Local AI code review
Use one agent to review code and a second to challenge it.
A single AI review can sound confident while missing evidence. HAICHI is built for an inspectable Reviewer + Verifier workflow: one agent finds risks, the second challenges each claim, and you decide what is actually worth fixing.
Why a verifier pass matters
AI code review is most useful when it stays concrete. The Reviewer should point to failure modes, missing tests, edge cases, and risky assumptions. The Verifier should reject vague findings and keep only claims supported by the change.
- Run a first review with a focused local agent role.
- Pass the result to a second agent that looks for unsupported or low-value claims.
- Retry only the weak step instead of starting the whole discussion over.
- Use workspace-scoped permissions only when you explicitly enable them.
Starter prompt
Review this change for failure modes and missing tests. Reviewer: list concrete risks. Verifier: challenge each risk, discard weak claims, and return the three fixes worth doing first.
That small loop is enough to test whether HAICHI helps your local AI workflow before you try larger agent chains.