Agents sound sure. They say tests passed, they only touched one file, Jane was emailed once. You can't check that from the thread. Half the work never shows up. Traceproof keeps a log in the project of every action, including the hidden ones, then tells you where the story stopped matching the files. Next chat reads that log first, so it doesn't continue from a lie.
Pick a lie the agent already told. Traceproof compares the recap to the tape.
Minimal tokens. The log stays on disk, not in the prompt.
A plugin for Claude Code. The same zip is a skill on ChatGPT. Open a project and say "set up Traceproof." Then ask what actually happened when something looks off.