sam3-segmentation 0.1.0

Install:

npx skills add lillyguisnet/sam3-segmentation                                   # any agent
/plugin install sam3-segmentation --marketplace lillyguisnet/sam3-segmentation  # Claude Code

First public release.

• Agent skill sam3-segmentation in the open Agent Skills
  format, installable with npx skills, as a Claude Code plugin, or by copying
  the folder (install.py).
• Direct SAM 3 calls: every instance of a phrase or example box, one object from
  clicks or a box, automatic "segment everything", and batches with one model load.
• Full workflow for detailed annotation: overlapping crops, raw candidates saved
  before any filter, score audits, review sheets and flags, recorded reversible
  review decisions, guarded boundary refinement, pixel-checked PNG / COCO RLE
  exports and a zoomable viewer.
• A demo on generated shapes that needs no GPU, model or account.
• Locked environments: a small base environment and a pinned SAM 3 / PyTorch
  2.10 (CUDA 12.8) environment, created by scripts/run outside the skill folder.
• Agent evals for claude plugin eval (evals/): when the skill should and
  should not be used, and four safety behaviours, each compared with no skill.
  They led to two rules: state the download size (about 11 GB) when asking to
  install, and never use or repeat a token the user pastes.
• Renamed from the unpublished detailed-segmentation. Saved run files keep
  their detailed-segmentation/... format names and stay readable.

Known limitations, planned for later versions:

• The skill has no rule for when a review is finished (for example "stop when a
  round finds nothing new") and no time or round budget.
• It does not tell the agent which reviewer to use; an agent may silently use its
  own default model as the "independent" reviewer.
• Nothing stops an agent from removing many proposals at once without having
  inspected each one.
• A single boundary refinement cannot be undone on its own; it can only be marked
  uncertain.
• Model inference needs Linux (or WSL2) and an NVIDIA GPU. CPU and Apple GPU
  inference are not implemented.