fix(terminal): stop proot "can't sanitize binding" fd warnings (#2878)

* fix(terminal): stop proot "can't sanitize binding" fd warnings

proot canonicalizes every -b host path with realpath(3). The /proc/self/fd/N magic links only resolve when the descriptor points to a real file, so piped stdio (what Acode's ProcessBuilder uses) made realpath(3) fail and proot drop the binding with `can't sanitize binding "/proc/self/fd/N"`.

The [ -e ] guard passed those links because stat(2) follows the magic link to the underlying pipe/socket inode, so it never filtered them out. Replace it with a can_bind helper that resolves the link target the way realpath(3) does, and narrow the LSP-side proot-warning filters to only this benign message so real proot warnings are no longer swallowed.

* fix(terminal): probe /proc/self/fd via shell pid in can_bind

readlink(1) runs as a child process, so for fd 2 its /proc/self/fd/2 is the /dev/null of the `2>/dev/null` redirect rather than the stderr proot inherits.  That made can_bind treat a piped stderr as bindable, leaving one `can't sanitize binding "/proc/self/fd/2"` warning.  Probe /proc/$$/fd/N instead; $$ survives the exec and names the process proot runs as.

Author: RohitKushvaha01

src/cm/lsp/serverLauncher.ts

@@ -40,6 +40,13 @@ const STATUS_FAILED: InstallStatus = "failed";
 
 const DONT_ASK_TERMINAL_REQUIRED_FOR_LSP = "dontAskTerminalRequiredForLsp";
 
+// PRoot prints `can't sanitize binding "/proc/self/fd/N"` when one of the
+// session's stdio descriptors is a pipe/socket and therefore cannot be
+// canonicalized (see init-sandbox.sh).  Drop only that known, harmless message
+// so genuine proot warnings are still surfaced in the LSP log.
+const PROOT_FD_BINDING_WARNING =
+  /can'?t sanitize binding "\/proc\/self\/fd(?:\/[012])?"/i;
+
 let alreadyInformed = false;
 
 function getTerminalRequiredMessage(): string {
@@ -911,7 +918,7 @@ async function startInteractiveServer(
 ): Promise<string> {
   const executor = getExecutor();
   const callback: ExecutorCallback = (type, data) => {
-    if (type === "stderr" && /proot warning/i.test(data)) return;
+    if (type === "stderr" && PROOT_FD_BINDING_WARNING.test(data)) return;
     if (type === "stdout" && /listening on/i.test(data)) {
       signalServerReady(serverId);
     }

src/plugins/terminal/scripts/init-sandbox.sh

@@ -69,19 +69,51 @@ ARGS="$ARGS -b $PREFIX/public:/root"
 ARGS="$ARGS -b $PREFIX/alpine/tmp:/dev/shm"
 
 
-if [ -e "/proc/self/fd" ]; then
+# PRoot canonicalizes every -b host path with realpath(3).  The magic links
+# under /proc/self/fd only resolve when the descriptor points at a real file:
+# for pipes, sockets, anon inodes or memfds the link target is not a path
+# ("pipe:[123]"), so realpath(3) fails with ENOENT, proot drops the binding and
+# prints `can't sanitize binding "/proc/self/fd/N"`.  `[ -e ]` follows the magic
+# link to the underlying inode and therefore returns true for those descriptors,
+# which is why it does not filter them out.  Test the link target the same way
+# realpath(3) does instead.
+#
+# Probe through this shell's own pid ($$), not /proc/self: readlink(1) runs as a
+# child process, so for fd 2 its /proc/self/fd/2 is the /dev/null of the
+# `2>/dev/null` redirect below rather than the stderr proot inherits.  $$ is
+# unchanged by the exec at the end of this script, so it always names the
+# process proot will run as.
+SELF_PID=$$
+
+can_bind() {
+    # Directories (e.g. /proc/self/fd) are canonicalizable as-is.
+    if [ -d "$1" ]; then
+        return 0
+    fi
+
+    # A /proc/self/fd/N magic link is only canonicalizable when it resolves to
+    # an existing absolute path; "pipe:[N]", "socket:[N]" and friends are not.
+    target=$(readlink "$1" 2>/dev/null) || return 1
+
+    case "$target" in
+        /*) [ -e "$target" ] ;;
+        *) return 1 ;;
+    esac
+}
+
+if can_bind "/proc/$SELF_PID/fd"; then
   ARGS="$ARGS -b /proc/self/fd:/dev/fd"
 fi
 
-if [ -e "/proc/self/fd/0" ]; then
+if can_bind "/proc/$SELF_PID/fd/0"; then
   ARGS="$ARGS -b /proc/self/fd/0:/dev/stdin"
 fi
 
-if [ -e "/proc/self/fd/1" ]; then
+if can_bind "/proc/$SELF_PID/fd/1"; then
   ARGS="$ARGS -b /proc/self/fd/1:/dev/stdout"
 fi
 
-if [ -e "/proc/self/fd/2" ]; then
+if can_bind "/proc/$SELF_PID/fd/2"; then
   ARGS="$ARGS -b /proc/self/fd/2:/dev/stderr"
 fi
 

src/settings/lspServerDetail.js

@@ -154,18 +154,19 @@ function formatStartupTimeoutValue(timeout) {
 		: strings["lsp-default"];
 }
 
+// PRoot prints `can't sanitize binding "/proc/self/fd/N"` when one of the
+// session's stdio descriptors is a pipe/socket and therefore cannot be
+// canonicalized (see init-sandbox.sh).  Drop only that known, harmless message
+// so genuine proot warnings still reach the UI.
+const PROOT_FD_BINDING_WARNING =
+	/can'?t sanitize binding "\/proc\/self\/fd(?:\/[012])?"/i;
+
 function sanitizeInstallMessage(message) {
 	const lines = String(message || "")
 		.split("\n")
 		.map((line) => line.trim())
 		.filter(Boolean)
-		.filter(
-			(line) =>
-				!/^proot warning:/i.test(line) &&
-				!line.includes(`"/proc/self/fd/0"`) &&
-				!line.includes(`"/proc/self/fd/1"`) &&
-				!line.includes(`"/proc/self/fd/2"`),
-		);
+		.filter((line) => !PROOT_FD_BINDING_WARNING.test(line));
 
 	return lines.join(" ");
 }