The text in Claude Code's "Extended Thinking" output is not authentic

Somebody actually went looking under the hood of Claude Code this weekend, and what they found should bother anyone who assumed those "extended thinking" logs were a faithful transcript of the model's reasoning. Spoiler: they aren't.

Here's the setup. Claude Code writes session logs to disk, and those logs include what Anthropic calls "thinking blocks" — supposedly a record of the model working through a problem before it acts. One developer went digging through their local logs to audit a session and instead found a 600-character encrypted signature where the reasoning should have been. No readable text. Just a blob only Anthropic can decrypt, because Anthropic holds the key and your machine never gets it.

Digging into Anthropic's own documentation confirmed the suspicion: the API returns a summary of the reasoning, not the reasoning itself. Want the actual, unabridged thinking trace? That requires an enterprise agreement. Researcher Matt Green apparently poked at these signature blocks too and has more detail on how they're structured. The developer who flagged this compared the summarization process to saving a bitmap as a JPEG, editing the JPEG, then trying to save it back as a bitmap — the compression step throws away information you can't recover, no matter how you repackage it afterward.

The practical fallout is uncomfortable for anyone building on Claude Code and hoping to produce an audit trail. You can scrape inputs, outputs, and the agent's actions well enough. But the actual logic that drove those actions during a session simply isn't accessible from your own local files, no matter how deep you dig. And the documentation doesn't exactly shout this limitation — the key sentence, that extended thinking