From b316edbaf954e5b5c2c0ba55b9bae6708a997f0f Mon Sep 17 00:00:00 2001
From: pi <pi@led-controller.local>
Date: Sun, 3 May 2026 21:27:31 +1200
Subject: [PATCH] fix(wifi): stagger driver ws dials and extend initial retry
 window

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 src/models/wifi_ws_clients.py | 63 ++++++++++++++++++++++++++++++-----
 src/settings.py               | 13 ++++++++
 2 files changed, 67 insertions(+), 9 deletions(-)

diff --git a/src/models/wifi_ws_clients.py b/src/models/wifi_ws_clients.py
index 1143d2a..fdd6a2c 100644
--- a/src/models/wifi_ws_clients.py
+++ b/src/models/wifi_ws_clients.py
@@ -195,6 +195,27 @@ async def _recv_forward_loop(ip: str, ws) -> None:
                 pass
 
 
+def _stagger_delay_s_for_ip(ip: str) -> float:
+    """0 .. wifi_driver_connect_stagger_max_s based on last IPv4 octet (deterministic spread)."""
+    global _settings
+    if _settings is None:
+        return 0.0
+    try:
+        max_s = float(_settings.get("wifi_driver_connect_stagger_max_s", 2.5))
+    except (TypeError, ValueError):
+        max_s = 2.5
+    if max_s <= 0:
+        return 0.0
+    parts = str(ip).strip().split(".")
+    if len(parts) != 4:
+        return 0.0
+    try:
+        last = int(parts[3]) % 256
+    except ValueError:
+        return 0.0
+    return (last / 255.0) * max_s
+
+
 async def _driver_connection_loop(ip: str) -> None:
     global _settings
     if _settings is None:
@@ -204,16 +225,37 @@ async def _driver_connection_loop(ip: str) -> None:
     if not path.startswith("/"):
         path = "/" + path
     uri = f"ws://{ip}:{port}{path}"
-    retry_interval_s = 2.0
-    retry_window_s = 30.0
-    deadline = asyncio.get_running_loop().time() + retry_window_s
+    try:
+        retry_interval_s = float(_settings.get("wifi_driver_connect_retry_interval_s", 2.0))
+    except (TypeError, ValueError):
+        retry_interval_s = 2.0
+    retry_interval_s = max(0.2, retry_interval_s)
+    try:
+        retry_window_s = float(_settings.get("wifi_driver_connect_retry_window_s", 120.0))
+    except (TypeError, ValueError):
+        retry_window_s = 120.0
+    retry_window_s = max(5.0, retry_window_s)
+    try:
+        open_timeout = float(_settings.get("wifi_driver_ws_open_timeout", 45.0))
+    except (TypeError, ValueError):
+        open_timeout = 45.0
+    open_timeout = max(5.0, open_timeout)
+
+    loop = asyncio.get_running_loop()
+    stagger = _stagger_delay_s_for_ip(ip)
+    if stagger > 0:
+        await asyncio.sleep(stagger)
+
+    # Only bound boot-time: after we have connected once, keep retrying (Wi-Fi drops, reboots).
+    connected_once = False
+    deadline = loop.time() + retry_window_s
     try:
         while True:
-            now = asyncio.get_running_loop().time()
-            if now >= deadline:
+            now = loop.time()
+            if not connected_once and now >= deadline:
                 print(
-                    f"[WS] driver {ip} still unreachable after {int(retry_window_s)}s; "
-                    "stopping retries until next hello"
+                    f"[WS] driver {ip} still unreachable after {int(retry_window_s)}s "
+                    f"(initial window); stopping until next UDP hello / registry prime"
                 )
                 break
             try:
@@ -222,8 +264,9 @@ async def _driver_connection_loop(ip: str) -> None:
                     uri,
                     ping_interval=20,
                     ping_timeout=15,
-                    open_timeout=30,
+                    open_timeout=open_timeout,
                 ) as ws:
+                    connected_once = True
                     _register_ws(ip, ws)
                     try:
                         await _recv_forward_loop(ip, ws)
@@ -239,7 +282,9 @@ async def _driver_connection_loop(ip: str) -> None:
                     n = _unreachable_counts.get(ip, 0) + 1
                     _unreachable_counts[ip] = n
                     if n == 1 or (n % 30) == 0:
-                        print(f"[WS] driver {ip} unreachable, retry in 2s: {e} (x{n})")
+                        print(
+                            f"[WS] driver {ip} unreachable, retry in {retry_interval_s}s: {e} (x{n})"
+                        )
                 else:
                     print(f"[WS] driver {ip} session error: {e!r}")
                     traceback.print_exception(type(e), e, e.__traceback__)
diff --git a/src/settings.py b/src/settings.py
index f7a74a9..43ab9c9 100644
--- a/src/settings.py
+++ b/src/settings.py
@@ -57,6 +57,19 @@ class Settings(dict):
         # down (0 disables). Helps drivers that reconnect after seeing traffic on 8766.
         if 'wifi_driver_hello_interval_s' not in self:
             self['wifi_driver_hello_interval_s'] = 10.0
+        # Outbound WebSocket dial: total seconds to keep trying before first success
+        # (many devices booting at once need more than a short window).
+        if 'wifi_driver_connect_retry_window_s' not in self:
+            self['wifi_driver_connect_retry_window_s'] = 120.0
+        # Spread outbound dials 0..N s by device IP so six+ drivers do not all hit the AP at once.
+        if 'wifi_driver_connect_stagger_max_s' not in self:
+            self['wifi_driver_connect_stagger_max_s'] = 2.5
+        # TCP/WebSocket open timeout per attempt (seconds).
+        if 'wifi_driver_ws_open_timeout' not in self:
+            self['wifi_driver_ws_open_timeout'] = 45.0
+        # Pause between outbound WebSocket dial attempts (seconds).
+        if 'wifi_driver_connect_retry_interval_s' not in self:
+            self['wifi_driver_connect_retry_interval_s'] = 2.0
         # UART to ESP32 ESP-NOW bridge; default off (Wi-Fi drivers need no serial).
         if 'serial_enabled' not in self:
             self['serial_enabled'] = False