chore(openai): use SpanAttributes.EMBEDDING_INVOCATION_PARAMETERS

[codefromthecrypt]

Remove hardcoded constant and outdated TODO comment. Use the official SpanAttributes.EMBEDDING_INVOCATION_PARAMETERS constant that was released in semantic conventions v0.1.25.

Bump openinference-semantic-conventions from >=0.1.23 to >=0.1.25.

---

Note

Serializes non-string tool outputs into message.content for agents and switches OpenAI instrumentation to the official EMBEDDING_INVOCATION_PARAMETERS attribute (bumping semantic conventions).

• OpenAI Agents:
  • Tool output extraction: custom_tool_call_output and function_call_output now set message.content; non-string outputs are JSON-serialized via safe_json_dumps.
  • Tests: Add cases for string/list/dict custom tool outputs; adjust function call output expectations.
• OpenAI Instrumentation:
  • Replace hardcoded embedding parameter key with SpanAttributes.EMBEDDING_INVOCATION_PARAMETERS.
  • Bump openinference-semantic-conventions to >=0.1.25.
• Tooling:
  • Add mypy override to disable warn_unused_ignores for tests.test_span_attribute_helpers.

^{Written by Cursor Bugbot for commit 7d97d96. This will update automatically on new commits. Configure here.}

[cursor]

Bug: Missing key existence check for output field

The function directly accesses obj["output"] without checking if the key exists, which could raise a KeyError if the field is optional. The old code used obj.get("output") for safe access, and the similar function _get_attributes_from_response_custom_tool_call_output_param checks if "output" in obj: before accessing. This inconsistency could cause runtime errors if FunctionCallOutput ever has a missing output key.