Tool ergonomics: dict schemas + keyword-only @tool name

[eliasposen]

Summary

• input_schema accepts JSON Schema dicts alongside Pydantic BaseModel classes. Dicts validated via jsonschema (Draft 2020-12); pydantic models still go through model_validate. Lets integrators wrap tools defined externally (MCP, OpenAPI, hand-written) without round-tripping through a synthetic pydantic model.
• output_schema accepts JSON Schema dicts alongside Python types / typing constructs. TypeAdapter path preserved for rich Python output validation (datetime, BaseModel subclasses, etc.).
• @tool decorator: name is now keyword-only; fn is the only positional argument. Enables @tool(input_schema=...) and @tool(output_schema=...) without forcing the user to also pass a name. Validators are compiled and cached at tool construction; malformed dict schemas raise jsonschema.SchemaError immediately.

Breaking change

@tool("custom_name") is no longer valid — use @tool(name="custom_name"). Migrated all in-repo call sites.

Test plan

• Pre-existing tool/decorator/schema tests still pass
• New tests for dict input/output schemas (validation success + failure paths)
• New tests for malformed dict schemas raising at construction
• New tests for explicit Pydantic model / Python type schemas overriding inference
• New test for combined tool(fn, name=..., input_schema=..., output_schema=...) direct-call form
• Async coverage for both schema forms
• Updated error-path tests to match the new TypeError-based dispatch

🤖 Generated with Claude Code