From bca0050dcb1dede2bce49bad2734927d4a8a5693 Mon Sep 17 00:00:00 2001
From: Eric <ciregenz@berkeley.edu>
Date: Tue, 16 Jun 2026 21:59:16 -0700
Subject: [PATCH] [eric] perf: start 9router in background instead of blocking
 the http bind

- service lifespan was awaiting ensure_9router (~7.4s, up to ~18s cold) on the boot path

- dispatch already ensures it lazily; serialize ensure_running so no double-spawn

- cuts warm backend-ready ~9-10s toward ~2-3s
---
 backend/apps/nine_router/process.py | 15 +++++++++++++++
 backend/apps/service/service.py     | 19 +++++++++++++++++--
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/backend/apps/nine_router/process.py b/backend/apps/nine_router/process.py
index b34a51e0..12935c00 100644
--- a/backend/apps/nine_router/process.py
+++ b/backend/apps/nine_router/process.py
@@ -57,6 +57,11 @@ NINE_ROUTER_NPM_VERSION = os.environ.get("OPENSWARM_ROUTER_VERSION", "0.3.60")
 
 _process: subprocess.Popen | None = None
 
+# Serializes ensure_running() so a background auto-start and a concurrent
+# dispatch-time ensure can't both spawn 9Router (double-bind on :20128). Lazily
+# created so module import doesn't require a running event loop.
+_start_lock: "asyncio.Lock | None" = None
+
 # Short TTL cache for positive is_running() results. The probe is a sync
 # httpx.get that blocks the event loop, and under load (9Router busy
 # streaming inference) it can exceed its 2s timeout and return False even
@@ -334,6 +339,16 @@ def _report_start_failure(reason: str, *, detail: str = "", **fields: Any) -> No
 
 
 async def ensure_running():
+    """Start 9Router if not already running. Serialized so concurrent callers
+    (the background auto-start + a dispatch-time ensure) can't double-spawn."""
+    global _start_lock
+    if _start_lock is None:
+        _start_lock = asyncio.Lock()
+    async with _start_lock:
+        await _ensure_running_impl()
+
+
+async def _ensure_running_impl():
     """Start 9Router if not already running."""
     global _process
     _is_packaged = os.environ.get("OPENSWARM_PACKAGED") == "1"
diff --git a/backend/apps/service/service.py b/backend/apps/service/service.py
index 532eab62..93c70851 100644
--- a/backend/apps/service/service.py
+++ b/backend/apps/service/service.py
@@ -33,6 +33,7 @@ logger = logging.getLogger(__name__)
 
 _pulse_task: asyncio.Task | None = None
 _drain_task: asyncio.Task | None = None
+_9r_start_task: asyncio.Task | None = None
 
 _last_9r_cost: float | None = None
 _last_9r_prompt_tokens: int | None = None
@@ -121,7 +122,7 @@ async def _drain_loop():
 
 @asynccontextmanager
 async def service_lifespan():
-    global _pulse_task, _drain_task
+    global _pulse_task, _drain_task, _9r_start_task
 
     try:
         from backend.apps.settings.settings import load_settings, _save_settings
@@ -193,7 +194,13 @@ async def service_lifespan():
 
     try:
         from backend.apps.nine_router import ensure_running as ensure_9router
-        await ensure_9router()
+        # Start 9Router in the BACKGROUND instead of awaiting it here. Awaiting
+        # it was ~7s (up to ~18s cold) of the startup critical path, blocking the
+        # HTTP bind and the whole UI behind it. 9Router is only needed when the
+        # user sends an agent message, and the dispatch path calls ensure_running()
+        # itself (now serialized, so no double-spawn), so the first message waits
+        # for readiness lazily. This is the single biggest warm-startup win.
+        _9r_start_task = asyncio.create_task(ensure_9router())
     except Exception as e:
         logger.debug(f"9Router auto-start skipped: {e}")
 
@@ -218,6 +225,14 @@ async def service_lifespan():
             pass
         _drain_task = None
 
+    if _9r_start_task and not _9r_start_task.done():
+        _9r_start_task.cancel()
+        try:
+            await _9r_start_task
+        except (asyncio.CancelledError, Exception):
+            pass
+    _9r_start_task = None
+
     try:
         from backend.apps.nine_router import stop as stop_9router
         stop_9router()