feat(smolmachines): rewrite Smolfile to smolvm 0.8.0 schema + drop gvproxy (PRD 0023 chunk 2a)
test / unit (pull_request) Successful in 21s
test / integration (pull_request) Successful in 39s

First sub-PR of chunk 2: rewrite the renderer chunk 1 shipped to
match smolvm 0.8.0's actual Smolfile shape, delete the dead
gvproxy renderer + its tests, simplify the prepare flow now that
there's no gvproxy socket + no loopback-port allocation.

Smolfile renderer:
- Old shape (under the abandoned gvproxy design): name = ...,
  command = [...], [[net]] attachment = "unixgram",
  socket = "...".
- New shape (smolvm 0.8.0): env = [...] (sorted K=V pairs),
  [network] allow_cidrs = ["<bundle-ip>/32"]. Nothing else.
  image / entrypoint / cmd come from the .smolmachine artifact
  built in chunk 2b; cpus / memory left at smolvm defaults.
- Tests assert no leakage of TSI's --outbound-localhost-only or
  the old gvproxy/unixgram keys.

util.py:
- smolmachines_gvproxy_subnet → smolmachines_bundle_subnet,
  returning (subnet, gateway, bundle_ip). bundle_ip is always
  at .2 (gateway .1); subnet is /24, third octet derived from
  the slug hash, skipping the docker-default 17 to avoid the
  common 192.168.17.x collision.
- allocate_loopback_port: deleted. The bundle gets a pinned
  docker IP now; the agent dials that IP directly through TSI.
- smolmachines_preflight: dropped the gvproxy check; only
  smolvm is required.

prepare.py:
- Drops the gvproxy.yaml render + the loopback port allocation
  + the gvproxy_socket field on the plan.
- Derives subnet / gateway / bundle_ip from the slug and
  populates the new SmolmachinesBottlePlan fields.
- Agent env now uses IP-literal URLs (http://<bundle-ip>:8888
  etc) since the guest will have no DNS resolver inside TSI's
  allowlist.

bottle_plan.py:
- Old fields: gvproxy_config_path, gvproxy_socket,
  gvproxy_subnet, gvproxy_gateway, host_port_map.
- New fields: bundle_subnet, bundle_gateway, bundle_ip,
  smolfile_path. (smolmachine artifact path lands in chunk 2b.)

Net: -410 lines. Full unit suite: 516 passing.

The VM lifecycle + bundle bringup + launch wiring + smoke tests
land in chunk 2b.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-05-27 04:01:07 -04:00
parent b57256789f
commit c73d717f71
8 changed files with 203 additions and 613 deletions
+37 -67
View File
@@ -1,90 +1,65 @@
"""Smolfile (TOML) renderer for the smolmachines backend (PRD 0023).
The Smolfile pins the per-bottle microVM's command + env + virtio-net
device. Three fields drive what we emit:
The Smolfile shape comes from `smolvm machine create --smolfile`
in smolvm 0.8.0. The renderer emits only the runtime overrides
that vary per bottle:
- `command` — the entrypoint claude-bottle runs inside the guest.
Chunk 1 ships a placeholder (`sleep infinity`); chunk 4 wires
the real `claude` entrypoint once provisioning is in place.
- `env = ["K=V", ...]` — agent's HTTPS_PROXY / NO_PROXY /
NODE_EXTRA_CA_CERTS, all using IP literals pointing at the
per-bottle sidecar bundle's pinned docker IP.
- `env` — the agent's HTTP_PROXY / NO_PROXY / CA paths, pointing
at `proxy.internal:<gateway-port>`. gvproxy resolves
`proxy.internal` to the gateway IP and port-forwards to the
host-side sidecar bundle.
- `[network] allow_cidrs = ["<bundle-ip>/32"]` — TSI's single-IP
allowlist. With this and no other `allow_*` or
`outbound-localhost-only`, the agent can dial exactly one IP:
the bundle. Host loopback, LAN, and the public internet
directly are all refused at the VMM layer.
- `[[net]]` — a virtio-net device backed by gvproxy's unixgram
socket via the VFKT handshake. This is the line that rejects
libkrun's TSI mode: TSI's CIDR allowlist permits the entire
127.0.0.0/8 of host loopback, which exposes every host-side
service; gvproxy's explicit port-forward list is the only thing
the guest can reach.
What the renderer does NOT emit:
The renderer is a pure function. Disk writes happen in
`prepare.py` via `smolfile_write`."""
- `image` / `entrypoint` / `cmd` — those come from the
`.smolmachine` artifact (produced by `smolvm pack create
--image claude-bottle:latest`) and don't vary across bottles
of the same agent image.
- `cpus` / `memory` — left at smolvm defaults until the
operator surfaces a need to override per bottle (the manifest
has no such field today).
Pure function; disk writes happen via `smolfile_write`."""
from __future__ import annotations
import json
from pathlib import Path
from typing import Any, Mapping
# Default port assignments INSIDE the gvproxy network — what the
# guest dials. The agent's HTTPS_PROXY etc. resolve to
# `proxy.internal:<one of these>`. Host-side mapping is dynamic
# (chunk 3 allocates loopback ports per bottle).
GVPROXY_PIPELOCK_GATEWAY_PORT = 8888
GVPROXY_GIT_GATE_GATEWAY_PORT = 8889
GVPROXY_SUPERVISE_GATEWAY_PORT = 8890
def smolfile_build(
*,
slug: str,
gvproxy_socket: Path,
env: Mapping[str, str],
command: tuple[str, ...] = ("sleep", "infinity"),
bundle_ip: str,
) -> dict[str, Any]:
"""Build the Smolfile config dict.
`gvproxy_socket` is the unixgram socket gvproxy listens on; the
guest's virtio-net device handshakes (VFKT magic) with it on
start. `env` is `{NAME: VALUE}` for the guest's process env.
`command` is the entrypoint argv inside the guest (placeholder
until chunk 4 — see module docstring).
Returns a TOML-shaped dict; render with `smolfile_render`."""
"""Build the Smolfile config dict. `env` is `{NAME: VALUE}` for
the guest's process env (IP-literal HTTPS_PROXY etc.).
`bundle_ip` is the pinned docker IP of the per-bottle sidecar
bundle; it lands in `[network] allow_cidrs` as a /32."""
return {
"name": f"claude-bottle-{slug}",
"command": list(command),
"env": [f"{k}={v}" for k, v in sorted(env.items())],
"net": [
{
"type": "virtio-net",
"attachment": "unixgram",
"socket": str(gvproxy_socket),
},
],
"network": {
"allow_cidrs": [f"{bundle_ip}/32"],
},
}
def smolfile_render(cfg: dict[str, Any]) -> str:
"""Render the Smolfile dict as TOML. Stdlib has `tomllib` for
reading TOML but no writer; the smolmachines schema we emit is
narrow enough (string scalars + string lists + one inline table
per net device) to render by hand. Avoids a `tomli_w` runtime
dep and keeps the project stdlib-only."""
"""Render the Smolfile dict as TOML. Schema is narrow (string
list + one table with a string list) so we render by hand and
stay stdlib-only."""
lines: list[str] = []
lines.append(f'name = {_toml_str(cfg["name"])}')
lines.append(f'command = {_toml_array(cfg["command"])}')
lines.append(f'env = {_toml_array(cfg["env"])}')
lines.append("")
for net in cfg.get("net", ()):
lines.append("[[net]]")
for key, value in net.items():
lines.append(f'{key} = {_toml_str(value)}')
lines.append("")
return "\n".join(lines).rstrip("\n") + "\n"
lines.append("[network]")
lines.append(f'allow_cidrs = {_toml_array(cfg["network"]["allow_cidrs"])}')
return "\n".join(lines) + "\n"
def smolfile_write(cfg: dict[str, Any], path: Path) -> Path:
@@ -96,16 +71,11 @@ def smolfile_write(cfg: dict[str, Any], path: Path) -> Path:
def _toml_str(value: Any) -> str:
"""TOML basic string: double-quoted with backslash + double-quote
escapes. The smolmachines fields we emit (slugs, paths, env
pairs) are ASCII-safe; the escape table covers what's reachable."""
escapes."""
s = str(value)
s = s.replace("\\", "\\\\").replace('"', '\\"')
return f'"{s}"'
def _toml_array(values: list[Any]) -> str:
"""TOML inline array. Uses `_toml_str` so quoting is consistent;
the alternative would be `json.dumps(values)` which renders
identical text for ASCII-only lists, but going through the same
quoter is one less surprise on future inputs."""
return "[" + ", ".join(_toml_str(v) for v in values) + "]"