fix: serialize Claude tool-use round trips with Anthropic message blocks

[auyua9]

Claude tool-use round trips need Anthropic-native message blocks

ReActAgent appends an assistant message containing tool_calls and one or more tool-result messages before the next provider request. ClaudeProvider.generate() currently serializes those messages with the generic OpenAI-shaped Message.to_dict() payload (tool_calls on the assistant and role: tool for results). Anthropic Messages API expects tool_use blocks on the assistant turn and tool_result blocks inside a user turn.

This issue tracks a focused fix to serialize the existing provider-neutral messages into the Anthropic wire format and cover the round trip with unit tests.

[Dante-dan]

I can take this as a focused Claude provider serialization fix.

The existing agent message model is provider-neutral, but the next Anthropic request must preserve the tool round trip as an assistant tool_use block followed by user tool_result content. I’ll trace the current ReActAgent message sequence into ClaudeProvider.generate(), add the conversion at the provider boundary, and keep ordinary text-only messages and other providers unchanged.

I’ll build synthetic single- and multi-tool fixtures that assert block ordering, tool IDs, names, JSON inputs, result/error content, and surrounding assistant/user text. I’ll also cover malformed or unmatched tool messages so the provider fails predictably rather than emitting an OpenAI-shaped payload to Anthropic. Before submitting, I’ll run the focused provider/agent tests plus the repository’s applicable test and conventional-commit checks, and recheck for overlapping work.

[Dante-dan]

I found that #3057 was already opened by the issue author before my comment and covers the same provider-boundary tool_use/tool_result serialization. The existing review threads also capture the remaining payload-validation and immutability follow-ups, so I'll step back rather than duplicate that work. Thanks for taking this on.