Skip to content

BetterWright 0.4.0

Choose a tag to compare

View all tags
@CuriosityOS CuriosityOS released this 15 Jul 13:38
v0.4.0
5e6c53a

Added

  • Native Pi Coding Agent integration — BetterWright is now a Pi package (pi install npm:betterwright). The pi.extensions manifest loads a persistent browser tool, an approval-gated browser_download tool, screenshot vision blocks, bounded step budgets, optional trace recording, and typed JavaScript exports, without an MCP hop.
  • Online-Mind2Web benchmark harness — deterministic runner with 50-task and full 300-task manifests, development/holdout partitions, v2 trajectory validation, resumable subset runs, and a strict local multimodal judge.
  • Recorded 300-task campaign report — 278/300 (92.7%) under the local strict judge, with public-safe results, difficulty breakdown, and dataset hashes under benchmarks/online-mind2web/.

Changed

  • connectOverCdp now has a display-aware default: on a desktop with a real Google Chrome, BetterWright attaches to that Chrome over CDP (real logins and extensions); headless or Chrome-less environments keep the managed Cloak sandbox with the network floor intact. Pass connectOverCdp: "" to force the sandbox.
  • publicSearchPolicy now defaults to "allow"; set it (or BETTERWRIGHT_PUBLIC_SEARCH_POLICY) to "block" to route broad discovery through a host-supplied search tool.

Fixed

  • Pi start-page navigation failures are recoverable and retained as trace evidence.
  • Challenge detection and Pi trace screenshots follow only the active page, so a stale CAPTCHA tab cannot replace current evidence.
  • The stock-Chromium fallback uses its own profile-chromium directory, preventing a shared-profile format upgrade from crashing the older Cloak binary; an already-poisoned profile now fails with a clear, actionable error.
  • Local benchmark judging retains the first trajectory frame.

Full details in CHANGELOG.md.

Assets 2
Loading