Nov 9, 2026
Automating the Dependency Review Nobody Had Time For
Our engineers were shipping more code than reviewers could audit. I built a Go control plane that checks every pull request's dependencies, secrets, and policy before a human looks at it.
The problem
At ZimWorx, the engineering organisation was producing code faster than it could be reviewed. More repositories and more pull requests all landed on the same few reviewers.
The part of review that suffered most was the part nobody enjoys: auditing dependencies. Is this package maintained? Is that transitive dependency in a fresh advisory? Did someone commit a key? Doing that by hand on every pull request across 50+ production repositories wasn't happening, and the gap was growing.
I wanted the machine to answer those questions so reviewers could spend their time on design and correctness.
What I built
Sentinel is a governance and security control plane for the whole GitHub organisation. It doesn't replace good scanners. It orchestrates them, correlates their findings, and enforces policy.
- Dependency intelligence. Builds a software bill of materials (CycloneDX or SPDX) for each repository, including transitive dependencies, and scores packages for maintenance risk and typosquatting.
- Live threat intelligence. Pulls advisories from GHSA, OSV.dev, NVD, and CISA's Known Exploited Vulnerabilities list, and matches them against every stored SBOM. A new advisory flags every affected repository, not just the next pull request.
- Secrets detection. Gitleaks catches credentials before they reach production.
- Policy as code. Rules are written in Rego and evaluated by Open Policy Agent, so "block critical vulnerabilities" or "no lockfile drift" is a reviewed policy file, not tribal knowledge.
- Pull request gates. A GitHub App and a reusable GitHub Action run a scan on every pull request and fail the check when policy returns
block. Overrides go through an authenticated API so exceptions are deliberate and recorded. - Health dashboard. Vulnerability age, per-repository health grades, and risk trends across the organisation.
It's written in Go, with PostgreSQL (SBOM graphs in JSONB), Redis and NATS for work distribution, and Trivy for scanning.
Design targets
- Under 10 seconds from webhook to a worker picking up the scan.
- Lightweight pull request checks (secrets, lockfiles, policy) in under 2 minutes.
- Full dependency-tree scans with SBOM generation in under 15 minutes.
- Horizontal scaling to 1,000+ repositories.
Results
Sentinel was onboarded to 45 repositories and uncovered more than 5,000 vulnerabilities across them: known CVEs, outdated libraries, inefficient code, exposed secrets, and more.
What I learned
Security tooling fails when it adds review work. Sentinel works because it removes it. The scan happens before the reviewer arrives, the policy explains why a check failed, and the human only looks at what genuinely needs judgment.