kylecui/petfish.ai/fish-guard

skill-trust-governance

Description

skill-trust-governance is a governance wrapper skill for the external trustskills engine.

Use it to evaluate a skill pack's executable behavior rather than trusting prose alone. It supports single-skill scans, multi-skill root scans, manifest generation, manifest verification, redline checks, and governance-level reporting.

Primary entrypoint:

uv run .opencode/skills/skill-trust-governance/scripts/trust_scan.py --path <skill-dir>

Triggers/Activation

Trigger this skill when the user asks for any of the following or clearly equivalent intent:

• skill trust
• skill安全
• 治理
• 可信度
• trust scan
• governance
• risk score
• redline check

Also trigger when the user wants to:

• scan one skill before publish
• batch-scan a skills root
• generate trust manifests
• verify existing manifests
• apply a custom governance policy YAML

Workflow

1. Decide scope: single skill with --path, or multi-skill root with --root.
2. Decide mode: default scan, --manifest for manifest generation, or --verify for manifest verification.
3. Run scripts/trust_scan.py instead of calling the external CLI ad hoc.
4. If a custom policy is provided, pass it with --policy.
5. Use --detail for expanded output and --output when a Markdown report is needed.
6. Prefer --json when the result will be consumed by quality-gate, CI, or other automation.
7. Report the governance level, overall risk score, triggered redlines, and any required follow-up action.

Must Do

• Use the bundled scripts/trust_scan.py wrapper as the default interface.
• Preserve trustskills as an external dependency; do not vendor or inline the engine.
• Surface the five governance levels clearly: allow, allow_with_ask, sandbox_required, manual_review_required, deny.
• Call out redline violations explicitly because they are hard DENY gates.
• If trustskills is missing, return install guidance: uv add git+https://github.com/kylecui/trustskills.git or uv pip install git+https://github.com/kylecui/trustskills.git.
• Use machine-readable JSON output when integrating with quality-gate or other automated checks.

Must Not Do

• Do not copy trustskills source code into the project.
• Do not import trustskills directly inside the wrapper; invoke it through subprocess only.
• Do not hide redline violations behind a summary-only response.
• Do not hardcode absolute paths.
• Do not create commands or agents for this pack.