@modusensus/dsh-mneme 0.9.4 → 0.10.0

package/test/sleep-catchup.test.js

@@ -0,0 +1,315 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+import { createVectorIndex } from "../src/vector-index.js";
+import { createSleepScheduler, runSleep } from "../src/dream/sleep.js";
+
+// Issue #529（续排）：本轮被预算截断（conflict 相位还有没进这一轮的对）且送出去的每一对
+// 都落地了时，调度器不该等满一个 sleepMinIntervalHours（默认 8h）才跑下一轮——池子没清空，
+// 剩下的对只能等到下一次到点，等于把"分批"变成"隔夜慢慢分批"。
+//
+// 这一组用例锁的是续排语义，防的回归是：
+//   ① 判据拿 status 反推 → 全判 keep 的轮次 applied 恒为 0、status 是 noop（永远到不了 ok），
+//      续排在最该追的场景静默失效；
+//   ② 只把冷却项改成 0、不在回调里补挂闹钟 → 续排只追一轮就断，剩下的对又回到隔夜等；
+//   ③ 不限制连击 → 用户持续写入时池子进得比追得快，变成永不落地的每五分钟一轮，
+//      而审计上每轮都"有产出"，肉眼很难发现；
+//   ④ 失败轮也追 → 与连败退避抢同一只闹钟，退避被自己踩空。
+
+const embedder = {
+  embedSingle: async () => [1, 0, 0],
+  embed: async () => [1, 0, 0],
+  schedule: () => {},
+  modelHash: "mock#1",
+  dimension: 3
+};
+
+function mockCtx(responder) {
+  let call = 0;
+  return {
+    logger: { warn: () => {}, info: () => {} },
+    agentDefaultModel: { currentSelection: () => ({ provider: "mock", model: "sleep-model" }) },
+    llm: {
+      async *stream(options) {
+        const userText = options.messages.find((m) => m.role === "user")?.content?.[0]?.text ?? "";
+        let reply = "[]";
+        if (options.purpose === "sleep-conflict") reply = responder(userText, ++call);
+        yield { type: "text-delta", index: 0, text: reply };
+        yield { type: "finish", reason: { kind: "stop" } };
+      }
+    }
+  };
+}
+
+function setup() {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const vectorIndex = createVectorIndex({ store });
+  service.setEmbedder(embedder);
+  service.setVectorIndex(vectorIndex);
+  return { store, service, vectorIndex };
+}
+
+// 全部推到同一个方向附近：任意两条的 cosine 都 ≈ 1，findPotentialConflicts 生成 C(n,2) 对，
+// 去重阶段按相似度留下 n/2 对（每记忆至多一对）。
+function seedSimilarPairs(service, vectorIndex, n) {
+  for (let i = 0; i < n; i++) {
+    const m = service.saveWithDedupe({ type: "project", title: `主题${i}`, content: `内容${i}`, importance: 3 }).memory;
+    vectorIndex.saveEmbedding(m.id, [1, 0, 0]);
+  }
+}
+
+function baseConfig(overrides = {}) {
+  return {
+    sleepModeEnabled: true,
+    sleepIdleMinutes: 5,
+    sleepMinIntervalHours: 8,
+    sleepConflictStrictness: "normal",
+    sleepArchiveDays: 30,
+    sleepCompressDays: 90,
+    sleepPatternMinMemories: 100,
+    sleepMaxPatternPerRun: 3,
+    ...overrides
+  };
+}
+
+// 与 test/sleep-failure-backoff.test.js 同款房型：时钟与定时器全部注入。
+function makeHarness(overrides = {}) {
+  let nowMs = 10_000_000;
+  const timers = [];
+  let seq = 1;
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  let runs = 0;
+  let outcome = () => ({ ok: true, status: "noop" });
+  // 运行**期间**要发生的事（写入 / 耗时）——用来复现"跑完那一刻已挂着的表只剩几秒"。
+  let duringRun = null;
+  const sched = createSleepScheduler({
+    service,
+    config: { sleepModeEnabled: true, sleepIdleMinutes: 5, sleepMinIntervalHours: 1, ...overrides },
+    logger: { warn: () => {} },
+    onRun: async () => {
+      runs++;
+      if (duringRun) await duringRun();
+      return outcome();
+    },
+    now: () => nowMs,
+    setTimeoutFn: (fn, delay) => {
+      const t = { id: seq++, at: nowMs + delay, fn };
+      timers.push(t);
+      return t.id;
+    },
+    clearTimeoutFn: (id) => {
+      const i = timers.findIndex((t) => t.id === id);
+      if (i >= 0) timers.splice(i, 1);
+    }
+  });
+  const fireNext = async () => {
+    timers.sort((a, b) => a.at - b.at);
+    const t = timers.shift();
+    if (!t) throw new Error("no armed timer to fire");
+    nowMs = t.at;
+    await t.fn();
+  };
+  return {
+    sched,
+    store,
+    timers,
+    fireNext,
+    runs: () => runs,
+    now: () => nowMs,
+    setOutcome: (fn) => { outcome = fn; },
+    setDuringRun: (fn) => { duringRun = fn; },
+    write: () => { nowMs += 1000; sched.noteWrite(); },
+    advance: (ms) => { nowMs += ms; }
+  };
+}
+
+const MIN = 60_000;
+const HOUR = 3_600_000;
+
+test("issue#529 续排：clean + 还有没进这一轮的对 ⇒ 下一个闹钟钉在静默窗口，不等最小间隔", async () => {
+  const h = makeHarness();
+  h.setOutcome(() => ({ ok: false, status: "noop", catchup: true, backlog: true }));
+  h.write();
+  await h.fireNext();
+  assert.equal(h.runs(), 1, "静默到点起跑一轮");
+  assert.equal(h.timers.length, 1, "关键：追 backlog 的轮次跑完必须自己补挂闹钟（此前一条都不挂）");
+  const delay = h.timers[0].at - h.now();
+  assert.equal(delay, 5 * MIN + 1000, "下限是一个静默窗口，不是 1 秒也不是 1 小时");
+
+  await h.fireNext();
+  assert.equal(h.runs(), 2, "关键：闹钟真的起跑了第二轮（只补挂不起跑等于没追）");
+  h.store.close();
+});
+
+test("issue#529 续排：不 clean 时不追，间隔回到最小间隔", async () => {
+  const h = makeHarness();
+  h.setOutcome(() => ({ ok: true, status: "degraded", catchup: false, backlog: true }));
+  h.write();
+  await h.fireNext();
+  assert.equal(h.runs(), 1);
+  assert.equal(h.timers.length, 0, "有对没落地 ⇒ 不补挂短闹钟，等下一次写入重新起算");
+  const t = h.now();
+  assert.equal(h.sched.shouldRun(t + HOUR - 1), false, "冷却仍是基数间隔");
+  assert.equal(h.sched.shouldRun(t + HOUR), true);
+  h.store.close();
+});
+
+test("issue#529 续排：运行期间用户写入挂的表要在收尾时重挂，不能剩几秒就开下一批", async () => {
+  const h = makeHarness();
+  h.setOutcome(() => ({ ok: false, status: "noop", catchup: true, backlog: true }));
+  h.write();
+  // 运行一开始就按静默窗口挂好了表（at = 起跑时刻 + 5min + 1s）。
+  const armedAt = h.timers[0].at;
+  h.setDuringRun(async () => {
+    // 运行早期用户又写了一次：noteWrite 撤表重挂，新表从**那一刻**起算。
+    h.advance(MIN);
+    h.write();
+    // 这一轮本身跑了很久（LLM 往返 + 落库），早就越过那只表的到点时刻。真实环境里
+    // 这只表此刻已经可以立即触发——收尾若不重挂，下一批就是"几秒后开跑"，
+    // 「让库安静一个静默窗口再追」形同虚设。
+    h.advance(10 * MIN);
+  });
+  await h.fireNext();
+  assert.equal(h.runs(), 1);
+  assert.ok(h.now() > armedAt, "前置：这一轮确实跑过了原来那只表的到点时刻");
+  assert.equal(h.timers.length, 1, "收尾必须只剩（重挂的）一只表");
+  assert.equal(h.timers[0].at - h.now(), 5 * MIN + 1000, "关键：从**收尾这一刻**重算整一个静默窗口，不是沿用运行期挂的旧表");
+  h.store.close();
+});
+
+test("issue#529 续排：台账写入失败要让这一轮不 clean，不能每个静默窗口重问同一个冲突", async () => {
+  const { service, store, vectorIndex } = setup();
+  seedSimilarPairs(service, vectorIndex, 8);
+  // 留痕写不进去：这一对既没被记成"已审"，下一轮还会原样回到候选池。若仍判 clean，
+  // 续排会一个静默窗口问一次 LLM、直到连击上限，而审计上每轮都是干净的成功轮。
+  service.saveConflictReviewed = () => {
+    throw new Error("ledger unavailable");
+  };
+  const ctx = mockCtx(() => "[]");
+  const config = baseConfig({ sleepConflictMaxPairs: 1, sleepConflictMaxPairsPerRun: 2 });
+  const result = await runSleep(ctx, service, config, ctx.logger, { embedder, vectorIndex }, null);
+  assert.ok(result.phases.conflicts.ledgerFailed > 0, "前置：确实有留痕写入失败");
+  assert.equal(result.phases.conflicts.clean, false, "关键：没记上账就不算落地");
+  assert.equal(result.catchup, false, "不 clean ⇒ 不追下一批，回到正常节奏");
+  store.close();
+});
+
+test("issue#529 续排：失败轮不追（与连败退避互斥）", async () => {
+  const h = makeHarness({ sleepFailureBackoff: true });
+  h.setOutcome(() => ({ ok: false, status: "failed", catchup: true, backlog: true }));
+  h.write();
+  await h.fireNext();
+  // 失败本身仍然要补挂（#187：不补挂就永远等不到下一轮），但补的是**退避后的**长表，
+  // 不是续排那张静默窗口的短表——这才是"互斥"的可观测形态。
+  assert.equal(h.timers.length, 1, "失败轮按 #187 仍补挂一张表");
+  assert.equal(h.timers[0].at - h.now(), 2 * HOUR + 1000, "补的是 1h × 2^1 的退避表，不是 5 分钟的续排表");
+  const t = h.now();
+  assert.equal(h.sched.shouldRun(t + 2 * HOUR - 1), false, "退避仍然生效（1h 基数 × 2^1）");
+  assert.equal(h.sched.shouldRun(t + 2 * HOUR), true);
+  h.store.close();
+});
+
+test("issue#529 续排：连击到上限即收手，闹钟回到正常间隔", async () => {
+  const h = makeHarness();
+  // 用 runSleep 的真实返回形状：全判 keep 的一轮 status 是 noop，而 ok 只在
+  // ok/degraded 时为真——桩里写 ok:true 会让 maybeSchedule 走进与真实路径不同的分支
+  // （!ran 那条），于是「到顶后没有定时器」测的其实是桩的偏差，不是实现的行为。
+  h.setOutcome(() => ({ ok: false, status: "noop", catchup: true, backlog: true }));
+  h.write();
+  await h.fireNext();
+  // 24 是实现里的 CATCHUP_STREAK_CAP：每轮都 clean 但池子永远追不完时，必须自己刹住。
+  for (let i = 2; i <= 24; i++) {
+    assert.equal(h.timers.length, 1, `第 ${i} 轮之前应当有一只已挂好的闹钟`);
+    await h.fireNext();
+  }
+  assert.equal(h.runs(), 24, "连追 24 轮");
+  const t = h.now();
+  assert.equal(h.sched.shouldRun(t + 5 * MIN), false, "收手后回到最小间隔，不再按静默窗口放行");
+  assert.equal(h.sched.shouldRun(t + HOUR), true);
+  // 收手后既不能是「一直按静默窗口重排」（空转链），也不能是「什么都不挂」
+  // （#187：CD 到期成了没有监听者的时刻）。正确形态是回到正常的最小间隔。
+  assert.equal(h.timers.length, 1, "按 #187 补挂一张表，但不能是续排的短表");
+  assert.equal(h.timers[0].at - h.now(), HOUR + 1000, "补的是正常最小间隔，不是 5 分钟的续排表");
+  h.store.close();
+});
+
+test("issue#529 续排：全判 keep 的一轮 status=noop 仍是 clean，续排判据不依赖 status", async () => {
+  const { service, store, vectorIndex } = setup();
+  seedSimilarPairs(service, vectorIndex, 8);
+  const ctx = mockCtx(() => "[]");
+  const config = baseConfig({ sleepConflictMaxPairs: 1, sleepConflictMaxPairsPerRun: 2 });
+  const result = await runSleep(ctx, service, config, ctx.logger, { embedder, vectorIndex }, null);
+  assert.equal(result.phases.conflicts.applied, 0, "全判 keep ⇒ 没有任何落地");
+  assert.equal(result.phases.conflicts.status, "noop", "关键：这一轮永远到不了 ok（拿 status 当判据就死在这里）");
+  assert.ok(result.phases.conflicts.perRunSkipped > 0, "前置：本轮确实被预算截断");
+  assert.equal(result.phases.conflicts.clean, true, "送出去的每一对都留了台账 ⇒ clean");
+  assert.equal(result.catchup, true, "关键：noop 轮次也要能追");
+  assert.equal(result.backlog, true);
+  store.close();
+});
+
+test("issue#529 续排：坏批让这一轮不 clean，即使还有 backlog 也不追", async () => {
+  const { service, store, vectorIndex } = setup();
+  seedSimilarPairs(service, vectorIndex, 8);
+  let call = 0;
+  // 第一批吐空（stream 正常结束但没有 JSON 数组 ⇒ 这批 failed），后面的批正常。
+  const ctx = mockCtx(() => (++call === 1 ? "" : "[]"));
+  const config = baseConfig({ sleepConflictMaxPairs: 1, sleepConflictMaxPairsPerRun: 2 });
+  const result = await runSleep(ctx, service, config, ctx.logger, { embedder, vectorIndex }, null);
+  assert.equal(result.backlog, true, "前置：确实有对没进这一轮");
+  assert.equal(result.phases.conflicts.clean, false, "有批没走通 ⇒ 不 clean");
+  assert.equal(result.catchup, false, "关键：不 clean 时宁可少追一轮，也不要在失败状态里把间隔压到静默窗口");
+  store.close();
+});
+
+test("issue#529 续排：连击到顶收手之后，下一次正常触发要能重新起算（刹车不能变成永久闸）", async () => {
+  const h = makeHarness();
+  h.setOutcome(() => ({ ok: false, status: "noop", catchup: true, backlog: true }));
+  h.write();
+  await h.fireNext();
+  for (let i = 2; i <= 24; i++) await h.fireNext();
+  assert.equal(h.runs(), 24);
+  assert.equal(h.timers[0].at - h.now(), HOUR + 1000, "前置：到顶后已收手，闹钟回到正常间隔");
+
+  // 收手若只把 catchupWanted 置 false、**不把连击清零**，这个进程此后每一轮都会被同一道闸
+  // 挡下（streak 只增不减），"等下一个静默窗口重新起算"就成了一句空话。
+  h.write();
+  for (let i = 0; i < 5 && h.runs() < 25; i++) await h.fireNext();
+  assert.equal(h.runs(), 25, "关键：收手之后仍能被正常触发救回来");
+  assert.equal(h.timers.length, 1, "重新起算后又该补挂续排闹钟");
+  assert.equal(h.timers[0].at - h.now(), 5 * MIN + 1000, "补的是静默窗口的短表，说明连击已从 0 重新算");
+  h.store.close();
+});
+
+test("issue#529 续排：一轮抛错要把连击一并清零，不能把上一轮的进度带进下一次起算", async () => {
+  const h = makeHarness();
+  // 同上面两条：runSleep 在 status==="noop" 时给 ok:false，桩要照它写。
+  const clean = () => ({ ok: false, status: "noop", catchup: true, backlog: true });
+  h.setOutcome(clean);
+  h.write();
+  await h.fireNext();
+  // 先追到 23（上限 24 的前一格）：此时连击若不被清掉，下一轮 clean 就会立刻撞上限。
+  for (let i = 2; i <= 23; i++) await h.fireNext();
+  assert.equal(h.runs(), 23);
+
+  h.setOutcome(() => {
+    throw new Error("boom");
+  });
+  h.write();
+  await h.fireNext();
+  assert.equal(h.runs(), 24, "第 24 轮抛错");
+  assert.equal(h.timers.length, 1, "抛错仍按 #187 补挂一张表（补的是退避长表）");
+  assert.equal(h.timers[0].at - h.now(), HOUR + 1000, "抛错后回落到最小间隔，续排的短表不再挂");
+
+  // 关键：抛错回合若不把 catchupStreak 清零，接下来这一轮 clean 的连击会从 23 累到 24，
+  // 于是"刚被救回来"的进程立刻又被同一道闸挡下。
+  h.setOutcome(clean);
+  await h.fireNext();
+  assert.equal(h.runs(), 25, "抛错后的第一次干净轮次仍要真的跑");
+  assert.equal(h.timers.length, 1, "关键：连击已清零，这一轮又能挂上续排闹钟");
+  assert.equal(h.timers[0].at - h.now(), 5 * MIN + 1000);
+  h.store.close();
+});

package/test/sleep-failure-backoff.test.js

@@ -0,0 +1,156 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+import { createSleepScheduler } from "../src/dream/sleep.js";
+
+// Issue #529（C 项）：sleep 连续失败退避（sleepFailureBackoff，opt-in）。
+//
+// 报告里 C 的原案是「同一 snapshot_hash 连续失败就拉长间隔」，但 sleep 的
+// snapshot_hash 是**跑完之后**对 phases 结果算的哈希（dream/sleep.js runSleep 里
+// createHash(JSON.stringify(phases))），不是输入快照——同一批输入跑两次结果不同就
+// 得到不同哈希，判据不成立。这里改用与 #292（autoDreamFailureBackoff）同构的
+// 连续失败计数：有效间隔 = sleepMinIntervalHours × 2^连败数，成功清零，封顶 24h。
+//
+// 这些用例锁的是「退避语义」，防的回归是：
+//   ① 退避闸失效/写反 → 恒定失败的模型按 sleepMinIntervalHours 的固定节奏连发
+//      （#529 实测连续 23 轮 llm failed）；
+//   ② 成功不清零 → 模型恢复后仍被旧失败拖着推迟，sleep 变相停摆；
+//   ③ 24h 封顶失效 → 指数无限翻倍（8h 基数 × 2^5 = 10 天），睡眠维护彻底静默；
+//   ④ 默认关路径被顺手改掉 → 未 opt-in 用户的重试间隔漂移；
+//   ⑤ noop 被当成失败 → LLM 通路正常、只是库里没活干的库被越退越长。
+//
+// 时钟与定时器全部注入（同 test/sleep-scheduler-rearm.test.js 的房型）。
+
+function makeHarness(overrides = {}) {
+  let nowMs = 10_000_000;
+  const timers = [];
+  let seq = 1;
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  let runs = 0;
+  let outcome = () => ({ ok: false, status: "failed" });
+  const sched = createSleepScheduler({
+    service,
+    config: {
+      sleepModeEnabled: true,
+      sleepIdleMinutes: 5,
+      sleepMinIntervalHours: 1,
+      ...overrides
+    },
+    logger: { warn: () => {} },
+    onRun: async () => {
+      runs++;
+      return outcome();
+    },
+    now: () => nowMs,
+    setTimeoutFn: (fn, delay) => {
+      const t = { id: seq++, at: nowMs + delay, fn };
+      timers.push(t);
+      return t.id;
+    },
+    clearTimeoutFn: (id) => {
+      const i = timers.findIndex((t) => t.id === id);
+      if (i >= 0) timers.splice(i, 1);
+    }
+  });
+  const fireNext = async () => {
+    timers.sort((a, b) => a.at - b.at);
+    const t = timers.shift();
+    if (!t) throw new Error("no armed timer to fire");
+    nowMs = t.at;
+    await t.fn();
+  };
+  return {
+    sched,
+    store,
+    timers,
+    fireNext,
+    runs: () => runs,
+    now: () => nowMs,
+    setOutcome: (fn) => { outcome = fn; },
+    write: () => { nowMs += 1000; sched.noteWrite(); }
+  };
+}
+
+const HOUR = 3_600_000;
+
+test("issue#529 默认关：失败一次后重试间隔仍是基数，不翻倍", async () => {
+  const h = makeHarness({ sleepFailureBackoff: false });
+  h.write();
+  await h.fireNext();
+  assert.equal(h.runs(), 1, "idle 到点起跑一次");
+  const t0 = h.now();
+  assert.equal(h.sched.shouldRun(t0 + HOUR - 1), false, "基数间隔内照旧不放行");
+  assert.equal(h.sched.shouldRun(t0 + HOUR), true, "关键：基数一到即放行——默认关不翻倍");
+  h.store.close();
+});
+
+test("issue#529 开启：连败把下次放行时刻指数推迟", async () => {
+  const h = makeHarness({ sleepFailureBackoff: true });
+  h.write();
+  await h.fireNext(); // 败 1 → 有效间隔 1h × 2^1 = 2h
+  assert.equal(h.runs(), 1);
+  const t1 = h.now();
+  assert.equal(h.sched.shouldRun(t1 + HOUR), false, "关键：基数到点不放行（现状会放行）");
+  assert.equal(h.sched.shouldRun(t1 + 2 * HOUR), true, "满 2h 才放行第 2 次");
+
+  await h.fireNext(); // 败 2 → 有效间隔 4h
+  assert.equal(h.runs(), 2);
+  const t2 = h.now();
+  assert.equal(h.sched.shouldRun(t2 + 3 * HOUR), false, "连败 2 次后 4h 内不放行");
+  assert.equal(h.sched.shouldRun(t2 + 4 * HOUR), true, "满 4h 放行第 3 次");
+  h.store.close();
+});
+
+test("issue#529 成功一次即清零、间隔回到基数", async () => {
+  const h = makeHarness({ sleepFailureBackoff: true });
+  h.write();
+  await h.fireNext(); // 败 1 → 2h
+  const t1 = h.now();
+  assert.equal(h.sched.shouldRun(t1 + HOUR), false, "前置：退避确实在挡");
+
+  h.setOutcome(() => ({ ok: true, status: "ok" }));
+  await h.fireNext(); // 成功 → 连败清零
+  const t2 = h.now();
+  assert.equal(h.sched.shouldRun(t2 + HOUR), true, "关键：清零后基数即放行（未清零要等 2h）");
+  h.store.close();
+});
+
+test("issue#529 封顶：连败 5 次推迟 24h 即止（未封顶需等 32h）", async () => {
+  const h = makeHarness({ sleepFailureBackoff: true });
+  h.write();
+  await h.fireNext();
+  for (let i = 2; i <= 5; i++) {
+    const wait = Math.min(HOUR * 2 ** (i - 1), 24 * HOUR); // 与实现同一口径的期望间隔
+    assert.equal(h.sched.shouldRun(h.now() + wait), true, `第 ${i} 次起跑前满 ${wait / HOUR}h`);
+    await h.fireNext();
+  }
+  assert.equal(h.runs(), 5);
+  const t = h.now();
+  assert.equal(h.sched.shouldRun(t + 24 * HOUR - 1), false, "关键：封顶后 24h 内仍不放行");
+  assert.equal(h.sched.shouldRun(t + 24 * HOUR), true, "封顶生效：24h 放行，不再翻倍");
+  h.store.close();
+});
+
+test("issue#529 noop 不算失败：没活干的库不该被退避越拖越稀", async () => {
+  const h = makeHarness({ sleepFailureBackoff: true });
+  h.setOutcome(() => ({ ok: false, status: "noop" }));
+  h.write();
+  await h.fireNext();
+  assert.equal(h.runs(), 1);
+  const t0 = h.now();
+  assert.equal(h.sched.shouldRun(t0 + HOUR), true, "关键：noop 后基数即放行（当作失败要等 2h）");
+  h.store.close();
+});
+
+test("issue#529 抛错同样计入连败", async () => {
+  const h = makeHarness({ sleepFailureBackoff: true });
+  h.setOutcome(() => { throw new Error("boom"); });
+  h.write();
+  await h.fireNext();
+  const t0 = h.now();
+  assert.equal(h.sched.shouldRun(t0 + HOUR), false, "关键：抛错也退避到 2h");
+  assert.equal(h.sched.shouldRun(t0 + 2 * HOUR), true);
+  h.store.close();
+});