fix(parsing): handle empty function tool call arguments

[sohumt123]

• I understand that this repository is auto-generated and my pull request may not be merged

This change is entirely within src/openai/lib/ and tests/lib/, which CONTRIBUTING.md states the generator will never modify ("The generator will never modify the contents of the src/openai/lib/ and examples/ directories."), so it should be safe from being overwritten on the next generation.

Changes being requested

parse_function_tool_arguments() raises when a strict function tool call has empty-string arguments.

Problem

The API returns empty-string arguments (rather than "{}") for strict function tools that take no arguments or whose parameters are all optional. Both copies of parse_function_tool_arguments() pass that empty string straight to the parser:

• src/openai/lib/_parsing/_completions.py — json.loads(function.arguments) for strict tools and model_parse_json(input_fn.model, function.arguments) for PydanticFunctionTool
• src/openai/lib/_parsing/_responses.py — the same two paths with function_call.arguments

json.loads("") raises json.JSONDecodeError: Expecting value: line 1 column 1 (char 0), and model_parse_json(model, "") raises a pydantic ValidationError. This affects chat.completions.parse(), responses.parse(), and the streaming done events, so with chat.completions.stream() the exception leaks out of the stream iterator before get_final_completion() can be reached.

Interestingly, the SDK's own incremental streaming parser already guards this exact case on the delta path:

# src/openai/lib/streaming/chat/_completions.py
if (
    input_tool
    and input_tool.get("function", {}).get("strict")
    and tool_call_snapshot.function.arguments
):
    tool_call_snapshot.function.parsed_arguments = from_json(...)

but the shared parse_function_tool_arguments() helper used by the non-streaming and done/final paths does not, which is what led me to believe the empty-arguments case is real and this omission is an oversight rather than intentional.

Minimal repro on a clean checkout:

from openai.lib._parsing._completions import parse_function_tool_arguments
from openai.types.chat.chat_completion_message_function_tool_call import Function

parse_function_tool_arguments(
    input_tools=[{
        "type": "function",
        "function": {
            "name": "noop",
            "strict": True,
            "parameters": {"type": "object", "properties": {}, "additionalProperties": False},
        },
    }],
    function=Function(name="noop", arguments=""),
)
# json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

Fix

In both parse_function_tool_arguments() implementations, normalize empty/whitespace-only arguments to "{}" before parsing:

args = function.arguments  # (function_call.arguments in _responses.py)
if not args or not args.strip():
    args = "{}"

The raw path then returns {} and the PydanticFunctionTool path parses "{}" into the model (valid for zero-param/all-optional strict schemas, which are the only ones for which the API emits empty-string arguments). Non-empty arguments are byte-for-byte unaffected, so there is no regression surface. The normalization is scoped strictly to empty/whitespace strings; if the API ever sent "" for a schema with required fields, the Pydantic path would now surface a clearer ValidationError (missing fields) rather than a JSON-syntax error.

Testing

Added offline, plain-assert unit tests (no network, no mock server, no snapshots):

• tests/lib/chat/test_completions.py — direct helper (empty / whitespace -> {}, non-empty regression guard, PydanticFunctionTool no-args model -> instance) and an end-to-end parse_chat_completion case that asserts parsed_arguments == {}.
• tests/lib/responses/test_responses.py — the responses variant of the helper.
• tests/lib/chat/test_completions_streaming.py — feeds hand-built chunks through ChatCompletionStreamState; pre-fix handle_chunk raised JSONDecodeError, post-fix the tool_calls.function.arguments.done event fires with parsed_arguments == {} and get_final_completion() agrees.

python -m pytest tests/lib/ -q
267 passed

The nine empty-argument assertions fail on a clean checkout and pass with the fix; the full tests/lib/ suite passes, and ruff check / ruff format --check / pyright / mypy are clean on the changed files.

Additional context & links

None — this is a self-discovered bug; I could not find an existing issue or PR describing empty-string arguments to parse_function_tool_arguments.