Merge pull request 'fix(smolmachines): docker push fails on Docker Desktop — daemon-side route differs from host loopback' (#74) from fix-local-registry-docker-desktop into main

README.md

@@ -190,6 +190,27 @@ The container is removed automatically when the session ends. If the script
 is killed with SIGKILL the exit trap won't fire and the container may be
 left running; remove it with `docker rm -f <container-name>`.
 
+### Smolmachines backend (experimental, macOS-only)
+
+A second backend runs the agent in a smolvm micro-VM (libkrun) with the
+sidecar bundle still in Docker. Selected via
+`CLAUDE_BOTTLE_BACKEND=smolmachines ./cli.py start <agent>`. Requires
+`smolvm` on PATH (`curl -sSL https://smolmachines.com/install.sh | sh`).
+
+**Known limitation, v1:** smolvm's TSI uses macOS networking, and
+Docker Desktop's container IPs aren't reachable from macOS, so the
+smolmachines bottle dials the sidecar bundle through host loopback
+port-forwards (`127.0.0.1:<random>`). TSI filters by IP only, so the
+allowlist is `127.0.0.1/32` — meaning the agent VM can reach **any
+service bound to macOS's loopback**, not just the bundle's published
+ports. Practical implication: while a smolmachines bottle is running,
+host-local dev services (postgres on 5432, dev servers, etc.) are
+reachable from inside the agent even if you intended them to be
+host-private. The docker backend keeps the bottle on a `--internal`
+docker network and doesn't have this issue. A future revision will
+narrow this via a per-bottle loopback alias + host-side proxy (see
+PRD 0023's "loopback scoping" section).
+
 ## Manifest
 
 Bottles and agents live as Markdown files with YAML frontmatter under

claude_bottle/backend/__init__.py

@@ -119,12 +119,20 @@ class Bottle(ABC):
     def exec_claude(self, argv: list[str], *, tty: bool = True) -> int: ...
 
     @abstractmethod
-    def exec(self, script: str) -> ExecResult:
-        """Run `script` as a POSIX shell script inside the bottle and
-        return the captured stdout/stderr/returncode. The bottle's
-        environment (including HTTPS_PROXY pointing at the pipelock
-        sidecar) is inherited by the child. Non-zero exit does not
-        raise — callers inspect `returncode` themselves."""
+    def exec(self, script: str, *, user: str = "node") -> ExecResult:
+        """Run `script` as a POSIX shell script inside the bottle as
+        `user` (default `node`, matching the agent image's USER
+        directive) and return the captured stdout/stderr/returncode.
+        The bottle's environment (including HTTPS_PROXY pointing at
+        the pipelock sidecar) is inherited by the child. Non-zero
+        exit does not raise — callers inspect `returncode`
+        themselves.
+
+        Pass `user="root"` for shell-outs that need privileged file
+        writes / package install — provisioning calls that need root
+        bypass `Bottle.exec` and use the backend-specific raw
+        machine-exec helper, but the tests have a legitimate use
+        case for arbitrary-user runs."""
 
     @abstractmethod
     def cp_in(self, host_path: str, container_path: str) -> None: ...

claude_bottle/backend/docker/bottle.py

@@ -51,12 +51,15 @@ class DockerBottle(Bottle):
             self.claude_docker_argv(argv, tty=tty), check=False,
         ).returncode
 
-    def exec(self, script: str) -> ExecResult:
+    def exec(self, script: str, *, user: str = "node") -> ExecResult:
         # Pipe via stdin to `sh -s` so the caller never has to worry
         # about quoting; the script source lands inside the container
-        # without crossing argv.
+        # without crossing argv. `-u <user>` overrides the image's
+        # default USER — defaults to `node` which is already the
+        # image's USER, so the explicit flag is a no-op there but
+        # keeps the cross-backend contract uniform.
         result = subprocess.run(
-            ["docker", "exec", "-i", self.name, "sh", "-s"],
+            ["docker", "exec", "-u", user, "-i", self.name, "sh", "-s"],
             input=script,
             capture_output=True,
             text=True,

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/bottle.py

@@ -4,63 +4,130 @@ Routes `exec_claude` / `exec` / `cp_in` through `smolvm machine
 exec` / `smolvm machine cp`. The handle is yielded by `launch`
 and torn down via the surrounding ExitStack on context exit;
 `close` is a no-op idempotent alias so the BottleBackend ABC's
-context-manager contract is satisfied."""
+context-manager contract is satisfied.
+
+User context: `smolvm machine exec` runs commands as root in the
+VM, but the agent image's USER is `node` and claude-code refuses
+to run as root with `--dangerously-skip-permissions`. Both
+`exec_claude` and `exec` switch to the requested user (default
+`node`) via `runuser -u <user> --` and set `HOME` / `USER`
+through `smolvm -e` — avoiding `runuser -l`'s login-shell wiring
+(PAM session setup, /etc/profile sourcing) which can hang on a
+minimal Debian VM with no PAM session config."""
 
 from __future__ import annotations
 
 import subprocess
-import sys
+from typing import Mapping
 
 from .. import Bottle, ExecResult
 from . import smolvm as _smolvm
 
 
+# Per-user env the agent image's USER (node) expects. claude
+# reads ~/.claude.json + writes session state under ~/.claude/;
+# bare `runuser -u` inherits root's HOME=/root, which claude
+# can't write to. Set HOME / USER explicitly through smolvm -e
+# so the child process sees them.
+_HOME_FOR = {
+    "node": "/home/node",
+    "root": "/root",
+}
+
+
+def _env_flags_for(user: str) -> list[str]:
+    home = _HOME_FOR.get(user, f"/home/{user}")
+    return ["-e", f"HOME={home}", "-e", f"USER={user}"]
+
+
+def _guest_env_flags(env: Mapping[str, str]) -> list[str]:
+    """Render `{K: V}` into a flat `-e K=V` argv slice for
+    `smolvm machine exec`. `smolvm machine create -e` set env
+    on PID 1 but it doesn't propagate to fresh exec process
+    trees, so we have to re-pass them every call."""
+    out: list[str] = []
+    for k, v in env.items():
+        out += ["-e", f"{k}={v}"]
+    return out
+
+
 class SmolmachinesBottle(Bottle):
     """Handle returned by `SmolmachinesBottleBackend.launch`. The
     underlying VM lifecycle (create / start / stop / delete) lives
     on the launch ExitStack — this class only routes runtime
     operations to the right `smolvm machine ...` subcommand."""
 
-    def __init__(self, machine_name: str, *, prompt_path: str | None = None) -> None:
+    def __init__(
+        self,
+        machine_name: str,
+        *,
+        prompt_path: str | None = None,
+        guest_env: Mapping[str, str] | None = None,
+    ) -> None:
         self.name = machine_name
         # In-VM path to the agent's prompt file. None when the
         # agent declared no prompt (file still exists; we just
         # don't pass --append-system-prompt-file).
         self._prompt_path = prompt_path
+        # Env vars the agent process needs (HTTPS_PROXY,
+        # CLAUDE_CODE_OAUTH_TOKEN, manifest-declared bottle env, …).
+        # Forwarded on every `smolvm machine exec` via `-e K=V`
+        # because exec doesn't inherit from machine_create's env.
+        self._guest_env = dict(guest_env or {})
 
     def exec_claude(self, argv: list[str], *, tty: bool = True) -> int:
-        """Run `claude` interactively inside the VM. Inherits the
-        operator's terminal (stdin / stdout / stderr) so the
-        session feels native. Blocks until claude exits; returns
-        the in-VM exit code.
+        """Run `claude` interactively inside the VM as the `node`
+        user. Inherits the operator's terminal (stdin / stdout /
+        stderr) so the session feels native. Blocks until claude
+        exits; returns the in-VM exit code.
 
         We bypass the captured-output `machine_exec` helper here
         because that one wraps stdout/stderr in pipes — fine for
         scripted exec, wrong for an interactive shell. Drop down
-        to `subprocess.run` with the TTY inherited."""
+        to `subprocess.run` with the TTY inherited.
+
+        UID switches via `runuser -u node --` (not `-l`) so we
+        avoid login-shell wiring. HOME / USER come from `smolvm
+        -e` instead, which sets them on the process env."""
         flags = ["smolvm", "machine", "exec", "--name", self.name]
         if tty:
             flags += ["-i", "-t"]
+        flags += _env_flags_for("node")
+        flags += _guest_env_flags(self._guest_env)
         claude_argv = ["claude"]
         if self._prompt_path:
             claude_argv += ["--append-system-prompt-file", self._prompt_path]
-        flags += ["--", *claude_argv, *argv]
+        claude_argv += argv
+        flags += ["--", "runuser", "-u", "node", "--", *claude_argv]
         result = subprocess.run(flags, check=False)
         return result.returncode
 
-    def exec(self, script: str) -> ExecResult:
-        """Run a POSIX shell script and capture the result. The
-        script runs under `/bin/sh -c`, matching what the docker
-        backend's `exec` does — callers can write shell-y test
-        helpers without worrying about argv splitting."""
-        r = _smolvm.machine_exec(
-            self.name,
-            ["/bin/sh", "-c", script],
+    def exec(self, script: str, *, user: str = "node") -> ExecResult:
+        """Run a POSIX shell script as `user` (default `node`) and
+        capture the result. Matches the docker backend's `exec`,
+        which defaults to the image's USER (also node) — so test
+        helpers / provision shell-outs run with the same identity
+        on both backends. Pass `user="root"` for tests that need
+        root.
+
+        `runuser -u <user> -- /bin/sh -c <script>` switches UID
+        without invoking a login shell; HOME / USER are set via
+        `smolvm -e` (see `_env_flags_for`)."""
+        argv = (
+            _env_flags_for(user)
+            + _guest_env_flags(self._guest_env)
+            + ["--", "runuser", "-u", user, "--", "/bin/sh", "-c", script]
+        )
+        # _smolvm.machine_exec expects argv (the bit after `--`);
+        # the -e flags go before, so call smolvm directly.
+        r = subprocess.run(
+            ["smolvm", "machine", "exec", "--name", self.name] + argv,
+            capture_output=True, text=True, check=False,
         )
         return ExecResult(
             returncode=r.returncode,
-            stdout=r.stdout,
-            stderr=r.stderr,
+            stdout=r.stdout or "",
+            stderr=r.stderr or "",
         )
 
     def cp_in(self, host_path: str, container_path: str) -> None:

claude_bottle/backend/smolmachines/bottle_plan.py

@@ -75,6 +75,17 @@ class SmolmachinesBottlePlan(BottlePlan):
     # None when bottle.supervise is False, matching the docker
     # backend's convention.
     supervise_plan: SupervisePlan | None
+    # Agent-side endpoints. On Docker Desktop the docker bridge
+    # IPs aren't reachable from the smolvm guest (TSI uses macOS
+    # networking; docker container IPs live in the daemon's VM),
+    # so the agent dials the bundle via host loopback +
+    # docker-published random ports. Empty at prepare time;
+    # launch populates these after bundle bringup via
+    # `dataclasses.replace`. Format: a `host:port` for git-gate
+    # (insteadOf URL prefix) + full URLs for proxy / supervise.
+    agent_proxy_url: str = ""
+    agent_git_gate_host: str = ""
+    agent_supervise_url: str = ""
 
     def print(self, *, remote_control: bool) -> None:
         """Compact y/N preflight. Same shape as the Docker
@@ -89,7 +100,10 @@ class SmolmachinesBottlePlan(BottlePlan):
         upstreams = [
             f"{g.Name} → {g.Upstream}" for g in bottle.git
         ]
-        routes = [r.host for r in bottle.egress.routes]
+        # Use the resolved egress_plan (lowercase `host` on the
+        # plan-level EgressRoute) rather than `bottle.egress.routes`,
+        # which is the manifest's capitalized-attr form.
+        routes = [r.host for r in self.egress_plan.routes]
 
         print(file=sys.stderr)
         info(f"agent           : {spec.agent_name}")

claude_bottle/backend/smolmachines/launch.py

@@ -21,6 +21,7 @@ from __future__ import annotations
 
 import dataclasses
 import os
+import time
 from contextlib import ExitStack, contextmanager
 from typing import Callable, Generator
 
@@ -34,6 +35,7 @@ from ...util import expand_tilde
 from ..docker.egress import (
     EGRESS_CA_IN_CONTAINER,
     EGRESS_PIPELOCK_CA_IN_CONTAINER,
+    EGRESS_PORT as _EGRESS_PORT,
     egress_tls_init,
 )
 from ..docker.git_gate import (
@@ -41,14 +43,28 @@ from ..docker.git_gate import (
     GIT_GATE_CREDS_DIR_IN_CONTAINER,
     GIT_GATE_ENTRYPOINT_IN_CONTAINER,
     GIT_GATE_HOOK_IN_CONTAINER,
+    GIT_GATE_PORT as _GIT_GATE_PORT,
+)
+from ..docker.pipelock import (
+    BUNDLE_LOCAL_PIPELOCK_URL,
+    PIPELOCK_PORT as _PIPELOCK_PORT_STR,
+    pipelock_tls_init,
 )
-from ..docker.pipelock import BUNDLE_LOCAL_PIPELOCK_URL, pipelock_tls_init
 from . import sidecar_bundle as _bundle
 from . import smolvm as _smolvm
 from .bottle import SmolmachinesBottle
 from .bottle_plan import SmolmachinesBottlePlan
 
 
+# Container-internal listening ports for each bundle daemon. The
+# bundle publishes each one on a random host loopback port (see
+# `_bundle.start_bundle`), and `_bundle.bundle_host_port` looks
+# them up post-start. Pipelock's port is an env-overridable string
+# in docker.pipelock; coerce to int here.
+_PIPELOCK_PORT = int(_PIPELOCK_PORT_STR)
+_SUPERVISE_PORT = SUPERVISE_PORT
+
+
 @contextmanager
 def launch(
     plan: SmolmachinesBottlePlan,
@@ -96,31 +112,129 @@ def launch(
         )
 
         # 3. Build the BundleLaunchSpec from the (now-resolved)
-        # inner Plans: daemon subset, env, bind-mounts.
+        # inner Plans: daemon subset, env, bind-mounts. The spec's
+        # ports_to_publish list expands depending on which daemons
+        # the agent needs to reach from the smolvm guest.
         bundle_spec = _bundle_launch_spec(plan, network)
         token_env = _resolve_token_env(plan, os.environ)
         _bundle.start_bundle(bundle_spec, env={**os.environ, **token_env})
         stack.callback(_bundle.stop_bundle, plan.slug)
 
-        # 4. smolvm VM. --from carries the pre-packed .smolmachine
+        # 4. Discover the host-side ports docker assigned for the
+        # bundle's published container ports, and bind the
+        # agent's URLs to `127.0.0.1:<host port>`. Docker container
+        # IPs (192.168.x.x in the daemon's bridge) aren't
+        # reachable from the smolvm guest on macOS — TSI uses
+        # macOS networking, and macOS sees the daemon's bridge
+        # via the published-port loopback forward only.
+        #
+        # Proxy hop order matches the docker backend: when the
+        # bottle declares egress routes, the agent's first hop is
+        # egress (for token injection), then pipelock. Without
+        # routes, the agent dials pipelock directly. Whichever
+        # one is "agent-facing" is the daemon whose port we
+        # publish on host loopback; the other stays bundle-
+        # internal as the upstream proxy.
+        if plan.egress_plan.routes:
+            agent_facing_port = _EGRESS_PORT
+        else:
+            agent_facing_port = _PIPELOCK_PORT
+        agent_facing_host_port = _bundle.bundle_host_port(
+            plan.slug, agent_facing_port,
+        )
+        agent_proxy_url = f"http://127.0.0.1:{agent_facing_host_port}"
+        agent_git_gate_host = ""
+        if plan.git_gate_plan.upstreams:
+            git_gate_host_port = _bundle.bundle_host_port(
+                plan.slug, _GIT_GATE_PORT,
+            )
+            agent_git_gate_host = f"127.0.0.1:{git_gate_host_port}"
+        agent_supervise_url = ""
+        if plan.supervise_plan is not None:
+            supervise_host_port = _bundle.bundle_host_port(
+                plan.slug, _SUPERVISE_PORT,
+            )
+            agent_supervise_url = f"http://127.0.0.1:{supervise_host_port}/"
+
+        # Stamp the URLs onto the plan + guest_env. provision_git
+        # and provision_supervise read the plan fields; the agent
+        # reads guest_env on every exec_claude.
+        guest_env = {
+            **plan.guest_env,
+            "HTTPS_PROXY": agent_proxy_url,
+            "HTTP_PROXY":  agent_proxy_url,
+        }
+        if agent_git_gate_host:
+            guest_env["GIT_GATE_URL"] = f"git://{agent_git_gate_host}"
+        if agent_supervise_url:
+            guest_env["MCP_SUPERVISE_URL"] = agent_supervise_url
+        plan = dataclasses.replace(
+            plan,
+            guest_env=guest_env,
+            agent_proxy_url=agent_proxy_url,
+            agent_git_gate_host=agent_git_gate_host,
+            agent_supervise_url=agent_supervise_url,
+        )
+
+        # 5. smolvm VM. --from carries the pre-packed .smolmachine
         # artifact (built by prepare); --allow-cidr + -e carry the
-        # per-bottle TSI allowlist + env. Smolfile isn't usable
-        # here — smolvm 0.8.0 makes `--from` and `--smolfile`
-        # mutually exclusive.
+        # per-bottle TSI allowlist + env. The allowlist is
+        # `127.0.0.1/32` because every bundle daemon the agent
+        # reaches is fronted by a host loopback port-forward.
+        # Smolfile isn't usable here — smolvm 0.8.0 makes `--from`
+        # and `--smolfile` mutually exclusive.
         _smolvm.machine_create(
             plan.machine_name,
             from_path=plan.agent_from_path,
-            allow_cidrs=[f"{plan.bundle_ip}/32"],
+            allow_cidrs=["127.0.0.1/32"],
             env=plan.guest_env,
         )
         stack.callback(_smolvm.machine_delete, plan.machine_name)
         _smolvm.machine_start(plan.machine_name)
         stack.callback(_smolvm.machine_stop, plan.machine_name)
 
-        # 5. Provision (CA / prompt / skills / git / supervise).
+        # 6. Repair filesystem ownership + perms that smolvm's
+        # pack process remapped to the host invoker's uid (501
+        # on macOS) rather than preserving the image's expected
+        # ownership.
+        #
+        #  - /home/node → node:node so the node user can write
+        #    its own dotfiles (claude appendFileSync on
+        #    ~/.claude.json otherwise bails with ENOENT/EPERM
+        #    and the TUI hangs without surfacing the error).
+        #  - /tmp + /var/tmp → root:root mode 1777 so non-root
+        #    processes can create their per-uid scratch dirs
+        #    (claude-code creates /tmp/claude-<uid>/ as soon as
+        #    it spawns a Bash tool call).
+        #
+        # All folded into one sh -c so we only pay one
+        # machine_exec round trip — back-to-back exec calls
+        # right after machine_start hit a SIGKILL race in
+        # libkrun's exec channel (see provision_ca for the
+        # other half of this same workaround).
+        _smolvm.machine_exec(plan.machine_name, [
+            "sh", "-c",
+            "chown -R node:node /home/node && "
+            "chown root:root /tmp /var/tmp && "
+            "chmod 1777 /tmp /var/tmp",
+        ])
+
+        # Wait briefly for the VM to settle. Back-to-back smolvm
+        # machine_exec calls immediately after machine_start
+        # occasionally SIGKILL the in-VM child at ~100ms (looks
+        # like a VM warm-up race in libkrun's exec channel).
+        # 1.5s is empirically enough to dodge it; provisioning
+        # already takes seconds so the wait is amortized.
+        time.sleep(1.5)
+
+        # 7. Provision (CA / prompt / skills / git / supervise).
         prompt_path = provision(plan, plan.machine_name)
 
-        yield SmolmachinesBottle(plan.machine_name, prompt_path=prompt_path)
+        yield SmolmachinesBottle(
+            plan.machine_name,
+            prompt_path=prompt_path,
+            guest_env=plan.guest_env,
+        )
     finally:
         stack.close()
 
@@ -144,9 +258,14 @@ def _bundle_launch_spec(
     env: list[str] = []
     volumes: list[tuple[str, str, bool]] = []
 
-    # PRD 0023 chunk 3: egress binds 127.0.0.1 inside the bundle
-    # so TSI's IP-only allowlist can't bypass pipelock.
-    env.append("EGRESS_LISTEN_HOST=127.0.0.1")
+    # In this Docker-Desktop-compatible topology, whichever daemon
+    # is "agent-facing" gets its port published on the host
+    # loopback (see `_ensure_smolmachine`'s discovery loop) and the
+    # other stays bundle-internal. The bundle is NOT reachable by
+    # bridge IP from the smolvm guest, so the
+    # PRD-0023-chunk-3 EGRESS_LISTEN_HOST=127.0.0.1 mitigation
+    # isn't needed: the agent can only dial whatever daemon's
+    # host port we publish, period.
 
     # --- pipelock ---------------------------------------------
     pp = plan.proxy_plan
@@ -201,6 +320,21 @@ def _bundle_launch_spec(
         ]
         volumes.append((str(sp.queue_dir), QUEUE_DIR_IN_CONTAINER, False))
 
+    # Container ports the agent reaches from the smolvm guest —
+    # published on host loopback so the guest can dial via TSI +
+    # macOS networking. The HTTP/HTTPS chokepoint is whichever
+    # daemon's port we publish: egress when routes are declared
+    # (token injection first, then forwards to bundle-internal
+    # pipelock), pipelock otherwise.
+    if ep.routes:
+        ports_to_publish: list[int] = [_EGRESS_PORT]
+    else:
+        ports_to_publish = [_PIPELOCK_PORT]
+    if gp.upstreams:
+        ports_to_publish.append(_GIT_GATE_PORT)
+    if sp is not None:
+        ports_to_publish.append(_SUPERVISE_PORT)
+
     return _bundle.BundleLaunchSpec(
         slug=plan.slug,
         network_name=network,
@@ -210,6 +344,7 @@ def _bundle_launch_spec(
         daemons_csv=",".join(daemons),
         environment=tuple(env),
         volumes=tuple(volumes),
+        ports_to_publish=tuple(ports_to_publish),
     )
 
 

claude_bottle/backend/smolmachines/local_registry.py

@@ -1,17 +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 on `127.0.0.1:<random>`,
-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.
 
-Loopback-only bind + the host's docker layer cache mean the round
-trip is fast (~5s) and there's no exposed surface on the LAN."""
+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).
+
+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
 
@@ -21,6 +41,7 @@ import subprocess
 import time
 import uuid
 from contextlib import contextmanager
+from dataclasses import dataclass
 from typing import Iterator
 
 from ...log import die
@@ -34,64 +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 RegistryHandle:
+    """Everything callers need to push to + pull from the ephemeral
+    registry.
+
+    `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[int]:
-    """Bring up a `registry:2.8.3` container on a random loopback
-    port, yield the port, 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,
-            # `127.0.0.1::5000` = bind to loopback, pick a random host
-            # port. No LAN exposure; the container hangs around just
-            # long enough for one push + one pack-create.
-            "-p", "127.0.0.1::5000",
-            REGISTRY_IMAGE,
-        ],
+        ["docker", "network", "create", network],
         check=True,
         capture_output=True,
     )
     try:
-        port = _host_port(name)
-        _wait_ready(port)
-        yield 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 _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; the loopback-only -p binding ensures we
-    get exactly `127.0.0.1:<port>`."""
+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.
+
+    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", "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>'}"
         )
-    # `127.0.0.1:54321\n` — split on the last colon to handle the
-    # `host:port` shape without parsing IP literals.
+    # `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:
@@ -102,12 +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."""
+    follows will land on a working server. We probe loopback
+    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
 
 
@@ -89,22 +89,23 @@ def resolve_plan(
 
     subnet, gateway, bundle_ip = smolmachines_bundle_subnet(slug)
 
-    # Agent's env. IP literals; no DNS resolution inside the guest
-    # (TSI allowlist contains only `<bundle_ip>/32` — no resolver).
+    # Agent's env: the prepare-time view doesn't yet know the
+    # host loopback ports the bundle's daemons get published on
+    # (those come from docker AFTER `docker run` returns), so
+    # HTTPS_PROXY / GIT_GATE_URL / MCP_SUPERVISE_URL are
+    # populated in launch.py and stamped onto guest_env there.
+    # What we set here is the part that doesn't depend on
+    # bundle bringup — bottle.env literals, the empty-NO_PROXY
+    # safe default, and the TLS trust env trio
+    # (NODE_EXTRA_CA_CERTS / SSL_CERT_FILE / REQUESTS_CA_BUNDLE)
+    # pointing at Debian's update-ca-certificates output bundle.
     guest_env: dict[str, str] = {
         **bottle.env,
-        "HTTPS_PROXY": f"http://{bundle_ip}:{_BUNDLE_PIPELOCK_PORT}",
-        "HTTP_PROXY":  f"http://{bundle_ip}:{_BUNDLE_PIPELOCK_PORT}",
         "NO_PROXY":    "localhost,127.0.0.1",
+        "NODE_EXTRA_CA_CERTS": "/etc/ssl/certs/ca-certificates.crt",
+        "SSL_CERT_FILE":       "/etc/ssl/certs/ca-certificates.crt",
+        "REQUESTS_CA_BUNDLE":  "/etc/ssl/certs/ca-certificates.crt",
     }
-    if bottle.git:
-        guest_env["GIT_GATE_URL"] = (
-            f"git://{bundle_ip}:{_BUNDLE_GIT_GATE_PORT}"
-        )
-    if bottle.supervise:
-        guest_env["MCP_SUPERVISE_URL"] = (
-            f"http://{bundle_ip}:{_BUNDLE_SUPERVISE_PORT}"
-        )
 
     # Inner Plans for the four bundle daemons. The ABCs are
     # platform-neutral — `.prepare()` writes config files + returns
@@ -124,6 +125,19 @@ def resolve_plan(
     egress_dir.mkdir(parents=True, exist_ok=True)
     egress_plan = Egress().prepare(bottle, slug, egress_dir)
 
+    # Claude-code refuses to start without *something* it
+    # recognises as a credential. When the bottle has an egress
+    # route carrying the `claude_code_oauth` role marker, egress
+    # strips + re-injects the real Authorization header on the
+    # outbound leg using a token held in egress's own environ — so
+    # the agent gets a non-secret placeholder here (matches the
+    # docker backend's forwarded_env logic in
+    # claude_bottle/backend/docker/prepare.py).
+    if any("claude_code_oauth" in r.roles for r in egress_plan.routes):
+        guest_env["CLAUDE_CODE_OAUTH_TOKEN"] = "egress-placeholder"
+        guest_env.setdefault("CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC", "1")
+        guest_env.setdefault("DISABLE_ERROR_REPORTING", "1")
+
     supervise_plan = None
     if bottle.supervise:
         supervise_dir = supervise_state_dir(slug)
@@ -184,14 +198,19 @@ def _ensure_smolmachine(image_ref: str) -> Path:
     a launcher binary at `.smolmachine` plus the sidecar alongside
     it; the sidecar is the actual artifact).
 
-    Conversion path: `docker build` (the existing layer cache makes
-    no-change rebuilds cheap) → `docker tag` with a
-    `localhost:<port>/...` ref → bring up the ephemeral registry
-    container → `docker push` into it → `smolvm pack create --image
-    <localhost ref>` → tear down the registry. Each pack-create
-    costs several seconds even on a hot cache, so we skip the whole
-    pipeline when the cached sidecar is already on disk for this
-    image ID."""
+    Conversion path: `docker build` (the existing layer cache
+    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
+    already on disk for this image ID."""
     _SMOLMACHINE_CACHE_DIR.mkdir(parents=True, exist_ok=True)
     docker_mod.build_image(image_ref, _REPO_DIR)
     # `sha256:abcd...` -> `abcd...` first 16 chars: short enough to
@@ -202,9 +221,17 @@ 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 port:
-        local_ref = f"localhost:{port}/claude-bottle:{digest}"
-        docker_mod.tag(image_ref, local_ref)
-        docker_mod.push(local_ref)
-        _smolvm.pack_create(local_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

claude_bottle/backend/smolmachines/provision/ca.py

@@ -66,8 +66,29 @@ def provision_ca(plan: SmolmachinesBottlePlan, target: str) -> None:
     # default. The env trio (NODE_EXTRA_CA_CERTS / SSL_CERT_FILE /
     # REQUESTS_CA_BUNDLE) on the guest_env covers Node + Python
     # `requests` / libraries that don't load the system bundle.
-    _smolvm.machine_exec(target, ["chmod", "644", AGENT_CA_PATH])
-    _smolvm.machine_exec(target, ["update-ca-certificates"])
+    #
+    # chown + chmod + update-ca-certificates run in one
+    # `sh -c` so we only pay one machine_exec round trip; the
+    # `&&` chaining surfaces the first failure as the return
+    # code.
+    r = _smolvm.machine_exec(target, [
+        "sh", "-c",
+        f"chown root:root {AGENT_CA_PATH} && "
+        f"chmod 644 {AGENT_CA_PATH} && "
+        f"update-ca-certificates",
+    ])
+    if r.returncode != 0 or "1 added" not in (r.stdout or ""):
+        # update-ca-certificates not adding our cert is fatal —
+        # claude-code's TLS handshake against the egress-MITM'd
+        # api.anthropic.com would fail downstream. Bail early
+        # with what we can see (output is captured by smolvm so
+        # we can surface it).
+        die(
+            f"update-ca-certificates didn't add the agent CA "
+            f"(exit {r.returncode}): "
+            f"stdout={(r.stdout or '').strip()!r} "
+            f"stderr={(r.stderr or '').strip()!r}"
+        )
 
     # Stdlib SHA-256 of the cert's DER bytes — the standard
     # fingerprint form. Never the private key.

claude_bottle/backend/smolmachines/provision/git.py

@@ -28,7 +28,6 @@ from pathlib import Path
 
 from ....git_gate import git_gate_render_gitconfig
 from ....log import info
-from ...docker.git_gate import GIT_GATE_PORT
 from .. import smolvm as _smolvm
 from ..bottle_plan import SmolmachinesBottlePlan
 
@@ -79,10 +78,12 @@ def _provision_git_gate_config(plan: SmolmachinesBottlePlan, target: str) -> Non
     if not bottle.git:
         return
 
-    # IP-literal form: the TSI allowlist passes <bundle_ip>/32 and
-    # nothing else, so the agent has to dial the gate by IP+port.
-    gate_host = f"{plan.bundle_ip}:{GIT_GATE_PORT}"
-    content = git_gate_render_gitconfig(bottle.git, gate_host)
+    # `127.0.0.1:<host port>` form: the bundle's git-gate port
+    # is published on host loopback at launch time so the
+    # smolvm guest (which can only reach macOS networking via
+    # TSI, not the docker bridge IP) can dial it. launch.py
+    # populates `plan.agent_git_gate_host` after bundle bringup.
+    content = git_gate_render_gitconfig(bottle.git, plan.agent_git_gate_host)
 
     guest_gitconfig = f"{_guest_home()}/.gitconfig"
     # Stage the file under the plan's stage_dir so `machine cp`

claude_bottle/backend/smolmachines/provision/supervise.py

@@ -14,7 +14,6 @@ short `supervise` alias (no DNS in the TSI-allowlisted guest)."""
 from __future__ import annotations
 
 from ....log import info, warn
-from ....supervise import SUPERVISE_PORT
 from .. import smolvm as _smolvm
 from ..bottle_plan import SmolmachinesBottlePlan
 
@@ -22,25 +21,33 @@ from ..bottle_plan import SmolmachinesBottlePlan
 _SUPERVISE_MCP_NAME = "supervise"
 
 
-def supervise_mcp_url(bundle_ip: str) -> str:
-    return f"http://{bundle_ip}:{SUPERVISE_PORT}/"
-
-
 def provision_supervise(plan: SmolmachinesBottlePlan, target: str) -> None:
     """Run `claude mcp add` inside the guest to register the
     supervise sidecar in claude-code's user config. No-op when
     bottle.supervise is False.
 
+    The URL is the agent-side endpoint launch.py populated after
+    bundle bringup — `http://127.0.0.1:<host port>/` rather than
+    the bundle's docker bridge IP, because that bridge isn't
+    reachable from the smolvm guest on macOS.
+
     Failure is logged but not fatal: the bottle still works (you
     just can't call supervise tools from the agent until the entry
     is added manually). The operator sees the warning at launch."""
     if plan.supervise_plan is None:
         return
-    url = supervise_mcp_url(plan.bundle_ip)
+    url = plan.agent_supervise_url
     info(f"registering supervise MCP server in agent claude config → {url}")
+    # `claude mcp add --scope user` writes to ~/.claude.json. The
+    # agent is the `node` user; smolvm machine_exec runs as root
+    # by default, so we have to switch user explicitly and set
+    # HOME so the config lands in /home/node/.claude.json (where
+    # the agent's claude actually reads it from).
     r = _smolvm.machine_exec(
         target,
         [
+            "runuser", "-u", "node", "--",
+            "env", "HOME=/home/node",
             "claude", "mcp", "add",
             "--scope", "user",
             "--transport", "http",
@@ -57,4 +64,4 @@ def provision_supervise(plan: SmolmachinesBottlePlan, target: str) -> None:
         )
 
 
-__all__ = ["provision_supervise", "supervise_mcp_url"]
+__all__ = ["provision_supervise"]

claude_bottle/backend/smolmachines/sidecar_bundle.py

@@ -70,6 +70,13 @@ class BundleLaunchSpec:
     environment: Sequence[str] = field(default_factory=tuple)
     # (host_path, container_path, read_only) bind mounts.
     volumes: Sequence[tuple[str, str, bool]] = field(default_factory=tuple)
+    # Container ports to publish on the host's 127.0.0.1, random
+    # host-side port per entry. The smolvm guest's TSI talks via
+    # macOS networking, so docker container IPs (192.168.x.x in
+    # the daemon's bridge) aren't directly reachable from the
+    # guest — host-loopback port-forwards are. Egress's port
+    # is bundle-internal and never published.
+    ports_to_publish: Sequence[int] = field(default_factory=tuple)
 
 
 def create_bundle_network(network_name: str, subnet: str, gateway: str) -> None:
@@ -135,6 +142,11 @@ def start_bundle(spec: BundleLaunchSpec, *,
     for host_path, container_path, read_only in spec.volumes:
         suffix = ":ro" if read_only else ""
         argv += ["-v", f"{host_path}:{container_path}{suffix}"]
+    # Loopback-only host port-forwards — the smolvm guest's TSI
+    # uses macOS networking, and macOS loopback is the only host
+    # surface that round-trips into Docker Desktop's daemon VM.
+    for port in spec.ports_to_publish:
+        argv += ["-p", f"127.0.0.1::{port}"]
     argv.append(spec.image)
     result = subprocess.run(
         argv, capture_output=True, text=True,
@@ -147,6 +159,33 @@ def start_bundle(spec: BundleLaunchSpec, *,
         )
 
 
+def bundle_host_port(slug: str, container_port: int) -> int:
+    """`docker port <bundle> <container_port>/tcp` → the random
+    host-side port docker assigned. Called after `start_bundle`
+    on each container port listed in `BundleLaunchSpec
+    .ports_to_publish` so the launch step can build the agent's
+    HTTPS_PROXY / GIT_GATE / SUPERVISE URLs in
+    `127.0.0.1:<host port>` form."""
+    container = bundle_container_name(slug)
+    result = subprocess.run(
+        ["docker", "port", container, f"{container_port}/tcp"],
+        capture_output=True, text=True, check=False,
+    )
+    if result.returncode != 0:
+        die(
+            f"docker port {container} {container_port}/tcp failed: "
+            f"{(result.stderr or '').strip() or '<no stderr>'}"
+        )
+    # `127.0.0.1:54321\n` — rpartition on last colon gives the port.
+    line = (result.stdout or "").splitlines()[0].strip()
+    _, _, port_str = line.rpartition(":")
+    try:
+        return int(port_str)
+    except ValueError:
+        die(f"unexpected `docker port` output: {line!r}")
+        return -1  # unreachable; die() never returns
+
+
 def stop_bundle(slug: str) -> None:
     """Idempotent: a missing container returns success."""
     container = bundle_container_name(slug)

claude_bottle/backend/smolmachines/smolvm.py

@@ -117,12 +117,21 @@ def machine_create(
     Smolfile because `--from` and `--smolfile` are themselves
     mutually exclusive in smolvm 0.8.0 — and we want `--from`'s
     no-pull-at-start property. The flag form gives the same
-    result without the Smolfile complication."""
+    result without the Smolfile complication.
+
+    `--net` is sent explicitly when `allow_cidrs` is non-empty.
+    smolvm 0.8.0's docs say `--allow-cidr` implies `--net`, but
+    empirically the implication only fires when no `--from` is
+    set — `--from PATH --allow-cidr X/32` silently produces a
+    machine with `network: false` and no routes in the guest, so
+    the agent can't reach the bundle's pinned IP."""
     args: list[str] = ["machine", "create"]
     if image is not None:
         args += ["--image", image]
     if from_path is not None:
         args += ["--from", str(from_path)]
+    if allow_cidrs:
+        args.append("--net")
     for cidr in allow_cidrs:
         args += ["--allow-cidr", cidr]
     if env:

docs/prds/0023-smolmachines-backend.md

@@ -600,6 +600,22 @@ PRD 0024's bundle image is a prerequisite — this PRD assumes
    the plan is to filter on a deterministic name prefix
    `claude-bottle-<slug>` + cross-reference with on-disk metadata
    under `state/<slug>/`.
+8. **Loopback scoping (Docker Desktop pivot).** The original
+   design pinned the bundle at a docker bridge IP and set TSI's
+   allowlist to `<bundle-ip>/32`. On Docker Desktop / macOS the
+   daemon runs inside its own Linux VM, so bridge IPs aren't
+   reachable from macOS networking — TSI's syscall impersonation
+   can't reach them. Resolution: publish each agent-facing bundle
+   port on host loopback (`-p 127.0.0.1::<port>`) and set TSI to
+   `127.0.0.1/32`. **This widens the TSI allowlist to anything
+   bound to macOS's loopback** — postgres, dev servers, other
+   bottles' published ports, mDNSResponder, etc. The agent can't
+   reach them by intent, but TSI can't filter by port. Follow-up
+   to scope back: bind each bottle's bundle ports on a per-bottle
+   loopback alias (e.g. `127.0.0.2` for bottle A, `127.0.0.3` for
+   B) added via `ifconfig lo0 alias`, set TSI to that single /32.
+   Needs sudo for alias setup; a small daemon-or-script we ship
+   alongside the launcher could handle it.
 
 ## References
 

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

@@ -1,15 +1,16 @@
 """Unit: ephemeral local-registry helper (PRD 0023 chunk 4c).
 
-The helper brings up a `registry:2.8.3` container on a random
-loopback port, yields the port, 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
 
 import subprocess
 import unittest
-from unittest.mock import call, patch
+from unittest.mock import patch
 
 from claude_bottle.backend.smolmachines import local_registry
 
@@ -20,38 +21,81 @@ def _ok(stdout: str = "", stderr: str = "") -> subprocess.CompletedProcess:
     )
 
 
+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(),                            # 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_host_port_parsed_from_docker_port(self):
-        # docker run + docker port + docker rm in that order; the
-        # port command returns `127.0.0.1:54321` for the loopback
-        # binding.
+    def test_yields_handle_with_network_and_endpoints(self):
         with patch.object(
             local_registry.subprocess, "run",
-            side_effect=[
-                _ok(stdout="<container-id>\n"),
-                _ok(stdout="127.0.0.1:54321\n"),
-                _ok(),
-            ],
+            side_effect=_stock_run_sequence(),
         ) as run, patch.object(
             local_registry.socket, "create_connection",
             return_value=_FakeSocket(),
         ):
-            with local_registry.ephemeral_registry() as port:
-                self.assertEqual(54321, port)
+            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.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)
 
-        # docker run, docker port, docker rm -f
-        self.assertEqual(3, run.call_count)
-        run_argv = run.call_args_list[0].args[0]
-        self.assertEqual(["docker", "run"], run_argv[:2])
-        self.assertIn("--rm", run_argv)
-        # Loopback-only port binding so the registry isn't exposed
-        # on the LAN even briefly.
-        self.assertIn("127.0.0.1::5000", run_argv)
-
-    def test_force_removes_container_on_clean_exit(self):
+    def test_registry_run_publishes_random_port_across_interfaces(self):
         with patch.object(
             local_registry.subprocess, "run",
-            side_effect=[_ok(stdout="cid\n"), _ok(stdout="127.0.0.1:1234\n"), _ok()],
+            side_effect=_stock_run_sequence(),
+        ) as run, patch.object(
+            local_registry.socket, "create_connection",
+            return_value=_FakeSocket(),
+        ):
+            with local_registry.ephemeral_registry():
+                pass
+        # 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)
+        # 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_and_network_on_clean_exit(self):
+        with patch.object(
+            local_registry.subprocess, "run",
+            side_effect=_stock_run_sequence(),
         ) as run, patch.object(
             local_registry.socket, "create_connection",
             return_value=_FakeSocket(),
@@ -59,14 +103,16 @@ 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=[_ok(stdout="cid\n"), _ok(stdout="127.0.0.1:1234\n"), _ok()],
+            side_effect=_stock_run_sequence(),
         ) as run, patch.object(
             local_registry.socket, "create_connection",
             return_value=_FakeSocket(),
@@ -75,15 +121,15 @@ 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=[_ok(stdout="cid\n"), _ok(stdout="127.0.0.1:1234\n"), _ok()],
+            side_effect=_stock_run_sequence(),
         ) as run, patch.object(
             local_registry.socket, "create_connection",
             side_effect=OSError("conn refused"),
@@ -95,18 +141,27 @@ 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])
-            return _ok(stdout="cid\n" if argv[:2] == ["docker", "run"]
-                       else "127.0.0.1:1\n")
+                # `--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")
+            return _ok()
 
         with patch.object(
             local_registry.subprocess, "run", side_effect=capture,
@@ -119,10 +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

@@ -40,32 +40,43 @@ 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 is a context manager yielding the port.
+        # 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 (
+            RegistryHandle,
+        )
+
         class _Reg:
             def __enter__(self_inner):
-                return 54321
+                return RegistryHandle(
+                    network="cb-net-xyz",
+                    push_endpoint="cb-registry-xyz:5000",
+                    pull_endpoint="localhost:54321",
+                )
             def __exit__(self_inner, *exc):
                 return False
 
@@ -83,13 +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",
@@ -97,23 +108,31 @@ 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 port.
-        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 goes from the source ref to a localhost:<port> ref
-        # with the digest as the tag suffix (so different builds
-        # land on different tags in the registry).
-        tag_args = tag.call_args.args
-        self.assertEqual("claude-bottle:latest", tag_args[0])
-        self.assertEqual(f"localhost:54321/claude-bottle:{digest}", tag_args[1])
-        # push targets the same localhost ref tag picks.
+        # 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"localhost:54321/claude-bottle:{digest}", push_args[0])
-        # pack_create reads from the registry ref, writes the
-        # binary alongside the cached sidecar.
+        self.assertEqual(
+            f"cb-registry-xyz:5000/claude-bottle:{digest}", push_args[2],
+        )
+
+        # 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])
+        self.assertEqual(
+            f"localhost:54321/claude-bottle:{digest}", pack_args[0],
+        )
         self.assertTrue(str(pack_args[1]).endswith(f"{digest}.smolmachine"))
 
 

tests/unit/test_smolmachines_provision.py

@@ -44,6 +44,8 @@ def _plan(
     pipelock_ca_path: Path = Path(),
     supervise: bool = False,
     bundle_ip: str = "192.168.50.2",
+    agent_git_gate_host: str = "127.0.0.1:55555",
+    agent_supervise_url: str = "http://127.0.0.1:55556/",
 ) -> SmolmachinesBottlePlan:
     bottle_json: dict = {}
     if git:
@@ -111,6 +113,8 @@ def _plan(
             mitmproxy_ca_cert_only_host_path=egress_ca_path,
         ),
         supervise_plan=supervise_plan,
+        agent_git_gate_host=agent_git_gate_host,
+        agent_supervise_url=agent_supervise_url,
     )
 
 
@@ -303,21 +307,38 @@ class TestProvisionCA(unittest.TestCase):
     def tearDown(self):
         self._tmp.cleanup()
 
+    # provision_ca dies hard if update-ca-certificates' stdout
+    # doesn't include "1 added"; supply a stock success return
+    # so the bulk of the tests below exercise the happy path.
+    _UPDATE_OK = SmolvmRunResult(
+        returncode=0,
+        stdout="Updating certificates in /etc/ssl/certs...\n1 added, 0 removed; done.\n",
+        stderr="",
+    )
+
     def test_pipelock_path_when_no_routes(self):
         plan = _plan(pipelock_ca_path=self.pipelock_ca)
         with patch(
             "claude_bottle.backend.smolmachines.provision.ca._smolvm.machine_cp"
         ) as cp, patch(
-            "claude_bottle.backend.smolmachines.provision.ca._smolvm.machine_exec"
+            "claude_bottle.backend.smolmachines.provision.ca._smolvm.machine_exec",
+            return_value=self._UPDATE_OK,
         ) as ex:
             _ca.provision_ca(plan, "claude-bottle-demo-abc12")
         cp.assert_called_once_with(
             str(self.pipelock_ca),
             "claude-bottle-demo-abc12:" + _ca.AGENT_CA_PATH,
         )
-        argvs = [c.args[1] for c in ex.call_args_list]
-        self.assertIn(["chmod", "644", _ca.AGENT_CA_PATH], argvs)
-        self.assertIn(["update-ca-certificates"], argvs)
+        # chmod + chown + update-ca-certificates are now folded
+        # into one `sh -c` invocation (working around a smolvm
+        # exec warm-up SIGKILL race), so we look at the single
+        # exec's argv rather than expecting separate calls.
+        ex.assert_called_once()
+        argv = ex.call_args.args[1]
+        self.assertEqual("sh", argv[0])
+        self.assertEqual("-c", argv[1])
+        self.assertIn("chmod 644", argv[2])
+        self.assertIn("update-ca-certificates", argv[2])
 
     def test_egress_path_when_routes_declared(self):
         plan = _plan(
@@ -328,7 +349,8 @@ class TestProvisionCA(unittest.TestCase):
         with patch(
             "claude_bottle.backend.smolmachines.provision.ca._smolvm.machine_cp"
         ) as cp, patch(
-            "claude_bottle.backend.smolmachines.provision.ca._smolvm.machine_exec"
+            "claude_bottle.backend.smolmachines.provision.ca._smolvm.machine_exec",
+            return_value=self._UPDATE_OK,
         ):
             _ca.provision_ca(plan, "claude-bottle-demo-abc12")
         # When routes are declared, egress is the agent's first hop,
@@ -412,9 +434,10 @@ class TestProvisionGit(unittest.TestCase):
         cp.assert_not_called()
 
     def test_writes_gitconfig_with_ip_port_form_for_smolmachines(self):
-        # Smolmachines's TSI-allowlisted guest has no DNS resolver,
-        # so the insteadOf URL has to be IP+port rather than the
-        # docker backend's `git-gate` short alias.
+        # Smolmachines's TSI-allowlisted guest dials git-gate via
+        # `127.0.0.1:<host port>` — the bundle's git-gate port is
+        # published on host loopback at launch time, and the plan
+        # carries the discovered host port (here mocked to 9418).
         plan = _plan(
             git=[GitEntry(
                 Name="claude-bottle",
@@ -422,7 +445,7 @@ class TestProvisionGit(unittest.TestCase):
                 IdentityFile="~/.ssh/id_ed25519",
             )],
             stage_dir=self.stage,
-            bundle_ip="192.168.99.2",
+            agent_git_gate_host="127.0.0.1:9418",
         )
         with patch(
             "claude_bottle.backend.smolmachines.provision.git._smolvm.machine_cp"
@@ -437,7 +460,7 @@ class TestProvisionGit(unittest.TestCase):
         self.assertEqual(self.stage, staged_path.parent)
         content = staged_path.read_text()
         self.assertIn(
-            '[url "git://192.168.99.2:9418/claude-bottle.git"]', content,
+            '[url "git://127.0.0.1:9418/claude-bottle.git"]', content,
         )
         self.assertIn(
             "\tinsteadOf = ssh://git@host/repo.git", content,
@@ -453,7 +476,10 @@ class TestProvisionSupervise(unittest.TestCase):
         ex.assert_not_called()
 
     def test_calls_claude_mcp_add_when_supervise_enabled(self):
-        plan = _plan(supervise=True, bundle_ip="192.168.50.2")
+        plan = _plan(
+            supervise=True,
+            agent_supervise_url="http://127.0.0.1:9100/",
+        )
         with patch(
             "claude_bottle.backend.smolmachines.provision.supervise._smolvm.machine_exec",
             return_value=SmolvmRunResult(returncode=0, stdout="", stderr=""),
@@ -461,14 +487,20 @@ class TestProvisionSupervise(unittest.TestCase):
             _supervise.provision_supervise(plan, "claude-bottle-demo-abc12")
         ex.assert_called_once()
         argv = ex.call_args.args[1]
-        # claude mcp add --scope user --transport http supervise <url>
+        # `claude mcp add --scope user` writes to ~/.claude.json,
+        # and the agent is the `node` user — switch UID + set
+        # HOME so the config lands in /home/node/.claude.json,
+        # not root's. URL is the agent-side endpoint (host
+        # loopback + discovered port), not the docker bridge IP.
         self.assertEqual(
             [
+                "runuser", "-u", "node", "--",
+                "env", "HOME=/home/node",
                 "claude", "mcp", "add",
                 "--scope", "user",
                 "--transport", "http",
                 "supervise",
-                "http://192.168.50.2:9100/",
+                "http://127.0.0.1:9100/",
             ],
             argv,
         )

tests/unit/test_smolmachines_smolvm.py

@@ -82,12 +82,20 @@ class TestArgvShapes(unittest.TestCase):
         self.assertEqual("smolvm", argv[0])
         self.assertIn("--from", argv)
         self.assertIn("/stage/agent.smolmachine", argv)
+        # `--net` is explicit because smolvm 0.8.0's implied-net
+        # from --allow-cidr doesn't fire when --from is set.
+        self.assertIn("--net", argv)
         self.assertIn("--allow-cidr", argv)
         self.assertIn("192.168.50.2/32", argv)
         self.assertIn("-e", argv)
         self.assertIn("HTTPS_PROXY=http://192.168.50.2:8888", argv)
         self.assertEqual("agent-xyz", argv[-1])
 
+    def test_machine_create_omits_net_when_no_allow_cidrs(self):
+        with self._patch_run() as m:
+            machine_create("agent-xyz", from_path=Path("/x.smolmachine"))
+        self.assertNotIn("--net", m.call_args.args[0])
+
     def test_machine_start_uses_dash_name(self):
         # `start` is the --name flag form, NOT positional.
         with self._patch_run() as m: