renor-engine 1.2.0 → 1.4.0
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 +34 -2
- package/THIRD-PARTY-NOTICES +2661 -0
- package/bin/renor-engine.cjs +38 -0
- package/dist/BUILD.json +6 -0
- package/dist/cli.js +1674 -516
- package/package.json +17 -80
- package/dist/.tsbuildinfo +0 -1
- package/dist/api/apiAuth.js +0 -43
- package/dist/api/server.js +0 -174
- package/dist/buildIdentity.js +0 -149
- package/dist/callPhases.js +0 -82
- package/dist/cli-extractors.js +0 -46
- package/dist/config.js +0 -65
- package/dist/context/composer.js +0 -320
- package/dist/context/readHonesty.js +0 -478
- package/dist/context/rowStore.js +0 -238
- package/dist/conversation/formatter.js +0 -81
- package/dist/conversation/parser.js +0 -193
- package/dist/conversation/types.js +0 -2
- package/dist/conversation/writer.js +0 -42
- package/dist/db/client.js +0 -149
- package/dist/db/handoffTable.js +0 -31
- package/dist/db/indexSalvage.js +0 -101
- package/dist/db/migrations.js +0 -670
- package/dist/db/readCache.js +0 -109
- package/dist/db/schema.js +0 -5
- package/dist/db/sqlite.js +0 -396
- package/dist/db/statements.js +0 -89
- package/dist/digests/digestCache.js +0 -146
- package/dist/digests/digests.js +0 -597
- package/dist/digests/worker.js +0 -350
- package/dist/embeddings/caching.js +0 -337
- package/dist/embeddings/coalescing.js +0 -177
- package/dist/embeddings/interactive.js +0 -103
- package/dist/embeddings/local.js +0 -70
- package/dist/embeddings/openai.js +0 -106
- package/dist/embeddings/rebuild.js +0 -166
- package/dist/embeddings/simdScan.js +0 -260
- package/dist/embeddings/space.js +0 -21
- package/dist/embeddings/types.js +0 -14
- package/dist/embeddings/vectorIndex.js +0 -674
- package/dist/embeddings/vertex.js +0 -334
- package/dist/eval/artifacts.js +0 -42
- package/dist/eval/cacheReport.js +0 -146
- package/dist/eval/comparison.js +0 -182
- package/dist/eval/context700-candidate.js +0 -145
- package/dist/eval/evalCache.js +0 -568
- package/dist/eval/hashEmbedding.js +0 -35
- package/dist/eval/indexSnapshots.js +0 -549
- package/dist/eval/longmemeval.js +0 -56
- package/dist/eval/metrics.js +0 -181
- package/dist/eval/pageDump.js +0 -67
- package/dist/eval/renorbenchBaseline.js +0 -486
- package/dist/eval/renorbenchDocuments.js +0 -134
- package/dist/eval/renorbenchProduct.js +0 -182
- package/dist/eval/runLedger.js +0 -279
- package/dist/eval/runManifest.js +0 -485
- package/dist/eval/scoring.js +0 -197
- package/dist/eval/textTokens.js +0 -17
- package/dist/evidence/attach.js +0 -264
- package/dist/evidence/background.js +0 -147
- package/dist/evidence/evidenceCache.js +0 -182
- package/dist/evidence/gate.js +0 -37
- package/dist/evidence/keyPool.js +0 -85
- package/dist/evidence/ollamaChat.js +0 -84
- package/dist/evidence/optin.js +0 -165
- package/dist/evidence/options.js +0 -151
- package/dist/evidence/secondPass.js +0 -294
- package/dist/extraction/classifyCapture.js +0 -34
- package/dist/extraction/decision-model.js +0 -174
- package/dist/extraction/eventExtractor.js +0 -113
- package/dist/extraction/extractionClient.js +0 -122
- package/dist/extraction/factExtractor.js +0 -46
- package/dist/extraction/factsWritePath.js +0 -179
- package/dist/extraction/jevDecisionModel.js +0 -307
- package/dist/extraction/preferenceExtractor.js +0 -150
- package/dist/extraction/secretGuard.js +0 -180
- package/dist/extraction/secretScanner.js +0 -665
- package/dist/extraction/typedFacts.js +0 -224
- package/dist/extraction/writePathAdjudication.js +0 -110
- package/dist/importing/apply.js +0 -672
- package/dist/importing/bundle-write.js +0 -80
- package/dist/importing/cli.js +0 -408
- package/dist/importing/existing.js +0 -51
- package/dist/importing/facts.js +0 -121
- package/dist/importing/forgetImported.js +0 -79
- package/dist/importing/hosted.js +0 -467
- package/dist/importing/journalLog.js +0 -96
- package/dist/importing/markdown.js +0 -98
- package/dist/importing/plan.js +0 -201
- package/dist/importing/secretScreen.js +0 -113
- package/dist/importing/serialize.js +0 -188
- package/dist/importing/source.js +0 -463
- package/dist/importing/types.js +0 -6
- package/dist/indexer/chunker.js +0 -319
- package/dist/indexer/dailyEntries.js +0 -63
- package/dist/indexer/hasher.js +0 -10
- package/dist/indexer/laneRuns.js +0 -48
- package/dist/indexer/o200k.js +0 -481
- package/dist/indexer/outsideRoot.js +0 -77
- package/dist/indexer/parser.js +0 -170
- package/dist/indexer/pipeline.js +0 -988
- package/dist/indexer/staleness.js +0 -201
- package/dist/indexer/tokenizer.js +0 -77
- package/dist/indexer/walker.js +0 -170
- package/dist/learning/cli.js +0 -113
- package/dist/learning/schema.js +0 -144
- package/dist/learning/settings.js +0 -45
- package/dist/learning/store.js +0 -267
- package/dist/learning/tap.js +0 -401
- package/dist/mcp/hive.js +0 -88
- package/dist/mcp/requestErrors.js +0 -47
- package/dist/mcp/server.js +0 -426
- package/dist/memory/agentLease.js +0 -43
- package/dist/memory/confinedWrite.js +0 -315
- package/dist/memory/controls.js +0 -73
- package/dist/memory/dailyNote.js +0 -265
- package/dist/memory/dreamer/adjudicate.js +0 -215
- package/dist/memory/dreamer/apply.js +0 -120
- package/dist/memory/dreamer/budget.js +0 -64
- package/dist/memory/dreamer/candidates.js +0 -149
- package/dist/memory/dreamer/collection.js +0 -104
- package/dist/memory/dreamer/dreamer.js +0 -352
- package/dist/memory/dreamer/review.js +0 -111
- package/dist/memory/durableAppend.js +0 -57
- package/dist/memory/factConflicts.js +0 -205
- package/dist/memory/factTerms.js +0 -321
- package/dist/memory/handoff.js +0 -363
- package/dist/memory/hiveNameHeader.cjs +0 -10
- package/dist/memory/inventory.js +0 -55
- package/dist/memory/invisibleText.js +0 -20
- package/dist/memory/lifecycle.js +0 -2014
- package/dist/memory/limits.js +0 -13
- package/dist/memory/requests.js +0 -183
- package/dist/memory/sourceName.js +0 -276
- package/dist/memory/statusFacts.js +0 -61
- package/dist/memory/whatsNew.js +0 -158
- package/dist/memory/writer.js +0 -176
- package/dist/memory/writerConflicts.js +0 -85
- package/dist/memorymap/memoryMap.js +0 -1594
- package/dist/memorymap/replacedJournal.js +0 -58
- package/dist/memorymap/replacements.js +0 -410
- package/dist/mvp/activity.js +0 -86
- package/dist/mvp/agentMemory.js +0 -1803
- package/dist/mvp/answerCheck.js +0 -211
- package/dist/mvp/auth.js +0 -531
- package/dist/mvp/cli.js +0 -86
- package/dist/mvp/client-store.js +0 -462
- package/dist/mvp/compileCache.js +0 -242
- package/dist/mvp/compileSeed.js +0 -33
- package/dist/mvp/config.js +0 -292
- package/dist/mvp/connectionStatus.js +0 -80
- package/dist/mvp/contextDistinct.js +0 -53
- package/dist/mvp/contextProfile.js +0 -87
- package/dist/mvp/contextRelevance.js +0 -413
- package/dist/mvp/directorySurface.js +0 -68
- package/dist/mvp/engine.js +0 -1078
- package/dist/mvp/exportAll.js +0 -393
- package/dist/mvp/factLanePack.js +0 -77
- package/dist/mvp/factMarkers.js +0 -65
- package/dist/mvp/factsLane.js +0 -203
- package/dist/mvp/hiddenSpans.js +0 -585
- package/dist/mvp/history.js +0 -414
- package/dist/mvp/hitDates.js +0 -31
- package/dist/mvp/hitFloor.js +0 -36
- package/dist/mvp/identity.js +0 -18
- package/dist/mvp/idle.js +0 -113
- package/dist/mvp/log.js +0 -118
- package/dist/mvp/memoryDir.js +0 -98
- package/dist/mvp/posthogMcp.js +0 -210
- package/dist/mvp/readScope.js +0 -104
- package/dist/mvp/readUsefulness.js +0 -222
- package/dist/mvp/sanitize.js +0 -48
- package/dist/mvp/schemaRejections.js +0 -91
- package/dist/mvp/scope.js +0 -127
- package/dist/mvp/server.js +0 -2712
- package/dist/mvp/serverCard.js +0 -38
- package/dist/mvp/storedSize.js +0 -139
- package/dist/mvp/telemetry.js +0 -474
- package/dist/net/poolFetch.js +0 -174
- package/dist/recall/actionRecall.js +0 -282
- package/dist/recall/autoRecall.js +0 -447
- package/dist/recall/searchRecall.js +0 -189
- package/dist/retrieval/adviceLines.js +0 -195
- package/dist/retrieval/adviceQuestion.js +0 -44
- package/dist/retrieval/conversationShape.js +0 -44
- package/dist/retrieval/dailyEntryTime.js +0 -55
- package/dist/retrieval/entityLookup.js +0 -162
- package/dist/retrieval/factsFirst.js +0 -188
- package/dist/retrieval/keyword.js +0 -84
- package/dist/retrieval/ledgerEventTime.js +0 -51
- package/dist/retrieval/ledgerLiveness.js +0 -83
- package/dist/retrieval/ledgerPresentation.js +0 -161
- package/dist/retrieval/noteIdentity.js +0 -57
- package/dist/retrieval/orchestrator.js +0 -1224
- package/dist/retrieval/plainWording.js +0 -236
- package/dist/retrieval/queryHelpers.js +0 -142
- package/dist/retrieval/questionSpan.js +0 -475
- package/dist/retrieval/rankingTrace.js +0 -184
- package/dist/retrieval/readerEvents.js +0 -690
- package/dist/retrieval/readerLayout.js +0 -422
- package/dist/retrieval/relativeTimeHints.js +0 -113
- package/dist/retrieval/semantic.js +0 -67
- package/dist/router/routerCache.js +0 -106
- package/dist/router/sessionRouter.js +0 -726
- package/dist/sharing/cli.js +0 -181
- package/dist/sharing/examples.js +0 -136
- package/dist/sharing/fields.js +0 -225
- package/dist/sharing/loop.js +0 -80
- package/dist/sharing/project.js +0 -327
- package/dist/sharing/send.js +0 -466
- package/dist/sharing/state.js +0 -222
- package/dist/sharing/wording.js +0 -48
- package/dist/toolRefusal.js +0 -120
- package/dist/watcher/watcher.js +0 -77
- /package/dist/{indexer/o200k.json → o200k.json} +0 -0
|
@@ -1,222 +0,0 @@
|
|
|
1
|
-
"use strict";
|
|
2
|
-
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
-
exports.ReadUsefulness = exports.DEFAULT_USE_WINDOW_MINUTES = void 0;
|
|
4
|
-
exports.noteItem = noteItem;
|
|
5
|
-
exports.factItem = factItem;
|
|
6
|
-
exports.contextItems = contextItems;
|
|
7
|
-
exports.searchItems = searchItems;
|
|
8
|
-
exports.retrieveItems = retrieveItems;
|
|
9
|
-
const node_crypto_1 = require("node:crypto");
|
|
10
|
-
const statements_1 = require("../db/statements");
|
|
11
|
-
const ledgerLiveness_1 = require("../retrieval/ledgerLiveness");
|
|
12
|
-
const telemetry_1 = require("./telemetry");
|
|
13
|
-
/** How long after a read a use still counts. Long enough for a multi-step task, short enough to stay one working session. */
|
|
14
|
-
exports.DEFAULT_USE_WINDOW_MINUTES = 30;
|
|
15
|
-
const MAX_WINDOW_MINUTES = 240;
|
|
16
|
-
/** Reads kept to be matched, so a busy workspace cannot grow the tracker without bound. */
|
|
17
|
-
const MAX_TRACKED_READS = 300;
|
|
18
|
-
/** Ids an event lists; `items` still counts them all. */
|
|
19
|
-
const MAX_EVENT_IDS = 20;
|
|
20
|
-
const norm = (value) => value.normalize('NFKC').toLowerCase().replace(/\s+/g, ' ').trim();
|
|
21
|
-
function noteItem(notePath) {
|
|
22
|
-
return { key: `note:${notePath}` };
|
|
23
|
-
}
|
|
24
|
-
function factItem(subject, predicate) {
|
|
25
|
-
return { key: `fact:${norm(subject)}\0${norm(predicate)}`, handles: [`subject:${norm(subject)}`] };
|
|
26
|
-
}
|
|
27
|
-
/** The note a memory id names (`note:<path>`, `entry:<path>#<hash>`), or the text itself when it names none. */
|
|
28
|
-
function noteOfId(id) {
|
|
29
|
-
if (id.startsWith('note:'))
|
|
30
|
-
return id.slice(5);
|
|
31
|
-
if (id.startsWith('entry:')) {
|
|
32
|
-
const hashAt = id.lastIndexOf('#');
|
|
33
|
-
return hashAt > 6 ? id.slice(6, hashAt) : id.slice(6);
|
|
34
|
-
}
|
|
35
|
-
return id;
|
|
36
|
-
}
|
|
37
|
-
function windowMs(env) {
|
|
38
|
-
const parsed = Number((env.RENOR_READ_USE_WINDOW_MINUTES ?? '').trim());
|
|
39
|
-
const minutes = Number.isFinite(parsed) && parsed >= 1 ? Math.min(MAX_WINDOW_MINUTES, Math.trunc(parsed)) : exports.DEFAULT_USE_WINDOW_MINUTES;
|
|
40
|
-
return minutes * 60_000;
|
|
41
|
-
}
|
|
42
|
-
class ReadUsefulness {
|
|
43
|
-
emit;
|
|
44
|
-
now;
|
|
45
|
-
reads = [];
|
|
46
|
-
secret = (0, node_crypto_1.randomBytes)(32);
|
|
47
|
-
windowMs;
|
|
48
|
-
constructor(emit = telemetry_1.emitTelemetry, now = Date.now, env = process.env) {
|
|
49
|
-
this.emit = emit;
|
|
50
|
-
this.now = now;
|
|
51
|
-
this.windowMs = windowMs(env);
|
|
52
|
-
}
|
|
53
|
-
opaque(key) {
|
|
54
|
-
return (0, node_crypto_1.createHmac)('sha256', this.secret).update(key).digest('hex').slice(0, 8);
|
|
55
|
-
}
|
|
56
|
-
/** Builds what a reply served and records it. Building reads the database, so it is inside the guard too. */
|
|
57
|
-
measure(tool, build) {
|
|
58
|
-
try {
|
|
59
|
-
const served = build();
|
|
60
|
-
this.served(tool, served.items, served.relevant, served.judged);
|
|
61
|
-
}
|
|
62
|
-
catch {
|
|
63
|
-
// Measuring must never break a read.
|
|
64
|
-
}
|
|
65
|
-
}
|
|
66
|
-
/** A reply was served. `items` are its returned items; `relevant` and `judged` as described above. */
|
|
67
|
-
served(tool, items, relevant, judged) {
|
|
68
|
-
try {
|
|
69
|
-
const at = this.now();
|
|
70
|
-
const unique = new Map();
|
|
71
|
-
for (const item of items)
|
|
72
|
-
if (!unique.has(item.key))
|
|
73
|
-
unique.set(item.key, item);
|
|
74
|
-
const id = (0, node_crypto_1.randomBytes)(4).toString('hex');
|
|
75
|
-
this.emit({
|
|
76
|
-
event: 'read',
|
|
77
|
-
tool,
|
|
78
|
-
read: id,
|
|
79
|
-
relevant,
|
|
80
|
-
judged,
|
|
81
|
-
items: unique.size,
|
|
82
|
-
ids: [...unique.keys()].slice(0, MAX_EVENT_IDS).map(key => this.opaque(key)),
|
|
83
|
-
});
|
|
84
|
-
if (!relevant || unique.size === 0)
|
|
85
|
-
return;
|
|
86
|
-
this.prune(at);
|
|
87
|
-
this.reads.push({
|
|
88
|
-
id,
|
|
89
|
-
tool,
|
|
90
|
-
at,
|
|
91
|
-
keys: new Set(unique.keys()),
|
|
92
|
-
handles: new Set([...unique.values()].flatMap(item => item.handles ?? [])),
|
|
93
|
-
used: false,
|
|
94
|
-
});
|
|
95
|
-
if (this.reads.length > MAX_TRACKED_READS)
|
|
96
|
-
this.reads.splice(0, this.reads.length - MAX_TRACKED_READS);
|
|
97
|
-
}
|
|
98
|
-
catch {
|
|
99
|
-
// Measuring must never break a read.
|
|
100
|
-
}
|
|
101
|
-
}
|
|
102
|
-
/** The agent read one memory in full (read_memory). */
|
|
103
|
-
readInFull(memoryId) {
|
|
104
|
-
this.use('read', read => read.keys.has(`note:${noteOfId(memoryId)}`));
|
|
105
|
-
}
|
|
106
|
-
/** A save named what it builds on. Each cite is a notePath or memory id a read returned, or a fact's subject. */
|
|
107
|
-
cited(cites) {
|
|
108
|
-
if (!cites?.length)
|
|
109
|
-
return;
|
|
110
|
-
const wanted = cites.slice(0, 20).map(cite => cite.trim()).filter(Boolean);
|
|
111
|
-
this.use('cite', read => wanted.some(cite => read.keys.has(`note:${noteOfId(cite)}`) || read.handles.has(`subject:${norm(cite)}`)));
|
|
112
|
-
}
|
|
113
|
-
/** A fact was saved with a new value for a subject and predicate (a correction or an update). */
|
|
114
|
-
updated(subject, predicate) {
|
|
115
|
-
const key = factItem(subject, predicate).key;
|
|
116
|
-
this.use('update', read => read.keys.has(key));
|
|
117
|
-
}
|
|
118
|
-
use(via, matches) {
|
|
119
|
-
try {
|
|
120
|
-
const at = this.now();
|
|
121
|
-
this.prune(at);
|
|
122
|
-
for (const read of this.reads) {
|
|
123
|
-
if (read.used || !matches(read))
|
|
124
|
-
continue;
|
|
125
|
-
read.used = true;
|
|
126
|
-
this.emit({
|
|
127
|
-
event: 'read_used',
|
|
128
|
-
tool: read.tool,
|
|
129
|
-
read: read.id,
|
|
130
|
-
via,
|
|
131
|
-
after_s: Math.max(0, Math.round((at - read.at) / 1000)),
|
|
132
|
-
});
|
|
133
|
-
}
|
|
134
|
-
}
|
|
135
|
-
catch {
|
|
136
|
-
// Measuring must never break a save.
|
|
137
|
-
}
|
|
138
|
-
}
|
|
139
|
-
prune(at) {
|
|
140
|
-
const cutoff = at - this.windowMs;
|
|
141
|
-
let drop = 0;
|
|
142
|
-
while (drop < this.reads.length && this.reads[drop].at < cutoff)
|
|
143
|
-
drop++;
|
|
144
|
-
if (drop > 0)
|
|
145
|
-
this.reads.splice(0, drop);
|
|
146
|
-
}
|
|
147
|
-
}
|
|
148
|
-
exports.ReadUsefulness = ReadUsefulness;
|
|
149
|
-
// ---- What each tool's reply serves ---------------------------------------------------------------------------------
|
|
150
|
-
/** get_context: its facts by subject and predicate, its memories by note. A memory of a saved fact names no note. */
|
|
151
|
-
function contextItems(body) {
|
|
152
|
-
const items = [];
|
|
153
|
-
for (const fact of Array.isArray(body.facts) ? body.facts : []) {
|
|
154
|
-
if (typeof fact.subject === 'string' && typeof fact.predicate === 'string')
|
|
155
|
-
items.push(factItem(fact.subject, fact.predicate));
|
|
156
|
-
}
|
|
157
|
-
for (const memory of Array.isArray(body.memories) ? body.memories : []) {
|
|
158
|
-
if (typeof memory.notePath === 'string' && memory.notePath !== '')
|
|
159
|
-
items.push(noteItem(memory.notePath));
|
|
160
|
-
else if (typeof memory.text === 'string')
|
|
161
|
-
items.push({ key: `text:${memory.text}` });
|
|
162
|
-
}
|
|
163
|
-
return verdictOf(body.relevance, items);
|
|
164
|
-
}
|
|
165
|
-
function verdictOf(relevance, items) {
|
|
166
|
-
const judged = typeof relevance === 'object' && relevance !== null && typeof relevance.relevant === 'boolean';
|
|
167
|
-
return { items, relevant: judged ? relevance.relevant : items.length > 0, judged };
|
|
168
|
-
}
|
|
169
|
-
/**
|
|
170
|
-
* search_memory: its hits as they are before the ledger's markup is taken off, so a fact is named by what it claims. A hit
|
|
171
|
-
* of the ledger names the facts of its events (the facts table knows each event's subject and predicate); anything else is
|
|
172
|
-
* a note.
|
|
173
|
-
*/
|
|
174
|
-
function searchItems(db, isLedger, hits, relevance) {
|
|
175
|
-
const events = new Set();
|
|
176
|
-
for (const hit of hits)
|
|
177
|
-
if (isLedger(hit.notePath))
|
|
178
|
-
for (const id of (0, ledgerLiveness_1.ledgerEventIds)(hit))
|
|
179
|
-
events.add(id);
|
|
180
|
-
const named = new Map();
|
|
181
|
-
if (events.size > 0) {
|
|
182
|
-
const ids = [...events];
|
|
183
|
-
const rows = (0, statements_1.listStatement)(db, ids.length, marks => `SELECT lower(f.event_id) AS event, e.canonical_name AS subject, f.predicate AS predicate
|
|
184
|
-
FROM facts f JOIN entities e ON e.id = f.subject_entity_id
|
|
185
|
-
WHERE lower(f.event_id) IN (${marks})`).all(...(0, statements_1.paddedList)(ids));
|
|
186
|
-
for (const row of rows)
|
|
187
|
-
named.set(row.event, row);
|
|
188
|
-
}
|
|
189
|
-
const items = [];
|
|
190
|
-
for (const hit of hits) {
|
|
191
|
-
if (!isLedger(hit.notePath)) {
|
|
192
|
-
items.push(hit.notePath ? noteItem(hit.notePath) : { key: `text:${hit.text}` });
|
|
193
|
-
continue;
|
|
194
|
-
}
|
|
195
|
-
const facts = (0, ledgerLiveness_1.ledgerEventIds)(hit).map(id => named.get(id)).filter((fact) => fact !== undefined);
|
|
196
|
-
if (facts.length > 0) {
|
|
197
|
-
for (const fact of facts)
|
|
198
|
-
items.push(factItem(fact.subject, fact.predicate));
|
|
199
|
-
continue;
|
|
200
|
-
}
|
|
201
|
-
// A fact the index has not caught up with is headed `subject > predicate`.
|
|
202
|
-
const [subject, predicate] = (hit.headingPath ?? '').split(' > ');
|
|
203
|
-
items.push(subject && predicate ? factItem(subject, predicate) : { key: `text:${hit.text}` });
|
|
204
|
-
}
|
|
205
|
-
return verdictOf(relevance, items);
|
|
206
|
-
}
|
|
207
|
-
/** retrieve_memory: the facts of the subject it was asked for, by the subject the facts table names them under. */
|
|
208
|
-
function retrieveItems(db, facts) {
|
|
209
|
-
const ids = [...new Set(facts.map(f => f.subject_entity_id).filter((id) => typeof id === 'number'))];
|
|
210
|
-
const names = new Map();
|
|
211
|
-
if (ids.length > 0) {
|
|
212
|
-
const rows = db.prepare(`SELECT id, canonical_name AS name FROM entities WHERE id IN (${ids.map(() => '?').join(',')})`)
|
|
213
|
-
.all(...ids);
|
|
214
|
-
for (const row of rows)
|
|
215
|
-
names.set(row.id, row.name);
|
|
216
|
-
}
|
|
217
|
-
const items = facts.flatMap(f => {
|
|
218
|
-
const subject = f.subject_entity_id === undefined ? undefined : names.get(f.subject_entity_id);
|
|
219
|
-
return subject === undefined ? [] : [factItem(subject, f.predicate)];
|
|
220
|
-
});
|
|
221
|
-
return { items, relevant: items.length > 0, judged: false };
|
|
222
|
-
}
|
package/dist/mvp/sanitize.js
DELETED
|
@@ -1,48 +0,0 @@
|
|
|
1
|
-
"use strict";
|
|
2
|
-
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
-
exports.sanitizeErrorMessage = sanitizeErrorMessage;
|
|
4
|
-
exports.escapeInvisibleControls = escapeInvisibleControls;
|
|
5
|
-
/**
|
|
6
|
-
* Privacy-safe error sanitizer: redacts host filesystem paths and bounds message length.
|
|
7
|
-
* Preserves error descriptions, status codes, and syscalls without disclosing server directories.
|
|
8
|
-
*/
|
|
9
|
-
function sanitizeErrorMessage(err) {
|
|
10
|
-
let message = err instanceof Error ? err.message : String(err);
|
|
11
|
-
if (err && typeof err === 'object' && 'path' in err && typeof err.path === 'string' && err.path.length > 0) {
|
|
12
|
-
message = message.split(err.path).join('[redacted-path]');
|
|
13
|
-
}
|
|
14
|
-
// Windows absolute paths: C:\...
|
|
15
|
-
const winPath = /(?:^|(?<=[\s(]))(['"`]?)[a-zA-Z]:\\(?:[^\s'"`:,;()\[\]\\]+\\)*[^\s'"`:,;()\[\]]*\1/g;
|
|
16
|
-
// Unix absolute paths: starts with / and has at least two segments (e.g. /home/user/vault or /tmp/x)
|
|
17
|
-
const unixPath = /(?:^|(?<=[\s(]))(['"`]?)\/(?:[^\s'"`:,;()\[\]/]+\/)+[^\s'"`:,;()\[\]]*\1/g;
|
|
18
|
-
return message.replace(winPath, '[redacted-path]').replace(unixPath, '[redacted-path]');
|
|
19
|
-
}
|
|
20
|
-
/**
|
|
21
|
-
* Characters JSON.stringify leaves literal although a reader may treat them as structure rather than
|
|
22
|
-
* data: the C1 controls (U+0085 NEXT LINE among them), the Unicode format controls — bidirectional
|
|
23
|
-
* overrides and zero-width characters a stored memory can use to reorder or hide text — and the line
|
|
24
|
-
* and paragraph separators. A memory is untrusted text; emitted raw, a bidirectional override makes
|
|
25
|
-
* the reply read backwards in an agent's transcript and a zero-width or line separator can hide an
|
|
26
|
-
* instruction from a human reviewer while the model still tokenizes it.
|
|
27
|
-
*
|
|
28
|
-
* Escaping them to a plain `\uXXXX` escape on the wire changes no tool's contract: JSON.stringify
|
|
29
|
-
* already escapes every C0 control, any JSON parser reads the same string back, so the decoded value
|
|
30
|
-
* is byte-identical while the raw reply an agent or a log reads shows the hidden character as an
|
|
31
|
-
* escape instead of applying it. Mirrors lifecycle.ts JSON_LEAVES_LITERAL, which does the same for the
|
|
32
|
-
* ledger's statement field. Surrogate pairs (astral format characters) are escaped unit by unit.
|
|
33
|
-
*
|
|
34
|
-
* The categories above are not the whole set a reader cannot see. Variation selectors (Mn: U+FE00–FE0F and the
|
|
35
|
-
* supplement U+E0100–E01EF, which spell out bytes an instruction can hide in) and blank fillers (Lo, So or Mn: the Hangul
|
|
36
|
-
* fillers, the braille blank, the combining grapheme joiner) are invisible in a transcript yet pass the categories
|
|
37
|
-
* above, so they are listed by code point (adv report renor-adv-inj6b-1012, F4). Emoji presentation selectors (U+FE0F
|
|
38
|
-
* after a symbol) are escaped too: still lossless after a JSON parse, a few bytes longer in the raw reply.
|
|
39
|
-
*/
|
|
40
|
-
const OUTPUT_INVISIBLE = new RegExp('[\\p{Cc}\\p{Cf}\\p{Zl}\\p{Zp}'
|
|
41
|
-
+ '\\u034F\\u115F\\u1160\\u2800\\u3164\\uFE00-\\uFE0F\\uFFA0\\u{E0100}-\\u{E01EF}]', 'gu');
|
|
42
|
-
/** A text of printable ASCII alone holds none of the characters above: found by a plain character class, not by the Unicode properties. */
|
|
43
|
-
const PRINTABLE_ASCII = /^[\x20-\x7e]*$/;
|
|
44
|
-
function escapeInvisibleControls(text) {
|
|
45
|
-
if (PRINTABLE_ASCII.test(text))
|
|
46
|
-
return text;
|
|
47
|
-
return text.replace(OUTPUT_INVISIBLE, ch => Array.from({ length: ch.length }, (_, i) => `\\u${ch.charCodeAt(i).toString(16).padStart(4, '0')}`).join(''));
|
|
48
|
-
}
|
|
@@ -1,91 +0,0 @@
|
|
|
1
|
-
"use strict";
|
|
2
|
-
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
-
exports.countSchemaRejections = countSchemaRejections;
|
|
4
|
-
const types_js_1 = require("@modelcontextprotocol/sdk/types.js");
|
|
5
|
-
const zod_compat_js_1 = require("@modelcontextprotocol/sdk/server/zod-compat.js");
|
|
6
|
-
const telemetry_1 = require("./telemetry");
|
|
7
|
-
/**
|
|
8
|
-
* Wraps `server` so that each call its tools' schemas refuse is counted as one `tool_rejected` event. `onRefused` hears the
|
|
9
|
-
* tool's name for each such call, before the event is written: a refused call is still a call (a session's first call).
|
|
10
|
-
*/
|
|
11
|
-
function countSchemaRejections(server, onRefused) {
|
|
12
|
-
const sdk = server;
|
|
13
|
-
const validate = sdk.validateToolInput;
|
|
14
|
-
if (typeof validate !== 'function')
|
|
15
|
-
return;
|
|
16
|
-
sdk.validateToolInput = async (tool, args, toolName) => {
|
|
17
|
-
try {
|
|
18
|
-
return await validate.call(server, tool, args, toolName);
|
|
19
|
-
}
|
|
20
|
-
catch (err) {
|
|
21
|
-
// Inside this method the SDK throws InvalidParams for one thing only: the arguments failed the schema (an unknown or
|
|
22
|
-
// disabled tool is refused before it is reached). Anything else is a defect of the schema itself, not a refusal.
|
|
23
|
-
if (err?.code === types_js_1.ErrorCode.InvalidParams) {
|
|
24
|
-
onRefused?.(toolName);
|
|
25
|
-
await recordRejection(tool, args, toolName);
|
|
26
|
-
}
|
|
27
|
-
throw err;
|
|
28
|
-
}
|
|
29
|
-
};
|
|
30
|
-
}
|
|
31
|
-
async function recordRejection(tool, args, toolName) {
|
|
32
|
-
// The SDK has refused the call, so it is counted whatever happens next: a failure to say more leaves the kind at `other`.
|
|
33
|
-
let described = { kind: 'other' };
|
|
34
|
-
try {
|
|
35
|
-
described = await describeFailure(tool, args);
|
|
36
|
-
}
|
|
37
|
-
catch {
|
|
38
|
-
// Observability must never change what the caller reads.
|
|
39
|
-
}
|
|
40
|
-
(0, telemetry_1.emitTelemetry)({ event: 'tool_rejected', tool: toolName, ...described });
|
|
41
|
-
}
|
|
42
|
-
/**
|
|
43
|
-
* Which argument the schema refused first and how, read from the issues the SDK's own parse yields (the same function, on
|
|
44
|
-
* the same schema and the same arguments, so the same verdict). Issues follow the schema's own order, so "first" is stable.
|
|
45
|
-
*/
|
|
46
|
-
async function describeFailure(tool, args) {
|
|
47
|
-
if (!tool.inputSchema)
|
|
48
|
-
return { kind: 'other' };
|
|
49
|
-
const parsed = await (0, zod_compat_js_1.safeParseAsync)(tool.inputSchema, args);
|
|
50
|
-
if (parsed.success)
|
|
51
|
-
return { kind: 'other' };
|
|
52
|
-
const issue = firstIssue(parsed.error);
|
|
53
|
-
if (!issue)
|
|
54
|
-
return { kind: 'other' };
|
|
55
|
-
// Only a name the tool's own schema declares can be an argument that failed: a key the caller invented is not one. A declared
|
|
56
|
-
// name that is not a plain identifier is left out too, so the call is still counted instead of the whole line being dropped.
|
|
57
|
-
const shape = (0, zod_compat_js_1.getObjectShape)(tool.inputSchema);
|
|
58
|
-
const head = issue.path[0];
|
|
59
|
-
const named = typeof head === 'string' && shape !== undefined && Object.prototype.hasOwnProperty.call(shape, head) && (0, telemetry_1.isSafeArgName)(head);
|
|
60
|
-
return { ...(named ? { arg: head } : {}), kind: kindOf(issue, args) };
|
|
61
|
-
}
|
|
62
|
-
function firstIssue(error) {
|
|
63
|
-
const issues = error?.issues;
|
|
64
|
-
if (!Array.isArray(issues))
|
|
65
|
-
return undefined;
|
|
66
|
-
const first = issues[0];
|
|
67
|
-
if (!first || typeof first.code !== 'string' || !Array.isArray(first.path))
|
|
68
|
-
return undefined;
|
|
69
|
-
return { code: first.code, path: first.path };
|
|
70
|
-
}
|
|
71
|
-
function kindOf(issue, args) {
|
|
72
|
-
switch (issue.code) {
|
|
73
|
-
// The schema says "invalid_type" both for an absent argument and for one of the wrong type; the call says which.
|
|
74
|
-
case 'invalid_type': return valueAt(args, issue.path) === undefined ? 'missing' : 'wrong_type';
|
|
75
|
-
case 'too_small': return 'too_small';
|
|
76
|
-
case 'too_big': return 'too_big';
|
|
77
|
-
case 'invalid_value': return 'invalid_value';
|
|
78
|
-
case 'invalid_format': return 'invalid_format';
|
|
79
|
-
default: return 'other';
|
|
80
|
-
}
|
|
81
|
-
}
|
|
82
|
-
/** Whatever sits at `path` in what the caller sent. Only ever compared with `undefined`: it is never kept or written. */
|
|
83
|
-
function valueAt(root, path) {
|
|
84
|
-
let value = root;
|
|
85
|
-
for (const key of path) {
|
|
86
|
-
if (value === null || typeof value !== 'object')
|
|
87
|
-
return undefined;
|
|
88
|
-
value = value[key];
|
|
89
|
-
}
|
|
90
|
-
return value;
|
|
91
|
-
}
|
package/dist/mvp/scope.js
DELETED
|
@@ -1,127 +0,0 @@
|
|
|
1
|
-
"use strict";
|
|
2
|
-
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
-
exports.normalizeScope = normalizeScope;
|
|
4
|
-
exports.dailyDirForScope = dailyDirForScope;
|
|
5
|
-
exports.notePathMatchesScope = notePathMatchesScope;
|
|
6
|
-
exports.scopeSqlPredicate = scopeSqlPredicate;
|
|
7
|
-
exports.scopeSpellings = scopeSpellings;
|
|
8
|
-
exports.savedScopeMatches = savedScopeMatches;
|
|
9
|
-
exports.savedScopeSqlPredicate = savedScopeSqlPredicate;
|
|
10
|
-
/**
|
|
11
|
-
* Optional memory scope for MVP tools (global vs named project/scope).
|
|
12
|
-
*
|
|
13
|
-
* Allowed forms:
|
|
14
|
-
* - "global" (default)
|
|
15
|
-
* - "scope:<slug>" or "project:<slug>"
|
|
16
|
-
* Slug: lowercase [a-z0-9-]{1,64} after normalization.
|
|
17
|
-
*/
|
|
18
|
-
const toolRefusal_1 = require("../toolRefusal");
|
|
19
|
-
/** The one directory a scope owns. Read by both the JS and the SQL scope rules. */
|
|
20
|
-
const SCOPES_ROOT = 'Memory/Scopes/';
|
|
21
|
-
const SLUG_RE = /^[a-z0-9-]{1,64}$/;
|
|
22
|
-
/** Normalize user/agent scope input. Throws on invalid values. */
|
|
23
|
-
function normalizeScope(raw) {
|
|
24
|
-
if (raw === undefined || raw === null || raw.trim() === '')
|
|
25
|
-
return 'global';
|
|
26
|
-
const s = raw.trim().toLowerCase();
|
|
27
|
-
if (s === 'global')
|
|
28
|
-
return 'global';
|
|
29
|
-
let kind = null;
|
|
30
|
-
let slug = s;
|
|
31
|
-
if (s.startsWith('scope:')) {
|
|
32
|
-
kind = 'scope';
|
|
33
|
-
slug = s.slice('scope:'.length);
|
|
34
|
-
}
|
|
35
|
-
else if (s.startsWith('project:')) {
|
|
36
|
-
kind = 'project';
|
|
37
|
-
slug = s.slice('project:'.length);
|
|
38
|
-
}
|
|
39
|
-
else {
|
|
40
|
-
// Bare slug → scope:<slug>
|
|
41
|
-
kind = 'scope';
|
|
42
|
-
slug = s;
|
|
43
|
-
}
|
|
44
|
-
// Normalize slug: spaces/underscores → hyphens; strip invalid chars
|
|
45
|
-
slug = slug
|
|
46
|
-
.replace(/[_\s]+/g, '-')
|
|
47
|
-
.replace(/[^a-z0-9-]/g, '')
|
|
48
|
-
.replace(/-+/g, '-')
|
|
49
|
-
.replace(/^-|-$/g, '');
|
|
50
|
-
if (!SLUG_RE.test(slug)) {
|
|
51
|
-
// The refusal names its code for telemetry (a closed set, never built from this input); the message the
|
|
52
|
-
// agent reads is unchanged, and it is the agent's own argument that is quoted in it.
|
|
53
|
-
throw (0, toolRefusal_1.refusal)('invalid_scope', `invalid scope ${JSON.stringify(raw)}: use "global" or "scope:<slug>" / "project:<slug>" with slug matching [a-z0-9-]{1,64}`);
|
|
54
|
-
}
|
|
55
|
-
return `${kind}:${slug}`;
|
|
56
|
-
}
|
|
57
|
-
/** Daily-note directory under the vault for this scope. */
|
|
58
|
-
function dailyDirForScope(scope) {
|
|
59
|
-
if (scope === 'global')
|
|
60
|
-
return 'Memory/Daily';
|
|
61
|
-
const slug = scope.includes(':') ? scope.slice(scope.indexOf(':') + 1) : scope;
|
|
62
|
-
return `Memory/Scopes/${slug}/Daily`;
|
|
63
|
-
}
|
|
64
|
-
/** True if a vault-relative note path belongs to the given scope. */
|
|
65
|
-
function notePathMatchesScope(notePath, scope) {
|
|
66
|
-
const norm = notePath.replace(/\\/g, '/');
|
|
67
|
-
if (scope === 'global') {
|
|
68
|
-
// Global notes: not under Memory/Scopes/
|
|
69
|
-
if (norm.startsWith('Memory/Scopes/') || norm.includes('/Memory/Scopes/'))
|
|
70
|
-
return false;
|
|
71
|
-
return true;
|
|
72
|
-
}
|
|
73
|
-
const slug = scope.slice(scope.indexOf(':') + 1);
|
|
74
|
-
const prefix = `Memory/Scopes/${slug}/`;
|
|
75
|
-
return norm === prefix.slice(0, -1) || norm.startsWith(prefix) || norm.includes('/' + prefix);
|
|
76
|
-
}
|
|
77
|
-
/**
|
|
78
|
-
* The same rule as notePathMatchesScope, as a SQL predicate over a stored path
|
|
79
|
-
* column, for the callers that must filter or count rows without materializing
|
|
80
|
-
* every path. Each arm below is one arm of that function — the three `dir`
|
|
81
|
-
* terms are its equality, startsWith and includes cases — and
|
|
82
|
-
* tests/mvp/scope-isolation.test.ts pins them against each other, so a scope
|
|
83
|
-
* can never mean two different things depending on whether it was checked in
|
|
84
|
-
* SQL or in JS. `dir` is built from a validated slug, so it contains no LIKE
|
|
85
|
-
* wildcard. Bind with better-sqlite3's positional params in `params`.
|
|
86
|
-
*/
|
|
87
|
-
function scopeSqlPredicate(scope, column) {
|
|
88
|
-
if (scope === 'global') {
|
|
89
|
-
return {
|
|
90
|
-
sql: `(${column} NOT LIKE ? AND instr(${column}, ?) = 0)`,
|
|
91
|
-
params: [`${SCOPES_ROOT}%`, `/${SCOPES_ROOT}`],
|
|
92
|
-
};
|
|
93
|
-
}
|
|
94
|
-
const dir = `${SCOPES_ROOT}${scope.slice(scope.indexOf(':') + 1)}`;
|
|
95
|
-
return {
|
|
96
|
-
sql: `(instr(${column}, ?) > 0 OR ${column} LIKE ? OR ${column} = ?)`,
|
|
97
|
-
params: [`/${dir}/`, `${dir}/%`, dir],
|
|
98
|
-
};
|
|
99
|
-
}
|
|
100
|
-
/**
|
|
101
|
-
* The scopes a fact or a ledger event can have been saved in and belong to `scope`: the one rule notes follow, for
|
|
102
|
-
* what is not a note. A named scope is its slug. `scope:<slug>`, `project:<slug>` and the bare `<slug>` are spellings
|
|
103
|
-
* of one scope: a note saved in any of them goes to the one folder of that slug (`dailyDirForScope`) and is read in
|
|
104
|
-
* all three (`notePathMatchesScope`). A fact keeps the spelling it was saved with (the ledger, its digest and the
|
|
105
|
-
* `scope` a read shows are what was written), and is read in every spelling of its slug, as its scope's notes are. A
|
|
106
|
-
* save corrects its subject and predicate in every spelling too: it closes the current value whatever spelling that
|
|
107
|
-
* was saved with (memory/lifecycle.ts, one lifecycle stream per slug).
|
|
108
|
-
* Global has one spelling. Distinct, so a caller that reads once per spelling never reads a scope twice.
|
|
109
|
-
*/
|
|
110
|
-
function scopeSpellings(scope) {
|
|
111
|
-
if (scope === 'global')
|
|
112
|
-
return ['global'];
|
|
113
|
-
const slug = scope.slice(scope.indexOf(':') + 1);
|
|
114
|
-
return [`scope:${slug}`, `project:${slug}`];
|
|
115
|
-
}
|
|
116
|
-
/** Whether a fact or ledger event saved in `saved` (a stored, normalized scope) belongs to `scope`. */
|
|
117
|
-
function savedScopeMatches(saved, scope) {
|
|
118
|
-
return scopeSpellings(scope).includes(saved);
|
|
119
|
-
}
|
|
120
|
-
/**
|
|
121
|
-
* The same rule as savedScopeMatches, as a SQL predicate over a stored scope column (`facts.scope`,
|
|
122
|
-
* `fact_terms.scope`). Bind with better-sqlite3's positional params in `params`.
|
|
123
|
-
*/
|
|
124
|
-
function savedScopeSqlPredicate(scope, column) {
|
|
125
|
-
const spellings = scopeSpellings(scope);
|
|
126
|
-
return { sql: `${column} IN (${spellings.map(() => '?').join(', ')})`, params: spellings };
|
|
127
|
-
}
|