When an AI message contains both text content and tool calls (e.g., [TextContent, ToolCall]), only the first part is serialized. The tool calls are silently dropped, causing Anthropic API errors when sending tool results.

Location

llms/anthropic/anthropicllm.go in handleAIMessage()

Problem

func handleAIMessage(msg llms.MessageContent) (anthropicclient.ChatMessage, error) {
    if toolCall, ok := msg.Parts[0].(llms.ToolCall); ok {  // Only checks Parts[0]
        // handles single ToolCall
        return ...
    }
    if textContent, ok := msg.Parts[0].(llms.TextContent); ok {  // Only checks Parts[0]
        // handles single TextContent
        return ...
    }
    return anthropicclient.ChatMessage{}, fmt.Errorf(...)
}

When Anthropic returns a response with both text ("I'll help you...") and a tool call, and we build an assistant message with Parts: [TextContent, ToolCall], only the TextContent is serialized. The ToolCall is dropped.

This causes the error:

messages.2.content.0: unexpected `tool_use_id` found in `tool_result` blocks: toolu_xxx. 
Each `tool_result` block must have a corresponding `tool_use` block in the previous message.

Expected Behavior

handleAIMessage should iterate through all Parts and include both text and tool_use content blocks in the Anthropic message.

Workaround

Put ToolCall first in Parts and skip text content when there are tool calls:

assistantMsg := llms.MessageContent{Role: llms.ChatMessageTypeAI}
// Put tool calls first (langchaingo only reads Parts[0])
for _, tc := range toolCalls {
    assistantMsg.Parts = append(assistantMsg.Parts, tc)
}
// Skip text content - langchaingo can't handle both

Related

This is similar to #1010 (multi-part tool results) but affects AI/assistant messages instead.