Skip to content
Vizzini

Getting Started

Vizzini has no tagged release yet, so build it from source. The engine, the pkg/reviewer Go API and the pr-review CLI are complete and tested.

Requirements

Section titled “Requirements”
Tool Version Used for
Go 1.25+ Building and testing (cuelang.org/go sets the floor)
cue CLI v0.17.1 make vet-cue
golangci-lint v2 make lint
curl, sha256sum any Seeding the tokenizer cache once
A C compiler any Only for make test, because Go’s race detector needs cgo
An OpenAI-compatible endpoint A LiteLLM proxy in front of the model you review with

Build and test

Section titled “Build and test”
Terminal window
git clone https://github.com/infrashift/vizzini-pr-reviewer
cd vizzini-pr-reviewer
make test      # seeds .cache/tiktoken on first run, then runs go test -race
make vet-cue   # cue vet ./schema/...
make lint      # golangci-lint + go mod tidy check
make build     # static bin/pr-review

The first make test downloads the two tokenizer encodings (about 5.3 MB) and checks them against pinned SHA-256 hashes. After that, everything runs offline. The tests actively block network access. See Running Offline.

Your first review

Section titled “Your first review”
Terminal window
export VIZZINI_LITELLM_URL=http://litellm:4000/v1
export VIZZINI_MODEL=<model name in your LiteLLM config>
export VIZZINI_API_KEY=<proxy key>
export TIKTOKEN_CACHE_DIR=$PWD/.cache/tiktoken


git diff main...HEAD | ./bin/pr-review --mode review

You get Markdown: a summary, a risk score, and findings sorted most severe first, each pinned to file:line. Add --format json for machines, or try --mode describe for a PR description and --mode improve for concrete code suggestions.

Next: the CLI Reference, Embedding in Go, or How a Review Works.