recipes
Guides shaped like your task, not like our features
The documentation answers «what does this key do». These pages answer «I have this problem, what do I do» — from the first review on your own machine to a gate the whole team stands on. Every recipe ends with a real run: its output, its duration and its price, measured on a named model schema and dated.
Nothing here is a mock-up. The excerpts come from runs on a public sample repository, so the numbers you read are the numbers we got — and, where a run went smaller than we asked, that is written down too.
01Your first local reviewInstall the binary, give it a model, review changes you have not committed — and learn what a green exit code does not mean.02The hook ladderFast review on commit, full one with the judge on push — with the thresholds, the exit codes and the four things that make the ladder expensive.03Rules your linter cannot checkWhich team agreements belong in the config, how to word them so the review acts on them, and the four different ways to say «do not flag this».04Which model finds, which one judgesGenerators, judges and the arbiter: what each role sees, why the judge must be the stronger model, and how a personal schema quietly replaces the team's.05Where the tokens goWhat a review bill is made of: the constant part your config sets, the cache that pays off only from the third run, and two settings that change the bill several times over.06The agent said «done»Three barriers between a connected MCP server and an agent that actually calls it — and the hook that catches what the agent skipped.07Review in a closed loopOllama or vLLM instead of a cloud model: which open model to run, measured; how to set up the server so it neither cuts the prompt nor calls home; and how to prove to your security officer that nothing leaves.08One key for the whole teamLocal review without a model key for every developer: the bot runs the review, and its key pays. Who gets access, how many slots to give developers, what the spend tracking records, and how to get the hooks running for the whole team.09«No findings» — did it look?An empty summary is four different messages. How to tell them apart, why to turn the diagnostics block on in advance, and how to read the reasons the judge gives for its drops.10No review came for this requestWhere the traces are and how to read them, what each one means, when to wait and when to look — including the «Resend» button that does nothing.11Threads resolved, code unchangedMeasure through the GitLab or GitHub API who closes the bot's threads and whether the code under them changed, tell three causes apart — and learn what closing a thread does to the gate.12Arguing with a findingHow to object so the bot answers, why it stays silent, and what an answer does not change — with a public exchange you can open and read.