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
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:
@@ -178,7 +178,7 @@ describe('AnthropicAdapter — 请求体契约', () => {
|
||||
expect(toolResultBlocks[0].tool_use_id).toBe('tc_1');
|
||||
});
|
||||
|
||||
it('A2: max_tokens 按模型上限钳制（63488 → sonnet 64000 / opus 32000）', async () => {
|
||||
it('A2: max_tokens 原样透传（v0.8.1：模型钳制已废除，设置面板是唯一上限来源）', async () => {
|
||||
const sonnet = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://a.test',
|
||||
@@ -194,9 +194,9 @@ describe('AnthropicAdapter — 请求体契约', () => {
|
||||
const { bodies } = captureFetch();
|
||||
await sonnet.send(makeRequest());
|
||||
await opus.send(makeRequest());
|
||||
// 引擎默认 63488 低于 sonnet 上限 64000 → 原样保留；opus 上限 32000 → 钳制生效
|
||||
// v0.8.1: 设置面板「最大输出上限」对一切模型原样透传，无任何按模型钳制
|
||||
expect(bodies[0].max_tokens).toBe(63_488);
|
||||
expect(bodies[1].max_tokens).toBe(32_000);
|
||||
expect(bodies[1].max_tokens).toBe(63_488);
|
||||
});
|
||||
|
||||
it('A3: 小 maxTokens 时 thinking budget 不跌破协议下限 1024（v0.6.4 边界加固）', async () => {
|
||||
@@ -516,8 +516,8 @@ describe('AnthropicAdapter — thinking budget 按 effort 映射矩阵', () => {
|
||||
['low', 1024],
|
||||
['medium', 4096],
|
||||
['high', 16384],
|
||||
// max=32768 但 sonnet 的 max_tokens 先钳到 64000 → budget 二次钳到 floor(64000/2)=32000
|
||||
['max', 32000],
|
||||
// v0.8.1: max_tokens 不再按模型钳制（100_000 原样透传）→ budget = min(32768, floor(100000/2)) = 32768
|
||||
['max', 32768],
|
||||
] as const)('effort=%s → budget 为该档值且 < max_tokens', async (effort, expectBudget) => {
|
||||
const adapter = makeAdapter();
|
||||
const { bodies } = captureFetch();
|
||||
@@ -570,13 +570,13 @@ describe('AnthropicAdapter — thinking budget 按 effort 映射矩阵', () => {
|
||||
});
|
||||
});
|
||||
|
||||
describe('AnthropicAdapter — max_tokens 钳制矩阵', () => {
|
||||
describe('AnthropicAdapter — max_tokens 透传矩阵（v0.8.1 无钳制）', () => {
|
||||
it.each([
|
||||
['claude-sonnet-4-5', 63_488, 63_488], // 引擎默认低于上限 → 原样
|
||||
['claude-sonnet-4-5', 70_000, 64_000], // 超上限 → 钳到 sonnet 64000
|
||||
['claude-opus-4-1', 63_488, 32_000], // opus 上限 32000
|
||||
['claude-haiku-4-5', 63_488, 32_000], // haiku 上限 32000
|
||||
['claude-sonnet-4-5', 500, 500], // 低于上限 → 原样
|
||||
['claude-sonnet-4-5', 63_488, 63_488],
|
||||
['claude-sonnet-4-5', 70_000, 70_000], // 超过任何旧元信息上限 → 原样透传
|
||||
['claude-opus-4-1', 63_488, 63_488],
|
||||
['claude-haiku-4-5', 63_488, 63_488],
|
||||
['claude-sonnet-4-5', 500, 500],
|
||||
])('%s maxTokens=%d → max_tokens=%d', async (model, requested, expected) => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
@@ -795,7 +795,7 @@ describe('DeepSeekAdapter — thinking 映射矩阵', () => {
|
||||
expect(bodies[0].stop).toEqual(['<END>']);
|
||||
});
|
||||
|
||||
it('max_tokens 按模型钳制（pro 384000 / vision 8192）', async () => {
|
||||
it('max_tokens 原样透传（v0.8.1：pro/vision 均不钳制）', async () => {
|
||||
const pro = makeAdapter('deepseek-v4-pro');
|
||||
const vision = makeAdapter('deepseek-v4-flash-vision-exp');
|
||||
const { bodies } = captureFetch();
|
||||
@@ -803,8 +803,8 @@ describe('DeepSeekAdapter — thinking 映射矩阵', () => {
|
||||
await vision.send(
|
||||
makeRequest({ params: { maxTokens: 63_488, temperature: 0, stream: false } }),
|
||||
);
|
||||
expect(bodies[0].max_tokens).toBe(384_000);
|
||||
expect(bodies[1].max_tokens).toBe(8_192);
|
||||
expect(bodies[0].max_tokens).toBe(500_000);
|
||||
expect(bodies[1].max_tokens).toBe(63_488);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -864,8 +864,8 @@ describe('AgnesAdapter — enable_thinking 对称性矩阵', () => {
|
||||
makeRequest({ params: { maxTokens: 70_000, temperature: 0.9, stream: false } }),
|
||||
);
|
||||
expect(bodies[0].temperature).toBe(0.9);
|
||||
// 65536 上限钳制
|
||||
expect(bodies[0].max_tokens).toBe(65_536);
|
||||
// v0.8.1: 原样透传，无 65536 钳制
|
||||
expect(bodies[0].max_tokens).toBe(70_000);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -912,21 +912,21 @@ describe('MimoAdapter — thinking 显式开关', () => {
|
||||
expect(bodies[0].top_p).toBeUndefined();
|
||||
});
|
||||
|
||||
it('max_completion_tokens 钳制矩阵（pro 131072 / standard 32768）', async () => {
|
||||
it('max_completion_tokens 原样透传（v0.8.1：pro/standard 均不钳制）', async () => {
|
||||
const pro = makeAdapter({ defaultModel: 'mimo-v2.5-pro' });
|
||||
const std = makeAdapter({ defaultModel: 'mimo-v2.5' });
|
||||
const { bodies } = captureFetch();
|
||||
await pro.send(makeRequest({ params: { maxTokens: 200_000, temperature: 0, stream: false } }));
|
||||
await std.send(makeRequest({ params: { maxTokens: 63_488, temperature: 0, stream: false } }));
|
||||
expect(bodies[0].max_completion_tokens).toBe(131_072);
|
||||
expect(bodies[1].max_completion_tokens).toBe(32_768);
|
||||
expect(bodies[0].max_completion_tokens).toBe(200_000);
|
||||
expect(bodies[1].max_completion_tokens).toBe(63_488);
|
||||
});
|
||||
|
||||
it('thinking 未关闭时未配置 maxTokens → 兜底 32768（思考占配额，防截断）', async () => {
|
||||
it('thinking 未关闭时未配置 maxTokens → 不下发该字段（v0.8.1：无写死兜底值）', async () => {
|
||||
const pro = makeAdapter({ defaultModel: 'mimo-v2.5-pro' });
|
||||
const { bodies } = captureFetch();
|
||||
await pro.send(makeRequest({ params: { temperature: 0, stream: false } }));
|
||||
expect(bodies[0].max_completion_tokens).toBe(32_768);
|
||||
expect(bodies[0].max_completion_tokens).toBeUndefined();
|
||||
});
|
||||
|
||||
it('enableWebSearch 且存在客户端 tools → web_search 服务端工具追加（不覆盖客户端工具）', async () => {
|
||||
@@ -1029,27 +1029,36 @@ describe('OpenAIAdapter — 推理模型字段路由（v0.6.4 P3-1）', () => {
|
||||
expect(bodies[0].reasoning_effort).toBeUndefined();
|
||||
});
|
||||
|
||||
it('gpt-4.1 长上下文 1M → getContextWindow 返回 1M（模型元信息表）', () => {
|
||||
const adapter = makeAdapter('gpt-4.1');
|
||||
it('gpt-4.1 → getContextWindow 返回设置面板配置值（v0.8.1：元信息不再承载窗口）', () => {
|
||||
const adapter = new OpenAIAdapter({
|
||||
provider: 'openai',
|
||||
baseURL: 'http://o.test',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'gpt-4.1',
|
||||
contextWindow: 1_000_000,
|
||||
});
|
||||
expect(adapter.getContextWindow()).toBe(1_000_000);
|
||||
// 未配置 → 0（引擎跳过压缩判定），无任何写死兜底
|
||||
const noCfg = makeAdapter('gpt-4.1');
|
||||
expect(noCfg.getContextWindow()).toBe(0);
|
||||
});
|
||||
|
||||
it('o3-mini max_completion_tokens 钳制到 100000', async () => {
|
||||
it('o3-mini max_completion_tokens 原样透传（v0.8.1：无 100000 钳制）', async () => {
|
||||
const adapter = makeAdapter('o3-mini');
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({ params: { maxTokens: 200_000, temperature: 0, stream: false } }),
|
||||
);
|
||||
expect(bodies[0].max_completion_tokens).toBe(100_000);
|
||||
expect(bodies[0].max_completion_tokens).toBe(200_000);
|
||||
});
|
||||
|
||||
it('gpt-4o max_tokens 钳制到 16384', async () => {
|
||||
it('gpt-4o max_tokens 原样透传（v0.8.1：无 16384 钳制）', async () => {
|
||||
const adapter = makeAdapter('gpt-4o');
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({ params: { maxTokens: 63_488, temperature: 0, stream: false } }),
|
||||
);
|
||||
expect(bodies[0].max_tokens).toBe(16_384);
|
||||
expect(bodies[0].max_tokens).toBe(63_488);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -1255,16 +1264,16 @@ describe('OllamaAdapter — options 缺省与工具定义', () => {
|
||||
|
||||
// ===== 跨 Provider maxTokens 钳制矩阵 =====
|
||||
|
||||
describe('跨 Provider — maxTokens 钳制矩阵汇总', () => {
|
||||
describe('跨 Provider — maxTokens 透传矩阵汇总（v0.8.1 无钳制）', () => {
|
||||
it.each([
|
||||
['anthropic', 'claude-opus-4-1', 100_000, 32_000],
|
||||
['anthropic', 'claude-sonnet-4-5', 100_000, 64_000],
|
||||
['deepseek', 'deepseek-v4-flash-vision-exp', 100_000, 8_192],
|
||||
['agnes', 'agnes-2.0-flash', 100_000, 65_536],
|
||||
['mimo', 'mimo-v2.5', 100_000, 32_768],
|
||||
['openai', 'gpt-4o', 100_000, 16_384],
|
||||
['anthropic', 'claude-opus-4-1', 100_000, 100_000],
|
||||
['anthropic', 'claude-sonnet-4-5', 100_000, 100_000],
|
||||
['deepseek', 'deepseek-v4-flash-vision-exp', 100_000, 100_000],
|
||||
['agnes', 'agnes-2.0-flash', 100_000, 100_000],
|
||||
['mimo', 'mimo-v2.5', 100_000, 100_000],
|
||||
['openai', 'gpt-4o', 100_000, 100_000],
|
||||
] as const)(
|
||||
'%s %s maxTokens=100000 → 钳制为 %d',
|
||||
'%s %s maxTokens=100000 → 原样透传 %d',
|
||||
async (provider, model, requested, expected) => {
|
||||
const adapterMap: Record<string, unknown> = {
|
||||
anthropic: new AnthropicAdapter({
|
||||
|
||||
Reference in New Issue
Block a user