Skip to content

jevpilot v0.3.0

Choose a tag to compare

View all tags
@bloudhood bloudhood released this 04 Oct 06:42
· 5 commits to main since this release
v0.3.0
aae0877

The agent can now see the page: browser_screenshot returns the page or one element as an image, or saves it for the user. Tools can be switched off by configuration. Existing configurations keep working; the new settings are optional. Install from the package attached below (see the README quick start); verify it with SHA256SUMS.txt.

Added

  • browser_screenshot captures the current tab of a session as a JPEG: the visible viewport, or one element by ref from the latest observation (scrolled into view first, also inside frames). It returns the image plus one line of JSON, or saves the file and returns only its path. It refuses sessions that use secret values, and says so when a JavaScript dialog is open or the page does not respond. The result carries no structured content, because some MCP clients show the model only the structured part and would drop the image. Checked with Claude Code and Codex: both models read text and colours from the returned image.
  • JEVPILOT_IMAGE_RESPONSES (allow or omit), JEVPILOT_SCREENSHOT_DIR (a directory where saved screenshots are kept) and JEVPILOT_DISABLED_TOOLS (tools to leave unregistered) — see docs/configuration.md.

Changed

  • Handoff screenshots are also skipped while a session holds unused secret_ref values, matching the screenshot tool's rule.

Fixed

  • Screenshot files are JPEG; the handoff and observation file extension changed from .png to .jpg.
Assets 4
Loading