The schema has the final say
LLM output is untrusted input. Closed CUE definitions reject unknown severities, invented fields and impossible line ranges, so callers always get typed data with a guaranteed shape.
Vizzini
Vizzini is a stateless AI code review engine written in pure Go. It reads a unified diff and drops generated files. It labels every line with its old and new line number, packs the hunks into token-bounded chunks, and asks an LLM behind your LiteLLM proxy for findings. Every answer is unified with a CUE schema before it reaches the caller. An answer that doesn’t fit is sent back to the model to fix.
Vizzini is built as a library first. Its primary consumer is the
mrman terminal review tool. It also
ships as a static pr-review binary for CI, UBI9 containers and offline QEMU
microVMs.
Getting StartedBuild Vizzini from source and review your first diff.
Why Vizzini
Section titled “Why Vizzini”Findings that point at real lines
The model sees every hunk line labeled with its old and new line
numbers, so a finding’s start_line refers to an actual line in the
new file.
Exact token budgets
Chunks stay within a tiktoken budget. Large files split only between hunks, and the token counts are exact, not estimated.
Small, static, offline
CGO_ENABLED=0, a short pinned dependency list, and checksum-pinned
tokenizer encodings, so it never needs the network at runtime.
Learn more
Section titled “Learn more”How a Review WorksFrom patch to validated findings, stage by stage.
CLI ReferenceFlags, environment, output formats and exit codes.
Embedding in GoUse pkg/reviewer from mrman or your own tool.
Review SchemaThe CUE contract every answer must satisfy.
Running OfflineSeeding the tokenizer cache for containers and microVMs.
Design DecisionsWhy pure Go, why CUE, why LiteLLM.