AI BriefingClaude Code's 'Extended Thinking' output is not the actual reasoning content
Key point
Claude Code's reasoning logs are merely encrypted signatures or summaries, not a complete record of the model's actual reasoning process.
Details
When Claude Code saves session records, it stores the model's reasoning process, the 'thinking blocks', to disk. However, checking the actual logs reveals that in some cases, only a 600-character signature is left instead of text.
Upon reviewing Anthropic's documentation, the following facts were revealed.
- Claude encrypts the reasoning content and stores it in the form of a signature, and the decryption key is held by Anthropic.
- The API returns only a SUMMARY of the reasoning, not the actual reasoning process.
- An Enterprise agreement is required to view the full reasoning output.
Therefore, the extended-thinking output obtained via ctrl+o is not the actual logic the model used to decide on its actions, but merely a summary of that logic. This is akin to converting a BMP file to JPEG and then saving it back as BMP, resulting in data loss.
As a result, it is currently impossible for users to fully record or audit the actual reasoning logic of a Claude Code agent through local files or scraping.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Claude Code Fails to Load AGENTS.md When Telemetry Is Disabled, With Workaround