config.yaml

# Scanner configuration. This is the only user-facing knob in the repo.
# Forking the scanner for your own catalogue is a single edit here plus
# enabling Actions and GitHub Pages on your fork.

# Schema version of THIS config file. Bumped when the shape changes.
config_version: 1

catalogue:
  # Skills are declared by per-namespace README.md files under
  # registry/<ns>/skills/ in the catalogue repo. Each README's
  # frontmatter lists sources[].repo plus per-skill overrides. This is
  # the canonical declaration; the in-tree .agents/skills/ format is
  # supported in scanner/enumerate.py for forks that need it but is
  # not enabled here because coder/registry duplicates the same
  # upstream skills across both layouts under different slugs.
  registry_repo:
    owner: coder
    repo: registry
    ref: main
  formats:
    external_sources:
      enabled: true
      # Path glob, relative to the cloned catalogue. Each matched README
      # has its frontmatter parsed for sources[].repo plus per-skill
      # overrides keyed by slug.
      readme_glob: registry/*/skills/README.md

scanners:
  skillspector:
    enabled: true
    # Pin by commit SHA. Bumped manually via PR; the workflow's pip
    # install reads this same pin so contributors do not have to keep
    # the version in two places. Dependabot does not parse this file,
    # so a bumper bot lives outside the loop until the upstream
    # publishes to PyPI and the pin can move into pyproject.toml.
    pin: "skillspector @ git+https://github.com/NVIDIA/SkillSpector.git@7bc9c0feb663375ced6e7436008d082e6e3486ea"
    # Empty so .github/workflows/scan.yaml can append --no-llm
    # dynamically based on whether the LLM credential secret is set.
    flags: []
    llm:
      provider: anthropic
      model: "claude-sonnet-4-6"

# Per-skill verdict policy. v1 has one input (SkillSpector risk_score).
# When more scanners join the pipeline we add new threshold fields here
# and a matching branch in scanner/verdict.py.
#
# Thresholds are aligned to SkillSpector's own severity bands so that the
# UI verdict matches the recommendation field SkillSpector returns:
#
#   0-20   LOW       SAFE             -> verdict: clean
#   21-50  MEDIUM    CAUTION          -> verdict: clean (advisory; see below)
#   51-80  HIGH      DO_NOT_INSTALL   -> verdict: suspicious
#   81-100 CRITICAL  DO_NOT_INSTALL   -> verdict: malicious
#
# Rationale: SkillSpector's static-analysis layer is loud on real
# catalogues (the ClawHub paper measured a ~49% positive rate on 67k
# skills) and is advisory rather than authoritative, so we only
# escalate above its HIGH cutoff. CAUTION-band findings still appear
# on the per-skill page so reviewers can see them; we just do not
# flag the skill as suspicious at the catalogue level.
verdict:
  malicious_risk_score: 81
  suspicious_risk_score: 51

publish:
  # One GitHub Release per scan. The rolling `latest` tag is the floor
  # the registry-server proxy falls back to when Pages is degraded.
  release:
    tag_format: "scan-{utc:%Y-%m-%dT%H-%MZ}"
    rolling_tag: latest
    retention_days: 90
  # Public CDN-cached JSON via actions/deploy-pages. The site reads
  # latest.json through a small proxy in coder/registry-server that
  # caches for 5 minutes and lets Cloudflare hold stale for 10 more.
  pages:
    enabled: true
    files:
      latest: latest.json
      history_dir: history
      history_retention_days: 30

# The schedule fires at :17 past the hour, every 6 hours, to dodge the
# top-of-hour cron-drift cluster on shared GitHub-hosted runners.
schedule:
  cron: "17 */6 * * *"