Claude Code reports February regression after 234,760 tool calls

• A GitHub issue cited by the primary evidence claims Claude Code has been unreliable on complex engineering work since February, based on analysis of 6,852 session logs, 17,871 thinking blocks, and 234,760 tool calls.
• According to the issue report, the reported failures were not vague quality complaints, they included ignored instructions, incorrect fixes, and cases where the agent did the opposite of what was requested.
• In Anthropic's response on Hacker News, Boris from the Claude Code team said the cited redact-thinking-2026-02-12 header only hides thinking in the UI and skips summaries to reduce latency, not the underlying thinking budget.
• The HN thread also surfaced adjacent hypotheses and field reports, including read-to-edit ratio shifts, phrase-level signs of shallow reasoning, and one user's claim that the 1M context window made complex coding worse.

You can read the full issue, Anthropic's settings docs, and the main Hacker News thread. The oddest detail is how operational this report is: log mining, tool-call counts, and even read-versus-edit ratios, not just vibes. The other useful reveal is that Anthropic's public reply narrowed one disputed point fast, the redact-thinking header was described as a UI and latency change, not a reasoning-budget change.

The issue was unusually concrete for a model quality complaint. It framed the regression as something visible in months of usage data from a stable, high-complexity environment with heavy CLAUDE.md conventions.

The report's core claim was that extended thinking is load-bearing for senior engineering workflows, and that behavior degraded after February. Its evidence bundle mixed counts and symptoms: 6,852 logs, 17,871 thinking blocks, 234,760 tool calls, plus repeated failures to follow instructions or produce correct fixes.

Boris from the Claude Code team said the issue was blending two separate things. His reply said redact-thinking-2026-02-12 hides thinking from the UI because most people do not read it, and avoids generating thinking summaries, which cuts latency.

That matters mainly because the issue had correlated the header rollout with lower-quality behavior. Anthropic's response said the header does not change thinking itself or extended reasoning budgets, and pointed to showThinkingSummaries: true in settings.json for people who want summaries visible again.

The surrounding discussion added three concrete leads:

• In the author's follow-up, the issue author said a stop-phrase guard was accidentally omitted from the original post, and suggested checking older logs for read:edit ratio shifts, thinking-character shifts, and post-redaction correlation.
• The same discussion summary also highlighted another user's report that Opus 4.6 felt similarly worse through Copilot and other surfaces, with phrases like "simplest fix" preceding bad code and new "too many tokens/turns" language.
• One linked comment in the HN thread raised a narrower operational theory, that many bad cases might be tied to the 1M context window, and said complex coding improved again after turning it off.