find_context_parameter treats a Context return annotation as a parameter

[otiscuilei]

Initial checks

• I tested the current main branch (6e30452).
• I searched existing issues and pull requests for this behavior.

Description

find_context_parameter() iterates over every entry returned by typing.get_type_hints(). That mapping includes the function return annotation under the special key "return", even though it is not a parameter.

As a result, a tool, prompt, or resource function whose return annotation is Context (or a union containing Context) is incorrectly recorded as having a context parameter named "return". The SDK later attempts to inject context as a keyword argument with that name, which can produce an unexpected-keyword error when the callable is invoked.

Expected: context detection only examines actual function parameters and ignores the return annotation.

Example code

from mcp.server.mcpserver import Context
from mcp.server.mcpserver.utilities.context_injection import find_context_parameter

def fn(value: int) -> Context:
    ...

print(find_context_parameter(fn))  # current: "return"; expected: None

Environment

• Python 3.12
• MCP Python SDK main at 6e30452

I used AI assistance while investigating this issue and reviewed and reproduced the behavior locally.

[nightcityblade]

Hi, I'd like to work on this. I'll submit a focused PR with regression coverage shortly.

[nightcityblade]

I noticed the contribution guide asks contributors to wait for maintainer feedback or a ready-for-work signal before implementation. I'm stepping back for now so this can be triaged first.

[hishammoizuddin]

So I ran into this and dug in a bit. The cause is that get_type_hints() hands back the return annotation under a "return" key alongside the real parameters, and find_context_parameter loops over the whole mapping, so a handler annotated -> Context ends up recorded as having a context parameter literally named "return". Since return is a keyword, injecting it can never work: the handler registers fine and then blows up the first time it's called.

One thing that surprised me is that it isn't just tools. Prompts and resource templates resolve their context through the same function, so all three break the same way.

Fix is to drop the "return" key before scanning the hints. I've put up a PR with that plus regression tests, I'd be happy to close it if you'd rather take a different approach.

Disclosure: I used AI assistance (Claude Code) on this. I read through the change myself and can explain it.