Skill 4 of 6 · Verify
deep-code-review
/deep-code-review
Fresh agents trace every promised behavior end to end - and the findings that reach you are the ones that survived an adversary trying to refute them.
What it does for you
Built to catch what looks finished but isn't
Works on any branch, PR, or set of files - with or without a plan behind it.
Hunts the last mile
Optimistic UI with no real call behind it, leftover mocks, swallowed errors, unregistered routes - every promised behavior is traced hop by hop through the full chain.
Evidence or it doesn't ship
"Save might not work" doesn't qualify as a finding. "The chain breaks at hop 4 - this handler calls a route that's never registered" does. No cited chain, no finding.
Fresh eyes on your diff
Reviewers never see the implementation conversation - an agent that watched the code get written rationalizes the same gaps. Independence is what makes the catch.
Scales with the stakes
One thorough agent for everyday diffs. For risky ones, a multi-agent fleet of specialized finders whose candidates are then adversarially verified. A maximum mode for release gates.
The standard
Evidence-gated, adversarially verified findings
In multi-agent mode, every significant candidate finding is handed to a fresh agent whose job is to prove it wrong. Only the findings that survive that attack reach your report.
The review never touches your code. It produces an evidence-backed report and stops. Deciding what to fix, defer, or reject is a separate, explicit step you control - so an unattended review can run as a gate without side effects.
Where it fits
The independent measurement point
Implementation says "done." deep-code-review checks whether the code agrees.