Skip to content
- Text files only. Binary files and files over
maxFileSizeKB are skipped.
- The replay is a reconstruction, not a recording. It shows each change as a person would plausibly type it, not the agent’s literal edit sequence (which is always an atomic write anyway).
- Every matching file needs a baseline for its first change to replay. In a git repository with git installed, clean tracked files use git as their baseline and cost no memory (
gitBaselines). Other files are kept in memory up to baselineBudgetMB, newest first. A file left out starts recording from its second change, and the Agent Typewriter output channel says how many files that is.
- The buffer keeps at most
historyLimit changes and historySizeMB of change text (before plus after, a character counted as a byte), whichever is reached first; the oldest changes are dropped. The newest change is always kept, even when it alone is bigger.
- A rename is only detected while its first half is still held for reordering; a delete and a create further apart replay separately.
- On Linux, VS Code’s file watcher does not look inside a folder tree made several levels deep at once (
mkdir -p, git worktree add into a new folder, a folder moved in): it reports the top folder and nothing written below it. Agent Typewriter watches those folders itself if they were made while it was recording in this window: up to 2000 per lane or new worktree waiting for its first change, and no more than a quarter of the inotify watches the system allows (fs.inotify.max_user_watches), which VS Code’s own watcher shares. Changes in a tree made before that are missed until the window is reloaded (Developer: Reload Window). So are changes in folders made at once inside a folder that was there before recording and is deleted and made again in one quick step (rm -rf dir && mkdir -p dir/...): VS Code reports that folder as merely changed, the same as a chmod.