@testdriverai/mcp 7.11.62-test → 7.11.63-test

package/agent/lib/redraw.js

@@ -232,7 +232,10 @@ const createRedraw = (
     
     // Capture initial image for screen stability monitoring
     if (currentOptions.screenRedraw) {
-      initialScreenImage = await system.captureScreenPNG(0.25, true);
+      // remoteResize=true: these frames exist only to be pixel-diffed at
+      // quarter scale, so let the runner shrink them before upload rather
+      // than shipping full-resolution PNGs on every poll.
+      initialScreenImage = await system.captureScreenPNG(0.25, true, false, true);
       lastScreenImage = initialScreenImage;
       emitter.emit(events.log.debug, `[redraw] start() - captured initial image: ${initialScreenImage}`);
     }
@@ -254,7 +257,7 @@ const createRedraw = (
       await updateNetwork();
     }
     
-    let nowImage = screenRedraw ? await system.captureScreenPNG(0.25, true) : null;
+    let nowImage = screenRedraw ? await system.captureScreenPNG(0.25, true, false, true) : null;
     let timeElapsed = Date.now() - startTime;
     let diffFromInitial = 0;
     let diffFromLast = 0;

package/agent/lib/system.js

@@ -57,13 +57,21 @@ const createSystem = (emitter, sandbox, config) => {
   // which is one of:
   //   { s3Key, width, height } — runner uploaded to S3 (Ably 64KB limit)
   //   { base64 }               — direct/local connection, bytes inline
-  const captureRaw = async () => {
+  //
+  // `target` ({ width, height }) asks the runner to do the downscale before it
+  // uploads. Only pass it when the bytes are going to be shrunk to that size
+  // on arrival regardless — it saves a multi-megabyte round-trip per capture,
+  // but it also moves the resample from Jimp (here) to sharp (there), so the
+  // pixels differ slightly. Anything feeding vision or the selector cache must
+  // keep resizing locally.
+  const captureRaw = async (target = null) => {
     return await sandbox.send({
       type: "system.screenshot",
+      ...(target ? { targetWidth: target.width, targetHeight: target.height } : {}),
     });
   };
 
-  const screenshot = async (options, rawResponse) => {
+  const screenshot = async (options, rawResponse, target = null) => {
     const MAX_RETRIES = 3;
     let lastError;
 
@@ -74,7 +82,7 @@ const createSystem = (emitter, sandbox, config) => {
         // capture fresh.
         let response = attempt === 0 && rawResponse
           ? rawResponse
-          : await captureRaw();
+          : await captureRaw(target);
 
         let base64;
 
@@ -125,7 +133,11 @@ const createSystem = (emitter, sandbox, config) => {
     return path.join(os.tmpdir(), `td-${Date.now()}-${randomUUID().slice(0, 8)}-${countImages}.png`);
   };
 
-  const captureAndResize = async (scale = 1, silent = false, mouse = false, rawResponse = null) => {
+  // `remoteResize` pushes the downscale onto the runner so the frame never
+  // crosses the wire at full resolution. Off by default: it changes which
+  // resampler produced the pixels, which callers feeding vision or the
+  // selector cache must not do. See `captureRaw`.
+  const captureAndResize = async (scale = 1, silent = false, mouse = false, rawResponse = null, remoteResize = false) => {
     try {
       if (!silent) {
         emitter.emit(events.screenCapture.start, {
@@ -138,21 +150,29 @@ const createSystem = (emitter, sandbox, config) => {
       let step1 = tmpFilename();
       let step2 = tmpFilename();
 
-      await screenshot({ filename: step1, format: "png" }, rawResponse);
+      const targetWidth = Math.floor(config.TD_RESOLUTION[0] * scale);
+      const targetHeight = Math.floor(config.TD_RESOLUTION[1] * scale);
+
+      await screenshot(
+        { filename: step1, format: "png" },
+        rawResponse,
+        remoteResize ? { width: targetWidth, height: targetHeight } : null,
+      );
 
       // Load the screenshot image with Jimp
       let image = await Jimp.read(step1);
-      
+
       // Validate the image was loaded correctly (not a 1x1 or tiny placeholder)
       if (image.getWidth() < 10 || image.getHeight() < 10) {
         throw new Error(`Screenshot appears corrupted: got ${image.getWidth()}x${image.getHeight()} pixels`);
       }
 
-      // Resize the image
-      image.resize(
-        Math.floor(config.TD_RESOLUTION[0] * scale),
-        Math.floor(config.TD_RESOLUTION[1] * scale),
-      );
+      // Resize the image. Still unconditional when the runner already resized:
+      // an older runner ignores targetWidth/targetHeight and returns a
+      // full-size frame, and Jimp treats a resize to the current size as a
+      // no-op, so this both stays correct across versions and guarantees every
+      // frame comes out at the same dimensions for pixel diffing.
+      image.resize(targetWidth, targetHeight);
 
       if (mouse) {
         // Only get mouse position when needed to avoid unnecessary websocket calls
@@ -196,8 +216,8 @@ const createSystem = (emitter, sandbox, config) => {
     return fs.readFileSync(step2, "base64");
   };
 
-  const captureScreenPNG = async (scale = 1, silent = false, mouse = false) => {
-    return await captureAndResize(scale, silent, mouse);
+  const captureScreenPNG = async (scale = 1, silent = false, mouse = false, remoteResize = false) => {
+    return await captureAndResize(scale, silent, mouse, null, remoteResize);
   };
 
   // Build the image payload to send to the API for a command (find/assert/etc).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@testdriverai/mcp",
-  "version": "7.11.62-test",
+  "version": "7.11.63-test",
   "description": "Next generation autonomous AI agent for end-to-end testing of web & desktop",
   "main": "sdk.js",
   "types": "sdk.d.ts",

package/vitest.unit.config.mjs

@@ -1,14 +1,28 @@
 import { defineConfig } from "vitest/config";
+import { dirname, resolve } from "path";
+import { fileURLToPath } from "url";
 
-// Unit tests for SDK internals — pure, no sandbox, no network.
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const monoRoot = resolve(__dirname, "..");
+
+// Unit tests for SDK and runner internals — pure, no sandbox, no network.
 //
 // Kept separate from vitest.config.mjs, which drives the real end-to-end
 // examples: that config provisions live sandboxes, resolves per-channel API
 // keys, and runs with an 8-minute timeout, so it can't host fast unit tests.
 // Run with `npm run test:unit`.
+//
+// Rooted at the mono repo so runner/ tests can live here too (same pattern as
+// vitest.runner.config.mjs) — dependencies still resolve from sdk/node_modules
+// for SDK tests and runner/node_modules for runner tests.
 export default defineConfig({
   test: {
-    include: ["agent/**/*.test.mjs", "lib/**/*.test.mjs"],
+    root: monoRoot,
+    include: [
+      "sdk/agent/**/*.test.mjs",
+      "sdk/lib/**/*.test.mjs",
+      "runner/lib/**/*.test.mjs",
+    ],
     // Don't let the e2e suites' fixtures/setup leak in.
     setupFiles: [],
   },