feat(wanda): implement artifact copying from built images

Extract artifacts from built images using `docker cp`. This approach is cross-platform reliable on Windows.

- Creates container from image without starting it
- Copies each artifact using `docker cp`
- Cleans up container when done
- Optional artifacts log warning instead of failing
- Extraction runs for root spec only, even on cache hit
- Logs extraction duration for performance monitoring

Example:

```shell
❯  PYTHON_VERSION=3.13 MANYLINUX_VERSION=260128.221a193 HOSTTYPE=aarch64 ARCH_SUFFIX=-aarch64 BUILDKITE_COMMIT=b5737cefc0 IS_LOCAL_BUILD=true /Users/andrew/devel/rayci-wanda-artifacts/_release/wanda-darwin-arm64 --artifacts_dir .whl/ ci/docker/ray-wheel.wanda.yaml

...

2026/02/04 09:10:39 extracting 1 artifact(s) from localhost:5000/rayci-work:ray-wheel-py3.13-aarch64
Successfully copied 72MB to /Users/andrew/devel/ray-local-wheel-build/.whl/ray-3.0.0.dev0-cp313-cp313-manylinux2014_aarch64.whl
2026/02/04 09:10:40 extracted 1 artifact(s) in 722ms:
2026/02/04 09:10:40   /Users/andrew/devel/ray-local-wheel-build/.whl/ray-3.0.0.dev0-cp313-cp313-manylinux2014_aarch64.whl
```

Topic: wanda-artifact-copy
Relative: wanda-artifact-spec
Labels: draft

Signed-off-by: andrew <andrew@anyscale.com>

Author: andrew-anyscale

wanda/docker_cmd.go

@@ -1,9 +1,11 @@
 package wanda
 
 import (
+	"archive/tar"
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"io"
 	"log"
 	"os"
 	"os/exec"
@@ -100,6 +102,11 @@ type dockerImageInfo struct {
 	ID          string `json:"Id"`
 	RepoDigests []string
 	RepoTags    []string
+	Config      *dockerImageConfig `json:"Config"`
+}
+
+type dockerImageConfig struct {
+	WorkingDir string `json:"WorkingDir"`
 }
 
 func (c *dockerCmd) inspectImage(tag string) (*dockerImageInfo, error) {
@@ -126,6 +133,68 @@ func (c *dockerCmd) tag(src, asTag string) error {
 	return c.run("tag", src, asTag)
 }
 
+// createContainer creates a container from an image without starting it.
+// Returns the container ID. A dummy command is provided for images without
+// CMD/ENTRYPOINT. The command doesn't need to exist since the container is
+// never started.
+func (c *dockerCmd) createContainer(image string) (string, error) {
+	cmd := c.cmd("create", image, "unused")
+	buf := new(bytes.Buffer)
+	cmd.Stdout = buf
+	if err := cmd.Run(); err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(buf.String()), nil
+}
+
+// copyFromContainer copies a file or directory from a container to the host.
+func (c *dockerCmd) copyFromContainer(containerID, src, dst string) error {
+	return c.run("cp", containerID+":"+src, dst)
+}
+
+// removeContainer removes a container quietly (no stdout).
+func (c *dockerCmd) removeContainer(containerID string) error {
+	cmd := exec.Command(c.bin, "rm", containerID)
+	cmd.Env = c.envs
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+// listContainerFiles lists all files in a container using docker export.
+func (c *dockerCmd) listContainerFiles(containerID string) ([]string, error) {
+	exportCmd := exec.Command(c.bin, "export", containerID)
+	exportCmd.Env = c.envs
+
+	stdout, err := exportCmd.StdoutPipe()
+	if err != nil {
+		return nil, fmt.Errorf("create stdout pipe: %w", err)
+	}
+
+	if err := exportCmd.Start(); err != nil {
+		return nil, fmt.Errorf("start docker export: %w", err)
+	}
+
+	var files []string
+	tr := tar.NewReader(stdout)
+	for {
+		header, err := tr.Next()
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			exportCmd.Process.Kill()
+			return nil, fmt.Errorf("read tar stream: %w", err)
+		}
+		files = append(files, "/"+strings.TrimPrefix(header.Name, "/"))
+	}
+
+	if err := exportCmd.Wait(); err != nil {
+		return nil, fmt.Errorf("docker export: %w", err)
+	}
+
+	return files, nil
+}
+
 func (c *dockerCmd) build(in *buildInput, core *buildInputCore, hints *buildInputHints) error {
 	if hints == nil {
 		hints = newBuildInputHints(nil, nil)

wanda/docker_cmd_test.go

@@ -1,9 +1,10 @@
 package wanda
 
 import (
-	"testing"
-
+	"os"
+	"path/filepath"
 	"strings"
+	"testing"
 
 	"github.com/google/go-containerregistry/pkg/name"
 	"github.com/google/go-containerregistry/pkg/v1/daemon"
@@ -135,3 +136,113 @@ func TestDockerCmdBuild_withHints(t *testing.T) {
 		t.Errorf("MESSAGE env got %q, want `MESSAGE=hint message`", messageEnv)
 	}
 }
+
+func TestDockerCmdCopyFromContainer(t *testing.T) {
+	cmd := newDockerCmd(&dockerCmdConfig{})
+
+	const testImage = "alpine:latest"
+
+	if err := cmd.run("pull", testImage); err != nil {
+		t.Fatalf("pull image: %v", err)
+	}
+
+	containerID, err := cmd.createContainer(testImage)
+	if err != nil {
+		t.Fatalf("createContainer: %v", err)
+	}
+	defer cmd.removeContainer(containerID)
+
+	tmpDir := t.TempDir()
+
+	// Copy a known file from the container
+	if err := cmd.copyFromContainer(containerID, "/etc/alpine-release", filepath.Join(tmpDir, "alpine-release")); err != nil {
+		t.Fatalf("copyFromContainer: %v", err)
+	}
+
+	if _, err := os.Stat(filepath.Join(tmpDir, "alpine-release")); os.IsNotExist(err) {
+		t.Error("alpine-release was not copied")
+	}
+}
+
+func TestDockerCmdCopyFromContainer_directory(t *testing.T) {
+	cmd := newDockerCmd(&dockerCmdConfig{})
+
+	const testImage = "alpine:latest"
+
+	if err := cmd.run("pull", testImage); err != nil {
+		t.Fatalf("pull image: %v", err)
+	}
+
+	containerID, err := cmd.createContainer(testImage)
+	if err != nil {
+		t.Fatalf("createContainer: %v", err)
+	}
+	defer cmd.removeContainer(containerID)
+
+	tmpDir := t.TempDir()
+
+	// Copy a directory from the container
+	if err := cmd.copyFromContainer(containerID, "/etc", filepath.Join(tmpDir, "etc")); err != nil {
+		t.Fatalf("copyFromContainer: %v", err)
+	}
+
+	if _, err := os.Stat(filepath.Join(tmpDir, "etc", "alpine-release")); os.IsNotExist(err) {
+		t.Error("alpine-release was not copied from /etc directory")
+	}
+}
+
+func TestDockerCmdCopyFromContainer_notFound(t *testing.T) {
+	cmd := newDockerCmd(&dockerCmdConfig{})
+
+	const testImage = "alpine:latest"
+
+	if err := cmd.run("pull", testImage); err != nil {
+		t.Fatalf("pull image: %v", err)
+	}
+
+	containerID, err := cmd.createContainer(testImage)
+	if err != nil {
+		t.Fatalf("createContainer: %v", err)
+	}
+	defer cmd.removeContainer(containerID)
+
+	tmpDir := t.TempDir()
+
+	// Copying a non-existent file should fail
+	if err := cmd.copyFromContainer(containerID, "/nonexistent/file", filepath.Join(tmpDir, "file")); err == nil {
+		t.Error("copyFromContainer should fail for non-existent file")
+	}
+}
+
+func TestDockerCmdListContainerFiles(t *testing.T) {
+	cmd := newDockerCmd(&dockerCmdConfig{})
+
+	const testImage = "alpine:latest"
+
+	if err := cmd.run("pull", testImage); err != nil {
+		t.Fatalf("pull image: %v", err)
+	}
+
+	containerID, err := cmd.createContainer(testImage)
+	if err != nil {
+		t.Fatalf("createContainer: %v", err)
+	}
+	defer cmd.removeContainer(containerID)
+
+	files, err := cmd.listContainerFiles(containerID)
+	if err != nil {
+		t.Fatalf("listContainerFiles: %v", err)
+	}
+
+	// Check that some expected files are present
+	found := false
+	for _, f := range files {
+		if f == "/etc/alpine-release" {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Error("/etc/alpine-release not found in file list")
+	}
+}

wanda/forge.go

@@ -7,6 +7,7 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
+	"time"
 
 	"github.com/google/go-containerregistry/pkg/authn"
 	cranename "github.com/google/go-containerregistry/pkg/name"
@@ -69,6 +70,17 @@ func Build(specFile string, config *ForgeConfig) error {
 		}
 	}
 
+	// Extract artifacts only for the root spec.
+	if config.ArtifactsDir != "" {
+		rootSpec := graph.Specs[graph.Root].Spec
+		if len(rootSpec.Artifacts) > 0 {
+			rootTag := forge.workTag(rootSpec.Name)
+			if err := forge.ExtractArtifacts(rootSpec, rootTag); err != nil {
+				return fmt.Errorf("extract artifacts: %w", err)
+			}
+		}
+	}
+
 	return nil
 }
 
@@ -182,6 +194,103 @@ func (f *Forge) resolveBases(froms []string) (map[string]*imageSource, error) {
 	return m, nil
 }
 
+// ExtractArtifacts copies Artifacts from a built image to ArtifactsDir.
+// Supports glob patterns in src paths (e.g., "/*.whl").
+//
+// NOTE(andrew-anyscale): We use `docker cp` for copying file-by-file rather than
+// a single more efficient method of extracting from `docker export` because
+// docker cp handles cross-platform issues reliably. If this becomes a bottleneck
+// indicated by the log-line below, we can consider using a different approach.
+func (f *Forge) ExtractArtifacts(spec *Spec, imageTag string) error {
+	d := f.newDockerCmd()
+	artifactsDir := f.config.ArtifactsDir
+
+	if err := os.MkdirAll(artifactsDir, 0755); err != nil {
+		return fmt.Errorf("create artifacts dir: %w", err)
+	}
+
+	log.Printf("extracting %d artifact(s) from %s", len(spec.Artifacts), imageTag)
+	extractStart := time.Now()
+
+	containerID, err := d.createContainer(imageTag)
+	if err != nil {
+		return fmt.Errorf("create container: %w", err)
+	}
+	defer func() {
+		if err := d.removeContainer(containerID); err != nil {
+			log.Printf("warning: failed to remove container %s: %v", containerID, err)
+		}
+	}()
+
+	// Lazily list container files only if needed for glob matching.
+	var containerFiles []string
+	var extracted []string
+
+	for _, a := range spec.Artifacts {
+		if err := a.Validate(); err != nil {
+			return fmt.Errorf("invalid artifact: %w", err)
+		}
+
+		if a.HasGlob() && containerFiles == nil {
+			var err error
+			containerFiles, err = d.listContainerFiles(containerID)
+			if err != nil {
+				return fmt.Errorf("list container files: %w", err)
+			}
+		}
+
+		srcs := a.ResolveSrcs(containerFiles)
+		if len(srcs) == 0 {
+			if a.Optional {
+				log.Printf("warning: no files matched pattern: %s", a.Src)
+				continue
+			}
+			return fmt.Errorf("no files matched pattern: %s", a.Src)
+		}
+
+		dstBase, err := a.ResolveDst(artifactsDir)
+		if err != nil {
+			return fmt.Errorf("resolve artifact dst: %w", err)
+		}
+		copyIntoDir := a.CopyIntoDir(len(srcs))
+
+		if copyIntoDir {
+			if err := os.MkdirAll(dstBase, 0755); err != nil {
+				return fmt.Errorf("create dir for artifact %s: %w", a.Dst, err)
+			}
+		} else {
+			if err := os.MkdirAll(filepath.Dir(dstBase), 0755); err != nil {
+				return fmt.Errorf("create dir for artifact %s: %w", a.Dst, err)
+			}
+		}
+
+		for _, src := range srcs {
+			dst := dstBase
+			if copyIntoDir {
+				dst = filepath.Join(dstBase, filepath.Base(src))
+			}
+
+			if err := d.copyFromContainer(containerID, src, dst); err != nil {
+				if a.Optional {
+					log.Printf("warning: optional artifact not found: %s", src)
+					continue
+				}
+				return fmt.Errorf("copy artifact %s: %w", src, err)
+			}
+			if abs, err := filepath.Abs(dst); err == nil {
+				dst = abs
+			}
+			extracted = append(extracted, dst)
+		}
+	}
+
+	log.Printf("extracted %d artifact(s) in %v:", len(extracted), time.Since(extractStart).Round(time.Millisecond))
+	for _, f := range extracted {
+		log.Printf("  %s", f)
+	}
+	return nil
+}
+
 // Build builds a container image from the given specification.
 func (f *Forge) Build(spec *Spec) error {
 	// Prepare the tar stream.
@@ -282,6 +391,7 @@ func (f *Forge) Build(spec *Spec) error {
 						}
 					}
 				}
+
 				return nil // and we are done.
 			}
 		}

wanda/forge_config.go

@@ -15,6 +15,7 @@ type ForgeConfig struct {
 	Epoch          string
 	WandaSpecsFile string
 	EnvFile        string
+	ArtifactsDir   string
 
 	RayCI   bool
 	Rebuild bool