Repository navigation
Renamed the repository from agy-cli-skill to agy-atelier. GitHub redirects the old URL. The marketplace
is now agy-atelier, so install with /plugin install agy-cli@agy-atelier.
Hardening
- agy runs with its terminal sandbox by default (
--no-sandbox/AGY_IMAGE_SANDBOX=0to opt out). Where the
system can't provide the sandbox, the run is retried without it, with a warning. - Hard timeout: a run still going 60 s after
--timeoutis stopped together with agy's helper processes. - The user's prompt is fenced in the instructions sent to agy, and agy is told to ignore any instructions inside it.
agy --versionis checked once per run; versions other than 1.3.x/1.4.x get a warning (AGY_SKIP_VERSION_CHECK=1).
Zero setup
- numpy + Pillow install themselves into a private venv (
~/.local/share/agy-atelier/venv) the first time a
command needs them (AGY_IMAGE_NO_AUTO_SETUP=1to disable). - PEP 723 inline dependency metadata, for script runners such as
uv run. - An interpreter that already has numpy + Pillow is used as is.
Windows
- Cross-platform installer
install.py(py install.pyon Windows);install.shnow calls it. - Windows-aware venv paths, re-exec and process-tree termination; UTF-8 decoding of agy output.
- CI runs the tests on Ubuntu (Python 3.9 and 3.12), macOS and Windows.
UX
--dry-runprints the planned agy commands, prompts and a time estimate without running anything.- The result JSON includes
agy_version,agy_runs(seconds, sandbox, token usage per run) andtotal_seconds. - Before 3 or more images, the remaining agy quota is logged, with a warning when a limit is almost used up.
--outvalues likeout.v2/orout\are treated as folders.
Docs
- New README with an animated demo, a third example (perfume bottle) and badges;
docs/make_demo.pyrebuilds the GIF.