Anthropic’s Claude Code logs each session to disk, storing “thinking blocks” that capture the model’s internal reasoning. A recent inspection revealed those blocks contain only a 600‑character encrypted signature, not the raw text. The platform’s docs confirm the API returns a summary of the reasoning while the full transcript remains encrypted and inaccessible to the client. This design choice limits transparency for teams auditing AI behavior.

Because Anthropic retains the decryption key, only customers with an enterprise agreement can request the complete thinking output. Independent researcher Matt Green documented the signature format and warned that promising an audit trail based on local logs is misleading. The “extended‑thinking” shortcut (Ctrl+O) merely presents a compressed view of Fable/Opus logic, not the actual decision‑making steps. Thus users cannot verify specific model actions.

Developers relying on Claude Code for compliance or debugging must record inputs, outputs, and actions externally, as the internal reasoning cannot be extracted. Scrappy scraping can capture surface interactions but offers no insight into the model’s hidden chain of thought. In practice, the feature behaves like saving a JPEG as a BMP—information is lost during conversion. Consequently, any compliance claim must rely on external logging.