Files
bot-bottle/claude_bottle/backend/smolmachines/bottle_plan.py
T
didericis-claude 4f136a9932
test / unit (pull_request) Successful in 26s
test / integration (pull_request) Successful in 39s
fix(smolmachines): agent dials bundle via host loopback ports, not docker bridge IP
Claude hung on outbound network calls under
CLAUDE_BOTTLE_BACKEND=smolmachines:

  Unable to connect to API (FailedToOpenSocket)

Root cause: the PRD-0023 design pinned the bundle at a docker
bridge IP (192.168.X.2) and set the smolvm guest's TSI allowlist
to `<bundle-ip>/32`. On native Linux this works — host shares
the docker bridge's network namespace, TSI's syscall
impersonation reaches the bridge IP directly. On Docker Desktop
(macOS), the daemon runs in its own Linux VM and docker bridge
IPs aren't reachable from macOS networking, so the smolvm
guest's TSI requests die "Network is unreachable" before they
hit pipelock.

Fix: publish each agent-facing bundle daemon's port on host
loopback (-p 127.0.0.1::PORT), discover the random host-side
ports after start, and route the agent through
`127.0.0.1:<host port>` instead of the bridge IP. macOS loopback
is the surface Docker Desktop's gvproxy forwards into the
daemon's VM, so the chain (guest TSI -> macOS loopback ->
daemon VM port-forward -> bundle container) works on both
Docker Desktop and native Linux.

Concrete changes:
- BundleLaunchSpec: add `ports_to_publish` so start_bundle adds
  `-p 127.0.0.1::PORT` for the agent-facing ports (pipelock
  always; git-gate when upstreams declared; supervise when
  enabled). Egress's port stays bundle-internal.
- sidecar_bundle.bundle_host_port(): wrap `docker port <bundle>
  <container_port>/tcp` so launch can look up the random
  host-side mapping after start.
- launch.py: discover the host ports, build URLs of the form
  `http://127.0.0.1:<host port>` / `git://127.0.0.1:<host port>`,
  stamp onto guest_env + new agent_*_url fields on the plan.
- launch.py: TSI allow_cidrs flips to `["127.0.0.1/32"]`. The
  bundle IP is no longer the agent's target.
- prepare.py: stop synthesizing HTTPS_PROXY / GIT_GATE_URL /
  MCP_SUPERVISE_URL at prepare time — launch owns those now
  (the values depend on a port docker hasn't assigned yet).
- provision_git: gate_host from plan.agent_git_gate_host.
- provision_supervise: URL from plan.agent_supervise_url.

End-to-end verified on Docker Desktop / macOS: guest dials
pipelock through TSI, pipelock forwards to api.anthropic.com,
the API responds with 401 (i.e. it received the request).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-27 15:31:44 -04:00

118 lines
4.9 KiB
Python

"""SmolmachinesBottlePlan — concrete BottlePlan for the smolmachines
backend (PRD 0023).
Slug + bundle docker subnet / gateway / pinned IP + smolvm
machine name + agent `.smolmachine` artifact + per-bottle guest
env. Provisioning fields (CA cert path, prompt path, etc.) land
in chunk 4."""
from __future__ import annotations
import sys
from dataclasses import dataclass
from pathlib import Path
from ...egress import EgressPlan
from ...git_gate import GitGatePlan
from ...log import info
from ...pipelock import PipelockProxyPlan
from ...supervise import SupervisePlan
from .. import BottlePlan
from ..print_util import print_multi
@dataclass(frozen=True)
class SmolmachinesBottlePlan(BottlePlan):
"""Resolved fields the launch step needs to bring up the bottle.
Inherits `spec` and `stage_dir` from BottlePlan."""
slug: str
# Per-bottle docker subnet for the sidecar bundle container.
# The bundle runs at `bundle_ip` (always `.2`); the gateway is
# at `.1`. smolvm's TSI allowlist is set to `bundle_ip/32`.
bundle_subnet: str
bundle_gateway: str
bundle_ip: str
# smolvm machine name + agent image source. machine_create
# boots from a packed `.smolmachine` artifact (pre-baked at
# prepare time via `smolvm pack create`); using `--from`
# instead of `--image` avoids the registry-pull race we hit
# when machine_start tried to fetch on-demand and the libkrun
# agent's network attempt got refused by macOS.
#
# Chunk 2d ships with a public placeholder image (alpine)
# since claude-bottle:latest lives in the operator's local
# docker daemon and smolvm's crane backend can't read from
# there; chunk 4 resolves the agent-image-conversion gap
# (push to a registry first, or smolvm grows a docker-daemon
# transport).
machine_name: str
agent_from_path: Path
# In-guest env vars (HTTPS_PROXY etc) — IP-literal URLs since
# the guest has no DNS resolver inside the TSI allowlist.
# Passed to `smolvm machine create` as `-e K=V` flags.
# Smolfile-rendering is gone (smolvm 0.8.0's
# `--smolfile` is mutually exclusive with `--from`, and
# `--from` is the path that avoids the registry-pull race).
guest_env: dict[str, str]
# Path to the agent's prompt file on the host. Always written
# (mode 0o600) so the in-VM path always exists; the file is
# empty when the agent has no prompt — claude-code reads it
# via --append-system-prompt-file only when non-empty.
prompt_file: Path
# Inner Plans for the four bundle daemons. The same shape the
# docker backend uses — same `.prepare()` calls produced
# them — but our launch step doesn't populate the
# docker-specific network fields (internal_network,
# egress_network) because the smolmachines bundle isn't on
# docker's `--internal` + egress bridge topology; it's on a
# per-bottle bridge with a pinned IP. The unused fields stay
# at their dataclass defaults.
proxy_plan: PipelockProxyPlan
git_gate_plan: GitGatePlan
egress_plan: EgressPlan
# 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
backend's so operators see one format across backends."""
del remote_control # not surfaced in the compact summary
spec = self.spec
manifest = spec.manifest
agent = manifest.agents[spec.agent_name]
bottle = manifest.bottle_for(spec.agent_name)
env_names = sorted(bottle.env.keys())
upstreams = [
f"{g.Name}{g.Upstream}" for g in bottle.git
]
# 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}")
print_multi("env ", env_names)
print_multi("skills ", list(agent.skills))
info(f"bottle : {agent.bottle}")
if upstreams:
print_multi(" git gate ", upstreams)
if routes:
print_multi(" egress ", routes)
print(file=sys.stderr)