#702 Fix(anthropic): route tool input_schema through Flyte type engine Pull request opened by AdilFayyaz Fix: Route Anthropic tool input_schema through Flyte type engine The Anthropic plugin was generating input_schema for Claude tools using a hand-written Python type mapper (

_python_type_to_json_schema

). This mapper was unaware of Flyte's internal type system, causing correctness failures - most critically,

Literal["C", "F"]

parameters produced

{"type": "string"}

instead of

{"type": "string", "enum": ["C", "F"]}

, because NativeInterface stores Literal as a dynamic Enum and the old code didn't recognise it. This PR replaces the ad-hoc mapper with a proper type-engine. Changes: •

src/flyte/_json_schema.py

(new) — converts a LiteralType protobuf to a JSON schema dict •

src/flyte/models.py

— adds

NativeInterface.json_schema

property, which calls

TypeEngine.to_literal_type()

per input and converts via

_json_schema.py

•

src/flyte/_task.py

— adds

AsyncFunctionTaskTemplate.json_schema

delegating to self.interface.json_schema •

plugins/anthropic/.../agents/_function_tools.py

— removes TYPE_MAP, _python_type_to_json_schema, _get_function_schema; replaces with func.json_schema / NativeInterface.from_callable(func).json_schema •

plugins/anthropic/tests/test_agents.py

— rewrites type schema tests to use the new path; adds regression tests for Literal, datetime, timedelta, Optional, nested collections, and true unions Design choices: • json_schema lives on NativeInterface, not TypedInterface — NativeInterface is the right home because it also holds parameter defaults, which are needed to build the required list. TypedInterface is a serialization-time artifact and not available during tool registration. • Optional[X] is simplified to just X's schema — rather than emitting a full oneOf + null variant. Claude tool use already communicates optionality through the required array, so the null variant adds noise without value. • Deferred imports in json_schema — TypeEngine and flyteidl2 (protobuf) are heavyweight. They are imported inside the property body rather than at module load time to avoid slowing down every import of flyte.models. • input_schema is computed once at function_tool() call time and stored as a plain dict on the FunctionTool dataclass — it is never recomputed during the agent loop. Example: • Literal["C","F"] → NativeInterface (Enum) → TypeEngine → LiteralType → literal_type_to_json_schema → {"type":"string","enum":["C","F"]} → FunctionTool.input_schema → Claude API --- Test Plan • Run the full test suite: cd plugins/anthropic && uv run pytest tests/ -v • All 58 tests pass with no failures flyteorg/flyte-sdk