Releases: junjunjunbong/agent-skill-hub
Releases · junjunjunbong/agent-skill-hub
Release list
v0.3.0 — first public release
First publishable release: the read-only Agent Skill inventory with Phase 6 security hardening.
Install
npx agent-skill-hubRequires Node.js ≥ 22. The server binds to 127.0.0.1 only and opens your browser with a one-time session link.
Security (Phase 6)
- Localhost session boundary: single-use bootstrap token in the URL fragment, exchanged once for an
HttpOnly; SameSite=Strictin-memory session cookie. Sessions die with the process. - Exact-origin allowlist — other localhost ports are rejected; CORS stays disabled.
- CSRF defense on every state-changing request: per-session
X-Agent-Skill-Hub-CSRFheader (constant-time compare) + JSON content type. - Approved extra roots (
/api/v1/approved-roots+ Settings UI) with breadth limits: no filesystem roots, no home ancestors, no overlaps. Symlinks out of the home resolve only after explicit approval. - Hardened read boundary: every SKILL.md read re-validates lexical path AND live realpath against the live boundary stores — symlink swaps (TOCTOU) and removed boundaries are refused immediately.
- Log masking (home →
~, secrets redacted) and opaque error bodies. pnpm audit --prod --audit-level high: zero findings at release.
Packaging & CI
- 3-OS CI matrix (Ubuntu/macOS/Windows): verify + packaged-tarball smoke, all green.
- npm tarball ships
dist+ license/notice documents only;npx agent-skill-hubverified from a clean temp project against a fresh temp home.
Fixed
- Prompt Center no longer enters an unbounded re-render loop while loading.
Product scope up to 0.3.0 (Phases 0–5)
59-platform registry, custom platforms, shared-root deduplication by real path, symlink-safe scan engine, snapshot history + diff, H001–H015 health rules, full inventory UI, and the deterministic read-only Prompt Center with metadata-only optional history.
See CHANGELOG.md and SECURITY.md.