fix: bound MCP stdio requests and shutdown across stdin waits

crates/mcp/src/stdio.rs

@@ -11,11 +11,11 @@ use secrets::SecretStore;
 use serde::{Deserialize, Serialize};
 use serde_json::{Value, json};
 use thiserror::Error;
-use tokio::io::{AsyncRead, AsyncReadExt, AsyncWriteExt, BufReader};
+use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt, BufReader};
-use tokio::process::{Child, ChildStderr, ChildStdin, ChildStdout, Command};
+use tokio::process::{Child, ChildStderr, ChildStdout, Command};
 use tokio::sync::{Mutex, mpsc};
 use tokio::task::JoinHandle;
-use tokio::time::{Instant, timeout};
+use tokio::time::{Instant, timeout, timeout_at};
 
 const MCP_PROTOCOL_VERSION: &str = "2025-11-25";
 const CLIENT_NAME: &str = "yoi";
@@ -31,8 +31,11 @@ pub struct McpStdioLimits {
     pub max_diagnostic_lines: usize,
     pub max_protocol_bytes: usize,
     pub startup_timeout: Duration,
+    /// One budget for stdin ownership, frame write/flush, and response wait.
     pub request_timeout: Duration,
+    /// Shared budget for stdin ownership/close and graceful process exit.
     pub shutdown_timeout: Duration,
+    /// Budget for each fallback wait: terminate, then forced kill/reap.
     pub kill_timeout: Duration,
 }
 
@@ -570,6 +573,30 @@ pub fn resolve_stdio_server(
     Ok(spec)
 }
 
+// The mutex orders entire frames, including replies emitted by the stdout reader.
+type SharedStdin = Arc<Mutex<Option<Box<dyn AsyncWrite + Send + Unpin>>>>;
+
+/// Last boundary reached by a failed request. A flushed frame does not prove
+/// that the server executed the method; callers must not blindly retry it.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum McpRequestStage {
+    WaitingForStdin,
+    WritingFrame,
+    FlushingFrame,
+    AwaitingResponse,
+}
+
+impl fmt::Display for McpRequestStage {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.write_str(match self {
+            Self::WaitingForStdin => "stdin ownership (method not sent)",
+            Self::WritingFrame => "frame write (delivery unknown; stdin closed)",
+            Self::FlushingFrame => "frame flush (delivery unknown; stdin closed)",
+            Self::AwaitingResponse => "response (frame flushed; execution unknown; do not retry)",
+        })
+    }
+}
+
 /// A running initialized stdio MCP client.
 pub struct McpStdioClient {
     server_name: String,
@@ -577,7 +604,7 @@ pub struct McpStdioClient {
     redactor: Redactor,
     diagnostics: Arc<Mutex<BoundedDiagnostics>>,
     list_changes: Arc<Mutex<BoundedListChanged>>,
-    stdin: Arc<Mutex<Option<ChildStdin>>>,
+    stdin: SharedStdin,
     child: Option<Child>,
     responses: mpsc::Receiver<ReaderEvent>,
     reader_task: JoinHandle<()>,
@@ -663,7 +690,7 @@ impl McpStdioClient {
             )
         })?;
 
-        let stdin = Arc::new(Mutex::new(Some(stdin)));
+        let stdin: SharedStdin = Arc::new(Mutex::new(Some(Box::new(stdin))));
         let diagnostics = Arc::new(Mutex::new(BoundedDiagnostics::new(
             spec.name.clone(),
             limits.max_diagnostic_lines,
@@ -747,6 +774,12 @@ impl McpStdioClient {
     /// The caller is responsible for applying Yoi tool permissions before this
     /// method is reached and for bounding/serializing the untrusted result before
     /// it is exposed to model-visible tool history.
+    ///
+    /// Timeout or cancellation after writing begins can leave execution unknown;
+    /// this method never retries. Interrupted write/flush closes stdin, and the
+    /// caller must shut down the process rather than reuse that connection. A
+    /// fully flushed frame permits later *different* requests with distinct IDs,
+    /// but is not permission to replay this call when its response is missing.
     pub async fn call_tool(
         &mut self,
         request: CallToolRequest,
@@ -902,23 +935,35 @@ impl McpStdioClient {
             method,
             params,
         };
-        self.write_protocol(phase, &request).await?;
+        if self.shutdown_started {
-        let response = match timeout(
+            return Err(McpClientError::new(
-            self.limits.request_timeout,
+                &self.server_name,
-            self.wait_for_response(id, phase),
+                phase,
-        )
+                McpErrorKind::Io("client shutdown has started".into()),
+            ));
+        }
+        let deadline = Instant::now() + self.limits.request_timeout;
+        let mut stage = McpRequestStage::WaitingForStdin;
+        // One budget covers ownership, write, flush, and the matching response.
+        let response = match timeout_at(deadline, async {
+            self.write_protocol(phase, &request, &mut stage).await?;
+            stage = McpRequestStage::AwaitingResponse;
+            self.wait_for_response(id, phase).await
+        })
         .await
         {
-            Ok(result) => result?,
+            Ok(Ok(response)) => response,
+            Ok(Err(err)) => return Err(err.with_request_stage(stage)),
             Err(_) => {
                 return Err(McpClientError::new(
                     &self.server_name,
                     phase,
                     McpErrorKind::Timeout {
-                        operation: method.to_owned(),
+                        operation: self.redactor.redact(method),
                         elapsed: self.limits.request_timeout,
                     },
                 )
+                .with_request_stage(stage)
                 .with_diagnostics(self.snapshot_diagnostics().await));
             }
         };
@@ -983,6 +1028,7 @@ impl McpStdioClient {
         method: &str,
         params: T,
     ) -> Result<(), McpClientError> {
+        let mut stage = McpRequestStage::WaitingForStdin;
         self.write_protocol(
             phase,
             &ClientNotification {
@@ -990,6 +1036,7 @@ impl McpStdioClient {
                 method,
                 params,
             },
+            &mut stage,
         )
         .await
     }
@@ -998,6 +1045,7 @@ impl McpStdioClient {
         &mut self,
         phase: McpPhase,
         value: &T,
+        stage: &mut McpRequestStage,
     ) -> Result<(), McpClientError> {
         write_json_line(
             &self.server_name,
@@ -1006,6 +1054,7 @@ impl McpStdioClient {
             value,
             self.limits.max_protocol_bytes,
             &self.redactor,
+            stage,
         )
         .await
     }
@@ -1013,15 +1062,23 @@ impl McpStdioClient {
     /// Close stdin and wait for process exit, falling back to terminate and kill.
     pub async fn shutdown(&mut self) -> Result<ShutdownReport, McpClientError> {
         self.shutdown_started = true;
-        {
+        let deadline = Instant::now() + self.limits.shutdown_timeout;
-            let mut stdin = self.stdin.lock().await;
+        // Ownership and EOF share one graceful budget. A reader reply may own
-            stdin.take();
+        // stdin indefinitely; its lock is never required to terminate/kill.
-        }
+        let closed = timeout_at(deadline, async {
+            self.stdin.lock().await.take();
+        })
+        .await
+        .is_ok();
 
         let mut child = match self.child.take() {
             Some(child) => child,
             None => return Ok(ShutdownReport::already_finished()),
         };
+        if !closed {
+            self.reader_task.abort();
+            return self.terminate_then_kill(child).await;
+        }
 
         if let Ok(Some(status)) = child.try_wait() {
             self.reader_task.abort();
@@ -1033,7 +1090,7 @@ impl McpStdioClient {
             });
         }
 
-        match timeout(self.limits.shutdown_timeout, child.wait()).await {
+        match timeout_at(deadline, child.wait()).await {
             Ok(Ok(status)) => {
                 self.reader_task.abort();
                 self.stderr_task.abort();
@@ -1195,12 +1252,13 @@ pub struct McpDiagnostics {
 }
 
 #[derive(Debug, Error, Clone)]
-#[error("MCP stdio server `{server_name}` failed during phase `{phase}`: {kind}")]
+#[error("MCP stdio server `{server_name}` failed during phase `{phase}`: {kind}{stage_detail}", stage_detail = request_stage.map(|stage| format!("; {stage}")).unwrap_or_default())]
 pub struct McpClientError {
     pub server_name: String,
     pub phase: McpPhase,
     pub kind: McpErrorKind,
     diagnostics: Option<McpDiagnostics>,
+    request_stage: Option<McpRequestStage>,
 }
 
 impl McpClientError {
@@ -1210,6 +1268,7 @@ impl McpClientError {
             phase,
             kind,
             diagnostics: None,
+            request_stage: None,
         }
     }
 
@@ -1223,6 +1282,15 @@ impl McpClientError {
         self
     }
 
+    fn with_request_stage(mut self, stage: McpRequestStage) -> Self {
+        self.request_stage = Some(stage);
+        self
+    }
+
+    pub fn request_stage(&self) -> Option<McpRequestStage> {
+        self.request_stage
+    }
+
     pub fn diagnostics(&self) -> Option<&McpDiagnostics> {
         self.diagnostics.as_ref()
     }
@@ -1311,7 +1379,7 @@ struct ErrorObject<'a> {
 fn spawn_stdout_reader(
     server_name: String,
     stdout: ChildStdout,
-    stdin: Arc<Mutex<Option<ChildStdin>>>,
+    stdin: SharedStdin,
     tx: mpsc::Sender<ReaderEvent>,
     limits: McpStdioLimits,
     redactor: Redactor,
@@ -1368,7 +1436,7 @@ fn spawn_stdout_reader(
 
 async fn handle_incoming_message(
     server_name: &str,
-    stdin: &Arc<Mutex<Option<ChildStdin>>>,
+    stdin: &SharedStdin,
     tx: &mpsc::Sender<ReaderEvent>,
     limits: &McpStdioLimits,
     redactor: &Redactor,
@@ -1392,6 +1460,7 @@ async fn handle_incoming_message(
                 &response,
                 limits.max_protocol_bytes,
                 redactor,
+                &mut McpRequestStage::WaitingForStdin,
             )
             .await;
         }
@@ -1484,10 +1553,11 @@ fn spawn_stderr_reader(
 async fn write_json_line<T: Serialize>(
     server_name: &str,
     phase: McpPhase,
-    stdin: &Arc<Mutex<Option<ChildStdin>>>,
+    stdin: &SharedStdin,
     value: &T,
     max_protocol_bytes: usize,
     redactor: &Redactor,
+    stage: &mut McpRequestStage,
 ) -> Result<(), McpClientError> {
     let mut bytes = serde_json::to_vec(value).map_err(|err| {
         McpClientError::new(
@@ -1507,27 +1577,34 @@ async fn write_json_line<T: Serialize>(
     }
     bytes.push(b'\n');
     let mut guard = stdin.lock().await;
-    let Some(stdin) = guard.as_mut() else {
+    let Some(mut writer) = guard.take() else {
         return Err(McpClientError::new(
             server_name,
             phase,
             McpErrorKind::Io("child stdin is closed".into()),
         ));
     };
-    stdin.write_all(&bytes).await.map_err(|err| {
+    // Keep ownership out of the slot across both awaits. Timeout/cancel or
+    // I/O error drops the writer, closing the connection instead of allowing
+    // the next frame to append to an unverified partial JSON line.
+    *stage = McpRequestStage::WritingFrame;
+    writer.write_all(&bytes).await.map_err(|err| {
         McpClientError::new(
             server_name,
             phase,
             McpErrorKind::Io(redactor.redact(&err.to_string())),
         )
     })?;
-    stdin.flush().await.map_err(|err| {
+    *stage = McpRequestStage::FlushingFrame;
+    writer.flush().await.map_err(|err| {
         McpClientError::new(
             server_name,
             phase,
             McpErrorKind::Io(redactor.redact(&err.to_string())),
         )
-    })
+    })?;
+    *guard = Some(writer);
+    Ok(())
 }
 
 async fn read_protocol_line<R: AsyncRead + Unpin>(
@@ -1684,3 +1761,6 @@ fn send_terminate(child: &mut Child) -> Result<(), ()> {
 fn send_terminate(child: &mut Child) -> Result<(), ()> {
     child.start_kill().map_err(|_| ())
 }
+
+#[cfg(test)]
+mod deadline_tests;

crates/mcp/src/stdio/deadline_tests.rs

@@ -0,0 +1,362 @@
+//! Module-boundary fixtures exercise the same writer/request path as ChildStdin.
+use super::*;
+use std::future::{Future, poll_fn};
+use std::pin::Pin;
+use std::sync::Mutex as StdMutex;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::task::{Context, Poll};
+
+#[derive(Clone, Copy)]
+enum Stall {
+    PartialWrite,
+    Flush,
+    Response,
+}
+
+struct ControlledWriter {
+    stall: Stall,
+    bytes: Arc<StdMutex<Vec<u8>>>,
+    reached: Option<mpsc::UnboundedSender<()>>,
+    dropped: Arc<AtomicBool>,
+}
+
+impl ControlledWriter {
+    fn signal(&mut self) {
+        if let Some(tx) = self.reached.take() {
+            tx.send(()).unwrap();
+        }
+    }
+}
+
+impl AsyncWrite for ControlledWriter {
+    fn poll_write(
+        mut self: Pin<&mut Self>,
+        _: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<std::io::Result<usize>> {
+        if matches!(self.stall, Stall::PartialWrite) {
+            if self.bytes.lock().unwrap().is_empty() {
+                self.bytes.lock().unwrap().extend_from_slice(&buf[..8]);
+                return Poll::Ready(Ok(8));
+            }
+            self.signal();
+            return Poll::Pending;
+        }
+        self.bytes.lock().unwrap().extend_from_slice(buf);
+        Poll::Ready(Ok(buf.len()))
+    }
+
+    fn poll_flush(mut self: Pin<&mut Self>, _: &mut Context<'_>) -> Poll<std::io::Result<()>> {
+        self.signal();
+        if matches!(self.stall, Stall::Flush) {
+            Poll::Pending
+        } else {
+            Poll::Ready(Ok(()))
+        }
+    }
+
+    fn poll_shutdown(self: Pin<&mut Self>, _: &mut Context<'_>) -> Poll<std::io::Result<()>> {
+        Poll::Ready(Ok(()))
+    }
+}
+
+impl Drop for ControlledWriter {
+    fn drop(&mut self) {
+        self.dropped.store(true, Ordering::SeqCst);
+    }
+}
+
+fn fixture_client(stdin: SharedStdin) -> (McpStdioClient, mpsc::Sender<ReaderEvent>) {
+    let (tx, responses) = mpsc::channel(16);
+    let redactor = Redactor::new(vec!["fixture-secret".into()]);
+    let client = McpStdioClient {
+        server_name: "deadline-fixture".into(),
+        limits: McpStdioLimits {
+            request_timeout: Duration::from_millis(100),
+            ..Default::default()
+        },
+        diagnostics: Arc::new(Mutex::new(BoundedDiagnostics::new(
+            "deadline-fixture".into(),
+            2,
+            redactor.clone(),
+        ))),
+        redactor,
+        list_changes: Arc::new(Mutex::new(BoundedListChanged::new(
+            "deadline-fixture".into(),
+        ))),
+        stdin,
+        child: None,
+        responses,
+        reader_task: tokio::spawn(std::future::pending()),
+        stderr_task: tokio::spawn(std::future::pending()),
+        next_id: 1,
+        initialized: None,
+        shutdown_started: false,
+    };
+    (client, tx)
+}
+
+async fn assert_pending<F: Future>(mut future: Pin<&mut F>) {
+    poll_fn(|cx| match future.as_mut().poll(cx) {
+        Poll::Pending => Poll::Ready(()),
+        Poll::Ready(_) => panic!("fixture boundary unexpectedly completed"),
+    })
+    .await;
+}
+
+#[tokio::test]
+async fn stdin_ownership_timeout_and_cancel_do_not_send_a_method() {
+    for cancel in [false, true] {
+        let bytes = Arc::new(StdMutex::new(Vec::new()));
+        let dropped = Arc::new(AtomicBool::new(false));
+        let stdin: SharedStdin = Arc::new(Mutex::new(Some(Box::new(ControlledWriter {
+            stall: Stall::Response,
+            bytes: bytes.clone(),
+            reached: None,
+            dropped: dropped.clone(),
+        }))));
+        let guard = stdin.lock().await;
+        let (mut client, _tx) = fixture_client(stdin.clone());
+        let mut request =
+            Box::pin(client.request::<Value>(McpPhase::Running, "tools/call", json!({})));
+        assert_pending(request.as_mut()).await;
+        if !cancel {
+            let err = timeout(Duration::from_secs(1), &mut request)
+                .await
+                .unwrap()
+                .unwrap_err();
+            assert!(matches!(err.kind, McpErrorKind::Timeout { .. }));
+            assert_eq!(err.request_stage(), Some(McpRequestStage::WaitingForStdin));
+            assert!(err.to_string().contains("method not sent"));
+        }
+        drop(request);
+        assert!(bytes.lock().unwrap().is_empty());
+        assert!(!dropped.load(Ordering::SeqCst));
+        assert!(
+            guard.is_some(),
+            "cancel before ownership must preserve the stream"
+        );
+    }
+}
+
+#[tokio::test]
+async fn interrupted_write_or_flush_closes_stdin_before_the_next_frame() {
+    for (stall, stage) in [
+        (Stall::PartialWrite, McpRequestStage::WritingFrame),
+        (Stall::Flush, McpRequestStage::FlushingFrame),
+    ] {
+        for cancel in [false, true] {
+            let bytes = Arc::new(StdMutex::new(Vec::new()));
+            let dropped = Arc::new(AtomicBool::new(false));
+            let (reached, mut rx) = mpsc::unbounded_channel();
+            let stdin: SharedStdin = Arc::new(Mutex::new(Some(Box::new(ControlledWriter {
+                stall,
+                bytes: bytes.clone(),
+                reached: Some(reached),
+                dropped: dropped.clone(),
+            }))));
+            let (mut client, _tx) = fixture_client(stdin.clone());
+            let mut request = Box::pin(client.request::<Value>(
+                McpPhase::Running,
+                "tools/call",
+                json!({"secret": "fixture-secret"}),
+            ));
+            // Poll all the way to the controlled Pending, then observe its channel.
+            assert_pending(request.as_mut()).await;
+            timeout(Duration::from_secs(1), rx.recv())
+                .await
+                .unwrap()
+                .unwrap();
+            if !cancel {
+                let err = timeout(Duration::from_secs(1), &mut request)
+                    .await
+                    .unwrap()
+                    .unwrap_err();
+                assert!(matches!(err.kind, McpErrorKind::Timeout { .. }));
+                assert_eq!(err.request_stage(), Some(stage));
+                let message = err.to_string();
+                assert!(message.contains("delivery unknown; stdin closed"));
+                assert!(!message.contains("fixture-secret"));
+            }
+            drop(request);
+            assert!(dropped.load(Ordering::SeqCst));
+            assert!(stdin.lock().await.is_none());
+            let before = bytes.lock().unwrap().clone();
+            if matches!(stall, Stall::PartialWrite) {
+                assert!(!before.contains(&b'\n'));
+            } else {
+                assert_eq!(before.last(), Some(&b'\n'));
+            }
+            let err = client
+                .request::<Value>(McpPhase::Running, "next-method", json!({}))
+                .await
+                .unwrap_err();
+            assert!(err.to_string().contains("stdin is closed"));
+            assert_eq!(
+                *bytes.lock().unwrap(),
+                before,
+                "no next frame can append after interruption"
+            );
+        }
+    }
+}
+
+#[tokio::test]
+async fn flushed_call_timeout_or_cancel_is_not_replayed_and_late_response_is_not_mixed() {
+    for cancel in [false, true] {
+        let bytes = Arc::new(StdMutex::new(Vec::new()));
+        let (reached, mut rx) = mpsc::unbounded_channel();
+        let stdin: SharedStdin = Arc::new(Mutex::new(Some(Box::new(ControlledWriter {
+            stall: Stall::Response,
+            bytes: bytes.clone(),
+            reached: Some(reached),
+            dropped: Arc::new(AtomicBool::new(false)),
+        }))));
+        let (mut client, tx) = fixture_client(stdin);
+        let mut request =
+            Box::pin(client.request::<Value>(McpPhase::Running, "tools/call", json!({})));
+        assert_pending(request.as_mut()).await;
+        timeout(Duration::from_secs(1), rx.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        if !cancel {
+            let err = timeout(Duration::from_secs(1), &mut request)
+                .await
+                .unwrap()
+                .unwrap_err();
+            assert_eq!(err.request_stage(), Some(McpRequestStage::AwaitingResponse));
+            assert!(err.to_string().contains("execution unknown; do not retry"));
+        }
+        drop(request);
+        // Reuse is safe only because the first frame completed and IDs are
+        // matched. This is a different method, not a replay of the unknown call.
+        tx.send(ReaderEvent::Response(ServerResponse {
+            id: 1,
+            result: Some(json!({"late": true})),
+            error: None,
+        }))
+        .await
+        .unwrap();
+        tx.send(ReaderEvent::Response(ServerResponse {
+            id: 2,
+            result: Some(json!({"next": true})),
+            error: None,
+        }))
+        .await
+        .unwrap();
+        let result = client
+            .request::<Value>(McpPhase::Running, "next-method", json!({}))
+            .await
+            .unwrap();
+        assert_eq!(result, json!({"next": true}));
+        let captured = bytes.lock().unwrap().clone();
+        let frames: Vec<Value> = captured
+            .split(|byte| *byte == b'\n')
+            .filter(|line| !line.is_empty())
+            .map(|line| serde_json::from_slice(line).unwrap())
+            .collect();
+        assert_eq!(frames.len(), 2);
+        assert_eq!(frames[0]["method"], "tools/call");
+        assert_eq!(frames[1]["method"], "next-method");
+    }
+}
+
+// Observes actual OS-pipe backpressure, not a guessed delay or just a mock
+// Pending. It forwards every operation to ChildStdin without changing framing.
+struct PendingObserver<W> {
+    writer: W,
+    accepted: usize,
+    reached: Option<mpsc::UnboundedSender<()>>,
+}
+impl<W: AsyncWrite + Unpin> AsyncWrite for PendingObserver<W> {
+    fn poll_write(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<std::io::Result<usize>> {
+        let result = Pin::new(&mut self.writer).poll_write(cx, buf);
+        if let Poll::Ready(Ok(count)) = &result {
+            self.accepted += count;
+        }
+        if result.is_pending() && self.accepted > 1024 {
+            if let Some(tx) = self.reached.take() {
+                tx.send(()).unwrap();
+            }
+        }
+        result
+    }
+    fn poll_flush(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
+        Pin::new(&mut self.writer).poll_flush(cx)
+    }
+    fn poll_shutdown(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
+        Pin::new(&mut self.writer).poll_shutdown(cx)
+    }
+}
+
+#[cfg(unix)]
+#[tokio::test]
+async fn blocked_reader_reply_does_not_prevent_shutdown_process_reaping() {
+    // A local stdio fixture: initialize, consume initialized, emit an unsupported
+    // server request whose echoed ID exceeds pipe capacity, then never read.
+    // The shell has no descendants and ignores TERM to also prove kill fallback.
+    let init = json!({"jsonrpc":"2.0", "id":1, "result": {"protocolVersion": MCP_PROTOCOL_VERSION, "capabilities":{}, "serverInfo":{"name":"fixture", "version":"1"}}});
+    let server_request =
+        json!({"jsonrpc":"2.0", "id":"x".repeat(256 * 1024), "method":"sampling/createMessage"});
+    // Avoid argv limits: generate the large protocol ID in the shell itself.
+    let script = format!(
+        "trap '' TERM; read -r init; printf '%s\\n' '{init}'; read -r initialized; printf '%s' '{{\"jsonrpc\":\"2.0\",\"method\":\"sampling/createMessage\",\"id\":\"'; i=0; while [ $i -lt 4096 ]; do printf '%s' '{}'; i=$((i+1)); done; printf '%s\\n' '\"}}'; while :; do :; done",
+        "x".repeat(64)
+    );
+    assert!(server_request.to_string().len() < McpStdioLimits::default().max_stdout_line_bytes);
+    let limits = McpStdioLimits {
+        shutdown_timeout: Duration::from_millis(50),
+        kill_timeout: Duration::from_millis(100),
+        ..Default::default()
+    };
+    let mut client = McpStdioClient::spawn(
+        McpStdioServerSpec::new("reader-blocked", "/bin/sh")
+            .arg("-c")
+            .arg(script),
+        limits,
+    )
+    .await
+    .unwrap();
+    let pid = client.child.as_ref().unwrap().id().unwrap();
+    let (reached, mut rx) = mpsc::unbounded_channel();
+    {
+        let mut guard = client.stdin.lock().await;
+        let writer = guard.take().unwrap();
+        *guard = Some(Box::new(PendingObserver {
+            writer,
+            accepted: 0,
+            reached: Some(reached),
+        }));
+    }
+    client.initialize().await.unwrap();
+    timeout(Duration::from_secs(3), rx.recv())
+        .await
+        .expect("reader reply reached OS pipe Pending")
+        .unwrap();
+    assert!(
+        client.stdin.try_lock().is_err(),
+        "reader owns stdin while its reply is blocked"
+    );
+    let report = timeout(Duration::from_secs(1), client.shutdown())
+        .await
+        .unwrap()
+        .unwrap();
+    assert!(report.terminated && report.killed);
+    assert!(report.exit_status.is_some());
+    assert!(client.child.is_none());
+    // Tokio wait already reaped the child; an additional wait cannot find it.
+    let mut status = 0;
+    assert_eq!(
+        unsafe { libc::waitpid(pid as libc::pid_t, &mut status, libc::WNOHANG) },
+        -1
+    );
+    assert_eq!(
+        std::io::Error::last_os_error().raw_os_error(),
+        Some(libc::ECHILD)
+    );
+}

crates/mcp/tests/fixtures/mock_server.rs

@@ -6,8 +6,11 @@ use std::time::Duration;
 use serde_json::{Value, json};
 
 fn main() {
-    let mode = env::var("YOI_MCP_MOCK_MODE").unwrap_or_else(|_| "success".to_string());
+    let mode = env::args()
+        .nth(1)
+        .unwrap_or_else(|| env::var("YOI_MCP_MOCK_MODE").unwrap_or_else(|_| "success".to_string()));
     match mode.as_str() {
+        "stdin-unread" => stdin_unread(),
         "success" => success(),
         "tools" => tools_list(),
         "tools-call-normal" => tools_call_normal(),
@@ -22,6 +25,20 @@ fn main() {
     }
 }
 
+fn stdin_unread() {
+    let init = read_json();
+    assert_eq!(init["method"], "initialize");
+    write_json(json!({"jsonrpc": "2.0", "id": init["id"], "result": initialize_result()}));
+    assert_eq!(read_json()["method"], "notifications/initialized");
+    // Protocol barrier: the host only writes its oversized call after this ack.
+    let ready = read_json();
+    assert_eq!(ready["method"], "fixture/ready");
+    write_json(json!({"jsonrpc": "2.0", "id": ready["id"], "result": {"ready": true}}));
+    loop {
+        thread::park();
+    }
+}
+
 fn success() {
     let init = read_json();
     assert_eq!(init["method"], "initialize");

crates/mcp/tests/stdio_lifecycle.rs

@@ -278,3 +278,48 @@ async fn sampling_requests_fail_closed_and_are_not_advertised() {
     let shutdown = client.shutdown().await.expect("shutdown succeeds");
     assert!(shutdown.exit_status.is_some_and(|status| status.success()));
 }
+
+#[tokio::test]
+async fn unread_stdin_request_times_out_during_write_and_closes_before_reuse() {
+    let limits = McpStdioLimits {
+        request_timeout: Duration::from_millis(100),
+        ..tight_limits()
+    };
+    let spec = McpStdioServerSpec::new("unread-stdin", env!("CARGO_BIN_EXE_mcp-stdio-mock-server"))
+        .arg("stdin-unread");
+    let mut client = McpStdioClient::connect(spec, limits).await.unwrap();
+    let ready: serde_json::Value = client
+        .request(McpPhase::Running, "fixture/ready", serde_json::json!({}))
+        .await
+        .unwrap();
+    assert_eq!(ready["ready"], true);
+    // Far beyond OS pipe capacity, but inside the existing protocol bound.
+    let call = CallToolRequest::new(
+        "write-blocked",
+        serde_json::json!({"payload": "x".repeat(768 * 1024)}),
+    );
+    let err = tokio::time::timeout(Duration::from_secs(1), client.call_tool(call))
+        .await
+        .expect("the whole request, not only response wait, is bounded")
+        .unwrap_err();
+    assert!(
+        matches!(err.kind, McpErrorKind::Timeout { ref operation, .. } if operation == "tools/call")
+    );
+    assert_eq!(
+        err.request_stage(),
+        Some(mcp::stdio::McpRequestStage::WritingFrame)
+    );
+    // The fixture barrier and reported stage establish actual write backpressure;
+    // the separate controlled fixture also verifies the exact partial bytes.
+    let next = client
+        .request::<serde_json::Value>(McpPhase::Running, "fixture/next", serde_json::json!({}))
+        .await
+        .unwrap_err();
+    assert!(next.to_string().contains("stdin is closed"));
+    let report = tokio::time::timeout(Duration::from_secs(1), client.shutdown())
+        .await
+        .unwrap()
+        .unwrap();
+    assert!(report.terminated || report.killed);
+    assert!(report.exit_status.is_some());
+}

crates/worker/src/feature/mcp.rs

@@ -1081,12 +1081,22 @@ impl Tool for McpStdioTool {
         &self,
         input_json: &str,
         _ctx: ToolExecutionContext,
+    ) -> Result<ToolOutput, ToolError> {
+        self.execute_with_limits(input_json, McpStdioLimits::default())
+            .await
+    }
+}
+
+impl McpStdioTool {
+    async fn execute_with_limits(
+        &self,
+        input_json: &str,
+        limits: McpStdioLimits,
     ) -> Result<ToolOutput, ToolError> {
         let arguments = parse_tool_arguments(input_json)?;
-        let mut client =
+        let mut client = McpStdioClient::connect(self.server_spec.clone(), limits)
-            McpStdioClient::connect(self.server_spec.clone(), McpStdioLimits::default())
+            .await
-                .await
+            .map_err(|err| ToolError::ExecutionFailed(mcp_call_error_message(&err)))?;
-                .map_err(|err| ToolError::ExecutionFailed(mcp_call_error_message(&err)))?;
 
         let call_result = client
             .call_tool(CallToolRequest::new(self.mcp_tool_name.clone(), arguments))
@@ -2208,6 +2218,64 @@ read exit_notification || true
         );
     }
 
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn stdio_tool_request_timeout_reports_unknown_execution_without_replay() {
+        let dir = tempfile::tempdir().unwrap();
+        let log = dir.path().join("frames.jsonl");
+        // Frame capture is observed only after shutdown has waited for process
+        // exit. No guessed sleep or process-global test-control variable.
+        let script = r#"
+printf '%s\n' 'connection' >> "$1"
+read -r init || exit 1
+printf '%s\n' '{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-11-25","capabilities":{"tools":{}},"serverInfo":{"name":"timeout-fixture","version":"1"}}}'
+read -r initialized || exit 1
+while IFS= read -r frame; do
+    printf '%s\n' "$frame" >> "$1"
+done
+"#;
+        let tool = McpStdioTool {
+            server_spec: McpStdioServerSpec::new("timeout-fixture", "/bin/sh").args([
+                "-c".to_string(),
+                script.to_string(),
+                "fixture".to_string(),
+                log.to_str().unwrap().to_string(),
+            ]),
+            mcp_tool_name: "demo-tool".to_string(),
+        };
+        let limits = McpStdioLimits {
+            request_timeout: std::time::Duration::from_millis(100),
+            shutdown_timeout: std::time::Duration::from_millis(100),
+            kill_timeout: std::time::Duration::from_millis(100),
+            ..Default::default()
+        };
+        let err = tokio::time::timeout(
+            std::time::Duration::from_secs(3),
+            tool.execute_with_limits("{}", limits),
+        )
+        .await
+        .unwrap()
+        .expect_err("response intentionally withheld");
+        let message = err.to_string();
+        assert!(
+            message.contains("tools/call transport/protocol failure"),
+            "{message}"
+        );
+        assert!(message.contains("timed out during tools/call"), "{message}");
+        assert!(
+            message.contains("frame flushed; execution unknown; do not retry"),
+            "{message}"
+        );
+        let capture = std::fs::read_to_string(log).unwrap();
+        let mut lines = capture.lines();
+        assert_eq!(lines.next(), Some("connection"));
+        let frames: Vec<Value> = lines
+            .map(|line| serde_json::from_str(line).unwrap())
+            .collect();
+        assert_eq!(frames.len(), 1, "must not reconnect or replay: {capture}");
+        assert_eq!(frames[0]["method"], "tools/call");
+    }
+
     #[test]
     fn call_tool_result_renderer_marks_untrusted_mcp_error() {
         let result = CallToolResult {