FunctionSchema has no strict field — OpenAI Responses tools always send strict: null

[mannyb223]

Description

pipecat.adapters.schemas.function_schema.FunctionSchema has no way to opt a tool into OpenAI's
strict function-calling mode (JSON-schema-guaranteed argument validation, which OpenAI's own
docs recommend enabling by default). This applies to both hand-written FunctionSchema tools and
direct functions (DirectFunctionWrapper.to_function_schema()), since both build the same
schema shape with no strict concept.

open_ai_responses_adapter.py's to_provider_tools_format reads a "strict" key off the dict
produced by FunctionSchema.to_default_dict():

tool: FunctionToolParam = {
    "type": "function",
    "name": d["name"],
    "parameters": d.get("parameters", {}),
    "strict": d.get("strict", None),
}

But to_default_dict() never sets that key, so every Responses-API tool ships with an explicit
"strict": null, never true, and there is no FunctionSchema constructor argument, decorator,
or ToolsSchema option to change that. (open_ai_adapter.py, the Chat Completions path, doesn't
even read/forward a strict key at all.)

The only existing workaround is ToolsSchema.custom_tools[AdapterType.OPENAI], which requires
hand-writing the full raw OpenAI tool dict (bypassing FunctionSchema entirely, and losing its
handler auto-registration) — that's documented as the escape hatch for provider-native tools that
don't fit the standard schema, not a way to add strict: true to an otherwise-standard tool.

Repro (stock pipecat, no custom code beyond this)

from pipecat.adapters.schemas.function_schema import FunctionSchema
from pipecat.adapters.schemas.tools_schema import ToolsSchema
from pipecat.adapters.services.open_ai_responses_adapter import OpenAIResponsesLLMAdapter

weather_function = FunctionSchema(
    name="get_weather",
    description="Get the weather for a location.",
    properties={"location": {"type": "string"}},
    required=["location"],
)
tools_schema = ToolsSchema(standard_tools=[weather_function])

adapter = OpenAIResponsesLLMAdapter()
tools = adapter.to_provider_tools_format(tools_schema)
print(tools[0]["strict"])  # -> None, not True

Expected

Some way to request strict: true (and get additionalProperties: false emitted) directly from
FunctionSchema/direct functions — e.g. a strict: bool = False constructor arg on
FunctionSchema that flows through to_default_dict() into both the Responses and Chat
Completions adapters, defaulting to False for backward compatibility.

Actual

strict is always null/absent regardless of the tool's schema shape, with no supported way to
change it short of bypassing FunctionSchema via custom_tools.

Environment

pipecat 1.8.1, openai Python SDK 2.54.0, OpenAI Responses API.

[omChauhanDev]

The strict read at open_ai_responses_adapter.py:133 is the only selective one. That adapter builds its tool dict field by field (:129-137), naming name, parameters, strict and description, so it picks up a key from to_default_dict() deliberately. The other consumers that build a payload take the dict whole:

• gemini_adapter.py:159-164 replaces func_dict["parameters"] and appends the rest verbatim into function_declarations.
• open_ai_adapter.py:211 casts the whole dict into ChatCompletionToolParam.function.
• bus/adapters/tools_schema_adapter.py:31 serialises it as-is.

So a strict key added to to_default_dict() would not stay on the OpenAI paths. It would ride into Gemini's function declarations and the bus payload for every tool, carrying its default rather than appearing only when something opted in.

FunctionSchema already has the shape for a field like this. handler is a constructor argument with a property and is deliberately not in to_default_dict(), because it is not part of a provider payload. A strict property read directly by the Responses adapter, func.strict rather than d.get("strict"), leaves the dict's shape untouched, which is what those other consumers rely on.

[mannyb223]

Agreed, a strict property on FunctionSchema read by the Responses adapter is the right shape, and it keeps the dict untouched for Gemini and the bus. Happy to test a PR on our side if one lands.