Claude Code’s “extended thinking” is a summary, not a full record
Claude Code saves on the local machine, but the thinking section does not contain readable reasoning text. It contains a roughly 600-character signature instead. Claude’s is encrypted, and Anthropic keeps the key; the user’s computer does not receive it.
The normal API returns a summary of the reasoning, not the full reasoning itself. Full access to the thinking output requires an . This means Claude Code cannot provide a complete audit trail showing exactly why the agent acted the way it did.
Inputs, outputs, and actions can still be logged separately, but those logs do not reveal the actual reasoning that drove the behavior. The concern is that the product and docs can make “” sound more complete than it really is.
Key points
- Claude Code’s local logs do not include the full readable reasoning text.
- The reasoning is encrypted, and Anthropic controls the key.
- The API returns a summary, not the complete .
- Full thinking access requires an .
- For important work, keep separate logs of prompts, outputs, actions, and approvals.