[Draft] Support functools.partial as tools, prompts, and resources

[strawgate]

Closes #3266

functools.partial objects failed with a PydanticSchemaGenerationError because inspect.isroutine() returns False for them, causing FastMCP to unwrap them via .__call__ into a method-wrapper that Pydantic can't handle. Additionally, functools.update_wrapper() sets __wrapped__ which causes Pydantic to follow it back to the original function, ignoring the partial's bound arguments.

This PR teaches FastMCP to recognize functools.partial as a first-class callable throughout the tool, prompt, and resource pipelines — skipping the __call__ unwrap and stripping __wrapped__ so Pydantic sees the partial's own signature.

import functools
from fastmcp import FastMCP

mcp = FastMCP("demo")

def add(x: int, y: int) -> int:
    return x + y

add_ten = functools.partial(add, y=10)
functools.update_wrapper(add_ten, add)

mcp.add_tool(add_ten)

Generated with Claude Code

[strawgate]

/marvin why do we need

        if isinstance(fn, functools.partial) and hasattr(fn, "__wrapped__"):
            fn = functools.partial(fn.func, *fn.args, **fn.keywords)

[marvin-context-protocol]

Claude finished @strawgate's task in 1m 34s —— View job

---

Why the __wrapped__ stripping is needed

The code:

if isinstance(fn, functools.partial) and hasattr(fn, "__wrapped__"):
    fn = functools.partial(fn.func, *fn.args, **fn.keywords)

exists to counteract an interaction between functools.update_wrapper() and Pydantic's schema generation.

The setup: When a user writes the common pattern:

add_ten = functools.partial(add, y=10)
functools.update_wrapper(add_ten, add)  # sets add_ten.__wrapped__ = add

update_wrapper copies attributes from add onto the partial, including setting add_ten.__wrapped__ = add (the original, unwrapped function).

The problem: Pydantic's schema generation follows __wrapped__ to introspect the function. When it sees __wrapped__ on the partial, it follows the attribute back to the original add(x: int, y: int) and builds a schema requiring both x and y — completely ignoring that y is already bound by the partial to 10.

Without this fix, Tool.from_function(add_ten) would generate an input schema demanding both x and y, and calling the tool with just {"x": 5} would fail validation.

The fix: Reconstructing a fresh functools.partial from the same components (.func, .args, .keywords) produces a functionally identical partial but without the __wrapped__ attribute that update_wrapper injected. Pydantic then correctly introspects the partial's own signature, sees that y is bound with default 10, and generates a schema where only x is required.

The guard hasattr(fn, "__wrapped__") ensures this reconstruction only happens when update_wrapper was actually called — a bare functools.partial(add, y=10) without update_wrapper already works correctly and doesn't need this treatment.

---

Marvin Context Protocol | Type /marvin to interact further

Give us feedback! React with 🚀 if perfect, 👍 if helpful, 👎 if not.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7ef84c0eab

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve partial metadata when stripping wrapped

Reconstructing the functools.partial object here removes attributes that functools.update_wrapper added (notably __annotations__ and module context), and downstream schema generation relies on that metadata for resolving postponed/forward type hints. For wrapped partial tools that use string or forward-referenced annotations, this can degrade or break input/output schema generation during registration. Prefer removing only __wrapped__ (or copying metadata onto the new partial) so bound-argument handling does not discard type information.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Keep resource docstring after partial normalization

This creates a fresh partial and drops the wrapped function metadata copied by update_wrapper, including __doc__; later description=metadata.description or inspect.getdoc(fn) will then use the generic functools.partial doc instead of the resource function's docstring. As a result, resources registered from wrapped partials expose incorrect descriptions unless callers always pass description explicitly.

Useful? React with 👍 / 👎.