diff --git a/Cargo.lock b/Cargo.lock
index 29b93e19..cda4cb77 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3475,7 +3475,7 @@ dependencies = [
 
 [[package]]
 name = "jjuicy-cli"
-version = "1.0.1"
+version = "1.0.2"
 dependencies = [
  "anyhow",
  "assert_matches",
diff --git a/src/worker/gui_util.rs b/src/worker/gui_util.rs
index 2d3099c0..6575e3bd 100644
--- a/src/worker/gui_util.rs
+++ b/src/worker/gui_util.rs
@@ -447,24 +447,26 @@ impl WorkspaceSession<'_> {
             return Ok(HashMap::new());
         }
 
-        // for each bookmark, BFS from its commit toward ancestors, stopping at log commits.
-        // those boundary log commits are the nearest visible ancestors (fork points).
-        // cost: O(depth_to_log) per bookmark instead of O(repo_size) per bookmark.
+        const MAX_BFS_VISITS: usize = 100_000;
         let mut fork_map: HashMap<CommitId, Vec<String>> = HashMap::new();
-        for (commit_id, labels) in &commits_to_labels {
+        let mut total_visits: usize = 0;
+        'outer: for (commit_id, labels) in &commits_to_labels {
             let mut queue = VecDeque::new();
             queue.push_back(commit_id.clone());
             let mut visited: HashSet<CommitId> = HashSet::new();
             visited.insert(commit_id.clone());
 
             while let Some(current) = queue.pop_front() {
+                total_visits += 1;
+                if total_visits > MAX_BFS_VISITS {
+                    break 'outer;
+                }
                 let commit = self.get_commit(&current)?;
                 for parent_id in commit.parent_ids() {
                     if !visited.insert(parent_id.clone()) {
                         continue;
                     }
                     if log_contains(parent_id)? {
-                        // nearest visible ancestor — record without walking further
                         let entry = fork_map.entry(parent_id.clone()).or_default();
                         for label in labels {
                             if !entry.contains(label) {
diff --git a/src/worker/tests/mod.rs b/src/worker/tests/mod.rs
index 285070d4..fb10298c 100644
--- a/src/worker/tests/mod.rs
+++ b/src/worker/tests/mod.rs
@@ -1,3 +1,4 @@
+mod perf;
 mod queries;
 mod session;
 
diff --git a/src/worker/tests/perf.rs b/src/worker/tests/perf.rs
new file mode 100644
index 00000000..8cf2c7f2
--- /dev/null
+++ b/src/worker/tests/perf.rs
@@ -0,0 +1,116 @@
+use std::path::PathBuf;
+use std::sync::mpsc;
+use std::time::{Duration, Instant};
+
+use anyhow::Result;
+
+use crate::worker::{WorkerSession, queries};
+
+const TIMEOUT_SECS: u64 = 30;
+
+struct RevsetBench {
+    name: &'static str,
+    revset: &'static str,
+}
+
+const REVSETS: &[RevsetBench] = &[
+    RevsetBench {
+        name: "all()",
+        revset: "all()",
+    },
+    RevsetBench {
+        name: "::@",
+        revset: "::@",
+    },
+    RevsetBench {
+        name: "default",
+        revset: r#"present(@) | ancestors(immutable_heads().., 2) | trunk()"#,
+    },
+    RevsetBench {
+        name: "mine",
+        revset: "trunk() | (trunk()..(bookmarks() & mine())) | parents(bookmarks() & mine()) | @",
+    },
+];
+
+/// Run with: BENCH_REPO=/path/to/repo cargo test query_bench -- --ignored --nocapture
+#[tokio::test]
+#[ignore]
+async fn query_bench() -> Result<()> {
+    let repo_path = match std::env::var("BENCH_REPO") {
+        Ok(p) => PathBuf::from(p),
+        Err(_) => {
+            eprintln!("SKIP: set BENCH_REPO=/path/to/jj/repo to run this benchmark");
+            return Ok(());
+        }
+    };
+
+    if !repo_path.join(".jj").exists() {
+        eprintln!("SKIP: no .jj directory found at {}", repo_path.display());
+        return Ok(());
+    }
+
+    eprintln!("repo: {}", repo_path.display());
+
+    let mut session = WorkerSession::default();
+
+    let t0 = Instant::now();
+    let mut ws = session.load_workspace(&repo_path).await?;
+    eprintln!("load_workspace: {}ms", t0.elapsed().as_millis());
+
+    let t1 = Instant::now();
+    ws.import_and_snapshot(false, true).await?;
+    eprintln!("import_and_snapshot: {}ms", t1.elapsed().as_millis());
+
+    eprintln!();
+    eprintln!(
+        "{:<12} {:>10} {:>10} {:>10} {:>10}  {:<6}",
+        "revset", "eval", "forks", "page", "total", "rows"
+    );
+    eprintln!("{}", "-".repeat(70));
+
+    for bench in REVSETS {
+        let (_cancel, rx) = mpsc::channel::<()>();
+        let bench_name = bench.name;
+
+        std::thread::spawn(move || {
+            if let Err(mpsc::RecvTimeoutError::Timeout) =
+                rx.recv_timeout(Duration::from_secs(TIMEOUT_SECS))
+            {
+                eprintln!("{bench_name:<12}  TIMEOUT (>{TIMEOUT_SECS}s)");
+                std::process::exit(1);
+            }
+        });
+
+        let t_total = Instant::now();
+
+        let t_eval = Instant::now();
+        let revset_result = ws.evaluate_revset_str(bench.revset);
+        let eval_ms = t_eval.elapsed().as_millis();
+
+        let revset = match revset_result {
+            Ok(r) => r,
+            Err(e) => {
+                eprintln!("{:<12}  ERROR: {e}", bench.name);
+                continue;
+            }
+        };
+
+        let t_forks = Instant::now();
+        let hidden_forks = ws.compute_hidden_forks(bench.revset).unwrap_or_default();
+        let forks_ms = t_forks.elapsed().as_millis();
+
+        let t_page = Instant::now();
+        let state = queries::QueryState::new(50);
+        let mut qs = queries::QuerySession::new(&ws, &*revset, state, hidden_forks);
+        let page = qs.get_page()?;
+        let page_ms = t_page.elapsed().as_millis();
+
+        let total_ms = t_total.elapsed().as_millis();
+        eprintln!(
+            "{:<12} {:>9}ms {:>9}ms {:>9}ms {:>9}ms  {:<6}",
+            bench.name, eval_ms, forks_ms, page_ms, total_ms, page.rows.len()
+        );
+    }
+
+    Ok(())
+}