Claude Code users face 'blank' reasoning blocks amid performance decline

Developers using Claude Code are reporting a critical failure in the tool's transparency, where the model's internal 'thinking blocks'—the reasoning chains used to plan code changes—appear blank or are stripped from the user interface. This loss of visibility prevents engineers from verifying the AI's logic before it executes edits to their codebase.

The issue is tied to a 'thinking redaction header' (redact-thinking-2026-02-12) implemented by Anthropic starting March 5, 2026, which hides these reasoning blocks from the UI. While the redaction hides the process, the underlying performance of the tool has also come under scrutiny. A community-developed tool called 'patch-cc,' available on PyPI, has emerged as an interactive patcher for the Claude Code native binary, allowing users to bypass these restrictions and enable 'live thinking.'

A Pattern of Regression

This UI failure follows a broader decline in the tool's autonomy and depth. Stella Laurenzo, Senior Director of AI at AMD, highlighted the regression in GitHub issue #42796 after conducting an extensive audit of 6,852 sessions. Laurenzo's analysis revealed a 67% drop in estimated thinking depth and a 70% decrease in the read-to-edit ratio following updates in February.

The regression is linked to the launch of 'adaptive thinking' for Opus 4.6 on February 9. Boris Cherny of Anthropic acknowledged that this adaptive approach could occasionally skip the reasoning phase entirely, a flaw that can lead to model hallucinations.

The Cost of Opacity

For professional developers, the thinking block serves as a vital safety mechanism. In an agentic 'plan → execute → verify' workflow, the ability to audit the plan is the primary defense against destructive or incorrect changes to a production codebase. When these blocks are redacted, the user loses the oversight required for trust, effectively turning the agent into a 'black box' that executes changes without a visible rationale.

This lack of transparency is particularly acute given the computational overhead associated with reasoning tokens. When the reasoning process is hidden, users are left in a position where they cannot verify the logic driving the agent's output.

What's Next

Attention now turns to whether Anthropic will roll back the redaction headers or provide more granular control over 'adaptive thinking' to restore the tool's previous depth. For now, power users are relying on third-party patches to regain the visibility necessary for complex engineering tasks. As Stella Laurenzo noted regarding the tool's current state, the decline in reliability means Claude cannot be trusted to perform complex engineering tasks without rigorous human oversight.