@proagentstore/cli 0.4.0 → 0.4.2

package/dist/browser-runner/runner.js

@@ -712,8 +712,7 @@ export class LocalRunner {
                 return mcp.callTool("browser_select_option", { element: label, target: this.refOf(a), values: [a.text ?? ""] });
             case "key":
                 return mcp.callTool("browser_press_key", { key: a.key || "Enter" });
-            case "scroll":
-                return mcp.callTool("browser_evaluate", { function: `() => window.scrollBy(0, ${Math.round(a.dy ?? 600)})` });
+            // scroll is handled before this mapper (mouse-wheel on the live page).
             case "wait":
                 return mcp.callTool("browser_wait_for", { time: Math.min(5, Math.max(0.1, (a.ms ?? 1000) / 1000)) });
             default:
@@ -764,9 +763,45 @@ export class LocalRunner {
         }
         if (this.applyResumePath)
             this.armFileAutoAttach(page, this.applyResumePath);
+        // Scroll is a VIEWPORT primitive, not an element action — no ref, no standard
+        // tool in @playwright/mcp 0.0.77. Use the mouse wheel on the live page (as the
+        // prior runtime did): it scrolls the actual hovered/inner scroll container,
+        // whereas window.scrollBy is a no-op on sites that scroll an inner panel (e.g.
+        // PageUp/Coles), which made the brain scroll-loop forever.
+        if (action.action === "scroll") {
+            await page.mouse.wheel(0, Math.round(action.dy ?? 600)).catch(() => undefined);
+            await page.waitForTimeout(300).catch(() => undefined);
+            const active = await this.getActivePage();
+            return {
+                ok: true,
+                url: active.url(),
+                title: await active.title().catch(() => ""),
+                challenge: await detectHumanChallenge(active),
+            };
+        }
         const mcp = await this.getMcp();
         const res = await this.callBrowserTool(mcp, action);
         const text = mcp.textOf(res).trim();
+        // A native page dialog (alert/confirm/beforeunload) puts the standard server
+        // into a modal state that blocks EVERY other tool until it's handled. The brain
+        // has no dialog vocabulary (the old runtime relied on Playwright's default
+        // auto-dismiss, which @playwright/mcp overrides), so clear it transparently and
+        // report it as a successful, informative step — the brain proceeds on the next
+        // snapshot instead of thrashing into a needless human handoff.
+        if (res.isError && /modal state/i.test(text) && /dialog/i.test(text)) {
+            const dialogMsg = (text.match(/dialog with message "([^"]*)"/)?.[1] || "").slice(0, 160);
+            await mcp.callTool("browser_handle_dialog", { accept: true }).catch(() => undefined);
+            await page.waitForTimeout(400).catch(() => undefined);
+            const settled = await this.getActivePage();
+            await settled.waitForLoadState("domcontentloaded", { timeout: 6_000 }).catch(() => undefined);
+            return {
+                ok: true,
+                url: settled.url(),
+                title: await settled.title().catch(() => ""),
+                challenge: await detectHumanChallenge(settled),
+                feedback: dialogMsg ? `a native dialog was accepted: "${dialogMsg}"` : "a native dialog was accepted",
+            };
+        }
         if (res.isError)
             throw new RunnerInputError(this.conciseFeedback(text) || `${action.action} failed`);
         // A click may trigger SPA navigation or open a new tab/popup. Give it a beat

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@proagentstore/cli",
-  "version": "0.4.0",
+  "version": "0.4.2",
   "description": "CLI for creating, publishing, and running ProAgentStore agents",
   "license": "MIT",
   "type": "module",