Meat — Reading Diff

Meat is a Go CLI that abridges a code diff into a reading diff: the same change, rewritten to keep only what a senior reviewer actually needs to read. The premise is that reviewing agent-written code no longer means sweating style, nil-checks, or imports — those are covered by the compiler and tests. What remains is the change to the program: what moved, where data came from, what new behavior appeared. Meat strips the mechanical noise and shows you the meat.

The architecture's central trust move is that the model never authors output text. Meat sends the model a numbered unified diff and asks for an edit plan — remove ranges, fold ranges, and single-line replace elisions, all in original line coordinates. A deterministic compiler then validates the plan (elision projections, move symmetry, Python structure rules, retained metadata) and mechanically renders the reading diff from the immutable input. The output is always a compressed subsequence of the original, so the model cannot invent code, comments, or behavior.

Around that core sit several subsystems: a per-language mandatory import pass that removes imports/includes/requires/use declarations automatically (Go, Python, JS/TS, Rust, C/C++, Java); exact move detection that pairs relocated blocks across hunks/files and enforces symmetric treatment so a move reads as relocation, not one-sided deletion; chunking that splits oversized diffs (up to 4 MB) at file→hunk→sub-hunk boundaries and merges per-chunk results; and a cache keyed by the SHA-256 of (rubric hash + model + diff) so unchanged re-runs are instant.

It is provider-agnostic with zero third-party dependencies: built-in backends speak the OpenAI Responses API (default gpt-5.6-sol) and Anthropic Messages API (default claude-opus-4-8) over the Go standard library. Install with go install meat.dev/cmd/meat@latest, then run meat to review the latest commit, or pipe any diff in.