agentd run with claude-code fails when prompts include tool_use blocks: Anthropic API 400 says assistant messages must start with thinking blocks when thinking is enabled. We drop thinking blocks when replaying tool_calls (Op.PromptMessage has no thinking blocks). Need to preserve thinking blocks or disable thinking for tool-call replay.
Disable extended thinking when tool_use blocks are present without preserved thinking blocks; avoids Anthropic 400 for claude-code/tool replay.