Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

Ask eval: document one fresh profile per compared build

Claude6163cc9fd7d7
diff --git a/scripts/check-webgpu-ask.mjs b/scripts/check-webgpu-ask.mjs
index 599c1e5e9..a1842c530 100644
--- a/scripts/check-webgpu-ask.mjs
+++ b/scripts/check-webgpu-ask.mjs
@@ -11,14 +11,13 @@
// answer, which is checked against a loose expectation. Follow-up tasks reuse
// the previous task's tab and conversation.
//
-// --extension DIR unpacked extension to load (default src/chrome). The
-// extension ID, and so the model cache, follows the path:
-// reuse one directory and profile to compare two builds.
-// --reset-service-worker
-// Chrome keeps running the service worker it registered
-// for the previous build. This removes the profile's
-// service-worker registrations and script cache (not the
-// model cache) so the build on disk is loaded.
+// --extension DIR unpacked extension to load (default src/chrome).
+// --profile DIR browser profile. Use a fresh profile for each build you
+// compare: a persistent profile keeps running the service
+// worker registered for the build it first loaded, even
+// after the files change, so a reused profile can silently
+// test the old code. The trace check below reports the
+// prompt variant that actually answered.
// --model ID WebGPU preset (default Bonsai 2 2B Vision)
// --download download the model if it is not cached yet
// --tasks a,b run only these tasks
@@ -103,11 +102,6 @@ const tasks = [
await fs.mkdir(path.dirname(output), { recursive: true });
const result = { model, extension, startedAt: new Date().toISOString(), tasks: [] };
-if (args.includes('--reset-service-worker')) {
- for (const name of ['ScriptCache', 'Database']) {
- await fs.rm(path.join(profile, 'Default', 'Service Worker', name), { recursive: true, force: true });
- }
-}
let context;
try {
context = await chromium.launchPersistentContext(profile, {