Automated PR Review: Exploring Always-On Review for Unbound Force

[sonupreetam]

Following up on the feedback loop gap identified in Discussion #158, I'd like to start a focused conversation about adding automated PR review to the meta-repo.

The Problem

Our review tooling (/review-pr, /review-council, Divisor agents) requires someone to be in an OpenCode session and manually run a command. This means:

• PRs can sit without any review pass until someone remembers to initiate one
• There's no consistent data trail to measure what reviews catch vs what they miss
• The feedback loop depends entirely on human initiation

The Idea

Every PR gets an automatic first-pass analysis -- no human initiation required. This would not replace human review or affect merge eligibility. PRs still need our approval. The analysis just surfaces insights so reviewers can focus on architecture and intent rather than catching routine issues.

Key constraints (from Slack feedback):

• Posts under a bot identity, never a human's name
• Informational only -- COMMENT event, never APPROVE or REQUEST_CHANGES
• One signal alongside deterministic CI checks, not a gating verdict

What I'd Like to Explore

There are open source tools in this space doing automated PR review and security analysis. Before we build anything or commit to an approach, I think we should:

1. Survey what's out there -- What open source review bots and GitHub Apps exist? What do they do well? What's missing?
2. Identify the gap -- Does anything in the ecosystem already combine multi-persona review, convention-pack awareness, and structured verdict artifacts the way the Divisor does? Or is there a tool we could just install?
3. Decide build vs adopt vs hybrid -- Should we adopt an existing tool, build on top of one, or extract our own from the Divisor logic?
4. Scope it -- Meta-repo only to start? What would a minimal pilot look like?

Why Now

This connects to several threads:

• The measurement baseline gap from @trevor-vaughan's harness analysis -- automated review on every PR creates the data stream needed to start measuring effectiveness
• The "advisory vs deterministic" gap -- an always-on review closes the "nobody ran the command" hole
• Existing OpenSpec proposals around /review-pr improvements (review-pr-enhancements, review-pr-state-awareness, review-pr-reliability) that are already evolving the review logic

Would love to hear thoughts -- especially if anyone has experience with existing tools in this space or opinions on build vs adopt.