jev-browse v0.2.0 — first PyPI release

The first release published to PyPI. The v0.1.0 tag predates the packaging and was never uploaded, so this is
the first version that uv tool install jev-browse or pipx install jev-browse can install.

Added

• CI on pull requests and main: pytest on Python 3.11–3.14 (with the mcp extra), ruff, a wheel build and
  clean-venv smoke test, and a full-history gitleaks scan, summarised by one ci check. Plus CodeQL, Dependabot
  (uv and GitHub Actions), OpenSSF Scorecard, and a CodeRabbit review config.
• PyPI packaging: uv tool install jev-browse (or pipx install jev-browse) gives a jev-browse command
  (install, uninstall, doctor, config, mcp). The wheel bundles the skill, and jev-browse install links it
  from the package.
• A Claude Code plugin and marketplace: /plugin marketplace add danielnc/jev-browse, then
  /plugin install jev-browse@jev-browse. It ships the skill plus /jev-browse:setup.
• A GitHub Actions workflow that publishes to PyPI on v* tags with Trusted Publishing (docs/releasing.md).
• jev-browse mcp: an MCP server on stdio for Cursor, Claude Desktop, Codex, and other MCP clients (install with
  uv tool install "jev-browse[mcp]"). Tools: fast_run, fast_run_status, jev_open, jev_find, jev_click,
  jev_check, jev_close, doctor. Each call runs through the browser-harness CLI with harness telemetry off.
  Results are concise text with a next: instruction. A fast_run longer than mcp.wait_s returns a resumable
  run_id. The MCP SDK is the optional mcp extra; the harness runtime stays stdlib-only. See docs/mcp.md.
• Settings mcp.harness_command and mcp.wait_s.
• A demo GIF at the top of the README: a real benchmark pair on Google Flights, the agent alone against the
  agent calling fast_run, with each run's time and list-price cost. scripts/make_demo_gif.py re-creates it
  (record, review, render; see docs/benchmarking.md).
• Codecov coverage upload from CI (one Python version), and README badges for PyPI, Python versions,
  Codecov, CodeRabbit reviews, and the license.
• Docstrings on the public API (fast_run, the jev_* helpers, the CLI, config, doctor, harness_api) and on
  scripts/make_demo_gif.py; a CodeRabbit pre-merge check for docstrings on public jev_browse/ functions.
• Offline tests for the jev-browse command dispatch and the browser-harness adapter (harness_api).

Changed

• The checkout launcher moved from bin/jev-browse to scripts/jev-browse. Claude Code puts a plugin's bin/
  on the Bash PATH, where the launcher would have shadowed a missing package install.
• doctor and the installer print jev-browse <command> hints in a package install.

Install: uv tool install jev-browse (add [mcp] for the MCP server) · https://pypi.org/project/jev-browse/0.2.0/

jev-browse v0.1.0

Added

• fast_run(url, goal, ...): a browser sub-task on an owned background tab, one TypeSafe Jev request per decision,
  with typed hand-backs (Reason), resume on the same tab, a commit gate, evidence, a trace, and a verify-ready
  JEV_BROWSE_RESULT line.
• Interactive helpers: jev_open, jev_adopt, jev_find, jev_click, jev_check, jev_close.
• Values from the caller, from goal-derived candidates chosen by Jev, or from a text backend after a miss, with a
  grounding gate. Personal fields are never sent to a text backend. Sensitive fields are never typed, read, or
  sent.
• Text backends: claude (subscription CLI, pre-started process), codex (subscription CLI, unmeasured),
  ollama, openai (any OpenAI-compatible endpoint), and none. Local and OpenAI-compatible backends run a
  known-answer canary, cached across processes, and a configurable fallback.
• One configuration surface: environment variables (including the browser-harness .env) over
  ~/.config/jev-browse/config.toml over defaults (docs/configuration.md).
• python3 -m jev_browse install | uninstall | doctor | config, and bin/jev-browse. The installer links the
  skill for Claude Code and Codex.
• The jev-browse agent skill, an agent-followable install.md, and a global-pointer snippet.
• A benchmark (bench/run_bench.py) with code-only verification and list-price cost, and a text-backend eval
  (bench/text_eval.py).

Ports code from browser-use/jev-ultrafast (MIT).