Line Numbers

GitAuto formats every file it reads with right-aligned line numbers inside code fences. Each line is prefixed with its number using the format {line_num:>width}:{line}, producing output like:

  1:import React from 'react'
  2:
  3:export const App = () => (
  4:  <div>Hello</div>
  5:)

Why This Exists

Without explicit line numbers, the model describes changes like "replace the import statement" instead of "replace line 1." When a file has multiple similar-looking lines - duplicate imports, repeated patterns, or common variable names - the model frequently targets the wrong location, producing broken diffs. Pre-computed line numbers give the model an unambiguous coordinate system so it never has to count. Diffs become precise: "at line 47, replace this with that." The right-alignment ensures consistent formatting regardless of file length - a 9-line file and a 999-line file both produce clean, readable output.

Why Models Can't Count Lines

Models don't see text line-by-line - they process chunks of tokens, where a single token might span part of a line or multiple short lines. Counting to "line 47" in a 200-line file requires tracking position across dozens of tokens, and the error rate grows with file length. Tokenization turns source code into subword pieces that have no alignment with newline characters, so the concept of "line number" doesn't exist natively in how the model represents text. The model has to count from the top of the file to figure out where it is - and it routinely miscounts, especially in longer files where the accumulated error compounds with each line.

How It Works

When GitAuto reads a file from the repository, it calculates the width needed for the largest line number, then pads every line number to that width. A 50-line file uses 2-character padding; a 500-line file uses 3. The colon separator between the number and the content makes it unambiguous where the line number ends and the code begins.

This formatting is applied before the file content enters the model's context window, so the model always sees numbered lines and can reference them in its tool calls for creating or modifying files.

Related Features

• Full File Reads - ensures entire files are loaded so line numbers cover all content
• Diff Hunk Repair - fixes diffs when line references are slightly off