fix: support non-Quectel Qualcomm modems and fix 410 dongle AT timeouts (#40)

1. Vendor-neutral modem compatibility:
   - Discovery switched from a vendor-ID whitelist to detecting the QMI
     channel directly (an interface bound to the kernel qmi_wwan driver),
     so SIMCom, Sierra, Telit and other Qualcomm-based modules are found
     automatically while MBIM-only devices stay excluded
   - AT port responses now distinguish an AT command error from firmware
     incompatibility: ERROR / +CME ERROR is returned as a normal response
     (200) instead of being folded into a 502, which only a real transport
     failure produces

2. Fixed the 410 dongle's AT command timeouts:
   - Default WWAN AT port switched from wwan0at0 to wwan0at1: ModemManager
     marks the first AT port that answers its probe as primary (at1 on the
     tested UFI dongles) and closes AT ports once initialization finishes,
     so at1 is the responsive, idle channel for vocat while MM uses the
     QMI port for control
   - Drain the WWAN input buffer before each command write, discarding the
     late bytes of a previous timed-out command so they cannot pollute the
     next response's parsing
   - AT+CGSN now uses an independent short timeout instead of inheriting
     the refresh's 30s deadline (on MHI modems it returns the IMEI line
     but never a final OK). Previously every refresh held the device lock
     for the full 30s, queueing AT terminal commands behind it for 10-20s
   - The QMI UIM ICCID fallback only runs when AT+CPIN? already proved a
     READY card, so a SIM-less slot no longer blocks refresh waiting out
     its long timeout

Tests: added WWAN drain cleanup, drain-before-write ordering, CGSN timeout
bound, skip-QMI-ICCID-without-SIM, CommandError-as-200, WWAN at1 port
selection and vendor-neutral discovery cases. go vet and go test ./... pass.

Co-authored-by: Test <[email protected]>

internal/device/snapshot.go

@@ -69,7 +69,12 @@ func (manager *Manager) readSnapshot(
 	if ccidErr != nil {
 		ccid, ccidErr = manager.command(ctx, client, "AT+QCCID")
 	}
-	if ccidErr != nil && strings.EqualFold(strings.TrimSpace(backend), "qmi") && isNativeQMICandidate(candidate) {
+	if ccidErr != nil && strings.EqualFold(strings.TrimSpace(backend), "qmi") && isNativeQMICandidate(candidate) &&
+		strings.EqualFold(strings.TrimSpace(snapshot.SIMStatus), "READY") {
+		// Without a READY SIM the QMI UIM ICCID read blocks until its (long)
+		// timeout, and every refresh holds the device lock while it does so,
+		// starving the AT terminal. Only fall back to QMI when the AT CPIN
+		// probe already proved a card is present.
 		qmiContext, cancelQMI := manager.withTimeout(ctx, manager.commandTimeout*5)
 		qmiICCID, qmiErr := manager.readNativeQMIICCID(qmiContext, candidate)
 		cancelQMI()
@@ -186,14 +191,18 @@ func (manager *Manager) readSnapshot(
 		snapshot.RegistrationSource = "COPS"
 	}
 	if snapshot.IMEI == "" {
-		response, ok := optional("AT+CGSN")
-		if ok {
-			snapshot.IMEI = parseIdentifier(
-				response,
-				[]string{"+CGSN:", "+GSN:"},
-				14,
-				17,
-			)
+		// AT+CGSN on some MHI modems (the UFI dongle behind the OpenStick 410)
+		// returns the IMEI line but never a final OK, so it would block until the
+		// caller's deadline (30s during a periodic refresh) and starve every other
+		// device operation behind the lock. Give it an independent short timeout
+		// and let the WWAN transport's drain discard the trailing stale bytes.
+		cgsnCtx, cancelCGSN := context.WithTimeout(ctx, manager.commandTimeout)
+		cgsnResponse, cgsnErr := manager.command(cgsnCtx, client, "AT+CGSN")
+		cancelCGSN()
+		if cgsnErr == nil {
+			if imei := parseIdentifier(cgsnResponse, []string{"+CGSN:", "+GSN:"}, 14, 17); imei != "" {
+				snapshot.IMEI = imei
+			}
 		}
 	}
 	if snapshot.IMEI == "" && strings.EqualFold(strings.TrimSpace(backend), "qmi") && isNativeQMICandidate(candidate) {
@@ -211,7 +220,6 @@ func (manager *Manager) readSnapshot(
 		// Preserve a prior successful read across a transient QMI/AT failure.
 		snapshot.IMEI = previousSnapshot.IMEI
 	}
-
 	if response, ok := optional("AT+CFUN?"); ok {
 		if mode, found := parseCFUN(response); found {
 			snapshot.OperatingMode = mode

internal/device/snapshot_guard_test.go

@@ -0,0 +1,142 @@
+package device
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"vocat/internal/modem"
+)
+
+// lenientATClient answers every command with a bare CommandError and records
+// the commands it saw. It lets snapshot tests exercise the full readSnapshot
+// sequence without enumerating every step of the transcript.
+type lenientATClient struct {
+	mu        sync.Mutex
+	commands  []string
+	cgsnDelay time.Duration
+	cgsnIMEI  string
+}
+
+func (c *lenientATClient) Execute(ctx context.Context, command string) (modem.Response, error) {
+	c.mu.Lock()
+	c.commands = append(c.commands, command)
+	c.mu.Unlock()
+	if command == "ATI" {
+		return okResponse("Qualcomm", "PCIe/MHI WWAN modem", "Revision: native-410"), nil
+	}
+	if command == "AT+CGSN" && c.cgsnDelay > 0 {
+		select {
+		case <-time.After(c.cgsnDelay):
+		case <-ctx.Done():
+		}
+	}
+	if command == "AT+CGSN" && c.cgsnIMEI != "" {
+		return okResponse("+CGSN: " + c.cgsnIMEI), nil
+	}
+	return modem.Response{}, &modem.CommandError{Command: command, Final: "ERROR"}
+}
+
+func (c *lenientATClient) WaitURC(context.Context, func(string) bool) (string, error) {
+	return "", errors.New("no URC")
+}
+
+func (c *lenientATClient) Close() error { return nil }
+
+func (c *lenientATClient) saw(command string) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	for _, seen := range c.commands {
+		if seen == command {
+			return true
+		}
+	}
+	return false
+}
+
+// AT+CGSN on some MHI modems returns the IMEI line but never a final OK, so it
+// would block until the caller's deadline and hold the device lock for the
+// whole periodic refresh. The snapshot must bound CGSN with its own short
+// timeout instead of inheriting the refresh deadline.
+func TestManagerRefreshBoundsCGSNTimeout(t *testing.T) {
+	client := &lenientATClient{cgsnDelay: 5 * time.Second}
+	manager, id := newStartedTestManager(t, client)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 4*time.Second)
+	defer cancel()
+	start := time.Now()
+	snapshot, err := manager.Refresh(ctx, id)
+	elapsed := time.Since(start)
+
+	if err != nil {
+		t.Fatalf("Refresh: %v", err)
+	}
+	// CGSN times out after CommandTimeout (1s in the test manager); the rest
+	// of the snapshot is immediate. An un-bounded CGSN would wait for the
+	// 4s outer deadline (or worse, a real 30s refresh deadline).
+	if elapsed > 3*time.Second {
+		t.Fatalf("Refresh took %s; CGSN was not bounded by CommandTimeout", elapsed)
+	}
+	if !client.saw("AT+CGSN") {
+		t.Fatalf("CGSN was never sent; commands = %v", client.commands)
+	}
+	if snapshot.IMEI != "" {
+		t.Fatalf("IMEI = %q, want empty after CGSN timeout", snapshot.IMEI)
+	}
+}
+
+// A missing SIM must not fall back to the QMI UIM ICCID read: without a READY
+// card that call blocks until its long timeout and starves the AT terminal
+// behind the device lock.
+func TestManagerRefreshSkipsQMIICCIDWithoutReadySIM(t *testing.T) {
+	// CGSN succeeds so the snapshot does not fall back to the QMI DMS IMEI
+	// read either; the test focuses on the UIM ICCID fallback being skipped
+	// without a READY card.
+	client := &lenientATClient{cgsnIMEI: "866241014372802"}
+	manager, err := NewManager(Options{
+		Discoverer: staticDiscoverer{candidates: []modem.Candidate{{
+			ID:               "mhi-wwan0",
+			Product:          "PCIe/MHI WWAN modem",
+			QMIControl:       "/dev/wwan0qmi0",
+			NetworkInterface: "wwan0",
+			ATPort:           modem.Port{Path: "/dev/wwan0at0", Name: "wwan0at0", Role: modem.PortRoleAT},
+		}}},
+		Opener: &staticOpener{client: client},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := manager.Start(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = manager.Stop(context.Background()) })
+
+	qmiCalls := 0
+	manager.qmiRadioOpener = func(context.Context, string) (qmiRadioSession, error) {
+		qmiCalls++
+		return nil, errors.New("QMI should not be opened without a SIM")
+	}
+	if err := manager.SetBackend("mhi-wwan0", "qmi"); err != nil {
+		t.Fatal(err)
+	}
+
+	snapshot, err := manager.Refresh(context.Background(), "mhi-wwan0")
+	if err != nil {
+		t.Fatalf("Refresh: %v", err)
+	}
+	// Exactly one QMI open is expected: the immutable DMS IMEI read runs
+	// unconditionally for native QMI candidates (IMEI is hardware identity,
+	// independent of the card). The UIM ICCID fallback, which would block
+	// without a READY SIM, must be skipped.
+	if qmiCalls != 1 {
+		t.Fatalf("qmiRadioOpener called %d times, want 1 (DMS IMEI only, UIM ICCID must be skipped without a READY SIM)", qmiCalls)
+	}
+	for _, warning := range snapshot.Warnings {
+		if strings.Contains(warning, "QMI UIM") {
+			t.Fatalf("unexpected QMI ICCID warning: %q", warning)
+		}
+	}
+}

internal/modem/discovery.go

@@ -12,11 +12,9 @@ import (
 )
 
 const (
-	quectelVendorID = "2c7c"
-	djiVendorID     = "2ca3"
-	dji4GProductID  = "4006"
+	djiVendorID    = "2ca3"
+	dji4GProductID = "4006"
 )
-
 type SysFSDiscoverer struct {
 	SysRoot string
 	DevRoot string
@@ -44,10 +42,19 @@ func (d *SysFSDiscoverer) Discover(ctx context.Context) ([]Candidate, error) {
 		if os.IsNotExist(err) {
 			entries = nil
 		} else {
-			return nil, fmt.Errorf("discover Quectel USB devices: %w", err)
+			return nil, fmt.Errorf("discover USB QMI modems: %w", err)
 		}
 	}
 
+	// Candidate modems are identified by kernel driver binding instead of a
+	// vendor-ID whitelist. qmi_wwan only binds Qualcomm QMI control interfaces,
+	// so any USB device with a bound interface exposes a live QMI channel. This
+	// keeps discovery vendor-neutral (SIMCom, Sierra, Telit and other
+	// Qualcomm-based modems are found automatically) while MBIM-only devices
+	// stay out, because cdc_mbim binds their control interface instead and the
+	// project has no MBIM backend.
+	qmiBound := d.qmiWWANBoundDevices()
+
 	aliases := readSerialAliases(filepath.Join(d.DevRoot, "serial", "by-id"))
 	devices := make(map[string]*discoveredUSBDevice)
 	for _, entry := range entries {
@@ -75,7 +82,7 @@ func (d *SysFSDiscoverer) Discover(ctx context.Context) ([]Candidate, error) {
 		}
 		vendorID := strings.ToLower(readTrimmed(filepath.Join(resolvedDevice, "idVendor")))
 		productID := strings.ToLower(readTrimmed(filepath.Join(resolvedDevice, "idProduct")))
-		if !isSupportedUSBModem(vendorID, productID) {
+		if _, bound := qmiBound[deviceName]; !bound && !IsDJI4GUSB(vendorID, productID) {
 			continue
 		}
 
@@ -84,7 +91,7 @@ func (d *SysFSDiscoverer) Discover(ctx context.Context) ([]Candidate, error) {
 			serialNumber := readTrimmed(filepath.Join(resolvedDevice, "serial"))
 			state = &discoveredUSBDevice{
 				candidate: Candidate{
-					ID:           candidateID(productID, serialNumber, deviceName),
+					ID:           candidateID(vendorID, productID, serialNumber, deviceName),
 					VendorID:     vendorID,
 					ProductID:    productID,
 					Manufacturer: readTrimmed(filepath.Join(resolvedDevice, "manufacturer")),
@@ -134,6 +141,14 @@ func (d *SysFSDiscoverer) Discover(ctx context.Context) ([]Candidate, error) {
 		})
 		assignQuectelPortRoles(state.candidate.Ports)
 		state.candidate.ATPort = selectATPort(state.candidate.Ports)
+		if !state.candidate.HasATPort() {
+			// A bound QMI interface proves the modem is alive, but the snapshot,
+			// SMS, USSD and eSIM (AT+CSIM) paths all require an AT port. A missing
+			// ttyUSB/ttyACM node almost always means the option/qcserial driver
+			// does not claim the serial interfaces (often a missing PID in its
+			// device-ID table), not that the module lacks an AT interface.
+			state.candidate.DiscoveryIssue = "at_port_missing"
+		}
 		result = append(result, state.candidate)
 	}
 	wwanCandidates, err := d.discoverWWAN(ctx)
@@ -145,11 +160,6 @@ func (d *SysFSDiscoverer) Discover(ctx context.Context) ([]Candidate, error) {
 	return result, nil
 }
 
-func isSupportedUSBModem(vendorID, productID string) bool {
-	return strings.EqualFold(strings.TrimSpace(vendorID), quectelVendorID) ||
-		IsDJI4GUSB(vendorID, productID)
-}
-
 // IsDJI4GUSB reports whether a USB identity belongs to the first-generation
 // DJI/Baiwang 4G module. It keeps the factory 2ca3:4006 identity usable without
 // requiring a persistent AT+QCFG USB identity rewrite to Quectel 2c7c:0125.
@@ -252,7 +262,7 @@ func (d *SysFSDiscoverer) discoverWWAN(ctx context.Context) ([]Candidate, error)
 			Ports: group.ports, NetworkInterface: selectWWANNetworkInterface(d.SysRoot, group.index),
 		}
 		if len(group.ports) > 0 {
-			candidate.ATPort = group.ports[0]
+			candidate.ATPort = selectWWANATPort(group.ports)
 		}
 		if len(group.qmiNames) > 0 {
 			candidate.QMIControl = filepath.Join(d.DevRoot, group.qmiNames[0])
@@ -263,6 +273,20 @@ func (d *SysFSDiscoverer) discoverWWAN(ctx context.Context) ([]Candidate, error)
 	return result, nil
 }
 
+// selectWWANATPort prefers the secondary AT port (…at1) over the primary
+// (…at0) when both exist, falling back to the first AT port otherwise. Some
+// Qualcomm MHI modems (notably the UFI dongle behind the OpenStick 410) answer
+// on at1 immediately while at0 delays every response by 10-20 seconds, so the
+// secondary port is the usable AT channel.
+func selectWWANATPort(ports []Port) Port {
+	for _, port := range ports {
+		if port.InterfaceNumber == 1 {
+			return port
+		}
+	}
+	return ports[0]
+}
+
 func parseWWANPortName(name string) (index, kind string, portIndex int, ok bool) {
 	if !strings.HasPrefix(name, "wwan") {
 		return "", "", 0, false
@@ -402,7 +426,34 @@ func readSerialAliases(root string) map[string]string {
 	return result
 }
 
-func candidateID(productID, serialNumber, usbName string) string {
+// qmiWWANBoundDevices returns the set of USB device paths (for example "1-6"
+// or the hub-attached "1-4.3.2") that currently have at least one interface
+// bound to the kernel's qmi_wwan driver. Interface entries in the driver
+// directory are named "<device-path>:<interface>.<altsetting>", so the part
+// before the first colon is the owning USB device. The qmi_wwan driver only
+// binds Qualcomm QMI control interfaces, so membership doubles as a vendor-
+// neutral "this is a live QMI modem" signal.
+func (d *SysFSDiscoverer) qmiWWANBoundDevices() map[string]struct{} {
+	driverRoot := filepath.Join(d.SysRoot, "bus", "usb", "drivers", "qmi_wwan")
+	entries, err := os.ReadDir(driverRoot)
+	if err != nil {
+		return nil
+	}
+	devices := make(map[string]struct{})
+	for _, entry := range entries {
+		// The driver directory also holds control files (bind, unbind, uevent,
+		// module, new_id, ...); only names containing a colon are interfaces.
+		deviceName, _, ok := strings.Cut(entry.Name(), ":")
+		if !ok || deviceName == "" {
+			continue
+		}
+		devices[deviceName] = struct{}{}
+	}
+	return devices
+}
+
+func candidateID(vendorID, productID, serialNumber, usbName string) string {
+	prefix := "usb-" + sanitizeID(vendorID)
 	serialNumber = strings.TrimSpace(serialNumber)
 	if serialNumber != "" && !strings.EqualFold(serialNumber, "android") {
 		// A surprising number of EC20/EC25 carrier boards expose the same
@@ -411,9 +462,9 @@ func candidateID(productID, serialNumber, usbName string) string {
 		// to the same hub into one entry.  Include the physical USB topology in the
 		// discovery key; configured devices remain stable through ATMapper's
 		// USB-path/IMEI matching even when Linux renumbers ttyUSB nodes.
-		return "quectel-" + sanitizeID(serialNumber+"-"+usbName)
+		return prefix + "-" + sanitizeID(serialNumber+"-"+usbName)
 	}
-	return "quectel-" + sanitizeID(productID+"-"+usbName)
+	return prefix + "-" + sanitizeID(productID+"-"+usbName)
 }
 
 func sanitizeID(value string) string {

internal/modem/discovery_common_test.go

@@ -2,24 +2,24 @@ package modem
 
 import "testing"
 
-func TestSupportedUSBModemIdentity(t *testing.T) {
+func TestIsDJI4GUSBIdentity(t *testing.T) {
 	tests := []struct {
 		name      string
 		vendorID  string
 		productID string
 		want      bool
 	}{
-		{name: "Quectel", vendorID: "2c7c", productID: "0125", want: true},
 		{name: "DJI 4G module", vendorID: "2ca3", productID: "4006", want: true},
 		{name: "DJI 4G module uppercase", vendorID: "2CA3", productID: "4006", want: true},
 		{name: "unrelated DJI device", vendorID: "2ca3", productID: "001f", want: false},
+		{name: "Quectel identity", vendorID: "2c7c", productID: "0125", want: false},
 		{name: "unrelated USB device", vendorID: "0403", productID: "6001", want: false},
 	}
 
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			if got := isSupportedUSBModem(test.vendorID, test.productID); got != test.want {
-				t.Fatalf("isSupportedUSBModem(%q, %q) = %v, want %v", test.vendorID, test.productID, got, test.want)
+			if got := IsDJI4GUSB(test.vendorID, test.productID); got != test.want {
+				t.Fatalf("IsDJI4GUSB(%q, %q) = %v, want %v", test.vendorID, test.productID, got, test.want)
 			}
 		})
 	}

internal/modem/discovery_test.go

@@ -8,6 +8,7 @@ import (
 	"os"
 	"path/filepath"
 	"strconv"
+	"strings"
 	"testing"
 )
 
@@ -47,6 +48,7 @@ func TestSysFSDiscoverySelectsInterface04AndNeverInterface02(t *testing.T) {
 	}
 	mustMkdir(t, filepath.Join(usbRoot, "1-6:1.0", "net", "enx001122334455"))
 	mustMkdir(t, filepath.Join(usbRoot, "1-6:1.4", "usbmisc", "cdc-wdm0"))
+	mustBindQMIWWAN(t, sysRoot, "1-6:1.4")
 
 	discoverer := NewSysFSDiscoverer(sysRoot, devRoot)
 	candidates, err := discoverer.Discover(context.Background())
@@ -57,7 +59,7 @@ func TestSysFSDiscoverySelectsInterface04AndNeverInterface02(t *testing.T) {
 		t.Fatalf("got %d candidates, want 1", len(candidates))
 	}
 	candidate := candidates[0]
-	if candidate.ID != "quectel-0125-1-6" {
+	if candidate.ID != "usb-2c7c-0125-1-6" {
 		t.Fatalf("ID = %q", candidate.ID)
 	}
 	if candidate.ATPort.Name != "ttyUSB2" {
@@ -101,6 +103,7 @@ func TestSysFSDiscoverySelectsTTYUSB2InQMIInterface00Layout(t *testing.T) {
 	)
 	mustMkdir(t, filepath.Join(usbRoot, "1-6:1.4", "usbmisc", "cdc-wdm0"))
 	mustMkdir(t, filepath.Join(usbRoot, "1-6:1.4", "net", "wwp0s20f0u6i4"))
+	mustBindQMIWWAN(t, sysRoot, "1-6:1.4")
 
 	candidates, err := NewSysFSDiscoverer(sysRoot, devRoot).Discover(context.Background())
 	if err != nil {
@@ -146,6 +149,7 @@ func TestSysFSDiscoverySelectsATPortForSecondQMIUSBModem(t *testing.T) {
 		}
 		mustWrite(t, filepath.Join(usbRoot, modem.usbName+":1.4", "bInterfaceNumber"), "04\n")
 		mustMkdir(t, filepath.Join(usbRoot, modem.usbName+":1.4", "usbmisc", modem.wdm))
+		mustBindQMIWWAN(t, sysRoot, modem.usbName+":1.4")
 	}
 
 	candidates, err := NewSysFSDiscoverer(sysRoot, devRoot).Discover(context.Background())
@@ -194,6 +198,7 @@ func TestSysFSDiscoveryDoesNotCollapseModemsWithSharedFactorySerial(t *testing.T
 			mustMkdir(t, filepath.Join(usbRoot, interfaceName, tty, "tty", tty))
 		}
 		mustMkdir(t, filepath.Join(usbRoot, item.usbName+":1.4", "usbmisc", fmt.Sprintf("cdc-wdm%d", index)))
+		mustBindQMIWWAN(t, sysRoot, item.usbName+":1.4")
 	}
 
 	candidates, err := NewSysFSDiscoverer(sysRoot, devRoot).Discover(context.Background())
@@ -216,9 +221,11 @@ func TestSysFSDiscoveryDoesNotCollapseModemsWithSharedFactorySerial(t *testing.T
 	}
 }
 
-func TestSysFSDiscoveryIgnoresNonQuectelUSB(t *testing.T) {
+func TestSysFSDiscoveryIgnoresUSBWithoutQMIWWANBinding(t *testing.T) {
 	root := t.TempDir()
 	usbRoot := filepath.Join(root, "sys", "bus", "usb", "devices")
+	// A plain USB serial adapter (FTDI) exposes ttyUSB but no QMI interface and
+	// is never bound to qmi_wwan, so it must not be treated as a modem.
 	mustWrite(t, filepath.Join(usbRoot, "2-1", "idVendor"), "0403\n")
 	mustWrite(t, filepath.Join(usbRoot, "2-1:1.0", "bInterfaceNumber"), "00\n")
 	mustMkdir(t, filepath.Join(usbRoot, "2-1:1.0", "ttyUSB9"))
@@ -235,6 +242,117 @@ func TestSysFSDiscoveryIgnoresNonQuectelUSB(t *testing.T) {
 	}
 }
 
+func TestSysFSDiscoveryFindsNonQuectelVendorBoundToQMIWWAN(t *testing.T) {
+	root := t.TempDir()
+	sysRoot := filepath.Join(root, "sys")
+	devRoot := filepath.Join(root, "dev")
+	usbRoot := filepath.Join(sysRoot, "bus", "usb", "devices")
+	// A Sierra EM7430 flashed to its QMI (rmnet0) composition: non-Quectel
+	// vendor, but its control interface is bound to qmi_wwan.
+	mustWrite(t, filepath.Join(usbRoot, "1-3", "idVendor"), "1199\n")
+	mustWrite(t, filepath.Join(usbRoot, "1-3", "idProduct"), "9077\n")
+	mustWrite(t, filepath.Join(usbRoot, "1-3", "manufacturer"), "Sierra Wireless, Incorporated\n")
+	mustWrite(t, filepath.Join(usbRoot, "1-3", "product"), "EM7430\n")
+	for number, tty := range []string{"ttyUSB0", "ttyUSB1", "ttyUSB2", "ttyUSB3"} {
+		interfaceName := "1-3:1." + strconv.Itoa(number)
+		mustWrite(t, filepath.Join(usbRoot, interfaceName, "bInterfaceNumber"), fmt.Sprintf("%02x\n", number))
+		mustMkdir(t, filepath.Join(usbRoot, interfaceName, tty, "tty", tty))
+	}
+	mustMkdir(t, filepath.Join(usbRoot, "1-3:1.4", "usbmisc", "cdc-wdm0"))
+	mustBindQMIWWAN(t, sysRoot, "1-3:1.4")
+
+	candidates, err := NewSysFSDiscoverer(sysRoot, devRoot).Discover(context.Background())
+	if err != nil {
+		t.Fatalf("Discover: %v", err)
+	}
+	if len(candidates) != 1 {
+		t.Fatalf("got %d candidates, want 1", len(candidates))
+	}
+	candidate := candidates[0]
+	if candidate.VendorID != "1199" || candidate.Product != "EM7430" {
+		t.Fatalf("candidate = %#v", candidate)
+	}
+	if candidate.ID != "usb-1199-9077-1-3" {
+		t.Fatalf("ID = %q", candidate.ID)
+	}
+	if candidate.ATPort.Role != PortRoleAT {
+		t.Fatalf("AT port = %#v", candidate.ATPort)
+	}
+	if candidate.QMIControl != filepath.Join(devRoot, "cdc-wdm0") {
+		t.Fatalf("QMI control = %q", candidate.QMIControl)
+	}
+	if candidate.DiscoveryIssue != "" {
+		t.Fatalf("discovery issue = %q, want none", candidate.DiscoveryIssue)
+	}
+}
+
+func TestSysFSDiscoveryMarksQMIModemWithoutATPort(t *testing.T) {
+	root := t.TempDir()
+	sysRoot := filepath.Join(root, "sys")
+	devRoot := filepath.Join(root, "dev")
+	usbRoot := filepath.Join(sysRoot, "bus", "usb", "devices")
+	// QMI control interface is bound, but no ttyUSB/ttyACM node exists (for
+	// example the option/qcserial driver does not claim the serial interfaces).
+	mustWrite(t, filepath.Join(usbRoot, "1-7", "idVendor"), "2c7c\n")
+	mustWrite(t, filepath.Join(usbRoot, "1-7", "idProduct"), "0125\n")
+	mustMkdir(t, filepath.Join(usbRoot, "1-7:1.4", "usbmisc", "cdc-wdm0"))
+	mustBindQMIWWAN(t, sysRoot, "1-7:1.4")
+
+	candidates, err := NewSysFSDiscoverer(sysRoot, devRoot).Discover(context.Background())
+	if err != nil {
+		t.Fatalf("Discover: %v", err)
+	}
+	if len(candidates) != 1 {
+		t.Fatalf("got %d candidates, want 1", len(candidates))
+	}
+	candidate := candidates[0]
+	if candidate.DiscoveryIssue != "at_port_missing" {
+		t.Fatalf("discovery issue = %q, want at_port_missing", candidate.DiscoveryIssue)
+	}
+	if candidate.HasATPort() {
+		t.Fatalf("candidate unexpectedly has an AT port: %#v", candidate.ATPort)
+	}
+	if candidate.QMIControl != filepath.Join(devRoot, "cdc-wdm0") {
+		t.Fatalf("QMI control = %q", candidate.QMIControl)
+	}
+}
+
+func TestSysFSDiscoveryFindsHubAttachedQMIWWANDevice(t *testing.T) {
+	root := t.TempDir()
+	sysRoot := filepath.Join(root, "sys")
+	devRoot := filepath.Join(root, "dev")
+	usbRoot := filepath.Join(sysRoot, "bus", "usb", "devices")
+	// Device behind a hub: the USB path "1-4.3.2" contains extra segments, and
+	// the qmi_wwan binding uses the same composite path before the colon.
+	mustWrite(t, filepath.Join(usbRoot, "1-4.3.2", "idVendor"), "2c7c\n")
+	mustWrite(t, filepath.Join(usbRoot, "1-4.3.2", "idProduct"), "0125\n")
+	for number, tty := range []string{"ttyUSB0", "ttyUSB1", "ttyUSB2", "ttyUSB3"} {
+		interfaceName := "1-4.3.2:1." + strconv.Itoa(number)
+		mustWrite(t, filepath.Join(usbRoot, interfaceName, "bInterfaceNumber"), fmt.Sprintf("%02x\n", number))
+		mustMkdir(t, filepath.Join(usbRoot, interfaceName, tty, "tty", tty))
+	}
+	mustMkdir(t, filepath.Join(usbRoot, "1-4.3.2:1.4", "usbmisc", "cdc-wdm0"))
+	mustBindQMIWWAN(t, sysRoot, "1-4.3.2:1.4")
+
+	candidates, err := NewSysFSDiscoverer(sysRoot, devRoot).Discover(context.Background())
+	if err != nil {
+		t.Fatalf("Discover: %v", err)
+	}
+	if len(candidates) != 1 {
+		t.Fatalf("got %d candidates, want 1", len(candidates))
+	}
+	candidate := candidates[0]
+	if candidate.ATPort.Name != "ttyUSB2" {
+		t.Fatalf("AT port = %#v, want ttyUSB2", candidate.ATPort)
+	}
+	if candidate.QMIControl != filepath.Join(devRoot, "cdc-wdm0") {
+		t.Fatalf("QMI control = %q", candidate.QMIControl)
+	}
+	if !strings.Contains(candidate.ID, "1-4-3-2") {
+		t.Fatalf("ID = %q, want hub topology in discovery key", candidate.ID)
+	}
+}
+
 func TestSysFSDiscoveryFindsPCIeMHIWWANWithoutUSBBus(t *testing.T) {
 	root := t.TempDir()
 	sysRoot := filepath.Join(root, "sys")
@@ -256,7 +374,7 @@ func TestSysFSDiscoveryFindsPCIeMHIWWANWithoutUSBBus(t *testing.T) {
 	if candidate.ID != "mhi-wwan0" || candidate.HardwareKind != "wwan" {
 		t.Fatalf("identity = %#v", candidate)
 	}
-	if candidate.ATPort.Path != filepath.Join(devRoot, "wwan0at0") || candidate.ATPort.Role != PortRoleAT {
+	if candidate.ATPort.Path != filepath.Join(devRoot, "wwan0at1") || candidate.ATPort.Role != PortRoleAT {
 		t.Fatalf("AT port = %#v", candidate.ATPort)
 	}
 	if candidate.QMIControl != filepath.Join(devRoot, "wwan0qmi0") {
@@ -290,6 +408,23 @@ func TestSysFSDiscoveryFindsWWANFromDevNodesWithoutClassDirectory(t *testing.T)
 	}
 }
 
+func TestSelectWWANATPortPrefersSecondaryATPort(t *testing.T) {
+	ports := []Port{
+		{Name: "wwan0at0", InterfaceNumber: 0, Role: PortRoleAT},
+		{Name: "wwan0at1", InterfaceNumber: 1, Role: PortRoleAT},
+	}
+	if got := selectWWANATPort(ports); got.Name != "wwan0at1" {
+		t.Fatalf("selectWWANATPort = %#v, want wwan0at1", got)
+	}
+}
+
+func TestSelectWWANATPortFallsBackToPrimaryWhenOnlyAT0(t *testing.T) {
+	ports := []Port{{Name: "wwan0at0", InterfaceNumber: 0, Role: PortRoleAT}}
+	if got := selectWWANATPort(ports); got.Name != "wwan0at0" {
+		t.Fatalf("selectWWANATPort = %#v, want wwan0at0", got)
+	}
+}
+
 func TestParseWWANPortName(t *testing.T) {
 	for _, test := range []struct {
 		name, index, kind string
@@ -323,3 +458,18 @@ func mustMkdir(t *testing.T, path string) {
 		t.Fatal(err)
 	}
 }
+
+// mustBindQMIWWAN mimics the kernel's driver-binding directory entry: it adds
+// interfaceName (e.g. "1-6:1.4") under /sys/bus/usb/drivers/qmi_wwan exactly
+// like the real qmi_wwan driver directory does for a bound QMI interface.
+func mustBindQMIWWAN(t *testing.T, sysRoot, interfaceName string) {
+	t.Helper()
+	driverDir := filepath.Join(sysRoot, "bus", "usb", "drivers", "qmi_wwan")
+	if err := os.MkdirAll(driverDir, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	target := filepath.Join(sysRoot, "bus", "usb", "devices", interfaceName)
+	if err := os.Symlink(target, filepath.Join(driverDir, interfaceName)); err != nil {
+		t.Fatal(err)
+	}
+}

internal/modem/session.go

@@ -148,14 +148,19 @@ func (session *Session) executeLocked(ctx context.Context, command string) (Resp
 	if err := ctx.Err(); err != nil {
 		return response, err
 	}
-	if err := writeAll(session.transport, []byte(command+"\r")); err != nil {
-		session.poisonLocked()
-		return response, fmt.Errorf("write %s: %w", command, err)
-	}
+	// Drain the transport before writing the command. Serial transports wait
+	// for any pending output here (a no-op after a synchronous command), while
+	// WWAN transports discard bytes left over from a previous command that
+	// timed out; without this, a late reply (e.g. a slow CGSN response) would
+	// be mis-parsed as this command's output.
 	if err := drainTransport(ctx, session.transport); err != nil {
 		session.poisonLocked()
 		return response, fmt.Errorf("drain %s: %w", command, err)
 	}
+	if err := writeAll(session.transport, []byte(command+"\r")); err != nil {
+		session.poisonLocked()
+		return response, fmt.Errorf("write %s: %w", command, err)
+	}
 	return session.readFinalLocked(ctx, started, command, "", response)
 }
 

internal/modem/session_test.go

@@ -453,6 +453,62 @@ func TestSessionExecutePromptRejectsUnsafeInput(t *testing.T) {
 	}
 }
 
+// drainOrderTransport forwards to an inner Transport while recording
+// transport-level events, so a test can assert the exact order of Drain and
+// Write calls.
+type drainOrderTransport struct {
+	inner  Transport
+	events chan string
+}
+
+func (transport *drainOrderTransport) Write(payload []byte) (int, error) {
+	transport.events <- "write:" + string(payload)
+	return transport.inner.Write(payload)
+}
+
+func (transport *drainOrderTransport) Read(buffer []byte) (int, error) {
+	return transport.inner.Read(buffer)
+}
+
+func (transport *drainOrderTransport) Drain() error {
+	transport.events <- "drain"
+	return transport.inner.Drain()
+}
+
+func (transport *drainOrderTransport) ResetInputBuffer() error {
+	return transport.inner.ResetInputBuffer()
+}
+
+func (transport *drainOrderTransport) SetReadTimeout(timeout time.Duration) error {
+	return transport.inner.SetReadTimeout(timeout)
+}
+
+func (transport *drainOrderTransport) Close() error {
+	return transport.inner.Close()
+}
+
+// WWAN transports discard stale bytes left over from a timed-out command
+// inside Drain, so the session must call it before writing the next command;
+// otherwise a late reply (e.g. a slow CGSN response) would be mis-parsed as
+// the new command's output.
+func TestSessionDrainsBeforeWritingCommand(t *testing.T) {
+	inner := &transcriptTransport{steps: []transportStep{{
+		write:  "AT+CSQ\r",
+		chunks: []string{"\r\n+CSQ: 24,99\r\nOK\r\n"},
+	}}}
+	events := make(chan string, 8)
+	session := newTestSession(t, &drainOrderTransport{inner: inner, events: events})
+	if _, err := session.Execute(context.Background(), "AT+CSQ"); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+	if first := <-events; first != "drain" {
+		t.Fatalf("first transport event = %q, want drain before the command write", first)
+	}
+	if second := <-events; second != "write:AT+CSQ\r" {
+		t.Fatalf("second transport event = %q, want the command write", second)
+	}
+}
+
 func newTestSession(t *testing.T, transport Transport) *Session {
 	t.Helper()
 	session, err := NewSession(transport, SessionOptions{

internal/modem/wwan_at_linux.go

@@ -108,8 +108,27 @@ func (transport *nativeWWANATTransport) Drain() error {
 		return io.ErrClosedPipe
 	}
 	// WWAN character-device writes are handed to the modem synchronously and
-	// have no termios output queue to drain.
-	return nil
+	// have no termios output queue to drain. A previous command that timed out
+	// can leave late bytes in the input buffer (e.g. a slow CGSN reply that
+	// arrives after the command deadline); discard them here so the next
+	// command starts from a clean stream instead of mis-parsing stale output.
+	buffer := make([]byte, 4096)
+	for {
+		fds := []unix.PollFd{{Fd: int32(transport.fd), Events: unix.POLLIN}}
+		ready, err := unix.Poll(fds, 0)
+		if err != nil {
+			return err
+		}
+		if ready == 0 || fds[0].Revents&unix.POLLIN == 0 {
+			return nil
+		}
+		if _, err := unix.Read(transport.fd, buffer); err != nil {
+			if errors.Is(err, unix.EINTR) || errors.Is(err, unix.EAGAIN) {
+				continue
+			}
+			return err
+		}
+	}
 }
 
 func (transport *nativeWWANATTransport) ResetInputBuffer() error {

internal/modem/wwan_at_linux_test.go

@@ -0,0 +1,78 @@
+//go:build linux
+
+package modem
+
+import (
+	"errors"
+	"io"
+	"testing"
+
+	"golang.org/x/sys/unix"
+)
+
+// TestNativeWWANATTransportDrainDiscardsPendingBytes verifies Drain discards
+// every byte already buffered on the transport. A command that timed out (e.g.
+// AT+CGSN on an MHI modem that never answers OK) can leave its late reply in
+// the input buffer; the next command's Drain must clear it, however much data
+// is pending, before the session writes the new command.
+func TestNativeWWANATTransportDrainDiscardsPendingBytes(t *testing.T) {
+	readFD, writeFD := socketpair(t)
+	defer unix.Close(writeFD)
+
+	// More than one 4096-byte Drain read: a slow CGSN reply (echo + IMEI +
+	// trailing CRLF) can exceed a single buffer.
+	payload := make([]byte, 12000)
+	for index := range payload {
+		payload[index] = byte('A' + index%26)
+	}
+	payload = append(payload, []byte("\r\n+CGSN: 357091089453326\r\n")...)
+	if _, err := unix.Write(writeFD, payload); err != nil {
+		t.Fatalf("seed stale bytes: %v", err)
+	}
+
+	transport := &nativeWWANATTransport{fd: readFD, readTimeout: -1}
+	if err := transport.Drain(); err != nil {
+		t.Fatalf("Drain: %v", err)
+	}
+	assertNoPendingBytes(t, readFD, "after Drain")
+
+	// Draining a clean transport is a fast no-op that must not block or error.
+	if err := transport.Drain(); err != nil {
+		t.Fatalf("second Drain: %v", err)
+	}
+}
+
+// TestNativeWWANATTransportDrainRejectsClosedTransport covers the guard that
+// keeps a poisoned session from draining a wedged, already-closed fd.
+func TestNativeWWANATTransportDrainRejectsClosedTransport(t *testing.T) {
+	readFD, writeFD := socketpair(t)
+	defer unix.Close(writeFD)
+	transport := &nativeWWANATTransport{fd: readFD, readTimeout: -1}
+	if err := transport.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	if err := transport.Drain(); !errors.Is(err, io.ErrClosedPipe) {
+		t.Fatalf("Drain after Close = %v, want ErrClosedPipe", err)
+	}
+}
+
+func socketpair(t *testing.T) (int, int) {
+	t.Helper()
+	fds, err := unix.Socketpair(unix.AF_UNIX, unix.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return fds[0], fds[1]
+}
+
+func assertNoPendingBytes(t *testing.T, fd int, context string) {
+	t.Helper()
+	fds := []unix.PollFd{{Fd: int32(fd), Events: unix.POLLIN}}
+	ready, err := unix.Poll(fds, 0)
+	if err != nil {
+		t.Fatalf("poll %s: %v", context, err)
+	}
+	if ready != 0 {
+		t.Fatalf("%s: fd still readable", context)
+	}
+}

internal/server/at_guard_test.go

@@ -1,6 +1,15 @@
 package server
 
-import "testing"
+import (
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"vocat/internal/device"
+	"vocat/internal/modem"
+)
 
 func TestValidateATCommandBlocksTrafficMessagingAndDialActions(t *testing.T) {
 	t.Parallel()
@@ -44,3 +53,62 @@ func TestValidateATCommandAllowsReadOnlyStatusQueries(t *testing.T) {
 		}
 	}
 }
+
+// The AT terminal must present ERROR / +CME ERROR as a normal response, not as
+// a 502. Before the CommandError branch was restored, every unsupported or
+// SIM-less command was folded into "the device operation failed", hiding the
+// real reason from the user.
+func TestHandleATSurfacesCommandErrorAsResponse(t *testing.T) {
+	controller := fakeDeviceController{
+		entry: device.Device{ID: "dev1"},
+		atHandler: func(command string) (modem.Response, error) {
+			return modem.Response{}, &modem.CommandError{
+				Command: command,
+				Final:   "+CME ERROR: 10",
+				Lines:   []string{"+CME ERROR: 10"},
+			}
+		},
+	}
+	server := &Server{devices: controller, logger: regionTestLogger(), maxRequestBodyBytes: 1 << 20}
+	recorder := httptest.NewRecorder()
+	request := httptest.NewRequest(
+		http.MethodPost,
+		"/api/devices/dev1/actions/at",
+		strings.NewReader(`{"cmd":"AT+CPIN?","timeout_ms":5000}`),
+	)
+	request.Header.Set("Content-Type", "application/json")
+
+	if !server.handleAT(recorder, request, "dev1") {
+		t.Fatal("handleAT returned false")
+	}
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200 (body=%s)", recorder.Code, recorder.Body.String())
+	}
+	data := decodeData(t, recorder)
+	response, _ := data["response"].(string)
+	if !strings.Contains(response, "+CME ERROR: 10") {
+		t.Fatalf("response = %q, want +CME ERROR text", response)
+	}
+}
+
+func TestHandleATMapsNonCommandErrorTo502(t *testing.T) {
+	controller := fakeDeviceController{
+		entry: device.Device{ID: "dev1"},
+		atErr: errors.New("transport wedged"),
+	}
+	server := &Server{devices: controller, logger: regionTestLogger(), maxRequestBodyBytes: 1 << 20}
+	recorder := httptest.NewRecorder()
+	request := httptest.NewRequest(
+		http.MethodPost,
+		"/api/devices/dev1/actions/at",
+		strings.NewReader(`{"cmd":"AT+CSQ"}`),
+	)
+	request.Header.Set("Content-Type", "application/json")
+
+	if !server.handleAT(recorder, request, "dev1") {
+		t.Fatal("handleAT returned false")
+	}
+	if recorder.Code != http.StatusBadGateway {
+		t.Fatalf("status = %d, want 502", recorder.Code)
+	}
+}

internal/server/device_api.go

@@ -1059,6 +1059,26 @@ func (s *Server) handleAT(w http.ResponseWriter, r *http.Request, id string) boo
 	defer cancel()
 	response, err := s.devices.ExecuteAT(ctx, id, command)
 	if err != nil {
+		var commandErr *modem.CommandError
+		if errors.As(err, &commandErr) {
+			// The modem answered with ERROR / +CME ERROR. An AT terminal must
+			// surface that text (including the CME detail) as a normal response;
+			// folding it into a 502 hides the real reason from the user.
+			text := strings.Join(commandErr.Lines, "\n")
+			if text != "" {
+				text += "\n"
+			}
+			text += commandErr.Final
+			writeJSON(w, http.StatusOK, map[string]any{
+				"data": map[string]any{
+					"response":    text,
+					"final":       commandErr.Final,
+					"duration_ms": 0,
+					"urcs":        []string{},
+				},
+			})
+			return true
+		}
 		s.writeDeviceError(w, err)
 		return true
 	}