@jigyasudham/veto 3.1.0 → 3.1.2

package/dist/router/implicit-outcomes.js

@@ -36,10 +36,10 @@ export function mineImplicitOutcomes() {
     const wmRow = db.prepare('SELECT pattern_val FROM patterns WHERE pattern_key = ?').get(WATERMARK_KEY);
     const since = wmRow?.pattern_val ?? '1970-01-01 00:00:00';
     const placeholders = toolNames.map(() => '?').join(',');
-    const rows = db.prepare(`SELECT id, session_id, tool_name, result_status, recorded_at
-     FROM tool_call_trace_log
-     WHERE recorded_at > ? AND tool_name IN (${placeholders})
-     ORDER BY recorded_at ASC
+    const rows = db.prepare(`SELECT id, session_id, tool_name, result_status, recorded_at
+     FROM tool_call_trace_log
+     WHERE recorded_at > ? AND tool_name IN (${placeholders})
+     ORDER BY recorded_at ASC
      LIMIT ${BATCH_LIMIT}`).all(since, ...toolNames);
     if (rows.length === 0)
         return { mined: 0, errors: 0, retries: 0 };

package/dist/router/learning-updater.js

@@ -5,7 +5,7 @@ import { getConfig } from '../memory/config.js';
 // ─── Record ───────────────────────────────────────────────────────────────────
 export function recordOutcome(taskType, complexity, modelTier, agent, outputQuality, tokensUsed = 0, fileExt) {
     const db = getDb();
-    db.prepare(`INSERT INTO learning_data (id, task_type, complexity, model_tier, output_quality, tokens_used, agent)
+    db.prepare(`INSERT INTO learning_data (id, task_type, complexity, model_tier, output_quality, tokens_used, agent)
      VALUES (?, ?, ?, ?, ?, ?, ?)`).run(randomUUID(), taskType, complexity, modelTier, outputQuality, tokensUsed, agent);
     // Predictive routing: track agent quality per file extension and task type keyword
     if (agent && agent !== 'dynamic' && outputQuality > 0) {
@@ -50,8 +50,8 @@ export function getRecommendedAgent(taskType, fileExt) {
     const db = getDb();
     // Try file extension match first (more specific)
     if (fileExt) {
-        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
-       WHERE pattern_key LIKE ? AND seen_count >= 3
+        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
+       WHERE pattern_key LIKE ? AND seen_count >= 3
        ORDER BY confidence DESC LIMIT 1`).all(`file_agent:${fileExt.toLowerCase()}:%`);
         if (rows.length > 0 && rows[0].confidence >= 0.7)
             return rows[0].pattern_val;
@@ -59,8 +59,8 @@ export function getRecommendedAgent(taskType, fileExt) {
     // Fall back to task keyword match
     const keyword = taskType.toLowerCase().split(/[\s_-]/)[0];
     if (keyword) {
-        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
-       WHERE pattern_key LIKE ? AND seen_count >= 3
+        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
+       WHERE pattern_key LIKE ? AND seen_count >= 3
        ORDER BY confidence DESC LIMIT 1`).all(`task_agent:${keyword}:%`);
         if (rows.length > 0 && rows[0].confidence >= 0.7)
             return rows[0].pattern_val;
@@ -70,7 +70,7 @@ export function getRecommendedAgent(taskType, fileExt) {
 // ─── Stats ────────────────────────────────────────────────────────────────────
 export function getLearningStats() {
     const db = getDb();
-    const rows = db.prepare(`SELECT model_tier, COUNT(*) as count, AVG(output_quality) as avg_quality
+    const rows = db.prepare(`SELECT model_tier, COUNT(*) as count, AVG(output_quality) as avg_quality
      FROM learning_data GROUP BY model_tier`).all();
     const total_tasks = rows.reduce((sum, r) => sum + r.count, 0);
     const tier_breakdown = {};
@@ -85,16 +85,16 @@ export function getLearningStats() {
 }
 export function getAgentPerformanceStats() {
     const db = getDb();
-    const rows = db.prepare(`
-    SELECT agent,
-           COUNT(*) as task_count,
-           AVG(output_quality) as avg_quality,
-           AVG(tokens_used) as avg_tokens,
-           SUM(CASE WHEN output_quality < 60 THEN 1 ELSE 0 END) as low_quality_count
-    FROM learning_data
-    WHERE agent IS NOT NULL
-    GROUP BY agent
-    ORDER BY avg_quality ASC
+    const rows = db.prepare(`
+    SELECT agent,
+           COUNT(*) as task_count,
+           AVG(output_quality) as avg_quality,
+           AVG(tokens_used) as avg_tokens,
+           SUM(CASE WHEN output_quality < 60 THEN 1 ELSE 0 END) as low_quality_count
+    FROM learning_data
+    WHERE agent IS NOT NULL
+    GROUP BY agent
+    ORDER BY avg_quality ASC
   `).all();
     return rows.map(r => ({
         agent: r.agent,
@@ -107,23 +107,23 @@ export function getAgentPerformanceStats() {
 }
 export function getTaskTypeBreakdown() {
     const db = getDb();
-    const rows = db.prepare(`
-    SELECT task_type,
-           COUNT(*) as count,
-           AVG(complexity) as avg_complexity,
-           AVG(output_quality) as avg_quality,
-           model_tier as dominant_tier
-    FROM learning_data
-    GROUP BY task_type, model_tier
-    HAVING COUNT(*) = (
-      SELECT MAX(c) FROM (
-        SELECT COUNT(*) as c FROM learning_data l2
-        WHERE l2.task_type = learning_data.task_type
-        GROUP BY l2.model_tier
-      )
-    )
-    ORDER BY count DESC
-    LIMIT 50
+    const rows = db.prepare(`
+    SELECT task_type,
+           COUNT(*) as count,
+           AVG(complexity) as avg_complexity,
+           AVG(output_quality) as avg_quality,
+           model_tier as dominant_tier
+    FROM learning_data
+    GROUP BY task_type, model_tier
+    HAVING COUNT(*) = (
+      SELECT MAX(c) FROM (
+        SELECT COUNT(*) as c FROM learning_data l2
+        WHERE l2.task_type = learning_data.task_type
+        GROUP BY l2.model_tier
+      )
+    )
+    ORDER BY count DESC
+    LIMIT 50
   `).all();
     return rows.map(r => {
         // Under-tiered: task scored high quality but at a tier below what complexity suggests
@@ -142,15 +142,15 @@ export function getTaskTypeBreakdown() {
 export function getCouncilInsights() {
     const db = getDb();
     // Correlate council approvals (GREEN/YELLOW) with debugging sessions that came after
-    const rows = db.prepare(`
-    SELECT co.verdict, co.task, co.debated_at,
-           d.decision as related_decision, d.made_at as decision_at
-    FROM council_outcomes co
-    LEFT JOIN decisions d ON d.session_id = co.session_id
-      AND d.decision LIKE '%debug%' OR d.decision LIKE '%fix%' OR d.decision LIKE '%bug%'
-    WHERE co.verdict IN ('GREEN', 'YELLOW')
-    ORDER BY co.debated_at DESC
-    LIMIT 30
+    const rows = db.prepare(`
+    SELECT co.verdict, co.task, co.debated_at,
+           d.decision as related_decision, d.made_at as decision_at
+    FROM council_outcomes co
+    LEFT JOIN decisions d ON d.session_id = co.session_id
+      AND d.decision LIKE '%debug%' OR d.decision LIKE '%fix%' OR d.decision LIKE '%bug%'
+    WHERE co.verdict IN ('GREEN', 'YELLOW')
+    ORDER BY co.debated_at DESC
+    LIMIT 30
   `).all();
     return rows.map(r => {
         let days_to_debug = null;
@@ -273,9 +273,9 @@ export function recordRoutingFeedback(opts) {
     const now = new Date().toISOString();
     const expiresAt = new Date(Date.now() + FEEDBACK_TTL_DAYS * 86_400_000).toISOString();
     const task_hash = opts.task.toLowerCase().replace(/\s+/g, '_').slice(0, 40);
-    db.prepare(`
-    INSERT INTO routing_feedback (id, task_hash, task_snippet, complexity, model_tier, agent, outcome, session_id, recorded_at, expires_at)
-    VALUES (?, ?, ?, ?, ?, ?, 'pending', ?, ?, ?)
+    db.prepare(`
+    INSERT INTO routing_feedback (id, task_hash, task_snippet, complexity, model_tier, agent, outcome, session_id, recorded_at, expires_at)
+    VALUES (?, ?, ?, ?, ?, ?, 'pending', ?, ?, ?)
   `).run(id, task_hash, opts.task.slice(0, 100), opts.complexity, opts.model_tier, opts.agent ?? null, opts.session_id ?? null, now, expiresAt);
     return id;
 }

package/dist/server.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-// Veto MCP Server — 92 tools, LLM council + auto-learning router
+// Veto MCP Server — 93 tools, LLM council + auto-learning router
 // Suppress node:sqlite experimental warning — it would corrupt the MCP stdio protocol
 process.removeAllListeners('warning');
 import { Server } from '@modelcontextprotocol/sdk/server/index.js';

package/dist/tools/compact.js

@@ -1,5 +1,5 @@
-// Compact tool surface — progressive disclosure for clients where 92 tool
-// schemas would burn ~25K context tokens before the user types a word.
+// Compact tool surface — progressive disclosure for clients where 93 tool
+// schemas would burn ~20K context tokens before the user types a word.
 //
 // In compact mode (VETO_COMPACT=1 or config compact_tools=true) ListTools
 // exposes only a small first-class set plus two meta-tools: veto_find_tools
@@ -22,7 +22,7 @@ const FIRST_CLASS = [
 export const META_TOOL_DEFINITIONS = [
     {
         name: 'veto_find_tools',
-        description: 'Searches the full Veto catalog (92 tools: agents, reviews, advisors, generators, memory, learning, git, workflow) by keyword and returns matching tool schemas. Compact mode exposes only core tools up front — call this first to discover the right tool, then invoke it via veto_call.',
+        description: 'Searches the full Veto catalog (93 tools: agents, reviews, advisors, generators, memory, learning, git, workflow) by keyword and returns matching tool schemas. Compact mode exposes only core tools up front — call this first to discover the right tool, then invoke it via veto_call.',
         inputSchema: {
             type: 'object',
             properties: {

package/dist/transcripts/archive.js

@@ -79,14 +79,14 @@ export async function captureSession(opts = {}) {
         if (existing) {
             // Content changed (grew): overwrite the row, reset the index watermark AND
             // parser_version so Step 6 re-derives events from the new bytes.
-            db.prepare(`UPDATE archives SET archive_path=?, content_sha256=?, source_bytes=?, archive_bytes=?,
-           project_dir=COALESCE(?, project_dir), veto_session_id=COALESCE(?, veto_session_id),
+            db.prepare(`UPDATE archives SET archive_path=?, content_sha256=?, source_bytes=?, archive_bytes=?,
+           project_dir=COALESCE(?, project_dir), veto_session_id=COALESCE(?, veto_session_id),
            updated_at=?, indexed_through_seq=0, parser_version=0 WHERE id=?`).run(archivePath, sha256, bytes, archiveBytes, proj, opts.vetoSessionId ?? null, now, id);
         }
         else {
-            db.prepare(`INSERT INTO archives (id, source, source_session_id, project_dir, veto_session_id, archive_path,
-           content_sha256, source_bytes, archive_bytes, source_format_hint, parser_version, indexed_through_seq,
-           captured_at, updated_at)
+            db.prepare(`INSERT INTO archives (id, source, source_session_id, project_dir, veto_session_id, archive_path,
+           content_sha256, source_bytes, archive_bytes, source_format_hint, parser_version, indexed_through_seq,
+           captured_at, updated_at)
          VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 0, 0, ?, ?)`).run(id, source, sessionId, proj, opts.vetoSessionId ?? null, archivePath, sha256, bytes, archiveBytes, formatHint, now, now);
         }
         return {

package/dist/transcripts/ingest.js

@@ -42,13 +42,13 @@ export function ingestArchive(archiveId) {
         if (sessionIds.size > 0 && !sessionIds.has(row.source_session_id)) {
             return { status: 'session_mismatch', reason: 'archive session id not present in transcript' };
         }
-        const ins = db.prepare(`INSERT INTO events (id, archive_id, source_session_id, seq, line_index, block_index, kind,
-         source_type, role, tool_name, text, secret_count, event_uuid, parent_uuid, is_sidechain,
-         ts_source, ts_utc, raw_offset, raw_length)
+        const ins = db.prepare(`INSERT INTO events (id, archive_id, source_session_id, seq, line_index, block_index, kind,
+         source_type, role, tool_name, text, secret_count, event_uuid, parent_uuid, is_sidechain,
+         ts_source, ts_utc, raw_offset, raw_length)
        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`);
-        const insDoc = db.prepare(`INSERT INTO search_docs (event_id, archive_id, source_session_id, project_dir, seq, kind, len)
+        const insDoc = db.prepare(`INSERT INTO search_docs (event_id, archive_id, source_session_id, project_dir, seq, kind, len)
        VALUES (?, ?, ?, ?, ?, ?, ?) RETURNING id`);
-        const insTerm = db.prepare(`INSERT INTO search_terms (term) VALUES (?)
+        const insTerm = db.prepare(`INSERT INTO search_terms (term) VALUES (?)
        ON CONFLICT(term) DO UPDATE SET term = term RETURNING id`);
         const insPosting = db.prepare(`INSERT INTO search_postings (term_id, doc_id, tf) VALUES (?, ?, ?)`);
         // Term ids are cached per ingest run — a session reuses its vocabulary
@@ -86,7 +86,7 @@ export function ingestArchive(archiveId) {
             }
             // Vectors were just dropped with their docs; clearing the marker lets
             // the lazy embed pass rebuild them without re-parsing the transcript.
-            db.prepare(`UPDATE archives SET indexed_through_seq = ?, parser_version = ?, chunker_version = 0, embed_model = NULL
+            db.prepare(`UPDATE archives SET indexed_through_seq = ?, parser_version = ?, chunker_version = 0, embed_model = NULL
           WHERE id = ?`).run(events.length, PARSER_VERSION, archiveId);
             db.exec('COMMIT');
             resetCaches();

package/dist/transcripts/mapping.js

@@ -17,11 +17,11 @@ export function recordSessionMapping(m) {
     const db = getTranscriptsDb();
     const now = new Date().toISOString();
     const proj = m.projectDir ? normalizeProjectDir(m.projectDir) : null;
-    db.prepare(`INSERT INTO session_map (source, source_session_id, transcript_path, project_dir, last_seen_at)
-     VALUES (?, ?, ?, ?, ?)
-     ON CONFLICT(source, source_session_id) DO UPDATE SET
-       transcript_path = excluded.transcript_path,
-       project_dir     = excluded.project_dir,
+    db.prepare(`INSERT INTO session_map (source, source_session_id, transcript_path, project_dir, last_seen_at)
+     VALUES (?, ?, ?, ?, ?)
+     ON CONFLICT(source, source_session_id) DO UPDATE SET
+       transcript_path = excluded.transcript_path,
+       project_dir     = excluded.project_dir,
        last_seen_at    = excluded.last_seen_at`).run(m.source ?? 'claude', m.sourceSessionId, m.transcriptPath, proj, now);
 }
 /** Look up the mapping for an exact session id. */

package/dist/transcripts/pyramid.js

@@ -15,8 +15,8 @@ export const ERROR_RE = /\b(error|errno|failed|failure|exception|traceback|not f
 /** L2 — the conversation spine (user + assistant text, in order). */
 export function buildSpine(archiveId) {
     const db = getTranscriptsDb();
-    const rows = db.prepare(`SELECT seq, role, kind, ts_utc, text FROM events
-     WHERE archive_id = ? AND kind IN ('user_message','assistant_message')
+    const rows = db.prepare(`SELECT seq, role, kind, ts_utc, text FROM events
+     WHERE archive_id = ? AND kind IN ('user_message','assistant_message')
      ORDER BY seq`).all(archiveId);
     return rows.map(r => ({
         seq: r.seq,

package/dist/transcripts/schema.js

@@ -17,39 +17,39 @@ export const MIGRATIONS = [
         // v1 — the two foundational anchor tables. Events / segments / FTS get their
         // own later migrations (Steps 6/9/10) so we don't big-bang the schema.
         version: 1,
-        up: `
-      -- L0 archive index: one row per captured source-CLI session.
-      CREATE TABLE IF NOT EXISTS archives (
-        id                  TEXT PRIMARY KEY,
-        source              TEXT NOT NULL DEFAULT 'claude',   -- claude|codex|gemini
-        source_session_id   TEXT NOT NULL,
-        project_dir         TEXT,                              -- normalized (lowercase drive)
-        veto_session_id     TEXT,                              -- link to veto.db sessions.id when known
-        archive_path        TEXT NOT NULL,                     -- absolute path to the .gz on disk
-        content_sha256      TEXT NOT NULL,                     -- of raw source bytes (dedup key)
-        source_bytes        INTEGER NOT NULL DEFAULT 0,        -- uncompressed size
-        archive_bytes       INTEGER NOT NULL DEFAULT 0,        -- gz size on disk
-        source_format_hint  TEXT,                              -- e.g. 'claude-jsonl'
-        parser_version      INTEGER NOT NULL DEFAULT 0,
-        indexed_through_seq INTEGER NOT NULL DEFAULT 0,        -- lazy-index watermark
-        captured_at         TEXT NOT NULL,
-        updated_at          TEXT NOT NULL,
-        UNIQUE(source, source_session_id)
-      );
-      CREATE INDEX IF NOT EXISTS idx_archives_project ON archives(project_dir);
-      CREATE INDEX IF NOT EXISTS idx_archives_sha     ON archives(content_sha256);
-
-      -- Live host-session -> transcript-file mapping, UPSERTed fire-and-forget by
-      -- the statusline (Step 4) so save-time capture knows which file to read.
-      CREATE TABLE IF NOT EXISTS session_map (
-        source            TEXT NOT NULL DEFAULT 'claude',
-        source_session_id TEXT NOT NULL,
-        transcript_path   TEXT NOT NULL,
-        project_dir       TEXT,
-        last_seen_at      TEXT NOT NULL,
-        PRIMARY KEY (source, source_session_id)
-      );
-      CREATE INDEX IF NOT EXISTS idx_session_map_project ON session_map(project_dir);
+        up: `
+      -- L0 archive index: one row per captured source-CLI session.
+      CREATE TABLE IF NOT EXISTS archives (
+        id                  TEXT PRIMARY KEY,
+        source              TEXT NOT NULL DEFAULT 'claude',   -- claude|codex|gemini
+        source_session_id   TEXT NOT NULL,
+        project_dir         TEXT,                              -- normalized (lowercase drive)
+        veto_session_id     TEXT,                              -- link to veto.db sessions.id when known
+        archive_path        TEXT NOT NULL,                     -- absolute path to the .gz on disk
+        content_sha256      TEXT NOT NULL,                     -- of raw source bytes (dedup key)
+        source_bytes        INTEGER NOT NULL DEFAULT 0,        -- uncompressed size
+        archive_bytes       INTEGER NOT NULL DEFAULT 0,        -- gz size on disk
+        source_format_hint  TEXT,                              -- e.g. 'claude-jsonl'
+        parser_version      INTEGER NOT NULL DEFAULT 0,
+        indexed_through_seq INTEGER NOT NULL DEFAULT 0,        -- lazy-index watermark
+        captured_at         TEXT NOT NULL,
+        updated_at          TEXT NOT NULL,
+        UNIQUE(source, source_session_id)
+      );
+      CREATE INDEX IF NOT EXISTS idx_archives_project ON archives(project_dir);
+      CREATE INDEX IF NOT EXISTS idx_archives_sha     ON archives(content_sha256);
+
+      -- Live host-session -> transcript-file mapping, UPSERTed fire-and-forget by
+      -- the statusline (Step 4) so save-time capture knows which file to read.
+      CREATE TABLE IF NOT EXISTS session_map (
+        source            TEXT NOT NULL DEFAULT 'claude',
+        source_session_id TEXT NOT NULL,
+        transcript_path   TEXT NOT NULL,
+        project_dir       TEXT,
+        last_seen_at      TEXT NOT NULL,
+        PRIMARY KEY (source, source_session_id)
+      );
+      CREATE INDEX IF NOT EXISTS idx_session_map_project ON session_map(project_dir);
     `,
     },
     {
@@ -58,33 +58,33 @@ export const MIGRATIONS = [
         // UNCOMPRESSED L0 bytes for exact expansion (Step 11). Unknown/unparsed lines
         // are stored+ordered (kind='unknown') so format drift never loses data.
         version: 2,
-        up: `
-      CREATE TABLE IF NOT EXISTS events (
-        id                TEXT PRIMARY KEY,
-        archive_id        TEXT NOT NULL,
-        source_session_id TEXT NOT NULL,
-        seq               INTEGER NOT NULL,        -- monotonic per archive, per event
-        line_index        INTEGER NOT NULL,        -- source JSONL line (0-based)
-        block_index       INTEGER NOT NULL DEFAULT 0,
-        kind              TEXT NOT NULL,           -- EVENT_KINDS
-        source_type       TEXT,                    -- verbatim source 'type' (user/assistant/system/…)
-        role              TEXT,
-        tool_name         TEXT,                    -- for tool_call/tool_result
-        text              TEXT,                    -- normalized + MASKED content for FTS
-        secret_count      INTEGER NOT NULL DEFAULT 0,
-        event_uuid        TEXT,                    -- native uuid (Claude)
-        parent_uuid       TEXT,                    -- native parentUuid (tree)
-        is_sidechain      INTEGER NOT NULL DEFAULT 0,
-        ts_source         TEXT,                    -- verbatim timestamp string
-        ts_utc            TEXT,                    -- normalized ISO
-        raw_offset        INTEGER NOT NULL DEFAULT 0,
-        raw_length        INTEGER NOT NULL DEFAULT 0,
-        UNIQUE(archive_id, seq),
-        FOREIGN KEY (archive_id) REFERENCES archives(id)
-      );
-      CREATE INDEX IF NOT EXISTS idx_events_archive ON events(archive_id, seq);
-      CREATE INDEX IF NOT EXISTS idx_events_session ON events(source_session_id);
-      CREATE INDEX IF NOT EXISTS idx_events_kind    ON events(kind);
+        up: `
+      CREATE TABLE IF NOT EXISTS events (
+        id                TEXT PRIMARY KEY,
+        archive_id        TEXT NOT NULL,
+        source_session_id TEXT NOT NULL,
+        seq               INTEGER NOT NULL,        -- monotonic per archive, per event
+        line_index        INTEGER NOT NULL,        -- source JSONL line (0-based)
+        block_index       INTEGER NOT NULL DEFAULT 0,
+        kind              TEXT NOT NULL,           -- EVENT_KINDS
+        source_type       TEXT,                    -- verbatim source 'type' (user/assistant/system/…)
+        role              TEXT,
+        tool_name         TEXT,                    -- for tool_call/tool_result
+        text              TEXT,                    -- normalized + MASKED content for FTS
+        secret_count      INTEGER NOT NULL DEFAULT 0,
+        event_uuid        TEXT,                    -- native uuid (Claude)
+        parent_uuid       TEXT,                    -- native parentUuid (tree)
+        is_sidechain      INTEGER NOT NULL DEFAULT 0,
+        ts_source         TEXT,                    -- verbatim timestamp string
+        ts_utc            TEXT,                    -- normalized ISO
+        raw_offset        INTEGER NOT NULL DEFAULT 0,
+        raw_length        INTEGER NOT NULL DEFAULT 0,
+        UNIQUE(archive_id, seq),
+        FOREIGN KEY (archive_id) REFERENCES archives(id)
+      );
+      CREATE INDEX IF NOT EXISTS idx_events_archive ON events(archive_id, seq);
+      CREATE INDEX IF NOT EXISTS idx_events_session ON events(source_session_id);
+      CREATE INDEX IF NOT EXISTS idx_events_kind    ON events(kind);
     `,
     },
     {
@@ -104,39 +104,39 @@ export const MIGRATIONS = [
         // loads its module, which throws on exactly the Nodes this redo supports.
         // Databases that ran the old dev-only v3 keep an inert orphan table.
         version: 3,
-        up: `
-      -- One row per indexed event. project_dir/session denormalized so the
-      -- hot query is a single join (same shape events_fts had). The INTEGER id
-      -- exists so postings never carry a 36-char uuid — measured, that uuid
-      -- (stored twice per posting: PK + purge index) was ~3x the entire
-      -- index's justified size.
-      CREATE TABLE IF NOT EXISTS search_docs (
-        id                INTEGER PRIMARY KEY,
-        event_id          TEXT NOT NULL UNIQUE,
-        archive_id        TEXT NOT NULL,
-        source_session_id TEXT NOT NULL,
-        project_dir       TEXT,
-        seq               INTEGER NOT NULL,
-        kind              TEXT NOT NULL,
-        len               INTEGER NOT NULL   -- token count, for BM25 length norm
-      );
-      CREATE INDEX IF NOT EXISTS idx_search_docs_archive ON search_docs(archive_id);
-      CREATE INDEX IF NOT EXISTS idx_search_docs_scope
-        ON search_docs(project_dir, source_session_id);
-
-      -- Term dictionary: term text stored once, compact integer join key.
-      CREATE TABLE IF NOT EXISTS search_terms (
-        id   INTEGER PRIMARY KEY,
-        term TEXT NOT NULL UNIQUE
-      );
-
-      CREATE TABLE IF NOT EXISTS search_postings (
-        term_id INTEGER NOT NULL,
-        doc_id  INTEGER NOT NULL,
-        tf      INTEGER NOT NULL,
-        PRIMARY KEY (term_id, doc_id)
-      ) WITHOUT ROWID;
-      CREATE INDEX IF NOT EXISTS idx_postings_doc ON search_postings(doc_id);
+        up: `
+      -- One row per indexed event. project_dir/session denormalized so the
+      -- hot query is a single join (same shape events_fts had). The INTEGER id
+      -- exists so postings never carry a 36-char uuid — measured, that uuid
+      -- (stored twice per posting: PK + purge index) was ~3x the entire
+      -- index's justified size.
+      CREATE TABLE IF NOT EXISTS search_docs (
+        id                INTEGER PRIMARY KEY,
+        event_id          TEXT NOT NULL UNIQUE,
+        archive_id        TEXT NOT NULL,
+        source_session_id TEXT NOT NULL,
+        project_dir       TEXT,
+        seq               INTEGER NOT NULL,
+        kind              TEXT NOT NULL,
+        len               INTEGER NOT NULL   -- token count, for BM25 length norm
+      );
+      CREATE INDEX IF NOT EXISTS idx_search_docs_archive ON search_docs(archive_id);
+      CREATE INDEX IF NOT EXISTS idx_search_docs_scope
+        ON search_docs(project_dir, source_session_id);
+
+      -- Term dictionary: term text stored once, compact integer join key.
+      CREATE TABLE IF NOT EXISTS search_terms (
+        id   INTEGER PRIMARY KEY,
+        term TEXT NOT NULL UNIQUE
+      );
+
+      CREATE TABLE IF NOT EXISTS search_postings (
+        term_id INTEGER NOT NULL,
+        doc_id  INTEGER NOT NULL,
+        tf      INTEGER NOT NULL,
+        PRIMARY KEY (term_id, doc_id)
+      ) WITHOUT ROWID;
+      CREATE INDEX IF NOT EXISTS idx_postings_doc ON search_postings(doc_id);
     `,
     },
     {
@@ -158,17 +158,17 @@ export const MIGRATIONS = [
         // parser_version, so ingest never depends on the model being installed and
         // installing it later backfills without re-parsing anything.
         version: 4,
-        up: `
-      CREATE TABLE IF NOT EXISTS search_vectors (
-        doc_id      INTEGER NOT NULL,
-        chunk_index INTEGER NOT NULL,
-        vec         BLOB NOT NULL,     -- int8, one byte per dimension
-        norm        REAL NOT NULL,     -- L2 norm of the int8 vector
-        PRIMARY KEY (doc_id, chunk_index)
-      );
-
-      ALTER TABLE archives ADD COLUMN chunker_version INTEGER NOT NULL DEFAULT 0;
-      ALTER TABLE archives ADD COLUMN embed_model TEXT;
+        up: `
+      CREATE TABLE IF NOT EXISTS search_vectors (
+        doc_id      INTEGER NOT NULL,
+        chunk_index INTEGER NOT NULL,
+        vec         BLOB NOT NULL,     -- int8, one byte per dimension
+        norm        REAL NOT NULL,     -- L2 norm of the int8 vector
+        PRIMARY KEY (doc_id, chunk_index)
+      );
+
+      ALTER TABLE archives ADD COLUMN chunker_version INTEGER NOT NULL DEFAULT 0;
+      ALTER TABLE archives ADD COLUMN embed_model TEXT;
     `,
     },
 ];

package/dist/transcripts/search.js

@@ -119,8 +119,8 @@ export function searchEvents(query, opts = {}) {
     // Metadata and text for the returned rows only.
     const docIds = top.map(([id]) => id);
     const marks2 = docIds.map(() => '?').join(', ');
-    const metaRows = db.prepare(`SELECT d.id, d.event_id, d.archive_id, d.source_session_id, d.seq, d.kind, e.text
-       FROM search_docs d JOIN events e ON e.id = d.event_id
+    const metaRows = db.prepare(`SELECT d.id, d.event_id, d.archive_id, d.source_session_id, d.seq, d.kind, e.text
+       FROM search_docs d JOIN events e ON e.id = d.event_id
       WHERE d.id IN (${marks2})`).all(...docIds);
     const metaById = new Map(metaRows.map(r => [r.id, r]));
     const termSet = new Set(kept.map(t => t.term));
@@ -261,7 +261,7 @@ export function searchEventsHybrid(query, opts = {}) {
     const db = getTranscriptsDb();
     // Resolve the semantic hits' doc ids to event metadata in one round trip.
     const docIds = semantic.map(s => s.docId);
-    const docRows = db.prepare(`SELECT id, event_id, archive_id, source_session_id, seq, kind
+    const docRows = db.prepare(`SELECT id, event_id, archive_id, source_session_id, seq, kind
        FROM search_docs WHERE id IN (${docIds.map(() => '?').join(', ')})`).all(...docIds);
     const docById = new Map(docRows.map(r => [r.id, r]));
     const fused = new Map();

package/dist/transcripts/vectors.js

@@ -68,8 +68,8 @@ export function embedArchive(archiveId) {
             return { status: 'already_embedded' };
         // Only indexed, searchable docs get vectors — the same set BM25 covers,
         // read with their masked text (C3: masked text is the only text there is).
-        const docs = db.prepare(`SELECT d.id AS doc_id, e.text AS text
-         FROM search_docs d JOIN events e ON e.id = d.event_id
+        const docs = db.prepare(`SELECT d.id AS doc_id, e.text AS text
+         FROM search_docs d JOIN events e ON e.id = d.event_id
         WHERE d.archive_id = ?`).all(archiveId);
         const insVec = db.prepare(`INSERT INTO search_vectors (doc_id, chunk_index, vec, norm) VALUES (?, ?, ?, ?)`);
         let chunks = 0;