Description

temperature and top_p were removed from the Interactions API's GenerationConfig / GenerationConfigParam in 2.14.0. The 2.14.0 changelog has no entry for this, and it wasn't flagged as a deprecation in any prior release.

The removal is type-only. GenerationConfig is extra="allow", so passing temperature still validates and still serializes onto the wire — code keeps working at runtime but fails type checking. That combination makes it easy to "fix" by deleting the field, silently changing sampling behavior.

This also leaves the two surfaces inconsistent: types.GenerateContentConfig (used by models.generate_content) still has both temperature and top_p as first-class fields in 2.18.1.

Reproduction (2.18.1)

from google.genai import version
from google.genai.interactions import GenerationConfigParam
from google.genai._gaos.types.interactions.generationconfig import GenerationConfig

print("google-genai:", version.__version__)
print("temperature in GenerationConfig:", "temperature" in GenerationConfig.model_fields)
print("extra policy:", GenerationConfig.model_config.get("extra"))
print("dump:", GenerationConfig.model_validate({"temperature": 0.0}).model_dump())

cfg: GenerationConfigParam = {"temperature": 0.0}  # pyright: reportAssignmentType

Runtime output:

google-genai: 2.18.1
temperature in GenerationConfig: False
extra policy: allow
dump: {'temperature': 0.0}

Type checker (pyright, standard mode):

error: Type "dict[str, float]" is not assignable to declared type "GenerationConfigParam"
    "temperature" is an undefined item in type "GenerationConfigParam" (reportAssignmentType)

Bisected across wheels: present in 2.13.0, absent from 2.14.0 onward.

Questions

1. Is this intentional — has the Interactions API stopped accepting temperature/top_p, or is it a codegen omission?
2. If intentional: is temperature now ignored server-side, or still honored? Callers doing deterministic structured extraction need to know whether removing it changes model behavior. A changelog entry and a deprecation window would help.
3. If a codegen bug: can the fields be restored?

Environment

• google-genai 2.18.1 (also reproduces on 2.14.0–2.18.0)
• Python 3.13, Linux