RFC: Multi-agent parallel review approach inspired by Claude Code's code-review plugin

[tfriedel]

Summary

Looking for opinions on integrating the multi-agent parallel review approach used by Claude Code's official code-review plugin.

This plugin is recommended by Boris Cherny, Creator & Head of Claude Code at Anthropic.

How It Works

Instead of a single agent reviewing everything, it runs 5 parallel agents with different review perspectives (bugs, CLAUDE.md compliance, git history context, etc.), then scores each finding for confidence and filters out likely false positives (keeping only ≥80% confidence).

Key Considerations

1. Agent portability: This relies on Claude Code's Task tool for spawning subagents. Other agents (codex, gemini) can't do this. Switching agents would give different review behavior.

2. Make prompts configurable?: Currently prompts are hardcoded. We could instead let users configure prompts per-agent, allowing them to bring their own multi-agent prompt for Claude.

3. Cost: Running 5+ agents per commit could be wasteful for routine small commits. Maybe reserve this for manual roborev review or a --thorough flag, not post-commit hooks.

Questions

• Is the reduced false-positive rate worth the added complexity?
• Should prompts be configurable per-agent instead of hardcoding this?
• How to handle the cost of multi-agent reviews on every commit?

Thoughts?

[wesm]

Why not, right? I think support custom prompts and opting in to this multi-agent/parallel review can all live in ~/.roborev/config.toml. Do you want to take a crack at writing up a more detailed spec that one of us can use to drive the implementation? i.e.

• what config.toml additions to add
• how to manage custom prompts (think through what the docs for this will look like on https://www.roborev.io/)
• what the results of the review will look like in the TUI (I assume just the output of Claude's "synthesis" that combines the different reviewer opinions)

This makes sense to me, in addition to having support for multiple adversarial agents (I've been switching back and forth between codex and gemini, but ideally I want both of their opinions on stuff that Claude is doing).

I also want to add reviewer modalities that focus on code simplification and refactoring, with the ability to target individual files, groups of files, or a whole folder for scrutiny