mirror of
https://github.com/openswarm-ai/openswarm.git
synced 2026-08-17 18:25:42 +02:00
[eric] agents: an agent's click hands the physical cursor straight back, so a canvas app can't hold the user's mouse (ENG-310)
This commit is contained in:
@@ -169,6 +169,7 @@ const getPort = require('get-port');
|
||||
const http = require('http');
|
||||
const affiliateTracking = require('./affiliateTracking');
|
||||
const cdpRoutes = require('./cdp-routes');
|
||||
const { releaseAgentPointerLock } = require('./releaseAgentPointerLock');
|
||||
const workflowsLifecycle = require('./workflowsLifecycle');
|
||||
|
||||
// Squirrel makes the APP create its own shortcuts: on --squirrel-install it must
|
||||
@@ -4335,6 +4336,8 @@ async function sendCdpCommandSerialized(wcId, method, params, sessionId) {
|
||||
ipcMain.handle('send-cdp-command', async (_event, wcId, method, params, sessionId) => {
|
||||
try {
|
||||
const result = await sendCdpCommandSerialized(wcId, method, params, sessionId);
|
||||
// A synthetic click can leave a canvas app holding the user's real cursor; hand it straight back (ENG-310).
|
||||
releaseAgentPointerLock(sendCdpCommandSerialized, wcId, method, params);
|
||||
return { ok: true, result };
|
||||
} catch (err) {
|
||||
return { ok: false, error: err && err.message ? err.message : String(err) };
|
||||
|
||||
@@ -0,0 +1,39 @@
|
||||
// The agent must never be able to swallow the user's real mouse (ENG-310).
|
||||
//
|
||||
// A CDP click is a trusted user gesture as far as Chromium is concerned, so a canvas app that calls
|
||||
// requestPointerLock() from its mousedown handler (every mouse-look game does) captures and HIDES the
|
||||
// physical cursor the instant the agent taps it. The user, who is off doing something else, finds
|
||||
// their pointer welded to a dashboard card for the rest of the run. Measured directly: an agent-
|
||||
// dispatched click on such a canvas left document.pointerLockElement=CANVAS.
|
||||
//
|
||||
// Denying the 'pointerLock' permission was the obvious fix and it is WRONG. Chromium caches that
|
||||
// decision per origin and asks exactly once, so whoever clicks first decides for the whole session:
|
||||
// deny for the agent and the user is denied too, forever, on their own app. Measured, both arms.
|
||||
//
|
||||
// So let the lock be granted and take it straight back off the agent's own click. Nothing about the
|
||||
// user's path changes: they still get pointer lock, from the same cached grant, whenever they click
|
||||
// it themselves. Exiting a lock needs no user gesture, which is what makes this side of it possible.
|
||||
|
||||
/** The one command that can end with the cursor captured: the release that completes a synthetic click. */
|
||||
function isSyntheticClickRelease(method, params) {
|
||||
return method === 'Input.dispatchMouseEvent' && !!params && params.type === 'mouseReleased';
|
||||
}
|
||||
|
||||
const EXIT_EXPRESSION = 'document.pointerLockElement ? (document.exitPointerLock(), true) : false';
|
||||
|
||||
/**
|
||||
* Give the cursor back after a synthetic click. Fire and forget: a failure here means the eval did
|
||||
* not land, which is no worse than not trying, and it must never fail the command it follows.
|
||||
*/
|
||||
function releaseAgentPointerLock(sendCdp, wcId, method, params) {
|
||||
if (!isSyntheticClickRelease(method, params)) return false;
|
||||
try {
|
||||
const p = sendCdp(wcId, 'Runtime.evaluate', { expression: EXIT_EXPRESSION, returnByValue: true });
|
||||
if (p && typeof p.catch === 'function') p.catch(() => {});
|
||||
} catch {
|
||||
// The surface went away mid-click; there is no cursor left to hand back.
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
module.exports = { releaseAgentPointerLock, isSyntheticClickRelease, EXIT_EXPRESSION };
|
||||
@@ -0,0 +1,94 @@
|
||||
// Run: node --test electron/releaseAgentPointerLock.test.js
|
||||
//
|
||||
// ENG-310: "the app agent takes over the physical cursor." Measured mechanism, not a guess: an
|
||||
// agent-dispatched click on a canvas that requests pointer lock left document.pointerLockElement=
|
||||
// CANVAS, i.e. the user's real mouse captured and hidden, and with this module wired the same click
|
||||
// logged ["locked","unlocked"] and ended at null.
|
||||
//
|
||||
// These pin the properties that decide whether it can come back: only the release half of a
|
||||
// synthetic CLICK triggers a handback (so reads and keystrokes cost nothing), a failure can never
|
||||
// break the command it follows, and main.js actually calls it.
|
||||
const test = require('node:test');
|
||||
const assert = require('node:assert/strict');
|
||||
const fs = require('node:fs');
|
||||
const path = require('node:path');
|
||||
|
||||
const { releaseAgentPointerLock, isSyntheticClickRelease, EXIT_EXPRESSION } = require('./releaseAgentPointerLock');
|
||||
|
||||
function recorder() {
|
||||
const calls = [];
|
||||
const send = (wcId, method, params) => {
|
||||
calls.push({ wcId, method, params });
|
||||
return Promise.resolve({});
|
||||
};
|
||||
return { calls, send };
|
||||
}
|
||||
|
||||
test('the release half of a synthetic click hands the cursor back', () => {
|
||||
const { calls, send } = recorder();
|
||||
assert.equal(releaseAgentPointerLock(send, 42, 'Input.dispatchMouseEvent', { type: 'mouseReleased' }), true);
|
||||
assert.equal(calls.length, 1);
|
||||
assert.equal(calls[0].wcId, 42, 'the handback must go to the surface that was clicked');
|
||||
assert.equal(calls[0].method, 'Runtime.evaluate');
|
||||
assert.equal(calls[0].params.expression, EXIT_EXPRESSION);
|
||||
});
|
||||
|
||||
test('the expression is a no-op when nothing is locked', () => {
|
||||
// It runs after every agent click, so it must cost nothing on the overwhelmingly common path.
|
||||
const doc = { pointerLockElement: null, exitPointerLock: () => { throw new Error('must not be called'); } };
|
||||
const run = new Function('document', `return (${EXIT_EXPRESSION});`);
|
||||
assert.equal(run(doc), false);
|
||||
});
|
||||
|
||||
test('the expression exits exactly one lock when there is one', () => {
|
||||
let exits = 0;
|
||||
const doc = { pointerLockElement: {}, exitPointerLock: () => { exits += 1; } };
|
||||
const run = new Function('document', `return (${EXIT_EXPRESSION});`);
|
||||
assert.equal(run(doc), true);
|
||||
assert.equal(exits, 1);
|
||||
});
|
||||
|
||||
test('nothing else in a click triggers a handback, so one click costs one eval', () => {
|
||||
const { calls, send } = recorder();
|
||||
for (const params of [{ type: 'mouseMoved' }, { type: 'mousePressed' }, { type: 'mouseWheel' }]) {
|
||||
assert.equal(releaseAgentPointerLock(send, 1, 'Input.dispatchMouseEvent', params), false);
|
||||
}
|
||||
assert.equal(calls.length, 0);
|
||||
});
|
||||
|
||||
test('reads and keystrokes never pay for it', () => {
|
||||
const { calls, send } = recorder();
|
||||
for (const method of ['Page.captureScreenshot', 'Runtime.evaluate', 'Accessibility.getFullAXTree', 'Input.dispatchKeyEvent']) {
|
||||
assert.equal(releaseAgentPointerLock(send, 1, method, { type: 'mouseReleased' }), false, `${method} must not trigger a handback`);
|
||||
}
|
||||
assert.equal(calls.length, 0);
|
||||
});
|
||||
|
||||
test('missing or malformed params are ignored rather than thrown on', () => {
|
||||
const { calls, send } = recorder();
|
||||
assert.equal(releaseAgentPointerLock(send, 1, 'Input.dispatchMouseEvent', undefined), false);
|
||||
assert.equal(releaseAgentPointerLock(send, 1, 'Input.dispatchMouseEvent', null), false);
|
||||
assert.equal(calls.length, 0);
|
||||
});
|
||||
|
||||
test('a handback that fails never breaks the click it follows', () => {
|
||||
const thrower = () => { throw new Error('debugger detached'); };
|
||||
assert.doesNotThrow(() => releaseAgentPointerLock(thrower, 1, 'Input.dispatchMouseEvent', { type: 'mouseReleased' }));
|
||||
const rejecter = () => Promise.reject(new Error('target closed'));
|
||||
assert.doesNotThrow(() => releaseAgentPointerLock(rejecter, 1, 'Input.dispatchMouseEvent', { type: 'mouseReleased' }));
|
||||
});
|
||||
|
||||
test('the classifier is exact, never a prefix guess', () => {
|
||||
assert.equal(isSyntheticClickRelease('Input.dispatchMouseEvent', { type: 'mouseReleased' }), true);
|
||||
assert.equal(isSyntheticClickRelease('Input.dispatchMouseEventExtra', { type: 'mouseReleased' }), false);
|
||||
assert.equal(isSyntheticClickRelease('Input.dispatchMouseEvent', { type: 'mouseReleasedish' }), false);
|
||||
});
|
||||
|
||||
test('main.js calls it, and only on the agent CDP path', () => {
|
||||
const src = fs.readFileSync(path.join(__dirname, 'main.js'), 'utf8');
|
||||
assert.match(src, /releaseAgentPointerLock\(sendCdpCommandSerialized, wcId, method, params\)/,
|
||||
'a handback nothing calls hands nothing back');
|
||||
const handler = src.slice(src.indexOf("ipcMain.handle('send-cdp-command'"));
|
||||
assert.ok(handler.indexOf('releaseAgentPointerLock') < handler.indexOf('return { ok: true, result }'),
|
||||
'the handback belongs inside the agent command path, where a human click never goes');
|
||||
});
|
||||
Reference in New Issue
Block a user