Skip to content

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

1.7 KB
Raw Blame

Repeated navigation without page progress

The navigation loop guard distinguishes a new browser document from useful page progress. A verified link can reload the same page with fresh accessibility references and a new one-use acceptance key. Repeating read, scroll and click must not consume the entire agent budget in that state.

The guard compares bounded root accessibility observations after a verified link navigation. It requires the actual control label and href, an independently reported URL-only effect, and the same page content and semantic route. A second unchanged return asks the agent to inspect the blocker; a fourth stops retries and requests an honest handoff. Re-reading or successfully scrolling between attempts does not establish navigation progress.

Only this retry identity collapses explicit nonce parameters with long hex values. Navigation, permissions, recipient binding, query and hash freshness continue using the full URL. Different search/page/hash views, changed content, successful field writes and new native chat deltas remain progress. Partial, subtree and conversation accessibility reads cannot assert an unchanged whole page. Fresh documents still discard stale targets and failure counters; a new run discards semantic retry history too.

test/semantic-page-retry.mjs replays a sanitized real thirty-tool signup rules loop in both production browser agents and standalone detectors. It verifies bounded recovery and the progress boundaries above. Public rules text, synthetic URLs, synthetic references and synthetic opaque values are used; no account or personal data is present. This corrects the product's retry behavior and does not establish that an external site's acceptance endpoint can advance.