@modusensus/dsh-mneme 0.9.1 → 0.9.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +7 -6
- package/bin/cli.mjs +61 -7
- package/docs/CONFIGURATION.md +23 -0
- package/lib/api.js +57 -2
- package/lib/client.js +72 -9
- package/lib/config.js +17 -0
- package/lib/dream/sleep.js +35 -12
- package/lib/dream.js +86 -14
- package/lib/embedding.js +52 -22
- package/lib/scope-state.js +277 -0
- package/lib/service.js +80 -3
- package/lib/settings.js +7 -0
- package/lib/store.js +139 -0
- package/lib/tools.js +71 -0
- package/package.json +1 -1
- package/src/api.js +57 -2
- package/src/config.js +17 -0
- package/src/dream/sleep.js +35 -12
- package/src/dream.js +86 -14
- package/src/embedding.js +52 -22
- package/src/scope-state.js +277 -0
- package/src/service.js +80 -3
- package/src/settings.js +7 -0
- package/src/store.js +139 -0
- package/src/tools.js +71 -0
- package/test/api.test.js +113 -1
- package/test/cli-flag-value.test.js +86 -0
- package/test/client.test.js +64 -3
- package/test/distill-tx-embed-gap.test.js +110 -0
- package/test/dream-maxtokens-fallback.test.js +132 -0
- package/test/embed-text-reasons.test.js +112 -0
- package/test/readme-claims.test.js +68 -0
- package/test/scope-state.test.js +481 -0
- package/test/tools-dto-schema-matrix.test.js +12 -3
- package/test/tools.test.js +6 -6
package/lib/service.js
CHANGED
|
@@ -2,6 +2,7 @@ import { createHash, randomUUID } from "node:crypto";
|
|
|
2
2
|
import { TYPE_FILE, MIRROR_READONLY_TYPES } from "./mirror.js";
|
|
3
3
|
import { updatedAtBounds } from "./store.js";
|
|
4
4
|
import { normalizeExplicitScope, scopeKeyOf } from "./scope.js";
|
|
5
|
+
import { createScopeStateGate } from "./scope-state.js";
|
|
5
6
|
import { STR, langOf } from "./lang.js";
|
|
6
7
|
import { computeHeat } from "./heat.js";
|
|
7
8
|
import { recallStats } from "./recall-stats.js";
|
|
@@ -185,6 +186,17 @@ export function computeRetrievalMetrics(actualIds, expectedIds) {
|
|
|
185
186
|
|
|
186
187
|
export function createService({ store, mirror, config, onWrite, logger, documentIndex, writeAdmission }) {
|
|
187
188
|
const language = langOf(config);
|
|
189
|
+
|
|
190
|
+
// issue #471 政策轴:作用域级可逆闸门。判定实现住 src/scope-state.js(与存储解耦,
|
|
191
|
+
// test 可以直接构造暂停集合、不必开库);这里只把 store 的读写函数接上去。
|
|
192
|
+
// 每次调用现读一次状态表,不做缓存:表恒小(行数 = 被人按过的作用域个数),
|
|
193
|
+
// 而缓存要配一套失效协议才能让面板/另一进程的改动立刻可见(见 scope-state.js)。
|
|
194
|
+
const scopeStateGate = createScopeStateGate({
|
|
195
|
+
listStates: () => store.listScopeStates(),
|
|
196
|
+
saveAudit: (entry) => store.saveLlmAudit(entry),
|
|
197
|
+
config,
|
|
198
|
+
logger
|
|
199
|
+
});
|
|
188
200
|
// Optional dream scheduler hook, installed via setDreamHook after creation
|
|
189
201
|
// (the scheduler holds a reference back to the service, so it cannot be
|
|
190
202
|
// passed in the constructor). Fired on the same write events as onWrite.
|
|
@@ -294,7 +306,11 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
294
306
|
|
|
295
307
|
function scheduleEmbed(memory) {
|
|
296
308
|
try {
|
|
297
|
-
|
|
309
|
+
// distill-tx-embed-gap: 入队而不是丢弃,transaction() 提交后补跑(#372 同款)。
|
|
310
|
+
if (txDepth > 0) {
|
|
311
|
+
pendingEmbed.push(memory);
|
|
312
|
+
return;
|
|
313
|
+
}
|
|
298
314
|
if (!embedder || !memory?.id) return;
|
|
299
315
|
|
|
300
316
|
// Readiness gate: embedder exposes `ready` (async init) and is not ready
|
|
@@ -346,6 +362,14 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
346
362
|
// 写入批量决定。回滚时由 transaction() 截断到事务前的基线(见 ROLLBACK 路径)。
|
|
347
363
|
const pendingEntity = [];
|
|
348
364
|
|
|
365
|
+
// distill-tx-embed-gap:scheduleEmbed 在事务内曾被首行 `if (txDepth > 0) return;`
|
|
366
|
+
// 直接丢弃(注释写 deferred,transaction() 从不补跑)——与 #372 的实体抽取是
|
|
367
|
+
// 同一个 bug 模板,只是受害者换成了嵌入:自动蒸馏的 session:* 行因此 100%
|
|
368
|
+
// 拿不到向量。照 #372 先例入队,事务提交后补跑;回滚时截断回本层基线,
|
|
369
|
+
// 绝不给已回滚的行嵌向量。队列同样不设上限:元素只是对已落库行的引用,
|
|
370
|
+
// cap 丢最旧等于把「向量被静默丢弃」换个量级带回来。
|
|
371
|
+
const pendingEmbed = [];
|
|
372
|
+
|
|
349
373
|
function scheduleEntityExtraction(memory) {
|
|
350
374
|
if (txDepth > 0) {
|
|
351
375
|
pendingEntity.push(memory);
|
|
@@ -393,10 +417,27 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
393
417
|
* activeStoreSize 的理由:两处各写一份过滤条件,漂移只在某条路径上显形)。
|
|
394
418
|
*
|
|
395
419
|
* strictScope 关(默认)或 scope 解析不到时原样返回:行为与改动前逐字节一致。
|
|
420
|
+
*
|
|
421
|
+
* issue #471 政策轴接在这里:判定收在 scope-state.js(同一个「过滤点不许漂移」
|
|
422
|
+
* 的理由),本函数只按顺序挂上去。政策轴排在事实轴**前面**——它更便宜(一次小
|
|
423
|
+
* 表读,无逐行字段比较),且语义上「这一路现在通不通」先于「你看不看得见」。
|
|
424
|
+
*
|
|
425
|
+
* 注意政策轴的短路条件与事实轴不同:它不看 `scope`。事实轴要有 scope 才知道
|
|
426
|
+
* 「以谁的眼睛看」,而政策轴问的是「这条路还开着吗」,与调用方是谁无关——会话
|
|
427
|
+
* 身份解析失败不该让闸门整条失效(故政策轴在 scope 为 null 时照样生效)。
|
|
428
|
+
*
|
|
429
|
+
* **覆盖面只到「喂给模型的那几条路」**:融合池 + `entity:`/`attr:` 前缀路 +
|
|
430
|
+
* 注入候选池,以及 dream/sleep 的整理读入口。`memory_list` / `memory_get` /
|
|
431
|
+
* 面板列表**刻意不过这道闸**(它们走 `strictVisibility` 与 store.list 的
|
|
432
|
+
* `visibility` 参数,与 `gateByScope` 平行的另一条链)。理由:按下之后人还要能
|
|
433
|
+
* 查「我到底按下了什么、里面有什么」,那是检视面不是供料面;把检视面也关掉,
|
|
434
|
+
* 人就没有任何入口找回被暂停的行、也没法确认该不该松开。这也是为什么按下的是
|
|
435
|
+
* 政策而不是事实:行还在,只是不再喂给模型。
|
|
396
436
|
*/
|
|
397
437
|
function gateByScope(rows, scope) {
|
|
398
|
-
|
|
399
|
-
|
|
438
|
+
const gated = scopeStateGate.filter(rows);
|
|
439
|
+
if (config?.strictScope !== true || !scope) return gated;
|
|
440
|
+
return gated.filter((m) => isVisibleInScope(m, scope));
|
|
400
441
|
}
|
|
401
442
|
|
|
402
443
|
/**
|
|
@@ -1152,6 +1193,8 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
1152
1193
|
// #372: 记住本层事务开始时的队列深度——回滚时截断回这里。本层(及更深的
|
|
1153
1194
|
// 嵌套层)入队的抽取随回滚一并丢弃,外层已提交未 drain 的条目不受牵连。
|
|
1154
1195
|
const baseline = pendingEntity.length;
|
|
1196
|
+
// distill-tx-embed-gap: 嵌入队列用同一套基线机制,与实体抽取保持一致。
|
|
1197
|
+
const embedBaseline = pendingEmbed.length;
|
|
1155
1198
|
let committed = false;
|
|
1156
1199
|
try {
|
|
1157
1200
|
const result = fn();
|
|
@@ -1164,6 +1207,8 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
1164
1207
|
// 复审抓出:仅靠 committed 门挡 drain 时,回滚条目会滞留队列、被下一个
|
|
1165
1208
|
// 成功事务误抽)。外层未提交的条目在 baseline 之前,原样保留。
|
|
1166
1209
|
pendingEntity.length = baseline;
|
|
1210
|
+
// distill-tx-embed-gap: 同理,绝不给已回滚的行嵌向量。
|
|
1211
|
+
pendingEmbed.length = embedBaseline;
|
|
1167
1212
|
throw error;
|
|
1168
1213
|
} finally {
|
|
1169
1214
|
txDepth--;
|
|
@@ -1180,6 +1225,10 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
1180
1225
|
// (txDepth 仍 >0 时 schedule 会重新入队,与 drain 形成自旋)——条目等最外层
|
|
1181
1226
|
// COMMIT 后统一补跑。
|
|
1182
1227
|
if (committed && txDepth === 0) drainPendingEntityExtraction();
|
|
1228
|
+
// distill-tx-embed-gap: 补跑事务内排队的嵌入——同一套嵌套规则:内层提交
|
|
1229
|
+
// 不 drain(txDepth 仍 >0 时 scheduleEmbed 会重新入队),等最外层 COMMIT
|
|
1230
|
+
// 后统一补跑。放在 drainPendingEntityExtraction 之后无依赖,仅保持顺序。
|
|
1231
|
+
if (committed && txDepth === 0) drainPendingEmbeds();
|
|
1183
1232
|
}
|
|
1184
1233
|
}
|
|
1185
1234
|
|
|
@@ -1189,6 +1238,14 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
1189
1238
|
}
|
|
1190
1239
|
}
|
|
1191
1240
|
|
|
1241
|
+
// distill-tx-embed-gap: drain 与 pendingEntity 同构。补跑走 scheduleEmbed 本体,
|
|
1242
|
+
// 因此 embedder 未 ready / 无 schedule 接口等既有分流逻辑全部复用,不另写一路。
|
|
1243
|
+
function drainPendingEmbeds() {
|
|
1244
|
+
while (pendingEmbed.length) {
|
|
1245
|
+
scheduleEmbed(pendingEmbed.shift());
|
|
1246
|
+
}
|
|
1247
|
+
}
|
|
1248
|
+
|
|
1192
1249
|
/**
|
|
1193
1250
|
* Embed an arbitrary query text and return its vector (null on failure / no
|
|
1194
1251
|
* embedder). Used by the injector to prefetch the semantic-first recall
|
|
@@ -1753,6 +1810,11 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
1753
1810
|
if (config?.strictScope === true && scope) {
|
|
1754
1811
|
candidates = candidates.filter((m) => isVisibleInScope(m, scope));
|
|
1755
1812
|
}
|
|
1813
|
+
// issue #471 政策轴:挂在事实轴之后、pin 池之前。位置与上面同一条纪律——必须在
|
|
1814
|
+
// 图召回合并之后,否则出局行会被 entityRecall 打上 graphHint 带回候选池。
|
|
1815
|
+
// 观察档(enabled 开、enforce 关)在这里留一条审计行说清「本应被拦下几条」,
|
|
1816
|
+
// 但不改 candidates;拦截档才真的摘掉。审计写失败只 warn,注入照常完成。
|
|
1817
|
+
candidates = scopeStateGate.gateInjection(candidates).rows;
|
|
1756
1818
|
// #249 第一批:B1 pin 池。取在相关性排序之后、轮换之前——取谁按此刻的候选
|
|
1757
1819
|
// 次序(即相关性次序),取到后从候选中摘除,于是下面的轮换重排碰不到它们
|
|
1758
1820
|
// (验收:pin 不参与跨轮轮换)。独立预算的两层意义:pin 既不占 maxItems
|
|
@@ -2516,6 +2578,16 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
2516
2578
|
mergeHumanEdits,
|
|
2517
2579
|
toApiList,
|
|
2518
2580
|
isVisibleInScope,
|
|
2581
|
+
// issue #471 政策轴:按下/松开与读取照 store 那三条件透传(纯 bookkeeping,
|
|
2582
|
+
// 不挂 write hook——它们不动任何记忆行,同 saveConflictPending 那批)。
|
|
2583
|
+
setScopeState: (r) => store.setScopeState(r),
|
|
2584
|
+
getScopeState: (d, v) => store.getScopeState(d, v),
|
|
2585
|
+
listScopeStates: (o) => store.listScopeStates(o),
|
|
2586
|
+
// 巩固(dream.js)侧的过滤出口。暴露出来而不是让 dream 自己读状态表,是因为
|
|
2587
|
+
// 判定只能有一份实现(gateByScope 注释里那条「过滤点漂移」的教训);dream 侧
|
|
2588
|
+
// 的三个读入口各自的附加条件不同(有的排 forgotten、有的不排),但「被按下的
|
|
2589
|
+
// 作用域不参与巩固」这一层对三处是一样的。enforce 关时本函数原样返回入参。
|
|
2590
|
+
filterPaused: (rows) => scopeStateGate.filter(rows),
|
|
2519
2591
|
transaction,
|
|
2520
2592
|
getDistillCursor: (sessionId) => store.getDistillCursor(sessionId),
|
|
2521
2593
|
setDistillCursor: (sessionId, lastSeq) => store.setDistillCursor(sessionId, lastSeq),
|
|
@@ -2592,6 +2664,11 @@ export function createService({ store, mirror, config, onWrite, logger, document
|
|
|
2592
2664
|
search: (q, o) => store.search(q, o),
|
|
2593
2665
|
searchVector: (v, o) => store.searchVector(v, o),
|
|
2594
2666
|
embeddedCount: () => store.embeddedCount(),
|
|
2667
|
+
// 向量出身的两个分组读数:状态端点 /semantic 用它们区分「有向量」和
|
|
2668
|
+
// 「可检索」(byModel 按行级 embedding_model,byDimension 按向量维数)。
|
|
2669
|
+
// passthrough 形状与 embeddedCount 同款,api 层不直接摸 store。
|
|
2670
|
+
embeddingModelBreakdown: () => store.embeddingModelBreakdown(),
|
|
2671
|
+
embeddingDimensionBreakdown: () => store.embeddingDimensionBreakdown(),
|
|
2595
2672
|
list: (o) => store.list(o),
|
|
2596
2673
|
all: () => store.all(),
|
|
2597
2674
|
count: (type, opts) => store.count(type, opts),
|
package/lib/settings.js
CHANGED
|
@@ -174,6 +174,13 @@ const FEATURE_FLAG_BOOLEANS = [
|
|
|
174
174
|
// Issue #17(v0.8.0 A3):strictScope 硬过滤——他 scope 完全不可见(关闭时
|
|
175
175
|
// 为 A2 软隔离:降权保留可见)。
|
|
176
176
|
"strictScope",
|
|
177
|
+
// Issue #471:政策轴——作用域级可逆闸门(scopeState)。与上面两个键是两条不同的
|
|
178
|
+
// 轴:那两个管「这条记忆属于谁」(事实轴),这两个管「这个作用域现在通不通」
|
|
179
|
+
// (政策轴,按作用域整体按下、跨会话存续、松开即恢复)。两级语义同
|
|
180
|
+
// writeAdmission/preInjectGate——enabled 跑判定(观察档,留审计行但照常注入)、
|
|
181
|
+
// enforce 真拦(不注入);都默认关。点号键平铺存、合并时展开回 scopeState 对象。
|
|
182
|
+
"scopeState.enabled",
|
|
183
|
+
"scopeState.enforce",
|
|
177
184
|
// 工具暴露开关(v0.8.5):记忆已每轮自动注入,memory_search/memory_archive
|
|
178
185
|
// 在慢/轻量模型上是多余往返,面板可关(默认关=行为不变)。
|
|
179
186
|
"disableMemorySearch",
|
package/lib/store.js
CHANGED
|
@@ -2,6 +2,15 @@ import { DatabaseSync } from "node:sqlite";
|
|
|
2
2
|
import { randomUUID } from "node:crypto";
|
|
3
3
|
import { statSync } from "node:fs";
|
|
4
4
|
import { contentHashOf } from "./content-hash.js";
|
|
5
|
+
// 政策轴(issue #471)的归一化函数。刻意只取归一化,不取判定:判定(buildPausedIndex /
|
|
6
|
+
// matchPausedScopes / createScopeStateGate)留在 scope-state.js,由 service 侧消费,
|
|
7
|
+
// 存储层只管「一格状态怎么落」。依赖方向 store → scope-state → scope 是单向的
|
|
8
|
+
// (scope.js 不 import 任何本包模块),所以不构成环。
|
|
9
|
+
import {
|
|
10
|
+
normalizeScopeStateDimension,
|
|
11
|
+
normalizeScopeStateState,
|
|
12
|
+
normalizeScopeStateValue
|
|
13
|
+
} from "./scope-state.js";
|
|
5
14
|
|
|
6
15
|
const SCHEMA = `
|
|
7
16
|
CREATE TABLE IF NOT EXISTS memories (
|
|
@@ -185,6 +194,25 @@ CREATE TABLE IF NOT EXISTS scope_changes (
|
|
|
185
194
|
);
|
|
186
195
|
CREATE INDEX IF NOT EXISTS idx_scope_changes_memory ON scope_changes(memory_id);
|
|
187
196
|
|
|
197
|
+
-- scope_state: v0.9.1 政策轴(issue #471)。与 scope_changes 那条事实轴相对:
|
|
198
|
+
-- scope_changes 记「某条记忆的归属被人工改过」(跟着记忆走),本表记「某个作用域
|
|
199
|
+
-- 现在通不通」(跟着作用域走)。按下/松开都是单方面动作,不触碰任何记忆行,
|
|
200
|
+
-- 松开即恢复——所以它是可逆的,而逐条 forget/archive 不是。
|
|
201
|
+
-- dimension 只认 'agent'|'workspace'(scopeKeyOf 那两个维度);scope_value='*'
|
|
202
|
+
-- 表示该维度全局。一行一格,重复按下同一状态不刷新 decided_at(见 setScopeState)。
|
|
203
|
+
-- Bookkeeping like the other audit tables: it never triggers write hooks.
|
|
204
|
+
CREATE TABLE IF NOT EXISTS scope_state (
|
|
205
|
+
id TEXT PRIMARY KEY,
|
|
206
|
+
dimension TEXT NOT NULL,
|
|
207
|
+
scope_value TEXT NOT NULL,
|
|
208
|
+
state TEXT NOT NULL, -- active | paused
|
|
209
|
+
note TEXT,
|
|
210
|
+
decided_at TEXT NOT NULL
|
|
211
|
+
);
|
|
212
|
+
-- 唯一性靠独立索引(本仓惯例:UNIQUE 不进 CREATE TABLE,见 idx_memory_provenance_link)。
|
|
213
|
+
-- 它同时是 setScopeState 那条 UPSERT 的冲突目标,所以不能省。
|
|
214
|
+
CREATE UNIQUE INDEX IF NOT EXISTS idx_scope_state_key ON scope_state(dimension, scope_value);
|
|
215
|
+
|
|
188
216
|
-- llm_audit_logs: every background LLM call (autoDream consolidation + summary,
|
|
189
217
|
-- autoSummarize compression) is recorded here — tokens in/out, duration, status
|
|
190
218
|
-- and the trigger that caused it (Bug8). Failures are captured as status='error'
|
|
@@ -604,6 +632,20 @@ function toScopeChange(row) {
|
|
|
604
632
|
};
|
|
605
633
|
}
|
|
606
634
|
|
|
635
|
+
// scope_state 行 → 读侧对象(issue #471)。NULL note 落 undefined,与
|
|
636
|
+
// toScopeChange 同口径。
|
|
637
|
+
function toScopeState(row) {
|
|
638
|
+
if (!row) return undefined;
|
|
639
|
+
return {
|
|
640
|
+
id: row.id,
|
|
641
|
+
dimension: row.dimension,
|
|
642
|
+
scope_value: row.scope_value,
|
|
643
|
+
state: row.state,
|
|
644
|
+
note: row.note ?? undefined,
|
|
645
|
+
decided_at: row.decided_at
|
|
646
|
+
};
|
|
647
|
+
}
|
|
648
|
+
|
|
607
649
|
function toRecallRun(row) {
|
|
608
650
|
if (!row) return undefined;
|
|
609
651
|
return {
|
|
@@ -1733,6 +1775,32 @@ export function createStore(path) {
|
|
|
1733
1775
|
).all();
|
|
1734
1776
|
}
|
|
1735
1777
|
|
|
1778
|
+
/**
|
|
1779
|
+
* Per-dimension counts over active embedded rows ('' = vector that failed to
|
|
1780
|
+
* parse as an array — legacy garbage rows, not a real dimension).
|
|
1781
|
+
*
|
|
1782
|
+
* Why per-dimension: embeddedCount() alone says "has a vector", but vectors
|
|
1783
|
+
* written by a different model/dimension are not comparable with the current
|
|
1784
|
+
* query embedding — searchVector skips them, so they are "present" without
|
|
1785
|
+
* being "retrievable". The status card needs this split to stop reporting a
|
|
1786
|
+
* coverage number that silently counts unsearchable rows.
|
|
1787
|
+
*/
|
|
1788
|
+
function embeddingDimensionBreakdown() {
|
|
1789
|
+
return db.prepare(
|
|
1790
|
+
`SELECT id, embedding FROM memories
|
|
1791
|
+
WHERE embedding IS NOT NULL AND embedding != ''
|
|
1792
|
+
AND forgotten = 0 AND archived = 0`
|
|
1793
|
+
).all().reduce((acc, row) => {
|
|
1794
|
+
let dim = "";
|
|
1795
|
+
try {
|
|
1796
|
+
const parsed = JSON.parse(row.embedding);
|
|
1797
|
+
dim = Array.isArray(parsed) ? String(parsed.length) : "";
|
|
1798
|
+
} catch { /* unparseable vector counts under "" */ }
|
|
1799
|
+
acc.set(dim, (acc.get(dim) ?? 0) + 1);
|
|
1800
|
+
return acc;
|
|
1801
|
+
}, new Map());
|
|
1802
|
+
}
|
|
1803
|
+
|
|
1736
1804
|
function search(query, { limit = 20, includeArchived = false } = {}) {
|
|
1737
1805
|
const q = String(query).trim();
|
|
1738
1806
|
if (!q) return [];
|
|
@@ -1789,6 +1857,11 @@ export function createStore(path) {
|
|
|
1789
1857
|
for (const row of rows) {
|
|
1790
1858
|
const v = getParsedEmbedding(row.id);
|
|
1791
1859
|
if (!v) continue;
|
|
1860
|
+
// distill-tx-embed-gap 顺带修:维度不匹配的行按维度分流,直接跳过。旧路径
|
|
1861
|
+
// 喂进 cosine 不等长就 return 0,换了模型/维度的旧向量以 0 分混进结果尾部
|
|
1862
|
+
// (threshold 默认 0 时照样入选),检索方拿到的分数是跨向量空间的假值。
|
|
1863
|
+
// 不等长的向量对本次查询没有可比分数,跳过是唯一诚实的处置。
|
|
1864
|
+
if (v.length !== vector.length) continue;
|
|
1792
1865
|
const score = cosine(vector, v);
|
|
1793
1866
|
if (score >= threshold) scored.push({ row: { id: row.id }, score });
|
|
1794
1867
|
}
|
|
@@ -2609,6 +2682,66 @@ export function createStore(path) {
|
|
|
2609
2682
|
return rows.map(toScopeChange);
|
|
2610
2683
|
}
|
|
2611
2684
|
|
|
2685
|
+
/**
|
|
2686
|
+
* 政策轴(issue #471):按下或松开一个作用域。一行一格,靠
|
|
2687
|
+
* idx_scope_state_key(dimension, scope_value) 做 UPSERT——重复按下同一状态
|
|
2688
|
+
* 不刷新 decided_at(CASE WHEN),因为「第一次按下的时刻」才是回执该报的东西;
|
|
2689
|
+
* 状态翻转则刷新,那是一次真的决策。
|
|
2690
|
+
*
|
|
2691
|
+
* note 的取舍与 decided_at 相反:非空的新 note 总是覆盖旧的,空的则原样留着。
|
|
2692
|
+
* 理由是这条记录里「当前通不通」由 state 那一列承担,note 是**理由**——理由被
|
|
2693
|
+
* 一次不带理由的重按抹掉是净损失(用户回看时只会看到「按下过」却不知道为什么),
|
|
2694
|
+
* 而「清空理由」不是一个真实需求(要改就再按一次带上新理由)。这与本仓防御段
|
|
2695
|
+
* 「宁可留着也不要静默丢」的取向一致。
|
|
2696
|
+
*
|
|
2697
|
+
* 归一化失败(维度/值/状态任一非法)返回 undefined,由调用方拒绝——本表只存
|
|
2698
|
+
* 可解释的三种枚举,与 memories 的 scope 列不同:那里脏输入落 NULL 是正当的
|
|
2699
|
+
* (「未标注」本身有语义),这里落 NULL 没有任何意思,会静默变成一行永远不
|
|
2700
|
+
* 生效的状态。
|
|
2701
|
+
*/
|
|
2702
|
+
function setScopeState({ dimension, scope_value, state, note, decided_at }) {
|
|
2703
|
+
const dim = normalizeScopeStateDimension(dimension);
|
|
2704
|
+
const val = normalizeScopeStateValue(scope_value);
|
|
2705
|
+
const st = normalizeScopeStateState(state);
|
|
2706
|
+
if (!dim || !val || !st) return undefined;
|
|
2707
|
+
const at = normalizeOccurredAt(decided_at) ?? nowIso();
|
|
2708
|
+
const noteText = normalizeScopeText(note);
|
|
2709
|
+
db.prepare(
|
|
2710
|
+
`INSERT INTO scope_state (id, dimension, scope_value, state, note, decided_at)
|
|
2711
|
+
VALUES (?, ?, ?, ?, ?, ?)
|
|
2712
|
+
ON CONFLICT(dimension, scope_value) DO UPDATE SET
|
|
2713
|
+
state = excluded.state,
|
|
2714
|
+
note = COALESCE(excluded.note, scope_state.note),
|
|
2715
|
+
decided_at = CASE WHEN scope_state.state = excluded.state THEN scope_state.decided_at ELSE excluded.decided_at END`
|
|
2716
|
+
).run(randomUUID(), dim, val, st, noteText, at);
|
|
2717
|
+
return toScopeState(
|
|
2718
|
+
db.prepare("SELECT * FROM scope_state WHERE dimension = ? AND scope_value = ?").get(dim, val)
|
|
2719
|
+
);
|
|
2720
|
+
}
|
|
2721
|
+
|
|
2722
|
+
/** 政策轴状态行(单格)。 */
|
|
2723
|
+
function getScopeState(dimension, scope_value) {
|
|
2724
|
+
const dim = normalizeScopeStateDimension(dimension);
|
|
2725
|
+
const val = normalizeScopeStateValue(scope_value);
|
|
2726
|
+
if (!dim || !val) return undefined;
|
|
2727
|
+
return toScopeState(
|
|
2728
|
+
db.prepare("SELECT * FROM scope_state WHERE dimension = ? AND scope_value = ?").get(dim, val)
|
|
2729
|
+
);
|
|
2730
|
+
}
|
|
2731
|
+
|
|
2732
|
+
/**
|
|
2733
|
+
* 政策轴状态行(全量)。不提供按维度/值过滤的参数——闸门判定要的就是全量
|
|
2734
|
+
* (一次现读一次索引),过滤留到 scope-state.js 的 buildPausedIndex 里做。
|
|
2735
|
+
* 排 decided_at DESC + rowid DESC 同 listScopeChanges(毫秒精度同毫秒写入
|
|
2736
|
+
* 时 UUID 排序随机,回执要报「最近一次决策」就得有确定的次键)。
|
|
2737
|
+
*/
|
|
2738
|
+
function listScopeStates({ state } = {}) {
|
|
2739
|
+
const rows = state
|
|
2740
|
+
? db.prepare("SELECT * FROM scope_state WHERE state = ? ORDER BY decided_at DESC, rowid DESC").all(state)
|
|
2741
|
+
: db.prepare("SELECT * FROM scope_state ORDER BY decided_at DESC, rowid DESC").all();
|
|
2742
|
+
return rows.map(toScopeState);
|
|
2743
|
+
}
|
|
2744
|
+
|
|
2612
2745
|
function getFailureStats({ since } = {}) {
|
|
2613
2746
|
const clause = since ? "WHERE created_at >= ?" : "";
|
|
2614
2747
|
const params = since ? [since] : [];
|
|
@@ -3233,6 +3366,7 @@ export function createStore(path) {
|
|
|
3233
3366
|
needsEmbeddingFor,
|
|
3234
3367
|
hasStaleEmbedding,
|
|
3235
3368
|
embeddingModelBreakdown,
|
|
3369
|
+
embeddingDimensionBreakdown,
|
|
3236
3370
|
searchVector,
|
|
3237
3371
|
saveDreamRun,
|
|
3238
3372
|
lastDreamRunAt,
|
|
@@ -3281,6 +3415,11 @@ export function createStore(path) {
|
|
|
3281
3415
|
countConflictPending,
|
|
3282
3416
|
saveScopeChange,
|
|
3283
3417
|
listScopeChanges,
|
|
3418
|
+
// 政策轴(issue #471):按下/松开一个作用域。纯 bookkeeping,不调
|
|
3419
|
+
// incrementGeneration()——它不动任何记忆行,动它只会让所有缓存无谓失效。
|
|
3420
|
+
setScopeState,
|
|
3421
|
+
getScopeState,
|
|
3422
|
+
listScopeStates,
|
|
3284
3423
|
createEntity,
|
|
3285
3424
|
findEntityByName,
|
|
3286
3425
|
findEntityById,
|
package/lib/tools.js
CHANGED
|
@@ -7,6 +7,7 @@ import { matchesPlatform } from "./runtime/closure.js";
|
|
|
7
7
|
import { verifyPayload } from "./runtime/verify.js";
|
|
8
8
|
import { TOOL_GUIDE, CONTINUITY_TOOL_RULE } from "./guide.js";
|
|
9
9
|
import { injectChildEnabled } from "./config.js";
|
|
10
|
+
import { normalizeScopeStateDimension, normalizeScopeStateValue, normalizeScopeStateState } from "./scope-state.js";
|
|
10
11
|
import { lastUserQuery } from "./inject.js";
|
|
11
12
|
|
|
12
13
|
const TEXT_OUTPUT = (text) => [{ type: "text", text }];
|
|
@@ -500,6 +501,76 @@ export function createTools(ctx, service, config, embedder) {
|
|
|
500
501
|
}
|
|
501
502
|
}),
|
|
502
503
|
|
|
504
|
+
// #471:政策轴——作用域级可逆闸门。与 memory_forget/memory_archive 的分工:
|
|
505
|
+
// 那两个是事实轴的单条操作(改的是「这条记忆属于谁 / 还算不算数」,一条条来);
|
|
506
|
+
// 本工具是政策轴的按作用域操作(改的是「这一路现在通不通」,跨会话存续、
|
|
507
|
+
// 松开即原样回来,一条记忆行都不动)。判断指引:整条工作线要临时静音用本工具;
|
|
508
|
+
// 单条记忆不该再被想起才用 memory_forget/memory_archive。
|
|
509
|
+
defineTool({
|
|
510
|
+
name: "set_scope_state",
|
|
511
|
+
description:
|
|
512
|
+
"Press or release the reversible scope gate for one scope (issue #471): while paused, that scope's memories stay stored and unchanged but are held back from injection and dream consolidation. " +
|
|
513
|
+
"This is not deletion and not forgetting — release with state=active and everything comes back exactly as it was, which is why it is the right tool for temporarily silencing a whole work stream rather than editing memories one by one. " +
|
|
514
|
+
"dimension is agent or workspace; scope_value is the label itself, or * for that dimension's global gate. " +
|
|
515
|
+
"Both switches (scopeState.enabled/enforce) are off by default, so setting a state has no effect on injection until the user turns them on in the panel.",
|
|
516
|
+
parameters: {
|
|
517
|
+
dimension: { type: "string", required: true, description: "Scope dimension to press: \"agent\" or \"workspace\"" },
|
|
518
|
+
scope_value: { type: "string", required: true, description: "The scope label, or \"*\" for that dimension's global gate" },
|
|
519
|
+
state: { type: "string", required: true, description: "\"paused\" presses the gate, \"active\" releases it" },
|
|
520
|
+
note: { type: "string", description: "Optional short reason, stored with the decision so the next reader knows why it was pressed" }
|
|
521
|
+
},
|
|
522
|
+
output: {
|
|
523
|
+
schema: {
|
|
524
|
+
type: "object",
|
|
525
|
+
additionalProperties: false,
|
|
526
|
+
properties: {
|
|
527
|
+
scope: {
|
|
528
|
+
type: "object",
|
|
529
|
+
additionalProperties: false,
|
|
530
|
+
properties: {
|
|
531
|
+
dimension: { type: "string", required: true },
|
|
532
|
+
scope_value: { type: "string", required: true },
|
|
533
|
+
state: { type: "string", required: true },
|
|
534
|
+
// 理由要能读回来:不带它,下次读的人只知道「按下过」却不知道为什么。
|
|
535
|
+
// 这与库层 setScopeState 的 note 取舍是同一条理由——不带理由的重按
|
|
536
|
+
// 不抹掉旧理由,正是为了让这里读得回来。可选:从没写过就没有。
|
|
537
|
+
note: { type: "string" },
|
|
538
|
+
decided_at: { type: "string", required: true }
|
|
539
|
+
}
|
|
540
|
+
}
|
|
541
|
+
}
|
|
542
|
+
},
|
|
543
|
+
// 回执如实回显当前值(不是回显入参):同一个状态重复按下时 decided_at 保持
|
|
544
|
+
// 首次值,回显它才看得出「这次没有真的改变什么」。
|
|
545
|
+
render: (_args, value) =>
|
|
546
|
+
TEXT_OUTPUT(
|
|
547
|
+
`Scope ${value.scope.dimension}:${value.scope.scope_value} is now ${value.scope.state} (decided at ${value.scope.decided_at}).`
|
|
548
|
+
)
|
|
549
|
+
},
|
|
550
|
+
async execute(args) {
|
|
551
|
+
const dimension = normalizeScopeStateDimension(args.dimension);
|
|
552
|
+
const scopeValue = normalizeScopeStateValue(args.scope_value);
|
|
553
|
+
const state = normalizeScopeStateState(args.state);
|
|
554
|
+
if (dimension === null) throw new Error("dimension must be \"agent\" or \"workspace\"");
|
|
555
|
+
if (scopeValue === null) throw new Error("scope_value must be a non-empty label or \"*\"");
|
|
556
|
+
if (state === null) throw new Error("state must be \"active\" or \"paused\"");
|
|
557
|
+
const row = service.setScopeState({ dimension, scope_value: scopeValue, state, note: args.note });
|
|
558
|
+
if (row === undefined) throw new Error("scope state could not be recorded");
|
|
559
|
+
return {
|
|
560
|
+
scope: {
|
|
561
|
+
dimension: row.dimension,
|
|
562
|
+
scope_value: row.scope_value,
|
|
563
|
+
state: row.state,
|
|
564
|
+
// 没写过理由时键整个不出现(而不是 note: undefined):output schema 是
|
|
565
|
+
// 闭合的,DTO 矩阵会拿真实返回值过 validateJsonSchemaValue,留一个
|
|
566
|
+
// 值为 undefined 的键等于留个空壳。
|
|
567
|
+
...(row.note === undefined ? {} : { note: row.note }),
|
|
568
|
+
decided_at: row.decided_at
|
|
569
|
+
}
|
|
570
|
+
};
|
|
571
|
+
}
|
|
572
|
+
}),
|
|
573
|
+
|
|
503
574
|
// #230(#164 设计稿评审线):document 型记忆注册——agent 产长文档的指针
|
|
504
575
|
// 行铸造口。写入权分离:全文归 agent(管线对文件零读零写零改),库里只
|
|
505
576
|
// 存摘要 + doc_path + evidence 三样;C2 比对去重、supersede 记账在
|
package/package.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@modusensus/dsh-mneme",
|
|
3
3
|
"description": "Cross-session memory plugin for DeepSeek Harness with autoDream consolidation: SQLite store, Markdown mirrors, 10 model tools, automatic injection, session summarization, user profile/rules, custom slash commands, vector (semantic) search, and a Web GUI panel",
|
|
4
|
-
"version": "0.9.
|
|
4
|
+
"version": "0.9.2",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"repository": {
|
|
7
7
|
"type": "git",
|
package/src/api.js
CHANGED
|
@@ -207,6 +207,15 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
|
|
|
207
207
|
// feature flags 快照(GET/PUT 共用):overrides 是持久化的用户显式覆盖;
|
|
208
208
|
// effective 是启动配置在白名单键上被 overrides 覆盖后的最终值。config 缺席
|
|
209
209
|
// (旧调用方直连、未传 cfg)时只报被覆盖的键,不把不存在的默认值编造给前端。
|
|
210
|
+
//
|
|
211
|
+
// pendingRestart 回答的是第三个问题:「我改的开关,现在真的生效了吗?」
|
|
212
|
+
// 功能开关的生效节奏与 panel_mode 相同——启动时由 index.js 合并进 cfg,运行中
|
|
213
|
+
// 那份 cfg 是冻结的(见 index.js 装配处的注释)。于是面板显示「已改」不等于
|
|
214
|
+
// 运行时已改:#481 实测人格总闸正是这样,面板开着、人格段却不注入,用户只能
|
|
215
|
+
// 靠反复试错发现。effective 表达的是「运行时取什么值」(overrides 优先),
|
|
216
|
+
// **反推不出待重启**,必须另比一次冻结值——这是本字段存在的唯一理由。
|
|
217
|
+
// 只报「显式覆盖 ≠ 冻结值」的键:未被覆盖的键运行时值就是 cfg 值,不存在落差;
|
|
218
|
+
// config 缺席(旧调用方直连、未传 cfg)时无从比较,返回空数组而不编造。
|
|
210
219
|
const flagKeys = [
|
|
211
220
|
...FEATURE_FLAG_SPEC.booleans,
|
|
212
221
|
...Object.keys(FEATURE_FLAG_SPEC.ints),
|
|
@@ -232,7 +241,11 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
|
|
|
232
241
|
// Issue #399:蒸馏原文留存 + 溯源。两个键都是嵌套子字段(enabled 开关 +
|
|
233
242
|
// retentionDays 整数),漏登记的后果同上——effective 静默缺键。
|
|
234
243
|
"distillRecords.enabled": ["distillRecords", "enabled"],
|
|
235
|
-
"distillRecords.retentionDays": ["distillRecords", "retentionDays"]
|
|
244
|
+
"distillRecords.retentionDays": ["distillRecords", "retentionDays"],
|
|
245
|
+
// Issue #471:政策轴——作用域级可逆闸门。两个键都是嵌套子字段,漏登记的后果
|
|
246
|
+
// 同上(effective 静默缺键、面板读不到也 PUT 不进)。
|
|
247
|
+
"scopeState.enabled": ["scopeState", "enabled"],
|
|
248
|
+
"scopeState.enforce": ["scopeState", "enforce"]
|
|
236
249
|
// sensitiveScanEnabled(#164 A2)是顶层扁平键,不需要进这张表——走 configFlagValue
|
|
237
250
|
// 的默认分支即可。
|
|
238
251
|
};
|
|
@@ -241,6 +254,19 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
|
|
|
241
254
|
if (path) return config?.[path[0]]?.[path[1]];
|
|
242
255
|
return config?.[key];
|
|
243
256
|
}
|
|
257
|
+
// 「显式覆盖 ≠ 冻结值」的键。用 Object.is 比而不是 !==:NaN 与 ±0 的边角上
|
|
258
|
+
// 两者会给出相反的结论(Object.is(NaN, NaN) 为真、Object.is(0, -0) 为假),
|
|
259
|
+
// 开关值里虽然不该出现它们,但比较语义与「值是否相同」对齐更省心。
|
|
260
|
+
// 未被覆盖的键不看:它的运行时值本来就取自 cfg,不存在落差。
|
|
261
|
+
function pendingRestartKeys(overrides) {
|
|
262
|
+
if (!config) return [];
|
|
263
|
+
const pending = [];
|
|
264
|
+
for (const key of flagKeys) {
|
|
265
|
+
if (overrides[key] === undefined) continue;
|
|
266
|
+
if (!Object.is(overrides[key], configFlagValue(key))) pending.push(key);
|
|
267
|
+
}
|
|
268
|
+
return pending;
|
|
269
|
+
}
|
|
244
270
|
function featureSnapshot() {
|
|
245
271
|
const overrides = settings.getFeatureFlags();
|
|
246
272
|
const effective = {};
|
|
@@ -251,7 +277,7 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
|
|
|
251
277
|
if (value !== undefined) effective[key] = value;
|
|
252
278
|
}
|
|
253
279
|
}
|
|
254
|
-
return { overrides, effective };
|
|
280
|
+
return { overrides, effective, pendingRestart: pendingRestartKeys(overrides) };
|
|
255
281
|
}
|
|
256
282
|
|
|
257
283
|
// Ensure the service has an embedder when the API layer was handed one
|
|
@@ -1054,6 +1080,25 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
|
|
|
1054
1080
|
const configured = embedder ? (typeof embedder.configured === "boolean" ? embedder.configured : true) : null;
|
|
1055
1081
|
const embedded = Number(stats?.embeddedCount ?? 0);
|
|
1056
1082
|
const total = Number(stats?.totalCount ?? 0);
|
|
1083
|
+
// Per-model / per-dimension provenance of stored vectors. "embedded"
|
|
1084
|
+
// counts rows with any vector; rows whose dimension differs from the
|
|
1085
|
+
// current embedder output are skipped by searchVector, so the honest
|
|
1086
|
+
// retrievable number is the current-dimension slice. modelBreakdown
|
|
1087
|
+
// groups by row-level embedding_model ('' = unknown origin).
|
|
1088
|
+
const dimensionBreakdown = typeof service.embeddingDimensionBreakdown === "function"
|
|
1089
|
+
? service.embeddingDimensionBreakdown()
|
|
1090
|
+
: null;
|
|
1091
|
+
const currentDim = embedder?.dimension != null ? String(embedder.dimension) : null;
|
|
1092
|
+
const byDimension = dimensionBreakdown
|
|
1093
|
+
? [...dimensionBreakdown.entries()].map(([dim, c]) => ({ dimension: dim === "" ? null : Number(dim), count: c }))
|
|
1094
|
+
: [];
|
|
1095
|
+
const unsearchable = (currentDim && dimensionBreakdown)
|
|
1096
|
+
? [...dimensionBreakdown.entries()]
|
|
1097
|
+
.filter(([dim]) => dim !== currentDim)
|
|
1098
|
+
.reduce((s, [, c]) => s + c, 0)
|
|
1099
|
+
: 0; const modelBreakdown = typeof service.embeddingModelBreakdown === "function"
|
|
1100
|
+
? service.embeddingModelBreakdown().map((r) => ({ model: r.model === "" ? null : r.model, count: r.c }))
|
|
1101
|
+
: [];
|
|
1057
1102
|
sendJson(res, 200, {
|
|
1058
1103
|
ready,
|
|
1059
1104
|
configured,
|
|
@@ -1064,6 +1109,16 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
|
|
|
1064
1109
|
: null,
|
|
1065
1110
|
// 空库时 ratio 必须是 null 而非 0 —— 0 会被读成「一条都没嵌上」。
|
|
1066
1111
|
coverage: { embedded, total, ratio: total > 0 ? embedded / total : null },
|
|
1112
|
+
// distill-tx-embed-gap 后续:向量「存在」≠「可检索」。unsearchable =
|
|
1113
|
+
// 维度与当前 embedder 输出不同的行数(searchVector 会跳过它们),
|
|
1114
|
+
// retrievable = 与当前维度同维、真正参与召回的行数。两者差值非零时
|
|
1115
|
+
// 面板必须能说出「有 N 条向量不可检索」,而不是只给一个虚高的 coverage。
|
|
1116
|
+
vectors: {
|
|
1117
|
+
byDimension,
|
|
1118
|
+
byModel: modelBreakdown,
|
|
1119
|
+
unsearchable,
|
|
1120
|
+
retrievable: currentDim ? (dimensionBreakdown.get(currentDim) ?? 0) : null,
|
|
1121
|
+
},
|
|
1067
1122
|
// 配了、库里有东西可嵌、却一条都没嵌上 = 真的降级(本机此前长期
|
|
1068
1123
|
// 处于 ready:true + embeddedCount:0,面板却显示正常)。
|
|
1069
1124
|
degraded: configured === true && total > 0 && embedded === 0,
|
package/src/config.js
CHANGED
|
@@ -863,6 +863,23 @@ export const Config = z.object({
|
|
|
863
863
|
// 数据、观察 A2 加权质量后再切硬过滤。
|
|
864
864
|
strictScope: z.boolean().default(false),
|
|
865
865
|
|
|
866
|
+
// scopeState(issue #471):政策轴——作用域级可逆闸门。与上面两个键是**两条不同
|
|
867
|
+
// 的轴**:scopeEnabled/strictScope 管「这条记忆属于谁」(事实轴:持久、跟着记忆
|
|
868
|
+
// 行走、改的是归属数据),本键管「这个作用域现在通不通」(政策轴:暂时、由人单
|
|
869
|
+
// 方面按下、松开即恢复、不改任何记忆行)。
|
|
870
|
+
// 两级语义同 writeAdmission/preInjectGate——enabled 打开时被按下的作用域其候选
|
|
871
|
+
// 照常注入,但留一条审计行说清「本应被拦下几条」(观察档,先在真实流量看会少掉
|
|
872
|
+
// 多少再决定真按);enforce 打开才真的不注入(拦截档)。两个都默认关,打开前
|
|
873
|
+
// 行为与今天逐字段一致。
|
|
874
|
+
// 不进 LIGHT_MODE_OFF:它只是查一张小表的判定,不引入重型后台路径、不加常驻
|
|
875
|
+
// 文本(同 actionPipelineEnabled 的定性)——轻量档预设该关的是重型能力,不该替
|
|
876
|
+
// 用户关掉一个纯判定入口,用户显式勾选仍然赢。
|
|
877
|
+
// 也走 feature_flags(FEATURE_FLAG_BOOLEANS 白名单),面板可启停=线上回滚开关。
|
|
878
|
+
scopeState: z.object({
|
|
879
|
+
enabled: z.boolean().default(false),
|
|
880
|
+
enforce: z.boolean().default(false)
|
|
881
|
+
}).default({}),
|
|
882
|
+
|
|
866
883
|
// --- v0.9.0 M.O.(Modus Operandi · 手法,beta)--------------------------
|
|
867
884
|
// 命名:M.O. = Modus Operandi,拉丁语「一个人一贯的做事方式」,缩写 M.O.;
|
|
868
885
|
// 也正是 org 词根 modus(modusensus)。界面主显示 M.O.,设置项 hint 第一句
|