fix(docs): remove corepack from docusaurus image build that is not available in latest Node

Co-authored-by: Miek Gieben <miek@miek.nl>

.github/workflows/go-tests.yml

@@ -14,12 +14,13 @@ on:
       - "**.proto"
       - "go.*"
       - "Makefile"
+      - "mise.toml"
       - "scripts/**"
       - "test/**"
 permissions:
   contents: read
 env:
-  MISE_VERSION: "2026.3.7"
+  MISE_VERSION: "2026.10.22"
 jobs:
   test:
     runs-on: ubuntu-latest
@@ -42,9 +43,9 @@ jobs:
 
       - name: Generate docs
         run: |
-          make cli-docs
+          mise run docs:cli
           git diff --exit-code ||
-            (echo "Documentation is out of date. Please run 'make cli-docs' and commit the changes." && exit 1)
+            (echo "Documentation is out of date. Please run 'mise run docs:cli' and commit the changes." && exit 1)
 
       - name: Run tests
         run: |
@@ -70,7 +71,7 @@ jobs:
           version: ${{ env.MISE_VERSION }}
 
       # TODO: figure out how to properly lock all platforms in the lockfile and remove this hack.
-      # `mise lock` (mise run mise:lock) doesn't generate checksums for some of the tools, e.g. protoc*
+      # `mise lock` doesn't generate checksums for some of the tools, e.g. protoc*
       - name: Reset changes in mise.lock after fresh mise install
         run: git checkout mise.lock
 

.github/workflows/lint.yml

@@ -14,7 +14,7 @@ on:
 permissions:
   contents: read
 env:
-  MISE_VERSION: "2026.3.7"
+  MISE_VERSION: "2026.10.22"
 jobs:
   lint:
     runs-on: ubuntu-latest
@@ -30,6 +30,6 @@ jobs:
         run: git checkout mise.lock
 
       - name: Lint
-        run: make lint
+        run: mise run lint
 
     timeout-minutes: 10

.github/workflows/nightly-build.yml

@@ -8,7 +8,7 @@ permissions:
   contents: write
 
 env:
-  MISE_VERSION: "2026.3.7"
+  MISE_VERSION: "2026.10.22"
 
 jobs:
   nightly:

.mockery.yml

@@ -1,4 +1,10 @@
 packages:
+  github.com/docker/docker/client:
+    config:
+      dir: internal/machine/docker
+      pkgname: docker
+    interfaces:
+      APIClient:
   github.com/psviderski/uncloud/internal/machine/caddyconfig:
     interfaces:
       CaddyfileValidator:

AGENTS.md

@@ -98,33 +98,6 @@ management.
 
 - **`misc/`**: Design documents and guides
 
-## Key Technologies
-
-### Core Dependencies
-
-```go
-// Networking and orchestration
-github.com/docker/docker // Docker API client
-github.com/docker/compose/v2 // Docker Compose integration
-golang.zx2c4.com/wireguard // WireGuard implementation
-github.com/hashicorp/serf // Gossip protocol
-
-// State management
-github.com/ipfs/go-ds-crdt // CRDT distributed storage
-github.com/dgraph-io/badger/v3 // Embedded database
-
-// Web proxy
-github.com/caddyserver/caddy/v2 // HTTP server and reverse proxy
-
-// CLI and UX
-github.com/spf13/cobra // CLI framework
-github.com/charmbracelet/huh // Interactive forms
-
-// gRPC and networking
-google.golang.org/grpc                     // gRPC framework
-github.com/siderolabs/grpc-proxy // gRPC proxy for forwarding
-```
-
 ## Development Workflow
 
 ### Build and Development
@@ -135,16 +108,18 @@ go build -o uc ./cmd/uc
 go build -o uncloudd ./cmd/uncloudd
 ```
 
-### Key Make Targets
+### Formatting and Linting
+
+- `mise run lint`: Check Go code and formatting without changing files
+- `mise run lint:fix`: Apply formatting and available lint fixes
+
+### Other Development Tasks
 
 - `proto`: Generate protobuf code
 - `ucind-cluster`: Create development cluster
 - `update-dev`: Deploy to development machines
 - `demo-reset`: Reset demo environment
-- `fmt`: Format code
 - `test`: Run all tests
-- `lint`: Lint the code using golangci-lint
-- `lint-and-fix`: Lint the code and fix issues whenever possible
 
 ## CLI Commands Structure
 

HACKING.md

@@ -35,10 +35,11 @@ Build the CLI:
 go build -o uc ./cmd/uc
 ```
 
-Or build and run the CLI with a single command:
+Or build and run the CLI with mise. It forwards arguments, including help flags, directly to the CLI:
 
 ```shell
-go run ./cmd/uc --help
+mise uc --help
+mise uc image rm -h
 ```
 
 The Uncloud daemon (`uncloudd`) only supports Linux, so you need to cross-compile it if you're developing on macOS or
@@ -99,10 +100,16 @@ mise run dev:reset user@host1 user@host2
 
 ## Linting and formatting
 
-Lint and format the code:
+Check the code and formatting without changing files:
 
 ```shell
-make lint-and-fix
+mise run lint
+```
+
+Apply formatting and available lint fixes:
+
+```shell
+mise run lint:fix
 ```
 
 ## Code generation
@@ -116,9 +123,31 @@ mise run proto
 Regenerate mocks:
 
 ```shell
-make mocks
+mise run mocks
 ```
 
+## Documentation
+
+Regenerate the CLI reference docs after changing commands or flags:
+
+```shell
+mise run docs:cli
+```
+
+To preview the documentation website, install its dependencies from the project root:
+
+```shell
+npm ci --prefix website
+```
+
+Then start the local development server:
+
+```shell
+mise run docs:dev
+```
+
+The server opens the site (http://localhost:3000/) in your browser and reloads it as you edit files in `website/`.
+
 ## Uncloud in Docker (UCinD)
 
 The [`ucind`](./cmd/ucind) CLI lets you run Uncloud clusters locally using Docker containers instead of real machines.

Makefile

@@ -21,10 +21,6 @@ corrosion-multiarch-image-push:
 ucind-multiarch-image-push:
 	docker buildx build --push --platform linux/amd64,linux/arm64 -t "$(UCIND_IMAGE)" --target ucind .
 
-.PHONY: mocks
-mocks:
-	@mockery
-
 .PHONY: test
 test:
 ifeq ($(TEST_NAME),)
@@ -51,25 +47,3 @@ test-clean:
 	else \
 		echo "Nothing to clean."; \
 	fi
-
-.PHONY: vet
-vet:
-	go vet ./...
-
-.PHONY: format fmt
-format fmt:
-	GOOS=linux golangci-lint fmt
-
-LINT_TARGETS := lint lint-and-fix
-.PHONY: $(LINT_TARGETS) _lint
-$(LINT_TARGETS): _lint
-lint: ARGS=
-lint-and-fix: ARGS=--fix
-_lint:
-# Explicitly set OS to Linux to not skip *_linux.go files when running on macOS.
-# Uncloud daemon won't likely support OS other than Linux anytime soon, so for now we can rely on that.
-	GOOS=linux golangci-lint run $(ARGS)
-
-.PHONY: cli-docs
-cli-docs:
-	go run ./cmd/uc docs

README.md

@@ -44,8 +44,8 @@ complexity of Kubernetes.
   [Uncloud DNS](https://github.com/psviderski/uncloud-dns) service.
 * **Automatic HTTPS**: Built-in Caddy reverse proxy handles TLS certificate provisioning and renewal using Let's
   Encrypt.
-* **Cluster storage for TLS certificates**: Native cluster storage for Caddy shares certificates and ACME challenge
+* **Cluster storage for TLS certificates**: [Cluster storage](https://github.com/unlabs-dev/caddy-uncloud) for Caddy
-  tokens across machines and coordinates issuance.
+  shares certificates and ACME challenge tokens across machines and coordinates issuance.
 * **Docker-like CLI**: Familiar commands for managing both infrastructure and applications.
 * **Remote management**: Control your entire infrastructure through SSH access to any single machine in the cluster.
 

api/pb/common.go

@@ -4,8 +4,16 @@ import (
 	"bytes"
 	"fmt"
 	"net/netip"
+
+	"google.golang.org/grpc/status"
 )
 
+// Err returns the upstream error, preserving its gRPC status code.
+// It returns nil for nil metadata or an absent or successful status.
+func (m *Metadata) Err() error {
+	return status.ErrorProto(m.GetStatus())
+}
+
 func NewIP(addr netip.Addr) *IP {
 	// MarshalBinary always returns a nil error.
 	ip, _ := addr.MarshalBinary()

api/pb/common.pb.go

@@ -87,7 +87,8 @@ type Metadata struct {
 	MachineName string `protobuf:"bytes,5,opt,name=machine_name,json=machineName,proto3" json:"machine_name,omitempty"`
 	// Address of the machine the response came from.
 	MachineAddr string `protobuf:"bytes,1,opt,name=machine_addr,json=machineAddr,proto3" json:"machine_addr,omitempty"`
-	// error is set if the request to upstream failed. The rest of the response is undefined.
+	// error is set if the request to upstream failed. The rest of the response apart from this metadata message
+	// is undefined.
 	Error string `protobuf:"bytes,2,opt,name=error,proto3" json:"error,omitempty"`
 	// error as a gRPC Status message.
 	Status *status.Status `protobuf:"bytes,3,opt,name=status,proto3" json:"status,omitempty"`

api/pb/common.proto

@@ -18,7 +18,8 @@ message Metadata {
   // Address of the machine the response came from.
   string machine_addr = 1;
 
-  // error is set if the request to upstream failed. The rest of the response is undefined.
+  // error is set if the request to upstream failed. The rest of the response apart from this metadata message
+  // is undefined.
   string error = 2;
   // error as a gRPC Status message.
   google.rpc.Status status = 3;

api/pb/docker.proto

@@ -24,6 +24,7 @@ service Docker {
   // Docker auth credentials if necessary.
   rpc InspectRemoteImage(InspectRemoteImageRequest) returns (InspectRemoteImageResponse);
   rpc ListImages(ListImagesRequest) returns (ListImagesResponse);
+  rpc RemoveImage(RemoveImageRequest) returns (RemoveImageResponse);
 
   rpc CreateVolume(CreateVolumeRequest) returns (CreateVolumeResponse);
   rpc ListVolumes(ListVolumesRequest) returns (ListVolumesResponse);
@@ -31,6 +32,8 @@ service Docker {
 
   rpc CreateServiceContainer(CreateServiceContainerRequest) returns (CreateContainerResponse);
   rpc InspectServiceContainer(InspectContainerRequest) returns (ServiceContainer);
+  // UpdateServiceContainer updates the requested properties of a service container in place without recreating it.
+  rpc UpdateServiceContainer(UpdateServiceContainerRequest) returns (google.protobuf.Empty);
   rpc ListServiceContainers(ListServiceContainersRequest) returns (ListServiceContainersResponse);
   rpc RemoveServiceContainer(RemoveContainerRequest) returns (google.protobuf.Empty);
 }
@@ -179,6 +182,40 @@ message ListImagesRequest {
   bytes options = 1;
 }
 
+message RemoveImageRequest {
+  string id = 1;
+  bool force = 2;
+  bool prune_children = 3;
+  repeated ImagePlatform platforms = 4;
+}
+
+// ImagePlatform describes an OCI image platform.
+message ImagePlatform {
+  string architecture = 1;
+  string os = 2;
+  string os_version = 3;
+  repeated string os_features = 4;
+  string variant = 5;
+}
+
+message RemoveImageResponse {
+  // Must contain only one repeated messages field to allow broadcasting RemoveImage requests to multiple machines.
+  repeated MachineImageRemoval messages = 1;
+}
+
+message MachineImageRemoval {
+  Metadata metadata = 1;
+  repeated ImageRemoval results = 2;
+}
+
+// ImageRemoval reports an image deletion or an untagged image reference.
+message ImageRemoval {
+  // The image ID of an image that was deleted.
+  string deleted = 1;
+  // The image reference of an image that was untagged.
+  string untagged = 2;
+}
+
 message ListImagesResponse {
   // Must contain only one repeated messages field to allow broadcasting ListImages requests to multiple machines.
   repeated MachineImages messages = 1;
@@ -246,6 +283,23 @@ message ServiceContainer {
   bytes service_spec = 2;
 }
 
+message UpdateServiceContainerRequest {
+  string id = 1;
+  // Fields that are not set are left unchanged. CPU and memory limits can be changed but not removed.
+  // CPU limit in nanocores.
+  optional int64 cpu = 2;
+  // Memory limit in bytes.
+  optional int64 memory = 3;
+  // Memory reservation in bytes.
+  optional int64 memory_reservation = 4;
+  // Service labels that replace all current service labels.
+  Labels service_labels = 5;
+}
+
+message Labels {
+  map<string, string> labels = 1;
+}
+
 message ListServiceContainersRequest {
   string service_id = 1;
   // JSON serialised container.ListOptions.

api/pb/docker_grpc.pb.go

@@ -32,11 +32,13 @@ const (
 	Docker_InspectImage_FullMethodName            = "/api.Docker/InspectImage"
 	Docker_InspectRemoteImage_FullMethodName      = "/api.Docker/InspectRemoteImage"
 	Docker_ListImages_FullMethodName              = "/api.Docker/ListImages"
+	Docker_RemoveImage_FullMethodName             = "/api.Docker/RemoveImage"
 	Docker_CreateVolume_FullMethodName            = "/api.Docker/CreateVolume"
 	Docker_ListVolumes_FullMethodName             = "/api.Docker/ListVolumes"
 	Docker_RemoveVolume_FullMethodName            = "/api.Docker/RemoveVolume"
 	Docker_CreateServiceContainer_FullMethodName  = "/api.Docker/CreateServiceContainer"
 	Docker_InspectServiceContainer_FullMethodName = "/api.Docker/InspectServiceContainer"
+	Docker_UpdateServiceContainer_FullMethodName  = "/api.Docker/UpdateServiceContainer"
 	Docker_ListServiceContainers_FullMethodName   = "/api.Docker/ListServiceContainers"
 	Docker_RemoveServiceContainer_FullMethodName  = "/api.Docker/RemoveServiceContainer"
 )
@@ -59,11 +61,14 @@ type DockerClient interface {
 	// Docker auth credentials if necessary.
 	InspectRemoteImage(ctx context.Context, in *InspectRemoteImageRequest, opts ...grpc.CallOption) (*InspectRemoteImageResponse, error)
 	ListImages(ctx context.Context, in *ListImagesRequest, opts ...grpc.CallOption) (*ListImagesResponse, error)
+	RemoveImage(ctx context.Context, in *RemoveImageRequest, opts ...grpc.CallOption) (*RemoveImageResponse, error)
 	CreateVolume(ctx context.Context, in *CreateVolumeRequest, opts ...grpc.CallOption) (*CreateVolumeResponse, error)
 	ListVolumes(ctx context.Context, in *ListVolumesRequest, opts ...grpc.CallOption) (*ListVolumesResponse, error)
 	RemoveVolume(ctx context.Context, in *RemoveVolumeRequest, opts ...grpc.CallOption) (*emptypb.Empty, error)
 	CreateServiceContainer(ctx context.Context, in *CreateServiceContainerRequest, opts ...grpc.CallOption) (*CreateContainerResponse, error)
 	InspectServiceContainer(ctx context.Context, in *InspectContainerRequest, opts ...grpc.CallOption) (*ServiceContainer, error)
+	// UpdateServiceContainer updates the requested properties of a service container in place without recreating it.
+	UpdateServiceContainer(ctx context.Context, in *UpdateServiceContainerRequest, opts ...grpc.CallOption) (*emptypb.Empty, error)
 	ListServiceContainers(ctx context.Context, in *ListServiceContainersRequest, opts ...grpc.CallOption) (*ListServiceContainersResponse, error)
 	RemoveServiceContainer(ctx context.Context, in *RemoveContainerRequest, opts ...grpc.CallOption) (*emptypb.Empty, error)
 }
@@ -217,6 +222,16 @@ func (c *dockerClient) ListImages(ctx context.Context, in *ListImagesRequest, op
 	return out, nil
 }
 
+func (c *dockerClient) RemoveImage(ctx context.Context, in *RemoveImageRequest, opts ...grpc.CallOption) (*RemoveImageResponse, error) {
+	cOpts := append([]grpc.CallOption{grpc.StaticMethod()}, opts...)
+	out := new(RemoveImageResponse)
+	err := c.cc.Invoke(ctx, Docker_RemoveImage_FullMethodName, in, out, cOpts...)
+	if err != nil {
+		return nil, err
+	}
+	return out, nil
+}
+
 func (c *dockerClient) CreateVolume(ctx context.Context, in *CreateVolumeRequest, opts ...grpc.CallOption) (*CreateVolumeResponse, error) {
 	cOpts := append([]grpc.CallOption{grpc.StaticMethod()}, opts...)
 	out := new(CreateVolumeResponse)
@@ -267,6 +282,16 @@ func (c *dockerClient) InspectServiceContainer(ctx context.Context, in *InspectC
 	return out, nil
 }
 
+func (c *dockerClient) UpdateServiceContainer(ctx context.Context, in *UpdateServiceContainerRequest, opts ...grpc.CallOption) (*emptypb.Empty, error) {
+	cOpts := append([]grpc.CallOption{grpc.StaticMethod()}, opts...)
+	out := new(emptypb.Empty)
+	err := c.cc.Invoke(ctx, Docker_UpdateServiceContainer_FullMethodName, in, out, cOpts...)
+	if err != nil {
+		return nil, err
+	}
+	return out, nil
+}
+
 func (c *dockerClient) ListServiceContainers(ctx context.Context, in *ListServiceContainersRequest, opts ...grpc.CallOption) (*ListServiceContainersResponse, error) {
 	cOpts := append([]grpc.CallOption{grpc.StaticMethod()}, opts...)
 	out := new(ListServiceContainersResponse)
@@ -305,11 +330,14 @@ type DockerServer interface {
 	// Docker auth credentials if necessary.
 	InspectRemoteImage(context.Context, *InspectRemoteImageRequest) (*InspectRemoteImageResponse, error)
 	ListImages(context.Context, *ListImagesRequest) (*ListImagesResponse, error)
+	RemoveImage(context.Context, *RemoveImageRequest) (*RemoveImageResponse, error)
 	CreateVolume(context.Context, *CreateVolumeRequest) (*CreateVolumeResponse, error)
 	ListVolumes(context.Context, *ListVolumesRequest) (*ListVolumesResponse, error)
 	RemoveVolume(context.Context, *RemoveVolumeRequest) (*emptypb.Empty, error)
 	CreateServiceContainer(context.Context, *CreateServiceContainerRequest) (*CreateContainerResponse, error)
 	InspectServiceContainer(context.Context, *InspectContainerRequest) (*ServiceContainer, error)
+	// UpdateServiceContainer updates the requested properties of a service container in place without recreating it.
+	UpdateServiceContainer(context.Context, *UpdateServiceContainerRequest) (*emptypb.Empty, error)
 	ListServiceContainers(context.Context, *ListServiceContainersRequest) (*ListServiceContainersResponse, error)
 	RemoveServiceContainer(context.Context, *RemoveContainerRequest) (*emptypb.Empty, error)
 	mustEmbedUnimplementedDockerServer()
@@ -358,6 +386,9 @@ func (UnimplementedDockerServer) InspectRemoteImage(context.Context, *InspectRem
 func (UnimplementedDockerServer) ListImages(context.Context, *ListImagesRequest) (*ListImagesResponse, error) {
 	return nil, status.Errorf(codes.Unimplemented, "method ListImages not implemented")
 }
+func (UnimplementedDockerServer) RemoveImage(context.Context, *RemoveImageRequest) (*RemoveImageResponse, error) {
+	return nil, status.Errorf(codes.Unimplemented, "method RemoveImage not implemented")
+}
 func (UnimplementedDockerServer) CreateVolume(context.Context, *CreateVolumeRequest) (*CreateVolumeResponse, error) {
 	return nil, status.Errorf(codes.Unimplemented, "method CreateVolume not implemented")
 }
@@ -373,6 +404,9 @@ func (UnimplementedDockerServer) CreateServiceContainer(context.Context, *Create
 func (UnimplementedDockerServer) InspectServiceContainer(context.Context, *InspectContainerRequest) (*ServiceContainer, error) {
 	return nil, status.Errorf(codes.Unimplemented, "method InspectServiceContainer not implemented")
 }
+func (UnimplementedDockerServer) UpdateServiceContainer(context.Context, *UpdateServiceContainerRequest) (*emptypb.Empty, error) {
+	return nil, status.Errorf(codes.Unimplemented, "method UpdateServiceContainer not implemented")
+}
 func (UnimplementedDockerServer) ListServiceContainers(context.Context, *ListServiceContainersRequest) (*ListServiceContainersResponse, error) {
 	return nil, status.Errorf(codes.Unimplemented, "method ListServiceContainers not implemented")
 }
@@ -591,6 +625,24 @@ func _Docker_ListImages_Handler(srv interface{}, ctx context.Context, dec func(i
 	return interceptor(ctx, in, info, handler)
 }
 
+func _Docker_RemoveImage_Handler(srv interface{}, ctx context.Context, dec func(interface{}) error, interceptor grpc.UnaryServerInterceptor) (interface{}, error) {
+	in := new(RemoveImageRequest)
+	if err := dec(in); err != nil {
+		return nil, err
+	}
+	if interceptor == nil {
+		return srv.(DockerServer).RemoveImage(ctx, in)
+	}
+	info := &grpc.UnaryServerInfo{
+		Server:     srv,
+		FullMethod: Docker_RemoveImage_FullMethodName,
+	}
+	handler := func(ctx context.Context, req interface{}) (interface{}, error) {
+		return srv.(DockerServer).RemoveImage(ctx, req.(*RemoveImageRequest))
+	}
+	return interceptor(ctx, in, info, handler)
+}
+
 func _Docker_CreateVolume_Handler(srv interface{}, ctx context.Context, dec func(interface{}) error, interceptor grpc.UnaryServerInterceptor) (interface{}, error) {
 	in := new(CreateVolumeRequest)
 	if err := dec(in); err != nil {
@@ -681,6 +733,24 @@ func _Docker_InspectServiceContainer_Handler(srv interface{}, ctx context.Contex
 	return interceptor(ctx, in, info, handler)
 }
 
+func _Docker_UpdateServiceContainer_Handler(srv interface{}, ctx context.Context, dec func(interface{}) error, interceptor grpc.UnaryServerInterceptor) (interface{}, error) {
+	in := new(UpdateServiceContainerRequest)
+	if err := dec(in); err != nil {
+		return nil, err
+	}
+	if interceptor == nil {
+		return srv.(DockerServer).UpdateServiceContainer(ctx, in)
+	}
+	info := &grpc.UnaryServerInfo{
+		Server:     srv,
+		FullMethod: Docker_UpdateServiceContainer_FullMethodName,
+	}
+	handler := func(ctx context.Context, req interface{}) (interface{}, error) {
+		return srv.(DockerServer).UpdateServiceContainer(ctx, req.(*UpdateServiceContainerRequest))
+	}
+	return interceptor(ctx, in, info, handler)
+}
+
 func _Docker_ListServiceContainers_Handler(srv interface{}, ctx context.Context, dec func(interface{}) error, interceptor grpc.UnaryServerInterceptor) (interface{}, error) {
 	in := new(ListServiceContainersRequest)
 	if err := dec(in); err != nil {
@@ -760,6 +830,10 @@ var Docker_ServiceDesc = grpc.ServiceDesc{
 			MethodName: "ListImages",
 			Handler:    _Docker_ListImages_Handler,
 		},
+		{
+			MethodName: "RemoveImage",
+			Handler:    _Docker_RemoveImage_Handler,
+		},
 		{
 			MethodName: "CreateVolume",
 			Handler:    _Docker_CreateVolume_Handler,
@@ -780,6 +854,10 @@ var Docker_ServiceDesc = grpc.ServiceDesc{
 			MethodName: "InspectServiceContainer",
 			Handler:    _Docker_InspectServiceContainer_Handler,
 		},
+		{
+			MethodName: "UpdateServiceContainer",
+			Handler:    _Docker_UpdateServiceContainer_Handler,
+		},
 		{
 			MethodName: "ListServiceContainers",
 			Handler:    _Docker_ListServiceContainers_Handler,

cmd/uc/caddy/cert/ls.go

@@ -17,6 +17,7 @@ import (
 	"github.com/psviderski/uncloud/internal/cli"
 	"github.com/psviderski/uncloud/internal/cli/completion"
 	"github.com/psviderski/uncloud/internal/cli/tui"
+	"github.com/psviderski/uncloud/internal/version"
 	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/psviderski/uncloud/pkg/client"
 	"github.com/spf13/cobra"
@@ -35,10 +36,13 @@ func NewListCommand() *cobra.Command {
 		Short:   "List certificates in cluster storage for Caddy.",
 		Long: `List certificates stored in the Uncloud cluster storage for Caddy.
 
-Caddy must use the Uncloud storage module (https://github.com/unlabs-dev/caddy-uncloud)
+Caddy uses cluster storage by default. Certificates don't appear here if Caddy uses local storage,
-configured with 'storage uncloud' in the global options for its certificates to appear here.
+for example, if it runs with a custom global config without 'storage uncloud' or was deployed with
+uc v0.21.0 or earlier and not upgraded.
 
-This inventory does not check whether Caddy currently serves or trusts a certificate.`,
+This inventory does not check whether Caddy currently serves or trusts a certificate.
+
+See ` + version.DocsURL + `/concepts/ingress/cluster-storage-for-caddy for details.`,
 		Args: cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, _ []string) error {
 			uncli := cmd.Context().Value("cli").(*cli.CLI)

cmd/uc/caddy/config.go

@@ -3,8 +3,8 @@ package caddy
 import (
 	"context"
 	"fmt"
-	"os"
 
+	"charm.land/lipgloss/v2"
 	"github.com/alecthomas/chroma/v2/quick"
 	"github.com/psviderski/uncloud/internal/cli"
 	"github.com/psviderski/uncloud/internal/cli/completion"
@@ -15,7 +15,6 @@ import (
 
 type configOptions struct {
 	machine string
-	noColor bool
 }
 
 func NewConfigCommand() *cobra.Command {
@@ -33,8 +32,6 @@ func NewConfigCommand() *cobra.Command {
 
 	cmd.Flags().StringVarP(&opts.machine, "machine", "m", "",
 		"Name or ID of the machine to get the configuration from. (default is connected machine)")
-	cmd.Flags().BoolVar(&opts.noColor, "no-color", false,
-		"Disable syntax highlighting for the output.")
 
 	completion.MachinesFlag(cmd)
 
@@ -54,13 +51,9 @@ func runConfig(ctx context.Context, uncli *cli.CLI, opts configOptions) error {
 	}
 
 	// Print the Caddyfile with syntax highlighting.
-	if opts.noColor {
+	if err = quick.Highlight(lipgloss.Writer, config.Caddyfile, "caddy", "terminal256", "monokai"); err != nil {
-		fmt.Print(config.Caddyfile)
-	} else {
-		if err = quick.Highlight(os.Stdout, config.Caddyfile, "caddy", "terminal256", "monokai"); err != nil {
 		// If highlighting fails, fall back to plain output.
-			fmt.Print(config.Caddyfile)
+		lipgloss.Print(config.Caddyfile)
-		}
 	}
 
 	if config.LastReconciliationError != "" {

cmd/uc/caddy/deploy.go

@@ -16,6 +16,7 @@ import (
 	"github.com/psviderski/uncloud/internal/cli"
 	"github.com/psviderski/uncloud/internal/cli/completion"
 	"github.com/psviderski/uncloud/internal/cli/tui"
+	"github.com/psviderski/uncloud/internal/version"
 	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/psviderski/uncloud/pkg/client"
 	"github.com/spf13/cobra"
@@ -25,6 +26,10 @@ type deployOptions struct {
 	caddyfile string
 	image     string
 	machines  []string
+
+	caddyfileChanged bool
+	imageChanged     bool
+	machinesChanged  bool
 }
 
 func NewDeployCommand() *cobra.Command {
@@ -33,21 +38,62 @@ func NewDeployCommand() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "deploy",
 		Short: "Deploy or upgrade Caddy reverse proxy across all machines in the cluster.",
-		Long: "Deploy or upgrade Caddy reverse proxy across all machines in the cluster.\n" +
+		Long: `Deploy or upgrade Caddy reverse proxy across all machines in the cluster.
-			"A rolling update is performed when updating existing containers to minimise disruption.",
+A rolling update is performed when updating existing containers to minimise disruption.
+
+By default, this deploys the latest version of the Uncloud Caddy image with a global config that enables cluster
+storage for TLS certificates. See https://github.com/unlabs-dev/caddy-uncloud for more details.
+
+Values set with --image, --caddyfile, and --machine are kept for later deployments. Pass an empty value to reset
+a flag to its default.
+
+If this command didn't deploy the current caddy service, for example, it was deployed with 'uc deploy'
+from a Compose file or an older version of uc, the deployment replaces its configuration.
+
+Requires Uncloud daemon ` + client.MinCaddyDaemonVersion + ` or newer on all machines.`,
+		Example: `  # Deploy Caddy or upgrade it to the latest version of the Uncloud Caddy image.
+  uc caddy deploy
+
+  # Deploy a custom image and keep it for later deployments.
+  uc caddy deploy --image ghcr.io/unlabs-dev/caddy-uncloud:0.1.3
+
+  # Go back to the latest version of the Uncloud Caddy image.
+  uc caddy deploy --image ""
+
+  # Deploy a custom image build and a custom global Caddy config and keep both for later deployments.
+  uc caddy deploy --image myorg/caddy-uncloud:latest --caddyfile global.Caddyfile
+
+  # Go back to the default global Caddy config but keep the custom image.
+  uc caddy deploy --caddyfile ""
+
+  # Deploy only to specific machines and keep them for later deployments.
+  uc caddy deploy --machine machine1,machine2
+
+  # Go back to deploying to all machines.
+  uc caddy deploy --machine ""`,
 		RunE: func(cmd *cobra.Command, args []string) error {
 			uncli := cmd.Context().Value("cli").(*cli.CLI)
+
+			opts.caddyfileChanged = cmd.Flag("caddyfile").Changed
+			opts.imageChanged = cmd.Flag("image").Changed
+			opts.machinesChanged = cmd.Flag("machine").Changed
+
 			return runDeploy(cmd.Context(), uncli, opts)
 		},
 	}
 
 	cmd.Flags().StringVar(&opts.caddyfile, "caddyfile", "",
-		"Path to a custom global Caddy config (Caddyfile) that will be prepended to the auto-generated Caddy config.")
+		"Path to a custom global Caddy config (Caddyfile) that will be prepended to the auto-generated Caddy config.\n"+
+			"An empty value resets it to the default config that enables cluster storage.\n"+
+			"(default is the current custom config or the default config)")
 	cmd.Flags().StringVar(&opts.image, "image", "",
-		"Caddy Docker image to deploy. (default caddy:LATEST_VERSION)")
+		"Caddy Docker image to deploy. It must include the Uncloud storage module if the Caddy config uses it.\n"+
+			"An empty value resets it to the latest version of the Uncloud Caddy image.\n"+
+			"(default is the current custom image or "+client.CaddyImage+":LATEST_VERSION)")
 	cmd.Flags().StringSliceVarP(&opts.machines, "machine", "m", nil,
-		"Machine names or IDs to deploy to. Can be specified multiple times or as a comma-separated "+
+		"Machine names or IDs to deploy to. Can be specified multiple times or as a comma-separated list.\n"+
-			"list. (default is all machines)")
+			"An empty value resets it to all machines.\n"+
+			"(default is the current machines or all machines)")
 
 	completion.MachinesFlag(cmd)
 
@@ -55,13 +101,13 @@ func NewDeployCommand() *cobra.Command {
 }
 
 func runDeploy(ctx context.Context, uncli *cli.CLI, opts deployOptions) error {
-	caddyfile := ""
+	config := ""
 	if opts.caddyfile != "" {
 		data, err := os.ReadFile(opts.caddyfile)
 		if err != nil {
 			return fmt.Errorf("read Caddyfile: %w", err)
 		}
-		caddyfile = strings.TrimSpace(string(data))
+		config = strings.TrimSpace(string(data))
 	}
 
 	clusterClient, err := uncli.ConnectCluster(ctx)
@@ -70,6 +116,10 @@ func runDeploy(ctx context.Context, uncli *cli.CLI, opts deployOptions) error {
 	}
 	defer clusterClient.Close()
 
+	if err = clusterClient.Caddy.CheckDaemonVersions(ctx); err != nil {
+		return err
+	}
+
 	svc, err := clusterClient.InspectService(ctx, client.CaddyServiceName)
 	if err != nil {
 		if !errors.Is(err, api.ErrNotFound) {
@@ -100,26 +150,48 @@ func runDeploy(ctx context.Context, uncli *cli.CLI, opts deployOptions) error {
 		}
 	}
 
-	if opts.image != "" {
+	current := svc.Spec()
-		fmt.Println(tui.Faint.Render("target image: ") + tui.FormatImage(opts.image, tui.Green))
+	deployOpts := resolveDeploymentOptions(current, opts, config)
+	configChanged := deployOpts.ConfigOrDefault() != current.CaddyConfig()
+
+	if svc.ID != "" && configChanged {
+		printCaddyConfig("current global Caddy config:", current.CaddyConfig())
 	}
 
 	fmt.Println()
 	fmt.Println("Preparing a deployment plan...")
 
-	placement := api.Placement{
+	d, err := clusterClient.Caddy.NewDeployment(ctx, deployOpts)
-		Machines: cli.ExpandCommaSeparatedValues(opts.machines),
-	}
-	d, err := clusterClient.Caddy.NewDeployment(ctx, client.CaddyDeploymentOptions{
-		Image: opts.image, Config: caddyfile, Placement: placement,
-	})
 	if err != nil {
 		return fmt.Errorf("create caddy deployment: %w", err)
 	}
+	target := d.Spec
 
-	if opts.image == "" {
+	// The Uncloud Caddy image resolves to the latest version allowed by this CLI, which is older than the running image
-		fmt.Println(tui.Faint.Render("target image: ") + tui.FormatImage(d.Spec.Container.Image,
+	// if a newer CLI deployed it. Continuing would downgrade the image and replace its default config.
-			tui.Green) + tui.Faint.Render(" (latest stable)"))
+	if deployOpts.Image == "" && client.IsNewerCaddyImage(current.Container.Image, target.Container.Image) {
+		return fmt.Errorf("caddy service runs image '%s' which is newer than the latest version '%s' supported by "+
+			"this uc version: upgrade uc or specify the image with --image",
+			current.Container.Image, target.Container.Image)
+	}
+
+	imageNote := " (latest)"
+	if deployOpts.Image != "" {
+		imageNote = " (custom)"
+		if !opts.imageChanged {
+			imageNote = " (custom, use --image \"\" for the latest version)"
+		}
+	}
+	fmt.Println(tui.Faint.Render("target image: ") + tui.FormatImage(target.Container.Image, tui.Green) +
+		tui.Faint.Render(imageNote))
+
+	if configChanged {
+		if deployOpts.Config != nil {
+			// The user provided this config so there is no need to print its content.
+			fmt.Println(tui.Faint.Render("target global Caddy config: ") + opts.caddyfile)
+		} else {
+			printCaddyConfig("target global Caddy config: (default)", deployOpts.ConfigOrDefault())
+		}
 	}
 
 	plan, err := d.Plan(ctx)
@@ -131,7 +203,7 @@ func runDeploy(ctx context.Context, uncli *cli.CLI, opts deployOptions) error {
 		fmt.Printf("%s service is up to date.\n", client.CaddyServiceName)
 	} else {
 		if svc.ID == "" {
-			if len(opts.machines) > 0 {
+			if len(deployOpts.Placement.Machines) > 0 {
 				fmt.Println("This will run a Caddy container on each selected machine.")
 			} else {
 				fmt.Println("This will run a Caddy container on each machine.")
@@ -163,6 +235,14 @@ func runDeploy(ctx context.Context, uncli *cli.CLI, opts deployOptions) error {
 		fmt.Println(summary)
 		fmt.Println()
 
+		warnings := deploymentWarnings(current, target, opts)
+		for _, warning := range warnings {
+			tui.PrintWarning(warning)
+		}
+		if len(warnings) > 0 {
+			fmt.Println()
+		}
+
 		title := "Proceed with deployment?"
 		// Include the direct connection or context name in the confirmation prompt to avoid accidentally
 		// deploying to the wrong cluster.
@@ -195,6 +275,98 @@ func runDeploy(ctx context.Context, uncli *cli.CLI, opts deployOptions) error {
 	return UpdateDomainRecords(ctx, clusterClient, uncli.ProgressOut())
 }
 
+// resolveDeploymentOptions returns the deployment options for the flags and the current service spec, which is zero if
+// the service doesn't exist. Each flag works on its own: a specified flag overrides the value and an empty one resets
+// it to the default. A flag that isn't specified keeps the explicitly set value.
+func resolveDeploymentOptions(
+	current api.ServiceSpec, opts deployOptions, config string,
+) client.CaddyDeploymentOptions {
+	var resolved client.CaddyDeploymentOptions
+
+	if opts.imageChanged {
+		resolved.Image = opts.image
+	} else if _, custom := current.Labels[api.LabelCaddyImageCustom]; custom {
+		resolved.Image = current.Container.Image
+	}
+
+	if opts.caddyfileChanged {
+		if opts.caddyfile != "" {
+			resolved.Config = new(config)
+		}
+	} else if _, custom := current.Labels[api.LabelCaddyConfigCustom]; custom {
+		resolved.Config = new(current.CaddyConfig())
+	}
+
+	if opts.machinesChanged {
+		resolved.Placement.Machines = cli.ExpandCommaSeparatedValues(opts.machines)
+	} else if _, managed := current.Labels[api.LabelCaddyManaged]; managed {
+		resolved.Placement = current.Placement
+	}
+
+	return resolved
+}
+
+// deploymentWarnings returns the consequences of deploying the target spec over the current one, which is zero if
+// the service doesn't exist.
+func deploymentWarnings(current, target api.ServiceSpec, opts deployOptions) []string {
+	var warnings []string
+	exists := current.Name != ""
+
+	if _, managed := current.Labels[api.LabelCaddyManaged]; exists && !managed {
+		warnings = append(warnings, "'uc caddy deploy' doesn't manage this caddy service yet, so this deployment "+
+			"replaces its configuration, including any custom global Caddy config.\n"+
+			"If you manage it with a Compose file, cancel and deploy with 'uc deploy' instead.")
+	}
+
+	currentStorage := client.CaddyStorageModule(current.CaddyConfig())
+	targetStorage := client.CaddyStorageModule(target.CaddyConfig())
+	if exists && currentStorage != targetStorage {
+		warning := fmt.Sprintf("Caddy certificate storage changes from %s to %s.\n"+
+			"Caddy doesn't automatically migrate certificates, so all of them will be reissued and HTTPS may fail "+
+			"for those domains until that finishes.\n"+
+			"To migrate them first, see %s/concepts/ingress/cluster-storage-for-caddy",
+			caddyStorageName(currentStorage), caddyStorageName(targetStorage), version.DocsURL)
+		if !opts.caddyfileChanged {
+			warning += "\nTo keep the current storage, specify a custom Caddyfile that uses it with --caddyfile."
+		}
+		warnings = append(warnings, warning)
+	}
+
+	imageOrConfigChanged := current.Container.Image != target.Container.Image ||
+		current.CaddyConfig() != target.CaddyConfig()
+	if targetStorage == client.CaddyStorageUncloud && imageOrConfigChanged &&
+		!client.IsDefaultCaddyImage(target.Container.Image) {
+		warnings = append(warnings, fmt.Sprintf("Image '%s' must include the Uncloud storage module to use "+
+			"'storage uncloud'. Otherwise, Caddy won't start or apply the new config.\n"+
+			"Check 'uc caddy config' after the deployment for errors.", target.Container.Image))
+	}
+
+	return warnings
+}
+
+// caddyStorageName returns a user-friendly name of the Caddy storage module.
+func caddyStorageName(module string) string {
+	switch module {
+	case client.CaddyStorageFileSystem:
+		return "local"
+	case client.CaddyStorageUncloud:
+		return "cluster"
+	default:
+		return "'" + module + "'"
+	}
+}
+
+// printCaddyConfig prints the title and the indented config or "none" if it's empty.
+func printCaddyConfig(title, config string) {
+	if config == "" {
+		lipgloss.Println(tui.Faint.Render(title) + " none")
+		return
+	}
+
+	lipgloss.Println(tui.Faint.Render(title))
+	fmt.Println(config)
+}
+
 func UpdateDomainRecords(ctx context.Context, clusterClient *client.Client, progressOut *streams.Out) error {
 	domain, err := clusterClient.GetDomain(ctx)
 	if err != nil {

cmd/uc/caddy/deploy_test.go

@@ -0,0 +1,263 @@
+package caddy
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/psviderski/uncloud/pkg/client"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestResolveDeploymentOptions(t *testing.T) {
+	t.Parallel()
+
+	const (
+		customImage  = "example.com/caddy:1.2.3"
+		customConfig = "{\n\tdebug\n}"
+	)
+	machines := api.Placement{Machines: []string{"m1", "m2"}}
+
+	// A service with a custom image, config, and machines that has the given labels.
+	service := func(labels ...string) api.ServiceSpec {
+		spec := api.ServiceSpec{
+			Name:      client.CaddyServiceName,
+			Caddy:     &api.CaddySpec{Config: customConfig},
+			Container: api.ContainerSpec{Image: customImage},
+			Placement: machines,
+		}
+		if len(labels) > 0 {
+			spec.Labels = make(map[string]string, len(labels))
+			for _, label := range labels {
+				spec.Labels[label] = ""
+			}
+		}
+		return spec
+	}
+
+	tests := []struct {
+		name    string
+		current api.ServiceSpec
+		opts    deployOptions
+		config  string
+		want    client.CaddyDeploymentOptions
+	}{
+		{
+			name: "no service gets the defaults",
+		},
+		{
+			name:    "service that isn't managed gets the defaults",
+			current: service(),
+		},
+		{
+			name:    "managed service follows the default image and config and keeps machines",
+			current: service(api.LabelCaddyManaged),
+			want:    client.CaddyDeploymentOptions{Placement: machines},
+		},
+		{
+			name:    "custom image is kept and config follows the default",
+			current: service(api.LabelCaddyManaged, api.LabelCaddyImageCustom),
+			want:    client.CaddyDeploymentOptions{Image: customImage, Placement: machines},
+		},
+		{
+			name:    "custom config is kept and image follows the latest version",
+			current: service(api.LabelCaddyManaged, api.LabelCaddyConfigCustom),
+			want:    client.CaddyDeploymentOptions{Config: new(customConfig), Placement: machines},
+		},
+		{
+			name:    "custom image and config are kept",
+			current: service(api.LabelCaddyManaged, api.LabelCaddyImageCustom, api.LabelCaddyConfigCustom),
+			want:    client.CaddyDeploymentOptions{Image: customImage, Config: new(customConfig), Placement: machines},
+		},
+		{
+			name: "custom empty config is kept",
+			current: api.ServiceSpec{
+				Name:   client.CaddyServiceName,
+				Labels: map[string]string{api.LabelCaddyManaged: "", api.LabelCaddyConfigCustom: ""},
+			},
+			want: client.CaddyDeploymentOptions{Config: new("")},
+		},
+		{
+			name:    "image flag overrides the current image",
+			current: service(api.LabelCaddyManaged, api.LabelCaddyImageCustom),
+			opts:    deployOptions{image: "caddy:2", imageChanged: true},
+			want:    client.CaddyDeploymentOptions{Image: "caddy:2", Placement: machines},
+		},
+		{
+			name:    "empty image flag resets to the latest version",
+			current: service(api.LabelCaddyManaged, api.LabelCaddyImageCustom),
+			opts:    deployOptions{imageChanged: true},
+			want:    client.CaddyDeploymentOptions{Placement: machines},
+		},
+		{
+			name:    "caddyfile flag overrides the current config",
+			current: service(api.LabelCaddyManaged),
+			opts:    deployOptions{caddyfile: "global.Caddyfile", caddyfileChanged: true},
+			config:  "{\n\tstorage redis\n}",
+			want:    client.CaddyDeploymentOptions{Config: new("{\n\tstorage redis\n}"), Placement: machines},
+		},
+		{
+			name:    "caddyfile flag with an empty file sets a custom empty config",
+			current: service(api.LabelCaddyManaged),
+			opts:    deployOptions{caddyfile: "empty.Caddyfile", caddyfileChanged: true},
+			want:    client.CaddyDeploymentOptions{Config: new(""), Placement: machines},
+		},
+		{
+			name:    "empty caddyfile flag resets to the default config",
+			current: service(api.LabelCaddyManaged, api.LabelCaddyConfigCustom),
+			opts:    deployOptions{caddyfileChanged: true},
+			want:    client.CaddyDeploymentOptions{Placement: machines},
+		},
+		{
+			name:    "machine flag overrides the current machines",
+			current: service(api.LabelCaddyManaged),
+			opts:    deployOptions{machines: []string{"m3,m4", "m5"}, machinesChanged: true},
+			want:    client.CaddyDeploymentOptions{Placement: api.Placement{Machines: []string{"m3", "m4", "m5"}}},
+		},
+		{
+			name:    "empty machine flag resets to all machines",
+			current: service(api.LabelCaddyManaged),
+			opts:    deployOptions{machines: []string{}, machinesChanged: true},
+		},
+		{
+			name:    "flags apply to a service that isn't managed",
+			current: service(),
+			opts: deployOptions{
+				image: "caddy:2", imageChanged: true,
+				caddyfile: "global.Caddyfile", caddyfileChanged: true,
+				machines: []string{"m3"}, machinesChanged: true,
+			},
+			config: customConfig,
+			want: client.CaddyDeploymentOptions{
+				Image:     "caddy:2",
+				Config:    new(customConfig),
+				Placement: api.Placement{Machines: []string{"m3"}},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, resolveDeploymentOptions(tt.current, tt.opts, tt.config))
+		})
+	}
+}
+
+func TestDeploymentWarnings(t *testing.T) {
+	t.Parallel()
+
+	const (
+		defaultImage = client.CaddyImage + ":0.1.3"
+		customImage  = "example.com/caddy:1.2.3"
+		localConfig  = "{\n\tdebug\n}"
+		override     = "doesn't manage this caddy service yet"
+		storage      = "certificate storage changes from"
+		keepStorage  = "To keep the current storage"
+		module       = "must include the Uncloud storage module"
+	)
+	managed := map[string]string{api.LabelCaddyManaged: ""}
+
+	service := func(image, config string, labels map[string]string) api.ServiceSpec {
+		spec := api.ServiceSpec{
+			Name:      client.CaddyServiceName,
+			Container: api.ContainerSpec{Image: image},
+			Labels:    labels,
+		}
+		if config != "" {
+			spec.Caddy = &api.CaddySpec{Config: config}
+		}
+		return spec
+	}
+
+	tests := []struct {
+		name    string
+		current api.ServiceSpec
+		target  api.ServiceSpec
+		opts    deployOptions
+		want    []string
+		notWant []string
+	}{
+		{
+			name:    "new service with defaults has no warnings",
+			target:  service(defaultImage, client.DefaultCaddyConfig, managed),
+			notWant: []string{override, storage, module},
+		},
+		{
+			name:    "new service with a custom image",
+			target:  service(customImage, client.DefaultCaddyConfig, managed),
+			want:    []string{module},
+			notWant: []string{override, storage},
+		},
+		{
+			name:    "service that isn't managed moves to cluster storage",
+			current: service("caddy:2.11.4", "", nil),
+			target:  service(defaultImage, client.DefaultCaddyConfig, managed),
+			want:    []string{override, "from local to cluster", keepStorage},
+			notWant: []string{module},
+		},
+		{
+			name:    "service that isn't managed already uses cluster storage",
+			current: service(defaultImage, client.DefaultCaddyConfig, nil),
+			target:  service(defaultImage, client.DefaultCaddyConfig, managed),
+			want:    []string{override},
+			notWant: []string{storage, module},
+		},
+		{
+			name:    "managed service without changes has no warnings",
+			current: service(defaultImage, client.DefaultCaddyConfig, managed),
+			target:  service(defaultImage, client.DefaultCaddyConfig, managed),
+			notWant: []string{override, storage, module},
+		},
+		{
+			name:    "managed service upgrades the default image",
+			current: service(client.CaddyImage+":0.1.2", client.DefaultCaddyConfig, managed),
+			target:  service(defaultImage, client.DefaultCaddyConfig, managed),
+			notWant: []string{override, storage, module},
+		},
+		{
+			name:    "custom caddyfile moves to local storage",
+			current: service(defaultImage, client.DefaultCaddyConfig, managed),
+			target:  service(defaultImage, localConfig, managed),
+			opts:    deployOptions{caddyfile: "global.Caddyfile", caddyfileChanged: true},
+			want:    []string{"from cluster to local"},
+			notWant: []string{override, keepStorage, module},
+		},
+		{
+			name:    "other storage module is named",
+			current: service(defaultImage, "{\n\tstorage redis\n}", managed),
+			target:  service(defaultImage, client.DefaultCaddyConfig, managed),
+			want:    []string{"from 'redis' to cluster", keepStorage},
+		},
+		{
+			name:    "custom image is unchanged with the same config",
+			current: service(customImage, client.DefaultCaddyConfig, managed),
+			target:  service(customImage, client.DefaultCaddyConfig, managed),
+			notWant: []string{override, storage, module},
+		},
+		{
+			name:    "custom image receives a changed default config",
+			current: service(customImage, "{\n\tstorage uncloud\n\tdebug\n}", managed),
+			target:  service(customImage, client.DefaultCaddyConfig, managed),
+			want:    []string{module},
+			notWant: []string{override, storage},
+		},
+		{
+			name:    "custom image with local storage",
+			current: service(defaultImage, localConfig, managed),
+			target:  service(customImage, localConfig, managed),
+			notWant: []string{override, storage, module},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			warnings := strings.Join(deploymentWarnings(tt.current, tt.target, tt.opts), "\n")
+			for _, want := range tt.want {
+				assert.Contains(t, warnings, want)
+			}
+			for _, notWant := range tt.notWant {
+				assert.NotContains(t, warnings, notWant)
+			}
+		})
+	}
+}

cmd/uc/context/show.go

@@ -22,15 +22,6 @@ func NewShowCommand() *cobra.Command {
 }
 
 func show(uncli *cli.CLI) error {
-	// discard errors, only show the current context, otherwise nothing
+	fmt.Println(uncli.ContextOverrideOrCurrent())
-	if uncli.Config == nil {
-		return nil
-	}
-
-	if len(uncli.Config.Contexts) == 0 {
-		return nil
-	}
-	fmt.Println(uncli.Config.CurrentContext)
-
 	return nil
 }

cmd/uc/deploy.go

@@ -18,7 +18,7 @@ import (
 	"github.com/psviderski/uncloud/pkg/client"
 	"github.com/psviderski/uncloud/pkg/client/compose"
 	"github.com/psviderski/uncloud/pkg/client/deploy"
-	"github.com/psviderski/uncloud/pkg/client/deploy/operation"
+	"github.com/psviderski/uncloud/pkg/client/operation"
 	"github.com/spf13/cobra"
 )
 
@@ -305,7 +305,7 @@ const defaultFailedContainerLogsTail = 10
 // UNCLOUD_FAILED_CONTAINER_LOGS_TAIL environment variable override when set and valid.
 func failedContainerLogsTail() int {
 	if v := os.Getenv("UNCLOUD_FAILED_CONTAINER_LOGS_TAIL"); v != "" {
-		if tail, err := logs.Tail(v); err == nil && (tail == -1 || tail > 0) {
+		if tail, err := logs.ParseTail(v); err == nil && (tail == -1 || tail > 0) {
 			return tail
 		}
 	}

cmd/uc/image/ls.go

@@ -13,6 +13,7 @@ import (
 	"github.com/charmbracelet/colorprofile"
 	"github.com/containerd/platforms"
 	"github.com/docker/docker/api/types/image"
+	"github.com/docker/docker/pkg/stringid"
 	"github.com/docker/go-units"
 
 	"github.com/psviderski/uncloud/internal/cli"
@@ -90,10 +91,15 @@ func list(ctx context.Context, uncli *cli.CLI, opts listOptions) error {
 
 	machines := cli.ExpandCommaSeparatedValues(opts.machines)
 
-	clusterImages, err := clusterClient.ListImages(ctx, api.ImageFilter{
+	var clusterImages []api.MachineImages
+	err = tui.RunSpinner(ctx, "Listing images...", func(ctx context.Context) error {
+		var err error
+		clusterImages, err = clusterClient.ListImages(ctx, api.ImageFilter{
 			Machines: machines,
 			Name:     opts.nameFilter,
 		})
+		return err
+	})
 	if err != nil {
 		return fmt.Errorf("list images: %w", err)
 	}
@@ -102,7 +108,7 @@ func list(ctx context.Context, uncli *cli.CLI, opts listOptions) error {
 	var rows []imageRow
 
 	for _, machineImages := range clusterImages {
-		if err := machineImages.Error(); err != nil {
+		if err := machineImages.Metadata.Err(); err != nil {
 			tui.PrintWarning(fmt.Sprintf("failed to list images on machine '%s': %s",
 				machineImages.Metadata.MachineName, err))
 			continue
@@ -116,49 +122,7 @@ func list(ctx context.Context, uncli *cli.CLI, opts listOptions) error {
 			store = "containerd"
 		}
 
-		// Process each image for this machine.
+		rows = append(rows, imageRows(machineImages.Images, machineName, store)...)
-		for _, img := range machineImages.Images {
-			// Show the first 12 chars without 'sha256:' as the image ID like Docker does.
-			id := strings.TrimPrefix(img.ID, "sha256:")[:12]
-
-			name := "<none>"
-			if len(img.RepoTags) > 0 && img.RepoTags[0] != "<none>:<none>" {
-				name = img.RepoTags[0]
-			}
-
-			imgPlatforms, _ := imagePlatforms(img)
-			formattedPlatforms := formatPlatforms(imgPlatforms)
-
-			created := ""
-			createdAt := time.Unix(img.Created, 0)
-			if !createdAt.IsZero() {
-				created = units.HumanDuration(time.Now().UTC().Sub(createdAt)) + " ago"
-			}
-
-			size := units.HumanSizeWithPrecision(float64(img.Size), 3)
-
-			// Check if the image is in use by any containers. Only supported by Docker API >=1.51
-			inUse := "-"
-			if img.Containers != -1 { // -1 means the info is not available.
-				if img.Containers > 0 {
-					inUse = lipgloss.NewStyle().Foreground(lipgloss.Color("10")).Render("●")
-				} else {
-					inUse = lipgloss.NewStyle().Foreground(lipgloss.Color("8")).Render("○")
-				}
-			}
-
-			rows = append(rows, imageRow{
-				id:           id,
-				name:         name,
-				platforms:    formattedPlatforms,
-				createdHuman: created,
-				createdUnix:  img.Created,
-				size:         size,
-				inUse:        inUse,
-				store:        store,
-				machine:      machineName,
-			})
-		}
 	}
 
 	if len(rows) == 0 {
@@ -184,13 +148,55 @@ func list(ctx context.Context, uncli *cli.CLI, opts listOptions) error {
 	return nil
 }
 
+// imageRows keeps every stored name visible and retains digest-only and unnamed images.
+func imageRows(images []image.Summary, machineName, store string) []imageRow {
+	var rows []imageRow
+	for _, img := range images {
+		// RepoTags includes stored digest names on containerd. RepoDigests can include derived names.
+		imageNames := img.RepoTags
+		if len(imageNames) == 0 {
+			imageNames = []string{"<untagged>"}
+		}
+
+		imgPlatforms, _ := imagePlatforms(img)
+		created := ""
+		if img.Created > 0 {
+			created = units.HumanDuration(time.Since(time.Unix(img.Created, 0))) + " ago"
+		}
+		// Check if the image is in use by any containers. Only supported by Docker API >=1.51
+		inUse := "-"
+		if img.Containers >= 0 { // -1 means the info is not available.
+			inUse = lipgloss.NewStyle().Foreground(lipgloss.Color("8")).Render("○")
+			if img.Containers > 0 {
+				inUse = lipgloss.NewStyle().Foreground(lipgloss.Color("10")).Render("●")
+			}
+		}
+
+		for _, name := range imageNames {
+			rows = append(rows, imageRow{
+				id:           stringid.TruncateID(img.ID),
+				name:         name,
+				platforms:    formatPlatforms(imgPlatforms),
+				createdHuman: created,
+				createdUnix:  img.Created,
+				size:         units.HumanSizeWithPrecision(float64(img.Size), 3),
+				inUse:        inUse,
+				store:        store,
+				machine:      machineName,
+			})
+		}
+	}
+
+	return rows
+}
+
 // imagePlatforms returns a list of platforms supported by the image and a boolean indicating if it's multi-platform.
 func imagePlatforms(img image.Summary) ([]string, bool) {
 	var formattedPlatforms []string
 	multiPlatform := false
 
 	for _, m := range img.Manifests {
-		if m.Kind != image.ManifestKindImage || !m.Available {
+		if m.Kind != image.ManifestKindImage || !m.Available || m.ImageData == nil {
 			continue
 		}
 

cmd/uc/image/rm.go

@@ -0,0 +1,137 @@
+package image
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+
+	"charm.land/lipgloss/v2"
+	"github.com/docker/compose/v2/pkg/progress"
+	"github.com/psviderski/uncloud/internal/cli"
+	"github.com/psviderski/uncloud/internal/cli/completion"
+	"github.com/psviderski/uncloud/internal/cli/tui"
+	"github.com/spf13/cobra"
+)
+
+type removeOptions struct {
+	machines []string
+	yes      bool
+}
+
+func NewRemoveCommand() *cobra.Command {
+	opts := removeOptions{}
+	cmd := &cobra.Command{
+		Use:     "rm IMAGE [IMAGE...]",
+		Aliases: []string{"remove", "delete"},
+		Short:   "Remove images from machines in the cluster.",
+		Long: `Remove images from all or selected machines in the cluster.
+
+An image name without a tag selects all its tags. Specify a tag, such as myapp:latest, to remove
+only that tag.
+
+Wildcard patterns are supported for image names and tags and should be quoted to prevent your
+shell from expanding them. Removing by image ID attempts to remove each name for that image
+independently.
+
+Images used by running or stopped containers cannot be removed. Unused tags can be removed even
+when the same image is used under another tag. Failed removals do not prevent other removals.`,
+		Example: `  # Remove a tag from all machines after confirming the plan.
+  uc image rm myapp:old
+
+  # Remove all tags for an image name from all machines.
+  uc image rm myapp
+
+  # Remove matching tags from two machines without confirmation.
+  uc image rm 'myapp:1.*' 'otherapp:*' -m machine1,machine2 --yes
+
+  # Remove all names for an image by ID from all machines.
+  uc image rm ee8fed0e9618`,
+		Args: cobra.MinimumNArgs(1),
+		PreRun: func(cmd *cobra.Command, args []string) {
+			cli.BindEnvToFlag(cmd, "yes", "UNCLOUD_AUTO_CONFIRM")
+		},
+		RunE: func(cmd *cobra.Command, args []string) error {
+			uncli := cmd.Context().Value("cli").(*cli.CLI)
+			return remove(cmd.Context(), uncli, args, opts)
+		},
+	}
+
+	cmd.Flags().StringSliceVarP(&opts.machines, "machine", "m", nil,
+		"Machine names or IDs to remove images from. Can be specified multiple times or as a comma-separated list.\n"+
+			"If not specified, the matching images will be removed from all machines.")
+	cmd.Flags().BoolVarP(&opts.yes, "yes", "y", false,
+		"Auto-confirm the removal plan. Should be explicitly set when running non-interactively,\n"+
+			"e.g., in CI/CD pipelines. [$UNCLOUD_AUTO_CONFIRM]")
+
+	completion.MachinesFlag(cmd)
+
+	return cmd
+}
+
+func remove(ctx context.Context, uncli *cli.CLI, images []string, opts removeOptions) error {
+	clusterClient, err := uncli.ConnectCluster(ctx)
+	if err != nil {
+		return fmt.Errorf("connect to cluster: %w", err)
+	}
+	defer clusterClient.Close()
+
+	plan, err := planImageRemoval(ctx, clusterClient, images, cli.ExpandCommaSeparatedValues(opts.machines))
+	if err != nil {
+		return fmt.Errorf("plan image removal: %w", err)
+	}
+	directConn := uncli.DirectConnection()
+	contextName := uncli.ContextOverrideOrCurrent()
+	cluster := ""
+	if directConn != "" {
+		cluster = directConn
+		lipgloss.Println(tui.Faint.Render("connection: ") + tui.NameStyle.Render(directConn))
+		lipgloss.Println()
+	} else if contextName != "" && len(uncli.Config.Contexts) > 1 {
+		// Only show context if there's more than one to avoid unnecessary clutter.
+		cluster = contextName
+		lipgloss.Println(tui.Faint.Render("context: ") + tui.NameStyle.Render(contextName))
+		lipgloss.Println()
+	}
+
+	lipgloss.Println(plan.Format())
+
+	if !opts.yes {
+		if !tui.IsTerminalAvailable() {
+			return errors.New("cannot ask to confirm image removal in non-interactive mode, " +
+				"use --yes flag or set UNCLOUD_AUTO_CONFIRM=true to auto-confirm")
+		}
+
+		title := "Proceed with image removal?"
+		if cluster != "" {
+			isDark := lipgloss.HasDarkBackground(os.Stdin, os.Stdout)
+			confirmStyle := tui.ThemeConfirm().Theme(isDark).Focused.Title
+			title = "Proceed with image removal on " + tui.NameStyle.Render(cluster) + confirmStyle.Render("?")
+		}
+
+		confirmed, err := tui.Confirm(title)
+		if err != nil {
+			return fmt.Errorf("confirm image removal: %w", err)
+		}
+		if !confirmed {
+			return cli.Cancelled("Image removal cancelled. No images were removed.")
+		}
+	}
+
+	untagged := make(map[string]bool)
+	var errs []error
+	for i, group := range plan.GroupPlans {
+		if i > 0 {
+			fmt.Println()
+		}
+
+		err := progress.RunWithTitle(ctx, func(ctx context.Context) error {
+			return group.Execute(ctx, clusterClient, untagged)
+		}, uncli.ProgressOut(), "Removing image "+group.Image)
+		if err != nil {
+			errs = append(errs, err)
+		}
+	}
+
+	return errors.Join(errs...)
+}

cmd/uc/image/rm_plan.go

@@ -0,0 +1,426 @@
+package image
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"path"
+	"regexp"
+	"slices"
+	"strings"
+
+	"charm.land/lipgloss/v2"
+	"github.com/distribution/reference"
+	"github.com/docker/docker/api/types/image"
+	"github.com/docker/docker/pkg/stringid"
+	"github.com/psviderski/uncloud/internal/cli/tui"
+	"github.com/psviderski/uncloud/pkg/api"
+)
+
+var imageIDRegex = regexp.MustCompile(`^(sha256:)?[a-f0-9]{4,64}$`)
+
+// imageRemovalPlan contains all image removal operations grouped by the user-supplied image name, ID, or pattern.
+type imageRemovalPlan struct {
+	GroupPlans []*removeImageGroupPlan
+}
+
+func (p *imageRemovalPlan) IsEmpty() bool { return len(p.GroupPlans) == 0 }
+
+func (p *imageRemovalPlan) Format() string {
+	var out strings.Builder
+	machines := make(map[string]struct{})
+	removeCount := 0
+	for _, group := range p.GroupPlans {
+		out.WriteString(group.Format())
+		for _, op := range group.Operations {
+			machines[op.MachineName] = struct{}{}
+			removeCount++
+		}
+	}
+
+	summary := fmt.Sprintf("%s remove %s across %s %s",
+		tui.BoldRed.Render(fmt.Sprint(removeCount)),
+		tui.Faint.Render("·"),
+		tui.Bold.Render(fmt.Sprint(len(machines))),
+		pluralMachines(len(machines)),
+	)
+	out.WriteString("\n" + tui.Faint.Render(strings.Repeat("─", lipgloss.Width(summary))) + "\n" + summary + "\n")
+
+	return out.String()
+}
+
+func pluralMachines(count int) string {
+	if count == 1 {
+		return "machine"
+	}
+	return "machines"
+}
+
+// Execute applies the frozen plan, collecting failures without discovering new targets.
+func (p *imageRemovalPlan) Execute(ctx context.Context, cli api.ImageClient) error {
+	untagged := make(map[string]bool)
+	var errs []error
+	for _, group := range p.GroupPlans {
+		if err := group.Execute(ctx, cli, untagged); err != nil {
+			errs = append(errs, err)
+		}
+	}
+	return errors.Join(errs...)
+}
+
+// removeImageGroupPlan groups removals for one image name, ID, or pattern supplied by the user.
+type removeImageGroupPlan struct {
+	// Image is the image name, ID, or pattern supplied by the user.
+	Image      string
+	Operations []*removeImageOperation
+}
+
+func (o *removeImageGroupPlan) Execute(ctx context.Context, cli api.ImageClient, untagged map[string]bool) error {
+	var errs []error
+	for _, op := range o.Operations {
+		if err := ctx.Err(); err != nil {
+			return errors.Join(append(errs, err)...)
+		}
+		if err := op.Execute(ctx, cli, untagged); err != nil {
+			errs = append(errs, err)
+		}
+	}
+	return errors.Join(errs...)
+}
+
+func (o *removeImageGroupPlan) Format() string {
+	var out strings.Builder
+	out.WriteString(tui.BoldRed.Render("-") + " remove image " + tui.NameStyle.Render(o.Image) + "\n")
+	for i, op := range o.Operations {
+		branch := "  ├── "
+		if i == len(o.Operations)-1 {
+			branch = "  ╰── "
+		}
+		out.WriteString(tui.Faint.Render(branch) + op.Format() + "\n")
+	}
+	return out.String()
+}
+
+func (o *removeImageGroupPlan) String() string {
+	return fmt.Sprintf("removeImageGroupPlan[image=%s]", o.Image)
+}
+
+// removeImageOperation removes an image name or an unnamed image by ID on a machine.
+type removeImageOperation struct {
+	MachineName string
+	ImageID     string
+	// Image is the exact image name or ID sent to Docker.
+	Image string
+}
+
+func (o *removeImageOperation) Execute(ctx context.Context, cli api.ImageClient, untagged map[string]bool) error {
+	if untagged[o.MachineName+"\x00"+o.Image] {
+		return nil
+	}
+	results, err := cli.RemoveImage(ctx, o.Image, api.RemoveImageOptions{Machines: []string{o.MachineName}})
+	// A removal may also untag other planned names. Track actual results across all groups, per machine.
+	for _, result := range results {
+		for _, removal := range result.Results {
+			if removal.Untagged != "" {
+				untagged[result.Metadata.MachineName+"\x00"+removal.Untagged] = true
+			}
+		}
+	}
+	if errors.Is(err, api.ErrNotFound) {
+		return nil
+	}
+	if err != nil {
+		return fmt.Errorf("remove image '%s': %w", o.Image, err)
+	}
+	return nil
+}
+
+func (o *removeImageOperation) Format() string {
+	return tui.BoldRed.Render("-") + " " + tui.FormatImage(o.Image, tui.NoStyle) +
+		tui.Faint.Render(" ("+stringid.TruncateID(o.ImageID)+") on ") + o.MachineName
+}
+
+func (o *removeImageOperation) String() string {
+	return fmt.Sprintf("removeImageOperation[machine_name=%s image=%s]", o.MachineName, o.Image)
+}
+
+type machineImageInventory struct {
+	machineName string
+	images      []image.Summary
+}
+
+// planImageRemoval builds a removal plan grouped by input argument from images on the selected machines.
+// An empty machines list selects all machines.
+//
+// Arguments expand as follows:
+//   - Name: without a tag or digest, select all stored names in that exact repository across image IDs.
+//     An explicit tag selects only that tag. Familiar names are normalised.
+//   - ID: select all RepoTags for the matching image, or its full ID if unnamed. A prefix must identify
+//     one unique full ID across the selected machines.
+//   - Pattern: match RepoTags only. A matching digest also selects that image's names in the same repository.
+//   - Digest: resolve through RepoTags, RepoDigests, or the image ID, then select that image's names in
+//     the same repository. If none exist but RepoDigests contains the reference, select the digest itself.
+//
+// Selecting all tags for a repository/image on a machine also adds its stored digest references after
+// the tags, even if those digests did not match an argument. Stored names are kept as is, and overlapping
+// selections produce only one operation per machine and name.
+//
+// Failed image listings produce warnings and exclude those machines from the plan. If no machine can
+// be queried, planning fails. Every argument must match on a successfully queried machine. Unmatched
+// or invalid arguments and ambiguous ID prefixes reject the whole plan before any removal.
+//
+// Planning does not check container usage. The backend checks each removal during execution, and failures
+// do not stop other operations. Tag removal may implicitly untag a used digest. Execution skips names
+// already untagged by earlier operations and tolerates names that disappeared after planning.
+func planImageRemoval(ctx context.Context, cli api.ImageClient, images, machines []string) (*imageRemovalPlan, error) {
+	if len(images) == 0 {
+		return nil, errors.New("at least one image is required")
+	}
+
+	var clusterImages []api.MachineImages
+	err := tui.RunSpinner(ctx, "Finding images to remove...", func(ctx context.Context) error {
+		var err error
+		clusterImages, err = cli.ListImages(ctx, api.ImageFilter{Machines: machines})
+		return err
+	})
+	if err != nil {
+		return nil, fmt.Errorf("list images: %w", err)
+	}
+
+	var inventory []*machineImageInventory
+	for _, machineImages := range clusterImages {
+		if err := machineImages.Metadata.Err(); err != nil {
+			tui.PrintWarning(fmt.Sprintf("failed to list images on machine '%s': %s",
+				machineImages.Metadata.MachineName, err))
+			continue
+		}
+		machine := &machineImageInventory{
+			machineName: machineImages.Metadata.MachineName,
+			images:      machineImages.Images,
+		}
+		inventory = append(inventory, machine)
+	}
+	if len(inventory) == 0 {
+		return nil, errors.New("failed to list images on any machine")
+	}
+
+	return planImageRemovalFromInventory(images, inventory)
+}
+
+func planImageRemovalFromInventory(images []string, inventory []*machineImageInventory) (*imageRemovalPlan, error) {
+	plan := &imageRemovalPlan{}
+	seen := make(map[string]bool)
+	var errs []error
+	for _, imageArg := range images {
+		group := &removeImageGroupPlan{Image: imageArg}
+		isID := imageIDRegex.MatchString(imageArg)
+		isPattern := strings.ContainsAny(imageArg, "*?[")
+		var namedRef reference.Named
+		if isPattern {
+			if _, err := path.Match(imageArg, ""); err != nil {
+				return nil, fmt.Errorf("invalid image pattern '%s': %w", imageArg, err)
+			}
+		} else if !isID {
+			var err error
+			namedRef, err = reference.ParseNormalizedNamed(imageArg)
+			if err != nil {
+				return nil, fmt.Errorf("invalid image '%s': %w", imageArg, err)
+			}
+		}
+		if isID {
+			matches := make(map[string]bool)
+			for _, machine := range inventory {
+				for _, img := range machine.images {
+					if imageIDMatches(img.ID, imageArg) {
+						matches[img.ID] = true
+					}
+				}
+			}
+			if len(matches) > 1 {
+				var candidates []string
+				for id := range matches {
+					candidates = append(candidates, id)
+				}
+				slices.Sort(candidates)
+				errs = append(errs,
+					fmt.Errorf("image ID '%s' is ambiguous: %s", imageArg, strings.Join(candidates, ", ")))
+				continue
+			}
+		}
+
+		matched := false
+		for _, machine := range inventory {
+			for _, img := range machine.images {
+				var imagesToRemove []string
+				switch {
+				case isID:
+					if !imageIDMatches(img.ID, imageArg) {
+						continue
+					}
+					matched = true
+					// Expand only stored names, matching uc image ls. RepoDigests remains available for digest lookup.
+					imagesToRemove = img.RepoTags
+					if len(imagesToRemove) == 0 {
+						imagesToRemove = []string{img.ID}
+					}
+				case isPattern:
+					for _, ref := range img.RepoTags {
+						if ok, _ := path.Match(imageArg, ref); ok {
+							named, _ := reference.ParseDockerRef(ref)
+							if _, digest := named.(reference.Canonical); digest {
+								imagesToRemove = append(imagesToRemove,
+									imageRepositoryReferences(img.RepoTags, named.Name())...)
+							} else {
+								imagesToRemove = append(imagesToRemove, ref)
+							}
+						}
+					}
+				case reference.IsNameOnly(namedRef):
+					imagesToRemove = imageRepositoryReferences(img.RepoTags, namedRef.Name())
+				default:
+					if canonical, digest := namedRef.(reference.Canonical); digest {
+						ref := reference.FamiliarString(namedRef)
+						if canonical.Digest().String() != img.ID && !slices.Contains(img.RepoTags, ref) &&
+							!slices.Contains(img.RepoDigests, ref) {
+							continue
+						}
+						imagesToRemove = imageRepositoryReferences(img.RepoTags, namedRef.Name())
+						if len(imagesToRemove) == 0 && slices.Contains(img.RepoDigests, ref) {
+							// An explicitly requested digest can resolve even when no stored name is listed.
+							imagesToRemove = []string{ref}
+						}
+					} else if slices.Contains(img.RepoTags, reference.FamiliarString(namedRef)) {
+						imagesToRemove = []string{reference.FamiliarString(namedRef)}
+					}
+				}
+				for _, ref := range imagesToRemove {
+					matched = true
+					key := machine.machineName + "\x00" + ref
+					if !seen[key] {
+						seen[key] = true
+						group.Operations = append(group.Operations, &removeImageOperation{
+							MachineName: machine.machineName, ImageID: img.ID, Image: ref,
+						})
+					}
+				}
+			}
+		}
+		if !matched {
+			errs = append(errs, fmt.Errorf("image '%s' not found on any machine", imageArg))
+		}
+		slices.SortFunc(group.Operations, compareImageRemovalOperations)
+		if len(group.Operations) > 0 {
+			plan.GroupPlans = append(plan.GroupPlans, group)
+		}
+	}
+
+	planDigestRemovals(plan, inventory)
+	if err := errors.Join(errs...); err != nil {
+		return nil, err
+	}
+	return plan, nil
+}
+
+func imageRepositoryReferences(refs []string, repository string) []string {
+	var matches []string
+	for _, ref := range refs {
+		if named, err := reference.ParseDockerRef(ref); err == nil && named.Name() == repository {
+			matches = append(matches, ref)
+		}
+	}
+	return matches
+}
+
+// planDigestRemovals includes stored digests with the group removing their last repository tag.
+// RepoDigests also contains synthesised references, so only digest names stored in RepoTags get operations here.
+func planDigestRemovals(plan *imageRemovalPlan, inventory []*machineImageInventory) {
+	type repository struct {
+		remainingTags int
+		digests       []string
+	}
+	repositories := make(map[string]*repository)
+	for _, machine := range inventory {
+		for _, img := range machine.images {
+			names := slices.Clone(img.RepoTags)
+			slices.Sort(names)
+			for _, ref := range slices.Compact(names) {
+				named, err := reference.ParseDockerRef(ref)
+				if err != nil {
+					continue
+				}
+				key := machine.machineName + "\x00" + img.ID + "\x00" + named.Name()
+				repo := repositories[key]
+				if repo == nil {
+					repo = &repository{}
+					repositories[key] = repo
+				}
+				if _, tagged := named.(reference.Tagged); tagged {
+					repo.remainingTags++
+				} else if _, digest := named.(reference.Canonical); digest {
+					repo.digests = append(repo.digests, ref)
+				}
+			}
+		}
+	}
+	emitted := make(map[string]bool)
+	var groups []*removeImageGroupPlan
+	for _, group := range plan.GroupPlans {
+		var operations []*removeImageOperation
+		for _, op := range group.Operations {
+			key := op.MachineName + "\x00" + op.Image
+			if emitted[key] {
+				continue
+			}
+			emitted[key] = true
+			operations = append(operations, op)
+			named, err := reference.ParseDockerRef(op.Image)
+			if err != nil {
+				continue
+			}
+			if _, tagged := named.(reference.Tagged); !tagged {
+				continue
+			}
+			repo := repositories[op.MachineName+"\x00"+op.ImageID+"\x00"+named.Name()]
+			if repo == nil {
+				continue
+			}
+			repo.remainingTags--
+			if repo.remainingTags != 0 {
+				continue
+			}
+			for _, digest := range repo.digests {
+				key := op.MachineName + "\x00" + digest
+				if !emitted[key] {
+					emitted[key] = true
+					operations = append(operations, &removeImageOperation{
+						MachineName: op.MachineName, ImageID: op.ImageID, Image: digest,
+					})
+				}
+			}
+		}
+		group.Operations = operations
+		slices.SortFunc(group.Operations, compareImageRemovalOperations)
+		if len(group.Operations) > 0 {
+			groups = append(groups, group)
+		}
+	}
+	plan.GroupPlans = groups
+}
+
+func compareImageRemovalOperations(a, b *removeImageOperation) int {
+	aDigest, bDigest := strings.Contains(a.Image, "@"), strings.Contains(b.Image, "@")
+	if aDigest != bDigest {
+		if aDigest {
+			return 1
+		}
+		return -1
+	}
+	if c := strings.Compare(a.Image, b.Image); c != 0 {
+		return c
+	}
+	return strings.Compare(a.MachineName, b.MachineName)
+}
+
+func imageIDMatches(id, prefix string) bool {
+	return strings.HasPrefix(strings.TrimPrefix(id, "sha256:"), strings.TrimPrefix(prefix, "sha256:"))
+}

cmd/uc/image/rm_plan_test.go

@@ -0,0 +1,405 @@
+package image
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/docker/docker/api/types/image"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestPlanImageRemovalFromInventory_Selectors(t *testing.T) {
+	idA, idB := testImageID("a"), testImageID("b")
+	digestRef := "myapp@" + testImageID("c")
+	tests := []struct {
+		name      string
+		selector  string
+		inventory []*machineImageInventory
+		want      []removeImageOperation
+	}{
+		{
+			name: "repository selects all stored names across images and machines", selector: "myapp",
+			inventory: []*machineImageInventory{
+				testMachineInventory("dev1", []image.Summary{
+					testImage(idA, "myapp:old", "other:keep"),
+					testImage(idB, "myapp:latest"),
+					testImage(testImageID("c"), digestRef),
+					testImage(testImageID("d"), "myapp-extra:latest", "team/myapp:latest", "myapp/worker:latest"),
+				}),
+				testMachineInventory("machine2", []image.Summary{testImage(idB, "myapp:test")}),
+			},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idB, Image: "myapp:latest"},
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "machine2", ImageID: idB, Image: "myapp:test"},
+				{MachineName: "dev1", ImageID: testImageID("c"), Image: digestRef},
+			},
+		},
+		{
+			name: "repository without latest includes its tags and stored digest", selector: "myapp",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", digestRef, "other:keep"),
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "dev1", ImageID: idA, Image: digestRef},
+			},
+		},
+		{
+			name: "explicit latest selects only that tag", selector: "myapp:latest",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", "myapp:latest", digestRef),
+			})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:latest"}},
+		},
+		{
+			name: "qualified repository selects familiar names", selector: "docker.io/library/myapp",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", "myapp:latest", "other:keep"),
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:latest"},
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+			},
+		},
+		{
+			name: "registry port does not count as an explicit tag", selector: "registry.example.com:5000/team/myapp",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "registry.example.com:5000/team/myapp:old", "registry.example.com:5000/team/myapp:new",
+					"registry.example.com:5001/team/myapp:keep", "registry.example.com:5000/team/myapp/worker:keep"),
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "registry.example.com:5000/team/myapp:new"},
+				{MachineName: "dev1", ImageID: idA, Image: "registry.example.com:5000/team/myapp:old"},
+			},
+		},
+		{
+			name: "explicit tag with registry port stays exact", selector: "registry.example.com:5000/team/myapp:old",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "registry.example.com:5000/team/myapp:old", "registry.example.com:5000/team/myapp:new"),
+			})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "registry.example.com:5000/team/myapp:old"}},
+		},
+		{
+			name: "tag uses each machines local image and allows missing copies", selector: "myapp:old",
+			inventory: []*machineImageInventory{
+				testMachineInventory("machine2", []image.Summary{testImage(idB, "myapp:old")}),
+				testMachineInventory("empty", nil),
+				testMachineInventory("dev1", []image.Summary{testImage(idA, "myapp:old")}),
+			},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "machine2", ImageID: idB, Image: "myapp:old"},
+			},
+		},
+		{
+			name: "qualified names use familiar references", selector: "docker.io/library/myapp:old",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old"),
+			})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:old"}},
+		},
+		{
+			name: "full ID expands all repositories and deduplicates references per machine", selector: idA,
+			inventory: []*machineImageInventory{
+				testMachineInventory("dev1", []image.Summary{
+					{
+						ID:       idA,
+						RepoTags: []string{"other:keep", "myapp:old", "myapp:old", digestRef, "other@" + testImageID("c")},
+						RepoDigests: []string{
+							digestRef, "other@" + testImageID("c"), "other@" + testImageID("c"), testImageID("c"),
+						},
+					},
+					testImage(idB, "unrelated:keep"),
+				}),
+				testMachineInventory("machine2", []image.Summary{{
+					ID: idA, RepoTags: []string{"myapp:old", digestRef}, RepoDigests: []string{digestRef},
+				}}),
+			},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "machine2", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "dev1", ImageID: idA, Image: "other:keep"},
+				{MachineName: "dev1", ImageID: idA, Image: digestRef},
+				{MachineName: "machine2", ImageID: idA, Image: digestRef},
+				{MachineName: "dev1", ImageID: idA, Image: "other@" + testImageID("c")},
+			},
+		},
+		{
+			name: "same full ID on multiple machines is unambiguous", selector: "aaaaaaaa",
+			inventory: []*machineImageInventory{
+				testMachineInventory("dev1", []image.Summary{testImage(idA, "myapp:old")}),
+				testMachineInventory("dev1", []image.Summary{testImage(idB, "myapp:latest")}),
+				testMachineInventory("machine2", []image.Summary{testImage(idA, "other:test")}),
+			},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "machine2", ImageID: idA, Image: "other:test"},
+			},
+		},
+		{
+			name: "short ID with algorithm prefix", selector: "sha256:aaaaaaaa",
+			inventory: []*machineImageInventory{testMachineInventory("dev1",
+				[]image.Summary{testImage(idA, "myapp:old")})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:old"}},
+		},
+		{
+			name: "unnamed image is removed by its full ID", selector: idA,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{ID: idA}})},
+			want:      []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: idA}},
+		},
+		{
+			name: "ID expands only stored names", selector: idA,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoTags: []string{"myapp:latest"},
+				RepoDigests: []string{"myapp@" + idA, "other@" + idA},
+			}})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:latest"}},
+		},
+		{
+			name: "ID removes an untagged image by ID even if RepoDigests is populated", selector: idA,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoDigests: []string{digestRef},
+			}})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: idA}},
+		},
+		{
+			name: "synthesized digest selects repository tags without a digest row", selector: digestRef,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoTags: []string{"myapp:old", "other:keep", "myapp:latest"},
+				RepoDigests: []string{digestRef, "other@" + testImageID("c")},
+			}})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:latest"},
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+			},
+		},
+		{
+			name: "explicit digest lookup works without stored names", selector: digestRef,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoDigests: []string{digestRef},
+			}})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: digestRef}},
+		},
+		{
+			name: "synthesized digest resolves by target and repository even if not listed", selector: "myapp@" + idA,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoTags: []string{"myapp:old", "other:keep"}, RepoDigests: []string{"other@" + idA},
+			}})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:old"}},
+		},
+		{
+			name: "stored digest follows its repository tags", selector: digestRef,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", "myapp:latest", digestRef, "other:keep"),
+				testImage(idB, "myapp@"+idB),
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:latest"},
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "dev1", ImageID: idA, Image: digestRef},
+			},
+		},
+		{
+			name: "digest pattern also expands matching repository tags", selector: "myapp@sha256:c*",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", "myapp:latest", digestRef, "other:keep"),
+				testImage(idB, "myapp@"+idB),
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:latest"},
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "dev1", ImageID: idA, Image: digestRef},
+			},
+		},
+		{
+			name: "digest pattern matches stored names and ignores derived names", selector: "myapp@sha256:*",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", digestRef),
+				{ID: idB, RepoTags: []string{"myapp:latest"}, RepoDigests: []string{"myapp@" + idB}},
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "dev1", ImageID: idA, Image: digestRef},
+			},
+		},
+		{
+			name: "last repository tag includes its stored digest", selector: "myapp:old",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", digestRef, "other:keep"),
+			})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:old"},
+				{MachineName: "dev1", ImageID: idA, Image: digestRef},
+			},
+		},
+		{
+			name: "last tag does not add a synthesized digest row", selector: "myapp:old",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoTags: []string{"myapp:old", "other:keep"}, RepoDigests: []string{digestRef},
+			}})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:old"}},
+		},
+		{
+			name: "remaining repository tag preserves its stored digest", selector: "myapp:old",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, "myapp:old", "myapp:latest", digestRef),
+			})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: "myapp:old"}},
+		},
+		{
+			name: "digest without repository tags remains executable", selector: digestRef,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+				testImage(idA, digestRef, "other:keep"),
+			})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: digestRef}},
+		},
+		{
+			name: "digest-only image keeps its named reference", selector: idA,
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoTags: []string{digestRef}, RepoDigests: []string{digestRef, testImageID("c")},
+			}})},
+			want: []removeImageOperation{{MachineName: "dev1", ImageID: idA, Image: digestRef}},
+		},
+		{
+			name: "tag pattern selects only matching tags", selector: "myapp:1.[12]",
+			inventory: []*machineImageInventory{testMachineInventory("dev1", []image.Summary{{
+				ID: idA, RepoTags: []string{"myapp:1.3", "myapp:1.2", "myapp:1.1"}, RepoDigests: []string{digestRef},
+			}})},
+			want: []removeImageOperation{
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:1.1"},
+				{MachineName: "dev1", ImageID: idA, Image: "myapp:1.2"},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			plan, err := planImageRemovalFromInventory([]string{tt.selector}, tt.inventory)
+			require.NoError(t, err)
+
+			require.Len(t, plan.GroupPlans, 1)
+			assert.Equal(t, tt.selector, plan.GroupPlans[0].Image)
+
+			var got []removeImageOperation
+			for _, op := range plan.GroupPlans[0].Operations {
+				got = append(got, *op)
+			}
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestPlanImageRemovalFromInventory_OverlappingSelectors(t *testing.T) {
+	inventory := []*machineImageInventory{
+		testMachineInventory("dev1", []image.Summary{testImage(testImageID("a"), "myapp:old", "myapp:latest")}),
+		testMachineInventory("machine2", []image.Summary{testImage(testImageID("b"), "myapp:test", "myapp:old")}),
+	}
+
+	plan, err := planImageRemovalFromInventory([]string{"myapp:old", "myapp", "myapp:*", testImageID("a"), "myapp:old"},
+		inventory)
+	require.NoError(t, err)
+
+	assert.Equal(t, &imageRemovalPlan{GroupPlans: []*removeImageGroupPlan{
+		{Image: "myapp:old", Operations: []*removeImageOperation{
+			{MachineName: "dev1", ImageID: testImageID("a"), Image: "myapp:old"},
+			{MachineName: "machine2", ImageID: testImageID("b"), Image: "myapp:old"},
+		}},
+		{Image: "myapp", Operations: []*removeImageOperation{
+			{MachineName: "dev1", ImageID: testImageID("a"), Image: "myapp:latest"},
+			{MachineName: "machine2", ImageID: testImageID("b"), Image: "myapp:test"},
+		}},
+	}}, plan, "the first selector owns each removal and empty groups are omitted")
+}
+
+func TestPlanImageRemovalFromInventory_DigestFollowsLastTagAcrossSelectors(t *testing.T) {
+	imageID := testImageID("a")
+	digestRef := "myapp@" + imageID
+	inventory := []*machineImageInventory{testMachineInventory("dev1", []image.Summary{
+		testImage(imageID, "myapp:old", "myapp:latest", "myapp:old", digestRef, digestRef, "other:keep"),
+	})}
+
+	plan, err := planImageRemovalFromInventory([]string{"myapp:old", "myapp:latest", digestRef}, inventory)
+	require.NoError(t, err)
+
+	assert.Equal(t, &imageRemovalPlan{GroupPlans: []*removeImageGroupPlan{
+		{Image: "myapp:old", Operations: []*removeImageOperation{
+			{MachineName: "dev1", ImageID: imageID, Image: "myapp:old"},
+		}},
+		{Image: "myapp:latest", Operations: []*removeImageOperation{
+			{MachineName: "dev1", ImageID: imageID, Image: "myapp:latest"},
+			{MachineName: "dev1", ImageID: imageID, Image: digestRef},
+		}},
+	}}, plan, "digest cleanup belongs to the last tag removal, so the later digest selector has no extra operation")
+}
+
+func TestPlanImageRemovalFromInventory_RejectsInvalidOrMissingSelectors(t *testing.T) {
+	idA := "sha256:abcdef" + strings.Repeat("a", 58)
+	idB := "sha256:abcdef" + strings.Repeat("b", 58)
+	inventory := []*machineImageInventory{
+		testMachineInventory("dev1", []image.Summary{{
+			ID: idA, RepoTags: []string{"myapp:old"}, RepoDigests: []string{"myapp@" + idA, "derived@" + idA},
+		}}),
+		testMachineInventory("machine2", []image.Summary{{ID: idB, RepoTags: []string{"myapp:old"}}}),
+	}
+
+	tests := []struct {
+		name   string
+		images []string
+		err    string
+	}{
+		{"missing repository", []string{"missing"}, "not found"},
+		{"repository only has a derived digest", []string{"derived"}, "not found"},
+		{"missing tag", []string{"myapp:missing"}, "not found"},
+		{"missing pattern", []string{"other:*"}, "not found"},
+		{"pattern only matches a derived digest", []string{"myapp@sha256:*"}, "not found"},
+		{"missing ID", []string{testImageID("c")}, "not found"},
+		{"ambiguous ID prefix", []string{"abcdef"}, "ambiguous"},
+		{"invalid name", []string{"MyApp:old"}, "invalid image"},
+		{"invalid pattern", []string{"myapp:["}, "invalid image pattern"},
+		{"one missing selector rejects the whole plan", []string{"myapp:old", "myapp:missing"}, "not found"},
+		{"one invalid selector rejects the whole plan", []string{"myapp:old", "myapp:["}, "invalid image pattern"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			plan, err := planImageRemovalFromInventory(tt.images, inventory)
+			require.ErrorContains(t, err, tt.err)
+			assert.Nil(t, plan, "a failed selector must not return a partial plan")
+		})
+	}
+
+	t.Run("ID prefix resolves within the supplied machines", func(t *testing.T) {
+		plan, err := planImageRemovalFromInventory([]string{"abcdef"}, inventory[:1])
+		require.NoError(t, err)
+		require.Len(t, plan.GroupPlans, 1)
+		require.Len(t, plan.GroupPlans[0].Operations, 1)
+		assert.Equal(t, idA, plan.GroupPlans[0].Operations[0].ImageID)
+	})
+}
+
+func TestPlanImageRemovalFromInventory_InUseImagesAreStillPlanned(t *testing.T) {
+	img := testImage(testImageID("a"), "myapp:old", "myapp:current")
+	img.Containers = 1
+
+	for _, selector := range []string{"myapp", "myapp:current", img.ID, "myapp:*"} {
+		t.Run(selector, func(t *testing.T) {
+			plan, err := planImageRemovalFromInventory([]string{selector}, []*machineImageInventory{
+				testMachineInventory("dev1", []image.Summary{img}),
+			})
+			require.NoError(t, err)
+			assert.False(t, plan.IsEmpty(), "usage protection belongs to the backend at execution time")
+		})
+	}
+}
+
+func testImageID(char string) string { return "sha256:" + strings.Repeat(char, 64) }
+
+func testImage(imageID string, tags ...string) image.Summary {
+	return image.Summary{ID: imageID, RepoTags: tags}
+}
+
+func testMachineInventory(name string, images []image.Summary) *machineImageInventory {
+	return &machineImageInventory{machineName: name, images: images}
+}

cmd/uc/image/root.go

@@ -13,6 +13,7 @@ func NewRootCommand() *cobra.Command {
 	cmd.AddCommand(
 		NewListCommand(),
 		NewPushCommand(),
+		NewRemoveCommand(),
 	)
 
 	return cmd

cmd/uc/machine/add.go

@@ -187,14 +187,12 @@ func add(ctx context.Context, uncli *cli.CLI, remoteMachine *cli.RemoteMachine,
 	}
 	fmt.Println("Machine joined the cluster.")
 
-	// TODO: scale the existing Caddy service to the new machine instead of running a new deployment
+	// Deploy the current caddy service spec to run a container on the added machine. This keeps the service
-	//  that may cause a small downtime.
+	// configuration however it was deployed and shouldn't change the containers on other machines if specs on them
-	// Deploy a Caddy service container to the added machine. If caddy service is already deployed on other machines,
+	// haven't diverged.
-	// use the deployed image version.
 	// NOTE: We use the cluster client to inspect and scale the Caddy service because the newly added machine may have
 	// issues accessing the API of existing machines in the cluster.
 	// See the issue for more details: https://github.com/psviderski/uncloud/issues/65.
-	caddyImage := ""
 	caddySvc, err := clusterClient.InspectService(ctx, client.CaddyServiceName)
 	if err != nil {
 		if errors.Is(err, api.ErrNotFound) {
@@ -203,26 +201,10 @@ func add(ctx context.Context, uncli *cli.CLI, remoteMachine *cli.RemoteMachine,
 		}
 		return fmt.Errorf("inspect caddy service: %w", err)
 	}
-	caddyImage = caddySvc.Containers[0].Container.Config.Image
-	// Find the latest created container and use its image.
-	var latestCreated time.Time
-	for _, c := range caddySvc.Containers[1:] {
-		created, err := time.Parse(time.RFC3339Nano, c.Container.Created)
-		if err != nil {
-			continue
-		}
-		if created.After(latestCreated) {
-			latestCreated = created
-			caddyImage = c.Container.Config.Image
-		}
-	}
 
 	fmt.Println()
 	fmt.Println("Preparing Caddy deployment...")
-	d, err := clusterClient.Caddy.NewDeployment(ctx, client.CaddyDeploymentOptions{Image: caddyImage})
+	d := clusterClient.NewDeployment(caddySvc.Spec(), nil)
-	if err != nil {
-		return fmt.Errorf("create caddy deployment: %w", err)
-	}
 
 	plan, err := d.Plan(ctx)
 	if err != nil {

cmd/uc/machine/logs.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"slices"
 	"strings"
+	"time"
 
 	"github.com/psviderski/uncloud/internal/cli"
 	"github.com/psviderski/uncloud/internal/cli/completion"
@@ -64,6 +65,15 @@ If no services are specified, streams logs from the uncloud service.`,
 }
 
 func runLogs(ctx context.Context, uncli *cli.CLI, services []string, opts logs.Options) error {
+	since, until, err := logs.TimeRange(opts.Since, opts.Until, time.Now())
+	if err != nil {
+		return err
+	}
+	tail, err := opts.TailLines()
+	if err != nil {
+		return err
+	}
+
 	if len(services) == 0 {
 		services = []string{api.SystemServiceUncloud}
 	}
@@ -74,11 +84,6 @@ func runLogs(ctx context.Context, uncli *cli.CLI, services []string, opts logs.O
 		}
 	}
 
-	tail, err := logs.Tail(opts.Tail)
-	if err != nil {
-		return err
-	}
-
 	c, err := uncli.ConnectCluster(ctx)
 	if err != nil {
 		return fmt.Errorf("connect to cluster: %w", err)
@@ -88,8 +93,8 @@ func runLogs(ctx context.Context, uncli *cli.CLI, services []string, opts logs.O
 	logsOpts := api.ServiceLogsOptions{
 		Follow:   opts.Follow,
 		Tail:     tail,
-		Since:    opts.Since,
+		Since:    since,
-		Until:    opts.Until,
+		Until:    until,
 		Machines: cli.ExpandCommaSeparatedValues(opts.Machines),
 	}
 

cmd/uc/machine/logs_test.go

@@ -0,0 +1,25 @@
+package machine
+
+import (
+	"context"
+	"testing"
+
+	"github.com/psviderski/uncloud/internal/cli/logs"
+	"github.com/stretchr/testify/require"
+)
+
+func TestRunLogsInvalidTimeFilters(t *testing.T) {
+	t.Parallel()
+
+	// Invalid filters must fail before connecting to the cluster.
+	for _, flag := range []string{"since", "until"} {
+		opts := logs.Options{}
+		if flag == "since" {
+			opts.Since = "invalid"
+		} else {
+			opts.Until = "invalid"
+		}
+		err := runLogs(context.Background(), nil, nil, opts)
+		require.ErrorContains(t, err, "invalid --"+flag+" value")
+	}
+}

cmd/uc/machine/rtt.go

@@ -54,8 +54,8 @@ func rtt(ctx context.Context, uncli *cli.CLI) error {
 			tui.PrintWarning("metadata is missing in response from unknown server")
 			continue
 		}
-		if m.Metadata.Error != "" {
+		if err := m.Metadata.Err(); err != nil {
-			tui.PrintWarning(fmt.Sprintf("failed to inspect machine '%s': %s", m.Metadata.MachineName, m.Metadata.Error))
+			tui.PrintWarning(fmt.Sprintf("failed to inspect machine '%s': %s", m.Metadata.MachineName, err))
 			continue
 		}
 		if m.Machine == nil {

cmd/uc/ps.go

@@ -207,9 +207,9 @@ func collectContainers(ctx context.Context, cli *client.Client) ([]containerInfo
 
 		machineName := msc.Metadata.MachineName
 
-		if msc.Metadata.Error != "" {
+		if err := msc.Metadata.Err(); err != nil {
 			tui.PrintWarning(fmt.Sprintf("failed to list service containers on machine %s: %s",
-				machineName, msc.Metadata.Error))
+				machineName, err))
 			continue
 		}
 

cmd/uc/service/inspect.go

@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"fmt"
+	"maps"
 	"slices"
 	"time"
 
@@ -58,6 +59,16 @@ func inspect(ctx context.Context, uncli *cli.CLI, opts inspectOptions) error {
 	fmt.Printf("Service ID: %s\n", svc.ID)
 	fmt.Printf("Name:       %s\n", svc.Name)
 	fmt.Printf("Mode:       %s\n", svc.Mode)
+
+	spec := svc.Spec()
+	if len(spec.Labels) > 0 {
+		fmt.Println("Labels:")
+		printLabels(spec.Labels)
+	}
+	if len(spec.Container.Labels) > 0 {
+		fmt.Println("Container labels:")
+		printLabels(spec.Container.Labels)
+	}
 	fmt.Println()
 
 	// Combine regular and hook containers.
@@ -130,3 +141,10 @@ func inspect(ctx context.Context, uncli *cli.CLI, opts inspectOptions) error {
 	lipgloss.Println(t)
 	return nil
 }
+
+// printLabels prints the labels sorted by key followed by a blank line.
+func printLabels(labels map[string]string) {
+	for _, key := range slices.Sorted(maps.Keys(labels)) {
+		lipgloss.Println("  " + key + tui.Faint.Render("=") + labels[key])
+	}
+}

cmd/uc/service/logs.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 	"fmt"
 	"strings"
+	"time"
 
 	mapset "github.com/deckarep/golang-set/v2"
 	"github.com/psviderski/uncloud/internal/cli"
@@ -78,6 +79,15 @@ If no services are specified, streams logs from all services defined in the Comp
 }
 
 func RunLogs(ctx context.Context, uncli *cli.CLI, args []string, opts logs.Options) error {
+	since, until, err := logs.TimeRange(opts.Since, opts.Until, time.Now())
+	if err != nil {
+		return err
+	}
+	tail, err := opts.TailLines()
+	if err != nil {
+		return err
+	}
+
 	serviceArgs, err := logs.ParseServiceArgs(args)
 	if err != nil {
 		return err
@@ -106,12 +116,6 @@ func RunLogs(ctx context.Context, uncli *cli.CLI, args []string, opts logs.Optio
 		}
 	}
 
-	// Parse tail option.
-	tail, err := logs.Tail(opts.Tail)
-	if err != nil {
-		return err
-	}
-
 	c, err := uncli.ConnectCluster(ctx)
 	if err != nil {
 		return fmt.Errorf("connect to cluster: %w", err)
@@ -121,8 +125,8 @@ func RunLogs(ctx context.Context, uncli *cli.CLI, args []string, opts logs.Optio
 	baseOpts := api.ServiceLogsOptions{
 		Follow:   opts.Follow,
 		Tail:     tail,
-		Since:    opts.Since,
+		Since:    since,
-		Until:    opts.Until,
+		Until:    until,
 		Machines: cli.ExpandCommaSeparatedValues(opts.Machines),
 	}
 

cmd/uc/service/logs_test.go

@@ -0,0 +1,25 @@
+package service
+
+import (
+	"context"
+	"testing"
+
+	"github.com/psviderski/uncloud/internal/cli/logs"
+	"github.com/stretchr/testify/require"
+)
+
+func TestRunLogsInvalidTimeFilters(t *testing.T) {
+	t.Parallel()
+
+	// Invalid filters must fail before loading Compose files or connecting to the cluster.
+	for _, flag := range []string{"since", "until"} {
+		opts := logs.Options{}
+		if flag == "since" {
+			opts.Since = "invalid"
+		} else {
+			opts.Until = "invalid"
+		}
+		err := RunLogs(context.Background(), nil, nil, opts)
+		require.ErrorContains(t, err, "invalid --"+flag+" value")
+	}
+}

cmd/uc/service/root.go

@@ -20,6 +20,7 @@ func NewRootCommand() *cobra.Command {
 		NewScaleCommand(""),
 		NewStartCommand(""),
 		NewStopCommand(""),
+		NewUpdateCommand(),
 	)
 	return cmd
 }

cmd/uc/service/run.go

@@ -21,13 +21,16 @@ import (
 type runOptions struct {
 	caddyfile         string
 	command           []string
+	containerLabels   []string
 	cpu               dockeropts.NanoCPUs
 	entrypoint        string
 	entrypointChanged bool
 	env               []string
 	image             string
+	labels            []string
 	machines          []string
 	memory            dockeropts.MemBytes
+	memoryReservation dockeropts.MemBytes
 	mode              string
 	shmSize           dockeropts.MemBytes
 	name              string
@@ -64,6 +67,9 @@ func NewRunCommand(groupID string) *cobra.Command {
 	cmd.Flags().StringVar(&opts.caddyfile, "caddyfile", "",
 		"Path to a custom Caddy config (Caddyfile) for the service. "+
 			"Cannot be used together with non-@host published ports.")
+	cmd.Flags().StringArrayVar(&opts.containerLabels, "container-label", nil,
+		"Set a Docker label on service containers. Can be specified multiple times.\n"+
+			"Format: key=value or just key for an empty value.")
 	cmd.Flags().VarP(&opts.cpu, "cpu", "",
 		"Maximum number of CPU cores a service container can use. Fractional values are allowed: "+
 			"0.5 for half a core or 2.25 for two and a quarter cores.")
@@ -72,6 +78,9 @@ func NewRunCommand(groupID string) *cobra.Command {
 	cmd.Flags().StringSliceVarP(&opts.env, "env", "e", nil,
 		"Set an environment variable for service containers. Can be specified multiple times.\n"+
 			"Format: VAR=value or just VAR to use the value from the local environment.")
+	cmd.Flags().StringArrayVarP(&opts.labels, "label", "l", nil,
+		"Set a service label. Service labels aren't applied to Docker containers, use --container-label for that.\n"+
+			"Can be specified multiple times. Format: key=value or just key for an empty value.")
 	cmd.Flags().StringVar(&opts.mode, "mode", api.ServiceModeReplicated,
 		fmt.Sprintf("Replication mode of the service: either '%s' (a specified number of containers across "+
 			"the machines) or '%s' (one container on every machine).",
@@ -83,6 +92,9 @@ func NewRunCommand(groupID string) *cobra.Command {
 		"Maximum amount of memory a service container can use. Value is a positive integer with optional unit suffix (b, k, m, g).\n"+
 			"Default unit is bytes if no suffix specified.\n"+
 			"Examples: 1073741824, 1024m, 1g (all equal 1 gibibyte)")
+	cmd.Flags().Var(&opts.memoryReservation, "memory-reservation",
+		"Minimum amount of memory reserved for a service container. Value is a positive integer with optional unit suffix (b, k, m, g).\n"+
+			"Default unit is bytes if no suffix specified.")
 	cmd.Flags().VarP(&opts.shmSize, "shm-size", "",
 		"Maximum amount of shared memory (mounted at /dev/shm) a service container can use. Value is a positive integer\n"+
 			"with optional unit suffix (b, k, m, g). Default unit is bytes if no suffix specified.\n"+
@@ -94,11 +106,12 @@ func NewRunCommand(groupID string) *cobra.Command {
 	cmd.Flags().StringSliceVarP(&opts.publish, "publish", "p", nil,
 		"Publish a service port to make it accessible outside the cluster. Can be specified multiple times.\n"+
 			"Format: [hostname:]container_port[/protocol] or [host_ip|host_prefix:]host_port:container_port[/protocol]@host\n"+
-			"Supported protocols: tcp, udp, http, https (default is tcp). If a hostname for http(s) port is not specified\n"+
+			"Supported protocols: tcp, udp, http, https. The protocol defaults to https if a hostname is specified,\n"+
-			"and a cluster domain is reserved, service-name.cluster-domain will be used as the hostname.\n"+
+			"otherwise tcp. http(s) ports without a hostname use 'service-name.cluster-domain' as the default hostname\n"+
+			"if a cluster domain is reserved.\n"+
 			"Examples:\n"+
 			"  -p 8080/https                       Publish port 8080 as HTTPS via reverse proxy with default service-name.cluster-domain hostname\n"+
-			"  -p app.example.com:8080/https       Publish port 8080 as HTTPS via reverse proxy with custom hostname\n"+
+			"  -p app.example.com:8080             Publish port 8080 as HTTPS via reverse proxy with custom hostname\n"+
 			// TODO: add support for publishing L4 tcp/udp ports.
 			//"  -p 9000:8080                   Publish port 8080 as TCP port 9000 via reverse proxy\n"+
 			"  -p 53:5353/udp@host                 Bind UDP port 5353 to host port 53\n"+
@@ -192,6 +205,18 @@ func prepareServiceSpec(opts runOptions) (api.ServiceSpec, error) {
 		return spec, err
 	}
 
+	labels, err := parseLabels(opts.labels)
+	if err != nil {
+		return spec, err
+	}
+	if err = api.ValidateLabels(labels); err != nil {
+		return spec, fmt.Errorf("invalid service labels: %w", err)
+	}
+	containerLabels, err := parseLabels(opts.containerLabels)
+	if err != nil {
+		return spec, err
+	}
+
 	switch opts.mode {
 	case api.ServiceModeReplicated, api.ServiceModeGlobal:
 	default:
@@ -232,17 +257,20 @@ func prepareServiceSpec(opts runOptions) (api.ServiceSpec, error) {
 			Command:    opts.command,
 			Env:        env,
 			Image:      opts.image,
+			Labels:     containerLabels,
 			Privileged: opts.privileged,
 			PullPolicy: opts.pull,
 			Resources: api.ContainerResources{
 				CPU:               opts.cpu.Value(),
 				Memory:            opts.memory.Value(),
+				MemoryReservation: opts.memoryReservation.Value(),
 				SharedMemory:      opts.shmSize.Value(),
 				Ulimits:           ulimits,
 			},
 			User:         opts.user,
 			VolumeMounts: mounts,
 		},
+		Labels:    labels,
 		Mode:      opts.mode,
 		Name:      opts.name,
 		Placement: placement,
@@ -301,6 +329,24 @@ func parseEnv(env []string) (api.EnvVars, error) {
 	return envVars, nil
 }
 
+// parseLabels parses labels from the command line arguments in the format "key=value" or just "key" for an empty value.
+func parseLabels(labels []string) (map[string]string, error) {
+	if len(labels) == 0 {
+		return nil, nil
+	}
+
+	result := make(map[string]string, len(labels))
+	for _, label := range labels {
+		key, value, _ := strings.Cut(label, "=")
+		if strings.TrimSpace(key) == "" {
+			return nil, fmt.Errorf("invalid label '%s': key must not be empty", label)
+		}
+		result[key] = value
+	}
+
+	return result, nil
+}
+
 // parseVolumeFlags parses volume flag values in Docker CLI format and returns VolumeSpecs and VolumeMounts.
 // It handles both named volumes (volume_name:/container/path[:ro|volume-nocopy])
 // and bind mounts (/host/path:/container/path[:ro]).

cmd/uc/service/run_test.go

@@ -0,0 +1,423 @@
+package service
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/docker/go-units"
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// validRunOptions returns the minimal options that produce a valid service spec.
+func validRunOptions() runOptions {
+	return runOptions{
+		image:    "nginx:1.27",
+		name:     "web",
+		mode:     api.ServiceModeReplicated,
+		pull:     api.PullPolicyMissing,
+		replicas: 1,
+	}
+}
+
+func TestParseLabels(t *testing.T) {
+	t.Parallel()
+
+	labels, err := parseLabels(nil)
+	require.NoError(t, err)
+	assert.Nil(t, labels)
+
+	labels, err = parseLabels([]string{"tier=frontend", "empty", "list=a,b", "expr=a=b", "tier=backend"})
+	require.NoError(t, err)
+	assert.Equal(t, map[string]string{
+		"tier":  "backend",
+		"empty": "",
+		"list":  "a,b",
+		"expr":  "a=b",
+	}, labels)
+
+	_, err = parseLabels([]string{"=value"})
+	require.ErrorContains(t, err, "key must not be empty")
+
+	_, err = parseLabels([]string{" \t =value"})
+	require.ErrorContains(t, err, "key must not be empty")
+
+	labels, err = parseLabels([]string{"key=value with spaces"})
+	require.NoError(t, err)
+	assert.Equal(t, map[string]string{"key": "value with spaces"}, labels)
+}
+
+func TestPrepareServiceSpec(t *testing.T) {
+	t.Parallel()
+
+	t.Run("minimal options", func(t *testing.T) {
+		t.Parallel()
+
+		spec, err := prepareServiceSpec(validRunOptions())
+		require.NoError(t, err)
+
+		assert.Equal(t, "web", spec.Name)
+		assert.Equal(t, api.ServiceModeReplicated, spec.Mode)
+		assert.EqualValues(t, 1, spec.Replicas)
+		assert.Equal(t, "nginx:1.27", spec.Container.Image)
+		assert.Equal(t, api.PullPolicyMissing, spec.Container.PullPolicy)
+		assert.Nil(t, spec.Caddy)
+		assert.Nil(t, spec.Labels)
+		assert.Nil(t, spec.Container.Labels)
+		assert.Nil(t, spec.Container.Entrypoint)
+		assert.Nil(t, spec.Container.Resources.Ulimits)
+		assert.Nil(t, spec.Placement.Machines)
+		assert.Empty(t, spec.Container.Env)
+		assert.Empty(t, spec.Ports)
+		assert.Empty(t, spec.Volumes)
+		assert.Empty(t, spec.Container.VolumeMounts)
+	})
+
+	t.Run("all options", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.command = []string{"nginx", "-g", "daemon off;"}
+		opts.env = []string{"FOO=bar", "EMPTY=", "EXPR=a=b"}
+		opts.labels = []string{"tier=frontend"}
+		opts.containerLabels = []string{"logging=enabled"}
+		opts.machines = []string{"m1,m2", "m3"}
+		opts.privileged = true
+		opts.pull = api.PullPolicyAlways
+		opts.publish = []string{"app.example.com:8080/https", "53:5353/udp@host"}
+		opts.replicas = 3
+		opts.ulimits = []string{"nofile=1024:2048", "nproc=65535"}
+		opts.user = "1000:1000"
+		opts.volumes = []string{"data:/var/lib/data:ro"}
+		require.NoError(t, opts.cpu.Set("0.5"))
+		require.NoError(t, opts.memory.Set("128m"))
+		require.NoError(t, opts.memoryReservation.Set("32m"))
+		require.NoError(t, opts.shmSize.Set("64m"))
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+
+		assert.Equal(t, api.ServiceSpec{
+			Container: api.ContainerSpec{
+				Command:    []string{"nginx", "-g", "daemon off;"},
+				Env:        api.EnvVars{"FOO": "bar", "EMPTY": "", "EXPR": "a=b"},
+				Image:      "nginx:1.27",
+				Labels:     map[string]string{"logging": "enabled"},
+				Privileged: true,
+				PullPolicy: api.PullPolicyAlways,
+				Resources: api.ContainerResources{
+					CPU:               500 * api.MilliCore,
+					Memory:            128 * units.MiB,
+					MemoryReservation: 32 * units.MiB,
+					SharedMemory:      64 * units.MiB,
+					Ulimits: map[string]api.Ulimit{
+						"nofile": {Soft: 1024, Hard: 2048},
+						"nproc":  {Soft: 65535, Hard: 65535},
+					},
+				},
+				User: "1000:1000",
+				VolumeMounts: []api.VolumeMount{
+					{VolumeName: "data", ContainerPath: "/var/lib/data", ReadOnly: true},
+				},
+			},
+			Labels:    map[string]string{"tier": "frontend"},
+			Mode:      api.ServiceModeReplicated,
+			Name:      "web",
+			Placement: api.Placement{Machines: []string{"m1", "m2", "m3"}},
+			Ports: []api.PortSpec{
+				{
+					Hostname:      "app.example.com",
+					ContainerPort: 8080,
+					Protocol:      api.ProtocolHTTPS,
+					Mode:          api.PortModeIngress,
+				},
+				{
+					PublishedPort: 53,
+					ContainerPort: 5353,
+					Protocol:      api.ProtocolUDP,
+					Mode:          api.PortModeHost,
+				},
+			},
+			Replicas: 3,
+			Volumes: []api.VolumeSpec{
+				{
+					Name:          "data",
+					Type:          api.VolumeTypeVolume,
+					VolumeOptions: &api.VolumeOptions{Name: "data"},
+				},
+			},
+		}, spec)
+	})
+
+	t.Run("generates a name when not specified", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.name = ""
+		opts.image = "bitnami/nginx:latest"
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Regexp(t, `^nginx-[a-z0-9]{4}$`, spec.Name)
+	})
+
+	t.Run("keeps the specified name", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.name = "custom"
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Equal(t, "custom", spec.Name)
+	})
+
+	t.Run("global mode", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.mode = api.ServiceModeGlobal
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Equal(t, api.ServiceModeGlobal, spec.Mode)
+	})
+
+	t.Run("entrypoint", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.entrypoint = "/bin/sh"
+		opts.entrypointChanged = true
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Equal(t, []string{"/bin/sh"}, spec.Container.Entrypoint)
+
+		// An explicitly empty entrypoint resets the one from the image.
+		opts = validRunOptions()
+		opts.entrypointChanged = true
+		spec, err = prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Equal(t, []string{""}, spec.Container.Entrypoint)
+	})
+
+	t.Run("env without a value is omitted when not set locally", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.env = []string{"UNCLOUD_TEST_ENV_THAT_IS_NOT_SET"}
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.NotContains(t, spec.Container.Env, "UNCLOUD_TEST_ENV_THAT_IS_NOT_SET")
+	})
+
+	t.Run("bind mount", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.volumes = []string{"/data/uploads:/app/uploads"}
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+
+		require.Len(t, spec.Volumes, 1)
+		assert.Regexp(t, `^bind-[a-z0-9]{4}$`, spec.Volumes[0].Name)
+		assert.Equal(t, api.VolumeTypeBind, spec.Volumes[0].Type)
+		require.NotNil(t, spec.Volumes[0].BindOptions)
+		assert.Equal(t, "/data/uploads", spec.Volumes[0].BindOptions.HostPath)
+		assert.True(t, spec.Volumes[0].BindOptions.CreateHostPath)
+
+		require.Len(t, spec.Container.VolumeMounts, 1)
+		assert.Equal(t, spec.Volumes[0].Name, spec.Container.VolumeMounts[0].VolumeName)
+		assert.Equal(t, "/app/uploads", spec.Container.VolumeMounts[0].ContainerPath)
+	})
+
+	t.Run("named volume mounted twice creates one volume", func(t *testing.T) {
+		t.Parallel()
+
+		opts := validRunOptions()
+		opts.volumes = []string{"data:/a", "data:/b"}
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Len(t, spec.Volumes, 1)
+		assert.Len(t, spec.Container.VolumeMounts, 2)
+	})
+
+	t.Run("caddyfile content is trimmed", func(t *testing.T) {
+		t.Parallel()
+
+		path := filepath.Join(t.TempDir(), "Caddyfile")
+		require.NoError(t, os.WriteFile(path, []byte("\nexample.com {\n}\n\n"), 0o600))
+
+		opts := validRunOptions()
+		opts.caddyfile = path
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		require.NotNil(t, spec.Caddy)
+		assert.Equal(t, "example.com {\n}", spec.Caddy.Config)
+	})
+
+	t.Run("empty caddyfile is ignored", func(t *testing.T) {
+		t.Parallel()
+
+		path := filepath.Join(t.TempDir(), "Caddyfile")
+		require.NoError(t, os.WriteFile(path, []byte(" \n"), 0o600))
+
+		opts := validRunOptions()
+		opts.caddyfile = path
+
+		spec, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+		assert.Nil(t, spec.Caddy)
+	})
+
+	t.Run("caddyfile can be used with host ports", func(t *testing.T) {
+		t.Parallel()
+
+		path := filepath.Join(t.TempDir(), "Caddyfile")
+		require.NoError(t, os.WriteFile(path, []byte("example.com {\n}"), 0o600))
+
+		opts := validRunOptions()
+		opts.caddyfile = path
+		opts.publish = []string{"53:5353/udp@host"}
+
+		_, err := prepareServiceSpec(opts)
+		require.NoError(t, err)
+	})
+}
+
+func TestPrepareServiceSpec_LocalEnv(t *testing.T) {
+	// Not parallel because t.Setenv changes the environment of the whole process.
+	t.Setenv("UNCLOUD_TEST_LOCAL_ENV", "from-local")
+
+	opts := validRunOptions()
+	opts.env = []string{"UNCLOUD_TEST_LOCAL_ENV"}
+
+	spec, err := prepareServiceSpec(opts)
+	require.NoError(t, err)
+	assert.Equal(t, "from-local", spec.Container.Env["UNCLOUD_TEST_LOCAL_ENV"])
+}
+
+func TestPrepareServiceSpec_Invalid(t *testing.T) {
+	t.Parallel()
+
+	caddyfile := filepath.Join(t.TempDir(), "Caddyfile")
+	require.NoError(t, os.WriteFile(caddyfile, []byte("example.com {\n}"), 0o600))
+
+	tests := []struct {
+		name    string
+		modify  func(opts *runOptions)
+		wantErr string
+	}{
+		{
+			name:    "invalid image",
+			modify:  func(opts *runOptions) { opts.image = "INVALID IMAGE" },
+			wantErr: "invalid image",
+		},
+		{
+			name:    "invalid name",
+			modify:  func(opts *runOptions) { opts.name = "Not_Valid" },
+			wantErr: "invalid service name",
+		},
+		{
+			name:    "invalid mode",
+			modify:  func(opts *runOptions) { opts.mode = "unknown" },
+			wantErr: "invalid replication mode: 'unknown'",
+		},
+		{
+			name:    "invalid pull policy",
+			modify:  func(opts *runOptions) { opts.pull = "sometimes" },
+			wantErr: "invalid pull policy: 'sometimes'",
+		},
+		{
+			name:    "invalid env",
+			modify:  func(opts *runOptions) { opts.env = []string{"=value"} },
+			wantErr: "invalid environment variable: '=value'",
+		},
+		{
+			name:    "invalid port",
+			modify:  func(opts *runOptions) { opts.publish = []string{"not-a-port"} },
+			wantErr: "invalid service port 'not-a-port'",
+		},
+		{
+			name:    "invalid ulimit",
+			modify:  func(opts *runOptions) { opts.ulimits = []string{"nofile"} },
+			wantErr: "invalid ulimit 'nofile'",
+		},
+		{
+			name:    "volume without a container path",
+			modify:  func(opts *runOptions) { opts.volumes = []string{"data"} },
+			wantErr: "invalid volume mount 'data'",
+		},
+		{
+			name:    "volume with a relative container path",
+			modify:  func(opts *runOptions) { opts.volumes = []string{"data:relative"} },
+			wantErr: "must be absolute path",
+		},
+		{
+			name:    "volume with an unknown option",
+			modify:  func(opts *runOptions) { opts.volumes = []string{"data:/data:unknown"} },
+			wantErr: "invalid option: 'unknown'",
+		},
+		{
+			name:    "missing caddyfile",
+			modify:  func(opts *runOptions) { opts.caddyfile = filepath.Join(os.TempDir(), "uncloud-test-no-such-file") },
+			wantErr: "read Caddyfile",
+		},
+		{
+			name: "caddyfile with an ingress port",
+			modify: func(opts *runOptions) {
+				opts.caddyfile = caddyfile
+				opts.publish = []string{"app.example.com:8080/https"}
+			},
+			wantErr: "cannot be specified simultaneously",
+		},
+		{
+			name:    "service label without a key",
+			modify:  func(opts *runOptions) { opts.labels = []string{"=value"} },
+			wantErr: "key must not be empty",
+		},
+		{
+			name:    "container label without a key",
+			modify:  func(opts *runOptions) { opts.containerLabels = []string{"=value"} },
+			wantErr: "key must not be empty",
+		},
+		{
+			name:    "service label with whitespace",
+			modify:  func(opts *runOptions) { opts.labels = []string{"my label=value"} },
+			wantErr: "invalid service labels: label 'my label' must not contain whitespace",
+		},
+		{
+			name:    "container label with whitespace",
+			modify:  func(opts *runOptions) { opts.containerLabels = []string{"my label=value"} },
+			wantErr: "invalid container labels: label 'my label' must not contain whitespace",
+		},
+		{
+			name:    "reserved service label",
+			modify:  func(opts *runOptions) { opts.labels = []string{api.LabelManaged + "=true"} },
+			wantErr: "invalid service labels",
+		},
+		{
+			name:    "reserved container label",
+			modify:  func(opts *runOptions) { opts.containerLabels = []string{api.LabelManaged + "=true"} },
+			wantErr: "invalid container labels",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			opts := validRunOptions()
+			tt.modify(&opts)
+
+			_, err := prepareServiceSpec(opts)
+			require.ErrorContains(t, err, tt.wantErr)
+		})
+	}
+}

cmd/uc/service/update.go

@@ -0,0 +1,183 @@
+package service
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"maps"
+	"slices"
+
+	dockeropts "github.com/docker/cli/opts"
+	"github.com/docker/compose/v2/pkg/progress"
+	"github.com/docker/docker/pkg/stringid"
+	"github.com/psviderski/uncloud/internal/cli"
+	"github.com/psviderski/uncloud/internal/cli/completion"
+	"github.com/psviderski/uncloud/internal/cli/tui"
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/psviderski/uncloud/pkg/client/deploy"
+	"github.com/spf13/cobra"
+)
+
+type updateOptions struct {
+	service string
+
+	cpu                dockeropts.NanoCPUs
+	cpuChanged         bool
+	memory             dockeropts.MemBytes
+	memoryChanged      bool
+	memoryReservation  dockeropts.MemBytes
+	reservationChanged bool
+	labelAdd           []string
+	labelRm            []string
+}
+
+func NewUpdateCommand() *cobra.Command {
+	opts := updateOptions{}
+	cmd := &cobra.Command{
+		Use:   "update SERVICE",
+		Short: "Update a service in place without recreating its containers.",
+		Long: `Update service configuration in place without recreating its containers.
+
+A CPU or memory limit can be changed but not removed. To remove a limit or change anything else,
+redeploy the service with 'uc deploy'`,
+		Example: `  # Change the CPU and memory limits.
+  uc service update web --cpu 0.5 --memory 512m
+
+  # Add or change a service label and remove another one.
+  uc service update web --label-add tier=frontend --label-rm legacy`,
+		Args: cobra.ExactArgs(1),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			uncli := cmd.Context().Value("cli").(*cli.CLI)
+
+			opts.service = args[0]
+			opts.cpuChanged = cmd.Flag("cpu").Changed
+			opts.memoryChanged = cmd.Flag("memory").Changed
+			opts.reservationChanged = cmd.Flag("memory-reservation").Changed
+
+			return update(cmd.Context(), uncli, opts)
+		},
+		ValidArgsFunction: func(
+			cmd *cobra.Command, args []string, toComplete string,
+		) ([]cobra.Completion, cobra.ShellCompDirective) {
+			if len(args) > 0 {
+				return nil, cobra.ShellCompDirectiveNoFileComp
+			}
+			uncli := cmd.Context().Value("cli").(*cli.CLI)
+			return completion.Services(cmd.Context(), uncli, args, toComplete)
+		},
+	}
+
+	cmd.Flags().Var(&opts.cpu, "cpu",
+		"Maximum number of CPU cores a service container can use.\n"+
+			"Fractional values are allowed: 0.5 for half a core or 2.25 for two and a quarter cores.")
+	cmd.Flags().Var(&opts.memory, "memory",
+		"Maximum amount of memory a service container can use. Value is a positive integer with\n"+
+			"optional unit suffix (b, k, m, g). Default unit is bytes if no suffix specified.\n"+
+			"Examples: 1073741824, 1024m, 1g (all equal 1 gibibyte)")
+	cmd.Flags().Var(&opts.memoryReservation, "memory-reservation",
+		"Minimum amount of memory reserved for a service container. Value is a positive integer with\n"+
+			"optional unit suffix (b, k, m, g). Default unit is bytes if no suffix specified.")
+	cmd.Flags().StringArrayVar(&opts.labelAdd, "label-add", nil,
+		"Add or change a service label. Can be specified multiple times.\n"+
+			"Format: key=value or just key for an empty value.")
+	cmd.Flags().StringArrayVar(&opts.labelRm, "label-rm", nil,
+		"Remove a service label by its key. Can be specified multiple times.")
+
+	return cmd
+}
+
+func update(ctx context.Context, uncli *cli.CLI, opts updateOptions) error {
+	if !opts.cpuChanged && !opts.memoryChanged && !opts.reservationChanged &&
+		len(opts.labelAdd) == 0 && len(opts.labelRm) == 0 {
+		return errors.New("no changes specified, set at least one of the flags: " +
+			"--cpu, --memory, --memory-reservation, --label-add, --label-rm")
+	}
+
+	labelAdd, err := parseLabels(opts.labelAdd)
+	if err != nil {
+		return err
+	}
+	if err = api.ValidateLabels(labelAdd); err != nil {
+		return fmt.Errorf("invalid service labels: %w", err)
+	}
+	for _, key := range opts.labelRm {
+		if _, ok := labelAdd[key]; ok {
+			return fmt.Errorf("label '%s' cannot be both added and removed", key)
+		}
+	}
+
+	clusterClient, err := uncli.ConnectCluster(ctx)
+	if err != nil {
+		return fmt.Errorf("connect to cluster: %w", err)
+	}
+	defer clusterClient.Close()
+
+	svc, err := clusterClient.InspectService(ctx, opts.service)
+	if err != nil {
+		return fmt.Errorf("inspect service '%s': %w", opts.service, err)
+	}
+
+	// Check all containers before changing any of them to avoid a partially updated service when a limit can't be
+	// changed in place. Containers may have different specs after a failed deployment so each one is updated from
+	// its own spec.
+	updates := make(map[string]api.UpdateContainerOptions, len(svc.Containers))
+	for _, mc := range svc.Containers {
+		spec := applyServiceUpdate(mc.Container.ServiceSpec, opts, labelAdd)
+
+		resources := mc.Container.AppliedResources()
+		if err = resources.ValidateUpdateInPlace(spec.Container.Resources); err != nil {
+			return fmt.Errorf("container '%s' on machine '%s' can't be updated without recreating it: %w",
+				stringid.TruncateID(mc.Container.ID), mc.MachineName, err)
+		}
+
+		containerOpts := deploy.ContainerUpdateOptions(mc.Container, spec)
+		if containerOpts != (api.UpdateContainerOptions{}) {
+			updates[mc.Container.ID] = containerOpts
+		}
+	}
+
+	if len(updates) == 0 {
+		fmt.Printf("Service %s is already up to date. No changes required.\n", tui.NameStyle.Render(svc.Name))
+		return nil
+	}
+
+	return progress.RunWithTitle(ctx, func(ctx context.Context) error {
+		// Keep updating the remaining containers if one fails, for example, because its machine is down.
+		// Running the command again only updates the containers that still differ.
+		var errs []error
+		for id, containerOpts := range updates {
+			if err := clusterClient.UpdateContainer(ctx, svc.ID, id, containerOpts); err != nil {
+				errs = append(errs, fmt.Errorf("update container '%s': %w", stringid.TruncateID(id), err))
+			}
+		}
+		return errors.Join(errs...)
+	}, uncli.ProgressOut(), "Updating service "+tui.NameStyle.Render(svc.Name))
+}
+
+// applyServiceUpdate returns a copy of the spec with the requested changes applied.
+func applyServiceUpdate(spec api.ServiceSpec, opts updateOptions, labelAdd map[string]string) api.ServiceSpec {
+	updated := spec.Clone()
+
+	if opts.cpuChanged {
+		updated.Container.Resources.CPU = opts.cpu.Value()
+	}
+	if opts.memoryChanged {
+		updated.Container.Resources.Memory = opts.memory.Value()
+	}
+	if opts.reservationChanged {
+		updated.Container.Resources.MemoryReservation = opts.memoryReservation.Value()
+	}
+
+	if len(labelAdd) > 0 && updated.Labels == nil {
+		updated.Labels = make(map[string]string, len(labelAdd))
+	}
+	maps.Copy(updated.Labels, labelAdd)
+
+	for key := range updated.Labels {
+		if slices.Contains(opts.labelRm, key) {
+			delete(updated.Labels, key)
+		}
+	}
+
+	return updated
+}

cmd/uc/service/update_test.go

@@ -0,0 +1,106 @@
+package service
+
+import (
+	"testing"
+
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/psviderski/uncloud/pkg/client/deploy"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestApplyServiceUpdate(t *testing.T) {
+	t.Parallel()
+
+	newSpec := func() api.ServiceSpec {
+		return api.ServiceSpec{
+			Name: "web",
+			Container: api.ContainerSpec{
+				Image:  "nginx",
+				Labels: map[string]string{"container": "keep"},
+				Resources: api.ContainerResources{
+					CPU:               100 * api.MilliCore,
+					Memory:            1024,
+					MemoryReservation: 512,
+				},
+			},
+			Labels: map[string]string{"tier": "frontend", "legacy": "true"},
+		}
+	}
+
+	t.Run("adds, changes and removes labels", func(t *testing.T) {
+		t.Parallel()
+
+		spec := newSpec()
+		updated := applyServiceUpdate(spec, updateOptions{labelRm: []string{"legacy", "missing"}},
+			map[string]string{"tier": "backend", "team": "core"})
+
+		assert.Equal(t, map[string]string{"tier": "backend", "team": "core"}, updated.Labels)
+		assert.Equal(t, map[string]string{"container": "keep"}, updated.Container.Labels)
+		assert.Equal(t, spec.Container.Resources, updated.Container.Resources)
+		assert.Equal(t, map[string]string{"tier": "frontend", "legacy": "true"}, spec.Labels,
+			"Expected the original spec to stay unchanged")
+	})
+
+	t.Run("adds labels to a spec without labels", func(t *testing.T) {
+		t.Parallel()
+
+		spec := newSpec()
+		spec.Labels = nil
+		updated := applyServiceUpdate(spec, updateOptions{}, map[string]string{"tier": "backend"})
+
+		assert.Equal(t, map[string]string{"tier": "backend"}, updated.Labels)
+	})
+
+	t.Run("changes only requested limits", func(t *testing.T) {
+		t.Parallel()
+
+		opts := updateOptions{cpuChanged: true}
+		require.NoError(t, opts.cpu.Set("0.5"))
+
+		updated := applyServiceUpdate(newSpec(), opts, nil)
+
+		assert.EqualValues(t, 500*api.MilliCore, updated.Container.Resources.CPU)
+		assert.EqualValues(t, 1024, updated.Container.Resources.Memory)
+		assert.EqualValues(t, 512, updated.Container.Resources.MemoryReservation)
+	})
+
+	t.Run("sends only changed fields to the container", func(t *testing.T) {
+		t.Parallel()
+
+		opts := updateOptions{memoryChanged: true}
+		require.NoError(t, opts.memory.Set("2048"))
+
+		ctr := api.ServiceContainer{ServiceSpec: newSpec()}
+		containerOpts := deploy.ContainerUpdateOptions(ctr, applyServiceUpdate(ctr.ServiceSpec, opts, nil))
+
+		require.NotNil(t, containerOpts.Memory)
+		assert.EqualValues(t, 2048, *containerOpts.Memory)
+		assert.Nil(t, containerOpts.CPU)
+		assert.Nil(t, containerOpts.MemoryReservation)
+		assert.Nil(t, containerOpts.ServiceLabels)
+	})
+
+	t.Run("removing the last label clears service labels", func(t *testing.T) {
+		t.Parallel()
+
+		ctr := api.ServiceContainer{ServiceSpec: newSpec()}
+		containerOpts := deploy.ContainerUpdateOptions(ctr, applyServiceUpdate(
+			ctr.ServiceSpec, updateOptions{labelRm: []string{"tier", "legacy"}}, nil,
+		))
+
+		require.NotNil(t, containerOpts.ServiceLabels)
+		assert.Empty(t, *containerOpts.ServiceLabels)
+	})
+
+	t.Run("no changes produce empty container options", func(t *testing.T) {
+		t.Parallel()
+
+		ctr := api.ServiceContainer{ServiceSpec: newSpec()}
+		containerOpts := deploy.ContainerUpdateOptions(ctr, applyServiceUpdate(
+			ctr.ServiceSpec, updateOptions{}, map[string]string{"tier": "frontend"},
+		))
+
+		assert.Equal(t, api.UpdateContainerOptions{}, containerOpts)
+	})
+}

internal/cli/logs/duration.go

@@ -0,0 +1,66 @@
+package logs
+
+import (
+	"fmt"
+	"math/big"
+	"strings"
+	"time"
+)
+
+// Duration is an extended standard time.Duration that also supports days as a unit.
+type Duration = time.Duration
+
+// ParseDuration parses a Go duration with the additional unit d, meaning exactly 24 hours.
+// Days can be fractional or combined with other units, such as "1.5d" or "2d3h".
+func ParseDuration(value string) (Duration, error) {
+	if !strings.Contains(value, "d") {
+		return time.ParseDuration(value)
+	}
+
+	rest := value
+	var normalised strings.Builder
+	if len(rest) > 0 && (rest[0] == '-' || rest[0] == '+') {
+		normalised.WriteByte(rest[0])
+		rest = rest[1:]
+	}
+	for len(rest) > 0 {
+		// Each component is a decimal number followed by a unit. Only the leading sign is allowed.
+		numberEnd := 0
+		for numberEnd < len(rest) && (isDigit(rest[numberEnd]) || rest[numberEnd] == '.') {
+			numberEnd++
+		}
+		unitEnd := numberEnd
+		for unitEnd < len(rest) && !isDigit(rest[unitEnd]) && rest[unitEnd] != '.' {
+			unitEnd++
+		}
+		if numberEnd == 0 || unitEnd == numberEnd {
+			return 0, fmt.Errorf("time: invalid duration '%s'", value)
+		}
+		number, unit := rest[:numberEnd], rest[numberEnd:unitEnd]
+		if unit == "d" {
+			// Use exact decimal arithmetic to avoid float rounding and retain nanosecond precision.
+			days, ok := new(big.Rat).SetString(number)
+			if !ok {
+				return 0, fmt.Errorf("time: invalid duration '%s'", value)
+			}
+			days.Mul(days, new(big.Rat).SetInt64(int64(24*time.Hour)))
+			nanoseconds := new(big.Int).Quo(days.Num(), days.Denom())
+			normalised.WriteString(nanoseconds.String())
+			normalised.WriteString("ns")
+		} else {
+			normalised.WriteString(rest[:unitEnd])
+		}
+		rest = rest[unitEnd:]
+	}
+
+	// The standard parser validates the remaining units and checks the total for overflow.
+	duration, err := time.ParseDuration(normalised.String())
+	if err != nil {
+		return 0, fmt.Errorf("time: invalid duration '%s': %w", value, err)
+	}
+	return duration, nil
+}
+
+func isDigit(c byte) bool {
+	return c >= '0' && c <= '9'
+}

internal/cli/logs/duration_test.go

@@ -0,0 +1,62 @@
+package logs
+
+import (
+	"math"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestParseDuration(t *testing.T) {
+	t.Parallel()
+
+	for _, tt := range []struct {
+		input string
+		want  time.Duration
+	}{
+		{"0", 0},
+		{"2m30s", 150 * time.Second},
+		{"1µs", time.Microsecond},
+		{"1μs", time.Microsecond},
+		{"2d", 48 * time.Hour},
+		{"010d", 240 * time.Hour},
+		{"08d", 192 * time.Hour},
+		{"2d3h", 51 * time.Hour},
+		{"3h2d", 51 * time.Hour},
+		{"1d1d", 48 * time.Hour},
+		{"1.5d", 36 * time.Hour},
+		{".5d", 12 * time.Hour},
+		{"1.d", 24 * time.Hour},
+		{"-2d3h", -51 * time.Hour},
+		{"+2d", 48 * time.Hour},
+		{"0d", 0},
+		{"0d1ns", time.Nanosecond},
+		{"0.000000000001d", 86 * time.Nanosecond},
+		{"-0.000000000001d", -86 * time.Nanosecond},
+		{"106751d23h47m16.854775807s", time.Duration(math.MaxInt64)},
+		{"-106751d23h47m16.854775808s", time.Duration(math.MinInt64)},
+	} {
+		t.Run(tt.input, func(t *testing.T) {
+			actual, err := ParseDuration(tt.input)
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, actual)
+		})
+	}
+}
+
+func TestParseDuration_Invalid(t *testing.T) {
+	t.Parallel()
+
+	for _, input := range []string{
+		"", "d", "2d3", "1..2d", ".d", "1d-2h", "1d+2h", "1d 2h", "1day", "2D", "1w",
+		"1e2d", "1d2w", "106752d", "-106752d", "106751d23h47m16.854775808s",
+		"-106751d23h47m16.854775809s", "999999999999999999999999999d",
+	} {
+		t.Run(input, func(t *testing.T) {
+			_, err := ParseDuration(input)
+			require.Error(t, err)
+		})
+	}
+}

internal/cli/logs/formatter.go

@@ -10,6 +10,7 @@ import (
 	"time"
 
 	"charm.land/lipgloss/v2"
+	"github.com/charmbracelet/colorprofile"
 	"github.com/docker/docker/pkg/stringid"
 	"github.com/psviderski/uncloud/internal/cli/tui"
 	"github.com/psviderski/uncloud/pkg/api"
@@ -24,6 +25,10 @@ type Formatter struct {
 	maxServiceWidth int
 
 	utc bool
+
+	// Cache each stream's terminal profile to avoid detection on every log entry.
+	stdout *colorprofile.Writer
+	stderr *colorprofile.Writer
 }
 
 func NewFormatter(machineNames, serviceNames []string, utc bool) *Formatter {
@@ -50,6 +55,8 @@ func NewFormatter(machineNames, serviceNames []string, utc bool) *Formatter {
 		maxMachineWidth: maxMachineWidth,
 		maxServiceWidth: maxServiceWidth,
 		utc:             utc,
+		stdout:          colorprofile.NewWriter(os.Stdout, os.Environ()),
+		stderr:          colorprofile.NewWriter(os.Stderr, os.Environ()),
 	}
 }
 
@@ -138,9 +145,9 @@ func (f *Formatter) PrintEntry(entry api.ServiceLogEntry) {
 
 	// Print to appropriate stream.
 	if entry.Stream == api.LogStreamStderr {
-		fmt.Fprint(os.Stderr, output.String())
+		fmt.Fprint(f.stderr, output.String())
 	} else {
-		fmt.Print(output.String())
+		fmt.Fprint(f.stdout, output.String())
 	}
 }
 
@@ -148,8 +155,7 @@ func (f *Formatter) PrintEntry(entry api.ServiceLogEntry) {
 func (f *Formatter) printError(entry api.ServiceLogEntry) {
 	if entry.Metadata.ServiceName == "" {
 		msg := fmt.Sprintf("ERROR: %v", entry.Err)
-		style := lipgloss.NewStyle().Bold(true).Foreground(lipgloss.BrightRed)
+		fmt.Fprintln(f.stderr, tui.BoldRed.Render(msg))
-		fmt.Fprintln(os.Stderr, style.Render(msg))
 		return
 	}
 
@@ -171,8 +177,7 @@ func (f *Formatter) printError(entry api.ServiceLogEntry) {
 		msg += fmt.Sprintf(": %v", entry.Err)
 	}
 
-	style := lipgloss.NewStyle().Bold(true).Foreground(lipgloss.BrightYellow)
+	fmt.Fprintln(f.stderr, tui.BoldYellow.Render(msg))
-	fmt.Fprintln(os.Stderr, style.Render(msg))
 }
 
 // palette is available colors for machine/service differentiation.

internal/cli/logs/logs.go

@@ -19,6 +19,18 @@ type Options struct {
 	Machines []string
 }
 
+// TailLines resolves the default tail limit after parsing flags. An empty opts.Tail means the user
+// did not specify a limit, so --since can select all matching logs without overriding an explicit --tail.
+func (opts Options) TailLines() (int, error) {
+	if opts.Tail == "" {
+		if opts.Since != "" {
+			return -1, nil
+		}
+		return 100, nil
+	}
+	return ParseTail(opts.Tail)
+}
+
 func Flags(options *Options) *pflag.FlagSet {
 	set := &pflag.FlagSet{}
 
@@ -29,14 +41,17 @@ func Flags(options *Options) *pflag.FlagSet {
 	set.StringVar(&options.Since, "since", "",
 		"Show logs generated on or after the given timestamp. Accepts relative duration, RFC 3339 date, or Unix timestamp.\n"+
 			"Examples:\n"+
-			"  --since 2m30s                      Relative duration (2 minutes 30 seconds ago)\n"+
+			"  --since 1h45m                      Relative duration (1 hour 45 minutes ago)\n"+
-			"  --since 1h                         Relative duration (1 hour ago)\n"+
+			"                                     Supported units: d (day = 24h), h (hour), m (minute),\n"+
-			"  --since 2025-11-24                 RFC 3339 date only (midnight using local timezone)\n"+
+			"                                                      s (second), ms (millisecond),\n"+
-			"  --since 2024-05-14T22:50:00        RFC 3339 date/time using local timezone\n"+
+			"                                                      us/µs (microsecond), ns (nanosecond)\n"+
+			"  --since 2025-11-24                 RFC 3339 date only (midnight using client local timezone)\n"+
+			"  --since 2024-05-14T22:50:00        RFC 3339 date/time using client local timezone\n"+
 			"  --since 2024-01-31T10:30:00Z       RFC 3339 date/time in UTC\n"+
 			"  --since 1763953966                 Unix timestamp (seconds since January 1, 1970)")
-	set.StringVarP(&options.Tail, "tail", "n", "100",
+	set.StringVarP(&options.Tail, "tail", "n", "",
-		"Show the most recent logs and limit the number of lines shown per replica. Use 'all' to show all logs.")
+		"Show the most recent logs and limit the number of lines shown per replica. Use 'all' to show all logs.\n"+
+			"Defaults to 100, or 'all' when --since is set.")
 	set.StringVar(&options.Until, "until", "",
 		"Show logs generated before the given timestamp. Accepts relative duration, RFC 3339 date, or Unix timestamp.\n"+
 			"See --since for examples.")
@@ -46,7 +61,7 @@ func Flags(options *Options) *pflag.FlagSet {
 	return set
 }
 
-func Tail(tail string) (int, error) {
+func ParseTail(tail string) (int, error) {
 	if tail == "all" {
 		return -1, nil
 	}

internal/cli/logs/logs_test.go

@@ -7,6 +7,42 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func TestOptionsTailLines(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name string
+		args []string
+		want int
+	}{
+		{"default", nil, 100},
+		{"follow", []string{"-f"}, 100},
+		{"since", []string{"--since", "1h"}, -1},
+		{"since and follow", []string{"--since", "1h", "-f"}, -1},
+		{"until only", []string{"--until", "1h"}, 100},
+		{"time range", []string{"--since", "3h", "--until", "1h"}, -1},
+		{"explicit default with since", []string{"--since", "1h", "--tail", "100"}, 100},
+		{"explicit limit with since", []string{"--since", "1h", "-n", "20"}, 20},
+		{"explicit limit before since", []string{"-n", "20", "--since", "1h"}, 20},
+		{"explicit all", []string{"--tail", "all"}, -1},
+		{"explicit zero with since", []string{"--since", "1h", "-n", "0", "-f"}, 0},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var options Options
+			require.NoError(t, Flags(&options).Parse(tt.args))
+			tail, err := options.TailLines()
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, tail)
+		})
+	}
+
+	var options Options
+	require.NoError(t, Flags(&options).Parse([]string{"--since", "1h", "--tail", "invalid"}))
+	_, err := options.TailLines()
+	require.ErrorContains(t, err, "invalid --tail value")
+}
+
 func TestParseServiceArgs(t *testing.T) {
 	t.Parallel()
 

internal/cli/logs/time.go

@@ -0,0 +1,72 @@
+package logs
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// TimeRange resolves log filters that could be RFC 3339, timestamp, or relative duration to UTC timestamps.
+// Dates without a timezone use now's location. Relative durations are computed from now.
+func TimeRange(since, until string, now time.Time) (string, string, error) {
+	var err error
+	since, err = timestamp(since, now)
+	if err != nil {
+		return "", "", fmt.Errorf("invalid --since value: %w", err)
+	}
+	until, err = timestamp(until, now)
+	if err != nil {
+		return "", "", fmt.Errorf("invalid --until value: %w", err)
+	}
+
+	return since, until, nil
+}
+
+func timestamp(value string, now time.Time) (string, error) {
+	if value == "" {
+		return "", nil
+	}
+	// A bare zero is the Unix epoch, matching Docker's log filters.
+	if duration, err := ParseDuration(value); value != "0" && err == nil {
+		return now.Add(-duration).UTC().Format(time.RFC3339Nano), nil
+	}
+
+	// Keep Docker's supported date layouts, but use the location's offset at the requested date.
+	for _, layout := range []string{
+		time.RFC3339Nano,
+		"2006-01-02T15:04Z07:00",
+		"2006-01-02T15Z07:00",
+		"2006-01-02Z07:00",
+		"2006-01-02T15:04:05.999999999",
+		"2006-01-02T15:04",
+		"2006-01-02T15",
+		"2006-01-02",
+	} {
+		if t, err := time.ParseInLocation(layout, value, now.Location()); err == nil {
+			return t.UTC().Format(time.RFC3339Nano), nil
+		}
+	}
+
+	seconds, fraction, hasFraction := strings.Cut(value, ".")
+	sec, err := strconv.ParseInt(seconds, 10, 64)
+	if err != nil {
+		return "", fmt.Errorf("failed to parse '%s' as a time or duration", value)
+	}
+
+	var nsec int64
+	if hasFraction {
+		if len(fraction) == 0 || len(fraction) > 9 || strings.ContainsAny(fraction, "+-") {
+			return "", fmt.Errorf("invalid Unix timestamp fraction in '%s'", value)
+		}
+		nsec, err = strconv.ParseInt(fraction+strings.Repeat("0", 9-len(fraction)), 10, 64)
+		if err != nil {
+			return "", fmt.Errorf("invalid Unix timestamp fraction in '%s': %w", value, err)
+		}
+	}
+	t := time.Unix(sec, nsec).UTC()
+	if t.Year() < 0 || t.Year() > 9999 {
+		return "", fmt.Errorf("Unix timestamp '%s' is outside the RFC 3339 date range", value)
+	}
+	return t.Format(time.RFC3339Nano), nil
+}

internal/cli/logs/time_test.go

@@ -0,0 +1,87 @@
+package logs
+
+import (
+	"testing"
+	"time"
+
+	timetypes "github.com/docker/docker/api/types/time"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestTimeRange(t *testing.T) {
+	t.Parallel()
+
+	location, err := time.LoadLocation("Australia/Sydney")
+	require.NoError(t, err)
+	// October is daylight-saving time, but July timestamps must use the winter offset.
+	// Daylight saving starts on 4 October. Day durations must still mean 24 elapsed hours.
+	now := time.Date(2026, 10, 5, 12, 0, 0, 123456789, location)
+	tests := []struct {
+		input string
+		want  string
+	}{
+		{"", ""},
+		{"2026-07-01", "2026-06-30T14:00:00Z"},
+		{"2026-07-01T10", "2026-07-01T00:00:00Z"},
+		{"2026-07-01T10:30", "2026-07-01T00:30:00Z"},
+		{"2026-07-01T10:30:45.123456789", "2026-07-01T00:30:45.123456789Z"},
+		{"2026-01-01T10:00:00", "2025-12-31T23:00:00Z"},
+		{"2026-07-01T10:30:45Z", "2026-07-01T10:30:45Z"},
+		{"2026-07-01T10:30:45+02:00", "2026-07-01T08:30:45Z"},
+		{"2026-07-01T10:30:45-04:00", "2026-07-01T14:30:45Z"},
+		{"2026-07-01T10+02:00", "2026-07-01T08:00:00Z"},
+		{"2026-07-01T10:30+02:00", "2026-07-01T08:30:00Z"},
+		{"2026-07-01+02:00", "2026-06-30T22:00:00Z"},
+		{"1763953966", "2025-11-24T03:12:46Z"},
+		{"1763953966.000000001", "2025-11-24T03:12:46.000000001Z"},
+		{"0", "1970-01-01T00:00:00Z"},
+		{"2m30s", "2026-10-05T00:57:30.123456789Z"},
+		{"1d", "2026-10-04T01:00:00.123456789Z"},
+		{"2d", "2026-10-03T01:00:00.123456789Z"},
+		{"2d3h", "2026-10-02T22:00:00.123456789Z"},
+		{"1.5d", "2026-10-03T13:00:00.123456789Z"},
+		{"-2d", "2026-10-07T01:00:00.123456789Z"},
+		{"-1h", "2026-10-05T02:00:00.123456789Z"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			since, until, err := TimeRange(tt.input, tt.input, now)
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, since)
+			assert.Equal(t, tt.want, until)
+
+			if since != "" {
+				// The server-side Docker SDK must preserve the cutoff even in another timezone.
+				actual, err := timetypes.GetTimestamp(since, now.In(time.UTC))
+				require.NoError(t, err)
+				expected, err := time.Parse(time.RFC3339Nano, tt.want)
+				require.NoError(t, err)
+				sec, nsec, err := timetypes.ParseTimestamps(actual, 0)
+				require.NoError(t, err)
+				assert.True(t, expected.Equal(time.Unix(sec, nsec)))
+			}
+		})
+	}
+
+	since, until, err := TimeRange("3h", "1h30m", now)
+	require.NoError(t, err)
+	assert.Equal(t, "2026-10-04T22:00:00.123456789Z", since)
+	assert.Equal(t, "2026-10-04T23:30:00.123456789Z", until)
+}
+
+func TestTimeRange_Invalid(t *testing.T) {
+	t.Parallel()
+
+	for _, input := range []string{
+		"invalid", "2026-02-30", "2026-01-01T25:00:00", "1763953966.xyz",
+		"1763953966.1234567890", "253402300800",
+	} {
+		t.Run(input, func(t *testing.T) {
+			_, _, err := TimeRange(input, "", time.Now())
+			require.ErrorContains(t, err, "invalid --since value")
+			_, _, err = TimeRange("", input, time.Now())
+			require.ErrorContains(t, err, "invalid --until value")
+		})
+	}
+}

internal/cli/progress/event.go

@@ -9,7 +9,7 @@ import (
 
 type eventIDKey struct{}
 
-// WithEventID returns a context that overrides the default event ID used by client methods.
+// WithEventID returns a context that overrides the event ID already set in the context.
 func WithEventID(ctx context.Context, eventID string) context.Context {
 	return context.WithValue(ctx, eventIDKey{}, eventID)
 }
@@ -48,7 +48,7 @@ func OldPreDeployHookEventID(serviceName, containerID, machineName string) strin
 		tui.Faint.Render(" on ") + machineName
 }
 
-// ImageEventID returns a progress event ID for image pull operations.
+// ImageEventID returns a progress event ID for image operations on a machine.
 func ImageEventID(image, machineName string) string {
 	return tui.Faint.Render("Image ") + image +
 		tui.Faint.Render(" on ") + machineName

internal/cli/tui/print.go

@@ -3,9 +3,11 @@ package tui
 import (
 	"fmt"
 	"os"
+
+	"charm.land/lipgloss/v2"
 )
 
 func PrintWarning(msg string) {
 	styledMsg := BoldYellow.Render(fmt.Sprintf("WARNING: %s", msg))
-	fmt.Fprintln(os.Stderr, styledMsg)
+	lipgloss.Fprintln(os.Stderr, styledMsg)
 }

internal/journal/journal.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"os/exec"
+	"time"
 
 	"github.com/psviderski/uncloud/pkg/api"
 )
@@ -31,11 +32,11 @@ func logs(ctx context.Context, unit string, opts api.ServiceLogsOptions) (io.Rea
 
 	if opts.Since != "" {
 		args = append(args, "-S")
-		args = append(args, opts.Since)
+		args = append(args, journalTimestamp(opts.Since))
 	}
 	if opts.Until != "" {
 		args = append(args, "-U")
-		args = append(args, opts.Until)
+		args = append(args, journalTimestamp(opts.Until))
 	}
 
 	cmd := commandContext(ctx, journalctl, args...)
@@ -51,6 +52,18 @@ func logs(ctx context.Context, unit string, opts api.ServiceLogsOptions) (io.Rea
 	return p, cmd.Wait, nil
 }
 
+// journalTimestamp formats normalised client timestamps for journalctl versions that do not
+// accept RFC 3339 timezone suffixes. Support for timestamps containing T and Z was added in
+// systemd 255 (6 December 2023). Debian 12 ships systemd 252, so it still needs this conversion.
+// Journald timestamps have microsecond precision.
+func journalTimestamp(value string) string {
+	if t, err := time.Parse(time.RFC3339Nano, value); err == nil {
+		return t.UTC().Format("2006-01-02 15:04:05.999999") + " UTC"
+	}
+	// Preserve raw filters from older clients and SDK callers.
+	return value
+}
+
 // follow synchronously follows the io.Reader, writing each new journal entry to channel.
 // It stops when the reader is exhausted or the context is cancelled.
 func follow(ctx context.Context, reader io.Reader, outCh chan api.LogEntry) {

internal/journal/logs_test.go

@@ -115,6 +115,9 @@ func TestEntry(t *testing.T) {
 }
 
 func TestLogs(t *testing.T) {
+	originalCommandContext := commandContext
+	t.Cleanup(func() { commandContext = originalCommandContext })
+
 	commandContext = func(ctx context.Context, _ string, _ ...string) *exec.Cmd {
 		return exec.CommandContext(ctx, "/usr/bin/tail", "testdata/logs")
 	}
@@ -150,3 +153,38 @@ func TestLogs(t *testing.T) {
 	// Still six because heartbeats are not written here and Tail is ignored as the command is overridden.
 	assert.Equal(t, 6, i)
 }
+
+func TestLogs_TimeFilters(t *testing.T) {
+	originalCommandContext := commandContext
+	t.Cleanup(func() { commandContext = originalCommandContext })
+
+	var args []string
+	commandContext = func(ctx context.Context, command string, commandArgs ...string) *exec.Cmd {
+		assert.Equal(t, "journalctl", command)
+		args = commandArgs
+		return exec.CommandContext(ctx, "/usr/bin/tail", "testdata/logs")
+	}
+
+	ch, err := Logs(context.Background(), "uncloud", api.ServiceLogsOptions{
+		Since: "2026-07-01T10:30:45.123456789+10:00",
+		Until: "2026-07-01T01:30:45Z",
+	})
+	require.NoError(t, err)
+	for range ch {
+	}
+	assert.Equal(t, []string{
+		"-u", "uncloud", "--no-hostname", "-n", "0", "-o", "short-unix",
+		"-S", "2026-07-01 00:30:45.123456 UTC",
+		"-U", "2026-07-01 01:30:45 UTC",
+	}, args)
+
+	// Older clients and SDK callers can still pass journalctl's native filters.
+	ch, err = Logs(context.Background(), "uncloud", api.ServiceLogsOptions{Since: "1h ago", Until: "today"})
+	require.NoError(t, err)
+	for range ch {
+	}
+	assert.Equal(t, []string{
+		"-u", "uncloud", "--no-hostname", "-n", "0", "-o", "short-unix",
+		"-S", "1h ago", "-U", "today",
+	}, args)
+}

internal/machine/caddyconfig/mocks_test.go

@@ -16,10 +16,19 @@ func NewMockCaddyfileValidator(t interface {
 	mock.TestingT
 	Cleanup(func())
 }) *MockCaddyfileValidator {
+	if helper, ok := t.(interface{ Helper() }); ok {
+		helper.Helper()
+	}
+
 	mock := &MockCaddyfileValidator{}
 	mock.Mock.Test(t)
 
-	t.Cleanup(func() { mock.AssertExpectations(t) })
+	t.Cleanup(func() {
+		if helper, ok := t.(interface{ Helper() }); ok {
+			helper.Helper()
+		}
+		mock.AssertExpectations(t)
+	})
 
 	return mock
 }

internal/machine/cluster/cluster.go

@@ -14,6 +14,7 @@ import (
 	"github.com/psviderski/uncloud/internal/machine/network"
 	"github.com/psviderski/uncloud/internal/machine/store"
 	"github.com/psviderski/uncloud/internal/secret"
+	"github.com/psviderski/uncloud/pkg/api"
 	"google.golang.org/grpc/codes"
 	"google.golang.org/grpc/status"
 	"google.golang.org/protobuf/types/known/emptypb"
@@ -87,6 +88,11 @@ func (c *Cluster) AddMachine(ctx context.Context, req *pb.AddMachineRequest) (*p
 func (c *Cluster) AddMachineWithoutReadyCheck(
 	ctx context.Context, req *pb.AddMachineRequest,
 ) (*pb.AddMachineResponse, error) {
+	if req.Name != "" {
+		if err := api.ValidateMachineName(req.Name); err != nil {
+			return nil, status.Error(codes.InvalidArgument, err.Error())
+		}
+	}
 	if req.Network == nil {
 		return nil, status.Error(codes.InvalidArgument, "network not set")
 	}

internal/machine/cluster/cluster_test.go

@@ -0,0 +1,28 @@
+package cluster
+
+import (
+	"context"
+	"testing"
+
+	"github.com/psviderski/uncloud/api/pb"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
+)
+
+func TestAddMachine_InvalidName(t *testing.T) {
+	t.Parallel()
+
+	ready := make(chan struct{})
+	close(ready)
+	// Invalid names must be rejected before any store access.
+	c := NewCluster(nil, nil, nil, ready)
+	for _, name := range []string{"VPS1", "worker.example", "rr", "nearest", "c337f00600de51ef4375c9a9a267dba5"} {
+		t.Run("add", func(t *testing.T) {
+			resp, err := c.AddMachine(context.Background(), &pb.AddMachineRequest{Name: name})
+			require.Nil(t, resp)
+			require.Equal(t, codes.InvalidArgument, status.Code(err))
+			require.ErrorContains(t, err, "invalid machine name")
+		})
+	}
+}

internal/machine/cluster/machine.go

@@ -6,6 +6,7 @@ import (
 	"strings"
 
 	"github.com/psviderski/uncloud/internal/secret"
+	"github.com/psviderski/uncloud/pkg/api"
 )
 
 // NewMachineID generates a new unique machine ID.
@@ -27,7 +28,7 @@ func NewRandomMachineName() (string, error) {
 // a numeric suffix ("-1", "-2", etc.) if needed.
 func DefaultMachineName(hostname string, existing []string) (string, error) {
 	name := machineNameFromHostname(hostname)
-	if name == "" {
+	if api.ValidateMachineName(name) != nil {
 		var err error
 		if name, err = NewRandomMachineName(); err != nil {
 			return "", err
@@ -38,7 +39,9 @@ func DefaultMachineName(hostname string, existing []string) (string, error) {
 		return name, nil
 	}
 	for i := 1; ; i++ {
-		candidate := fmt.Sprintf("%s-%d", name, i)
+		suffix := fmt.Sprintf("-%d", i)
+		base := strings.TrimRight(name[:min(len(name), 63-len(suffix))], "-")
+		candidate := base + suffix
 		if !slices.Contains(existing, candidate) {
 			return candidate, nil
 		}
@@ -55,12 +58,13 @@ func machineNameFromHostname(hostname string) string {
 	var b strings.Builder
 	for _, r := range label {
 		switch {
-		case r >= 'a' && r <= 'z', r >= '0' && r <= '9', r == '-', r == '_':
+		case r >= 'a' && r <= 'z', r >= '0' && r <= '9', r == '-':
 			b.WriteRune(r)
 		default:
 			b.WriteRune('-')
 		}
 	}
 	// Trim leading and trailing hyphens that may result from the sanitisation above.
-	return strings.Trim(b.String(), "-")
+	name := strings.Trim(b.String(), "-")
+	return strings.TrimRight(name[:min(len(name), 63)], "-")
 }

internal/machine/cluster/machine_test.go

@@ -1,8 +1,10 @@
 package cluster
 
 import (
+	"strings"
 	"testing"
 
+	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -18,9 +20,11 @@ func TestMachineNameFromHostname(t *testing.T) {
 		{"simple", "web", "web"},
 		{"fqdn uses first label", "web-1.example.com", "web-1"},
 		{"uppercase lowercased", "Web-Server", "web-server"},
-		{"invalid chars replaced", "host_name@1", "host_name-1"},
+		{"invalid chars replaced", "host_name@1", "host-name-1"},
-		{"trim surrounding hyphens", "-_host_-", "_host_"},
+		{"trim surrounding hyphens", "-_host_-", "host"},
 		{"whitespace trimmed", "  myhost  ", "myhost"},
+		{"long hostname truncated", strings.Repeat("a", 64), strings.Repeat("a", 63)},
+		{"truncation trims trailing hyphen", strings.Repeat("a", 62) + "-b", strings.Repeat("a", 62)},
 		{"empty", "", ""},
 		{"only invalid chars", "@#", ""},
 		{"dot only", ".example.com", ""},
@@ -37,10 +41,12 @@ func TestDefaultMachineName(t *testing.T) {
 	t.Parallel()
 
 	t.Run("falls back to random", func(t *testing.T) {
-		// An empty or fully invalid hostname falls back to a random "machine-xxxx" name.
+		for _, hostname := range []string{"***", "rr", "nearest", strings.Repeat("a", 32)} {
-		got, err := DefaultMachineName("***", nil)
+			got, err := DefaultMachineName(hostname, nil)
 			require.NoError(t, err)
-		assert.Regexp(t, `^machine-[a-zA-Z0-9]{4}$`, got)
+			assert.Regexp(t, `^machine-[a-z0-9]{4}$`, got)
+			require.NoError(t, api.ValidateMachineName(got))
+		}
 	})
 
 	tests := []struct {
@@ -49,10 +55,45 @@ func TestDefaultMachineName(t *testing.T) {
 		existing []string
 		want     string
 	}{
-		{"from hostname", "web-1.example.com", nil, "web-1"},
+		{
-		{"dedup against existing", "web", []string{"web"}, "web-1"},
+			name:     "from hostname",
-		{"dedup multiple", "web", []string{"web", "web-1"}, "web-2"},
+			hostname: "web-1.example.com",
-		{"sanitized", "My_Host", nil, "my_host"},
+			want:     "web-1",
+		},
+		{
+			name:     "dedup against existing",
+			hostname: "web",
+			existing: []string{"web"},
+			want:     "web-1",
+		},
+		{
+			name:     "dedup multiple",
+			hostname: "web",
+			existing: []string{"web", "web-1"},
+			want:     "web-2",
+		},
+		{
+			name:     "sanitized",
+			hostname: "My_Host",
+			want:     "my-host",
+		},
+		{
+			name:     "long hostname",
+			hostname: strings.Repeat("a", 64),
+			want:     strings.Repeat("a", 63),
+		},
+		{
+			name:     "dedup maximum length",
+			hostname: strings.Repeat("a", 63),
+			existing: []string{strings.Repeat("a", 63)},
+			want:     strings.Repeat("a", 61) + "-1",
+		},
+		{
+			name:     "dedup trims trailing hyphen",
+			hostname: strings.Repeat("a", 60) + "-bb",
+			existing: []string{strings.Repeat("a", 60) + "-bb"},
+			want:     strings.Repeat("a", 60) + "-1",
+		},
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
@@ -60,6 +101,7 @@ func TestDefaultMachineName(t *testing.T) {
 			got, err := DefaultMachineName(tt.hostname, tt.existing)
 			require.NoError(t, err)
 			assert.Equal(t, tt.want, got)
+			require.NoError(t, api.ValidateMachineName(got))
 		})
 	}
 }

internal/machine/dns/resolver.go

@@ -10,6 +10,8 @@ import (
 	"sync"
 	"time"
 
+	"github.com/psviderski/uncloud/api/pb"
+	"github.com/psviderski/uncloud/internal/machine/network"
 	"github.com/psviderski/uncloud/internal/machine/store"
 )
 
@@ -19,7 +21,9 @@ type ClusterResolver struct {
 	store *store.Store
 	// serviceIPs maps service names to container IPs.
 	serviceIPs map[string][]netip.Addr
-	// mu protects the serviceIPs map.
+	// machineIPs maps machine IDs and names to their IP.
+	machineIPs map[string]netip.Addr
+	// mu protects the serviceIPs and machineIPs map.
 	mu sync.RWMutex
 	// lastUpdate tracks when records were last updated.
 	lastUpdate time.Time
@@ -46,6 +50,13 @@ func (r *ClusterResolver) Run(ctx context.Context) error {
 	// TODO: implement machine membership check using Corrossion Admin client to filter available containers.
 	r.updateServiceIPs(containers)
 
+	machines, mchanges, err := r.store.SubscribeMachines(ctx)
+	if err != nil {
+		return fmt.Errorf("subscribe to machine changes: %w", err)
+	}
+	r.log.Info("Subscribed to machine changes in the clsuter to keep machine DNS records updated.")
+	r.updateMachineIPs(machines)
+
 	for {
 		select {
 		case _, ok := <-changes:
@@ -59,9 +70,22 @@ func (r *ClusterResolver) Run(ctx context.Context) error {
 				r.log.Error("Failed to list containers.", "err", err)
 				continue
 			}
-
 			// TODO: implement machine membership check using Corrossion Admin client to filter available containers.
 			r.updateServiceIPs(containers)
+
+		case _, ok := <-mchanges:
+			if !ok {
+				return fmt.Errorf("machine subscription failed")
+			}
+			r.log.Debug("Cluster machines changed, updating DNS records.")
+
+			machines, err := r.store.ListMachines(ctx)
+			if err != nil {
+				r.log.Error("Failed to list machines.", "err", err)
+				continue
+			}
+			r.updateMachineIPs(machines)
+
 		case <-ctx.Done():
 			return nil
 		}
@@ -121,19 +145,49 @@ func (r *ClusterResolver) updateServiceIPs(containers []store.ContainerRecord) {
 	r.log.Info("DNS records updated.", "services", len(newServiceIPs)/3, "containers", containersCount)
 }
 
-// Resolve returns IP addresses of the service containers.
+func (r *ClusterResolver) updateMachineIPs(machines []*pb.MachineInfo) {
-func (r *ClusterResolver) Resolve(serviceName string) []netip.Addr {
+	newMachineIPs := make(map[string]netip.Addr, len(machines))
+
+	for _, machine := range machines {
+		subnet, err := machine.Network.Subnet.ToPrefix()
+		if err != nil {
+			continue
+		}
+		addr := network.MachineIP(subnet)
+		newMachineIPs[machine.Name+".m"] = addr
+		newMachineIPs[machine.Id+".m"] = addr
+	}
+	r.mu.Lock()
+	r.machineIPs = newMachineIPs
+	r.mu.Unlock()
+
+	r.log.Info("DNS records updated.", "machines", len(machines))
+}
+
+// Resolve returns IP addresses of the service containers or machines.
+func (r *ClusterResolver) Resolve(name string) []netip.Addr {
 	r.mu.RLock()
 	defer r.mu.RUnlock()
 
-	ips, ok := r.serviceIPs[serviceName]
+	// Return a copy of the IPs slice to prevent modification of the original.
-	if !ok || len(ips) == 0 {
+	ips, ok := r.serviceIPs[name]
-		return nil
+	if ok && len(ips) > 0 {
+		return slices.Clone(ips)
 	}
 
-	// Return a copy of the IPs slice to prevent modification of the original.
+	ip, ok := r.machineIPs[name]
-	ipsCopy := make([]netip.Addr, len(ips))
+	if ok {
-	copy(ipsCopy, ips)
+		return slices.Clone([]netip.Addr{ip})
+	}
 
-	return ipsCopy
+	if name == "m" { // all machines subdomain
+		// collect all unique IP address
+		uniq := map[string]netip.Addr{}
+		for _, addr := range r.machineIPs {
+			uniq[addr.String()] = addr
+		}
+		return slices.Clone(slices.Collect(maps.Values(uniq)))
+	}
+
+	return nil
 }

internal/machine/dns/resolver_test.go

@@ -1,11 +1,13 @@
 package dns
 
 import (
+	"net/netip"
 	"reflect"
 	"testing"
 
 	"github.com/docker/docker/api/types/container"
 	"github.com/docker/docker/api/types/network"
+	"github.com/psviderski/uncloud/api/pb"
 	"github.com/psviderski/uncloud/internal/machine/store"
 	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/stretchr/testify/assert"
@@ -75,3 +77,32 @@ func newRecord(serviceID, serviceName, ip, machineID string) store.ContainerReco
 		MachineID: machineID,
 	}
 }
+
+func TestClusterResolver_UpdateMachineIPs(t *testing.T) {
+	t.Parallel()
+
+	machines := []*pb.MachineInfo{
+		newMachineRecord("x0y0z0", "mach-1", "10.210.0.0/24"),
+		newMachineRecord("x1y1z1", "mach-2", "10.210.1.0/24"),
+		newMachineRecord("x2y2z2", "mach-3", "10.210.2.0/24"),
+	}
+
+	r := NewClusterResolver(nil)
+	r.updateMachineIPs(machines)
+
+	assert.NotEmpty(t, r.Resolve("mach-1.m"))
+	assert.NotEmpty(t, r.Resolve("mach-3.m"))
+	assert.NotEmpty(t, r.Resolve("x0y0z0.m"))
+
+	assert.Equal(t, 3, len(r.Resolve("m")))
+}
+
+func newMachineRecord(machineID, machineName, prefix string) *pb.MachineInfo {
+	return &pb.MachineInfo{
+		Id:   machineID,
+		Name: machineName,
+		Network: &pb.NetworkConfig{
+			Subnet: pb.NewIPPrefix(netip.MustParsePrefix(prefix)),
+		},
+	}
+}

internal/machine/dns/server.go

@@ -293,13 +293,13 @@ func (s *Server) forwardRequest(req *dns.Msg, proto string) (*dns.Msg, error) {
 // handleAQuery processes an A query for the internal domain and returns A records for the requested name.
 // The internal domain suffix is already stripped from the name. An empty list is returned if no records are found.
 func (s *Server) handleAQuery(name string) []dns.RR {
-	serviceName, mode := extractModeFromDomain(trimInternalDomain(name))
+	unname, mode := extractModeFromDomain(trimInternalDomain(name))
-	ips := s.resolver.Resolve(serviceName)
+	ips := s.resolver.Resolve(unname)
 	if len(ips) == 0 {
-		s.log.Debug("Failed to resolve service name.", "service", serviceName)
+		s.log.Debug("Failed to resolve internal name.", "name", name)
 		return nil
 	}
-	s.log.Debug("Resolved service name.", "service", serviceName, "ips", ips)
+	s.log.Debug("Resolved service internal name.", "name", name, "ips", ips)
 
 	if len(ips) > 1 {
 		// Shuffle the IPs to approximate round-robin.

internal/machine/docker/client.go

@@ -167,7 +167,7 @@ func (c *Client) ListContainers(ctx context.Context, opts container.ListOptions)
 	machineContainers := make([]MachineContainers, len(resp.Messages))
 	for i, msg := range resp.Messages {
 		machineContainers[i].Metadata = msg.Metadata
-		if msg.Metadata != nil && msg.Metadata.Error != "" {
+		if msg.Metadata.Err() != nil {
 			continue
 		}
 
@@ -258,7 +258,7 @@ func (c *Client) InspectImage(ctx context.Context, id string) ([]api.MachineImag
 
 	notFoundCount := 0
 	for _, msg := range resp.Messages {
-		if msg.Metadata != nil && msg.Metadata.Status != nil && codes.Code(msg.Metadata.Status.Code) == codes.NotFound {
+		if status.Code(msg.Metadata.Err()) == codes.NotFound {
 			notFoundCount++
 		}
 	}
@@ -269,7 +269,7 @@ func (c *Client) InspectImage(ctx context.Context, id string) ([]api.MachineImag
 	images := make([]api.MachineImage, len(resp.Messages))
 	for i, msg := range resp.Messages {
 		images[i].Metadata = msg.Metadata
-		if msg.Metadata != nil && msg.Metadata.Error != "" {
+		if msg.Metadata.Err() != nil {
 			continue
 		}
 
@@ -312,7 +312,7 @@ func parseRemoteImageMessage(msg *pb.RemoteImage) (api.MachineRemoteImage, error
 	mri := api.MachineRemoteImage{
 		Metadata: msg.Metadata,
 	}
-	if msg.Metadata != nil && msg.Metadata.Error != "" {
+	if msg.Metadata.Err() != nil {
 		return mri, nil
 	}
 
@@ -393,7 +393,7 @@ func (c *Client) ListVolumes(ctx context.Context, opts volume.ListOptions) ([]Ma
 	machineVolumes := make([]MachineVolumes, len(resp.Messages))
 	for i, msg := range resp.Messages {
 		machineVolumes[i].Metadata = msg.Metadata
-		if msg.Metadata != nil && msg.Metadata.Error != "" {
+		if msg.Metadata.Err() != nil {
 			continue
 		}
 
@@ -472,7 +472,7 @@ func (c *Client) ListServiceContainers(
 	machineContainers := make([]MachineServiceContainers, len(resp.Messages))
 	for i, msg := range resp.Messages {
 		machineContainers[i].Metadata = msg.Metadata
-		if msg.Metadata != nil && msg.Metadata.Error != "" {
+		if msg.Metadata.Err() != nil {
 			continue
 		}
 

internal/machine/docker/controller.go

@@ -27,7 +27,7 @@ const (
 type Controller struct {
 	// machineID is the ID of the machine where the managed Docker daemon is running.
 	machineID string
-	client    *client.Client
+	client    client.APIClient
 	service   *Service
 	store     *store.Store
 }
@@ -129,6 +129,11 @@ func (c *Controller) WatchAndSyncContainers(ctx context.Context) error {
 				"container_name", e.Actor.Attributes["name"],
 				"action", e.Action)
 
+			if err := c.syncContainersToStore(ctx); err != nil {
+				return fmt.Errorf("sync containers to cluster store: %w", err)
+			}
+		case <-c.service.ContainersChanged():
+			slog.Debug("Syncing containers to cluster store triggered by a container change in machine database.")
 			if err := c.syncContainersToStore(ctx); err != nil {
 				return fmt.Errorf("sync containers to cluster store: %w", err)
 			}

internal/machine/docker/image.go

@@ -0,0 +1,79 @@
+package docker
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/containerd/errdefs"
+	"github.com/distribution/reference"
+	"github.com/docker/docker/api/types/container"
+	"github.com/docker/docker/api/types/filters"
+	"github.com/docker/docker/api/types/image"
+	"github.com/docker/docker/pkg/stringid"
+	"github.com/psviderski/uncloud/pkg/api"
+)
+
+// RemoveImage removes an image while protecting names used by containers, even when force is requested.
+func (s *Service) RemoveImage(ctx context.Context, id string, opts image.RemoveOptions) ([]image.DeleteResponse, error) {
+	if err := s.checkImageRemovalConflicts(ctx, id); err != nil {
+		return nil, err
+	}
+	// Docker enforces its own conflicts, including protection of a used image's final reference.
+	return s.Client.ImageRemove(ctx, id, opts)
+}
+
+// checkImageRemovalConflicts rejects an image ID used by a running or stopped container, or a name used by a
+// container while it still resolves to that container's image. A digest request also checks the repository tags
+// Docker would remove with it. Unused aliases are allowed.
+func (s *Service) checkImageRemovalConflicts(ctx context.Context, id string) error {
+	img, err := s.Client.ImageInspect(ctx, id)
+	if err != nil {
+		return err
+	}
+	// Match Docker's resolved ID before parsing names, since a short ID also parses as a repository name.
+	isID := strings.HasPrefix(strings.TrimPrefix(img.ID, "sha256:"), strings.TrimPrefix(id, "sha256:"))
+	var named reference.Named
+	if !isID {
+		named, err = reference.ParseDockerRef(id)
+		if err != nil {
+			return errdefs.ErrInvalidArgument.WithMessage(err.Error())
+		}
+	}
+	_, isDigest := named.(reference.Canonical)
+
+	containers, err := s.Client.ContainerList(ctx, container.ListOptions{
+		All: true, Filters: filters.NewArgs(filters.Arg("ancestor", img.ID)),
+	})
+	if err != nil {
+		return fmt.Errorf("list containers using image: %w", err)
+	}
+
+	for _, ctr := range containers {
+		// The ancestor filter can include descendants. A retagged name must also be checked against its current ID.
+		if ctr.ImageID != img.ID {
+			continue
+		}
+
+		usedRef, err := reference.ParseDockerRef(ctr.Image)
+		// Familiar names make "alpine:latest" and "docker.io/library/alpine:latest" identify the same tag.
+		removesUsedName := err == nil && named != nil && reference.FamiliarString(usedRef) == reference.FamiliarString(named)
+		if isDigest && err == nil && usedRef.Name() == named.Name() {
+			// Removing a digest also removes this repository's tags pointing at the same image.
+			removesUsedName = true
+		}
+		if isID || removesUsedName {
+			ctrID, state := stringid.TruncateID(ctr.ID), "stopped"
+			if serviceName := ctr.Labels[api.LabelServiceName]; serviceName != "" {
+				ctrID = serviceName + "/" + ctrID
+			}
+			if ctr.State == container.StateRunning || ctr.State == container.StatePaused {
+				state = container.StateRunning
+			}
+			return errdefs.ErrConflict.WithMessage(
+				fmt.Sprintf("image '%s' is used by container '%s' (%s)", id, ctrID, state))
+		}
+	}
+
+	return nil
+}

internal/machine/docker/image_test.go

@@ -0,0 +1,175 @@
+package docker
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+
+	"github.com/containerd/errdefs"
+	"github.com/docker/docker/api/types/container"
+	"github.com/docker/docker/api/types/image"
+	ocispec "github.com/opencontainers/image-spec/specs-go/v1"
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/mock"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCheckImageRemovalConflicts(t *testing.T) {
+	imageID := "sha256:" + strings.Repeat("a", 64)
+	otherID := "sha256:" + strings.Repeat("b", 64)
+	digestRef := "myapp@" + imageID
+
+	tests := []struct {
+		name             string
+		requested        string
+		configured       string
+		containerImageID string
+		serviceName      string
+		blocked          bool
+	}{
+		{name: "used tag", requested: "myapp:current", configured: "myapp:current", serviceName: "web", blocked: true},
+		{name: "qualified request", requested: "docker.io/library/myapp:current", configured: "myapp:current", blocked: true},
+		{name: "qualified container reference", requested: "myapp:current", configured: "docker.io/library/myapp:current", blocked: true},
+		{name: "implicit latest", requested: "myapp", configured: "myapp:latest", blocked: true},
+		{name: "unused alias", requested: "myapp:old", configured: "myapp:current"},
+		{name: "unused alias of ID-used image", requested: "myapp:old", configured: imageID},
+		{name: "tag may clean up a used digest", requested: "myapp:old", configured: digestRef},
+		{name: "used digest", requested: digestRef, configured: digestRef, blocked: true},
+		{name: "digest also removes used tag", requested: digestRef, configured: "myapp:current", blocked: true},
+		{name: "unused repository digest", requested: digestRef, configured: "other:keep"},
+		{name: "image ID", requested: imageID, configured: "myapp:current", blocked: true},
+		{name: "short image ID", requested: "aaaaaaaa", configured: "myapp:current", blocked: true},
+		{name: "short image ID with algorithm", requested: "sha256:aaaaaaaa", configured: "myapp:current", blocked: true},
+		{name: "retagged name", requested: "myapp:current", configured: "myapp:current", containerImageID: otherID},
+	}
+
+	for _, tt := range tests {
+		for _, state := range []string{container.StateExited, container.StateRunning} {
+			t.Run(fmt.Sprintf("%s/%s", tt.name, state), func(t *testing.T) {
+				containerImageID := tt.containerImageID
+				if containerImageID == "" {
+					containerImageID = imageID
+				}
+
+				cli := NewMockAPIClient(t)
+				cli.EXPECT().ImageInspect(mock.Anything, tt.requested).
+					Return(image.InspectResponse{ID: imageID}, nil).Once()
+				cli.EXPECT().ContainerList(mock.Anything, mock.MatchedBy(func(opts container.ListOptions) bool {
+					return opts.All // Stopped containers must be checked too.
+				})).Return([]container.Summary{{
+					ID:      strings.Repeat("c", 64),
+					Labels:  map[string]string{api.LabelServiceName: tt.serviceName},
+					Names:   []string{"/web"},
+					ImageID: containerImageID,
+					Image:   tt.configured,
+					State:   state,
+				}}, nil).Once()
+
+				service := NewService(cli, nil)
+				err := service.checkImageRemovalConflicts(context.Background(), tt.requested)
+
+				if tt.blocked {
+					require.True(t, errdefs.IsConflict(err), "%v", err)
+					assert.ErrorContains(t, err, tt.requested)
+					if tt.serviceName != "" {
+						assert.ErrorContains(t, err, "container 'web/cccccccccccc'")
+					} else {
+						assert.ErrorContains(t, err, "container 'cccccccccccc'")
+					}
+					wantState := "stopped"
+					if state == container.StateRunning {
+						wantState = state
+					}
+					assert.ErrorContains(t, err, wantState)
+				} else {
+					require.NoError(t, err)
+				}
+			})
+		}
+	}
+}
+
+func TestCheckImageRemovalConflicts_PropagatesLookupErrors(t *testing.T) {
+	imageID := "sha256:" + strings.Repeat("a", 64)
+	t.Run("image lookup fails", func(t *testing.T) {
+		cli := NewMockAPIClient(t)
+		lookupErr := errdefs.ErrNotFound.WithMessage("no such image")
+		cli.EXPECT().ImageInspect(mock.Anything, "myapp:old").Return(image.InspectResponse{}, lookupErr).Once()
+
+		service := NewService(cli, nil)
+		require.ErrorIs(t, service.checkImageRemovalConflicts(context.Background(), "myapp:old"), lookupErr)
+	})
+
+	t.Run("container lookup fails", func(t *testing.T) {
+		cli := NewMockAPIClient(t)
+		lookupErr := errors.New("Docker unavailable")
+		cli.EXPECT().ImageInspect(mock.Anything, "myapp:old").Return(image.InspectResponse{ID: imageID}, nil).Once()
+		cli.EXPECT().ContainerList(mock.Anything, mock.Anything).Return(nil, lookupErr).Once()
+
+		service := NewService(cli, nil)
+		err := service.checkImageRemovalConflicts(context.Background(), "myapp:old")
+
+		require.ErrorIs(t, err, lookupErr)
+		assert.ErrorContains(t, err, "list containers using image")
+	})
+}
+
+func TestServiceRemoveImage_RejectsUsedNameBeforeDeletion(t *testing.T) {
+	imageID := "sha256:" + strings.Repeat("a", 64)
+	cli := NewMockAPIClient(t)
+	cli.EXPECT().ImageInspect(mock.Anything, "myapp:current").Return(image.InspectResponse{ID: imageID}, nil).Once()
+	cli.EXPECT().ContainerList(mock.Anything, mock.Anything).Return([]container.Summary{{
+		ID: "container-id", ImageID: imageID, Image: "myapp:current", State: container.StateRunning,
+	}}, nil).Once()
+
+	service := NewService(cli, nil)
+	// ImageRemove has no expectation: any deletion call would fail this test.
+	results, err := service.RemoveImage(context.Background(), "myapp:current", image.RemoveOptions{Force: true})
+
+	require.True(t, errdefs.IsConflict(err), "%v", err)
+	assert.Nil(t, results)
+}
+
+func TestServiceRemoveImage_DelegatesToDocker(t *testing.T) {
+	imageID := "sha256:" + strings.Repeat("a", 64)
+	tests := []struct {
+		name    string
+		opts    image.RemoveOptions
+		results []image.DeleteResponse
+		err     error
+	}{
+		{
+			name:    "options and results pass through",
+			opts:    image.RemoveOptions{Force: true, PruneChildren: true, Platforms: []ocispec.Platform{{OS: "linux", Architecture: "amd64"}}},
+			results: []image.DeleteResponse{{Deleted: imageID}},
+		},
+		{
+			name: "Docker rejects final reference removal",
+			opts: image.RemoveOptions{PruneChildren: true},
+			err:  errdefs.ErrConflict.WithMessage("container is using its referenced image"),
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cli := NewMockAPIClient(t)
+			cli.EXPECT().ImageInspect(mock.Anything, "myapp:old").Return(image.InspectResponse{ID: imageID}, nil).Once()
+			cli.EXPECT().ContainerList(mock.Anything, mock.Anything).Return([]container.Summary{{
+				ID: "container-id", ImageID: imageID, Image: "myapp@" + imageID, State: container.StateRunning,
+			}}, nil).Once()
+			cli.EXPECT().ImageRemove(mock.Anything, "myapp:old", tt.opts).Return(tt.results, tt.err).Once()
+
+			service := NewService(cli, nil)
+			results, err := service.RemoveImage(context.Background(), "myapp:old", tt.opts)
+			if tt.err != nil {
+				require.ErrorIs(t, err, tt.err)
+			} else {
+				require.NoError(t, err)
+			}
+			assert.Equal(t, tt.results, results)
+		})
+	}
+}

internal/machine/docker/server.go

@@ -56,7 +56,7 @@ var fullDockerIDRegex = regexp.MustCompile(`^[a-f0-9]{64}$`)
 // Server implements the gRPC Docker service that proxies requests to the Docker daemon.
 type Server struct {
 	pb.UnimplementedDockerServer
-	client  *client.Client
+	client  client.APIClient
 	service *Service
 	db      *sqlx.DB
 	// internalDNSIP is a function that returns the IP address of the internal DNS server. It may return an empty
@@ -349,6 +349,41 @@ func (s *Server) InspectImage(ctx context.Context, req *pb.InspectImageRequest)
 	}, nil
 }
 
+// RemoveImage removes an image while protecting names used by containers, even when force is requested.
+func (s *Server) RemoveImage(ctx context.Context, req *pb.RemoveImageRequest) (*pb.RemoveImageResponse, error) {
+	opts := image.RemoveOptions{Force: req.Force, PruneChildren: req.PruneChildren}
+	for _, platform := range req.Platforms {
+		opts.Platforms = append(opts.Platforms, ocispec.Platform{
+			Architecture: platform.Architecture,
+			OS:           platform.Os,
+			OSVersion:    platform.OsVersion,
+			OSFeatures:   platform.OsFeatures,
+			Variant:      platform.Variant,
+		})
+	}
+
+	removals, err := s.service.RemoveImage(ctx, req.Id, opts)
+	if err != nil {
+		switch {
+		case errdefs.IsNotFound(err):
+			return nil, status.Error(codes.NotFound, err.Error())
+		case errdefs.IsConflict(err):
+			return nil, status.Error(codes.FailedPrecondition, err.Error())
+		case errdefs.IsInvalidArgument(err):
+			return nil, status.Error(codes.InvalidArgument, err.Error())
+		default:
+			return nil, status.Error(codes.Internal, err.Error())
+		}
+	}
+
+	result := &pb.MachineImageRemoval{}
+	for _, removal := range removals {
+		result.Results = append(result.Results, &pb.ImageRemoval{Deleted: removal.Deleted, Untagged: removal.Untagged})
+	}
+
+	return &pb.RemoveImageResponse{Messages: []*pb.MachineImageRemoval{result}}, nil
+}
+
 // InspectRemoteImage returns the image metadata for an image in a remote registry using the machine's Docker auth
 // credentials if necessary.
 func (s *Server) InspectRemoteImage(
@@ -559,18 +594,23 @@ func (s *Server) CreateServiceContainer(
 		}
 	}
 
+	// Apply user-defined labels first so that the system labels always take precedence.
+	labels := maps.Clone(spec.Container.Labels)
+	if labels == nil {
+		labels = make(map[string]string)
+	}
+	labels[api.LabelServiceID] = req.ServiceId
+	labels[api.LabelServiceName] = spec.Name
+	labels[api.LabelServiceMode] = spec.Mode
+	labels[api.LabelManaged] = ""
+
 	config := &container.Config{
 		Cmd:        spec.Container.Command,
 		Env:        envVars.ToSlice(),
 		Entrypoint: spec.Container.Entrypoint,
 		Hostname:   containerName,
 		Image:      spec.Container.Image,
-		Labels: map[string]string{
+		Labels:     labels,
-			api.LabelServiceID:   req.ServiceId,
-			api.LabelServiceName: spec.Name,
-			api.LabelServiceMode: spec.Mode,
-			api.LabelManaged:     "",
-		},
 		User:       spec.Container.User,
 		Tty:        spec.Container.Tty,
 		OpenStdin:  spec.Container.OpenStdin,
@@ -694,12 +734,10 @@ func (s *Server) CreateServiceContainer(
 
 	// Strip the container configuration that doesn't make sense for the pre-deploy hook and apply its overrides.
 	if spec.PreDeploy != nil && req.ContainerType == pb.CreateServiceContainerRequest_PRE_DEPLOY {
-		config.Labels = map[string]string{
+		// Keep the user-defined labels but replace the service mode and ports labels with the hook label.
-			api.LabelServiceID:   req.ServiceId,
+		delete(config.Labels, api.LabelServiceMode)
-			api.LabelServiceName: spec.Name,
+		delete(config.Labels, api.LabelServicePorts)
-			api.LabelHook:        api.LabelHookPreDeploy,
+		config.Labels[api.LabelHook] = api.LabelHookPreDeploy
-			api.LabelManaged:     "",
-		}
 		config.Healthcheck = &container.HealthConfig{
 			Test: []string{"NONE"},
 		}
@@ -1144,6 +1182,36 @@ func (s *Server) RemoveServiceContainer(ctx context.Context, req *pb.RemoveConta
 	return resp, nil
 }
 
+// UpdateServiceContainer updates the requested properties of a service container in place without recreating it.
+func (s *Server) UpdateServiceContainer(
+	ctx context.Context, req *pb.UpdateServiceContainerRequest,
+) (*emptypb.Empty, error) {
+	opts := api.UpdateContainerOptions{
+		CPU:               req.Cpu,
+		Memory:            req.Memory,
+		MemoryReservation: req.MemoryReservation,
+	}
+	if req.ServiceLabels != nil {
+		labels := req.ServiceLabels.Labels
+		opts.ServiceLabels = &labels
+	}
+
+	if err := s.service.UpdateServiceContainer(ctx, req.Id, opts); err != nil {
+		switch {
+		case errdefs.IsNotFound(err):
+			return nil, status.Error(codes.NotFound, err.Error())
+		case errdefs.IsInvalidArgument(err):
+			return nil, status.Error(codes.InvalidArgument, err.Error())
+		case errdefs.IsFailedPrecondition(err):
+			return nil, status.Error(codes.FailedPrecondition, err.Error())
+		default:
+			return nil, status.Error(codes.Internal, err.Error())
+		}
+	}
+
+	return &emptypb.Empty{}, nil
+}
+
 // logsHeartbeatInterval is the interval at which heartbeat entries are sent when there are no logs to stream.
 const logsHeartbeatInterval = 200 * time.Millisecond
 

internal/machine/docker/service.go

@@ -15,6 +15,7 @@ import (
 	"time"
 
 	"github.com/cenkalti/backoff/v4"
+	"github.com/containerd/errdefs"
 	"github.com/docker/docker/api/types/container"
 	"github.com/docker/docker/api/types/events"
 	"github.com/docker/docker/api/types/filters"
@@ -31,19 +32,39 @@ import (
 // from the machine database.
 type Service struct {
 	// Client is a Docker client for managing Docker resources.
-	Client *client.Client
+	Client client.APIClient
 	// db is a connection to the machine database.
 	db *sqlx.DB
+	// containersChanged signals that service containers changed in the machine database without a corresponding
+	// Docker event, for example, when a container spec is updated in place.
+	containersChanged chan struct{}
 }
 
 // NewService creates a new Docker service instance.
-func NewService(client *client.Client, db *sqlx.DB) *Service {
+func NewService(client client.APIClient, db *sqlx.DB) *Service {
 	return &Service{
 		Client: client,
 		db:     db,
+		// Buffering as a pending signal is enough for the consumer to pick up all changes made before it reacts.
+		containersChanged: make(chan struct{}, 1),
 	}
 }
 
+// NotifyContainersChanged signals that service containers changed in the machine database. It doesn't block
+// if a signal is already pending.
+func (s *Service) NotifyContainersChanged() {
+	select {
+	case s.containersChanged <- struct{}{}:
+	default:
+	}
+}
+
+// ContainersChanged returns a channel that receives a signal when service containers change in the machine database
+// without a corresponding Docker event.
+func (s *Service) ContainersChanged() <-chan struct{} {
+	return s.containersChanged
+}
+
 // WatchDaemonRestart watches the Docker engine and signals on the returned channel whenever the daemon
 // becomes reachable again after a disconnect, which typically indicates a restart.
 func (s *Service) WatchDaemonRestart(ctx context.Context) <-chan struct{} {
@@ -141,6 +162,90 @@ func (s *Service) InspectServiceContainer(ctx context.Context, nameOrID string)
 	return serviceCtr, nil
 }
 
+// UpdateServiceContainer updates the requested properties of a service container in place without recreating it.
+// It applies the CPU and memory limits to the container and updates the service spec stored for the container.
+func (s *Service) UpdateServiceContainer(ctx context.Context, id string, opts api.UpdateContainerOptions) error {
+	if opts.ServiceLabels != nil {
+		if err := api.ValidateLabelKeys(*opts.ServiceLabels); err != nil {
+			return errdefs.ErrInvalidArgument.WithMessage(fmt.Sprintf("invalid service labels: %v", err))
+		}
+	}
+
+	ctr, err := s.InspectServiceContainer(ctx, id)
+	if err != nil {
+		return err
+	}
+	spec := ctr.ServiceSpec
+
+	resources := spec.Container.Resources
+	if opts.CPU != nil {
+		resources.CPU = *opts.CPU
+	}
+	if opts.Memory != nil {
+		resources.Memory = *opts.Memory
+	}
+	if opts.MemoryReservation != nil {
+		resources.MemoryReservation = *opts.MemoryReservation
+	}
+	if resources.CPU < 0 || resources.Memory < 0 || resources.MemoryReservation < 0 {
+		return errdefs.ErrInvalidArgument.WithMessage("CPU and memory limits must not be negative")
+	}
+	// Compare with the limits Docker applies to the container rather than the stored spec. They may differ if
+	// a previous update applied the limits but failed to store the spec, or the container was updated outside
+	// of Uncloud.
+	applied := ctr.AppliedResources()
+	if err = applied.ValidateUpdateInPlace(resources); err != nil {
+		return errdefs.ErrFailedPrecondition.WithMessage(
+			fmt.Sprintf("container '%s' can't be updated without recreating it: %v", ctr.ID, err))
+	}
+
+	// Only CPU and memory limits can differ after the in-place update check above.
+	if !applied.Equals(resources) {
+		updateResources := container.Resources{
+			NanoCPUs:          resources.CPU,
+			Memory:            resources.Memory,
+			MemoryReservation: resources.MemoryReservation,
+		}
+		// Docker defaults the swap limit to twice the memory limit when creating a container and rejects raising
+		// the memory limit above the current swap limit, so update both to keep the same semantics.
+		if resources.Memory > 0 {
+			updateResources.MemorySwap = 2 * resources.Memory
+		}
+		if _, err = s.Client.ContainerUpdate(ctx, ctr.ID, container.UpdateConfig{
+			Resources: updateResources,
+		}); err != nil {
+			return fmt.Errorf("update container resources: %w", err)
+		}
+	}
+	spec.Container.Resources = resources
+
+	if opts.ServiceLabels != nil {
+		spec.Labels = nil
+		if len(*opts.ServiceLabels) > 0 {
+			spec.Labels = *opts.ServiceLabels
+		}
+	}
+
+	specBytes, err := json.Marshal(spec)
+	if err != nil {
+		return fmt.Errorf("marshal service spec: %w", err)
+	}
+	res, err := s.db.ExecContext(ctx,
+		`UPDATE containers SET service_spec = $1, updated_at = datetime('subsecond') WHERE id = $2`,
+		string(specBytes), ctr.ID)
+	if err != nil {
+		return fmt.Errorf("update container in database: %w", err)
+	}
+	if n, err := res.RowsAffected(); err == nil && n == 0 {
+		return errdefs.ErrNotFound.WithMessage(fmt.Sprintf("container '%s' not found in machine database", ctr.ID))
+	}
+
+	// Propagate the updated spec to the cluster store without waiting for the next regular sync.
+	s.NotifyContainersChanged()
+
+	return nil
+}
+
 // ListServiceContainersResult holds the result of listing service containers, split into regular
 // service containers and one-off hook containers.
 type ListServiceContainersResult struct {

internal/machine/machine.go

@@ -836,6 +836,11 @@ func (m *Machine) InitCluster(ctx context.Context, req *pb.InitClusterRequest) (
 	if m.Initialised() {
 		return nil, status.Error(codes.FailedPrecondition, "machine is already configured as a cluster member")
 	}
+	if req.MachineName != "" {
+		if err := api.ValidateMachineName(req.MachineName); err != nil {
+			return nil, status.Error(codes.InvalidArgument, err.Error())
+		}
+	}
 
 	clusterNetwork, err := req.Network.ToPrefix()
 	if err != nil {
@@ -1220,8 +1225,8 @@ func (m *Machine) applyMachineUpdate(ctx context.Context, req *pb.UpdateMachineR
 	defer m.state.mu.Unlock()
 
 	if req.Name != nil {
-		if *req.Name == "" {
+		if err := api.ValidateMachineName(*req.Name); err != nil {
-			return status.Error(codes.InvalidArgument, "machine name cannot be empty")
+			return status.Error(codes.InvalidArgument, err.Error())
 		}
 		// Check for duplicate names across the cluster, excluding this machine.
 		if *req.Name != m.state.Name {

internal/machine/machine_test.go

@@ -0,0 +1,38 @@
+package machine
+
+import (
+	"context"
+	"testing"
+
+	"github.com/psviderski/uncloud/api/pb"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
+)
+
+func TestInitCluster_InvalidMachineName(t *testing.T) {
+	t.Parallel()
+
+	for _, name := range []string{"VPS1", "rr", "nearest", "c337f00600de51ef4375c9a9a267dba5"} {
+		// Invalid names must be rejected before initializing cluster state.
+		m := &Machine{state: &State{}}
+
+		resp, err := m.InitCluster(context.Background(), &pb.InitClusterRequest{MachineName: name})
+		require.Nil(t, resp)
+		require.Equal(t, codes.InvalidArgument, status.Code(err))
+		require.ErrorContains(t, err, "invalid machine name")
+	}
+}
+
+func TestUpdateMachine_InvalidName(t *testing.T) {
+	t.Parallel()
+
+	for _, name := range []string{"", "VPS1", "worker.example", "rr", "nearest", "c337f00600de51ef4375c9a9a267dba5"} {
+		m := &Machine{state: &State{ID: "machine-id", Name: "worker"}}
+		resp, err := m.UpdateMachine(context.Background(), &pb.UpdateMachineRequest{Name: &name})
+		require.Nil(t, resp)
+		require.Equal(t, codes.InvalidArgument, status.Code(err))
+		require.ErrorContains(t, err, "invalid machine name")
+		require.Equal(t, "worker", m.state.Name)
+	}
+}

internal/machine/store/container.go

@@ -23,6 +23,10 @@ const (
 	// SyncStatusOutdated indicates that a container record may be outdated, for example, due to being unable
 	// to retrieve the container's state from the Docker daemon or when the machine is being stopped or restarted.
 	SyncStatusOutdated = "outdated"
+
+	// containerChangesDebounceInterval defines how long to wait before notifying subscribers about container changes.
+	// Multiple changes within this window are grouped into a single notification to prevent system overload.
+	containerChangesDebounceInterval = 250 * time.Millisecond
 )
 
 type ContainerRecord struct {
@@ -92,6 +96,9 @@ func normaliseContainerForStore(ctr *api.ServiceContainer) {
 	// Remove the environment variables to avoid leaking secrets.
 	ctr.Config.Env = nil
 	ctr.ServiceSpec.Container.Env = nil
+	if ctr.ServiceSpec.PreDeploy != nil {
+		ctr.ServiceSpec.PreDeploy.Env = nil
+	}
 
 	// Docker returns Mounts in a non-deterministic order so sort them.
 	slices.SortFunc(ctr.Mounts, func(a, b container.MountPoint) int {
@@ -267,6 +274,13 @@ func (s *Store) SubscribeContainers(ctx context.Context) ([]ContainerRecord, <-c
 	changes := make(chan struct{})
 	go func() {
 		defer close(changes)
+		// Coalesce bursts of rapid-fire row-level change events (e.g. from frequent health-check updates
+		// across many services) into a single signal per debounce window, instead of forwarding one signal
+		// per event. Every subscriber of this channel (e.g. the Caddy and DNS reconcilers) otherwise reruns
+		// its full reconciliation on every single event, which can burn significant CPU across the cluster
+		// when there's a lot of container churn.
+		var debouncer *time.Timer
+		var debouncerCh <-chan time.Time
 		for {
 			select {
 			case <-ctx.Done():
@@ -279,8 +293,21 @@ func (s *Store) SubscribeContainers(ctx context.Context) ([]ContainerRecord, <-c
 					}
 					return
 				}
-				// Just signal that there is a change in the containers list.
+				if debouncerCh == nil {
-				changes <- struct{}{}
+					if debouncer == nil {
+						debouncer = time.NewTimer(containerChangesDebounceInterval)
+					} else {
+						debouncer.Reset(containerChangesDebounceInterval)
+					}
+					debouncerCh = debouncer.C
+				}
+			case <-debouncerCh:
+				select {
+				case changes <- struct{}{}:
+				case <-ctx.Done():
+					return
+				}
+				debouncerCh = nil
 			}
 		}
 	}()

internal/machine/store/container_test.go

@@ -0,0 +1,67 @@
+package store
+
+import (
+	"testing"
+
+	"github.com/docker/docker/api/types/container"
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestNormaliseContainerForStore_StripsEnv pins the behaviour that all fields carrying environment variables
+// are cleared before a container record is written to the replicated store, regardless of which part of the
+// spec they live in. See https://github.com/psviderski/uncloud/issues/422.
+func TestNormaliseContainerForStore_StripsEnv(t *testing.T) {
+	t.Parallel()
+
+	newContainer := func(preDeploy *api.PreDeployHook) api.ServiceContainer {
+		return api.ServiceContainer{
+			Container: api.Container{
+				InspectResponse: container.InspectResponse{
+					Config: &container.Config{
+						Env: []string{"CONTROL_TOKEN=control-value-aaa"},
+					},
+				},
+			},
+			ServiceSpec: api.ServiceSpec{
+				Container: api.ContainerSpec{
+					Env: api.EnvVars{"CONTROL_TOKEN": "control-value-aaa"},
+				},
+				PreDeploy: preDeploy,
+			},
+		}
+	}
+
+	t.Run("service with pre-deploy hook", func(t *testing.T) {
+		t.Parallel()
+
+		ctr := newContainer(&api.PreDeployHook{
+			Command: []string{"/bin/true"},
+			Env:     api.EnvVars{"HOOK_TOKEN": "hook-value-bbb"},
+		})
+
+		normaliseContainerForStore(&ctr)
+
+		assert.Nil(t, ctr.Config.Env, "Config.Env should be stripped")
+		assert.Nil(t, ctr.ServiceSpec.Container.Env, "ServiceSpec.Container.Env should be stripped")
+		require.NotNil(t, ctr.ServiceSpec.PreDeploy, "PreDeploy itself should be left in place")
+		assert.Nil(t, ctr.ServiceSpec.PreDeploy.Env, "ServiceSpec.PreDeploy.Env should be stripped")
+		// Non-env fields of the hook must survive the normalisation.
+		assert.Equal(t, []string{"/bin/true"}, ctr.ServiceSpec.PreDeploy.Command)
+	})
+
+	t.Run("service without pre-deploy hook", func(t *testing.T) {
+		t.Parallel()
+
+		ctr := newContainer(nil)
+
+		require.NotPanics(t, func() {
+			normaliseContainerForStore(&ctr)
+		})
+
+		assert.Nil(t, ctr.Config.Env)
+		assert.Nil(t, ctr.ServiceSpec.Container.Env)
+		assert.Nil(t, ctr.ServiceSpec.PreDeploy)
+	})
+}

mise.lock

@@ -1,40 +1,58 @@
 # @generated - this file is auto-generated by `mise lock` https://mise.jdx.dev/dev-tools/mise-lock.html
 
+lockfile_version = 3
+
 [[tools."aqua:vektra/mockery"]]
-version = "3.5.3"
+version = "3.8.1"
 backend = "aqua:vektra/mockery"
+specifiers = ["3.8.1"]
 
 [tools."aqua:vektra/mockery"."platforms.linux-arm64"]
-checksum = "sha256:ec02ca72b7ad1ff5a1de5503a27d3b1b3f230d722eb35c5b92e43e295a8e1290"
+checksum = "sha256:d1f481f3a6e9868dd3ad65da265238ee109ab3a5dd7508958d5811953696f452"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Linux_arm64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Linux_arm64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915757"
 
 [tools."aqua:vektra/mockery"."platforms.linux-arm64-musl"]
-checksum = "sha256:ec02ca72b7ad1ff5a1de5503a27d3b1b3f230d722eb35c5b92e43e295a8e1290"
+checksum = "sha256:d1f481f3a6e9868dd3ad65da265238ee109ab3a5dd7508958d5811953696f452"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Linux_arm64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Linux_arm64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915757"
 
 [tools."aqua:vektra/mockery"."platforms.linux-x64"]
-checksum = "sha256:ebce416b0175338525246c376885a1579ca6cd4d4015140ba0c70e6b5339a39c"
+checksum = "sha256:953c7d38b58dd6398b1c595f27caf8bbf27e9b733b49c404d2e124e86d2ae72a"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Linux_x86_64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Linux_x86_64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915756"
 
 [tools."aqua:vektra/mockery"."platforms.linux-x64-musl"]
-checksum = "sha256:ebce416b0175338525246c376885a1579ca6cd4d4015140ba0c70e6b5339a39c"
+checksum = "sha256:953c7d38b58dd6398b1c595f27caf8bbf27e9b733b49c404d2e124e86d2ae72a"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Linux_x86_64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Linux_x86_64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915756"
 
 [tools."aqua:vektra/mockery"."platforms.macos-arm64"]
-checksum = "sha256:a3a94b14c7414e148f2252199ffc4a0108d311358f3d336cbe05bb73cb203704"
+checksum = "sha256:303a794484243906b37da0c37056f974bf94278c2e10d7b1372e6bd53a2d5034"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Darwin_arm64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Darwin_arm64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915806"
 
 [tools."aqua:vektra/mockery"."platforms.macos-x64"]
-checksum = "sha256:8bd2b3e4722e476400395920fb5bbd68be4ea22ec12cddf1e424c40610fda863"
+checksum = "sha256:6882331344e946268b304dce042b01371a0213a896536413be8abe49d532ca6f"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Darwin_x86_64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Darwin_x86_64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915746"
 
 [tools."aqua:vektra/mockery"."platforms.windows-x64"]
-checksum = "sha256:3b4264066db615a56a8b59abda992f6db82b8cd7666937ed9dd8f4a4bfc424a0"
+checksum = "sha256:7a7458756e9d7bce5f9936b7a63394000f3affa751b9740e65350c4bdcb627dc"
-url = "https://github.com/vektra/mockery/releases/download/v3.5.3/mockery_3.5.3_Windows_x86_64.tar.gz"
+checksum_source = "mise-versions"
+url = "https://github.com/vektra/mockery/releases/download/v3.8.1/mockery_3.8.1_Windows_x86_64.tar.gz"
+url_api = "https://api.github.com/repos/vektra/mockery/releases/assets/632915759"
 
 [[tools.go]]
 version = "1.26.1"
 backend = "core:go"
+specifiers = ["1.26.1"]
 
 [tools.go."platforms.linux-arm64"]
 checksum = "sha256:a290581cfe4fe28ddd737dde3095f3dbeb7f2e4065cab4eae44dfc53b760c2f7"
@@ -67,95 +85,118 @@ url = "https://dl.google.com/go/go1.26.1.windows-amd64.zip"
 [[tools.golangci-lint]]
 version = "2.11.2"
 backend = "aqua:golangci/golangci-lint"
+specifiers = ["2.11.2"]
 
 [tools.golangci-lint."platforms.linux-arm64"]
 checksum = "sha256:c3c83c7ed003e881991af21361c487151abd37969927ed9062cd0b48f3cd1d8d"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-linux-arm64.tar.gz"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099658"
+provenance = "github-attestations"
 
 [tools.golangci-lint."platforms.linux-arm64-musl"]
 checksum = "sha256:c3c83c7ed003e881991af21361c487151abd37969927ed9062cd0b48f3cd1d8d"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-linux-arm64.tar.gz"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099658"
+provenance = "github-attestations"
 
 [tools.golangci-lint."platforms.linux-x64"]
 checksum = "sha256:852741ea769f4543ab4b6cf86b313cb7173f2e8d2d39f00eb0cd5ef5d4d5eda9"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-linux-amd64.tar.gz"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099674"
 provenance = "github-attestations"
 
 [tools.golangci-lint."platforms.linux-x64-musl"]
 checksum = "sha256:852741ea769f4543ab4b6cf86b313cb7173f2e8d2d39f00eb0cd5ef5d4d5eda9"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-linux-amd64.tar.gz"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099674"
+provenance = "github-attestations"
 
 [tools.golangci-lint."platforms.macos-arm64"]
 checksum = "sha256:11e9549b955bae4bfb71aa3b5c2762e446e2d3f975fe963d6d767b8f745d9344"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-darwin-arm64.tar.gz"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099671"
 provenance = "github-attestations"
 
 [tools.golangci-lint."platforms.macos-x64"]
 checksum = "sha256:1dbbbe745eadba9cf012d49bd0b8c9290cbcb50dbf2cf9611f75a715b46ae3af"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-darwin-amd64.tar.gz"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099653"
+provenance = "github-attestations"
 
 [tools.golangci-lint."platforms.windows-x64"]
 checksum = "sha256:60e50178253d276d1c808ef00845eddcabca63b7d69d86a6dbaa41e88bc843f3"
 url = "https://github.com/golangci/golangci-lint/releases/download/v2.11.2/golangci-lint-2.11.2-windows-amd64.zip"
+url_api = "https://api.github.com/repos/golangci/golangci-lint/releases/assets/369099651"
+provenance = "github-attestations"
 
 [[tools.goreleaser]]
 version = "2.15.2"
 backend = "aqua:goreleaser/goreleaser"
+specifiers = ["2.15.2"]
 
 [tools.goreleaser."platforms.linux-arm64"]
 checksum = "sha256:5db66761a98f6693161e49e1a95d28d2673a892ba60cb4a5e16736cafd41c4c9"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Linux_arm64.tar.gz"
-provenance = "cosign"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702105"
+provenance = "github-attestations"
 
 [tools.goreleaser."platforms.linux-arm64-musl"]
 checksum = "sha256:5db66761a98f6693161e49e1a95d28d2673a892ba60cb4a5e16736cafd41c4c9"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Linux_arm64.tar.gz"
-provenance = "cosign"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702105"
+provenance = "github-attestations"
 
 [tools.goreleaser."platforms.linux-x64"]
 checksum = "sha256:0ebdbf0353aba566b969dde746cc4e4806f96c27aa2f3971b229a9df7611fedc"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Linux_x86_64.tar.gz"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702106"
 provenance = "github-attestations"
 
 [tools.goreleaser."platforms.linux-x64-musl"]
 checksum = "sha256:0ebdbf0353aba566b969dde746cc4e4806f96c27aa2f3971b229a9df7611fedc"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Linux_x86_64.tar.gz"
-provenance = "cosign"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702106"
+provenance = "github-attestations"
 
 [tools.goreleaser."platforms.macos-arm64"]
 checksum = "sha256:0e6bd67688ac949780bf1166813a91f89856898ef4c40d7d46c2c74ebaa4b9ee"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Darwin_all.tar.gz"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702148"
 provenance = "github-attestations"
 
 [tools.goreleaser."platforms.macos-x64"]
 checksum = "sha256:0e6bd67688ac949780bf1166813a91f89856898ef4c40d7d46c2c74ebaa4b9ee"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Darwin_all.tar.gz"
-provenance = "cosign"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702148"
+provenance = "github-attestations"
 
 [tools.goreleaser."platforms.windows-x64"]
 checksum = "sha256:7459832946dbe122c144f8d7f87484d8572ca005b779310aa6bb03346e8de17a"
 url = "https://github.com/goreleaser/goreleaser/releases/download/v2.15.2/goreleaser_Windows_x86_64.zip"
-provenance = "cosign"
+url_api = "https://api.github.com/repos/goreleaser/goreleaser/releases/assets/385702135"
+provenance = "github-attestations"
 
 [[tools.node]]
 version = "22.20.0"
 backend = "core:node"
+specifiers = ["22.20.0"]
 
 [tools.node."platforms.linux-arm64"]
 checksum = "sha256:4181609e03dcb9880e7e5bf956061ecc0503c77a480c6631d868cb1f65a2c7dd"
 url = "https://nodejs.org/dist/v22.20.0/node-v22.20.0-linux-arm64.tar.gz"
 
 [tools.node."platforms.linux-arm64-musl"]
-checksum = "sha256:4181609e03dcb9880e7e5bf956061ecc0503c77a480c6631d868cb1f65a2c7dd"
+checksum = "sha256:79a087c1e23fc97d173ad6676df5b1a4e1a02daf1acca05ee25662747c5d492c"
-url = "https://nodejs.org/dist/v22.20.0/node-v22.20.0-linux-arm64.tar.gz"
+url = "https://nodejs.org/dist/v22.20.0/node-v22.20.0.tar.gz"
+install = "source"
 
 [tools.node."platforms.linux-x64"]
 checksum = "sha256:eeaccb0378b79406f2208e8b37a62479c70595e20be6b659125eb77dd1ab2a29"
 url = "https://nodejs.org/dist/v22.20.0/node-v22.20.0-linux-x64.tar.gz"
 
 [tools.node."platforms.linux-x64-musl"]
-checksum = "sha256:eeaccb0378b79406f2208e8b37a62479c70595e20be6b659125eb77dd1ab2a29"
+checksum = "sha256:461aafed5eb607357479c6de5df978515d74daf7e0169e972e61c8390c035f1a"
-url = "https://nodejs.org/dist/v22.20.0/node-v22.20.0-linux-x64.tar.gz"
+url = "https://unofficial-builds.nodejs.org/download/release/v22.20.0/node-v22.20.0-linux-x64-musl.tar.gz"
 
 [tools.node."platforms.macos-arm64"]
 checksum = "sha256:cc04a76a09f79290194c0646f48fec40354d88969bec467789a5d55dd097f949"
@@ -172,41 +213,119 @@ url = "https://nodejs.org/dist/v22.20.0/node-v22.20.0-win-x64.zip"
 [[tools.protoc]]
 version = "27.3"
 backend = "aqua:protocolbuffers/protobuf/protoc"
+specifiers = ["27.3"]
+
+[tools.protoc."platforms.linux-arm64"]
+checksum = "sha256:bdad36f3ad7472281d90568c4956ea2e203c216e0de005c6bd486f1920f2751c"
+url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-linux-aarch_64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006374"
+
+[tools.protoc."platforms.linux-arm64-musl"]
+checksum = "sha256:bdad36f3ad7472281d90568c4956ea2e203c216e0de005c6bd486f1920f2751c"
+url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-linux-aarch_64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006374"
 
 [tools.protoc."platforms.linux-x64"]
-checksum = "blake3:8c2938accf7061164ea3968529a47ced1db4824ca553857f6903c68ae4347739"
+checksum = "sha256:6dab2adab83f915126cab53540d48957c40e9e9023969c3e84d44bfb936c7741"
 url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-linux-x86_64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006380"
+
+[tools.protoc."platforms.linux-x64-musl"]
+checksum = "sha256:6dab2adab83f915126cab53540d48957c40e9e9023969c3e84d44bfb936c7741"
+url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-linux-x86_64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006380"
+
+[tools.protoc."platforms.macos-arm64"]
+checksum = "sha256:b22116bd97cdbd7ea25346abe635a9df268515fe5ef5afa93cd9a68fc2513f84"
+url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-osx-aarch_64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006383"
+
+[tools.protoc."platforms.macos-x64"]
+checksum = "sha256:ce282648fed0e7fbd6237d606dc9ec168dd2c1863889b04efa0b19c47da65d1b"
+url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-osx-x86_64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006381"
+
+[tools.protoc."platforms.windows-x64"]
+checksum = "sha256:bdccd341a0c25ff4f65d8530f170c78ce5681e42920aa5d892f27ac06cc3bba0"
+url = "https://github.com/protocolbuffers/protobuf/releases/download/v27.3/protoc-27.3-win64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf/releases/assets/183006385"
 
 [[tools.protoc-gen-go]]
 version = "1.34.2"
 backend = "aqua:protocolbuffers/protobuf-go/protoc-gen-go"
+specifiers = ["1.34.2"]
 
 [tools.protoc-gen-go."platforms.linux-arm64"]
+checksum = "sha256:63d400167e75ab9f6690688f6fdc6a9455aa20bc1faa71e32149dbd322f7f198"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.linux.arm64.tar.gz"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109690"
 
 [tools.protoc-gen-go."platforms.linux-arm64-musl"]
+checksum = "sha256:63d400167e75ab9f6690688f6fdc6a9455aa20bc1faa71e32149dbd322f7f198"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.linux.arm64.tar.gz"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109690"
 
 [tools.protoc-gen-go."platforms.linux-x64"]
-checksum = "blake3:f765e393e0f0b25070c13848d88f5e6303e66090132afa1b753e981507a4174e"
+checksum = "sha256:b87bc134dee55576a842141bf0ed27761c635d746780fce5dee038c6dd16554f"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.linux.amd64.tar.gz"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109692"
 
 [tools.protoc-gen-go."platforms.linux-x64-musl"]
+checksum = "sha256:b87bc134dee55576a842141bf0ed27761c635d746780fce5dee038c6dd16554f"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.linux.amd64.tar.gz"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109692"
 
 [tools.protoc-gen-go."platforms.macos-arm64"]
+checksum = "sha256:17aca7f948dbb624049030cf841e35895cf34183ba006e721247fdeb95ff2780"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.darwin.arm64.tar.gz"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109697"
 
 [tools.protoc-gen-go."platforms.macos-x64"]
+checksum = "sha256:9b48d8f90add02e8e94e14962fed74e7ce2b2d6bda4dd42f1f4fbccf0f766f1a"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.darwin.amd64.tar.gz"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109698"
 
 [tools.protoc-gen-go."platforms.windows-x64"]
+checksum = "sha256:abafd39612177dd4e9a65207cadd5374a9352d8611e8e040f8462fcfa3010daf"
 url = "https://github.com/protocolbuffers/protobuf-go/releases/download/v1.34.2/protoc-gen-go.v1.34.2.windows.amd64.zip"
+url_api = "https://api.github.com/repos/protocolbuffers/protobuf-go/releases/assets/173109685"
 
 [[tools.protoc-gen-go-grpc]]
 version = "1.5.1"
 backend = "aqua:grpc/grpc-go/protoc-gen-go-grpc"
+specifiers = ["1.5.1"]
+
+[tools.protoc-gen-go-grpc."platforms.linux-arm64"]
+checksum = "sha256:370a7cfa5786e61e6fdc5aa8f689e35c2d27d036a49caac7aa25f962c89a01e8"
+url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.linux.arm64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572238"
+
+[tools.protoc-gen-go-grpc."platforms.linux-arm64-musl"]
+checksum = "sha256:370a7cfa5786e61e6fdc5aa8f689e35c2d27d036a49caac7aa25f962c89a01e8"
+url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.linux.arm64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572238"
 
 [tools.protoc-gen-go-grpc."platforms.linux-x64"]
-checksum = "blake3:197739132f0408f017d0c83773ac1ace85c0572473fdd658ec6e4ccbf20e0552"
+checksum = "sha256:a6cac4ea731e54aea304ad44d704a69d1cdc82997084b33637e21a89dc9229d6"
 url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.linux.amd64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572229"
+
+[tools.protoc-gen-go-grpc."platforms.linux-x64-musl"]
+checksum = "sha256:a6cac4ea731e54aea304ad44d704a69d1cdc82997084b33637e21a89dc9229d6"
+url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.linux.amd64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572229"
+
+[tools.protoc-gen-go-grpc."platforms.macos-arm64"]
+checksum = "sha256:d6083feb51dcfe59f26793e99ef01ef5eac68b64294ec2546711f614ac5878f3"
+url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.darwin.arm64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572243"
+
+[tools.protoc-gen-go-grpc."platforms.macos-x64"]
+checksum = "sha256:c12c259d59c8b451f24b6abdad6d055fd0884c5268b7247fd7e5ec9d2432d1f1"
+url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.darwin.amd64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572208"
+
+[tools.protoc-gen-go-grpc."platforms.windows-x64"]
+checksum = "sha256:c2ad1460c9e34b5b37759826da61220be6ee0a8a11431723d205512a69319991"
+url = "https://github.com/grpc/grpc-go/releases/download/cmd/protoc-gen-go-grpc/v1.5.1/protoc-gen-go-grpc.v1.5.1.windows.amd64.tar.gz"
+url_api = "https://api.github.com/repos/grpc/grpc-go/releases/assets/182572216"

mise.toml

@@ -1,8 +1,9 @@
 [settings]
 experimental = true
+lockfile = true
 
 [tools]
-"aqua:vektra/mockery" = "3.5.3"
+"aqua:vektra/mockery" = "3.8.1"
 go = "1.26.1"
 golangci-lint = "2.11.2"
 goreleaser = "2.15.2"
@@ -36,9 +37,37 @@ GOOS=linux GOARCH=arm64 go build -ldflags "-X github.com/psviderski/uncloud/inte
 """
 outputs = ["uncloudd-linux-arm64"]
 
-[tasks."mise:lock"]
+[tasks."docs:cli"]
-description = "Update mise.lock with hashes for all platforms"
+description = "Regenerate the CLI reference docs"
-run = "mise lock --platform linux-x64,macos-arm64"
+run = "go run ./cmd/uc docs"
+
+[tasks."docs:dev"]
+description = "Run the documentation website locally"
+dir = "website"
+run = "npm start --"
+
+[tasks.fmt]
+description = "Format Go code and imports"
+alias = "format"
+env.GOOS = "linux"
+run = "golangci-lint fmt"
+
+[tasks.lint]
+description = "Check Go code and formatting"
+# Explicitly set OS to Linux to not skip *_linux.go files when running on macOS.
+# Uncloud daemon won't likely support OS other than Linux anytime soon, so for now we can rely on that.
+env.GOOS = "linux"
+run = "golangci-lint run"
+
+[tasks."lint:fix"]
+description = "Format Go code and apply available lint fixes"
+alias = "lint-and-fix"
+env.GOOS = "linux"
+run = "golangci-lint run --fix"
+
+[tasks.mocks]
+description = "Regenerate Go test mocks"
+run = "mockery"
 
 [tasks.proto]
 description = "Regenerate gRPC API code from .proto files"
@@ -47,12 +76,20 @@ protoc --go_out=. --go_opt=paths=source_relative --go-grpc_out=. --go-grpc_opt=p
   --proto_path=. --proto_path=api/pb/vendor api/pb/*.proto pkg/distlock/grpc/*.proto
 """
 
+[tasks.release]
+description = "Create a draft GitHub release with GoReleaser"
+# The 1Password CLI resolves op:// secret references in the environment, e.g. GITHUB_TOKEN loaded from .env.
+# Extra arguments are passed to goreleaser, e.g. 'mise run release --release-notes data/release-notes-X.Y.Z.md'.
+run = "op run -- goreleaser release --clean --draft"
+
 [tasks.uc]
 description = "Run the Uncloud CLI"
+raw_args = true
 run = "go run ./cmd/uc"
 
 [tasks.ucind]
 description = "Run ucind CLI for managing local Docker-based clusters"
+raw_args = true
 run = "go run ./cmd/ucind"
 
 [tasks."ucind:image"]

pkg/api/caddy.go

@@ -5,6 +5,15 @@ import (
 	"time"
 )
 
+const (
+	// LabelCaddyManaged marks the caddy service as managed by 'uc caddy deploy'.
+	LabelCaddyManaged = LabelReservedPrefix + "caddy.managed"
+	// LabelCaddyImageCustom marks the caddy image as explicitly set instead of following the latest default image.
+	LabelCaddyImageCustom = LabelReservedPrefix + "caddy.image.custom"
+	// LabelCaddyConfigCustom marks the global Caddy config as explicitly set instead of following the default config.
+	LabelCaddyConfigCustom = LabelReservedPrefix + "caddy.config.custom"
+)
+
 // CaddyConfig is the saved Caddy configuration on a machine.
 type CaddyConfig struct {
 	Caddyfile string

pkg/api/client.go

@@ -30,6 +30,7 @@ type ContainerClient interface {
 	StartContainer(ctx context.Context, serviceNameOrID, containerNameOrID string) error
 	StopContainer(ctx context.Context, serviceNameOrID, containerNameOrID string, opts container.StopOptions) error
 	RemoveContainer(ctx context.Context, serviceNameOrID, containerNameOrID string, opts container.RemoveOptions) error
+	UpdateContainer(ctx context.Context, serviceNameOrID, containerNameOrID string, opts UpdateContainerOptions) error
 	WaitContainerHealthy(
 		ctx context.Context, serviceNameOrID, containerNameOrID string, opts WaitContainerHealthyOptions,
 	) error
@@ -42,6 +43,8 @@ type DNSClient interface {
 type ImageClient interface {
 	InspectImage(ctx context.Context, id string) ([]MachineImage, error)
 	InspectRemoteImage(ctx context.Context, id string) ([]MachineRemoteImage, error)
+	ListImages(ctx context.Context, filter ImageFilter) ([]MachineImages, error)
+	RemoveImage(ctx context.Context, image string, opts RemoveImageOptions) ([]MachineImageRemoval, error)
 }
 
 type LogsClient interface {

pkg/api/container.go

@@ -19,6 +19,11 @@ const (
 	// in internal/machine/docker/manager.go.
 	DockerNetworkName = "uncloud"
 
+	// LabelReservedPrefix is the prefix of labels reserved for Uncloud that users can't set on services and containers.
+	LabelReservedPrefix = "uncloud."
+	// LabelDaemonReservedPrefix is the prefix of labels reserved for the Uncloud daemon system containers.
+	LabelDaemonReservedPrefix = "uncloudd."
+
 	// LabelDaemonManaged marks system containers as managed by the Uncloud daemon.
 	LabelDaemonManaged = "uncloudd.managed"
 
@@ -175,6 +180,19 @@ type CreateContainerResponse struct {
 	Name string
 }
 
+// UpdateContainerOptions defines the properties of a service container to update in place without recreating it.
+// Nil fields are left unchanged.
+type UpdateContainerOptions struct {
+	// CPU is the new CPU limit in nanocores. The limit can be changed but not removed.
+	CPU *int64
+	// Memory is the new memory limit in bytes. The limit can be changed but not removed.
+	Memory *int64
+	// MemoryReservation is the new memory reservation in bytes. The reservation can be changed but not removed.
+	MemoryReservation *int64
+	// ServiceLabels replaces all service labels. A pointer to an empty map removes all service labels.
+	ServiceLabels *map[string]string
+}
+
 type ServiceContainer struct {
 	Container
 	ServiceSpec ServiceSpec
@@ -195,6 +213,21 @@ func (c *ServiceContainer) ServiceName() string {
 	return c.Config.Labels[LabelServiceName]
 }
 
+// AppliedResources returns the resources from the service spec with the CPU and memory limits that Docker currently
+// applies to the container. The applied limits may differ from the spec if an in-place update was interrupted or
+// the container was updated outside of Uncloud.
+func (c *ServiceContainer) AppliedResources() ContainerResources {
+	resources := c.ServiceSpec.Container.Resources
+	if c.ContainerJSONBase == nil || c.HostConfig == nil {
+		return resources
+	}
+
+	resources.CPU = c.HostConfig.NanoCPUs
+	resources.Memory = c.HostConfig.Memory
+	resources.MemoryReservation = c.HostConfig.MemoryReservation
+	return resources
+}
+
 // ServiceMode returns the replication mode of the service this container belongs to.
 func (c *ServiceContainer) ServiceMode() string {
 	return c.ServiceSpec.Mode

pkg/api/image.go

@@ -1,8 +1,6 @@
 package api
 
 import (
-	"errors"
-
 	"github.com/distribution/reference"
 	"github.com/docker/docker/api/types/image"
 	v1 "github.com/opencontainers/image-spec/specs-go/v1"
@@ -24,14 +22,6 @@ type MachineImages struct {
 	ContainerdStore bool
 }
 
-// Error returns a non-nil error if listing images on this machine failed.
-func (mi MachineImages) Error() error {
-	if mi.Metadata != nil && mi.Metadata.Error != "" {
-		return errors.New(mi.Metadata.Error)
-	}
-	return nil
-}
-
 // ImageFilter defines criteria to filter images in ListImages.
 type ImageFilter struct {
 	// Machines filters images to those present on the specified machines (names or IDs).
@@ -42,6 +32,18 @@ type ImageFilter struct {
 	Name string
 }
 
+// RemoveImageOptions defines options for removing an image from cluster machines.
+type RemoveImageOptions struct {
+	// Machines selects machines to remove the image from (names or IDs). If empty, all cluster machines are selected.
+	Machines []string
+}
+
+// MachineImageRemoval contains the outcome of removing an image on one machine.
+type MachineImageRemoval struct {
+	Metadata *pb.Metadata
+	Results  []image.DeleteResponse
+}
+
 // MachineRemoteImage represents an image in a remote registry fetched by a particular machine.
 type MachineRemoteImage struct {
 	Metadata *pb.Metadata

pkg/api/machine.go

@@ -1,12 +1,33 @@
 package api
 
 import (
+	"fmt"
 	"net/netip"
 	"strings"
 
 	"github.com/psviderski/uncloud/api/pb"
 )
 
+// ValidateMachineName checks that a machine name is a lowercase DNS label and doesn't conflict with
+// internal DNS query modes or machine IDs.
+func ValidateMachineName(name string) error {
+	if !DNSLabelRegex.MatchString(name) {
+		return fmt.Errorf("invalid machine name %q: must be 1-63 characters, lowercase letters, numbers, "+
+			"and hyphens only, starting and ending with a letter or number", name)
+	}
+
+	switch name {
+	case "rr", "nearest":
+		return fmt.Errorf("invalid machine name %q: reserved for internal DNS query modes", name)
+	}
+	if IDRegex.MatchString(name) {
+		return fmt.Errorf(
+			"invalid machine name %q: must not match the machine ID format (32 hexadecimal characters)", name)
+	}
+
+	return nil
+}
+
 // MachineFilter defines criteria to filter machines in ListMachines.
 type MachineFilter struct {
 	// Available filters machines that are not DOWN.

pkg/api/machine_test.go

@@ -3,6 +3,7 @@ package api
 import (
 	"encoding/json"
 	"net/netip"
+	"strings"
 	"testing"
 
 	"github.com/psviderski/uncloud/api/pb"
@@ -10,6 +11,52 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func TestValidateMachineName(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		input   string
+		wantErr string
+	}{
+		{name: "single letter", input: "a"},
+		{name: "single digit", input: "1"},
+		{name: "two characters", input: "a1"},
+		{name: "generated name", input: "machine-ab12"},
+		{name: "hyphens and digits", input: "1-worker-2"},
+		{name: "maximum length", input: strings.Repeat("a", 63)},
+		{name: "maximum length with hyphens", input: "a" + strings.Repeat("-", 61) + "1"},
+		{name: "machine namespace label", input: "m"},
+		{name: "mode prefix", input: "nearest-worker"},
+		{name: "short hexadecimal name", input: strings.Repeat("a", 31)},
+		{name: "long hexadecimal name", input: strings.Repeat("a", 33)},
+		{name: "non-hexadecimal 32 characters", input: strings.Repeat("g", 32)},
+		{name: "empty", wantErr: "must be 1-63 characters"},
+		{name: "too long", input: strings.Repeat("a", 64), wantErr: "must be 1-63 characters"},
+		{name: "uppercase", input: "VPS1", wantErr: "lowercase letters"},
+		{name: "leading hyphen", input: "-worker", wantErr: "starting and ending"},
+		{name: "trailing hyphen", input: "worker-", wantErr: "starting and ending"},
+		{name: "underscore", input: "worker_1", wantErr: "hyphens only"},
+		{name: "dot", input: "worker.example", wantErr: "hyphens only"},
+		{name: "space", input: "worker 1", wantErr: "hyphens only"},
+		{name: "leading whitespace", input: " worker", wantErr: "hyphens only"},
+		{name: "trailing whitespace", input: "worker\t", wantErr: "hyphens only"},
+		{name: "non-ASCII", input: "wörker", wantErr: "lowercase letters"},
+		{name: "round-robin mode", input: "rr", wantErr: "reserved for internal DNS query modes"},
+		{name: "nearest mode", input: "nearest", wantErr: "reserved for internal DNS query modes"},
+		{name: "machine ID", input: "c337f00600de51ef4375c9a9a267dba5", wantErr: "machine ID format"},
+	}
+
+	for _, tt := range tests {
+		err := ValidateMachineName(tt.input)
+		if tt.wantErr == "" {
+			require.NoError(t, err)
+		} else {
+			require.ErrorContains(t, err, tt.wantErr)
+		}
+	}
+}
+
 func TestMachineMembersList_Info(t *testing.T) {
 	t.Parallel()
 

pkg/api/resources.go

@@ -1,7 +1,11 @@
 package api
 
 import (
+	"errors"
+
 	"github.com/docker/docker/api/types/container"
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
 )
 
 const (
@@ -27,6 +31,35 @@ type ContainerResources struct {
 	Ulimits map[string]Ulimit
 }
 
+// Equals returns true if the resources are equal, treating nil and empty slices and maps as equal.
+func (r ContainerResources) Equals(other ContainerResources) bool {
+	return cmp.Equal(r, other, cmpopts.EquateEmpty())
+}
+
+// ValidateUpdateInPlace returns an error explaining why a container created with these resources can't be updated
+// to the new resources without recreating it. It returns nil if the container can be updated in place.
+// Docker can only change the CPU and memory limits of an existing container and can't remove a limit once it's set.
+func (r ContainerResources) ValidateUpdateInPlace(new ContainerResources) error {
+	// Docker treats zero values as unchanged when updating a container.
+	if r.CPU != 0 && new.CPU == 0 {
+		return errors.New("CPU limit can't be removed")
+	}
+	if r.Memory != 0 && new.Memory == 0 {
+		return errors.New("memory limit can't be removed")
+	}
+	if r.MemoryReservation != 0 && new.MemoryReservation == 0 {
+		return errors.New("memory reservation can't be removed")
+	}
+
+	r.CPU = new.CPU
+	r.Memory = new.Memory
+	r.MemoryReservation = new.MemoryReservation
+	if !r.Equals(new) {
+		return errors.New("only CPU and memory limits can be changed")
+	}
+	return nil
+}
+
 // DeviceMapping represents a device mapping between host and container.
 type DeviceMapping struct {
 	// HostPath is the path to the device on the host.

pkg/api/resources_test.go

@@ -0,0 +1,72 @@
+package api
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestContainerResources_ValidateUpdateInPlace(t *testing.T) {
+	t.Parallel()
+
+	current := ContainerResources{
+		CPU:               Core,
+		Memory:            512 * 1024 * 1024,
+		MemoryReservation: 256 * 1024 * 1024,
+	}
+
+	tests := []struct {
+		name    string
+		new     func(r *ContainerResources)
+		wantErr string
+	}{
+		{
+			name: "same resources",
+			new:  func(*ContainerResources) {},
+		},
+		{
+			name: "change CPU and memory limits",
+			new: func(r *ContainerResources) {
+				r.CPU = 2 * Core
+				r.Memory = 1024 * 1024 * 1024
+				r.MemoryReservation = 128 * 1024 * 1024
+			},
+		},
+		{
+			name:    "remove CPU limit",
+			new:     func(r *ContainerResources) { r.CPU = 0 },
+			wantErr: "CPU limit can't be removed",
+		},
+		{
+			name:    "remove memory limit",
+			new:     func(r *ContainerResources) { r.Memory = 0 },
+			wantErr: "memory limit can't be removed",
+		},
+		{
+			name:    "remove memory reservation",
+			new:     func(r *ContainerResources) { r.MemoryReservation = 0 },
+			wantErr: "memory reservation can't be removed",
+		},
+		{
+			name:    "change shared memory",
+			new:     func(r *ContainerResources) { r.SharedMemory = 64 * 1024 * 1024 },
+			wantErr: "only CPU and memory limits can be changed",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			newResources := current
+			tt.new(&newResources)
+
+			err := current.ValidateUpdateInPlace(newResources)
+			if tt.wantErr == "" {
+				assert.NoError(t, err)
+			} else {
+				assert.EqualError(t, err, tt.wantErr)
+			}
+		})
+	}
+}

pkg/api/service.go

@@ -8,6 +8,7 @@ import (
 	"slices"
 	"strings"
 	"time"
+	"unicode"
 
 	mapset "github.com/deckarep/golang-set/v2"
 	"github.com/distribution/reference"
@@ -41,12 +42,30 @@ const (
 )
 
 var (
-	serviceIDRegexp = regexp.MustCompile("^[0-9a-f]{32}$")
+	IDRegex       = regexp.MustCompile("^[0-9a-f]{32}$")
-	dnsLabelRegexp  = regexp.MustCompile(`^[a-z0-9]([-a-z0-9]*[a-z0-9])?$`)
+	DNSLabelRegex = regexp.MustCompile(`^[a-z0-9]([-a-z0-9]{0,61}[a-z0-9])?$`)
 )
 
 func ValidateServiceID(id string) bool {
-	return serviceIDRegexp.MatchString(id)
+	return IDRegex.MatchString(id)
+}
+
+// ValidateServiceName checks that a service name is a lowercase DNS label and doesn't conflict with
+// the machine DNS namespace or service IDs.
+func ValidateServiceName(name string) error {
+	if !DNSLabelRegex.MatchString(name) {
+		return fmt.Errorf("invalid service name %q: must be 1-63 characters, lowercase letters, numbers, "+
+			"and hyphens only, starting and ending with a letter or number", name)
+	}
+	if name == "m" {
+		return fmt.Errorf("invalid service name %q: reserved for the machine DNS namespace", name)
+	}
+	if IDRegex.MatchString(name) {
+		return fmt.Errorf(
+			"invalid service name %q: must not match the service ID format (32 hexadecimal characters)", name)
+	}
+
+	return nil
 }
 
 // ServiceSpec defines the desired state of a service.
@@ -59,6 +78,9 @@ type ServiceSpec struct {
 	Configs []ConfigSpec
 	// Container defines the desired state of each container in the service.
 	Container ContainerSpec
+	// Labels are key-value metadata attached to the service. Unlike container labels, they're not applied to Docker
+	// containers. The "uncloud." key prefix is reserved for labels set by Uncloud.
+	Labels map[string]string
 	// Mode is the replication mode of the service. Default is ServiceModeReplicated if empty.
 	Mode string
 	Name string
@@ -139,6 +161,11 @@ func (s *ServiceSpec) Validate() error {
 	if err := s.Container.Validate(); err != nil {
 		return err
 	}
+	// We want to allow reserved uncloud. labels in the service spec, because we can't distinguish between user-defined
+	// in a Compose file and those set by a uc command.
+	if err := ValidateLabelKeys(s.Labels); err != nil {
+		return fmt.Errorf("invalid service labels: %w", err)
+	}
 
 	switch s.Mode {
 	case "", ServiceModeGlobal, ServiceModeReplicated:
@@ -147,12 +174,8 @@ func (s *ServiceSpec) Validate() error {
 	}
 
 	if s.Name != "" {
-		if len(s.Name) > 63 {
+		if err := ValidateServiceName(s.Name); err != nil {
-			return fmt.Errorf("service name too long (max 63 characters): %q", s.Name)
+			return err
-		}
-		if !dnsLabelRegexp.MatchString(s.Name) {
-			return fmt.Errorf("invalid service name: %q. must be 1-63 characters, lowercase letters, numbers, "+
-				"and dashes only; must start and end with a letter or number", s.Name)
 		}
 	}
 
@@ -229,6 +252,7 @@ func (s *ServiceSpec) Clone() ServiceSpec {
 		spec.Caddy = &caddyCopy
 	}
 	spec.Container = s.Container.Clone()
+	spec.Labels = maps.Clone(s.Labels)
 	spec.PreDeploy = s.PreDeploy.Clone()
 
 	if s.Ports != nil {
@@ -265,6 +289,10 @@ type ContainerSpec struct {
 	Image       string
 	// Run a custom init inside the container. If nil, use the daemon's configured settings.
 	Init *bool
+	// Labels are user-defined key-value metadata applied as Docker labels to each container in the service.
+	// Docker labels are immutable so changing them requires recreating the containers.
+	// Keys must not use the reserved "uncloud." prefix.
+	Labels map[string]string
 	// LogDriver overrides the default logging driver for the container. Each Docker daemon can have its own default.
 	LogDriver *LogDriver
 	// PidMode sets the PID namespace mode for the container. Currently only "" or "host" is supported.
@@ -319,6 +347,9 @@ func (s *ContainerSpec) Validate() error {
 	if _, err := reference.ParseDockerRef(s.Image); err != nil {
 		return fmt.Errorf("invalid image '%s': %w", s.Image, err)
 	}
+	if err := ValidateLabels(s.Labels); err != nil {
+		return fmt.Errorf("invalid container labels: %w", err)
+	}
 
 	for _, m := range s.VolumeMounts {
 		if err := m.Validate(); err != nil {
@@ -376,6 +407,7 @@ func (s *ContainerSpec) Clone() ContainerSpec {
 		hc.Test = slices.Clone(s.Healthcheck.Test)
 		spec.Healthcheck = &hc
 	}
+	spec.Labels = maps.Clone(s.Labels)
 	if s.LogDriver != nil {
 		logDriver := *s.LogDriver
 		if s.LogDriver.Options != nil {
@@ -414,6 +446,34 @@ func (s *ContainerSpec) Clone() ContainerSpec {
 	return spec
 }
 
+// ValidateLabelKeys checks that label keys are non-empty and don't contain whitespace.
+func ValidateLabelKeys(labels map[string]string) error {
+	for k := range labels {
+		if strings.TrimSpace(k) == "" {
+			return fmt.Errorf("label key must not be empty")
+		}
+		if strings.ContainsFunc(k, unicode.IsSpace) {
+			return fmt.Errorf("label '%s' must not contain whitespace", k)
+		}
+	}
+	return nil
+}
+
+// ValidateLabels checks that label keys are valid and don't use the reserved Uncloud prefixes. Use it for labels
+// provided by users and for container labels.
+func ValidateLabels(labels map[string]string) error {
+	if err := ValidateLabelKeys(labels); err != nil {
+		return err
+	}
+	for k := range labels {
+		if strings.HasPrefix(k, LabelReservedPrefix) || strings.HasPrefix(k, LabelDaemonReservedPrefix) {
+			return fmt.Errorf("label '%s' cannot use the reserved prefix '%s' or '%s'",
+				k, LabelReservedPrefix, LabelDaemonReservedPrefix)
+		}
+	}
+	return nil
+}
+
 type EnvVars map[string]string
 
 // ToSlice converts the environment variables to a slice of strings in the format "key=value".
@@ -552,6 +612,36 @@ func (s *Service) FindContainer(nameOrID string) (MachineServiceContainer, error
 	return MachineServiceContainer{}, ErrNotFound
 }
 
+// Spec returns the service spec derived from its containers.
+// Containers may have different specs, for example, after a failed deployment. The result is the spec of the newest
+// running container, or the newest container if none are running. Replicas is set to the actual number of containers
+// for a replicated service.
+func (s *Service) Spec() ServiceSpec {
+	if len(s.Containers) == 0 {
+		return ServiceSpec{}
+	}
+
+	isRunning := func(c MachineServiceContainer) bool {
+		return c.Container.State.Running && !c.Container.State.Paused
+	}
+	selected := slices.MaxFunc(s.Containers, func(a, b MachineServiceContainer) int {
+		if aRunning, bRunning := isRunning(a), isRunning(b); aRunning != bRunning {
+			if aRunning {
+				return 1
+			}
+			return -1
+		}
+		return a.Container.CreatedTime().Compare(b.Container.CreatedTime())
+	})
+
+	spec := selected.Container.ServiceSpec.Clone()
+	// The replicas stored in a container spec go stale when the service is scaled without recreating the container.
+	if spec.Mode == ServiceModeReplicated {
+		spec.Replicas = uint(len(s.Containers))
+	}
+	return spec
+}
+
 // MachineIDs returns a list of unique machine IDs where the service containers are running.
 func (s *Service) MachineIDs() []string {
 	ids := mapset.NewSet[string]()

pkg/api/service_test.go

@@ -2,6 +2,7 @@ package api
 
 import (
 	"os"
+	"strings"
 	"testing"
 
 	"github.com/docker/docker/api/types/container"
@@ -9,6 +10,137 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+func TestServiceSpec_Validate_Name(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		input   string
+		wantErr string
+	}{
+		{name: "empty allows generated name"},
+		{name: "single character", input: "a"},
+		{name: "single digit", input: "1"},
+		{name: "two characters", input: "a1"},
+		{name: "hyphens and digits", input: "1-web-2"},
+		{name: "consecutive hyphens", input: "web--1"},
+		{name: "maximum length", input: strings.Repeat("a", 63)},
+		{name: "maximum length with hyphens", input: "a" + strings.Repeat("-", 61) + "1"},
+		{name: "round-robin mode is a valid service name", input: "rr"},
+		{name: "nearest mode is a valid service name", input: "nearest"},
+		{name: "machine namespace prefix", input: "m-service"},
+		{name: "short hexadecimal name", input: strings.Repeat("a", 31)},
+		{name: "long hexadecimal name", input: strings.Repeat("a", 33)},
+		{name: "non-hexadecimal 32 characters", input: strings.Repeat("g", 32)},
+		{name: "too long", input: strings.Repeat("a", 64), wantErr: "must be 1-63 characters"},
+		{name: "uppercase", input: "WEB1", wantErr: "lowercase letters"},
+		{name: "leading hyphen", input: "-web", wantErr: "starting and ending"},
+		{name: "trailing hyphen", input: "web-", wantErr: "starting and ending"},
+		{name: "hyphen only", input: "-", wantErr: "starting and ending"},
+		{name: "underscore", input: "web_1", wantErr: "hyphens only"},
+		{name: "dot", input: "web.example", wantErr: "hyphens only"},
+		{name: "slash", input: "web/api", wantErr: "hyphens only"},
+		{name: "space", input: "web 1", wantErr: "hyphens only"},
+		{name: "leading whitespace", input: " web", wantErr: "hyphens only"},
+		{name: "trailing whitespace", input: "web ", wantErr: "hyphens only"},
+		{name: "tab", input: "web\t1", wantErr: "hyphens only"},
+		{name: "newline", input: "web\n", wantErr: "hyphens only"},
+		{name: "non-ASCII", input: "wéb", wantErr: "lowercase letters"},
+		{name: "machine DNS namespace", input: "m", wantErr: "reserved for the machine DNS namespace"},
+		{name: "service ID", input: "c337f00600de51ef4375c9a9a267dba5", wantErr: "service ID format"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			spec := ServiceSpec{Name: tt.input, Container: ContainerSpec{Image: "nginx:latest"}}
+			err := spec.Validate()
+			if tt.wantErr != "" {
+				require.ErrorContains(t, err, tt.wantErr)
+			} else {
+				require.NoError(t, err)
+			}
+		})
+	}
+}
+
+func TestServiceSpec_Validate_Labels(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		labels  map[string]string
+		wantErr string
+	}{
+		{name: "nil labels"},
+		{name: "valid labels", labels: map[string]string{"com.example.team": "web", "tier": ""}},
+		{name: "prefix without dot", labels: map[string]string{"uncloudfoo": "bar"}},
+		{name: "whitespace in value", labels: map[string]string{"tier": "value with spaces"}},
+		{name: "empty key", labels: map[string]string{"": "bar"}, wantErr: "must not be empty"},
+		{name: "only whitespace in key", labels: map[string]string{" ": "bar"}, wantErr: "must not be empty"},
+		{name: "space in key", labels: map[string]string{"my label": "bar"}, wantErr: "must not contain whitespace"},
+		{name: "leading space in key", labels: map[string]string{" tier": "bar"}, wantErr: "must not contain whitespace"},
+		{name: "trailing space in key", labels: map[string]string{"tier ": "bar"}, wantErr: "must not contain whitespace"},
+		{name: "tab in key", labels: map[string]string{"my\tlabel": "bar"}, wantErr: "must not contain whitespace"},
+		{name: "newline in key", labels: map[string]string{"my\nlabel": "bar"}, wantErr: "must not contain whitespace"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name+"/service", func(t *testing.T) {
+			t.Parallel()
+			spec := ServiceSpec{Container: ContainerSpec{Image: "nginx:latest"}, Labels: tt.labels}
+			err := spec.Validate()
+			if tt.wantErr != "" {
+				require.ErrorContains(t, err, "invalid service labels")
+				require.ErrorContains(t, err, tt.wantErr)
+			} else {
+				require.NoError(t, err)
+			}
+		})
+		t.Run(tt.name+"/container", func(t *testing.T) {
+			t.Parallel()
+			spec := ServiceSpec{Container: ContainerSpec{Image: "nginx:latest", Labels: tt.labels}}
+			err := spec.Validate()
+			if tt.wantErr != "" {
+				require.ErrorContains(t, err, "invalid container labels")
+				require.ErrorContains(t, err, tt.wantErr)
+			} else {
+				require.NoError(t, err)
+			}
+		})
+	}
+}
+
+func TestServiceSpec_Validate_ReservedLabels(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name   string
+		labels map[string]string
+	}{
+		{name: "reserved prefix", labels: map[string]string{LabelManaged: ""}},
+		{name: "reserved daemon prefix", labels: map[string]string{"uncloudd.foo": "bar"}},
+		{name: "managed service", labels: map[string]string{
+			LabelReservedPrefix + "service.managed": "",
+			LabelReservedPrefix + "custom.key":      "value",
+		}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			// Uncloud sets its own labels on services, so a service spec can have them.
+			spec := ServiceSpec{Container: ContainerSpec{Image: "nginx:latest"}, Labels: tt.labels}
+			require.NoError(t, spec.Validate())
+
+			// Container labels and labels provided by users can't use the reserved prefixes.
+			spec = ServiceSpec{Container: ContainerSpec{Image: "nginx:latest", Labels: tt.labels}}
+			err := spec.Validate()
+			require.ErrorContains(t, err, "invalid container labels")
+			require.ErrorContains(t, err, "reserved prefix")
+			require.ErrorContains(t, ValidateLabels(tt.labels), "reserved prefix")
+		})
+	}
+}
+
 func TestServiceSpec_Validate_CaddyAndPorts(t *testing.T) {
 	tests := []struct {
 		name    string
@@ -215,6 +347,7 @@ func TestContainerSpec_Clone(t *testing.T) {
 		},
 		Image:  "nginx:latest",
 		Init:   new(true),
+		Labels: map[string]string{"com.example.team": "web"},
 		LogDriver: &LogDriver{
 			Name: "json-file",
 			Options: map[string]string{
@@ -259,6 +392,7 @@ func TestContainerSpec_Clone(t *testing.T) {
 	original.Command[0] = stringModified
 	original.Entrypoint[0] = stringModified
 	original.Env["FOO"] = stringModified
+	original.Labels["com.example.team"] = stringModified
 	original.LogDriver.Options["max-size"] = stringModified
 	original.Volumes[0] = stringModified
 	original.VolumeMounts[0].ContainerPath = stringModified
@@ -274,6 +408,7 @@ func TestContainerSpec_Clone(t *testing.T) {
 	assert.Equal(t, "NET_ADMIN", cloned.CapAdd[0])
 	assert.Equal(t, "ALL", cloned.CapDrop[0])
 	assert.Equal(t, "sh", cloned.Command[0])
+	assert.Equal(t, "web", cloned.Labels["com.example.team"])
 	assert.Equal(t, "/bin/bash", cloned.Entrypoint[0])
 	assert.Equal(t, "bar", cloned.Env["FOO"])
 	assert.Equal(t, "qux", cloned.Env["BAZ"])
@@ -299,3 +434,110 @@ func TestContainerSpec_Clone(t *testing.T) {
 	assert.Equal(t, os.FileMode(0o644), *cloned.ConfigMounts[0].Mode, "Mode should be deep copied")
 	assert.Equal(t, "1", cloned.Sysctls["net.ipv4.ip_forward"])
 }
+
+func TestService_Spec(t *testing.T) {
+	t.Parallel()
+
+	newContainer := func(name, created string, running bool, mode string, replicas uint) MachineServiceContainer {
+		return MachineServiceContainer{Container: ServiceContainer{
+			Container: Container{InspectResponse: container.InspectResponse{
+				ContainerJSONBase: &container.ContainerJSONBase{
+					Created: created,
+					State:   &container.State{Running: running},
+				},
+			}},
+			ServiceSpec: ServiceSpec{
+				Labels:   map[string]string{"from": name},
+				Mode:     mode,
+				Replicas: replicas,
+			},
+		}}
+	}
+	const (
+		older = "2026-01-01T00:00:00Z"
+		newer = "2026-01-02T00:00:00Z"
+	)
+
+	t.Run("no containers", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{HookContainers: []MachineServiceContainer{
+			newContainer("hook", newer, true, ServiceModeReplicated, 1),
+		}}
+
+		assert.Equal(t, ServiceSpec{}, svc.Spec())
+	})
+
+	t.Run("newest running container wins", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{Containers: []MachineServiceContainer{
+			newContainer("old-running", older, true, ServiceModeReplicated, 1),
+			newContainer("new-running", newer, true, ServiceModeReplicated, 1),
+		}}
+
+		spec := svc.Spec()
+		assert.Equal(t, "new-running", spec.Labels["from"])
+	})
+
+	t.Run("running container wins over newer stopped one", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{Containers: []MachineServiceContainer{
+			newContainer("new-stopped", newer, false, ServiceModeReplicated, 1),
+			newContainer("old-running", older, true, ServiceModeReplicated, 1),
+		}}
+
+		spec := svc.Spec()
+		assert.Equal(t, "old-running", spec.Labels["from"])
+	})
+
+	t.Run("newest container wins when none are running", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{Containers: []MachineServiceContainer{
+			newContainer("old-stopped", older, false, ServiceModeReplicated, 1),
+			newContainer("new-stopped", newer, false, ServiceModeReplicated, 1),
+		}}
+
+		spec := svc.Spec()
+		assert.Equal(t, "new-stopped", spec.Labels["from"])
+	})
+
+	t.Run("replicas match the number of containers", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{Containers: []MachineServiceContainer{
+			newContainer("a", older, true, ServiceModeReplicated, 5),
+			newContainer("b", newer, true, ServiceModeReplicated, 5),
+			newContainer("c", newer, false, ServiceModeReplicated, 5),
+		}}
+
+		spec := svc.Spec()
+		assert.EqualValues(t, 3, spec.Replicas)
+	})
+
+	t.Run("global service has no replicas", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{Containers: []MachineServiceContainer{
+			newContainer("a", older, true, ServiceModeGlobal, 0),
+			newContainer("b", newer, true, ServiceModeGlobal, 0),
+		}}
+
+		spec := svc.Spec()
+		assert.Zero(t, spec.Replicas)
+	})
+
+	t.Run("result is independent of the container spec", func(t *testing.T) {
+		t.Parallel()
+
+		svc := Service{Containers: []MachineServiceContainer{
+			newContainer("a", older, true, ServiceModeReplicated, 1),
+		}}
+
+		spec := svc.Spec()
+		spec.Labels["from"] = "changed"
+		assert.Equal(t, "a", svc.Containers[0].Container.ServiceSpec.Labels["from"])
+	})
+}

pkg/client/caddy.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"regexp"
+	"strings"
 	"time"
 
 	"github.com/Masterminds/semver"
@@ -11,6 +12,7 @@ import (
 	"github.com/google/go-containerregistry/pkg/name"
 	"github.com/google/go-containerregistry/pkg/v1/remote"
 	"github.com/psviderski/uncloud/api/pb"
+	"github.com/psviderski/uncloud/internal/version"
 	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/psviderski/uncloud/pkg/client/deploy"
 	"google.golang.org/protobuf/types/known/emptypb"
@@ -18,11 +20,33 @@ import (
 
 const (
 	CaddyServiceName = "caddy"
-	// CaddyImage is the official Caddy Docker image on Docker Hub: https://hub.docker.com/_/caddy
+	// CaddyImage is the Uncloud Caddy image that is deployed by default. It includes the Uncloud storage module for
-	CaddyImage = "caddy"
+	// sharing TLS certificates across machines: https://github.com/unlabs-dev/caddy-uncloud
+	CaddyImage = "ghcr.io/unlabs-dev/caddy-uncloud"
+	// DefaultCaddyConfig is the default global Caddyfile. It must be compatible with all versions of CaddyImage allowed
+	// by caddyImageConstraint.
+	DefaultCaddyConfig = "{\n\tstorage uncloud\n}"
+	// CaddyStorageFileSystem is the name of the default Caddy storage module that stores TLS assets locally.
+	CaddyStorageFileSystem = "file_system"
+	// CaddyStorageUncloud is the name of the Caddy storage module that stores TLS assets in the cluster.
+	CaddyStorageUncloud = "uncloud"
+	// MinCaddyDaemonVersion is the minimum daemon version required for 'uc caddy deploy'. Older daemons don't store
+	// service labels that mark the service as managed by this command. Distributed locks were introduced in 0.21.0,
+	// which is the minimum required by the cluster storage even on machines that don't run the Caddy service.
+	MinCaddyDaemonVersion = "0.22.0"
+
+	// officialCaddyImage is the official Caddy Docker image on Docker Hub: https://hub.docker.com/_/caddy
+	officialCaddyImage = "docker.io/library/caddy"
 )
 
-var caddyImageTagRegex = regexp.MustCompile(`^2\.\d+\.\d+$`)
+var (
+	// caddyImageConstraint limits CaddyImage to the major version that is known to be compatible with
+	// DefaultCaddyConfig.
+	caddyImageConstraint = mustNewConstraint("<1.0.0")
+	// caddyImageTagRegex matches the exact version tags of CaddyImage, skipping floating tags and image variants.
+	caddyImageTagRegex = regexp.MustCompile(`^\d+\.\d+\.\d+$`)
+	caddyStorageRegex  = regexp.MustCompile(`(?m)^[ \t]*storage[ \t]+(\S+)`)
+)
 
 // CaddyClient provides Caddy operations over its parent Client's connection.
 // The parent client owns the connection and must be used to close it.
@@ -64,19 +88,28 @@ func (c *CaddyClient) Config(ctx context.Context, opts CaddyConfigOptions) (api.
 	return config, nil
 }
 
-// CaddyDeploymentOptions configures a Caddy reverse proxy deployment.
+// CaddyDeploymentOptions configures a Caddy reverse proxy deployment. The zero value deploys the latest version of
+// CaddyImage with DefaultCaddyConfig to all machines.
 type CaddyDeploymentOptions struct {
-	// Image defaults to the latest stable 2.x.x official Caddy image.
+	// Image defaults to the latest version of CaddyImage. A non-empty value is marked as custom.
-	// Custom images must include curl and start Caddy with /etc/caddy/Caddyfile.
+	// Custom images must include curl and start Caddy with /etc/caddy/Caddyfile. They must also include the Uncloud
+	// storage module if the config uses 'storage uncloud'.
 	Image string
-	// Config contains an optional global Caddyfile.
+	// Config is the global Caddyfile. Nil means DefaultCaddyConfig. Any other value, even empty, is marked as custom.
-	Config    string
+	Config    *string
 	Placement api.Placement
 }
 
-// NewDeployment creates a new deployment for a Caddy reverse proxy service.
+// ConfigOrDefault returns the global Caddyfile of the deployment.
-// The service is deployed in global mode to all machines in the cluster. If the image is not provided, the latest
+func (o CaddyDeploymentOptions) ConfigOrDefault() string {
-// version of the official Caddy Docker image is used.
+	if o.Config != nil {
+		return *o.Config
+	}
+	return DefaultCaddyConfig
+}
+
+// NewDeployment creates a new managed deployment for a Caddy reverse proxy service.
+// The service is deployed in global mode to all machines in the cluster or the machines set in the placement.
 func (c *CaddyClient) NewDeployment(ctx context.Context, opts CaddyDeploymentOptions) (*deploy.Deployment, error) {
 	if err := ctx.Err(); err != nil {
 		return nil, err
@@ -90,6 +123,10 @@ func (c *CaddyClient) NewDeployment(ctx context.Context, opts CaddyDeploymentOpt
 
 		image = reference.FamiliarString(latest)
 	}
+	config := opts.ConfigOrDefault()
+	if err := validateCaddyImageStorage(image, config); err != nil {
+		return nil, err
+	}
 
 	spec := api.ServiceSpec{
 		Container: api.ContainerSpec{
@@ -186,17 +223,119 @@ func (c *CaddyClient) NewDeployment(ctx context.Context, opts CaddyDeploymentOpt
 		},
 	}
 
-	if opts.Config != "" {
+	if config != "" {
 		spec.Caddy = &api.CaddySpec{
-			Config: opts.Config,
+			Config: config,
 		}
 	}
 
+	spec.Labels = map[string]string{api.LabelCaddyManaged: ""}
+	if opts.Image != "" {
+		spec.Labels[api.LabelCaddyImageCustom] = ""
+	}
+	if opts.Config != nil {
+		spec.Labels[api.LabelCaddyConfigCustom] = ""
+	}
+
 	return c.client.NewDeployment(spec, nil), nil
 }
 
-// latestCaddyImage returns the latest image of the official Caddy Docker image on Docker Hub.
+// CheckDaemonVersions returns an error naming the available machines whose daemons are older than
-// The latest image is determined by the latest version tag 2.x.x.
+// MinCaddyDaemonVersion. Machines that are down are skipped because a deployment doesn't touch them.
+func (c *CaddyClient) CheckDaemonVersions(ctx context.Context) error {
+	machines, err := c.client.ListMachines(ctx, &api.MachineFilter{Available: true})
+	if err != nil {
+		return fmt.Errorf("list machines: %w", err)
+	}
+
+	if outdated := outdatedCaddyDaemons(machines); len(outdated) > 0 {
+		return fmt.Errorf("deploying Caddy requires Uncloud daemon %s or newer on all machines, upgrade these "+
+			"machines and try again: %s. See https://github.com/psviderski/uncloud/releases/latest",
+			MinCaddyDaemonVersion, strings.Join(outdated, ", "))
+	}
+	return nil
+}
+
+// outdatedCaddyDaemons returns descriptions of the machines with outdated daemons. Prereleases and development builds
+// aren't considered outdated because their versions don't show whether they're older than a stable release.
+func outdatedCaddyDaemons(machines api.MachineMembersList) []string {
+	minVersion := semver.MustParse(MinCaddyDaemonVersion)
+
+	var outdated []string
+	for _, m := range machines {
+		v, err := semver.NewVersion(m.Machine.DaemonVersion)
+		if err != nil || v.Prerelease() != "" {
+			continue
+		}
+		if v.LessThan(minVersion) {
+			outdated = append(outdated, fmt.Sprintf("%s (%s)", m.Machine.Name, m.Machine.DaemonVersion))
+		}
+	}
+
+	return outdated
+}
+
+// CaddyStorageModule returns the name of the storage module set in the global options of a Caddyfile.
+func CaddyStorageModule(config string) string {
+	if match := caddyStorageRegex.FindStringSubmatch(config); match != nil {
+		return match[1]
+	}
+	return CaddyStorageFileSystem
+}
+
+// IsDefaultCaddyImage returns whether the image is any version of CaddyImage.
+func IsDefaultCaddyImage(image string) bool {
+	return imageName(image) == CaddyImage
+}
+
+// IsNewerCaddyImage returns whether the image is a newer version of CaddyImage than the other image.
+// It returns false if any of the images isn't a version of CaddyImage.
+func IsNewerCaddyImage(image, other string) bool {
+	version, otherVersion := caddyImageVersion(image), caddyImageVersion(other)
+	return version != nil && otherVersion != nil && version.GreaterThan(otherVersion)
+}
+
+// caddyImageVersion returns the version in the image tag if the image is a version of CaddyImage, or nil otherwise.
+func caddyImageVersion(image string) *semver.Version {
+	named, err := reference.ParseNormalizedNamed(image)
+	if err != nil || named.Name() != CaddyImage {
+		return nil
+	}
+	tagged, ok := named.(reference.Tagged)
+	if !ok {
+		return nil
+	}
+	version, err := semver.NewVersion(tagged.Tag())
+	if err != nil {
+		return nil
+	}
+
+	return version
+}
+
+// imageName returns the normalised name of the image without its tag or digest.
+// It returns an empty string if the image reference is invalid.
+func imageName(image string) string {
+	named, err := reference.ParseNormalizedNamed(image)
+	if err != nil {
+		return ""
+	}
+	return named.Name()
+}
+
+// validateCaddyImageStorage rejects the official Caddy image with a config that uses the Uncloud storage module
+// because it doesn't include this module. Otherwise, a new container could start from the previously saved Caddyfile
+// and pass its health check, but its routes would stop updating as Caddy would reject the new config.
+func validateCaddyImageStorage(image, config string) error {
+	if imageName(image) == officialCaddyImage && CaddyStorageModule(config) == CaddyStorageUncloud {
+		return fmt.Errorf("official Caddy image '%s' doesn't support 'storage uncloud' in the global Caddy config: "+
+			"use an image that includes the Uncloud storage module or a Caddyfile without this option. "+
+			"See %s/concepts/ingress/cluster-storage-for-caddy", image, version.DocsURL)
+	}
+	return nil
+}
+
+// latestCaddyImage returns the latest version of CaddyImage allowed by caddyImageConstraint.
 func latestCaddyImage(ctx context.Context) (reference.NamedTagged, error) {
 	if err := ctx.Err(); err != nil {
 		return nil, err
@@ -222,7 +361,7 @@ func latestCaddyImage(ctx context.Context) (reference.NamedTagged, error) {
 	return imageWithTag, nil
 }
 
-// latestCaddyTag selects the newest stable 2.x.x tag, falling back to latest.
+// latestCaddyTag selects the newest tag matching caddyImageTagRegex and caddyImageConstraint, falling back to latest.
 func latestCaddyTag(tags []string) string {
 	latestTag := "latest"
 	var latestVersion *semver.Version
@@ -232,7 +371,7 @@ func latestCaddyTag(tags []string) string {
 		}
 
 		v, err := semver.NewVersion(t)
-		if err != nil {
+		if err != nil || !caddyImageConstraint.Check(v) {
 			continue
 		}
 		if latestVersion == nil || v.GreaterThan(latestVersion) {
@@ -243,3 +382,12 @@ func latestCaddyTag(tags []string) string {
 
 	return latestTag
 }
+
+// mustNewConstraint parses a version constraint and panics if it's invalid.
+func mustNewConstraint(constraint string) *semver.Constraints {
+	c, err := semver.NewConstraint(constraint)
+	if err != nil {
+		panic(fmt.Sprintf("invalid version constraint '%s': %v", constraint, err))
+	}
+	return c
+}

pkg/client/caddy_test.go

@@ -5,6 +5,8 @@ import (
 	"testing"
 
 	"github.com/distribution/reference"
+	"github.com/psviderski/uncloud/api/pb"
+	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
@@ -15,7 +17,7 @@ func TestLatestCaddyImage(t *testing.T) {
 	image, err := latestCaddyImage(context.Background())
 	require.NoError(t, err)
 
-	assert.Regexp(t, `^caddy:2\.\d+\.\d+$`, reference.FamiliarString(image))
+	assert.Regexp(t, `^ghcr\.io/unlabs-dev/caddy-uncloud:0\.\d+\.\d+$`, reference.FamiliarString(image))
 }
 
 func TestLatestCaddyTag(t *testing.T) {
@@ -27,14 +29,209 @@ func TestLatestCaddyTag(t *testing.T) {
 		want string
 	}{
 		{name: "empty", want: "latest"},
-		{name: "semantic ordering", tags: []string{"2.9.9", "2.11.4", "2.10.0", "2.11.3"}, want: "2.11.4"},
+		{name: "semantic ordering", tags: []string{"0.1.9", "0.1.11", "0.1.10", "0.1.2"}, want: "0.1.11"},
-		{name: "ignore other versions and variants", tags: []string{
+		{name: "newest minor version", tags: []string{"0.1.3", "0.2.0", "0.0.9"}, want: "0.2.0"},
-			"latest", "1.0.0", "3.0.0", "2.12.0-rc.1", "2.12.0-alpine", "2.12", "v2.12.0", "2.11.4",
+		{name: "ignore newer major versions", tags: []string{"1.0.0", "0.1.3", "2.3.4", "1.0.1"}, want: "0.1.3"},
-		}, want: "2.11.4"},
+		{name: "ignore variants", tags: []string{
-		{name: "no stable version", tags: []string{"builder", "2.12.0-beta.1"}, want: "latest"},
+			"latest", "0.1.4-rc.1", "0.1.4-alpine", "0.1", "v0.1.4", "1.0.0-rc.1", "0.1.3",
+		}, want: "0.1.3"},
+		{name: "no stable version", tags: []string{"builder", "0.1.4-beta.1"}, want: "latest"},
 	} {
 		t.Run(tt.name, func(t *testing.T) {
 			assert.Equal(t, tt.want, latestCaddyTag(tt.tags))
 		})
 	}
 }
+
+func TestDefaultCaddyConfig_UsesUncloudStorage(t *testing.T) {
+	t.Parallel()
+
+	assert.Equal(t, CaddyStorageUncloud, CaddyStorageModule(DefaultCaddyConfig))
+}
+
+func TestCaddyStorageModule(t *testing.T) {
+	t.Parallel()
+
+	for _, tt := range []struct {
+		name   string
+		config string
+		want   string
+	}{
+		{name: "empty config", want: CaddyStorageFileSystem},
+		{name: "no storage option", config: "{\n\tdebug\n}", want: CaddyStorageFileSystem},
+		{name: "uncloud", config: "{\n\tstorage uncloud\n}", want: CaddyStorageUncloud},
+		{name: "uncloud with options", config: "{\n\tstorage uncloud {\n\t\tfoo bar\n\t}\n}", want: CaddyStorageUncloud},
+		{name: "spaces around", config: "{\n    storage   uncloud  \n}", want: CaddyStorageUncloud},
+		{name: "other module", config: "{\n\tstorage redis\n}", want: "redis"},
+		{name: "explicit file system", config: "{\n\tstorage file_system /data\n}", want: CaddyStorageFileSystem},
+		{name: "commented out", config: "{\n\t# storage uncloud\n}", want: CaddyStorageFileSystem},
+		{name: "similar option", config: "{\n\tstorage_clean_interval 24h\n}", want: CaddyStorageFileSystem},
+		{name: "no module on the same line", config: "{\n\tstorage\n\tdebug\n}", want: CaddyStorageFileSystem},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, CaddyStorageModule(tt.config))
+		})
+	}
+}
+
+func TestIsDefaultCaddyImage(t *testing.T) {
+	t.Parallel()
+
+	assert.True(t, IsDefaultCaddyImage("ghcr.io/unlabs-dev/caddy-uncloud"))
+	assert.True(t, IsDefaultCaddyImage("ghcr.io/unlabs-dev/caddy-uncloud:0.1.3"))
+	assert.False(t, IsDefaultCaddyImage("caddy:2.11.4"))
+	assert.False(t, IsDefaultCaddyImage("example.com/unlabs-dev/caddy-uncloud:0.1.3"))
+	assert.False(t, IsDefaultCaddyImage(""))
+}
+
+func TestIsNewerCaddyImage(t *testing.T) {
+	t.Parallel()
+
+	const repo = "ghcr.io/unlabs-dev/caddy-uncloud"
+	for _, tt := range []struct {
+		name  string
+		image string
+		other string
+		want  bool
+	}{
+		{name: "newer major version", image: repo + ":1.0.0", other: repo + ":0.1.3", want: true},
+		{name: "newer minor version", image: repo + ":0.2.0", other: repo + ":0.1.3", want: true},
+		{name: "same", image: repo + ":0.1.3", other: repo + ":0.1.3"},
+		{name: "older", image: repo + ":0.1.2", other: repo + ":0.1.3"},
+		{name: "different repo", image: "example.com/custom:1.0.0", other: repo + ":0.1.3"},
+		{name: "official image", image: "caddy:2.11.4", other: repo + ":0.1.3"},
+		{name: "tag without version", image: repo + ":latest", other: repo + ":0.1.3"},
+		{name: "no tag", image: repo, other: repo + ":0.1.3"},
+		{name: "fallback to latest tag", image: repo + ":1.0.0", other: repo + ":latest"},
+		{name: "no current image", other: repo + ":0.1.3"},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, IsNewerCaddyImage(tt.image, tt.other))
+		})
+	}
+}
+
+func TestCaddyClientNewDeployment(t *testing.T) {
+	t.Parallel()
+
+	const customImage = "example.com/caddy:1.2.3"
+	for _, tt := range []struct {
+		name       string
+		opts       CaddyDeploymentOptions
+		wantConfig string
+		wantLabels map[string]string
+	}{
+		{
+			name:       "defaults",
+			wantConfig: DefaultCaddyConfig,
+			wantLabels: map[string]string{api.LabelCaddyManaged: ""},
+		},
+		{
+			name:       "custom image",
+			opts:       CaddyDeploymentOptions{Image: customImage},
+			wantConfig: DefaultCaddyConfig,
+			wantLabels: map[string]string{api.LabelCaddyManaged: "", api.LabelCaddyImageCustom: ""},
+		},
+		{
+			name:       "custom config",
+			opts:       CaddyDeploymentOptions{Config: new("{\n\tdebug\n}")},
+			wantConfig: "{\n\tdebug\n}",
+			wantLabels: map[string]string{api.LabelCaddyManaged: "", api.LabelCaddyConfigCustom: ""},
+		},
+		{
+			name:       "custom empty config",
+			opts:       CaddyDeploymentOptions{Config: new("")},
+			wantLabels: map[string]string{api.LabelCaddyManaged: "", api.LabelCaddyConfigCustom: ""},
+		},
+		{
+			name:       "custom config matching the default is still custom",
+			opts:       CaddyDeploymentOptions{Config: new(DefaultCaddyConfig)},
+			wantConfig: DefaultCaddyConfig,
+			wantLabels: map[string]string{api.LabelCaddyManaged: "", api.LabelCaddyConfigCustom: ""},
+		},
+		{
+			name: "custom image and config",
+			opts: CaddyDeploymentOptions{
+				Image:  customImage,
+				Config: new("{\n\tdebug\n}"),
+			},
+			wantConfig: "{\n\tdebug\n}",
+			wantLabels: map[string]string{
+				api.LabelCaddyManaged:      "",
+				api.LabelCaddyImageCustom:  "",
+				api.LabelCaddyConfigCustom: "",
+			},
+		},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			c := &CaddyClient{client: &Client{}}
+			d, err := c.NewDeployment(context.Background(), tt.opts)
+			require.NoError(t, err)
+
+			assert.Equal(t, tt.wantConfig, d.Spec.CaddyConfig())
+			assert.Equal(t, tt.wantLabels, d.Spec.Labels)
+		})
+	}
+
+	t.Run("official image needs a config without cluster storage", func(t *testing.T) {
+		c := &CaddyClient{client: &Client{}}
+		_, err := c.NewDeployment(context.Background(), CaddyDeploymentOptions{Image: "caddy:2.11.4"})
+		require.ErrorContains(t, err, "doesn't support 'storage uncloud'")
+
+		_, err = c.NewDeployment(context.Background(), CaddyDeploymentOptions{Image: "caddy:2.11.4", Config: new("")})
+		require.NoError(t, err)
+	})
+}
+
+func TestValidateCaddyImageStorage(t *testing.T) {
+	t.Parallel()
+
+	const storageConfig = "{\n\tstorage uncloud\n}"
+	for _, tt := range []struct {
+		name    string
+		image   string
+		config  string
+		wantErr bool
+	}{
+		{name: "official image with uncloud storage", image: "caddy:2.11.4", config: storageConfig, wantErr: true},
+		{name: "official image with registry", image: "docker.io/library/caddy:2", config: storageConfig, wantErr: true},
+		{name: "official image without config", image: "caddy:2.11.4"},
+		{name: "official image with other config", image: "caddy:2.11.4", config: "{\n\tdebug\n}"},
+		{name: "default image with uncloud storage", image: CaddyImage + ":0.1.3", config: storageConfig},
+		{name: "custom image with uncloud storage", image: "example.com/caddy:1", config: storageConfig},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			err := validateCaddyImageStorage(tt.image, tt.config)
+			if tt.wantErr {
+				require.ErrorContains(t, err, "doesn't support 'storage uncloud'")
+			} else {
+				require.NoError(t, err)
+			}
+		})
+	}
+}
+
+func TestOutdatedCaddyDaemons(t *testing.T) {
+	t.Parallel()
+
+	machine := func(name, version string) *pb.MachineMember {
+		return &pb.MachineMember{Machine: &pb.MachineInfo{Name: name, DaemonVersion: version}}
+	}
+
+	assert.Equal(t, []string{"old (0.21.0)", "older (0.20.0)", "unsupported (0.9.3)"},
+		outdatedCaddyDaemons(api.MachineMembersList{
+			machine("old", "0.21.0"),
+			machine("older", "0.20.0"),
+			machine("unsupported", "0.9.3"),
+			machine("no-version", ""),
+			machine("required", "0.22.0"),
+			machine("newer", "0.23.1"),
+			machine("major", "1.0.0"),
+			machine("nightly", "0.22.0-nightly-abc1234"),
+			machine("old-dev-build", "v0.21.0-5-g6c7ddeb-dirty"),
+			machine("local-build", "999.0.0-dev"),
+			machine("not-semver", "6c7ddeb"),
+		}))
+	assert.Empty(t, outdatedCaddyDaemons(api.MachineMembersList{machine("required", "0.22.0")}))
+	assert.Empty(t, outdatedCaddyDaemons(nil))
+}

pkg/client/client.go

@@ -86,6 +86,7 @@ func (cli *Client) progressOut() *streams.Out {
 // ProxyMachinesContext returns a new context that proxies gRPC requests to the specified machines.
 // If namesOrIDs is nil or empty, all machines are included.
 // This triggers One2Many proxying, which always injects metadata into the response.
+// TODO: validate namesOrIDs can't contain "*" as this is reserved as internal value and may break the caller logic.
 func ProxyMachinesContext(ctx context.Context, namesOrIDs []string) context.Context {
 	md := outgoingMetadataWithoutProxyTargets(ctx)
 	if len(namesOrIDs) == 0 {

pkg/client/compose/deploy.go

@@ -13,8 +13,8 @@ import (
 	"github.com/docker/docker/api/types/volume"
 	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/psviderski/uncloud/pkg/client/deploy"
-	"github.com/psviderski/uncloud/pkg/client/deploy/operation"
 	"github.com/psviderski/uncloud/pkg/client/deploy/scheduler"
+	"github.com/psviderski/uncloud/pkg/client/operation"
 )
 
 type Client interface {

pkg/client/compose/plan.go

@@ -10,7 +10,7 @@ import (
 	"github.com/psviderski/uncloud/internal/cli/tui"
 	"github.com/psviderski/uncloud/pkg/api"
 	"github.com/psviderski/uncloud/pkg/client/deploy"
-	"github.com/psviderski/uncloud/pkg/client/deploy/operation"
+	"github.com/psviderski/uncloud/pkg/client/operation"
 )
 
 // Plan holds the compose-level deployment plan with volume and service operations.
@@ -55,7 +55,7 @@ func (p *Plan) Format() string {
 
 // formatSummary counts all operations across the plan and renders the summary footer.
 func (p *Plan) formatSummary() string {
-	var createCount, startFirstCount, stopFirstCount, removeCount int
+	var createCount, startFirstCount, stopFirstCount, updateCount, removeCount int
 	machines := make(map[string]struct{})
 
 	for _, op := range p.Volumes {
@@ -76,12 +76,15 @@ func (p *Plan) formatSummary() string {
 				} else {
 					startFirstCount++
 				}
+			case *operation.UpdateContainerOperation:
+				machines[o.MachineID] = struct{}{}
+				updateCount++
 			case *operation.RemoveContainerOperation:
 				machines[o.MachineID] = struct{}{}
 				removeCount++
 			case *operation.StopContainerOperation:
+				// Stop operations aren't counted because a later operation in the plan removes the stopped container.
 				machines[o.MachineID] = struct{}{}
-				removeCount++
 			}
 		}
 	}
@@ -99,6 +102,10 @@ func (p *Plan) formatSummary() string {
 		parts = append(parts,
 			tui.BoldYellow.Render(strconv.Itoa(stopFirstCount))+" "+tui.Yellow.Render("replace (stop-first)"))
 	}
+	if updateCount > 0 {
+		parts = append(parts,
+			tui.BoldYellow.Render(strconv.Itoa(updateCount))+" "+tui.Yellow.Render("update (in place)"))
+	}
 	if removeCount > 0 {
 		parts = append(parts,
 			tui.BoldRed.Render(strconv.Itoa(removeCount))+" "+tui.Red.Render("remove"))

pkg/client/compose/plan_test.go

@@ -0,0 +1,47 @@
+package compose
+
+import (
+	"testing"
+
+	"github.com/charmbracelet/x/ansi"
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/psviderski/uncloud/pkg/client/deploy"
+	"github.com/psviderski/uncloud/pkg/client/operation"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestPlan_FormatSummary(t *testing.T) {
+	t.Parallel()
+
+	web := &deploy.ServicePlan{}
+	web.Operations = []operation.Operation{
+		&operation.UpdateContainerOperation{MachineID: "m-1"},
+		&operation.RemoveContainerOperation{MachineID: "m-2"},
+	}
+	db := &deploy.ServicePlan{}
+	db.Operations = []operation.Operation{
+		&operation.UpdateContainerOperation{MachineID: "m-2"},
+	}
+	plan := &Plan{
+		Volumes:  []*operation.CreateVolumeOperation{{MachineID: "m-3"}},
+		Services: []*deploy.ServicePlan{web, db},
+	}
+
+	assert.Equal(t, "1 create · 2 update (in place) · 1 remove · across 3 machines", ansi.Strip(plan.formatSummary()))
+}
+
+func TestPlan_FormatSummary_StopOperationsNotCounted(t *testing.T) {
+	t.Parallel()
+
+	svc := &deploy.ServicePlan{}
+	svc.Operations = []operation.Operation{
+		&operation.StopContainerOperation{MachineID: "m-1"},
+		&operation.StopContainerOperation{MachineID: "m-1"},
+		&operation.ReplaceContainerOperation{MachineID: "m-1", Order: api.UpdateOrderStopFirst},
+		&operation.RemoveContainerOperation{MachineID: "m-1"},
+	}
+	plan := &Plan{Services: []*deploy.ServicePlan{svc}}
+
+	// Stop operations aren't counted because a later operation in the plan removes the stopped container.
+	assert.Equal(t, "1 replace (stop-first) · 1 remove · across 1 machine", ansi.Strip(plan.formatSummary()))
+}

pkg/client/compose/project_test.go

@@ -344,20 +344,6 @@ configs:
 			warnCount:    1,
 			warnContains: []string{"container_name"},
 		},
-		{
-			name: "deploy labels",
-			composeYAML: `services:
-  app:
-    image: myapp:latest
-    deploy:
-      mode: replicated
-      labels:
-        - foo
-      replicas: 2
-`,
-			warnCount:    1,
-			warnContains: []string{"deploy labels"},
-		},
 		{
 			name: "deploy restart_policy",
 			composeYAML: `services:

pkg/client/compose/secret.go

@@ -89,9 +89,28 @@ func ResolveSecrets(ctx context.Context, project *types.Project) error {
 		return v, nil
 	}
 
-	// Resolve only secret references set as values for environment variables in enabled services.
 	for _, service := range project.Services {
-		for k, v := range service.Environment {
+		// Resolve secret references set as values for environment variables in the service.
+		if err := resolveEnvSecrets(service.Environment, resolve); err != nil {
+			return err
+		}
+
+		// Resolve secret references set as values for the pre-deploy hook's environment variables, if any.
+		if hook, ok := service.Extensions[PreDeployHookExtensionKey].(PreDeployHook); ok {
+			if err := resolveEnvSecrets(hook.Environment, resolve); err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
+}
+
+// resolveEnvSecrets resolves 'secret://name' references found as values in env to actual secret values using
+// resolve, setting each referenced variable to the secret's value in place. Variables that are not secret
+// references are left untouched.
+func resolveEnvSecrets(env types.MappingWithEquals, resolve func(string) (string, error)) error {
+	for k, v := range env {
 		if v == nil {
 			continue
 		}
@@ -103,18 +122,31 @@ func ResolveSecrets(ctx context.Context, project *types.Project) error {
 		if err != nil {
 			return err
 		}
-			service.Environment[k] = &value
+		env[k] = &value
 	}
-	}
-
 	return nil
 }
 
-// HasCommandSecretRefs reports whether any service environment references a secret that is resolved by running
+// HasCommandSecretRefs reports whether any service environment, including a pre-deploy hook's environment,
-// a command.
+// references a secret that is resolved by running a command.
 func HasCommandSecretRefs(project *types.Project) bool {
 	for _, service := range project.Services {
-		for _, v := range service.Environment {
+		if envHasCommandSecretRef(service.Environment, project) {
+			return true
+		}
+		if hook, ok := service.Extensions[PreDeployHookExtensionKey].(PreDeployHook); ok {
+			if envHasCommandSecretRef(hook.Environment, project) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// envHasCommandSecretRef reports whether env contains a 'secret://name' reference to a secret resolved
+// by running a command.
+func envHasCommandSecretRef(env types.MappingWithEquals, project *types.Project) bool {
+	for _, v := range env {
 		if v == nil {
 			continue
 		}
@@ -126,7 +158,6 @@ func HasCommandSecretRefs(project *types.Project) bool {
 			return true
 		}
 	}
-	}
 	return false
 }
 

pkg/client/compose/secret_test.go

@@ -328,6 +328,74 @@ secrets:
 	assert.Equal(t, "run\n", string(runs), "command should run exactly once across services and repeated resolutions")
 }
 
+// TestResolveSecrets_PreDeployHook pins the behaviour that 'secret://name' references in the pre-deploy hook's
+// environment are resolved just like references in the service's own environment.
+// See https://github.com/psviderski/uncloud/issues/422.
+func TestResolveSecrets_PreDeployHook(t *testing.T) {
+	t.Parallel()
+
+	content := `
+services:
+  foo:
+    image: foo
+    environment:
+      CONTROL_TOKEN: secret://control_token
+    x-pre_deploy:
+      command: ["./migrate.sh"]
+      environment:
+        HOOK_TOKEN: secret://hook_token
+        PLAIN: hello
+secrets:
+  control_token:
+    x-command: printf control-value
+  hook_token:
+    x-command: printf hook-value
+`
+	project := loadProject(t, content)
+	require.NoError(t, ResolveSecrets(context.Background(), project))
+
+	assert.Equal(t, env(map[string]string{"CONTROL_TOKEN": "control-value"}), project.Services["foo"].Environment)
+
+	hook, ok := project.Services["foo"].Extensions[PreDeployHookExtensionKey].(PreDeployHook)
+	require.True(t, ok, "x-pre_deploy extension not found")
+	assert.Equal(t, env(map[string]string{"HOOK_TOKEN": "hook-value", "PLAIN": "hello"}), hook.Environment)
+}
+
+// TestResolveSecrets_PreDeployHook_ResolvedOnce verifies that a secret referenced from both the service's
+// environment and its pre-deploy hook's environment is only resolved once.
+func TestResolveSecrets_PreDeployHook_ResolvedOnce(t *testing.T) {
+	t.Parallel()
+
+	counter := filepath.Join(t.TempDir(), "runs")
+	content := fmt.Sprintf(`
+services:
+  foo:
+    image: foo
+    environment:
+      TOKEN: secret://token
+    x-pre_deploy:
+      command: ["./migrate.sh"]
+      environment:
+        TOKEN: secret://token
+secrets:
+  token:
+    x-command: "sh -c 'echo run >> %s; printf abc'"
+`, counter)
+
+	project := loadProject(t, content)
+	require.NoError(t, ResolveSecrets(context.Background(), project))
+
+	assert.Equal(t, "abc", *project.Services["foo"].Environment["TOKEN"])
+	hook, ok := project.Services["foo"].Extensions[PreDeployHookExtensionKey].(PreDeployHook)
+	require.True(t, ok, "x-pre_deploy extension not found")
+	assert.Equal(t, "abc", *hook.Environment["TOKEN"])
+
+	runs, err := os.ReadFile(counter)
+	require.NoError(t, err)
+	assert.Equal(t, "run\n", string(runs),
+		"command should run exactly once across the service and hook environments")
+}
+
 func TestHasCommandSecretRefs(t *testing.T) {
 	t.Parallel()
 
@@ -387,6 +455,22 @@ services:
 `,
 			want: false,
 		},
+		{
+			name: "command secret referenced from pre-deploy hook",
+			content: `
+services:
+  foo:
+    image: foo
+    x-pre_deploy:
+      command: ["./migrate.sh"]
+      environment:
+        TOKEN: secret://token
+secrets:
+  token:
+    x-command: printf abc
+`,
+			want: true,
+		},
 	}
 
 	for _, tt := range tests {

pkg/client/compose/service.go

@@ -66,6 +66,10 @@ func ServiceSpecFromCompose(project *types.Project, serviceName string) (api.Ser
 		Mode: api.ServiceModeReplicated,
 	}
 
+	if labels := mergeLabels(service.Labels, service.CustomLabels); len(labels) > 0 {
+		spec.Container.Labels = labels
+	}
+
 	// Map x-caddy extension to spec.Caddy if specified.
 	if caddy, ok := service.Extensions[CaddyExtensionKey].(Caddy); ok && caddy.Config != "" {
 		spec.Caddy = &api.CaddySpec{
@@ -124,6 +128,13 @@ func ServiceSpecFromCompose(project *types.Project, serviceName string) (api.Ser
 			d := time.Duration(cfg.Monitor)
 			spec.UpdateConfig.MonitorPeriod = &d
 		}
+
+		if len(service.Deploy.Labels) > 0 {
+			if err = api.ValidateLabels(service.Deploy.Labels); err != nil {
+				return spec, fmt.Errorf("invalid deploy.labels: %w", err)
+			}
+			spec.Labels = maps.Clone(service.Deploy.Labels)
+		}
 	}
 
 	// TODO: can service.tmpfs be handled as tmpfs volume mounts as well?
@@ -480,9 +491,6 @@ func validateServicesFeatures(project *types.Project) []error {
 		if service.DNSSearch != nil {
 			errs = append(errs, err(service.Name, "dns_search"))
 		}
-		if service.Labels != nil {
-			errs = append(errs, err(service.Name, "labels"))
-		}
 		if service.Links != nil {
 			errs = append(errs, err(service.Name, "links"))
 		}
@@ -517,9 +525,6 @@ func validateServicesFeatures(project *types.Project) []error {
 			if service.Deploy.RollbackConfig != nil {
 				errs = append(errs, err(service.Name, "deploy rollback_config"))
 			}
-			if len(service.Deploy.Labels) != 0 {
-				errs = append(errs, err(service.Name, "deploy labels"))
-			}
 		}
 		// we only allow the 'default' network, nothing else.
 		if x := service.Networks; x != nil {

pkg/client/compose/service_test.go

@@ -875,6 +875,117 @@ services:
 	}
 }
 
+func TestServiceSpecFromCompose_Labels(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name                    string
+		composeYAML             string
+		expectedContainerLabels map[string]string
+		expectedServiceLabels   map[string]string
+	}{
+		{
+			name: "no labels",
+			composeYAML: `
+services:
+  web:
+    image: nginx
+`,
+		},
+		{
+			name: "container labels as list",
+			composeYAML: `
+services:
+  web:
+    image: nginx
+    labels:
+      - com.example.team=web
+      - com.example.empty
+`,
+			expectedContainerLabels: map[string]string{
+				"com.example.team":  "web",
+				"com.example.empty": "",
+			},
+		},
+		{
+			name: "container and service labels as maps",
+			composeYAML: `
+services:
+  web:
+    image: nginx
+    labels:
+      com.example.team: web
+    deploy:
+      labels:
+        tier: frontend
+`,
+			expectedContainerLabels: map[string]string{"com.example.team": "web"},
+			expectedServiceLabels:   map[string]string{"tier": "frontend"},
+		},
+		{
+			name: "service labels only",
+			composeYAML: `
+services:
+  web:
+    image: nginx
+    deploy:
+      labels:
+        - tier=frontend
+`,
+			expectedServiceLabels: map[string]string{"tier": "frontend"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			project, err := LoadProjectFromContent(context.Background(), tt.composeYAML)
+			require.NoError(t, err)
+
+			spec, err := ServiceSpecFromCompose(project, "web")
+			require.NoError(t, err)
+
+			assert.Equal(t, tt.expectedContainerLabels, spec.Container.Labels)
+			assert.Equal(t, tt.expectedServiceLabels, spec.Labels)
+		})
+	}
+}
+
+func TestServiceSpecFromCompose_InvalidServiceLabels(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		label   string
+		wantErr string
+	}{
+		{name: "reserved prefix", label: api.LabelManaged, wantErr: "reserved prefix"},
+		{name: "reserved daemon prefix", label: "uncloudd.foo", wantErr: "reserved prefix"},
+		{name: "managed service", label: api.LabelReservedPrefix + "service.managed", wantErr: "reserved prefix"},
+		{name: "whitespace in key", label: "my label", wantErr: "must not contain whitespace"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			composeYAML := `
+services:
+  web:
+    image: nginx
+    deploy:
+      labels:
+        - ` + tt.label + `=value
+`
+			project, err := LoadProjectFromContent(context.Background(), composeYAML)
+			require.NoError(t, err)
+
+			_, err = ServiceSpecFromCompose(project, "web")
+			require.ErrorContains(t, err, "invalid deploy.labels")
+			require.ErrorContains(t, err, tt.wantErr)
+		})
+	}
+}
+
 func TestServiceSpecFromCompose_UpdateConfig(t *testing.T) {
 	t.Parallel()
 

pkg/client/container.go

@@ -343,6 +343,39 @@ func (cli *Client) RemoveContainer(
 	return nil
 }
 
+// UpdateContainer updates the requested properties of the container within the service in place without recreating it.
+func (cli *Client) UpdateContainer(
+	ctx context.Context, serviceNameOrID, containerNameOrID string, opts api.UpdateContainerOptions,
+) error {
+	op, err := cli.resolveContainerOperation(ctx, serviceNameOrID, containerNameOrID)
+	if err != nil {
+		return err
+	}
+
+	req := &pb.UpdateServiceContainerRequest{
+		Id:                op.containerID,
+		Cpu:               opts.CPU,
+		Memory:            opts.Memory,
+		MemoryReservation: opts.MemoryReservation,
+	}
+	if opts.ServiceLabels != nil {
+		req.ServiceLabels = &pb.Labels{Labels: *opts.ServiceLabels}
+	}
+
+	pw := progress.ContextWriter(op.ctx)
+	pw.Event(progress.NewEvent(op.eventID, progress.Working, "Updating"))
+	if _, err = cli.Docker.GRPCClient.UpdateServiceContainer(op.ctx, req); err != nil {
+		pw.Event(progress.ErrorEvent(op.eventID))
+		if status.Code(err) == codes.NotFound {
+			return api.ErrNotFound
+		}
+		return err
+	}
+	pw.Event(progress.NewEvent(op.eventID, progress.Done, "Updated"))
+
+	return nil
+}
+
 // ExecContainer executes a command in a container within the service.
 // If containerNameOrID is empty, the first container in the service will be used.
 func (cli *Client) ExecContainer(

pkg/client/deploy/container.go

@@ -1,7 +1,7 @@
 package deploy
 
 import (
-	"reflect"
+	"maps"
 	"sort"
 
 	"github.com/google/go-cmp/cmp"
@@ -13,6 +13,7 @@ type ContainerSpecStatus string
 
 const (
 	ContainerUpToDate ContainerSpecStatus = "up-to-date"
+	// ContainerNeedsUpdate means the container can be updated in place without recreating it.
 	ContainerNeedsUpdate   ContainerSpecStatus = "needs-update"
 	ContainerNeedsRecreate ContainerSpecStatus = "needs-recreate"
 )
@@ -83,15 +84,7 @@ func EvalContainerSpecChange(current api.ServiceSpec, new api.ServiceSpec) Conta
 		}
 	}
 
-	// Device reservations and mappings are immutable, so we'll need to recreate if any have changed
+	if current.Container.Resources.ValidateUpdateInPlace(newResources) != nil {
-	if !reflect.DeepEqual(current.Container.Resources.DeviceReservations, newResources.DeviceReservations) {
-		return ContainerNeedsRecreate
-	}
-	if !reflect.DeepEqual(current.Container.Resources.Devices, newResources.Devices) {
-		return ContainerNeedsRecreate
-	}
-	// Ulimits are immutable, so we'll need to recreate if any have changed.
-	if !reflect.DeepEqual(current.Container.Resources.Ulimits, newResources.Ulimits) {
 		return ContainerNeedsRecreate
 	}
 
@@ -100,8 +93,10 @@ func EvalContainerSpecChange(current api.ServiceSpec, new api.ServiceSpec) Conta
 		return ContainerNeedsRecreate
 	}
 
-	// Remaining resources are mutable.
+	// CPU and memory limits are updated on the running container. Service labels are only stored in the service spec
-	if !reflect.DeepEqual(current.Container.Resources, newResources) {
+	// and not applied to Docker containers.
+	if !current.Container.Resources.Equals(newResources) ||
+		!cmp.Equal(current.Labels, new.Labels, cmpopts.EquateEmpty()) {
 		return ContainerNeedsUpdate
 	}
 
@@ -119,3 +114,53 @@ func sortConfigs(configs []api.ConfigSpec) {
 		return configs[i].Name < configs[j].Name
 	})
 }
+
+// EvalContainerChange evaluates whether the running container matches the new spec, can be updated in place to
+// match it, or needs to be recreated. Unlike EvalContainerSpecChange, it also compares the new spec with the CPU and
+// memory limits that Docker applies to the container, so the container is updated when they drifted from its spec.
+func EvalContainerChange(current api.ServiceContainer, new api.ServiceSpec) ContainerSpecStatus {
+	status := EvalContainerSpecChange(current.ServiceSpec, new)
+	if status == ContainerNeedsRecreate {
+		return status
+	}
+
+	applied := current.AppliedResources()
+	newResources := new.Container.Resources
+	if applied.ValidateUpdateInPlace(newResources) != nil {
+		return ContainerNeedsRecreate
+	}
+	// Only CPU and memory limits can differ after the in-place update check above.
+	if !applied.Equals(newResources) {
+		return ContainerNeedsUpdate
+	}
+	return status
+}
+
+// ContainerUpdateOptions returns the container properties that can be updated in place and differ between the new
+// spec and either the container spec or the limits Docker applies to the container. It should only be used when
+// EvalContainerChange returns ContainerNeedsUpdate.
+func ContainerUpdateOptions(current api.ServiceContainer, new api.ServiceSpec) api.UpdateContainerOptions {
+	var opts api.UpdateContainerOptions
+
+	stored, applied := current.ServiceSpec.Container.Resources, current.AppliedResources()
+	res := new.Container.Resources
+	if res.CPU != stored.CPU || res.CPU != applied.CPU {
+		opts.CPU = &res.CPU
+	}
+	if res.Memory != stored.Memory || res.Memory != applied.Memory {
+		opts.Memory = &res.Memory
+	}
+	if res.MemoryReservation != stored.MemoryReservation || res.MemoryReservation != applied.MemoryReservation {
+		opts.MemoryReservation = &res.MemoryReservation
+	}
+
+	if !cmp.Equal(current.ServiceSpec.Labels, new.Labels, cmpopts.EquateEmpty()) {
+		labels := maps.Clone(new.Labels)
+		if labels == nil {
+			labels = map[string]string{}
+		}
+		opts.ServiceLabels = &labels
+	}
+
+	return opts
+}

pkg/client/deploy/container_test.go

@@ -86,6 +86,223 @@ func TestEvalContainerSpecChange_ContainerTty(t *testing.T) {
 	assert.Equal(t, ContainerNeedsRecreate, EvalContainerSpecChange(newSpec, currentSpec))
 }
 
+func TestEvalContainerSpecChange_ContainerLabels(t *testing.T) {
+	t.Parallel()
+
+	currentSpec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image: "nginx:latest",
+		},
+	}
+	newSpec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image:  "nginx:latest",
+			Labels: map[string]string{"com.example.team": "web"},
+		},
+	}
+	emptySpec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image:  "nginx:latest",
+			Labels: map[string]string{},
+		},
+	}
+
+	assert.Equal(t, ContainerNeedsRecreate, EvalContainerSpecChange(currentSpec, newSpec))
+	assert.Equal(t, ContainerNeedsRecreate, EvalContainerSpecChange(newSpec, currentSpec))
+	assert.Equal(t, ContainerUpToDate, EvalContainerSpecChange(newSpec, newSpec))
+	assert.Equal(t, ContainerUpToDate, EvalContainerSpecChange(currentSpec, emptySpec))
+}
+
+func TestEvalContainerSpecChange_ServiceLabels(t *testing.T) {
+	t.Parallel()
+
+	currentSpec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image: "nginx:latest",
+		},
+	}
+	newSpec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image: "nginx:latest",
+		},
+		Labels: map[string]string{"tier": "frontend"},
+	}
+	emptySpec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image: "nginx:latest",
+		},
+		Labels: map[string]string{},
+	}
+
+	assert.Equal(t, ContainerNeedsUpdate, EvalContainerSpecChange(currentSpec, newSpec))
+	assert.Equal(t, ContainerNeedsUpdate, EvalContainerSpecChange(newSpec, currentSpec))
+	assert.Equal(t, ContainerUpToDate, EvalContainerSpecChange(newSpec, newSpec))
+	assert.Equal(t, ContainerUpToDate, EvalContainerSpecChange(currentSpec, emptySpec))
+
+	// Changes that require recreating the container take precedence over the in-place update.
+	newSpecWithImage := newSpec.Clone()
+	newSpecWithImage.Container.Image = "nginx:alpine"
+	assert.Equal(t, ContainerNeedsRecreate, EvalContainerSpecChange(currentSpec, newSpecWithImage))
+}
+
+func TestEvalContainerChange(t *testing.T) {
+	t.Parallel()
+
+	spec := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image:     "nginx:latest",
+			Resources: api.ContainerResources{CPU: api.Core, Memory: 256 * 1024 * 1024},
+		},
+	}
+	appliedSpec := container.Resources{NanoCPUs: api.Core, Memory: 256 * 1024 * 1024}
+
+	tests := []struct {
+		name       string
+		hostConfig *container.HostConfig
+		change     func(spec *api.ServiceSpec)
+		expected   ContainerSpecStatus
+	}{
+		{
+			name:       "applied limits match spec",
+			hostConfig: &container.HostConfig{Resources: appliedSpec},
+			change:     func(*api.ServiceSpec) {},
+			expected:   ContainerUpToDate,
+		},
+		{
+			name:     "unknown host config falls back to spec limits",
+			change:   func(*api.ServiceSpec) {},
+			expected: ContainerUpToDate,
+		},
+		{
+			name:       "applied memory limit drifted from spec",
+			hostConfig: &container.HostConfig{Resources: container.Resources{NanoCPUs: api.Core, Memory: 512 * 1024 * 1024}},
+			change:     func(*api.ServiceSpec) {},
+			expected:   ContainerNeedsUpdate,
+		},
+		{
+			name:       "applied limits already match new spec but stored spec doesn't",
+			hostConfig: &container.HostConfig{Resources: container.Resources{NanoCPUs: api.Core, Memory: 512 * 1024 * 1024}},
+			change:     func(spec *api.ServiceSpec) { spec.Container.Resources.Memory = 512 * 1024 * 1024 },
+			expected:   ContainerNeedsUpdate,
+		},
+		{
+			name: "applied limit not in spec can't be removed in place",
+			hostConfig: &container.HostConfig{Resources: container.Resources{
+				NanoCPUs: api.Core, Memory: 256 * 1024 * 1024, MemoryReservation: 128 * 1024 * 1024,
+			}},
+			change:   func(*api.ServiceSpec) {},
+			expected: ContainerNeedsRecreate,
+		},
+		{
+			name:       "spec change that requires recreating",
+			hostConfig: &container.HostConfig{Resources: appliedSpec},
+			change:     func(spec *api.ServiceSpec) { spec.Container.Image = "nginx:alpine" },
+			expected:   ContainerNeedsRecreate,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			newSpec := spec.Clone()
+			tt.change(&newSpec)
+			assert.Equal(t, tt.expected, EvalContainerChange(serviceContainer(spec, tt.hostConfig), newSpec))
+		})
+	}
+}
+
+func TestContainerUpdateOptions(t *testing.T) {
+	t.Parallel()
+
+	current := api.ServiceSpec{
+		Container: api.ContainerSpec{
+			Image:     "nginx:latest",
+			Resources: api.ContainerResources{CPU: api.Core, Memory: 256 * 1024 * 1024},
+		},
+		Labels: map[string]string{"tier": "frontend"},
+	}
+	applied := &container.HostConfig{Resources: container.Resources{NanoCPUs: api.Core, Memory: 256 * 1024 * 1024}}
+
+	tests := []struct {
+		name       string
+		hostConfig *container.HostConfig
+		change     func(spec *api.ServiceSpec)
+		expected   api.UpdateContainerOptions
+	}{
+		{
+			name:       "no changes",
+			hostConfig: applied,
+			change:     func(*api.ServiceSpec) {},
+		},
+		{
+			name:       "CPU only",
+			hostConfig: applied,
+			change:     func(spec *api.ServiceSpec) { spec.Container.Resources.CPU = 2 * api.Core },
+			expected:   api.UpdateContainerOptions{CPU: new(int64(2 * api.Core))},
+		},
+		{
+			name:       "memory and reservation",
+			hostConfig: applied,
+			change: func(spec *api.ServiceSpec) {
+				spec.Container.Resources.Memory = 512 * 1024 * 1024
+				spec.Container.Resources.MemoryReservation = 128 * 1024 * 1024
+			},
+			expected: api.UpdateContainerOptions{
+				Memory:            new(int64(512 * 1024 * 1024)),
+				MemoryReservation: new(int64(128 * 1024 * 1024)),
+			},
+		},
+		{
+			name: "applied memory limit drifted from spec",
+			hostConfig: &container.HostConfig{Resources: container.Resources{
+				NanoCPUs: api.Core, Memory: 512 * 1024 * 1024,
+			}},
+			change:   func(*api.ServiceSpec) {},
+			expected: api.UpdateContainerOptions{Memory: new(int64(256 * 1024 * 1024))},
+		},
+		{
+			name: "applied memory limit already matches new spec",
+			hostConfig: &container.HostConfig{Resources: container.Resources{
+				NanoCPUs: api.Core, Memory: 512 * 1024 * 1024,
+			}},
+			change:   func(spec *api.ServiceSpec) { spec.Container.Resources.Memory = 512 * 1024 * 1024 },
+			expected: api.UpdateContainerOptions{Memory: new(int64(512 * 1024 * 1024))},
+		},
+		{
+			name:       "change service labels",
+			hostConfig: applied,
+			change:     func(spec *api.ServiceSpec) { spec.Labels = map[string]string{"tier": "backend"} },
+			expected:   api.UpdateContainerOptions{ServiceLabels: &map[string]string{"tier": "backend"}},
+		},
+		{
+			name:       "remove all service labels",
+			hostConfig: applied,
+			change:     func(spec *api.ServiceSpec) { spec.Labels = nil },
+			expected:   api.UpdateContainerOptions{ServiceLabels: &map[string]string{}},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			newSpec := current.Clone()
+			tt.change(&newSpec)
+			assert.Equal(t, tt.expected, ContainerUpdateOptions(serviceContainer(current, tt.hostConfig), newSpec))
+		})
+	}
+}
+
+func serviceContainer(spec api.ServiceSpec, hostConfig *container.HostConfig) api.ServiceContainer {
+	return api.ServiceContainer{
+		Container: api.Container{InspectResponse: container.InspectResponse{
+			ContainerJSONBase: &container.ContainerJSONBase{HostConfig: hostConfig},
+		}},
+		ServiceSpec: spec,
+	}
+}
+
 func TestEvalContainerSpecChange_ContainerResources(t *testing.T) {
 	t.Parallel()
 
@@ -127,7 +344,7 @@ func TestEvalContainerSpecChange_ContainerResources(t *testing.T) {
 				CPU: 1000000000,
 			},
 			new:  api.ContainerResources{},
-			want: ContainerNeedsUpdate,
+			want: ContainerNeedsRecreate,
 		},
 
 		// Memory
@@ -189,19 +406,33 @@ func TestEvalContainerSpecChange_ContainerResources(t *testing.T) {
 			name:    "unset memory",
 			current: api.ContainerResources{Memory: 100 * 1024 * 1024, MemoryReservation: 50 * 1024 * 1024},
 			new:     api.ContainerResources{MemoryReservation: 50 * 1024 * 1024},
-			want:    ContainerNeedsUpdate,
+			want:    ContainerNeedsRecreate,
 		},
 		{
 			name:    "unset reservation",
 			current: api.ContainerResources{Memory: 100 * 1024 * 1024, MemoryReservation: 50 * 1024 * 1024},
 			new:     api.ContainerResources{Memory: 100 * 1024 * 1024},
-			want:    ContainerNeedsUpdate,
+			want:    ContainerNeedsRecreate,
 		},
 		{
 			name:    "unset memory and reservation",
 			current: api.ContainerResources{Memory: 100 * 1024 * 1024, MemoryReservation: 50 * 1024 * 1024},
 			new:     api.ContainerResources{},
-			want:    ContainerNeedsUpdate,
+			want:    ContainerNeedsRecreate,
+		},
+
+		// Shared memory
+		{
+			name:    "change shared memory",
+			current: api.ContainerResources{SharedMemory: 64 * 1024 * 1024},
+			new:     api.ContainerResources{SharedMemory: 128 * 1024 * 1024},
+			want:    ContainerNeedsRecreate,
+		},
+		{
+			name:    "change memory and shared memory",
+			current: api.ContainerResources{Memory: 100 * 1024 * 1024},
+			new:     api.ContainerResources{Memory: 200 * 1024 * 1024, SharedMemory: 64 * 1024 * 1024},
+			want:    ContainerNeedsRecreate,
 		},
 
 		// Memory and CPU

pkg/client/deploy/deploy.go

@@ -1,19 +1,23 @@
 package deploy
 
 import (
+	"cmp"
 	"context"
 	"errors"
 	"fmt"
+	"maps"
+	"slices"
 	"strconv"
 	"strings"
+	"unicode"
 
 	"charm.land/lipgloss/v2"
 	"charm.land/lipgloss/v2/table"
 	"github.com/distribution/reference"
 	"github.com/psviderski/uncloud/internal/cli/tui"
 	"github.com/psviderski/uncloud/pkg/api"
-	"github.com/psviderski/uncloud/pkg/client/deploy/operation"
 	"github.com/psviderski/uncloud/pkg/client/deploy/scheduler"
+	"github.com/psviderski/uncloud/pkg/client/operation"
 )
 
 type Client interface {
@@ -50,25 +54,31 @@ type ServicePlan struct {
 // Format renders the service plan as a styled block with a spec diff and nested container operations.
 func (sp *ServicePlan) Format() string {
 	// Determine service-level operation type and extract the old spec from container operations.
-	// Assume replace operations precede remove operations (rolling strategy) so the first replace operation
+	// Containers may have different specs after a failed deployment. The first replaced container determines the old
-	// (if exists) determines the old spec for the diff. Otherwise, fallback to the first remove operation.
+	// spec for the diff because a container updated in place already has the new image. Otherwise, fallback to the
+	// first updated container and then the first removed one.
 	var hasRun, hasRemove bool
-	var oldSpec *api.ServiceSpec
+	var replacedSpec, updatedSpec, removedSpec *api.ServiceSpec
 	for _, op := range sp.Operations {
 		switch o := op.(type) {
 		case *operation.RunContainerOperation:
 			hasRun = true
 		case *operation.ReplaceContainerOperation:
-			if oldSpec == nil {
+			if replacedSpec == nil {
-				oldSpec = &o.OldContainer.ServiceSpec
+				replacedSpec = &o.OldContainer.ServiceSpec
+			}
+		case *operation.UpdateContainerOperation:
+			if updatedSpec == nil {
+				updatedSpec = &o.Container.ServiceSpec
 			}
 		case *operation.RemoveContainerOperation:
 			hasRemove = true
-			if oldSpec == nil {
+			if removedSpec == nil {
-				oldSpec = &o.Container.ServiceSpec
+				removedSpec = &o.Container.ServiceSpec
 			}
 		}
 	}
+	oldSpec := cmp.Or(replacedSpec, updatedSpec, removedSpec)
 
 	// Service line modifier and verb.
 	var modifier, verb string
@@ -91,7 +101,7 @@ func (sp *ServicePlan) Format() string {
 	out.WriteString("\n")
 
 	// Build spec diff table: columns are [modifier, attribute, value or change].
-	// TODO: print diff for all changed attributes, not just image and replicas.
+	// TODO: print diff for all changed attributes, not just image, replicas, and labels.
 	//  Consider reusing the logic in EvalContainerSpecChange to return a structured diff.
 	specTable := table.New().
 		Border(lipgloss.Border{}).
@@ -124,6 +134,22 @@ func (sp *ServicePlan) Format() string {
 		specTable.Row(mod, "image:", formatImageDiff(oldSpec.Container.Image, sp.Spec.Container.Image))
 	}
 
+	// An existing service without an old spec only has new containers added to up-to-date ones so its Caddy config and
+	// labels didn't change.
+	specChanged := sp.IsNewService || oldSpec != nil
+	var old api.ServiceSpec
+	if oldSpec != nil {
+		old = *oldSpec
+	}
+
+	// Caddy config row.
+	if specChanged {
+		mod, value := formatCaddyConfigDiff(old.CaddyConfig(), sp.Spec.CaddyConfig(), sp.IsNewService)
+		if value != "" {
+			specTable.Row(mod, "caddy config:", value)
+		}
+	}
+
 	// Replicas row for replicated services.
 	if sp.Spec.Mode == api.ServiceModeReplicated {
 		replicasStr := fmt.Sprintf("%d", sp.Spec.Replicas)
@@ -141,17 +167,19 @@ func (sp *ServicePlan) Format() string {
 		}
 	}
 
-	// Stack "  │ " tree prefixes vertically, then join horizontally with the table.
+	specDiff := specTable.String() + "\n"
-	tableStr := specTable.String()
-	treePrefix := tui.Faint.Render("  │ ")
-	treeColRows := make([]string, specTable.GetData().Rows())
-	for i := range treeColRows {
-		treeColRows[i] = treePrefix
-	}
-	treeCol := lipgloss.JoinVertical(lipgloss.Left, treeColRows...)
 
-	out.WriteString(lipgloss.JoinHorizontal(lipgloss.Top, treeCol, tableStr))
+	// Render the labels outside the table because they span multiple lines.
-	out.WriteString("\n")
+	if specChanged {
+		specDiff += formatLabelsDiff("labels:", old.Labels, sp.Spec.Labels, sp.IsNewService)
+		specDiff += formatLabelsDiff("container labels:", old.Container.Labels, sp.Spec.Container.Labels,
+			sp.IsNewService)
+	}
+
+	treePrefix := tui.Faint.Render("  │ ")
+	for line := range strings.Lines(specDiff) {
+		out.WriteString(treePrefix + line)
+	}
 
 	// Blank separator line before container operations.
 	out.WriteString(tui.Faint.Render("  │"))
@@ -173,7 +201,7 @@ func (sp *ServicePlan) Format() string {
 
 // FormatSummary counts operations in the service plan and renders a styled summary line.
 func (sp *ServicePlan) FormatSummary() string {
-	var createCount, startFirstCount, stopFirstCount, removeCount int
+	var createCount, startFirstCount, stopFirstCount, updateCount, removeCount int
 	machines := make(map[string]struct{})
 
 	for _, op := range sp.Operations {
@@ -188,12 +216,15 @@ func (sp *ServicePlan) FormatSummary() string {
 			} else {
 				startFirstCount++
 			}
+		case *operation.UpdateContainerOperation:
+			machines[o.MachineID] = struct{}{}
+			updateCount++
 		case *operation.RemoveContainerOperation:
 			machines[o.MachineID] = struct{}{}
 			removeCount++
 		case *operation.StopContainerOperation:
+			// Stop operations aren't counted because a later operation in the plan removes the stopped container.
 			machines[o.MachineID] = struct{}{}
-			removeCount++
 		}
 	}
 
@@ -210,6 +241,10 @@ func (sp *ServicePlan) FormatSummary() string {
 		parts = append(parts,
 			tui.BoldYellow.Render(strconv.Itoa(stopFirstCount))+" "+tui.Yellow.Render("replace (stop-first)"))
 	}
+	if updateCount > 0 {
+		parts = append(parts,
+			tui.BoldYellow.Render(strconv.Itoa(updateCount))+" "+tui.Yellow.Render("update (in place)"))
+	}
 	if removeCount > 0 {
 		parts = append(parts,
 			tui.BoldRed.Render(strconv.Itoa(removeCount))+" "+tui.Red.Render("remove"))
@@ -265,6 +300,127 @@ func formatImageDiff(oldImage, newImage string) string {
 		tui.FormatImage(newImage, tui.Green)
 }
 
+// formatCaddyConfigDiff returns the modifier and the number of added and removed lines in the Caddy config. It omits
+// the config content because it can be large. It returns an empty value if the configs are the same.
+func formatCaddyConfigDiff(oldConfig, newConfig string, isNewService bool) (mod, value string) {
+	if oldConfig == newConfig {
+		return "", ""
+	}
+
+	// Count lines regardless of their order, which is enough for a summary and doesn't require a diff algorithm.
+	remaining := make(map[string]int)
+	var removed int
+	if oldConfig != "" {
+		for line := range strings.SplitSeq(oldConfig, "\n") {
+			remaining[line]++
+			removed++
+		}
+	}
+	var added int
+	if newConfig != "" {
+		for line := range strings.SplitSeq(newConfig, "\n") {
+			if remaining[line] > 0 {
+				remaining[line]--
+				removed--
+			} else {
+				added++
+			}
+		}
+	}
+
+	var parts []string
+	if added > 0 {
+		parts = append(parts, tui.Green.Render("+"+strconv.Itoa(added)))
+	}
+	if removed > 0 {
+		parts = append(parts, tui.Red.Render("-"+strconv.Itoa(removed)))
+	}
+
+	switch {
+	case added+removed == 0:
+		value = "changed"
+	case added+removed == 1:
+		value = parts[0] + " line"
+	default:
+		value = strings.Join(parts, " ") + " lines"
+	}
+
+	switch {
+	case isNewService:
+		mod = ""
+	case oldConfig == "":
+		mod = tui.Green.Render("+")
+	case newConfig == "":
+		mod = tui.Red.Render("-")
+	default:
+		mod = "~"
+	}
+
+	return mod, value
+}
+
+// formatLabelsDiff renders a block with the attribute name followed by a nested line for each added, changed, or
+// removed label, sorted by key. It returns an empty string if the labels are the same.
+func formatLabelsDiff(attribute string, oldLabels, newLabels map[string]string, isNewService bool) string {
+	keys := slices.Concat(slices.Collect(maps.Keys(oldLabels)), slices.Collect(maps.Keys(newLabels)))
+	slices.Sort(keys)
+
+	var lines strings.Builder
+	for _, key := range slices.Compact(keys) {
+		oldValue, inOld := oldLabels[key]
+		newValue, inNew := newLabels[key]
+
+		var mod, diff string
+		switch {
+		case !inOld:
+			mod = tui.Green.Render("+")
+			// All attributes of a new service are new so only the green color marks them, without a modifier.
+			if isNewService {
+				mod = " "
+			}
+			diff = tui.Green.Render(key+tui.Faint.Render("=")) + tui.Green.Render(formatLabelValue(newValue))
+		case !inNew:
+			mod = tui.Red.Render("-")
+			diff = tui.Red.Render(key+tui.Faint.Render("=")) + tui.Red.Render(formatLabelValue(oldValue))
+		case oldValue != newValue:
+			mod = tui.Yellow.Render("~")
+			diff = key +
+				tui.Faint.Render("=") +
+				tui.Red.Render(formatLabelValue(oldValue)) + " " +
+				tui.Faint.Render("→") + " " +
+				tui.Green.Render(formatLabelValue(newValue))
+		default:
+			continue
+		}
+
+		lines.WriteString("    " + mod + " " + diff + "\n")
+	}
+	if lines.Len() == 0 {
+		return ""
+	}
+
+	blockMod := tui.Yellow.Render("~")
+	switch {
+	case isNewService:
+		blockMod = " "
+	case len(oldLabels) == 0:
+		blockMod = tui.Green.Render("+")
+	case len(newLabels) == 0:
+		blockMod = tui.Red.Render("-")
+	}
+
+	return blockMod + " " + tui.Faint.Render(attribute) + "\n" + lines.String()
+}
+
+// formatLabelValue returns the label value on a single line. It quotes and escapes empty values and values with
+// non-printable characters such as newlines so that they're visible in the diff. Other values are returned as is.
+func formatLabelValue(value string) string {
+	if value == "" || strings.ContainsFunc(value, func(r rune) bool { return !unicode.IsPrint(r) }) {
+		return strconv.Quote(value)
+	}
+	return value
+}
+
 // NewDeployment creates a new deployment for the given service specification.
 // If strategy is nil, a default RollingStrategy will be used.
 func NewDeployment(cli Client, spec api.ServiceSpec, strategy Strategy) *Deployment {

pkg/client/deploy/deploy_test.go

@@ -0,0 +1,406 @@
+package deploy
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/charmbracelet/x/ansi"
+	"github.com/docker/docker/api/types/container"
+	"github.com/psviderski/uncloud/pkg/api"
+	"github.com/psviderski/uncloud/pkg/client/operation"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestServicePlan_FormatSummary(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name     string
+		ops      []operation.Operation
+		expected string
+	}{
+		{
+			name: "in-place updates only",
+			ops: []operation.Operation{
+				&operation.UpdateContainerOperation{MachineID: "m-1"},
+				&operation.UpdateContainerOperation{MachineID: "m-2"},
+			},
+			expected: "2 update (in place) · across 2 machines",
+		},
+		{
+			name: "stop operations not counted as they are followed by remove",
+			ops: []operation.Operation{
+				&operation.StopContainerOperation{MachineID: "m-1"},
+				&operation.StopContainerOperation{MachineID: "m-1"},
+				&operation.ReplaceContainerOperation{MachineID: "m-1", Order: api.UpdateOrderStopFirst},
+				&operation.RemoveContainerOperation{MachineID: "m-1"},
+			},
+			expected: "1 replace (stop-first) · 1 remove · across 1 machine",
+		},
+		{
+			name: "all operation types",
+			ops: []operation.Operation{
+				&operation.RunContainerOperation{MachineID: "m-1"},
+				&operation.ReplaceContainerOperation{MachineID: "m-1", Order: api.UpdateOrderStartFirst},
+				&operation.ReplaceContainerOperation{MachineID: "m-2", Order: api.UpdateOrderStopFirst},
+				&operation.UpdateContainerOperation{MachineID: "m-2"},
+				&operation.RemoveContainerOperation{MachineID: "m-3"},
+			},
+			expected: "1 create · 1 replace (start-first) · 1 replace (stop-first) · 1 update (in place) · 1 remove · " +
+				"across 3 machines",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			plan := ServicePlan{}
+			plan.Operations = tt.ops
+			assert.Equal(t, tt.expected, ansi.Strip(plan.FormatSummary()))
+		})
+	}
+}
+
+func TestServicePlan_Format(t *testing.T) {
+	t.Parallel()
+
+	newContainer := func(spec api.ServiceSpec) api.ServiceContainer {
+		return api.ServiceContainer{
+			Container: api.Container{InspectResponse: container.InspectResponse{
+				ContainerJSONBase: &container.ContainerJSONBase{ID: "0f6df6593e1b0123456789abcdef"},
+			}},
+			ServiceSpec: spec,
+		}
+	}
+	spec := func(image string, labels, containerLabels map[string]string) api.ServiceSpec {
+		return api.ServiceSpec{
+			Name:      "web",
+			Mode:      api.ServiceModeReplicated,
+			Replicas:  1,
+			Labels:    labels,
+			Container: api.ContainerSpec{Image: image, Labels: containerLabels},
+		}
+	}
+
+	withCaddy := func(config string) api.ServiceSpec {
+		s := spec("nginx:1.25", nil, nil)
+		s.Caddy = &api.CaddySpec{Config: config}
+		return s
+	}
+	replaceFrom := func(old api.ServiceSpec) func(api.ServiceSpec) []operation.Operation {
+		return func(new api.ServiceSpec) []operation.Operation {
+			return []operation.Operation{&operation.ReplaceContainerOperation{
+				Spec:         new,
+				MachineName:  "machine-1",
+				OldContainer: newContainer(old),
+			}}
+		}
+	}
+
+	tests := []struct {
+		name         string
+		isNewService bool
+		spec         api.ServiceSpec
+		ops          func(spec api.ServiceSpec) []operation.Operation
+		expected     []string
+	}{
+		{
+			name: "caddy config lines added and removed",
+			spec: withCaddy("{\n\tstorage uncloud\n}"),
+			ops:  replaceFrom(withCaddy("{\n\tdebug\n\tauto_https off\n}")),
+			expected: []string{
+				"~ update service web",
+				"  │   image:        nginx:1.25",
+				"  │ ~ caddy config: +1 -2 lines",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "caddy config single line added",
+			spec: withCaddy("{\n\tstorage uncloud\n\tdebug\n}"),
+			ops:  replaceFrom(withCaddy("{\n\tstorage uncloud\n}")),
+			expected: []string{
+				"~ update service web",
+				"  │   image:        nginx:1.25",
+				"  │ ~ caddy config: +1 line",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "caddy config added to existing service",
+			spec: withCaddy("{\n\tstorage uncloud\n}"),
+			ops:  replaceFrom(spec("nginx:1.25", nil, nil)),
+			expected: []string{
+				"~ update service web",
+				"  │   image:        nginx:1.25",
+				"  │ + caddy config: +3 lines",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "caddy config removed",
+			spec: spec("nginx:1.25", nil, nil),
+			ops:  replaceFrom(withCaddy("{\n\tstorage uncloud\n}")),
+			expected: []string{
+				"~ update service web",
+				"  │   image:        nginx:1.25",
+				"  │ - caddy config: -3 lines",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "caddy config with only reordered lines",
+			spec: withCaddy("a {\n}\nb {\n}"),
+			ops:  replaceFrom(withCaddy("b {\n}\na {\n}")),
+			expected: []string{
+				"~ update service web",
+				"  │   image:        nginx:1.25",
+				"  │ ~ caddy config: changed",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "caddy config unchanged ignoring surrounding whitespace",
+			spec: withCaddy("{\n\tdebug\n}"),
+			ops:  replaceFrom(withCaddy("\n{\n\tdebug\n}\n")),
+			expected: []string{
+				"~ update service web",
+				"  │   image: nginx:1.25",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name:         "new service with caddy config",
+			isNewService: true,
+			spec:         withCaddy("{\n\tstorage uncloud\n}"),
+			ops: func(new api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.RunContainerOperation{Spec: new, MachineName: "machine-1"}}
+			},
+			expected: []string{
+				"+ create service web",
+				"  │   image:        nginx:1.25",
+				"  │   caddy config: +3 lines",
+				"  │",
+				"  ╰── +   run container web on machine-1",
+			},
+		},
+		{
+			name:         "caddy config before replicas",
+			isNewService: true,
+			spec: func() api.ServiceSpec {
+				s := withCaddy("{\n\tstorage uncloud\n}")
+				s.Replicas = 3
+				return s
+			}(),
+			ops: func(new api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.RunContainerOperation{Spec: new, MachineName: "machine-1"}}
+			},
+			expected: []string{
+				"+ create service web",
+				"  │   image:        nginx:1.25",
+				"  │   caddy config: +3 lines",
+				"  │   replicas:     3",
+				"  │",
+				"  ╰── +   run container web on machine-1",
+			},
+		},
+		{
+			name: "service labels added, changed, and removed in place",
+			spec: spec("nginx:1.25", map[string]string{"env": "prod", "team": "platform", "same": "1"}, nil),
+			ops: func(api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.UpdateContainerOperation{
+					MachineName: "machine-1",
+					Container: newContainer(spec(
+						"nginx:1.25", map[string]string{"old": "legacy", "team": "backend", "same": "1"}, nil,
+					)),
+				}}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │   image: nginx:1.25",
+				"  │ ~ labels:",
+				"  │     + env=prod",
+				"  │     - old=legacy",
+				"  │     ~ team=backend → platform",
+				"  │",
+				"  ╰── ~   update container web/0f6df6593e1b on machine-1 (in place)",
+			},
+		},
+		{
+			name: "container labels changed with image on replace",
+			spec: spec("nginx:1.26", nil, map[string]string{"proxy": "on"}),
+			ops: func(new api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.ReplaceContainerOperation{
+					Spec:         new,
+					MachineName:  "machine-1",
+					OldContainer: newContainer(spec("nginx:1.25", nil, map[string]string{"proxy": "off"})),
+				}}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │ ~ image: nginx:1.25 → 1.26",
+				"  │ ~ container labels:",
+				"  │     ~ proxy=off → on",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "no label blocks when labels are the same",
+			spec: spec("nginx:1.26", map[string]string{}, map[string]string{"proxy": "on"}),
+			ops: func(new api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.ReplaceContainerOperation{
+					Spec:         new,
+					MachineName:  "machine-1",
+					OldContainer: newContainer(spec("nginx:1.25", nil, map[string]string{"proxy": "on"})),
+				}}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │ ~ image: nginx:1.25 → 1.26",
+				"  │",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-1",
+			},
+		},
+		{
+			name: "replaced container determines old spec over updated one",
+			spec: spec("nginx:1.26", map[string]string{"team": "platform"}, nil),
+			ops: func(new api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{
+					&operation.UpdateContainerOperation{
+						MachineName: "machine-1",
+						Container:   newContainer(spec("nginx:1.26", map[string]string{"team": "backend"}, nil)),
+					},
+					&operation.ReplaceContainerOperation{
+						Spec:         new,
+						MachineName:  "machine-2",
+						OldContainer: newContainer(spec("nginx:1.25", map[string]string{"team": "backend"}, nil)),
+					},
+				}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │ ~ image: nginx:1.25 → 1.26",
+				"  │ ~ labels:",
+				"  │     ~ team=backend → platform",
+				"  │",
+				"  ├── ~   update container web/0f6df6593e1b on machine-1 (in place)",
+				"  ╰── +/- replace container web/0f6df6593e1b on machine-2",
+			},
+		},
+		{
+			name: "all labels removed",
+			spec: spec("nginx:1.25", map[string]string{}, nil),
+			ops: func(api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.UpdateContainerOperation{
+					MachineName: "machine-1",
+					Container:   newContainer(spec("nginx:1.25", map[string]string{"a": "1", "b": "2"}, nil)),
+				}}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │   image: nginx:1.25",
+				"  │ - labels:",
+				"  │     - a=1",
+				"  │     - b=2",
+				"  │",
+				"  ╰── ~   update container web/0f6df6593e1b on machine-1 (in place)",
+			},
+		},
+		{
+			name: "first labels added to existing service",
+			spec: spec("nginx:1.25", map[string]string{"a": "1"}, nil),
+			ops: func(api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.UpdateContainerOperation{
+					MachineName: "machine-1",
+					Container:   newContainer(spec("nginx:1.25", nil, nil)),
+				}}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │   image: nginx:1.25",
+				"  │ + labels:",
+				"  │     + a=1",
+				"  │",
+				"  ╰── ~   update container web/0f6df6593e1b on machine-1 (in place)",
+			},
+		},
+		{
+			name: "empty and non-printable label values are escaped on a single line",
+			spec: spec("nginx:1.25", map[string]string{
+				"added":    "1\n2",
+				"cleared":  "",
+				"json":     `{"a":1}`,
+				"multi":    "a\nb\nc",
+				"tab":      "a\tb",
+				"trailing": "hello\n",
+				"unicode":  "héllo wörld 🚀",
+			}, nil),
+			ops: func(api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.UpdateContainerOperation{
+					MachineName: "machine-1",
+					Container: newContainer(spec("nginx:1.25", map[string]string{
+						"cleared":  "x",
+						"multi":    "a\nb",
+						"trailing": "hello",
+					}, nil)),
+				}}
+			},
+			expected: []string{
+				"~ update service web",
+				"  │   image: nginx:1.25",
+				"  │ ~ labels:",
+				`  │     + added="1\n2"`,
+				`  │     ~ cleared=x → ""`,
+				`  │     + json={"a":1}`,
+				`  │     ~ multi="a\nb" → "a\nb\nc"`,
+				`  │     + tab="a\tb"`,
+				`  │     ~ trailing=hello → "hello\n"`,
+				"  │     + unicode=héllo wörld 🚀",
+				"  │",
+				"  ╰── ~   update container web/0f6df6593e1b on machine-1 (in place)",
+			},
+		},
+		{
+			name:         "new service with labels",
+			isNewService: true,
+			spec:         spec("nginx:1.25", map[string]string{"team": "platform"}, map[string]string{"proxy": "on"}),
+			ops: func(new api.ServiceSpec) []operation.Operation {
+				return []operation.Operation{&operation.RunContainerOperation{Spec: new, MachineName: "machine-1"}}
+			},
+			expected: []string{
+				"+ create service web",
+				"  │   image: nginx:1.25",
+				"  │   labels:",
+				"  │       team=platform",
+				"  │   container labels:",
+				"  │       proxy=on",
+				"  │",
+				"  ╰── +   run container web on machine-1",
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			plan := ServicePlan{ServiceName: "web", IsNewService: tt.isNewService, Spec: tt.spec}
+			plan.Operations = tt.ops(tt.spec)
+
+			// The table pads cells to the column width so ignore trailing whitespace.
+			var lines []string
+			for line := range strings.Lines(ansi.Strip(plan.Format())) {
+				lines = append(lines, strings.TrimRight(line, " \n"))
+			}
+			assert.Equal(t, tt.expected, lines)
+		})
+	}
+}

pkg/client/deploy/strategy.go

@@ -8,8 +8,8 @@ import (
 	"github.com/psviderski/uncloud/api/pb"
 	"github.com/psviderski/uncloud/internal/secret"
 	"github.com/psviderski/uncloud/pkg/api"
-	"github.com/psviderski/uncloud/pkg/client/deploy/operation"
 	"github.com/psviderski/uncloud/pkg/client/deploy/scheduler"
+	"github.com/psviderski/uncloud/pkg/client/operation"
 )
 
 // Strategy defines how a service should be deployed or updated. Different implementations can provide various
@@ -92,7 +92,7 @@ func (s *RollingStrategy) planReplicated(svc *api.Service, spec api.ServiceSpec)
 
 	// Organise existing containers by machine.
 	containersOnMachine := make(map[string][]api.ServiceContainer)
-	upToDateContainersOnMachine := make(map[string]int)
+	reusableContainersOnMachine := make(map[string]int)
 	containerSpecStatuses := make(map[string]ContainerSpecStatus)
 	if svc != nil {
 		for _, c := range svc.Containers {
@@ -105,40 +105,50 @@ func (s *RollingStrategy) planReplicated(svc *api.Service, spec api.ServiceSpec)
 			if s.ForceRecreate {
 				status = ContainerNeedsRecreate
 			} else {
-				status = EvalContainerSpecChange(c.Container.ServiceSpec, spec)
+				status = EvalContainerChange(c.Container, spec)
 			}
 			containerSpecStatuses[c.Container.ID] = status
 
-			if status == ContainerUpToDate {
+			if status != ContainerNeedsRecreate {
-				upToDateContainersOnMachine[c.MachineID] += 1
+				reusableContainersOnMachine[c.MachineID] += 1
 			}
 		}
 
-		// Sort containers such that running containers with the desired spec are first.
+		// Sort containers such that running containers with the desired spec or that can be updated in place to
-		slices.SortFunc(svc.Containers, func(c1, c2 api.MachineServiceContainer) int {
+		// the desired spec are first.
-			if status, ok := containerSpecStatuses[c1.Container.ID]; ok && status == ContainerUpToDate {
+		// TODO: consider ranking up-to-date containers above ones that need an in-place update (here, when sorting
-				return -1
+		//  machines, and in reconcileGlobalContainer) to avoid updating a container while removing an up-to-date one.
+		isReusable := func(c api.MachineServiceContainer) bool {
+			// Not running containers have no status.
+			status, ok := containerSpecStatuses[c.Container.ID]
+			return ok && status != ContainerNeedsRecreate
 		}
-			if status, ok := containerSpecStatuses[c2.Container.ID]; ok && status == ContainerUpToDate {
+		slices.SortFunc(svc.Containers, func(c1, c2 api.MachineServiceContainer) int {
+			r1, r2 := isReusable(c1), isReusable(c2)
+			switch {
+			case r1 == r2:
+				return 0
+			case r1:
+				return -1
+			default:
 				return 1
 			}
-			return 0
 		})
 
 		for _, c := range svc.Containers {
 			containersOnMachine[c.MachineID] = append(containersOnMachine[c.MachineID], c.Container)
 		}
 
-		// Sort machines such that machines with the most up-to-date containers are first, followed by machines with
+		// Sort machines such that machines with the most containers that are up to date or can be updated in place
-		// existing containers, and finally machines without containers.
+		// are first, followed by machines with existing containers, and finally machines without containers.
 		slices.SortFunc(matchedMachines, func(m1, m2 *pb.MachineInfo) int {
-			if upToDateContainersOnMachine[m1.Id] > 0 && upToDateContainersOnMachine[m2.Id] > 0 {
+			if reusableContainersOnMachine[m1.Id] > 0 && reusableContainersOnMachine[m2.Id] > 0 {
-				return upToDateContainersOnMachine[m2.Id] - upToDateContainersOnMachine[m1.Id]
+				return reusableContainersOnMachine[m2.Id] - reusableContainersOnMachine[m1.Id]
 			}
-			if upToDateContainersOnMachine[m1.Id] > 0 {
+			if reusableContainersOnMachine[m1.Id] > 0 {
 				return -1
 			}
-			if upToDateContainersOnMachine[m2.Id] > 0 {
+			if reusableContainersOnMachine[m2.Id] > 0 {
 				return 1
 			}
 			return len(containersOnMachine[m2.Id]) - len(containersOnMachine[m1.Id])
@@ -170,7 +180,16 @@ func (s *RollingStrategy) planReplicated(svc *api.Service, spec api.ServiceSpec)
 			if status == ContainerUpToDate {
 				continue
 			}
-			// TODO: handle ContainerNeedsUpdate when update of mutable fields on a container is supported.
+			if status == ContainerNeedsUpdate {
+				plan.Operations = append(plan.Operations, &operation.UpdateContainerOperation{
+					ServiceID:   plan.ServiceID,
+					MachineID:   m.Id,
+					MachineName: m.Name,
+					Container:   ctr,
+					Options:     ContainerUpdateOptions(ctr, spec),
+				})
+				continue
+			}
 		}
 
 		// Replace the old container with a new one.
@@ -301,7 +320,7 @@ func reconcileGlobalContainer(
 	}
 
 	// Check if there is a container with the same spec already running. If so, remove the rest.
-	upToDate := false
+	reused := false
 	for i, c := range containers {
 		if !c.Container.State.Running || c.Container.State.Paused {
 			// Skip containers that are not running.
@@ -312,12 +331,21 @@ func reconcileGlobalContainer(
 		if forceRecreate {
 			status = ContainerNeedsRecreate
 		} else {
-			status = EvalContainerSpecChange(c.Container.ServiceSpec, spec)
+			status = EvalContainerChange(c.Container, spec)
 		}
 
-		if status == ContainerUpToDate {
+		if status != ContainerNeedsRecreate {
-			// The container is already running with the same spec.
+			// The container is already running with the desired spec or can be updated in place to it.
-			upToDate = true
+			reused = true
+			if status == ContainerNeedsUpdate {
+				ops = append(ops, &operation.UpdateContainerOperation{
+					ServiceID:   serviceID,
+					MachineID:   machine.Id,
+					MachineName: machine.Name,
+					Container:   c.Container,
+					Options:     ContainerUpdateOptions(c.Container, spec),
+				})
+			}
 			for j, old := range containers {
 				if i == j {
 					continue
@@ -331,10 +359,8 @@ func reconcileGlobalContainer(
 			}
 			break
 		}
-		// TODO: handle ContainerNeedsUpdate when update of mutable fields on a container is supported.
-		//  Make sure to update preDeployOperation accordingly.
 	}
-	if upToDate {
+	if reused {
 		return ops, nil
 	}
 
@@ -360,6 +386,7 @@ func reconcileGlobalContainer(
 			if err != nil || len(conflictingPorts) > 0 {
 				ops = append(ops, &operation.StopContainerOperation{
 					ServiceID:       serviceID,
+					ServiceName:     c.Container.ServiceSpec.Name,
 					ContainerID:     c.Container.ID,
 					MachineID:       machine.Id,
 					MachineName:     machine.Name,

pkg/client/deploy/strategy_test.go

@@ -1,6 +1,8 @@
 package deploy
 
 import (
+	"slices"
+	"strings"
 	"testing"
 
 	"github.com/docker/docker/api/types/container"
@@ -8,9 +10,10 @@ import (
 	"github.com/google/go-cmp/cmp/cmpopts"
 	"github.com/psviderski/uncloud/api/pb"
 	"github.com/psviderski/uncloud/pkg/api"
-	"github.com/psviderski/uncloud/pkg/client/deploy/operation"
 	"github.com/psviderski/uncloud/pkg/client/deploy/scheduler"
+	"github.com/psviderski/uncloud/pkg/client/operation"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestDetermineUpdateOrder(t *testing.T) {
@@ -360,6 +363,7 @@ func TestReconcileGlobalContainer(t *testing.T) {
 			expectedOps: []operation.Operation{
 				&operation.StopContainerOperation{
 					ServiceID:   "service-1",
+					ServiceName: container2WithPort9090.ServiceSpec.Name,
 					ContainerID: "container-2",
 					MachineID:   "machine-1",
 					MachineName: "machine-1",
@@ -424,6 +428,121 @@ func TestReconcileGlobalContainer(t *testing.T) {
 	}
 }
 
+func TestRollingStrategy_Plan_UpdateInPlace(t *testing.T) {
+	t.Parallel()
+
+	state := &scheduler.ClusterState{
+		Machines: []*scheduler.Machine{
+			{Info: &pb.MachineInfo{Id: "m-1", Name: "machine-1"}},
+			{Info: &pb.MachineInfo{Id: "m-2", Name: "machine-2"}},
+		},
+	}
+	newContainer := func(id string, spec api.ServiceSpec) api.ServiceContainer {
+		ctr := newServiceContainer(id, container.State{Running: true, Status: "running"})
+		ctr.ServiceSpec = spec
+		return ctr
+	}
+
+	tests := []struct {
+		name   string
+		change func(spec *api.ServiceSpec)
+		// options are the expected in-place update options. Nil means the containers must be recreated.
+		options *api.UpdateContainerOptions
+	}{
+		{
+			name:    "service labels",
+			change:  func(spec *api.ServiceSpec) { spec.Labels = map[string]string{"tier": "backend"} },
+			options: &api.UpdateContainerOptions{ServiceLabels: &map[string]string{"tier": "backend"}},
+		},
+		{
+			name: "CPU and memory limits",
+			change: func(spec *api.ServiceSpec) {
+				spec.Container.Resources.CPU = 2 * api.Core
+				spec.Container.Resources.Memory = 512 * 1024 * 1024
+			},
+			options: &api.UpdateContainerOptions{
+				CPU:    new(int64(2 * api.Core)),
+				Memory: new(int64(512 * 1024 * 1024)),
+			},
+		},
+		{
+			name:   "remove memory limit",
+			change: func(spec *api.ServiceSpec) { spec.Container.Resources.Memory = 0 },
+		},
+	}
+
+	for _, tt := range tests {
+		for _, mode := range []string{api.ServiceModeReplicated, api.ServiceModeGlobal} {
+			t.Run(tt.name+"/"+mode, func(t *testing.T) {
+				currentSpec := api.ServiceSpec{
+					Name: "app",
+					Mode: mode,
+					Container: api.ContainerSpec{
+						Image: "nginx:latest",
+						Resources: api.ContainerResources{
+							CPU:    api.Core,
+							Memory: 256 * 1024 * 1024,
+						},
+					},
+					Labels:    map[string]string{"tier": "frontend"},
+					PreDeploy: &api.PreDeployHook{Command: []string{"db", "migrate"}},
+				}
+				if mode == api.ServiceModeReplicated {
+					currentSpec.Replicas = 2
+				}
+				newSpec := currentSpec.Clone()
+				tt.change(&newSpec)
+
+				ctr1 := newContainer("container-1", currentSpec)
+				ctr2 := newContainer("container-2", currentSpec)
+				svc := &api.Service{
+					ID:   "svc-1",
+					Name: "app",
+					Mode: mode,
+					Containers: []api.MachineServiceContainer{
+						{MachineID: "m-1", MachineName: "machine-1", Container: ctr1},
+						{MachineID: "m-2", MachineName: "machine-2", Container: ctr2},
+					},
+				}
+
+				plan, err := (&RollingStrategy{}).Plan(state, svc, newSpec)
+				require.NoError(t, err)
+
+				if tt.options == nil {
+					require.NotEmpty(t, plan.Operations)
+					for _, op := range plan.Operations {
+						_, isUpdate := op.(*operation.UpdateContainerOperation)
+						assert.False(t, isUpdate, "Expected containers to be recreated, got %s", op)
+					}
+					return
+				}
+				// The replicated strategy shuffles machines so sort the operations to compare them deterministically.
+				ops := slices.Clone(plan.Operations)
+				slices.SortFunc(ops, func(a, b operation.Operation) int {
+					return strings.Compare(a.String(), b.String())
+				})
+				// The containers are updated in place and the pre-deploy hook is not run.
+				assertOperationsEqual(t, []operation.Operation{
+					&operation.UpdateContainerOperation{
+						ServiceID:   "svc-1",
+						MachineID:   "m-1",
+						MachineName: "machine-1",
+						Container:   ctr1,
+						Options:     *tt.options,
+					},
+					&operation.UpdateContainerOperation{
+						ServiceID:   "svc-1",
+						MachineID:   "m-2",
+						MachineName: "machine-2",
+						Container:   ctr2,
+						Options:     *tt.options,
+					},
+				}, ops)
+			})
+		}
+	}
+}
+
 func TestPreDeployOperations(t *testing.T) {
 	hook := &api.PreDeployHook{
 		Command: []string{"db", "migrate"},
@@ -502,6 +621,20 @@ func TestPreDeployOperations(t *testing.T) {
 				},
 			},
 		},
+		{
+			name: "plan has only UpdateContainerOperation",
+			plan: ServicePlan{
+				ServiceID:   "svc-1",
+				ServiceName: "app",
+				Spec:        api.ServiceSpec{PreDeploy: hook},
+				SequenceOperation: operation.SequenceOperation{
+					Operations: []operation.Operation{
+						&operation.UpdateContainerOperation{MachineID: "m-1", MachineName: "machine-1"},
+					},
+				},
+			},
+			expected: nil,
+		},
 		{
 			name: "plan has only RemoveContainerOperation",
 			plan: ServicePlan{
@@ -614,7 +747,8 @@ func TestPreDeployOperations(t *testing.T) {
 			},
 			svc: &api.Service{
 				HookContainers: []api.MachineServiceContainer{
-					{MachineID: "m-1", Container: newServiceContainer("stopped-hook", container.State{Status: "exited"})},
+					{MachineID: "m-1", Container: newServiceContainer("stopped-hook",
+						container.State{Status: "exited"})},
 					{MachineID: "m-1", Container: runningHook1},
 					{MachineID: "m-3", Container: runningHook2},
 				},
@@ -660,6 +794,7 @@ func newServiceContainer(id string, state container.State) api.ServiceContainer
 				ID:    id,
 				State: &state,
 			},
+			Config: &container.Config{},
 		},
 	}}
 }

pkg/client/image.go

@@ -23,9 +23,11 @@ import (
 	"github.com/docker/docker/api/types/image"
 	dockerclient "github.com/docker/docker/client"
 	"github.com/docker/docker/pkg/jsonmessage"
+	"github.com/docker/docker/pkg/stringid"
 	"github.com/docker/go-connections/nat"
 	ocispec "github.com/opencontainers/image-spec/specs-go/v1"
 	"github.com/psviderski/uncloud/api/pb"
+	cliprogress "github.com/psviderski/uncloud/internal/cli/progress"
 	"github.com/psviderski/uncloud/internal/cli/tui"
 	"github.com/psviderski/uncloud/internal/docker"
 	"github.com/psviderski/uncloud/internal/machine/constants"
@@ -34,6 +36,8 @@ import (
 	"github.com/psviderski/uncloud/internal/secret"
 	"github.com/psviderski/uncloud/pkg/api"
 	netproxy "golang.org/x/net/proxy"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
 )
 
 // This is the container image used to run socat proxy containers.
@@ -61,7 +65,7 @@ func (cli *Client) ListImages(ctx context.Context, filter api.ImageFilter) ([]ap
 	// Broadcast the image list request to the specified machines or all machines if none specified.
 	listCtx := ProxyMachinesContext(ctx, filter.Machines)
 
-	opts := image.ListOptions{Manifests: true}
+	opts := image.ListOptions{All: true, Manifests: true}
 	if filter.Name != "" {
 		opts.Filters = filters.NewArgs(
 			filters.Arg("reference", filter.Name),
@@ -92,7 +96,7 @@ func (cli *Client) ListImages(ctx context.Context, filter api.ImageFilter) ([]ap
 			ContainerdStore: msg.ContainerdStore,
 		}
 
-		if msg.Metadata.Error == "" && len(msg.Images) > 0 {
+		if msg.Metadata.Err() == nil && len(msg.Images) > 0 {
 			if err = json.Unmarshal(msg.Images, &mi.Images); err != nil {
 				return nil, fmt.Errorf("unmarshal images: %w", err)
 			}
@@ -104,6 +108,81 @@ func (cli *Client) ListImages(ctx context.Context, filter api.ImageFilter) ([]ap
 	return machineImages, nil
 }
 
+// RemoveImage removes one exact image name or ID on the selected machines with protection of used references.
+// Results include the outcome of removing an image on each machine.
+//
+// An absent image on some machines is normal if removal succeeds on at least one machine. If every selected machine
+// reports NotFound, the returned error is api.ErrNotFound. Other machine failures are combined, with successful
+// results still returned.
+func (cli *Client) RemoveImage(
+	ctx context.Context, imageNameOrID string, opts api.RemoveImageOptions,
+) ([]api.MachineImageRemoval, error) {
+	if strings.TrimSpace(imageNameOrID) == "" {
+		return nil, errors.New("image name or ID is required")
+	}
+
+	resp, err := cli.Docker.GRPCClient.RemoveImage(ProxyMachinesContext(ctx, opts.Machines),
+		&pb.RemoveImageRequest{Id: imageNameOrID, PruneChildren: true})
+	if err != nil {
+		return nil, err
+	}
+
+	pw := progress.ContextWriter(ctx)
+	var results []api.MachineImageRemoval
+	var errs []error
+	notFoundCount := 0
+	for _, msg := range resp.Messages {
+		result := api.MachineImageRemoval{Metadata: msg.Metadata}
+		for _, removal := range msg.Results {
+			result.Results = append(result.Results, image.DeleteResponse{
+				Deleted:  removal.Deleted,
+				Untagged: removal.Untagged,
+			})
+		}
+		results = append(results, result)
+
+		eventID := cliprogress.ImageEventID(imageNameOrID, msg.Metadata.MachineName)
+		if err = result.Metadata.Err(); err != nil {
+			event := progress.NewEvent(eventID, progress.Done, "Not found")
+			if status.Code(err) == codes.NotFound {
+				notFoundCount++
+			} else {
+				errs = append(errs, fmt.Errorf("machine '%s': %w", msg.Metadata.MachineName, err))
+				event.Status = progress.Error
+				event.StatusText = status.Convert(err).Message()
+			}
+			pw.Event(event)
+			continue
+		}
+
+		// An unnamed image can return Deleted without Untagged. Use the requested name or ID as its row fallback.
+		deleted := false
+		for _, removal := range result.Results {
+			if removal.Untagged != "" {
+				eventID = cliprogress.ImageEventID(removal.Untagged, msg.Metadata.MachineName)
+				pw.Event(progress.NewEvent(eventID, progress.Done, "Untagged"))
+			}
+			if removal.Deleted != "" && !deleted {
+				// Docker reports the deletion of the selected image before any pruned parents or released layers.
+				// Only that first deletion updates the last untagged row marking the image as Removed.
+				deleted = true
+				pw.Event(progress.NewEvent(eventID, progress.Done,
+					fmt.Sprintf("Removed (%s)", stringid.TruncateID(removal.Deleted))))
+			}
+		}
+	}
+
+	if err = errors.Join(errs...); err != nil {
+		return results, err
+	}
+	if notFoundCount == len(results) {
+		return results, api.ErrNotFound
+	}
+
+	return results, nil
+}
+
+// TODO: move PushImageOptions to api package and include PushImage in ImageClient interface.
 type PushImageOptions struct {
 	// AllMachines pushes the image to all machines in the cluster. Takes precedence over Machines field.
 	AllMachines bool
@@ -215,7 +294,7 @@ func (cli *Client) pushImageToMachine(
 		Name:     "%invalid-name-to-only-check-store-type%",
 	})
 	if err == nil {
-		err = images[0].Error()
+		err = images[0].Metadata.Err()
 	}
 	if err != nil {
 		return fmt.Errorf("check Docker image store type on machine '%s': %w", machine.Name, err)

pkg/client/operation/container.go

@@ -83,6 +83,8 @@ func (o *RunContainerOperation) String() string {
 // StopContainerOperation stops a container on a specific machine.
 type StopContainerOperation struct {
 	ServiceID string
+	// ServiceName is used for formatting the operation as part of the deployment plan.
+	ServiceName string
 	ContainerID string
 	MachineID   string
 	// MachineName is used for formatting the operation as part of the deployment plan.
@@ -98,8 +100,7 @@ func (o *StopContainerOperation) Execute(ctx context.Context, cli Client) error
 }
 
 func (o *StopContainerOperation) Format() string {
-	// TODO: pass service name to format the display name consistently with other operations.
+	displayName := o.ServiceName + tui.Faint.Render("/") + stringid.TruncateID(o.ContainerID)
-	displayName := stringid.TruncateID(o.ContainerID)
 
 	return tui.BoldRed.Render("-") + "   " +
 		tui.Faint.Render("stop container") + " " +
@@ -173,6 +174,16 @@ type ReplaceContainerOperation struct {
 func (o *ReplaceContainerOperation) Execute(ctx context.Context, cli Client) error {
 	stopFirst := o.Order == api.UpdateOrderStopFirst
 
+	// Create the new container before stopping the old one. CreateContainer also pulls the image according to the
+	// service pull policy, so the stop-first downtime only includes stopping the old container and starting and
+	// monitoring the new one.
+	resp, err := cli.CreateContainer(ctx, o.ServiceID, o.Spec, o.MachineID)
+	if err != nil {
+		return fmt.Errorf("create new container: %w", err)
+	}
+	// Override event ID so StartContainer and WaitContainerHealthy update the same progress line as creation.
+	newCtx := cliprogress.WithEventID(ctx, cliprogress.NewContainerEventID(ctx, resp.Name, o.MachineName))
+
 	wasRunning := false
 	if stopFirst {
 		// Inspect the old container to remember its running state before stopping.
@@ -189,12 +200,6 @@ func (o *ReplaceContainerOperation) Execute(ctx context.Context, cli Client) err
 		}
 	}
 
-	resp, err := cli.CreateContainer(ctx, o.ServiceID, o.Spec, o.MachineID)
-	if err != nil {
-		return fmt.Errorf("create new container: %w", err)
-	}
-	// Override event ID so StartContainer and WaitContainerHealthy update the same progress line as creation.
-	newCtx := cliprogress.WithEventID(ctx, cliprogress.NewContainerEventID(ctx, resp.Name, o.MachineName))
 	if err = cli.StartContainer(newCtx, o.ServiceID, resp.ID); err != nil {
 		return fmt.Errorf("start new container: %w", err)
 	}
@@ -281,6 +286,39 @@ func (o *ReplaceContainerOperation) String() string {
 		o.MachineID, o.ServiceID, o.OldContainer.ID, o.Order)
 }
 
+// UpdateContainerOperation updates the requested properties of a running container in place without recreating it.
+type UpdateContainerOperation struct {
+	ServiceID string
+	MachineID string
+	// MachineName is used for formatting the operation as part of the deployment plan.
+	MachineName string
+	Container   api.ServiceContainer
+	Options     api.UpdateContainerOptions
+}
+
+func (o *UpdateContainerOperation) Execute(ctx context.Context, cli Client) error {
+	if err := cli.UpdateContainer(ctx, o.ServiceID, o.Container.ID, o.Options); err != nil {
+		return fmt.Errorf("update container: %w", err)
+	}
+	return nil
+}
+
+func (o *UpdateContainerOperation) Format() string {
+	displayName := o.Container.ServiceSpec.Name + tui.Faint.Render("/") + o.Container.ShortID()
+
+	return tui.BoldYellow.Render("~") + "   " +
+		tui.Faint.Render("update container") + " " +
+		displayName + " " +
+		tui.Faint.Render("on") + " " +
+		o.MachineName + " " +
+		tui.Faint.Render("(in place)")
+}
+
+func (o *UpdateContainerOperation) String() string {
+	return fmt.Sprintf("UpdateContainerOperation[machine_id=%s service_id=%s container_id=%s]",
+		o.MachineID, o.ServiceID, o.Container.ID)
+}
+
 // stopOptions converts a stop grace period duration to Docker container stop options.
 func stopOptions(gracePeriod *time.Duration) container.StopOptions {
 	if gracePeriod == nil {