From 1eb4d0afe0780d6ae2deb4545332c52569ee7ffc Mon Sep 17 00:00:00 2001
From: Rares Cristea <rares.mihail.cristea@gmail.com>
Date: Wed, 4 Nov 2026 14:29:24 +0200
Subject: [PATCH] Read a plugin's whole tools/list line, however long (#797)

An MCP plugin whose `tools/list` answer is one JSON-RPC line over 1 MiB never loaded: the process reader cut long lines into fragments that never parsed, so the plugin timed out. The partial-line cap is now the 16 MiB the MCP client already allows, with a regression test for a multi-megabyte line.
---
 crates/platform/src/process.rs | 27 +++++++++++++++++++++++++--
 docs/pitfalls.md               |  8 ++++++++
 2 files changed, 33 insertions(+), 2 deletions(-)

diff --git a/crates/platform/src/process.rs b/crates/platform/src/process.rs
index 3d6076a1..70501180 100644
--- a/crates/platform/src/process.rs
+++ b/crates/platform/src/process.rs
@@ -59,8 +59,10 @@ pub const DEFAULT_STDERR_CAP: usize = 1024 * 1024;
 /// How long to wait for the output threads after the child exits (an escaped grandchild may hold
 /// the pipe open; we stop waiting rather than hang).
 const DRAIN_GRACE: Duration = Duration::from_secs(2);
-/// Longest partial line kept before it is delivered anyway.
-const MAX_LINE: usize = 1024 * 1024;
+/// Longest partial line kept before it is delivered anyway. MCP servers answer `tools/list` as
+/// one JSON-RPC line that can be megabytes (a plugin with many richly described tools), so this
+/// must cover such a whole line: a fragment is not JSON, and the answer would never parse.
+const MAX_LINE: usize = 16 * 1024 * 1024;
 
 /// A program to run, built fluently.
 #[derive(Clone)]
@@ -1445,6 +1447,27 @@ mod tests {
         assert!(err.message.contains("bad thing"), "{}", err.message);
     }
 
+    /// An MCP plugin answers `tools/list` as one JSON-RPC line that can be megabytes (a server with
+    /// many richly described tools). A line past the old 1 MiB cap arrived in fragments, so the
+    /// answer never parsed and the plugin timed out: the whole line must arrive as one line.
+    #[cfg(unix)]
+    #[test]
+    fn a_multimegabyte_line_arrives_whole() {
+        let (tx, rx) = std::sync::mpsc::channel();
+        let wanted = 3 * 1024 * 1024 + 7; // past the old cap, and not a round number
+        let out = Process::new("sh")
+            .args(["-c", &format!("head -c {wanted} /dev/zero | tr '\\0' a; echo")])
+            .on_stdout_line(move |l| {
+                let _ = tx.send(l.len());
+            })
+            .timeout(Duration::from_secs(30))
+            .run()
+            .unwrap();
+        assert!(out.success);
+        let lines: Vec<usize> = rx.into_iter().collect();
+        assert_eq!(lines, vec![wanted], "one line of {wanted} bytes, not {} fragments of {:?}", lines.len(), lines);
+    }
+
     #[test]
     fn missing_program_is_a_clear_error() {
         let err = Process::new("no-such-program-craft-xyz").run().unwrap_err();
diff --git a/docs/pitfalls.md b/docs/pitfalls.md
index 396924ea..6bb29639 100644
--- a/docs/pitfalls.md
+++ b/docs/pitfalls.md
@@ -734,3 +734,11 @@ sum the same payload repeatedly. Count checksum byte visits across the whole sca
 cancellation callback inside long sums, and distinguish skipped verification from a mismatch.
 Test overlapping candidates with correct checksums and a complete nested archive so a structural
 trailer cannot turn unchecked data into a high-confidence result.
+
+**A line cap that cut a JSON-RPC answer in pieces.** MCP over stdio answers one message per line,
+and a server with many richly described tools answers `tools/list` as several megabytes on that
+one line. The process reader capped a partial line at 1 MiB and delivered it anyway, so the answer
+arrived as fragments that never parsed and the plugin failed as a `tools/list` timeout — while the
+same server answered a direct probe in under a second. A cap protects memory, but a JSON-RPC line
+must arrive whole to parse at all: keep the cap at the size the MCP client already allows, and test
+with a multi-megabyte line, not a short one.