@polycode-projects/the-mechanical-code-talker 1.4.1 → 1.5.2

package/src/interpret/strategies/constructions.mjs

@@ -0,0 +1,207 @@
+// interpret/strategies/constructions.mjs — strategy N+1: construction-grammar
+// template banks (PLAN_ADVANCED_GRAMMAR.md track (d)). Per-construction closed
+// template families loaded as DATA from data/templates/constructions/*.toml
+// (pattern -> AST skeleton, slot types validated against the closed RELATIONS/
+// ENTITY_TO_TYPE vocabulary ask-vocab.mjs already owns), registered here as its
+// OWN additive class ("construction") — the same "own-class strategy" pattern
+// interpret/strategies/ace.mjs and noise-strip.mjs already use, so a construction
+// match outranks a same-text keyword-spot GUESS outright (interpret/merge.mjs
+// picks the highest-confidence CLASS; within-class disagreement is the honest
+// {ambiguousParse} tie, which this strategy deliberately avoids triggering
+// against keyword-spot by living in its own class) rather than colliding with it.
+//
+// The point (mirrors grammar.mjs's own file-header precedent, "same shape, new
+// grammatical coverage, not a new mechanism"): grammar GROWTH as committed data,
+// not more normalize.mjs/grammar.mjs code — data/templates/grammar-rules.toml
+// and data/templates/responses.jsonl already work this way. Continues
+// grammar.mjs's T1-T10 numbering (T11+, see the TOML file's own [[construction]]
+// `id` fields) without renumbering anything grammar.mjs already owns.
+//
+// Loader discipline (mirrors src/finish.mjs's loadGrammarRules/grammarRules
+// pattern exactly): synchronous (the pipeline is sync-capable), cached once per
+// process, and DEFENSIVE — a missing directory, unparseable TOML, or an entry
+// that fails validation (an unrecognized `kind`/`entityType`, a malformed
+// pattern) is silently DROPPED, never thrown and never guessed into the nearest
+// match. "One broken strategy/entry never takes the pipeline down"
+// (interpret/pipeline.mjs's own file-header discipline) extends here to one
+// broken DATA ROW never taking the strategy down.
+
+import { readFileSync, readdirSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { join, dirname } from "node:path";
+import { parse as parseToml } from "smol-toml";
+
+import { RELATIONS, ENTITY_TO_TYPE } from "../../ask-vocab.mjs";
+import { escapeRegex } from "../normalize.mjs";
+
+const STRATEGY_DIR = dirname(fileURLToPath(import.meta.url));
+/** The construction-bank directory (data, not code) — every *.toml file inside
+ *  is loaded, in filename order, so a future bank is a new committed file, not
+ *  an edit to this loader. */
+export const CONSTRUCTIONS_DIR = join(STRATEGY_DIR, "..", "..", "..", "data", "templates", "constructions");
+
+const VALID_KINDS = new Set(Object.keys(RELATIONS));
+const VALID_ENTITY_TYPES = new Set(Object.values(ENTITY_TO_TYPE));
+const VALID_SHAPES = new Set(["ask", "reverse", "forward", "where", "when", "meta", "mentions"]);
+
+/** Read every *.toml file in `dir` (sorted, deterministic) and return the raw
+ *  parsed tables concatenated: {relations:[...], constructions:[...]}. A
+ *  missing directory or an unparseable file is DEFENSIVE (per-file: a broken
+ *  file is skipped, not fatal to the others) — callers get whatever validly
+ *  parsed, never a thrown error from a data-authoring mistake. */
+export function readConstructionFiles(dir = CONSTRUCTIONS_DIR) {
+  let files;
+  try {
+    files = readdirSync(dir).filter((f) => f.endsWith(".toml")).sort();
+  } catch {
+    return { relations: [], constructions: [] };
+  }
+  const relations = [];
+  const constructions = [];
+  for (const file of files) {
+    let parsed;
+    try {
+      parsed = parseToml(readFileSync(join(dir, file), "utf8"));
+    } catch {
+      continue; // one malformed file never takes the others down
+    }
+    if (Array.isArray(parsed.relation)) relations.push(...parsed.relation);
+    if (Array.isArray(parsed.construction)) constructions.push(...parsed.construction);
+  }
+  return { relations, constructions };
+}
+
+/** Validate + index the raw [[relation]] rows into noun -> {kind, entityType}.
+ *  Closed-vocabulary validation (the whole point of track (d)'s "slot types
+ *  validated against ENTITY_TO_TYPE/VERB_TO_KIND" deliverable): `kind` MUST be
+ *  one of RELATIONS' own keys and `entityType` (when present) MUST be one of
+ *  ENTITY_TO_TYPE's canonical class names — an entry failing either check is
+ *  dropped, never coerced to the nearest-looking valid value. First occurrence
+ *  of a noun wins (closed-table "first match" discipline, same as every other
+ *  table in this codebase); a later duplicate is silently ignored. */
+export function buildAgentNounTable(relations) {
+  const table = {};
+  for (const r of relations || []) {
+    if (!r || typeof r.noun !== "string" || !r.noun.trim()) continue;
+    if (typeof r.kind !== "string" || !VALID_KINDS.has(r.kind)) continue;
+    if (r.entityType !== undefined && (typeof r.entityType !== "string" || !VALID_ENTITY_TYPES.has(r.entityType))) continue;
+    const noun = r.noun.trim().toLowerCase();
+    if (table[noun]) continue;
+    table[noun] = { kind: r.kind, entityType: r.entityType || null };
+  }
+  return table;
+}
+
+/** Compile one pattern string ("<AGENT> of <TERM>") against the closed agent-
+ *  noun alternation into {re, agentIndex, termIndex}, or null when the pattern
+ *  doesn't carry exactly one <AGENT> and one <TERM> token (a malformed pattern
+ *  — dropped, not guessed at). Literal text is escaped and whitespace-
+ *  normalized (\s+), matching every other anchored-template regex in this
+ *  codebase (grammar.mjs's own TEMPLATES). Case-insensitive; tolerates one
+ *  optional trailing "?", same as grammar.mjs's own templates. */
+function compilePattern(pattern, agentNouns) {
+  if (typeof pattern !== "string" || !pattern.trim() || !agentNouns.length) return null;
+  const agentAlt = agentNouns.slice().sort((a, b) => b.length - a.length).map(escapeRegex).join("|");
+  const parts = pattern.split(/(<AGENT>|<TERM>)/).filter((p) => p !== "");
+  let source = "";
+  const slots = [];
+  for (const part of parts) {
+    if (part === "<AGENT>") {
+      slots.push("agent");
+      source += `(${agentAlt})`;
+    } else if (part === "<TERM>") {
+      slots.push("term");
+      source += `(.+?)`;
+    } else {
+      source += part.split(/(\s+)/).map((seg) => (/^\s+$/.test(seg) ? "\\s+" : escapeRegex(seg))).join("");
+    }
+  }
+  const agentCount = slots.filter((s) => s === "agent").length;
+  const termCount = slots.filter((s) => s === "term").length;
+  if (agentCount !== 1 || termCount !== 1) return null;
+  return {
+    re: new RegExp(`^${source}\\??$`, "i"),
+    agentIndex: slots.indexOf("agent") + 1,
+    termIndex: slots.indexOf("term") + 1,
+  };
+}
+
+/** Validate + compile the raw [[construction]] rows into runnable templates:
+ *  {id, name, shape, re, agentIndex, termIndex}. Requires a valid `id`
+ *  (non-empty string, first-occurrence-wins on duplicates), a `shape` from the
+ *  anchored-template shape vocabulary (VALID_SHAPES), and a pattern that
+ *  compiles cleanly against the agent-noun table — anything else is dropped. */
+export function buildConstructionTemplates(constructions, agentNounTable) {
+  const agentNouns = Object.keys(agentNounTable);
+  const seen = new Set();
+  const out = [];
+  for (const c of constructions || []) {
+    if (!c || typeof c.id !== "string" || !c.id.trim() || seen.has(c.id)) continue;
+    if (typeof c.shape !== "string" || !VALID_SHAPES.has(c.shape)) continue;
+    const compiled = compilePattern(c.pattern, agentNouns);
+    if (!compiled) continue;
+    seen.add(c.id);
+    out.push({ id: c.id, name: c.name || c.id, shape: c.shape, ...compiled });
+  }
+  return out;
+}
+
+let bankCache = null;
+
+/** The cached, compiled construction bank (relations table + runnable
+ *  templates), loaded once per process. Defensive: any failure anywhere in the
+ *  load/validate/compile chain degrades to an empty bank (the strategy simply
+ *  never fires) rather than crashing the pipeline that imports this module. */
+export function constructionBank(dir = CONSTRUCTIONS_DIR) {
+  if (bankCache !== null && dir === CONSTRUCTIONS_DIR) return bankCache;
+  let bank;
+  try {
+    const { relations, constructions } = readConstructionFiles(dir);
+    const agentNounTable = buildAgentNounTable(relations);
+    const templates = buildConstructionTemplates(constructions, agentNounTable);
+    bank = { agentNounTable, templates };
+  } catch {
+    bank = { agentNounTable: {}, templates: [] };
+  }
+  if (dir === CONSTRUCTIONS_DIR) bankCache = bank;
+  return bank;
+}
+
+/** Strategy 1: scan the compiled construction templates, first match wins
+ *  (mirrors grammar.mjs's parseAnchored exactly). A structural regex match
+ *  whose agent noun somehow isn't in the table (shouldn't happen — the
+ *  alternation is built FROM the table) falls through defensively rather than
+ *  building a half-formed parse. Pure. */
+export function parseConstruction(text, bank = constructionBank()) {
+  for (const t of bank.templates) {
+    const m = text.match(t.re);
+    if (!m) continue;
+    const agentText = m[t.agentIndex].toLowerCase();
+    const relation = bank.agentNounTable[agentText];
+    if (!relation) continue;
+    const object = m[t.termIndex].trim();
+    if (!object) continue;
+    return {
+      shape: t.shape, entityType: relation.entityType || null,
+      modifier: "direct", kind: relation.kind, object,
+    };
+  }
+  return null;
+}
+
+/** Pipeline registration (interpret/pipeline.mjs): construction-grammar
+ *  templates as their OWN class ("construction"), confidence 0.9 — the same
+ *  evidentiary weight as grammar.mjs's anchored T1-T10 (an anchored, closed
+ *  pattern match), so it outright outranks a same-text "graph-query"-class
+ *  keyword-spot guess (0.7) instead of triggering a same-class {ambiguousParse}
+ *  tie against it (see this file's header). */
+export const constructionsStrategy = {
+  id: "constructions",
+  class: "construction",
+  run(text) {
+    const parsed = parseConstruction(text);
+    return parsed
+      ? { strategyId: "constructions", class: "construction", candidates: [{ parsed, confidence: 0.9 }] }
+      : null;
+  },
+};

package/src/interpret/strategies/grammar.mjs

@@ -7,7 +7,7 @@
 import {
   VERB_TO_KIND, ENTITY_TO_TYPE, MODIFIER_TO_KIND,
   META_MEANING_VERBS, WHERE_MARKERS, MENTION_MARKERS,
-  INHERITS_REVERSE_VERBS, stripTrailingScopeFiller,
+  INHERITS_REVERSE_VERBS, stripTrailingScopeFiller, stripTrailingDiscourseTag,
 } from "../../ask-vocab.mjs";
 import { escapeRegex } from "../normalize.mjs";
 
@@ -96,12 +96,15 @@ const TEMPLATES = [
   // Fix 3: stripTrailingScopeFiller (ask-vocab.mjs) trims a curated trailing clause
   // ("what is a Module in this graph" -> "Module") off the object before it's
   // returned, so a scoping tail never corrupts the lookup term either the bare-form
-  // check above or downstream resolution/rendering perform.
+  // check above or downstream resolution/rendering perform. HANDOVER.md 2026-07-10
+  // item 8: stripTrailingDiscourseTag trims a bare trailing "then"/"though" the
+  // same way ("what is a component then" -> "component") — applied first, since a
+  // discourse tag sits outermost when both happen to stack.
   {
     name: "meta-whatis",
     re: new RegExp(`^what\\s+(?:is|are)\\s+(?:(an?)\\s+)?(.+?)\\??$`, "i"),
     build: (m) => {
-      const object = m[2].trim();
+      const object = stripTrailingDiscourseTag(m[2].trim());
       if (!m[1] && !ENTITY_TO_TYPE[object.toLowerCase()]) return null; // bare form: closed-set only
       return { shape: "meta", entityType: null, modifier: "direct", kind: "meta", object: stripTrailingScopeFiller(object) };
     },
@@ -135,6 +138,24 @@ const TEMPLATES = [
       ? { shape: "when", entityType: null, modifier: "direct", kind: "touches", object: m[1].trim() }
       : null),
   },
+  // T9 commit-history NP (PLAN_CHAT_FEEL item 6 remainder): "the commit history of
+  // X" / "commit history for X" — an NP form of T8's SAME "when did X change"
+  // intent; reuses shape="when" verbatim so evaluation/rendering are byte-
+  // identical, only the recognizer surface differs.
+  {
+    name: "commit-history",
+    re: /^(?:the\s+)?commit\s+history\s+(?:of|for)\s+(.+?)\??$/i,
+    build: (m) => ({ shape: "when", entityType: null, modifier: "direct", kind: "touches", object: m[1].trim() }),
+  },
+  // T10 cochange-partners NP (PLAN_CHAT_FEEL item 6 remainder): "cochange partners
+  // of X" — an NP form of the existing "which modules cochange with X" verb-phrase
+  // shape (ask-vocab.mjs's cochange verb table); reuses shape="reverse"/
+  // kind="cochange" so evaluation is byte-identical.
+  {
+    name: "cochange-partners",
+    re: /^co-?change\s+partners\s+(?:of|for|with)\s+(.+?)\??$/i,
+    build: (m) => ({ shape: "reverse", entityType: "Module", modifier: "direct", kind: "cochange", object: m[1].trim() }),
+  },
 ];
 
 /** Strategy 1: the original P0 anchored grammar — the whole (normalized) string

package/src/interpret/strategies/keywords.mjs

@@ -96,6 +96,27 @@ export function parseKeywordSpot(text, nlp = null) {
       return { shape: kind, entityType: null, modifier: "direct", kind, object: objText };
     }
   }
+  // has/have/had-changed carve-out (PLAN_CHAT_FEEL item 6 remainder): "has X
+  // changed" / "have X touched" / "had X ever been updated" — the present/past-
+  // perfect yes/no frame over the SAME touches-family verb the "when" branch below
+  // already answers ("when did X change"). Routed BEFORE the general verb-priority
+  // scan: "has"/"have"/"had" is ALSO a curated `defines` verb ("this module HAS
+  // three functions") and, sitting first in the sentence, would otherwise win that
+  // scan outright, misreading "has X changed" as a `defines` reverse-query over
+  // the object "X changed" — a confidently-wrong grain, not an honest miss. Gated
+  // on the sentence's OWN final word being a genuine touches verb (never a guess
+  // at which verb the sentence "really" means) — an ordinary defines question
+  // ("has app.mjs three functions") never ends in one, so this can't shadow it.
+  const PERFECT_AUX = new Set(["has", "have", "had"]);
+  if (PERFECT_AUX.has(lcWords[0])) {
+    let end = lcWords.length;
+    while (end > 1 && (lcWords[end - 1] === "ever" || lcWords[end - 1] === "been")) end -= 1;
+    const tailVerb = end > 1 ? VERB_TO_KIND[lcWords[end - 1]] : null;
+    if (tailVerb === "touches") {
+      const objText = words.slice(1, end - 1).filter((_, j) => !STOPWORDS.has(lcWords[1 + j])).join(" ").trim();
+      if (objText) return { shape: "when", entityType: null, modifier: "direct", kind: "touches", object: objText };
+    }
+  }
   let canonWords = lcWords;
   let verbHit = findPhrase(lcWords, VERB_TO_KIND);
   if (!verbHit && nlp) {
@@ -192,6 +213,19 @@ export function parseKeywordSpot(text, nlp = null) {
     if (objText) return { shape: "when", entityType: null, modifier: "direct", kind: "touches", object: objText };
   }
 
+  // who-last shape (HANDOVER.md 2026-07-10 item 5): "who last touched X" / "who
+  // touched X last" used to fall into the ordinary reverse-list shape below and
+  // list EVERY touching commit's author, ignoring "last" entirely — the mirror
+  // gap of the "when" shape just above, which already answers a single newest
+  // commit. "who"/"last" are both STOPWORDS (normalize.mjs), so they never
+  // survive into beforeText/afterText either way — checked here, before they're
+  // stripped, the same way the "when" check above reads `lcWords` directly
+  // rather than the post-strip text.
+  if (kind === "touches" && lcWords.includes("who") && lcWords.includes("last")) {
+    const objText = beforeText || afterText;
+    if (objText) return { shape: "whoLast", entityType: null, modifier: "direct", kind: "touches", object: objText };
+  }
+
   // reversible passive (Cycle 6, archive/PLAN_CYCLE_4.md): "PATIENT is VERBed BY AGENT" — an
   // agent-marking "by" plus a passive auxiliary flips the active reading, so the AGENT
   // (after "by") is the edge SUBJECT and the PATIENT the edge OBJECT. Object-first

package/src/memory/blocks.mjs

@@ -43,7 +43,7 @@ const INDEX_NAME = "index.json";
 
 export const PAGERANK_DAMPING = 0.85;
 export const PAGERANK_ITERATIONS = 20;
-const OVERLAP_MIN = 2;          // shared tokens for a similarity edge
+export const OVERLAP_MIN = 2;   // shared tokens for a similarity edge
 const MAX_TOKENS_PER_BLOCK = 800; // beyond tokenizeProse's per-doc cap: union over lines
 
 const blocksDir = (dir) => join(dir, BLOCKS_DIR_REL);
@@ -87,8 +87,13 @@ export async function loadBlockIndex(dir) {
  * `tokensById` is a plain { id: tokens[] } map; an undirected edge joins two
  * blocks sharing at least `overlapMin` tokens. Returns { ids, neighbours }
  * (neighbours[i] is an array of adjacent indices into ids).
+ *
+ * Exported (as of Stage 0 of PLAN_COMPLETIONS.md) so src/completions/group.mjs can build
+ * connected components over the same similarity graph rankBlocks/degreeOf already use,
+ * rather than re-deriving the shared-token-overlap logic — this IS the clustering
+ * primitive this graph was missing; grouping is layered on top of it, not duplicated.
  */
-function buildNeighbours(tokensById, overlapMin) {
+export function buildNeighbours(tokensById, overlapMin) {
   const ids = Object.keys(tokensById || {});
   const N = ids.length;
   const sets = ids.map((id) => new Set(tokensById[id] || []));