Run multi-stage PR reviews with parallel sub-agents
Multi-stage code review for Claude Code — parallel sub-agent detection, validation passes, persistent JSON state, and an automated fix loop that re-reviews and reverts regressions before committing. Task: The author built a six-command pipeline for Claude Code using parallel sub-agents for multi-lens detection, validation gates, and an automated fix loop that reverts regressions. The workflow includes commands for review, adding external findings, interactive walkthroughs, and automated fixes with human override options. The author uses this on their own PRs, integrating human judgment via the walkthrough and promote commands to handle uncertain findings.; tools: Claude Code, Codex. Human role: The original post does not say how a person stayed involved.; limitations: Requires a Claude Code subscription (Max recommended) and specific runtime dependencies like uv, jq, and gh. Full Disk Access is recommended for accurate token tallying. The workflow's effectiveness is currently supported only by the author's personal experience, lacking independent validation..
Tools
Published on Agent Work Atlas:
Overview
Multi-stage code review for Claude Code — parallel sub-agent detection, validation passes, persistent JSON state, and an automated fix loop that re-reviews and reverts regressions before committing.
Workflow
The author built a six-command pipeline for Claude Code using parallel sub-agents for multi-lens detection, validation gates, and an automated fix loop that reverts regressions. The workflow includes commands for review, adding external findings, interactive walkthroughs, and automated fixes with human override options. The author uses this on their own PRs, integrating human judgment via the walkthrough and promote commands to handle uncertain findings.
Reported result
The author reports an anecdotal improvement in catching real bugs and reducing false positives compared to built-in tools and other reviewers like CodeRabbit. No measured metrics or broader success data are provided. A measurable acceptance check would be to track bug detection rates and false positive counts against a baseline over a set number of PRs.
Human role
The original post does not say how a person stayed involved.
Limitations
Requires a Claude Code subscription (Max recommended) and specific runtime dependencies like uv, jq, and gh. Full Disk Access is recommended for accurate token tallying. The workflow's effectiveness is currently supported only by the author's personal experience, lacking independent validation.
Try it
I want to implement the adamsreview multi-stage code review workflow for my project. First, install the plugin using /plugin marketplace add adamjgmiller/adamsreview and /plugin install adamsreview@adamsreview. Ensure dependencies like uv, jq, and gh are installed. For a sample PR [replaceable PR URL], run /adamsreview:review to generate the initial findings. Then, use /adamsreview:walkthrough to interactively review uncertain findings, promoting those you want auto-fixed. Finally, run /adamsreview:fix to apply the fixes. Please show me the output of each step and confirm before proceeding to the next. Do not commit any changes until I explicitly approve the final diff.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.