fix(containerd): prevent silent network failures from leaving containers unreachable

devenv/server-entrypoint.sh

@@ -1,6 +1,12 @@
 #!/bin/sh
 set -e
 
+# Clean up stale CNI state from previous runs. After a container restart the
+# cni0 bridge may linger with a zeroed MAC (00:00:00:00:00:00), causing the
+# bridge plugin to fail with "could not set bridge's mac: invalid argument".
+ip link delete cni0 2>/dev/null || true
+rm -rf /var/lib/cni/networks/* /var/lib/cni/results/* 2>/dev/null || true
+
 # Ensure IP forwarding and subnet MASQUERADE for CNI.
 sysctl -w net.ipv4.ip_forward=1 2>/dev/null || true
 iptables -t nat -C POSTROUTING -s 10.88.0.0/16 ! -o cni0 -j MASQUERADE 2>/dev/null || \

docker/server-entrypoint.sh

@@ -3,6 +3,12 @@ set -e
 
 MCP_IMAGE="${MCP_IMAGE:-docker.io/library/memoh-mcp:latest}"
 
+# ---- Clean up stale CNI state from previous runs ----
+# After a container restart the cni0 bridge may linger with a zeroed MAC
+# (00:00:00:00:00:00), causing "could not set bridge's mac: invalid argument".
+ip link delete cni0 2>/dev/null || true
+rm -rf /var/lib/cni/networks/* /var/lib/cni/results/* 2>/dev/null || true
+
 # ---- Ensure IP forwarding and subnet MASQUERADE for CNI ----
 sysctl -w net.ipv4.ip_forward=1 2>/dev/null || true
 iptables -t nat -C POSTROUTING -s 10.88.0.0/16 ! -o cni0 -j MASQUERADE 2>/dev/null || \

internal/containerd/network.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"strconv"
 	"strings"
@@ -51,19 +52,26 @@ func setupCNINetwork(ctx context.Context, task client.Task, containerID string,
 	}
 	result, err := cni.Setup(ctx, containerID, netnsPath)
 	if err != nil {
-		if !isDuplicateAllocationError(err) && !isVethExistsError(err) {
+		retryable := isDuplicateAllocationError(err) || isVethExistsError(err) || isBridgeMACError(err)
+		if !retryable {
 			return "", err
 		}
-		// Stale IPAM allocation or veth exists (e.g. after container restart with persisted
-		// /var/lib/cni). Remove may fail if the previous iptables/veth state
-		// is already gone; ignore the error so the retry Setup still runs.
+		if isBridgeMACError(err) {
+			// Stale bridge with zeroed MAC after container restart; delete it so
+			// the plugin can recreate a healthy one.
+			_ = exec.CommandContext(ctx, "ip", "link", "delete", "cni0").Run()
+		}
 		_ = cni.Remove(ctx, containerID, netnsPath)
 		result, err = cni.Setup(ctx, containerID, netnsPath)
 		if err != nil {
 			return "", err
 		}
 	}
-	return extractIP(result), nil
+	ip := extractIP(result)
+	if ip == "" {
+		return "", fmt.Errorf("cni setup returned no usable IP for %s", containerID)
+	}
+	return ip, nil
 }
 
 func extractIP(result *gocni.Result) string {
@@ -139,3 +147,13 @@ func isVethExistsError(err error) bool {
 	}
 	return strings.Contains(err.Error(), "already exists")
 }
+
+// isBridgeMACError returns true if the CNI bridge plugin failed because the
+// stale cni0 bridge has a zeroed MAC address (common after container restart).
+func isBridgeMACError(err error) bool {
+	if err == nil {
+		return false
+	}
+	msg := err.Error()
+	return strings.Contains(msg, "set bridge") && strings.Contains(msg, "mac")
+}

internal/handlers/containerd.go

@@ -247,15 +247,8 @@ func (h *ContainerdHandler) ensureContainerAndTask(ctx context.Context, containe
 	}
 	if len(tasks) > 0 {
 		if tasks[0].Status == ctr.TaskStatusRunning {
-			if netResult, netErr := h.service.SetupNetwork(ctx, ctr.NetworkSetupRequest{
-				ContainerID: containerID,
-				CNIBinDir:   h.cfg.CNIBinaryDir,
-				CNIConfDir:  h.cfg.CNIConfigDir,
-			}); netErr != nil {
-				h.logger.Warn("network re-setup failed for running task",
-					slog.String("container_id", containerID), slog.Any("error", netErr))
-			} else if netResult.IP != "" && h.manager != nil {
-				h.manager.SetContainerIP(botID, netResult.IP)
+			if err := h.setupNetworkOrFail(ctx, containerID, botID); err != nil {
+				return err
 			}
 			return nil
 		}
@@ -270,17 +263,37 @@ func (h *ContainerdHandler) ensureContainerAndTask(ctx context.Context, containe
 	if err := h.service.StartContainer(ctx, containerID, nil); err != nil {
 		return err
 	}
-	if netResult, netErr := h.service.SetupNetwork(ctx, ctr.NetworkSetupRequest{
-		ContainerID: containerID,
-		CNIBinDir:   h.cfg.CNIBinaryDir,
-		CNIConfDir:  h.cfg.CNIConfigDir,
-	}); netErr != nil {
-		h.logger.Warn("network setup failed, task kept running",
-			slog.String("container_id", containerID), slog.Any("error", netErr))
-	} else if netResult.IP != "" && h.manager != nil {
-		h.manager.SetContainerIP(botID, netResult.IP)
+	return h.setupNetworkOrFail(ctx, containerID, botID)
+}
+
+// setupNetworkOrFail attempts CNI network setup with one retry. Returns an error
+// if no usable IP is obtained — callers must not silently ignore this.
+func (h *ContainerdHandler) setupNetworkOrFail(ctx context.Context, containerID, botID string) error {
+	var lastErr error
+	for attempt := 0; attempt < 2; attempt++ {
+		netResult, err := h.service.SetupNetwork(ctx, ctr.NetworkSetupRequest{
+			ContainerID: containerID,
+			CNIBinDir:   h.cfg.CNIBinaryDir,
+			CNIConfDir:  h.cfg.CNIConfigDir,
+		})
+		if err != nil {
+			lastErr = err
+			h.logger.Warn("network setup attempt failed",
+				slog.String("container_id", containerID),
+				slog.Int("attempt", attempt+1),
+				slog.Any("error", err))
+			continue
+		}
+		if netResult.IP == "" {
+			lastErr = fmt.Errorf("network setup returned no IP for %s", containerID)
+			continue
+		}
+		if h.manager != nil {
+			h.manager.SetContainerIP(botID, netResult.IP)
+		}
+		return nil
 	}
-	return nil
+	return fmt.Errorf("network setup failed for container %s: %w", containerID, lastErr)
 }
 
 // botContainerID resolves container_id for a bot from the database.
@@ -967,20 +980,15 @@ func (h *ContainerdHandler) ReconcileContainers(ctx context.Context) {
 						slog.String("bot_id", botID), slog.Any("error", dbErr))
 				}
 			}
-			if netResult, netErr := h.service.SetupNetwork(ctx, ctr.NetworkSetupRequest{
-				ContainerID: containerID,
-				CNIBinDir:   h.cfg.CNIBinaryDir,
-				CNIConfDir:  h.cfg.CNIConfigDir,
-			}); netErr != nil {
-				h.logger.Warn("reconcile: network re-setup failed for running task",
+			if netErr := h.setupNetworkOrFail(ctx, containerID, botID); netErr != nil {
+				h.logger.Error("reconcile: network setup failed for running task, container unreachable",
 					slog.String("bot_id", botID),
 					slog.String("container_id", containerID),
 					slog.Any("error", netErr))
-			} else if netResult.IP != "" && h.manager != nil {
-				h.manager.SetContainerIP(botID, netResult.IP)
+			} else {
+				h.logger.Info("reconcile: container healthy",
+					slog.String("bot_id", botID), slog.String("container_id", containerID))
 			}
-			h.logger.Info("reconcile: container healthy",
-				slog.String("bot_id", botID), slog.String("container_id", containerID))
 			continue
 		}
 

internal/mcp/manager.go

@@ -116,34 +116,39 @@ func (m *Manager) SetContainerIP(botID, ip string) {
 }
 
 // recoverContainerIP attempts to restore the container IP by re-running CNI setup.
-// CNI plugins are idempotent - calling Setup again returns the existing IP allocation.
+// CNI plugins are idempotent — calling Setup again returns the existing IP allocation.
+// Retries up to 2 times to tolerate transient CNI failures (IPAM lock contention, etc.).
 func (m *Manager) recoverContainerIP(botID string) (string, error) {
 	ctx := context.Background()
 	containerID := m.containerID(botID)
 
-	// First check if container exists and get basic info
 	info, err := m.service.GetContainer(ctx, containerID)
 	if err != nil {
 		return "", err
 	}
 
-	// Check if IP is stored in labels (if we ever add label persistence)
 	if ip, ok := info.Labels["mcp.container_ip"]; ok {
 		return ip, nil
 	}
 
-	// Container exists but IP not cached - need to re-setup network to get IP
-	// This happens after server restart when in-memory cache is lost
-	netResult, err := m.service.SetupNetwork(ctx, ctr.NetworkSetupRequest{
-		ContainerID: containerID,
-		CNIBinDir:   m.cfg.CNIBinaryDir,
-		CNIConfDir:  m.cfg.CNIConfigDir,
-	})
-	if err != nil {
-		return "", fmt.Errorf("network setup for IP recovery: %w", err)
+	const maxAttempts = 2
+	var lastErr error
+	for i := 0; i < maxAttempts; i++ {
+		netResult, err := m.service.SetupNetwork(ctx, ctr.NetworkSetupRequest{
+			ContainerID: containerID,
+			CNIBinDir:   m.cfg.CNIBinaryDir,
+			CNIConfDir:  m.cfg.CNIConfigDir,
+		})
+		if err != nil {
+			lastErr = err
+			m.logger.Warn("IP recovery attempt failed",
+				slog.String("bot_id", botID), slog.Int("attempt", i+1), slog.Any("error", err))
+			time.Sleep(time.Duration(i+1) * 500 * time.Millisecond)
+			continue
+		}
+		return netResult.IP, nil
 	}
-
-	return netResult.IP, nil
+	return "", fmt.Errorf("network setup for IP recovery after %d attempts: %w", maxAttempts, lastErr)
 }
 
 // MCPClient returns a gRPC client for the given bot's container.
@@ -295,12 +300,14 @@ func (m *Manager) Start(ctx context.Context, botID string) error {
 		}
 		return err
 	}
-	if netResult.IP != "" {
-		m.mu.Lock()
-		m.containerIPs[botID] = netResult.IP
-		m.mu.Unlock()
-		m.logger.Info("container network ready", slog.String("bot_id", botID), slog.String("ip", netResult.IP))
+	if netResult.IP == "" {
+		if stopErr := m.service.StopContainer(ctx, m.containerID(botID), &ctr.StopTaskOptions{Force: true}); stopErr != nil {
+			m.logger.Warn("cleanup: stop task failed", slog.String("container_id", m.containerID(botID)), slog.Any("error", stopErr))
+		}
+		return fmt.Errorf("network setup returned no IP for bot %s", botID)
 	}
+	m.SetContainerIP(botID, netResult.IP)
+	m.logger.Info("container network ready", slog.String("bot_id", botID), slog.String("ip", netResult.IP))
 	return nil
 }
 

internal/mcp/mcpclient/client.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"io"
 	"sync"
+	"time"
 
 	"google.golang.org/grpc"
 	"google.golang.org/grpc/connectivity"
@@ -20,11 +21,14 @@ import (
 	pb "github.com/memohai/memoh/internal/mcp/mcpcontainer"
 )
 
+const connectingTimeout = 30 * time.Second
+
 // Client wraps a gRPC connection to a single MCP container.
 type Client struct {
-	conn   *grpc.ClientConn
-	svc    pb.ContainerServiceClient
-	target string
+	conn      *grpc.ClientConn
+	svc       pb.ContainerServiceClient
+	target    string
+	createdAt time.Time
 }
 
 // NewClientFromConn wraps an existing gRPC connection into a Client.
@@ -47,9 +51,10 @@ func Dial(_ context.Context, ip string) (*Client, error) {
 		return nil, fmt.Errorf("grpc dial %s: %w", target, err)
 	}
 	return &Client{
-		conn:   conn,
-		svc:    pb.NewContainerServiceClient(conn),
-		target: target,
+		conn:      conn,
+		svc:       pb.NewContainerServiceClient(conn),
+		target:    target,
+		createdAt: time.Now(),
 	}, nil
 }
 
@@ -308,12 +313,14 @@ func (p *Pool) MCPClient(ctx context.Context, botID string) (*Client, error) {
 }
 
 // Get returns a cached client or dials a new one.
-// Stale connections (Shutdown / TransientFailure) are evicted automatically.
+// Stale connections (Shutdown / TransientFailure / stuck Connecting) are evicted automatically.
 func (p *Pool) Get(ctx context.Context, botID string) (*Client, error) {
 	p.mu.RLock()
 	if c, ok := p.clients[botID]; ok {
 		state := c.conn.GetState()
-		if state != connectivity.Shutdown && state != connectivity.TransientFailure {
+		stale := state == connectivity.Shutdown || state == connectivity.TransientFailure ||
+			(state == connectivity.Connecting && time.Since(c.createdAt) > connectingTimeout)
+		if !stale {
 			p.mu.RUnlock()
 			return c, nil
 		}