artifact-hygiene

Artifact Hygiene

Run the local-only proof of concept that checks whether repository content is safe to publish. The helper is the only component allowed to read candidate content. The active model receives normalized, redacted findings and coverage—not raw candidate content. On Claude Code this skill carries no model: pin on purpose: it rides the session model rather than downgrading to the standard alias, while Pi routes by model-tier.

This skill is advisory and read-only. It never fetches, follows links, calls remote services, validates credentials, installs hooks, blocks CI, edits files, changes Git state, rewrites history, or creates tracker or remote state.

Scope

The proof of concept scans:

• the full publishable working tree: tracked regular text files plus staged, unstaged, and untracked-not-ignored regular text files; and
• unpublished branch commit messages and per-path patches relative to a locally available default-branch ref, without fetching. If no remote-backed base is available, it scans all commits reachable from HEAD rather than treating the current local branch as published.

It uses an audit-owned Gitleaks configuration and empty ignore file, scrubs scanner configuration from the environment, ignores inline scanner allow-comments, and never passes a baseline. Repository scanner configuration cannot suppress the audit. Before any clean result, a private runtime canary must prove that the selected scanner and audit configuration can detect the pinned secret shape. The helper also reports known session-share links, Bead references, email addresses, name-like personal data (personal names), AI attribution trailers or boilerplate, and scanner-suppression controls without exposing matched values.

The custom-detectors coverage source proves every built-in non-secret detector still matches its private canary. Missing detector coverage produces a partial result, never a clean result. Personal-name detection is advisory and contextual: it detects direct requests to contact a named person or handle, not capitalized technical phrases. Expected Co-authored-by: and Signed-off-by: trailers are excluded from personal-data findings. Bead-reference, personal-data, and AI-attribution detectors scan unpublished history only—commit messages and added patch lines—rather than re-reporting content already published on the base branch. Bead references and personal-data matches in dependency lockfiles are ignored, as are email addresses under reserved example domains. Email addresses in the owner, created_by, and assignee fields of Beads .beads/issues.jsonl records are treated as structural attribution; unrelated fields remain reportable.