fix: save intermediate tool messages to prevent missing tool_result errors

[ahmed-sekka]

Summary

Fixes #396

When BedrockLLMAgent uses tools, the intermediate toolUse/toolResult conversation messages were only kept in a local list and never persisted to storage. This caused subsequent requests to fail with Bedrock validation errors about missing tool_result blocks.

• BedrockLLMAgent now collects intermediate tool messages in tool_conversation during both single and streaming responses
• Orchestrator saves these intermediate messages to storage alongside user input and final response
• When max_recursions is exhausted, toolUse blocks are stripped from the response to prevent saving orphaned toolUse without toolResult

Test plan

• Added unit tests for BedrockLLMAgent tool message collection (test_bedrock_llm_agent.py)
• Added unit tests for Orchestrator intermediate message persistence (test_orchestrator.py)
• Manual test: verify multi-turn conversations with tool-using BedrockLLMAgent + DynamoDBStorage no longer raise ValidationException

🤖 Generated with Claude Code

[brnaba-aws]

Thanks for you PR, we will review it and provide feedback. thanks

[cornelcroi]

Hey @ahmed-sekka, the fix solves a real problem but there are two things that need to be addressed before this can merge.

dispatch_to_agent is a public method and changing its return type from ConversationMessage | AsyncIterable[Any] to a tuple is a breaking change — any caller who has subclassed AgentSquad or called this method directly will get a silent runtime error. You need a backward-compatible approach, for example returning the tool conversation as part of the existing response object or through a separate optional channel. Second, in the streaming path tool_conversation is populated inside the stream_generator coroutine, but the orchestrator saves it before the stream is consumed, so it will always be empty at that point — the save inside process_stream is correct, the one before it is dead code and should be removed.