secrets-scan

Quick reference

When to use

• Pre-push audit — run before opening a PR or pushing to a remote branch
• Open-source prep — before making a private repo public, scan the full history surface
• CI gate — catch secrets that slipped past .gitignore rules
• Post-incident — verify no additional credentials are exposed after a breach
• Onboarding — audit a legacy codebase you’ve just inherited

Examples

# Scan only — full repo, severity-ranked report
/brewtools:secrets-scan

# Scan and remediate interactively (CRITICAL + HIGH findings prompted)
/brewtools:secrets-scan --fix

# Free-form intent after the command (the skill is user-invoked only)
/brewtools:secrets-scan check for secrets before I open-source this repo

Expected output after scan:

## Secrets Scan Complete — FINDINGS

| Metric    | Value |
|-----------|-------|
| Files     | 342   |
| Unscanned | 0     |
| CRITICAL  | 1     |
| HIGH      | 3     |
| MEDIUM    | 7     |
| LOW       | 12    |

Report: .claude/reports/20260417-143022_secrets-scan/report.md (mode 600)

Verdict is one of CLEAN, FINDINGS, or INCOMPLETE — never CLEAN while any chunk failed reconciliation.

Flow

Delegation

Each chunk agent receives the same six-field brief, reused verbatim per chunk:

A bare one-line task is never enough. The failure mode is concrete: a malformed or chatty response used to silently remove an entire chunk from the merged report, and a missed CRITICAL there is a credential shipped to a remote. Reconciliation is what closes that gap — the chunk either gets accounted for or the whole scan is marked INCOMPLETE.

The CONTEXT line has to be true. The brief tells each agent the list is already filtered and instructs it not to skip anything — so when Phase 1 handed over an unfiltered git ls-files, ten agents spent their chunk budget reading .png, .jar, lockfiles and vendored code they had been told not to skip. Phase 1 now filters for real.

Redaction is mandatory, not optional. Agents never quote, echo, or summarise a matched value — for every hit they run the skill’s redaction script and copy back only a length, a sha256[:12] fingerprint, and a preview of at most 4 leading characters. The raw secret never enters an agent’s response, the report, or this chat.