Feat/mcp transport

CHANGELOG.md

@@ -9,14 +9,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/2.0.0.
 
 ### Added
 
-- `mcp` feature + `loopctl::mcp` module — adapt any MCP server's tools as loopctl `Tool` implementations. New public types: `McpClient` (a connected client handle), `McpToolProvider` (discovers a server's tools and registers them into a `ToolRegistry`), `McpTool` (one server tool as a `Tool`), `McpError`. The adapter is transport-agnostic; `McpClient::in_process` connects an in-process rmcp server for tests and bundled-server use. Real transports (stdio, HTTP/SSE) arrive in a later release. The optional `rmcp` dependency is pulled in only by the `mcp` feature (`default = []` is unchanged). A runnable end-to-end demo ships at `examples/mcp-adapter.rs` (`cargo run --example mcp-adapter --features mcp`).
+- `McpToolProvider::with_call_timeout(Duration)` — a per-call budget for every adapted tool's `tools/call` round-trip (default 60s). A call that exceeds it resolves to a *soft* error result naming the tool and the budget, so a wedged MCP server costs one tool result instead of hanging the agent loop indefinitely. Updates already-discovered tools and applies to later refreshes. Pinned by `call_timeout_cuts_a_slow_tool_with_a_soft_error` / `default_timeout_lets_a_quick_tool_through` in `tests/mcp_tool_provider.rs`.
+- MCP transports (stdio + Streamable HTTP/SSE) — `McpClient::stdio(command)` spawns an MCP server as a child process over stdio; `McpClient::http_sse(endpoint)` and `McpClient::http_sse_with_client(endpoint, reqwest::Client)` connect via the Streamable HTTP transport (rmcp handles `Mcp-Session-Id`, JSON-vs-SSE response splitting, and DELETE-on-close). `McpClient::reconnect(&StreamRetryConfig)` re-establishes a dropped connection using the crate's existing backoff strategy (the one retry strategy for the crate, not a second one). New public `CommandSpec` describes what to spawn. A stdio example server ships at `examples/mcp-stdio-server.rs` for transport testing. The `mcp` feature now enables `streaming` (for `StreamRetryConfig`) and `reqwest` (for the HTTP client); `default = []` is unchanged.
+- `mcp` feature + `loopctl::mcp` module — adapt any MCP server's tools as loopctl `Tool` implementations. New public types: `McpClient` (a connected client handle), `McpToolProvider` (discovers a server's tools and registers them into a `ToolRegistry`), `McpTool` (one server tool as a `Tool`), `McpError`. Real-server transports ship alongside (stdio child processes and Streamable HTTP/SSE — see the transports entry above); `McpClient::in_process` connects an in-process rmcp server for tests and bundled-server use. The optional `rmcp` dependency is pulled in only by the `mcp` feature (`default = []` is unchanged). A runnable end-to-end demo ships at `examples/mcp-adapter.rs` (`cargo run --example mcp-adapter --features mcp`).
 - `LoopError::ToolRecoveryExhausted { tool, attempts }` — the driver now enforces `MAX_RECOVERY_ATTEMPTS` (5) as a hard ceiling. A recovery strategy that always returns `Retry` is stopped after 5 retries (attempt 6), returning this variant instead of looping forever. Pinned by `recovery_ceiling_stops_retry_forever_strategy`.
 - `RunConfig::memory_top_k` — configurable number of memory entries retrieved and injected per turn (default 3; was a hardcoded magic number).
 - `MachineStep::CallTools { turn, calls }` — the machine now emits the 0-indexed turn number on `CallTools` (matching `CallLLM`), so both handlers source the turn identically from the machine rather than one reading a field and the other querying a counter.
 - `parallel_hard_error_discards_sibling_results` test — pins the documented contract that a hard error in a parallel wave aborts the batch and discards already-completed sibling results.
 
 ### Changed
 
+- Provider HTTP clients now set a **read timeout** (maximum gap between response bytes) instead of a total request timeout, and `with_timeout` configures that read timeout. A total HTTP-layer cap aborted every SSE stream longer than the configured duration (default 2 minutes) — pre-empting the `StreamHandler`'s per-event/total-stream deadlines and the engine's turn timeout, which own generation-length budgets. Long healthy streams now run as long as they keep producing bytes; a server silent for the configured gap (default 120s) is still aborted. Behavior change: generations longer than the old total cap no longer fail at the HTTP layer.
+
 - **Breaking:** Machine turn indices are now 0-indexed (`CallLLM { turn: 0 }` for the first turn). Previously 1-indexed (`turn: 1`). `AwaitingModel { turn }` and `AwaitingTools { turn }` follow the same convention. Callers matching on these variants in tests or drivers must adjust.
 - **Breaking:** `LoopError` gains `ToolRecoveryExhausted` variant. Exhaustive matches on `LoopError` must add this arm.
 - **Breaking:** `RunConfig` gains `memory_top_k` field. Struct-literal construction must add it (use `..Default::default()` or the builders).
@@ -37,6 +41,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/2.0.0.
 
 ### Fixed
 
+- The non-streaming fallback forwards the turn's `RequestOptions` and is bounded by the total-stream deadline (`StreamHandler::fallback_non_streaming`): it previously called `create_message` (dropping any configured `response_format`/`tool_constraint`) and raced only the cancel signal, so a hanging fallback request hung the turn. Pinned by `fallback_non_streaming_forwards_request_options` and `fallback_non_streaming_honors_the_total_deadline`.
+- MCP `tools/call` round-trips are bounded by the new per-call timeout (see `McpToolProvider::with_call_timeout` above) — previously a wedged server hung the agent loop with cancellation as the only exit.
 - Tool-result parts in a `CallTools` turn preserve **model request order** across preresolved (unknown-tool) and dispatched calls. Previously the turn's results were assembled as `[all preresolved, then all dispatched]`, which reordered the parts the model saw relative to the calls it made. Provider-safe in practice (providers match by `tool_call_id`, not position), but order-non-preserving and surprising to hosts that assume positional alignment. Pinned by `test_mixed_known_unknown_tools_preserve_request_order`.
 - The `run()` `Done` arm now matches every `MachineOutcome` variant explicitly (`Completed`, `MaxTurnsExceeded`, `Cancelled`, `Failed`) instead of using a wildcard `other => ... unwrap_or(Cancelled)` fallback. `MachineOutcome` is `#[non_exhaustive]` but defined in this crate, so the compiler proves this exhaustive — a future variant forces a compile error here rather than being silently mislabelled as `Cancelled`.
 - `handle_call_tools` doc corrected to state where cancellation is actually honored (the in-flight tool call is raced against the cancel signal in `execute_tool_call`'s `select!`; the sequential path checks the signal between calls), instead of claiming a `select!` that does not exist in the function itself.

Cargo.toml

@@ -33,7 +33,7 @@ bytes = { version = "1", optional = true }
 async-stream = { version = "0.3", optional = true }
 httpdate = { version = "1", optional = true }
 jsonschema = { version = "0.49", optional = true }
-rmcp = { version = "3", optional = true, default-features = false, features = ["client", "server", "macros", "transport-async-rw"] }
+rmcp = { version = "3", optional = true, default-features = false, features = ["client", "server", "macros", "transport-async-rw", "transport-child-process", "transport-streamable-http-client-reqwest", "transport-io"] }
 
 [dev-dependencies]
 tokio = { version = "1.52.3", features = ["rt-multi-thread", "macros", "sync", "time", "signal"] }
@@ -62,8 +62,10 @@ zai = ["providers", "anthropic"]
 grammar = ["providers"]
 schema_validation = ["dep:jsonschema"]
 
-# MCP client adapter
-mcp = ["dep:rmcp"]
+# MCP client adapter + transports (stdio, Streamable HTTP/SSE). Enables
+# `streaming` for `StreamRetryConfig` (reconnect backoff) and `reqwest` for the
+# HTTP transport's caller-supplied client.
+mcp = ["dep:rmcp", "streaming", "dep:reqwest"]
 
 [[example]]
 name = "hello-cli"
@@ -85,6 +87,10 @@ required-features = ["testing", "providers"]
 name = "mcp-adapter"
 required-features = ["mcp"]
 
+[[example]]
+name = "mcp-stdio-server"
+required-features = ["mcp"]
+
 [lints.clippy]
 pedantic = { level = "warn", priority = -1 }
 unwrap_used = "deny"

README.md

@@ -158,7 +158,7 @@ let agent = BareLoop::new(
 | `zai` | No | `providers`, `anthropic` | Z.AI API client (Anthropic-compatible) |
 | `grammar` | No | `providers` | Tool-call grammar providers for grammar-aware samplers (vLLM `guided_json`); enables the `Grammar` mode of `ToolConstraint` |
 | `schema_validation` | No | — | JSON Schema validation of `Correction::modified_input` in `LlmReflector` (pulls `jsonschema`); when off, validation is skipped |
-| `mcp` | No | `rmcp` | MCP client adapter (`mcp::McpToolProvider`) — adapt an MCP server's tools as loopctl `Tool` impls (in-process; stdio/HTTP/SSE transports land in a later release) |
+| `mcp` | No | `rmcp`, `reqwest`, `async-stream` | MCP client adapter (`mcp::McpToolProvider`) — adapt an MCP server's tools as loopctl `Tool` impls over stdio, Streamable HTTP/SSE, or in-process (`McpClient::stdio`/`http_sse`/`in_process`) |
 
 ### Streaming vs non-streaming
 

examples/mcp-stdio-server.rs

@@ -0,0 +1,58 @@
+//! A minimal MCP server over stdio — the subprocess target for the L-13 stdio
+//! transport tests.
+//!
+//! Built with rmcp's `#[tool_router]` / `#[tool]` macros, exposing one `greet`
+//! tool. Runs the server on stdin/stdout (the MCP stdio transport); stderr is
+//! inherited so server logs surface during tests. Exits when the client
+//! disconnects.
+//!
+//! ```sh
+//! cargo run --example mcp-stdio-server --features mcp
+//! ```
+
+#![allow(
+    clippy::expect_used,
+    clippy::panic,
+    clippy::missing_errors_doc,
+    clippy::missing_panics_doc,
+    dead_code
+)]
+
+use rmcp::ServiceExt;
+use rmcp::handler::server::ServerHandler;
+use rmcp::handler::server::router::tool::ToolRouter;
+use rmcp::{tool, tool_handler, tool_router};
+
+/// A server exposing one `greet` tool.
+#[derive(Clone)]
+struct StdioServer {
+    router: ToolRouter<Self>,
+}
+
+#[tool_router]
+impl StdioServer {
+    fn new() -> Self {
+        Self {
+            router: Self::tool_router(),
+        }
+    }
+
+    #[tool(description = "Return a friendly greeting")]
+    async fn greet(&self) -> String {
+        "hello from stdio".to_string()
+    }
+}
+
+#[tool_handler]
+impl ServerHandler for StdioServer {}
+
+#[tokio::main]
+async fn main() {
+    let server = StdioServer::new();
+    let transport = rmcp::transport::io::stdio();
+    let running = server
+        .serve(transport)
+        .await
+        .expect("stdio server initialize");
+    running.waiting().await.ok();
+}

src/engine/bare/config.rs

@@ -284,16 +284,30 @@ impl<C: ApiClient> BareLoop<C> {
     /// automatically from `self.tools` so that schema generation and dispatch
     /// always share the same underlying registry:
     ///
+    /// The pipeline is also the injection point for per-dispatch host state:
+    /// a middleware may augment `ctx.tool_context` — set extensions, `cwd`,
+    /// `is_non_interactive` — before the tool runs. See
+    /// [`ToolContext`](crate::tool::ToolContext) ("Passing host state to
+    /// tools") and the `host-state` example for the full pattern.
+    ///
     /// # Example
     ///
-    /// ```rust,ignore
-    /// use loopctl::engine::middleware::{ToolPipeline, TimeoutMiddleware};
+    /// ```rust,no_run
+    /// # use std::sync::Arc;
+    /// # use loopctl::config::SessionConfig;
+    /// # use loopctl::engine::BareLoop;
+    /// # use loopctl::testing::MockApiClient;
+    /// # use loopctl::tool::ToolRegistry;
+    /// use loopctl::middleware::{ToolPipeline, TimeoutMiddleware};
     ///
     /// let builder = ToolPipeline::builder()
     ///     .with_middleware(TimeoutMiddleware::from_secs(30));
     ///
-    /// let mut agent = BareLoop::new(client, registry, config);
-    /// agent.set_pipeline(builder)?;
+    /// # let client = MockApiClient::new("demo");
+    /// # let registry = ToolRegistry::new();
+    /// # let config = SessionConfig::default();
+    /// let mut agent = BareLoop::new(Arc::new(client), registry, config);
+    /// agent.set_pipeline(builder).expect("static composition is valid");
     /// ```
     ///
     /// # Errors

src/managers.rs

@@ -357,7 +357,10 @@ impl LoopManagers {
     /// Set the middleware pipeline for tool dispatch.
     ///
     /// Non-consuming variant of [`with_pipeline`](Self::with_pipeline) for
-    /// cases where the managers is already constructed.
+    /// cases where a `LoopManagers` is already constructed. If you are installing
+    /// onto a [`BareLoop`](crate::engine::BareLoop), prefer
+    /// [`BareLoop::set_pipeline`](crate::engine::BareLoop::set_pipeline), which
+    /// shares the loop's tool registry with the pipeline core automatically.
     pub fn set_pipeline(&mut self, pipeline: ToolPipeline) {
         self.tool_pipeline = Some(pipeline);
     }