bug(genai): nested object schemas have required overwritten and nullable flag dropped #1725

Please excuse the robot-generated issue text. I've isolated and confirmed the bug, and that the monkey patches fix it.

When converting an MCP-style tool dict (or any JSON Schema with nested object properties) to Gemini's tool format, _function_utils._get_properties_from_schema corrupts nested objects in two distinct ways.

Affected versions: langchain-google-genai==4.2.2 (latest); google-genai==1.73.1.

Bug A — required is overwritten with "every property without a default"

_function_utils.py lines 573–576:

if isinstance(v_properties, dict):
    properties_item["required"] = [
        k for k, v in v_properties.items() if "default" not in v
    ]

For any nested OBJECT property, the converter discards the source schema's own required array and rebuilds it as "every property that lacks an explicit default key". A schema with required: ["name", "role"] and 6 nullable-optional fields ends up with all 8 fields marked required.

No existing issue or PR I could find addresses this.

Bug B — nullable flag dropped on the second pass

The converter is invoked twice during a single tool conversion: once by _format_json_schema_to_gapic on the outer schema, and again when _dict_to_genai_schema recurses into the inner object. The first pass correctly converts anyOf:[T, {"type":"null"}] → {type: T, nullable: true}. The second pass calls _is_nullable_schema, which only recognizes anyOf shape and the NULL type — not an explicit nullable: true key. So the flag silently disappears on second pass.

PR #1470 fixes this from the other side (avoids producing the explicit nullable: true shape in the first place); my fix below patches _is_nullable_schema to recognize the shape that already gets emitted. Either approach works.

Minimum reproduction

Self-contained, copy-paste runnable. Requires only pip install langchain-google-genai==4.2.2.

from langchain_google_genai._function_utils import _format_to_genai_function_declaration

tool = {
    "name": "create_contact",
    "description": "Create an administrative contact under a group.",
    "parameters": {
        "type": "object",
        "properties": {
            "contact": {
                "type": "object",
                "properties": {
                    "name":  {"type": "string"},
                    "role":  {"type": "string"},
                    "email": {"anyOf": [{"type": "string"}, {"type": "null"}]},
                    "phone": {"anyOf": [{"type": "string"}, {"type": "null"}]},
                },
                "required": ["name", "role"],
            },
        },
        "required": ["contact"],
    },
}

fn = _format_to_genai_function_declaration(tool)
contact = fn.parameters.properties["contact"]

# Bug A: required is overwritten with all properties (not just the input list)
print("required:", contact.required)
assert contact.required == ["name", "role"], \
    f"Bug A: expected ['name', 'role'], got {contact.required}"

# Bug B: nullable flag is dropped on nested anyOf:[T, null]
print("email.nullable:", contact.properties["email"].nullable)
assert contact.properties["email"].nullable is True, \
    f"Bug B: expected nullable=True, got {contact.properties['email'].nullable}"

Output on 4.2.2:

required: ['name', 'role', 'email', 'phone']
AssertionError: Bug A: expected ['name', 'role'], got ['name', 'role', 'email', 'phone']

(Bug B asserts after Bug A is fixed; the demo shows both fail.)

Proposed fix

Two surgical changes. Running these as monkey patches in production: both bugs disappear, our 9-case regression suite (top-level happy paths + nested + the real create_contact shape) goes 4/9 → 9/9 passing.

Bug A — honor source required when present

In _get_properties_from_schema, replace lines 573–576 with:

if isinstance(v_properties, dict):
    properties_item["required"] = (
        list(v["required"])
        if isinstance(v.get("required"), list)
        else []
    )

(The discarded synthesis branch was a heuristic for "Optional fields have a default set"; in practice both Pydantic's model_json_schema() and raw JSON Schema specify required explicitly, so a missing required should be honored as "no required fields", not "everything that lacks default". My original proposed fix above kept the synthesis as a fallback, but verifying against the real production case — a nested address object whose source had no required — showed the fallback still produces the bug.)

Bug B — detect explicit nullable: true

In _is_nullable_schema, add at the top:

def _is_nullable_schema(schema: dict[str, Any]) -> bool:
    if schema.get("nullable") is True:
        return True
    # ... existing logic ...

(Or merge with PR #1470's approach, which fixes the other end of the same round-trip.)

Happy to send a PR with both fixes plus the repro converted into a regression test if it'd help — let me know.