browser-automation-cli 0.7.2__tar.gz → 0.8.2__tar.gz

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/Cargo.lock

@@ -34,7 +34,7 @@ dependencies = [
 
 [[package]]
 name = "browser-cli"
-version = "0.7.2"
+version = "0.8.2"
 dependencies = [
  "base64",
  "futures-util",

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "browser-cli"
-version = "0.7.2"
+version = "0.8.2"
 edition = "2021"
 description = "Browser automation daemon + CLI for coding agents. Persistent sessions, no MCP, no extensions."
 license = "MIT"

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: browser-automation-cli
-Version: 0.7.2
+Version: 0.8.2
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: MacOS
@@ -141,7 +141,7 @@ All print JSON (snapshot prints text). Add `-s` / `--snapshot` to any action to
 | :-- | :-- |
 | `navigate <url> [--wait load\|domcontentloaded\|networkidle]` | Returns as soon as the page is usable; never fails on a slow `networkidle` |
 | `snapshot [scope] [--all] [--max N] [--json]` | Visible interactive elements + headings. `--all` adds text blocks, `--json` gives boxes and unique selectors |
-| `click <target> [--double]` | |
+| `click <target> [--double]` | Also `click --at X,Y` to click raw viewport pixels (canvas / vision — no DOM target; screenshot first, pixels map 1:1) |
 | `type <target> <text> [--sequential] [--submit]` | `fill()` by default; `--sequential` sends key events (autocomplete); `--submit` presses Enter |
 | `press <key> [target]` | `Enter`, `Tab`, `Control+a`, … |
 | `hover <target>` | |
@@ -155,7 +155,7 @@ All print JSON (snapshot prints text). Add `-s` / `--snapshot` to any action to
 | `back` / `forward` | |
 | `batch` | JSON lines on stdin, run in one round-trip, stop at first failure |
 
-**Targets:** `@e12` (ref from snapshot — preferred) · CSS selector · `text=Create` · `role=button[name=Create]` · `label=Email` · `placeholder=Search` · or flags `--text / --role [--name] / --label / --placeholder`. Ambiguous CSS selectors are refused (strict mode) instead of clicking the first match.
+**Targets:** `@e12` (ref from snapshot — preferred) · CSS selector · `text=Create` · `role=button[name=Create]` · `label=Email` · `placeholder=Search` · or flags `--text / --role [--name] / --label / --placeholder`. Ambiguous CSS selectors are refused (strict mode) instead of clicking the first match. For canvas / vision cases with no DOM target, `click --at X,Y` clicks raw viewport pixels (take a `screenshot` first; its pixels map 1:1 to click coordinates).
 
 ***
 

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/README.md

@@ -126,7 +126,7 @@ All print JSON (snapshot prints text). Add `-s` / `--snapshot` to any action to
 | :-- | :-- |
 | `navigate <url> [--wait load\|domcontentloaded\|networkidle]` | Returns as soon as the page is usable; never fails on a slow `networkidle` |
 | `snapshot [scope] [--all] [--max N] [--json]` | Visible interactive elements + headings. `--all` adds text blocks, `--json` gives boxes and unique selectors |
-| `click <target> [--double]` | |
+| `click <target> [--double]` | Also `click --at X,Y` to click raw viewport pixels (canvas / vision — no DOM target; screenshot first, pixels map 1:1) |
 | `type <target> <text> [--sequential] [--submit]` | `fill()` by default; `--sequential` sends key events (autocomplete); `--submit` presses Enter |
 | `press <key> [target]` | `Enter`, `Tab`, `Control+a`, … |
 | `hover <target>` | |
@@ -140,7 +140,7 @@ All print JSON (snapshot prints text). Add `-s` / `--snapshot` to any action to
 | `back` / `forward` | |
 | `batch` | JSON lines on stdin, run in one round-trip, stop at first failure |
 
-**Targets:** `@e12` (ref from snapshot — preferred) · CSS selector · `text=Create` · `role=button[name=Create]` · `label=Email` · `placeholder=Search` · or flags `--text / --role [--name] / --label / --placeholder`. Ambiguous CSS selectors are refused (strict mode) instead of clicking the first match.
+**Targets:** `@e12` (ref from snapshot — preferred) · CSS selector · `text=Create` · `role=button[name=Create]` · `label=Email` · `placeholder=Search` · or flags `--text / --role [--name] / --label / --placeholder`. Ambiguous CSS selectors are refused (strict mode) instead of clicking the first match. For canvas / vision cases with no DOM target, `click --at X,Y` clicks raw viewport pixels (take a `screenshot` first; its pixels map 1:1 to click coordinates).
 
 ***
 

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/SKILL.md

@@ -73,6 +73,7 @@ browser shutdown
 browser <id> navigate <url> [-s]        # -s on any action: append a fresh snapshot
 browser <id> snapshot [scope] [--all] [--max N] [--json]
 browser <id> click <target> [--double] [-s]
+browser <id> click --at X,Y [--double]   # click raw viewport pixels (canvas / vision — no DOM target)
 browser <id> type <target> <text> [--sequential] [--submit] [-s]   # alias: fill
 browser <id> press <key> [target]       # Enter, Tab, Escape, Control+a
 browser <id> hover <target>
@@ -87,7 +88,7 @@ browser <id> back | forward
 browser <id> batch                      # JSON lines on stdin, one round-trip
 ```
 
-**Targets:** `@e12` (ref from snapshot — preferred) · `--text "Create"` · `--role button --name Create` · `--label "Email"` · `--placeholder Search` · CSS selector. Ambiguous CSS selectors are refused; use a ref or text.
+**Targets:** `@e12` (ref from snapshot — preferred) · `--text "Create"` · `--role button --name Create` · `--label "Email"` · `--placeholder Search` · CSS selector. Ambiguous CSS selectors are refused; use a ref or text. For canvas / vision cases with no DOM target, `click --at X,Y` clicks raw viewport pixels (take a `screenshot` first; its pixels map 1:1 to click coordinates).
 
 ## Snapshot Format
 

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/install.sh

@@ -1,7 +1,7 @@
 #!/bin/sh
 # Install the Rust browser/browser-daemon binaries from the GitHub release into ~/.local/bin.
 set -e
-V="${BROWSER_CLI_VERSION:-0.7.2}"; OS=$(uname -s | tr A-Z a-z); ARCH=$(uname -m)
+V="${BROWSER_CLI_VERSION:-0.8.2}"; OS=$(uname -s | tr A-Z a-z); ARCH=$(uname -m)
 URL="https://github.com/jshan9078/browser-automation-cli/releases/download/v$V/browser-cli-$V-$OS-$ARCH.tar.gz"
 DEST="${BROWSER_CLI_BIN:-$HOME/.local/bin}"; mkdir -p "$DEST"
 echo "Downloading $URL"; curl -fsSL "$URL" | tar -C "$DEST" -xzf -

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "browser-automation-cli"
-version = "0.7.2"
+version = "0.8.2"
 description = "Browser automation daemon + CLI for coding agents. Persistent sessions, no MCP, no extensions."
 readme = "README.md"
 requires-python = ">=3.9"

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/src/actions.rs

@@ -61,6 +61,17 @@ impl Page {
         }
     }
 
+    /// Recover a wedged session. A stuck renderer main thread makes Runtime.evaluate (snapshot/eval/
+    /// click) hang until the CDP timeout, and it never comes back on its own. Page.reload is handled by
+    /// the browser process (not the stuck renderer), so it replaces the renderer with a fresh one and
+    /// un-wedges the session. Bounded so recovery itself can't hang. Loses page state (the session was
+    /// already unusable), but restores a working session for the caller to retry.
+    pub async fn recover(&self) -> Result<(), String> {
+        self.send("Page.reload", json!({"ignoreCache": false})).await?;
+        let _ = self.wait_ready("interactive", 15000).await;
+        Ok(())
+    }
+
     async fn after(&self, mut result: Map<String, Value>, snap: bool, settle: bool) -> Value {
         if settle || snap { self.settle().await; }
         result.insert("url".into(), json!(self.url().await));
@@ -189,8 +200,14 @@ pub async fn snapshot_p(page: &Page, p: &Map<String, Value>) -> Value {
 }
 
 pub async fn click(page: &Page, p: &Map<String, Value>) -> Value {
-    let r = match page.resolve(&target_of(p), ACTION_TIMEOUT_MS).await { Ok(r) => r, Err(e) => return fail(e) };
-    let (x, y) = (r["x"].as_f64().unwrap_or(0.0), r["y"].as_f64().unwrap_or(0.0));
+    // Raw viewport pixel click (`click --at X,Y`) when x/y are given; else resolve a DOM target.
+    let (x, y) = match (p.get("x").and_then(|v| v.as_f64()), p.get("y").and_then(|v| v.as_f64())) {
+        (Some(x), Some(y)) => (x, y),
+        _ => {
+            let r = match page.resolve(&target_of(p), ACTION_TIMEOUT_MS).await { Ok(r) => r, Err(e) => return fail(e) };
+            (r["x"].as_f64().unwrap_or(0.0), r["y"].as_f64().unwrap_or(0.0))
+        }
+    };
     if let Err(e) = page.click_at(x, y, if flag(p, "double") { 2 } else { 1 }).await { return fail(e); }
     page.after(ok(), flag(p, "snap"), true).await
 }

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/src/cdp.rs

@@ -63,7 +63,11 @@ impl Cdp {
         let mut msg = json!({"id": id, "method": method, "params": params});
         if let Some(sid) = session { msg["sessionId"] = json!(sid); }
         self.tx.send(msg.to_string()).map_err(|_| "cdp send: connection closed".to_string())?;
-        match tokio::time::timeout(std::time::Duration::from_secs(60), r).await {
+        // Per-command timeout. Configurable via BROWSER_CDP_TIMEOUT_MS so a wedged renderer fails fast
+        // (default 60s, unchanged). On timeout the pending entry is dropped and the caller gets an error
+        // it can act on (server.rs then attempts a reload-based recovery).
+        let timeout_ms: u64 = std::env::var("BROWSER_CDP_TIMEOUT_MS").ok().and_then(|v| v.parse().ok()).filter(|&v| v > 0).unwrap_or(60_000);
+        match tokio::time::timeout(std::time::Duration::from_millis(timeout_ms), r).await {
             Ok(Ok(v)) => v,
             Ok(Err(_)) => Err("cdp: response channel dropped".into()),
             Err(_) => { self.pending.lock().unwrap().remove(&id); Err(format!("cdp: {method} timed out")) }

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/src/main.rs

@@ -42,7 +42,7 @@ Page commands (all print JSON; add -s/--snapshot to include a fresh snapshot in
   browser <id> snapshot [scope-selector] [--all] [--max N] [--json]
                                          Interactive elements as "@e12 button "Create"" lines.
                                          --all adds text blocks; --json gives structured output
-  browser <id> click <target> [--double]
+  browser <id> click <target> [--double]      (or `--at X,Y` to click raw viewport pixels — canvas/vision)
   browser <id> type <target> <text> [--sequential] [--submit]   (alias: fill)
   browser <id> press <key> [target]      e.g. Enter, Tab, Control+a
   browser <id> hover <target>
@@ -128,7 +128,7 @@ fn parse_flags(args: &[String], bools: &[&str], valued: &[&str]) -> (Vec<String>
 }
 
 const BOOLS: &[&str] = &["-s", "--snapshot", "--all", "--json", "--double", "--sequential", "--submit", "--gone", "-f", "--full-page", "--clear", "--table"];
-const VALUED: &[&str] = &["--text", "--role", "--name", "--label", "--placeholder", "--wait", "--max", "--timeout", "-o", "--output", "-q", "--quality", "--selector", "--format"];
+const VALUED: &[&str] = &["--text", "--role", "--name", "--label", "--placeholder", "--wait", "--max", "--timeout", "-o", "--output", "-q", "--quality", "--selector", "--format", "--at"];
 
 fn target_params(f: &HashMap<String, Value>) -> Map<String, Value> {
     let mut p = Map::new();
@@ -166,7 +166,19 @@ fn build(sid: &str, action: &str, rest: &[String]) -> Option<Value> {
             req("snapshot", p)
         }
         "click" => {
-            p.insert("selector".into(), json!(sel(0)));
+            // `--at X,Y` clicks raw viewport pixels (for canvas / vision tasks with no DOM target);
+            // otherwise resolve a normal target (@ref / selector / --text / ...).
+            if let Some(at) = f.get("at").and_then(|v| v.as_str()) {
+                let mut it = at.split(',');
+                let x = it.next().and_then(|s| s.trim().parse::<f64>().ok());
+                let y = it.next().and_then(|s| s.trim().parse::<f64>().ok());
+                match (x, y) {
+                    (Some(x), Some(y)) => { p.insert("x".into(), json!(x)); p.insert("y".into(), json!(y)); }
+                    _ => return None,
+                }
+            } else {
+                p.insert("selector".into(), json!(sel(0)));
+            }
             if f.contains_key("double") { p.insert("double".into(), json!(true)); }
             req("click", p)
         }

{browser_automation_cli-0.7.2 → browser_automation_cli-0.8.2}/src/server.rs

@@ -39,6 +39,13 @@ async fn dispatch(page: &Page, session_id: &str, action: &str, params: &Map<Stri
     }
 }
 
+/// True if an action failed specifically because a CDP command timed out (a wedged renderer), which is
+/// recoverable via reload — as opposed to an ordinary action failure (bad selector, JS exception, etc.).
+fn is_cdp_timeout(result: &Value) -> bool {
+    result.get("success").and_then(|v| v.as_bool()) == Some(false)
+        && result.get("error").and_then(|v| v.as_str()).map(|e| e.contains("timed out")).unwrap_or(false)
+}
+
 pub async fn process(shared: &Shared, request: &Value, shutdown: &tokio::sync::watch::Sender<bool>) -> Value {
     let action = request.get("action").and_then(|v| v.as_str()).unwrap_or("");
     let session_id = request.get("session_id").and_then(|v| v.as_str()).map(String::from);
@@ -70,7 +77,16 @@ pub async fn process(shared: &Shared, request: &Value, shutdown: &tokio::sync::w
                 _ => {
                     let page = match shared.lock().await.wake(&sid).await { Ok(p) => p, Err(e) => return json!({"success": false, "error": e}) };
                     // the manager lock is NOT held while the action runs, so other sessions proceed in parallel
-                    let result = dispatch(&page, &sid, action, &params, shared).await;
+                    let mut result = dispatch(&page, &sid, action, &params, shared).await;
+                    // Wedge recovery: a CDP timeout means the renderer main thread is stuck (commands hang
+                    // and never recover on their own). Page.reload replaces the renderer with a fresh one;
+                    // then retry the command ONCE. Skipped for navigate (it already loads a fresh document).
+                    if is_cdp_timeout(&result) && action != "navigate" {
+                        if page.recover().await.is_ok() {
+                            result = dispatch(&page, &sid, action, &params, shared).await;
+                            if let Some(o) = result.as_object_mut() { o.insert("recovered".into(), json!(true)); }
+                        }
+                    }
                     let title = page.title().await;
                     let mut m = shared.lock().await;
                     if let Some(s) = m.sessions.get_mut(&sid) { s.title = title; }