Review finding triage is a step inserted between an AI code review and the fixes that follow it: instead of addressing every finding the reviewer raised, an agent first evaluates each finding critically — whether it really needs to be fixed, given the project’s requirements and existing code — fixes only the ones it accepts, and records why it declined the rest. It is described in AIによる実装の品質が微妙で毎回自分で指摘しまくる必要があったので、確認前に自動で品質を上げさせるようにした, whose author credits the idea to an article by another developer, kawarimidoll.
Usage
The problem it answers is that AI review output mixes useful findings with off-target and excessive ones. The post reports that having a subagent or Codex CLI review Claude Code’s code and then fix everything it raised sometimes left the code messier than before. With triage in place, the same review output is filtered before any change is made.
In the post’s setup the triage runs as a /fix-review-comments skill after three reviewer agents with
different focuses have reviewed the change in parallel. Findings that several reviewers raised on the same
spot are read as higher priority. In the example shown, ten findings came in; six were addressed (as five fixes, one of them covering a
point two reviewers shared) and four were declined with reasons that refer to the project’s context — for instance that the behaviour
was intentionally specified in the requirements, or that the point was already covered by existing text.
When It Applies
- It applies where an agent both reviews and fixes code without a human in between, so that whatever the review raises would otherwise be acted on automatically.
- It assumes the triaging agent has enough context — requirements, the surrounding code — to tell an appropriate finding from an inappropriate one; the declined items in the example are justified by reference to exactly that context.
- The recorded reasons for declined findings are what make the result checkable: the author presents this selection as the reason he can leave the fixes to run automatically, and still reads the code himself afterwards.
- How well established it is: the source is one developer’s account of his own configuration, adapted from another developer’s idea, with an illustrative session rather than any measurement.
Related Terms
- Critical Dialogue Review — a two-model review loop in which the implementing agent accepts, rejects or defers each finding with a recorded reason
- Agentic Code Review
- Code Review Agent
- LLM-as-a-Judge