chore(release): 1.5.16 — strict multi-agent tab isolation on the relay

Several agents (and other tools opening tabs) share one real Chrome via the
relay. Previously every session adopted ALL tabs from Target.getTargets, so
another agent's tab churn polluted the list, dropped the tab being driven, and
drifted commands onto the wrong page (the W-8BEN tax tab vanished mid-flow when a
concurrent iphone-use agent opened tabs).

A tab group belongs to exactly one agent. On the relay a session now tracks and
drives ONLY the tabs it created (its own colored group) plus pop-ups its own
clicks open — it never adopts the user's or other agents' tabs:

- discover_and_attach (relay): create the session's own tab and pin it; do not
  adopt any existing foreign tab.
- resync_targets (relay): never adopt unknown targets; never prune the session's
  tabs on a single getTargets snapshot (multi-agent churn / cross-process-nav
  gaps routinely omit live tabs) — prune only after RELAY_PRUNE_MISSES
  consecutive absent snapshots (debounced), pinned active always protected.
- adopt_newly_opened: a tab that appears right after this session's action is a
  pop-up we opened — record it as owned.

Launched browsers (every tab ours) keep adopting all tabs. Adds debounced_prune_ids
+ unit tests for the churn tolerance.

cli/Cargo.lock

@@ -290,7 +290,7 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 
 [[package]]
 name = "chrome-use"
-version = "1.5.11"
+version = "1.5.16"
 dependencies = [
  "aes",
  "aes-gcm",

cli/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "chrome-use"
-version = "1.5.11"
+version = "1.5.16"
 edition = "2021"
 description = "Fast browser automation CLI for AI agents"
 license = "Apache-2.0"

cli/src/commands.rs

@@ -34,11 +34,50 @@ pub enum ParseError {
 /// suggestions on an unknown command (issue #29). Not exhaustive — just the
 /// common verbs plus a few known wrong-guesses mapped to the real command.
 const KNOWN_COMMANDS: &[&str] = &[
-    "open", "navigate", "click", "fill", "type", "press", "snapshot", "screenshot", "eval", "get",
-    "text", "html", "frames", "find", "wait", "scroll", "hover", "select", "check", "uncheck",
-    "tab", "tabs", "close", "back", "forward", "reload", "sessions", "status", "daemon", "doctor",
-    "upgrade", "connect", "cookies", "mouse", "keyboard", "stream", "frame", "profiles", "title",
-    "url", "is", "drag", "dialog", "upload",
+    "open",
+    "navigate",
+    "click",
+    "fill",
+    "type",
+    "press",
+    "snapshot",
+    "screenshot",
+    "eval",
+    "get",
+    "text",
+    "html",
+    "frames",
+    "find",
+    "wait",
+    "scroll",
+    "hover",
+    "select",
+    "check",
+    "uncheck",
+    "tab",
+    "tabs",
+    "close",
+    "back",
+    "forward",
+    "reload",
+    "sessions",
+    "status",
+    "daemon",
+    "doctor",
+    "upgrade",
+    "connect",
+    "cookies",
+    "mouse",
+    "keyboard",
+    "stream",
+    "frame",
+    "profiles",
+    "title",
+    "url",
+    "is",
+    "drag",
+    "dialog",
+    "upload",
 ];
 
 /// Levenshtein distance, capped — small inputs only (command names).
@@ -525,20 +564,31 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
             Ok(json!({ "id": id, "action": "fill", "selector": sel, "value": rest[1..].join(" ") }))
         }
         "type" => {
+            // `--key-events` (alias `--keys`): send real per-character keystrokes
+            // instead of Input.insertText, so autocomplete/combobox widgets that
+            // only react to key events fire (e.g. Google address postal lookup).
+            let key_events = rest.iter().any(|a| *a == "--key-events" || *a == "--keys");
+            let rest: Vec<&str> = rest
+                .iter()
+                .copied()
+                .filter(|a| *a != "--key-events" && *a != "--keys")
+                .collect();
             // `type --focused <text>` types into whatever element currently has
             // focus (no selector) — for custom widgets that move focus to a hidden
             // input after you open them.
             if rest.first() == Some(&"--focused") {
                 return Ok(json!({
                     "id": id, "action": "type", "focused": true,
-                    "text": rest[1..].join(" "),
+                    "text": rest[1..].join(" "), "keyEvents": key_events,
                 }));
             }
             let sel = rest.first().ok_or_else(|| ParseError::MissingArguments {
                 context: "type".to_string(),
-                usage: "type <selector> <text>   (or: type --focused <text>)",
+                usage: "type <selector> <text>   (or: type --focused <text>) [--key-events]",
             })?;
-            Ok(json!({ "id": id, "action": "type", "selector": sel, "text": rest[1..].join(" ") }))
+            Ok(
+                json!({ "id": id, "action": "type", "selector": sel, "text": rest[1..].join(" "), "keyEvents": key_events }),
+            )
         }
         "pick" => {
             // pick <selector|@ref> --option "<text>"  — atomic combobox select:
@@ -752,7 +802,8 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
                             _ => {
                                 return Err(ParseError::InvalidValue {
                                     message: format!("scroll --at: invalid coordinate `{}`", val),
-                                    usage: "scroll [direction] [amount] --at <x,y> (e.g. --at 640,400)",
+                                    usage:
+                                        "scroll [direction] [amount] --at <x,y> (e.g. --at 640,400)",
                                 })
                             }
                         }
@@ -961,17 +1012,21 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
                     "--full" | "-f" => full_page = true,
                     // `--clip x,y,w,h` captures a pixel region (issue #34).
                     "--clip" => {
-                        let raw = rest.get(i + 1).ok_or_else(|| ParseError::MissingArguments {
-                            context: "screenshot --clip".to_string(),
-                            usage: "screenshot --clip <x,y,w,h> [path]",
-                        })?;
+                        let raw = rest
+                            .get(i + 1)
+                            .ok_or_else(|| ParseError::MissingArguments {
+                                context: "screenshot --clip".to_string(),
+                                usage: "screenshot --clip <x,y,w,h> [path]",
+                            })?;
                         let nums: Vec<f64> = raw
                             .split(',')
                             .filter_map(|n| n.trim().parse::<f64>().ok())
                             .collect();
                         if nums.len() != 4 {
                             return Err(ParseError::InvalidValue {
-                                message: format!("--clip expects 'x,y,w,h' (4 numbers), got '{raw}'"),
+                                message: format!(
+                                    "--clip expects 'x,y,w,h' (4 numbers), got '{raw}'"
+                                ),
                                 usage: "screenshot --clip <x,y,w,h> [path]",
                             });
                         }
@@ -4112,6 +4167,28 @@ mod tests {
         assert_eq!(cmd["action"], "type");
         assert_eq!(cmd["selector"], "#input");
         assert_eq!(cmd["text"], "some text");
+        assert_eq!(cmd["keyEvents"], false);
+    }
+
+    #[test]
+    fn test_type_key_events() {
+        // --key-events sends real keystrokes (for autocomplete/combobox) and must
+        // not be swallowed into the typed text.
+        let cmd = parse_command(
+            &args("type #postal 201-0001 --key-events"),
+            &default_flags(),
+        )
+        .unwrap();
+        assert_eq!(cmd["action"], "type");
+        assert_eq!(cmd["selector"], "#postal");
+        assert_eq!(cmd["text"], "201-0001");
+        assert_eq!(cmd["keyEvents"], true);
+
+        let focused =
+            parse_command(&args("type --focused 201-0001 --keys"), &default_flags()).unwrap();
+        assert_eq!(focused["focused"], true);
+        assert_eq!(focused["text"], "201-0001");
+        assert_eq!(focused["keyEvents"], true);
     }
 
     #[test]
@@ -4399,8 +4476,11 @@ mod tests {
     #[test]
     fn test_screenshot_clip() {
         // `--clip x,y,w,h` captures a pixel region (issue #34); the path still parses.
-        let cmd = parse_command(&args("screenshot --clip 10,20,200,40 out.png"), &default_flags())
-            .unwrap();
+        let cmd = parse_command(
+            &args("screenshot --clip 10,20,200,40 out.png"),
+            &default_flags(),
+        )
+        .unwrap();
         assert_eq!(cmd["action"], "screenshot");
         assert_eq!(cmd["clip"]["x"], 10.0);
         assert_eq!(cmd["clip"]["y"], 20.0);
@@ -4978,7 +5058,10 @@ mod tests {
         assert_eq!(nearest_command("sesions").as_deref(), Some("sessions"));
         assert_eq!(nearest_command("session").as_deref(), Some("sessions"));
         assert_eq!(nearest_command("clik").as_deref(), Some("click"));
-        assert_eq!(nearest_command("screenshits").as_deref(), Some("screenshot"));
+        assert_eq!(
+            nearest_command("screenshits").as_deref(),
+            Some("screenshot")
+        );
         // Nonsense with no close match stays silent.
         assert_eq!(nearest_command("xyzzy"), None);
         // The unknown-command error embeds the suggestion.

cli/src/native/actions.rs

@@ -3225,6 +3225,14 @@ async fn handle_type(cmd: &Value, state: &mut DaemonState) -> Result<Value, Stri
     let mgr = state.browser.as_ref().ok_or("Browser not launched")?;
     let session_id = mgr.active_session_id()?.to_string();
 
+    // `--key-events`: dispatch real per-character keyDown/keyUp instead of
+    // Input.insertText, so autocomplete/combobox widgets that only react to key
+    // events fire (e.g. Google's address postal-code lookup) (issue #4/#36).
+    let key_events = cmd
+        .get("keyEvents")
+        .and_then(|v| v.as_bool())
+        .unwrap_or(false);
+
     // `type --focused <text>`: type into the currently-focused element without a
     // selector (custom widgets that move focus to a hidden input on open).
     if cmd
@@ -3236,7 +3244,14 @@ async fn handle_type(cmd: &Value, state: &mut DaemonState) -> Result<Value, Stri
             .get("text")
             .and_then(|v| v.as_str())
             .ok_or("Missing 'text' parameter")?;
-        interaction::type_text_into_active_context(&mgr.client, &session_id, text, None).await?;
+        interaction::type_text_into_active_context(
+            &mgr.client,
+            &session_id,
+            text,
+            None,
+            key_events,
+        )
+        .await?;
         return Ok(json!({ "typed": text, "focused": true }));
     }
 
@@ -3260,6 +3275,7 @@ async fn handle_type(cmd: &Value, state: &mut DaemonState) -> Result<Value, Stri
         clear,
         delay,
         &state.iframe_sessions,
+        key_events,
     )
     .await?;
     Ok(json!({ "typed": text }))
@@ -3483,27 +3499,40 @@ async fn handle_scroll(cmd: &Value, state: &mut DaemonState) -> Result<Value, St
         return Ok(json!({ "scrolled": true, "via": "selector" }));
     }
 
-    // No selector: dispatch a real (isTrusted) wheel at a viewport coordinate.
-    // This hits the compositor and scrolls whatever scroll container is under the
-    // pointer — including cross-origin iframes that `window.scrollBy` on the top
-    // document silently no-ops on (issue #36). The coordinate is, in priority:
-    //   --at x,y  → that exact pixel
-    //   --frame n → the center of frame n from `chrome-use frames`
-    //   default   → the viewport center
-    let (x, y, via) = if let Some(at) = cmd.get("at").and_then(|v| v.as_array()) {
-        let x = at.first().and_then(|v| v.as_f64()).unwrap_or(0.0);
-        let y = at.get(1).and_then(|v| v.as_f64()).unwrap_or(0.0);
-        (x, y, "at")
-    } else if let Some(n) = cmd.get("frame").and_then(|v| v.as_u64()) {
-        let (x, y) = frame_center(mgr, &session_id, &state.iframe_sessions, n as usize).await?;
-        (x, y, "frame")
-    } else {
-        let (x, y) = viewport_center(mgr, &session_id).await?;
-        (x, y, "center")
-    };
+    // `--at x,y` / `--frame n`: dispatch a real (isTrusted) wheel at a viewport
+    // coordinate. This hits the compositor and scrolls whatever scroll container
+    // is under the pointer — including cross-origin iframes that `window.scrollBy`
+    // on the top document silently no-ops on (issue #36).
+    if cmd.get("at").is_some() || cmd.get("frame").is_some() {
+        let (x, y, via) = if let Some(at) = cmd.get("at").and_then(|v| v.as_array()) {
+            let x = at.first().and_then(|v| v.as_f64()).unwrap_or(0.0);
+            let y = at.get(1).and_then(|v| v.as_f64()).unwrap_or(0.0);
+            (x, y, "at")
+        } else {
+            let n = cmd.get("frame").and_then(|v| v.as_u64()).unwrap_or(0);
+            let (x, y) = frame_center(mgr, &session_id, &state.iframe_sessions, n as usize).await?;
+            (x, y, "frame")
+        };
+        dispatch_wheel(&mgr.client, &session_id, x, y, dx, dy).await?;
+        return Ok(json!({ "scrolled": true, "via": via, "at": [x, y] }));
+    }
 
-    dispatch_wheel(&mgr.client, &session_id, x, y, dx, dy).await?;
-    Ok(json!({ "scrolled": true, "via": via, "at": [x, y] }))
+    // Default (no selector/at/frame): scroll the page with `window.scrollBy`. This
+    // is the reliable path for ordinary page scrolling; a coordinate wheel at the
+    // viewport centre is NOT a dependable substitute (it no-ops on some pages,
+    // e.g. headless), so the wheel stays opt-in via `--at`/`--frame` for the
+    // cross-origin-iframe case (issue #36).
+    interaction::scroll(
+        &mgr.client,
+        &session_id,
+        &state.ref_map,
+        None,
+        dx,
+        dy,
+        &state.iframe_sessions,
+    )
+    .await?;
+    Ok(json!({ "scrolled": true, "via": "page" }))
 }
 
 /// Viewport center in CSS pixels, used as the default wheel landing point for
@@ -3826,12 +3855,9 @@ async fn handle_gettext(cmd: &Value, state: &mut DaemonState) -> Result<Value, S
 async fn handle_frames(_cmd: &Value, state: &mut DaemonState) -> Result<Value, String> {
     let mgr = state.browser.as_ref().ok_or("Browser not launched")?;
     let session_id = mgr.active_session_id()?.to_string();
-    let frames = super::element::collect_all_frames_text(
-        &mgr.client,
-        &session_id,
-        &state.iframe_sessions,
-    )
-    .await?;
+    let frames =
+        super::element::collect_all_frames_text(&mgr.client, &session_id, &state.iframe_sessions)
+            .await?;
     let list: Vec<Value> = frames
         .iter()
         .enumerate()
@@ -4327,7 +4353,10 @@ async fn handle_cf_status(_cmd: &Value, state: &mut DaemonState) -> Result<Value
     let url = mgr.get_url().await.unwrap_or_default();
 
     // 1. Is the page a Cloudflare challenge right now?
-    let probe_raw = mgr.evaluate(CF_CHALLENGE_JS, None).await.unwrap_or(Value::Null);
+    let probe_raw = mgr
+        .evaluate(CF_CHALLENGE_JS, None)
+        .await
+        .unwrap_or(Value::Null);
     let probe = parse_json_string(probe_raw, "cf challenge probe").unwrap_or(Value::Null);
     let challenged = probe
         .get("challenged")
@@ -4722,8 +4751,18 @@ async fn handle_keyboard(cmd: &Value, state: &DaemonState) -> Result<Value, Stri
                 .get("text")
                 .and_then(|v| v.as_str())
                 .ok_or("Missing 'text' parameter")?;
-            interaction::type_text_into_active_context(&mgr.client, &session_id, text, None)
-                .await?;
+            let key_events = cmd
+                .get("keyEvents")
+                .and_then(|v| v.as_bool())
+                .unwrap_or(false);
+            interaction::type_text_into_active_context(
+                &mgr.client,
+                &session_id,
+                text,
+                None,
+                key_events,
+            )
+            .await?;
             return Ok(json!({ "typed": text }));
         }
         Some("insertText") => {
@@ -7252,6 +7291,28 @@ async fn handle_drag(cmd: &Value, state: &mut DaemonState) -> Result<Value, Stri
         .and_then(|v| v.as_str())
         .ok_or("Missing 'target' parameter")?;
 
+    // Over the relay (or into an iframe) a coordinate drag drifts to the
+    // foreground tab and can't reach an OOPIF — DOM-dispatch an HTML5 drag in the
+    // element's own session instead (issues #31/#36). `coord` mode forces the
+    // coordinate path for pointer-driven drags (canvas/sliders) on a launched
+    // browser.
+    if std::env::var("AGENT_BROWSER_CLICK_MODE").as_deref() != Ok("coord")
+        && (crate::connect::relay_url().is_some()
+            || state.ref_map.ref_is_in_iframe(source)
+            || state.ref_map.ref_is_in_iframe(target))
+    {
+        super::interaction::dom_drag(
+            &mgr.client,
+            &session_id,
+            &state.ref_map,
+            source,
+            target,
+            &state.iframe_sessions,
+        )
+        .await?;
+        return Ok(json!({ "dragged": { "source": source, "target": target }, "via": "dom" }));
+    }
+
     let (sx, sy, _, _, source_session_id) = super::element::resolve_element_center(
         &mgr.client,
         &session_id,

cli/src/native/browser.rs

@@ -235,6 +235,43 @@ fn prunable_target_ids(
         .collect()
 }
 
+/// Consecutive missing `getTargets` snapshots before an owned relay tab is
+/// pruned. >1 so a single churning/partial snapshot (other agents opening/closing
+/// tabs) or a brief cross-process-nav gap can't drop the tab the agent is driving.
+const RELAY_PRUNE_MISSES: u32 = 3;
+
+/// Debounced prune for the relay: target ids to drop, mutating per-target miss
+/// counters. A tab in `live_ids` resets to 0; an absent (non-pinned) tab
+/// increments and is pruned only at `RELAY_PRUNE_MISSES`. Counters for
+/// no-longer-tracked targets are forgotten. Pure, so the multi-agent churn
+/// tolerance is unit-testable without a live browser.
+fn debounced_prune_ids(
+    pages: &[PageInfo],
+    live_ids: &HashSet<String>,
+    pinned: Option<&str>,
+    misses: &mut HashMap<String, u32>,
+) -> Vec<String> {
+    let tracked: HashSet<&str> = pages.iter().map(|p| p.target_id.as_str()).collect();
+    misses.retain(|tid, _| tracked.contains(tid.as_str()));
+    let mut prune = Vec::new();
+    for p in pages {
+        let tid = p.target_id.as_str();
+        if live_ids.contains(tid) {
+            misses.remove(tid);
+            continue;
+        }
+        if pinned == Some(tid) {
+            continue;
+        }
+        let c = misses.entry(p.target_id.clone()).or_insert(0);
+        *c += 1;
+        if *c >= RELAY_PRUNE_MISSES {
+            prune.push(p.target_id.clone());
+        }
+    }
+    prune
+}
+
 /// Whether the resolved active page is a tab the session created (its target_id
 /// is in `created_targets`). Pure core of [`BrowserManager::active_is_session_owned`]
 /// so the relay no-hijack rule is unit-testable without a live browser.
@@ -464,6 +501,14 @@ pub struct BrowserManager {
     /// the session's commands onto the wrong page — the wrong-origin-fetch hazard
     /// in the dogfood reports. Falls back to the index if the pinned tab is gone.
     active_target_id: Option<String>,
+    /// Per-target count of CONSECUTIVE `resync_targets` snapshots in which an
+    /// owned tab was missing from `Target.getTargets`. Over the relay a single
+    /// snapshot routinely omits live tabs (multi-agent churn, a cross-process nav
+    /// briefly dropping the target), so we must not prune on one miss — that lost
+    /// the tab the agent was driving. A tab is removed only after it's been absent
+    /// for `RELAY_PRUNE_MISSES` consecutive snapshots; any snapshot that includes
+    /// it resets the counter. Keyed by stable target_id.
+    relay_target_misses: HashMap<String, u32>,
     next_tab_id: u32,
     /// Whether to enable the CDP `Runtime` domain (console / error / exception capture).
     /// OFF by default for stealth: a live `Runtime.enable` is a detectable CDP signal
@@ -579,7 +624,10 @@ impl BrowserManager {
         crate::connect::log_connect_mode(
             &ws_url,
             true,
-            DAEMON_SESSION.get().map(String::as_str).unwrap_or("default"),
+            DAEMON_SESSION
+                .get()
+                .map(String::as_str)
+                .unwrap_or("default"),
         );
         let manager = if engine == "lightpanda" {
             initialize_lightpanda_manager(ws_url, process).await?
@@ -597,6 +645,7 @@ impl BrowserManager {
                 visited_origins: HashSet::new(),
                 created_targets: HashSet::new(),
                 active_target_id: None,
+                relay_target_misses: HashMap::new(),
                 next_tab_id: 1,
                 capture_console: console_capture_enabled(),
             };
@@ -681,7 +730,10 @@ impl BrowserManager {
         crate::connect::log_connect_mode(
             &ws_url,
             false,
-            DAEMON_SESSION.get().map(String::as_str).unwrap_or("default"),
+            DAEMON_SESSION
+                .get()
+                .map(String::as_str)
+                .unwrap_or("default"),
         );
         let client = Arc::new(CdpClient::connect_with_headers(&ws_url, headers).await?);
         let mut manager = Self {
@@ -696,6 +748,7 @@ impl BrowserManager {
             visited_origins: HashSet::new(),
             created_targets: HashSet::new(),
             active_target_id: None,
+            relay_target_misses: HashMap::new(),
             next_tab_id: 1,
             capture_console: console_capture_enabled(),
         };
@@ -725,6 +778,43 @@ impl BrowserManager {
         Self::connect_cdp(&ws_url).await
     }
 
+    /// Page targets to adopt, merging several `Target.getTargets` snapshots over
+    /// the extension relay. A single relay snapshot is flaky on a busy real Chrome
+    /// — it can omit live tabs (a different window's set, or a partial list; issue
+    /// #31) — so a tab the daemon should adopt would silently vanish (e.g. after a
+    /// daemon restart the page being driven disappeared from the tab list). Taking
+    /// the union of a few snapshots makes adoption resilient to a transient miss.
+    /// Off the relay (a browser we launched) one snapshot is authoritative.
+    async fn collect_page_targets(&self) -> Result<Vec<TargetInfo>, String> {
+        let rounds = if crate::connect::relay_url().is_some() {
+            3
+        } else {
+            1
+        };
+        let mut by_id: HashMap<String, TargetInfo> = HashMap::new();
+        let mut any_ok = false;
+        for i in 0..rounds {
+            if i > 0 {
+                tokio::time::sleep(Duration::from_millis(150)).await;
+            }
+            match self
+                .client
+                .send_command_typed::<_, GetTargetsResult>("Target.getTargets", &json!({}), None)
+                .await
+            {
+                Ok(result) => {
+                    any_ok = true;
+                    for t in result.target_infos.into_iter().filter(should_track_target) {
+                        by_id.entry(t.target_id.clone()).or_insert(t);
+                    }
+                }
+                Err(e) if i == rounds - 1 && !any_ok => return Err(e),
+                Err(_) => {}
+            }
+        }
+        Ok(by_id.into_values().collect())
+    }
+
     async fn discover_and_attach_targets(&mut self) -> Result<(), String> {
         self.client
             .send_command_typed::<_, Value>(
@@ -734,16 +824,7 @@ impl BrowserManager {
             )
             .await?;
 
-        let result: GetTargetsResult = self
-            .client
-            .send_command_typed("Target.getTargets", &json!({}), None)
-            .await?;
-
-        let page_targets: Vec<TargetInfo> = result
-            .target_infos
-            .into_iter()
-            .filter(should_track_target)
-            .collect();
+        let page_targets: Vec<TargetInfo> = self.collect_page_targets().await?;
 
         if page_targets.is_empty() {
             // Create a new tab
@@ -789,7 +870,19 @@ impl BrowserManager {
             self.active_page_index = 0;
             self.pin_active_target();
             self.enable_domains(&attach_result.session_id).await?;
+        } else if self.agent_group().is_some() {
+            // STRICT MULTI-AGENT ISOLATION (relay / the user's real Chrome).
+            // `page_targets` here are the USER's and OTHER agents' tabs. A tab
+            // group belongs to exactly ONE agent, so this session must NOT adopt
+            // any of them — it tracks ONLY tabs it creates (its own colored group)
+            // plus popups it opens. Adopting foreign tabs is precisely what let
+            // another concurrent agent's tab churn drop the tab we were driving and
+            // drift eval/click onto the wrong page (multi-agent failure). Open our
+            // own dedicated background tab in the session's group and pin it; the
+            // user's / other agents' tabs stay invisible to us.
+            self.tab_new(None, None).await?;
         } else {
+            // A browser WE launched: every tab is ours, so adopt them all.
             for target in &page_targets {
                 let attach_result: AttachToTargetResult = self
                     .client
@@ -815,7 +908,6 @@ impl BrowserManager {
                     target_type: target.target_type.clone(),
                 });
             }
-
             self.active_page_index = 0;
             self.pin_active_target();
             let session_id = self.pages[0].session_id.clone();
@@ -1161,15 +1253,31 @@ impl BrowserManager {
     pub async fn evaluate(&self, script: &str, _args: Option<Value>) -> Result<Value, String> {
         let session_id = self.active_session_id()?.to_string();
 
+        // `replMode: true` lets successive `eval`s re-declare top-level
+        // `let`/`const` instead of throwing "Identifier 'x' has already been
+        // declared" (issue #38 — independent `eval` steps in a test suite collided
+        // in the page's shared lexical scope). BUT replMode and `awaitPromise` are
+        // mutually exclusive in Chrome: under replMode a returned promise is NOT
+        // awaited (it serialises to `{}`), which breaks `fetch(...).then(...)` and
+        // every other async eval. So enable replMode ONLY for synchronous scripts
+        // that declare a top-level `let`/`const`; promise-returning scripts keep
+        // `awaitPromise` (no replMode) — exactly the pre-#38 behaviour.
+        let mentions_async = script.contains("await")
+            || script.contains(".then(")
+            || script.contains("fetch(")
+            || script.contains("Promise");
+        let declares = script.contains("let ") || script.contains("const ");
+        let repl_mode = declares && !mentions_async;
         let result: EvaluateResult = self
             .client
             .send_command_typed(
                 "Runtime.evaluate",
-                &EvaluateParams {
-                    expression: script.to_string(),
-                    return_by_value: Some(true),
-                    await_promise: Some(true),
-                },
+                &json!({
+                    "expression": script,
+                    "returnByValue": true,
+                    "awaitPromise": !repl_mode,
+                    "replMode": repl_mode,
+                }),
                 Some(&session_id),
             )
             .await?;
@@ -1494,6 +1602,11 @@ impl BrowserManager {
                 title: sanitize_title(&target.title),
                 target_type: target.target_type.clone(),
             };
+            // A tab that appeared right after THIS session's action (a click that
+            // opened a popup/new tab) is ours — record it as owned so it's tracked,
+            // protected from churn-pruning, and cleaned up on close, consistent with
+            // strict multi-agent isolation (we only ever own tabs we created/opened).
+            self.created_targets.insert(target.target_id.clone());
             self.add_background_page(page.clone());
             let _ = self.enable_domains(&attach.session_id).await;
             if opened.is_none() {
@@ -1521,13 +1634,19 @@ impl BrowserManager {
             .filter(should_track_target)
             .collect();
         let live_ids: HashSet<String> = live.iter().map(|t| t.target_id.clone()).collect();
+        let on_relay = self.agent_group().is_some();
 
         for target in &live {
             if self.update_page_target_info(target) {
                 continue;
             }
-            // A target this session hasn't tracked yet — attach and add it in the
-            // background so it's listable/adoptable without stealing the active tab.
+            // STRICT MULTI-AGENT ISOLATION: on the relay (the user's real Chrome,
+            // shared with other agents), NEVER adopt a tab this session didn't
+            // create — it belongs to the user or another agent's group. Only a
+            // browser we launched (every tab ours) adopts unknown targets.
+            if on_relay {
+                continue;
+            }
             let attach_result: AttachToTargetResult = match self
                 .client
                 .send_command_typed(
@@ -1558,12 +1677,26 @@ impl BrowserManager {
             let _ = self.enable_domains(&attach_result.session_id).await;
         }
 
-        // Drop tabs that no longer exist so `tab list` doesn't show phantom rows —
-        // but never prune the explicitly-pinned active target on a transient
-        // getTargets snapshot (issue #31; see `prunable_target_ids`).
-        let gone = prunable_target_ids(&self.pages, &live_ids, self.active_target_id.as_deref());
-        for tid in gone {
-            self.remove_page_by_target_id(&tid);
+        // Prune tabs that are gone. On a LAUNCHED browser a missing target really
+        // is closed, so prune immediately. On the RELAY a single `getTargets`
+        // snapshot routinely omits live tabs (multi-agent churn, a brief
+        // cross-process-nav gap) — dropping the tab we're driving on one bad
+        // snapshot is the failure we're fixing — so prune only after the tab has
+        // been absent for several CONSECUTIVE snapshots (debounced). The pinned
+        // active target is protected either way (issue #31).
+        let gone = if on_relay {
+            debounced_prune_ids(
+                &self.pages,
+                &live_ids,
+                self.active_target_id.as_deref(),
+                &mut self.relay_target_misses,
+            )
+        } else {
+            prunable_target_ids(&self.pages, &live_ids, self.active_target_id.as_deref())
+        };
+        for tid in &gone {
+            self.relay_target_misses.remove(tid);
+            self.remove_page_by_target_id(tid);
         }
 
         // Refresh url/title from each live tab. The relay only stamps target_info
@@ -2473,6 +2606,7 @@ async fn initialize_lightpanda_manager(
             visited_origins: HashSet::new(),
             created_targets: HashSet::new(),
             active_target_id: None,
+            relay_target_misses: HashMap::new(),
             next_tab_id: 1,
             capture_console: console_capture_enabled(),
         };
@@ -2766,7 +2900,9 @@ mod tests {
              its tab is gone (closed, navigated across processes, or lost after an extension \
              restart). Re-attach by re-opening your target URL before retrying."
         ));
-        assert!(is_stale_target_error("unknown sessionId cb-tab-7 for Page.navigate"));
+        assert!(is_stale_target_error(
+            "unknown sessionId cb-tab-7 for Page.navigate"
+        ));
         assert!(is_stale_target_error("no attached tab for Page.navigate"));
     }
 
@@ -2774,8 +2910,12 @@ mod tests {
     fn stale_target_error_ignores_unrelated_failures() {
         // A genuine navigation failure (bad URL, DNS, blocked) must NOT trigger
         // the open-a-fresh-tab recovery — that would mask the real error.
-        assert!(!is_stale_target_error("Navigation failed: net::ERR_NAME_NOT_RESOLVED"));
-        assert!(!is_stale_target_error("CDP command timed out: Page.navigate"));
+        assert!(!is_stale_target_error(
+            "Navigation failed: net::ERR_NAME_NOT_RESOLVED"
+        ));
+        assert!(!is_stale_target_error(
+            "CDP command timed out: Page.navigate"
+        ));
     }
 
     fn page(target_id: &str) -> PageInfo {
@@ -2882,7 +3022,10 @@ mod tests {
         let dirty = "\u{200d}\u{2061}\u{200d}\u{2063}\u{200b}\u{2062}\u{feff}GitHub";
         assert_eq!(sanitize_title(dirty), "GitHub");
         // Clean titles (incl. CJK + normal punctuation) pass through untouched.
-        assert_eq!(sanitize_title("購入手続きへ - メルカリ"), "購入手続きへ - メルカリ");
+        assert_eq!(
+            sanitize_title("購入手続きへ - メルカリ"),
+            "購入手続きへ - メルカリ"
+        );
         assert_eq!(sanitize_title("  Hello World  "), "Hello World");
         // Emoji and real content survive; only the invisibles are dropped.
         assert_eq!(sanitize_title("✓ Done\u{200b}"), "✓ Done");
@@ -2914,7 +3057,47 @@ mod tests {
         // A pinned target that IS in the live set is simply not prunable anyway.
         let mut live2 = HashSet::new();
         live2.insert("A".to_string());
-        assert_eq!(prunable_target_ids(&pages, &live2, Some("A")), vec!["B".to_string()]);
+        assert_eq!(
+            prunable_target_ids(&pages, &live2, Some("A")),
+            vec!["B".to_string()]
+        );
+    }
+
+    #[test]
+    fn debounced_prune_tolerates_transient_churn() {
+        // Multi-agent churn: a single getTargets snapshot omits our owned tab "B"
+        // (another agent opened/closed tabs). It must NOT be pruned on one miss.
+        let pages = vec![page("A"), page("B")];
+        let mut misses = HashMap::new();
+        let empty: HashSet<String> = HashSet::new();
+        // Misses 1 and 2: B absent but under threshold → not pruned.
+        assert!(debounced_prune_ids(&pages, &empty, Some("A"), &mut misses).is_empty());
+        assert!(debounced_prune_ids(&pages, &empty, Some("A"), &mut misses).is_empty());
+        // Miss 3 (== RELAY_PRUNE_MISSES): genuinely gone → pruned.
+        assert_eq!(
+            debounced_prune_ids(&pages, &empty, Some("A"), &mut misses),
+            vec!["B".to_string()]
+        );
+    }
+
+    #[test]
+    fn debounced_prune_resets_on_reappearance_and_protects_pin() {
+        let pages = vec![page("A"), page("B")];
+        let mut misses = HashMap::new();
+        let empty: HashSet<String> = HashSet::new();
+        let mut live_b: HashSet<String> = HashSet::new();
+        live_b.insert("B".to_string());
+        // Two misses for B, then it reappears → counter resets, so it survives
+        // indefinitely under intermittent churn.
+        debounced_prune_ids(&pages, &empty, Some("A"), &mut misses);
+        debounced_prune_ids(&pages, &empty, Some("A"), &mut misses);
+        assert!(debounced_prune_ids(&pages, &live_b, Some("A"), &mut misses).is_empty());
+        assert!(debounced_prune_ids(&pages, &empty, Some("A"), &mut misses).is_empty()); // back to miss 1
+                                                                                         // The pinned active "A" is never pruned no matter how many misses.
+        for _ in 0..5 {
+            let gone = debounced_prune_ids(&pages, &empty, Some("A"), &mut misses);
+            assert!(!gone.contains(&"A".to_string()));
+        }
     }
 
     #[test]

cli/src/native/e2e_tests.rs

@@ -35,6 +35,7 @@ fn native_test_fixture_html(name: &str) -> &'static str {
         "html5_drag_probe" => include_str!("test_fixtures/html5_drag_probe.html"),
         "pointer_capture_probe" => include_str!("test_fixtures/pointer_capture_probe.html"),
         "upload_probe" => include_str!("test_fixtures/upload_probe.html"),
+        "iframe_button_probe" => include_str!("test_fixtures/iframe_button_probe.html"),
         _ => panic!("Unknown native test fixture: {}", name),
     }
 }
@@ -573,6 +574,76 @@ async fn e2e_snapshot_and_click_ref() {
     assert_success(&resp);
 }
 
+/// Clicking a button INSIDE an iframe by `@ref` must deliver a TRUSTED activation
+/// (`event.isTrusted === true`), not a synthetic DOM `.click()`. Security-sensitive
+/// embedded forms (Google Payments' `保存`) reject `isTrusted:false` clicks, so an
+/// enabled submit button silently no-op'd (issue #39). The fix routes iframe-ref
+/// clicks to a real `Input.dispatchMouseEvent` on the element's own frame session.
+/// The fixture's iframe button writes `clicked:<isTrusted>` into its own text on
+/// click, which the cross-frame snapshot reads back.
+#[tokio::test]
+#[ignore]
+async fn e2e_iframe_button_click_is_trusted() {
+    let mut state = DaemonState::new();
+
+    let resp = execute_command(
+        &json!({ "id": "1", "action": "launch", "headless": true }),
+        &mut state,
+    )
+    .await;
+    assert_success(&resp);
+
+    let resp = execute_command(
+        &json!({ "id": "2", "action": "navigate", "url": native_test_fixture_url("iframe_button_probe") }),
+        &mut state,
+    )
+    .await;
+    assert_success(&resp);
+
+    // Snapshot (interactive) — the button lives in the iframe and must appear with
+    // a ref; that ref carries the frame_id so the click resolves into the frame.
+    let resp = execute_command(
+        &json!({ "id": "3", "action": "snapshot", "interactive": true }),
+        &mut state,
+    )
+    .await;
+    assert_success(&resp);
+    let snapshot = get_data(&resp)["snapshot"].as_str().unwrap_or("");
+    let ref_id = snapshot
+        .lines()
+        .find(|l| l.contains("button \"save\""))
+        .and_then(|l| l.split("ref=").nth(1))
+        .map(|r| r.trim_end_matches(']').trim())
+        .unwrap_or_else(|| panic!("iframe button not found in snapshot:\n{snapshot}"));
+
+    // Click it by ref.
+    let resp = execute_command(
+        &json!({ "id": "4", "action": "click", "selector": ref_id }),
+        &mut state,
+    )
+    .await;
+    assert_success(&resp);
+
+    tokio::time::sleep(tokio::time::Duration::from_millis(300)).await;
+
+    // The button rewrote its own text with the click's isTrusted flag; read it
+    // back across frames.
+    let resp = execute_command(
+        &json!({ "id": "5", "action": "snapshot", "interactive": true }),
+        &mut state,
+    )
+    .await;
+    assert_success(&resp);
+    let after = get_data(&resp)["snapshot"].as_str().unwrap_or("");
+    assert!(
+        after.contains("clicked:true"),
+        "iframe button click must be trusted (isTrusted:true); snapshot:\n{after}"
+    );
+
+    let resp = execute_command(&json!({ "id": "99", "action": "close" }), &mut state).await;
+    assert_success(&resp);
+}
+
 // ---------------------------------------------------------------------------
 // Screenshot
 // ---------------------------------------------------------------------------

cli/src/native/element.rs

@@ -100,6 +100,15 @@ impl RefMap {
         self.map.get(ref_id)
     }
 
+    /// Whether `selector_or_ref` is a `@ref` whose snapshot entry lives inside an
+    /// iframe (has a `frame_id`). Pointer interactions use this to choose
+    /// DOM-dispatch over coordinates for OOPIF elements (issue #36).
+    pub fn ref_is_in_iframe(&self, selector_or_ref: &str) -> bool {
+        parse_ref(selector_or_ref)
+            .and_then(|r| self.map.get(&r).map(|e| e.frame_id.is_some()))
+            .unwrap_or(false)
+    }
+
     pub fn entries_sorted(&self) -> Vec<(String, RefEntry)> {
         let mut entries = self
             .map
@@ -1019,7 +1028,9 @@ async fn eval_text_in_frame(client: &CdpClient, session_id: &str, frame_id: &str
         .await
         .ok()
         .and_then(|v| v.get("executionContextId").and_then(|c| c.as_i64()));
-    let Some(ctx_id) = ctx else { return String::new() };
+    let Some(ctx_id) = ctx else {
+        return String::new();
+    };
     let res = client
         .send_command(
             "Runtime.evaluate",
@@ -1090,7 +1101,10 @@ pub async fn collect_all_frames_text(
         let (kind, text) = if is_top {
             ("top", eval_text_default(client, top_session).await)
         } else {
-            ("inline", eval_text_in_frame(client, top_session, &fid).await)
+            (
+                "inline",
+                eval_text_in_frame(client, top_session, &fid).await,
+            )
         };
         out.push(FrameText {
             frame_id: fid,

cli/src/native/interaction.rs

@@ -7,6 +7,17 @@ use super::cdp::types::*;
 use super::element::{parse_ref, resolve_element_center, resolve_element_object_id, RefMap};
 use super::humanize;
 
+/// Whether a pointer interaction should be DOM-dispatched (invoke the event on
+/// the element in its own session) rather than dispatched at a viewport
+/// coordinate via `Input.dispatchMouseEvent`. True when the target is inside an
+/// iframe (an OOPIF element's box can't be mapped to a top-viewport point) or we
+/// drive over the extension relay (a coordinate Input event isn't confined to the
+/// target tab on a busy real Chrome — it drifts onto the foreground tab; issues
+/// #31/#36). DOM-dispatch always hits the right element in the right tab.
+fn prefer_dom_dispatch(ref_map: &RefMap, selector_or_ref: &str) -> bool {
+    ref_map.ref_is_in_iframe(selector_or_ref) || crate::connect::relay_url().is_some()
+}
+
 pub async fn click(
     client: &CdpClient,
     session_id: &str,
@@ -45,29 +56,43 @@ pub async fn click(
         .await;
     }
 
-    // Over the extension relay we drive the user's real, in-use Chrome, where a
-    // coordinate `Input.dispatchMouseEvent` is NOT reliably confined to our target
-    // tab — it can be delivered to whatever tab is in the foreground, and an OOPIF
-    // element's box can't be mapped to a top-viewport point at all. This twice
-    // opened an unrelated tab on the user's busy Chrome (issues #31/#36). So on the
-    // relay, never use coordinates for a normal left click: DOM-dispatch invokes
-    // the element's click in its own (frame) session, always hitting the right
-    // element in the right tab. Double/right clicks still need true pointer
-    // semantics, and `coord` mode is an explicit opt-out.
-    if mode != "coord" && button == "left" && click_count == 1 {
-        let in_iframe = parse_ref(selector_or_ref)
-            .and_then(|r| ref_map.get(&r).map(|e| e.frame_id.is_some()))
-            .unwrap_or(false);
-        if in_iframe || crate::connect::relay_url().is_some() {
-            return dom_click(
-                client,
-                session_id,
-                ref_map,
-                selector_or_ref,
-                iframe_sessions,
-            )
-            .await;
-        }
+    // An element INSIDE an iframe needs a TRUSTED activation: a DOM `.click()` is
+    // `isTrusted:false`, which security-sensitive embedded forms reject — Google
+    // Payments' enabled `保存` button silently no-ops on a synthetic click (issue
+    // #39). A coordinate `Input.dispatchMouseEvent` can't help either: `getBoxModel`
+    // for a sub-frame node returns frame-local coordinates that don't compose the
+    // iframe's offset, so the click lands in the wrong place. The frame-agnostic
+    // trusted path is keyboard activation — focus the element in its own frame, then
+    // dispatch a real Enter on the page session; Chrome routes the key to the
+    // focused element regardless of frame (same as `type --focused`), and Enter on a
+    // focused button/link fires a trusted `click`. `coord` mode opts out.
+    let in_iframe = ref_map.ref_is_in_iframe(selector_or_ref);
+    if mode != "coord" && button == "left" && click_count == 1 && in_iframe {
+        return dom_activate(
+            client,
+            session_id,
+            ref_map,
+            selector_or_ref,
+            iframe_sessions,
+        )
+        .await;
+    }
+    // On the relay (the user's real Chrome) a TOP-document coordinate click used to
+    // drift onto the foreground tab; that root cause is fixed (#5: the agent drives
+    // its own pinned tab), but DOM-dispatch stays the conservative default here.
+    if mode != "coord"
+        && button == "left"
+        && click_count == 1
+        && crate::connect::relay_url().is_some()
+    {
+        return dom_click(
+            client,
+            session_id,
+            ref_map,
+            selector_or_ref,
+            iframe_sessions,
+        )
+        .await;
     }
 
     let resolved = resolve_element_center(
@@ -260,6 +285,112 @@ async fn dom_click(
     Ok(())
 }
 
+/// Trusted activation of an element inside an iframe (issue #39). Focuses the
+/// element in its own frame session, then dispatches a real Enter/Space on the
+/// page session — Chrome routes the key to the focused element across frames, and
+/// Enter/Space on a focused button/link/checkbox fires a `click` with
+/// `isTrusted: true`, which security-sensitive embedded forms (Google Payments
+/// `保存`) require. Non-activatable roles (a `div[onclick]`) can't be keyboard-
+/// activated, so they fall back to a DOM `.click()`.
+async fn dom_activate(
+    client: &CdpClient,
+    session_id: &str,
+    ref_map: &RefMap,
+    selector_or_ref: &str,
+    iframe_sessions: &HashMap<String, String>,
+) -> Result<(), String> {
+    let role = parse_ref(selector_or_ref)
+        .and_then(|r| ref_map.get(&r).map(|e| e.role.clone()))
+        .unwrap_or_default();
+    // Space toggles checkbox-like controls; Enter activates buttons/links/menus.
+    let key = match role.as_str() {
+        "checkbox" | "radio" | "switch" | "option" | "menuitemcheckbox" | "menuitemradio" => {
+            Some("space")
+        }
+        "button" | "link" | "menuitem" | "tab" | "treeitem" => Some("enter"),
+        _ => None,
+    };
+    let Some(key) = key else {
+        // Not keyboard-activatable — best effort via DOM .click() (untrusted).
+        return dom_click(
+            client,
+            session_id,
+            ref_map,
+            selector_or_ref,
+            iframe_sessions,
+        )
+        .await;
+    };
+
+    let (object_id, effective_session_id) = resolve_element_object_id(
+        client,
+        session_id,
+        ref_map,
+        selector_or_ref,
+        iframe_sessions,
+    )
+    .await?;
+    // Focus the element in its OWN frame session so the keystroke lands on it.
+    client
+        .send_command_typed::<_, Value>(
+            "Runtime.callFunctionOn",
+            &CallFunctionOnParams {
+                function_declaration: "function() { this.focus(); }".to_string(),
+                object_id: Some(object_id),
+                arguments: None,
+                return_by_value: Some(true),
+                await_promise: Some(false),
+            },
+            Some(&effective_session_id),
+        )
+        .await?;
+    // Trusted key on the page session — routed to the focused (in-frame) element.
+    press_key(client, session_id, key).await?;
+    wait_for_paint_settled(client, &effective_session_id).await;
+    Ok(())
+}
+
+/// DOM-dispatch a double-click on the element in its own session (no coordinates)
+/// — the relay/iframe-safe counterpart to a coordinate dblclick. Fires the full
+/// click,click,dblclick sequence so handlers bound to any of them respond.
+async fn dom_dblclick(
+    client: &CdpClient,
+    session_id: &str,
+    ref_map: &RefMap,
+    selector_or_ref: &str,
+    iframe_sessions: &HashMap<String, String>,
+) -> Result<(), String> {
+    let (object_id, effective_session_id) = resolve_element_object_id(
+        client,
+        session_id,
+        ref_map,
+        selector_or_ref,
+        iframe_sessions,
+    )
+    .await?;
+    client
+        .send_command_typed::<_, Value>(
+            "Runtime.callFunctionOn",
+            &CallFunctionOnParams {
+                function_declaration: r#"function() {
+                    const opts = { bubbles: true, cancelable: true, view: window };
+                    this.dispatchEvent(new MouseEvent('click', opts));
+                    this.dispatchEvent(new MouseEvent('click', { ...opts, detail: 2 }));
+                    this.dispatchEvent(new MouseEvent('dblclick', opts));
+                }"#
+                .to_string(),
+                object_id: Some(object_id),
+                arguments: None,
+                return_by_value: Some(true),
+                await_promise: Some(false),
+            },
+            Some(&effective_session_id),
+        )
+        .await?;
+    wait_for_paint_settled(client, &effective_session_id).await;
+    Ok(())
+}
+
 pub async fn dblclick(
     client: &CdpClient,
     session_id: &str,
@@ -267,6 +398,20 @@ pub async fn dblclick(
     selector_or_ref: &str,
     iframe_sessions: &HashMap<String, String>,
 ) -> Result<(), String> {
+    // Same relay/iframe drift hazard as a single click — DOM-dispatch the
+    // double-click there instead of a coordinate one (issues #31/#36).
+    if std::env::var("AGENT_BROWSER_CLICK_MODE").as_deref() != Ok("coord")
+        && prefer_dom_dispatch(ref_map, selector_or_ref)
+    {
+        return dom_dblclick(
+            client,
+            session_id,
+            ref_map,
+            selector_or_ref,
+            iframe_sessions,
+        )
+        .await;
+    }
     click(
         client,
         session_id,
@@ -279,6 +424,50 @@ pub async fn dblclick(
     .await
 }
 
+/// DOM-dispatch a hover (pointer/mouse enter+move) on the element in its own
+/// session — reaches OOPIF elements and never drifts to the foreground tab over
+/// the relay, unlike a coordinate `mouseMoved` (issues #31/#36).
+async fn dom_hover(
+    client: &CdpClient,
+    session_id: &str,
+    ref_map: &RefMap,
+    selector_or_ref: &str,
+    iframe_sessions: &HashMap<String, String>,
+) -> Result<(), String> {
+    let (object_id, effective_session_id) = resolve_element_object_id(
+        client,
+        session_id,
+        ref_map,
+        selector_or_ref,
+        iframe_sessions,
+    )
+    .await?;
+    client
+        .send_command_typed::<_, Value>(
+            "Runtime.callFunctionOn",
+            &CallFunctionOnParams {
+                function_declaration: r#"function() {
+                    const r = this.getBoundingClientRect();
+                    const cx = r.left + r.width / 2, cy = r.top + r.height / 2;
+                    const base = { bubbles: true, cancelable: true, view: window, clientX: cx, clientY: cy };
+                    this.dispatchEvent(new PointerEvent('pointerover', base));
+                    this.dispatchEvent(new PointerEvent('pointerenter', { ...base, bubbles: false }));
+                    this.dispatchEvent(new MouseEvent('mouseover', base));
+                    this.dispatchEvent(new MouseEvent('mouseenter', { ...base, bubbles: false }));
+                    this.dispatchEvent(new MouseEvent('mousemove', base));
+                }"#
+                .to_string(),
+                object_id: Some(object_id),
+                arguments: None,
+                return_by_value: Some(true),
+                await_promise: Some(false),
+            },
+            Some(&effective_session_id),
+        )
+        .await?;
+    Ok(())
+}
+
 pub async fn hover(
     client: &CdpClient,
     session_id: &str,
@@ -286,6 +475,18 @@ pub async fn hover(
     selector_or_ref: &str,
     iframe_sessions: &HashMap<String, String>,
 ) -> Result<(), String> {
+    // Coordinate `mouseMoved` drifts to the foreground tab over the relay and
+    // can't reach an OOPIF — DOM-dispatch the hover there (issues #31/#36).
+    if prefer_dom_dispatch(ref_map, selector_or_ref) {
+        return dom_hover(
+            client,
+            session_id,
+            ref_map,
+            selector_or_ref,
+            iframe_sessions,
+        )
+        .await;
+    }
     let (x, y, _w, _h, effective_session_id) = resolve_element_center(
         client,
         session_id,
@@ -314,6 +515,63 @@ pub async fn hover(
     Ok(())
 }
 
+/// DOM-dispatch an HTML5 drag-and-drop from `source` to `target` in their shared
+/// session — the relay/iframe-safe counterpart to the coordinate drag, which
+/// drifts to the foreground tab over the relay and can't reach an OOPIF (issues
+/// #31/#36). Covers HTML5 DnD (sortable lists, file/card boards); pointer-driven
+/// drag (canvas, sliders) still needs the coordinate path. Errors if source and
+/// target live in different frames — a synthetic cross-frame DnD isn't reliable.
+pub async fn dom_drag(
+    client: &CdpClient,
+    session_id: &str,
+    ref_map: &RefMap,
+    source: &str,
+    target: &str,
+    iframe_sessions: &HashMap<String, String>,
+) -> Result<(), String> {
+    let (src_obj, src_session) =
+        resolve_element_object_id(client, session_id, ref_map, source, iframe_sessions).await?;
+    let (tgt_obj, tgt_session) =
+        resolve_element_object_id(client, session_id, ref_map, target, iframe_sessions).await?;
+    if src_session != tgt_session {
+        return Err(
+            "drag source and target are in different frames; cross-frame drag-and-drop over the \
+             relay isn't supported — drag within a single frame, or use a launched browser with \
+             AGENT_BROWSER_CLICK_MODE=coord"
+                .to_string(),
+        );
+    }
+    client
+        .send_command_typed::<_, Value>(
+            "Runtime.callFunctionOn",
+            &CallFunctionOnParams {
+                function_declaration: r#"function(target) {
+                    const dt = new DataTransfer();
+                    const ev = (type, el) => el.dispatchEvent(
+                        new DragEvent(type, { bubbles: true, cancelable: true, dataTransfer: dt }));
+                    ev('dragstart', this);
+                    ev('drag', this);
+                    ev('dragenter', target);
+                    ev('dragover', target);
+                    ev('drop', target);
+                    ev('dragend', this);
+                }"#
+                .to_string(),
+                object_id: Some(src_obj),
+                arguments: Some(vec![CallArgument {
+                    value: None,
+                    object_id: Some(tgt_obj),
+                }]),
+                return_by_value: Some(true),
+                await_promise: Some(false),
+            },
+            Some(&src_session),
+        )
+        .await?;
+    wait_for_paint_settled(client, &src_session).await;
+    Ok(())
+}
+
 pub async fn fill(
     client: &CdpClient,
     session_id: &str,
@@ -397,6 +655,7 @@ pub async fn type_text(
     clear: bool,
     delay_ms: Option<u64>,
     iframe_sessions: &HashMap<String, String>,
+    key_events: bool,
 ) -> Result<(), String> {
     let (object_id, effective_session_id) = resolve_element_object_id(
         client,
@@ -443,7 +702,7 @@ pub async fn type_text(
             .await?;
     }
 
-    type_text_into_active_context(client, session_id, text, delay_ms).await
+    type_text_into_active_context(client, session_id, text, delay_ms, key_events).await
 }
 
 pub async fn type_text_into_active_context(
@@ -451,6 +710,7 @@ pub async fn type_text_into_active_context(
     session_id: &str,
     text: &str,
     delay_ms: Option<u64>,
+    key_events: bool,
 ) -> Result<(), String> {
     // Per-character timing: an explicit `delay_ms` wins (caller asked for a
     // fixed cadence); otherwise fall back to humanize — variable, human-like
@@ -500,6 +760,46 @@ pub async fn type_text_into_active_context(
                     Some(session_id),
                 )
                 .await?;
+        } else if key_events {
+            // Real keystrokes (keyDown+keyUp carrying `text`) for autocomplete /
+            // combobox widgets that only react to key events and ignore the
+            // `input` that `Input.insertText` fires — e.g. Google's address
+            // postal-code → city/prefecture lookup (issue #36 / #4). The keyDown's
+            // `text` still inserts the character, so the field also fills.
+            let (key, code, key_code) = char_to_key_info(ch);
+            let s = ch.to_string();
+            client
+                .send_command_typed::<_, Value>(
+                    "Input.dispatchKeyEvent",
+                    &DispatchKeyEventParams {
+                        event_type: "keyDown".to_string(),
+                        key: Some(key.clone()),
+                        code: Some(code.clone()),
+                        text: Some(s.clone()),
+                        unmodified_text: Some(s),
+                        windows_virtual_key_code: Some(key_code),
+                        native_virtual_key_code: Some(key_code),
+                        modifiers: None,
+                    },
+                    Some(session_id),
+                )
+                .await?;
+            client
+                .send_command_typed::<_, Value>(
+                    "Input.dispatchKeyEvent",
+                    &DispatchKeyEventParams {
+                        event_type: "keyUp".to_string(),
+                        key: Some(key),
+                        code: Some(code),
+                        text: None,
+                        unmodified_text: None,
+                        windows_virtual_key_code: Some(key_code),
+                        native_virtual_key_code: Some(key_code),
+                        modifiers: None,
+                    },
+                    Some(session_id),
+                )
+                .await?;
         } else {
             // VS Code/Electron webviews reject repeated dispatchKeyEvent calls
             // carrying printable `text`. Insert printable characters directly

cli/src/native/snapshot.rs

@@ -330,6 +330,13 @@ impl RoleNameTracker {
     }
 }
 
+/// Max iframe nesting depth `take_snapshot` expands. Embedded payment/checkout
+/// widgets nest a few frames deep (e.g. AdSense → payments.google.com → an inner
+/// form frame); expanding past the first level is what gives those inner refs a
+/// `frame_id` so clicks resolve into the right frame (issue #36). Capped to keep
+/// a pathological frame tree from blowing up the snapshot.
+const MAX_IFRAME_DEPTH: usize = 3;
+
 pub async fn take_snapshot(
     client: &CdpClient,
     session_id: &str,
@@ -337,6 +344,28 @@ pub async fn take_snapshot(
     ref_map: &mut RefMap,
     frame_id: Option<&str>,
     iframe_sessions: &HashMap<String, String>,
+) -> Result<String, String> {
+    take_snapshot_at_depth(
+        client,
+        session_id,
+        options,
+        ref_map,
+        frame_id,
+        iframe_sessions,
+        0,
+    )
+    .await
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn take_snapshot_at_depth(
+    client: &CdpClient,
+    session_id: &str,
+    options: &SnapshotOptions,
+    ref_map: &mut RefMap,
+    frame_id: Option<&str>,
+    iframe_sessions: &HashMap<String, String>,
+    depth: usize,
 ) -> Result<String, String> {
     client
         .send_command_no_params("DOM.enable", Some(session_id))
@@ -606,10 +635,11 @@ pub async fn take_snapshot(
     }
 
     // Recurse into child iframes: for each Iframe node with a backend_node_id,
-    // resolve the child frame ID and take a snapshot of its content.
-    // We only recurse from the main frame (frame_id == None) to avoid
-    // unbounded depth; nested iframes within iframes are not expanded.
-    if frame_id.is_none() {
+    // resolve the child frame ID and snapshot its content. Recurse to
+    // MAX_IFRAME_DEPTH (not just the main frame) so refs inside nested
+    // payment/checkout widgets get a `frame_id` and clicks resolve into the right
+    // frame (issue #36); the cap bounds a pathological frame tree.
+    if depth < MAX_IFRAME_DEPTH {
         let mut iframe_snapshots: Vec<(String, String)> = Vec::new(); // (ref_id, child_snapshot)
         for node in tree_nodes.iter() {
             if node.role != "Iframe" || !node.has_ref {
@@ -622,13 +652,14 @@ pub async fn take_snapshot(
             if let Ok(child_fid) = resolve_iframe_frame_id(client, session_id, bid).await {
                 // Snapshot the child frame; errors are silently ignored
                 // (e.g. cross-origin iframes)
-                if let Ok(child_text) = Box::pin(take_snapshot(
+                if let Ok(child_text) = Box::pin(take_snapshot_at_depth(
                     client,
                     session_id,
                     options,
                     ref_map,
                     Some(&child_fid),
                     iframe_sessions,
+                    depth + 1,
                 ))
                 .await
                 {

cli/src/native/test_fixtures/iframe_button_probe.html

@@ -0,0 +1,28 @@
+<!doctype html>
+<html>
+  <head>
+    <meta charset="utf-8" />
+    <title>iframe button probe</title>
+  </head>
+  <body>
+    <h1>iframe button probe</h1>
+    <iframe
+      id="frame"
+      width="320"
+      height="140"
+      srcdoc="
+        <!doctype html>
+        <html>
+          <body style='margin:24px'>
+            <button id='b' style='padding:24px;font-size:22px'>save</button>
+            <script>
+              document.getElementById('b').addEventListener('click', function (e) {
+                this.textContent = 'clicked:' + e.isTrusted;
+              });
+            </script>
+          </body>
+        </html>
+      "
+    ></iframe>
+  </body>
+</html>

cli/src/output.rs

@@ -228,13 +228,28 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
         // because its response carries `url`/`title`, which later generic
         // renderers would otherwise swallow.
         if action == Some("cf_status") {
-            let challenged = data.get("challenged").and_then(|v| v.as_bool()).unwrap_or(false);
-            let rec = data.get("recommendation").and_then(|v| v.as_str()).unwrap_or("?");
+            let challenged = data
+                .get("challenged")
+                .and_then(|v| v.as_bool())
+                .unwrap_or(false);
+            let rec = data
+                .get("recommendation")
+                .and_then(|v| v.as_str())
+                .unwrap_or("?");
             let cl = data.get("clearance");
-            let present = cl.and_then(|c| c.get("present")).and_then(|v| v.as_bool()).unwrap_or(false);
-            let expired = cl.and_then(|c| c.get("expired")).and_then(|v| v.as_bool()).unwrap_or(false);
+            let present = cl
+                .and_then(|c| c.get("present"))
+                .and_then(|v| v.as_bool())
+                .unwrap_or(false);
+            let expired = cl
+                .and_then(|c| c.get("expired"))
+                .and_then(|v| v.as_bool())
+                .unwrap_or(false);
             let expires_in = cl.and_then(|c| c.get("expiresIn")).and_then(|v| v.as_i64());
-            let device = data.get("deviceVerified").and_then(|v| v.as_bool()).unwrap_or(false);
+            let device = data
+                .get("deviceVerified")
+                .and_then(|v| v.as_bool())
+                .unwrap_or(false);
 
             let (icon, headline) = match rec {
                 "proceed" => (color::success_indicator().to_string(), "cleared — no challenge, proceed"),
@@ -243,7 +258,10 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
                 _ => (color::cyan("•").to_string(), "unknown"),
             };
             println!("{} {}", icon, headline);
-            println!("  challenged:     {}", if challenged { "yes" } else { "no" });
+            println!(
+                "  challenged:     {}",
+                if challenged { "yes" } else { "no" }
+            );
             let cl_desc = if !present {
                 "absent".to_string()
             } else if expired {
@@ -254,7 +272,14 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
                 "present (session)".to_string()
             };
             println!("  cf_clearance:   {}", cl_desc);
-            println!("  device trusted: {}", if device { "yes (CF_VERIFIED_DEVICE)" } else { "no" });
+            println!(
+                "  device trusted: {}",
+                if device {
+                    "yes (CF_VERIFIED_DEVICE)"
+                } else {
+                    "no"
+                }
+            );
             return;
         }
 
@@ -382,7 +407,11 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
                 let count = list.len();
                 println!(
                     "{}",
-                    color::bold(&format!("{} frame{}", count, if count == 1 { "" } else { "s" }))
+                    color::bold(&format!(
+                        "{} frame{}",
+                        count,
+                        if count == 1 { "" } else { "s" }
+                    ))
                 );
                 for f in list {
                     let idx = f.get("index").and_then(|v| v.as_i64()).unwrap_or(0);
@@ -1464,6 +1493,12 @@ Usage: chrome-use type <selector> <text>
 Types text into the specified element character by character.
 Unlike fill, this does not clear existing content first.
 
+Options:
+  --key-events         Send real per-character keyDown/keyUp instead of
+  (alias --keys)       Input.insertText. Use for autocomplete / combobox fields
+                       that only react to key events — e.g. a postal-code box
+                       that auto-fills city/prefecture, or Google Places.
+
 Global Options:
   --json               Output as JSON
   --session <name>     Use specific session
@@ -1471,6 +1506,7 @@ Global Options:
 Examples:
   chrome-use type "#search" "hello"
   chrome-use type @e2 "additional text"
+  chrome-use type @e5 "201-0001" --key-events   # trigger the address autocomplete
 
 See Also:
   For typing into contenteditable editors (Lexical, ProseMirror, etc.)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "chrome-use",
-  "version": "1.5.11",
+  "version": "1.5.16",
   "description": "chrome-use — drive your real, logged-in Chrome from any AI agent, stealth by default",
   "type": "module",
   "packageManager": "pnpm@11.1.3",

skill-data/core/SKILL.md

@@ -36,6 +36,29 @@ Refs (`@e1`, `@e2`, ...) are assigned fresh on every snapshot. They become
 submits, dynamic re-renders, dialog opens. Always re-snapshot before your
 next ref interaction.
 
+> **Hard rule: snapshot-first, never screenshot-to-locate.** For form fields and
+> buttons, ALWAYS `snapshot -i` and act on refs/selectors. Do **not** reach for
+> `screenshot` + coordinate clicks to find or hit an element — `snapshot -i` now
+> pierces **cross-origin iframes** (embedded Google Payments / Stripe / checkout /
+> KYC forms) and lists their elements by `@ref`, including input values. Use
+> coordinates only for canvas/WebGL, or when `snapshot` genuinely returns nothing
+> for your target. Screenshots are for *visual verification you report*, never the
+> agent's own input — and a full-page `screenshot` of a real retina browser is
+> often too large for an image reader anyway. (If you ever feel you *need* a
+> screenshot to read state or locate something, that's a bug — please file it.)
+
+> **Snapshot-first, always. Never default to `screenshot` + coordinate clicking
+> for form fields or buttons.** Run `snapshot -i` and act on `@refs`. Use
+> coordinates only for canvas/WebGL, or when `snapshot` genuinely returns nothing
+> for your target. This holds **even inside cross-origin embedded iframes** —
+> since v1.5.12 `snapshot -i` pierces out-of-process iframes (Google Payments,
+> Stripe, embedded checkout/KYC) and lists their elements with refs, so
+> `click @e` / `type @e` / `fill @e` work directly. A screenshot is for a genuine
+> *visual* check you report to the user — not your own input. (Full-page
+> screenshots of a real retina Chrome are often too large for the image reader
+> anyway.) Driving off pixels on the relay also risks a coordinate event drifting
+> onto the user's foreground tab — refs never do. See issue #37.
+
 ## Before you automate: pick the cheapest tool
 
 Driving a browser is the heavy option. chrome-use earns its keep when you
@@ -108,7 +131,17 @@ Each `--session` that connects gets its **own colored Chrome tab group** (named
 after the session) and drives only its own tabs — multiple agents share the one
 real browser without cross-talk, and the user's own tabs are never grouped. CDP
 drives the page without moving the user's mouse/keyboard, so it doesn't fight
-them for control. **Anti-detection ranking: this real logged-in Chrome (extension
+them for control.
+
+**Strict multi-agent isolation.** A session over the relay tracks and drives
+**only the tabs it created** (its own group) plus pop-ups its own clicks open. It
+does **not** adopt the user's existing tabs or other agents' tabs, so several
+agents (and other tools opening tabs) can work in the same real Chrome
+concurrently without ever dropping or stealing each other's tabs — another
+agent's tab churn can't make your bound tab vanish or drift your commands onto the
+wrong page. Consequence: `tab list` shows only *your* session's tabs; to drive a
+specific pre-existing tab, navigate to it in your own tab instead of expecting it
+in the list. **Anti-detection ranking: this real logged-in Chrome (extension
 connect) > a headed launched browser > headless (forbidden).** A genuine human
 browser has no headless/automation tells at all, so prefer it for anything
 anti-bot-sensitive.
@@ -264,6 +297,10 @@ chrome-use hover @e1                   # hover
 chrome-use focus @e1                   # focus (useful before keyboard input)
 chrome-use fill @e2 "hello"            # clear then type
 chrome-use type @e2 " world"           # type without clearing
+chrome-use type @e5 "201-0001" --key-events  # real keystrokes (not insertText) —
+                                          # use for autocomplete/combobox fields that
+                                          # only react to key events (e.g. a postal box
+                                          # that auto-fills city/prefecture, Google Places)
 chrome-use press Enter                 # press a key at current focus (down+up)
 chrome-use press Control+a             # key combination
 chrome-use keydown d                   # HOLD a key down (no auto-release)
@@ -295,6 +332,18 @@ chrome-use scrollintoview @e1          # scroll element into view
 chrome-use drag @e1 @e2                # drag and drop
 ```
 
+**Cross-origin iframes (embedded payment / checkout / KYC widgets — Google
+Payments, Stripe, etc.) — drive them by ref, never by screenshot.** `snapshot -i`
+pierces these out-of-process iframes and lists their elements by `@ref`
+(including input values); `get text --all-frames` reads their text. Then just act
+on the refs: `click @e`, `type @e`, `hover @e`, `dblclick @e`, `drag @a @b` all
+work into the iframe. Over the extension relay these are dispatched through the
+DOM (in the element's own frame), so they hit the right element in the right tab
+— a coordinate click/scroll there can drift onto whatever tab is in the
+foreground, so prefer refs. For below-the-fold content in such a frame, scroll it
+with `scroll down N --at x,y` (a pixel over the frame) or `--frame n`. For a
+postal/autocomplete box inside the frame, `type @e "…" --key-events`.
+
 ### When refs don't work or you don't want to snapshot
 
 Use semantic locators: