diff --git a/gateway/control_socket.py b/gateway/control_socket.py
index 383b3c8769..50c516f393 100644
--- a/gateway/control_socket.py
+++ b/gateway/control_socket.py
@@ -543,3 +543,18 @@ def _query_windows_pipe(
 def identify_gateway(home: Path, *, timeout: float = _DEFAULT_CLIENT_TIMEOUT) -> Optional[dict[str, Any]]:
     """Convenience wrapper: ``identify`` the gateway serving ``home``."""
     return query_gateway_control(home, "identify", timeout=timeout)
+
+
+def pause_gateway_for_update(
+    home: Path, *, timeout: float = _DEFAULT_CLIENT_TIMEOUT
+) -> Optional[dict[str, Any]]:
+    """Ask the gateway serving ``home`` to drain and exit for an update.
+
+    Step 2 of the socket migration (#92091). Returns the gateway's ACK —
+    ``{"pausing": bool, "already_stopping": bool, "pid": int,
+    "drain_timeout": float}`` — or None when no gateway answers (older
+    gateway without the verb, no socket, dead socket). None means the
+    caller falls back to the legacy pause path (signals / tree-kill),
+    exactly as before this verb existed.
+    """
+    return query_gateway_control(home, "pause-for-update", timeout=timeout)
diff --git a/gateway/run.py b/gateway/run.py
index 915eee4d49..7e16191321 100644
--- a/gateway/run.py
+++ b/gateway/run.py
@@ -31437,7 +31437,47 @@ async def start_gateway(config: Optional[GatewayConfig] = None, replace: bool =
     try:
         from gateway.control_socket import GatewayControlServer
 
-        _control_server = GatewayControlServer()
+        # pause-for-update (#92091 step 2): the updater asks this gateway to
+        # drain in-flight turns and exit cleanly — releasing every venv file
+        # handle — instead of being tree-killed mid-turn. Same drain path as
+        # SIGUSR1/service restarts (request_restart(via_service=True)); the
+        # updater (or the service manager) relaunches after the code swap.
+        # The handler runs on the socket's executor thread, so the restart
+        # request is marshalled onto the loop thread; the ACK returns the
+        # drain budget so the caller knows how long to wait for exit.
+        _main_loop = asyncio.get_running_loop()
+
+        def _pause_for_update_handler() -> dict:
+            try:
+                from hermes_cli.gateway import _get_restart_drain_timeout
+
+                _drain = float(_get_restart_drain_timeout())
+            except Exception:
+                _drain = 30.0
+            accepted_box: list[bool] = []
+            _done = threading.Event()
+
+            def _request() -> None:
+                try:
+                    accepted_box.append(
+                        runner.request_restart(detached=False, via_service=True)
+                    )
+                finally:
+                    _done.set()
+
+            _main_loop.call_soon_threadsafe(_request)
+            _done.wait(timeout=5.0)
+            accepted = bool(accepted_box and accepted_box[0])
+            return {
+                "pausing": accepted,
+                "already_stopping": not accepted,
+                "pid": os.getpid(),
+                "drain_timeout": _drain,
+            }
+
+        _control_server = GatewayControlServer(
+            verb_handlers={"pause-for-update": _pause_for_update_handler}
+        )
         if not await _control_server.start():
             _control_server = None
         else:
diff --git a/hermes_cli/update_cmd.py b/hermes_cli/update_cmd.py
index fa527fe55a..3a62699bc3 100644
--- a/hermes_cli/update_cmd.py
+++ b/hermes_cli/update_cmd.py
@@ -5116,6 +5116,7 @@ def _pause_windows_gateways_for_update() -> dict | None:
 
     profiles: dict[str, int] = {}
     mapped_pids = []
+    socket_acks: list[dict] = []
     for pid in running_pids:
         proc = profile_processes.get(pid)
         if proc is None:
@@ -5123,6 +5124,23 @@ def _pause_windows_gateways_for_update() -> dict | None:
         profiles[str(proc.profile)] = int(pid)
         mapped_pids.append(int(pid))
         _write_update_planned_stop_marker(Path(proc.path), int(pid))
+        # Socket-first pause (#92091 step 2): ask the gateway to drain and
+        # exit itself instead of relying on the marker poll + force-kill
+        # ladder. A positive ACK means the gateway is running its own
+        # graceful restart path (same drain as SIGUSR1/service restarts) and
+        # will release its venv handles on the way out. No answer (older
+        # gateway, no socket) → the marker watcher / force-kill fallback
+        # below behaves exactly as before this verb existed.
+        try:
+            from gateway.control_socket import pause_gateway_for_update
+
+            ack = pause_gateway_for_update(Path(proc.path))
+            if ack and (ack.get("pausing") or ack.get("already_stopping")):
+                socket_acks.append(ack)
+        except Exception as exc:
+            logger.debug(
+                "Socket pause unavailable for gateway %s: %s", pid, exc
+            )
 
     # Resolve each mapped worker's venv-side launcher BEFORE draining: the
     # drain stops tracking a PID exactly when it dies, so a gracefully
@@ -5143,6 +5161,23 @@ def _pause_windows_gateways_for_update() -> dict | None:
         drain_timeout = max(float(_get_restart_drain_timeout()), 1.0)
     except Exception:
         drain_timeout = 10.0
+    if socket_acks:
+        # A socket-paused gateway drains its ACTIVE TURN before exiting; give
+        # it the budget it declared (plus teardown grace) rather than only
+        # the local default, so a mid-turn gateway isn't force-killed at the
+        # end of a too-short wait — the exact outcome the verb exists to
+        # prevent.
+        try:
+            declared = max(
+                float(a.get("drain_timeout") or 0.0) for a in socket_acks
+            )
+            drain_timeout = max(drain_timeout, declared + 10.0)
+        except Exception:
+            pass
+        print(
+            f"  → {len(socket_acks)} gateway(s) ACKed socket pause; "
+            f"waiting up to {int(drain_timeout)}s for graceful exit"
+        )
     survivors = _m()._wait_for_windows_update_gateway_exit(
         mapped_pids,
         timeout=drain_timeout,
diff --git a/tests/gateway/test_control_socket_pause.py b/tests/gateway/test_control_socket_pause.py
new file mode 100644
index 0000000000..07998ea36a
--- /dev/null
+++ b/tests/gateway/test_control_socket_pause.py
@@ -0,0 +1,109 @@
+"""pause-for-update control-socket verb (#92091 step 2, campaign #91277).
+
+The updater asks a running gateway to drain and exit cleanly (releasing its
+venv file handles) instead of tree-killing it mid-turn. Fallback contract:
+older gateways without the verb answer nothing, and callers keep the legacy
+marker/force-kill path.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from gateway.control_socket import (
+    GatewayControlServer,
+    pause_gateway_for_update,
+    query_gateway_control,
+)
+
+
+def _make_server(tmp_path, handler):
+    server = GatewayControlServer(
+        home=tmp_path, verb_handlers={"pause-for-update": handler}
+    )
+    return server
+
+
+def test_pause_verb_dispatches_and_returns_ack(tmp_path):
+    calls = []
+
+    def handler():
+        calls.append(1)
+        return {"pausing": True, "already_stopping": False, "pid": 111,
+                "drain_timeout": 30.0}
+
+    server = _make_server(tmp_path, handler)
+    raw = json.dumps({"verb": "pause-for-update", "id": 7}).encode()
+    response = json.loads(server.handle_request_line(raw).decode())
+    assert response["ok"] is True
+    assert response["result"]["pausing"] is True
+    assert response["result"]["drain_timeout"] == 30.0
+    assert response["id"] == 7
+    assert calls == [1]
+
+
+def test_unknown_verb_still_lists_pause(tmp_path):
+    server = _make_server(tmp_path, lambda: {})
+    raw = json.dumps({"verb": "nope"}).encode()
+    response = json.loads(server.handle_request_line(raw).decode())
+    assert response["ok"] is False
+    assert "pause-for-update" in response["supported_verbs"]
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason="unix socket transport")
+def test_pause_client_roundtrip_over_real_socket(tmp_path):
+    """Full client→socket→handler→ACK path over a REAL unix socket."""
+
+    async def scenario():
+        acks = []
+
+        def handler():
+            acks.append(1)
+            return {"pausing": True, "already_stopping": False,
+                    "pid": 4242, "drain_timeout": 12.5}
+
+        server = _make_server(tmp_path, handler)
+        assert await server.start()
+        try:
+            loop = asyncio.get_running_loop()
+            result = await loop.run_in_executor(
+                None, lambda: pause_gateway_for_update(tmp_path, timeout=5.0)
+            )
+        finally:
+            await server.stop()
+        return result, acks
+
+    result, acks = asyncio.run(scenario())
+    assert acks == [1]
+    assert result is not None
+    assert result["pausing"] is True and result["drain_timeout"] == 12.5
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason="unix socket transport")
+def test_pause_client_none_when_gateway_lacks_verb(tmp_path):
+    """Back-compat: a step-1 gateway (identify/status only) answers ok:false
+    for the unknown verb → the client returns None → caller keeps the legacy
+    kill path."""
+
+    async def scenario():
+        server = GatewayControlServer(home=tmp_path)  # no pause handler
+        assert await server.start()
+        try:
+            loop = asyncio.get_running_loop()
+            return await loop.run_in_executor(
+                None, lambda: pause_gateway_for_update(tmp_path, timeout=5.0)
+            )
+        finally:
+            await server.stop()
+
+    assert asyncio.run(scenario()) is None
+
+
+def test_pause_client_none_when_no_socket(tmp_path):
+    assert pause_gateway_for_update(tmp_path, timeout=0.5) is None