@deeplake/hivemind 0.6.48 → 0.7.4

package/cursor/bundle/session-end.js

@@ -23,6 +23,9 @@ import { join } from "node:path";
 import { homedir } from "node:os";
 var DEBUG = process.env.HIVEMIND_DEBUG === "1";
 var LOG = join(homedir(), ".deeplake", "hook-debug.log");
+function utcTimestamp(d = /* @__PURE__ */ new Date()) {
+  return d.toISOString().replace("T", " ").slice(0, 19) + " UTC";
+}
 function log(tag, msg) {
   if (!DEBUG)
     return;
@@ -30,14 +33,232 @@ function log(tag, msg) {
 `);
 }
 
+// dist/src/config.js
+import { readFileSync, existsSync } from "node:fs";
+import { join as join2 } from "node:path";
+import { homedir as homedir2, userInfo } from "node:os";
+function loadConfig() {
+  const home = homedir2();
+  const credPath = join2(home, ".deeplake", "credentials.json");
+  let creds = null;
+  if (existsSync(credPath)) {
+    try {
+      creds = JSON.parse(readFileSync(credPath, "utf-8"));
+    } catch {
+      return null;
+    }
+  }
+  const token = process.env.HIVEMIND_TOKEN ?? creds?.token;
+  const orgId = process.env.HIVEMIND_ORG_ID ?? creds?.orgId;
+  if (!token || !orgId)
+    return null;
+  return {
+    token,
+    orgId,
+    orgName: creds?.orgName ?? orgId,
+    userName: creds?.userName || userInfo().username || "unknown",
+    workspaceId: process.env.HIVEMIND_WORKSPACE_ID ?? creds?.workspaceId ?? "default",
+    apiUrl: process.env.HIVEMIND_API_URL ?? creds?.apiUrl ?? "https://api.deeplake.ai",
+    tableName: process.env.HIVEMIND_TABLE ?? "memory",
+    sessionsTableName: process.env.HIVEMIND_SESSIONS_TABLE ?? "sessions",
+    memoryPath: process.env.HIVEMIND_MEMORY_PATH ?? join2(home, ".deeplake", "memory")
+  };
+}
+
+// dist/src/hooks/summary-state.js
+import { readFileSync as readFileSync2, writeFileSync, writeSync, mkdirSync, renameSync, existsSync as existsSync2, unlinkSync, openSync, closeSync } from "node:fs";
+import { homedir as homedir3 } from "node:os";
+import { join as join3 } from "node:path";
+var dlog = (msg) => log("summary-state", msg);
+var STATE_DIR = join3(homedir3(), ".claude", "hooks", "summary-state");
+var YIELD_BUF = new Int32Array(new SharedArrayBuffer(4));
+function lockPath(sessionId) {
+  return join3(STATE_DIR, `${sessionId}.lock`);
+}
+function tryAcquireLock(sessionId, maxAgeMs = 10 * 60 * 1e3) {
+  mkdirSync(STATE_DIR, { recursive: true });
+  const p = lockPath(sessionId);
+  if (existsSync2(p)) {
+    try {
+      const ageMs = Date.now() - parseInt(readFileSync2(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);
+    } catch (unlinkErr) {
+      dlog(`could not unlink stale lock for ${sessionId}: ${unlinkErr.message}`);
+      return false;
+    }
+  }
+  try {
+    const fd = openSync(p, "wx");
+    try {
+      writeSync(fd, String(Date.now()));
+    } finally {
+      closeSync(fd);
+    }
+    return true;
+  } catch (e) {
+    if (e.code === "EEXIST")
+      return false;
+    throw e;
+  }
+}
+
+// dist/src/hooks/cursor/spawn-wiki-worker.js
+import { spawn, execSync } from "node:child_process";
+import { fileURLToPath } from "node:url";
+import { dirname, join as join5 } from "node:path";
+import { writeFileSync as writeFileSync2, mkdirSync as mkdirSync3 } from "node:fs";
+import { homedir as homedir4, tmpdir } from "node:os";
+
+// dist/src/utils/wiki-log.js
+import { mkdirSync as mkdirSync2, appendFileSync as appendFileSync2 } from "node:fs";
+import { join as join4 } from "node:path";
+function makeWikiLogger(hooksDir, filename = "deeplake-wiki.log") {
+  const path = join4(hooksDir, filename);
+  return {
+    path,
+    log(msg) {
+      try {
+        mkdirSync2(hooksDir, { recursive: true });
+        appendFileSync2(path, `[${utcTimestamp()}] ${msg}
+`);
+      } catch {
+      }
+    }
+  };
+}
+
+// dist/src/hooks/cursor/spawn-wiki-worker.js
+var HOME = homedir4();
+var wikiLogger = makeWikiLogger(join5(HOME, ".cursor", "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.
+
+SESSION JSONL path: __JSONL__
+SUMMARY FILE to write: __SUMMARY__
+SESSION ID: __SESSION_ID__
+PROJECT: __PROJECT__
+PREVIOUS JSONL OFFSET (lines already processed): __PREV_OFFSET__
+CURRENT JSONL LINES: __JSONL_LINES__
+
+Steps:
+1. Read the session JSONL at the path above.
+   - If PREVIOUS JSONL OFFSET > 0, this is a resumed session. Read the existing summary file first,
+     then focus on lines AFTER the offset for new content. Merge new facts into the existing summary.
+   - If offset is 0, generate from scratch.
+
+2. Write the summary file at the path above with this EXACT format:
+
+# Session __SESSION_ID__
+- **Source**: __JSONL_SERVER_PATH__
+- **Started**: <extract from JSONL>
+- **Ended**: <now>
+- **Project**: __PROJECT__
+- **JSONL offset**: __JSONL_LINES__
+
+## What Happened
+<2-3 dense sentences. What was the goal, what was accomplished, what's left.>
+
+## People
+<For each person mentioned: name, role, what they did/said. Format: **Name** \u2014 role \u2014 action>
+
+## Entities
+<Every named thing: repos, branches, files, APIs, tools, services, tables, features, bugs.
+Format: **entity** (type) \u2014 what was done with it, its current state>
+
+## Decisions & Reasoning
+<Every decision made and WHY.>
+
+## Key Facts
+<Bullet list of atomic facts that could answer future questions.>
+
+## Files Modified
+<bullet list: path (new/modified/deleted) \u2014 what changed>
+
+## Open Questions / TODO
+<Anything unresolved, blocked, or explicitly deferred>
+
+IMPORTANT: Be exhaustive. Extract EVERY entity, decision, and fact.
+PRIVACY: Never include absolute filesystem paths in the summary.
+LENGTH LIMIT: Keep the total summary under 4000 characters.`;
+var wikiLog = wikiLogger.log;
+function findCursorBin() {
+  try {
+    return execSync("which cursor-agent 2>/dev/null", { encoding: "utf-8" }).trim() || "cursor-agent";
+  } catch {
+    return "cursor-agent";
+  }
+}
+function spawnCursorWikiWorker(opts) {
+  const { config, sessionId, cwd, bundleDir, reason } = opts;
+  const projectName = cwd.split("/").pop() || "unknown";
+  const tmpDir = join5(tmpdir(), `deeplake-wiki-${sessionId}-${Date.now()}`);
+  mkdirSync3(tmpDir, { recursive: true });
+  const configFile = join5(tmpDir, "config.json");
+  writeFileSync2(configFile, JSON.stringify({
+    apiUrl: config.apiUrl,
+    token: config.token,
+    orgId: config.orgId,
+    workspaceId: config.workspaceId,
+    memoryTable: config.tableName,
+    sessionsTable: config.sessionsTableName,
+    sessionId,
+    userName: config.userName,
+    project: projectName,
+    tmpDir,
+    cursorBin: findCursorBin(),
+    cursorModel: process.env.HIVEMIND_CURSOR_MODEL ?? "auto",
+    wikiLog: WIKI_LOG,
+    hooksDir: join5(HOME, ".cursor", "hooks"),
+    promptTemplate: WIKI_PROMPT_TEMPLATE
+  }));
+  wikiLog(`${reason}: spawning summary worker for ${sessionId}`);
+  const workerPath = join5(bundleDir, "wiki-worker.js");
+  spawn("nohup", ["node", workerPath, configFile], {
+    detached: true,
+    stdio: ["ignore", "ignore", "ignore"]
+  }).unref();
+  wikiLog(`${reason}: spawned summary worker for ${sessionId}`);
+}
+function bundleDirFromImportMeta(importMetaUrl) {
+  return dirname(fileURLToPath(importMetaUrl));
+}
+
 // dist/src/hooks/cursor/session-end.js
 var log2 = (msg) => log("cursor-session-end", msg);
 async function main() {
   if (process.env.HIVEMIND_WIKI_WORKER === "1")
     return;
   const input = await readStdin();
-  const sessionId = input.conversation_id ?? input.session_id ?? "?";
-  log2(`session=${sessionId} reason=${input.reason ?? "?"} status=${input.final_status ?? "?"}`);
+  const sessionId = input.conversation_id ?? input.session_id ?? "";
+  log2(`session=${sessionId || "?"} reason=${input.reason ?? "?"} status=${input.final_status ?? "?"}`);
+  if (!sessionId)
+    return;
+  if (!tryAcquireLock(sessionId)) {
+    wikiLog(`SessionEnd: periodic worker already running for ${sessionId}, skipping final`);
+    return;
+  }
+  try {
+    const config = loadConfig();
+    if (!config) {
+      wikiLog(`SessionEnd: no config, skipping summary`);
+      return;
+    }
+    spawnCursorWikiWorker({
+      config,
+      sessionId,
+      cwd: process.cwd(),
+      bundleDir: bundleDirFromImportMeta(import.meta.url),
+      reason: "SessionEnd"
+    });
+  } catch (e) {
+    wikiLog(`SessionEnd: spawn failed: ${e?.message ?? e}`);
+  }
 }
 main().catch((e) => {
   log2(`fatal: ${e.message}`);

package/cursor/bundle/session-start.js

@@ -1,35 +1,100 @@
 #!/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 as mkdirSync2, readFileSync as readFileSync3, writeFileSync as writeFileSync2 } from "node:fs";
+import { join as join4 } from "node:path";
+import { tmpdir } from "node:os";
+function getIndexMarkerDir() {
+  return process.env.HIVEMIND_INDEX_MARKER_DIR ?? join4(tmpdir(), "hivemind-deeplake-indexes");
+}
+function buildIndexMarkerPath(workspaceId, orgId, table, suffix) {
+  const markerKey = [workspaceId, orgId, table, suffix].join("__").replace(/[^a-zA-Z0-9_.-]/g, "_");
+  return join4(getIndexMarkerDir(), `${markerKey}.json`);
+}
+function hasFreshIndexMarker(markerPath) {
+  if (!existsSync2(markerPath))
+    return false;
+  try {
+    const raw = JSON.parse(readFileSync3(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) {
+  mkdirSync2(getIndexMarkerDir(), { recursive: true });
+  writeFileSync2(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/hooks/cursor/session-start.js
 import { fileURLToPath } from "node:url";
 import { dirname as dirname2, join as join6 } from "node:path";
 
 // dist/src/commands/auth.js
-import { readFileSync, writeFileSync, existsSync, mkdirSync, unlinkSync } from "node:fs";
+import { execSync } from "node:child_process";
+
+// 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/commands/auth-creds.js
+import { readFileSync, writeFileSync, mkdirSync, unlinkSync } from "node:fs";
 import { join } from "node:path";
 import { homedir } from "node:os";
-import { execSync } from "node:child_process";
-var CONFIG_DIR = join(homedir(), ".deeplake");
-var CREDS_PATH = join(CONFIG_DIR, "credentials.json");
+function configDir() {
+  return join(homedir(), ".deeplake");
+}
+function credsPath() {
+  return join(configDir(), "credentials.json");
+}
 function loadCredentials() {
-  if (!existsSync(CREDS_PATH))
-    return null;
   try {
-    return JSON.parse(readFileSync(CREDS_PATH, "utf-8"));
+    return JSON.parse(readFileSync(credsPath(), "utf-8"));
   } catch {
     return null;
   }
 }
 
 // dist/src/config.js
-import { readFileSync as readFileSync2, existsSync as existsSync2 } from "node:fs";
+import { readFileSync as readFileSync2, existsSync } from "node:fs";
 import { join as join2 } from "node:path";
 import { homedir as homedir2, userInfo } from "node:os";
 function loadConfig() {
   const home = homedir2();
   const credPath = join2(home, ".deeplake", "credentials.json");
   let creds = null;
-  if (existsSync2(credPath)) {
+  if (existsSync(credPath)) {
     try {
       creds = JSON.parse(readFileSync2(credPath, "utf-8"));
     } catch {
@@ -55,9 +120,6 @@ function loadConfig() {
 
 // dist/src/deeplake-api.js
 import { randomUUID } from "node:crypto";
-import { existsSync as existsSync3, mkdirSync as mkdirSync2, readFileSync as readFileSync3, writeFileSync as writeFileSync2 } from "node:fs";
-import { join as join4 } from "node:path";
-import { tmpdir } from "node:os";
 
 // dist/src/utils/debug.js
 import { appendFileSync } from "node:fs";
@@ -77,7 +139,17 @@ 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/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();
@@ -97,7 +169,6 @@ 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));
 }
@@ -117,9 +188,6 @@ 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 ?? join4(tmpdir(), "hivemind-deeplake-indexes");
-}
 var Semaphore = class {
   max;
   waiting = [];
@@ -188,7 +256,8 @@ 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 })
@@ -215,7 +284,8 @@ 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 ${delay.toFixed(0)}ms`);
         await sleep(delay);
@@ -249,7 +319,7 @@ var DeeplakeApi = class {
     const lud = row.lastUpdateDate ?? ts;
     const exists = await this.query(`SELECT path FROM "${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 = '${sqlStr(row.project)}'`;
       if (row.description !== void 0)
@@ -257,8 +327,8 @@ var DeeplakeApi = class {
       await this.query(`UPDATE "${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)}'`;
@@ -283,48 +353,83 @@ 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 join4(getIndexMarkerDir(), `${markerKey}.json`);
-  }
-  hasFreshLookupIndexMarker(table, suffix) {
-    const markerPath = this.getLookupIndexMarkerPath(table, suffix);
-    if (!existsSync3(markerPath))
-      return false;
-    try {
-      const raw = JSON.parse(readFileSync3(markerPath, "utf-8"));
-      const updatedAt = raw.updatedAt ? new Date(raw.updatedAt).getTime() : NaN;
-      if (!Number.isFinite(updatedAt) || Date.now() - updatedAt > INDEX_MARKER_TTL_MS)
-        return false;
-      return true;
-    } catch {
-      return false;
-    }
-  }
-  markLookupIndexReady(table, suffix) {
-    mkdirSync2(getIndexMarkerDir(), { recursive: true });
-    writeFileSync2(this.getLookupIndexMarkerPath(table, suffix), JSON.stringify({ updatedAt: (/* @__PURE__ */ new Date()).toISOString() }), "utf-8");
-  }
   async ensureLookupIndex(table, suffix, columnsSql) {
-    if (this.hasFreshLookupIndexMarker(table, suffix))
+    const markers = await getIndexMarkerStore();
+    const markerPath = markers.buildIndexMarkerPath(this.workspaceId, this.orgId, table, suffix);
+    if (markers.hasFreshIndexMarker(markerPath))
       return;
     const indexName = this.buildLookupIndexName(table, suffix);
     try {
       await this.query(`CREATE INDEX IF NOT EXISTS "${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} ${sqlType}`);
+    } catch (e) {
+      const msg = e instanceof Error ? e.message : String(e);
+      if (!/already exists/i.test(msg))
+        throw e;
+      const recheck = await this.query(colCheck);
+      if (recheck.length === 0)
+        throw e;
+    }
+    markers.writeIndexMarker(markerPath);
+  }
   /** List all tables in the workspace (with retry). */
   async listTables(forceRefresh = false) {
     if (!forceRefresh && this._tablesCache)
@@ -340,7 +445,8 @@ var DeeplakeApi = class {
         const resp = await fetch(`${this.apiUrl}/workspaces/${this.workspaceId}/tables`, {
           headers: {
             Authorization: `Bearer ${this.token}`,
-            "X-Activeloop-Org-Id": this.orgId
+            "X-Activeloop-Org-Id": this.orgId,
+            ...deeplakeClientHeader()
           }
         });
         if (resp.ok) {
@@ -365,28 +471,60 @@ var DeeplakeApi = class {
     }
     return { tables: [], cacheable: false };
   }
+  /**
+   * Run a `CREATE TABLE` with an extra outer retry budget. The base
+   * `query()` already retries 3 times on fetch errors (~3.5s total), but a
+   * failed CREATE is permanent corruption — every subsequent SELECT against
+   * the missing table fails. Wrapping in an outer loop with longer backoff
+   * (2s, 5s, then 10s) gives us ~17s of reach across transient network
+   * blips before giving up. Failures still propagate; getApi() resets its
+   * cache on init failure (openclaw plugin) so the next call retries the
+   * whole init flow.
+   */
+  async createTableWithRetry(sql, label) {
+    const OUTER_BACKOFFS_MS = [2e3, 5e3, 1e4];
+    let lastErr = null;
+    for (let attempt = 0; attempt <= OUTER_BACKOFFS_MS.length; attempt++) {
+      try {
+        await this.query(sql);
+        return;
+      } catch (err) {
+        lastErr = err;
+        const msg = err instanceof Error ? err.message : String(err);
+        log2(`CREATE TABLE "${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")`);
   }
 };