BerriAI/litellm · error · OCIError

Tool call `function.arguments` must be a JSON string

Error message

Tool call `function.arguments` must be a JSON string

What it means

Tool call arguments must be a JSON-encoded string (defaulting to '{}' when absent), matching the OpenAI wire format. Passing arguments as a dict/object instead of a serialized string raises OCIError(400) "Tool call `function.arguments` must be a JSON string" in the OCI GENERIC adapter.

Source

Thrown at litellm/llms/oci/chat/generic.py:137

            raise OCIError(status_code=400, message="Each tool call must be a dictionary")
        if tool_call.get("type") != "function":
            raise OCIError(status_code=400, message="OCI only supports function tool calls")

        tool_call_id = tool_call.get("id")
        if not isinstance(tool_call_id, str):
            raise OCIError(status_code=400, message="Tool call `id` must be a string")

        tool_function = tool_call.get("function")
        if not isinstance(tool_function, dict):
            raise OCIError(status_code=400, message="Tool call `function` must be a dictionary")

        function_name = tool_function.get("name")
        if not isinstance(function_name, str):
            raise OCIError(status_code=400, message="Tool call `function.name` must be a string")

        arguments = tool_call["function"].get("arguments", "{}")
        if not isinstance(arguments, str):
            raise OCIError(
                status_code=400,
                message="Tool call `function.arguments` must be a JSON string",
            )

        tool_calls_formatted.append(
            OCIToolCall(
                id=tool_call_id,
                type="FUNCTION",
                name=function_name,
                arguments=arguments,
            )
        )

    return OCIMessage(
        role=open_ai_to_generic_oci_role_map[role],
        content=None,
        toolCalls=tool_calls_formatted,
        toolCallId=None,

View on GitHub (pinned to 6c2dcb801b)

Solutions

1. Serialize: arguments=json.dumps({...}).
2. Echo tool calls back verbatim from the model's own response instead of reconstructing them by hand.
3. Pre-flight: coerce with json.dumps if not isinstance(arguments, str).

Example fix

# before
{'id':'c1','type':'function','function':{'name':'get_weather','arguments':{'city':'SF'}}}

# after
import json
{'id':'c1','type':'function','function':{'name':'get_weather','arguments':json.dumps({'city':'SF'})}}

import json

for tc in msg.get('tool_calls') or []:
    args = tc.get('function', {}).get('arguments', '{}')
    if not isinstance(args, str):
        tc['function']['arguments'] = json.dumps(args)

def tool_call_args_are_json_string(tc: object) -> bool:
    args = tc.get('function', {}).get('arguments', '{}') if isinstance(tc, dict) else None
    if not isinstance(args, str):
        return False
    try:
        json.loads(args)
        return True
    except (TypeError, ValueError):
        return False

When it happens

Trigger: Sending {'function': {'name':'f','arguments': {'city':'SF'}}} — a live dict — in assistant tool-call history to an oci/ GENERIC model. This is the classic mistake when hand-building history instead of echoing back what the model produced.

Common situations: Developers writing arguments as objects because it reads nicer; frameworks that auto-serialize dicts and stop doing so after an upgrade; mixing conventions between request 'tools' (where parameters is a dict) and history 'tool_calls' (where arguments is a string).

Related errors

• Each tool call must be a dictionary
• OCI only supports function tool calls
• Tool call `id` must be a string
• Tool call `function` must be a dictionary
• Tool call `function.name` must be a string

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/8554522c47658361. Report an issue: GitHub.