@deeplake/hivemind 0.6.47 → 0.7.4

package/codex/bundle/capture.js

@@ -1,4 +1,56 @@
 #!/usr/bin/env node
+var __defProp = Object.defineProperty;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __esm = (fn, res) => function __init() {
+  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+};
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+
+// dist/src/index-marker-store.js
+var index_marker_store_exports = {};
+__export(index_marker_store_exports, {
+  buildIndexMarkerPath: () => buildIndexMarkerPath,
+  getIndexMarkerDir: () => getIndexMarkerDir,
+  hasFreshIndexMarker: () => hasFreshIndexMarker,
+  writeIndexMarker: () => writeIndexMarker
+});
+import { existsSync as existsSync2, mkdirSync, readFileSync as readFileSync2, writeFileSync } from "node:fs";
+import { join as join3 } from "node:path";
+import { tmpdir } from "node:os";
+function getIndexMarkerDir() {
+  return process.env.HIVEMIND_INDEX_MARKER_DIR ?? join3(tmpdir(), "hivemind-deeplake-indexes");
+}
+function buildIndexMarkerPath(workspaceId, orgId, table, suffix) {
+  const markerKey = [workspaceId, orgId, table, suffix].join("__").replace(/[^a-zA-Z0-9_.-]/g, "_");
+  return join3(getIndexMarkerDir(), `${markerKey}.json`);
+}
+function hasFreshIndexMarker(markerPath) {
+  if (!existsSync2(markerPath))
+    return false;
+  try {
+    const raw = JSON.parse(readFileSync2(markerPath, "utf-8"));
+    const updatedAt = raw.updatedAt ? new Date(raw.updatedAt).getTime() : NaN;
+    if (!Number.isFinite(updatedAt) || Date.now() - updatedAt > INDEX_MARKER_TTL_MS)
+      return false;
+    return true;
+  } catch {
+    return false;
+  }
+}
+function writeIndexMarker(markerPath) {
+  mkdirSync(getIndexMarkerDir(), { recursive: true });
+  writeFileSync(markerPath, JSON.stringify({ updatedAt: (/* @__PURE__ */ new Date()).toISOString() }), "utf-8");
+}
+var INDEX_MARKER_TTL_MS;
+var init_index_marker_store = __esm({
+  "dist/src/index-marker-store.js"() {
+    "use strict";
+    INDEX_MARKER_TTL_MS = Number(process.env.HIVEMIND_INDEX_MARKER_TTL_MS ?? 6 * 60 * 6e4);
+  }
+});
 
 // dist/src/utils/stdin.js
 function readStdin() {
@@ -51,9 +103,6 @@ function loadConfig() {
 
 // dist/src/deeplake-api.js
 import { randomUUID } from "node:crypto";
-import { existsSync as existsSync2, mkdirSync, readFileSync as readFileSync2, writeFileSync } from "node:fs";
-import { join as join3 } from "node:path";
-import { tmpdir } from "node:os";
 
 // dist/src/utils/debug.js
 import { appendFileSync } from "node:fs";
@@ -76,7 +125,26 @@ function sqlStr(value) {
   return value.replace(/\\/g, "\\\\").replace(/'/g, "''").replace(/\0/g, "").replace(/[\x01-\x08\x0b\x0c\x0e-\x1f\x7f]/g, "");
 }
 
+// dist/src/embeddings/columns.js
+var SUMMARY_EMBEDDING_COL = "summary_embedding";
+var MESSAGE_EMBEDDING_COL = "message_embedding";
+
+// dist/src/utils/client-header.js
+var DEEPLAKE_CLIENT_HEADER = "X-Deeplake-Client";
+function deeplakeClientValue() {
+  return "hivemind";
+}
+function deeplakeClientHeader() {
+  return { [DEEPLAKE_CLIENT_HEADER]: deeplakeClientValue() };
+}
+
 // dist/src/deeplake-api.js
+var indexMarkerStorePromise = null;
+function getIndexMarkerStore() {
+  if (!indexMarkerStorePromise)
+    indexMarkerStorePromise = Promise.resolve().then(() => (init_index_marker_store(), index_marker_store_exports));
+  return indexMarkerStorePromise;
+}
 var log2 = (msg) => log("sdk", msg);
 function summarizeSql(sql, maxLen = 220) {
   const compact = sql.replace(/\s+/g, " ").trim();
@@ -96,7 +164,6 @@ var MAX_RETRIES = 3;
 var BASE_DELAY_MS = 500;
 var MAX_CONCURRENCY = 5;
 var QUERY_TIMEOUT_MS = Number(process.env.HIVEMIND_QUERY_TIMEOUT_MS ?? 1e4);
-var INDEX_MARKER_TTL_MS = Number(process.env.HIVEMIND_INDEX_MARKER_TTL_MS ?? 6 * 60 * 6e4);
 function sleep(ms) {
   return new Promise((resolve) => setTimeout(resolve, ms));
 }
@@ -116,9 +183,6 @@ function isTransientHtml403(text) {
   const body = text.toLowerCase();
   return body.includes("<html") || body.includes("403 forbidden") || body.includes("cloudflare") || body.includes("nginx");
 }
-function getIndexMarkerDir() {
-  return process.env.HIVEMIND_INDEX_MARKER_DIR ?? join3(tmpdir(), "hivemind-deeplake-indexes");
-}
 var Semaphore = class {
   max;
   waiting = [];
@@ -187,7 +251,8 @@ var DeeplakeApi = class {
           headers: {
             Authorization: `Bearer ${this.token}`,
             "Content-Type": "application/json",
-            "X-Activeloop-Org-Id": this.orgId
+            "X-Activeloop-Org-Id": this.orgId,
+            ...deeplakeClientHeader()
           },
           signal,
           body: JSON.stringify({ query: sql })
@@ -214,7 +279,8 @@ var DeeplakeApi = class {
       }
       const text = await resp.text().catch(() => "");
       const retryable403 = isSessionInsertQuery(sql) && (resp.status === 401 || resp.status === 403 && (text.length === 0 || isTransientHtml403(text)));
-      if (attempt < MAX_RETRIES && (RETRYABLE_CODES.has(resp.status) || retryable403)) {
+      const alreadyExists = resp.status === 500 && isDuplicateIndexError(text);
+      if (!alreadyExists && attempt < MAX_RETRIES && (RETRYABLE_CODES.has(resp.status) || retryable403)) {
         const delay = BASE_DELAY_MS * Math.pow(2, attempt) + Math.random() * 200;
         log2(`query retry ${attempt + 1}/${MAX_RETRIES} (${resp.status}) in ${delay.toFixed(0)}ms`);
         await sleep(delay);
@@ -248,7 +314,7 @@ var DeeplakeApi = class {
     const lud = row.lastUpdateDate ?? ts;
     const exists = await this.query(`SELECT path FROM "${this.tableName}" WHERE path = '${sqlStr(row.path)}' LIMIT 1`);
     if (exists.length > 0) {
-      let setClauses = `summary = E'${sqlStr(row.contentText)}', mime_type = '${sqlStr(row.mimeType)}', size_bytes = ${row.sizeBytes}, last_update_date = '${lud}'`;
+      let setClauses = `summary = E'${sqlStr(row.contentText)}', ${SUMMARY_EMBEDDING_COL} = NULL, mime_type = '${sqlStr(row.mimeType)}', size_bytes = ${row.sizeBytes}, last_update_date = '${lud}'`;
       if (row.project !== void 0)
         setClauses += `, project = '${sqlStr(row.project)}'`;
       if (row.description !== void 0)
@@ -256,8 +322,8 @@ var DeeplakeApi = class {
       await this.query(`UPDATE "${this.tableName}" SET ${setClauses} WHERE path = '${sqlStr(row.path)}'`);
     } else {
       const id = randomUUID();
-      let cols = "id, path, filename, summary, mime_type, size_bytes, creation_date, last_update_date";
-      let vals = `'${id}', '${sqlStr(row.path)}', '${sqlStr(row.filename)}', E'${sqlStr(row.contentText)}', '${sqlStr(row.mimeType)}', ${row.sizeBytes}, '${cd}', '${lud}'`;
+      let cols = `id, path, filename, summary, ${SUMMARY_EMBEDDING_COL}, mime_type, size_bytes, creation_date, last_update_date`;
+      let vals = `'${id}', '${sqlStr(row.path)}', '${sqlStr(row.filename)}', E'${sqlStr(row.contentText)}', NULL, '${sqlStr(row.mimeType)}', ${row.sizeBytes}, '${cd}', '${lud}'`;
       if (row.project !== void 0) {
         cols += ", project";
         vals += `, '${sqlStr(row.project)}'`;
@@ -282,48 +348,83 @@ var DeeplakeApi = class {
   buildLookupIndexName(table, suffix) {
     return `idx_${table}_${suffix}`.replace(/[^a-zA-Z0-9_]/g, "_");
   }
-  getLookupIndexMarkerPath(table, suffix) {
-    const markerKey = [
-      this.workspaceId,
-      this.orgId,
-      table,
-      suffix
-    ].join("__").replace(/[^a-zA-Z0-9_.-]/g, "_");
-    return join3(getIndexMarkerDir(), `${markerKey}.json`);
-  }
-  hasFreshLookupIndexMarker(table, suffix) {
-    const markerPath = this.getLookupIndexMarkerPath(table, suffix);
-    if (!existsSync2(markerPath))
-      return false;
-    try {
-      const raw = JSON.parse(readFileSync2(markerPath, "utf-8"));
-      const updatedAt = raw.updatedAt ? new Date(raw.updatedAt).getTime() : NaN;
-      if (!Number.isFinite(updatedAt) || Date.now() - updatedAt > INDEX_MARKER_TTL_MS)
-        return false;
-      return true;
-    } catch {
-      return false;
-    }
-  }
-  markLookupIndexReady(table, suffix) {
-    mkdirSync(getIndexMarkerDir(), { recursive: true });
-    writeFileSync(this.getLookupIndexMarkerPath(table, suffix), JSON.stringify({ updatedAt: (/* @__PURE__ */ new Date()).toISOString() }), "utf-8");
-  }
   async ensureLookupIndex(table, suffix, columnsSql) {
-    if (this.hasFreshLookupIndexMarker(table, suffix))
+    const markers = await getIndexMarkerStore();
+    const markerPath = markers.buildIndexMarkerPath(this.workspaceId, this.orgId, table, suffix);
+    if (markers.hasFreshIndexMarker(markerPath))
       return;
     const indexName = this.buildLookupIndexName(table, suffix);
     try {
       await this.query(`CREATE INDEX IF NOT EXISTS "${indexName}" ON "${table}" ${columnsSql}`);
-      this.markLookupIndexReady(table, suffix);
+      markers.writeIndexMarker(markerPath);
     } catch (e) {
       if (isDuplicateIndexError(e)) {
-        this.markLookupIndexReady(table, suffix);
+        markers.writeIndexMarker(markerPath);
         return;
       }
       log2(`index "${indexName}" skipped: ${e.message}`);
     }
   }
+  /**
+   * Ensure a vector column exists on the given table.
+   *
+   * The previous implementation always issued `ALTER TABLE ADD COLUMN IF NOT
+   * EXISTS …` on every SessionStart. On a long-running workspace that's
+   * already migrated, every call returns 500 "Column already exists" — noisy
+   * in the log and a wasted round-trip. Worse, the very first call after the
+   * column is genuinely added triggers Deeplake's post-ALTER `vector::at`
+   * window (~30s) during which subsequent INSERTs fail; minimising the
+   * number of ALTER calls minimises exposure to that window.
+   *
+   * New flow:
+   *   1. Check the local marker file (mirrors ensureLookupIndex). If fresh,
+   *      return — zero network calls.
+   *   2. SELECT 1 FROM information_schema.columns WHERE table_name = T AND
+   *      column_name = C. Read-only, idempotent, can't tickle the post-ALTER
+   *      bug. If the column is present → mark + return.
+   *   3. Only if step 2 says the column is missing, fall back to ALTER ADD
+   *      COLUMN IF NOT EXISTS. Mark on success, also mark if Deeplake reports
+   *      "already exists" (race: another client added it between our SELECT
+   *      and ALTER).
+   *
+   * Marker uses the same dir / TTL as ensureLookupIndex so both schema
+   * caches share an opt-out (HIVEMIND_INDEX_MARKER_DIR) and a TTL knob.
+   */
+  async ensureEmbeddingColumn(table, column) {
+    await this.ensureColumn(table, column, "FLOAT4[]");
+  }
+  /**
+   * Generic marker-gated column migration. Same SELECT-then-ALTER flow as
+   * ensureEmbeddingColumn, parameterized by SQL type so it can patch up any
+   * column that was added to the schema after the table was originally
+   * created. Used today for `summary_embedding`, `message_embedding`, and
+   * the `agent` column (added 2026-04-11) — the latter has no fallback if
+   * a user upgraded over a pre-2026-04-11 table, so every INSERT fails
+   * with `column "agent" does not exist`.
+   */
+  async ensureColumn(table, column, sqlType) {
+    const markers = await getIndexMarkerStore();
+    const markerPath = markers.buildIndexMarkerPath(this.workspaceId, this.orgId, table, `col_${column}`);
+    if (markers.hasFreshIndexMarker(markerPath))
+      return;
+    const colCheck = `SELECT 1 FROM information_schema.columns WHERE table_name = '${sqlStr(table)}' AND column_name = '${sqlStr(column)}' AND table_schema = '${sqlStr(this.workspaceId)}' LIMIT 1`;
+    const rows = await this.query(colCheck);
+    if (rows.length > 0) {
+      markers.writeIndexMarker(markerPath);
+      return;
+    }
+    try {
+      await this.query(`ALTER TABLE "${table}" ADD COLUMN ${column} ${sqlType}`);
+    } catch (e) {
+      const msg = e instanceof Error ? e.message : String(e);
+      if (!/already exists/i.test(msg))
+        throw e;
+      const recheck = await this.query(colCheck);
+      if (recheck.length === 0)
+        throw e;
+    }
+    markers.writeIndexMarker(markerPath);
+  }
   /** List all tables in the workspace (with retry). */
   async listTables(forceRefresh = false) {
     if (!forceRefresh && this._tablesCache)
@@ -339,7 +440,8 @@ var DeeplakeApi = class {
         const resp = await fetch(`${this.apiUrl}/workspaces/${this.workspaceId}/tables`, {
           headers: {
             Authorization: `Bearer ${this.token}`,
-            "X-Activeloop-Org-Id": this.orgId
+            "X-Activeloop-Org-Id": this.orgId,
+            ...deeplakeClientHeader()
           }
         });
         if (resp.ok) {
@@ -364,28 +466,60 @@ var DeeplakeApi = class {
     }
     return { tables: [], cacheable: false };
   }
+  /**
+   * Run a `CREATE TABLE` with an extra outer retry budget. The base
+   * `query()` already retries 3 times on fetch errors (~3.5s total), but a
+   * failed CREATE is permanent corruption — every subsequent SELECT against
+   * the missing table fails. Wrapping in an outer loop with longer backoff
+   * (2s, 5s, then 10s) gives us ~17s of reach across transient network
+   * blips before giving up. Failures still propagate; getApi() resets its
+   * cache on init failure (openclaw plugin) so the next call retries the
+   * whole init flow.
+   */
+  async createTableWithRetry(sql, label) {
+    const OUTER_BACKOFFS_MS = [2e3, 5e3, 1e4];
+    let lastErr = null;
+    for (let attempt = 0; attempt <= OUTER_BACKOFFS_MS.length; attempt++) {
+      try {
+        await this.query(sql);
+        return;
+      } catch (err) {
+        lastErr = err;
+        const msg = err instanceof Error ? err.message : String(err);
+        log2(`CREATE TABLE "${label}" attempt ${attempt + 1}/${OUTER_BACKOFFS_MS.length + 1} failed: ${msg}`);
+        if (attempt < OUTER_BACKOFFS_MS.length) {
+          await sleep(OUTER_BACKOFFS_MS[attempt]);
+        }
+      }
+    }
+    throw lastErr;
+  }
   /** Create the memory table if it doesn't already exist. Migrate columns on existing tables. */
   async ensureTable(name) {
     const tbl = name ?? this.tableName;
     const tables = await this.listTables();
     if (!tables.includes(tbl)) {
       log2(`table "${tbl}" not found, creating`);
-      await this.query(`CREATE TABLE IF NOT EXISTS "${tbl}" (id TEXT NOT NULL DEFAULT '', path TEXT NOT NULL DEFAULT '', filename TEXT NOT NULL DEFAULT '', summary TEXT NOT NULL DEFAULT '', author TEXT NOT NULL DEFAULT '', mime_type TEXT NOT NULL DEFAULT 'text/plain', size_bytes BIGINT NOT NULL DEFAULT 0, project TEXT NOT NULL DEFAULT '', description TEXT NOT NULL DEFAULT '', agent TEXT NOT NULL DEFAULT '', creation_date TEXT NOT NULL DEFAULT '', last_update_date TEXT NOT NULL DEFAULT '') USING deeplake`);
+      await this.createTableWithRetry(`CREATE TABLE IF NOT EXISTS "${tbl}" (id TEXT NOT NULL DEFAULT '', path TEXT NOT NULL DEFAULT '', filename TEXT NOT NULL DEFAULT '', summary TEXT NOT NULL DEFAULT '', summary_embedding FLOAT4[], author TEXT NOT NULL DEFAULT '', mime_type TEXT NOT NULL DEFAULT 'text/plain', size_bytes BIGINT NOT NULL DEFAULT 0, project TEXT NOT NULL DEFAULT '', description TEXT NOT NULL DEFAULT '', agent TEXT NOT NULL DEFAULT '', creation_date TEXT NOT NULL DEFAULT '', last_update_date TEXT NOT NULL DEFAULT '') USING deeplake`, tbl);
       log2(`table "${tbl}" created`);
       if (!tables.includes(tbl))
         this._tablesCache = [...tables, tbl];
     }
+    await this.ensureEmbeddingColumn(tbl, SUMMARY_EMBEDDING_COL);
+    await this.ensureColumn(tbl, "agent", "TEXT NOT NULL DEFAULT ''");
   }
   /** Create the sessions table (uses JSONB for message since every row is a JSON event). */
   async ensureSessionsTable(name) {
     const tables = await this.listTables();
     if (!tables.includes(name)) {
       log2(`table "${name}" not found, creating`);
-      await this.query(`CREATE TABLE IF NOT EXISTS "${name}" (id TEXT NOT NULL DEFAULT '', path TEXT NOT NULL DEFAULT '', filename TEXT NOT NULL DEFAULT '', message JSONB, author TEXT NOT NULL DEFAULT '', mime_type TEXT NOT NULL DEFAULT 'application/json', size_bytes BIGINT NOT NULL DEFAULT 0, project TEXT NOT NULL DEFAULT '', description TEXT NOT NULL DEFAULT '', agent TEXT NOT NULL DEFAULT '', creation_date TEXT NOT NULL DEFAULT '', last_update_date TEXT NOT NULL DEFAULT '') USING deeplake`);
+      await this.createTableWithRetry(`CREATE TABLE IF NOT EXISTS "${name}" (id TEXT NOT NULL DEFAULT '', path TEXT NOT NULL DEFAULT '', filename TEXT NOT NULL DEFAULT '', message JSONB, message_embedding FLOAT4[], author TEXT NOT NULL DEFAULT '', mime_type TEXT NOT NULL DEFAULT 'application/json', size_bytes BIGINT NOT NULL DEFAULT 0, project TEXT NOT NULL DEFAULT '', description TEXT NOT NULL DEFAULT '', agent TEXT NOT NULL DEFAULT '', creation_date TEXT NOT NULL DEFAULT '', last_update_date TEXT NOT NULL DEFAULT '') USING deeplake`, name);
       log2(`table "${name}" created`);
       if (!tables.includes(name))
         this._tablesCache = [...tables, name];
     }
+    await this.ensureEmbeddingColumn(name, MESSAGE_EMBEDDING_COL);
+    await this.ensureColumn(name, "agent", "TEXT NOT NULL DEFAULT ''");
     await this.ensureLookupIndex(name, "path_creation_date", `("path", "creation_date")`);
   }
 };
@@ -396,25 +530,306 @@ function buildSessionPath(config, sessionId) {
   return `/sessions/${config.userName}/${config.userName}_${config.orgName}_${workspace}_${sessionId}.jsonl`;
 }
 
-// dist/src/hooks/summary-state.js
-import { readFileSync as readFileSync3, writeFileSync as writeFileSync2, writeSync, mkdirSync as mkdirSync2, renameSync, existsSync as existsSync3, unlinkSync, openSync, closeSync } from "node:fs";
+// dist/src/embeddings/client.js
+import { connect } from "node:net";
+import { spawn } from "node:child_process";
+import { openSync, closeSync, writeSync, unlinkSync, existsSync as existsSync3, readFileSync as readFileSync3 } from "node:fs";
 import { homedir as homedir3 } from "node:os";
 import { join as join4 } from "node:path";
+
+// dist/src/embeddings/protocol.js
+var DEFAULT_SOCKET_DIR = "/tmp";
+var DEFAULT_IDLE_TIMEOUT_MS = 10 * 60 * 1e3;
+var DEFAULT_CLIENT_TIMEOUT_MS = 2e3;
+function socketPathFor(uid, dir = DEFAULT_SOCKET_DIR) {
+  return `${dir}/hivemind-embed-${uid}.sock`;
+}
+function pidPathFor(uid, dir = DEFAULT_SOCKET_DIR) {
+  return `${dir}/hivemind-embed-${uid}.pid`;
+}
+
+// dist/src/embeddings/client.js
+var SHARED_DAEMON_PATH = join4(homedir3(), ".hivemind", "embed-deps", "embed-daemon.js");
+var log3 = (m) => log("embed-client", m);
+function getUid() {
+  const uid = typeof process.getuid === "function" ? process.getuid() : void 0;
+  return uid !== void 0 ? String(uid) : process.env.USER ?? "default";
+}
+var EmbedClient = class {
+  socketPath;
+  pidPath;
+  timeoutMs;
+  daemonEntry;
+  autoSpawn;
+  spawnWaitMs;
+  nextId = 0;
+  constructor(opts = {}) {
+    const uid = getUid();
+    const dir = opts.socketDir ?? "/tmp";
+    this.socketPath = socketPathFor(uid, dir);
+    this.pidPath = pidPathFor(uid, dir);
+    this.timeoutMs = opts.timeoutMs ?? DEFAULT_CLIENT_TIMEOUT_MS;
+    this.daemonEntry = opts.daemonEntry ?? process.env.HIVEMIND_EMBED_DAEMON ?? (existsSync3(SHARED_DAEMON_PATH) ? SHARED_DAEMON_PATH : void 0);
+    this.autoSpawn = opts.autoSpawn ?? true;
+    this.spawnWaitMs = opts.spawnWaitMs ?? 5e3;
+  }
+  /**
+   * Returns an embedding vector, or null on timeout/failure. Hooks MUST treat
+   * null as "skip embedding column" — never block the write path on us.
+   *
+   * Fire-and-forget spawn on miss: if the daemon isn't up, this call returns
+   * null AND kicks off a background spawn. The next call finds a ready daemon.
+   */
+  async embed(text, kind = "document") {
+    let sock;
+    try {
+      sock = await this.connectOnce();
+    } catch {
+      if (this.autoSpawn)
+        this.trySpawnDaemon();
+      return null;
+    }
+    try {
+      const id = String(++this.nextId);
+      const req = { op: "embed", id, kind, text };
+      const resp = await this.sendAndWait(sock, req);
+      if (resp.error || !("embedding" in resp) || !resp.embedding) {
+        log3(`embed err: ${resp.error ?? "no embedding"}`);
+        return null;
+      }
+      return resp.embedding;
+    } catch (e) {
+      const err = e instanceof Error ? e.message : String(e);
+      log3(`embed failed: ${err}`);
+      return null;
+    } finally {
+      try {
+        sock.end();
+      } catch {
+      }
+    }
+  }
+  /**
+   * Wait up to spawnWaitMs for the daemon to accept connections, spawning if
+   * necessary. Meant for SessionStart / long-running batches — not the hot path.
+   */
+  async warmup() {
+    try {
+      const s = await this.connectOnce();
+      s.end();
+      return true;
+    } catch {
+      if (!this.autoSpawn)
+        return false;
+      this.trySpawnDaemon();
+      try {
+        const s = await this.waitForSocket();
+        s.end();
+        return true;
+      } catch {
+        return false;
+      }
+    }
+  }
+  connectOnce() {
+    return new Promise((resolve, reject) => {
+      const sock = connect(this.socketPath);
+      const to = setTimeout(() => {
+        sock.destroy();
+        reject(new Error("connect timeout"));
+      }, this.timeoutMs);
+      sock.once("connect", () => {
+        clearTimeout(to);
+        resolve(sock);
+      });
+      sock.once("error", (e) => {
+        clearTimeout(to);
+        reject(e);
+      });
+    });
+  }
+  trySpawnDaemon() {
+    let fd;
+    try {
+      fd = openSync(this.pidPath, "wx", 384);
+      writeSync(fd, String(process.pid));
+    } catch (e) {
+      if (this.isPidFileStale()) {
+        try {
+          unlinkSync(this.pidPath);
+        } catch {
+        }
+        try {
+          fd = openSync(this.pidPath, "wx", 384);
+          writeSync(fd, String(process.pid));
+        } catch {
+          return;
+        }
+      } else {
+        return;
+      }
+    }
+    if (!this.daemonEntry || !existsSync3(this.daemonEntry)) {
+      log3(`daemonEntry not configured or missing: ${this.daemonEntry}`);
+      try {
+        closeSync(fd);
+        unlinkSync(this.pidPath);
+      } catch {
+      }
+      return;
+    }
+    try {
+      const child = spawn(process.execPath, [this.daemonEntry], {
+        detached: true,
+        stdio: "ignore",
+        env: process.env
+      });
+      child.unref();
+      log3(`spawned daemon pid=${child.pid}`);
+    } finally {
+      closeSync(fd);
+    }
+  }
+  isPidFileStale() {
+    try {
+      const raw = readFileSync3(this.pidPath, "utf-8").trim();
+      const pid = Number(raw);
+      if (!pid || Number.isNaN(pid))
+        return true;
+      try {
+        process.kill(pid, 0);
+        return false;
+      } catch {
+        return true;
+      }
+    } catch {
+      return true;
+    }
+  }
+  async waitForSocket() {
+    const deadline = Date.now() + this.spawnWaitMs;
+    let delay = 30;
+    while (Date.now() < deadline) {
+      await sleep2(delay);
+      delay = Math.min(delay * 1.5, 300);
+      if (!existsSync3(this.socketPath))
+        continue;
+      try {
+        return await this.connectOnce();
+      } catch {
+      }
+    }
+    throw new Error("daemon did not become ready within spawnWaitMs");
+  }
+  sendAndWait(sock, req) {
+    return new Promise((resolve, reject) => {
+      let buf = "";
+      const to = setTimeout(() => {
+        sock.destroy();
+        reject(new Error("request timeout"));
+      }, this.timeoutMs);
+      sock.setEncoding("utf-8");
+      sock.on("data", (chunk) => {
+        buf += chunk;
+        const nl = buf.indexOf("\n");
+        if (nl === -1)
+          return;
+        const line = buf.slice(0, nl);
+        clearTimeout(to);
+        try {
+          resolve(JSON.parse(line));
+        } catch (e) {
+          reject(e);
+        }
+      });
+      sock.on("error", (e) => {
+        clearTimeout(to);
+        reject(e);
+      });
+      sock.on("end", () => {
+        clearTimeout(to);
+        reject(new Error("connection closed without response"));
+      });
+      sock.write(JSON.stringify(req) + "\n");
+    });
+  }
+};
+function sleep2(ms) {
+  return new Promise((r) => setTimeout(r, ms));
+}
+
+// dist/src/embeddings/sql.js
+function embeddingSqlLiteral(vec) {
+  if (!vec || vec.length === 0)
+    return "NULL";
+  const parts = [];
+  for (const v of vec) {
+    if (!Number.isFinite(v))
+      return "NULL";
+    parts.push(String(v));
+  }
+  return `ARRAY[${parts.join(",")}]::float4[]`;
+}
+
+// dist/src/embeddings/disable.js
+import { createRequire } from "node:module";
+import { homedir as homedir4 } from "node:os";
+import { join as join5 } from "node:path";
+import { pathToFileURL } from "node:url";
+var cachedStatus = null;
+function defaultResolveTransformers() {
+  try {
+    createRequire(import.meta.url).resolve("@huggingface/transformers");
+    return;
+  } catch {
+  }
+  const sharedDir = join5(homedir4(), ".hivemind", "embed-deps");
+  createRequire(pathToFileURL(`${sharedDir}/`).href).resolve("@huggingface/transformers");
+}
+var _resolve = defaultResolveTransformers;
+function detectStatus() {
+  if (process.env.HIVEMIND_EMBEDDINGS === "false")
+    return "env-disabled";
+  try {
+    _resolve();
+    return "enabled";
+  } catch {
+    return "no-transformers";
+  }
+}
+function embeddingsStatus() {
+  if (cachedStatus !== null)
+    return cachedStatus;
+  cachedStatus = detectStatus();
+  return cachedStatus;
+}
+function embeddingsDisabled() {
+  return embeddingsStatus() !== "enabled";
+}
+
+// dist/src/hooks/codex/capture.js
+import { fileURLToPath as fileURLToPath2 } from "node:url";
+import { dirname as dirname2, join as join9 } from "node:path";
+
+// dist/src/hooks/summary-state.js
+import { readFileSync as readFileSync4, writeFileSync as writeFileSync2, writeSync as writeSync2, mkdirSync as mkdirSync2, renameSync, existsSync as existsSync4, unlinkSync as unlinkSync2, openSync as openSync2, closeSync as closeSync2 } from "node:fs";
+import { homedir as homedir5 } from "node:os";
+import { join as join6 } from "node:path";
 var dlog = (msg) => log("summary-state", msg);
-var STATE_DIR = join4(homedir3(), ".claude", "hooks", "summary-state");
+var STATE_DIR = join6(homedir5(), ".claude", "hooks", "summary-state");
 var YIELD_BUF = new Int32Array(new SharedArrayBuffer(4));
 function statePath(sessionId) {
-  return join4(STATE_DIR, `${sessionId}.json`);
+  return join6(STATE_DIR, `${sessionId}.json`);
 }
 function lockPath(sessionId) {
-  return join4(STATE_DIR, `${sessionId}.lock`);
+  return join6(STATE_DIR, `${sessionId}.lock`);
 }
 function readState(sessionId) {
   const p = statePath(sessionId);
-  if (!existsSync3(p))
+  if (!existsSync4(p))
     return null;
   try {
-    return JSON.parse(readFileSync3(p, "utf-8"));
+    return JSON.parse(readFileSync4(p, "utf-8"));
   } catch {
     return null;
   }
@@ -433,14 +848,14 @@ function withRmwLock(sessionId, fn) {
   let fd = null;
   while (fd === null) {
     try {
-      fd = openSync(rmwLock, "wx");
+      fd = openSync2(rmwLock, "wx");
     } catch (e) {
       if (e.code !== "EEXIST")
         throw e;
       if (Date.now() > deadline) {
         dlog(`rmw lock deadline exceeded for ${sessionId}, reclaiming stale lock`);
         try {
-          unlinkSync(rmwLock);
+          unlinkSync2(rmwLock);
         } catch (unlinkErr) {
           dlog(`stale rmw lock unlink failed for ${sessionId}: ${unlinkErr.message}`);
         }
@@ -452,9 +867,9 @@ function withRmwLock(sessionId, fn) {
   try {
     return fn();
   } finally {
-    closeSync(fd);
+    closeSync2(fd);
     try {
-      unlinkSync(rmwLock);
+      unlinkSync2(rmwLock);
     } catch (unlinkErr) {
       dlog(`rmw lock cleanup failed for ${sessionId}: ${unlinkErr.message}`);
     }
@@ -491,27 +906,27 @@ function shouldTrigger(state, cfg, now = Date.now()) {
 function tryAcquireLock(sessionId, maxAgeMs = 10 * 60 * 1e3) {
   mkdirSync2(STATE_DIR, { recursive: true });
   const p = lockPath(sessionId);
-  if (existsSync3(p)) {
+  if (existsSync4(p)) {
     try {
-      const ageMs = Date.now() - parseInt(readFileSync3(p, "utf-8"), 10);
+      const ageMs = Date.now() - parseInt(readFileSync4(p, "utf-8"), 10);
       if (Number.isFinite(ageMs) && ageMs < maxAgeMs)
         return false;
     } catch (readErr) {
       dlog(`lock file unreadable for ${sessionId}, treating as stale: ${readErr.message}`);
     }
     try {
-      unlinkSync(p);
+      unlinkSync2(p);
     } catch (unlinkErr) {
       dlog(`could not unlink stale lock for ${sessionId}: ${unlinkErr.message}`);
       return false;
     }
   }
   try {
-    const fd = openSync(p, "wx");
+    const fd = openSync2(p, "wx");
     try {
-      writeSync(fd, String(Date.now()));
+      writeSync2(fd, String(Date.now()));
     } finally {
-      closeSync(fd);
+      closeSync2(fd);
     }
     return true;
   } catch (e) {
@@ -522,7 +937,7 @@ function tryAcquireLock(sessionId, maxAgeMs = 10 * 60 * 1e3) {
 }
 function releaseLock(sessionId) {
   try {
-    unlinkSync(lockPath(sessionId));
+    unlinkSync2(lockPath(sessionId));
   } catch (e) {
     if (e?.code !== "ENOENT") {
       dlog(`releaseLock unlink failed for ${sessionId}: ${e.message}`);
@@ -531,17 +946,17 @@ function releaseLock(sessionId) {
 }
 
 // dist/src/hooks/codex/spawn-wiki-worker.js
-import { spawn, execSync } from "node:child_process";
+import { spawn as spawn2, execSync } from "node:child_process";
 import { fileURLToPath } from "node:url";
-import { dirname, join as join6 } from "node:path";
+import { dirname, join as join8 } from "node:path";
 import { writeFileSync as writeFileSync3, mkdirSync as mkdirSync4 } from "node:fs";
-import { homedir as homedir4, tmpdir as tmpdir2 } from "node:os";
+import { homedir as homedir6, tmpdir as tmpdir2 } from "node:os";
 
 // dist/src/utils/wiki-log.js
 import { mkdirSync as mkdirSync3, appendFileSync as appendFileSync2 } from "node:fs";
-import { join as join5 } from "node:path";
+import { join as join7 } from "node:path";
 function makeWikiLogger(hooksDir, filename = "deeplake-wiki.log") {
-  const path = join5(hooksDir, filename);
+  const path = join7(hooksDir, filename);
   return {
     path,
     log(msg) {
@@ -556,8 +971,8 @@ function makeWikiLogger(hooksDir, filename = "deeplake-wiki.log") {
 }
 
 // dist/src/hooks/codex/spawn-wiki-worker.js
-var HOME = homedir4();
-var wikiLogger = makeWikiLogger(join6(HOME, ".codex", "hooks"));
+var HOME = homedir6();
+var wikiLogger = makeWikiLogger(join8(HOME, ".codex", "hooks"));
 var WIKI_LOG = wikiLogger.path;
 var WIKI_PROMPT_TEMPLATE = `You are building a personal wiki from a coding session. Your goal is to extract every piece of knowledge \u2014 entities, decisions, relationships, and facts \u2014 into a structured, searchable wiki entry.
 
@@ -619,9 +1034,9 @@ function findCodexBin() {
 function spawnCodexWikiWorker(opts) {
   const { config, sessionId, cwd, bundleDir, reason } = opts;
   const projectName = cwd.split("/").pop() || "unknown";
-  const tmpDir = join6(tmpdir2(), `deeplake-wiki-${sessionId}-${Date.now()}`);
+  const tmpDir = join8(tmpdir2(), `deeplake-wiki-${sessionId}-${Date.now()}`);
   mkdirSync4(tmpDir, { recursive: true });
-  const configFile = join6(tmpDir, "config.json");
+  const configFile = join8(tmpDir, "config.json");
   writeFileSync3(configFile, JSON.stringify({
     apiUrl: config.apiUrl,
     token: config.token,
@@ -635,12 +1050,12 @@ function spawnCodexWikiWorker(opts) {
     tmpDir,
     codexBin: findCodexBin(),
     wikiLog: WIKI_LOG,
-    hooksDir: join6(HOME, ".codex", "hooks"),
+    hooksDir: join8(HOME, ".codex", "hooks"),
     promptTemplate: WIKI_PROMPT_TEMPLATE
   }));
   wikiLog(`${reason}: spawning summary worker for ${sessionId}`);
-  const workerPath = join6(bundleDir, "wiki-worker.js");
-  spawn("nohup", ["node", workerPath, configFile], {
+  const workerPath = join8(bundleDir, "wiki-worker.js");
+  spawn2("nohup", ["node", workerPath, configFile], {
     detached: true,
     stdio: ["ignore", "ignore", "ignore"]
   }).unref();
@@ -651,7 +1066,10 @@ function bundleDirFromImportMeta(importMetaUrl) {
 }
 
 // dist/src/hooks/codex/capture.js
-var log3 = (msg) => log("codex-capture", msg);
+var log4 = (msg) => log("codex-capture", msg);
+function resolveEmbedDaemonPath() {
+  return join9(dirname2(fileURLToPath2(import.meta.url)), "embeddings", "embed-daemon.js");
+}
 var CAPTURE = process.env.HIVEMIND_CAPTURE !== "false";
 async function main() {
   if (!CAPTURE)
@@ -659,7 +1077,7 @@ async function main() {
   const input = await readStdin();
   const config = loadConfig();
   if (!config) {
-    log3("no config");
+    log4("no config");
     return;
   }
   const sessionsTable = config.sessionsTableName;
@@ -676,7 +1094,7 @@ async function main() {
   };
   let entry;
   if (input.hook_event_name === "UserPromptSubmit" && input.prompt !== void 0) {
-    log3(`user session=${input.session_id}`);
+    log4(`user session=${input.session_id}`);
     entry = {
       id: crypto.randomUUID(),
       ...meta,
@@ -684,7 +1102,7 @@ async function main() {
       content: input.prompt
     };
   } else if (input.hook_event_name === "PostToolUse" && input.tool_name !== void 0) {
-    log3(`tool=${input.tool_name} session=${input.session_id}`);
+    log4(`tool=${input.tool_name} session=${input.session_id}`);
     entry = {
       id: crypto.randomUUID(),
       ...meta,
@@ -695,28 +1113,30 @@ async function main() {
       tool_response: JSON.stringify(input.tool_response)
     };
   } else {
-    log3(`unknown event: ${input.hook_event_name}, skipping`);
+    log4(`unknown event: ${input.hook_event_name}, skipping`);
     return;
   }
   const sessionPath = buildSessionPath(config, input.session_id);
   const line = JSON.stringify(entry);
-  log3(`writing to ${sessionPath}`);
+  log4(`writing to ${sessionPath}`);
   const projectName = (input.cwd ?? "").split("/").pop() || "unknown";
   const filename = sessionPath.split("/").pop() ?? "";
-  const jsonForSql = sqlStr(line);
-  const insertSql = `INSERT INTO "${sessionsTable}" (id, path, filename, message, author, size_bytes, project, description, agent, creation_date, last_update_date) VALUES ('${crypto.randomUUID()}', '${sqlStr(sessionPath)}', '${sqlStr(filename)}', '${jsonForSql}'::jsonb, '${sqlStr(config.userName)}', ${Buffer.byteLength(line, "utf-8")}, '${sqlStr(projectName)}', '${sqlStr(input.hook_event_name ?? "")}', 'codex', '${ts}', '${ts}')`;
+  const jsonForSql = line.replace(/'/g, "''");
+  const embedding = embeddingsDisabled() ? null : await new EmbedClient({ daemonEntry: resolveEmbedDaemonPath() }).embed(line, "document");
+  const embeddingSql = embeddingSqlLiteral(embedding);
+  const insertSql = `INSERT INTO "${sessionsTable}" (id, path, filename, message, message_embedding, author, size_bytes, project, description, agent, creation_date, last_update_date) VALUES ('${crypto.randomUUID()}', '${sqlStr(sessionPath)}', '${sqlStr(filename)}', '${jsonForSql}'::jsonb, ${embeddingSql}, '${sqlStr(config.userName)}', ${Buffer.byteLength(line, "utf-8")}, '${sqlStr(projectName)}', '${sqlStr(input.hook_event_name ?? "")}', 'codex', '${ts}', '${ts}')`;
   try {
     await api.query(insertSql);
   } catch (e) {
     if (e.message?.includes("permission denied") || e.message?.includes("does not exist")) {
-      log3("table missing, creating and retrying");
+      log4("table missing, creating and retrying");
       await api.ensureSessionsTable(sessionsTable);
       await api.query(insertSql);
     } else {
       throw e;
     }
   }
-  log3("capture ok");
+  log4("capture ok");
   maybeTriggerPeriodicSummary(input.session_id, input.cwd ?? "", config);
 }
 function maybeTriggerPeriodicSummary(sessionId, cwd, config) {
@@ -728,7 +1148,7 @@ function maybeTriggerPeriodicSummary(sessionId, cwd, config) {
     if (!shouldTrigger(state, cfg))
       return;
     if (!tryAcquireLock(sessionId)) {
-      log3(`periodic trigger suppressed (lock held) session=${sessionId}`);
+      log4(`periodic trigger suppressed (lock held) session=${sessionId}`);
       return;
     }
     wikiLog(`Periodic: threshold hit (total=${state.totalCount}, since=${state.totalCount - state.lastSummaryCount}, N=${cfg.everyNMessages}, hours=${cfg.everyHours})`);
@@ -741,19 +1161,19 @@ function maybeTriggerPeriodicSummary(sessionId, cwd, config) {
         reason: "Periodic"
       });
     } catch (e) {
-      log3(`periodic spawn failed: ${e.message}`);
+      log4(`periodic spawn failed: ${e.message}`);
       try {
         releaseLock(sessionId);
       } catch (releaseErr) {
-        log3(`releaseLock after periodic spawn failure also failed: ${releaseErr.message}`);
+        log4(`releaseLock after periodic spawn failure also failed: ${releaseErr.message}`);
       }
       throw e;
     }
   } catch (e) {
-    log3(`periodic trigger error: ${e.message}`);
+    log4(`periodic trigger error: ${e.message}`);
   }
 }
 main().catch((e) => {
-  log3(`fatal: ${e.message}`);
+  log4(`fatal: ${e.message}`);
   process.exit(0);
 });