Developer Patrick McAnna spent part of his weekend poking around Claude Code's session logs and found something unsettling: the "thinking blocks" that should contain an AI agent's reasoning are encrypted into a 600-character signature with no readable text attached.

The Encryption Problem

McAnna dug into Anthropic's own documentation and discovered that Claude encrypts its full reasoning process before it ever reaches your machine. The company holds the decryption key—not you. When developers call the extended-thinking API, they're receiving a summary of what happened, not the actual thought chain that drove the model's behavior. "This is like saving a jpeg as a .bmp and then editing the .bmp and presenting it as a .jpeg," McAnna wrote. "The conversion produces data loss."

What You're Actually Getting

The ctrl+o interface labeled "extended thinking" returns a summary of Claude's reasoning logic, not the genuine chain-of-thought that led to specific actions in your session. If you're relying on this for audit trails, debugging, or security review—bad news. The local files offer no readable insight into what the model actually considered.

Enterprise Lock-In

Want the real thinking output? That requires an enterprise agreement with Anthropic. For everyone else—individual developers, small teams, startups—you're locked out of your own agent's decision-making process by design. Security researcher Matt Green has also examined these signature blocks and raised additional concerns about transparency. The language in Anthropic's docs doesn't exactly scream "you can't see what we did here" either—it's buried in caveats that are easy to miss if you're skimming documentation after your morning coffee. "I'm underwhelmed by how Anthropic is presenting the behavior of their application," McAnna noted.

Key Takeaways

  • Your Claude Code session logs contain encrypted signatures, not readable reasoning
  • The extended-thinking output is a summary—useful context but not an audit trail
  • Full access to actual thinking requires an enterprise contract with Anthropic
  • You can log inputs and outputs through scraping, but never the true decision logic

The Bottom Line

This isn't just a documentation gap—it's a fundamental transparency problem for anyone building production systems on Claude Code. If you need verifiable reasoning trails from your agents, the open-source alternatives are looking more attractive by the day. Anthropic needs to come clean about what developers are actually buying.