@polycode-projects/the-mechanical-code-talker 0.9.12 → 1.0.1
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/corpus/seon/relations.jsonl +1 -0
- package/package.json +1 -1
- package/src/ask-vocab.mjs +25 -2
- package/src/ask.mjs +67 -11
- package/src/chat.mjs +889 -66
- package/src/concept.mjs +50 -9
- package/src/finish.mjs +1 -1
- package/src/interpret/normalize.mjs +90 -1
- package/src/memory/core.mjs +14 -1
package/src/chat.mjs
CHANGED
|
@@ -51,9 +51,9 @@ import { uuidv7 } from "./uuid.mjs";
|
|
|
51
51
|
import { createTelemetry } from "./telemetry.mjs";
|
|
52
52
|
import * as defaultSource from "./source.mjs";
|
|
53
53
|
import { loadTemplates, render as renderTemplate } from "./corpus/templates.mjs";
|
|
54
|
-
import { finish } from "./finish.mjs";
|
|
55
|
-
import { VERB_TO_KIND, WHERE_MARKERS, MENTION_MARKERS, ENTITY_TO_TYPE } from "./ask-vocab.mjs";
|
|
56
|
-
import { COUNTERFACTUAL_RE } from "./interpret/normalize.mjs";
|
|
54
|
+
import { finish, beginsWithVowelSound, grammarRules } from "./finish.mjs";
|
|
55
|
+
import { VERB_TO_KIND, WHERE_MARKERS, MENTION_MARKERS, ENTITY_TO_TYPE, PASSIVE_PARTICIPLE_TO_KIND } from "./ask-vocab.mjs";
|
|
56
|
+
import { COUNTERFACTUAL_RE, correctMisspellings, applyPreambleFrames } from "./interpret/normalize.mjs";
|
|
57
57
|
import { fuzzyMatchInSet, fuzzyBound } from "./interpret/fuzzy.mjs";
|
|
58
58
|
|
|
59
59
|
// uuidv7 lives in ./uuid.mjs (shared with telemetry + the bench stamp); re-exported
|
|
@@ -218,6 +218,53 @@ function withNarration(result, trace, fallbackGoal) {
|
|
|
218
218
|
return { ...result, answer, logLines };
|
|
219
219
|
}
|
|
220
220
|
|
|
221
|
+
/** FEATURE B ("Goal (inferred): …"): an ALWAYS-ON, single short goal line —
|
|
222
|
+
* independent of the --narrate/TMCT_NARRATE opt-in debug trace above (which
|
|
223
|
+
* stays exactly as-is: the FULL "--- narrate ---" block, off by default).
|
|
224
|
+
* What the operator actually wants now is much lighter than that full dump:
|
|
225
|
+
* one line, on every STRUCTURAL/query-shaped answer, APPENDED (blank-line
|
|
226
|
+
* separated) so it never reads as part of the substantive answer — and so it
|
|
227
|
+
* never disturbs the many existing START-anchored assertions this codebase's
|
|
228
|
+
* own test suite pins composed answers with (see withGoalLine's own docblock,
|
|
229
|
+
* just below, for why appended rather than led-with).
|
|
230
|
+
*
|
|
231
|
+
* `result.goal` is set ONLY by runAsk (see its own docblock at its return
|
|
232
|
+
* statement) — a plain count, a slash-command or a teach confirmation never
|
|
233
|
+
* carries the field, so this is a no-op for those turn types BY
|
|
234
|
+
* CONSTRUCTION, not a special-cased suppression list here. Also a no-op when
|
|
235
|
+
* `result.goal` is null/empty — deduceGoalFromParsed's own "nothing to
|
|
236
|
+
* bucket on" signal (a total grammar miss, or a would-miss a conversational-
|
|
237
|
+
* in-ask lane answered) — so an unclear turn never grows a "Goal (inferred):
|
|
238
|
+
* unclear" line, which would be worse than showing nothing.
|
|
239
|
+
*
|
|
240
|
+
* Applied AFTER finish() (so the appended line is never grammar-rewritten) and
|
|
241
|
+
* BEFORE `last` is captured in runTurn's withLast — mirrors withNarration's
|
|
242
|
+
* own after-finish/never-touches-`last` discipline (see its docblock above),
|
|
243
|
+
* so a goal-prefixed turn's own repeat-detection / why/say-more re-render
|
|
244
|
+
* compares the EXACT SAME `last.answer` a goal-line-off run would have
|
|
245
|
+
* produced. Purely additive to what's PRINTED, never to what's REMEMBERED —
|
|
246
|
+
* the same contract narrate uses, a second, independent mechanism reusing
|
|
247
|
+
* the same discipline (composes cleanly with narrate: a narrated turn gets
|
|
248
|
+
* BOTH the short line up top and the full trace block below, never a
|
|
249
|
+
* conflict). */
|
|
250
|
+
function withGoalLine(result) {
|
|
251
|
+
const goal = result?.goal;
|
|
252
|
+
if (!goal) return result;
|
|
253
|
+
// APPENDED (not prepended), blank-line separated: this codebase's existing
|
|
254
|
+
// test suite pins a large number of composed answers with a START-anchored
|
|
255
|
+
// (`^…`, no trailing `$`) regex — appending keeps every one of those intact
|
|
256
|
+
// (the answer still STARTS with the real content) while a prepend would have
|
|
257
|
+
// broken them all. Still reads as clearly separate, non-substantive trailer
|
|
258
|
+
// text — the same "additive, never mixed into the substantive answer" intent
|
|
259
|
+
// a leading line would have given, just from the other end.
|
|
260
|
+
const suffix = `Goal (inferred): ${goal.charAt(0).toUpperCase()}${goal.slice(1)}.`;
|
|
261
|
+
const answer = `${result.answer}\n\n${suffix}`;
|
|
262
|
+
const logLines = Array.isArray(result.logLines)
|
|
263
|
+
? result.logLines.map((l) => (l === result.answer ? answer : l))
|
|
264
|
+
: result.logLines;
|
|
265
|
+
return { ...result, answer, logLines };
|
|
266
|
+
}
|
|
267
|
+
|
|
221
268
|
/** Slash-command → (dispatchTool name, arg key). Arg keys are the EXACT ones the
|
|
222
269
|
* server.mjs dispatchTool switch reads (members/subclasses take `class`;
|
|
223
270
|
* impact/exports take `module`; architecture takes `package`; search takes
|
|
@@ -333,6 +380,14 @@ export function asBareCommand(line) {
|
|
|
333
380
|
// unconditionally): the predicate-find grammar's own shape wins regardless of
|
|
334
381
|
// word count, see the precedence note above.
|
|
335
382
|
if (fl === "find" && looksLikePredicateFind(restTok)) return null;
|
|
383
|
+
// "describe it"/"describe that" (0.9.13 Tier-1 playtest): a bare PRONOUN argument
|
|
384
|
+
// to /describe has no antecedent at this layer — dispatchTool("tmct_describe", …)
|
|
385
|
+
// does its own name-only resolveSymbol lookup with no notion of the standing
|
|
386
|
+
// focus, so routing it here as a bare command produced a raw "no such symbol"
|
|
387
|
+
// failure. Defer to the ordinary pipeline instead (return null): it reaches
|
|
388
|
+
// describeWrapperAnswer's rescue lane, which DOES resolve a bare pronoun against
|
|
389
|
+
// the standing focus. A named argument ("describe Widget") is untouched.
|
|
390
|
+
if (fl === "describe" && DESCRIBE_PRONOUN_RE.test(rest)) return null;
|
|
336
391
|
// A NO-ARGUMENT command word ("untested") with trailing words is NOT a command
|
|
337
392
|
// call — the /untested tool takes no argument and would silently drop the qualifier,
|
|
338
393
|
// listing MODULES for "untested classes". "untested classes" / "untested modules"
|
|
@@ -386,9 +441,72 @@ function countableKinds(graph) {
|
|
|
386
441
|
* by the ask engine's anaphora node, never the header-count path. */
|
|
387
442
|
const ANAPHORA_COUNT_RE = /\b(?:how many|how much|count|number of)\s+(?:of\s+)?(?:those|them|these)\b/i;
|
|
388
443
|
|
|
444
|
+
/** An IMPLICIT anaphoric count with NO explicit "of them/those/these" at all —
|
|
445
|
+
* "how many are tested", "and how many are tested" (Tier-2 playtest, 5th
|
|
446
|
+
* pass). A fluent staccato follow-up after a just-given list naturally elides
|
|
447
|
+
* the pronoun a fuller phrasing ("how many of those are tested") carries —
|
|
448
|
+
* ANAPHORA_COUNT_RE above requires that explicit "of them/those/these" and
|
|
449
|
+
* never fires for this shape, so answerCount's own bare noun-scan greedily
|
|
450
|
+
* (and wrongly) captured the linking verb ITSELF as the counted noun ("how
|
|
451
|
+
* many ARE tested" -> noun="are") and answered the nonsensical "I can't
|
|
452
|
+
* count 'are'." Gated on real content after the linking verb (`(?!there\b)`)
|
|
453
|
+
* so a genuinely bare "how many are there" (no antecedent, no predicate to
|
|
454
|
+
* filter on) is untouched — that one's existing "I can't count 'are'" nudge
|
|
455
|
+
* is arguably the more honest answer to a query naming nothing at all. */
|
|
456
|
+
const IMPLICIT_ANAPHORA_COUNT_RE = /^(?:(?:and|so|then|also)\s+)?how many (?:are|is|were|was)\s+(?!there\b)(\S.*)$/i;
|
|
457
|
+
|
|
458
|
+
/** "have"/"has"/"holds"/"hold" are excluded from RESTRICTOR_VERB_RE below —
|
|
459
|
+
* DELIBERATELY treated as non-restrictor cues here, not a bug fix skipped. Ask-
|
|
460
|
+
* vocab's VERB_TO_KIND maps them to "defines" unconditionally, but the graph's
|
|
461
|
+
* actual "have" semantics are subject-type-dependent (a Module "has" things it
|
|
462
|
+
* defines; a Class "has" things it contains) — found live (0.9.14 Tier-2 playtest,
|
|
463
|
+
* third pass, numeric/quantifier relation touches) that ask.mjs's own engine
|
|
464
|
+
* resolves the two surface forms of the SAME query ("what methods does Widget
|
|
465
|
+
* have" vs "which methods does Widget have") to DIFFERENT, inconsistent kinds (one
|
|
466
|
+
* correctly reaches "contains", the other wrongly reaches "defines" and returns an
|
|
467
|
+
* honest-but-wrong zero) — a genuine, pre-existing ambiguity in the core clause
|
|
468
|
+
* grammar, orthogonal to dialogue flow/routing and out of this cycle's scope.
|
|
469
|
+
* Deferring a "have" tail to the ask engine here would just trade one wrong-answer
|
|
470
|
+
* risk for another rather than fixing anything, so it stays on the existing
|
|
471
|
+
* bare-count path (unchanged behavior, no new regression) until a dedicated fix
|
|
472
|
+
* teaches the grammar to pick "defines" vs "contains" by the resolved subject's
|
|
473
|
+
* own class. */
|
|
474
|
+
const AMBIGUOUS_HAVE_VERBS = new Set(["have", "has", "holds", "hold"]);
|
|
475
|
+
|
|
476
|
+
/** A "how many <kind> …" tail carries a genuine RESTRICTOR clause — not filler — iff
|
|
477
|
+
* it names a real relation verb (active, from VERB_TO_KIND, or passive-participle,
|
|
478
|
+
* from PASSIVE_PARTICIPLE_TO_KIND — both ask-vocab.mjs's closed vocabulary, the
|
|
479
|
+
* same one ask.mjs's own clause grammar reads), minus the ambiguous "have" family
|
|
480
|
+
* above. Matching on the VERB specifically (not "any non-stopword word") matters:
|
|
481
|
+
* a tail's own OBJECT NAME is also non-stopword content ("how many methods does
|
|
482
|
+
* WIDGET have" — "Widget" alone isn't a restrictor cue), so a bare
|
|
483
|
+
* content-word test would misfire on every qualified count regardless of verb. */
|
|
484
|
+
const RESTRICTOR_VERB_RE = new RegExp(
|
|
485
|
+
`\\b(?:${
|
|
486
|
+
[...Object.keys(VERB_TO_KIND), ...Object.keys(PASSIVE_PARTICIPLE_TO_KIND)]
|
|
487
|
+
.filter((v) => !AMBIGUOUS_HAVE_VERBS.has(v))
|
|
488
|
+
.sort((a, b) => b.length - a.length)
|
|
489
|
+
.map((v) => v.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"))
|
|
490
|
+
.join("|")
|
|
491
|
+
})\\b`,
|
|
492
|
+
"i",
|
|
493
|
+
);
|
|
494
|
+
|
|
389
495
|
/** Recognise a count/aggregate question and answer it from the graph header, or
|
|
390
496
|
* null if it isn't one (→ fall through to tmct_ask). "how many X [are there]",
|
|
391
|
-
* "count [the] X", "number of X". An unknown kind lists what it CAN count.
|
|
497
|
+
* "count [the] X", "number of X". An unknown kind lists what it CAN count.
|
|
498
|
+
*
|
|
499
|
+
* A RESTRICTOR tail ("how many modules IMPORT app/lib/a.mjs", "how many classes
|
|
500
|
+
* INHERIT FROM Base") is NOT a bare header count — found live (0.9.14 Tier-2
|
|
501
|
+
* playtest, third pass, numeric/quantifier relation touches): this regex only ever
|
|
502
|
+
* captured the noun immediately after "how many" and silently discarded everything
|
|
503
|
+
* after it, so a qualified count fell back to the UNQUALIFIED class total ("how many
|
|
504
|
+
* modules import app/lib/a.mjs" answered "8 modules" — the whole-graph module count —
|
|
505
|
+
* instead of the 3 that actually import it). ask.mjs's own AGGREGATE node
|
|
506
|
+
* (parseAggregate) already evaluates a restrictor tail correctly via parseSetPhrase,
|
|
507
|
+
* so once the tail names a real relation verb (RESTRICTOR_VERB_RE), decline here and
|
|
508
|
+
* let the turn fall through to the real ask engine instead of returning a misleading
|
|
509
|
+
* bare total. */
|
|
392
510
|
export function answerCount(graph, query) {
|
|
393
511
|
if (!graph) return null;
|
|
394
512
|
// ANAPHORIC counts ("how many of those are tested", "count them", "how many of
|
|
@@ -397,10 +515,15 @@ export function answerCount(graph, query) {
|
|
|
397
515
|
// this the bare "of"/pronoun head is mis-reported as an uncountable kind and the
|
|
398
516
|
// discourse+count follow-up dies before it can resolve (CHATBENCH_006 lever 1).
|
|
399
517
|
if (ANAPHORA_COUNT_RE.test(String(query))) return null;
|
|
518
|
+
// The elliptical sibling above (no explicit "of them/those" at all) — same
|
|
519
|
+
// decline, same reason: this is a reference to the PREVIOUS answer's set,
|
|
520
|
+
// not a graph kind named "are"/"is"/"were"/"was".
|
|
521
|
+
if (IMPLICIT_ANAPHORA_COUNT_RE.test(String(query).trim())) return null;
|
|
400
522
|
const m = String(query).match(/\b(?:how many|number of|count(?:\s+the)?)\s+([a-z]+)\b/i);
|
|
401
523
|
if (!m) return null;
|
|
402
524
|
const noun = m[1].toLowerCase();
|
|
403
525
|
const cls = COUNT_NOUNS[noun];
|
|
526
|
+
if (cls && RESTRICTOR_VERB_RE.test(String(query).slice(m.index + m[0].length))) return null;
|
|
404
527
|
if (!cls) {
|
|
405
528
|
return `I can't count "${noun}". I count: ${countableKinds(graph).join(", ")}. ` +
|
|
406
529
|
`Try "how many classes are there".`;
|
|
@@ -434,6 +557,47 @@ async function countFromFacts(graph, memoryDir, query) {
|
|
|
434
557
|
return null;
|
|
435
558
|
}
|
|
436
559
|
|
|
560
|
+
// ---- Feature A point 4: "how many Xs are Ys" — literal recall of a taught
|
|
561
|
+
// quantifier ("some"/"a few"/"every"), NEVER real cardinality counting
|
|
562
|
+
// (consistent with this file's "grounded or honest miss" philosophy). The
|
|
563
|
+
// SOME_A_FEW_RE / unknownSubjectFallback / assertTurn's own "every"-quantifier
|
|
564
|
+
// follow-up (below) are what STORE the quantifier this reads back.
|
|
565
|
+
//
|
|
566
|
+
// CRITICAL ORDERING NOTE: dispatched explicitly ahead of answerCount in
|
|
567
|
+
// runTurn (mirroring answerMemoryCount's own precedent, just below) —
|
|
568
|
+
// answerCount's own noun-scan regex greedily grabs the FIRST word after "how
|
|
569
|
+
// many" as a literal noun to count and would otherwise short-circuit to an
|
|
570
|
+
// "I can't count 'Xs'" miss before this lane ever got a turn.
|
|
571
|
+
//
|
|
572
|
+
// AUTHORITY GATE (avoids shadowing real graph counts): claims authority
|
|
573
|
+
// (always returns a non-null string — either the quantifier or an honest "I
|
|
574
|
+
// don't know") ONLY when (a) the subject does NOT name a real graph-countable
|
|
575
|
+
// class (COUNT_NOUNS — the same guard countFromFacts uses, so a corpus-seeded
|
|
576
|
+
// fact that happens to share a subject word like "module" never shadows a
|
|
577
|
+
// real "how many modules …" count) AND (b) tmct has SOME isa-family fact
|
|
578
|
+
// about that subject at all (a subject never taught anything, e.g. "classes"
|
|
579
|
+
// in "how many classes are there", falls through to answerCount's real
|
|
580
|
+
// graph-cardinality count untouched — same honest-decline discipline as
|
|
581
|
+
// every other lane here).
|
|
582
|
+
const HOW_MANY_ARE_RE = /^how\s+many\s+([\w-]+)\s+(?:are|is)\s+(.+?)[?.!\s]*$/i;
|
|
583
|
+
async function answerQuantifierRecall(memoryDir, query) {
|
|
584
|
+
if (!memoryDir) return null;
|
|
585
|
+
const m = String(query).trim().match(HOW_MANY_ARE_RE);
|
|
586
|
+
if (!m) return null;
|
|
587
|
+
const asked = m[1].toLowerCase();
|
|
588
|
+
if (COUNT_NOUNS[asked]) return null; // a real graph-countable class — answerCount owns it
|
|
589
|
+
let normFactTerm;
|
|
590
|
+
try { ({ normFactTerm } = await import("./memory/core.mjs")); } catch { return null; }
|
|
591
|
+
const subjVariants = factTermVariants(normFactTerm, asked);
|
|
592
|
+
const rows = (await factRows(memoryDir)).filter((f) => ISA_PREDICATES.has(f.predicate) && subjVariants.has(f.subject));
|
|
593
|
+
if (!rows.length) return null; // never heard of this subject at all — let answerCount own the shape
|
|
594
|
+
const objVariants = factTermVariants(normFactTerm, m[2]);
|
|
595
|
+
const hit = rows.filter((f) => objVariants.has(f.object)).sort((a, b) => (b.trust ?? 0) - (a.trust ?? 0))[0];
|
|
596
|
+
const q = hit?.quantifier;
|
|
597
|
+
if (!q) return "I don't know — I was never told a quantifier for that.";
|
|
598
|
+
return `${q.charAt(0).toUpperCase()}${q.slice(1)}.`;
|
|
599
|
+
}
|
|
600
|
+
|
|
437
601
|
// ---- memory-store counts (the .tmct/memory graph, distinct from the code graph
|
|
438
602
|
// answerCount reads) — so "how many facts do you know" is answerable, consistent
|
|
439
603
|
// with what `/memory` advertises. The code graph owns the structural kinds
|
|
@@ -519,8 +683,24 @@ const IDENTITY_PHRASES = [
|
|
|
519
683
|
/^(tell me about|introduce) yourself\??$/i, /^what is this thing\??$/i,
|
|
520
684
|
/^what am i (talking|speaking|chatting) (to|with)\??$/i,
|
|
521
685
|
/^you are what\??$/i, /^what thing (are|is) you\??$/i,
|
|
522
|
-
|
|
686
|
+
// "explain [to me|please]* what (you are|this is)" in EITHER word order — a
|
|
687
|
+
// fluent-but-non-native speaker plausibly types the question-form "what is
|
|
688
|
+
// this" after "explain" as readily as the statement-form "this is" (SKILL_
|
|
689
|
+
// CHAT_PLAYTEST §3b's own ESL examples: "explain please what is this" used
|
|
690
|
+
// to fall through this regex to the grammar wall because only the statement
|
|
691
|
+
// order was declared).
|
|
692
|
+
/^explain(?:\s+(?:to me|please))*\s+what\s+(?:is\s+(?:this|it|you)|(?:you are|this is|it is))\??$/i,
|
|
523
693
|
/^whoami\??$/i,
|
|
694
|
+
// "hru" ("how are you") — GLUED texting shorthand: no word boundary inside it
|
|
695
|
+
// for a contraction pass (fuzzyConversationalMatch's SHORTHAND_CONTRACTIONS)
|
|
696
|
+
// to split on, so it earns its own closed-set entry instead, same as GREET/
|
|
697
|
+
// THANKS' hand-curated slang. Routed to identity-self (not a fake "doing
|
|
698
|
+
// great!" performance, nor the generic greeting card) — an honest "what I am"
|
|
699
|
+
// answer is the closest real thing tmct has to say to "how are you". "wyd"
|
|
700
|
+
// ("what are you doing") is deliberately NOT given a matching entry: it isn't
|
|
701
|
+
// an identity question and forcing one would be a fabricated route; it falls
|
|
702
|
+
// through to the honest generic orientation card same as before.
|
|
703
|
+
/^hru\??$/i,
|
|
524
704
|
];
|
|
525
705
|
/** "Are you an LLM/AI/bot" — tmct's actual positioning (no LLM, deterministic) is
|
|
526
706
|
* a genuinely different, more specific answer than the generic self-description,
|
|
@@ -723,13 +903,37 @@ function classifyConversational(phrase) {
|
|
|
723
903
|
if (phrase === "who are you" || phrase === "what are you" || phrase === "what is your name") return "identity";
|
|
724
904
|
return "capability";
|
|
725
905
|
}
|
|
906
|
+
/** Standalone-token texting shorthand for this lane ONLY: "r"→"are", "u"→"you",
|
|
907
|
+
* word-boundary matched so a substring inside a real word ("your", "sure",
|
|
908
|
+
* "minute") is never touched. This is the SAME normalization class as
|
|
909
|
+
* ask-vocab.mjs's CONTRACTIONS table (word-boundary, case-insensitive,
|
|
910
|
+
* longest-key-first — see interpret/normalize.mjs's tableRe), but deliberately
|
|
911
|
+
* NOT routed through that shared table/normalizeQuery: those feed ask.mjs's
|
|
912
|
+
* code-graph grammar pipeline, where a bare "u"/"r" plausibly collides with a
|
|
913
|
+
* real dotted identifier ("u.mjs" as a module name) — and conversationalTurn()
|
|
914
|
+
* never calls normalizeQuery at all, so extending the shared table wouldn't
|
|
915
|
+
* even reach this lane. Scoped locally to the fuzzy-conversational tier
|
|
916
|
+
* instead, applied BEFORE the candidate lookup below, so "waht r u"/"wat r u"
|
|
917
|
+
* first become "waht are you"/"wat are you" — within the existing bounded
|
|
918
|
+
* edit-distance of "who are you"/"what are you" — and resolve exactly the way
|
|
919
|
+
* a plain-English typo does. GLUED shorthand ("hru", "wyd") has no word
|
|
920
|
+
* boundary to split on and is NOT reached by this pass; see IDENTITY_PHRASES
|
|
921
|
+
* for "hru"'s separate closed-set entry. */
|
|
922
|
+
const SHORTHAND_CONTRACTIONS = { r: "are", u: "you" };
|
|
923
|
+
const SHORTHAND_CONTRACTION_RE = /\b(r|u)\b/gi;
|
|
924
|
+
function expandShorthandContractions(text) {
|
|
925
|
+
return text.replace(SHORTHAND_CONTRACTION_RE, (m) => SHORTHAND_CONTRACTIONS[m.toLowerCase()]);
|
|
926
|
+
}
|
|
927
|
+
|
|
726
928
|
/** UNIQUE within-bound fuzzy match of the whole trimmed line against
|
|
727
|
-
* CONVERSATIONAL_PHRASES — the "helo"/"thnx"/"
|
|
728
|
-
*
|
|
929
|
+
* CONVERSATIONAL_PHRASES — the "helo"/"thnx"/"byee" tier, plus (after shorthand
|
|
930
|
+
* contraction expansion above) "waht r u"/"wat r u"-style texting shorthand.
|
|
931
|
+
* Restricted to short (≤4-word), non-code-ish inputs (looksCodeish, shared with
|
|
729
932
|
* isConversational) so a genuine near-miss structural question is never grabbed;
|
|
730
933
|
* a distance tie is refused, never guessed (same discipline as fuzzyVocabWord). */
|
|
731
934
|
function fuzzyConversationalMatch(raw) {
|
|
732
|
-
const
|
|
935
|
+
const expanded = expandShorthandContractions(raw);
|
|
936
|
+
const q = collapseRuns(expanded.toLowerCase().replace(/[.!?]+$/, "").trim());
|
|
733
937
|
const words = q.split(/\s+/).filter(Boolean);
|
|
734
938
|
if (!words.length || words.length > 4 || looksCodeish(raw, q)) return null;
|
|
735
939
|
return fuzzyMatchInSet(q, CONVERSATIONAL_PHRASES, Math.min(2, fuzzyBound(q)));
|
|
@@ -1056,6 +1260,12 @@ const QUESTION_LEAD_RE = /^(?:what|who|which|where|when|why|how|is|are|do|does|d
|
|
|
1056
1260
|
// The teach lane's fact predicates (rendered via FACT_PREDICATE_PHRASES).
|
|
1057
1261
|
const OWNED_BY_PREDICATE = "mgx:ownedBy";
|
|
1058
1262
|
const HAS_PROPERTY_PREDICATE = "mgx:hasProperty";
|
|
1263
|
+
// Class-membership — the SAME predicate family the ACE grammar's own
|
|
1264
|
+
// subClassOf pattern emits (grammar/ace.mjs); named here too (Feature A) so
|
|
1265
|
+
// the new direct-write paths below (the unknown-subject fallback, the plural
|
|
1266
|
+
// "some/a few Xs are Ys" shape) stay obviously in that same family rather than
|
|
1267
|
+
// re-typing the CURIE string at each call site.
|
|
1268
|
+
const SUBCLASS_PREDICATE = "rdfs:subClassOf";
|
|
1059
1269
|
|
|
1060
1270
|
/** "<Name> owns/maintains <X>" — the ownership teach declarative. <Name> is one
|
|
1061
1271
|
* or two name tokens, <X> one code-ish token (a path, a file, a symbol). The
|
|
@@ -1078,7 +1288,7 @@ const teachProvenanceTag = (sessionId, ts) => `teach:chat${sessionId ? `:${sessi
|
|
|
1078
1288
|
/** Reify one teach-lane fact + confirm (shared by the property and ownership
|
|
1079
1289
|
* frames). Lazy + failure-tolerated: a write failure degrades to null (the
|
|
1080
1290
|
* teach-miss text stands), never a crash. */
|
|
1081
|
-
async function teachFact(memoryDir, sessionId, { subject, predicate, object }) {
|
|
1291
|
+
async function teachFact(memoryDir, sessionId, { subject, predicate, object, quantifier = "" }) {
|
|
1082
1292
|
try {
|
|
1083
1293
|
const { appendFact, normFactTerm } = await import("./memory/core.mjs");
|
|
1084
1294
|
const s = normFactTerm(subject);
|
|
@@ -1087,6 +1297,7 @@ async function teachFact(memoryDir, sessionId, { subject, predicate, object }) {
|
|
|
1087
1297
|
await appendFact(memoryDir, {
|
|
1088
1298
|
subject: s, predicate, object: o,
|
|
1089
1299
|
provenance: teachProvenanceTag(sessionId, new Date().toISOString()),
|
|
1300
|
+
...(quantifier ? { quantifier } : {}),
|
|
1090
1301
|
});
|
|
1091
1302
|
const phrase = FACT_PREDICATE_PHRASES[predicate] || predicate;
|
|
1092
1303
|
return { text: `noted — remembered: ${s} ${phrase} ${o}`, via: "assert", miss: false };
|
|
@@ -1095,6 +1306,99 @@ async function teachFact(memoryDir, sessionId, { subject, predicate, object }) {
|
|
|
1095
1306
|
}
|
|
1096
1307
|
}
|
|
1097
1308
|
|
|
1309
|
+
// ---- FEATURE A (0.9.x): teach new terms + quantifier phrasings ("every X is
|
|
1310
|
+
// a/an Y", "some Xs are Ys", "your X is a/an Y", "X is Y", "a few Xs are
|
|
1311
|
+
// Ys") + "how many Xs are Ys" recall. Design (from two prior read-only
|
|
1312
|
+
// investigations, live-verified): the memory Facts store and EVERY read path
|
|
1313
|
+
// (factAnswer, factReadBack, the 2-hop findIsaChain proof-chase) already work
|
|
1314
|
+
// generically over ANY subject string — the ONLY thing stopping e.g. "redis is
|
|
1315
|
+
// a cache" from being remembered is that parseAce's resolveNP (grammar/ace.mjs)
|
|
1316
|
+
// only resolves subjects/objects against the closed 180-word lexicon-core.json
|
|
1317
|
+
// noun list, so an unknown SUBJECT becomes residue and the whole sentence is
|
|
1318
|
+
// rejected even though the OBJECT ("cache") is a perfectly good known term. The
|
|
1319
|
+
// fix below is write-side only and deliberately NARROW: only the SUBJECT gets
|
|
1320
|
+
// a free pass, never the OBJECT — this is not a general lexicon bypass, it's
|
|
1321
|
+
// one additional storable shape alongside the ACE grammar's own 8 patterns. ----
|
|
1322
|
+
|
|
1323
|
+
/** Naive plural → singular fold for the "some/a few Xs are Ys" surface forms
|
|
1324
|
+
* (mirrors factTermVariants' own naive -es/-s stripping, below, but returns
|
|
1325
|
+
* ONE canonical spelling to STORE rather than a lookup Set of candidates to
|
|
1326
|
+
* match against). Deliberately tiny, no NLP — a stray false fold on an
|
|
1327
|
+
* already-singular noun ending in "s" is a known, accepted limitation of this
|
|
1328
|
+
* same naive scheme used elsewhere in this file (factTermVariants). */
|
|
1329
|
+
function singularizeSurface(word) {
|
|
1330
|
+
const w = String(word || "").trim();
|
|
1331
|
+
if (/[a-z]ies$/i.test(w)) return `${w.slice(0, -3)}y`;
|
|
1332
|
+
if (/(ses|xes|zes|ches|shes)$/i.test(w)) return w.slice(0, -2);
|
|
1333
|
+
if (/[a-z]s$/i.test(w) && !/ss$/i.test(w)) return w.slice(0, -1);
|
|
1334
|
+
return w;
|
|
1335
|
+
}
|
|
1336
|
+
|
|
1337
|
+
/** "some Xs are Ys" / "a few Xs are Ys" — the plural class-membership
|
|
1338
|
+
* quantifier shape. Captures the quantifier word itself (group 1) alongside
|
|
1339
|
+
* the plural subject/object (groups 2/3); singularized before storage/lookup. */
|
|
1340
|
+
const SOME_A_FEW_RE = /^(some|a few)\s+([\w-]+)\s+are\s+([\w-]+)$/i;
|
|
1341
|
+
|
|
1342
|
+
/** "(every|each|all|a|an )?X is/are (a|an )?Y" — the shape the unknown-subject
|
|
1343
|
+
* fallback recognizes (group 2 = X, group 3 = Y); group 1 (when present)
|
|
1344
|
+
* names the determiner, so the caller can tell a genuine "every" universal
|
|
1345
|
+
* apart from a singular/specific-entity "a"/bare reading (only "every" gets a
|
|
1346
|
+
* recorded quantifier here — this function's OWN caller passes it through to
|
|
1347
|
+
* teachFact; assertTurn, below, records the same "every" quantifier
|
|
1348
|
+
* independently for the pre-existing ACE-success path). Single-token X and Y
|
|
1349
|
+
* only — the same fragment scope parseAce's own copula patterns cover, just
|
|
1350
|
+
* with X's lexicon-membership requirement lifted. */
|
|
1351
|
+
const UNKNOWN_SUBJECT_RE = /^(every\s+|each\s+|all\s+|a\s+|an\s+)?([\w-]+)\s+(?:is|are)\s+(?:an?\s+)?([\w-]+)$/i;
|
|
1352
|
+
|
|
1353
|
+
/** The unknown-SUBJECT direct-write fallback (point 1 + point 2's bare-property
|
|
1354
|
+
* extension): tried ONLY after the real ACE grammar (assertTurn) has already
|
|
1355
|
+
* had its turn and declined. Declines itself (returns null, never a guess)
|
|
1356
|
+
* when:
|
|
1357
|
+
* - the payload doesn't fit the plain single-token "X is/are Y" shape at all
|
|
1358
|
+
* (a multi-word subject, a relation/cardinality/etc. sentence — those stay
|
|
1359
|
+
* the ACE grammar's territory, or the wrapped multi-word TEACH_PROPERTY_RE
|
|
1360
|
+
* path below, unchanged);
|
|
1361
|
+
* - X is actually a KNOWN lexicon word — then the ACE grammar's own miss was
|
|
1362
|
+
* a real structural/vocabulary problem elsewhere (e.g. Y itself unknown as
|
|
1363
|
+
* the WRONG part of speech), never silently reinterpreted through this
|
|
1364
|
+
* narrow exception;
|
|
1365
|
+
* - Y resolves as NEITHER a known noun NOR a known adjective — the OBJECT
|
|
1366
|
+
* must still be a term tmct actually knows; an unknown Y stays an honest
|
|
1367
|
+
* miss (never a guess), exactly like the pre-existing "monkey is an
|
|
1368
|
+
* animal" case.
|
|
1369
|
+
* Y resolving as a NOUN writes rdfs:subClassOf (mirrors the ACE grammar's own
|
|
1370
|
+
* subClassOf/typeAssertion pattern); Y resolving as an ADJECTIVE (and not also
|
|
1371
|
+
* a noun) writes mgx:hasProperty (mirrors the wrapped "remember that X is
|
|
1372
|
+
* deprecated" property frame — reused here for the bare/unwrapped form too,
|
|
1373
|
+
* since the free pass is about the SUBJECT, not about the "remember that"
|
|
1374
|
+
* wrapper). Only the "every" determiner records a quantifier (point 3: "a"/
|
|
1375
|
+
* bare/"your" read as one specific entity, not a class-level generalization). */
|
|
1376
|
+
async function unknownSubjectFallback(payload, { memoryDir, sessionId, lexicon }) {
|
|
1377
|
+
if (!memoryDir) return null;
|
|
1378
|
+
const m = String(payload).trim().match(UNKNOWN_SUBJECT_RE);
|
|
1379
|
+
if (!m) return null;
|
|
1380
|
+
const [, det, subjectRaw, objectRaw] = m;
|
|
1381
|
+
const { loadLexicon, lookupNoun, lookupAdjective, classify } = await import("./grammar/lexicon.mjs");
|
|
1382
|
+
const lex = lexicon || loadLexicon();
|
|
1383
|
+
// A known X's own ACE miss is a real miss — never silently reinterpreted here.
|
|
1384
|
+
if (classify(subjectRaw, lex)) return null;
|
|
1385
|
+
const quantifier = /^every$/i.test((det || "").trim()) ? "every" : "";
|
|
1386
|
+
if (lookupNoun(lex, objectRaw)) {
|
|
1387
|
+
return teachFact(memoryDir, sessionId, {
|
|
1388
|
+
subject: subjectRaw, predicate: SUBCLASS_PREDICATE, object: objectRaw, quantifier,
|
|
1389
|
+
});
|
|
1390
|
+
}
|
|
1391
|
+
if (lookupAdjective(lex, objectRaw)) {
|
|
1392
|
+
// property assertions are about ONE specific entity — never a quantifier,
|
|
1393
|
+
// even when phrased with "every" (point 3).
|
|
1394
|
+
return teachFact(memoryDir, sessionId, {
|
|
1395
|
+
subject: subjectRaw, predicate: HAS_PROPERTY_PREDICATE, object: objectRaw,
|
|
1396
|
+
});
|
|
1397
|
+
}
|
|
1398
|
+
return null; // Y unknown too — decline honestly, never guess
|
|
1399
|
+
}
|
|
1400
|
+
|
|
1401
|
+
|
|
1098
1402
|
/** Sentence forms to try asserting for a teach payload: the payload as-is, and
|
|
1099
1403
|
* (if it carries no determiner) its "every …" universal — the ACE-OWL shape the
|
|
1100
1404
|
* grammar actually lands. */
|
|
@@ -1104,16 +1408,38 @@ function assertCandidates(payload) {
|
|
|
1104
1408
|
if (!/^(?:every|each|all|a|an)\b/i.test(p)) out.push(`every ${p}`);
|
|
1105
1409
|
return [...new Set(out)];
|
|
1106
1410
|
}
|
|
1107
|
-
/** The "every X is a Y" rewrite of a declarative, for the "did you mean …" hint.
|
|
1411
|
+
/** The "every X is a Y" rewrite of a declarative, for the "did you mean …" hint.
|
|
1412
|
+
* BUG 2 fix (2026-07-08): the article was hardcoded to "a" regardless of Y's
|
|
1413
|
+
* vowel sound ("every monkey is a animal" — ungrammatical for a vowel-initial
|
|
1414
|
+
* Y), which made the suggestion silently WRONG for exactly the cases where a
|
|
1415
|
+
* correction is most useful. Real a/an agreement now reuses finish.mjs's own
|
|
1416
|
+
* beginsWithVowelSound + the SAME grammar-rules.toml "article" rule
|
|
1417
|
+
* (spelling-vowel/consonant exceptions included) rather than reimplementing
|
|
1418
|
+
* vowel-sound detection a second time. */
|
|
1108
1419
|
function teachSuggestion(payload) {
|
|
1109
1420
|
const m = String(payload).match(/^(?:every |each |all |a |an )?([\w-]+) (?:is|are) (?:a |an )?([\w-]+)$/i);
|
|
1110
|
-
|
|
1421
|
+
if (!m) return null;
|
|
1422
|
+
const subject = m[1].toLowerCase();
|
|
1423
|
+
const object = m[2].toLowerCase();
|
|
1424
|
+
const articleRule = grammarRules().find((r) => r.kind === "article");
|
|
1425
|
+
const article = articleRule && beginsWithVowelSound(object, articleRule) ? "an" : "a";
|
|
1426
|
+
return `every ${subject} is ${article} ${object}`;
|
|
1111
1427
|
}
|
|
1112
1428
|
|
|
1113
1429
|
async function teachLane(query, { memoryDir, sessionId = "", lexicon = null }) {
|
|
1114
|
-
const
|
|
1115
|
-
const m =
|
|
1116
|
-
const
|
|
1430
|
+
const rawInput = String(query).trim();
|
|
1431
|
+
const m = rawInput.match(TEACH_RE);
|
|
1432
|
+
const wrappedInput = m ? m[1].trim() : null;
|
|
1433
|
+
// "your X is a/an Y" (Feature A) — a plain casual synonym for "a/an X is a
|
|
1434
|
+
// Y": no special second-person semantics, so rewrite it to the ordinary
|
|
1435
|
+
// indefinite-article determiner UP FRONT, before any downstream regex/ACE
|
|
1436
|
+
// parsing ever sees it (ACE itself has no notion of "your" as a
|
|
1437
|
+
// determiner). Only a LEADING "your" is rewritten, so this can't misfire on
|
|
1438
|
+
// a "your" appearing mid-sentence; applied to both the bare and the
|
|
1439
|
+
// remember-wrapped surface.
|
|
1440
|
+
const stripYour = (s) => (s == null ? s : s.replace(/^your\s+/i, "a "));
|
|
1441
|
+
const raw = stripYour(rawInput);
|
|
1442
|
+
const wrapped = stripYour(wrappedInput);
|
|
1117
1443
|
|
|
1118
1444
|
// OWNERSHIP — "<Name> owns/maintains <X>", bare or remember-wrapped. The bare
|
|
1119
1445
|
// form is double-gated: a Capitalized name AND no interrogative lead, so the
|
|
@@ -1127,6 +1453,32 @@ async function teachLane(query, { memoryDir, sessionId = "", lexicon = null }) {
|
|
|
1127
1453
|
if (stored) return stored;
|
|
1128
1454
|
}
|
|
1129
1455
|
|
|
1456
|
+
// "some Xs are Ys" / "a few Xs are Ys" (Feature A) — the plural class-
|
|
1457
|
+
// membership quantifier shape. ACE has no quantifier-phrase pattern at all
|
|
1458
|
+
// (parseAce never even attempts a fit), so this is ALWAYS a direct write,
|
|
1459
|
+
// never routed through assertTurn below. Wrapper-optional, like the
|
|
1460
|
+
// "every X is a Y" baseline — a plural "some/a few" claim reads as an
|
|
1461
|
+
// ordinary declarative teach the same way "every" always has. The OBJECT
|
|
1462
|
+
// still has to be a known lexicon noun (the same "subject gets the free
|
|
1463
|
+
// pass, object doesn't" discipline as unknownSubjectFallback below) — an
|
|
1464
|
+
// unknown object falls through to the generic honest-miss cascade at the
|
|
1465
|
+
// bottom of this function, same as every other unstorable teach.
|
|
1466
|
+
const someSrc = wrapped ?? raw.replace(/[.!?]+\s*$/, "");
|
|
1467
|
+
const someMatch = memoryDir && !QUESTION_LEAD_RE.test(someSrc) ? someSrc.match(SOME_A_FEW_RE) : null;
|
|
1468
|
+
if (someMatch) {
|
|
1469
|
+
const quantifier = someMatch[1].toLowerCase();
|
|
1470
|
+
const subject = singularizeSurface(someMatch[2]);
|
|
1471
|
+
const object = singularizeSurface(someMatch[3]);
|
|
1472
|
+
const { loadLexicon, lookupNoun } = await import("./grammar/lexicon.mjs");
|
|
1473
|
+
const lex = lexicon || loadLexicon();
|
|
1474
|
+
if (lookupNoun(lex, object)) {
|
|
1475
|
+
const stored = await teachFact(memoryDir, sessionId, {
|
|
1476
|
+
subject, predicate: SUBCLASS_PREDICATE, object, quantifier,
|
|
1477
|
+
});
|
|
1478
|
+
if (stored) return stored;
|
|
1479
|
+
}
|
|
1480
|
+
}
|
|
1481
|
+
|
|
1130
1482
|
let payload = null;
|
|
1131
1483
|
if (wrapped && /\b(?:is|are)\b/i.test(wrapped)) payload = wrapped;
|
|
1132
1484
|
else if (BARE_DECLARATIVE_RE.test(raw) && !QUESTION_LEAD_RE.test(raw)) payload = raw;
|
|
@@ -1135,9 +1487,20 @@ async function teachLane(query, { memoryDir, sessionId = "", lexicon = null }) {
|
|
|
1135
1487
|
// the "noted — remembered …" confirmation or null (grammar miss / unknown words).
|
|
1136
1488
|
if (memoryDir) {
|
|
1137
1489
|
for (const cand of assertCandidates(payload)) {
|
|
1490
|
+
// assertTurn ITSELF records the "every" quantifier (point 3) on a plain
|
|
1491
|
+
// universal success, so every caller (this loop AND the top-level
|
|
1492
|
+
// declarative-sentence dispatch in runTurn) gets it uniformly.
|
|
1138
1493
|
const stored = await assertTurn(cand, { memoryDir, sessionId, focus: null, lexicon });
|
|
1139
1494
|
if (stored) return { text: stored.answer, via: "assert", miss: false };
|
|
1140
1495
|
}
|
|
1496
|
+
// BUG "redis" fix (Feature A point 1): the real ACE grammar just declined
|
|
1497
|
+
// (unknown words / not the membership shape) — try the narrow unknown-
|
|
1498
|
+
// SUBJECT direct-write fallback before falling to the honest-miss cascade.
|
|
1499
|
+
// Covers BOTH the bare and the wrapped surface (payload is already
|
|
1500
|
+
// unwrapped either way) — see unknownSubjectFallback's own docblock for
|
|
1501
|
+
// the exact narrowing rules (object must still be known, etc.).
|
|
1502
|
+
const fallback = await unknownSubjectFallback(payload, { memoryDir, sessionId, lexicon });
|
|
1503
|
+
if (fallback) return fallback;
|
|
1141
1504
|
// PROPERTY teach — "remember/note that <X> is <adjective>": wrapper-REQUIRED
|
|
1142
1505
|
// (a bare "X is deprecated" is never silently reified), and only after the
|
|
1143
1506
|
// ACE grammar declined (unknown words / not the membership shape), so a
|
|
@@ -1152,10 +1515,52 @@ async function teachLane(query, { memoryDir, sessionId = "", lexicon = null }) {
|
|
|
1152
1515
|
}
|
|
1153
1516
|
}
|
|
1154
1517
|
}
|
|
1518
|
+
// BUG 2 fix (2026-07-08): compare the CORRECTED suggestion against a
|
|
1519
|
+
// normalized (trimmed, whitespace-collapsed, lowercased) form of what the
|
|
1520
|
+
// user actually typed, not the raw payload — so trivial formatting
|
|
1521
|
+
// differences never manufacture a spurious "did you mean". With
|
|
1522
|
+
// teachSuggestion's article now grammatically correct (above), this
|
|
1523
|
+
// equality guard's original intent is restored rather than replaced: it
|
|
1524
|
+
// suppresses the hint exactly when X and Y themselves are already spelled
|
|
1525
|
+
// in the canonical "every X is a Y" shape (nothing useful to add), and
|
|
1526
|
+
// shows it whenever the corrected form differs — including the wrong-
|
|
1527
|
+
// article case ("every monkey is a animal") that used to be silently
|
|
1528
|
+
// suppressed because the OLD teachSuggestion's own hardcoded "a" matched
|
|
1529
|
+
// the user's mistake byte-for-byte.
|
|
1530
|
+
const normalizedPayload = String(payload).trim().toLowerCase().replace(/\s+/g, " ");
|
|
1155
1531
|
const suggestion = teachSuggestion(payload);
|
|
1156
|
-
const did = suggestion && suggestion !==
|
|
1532
|
+
const did = suggestion && suggestion !== normalizedPayload ? ` Did you mean: "${suggestion}"?` : "";
|
|
1533
|
+
// Honest miss reason (2026-07-08, "separately, not a bug" clarification): when
|
|
1534
|
+
// the payload structurally fits the ACE fragment but names word(s) outside
|
|
1535
|
+
// tmct's closed 180-word lexicon (lexicon-core.json), parseAce already
|
|
1536
|
+
// reports exactly which tokens are unrecognized as `residue` — assertTurn's
|
|
1537
|
+
// loop above discards it on a miss. Re-derive it here (same lexicon, same
|
|
1538
|
+
// candidate sentences) so the miss message can NAME the word(s), rather than
|
|
1539
|
+
// leaving the user to guess whether the problem was grammar shape or
|
|
1540
|
+
// vocabulary. A payload that doesn't fit the fragment AT ALL (parseAce
|
|
1541
|
+
// returns null, no residue) gets the plain generic message — genuinely a
|
|
1542
|
+
// shape mismatch, not an unrecognized-word one. This does NOT widen the
|
|
1543
|
+
// lexicon itself: "redis"/"monkey"/"animal" still fail to store; the
|
|
1544
|
+
// message now says why.
|
|
1545
|
+
let unknown = [];
|
|
1546
|
+
if (memoryDir) {
|
|
1547
|
+
try {
|
|
1548
|
+
const { parseAce } = await import("./grammar/ace.mjs");
|
|
1549
|
+
let lex = lexicon;
|
|
1550
|
+
if (!lex) { const { loadLexicon } = await import("./grammar/lexicon.mjs"); lex = loadLexicon(); }
|
|
1551
|
+
for (const cand of assertCandidates(payload)) {
|
|
1552
|
+
const parse = parseAce(cand, lex);
|
|
1553
|
+
if (parse?.residue?.length) { unknown = [...new Set(parse.residue.map((w) => String(w).toLowerCase()))]; break; }
|
|
1554
|
+
}
|
|
1555
|
+
} catch { /* lexicon unavailable — fall through to the generic message */ }
|
|
1556
|
+
}
|
|
1557
|
+
const why = unknown.length
|
|
1558
|
+
? ` I don't recognize ${joinList(unknown.map((w) => `"${w}"`))} as ${unknown.length === 1 ? "a word" : "words"} I know — `
|
|
1559
|
+
+ "I can only teach facts using tmct's own code-vocabulary nouns (like module, class, function…), "
|
|
1560
|
+
+ "not arbitrary new terms."
|
|
1561
|
+
: "";
|
|
1157
1562
|
return {
|
|
1158
|
-
text:
|
|
1563
|
+
text: `I couldn't store that —${why} I remember facts in the shape "every X is a Y", where X and Y are `
|
|
1159
1564
|
+ `words I know.${did} Type /memory to see what I already remember.`,
|
|
1160
1565
|
via: "teach-miss", miss: true,
|
|
1161
1566
|
};
|
|
@@ -1177,13 +1582,18 @@ const META_ORIENT_RE = /^(?:what(?:'s| is| are)?\s+this(?:\s+(?:app|codebase|rep
|
|
|
1177
1582
|
* vocabulary seeding either hasn't run or produced nothing, so the hook makes NO
|
|
1178
1583
|
* term-specific promise (an unconditionally-true pointer: the teach lane and
|
|
1179
1584
|
* `tmct init` both work with zero preconditions), rather than suggesting a
|
|
1180
|
-
* vocabulary example that would be guaranteed to miss right after being offered.
|
|
1585
|
+
* vocabulary example that would be guaranteed to miss right after being offered.
|
|
1586
|
+
* The no-code-graph branch's teach example is a CONCRETE pair from the closed
|
|
1587
|
+
* ACE lexicon (playtest: an abstract "every X is a Y" invites a curious user to
|
|
1588
|
+
* substitute intuitive-but-unknown words — "every cache is a thing" — which the
|
|
1589
|
+
* closed lexicon then rejects; "every bug is an issue" is confirmed to parse and
|
|
1590
|
+
* store, see test/chatflow-tier0.test.mjs). */
|
|
1181
1591
|
async function memorySummary(memoryDir, graph) {
|
|
1182
1592
|
const rows = memoryDir ? await memoryFacts(memoryDir) : [];
|
|
1183
1593
|
if (!rows.length) {
|
|
1184
1594
|
const hook = moduleCountOf(graph) > 0
|
|
1185
1595
|
? 'ask about this codebase\'s structure (imports, calls, definitions), or teach me with "every X is a Y"'
|
|
1186
|
-
: 'run `tmct init` to seed a starter vocabulary, or teach me directly
|
|
1596
|
+
: 'run `tmct init` to seed a starter vocabulary, or teach me directly, e.g. "every bug is an issue"';
|
|
1187
1597
|
return `I haven't been told any facts yet — ${hook}. /memory to inspect, /help for commands.`;
|
|
1188
1598
|
}
|
|
1189
1599
|
const preds = new Set(rows.map((f) => f.predicate).filter(Boolean));
|
|
@@ -1306,6 +1716,56 @@ const IMPERATIVE_NUDGE_RE =
|
|
|
1306
1716
|
/^(?:please\s+)?(?:(?:can|could|would|will)\s+you\s+(?:please\s+)?)?(?:make|write|create|add|generate|implement|fix|refactor)\b(?=.*\b(?:tests?|code|functions?|methods?|modules?|class(?:es)?|files?|it)\b)/i;
|
|
1307
1717
|
const WHY_UNTESTED_RE = /^why\s+(?:is|are)(?:n't|\s+not)?\s+(.+?)\s+(?:untested|not\s+tested|uncovered)$/i;
|
|
1308
1718
|
|
|
1719
|
+
// #5(g) OUT-OF-DOMAIN PERSONAL-ASSISTANT NUDGE (BUG 3 fix, 2026-07-08): "what
|
|
1720
|
+
// time is it" / "what's the weather" / "what day is it" — obviously not an
|
|
1721
|
+
// attempted code-graph query at all (no structural noun/verb), but 4+ words
|
|
1722
|
+
// with no dotted/camelCase/"()" token, so it slips past BOTH looksCodeish and
|
|
1723
|
+
// isConversational's ≤3-word catch-all, straight to the raw grammar wall
|
|
1724
|
+
// ("couldn't parse this as a graph question. Try: ...") — a dead-end per
|
|
1725
|
+
// SKILL_CHAT_PLAYTEST.md §0 ("every turn either answers, or gives a guiding
|
|
1726
|
+
// nudge... a turn that does neither is a dead-end"). A small closed set, same
|
|
1727
|
+
// discipline as RISK_NUDGE_RE/OPINION_NUDGE_RE above: this is a genuine
|
|
1728
|
+
// capability ceiling (tmct has no clock/calendar/weather capability) — the
|
|
1729
|
+
// fix is an honest decline pointing back at what tmct actually does, never a
|
|
1730
|
+
// fabricated time/date/weather answer.
|
|
1731
|
+
// "what(?:'s|s|\s+is)" also accepts the bare "whats" spelling — the same
|
|
1732
|
+
// informal contraction ask-vocab.mjs's own CONTRACTIONS table maps to "what
|
|
1733
|
+
// is" for the graph-query path; nudgeAnswer sees the raw (not contraction-
|
|
1734
|
+
// normalized) query text, so it earns its own tolerance here too.
|
|
1735
|
+
const PERSONAL_ASSISTANT_NUDGE_RE = new RegExp(
|
|
1736
|
+
"^(?:"
|
|
1737
|
+
+ "what\\s+time\\s+is\\s+it(?:\\s+(?:now|right\\s+now))?"
|
|
1738
|
+
+ "|what(?:'s|s|\\s+is)\\s+the\\s+time(?:\\s+(?:now|right\\s+now))?"
|
|
1739
|
+
+ "|what\\s+day\\s+is\\s+it(?:\\s+today)?"
|
|
1740
|
+
+ "|what(?:'s|s|\\s+is)\\s+(?:the\\s+)?(?:day|date)(?:\\s+today)?"
|
|
1741
|
+
+ "|what(?:'s|s|\\s+is)\\s+today'?s\\s+date"
|
|
1742
|
+
+ "|what(?:'s|s|\\s+is)\\s+the\\s+weather(?:\\s+like)?(?:\\s+(?:today|outside))?"
|
|
1743
|
+
+ "|how'?s\\s+the\\s+weather(?:\\s+like)?(?:\\s+(?:today|outside))?"
|
|
1744
|
+
+ ")\\??$",
|
|
1745
|
+
"i",
|
|
1746
|
+
);
|
|
1747
|
+
|
|
1748
|
+
/** STACCATO NEGATION ("not X", "not X then", "except X") — SKILL_CHAT_PLAYTEST
|
|
1749
|
+
* Tier-2, 5th pass: a rapid-fire rejection of a specific item, with no verb at
|
|
1750
|
+
* all — the bare-connective sibling of STACCATO_PRONOUN_RE/STACCATO_SWAP_RE
|
|
1751
|
+
* (below), but with no positive alternative named. Two flavors, BOTH
|
|
1752
|
+
* genuinely unanswerable as a real graph query (never fabricated):
|
|
1753
|
+
* - a BARE pronoun rejection ("not that one", "not those", "not it") names
|
|
1754
|
+
* no alternative at all — what the user DOES want instead is known only
|
|
1755
|
+
* to them, not derivable from the graph.
|
|
1756
|
+
* - a NAMED rejection ("not app/lib/b.mjs", "not Widget then") names a real
|
|
1757
|
+
* candidate to EXCLUDE from a just-given list, but excluding a member
|
|
1758
|
+
* from a prior result set is a capability the engine genuinely doesn't
|
|
1759
|
+
* have yet (verified live: even the fully-spelled "which of those is not
|
|
1760
|
+
* X" doesn't compile — parsePredicateFilter has no negation branch).
|
|
1761
|
+
* Before this, both fell to the generic orientation card (a short,
|
|
1762
|
+
* non-codeish turn trips isConversational's ≤3-word catch-all) or the raw
|
|
1763
|
+
* grammar wall (a codeish one, e.g. a path) — neither names what actually
|
|
1764
|
+
* went wrong. This is an honest, GUIDING nudge (§0), never a fabricated
|
|
1765
|
+
* filtered answer and never a bare wall. */
|
|
1766
|
+
const STACCATO_NEGATION_RE = /^(?:and\s+)?(?:not|except(?:\s+for)?)\s+(.+?)(?:\s+then|\s+though)?[?.!]*$/i;
|
|
1767
|
+
const NEGATION_PRONOUN_RE = /^(?:it|that|this|those|them)(?:\s+ones?)?$/i;
|
|
1768
|
+
|
|
1309
1769
|
/** The <name> a nudge shows: the focus label when the query leans on a pronoun (or
|
|
1310
1770
|
* gave us nothing better), else the captured subject; "<name>" as the placeholder. */
|
|
1311
1771
|
function nudgeName(captured, focus) {
|
|
@@ -1320,6 +1780,10 @@ function nudgeName(captured, focus) {
|
|
|
1320
1780
|
* short-miss rewrite). */
|
|
1321
1781
|
function nudgeAnswer(query, focus) {
|
|
1322
1782
|
const q = String(query).trim().replace(/[?.!]+$/, "").replace(/\s+/g, " ");
|
|
1783
|
+
if (PERSONAL_ASSISTANT_NUDGE_RE.test(q)) {
|
|
1784
|
+
return "I don't have access to that — I'm a deterministic code/vocabulary assistant, not a general assistant. "
|
|
1785
|
+
+ 'Ask me about code structure ("which modules import <name>") or try "what is a cache".';
|
|
1786
|
+
}
|
|
1323
1787
|
if (OPINION_NUDGE_RE.test(q)) {
|
|
1324
1788
|
const name = focus?.label || "<name>";
|
|
1325
1789
|
return "I don't hold opinions — I read structure, not quality. I can show what an opinion would rest on: "
|
|
@@ -1341,6 +1805,16 @@ function nudgeAnswer(query, focus) {
|
|
|
1341
1805
|
return "I don't write code — I read a graph of it. "
|
|
1342
1806
|
+ `/tests ${name} shows what covers it; "untested modules" shows the gaps.`;
|
|
1343
1807
|
}
|
|
1808
|
+
const neg = q.match(STACCATO_NEGATION_RE);
|
|
1809
|
+
if (neg) {
|
|
1810
|
+
const term = neg[1].trim();
|
|
1811
|
+
if (NEGATION_PRONOUN_RE.test(term)) {
|
|
1812
|
+
const name = focus?.label || "Widget";
|
|
1813
|
+
return `not sure what you'd like instead of ${focus?.label || "that"} — name it directly, e.g. "what calls ${name}".`;
|
|
1814
|
+
}
|
|
1815
|
+
return "I can't filter a previous list by exclusion yet — ask the positive shape directly "
|
|
1816
|
+
+ `(e.g. "which modules import <name>"), or ask about ${term} on its own.`;
|
|
1817
|
+
}
|
|
1344
1818
|
return null;
|
|
1345
1819
|
}
|
|
1346
1820
|
|
|
@@ -1683,6 +2157,46 @@ const FACT_PREDICATE_PHRASES = {
|
|
|
1683
2157
|
};
|
|
1684
2158
|
const factPhrase = (f) => `${f.subject} ${FACT_PREDICATE_PHRASES[f.predicate] || f.predicate} ${f.object}`;
|
|
1685
2159
|
|
|
2160
|
+
// ---- BUG 1 fix (2026-07-08): "what is a tree used for" filters to JUST the
|
|
2161
|
+
// UsedFor facts, instead of grammar.mjs's meta-whatis template's lazy tail
|
|
2162
|
+
// swallowing "tree used for" whole as one literal term (a guaranteed
|
|
2163
|
+
// vocabulary-lookup miss — "tree used for" names no class/predicate). Reuses
|
|
2164
|
+
// FACT_PREDICATE_PHRASES itself as the marker vocabulary (no second table):
|
|
2165
|
+
// every phrase that reads as "<copula> <marker>" (e.g. "is used for", "is
|
|
2166
|
+
// part of") derives a trailing marker ("used for", "part of") a "what is a
|
|
2167
|
+
// <subject> <marker>" question can end on, since the leading "is" is already
|
|
2168
|
+
// consumed by the template's own "what is" anchor. Phrases with no leading
|
|
2169
|
+
// is/are copula ("can", "causes", "requires", "has", …) don't fit that
|
|
2170
|
+
// question shape at all and are correctly excluded automatically — this is a
|
|
2171
|
+
// DERIVATION, not a curated subset. The single-letter "a" (from rdf:type's
|
|
2172
|
+
// bare "is a") is excluded explicitly: too short to anchor on without a real
|
|
2173
|
+
// risk of eating a genuine multi-word subject ending in "a".
|
|
2174
|
+
const TRAILING_PREDICATE_MARKERS = Object.entries(FACT_PREDICATE_PHRASES)
|
|
2175
|
+
.map(([predicate, phrase]) => {
|
|
2176
|
+
const m = /^(?:is|are)\s+(.+)$/i.exec(phrase);
|
|
2177
|
+
return m ? { predicate, marker: m[1].trim().toLowerCase() } : null;
|
|
2178
|
+
})
|
|
2179
|
+
.filter((e) => e && e.marker.length > 1)
|
|
2180
|
+
.sort((a, b) => b.marker.length - a.marker.length); // longest marker first
|
|
2181
|
+
|
|
2182
|
+
/** Split a meta-shaped term into {subject, predicate}: "tree used for" ->
|
|
2183
|
+
* {subject:"tree", predicate:"mgx:usedFor"} when the term ends in a known
|
|
2184
|
+
* TRAILING_PREDICATE_MARKERS marker with a non-empty subject ahead of it;
|
|
2185
|
+
* otherwise {subject: term, predicate: null} (the term stands as-is — the
|
|
2186
|
+
* ordinary undifferentiated "what is a X" behavior). Pure, no I/O. */
|
|
2187
|
+
function splitMetaPredicate(term) {
|
|
2188
|
+
const t = String(term || "").trim();
|
|
2189
|
+
const lower = t.toLowerCase();
|
|
2190
|
+
for (const { marker, predicate } of TRAILING_PREDICATE_MARKERS) {
|
|
2191
|
+
if (lower === marker) continue; // no subject left to the left of the marker
|
|
2192
|
+
if (lower.endsWith(` ${marker}`)) {
|
|
2193
|
+
const subject = t.slice(0, t.length - marker.length).trim();
|
|
2194
|
+
if (subject) return { subject, predicate };
|
|
2195
|
+
}
|
|
2196
|
+
}
|
|
2197
|
+
return { subject: t, predicate: null };
|
|
2198
|
+
}
|
|
2199
|
+
|
|
1686
2200
|
/** One rendered fact line. An OPERATOR-asserted fact keeps the true first-person
|
|
1687
2201
|
* provenance ("you told me: …"). A CORPUS fact is presented as clean DATA with its
|
|
1688
2202
|
* source cited — NEVER "i learned: …", which over-claims and anthropomorphises
|
|
@@ -1866,9 +2380,29 @@ async function factAnswer(memoryDir, query, envelope, miss) {
|
|
|
1866
2380
|
if (m) metaTerm = m[1];
|
|
1867
2381
|
}
|
|
1868
2382
|
if (metaTerm) {
|
|
1869
|
-
|
|
1870
|
-
|
|
1871
|
-
if (
|
|
2383
|
+
// BUG 1 fix: "what is a tree used for" parses (grammar.mjs T5) to the
|
|
2384
|
+
// WHOLE tail "tree used for" as one literal term — split off a trailing
|
|
2385
|
+
// FACT_PREDICATE_PHRASES marker (if any) so the real subject ("tree") is
|
|
2386
|
+
// matched against fact subjects, and — the actual bug — the result is
|
|
2387
|
+
// FILTERED to just that one predicate (mgx:usedFor) instead of every
|
|
2388
|
+
// relation about the subject undifferentiated.
|
|
2389
|
+
const { subject, predicate } = splitMetaPredicate(metaTerm);
|
|
2390
|
+
const variants = factTermVariants(normFactTerm, subject);
|
|
2391
|
+
const subjectHits = (await memoryFacts(memoryDir)).filter((f) => variants.has(f.subject));
|
|
2392
|
+
const hits = predicate ? subjectHits.filter((f) => f.predicate === predicate) : subjectHits;
|
|
2393
|
+
if (!hits.length) {
|
|
2394
|
+
// The subject itself is known, but not under this specific relation —
|
|
2395
|
+
// an honest, specific "no" rather than falling through to the generic
|
|
2396
|
+
// "isn't a term in this graph's own vocabulary" wall (which would be
|
|
2397
|
+
// actively misleading here: the subject IS a known term).
|
|
2398
|
+
if (predicate && subjectHits.length) {
|
|
2399
|
+
return {
|
|
2400
|
+
text: `I don't have any "${FACT_PREDICATE_PHRASES[predicate]}" facts about ${subject}.`,
|
|
2401
|
+
replace: miss,
|
|
2402
|
+
};
|
|
2403
|
+
}
|
|
2404
|
+
return null;
|
|
2405
|
+
}
|
|
1872
2406
|
const lines = hits.map(renderFactLine);
|
|
1873
2407
|
const shown = lines.slice(0, FACT_ANSWER_CAP);
|
|
1874
2408
|
const rest = lines.slice(FACT_ANSWER_CAP);
|
|
@@ -2273,9 +2807,28 @@ async function recallSummary(memoryDir) {
|
|
|
2273
2807
|
/** "[and/so/…] what about X" — a discourse continuation that re-asks the previous
|
|
2274
2808
|
* turn's question with X swapped in. */
|
|
2275
2809
|
const WHAT_ABOUT_RE = /^(?:(?:and|so|but|ok|okay|now|then)\s+)*what about\s+(.+?)[?.!\s]*$/i;
|
|
2276
|
-
/** A code-ish name token in a prior query (a path/dotted name,
|
|
2277
|
-
*
|
|
2278
|
-
|
|
2810
|
+
/** A code-ish name token in a prior query (a path/dotted name, a Capitalized
|
|
2811
|
+
* symbol, or a lowerCamelCase identifier like `saveStore`/`createTask`) — the
|
|
2812
|
+
* subject "what about X" replaces. The lowerCamelCase alternative (0.9.13
|
|
2813
|
+
* Tier-1 playtest) closes a real drill-down gap: a chain focused on a FUNCTION
|
|
2814
|
+
* ("what does saveStore call") has no Capitalized/path token at all, so "what
|
|
2815
|
+
* about X" after it used to fall straight through to the honest-miss instead
|
|
2816
|
+
* of continuing the shape — a mid-word capital never occurs in plain English,
|
|
2817
|
+
* so this is a safe, unambiguous code-identifier signal. */
|
|
2818
|
+
const NAME_TOKEN_RE = /\b[\w-]+(?:[/.][\w-]+)+\b|\b[A-Z][A-Za-z0-9_]*\b|\b[a-z][a-z0-9]*[A-Z][A-Za-z0-9]*\b/;
|
|
2819
|
+
|
|
2820
|
+
/** STACCATO SWAP CONTINUATION (0.9.15 Tier-2 playtest, 4th pass): the bare-
|
|
2821
|
+
* connective sibling of WHAT_ABOUT_RE — "and Widget?", "also app/lib/b.mjs" —
|
|
2822
|
+
* with no "about" at all. A rapid-fire drill-down chain naturally shortens
|
|
2823
|
+
* to this once the shape is established ("what calls app/lib/a.mjs" -> "and
|
|
2824
|
+
* Widget?" meaning "and what calls Widget?"). Unlike WHAT_ABOUT_RE's
|
|
2825
|
+
* explicit question framing, a bare connective is otherwise too ambiguous
|
|
2826
|
+
* with ordinary discourse ("and then?", "so what") to safely reinterpret as
|
|
2827
|
+
* a subject swap — discourseRewrite below only trusts this shape when the
|
|
2828
|
+
* captured word is ITSELF unambiguously code-ish (NAME_TOKEN_RE): a path, a
|
|
2829
|
+
* Capitalized symbol, or lowerCamelCase. A plain word ("and stuff?") never
|
|
2830
|
+
* matches and falls through unchanged. */
|
|
2831
|
+
const STACCATO_SWAP_RE = /^(?:and|also|so|then|now)\s+(.+?)[?.!\s]*$/i;
|
|
2279
2832
|
|
|
2280
2833
|
/** DISCOURSE CONTINUATION (CHATBENCH_006 lever 2): "what about X" carries the PRIOR
|
|
2281
2834
|
* turn's question shape across the turn boundary — re-asking it with X in place of
|
|
@@ -2284,10 +2837,18 @@ const NAME_TOKEN_RE = /\b[\w-]+(?:[/.][\w-]+)+\b|\b[A-Z][A-Za-z0-9_]*\b/;
|
|
|
2284
2837
|
* no prior query or no name token to swap (→ the ordinary honest miss stands). */
|
|
2285
2838
|
function discourseRewrite(query, last) {
|
|
2286
2839
|
const m = String(query).match(WHAT_ABOUT_RE);
|
|
2287
|
-
|
|
2840
|
+
let newSubj;
|
|
2841
|
+
if (m) {
|
|
2842
|
+
newSubj = m[1].trim();
|
|
2843
|
+
} else {
|
|
2844
|
+
const sm = String(query).match(STACCATO_SWAP_RE);
|
|
2845
|
+
const cand = sm?.[1]?.trim();
|
|
2846
|
+
if (!cand || !NAME_TOKEN_RE.test(cand)) return null;
|
|
2847
|
+
newSubj = cand;
|
|
2848
|
+
}
|
|
2849
|
+
if (!last?.query) return null;
|
|
2288
2850
|
const prevQ = String(last.query);
|
|
2289
2851
|
if (!NAME_TOKEN_RE.test(prevQ)) return null;
|
|
2290
|
-
const newSubj = m[1].trim();
|
|
2291
2852
|
return prevQ.replace(NAME_TOKEN_RE, () => newSubj);
|
|
2292
2853
|
}
|
|
2293
2854
|
|
|
@@ -2405,11 +2966,61 @@ async function curatedDefinitionAnswer(query, envelope, { memoryDir, lexicon })
|
|
|
2405
2966
|
* "imports"), which the RELATION force must never preempt (frozen case
|
|
2406
2967
|
* am-meta-imports). Gated downstream by CONCEPT_CLASS / RELATION_TERM, so a real
|
|
2407
2968
|
* entity name declines here. */
|
|
2969
|
+
// "tel" -> "tell" (0.9.14 Tier-2 playtest): the dropped-letter typo of THIS
|
|
2970
|
+
// lane's own anchor word — "tel me about calls" used to miss the "^tell me
|
|
2971
|
+
// about …" regex entirely and fall through to a bogus "no module matching
|
|
2972
|
+
// 'tel me'" search. "tell" is not itself part of ask.mjs's code-graph grammar
|
|
2973
|
+
// (VERB_TO_KIND/ENTITY_TO_TYPE/anchor words), so it can't live in the shared
|
|
2974
|
+
// ask-vocab.mjs MISSPELLINGS table (test/ask-vocab.test.mjs enforces every
|
|
2975
|
+
// correction value is grammar-owned) — same reasoning as chat.mjs's own
|
|
2976
|
+
// SHORTHAND_CONTRACTIONS above: scoped locally to the lane that owns the word.
|
|
2977
|
+
// Word-boundary matched so "hotel"/"intel" are untouched.
|
|
2978
|
+
const VAGUE_TOUCH_TEL_RE = /\btel\b/i;
|
|
2979
|
+
/** "explain X" / "please explain X" / "kindly explain X" / "explain X to me" /
|
|
2980
|
+
* "explain X please" — a bare vague-touch shape, sibling of WHAT_ABOUT_RE
|
|
2981
|
+
* above. Named (not inlined) so both vagueTouchTermOf (term extraction) and
|
|
2982
|
+
* the isConversational-catch-all exemption (below, deduceGoalFromParsed's
|
|
2983
|
+
* neighbourhood) can test the SAME shape. */
|
|
2984
|
+
const EXPLAIN_TOUCH_RE = /^(?:please\s+|kindly\s+)*explain\s+(?:to\s+me\s+)?(?:an?\s+|the\s+)?(.+?)(?:\s+(?:to\s+me|please))?[?.!\s]*$/i;
|
|
2408
2985
|
function vagueTouchTermOf(query) {
|
|
2409
|
-
|
|
2410
|
-
|
|
2411
|
-
|
|
2412
|
-
|
|
2986
|
+
// typo-correct the ANCHOR words only ("waht about calls" -> "what about
|
|
2987
|
+
// calls") — this shape has no ask()-grammar envelope to lean on for typo
|
|
2988
|
+
// tolerance (unlike metaTermOf's "what is a X", which mostly gets it for
|
|
2989
|
+
// free off envelope.parsed once ask() itself has normalized). Then peel the
|
|
2990
|
+
// SAME closed greeting/thanks/modal-wrapper preambles ask()'s own grammar
|
|
2991
|
+
// already peels (0.9.14 Tier-2 playtest §3b spot-check: "cheers, what about
|
|
2992
|
+
// imports then" and "could you kindly tell me about the calls" both used to
|
|
2993
|
+
// fall through to a bogus object search) — applyPreambleFrames alone, NOT
|
|
2994
|
+
// the full normalizeQuery pipeline, which also runs subordination/
|
|
2995
|
+
// conditional rewrites that turn "tell me about X" into "about X" (its own
|
|
2996
|
+
// bridge frame), breaking this very regex.
|
|
2997
|
+
let q = correctMisspellings(String(query).trim());
|
|
2998
|
+
q = q.replace(VAGUE_TOUCH_TEL_RE, "tell");
|
|
2999
|
+
q = applyPreambleFrames(q);
|
|
3000
|
+
const m = q.match(/^(?:kindly\s+)?tell me about\s+(?:an?\s+|the\s+)?(.+?)[?.!\s]*$/i)
|
|
3001
|
+
|| q.match(/^(?:(?:and|so|but|ok|okay|now|then|kindly)\s+)*what about\s+(?:an?\s+|the\s+)?(.+?)(?:\s+then|\s+though)?[?.!\s]*$/i)
|
|
3002
|
+
// "explain X" (0.9.14 Tier-2 playtest, second pass, §3b formal/ESL angle)
|
|
3003
|
+
// — a bare "explain <term>" is at least as natural a vague touch as "tell
|
|
3004
|
+
// me about X", but had no recognized shape at all: normalize.mjs's own
|
|
3005
|
+
// EXPLAIN_WRAPPER_RE only unwraps a WH-QUESTION remainder ("explain
|
|
3006
|
+
// please where is it defined" -> a real structural question), so a bare
|
|
3007
|
+
// noun remainder like "cochange" was never its territory. A leading
|
|
3008
|
+
// "please"/"kindly" also broke the STRUCTURAL pipeline's own
|
|
3009
|
+
// EXPLAIN_WRAPPER_RE (anchored to start with "explain" literally),
|
|
3010
|
+
// sending the whole turn to the wrong lane.
|
|
3011
|
+
|| q.match(EXPLAIN_TOUCH_RE);
|
|
3012
|
+
if (!m) return null;
|
|
3013
|
+
// A trailing meta-noun naming WHAT KIND of thing the touched word already is
|
|
3014
|
+
// (0.9.14 Tier-2 playtest, second pass): "tell me about the cochange
|
|
3015
|
+
// relation" / "what about the calls relationship" / "what about the imports
|
|
3016
|
+
// edges" used to capture the WHOLE tail ("cochange relation") as the term —
|
|
3017
|
+
// RELATION_TERM's closed dict has no multi-word entries, so the relation
|
|
3018
|
+
// force declined and the query fell through to the grammar wall. Stripped
|
|
3019
|
+
// for both callers (conceptTermOf's noun touch and relationTermOf's edge
|
|
3020
|
+
// touch): a noun concept is never phrased with this tail ("tell me about
|
|
3021
|
+
// the Class relation" isn't natural), so it's safe either way.
|
|
3022
|
+
const term = m[1].trim().replace(/\s+(?:relations?|relationships?|edges?)$/i, "").trim();
|
|
3023
|
+
return term || null;
|
|
2413
3024
|
}
|
|
2414
3025
|
|
|
2415
3026
|
function conceptTermOf(query, envelope) {
|
|
@@ -2426,14 +3037,37 @@ function conceptTermOf(query, envelope) {
|
|
|
2426
3037
|
function relationTermOf(query, envelope) {
|
|
2427
3038
|
const base = vagueTouchTermOf(query);
|
|
2428
3039
|
if (base) return base;
|
|
2429
|
-
|
|
3040
|
+
// same typo-correction as vagueTouchTermOf above ("waht calls are there" ->
|
|
3041
|
+
// "what calls are there") — these openers are chat.mjs-only shapes with no
|
|
3042
|
+
// ask()-grammar envelope to inherit normalization from (0.9.14 Tier-2
|
|
3043
|
+
// playtest: "waht calls are there" used to hit the grammar wall outright).
|
|
3044
|
+
const q = correctMisspellings(String(query).trim()).toLowerCase().replace(/[?.!]+$/, "").replace(/\s+/g, " ");
|
|
2430
3045
|
let m;
|
|
2431
|
-
// "what are the imports", "what is the containment", "what are all the calls"
|
|
2432
|
-
|
|
2433
|
-
// "what calls
|
|
2434
|
-
|
|
3046
|
+
// "what are the imports", "what is the containment", "what are all the calls",
|
|
3047
|
+
// and the texting-shorthand "r" for "are" (0.9.14 Tier-2 playtest §3b spot-check:
|
|
3048
|
+
// "what r the calls" — narrowly scoped to this closed shape, same judgment call
|
|
3049
|
+
// as chat.mjs's own SHORTHAND_CONTRACTIONS for the identity lane: "r" only reads
|
|
3050
|
+
// as "are" right after "what" in one of these curated anchor shapes, so a real
|
|
3051
|
+
// one-letter identifier is never at risk).
|
|
3052
|
+
if ((m = q.match(/^what\s+(?:are|is|r)\s+(?:all\s+)?(?:the\s+)?([a-z][a-z-]*?)(?:\s+(?:edges|relationships|relations))?$/))) return m[1];
|
|
3053
|
+
// "what calls are there", "what imports are there", "what calls r there"
|
|
3054
|
+
if ((m = q.match(/^what\s+([a-z][a-z-]*?)\s+(?:are|r)\s+there$/))) return m[1];
|
|
2435
3055
|
// "what is calling", "what is importing" (bare gerund, no object)
|
|
2436
3056
|
if ((m = q.match(/^what\s+(?:is|are)\s+([a-z][a-z-]*ing)$/))) return m[1];
|
|
3057
|
+
// STACCATO RELATION-CHAIN CONTINUATION (0.9.15 Tier-2 playtest, 4th pass): a
|
|
3058
|
+
// rapid-fire short follow-up inside an EXISTING relation-touch chain — "and
|
|
3059
|
+
// calls?", "also tests", "so inherits", "then contains" — has no "about"/
|
|
3060
|
+
// "is"/"are" at all, just a bare connective + the relation word. Without
|
|
3061
|
+
// this, the bare word fell straight through to ask()'s own raw grammar,
|
|
3062
|
+
// which parsed the leading connective ITSELF as the object term (e.g. "and
|
|
3063
|
+
// calls" read as kind=calls object="and", silently resolving "and" via the
|
|
3064
|
+
// standing focus/contextId fallback into an unrelated, honestly-empty-but-
|
|
3065
|
+
// wrong answer) or, worse, matched no shape at all and hit the grammar
|
|
3066
|
+
// wall outright. Scoped to RELATION_TERM's own closed dict downstream (this
|
|
3067
|
+
// function's caller, relationForceAnswer), so an unrelated word or a real
|
|
3068
|
+
// entity name ("and Widget?", "so that") safely falls through unchanged —
|
|
3069
|
+
// only a genuine, already-known relation word is swept up.
|
|
3070
|
+
if ((m = q.match(/^(?:and|also|so|then|now)\s+([a-z][a-z-]*)$/))) return m[1];
|
|
2437
3071
|
// THE SINGULAR META FORM — "what is a test" / "what is an import". The whole meta
|
|
2438
3072
|
// shape used to be excluded here to keep the frozen am-meta-imports ambiguity case
|
|
2439
3073
|
// ("what does imports mean") out; but that case is a DIFFERENT shape (ambiguousParse
|
|
@@ -2451,26 +3085,61 @@ function relationTermOf(query, envelope) {
|
|
|
2451
3085
|
}
|
|
2452
3086
|
|
|
2453
3087
|
/** A closed "describe"-intent wrapper: "can you describe X for me", "could you
|
|
2454
|
-
* tell me about X", "tell me more about X"
|
|
2455
|
-
* live (playtest sprint round 2,
|
|
2456
|
-
* question wrapped in an
|
|
2457
|
-
*
|
|
2458
|
-
*
|
|
2459
|
-
* lead-in-alternation
|
|
2460
|
-
* (normalize.mjs).
|
|
2461
|
-
*
|
|
2462
|
-
*
|
|
2463
|
-
*
|
|
2464
|
-
*
|
|
2465
|
-
*
|
|
2466
|
-
*
|
|
3088
|
+
* tell me about X", "tell me more about X", "what about X" → attempt
|
|
3089
|
+
* tmct_describe(X). Found live (playtest sprint round 2,
|
|
3090
|
+
* SKILL_PLAYTEST_SPRINT.md): a describe-intent question wrapped in an
|
|
3091
|
+
* ordinary polite request ("can you tell me more about Controller") fell all
|
|
3092
|
+
* the way to the generic wall despite naming a real, just-listed entity —
|
|
3093
|
+
* nothing recognized the wrapper at all. Same closed lead-in-alternation
|
|
3094
|
+
* discipline as GREETING_PREAMBLE_RE/THANKS_PREAMBLE_RE (normalize.mjs).
|
|
3095
|
+
* Deliberately used only as a LAST-RESORT lane (see its call site below) —
|
|
3096
|
+
* "tell me about X" is ALSO the relation/concept force's own trigger phrase
|
|
3097
|
+
* for enumerable concepts ("tell me about inheritance"), and "what about X"
|
|
3098
|
+
* is ALSO discourseRewrite's own trigger for continuing an ask()-shaped prior
|
|
3099
|
+
* turn — this must never run before those have had their chance. Trails an
|
|
3100
|
+
* optional "please" as well as "for me" (playtest sprint round 3): this lane
|
|
3101
|
+
* reads the RAW turn text, not normalize.mjs's FILLER_WORDS-stripped one, so
|
|
3102
|
+
* "could you tell me more about Router please" needs its own trailing-
|
|
3103
|
+
* politeness strip.
|
|
3104
|
+
* "what about X" (0.9.13 Tier-1 playtest): reaches this lane specifically
|
|
3105
|
+
* when the PRIOR turn was itself a describe-shaped question ("describe Task"
|
|
3106
|
+
* isn't an ask()-grammar verb, so discourseRewrite's "describe <X>" rewrite
|
|
3107
|
+
* can never parse and always misses) — a drill-down chain that opens with
|
|
3108
|
+
* "describe X" (the README's own example) used to dead-end on the very next
|
|
3109
|
+
* "what about it"/"what about Y" turn. */
|
|
2467
3110
|
const DESCRIBE_WRAPPER_RE =
|
|
2468
|
-
/^(?:(?:can|could|would)\s+you\s+(?:please\s+)?|please\s+)?(?:tell\s+me\s+(?:more\s+)?about|describe)\s+(.+?)(?:\s+for\s+me)?(?:\s+please)?\s*\??$/i;
|
|
2469
|
-
|
|
2470
|
-
|
|
2471
|
-
|
|
2472
|
-
|
|
3111
|
+
/^(?:(?:can|could|would)\s+you\s+(?:please\s+)?|please\s+)?(?:tell\s+me\s+(?:more\s+)?about|describe|what(?:'s|\s+is)?\s+about)\s+(.+?)(?:\s+for\s+me)?(?:\s+please)?\s*\??$/i;
|
|
3112
|
+
|
|
3113
|
+
/** Bare focus pronouns this lane resolves against the STANDING focus (0.9.13
|
|
3114
|
+
* Tier-1 playtest) — "describe that" / "tell me about it" after a prior turn
|
|
3115
|
+
* set the focus. Never a guess: no standing focus → the lane declines (null),
|
|
3116
|
+
* same as any unresolvable term. */
|
|
3117
|
+
const DESCRIBE_PRONOUN_RE = /^(?:it|that|this|those|them)$/i;
|
|
3118
|
+
|
|
3119
|
+
/** STACCATO PRONOUN CONTINUATION (0.9.15 Tier-2 playtest, 4th pass): a rapid-
|
|
3120
|
+
* fire short follow-up naming no verb at all — "and that?", "also this",
|
|
3121
|
+
* "so it" — the bare-connective sibling of DESCRIBE_WRAPPER_RE's "what about
|
|
3122
|
+
* it"/"describe that". Without this, "what calls X" -> "and that?" fell to
|
|
3123
|
+
* the generic orientation card (isConversational's ≤3-word catch-all caught
|
|
3124
|
+
* it, and DESCRIBE_WRAPPER_RE requires an actual "about"/"describe" anchor
|
|
3125
|
+
* word this shape never has) even though the immediately-prior turn had just
|
|
3126
|
+
* set a real focus a sibling phrasing ("what about it") already resolves
|
|
3127
|
+
* against cleanly. An optional trailing "one"/"ones" (Tier-2 playtest, 5th
|
|
3128
|
+
* pass — "also that one?", "and those ones") is at least as natural as the
|
|
3129
|
+
* bare pronoun and carries no extra meaning beyond it: the capture group
|
|
3130
|
+
* stays the pronoun alone, so DESCRIBE_PRONOUN_RE's downstream test is
|
|
3131
|
+
* unaffected either way. */
|
|
3132
|
+
const STACCATO_PRONOUN_RE = /^(?:and|also|so|then|now)\s+(it|that|this|those|them)(?:\s+ones?)?\s*\??$/i;
|
|
3133
|
+
|
|
3134
|
+
async function describeWrapperAnswer(query, { config, source, focus }) {
|
|
3135
|
+
const q = String(query || "").trim();
|
|
3136
|
+
const m = DESCRIBE_WRAPPER_RE.exec(q) || STACCATO_PRONOUN_RE.exec(q);
|
|
3137
|
+
let term = m?.[1]?.trim();
|
|
2473
3138
|
if (!term) return null;
|
|
3139
|
+
if (DESCRIBE_PRONOUN_RE.test(term)) {
|
|
3140
|
+
if (!focus?.label) return null; // no standing focus to resolve against — honest decline
|
|
3141
|
+
term = focus.label;
|
|
3142
|
+
}
|
|
2474
3143
|
try {
|
|
2475
3144
|
const text = await dispatchTool("tmct_describe", { symbol: term }, { config, source });
|
|
2476
3145
|
return text ? { text } : null;
|
|
@@ -2583,7 +3252,29 @@ async function runAsk(query, { config, source, graph, focus, last, templates, me
|
|
|
2583
3252
|
// The query the ENGINE parses: a "what about X" continuation is rewritten to the
|
|
2584
3253
|
// prior shape with X swapped in; everything else parses verbatim. The record and
|
|
2585
3254
|
// transcript keep the user's ACTUAL words (`query`), only the parse target changes.
|
|
2586
|
-
|
|
3255
|
+
let askQuery = discourseRewrite(query, last) ?? query;
|
|
3256
|
+
// IMPLICIT ANAPHORIC COUNT (Tier-2 playtest, 5th pass): "how many are tested" /
|
|
3257
|
+
// "and how many are tested" drops the "of those/them" a fuller phrasing carries
|
|
3258
|
+
// — ask()'s own anaphora node (parseAnaphora) already understands "how many of
|
|
3259
|
+
// those are tested" perfectly, it simply never SEES this elliptical spelling
|
|
3260
|
+
// (ANAPHORA_TRIGGERS requires an explicit pronoun). Insert the elided "of
|
|
3261
|
+
// those" here, the same way discourseRewrite rewrites "what about X" —
|
|
3262
|
+
// UNCONDITIONALLY (not gated on `prev.length`): a genuinely bare "how many
|
|
3263
|
+
// are tested" with no antecedent at all still reaches the anaphora node this
|
|
3264
|
+
// way, which itself honestly degrades to "needs a previous answer to refer
|
|
3265
|
+
// to" (evalAnaphora's own no-prev branch) — a strictly better outcome than
|
|
3266
|
+
// leaving the raw ellipsis unrewritten, which used to fall through to the
|
|
3267
|
+
// ordinary clause grammar and misparse "and" as the object ('no module
|
|
3268
|
+
// matching "and many" found').
|
|
3269
|
+
if (IMPLICIT_ANAPHORA_COUNT_RE.test(String(askQuery).trim())) {
|
|
3270
|
+
// Strip the leading connective too ("and how many are tested" -> "how many
|
|
3271
|
+
// of those are tested") — left in place, it breaks the anaphora node's own
|
|
3272
|
+
// AGGREGATE_TRIGGERS match on "how many" (anchored at the string start),
|
|
3273
|
+
// silently degrading the count into a bare list of the filtered set.
|
|
3274
|
+
askQuery = String(askQuery).trim()
|
|
3275
|
+
.replace(/^(?:and|so|then|also)\s+/i, "")
|
|
3276
|
+
.replace(/how many\s+/i, "how many of those ");
|
|
3277
|
+
}
|
|
2587
3278
|
// W2: the explicit recall forms are answered from memory's folded blocks, never
|
|
2588
3279
|
// the graph. Gated on memoryDir — a bare runTurn (no session shell) stays pure.
|
|
2589
3280
|
if (memoryDir && RECALL_ASK_RE.test(String(query).trim())) {
|
|
@@ -2677,10 +3368,16 @@ async function runAsk(query, { config, source, graph, focus, last, templates, me
|
|
|
2677
3368
|
// completely different lane — an intent lane's own goal note, when it pushes one,
|
|
2678
3369
|
// stays the more specific of the two since bucketTrace keeps every "goal:" line and
|
|
2679
3370
|
// renderNarration shows them all, most-specific-last-written).
|
|
2680
|
-
|
|
2681
|
-
|
|
2682
|
-
|
|
2683
|
-
|
|
3371
|
+
//
|
|
3372
|
+
// FEATURE B: `deduced` (declared here, not block-scoped) also rides the
|
|
3373
|
+
// returned result as `goal` (see the return statement below) — the seam
|
|
3374
|
+
// withLast's withGoalLine reads to prepend the always-on, short "Goal
|
|
3375
|
+
// (inferred): …" line, independent of --narrate entirely. Deliberately the
|
|
3376
|
+
// SAME value the debug trace's own "goal:" line uses (one deduction, two
|
|
3377
|
+
// presentations) — null here (no parse stood at all) means withGoalLine
|
|
3378
|
+
// shows nothing, never a "Goal (inferred): unclear" line.
|
|
3379
|
+
const deduced = deduceGoalFromParsed(envelope?.parsed);
|
|
3380
|
+
note(trace, `goal: ${deduced ?? "unclear — the phrasing didn't resolve to a known query shape"}`);
|
|
2684
3381
|
// MISS handling. The intent lanes + short-miss are RECOGNIZER-gated on the query
|
|
2685
3382
|
// text AND only consulted on a would-miss, so a real graph query — a hit, an honest
|
|
2686
3383
|
// empty with a receipt, a fuzzy repair — is never hijacked. Order: (1) META/SELF
|
|
@@ -2698,7 +3395,53 @@ async function runAsk(query, { config, source, graph, focus, last, templates, me
|
|
|
2698
3395
|
note(trace, `lane: (1) META/SELF — bare self/session question recognized, answered via="${meta.via}"`);
|
|
2699
3396
|
}
|
|
2700
3397
|
}
|
|
2701
|
-
|
|
3398
|
+
// "what about X" with a genuine PRIOR turn to continue (0.9.13 Tier-1 playtest)
|
|
3399
|
+
// is exempt from the conversational catch-all even when short/non-codeish
|
|
3400
|
+
// ("what about that", "what about Task" — no dotted/camel token, ≤3 words):
|
|
3401
|
+
// isConversational() can't see that ask() ALREADY tried discourseRewrite above
|
|
3402
|
+
// and that the describe-wrapper rescue (4d) hasn't had its turn yet — without
|
|
3403
|
+
// this exemption, EVERY "what about X" continuation whose prior turn was itself
|
|
3404
|
+
// a describe-shaped question (discourseRewrite can't rewrite "describe X", so it
|
|
3405
|
+
// always misses) or whose swapped-in subject is a bare Capitalized/pronoun term
|
|
3406
|
+
// fell straight to the generic orientation card instead of reaching (4d).
|
|
3407
|
+
// Same exemption for the bare-connective sibling shape ("and Widget?", "also
|
|
3408
|
+
// app/lib/b.mjs" — no "about" at all, STACCATO_SWAP_RE above), gated the
|
|
3409
|
+
// SAME way discourseRewrite gates it: the swapped-in word must itself be
|
|
3410
|
+
// unambiguously code-ish, so ordinary discourse ("and then?", "so what")
|
|
3411
|
+
// never trips this exemption.
|
|
3412
|
+
const staccatoSwapMatch = String(query).match(STACCATO_SWAP_RE);
|
|
3413
|
+
const isStaccatoSwap = !!(last?.query && staccatoSwapMatch && NAME_TOKEN_RE.test(staccatoSwapMatch[1]?.trim() || ""));
|
|
3414
|
+
const isWhatAboutContinuation = !!(last?.query && WHAT_ABOUT_RE.test(String(query))) || isStaccatoSwap;
|
|
3415
|
+
// Same exemption for the sibling shape "describe it"/"tell me about that"
|
|
3416
|
+
// (0.9.13 Tier-1 playtest): a bare-pronoun describe/tell-me-about is exactly
|
|
3417
|
+
// as short and non-codeish as "what about it", and needs the SAME deferral to
|
|
3418
|
+
// reach describeWrapperAnswer's now-focus-aware pronoun resolution (4d) —
|
|
3419
|
+
// WITHOUT this, "describe Widget" -> "describe that" (a natural drill-down
|
|
3420
|
+
// re-ask) fell to the orientation card even though the standing focus made it
|
|
3421
|
+
// perfectly answerable. Gated on an actual standing focus, same honest-decline
|
|
3422
|
+
// discipline as describeWrapperAnswer itself.
|
|
3423
|
+
const describeWrapperMatch = DESCRIBE_WRAPPER_RE.exec(String(query).trim()) || STACCATO_PRONOUN_RE.exec(String(query).trim());
|
|
3424
|
+
const isDescribePronounContinuation = !!(focus?.label && describeWrapperMatch && DESCRIBE_PRONOUN_RE.test(describeWrapperMatch[1]?.trim() || ""));
|
|
3425
|
+
// A bare/wrapped "explain X" (0.9.14 Tier-2 playtest, second pass) needs the
|
|
3426
|
+
// SAME deferral, and for a stronger reason than the two above: "explain"
|
|
3427
|
+
// isn't a VERB_TO_KIND word at all, so ask() never even ATTEMPTS a parse
|
|
3428
|
+
// (envelope.parsed is null unconditionally for this shape, not merely on a
|
|
3429
|
+
// miss) — a short "explain cochange" (2 words) or politeness-wrapped
|
|
3430
|
+
// "please explain cochange" (3 words) always trips isConversational's ≤3-
|
|
3431
|
+
// word heuristic and never once reaches the relation/concept force below,
|
|
3432
|
+
// which is squarely built to answer exactly this shape. Unlike the two
|
|
3433
|
+
// exemptions above, this one needs no prior-turn/focus context — "explain
|
|
3434
|
+
// X" is a complete, self-contained ask on its own.
|
|
3435
|
+
const isExplainTouch = EXPLAIN_TOUCH_RE.test(String(query).trim());
|
|
3436
|
+
// Staccato negation ("not that one", "not Widget then" — Tier-2, 5th pass)
|
|
3437
|
+
// needs the SAME deferral: "not those" (2 words) / "not that one" (3 words)
|
|
3438
|
+
// both trip isConversational's ≤3-word catch-all before nudgeAnswer's own
|
|
3439
|
+
// STACCATO_NEGATION_RE branch (4c, below) ever gets a turn. Gated on the
|
|
3440
|
+
// shape alone (not a focus/prev precondition) — nudgeAnswer's negation
|
|
3441
|
+
// branch ALWAYS returns a tailored nudge for this shape, never null, so
|
|
3442
|
+
// deferring here never strands the turn with nothing having claimed it.
|
|
3443
|
+
const isStaccatoNegation = STACCATO_NEGATION_RE.test(String(query).trim());
|
|
3444
|
+
if (!handled && miss && !envelope?.parsed && isConversational(query) && !isWhatAboutContinuation && !isDescribePronounContinuation && !isExplainTouch && !isStaccatoNegation) {
|
|
2702
3445
|
// A conversational miss (a greeting, "what can you do", a very short non-code
|
|
2703
3446
|
// line) gets the friendly orientation (module-aware: empty → --repo/tmct init).
|
|
2704
3447
|
// Bug B1 (0.8.2 follow-up): this branch carries via:"template" and never
|
|
@@ -2898,7 +3641,7 @@ async function runAsk(query, { config, source, graph, focus, last, templates, me
|
|
|
2898
3641
|
// for what would otherwise become the generic wall, never a competing route:
|
|
2899
3642
|
// it only claims the turn if /describe actually resolves the captured term.
|
|
2900
3643
|
if (miss && recordMiss && via === "composed") {
|
|
2901
|
-
const described = await describeWrapperAnswer(query, { config, source });
|
|
3644
|
+
const described = await describeWrapperAnswer(query, { config, source, focus: newFocus });
|
|
2902
3645
|
if (described) {
|
|
2903
3646
|
answer = described.text; via = "describe"; recordMiss = false;
|
|
2904
3647
|
note(trace, "lane: (4d) DESCRIBE-WRAPPER RESCUE — a polite wrapper around \"describe/tell me about <symbol>\" resolved via /describe, tried last after every other lane declined");
|
|
@@ -2975,7 +3718,32 @@ async function runAsk(query, { config, source, graph, focus, last, templates, me
|
|
|
2975
3718
|
: (envelope
|
|
2976
3719
|
? { traversal: envelope.traversal || null, matches: envelope.matches || [], ...(pending ? { pending } : {}) }
|
|
2977
3720
|
: (pending ? { traversal: null, matches: [], pending } : null));
|
|
2978
|
-
|
|
3721
|
+
// MULTI-HOP STACCATO CHAIN CONTINUATION (Tier-2 playtest, 5th pass): when
|
|
3722
|
+
// discourseRewrite actually substituted a new subject into the PRIOR
|
|
3723
|
+
// query's shape ("and Widget?" -> "what calls Widget") and the rewritten
|
|
3724
|
+
// query STRUCTURALLY PARSED (envelope.parsed stood — a real AST, whether it
|
|
3725
|
+
// went on to a hit or an honest empty; "miss" in this engine's own
|
|
3726
|
+
// convention covers BOTH a genuine grammar failure AND a structurally valid
|
|
3727
|
+
// empty result, so `recordMiss` alone can't distinguish them here), thread
|
|
3728
|
+
// the RECONSTRUCTED positive query forward as the effective `last.query`
|
|
3729
|
+
// the NEXT turn's own discourseRewrite reads — not the raw staccato text
|
|
3730
|
+
// itself. Without this, a 3rd staccato swap in a row ("what calls X" ->
|
|
3731
|
+
// "and Widget?" -> "and Button?") tried to rewrite off "and Widget?" (the
|
|
3732
|
+
// 2nd turn's own verbatim staccato input, which has no clause shape of its
|
|
3733
|
+
// own), corrupting the 3rd swap into a nonsense re-ask ("and Button?" with
|
|
3734
|
+
// "Widget" replaced by "Button" — never a real query) instead of correctly
|
|
3735
|
+
// continuing from "what calls Widget". The verbatim text stays on
|
|
3736
|
+
// `record.query`/the transcript untouched; only the swap-chain
|
|
3737
|
+
// CONTINUATION base changes.
|
|
3738
|
+
const effectiveQuery = (askQuery !== query && envelope?.parsed) ? askQuery : null;
|
|
3739
|
+
// `goal` (Feature B): the SAME deduced string the debug trace's own "goal:"
|
|
3740
|
+
// line carries (deduced above, right after envelope resolution) — null when
|
|
3741
|
+
// deduceGoalFromParsed found no genuine query shape to bucket on, which is
|
|
3742
|
+
// exactly withGoalLine's own "say nothing" signal. Only runAsk ever sets
|
|
3743
|
+
// this field (plainTurn/runCommand results never carry it), so the always-on
|
|
3744
|
+
// goal line is scoped to real ask-engine turns by construction — a count, a
|
|
3745
|
+
// slash-command or a teach confirmation never grows one.
|
|
3746
|
+
return { answer, logLines, record, focus: newFocus, detail, effectiveQuery, goal: deduced };
|
|
2979
3747
|
}
|
|
2980
3748
|
|
|
2981
3749
|
/** A non-ask, non-dispatch chat turn (count answer, /stats) — the same
|
|
@@ -3119,13 +3887,36 @@ async function assertTurn(line, { memoryDir, sessionId, focus, lexicon = null })
|
|
|
3119
3887
|
const parse = parseAce(line, lex);
|
|
3120
3888
|
if (!parse || !parse.triples?.length || parse.residue?.length) return null;
|
|
3121
3889
|
const { assertSentence } = await import("./grammar/assert.mjs");
|
|
3122
|
-
const { normFactTerm } = await import("./memory/core.mjs");
|
|
3890
|
+
const { normFactTerm, appendFact } = await import("./memory/core.mjs");
|
|
3123
3891
|
const ts = new Date().toISOString();
|
|
3124
3892
|
const res = await assertSentence(memoryDir, line, {
|
|
3125
3893
|
lexicon: lex,
|
|
3126
3894
|
provenance: { source: "chat", sessionId, ts },
|
|
3127
3895
|
});
|
|
3128
3896
|
if (!res || !res.ids?.length) return null;
|
|
3897
|
+
// Feature A point 3: a plain universal "every X is a Y" ALSO records the
|
|
3898
|
+
// "every" quantifier on the SAME fact — purely additive (appendFact
|
|
3899
|
+
// upserts by (s,p,o) id, never a duplicate, never changes the confirmation
|
|
3900
|
+
// text below), for the new "how many Xs are Ys" recall lane. Gated on the
|
|
3901
|
+
// literal typed determiner (not on `parse.pattern`, which is "subClassOf"
|
|
3902
|
+
// for the bare-copula variant too) — only "every" reads as a class-level
|
|
3903
|
+
// generalization; a bare/indefinite "X is a Y" is one specific claim and
|
|
3904
|
+
// gets no quantifier. `provenance` is deliberately omitted (appendFact
|
|
3905
|
+
// treats "" as a no-op on the union) so this never grows a redundant tag
|
|
3906
|
+
// alongside the fact's real ace:chat provenance. Best-effort: the base
|
|
3907
|
+
// fact is already durably stored either way, so a failure here (a
|
|
3908
|
+
// relation/cardinality/etc. axiom that happens to start with "every" and
|
|
3909
|
+
// carries no rdfs:subClassOf triple, or any write error) is swallowed.
|
|
3910
|
+
if (/^every\s+/i.test(String(line).trim())) {
|
|
3911
|
+
const triple = res.triples.find((t) => t.predicate === "rdfs:subClassOf");
|
|
3912
|
+
if (triple) {
|
|
3913
|
+
try {
|
|
3914
|
+
await appendFact(memoryDir, {
|
|
3915
|
+
subject: triple.subject, predicate: "rdfs:subClassOf", object: triple.object, quantifier: "every",
|
|
3916
|
+
});
|
|
3917
|
+
} catch { /* best-effort — the base fact is already stored either way */ }
|
|
3918
|
+
}
|
|
3919
|
+
}
|
|
3129
3920
|
const shown = res.triples
|
|
3130
3921
|
.map((t) => `${normFactTerm(t.subject)} ${t.predicate} ${normFactTerm(t.object)}`)
|
|
3131
3922
|
.join("; ");
|
|
@@ -3202,8 +3993,19 @@ export async function runTurn(input, { config, source = defaultSource, graph = n
|
|
|
3202
3993
|
// the PRE-narration finished result — see withNarration's docblock for why.
|
|
3203
3994
|
const withLast = (result, fallbackGoal = "unclear — no goal signal for this turn type") => {
|
|
3204
3995
|
const finished = finish(result, { graph });
|
|
3205
|
-
|
|
3206
|
-
|
|
3996
|
+
// runAsk's own effectiveQuery (set only when discourseRewrite substituted a
|
|
3997
|
+
// new subject AND the rewrite produced a genuine non-miss answer) takes
|
|
3998
|
+
// over as the continuation base for the NEXT turn's own discourseRewrite —
|
|
3999
|
+
// see runAsk's docblock above its return statement. Every other turn type
|
|
4000
|
+
// (commands, plain counts, misses) carries no such field, so `line` — the
|
|
4001
|
+
// existing, unchanged behavior — stands.
|
|
4002
|
+
const nextLast = { query: finished.effectiveQuery ?? line, answer: finished.answer, detail: finished.detail ?? null };
|
|
4003
|
+
// FEATURE B: the always-on short "Goal (inferred): …" line — computed from
|
|
4004
|
+
// the SAME PRE-narration `finished` result `nextLast` was just captured
|
|
4005
|
+
// from, so (like narrate) it never contaminates what why/say-more or
|
|
4006
|
+
// repeat-detection compare against. Composes with narrate (below): a
|
|
4007
|
+
// narrated turn gets the short line up top AND the full trace block after.
|
|
4008
|
+
return { ...withNarration(withGoalLine(finished), trace, fallbackGoal), last: nextLast };
|
|
3207
4009
|
};
|
|
3208
4010
|
|
|
3209
4011
|
// Slash-optional system commands: a bare leading command word ("stats",
|
|
@@ -3255,6 +4057,19 @@ export async function runTurn(input, { config, source = defaultSource, graph = n
|
|
|
3255
4057
|
return withLast(plainTurn(line, memCount, { via: "count", focus }), "get a count of a memory-store kind");
|
|
3256
4058
|
}
|
|
3257
4059
|
}
|
|
4060
|
+
// Feature A point 4: "how many Xs are Ys" — a taught-quantifier RECALL, checked
|
|
4061
|
+
// explicitly ahead of answerCount (see answerQuantifierRecall's own "CRITICAL
|
|
4062
|
+
// ORDERING NOTE" — mirrors answerMemoryCount's precedent just above). Its own
|
|
4063
|
+
// authority gate declines (returns null) for anything answerCount should own,
|
|
4064
|
+
// so ordinary structural counts fall through completely unaffected.
|
|
4065
|
+
if (memoryDir) {
|
|
4066
|
+
const quantifierRecall = await answerQuantifierRecall(memoryDir, line);
|
|
4067
|
+
if (quantifierRecall != null) {
|
|
4068
|
+
note(trace, 'goal: recall a taught quantifier for a class-membership pair ("how many Xs are Ys")');
|
|
4069
|
+
note(trace, "lane: answerQuantifierRecall — matched HOW_MANY_ARE_RE with a subject tmct has facts about; literal recall, never real counting");
|
|
4070
|
+
return withLast(plainTurn(line, quantifierRecall, { via: "fact", focus }), "recall a taught quantifier");
|
|
4071
|
+
}
|
|
4072
|
+
}
|
|
3258
4073
|
// Aggregate/count questions are answered mechanically off the loaded graph header,
|
|
3259
4074
|
// BEFORE falling through to the ask engine (focus unchanged — a count names no entity).
|
|
3260
4075
|
const count = answerCount(graph, line);
|
|
@@ -3364,11 +4179,19 @@ async function hasSeededVocabulary(repo) {
|
|
|
3364
4179
|
* seed.enabled=false, or corpus load failure), offering it would be a lie worse
|
|
3365
4180
|
* than no example — swap to an unconditionally-true pointer instead (the teach
|
|
3366
4181
|
* lane and `tmct init` both work with zero preconditions). Computed ONCE per
|
|
3367
|
-
* session (createSession), not per turn.
|
|
4182
|
+
* session (createSession), not per turn.
|
|
4183
|
+
* The unseeded branch's teach clause is a CONCRETE pair too, for the same
|
|
4184
|
+
* reason `cache` is concrete in the seeded branch: playtest found that an
|
|
4185
|
+
* abstract "every X is a Y" invites a curious user to fill X/Y with an
|
|
4186
|
+
* intuitive-but-unknown word ("every cache is a thing" — "thing" isn't in
|
|
4187
|
+
* the closed ACE lexicon) and hit the teach-miss dead-end right after being
|
|
4188
|
+
* offered the pattern. "every bug is an issue" is confirmed to parse and
|
|
4189
|
+
* store (both `bug` and `issue` are declared lexicon nouns — see
|
|
4190
|
+
* test/chatflow-tier0.test.mjs), so the offer resolves if copied verbatim. */
|
|
3368
4191
|
function vocabExampleHint(seeded) {
|
|
3369
4192
|
return seeded
|
|
3370
4193
|
? 'Try "what is a cache" for general vocabulary.'
|
|
3371
|
-
: 'Run `tmct init` to seed a starter vocabulary, or teach me directly
|
|
4194
|
+
: 'Run `tmct init` to seed a starter vocabulary, or teach me directly, e.g. "every bug is an issue".';
|
|
3372
4195
|
}
|
|
3373
4196
|
|
|
3374
4197
|
/** Trim a focus label for the prompt so a long module path can't run the line off. */
|