Merge pull request #184 from LockInTime/fix/supervised-owner-eagain

fix(sdk): ignore retryable reads on the supervised owner pipe

Author: SarthakWade

apps/headless/Sources/HeadlessCLI/main.swift

@@ -142,7 +142,7 @@ private struct HostLauncher {
                 let count = withUnsafeMutableBytes(of: &byte) { buffer in
                     read(STDIN_FILENO, buffer.baseAddress, 1)
                 }
-                if count >= 0 || errno != EINTR { break }
+                if supervisedOwnerChannelClosed(readCount: count, errnoValue: errno) { break }
             } else if status < 0, errno != EINTR {
                 break
             }

apps/headless/Sources/HeadlessProtocol/SupervisedHost.swift

@@ -6,6 +6,14 @@ import Darwin
 import Glibc
 #endif
 
+/// Owner-pipe reads that mean the SDK/CLI is gone. EOF and unexpected data
+/// close the host. `EINTR` / `EAGAIN` / `EWOULDBLOCK` must not, or a
+/// spurious DispatchSource wakeup kills a live supervised session.
+public func supervisedOwnerChannelClosed(readCount: Int, errnoValue: Int32) -> Bool {
+    if readCount >= 0 { return true }
+    return errnoValue != EINTR && errnoValue != EAGAIN && errnoValue != EWOULDBLOCK
+}
+
 public final class SupervisedHostOwnerMonitor: @unchecked Sendable {
     private let source: DispatchSourceRead
     private let lock = NSLock()
@@ -28,7 +36,7 @@ public final class SupervisedHostOwnerMonitor: @unchecked Sendable {
             let count = withUnsafeMutableBytes(of: &byte) { buffer in
                 read(STDIN_FILENO, buffer.baseAddress, 1)
             }
-            if count >= 0 || errno != EINTR {
+            if supervisedOwnerChannelClosed(readCount: count, errnoValue: errno) {
                 self.stop()
                 onOwnerExit()
             }

apps/headless/Tests/HeadlessProtocolTests/ProtocolTests.swift

@@ -2766,6 +2766,33 @@ struct ProtocolTests {
         )
     }
 
+    static func supervisedOwnerChannelIgnoresRetryableReads() throws {
+        try expect(
+            supervisedOwnerChannelClosed(readCount: 0, errnoValue: 0),
+            "EOF on the owner pipe must stop the host"
+        )
+        try expect(
+            supervisedOwnerChannelClosed(readCount: 1, errnoValue: 0),
+            "unexpected owner-pipe data must stop the host"
+        )
+        try expect(
+            !supervisedOwnerChannelClosed(readCount: -1, errnoValue: EINTR),
+            "EINTR must not be treated as owner exit"
+        )
+        try expect(
+            !supervisedOwnerChannelClosed(readCount: -1, errnoValue: EAGAIN),
+            "EAGAIN must not be treated as owner exit"
+        )
+        try expect(
+            !supervisedOwnerChannelClosed(readCount: -1, errnoValue: EWOULDBLOCK),
+            "EWOULDBLOCK must not be treated as owner exit"
+        )
+        try expect(
+            supervisedOwnerChannelClosed(readCount: -1, errnoValue: EBADF),
+            "a broken owner descriptor must stop the host"
+        )
+    }
+
     static func oversizedSocketRequestIsRejected() throws {
         try LocalRuntime.preparePrivateDirectory()
         let socketPath = LocalRuntime.directoryURL
@@ -4115,6 +4142,7 @@ struct ProtocolTests {
             ("host authentication orchestration", hostAuthenticationOrchestration),
             ("docs command reference matches help", docsCommandReferenceMatchesHelp),
             ("menu shortcuts have unique chords", menuShortcutsHaveUniqueChords),
+            ("supervised owner channel ignores retryable reads", supervisedOwnerChannelIgnoresRetryableReads),
             ("artifact file upload boundaries", artifactUploadCommands),
         ]
 

packages/headless-python/tests/test_lifecycle.py

@@ -479,8 +479,9 @@ async def scenario() -> None:
                     HEADLESS_TEST_PID_FILE=str(pid_file),
                 ),
             )
-        # Cleanup has its own 100 ms budget after the single 300 ms startup budget.
-        assert time.monotonic() - started < 0.5
+        # Cleanup has its own 100 ms budget after the single 300 ms startup
+        # budget. Leave slack for slow macOS 3.14 CI runners.
+        assert time.monotonic() - started < 1.0
 
     asyncio.run(scenario())
     assert process_is_gone(int(pid_file.read_text()))