GuardDOG
Hybrid repo scanner for the AI-generated-code era — local layers for free, LLM only where it pays.
- Period
- Oct 2026
- Role
- Design & implementation
- Commits
- 13
- Key metric
- 17 regex rules · ~15 AST checks · exit-code CI gate
- Links
- Repo ↗
Problem
AI-generated code looks clean: correct syntax, plausible names, green tests. Underneath — a hardcoded password shipped to prod, print(entered_pass) in a log path, verify=False on a payment request. Linters check style, not semantics. An LLM can see it — but streaming a whole repo into a model on every commit is slow and burns tokens.
Architecture
GuardDOG sits in between. A local layer — 17 language-agnostic regex rules plus ~15 Python AST checks — catches the obvious instantly and for free. The LLM layer audits only files whose SHA256 changed since the last run, and only when an API key is present. No key → an honest skipped in the report; the local layer still runs. No fake-green, ever.
Key decisions
- Findings land in the exit code — a CRITICAL hit fails the build or blocks the commit. It’s a gate, not a suggestion.
- SHA256 change-detection cache — the expensive layer runs only on what changed.
- Zero runtime dependencies — drops into any repo: one command, a
run_all.bat, a pre-commit hook, or a CI stage. - Reports in md/json/html — human review and machine gates read the same truth.
- 26 tests covering the rule layers — the scanner itself is tested like production code.
What it proves
The thesis behind how I work with AI, made into a tool: use the cheap deterministic check everywhere, use the expensive probabilistic check exactly where it earns its tokens — and wire both into the pipeline so it’s enforced, not aspirational.