WebVoyager: 643-task live-web agent benchmark on Kernel browsers - #42

Open

rgarcia wants to merge 11 commits into

hypeship/cua-bench-online-mind2webfrom

hypeship/bench-webvoyager

Commits on Jun 28, 2026

Add WebVoyager adapter for the Kernel env + cua agent
rgarcia
committed
533a8fb
webvoyager: make the verifier judge pip-free (stdlib HTTPS to Anthropic)

rgarcia
and
claude
committed
f5455b4
webvoyager: record 20-task live smoke (10/20) + harden the judge
rgarcia
committed
1b0ac9c
webvoyager: reconcile SMOKE.md with harbor's final summary
rgarcia
committed
eb921de
webvoyager: record the opus-4-8 live smoke (14/20, Mean 0.70)

rgarcia
and
claude
committed
41d7733
webvoyager: raise MAX_IMAGES judge default 3 -> 15 to match canonical auto-eval

rgarcia
and
claude
committed
b4e7e16
webvoyager: convert the verifier judge to a pi-ai node bin

rgarcia
and
claude
committed
729ec25
webvoyager: send o-series reasoning effort, drop temperature in the pi-ai judge

rgarcia
and
claude
committed
0e859b2
webvoyager: validate --max-images and reject negative --limit

rgarcia
and
claude
committed
cad3863
webvoyager: fix default --output-dir to adapters/webvoyager/.tasks

rgarcia
and
claude
committed
b8d8bba
webvoyager: fail closed on artifact-read errors and plumb OPENAI_API_KEY

rgarcia
and
claude
committed
a15dafe