Claude Code's extended thinking output is a summary, not the real reasoning

Patrick McCanna argues that the "Extended Thinking" text Claude Code shows you while it works is not the model's real reasoning. It is a summary. The actual chain of thought is encrypted into a signature that your machine receives but cannot read, because Anthropic holds the decryption key. McCanna says he went looking for the raw thinking in his own session logs and found only a 600-character signature with no readable text behind it. Anthropic's own documentation backs the basic point: extended thinking "returns a summary of Claude's full thinking process," a line buried deep enough that most users never see it.

His worry is about audit trails. If you run an agent and want a faithful record of why it did something, the local files do not give you one. What you have is a paraphrase produced by another pass of the model, not the steps it actually took. He compares it to saving a JPEG as a BMP, editing the BMP, and then calling it the original: the conversion already lost information. Full access to the underlying reasoning, he notes, requires an enterprise agreement.

This is a presentation problem more than a bug. The output looks like a transcript of thought, so people read it as one. McCanna's post is a reminder that the words scrolling past during a long agent run are a reconstruction, and that treating them as ground truth for debugging or compliance is a mistake.

Why it matters

If you rely on Claude Code's thinking output to debug agent behavior or to satisfy an audit, you are trusting a summary, not a log. For anything that has to hold up later, capture the inputs, tool calls, and outputs yourself rather than the thinking text.