​

AI Skill Governance — Approval & Version History

Audience: reviewers who approve org-authored AI skills. Spec: specs/pf/specs/PF-120-org-authored-ai-skill-governance.md.

Governance covers the lifecycle of the org-authored and org-customized AI skills the Create Custom Skill wizard produces: an approval queue where a reviewer signs off before a skill can drive AI, and an immutable version history with one-click rollback. It is the reviewer/approver companion to the authoring guide.

​

Approval queue

Reach the queue at /settings/ai-skills/approvals (heading AI Skill Approvals), guarded by the pf.ai_skills.approve permission. Regulated categories (clinical, compliance, HR, housing) require second-person approval — a pf.ai_skills.approve holder who is not the submitter — so a skill cannot drive AI until someone else signs off. Non-regulated skills self-activate at submission; imported skills route to review with a mandatory sensitivity class and a no-PHI attestation.

​

Version history and rollback

Every create and every prompt/config edit inserts an immutable snapshot into pf_ai_skill_versions (insert-only — no update or delete). Open the version-history drawer from a skill’s Configure dialog on /settings/ai-skills. Rollback writes a new forward version capturing the rolled-back state — history is never rewound — and a rolled-back regulated skill re-enters review and needs fresh approval.

​

What gets audited

Every lifecycle transition emits a platform audit entry — pf.ai_skill.submitted, pf.ai_skill.approved, pf.ai_skill.rejected, pf.ai_skill.rolled_back — with actor, org, skill id, and from/to state. Write access gates on pf.ai_skills.manage and approve/reject on pf.ai_skills.approve, so the UI controls and the row-level security agree.

​

See also