diff --git a/src/db.rs b/src/db.rs
index 43659f9..e7b3052 100644
--- a/src/db.rs
+++ b/src/db.rs
@@ -28,7 +28,7 @@ use std::thread;
 
 use rusqlite::Connection;
 use tokio::sync::{mpsc, oneshot};
-use tracing::{debug, info};
+use tracing::{debug, info, warn};
 
 use crate::dsl::action::ReferentialAction;
 use crate::dsl::command::{
@@ -46,6 +46,7 @@ use crate::persistence::{
     SchemaSnapshot, TableSchema, TableSnapshot, decode_cell, parse_csv, parse_schema,
 };
 use crate::project::{DATA_DIR, PROJECT_YAML};
+use crate::undo::{DEFAULT_RING_CAPACITY, SnapshotError, SnapshotMeta, SnapshotStore, Staged};
 
 /// Inbox capacity. The worker is fast enough that this rarely
 /// matters; `64` is a generous head-room for bursts.
@@ -681,6 +682,40 @@ enum Request {
         source: crate::dsl::grammar::IdentSource,
         reply: oneshot::Sender<Result<Vec<String>, DbError>>,
     },
+    /// Restore the most recent undo snapshot (ADR-0006 Amendment 1).
+    /// Replies with the metadata of the command that was undone, or
+    /// `None` if there is nothing to undo (or undo is disabled).
+    Undo {
+        reply: oneshot::Sender<Result<Option<SnapshotMeta>, DbError>>,
+    },
+    /// Re-apply the most recently undone snapshot. `None` if there is
+    /// nothing to redo.
+    Redo {
+        reply: oneshot::Sender<Result<Option<SnapshotMeta>, DbError>>,
+    },
+    /// Read — without restoring — the snapshot `undo` would restore.
+    /// Used to build the confirmation prompt. `None` if the ring is
+    /// empty or undo is disabled.
+    PeekUndo {
+        reply: oneshot::Sender<Result<Option<SnapshotMeta>, DbError>>,
+    },
+    /// Read — without restoring — the snapshot `redo` would restore.
+    PeekRedo {
+        reply: oneshot::Sender<Result<Option<SnapshotMeta>, DbError>>,
+    },
+    /// Open a batch (ADR-0006 Amendment 1): take one boundary
+    /// snapshot for the whole batch and suppress per-command
+    /// snapshots until `EndBatch`. Used by `replay` so a multi-command
+    /// replay is a single undo step. `source` is the batch command.
+    BeginBatch {
+        source: Option<String>,
+        reply: oneshot::Sender<()>,
+    },
+    /// Close a batch: finalise the boundary snapshot into the ring if
+    /// any mutation committed during the batch, else discard it.
+    EndBatch {
+        reply: oneshot::Sender<()>,
+    },
 }
 
 impl Database {
@@ -693,7 +728,7 @@ impl Database {
     /// are skipped — useful for unit tests that exercise the
     /// SQLite layer in isolation.
     pub fn open<P: AsRef<Path>>(path: P) -> Result<Self, DbError> {
-        Self::open_inner(path, None)
+        Self::open_inner(path, None, false)
     }
 
     /// Open a database with per-command persistence wired in
@@ -705,12 +740,24 @@ impl Database {
         path: P,
         persistence: Persistence,
     ) -> Result<Self, DbError> {
-        Self::open_inner(path, Some(persistence))
+        Self::open_inner(path, Some(persistence), false)
+    }
+
+    /// Open with per-command persistence *and* the undo/snapshot ring
+    /// (ADR-0006 Amendment 1). `undo_enabled` is `false` under the
+    /// `--no-undo` CLI flag, in which case no snapshots are taken.
+    pub fn open_with_persistence_and_undo<P: AsRef<Path>>(
+        path: P,
+        persistence: Persistence,
+        undo_enabled: bool,
+    ) -> Result<Self, DbError> {
+        Self::open_inner(path, Some(persistence), undo_enabled)
     }
 
     fn open_inner<P: AsRef<Path>>(
         path: P,
         persistence: Option<Persistence>,
+        undo_enabled: bool,
     ) -> Result<Self, DbError> {
         let path_display = path.as_ref().to_string_lossy().into_owned();
         let conn = match path.as_ref().to_str() {
@@ -721,10 +768,27 @@ impl Database {
         info!(path = %path_display, "opened database");
         configure_connection(&conn).map_err(DbError::from_rusqlite)?;
 
+        // The undo ring needs the project directory; it is only
+        // available when persistence is wired and undo is enabled.
+        let snapshots = if undo_enabled {
+            persistence
+                .as_ref()
+                .map(|p| SnapshotStore::new(p.project_path(), DEFAULT_RING_CAPACITY))
+        } else {
+            None
+        };
+        if let Some(store) = &snapshots {
+            // Sweep crash leftovers (`.staging/`, orphan payloads).
+            if let Err(e) = store.cleanup() {
+                warn!(error = %e, "undo snapshot cleanup on open failed");
+            }
+            info!("undo snapshots enabled");
+        }
+
         let (tx, rx) = mpsc::channel::<Request>(REQUEST_CHANNEL_CAPACITY);
         thread::Builder::new()
             .name("rdbms-db-worker".to_string())
-            .spawn(move || worker_loop(conn, persistence, rx))
+            .spawn(move || worker_loop(conn, persistence, snapshots, rx))
             .map_err(|e| DbError::Io(e.to_string()))?;
 
         Ok(Self { inbox: tx })
@@ -1238,6 +1302,57 @@ impl Database {
         recv.await.map_err(|_| DbError::WorkerGone)?
     }
 
+    /// Restore the most recent undo snapshot (ADR-0006 Amendment 1).
+    /// `Ok(Some(meta))` reports the command that was undone;
+    /// `Ok(None)` means nothing to undo (or undo is disabled).
+    pub async fn undo(&self) -> Result<Option<SnapshotMeta>, DbError> {
+        let (reply, recv) = oneshot::channel();
+        self.send(Request::Undo { reply }).await?;
+        recv.await.map_err(|_| DbError::WorkerGone)?
+    }
+
+    /// Re-apply the most recently undone snapshot. `Ok(None)` means
+    /// nothing to redo.
+    pub async fn redo(&self) -> Result<Option<SnapshotMeta>, DbError> {
+        let (reply, recv) = oneshot::channel();
+        self.send(Request::Redo { reply }).await?;
+        recv.await.map_err(|_| DbError::WorkerGone)?
+    }
+
+    /// Metadata of the snapshot `undo` would restore, without
+    /// restoring it — for the confirmation prompt.
+    pub async fn peek_undo(&self) -> Result<Option<SnapshotMeta>, DbError> {
+        let (reply, recv) = oneshot::channel();
+        self.send(Request::PeekUndo { reply }).await?;
+        recv.await.map_err(|_| DbError::WorkerGone)?
+    }
+
+    /// Metadata of the snapshot `redo` would restore.
+    pub async fn peek_redo(&self) -> Result<Option<SnapshotMeta>, DbError> {
+        let (reply, recv) = oneshot::channel();
+        self.send(Request::PeekRedo { reply }).await?;
+        recv.await.map_err(|_| DbError::WorkerGone)?
+    }
+
+    /// Open a batch so a multi-command operation (`replay`, future
+    /// batch commands) records a single undo step (ADR-0006
+    /// Amendment 1). Pair with [`Database::end_batch`]. `source` is
+    /// the batch command text recorded on the boundary snapshot.
+    pub async fn begin_batch(&self, source: Option<String>) -> Result<(), DbError> {
+        let (reply, recv) = oneshot::channel();
+        self.send(Request::BeginBatch { source, reply }).await?;
+        recv.await.map_err(|_| DbError::WorkerGone)
+    }
+
+    /// Close a batch opened with [`Database::begin_batch`]. The
+    /// boundary snapshot is kept iff a mutation committed during the
+    /// batch.
+    pub async fn end_batch(&self) -> Result<(), DbError> {
+        let (reply, recv) = oneshot::channel();
+        self.send(Request::EndBatch { reply }).await?;
+        recv.await.map_err(|_| DbError::WorkerGone)
+    }
+
     async fn send(&self, req: Request) -> Result<(), DbError> {
         self.inbox.send(req).await.map_err(|_| DbError::WorkerGone)
     }
@@ -1325,16 +1440,192 @@ fn iso8601_now() -> String {
 fn worker_loop(
     conn: Connection,
     persistence: Option<Persistence>,
+    snapshots: Option<SnapshotStore>,
     mut rx: mpsc::Receiver<Request>,
 ) {
     debug!("db worker started");
+    // `conn` must be mutable: restoring a snapshot (undo/redo) writes
+    // into the live connection via the backup API (`&mut`).
+    let mut conn = conn;
+    let snap = snapshots.as_ref();
+    let mut batch = BatchState::default();
     while let Some(req) = rx.blocking_recv() {
-        handle_request(&conn, persistence.as_ref(), req);
+        // Undo/redo/peek/batch are handled here: undo/redo need
+        // `&mut conn` for the restore, and batch state lives across
+        // requests. Everything else goes to `handle_request`, which
+        // brackets mutations with a pre-op snapshot.
+        match req {
+            Request::Undo { reply } => {
+                let _ = reply.send(do_undo(snap, &mut conn));
+            }
+            Request::Redo { reply } => {
+                let _ = reply.send(do_redo(snap, &mut conn));
+            }
+            Request::PeekUndo { reply } => {
+                let _ = reply.send(peek_undo_op(snap));
+            }
+            Request::PeekRedo { reply } => {
+                let _ = reply.send(peek_redo_op(snap));
+            }
+            Request::BeginBatch { source, reply } => {
+                begin_batch(snap, &conn, &mut batch, source.as_deref());
+                let _ = reply.send(());
+            }
+            Request::EndBatch { reply } => {
+                end_batch(snap, &mut batch);
+                let _ = reply.send(());
+            }
+            other => handle_request(&conn, persistence.as_ref(), snap, &mut batch, other),
+        }
     }
     debug!("db worker exiting");
 }
 
-fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Request) {
+/// Worker-side undo bracketing state for the request stream.
+/// `active` is set between `BeginBatch`/`EndBatch` so per-command
+/// snapshots are suppressed in favour of one boundary snapshot for
+/// the whole batch (ADR-0006 Amendment 1).
+#[derive(Default)]
+struct BatchState {
+    active: bool,
+    dirty: bool,
+    staged: Option<Staged>,
+}
+
+fn snapshot_to_db_error(e: SnapshotError) -> DbError {
+    DbError::Io(e.to_string())
+}
+
+/// Stage a pre-mutation snapshot, never failing the user's command if
+/// the snapshot itself can't be taken (the real persistence is the
+/// durable state — the snapshot is a best-effort safety net). Returns
+/// `None` when undo is off, the command has no user `source` (an
+/// internal op, e.g. open-time rebuild — not undoable), or staging
+/// failed.
+fn stage_pre_mutation(
+    snap: Option<&SnapshotStore>,
+    conn: &Connection,
+    source: Option<&str>,
+) -> Option<Staged> {
+    let store = snap?;
+    let src = source?;
+    match store.stage(conn, src) {
+        Ok(staged) => Some(staged),
+        Err(e) => {
+            warn!(error = %e, "could not stage undo snapshot; command proceeds without undo");
+            None
+        }
+    }
+}
+
+/// Run a mutating handler with undo bracketing: stage before, then
+/// finalise on success / discard on failure — or, inside a batch,
+/// just mark the batch dirty so its single boundary snapshot is kept.
+/// The command result is always sent; snapshot bookkeeping never
+/// fails the user's actual work.
+fn snapshot_then<T>(
+    snap: Option<&SnapshotStore>,
+    batch: &mut BatchState,
+    conn: &Connection,
+    source: Option<&str>,
+    reply: oneshot::Sender<Result<T, DbError>>,
+    run: impl FnOnce() -> Result<T, DbError>,
+) {
+    let staged = if batch.active {
+        None
+    } else {
+        stage_pre_mutation(snap, conn, source)
+    };
+    let result = run();
+    let committed = result.is_ok();
+    if batch.active {
+        if committed {
+            batch.dirty = true;
+        }
+    } else if let (Some(store), Some(st)) = (snap, staged) {
+        let outcome = if committed {
+            store.finalize(st).map(|_| ())
+        } else {
+            store.discard(st)
+        };
+        if let Err(e) = outcome {
+            warn!(error = %e, "undo snapshot bookkeeping failed (command already applied)");
+        }
+    }
+    let _ = reply.send(result);
+}
+
+/// Open a batch: one boundary snapshot, then suppress per-command
+/// snapshots until `end_batch`.
+fn begin_batch(
+    snap: Option<&SnapshotStore>,
+    conn: &Connection,
+    batch: &mut BatchState,
+    source: Option<&str>,
+) {
+    if batch.active {
+        warn!("BeginBatch while a batch is active; ignoring (no nested batches)");
+        return;
+    }
+    batch.staged = stage_pre_mutation(snap, conn, source);
+    batch.active = true;
+    batch.dirty = false;
+}
+
+/// Close a batch: keep the boundary snapshot iff a mutation committed
+/// during it, else discard it (an all-skips batch leaves no undo step).
+fn end_batch(snap: Option<&SnapshotStore>, batch: &mut BatchState) {
+    if !batch.active {
+        warn!("EndBatch with no active batch; ignoring");
+        return;
+    }
+    if let (Some(store), Some(st)) = (snap, batch.staged.take()) {
+        let outcome = if batch.dirty {
+            store.finalize(st).map(|_| ())
+        } else {
+            store.discard(st)
+        };
+        if let Err(e) = outcome {
+            warn!(error = %e, "batch undo snapshot bookkeeping failed");
+        }
+    }
+    batch.active = false;
+    batch.dirty = false;
+}
+
+fn do_undo(
+    snap: Option<&SnapshotStore>,
+    conn: &mut Connection,
+) -> Result<Option<SnapshotMeta>, DbError> {
+    snap.map_or(Ok(None), |store| {
+        store.undo(conn).map_err(snapshot_to_db_error)
+    })
+}
+
+fn do_redo(
+    snap: Option<&SnapshotStore>,
+    conn: &mut Connection,
+) -> Result<Option<SnapshotMeta>, DbError> {
+    snap.map_or(Ok(None), |store| {
+        store.redo(conn).map_err(snapshot_to_db_error)
+    })
+}
+
+fn peek_undo_op(snap: Option<&SnapshotStore>) -> Result<Option<SnapshotMeta>, DbError> {
+    snap.map_or(Ok(None), |s| s.peek_undo().map_err(snapshot_to_db_error))
+}
+
+fn peek_redo_op(snap: Option<&SnapshotStore>) -> Result<Option<SnapshotMeta>, DbError> {
+    snap.map_or(Ok(None), |s| s.peek_redo().map_err(snapshot_to_db_error))
+}
+
+fn handle_request(
+    conn: &Connection,
+    persistence: Option<&Persistence>,
+    snap: Option<&SnapshotStore>,
+    batch: &mut BatchState,
+    req: Request,
+) {
     match req {
         Request::CreateTable {
             name,
@@ -1343,7 +1634,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_create_table(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_create_table(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1357,7 +1648,9 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_drop_table(conn, persistence, source.as_deref(), &name));
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || {
+                do_drop_table(conn, persistence, source.as_deref(), &name)
+            });
         }
         Request::AddColumn {
             table,
@@ -1365,7 +1658,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_add_column(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_add_column(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1380,7 +1673,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_drop_column(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_drop_column(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1396,7 +1689,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_rename_column(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_rename_column(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1413,7 +1706,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_change_column_type(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_change_column_type(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1450,7 +1743,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_add_relationship(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_add_relationship(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1469,7 +1762,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_drop_relationship(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_drop_relationship(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1483,7 +1776,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_add_index(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_add_index(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1497,7 +1790,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_drop_index(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_drop_index(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1511,7 +1804,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_add_constraint(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_add_constraint(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1527,7 +1820,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_drop_constraint(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_drop_constraint(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1543,7 +1836,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_insert(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_insert(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1559,7 +1852,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_update(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_update(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1574,7 +1867,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_delete(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_delete(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1615,7 +1908,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             returning,
             reply,
         } => {
-            let _ = reply.send(do_sql_insert(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_sql_insert(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1633,7 +1926,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             returning,
             reply,
         } => {
-            let _ = reply.send(do_sql_update(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_sql_update(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1649,7 +1942,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             returning,
             reply,
         } => {
-            let _ = reply.send(do_sql_delete(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_sql_delete(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1663,7 +1956,7 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             source,
             reply,
         } => {
-            let _ = reply.send(do_rebuild_from_text(
+            snapshot_then(snap, batch, conn, source.as_deref(), reply, || do_rebuild_from_text(
                 conn,
                 persistence,
                 source.as_deref(),
@@ -1691,6 +1984,18 @@ fn handle_request(conn: &Connection, persistence: Option<&Persistence>, req: Req
             let result = do_list_names_for(conn, source);
             let _ = reply.send(result);
         }
+        // Undo/redo/peek/batch are intercepted in `worker_loop` (they
+        // need `&mut conn` or persistent batch state) and never reach
+        // here. Listed explicitly so a new variant still forces a
+        // decision at compile time.
+        Request::Undo { .. }
+        | Request::Redo { .. }
+        | Request::PeekUndo { .. }
+        | Request::PeekRedo { .. }
+        | Request::BeginBatch { .. }
+        | Request::EndBatch { .. } => {
+            unreachable!("undo/redo/peek/batch are handled in worker_loop")
+        }
     }
 }
 
diff --git a/tests/undo_snapshots.rs b/tests/undo_snapshots.rs
new file mode 100644
index 0000000..1798385
--- /dev/null
+++ b/tests/undo_snapshots.rs
@@ -0,0 +1,236 @@
+//! Tier-3 integration tests for the undo/snapshot ring wired into
+//! the db worker (ADR-0006 Amendment 1, §8 step 3).
+//!
+//! These drive the real `Database` worker: a mutation takes a
+//! pre-op snapshot, `undo` restores it through the live connection,
+//! `redo` re-applies, a batch collapses to a single undo step, and
+//! `--no-undo` (undo disabled) takes no snapshots at all.
+
+use std::path::Path;
+
+use rdbms_playground::db::Database;
+use rdbms_playground::dsl::{ColumnSpec, RowFilter, Type, Value};
+use rdbms_playground::persistence::Persistence;
+use rdbms_playground::project;
+
+fn tempdir() -> tempfile::TempDir {
+    tempfile::tempdir().expect("create tempdir")
+}
+
+fn rt() -> tokio::runtime::Runtime {
+    tokio::runtime::Builder::new_current_thread()
+        .enable_all()
+        .build()
+        .expect("tokio rt")
+}
+
+/// Open a fresh temp project with undo enabled (or not).
+fn open_project(
+    data: &tempfile::TempDir,
+    undo_enabled: bool,
+) -> (project::Project, Database, std::path::PathBuf) {
+    let project = project::open_or_create(None, Some(data.path())).expect("open project");
+    let path = project.path().to_path_buf();
+    let persistence = Persistence::new(path.clone());
+    let db = Database::open_with_persistence_and_undo(project.db_path(), persistence, undo_enabled)
+        .expect("open db");
+    (project, db, path)
+}
+
+async fn make_customers(db: &Database) {
+    db.create_table(
+        "Customers".to_string(),
+        vec![
+            ColumnSpec::new("id".to_string(), Type::Serial),
+            ColumnSpec::new("Name".to_string(), Type::Text),
+        ],
+        vec!["id".to_string()],
+        Some("create table Customers with pk id(serial)".to_string()),
+    )
+    .await
+    .unwrap();
+}
+
+async fn insert_named(db: &Database, name: &str) {
+    db.insert(
+        "Customers".to_string(),
+        None,
+        vec![Value::Text(name.to_string())],
+        Some(format!("insert into Customers ('{name}')")),
+    )
+    .await
+    .unwrap();
+}
+
+async fn row_count(db: &Database) -> usize {
+    db.query_data("Customers".to_string(), None, None, None)
+        .await
+        .unwrap()
+        .rows
+        .len()
+}
+
+fn snapshots_dir(path: &Path) -> std::path::PathBuf {
+    path.join(".snapshots")
+}
+
+#[test]
+fn mutation_snapshots_and_undo_restores_through_the_worker() {
+    let data = tempdir();
+    let (_p, db, path) = open_project(&data, true);
+
+    rt().block_on(async {
+        make_customers(&db).await;
+        insert_named(&db, "Alice").await;
+        insert_named(&db, "Bob").await;
+        assert_eq!(row_count(&db).await, 2);
+
+        // Destructive op: delete Bob (id = 2).
+        db.delete(
+            "Customers".to_string(),
+            RowFilter::eq("id", Value::Number("2".to_string())),
+            Some("delete from Customers where id = 2".to_string()),
+        )
+        .await
+        .unwrap();
+        assert_eq!(row_count(&db).await, 1);
+
+        // The pending undo names the delete.
+        let peek = db.peek_undo().await.unwrap().expect("an undo entry");
+        assert_eq!(peek.command, "delete from Customers where id = 2");
+
+        // Undo restores Bob.
+        let undone = db.undo().await.unwrap().expect("undo applied");
+        assert_eq!(undone.command, "delete from Customers where id = 2");
+        assert_eq!(row_count(&db).await, 2, "Bob restored by undo");
+    });
+
+    assert!(snapshots_dir(&path).exists(), "snapshots dir created");
+}
+
+#[test]
+fn redo_reapplies_the_undone_command() {
+    let data = tempdir();
+    let (_p, db, _path) = open_project(&data, true);
+
+    rt().block_on(async {
+        make_customers(&db).await;
+        insert_named(&db, "Alice").await;
+        insert_named(&db, "Bob").await;
+        db.delete(
+            "Customers".to_string(),
+            RowFilter::eq("id", Value::Number("2".to_string())),
+            Some("delete from Customers where id = 2".to_string()),
+        )
+        .await
+        .unwrap();
+        assert_eq!(row_count(&db).await, 1);
+
+        db.undo().await.unwrap();
+        assert_eq!(row_count(&db).await, 2);
+
+        let redone = db.redo().await.unwrap().expect("redo applied");
+        assert_eq!(redone.command, "delete from Customers where id = 2");
+        assert_eq!(row_count(&db).await, 1, "delete re-applied by redo");
+    });
+}
+
+#[test]
+fn new_work_after_undo_clears_redo() {
+    let data = tempdir();
+    let (_p, db, _path) = open_project(&data, true);
+
+    rt().block_on(async {
+        make_customers(&db).await;
+        insert_named(&db, "Alice").await;
+        db.delete(
+            "Customers".to_string(),
+            RowFilter::AllRows,
+            Some("delete from Customers --all-rows".to_string()),
+        )
+        .await
+        .unwrap();
+        db.undo().await.unwrap();
+        assert!(db.peek_redo().await.unwrap().is_some(), "redo available");
+
+        // New destructive work.
+        insert_named(&db, "Carol").await;
+        assert!(
+            db.peek_redo().await.unwrap().is_none(),
+            "new work discards the redo stack"
+        );
+    });
+}
+
+#[test]
+fn undo_disabled_takes_no_snapshots() {
+    let data = tempdir();
+    let (_p, db, path) = open_project(&data, false);
+
+    rt().block_on(async {
+        make_customers(&db).await;
+        insert_named(&db, "Alice").await;
+        db.delete(
+            "Customers".to_string(),
+            RowFilter::AllRows,
+            Some("delete from Customers --all-rows".to_string()),
+        )
+        .await
+        .unwrap();
+
+        // Nothing to undo, and no snapshot machinery on disk.
+        assert!(db.undo().await.unwrap().is_none(), "undo is a no-op when disabled");
+        assert!(db.peek_undo().await.unwrap().is_none());
+    });
+
+    assert!(
+        !snapshots_dir(&path).exists(),
+        "no .snapshots dir when undo is disabled"
+    );
+}
+
+#[test]
+fn batch_records_a_single_undo_step() {
+    let data = tempdir();
+    let (_p, db, _path) = open_project(&data, true);
+
+    rt().block_on(async {
+        make_customers(&db).await; // one undo entry (the create)
+
+        // A batch of three inserts → one boundary snapshot.
+        db.begin_batch(Some("replay history.log".to_string()))
+            .await
+            .unwrap();
+        insert_named(&db, "Alice").await;
+        insert_named(&db, "Bob").await;
+        insert_named(&db, "Carol").await;
+        db.end_batch().await.unwrap();
+        assert_eq!(row_count(&db).await, 3);
+
+        // The single batch undo names the batch command.
+        let peek = db.peek_undo().await.unwrap().expect("batch undo entry");
+        assert_eq!(peek.command, "replay history.log");
+
+        // One undo rolls the whole batch back to the pre-batch state
+        // (table exists, no rows).
+        db.undo().await.unwrap();
+        assert_eq!(row_count(&db).await, 0, "whole batch undone in one step");
+
+        // The next undo is the create_table (table gone).
+        let next = db.peek_undo().await.unwrap().expect("create entry");
+        assert_eq!(next.command, "create table Customers with pk id(serial)");
+    });
+}
+
+#[test]
+fn empty_undo_and_redo_are_no_ops() {
+    let data = tempdir();
+    let (_p, db, _path) = open_project(&data, true);
+
+    rt().block_on(async {
+        assert!(db.undo().await.unwrap().is_none());
+        assert!(db.redo().await.unwrap().is_none());
+        assert!(db.peek_undo().await.unwrap().is_none());
+        assert!(db.peek_redo().await.unwrap().is_none());
+    });
+}