Concurrent Sub-Agent Review with different personas

[fafg]

I was thinking how repos (to be reviewed) could have a .ocr/rules.json and in this file we could define agents to run concurrently against the PR. one context, multiple personas reviewing with its own rules, and strategies.

Personas:
1 - Security Reviewer
2 - Code Quality
3 - Performance
4 - Documentation

configured models per agent, with fall back model...

idea came from: https://blog.cloudflare.com/ai-code-review/

[stay-foolish-forever]

Thanks for the proposal, and for the Cloudflare link — good read.

We're not planning a multi-persona / multi-agent review mode in the short term. Running several personas over the same context multiplies token cost, and the outputs overlap enough that we'd need a dedup + merge layer before they're usable as inline comments.

The good news is that the existing review rules already cover most of this. Instead of personas as agents, write each concern as its own rule: your security checklist in one rule, performance in another, documentation in another. The path glob on each rule then decides which files that concern applies to — so you can scope the security rule to **/{auth,crypto}/** and .github/workflows/**, the performance rule to your hot paths, the docs rule to **/*.md, and so on. Rules can be defined at the global or project level, and the rule content can be either inline text or a path to a Markdown checklist, so the persona prompts stay reviewable in-repo.

Worth noting that rules are evaluated in declaration order and the first matching path wins, so today multiple user rules cannot apply to the same file — concerns are partitioned by glob rather than stacked. (merge_system_rule: true only merges OCR's built-in language rule with yours, and ocr rules check <file> shows which rule a given file resolves to.) So for genuinely overlapping personas on the same file, the way to go is separate passes — ocr review --rule ./security-only.json, one per persona, optionally with a different --model each. If you're invoking ocr from a coding agent, just ask the agent to automate that: it can fan out one pass per persona and merge the results, which is a natural fit for its sub-agent capabilities.

See the docs for details: https://open-codereview.ai/docs/review-rules

If you hit something rules genuinely can't express, follow up here — that's the kind of gap that would change our mind.