AunySillyMelean-browser-use
Documentation

From CHANGELOG.md · e31dc91 · Built 2026-11-17

Changelog

All notable changes to this project are documented here, following Keep a Changelog.

[Unreleased]

Fixed

  • The runner agent no longer lists shortcuts_list and shortcuts_execute, which Claude Code 2.1.319 removed from Claude in Chrome.

0.1.4 - 2026-11-10

Changed

  • Hardened workflow credentials and published the verified package artifact.
  • Documented security response targets and current repository protections.
  • Test jobs run on GitHub-hosted runners.

0.1.3 - 2026-10-31

Fixed

  • Docs, agent and skill text run the CLI as npx lean-browser-use <verb>, since npx installs nothing on PATH.
  • Installed files no longer link to docs that are not installed.
  • The privacy gate accepts Dependabot's sign-off trailer; every other email stays blocked.
  • The privacy gate accepts the Hermes Agent co-author trailer.

Changed

  • Code scanning findings hardened and release tooling pinned; Dependabot keeps release tooling updated.
  • The release workflow creates the GitHub release page.
  • Test jobs run on instabox-medium runners.

0.1.2 - 2026-10-28

Added

  • PostToolUse touch hook records browser actions and successful closes per session and subagent.
  • Stop and runner SubagentStop gate requests closure once per dirty stretch, with the exact close tool.
  • keep --reason command holds browser surfaces for the user, including subagent surfaces.
  • doctor open-surfaces check reports dirty browser surfaces older than ten minutes.
  • --no-stop-gate install flag disables touch and stop-gate registration. The gate is on by default.

Changed

  • Cleanup guide and runner steps explain that closing the run's tabs ends browser control.

0.1.1 - 2026-10-27

Added

  • Full after-use cleanup steps for the runner and main session, including tabs, Browser panes, shell tasks, scratch files and verification.
  • Cleanup guide on the website, with links from the README, runner agent, skill reference and llms.txt.
  • cleanup command: dry run by default, --apply ends only leased automation processes, children first, with SIGTERM, a 3 second wait, then SIGKILL for the same process. It prints one summary line.
  • spawn command: starts a process detached and records a lease (process id, start time, command line, session).
  • Cleanup hooks for SubagentStop, Stop and SessionEnd, registered by the installer and removed by uninstall; each exits 0 on every path.
  • doctor check that reports leftover leased processes and exits 1 when any exist.
  • Runner agent and cleanup reference end with the exact cleanup command.

Changed

  • uninstall removes hook events it leaves empty instead of keeping empty lists.

0.1.0 - 2026-10-25

Added

  • Dollar cost rows and an about 25% cheaper headline, the best of two measured runs on one setup.
  • Whole-run context totals of 2,657,290 plain and 2,846,792 runner, with the 15.54-step break-even flagged as an estimate.
  • One paired browser-token benchmark with static proof, exact counts, and reproduction guidance.
  • Browser method skill, focused references, and a fresh-context runner.
  • Browser and cheap-helper packet templates with evidence and cleanup contracts.
  • Pick and retry guards plus a batching nudge.
  • Local installer, doctor, hash-aware uninstall, and transcript meter.
  • Optional compile-once page reader and SearXNG companion.
  • Offline checks, package measurements, provenance, and contribution guides.