@poncho-ai/browser 0.6.21 → 0.6.23

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/browser@0.6.21 build /home/runner/work/poncho-ai/poncho-ai/packages/browser
+> @poncho-ai/browser@0.6.23 build /home/runner/work/poncho-ai/poncho-ai/packages/browser
 > tsup src/index.ts --format esm --dts
 
 CLI Building entry: src/index.ts
@@ -7,8 +7,8 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js 46.07 KB
-ESM ⚡️ Build success in 67ms
+ESM dist/index.js 46.85 KB
+ESM ⚡️ Build success in 85ms
 DTS Build start
-DTS ⚡️ Build success in 5074ms
+DTS ⚡️ Build success in 5359ms
 DTS dist/index.d.ts 13.69 KB

package/CHANGELOG.md

@@ -1,5 +1,28 @@
 # @poncho-ai/browser
 
+## 0.6.23
+
+### Patch Changes
+
+- [`2d518d1`](https://github.com/cesr/poncho-ai/commit/2d518d1666e69d63a94be8a781d11d0569e6af7b) Thanks [@cesr](https://github.com/cesr)! - Force the configured viewport on remote browsers (cloud provider / cdpUrl).
+  `launchOpts.viewport` is only honored when launching a local context, so a
+  Browserbase/Kernel/CDP session rendered at the provider's large default — the
+  page looked huge, content tiny, scrolling appeared broken, and tap coordinates
+  mismatched the frame after reconnect. After connecting, call
+  `setViewport(width, height)` so the page renders at the intended size and frames
+  - input stay consistent.
+
+## 0.6.22
+
+### Patch Changes
+
+- [`3588b19`](https://github.com/cesr/poncho-ai/commit/3588b19cf8e8fb112df3642b93e8a6aa4d4e3021) Thanks [@cesr](https://github.com/cesr)! - Steer the agent to use `browser_open` only as a last resort. The description now
+  tells it to prefer `web_fetch` for reading pages and a dedicated API/MCP
+  integration when one exists, and to reach for the browser only when those can't
+  do the job — a page web_fetch can't render, or operating a site/web app that has
+  no API and no MCP (e.g. logging in and clicking through a UI). Reinforces that
+  credentials are entered by the user in the live view, never asked for in chat.
+
 ## 0.6.21
 
 ### Patch Changes

package/dist/index.js

@@ -427,6 +427,13 @@ var BrowserSession = class {
       launchOpts.args = baseArgs;
     }
     await mgr.launch(launchOpts);
+    if (this.isRemote) {
+      try {
+        await mgr.setViewport(viewport.width ?? 1280, viewport.height ?? 720);
+      } catch (e) {
+        console.warn(`[poncho][browser] setViewport failed: ${e?.message ?? e}`);
+      }
+    }
     this._contextStealthInstalled = false;
     this._uaOverrideApplied.clear();
     if (this.stealthEnabled) {
@@ -1079,7 +1086,7 @@ function createBrowserTools(getSession) {
   return [
     {
       name: "browser_open",
-      description: "Open a URL in a headless browser. Returns the page title. Use this to navigate to websites and web applications. To open files from the virtual filesystem, use /api/vfs/{path} (e.g. /api/vfs/downloads/report.pdf).",
+      description: "Open a URL in a real browser whose live view the user can watch and interact with. Returns the page title. This is a HEAVY, last-resort tool \u2014 prefer cheaper options first: use `web_fetch` to read page content, and use a dedicated API or MCP integration when one exists for the service. Only reach for the browser when those can't do the job: a page `web_fetch` can't render (JS-heavy/SPA), or a task that requires operating a site or web app that has no API and no MCP integration (e.g. logging in and clicking through a UI). When a site needs credentials, navigate to its login page and let the user sign in directly in the live view \u2014 never ask for passwords in chat. To open files from the virtual filesystem, use /api/vfs/{path} (e.g. /api/vfs/downloads/report.pdf).",
       inputSchema: {
         type: "object",
         properties: {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/browser",
-  "version": "0.6.21",
+  "version": "0.6.23",
   "description": "Browser automation for Poncho agents, powered by agent-browser",
   "repository": {
     "type": "git",

package/src/session.ts

@@ -365,6 +365,18 @@ export class BrowserSession {
 
     await mgr.launch(launchOpts as Parameters<BrowserManagerInstance["launch"]>[0]);
 
+    // Remote browsers (cloud provider / cdpUrl) ignore launchOpts.viewport —
+    // that's only applied when launching a local context — so the page renders
+    // at the provider's (often huge) default and the screencast shows it shrunk.
+    // Force our configured viewport on the connected page.
+    if (this.isRemote) {
+      try {
+        await mgr.setViewport(viewport.width ?? 1280, viewport.height ?? 720);
+      } catch (e) {
+        console.warn(`[poncho][browser] setViewport failed: ${(e as Error)?.message ?? e}`);
+      }
+    }
+
     // Reset stealth tracking for fresh browser
     this._contextStealthInstalled = false;
     this._uaOverrideApplied.clear();

package/src/tools.ts

@@ -10,7 +10,10 @@ export function createBrowserTools(
     {
       name: "browser_open",
       description:
-        "Open a URL in a headless browser. Returns the page title. Use this to navigate to websites and web applications. " +
+        "Open a URL in a real browser whose live view the user can watch and interact with. Returns the page title. " +
+        "This is a HEAVY, last-resort tool — prefer cheaper options first: use `web_fetch` to read page content, and use a dedicated API or MCP integration when one exists for the service. " +
+        "Only reach for the browser when those can't do the job: a page `web_fetch` can't render (JS-heavy/SPA), or a task that requires operating a site or web app that has no API and no MCP integration (e.g. logging in and clicking through a UI). " +
+        "When a site needs credentials, navigate to its login page and let the user sign in directly in the live view — never ask for passwords in chat. " +
         "To open files from the virtual filesystem, use /api/vfs/{path} (e.g. /api/vfs/downloads/report.pdf).",
       inputSchema: {
         type: "object",