The text in Claude Code’s “Extended Thinking” output
Points and comments are a snapshot, not live.
Claude Code encrypts its reasoning logs into signatures users cannot decrypt locally.
Claude Code records sessions to disk with 'thinking blocks' that are encrypted into 600-character signatures. Anthropic holds the decryption key; the machine never receives it. The API returns only a summary of reasoning, not the actual reasoning. Full thinking output requires an enterprise agreement. Matt Green's blog provides further observations on the signature blocks. The author warns this undermines promises of audit trails and notes that conversion to a summary introduces data loss.
What commenters are saying
Commenters generally agree this is unsurprising: OpenAI and Google similarly hide raw reasoning, largely to prevent model distillation by competitors. Several note that reasoning traces are already one level removed from internal processes (KVCache states), and summaries are lossy but practically useful. One correction: saving a JPEG as BMP and back is lossy, not the reverse (author admitted the error). A minority view argues the practice reflects hubris or a weak moat.
There is no strong protest; the thread is analytical rather than outraged.