AssistantMessage

A response from the LLM, optionally containing text, tool-call requests, or both, and the reasoning the model produced before them.

content always returns a non-null String; it returns "" when the LLM response contains only tool calls and no accompanying text (contentOpt is None). Code that displays assistant output should check for an empty string rather than null-guarding.

A well-formed AssistantMessage must satisfy at least one of:

• contentOpt.exists(_.trim.nonEmpty) — the LLM produced text.
• toolCalls.nonEmpty — the LLM requested one or more tool invocations.

Thinking alone does not make a message well-formed.

Construct one with the companion apply and change it with the with* setters.

'''Sealed thinking.''' A signed ThinkingBlock.Text, a ThinkingBlock.Redacted or a ThinkingBlock.Opaque block is valid only in the conversation the provider produced it in. Anthropic signs a thinking block over everything sent before it - the system prompt, the tools and every earlier message - and rejects it once any of those changes; Bedrock documents its reasoning signature as a hash of all the messages in the conversation. Both also reject a block sent beside content or tool calls other than the ones it came with. Two rules keep llm4s from sending a signed turn that no longer holds:

• '''Its own message.''' withContent and withToolCalls, given a value different from the current one, '''unseal''' the thinking: they drop redacted and opaque blocks, the signatures of text blocks and the binding, keeping the reasoning text.
• '''Everything before it.''' The client that receives sealed thinking records, in thinkingBinding, a fingerprint of the request it answered and of who answered it: the provider and model, the system messages, the tools, the response format and every earlier message. When the message is sent again, the client replays its sealed thinking only if it is calling that provider and model and the conversation before it still has that fingerprint, and sends it unsealed otherwise - so a signature is never sent to a provider or model that did not produce it. So pruning, compression, summarisation, an edit or an insertion anywhere earlier in the history - by any code, through any API - unseals every later turn, without the code that made the change having to know about thinking.

Unsealed thinking keeps its text, which Anthropic and Bedrock leave out of the request and the other providers still receive. Earlier turns whose history is unchanged keep their sealed thinking, as Anthropic recommends: the blocks it accepts back are an unbroken run of the original sequence, and these rules never leave a gap, because unsealing one turn changes the history of every later one. To keep signed thinking, leave the message and everything before it as they are; to replace it deliberately, use withThinking.