Skip to content

feat(security): SkillSpector severity gate script with reviewed acceptances #213

Closed

Description

@DevSecNinja

Part of #211 (rescues the scanner half of #125).

Problem

#125's Scan AI primitives job fails (run 35013183566). Upgrading to SkillSpector v2.13.4 and running the scan locally shows the cause is real report content, not infrastructure:

  • 16 HIGH findings, all false positives on review: Tool Guardian's deny-list regexes (rm -rf /, git push --force, …), test docs saying "no warning", os.environ.copy() passed to gh, and two external awesome-copilot agents (draw.io XML comment, a CWE example list).
  • 3 targets with incomplete inspection (bounded-parser limits on formula examples / code snippets).

#125's gate has no way for the owner to approve a reviewed finding, so it can never go green without weakening the policy.

Proposed solution

scripts/scan-skills.py + unit tests:

  • Same hardened container command, symlink/target validation and report contract as ci(security): add fail-closed SkillSpector scans #125 (fail closed on HIGH/CRITICAL, scanner errors, malformed reports, scope exclusions, suppressions).
  • Scan roots also cover .agents/**, .github/instructions/**, .github/hooks/**.
  • Owner-reviewed .github/skillspector-accepted.json: each accepted HIGH/CRITICAL finding (path + rule + severity, optionally exact evidence) or incomplete-inspection limitation (path + reason code) needs a written reason. Anything new, escalated or changed still blocks; stale entries fail the gate.
  • Handle v2.13.x complete_with_caveats and parser-diagnostic line numbers.
  • Write summary.json (gate result) for the reporting layer (separate PR).

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions