chore(release): 1.4.1 — hold-to-move (keydown/keyup full descriptor, #game) + expects ab-connect 0.4.8 (#23 reattach hardening)

claude-in-chrome completes the Rakuten cart→購入手続き→checkout flow that
chrome-use 1.2.3 couldn't, because it binds to the browser-level tabId (survives
renderer-process swaps) rather than a CDP target/sessionId (torn down by the
cross-origin OAuth/SSO nav). chrome-use's relay is already keyed to the stable
tabId (cb-tab-<tabId>, #17) and sends commands by {tabId} — the gap was purely
that the extension treated the session→tab map as the source of truth and only
reactively re-attached after a failed lookup.

Make the tabId the PRIMARY resolution path: derive it straight from the session
id (tabIdFromSession), ensure-attach with short retries across the swap window
(recoverSessionTab now loops), and route every send through sendCdpToTab, which
on a detached-style error drops the stale handle, re-attaches the stable tab, and
retries once. So a cross-process nav never surfaces as a hard error — there's no
'session gone' window, matching claude-in-chrome. Builds on 0.4.5/0.4.6 reattach;
makes it primary + bulletproof rather than a fallback.

cli/Cargo.lock

@@ -290,7 +290,7 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 
 [[package]]
 name = "chrome-use"
-version = "1.3.0"
+version = "1.4.1"
 dependencies = [
  "aes",
  "aes-gcm",

cli/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "chrome-use"
-version = "1.3.0"
+version = "1.4.1"
 edition = "2021"
 description = "Fast browser automation CLI for AI agents"
 license = "Apache-2.0"

cli/build.rs

@@ -3,6 +3,23 @@ use std::env;
 use std::fs;
 use std::path::Path;
 
+/// Embed the version of the `ab-connect` extension this CLI ships alongside, so
+/// `doctor` can tell a connected extension "you're older than what this CLI
+/// expects, update it." Read from the extension manifest at build time so it
+/// stays in sync with whatever extension version is in the same checkout/release
+/// (the ext is on its own 0.4.x line, separate from the CLI version). Falls back
+/// to "unknown" if the manifest can't be read.
+fn embed_extension_version() {
+    let manifest = Path::new("../extensions/ab-connect/manifest.json");
+    println!("cargo:rerun-if-changed=../extensions/ab-connect/manifest.json");
+    let version = fs::read_to_string(manifest)
+        .ok()
+        .and_then(|s| serde_json::from_str::<serde_json::Value>(&s).ok())
+        .and_then(|v| v.get("version").and_then(|x| x.as_str()).map(String::from))
+        .unwrap_or_else(|| "unknown".to_string());
+    println!("cargo:rustc-env=AB_CONNECT_VERSION={}", version);
+}
+
 /// Ensure `packages/dashboard/out/` exists so `rust-embed` doesn't fail during
 /// Rust-only dev builds where the dashboard hasn't been built. The placeholder
 /// `index.html` is only written when the directory is completely absent.
@@ -20,6 +37,7 @@ fn ensure_dashboard_dir() {
 
 fn main() {
     ensure_dashboard_dir();
+    embed_extension_version();
 
     let protocol_dir = Path::new("cdp-protocol");
     let out_dir = env::var("OUT_DIR").unwrap();

cli/src/connect.rs

@@ -449,6 +449,28 @@ pub fn relay_url() -> Option<String> {
     }
 }
 
+/// 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.
+fn relay_ext_version_path() -> PathBuf {
+    relay_url_path().with_file_name("relay-ext-version")
+}
+
+/// Version of the connected `ab-connect` extension, if the host learned it from
+/// the extension's `hello`. `None` when no extension has connected since the
+/// host started, or the extension predates version reporting.
+pub fn relay_ext_version() -> Option<String> {
+    let s = std::fs::read_to_string(relay_ext_version_path())
+        .ok()?
+        .trim()
+        .to_string();
+    if s.is_empty() {
+        None
+    } else {
+        Some(s)
+    }
+}
+
 /// Hidden `__nm-host` mode: launched by Chrome for the ab-connect extension.
 ///
 /// Bridges the extension (native-messaging stdio, envelope protocol) to a local
@@ -570,6 +592,15 @@ async fn nm_host_main() {
             Ok(v) => v,
             Err(_) => continue,
         };
+        // Extension version handshake: record it next to the relay URL so
+        // `doctor` can report which extension build is live (and whether it's
+        // behind). Best-effort; the message carries no CDP payload.
+        if v.get("method").and_then(|m| m.as_str()) == Some("hello") {
+            if let Some(ver) = v.get("version").and_then(|x| x.as_str()) {
+                let _ = std::fs::write(relay_ext_version_path(), ver);
+            }
+            continue;
+        }
         let outs = {
             let mut s = state.lock().await;
             s.handle_ext_message(&v, "")
@@ -602,6 +633,7 @@ async fn nm_host_main() {
     }
     nm_log("[nm-host] stdin EOF — Chrome closed the port");
     let _ = std::fs::remove_file(relay_url_path());
+    let _ = std::fs::remove_file(relay_ext_version_path());
 }
 
 #[allow(clippy::too_many_arguments)]

cli/src/doctor/mod.rs

@@ -18,6 +18,7 @@ mod launch;
 mod network;
 mod providers;
 mod security;
+mod versions;
 
 use serde_json::{json, Value};
 
@@ -97,6 +98,7 @@ pub fn run_doctor(opts: DoctorOptions) -> i32 {
     let mut fixed: Vec<String> = Vec::new();
 
     environment::check(&mut checks);
+    versions::check(&mut checks);
     chrome::check(&mut checks);
     daemon::check(&mut checks);
     config::check(&mut checks);

cli/src/doctor/versions.rs

@@ -0,0 +1,89 @@
+//! Version-coherence checks across all four moving parts: the CLI binary, the
+//! per-session daemons (covered by `daemon.rs`), the connected `ab-connect`
+//! extension, and the bundled skill. The extension was previously a black box —
+//! nothing reported which build was live — so a user could sit on an old
+//! extension with no signal. The extension now reports its version over the
+//! relay (`hello`), the host records it, and this surfaces it in one place.
+
+use super::{Check, Status};
+use crate::{connect, upgrade};
+
+pub(super) fn check(checks: &mut Vec<Check>) {
+    let category = "Versions";
+    let cli_version = env!("CARGO_PKG_VERSION");
+
+    // CLI — compare against the latest seen by the background update check.
+    match upgrade::cached_latest_version() {
+        Some(latest) if upgrade::version_is_newer(&latest, cli_version) => {
+            checks.push(
+                Check::new(
+                    "versions.cli",
+                    category,
+                    Status::Warn,
+                    format!("CLI {cli_version} (newer available: {latest})"),
+                )
+                .with_fix("chrome-use upgrade".to_string()),
+            );
+        }
+        _ => {
+            checks.push(Check::new(
+                "versions.cli",
+                category,
+                Status::Pass,
+                format!("CLI {cli_version}"),
+            ));
+        }
+    }
+
+    // Extension — the build this CLI shipped alongside (embedded at compile time
+    // from the extension manifest) is what we expect to be running.
+    let expected_ext = env!("AB_CONNECT_VERSION");
+    match connect::relay_ext_version() {
+        Some(ext) if upgrade::version_is_newer(expected_ext, &ext) => {
+            checks.push(
+                Check::new(
+                    "versions.extension",
+                    category,
+                    Status::Warn,
+                    format!("extension {ext} is behind the bundled {expected_ext}"),
+                )
+                .with_fix(
+                    "update ab-connect in Chrome: chrome://extensions \u{2192} reload \
+                     (or wait for the Web Store auto-update)"
+                        .to_string(),
+                ),
+            );
+        }
+        Some(ext) => {
+            checks.push(Check::new(
+                "versions.extension",
+                category,
+                Status::Pass,
+                format!("extension {ext}"),
+            ));
+        }
+        None => {
+            checks.push(Check::new(
+                "versions.extension",
+                category,
+                Status::Info,
+                format!(
+                    "extension not connected (or it predates version reporting — \
+                     expected {expected_ext})"
+                ),
+            ));
+        }
+    }
+
+    // Skill — ships inside the same release artifact as the binary, so it's
+    // version-locked here. Copies made elsewhere via `skills add` aren't.
+    checks.push(Check::new(
+        "versions.skill",
+        category,
+        Status::Info,
+        format!(
+            "skills bundled with this CLI ({cli_version}); copies made via `skills add` \
+             elsewhere may be stale — re-run to refresh"
+        ),
+    ));
+}

cli/src/native/actions.rs

@@ -2888,7 +2888,32 @@ async fn handle_snapshot(cmd: &Value, state: &mut DaemonState) -> Result<Value,
         })
         .collect();
 
-    Ok(json!({ "snapshot": tree, "origin": url, "refs": refs }))
+    let ref_count = refs.len();
+    let mut out = json!({ "snapshot": tree, "origin": url, "refs": refs });
+
+    // Canvas/WebGL apps (games, map/3D viewers, drawing tools) paint to a
+    // <canvas> and expose almost no accessibility tree, so `snapshot` comes back
+    // near-empty and agents get stuck looking for refs that will never exist
+    // (dogfood: the Dead Cell game). When the tree is sparse but a canvas
+    // dominates the viewport, tell them to switch to the screenshot-driven path.
+    if ref_count < 3 {
+        let canvas_js =
+            "(() => { const c = document.querySelector('canvas'); if (!c) return false; \
+                         const r = c.getBoundingClientRect(); \
+                         return r.width * r.height > innerWidth * innerHeight * 0.5; })()";
+        if let Ok(v) = mgr.evaluate(canvas_js, None).await {
+            if v.as_bool() == Some(true) {
+                out["note"] = json!(
+                    "This page renders to a <canvas> (game / WebGL / editor) and exposes almost no \
+                     accessibility tree — refs won't help. Use `screenshot` to see it, coordinate \
+                     `click <x> <y>` to interact, and `keydown`/`keyup`/`press` for keyboard \
+                     (hold-to-move: `keydown d` … `keyup d`)."
+                );
+            }
+        }
+    }
+
+    Ok(out)
 }
 
 /// Resolve a (possibly relative) saved-file path to an absolute one so the CLI
@@ -8760,13 +8785,7 @@ async fn handle_keydown(cmd: &Value, state: &DaemonState) -> Result<Value, Strin
         .and_then(|v| v.as_str())
         .ok_or("Missing 'key' parameter")?;
 
-    mgr.client
-        .send_command(
-            "Input.dispatchKeyEvent",
-            Some(json!({ "type": "keyDown", "key": key })),
-            Some(&session_id),
-        )
-        .await?;
+    interaction::dispatch_single_key(&mgr.client, &session_id, key, "keyDown").await?;
     Ok(json!({ "keydown": key }))
 }
 
@@ -8778,13 +8797,7 @@ async fn handle_keyup(cmd: &Value, state: &DaemonState) -> Result<Value, String>
         .and_then(|v| v.as_str())
         .ok_or("Missing 'key' parameter")?;
 
-    mgr.client
-        .send_command(
-            "Input.dispatchKeyEvent",
-            Some(json!({ "type": "keyUp", "key": key })),
-            Some(&session_id),
-        )
-        .await?;
+    interaction::dispatch_single_key(&mgr.client, &session_id, key, "keyUp").await?;
     Ok(json!({ "keyup": key }))
 }
 

cli/src/native/browser.rs

@@ -166,6 +166,25 @@ fn resolve_active_index(
     active_page_index
 }
 
+/// 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.
+fn active_index_is_owned(
+    pages: &[PageInfo],
+    active_target_id: Option<&str>,
+    active_page_index: usize,
+    created_targets: &HashSet<String>,
+) -> bool {
+    pages
+        .get(resolve_active_index(
+            pages,
+            active_target_id,
+            active_page_index,
+        ))
+        .map(|p| created_targets.contains(&p.target_id))
+        .unwrap_or(false)
+}
+
 /// Converts common error messages into AI-friendly, actionable descriptions.
 pub fn to_ai_friendly_error(error: &str) -> String {
     let lower = error.to_lowercase();
@@ -799,6 +818,20 @@ impl BrowserManager {
         )
     }
 
+    /// Whether the resolved active page is a tab THIS session created (via
+    /// `Target.createTarget` — `tab new`, `ensure_page`, or the first `open`).
+    /// On the shared real browser a fresh session also passively attaches to the
+    /// user's existing tabs; those are NOT owned, and navigating one would
+    /// clobber the user's page. Used to gate `navigate` on the relay.
+    fn active_is_session_owned(&self) -> bool {
+        active_index_is_owned(
+            &self.pages,
+            self.active_target_id.as_deref(),
+            self.active_page_index,
+            &self.created_targets,
+        )
+    }
+
     /// 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) {
@@ -816,6 +849,18 @@ impl BrowserManager {
     }
 
     pub async fn navigate(&mut self, url: &str, wait_until: WaitUntil) -> Result<Value, String> {
+        // On the shared real browser (extension relay), a fresh session only
+        // passively attached to the user's existing tabs — it doesn't own any. The
+        // pre-fix code made one of those the active tab, so the first `open` then
+        // navigated (clobbered) the user's page: in dogfooding an `open` replaced a
+        // half-filled form with the target site. If the active tab isn't one we
+        // created, open our own tab in this session's group and navigate THAT, so
+        // the user's (and other sessions') tabs are never hijacked. Off the relay
+        // (a browser we launched) reusing the active tab is correct, so this is
+        // gated on `agent_group()`.
+        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 lifecycle_rx = self.client.subscribe();
 
@@ -2431,6 +2476,35 @@ mod tests {
         assert_eq!(resolve_active_index(&pages, Some("CLOSED"), 1), 1);
     }
 
+    // --- issue: `open` must not hijack a user's tab on the relay (dogfood) ---
+
+    #[test]
+    fn active_not_owned_when_only_user_tabs_discovered() {
+        // A fresh relay session passively attached to the user's tabs but created
+        // none — so navigate must NOT reuse the active tab (it'd clobber the
+        // user's page); it has to open its own first.
+        let pages = vec![page("USER_A"), page("USER_B")];
+        let created = HashSet::new();
+        assert!(!active_index_is_owned(&pages, Some("USER_A"), 0, &created));
+    }
+
+    #[test]
+    fn active_owned_when_session_created_the_tab() {
+        let pages = vec![page("USER_A"), page("OURS")];
+        let mut created = HashSet::new();
+        created.insert("OURS".to_string());
+        // Active pinned to the tab we created → safe to navigate it.
+        assert!(active_index_is_owned(&pages, Some("OURS"), 1, &created));
+        // But pinned to the user's tab → not owned, even though we own another.
+        assert!(!active_index_is_owned(&pages, Some("USER_A"), 0, &created));
+    }
+
+    #[test]
+    fn active_not_owned_when_no_pages() {
+        let created = HashSet::new();
+        assert!(!active_index_is_owned(&[], None, 0, &created));
+    }
+
     #[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/interaction.rs

@@ -557,6 +557,48 @@ pub async fn press_key_with_modifiers(
     Ok(())
 }
 
+/// Dispatch a SINGLE key event (`keyDown` or `keyUp`) carrying the full key
+/// descriptor — `key`, `code`, `windowsVirtualKeyCode`/`nativeVirtualKeyCode`,
+/// and (on key-down) printable `text`. Powers the `keydown`/`keyup` commands.
+///
+/// The previous implementation sent only `{key}`, so games and shortcut handlers
+/// that read `event.code` (e.g. `"KeyD"`, `"ArrowRight"`) or `event.keyCode` saw
+/// nothing — a held key set no movement flag and did nothing (dogfood: holding a
+/// direction in a canvas platformer barely nudged the player). Sending the same
+/// descriptor `press` uses makes hold-to-move work regardless of which field the
+/// page keys off.
+pub async fn dispatch_single_key(
+    client: &CdpClient,
+    session_id: &str,
+    key: &str,
+    event_type: &str,
+) -> Result<(), String> {
+    let (key_name, code, key_code) = named_key_info(key);
+    // Printable text is only meaningful on key-down; key-up never inserts.
+    let text = if event_type == "keyDown" {
+        key_text(&key_name)
+    } else {
+        None
+    };
+    client
+        .send_command_typed::<_, Value>(
+            "Input.dispatchKeyEvent",
+            &DispatchKeyEventParams {
+                event_type: event_type.to_string(),
+                key: Some(key_name),
+                code: Some(code),
+                text: text.clone(),
+                unmodified_text: text,
+                windows_virtual_key_code: Some(key_code),
+                native_virtual_key_code: Some(key_code),
+                modifiers: None,
+            },
+            Some(session_id),
+        )
+        .await?;
+    Ok(())
+}
+
 pub async fn scroll(
     client: &CdpClient,
     session_id: &str,

cli/src/output.rs

@@ -297,6 +297,11 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
         // Snapshot
         if let Some(snapshot) = data.get("snapshot").and_then(|v| v.as_str()) {
             print_with_boundaries(snapshot, origin, opts);
+            // Canvas-app hint: the tree was near-empty but the page paints to a
+            // <canvas>, so refs are a dead end — point at the screenshot path.
+            if let Some(note) = data.get("note").and_then(|v| v.as_str()) {
+                eprintln!("{}", color::dim(note));
+            }
             return;
         }
         // Title
@@ -3066,6 +3071,9 @@ Core Commands:
   type <sel> <text>          Type into element
   fill <sel> <text>          Clear and fill
   press <key>                Press key (Enter, Tab, Control+a)
+  keydown <key>              Hold a key down (no auto-release) — for games/shortcuts
+  keyup <key>                Release a held key. Pair with keydown to hold-to-move:
+                             `keydown d` … `keyup d`
   keyboard type <text>       Type text with real keystrokes (no selector)
   keyboard inserttext <text> Insert text without key events
   hover <sel>                Hover element

cli/src/upgrade.rs

@@ -52,6 +52,27 @@ fn is_newer(latest: &str, current: &str) -> bool {
     matches!((parse_version(latest), parse_version(current)), (Some(l), Some(c)) if l > c)
 }
 
+/// Public semver-ish comparison (`latest` strictly newer than `current`), so
+/// `doctor` can flag a stale extension/CLI without re-implementing parsing.
+pub fn version_is_newer(latest: &str, current: &str) -> bool {
+    is_newer(latest, current)
+}
+
+/// The latest CLI version recorded by the background update check, if any.
+/// `doctor` uses it to show "a newer chrome-use is available" without a network
+/// call (the `__update-check` worker refreshes the cache out of band).
+pub fn cached_latest_version() -> Option<String> {
+    std::fs::read_to_string(update_cache_path())
+        .ok()
+        .and_then(|s| serde_json::from_str::<serde_json::Value>(&s).ok())
+        .and_then(|j| {
+            j.get("latest")
+                .and_then(|v| v.as_str())
+                .map(|s| s.to_string())
+        })
+        .filter(|s| !s.is_empty())
+}
+
 /// Hidden `__update-check` subcommand: fetch the latest release tag and cache it.
 /// Spawned detached by [`maybe_notify_update`] so the network call never blocks a
 /// real command. Uses `curl` (no extra deps, matches `upgrade`).

extensions/ab-connect/background.js

@@ -108,6 +108,12 @@ function connectHost() {
     // reconnect. Keep chrome.debugger attached so reconnect is cheap.
     for (const tabId of tabs.keys()) setBadge(tabId, 'connecting')
   })
+  // Report our version so the host can tell the CLI/`doctor` which extension
+  // build is live (otherwise the extension version is a black box — the user
+  // can't tell they're on an old one). Best-effort; ignored by older hosts.
+  try {
+    postToHost({ method: 'hello', version: chrome.runtime.getManifest().version })
+  } catch {}
   // Tell the daemon about everything we already have attached, then attach
   // anything new.
   reannounceAttachedTabs()
@@ -149,23 +155,56 @@ function tabForTarget(targetId) {
   return null
 }
 
-// Best-effort recovery for a stale `cb-tab-<tabId>` session: the handle is gone
-// from our maps, but if the underlying Chrome tab still exists and is eligible,
-// re-attach to it and return its id so the in-flight command can be retried.
-// Returns null when the tab is genuinely gone (closed / restricted), in which
-// case the caller surfaces the stale-session error. (issue #20.1)
+// The STABLE Chrome tabId encoded in a `cb-tab-<tabId>` session id (#17), or
+// null for any other session shape (child/iframe sessions). The tabId is the
+// real source of truth: it survives the renderer-process swaps (cross-origin
+// OAuth/SSO navs) that tear down the page's CDP target — which is why binding to
+// it (like claude-in-chrome) rides through the hop that killed the old
+// target/sessionId binding (issue #23).
+function tabIdFromSession(sessionId) {
+  const m = /^cb-tab-(\d+)$/.exec(sessionId || '')
+  return m ? Number(m[1]) : null
+}
+
+// Ensure the debugger is attached to a `cb-tab-<tabId>` session's tab, re-attaching
+// across the transient window of a process swap (with a couple of short retries).
+// Returns the tabId on success, or null when the tab is genuinely gone
+// (closed / restricted). (issues #20.1, #23)
 async function recoverSessionTab(sessionId) {
-  const m = /^cb-tab-(\d+)$/.exec(sessionId)
-  if (!m) return null
-  const tabId = Number(m[1])
-  const tab = await chrome.tabs.get(tabId).catch(() => null)
-  if (!eligible(tab)) return null
-  try {
-    await attachTab(tabId)
-  } catch {
-    return null
+  const tabId = tabIdFromSession(sessionId)
+  if (tabId == null) return null
+  for (let i = 0; i < 3; i++) {
+    const tab = await chrome.tabs.get(tabId).catch(() => null)
+    if (!eligible(tab)) return null
+    try {
+      await attachTab(tabId)
+      if (tabs.has(tabId)) return tabId
+    } catch {
+      // mid-swap: the tab exists but isn't attachable yet — back off and retry.
+    }
+    await new Promise((r) => setTimeout(r, 120 + i * 150))
+  }
+  return null
+}
+
+// Send a CDP command to a tab, riding a debugger detach that can happen between
+// our attach check and the command itself (a renderer-process swap mid-flight).
+// On a detached-style failure, drop the stale handle, re-attach the stable tab,
+// and retry once — so a cross-process nav never surfaces as a hard error (#23).
+async function sendCdpToTab(tabId, method, params) {
+  const dbg = { tabId }
+  try {
+    return await chrome.debugger.sendCommand(dbg, method, params)
+  } catch (e) {
+    const msg = String((e && e.message) || e)
+    if (!/detached|not attached|target.*(closed|gone)|no target|cannot access|frame.*detached/i.test(msg)) {
+      throw e
+    }
+    detachTab(tabId, false)
+    const ok = await recoverSessionTab(`cb-tab-${tabId}`)
+    if (!ok) throw e
+    return await chrome.debugger.sendCommand(dbg, method, params)
   }
-  return tabs.has(tabId) ? tabId : null
 }
 
 function anyConnectedTab() {
@@ -226,24 +265,26 @@ async function handleForwardCdpCommand(msg) {
   // Fail loudly instead so the agent sees an actionable error, not bad data.
   let tabId
   if (sessionId) {
-    tabId = tabForSession(sessionId)
-    if (!tabId) {
-      // The session's debugger handle is gone, but `cb-tab-<tabId>` encodes the
-      // STABLE Chrome tabId (#17). A cross-process navigation (e.g. an SSO
-      // redirect to another origin), a service-worker restart, or DevTools
-      // briefly stealing the debugger all tear the handle down while the tab
-      // itself lives on. Before failing, try to transparently re-attach to that
-      // same tab and retry — so `open`/`navigate`/`eval` self-heal instead of
-      // dead-ending the agent (issue #20.1). attachTab re-mints the identical
-      // `cb-tab-<tabId>` session, so the daemon's binding stays valid.
-      tabId = await recoverSessionTab(sessionId)
-      if (!tabId) {
+    // The stable Chrome tabId encoded in `cb-tab-<tabId>` is the source of truth
+    // (it survives renderer-process swaps; the CDP target/sessionId does not).
+    // Resolve via it primarily — don't depend on a session→tab map entry that the
+    // detach handler may have cleared — and ensure the debugger is attached,
+    // re-attaching across a cross-process nav before failing (issues #20.1, #23).
+    // `tabForSession` still covers child/iframe sessions that aren't `cb-tab-*`.
+    tabId = tabIdFromSession(sessionId) ?? tabForSession(sessionId)
+    if (tabId == null) {
+      throw new Error(`unknown sessionId ${sessionId} for ${method}`)
+    }
+    if (!tabs.has(tabId)) {
+      const recovered = await recoverSessionTab(sessionId)
+      if (!recovered) {
         throw new Error(
           `stale sessionId ${sessionId} for ${method}: its tab is gone (closed, ` +
             `navigated across processes, or lost after an extension restart). ` +
             `Re-attach by re-opening your target URL before retrying.`,
         )
       }
+      tabId = recovered
     }
   } else if (typeof params?.targetId === 'string') {
     tabId = tabForTarget(params.targetId)
@@ -253,18 +294,17 @@ async function handleForwardCdpCommand(msg) {
     // applies to any attached tab.
     tabId = anyConnectedTab()
   }
-  if (!tabId) throw new Error(`no attached tab for ${method}`)
-  const dbg = { tabId }
+  if (tabId == null) throw new Error(`no attached tab for ${method}`)
 
   // Re-enabling Runtime can leave a stale state; bounce it (matches upstream).
   if (method === 'Runtime.enable') {
     try {
-      await chrome.debugger.sendCommand(dbg, 'Runtime.disable')
+      await sendCdpToTab(tabId, 'Runtime.disable', undefined)
       await new Promise((r) => setTimeout(r, 30))
     } catch {}
-    return await chrome.debugger.sendCommand(dbg, 'Runtime.enable', params)
+    return await sendCdpToTab(tabId, 'Runtime.enable', params)
   }
-  return await chrome.debugger.sendCommand(dbg, method, params)
+  return await sendCdpToTab(tabId, method, params)
 }
 
 // ---- attach / detach ------------------------------------------------------

extensions/ab-connect/manifest.json

@@ -1,7 +1,7 @@
 {
   "manifest_version": 3,
   "name": "chrome-use",
-  "version": "0.4.6",
+  "version": "0.4.8",
   "description": "Let chrome-use drive your logged-in Chrome \u2014 install once, no token, no per-use confirmation.",
   "key": "MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEA6vQIyscGIPYPZdSpPwPL0+0gxUROyRgCpmvCSDoc8XUm4qm97VbKnD9Ijc1lV22lNWZtE78gaRjt6BeSfuMgnBymnhLKjN1gU6AI5QUU0mrJyeHdWKvrKQR5FmsM2A7Xr1ykE2SiiS8zNUS3Y/6O5l+Nva7wrVy6E4a2dkBVQkOsu+DV+nEZvhIyuDY5D5SPXqNwUTWTaglwj5mjvHz36xSwCWlPmrtJ+ED0AUyrb2z4GIOmvk4kqtBVrh/UD058klLo4CkYOnIybB5aV6WYuwarfPY4bF/dLggPem+ewLNTUNBuwrxj/A4nUv0LJTuRO8rR7f8WR9qnRCY0Ic5saQIDAQAB",
   "icons": {

package.json

@@ -1,6 +1,6 @@
 {
   "name": "chrome-use",
-  "version": "1.3.0",
+  "version": "1.4.1",
   "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

@@ -226,8 +226,11 @@ 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 press Enter                 # press a key at current focus
+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)
+chrome-use keyup d                     # release it — pair them to hold-to-move
+                                          # in a game: `keydown d; sleep; keyup d`
 chrome-use check @e3                   # check checkbox
 chrome-use uncheck @e3                 # uncheck
 chrome-use select @e4 "option-value"   # native <select> only
@@ -296,6 +299,23 @@ chrome-use click --coords 449,320   # same, explicit flag
 
 A bare-number argument is always a coordinate, never a selector.
 
+### Canvas / WebGL apps (games, map & 3D viewers, drawing tools)
+
+These paint everything to a `<canvas>` and expose **almost no accessibility
+tree**, so `snapshot` comes back near-empty and refs are a dead end. `snapshot`
+detects this and prints a one-line hint. Drive them the screenshot way:
+
+```bash
+chrome-use screenshot /tmp/s.png       # SEE the state (your only read path —
+                                          # eval/get text return nothing useful)
+chrome-use click 640 360               # interact by viewport coordinate
+chrome-use keydown d; sleep 0.6; chrome-use keyup d   # hold-to-move
+chrome-use press Space                 # discrete actions (jump/attack/confirm)
+```
+
+Each command is a ~250ms round-trip, so this is fine for turn-based / canvas
+*apps* but too slow to play a real-time 60fps action game frame-by-frame.
+
 ## Waiting (read this)
 
 Agents fail more often from bad waits than from bad selectors. Pick the