[BUG] Handle escaped JSON in tool call arguments

[paxiaatucsdedu]

Problem

LLMs occasionally return tool call arguments as escaped JSON strings (e.g., '"{\\"month\\": \\"August\\"}"') instead of plain JSON. This caused validation errors when creating AIMessage objects, as the arguments remained strings after a single json.loads() instead of being parsed into dictionaries.

Issue link: #52

Solution

Added double-parsing logic in OCIUtils.convert_oci_tool_call_to_langchain() to detect and handle escaped JSON.
Added escaped JSON test case in unit test.

Testing

Validated with real LLM responses showing the escaped JSON pattern:

system_msg = SystemMessage(content="You are a finance assistant. \nYou have access to these tools:\n- MonthlyVarianceSQL: fetch per-customer variance for a specific month and year\n- MultiMonthVarianceSQL: fetch variance summary for a year\nQuestion: {input}\nAlways use the appropriate tool first to get data, then explain in natural language.\n")

# REQUEST 1
print("REQUEST 1:")
messages1 = [
    system_msg,
    HumanMessage(content="Explain why cash inflow in August 2025 was below forecast.")
]
response1 = chat_with_tools.invoke(messages1)
print(f"Tool calls: {response1.tool_calls}\n")

# REQUEST 2 (with conversation history)
print("REQUEST 2:")
messages2 = [
    system_msg,
    HumanMessage(content="Who are my customers?\n"),
    AIMessage(content="Your request is incomplete. Please provide more details or specify the task you need help with."),
    HumanMessage(content="Explain why cash inflow in August 2025 was below forecast.\n"),
    AIMessage(content="", tool_calls=[{"id": "call_ed2635c686f449eea25915b2", "name": "monthly_variance_sql", "args": {"month": "August", "year": "2025"}}]),
    ToolMessage(content="<tool_response!>", tool_call_id="call_ed2635c686f449eea25915b2"),
    AIMessage(content="<Agent_reponse>"),
    HumanMessage(content="Tell me more information about Acme Corp  in this period")
]

try:
    response2 = chat_with_tools.invoke(messages2)
    print(f"Tool calls: {response2.tool_calls}")
except Exception as e:
    print(f"Error (this is the bug): {e}")

Output:

REQUEST 1:
Tool calls: [{'name': 'monthly_variance_sql', 'args': {'month': 'August', 'year': '2025'}, 'id': 'call_9a19748f1068440dbb0aee24', 'type': 'tool_call'}]

REQUEST 2:
Tool calls: [{'name': 'monthly_variance_sql', 'args': {'month': 'August', 'year': '2025'}, 'id': 'call_db1c9eb7f6514348b3d0983a', 'type': 'tool_call'}]

[kirankumarjoseph]

Didn't you need to address this in stream case too?

Suggested change

    def process_stream_tool_calls(
        self, event_data: Dict, tool_call_ids: Set[str]
    ) -> List[ToolCallChunk]:
        """
        Process Cohere stream tool calls and return them as ToolCallChunk objects.

        Args:
            event_data: The event data from the stream
            tool_call_ids: Set of existing tool call IDs for index tracking

        Returns:
            List of ToolCallChunk objects
        """
        tool_call_chunks = []
        tool_call_response = self.chat_stream_tool_calls(event_data)

        if not tool_call_response:
            return tool_call_chunks

        for tool_call in self.format_stream_tool_calls(tool_call_response):
            tool_id = tool_call.get("id")
            if tool_id:
                tool_call_ids.add(tool_id)

            original_args = tool_call["function"].get("arguments")
            unescaped_args = json.loads(original_args)
            # If the args are parsable into anything other than str,
            # then it is not double escaped, no need to json.load here.
            # Use original as is
            if isinstance(unescaped_args, str):
                final_args = unescaped_args
            else:
                final_args = original_args

            tool_call_chunks.append(
                tool_call_chunk(
                    name=tool_call["function"].get("name"),
                    args=final_args,
                    id=tool_id,
                    index=len(tool_call_ids) - 1,  # index tracking
                )
            )
        return tool_call_chunks