grumpy.dev 发布 PR 合并门禁：用 LLM 出题检验作者是否理解改动

grumpy - the AI that vibe checks your PRs

To put it plainly: Grumpy is a merge gate that checks the author of the PR understands the change they are putting in.

grumpy sits in your PR pipeline as a GitHub Action. Before a PR can merge, grumpy walks whoever opened it through their own diff one section at a time, asking a question about each — the first always "How does this change work on a high level?" — and blocks the merge until they answer enough of them well enough to convince an LLM grader.

Why should you care?

AI is here, and more PRs are AI-generated or AI-assisted. People are literally taking Jira tickets, shoving them into Claude, and reviewers have to deal with the 1K+ line PRs.

The result? More burnout. People just say: "LGTM, CI is green" and now you are shipping terrible software! For business critical software, this is unacceptable.

Features (so far)

An exam sheet for every PR. grumpy reads the diff once to form its own blind interpretation, then writes a short sheet: the fixed high-level question plus scoped questions, each anchored to a specific part of the diff. The author sees one question per screen, next to just the slice of the diff it is about.
Sized to the change. One scoped question per 15 changed lines, up to EXAM_QUESTION_COUNT. A one-line PR gets one question, a 1,000-line monstrosity gets picked apart section by section, and nobody is quizzed on git headers.
Graded on submit, leniently. Claude marks each answer against the blind interpretation. A terse but correct answer passes; contradicting the diff fails. Wrong answers can be retried, then grumpy reveals the answer and moves on. "Explain it for me" gives a plain-language walkthrough of a section with no effect on the mark, and "Skip" gives up on a question.
No understanding? No merge. The session passes once PASSINGMARKS screens are correct and fails as soon as that is out of reach. grumpy reports it as a grumpy/verdict commit status; require it in branch protection and the PR cannot merge until the author passes.
Good development practices check. Oversized PRs and diffs are rejected before any question is asked. More coming soon.
Local first. Self-hosted only: one Docker Compose command and it's running, and it works in enterprise environments. BYOK of course.
No GitHub App, no OAuth, no webhooks. It's a FastAPI service that a GitHub Action talks to over a bearer token. Nothing to install on GitHub beyond a workflow file. Just keep the grumpy token a secret.
Sarcastic mode. For the thick skinned, grumpy will brutally tear your understanding. I used to have a tech lead who did this. I miss them.

Need a demo?

See it in action on real PRs:

Demo 1: Pass a PR: answer well and grumpy/verdict goes green.
Demo 2: Fail a PR: answer badly and the merge stays blocked.

Contributing

I am well aware of the irony of having vibe slop to deal with the problem of vibe slop.

I encourage AI generated PRs over human ones. Why? Because the entire point is to help devs integrate vibe coding more safely.

I still think humans > AI any day though.

Documentation

Full docs live at https://bhurghundii.com/grumpy.dev/ (source in docs/):

Quickstart: self-host it with Docker Compose
How it works
Wire it into a repo and block PRs until they pass
Configuration
Deploying to Railway and the pre-launch security checklist
Local development and limitations

Quick look

git clone https://github.com/bhurghundii/grumpy.dev.git
cd grumpy.dev
cp .env.example .env
# edit .env: set GRUMPY_TOKEN to a real secret and GRUMPY_BASE_URL
docker compose up --build
curl localhost:8000/healthz
# {"status": "ok"}

License

MIT