chore(release): 1.5.12 — relay-safe hover/dblclick/drag, deeper iframe snapshot, key-events typing (#37)

Follow-up to #36 — make the whole interaction surface reach cross-origin OOPIFs
and stop coordinate events drifting onto the user's foreground tab over the relay.

- hover/dblclick/drag now DOM-dispatch over the relay or into an iframe (like
  click already did): a coordinate Input event isn't confined to the target tab
  on a busy real Chrome and can't map an OOPIF element's box to a top-viewport
  point. drag does an HTML5 DnD in the element's frame; cross-frame drag errors
  loudly instead of drifting.
- snapshot recurses iframes to MAX_IFRAME_DEPTH (3) instead of one level, so refs
  inside nested payment/checkout widgets get a frame_id and resolve into the
  right frame.
- relay tab adoption merges several Target.getTargets snapshots — a single flaky
  relay snapshot was dropping live tabs (a driven tab vanished after restart).
- `type --key-events` (alias --keys) sends real per-character keyDown/keyUp
  instead of Input.insertText, so autocomplete/combobox widgets that ignore the
  insertText input event fire (Google address postal lookup; commits Angular
  reactive forms so Save enables).
- SKILL: hard "snapshot-first, never default to screenshot+coordinates" rule;
  snapshot -i pierces cross-origin iframes since v1.5.12; cross-origin iframe
  driving guidance (#37).

cli/Cargo.lock

@@ -290,7 +290,7 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 
 [[package]]
 name = "chrome-use"
-version = "1.5.3"
+version = "1.5.12"
 dependencies = [
  "aes",
  "aes-gcm",

cli/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "chrome-use"
-version = "1.5.3"
+version = "1.5.12"
 edition = "2021"
 description = "Fast browser automation CLI for AI agents"
 license = "Apache-2.0"

cli/src/commands.rs

@@ -525,20 +525,29 @@ 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:
@@ -729,10 +738,56 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
                         } else {
                             return Err(ParseError::MissingArguments {
                                 context: "scroll --selector".to_string(),
-                                usage: "scroll [direction] [amount] [--selector <sel>]",
+                                usage: "scroll [direction] [amount] [--selector <sel>] [--at <x,y>] [--frame <n>]",
                             });
                         }
                     }
+                    "--at" => {
+                        // `--at x,y`: dispatch the wheel at this viewport pixel, so it
+                        // scrolls whatever element/iframe is under the pointer — including
+                        // cross-origin iframes that `window.scrollBy` can't reach (#36).
+                        let val = rest.get(i + 1).ok_or(ParseError::MissingArguments {
+                            context: "scroll --at".to_string(),
+                            usage: "scroll [direction] [amount] --at <x,y>",
+                        })?;
+                        let mut parts = val.split(',');
+                        match (
+                            parts.next().and_then(|s| s.trim().parse::<f64>().ok()),
+                            parts.next().and_then(|s| s.trim().parse::<f64>().ok()),
+                        ) {
+                            (Some(x), Some(y)) => {
+                                obj.insert("at".to_string(), json!([x, y]));
+                            }
+                            _ => {
+                                return Err(ParseError::InvalidValue {
+                                    message: format!("scroll --at: invalid coordinate `{}`", val),
+                                    usage: "scroll [direction] [amount] --at <x,y> (e.g. --at 640,400)",
+                                })
+                            }
+                        }
+                        i += 1;
+                    }
+                    "--frame" => {
+                        // `--frame n`: scroll the n-th frame from `chrome-use frames` by
+                        // dispatching the wheel at that frame's center — reaches content in
+                        // a cross-origin iframe without needing a selector into it (#36).
+                        let val = rest.get(i + 1).ok_or(ParseError::MissingArguments {
+                            context: "scroll --frame".to_string(),
+                            usage: "scroll [direction] [amount] --frame <n>",
+                        })?;
+                        match val.trim().parse::<usize>() {
+                            Ok(n) => {
+                                obj.insert("frame".to_string(), json!(n));
+                            }
+                            Err(_) => {
+                                return Err(ParseError::InvalidValue {
+                                    message: format!("scroll --frame: invalid index `{}`", val),
+                                    usage: "scroll [direction] [amount] --frame <n> (index from `chrome-use frames`)",
+                                })
+                            }
+                        }
+                        i += 1;
+                    }
                     arg if arg.starts_with('-') => {}
                     _ => {
                         match positional_index {
@@ -907,17 +962,37 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
             // selector: @ref or CSS selector
             // path: file path (contains / or . or ends with known extension)
             let mut full_page = false;
-            let positional: Vec<&str> = rest
-                .iter()
-                .filter(|arg| match **arg {
-                    "--full" | "-f" => {
-                        full_page = true;
-                        false
+            let mut clip: Option<Value> = None;
+            let mut positional: Vec<&str> = Vec::new();
+            let mut i = 0;
+            while i < rest.len() {
+                match rest[i] {
+                    "--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 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}'"),
+                                usage: "screenshot --clip <x,y,w,h> [path]",
+                            });
+                        }
+                        clip = Some(json!({
+                            "x": nums[0], "y": nums[1], "width": nums[2], "height": nums[3]
+                        }));
+                        i += 1;
                     }
-                    _ => true,
-                })
-                .copied()
-                .collect();
+                    other => positional.push(other),
+                }
+                i += 1;
+            }
             let (selector, path) = match (positional.first(), positional.get(1)) {
                 (Some(first), Some(second)) => {
                     // Two args: first is selector, second is path
@@ -948,6 +1023,9 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
                 "path": path, "selector": selector,
                 "fullPage": full_page, "annotate": flags.annotate
             });
+            if let Some(c) = clip {
+                cmd["clip"] = c;
+            }
             if let Some(ref fmt) = flags.screenshot_format {
                 cmd["format"] = json!(fmt);
             }
@@ -1078,6 +1156,14 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
             Ok(json!({ "id": id, "action": "stealth_status" }))
         }
 
+        // `cf-status` — Cloudflare challenge/clearance preflight: is the page
+        // currently a CF challenge, and is there a still-valid cf_clearance (the
+        // HttpOnly persistence cookie)? Lets an agent SKIP re-solving when already
+        // cleared, and know when it must solve. Persistence optimization.
+        "cf-status" | "cf" | "cloudflare-status" | "clearance" => {
+            Ok(json!({ "id": id, "action": "cf_status" }))
+        }
+
         // === Close ===
         "close" | "quit" | "exit" => {
             // `close <tab>` closes only that tab (and the output says "Tab
@@ -2455,6 +2541,15 @@ fn parse_get(rest: &[&str], id: &str) -> Result<Value, ParseError> {
             if all_frames {
                 return Ok(json!({ "id": id, "action": "gettext", "allFrames": true }));
             }
+            // `get text --pierce` reads through CLOSED shadow DOM / child docs
+            // via the CDP DOM tree — content eval/innerText can't reach, e.g. an
+            // extension's injected panel in a closed shadow root (issue #30).
+            let pierce = rest[1..]
+                .iter()
+                .any(|a| matches!(*a, "--pierce" | "--shadow" | "--deep"));
+            if pierce {
+                return Ok(json!({ "id": id, "action": "gettext", "pierce": true }));
+            }
             // `get text --main` returns the main-content region (readability),
             // skipping header/nav/footer/sidebar boilerplate (issue #27).
             let main = rest[1..]
@@ -4026,6 +4121,24 @@ 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]
@@ -4310,6 +4423,21 @@ mod tests {
         assert_eq!(cmd["fullPage"], true);
     }
 
+    #[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();
+        assert_eq!(cmd["action"], "screenshot");
+        assert_eq!(cmd["clip"]["x"], 10.0);
+        assert_eq!(cmd["clip"]["y"], 20.0);
+        assert_eq!(cmd["clip"]["width"], 200.0);
+        assert_eq!(cmd["clip"]["height"], 40.0);
+        assert_eq!(cmd["path"], "out.png");
+        // Bad clip is a clear error, not silent.
+        assert!(parse_command(&args("screenshot --clip 1,2,3"), &default_flags()).is_err());
+    }
+
     #[test]
     fn test_screenshot_with_ref() {
         let cmd = parse_command(&args("screenshot @e1"), &default_flags()).unwrap();
@@ -4897,6 +5025,16 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_get_text_pierce() {
+        for variant in ["get text --pierce", "get text --shadow", "text --deep"] {
+            let cmd = parse_command(&args(variant), &default_flags()).unwrap();
+            assert_eq!(cmd["action"], "gettext", "{variant}");
+            assert_eq!(cmd["pierce"], true, "{variant}");
+            assert!(cmd.get("selector").is_none(), "{variant}");
+        }
+    }
+
     #[test]
     fn test_tab_activate_flag() {
         let plain = parse_command(&args("tab t3"), &default_flags()).unwrap();
@@ -5803,6 +5941,35 @@ mod tests {
         assert_eq!(cmd["selector"], ".sidebar");
     }
 
+    #[test]
+    fn test_scroll_at_coordinate() {
+        // `--at x,y` carries a [x, y] array for a wheel dispatched at that pixel
+        // (issue #36: cross-origin iframe scroll).
+        let cmd = parse_command(&args("scroll down 700 --at 640,400"), &default_flags()).unwrap();
+        assert_eq!(cmd["action"], "scroll");
+        assert_eq!(cmd["direction"], "down");
+        assert_eq!(cmd["amount"], 700);
+        assert_eq!(cmd["at"], json!([640.0, 400.0]));
+    }
+
+    #[test]
+    fn test_scroll_at_rejects_garbage() {
+        assert!(parse_command(&args("scroll --at nope"), &default_flags()).is_err());
+        assert!(parse_command(&args("scroll --at 1"), &default_flags()).is_err());
+    }
+
+    #[test]
+    fn test_scroll_frame_index() {
+        let cmd = parse_command(&args("scroll down 700 --frame 2"), &default_flags()).unwrap();
+        assert_eq!(cmd["action"], "scroll");
+        assert_eq!(cmd["frame"], 2);
+    }
+
+    #[test]
+    fn test_scroll_frame_rejects_non_integer() {
+        assert!(parse_command(&args("scroll --frame two"), &default_flags()).is_err());
+    }
+
     #[test]
     fn test_scroll_selector_before_positional() {
         let cmd =

cli/src/connect.rs

@@ -449,6 +449,47 @@ pub fn relay_url() -> Option<String> {
     }
 }
 
+/// Append a one-line record of how a CDP connection was established, to
+/// `~/.chrome-use/connect-mode.log`. This is the smoking-gun detector for the
+/// "Allow remote debugging?" consent modal: that modal ONLY appears on a raw
+/// remote-debugging attach / a browser we launched with a debug port — NEVER on
+/// the extension relay. When the modal reappears, this log says which session
+/// took which path and when, so we can tell a code regression (`raw-port` /
+/// `launched` while the relay was up) from Chrome's own extension-debugger
+/// consent UX. Low volume (one line per connection); best-effort, never fails a
+/// connection.
+pub fn log_connect_mode(ws_url: &str, launched: bool, session: &str) {
+    let relay = relay_url();
+    let relay_up = relay.is_some();
+    let mode = if launched {
+        "launched(debug-port)"
+    } else if relay.as_deref() == Some(ws_url) {
+        "relay"
+    } else if ws_url.contains("127.0.0.1") || ws_url.contains("localhost") {
+        "raw-port-attach"
+    } else {
+        "remote-ws"
+    };
+    // A raw-port attach or a self-launch while the relay was available is the
+    // exact thing that pops the consent modal — flag it loudly in the line.
+    let suspect = (mode == "raw-port-attach" || launched) && relay_up;
+    let line = format!(
+        "session={session} mode={mode} relay_up={relay_up}{} ws={ws_url}\n",
+        if suspect { " CONSENT-MODAL-RISK" } else { "" }
+    );
+    if let Some(home) = dirs::home_dir() {
+        let path = home.join(".chrome-use").join("connect-mode.log");
+        use std::io::Write;
+        if let Ok(mut f) = std::fs::OpenOptions::new()
+            .create(true)
+            .append(true)
+            .open(&path)
+        {
+            let _ = f.write_all(line.as_bytes());
+        }
+    }
+}
+
 /// Sidecar recording the connected extension's version, written by the host when
 /// it receives the extension's `hello` (sibling of `relay-cdp-url`). Lets
 /// `doctor` surface which extension build is live without a CDP round-trip.

cli/src/main.rs

@@ -1369,6 +1369,20 @@ fn main() {
         && flags.provider.is_none()
         && (flags.force_launch || !flags.auto_connect)
     {
+        // Launching a debug-port Chrome pops Chrome's "Allow remote debugging?"
+        // consent modal (Chrome 136+). When the ab-connect relay is already up,
+        // this is almost always unintended — the relay drives the user's real
+        // Chrome with NO modal. Warn so the modal is self-explained and the
+        // caller (often a stray --launch / --no-auto-connect) is fixable (#32).
+        if !flags.json && connect::relay_url().is_some() {
+            eprintln!(
+                "{} launching a new Chrome with a debug port — this pops Chrome's \
+                 \"Allow remote debugging?\" modal.\n  The ab-connect relay is up; \
+                 drop --launch/--new (and don't pass --no-auto-connect) to drive your \
+                 real Chrome with no modal.",
+                color::warning_indicator()
+            );
+        }
         let mut launch_cmd = json!({
             "id": gen_id(),
             "action": "launch",

cli/src/native/actions.rs

@@ -1341,6 +1341,7 @@ pub async fn execute_command(cmd: &Value, state: &mut DaemonState) -> Value {
         "forward" => handle_forward(state).await,
         "reload" => handle_reload(state).await,
         "cookies_get" => handle_cookies_get(cmd, state).await,
+        "cf_status" => handle_cf_status(cmd, state).await,
         "cookies_set" => handle_cookies_set(cmd, state).await,
         "cookies_clear" => handle_cookies_clear(state).await,
         "storage_get" => handle_storage_get(cmd, state).await,
@@ -2999,6 +3000,14 @@ async fn handle_screenshot(cmd: &Value, state: &mut DaemonState) -> Result<Value
             .get("screenshotDir")
             .and_then(|v| v.as_str())
             .map(String::from),
+        clip: cmd.get("clip").and_then(|c| {
+            Some((
+                c.get("x")?.as_f64()?,
+                c.get("y")?.as_f64()?,
+                c.get("width")?.as_f64()?,
+                c.get("height")?.as_f64()?,
+            ))
+        }),
     };
 
     if annotate {
@@ -3216,6 +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
@@ -3227,7 +3244,8 @@ 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 }));
     }
 
@@ -3251,6 +3269,7 @@ async fn handle_type(cmd: &Value, state: &mut DaemonState) -> Result<Value, Stri
         clear,
         delay,
         &state.iframe_sessions,
+        key_events,
     )
     .await?;
     Ok(json!({ "typed": text }))
@@ -3458,17 +3477,169 @@ async fn handle_scroll(cmd: &Value, state: &mut DaemonState) -> Result<Value, St
         }
     }
 
-    interaction::scroll(
-        &mgr.client,
-        &session_id,
-        &state.ref_map,
-        selector,
-        dx,
-        dy,
-        &state.iframe_sessions,
-    )
-    .await?;
-    Ok(json!({ "scrolled": true }))
+    // An explicit `--selector` keeps the precise element-scroll path (scrollBy on
+    // the resolved node, same-origin only).
+    if let Some(sel) = selector {
+        interaction::scroll(
+            &mgr.client,
+            &session_id,
+            &state.ref_map,
+            Some(sel),
+            dx,
+            dy,
+            &state.iframe_sessions,
+        )
+        .await?;
+        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")
+    };
+
+    dispatch_wheel(&mgr.client, &session_id, x, y, dx, dy).await?;
+    Ok(json!({ "scrolled": true, "via": via, "at": [x, y] }))
+}
+
+/// Viewport center in CSS pixels, used as the default wheel landing point for
+/// `scroll` (issue #36). Falls back to a sane 640×400 center if the page can't
+/// be evaluated (e.g. a restricted document).
+async fn viewport_center(mgr: &BrowserManager, session_id: &str) -> Result<(f64, f64), String> {
+    let dims = mgr
+        .client
+        .send_command_typed::<_, Value>(
+            "Runtime.evaluate",
+            &super::cdp::types::EvaluateParams {
+                expression: "[window.innerWidth, window.innerHeight]".to_string(),
+                return_by_value: Some(true),
+                await_promise: Some(false),
+            },
+            Some(session_id),
+        )
+        .await
+        .ok();
+    let arr = dims
+        .as_ref()
+        .and_then(|v| v.get("result"))
+        .and_then(|v| v.get("value"))
+        .and_then(|v| v.as_array());
+    let w = arr
+        .and_then(|a| a.first())
+        .and_then(|v| v.as_f64())
+        .filter(|w| *w > 0.0)
+        .unwrap_or(1280.0);
+    let h = arr
+        .and_then(|a| a.get(1))
+        .and_then(|v| v.as_f64())
+        .filter(|h| *h > 0.0)
+        .unwrap_or(800.0);
+    Ok((w / 2.0, h / 2.0))
+}
+
+/// Center of the `n`-th frame (as listed by `chrome-use frames`) in top-viewport
+/// CSS pixels, so `scroll --frame n` lands its wheel inside a cross-origin iframe
+/// without needing a selector into it (issue #36). Resolves the frame's owning
+/// `<iframe>` element box via `DOM.getFrameOwner` + `DOM.getBoxModel` — exact for
+/// a frame nested directly under the top document; for a deeper nesting the box is
+/// relative to the intermediate frame, so prefer `--at x,y` from a screenshot.
+async fn frame_center(
+    mgr: &BrowserManager,
+    session_id: &str,
+    iframe_sessions: &HashMap<String, String>,
+    n: usize,
+) -> Result<(f64, f64), String> {
+    let frames =
+        super::element::collect_all_frames_text(&mgr.client, session_id, iframe_sessions).await?;
+    let frame = frames.get(n).ok_or_else(|| {
+        format!(
+            "frame index {} out of range (run `chrome-use frames`: {} frame(s))",
+            n,
+            frames.len()
+        )
+    })?;
+    if n == 0 {
+        // Frame 0 is the top document — there's no owner element; scroll its center.
+        return viewport_center(mgr, session_id).await;
+    }
+    let owner = mgr
+        .client
+        .send_command_typed::<_, Value>(
+            "DOM.getFrameOwner",
+            &json!({ "frameId": frame.frame_id }),
+            Some(session_id),
+        )
+        .await
+        .map_err(|e| format!("can't locate frame {}'s owner element: {}", n, e))?;
+    let backend_node_id = owner
+        .get("backendNodeId")
+        .and_then(|v| v.as_i64())
+        .ok_or_else(|| format!("frame {} has no owner <iframe> element", n))?;
+    let box_model = mgr
+        .client
+        .send_command_typed::<_, Value>(
+            "DOM.getBoxModel",
+            &json!({ "backendNodeId": backend_node_id }),
+            Some(session_id),
+        )
+        .await
+        .map_err(|e| format!("can't measure frame {}'s box: {}", n, e))?;
+    let content = box_model
+        .get("model")
+        .and_then(|m| m.get("content"))
+        .and_then(|c| c.as_array())
+        .ok_or_else(|| format!("frame {} box model has no content quad", n))?;
+    let coord = |i: usize| content.get(i).and_then(|v| v.as_f64()).unwrap_or(0.0);
+    // content quad is [x1,y1, x2,y2, x3,y3, x4,y4]; opposite corners are 0 and 2.
+    let cx = (coord(0) + coord(4)) / 2.0;
+    let cy = (coord(1) + coord(5)) / 2.0;
+    Ok((cx, cy))
+}
+
+/// Dispatch a trusted mouse wheel at `(x, y)`, humanized like `handle_wheel`.
+async fn dispatch_wheel(
+    client: &super::cdp::client::CdpClient,
+    session_id: &str,
+    x: f64,
+    y: f64,
+    delta_x: f64,
+    delta_y: f64,
+) -> Result<(), String> {
+    let level = humanize::active_level();
+    let seed = humanize::next_seed();
+    for (dx, dy, delay) in humanize::scroll_segments(delta_x, delta_y, level, seed) {
+        client
+            .send_command(
+                "Input.dispatchMouseEvent",
+                Some(json!({
+                    "type": "mouseWheel",
+                    "x": x,
+                    "y": y,
+                    "deltaX": dx,
+                    "deltaY": dy,
+                })),
+                Some(session_id),
+            )
+            .await?;
+        if !delay.is_zero() {
+            tokio::time::sleep(delay).await;
+        }
+    }
+    Ok(())
 }
 
 async fn handle_select(cmd: &Value, state: &mut DaemonState) -> Result<Value, String> {
@@ -3628,6 +3799,15 @@ async fn handle_gettext(cmd: &Value, state: &mut DaemonState) -> Result<Value, S
         }));
     }
 
+    // `get text --pierce` reads text through CLOSED shadow roots and child
+    // documents via the CDP DOM tree — content `innerText`/`eval` can't see,
+    // e.g. an extension's injected panel in a closed shadow DOM (#30).
+    if cmd.get("pierce").and_then(|v| v.as_bool()) == Some(true) {
+        let text = super::element::get_pierced_text(&mgr.client, &session_id).await?;
+        let url = mgr.get_url().await.unwrap_or_default();
+        return Ok(json!({ "text": text, "origin": url, "pierce": true }));
+    }
+
     // `get text --main` returns the page's main-content region (readability-lite),
     // skipping global header/nav/footer/sidebar boilerplate (#27).
     if cmd.get("main").and_then(|v| v.as_bool()) == Some(true) {
@@ -4113,6 +4293,108 @@ async fn handle_cookies_clear(state: &DaemonState) -> Result<Value, String> {
     Ok(json!({ "cleared": true }))
 }
 
+// Detect whether the active page is *currently* a Cloudflare challenge
+// (the full-page "Just a moment…" / "正在进行安全验证" interstitial), so an agent
+// knows whether it must solve or can proceed. Runs in the top frame main world.
+const CF_CHALLENGE_JS: &str = r#"(function(){
+  var t = document.title || '';
+  var challenged =
+    /just a moment|attention required|checking (your|if)|verify you are human|正在进行安全验证|安全验证|请稍候|请完成|人机验证/i.test(t) ||
+    !!document.querySelector('#challenge-form, #challenge-running, #cf-challenge-running, [id^="cf-chl"], script[src*="/cdn-cgi/challenge-platform/"]');
+  var turnstile = !!document.querySelector('.cf-turnstile, [data-sitekey]');
+  return JSON.stringify({ title: t, challenged: challenged, turnstile: turnstile, readyState: document.readyState });
+})()"#;
+
+/// Recommendation for a Cloudflare-gated page, from the current challenge state
+/// and whether a still-valid `cf_clearance` exists. Pure so it's unit-testable.
+/// - not challenged            → "proceed" (the page is cleared/loaded)
+/// - challenged, valid cookie  → "reissue" (clearance present but page still
+///   blocks → it's stale or the IP/UA no longer matches what it was issued for)
+/// - challenged, no cookie     → "solve"
+fn cf_recommendation(challenged: bool, clearance_valid: bool) -> &'static str {
+    if !challenged {
+        "proceed"
+    } else if clearance_valid {
+        "reissue"
+    } else {
+        "solve"
+    }
+}
+
+/// `cf_clearance` validity for a cookie's expiry (epoch seconds; <=0 = session
+/// cookie, treated as non-expiring). Returns (present, expired). Pure.
+fn clearance_state(expires: Option<f64>, now: f64) -> (bool, bool) {
+    match expires {
+        None => (false, false),
+        Some(e) if e <= 0.0 => (true, false), // session cookie: no expiry
+        Some(e) => (true, e < now),
+    }
+}
+
+async fn handle_cf_status(_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 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 = parse_json_string(probe_raw, "cf challenge probe").unwrap_or(Value::Null);
+    let challenged = probe
+        .get("challenged")
+        .and_then(|v| v.as_bool())
+        .unwrap_or(false);
+    let turnstile = probe
+        .get("turnstile")
+        .and_then(|v| v.as_bool())
+        .unwrap_or(false);
+    let title = probe
+        .get("title")
+        .and_then(|v| v.as_str())
+        .unwrap_or("")
+        .to_string();
+
+    // 2. Persistence artifacts: cf_clearance (HttpOnly → must read via CDP, not
+    //    document.cookie) + CF_VERIFIED_DEVICE. Scope to the current URL.
+    let urls = if url.is_empty() {
+        None
+    } else {
+        Some(vec![url.clone()])
+    };
+    let cookies = super::cookies::get_cookies(&mgr.client, &session_id, urls)
+        .await
+        .unwrap_or_default();
+    let clearance = cookies.iter().find(|c| c.name == "cf_clearance");
+    let now = std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .map(|d| d.as_secs_f64())
+        .unwrap_or(0.0);
+    let (present, expired) = clearance_state(clearance.map(|c| c.expires), now);
+    let expires_in = clearance
+        .filter(|_| present && !expired)
+        .map(|c| (c.expires - now).max(0.0) as i64);
+    let device_verified = cookies
+        .iter()
+        .any(|c| c.name.starts_with("CF_VERIFIED_DEVICE"));
+
+    let clearance_valid = present && !expired;
+    let recommendation = cf_recommendation(challenged, clearance_valid);
+
+    Ok(json!({
+        "url": url,
+        "title": title,
+        "challenged": challenged,
+        "turnstile": turnstile,
+        "clearance": {
+            "present": present,
+            "expired": expired,
+            "expiresIn": expires_in,
+            "httpOnly": clearance.map(|c| c.http_only).unwrap_or(false),
+        },
+        "deviceVerified": device_verified,
+        "recommendation": recommendation,
+    }))
+}
+
 async fn handle_storage_get(cmd: &Value, state: &DaemonState) -> Result<Value, String> {
     let mgr = state.browser.as_ref().ok_or("Browser not launched")?;
     let session_id = mgr.active_session_id()?.to_string();
@@ -4450,8 +4732,18 @@ 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") => {
@@ -6980,6 +7272,28 @@ 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,
@@ -7334,6 +7648,7 @@ async fn handle_diff_screenshot(cmd: &Value, state: &DaemonState) -> Result<Valu
         quality: None,
         annotate: false,
         output_dir: None,
+        clip: None,
     };
 
     let result = screenshot::take_screenshot(
@@ -9060,6 +9375,26 @@ mod tests {
     use crate::test_utils::EnvGuard;
     use std::fs;
 
+    #[test]
+    fn test_cf_recommendation() {
+        assert_eq!(cf_recommendation(false, false), "proceed");
+        assert_eq!(cf_recommendation(false, true), "proceed");
+        assert_eq!(cf_recommendation(true, false), "solve");
+        assert_eq!(cf_recommendation(true, true), "reissue");
+    }
+
+    #[test]
+    fn test_clearance_state() {
+        // no cookie
+        assert_eq!(clearance_state(None, 1000.0), (false, false));
+        // session cookie (expires <= 0) → present, never expired
+        assert_eq!(clearance_state(Some(-1.0), 1000.0), (true, false));
+        // valid: expiry in the future
+        assert_eq!(clearance_state(Some(2000.0), 1000.0), (true, false));
+        // expired: expiry in the past
+        assert_eq!(clearance_state(Some(500.0), 1000.0), (true, true));
+    }
+
     #[test]
     fn test_url_glob_to_regex() {
         assert_eq!(url_glob_to_regex("**/dashboard"), "^.*/dashboard$");

cli/src/native/browser.rs

@@ -121,7 +121,7 @@ fn normalize_url_for_match(url: &str) -> String {
 fn update_page_target_info_in_pages(pages: &mut [PageInfo], target: &TargetInfo) -> bool {
     if let Some(page) = pages.iter_mut().find(|p| p.target_id == target.target_id) {
         page.url = target.url.clone();
-        page.title = target.title.clone();
+        page.title = sanitize_title(&target.title);
         page.target_type = target.target_type.clone();
         return true;
     }
@@ -166,6 +166,75 @@ fn resolve_active_index(
     active_page_index
 }
 
+/// Strip zero-width / invisible / bidi-format Unicode from a page title before
+/// we store it. Some sites prepend runs of ZWJ / word-joiner / invisible-times /
+/// BOM to `document.title` (badging, watermarking, anti-scrape); left in, they
+/// pollute `tab list`, break text matching, and wreck column alignment (#33).
+fn sanitize_title(s: &str) -> String {
+    s.chars()
+        .filter(|&c| {
+            !matches!(c as u32,
+                0x00AD            // soft hyphen
+                | 0x200B..=0x200F // ZWSP, ZWNJ, ZWJ, LRM, RLM
+                | 0x2028 | 0x2029 // line / paragraph separators
+                | 0x202A..=0x202E // bidi embedding/override
+                | 0x2060..=0x2064 // word joiner, invisible operators
+                | 0x2066..=0x2069 // bidi isolates
+                | 0x180E          // Mongolian vowel separator
+                | 0xFEFF          // BOM / ZW no-break space
+            )
+        })
+        .collect::<String>()
+        .trim()
+        .to_string()
+}
+
+/// Best-effort MIME type from a filename extension, for the relay file-upload
+/// fallback (the page-constructed `File` needs a sensible `type`). Covers the
+/// common upload kinds; anything unknown falls back to a generic binary type.
+fn mime_for_path(name: &str) -> &'static str {
+    let ext = name.rsplit('.').next().unwrap_or("").to_lowercase();
+    match ext.as_str() {
+        "png" => "image/png",
+        "jpg" | "jpeg" => "image/jpeg",
+        "gif" => "image/gif",
+        "webp" => "image/webp",
+        "svg" => "image/svg+xml",
+        "bmp" => "image/bmp",
+        "pdf" => "application/pdf",
+        "txt" => "text/plain",
+        "csv" => "text/csv",
+        "json" => "application/json",
+        "mp4" => "video/mp4",
+        "webm" => "video/webm",
+        "mov" => "video/quicktime",
+        "mp3" => "audio/mpeg",
+        "zip" => "application/zip",
+        _ => "application/octet-stream",
+    }
+}
+
+/// Target ids to prune after a `Target.getTargets` resync: tracked pages whose
+/// target is no longer in the live set — EXCEPT the explicitly-pinned active
+/// target, which is protected. The relay against a busy real Chrome occasionally
+/// returns a different window's tabs for a single `getTargets` call ("tab list
+/// hops windows", issue #31); pruning on that transient snapshot would drop the
+/// agent's adopted tab and drift subsequent eval/click onto a foreign tab. A
+/// genuine close still arrives as `Target.targetDestroyed` (handled in the event
+/// drain), which removes the pin properly — so protecting it here only guards
+/// against flaky snapshots, not real closures.
+fn prunable_target_ids(
+    pages: &[PageInfo],
+    live_ids: &HashSet<String>,
+    pinned: Option<&str>,
+) -> Vec<String> {
+    pages
+        .iter()
+        .map(|p| p.target_id.clone())
+        .filter(|tid| !live_ids.contains(tid) && pinned != Some(tid.as_str()))
+        .collect()
+}
+
 /// 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.
@@ -185,6 +254,21 @@ fn active_index_is_owned(
         .unwrap_or(false)
 }
 
+/// Whether a CDP error means the bound relay target is gone — the tab was
+/// closed, navigated across processes (renderer swap), or lost after an
+/// extension/service-worker restart, and the relay could not re-attach. The
+/// ab-connect relay surfaces these as `stale sessionId … its tab is gone`,
+/// `unknown sessionId …`, or `no attached tab …`. `navigate` keys its
+/// auto-reattach recovery off this (issue #35) so a dead session rebinds to a
+/// fresh tab instead of erroring on every command until the user runs `tab new`.
+fn is_stale_target_error(error: &str) -> bool {
+    let lower = error.to_lowercase();
+    lower.contains("its tab is gone")
+        || lower.contains("stale sessionid")
+        || lower.contains("unknown sessionid")
+        || lower.contains("no attached tab")
+}
+
 /// Converts common error messages into AI-friendly, actionable descriptions.
 pub fn to_ai_friendly_error(error: &str) -> String {
     let lower = error.to_lowercase();
@@ -490,6 +574,13 @@ impl BrowserManager {
             }
         };
 
+        // A launched browser carries a debug port → it's the other path that can
+        // pop Chrome's consent modal; record it for #31 diagnosis.
+        crate::connect::log_connect_mode(
+            &ws_url,
+            true,
+            DAEMON_SESSION.get().map(String::as_str).unwrap_or("default"),
+        );
         let manager = if engine == "lightpanda" {
             initialize_lightpanda_manager(ws_url, process).await?
         } else {
@@ -585,6 +676,13 @@ impl BrowserManager {
         headers: Option<Vec<(String, String)>>,
     ) -> Result<Self, String> {
         let ws_url = resolve_cdp_url(url).await?;
+        // Record the transport so a reappearing "Allow remote debugging?" modal
+        // can be traced to a raw-port attach vs the consent-free relay (#31).
+        crate::connect::log_connect_mode(
+            &ws_url,
+            false,
+            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 {
             client,
@@ -627,6 +725,43 @@ 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>(
@@ -636,16 +771,7 @@ 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
@@ -713,15 +839,28 @@ impl BrowserManager {
                     target_id: target.target_id.clone(),
                     session_id: attach_result.session_id.clone(),
                     url: target.url.clone(),
-                    title: target.title.clone(),
+                    title: sanitize_title(&target.title),
                     target_type: target.target_type.clone(),
                 });
             }
 
-            self.active_page_index = 0;
-            self.pin_active_target();
-            let session_id = self.pages[0].session_id.clone();
-            self.enable_domains(&session_id).await?;
+            if self.agent_group().is_some() {
+                // Relay: the adopted tabs above are the USER's, in their real
+                // Chrome. NEVER make one of them the agent's working tab — that is
+                // how commands drifted onto whatever page the user was viewing
+                // between steps (eval/click/get landed on the user's foreground
+                // tab; #35). Open our own dedicated background tab in the session's
+                // group and pin THAT as active. The user's tabs stay adopted (so
+                // `tab list` / explicit `tab switch` can reach them) but are never
+                // auto-selected — the agent only ever drives a tab it owns.
+                self.tab_new(None, None).await?;
+            } else {
+                // A browser we launched: every tab is ours, so the first is fine.
+                self.active_page_index = 0;
+                self.pin_active_target();
+                let session_id = self.pages[0].session_id.clone();
+                self.enable_domains(&session_id).await?;
+            }
         }
 
         Ok(())
@@ -832,6 +971,24 @@ impl BrowserManager {
         )
     }
 
+    /// Drop the page bound to `session_id` from the tracked list — used when the
+    /// relay reports its tab is gone (issue #35) so the stale entry can't keep
+    /// resolving as active. Forgets ownership, unpins it if it was pinned, and
+    /// keeps `active_page_index` in range.
+    fn drop_page_by_session(&mut self, session_id: &str) {
+        let Some(pos) = self.pages.iter().position(|p| p.session_id == session_id) else {
+            return;
+        };
+        let target_id = self.pages[pos].target_id.clone();
+        self.pages.remove(pos);
+        self.created_targets.remove(&target_id);
+        if self.active_target_id.as_deref() == Some(target_id.as_str()) {
+            self.active_target_id = None;
+        }
+        self.active_page_index =
+            active_page_index_after_removal(self.active_page_index, pos, self.pages.len());
+    }
+
     /// Pin the current active page by target_id so later commands stick to it.
     /// Call after any explicit open / tab new / tab switch.
     fn pin_active_target(&mut self) {
@@ -861,10 +1018,10 @@ impl BrowserManager {
         if self.agent_group().is_some() && !self.active_is_session_owned() {
             self.tab_new(None, None).await?;
         }
-        let session_id = self.active_session_id()?.to_string();
+        let mut session_id = self.active_session_id()?.to_string();
         let mut lifecycle_rx = self.client.subscribe();
 
-        let nav_result: PageNavigateResult = self
+        let nav_result: PageNavigateResult = match self
             .client
             .send_command_typed(
                 "Page.navigate",
@@ -874,7 +1031,38 @@ impl BrowserManager {
                 },
                 Some(&session_id),
             )
-            .await?;
+            .await
+        {
+            Ok(r) => r,
+            // Auto-reattach when the bound tab is gone (issue #35). On the shared
+            // real browser the human can close/swap the agent's tab, and a
+            // cross-process nav can destroy the target without a re-attachable
+            // tabId — both leave the cached `cb-tab-<id>` session stale, so every
+            // command (including `open`) failed on it and only `tab new`
+            // recovered. The relay error literally says "re-open your target URL
+            // to re-attach"; fulfil that here: drop the dead page, open a fresh
+            // owned tab in this session's group, and navigate THAT. Gated on the
+            // relay (`agent_group`) and on the explicit navigation intent — read
+            // commands deliberately still fail loudly rather than silently
+            // recover onto a blank tab and return wrong data (issue #8.1).
+            Err(e) if self.agent_group().is_some() && is_stale_target_error(&e) => {
+                self.drop_page_by_session(&session_id);
+                self.tab_new(None, None).await?;
+                session_id = self.active_session_id()?.to_string();
+                lifecycle_rx = self.client.subscribe();
+                self.client
+                    .send_command_typed(
+                        "Page.navigate",
+                        &PageNavigateParams {
+                            url: url.to_string(),
+                            referrer: None,
+                        },
+                        Some(&session_id),
+                    )
+                    .await?
+            }
+            Err(e) => return Err(e),
+        };
 
         if let Some(ref error_text) = nav_result.error_text {
             return Err(format!("Navigation failed: {}", error_text));
@@ -939,7 +1127,7 @@ impl BrowserManager {
         self.active_page_index = self.resolved_active_index();
         if let Some(page) = self.pages.get_mut(self.active_page_index) {
             page.url = page_url.clone();
-            page.title = title.clone();
+            page.title = sanitize_title(&title);
         }
         self.pin_active_target();
 
@@ -1001,7 +1189,7 @@ impl BrowserManager {
 
     pub async fn get_title(&self) -> Result<String, String> {
         let result = self.evaluate_simple("document.title").await?;
-        Ok(result.as_str().unwrap_or("").to_string())
+        Ok(sanitize_title(result.as_str().unwrap_or("")))
     }
 
     pub async fn get_content(&self) -> Result<String, String> {
@@ -1344,7 +1532,7 @@ impl BrowserManager {
                 target_id: target.target_id.clone(),
                 session_id: attach.session_id.clone(),
                 url: target.url.clone(),
-                title: target.title.clone(),
+                title: sanitize_title(&target.title),
                 target_type: target.target_type.clone(),
             };
             self.add_background_page(page.clone());
@@ -1405,19 +1593,16 @@ impl BrowserManager {
                 target_id: target.target_id.clone(),
                 session_id: attach_result.session_id.clone(),
                 url: target.url.clone(),
-                title: target.title.clone(),
+                title: sanitize_title(&target.title),
                 target_type: target.target_type.clone(),
             });
             let _ = self.enable_domains(&attach_result.session_id).await;
         }
 
-        // Drop tabs that no longer exist so `tab list` doesn't show phantom rows.
-        let gone: Vec<String> = self
-            .pages
-            .iter()
-            .map(|p| p.target_id.clone())
-            .filter(|tid| !live_ids.contains(tid))
-            .collect();
+        // 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);
         }
@@ -1451,7 +1636,7 @@ impl BrowserManager {
                             }
                         }
                         if let Some(t) = ti.get("title").and_then(|v| v.as_str()) {
-                            page.title = t.to_string();
+                            page.title = sanitize_title(t);
                         }
                     }
                 }
@@ -1639,7 +1824,7 @@ impl BrowserManager {
 
         if let Some(page) = self.pages.get_mut(index) {
             page.url = url.clone();
-            page.title = title.clone();
+            page.title = sanitize_title(&title);
         }
 
         let page = &self.pages[index];
@@ -1894,7 +2079,8 @@ impl BrowserManager {
             .and_then(|v| v.as_i64())
             .ok_or("Could not get backendNodeId for file input")?;
 
-        self.client
+        let set_files = self
+            .client
             .send_command(
                 "DOM.setFileInputFiles",
                 Some(json!({
@@ -1903,26 +2089,153 @@ impl BrowserManager {
                 })),
                 Some(&effective_session_id),
             )
-            .await
-            .map_err(|e| {
-                // Chrome's chrome.debugger API (the extension-relay transport)
-                // forbids DOM.setFileInputFiles for security, surfacing as an
-                // opaque `-32000 "Not allowed"`. Translate it into an actionable
-                // message rather than leaking the raw CDP error (issue #13).
-                if e.contains("Not allowed") || e.contains("-32000") {
-                    "file upload isn't supported over the extension relay — \
-                     Chrome's chrome.debugger API forbids DOM.setFileInputFiles. \
-                     Use a direct-CDP session instead: \
-                     `chrome-use --session up --launch open <url>` (carry your \
-                     login over with `cookies export` | `cookies set --curl`), \
-                     then run `upload` in that session. \
-                     See https://github.com/leeguooooo/chrome-use/issues/13"
-                        .to_string()
-                } else {
-                    e
-                }
-            })?;
+            .await;
 
+        if let Err(e) = set_files {
+            // Chrome's chrome.debugger API (the extension-relay transport) forbids
+            // DOM.setFileInputFiles for security, surfacing as an opaque
+            // `-32000 "Not allowed"`. Fall back to constructing the File entirely
+            // IN THE PAGE and assigning it to the input — the standard
+            // Playwright/Cypress trick, which needs no privileged CDP and so works
+            // over the relay (issue #13).
+            if e.contains("Not allowed") || e.contains("-32000") {
+                return self
+                    .upload_files_via_page(object_id, files, &effective_session_id)
+                    .await;
+            }
+            return Err(e);
+        }
+
+        Ok(())
+    }
+
+    /// Relay-safe file upload: read each file locally, hand its bytes to the page
+    /// as base64, and rebuild a `File` there — then either assign it to a file
+    /// `<input>` (Chrome allows `input.files = dataTransfer.files`) or, for a
+    /// dropzone/composer, dispatch synthetic `paste`/`drop` events carrying the
+    /// `DataTransfer`. No `DOM.setFileInputFiles`, so chrome.debugger permits it.
+    async fn upload_files_via_page(
+        &self,
+        object_id: String,
+        files: &[String],
+        session_id: &str,
+    ) -> Result<(), String> {
+        use base64::Engine;
+        // The relay tunnels every CDP message through Chrome native messaging,
+        // which caps a single message at ~1 MiB. A whole image's base64 blows
+        // past that ("CDP response channel closed"), so we STREAM the bytes into
+        // a page-side buffer in sub-limit chunks, then assemble the File from it.
+        const CHUNK: usize = 96 * 1024; // base64 chars per message; safe under 1 MiB
+
+        // Reset the staging buffer.
+        self.client
+            .send_command(
+                "Runtime.evaluate",
+                Some(json!({ "expression": "window.__cuUpload = [];", "returnByValue": true })),
+                Some(session_id),
+            )
+            .await
+            .map_err(|e| format!("relay upload (reset) failed: {}", e))?;
+
+        for path in files {
+            let bytes = std::fs::read(path).map_err(|e| format!("cannot read {}: {}", path, e))?;
+            let name = std::path::Path::new(path)
+                .file_name()
+                .and_then(|n| n.to_str())
+                .unwrap_or("upload.bin")
+                .to_string();
+            let mime = mime_for_path(&name);
+            let b64 = base64::engine::general_purpose::STANDARD.encode(&bytes);
+
+            // Push the file's metadata with an empty buffer.
+            let init = format!(
+                "window.__cuUpload.push({{ name: {}, type: {}, b64: '' }});",
+                serde_json::to_string(&name).unwrap_or_default(),
+                serde_json::to_string(mime).unwrap_or_default(),
+            );
+            self.client
+                .send_command(
+                    "Runtime.evaluate",
+                    Some(json!({ "expression": init, "returnByValue": true })),
+                    Some(session_id),
+                )
+                .await
+                .map_err(|e| format!("relay upload (init) failed: {}", e))?;
+
+            // Stream the base64 in chunks. base64's alphabet (A–Za–z0–9+/=) needs
+            // no escaping inside a single-quoted JS string, so concatenation is safe.
+            let idx = "window.__cuUpload[window.__cuUpload.length-1].b64";
+            let mut start = 0;
+            while start < b64.len() {
+                let end = (start + CHUNK).min(b64.len());
+                let chunk = &b64[start..end];
+                let expr = format!("{idx} += '{chunk}';");
+                self.client
+                    .send_command(
+                        "Runtime.evaluate",
+                        Some(json!({ "expression": expr, "returnByValue": true })),
+                        Some(session_id),
+                    )
+                    .await
+                    .map_err(|e| format!("relay upload (chunk) failed: {}", e))?;
+                start = end;
+            }
+        }
+
+        // Assemble the Files from the buffer and attach to the element, then clean up.
+        let func = r#"function() {
+            const filesData = window.__cuUpload || [];
+            const dt = new DataTransfer();
+            for (const f of filesData) {
+                const bin = atob(f.b64);
+                const arr = new Uint8Array(bin.length);
+                for (let i = 0; i < bin.length; i++) arr[i] = bin.charCodeAt(i);
+                dt.items.add(new File([arr], f.name, { type: f.type }));
+            }
+            try { delete window.__cuUpload; } catch (e) { window.__cuUpload = undefined; }
+            const el = this;
+            if (el.tagName === 'INPUT' && el.type === 'file') {
+                el.files = dt.files;
+                el.dispatchEvent(new Event('input', { bubbles: true }));
+                el.dispatchEvent(new Event('change', { bubbles: true }));
+                return 'input:' + dt.files.length;
+            }
+            // Dropzone / rich composer: replay paste then drop with the files.
+            try { el.dispatchEvent(new ClipboardEvent('paste', { bubbles: true, clipboardData: dt })); } catch (e) {}
+            try {
+                const ev = new DragEvent('drop', { bubbles: true, cancelable: true });
+                Object.defineProperty(ev, 'dataTransfer', { value: dt });
+                el.dispatchEvent(ev);
+            } catch (e) {}
+            return 'event:' + dt.files.length;
+        }"#;
+
+        let result: EvaluateResult = self
+            .client
+            .send_command_typed(
+                "Runtime.callFunctionOn",
+                &CallFunctionOnParams {
+                    function_declaration: func.to_string(),
+                    object_id: Some(object_id),
+                    arguments: None,
+                    return_by_value: Some(true),
+                    await_promise: Some(false),
+                },
+                Some(session_id),
+            )
+            .await
+            .map_err(|e| format!("relay file-injection failed: {}", e))?;
+
+        if let Some(ref details) = result.exception_details {
+            return Err(format!(
+                "relay file-injection threw: {}",
+                details
+                    .exception
+                    .as_ref()
+                    .and_then(|ex| ex.description.as_deref())
+                    .unwrap_or(&details.text)
+            ));
+        }
         Ok(())
     }
 
@@ -2485,6 +2798,27 @@ mod tests {
         assert_eq!(active_page_index_after_removal(0, 0, 0), 0);
     }
 
+    #[test]
+    fn stale_target_error_matches_relay_signatures() {
+        // The exact relay error `open` must recover from (issue #35), as wrapped
+        // by send_command's `CDP error (Page.navigate): …` prefix.
+        assert!(is_stale_target_error(
+            "CDP error (Page.navigate): stale sessionId cb-tab-1655244623 for Page.navigate: \
+             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("no attached tab for Page.navigate"));
+    }
+
+    #[test]
+    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"));
+    }
+
     fn page(target_id: &str) -> PageInfo {
         PageInfo {
             tab_id: 1,
@@ -2582,6 +2916,48 @@ mod tests {
         assert!(!active_index_is_owned(&[], None, 0, &created));
     }
 
+    #[test]
+    fn test_sanitize_title() {
+        // The exact pollution from #33: ZWJ / word-joiner / invisible-times / BOM
+        // prepended to "GitHub".
+        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("  Hello World  "), "Hello World");
+        // Emoji and real content survive; only the invisibles are dropped.
+        assert_eq!(sanitize_title("✓ Done\u{200b}"), "✓ Done");
+    }
+
+    #[test]
+    fn test_mime_for_path() {
+        assert_eq!(mime_for_path("a.png"), "image/png");
+        assert_eq!(mime_for_path("PHOTO.JPG"), "image/jpeg");
+        assert_eq!(mime_for_path("clip.webp"), "image/webp");
+        assert_eq!(mime_for_path("doc.pdf"), "application/pdf");
+        assert_eq!(mime_for_path("noext"), "application/octet-stream");
+        assert_eq!(mime_for_path("weird.xyz"), "application/octet-stream");
+    }
+
+    #[test]
+    fn prune_protects_pinned_target_on_transient_snapshot() {
+        // The relay returned a getTargets snapshot missing the pinned tab "A"
+        // (it hopped to another window). "B" is also absent. Without protection
+        // both would be pruned and the next command would drift; with the pin
+        // protected, only the genuinely-unpinned "B" is dropped (issue #31).
+        let pages = vec![page("A"), page("B")];
+        let live: HashSet<String> = HashSet::new(); // snapshot returned neither
+        let gone = prunable_target_ids(&pages, &live, Some("A"));
+        assert_eq!(gone, vec!["B".to_string()]);
+        // With no pin, both are prunable (unchanged behavior).
+        let gone_unpinned = prunable_target_ids(&pages, &live, None);
+        assert_eq!(gone_unpinned.len(), 2);
+        // 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()]);
+    }
+
     #[test]
     fn resolve_active_index_pin_survives_passive_background_tab() {
         // A foreign tab ("Z") gets appended by passive discovery after we pinned

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
@@ -1169,6 +1178,69 @@ pub async fn get_main_content_text(client: &CdpClient, session_id: &str) -> Resu
         .to_string())
 }
 
+// Text nodes whose parent is one of these carry no visible content.
+fn is_noise_tag(name: &str) -> bool {
+    matches!(name, "SCRIPT" | "STYLE" | "NOSCRIPT" | "TEMPLATE" | "HEAD")
+}
+
+// Walk a CDP DOM.Node tree, collecting text-node values. Unlike `innerText`
+// (JS, blocked by CLOSED shadow roots), the CDP DOM tree from
+// `DOM.getDocument(pierce:true)` includes closed shadow roots and child
+// documents — so this reaches text JS can't. `parent_noise` carries whether an
+// ancestor was <script>/<style>/etc so their text is skipped.
+fn collect_dom_text(node: &Value, parent_noise: bool, out: &mut String) {
+    let node_type = node.get("nodeType").and_then(|v| v.as_i64()).unwrap_or(0);
+    let node_name = node.get("nodeName").and_then(|v| v.as_str()).unwrap_or("");
+    if node_type == 3 {
+        if !parent_noise {
+            if let Some(t) = node.get("nodeValue").and_then(|v| v.as_str()) {
+                let t = t.trim();
+                if !t.is_empty() {
+                    if !out.is_empty() {
+                        out.push(' ');
+                    }
+                    out.push_str(t);
+                }
+            }
+        }
+        return;
+    }
+    let noise = parent_noise || is_noise_tag(node_name);
+    if let Some(children) = node.get("children").and_then(|v| v.as_array()) {
+        for child in children {
+            collect_dom_text(child, noise, out);
+        }
+    }
+    if let Some(shadow) = node.get("shadowRoots").and_then(|v| v.as_array()) {
+        for sr in shadow {
+            collect_dom_text(sr, noise, out);
+        }
+    }
+    if let Some(doc) = node.get("contentDocument") {
+        collect_dom_text(doc, noise, out);
+    }
+}
+
+/// Extract text from the page via the CDP DOM tree with `pierce:true`, which
+/// reaches into CLOSED shadow roots and child documents that `innerText`/`eval`
+/// cannot. Lets an agent read content rendered into a closed shadow DOM (e.g. an
+/// extension's injected debug panel) without any extra Chrome permission — it
+/// rides the per-tab debugger session that's already attached (#30).
+pub async fn get_pierced_text(client: &CdpClient, session_id: &str) -> Result<String, String> {
+    let doc = client
+        .send_command(
+            "DOM.getDocument",
+            Some(serde_json::json!({ "depth": -1, "pierce": true })),
+            Some(session_id),
+        )
+        .await?;
+    let mut out = String::new();
+    if let Some(root) = doc.get("root") {
+        collect_dom_text(root, false, &mut out);
+    }
+    Ok(out)
+}
+
 pub async fn get_element_attribute(
     client: &CdpClient,
     session_id: &str,
@@ -1642,6 +1714,31 @@ mod tests {
         assert_eq!(parse_ref("@e123"), Some("e123".to_string()));
     }
 
+    #[test]
+    fn test_collect_dom_text_pierces_closed_shadow_and_skips_noise() {
+        // A CDP DOM.Node tree: a host element whose CLOSED shadow root holds the
+        // text, plus a <script> whose text must be skipped.
+        let tree = serde_json::json!({
+            "nodeType": 1, "nodeName": "BODY",
+            "children": [
+                { "nodeType": 1, "nodeName": "SCRIPT",
+                  "children": [ { "nodeType": 3, "nodeName": "#text", "nodeValue": "var secret=1;" } ] },
+                { "nodeType": 1, "nodeName": "DIV",
+                  "shadowRoots": [
+                    { "nodeType": 11, "nodeName": "#document-fragment",
+                      "children": [
+                        { "nodeType": 1, "nodeName": "SPAN",
+                          "children": [ { "nodeType": 3, "nodeName": "#text", "nodeValue": "DECRYPTED 42" } ] }
+                      ] }
+                  ] }
+            ]
+        });
+        let mut out = String::new();
+        collect_dom_text(&tree, false, &mut out);
+        assert_eq!(out, "DECRYPTED 42");
+        assert!(!out.contains("secret"), "script text must be skipped");
+    }
+
     #[test]
     fn test_parse_ref_equals_prefix() {
         assert_eq!(parse_ref("ref=e1"), Some("e1".to_string()));

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,
@@ -45,6 +56,26 @@ 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 && prefer_dom_dispatch(ref_map, selector_or_ref) {
+        return dom_click(
+            client,
+            session_id,
+            ref_map,
+            selector_or_ref,
+            iframe_sessions,
+        )
+        .await;
+    }
+
     let resolved = resolve_element_center(
         client,
         session_id,
@@ -235,6 +266,47 @@ async fn dom_click(
     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,
@@ -242,6 +314,13 @@ 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,
@@ -254,6 +333,50 @@ 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,
@@ -261,6 +384,11 @@ 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,
@@ -289,6 +417,63 @@ 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,
@@ -372,6 +557,7 @@ 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,
@@ -418,7 +604,7 @@ 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(
@@ -426,6 +612,7 @@ 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
@@ -475,6 +662,46 @@ 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/screenshot.rs

@@ -60,6 +60,9 @@ pub struct ScreenshotOptions {
     pub quality: Option<i32>,
     pub annotate: bool,
     pub output_dir: Option<String>,
+    /// Explicit pixel region (x, y, width, height) — `--clip` (issue #34). Takes
+    /// precedence over selector/full_page.
+    pub clip: Option<(f64, f64, f64, f64)>,
 }
 
 impl Default for ScreenshotOptions {
@@ -72,6 +75,7 @@ impl Default for ScreenshotOptions {
             quality: None,
             annotate: false,
             output_dir: None,
+            clip: None,
         }
     }
 }
@@ -187,7 +191,16 @@ async fn capture_screenshot_base64(
         capture_beyond_viewport: if options.full_page { Some(true) } else { None },
     };
 
-    if options.full_page {
+    if let Some((x, y, width, height)) = options.clip {
+        // Explicit pixel region wins over selector/full_page (issue #34).
+        params.clip = Some(Viewport {
+            x,
+            y,
+            width,
+            height,
+            scale: 1.0,
+        });
+    } else if options.full_page {
         let metrics: Value = client
             .send_command_no_params("Page.getLayoutMetrics", Some(session_id))
             .await?;

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 +344,19 @@ 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 +626,11 @@ 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 +643,14 @@ 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/output.rs

@@ -224,6 +224,40 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
             return;
         }
 
+        // Cloudflare challenge/clearance preflight (`cf-status`). Checked early
+        // 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 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 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 (icon, headline) = match rec {
+                "proceed" => (color::success_indicator().to_string(), "cleared — no challenge, proceed"),
+                "solve" => (color::warning_indicator().to_string(), "Cloudflare challenge active, no valid clearance — solve it"),
+                "reissue" => (color::warning_indicator().to_string(), "challenge active but a clearance cookie exists — stale (IP/UA changed?), re-solve"),
+                _ => (color::cyan("•").to_string(), "unknown"),
+            };
+            println!("{} {}", icon, headline);
+            println!("  challenged:     {}", if challenged { "yes" } else { "no" });
+            let cl_desc = if !present {
+                "absent".to_string()
+            } else if expired {
+                "present but EXPIRED".to_string()
+            } else if let Some(s) = expires_in {
+                format!("valid, expires in {}m {}s", s / 60, s % 60)
+            } else {
+                "present (session)".to_string()
+            };
+            println!("  cf_clearance:   {}", cl_desc);
+            println!("  device trusted: {}", if device { "yes (CF_VERIFIED_DEVICE)" } else { "no" });
+            return;
+        }
+
         // Dialog status response
         if action == Some("dialog") {
             if let Some(has_dialog) = data.get("hasDialog").and_then(|v| v.as_bool()) {
@@ -1430,6 +1464,12 @@ 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
@@ -1437,6 +1477,7 @@ 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.)
@@ -1701,12 +1742,23 @@ Usage: chrome-use scroll [direction] [amount] [options]
 
 Scrolls the page or a specific element in the specified direction.
 
+Without --selector, scroll dispatches a real (isTrusted) mouse wheel at a
+viewport coordinate, so it scrolls whatever container is under the pointer —
+including cross-origin iframes (Google Payments, Stripe, embedded checkout/KYC)
+that plain page scroll can't reach.
+
 Arguments:
   direction            up, down, left, right (default: down)
   amount               Pixels to scroll (default: 300)
 
 Options:
-  -s, --selector <sel> CSS selector for a scrollable container
+  -s, --selector <sel> CSS selector for a scrollable container (same-origin)
+  --at <x,y>           Dispatch the wheel at this viewport pixel (read it from a
+                       screenshot) — precise way into a cross-origin iframe
+  --frame <n>          Scroll the n-th frame from `chrome-use frames` (wheel at
+                       that frame's center)
+
+Without --selector/--at/--frame the wheel lands at the viewport center.
 
 Global Options:
   --json               Output as JSON
@@ -1718,6 +1770,8 @@ Examples:
   chrome-use scroll up 200
   chrome-use scroll left 100
   chrome-use scroll down 500 --selector "div.scroll-container"
+  chrome-use scroll down 700 --at 640,400      # wheel at a pixel over an iframe
+  chrome-use scroll down 700 --frame 2         # scroll frame 2 from `frames`
 "##
         }
         "scrollintoview" | "scrollinto" => {
@@ -1798,6 +1852,8 @@ Pass --hide-scrollbars false when launching to keep native scrollbars visible.
 
 Options:
   --full, -f           Capture full page (not just viewport)
+  [selector]           Capture just an element (CSS or @ref), e.g. `screenshot ".header" h.png`
+  --clip <x,y,w,h>     Capture a pixel region, e.g. `screenshot --clip 0,0,200,40 corner.png`
   --annotate           Overlay numbered labels on interactive elements.
                        Each label [N] corresponds to ref @eN from snapshot.
                        Prints a legend mapping labels to element roles/names.
@@ -1818,6 +1874,8 @@ Examples:
   chrome-use screenshot
   chrome-use screenshot ./screenshot.png
   chrome-use screenshot --full ./full-page.png
+  chrome-use screenshot ".header .indicator" corner.png  # just one element
+  chrome-use screenshot --clip 1600,0,200,40 corner.png  # a pixel region
   chrome-use screenshot --annotate              # Labeled screenshot + legend
   chrome-use screenshot --annotate ./page.png   # Save annotated screenshot
   chrome-use screenshot --annotate --json       # JSON output with annotations
@@ -1961,6 +2019,7 @@ Retrieves various types of information from elements or the page.
 Subcommands:
   text [selector]            Element text; no selector = WHOLE PAGE, all frames
   text --main                Main-content text only (skip nav/header/sidebar)
+  text --pierce              Read through CLOSED shadow DOM (injected panels)
   html <selector>            Get inner HTML of element
   value <selector>           Get value of input element
   attr <selector> <name>     Get attribute value
@@ -3194,6 +3253,10 @@ Get Info:  chrome-use get <what> [selector]
 Check State:  chrome-use is <what> <selector>
   visible, enabled, checked
 
+Anti-bot:  chrome-use stealth | cf-status
+  stealth         stealth self-check (webdriver/UA/plugins + overrides)
+  cf-status       Cloudflare challenge + cf_clearance preflight (skip re-solving)
+
 Find Elements:  chrome-use find <locator> <value> <action> [text]
   role, text, label, placeholder, alt, title, testid, first, last, nth
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "chrome-use",
-  "version": "1.5.3",
+  "version": "1.5.12",
   "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 +36,18 @@ 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.
 
+> **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
@@ -127,6 +139,19 @@ cadence, and scroll/drag ease. Default `off`; a per-navigation detector
 auto-escalates pages guarded by Akamai/PerimeterX/DataDome to `human`. Leave it
 on auto; force `human` only when you already know the target scores behaviour.
 
+**Cloudflare clearance — solve once, reuse.** Passing a Cloudflare challenge
+mints a `cf_clearance` cookie (HttpOnly — invisible to `eval`/`document.cookie`;
+read it via `chrome-use cookies`). It's bound to your **IP + User-Agent**: reuse
+the same exit IP and UA and you skip the challenge until it expires. Driving the
+user's real Chrome (relay) persists it natively; for isolated sessions,
+`--session-name <name>` save/restores it. Before spending effort solving, run
+`chrome-use cf-status` (aliases `cf`, `clearance`): it reports whether the page
+is *currently* a Cloudflare challenge and whether a still-valid `cf_clearance`
+exists, with a recommendation — `proceed` (already cleared, don't re-solve),
+`solve` (challenge up, no clearance), or `reissue` (clearance present but page
+still blocks → IP/UA drifted, re-solve). Use it as a preflight to avoid
+re-solving what you already cleared.
+
 ## Two ways to drive a page — and when to drop to `eval`
 
 You have a **real Chrome with the user's DOM**. Two layers, mix them freely:
@@ -210,6 +235,7 @@ For unstructured reading (no refs needed):
 chrome-use get text                    # WHOLE PAGE — all frames by default (see below)
 chrome-use get text @e1                # visible text of one element (or a CSS selector)
 chrome-use get text --main             # main content only — skip nav/header/sidebar
+chrome-use get text --pierce           # read through CLOSED shadow DOM (injected panels)
 chrome-use frames                      # list every frame + where the text lives
 chrome-use get html @e1                # innerHTML
 chrome-use get attr @e1 href           # any attribute
@@ -234,6 +260,12 @@ holds what), run `chrome-use frames`. To **cut boilerplate** (global nav/header/
 footer, "related items" sidebars), use `chrome-use get text --main`. If content
 is lazy-loaded, `scroll` it into view first, then read.
 
+**Closed shadow DOM.** Some injected UI (browser-extension debug panels, web
+components) renders into a *closed* shadow root that `eval`/`innerText` cannot
+read. `chrome-use get text --pierce` reads through closed shadow roots and child
+documents via the CDP DOM tree — use it when content is clearly on screen (you
+see it in a screenshot) but `get text`/`eval` come back empty.
+
 ## Interacting
 
 ```bash
@@ -244,6 +276,10 @@ 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)
@@ -261,16 +297,32 @@ chrome-use pick @e4 --option "Europe"  # ANY combobox (react-select / ARIA /
                                           # (no silent no-op). Use this for custom
                                           # dropdowns where `select` returns ✓ but
                                           # changes nothing.
-chrome-use upload @e5 file1.pdf        # upload file(s) — NOTE: needs a --launch/direct-CDP
-                                          # session. Over the extension relay it CANNOT work
-                                          # (Chrome's chrome.debugger forbids it); chrome-use
-                                          # errors with a hint. Carry your login into a launched
-                                          # session via `cookies export` | `cookies set --curl`.
+chrome-use upload @e5 file1.pdf        # upload file(s) — works over the extension relay too:
+                                          # chrome.debugger forbids setFileInputFiles, so the
+                                          # file's bytes are streamed into the page and rebuilt as
+                                          # a File there (chunked under native-messaging's 1 MiB cap).
+                                          # Works on file <input>s and drop/paste composers (e.g. X).
 chrome-use scroll down 500             # scroll page (up/down/left/right)
+chrome-use scroll down 700 --at 640,400 # wheel at a pixel — scrolls a cross-origin
+                                          # iframe (Payments/Stripe/checkout/KYC) that
+                                          # plain page scroll can't reach
+chrome-use scroll down 700 --frame 2    # scroll frame 2 from `chrome-use frames`
 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: