What’s new
- Added a fully synthetic public demo lab with sentinel-owned prepare, inventory, stale-preview exercise, post-apply checks, and safe reset commands.
- Added a staged five-minute walkthrough that preserves both human approval gates and never auto-extracts or submits a preview token.
- Added wrong-token and stale-preview refusal tests proving zero partial approved edits, plus protected P0/P1 and ordinary-empty-directory invariants on the success path.
- Added an explicit non-zero-risk safety boundary, a redacted 16:9 evidence card, and updated Chinese self-media materials.
- Expanded the full regression suite from 195 to 214 tests.
Verification
- 214/214 full regression tests passed locally in strict
ResourceWarningmode. - GitHub Actions passed on Python 3.9, 3.11, and 3.13 for both
mainand thev0.1.2tag. - A fresh-clone public-demo rehearsal passed, including discover, audit, preview, approval, apply, verify, check, and reset.
- Skill validation, compile, JSON, Markdown-link, image-dimension, and privacy checks passed.
No production Skill behavior changed from v0.1.1.