Repository navigation
validate_schema skips validation for non-BaseModel schemas (list[str], dict[…], raw JSON Schema) #7280
Description
Description
validate_schema (google/adk/utils/_schema_utils.py) and SchemaType advertise support for list[str], list[int], dict[str, T], raw dict JSON schemas, and types.Schema.
Only BaseModel / list[BaseModel] go through Pydantic. The else branch is a bare safe_json_loads — no TypeAdapter / jsonschema check — so invalid values are silently accepted into session state / tool results.
Environment
google-adk==2.9.2- Offline unit repro (no live API / GCP)
Repro
from google.adk.utils._schema_utils import validate_schema
from pydantic import BaseModel, ValidationError
class M(BaseModel):
a: str
# Control: BaseModel still validates
try:
validate_schema(M, '{"a": 1}')
except ValidationError:
print("basemodel rejects bad: OK")
print(validate_schema(list[str], "[1, 2, 3]")) # → [1, 2, 3] WRONG
print(validate_schema(dict[str, int], '{"a": "x"}')) # → {'a': 'x'} WRONG
schema = {"type": "object", "properties": {"a": {"type": "string"}}, "required": ["a"]}
print(validate_schema(schema, '{"a": 1}')) # → {'a': 1} WRONGObserved (google-adk==2.9.2)
list[str]+"[1, 2, 3]"→[1, 2, 3]dict[str, int]+'{"a": "x"}'→{'a': 'x'}- Raw JSON Schema requiring string
a+'{"a": 1}'→{'a': 1} BaseModelpath still raisesValidationError(control OK)
Expected
Non-BaseModel schemas should be validated (e.g. TypeAdapter(schema).validate_json(...) for GenericAlias types; jsonschema for raw dict / types.Schema), or unsupported types should fail loudly — not silently accept invalid values.
Impact
Agents with output_schema=list[str] (etc.) + output_key write invalid values into session state. AgentTool returns unvalidated payloads to the parent. Downstream agents/tools treat garbage as schema-valid.
Suggested fix
For GenericAlias / non-dict schema types, use TypeAdapter(schema).validate_json(json_text) (same as list[BaseModel]). For raw dict / types.Schema JSON schemas, validate with jsonschema (already a dependency) or reject unsupported types loudly.
Notes
Nearby #5054 (output_schema=str + tools loop) and #6747 (python-mode dump) are different root causes. Happy to open a PR.
Activity
chelsealong commentedon Sep 25, 2026
Picking this one up now — opening a PR shortly. Flagging it here so nobody duplicates the work; if someone is already on it, say so and I will drop mine.
michaelkillgta commentedon Sep 25, 2026
Thanks for picking this up so quickly — and for the PRs (#7285 / #7284 / #7283). Happy to review or clarify anything from the original repros if useful.
llalitkumarrr commentedon Sep 29, 2026
Hello @michaelkillgta,
Your observations in this report are correct validation is being skipped which may cause issues for agents with an
output_schema. The linked PR addresses this concern but it requires an in-depth review. We appreciate your patience while the reviewer examines it.Additionally if you could share your observations on the PR, that would be a great help.
michaelkillgta commentedon Sep 30, 2026
Looked at #7285. Routing generic aliases through TypeAdapter, and dict /
types.Schemathrough jsonschema, matches the report. The new tests cover thelist[str]anddict[str, int]rejections, plus raw JSON Schema andtypes.Schema.Callers that only catch pydantic's ValidationError will still miss
jsonschema.ValidationErroron the dict andtypes.Schemapath, so it may be worth wrapping those into one error type. I reviewed the diff only and have not run the suite.michaelkillgta commentedon Sep 30, 2026
Follow-up after running the tests. On #7285 (
8dff62f),test_schema_utils.pypassed 71 tests.The split is real.
list[str]anddict[str, int]raise pydantic'sValidationError. A raw JSON Schema dict and atypes.Schemaraisejsonschema.ValidationErrorfor the same kind of bad value.