think-block-leakage#

The model's internal reasoning escapes into the reply the user sees. This usually means an unclosed <think> / <thinking> tag, or a reply that begins with planning prose instead of the answer itself.

Symptoms#

• Reply contains literal <think>, <thinking>, or similar reasoning tags.
• Reply opens with "Let me think...", "Okay, the user wants...", "First I'll need to...", or other planning preamble.
• Reply is cut off mid-sentence and an opening reasoning tag has no matching close.
• A </think> appears with only a few dozen characters of content after it.

What to do#

• Inspect the raw LLM output for unmatched reasoning tags before returning it. Strip or redact any content inside reasoning tags.
• If the provider supports a separate reasoning channel, emit reasoning there and keep it out of the reply body entirely.
• If leakage is detected, regenerate the reply. Do not ship reasoning-as-answer.
• If the reply starts with planning language, trim the preamble. The user should see the answer first.
• For persistent leakage, tighten the system prompt to forbid meta-commentary in the reply body.