BrowserSkill Files

BrowserSkill 0.4.1 coordinates the CLI/daemon, browser extension and DSH plugin.

Changed

• Shared-control default: new CLI and DSH sync calls select target policy. DOM actions can reuse the returned control context across completed actions and supplementary reads; unrelated page updates no longer require a full-page resync. Human intervention or navigation revokes control. Use --expect / DSH expect to protect observed business facts. Scripts, untargeted keys and visual actions still require the sync's --observation / observationId and strict observation checks. Select --policy strict / policy: "strict" on the first sync to retain the one-action-per-context workflow; later syncs and steps preserve that session's policy.
• CLI/daemon and extension now use protocol 1.8. The new target policy requires both peers to support 1.8; update the CLI, running daemon, extension and DSH plugin together rather than relying on a successful mixed-version connection.
• DOM references remain attached to the same nodes across supplementary observations. Removed nodes are not silently replaced, and navigation invalidates the document's refs. DOM actions can wait for an existing target to become ready within their timeout without replaying input.

Fixed

• Interaction capture initializes on ordinary HTTP intranet pages and their child frames when crypto.randomUUID() is unavailable. A compatible UUID fallback restores the frame coverage needed for shared control.
• Shared-control synchronization waits for cancelled observation thumbnails to settle. Viewport capture reads have a 10-second deadline with separate bounded cleanup; reaching that deadline returns timeout (exit 4) with screenshot-specific guidance. Browser/CDP capture failures remain cdp_failed (exit 3).
• Completed inputs retain their successful receipts when they navigate, including clicks, Enter submissions, drag submissions, and forms submitted after filling or selecting a value. After navigation, synchronize before the next action.
• CLI cancellation is armed before sending a browser command, so an immediate Ctrl-C can cancel the operation and return its outcome instead of terminating silently.
• Pointer hit tests stay within the target's document, and human input during frame changes continues to revoke control without discarding completed actions.
• Windows daemon startup errors distinguish verified Job Object restrictions, unverified breakaway failures, executable access failures, partial breakaway and resume failures, with recovery guidance for the actual failure stage.
• DSH-injected messages use native durable-format v4 plugin sources while retaining the legacy source shape for older formats.

Removed

• Automatic native Agent Window tab groups and the tabGroups permission, to prevent new BrowserSkill groups from accumulating across browsers through Chrome sync. Existing saved groups are not deleted; page control indicators, background window behavior and user-tab routing remain available.

Upgrade

1. Finish active tasks, install CLI 0.4.1 and restart the running daemon.
2. Install extension 0.4.1 before upgrading the DSH plugin to 0.4.1. Store review can delay availability; verify the installed extension, not just its submission.
3. Restart the DSH profile and invoke /browser-skill again, or start a new CLI agent conversation to load the updated skill. Custom tool integrations need the new observation and expectation fields for target-policy actions.
4. Check bsk --version, bsk status and bsk doctor; the running daemon and extension should both report protocol 1.8.

Full changelog: https://github.com/Tencent/BrowserSkill/blob/b8f2d4f0fa35c8c7c866229e1cf3b3329ae39de5/CHANGELOG.md