MetonaTeam/metona-ollama-desktop
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

v0.14.9: fix prompt injection, context compression, auto memory, icon path

Browse Source
This commit is contained in:
紫影233
2026-07-07 14:05:17 +08:00
parent 7c78a65a9c
commit e9e64eb9c8
16 changed files with 424 additions and 407 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/renderer/services/context-manager.ts
+59 -15
View File
@@ -12,16 +12,25 @@ import { logInfo, logWarn, logSuccess, logError } from './log-service.js';
/** 校准比例：actualTokens / estimatedTokens，基于 Ollama 返回的实际计数动态修正 */
let _tokenCalibrationRatio = 1.0;
let _calibrationSamples = 0;
let _calibrationModel = ''; // C8: 记录校准时的模型名
const MIN_CALIBRATION_SAMPLES = 3;
/**
* 记录 Ollama 返回的实际 token 计数，用于校准估算器。
* 在 agent-engine.ts 每轮流式完成后调用。
* C8: 模型切换时自动重置校准比例，避免不同 tokenizer 导致估算失真
* @param actualInputTokens Ollama 返回的 prompt_eval_count
* @param actualOutputTokens Ollama 返回的 eval_count
* @param estimatedTokens 本轮消息调用 estimateTokens 的合计值
* @param modelName 当前使用的模型名
*/
export function recordActualTokens(actualInputTokens: number, actualOutputTokens: number, estimatedCount: number): void {
export function recordActualTokens(actualInputTokens: number, actualOutputTokens: number, estimatedCount: number, modelName?: string): void {
// C8: 模型切换时重置校准
if (modelName && modelName !== _calibrationModel) {
_calibrationModel = modelName;
_calibrationRatio = 1.0;
_calibrationSamples = 0;
}
if (actualInputTokens <= 0 || estimatedCount <= 0) return;
const actualTotal = actualInputTokens + actualOutputTokens;
const sampleRatio = actualTotal / Math.max(1, estimatedCount);
@@ -68,9 +77,11 @@ export function scoreMessageImportance(msg: OllamaMessage): number {
let score = 5; // 默认中等
// SOUL.md / 日期 / 环境 消息永远不可压缩
if (msg.content?.startsWith('[SOUL.md]')) return 10;
if (msg.content?.includes('[SOUL.md]') || msg.content?.includes('<<<REFERENCE_DATA_START>>>')) return 10;
if (msg.content?.startsWith('[日期]')) return 10;
if (msg.content?.startsWith('[环境]')) return 10;
// 安全规则提示也不可压缩
if (msg.content?.includes('安全规则')) return 10;
// 角色权重
if (msg.role === 'user') score += 2; // 用户消息最重要
@@ -79,8 +90,8 @@ export function scoreMessageImportance(msg: OllamaMessage): number {
// ephemeral 临时消息 → 最低权重，优先丢弃
if (msg.ephemeral) return 0;
// 已压缩标记 → 低权重（避免压缩产物堆积）
if (msg.compressed) score = 1;
// 已压缩标记 → 中等权重（C6: 提高从 1 到 4，避免快速摘要被立即丢弃）
if (msg.compressed) score = 4;
const content = msg.content || '';
@@ -210,9 +221,16 @@ 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
*/
export function shouldAutoCompress(messages: OllamaMessage[], numCtx: number): boolean {
const totalTokens = estimateTokens(messages.map(m => m.content || '').join(''));
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;
if (m.images?.length) totalTokens += m.images.length * 100;
}
const threshold = numCtx * AUTO_COMPRESS_THRESHOLD;
return totalTokens > threshold;
}
@@ -257,9 +275,18 @@ export async function compressWithLLM(
const keepTail = options.keepTail ?? COMPRESS_KEEP_TAIL;
const maxSummaryTokens = options.maxSummaryTokens ?? 500;
// 分离 system 和非 system 消息
// C1: 分离 system 和非 system 消息，再细分不可压缩的 system 消息
const systemMsgs = messages.filter(m => m.role === 'system');
const nonSystemMsgs = messages.filter(m => m.role !== 'system');
const incompressibleSysMsgs: OllamaMessage[] = [];
const compressibleSysMsgs: OllamaMessage[] = [];
for (const m of systemMsgs) {
if (scoreMessageImportance(m) >= 10) {
incompressibleSysMsgs.push(m);
} else {
compressibleSysMsgs.push(m);
}
}
if (nonSystemMsgs.length <= keepHead + keepTail + 2) {
logInfo('上下文压缩: 消息太少，跳过压缩');
@@ -345,8 +372,9 @@ export async function compressWithLLM(
if (parsed.toolResults.length) parts.push(`🔧 工具结果: ${parsed.toolResults.join('；')}`);
// 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 = {
role: 'system',
role: 'user',
content: `📋 以下是对之前对话的摘要（已压缩 ${uncompressedMiddle.length} 条消息）：\n\n${parts.join('\n')}`,
compressed: true
};
@@ -354,11 +382,11 @@ export async function compressWithLLM(
// 保留已压缩的中间消息 + 新摘要
const alreadyCompressed = middle.filter(m => m.compressed);
// 合并所有 system 消息为一条，避免多条 system 消息触发 Qwen3 等模型 400 错误
// 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 = [
...systemMsgs.map(m => m.content || ''),
...incompressibleSysMsgs.map(m => m.content || ''),
...compressibleSysMsgs.map(m => m.content || ''),
...alreadyCompressed.filter(m => m.role === 'system').map(m => m.content || ''),
summaryMsg.content || '',
].filter(Boolean).join('\n\n');
const nonSystemCompressed = alreadyCompressed.filter(m => m.role !== 'system');
@@ -367,6 +395,7 @@ export async function compressWithLLM(
{ role: 'system', content: mergedSystemContent },
...head,
...nonSystemCompressed,
summaryMsg,
...tail
];
@@ -437,6 +466,7 @@ function createQuickSummary(messages: OllamaMessage[]): string {
/**
* 根据 token 限制裁剪消息。
* v6.0: 按重要性评分决定保留顺序——重要性低的优先被丢弃。
* C2: 始终保留最近 N 条消息（时间窗口保护），避免丢失关键上下文
*/
function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaMessage[] {
// 分离 system 和非 system 消息
@@ -445,8 +475,13 @@ function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaM
if (nonSystemMsgs.length <= 4) return messages; // 消息太少不裁剪
// C2: 最近 6 条消息始终保留（时间窗口保护）
const PROTECT_RECENT = 6;
const recentMsgs = nonSystemMsgs.slice(-PROTECT_RECENT);
const olderMsgs = nonSystemMsgs.slice(0, -PROTECT_RECENT);
// 计算每条消息的 token + 重要性
const scored = nonSystemMsgs.map(m => ({
const scored = olderMsgs.map(m => ({
msg: m,
tokens: estimateTokens(m.content || '') +
(m.images ? m.images.length * 100 : 0) +
@@ -456,7 +491,16 @@ function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaM
// system 消息的 token 消耗
const systemTokens = systemMsgs.reduce((sum, m) => sum + estimateTokens(m.content || ''), 0);
const availableTokens = maxTokens - systemTokens;
const recentTokens = recentMsgs.reduce((sum, m) =>
sum + estimateTokens(m.content || '') +
(m.images ? m.images.length * 100 : 0) +
(m.tool_calls ? m.tool_calls.length * 50 : 0), 0);
const availableTokens = maxTokens - systemTokens - recentTokens;
if (availableTokens <= 0) {
// 连 system + recent 都超了，只保留 system + recent
return [...systemMsgs, ...recentMsgs];
}
// 按重要性降序排列，取能装下的最大数量
scored.sort((a, b) => b.importance - a.importance);
@@ -469,13 +513,13 @@ function trimByTokenLimit(messages: OllamaMessage[], maxTokens: number): OllamaM
kept.add(i);
}
// 按原始顺序重建：system → 按重要性保留的非system
// 按原始顺序重建：system → 按重要性保留的旧消息 → 最近的 protected 消息
const result: OllamaMessage[] = [...systemMsgs];
// 使用原始顺序而非重要性顺序，保持对话的时间线
const keptSet = new Set(scored.filter((_, i) => kept.has(i)).map(s => s.msg));
for (const m of nonSystemMsgs) {
for (const m of olderMsgs) {
if (keptSet.has(m)) result.push(m);
}
result.push(...recentMsgs);
return result;
}