v0.17.1: 退出释放显存修正 + 备份完整性 + 核心逻辑测试补课 + 上下文逻辑收敛 + 记忆日志可读性

修复： - main.ts 退出释放模型显存改用 getSetting(serverUrl)，不再硬编码 127.0.0.1:11434（避免非默认地址时释放请求打到错误端口） - 备份导出/导入并入 localStorage 持久化状态（会话摘要、度量历史、轨迹降级缓存、主题），版本升级到 v2，实现完整备份 - 工具数量改为 getEnabledToolDefinitions().length 动态计算，删除写死"32 个"的硬编码 - 记忆日志区分操作来源：memory:write 透传 reason，标注"新增记忆/替换/删除/清空/TTL 衰减清理/访问统计写回（无新条目）"，避免"写了但看不到新记忆"的困惑可维护性： - 上下文压力逻辑收敛到统一 calculateContextStats，删除 getContextPressureLevel / getTrendAwareCompressThreshold 的重复实现 - 消除 validateToolArgs 同名碰撞（agent-engine 本地版改名 validateToolArgsQuick） - 子代理工具集改用 getEnabledToolDefinitions() 基线，跟随全局启用开关与 Plan 模式 - 抽取 html-utils.ts 纯函数模块（实体解码/HTML→文本/HTML→Markdown/拦截页检测/相关性评分），tool-handlers-system 净减约 190 行重复代码 - 统一静态导入（savePlanTracker/setPlanModeActive/collectDiagnostics/addWrittenFile） - console.* 使用处补充豁免说明（启动/退出/刷盘阶段无渲染进程可推送日志） - run_command 工具描述改为反映可配置执行模式测试： - 新增 7 个测试文件 + 扩展 2 个，共 273 个测试（原 34 → 273） - 覆盖 agent-engine / agent-safety / context-manager / tool-registry / result-formatter / tool-parsing / memory-service / crypto / build-context / html-utils / utils / tool-handlers-fs - 全部通过 npm run typecheck && npm test && npm run build
2026-08-26 15:02:47 +08:00
parent 0b172d30c0
commit b66945c8a7
32 changed files with 2082 additions and 326 deletions
@@ -0,0 +1,181 @@
+import { describe, it, expect, beforeEach } from 'vitest';
+import {
+  estimateTokens,
+  recordActualTokens,
+  scoreMessageImportance,
+  mergeConsecutiveMessages,
+  chooseCompressionStrategy,
+  getAdaptiveCompressThreshold,
+  shouldAutoCompress,
+  calculateContextStats,
+} from '../src/renderer/services/context-manager.js';
+import type { OllamaMessage } from '../src/renderer/types.js';
+
+describe('estimateTokens', () => {
+  it('空文本为 0', () => {
+    expect(estimateTokens('')).toBe(0);
+  });
+
+  it('估算中文与英文差异', () => {
+    const zh = estimateTokens('你好世界');
+    const en = estimateTokens('hello world');
+    expect(zh).toBeGreaterThan(0);
+    expect(en).toBeGreaterThan(0);
+    // 中文按 1.5 字/token，4 字约 2-3 token
+    expect(zh).toBeLessThanOrEqual(3);
+  });
+
+  it('校准样本不足时不应用比例（保持原始估算）', () => {
+    // 未调 recordActualTokens 前，校准样本为 0，原始估算
+    expect(estimateTokens('abc')).toBe(Math.ceil(3 / 4));
+  });
+});
+
+describe('getAdaptiveCompressThreshold', () => {
+  it('小上下文模型更早触发', () => {
+    expect(getAdaptiveCompressThreshold(4096)).toBe(0.55);
+  });
+  it('中上下文使用标准阈值', () => {
+    expect(getAdaptiveCompressThreshold(16384)).toBe(0.5);
+  });
+  it('大上下文稍晚触发', () => {
+    expect(getAdaptiveCompressThreshold(65536)).toBe(0.45);
+  });
+});
+
+describe('scoreMessageImportance', () => {
+  it('SOUL.md 与参考数据不可压缩（满 10 分）', () => {
+    const m: OllamaMessage = { role: 'system', content: '[SOUL.md]\nxxx' };
+    expect(scoreMessageImportance(m)).toBe(10);
+  });
+
+  it('含 REFERENCE_DATA 标记的满 10 分', () => {
+    const m: OllamaMessage = { role: 'system', content: '<<<REFERENCE_DATA_START>>>' };
+    expect(scoreMessageImportance(m)).toBe(10);
+  });
+
+  it('日期/环境消息满 10 分', () => {
+    expect(scoreMessageImportance({ role: 'system', content: '[日期] 2026年' })).toBe(10);
+    expect(scoreMessageImportance({ role: 'system', content: '[环境] 运行环境' })).toBe(10);
+  });
+
+  it('ephemeral 消息权重为 0（优先丢弃）', () => {
+    const m: OllamaMessage = { role: 'user', content: '临时提醒', ephemeral: true };
+    expect(scoreMessageImportance(m)).toBe(0);
+  });
+
+  it('用户消息高于默认权重', () => {
+    const user = scoreMessageImportance({ role: 'user', content: '普通用户消息' });
+    const assistant = scoreMessageImportance({ role: 'assistant', content: '普通助手消息' });
+    expect(user).toBeGreaterThan(assistant);
+  });
+
+  it('工具调用消息加分', () => {
+    const m: OllamaMessage = { role: 'assistant', content: '', tool_calls: [{ type: 'function', function: { name: 'read_file', arguments: {} } }] };
+    const base = scoreMessageImportance({ role: 'assistant', content: 'hello' });
+    expect(scoreMessageImportance(m)).toBeGreaterThan(base);
+  });
+});
+
+describe('mergeConsecutiveMessages', () => {
+  it('合并连续 user 消息', () => {
+    const msgs: OllamaMessage[] = [
+      { role: 'user', content: 'a' },
+      { role: 'user', content: 'b' },
+      { role: 'assistant', content: 's' },
+    ];
+    const out = mergeConsecutiveMessages(msgs);
+    expect(out).toHaveLength(2);
+    expect(out[0].content).toContain('a');
+    expect(out[0].content).toContain('b');
+  });
+
+  it('不合并 tool / system / ephemeral / compressed 消息', () => {
+    const msgs: OllamaMessage[] = [
+      { role: 'tool', content: 't1', tool_name: 'read_file' },
+      { role: 'tool', content: 't2', tool_name: 'read_file' },
+    ];
+    expect(mergeConsecutiveMessages(msgs)).toHaveLength(2);
+  });
+
+  it('超过 3000 字符限制时不合并', () => {
+    const long = 'x'.repeat(2000);
+    const msgs: OllamaMessage[] = [
+      { role: 'user', content: long },
+      { role: 'user', content: long },
+    ];
+    expect(mergeConsecutiveMessages(msgs)).toHaveLength(2);
+  });
+
+  it('空/单消息原样返回', () => {
+    expect(mergeConsecutiveMessages([])).toEqual([]);
+    expect(mergeConsecutiveMessages([{ role: 'user', content: 'a' }])).toHaveLength(1);
+  });
+});
+
+describe('chooseCompressionStrategy', () => {
+  const numCtx = 131072;
+
+  it('压力低且消息少时跳过压缩', () => {
+    const d = chooseCompressionStrategy([], numCtx, 'low');
+    expect(d.strategy).toBe('skip');
+  });
+
+  it('工具结果占比高且非 critical 时用 fast', () => {
+    const msgs: OllamaMessage[] = [
+      { role: 'tool', content: 'x'.repeat(50), tool_name: 'read_file' },
+      { role: 'tool', content: 'y'.repeat(50), tool_name: 'read_file' },
+      { role: 'user', content: 'q' },
+    ];
+    const d = chooseCompressionStrategy(msgs, numCtx, 'high');
+    expect(d.strategy).toBe('fast');
+  });
+
+  it('critical 压力用 llm', () => {
+    const d = chooseCompressionStrategy([{ role: 'user', content: 'x' }, { role: 'assistant', content: 'y' }, { role: 'user', content: 'z' }], numCtx, 'critical');
+    expect(d.strategy).toBe('llm');
+  });
+
+  it('中等压力用 medium', () => {
+    const d = chooseCompressionStrategy([{ role: 'user', content: 'x' }, { role: 'assistant', content: 'y' }], numCtx, 'medium');
+    expect(d.strategy).toBe('medium');
+  });
+});
+
+describe('shouldAutoCompress', () => {
+  it('超阈值触发', () => {
+    // 构造大量内容使 token 超 50% numCtx
+    const msgs: OllamaMessage[] = [];
+    for (let i = 0; i < 50; i++) msgs.push({ role: 'assistant', content: '内容'.repeat(400) });
+    expect(shouldAutoCompress(msgs, 8192)).toBe(true);
+  });
+
+  it('少量消息不触发', () => {
+    const msgs: OllamaMessage[] = [{ role: 'user', content: 'hello' }];
+    expect(shouldAutoCompress(msgs, 131072)).toBe(false);
+  });
+});
+
+describe('calculateContextStats', () => {
+  it('计算总 token 与使用率', () => {
+    const msgs: OllamaMessage[] = [{ role: 'user', content: 'hello world' }];
+    const stats = calculateContextStats(msgs, 131072);
+    expect(stats.totalTokens).toBeGreaterThan(0);
+    expect(stats.usageRatio).toBeGreaterThan(0);
+    expect(stats.usageRatio).toBeLessThan(0.01);
+    expect(stats.messageCount).toBe(1);
+  });
+
+  it('空消息列表给出低压力', () => {
+    const stats = calculateContextStats([], 131072);
+    expect(stats.pressureInfo.level).toBe('low');
+    expect(stats.compressDecision.shouldCompress).toBe(false);
+  });
+});
+
+// 校准记录后的估算比例（重置校准状态：通过重新导入不可行，这里仅验证不抛错）
+describe('recordActualTokens', () => {
+  it('记录实际 token 不抛错', () => {
+    expect(() => recordActualTokens(100, 50, 90, 'test-model')).not.toThrow();
+  });
+});