agentic-engineering / ai-code-review
AI Code Review
AI code review here does not mean "an assistant that leaves comments". It means a gate: an adversarial reviewer — from a different model family than the builder — whose verdict every PR must survive before a human merges it.
The verdict is machine-parseable, and a missing verdict counts as a rejection.
Why a different model family
The reviewer is never the builder's family. Same-family review inherits the same training biases and the same blind spots — it is self-review with extra steps. Cross-model review is the cheapest diversity you can buy: the two models fail differently, and their disagreements mark exactly the code a human should look at first.
The verdict protocol
VERDICT: APPROVED VERDICT: REVISE — <reason>
Reviewers must end with exactly one verdict line. Scripts parse it. Anything else — crash, timeout, empty output, malformed line — is treated as a rejection. The gate is fail-closed by construction, because the failure mode of a lenient gate is silent bad code reaching production.
What the reviewer hunts
- Correctness bugs in the diff — logic, boundaries, error paths.
- Security issues: injection, leaked secrets, unsafe handling of untrusted text.
- Broken error handling and missing tests for changed behavior.
- Scope violations: changes that reach outside what the PR claims to do.
- Self-review violations: did the builder approve or merge anything itself?
Style is explicitly out of scope. The gate catches what breaks things, not what a linter should.
The log is the truth
Every gate run records its lane, token usage, and duration in an append-only log. That log is the heartbeat of the whole rig — 967 entries in the audited September 2026 week — and it is what makes the workflow auditable after the fact: you can reconstruct not just what merged, but what every reviewer said about everything that did not.
Figures below come from one audited production week (September 21–28, 2026) across my own GitHub account — merged-PR counts, gate logs, and token accounting, published weekly on the Turbo Rig stats page. They are measurements of my own workflow, not industry averages.
Adopting it
The gate that runs my workflow is one portable bash script plus reviewer CLIs — no daemons, no services, no new platform to babysit. If you want one for your team, that is part of team onboarding and the stabilization engagement.
Talk it through with someone who runs this stack on his own systems every day.