OpenAI tool-call arguments should normalize on parse

[davidahmann]

Problem observed

OpenAI chat and responses parse paths currently hand back function-call arguments as raw JSON strings, while the other backends in this repo already normalize equivalent tool-call payloads to mappings.

Why it matters operationally

That backend-specific envelope shape leaks into downstream logic and tests. Tool-call handling is supposed to be provider-agnostic at the framework boundary, but callers end up needing to special-case OpenAI responses or remember to re-parse arguments defensively. The inconsistency is small, but it undermines deterministic workflow behavior and makes backend swaps harder to reason about.

Minimal repro

1. Parse an OpenAI function-call payload whose arguments field is a JSON object string.
2. Inspect the resulting Content.from_function_call(...).arguments value.
3. Parse an equivalent tool call from a backend that already emits mappings.
4. Expected: equivalent mapping-shaped arguments. Actual before the fix: OpenAI kept the raw string.

Fix approach

The change stays at the parse boundary. A helper normalizes provider arguments to a mapping when the payload is a JSON object, and the OpenAI chat/responses parsers use it immediately on ingest. That avoids changing outgoing provider payload formatting while still removing the internal shape mismatch that downstream code sees.

Validation evidence

• uv run pytest packages/core/tests/openai/test_openai_chat_client.py::test_parse_tool_calls_from_openai_normalizes_json_object_arguments packages/core/tests/openai/test_openai_responses_client.py::test_response_content_creation_with_function_call packages/core/tests/openai/test_openai_responses_client.py::test_parse_response_from_openai_function_call_includes_status -m "not integration" passed.
• The focused tests now assert mapping-shaped arguments on the OpenAI parse path.

Open follow-up question for maintainers

Do you want the same normalization applied to any remaining Azure/OpenAI ingestion surfaces that intentionally preserve raw strings today?

Inspired by research context: CAISI publishes independent, reproducible AI agent governance research: https://caisi.dev