`#[tool(description = …)]` accepts only a bare literal, while `#[doc]` and `#[schemars]` on the same items accept `include_str!` and `const` paths

[SergiusDu]

Summary

description in #[tool] accepts only a bare string literal. Two attributes that sit on the very same items are more permissive: #[doc] on the function accepts include_str! and that value does reach the published description, and #[schemars(extend(…))] on the argument struct accepts const paths.

So a server can already source a tool's published text from outside the attribute — through #[doc], or through a literal handed in by a wrapper macro — just not through description, which is the field that names it. Accepting a const path and concat! there would remove the inconsistency.

This is an enhancement request, not a blocker: the workarounds below all work.

Version

rmcp 3.1.2, rustc 1.96.1, edition 2024.

What was measured

form	result
#[tool(description = "…")]	works
#[tool(description = my_crate::caps::TEXT)]	error: Unexpected type \path``
#[tool(description = concat!("a", "b"))]	error: Unexpected type \macro``
/// doc comment + #[tool]	works — becomes the description
#[doc = include_str!("text.md")] + #[tool]	works — becomes the description
#[schemars(extend("maximum" = MOST_STARTS_AT_ONCE))] on the args struct	works — a const path, in the same attribute stack
doc comment and description together	description wins (no surprise here)
macro_rules! wrapping the whole #[tool_router] impl	works
macro_rules! wrapping only the fn inside the impl	compiles, registers nothing — filed separately

The schemars row is the one that makes this feel like an oversight rather than a decision: a const is already acceptable one attribute away, for a value that is likewise baked into the published schema.

Why it matters on a larger surface

On a server with many tools the published text is an interface, and it usefully has two halves: a short one broadcast in tools/list before the first call, and a longer one fetched on demand. Ours measures 82 953 B of description across 66 tools, against a 13 590 B fetchable handbook — so keeping the split is worth real bytes, and it is far easier to keep the two halves honest when both can be declared together and the short one referenced from the attribute:

// what one would like to write
const GET_OR_CREATE_PROJECT: Capability = Capability {
    broadcast: "…what prevents a wrong call…",
    taught:    "…the full explanation, fetched on demand…",
};

#[tool(description = GET_OR_CREATE_PROJECT.broadcast)]

Today that requires either moving the text into a file and using #[doc = include_str!(…)], or keeping a second copy plus a test comparing it to the literal.

Suggested change

Accept in description what neighbouring attributes already accept: a path to a const &'static str, and concat!. The generated code stores a &'static str either way.

If the literal is required because the value is needed during macro expansion, a line saying so in the docs — and an error message reading "only a string literal is accepted here" rather than naming the token type — would save the next person the experiment.

What this report does not establish

It does not claim the current behaviour is wrong, only that it is inconsistent with #[doc] and #[schemars] on the same items. name was not tested against a non-literal; only description was.