Add reusable reviewing-openmc-code skill and Copilot Review agent

[paulromano]

Description

This PR introduces AI-assisted code review tooling so that the review guidelines live in a single, tool-agnostic skill that can be used by Copilot, Claude Code, Codex, and similar agents. It makes the following updates:

• .claude/skills/openmc-code-review/SKILL.md: A reusable skill containing the code review procedure. Supports both reviewer mode (PR/branch review) and self-review mode (author preparing changes before submission). Note that I've chosen to put this in the .claude directory because it is recognized by both Claude Code and Copilot.
• .github/agents/Review.agent.md: A thin Github Copilot custom-agent wrapper that delegates to the openmc-code-review skill, providing a named "Review" persona in Copilot's agent dropdown.
• .github/copilot-instructions.md: Instructs Copilot to use the openmc-code-review skill when performing automated code reviews triggered directly on the GitHub repository.
• AGENTS.md (updated): Replaced the inline code-review guidance with a pointer to the openmc-code-review skill, avoiding duplication.

Checklist

• I have performed a self-review of my own code
• I have run clang-format (version 15) on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (if applicable)
• I have made corresponding changes to the documentation (if applicable)
• I have added tests that prove my fix is effective or that my feature works (if applicable)

[jtramm]

This is great @paulromano!!! I had a lot of fun reviewing this and experimenting with making custom skills. I'm guessing this will be the first of many agentic skills for OpenMC in the near future.

I made a few suggestions for you to consider, but overall it seems like a great starting point for us.

For posterity, I've been benchmarking the claude code agentic review on my upcoming shared secondary branch draft PR (https://github.com/jtramm/openmc/tree/another_shared_secondary_branch). There's at least 10 legitimate logic bugs in there (not style/maintainability issues), but the default claude code review and this skill only find about 1 in 10 of the bugs on any given run. So, my own personal baseline expectation is that the agent will likely flag super obvious issues, but may only find about 10% of other more complex bugs. The default claude code /review skill doesn't do any better.

One trick I found is that due to random seeding, for complex PRs there is actually a huge divergence in the results you get (provided you run the review from clean/fresh contexts that aren't polluted from a previous review tool). As such, I've found that if I run multiple fully independent agentic reviews I'm able to expose way more issues than just running one. There's a paper here that shows this effect: https://arxiv.org/html/2509.01494v1 where you see significant improvements in PR reviewing if you collate 3-5 independent reviews, with greatly diminished returns going up to 10 or 20 reviews.

To that end, I may follow this up with a multi-agent skill that launches multiple such reviews in parallel so as to potentially uncover more bugs (as empirically this seems to be helpful, as discussed in https://arxiv.org/html/2509.01494v1, and in my findings as well). Sounds complex but claude already has an agent tool, so it's not too hard to get it to spawn a bunch of subagents to do parallel reviews (though it is more expensive!)