    }
+  }
+
+  /**
+   * P1.1: 异步跑 Auto-Compact (LLM 摘要).
+   * 入口: promptStream 入口, 在 computeJudgmentGate 之后, onSessionStart 之前.
+   *
+   * 逻辑:
+   *   1. 跑完整 compactPipeline (5 层, 异步)
+   *   2. 第 5 层 (Auto-Compact) 需要 LLM, 通过 getMinimax().chat 注入
+   *   3. 如果 budgetGate 不超限, 5 层短路在前 3 层, 不会调 LLM → 零开销
+   *   4. 失败静默: 任何异常 → console.warn + 保留原 messageHistory
+   *
+   * onStream 广播: 跟 computeJudgmentGate 风格一致 (phase 事件供 UI timeline 显示)
+   */
+  private async maybeAutoCompact(
+    onStream?: (chunk: any) => void,
+    signal?: AbortSignal
+  ): Promise<void> {
+    if (this.messageHistory.length < 10) return;  // 历史太短, 不值得压
+
+    onStream?.({ type: 'status', content: '🗜️ 评估是否需要压缩上下文...', tool: 'compactor' });
+
+    // 注入 LLM (用 getMinimax().chat, 与 judgment 注入门 / ReAct 循环同一来源)
+    // 给 Context Collapse (虚拟投影) 和 Auto-Compact (摘要) 共用
+    const llm = getMinimax();
+    const llmChat = async (systemPrompt: string, userPrompt: string): Promise<string> => {
+      const r = await llm.chat(userPrompt, systemPrompt, signal);
+      return r.reply;
+    };
+
+    const { compactPipeline, isContextCollapseEnabled } = await import('../context-compaction/index.js');
+    const result = await compactPipeline(this.messageHistory as any, {
+      maxTokens: 8000,
+      llmChat,
+      collapseLlmChat: llmChat,  // P1.2: Context Collapse 投影也用同一 LLM
+      cacheScope: this.currentChannelId || 'default',
+    });
+
+    if (result.compacted && result.history.length < this.messageHistory.length) {
+      const saved = this.messageHistory.length - result.history.length;
+      const stagesApplied = result.stages.filter((s) => s.applied).map((s) => s.stage).join(' → ');
+      onStream?.({
+        type: 'status',
+        content: `🗜️ 上下文压缩: ${stagesApplied || 'no-op'} | 节省 ${saved} 条 (剩余 ${result.history.length}, collapse=${isContextCollapseEnabled() ? 'on' : 'off'})`,
+        tool: 'compactor',
+      });
+      // 关键: 第 4 层 (Context Collapse) 是读时投影 (非破坏)
+      //       第 5 层 (Auto-Compact) 是破坏性折叠
+      //       这里用 if-else 区分: collapse on → 仅 buildContext 用; collapse off → 真更新
+      if (isContextCollapseEnabled()) {
+        this.projectedHistory = result.history as Message[];  // buildContext 用
+        // messageHistory 不变 (非破坏)
+      } else {
+        this.messageHistory = result.history as Message[];  // 真破坏性更新
+        this.projectedHistory = null;
+      }
+    }
+  }
+
   private isFinalResponse(content: string): boolean {
     // 只有明确输出 <final gen> 才认为是最终回答
     return content.includes('<final gen>');
@@ -1755,6 +2173,9 @@ Workspace root folder: ${this.cwd}
       /\btool\s*=>\s*["'](\w+)["']/,
       // 兼容: [TOOL_CALL] 块内 JSON 形式 {"name": "x", "args": {...}}
       /\[TOOL_CALL\][\s\S]*?\{\s*"name"\s*:\s*"(\w+)"\s*,\s*"args"\s*:\s*(\{[\s\S]*?\})/i,
+      // 2026-06-15 修: 兼容 LLM 输出的 XML 格式 <tool_name>...<arg>value</arg>...</tool_name>
+      //   实际 LLM 习惯: <shell_exec>\n<command>ls</command>\n<args>["-la", "..."]</args>\n</shell_exec>
+      /<(\w+)>([\s\S]*?)<\/\1>/,
     ];
 
     for (const pattern of patterns) {
@@ -1779,6 +2200,18 @@ Workspace root folder: ${this.cwd}
               if (key) args[key] = valueParts.join(':') || '';
             }
           }
+        } else if (rawArgs && /<\w+>[\s\S]*<\/\w+>/.test(rawArgs)) {
+          // 2026-06-15 修: XML 格式, 解析内嵌子标签 <argname>value</argname>
+          //   例: <command>ls</command>\n<args>["-la","~/.bolloon/skills"]</args>
+          const xmlArgPattern = /<(\w+)>([\s\S]*?)<\/\1>/g;
+          let xmlMatch: RegExpExecArray | null;
+          while ((xmlMatch = xmlArgPattern.exec(rawArgs)) !== null) {
+            const argName = xmlMatch[1];
+            const argValue = xmlMatch[2].trim();
+            if (argName && argValue) {
+              args[argName] = argValue;
+            }
+          }
         } else if (rawArgs) {
           // 形参串, 形如 key: value, key2: value2
           const argPairs = rawArgs.split(',').map(s => s.trim()).filter(Boolean);