$ cd ~/nutrx-labs/tools/rtfd
rtfd will watch your pull requests, hand each one to the agent you choose, and post one review with inline comments. The agent reads and rtfd does the writing, so a hostile PR can't talk its way into your repo. It's in the lab: the design is settled, the build is in progress.
// what it does
rtfd fetches the PR and posts the review itself. The agent gets a checkout, the diff, CI state and open threads as files, and no provider token. The worst a malicious PR can get out of it is a bad review, never a bad write.
Git history comes through rtfd-git: four read commands with fixed command lines. No plugins, none of your MCP servers, and agent config in the PR is reset to the base branch first. rtfd validate lists every tool the agent ends up with.
The agent writes findings; rtfd decides the verdict. Any open Should means not ready, so a review can never say "ready" next to two Shoulds. Every comment is scanned for secrets before it goes out.
After the author pushes fixes, rtfd reviews only what changed, replies "Looks good now." on each fixed thread, resolves it, and leaves the rest alone.
Should, Could, MINOR. Short, and no bot marker. The tone lives in a per-repo prompt that rtfd re-reads on every run, so you edit the brief, not the code.
When the agent hits its plan's limit, rtfd pauses until the stated reset time, for every repo on that account, and gives the attempt back. No more three failed attempts in eleven minutes.
// how it works
One YAML file says which repos to watch, which agent reviews each, which prompt it follows and whose PRs count. The token stays in an environment variable; the config only names it. This is the draft config from the design and may change.
rtfd polls GitHub, waits for the pushes to settle and for CI to finish, then starts one review per new head. PRs from authors outside the allowlist, and from forks, are skipped.
It fetches the PR into its own mirror, so your clones aren't touched, and checks it out. Agent config files go back to the base branch's version. The diff, CI state and open threads land in a run folder for the agent.
Your agent runs read-only, in a scrubbed environment, and returns a findings document. Any CLI that can print JSON can fill the role: Claude Code first, Copilot CLI next.
It validates the findings, works out the verdict, posts a single review with inline comments and reads it back to confirm. One long-running process covers every repo, and rtfd status answers "is it watching?"
// good to know
// the full manual
Everything else is in the README.
Configuration, every command, troubleshooting and the design notes.
Read the README ↗