v0.15.0: Agent ReAct Loop 核心引擎深度审计与生产级增强 (R1-R50)
核心引擎健壮性(R1-R10)、上下文管理优化(R11-R20)、工具安全与验证(R21-R30)、UI渲染性能(R31-R40)、基础设施与监控(R41-R50)、版本号升级
This commit is contained in:
thzxx
@@ -7,6 +7,159 @@
|
||||
import type { OllamaMessage, OllamaStreamChunk } from '../types.js';
|
||||
import { logInfo, logWarn, logSuccess, logError } from './log-service.js';
|
||||
|
||||
// ── R12: 压缩去重 — 内容指纹追踪 ──
|
||||
|
||||
/** DJB2 哈希函数，用于消息内容指纹 */
|
||||
function djb2Hash(str: string): string {
|
||||
let hash = 5381;
|
||||
for (let i = 0; i < str.length; i++) {
|
||||
hash = ((hash << 5) + hash + str.charCodeAt(i)) & 0x7fffffff;
|
||||
}
|
||||
return hash.toString(36);
|
||||
}
|
||||
|
||||
/** 记录已压缩消息批次的内容指纹，避免重复压缩相同内容 */
|
||||
const _compressedContentHashes = new Set<string>();
|
||||
const MAX_COMPRESSED_HASHES = 100;
|
||||
|
||||
/** R12: 计算消息批次的内容指纹 */
|
||||
function computeMessagesHash(messages: OllamaMessage[]): string {
|
||||
const content = messages.map(m => `${m.role}:${(m.content || '').slice(0, 200)}`).join('|');
|
||||
return djb2Hash(content);
|
||||
}
|
||||
|
||||
/** R12: 检查消息批次是否已被压缩过 */
|
||||
function isAlreadyCompressed(messages: OllamaMessage[]): boolean {
|
||||
if (messages.length === 0) return true;
|
||||
const hash = computeMessagesHash(messages);
|
||||
return _compressedContentHashes.has(hash);
|
||||
}
|
||||
|
||||
/** R12: 记录已压缩的消息批次指纹 */
|
||||
function markAsCompressed(messages: OllamaMessage[]): void {
|
||||
const hash = computeMessagesHash(messages);
|
||||
_compressedContentHashes.add(hash);
|
||||
// LRU 式淘汰：超过上限时移除最早的
|
||||
if (_compressedContentHashes.size > MAX_COMPRESSED_HASHES) {
|
||||
const firstKey = _compressedContentHashes.values().next().value;
|
||||
if (firstKey) _compressedContentHashes.delete(firstKey);
|
||||
}
|
||||
}
|
||||
|
||||
// ── R18: 上下文使用率预测 ──
|
||||
|
||||
/** 上下文使用率趋势数据点 */
|
||||
interface TokenUsagePoint {
|
||||
turn: number;
|
||||
tokens: number;
|
||||
numCtx: number;
|
||||
timestamp: number;
|
||||
}
|
||||
|
||||
const _tokenUsageTrend: TokenUsagePoint[] = [];
|
||||
const MAX_TREND_POINTS = 20;
|
||||
let _currentTurn = 0;
|
||||
|
||||
/** R18: 上下文预警级别 */
|
||||
export type ContextWarningLevel = 'safe' | 'notice' | 'warning' | 'critical';
|
||||
|
||||
/** R18: 上下文预测结果 */
|
||||
export interface ContextPrediction {
|
||||
level: ContextWarningLevel;
|
||||
currentUsage: number; // 0-1
|
||||
predictedUsage: number; // 预测下一轮的使用率
|
||||
turnsToOverflow: number; // 预计多少轮后溢出（-1 表示不会溢出）
|
||||
message: string;
|
||||
}
|
||||
|
||||
/** R18: 记录当前轮次的 token 使用量 */
|
||||
export function recordTokenUsage(tokens: number, numCtx: number): void {
|
||||
_currentTurn++;
|
||||
_tokenUsageTrend.push({
|
||||
turn: _currentTurn,
|
||||
tokens,
|
||||
numCtx,
|
||||
timestamp: Date.now(),
|
||||
});
|
||||
if (_tokenUsageTrend.length > MAX_TREND_POINTS) {
|
||||
_tokenUsageTrend.shift();
|
||||
}
|
||||
}
|
||||
|
||||
/** R18: 预测上下文溢出风险 */
|
||||
export function predictContextOverflow(numCtx: number): ContextPrediction {
|
||||
const currentTokens = _tokenUsageTrend.length > 0
|
||||
? _tokenUsageTrend[_tokenUsageTrend.length - 1].tokens
|
||||
: 0;
|
||||
const currentUsage = currentTokens / numCtx;
|
||||
|
||||
// 至少需要 3 个数据点才能做线性回归预测
|
||||
if (_tokenUsageTrend.length < 3) {
|
||||
const level: ContextWarningLevel = currentUsage > 0.8 ? 'critical'
|
||||
: currentUsage > 0.6 ? 'warning'
|
||||
: currentUsage > 0.4 ? 'notice'
|
||||
: 'safe';
|
||||
return {
|
||||
level,
|
||||
currentUsage,
|
||||
predictedUsage: currentUsage,
|
||||
turnsToOverflow: -1,
|
||||
message: level === 'safe' ? '' : `当前上下文使用率 ${(currentUsage * 100).toFixed(0)}%`,
|
||||
};
|
||||
}
|
||||
|
||||
// 线性回归：y = ax + b，预测未来 token 增长
|
||||
const n = _tokenUsageTrend.length;
|
||||
const xs = _tokenUsageTrend.map(p => p.turn);
|
||||
const ys = _tokenUsageTrend.map(p => p.tokens);
|
||||
const xMean = xs.reduce((s, x) => s + x, 0) / n;
|
||||
const yMean = ys.reduce((s, y) => s + y, 0) / n;
|
||||
let numerator = 0, denominator = 0;
|
||||
for (let i = 0; i < n; i++) {
|
||||
numerator += (xs[i] - xMean) * (ys[i] - yMean);
|
||||
denominator += (xs[i] - xMean) ** 2;
|
||||
}
|
||||
const slope = denominator !== 0 ? numerator / denominator : 0;
|
||||
const intercept = yMean - slope * xMean;
|
||||
|
||||
// 预测下一轮
|
||||
const nextTurn = _currentTurn + 1;
|
||||
const predictedTokens = Math.max(0, slope * nextTurn + intercept);
|
||||
const predictedUsage = predictedTokens / numCtx;
|
||||
|
||||
// 计算预计多少轮后溢出
|
||||
let turnsToOverflow = -1;
|
||||
if (slope > 0) {
|
||||
turnsToOverflow = Math.ceil((numCtx - intercept) / slope - _currentTurn);
|
||||
if (turnsToOverflow < 0) turnsToOverflow = 0;
|
||||
}
|
||||
|
||||
// 确定预警级别
|
||||
const maxUsage = Math.max(currentUsage, predictedUsage);
|
||||
let level: ContextWarningLevel;
|
||||
let message = '';
|
||||
|
||||
if (maxUsage > 0.85 || turnsToOverflow === 0) {
|
||||
level = 'critical';
|
||||
message = `⚠️ 上下文即将溢出！当前 ${currentTokens}/${numCtx} tokens (${(currentUsage * 100).toFixed(0)}%)，预计 ${turnsToOverflow} 轮后溢出`;
|
||||
} else if (maxUsage > 0.7 || (turnsToOverflow >= 1 && turnsToOverflow <= 3)) {
|
||||
level = 'warning';
|
||||
message = `⚠️ 上下文使用率较高 (${(currentUsage * 100).toFixed(0)}%)，预计 ${turnsToOverflow} 轮后可能溢出，建议压缩`;
|
||||
} else if (maxUsage > 0.5) {
|
||||
level = 'notice';
|
||||
message = `上下文使用率 ${(currentUsage * 100).toFixed(0)}%，趋势正常`;
|
||||
} else {
|
||||
level = 'safe';
|
||||
}
|
||||
|
||||
return { level, currentUsage, predictedUsage, turnsToOverflow, message };
|
||||
}
|
||||
|
||||
/** R18: 获取 token 使用趋势数据（供调试用） */
|
||||
export function getTokenUsageTrend(): TokenUsagePoint[] {
|
||||
return [..._tokenUsageTrend];
|
||||
}
|
||||
|
||||
// ── Token 校准系统 ──
|
||||
|
||||
/** 校准比例：actualTokens / estimatedTokens，基于 Ollama 返回的实际计数动态修正 */
|
||||
@@ -67,6 +220,40 @@ export function getTokenCalibration(): { ratio: number; samples: number } {
|
||||
/** 自动压缩阈值：当消息 token 占 context window 比例超过此值时触发自动压缩 */
|
||||
export const AUTO_COMPRESS_THRESHOLD = 0.3;
|
||||
|
||||
/** R14: 自适应压缩阈值 — 根据模型上下文长度动态调整 */
|
||||
export function getAdaptiveCompressThreshold(numCtx: number): number {
|
||||
// 小上下文模型（<8K）：更早触发压缩（40%），留更多余量
|
||||
// 中等上下文（8K-32K）：标准阈值（30%）
|
||||
// 大上下文（>32K）：稍晚触发（25%），避免过于频繁压缩
|
||||
if (numCtx < 8192) return 0.4;
|
||||
if (numCtx > 32768) return 0.25;
|
||||
return AUTO_COMPRESS_THRESHOLD;
|
||||
}
|
||||
|
||||
/** R19: 压缩效果指标 */
|
||||
export interface CompressionMetrics {
|
||||
beforeMessages: number;
|
||||
afterMessages: number;
|
||||
beforeTokens: number;
|
||||
afterTokens: number;
|
||||
compressionRatio: number;
|
||||
timestamp: number;
|
||||
}
|
||||
const _compressionHistory: CompressionMetrics[] = [];
|
||||
const MAX_COMPRESSION_HISTORY = 20;
|
||||
|
||||
/** R19: 获取压缩历史指标 */
|
||||
export function getCompressionHistory(): CompressionMetrics[] {
|
||||
return [..._compressionHistory];
|
||||
}
|
||||
|
||||
/** R19: 获取平均压缩率 */
|
||||
export function getAverageCompressionRatio(): number {
|
||||
if (_compressionHistory.length === 0) return 1.0;
|
||||
const sum = _compressionHistory.reduce((s, m) => s + m.compressionRatio, 0);
|
||||
return sum / _compressionHistory.length;
|
||||
}
|
||||
|
||||
/** 压缩后保留首尾消息数 */
|
||||
const COMPRESS_KEEP_HEAD = 5;
|
||||
const COMPRESS_KEEP_TAIL = 8;
|
||||
@@ -104,9 +291,11 @@ export function scoreMessageImportance(msg: OllamaMessage): number {
|
||||
/完成|done|✓|success|成功|结果|result/i,
|
||||
/项目|project|工作空间|workspace|git|repo|仓库/i,
|
||||
];
|
||||
const lowValuePatterns = [/好的|明白|ok|知道了|嗯|哦|好/i,
|
||||
/继续|请|帮我|可以吗/i,
|
||||
/谢谢|感谢|不客气/i,
|
||||
// R10: 修复低价值模式误报 — 移除单字“好”（匹配几乎所有中文文本）
|
||||
// 仅匹配明确的短语回复，且仅对短消息（<100字）生效
|
||||
const lowValuePatterns = [/^(好的|明白|ok|知道了|嗯|哦|好[的呀吧]|收到)$/i,
|
||||
/^(继续|请继续|帮帮我|可以吗|行吗|好的谢谢)$/i,
|
||||
/^(谢谢|感谢|不客气|多谢|thanks?)$/i,
|
||||
];
|
||||
|
||||
for (const p of highValuePatterns) {
|
||||
@@ -119,6 +308,26 @@ export function scoreMessageImportance(msg: OllamaMessage): number {
|
||||
// 工具调用 → 高价值
|
||||
if (msg.tool_calls?.length) score += 2;
|
||||
|
||||
// R20: 工具结果类型感知 — 不同工具结果的价值不同
|
||||
if (msg.role === 'tool' && msg.tool_name) {
|
||||
// 写类工具结果：高价值（记录了操作结果）
|
||||
if (/write_file|edit_file|delete_file|create_directory|move_file|copy_file/.test(msg.tool_name)) {
|
||||
score += 2;
|
||||
}
|
||||
// 搜索类工具结果：中等价值
|
||||
if (/search_files|web_search/.test(msg.tool_name)) {
|
||||
score += 1;
|
||||
}
|
||||
// 读取类工具结果：中等价值
|
||||
if (/read_file|list_directory|tree/.test(msg.tool_name)) {
|
||||
score += 1;
|
||||
}
|
||||
// 时间/计算类工具结果：低价值（时效性强，很快过期）
|
||||
if (/datetime|random|uuid|hash/.test(msg.tool_name)) {
|
||||
score -= 2;
|
||||
}
|
||||
}
|
||||
|
||||
// 长度加分：长消息通常包含更多信息
|
||||
if (content.length > 500) score += 1;
|
||||
if (content.length > 2000) score += 1;
|
||||
@@ -142,6 +351,111 @@ export interface ContextBuildOptions {
|
||||
workspaceContext?: string;
|
||||
}
|
||||
|
||||
// ── R16: 增量摘要合并工具函数 ──
|
||||
|
||||
/** R16: 从已压缩消息中提取结构化摘要 */
|
||||
function extractStructuredSummary(compressedMsgs: OllamaMessage[]): StructuredSummary {
|
||||
const summary: StructuredSummary = {
|
||||
topics: [], decisions: [], pendingTasks: [], constraints: [], knowledge: [], toolResults: [],
|
||||
};
|
||||
for (const msg of compressedMsgs) {
|
||||
const content = msg.content || '';
|
||||
// 解析结构化摘要中的各部分
|
||||
const topicMatch = content.match(/📌 主题:\s*(.+)/);
|
||||
if (topicMatch) summary.topics.push(...topicMatch[1].split('；').filter(Boolean));
|
||||
const decisionMatch = content.match(/✅ 决策:\s*(.+)/);
|
||||
if (decisionMatch) summary.decisions.push(...decisionMatch[1].split('；').filter(Boolean));
|
||||
const pendingMatch = content.match(/⏳ 待办:\s*(.+)/);
|
||||
if (pendingMatch) summary.pendingTasks.push(...pendingMatch[1].split('；').filter(Boolean));
|
||||
const constraintMatch = content.match(/📏 约束:\s*(.+)/);
|
||||
if (constraintMatch) summary.constraints.push(...constraintMatch[1].split('；').filter(Boolean));
|
||||
const knowledgeMatch = content.match(/🧠 知识点:\s*(.+)/);
|
||||
if (knowledgeMatch) summary.knowledge.push(...knowledgeMatch[1].split('；').filter(Boolean));
|
||||
const toolMatch = content.match(/🔧 工具结果:\s*(.+)/);
|
||||
if (toolMatch) summary.toolResults.push(...toolMatch[1].split('；').filter(Boolean));
|
||||
}
|
||||
return summary;
|
||||
}
|
||||
|
||||
/** R16: 合并两个结构化摘要，去重并限制条目数 */
|
||||
function mergeSummaries(old_: StructuredSummary, new_: StructuredSummary): StructuredSummary {
|
||||
const mergeArrays = (oldArr: string[], newArr: string[], max: number): string[] => {
|
||||
// 合并、去重、限制数量（新摘要优先）
|
||||
const combined = [...new Set([...newArr, ...oldArr])];
|
||||
return combined.slice(0, max);
|
||||
};
|
||||
return {
|
||||
topics: mergeArrays(old_.topics, new_.topics, 4),
|
||||
decisions: mergeArrays(old_.decisions, new_.decisions, 3),
|
||||
pendingTasks: mergeArrays(old_.pendingTasks, new_.pendingTasks, 3),
|
||||
constraints: mergeArrays(old_.constraints, new_.constraints, 3),
|
||||
knowledge: mergeArrays(old_.knowledge, new_.knowledge, 3),
|
||||
toolResults: mergeArrays(old_.toolResults, new_.toolResults, 3),
|
||||
};
|
||||
}
|
||||
|
||||
// ── R17: System 消息分区优化 ──
|
||||
|
||||
/** R17: 判断 system 消息是否为稳定前缀（不会在会话中改变） */
|
||||
function isStableSystemMessage(content: string): boolean {
|
||||
// SOUL.md、安全规则、日期、环境信息等 — 在整个会话中不会改变
|
||||
return content.includes('[SOUL.md]')
|
||||
|| content.includes('<<<REFERENCE_DATA_START>>>')
|
||||
|| content.startsWith('[日期]')
|
||||
|| content.startsWith('[环境]')
|
||||
|| content.includes('安全规则')
|
||||
|| content.includes('系统提示')
|
||||
|| content.includes('You are'); // 通用系统 prompt
|
||||
}
|
||||
|
||||
/** R17: 对 system 消息排序，稳定部分在前，动态部分在后，支持 LLM Prefix Caching */
|
||||
function reorderSystemMessagesForPrefixCaching(messages: OllamaMessage[]): void {
|
||||
// 找到所有 system 消息
|
||||
const systemIndices: number[] = [];
|
||||
for (let i = 0; i < messages.length; i++) {
|
||||
if (messages[i].role === 'system') systemIndices.push(i);
|
||||
}
|
||||
if (systemIndices.length <= 1) return;
|
||||
|
||||
// 提取并分类
|
||||
const stableContents: string[] = [];
|
||||
const dynamicContents: string[] = [];
|
||||
for (const idx of systemIndices) {
|
||||
const content = messages[idx].content || '';
|
||||
if (isStableSystemMessage(content)) {
|
||||
stableContents.push(content);
|
||||
} else {
|
||||
dynamicContents.push(content);
|
||||
}
|
||||
}
|
||||
|
||||
if (stableContents.length === 0 || dynamicContents.length === 0) return;
|
||||
|
||||
// 合并稳定部分和动态部分
|
||||
const stableContent = stableContents.join('\n\n');
|
||||
const dynamicContent = dynamicContents.join('\n\n');
|
||||
|
||||
// 重写第一条 system 消息为稳定部分，其余 system 消息合并为动态部分
|
||||
const firstSysIdx = systemIndices[0];
|
||||
messages[firstSysIdx].content = stableContent;
|
||||
|
||||
// 将其余 system 消息合并为一条动态 system 消息，放在最后一条 system 消息位置
|
||||
const lastSysIdx = systemIndices[systemIndices.length - 1];
|
||||
if (firstSysIdx !== lastSysIdx) {
|
||||
messages[lastSysIdx].content = dynamicContent;
|
||||
// 删除中间的 system 消息
|
||||
const middleSysIndices = systemIndices.slice(1, -1);
|
||||
for (let i = middleSysIndices.length - 1; i >= 0; i--) {
|
||||
messages.splice(middleSysIndices[i], 1);
|
||||
}
|
||||
} else {
|
||||
// 只有一条 system 消息时，追加动态内容
|
||||
messages[firstSysIdx].content = stableContent + '\n\n' + dynamicContent;
|
||||
}
|
||||
|
||||
logInfo(`R17: System 消息分区完成 — 稳定前缀 ${estimateTokens(stableContent)} tokens, 动态部分 ${estimateTokens(dynamicContent)} tokens`);
|
||||
}
|
||||
|
||||
const DEFAULT_OPTIONS: Required<ContextBuildOptions> = {
|
||||
windowSize: 40,
|
||||
summaryBatchSize: 30,
|
||||
@@ -164,20 +478,29 @@ export function buildContext(
|
||||
const opts = { ...DEFAULT_OPTIONS, ...options };
|
||||
const result: OllamaMessage[] = [];
|
||||
|
||||
// 系统 prompt
|
||||
let systemContent = '';
|
||||
if (opts.memoryContext) {
|
||||
systemContent += opts.memoryContext + '\n\n';
|
||||
}
|
||||
if (opts.workspaceContext) {
|
||||
systemContent += opts.workspaceContext + '\n\n';
|
||||
}
|
||||
|
||||
// 提取已有的 system 消息
|
||||
// R17: System 消息分区 — 稳定部分在前，动态部分在后，支持 LLM Prefix Caching
|
||||
// 稳定部分：已有的 system 消息（SOUL.md、规则等，不随会话变化）
|
||||
// 动态部分：memoryContext、workspaceContext（每轮可能变化）
|
||||
const existingSystem = allMessages.filter(m => m.role === 'system');
|
||||
const stableParts: string[] = [];
|
||||
const dynamicParts: string[] = [];
|
||||
|
||||
for (const sys of existingSystem) {
|
||||
systemContent += sys.content + '\n';
|
||||
if (isStableSystemMessage(sys.content || '')) {
|
||||
stableParts.push(sys.content || '');
|
||||
} else {
|
||||
dynamicParts.push(sys.content || '');
|
||||
}
|
||||
}
|
||||
// memoryContext 和 workspaceContext 是动态的
|
||||
if (opts.memoryContext) dynamicParts.push(opts.memoryContext);
|
||||
if (opts.workspaceContext) dynamicParts.push(opts.workspaceContext);
|
||||
|
||||
let systemContent = '';
|
||||
// 稳定前缀优先
|
||||
if (stableParts.length > 0) systemContent += stableParts.join('\n\n') + '\n\n';
|
||||
// 动态部分在后
|
||||
if (dynamicParts.length > 0) systemContent += dynamicParts.join('\n\n') + '\n\n';
|
||||
|
||||
if (systemContent.trim()) {
|
||||
result.push({ role: 'system', content: systemContent.trim() });
|
||||
@@ -223,16 +546,23 @@ export function buildContext(
|
||||
* 判断是否需要自动压缩
|
||||
* 当总 token 数超过 context window 的 AUTO_COMPRESS_THRESHOLD 比例时返回 true
|
||||
* C3: 包含 tool_calls 和 images 的 token 开销
|
||||
* R13: tool_calls 开销按实际参数大小估算而非固定 50
|
||||
*/
|
||||
export function shouldAutoCompress(messages: OllamaMessage[], numCtx: number): boolean {
|
||||
let totalTokens = 0;
|
||||
for (const m of messages) {
|
||||
totalTokens += estimateTokens(m.content || '');
|
||||
// C3: tool_calls 和 images 也消耗大量 token
|
||||
if (m.tool_calls?.length) totalTokens += m.tool_calls.length * 50;
|
||||
// R13: tool_calls 开销按实际 JSON 参数大小估算
|
||||
if (m.tool_calls?.length) {
|
||||
for (const tc of m.tool_calls) {
|
||||
const argsSize = JSON.stringify(tc.function.arguments || {}).length;
|
||||
totalTokens += estimateTokens(tc.function.name) + Math.ceil(argsSize / 4) + 20; // 20 tokens overhead per call
|
||||
}
|
||||
}
|
||||
if (m.images?.length) totalTokens += m.images.length * 100;
|
||||
}
|
||||
const threshold = numCtx * AUTO_COMPRESS_THRESHOLD;
|
||||
// R14: 使用自适应压缩阈值
|
||||
const threshold = numCtx * getAdaptiveCompressThreshold(numCtx);
|
||||
return totalTokens > threshold;
|
||||
}
|
||||
|
||||
@@ -298,6 +628,28 @@ export async function compressWithLLM(
|
||||
const tail = nonSystemMsgs.slice(-keepTail);
|
||||
const middle = nonSystemMsgs.slice(keepHead, nonSystemMsgs.length - keepTail);
|
||||
|
||||
// R15: 对话轮次边界保护 — 调整 head/tail 切分点，避免在对话轮次中间切割
|
||||
// 如果 head 末尾是带 tool_calls 的 assistant，将后续 tool 消息也纳入 head
|
||||
if (head.length > 0 && head[head.length - 1].tool_calls?.length) {
|
||||
let extendIdx = 0;
|
||||
while (extendIdx < middle.length && middle[extendIdx].role === 'tool') {
|
||||
head.push(middle[extendIdx]);
|
||||
extendIdx++;
|
||||
}
|
||||
middle.splice(0, extendIdx);
|
||||
}
|
||||
// 如果 tail 开头是 tool 消息（无对应 assistant），向前扩展到包含 assistant
|
||||
if (tail.length > 0 && tail[0].role === 'tool') {
|
||||
let extendBack = middle.length - 1;
|
||||
while (extendBack >= 0 && middle[extendBack].role !== 'assistant') {
|
||||
extendBack--;
|
||||
}
|
||||
if (extendBack >= 0) {
|
||||
const moved = middle.splice(extendBack);
|
||||
tail.unshift(...moved);
|
||||
}
|
||||
}
|
||||
|
||||
// 过滤掉已经压缩过的消息（避免重复压缩）
|
||||
const uncompressedMiddle = middle.filter(m => !m.compressed);
|
||||
if (uncompressedMiddle.length === 0) {
|
||||
@@ -305,6 +657,12 @@ export async function compressWithLLM(
|
||||
return messages;
|
||||
}
|
||||
|
||||
// R12: 压缩去重 — 检查这批消息是否已被压缩过（内容指纹匹配）
|
||||
if (isAlreadyCompressed(uncompressedMiddle)) {
|
||||
logInfo('R12: 消息批次内容指纹匹配已压缩记录，跳过重复压缩');
|
||||
return messages;
|
||||
}
|
||||
|
||||
// 构建对话文本
|
||||
const conversationText = uncompressedMiddle.map(m => {
|
||||
const role = m.role === 'user' ? '用户' : 'AI';
|
||||
@@ -363,6 +721,14 @@ export async function compressWithLLM(
|
||||
parsed = { topics: [summaryJson.slice(0, 100)], decisions: [], pendingTasks: [], constraints: [], knowledge: [], toolResults: [] };
|
||||
}
|
||||
|
||||
// R16: 增量摘要合并 — 提取已有压缩摘要并与新摘要合并
|
||||
const alreadyCompressedMiddle = middle.filter(m => m.compressed && m.role !== 'system');
|
||||
if (alreadyCompressedMiddle.length > 0) {
|
||||
const oldSummary = extractStructuredSummary(alreadyCompressedMiddle);
|
||||
parsed = mergeSummaries(oldSummary, parsed);
|
||||
logInfo(`R16: 合并了 ${alreadyCompressedMiddle.length} 条旧摘要到新摘要`);
|
||||
}
|
||||
|
||||
// 构建结构化摘要消息文本
|
||||
const parts: string[] = [];
|
||||
if (parsed.topics.length) parts.push(`📌 主题: ${parsed.topics.join('；')}`);
|
||||
@@ -372,6 +738,18 @@ export async function compressWithLLM(
|
||||
if (parsed.knowledge.length) parts.push(`🧠 知识点: ${parsed.knowledge.join('；')}`);
|
||||
if (parsed.toolResults.length) parts.push(`🔧 工具结果: ${parsed.toolResults.join('；')}`);
|
||||
|
||||
// R11: 压缩质量验证 — 如果摘要为空或过短，回退到文本摘要
|
||||
if (parts.length === 0 || parts.join('').length < 50) {
|
||||
logWarn('R11: 压缩摘要质量不足，回退到文本摘要');
|
||||
const textSummary = uncompressedMiddle.map(m => {
|
||||
const role = m.role === 'user' ? '用户' : 'AI';
|
||||
const content = (m.content || '').slice(0, 200);
|
||||
return `${role}: ${content}`;
|
||||
}).join('\n').slice(0, 1000);
|
||||
parts.length = 0;
|
||||
parts.push(`📌 主题: ${textSummary}`);
|
||||
}
|
||||
|
||||
// 构建压缩后的摘要消息
|
||||
// C5: 使用 user role 而非 system role，避免部分模型拒绝多个 system 消息
|
||||
const summaryMsg: OllamaMessage = {
|
||||
@@ -380,8 +758,11 @@ export async function compressWithLLM(
|
||||
compressed: true
|
||||
};
|
||||
|
||||
// 保留已压缩的中间消息 + 新摘要
|
||||
const alreadyCompressed = middle.filter(m => m.compressed);
|
||||
// R16: 保留已压缩的中间消息（system 消息单独处理）+ 新摘要
|
||||
// R12: 标记这批消息为已压缩
|
||||
markAsCompressed(uncompressedMiddle);
|
||||
// R16: 已压缩的旧摘要已被合并到新摘要中，不再保留它们
|
||||
const alreadyCompressed = middle.filter(m => m.compressed && m.role === 'system');
|
||||
|
||||
// C1: 合并 system 消息为一条，但保留不可压缩的 system 消息完整内容
|
||||
const mergedSystemContent = [
|
||||
@@ -400,10 +781,29 @@ export async function compressWithLLM(
|
||||
...tail
|
||||
];
|
||||
|
||||
// R17: System 消息分区优化 — 将 system 消息按稳定性排序，支持 LLM Prefix Caching
|
||||
// 稳定部分（SOUL.md、规则等）放在最前面，动态部分（工作空间、记忆）放在后面
|
||||
// 这样 Ollama 可以缓存稳定前缀，只重新处理动态部分
|
||||
reorderSystemMessagesForPrefixCaching(result);
|
||||
|
||||
const beforeTokens = estimateTokens(messages.map(m => m.content || '').join(''));
|
||||
const afterTokens = estimateTokens(result.map(m => m.content || '').join(''));
|
||||
|
||||
logSuccess(`上下文压缩完成: ${messages.length} 条 → ${result.length} 条, tokens: ${beforeTokens} → ${afterTokens}`);
|
||||
// R19: 记录压缩指标
|
||||
const metrics: CompressionMetrics = {
|
||||
beforeMessages: messages.length,
|
||||
afterMessages: result.length,
|
||||
beforeTokens,
|
||||
afterTokens,
|
||||
compressionRatio: beforeTokens > 0 ? afterTokens / beforeTokens : 1.0,
|
||||
timestamp: Date.now(),
|
||||
};
|
||||
_compressionHistory.push(metrics);
|
||||
if (_compressionHistory.length > MAX_COMPRESSION_HISTORY) {
|
||||
_compressionHistory.shift();
|
||||
}
|
||||
|
||||
logSuccess(`上下文压缩完成: ${messages.length} 条 → ${result.length} 条, tokens: ${beforeTokens} → ${afterTokens} (压缩率: ${(metrics.compressionRatio * 100).toFixed(0)}%)`);
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user