diff --git a/src/HalModule.cpp b/src/HalModule.cpp
index 8667d9a..6633fb5 100644
--- a/src/HalModule.cpp
+++ b/src/HalModule.cpp
@@ -67,7 +67,7 @@ HalModule::HalModule(
       _eepromManager{eepromManager}, _logger{logger} {}
 
 bool HalModule::rtw_hal_init(SelectedChannel selectedChannel) {
-  auto status = rtl8812au_hal_init();
+  auto status = rtl8812au_hal_init(selectedChannel.Channel);
 
   if (status) {
     _radioManagementModule->init_hw_mlme_ext(selectedChannel);
@@ -79,7 +79,7 @@ bool HalModule::rtw_hal_init(SelectedChannel selectedChannel) {
   return status;
 }
 
-bool HalModule::rtl8812au_hal_init() {
+bool HalModule::rtl8812au_hal_init(uint8_t init_channel) {
   // Check if MAC has already power on. by tynli. 2011.05.27.
   auto value8 = _device.rtw_read8(REG_SYS_CLKR + 1);
   auto regCr = _device.rtw_read8(REG_CR);
@@ -319,14 +319,23 @@ bool HalModule::rtl8812au_hal_init() {
     _device.phy_set_bb_reg(rTxPath_Jaguar, bMaskLWord, 0x1111);
   }
 
-  if (registry_priv::channel <= 14) {
+  /* Init directly at the user's selected channel so we only run the
+   * per-rate TX-power write loop once. The redundant prior pattern
+   * (init at registry_priv::channel = 36, then re-channel-set to the
+   * user's channel from `init_hw_mlme_ext`) wedged 8821AU at ch100
+   * mid-second-channel-set TX-power loop — the chip stopped ACK'ing
+   * USB control transfers after 2 BB writes, leaving the demo
+   * deadlocked on libusb_control_transfer until SIGKILL. See kaeru
+   * cite "8821AU ch100 wedge — confirmed second-channel-set is the
+   * trigger, not band-switch 2026-06-02". */
+  if (init_channel <= 14) {
     _radioManagementModule->PHY_SwitchWirelessBand8812(BandType::BAND_ON_2_4G);
   } else {
     _radioManagementModule->PHY_SwitchWirelessBand8812(BandType::BAND_ON_5G);
   }
 
   _radioManagementModule->rtw_hal_set_chnl_bw(
-      registry_priv::channel, ChannelWidth_t::CHANNEL_WIDTH_20,
+      init_channel, ChannelWidth_t::CHANNEL_WIDTH_20,
       HAL_PRIME_CHNL_OFFSET_DONT_CARE, HAL_PRIME_CHNL_OFFSET_DONT_CARE);
 
   // HW SEQ CTRL
diff --git a/src/HalModule.h b/src/HalModule.h
index eaa7373..9fc8083 100644
--- a/src/HalModule.h
+++ b/src/HalModule.h
@@ -59,7 +59,7 @@ class HalModule {
   bool rtw_hal_init(SelectedChannel selectedChannel);
 
 private:
-  bool rtl8812au_hal_init();
+  bool rtl8812au_hal_init(uint8_t init_channel);
   bool InitPowerOn();
   bool InitLLTTable8812A(uint8_t txpktbuf_bndy);
   bool InitLLTTable8814A();
diff --git a/src/RadioManagementModule.cpp b/src/RadioManagementModule.cpp
index e18d1f1..fe2abb1 100644
--- a/src/RadioManagementModule.cpp
+++ b/src/RadioManagementModule.cpp
@@ -906,6 +906,29 @@ uint32_t RadioManagementModule::phy_get_tx_bb_swing_8812a(BandType Band,
 }
 
 void RadioManagementModule::init_hw_mlme_ext(SelectedChannel pmlmeext) {
+  /* If `HalModule::rtl8812au_hal_init` already programmed the same
+   * (channel, bw, offset) AND the chip is an 8821AU, skip the
+   * reset-and-redo. The second channel-set wedges 8821AU at ch100
+   * mid-TX-power loop (chip stops ACK'ing USB control transfers
+   * after 2 BB writes — see kaeru cite "8821AU ch100 wedge —
+   * confirmed second-channel-set is the trigger, not band-switch
+   * 2026-06-02"). Other chips (8812AU, 8814AU) tolerate the second
+   * pass; some 8814AU init steps appear to depend on it (gating the
+   * skip caused 8814 ch6 to lose its RX loop entry).
+   *
+   * Falls through to the historical reset+redo path for everything
+   * else. */
+  bool same_target =
+      (_currentChannel == pmlmeext.Channel) &&
+      (_currentChannelBw == pmlmeext.ChannelWidth) &&
+      (_cur40MhzPrimeSc == pmlmeext.ChannelOffset) &&
+      (current_band_type != BandType::BAND_MAX);
+  bool is_8821 = _eepromManager->version_id.ICType == CHIP_8821;
+  if (same_target && is_8821) {
+    Set_HW_VAR_ENABLE_RX_BAR(true);
+    return;
+  }
+
   /* Modify to make sure first time change channel(band) would be done properly
    */
   _currentChannel = 0;