fix(openapi): preserve anyOf for non-null union body params instead of injecting an arbitrary type

[K4bain]

fix(openapi): preserve anyOf for non-null union body params instead of injecting an arbitrary type

Problem

For a request body field typed as a union of two real types (e.g. dict[str, list[str]] | list[str]), the projected MCP tool schema gets a single arbitrary "type" injected alongside anyOf. The MCP server's jsonschema validator then enforces that "type" strictly and rejects every call using the variant that lost the pick, with a misleading error like:

Input validation error: {'Skills': [...]} is not of type 'array'

The injected type is also non-deterministic (next(iter(set))), so which variant breaks can differ per run. Fixes #307.

Root cause

In convert_openapi_to_mcp_tools, both the query-param and body-param loops do:

if "type" not in properties[param_name]:
    properties[param_name]["type"] = get_single_param_type_from_schema(param_schema)

For T | U (both real types) Pydantic emits {"anyOf": [...]} with no top-level "type", so the branch fires and one arbitrary variant type wins.

Fix

Only fall back to get_single_param_type_from_schema when the union has at most one non-null variant. T | None (nullable) keeps the existing collapse behavior — its existing tests still pass unchanged — while T | U keeps anyOf verbatim, which the JSON Schema validator at the MCP layer handles correctly.

Verification

• New test test_union_type_schema_not_collapsed asserts anyOf is preserved with no injected "type" and both object/array variants present.
• Repro from the issue validated with jsonschema: both the dict variant and the list variant pass validation (previously one of them always failed).
• Full suite: 70 passed (the 20 test_sse_real_transport errors are the pre-existing Windows os.fork limitation tracked in test: skip real-transport tests on Windows (they require os.fork) #332).

[K4bain]

Follow-up pushed: the guard now also covers oneOf-only unions (hand-written OpenAPI specs), which previously still received a bogus single injected type alongside the union — same rejection risk as the anyOf case. New test test_oneof_body_schema_not_collapsed covers it. 79/79 tests green.

[K4bain]

Second follow-up: the path-param loop had the same injection bug (path param typed int | str got both anyOf and a bogus injected type — verified live before the fix). Same guard applied there; test_union_path_param_not_collapsed covers it. All three param loops (path/query/body) now treat unions uniformly. 80/80 tests green.