fix: checkpoint WAL on live per-agent DB after publishMemoryDatabaseTables

Karl (via Dazlarus) · Karl (via Dazlarus) · commit ee66b808412a · 2026-07-03T22:21:59.000-05:00
The WAL checkpoint was inside writeMeta() which runs against the
shadow/temp database during full reindex. The meta row gets copied
to the live per-agent DB by publishMemoryDatabaseTables, so the
checkpoint must happen on originalDb (the live DB) after publish
returns — not on the shadow DB where writeMeta wrote it.

This ensures the meta row is durable on disk before the process
moves on. Without it, a crash between publish and close can leave
meta only in the WAL file, causing the next startup to declare the
index missing.

Updates test comments to reflect the new checkpoint location.
diff --git a/extensions/memory-core/src/memory/manager-sync-ops.ts b/extensions/memory-core/src/memory/manager-sync-ops.ts
@@ -2793,6 +2793,18 @@ export abstract class MemoryManagerSyncOps {
         vectorExtensionPath: this.vector.extensionPath,
       });
 
+      // Force WAL checkpoint on the live per-agent DB so the newly
+      // published meta row survives process crashes. Without this, the
+      // meta may live only in the WAL file; if the process is killed
+      // before closeMemoryDatabase() checkpoints, the next startup reads
+      // no meta and declares the index missing.
+      try {
+        originalDb.exec("PRAGMA wal_checkpoint(TRUNCATE)");
+      } catch {
+        // Non-fatal: checkpoint may fail in read-only or edge cases.
+        // The normal close path will still attempt a checkpoint.
+      }
+
       this.db = originalDb;
       this.resetVectorState();
       this.fts.available = nextFtsState.available;
@@ -2859,16 +2871,6 @@ export abstract class MemoryManagerSyncOps {
         `INSERT INTO memory_index_meta (key, value) VALUES (?, ?) ON CONFLICT(key) DO UPDATE SET value=excluded.value`,
       )
       .run(META_KEY, value);
-    // Force WAL checkpoint so meta survives process crashes.
-    // Without this, the meta row may live only in the WAL file; if the
-    // process is killed before closeMemoryDatabase() checkpoints, the
-    // next startup reads no meta and declares the index missing.
-    try {
-      this.db.exec("PRAGMA wal_checkpoint(TRUNCATE)");
-    } catch {
-      // Non-fatal: checkpoint may fail in read-only or edge cases.
-      // The normal close path will still attempt a checkpoint.
-    }
     this.lastMetaSerialized = value;
   }
 }
diff --git a/extensions/memory-core/src/memory/manager.wal-checkpoint.test.ts b/extensions/memory-core/src/memory/manager.wal-checkpoint.test.ts
@@ -112,10 +112,11 @@ describe("writeMeta WAL checkpoint", () => {
   }
 
   it("writeMeta calls PRAGMA wal_checkpoint(TRUNCATE) during sync", async () => {
-    // Proves the fix: writeMeta itself forces a WAL checkpoint after writing
-    // the meta row, independent of closeMemoryDatabase. On unpatched main,
-    // the only checkpoint happens during close — meta can be lost if the
-    // process is killed before close.
+    // Proves the fix: after publishMemoryDatabaseTables copies the shadow
+    // DB contents (including the meta row) to the live per-agent DB, a
+    // WAL checkpoint is forced on the live DB — not on the shadow DB.
+    // On unpatched main, the checkpoint was inside writeMeta on the shadow
+    // DB and did nothing useful for crash durability of the live index.
     //
     // This test exercises the real per-agent DB path resolved through
     // resolveOpenClawAgentSqlitePath (not a legacy store.path override).
@@ -132,8 +133,9 @@ describe("writeMeta WAL checkpoint", () => {
       ([sql]) => typeof sql === "string" && sql === "PRAGMA wal_checkpoint(TRUNCATE)",
     );
 
-    // At least one checkpoint fired during sync (from writeMeta).
-    // Close-time checkpoint may add more, but the sync-time call is the fix.
+    // At least one checkpoint fired during sync (from the post-publish
+    // checkpoint on the live per-agent DB). Close-time checkpoint may add
+    // more, but the sync-time call is the fix.
     expect(checkpointCalls.length).toBeGreaterThan(0);
 
     execSpy.mockRestore();
