@q-agent/agent 0.1.9 → 0.1.11

package/dist/src/runner.js

@@ -854,12 +854,40 @@ async function processExplorationJob(cfg, session) {
     const args = [script, session.baseUrl];
     if (storageState)
         args.push(storageState);
-    const child = (0, child_process_1.spawn)((0, paths_1.nodeBin)(), args, { cwd: nm, env: nodePathEnv(nm) });
+    // Run the explore browser HEADED on the paired device: the user watches the
+    // session, and a headed browser dodges WAF/bot-protection that blocks headless.
+    const child = (0, child_process_1.spawn)((0, paths_1.nodeBin)(), args, {
+        cwd: nm,
+        env: { ...nodePathEnv(nm), QAGENT_EXPLORE_HEADED: "1" },
+    });
     activeChild = child;
+    // Capture the driver's stderr so a launch failure (missing browser, bad
+    // Playwright resolution, …) is visible instead of a silent "complete".
+    let driverStderr = "";
+    child.stderr?.on("data", (d) => {
+        driverStderr += String(d);
+    });
     const driver = makeExploreDriver(child);
     try {
-        // Wait for the driver's readiness signal before the first observe.
-        await driver.ready();
+        // Wait for the driver's readiness signal before the first observe. If the
+        // driver died before signalling ready, finalize with a clear driver-error
+        // (carrying its stderr) rather than letting the loop break silently.
+        const ready = await driver.ready();
+        if (!ready || ready.ok === false) {
+            const error = (ready && ready.error) || "driver failed to start";
+            const detail = driverStderr.trim();
+            await api
+                .postExploreFinalize(cfg, session.sessionId, {
+                discovered: { routes: [], selectors: [] },
+                log: [{ phase: "driver", error, stderr: detail || undefined }],
+                stopReason: "driver-error",
+                stepsTaken: 0,
+            })
+                .catch((err) => console.error("postExploreFinalize failed:", err));
+            (0, bus_1.emit)("explore-complete", { sessionId: session.sessionId, stopReason: "driver-error", error });
+            console.error(`Exploration ${session.sessionId} driver-error: ${error}${detail ? ` — ${detail}` : ""}`);
+            return;
+        }
         await runExplorationLoop(cfg, session, driver, api, bus_1.emit);
     }
     finally {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@q-agent/agent",
-  "version": "0.1.9",
+  "version": "0.1.11",
   "description": "Q-Agent Local Agent — claims execution jobs from the Q-Agent server and runs Playwright locally, so manual login/MFA happens on the user's own machine and session credentials never leave it.",
   "license": "UNLICENSED",
   "bin": {

package/vendor/explore_session.cjs

@@ -18,7 +18,7 @@
 //
 // Protocol (one JSON object per line, each way):
 //   {"cmd":"observe"}
-//     -> {"ok":true,"url":..,"path":..,"a11y":<accessibility.snapshot()>,
+//     -> {"ok":true,"url":..,"path":..,"a11y":<ariaSnapshot() role+name tree>,
 //         "elements":[<distilled interactive DOM>]}
 //   {"cmd":"act","action":"goto|click|fill|expectVisible","args":{...}}
 //     -> {"ok":bool,"error":str|null,"changed":bool}
@@ -121,7 +121,11 @@ async function doAct(action, args) {
 }
 
 async function doObserve() {
-  const a11y = await page.accessibility.snapshot();
+  // Playwright removed page.accessibility; the supported role+name tree is
+  // locator.ariaSnapshot() (a compact YAML string, ideal for the model and
+  // maps directly to getByRole). Best-effort — never fail observe on it.
+  let a11y = '';
+  try { a11y = await page.locator('body').ariaSnapshot(); } catch { a11y = ''; }
   const elements = await page.evaluate(DISTILL);
   const url = page.url();
   let path = '';
@@ -155,12 +159,21 @@ async function handle(line) {
 
 (async () => {
   if (!baseURL) { console.error('explore_session: missing baseURL arg'); process.exit(1); }
-  browser = await chromium.launch({ headless: true });
+  // Headed when QAGENT_EXPLORE_HEADED=1 (the Local Agent sets it so the user can
+  // watch, and a headed browser trips WAF/bot-protection far less than headless);
+  // headless otherwise (e.g. the server, which has no display).
+  browser = await chromium.launch({ headless: process.env.QAGENT_EXPLORE_HEADED !== '1' });
   const contextOpts = { baseURL };
   if (storageState) contextOpts.storageState = storageState;
   const context = await browser.newContext(contextOpts);
   page = await context.newPage();
 
+  // Land on the app before the first observe so step 1 sees the real page,
+  // not about:blank. Best-effort — the model can still goto elsewhere.
+  if (baseURL) {
+    try { await page.goto(baseURL, { waitUntil: 'domcontentloaded' }); } catch {}
+  }
+
   // Serialize commands: one line in -> one response out, in order.
   const rl = readline.createInterface({ input: process.stdin });
   let chain = Promise.resolve();