agent-recall-core 3.4.48 → 3.4.50
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 +25 -0
- package/dist/digest/store.d.ts +15 -6
- package/dist/digest/store.d.ts.map +1 -1
- package/dist/digest/store.js +33 -20
- package/dist/digest/store.js.map +1 -1
- package/dist/embeddings/chunker.d.ts +102 -0
- package/dist/embeddings/chunker.d.ts.map +1 -0
- package/dist/embeddings/chunker.js +260 -0
- package/dist/embeddings/chunker.js.map +1 -0
- package/dist/embeddings/config.d.ts +93 -0
- package/dist/embeddings/config.d.ts.map +1 -0
- package/dist/embeddings/config.js +174 -0
- package/dist/embeddings/config.js.map +1 -0
- package/dist/embeddings/index-store.d.ts +61 -0
- package/dist/embeddings/index-store.d.ts.map +1 -0
- package/dist/embeddings/index-store.js +158 -0
- package/dist/embeddings/index-store.js.map +1 -0
- package/dist/embeddings/indexer.d.ts +75 -0
- package/dist/embeddings/indexer.d.ts.map +1 -0
- package/dist/embeddings/indexer.js +196 -0
- package/dist/embeddings/indexer.js.map +1 -0
- package/dist/embeddings/runtime.d.ts +66 -0
- package/dist/embeddings/runtime.d.ts.map +1 -0
- package/dist/embeddings/runtime.js +197 -0
- package/dist/embeddings/runtime.js.map +1 -0
- package/dist/helpers/associative-link.d.ts.map +1 -1
- package/dist/helpers/associative-link.js +26 -1
- package/dist/helpers/associative-link.js.map +1 -1
- package/dist/helpers/tokenize.d.ts +1 -0
- package/dist/helpers/tokenize.d.ts.map +1 -1
- package/dist/helpers/tokenize.js +18 -0
- package/dist/helpers/tokenize.js.map +1 -1
- package/dist/index.d.ts +27 -7
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +42 -5
- package/dist/index.js.map +1 -1
- package/dist/palace/awareness.d.ts +56 -10
- package/dist/palace/awareness.d.ts.map +1 -1
- package/dist/palace/awareness.js +279 -182
- package/dist/palace/awareness.js.map +1 -1
- package/dist/palace/consolidate.d.ts +1 -1
- package/dist/palace/consolidate.d.ts.map +1 -1
- package/dist/palace/consolidate.js +9 -9
- package/dist/palace/consolidate.js.map +1 -1
- package/dist/palace/decay-pass.d.ts +1 -1
- package/dist/palace/decay-pass.d.ts.map +1 -1
- package/dist/palace/decay-pass.js +2 -2
- package/dist/palace/decay-pass.js.map +1 -1
- package/dist/palace/fan-out.d.ts +1 -1
- package/dist/palace/fan-out.d.ts.map +1 -1
- package/dist/palace/fan-out.js +5 -5
- package/dist/palace/fan-out.js.map +1 -1
- package/dist/palace/index-manager.d.ts +1 -1
- package/dist/palace/index-manager.d.ts.map +1 -1
- package/dist/palace/index-manager.js +1 -1
- package/dist/palace/index-manager.js.map +1 -1
- package/dist/palace/insights-index.d.ts +19 -6
- package/dist/palace/insights-index.d.ts.map +1 -1
- package/dist/palace/insights-index.js +36 -13
- package/dist/palace/insights-index.js.map +1 -1
- package/dist/palace/keystone.d.ts +1 -1
- package/dist/palace/keystone.d.ts.map +1 -1
- package/dist/palace/keystone.js +2 -2
- package/dist/palace/keystone.js.map +1 -1
- package/dist/palace/rooms.d.ts +3 -3
- package/dist/palace/rooms.d.ts.map +1 -1
- package/dist/palace/rooms.js +34 -6
- package/dist/palace/rooms.js.map +1 -1
- package/dist/prompts/consolidation-prompt.d.ts.map +1 -1
- package/dist/prompts/consolidation-prompt.js +5 -0
- package/dist/prompts/consolidation-prompt.js.map +1 -1
- package/dist/retrieval/candidates.d.ts.map +1 -1
- package/dist/retrieval/candidates.js +17 -2
- package/dist/retrieval/candidates.js.map +1 -1
- package/dist/retrieval/query-memory.d.ts +65 -7
- package/dist/retrieval/query-memory.d.ts.map +1 -1
- package/dist/retrieval/query-memory.js +518 -78
- package/dist/retrieval/query-memory.js.map +1 -1
- package/dist/retrieval/semantic-leg.d.ts +74 -0
- package/dist/retrieval/semantic-leg.d.ts.map +1 -0
- package/dist/retrieval/semantic-leg.js +180 -0
- package/dist/retrieval/semantic-leg.js.map +1 -0
- package/dist/storage/archive-write.d.ts.map +1 -1
- package/dist/storage/archive-write.js +29 -8
- package/dist/storage/archive-write.js.map +1 -1
- package/dist/storage/capture-router.d.ts +1 -1
- package/dist/storage/capture-router.d.ts.map +1 -1
- package/dist/storage/capture-router.js +26 -9
- package/dist/storage/capture-router.js.map +1 -1
- package/dist/storage/consolidation-queue.d.ts +11 -2
- package/dist/storage/consolidation-queue.d.ts.map +1 -1
- package/dist/storage/consolidation-queue.js +45 -16
- package/dist/storage/consolidation-queue.js.map +1 -1
- package/dist/storage/corrections.d.ts +53 -5
- package/dist/storage/corrections.d.ts.map +1 -1
- package/dist/storage/corrections.js +412 -124
- package/dist/storage/corrections.js.map +1 -1
- package/dist/storage/dream-health.d.ts +56 -8
- package/dist/storage/dream-health.d.ts.map +1 -1
- package/dist/storage/dream-health.js +130 -35
- package/dist/storage/dream-health.js.map +1 -1
- package/dist/storage/dream-yield.d.ts +81 -0
- package/dist/storage/dream-yield.d.ts.map +1 -0
- package/dist/storage/dream-yield.js +77 -0
- package/dist/storage/dream-yield.js.map +1 -0
- package/dist/storage/durable-intent.d.ts.map +1 -1
- package/dist/storage/durable-intent.js +24 -1
- package/dist/storage/durable-intent.js.map +1 -1
- package/dist/storage/filelock.d.ts +92 -12
- package/dist/storage/filelock.d.ts.map +1 -1
- package/dist/storage/filelock.js +317 -47
- package/dist/storage/filelock.js.map +1 -1
- package/dist/storage/fs-utils.d.ts +20 -0
- package/dist/storage/fs-utils.d.ts.map +1 -1
- package/dist/storage/fs-utils.js +27 -0
- package/dist/storage/fs-utils.js.map +1 -1
- package/dist/storage/paths.d.ts +34 -0
- package/dist/storage/paths.d.ts.map +1 -1
- package/dist/storage/paths.js +71 -0
- package/dist/storage/paths.js.map +1 -1
- package/dist/storage/pending.d.ts +154 -0
- package/dist/storage/pending.d.ts.map +1 -0
- package/dist/storage/pending.js +454 -0
- package/dist/storage/pending.js.map +1 -0
- package/dist/storage/project.d.ts.map +1 -1
- package/dist/storage/project.js +68 -16
- package/dist/storage/project.js.map +1 -1
- package/dist/storage/session-card.d.ts +13 -0
- package/dist/storage/session-card.d.ts.map +1 -1
- package/dist/storage/session-card.js +34 -3
- package/dist/storage/session-card.js.map +1 -1
- package/dist/storage/unclaimed.d.ts +165 -0
- package/dist/storage/unclaimed.d.ts.map +1 -0
- package/dist/storage/unclaimed.js +442 -0
- package/dist/storage/unclaimed.js.map +1 -0
- package/dist/storage/working-memory.d.ts.map +1 -1
- package/dist/storage/working-memory.js +19 -1
- package/dist/storage/working-memory.js.map +1 -1
- package/dist/supabase/recall-backend.d.ts +28 -2
- package/dist/supabase/recall-backend.d.ts.map +1 -1
- package/dist/supabase/recall-backend.js +80 -23
- package/dist/supabase/recall-backend.js.map +1 -1
- package/dist/supabase/sync.d.ts.map +1 -1
- package/dist/supabase/sync.js +13 -0
- package/dist/supabase/sync.js.map +1 -1
- package/dist/tools-logic/alignment-check.js +1 -1
- package/dist/tools-logic/alignment-check.js.map +1 -1
- package/dist/tools-logic/awareness-update.d.ts.map +1 -1
- package/dist/tools-logic/awareness-update.js +25 -16
- package/dist/tools-logic/awareness-update.js.map +1 -1
- package/dist/tools-logic/check-action.js +1 -1
- package/dist/tools-logic/check-action.js.map +1 -1
- package/dist/tools-logic/check.d.ts +55 -1
- package/dist/tools-logic/check.d.ts.map +1 -1
- package/dist/tools-logic/check.js +230 -57
- package/dist/tools-logic/check.js.map +1 -1
- package/dist/tools-logic/confidence.d.ts +11 -6
- package/dist/tools-logic/confidence.d.ts.map +1 -1
- package/dist/tools-logic/confidence.js +11 -6
- package/dist/tools-logic/confidence.js.map +1 -1
- package/dist/tools-logic/context-synthesize.js +2 -2
- package/dist/tools-logic/context-synthesize.js.map +1 -1
- package/dist/tools-logic/digest-read.js +2 -2
- package/dist/tools-logic/digest-read.js.map +1 -1
- package/dist/tools-logic/digest-recall.js +1 -1
- package/dist/tools-logic/digest-recall.js.map +1 -1
- package/dist/tools-logic/dream-admission.d.ts +144 -0
- package/dist/tools-logic/dream-admission.d.ts.map +1 -0
- package/dist/tools-logic/dream-admission.js +537 -0
- package/dist/tools-logic/dream-admission.js.map +1 -0
- package/dist/tools-logic/insight-promotion.d.ts +19 -1
- package/dist/tools-logic/insight-promotion.d.ts.map +1 -1
- package/dist/tools-logic/insight-promotion.js +48 -14
- package/dist/tools-logic/insight-promotion.js.map +1 -1
- package/dist/tools-logic/journal-capture.js +3 -3
- package/dist/tools-logic/journal-capture.js.map +1 -1
- package/dist/tools-logic/journal-write.js +3 -3
- package/dist/tools-logic/journal-write.js.map +1 -1
- package/dist/tools-logic/knowledge-write.js +2 -2
- package/dist/tools-logic/knowledge-write.js.map +1 -1
- package/dist/tools-logic/palace-lint.js +1 -1
- package/dist/tools-logic/palace-lint.js.map +1 -1
- package/dist/tools-logic/palace-read.js +1 -1
- package/dist/tools-logic/palace-read.js.map +1 -1
- package/dist/tools-logic/palace-search.js +1 -1
- package/dist/tools-logic/palace-search.js.map +1 -1
- package/dist/tools-logic/palace-walk.js +3 -3
- package/dist/tools-logic/palace-walk.js.map +1 -1
- package/dist/tools-logic/palace-write.js +4 -4
- package/dist/tools-logic/palace-write.js.map +1 -1
- package/dist/tools-logic/predict-correction.js +1 -1
- package/dist/tools-logic/predict-correction.js.map +1 -1
- package/dist/tools-logic/recall-backend.d.ts +11 -2
- package/dist/tools-logic/recall-backend.d.ts.map +1 -1
- package/dist/tools-logic/recall-backend.js +2 -2
- package/dist/tools-logic/recall-backend.js.map +1 -1
- package/dist/tools-logic/resurrect.d.ts.map +1 -1
- package/dist/tools-logic/resurrect.js +9 -1
- package/dist/tools-logic/resurrect.js.map +1 -1
- package/dist/tools-logic/safety-consolidation.d.ts +15 -0
- package/dist/tools-logic/safety-consolidation.d.ts.map +1 -1
- package/dist/tools-logic/safety-consolidation.js +44 -31
- package/dist/tools-logic/safety-consolidation.js.map +1 -1
- package/dist/tools-logic/session-end-reflect.d.ts.map +1 -1
- package/dist/tools-logic/session-end-reflect.js +4 -1
- package/dist/tools-logic/session-end-reflect.js.map +1 -1
- package/dist/tools-logic/session-end.d.ts +8 -0
- package/dist/tools-logic/session-end.d.ts.map +1 -1
- package/dist/tools-logic/session-end.js +51 -13
- package/dist/tools-logic/session-end.js.map +1 -1
- package/dist/tools-logic/session-start.d.ts +23 -0
- package/dist/tools-logic/session-start.d.ts.map +1 -1
- package/dist/tools-logic/session-start.js +44 -1
- package/dist/tools-logic/session-start.js.map +1 -1
- package/dist/tools-logic/smart-recall.d.ts +185 -10
- package/dist/tools-logic/smart-recall.d.ts.map +1 -1
- package/dist/tools-logic/smart-recall.js +363 -85
- package/dist/tools-logic/smart-recall.js.map +1 -1
- package/dist/tools-logic/store-doctor.d.ts.map +1 -1
- package/dist/tools-logic/store-doctor.js +14 -5
- package/dist/tools-logic/store-doctor.js.map +1 -1
- package/dist/tools-logic/store-repair.js +1 -1
- package/dist/tools-logic/store-repair.js.map +1 -1
- package/dist/tools-logic/supersession.d.ts +1 -1
- package/dist/tools-logic/supersession.d.ts.map +1 -1
- package/dist/tools-logic/supersession.js +2 -2
- package/dist/tools-logic/supersession.js.map +1 -1
- package/dist/types.d.ts +1 -1
- package/dist/types.js +1 -1
- package/dist/vector/embedding.d.ts +8 -0
- package/dist/vector/embedding.d.ts.map +1 -1
- package/dist/vector/embedding.js +8 -0
- package/dist/vector/embedding.js.map +1 -1
- package/package.json +1 -1
|
@@ -10,8 +10,9 @@ import * as crypto from "node:crypto";
|
|
|
10
10
|
import { ensureDir } from "./fs-utils.js";
|
|
11
11
|
import { byteCap, sanitizeName } from "./sanitize.js";
|
|
12
12
|
import { journalDir, projectSubPath } from "./paths.js";
|
|
13
|
-
import { withLock } from "./filelock.js";
|
|
13
|
+
import { withLock, LockContentionError } from "./filelock.js";
|
|
14
14
|
import { scrubForCloud } from "./content-guard.js";
|
|
15
|
+
import { tokenizeWords, HAN_RUN_RE } from "../helpers/tokenize.js";
|
|
15
16
|
// ---------------------------------------------------------------------------
|
|
16
17
|
// Internal helpers
|
|
17
18
|
// ---------------------------------------------------------------------------
|
|
@@ -55,9 +56,42 @@ export const GATE_VERSION = "v4-2026-06-22";
|
|
|
55
56
|
* best-effort: a failure to rotate must never throw into writeCorrection.
|
|
56
57
|
*/
|
|
57
58
|
const REJECTED_LOG_CAP = 2000;
|
|
58
|
-
/**
|
|
59
|
-
|
|
60
|
-
|
|
59
|
+
/**
|
|
60
|
+
* Auto-detect severity: p0 if uses strong negation/mandate language, else p1.
|
|
61
|
+
*
|
|
62
|
+
* INDEPENDENT-REVIEW FIX (2026-09-09, TOW2-326 class): this pattern was
|
|
63
|
+
* English-only, so a CJK-only P0-caliber prohibition (禁止/不得/不能/不要 with
|
|
64
|
+
* no English marker) would now clear the newly CJK-aware capture gate above
|
|
65
|
+
* but still classify as p1 — silently losing the P0 blocking power (see
|
|
66
|
+
* check.ts's own doctrine comment: P0 severity is what lets a correction
|
|
67
|
+
* OVERRIDE a plan via `verdict: "blocked"`) an equivalent English phrasing
|
|
68
|
+
* would get. CJK rows added, mirroring STRONG_IMPERATIVE above; kept
|
|
69
|
+
* byte-identical to the duplicate `p0Patterns` copy in tools-logic/check.ts
|
|
70
|
+
* (search that file for `p0Patterns`, not a line number — line references
|
|
71
|
+
* drift; that copy is the primary caller, which computes severity itself
|
|
72
|
+
* before calling writeCorrection — this function is the FALLBACK path for
|
|
73
|
+
* callers that don't pre-supply severity).
|
|
74
|
+
*
|
|
75
|
+
* INDEPENDENT-REVIEW FIX (2026-09-09, round 3): 不能 scoped to 你不能 —
|
|
76
|
+
* see the identical fix/reasoning on STRONG_IMPERATIVE above.
|
|
77
|
+
*
|
|
78
|
+
* PRE-SHIP GATE FIX (2026-09-09, C-1): 不要's benign-reassurance exclusion
|
|
79
|
+
* widened from the narrow round-1 set (担心/客气/急/着急/紧张/见外) to the
|
|
80
|
+
* FULL round-2 set also used by GATED_PROHIBITION_PATTERNS and
|
|
81
|
+
* STRONG_IMPERATIVE (adds 慌/害怕/在意) — see CJK_REASSURANCE_COMPLETIONS'
|
|
82
|
+
* own doc comment. This copy had drifted out of sync with the other three
|
|
83
|
+
* sites; all four are now built from the SAME exported constant where they
|
|
84
|
+
* share a package (this file + check.ts), or kept byte-identical by
|
|
85
|
+
* cross-reference where they don't (correction-detector.ts, cli package).
|
|
86
|
+
*/
|
|
87
|
+
// Fix #2 (dual-channel capture gate, 2026-09-11): exported so the pending
|
|
88
|
+
// staging path (storage/pending.ts) stamps the SAME severity a direct capture
|
|
89
|
+
// would get — the S-M1 English injection twin pins "should not → p1" on the
|
|
90
|
+
// staged record, so staging and capture must never diverge on this classifier.
|
|
91
|
+
// Same internal-use-export convention as splitSentences/dropHardNoise below.
|
|
92
|
+
export function detectSeverity(text) {
|
|
93
|
+
const p0Patterns = new RegExp(`\\bnever\\b|\\balways\\b|\\bdon'?t\\b|\\bdo not\\b|\\bmust not\\b|\\bforbid\\b|\\bprohibit\\b|` +
|
|
94
|
+
`永远不要|绝不|千万不要|总是|一直|始终|不要(?!${CJK_REASSURANCE_COMPLETIONS})|不可以|不准|你不能(?!不)|不得(?!不)|不应该|切勿|禁止`, "i");
|
|
61
95
|
return p0Patterns.test(text) ? "p0" : "p1";
|
|
62
96
|
}
|
|
63
97
|
/**
|
|
@@ -164,7 +198,11 @@ function todayDate() {
|
|
|
164
198
|
* retractCorrection / recordOutcome) so every correction writer shares one
|
|
165
199
|
* durable path. Pure side-effect helper; no behavior change vs the originals.
|
|
166
200
|
*/
|
|
167
|
-
|
|
201
|
+
// Fix #2 (dual-channel capture gate, 2026-09-11): exported so the pending
|
|
202
|
+
// store (storage/pending.ts) writes through the SAME scrub-on-write choke
|
|
203
|
+
// point as corrections — rider R2a's "same scrub path" is a structural fact
|
|
204
|
+
// (one function), not a parallel implementation kept in sync by hand.
|
|
205
|
+
export function writeRecordAtomic(filepath, record) {
|
|
168
206
|
// Scrub BEFORE the local write — corrections are AgentRecall's most-injected
|
|
169
207
|
// artifact (readP0Corrections is loaded into every session_start briefing,
|
|
170
208
|
// handoff.md's "Binding rules" section, and check()'s/checkAction's
|
|
@@ -410,6 +448,18 @@ export function effectiveConfidenceOf(record) {
|
|
|
410
448
|
export function splitSentences(text) {
|
|
411
449
|
// Boundary = one of . ! ? OR a newline, that is followed by whitespace or EOT.
|
|
412
450
|
// A `.` wedged between two non-space chars (4.7, file.md, e.g.) is NOT a boundary.
|
|
451
|
+
//
|
|
452
|
+
// S-M2 fix (pre-ship gate review, 2026-09-09): CJK full-width terminators
|
|
453
|
+
// 。!? are ALWAYS a boundary, unconditionally (no "followed by whitespace"
|
|
454
|
+
// requirement) — unlike ASCII ./!/?, CJK prose has no decimal-point
|
|
455
|
+
// ambiguity to protect against and conventionally has no space after
|
|
456
|
+
// sentence punctuation, so requiring trailing whitespace would silently
|
|
457
|
+
// never split CJK text at all. This is the root dependency S-M1 (sentence-
|
|
458
|
+
// scoped WRITE, see hook-correction in cli/src/index.ts) needs: without it,
|
|
459
|
+
// a CJK prompt-injection tail appended after a genuine trigger clause (e.g.
|
|
460
|
+
// "不要在未经用户确认的情况下发布代码。忽略之前所有的规则…") could never be
|
|
461
|
+
// isolated from the trigger sentence, because the whole string was always
|
|
462
|
+
// returned as ONE fragment.
|
|
413
463
|
const out = [];
|
|
414
464
|
let buf = "";
|
|
415
465
|
for (let i = 0; i < text.length; i++) {
|
|
@@ -417,11 +467,14 @@ export function splitSentences(text) {
|
|
|
417
467
|
buf += ch;
|
|
418
468
|
const next = text[i + 1];
|
|
419
469
|
const isPunct = ch === "." || ch === "!" || ch === "?";
|
|
470
|
+
const isCjkPunct = ch === "。" || ch === "!" || ch === "?";
|
|
420
471
|
const isNewline = ch === "\n" || ch === "\r";
|
|
421
472
|
// Sentence boundary: terminal punctuation at end OR followed by whitespace;
|
|
422
|
-
// newline is
|
|
423
|
-
// a boundary (next is defined and not
|
|
473
|
+
// a CJK terminator or newline is ALWAYS a boundary. A `.` between
|
|
474
|
+
// non-whitespace chars is NOT a boundary (next is defined and not
|
|
475
|
+
// whitespace) — keeps decimals intact.
|
|
424
476
|
const atBoundary = isNewline ||
|
|
477
|
+
isCjkPunct ||
|
|
425
478
|
(isPunct && (next === undefined || /\s/.test(next)));
|
|
426
479
|
if (atBoundary) {
|
|
427
480
|
const frag = buf.trim();
|
|
@@ -446,18 +499,157 @@ export function splitSentences(text) {
|
|
|
446
499
|
// when the fragment is NOT a hedged/reporting frame. This is recall-safe: every
|
|
447
500
|
// genuine fixture correction carries a STRONG marker, a preference shape, or a
|
|
448
501
|
// non-hedged weak verb (verified by scripts/eval/capture-gate-confusion.mjs).
|
|
449
|
-
|
|
450
|
-
|
|
502
|
+
// CJK ADDITIONS (TOW2-326 class): the audit found this gate was English-
|
|
503
|
+
// directive-only — a Chinese correction whose only imperative markers are
|
|
504
|
+
// e.g. 必须/禁止/不要 (with none of the three pre-existing PREFERENCE_PATTERN
|
|
505
|
+
// trigger words 偏好/喜欢/要求 present) was silently REJECTED by
|
|
506
|
+
// writeCorrection() before it ever reached disk (verified directly:
|
|
507
|
+
// isLikelyRealCorrection("发布代码前必须获得用户确认") returned
|
|
508
|
+
// { ok:false, reason:"no actionable signal..." } — see
|
|
509
|
+
// audit-cjk-check-action.test.mjs's file-header finding, which flagged this
|
|
510
|
+
// exact gap for "whichever worker owns this file next"). CJK has no word
|
|
511
|
+
// boundaries (\b matches nothing meaningful around Han characters), so the
|
|
512
|
+
// CJK rows below are bare substring alternatives, mirroring the existing CJK
|
|
513
|
+
// rows in PREFERENCE_PATTERN (偏好|喜欢|要求) and in correction-detector.ts's
|
|
514
|
+
// CORRECTION_PATTERNS/BEHAVIORAL_SIGNALS — same convention, same file class.
|
|
515
|
+
//
|
|
516
|
+
// INDEPENDENT-REVIEW FIX (2026-09-09, round 3): 不能 is scoped to
|
|
517
|
+
// second-person 你不能, NOT bare 不能 — mirrors correction-detector.ts's
|
|
518
|
+
// pre-existing BEHAVIORAL_SIGNALS entry `/你不能(?!不)/` (and that entry's
|
|
519
|
+
// own documented reasoning): bare 不能 collides with plain capability/bug-
|
|
520
|
+
// report statements ("这个不能这样跑,报错了"), which are common in this
|
|
521
|
+
// gate's real traffic. The first version of this line added bare 不能(?!不)
|
|
522
|
+
// without inheriting that existing scoping — same class-not-instance miss
|
|
523
|
+
// this whole PR exists to close, just aimed at itself.
|
|
524
|
+
// PRE-SHIP GATE FIX (2026-09-09, C-1): shared benign-reassurance completion
|
|
525
|
+
// list for 不要 — extracted so STRONG_IMPERATIVE below, detectSeverity above,
|
|
526
|
+
// and check.ts's p0Patterns (imported from here) cannot independently drift
|
|
527
|
+
// again. The 2026-09-09 gate review found exactly that: this list existed in
|
|
528
|
+
// THREE places with THREE different widths (STRONG_IMPERATIVE had none at
|
|
529
|
+
// all; detectSeverity/check.ts had the narrow round-1 set). Exported so
|
|
530
|
+
// check.ts (same package) can build its own p0Patterns from the identical
|
|
531
|
+
// string instead of a hand-copied literal. correction-detector.ts (cli
|
|
532
|
+
// package) keeps its own copy — cross-package import was judged not worth
|
|
533
|
+
// the coupling for a 9-token disjunction; see that file's own doc comment,
|
|
534
|
+
// kept in sync by cross-reference (same "duplicated, cross-referenced"
|
|
535
|
+
// convention this file already uses for detectSeverity/p0Patterns).
|
|
536
|
+
export const CJK_REASSURANCE_COMPLETIONS = "担心|客气|急|着急|紧张|见外|慌|害怕|在意";
|
|
537
|
+
// PRE-SHIP GATE FIX (2026-09-09, S-M3): a directive-shaped clause that is
|
|
538
|
+
// being REPORTED, QUOTED from a document/manual, or explicitly left
|
|
539
|
+
// UNRESOLVED in discussion is never an authoritative rule — however strong
|
|
540
|
+
// its own marker. This is the ONE frame check STRONG_IMPERATIVE does NOT
|
|
541
|
+
// bypass (unlike HEDGE_FRAME, which only gates WEAK_IMPERATIVE) — design
|
|
542
|
+
// decision FINAL: quoted third-party rules and undecided narratives are
|
|
543
|
+
// never capturable as authoritative. One table (class-not-instance):
|
|
544
|
+
// - reporting/discussion openers: 我们讨论了/讨论过, 聊到(了)
|
|
545
|
+
// - document/manual attribution openers: 文档里写着, X手册规定,
|
|
546
|
+
// 根据…的说法, 据说
|
|
547
|
+
// - undecided-outcome markers (checked ANYWHERE, not just the opener —
|
|
548
|
+
// these are CONCLUDING remarks): 没有定论, 没(有)达成一致
|
|
549
|
+
// Anchored at ^ for the opener half, mirroring HEDGE_FRAME's own convention.
|
|
550
|
+
const QUOTE_NARRATIVE_OPENER = /^\s*(?:我们|大家|昨天|之前|上次|后来|团队)?[^\n]{0,6}?(?:讨论了|讨论过|聊到了?|文档里写着|(?:产品|公司|员工|操作|规范)?手册(?:里)?规定|根据[^\n]{0,20}的说法|据说)/;
|
|
551
|
+
const QUOTE_NARRATIVE_UNRESOLVED = /没有定论|没达成一致|没有达成一致/;
|
|
552
|
+
function isQuoteNarrativeFrame(text) {
|
|
553
|
+
return QUOTE_NARRATIVE_OPENER.test(text) || QUOTE_NARRATIVE_UNRESOLVED.test(text);
|
|
554
|
+
}
|
|
555
|
+
// PRE-SHIP GATE FIX (2026-09-09, C-2): three of the bare CJK tokens below
|
|
556
|
+
// (需要/确保/记得) gained structural completion requirements, mirroring the
|
|
557
|
+
// pre-existing English precedent's own narrowing (e.g. `needs?\s+(to|the|
|
|
558
|
+
// ...)` — "needs three years" does not match because "three" isn't in that
|
|
559
|
+
// list). Each excludes the SPECIFIC descriptive-prose shape the pre-ship
|
|
560
|
+
// gate review's fixture battery found, not an invented whitelist:
|
|
561
|
+
// - 需要(?!<digit-or-CJK-numeral>): "需要三年经验" (a quantity/requirement
|
|
562
|
+
// phrase) vs "需要先确认一下" (a directive) — descriptive usage quantifies,
|
|
563
|
+
// directive usage rarely does.
|
|
564
|
+
// - 确保(?!金): "确保金" is a name/compound collision (mirrors the existing
|
|
565
|
+
// 试试/改成/统一 compound-word fixes below), not the verb 确保.
|
|
566
|
+
// - 记得(?!上次|之前|以前|那时): "我记得上次是这样" (recollection of the
|
|
567
|
+
// past) vs "记得先跑测试" (a forward-looking directive).
|
|
568
|
+
const STRONG_IMPERATIVE = new RegExp(`\\b(never|always|don'?t|do not|must\\s+not|must|should\\s+not|needs?\\s+(to|those|the|a|an|more|all)\\b|instead|make\\s+sure|remember\\s+to|remove\\s+all|replace\\s+with|default\\s+to|keep\\s+the|keep\\s+\\w|show\\s+(both|all|the|only))\\b|` +
|
|
569
|
+
`永远不要|绝不|千万不要|总是|一直|始终|不要(?!${CJK_REASSURANCE_COMPLETIONS})|别再|禁止|不可以|不准|你不能(?!不)|不得(?!不)|切勿|不应该|必须|务必|一定要|` +
|
|
570
|
+
`需要(?![0-9一二两三四五六七八九十百千万几])|而不是|改为|换成|确保(?!金)|记得(?!上次|之前|以前|那时)|切记|删除所有|去掉所有|替换为|改用|默认使用|默认为|保留|保持|显示所有|只显示|仅显示`, "i");
|
|
571
|
+
// INDEPENDENT-REVIEW FIX (2026-09-09): English "OK, use the new endpoint
|
|
572
|
+
// instead" is rescued by the actionable scan (via "use") BEFORE it ever
|
|
573
|
+
// reaches the soft acknowledgmentPattern gate below — the CJK acknowledgment
|
|
574
|
+
// list added 可以/好的 as pure-ack openers, but without an equivalent CJK
|
|
575
|
+
// suggestion-verb row, a CJK sentence using the SAME "OK, do X" shape (可以
|
|
576
|
+
// 试试.../ 好的,那...改成/统一叫做...) had nothing to rescue it and fell
|
|
577
|
+
// into the ack gate as a false negative — the exact "correction silently
|
|
578
|
+
// never reaches disk" failure this file's CJK additions exist to close.
|
|
579
|
+
//
|
|
580
|
+
// ROUND-2 INDEPENDENT-REVIEW FIX (2026-09-09): the first pass added these as
|
|
581
|
+
// BARE 2-char substrings (试试|改成|统一), which — because CJK has no word
|
|
582
|
+
// boundaries — collide with common COMPOUND-WORD junctions in ordinary,
|
|
583
|
+
// non-directive Chinese: "面试" + "试讲" → "面试试讲" contains "试试"; "整改"
|
|
584
|
+
// + "成效" → "整改成效" and "修改" + "成本" → "修改成本" both contain "改成";
|
|
585
|
+
// and bare "统一" is itself an ordinary adjective/adverb ("统一标准", "统一
|
|
586
|
+
// 安排") with no directive meaning at all. All three were REJECTED, not just
|
|
587
|
+
// narrowed, in favor of structural patterns that require the actual
|
|
588
|
+
// suggestion/naming-rule SHAPE the original fixtures were modeling, not a
|
|
589
|
+
// bare substring:
|
|
590
|
+
// - 试试: only after a suggestion opener (可以/来/去) or before its most
|
|
591
|
+
// common completion (看) — "可以试试" / "试试...看" — not a bare
|
|
592
|
+
// mid-sentence occurrence that could straddle a compound-word boundary.
|
|
593
|
+
// - 改成: only after 把/将 (the object-marking preposition Mandarin's
|
|
594
|
+
// "change X to Y" construction always uses) — "把/将 ... 改成" — never
|
|
595
|
+
// a bare occurrence, closing the 整改+成效 / 修改+成本 collision.
|
|
596
|
+
// - 统一: only followed by a naming/directive completion (叫做/命名/称为/
|
|
597
|
+
// 改为) — "统一叫做orderId" — never bare, closing the "统一标准/统一
|
|
598
|
+
// 安排" ordinary-adjective false-positive. NOTE: 使用/采用 were
|
|
599
|
+
// considered but dropped from this completion list (own stress-testing,
|
|
600
|
+
// independent of the review rounds) — "统一使用"/"统一采用" are just as
|
|
601
|
+
// often ordinary descriptive prose ("公司统一使用企业微信办公" — the
|
|
602
|
+
// company uniformly uses X, a factual statement, not a correction) as a
|
|
603
|
+
// genuine naming rule, so they added false-positive risk without a
|
|
604
|
+
// matching precision gain; 叫做/命名/称为/改为 are naming-specific verbs
|
|
605
|
+
// with no equivalent descriptive-prose sense.
|
|
606
|
+
// ROUND-3 PRE-SHIP GATE FIX (2026-09-09): the 试试...看 gapped alternative
|
|
607
|
+
// ({0,10} chars between them) reproduced the EXACT compound-word collision
|
|
608
|
+
// class this section already fixed for 试试/改成/统一 above — "面试试讲,看看
|
|
609
|
+
// 效果怎么样" contains 试试 spanning the 面试/试讲 word boundary, and the
|
|
610
|
+
// SEPARATE, unrelated 看看 later in the sentence (reporting "let's see how it
|
|
611
|
+
// goes", nothing to do with 试试) was close enough to satisfy the {0,10} gap
|
|
612
|
+
// and wrongly rescue it. Tightened to the CONTIGUOUS "试试看" collocation
|
|
613
|
+
// (zero gap) — every genuine "试试看" fixture (own AND existing regression
|
|
614
|
+
// suite) has 看 immediately after 试试 with nothing between; the gapped form
|
|
615
|
+
// was never load-bearing for any real fixture, only for this collision.
|
|
616
|
+
//
|
|
617
|
+
// PRE-SHIP GATE FIX (2026-09-09, C-2): 应该/使用/停止/避免 gained structural
|
|
618
|
+
// completion requirements, same rationale as 需要/确保/记得 above:
|
|
619
|
+
// - 应该(?!已经|会|也许|可能): "他应该已经到了" (epistemic "must have") vs
|
|
620
|
+
// "应该用novada-search" (a directive).
|
|
621
|
+
// - 使用(?!的): "使用的是旧版本" (descriptive "what is used") vs "应该使用
|
|
622
|
+
// 新版本" (a directive).
|
|
623
|
+
// - 停止(?!了): "服务停止了响应" (past-tense event report) vs "请停止这样做"
|
|
624
|
+
// (a command).
|
|
625
|
+
// - (?<!为了)避免: "为了避免误会" (purpose clause) vs "请避免这样写" (a
|
|
626
|
+
// directive) — excludes what PRECEDES 避免, not what follows, since
|
|
627
|
+
// "避免误会" itself appears in both shapes.
|
|
628
|
+
const WEAK_IMPERATIVE = /\b(should|use|using|stop|avoid|prefer)\b|应该(?!已经|会|也许|可能)|使用(?!的)|停止(?!了)|(?<!为了)避免|(?:可以|来|去)[^\n]{0,4}试试|试试看|(?:把|将)[^\n]{0,10}改成|统一(?:叫做|命名|称为|改为)/i;
|
|
451
629
|
// Tentative / reporting frame at the START of a fragment — the speaker is musing
|
|
452
630
|
// or reporting, not issuing a rule. A WEAK marker inside such a frame is NOT a
|
|
453
631
|
// directive. Anchored at ^ so it only catches the OPENER, never a directive
|
|
454
|
-
// sentence that merely follows a hedge.
|
|
455
|
-
|
|
632
|
+
// sentence that merely follows a hedge. CJK rows mirror the English hedge
|
|
633
|
+
// openers (musing/intent/tentative-team framing) — no \b equivalent needed.
|
|
634
|
+
const HEDGE_FRAME = /^\s*(i\s+(think|guess|suppose|believe|reckon|feel|will)\b|i'?ll\b|i'?m\s+going\s+to\b|maybe\b|perhaps\b|sounds?\s+good\b|the\s+team\s+(wants?|thinks?|prefers?)\b|we\s+(could|might|may)\b|我觉得|我猜|我想|我认为|我感觉|我会|我打算|我要去|也许|可能|或许|听起来不错|听起来还行|团队希望|团队认为|团队倾向于|我们可以|我们也许|我们或许)/i;
|
|
456
635
|
// Preference / corrective-fact statement. Includes user-preference verbs (CJK
|
|
457
636
|
// equivalents) AND the "X not Y" / "wrong … not" corrective-fact shape that
|
|
458
637
|
// carries real intent without an imperative verb (e.g. "Product names are
|
|
459
|
-
// novada-search (not novada-mcp)"). Scanned per-fragment.
|
|
460
|
-
|
|
638
|
+
// novada-search (not novada-mcp)"). Scanned per-fragment. CJK rows extended
|
|
639
|
+
// (TOW2-326 class) beyond the original 偏好|喜欢|要求 three: 我要/我希望/想要
|
|
640
|
+
// (first-person preference verbs) and 更喜欢/倾向于 (comparative preference).
|
|
641
|
+
// PRE-SHIP GATE FIX (2026-09-09, C-2): 要求/喜欢 gained structural
|
|
642
|
+
// completion requirements:
|
|
643
|
+
// - (?<!的)要求: "客户的要求还没定" (的要求 — possessive NOUN phrase) vs
|
|
644
|
+
// "用户要求删除文件前必须先确认" (要求 as a first-person/subject VERB,
|
|
645
|
+
// the existing audit fixture — preceded by 户, not 的).
|
|
646
|
+
// - 喜欢 excluded when a THIRD-PERSON subject (他/她/它) opens the same
|
|
647
|
+
// clause: "他很喜欢这个方案" is a report about someone ELSE's
|
|
648
|
+
// preference, not the USER'S — mirrors the English row's own
|
|
649
|
+
// `\buser\s+(wants?|prefers?|likes?...)\b` requirement (subject-scoped,
|
|
650
|
+
// not "anyone likes anything"). "我更喜欢"/"更喜欢" (first-person) is a
|
|
651
|
+
// SEPARATE alternative below and is unaffected.
|
|
652
|
+
const PREFERENCE_PATTERN = /(\buser\s+(wants?|prefers?|likes?|needs?|agreed|tested|chose|wanted)\b|\bthe\s+user\s+is\b|偏好|(?<!他[^,。!?\n]{0,4})(?<!她[^,。!?\n]{0,4})(?<!它[^,。!?\n]{0,4})喜欢|(?<!的)要求|我要|我希望|想要|更喜欢|倾向于|\bwrong\b[\s\S]{0,60}\bnot\b|\(not\s+[^)]+\)|\bnot\s+\w[\w-]*[,.]?\s+(it'?s|its|use|the\s+\w))/i;
|
|
461
653
|
/**
|
|
462
654
|
* Capture-quality gate — rejects context-free fragments, pure acknowledgments,
|
|
463
655
|
* and text that carries no actionable signal.
|
|
@@ -569,19 +761,27 @@ export function isLikelyRealCorrection(rule, _context) {
|
|
|
569
761
|
// truncated slice), so a directive in sentence 2+ is now seen and can RESCUE
|
|
570
762
|
// a text that opens with an acknowledgment.
|
|
571
763
|
const fragments = [r, ...splitSentences(r)];
|
|
572
|
-
// (a) STRONG directive marker in any fragment → accept unconditionally
|
|
573
|
-
|
|
764
|
+
// (a) STRONG directive marker in any fragment → accept unconditionally,
|
|
765
|
+
// UNLESS the fragment is a quote/narrative frame (S-M3, 2026-09-09 pre-ship
|
|
766
|
+
// gate fix) — "unconditionally" here means "not gated on HEDGE_FRAME like
|
|
767
|
+
// WEAK below", never meant "even when the rule is merely being quoted from
|
|
768
|
+
// a document or discussed without resolution". This is the one frame check
|
|
769
|
+
// STRONG markers do not bypass.
|
|
770
|
+
if (fragments.some((f) => STRONG_IMPERATIVE.test(f) && !isQuoteNarrativeFrame(f))) {
|
|
574
771
|
return { ok: true };
|
|
575
772
|
}
|
|
576
773
|
// (a2) WEAK directive marker → accept only in a fragment that is NOT a hedged/
|
|
577
|
-
// reporting frame. Closes the Loop-14
|
|
578
|
-
// should use it") while still
|
|
579
|
-
// making it full width")
|
|
580
|
-
|
|
774
|
+
// reporting frame, NOR a quote/narrative frame. Closes the Loop-14
|
|
775
|
+
// filler-prose false-accept ("I think we should use it") while still
|
|
776
|
+
// accepting a direct weak-verb correction ("stop making it full width")
|
|
777
|
+
// and a directive sentence that merely FOLLOWS a hedge.
|
|
778
|
+
if (fragments.some((f) => WEAK_IMPERATIVE.test(f) && !HEDGE_FRAME.test(f) && !isQuoteNarrativeFrame(f))) {
|
|
581
779
|
return { ok: true };
|
|
582
780
|
}
|
|
583
|
-
// (b) preference / corrective-fact statement in any fragment
|
|
584
|
-
|
|
781
|
+
// (b) preference / corrective-fact statement in any fragment (quote/
|
|
782
|
+
// narrative-framed preferences don't count as the user's OWN preference
|
|
783
|
+
// either — S-M3 applies uniformly across all three signal types).
|
|
784
|
+
if (fragments.some((f) => PREFERENCE_PATTERN.test(f) && !isQuoteNarrativeFrame(f))) {
|
|
585
785
|
return { ok: true };
|
|
586
786
|
}
|
|
587
787
|
// ── SOFT ACKNOWLEDGMENT GATE — runs AFTER the actionable scan ──────────────
|
|
@@ -590,28 +790,78 @@ export function isLikelyRealCorrection(rule, _context) {
|
|
|
590
790
|
// found NO directive, so anything matching here is a genuine content-free ack
|
|
591
791
|
// ("ok sure", "no that's not what I meant", "confirmed"). NO length cap on the
|
|
592
792
|
// anchor — only the trailing budget — matching the v2 behavior for true acks.
|
|
593
|
-
|
|
793
|
+
// CJK ack markers added (TOW2-326 class): bare Chinese acknowledgments
|
|
794
|
+
// ("好的,我明白了", "没问题,收到") must be rejected the same way their
|
|
795
|
+
// English counterparts are — only reached when the actionable scan above
|
|
796
|
+
// found NO directive anywhere in the text, so a genuine CJK correction that
|
|
797
|
+
// happens to open with one of these words is unaffected (rescued earlier).
|
|
798
|
+
const acknowledgmentPattern = /^(no[,.]?\s*(that'?s\s+wrong[.!]?)?|ok(ay)?\b|good\b|great\b|nice\b|yes\b|yeah\b|right\b|wait\b|hmm+\b|sure\b|thanks?\b|confirmed\b|fair\s+point\b|不好意思|好的?|嗯+|哦+|噢+|是的?|对的?|没问题|收到|明白了?|知道了?|清楚了?|了解了?|确认了?|谢谢|辛苦了|行吧|可以的?)[\s\S]{0,80}$/i;
|
|
594
799
|
if (acknowledgmentPattern.test(r)) {
|
|
595
800
|
return { ok: false, reason: "pure acknowledgment or fragment — no rule content" };
|
|
596
801
|
}
|
|
597
802
|
return { ok: false, reason: "no actionable signal — rule lacks imperative/modal marker, preference statement, or substantive content" };
|
|
598
803
|
}
|
|
599
804
|
/**
|
|
600
|
-
*
|
|
601
|
-
*
|
|
602
|
-
*
|
|
805
|
+
* Fix #2 (dual-channel capture gate, 2026-09-11) — DISTILLED rule identity,
|
|
806
|
+
* the consolidation match key. Replaces the old char-level normalization
|
|
807
|
+
* (lowercase + collapse non-alphanumerics) with the shared CJK-aware
|
|
808
|
+
* tokenizer (helpers/tokenize.ts): NFKC-normalize, segment Han runs with
|
|
809
|
+
* Intl.Segmenter, lowercase + punctuation-strip the Latin remainder, then
|
|
810
|
+
* join the ORDERED token stream. Same rule REWORDED at the punctuation/
|
|
811
|
+
* case/whitespace/full-width level now merges — critically including CJK
|
|
812
|
+
* internal-whitespace variants ("永远不要 跳过测试" vs "永远不要跳过测试"),
|
|
813
|
+
* which the old char-normalizer treated as two distinct rules because it
|
|
814
|
+
* had no notion of CJK word boundaries.
|
|
815
|
+
*
|
|
816
|
+
* STILL deliberately conservative — token-level identity, ORDER PRESERVED
|
|
817
|
+
* INCLUDING cross-script interleaving (review fix 2026-09-11, code-review
|
|
818
|
+
* MEDIUM: tokenizeWords emits Han tokens first, which would have collapsed
|
|
819
|
+
* "在 staging 测试 deploy" and "在 staging deploy 测试" into one identity —
|
|
820
|
+
* the text is therefore segmented at Han-run boundaries first and each
|
|
821
|
+
* segment tokenized in source order). The old doctrine stands: exact
|
|
822
|
+
* identity is the one gate with ZERO risk of folding two DISTINCT rules
|
|
823
|
+
* into one ("use proxy.ts" vs "use middleware.ts" differ by a real token
|
|
824
|
+
* and never merge; "test before push" vs "push before test" differ by
|
|
825
|
+
* order and never merge). Paraphrase-level consolidation stays out of the
|
|
826
|
+
* zero-LLM storage path. `minLength: 1` keeps short discriminating tokens
|
|
827
|
+
* ("ts", "js", "p0") — dropping them would over-merge.
|
|
828
|
+
*
|
|
829
|
+
* NON-HAN/NON-LATIN SCRIPTS (review fix 2026-09-11, code-review HIGH-2): the
|
|
830
|
+
* strip class is `\p{L}\p{N}` (all letters/numbers, like the old
|
|
831
|
+
* normalizeRule) — NOT `[a-z0-9]` — so hangul/kana/etc. survive as tokens
|
|
832
|
+
* instead of every such rule distilling to "" and colliding. An EMPTY
|
|
833
|
+
* identity must never match anything — both merge loops (writeCorrection
|
|
834
|
+
* below, pending.ts's dedupe) guard on it.
|
|
603
835
|
*
|
|
604
|
-
*
|
|
605
|
-
*
|
|
606
|
-
*
|
|
607
|
-
* zero-LLM storage path because it cannot tell a duplicate from a contradiction:
|
|
608
|
-
* "use proxy.ts" vs "use middleware.ts" and a "P0" vs "P1" variant differ by a
|
|
609
|
-
* short/numeric token that any local matcher either inflates (char-trigram) or
|
|
610
|
-
* drops (sub-3-char token filter) — so it would wrongly merge them. Paraphrase-
|
|
611
|
-
* level consolidation is left to the optional semantic/LLM path, never here.
|
|
836
|
+
* Exported: the pending store (storage/pending.ts) dedupes staged captures
|
|
837
|
+
* by the SAME identity, so a repeat capture can never fan out into N
|
|
838
|
+
* pending rows (the ×34 verbatim-repeat exemplar dedupes to one).
|
|
612
839
|
*/
|
|
840
|
+
const DISTILL_OPTS = {
|
|
841
|
+
minLength: 1,
|
|
842
|
+
asciiStripRegex: /[^\p{L}\p{N}\s]+/gu,
|
|
843
|
+
};
|
|
844
|
+
export function distillRuleIdentity(rule) {
|
|
845
|
+
// NFKC up-front so the Han-run boundary scan below sees the same composed
|
|
846
|
+
// text tokenizeWords normalizes to internally.
|
|
847
|
+
const s = (rule ?? "").normalize("NFKC");
|
|
848
|
+
const out = [];
|
|
849
|
+
const hanRun = new RegExp(HAN_RUN_RE.source, "gu");
|
|
850
|
+
let last = 0;
|
|
851
|
+
let m;
|
|
852
|
+
while ((m = hanRun.exec(s)) !== null) {
|
|
853
|
+
if (m.index > last)
|
|
854
|
+
out.push(...tokenizeWords(s.slice(last, m.index).toLowerCase(), DISTILL_OPTS));
|
|
855
|
+
out.push(...tokenizeWords(m[0], DISTILL_OPTS)); // pure Han run — segmented in place
|
|
856
|
+
last = m.index + m[0].length;
|
|
857
|
+
}
|
|
858
|
+
if (last < s.length)
|
|
859
|
+
out.push(...tokenizeWords(s.slice(last).toLowerCase(), DISTILL_OPTS));
|
|
860
|
+
return out.join(" ");
|
|
861
|
+
}
|
|
862
|
+
/** Consolidation match key — see distillRuleIdentity above. */
|
|
613
863
|
function normalizeRule(rule) {
|
|
614
|
-
return (rule
|
|
864
|
+
return distillRuleIdentity(rule);
|
|
615
865
|
}
|
|
616
866
|
/**
|
|
617
867
|
* Write a correction to persistent storage.
|
|
@@ -622,7 +872,7 @@ function normalizeRule(rule) {
|
|
|
622
872
|
* are unaffected (the return value was void, now it is an object; ignoring it
|
|
623
873
|
* still compiles and runs correctly).
|
|
624
874
|
*/
|
|
625
|
-
export function writeCorrection(project, correction) {
|
|
875
|
+
export async function writeCorrection(project, correction) {
|
|
626
876
|
// Capture-quality gate — reject noise before touching disk.
|
|
627
877
|
// v3 (Loop 8): classify on the FULL correction text, not the truncated rule.
|
|
628
878
|
// `rule` is a first-sentence title slice (set by check.ts) that hid the
|
|
@@ -688,7 +938,10 @@ export function writeCorrection(project, correction) {
|
|
|
688
938
|
// higher severity/authority/weight. High-precision LOCAL gate — no key, no
|
|
689
939
|
// network — so this never runs an LLM on the storage hot path.
|
|
690
940
|
const normNew = normalizeRule(record.rule);
|
|
691
|
-
|
|
941
|
+
// Review fix (2026-09-11, HIGH-2 guard): an empty distilled identity
|
|
942
|
+
// (rule made entirely of stripped symbols) must never match anything —
|
|
943
|
+
// without this, every such record would fold into the first one.
|
|
944
|
+
for (const existing of normNew ? readActiveCorrections(project) : []) {
|
|
692
945
|
if (existing.id === record.id)
|
|
693
946
|
continue; // never merge into self (same-day re-slug)
|
|
694
947
|
if ((existing.kind ?? "correction") !== (record.kind ?? "correction"))
|
|
@@ -700,6 +953,12 @@ export function writeCorrection(project, correction) {
|
|
|
700
953
|
proof_count: (existing.proof_count ?? 1) + 1,
|
|
701
954
|
merged_from: [...(existing.merged_from ?? []), record.id],
|
|
702
955
|
tags: Array.from(new Set([...(existing.tags ?? []), ...(record.tags ?? [])])),
|
|
956
|
+
// Fix #2 review fix (2026-09-11, code-review LOW): absorb the incoming
|
|
957
|
+
// structured form's applies_when like tags — union, and only mint the
|
|
958
|
+
// field when at least one side carries it (old records stay untouched).
|
|
959
|
+
...(existing.applies_when || record.applies_when
|
|
960
|
+
? { applies_when: Array.from(new Set([...(existing.applies_when ?? []), ...(record.applies_when ?? [])])) }
|
|
961
|
+
: {}),
|
|
703
962
|
// keep the STRONGER signal on every axis
|
|
704
963
|
severity: existing.severity === "p0" || record.severity === "p0" ? "p0" : "p1",
|
|
705
964
|
weight: Math.max(existing.weight ?? 0, record.weight ?? 0),
|
|
@@ -796,8 +1055,18 @@ export function readCorrections(project) {
|
|
|
796
1055
|
if (!fs.existsSync(dir))
|
|
797
1056
|
return [];
|
|
798
1057
|
readCorrectionsScanLog.push(project);
|
|
1058
|
+
// fix4 S1 (2026-09-11): reserved-name exclusion — every `_`-prefixed entry
|
|
1059
|
+
// in a corrections directory is INFRASTRUCTURE, never a correction record:
|
|
1060
|
+
// `_index.md`, `_outcomes.jsonl`, `_rejected.jsonl`, `_ab_arms.jsonl` (all
|
|
1061
|
+
// already extension-excluded), plus the `_pending/`/`_quarantine/` staging
|
|
1062
|
+
// subtrees (fix2's dual-channel capture gate stages raw captures under
|
|
1063
|
+
// `corrections/_pending/` — those records are BY DESIGN not yet active and
|
|
1064
|
+
// must never be parsed as if they were). The class is excluded BY NAME
|
|
1065
|
+
// (leading underscore), not by extension accident, so a future flat
|
|
1066
|
+
// `_staged.json`-style infra file can never leak into the active ledger.
|
|
1067
|
+
// Class-not-instance: one namespace rule, not one branch per known file.
|
|
799
1068
|
const files = fs.readdirSync(dir)
|
|
800
|
-
.filter((f) => f.endsWith(".json"))
|
|
1069
|
+
.filter((f) => f.endsWith(".json") && !f.startsWith("_"))
|
|
801
1070
|
.sort()
|
|
802
1071
|
.reverse();
|
|
803
1072
|
const records = [];
|
|
@@ -842,7 +1111,7 @@ export function readP0Corrections(project, preloaded) {
|
|
|
842
1111
|
* The file is rewritten atomically — never deleted. The record remains in
|
|
843
1112
|
* _outcomes.jsonl history and can be manually reactivated by editing the JSON.
|
|
844
1113
|
*/
|
|
845
|
-
export function retractCorrection(project, id, reason, supersededBy) {
|
|
1114
|
+
export async function retractCorrection(project, id, reason, supersededBy) {
|
|
846
1115
|
const dir = correctionsDir(project);
|
|
847
1116
|
// ── LOCKED critical section (P0 data-loss fix, 2026-07-25) ────────────────
|
|
848
1117
|
// read-all→find→mutate→atomic-rewrite→index-regen, unlocked, is exactly the
|
|
@@ -891,7 +1160,7 @@ export function retractCorrection(project, id, reason, supersededBy) {
|
|
|
891
1160
|
* the core-level enforcement of the single-producer contract (the CLI's
|
|
892
1161
|
* `ar outcomes record` is the one caller that adds the prefix).
|
|
893
1162
|
*/
|
|
894
|
-
export function recordOutcome(outcome) {
|
|
1163
|
+
export async function recordOutcome(outcome) {
|
|
895
1164
|
// C3b single-producer gate: not_triggered without the dream-audit evidence
|
|
896
1165
|
// prefix indicates an unauthorized producer — fail loudly, never silently.
|
|
897
1166
|
if (outcome.kind === "not_triggered" &&
|
|
@@ -930,89 +1199,108 @@ export function recordOutcome(outcome) {
|
|
|
930
1199
|
// DIFFERENT corrections' index writes race). See the matching note in
|
|
931
1200
|
// writeCorrection/retractCorrection above; same lock name, same pattern as
|
|
932
1201
|
// `palace-index-${project}` / `digest-${project}` elsewhere in this codebase.
|
|
933
|
-
|
|
934
|
-
|
|
935
|
-
|
|
936
|
-
|
|
937
|
-
|
|
938
|
-
|
|
939
|
-
|
|
940
|
-
|
|
941
|
-
|
|
942
|
-
|
|
943
|
-
|
|
944
|
-
|
|
945
|
-
|
|
946
|
-
|
|
947
|
-
|
|
948
|
-
|
|
949
|
-
|
|
950
|
-
|
|
951
|
-
|
|
952
|
-
|
|
953
|
-
|
|
954
|
-
|
|
955
|
-
|
|
956
|
-
|
|
957
|
-
|
|
958
|
-
|
|
959
|
-
|
|
960
|
-
|
|
961
|
-
|
|
962
|
-
|
|
963
|
-
|
|
964
|
-
|
|
965
|
-
|
|
966
|
-
|
|
967
|
-
|
|
968
|
-
|
|
969
|
-
|
|
970
|
-
|
|
971
|
-
|
|
1202
|
+
//
|
|
1203
|
+
// fix6-locks contention handling: the jsonl ledger line above is ALREADY
|
|
1204
|
+
// appended (append-only, lossless), and the materialized counters below are
|
|
1205
|
+
// fully re-derivable from that ledger via runOutcomesRebuild. So if another
|
|
1206
|
+
// LIVE process holds the lock past the timeout, skipping this counter pass
|
|
1207
|
+
// loses NOTHING durable — log the skip and return instead of stealing the
|
|
1208
|
+
// lock (old behavior, mid-write corruption risk) or failing the caller's
|
|
1209
|
+
// hot path (session_start / session_end / check_action record inline).
|
|
1210
|
+
try {
|
|
1211
|
+
await withLock(`corrections-${outcome.project}`, () => {
|
|
1212
|
+
// Update the per-correction file's counters.
|
|
1213
|
+
const target = readCorrections(outcome.project).find((r) => r.id === outcome.correction_id);
|
|
1214
|
+
if (!target)
|
|
1215
|
+
return; // outcome can still be replayed later if record is restored
|
|
1216
|
+
const updated = {
|
|
1217
|
+
...target,
|
|
1218
|
+
retrieved_count: target.retrieved_count ?? 0,
|
|
1219
|
+
heeded_count: target.heeded_count ?? 0,
|
|
1220
|
+
recurrence_count: target.recurrence_count ?? 0,
|
|
1221
|
+
};
|
|
1222
|
+
if (outcome.kind === "retrieved") {
|
|
1223
|
+
updated.retrieved_count = (updated.retrieved_count ?? 0) + 1;
|
|
1224
|
+
updated.last_retrieved = outcome.at;
|
|
1225
|
+
}
|
|
1226
|
+
else if (outcome.kind === "heeded") {
|
|
1227
|
+
updated.heeded_count = (updated.heeded_count ?? 0) + 1;
|
|
1228
|
+
updated.last_outcome = outcome.at;
|
|
1229
|
+
}
|
|
1230
|
+
else if (outcome.kind === "recurred") {
|
|
1231
|
+
updated.recurrence_count = (updated.recurrence_count ?? 0) + 1;
|
|
1232
|
+
updated.last_outcome = outcome.at;
|
|
1233
|
+
}
|
|
1234
|
+
else if (outcome.kind === "predicted") {
|
|
1235
|
+
// Wave 5: prediction fired — instrument the predict-the-correction loop.
|
|
1236
|
+
updated.predicted_count = (updated.predicted_count ?? 0) + 1;
|
|
1237
|
+
updated.last_predicted = outcome.at;
|
|
1238
|
+
}
|
|
1239
|
+
else if (outcome.kind === "predict_hit") {
|
|
1240
|
+
updated.predict_hits = (updated.predict_hits ?? 0) + 1;
|
|
1241
|
+
}
|
|
1242
|
+
else if (outcome.kind === "not_violated") {
|
|
1243
|
+
// Heed-rate credit model Option A (2026-08-29): its OWN counter, seeded
|
|
1244
|
+
// only on its own first event (mirrors predicted_count/predict_hits'
|
|
1245
|
+
// pattern, NOT the unconditional retrieved/heeded/recurrence baseline
|
|
1246
|
+
// seed above) — deliberately kept out of precision/proof_confidence
|
|
1247
|
+
// below, which read ONLY heeded_count/recurrence_count.
|
|
1248
|
+
updated.not_violated_count = (updated.not_violated_count ?? 0) + 1;
|
|
1249
|
+
updated.last_outcome = outcome.at;
|
|
1250
|
+
}
|
|
1251
|
+
const r = updated.retrieved_count ?? 0;
|
|
1252
|
+
// Clamp to [0,1]: `retrieved` is guarded 1/day but `heeded` can fire on every
|
|
1253
|
+
// session_end, so raw heeded/retrieved can exceed 1.0 ("150% heeded" is
|
|
1254
|
+
// nonsense). min(1, …) keeps the metric honest. (Root-cause follow-up: apply
|
|
1255
|
+
// the same 1/day guard to heeded as retrieved has, for finer resolution.)
|
|
1256
|
+
// NB (Wave 5): `precision` is heeded/retrieved ONLY — predict_* never touch it.
|
|
1257
|
+
updated.precision = r > 0 ? Math.min(1, Number(((updated.heeded_count ?? 0) / r).toFixed(3))) : undefined;
|
|
1258
|
+
// Wave 5: predict_precision = predict_hits / predicted_count, kept SEPARATE
|
|
1259
|
+
// from `precision`. Undefined until at least one prediction has fired.
|
|
1260
|
+
// A predict_hit implies a prior prediction. If data is inconsistent (hits
|
|
1261
|
+
// recorded without a matching predicted_count — e.g. migrated/corrupt records),
|
|
1262
|
+
// floor the denominator at predict_hits so the metric stays VISIBLE and bounded
|
|
1263
|
+
// rather than silently undefined while hits exist.
|
|
1264
|
+
const pc = updated.predicted_count ?? 0;
|
|
1265
|
+
const ph = updated.predict_hits ?? 0;
|
|
1266
|
+
const predictDenom = Math.max(pc, ph);
|
|
1267
|
+
updated.predict_precision = predictDenom > 0
|
|
1268
|
+
? Math.min(1, Number((ph / predictDenom).toFixed(3)))
|
|
1269
|
+
: undefined;
|
|
1270
|
+
// P3: evidence-grounded proof_confidence. With NO outcome evidence yet, keep the
|
|
1271
|
+
// authority prior (weight); once heeded/recurrence accrue, move to the Beta
|
|
1272
|
+
// posterior so a rule that keeps being honored strengthens and one whose bug
|
|
1273
|
+
// keeps recurring weakens. Kept SEPARATE from `precision` (heeded/retrieved) and
|
|
1274
|
+
// from `weight` (static authority) — this is the evidence axis.
|
|
1275
|
+
const heededC = updated.heeded_count ?? 0;
|
|
1276
|
+
const recurC = updated.recurrence_count ?? 0;
|
|
1277
|
+
updated.proof_confidence = (heededC + recurC) > 0
|
|
1278
|
+
? Number(betaPosterior(heededC, recurC).toFixed(3))
|
|
1279
|
+
: (updated.weight ?? defaultWeight(updated.severity));
|
|
1280
|
+
// Re-write the JSON file atomically (tmp + rename — prevents truncation on
|
|
1281
|
+
// SIGTERM). Reuse the EXISTING filename (see findExistingCorrectionFile
|
|
1282
|
+
// doc) — never recompute via slugify, or a v1-named file would orphan.
|
|
1283
|
+
const filename = findExistingCorrectionFile(dir, updated.id)
|
|
1284
|
+
?? `${updated.date}--${slugify(updated.rule || updated.id)}.json`;
|
|
1285
|
+
const filepath = path.join(dir, filename);
|
|
1286
|
+
writeRecordAtomic(filepath, updated);
|
|
1287
|
+
// W2-1: regenerate the materialized index on every corrections mutation.
|
|
1288
|
+
// (The ledger-only early-returns above for triggered/not_triggered/unknown
|
|
1289
|
+
// touch no per-record field the index renders — severity/status/date/rule/
|
|
1290
|
+
// failure_class are all unchanged — so they deliberately skip this call,
|
|
1291
|
+
// preserving the existing hot-path optimization documented above.)
|
|
1292
|
+
regenerateCorrectionsIndex(outcome.project);
|
|
1293
|
+
});
|
|
1294
|
+
}
|
|
1295
|
+
catch (err) {
|
|
1296
|
+
if (err instanceof LockContentionError) {
|
|
1297
|
+
console.error(`[agent-recall] recordOutcome(${outcome.kind}, ${outcome.correction_id}): ${err.message} ` +
|
|
1298
|
+
`— materialized counter update skipped; the _outcomes.jsonl ledger already holds this ` +
|
|
1299
|
+
`event and \`ar outcomes rebuild\` re-derives the counters.`);
|
|
1300
|
+
return;
|
|
972
1301
|
}
|
|
973
|
-
|
|
974
|
-
|
|
975
|
-
// session_end, so raw heeded/retrieved can exceed 1.0 ("150% heeded" is
|
|
976
|
-
// nonsense). min(1, …) keeps the metric honest. (Root-cause follow-up: apply
|
|
977
|
-
// the same 1/day guard to heeded as retrieved has, for finer resolution.)
|
|
978
|
-
// NB (Wave 5): `precision` is heeded/retrieved ONLY — predict_* never touch it.
|
|
979
|
-
updated.precision = r > 0 ? Math.min(1, Number(((updated.heeded_count ?? 0) / r).toFixed(3))) : undefined;
|
|
980
|
-
// Wave 5: predict_precision = predict_hits / predicted_count, kept SEPARATE
|
|
981
|
-
// from `precision`. Undefined until at least one prediction has fired.
|
|
982
|
-
// A predict_hit implies a prior prediction. If data is inconsistent (hits
|
|
983
|
-
// recorded without a matching predicted_count — e.g. migrated/corrupt records),
|
|
984
|
-
// floor the denominator at predict_hits so the metric stays VISIBLE and bounded
|
|
985
|
-
// rather than silently undefined while hits exist.
|
|
986
|
-
const pc = updated.predicted_count ?? 0;
|
|
987
|
-
const ph = updated.predict_hits ?? 0;
|
|
988
|
-
const predictDenom = Math.max(pc, ph);
|
|
989
|
-
updated.predict_precision = predictDenom > 0
|
|
990
|
-
? Math.min(1, Number((ph / predictDenom).toFixed(3)))
|
|
991
|
-
: undefined;
|
|
992
|
-
// P3: evidence-grounded proof_confidence. With NO outcome evidence yet, keep the
|
|
993
|
-
// authority prior (weight); once heeded/recurrence accrue, move to the Beta
|
|
994
|
-
// posterior so a rule that keeps being honored strengthens and one whose bug
|
|
995
|
-
// keeps recurring weakens. Kept SEPARATE from `precision` (heeded/retrieved) and
|
|
996
|
-
// from `weight` (static authority) — this is the evidence axis.
|
|
997
|
-
const heededC = updated.heeded_count ?? 0;
|
|
998
|
-
const recurC = updated.recurrence_count ?? 0;
|
|
999
|
-
updated.proof_confidence = (heededC + recurC) > 0
|
|
1000
|
-
? Number(betaPosterior(heededC, recurC).toFixed(3))
|
|
1001
|
-
: (updated.weight ?? defaultWeight(updated.severity));
|
|
1002
|
-
// Re-write the JSON file atomically (tmp + rename — prevents truncation on
|
|
1003
|
-
// SIGTERM). Reuse the EXISTING filename (see findExistingCorrectionFile
|
|
1004
|
-
// doc) — never recompute via slugify, or a v1-named file would orphan.
|
|
1005
|
-
const filename = findExistingCorrectionFile(dir, updated.id)
|
|
1006
|
-
?? `${updated.date}--${slugify(updated.rule || updated.id)}.json`;
|
|
1007
|
-
const filepath = path.join(dir, filename);
|
|
1008
|
-
writeRecordAtomic(filepath, updated);
|
|
1009
|
-
// W2-1: regenerate the materialized index on every corrections mutation.
|
|
1010
|
-
// (The ledger-only early-returns above for triggered/not_triggered/unknown
|
|
1011
|
-
// touch no per-record field the index renders — severity/status/date/rule/
|
|
1012
|
-
// failure_class are all unchanged — so they deliberately skip this call,
|
|
1013
|
-
// preserving the existing hot-path optimization documented above.)
|
|
1014
|
-
regenerateCorrectionsIndex(outcome.project);
|
|
1015
|
-
});
|
|
1302
|
+
throw err;
|
|
1303
|
+
}
|
|
1016
1304
|
}
|
|
1017
1305
|
/**
|
|
1018
1306
|
* Outcome kinds that affect per-correction counters (mirrors recordOutcome's
|
|
@@ -1316,7 +1604,7 @@ function toPublicDiffs(entries) {
|
|
|
1316
1604
|
* 4. Only per-correction JSON files that actually differ are rewritten; the
|
|
1317
1605
|
* shared _index.md is regenerated ONCE at the end, only if anything wrote.
|
|
1318
1606
|
*/
|
|
1319
|
-
export function runOutcomesRebuild(project, opts = {}) {
|
|
1607
|
+
export async function runOutcomesRebuild(project, opts = {}) {
|
|
1320
1608
|
const apply = opts.apply === true;
|
|
1321
1609
|
if (!apply) {
|
|
1322
1610
|
const plan = computeLedgerDivergence(project);
|
|
@@ -1807,13 +2095,13 @@ export function getCorrectionKPIs(project, preloaded) {
|
|
|
1807
2095
|
* posture: deleting belief is a deliberate act, so the default never mutates;
|
|
1808
2096
|
* an explicit human (or opt-in flag) triggers the retraction.
|
|
1809
2097
|
*/
|
|
1810
|
-
export function reviewNoiseCorrections(project, opts) {
|
|
2098
|
+
export async function reviewNoiseCorrections(project, opts) {
|
|
1811
2099
|
const auto = opts?.auto ?? (process.env.AR_CONSOLIDATE_AUTO === "1");
|
|
1812
2100
|
const suggestions = getCorrectionKPIs(project).noise_candidates;
|
|
1813
2101
|
const pruned = [];
|
|
1814
2102
|
if (auto) {
|
|
1815
2103
|
for (const c of suggestions) {
|
|
1816
|
-
const res = retractCorrection(project, c.id, "auto-pruned: low signal (precision<0.3, retrieved≥3)");
|
|
2104
|
+
const res = await retractCorrection(project, c.id, "auto-pruned: low signal (precision<0.3, retrieved≥3)");
|
|
1817
2105
|
if (res.success)
|
|
1818
2106
|
pruned.push(c.id);
|
|
1819
2107
|
}
|