feat(supervise): list-egress-proxy-routes MCP tool, defaults on egress-proxy
test / unit (pull_request) Successful in 17s
test / integration (pull_request) Successful in 1m7s

Reshape the allowlist topology so the egress-proxy is the bottle's
single allowlist surface, and replace the agent-side
routes/allowlist file mounts with a live MCP tool.

Policy change (move defaults to egress-proxy):

  - `egress_proxy_routes_for_bottle(bottle)` now folds in
    DEFAULT_ALLOWLIST (the claude-code defaults) and
    `bottle.egress.allowlist` (user adds) as bare-pass routes (no
    auth, no path filter), on top of the bottle's
    `egress_proxy.routes`. Manifest routes win on host collision.
  - `pipelock_effective_allowlist(bottle)` mirrors egress-proxy's
    effective host set when egress-proxy is in use. Pipelock is
    no longer the bottle's primary allowlist authority; it
    enforces a downstream copy as defense-in-depth + does DLP body
    scanning.
  - Split out `egress_proxy_manifest_routes(bottle)` for callers
    that want just the manifest entries (tests, internal use).
  - DEFAULT_ALLOWLIST moves from `pipelock.py` to `egress_proxy.py`
    (pipelock re-imports for the no-egress-proxy fallback path).
  - Dropped the `egress-proxy` auto-allow on pipelock's allowlist
    — the agent never dials egress-proxy via the proxy mechanism;
    pipelock only sees upstream hostnames from egress-proxy's
    CONNECTs.

Introspection endpoint (existing mitmproxy feature):

  - Egress-proxy addon recognises requests to the magic host
    `_egress-proxy.local` and synthesizes responses via
    `flow.response = http.Response.make(...)` — no upstream
    connection, no allowlist enforcement on the magic host.
  - `GET /allowlist` returns the in-memory route table as JSON
    (host + path_allowlist + auth_scheme + token_env per route;
    no token VALUES).
  - Smoke-tested end-to-end against a real egress-proxy container.

MCP tool (existing supervise plumbing):

  - New `list-egress-proxy-routes` tool (no inputs, no operator
    approval). Handler fetches via egress-proxy's introspection
    endpoint using urllib's ProxyHandler against
    `EGRESS_PROXY_FORWARD_PROXY`. Returns the JSON payload as the
    tool's text content; `isError: true` if the proxy is
    unreachable.
  - `egress-proxy-block` description now points the agent at
    `list-egress-proxy-routes` instead of a staged file path.
  - `pipelock-block` description acknowledges the mirror — agents
    should prefer `egress-proxy-block` to add hosts; pipelock-block
    stays for the rare divergence case.

Drop agent-side file mounts:

  - Supervise's `current-config` dir staging no longer writes
    routes.yaml / allowlist. Only `Dockerfile` remains
    (capability-block still reads it from
    `/etc/claude-bottle/current-config/Dockerfile`).
  - `prepare.py` stops passing `routes_content` /
    `allowlist_content` to `supervise.prepare`.
  - `Supervise.prepare` signature simplified to one
    `dockerfile_content` kwarg.

Tests: 400 unit + integration pass. Added coverage for
defaults-folding (`TestRoutesForBottleFoldsDefaults`), the new
tool definition + handler, and the updated supervise.prepare
shape.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-05-25 18:23:01 -04:00
parent 1cec0d9aa6
commit 3be70eb07a
12 changed files with 410 additions and 144 deletions
+42 -41
View File
@@ -22,21 +22,14 @@ from dataclasses import dataclass
from pathlib import Path
from typing import cast
from .egress_proxy import EGRESS_PROXY_HOSTNAME
from .egress_proxy import (
DEFAULT_ALLOWLIST,
EGRESS_PROXY_HOSTNAME,
egress_proxy_routes_for_bottle,
)
from .supervise import SUPERVISE_HOSTNAME
from .manifest import Bottle
# Baked-in default allowlist for hosts Claude Code itself needs.
DEFAULT_ALLOWLIST: tuple[str, ...] = (
"api.anthropic.com",
"statsig.anthropic.com",
"sentry.io",
"claude.ai",
"platform.claude.com",
"downloads.claude.ai",
"raw.githubusercontent.com",
)
# Hosts pipelock should NOT TLS-MITM, even when tls_interception is
# enabled. The Claude API endpoint is an LLM provider — its request
# bodies are user-authored conversation text that legitimately can
@@ -64,43 +57,51 @@ def pipelock_bottle_allowlist(bottle: Bottle) -> list[str]:
def pipelock_route_hosts(bottle: Bottle) -> list[str]:
"""Hostnames declared in `bottle.egress_proxy.routes`. Returned
sorted + deduped.
Post-cutover topology (PRD 0017): the agent's HTTPS_PROXY points
at egress-proxy, not pipelock; egress-proxy's outbound leg sets
`HTTPS_PROXY=pipelock`. So pipelock no longer terminates the
agent's connections — it sees the egress-proxy → upstream leg
only. Each declared route's host still needs to be on pipelock's
allowlist so that leg can leave the egress network."""
sorted + deduped. Used by the no-egress-proxy fallback path
below; bottles that DO use egress-proxy include the same hosts
via `egress_proxy_routes_for_bottle`."""
hosts = {r.Host for r in bottle.egress_proxy.routes if r.Host}
return sorted(hosts)
def pipelock_effective_allowlist(bottle: Bottle) -> list[str]:
"""Deduplicated union of: baked-in defaults, bottle.egress.allowlist,
the egress-proxy route hosts (from bottle.egress_proxy.routes), the
egress-proxy sidecar's own hostname when any route is declared, and
the supervise sidecar's hostname when bottle.supervise is enabled.
Sorted for stability. Git upstreams declared in `bottle.git` do NOT
contribute here — git traffic flows through the per-agent git-gate
sidecar (PRD 0008), not pipelock.
"""Hostnames pipelock allows. Sorted for stability.
The egress-proxy + supervise hostnames are auto-added because the
sidecars sit on the bottle's internal network alongside the agent;
requests that pass through pipelock for `egress-proxy:9099` or
`supervise:9100` (e.g. when egress-proxy uses HTTPS_PROXY=pipelock
on its upstream leg) would otherwise be 403'd by pipelock's
hostname gate."""
Two paths, depending on whether the bottle uses egress-proxy:
- Bottle declares `egress_proxy.routes[]` → agent's HTTPS_PROXY
points at egress-proxy. Egress-proxy is the bottle's primary
allowlist gate (DEFAULT_ALLOWLIST + bottle.egress.allowlist +
manifest routes all live there as bare-pass or full routes,
folded in by `egress_proxy_routes_for_bottle`). Pipelock's
allowlist is then a MIRROR of egress-proxy's hosts — same
set, just serving as the defense-in-depth hostname gate +
DLP scanner on the upstream leg.
- Bottle has no `egress_proxy.routes[]` → agent talks straight
to pipelock. Pipelock keeps its previous behavior: bake in
DEFAULT_ALLOWLIST + bottle.egress.allowlist for claude-code
defaults.
The supervise sidecar's hostname is auto-added when supervise
is enabled (sibling-sidecar traffic that flows through pipelock
would otherwise be 403'd). Git upstreams declared in
`bottle.git` do NOT contribute here — git traffic flows
through git-gate (PRD 0008), not pipelock."""
seen: dict[str, None] = {}
for h in DEFAULT_ALLOWLIST:
seen.setdefault(h, None)
for h in pipelock_bottle_allowlist(bottle):
if h:
seen.setdefault(h, None)
for h in pipelock_route_hosts(bottle):
seen.setdefault(h, None)
if bottle.egress_proxy.routes:
seen.setdefault(EGRESS_PROXY_HOSTNAME, None)
# Mirror egress-proxy's effective host set — same defaults
# and bottle.egress.allowlist entries are already folded in
# at the egress-proxy layer; we don't add them twice.
for r in egress_proxy_routes_for_bottle(bottle):
if r.host:
seen.setdefault(r.host, None)
else:
for h in DEFAULT_ALLOWLIST:
seen.setdefault(h, None)
for h in pipelock_bottle_allowlist(bottle):
if h:
seen.setdefault(h, None)
if bottle.supervise:
seen.setdefault(SUPERVISE_HOSTNAME, None)
return sorted(seen.keys())