There is no free()

The text Claude Code prints when you press ctrl+o, the block it labels extended thinking, is a summary of the model’s reasoning. It is not the reasoning that drove the agent’s actions in your session. Anthropic says so in its own documentation, in a line that is easy to read past: “extended thinking returns a summary of Claude’s full thinking process.”

Patrick McCanna describes it as a lossy round trip. Save a bitmap as a JPEG, edit the JPEG, then save it back as a bitmap. What you get looks like the original format, but the conversion discarded data you cannot recover. The extended-thinking block is the JPEG. The full thinking process that Fable/Opus actually ran is the bitmap you never see.

That matters the moment you treat the output as a record. If you need to show what logic your agent used during a session, and that requirement turns up in regulated work, incident review, and any audit worth the name, the files on your machine will not give it to you. The reasoning logs are not accessible to you. You can scrape the inputs, the outputs, and the actions a running Claude Code takes, and doing so is worth the effort. A transcript of prompts, tool calls, and results is a real record of what the agent did. It is not the reasoning that produced those actions, and the summary you can read is a paraphrase generated alongside the work, not a faithful trace of it.

So before you promise anyone an audit trail, be precise about what you can actually hand over. Log the I/O and the tool invocations, because those are real and reconstructable. Treat the extended-thinking text as a readability aid for whoever is watching the session, not as evidence of how a decision was reached. If your compliance story depends on reproducing the decision logic behind an automated action, this does not support it, and the indirect phrasing in the docs makes it easy to assume otherwise.

McCanna’s own takeaway is that performance improvements in open-source models need to come faster. A model you host is a model whose reasoning tokens you can capture yourself, rather than reading a summary of them after the fact.