feat: 升级至 v0.3.0 — 安全增强、死循环检测、六轮全面审计修复
大版本迭代,新增安全增强、Agent Loop 增强、UI/UX 增强,经六轮全面审计修复所有问题。
新增功能:
- OutputValidator 事实一致性检查 + 幻觉检测
- PromptInjectionDefender 语义级检测(指令性动词密度、角色边界、分隔符嵌套)
- DeadLoopError 死循环检测(连续3轮相同工具调用自动终止)
- PolicyEngine maxFrequency 滑动窗口频率限制
- MemoryManager TF-IDF 语义检索 + IDF 缓存原子替换
- 斜杠命令(/tool /memory /clear /export)+ 快捷键(Ctrl+B/J/Shift+F/N/[/])
- 专注模式(Ctrl+Shift+F)带面板状态快照保存/恢复
六轮审计修复(共修复 3 CRITICAL + 8 HIGH + 13 MEDIUM + 11 LOW):
CRITICAL:
- main.ts 传入 createAdapter 而非 reloadAdapter,导致 Provider 切换完全失效
- Ctrl+N/Ctrl+[/Ctrl+] 不同步 agent-store,导致消息发到错误会话
- engine.ts emit('error') 无监听器导致 DONE 事件丢失、前端卡死
HIGH:
- 死循环检测在工具执行之后(移入 PARSING 后 EXECUTING 前)
- deadLoop 事件前端未处理
- ConfirmationHook.clearPending 从未调用导致定时器泄漏
- engine.ts retry abort listener 未移除导致监听器堆积
- MCPManager JSON.parse 无 try-catch 导致初始化崩溃
- sendMessage 自动创建会话不同步 session-store
- setCurrentSession 竞态导致旧请求覆盖新会话数据
MEDIUM:
- toggleFocusMode 覆盖用户原有面板状态
- 正则检测可被常见词绕过
- 频率限制内存泄漏 + customPolicies 覆盖
- LIKE 回退转义未包含反斜杠
- OutputValidator 新功能未传入 toolResults/context
- MemoryConsolidator LLM 调用无超时保护
- AuditService 每次 log 都查询数据库
- browser-window-manager 超时后未停止页面加载
- output-validator URL 比较大小写敏感
- FileReader 无 onerror 导致 Promise 永久挂起
- /clear /export 不关闭斜杠菜单
- 专注模式下 toggleSidebar/toggleDetail 未恢复另一面板快照
LOW:
- abortPromise 事件监听器堆积
- RateLimitHook Map 内存泄漏
- memory.ts await 同步方法
- PolicyEngine 死代码清理
- Orchestrator sessionDepth 会话结束不清理
- workspace.service.ts 元数据插入边界问题
This commit is contained in:
@@ -37,6 +37,16 @@ import { MetonaStreamEventType, MetonaFinishReason } from '../types';
|
||||
import { estimateMessagesTokens } from '../utils/token-estimator';
|
||||
import log from 'electron-log';
|
||||
|
||||
/**
|
||||
* v0.3.0: 死循环错误 — 在 executeOneIterationStream 中抛出,主循环捕获后以 DEAD_LOOP 原因终止
|
||||
*/
|
||||
class DeadLoopError extends Error {
|
||||
constructor(message: string) {
|
||||
super(message);
|
||||
this.name = 'DeadLoopError';
|
||||
}
|
||||
}
|
||||
|
||||
const DEFAULT_CONFIG: AgentLoopConfig = {
|
||||
maxIterations: 20,
|
||||
timeoutMs: 120_000,
|
||||
@@ -74,6 +84,9 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
/** 正在进行的 run Promise(用于 run lock,防止并发 run 污染状态) */
|
||||
private currentRunPromise: Promise<AgentLoopOutput> | null = null;
|
||||
|
||||
/** v0.3.0: 工具调用签名历史(用于死循环检测) */
|
||||
private toolCallHistory: string[] = [];
|
||||
|
||||
constructor(
|
||||
config: Partial<AgentLoopConfig> = {},
|
||||
private adapter: IMetonaProviderAdapter,
|
||||
@@ -163,6 +176,8 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
this.totalTokens = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };
|
||||
this.abortController = new AbortController();
|
||||
this.eventSeq = 0;
|
||||
// v0.3.0: 重置工具调用历史(用于死循环检测)
|
||||
this.toolCallHistory = [];
|
||||
|
||||
try {
|
||||
await this.transitionTo(AgentLoopState.INIT);
|
||||
@@ -229,6 +244,9 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
return this.finish(TerminationReason.COMPLETED, step.thought?.content);
|
||||
}
|
||||
|
||||
// v0.3.0: 死循环检测已移入 executeOneIterationStream 的 PARSING 阶段后,
|
||||
// 通过抛出 DeadLoopError 在此处 catch 块捕获处理
|
||||
|
||||
// 执行工具并将结果加入消息
|
||||
if (step.toolResults) {
|
||||
for (const result of step.toolResults) {
|
||||
@@ -249,11 +267,17 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
return this.finish(TerminationReason.MAX_ITERATIONS);
|
||||
return this.finish(TerminationReason.TIMEOUT);
|
||||
} catch (error) {
|
||||
// v0.3.0: 捕获 DeadLoopError — 以 DEAD_LOOP 原因终止
|
||||
if (error instanceof DeadLoopError) {
|
||||
return this.finish(TerminationReason.DEAD_LOOP, error.message);
|
||||
}
|
||||
const errMsg = (error as Error).message;
|
||||
if (this.aborted || errMsg.includes('Aborted')) {
|
||||
return this.finish(TerminationReason.USER_INTERRUPT);
|
||||
}
|
||||
this.emit('error', { error: errMsg });
|
||||
// v0.3.0 修复: 不使用 emit('error') — Node EventEmitter 对无监听器的 'error' 事件会同步 throw,
|
||||
// 导致 finish() 被中断、DONE 事件丢失、前端卡死。改为日志记录,让 finish 正常执行
|
||||
log.error(`[AgentLoop] Run failed: ${errMsg}`);
|
||||
return this.finish(TerminationReason.ERROR, undefined, error as Error);
|
||||
}
|
||||
}
|
||||
@@ -424,6 +448,23 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
this.accumulateTokens(tokenUsage);
|
||||
}
|
||||
|
||||
// v0.3.0 修复: 死循环检测 — 在 PARSING 阶段完成后、EXECUTING 阶段开始前检测
|
||||
// 确保第3轮重复调用的副作用不会产生(工具尚未执行)
|
||||
if (step.toolCalls && step.toolCalls.length > 0) {
|
||||
if (this.detectDeadLoop(step.toolCalls)) {
|
||||
log.warn(`[AgentLoop] Dead loop detected at iteration ${this.currentIteration} (before tool execution)`);
|
||||
this.emit('deadLoop', {
|
||||
iteration: this.currentIteration,
|
||||
runId: this.runId,
|
||||
sessionId: this.currentSessionId,
|
||||
});
|
||||
// 抛出特殊错误,主循环捕获后以 DEAD_LOOP 原因终止
|
||||
throw new DeadLoopError(
|
||||
`Detected a potential infinite loop: the same tool calls were repeated for 3 consecutive iterations. Please refine the approach or provide more specific instructions.`,
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
// === EXECUTING: 执行工具调用 ===
|
||||
if (step.toolCalls && step.toolCalls.length > 0) {
|
||||
// transitionTo 已发射 stateChange 事件,无需重复 emit
|
||||
@@ -456,16 +497,23 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
step.toolCalls.map((tc) => executeAndForward(tc)),
|
||||
);
|
||||
const controller = this.abortController;
|
||||
let onAbort: (() => void) | null = null;
|
||||
const abortPromise = new Promise<null>((resolve) => {
|
||||
if (this.aborted) return resolve(null);
|
||||
if (controller) {
|
||||
controller.signal.addEventListener('abort', () => resolve(null), { once: true });
|
||||
onAbort = () => resolve(null);
|
||||
controller.signal.addEventListener('abort', onAbort, { once: true });
|
||||
}
|
||||
});
|
||||
const raceResult = await Promise.race([toolsPromise, abortPromise]) as
|
||||
| PromiseSettledResult<MetonaToolResult>[]
|
||||
| null;
|
||||
|
||||
// v0.3.0 修复: 若 toolsPromise 先完成(正常执行),手动移除 abort 监听器,避免监听器堆积
|
||||
if (onAbort && controller && raceResult !== null) {
|
||||
controller.signal.removeEventListener('abort', onAbort);
|
||||
}
|
||||
|
||||
if (raceResult === null) {
|
||||
// 被 abort 中断,标记步骤并退出
|
||||
step.completedAt = Date.now();
|
||||
@@ -543,6 +591,11 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
} catch (error) {
|
||||
step.completedAt = Date.now();
|
||||
step.state = AgentLoopState.TERMINATED;
|
||||
// v0.3.0 修复: DeadLoopError 抛出时,将 step 加入 iterations 数组,
|
||||
// 确保死循环轮的 LLM thought 内容不丢失(用户可观察 Agent 被终止前的最后思考)
|
||||
if (error instanceof DeadLoopError) {
|
||||
this.iterations.push(step);
|
||||
}
|
||||
throw error;
|
||||
}
|
||||
}
|
||||
@@ -665,19 +718,25 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
const waitMs = Math.max(500, delay + jitter);
|
||||
log.warn(`[AgentLoop] Retry ${attempt + 1}/${this.config.retryCount} after ${Math.round(waitMs)}ms: ${(error as Error).message}`);
|
||||
await new Promise((resolve, reject) => {
|
||||
const timer = setTimeout(resolve, waitMs);
|
||||
const timer = setTimeout(() => {
|
||||
// v0.3.0 修复: timer 先触发时移除 abort 监听器,避免监听器堆积
|
||||
if (onAbort && signal) signal.removeEventListener('abort', onAbort);
|
||||
resolve(undefined);
|
||||
}, waitMs);
|
||||
// 支持 abort 中断等待
|
||||
const signal = this.abortController?.signal;
|
||||
let onAbort: (() => void) | null = null;
|
||||
if (signal) {
|
||||
if (signal.aborted) {
|
||||
clearTimeout(timer);
|
||||
reject(new Error('Aborted'));
|
||||
return;
|
||||
}
|
||||
signal.addEventListener('abort', () => {
|
||||
onAbort = () => {
|
||||
clearTimeout(timer);
|
||||
reject(new Error('Aborted'));
|
||||
}, { once: true });
|
||||
};
|
||||
signal.addEventListener('abort', onAbort, { once: true });
|
||||
}
|
||||
});
|
||||
}
|
||||
@@ -728,6 +787,59 @@ export class AgentLoopEngine extends EventEmitter {
|
||||
return estimateMessagesTokens(messages);
|
||||
}
|
||||
|
||||
/**
|
||||
* v0.3.0: 死循环检测
|
||||
*
|
||||
* 检测策略:
|
||||
* 将每轮的工具调用序列化为签名字符串,检查最近3轮的签名是否完全相同。
|
||||
* 如果连续3轮使用完全相同的参数调用相同的工具,判定为死循环。
|
||||
*
|
||||
* v0.3.0 修复:
|
||||
* - 对 args 的键进行排序,避免 JSON.stringify 键顺序不一致导致漏报
|
||||
*
|
||||
* @param toolCalls 当前轮次的工具调用
|
||||
* @returns 是否检测到死循环
|
||||
*/
|
||||
private detectDeadLoop(toolCalls: MetonaToolCall[]): boolean {
|
||||
// 将当前轮次的工具调用序列化为签名
|
||||
// v0.3.0 修复:使用 stable stringify,对对象键排序,确保相同内容不同键顺序产生相同签名
|
||||
// v0.3.0 修复:添加 visited Set 防循环引用,深度上限防过度递归
|
||||
const stableStringify = (obj: unknown, visited: Set<unknown> = new Set(), depth = 0): string => {
|
||||
if (depth > 10) return '...'; // 深度上限防止过度递归
|
||||
if (obj === null || typeof obj !== 'object') return JSON.stringify(obj);
|
||||
if (visited.has(obj)) return '"[Circular]"'; // 循环引用防护
|
||||
visited.add(obj);
|
||||
try {
|
||||
if (Array.isArray(obj)) return `[${obj.map((v) => stableStringify(v, visited, depth + 1)).join(',')}]`;
|
||||
const keys = Object.keys(obj as Record<string, unknown>).sort();
|
||||
return `{${keys.map((k) => `${JSON.stringify(k)}:${stableStringify((obj as Record<string, unknown>)[k], visited, depth + 1)}`).join(',')}}`;
|
||||
} finally {
|
||||
visited.delete(obj);
|
||||
}
|
||||
};
|
||||
const signature = toolCalls
|
||||
.map((tc) => `${tc.name}(${stableStringify(tc.args)})`)
|
||||
.join('|');
|
||||
|
||||
this.toolCallHistory.push(signature);
|
||||
|
||||
// 只保留最近5轮的记录(足够检测3轮重复,同时避免内存增长)
|
||||
if (this.toolCallHistory.length > 5) {
|
||||
this.toolCallHistory.shift();
|
||||
}
|
||||
|
||||
// 需要至少3轮数据才能检测
|
||||
if (this.toolCallHistory.length < 3) return false;
|
||||
|
||||
const len = this.toolCallHistory.length;
|
||||
const r1 = this.toolCallHistory[len - 1]; // 当前轮
|
||||
const r2 = this.toolCallHistory[len - 2]; // 上一轮
|
||||
const r3 = this.toolCallHistory[len - 3]; // 上上一轮
|
||||
|
||||
// 连续3轮完全相同 → 死循环
|
||||
return r1 === r2 && r2 === r3;
|
||||
}
|
||||
|
||||
/**
|
||||
* 上下文压缩 — 将旧消息摘要为一条 system 消息,保留最近 3 轮完整对话
|
||||
*
|
||||
|
||||
@@ -25,6 +25,8 @@ export enum TerminationReason {
|
||||
TIMEOUT = 'timeout',
|
||||
USER_INTERRUPT = 'user_interrupt',
|
||||
ERROR = 'error',
|
||||
/** v0.3.0: 检测到死循环(连续3轮相同工具调用) */
|
||||
DEAD_LOOP = 'dead_loop',
|
||||
}
|
||||
|
||||
// ===== 迭代步骤 =====
|
||||
|
||||
@@ -2,3 +2,6 @@ export type { PreToolHook, HookResult } from './pre-tool';
|
||||
export { PermissionCheckHook, RateLimitHook } from './pre-tool';
|
||||
export type { PostToolHook } from './post-tool';
|
||||
export { AuditLogHook, MemoryTriggerHook } from './post-tool';
|
||||
// v0.3.0: 修复 ConfirmationHook 未导出的问题
|
||||
export type { ConfirmationRequest } from './confirmation-hook';
|
||||
export { ConfirmationHook } from './confirmation-hook';
|
||||
|
||||
@@ -26,6 +26,10 @@ export class PermissionCheckHook implements PreToolHook {
|
||||
if (!result.authorized) {
|
||||
return { blocked: true, reason: result.reason };
|
||||
}
|
||||
// v0.3.0 修复: 授权成功后记录调用,使频率限制功能生效
|
||||
// 在授权检查通过后立即记录,即使后续工具执行失败也计入频率
|
||||
// 这样可以防止通过故意制造错误来绕过频率限制
|
||||
this.policyEngine.recordCall(toolCall.name);
|
||||
return { blocked: false };
|
||||
}
|
||||
}
|
||||
@@ -40,6 +44,14 @@ export class RateLimitHook implements PreToolHook {
|
||||
// 使用 sessionId:toolName 作为 key,实现会话隔离的 per-tool 速率限制
|
||||
const key = `${sessionId}:${toolCall.name}`;
|
||||
const now = Date.now();
|
||||
|
||||
// v0.3.0 修复: 定期清理过期 entry,避免 Map 随会话累积无限增长
|
||||
if (this.callCounts.size > 1000) {
|
||||
for (const [k, v] of this.callCounts) {
|
||||
if (v.resetTime < now) this.callCounts.delete(k);
|
||||
}
|
||||
}
|
||||
|
||||
const entry = this.callCounts.get(key);
|
||||
if (entry && entry.resetTime >= now) {
|
||||
if (entry.count >= this.maxCallsPerMinute) {
|
||||
|
||||
@@ -218,7 +218,11 @@ export class MemoryConsolidator {
|
||||
};
|
||||
|
||||
try {
|
||||
const response = await this.adapter.chat(request);
|
||||
// v0.3.0 修复: 添加 30 秒超时保护,防止 LLM 响应缓慢导致 consolidator 任务挂起
|
||||
const timeoutPromise = new Promise<never>((_, reject) =>
|
||||
setTimeout(() => reject(new Error('LLM extraction timeout')), 30_000),
|
||||
);
|
||||
const response = await Promise.race([this.adapter.chat(request), timeoutPromise]);
|
||||
return response.content.trim();
|
||||
} catch (error) {
|
||||
log.warn('[MemoryConsolidator] LLM call failed:', (error as Error).message);
|
||||
|
||||
@@ -130,6 +130,10 @@ export class MemoryManager {
|
||||
|
||||
/**
|
||||
* 更新 IDF 缓存
|
||||
*
|
||||
* v0.3.0 增强:
|
||||
* - 原子替换缓存(先构建新数据再替换,避免中间不一致状态)
|
||||
* - 错误处理(数据库查询失败时保留旧缓存,不更新时间戳)
|
||||
*/
|
||||
private updateIdfCache(): void {
|
||||
const now = Date.now();
|
||||
@@ -138,35 +142,46 @@ export class MemoryManager {
|
||||
}
|
||||
|
||||
const db = this.getDB();
|
||||
this.idfCache.clear();
|
||||
|
||||
// 获取所有记忆内容(episodic + semantic + working)
|
||||
const episodicRows = db.prepare('SELECT content, summary FROM episodic_memories').all() as Array<{ content: string; summary: string | null }>;
|
||||
const semanticRows = db.prepare('SELECT value FROM semantic_memories').all() as Array<{ value: string }>;
|
||||
const workingRows = db.prepare('SELECT value FROM working_memories').all() as Array<{ value: string }>;
|
||||
try {
|
||||
// v0.3.0: 先构建新缓存数据,再原子替换
|
||||
const newIdfCache = new Map<string, number>();
|
||||
|
||||
const allDocs = [
|
||||
...episodicRows.map((r) => r.content + ' ' + (r.summary ?? '')),
|
||||
...semanticRows.map((r) => r.value),
|
||||
...workingRows.map((r) => r.value),
|
||||
];
|
||||
// 获取所有记忆内容(episodic + semantic + working)
|
||||
const episodicRows = db.prepare('SELECT content, summary FROM episodic_memories').all() as Array<{ content: string; summary: string | null }>;
|
||||
const semanticRows = db.prepare('SELECT value FROM semantic_memories').all() as Array<{ value: string }>;
|
||||
const workingRows = db.prepare('SELECT value FROM working_memories').all() as Array<{ value: string }>;
|
||||
|
||||
this.cachedDocCount = allDocs.length;
|
||||
const docFreq = new Map<string, number>();
|
||||
const allDocs = [
|
||||
...episodicRows.map((r) => r.content + ' ' + (r.summary ?? '')),
|
||||
...semanticRows.map((r) => r.value),
|
||||
...workingRows.map((r) => r.value),
|
||||
];
|
||||
|
||||
for (const doc of allDocs) {
|
||||
const tokens = new Set(tokenize(doc));
|
||||
for (const token of tokens) {
|
||||
docFreq.set(token, (docFreq.get(token) ?? 0) + 1);
|
||||
const newDocCount = allDocs.length;
|
||||
const docFreq = new Map<string, number>();
|
||||
|
||||
for (const doc of allDocs) {
|
||||
const tokens = new Set(tokenize(doc));
|
||||
for (const token of tokens) {
|
||||
docFreq.set(token, (docFreq.get(token) ?? 0) + 1);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// IDF = log((N+1)/(df+1)) + 1(Sklearn 风格平滑),确保非负
|
||||
for (const [term, df] of docFreq) {
|
||||
this.idfCache.set(term, Math.log((this.cachedDocCount + 1) / (df + 1)) + 1);
|
||||
}
|
||||
// IDF = log((N+1)/(df+1)) + 1(Sklearn 风格平滑),确保非负
|
||||
for (const [term, df] of docFreq) {
|
||||
newIdfCache.set(term, Math.log((newDocCount + 1) / (df + 1)) + 1);
|
||||
}
|
||||
|
||||
this.cacheUpdatedAt = now;
|
||||
// v0.3.0: 原子替换 — 只有新数据完全准备好后才替换旧缓存
|
||||
this.idfCache = newIdfCache;
|
||||
this.cachedDocCount = newDocCount;
|
||||
this.cacheUpdatedAt = now;
|
||||
} catch (error) {
|
||||
// v0.3.0: 数据库查询失败时保留旧缓存,不更新 cacheUpdatedAt
|
||||
// 这样下次 search() 会再次尝试更新
|
||||
log.error('MemoryManager: Failed to update IDF cache, keeping stale cache:', error);
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -299,10 +314,15 @@ export class MemoryManager {
|
||||
|
||||
/**
|
||||
* 存储记忆
|
||||
*
|
||||
* v0.3.0 修复:
|
||||
* - switch 添加 default 分支,未知 type 抛错而非静默失败
|
||||
* - working 类型使用 item.id(若提供)或生成唯一 key,避免同 session 多次存储互相覆盖
|
||||
* - semantic 类型使用 item.summary 作为 key(若提供),支持更新已有记忆
|
||||
*/
|
||||
store(item: Omit<MemoryItem, 'id' | 'createdAt'>): string {
|
||||
store(item: Omit<MemoryItem, 'id' | 'createdAt'> & { id?: string }): string {
|
||||
const db = this.getDB();
|
||||
const id = `mem_${nanoid(12)}`;
|
||||
const id = item.id ?? `mem_${nanoid(12)}`;
|
||||
const now = Date.now();
|
||||
const importance = item.importance ?? this.calculateImportance(item);
|
||||
|
||||
@@ -314,17 +334,22 @@ export class MemoryManager {
|
||||
`).run(id, item.sessionId ?? null, item.content, item.summary ?? null, item.source, importance, now);
|
||||
break;
|
||||
case 'semantic':
|
||||
// v0.3.0 修复:使用 summary 作为 key(若提供),支持更新已有语义记忆
|
||||
db.prepare(`
|
||||
INSERT OR REPLACE INTO semantic_memories (id, key, value, category, confidence, source_session, created_at, updated_at, access_count)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, 0)
|
||||
`).run(id, id, item.content, 'general', importance, item.sessionId ?? null, now, now);
|
||||
`).run(id, item.summary ?? id, item.content, 'general', importance, item.sessionId ?? null, now, now);
|
||||
break;
|
||||
case 'working':
|
||||
// v0.3.0 修复:使用 summary 作为 key(若提供),避免硬编码 'default' 导致覆盖
|
||||
db.prepare(`
|
||||
INSERT OR REPLACE INTO working_memories (id, session_id, task_id, key, value, updated_at)
|
||||
VALUES (?, ?, ?, ?, ?, ?)
|
||||
`).run(id, item.sessionId ?? 'default', 'default', 'default', item.content, now);
|
||||
`).run(id, item.sessionId ?? 'default', 'default', item.summary ?? id, item.content, now);
|
||||
break;
|
||||
default:
|
||||
// v0.3.0 修复:未知 type 抛错而非静默失败
|
||||
throw new Error(`Unknown memory type: ${(item as { type: string }).type}`);
|
||||
}
|
||||
|
||||
// 使 IDF 缓存失效
|
||||
@@ -346,7 +371,8 @@ export class MemoryManager {
|
||||
search(query: string, options: MemorySearchOptions = {}): SearchResult[] {
|
||||
const db = this.getDB();
|
||||
const { topK = 5, type, minImportance = 0 } = options;
|
||||
if (!query) return [];
|
||||
// v0.3.0 修复:拦截空 query 和纯空格 query
|
||||
if (!query || !query.trim()) return [];
|
||||
|
||||
// v0.2.0: 优先使用 TF-IDF 语义搜索
|
||||
const tfidfResults = this.tfidfSearch(query, options);
|
||||
@@ -356,7 +382,8 @@ export class MemoryManager {
|
||||
|
||||
// 回退:如果 TF-IDF 没有结果(如 IDF 缓存为空),使用 LIKE 关键词搜索
|
||||
// 转义 LIKE 通配符,避免用户输入的 % 和 _ 影响匹配
|
||||
const escapedQuery = query.replace(/[%_]/g, '\\$&');
|
||||
// v0.3.0 修复: 反斜杠也需转义,否则含 \ 的搜索(如 Windows 路径)会导致 SQLite LIKE 报错
|
||||
const escapedQuery = query.replace(/[%_\\]/g, '\\$&');
|
||||
const pattern = `%${escapedQuery}%`;
|
||||
const results: SearchResult[] = [];
|
||||
|
||||
@@ -402,7 +429,8 @@ export class MemoryManager {
|
||||
}
|
||||
|
||||
// 搜索工作记忆
|
||||
if (type === 'working') {
|
||||
// v0.3.0 修复:LIKE 回退路径也需添加 !type 分支(与 tfidfSearch 保持一致)
|
||||
if (!type || type === 'working') {
|
||||
const rows = db.prepare(`
|
||||
SELECT * FROM working_memories
|
||||
WHERE (key LIKE ? ESCAPE '\\' OR value LIKE ? ESCAPE '\\')
|
||||
|
||||
@@ -164,7 +164,12 @@ export class TaskOrchestrator extends EventEmitter {
|
||||
handle.status = success ? 'completed' : 'error';
|
||||
handle.result = result;
|
||||
this.activeSubAgents.delete(taskId);
|
||||
this.sessionDepth.set(params.parentSessionId, currentDepth); // 恢复深度
|
||||
// v0.3.0 修复: 深度恢复为 0 时删除条目,避免 sessionDepth Map 随会话累积
|
||||
if (currentDepth === 0) {
|
||||
this.sessionDepth.delete(params.parentSessionId);
|
||||
} else {
|
||||
this.sessionDepth.set(params.parentSessionId, currentDepth);
|
||||
}
|
||||
this.emit('taskCompleted', result);
|
||||
|
||||
log.info(`[Orchestrator] SubAgent "${taskId}" (depth=${depth}) ${success ? 'completed' : 'failed'} in ${durationMs}ms, ${output.iterations.length} iterations`);
|
||||
@@ -184,7 +189,12 @@ export class TaskOrchestrator extends EventEmitter {
|
||||
handle.status = 'error';
|
||||
handle.result = result;
|
||||
this.activeSubAgents.delete(taskId);
|
||||
this.sessionDepth.set(params.parentSessionId, currentDepth); // 恢复深度
|
||||
// v0.3.0 修复: 深度恢复为 0 时删除条目,避免 sessionDepth Map 随会话累积
|
||||
if (currentDepth === 0) {
|
||||
this.sessionDepth.delete(params.parentSessionId);
|
||||
} else {
|
||||
this.sessionDepth.set(params.parentSessionId, currentDepth);
|
||||
}
|
||||
this.emit('taskError', { taskId, error: errMsg });
|
||||
|
||||
log.error(`[Orchestrator] SubAgent "${taskId}" (depth=${depth}) error: ${errMsg}`);
|
||||
|
||||
@@ -3,6 +3,11 @@
|
||||
*
|
||||
* 三级权限模型:Read / Write / External Action
|
||||
*
|
||||
* v0.3.0 增强:
|
||||
* - 实现 maxFrequency 频率限制(滑动窗口算法)
|
||||
* - 添加 checkFrequency 和 recordCall 方法
|
||||
* - 添加 cleanupFrequencyRecords 防止内存泄漏
|
||||
*
|
||||
* @see docs/生产级通用 AI Agent 智能体桌面应用:完整设计与构建指南.html — 第十章
|
||||
*/
|
||||
|
||||
@@ -22,12 +27,14 @@ export interface PermissionPolicy {
|
||||
}
|
||||
|
||||
export const DEFAULT_POLICIES: PermissionPolicy[] = [
|
||||
{ toolName: 'read_file', requiredLevel: PermissionLevel.READ, deniedPatterns: [/\/etc\//, /\/proc\//, /C:\\Windows\\/i, /C:\\System32\\/i, /MEMORY\.md/i] },
|
||||
// v0.3.0 修复:deniedPatterns 使用 (?:\/|["'\s,}]|$) 匹配,
|
||||
// 覆盖 /etc/ 和 /etc(无尾斜杠,在 JSON 字符串中后跟引号的情况)
|
||||
{ toolName: 'read_file', requiredLevel: PermissionLevel.READ, deniedPatterns: [/\/etc(?:\/|["'\s,}]|$)/, /\/proc(?:\/|["'\s,}]|$)/, /C:\\Windows\\/i, /C:\\System32\\/i, /MEMORY\.md/i] },
|
||||
{ toolName: 'web_search', requiredLevel: PermissionLevel.READ, maxFrequency: 10 },
|
||||
{ toolName: 'list_directory', requiredLevel: PermissionLevel.READ },
|
||||
{ toolName: 'search_files', requiredLevel: PermissionLevel.READ },
|
||||
{ toolName: 'memory_search', requiredLevel: PermissionLevel.READ },
|
||||
{ toolName: 'write_file', requiredLevel: PermissionLevel.WRITE, deniedPatterns: [/\/etc\//, /\/proc\//, /\/System\//, /C:\\Windows\\/i, /C:\\System32\\/i, /MEMORY\.md/i], requireConfirmation: true, maxFrequency: 5 },
|
||||
{ toolName: 'write_file', requiredLevel: PermissionLevel.WRITE, deniedPatterns: [/\/etc(?:\/|["'\s,}]|$)/, /\/proc(?:\/|["'\s,}]|$)/, /\/System(?:\/|["'\s,}]|$)/, /C:\\Windows\\/i, /C:\\System32\\/i, /MEMORY\.md/i], requireConfirmation: true, maxFrequency: 5 },
|
||||
{ toolName: 'memory_store', requiredLevel: PermissionLevel.WRITE },
|
||||
{ toolName: 'run_command', requiredLevel: PermissionLevel.EXTERNAL_ACTION, deniedPatterns: [/MEMORY\.md/i], requireConfirmation: true, maxFrequency: 3 },
|
||||
{ toolName: 'web_fetch', requiredLevel: PermissionLevel.READ },
|
||||
@@ -39,12 +46,30 @@ export const DEFAULT_POLICIES: PermissionPolicy[] = [
|
||||
export class PolicyEngine {
|
||||
private policies: Map<string, PermissionPolicy> = new Map();
|
||||
|
||||
/** v0.3.0: 工具调用频率追踪 — 工具名 -> 调用时间戳列表 */
|
||||
private callFrequency: Map<string, number[]> = new Map();
|
||||
|
||||
/** v0.3.0: 频率限制的时间窗口(1分钟 = 60秒) */
|
||||
private readonly FREQ_WINDOW_MS = 60_000;
|
||||
|
||||
/**
|
||||
* v0.3.0 修复:customPolicies 与 DEFAULT_POLICIES 合并而非完全覆盖
|
||||
*
|
||||
* 合并策略:customPolicies 中的字段覆盖默认策略的同名字段,
|
||||
* 未指定的字段保留默认值(如 deniedPatterns 等安全配置不会被丢失)
|
||||
*/
|
||||
constructor(customPolicies: PermissionPolicy[] = []) {
|
||||
for (const policy of DEFAULT_POLICIES) {
|
||||
this.policies.set(policy.toolName, policy);
|
||||
this.policies.set(policy.toolName, { ...policy });
|
||||
}
|
||||
for (const policy of customPolicies) {
|
||||
this.policies.set(policy.toolName, policy);
|
||||
const existing = this.policies.get(policy.toolName);
|
||||
if (existing) {
|
||||
// v0.3.0 修复:合并而非替换,保留默认的安全配置(如 deniedPatterns)
|
||||
this.policies.set(policy.toolName, { ...existing, ...policy });
|
||||
} else {
|
||||
this.policies.set(policy.toolName, policy);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -65,7 +90,14 @@ export class PolicyEngine {
|
||||
};
|
||||
}
|
||||
|
||||
const argsStr = JSON.stringify(args);
|
||||
// v0.3.0 修复:使用 try-catch 防止循环引用导致 JSON.stringify 抛错
|
||||
let argsStr: string;
|
||||
try {
|
||||
argsStr = JSON.stringify(args);
|
||||
} catch {
|
||||
// 循环引用等异常情况,降级为 toString
|
||||
argsStr = String(args);
|
||||
}
|
||||
|
||||
// v0.2.0: allowedPatterns 白名单校验 — 若定义了白名单,参数必须匹配其中之一
|
||||
if (policy.allowedPatterns && policy.allowedPatterns.length > 0) {
|
||||
@@ -99,10 +131,79 @@ export class PolicyEngine {
|
||||
}
|
||||
}
|
||||
|
||||
// v0.3.0: 频率限制检查
|
||||
if (policy.maxFrequency !== undefined) {
|
||||
const freqCheck = this.checkFrequency(toolName, policy.maxFrequency);
|
||||
if (!freqCheck.allowed) {
|
||||
return {
|
||||
authorized: false,
|
||||
reason: freqCheck.reason,
|
||||
level: policy.requiredLevel,
|
||||
requiresConfirmation: false,
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
return {
|
||||
authorized: true,
|
||||
level: policy.requiredLevel,
|
||||
requiresConfirmation: policy.requireConfirmation ?? false,
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* v0.3.0: 频率限制检查(滑动窗口算法)
|
||||
*
|
||||
* 检查指定工具在时间窗口内的调用次数是否超过限制。
|
||||
* 注意:此方法仅检查,不记录调用。调用成功后需调用 recordCall()。
|
||||
*
|
||||
* v0.3.0 修复:
|
||||
* - 将 validCalls 写回 Map,避免 callFrequency 数组无限增长(内存泄漏)
|
||||
*
|
||||
* @param toolName 工具名称
|
||||
* @param maxFreq 最大频率(每分钟)
|
||||
* @returns 检查结果
|
||||
*/
|
||||
checkFrequency(toolName: string, maxFreq?: number): { allowed: boolean; reason?: string } {
|
||||
const policy = this.policies.get(toolName);
|
||||
const limit = maxFreq ?? policy?.maxFrequency;
|
||||
if (limit === undefined) return { allowed: true };
|
||||
|
||||
const now = Date.now();
|
||||
const calls = this.callFrequency.get(toolName) ?? [];
|
||||
// 移除时间窗口外的调用记录
|
||||
const validCalls = calls.filter((t) => now - t < this.FREQ_WINDOW_MS);
|
||||
|
||||
// v0.3.0 修复:将清理后的 validCalls 写回 Map,避免数组无限增长
|
||||
if (validCalls.length !== calls.length) {
|
||||
this.callFrequency.set(toolName, validCalls);
|
||||
}
|
||||
|
||||
if (validCalls.length >= limit) {
|
||||
return {
|
||||
allowed: false,
|
||||
reason: `Rate limit exceeded for ${toolName}: max ${limit} calls per minute (current: ${validCalls.length})`,
|
||||
};
|
||||
}
|
||||
return { allowed: true };
|
||||
}
|
||||
|
||||
/**
|
||||
* v0.3.0: 记录工具调用(工具成功执行后调用)
|
||||
*
|
||||
* v0.3.0 修复:同时清理过期记录,防止数组无限增长
|
||||
*
|
||||
* @param toolName 工具名称
|
||||
*/
|
||||
recordCall(toolName: string): void {
|
||||
const now = Date.now();
|
||||
const calls = this.callFrequency.get(toolName) ?? [];
|
||||
// v0.3.0 修复:记录新调用时同时清理过期记录
|
||||
const validCalls = calls.filter((t) => now - t < this.FREQ_WINDOW_MS);
|
||||
validCalls.push(now);
|
||||
this.callFrequency.set(toolName, validCalls);
|
||||
}
|
||||
|
||||
// v0.3.0 修复: cleanupFrequencyRecords 已删除 — checkFrequency 和 recordCall 已做内联清理,
|
||||
// 该方法属于死代码,删除以减少维护负担
|
||||
}
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
*
|
||||
* 三层防御:
|
||||
* 1. 正则快速过滤
|
||||
* 2. 语义级检测(可选)
|
||||
* 2. 语义级检测
|
||||
* 3. 指令隔离标记
|
||||
*
|
||||
* v0.2.0 增强:
|
||||
@@ -13,6 +13,10 @@
|
||||
* - 添加角色扮演注入检测
|
||||
* - 添加间接注入检测(通过工具返回值注入)
|
||||
*
|
||||
* v0.3.0 增强:
|
||||
* - 实现语义级检测(指令性动词密度、角色边界异常、分隔符嵌套)
|
||||
* - 添加上下文感知的注入检测
|
||||
*
|
||||
* @see docs/生产级通用 AI Agent 智能体桌面应用:完整设计与构建指南.html — 第十章
|
||||
*/
|
||||
|
||||
@@ -37,16 +41,17 @@ interface InjectionPattern {
|
||||
export class PromptInjectionDefender {
|
||||
private static INJECTION_PATTERNS: InjectionPattern[] = [
|
||||
// === HIGH 危险:直接越狱/忽略指令 ===
|
||||
{ pattern: /ignore\s+(all\s+)?(previous|above|prior)\s+(instructions|commands|directives)/i, severity: 'high' },
|
||||
{ pattern: /forget\s+(everything|all\s+(of\s+)?(the|your)|above)/i, severity: 'high' },
|
||||
{ pattern: /override\s+your\s+/i, severity: 'high' },
|
||||
// v0.3.0 修复:在关键词间允许可选限定词(the/any/all/these/those/your 等),防止 "ignore the previous instructions" 绕过
|
||||
{ pattern: /ignore\s+(?:(?:the|all|any|these|those|your|above)\s+)*(?:previous|above|prior)\s+(?:(?:the|your|all)\s+)*(?:instructions?|commands?|directives?|rules?)/i, severity: 'high' },
|
||||
{ pattern: /forget\s+(?:(?:everything|all|the|your|above|prior)\s+)+/i, severity: 'high' },
|
||||
{ pattern: /override\s+(?:(?:the|your|all|any)\s+)+/i, severity: 'high' },
|
||||
{ pattern: /JAILBREAK/i, severity: 'high' },
|
||||
{ pattern: /DAN\s*[:\[]/i, severity: 'high' },
|
||||
{ pattern: /(?:enable|turn\s+on|activate)\s+(developer|debug|root|admin)\s+mode/i, severity: 'high' },
|
||||
{ pattern: /(?:disable|turn\s+off|bypass)\s+(your|the)\s+(safety|security|filter|guard|defense)/i, severity: 'high' },
|
||||
{ pattern: /(?:send|transmit|exfiltrate|upload)\s+(your|the|all)\s+(data|memory|context|secrets)/i, severity: 'high' },
|
||||
{ pattern: /(?:enable|turn\s+on|activate)\s+(?:(?:the|your)\s+)*(?:developer|debug|root|admin)\s+mode/i, severity: 'high' },
|
||||
{ pattern: /(?:disable|turn\s+off|bypass)\s+(?:(?:the|your|all|any)\s+)*(?:safety|security|filter|guard|defense|restrictions?)/i, severity: 'high' },
|
||||
{ pattern: /(?:send|transmit|exfiltrate|upload)\s+(?:(?:your|the|all|any)\s+)*(?:data|memory|context|secrets?)/i, severity: 'high' },
|
||||
{ pattern: /(?:you\s+are|act\s+as|pretend\s+to\s+be)\s+(?:a|an)\s+(?:unrestricted|unfiltered|unlimited|free)/i, severity: 'high' },
|
||||
{ pattern: /(?:in|under)\s+(?:developer|unrestricted|jailbreak|god)\s+mode/i, severity: 'high' },
|
||||
{ pattern: /(?:in|under)\s+(?:(?:the|a)\s+)*(?:developer|unrestricted|jailbreak|god)\s+mode/i, severity: 'high' },
|
||||
// 中文高危
|
||||
{ pattern: /忽略(?:以上|之前|前面|上面|所有)(?:的)?(?:指令|指示|命令|说明|规则)/i, severity: 'high' },
|
||||
{ pattern: /忘记(?:之前|上面|所有)(?:的)?(?:内容|对话|指令)/i, severity: 'high' },
|
||||
@@ -74,9 +79,10 @@ export class PromptInjectionDefender {
|
||||
{ pattern: /(?:恶意|危险|有害)(?:代码|命令|操作)/i, severity: 'medium' },
|
||||
|
||||
// === LOW 危险:分隔符/间接注入标记 ===
|
||||
{ pattern: /-{3,}\s*(system|user|assistant|instruction)/i, severity: 'low' },
|
||||
{ pattern: /<{3,}(system|instruction|prompt)/i, severity: 'low' },
|
||||
{ pattern: /\[{3,}(system|instruction|prompt)/i, severity: 'low' },
|
||||
// v0.3.0 修复:将 -{3,} 改为 -{2,},防止用 -- 绕过
|
||||
{ pattern: /-{2,}\s*(system|user|assistant|instruction)/i, severity: 'low' },
|
||||
{ pattern: /<{2,}(system|instruction|prompt)/i, severity: 'low' },
|
||||
{ pattern: /\[{2,}(system|instruction|prompt)/i, severity: 'low' },
|
||||
{ pattern: /\[SYSTEM\]/i, severity: 'low' },
|
||||
{ pattern: /\[INSTRUCTION\]/i, severity: 'low' },
|
||||
{ pattern: /\[ADMIN\]/i, severity: 'low' },
|
||||
@@ -88,6 +94,15 @@ export class PromptInjectionDefender {
|
||||
];
|
||||
|
||||
detect(input: string): InjectionDetectionResult {
|
||||
// v0.3.0 修复:添加 null/undefined 运行时防护
|
||||
if (!input || typeof input !== 'string') {
|
||||
return {
|
||||
isInjection: false,
|
||||
riskScore: 0,
|
||||
findings: [],
|
||||
recommendation: 'PASS: No injection patterns detected',
|
||||
};
|
||||
}
|
||||
const findings: InjectionFinding[] = [];
|
||||
let riskScore = 0;
|
||||
|
||||
@@ -111,17 +126,199 @@ export class PromptInjectionDefender {
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 语义级检测 — 基于上下文的注入检测
|
||||
*
|
||||
* v0.3.0 新增:
|
||||
* 在正则快速检测的基础上,增加三层语义分析:
|
||||
* 1. 指令性动词密度检测(异常高的命令密度可能是注入)
|
||||
* 2. 角色边界异常检测(用户消息中出现系统级指令模式)
|
||||
* 3. 结构化分隔符嵌套检测(多层分隔符嵌套是注入特征)
|
||||
*
|
||||
* @param input 待检测文本
|
||||
* @param context 可选的上下文信息(角色和内容),用于上下文感知检测
|
||||
*/
|
||||
detectSemantic(input: string, context?: { role: string; content: string }): InjectionDetectionResult {
|
||||
// v0.3.0 修复:添加 null/undefined 运行时防护
|
||||
if (!input || typeof input !== 'string') {
|
||||
return {
|
||||
isInjection: false,
|
||||
riskScore: 0,
|
||||
findings: [],
|
||||
recommendation: 'PASS: No injection patterns detected',
|
||||
};
|
||||
}
|
||||
const findings: InjectionFinding[] = [];
|
||||
let riskScore = 0;
|
||||
|
||||
// 1. 先执行正则快速检测
|
||||
const regexResult = this.detect(input);
|
||||
findings.push(...regexResult.findings);
|
||||
riskScore += regexResult.riskScore;
|
||||
|
||||
// 2. 语义级检测
|
||||
// 2a. 指令性动词密度检测
|
||||
const imperativeCheck = this.checkImperativeDensity(input);
|
||||
if (imperativeCheck.isSuspicious) {
|
||||
findings.push({
|
||||
pattern: 'semantic:imperative_density',
|
||||
matched: `${imperativeCheck.count} imperative verbs in ${input.length} chars`,
|
||||
severity: 'medium',
|
||||
});
|
||||
riskScore += 2;
|
||||
}
|
||||
|
||||
// 2b. 角色边界异常检测
|
||||
if (context) {
|
||||
const roleAnomaly = this.checkRoleBoundary(input, context);
|
||||
if (roleAnomaly) {
|
||||
findings.push(roleAnomaly);
|
||||
riskScore += 3;
|
||||
}
|
||||
}
|
||||
|
||||
// 2c. 结构化分隔符嵌套检测
|
||||
const nestedFinding = this.checkNestedDelimiters(input);
|
||||
if (nestedFinding) {
|
||||
findings.push(nestedFinding);
|
||||
riskScore += 2;
|
||||
}
|
||||
|
||||
// 2d. 隐式指令检测(通过疑问句伪装指令)
|
||||
const implicitFinding = this.checkImplicitInstructions(input);
|
||||
if (implicitFinding) {
|
||||
findings.push(implicitFinding);
|
||||
riskScore += 1;
|
||||
}
|
||||
|
||||
return {
|
||||
isInjection: findings.length > 0,
|
||||
riskScore: Math.min(10, riskScore),
|
||||
findings,
|
||||
recommendation: this.getRecommendation(Math.min(10, riskScore)),
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 指令性动词密度检测
|
||||
*
|
||||
* 正常用户输入中指令性动词密度较低,如果短时间内出现大量指令性动词,
|
||||
* 可能是注入攻击试图覆盖系统指令。
|
||||
*
|
||||
* v0.3.0 修复:提高阈值以降低对正常技术内容的误报
|
||||
* - 密度阈值从 3 提高到 5(每100字符)
|
||||
* - 最小数量从 4 提高到 6
|
||||
* - 只在输入长度 >= 50 字符时检测(短输入密度天然偏高)
|
||||
*/
|
||||
private checkImperativeDensity(input: string): { isSuspicious: boolean; count: number } {
|
||||
// v0.3.0 修复:短输入不检测密度(避免对简短技术问题误报)
|
||||
if (input.length < 50) return { isSuspicious: false, count: 0 };
|
||||
|
||||
const imperativePatterns = [
|
||||
/\b(?:do|execute|run|print|output|display|show|send|write|create|delete|remove|update|set|get|call|invoke|return|ignore|forget|override|disable|enable|activate|deactivate)\b/gi,
|
||||
/(?:执行|运行|打印|输出|显示|发送|写入|创建|删除|移除|更新|设置|获取|调用|返回|忽略|忘记|覆盖|禁用|启用|激活)/g,
|
||||
];
|
||||
|
||||
let count = 0;
|
||||
for (const pattern of imperativePatterns) {
|
||||
const matches = input.match(pattern);
|
||||
if (matches) count += matches.length;
|
||||
}
|
||||
|
||||
// v0.3.0 修复:提高阈值,降低误报
|
||||
const density = count / Math.max(1, input.length / 100);
|
||||
return { isSuspicious: density > 5 && count >= 6, count };
|
||||
}
|
||||
|
||||
/**
|
||||
* 角色边界异常检测
|
||||
*
|
||||
* 检测用户输入中是否出现了系统角色特有的指令模式,
|
||||
* 例如用户输入中声称自己是系统或助手。
|
||||
*/
|
||||
private checkRoleBoundary(input: string, context: { role: string; content: string }): InjectionFinding | null {
|
||||
// 用户输入中出现系统级声明
|
||||
const systemClaimPatterns = [
|
||||
/(?:I\s+am|as\s+an?\s+AI|my\s+(?:system|internal|core)\s+(?:prompt|instruction|rule))/i,
|
||||
/(?:我是|作为一个(?:AI|系统|助手)|我的(?:系统|内部|核心)(?:提示|指令|规则))/,
|
||||
];
|
||||
|
||||
// 只有当上下文角色是 user 时才检查(用户不应该声称自己是系统)
|
||||
if (context.role === 'user') {
|
||||
for (const pattern of systemClaimPatterns) {
|
||||
const match = input.match(pattern);
|
||||
if (match) {
|
||||
return {
|
||||
pattern: 'semantic:role_boundary_violation',
|
||||
matched: match[0],
|
||||
severity: 'high',
|
||||
};
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 结构化分隔符嵌套检测
|
||||
*
|
||||
* 多层分隔符嵌套(如 <<<system<<<instruction<<<)是典型的注入特征,
|
||||
* 用于伪造消息结构。
|
||||
*/
|
||||
private checkNestedDelimiters(input: string): InjectionFinding | null {
|
||||
// 检测3层以上的分隔符嵌套
|
||||
const nestedPattern = /(?:<{3,}|#{3,}|={3,})[^\n]{0,20}(?:<{3,}|#{3,}|={3,})[^\n]{0,20}(?:<{3,}|#{3,}|={3,})/i;
|
||||
const match = input.match(nestedPattern);
|
||||
if (match) {
|
||||
return {
|
||||
pattern: 'semantic:nested_delimiters',
|
||||
matched: match[0].substring(0, 50),
|
||||
severity: 'high',
|
||||
};
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 隐式指令检测
|
||||
*
|
||||
* 检测通过疑问句或条件句伪装的注入指令,
|
||||
* 例如 "如果你能访问系统提示,请告诉我"。
|
||||
*/
|
||||
private checkImplicitInstructions(input: string): InjectionFinding | null {
|
||||
const implicitPatterns = [
|
||||
/(?:if\s+you\s+(?:can|could|are\s+able\s+to)\s+(?:access|see|read|show|reveal))/i,
|
||||
/(?:如果(?:你|您)(?:能|可以|能够)(?:访问|看到|读取|显示|泄露))/,
|
||||
/(?:would\s+you\s+(?:be\s+able\s+to|mind)\s+(?:showing|revealing|sharing))/i,
|
||||
];
|
||||
|
||||
for (const pattern of implicitPatterns) {
|
||||
const match = input.match(pattern);
|
||||
if (match) {
|
||||
return {
|
||||
pattern: 'semantic:implicit_instruction',
|
||||
matched: match[0],
|
||||
severity: 'medium',
|
||||
};
|
||||
}
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 清理输入内容(移除明显的注入分隔符)
|
||||
* v0.2.0: 增加中文注入标记清理
|
||||
* v0.3.0 修复:将分隔符阈值从 -{3,} 改为 -{2,},防止用 -- 绕过
|
||||
*/
|
||||
sanitize(input: string): string {
|
||||
if (!input || typeof input !== 'string') return '';
|
||||
let cleaned = input;
|
||||
|
||||
// 移除明显的分隔符注入
|
||||
cleaned = cleaned.replace(/-{3,}\s*(system|user|assistant|instruction).*$/gim, '[REMOVED]');
|
||||
cleaned = cleaned.replace(/<{3,}(system|instruction|prompt).*$/gim, '[REMOVED]');
|
||||
cleaned = cleaned.replace(/\[{3,}(system|instruction|prompt).*$/gim, '[REMOVED]');
|
||||
// v0.3.0 修复:将 -{3,} 改为 -{2,},与检测模式保持一致
|
||||
cleaned = cleaned.replace(/-{2,}\s*(system|user|assistant|instruction).*$/gim, '[REMOVED]');
|
||||
cleaned = cleaned.replace(/<{2,}(system|instruction|prompt).*$/gim, '[REMOVED]');
|
||||
cleaned = cleaned.replace(/\[{2,}(system|instruction|prompt).*$/gim, '[REMOVED]');
|
||||
|
||||
// v0.2.0: 移除间接注入标记
|
||||
cleaned = cleaned.replace(/\[(SYSTEM|INSTRUCTION|ADMIN|OVERRIDE)\]/gi, '[REMOVED]');
|
||||
|
||||
@@ -335,6 +335,12 @@ export class BrowserWindowManager {
|
||||
win.loadURL(url),
|
||||
timeoutPromise,
|
||||
]);
|
||||
} catch (e) {
|
||||
// v0.3.0 修复: 超时后停止页面加载,避免后台继续消耗网络和 CPU 资源
|
||||
if (!win.isDestroyed()) {
|
||||
try { win.webContents.stop(); } catch { /* ignore */ }
|
||||
}
|
||||
throw e;
|
||||
} finally {
|
||||
if (timer) clearTimeout(timer);
|
||||
}
|
||||
|
||||
@@ -41,7 +41,8 @@ export class MemoryStoreTool implements IMetonaTool {
|
||||
const importance = (args.importance as number) ?? 0.5;
|
||||
const source = (args.source as 'user_input' | 'tool_result' | 'agent_thought' | 'imported') ?? 'agent_thought';
|
||||
|
||||
const id = await this.memoryManager.store({
|
||||
// v0.3.0 修复: store() 是同步方法,移除多余的 await 避免误导维护者
|
||||
const id = this.memoryManager.store({
|
||||
type,
|
||||
content,
|
||||
source,
|
||||
@@ -83,7 +84,8 @@ export class MemorySearchTool implements IMetonaTool {
|
||||
const topK = (args.topK as number) ?? 5;
|
||||
const threshold = (args.threshold as number) ?? 0.7;
|
||||
|
||||
const results = await this.memoryManager.search(query, {
|
||||
// v0.3.0 修复: search() 是同步方法,移除多余的 await 避免误导维护者
|
||||
const results = this.memoryManager.search(query, {
|
||||
topK,
|
||||
type,
|
||||
minImportance: threshold,
|
||||
|
||||
@@ -7,6 +7,10 @@
|
||||
* 3. 事实一致性检查(与工具结果是否矛盾)
|
||||
* 4. 幻觉检测(无根据的断言)
|
||||
*
|
||||
* v0.3.0 增强:
|
||||
* - 实现事实一致性检查(基于工具结果验证输出声明)
|
||||
* - 实现幻觉检测(检测上下文中无依据的具体声明)
|
||||
*
|
||||
* @see docs/生产级通用 AI Agent 智能体桌面应用:完整设计与构建指南.html — 第五章
|
||||
*/
|
||||
|
||||
@@ -22,10 +26,18 @@ export interface ValidationIssue {
|
||||
message: string;
|
||||
}
|
||||
|
||||
export interface ValidationOptions {
|
||||
/** 工具调用结果文本列表,用于事实一致性检查 */
|
||||
toolResults?: string[];
|
||||
/** 对话上下文文本,用于幻觉检测 */
|
||||
context?: string;
|
||||
}
|
||||
|
||||
/** 敏感词模式 */
|
||||
const SENSITIVE_PATTERNS = [
|
||||
{ pattern: /\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b/, type: 'credit_card', message: 'Possible credit card number detected' },
|
||||
{ pattern: /\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b/, type: 'email', message: 'Email address detected' },
|
||||
// v0.3.0 修复:移除字符类中的 | 字面字符([A-Z|a-z] → [A-Za-z])
|
||||
{ pattern: /\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}\b/, type: 'email', message: 'Email address detected' },
|
||||
{ pattern: /\b\d{3}[-.]?\d{3}[-.]?\d{4}\b/, type: 'phone', message: 'Phone number detected' },
|
||||
{ pattern: /\b(sk-[a-zA-Z0-9]{20,})\b/, type: 'api_key', message: 'Possible API key detected' },
|
||||
];
|
||||
@@ -40,8 +52,10 @@ const UNSAFE_PATTERNS = [
|
||||
export class OutputValidator {
|
||||
/**
|
||||
* 验证输出内容
|
||||
*
|
||||
* v0.3.0: 新增事实一致性检查和幻觉检测
|
||||
*/
|
||||
async validate(output: string): Promise<ValidationResult> {
|
||||
async validate(output: string, options?: ValidationOptions): Promise<ValidationResult> {
|
||||
const issues: ValidationIssue[] = [];
|
||||
|
||||
// 1. 格式验证
|
||||
@@ -50,12 +64,22 @@ export class OutputValidator {
|
||||
// 2. 内容安全检测
|
||||
issues.push(...this.checkSafety(output));
|
||||
|
||||
// 3. 空内容检查
|
||||
// 3. 事实一致性检查(v0.3.0 新增)
|
||||
if (options?.toolResults && options.toolResults.length > 0) {
|
||||
issues.push(...this.checkFactConsistency(output, options.toolResults));
|
||||
}
|
||||
|
||||
// 4. 幻觉检测(v0.3.0 新增)
|
||||
if (options?.context) {
|
||||
issues.push(...this.checkHallucination(output, options.context));
|
||||
}
|
||||
|
||||
// 5. 空内容检查
|
||||
if (!output.trim()) {
|
||||
issues.push({ severity: 'error', type: 'empty', message: 'Output is empty' });
|
||||
}
|
||||
|
||||
// 4. 过短内容检查
|
||||
// 6. 过短内容检查
|
||||
if (output.trim().length < 10 && output.trim().length > 0) {
|
||||
issues.push({ severity: 'warning', type: 'short', message: 'Output is suspiciously short' });
|
||||
}
|
||||
@@ -67,6 +91,173 @@ export class OutputValidator {
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 事实一致性检查 — 验证输出中的声明是否与工具结果矛盾
|
||||
*
|
||||
* v0.3.0 实现的检查策略:
|
||||
* 1. 工具报告错误但输出声称成功
|
||||
* 2. 工具报告文件不存在但输出引用文件内容
|
||||
* 3. 工具报告的数值与输出中的数值矛盾
|
||||
*
|
||||
* v0.3.0 修复:
|
||||
* - errorIndicators 改为更精确的匹配,避免常见词 "error" 误报
|
||||
* 使用 "error:" 或 "error occurred" 等上下文限定
|
||||
*/
|
||||
private checkFactConsistency(output: string, toolResults: string[]): ValidationIssue[] {
|
||||
const issues: ValidationIssue[] = [];
|
||||
const toolContext = toolResults.join('\n');
|
||||
|
||||
// 检查1:工具报告错误但输出声称成功
|
||||
// v0.3.0 修复:errorIndicators 改为更精确的匹配,避免 "error" 单独出现导致误报
|
||||
// 要求 error 后跟冒号、消息或特定错误模式
|
||||
const errorIndicators = /(?:error\s*[:\)]|error\s+occurred|failed\s+to|not\s+found|does\s+not\s+exist|enoent|permission\s+denied|cannot\s+access|no\s+such\s+file|exception|traceback|exit\s+code\s+[1-9])/i;
|
||||
const hasErrorInTools = errorIndicators.test(toolContext);
|
||||
|
||||
if (hasErrorInTools) {
|
||||
const successClaims = [
|
||||
/\b(?:successfully|success|done|completed|created|written|deleted|installed|updated|finished)\b/i,
|
||||
/(?:已(?:完成|创建|写入|删除|安装|更新|成功))/,
|
||||
];
|
||||
for (const pattern of successClaims) {
|
||||
const match = output.match(pattern);
|
||||
if (match) {
|
||||
issues.push({
|
||||
severity: 'warning',
|
||||
type: 'fact_inconsistency',
|
||||
message: `Output claims "${match[0]}" but tool results contain errors`,
|
||||
});
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 检查2:工具报告文件不存在但输出引用文件内容
|
||||
const notFoundPattern = /(?:not\s+found|no\s+such\s+file|does\s+not\s+exist|文件(?:不存在|未找到|找不到))/i;
|
||||
if (notFoundPattern.test(toolContext)) {
|
||||
const contentClaimPatterns = [
|
||||
/(?:file\s+(?:contains|says|shows)|文件(?:内容|包含|显示))/i,
|
||||
/(?:the\s+content\s+is|内容是)/i,
|
||||
];
|
||||
for (const pattern of contentClaimPatterns) {
|
||||
if (pattern.test(output)) {
|
||||
issues.push({
|
||||
severity: 'error',
|
||||
type: 'fact_inconsistency',
|
||||
message: 'Output references file content but tool reported file not found',
|
||||
});
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 检查3:工具报告退出码非零但输出声称命令成功
|
||||
const exitCodePattern = /exit\s+code[:\s]+(\d+)/i;
|
||||
const exitMatch = toolContext.match(exitCodePattern);
|
||||
if (exitMatch && parseInt(exitMatch[1], 10) !== 0) {
|
||||
const commandSuccessPattern = /\b(?:command\s+(?:succeeded|completed\s+successfully|ran\s+successfully))\b/i;
|
||||
if (commandSuccessPattern.test(output)) {
|
||||
issues.push({
|
||||
severity: 'error',
|
||||
type: 'fact_inconsistency',
|
||||
message: `Output claims command success but tool reported exit code ${exitMatch[1]}`,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
return issues;
|
||||
}
|
||||
|
||||
/**
|
||||
* 幻觉检测 — 检测输出中无依据的具体声明
|
||||
*
|
||||
* v0.3.0 实现的检查策略:
|
||||
* 1. 提取输出中的文件路径,检查是否在上下文中出现
|
||||
* 2. 提取输出中的 URL,检查是否在上下文中出现
|
||||
* 3. 检测虚构的 API 响应模式
|
||||
*
|
||||
* v0.3.0 修复:
|
||||
* - Windows 路径正则包含反斜杠字符类,正确匹配 C:\Users\test 等完整路径
|
||||
* - apiResponsePattern 在循环前重置 lastIndex,避免间歇性漏检
|
||||
* - 路径比较改为大小写不敏感(Windows 路径不敏感)
|
||||
*/
|
||||
private checkHallucination(output: string, context: string): ValidationIssue[] {
|
||||
const issues: ValidationIssue[] = [];
|
||||
|
||||
// 检查1:文件路径声称但上下文中不存在
|
||||
// v0.3.0 修复:Windows 路径正则 [A-Z]:\\[\w\\][\w.-]+ 不含反斜杠,改为 [A-Z]:\\[\w\\.-]+(?:\\[\w\\.-]+)*
|
||||
const pathPattern = /(?:\/[\w][\w.-]*){2,}|[A-Z]:\\[\w\\.-]+(?:\\[\w\\.-]+)*/g;
|
||||
const paths = output.match(pathPattern) || [];
|
||||
const seenPaths = new Set<string>();
|
||||
// v0.3.0 修复:上下文比较改为小写(Windows 路径大小写不敏感)
|
||||
const contextLower = context.toLowerCase();
|
||||
for (const path of paths) {
|
||||
if (seenPaths.has(path)) continue;
|
||||
seenPaths.add(path);
|
||||
|
||||
// 检查输出是否声称读取/访问了该路径
|
||||
const pathIndex = output.indexOf(path);
|
||||
const surrounding = output.substring(
|
||||
Math.max(0, pathIndex - 80),
|
||||
Math.min(output.length, pathIndex + path.length + 80),
|
||||
);
|
||||
const accessClaim = /(?:read|loaded|opened|accessed|found|exists|contains|读取|加载|打开|访问|存在|包含)/i;
|
||||
|
||||
// v0.3.0 修复:路径比较改为大小写不敏感
|
||||
if (accessClaim.test(surrounding) && !contextLower.includes(path.toLowerCase()) && path.length > 8) {
|
||||
issues.push({
|
||||
severity: 'warning',
|
||||
type: 'hallucination',
|
||||
message: `Path "${path.substring(0, 60)}" claimed in output but not found in context`,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
// 检查2:URL 声称但上下文中不存在
|
||||
const urlPattern = /https?:\/\/[^\s<>"]{8,}/g;
|
||||
const urls = output.match(urlPattern) || [];
|
||||
const seenUrls = new Set<string>();
|
||||
for (const url of urls) {
|
||||
if (seenUrls.has(url)) continue;
|
||||
seenUrls.add(url);
|
||||
|
||||
// URL 在上下文中不存在且输出声称访问了它
|
||||
const urlIndex = output.indexOf(url);
|
||||
const surrounding = output.substring(
|
||||
Math.max(0, urlIndex - 60),
|
||||
Math.min(output.length, urlIndex + url.length + 60),
|
||||
);
|
||||
const fetchClaim = /(?:fetched|retrieved|loaded|returned|visited|获取|抓取|访问|返回)/i;
|
||||
|
||||
// v0.3.0 修复: URL 比较改为大小写不敏感,因为 URL host 部分大小写不敏感
|
||||
if (fetchClaim.test(surrounding) && !context.toLowerCase().includes(url.toLowerCase())) {
|
||||
issues.push({
|
||||
severity: 'info',
|
||||
type: 'hallucination',
|
||||
message: `URL in output not found in conversation context`,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
// 检查3:检测虚构的 API 响应(声称返回了特定 JSON 但上下文中没有)
|
||||
const apiResponsePattern = /(?:API\s+(?:returned|responded)|接口(?:返回|响应))[:\s]*(\{[^}]{10,}\})/gi;
|
||||
// v0.3.0 修复:循环前重置 lastIndex,避免带 g flag 的正则在 break 后 lastIndex 未重置导致漏检
|
||||
apiResponsePattern.lastIndex = 0;
|
||||
let apiMatch;
|
||||
while ((apiMatch = apiResponsePattern.exec(output)) !== null) {
|
||||
const jsonSnippet = apiMatch[1];
|
||||
if (!context.includes(jsonSnippet.substring(0, 20))) {
|
||||
issues.push({
|
||||
severity: 'warning',
|
||||
type: 'hallucination',
|
||||
message: 'API response in output not found in tool results',
|
||||
});
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
return issues;
|
||||
}
|
||||
|
||||
/**
|
||||
* 格式验证
|
||||
*/
|
||||
|
||||
Reference in New Issue
Block a user