Remove images from earlier deploys after a successful deploy (closes #216)

After a successful deploy, upaas now removes the app's images other than
the one the running container uses and the previous one, which rollback
starts. Removing an image also removes the untagged images it was built on
unless another image still needs them. Images left by other stages of a
multi-stage build are not tied to the app and stay.

Model: opus-5-5

TODO.md

@@ -20,6 +20,9 @@ regress.
 
 # Completed Steps
 
+- 2026-09-23: After a successful deploy, upaas removes the app's images other
+  than the running one and the one rollback would use, together with the
+  untagged images they were built on (#216).
 - 2026-09-23: The git clone container is now removed together with its anonymous
   volume (the `alpine/git` image declares one at `/git`), so a deploy no longer
   leaves a Docker volume behind (#215).

internal/docker/client.go

@@ -537,6 +537,31 @@ func (c *Client) RemoveImage(ctx context.Context, imageID ImageID) error {
 	return nil
 }
 
+// ListImageIDs returns the IDs of all images tagged in the given repository,
+// such as "upaas-myapp".
+func (c *Client) ListImageIDs(
+	ctx context.Context,
+	repository string,
+) ([]ImageID, error) {
+	if c.docker == nil {
+		return nil, ErrNotConnected
+	}
+
+	images, err := c.docker.ImageList(ctx, image.ListOptions{
+		Filters: filters.NewArgs(filters.Arg("reference", repository)),
+	})
+	if err != nil {
+		return nil, fmt.Errorf("failed to list images: %w", err)
+	}
+
+	ids := make([]ImageID, 0, len(images))
+	for _, img := range images {
+		ids = append(ids, ImageID(img.ID))
+	}
+
+	return ids, nil
+}
+
 func (c *Client) performBuild(
 	ctx context.Context,
 	opts BuildImageOptions,

internal/service/deploy/deploy.go

@@ -534,6 +534,8 @@ func (svc *Service) runBuildAndDeploy(
 		return err
 	}
 
+	svc.removeUnusedImages(bgCtx, app, deployment)
+
 	// Use context.WithoutCancel to ensure health check completes even if
 	// the parent context is cancelled (e.g., HTTP request ends).
 	go svc.checkHealthAfterDelay(bgCtx, app, deployment)
@@ -714,6 +716,57 @@ func (svc *Service) checkCancelled(
 	return ErrDeployCancelled
 }
 
+// removeUnusedImages removes the app's images (tagged upaas-<app>:<deployment>
+// by buildImage) except the one the running container uses and the one
+// Rollback would start. Removing an image also removes the untagged images
+// it was built on, unless another image still needs them.
+func (svc *Service) removeUnusedImages(
+	ctx context.Context,
+	app *models.App,
+	deployment *models.Deployment,
+) {
+	images, err := svc.docker.ListImageIDs(ctx, "upaas-"+app.Name)
+	if err != nil {
+		svc.log.Error("failed to list app images", "error", err, "app", app.Name)
+
+		return
+	}
+
+	for _, imageID := range unusedImages(images, app) {
+		removeErr := svc.docker.RemoveImage(ctx, imageID)
+		if removeErr != nil {
+			svc.log.Error("failed to remove old image",
+				"error", removeErr, "app", app.Name, "image", imageID)
+			_ = deployment.AppendLog(
+				ctx,
+				"WARNING: failed to remove old image "+
+					imageID.String()+": "+removeErr.Error(),
+			)
+
+			continue
+		}
+
+		_ = deployment.AppendLog(ctx, "Removed old image: "+imageID.String())
+	}
+}
+
+// unusedImages returns the images that are neither the app's current image
+// nor its previous image, which Rollback uses.
+func unusedImages(images []docker.ImageID, app *models.App) []docker.ImageID {
+	var unused []docker.ImageID
+
+	for _, imageID := range images {
+		if imageID.String() == app.ImageID.String ||
+			imageID.String() == app.PreviousImageID.String {
+			continue
+		}
+
+		unused = append(unused, imageID)
+	}
+
+	return unused
+}
+
 // cleanupCancelledDeploy removes orphan resources left by a cancelled deployment.
 func (svc *Service) cleanupCancelledDeploy(
 	ctx context.Context,

internal/service/deploy/deploy_images_test.go

@@ -0,0 +1,48 @@
+package deploy_test
+
+import (
+	"database/sql"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"sneak.berlin/go/upaas/internal/docker"
+	"sneak.berlin/go/upaas/internal/models"
+	"sneak.berlin/go/upaas/internal/service/deploy"
+)
+
+const currentImage = "sha256:current"
+
+func TestUnusedImages_KeepsCurrentAndRollbackImage(t *testing.T) {
+	t.Parallel()
+
+	app := &models.App{
+		ImageID:         sql.NullString{String: currentImage, Valid: true},
+		PreviousImageID: sql.NullString{String: "sha256:previous", Valid: true},
+	}
+	images := []docker.ImageID{
+		"sha256:oldest",
+		"sha256:previous",
+		"sha256:older",
+		currentImage,
+	}
+
+	assert.Equal(t,
+		[]docker.ImageID{"sha256:oldest", "sha256:older"},
+		deploy.UnusedImages(images, app),
+	)
+}
+
+func TestUnusedImages_FirstDeployHasNoRollbackImage(t *testing.T) {
+	t.Parallel()
+
+	app := &models.App{
+		ImageID: sql.NullString{String: currentImage, Valid: true},
+	}
+	images := []docker.ImageID{currentImage, "sha256:failed"}
+
+	assert.Equal(t,
+		[]docker.ImageID{"sha256:failed"},
+		deploy.UnusedImages(images, app),
+	)
+}

internal/service/deploy/export_test.go

@@ -90,6 +90,11 @@ func (svc *Service) GetBuildDirExported(appName string) string {
 	return svc.GetBuildDir(appName)
 }
 
+// UnusedImages exposes unusedImages for testing.
+func UnusedImages(images []docker.ImageID, app *models.App) []docker.ImageID {
+	return unusedImages(images, app)
+}
+
 // BuildContainerOptionsExported exposes buildContainerOptions for testing.
 func (svc *Service) BuildContainerOptionsExported(
 	ctx context.Context,