fix(smolmachines): use containerized crane to push, bypassing docker daemon's HTTPS preference

The previous fix (`host.docker.internal:<port>` for daemon-side
push) still failed:

  Get "https://host.docker.internal:53958/v2/":
    http: server gave HTTP response to HTTPS client

`host.docker.internal` is reachable from Docker Desktop's daemon
VM but isn't in the daemon's default insecure-registries CIDRs
(only `::1/128` and `127.0.0.0/8` are), so docker push tries
HTTPS, hits a plain-HTTP registry, and refuses. The daemon.json
fix (`"insecure-registries": ["host.docker.internal"]`) works
but is a one-time manual step in Docker Desktop's UI — not
something we can do for the user.

Sidestep the daemon push entirely:

  1. docker build (as before) — local layer cache makes
     no-change rebuilds cheap.
  2. docker save the image to a per-digest tarball alongside the
     cached `.smolmachine`.
  3. Start an ephemeral registry container on a per-session
     docker network, with `-p :5000` so the host can also reach
     it for the pack step.
  4. docker run a one-shot crane container on the SAME network,
     mount the tarball, `crane push --insecure /img.tar
     <registry-container>:5000/...`. Container DNS resolves the
     registry on the network; `--insecure` forces plain HTTP.
  5. `smolvm pack create --image localhost:<host port>/...` from
     the host. smolvm's bundled crane auto-falls-back to HTTP
     for localhost addresses, so no insecure-registries config
     is needed on that side.
  6. Tear down everything; reap the tarball (registries hold the
     same bytes, no need to keep both around).

Net effect: the docker daemon never does an HTTP/HTTPS-policy
decision on our behalf. `docker push` is gone from the prepare
path; `docker save`, `docker network create`, `docker run` (for
registry + crane) replace it.

Tested end-to-end on Docker Desktop / macOS: `_ensure_smolmachine
("claude-bottle:latest")` produces a 204MB
`.smolmachine.smolmachine` artifact.

Adds:
- backend/docker/util.py:save() — thin docker save wrapper.
- local_registry.crane_push_tarball() — one-shot crane run on
  the registry's network.
- CRANE_IMAGE constant pinned by digest
  (gcr.io/go-containerregistry/crane@sha256:0ae17ecb...).

Removes:
- backend/docker/util.py:tag() / push() — unused without daemon
  push.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

claude_bottle/backend/docker/util.py

@@ -166,18 +166,15 @@ def image_id(ref: str) -> str:
     return r.stdout.strip()
 
 
-def tag(src: str, dst: str) -> None:
-    """`docker tag SRC DST`. Idempotent. Used by smolmachines prepare
-    to retag the locally-built image into a localhost:<port>/... ref
-    that the ephemeral registry will accept."""
-    subprocess.run(["docker", "tag", src, dst], check=True)
-
-
-def push(ref: str) -> None:
-    """`docker push REF`. Used by smolmachines prepare to push the
-    agent image into the ephemeral local registry so smolvm's crane
-    backend can pull it."""
-    subprocess.run(["docker", "push", ref], check=True)
+def save(ref: str, output: str) -> None:
+    """`docker save REF -o OUTPUT`. Writes a tarball of the image
+    layers + manifest to the host path. Used by smolmachines
+    prepare to hand the agent image to a containerized crane that
+    pushes it to the ephemeral registry — bypassing the docker
+    daemon's `docker push` (which on Docker Desktop can't reach a
+    host-loopback registry and refuses plain-HTTP pushes to
+    non-loopback hosts)."""
+    subprocess.run(["docker", "save", ref, "-o", output], check=True)
 
 
 def _silent_run(cmd: Iterable[str]) -> int:

claude_bottle/backend/smolmachines/local_registry.py

@@ -1,40 +1,37 @@
 """Ephemeral local OCI registry for the smolmachines agent-image
 conversion path (PRD 0023 chunk 4c).
 
-`smolvm pack create --image <ref>` only accepts registry refs — it
-can't read the local docker daemon's image cache, an OCI layout
-directory, or a `docker save` tarball. To convert the agent's
-Dockerfile-built image into a `.smolmachine` artifact we run a
-short-lived `registry:2.8.3` container, push the locally-tagged
-image into it, and let smolvm pull from there. The registry
-container is torn down as soon as the pack completes.
+`smolvm pack create --image <ref>` only accepts OCI registry refs
+— it can't read the local docker daemon's image cache, an OCI
+layout directory, or a `docker save` tarball. To convert the
+agent's Dockerfile-built image into a `.smolmachine` artifact we
+spin up a short-lived `registry:2.8.3` container alongside a
+`crane` helper container on a private docker network, push via
+`crane push --insecure <tarball> <registry-container>:5000/...`,
+and let smolvm pull from the registry's published host port. The
+network + both containers are torn down after the pack completes.
 
-Two routing hostnames, one registry container. On Docker Desktop
-(macOS/Windows) the docker daemon runs inside its own Linux VM,
-so its `localhost` is *not* the host's loopback — a registry
-bound to `127.0.0.1::<port>` on the host is unreachable from the
-daemon side, and `docker push` fails with `context deadline
-exceeded`. The fix: bind to all interfaces so both routes work,
-and yield two refs:
+Why this two-container dance instead of plain `docker push`:
+- Docker Desktop's daemon runs in its own Linux VM, so its
+  `localhost` is not the host's loopback. A registry bound to
+  the host's 127.0.0.1 is unreachable from the daemon side.
+- `host.docker.internal` is reachable from the daemon but isn't
+  in Docker's default insecure-registries CIDRs (only `::1/128`
+  and `127.0.0.0/8` are), so `docker push` to it tries HTTPS,
+  hits a plain-HTTP registry, and dies with
+  `http: server gave HTTP response to HTTPS client`. Adding
+  `host.docker.internal` to daemon.json works but is a one-time
+  manual step the user has to do in Docker Desktop's UI.
+- Going through a docker network sidesteps the host-vs-daemon
+  loopback mismatch (crane and registry containers see each
+  other on the network) AND the HTTPS preference (crane has an
+  `--insecure` flag that forces plain HTTP).
 
-  - `daemon_endpoint`: how the docker CLI/daemon dials the
-    registry (`host.docker.internal:<port>` on Docker Desktop,
-    `localhost:<port>` on a native Linux daemon that shares the
-    host's network namespace).
-  - `host_endpoint`: how `smolvm pack create` (a host process)
-    dials the registry. Always `localhost:<port>` — the port
-    binding includes loopback either way.
-
-The registry stores images by repo+tag; the hostname in the ref
-is just routing, so a push to `host.docker.internal:<port>/cb:abc`
-and a pull of `localhost:<port>/cb:abc` hit the same stored
-blob.
-
-Trade-off: binding to all interfaces puts the registry on every
-network interface briefly (~5-10s during prepare). The agent
-image we push is built from the repo's public Dockerfile — no
-secrets in it — and the user is on their own machine; the LAN
-exposure window is short and the contents non-sensitive."""
+The registry is also published on a random host port so smolvm
+— a host process — can pull from `localhost:<port>` via Docker's
+port-forward. smolvm's bundled crane auto-falls-back to HTTP for
+localhost addresses, so no insecure-registries config is needed
+on that side either."""
 
 from __future__ import annotations
 
@@ -58,106 +55,150 @@ REGISTRY_IMAGE = os.environ.get(
 )
 
 
+# gcr.io/go-containerregistry/crane:latest, pinned by digest. ~10MB,
+# stable upstream from Google; we only invoke `crane push --insecure`
+# against a localhost-equivalent registry, so the trust surface is
+# narrow.
+CRANE_IMAGE = os.environ.get(
+    "CLAUDE_BOTTLE_CRANE_IMAGE",
+    "gcr.io/go-containerregistry/crane@sha256:0ae17ecb34315aa7cbff28f6eddee3b7adae0b2f90101260d990804db1eb0084",
+)
+
+
+# Internal port the registry binds to inside its container — fixed
+# by the registry:2 image. The host-side mapping is random.
+_REGISTRY_CONTAINER_PORT = "5000"
+
+
 # How long to wait for the registry's HTTP layer to bind before
-# giving up. Two seconds is empirically enough; bumping to 10s leaves
-# headroom for slow CI runners without making the failure mode chatty.
+# giving up. Two seconds is empirically enough; 10s leaves headroom
+# for slow CI runners without making the failure mode chatty.
 _READY_TIMEOUT_S = 10.0
 
 
 @dataclass(frozen=True)
-class RegistryEndpoints:
-    """The two `<host>:<port>` strings to embed in image refs. They
-    point at the same registry container; only the routing
-    hostname differs."""
+class RegistryHandle:
+    """Everything callers need to push to + pull from the ephemeral
+    registry.
 
-    daemon_endpoint: str
-    host_endpoint: str
+    `network` is the per-session docker network — a `crane push`
+    container has to join it to reach the registry by name.
+    `push_endpoint` is the `<host>:<port>` form to embed in image
+    refs given to the crane push container (resolves via docker
+    network DNS). `pull_endpoint` is the `<host>:<port>` form a
+    host process (smolvm) uses; the registry's host port mapping
+    backs this."""
+
+    network: str
+    push_endpoint: str
+    pull_endpoint: str
 
 
 @contextmanager
-def ephemeral_registry() -> Iterator[RegistryEndpoints]:
-    """Bring up a `registry:2.8.3` container on a random host port,
-    yield the daemon-side + host-side endpoints, force-remove the
-    container on exit.
+def ephemeral_registry() -> Iterator[RegistryHandle]:
+    """Bring up a per-session docker network + a `registry:2.8.3`
+    container on it (published on a random host port), yield a
+    `RegistryHandle`, force-remove both on exit.
 
     The container is started with `--rm` so a clean exit cleans up
     on its own; the `finally` block force-removes on abnormal exit
     (the calling process crashes between yield and close)."""
-    name = f"claude-bottle-registry-{uuid.uuid4().hex[:12]}"
+    session_id = uuid.uuid4().hex[:12]
+    network = f"claude-bottle-registry-net-{session_id}"
+    registry_name = f"claude-bottle-registry-{session_id}"
+
     subprocess.run(
-        [
-            "docker", "run", "-d", "--rm",
-            "--name", name,
-            # `-p :5000` (no IP prefix) binds the container's port
-            # 5000 on a random host port across all interfaces. The
-            # registry container itself listens on 0.0.0.0:5000
-            # internally; binding to all interfaces is necessary for
-            # Docker Desktop's daemon to reach it via
-            # host.docker.internal — a 127.0.0.1-only host binding
-            # is invisible to a daemon running in its own VM.
-            "-p", "5000",
-            REGISTRY_IMAGE,
-        ],
+        ["docker", "network", "create", network],
         check=True,
         capture_output=True,
     )
     try:
-        port = _host_port(name)
-        _wait_ready(port)
-        daemon_host = _daemon_side_hostname()
-        yield RegistryEndpoints(
-            daemon_endpoint=f"{daemon_host}:{port}",
-            host_endpoint=f"localhost:{port}",
+        subprocess.run(
+            [
+                "docker", "run", "-d", "--rm",
+                "--name", registry_name,
+                "--network", network,
+                # `-p :5000` (no IP prefix) binds the container's
+                # port 5000 on a random host port across all
+                # interfaces. The host side reaches the registry
+                # via this port — smolvm's `pack create` pulls from
+                # `localhost:<port>` and the docker port-forward
+                # routes there.
+                "-p", _REGISTRY_CONTAINER_PORT,
+                REGISTRY_IMAGE,
+            ],
+            check=True,
+            capture_output=True,
         )
+        try:
+            port = _host_port(registry_name)
+            _wait_ready(port)
+            yield RegistryHandle(
+                network=network,
+                push_endpoint=f"{registry_name}:{_REGISTRY_CONTAINER_PORT}",
+                pull_endpoint=f"localhost:{port}",
+            )
+        finally:
+            subprocess.run(
+                ["docker", "rm", "-f", registry_name],
+                check=False,
+                capture_output=True,
+            )
     finally:
         subprocess.run(
-            ["docker", "rm", "-f", name],
+            ["docker", "network", "rm", network],
             check=False,
             capture_output=True,
         )
 
 
-def _daemon_side_hostname() -> str:
-    """Pick the hostname the docker daemon should use to dial the
-    registry. On Docker Desktop the daemon runs in its own Linux
-    VM and only sees the host via `host.docker.internal`; on
-    native Linux the daemon shares the host's network namespace
-    and `localhost` works.
+def crane_push_tarball(handle: RegistryHandle, tarball_path: str, ref: str) -> None:
+    """Run `crane push --insecure <tarball> <ref>` inside a one-shot
+    container on the registry's docker network. `ref` should
+    reference the registry by `handle.push_endpoint` so the crane
+    container resolves it via docker network DNS.
 
-    `docker info --format '{{.OperatingSystem}}'` returns
-    `"Docker Desktop"` on macOS / Windows Desktop installs (and on
-    Linux Desktop, which also uses a VM). Anything else (e.g.
-    `"Debian GNU/Linux 12 (bookworm)"`) is a native daemon."""
+    Doesn't go through `docker push` to avoid the Docker-Desktop
+    daemon's HTTPS preference for non-loopback hostnames — crane's
+    `--insecure` flag forces plain HTTP, which is what the
+    registry container speaks."""
     r = subprocess.run(
-        ["docker", "info", "--format", "{{.OperatingSystem}}"],
-        capture_output=True,
-        text=True,
-        check=False,
-    )
-    operating_system = (r.stdout or "").strip()
-    if operating_system == "Docker Desktop":
-        return "host.docker.internal"
-    return "localhost"
-
-
-def _host_port(name: str) -> int:
-    """Resolve the host-side port docker mapped to the registry's
-    container port 5000. `docker port <name> 5000/tcp` returns one
-    or more `host:port` lines (one per address family) — we take
-    the first IPv4 line."""
-    r = subprocess.run(
-        ["docker", "port", name, "5000/tcp"],
+        [
+            "docker", "run", "--rm",
+            "--network", handle.network,
+            "-v", f"{tarball_path}:/img.tar:ro",
+            CRANE_IMAGE,
+            "push", "--insecure", "/img.tar", ref,
+        ],
         capture_output=True,
         text=True,
         check=False,
     )
     if r.returncode != 0:
         die(
-            f"docker port {name} 5000/tcp failed: "
+            f"crane push of {tarball_path!r} to {ref!r} failed: "
+            f"{(r.stderr or r.stdout or '').strip() or '<no output>'}"
+        )
+
+
+def _host_port(name: str) -> int:
+    """Resolve the host-side port docker mapped to the registry's
+    container port. `docker port <name> 5000/tcp` returns one or
+    more `host:port` lines (one per address family) — we take the
+    first."""
+    r = subprocess.run(
+        ["docker", "port", name, f"{_REGISTRY_CONTAINER_PORT}/tcp"],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if r.returncode != 0:
+        die(
+            f"docker port {name} {_REGISTRY_CONTAINER_PORT}/tcp failed: "
             f"{(r.stderr or '').strip() or '<no stderr>'}"
         )
-    # `0.0.0.0:54321\n[::]:54321\n` — take the first line, split
-    # on the last colon to handle either IPv4 or IPv6 host syntax.
+    # `0.0.0.0:54321\n[::]:54321\n` — split on the last colon to
+    # handle either IPv4 or IPv6 host syntax.
     line = (r.stdout or "").splitlines()[0].strip()
     _, _, port_str = line.rpartition(":")
     try:
@@ -168,15 +209,15 @@ def _host_port(name: str) -> int:
 
 
 def _wait_ready(port: int) -> None:
-    """Block until the registry's HTTP layer accepts a TCP connection
-    on `127.0.0.1:<port>`, or `_READY_TIMEOUT_S` elapses.
+    """Block until the registry's HTTP layer accepts a TCP
+    connection on `127.0.0.1:<port>`, or `_READY_TIMEOUT_S`
+    elapses.
 
     A successful TCP connect is sufficient — registry:2.8.3 binds
     after it's ready to serve `/v2/` requests, so the push that
     follows will land on a working server. We probe loopback
-    specifically (not host.docker.internal) because this helper
-    runs on the host, and 0.0.0.0-bound ports are reachable via
-    127.0.0.1 too."""
+    specifically (not via the docker network) because this helper
+    runs on the host."""
     deadline = time.monotonic() + _READY_TIMEOUT_S
     last_err: Exception | None = None
     while time.monotonic() < deadline:

claude_bottle/backend/smolmachines/prepare.py

@@ -34,7 +34,7 @@ from ...pipelock import PipelockProxy
 from ...supervise import Supervise
 from . import smolvm as _smolvm
 from .bottle_plan import SmolmachinesBottlePlan
-from .local_registry import ephemeral_registry
+from .local_registry import crane_push_tarball, ephemeral_registry
 from .util import smolmachines_bundle_subnet, smolmachines_preflight
 
 
@@ -185,15 +185,14 @@ def _ensure_smolmachine(image_ref: str) -> Path:
     it; the sidecar is the actual artifact).
 
     Conversion path: `docker build` (the existing layer cache
-    makes no-change rebuilds cheap) → `docker tag` + `docker push`
-    using the daemon-side endpoint (`host.docker.internal:<port>`
-    on Docker Desktop, `localhost:<port>` on native Linux) →
-    `smolvm pack create --image <host endpoint>` using the
-    host-side endpoint (always `localhost:<port>` — smolvm is a
-    host process) → tear down the registry. The two endpoints
-    route to the same registry container; only the hostname
-    differs because the docker daemon (on Docker Desktop) doesn't
-    share the host's loopback.
+    makes no-change rebuilds cheap) → `docker save` to a tarball
+    → spin up an ephemeral registry on a private docker network →
+    `crane push --insecure` from a one-shot container on the same
+    network → `smolvm pack create --image localhost:<host port>/...`
+    → tear down the registry + network. The crane push detour
+    sidesteps the Docker-Desktop daemon's HTTPS preference for
+    non-loopback registries — see the `local_registry` module
+    docstring for the gory details.
 
     Each pack-create costs several seconds even on a hot cache,
     so we skip the whole pipeline when the cached sidecar is
@@ -208,10 +207,17 @@ def _ensure_smolmachine(image_ref: str) -> Path:
     sidecar = _SMOLMACHINE_CACHE_DIR / f"{digest}.smolmachine.smolmachine"
     if sidecar.is_file():
         return sidecar
-    with ephemeral_registry() as endpoints:
-        push_ref = f"{endpoints.daemon_endpoint}/claude-bottle:{digest}"
-        pack_ref = f"{endpoints.host_endpoint}/claude-bottle:{digest}"
-        docker_mod.tag(image_ref, push_ref)
-        docker_mod.push(push_ref)
-        _smolvm.pack_create(pack_ref, binary)
+    tarball = _SMOLMACHINE_CACHE_DIR / f"{digest}.image.tar"
+    docker_mod.save(image_ref, str(tarball))
+    try:
+        with ephemeral_registry() as handle:
+            push_ref = f"{handle.push_endpoint}/claude-bottle:{digest}"
+            pack_ref = f"{handle.pull_endpoint}/claude-bottle:{digest}"
+            crane_push_tarball(handle, str(tarball), push_ref)
+            _smolvm.pack_create(pack_ref, binary)
+    finally:
+        # Tarball is ~500MB-1GB for the agent image; reclaim once
+        # the smolmachine artifact exists. The artifact itself is
+        # the long-lived cache entry.
+        tarball.unlink(missing_ok=True)
     return sidecar

tests/unit/test_docker_util_image.py

@@ -54,26 +54,18 @@ class TestImageId(unittest.TestCase):
             self.assertIn("missing:tag", die.call_args.args[0])
 
 
-class TestTagPush(unittest.TestCase):
-    def test_tag_runs_docker_tag(self):
+class TestSave(unittest.TestCase):
+    def test_save_runs_docker_save(self):
         with patch.object(
             docker_mod.subprocess, "run", return_value=_ok(),
         ) as run:
-            docker_mod.tag("claude-bottle:latest", "localhost:5000/cb:abc")
+            docker_mod.save("claude-bottle:latest", "/tmp/img.tar")
         argv = run.call_args.args[0]
         self.assertEqual(
-            ["docker", "tag", "claude-bottle:latest", "localhost:5000/cb:abc"],
+            ["docker", "save", "claude-bottle:latest", "-o", "/tmp/img.tar"],
             argv,
         )
 
-    def test_push_runs_docker_push(self):
-        with patch.object(
-            docker_mod.subprocess, "run", return_value=_ok(),
-        ) as run:
-            docker_mod.push("localhost:5000/cb:abc")
-        argv = run.call_args.args[0]
-        self.assertEqual(["docker", "push", "localhost:5000/cb:abc"], argv)
-
 
 if __name__ == "__main__":
     unittest.main()

tests/unit/test_smolmachines_local_registry.py

@@ -1,9 +1,10 @@
 """Unit: ephemeral local-registry helper (PRD 0023 chunk 4c).
 
-The helper brings up a `registry:2.8.3` container on a random
-host port, yields a `(daemon_endpoint, host_endpoint)` pair, and
-tears the container down on exit. Tests mock `subprocess.run` +
-`socket.create_connection` so they run without docker."""
+The helper brings up a `registry:2.8.3` container on a private
+docker network with a random host-side port, yields a
+`RegistryHandle`, and tears the container + network down on exit.
+Tests mock `subprocess.run` + `socket.create_connection` so they
+run without docker."""
 
 from __future__ import annotations
 
@@ -20,71 +21,56 @@ def _ok(stdout: str = "", stderr: str = "") -> subprocess.CompletedProcess:
     )
 
 
-# `docker info` always runs once per ephemeral_registry() to pick
-# the daemon-side hostname; the run sequence is therefore
-# (docker run, docker port, docker info, docker rm). Helpers below
-# build a stock side_effect that covers all four.
-def _stock_run_sequence(
-    *,
-    port: str = "0.0.0.0:54321\n",
-    operating_system: str = "Docker Desktop\n",
-):
+def _fail(stderr: str = "boom") -> subprocess.CompletedProcess:
+    return subprocess.CompletedProcess(
+        args=[], returncode=1, stdout="", stderr=stderr,
+    )
+
+
+# Run sequence per ephemeral_registry() call:
+#   docker network create  -> ok
+#   docker run -d (registry)  -> ok (container id)
+#   docker port (host port)  -> ok (mapping line)
+#   docker rm -f (registry)  -> ok (in finally)
+#   docker network rm  -> ok (in finally)
+def _stock_run_sequence(port_line: str = "0.0.0.0:54321\n"):
     return [
-        _ok(stdout="<container-id>\n"),  # docker run
-        _ok(stdout=port),                # docker port
-        _ok(stdout=operating_system),    # docker info
-        _ok(),                           # docker rm -f
+        _ok(),                            # docker network create
+        _ok(stdout="<container-id>\n"),   # docker run
+        _ok(stdout=port_line),            # docker port
+        _ok(),                            # docker rm -f
+        _ok(),                            # docker network rm
     ]
 
 
 class TestEphemeralRegistry(unittest.TestCase):
-    def test_yields_endpoints_with_docker_desktop_routing(self):
-        # On Docker Desktop the daemon runs in its own VM, so the
-        # registry has to be addressed by host.docker.internal for
-        # docker push to work; smolvm (host process) still uses
-        # localhost.
+    def test_yields_handle_with_network_and_endpoints(self):
         with patch.object(
             local_registry.subprocess, "run",
-            side_effect=_stock_run_sequence(operating_system="Docker Desktop\n"),
-        ), patch.object(
+            side_effect=_stock_run_sequence(),
+        ) as run, patch.object(
             local_registry.socket, "create_connection",
             return_value=_FakeSocket(),
         ):
-            with local_registry.ephemeral_registry() as endpoints:
-                self.assertEqual(
-                    "host.docker.internal:54321", endpoints.daemon_endpoint,
+            with local_registry.ephemeral_registry() as handle:
+                # push_endpoint points at the registry container by
+                # its docker-network name on its container port.
+                self.assertTrue(
+                    handle.push_endpoint.startswith(
+                        "claude-bottle-registry-"
+                    )
                 )
-                self.assertEqual(
-                    "localhost:54321", endpoints.host_endpoint,
+                self.assertTrue(handle.push_endpoint.endswith(":5000"))
+                # pull_endpoint is the host-side mapping for smolvm.
+                self.assertEqual("localhost:54321", handle.pull_endpoint)
+                # network name is the per-session bridge crane joins.
+                self.assertTrue(
+                    handle.network.startswith("claude-bottle-registry-net-")
                 )
+        # docker network create + docker run + docker port + rm -f + network rm
+        self.assertEqual(5, run.call_count)
 
-    def test_yields_endpoints_with_native_linux_routing(self):
-        # On a native Linux daemon the daemon shares the host's
-        # network namespace, so localhost reaches the registry from
-        # both sides.
-        with patch.object(
-            local_registry.subprocess, "run",
-            side_effect=_stock_run_sequence(
-                operating_system="Debian GNU/Linux 12 (bookworm)\n",
-            ),
-        ), patch.object(
-            local_registry.socket, "create_connection",
-            return_value=_FakeSocket(),
-        ):
-            with local_registry.ephemeral_registry() as endpoints:
-                self.assertEqual(
-                    "localhost:54321", endpoints.daemon_endpoint,
-                )
-                self.assertEqual(
-                    "localhost:54321", endpoints.host_endpoint,
-                )
-
-    def test_runs_docker_with_all_interface_bind(self):
-        # `-p 5000` (no IP prefix) binds the container's port 5000
-        # on a random host port across all interfaces — needed so
-        # Docker Desktop's daemon can reach the registry via
-        # host.docker.internal. The 127.0.0.1-only bind we used
-        # previously was invisible to the daemon's VM.
+    def test_registry_run_publishes_random_port_across_interfaces(self):
         with patch.object(
             local_registry.subprocess, "run",
             side_effect=_stock_run_sequence(),
@@ -94,16 +80,19 @@ class TestEphemeralRegistry(unittest.TestCase):
         ):
             with local_registry.ephemeral_registry():
                 pass
-
-        run_argv = run.call_args_list[0].args[0]
+        # second call is the docker run for the registry
+        run_argv = run.call_args_list[1].args[0]
         self.assertEqual(["docker", "run"], run_argv[:2])
         self.assertIn("--rm", run_argv)
+        # `-p 5000` (no IP prefix) — needed so the host-published
+        # port is reachable from BOTH the host (for smolvm) and the
+        # docker daemon (for the docker port command to find it).
         self.assertIn("5000", run_argv)
-        # Explicitly NOT the loopback-only form — that one's broken
-        # under Docker Desktop.
-        self.assertNotIn("127.0.0.1::5000", run_argv)
+        # And the registry is attached to the same per-session
+        # network the crane push container joins.
+        self.assertIn("--network", run_argv)
 
-    def test_force_removes_container_on_clean_exit(self):
+    def test_force_removes_container_and_network_on_clean_exit(self):
         with patch.object(
             local_registry.subprocess, "run",
             side_effect=_stock_run_sequence(),
@@ -114,11 +103,13 @@ class TestEphemeralRegistry(unittest.TestCase):
             with local_registry.ephemeral_registry():
                 pass
 
-        # Last call is `docker rm -f <name>`.
-        last_argv = run.call_args_list[-1].args[0]
-        self.assertEqual(["docker", "rm", "-f"], last_argv[:3])
+        # Last two calls are `docker rm -f <container>` then
+        # `docker network rm <network>`.
+        argvs = [c.args[0] for c in run.call_args_list]
+        self.assertEqual(["docker", "rm", "-f"], argvs[-2][:3])
+        self.assertEqual(["docker", "network", "rm"], argvs[-1][:3])
 
-    def test_force_removes_container_on_exception_inside_with(self):
+    def test_force_removes_on_exception_inside_with(self):
         with patch.object(
             local_registry.subprocess, "run",
             side_effect=_stock_run_sequence(),
@@ -130,12 +121,12 @@ class TestEphemeralRegistry(unittest.TestCase):
                 with local_registry.ephemeral_registry():
                     raise RuntimeError("inside with")
 
-        # rm -f still ran on exception.
-        last_argv = run.call_args_list[-1].args[0]
-        self.assertEqual(["docker", "rm", "-f"], last_argv[:3])
+        # Both teardowns still ran.
+        argvs = [c.args[0] for c in run.call_args_list]
+        self.assertEqual(["docker", "rm", "-f"], argvs[-2][:3])
+        self.assertEqual(["docker", "network", "rm"], argvs[-1][:3])
 
-    def test_wait_ready_times_out_when_socket_never_connects(self):
-        # Drop the timeout to a value that fits the test budget.
+    def test_wait_ready_times_out(self):
         with patch.object(local_registry, "_READY_TIMEOUT_S", 0.1), patch.object(
             local_registry.subprocess, "run",
             side_effect=_stock_run_sequence(),
@@ -150,21 +141,26 @@ class TestEphemeralRegistry(unittest.TestCase):
                 with local_registry.ephemeral_registry():
                     self.fail("yield reached despite unreachable registry")
             die.assert_called_once()
-        # rm -f still ran (cleanup goes through the finally block).
-        last_argv = run.call_args_list[-1].args[0]
-        self.assertEqual(["docker", "rm", "-f"], last_argv[:3])
+        # Teardown still ran via the finally blocks.
+        argvs = [c.args[0] for c in run.call_args_list]
+        self.assertEqual(["docker", "rm", "-f"], argvs[-2][:3])
+        self.assertEqual(["docker", "network", "rm"], argvs[-1][:3])
 
-    def test_unique_container_name_per_call(self):
-        names: list[str] = []
+    def test_unique_session_ids_per_call(self):
+        sessions: list[tuple[str, str]] = []
 
         def capture(argv, *a, **kw):
+            if argv[:3] == ["docker", "network", "create"]:
+                return _ok()
             if argv[:2] == ["docker", "run"]:
-                names.append(argv[argv.index("--name") + 1])
+                # `--name <registry-name>` and `--network <net-name>`
+                # both encode the session id.
+                name = argv[argv.index("--name") + 1]
+                network = argv[argv.index("--network") + 1]
+                sessions.append((name, network))
                 return _ok(stdout="cid\n")
             if argv[:2] == ["docker", "port"]:
                 return _ok(stdout="0.0.0.0:1\n")
-            if argv[:2] == ["docker", "info"]:
-                return _ok(stdout="Docker Desktop\n")
             return _ok()
 
         with patch.object(
@@ -178,10 +174,64 @@ class TestEphemeralRegistry(unittest.TestCase):
             with local_registry.ephemeral_registry():
                 pass
 
-        self.assertEqual(2, len(names))
-        self.assertNotEqual(names[0], names[1])
-        for n in names:
-            self.assertTrue(n.startswith("claude-bottle-registry-"))
+        self.assertEqual(2, len(sessions))
+        self.assertNotEqual(sessions[0], sessions[1])
+
+
+class TestCranePushTarball(unittest.TestCase):
+    def test_runs_crane_container_on_registry_network_with_insecure_flag(self):
+        handle = local_registry.RegistryHandle(
+            network="cb-registry-net-x",
+            push_endpoint="cb-registry-x:5000",
+            pull_endpoint="localhost:54321",
+        )
+        with patch.object(
+            local_registry.subprocess, "run", return_value=_ok(),
+        ) as run:
+            local_registry.crane_push_tarball(
+                handle, "/tmp/img.tar", "cb-registry-x:5000/cb:abc",
+            )
+
+        argv = run.call_args.args[0]
+        # Joined to the same docker network so it can reach the
+        # registry by container name (no host port-forward needed
+        # for the push leg).
+        self.assertEqual("docker", argv[0])
+        self.assertEqual("run", argv[1])
+        self.assertIn("--rm", argv)
+        self.assertIn("--network", argv)
+        self.assertEqual(
+            "cb-registry-net-x", argv[argv.index("--network") + 1],
+        )
+        # The tarball is mounted read-only at /img.tar.
+        self.assertIn("-v", argv)
+        self.assertIn("/tmp/img.tar:/img.tar:ro", argv)
+        # And the crane command itself uses --insecure so plain
+        # HTTP is allowed against the registry container.
+        self.assertIn("push", argv)
+        self.assertIn("--insecure", argv)
+        self.assertIn("/img.tar", argv)
+        self.assertIn("cb-registry-x:5000/cb:abc", argv)
+
+    def test_dies_when_crane_returns_non_zero(self):
+        handle = local_registry.RegistryHandle(
+            network="cb-net", push_endpoint="cb:5000", pull_endpoint="localhost:1",
+        )
+        with patch.object(
+            local_registry.subprocess, "run",
+            return_value=_fail("push failed"),
+        ), patch.object(
+            local_registry, "die", side_effect=SystemExit("die"),
+        ) as die:
+            with self.assertRaises(SystemExit):
+                local_registry.crane_push_tarball(
+                    handle, "/tmp/img.tar", "cb:5000/cb:abc",
+                )
+            die.assert_called_once()
+            # Error message names what was being pushed where.
+            msg = die.call_args.args[0]
+            self.assertIn("/tmp/img.tar", msg)
+            self.assertIn("cb:5000/cb:abc", msg)
 
 
 class _FakeSocket:

tests/unit/test_smolmachines_prepare_image.py

@@ -40,40 +40,42 @@ class TestEnsureSmolmachine(unittest.TestCase):
             _prepare.docker_mod, "image_id",
             return_value=f"sha256:{digest}fffffffffffffffff",
         ), patch.object(
+            _prepare.docker_mod, "save",
+        ) as save, patch.object(
             _prepare, "ephemeral_registry",
         ) as registry, patch.object(
-            _prepare.docker_mod, "tag",
-        ) as tag, patch.object(
-            _prepare.docker_mod, "push",
+            _prepare, "crane_push_tarball",
         ) as push, patch.object(
             _prepare._smolvm, "pack_create",
         ) as pack:
             result = _prepare._ensure_smolmachine("claude-bottle:latest")
 
         self.assertEqual(sidecar, result)
-        # build still runs (Dockerfile edits land without manual rmi)
+        # build still runs (Dockerfile edits land without manual rmi).
         build.assert_called_once()
-        # No registry, no tag, no push, no pack on cache hit.
+        # No save (500MB tarball), no registry, no push, no pack on
+        # cache hit.
+        save.assert_not_called()
         registry.assert_not_called()
-        tag.assert_not_called()
         push.assert_not_called()
         pack.assert_not_called()
 
-    def test_cache_miss_runs_build_tag_push_pack_in_order(self):
+    def test_cache_miss_runs_build_save_push_pack_in_order(self):
         digest = "0123456789abcdef"
 
-        # ephemeral_registry yields a RegistryEndpoints with two
-        # routing hostnames — daemon-side for docker push,
-        # host-side for smolvm pack create.
+        # ephemeral_registry yields a RegistryHandle with the
+        # docker network + a push endpoint (container DNS) and
+        # pull endpoint (host port-forward).
         from claude_bottle.backend.smolmachines.local_registry import (
-            RegistryEndpoints,
+            RegistryHandle,
         )
 
         class _Reg:
             def __enter__(self_inner):
-                return RegistryEndpoints(
-                    daemon_endpoint="host.docker.internal:54321",
-                    host_endpoint="localhost:54321",
+                return RegistryHandle(
+                    network="cb-net-xyz",
+                    push_endpoint="cb-registry-xyz:5000",
+                    pull_endpoint="localhost:54321",
                 )
             def __exit__(self_inner, *exc):
                 return False
@@ -92,13 +94,13 @@ class TestEnsureSmolmachine(unittest.TestCase):
             _prepare.docker_mod, "image_id",
             return_value=f"sha256:{digest}fffffffffffffffff",
         ), patch.object(
+            _prepare.docker_mod, "save",
+            side_effect=record("save"),
+        ) as save, patch.object(
             _prepare, "ephemeral_registry",
             return_value=_Reg(),
         ), patch.object(
-            _prepare.docker_mod, "tag",
-            side_effect=record("tag"),
-        ) as tag, patch.object(
-            _prepare.docker_mod, "push",
+            _prepare, "crane_push_tarball",
             side_effect=record("push"),
         ) as push, patch.object(
             _prepare._smolvm, "pack_create",
@@ -106,26 +108,27 @@ class TestEnsureSmolmachine(unittest.TestCase):
         ) as pack:
             _prepare._ensure_smolmachine("claude-bottle:latest")
 
-        # build first (no point pushing if the build fails), then
-        # tag → push → pack against the registry endpoints.
-        self.assertEqual(["build", "tag", "push", "pack"], calls)
+        # Build → save → push → pack in that order. No `docker
+        # push` (the daemon's HTTPS-by-default path is what we're
+        # sidestepping).
+        self.assertEqual(["build", "save", "push", "pack"], calls)
 
-        # tag + push target the daemon-side endpoint (host.docker
-        # .internal on Docker Desktop, since the daemon's
-        # localhost is its own VM's loopback).
-        tag_args = tag.call_args.args
-        self.assertEqual("claude-bottle:latest", tag_args[0])
-        self.assertEqual(
-            f"host.docker.internal:54321/claude-bottle:{digest}", tag_args[1],
-        )
+        # docker save targets a per-digest tarball alongside the
+        # cached sidecar.
+        save_args = save.call_args.args
+        self.assertEqual("claude-bottle:latest", save_args[0])
+        self.assertTrue(save_args[1].endswith(f"{digest}.image.tar"))
+
+        # crane push runs against the push_endpoint (container DNS
+        # on the registry network) with the digest as the tag.
         push_args = push.call_args.args
         self.assertEqual(
-            f"host.docker.internal:54321/claude-bottle:{digest}", push_args[0],
+            f"cb-registry-xyz:5000/claude-bottle:{digest}", push_args[2],
         )
-        # pack_create reads from the host-side endpoint (smolvm is
-        # a host process and can only resolve localhost). The
-        # registry stores images by repo+tag, so both endpoints
-        # hit the same blob.
+
+        # pack_create reads from the pull_endpoint (host port-
+        # forward, smolvm is on the host). Same repo+tag, just a
+        # different routing hostname — the registry stores one blob.
         pack_args = pack.call_args.args
         self.assertEqual(
             f"localhost:54321/claude-bottle:{digest}", pack_args[0],