26 Unix tools. One binary. Zero dependencies.·the missing coreutils for the agent era·vrk mcp - list all 26 tools to any AI agent·curl vrk.sh/install.sh | sh - ready in 5 seconds·

Diff review budget

A huge diff fails fast and cheap, instead of getting cut off halfway through the review.

Gate a git diff on token count before it goes to the model.

Pipeline

git diff main | vrk tok --check 20000 | vrk prompt --system "Review this diff"

The problem

Asking a model to review a diff is one of the most useful things you can do with one. The trouble is the diff nobody planned for: a lockfile update, a generated file, a rename across the tree. It may be 80,000 tokens. The model either refuses, truncates it, or reviews the first part and sounds confident about the rest.

How the pipeline works

git diff main prints your changes against main. vrk tok --check 20000 counts the tokens. At 20,000 or fewer, it passes the diff through unchanged. Over that, it exits 1 with nothing on stdout, and vrk prompt has nothing to send.

vrk prompt sends the diff with the system prompt Review this diff. Put your own review rules in the system prompt, or load them from a file with --system @review-rules.txt.

When the diff is too big

Pick one of these:

  • Review part of it: git diff main -- src/, or look at git diff main --stat first.
  • Leave out noise: git diff main -- . ':!package-lock.json'.
  • Split it with vrk chunk and review each piece.

In a script, test the exit code:

if ! git diff main | vrk tok --check 20000 > review-input.diff; then
  echo "diff too big for one review, see git diff --stat" >&2
  exit 1
fi
vrk prompt --system "Review this diff" < review-input.diff

Set the number from your model

20,000 is an example. Use your model’s context window, minus room for the system prompt and the reply.

Tools used