Agent forensics reconstructs what the model saw, what it decided and what actually executed. Collect session transcripts, checkpoints, shell history, git reflog and provider logs, and compare them, because each can be incomplete. LLMs can help triage large logs, but keep a human verifying every conclusion.
Read in this order
- OfficialExplore the .claude directory, including session transcripts (Claude Code docs)Start here: where Claude Code stores session transcripts on disk.
- Open sourceReconstructing AI agent activity: Two new scripts for forensic review (SANS ISC)Free scripts that rebuild what an agent did from its local files, October 2026.
- PaperDigital Forensics in the Age of Large Language ModelsWhat LLMs can and cannot do in forensic analysis.