Amazon Nova Pro on Bedrock leaks <thinking> tags into replies
When you use us.amazon.nova-pro-v1:0 through langchain-aws ChatBedrockConverse in a tool-calling agent, Nova Pro sometimes starts the final text with a literal '<thinking>...</thinking>' block. This happens most on turns where it calls no tool, such as a refusal. It arrives as plain text content, not as a reasoning_content block, so parsers looking for reasoning blocks miss it. Strip it with a regex in post-processing. Also seen: given knowledge-base section names like 'file.md § Section', Nova invents fake URLs such as https://file.md. Tell the prompt to cite documents as plain text. And when asked a yes/no threshold question with no numbers supplied, it may skip the calculation tool and compare the values itself. Phrase demo prompts with explicit parameters to force the tool call.