Anthropic tool use argument error appending to nil

[ci]

Not entirely sure how to repro this yet, but logging in case others run into, or I can repro with verbose again.

I ran into a case where seemingly message.tool_calls was nil and we're trying to append to it in here:

** (ArgumentError) argument error
    :erlang.++(nil, [%LangChain.Message.ToolCall{status: :complete, type: :function, call_id: "toolu_01TAKvP91x1UE9BKHrVwVH5r", name: "my_function", arguments: nil, index: nil}])
    (langchain 0.3.1) lib/chat_models/chat_anthropic.ex:702: LangChain.ChatModels.ChatAnthropic.do_process_content_response/2
    (elixir 1.18.2) lib/enum.ex:2546: Enum."-reduce/3-lists^foldl/2-0-"/3
    (langchain 0.3.1) lib/chat_models/chat_anthropic.ex:389: LangChain.ChatModels.ChatAnthropic.do_api_request/4
    (langchain 0.3.1) lib/chat_models/chat_anthropic.ex:324: LangChain.ChatModels.ChatAnthropic.call/3
    (langchain 0.3.1) lib/chains/llm_chain.ex:552: LangChain.Chains.LLMChain.do_run/1
    (langchain 0.3.1) lib/chains/llm_chain.ex:398: LangChain.Chains.LLMChain.run/2
    iex:1: (file)

It somehow looks like Anthropic didn't respect the tool input_schema?

Sample of input/output from Anthropic:

Input

{
  "messages": [
    {
      "content": [
        {
          "text": "..",
          "type": "text"
        }
      ],
      "role": "user"
    },
    ..longchain..
    {
      "content": [
        {
          "cache_control": {
            "type": "ephemeral"
          },
          "content": "..output of previous tool",
          "is_error": false,
          "tool_use_id": "toolu_01TznQYPWgXE1Wa6EpFsN3RT",
          "type": "tool_result"
        }
      ],
      "role": "user"
    }
  ],
  "stream": false,
  "system": [
    {
      "cache_control": {
        "type": "ephemeral"
      },
      "text": "..systemprompt..",
      "type": "text"
    }
  ],
  "tools": [
    {
      "description": "description",
      "input_schema": {
        "type": "object",
        "required": [
          "arg1",
          "arg2"
        ],
        "properties": {
          "arg2": {
            "type": "string",
            "description": "arg2"
          },
          "arg3": {
            "type": "string",
            "description": "optional"
          },
          "arg1": {
            "type": "string",
            "description": "arg1"
          }
        }
      },
      "name": "my_function"
    },
    ..othertools..
  ],
  "model": "claude-3-7-sonnet-20250219",
  "max_tokens": 4096,
  "temperature": 1
}

Output

{
    "id": "msg_01EiTqy9J4Sow2d77DX6Km1R",
    "type": "message",
    "role": "assistant",
    "model": "claude-3-7-sonnet-20250219",
    "content": [
        {
            "type": "text",
            "text": "..ai.. let's do something:"
        },
        {
            "type": "tool_use",
            "id": "toolu_01TAKvP91x1UE9BKHrVwVH5r",
            "name": "my_function",
            "input": {}
        }
    ],
    "stop_reason": "max_tokens",
    "stop_sequence": null,
    "usage": {
        "input_tokens": 0,
        "cache_creation_input_tokens": 2179,
        "cache_read_input_tokens": 42095,
        "output_tokens": 4096
    }
}

[ci]

actually nvm.. just realized the "stop_reason": "max_tokens", 🤦

maybe we should handle this stop_reason with an error or something? 🤔

[brainlid]

Hi @ci! Ah, interesting. Yes, there is handling for the stop_reason max_tokens, but I think this surfaced an issue when the it was return during a tool_use message from the assistant. 🤔

The stop_reason should take precedence here. I think we can setup a failing test for this now.

[rubiagatra]

Where should I put in a test for this stop reason, and how do you want to make it @brainlid? maybe I can help

[brainlid]

I encountered this recently and have pushed up a commit on the main branch, which is after 0.4.0-rc.0. It will be officially released in the next RC or final version.