Fix SQLite WAL corruption: checkpoint before export, copy WAL/SHM files, preserve FTS schema via iterdump

sync.sh

@@ -13,25 +13,46 @@ echo "[$(date '+%H:%M:%S')] Sync from $HOSTNAME..."
 # with remote when pulling. Remove it before fetch+merge.
 rm -f "$SYNC_DIR/state_merged.db"
 
-# ── Step 1: Export local state.db (via temp dir to avoid lock) ────────────
+# ── Step 1: Export local state.db (via temp dir to avoid lock & WAL issues) ─
-python3 << PYEOF
+python3 << 'PYEOF'
 import sqlite3, os, shutil, tempfile
 
 local_db = os.path.join(os.path.expanduser('~/.hermes'), 'state.db')
 export_db = os.path.join(os.path.expanduser('~/.hermes-sync'), f"state_{os.environ.get('HOSTNAME') or __import__('socket').gethostname()}.db")
+hostname = os.environ.get('HOSTNAME') or __import__('socket').gethostname()
 tmpdir = tempfile.mkdtemp(prefix='hs_exp_')
 
 try:
-    tmp_db = os.path.join(tmpdir, 'db')
+    # Step 1a: Open with WAL mode and checkpoint (forces WAL -> main db)
-    shutil.copy2(local_db, tmp_db)
+    conn = sqlite3.connect(f'file:{local_db}?mode=ro', uri=True)
+    jm = conn.execute('PRAGMA journal_mode').fetchone()[0]
 
-    test = sqlite3.connect(tmp_db)
+    if jm == 'wal':
+        # WAL mode: checkpoint first to flush WAL into main db
+        conn.execute('PRAGMA wal_checkpoint(TRUNCATE)')
+        print(f'WAL checkpointed ({jm})')
+    conn.close()
+
+    # Step 1b: Copy all three files (db + -wal + -shm) if they exist
+    tmp_db = os.path.join(tmpdir, 'db')
+    for suffix in ['', '-wal', '-shm']:
+        src = local_db + suffix
+        if os.path.exists(src):
+            shutil.copy2(src, tmp_db + suffix)
+
+    # Step 1c: Verify (open read-only to confirm it's not corrupted)
+    test = sqlite3.connect(f'file:{tmp_db}?mode=ro', uri=True)
     r = test.execute('SELECT COUNT(*) FROM sessions').fetchone()[0]
     m = test.execute('SELECT COUNT(*) FROM messages').fetchone()[0]
     test.close()
 
     shutil.copy2(tmp_db, export_db)
-    print(f'Exported: {r}s/{m}m')
+    if os.path.exists(tmp_db + '-wal'):
+        shutil.copy2(tmp_db + '-wal', export_db + '-wal')
+    if os.path.exists(tmp_db + '-shm'):
+        shutil.copy2(tmp_db + '-shm', export_db + '-shm')
+
+    print(f'Exported: {r}s/{m}m (journal={jm})')
 finally:
     shutil.rmtree(tmpdir, ignore_errors=True)
 PYEOF
@@ -85,53 +106,54 @@ tmpdir = tempfile.mkdtemp(prefix='hs_merge_')
 tmp_merged = os.path.join(tmpdir, 'merged.db')
 
 try:
-    # Create merged DB with proper schema
+    # Create merged DB (DELETE journal avoids WAL complications during merge)
     conn = sqlite3.connect(tmp_merged)
     conn.execute('PRAGMA journal_mode=DELETE')
     conn.execute('PRAGMA locking_mode=NORMAL')
     conn.execute('PRAGMA synchronous=FULL')
 
-    conn.execute('''
+    # Use one reference DB to get exact schema (handles FTS, etc.)
-    CREATE TABLE sessions (
+    ref_db = db_files[0]
-        id TEXT PRIMARY KEY, source TEXT NOT NULL, user_id TEXT,
+    ref_copy = os.path.join(tmpdir, 'ref.db')
-        model TEXT, model_config TEXT, system_prompt TEXT,
+    shutil.copy2(ref_db, ref_copy)
-        parent_session_id TEXT, started_at REAL, ended_at REAL,
+    if os.path.exists(ref_db + '-wal'):
-        end_reason TEXT, message_count INTEGER DEFAULT 0,
+        shutil.copy2(ref_db + '-wal', ref_copy + '-wal')
-        tool_call_count INTEGER DEFAULT 0, input_tokens INTEGER DEFAULT 0,
+    if os.path.exists(ref_db + '-shm'):
-        output_tokens INTEGER DEFAULT 0, cache_read_tokens INTEGER DEFAULT 0,
+        shutil.copy2(ref_db + '-shm', ref_copy + '-shm')
-        cache_write_tokens INTEGER DEFAULT 0, reasoning_tokens INTEGER DEFAULT 0,
-        billing_provider TEXT, billing_base_url TEXT, billing_mode TEXT,
-        estimated_cost_usd REAL, actual_cost_usd REAL,
-        cost_status TEXT, cost_source TEXT, pricing_version TEXT, title TEXT
-    )''')
 
-    conn.execute('''
+    ref = sqlite3.connect(f'file:{ref_copy}?mode=ro', uri=True)
-    CREATE TABLE messages (
+    for line in ref.iterdump():
-        id INTEGER PRIMARY KEY,
+        # Skip FTS virtual table data rows (they are derived, not real data)
-        session_id TEXT NOT NULL, role TEXT NOT NULL, content TEXT,
+        if line.startswith('INSERT INTO messages_fts') or \
-        tool_call_id TEXT, tool_calls TEXT, tool_name TEXT,
+           line.startswith('DELETE FROM messages_fts'):
-        timestamp REAL NOT NULL, token_count INTEGER,
+            continue
-        finish_reason TEXT, reasoning TEXT, reasoning_details TEXT,
+        try:
-        codex_reasoning_items TEXT
+            conn.execute(line)
-    )''')
+        except Exception as e:
-
+            pass  # Ignore schema errors from partial DDL
-    conn.execute('CREATE INDEX idx_msg_session ON messages(session_id)')
+    ref.close()
-    conn.execute('CREATE INDEX idx_msg_ts ON messages(timestamp)')
+    os.remove(ref_copy)
+    if os.path.exists(ref_copy + '-wal'):
+        os.remove(ref_copy + '-wal')
 
+    # Copy sessions (INSERT OR REPLACE to dedupe by PK)
     for db_file in db_files:
         name = os.path.basename(db_file)
-        # Copy to temp file first (avoids WAL lock issues with open connections)
+
+        # Copy to temp with WAL files
         tmp_copy = os.path.join(tmpdir, name)
         shutil.copy2(db_file, tmp_copy)
-        # Also copy WAL if exists
         if os.path.exists(db_file + '-wal'):
             shutil.copy2(db_file + '-wal', tmp_copy + '-wal')
+        if os.path.exists(db_file + '-shm'):
+            shutil.copy2(db_file + '-shm', tmp_copy + '-shm')
 
-        src = sqlite3.connect(tmp_copy)
+        src = sqlite3.connect(f'file:{tmp_copy}?mode=ro', uri=True)
 
         s_cnt = src.execute('SELECT COUNT(*) FROM sessions').fetchone()[0]
         m_cnt = src.execute('SELECT COUNT(*) FROM messages').fetchone()[0]
-        print(f'  {name}: {s_cnt}s/{m_cnt}m')
+        jm = src.execute('PRAGMA journal_mode').fetchone()[0]
+        print(f'  {name}: {s_cnt}s/{m_cnt}m journal={jm}')
 
         sess_rows = src.execute('SELECT * FROM sessions').fetchall()
         sess_cols = len(src.execute('PRAGMA table_info(sessions)').fetchall())
@@ -145,8 +167,9 @@ try:
 
         src.close()
         os.remove(tmp_copy)
-        if os.path.exists(tmp_copy + '-wal'):
+        for suf in ['-wal', '-shm']:
-            os.remove(tmp_copy + '-wal')
+            if os.path.exists(tmp_copy + suf):
+                os.remove(tmp_copy + suf)
 
     conn.commit()
     conn.close()
@@ -154,6 +177,10 @@ try:
     if os.path.exists(merged_path):
         os.remove(merged_path)
     shutil.copy2(tmp_merged, merged_path)
+    # Ensure no WAL on merged (merge output should be clean DELETE)
+    for suf in ['-wal', '-shm']:
+        if os.path.exists(merged_path + suf):
+            os.remove(merged_path + suf)
     print(f'Merged: {os.path.getsize(merged_path)/1024:.0f} KB')
 
 finally:
@@ -188,16 +215,36 @@ if not os.path.exists(merged_path):
 else:
     tmpdir = tempfile.mkdtemp(prefix='hs_rest_')
     try:
+        # Step 7a: Ensure merged db is fully checkpointed (DELETE journal mode)
+        merge_conn = sqlite3.connect(f'file:{merged_path}?mode=ro', uri=True)
+        merge_jm = merge_conn.execute('PRAGMA journal_mode').fetchone()[0]
+        if merge_jm == 'wal':
+            merge_conn.execute('PRAGMA wal_checkpoint(TRUNCATE)')
+            print(f'Restored merged DB had WAL, checkpointed ({merge_jm})')
+        merge_conn.close()
+
         tmp_db = os.path.join(tmpdir, 'db')
         shutil.copy2(merged_path, tmp_db)
+        # Also copy -wal/-shm if present
+        for suf in ['-wal', '-shm']:
+            if os.path.exists(merged_path + suf):
+                shutil.copy2(merged_path + suf, tmp_db + suf)
 
-        test = sqlite3.connect(tmp_db)
+        # Step 7b: Verify before overwriting
+        test = sqlite3.connect(f'file:{tmp_db}?mode=ro', uri=True)
         r = test.execute('SELECT COUNT(*) FROM sessions').fetchone()[0]
         m = test.execute('SELECT COUNT(*) FROM messages').fetchone()[0]
+        test.execute('PRAGMA integrity_check')  # verify
         test.close()
 
         shutil.copy2(local_db, local_db + '.bak')
         shutil.copy2(tmp_db, local_db)
+        # Also copy WAL files to local
+        for suf in ['-wal', '-shm']:
+            if os.path.exists(tmp_db + suf):
+                shutil.copy2(tmp_db + suf, local_db + suf)
+            elif os.path.exists(local_db + suf):
+                os.remove(local_db + suf)
         os.remove(local_db + '.bak')
         print(f'Restored: {r}s/{m}m')
     finally: