@lh8ppl/claude-memory-kit 0.2.4 → 0.3.0

package/src/index-rebuild.mjs

@@ -47,6 +47,7 @@ import { existsSync, readdirSync, readFileSync, statSync } from 'node:fs';
 import { basename, join, relative } from 'node:path';
 import chokidar from 'chokidar';
 import { INDEX_DB_SCHEMA } from './index-db.mjs';
+import { syncTranscriptChunks } from './transcript-index.mjs';
 import { readBullet, parseBulletProvenance } from './provenance.mjs';
 import { parse as parseFrontmatter } from './frontmatter.mjs';
 import {
@@ -145,7 +146,10 @@ export function parseObservationsFromScratchpad({
   projectRoot,
   userDir,
 }) {
-  const lines = content.split('\n');
+  // Task 139 (D-126): CRLF-tolerant read — autocrlf clones rewrite the
+  // committed memory files; a strict-\n split left \r on every line and
+  // the bullet/provenance regexes went blind.
+  const lines = content.split(/\r?\n/);
   const sha1 = sha1OfContent(content);
   const source_file = relativeSource(path, { projectRoot, userDir });
   const baseName = basename(path);
@@ -435,6 +439,12 @@ export function reindexBoot({ projectRoot, userDir, db, now }) {
     });
     const knownPaths = db.prepare('SELECT path FROM files').all();
     for (const { path: relPath } of knownPaths) {
+      // Task 104.2 composition guard: 'transcript:'-prefixed checkpoints
+      // belong to the transcript scope (transcript-index.mjs) — they are
+      // never in the observation live-set and pruning them here would
+      // defeat that scope's checkpoint on every boot. Its own sync prunes
+      // its own orphans.
+      if (relPath.startsWith('transcript:')) continue;
       if (liveRelPaths.has(relPath)) continue;
       const obsCount = db
         .prepare('SELECT COUNT(*) AS n FROM observations WHERE source_file = ?')
@@ -443,12 +453,24 @@ export function reindexBoot({ projectRoot, userDir, db, now }) {
     }
   }
 
+  // Task 104.2 — sync the transcript scope (the L3 raw tier) in the same
+  // boot pass. Cheap: per-file sha1 checkpoint; best-effort — a transcript
+  // sync hiccup must not fail the observation reindex.
+  let transcripts = { files: 0, chunks: 0 };
+  try {
+    transcripts = syncTranscriptChunks({ db, projectRoot, now: ts });
+  } catch {
+    // best-effort; the next boot retries
+  }
+
   return {
     filesScanned,
     filesReindexed,
     observationsAffected,
     filesPruned,
     observationsPruned,
+    transcriptFiles: transcripts.files,
+    transcriptChunks: transcripts.chunks,
     durationMs: Date.now() - t0,
     skipped,
   };
@@ -464,13 +486,20 @@ export function reindexBoot({ projectRoot, userDir, db, now }) {
 export function reindexFull({ projectRoot, userDir, db, now }) {
   const t0 = Date.now();
   const ts = now ?? t0;
-  // Drop + recreate (faster than per-row DELETE).
+  // Drop + recreate (faster than per-row DELETE). Task 104.2: the transcript
+  // scope drops + rebuilds with everything else — `files` carries its
+  // checkpoints, so a full reindex must re-chunk from scratch too.
   db.exec(`
     DROP TABLE IF EXISTS observations_fts;
     DROP TRIGGER IF EXISTS obs_after_insert;
     DROP TRIGGER IF EXISTS obs_after_update;
     DROP TRIGGER IF EXISTS obs_after_delete;
     DROP TABLE IF EXISTS observations;
+    DROP TABLE IF EXISTS transcript_chunks_fts;
+    DROP TRIGGER IF EXISTS tch_after_insert;
+    DROP TRIGGER IF EXISTS tch_after_update;
+    DROP TRIGGER IF EXISTS tch_after_delete;
+    DROP TABLE IF EXISTS transcript_chunks;
     DROP TABLE IF EXISTS files;
   `);
   db.exec(INDEX_DB_SCHEMA);
@@ -514,9 +543,20 @@ export function reindexFull({ projectRoot, userDir, db, now }) {
     observationsAffected += txn(source, sha1);
   }
 
+  // Task 104.2 — rebuild the transcript scope from scratch (its tables were
+  // dropped above). Best-effort, same contract as the boot-path sync.
+  let transcripts = { files: 0, chunks: 0 };
+  try {
+    transcripts = syncTranscriptChunks({ db, projectRoot, now: ts });
+  } catch {
+    // best-effort; the next reindex retries
+  }
+
   return {
     filesScanned,
     observationsAffected,
+    transcriptFiles: transcripts.files,
+    transcriptChunks: transcripts.chunks,
     durationMs: Date.now() - t0,
     skipped,
   };

package/src/inject-context.mjs

@@ -61,6 +61,33 @@ function trustLabel(rank) {
 const DEFAULT_CAP_BYTES = 13_000;
 const HOOK_EVENT_NAME = 'SessionStart';
 
+// Task 75.0 (D-64 / memory-os Layer-07 "Ground Truth", D-73 near-verbatim):
+// injecting memory is insufficient — the agent must be TOLD the injected
+// context is authoritative, or it re-derives from code what the snapshot
+// already answers (the D-40 cold-open failure). This preamble leads every
+// non-empty snapshot. It is code-generated (not template-scaffolded) on
+// purpose: always present, never consolidated/evicted/graduated, and
+// existing installs pick it up on upgrade with no re-scaffold (avoids the
+// Task-73 stale-template class).
+//
+// §7.1 composition: the preamble + its 2 joining newlines must fit the
+// 725-byte slack between Σ TIER_BUDGETS (12,275) and DEFAULT_CAP_BYTES
+// (13,000) — worst case 12,275 + len + 2 ≤ 13,000, i.e. len ≤ 723. The
+// boundary test pins len ≤ 700. injectContext also subtracts the reserve
+// from the cap handed to enforceCap, so custom capBytes stay honored.
+export const AUTHORITATIVE_MEMORY_PREAMBLE = [
+  '# Injected memory — AUTHORITATIVE (claude-memory-kit)',
+  '',
+  'Ground-truth ranking: (1) terminal/tool output → live system state;',
+  '(2) THIS snapshot + `cmk search` → documented knowledge & prior decisions;',
+  '(3) official docs → version-specifics; (4) training knowledge → verify against 1-3.',
+  '',
+  'When injected memory contradicts your assumptions, injected memory wins.',
+  'Lead with memory — never re-derive from code what it already answers, and',
+  'never treat a question as novel when the answer is already in your prompt.',
+  'This snapshot is a bounded hot index; `cmk search "<topic>"` reaches the facts not shown here.',
+].join('\n');
+
 // Match any line containing a `(P-XXXXXXXX)`-shaped citation id. Looser
 // than ID_PATTERN on purpose — alphabet-validation is the writer's job;
 // here we just want to recognize "any line that LOOKS like it carries a
@@ -520,7 +547,12 @@ function truncateTierToBudget(blockText, budget, valueById = new Map()) {
 // lowest-priority tier wholesale, logged as a dropped_tiers event.
 // This shouldn't fire under the documented budget table (1500+4500+
 // 4000 = 10000 ≤ 10240 default cap), but the safety net is cheap.
-function enforceCap(orderedBlocks, capBytes, ts) {
+// `reportCapBytes` (Task 75.0): the CALLER-facing cap for Door-4 events.
+// injectContext hands enforceCap a cap reduced by the preamble reserve;
+// truncation.log must still report the capBytes the user configured, not
+// the internal effective value, or the log reads as nonsense (411 when
+// the user set 1024).
+function enforceCap(orderedBlocks, capBytes, ts, reportCapBytes = capBytes) {
   const tierEvents = [];
   // Step 1: per-tier budget enforcement (section-granular).
   for (const block of orderedBlocks) {
@@ -559,7 +591,7 @@ function enforceCap(orderedBlocks, capBytes, ts) {
     bytes -= Buffer.byteLength(dropped.text, 'utf8');
     let event = dropEvents[dropEvents.length - 1];
     if (!event) {
-      event = { ts, capBytes, dropped_tiers: [] };
+      event = { ts, capBytes: reportCapBytes, dropped_tiers: [] };
       dropEvents.push(event);
     }
     event.dropped_tiers.push(dropped.tier);
@@ -707,15 +739,26 @@ export function injectContext({
   }
 
   // 3. Cap enforcement: drop whole tier blocks from the tail until within
-  // capBytes. Each drop emits one truncation event.
+  // capBytes. Each drop emits one truncation event. The authoritative-memory
+  // preamble (Task 75.0) is reserved out of the cap up front so the final
+  // snapshot (preamble + blocks) still honors capBytes exactly.
+  const preambleReserve =
+    rawBlocks.length > 0
+      ? Buffer.byteLength(AUTHORITATIVE_MEMORY_PREAMBLE, 'utf8') + 2
+      : 0;
   const { blocks: keptBlocks, truncationEvents } = enforceCap(
     rawBlocks,
-    cap,
+    Math.max(0, cap - preambleReserve),
     ts,
+    cap,
   );
 
-  // 4. Concatenate.
-  const snapshot = keptBlocks.map((b) => b.text).join('\n');
+  // 4. Concatenate. The preamble leads every non-empty snapshot; an empty
+  // snapshot stays empty (don't claim authoritative memory with nothing
+  // behind it).
+  const body = keptBlocks.map((b) => b.text).join('\n');
+  const snapshot =
+    body === '' ? '' : `${AUTHORITATIVE_MEMORY_PREAMBLE}\n\n${body}`;
 
   // 5. Persist side-effect logs under <projectRoot>/context/.locks/. We
   // only write the project-tier .locks file (which is the well-known

package/src/install.mjs

@@ -39,6 +39,7 @@ import {
   writeFileSync,
 } from 'node:fs';
 import { homedir } from 'node:os';
+import { spawnSync } from 'node:child_process';
 import { basename, dirname, join, relative, resolve } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import { injectClaudeMdBlock } from './claude-md.mjs';
@@ -411,7 +412,112 @@ export async function install(options = {}) {
     }
   }
 
-  return { projectRoot, userTier, created, skipped, gitignore, claudeMd, hooks, mcpServer, errors };
+  // Task 46 — semantic-recall opt-in/out. `--with-semantic`: install the
+  // optional embedder (~260 MB once, fully local), flip the project's
+  // default search mode to hybrid, and pre-warm the model so the one-time
+  // download happens NOW, not as a surprise on the first search.
+  // `--no-semantic`: pin keyword explicitly. Neither flag → settings
+  // untouched (keyword by absence). The npm spawn is injectable
+  // (options.spawnNpm) so tests assert the argv without touching the host.
+  // Both flags together → withSemantic wins (the affirmative opt-in beats
+  // the pin-off; checked first below).
+  let semantic = { action: 'skipped' };
+  if (options.withSemantic) {
+    semantic = await enableSemantic({ projectRoot, spawnNpm: options.spawnNpm, warm: options.warmEmbedder });
+    if (semantic.action === 'error') errors.push({ path: 'semantic', error: semantic.error });
+  } else if (options.noSemantic) {
+    const r = mergeProjectSettings(projectRoot, { search: { default_mode: 'keyword' } });
+    semantic = r.ok
+      ? { action: 'disabled', path: r.path }
+      : { action: 'error', error: r.error };
+    if (!r.ok) errors.push({ path: r.path, error: r.error });
+  }
+
+  return { projectRoot, userTier, created, skipped, gitignore, claudeMd, hooks, mcpServer, semantic, errors };
+}
+
+/**
+ * Read-merge-write <projectRoot>/context/settings.json, preserving every
+ * key the user already has (over-mutation-safe; deep-merges one level).
+ */
+export function mergeProjectSettings(projectRoot, patch) {
+  const path = join(projectRoot, 'context', 'settings.json');
+  try {
+    let current = {};
+    if (existsSync(path)) {
+      current = JSON.parse(readFileSync(path, 'utf8'));
+    }
+    const next = { ...current };
+    for (const [key, value] of Object.entries(patch)) {
+      next[key] =
+        value && typeof value === 'object' && !Array.isArray(value)
+          ? { ...(current[key] ?? {}), ...value }
+          : value;
+    }
+    mkdirSync(dirname(path), { recursive: true });
+    writeFileSync(path, JSON.stringify(next, null, 2) + '\n', 'utf8');
+    return { ok: true, path };
+  } catch (err) {
+    return { ok: false, path, error: err?.message ?? String(err) };
+  }
+}
+
+/**
+ * The production npm-spawn closure, as an injectable-seam factory
+ * (Task 125.4) so its argv/shell/timeout contract is testable without
+ * running a real `npm install -g` (which stays a machine-level step
+ * tests must never take).
+ */
+export function buildDefaultNpmRunner({ spawnSyncImpl = spawnSync } = {}) {
+  return () => {
+    // One constant command string under shell:true (no user input — and
+    // an args array + shell:true trips Node's DEP0190). npm is npm.cmd
+    // on Windows; the shell resolves it cross-platform.
+    const r = spawnSyncImpl('npm install -g @huggingface/transformers', {
+      encoding: 'utf8',
+      stdio: 'inherit',
+      shell: true,
+      // spawn-discipline (design §8.5): a hung registry shouldn't hang
+      // install forever; 10 min covers the ~46 MB package on slow links.
+      timeout: 600_000,
+    });
+    return { status: r.status, error: r.error?.message };
+  };
+}
+
+async function enableSemantic({ projectRoot, spawnNpm, warm }) {
+  // 1. Install the optional embedder globally (it resolves as a sibling of
+  // the globally-installed kit). Injectable for tests.
+  const runNpm = spawnNpm ?? buildDefaultNpmRunner();
+  const npm = runNpm();
+  if (npm.status !== 0) {
+    return {
+      action: 'error',
+      error: `npm install -g @huggingface/transformers failed (${npm.error ?? `exit ${npm.status}`}) — semantic recall NOT enabled; keyword search is unaffected`,
+    };
+  }
+  // 2. Flip the project default to hybrid ONLY after the dependency landed
+  // (no half-state: a hybrid default without an embedder would degrade
+  // every search to a fallback warning).
+  const settings = mergeProjectSettings(projectRoot, { search: { default_mode: 'hybrid' } });
+  if (!settings.ok) {
+    return { action: 'error', error: settings.error };
+  }
+  // 3. Pre-warm (best-effort): the one-time model download happens during
+  // install, not on the first search. Injectable for tests.
+  let warmed = { ok: false, reason: 'skipped' };
+  try {
+    const warmFn =
+      warm ??
+      (async () => {
+        const { warmEmbedder } = await import('./semantic-backend.mjs');
+        return warmEmbedder();
+      });
+    warmed = await warmFn();
+  } catch (err) {
+    warmed = { ok: false, reason: err?.message ?? String(err) };
+  }
+  return { action: 'enabled', path: settings.path, defaultMode: 'hybrid', warmed };
 }
 
 /**

package/src/mcp-server.mjs

@@ -106,16 +106,60 @@ export function validatePath(p, { projectRoot, userDir }) {
 
 // --- Tool handlers ----------------------------------------------------
 
-function makeMkSearch({ db, semanticBackend }) {
-  return async ({ query, mode, tier, since, limit, min_trust }) => {
+function makeMkSearch({ db, semanticBackend, projectRoot }) {
+  return async ({ query, mode, scope, tier, since, limit, min_trust }) => {
+    // Task 46: explicit mode wins; otherwise the project's configured
+    // default (search.default_mode — set by `cmk install --with-semantic`).
+    const { prepareSemanticBackend, resolveDefaultSearchMode } = await import(
+      './semantic-backend.mjs'
+    );
+    let wantMode =
+      mode ??
+      (projectRoot ? resolveDefaultSearchMode({ projectRoot }) : SEARCH_MODES.KEYWORD);
+    // Task 65: when the caller asks for semantic/hybrid and no test seam is
+    // injected, prepare the REAL embedded backend (lazy-optional — an absent
+    // embedder degrades to the actionable error below; keyword unaffected).
+    let backend = semanticBackend;
+    let degradedNote = null;
+    if (
+      backend === undefined &&
+      (wantMode === SEARCH_MODES.SEMANTIC || wantMode === SEARCH_MODES.HYBRID)
+    ) {
+      const prep = await prepareSemanticBackend({ db, query, scope: scope ?? 'facts' });
+      if (!prep.ok && mode) {
+        // Explicitly requested — surface the actionable error.
+        return {
+          content: [
+            {
+              type: 'text',
+              text: `error: semantic backend unavailable (${prep.reason}). ${prep.hint ?? 'Use mode "keyword".'}`,
+            },
+          ],
+          isError: true,
+        };
+      }
+      if (!prep.ok) {
+        // Configured default can't run — degrade gracefully to keyword,
+        // but NOT silently (Task 125.1, the user's call reversing the
+        // Task-46 review skip): the note below tells the model what it
+        // got, so it can relay the fix to the user.
+        wantMode = SEARCH_MODES.KEYWORD;
+        degradedNote =
+          `note: this project's configured default search is semantic (hybrid), but the embedder is unavailable (${prep.reason}) — these are keyword-only results. ` +
+          'Suggest the user run `cmk install --with-semantic` to restore semantic recall.';
+      } else {
+        backend = prep.backend;
+      }
+    }
     const r = search({
       db, query,
-      mode: mode ?? SEARCH_MODES.KEYWORD,
+      mode: wantMode,
+      scope,
       tier,
       since,
       limit,
       minTrust: min_trust,
-      semanticBackend,
+      semanticBackend: backend,
     });
     if (r.action === 'error') {
       return {
@@ -124,7 +168,12 @@ function makeMkSearch({ db, semanticBackend }) {
       };
     }
     return {
-      content: [{ type: 'text', text: JSON.stringify(r.results, null, 2) }],
+      content: [
+        { type: 'text', text: JSON.stringify(r.results, null, 2) },
+        // Results stay content[0] (shape-compatible); the degradation note,
+        // when present, rides as a second block.
+        ...(degradedNote ? [{ type: 'text', text: degradedNote }] : []),
+      ],
     };
   };
 }
@@ -505,17 +554,18 @@ export function buildMcpServer({ projectRoot, userDir, db, semanticBackend }) {
   server.registerTool(
     'mk_search',
     {
-      description: 'Search kit memory (FTS5 keyword by default; semantic + hybrid require the Layer-5b semantic backend, not yet shipped).',
+      description: 'Search kit memory. FTS5 keyword by default; semantic + hybrid use the embedded Layer-5b backend (sqlite-vec + a local ONNX embedder — needs the optional @huggingface/transformers install).',
       inputSchema: {
         query: z.string().min(1).describe('search query'),
         mode: z.enum(['keyword', 'semantic', 'hybrid']).optional(),
+        scope: z.enum(['facts', 'transcripts']).optional().describe("'facts' (default) = curated memory; 'transcripts' = the raw session record — the LAST-RESORT recall tier, search it only when curated memory has no answer"),
         tier: z.enum(['U', 'P', 'L']).optional(),
         since: z.string().optional().describe('ISO 8601 timestamp'),
         limit: z.number().int().positive().max(1000).optional(),
         min_trust: z.enum(['low', 'medium', 'high']).optional(),
       },
     },
-    makeMkSearch({ db, semanticBackend }),
+    makeMkSearch({ db, semanticBackend, projectRoot }),
   );
 
   // mk_get

package/src/merge-facts.mjs

@@ -26,6 +26,7 @@ import { parse, format } from './frontmatter.mjs';
 import { appendAuditEntry, nowIso, REASON_CODES } from './audit-log.mjs';
 import { ERROR_CATEGORIES, errorResult, notFoundResult } from './result-shapes.mjs';
 import { writeFact } from './write-fact.mjs';
+import { reindex } from './reindex.mjs';
 
 function listLiveFactFiles(factDir) {
   if (!existsSync(factDir)) return [];
@@ -193,6 +194,17 @@ export function mergeFacts(opts = {}) {
   const supersededA = moveToSuperseded(matchA, writeResult.id);
   const supersededB = moveToSuperseded(matchB, writeResult.id);
 
+  // Task 124 (the D-112 class): writeFact refreshed INDEX.md when C was
+  // created — but A and B left the fact dir AFTER that, so the index kept
+  // both as dangling lines until a manual `cmk reindex`. The writer owns
+  // the derived view on the removal side too. Best-effort, same contract
+  // as writeFact's: the merge is already durable on disk.
+  try {
+    reindex({ tier, projectRoot, userDir, warn: () => {} });
+  } catch {
+    // index rebuild is best-effort; the merge already succeeded
+  }
+
   const ts = now ?? nowIso();
   appendAuditEntry(tierRoot, {
     ts,

package/src/provenance.mjs

@@ -44,6 +44,10 @@ const VALID_WRITE_SOURCES = new Set([
   'compressor',
   'manual-edit',
   'imported',
+  // Task 138 review finding: the conflict-queue merge-both action writes a
+  // merged bullet to the scratchpad; its provenance needs a valid write key
+  // (the old hand-rolled comment had none and broke reindex - D-125 class).
+  'merged',
 ]);
 const REQUIRED_PROVENANCE_FIELDS = [
   'source',

package/src/result-shapes.mjs

@@ -104,7 +104,7 @@ export const ERROR_CATEGORIES = Object.freeze({
   POISON_GUARD: 'poison_guard',
 
   // `cmk search` requested --mode=semantic or --mode=hybrid but the
-  // Layer-5b semantic backend is not yet shipped (Task 30, design
+  // Layer-5b semantic backend's optional embedder is not installed (Task 30/65, design
   // §9.3). Pairs with `process.exitCode = 2` in subcommands.mjs per
   // tasks.md 30.2's explicit "exit 2 when unavailable" contract.
   // NO silent fallback to keyword — the user asked for semantic,

package/src/scratchpad.mjs

@@ -180,7 +180,9 @@ function findSectionRange(lines, sectionTitle) {
 }
 
 function insertIntoSection(text, sectionTitle, bullet) {
-  const lines = text.split('\n');
+  // Task 139 (D-126): CRLF-tolerant read; the join below re-emits \n,
+  // so a CRLF-converted scratchpad self-heals on the next write.
+  const lines = text.split(/\r?\n/);
   const range = findSectionRange(lines, sectionTitle);
   if (!range) return null;
   // Insert before the next `## ` heading; skip trailing blank lines so the
@@ -208,7 +210,7 @@ function insertIntoSection(text, sectionTitle, bullet) {
 export function ensureSectionExists(scratchpadPath, sectionTitle) {
   if (!existsSync(scratchpadPath)) return { created: false, error: 'no-file' };
   const text = readFileSync(scratchpadPath, 'utf8');
-  if (findSectionRange(text.split('\n'), sectionTitle)) return { created: false };
+  if (findSectionRange(text.split(/\r?\n/), sectionTitle)) return { created: false }; // Task 139: CRLF-tolerant
   const body = text.trimEnd(); // drop trailing whitespace/blank lines (no `\s+$` regex — trips ReDoS heuristics)
   // No leading blank lines for an empty/whitespace-only file (the scaffolded
   // scratchpads are never empty, but keep the output clean if one ever is).
@@ -220,7 +222,7 @@ export function ensureSectionExists(scratchpadPath, sectionTitle) {
 const EVICTED_ID_RE = /^- \(([PUL]-[A-Za-z0-9]+)\)/;
 
 function consolidate(text, { nowDate }) {
-  const lines = text.split('\n');
+  const lines = text.split(/\r?\n/); // Task 139: CRLF-tolerant
   const removeIdx = new Set();
   const evicted = [];
   const staleCutoff = new Date(nowDate.getTime() - STALE_AFTER_DAYS * 24 * 60 * 60 * 1000);

package/src/search.mjs

@@ -11,7 +11,7 @@
 //             ~100ms for 10k bullets. Always available — the keyword
 //             backend ships in v0.1.0 with no extra install.
 //
-//   semantic  the Layer-5b semantic backend (not yet shipped — the embedded
+//   semantic  the Layer-5b semantic backend (Task 65: sqlite-vec + local ONNX embedder; the embedded
 //             vector backend is a future release; the DI seam below is the
 //             drop-in point). Until then this mode errors with
 //             ERROR_CATEGORIES.SEMANTIC_UNAVAILABLE when the caller
@@ -54,6 +54,15 @@ export const SEARCH_MODES = Object.freeze({
 export const DEFAULT_LIMIT = 20;
 const MAX_LIMIT = 1000;
 
+// Task 104.2 (D-117) — search scopes. 'facts' = the curated observation
+// index (L1, the default). 'transcripts' = the SEPARATE raw-transcript
+// chunk index (the L3 last-resort tier) — reached ONLY when explicitly
+// asked, so raw history never pollutes curated results.
+export const SEARCH_SCOPES = Object.freeze({
+  FACTS: 'facts',
+  TRANSCRIPTS: 'transcripts',
+});
+
 const TRUST_ORDINAL = Object.freeze({
   low: 1,
   medium: 2,
@@ -107,7 +116,24 @@ function validateInput(opts) {
       errors.push(`limit: must be a positive integer ≤ ${MAX_LIMIT}`);
     }
   }
-  return { errors, mode };
+  const scope = opts.scope ?? SEARCH_SCOPES.FACTS;
+  if (scope !== SEARCH_SCOPES.FACTS && scope !== SEARCH_SCOPES.TRANSCRIPTS) {
+    errors.push(`scope: must be one of facts/transcripts (got ${JSON.stringify(scope)})`);
+  }
+  if (scope === SEARCH_SCOPES.TRANSCRIPTS) {
+    // Chunks carry no tier/trust/created_at — rejecting these is more honest
+    // than silently ignoring them (the explicit-vs-configured asymmetry rule).
+    for (const [key, label] of [
+      ['tier', 'tier'],
+      ['minTrust', 'minTrust'],
+      ['since', 'since'],
+    ]) {
+      if (opts[key] !== undefined) {
+        errors.push(`${label}: not supported under the transcripts scope (raw chunks carry no ${label})`);
+      }
+    }
+  }
+  return { errors, mode, scope };
 }
 
 // --- Keyword (FTS5 BM25) backend --------------------------------------
@@ -211,6 +237,60 @@ function runKeywordSearch(db, opts) {
   }));
 }
 
+// --- Transcript-scope keyword backend (Task 104.2, the L3 raw tier) ----
+
+const TRANSCRIPT_KEYWORD_SQL = `
+SELECT
+  t.source_file AS source_file,
+  t.source_line AS source_line,
+  t.heading AS heading,
+  transcript_chunks_fts.rank AS score,
+  snippet(transcript_chunks_fts, 0, '<b>', '</b>', '...', 16) AS snippet
+FROM transcript_chunks_fts
+JOIN transcript_chunks t ON t.rowid = transcript_chunks_fts.rowid
+WHERE transcript_chunks_fts MATCH @query
+ORDER BY transcript_chunks_fts.rank
+LIMIT @limit
+`;
+
+// Synthetic, readable id for a raw chunk (chunks are locations, not curated
+// facts — no [PUL]-XXXXXXXX identity). Also the RRF fusion key in hybrid
+// mode and the drill-back handle the memory-search skill surfaces.
+function transcriptHitId(row) {
+  return `T:${row.source_file}:${row.source_line}`;
+}
+
+function runTranscriptKeywordSearch(db, opts) {
+  let rows;
+  try {
+    rows = db
+      .prepare(TRANSCRIPT_KEYWORD_SQL)
+      .all({ query: opts.query, limit: opts.limit ?? DEFAULT_LIMIT });
+  } catch (err) {
+    if (err?.code === 'SQLITE_ERROR' || /fts5:|no such column:/i.test(err?.message ?? '')) {
+      throw new FTS5ParseError(err, opts.query);
+    }
+    throw err;
+  }
+  return rows.map((r) => ({
+    id: transcriptHitId(r),
+    // Raw turns contain newlines (dialogue + Tools blocks) — flatten so the
+    // one-line-per-hit output contract holds across scopes.
+    snippet: flattenSnippet(r.snippet),
+    source_file: r.source_file,
+    source_line: r.source_line,
+    heading: r.heading,
+    score: r.score,
+  }));
+}
+
+const TRANSCRIPT_SNIPPET_MAX = 240;
+
+function flattenSnippet(s) {
+  const flat = String(s ?? '').replace(/\s+/g, ' ').trim();
+  return flat.length > TRANSCRIPT_SNIPPET_MAX ? flat.slice(0, TRANSCRIPT_SNIPPET_MAX) + '…' : flat;
+}
+
 // --- Reciprocal-rank fusion (hybrid mode) -----------------------------
 
 /**
@@ -255,10 +335,15 @@ export function reciprocalRankFusion({
 // --- Public boundary --------------------------------------------------
 
 export function search(opts = {}) {
-  const { errors, mode } = validateInput(opts);
+  const { errors, mode, scope } = validateInput(opts);
   if (errors.length > 0) {
     return errorResult({ category: ERROR_CATEGORIES.SCHEMA, errors });
   }
+  // Scope dispatch (Task 104.2): the transcripts scope swaps the keyword
+  // backend; semantic/hybrid use the caller-prepared backend exactly like
+  // the facts scope (prepareSemanticBackend({scope}) embeds the right table).
+  const keywordBackend =
+    scope === SEARCH_SCOPES.TRANSCRIPTS ? runTranscriptKeywordSearch : runKeywordSearch;
 
   // Semantic + hybrid require an injected backend. Production v0.1.0
   // passes undefined → error with the not-yet-shipped hint. A future
@@ -268,7 +353,8 @@ export function search(opts = {}) {
       return errorResult({
         category: ERROR_CATEGORIES.SEMANTIC_UNAVAILABLE,
         errors: [
-          'the Layer-5b semantic backend is not yet shipped — semantic/hybrid search will land in a future release. ' +
+          'no semantic backend provided — semantic/hybrid need the embedded Layer-5b backend prepared by the caller ' +
+            '(the CLI/MCP do this automatically when the optional @huggingface/transformers embedder is installed). ' +
             'Use --mode=keyword for the always-available FTS5 search.',
         ],
       });
@@ -278,15 +364,16 @@ export function search(opts = {}) {
   let results;
   try {
     if (mode === SEARCH_MODES.KEYWORD) {
-      results = runKeywordSearch(opts.db, opts);
+      results = keywordBackend(opts.db, opts);
     } else if (mode === SEARCH_MODES.SEMANTIC) {
       // The semantic backend is an injected callable returning the same
-      // shape as runKeywordSearch (array of {id, snippet, source_file,
-      // source_line, tier, trust, score}).
+      // shape as the scope's keyword backend (facts: {id, snippet,
+      // source_file, source_line, tier, trust, score}; transcripts: the
+      // synthetic-T:-id shape without tier/trust).
       results = opts.semanticBackend(opts);
     } else {
       // hybrid: run both backends + fuse.
-      const keywordResults = runKeywordSearch(opts.db, opts);
+      const keywordResults = keywordBackend(opts.db, opts);
       const semanticResults = opts.semanticBackend(opts);
       const fused = reciprocalRankFusion({
         keywordResults,
@@ -308,5 +395,5 @@ export function search(opts = {}) {
     throw err;
   }
 
-  return { action: 'found', mode, results };
+  return { action: 'found', mode, scope, results };
 }