Four projects verify AI code before the pull request

Four separate projects released within 48 hours, and none of them review a pull request.

All four assume the same thing: an agent produces more diff than a human will read, so checking has to happen while the code gets written, not after.

GitWarren checks the dirty worktree before a single line gets committed. It ships 17 MCP tools so the reviewing agent can answer questions about its own diff, without a human opening a PR first.

Twinrun makes a related bet with less machinery. It runs the old and new versions of a changed function on identical inputs and diffs the outputs. That needs no test assertions, just a callable and its predecessor.

Ponytail goes further upstream. It applies a stop rule before generation happens at all, rather than checking the code after it exists.

webmcp-stack takes a fourth approach. It generates WebMCP tools straight from an OpenAPI spec and bakes confirmation prompts into the regenerated code regions. The check lives inside the codegen step, not a separate review pass.

Four different mechanisms, and none of them touch a pull request. If review tooling built around diffs and PR comments is solving a problem that already moved upstream, the next question is adoption. All four are new enough that nobody outside their own repo has picked them up yet.