When Patrick McCanna opened his Claude Code session logs over the weekend, he expected to find the model’s reasoning. What he found instead was a 600-character encrypted signature and no readable text. His investigation and the resulting post on patrickmccanna.net, published June 22, 2026, raises a question that every builder using visible chain-of-thought as a debugging tool should sit with: what exactly is being displayed?
The core claim McCanna makes, drawing from Anthropic’s own documentation at platform.claude.com, is this: Claude encrypts its reasoning into an opaque signature block. Anthropic holds the decryption key. The developer’s machine does not receive it. What the API returns is a summary of the reasoning, not the reasoning itself. And accessing the full thinking output, McCanna says the docs indicate, requires an enterprise agreement.
Some of this is independently verifiable. The Anthropic documentation for extended thinking does describe a “summarized thinking” mode, and it notes that what the API hands back is a condensed version of the model’s complete reasoning rather than the raw trace. That language appears in the public docs. McCanna’s characterization that this distinction is easy to miss if you skim the page is fair. The phrasing is indirect.
What is harder to independently verify, and where McCanna is expressing his own read of the situation, is the severity of the gap. He compares what users receive to saving a bitmap image as a JPEG, then editing the JPEG and saving back to bitmap: the conversion introduces lossy compression. The analogy captures the directional concern. A summary is not a lossless copy of the source reasoning. But how much fidelity is lost, and whether the summary reliably reflects the actual decision logic, is not something McCanna’s post establishes with precision. He is stating a structural concern, not providing a technical measurement.
That structural concern is worth taking seriously on its own terms.
Builders who use the “thinking” output in Claude Code have been treating it as a debugging surface: a way to understand why the model took a particular action, to audit unexpected behavior, or to satisfy a stakeholder who wants to see the reasoning behind an agent’s decision. If that text is a summary generated after the fact rather than the actual step-by-step reasoning that produced the action, the debugging surface is significantly weaker than assumed.
The audit trail problem is the sharper edge here. McCanna flags it directly: understand this constraint before you guarantee anyone a reasoning record. If you have committed to a customer, a regulator, or an internal compliance team that your AI-assisted workflow produces a record of how the model reasoned through a decision, the local session logs as currently structured do not deliver that. You can capture inputs, outputs, and actions. You cannot, on a standard account, capture the actual reasoning chain that connected them.
Anthropic has not published a detailed technical response to McCanna’s post, and this article does not attribute any Anthropic statement not found in the public documentation. The company’s docs do describe encrypted reasoning signatures. They do describe summarized thinking as the default return value. Whether the summary is adequate for the debugging and audit purposes developers have been using it for is a product question Anthropic has not publicly addressed in those terms.
The practical implication for builders is narrow but real. If you have built a workflow where the visible chain-of-thought is load-bearing, either as a trust signal for end users or as an audit artifact for compliance, you should re-examine that assumption now. The text you are reading is Anthropic’s summary of the model’s thinking. It is not a window into the reasoning. Whether that distinction matters for your use case depends on what you were relying on it for.
Teams building on Claude Code for enterprise or regulated applications should clarify with Anthropic what the enterprise extended thinking agreement actually unlocks, and whether it provides a verifiable reasoning record or a more detailed summary. Those are different things, and the distinction matters before you make a compliance commitment you cannot support.
Source: Patrick McCanna, writing on his personal blog at patrickmccanna.net, published June 22, 2026.
