Rationalize feature flags for v3.0

[DaleSeo]

Labels: T-enhancement, T-config, T-documentation

Problem

rmcp exposes 26 feature flags (including implicit ones). They have grown organically — roughly one flag per new optional dependency — rather than from a user-facing design. Concretely:

• Names mix four unrelated concerns. Role (client, server), protocol capability (elicitation, request-state, auth), transport (transport-io, transport-streamable-http-*), and raw dependency crate names (base64, uuid, schemars, tower, reqwest) all sit in one flat list with no way to tell them apart.
• Internal building blocks look like public flags. client-side-sse, server-side-http, transport-streamable-http-server-session, base64, uuid all appear on docs.rs but should never be enabled directly. Meanwhile __reqwest is correctly hidden — so the convention exists but is applied to exactly one flag.
• Common use cases need 3+ flags, and one of them is a silent trap. A Streamable HTTP client needs client + transport-streamable-http-client-reqwest + reqwest. Omit the last one and it still compiles — reqwest is pulled without any TLS backend, so every https:// request fails at runtime. Our own conformance/Cargo.toml does exactly this; it survives only because it targets http://localhost.
• Protocol capabilities are gated inconsistently. elicitation is opt-in (only to gate the url crate), while sampling, tasks (SEP-2663), and response caching (SEP-2549) are always on. There is no principle a user can infer.
• local violates Cargo's additivity contract. It removes Send bounds crate-wide for WASM/single-threaded use. Because Cargo features are additive, any crate anywhere in the dependency graph enabling local silently changes trait bounds for everyone. The name gives no hint of this, and it is undocumented.

Use cases today

What a user must put in Cargo.toml right now, per scenario:

Use case	Required features today	Pain
Stdio server (most common server)	default (server, macros, base64) + transport-io	transport-io doesn't say "stdio"; spec's term is stdio transport
Stdio client (spawns a server)	client, transport-child-process, default-features = false	default drags in server + schemars + pastey the client never uses
Streamable HTTP server	server, macros, transport-streamable-http-server	OK; tower integration arrives via an internal flag
Streamable HTTP client	client, transport-streamable-http-client-reqwest, reqwest	Trap: omit reqwest → compiles, but all TLS fails at runtime
HTTP client, non-default TLS	above but reqwest-native-tls / reqwest-tls-no-provider	reqwest actually means "reqwest + rustls"; segment order inconsistent (reqwest-native-tls vs reqwest-tls-no-provider)
HTTP client, custom HTTP stack	client, transport-streamable-http-client	fine
HTTP client over Unix socket	client, transport-streamable-http-client-unix-socket	six-segment name; undocumented in README
OAuth client	+ auth (+ auth-client-credentials-jwt)	auth pulls reqwest via __reqwest — same silent no-TLS trap
Server using elicitation	+ elicitation	the only protocol capability behind a flag; exists only to gate url
Sealed request state (SEP-2322)	+ request-state	fine (real crypto deps), but undocumented in README
WASM / single-threaded	local (+ transport-worker)	non-additive; opaque name; undocumented

Dead remnant: transport-ws exists only as a comment in Cargo.toml, yet src/transport/ws.rs and a commented #[cfg] in transport.rs remain.

Proposal

A naming taxonomy — one rule per axis

Axis	Convention	Examples
Role	bare word	client, server
Protocol capability	spec terminology, kebab-case	request-state, auth
Transport	transport- prefix, spec terminology	transport-stdio, transport-streamable-http-server
Ecosystem integration	crate name (established Rust convention)	tower
TLS backend	tls- prefix	tls-rustls, tls-native, tls-no-provider
Use-case bundle	<role>-<transport>	client-http, server-stdio
Internal building block	__ prefix, hidden from docs	__server-side-http

1. Renames (v3.0, with one-release aliases)

Today	Proposed	Rationale
transport-io	transport-stdio	match spec vocabulary; users search "stdio"
reqwest	tls-rustls	the flag's real meaning is the TLS choice, not the HTTP crate
reqwest-native-tls	tls-native	consistent tls- prefix + order
reqwest-tls-no-provider	tls-no-provider	same
local	unsync (or replace with cfg(target_family = "wasm"))	name should say what it does; ideally not a feature at all, since it is non-additive
which-command	fold into transport-child-process (or command-path-lookup)	named after the which crate, not the behavior

Old names stay as empty forwarding aliases (transport-io = ["transport-stdio"]) for one major cycle, then are removed.

2. Un-gate capabilities that only exist to hide a trivial dep (v3.0)

Today	Proposed	Rationale
elicitation	always on	gates a core capability while nothing else (sampling/tasks/caching) is gated. Correction: url is not a trivial dep — making it mandatory adds 29 crates to a minimal build via idna → ICU4X (37 → 66). Un-gating therefore also requires the elicitation API to take URLs as plain strings, so url stays behind auth. New rule: gate a capability only when it carries a heavy dep — request-state (hmac/sha2) and auth (oauth2) stay opt-in
base64	always on (already in defaults)	implicit dep-named flag gating three image-content helpers
uuid	hidden via dep:uuid	implicit dep-named flag; pure implementation detail

3. Hide internal building blocks (v3.0)

client-side-sse, server-side-http, transport-streamable-http-server-session get the __ prefix (matching __reqwest) and leave the docs.rs feature list. Users compose transports; they don't assemble the parts.

transport-worker was originally listed here too, but it stays public: crates/rmcp/README.md and the transport module docs both present implementing Worker as one of the ways to build a transport, so hiding the flag would make a documented extension point unreachable.

4. Use-case meta-features (non-breaking, can ship now)

server-stdio = ["server", "macros", "transport-stdio"]
server-http  = ["server", "macros", "transport-streamable-http-server"]
client-stdio = ["client", "transport-child-process"]
client-http  = ["client", "transport-streamable-http-client-reqwest", "tls-rustls"]

Getting-started docs become one flag per scenario. client-http bakes in rustls; the raw no-TLS combination stays legal (plain-http:// deployments need it) but becomes an opt-in, not a pitfall.

5. Housekeeping (non-breaking, can ship now)

• README feature table documents ~13 of 26 flags; document all public ones (missing: request-state, auth-client-credentials-jwt, which-command, tower, local, unix-socket transport).
• Delete src/transport/ws.rs and the commented transport-ws remnants (recoverable from git if WebSocket ever lands).
• Document the TLS pitfall until the meta-features land.

Use cases after this proposal

Use case	Features after
Stdio server	server-stdio (or defaults + transport-stdio)
Stdio client	client-stdio, default-features = false
Streamable HTTP server	server-http
Streamable HTTP client	client-http
HTTP client, native TLS	client, transport-streamable-http-client-reqwest, tls-native
HTTP client, custom stack	client, transport-streamable-http-client
OAuth client	client-http, auth
Server using elicitation	nothing extra — always available
Sealed request state	+ request-state
WASM / single-threaded	unsync (or automatic via target cfg)

Rollout

Ship now (minor, non-breaking):

• Use-case meta-features (§4)
• README coverage + TLS-pitfall docs (§5)
• Delete ws.rs / transport-ws remnants (§5)

v3.0 (breaking):

• Renames with forwarding aliases (§1)
• Un-gate elicitation / base64, hide uuid (§2)
• __-prefix internal flags (§3)

v4.0:

• Drop the forwarding aliases

Open questions

1. local → renamed feature vs. cfg(target_family = "wasm")? Does anyone use local on native targets (e.g. LocalSet-based single-threaded servers)? If yes, a renamed feature stays; if wasm-only, target cfg is strictly better — it removes the additivity violation entirely.
2. Should defaults shrink? Shrinking away from server/macros is a bigger break for the majority (server authors). This proposal leaves defaults unchanged.
3. tls-* vs keeping reqwest-tls-*? If a second HTTP backend ever lands, tls-* generalizes better; reqwest-tls-* is more honest today. Proposal prefers tls-*.

[DaleSeo]

Implementation is up on feat/rationalize-feature-flags; PR to follow. §1–§5 are all in. Two things came out differently from the proposal above, and I have edited the body to match.

1. transport-worker stays public

§3 listed it among the internals to hide. It should not be: crates/rmcp/README.md and the transport module docs both present implementing Worker as one of the ways to build a transport, so hiding the flag would make a documented extension point unreachable. The other three internals are __-prefixed as proposed.

2. "url is a trivial dep" was wrong

Un-gating elicitation makes url mandatory, and url pulls idna → ICU4X. Measured against main with an identical Cargo.lock:

build	before	with url mandatory
--no-default-features	37 crates	66
default	54 crates	83

The fix keeps §2's promise — elicitation needs no flag — without the weight. elicit_url and elicit_url_with_timeout now take impl Into<String> rather than impl Into<Url>; the wire model already stored a String and the value was stringified immediately, so url can stay behind auth. Callers passing a Url still compile unchanged via impl From<Url> for String, and &str/String are newly accepted. A minimal build now adds only base64, a leaf crate with no transitive deps.

Worth reviewing on its own terms: this is an API change rather than a feature-flag change, and rmcp no longer validates the URL.

Two guards added

Neither --all-features nor the aggregate feature list catches these, so both are new:

• A feature-matrix CI job that builds every public feature on its own. This catches a missing dep: entry that a sibling feature would otherwise supply — invisible when everything is enabled at once. All 33 public features currently build standalone.
• tests/test_feature_aliases.rs, static assertions that each deprecated alias still forwards. Building everything at once cannot catch transport-io = [], because transport-stdio is enabled independently anyway. Verified by injecting exactly that typo: the build fails with evaluation panicked: `transport-io` must forward to `transport-stdio` .

A coordination note for the review window

While the branch was open, #1068 merged code naming the pre-rename local flag: the ClientHandler bounds in crates/rmcp/src/handler/client.rs and a test_local_client_handler target. Rebasing produced no conflict there — git merges text, not renames — and the result compiles either way, because local survives as an alias and Send + Sync is merely the stricter bound. It would have surfaced only in downstream code that enables the canonical unsync and implements a !Send handler.

So while this is in review: new feature gates should use the v3 names, and rebasing this branch needs a grep for the old ones rather than trusting a clean merge.

This does not close the issue

Two items remain:

• v4.0: drop the forwarding aliases
• Open question 1: unsync as a feature vs cfg(target_family = "wasm")

On the second — the rename made the name honest but left the mechanism intact. Because Cargo unifies features across the dependency graph, any crate anywhere enabling unsync drops the Send bounds for every other rmcp user in that build, and also removes StreamableHttpService from the public API. Neither is something an affected crate can opt out of. A target cfg would eliminate the violation outright, but only if nobody relies on unsync for single-threaded native targets, which is a call about real usage rather than about the code.

Open questions 2 and 3 are implemented as proposed: defaults unchanged, tls-* naming adopted.