Compare commits

...
6 Commits
Author SHA1 Message Date
leeguooooo 3d82f11ff2 chore(release): 1.5.0 — text-selector click + get text→body + tab --activate + click --follow/openedTab (#24); fill fires input/change/blur (#25); close <tab> wording + chrome-use current (#26)
Release binaries / Build macOS ARM64 (push) Has been cancelled
Release binaries / Build macOS x64 (push) Has been cancelled
Release binaries / Build Linux ARM64 (push) Has been cancelled
Release binaries / Build Linux musl ARM64 (push) Has been cancelled
Release binaries / Build Linux musl x64 (push) Has been cancelled
Release binaries / Build Linux x64 (push) Has been cancelled
Release binaries / Build Windows x64 (push) Has been cancelled
Release binaries / Attach binaries to GitHub Release (push) Has been cancelled
2026-06-15 11:32:30 +09:00
leeguooooo 33269adc1a fix(fill/tabs): dispatch real input/change/blur (#25); close <tab> wording + chrome-use current (#26)
#25 — fill() didn't fire the events framework inputs / site autocomplete need:
it set value directly (bypassing React's value-tracker) and typed via
Input.insertText, so controlled components and input/change/blur listeners (e.g.
Mercari's postal-code → 都道府県 lookup) never ran though the value showed. fill
now emulates a real edit: focus, set through the element's prototype value setter
(React _valueTracker registers), then dispatch input → input → change → blur/
focusout. SELECT and contenteditable handled too. type <sel> <text> remains for
per-keystroke sites. Verified live: an input wired with input/change/blur fired
'IICB' from one fill.

#26 (ergonomics):
- 'close <tab>' now closes just that tab and prints 'Tab [tN] closed'; bare
  'close' still closes the browser. Previously 'close t12' ran a browser close
  and alarmingly printed 'Browser closed'.
- new 'chrome-use current': prints the active tab's stable handle (tabId + CDP
  targetId + url/title), refreshed live — so an agent holds the targetId (which
  survives cross-process nav) instead of re-deriving 'which tab is live' from
  'tabs' every step. The deeper tab-id churn is the #21/#23 stable-targetId story.

Tests cover fill events (live), close tab-vs-browser parse, and current.
2026-06-15 11:26:34 +09:00
leeguooooo 9ab8753b48 feat(click): report (and optionally --follow) a tab opened by a click (#24-A)
A click on a target=_blank link / window.open opened a new tab, but the active
tab stayed put, so the post-click snapshot showed the OLD page — looking exactly
like the click failed. On the relay the new tab is discovered only via getTargets
(the relay doesn't push target events to the daemon), so it went unsurfaced.

handle_click now snapshots tracked targets before the click and, after, runs a
lightweight BrowserManager::adopt_newly_opened (one getTargets, attaches only the
new target — far cheaper than a full resync) to detect a freshly-opened tab. It's
reported as openedTab {tabId,url,title} in the response (and a '→ opened new tab
[tN] <url>' hint in text mode). Default keeps focus on the current tab (so
multi-tab flows aren't hijacked, per #7/#8.1); 'click <sel> --follow' switches to
the new tab. Verified live: clicking a _blank link prints
'→ opened new tab [t13] https://example.org/'.

Completes the #24 friction items (B/C/D shipped in 770708b).
2026-06-15 11:13:34 +09:00
leeguooooo 770708b8e6 fix(cli): text-selector click by visible label + get text→body + tab --activate (#24)
Three CLI gaps surfaced driving a Mercari signup→checkout flow:

- #24-B (correctness): a bare label like 'click 購入手続きへ' was fed straight to
  document.querySelector as CSS and failed as an invalid selector, even though
  snapshot listed the button by that exact name. build_find_element_js now tries
  CSS first, then falls back to matching an interactive element by visible text
  (exact then contains) — nested and non-ASCII labels resolve. 'text=<label>'
  forces the text path. CSS still wins when it matches.
- #24-D: 'get text' with no selector now returns the whole page (body).
- #24-C: 'tab <ref> --activate' (alias --front) switches to the tab AND raises it
  to the foreground — to surface a specific tab for the human.

Tests cover the text fallback / text= / xpath builder, body default, activate
flag. The core stale-sessionId-after-cross-process-nav bug is the #20/#23 class,
already fixed in ext 0.4.8 — needs that extension deployed.
2026-06-15 11:00:49 +09:00
leeguooooo 7c594820da docs(stream): document the bidirectional WS as the real-time driving path
Dogfooding (driving a canvas game) showed the slow, low-fidelity way — one
screenshot + one CLI call per action — when chrome-use already ships the right
tool: the session WebSocket is BIDIRECTIONAL. It streams ~60fps screencast frames
AND accepts input_keyboard/input_mouse/input_touch on the same socket, straight
to CDP Input.dispatch* — verified live over the extension relay (217 frames in
3.4s, ~64fps, and the input drove the game). But the inbound input protocol was
undocumented, so agents default to the CLI-per-action grind.

Document it in --help (stream) and the core skill: the frame + input message
schemas and the 'connect once, read frames, send timed input' loop, with a node
snippet. Reserve screenshots for one-off checks; use the WS for sustained
real-time control.
2026-06-14 00:54:30 +09:00
leeguooooo 81d18bbd2e feat(input): press --hold <ms> for precise timed key-holds + document timed-driving pattern
Dogfooding by driving a canvas game surfaced that per-action shell round-trips
(keydown; sleep; keyup) are the slowest, lowest-fidelity way to drive anything
timed — each is a process spawn + relay round-trip with ~250ms jitter, so a
'0.8s hold' is anything but.

- 'press <key> --hold <ms>': keyDown, wait, keyUp all inside the daemon, so the
  hold duration is precise and it's one round-trip. For games (hold-to-move/
  charge) and any press-and-hold.
- Documented the real driving pattern in the core skill + --help: script a timed
  sequence in ONE round-trip with 'batch "press d --hold 900" "press j" "wait 200"'
  (batch sends each step to the running daemon; --hold/wait block in-daemon), and
  prefer reading engine state via main-world 'eval' over guessing from pixels.

Parser test covers plain/held/missing-duration. Builds on the keydown/keyup full
descriptor fix.
2026-06-14 00:43:52 +09:00
10 changed files with 467 additions and 75 deletions
+1 -1
View File
@@ -290,7 +290,7 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
[[package]]
name = "chrome-use"
version = "1.4.1"
version = "1.5.0"
dependencies = [
"aes",
"aes-gcm",
+1 -1
View File
@@ -1,6 +1,6 @@
[package]
name = "chrome-use"
version = "1.4.1"
version = "1.5.0"
edition = "2021"
description = "Fast browser automation CLI for AI agents"
license = "Apache-2.0"
+140 -26
View File
@@ -424,6 +424,9 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
// === Core Actions ===
"click" => {
let new_tab = rest.contains(&"--new-tab");
// `--follow`: if the click opens a new tab, switch the active tab to
// it (default reports the opened tab but stays put) (issue #24-A).
let follow = rest.contains(&"--follow");
// Coordinate click as a first-class form (issue #8.4): when the only
// handle is a pixel position, no element/selector is needed.
// click <x> <y> e.g. click 449 320
@@ -432,23 +435,27 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
let coord_args: Vec<&str> = rest
.iter()
.copied()
.filter(|a| *a != "--new-tab" && *a != "--coords")
.filter(|a| !a.starts_with("--"))
.collect();
if let Some((x, y)) = parse_coords(&coord_args) {
return Ok(json!({ "id": id, "action": "click", "x": x, "y": y }));
}
let sel = rest
.iter()
.find(|arg| **arg != "--new-tab")
.find(|arg| !arg.starts_with("--"))
.ok_or_else(|| ParseError::MissingArguments {
context: "click".to_string(),
usage: "click <selector> | click <x> <y> | click --coords <x>,<y> [--new-tab]",
usage:
"click <selector> | click <x> <y> | click --coords <x>,<y> [--new-tab] [--follow]",
})?;
let mut cmd = json!({ "id": id, "action": "click", "selector": sel });
if new_tab {
Ok(json!({ "id": id, "action": "click", "selector": sel, "newTab": true }))
} else {
Ok(json!({ "id": id, "action": "click", "selector": sel }))
cmd["newTab"] = json!(true);
}
if follow {
cmd["follow"] = json!(true);
}
Ok(cmd)
}
"dblclick" => {
let sel = rest.first().ok_or_else(|| ParseError::MissingArguments {
@@ -583,11 +590,27 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
// === Keyboard ===
"press" | "key" => {
let key = rest.first().ok_or_else(|| ParseError::MissingArguments {
context: "press".to_string(),
usage: "press <key>",
let key = rest.iter().find(|a| !a.starts_with("--")).ok_or_else(|| {
ParseError::MissingArguments {
context: "press".to_string(),
usage: "press <key> [--hold <ms>]",
}
})?;
Ok(json!({ "id": id, "action": "press", "key": key }))
let mut c = json!({ "id": id, "action": "press", "key": key });
// `--hold <ms>`: hold the key down for <ms> then release, timed inside
// the daemon (one round-trip, no shell-sleep jitter) — for games and
// hold-to-charge where keydown+sleep+keyup over 3 round-trips is too
// imprecise.
if let Some(i) = rest.iter().position(|a| *a == "--hold") {
let ms = rest.get(i + 1).and_then(|s| s.parse::<u64>().ok()).ok_or(
ParseError::MissingArguments {
context: "press --hold".to_string(),
usage: "press <key> --hold <ms>",
},
)?;
c["hold"] = json!(ms);
}
Ok(c)
}
"keydown" => {
let key = rest.first().ok_or_else(|| ParseError::MissingArguments {
@@ -1003,7 +1026,22 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
}
// === Close ===
"close" | "quit" | "exit" => Ok(json!({ "id": id, "action": "close" })),
"close" | "quit" | "exit" => {
// `close <tab>` closes only that tab (and the output says "Tab
// closed"); bare `close` closes the browser/session. `close --all` is
// intercepted earlier in the dispatcher. Previously `close t12` still
// ran a browser close and alarmingly printed "Browser closed" (#26).
if let Some(tab_ref) = rest.iter().find(|a| !a.starts_with("--")) {
Ok(json!({ "id": id, "action": "tab_close", "tabId": tab_ref }))
} else {
Ok(json!({ "id": id, "action": "close" }))
}
}
// The active tab's stable handle — `targetId` survives cross-process
// navigation and is reusable across sessions, so an agent can hold it
// instead of re-deriving "which tab is live" from `tabs` each step (#26).
"current" => Ok(json!({ "id": id, "action": "current" })),
// === Inspect ===
"inspect" => Ok(json!({ "id": id, "action": "inspect" })),
@@ -1559,11 +1597,16 @@ fn parse_command_inner(args: &[String], flags: &Flags) -> Result<Value, ParseErr
}
Ok(cmd)
}
Some(tab_ref) => Ok(json!({
"id": id,
"action": "tab_switch",
"tabId": tab_ref,
})),
Some(tab_ref) => {
// `tab <ref> --activate` (alias `--front`) switches to the tab
// AND raises it to the foreground — for handing a specific tab
// to the human (SMS code, captcha) (issue #24-C).
let mut cmd = json!({ "id": id, "action": "tab_switch", "tabId": tab_ref });
if rest.iter().any(|a| *a == "--activate" || *a == "--front") {
cmd["activate"] = json!(true);
}
Ok(cmd)
}
None => {
let mut cmd = json!({ "id": id, "action": "tab_list" });
if full {
@@ -2345,10 +2388,10 @@ fn parse_get(rest: &[&str], id: &str) -> Result<Value, ParseError> {
match rest.first().copied() {
Some("text") => {
let sel = rest.get(1).ok_or_else(|| ParseError::MissingArguments {
context: "get text".to_string(),
usage: "get text <selector>",
})?;
// `get text` with no selector returns the whole page's text (body) —
// a common convenience; previously it errored without a selector
// (issue #24-D).
let sel = rest.get(1).copied().unwrap_or("body");
Ok(json!({ "id": id, "action": "gettext", "selector": sel }))
}
Some("html") => {
@@ -3581,6 +3624,22 @@ mod tests {
assert_eq!(cmd["url"], "https://example.com");
}
#[test]
fn test_press_plain_and_hold() {
let cmd = parse_command(&args("press d"), &default_flags()).unwrap();
assert_eq!(cmd["action"], "press");
assert_eq!(cmd["key"], "d");
assert!(cmd.get("hold").is_none());
let held = parse_command(&args("press d --hold 800"), &default_flags()).unwrap();
assert_eq!(held["key"], "d");
assert_eq!(held["hold"], 800);
// Missing/invalid duration is an error, not a silent no-hold.
assert!(parse_command(&args("press d --hold"), &default_flags()).is_err());
assert!(parse_command(&args("press d --hold abc"), &default_flags()).is_err());
}
#[test]
fn test_navigate_reuse_tab_flag() {
let cmd = parse_command(
@@ -3747,6 +3806,21 @@ mod tests {
assert!(cmd.get("x").is_none());
}
#[test]
fn test_click_follow_flag() {
// `--follow` sets the flag; the selector is still found even with the flag
// before it (issue #24-A).
let cmd = parse_command(&args("click @e5 --follow"), &default_flags()).unwrap();
assert_eq!(cmd["selector"], "@e5");
assert_eq!(cmd["follow"], true);
let cmd2 = parse_command(&args("click --follow @e5"), &default_flags()).unwrap();
assert_eq!(cmd2["selector"], "@e5");
assert_eq!(cmd2["follow"], true);
// Absent by default.
let plain = parse_command(&args("click @e5"), &default_flags()).unwrap();
assert!(plain.get("follow").is_none());
}
#[test]
fn test_tabs_alias_lists() {
assert_eq!(
@@ -3955,6 +4029,28 @@ mod tests {
assert_eq!(cmd["tabId"], "docs");
}
#[test]
fn test_close_tab_vs_browser() {
// `close <tab>` closes that tab (says "Tab closed"); bare `close` closes
// the browser (#26).
let tab = parse_command(&args("close t12"), &default_flags()).unwrap();
assert_eq!(tab["action"], "tab_close");
assert_eq!(tab["tabId"], "t12");
let browser = parse_command(&args("close"), &default_flags()).unwrap();
assert_eq!(browser["action"], "close");
// `quit`/`exit` aliases still browser-close.
assert_eq!(
parse_command(&args("quit"), &default_flags()).unwrap()["action"],
"close"
);
}
#[test]
fn test_current_command() {
let cmd = parse_command(&args("current"), &default_flags()).unwrap();
assert_eq!(cmd["action"], "current");
}
#[test]
fn test_tab_sends_string_tab_id() {
let cmd = parse_command(&args("tab t2"), &default_flags()).unwrap();
@@ -4654,12 +4750,30 @@ mod tests {
}
#[test]
fn test_get_text_missing_selector() {
let result = parse_command(&args("get text"), &default_flags());
assert!(result.is_err());
let err = result.unwrap_err();
assert!(matches!(err, ParseError::MissingArguments { .. }));
assert!(err.format().contains("get text"));
fn test_get_text_defaults_to_body() {
// `get text` with no selector now returns the whole page (body) instead
// of erroring (issue #24-D).
let cmd = parse_command(&args("get text"), &default_flags()).unwrap();
assert_eq!(cmd["action"], "gettext");
assert_eq!(cmd["selector"], "body");
// An explicit selector still wins.
let cmd2 = parse_command(&args("get text h1"), &default_flags()).unwrap();
assert_eq!(cmd2["selector"], "h1");
}
#[test]
fn test_tab_activate_flag() {
let plain = parse_command(&args("tab t3"), &default_flags()).unwrap();
assert_eq!(plain["action"], "tab_switch");
assert!(plain.get("activate").is_none());
let act = parse_command(&args("tab t3 --activate"), &default_flags()).unwrap();
assert_eq!(act["action"], "tab_switch");
assert_eq!(act["tabId"], "t3");
assert_eq!(act["activate"], true);
// `--front` alias.
let front = parse_command(&args("tab t3 --front"), &default_flags()).unwrap();
assert_eq!(front["activate"], true);
}
// === Protocol alignment tests ===
+63 -1
View File
@@ -1395,6 +1395,7 @@ pub async fn execute_command(cmd: &Value, state: &mut DaemonState) -> Value {
"count" => handle_count(cmd, state).await,
"styles" => handle_styles(cmd, state).await,
"bringtofront" => handle_bringtofront(state).await,
"current" => handle_current(state).await,
"timezone" => handle_timezone(cmd, state).await,
"locale" => handle_locale(cmd, state).await,
"geolocation" => handle_geolocation(cmd, state).await,
@@ -3116,6 +3117,15 @@ async fn handle_click(cmd: &Value, state: &mut DaemonState) -> Result<Value, Str
let button = cmd.get("button").and_then(|v| v.as_str()).unwrap_or("left");
let click_count = cmd.get("clickCount").and_then(|v| v.as_i64()).unwrap_or(1) as i32;
let follow = cmd.get("follow").and_then(|v| v.as_bool()).unwrap_or(false);
// Snapshot tracked targets so we can tell if this click opened a NEW tab
// (target=_blank link / window.open). On the relay the new tab is discovered
// passively and doesn't steal focus (#7/#8.1), so without surfacing it the
// post-click snapshot shows the OLD page and looks like the click failed
// (issue #24-A).
let before: std::collections::HashSet<String> =
mgr.pages_list().into_iter().map(|p| p.target_id).collect();
interaction::click(
&mgr.client,
@@ -3128,7 +3138,26 @@ async fn handle_click(cmd: &Value, state: &mut DaemonState) -> Result<Value, Str
)
.await?;
Ok(json!({ "clicked": selector }))
// Give a just-opened tab a moment to register, then look for it.
let mgr = state.browser.as_mut().ok_or("Browser not launched")?;
tokio::time::sleep(std::time::Duration::from_millis(150)).await;
let opened = mgr.adopt_newly_opened(&before).await;
let mut out = json!({ "clicked": selector });
if let Some(page) = opened {
let tab_id = super::browser::format_tab_id(page.tab_id);
out["openedTab"] = json!({ "tabId": tab_id, "url": page.url, "title": page.title });
// `--follow`: switch the active tab to the newly-opened one (default is
// to report it but stay put, so multi-tab flows aren't hijacked).
if follow {
state.ref_map.clear();
state.iframe_sessions.clear();
state.active_frame_id = None;
let _ = mgr.tab_switch_by_id(page.tab_id).await;
out["followed"] = json!(true);
}
}
Ok(out)
}
async fn handle_dblclick(cmd: &Value, state: &mut DaemonState) -> Result<Value, String> {
@@ -3334,6 +3363,16 @@ async fn handle_press(cmd: &Value, state: &mut DaemonState) -> Result<Value, Str
// Parse modifier+key chords like "Control+a", "Shift+Enter", "Control+Shift+a"
let (actual_key, modifiers) = parse_key_chord(key);
// `--hold <ms>`: keyDown, wait, keyUp — all inside the daemon so the hold
// duration is precise (no shell-sleep / round-trip jitter). For games
// (hold-to-move/charge) and any press-and-hold interaction.
if let Some(ms) = cmd.get("hold").and_then(|v| v.as_u64()) {
interaction::dispatch_single_key(&mgr.client, &session_id, &actual_key, "keyDown").await?;
tokio::time::sleep(std::time::Duration::from_millis(ms)).await;
interaction::dispatch_single_key(&mgr.client, &session_id, &actual_key, "keyUp").await?;
return Ok(json!({ "pressed": key, "heldMs": ms }));
}
interaction::press_key_with_modifiers(&mgr.client, &session_id, &actual_key, modifiers).await?;
Ok(json!({ "pressed": key }))
}
@@ -4456,6 +4495,17 @@ async fn handle_tab_switch(cmd: &Value, state: &mut DaemonState) -> Result<Value
state.active_frame_id = None;
let result = mgr.tab_switch_by_id(tab_id).await?;
// `--activate`: raise this tab to the foreground (the switch made it active;
// bring_to_front acts on the active tab) — for handing a specific tab to the
// human (issue #24-C). Best-effort; don't fail the switch if it can't.
if cmd
.get("activate")
.and_then(|v| v.as_bool())
.unwrap_or(false)
{
let _ = mgr.bring_to_front().await;
}
if let Some(ref server) = state.stream_server {
if let Ok(dims) = mgr
.evaluate(
@@ -5284,6 +5334,18 @@ async fn handle_bringtofront(state: &DaemonState) -> Result<Value, String> {
Ok(json!({ "broughtToFront": true }))
}
async fn handle_current(state: &mut DaemonState) -> Result<Value, String> {
let mgr = state.browser.as_mut().ok_or("Browser not launched")?;
// Refresh so `current` reflects the live URL/title even after a cross-process
// nav (the relay's cached target_info can lag) (#26).
mgr.resync_targets().await.ok();
let mut info = mgr.active_page_info().ok_or("No active tab")?;
if let Some(obj) = info.as_object_mut() {
obj.insert("current".to_string(), json!(true));
}
Ok(info)
}
async fn handle_timezone(cmd: &Value, state: &DaemonState) -> Result<Value, String> {
let mgr = state.browser.as_ref().ok_or("Browser not launched")?;
let timezone = cmd
+77
View File
@@ -1263,6 +1263,22 @@ impl BrowserManager {
.collect()
}
/// The active tab's stable handle + current location, for `chrome-use
/// current` (#26). `targetId` survives cross-process navigation, so it's the
/// handle an agent should hold across a multi-step flow.
pub fn active_page_info(&self) -> Option<Value> {
let i = self.resolved_active_index();
self.pages.get(i).map(|p| {
json!({
"tabId": format_tab_id(p.tab_id),
"targetId": p.target_id,
"label": p.label,
"url": p.url,
"title": p.title,
})
})
}
/// Stable `tab_id` for a page identified by its CDP `targetId`, if tracked.
/// Lets callers adopt a tab by the cross-session-stable target id.
pub fn tab_id_for_target(&self, target_id: &str) -> Option<u32> {
@@ -1279,6 +1295,67 @@ impl BrowserManager {
/// the active tab is preserved, and re-pinned if it was pruned. Powers a live
/// `tab list` and adopt-by-targetId so a fresh session can reach a stranded,
/// still-filled tab without reloading it (issue #21).
/// Detect targets that appeared since the `before` set (e.g. a click that
/// opened a new tab via a `target=_blank` link or `window.open`), attach +
/// track each in the background, and return the first newly-opened page.
///
/// Lighter than [`resync_targets`] — one `getTargets` and work only on the
/// new targets, no whole-tab url/title refresh — so it's cheap enough to run
/// after every click. The new tab is added in the background (never steals
/// the active tab, per #7/#8.1); the caller surfaces it so the agent knows a
/// tab opened instead of seeing the old page (issue #24-A).
pub async fn adopt_newly_opened(&mut self, before: &HashSet<String>) -> Option<PageInfo> {
let result: GetTargetsResult = self
.client
.send_command_typed("Target.getTargets", &json!({}), None)
.await
.ok()?;
let live: Vec<TargetInfo> = result
.target_infos
.into_iter()
.filter(should_track_target)
.collect();
let mut opened: Option<PageInfo> = None;
for target in &live {
if before.contains(&target.target_id)
|| self.pages.iter().any(|p| p.target_id == target.target_id)
{
continue;
}
let attach: AttachToTargetResult = match self
.client
.send_command_typed(
"Target.attachToTarget",
&AttachToTargetParams {
target_id: target.target_id.clone(),
flatten: true,
},
None,
)
.await
{
Ok(r) => r,
Err(_) => continue,
};
let tab_id = self.assign_tab_id();
let page = PageInfo {
tab_id,
label: None,
target_id: target.target_id.clone(),
session_id: attach.session_id.clone(),
url: target.url.clone(),
title: target.title.clone(),
target_type: target.target_type.clone(),
};
self.add_background_page(page.clone());
let _ = self.enable_domains(&attach.session_id).await;
if opened.is_none() {
opened = Some(page);
}
}
opened
}
pub async fn resync_targets(&mut self) -> Result<(), String> {
self.client
.send_command_typed::<_, Value>(
+55 -8
View File
@@ -796,16 +796,43 @@ pub(super) fn extract_ax_string(value: &Option<AXValue>) -> String {
/// Build a JS expression that finds a DOM element by CSS selector or XPath.
fn build_find_element_js(selector: &str) -> String {
if let Some(xpath) = selector.strip_prefix("xpath=") {
format!(
return format!(
"document.evaluate({}, document, null, XPathResult.FIRST_ORDERED_NODE_TYPE, null).singleNodeValue",
serde_json::to_string(xpath).unwrap_or_default()
)
} else {
format!(
"document.querySelector({})",
serde_json::to_string(selector).unwrap_or_default()
)
);
}
// Bare string (or explicit `text=`): try CSS first, then fall back to
// matching an interactive element by its VISIBLE TEXT. snapshot exposes
// buttons/links by their name, so `click "購入手続きへ"` should resolve by
// that label — previously it was fed straight to `querySelector` as CSS and
// failed as an invalid selector even though the button was right there
// (issue #24-B). CSS still wins when it matches, so existing selectors are
// unaffected; nested/non-ASCII labels now resolve too.
let text_only = selector.strip_prefix("text=");
let force_text = text_only.is_some();
let sel_json = serde_json::to_string(selector).unwrap_or_default();
let want_json = serde_json::to_string(text_only.unwrap_or(selector)).unwrap_or_default();
format!(
r#"(() => {{
const sel = {sel};
const css = {force_text} ? null : (() => {{ try {{ return document.querySelector(sel); }} catch (_e) {{ return null; }} }})();
if (css) return css;
const norm = s => (s == null ? '' : String(s)).replace(/\s+/g, ' ').trim();
const w = norm({want}); if (!w) return null;
const wl = w.toLowerCase();
const interactive = Array.from(document.querySelectorAll(
'button,a,[role=button],[role=link],[role=menuitem],[role=tab],[role=option],input[type=submit],input[type=button],input[type=reset],summary,label,[onclick]'));
const textOf = e => norm(e.innerText || e.textContent) || norm(e.value) ||
norm(e.getAttribute && e.getAttribute('aria-label')) || norm(e.getAttribute && e.getAttribute('title'));
let hit = interactive.find(e => textOf(e) === w) || interactive.find(e => textOf(e).toLowerCase().includes(wl));
if (hit) return hit;
const leaves = Array.from(document.querySelectorAll('*')).filter(e => !e.children.length);
return leaves.find(e => norm(e.textContent) === w) || leaves.find(e => norm(e.textContent).toLowerCase().includes(wl)) || null;
}})()"#,
sel = sel_json,
want = want_json,
force_text = force_text
)
}
/// Build a JS expression that counts matching DOM elements by CSS selector or XPath.
@@ -1452,10 +1479,30 @@ mod tests {
#[test]
fn test_build_selector_js_css() {
let js = build_selector_js("#submit-btn");
assert!(js.contains("document.querySelector(\"#submit-btn\")"));
// CSS is now tried via a `sel` variable, with a visible-text fallback
// appended (issue #24-B). It must still use querySelector (not xpath).
assert!(js.contains("const sel = \"#submit-btn\""));
assert!(js.contains("document.querySelector(sel)"));
assert!(!js.contains("document.evaluate"));
}
#[test]
fn test_build_find_element_js_text_fallback() {
// A bare label gets a text-matching fallback so `click "購入手続きへ"`
// resolves by visible text, not just CSS (issue #24-B).
let js = build_find_element_js("購入手続きへ");
assert!(js.contains("購入手続きへ"));
assert!(js.contains("interactive")); // the text-match branch
assert!(js.contains("textOf"));
// `text=` forces the text path (skips CSS).
let forced = build_find_element_js("text=Buy now");
assert!(forced.contains("true ? null")); // force_text => css skipped
// xpath is unchanged.
let xp = build_find_element_js("xpath=//button");
assert!(xp.contains("document.evaluate"));
assert!(!xp.contains("interactive"));
}
#[test]
fn test_build_selector_js_xpath() {
let js = build_selector_js("xpath=//button[@id='ok']");
+39 -32
View File
@@ -306,32 +306,50 @@ pub async fn fill(
)
.await?;
// Focus the element
client
.send_command_typed::<_, Value>(
"Runtime.callFunctionOn",
&CallFunctionOnParams {
function_declaration: "function() { this.focus(); }".to_string(),
object_id: Some(object_id.clone()),
arguments: None,
return_by_value: Some(true),
await_promise: Some(false),
},
Some(&effective_session_id),
)
.await?;
// Emulate a real edit so framework-controlled inputs (React/Vue) and
// site-side listeners actually see the change (issue #25): the old path set
// `this.value` directly and used Input.insertText, which left React's
// internal value-tracker out of sync and never fired change/blur — so
// dependent logic (e.g. Mercari's postal-code → 都道府県 autocomplete) never
// ran even though the value was visible. Set the value through the element's
// PROTOTYPE setter (which React's _valueTracker hooks), then dispatch
// input → change → blur/focusout. `type <sel> <text>` remains for sites that
// need per-keystroke events.
let fill_js = format!(
r#"function() {{
const el = this;
const v = {val};
try {{ el.focus(); }} catch (e) {{}}
const tag = el.tagName;
const fire = (type, ctor) => el.dispatchEvent(new (ctor || Event)(type, {{ bubbles: true }}));
if (tag === 'SELECT') {{
el.value = v; fire('input'); fire('change'); return true;
}}
if (el.isContentEditable) {{
el.textContent = v; fire('input', window.InputEvent || Event); fire('change');
try {{ el.blur(); }} catch (e) {{}} fire('focusout'); return true;
}}
const proto = tag === 'TEXTAREA' ? window.HTMLTextAreaElement.prototype
: window.HTMLInputElement.prototype;
const desc = Object.getOwnPropertyDescriptor(proto, 'value');
const set = desc && desc.set ? (x) => desc.set.call(el, x) : (x) => {{ el.value = x; }};
set(''); // reset the framework tracker
fire('input', window.InputEvent || Event);
set(v); // native setter → React/Vue registers
fire('input', window.InputEvent || Event);
fire('change');
try {{ el.blur(); }} catch (e) {{}}
fire('focusout'); // blur-triggered lookups/validation
return true;
}}"#,
val = serde_json::to_string(value).unwrap_or_default()
);
// Select all + delete to clear
client
.send_command_typed::<_, Value>(
"Runtime.callFunctionOn",
&CallFunctionOnParams {
function_declaration: r#"function() {
this.select && this.select();
this.value = '';
this.dispatchEvent(new Event('input', { bubbles: true }));
}"#
.to_string(),
function_declaration: fill_js,
object_id: Some(object_id),
arguments: None,
return_by_value: Some(true),
@@ -341,17 +359,6 @@ pub async fn fill(
)
.await?;
// Insert text (keyboard input dispatched at page level, use parent session_id)
client
.send_command_typed::<_, Value>(
"Input.insertText",
&InsertTextParams {
text: value.to_string(),
},
Some(session_id),
)
.await?;
Ok(())
}
+55 -1
View File
@@ -186,6 +186,44 @@ pub fn print_response_with_opts(resp: &Response, action: Option<&str>, opts: &Ou
}
if let Some(data) = &resp.data {
// A click that opened a new tab: surface it so the agent doesn't read the
// unchanged old page as a failed click (issue #24-A).
if let Some(opened) = data.get("openedTab") {
let tid = opened.get("tabId").and_then(|v| v.as_str()).unwrap_or("?");
let url = opened.get("url").and_then(|v| v.as_str()).unwrap_or("");
let followed = data
.get("followed")
.and_then(|v| v.as_bool())
.unwrap_or(false);
let verb = if followed {
"switched to new tab"
} else {
"opened new tab"
};
eprintln!(
"{} {} [{}] {}",
color::cyan(""),
verb,
tid,
color::dim(url)
);
}
// `current`: the active tab's stable handle (#26).
if data
.get("current")
.and_then(|v| v.as_bool())
.unwrap_or(false)
{
let tid = data.get("tabId").and_then(|v| v.as_str()).unwrap_or("?");
let title = data.get("title").and_then(|v| v.as_str()).unwrap_or("");
let url = data.get("url").and_then(|v| v.as_str()).unwrap_or("");
let target = data.get("targetId").and_then(|v| v.as_str()).unwrap_or("");
println!("{} [{}] {} - {}", color::cyan(""), tid, title, url);
println!(" {}", color::dim(&format!("target: {}", target)));
return;
}
// Dialog status response
if action == Some("dialog") {
if let Some(has_dialog) = data.get("hasDialog").and_then(|v| v.as_bool()) {
@@ -2778,6 +2816,20 @@ Notes:
- Streaming is always enabled. Set AGENT_BROWSER_STREAM_PORT to bind to a
specific port instead of the default OS-assigned port.
The WS is BIDIRECTIONAL the high-throughput way to drive a live/real-time page
(games, canvas apps) instead of one screenshot + one CLI call per action:
- Server -> client (JSON text frames):
{"type":"frame","data":"<base64 jpeg>"} live screencast (~60fps)
plus status / tabs messages.
- Client -> server (send JSON text):
{"type":"input_keyboard","eventType":"keyDown|keyUp","key":" ","code":"Space",
"windowsVirtualKeyCode":32}
{"type":"input_mouse","eventType":"mousePressed|mouseReleased|mouseMoved",
"x":640,"y":360,"button":"left","clickCount":1}
{"type":"input_touch","eventType":"touchStart|touchEnd","touchPoints":[...]}
Connect once and run a tight local loop: read frames, send timed input no
per-action process spawn, no round-trip. Works over the extension relay too.
Global Options:
--json Output as JSON
--session <name> Use specific session
@@ -3070,7 +3122,9 @@ Core Commands:
dblclick <sel> Double-click element
type <sel> <text> Type into element
fill <sel> <text> Clear and fill
press <key> Press key (Enter, Tab, Control+a)
press <key> [--hold <ms>] Press key (Enter, Tab, Control+a). --hold keeps it
down <ms> then releases precise (in-daemon), for
games/charge: `press d --hold 800`
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`
+2 -2
View File
@@ -1,7 +1,7 @@
{
"name": "chrome-use",
"version": "1.4.1",
"description": "chrome-use drive your real, logged-in Chrome from any AI agent, stealth by default",
"version": "1.5.0",
"description": "chrome-use \u2014 drive your real, logged-in Chrome from any AI agent, stealth by default",
"type": "module",
"packageManager": "pnpm@11.1.3",
"files": [
+34 -3
View File
@@ -309,12 +309,43 @@ detects this and prints a one-line hint. Drive them the screenshot way:
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 d --hold 800 # hold-to-move, precise (timed in-daemon —
# NOT keydown+shell-sleep+keyup, which
# adds ~250ms jitter per round-trip)
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.
**Don't drive frame-by-frame with one CLI call per action** — that's the slowest,
lowest-fidelity way (each call is a process spawn + round-trip). Script a *timed
sequence in a single round-trip* with `batch` (it sends each step to the running
daemon; `press --hold` and `wait` block in-daemon, so timing is precise):
```bash
chrome-use batch "press d --hold 900" "press j" "press j" "wait 200" "press d --hold 500"
```
Also try reading real state instead of pixels: `eval` runs in the page's main
world, so for a framework/engine game you can often reach its globals (e.g. a
Phaser/PIXI/Three instance, a store, `window.__GAME__`) and read positions/score
directly — far better than guessing from a screenshot.
**For genuinely real-time driving, drop the CLI entirely and use the WebSocket.**
`chrome-use stream enable` opens a bidirectional WS (`stream status` prints the
`ws://127.0.0.1:<port>`). Connect once and you get a live ~60fps screencast AND
can send input on the same socket — no per-action process spawn, no round-trip,
works over the extension relay:
```js
// node (global WebSocket): live frames + locally-timed input
const ws = new WebSocket("ws://127.0.0.1:PORT")
ws.onmessage = e => { const m = JSON.parse(e.data); if (m.type==="frame") {/* base64 jpeg */} }
const k = (eventType,key,code,vk) => ws.send(JSON.stringify({type:"input_keyboard",eventType,key,code,windowsVirtualKeyCode:vk}))
k("keyDown"," ","Space",32); setTimeout(()=>k("keyUp"," ","Space",32), 80) // a jump
// also: {type:"input_mouse",eventType:"mousePressed",x,y,button:"left",clickCount:1}
```
This is the difference between watching a slideshow and playing the game. Reserve
screenshots for one-off checks; use the WS for any sustained real-time control.
## Waiting (read this)