FAQ¶

Do I need Python installed?¶

Python 3.11 or newer. grip is a Python package (grip-hook on PyPI). uv tool install grip-hook fetches a suitable Python by itself if you do not have one, and the pre-commit framework route manages its own environment.

How do I upgrade or uninstall?¶

pipx upgrade grip-hook, or uv tool upgrade grip-hook. Uninstalling the package leaves the git hooks, .grip.toml and the agent rule files behind; Removing grip takes each of them out in order.

Can I bypass it?¶

Yes. GRIP_SKIP=1 git push skips once, and git's own --no-verify skips every hook. grip is a forcing function for honest developers, not a security control.

Does it work with GUI git clients and IDEs?¶

If the client runs hooks without a terminal, grip cannot ask questions. By default it prints a notice and lets the operation through; set require_tty = true to block instead. Most IDE terminals work fine.

GUI clients often run hooks with a shorter PATH than your shell. The hook does not depend on it: it runs grip by the absolute path recorded at grip install. If you later move or reinstall grip, run grip install again; grip status shows grip missing until you do.

Does it run in CI?¶

The hooks do not: when CI is set they exit immediately, because nobody is there to answer. The GitHub Action does something else in CI: it posts the five questions on the pull request for the author or a reviewer to answer in review, and grades nothing.

Why do I keep getting the same questions?¶

Because the diff is the same. The questions written for a diff are reused for cache_hours (24 by default) when that exact diff is quizzed again, so a retry after a failed quiz is graded against the same five questions and costs one model call instead of two. The feedback says where to look again, not what the answer was. Change the diff, set retry_questions = "new" for new questions after a failed attempt, pass --fresh for one run, or set cache_hours = 0 to get new questions every time.

A diff that passed is remembered for remember_passes_hours, so an identical diff is not quizzed again at all. grip forget clears that memory.

Why did a failing provider block my commit?¶

Because a silently bypassed hook is worse than a loud one. Set fail_open = true if you prefer availability over strictness.

How much does it cost, and how long does it take?¶

Two requests per quiz: one writes the questions (or draws the whiteboard), one grades the answers. A retry on the same diff reuses the questions, so it costs only the grading. Input is dominated by the diff, sent with each request, plus a few hundred tokens of prompt; output is a few hundred tokens.

grip measures it instead of guessing. After every graded quiz it prints one line:

took 9.8 s · 3,140 tokens

grip last repeats it with the time of each request, and grip status shows the average over the last 20 quizzes. The Anthropic and OpenAI-compatible APIs report token counts; the coding-agent CLIs (claude-code, codex, gemini) do not, so for them grip shows the time only. grip last --json lists every request with its input_tokens and output_tokens.

To turn tokens into money, use the price your provider publishes for the model, per million tokens, for input and output separately:

cost ≈ input_tokens × input price / 1,000,000 + output_tokens × output price / 1,000,000

grip carries no price list: prices change, and a gateway or a plan may charge differently. With a coding agent the quiz counts against that agent's plan, and Ollama costs nothing. Use exclude and max_diff_bytes to keep large diffs in check, and see Cost and privacy to estimate the tokens for your own diffs. A grip appeal is one more request, recorded with the quiz it regrades.

Is my code sent anywhere else?¶

Only to the provider you configured, and only the diff plus your answers. With a coding agent as the provider, that is the agent's vendor, under the terms of your account with them. See Cost and privacy, How it works and SECURITY.md.

grip also keeps a local log of every graded quiz in .git/grip/history.jsonl and a list of quizzed repositories under ~/.local/share/grip/, so that grip study export can build one anonymised file of your scores across repositories, for a study platform. That file is only written when you run the command, contains no code, paths, question text, answers or feedback, and goes nowhere unless you upload it. See grip study.

Can questions be multiple choice?¶

Not currently. Free-text answers are a much better signal of understanding than recognition, and grading against a rubric keeps it fair. Open an issue if you have a use case.

Why exactly five questions?¶

It is enough to cover behaviour, motivation, an edge case, a risk and verification, and short enough that people answer rather than bypass. It is fixed on purpose so scores are comparable across repositories.

How do I cite grip?¶

GitHub's "Cite this repository" button reads CITATION.cff. In BibTeX:

@software{lejeune2026grip,
  author = {L., Erwin},
  title  = {grip: keep a grip on your code},
  year   = {2026},
  url    = {https://github.com/guilyx/grip}
}