feat(tools): add browser_press_key

Author: rexleimo

crates/rexos-tools/src/browser_bridge.py

@@ -90,6 +90,32 @@ def main() -> int:
                         continue
                     page.fill(selector, text, timeout=timeout_ms)
                     respond({"success": True, "data": {"selector": selector, "typed": text}})
+                elif action == "PressKey":
+                    key = cmd.get("key", "")
+                    selector = cmd.get("selector", "")
+                    if not key:
+                        respond({"success": False, "error": "missing key"})
+                        continue
+                    if selector:
+                        page.press(selector, key, timeout=timeout_ms)
+                    else:
+                        page.keyboard.press(key)
+                    try:
+                        page.wait_for_load_state("domcontentloaded", timeout=timeout_ms)
+                    except Exception:
+                        pass
+                    current_url = page.url
+                    respond(
+                        {
+                            "success": True,
+                            "data": {
+                                "key": key,
+                                "selector": selector or None,
+                                "title": page.title(),
+                                "url": current_url,
+                            },
+                        }
+                    )
                 elif action == "ReadPage":
                     content = "(empty)"
                     try:
@@ -145,4 +171,3 @@ def respond(obj) -> None:
 
 if __name__ == "__main__":
     raise SystemExit(main())
-

crates/rexos-tools/src/lib.rs

@@ -89,6 +89,7 @@ impl Toolset {
             browser_navigate_def(),
             browser_click_def(),
             browser_type_def(),
+            browser_press_key_def(),
             browser_read_page_def(),
             browser_screenshot_def(),
             browser_close_def(),
@@ -263,6 +264,12 @@ impl Toolset {
                     serde_json::from_str(arguments_json).context("parse browser_type arguments")?;
                 self.browser_type(&args.selector, &args.text).await
             }
+            "browser_press_key" => {
+                let args: BrowserPressKeyArgs = serde_json::from_str(arguments_json)
+                    .context("parse browser_press_key arguments")?;
+                self.browser_press_key(args.selector.as_deref(), &args.key)
+                    .await
+            }
             "browser_read_page" => {
                 let _args: serde_json::Value = serde_json::from_str(arguments_json)
                     .context("parse browser_read_page arguments")?;
@@ -1289,6 +1296,24 @@ impl Toolset {
         Ok(resp.into_tool_output()?)
     }
 
+    async fn browser_press_key(&self, selector: Option<&str>, key: &str) -> anyhow::Result<String> {
+        let mut guard = self.browser.lock().await;
+        let session = guard
+            .as_mut()
+            .context("browser session not started; call browser_navigate first")?;
+
+        let mut cmd = serde_json::json!({
+            "action": "PressKey",
+            "key": key,
+        });
+        if let Some(sel) = selector {
+            cmd["selector"] = serde_json::Value::String(sel.to_string());
+        }
+
+        let resp = session.send(cmd).await?;
+        Ok(resp.into_tool_output()?)
+    }
+
     async fn browser_read_page(&self) -> anyhow::Result<String> {
         let mut guard = self.browser.lock().await;
         let session = guard
@@ -1737,6 +1762,13 @@ struct BrowserTypeArgs {
     text: String,
 }
 
+#[derive(Debug, serde::Deserialize)]
+struct BrowserPressKeyArgs {
+    key: String,
+    #[serde(default)]
+    selector: Option<String>,
+}
+
 #[derive(Debug, serde::Deserialize)]
 struct BrowserScreenshotArgs {
     #[serde(default)]
@@ -3184,6 +3216,26 @@ fn browser_type_def() -> ToolDefinition {
     }
 }
 
+fn browser_press_key_def() -> ToolDefinition {
+    ToolDefinition {
+        kind: "function".to_string(),
+        function: ToolFunctionDefinition {
+            name: "browser_press_key".to_string(),
+            description: "Press a key in the browser (optionally on a target element)."
+                .to_string(),
+            parameters: serde_json::json!({
+                "type": "object",
+                "properties": {
+                    "key": { "type": "string", "description": "Key to press (example: Enter, Escape, ArrowDown, Control+A)." },
+                    "selector": { "type": "string", "description": "Optional CSS selector to target before pressing the key." }
+                },
+                "required": ["key"],
+                "additionalProperties": false
+            }),
+        },
+    }
+}
+
 fn browser_read_page_def() -> ToolDefinition {
     ToolDefinition {
         kind: "function".to_string(),
@@ -3332,6 +3384,7 @@ mod tests {
             "browser_navigate",
             "browser_click",
             "browser_type",
+            "browser_press_key",
             "browser_read_page",
             "browser_screenshot",
             "browser_close",
@@ -4005,6 +4058,22 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn browser_press_key_requires_session() {
+        let tmp = tempfile::tempdir().unwrap();
+        let tools = Toolset::new(tmp.path().to_path_buf()).unwrap();
+
+        let err = tools
+            .call("browser_press_key", r#"{ "key": "Enter" }"#)
+            .await
+            .unwrap_err();
+        let msg = err.to_string();
+        assert!(
+            msg.contains("browser_navigate") || msg.contains("session"),
+            "{msg}"
+        );
+    }
+
     #[tokio::test]
     async fn browser_read_page_requires_session() {
         let tmp = tempfile::tempdir().unwrap();
@@ -4059,6 +4128,13 @@ mod tests {
             .await
             .unwrap();
 
+        let out = tools
+            .call("browser_press_key", r#"{ "key": "Enter" }"#)
+            .await
+            .unwrap();
+        let v: serde_json::Value = serde_json::from_str(&out).unwrap();
+        assert_eq!(v["key"], "Enter");
+
         let page = tools.call("browser_read_page", r#"{}"#).await.unwrap();
         let v: serde_json::Value = serde_json::from_str(&page).unwrap();
         assert_eq!(v["title"], "Stub");
@@ -4137,6 +4213,8 @@ for line in sys.stdin:
         resp = {"success": True, "data": {"clicked": cmd.get("selector", "")}}
     elif action == "Type":
         resp = {"success": True, "data": {"typed": cmd.get("text", ""), "selector": cmd.get("selector", "")}}
+    elif action == "PressKey":
+        resp = {"success": True, "data": {"key": cmd.get("key", ""), "selector": cmd.get("selector", "")}}
     elif action == "Close":
         resp = {"success": True, "data": {"status": "closed"}}
         sys.stdout.write(json.dumps(resp) + "\n")

docs-site/how-to/browser-automation.md

@@ -18,6 +18,7 @@ If your Python executable isn't `python3`, set `REXOS_BROWSER_PYTHON` (example:
 - `browser_navigate` — open a URL (SSRF-protected by default)
 - `browser_click` — click by CSS selector (best-effort text fallback)
 - `browser_type` — fill an input
+- `browser_press_key` — press a key (example: `Enter` to submit a form)
 - `browser_read_page` — return `{title,url,content}` (content is truncated)
 - `browser_screenshot` — write a PNG to a workspace-relative path
 - `browser_close` — close the session (idempotent)
@@ -27,6 +28,7 @@ If your Python executable isn't `python3`, set `REXOS_BROWSER_PYTHON` (example:
 1. `browser_navigate` to the entry page
 2. `browser_read_page` to confirm state
 3. One small action: `browser_click` or `browser_type`
+   - If you need to submit a form, use `browser_press_key` with `Enter`.
 4. `browser_read_page` again to confirm the page changed
 5. Repeat until done, then `browser_screenshot` for evidence and `browser_close`
 
@@ -46,7 +48,7 @@ If a CSS selector fails, `browser_click` will try a **best-effort visible-text f
 Use this as a starting point for agent prompts:
 
 ```text
-You may use RexOS browser tools (browser_navigate/click/type/read_page/screenshot/close).
+You may use RexOS browser tools (browser_navigate/click/type/press_key/read_page/screenshot/close).
 
 Rules:
 - Always call browser_read_page after navigate/click/type to verify page state before the next step.

docs-site/reference/tools.md

@@ -35,7 +35,7 @@ By default it rejects loopback/private IPs (basic SSRF protection). For local te
 
 Browser tools enable headless browser automation via a Python Playwright bridge:
 
-- `browser_navigate` / `browser_click` / `browser_type` / `browser_read_page` / `browser_screenshot` / `browser_close`
+- `browser_navigate` / `browser_click` / `browser_type` / `browser_press_key` / `browser_read_page` / `browser_screenshot` / `browser_close`
 
 Notes:
 

docs-site/zh/how-to/browser-automation.md

@@ -18,6 +18,7 @@ python3 -m playwright install chromium
 - `browser_navigate`：打开 URL（默认带 SSRF 防护）
 - `browser_click`：按 CSS selector 点击（会做尽力的可见文本 fallback）
 - `browser_type`：填写输入框
+- `browser_press_key`：按键（例如用 `Enter` 提交表单）
 - `browser_read_page`：返回 `{title,url,content}`（content 会被截断）
 - `browser_screenshot`：把 PNG 写入 workspace 相对路径
 - `browser_close`：关闭 session（可重复调用）
@@ -27,6 +28,7 @@ python3 -m playwright install chromium
 1. `browser_navigate` 打开入口页面
 2. `browser_read_page` 确认状态
 3. 每次只做一个小动作：`browser_click` 或 `browser_type`
+   - 需要提交表单时，用 `browser_press_key` 按 `Enter`。
 4. 再次 `browser_read_page` 确认页面确实变化
 5. 直到完成，最后 `browser_screenshot` 留证并 `browser_close`
 
@@ -44,7 +46,7 @@ python3 -m playwright install chromium
 ## Prompt 模板（可直接复制）
 
 ```text
-你可以使用 RexOS 的 browser 工具（browser_navigate/click/type/read_page/screenshot/close）。
+你可以使用 RexOS 的 browser 工具（browser_navigate/click/type/press_key/read_page/screenshot/close）。
 
 规则：
 - navigate/click/type 之后必须立刻 browser_read_page，先验证页面状态再做下一步。

docs-site/zh/reference/tools.md

@@ -35,7 +35,7 @@ RexOS 会强制超时，并使用尽量最小的环境。
 
 浏览器工具通过 Python Playwright bridge 提供无头浏览器自动化能力：
 
-- `browser_navigate` / `browser_click` / `browser_type` / `browser_read_page` / `browser_screenshot` / `browser_close`
+- `browser_navigate` / `browser_click` / `browser_type` / `browser_press_key` / `browser_read_page` / `browser_screenshot` / `browser_close`
 
 说明：
 

docs/alignment.md

@@ -106,9 +106,9 @@
 
 新增：
 - `web_fetch`：SSRF 防护（默认拒绝 loopback/private/link-local），支持 `allow_private=true`（便于本地测试/内网）
-- 兼容工具面（aliases + stubs）：
-  - 已实现：`file_read/file_write/file_list`、`apply_patch`、`shell_exec`、`web_search`、`memory_store/memory_recall`
-  - 已对齐命名但未实现（stub）：`agent_*` / `task_*` / `schedule_*` / `knowledge_*` / `cron_*` / `channel_send` / `a2a_*` / `process_*` 等（会返回 not implemented）
+- 兼容工具面（aliases + reserved names）：
+  - Toolset：`file_*`、`apply_patch`、`shell_exec`、`web_search`、`web_fetch`、`browser_*`、`a2a_*`、`docker_exec`、`process_*`、`speech_to_text`、`text_to_speech`、`canvas_present`
+  - Runtime：`memory_*`、`agent_*`、`hand_*`、`task_*`、`event_publish`、`schedule_*`、`knowledge_*`、`cron_*`、`channel_send`
 
 对应实现：
 - `crates/rexos-tools/src/lib.rs`：`fs_read/fs_write/shell/web_fetch`