[Bug] Extended thinking blocks leak into visible assistant output instead of being collapsed

Bug Description
[Bug] Extended thinking / internal reasoning blocks leak into the visible assistant output instead of being collapsed. Environment: - Claude Code 2.1.218 (Windows 11, PowerShell) - Model: opus[1m] (Opus 4.8 1M context) - effortLevel: "xhigh" (in settings.json) Symptom: The model's English internal reasoning (thinking block) is rendered in full at the top of the final answer, mixed with the actual (Korean) response. Occurs repeatedly within a long session. Trigger correlation: Frequency appears tied to effortLevel "xhigh" — higher effort produces longer thinking, which leaks more often. claude update (already on latest 2.1.218) did not fix it. Lowering effort to "high" via /effort is being tested as a workaround. Expected: thinking blocks stay hidden/collapsed, never inline in the answer.

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.218
• Feedback ID: 6c53186e-56d6-43a5-8da3-772d881105e3

Errors

[]