feat: v0.8.1 记忆深化 · 观测闭环 · 体验收口 — 窗口/输出上限全局单一配置 · 2478 用例全量回归 + E2E 冒烟
CI / 类型检查 + Lint + 单元测试 (push) Failing after 9m8s
CI / 全量测试 (Electron ABI) (push) Failing after 6m0s
CI / 产物编译验证 (push) Successful in 10m58s

硬性契约:删除代码中一切写死的上下文窗口与最大输出上限(含六家模型元信息
钳制与全部兜底值)——唯一合法来源是设置面板「上下文长度」(llm.contextWindow)
与「最大输出上限」(llm.maxTokens),跨 Provider/模型原样透传。

P0 正确性收口:
- 迁移 11/12(SCHEMA_VERSION 5):记忆表 embedding 列 + 分 Provider 窗口键清理
- 记忆生命周期接线:会话终态清理 working memory / episodic 90 天 TTL / access_count 回写
- 回放缓冲模块化 + 会话终态清理(杜绝 4MB/会话内存滞留)
- i18n 收口:主进程 main-locale(zh/en,ui.locale 热切换)+ 渲染层 17 处出层

P1 能力演进:
- 本地向量混合检索:0.6×向量余弦 + 0.4×TF-IDF,Ollama embeddings 首次投产,
  存量记忆惰性回填,嵌入不可用自动回退 TF-IDF
- MEMORY.md 维护闭环:固化去重消除截断盲区;两阶段维护(AI 建议 → 用户确认 →
  原子改写 + 语义记忆双轨同步 + 审计);>50KB 告警
- 可观测闭环:cacheTokens 引擎→前端透传(Token 面板命中率/成本行)+ 输入框
  上下文占用指示条
- MCP Prompts/Resources 对话可用:/mcp:{server}:{prompt} 与 @mcp:{server}:{uri}

P2 体验补全:
- 工具自定义策略(正则白/黑名单 + 频率 + 强制确认,热生效)
- 连续 ≥3 同类工具确认聚合为单弹框
- 会话消息游标分页(首屏 200 条向上翻页)
- 开机自启;Playwright + Electron E2E 冒烟(本地 mock LLM 零外联)

Review 回归修复:MCP 大小写失配 / 分页状态复位 / 清空=未配置语义(Number(null)=0
隐患)/ MEMORY.md 告警位置 / working_memories FK(迁移 13)/ 全局配置层废键清理;
附带根治权限加固启动时序、代理回环放行、safeStorage 降级、悬空 symlink 逃逸。

验证:typecheck/lint 0 问题;test:electron 2478/2478(0 跳过);E2E 2/2;
docs/v0.8.1-迭代实施清单.md 全项留档。
This commit is contained in:
2026-09-08 09:35:58 +08:00
parent 839860083f
commit 9b45c445bf
85 changed files with 5286 additions and 1158 deletions
+1 -1
View File
@@ -127,7 +127,7 @@ function makeCtx(overrides: Record<string, unknown> = {}) {
logSessionEnd: vi.fn(),
log: vi.fn(),
},
memoryManager: { search: vi.fn(() => []) },
memoryManager: { search: vi.fn(() => []), clearWorkingMemory: vi.fn() },
promptInjectionDefender: {
detect: vi.fn(() => ({
isInjection: false,
@@ -0,0 +1,80 @@
/**
* Replay Buffer 单元测试(v0.8.1 P0-3 模块化收口)
*
* 锁定有界缓冲契约:条数/字节双上限溢出丢最旧、truncated 标记、runId 记录、
* TERMINATED 保留 + 新内容兜底重置、终态清除(会话删除联动)。
*/
import { describe, it, expect, beforeEach } from 'vitest';
import {
appendReplay,
resetReplayBuffer,
markReplayTerminated,
clearReplayBuffer,
getReplayBufferData,
} from '../replay-buffer';
describe('replay-buffer — 有界回放缓冲(v0.8.1 P0-3', () => {
beforeEach(() => {
clearReplayBuffer('s1');
});
it('append + get:按序返回事件与 runId', () => {
appendReplay('s1', { channel: 'stateChange', payload: { runId: 'run_a' }, ts: 1 });
appendReplay('s1', { channel: 'streamEvent', payload: { type: 'text_delta' }, ts: 2 });
const data = getReplayBufferData('s1');
expect(data.events).toHaveLength(2);
expect(data.events[0].channel).toBe('stateChange');
expect(data.runId).toBe('run_a');
expect(data.truncated).toBe(false);
});
it('条数上限:超过 2000 条丢弃最旧并置 truncated', () => {
for (let i = 0; i < 2005; i++) {
appendReplay('s1', { channel: 'streamEvent', payload: { i }, ts: i });
}
const data = getReplayBufferData('s1');
expect(data.events.length).toBeLessThanOrEqual(2000);
expect(data.truncated).toBe(true);
// 最旧的已被丢弃
expect((data.events[0].payload as { i: number }).i).toBeGreaterThan(0);
});
it('字节上限:单条 1MB 的超大 payload 被丢弃并置 truncated', () => {
const big = 'x'.repeat(1024 * 1024);
for (let i = 0; i < 6; i++) {
appendReplay('s1', { channel: 'streamEvent', payload: { big }, ts: i });
}
const data = getReplayBufferData('s1');
expect(data.truncated).toBe(true);
// 缓冲字节被控制在 4MB 上限附近(至少留最后一条)
expect(data.events.length).toBeGreaterThanOrEqual(1);
expect(data.events.length).toBeLessThan(6);
});
it('TERMINATED 保留缓冲;新内容到达兜底重置', () => {
appendReplay('s1', { channel: 'stateChange', payload: { state: 'TERMINATED' }, ts: 1 });
markReplayTerminated('s1');
// 完成后仍可回放最终内容
expect(getReplayBufferData('s1').events).toHaveLength(1);
// 下一次 run 的新内容 → 缓冲重置
appendReplay('s1', { channel: 'streamEvent', payload: { type: 'text_delta' }, ts: 2 });
const data = getReplayBufferData('s1');
expect(data.events).toHaveLength(1);
expect(data.truncated).toBe(false);
});
it('INIT 重置:清除上一 run 的缓冲', () => {
appendReplay('s1', { channel: 'streamEvent', payload: { a: 1 }, ts: 1 });
resetReplayBuffer('s1');
expect(getReplayBufferData('s1').events).toHaveLength(0);
});
it('clearReplayBuffer:终态清除(会话删除联动)', () => {
appendReplay('s1', { channel: 'streamEvent', payload: { a: 1 }, ts: 1 });
clearReplayBuffer('s1');
const data = getReplayBufferData('s1');
expect(data.events).toHaveLength(0);
expect(data.runId).toBeNull();
});
});
@@ -115,7 +115,8 @@ describe('sessions 域 — 参数校验矩阵', () => {
expect(await getHandler('sessions:getMessages')(null, 42)).toEqual([]);
await getHandler('sessions:getMessages')(null, 's1');
expect(svc.getMessages).toHaveBeenCalledWith('s1');
// v0.8.1 P2-3: 未传选项时以 undefined 透传(全量语义向后兼容)
expect(svc.getMessages).toHaveBeenCalledWith('s1', undefined);
});
it('saveTrace 严格校验:traceSteps 必须为数组且 tokenUsage 必填', async () => {
@@ -149,6 +150,8 @@ describe('sessions:clearMessages — v0.7.2 A1 语义修正', () => {
function makeCtx(): IPCContext {
return {
sessionService: { clearMessages: vi.fn() },
// v0.8.1 P0-2: clearMessages 联动清理工作记忆
memoryManager: { clearWorkingMemory: vi.fn() },
} as unknown as IPCContext;
}
@@ -442,6 +445,8 @@ describe('sessions 域 — 补充校验矩阵', () => {
disposeEngine: vi.fn(),
},
confirmationHook: { forgetSession: vi.fn() },
// v0.8.1 P0-2: delete/purge/clearMessages 联动清理工作记忆
memoryManager: { clearWorkingMemory: vi.fn() },
};
return { ctx: raw as unknown as IPCContext, raw };
}