[ahamsel]

Why

The benchmark leaned on email and shopping, so "powerful for any task" wasn't being measured. A new everyday suite of 13 common asks on public sites passed 11/13 on its first run in the user's Edge. The traces showed what to fix:

Task	Problem
Best Buy price	20s of a 28.6s task spent waiting for ads and trackers to finish loading
YouTube trailer	The answer had no link (the writer only saw page text)
Stack Overflow	A CAPTCHA led the planner through 20 calls of Google caches and repeat searches (47.7s)
Chat replies	Occasionally ended with filler

Separately, the user asked for obstacles and declines to be handled in one general way rather than case by case.

What changed

• Everyday suite (bench --suite everyday):
  • Tasks: weather, a driving time, a YouTube trailer, a product price, a headline, MDN, a recipe, currency, GitHub stars, Stack Overflow, and a few facts.
  • Currency and GitHub stars are checked against live APIs.
• Heavy sites: after a navigation, WebPilot proceeds once the DOM is ready, allowing at most 2.5s more for the full load. This applies to both drivers and all tab waits.
• Linked answers: each page read keeps its most task-relevant links (absolute URLs, ranked by relevance) for the answer writer and the answer check. The writer includes links for the items it mentions.
• One rule for obstacles, anything only the user can clear:
  • covers sign-in, CAPTCHA, paywall, verification code, payment details, consent, and declined actions;
  • never get past it or fake it; do everything else first; then say exactly what's needed.
  • Replaces the sign-in and CAPTCHA special cases.
• One rule for declines: any declined confirmation hands the decision to the model. It judges whether anything can still move forward (other parts, other sources) or whether the rest depends on it. A declined action is never offered to Jev again.
• Resilience: an answer stream that fails partway is withdrawn and retried once. It happened in a real run: "JSON error injected into SSE stream".
• Chat replies are direct, with no filler.

Results (user's real browsers)

Suite	Chrome	Edge
Everyday	13/13, median 6.3s	13/13, median 7.3s
Realistic (regression)	11/11, median 9.1s	11/11, median 9.3s

• Stack Overflow with a CAPTCHA: 14–18s (was 47.7s); it goes straight to another source.
• Best Buy: 15–16s (was 28.7s).
• The YouTube answer includes the watch link.

Also:

• npm test: 76/76.
• npm run extension:smoke: 11/11.
• npm run build passes.

🤖 Generated with Claude Code