fix(anthropic): strip content from assistant messages with tool call for bedrock

[Pnkcaht]

What I Did

• In pkg/model/provider/anthropic/client.go, updated the convertMessages function to strictly separate content (text) and tool_use blocks in assistant messages.
• When msg.ToolCalls exists, we now ignore any msg.Content (text) to comply with Anthropic / AWS Bedrock strict protocol (no mixing allowed).
• Added logging (warn/error) for:
  • Ignored content (helps debug when LLM ignores system prompt)
  • Missing tool call IDs
  • JSON unmarshal failures on tool arguments
• Preserved thinking blocks (allowed with tool_use)
• This fixes the ValidationException on Bedrock when using transfer_task in multi-agent setups.

Releated Issue

Closes #1593

Before & After

Before

Before (caused ValidationException on Bedrock):

{
  "role": "assistant",
  "content": "Let me proceed with analyzing the issue...",
  "tool_calls": [
    {
      "id": "tooluse_ABC123",
      "type": "function",
      "function": {
        "name": "transfer_task",
        "arguments": "{\"agent\": \"scout\", \"task\": \"Retrieve data from source A\", \"expected_output\": \"Summary of findings\"}"
      }
    },
    {
      "id": "tooluse_XYZ789",
      "type": "function",
      "function": {
        "name": "transfer_task",
        "arguments": "{\"agent\": \"coder\", \"task\": \"Analyze the implementation\", \"expected_output\": \"Technical analysis\"}"
      }
    }
  ]
}

After

After (now compliant with Anthropic/Bedrock protocol):

{
  "role": "assistant",
  "tool_calls": [
    {
      "id": "tooluse_ABC123",
      "type": "function",
      "function": {
        "name": "transfer_task",
        "arguments": "{\"agent\": \"scout\", \"task\": \"Retrieve data from source A\", \"expected_output\": \"Summary of findings\"}"
      }
    },
    {
      "id": "tooluse_XYZ789",
      "type": "function",
      "function": {
        "name": "transfer_task",
        "arguments": "{\"agent\": \"coder\", \"task\": \"Analyze the implementation\", \"expected_output\": \"Technical analysis\"}"
      }
    }
  ]
}

Diagram

sequenceDiagram
    participant Coordinator as Coordinator Agent
    participant LLM as Anthropic/Bedrock
    participant Subagents as Subagents (scout, coder)

    Note over Coordinator,Subagents: Before (broken)
    Coordinator->>LLM: Send message with text + transfer_task calls
    LLM-->>Coordinator: Returns assistant msg with content + tool_calls
    Coordinator->>Bedrock: Send conversation history
    Bedrock-->>Coordinator: ValidationException (mixing forbidden)

    Note over Coordinator,Subagents: After (fixed)
    Coordinator->>LLM: Send prompt (no text + tool calls)
    LLM-->>Coordinator: Returns assistant msg with only tool_use blocks
    Coordinator->>Subagents: Execute transfer_task
    Subagents-->>Coordinator: Return tool results
    Coordinator->>LLM: Send next turn with tool_results
    LLM-->>Coordinator: Final analysis (only content, no tool_calls)
    Coordinator->>User: Response

Loading

[krissetto]

/review

[github-actions]

Review Summary

Found 1 potential issue in the changed code related to tool call handling that could cause message sequencing problems.

Key Finding:
The new validation logic skips tool calls with missing IDs but still appends the message, which could cause mismatches between tool_use and tool_result blocks.

Note: The JSON unmarshal fallback behavior exists in both old and new code - the PR only adds logging, so that's not a new issue.

[krissetto]

we now ignore any msg.Content

I'm not sure we should do this, maybe it'd be better to split it out into its own message before the tool calls

[Pnkcaht]

we now ignore any msg.Content

I'm not sure we should do this, maybe it'd be better to split it out into its own message before the tool calls

Do you want me to do it? I'll do it without a problem, it needs to be perfect :)

[Pnkcaht]

@krissetto Done! I implemented the split: if there's content + tool calls, I send the message first with just the text/reasoning (preserving the logic) and then only with the use of the tool.

Full history, Bedrock protocol ok. Suggestion made by the AI ​​also resolved in the other commit.