Claude Code's Extended Thinking Is a Summary, Not Raw Reasoning

A 600-character signature is all that sits in Claude Code’s local session logs where the model’s reasoning should be. No text, no chain-of-thought, just a blob you cannot decrypt.

I went to inspect Claude Code’s reasoning this weekend after reading that it writes thinking blocks to disk. The logs contain a 600-character signature and nothing else. Anthropic holds the key. Your machine never receives the actual reasoning.

What the API Actually Returns

Reading the docs clarifies the deception: “Extended thinking returns a summary of Claude’s full thinking process.” Not the thinking itself. A summary. Anthropic encrypts the model’s internal reasoning, hands you a compressed version, and calls it extended thinking.

Matt Green has documented the signature blocks in more detail. The language in Anthropic’s documentation is deliberately indirect. If you haven’t had coffee, you might miss that “summary” part entirely.

The Enterprise Paywall on Transparency

Full thinking output requires an enterprise agreement. If you are a regular user or a team evaluating Claude Code for agentic workflows, you cannot access the reasoning that drove your agent’s actions. The logs on your system are not accessible to you.

This is like saving a JPEG as a BMP, editing the BMP, and presenting it as the original JPEG. Data loss is baked in. You can scrape inputs, outputs, and actions of a running Claude Code session, but you will never capture the actual reasoning that produced those outputs.

Why This Kills the Audit Trail

If you ever need a record of the logic your agent used in a session, you cannot produce one using the local files. The reasoning logs on your system are not yours. Anthropic controls the key.

Anyone promising an audit trail for Claude Code agents is promising something that does not exist. The tool outputs a summary, not the reasoning. For safety-critical or regulated applications, this is a non-starter.

Open source models need to catch up faster. Until then, trust in Claude Code requires ignoring what Anthropic actually ships.

Source: Claude Code's "extended thinking" is a summary- not authentic thinking
Domain: patrickmccanna.net