How much of your code did an AI write?
Three steps, one repository, a measured answer.
Step 1 Install the recorder
One command, nothing leaves the machine.
Step 2 Let it backfill
It reads the history your tools already wrote.
Step 3 Read the result together
Including the part it cannot attribute.
Most teams guess, and guess low.
The raw share of lines a model produced is usually a large number and a misleading one. The figure that matters is how much AI-written code survived review and is still in production — and how much of the file nobody can account for at all.
What you get out of the session
AI-authored share by repository and by file
Which coding agents were actually used
Which models, where the tool records them
The proportion of lines that cannot be attributed
How much of the history predates any recorder
Whether hand edits were being counted as AI work
Directories where AI authorship is concentrated
A dated report you can keep
The definition behind every figure
An honest view of what the method cannot see
Bring the repository you are least comfortable about.
Thirty minutes, one codebase, and a number you can defend afterwards.