What Agent Touched

See which files your AI agent changed, right in the VS Code explorer.

• Explorer colors: M yellow (modified), A green (added), D red (deleted; shown on parent folders and in the list).
• Open a changed file -> automatic diff of original before the agent touched it vs now (once per change; the editor title diff button reopens it).
• "What Agent Touched" list in the explorer and a status bar summary (👁 3M 1A 0D) of the files in the open workspace. Files the agent changed outside the workspace (e.g. its own config or memory) are kept in a collapsed Outside workspace group with their own counts. Clear starts a new round.
• Works in any permission mode, including fully automatic ones. No review workflow, no Keep/Undo: it only shows.
• Cost is proportional to the files the agent touches. Nothing is scanned, so huge workspaces and network file systems are fine.

Supported agents: Claude Code (VS Code extension and terminal CLI alike). The store format is agent-neutral, so other agents only need a hook adapter.

How it works

Claude Code (CLI / VS Code)          VS Code extension
        |                                  ^
  PreToolUse hook (Edit/Write/...)          | polls the store, compares
  copies the file's original the            | originals with files on disk
  first time a session touches it           |
        v                                  |
  ~/.what-agent-touched/sessions/<agent>-<session_id>/
      session.json      files/<key>.json      base/<original bytes>

This is Cursor-style copy-on-write: the original is saved right before the agent's first write to a file, so diffs are exact without any baseline snapshot. A round is one agent session; files touched by several sessions are compared with the earliest original. A file whose content returns to its original drops off the list.

Not covered: files changed through shell commands (sed -i, scripts, make, ...). Like Claude Code's own checkpoints, only the file editing tools are hooked. Your own manual edits to a touched file do show up in its diff.

Install

Requirements: VS Code 1.80+, Python 3.6+ (hook, standard library only).

1. Install the extension (.vsix, see Build): code --install-extension what-agent-touched-<version>.vsix With Remote-SSH / WSL, run it in the remote terminal so it installs where the agent runs.

2. Install the Claude Code hook, either with the command palette: What Agent Touched: Install Claude Code Hook, or by hand:

   python3 hooks/install_claude_code_hook.py     # Linux / macOS
   py hooks\install_claude_code_hook.py          # Windows
   

   It copies the hook to ~/.what-agent-touched/hooks/ and adds one PreToolUse entry to ~/.claude/settings.json (or $CLAUDE_CONFIG_DIR), keeping everything else and writing a timestamped backup first. Re-running it replaces the entry; --uninstall removes it.

3. Start a new Claude Code session. Edits show up within the poll interval.

Settings

Colors can be changed through workbench.colorCustomizations: whatAgentTouched.modifiedForeground, .addedForeground, .deletedForeground.

Build and test

Plain JavaScript, no build step, no dependencies.

python3 scripts/package_vsix.py              # -> what-agent-touched-<version>.vsix
python3 -m unittest discover -s test         # hook + installer
node test/test_store.js                      # store logic (drives the real hook)
node test/smoke_extension.js                 # extension against a fake vscode API

Set PYTHON=... for the Node tests if python3 (python on Windows) is not the right interpreter.

Adding another agent

Write a hook that, right before the agent writes a file, creates sessions/<agent>-<session_id>/files/<sha1(normcase(path))>.json with O_EXCL (first capture wins) containing {"path", "existed", "blob", "size", "mtime_ms", "captured"}, copies the original to base/ and touches session.json. See hooks/claude_code.py.

License

MIT