[Shreyavemana]

Summary

Registered output schemas lose the null branch of Optional fields. A tool that returns None for an X | None field produces a result that a client validating against the advertised outputSchema rejects.

Observed

A tool declares an output model with Optional fields (str | None = None, int | None = None) and returns None for them. A client that validates structuredContent against the advertised outputSchema fails the call with errors such as:

data/<field> must be string
data/<field> must be integer

Cause

In nitrostack/core/app.py:

• inspector_friendly_schema (via _unwrap_nullable_schema) rewrites anyOf: [T, null] to plain T. Its docstring says this is so MCP Inspector can render form fields.
• tool_json_schema applies the same rewrite to the output schema, so the registered outputSchema no longer allows null.

Repro

Run on main (0.4.0):

from typing import Optional
from pydantic import BaseModel
from nitrostack.core.app import tool_json_schema

class Out(BaseModel):
    maybe: Optional[str] = None
    n: Optional[int] = None

print(Out.model_json_schema()["properties"])
print(tool_json_schema(Out)["properties"])

Pydantic emits anyOf: [{type: string}, {type: null}] for maybe. tool_json_schema returns:

{'maybe': {'type': 'string', 'title': 'Maybe'}, 'n': {'type': 'integer', 'title': 'N'}}

NitroTestingModule.call_tool returns the null value without validating it against the schema, so the in-process harness does not catch this.

Expected

The registered outputSchema keeps the null branch, because it describes what the tool returns. The Inspector rewrite is only needed for input forms.

Workaround

Remove output_schema from @tool.

Proposal

1. Apply inspector_friendly_schema to input schemas only.
2. Add a test that registers an output model with an Optional field and checks the schema still accepts null.

[Shreyavemana]

The fix makes tool_json_schema keep the null option for Optional fields in output schemas, so tools can safely return None. Input schemas still use the Inspector rewrite, and all regression tests passed successfully (the new tests created - 5 along with the all the test cases - (818 total, 0 failed)).

[manish-wekan]

Install the entire nitrostack-py package from output-schema-keep-null branch, either locally or in /tmp/<name_of_dir>/
then import that package
and test it against #42

#42
If the tests pass, add the test report as comments. here.

[Shreyavemana]

Verification of the proposed fix (fix/output-schema-keep-null, 416fe98)

Changes: Removed 'inspectore_friendly_schema' from output - schema processing and added regression tests.

Installation and test environment

Cloned the branch into a temporary directory, created a fresh virtual environment, and installed the package using pip install <repo>.

The targeted tests were run from outside the repository, and the import resolved to venv\Lib\site-packages\nitrostack, confirming that the installed package was being tested rather than the source checkout.

Results:

• 31 tests passed against the installed package, covering output schemas, input schemas, and MCP contracts.
• 818 tests passed in the source checkout.
• End-to-end test passed: The original main branch fails to validate nullable output, while the fix branch correctly preserves nullability in the advertised JSON Schema.
• Input behavior unchanged: Optional fields are still handled correctly for Inspector form rendering.

Conclusion: The fix resolves the nullable output-schema issue, and all reported tests passed in their respective environments.