Things I've built, mostly so I'd want to use them.

reviewgraph routes a GitHub PR through staged reviewer agents and treats the PR conversation as shared review memory. Reviewers get introduced under explicit graph conditions: size and risk routing, path triggers, diff patterns, each with a recorded trigger reason and a scoped context so it only sees what it needs.

Every GitHub write sits behind item-level human approval. The CLI is dry-run by default and emits markdown plus JSON; GitHub PR references only run against deterministic fake transport data, and live LLM calls are an explicit opt-in. It’s an implementation prototype and harness suite rather than a hosted service, with fixture PRs keeping runs deterministic.