fix: make sure daemons are brought down cleanly

.github/workflows/pre-commit.yml

@@ -29,5 +29,8 @@ jobs:
       - name: Install dependencies
         run: uv sync --python ${{ matrix.python-version }} --python-preference only-managed
 
-      - name: Run prek
-        run: uv run prek run --all-files
+      - name: Lint & format (prek)
+        run: uv run prek run --all-files --verbose --show-diff-on-failure --skip pytest
+
+      - name: Test (pytest)
+        run: uv run pytest tests/ -v

src/cocoindex_code/cli.py

@@ -467,26 +467,26 @@ def daemon_restart() -> None:
 @daemon_app.command("stop")
 def daemon_stop() -> None:
     """Stop the daemon."""
-    from .client import stop_daemon
+    from .client import is_daemon_running, stop_daemon
     from .daemon import daemon_pid_path
 
     pid_path = daemon_pid_path()
-    if not pid_path.exists():
+    if not pid_path.exists() and not is_daemon_running():
         _typer.echo("Daemon is not running.")
         return
 
     stop_daemon()
 
-    # Wait for process to exit
+    # Wait for process to exit (check both pid file and socket)
     import time
 
     deadline = time.monotonic() + 5.0
     while time.monotonic() < deadline:
-        if not pid_path.exists():
+        if not pid_path.exists() and not is_daemon_running():
             break
         time.sleep(0.1)
 
-    if pid_path.exists():
+    if pid_path.exists() or is_daemon_running():
         _typer.echo("Warning: daemon may not have stopped cleanly.", err=True)
     else:
         _typer.echo("Daemon stopped.")

src/cocoindex_code/client.py

@@ -183,13 +183,26 @@ def start_daemon() -> None:
         cmd = [sys.executable, "-m", "cocoindex_code.cli", "run-daemon"]
 
     log_fd = open(log_path, "a")
-    subprocess.Popen(
-        cmd,
-        start_new_session=True,
-        stdout=log_fd,
-        stderr=log_fd,
-        stdin=subprocess.DEVNULL,
-    )
+    if sys.platform == "win32":
+        # DETACHED_PROCESS fully detaches the daemon from the parent console,
+        # preventing its exit code from leaking back to the calling shell.
+        _create_new_process_group = 0x00000200
+        _detached_process = 0x00000008
+        subprocess.Popen(
+            cmd,
+            stdout=log_fd,
+            stderr=log_fd,
+            stdin=subprocess.DEVNULL,
+            creationflags=_create_new_process_group | _detached_process,
+        )
+    else:
+        subprocess.Popen(
+            cmd,
+            start_new_session=True,
+            stdout=log_fd,
+            stderr=log_fd,
+            stdin=subprocess.DEVNULL,
+        )
     log_fd.close()
 
 
@@ -205,12 +218,49 @@ def _find_ccc_executable() -> str | None:
     return None
 
 
+def _pid_alive(pid: int) -> bool:
+    """Return True if *pid* is still running."""
+    if sys.platform == "win32":
+        # Avoid os.kill(pid, 0) on Windows — it has a CPython bug that corrupts
+        # the C-level exception state, causing subsequent C function calls
+        # (time.monotonic, time.sleep) to raise SystemError even after the
+        # OSError is caught.  Use OpenProcess via ctypes instead.
+        import ctypes
+
+        kernel32 = getattr(ctypes, "windll").kernel32
+        handle = kernel32.OpenProcess(0x1000, False, pid)  # PROCESS_QUERY_LIMITED_INFORMATION
+        if handle:
+            kernel32.CloseHandle(handle)
+            return True
+        return False
+    try:
+        os.kill(pid, 0)  # signal 0: check existence without killing
+        return True
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True  # process exists but we can't signal it
+
+
 def stop_daemon() -> None:
     """Stop the daemon gracefully.
 
-    Sends a StopRequest, waits for the process to exit, falls back to SIGTERM.
+    Sends a StopRequest, waits for the process to exit, falls back to
+    SIGTERM → SIGKILL.  Only removes the PID file after confirming that
+    the specific PID is no longer alive.
     """
-    # Step 1: try sending StopRequest
+    pid_path = daemon_pid_path()
+
+    # Read the PID early so we can track the actual process.
+    pid: int | None = None
+    try:
+        pid = int(pid_path.read_text().strip())
+        if pid == os.getpid():
+            pid = None  # safety: never kill ourselves
+    except (FileNotFoundError, ValueError):
+        pass
+
+    # Step 1: try sending StopRequest via socket
     try:
         client = DaemonClient.connect()
         client.handshake()
@@ -220,65 +270,78 @@ def stop_daemon() -> None:
         pass
 
     # Step 2: wait for process to exit (up to 5s)
-    pid_path = daemon_pid_path()
-    deadline = time.monotonic() + 5.0
-    while time.monotonic() < deadline and pid_path.exists():
-        time.sleep(0.1)
-
-    if not pid_path.exists():
-        return  # Clean exit
+    if pid is not None:
+        deadline = time.monotonic() + 5.0
+        while time.monotonic() < deadline and _pid_alive(pid):
+            time.sleep(0.1)
+        if not _pid_alive(pid):
+            _cleanup_stale_files(pid_path, pid)
+            return
 
     # Step 3: if still running, try SIGTERM
-    pid: int | None = None
-    if pid_path.exists():
+    if pid is not None and _pid_alive(pid):
         try:
-            pid = int(pid_path.read_text().strip())
-            if pid != os.getpid():
-                os.kill(pid, signal.SIGTERM)
-            else:
-                pid = None
-        except (ValueError, ProcessLookupError, PermissionError):
+            os.kill(pid, signal.SIGTERM)
+        except (ProcessLookupError, PermissionError):
             pass
 
-        # Wait a bit more
         deadline = time.monotonic() + 2.0
-        while time.monotonic() < deadline and pid_path.exists():
+        while time.monotonic() < deadline and _pid_alive(pid):
             time.sleep(0.1)
 
-    # Step 4: if still running, escalate to SIGKILL (Unix only;
+        if not _pid_alive(pid):
+            _cleanup_stale_files(pid_path, pid)
+            return
+
+    # Step 4: escalate to SIGKILL (Unix only;
     # on Windows SIGTERM already calls TerminateProcess)
-    if sys.platform != "win32" and pid_path.exists():
+    if sys.platform != "win32" and pid is not None and _pid_alive(pid):
         try:
-            pid = int(pid_path.read_text().strip())
-            if pid != os.getpid():
-                os.kill(pid, signal.SIGKILL)
-            else:
-                pid = None
-        except (ValueError, ProcessLookupError, PermissionError):
+            os.kill(pid, signal.SIGKILL)
+        except (ProcessLookupError, PermissionError):
             pass
 
+        # SIGKILL is async; give the kernel a moment to reap
+        deadline = time.monotonic() + 1.0
+        while time.monotonic() < deadline and _pid_alive(pid):
+            time.sleep(0.1)
+
     # Step 4b: on Windows, wait for the process to fully exit after TerminateProcess
     # so that named pipe handles are released before starting a new daemon.
     if sys.platform == "win32" and pid is not None:
         deadline = time.monotonic() + 3.0
-        while time.monotonic() < deadline:
-            try:
-                os.kill(pid, 0)  # Check if process still exists
-                time.sleep(0.1)
-            except (ProcessLookupError, PermissionError, OSError):
-                break  # Process has exited
+        while time.monotonic() < deadline and _pid_alive(pid):
+            time.sleep(0.1)
 
     # Step 5: clean up stale files
+    _cleanup_stale_files(pid_path, pid)
+
+
+def _cleanup_stale_files(pid_path: Path, pid: int | None) -> None:
+    """Remove socket and PID file after the daemon has exited.
+
+    Only removes the PID file when *pid* matches what is on disk, to
+    avoid accidentally deleting a newer daemon's PID file.
+    """
     if sys.platform != "win32":
         sock = daemon_socket_path()
         try:
             Path(sock).unlink(missing_ok=True)
         except Exception:
             pass
-    try:
-        pid_path.unlink(missing_ok=True)
-    except Exception:
-        pass
+    if pid is not None:
+        try:
+            stored = pid_path.read_text().strip()
+            if stored == str(pid):
+                pid_path.unlink(missing_ok=True)
+        except (FileNotFoundError, ValueError):
+            pass
+    else:
+        # No PID known — cautiously remove if file exists
+        try:
+            pid_path.unlink(missing_ok=True)
+        except Exception:
+            pass
 
 
 def _wait_for_daemon(timeout: float = 30.0) -> None:

src/cocoindex_code/daemon.py

@@ -447,17 +447,24 @@ def run_daemon() -> None:
     try:
         asyncio.run(_async_daemon_main(embedder))
     finally:
-        # Clean up PID file and socket (named pipes on Windows clean up automatically)
-        try:
-            pid_path.unlink(missing_ok=True)
-        except Exception:
-            pass
+        # Clean up socket first, then PID file last.
+        # The PID file is the authoritative "daemon is alive" indicator, so it
+        # must be the very last thing removed to avoid races where a client
+        # sees the PID gone but the socket (or process) is still lingering.
         if sys.platform != "win32":
             sock = daemon_socket_path()
             try:
                 Path(sock).unlink(missing_ok=True)
             except Exception:
                 pass
+        # Only remove the PID file if it still contains *our* PID.
+        # A new daemon may have already overwritten it during a restart race.
+        try:
+            stored = pid_path.read_text().strip()
+            if stored == str(os.getpid()):
+                pid_path.unlink(missing_ok=True)
+        except Exception:
+            pass
         logger.info("Daemon stopped")
 
 

tests/test_daemon.py

@@ -29,6 +29,7 @@
     RemoveProjectRequest,
     Response,
     SearchRequest,
+    StopRequest,
     decode_response,
     encode_request,
 )
@@ -92,6 +93,18 @@ def daemon_sock() -> Iterator[str]:
 
     yield sock_path
 
+    # Gracefully shut down the daemon thread so named pipes are released on Windows
+    try:
+        conn = Client(sock_path, family=_connection_family())
+        conn.send_bytes(encode_request(HandshakeRequest(version=__version__)))
+        conn.recv_bytes()
+        conn.send_bytes(encode_request(StopRequest()))
+        conn.recv_bytes()
+        conn.close()
+    except Exception:
+        pass
+    thread.join(timeout=5)
+
     # Restore patches and env var
     dm.create_embedder = _orig_create_embedder  # type: ignore[attr-defined]
     if old_env is None: