Iterate Code Iteration

Problem

Code risks in real projects are rarely isolated typos. They are spread across correctness, security, performance, architecture, dependency boundaries, and test gaps. A single prompt-driven edit lacks global constraints, while manual review is hard to sustain across a multi-module repository.

How it works

Iterate turns review into a repeatable loop. It first builds or reads ITERATE.md to capture the tech stack, module boundaries, and validation commands, then runs parallel dimension-based review. Findings are split: atomic issues are fixed directly by the main flow, while architectural issues require user approval and are executed serially by sub-agents. Each round runs validation before continuing until zero findings or the rounds cap is reached. For read-only audits, review-only mode performs review, convergence checks, and report validation without modifying files.

Fit and limits

It suits refactor wrap-up, pre-release health checks, security hardening, and cross-file quality repair, but not one-off simple edits. If validation.commands is missing, the repair loop has fewer reliable signals. If the repo has protected paths, risk areas, or known exemptions, declare them in personalization first so the model does not bypass constraints.