Files
bot-bottle/claude_bottle/pipelock.py
T
didericis 3be70eb07a
test / unit (pull_request) Successful in 17s
test / integration (pull_request) Successful in 1m7s
feat(supervise): list-egress-proxy-routes MCP tool, defaults on egress-proxy
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>
2026-05-25 18:23:01 -04:00

391 lines
16 KiB
Python

"""Pipelock sidecar lifecycle for the per-agent egress topology.
Pipelock (https://github.com/luckyPipewrench/pipelock) is an HTTP
forward proxy with hostname allowlisting + DLP scanning + URL-entropy
checks. One sidecar per agent, attached to the agent's --internal
network and a per-agent user-defined egress bridge.
Post-PRD-0017 topology: the agent's HTTP_PROXY points at egress-proxy
(not pipelock); egress-proxy sets `HTTPS_PROXY=pipelock` on its
outbound leg. So pipelock no longer sees the agent's connections
directly — it sees the egress-proxy → upstream leg, applies the
hostname allowlist + DLP body scan there, and forwards to the real
upstream.
Image pin: ghcr.io/luckypipewrench/pipelock@sha256:<digest> for tag 2.3.0.
"""
from __future__ import annotations
from abc import ABC, abstractmethod
from dataclasses import dataclass
from pathlib import Path
from typing import cast
from .egress_proxy import (
DEFAULT_ALLOWLIST,
EGRESS_PROXY_HOSTNAME,
egress_proxy_routes_for_bottle,
)
from .supervise import SUPERVISE_HOSTNAME
from .manifest import Bottle
# 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
# trigger DLP scanners (notably the BIP-39 seed-phrase detector, which
# fires on any 12+ consecutive English words that happen to be on the
# BIP-39 wordlist and pass the checksum). Per pipelock's own
# configuration.md, the recommended treatment for LLM API endpoints is
# `passthrough_domains`: pipelock still proxies the CONNECT (so the
# api_allowlist gate applies), but it does not generate a leaf cert or
# decrypt the body. Body scanning happens on hosts that aren't
# passthrough'd, so DLP protection against agent exfil to other
# allowlisted hosts is unchanged.
DEFAULT_TLS_PASSTHROUGH: tuple[str, ...] = (
"api.anthropic.com",
)
# --- Allowlist resolution --------------------------------------------------
def pipelock_bottle_allowlist(bottle: Bottle) -> list[str]:
"""Hostnames in bottle.egress.allowlist."""
return list(bottle.egress.allowlist)
def pipelock_route_hosts(bottle: Bottle) -> list[str]:
"""Hostnames declared in `bottle.egress_proxy.routes`. Returned
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]:
"""Hostnames pipelock allows. Sorted for stability.
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] = {}
if bottle.egress_proxy.routes:
# 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())
def pipelock_seed_phrase_detection_enabled(bottle: Bottle) -> bool:
"""Whether pipelock's BIP-39 seed-phrase detector stays on for
this bottle.
LLM conversation bodies legitimately trip the detector — any 12+
English words that pass the BIP-39 checksum match — so any
bottle that routes claude through pipelock's body scanner gets
blocked on the first real chat. We tried two narrower knobs
first:
- `suppress: [{rule, path}]` — pipelock accepts the schema
but the entry only silences the alert; the body_dlp block
still fires.
- `rules.disabled: ["dlp:BIP-39 Seed Phrase"]` — same shape,
same outcome: 403 still returned.
Empirically only `seed_phrase_detection.enabled: false`
actually stops the block (verified by sending a 12-word BIP-39
body through three pipelock instances). It is a global toggle —
no per-path / per-host knob in pipelock 2.3.0 — so we turn the
detector off for the entire bottle when the bottle declares an
egress-proxy route to `api.anthropic.com`. The trade-off is
accepted: BIP-39 detection has little value in claude-bottle's
threat model (the agent has no access to a user's crypto wallet
seeds; the patterns that matter — gh*_, sk-ant-, AKIA, etc. —
keep firing)."""
return not any(
r.Host == "api.anthropic.com" for r in bottle.egress_proxy.routes
)
def pipelock_effective_tls_passthrough(bottle: Bottle) -> list[str]:
"""Hostnames pipelock should pass through (no TLS MITM, no body
scan). Default carries the LLM API endpoint — its request bodies
are user-authored conversation text that legitimately trips DLP
scanners (notably pipelock's BIP-39 seed-phrase detector). Every
other allowlisted host is MITM'd by pipelock's per-bottle CA so
its body scanner sees the cleartext.
egress-proxy route hosts (github, gitea, npm) are deliberately
NOT auto-added here. egress-proxy's HTTPS client trusts pipelock's
CA at runtime (folded into its trust store via docker cp), so
pipelock MITMs and body-scans the egress-proxy → upstream leg the
same way it body-scanned the agent's direct HTTPS traffic before
the PRD 0017 cutover.
`bottle` is kept on the signature for forward-compat (a future
knob might let a manifest opt a host into passthrough); today
the returned list is independent of the bottle."""
del bottle # not consulted; see docstring.
return sorted(DEFAULT_TLS_PASSTHROUGH)
def pipelock_allowlist_summary(bottle: Bottle) -> str:
"""One-line summary for the y/N preflight display:
"<N> hosts allowed (host1, host2, host3, +M more)"."""
hosts = pipelock_effective_allowlist(bottle)
count = len(hosts)
if count == 0:
return "0 hosts allowed (none)"
show = count
more = 0
if count > 5:
show = 3
more = count - show
joined = ", ".join(hosts[:show])
if more > 0:
return f"{count} hosts allowed ({joined}, +{more} more)"
return f"{count} hosts allowed ({joined})"
# --- Config build + YAML render --------------------------------------------
def pipelock_build_config(
bottle: Bottle,
*,
ca_cert_path: str = "",
ca_key_path: str = "",
ssrf_ip_allowlist: tuple[str, ...] = (),
) -> dict[str, object]:
"""Build the structured pipelock config dict the sidecar will load.
Deliberately carries no env values, no secrets, no per-agent
customization beyond the resolved hostname list. The shape mirrors
the YAML pipelock expects on disk; `pipelock_render_yaml` serializes
it. Tests assert on this dict; production code renders it.
`ca_cert_path` / `ca_key_path` are the **in-container** paths the
pipelock sidecar will read its CA from at runtime (they're
populated into the container at start time via `docker cp`).
Pass both or neither: both → emit `tls_interception` block with
`enabled: true`; neither → omit the block entirely (pipelock
falls back to its built-in default of `enabled: false`). Used
by PRD 0006 to turn on pipelock's native TLS interception.
`ssrf_ip_allowlist` is the list of IPs / CIDRs that bypass
pipelock's SSRF guard. Pipelock blocks RFC1918-resolved
destinations by default, which would catch sibling-sidecar
traffic on the bottle's internal Docker network in 172.x space
(e.g. egress-proxy → pipelock on the upstream leg). Pass the
bottle's internal network CIDR here so internal-network requests
pass through pipelock while api_allowlist + body-scanning still
apply. Empty by default; omitted from the rendered yaml when
empty so pipelock keeps its built-in SSRF defaults."""
cfg: dict[str, object] = {
"version": 1,
"mode": "strict",
"enforce": True,
"api_allowlist": pipelock_effective_allowlist(bottle),
"forward_proxy": {"enabled": True},
}
if not pipelock_seed_phrase_detection_enabled(bottle):
cfg["seed_phrase_detection"] = {"enabled": False}
cfg["dlp"] = {"include_defaults": True, "scan_env": True}
# Body-scan enforcement is a separate pipelock section (each DLP
# "surface" — body, MCP, response — has its own action). Pipelock's
# built-in default for request_body_scanning is "warn" (forward
# with a log line); claude-bottle's default is "block" so a hit
# actually stops the request from leaving the egress network.
cfg["request_body_scanning"] = {"action": bottle.egress.dlp_action}
if ca_cert_path or ca_key_path:
if not (ca_cert_path and ca_key_path):
raise ValueError(
"pipelock_build_config: pass both ca_cert_path and ca_key_path "
"to enable tls_interception, or neither to leave it off"
)
cfg["tls_interception"] = {
"enabled": True,
"ca_cert": ca_cert_path,
"ca_key": ca_key_path,
"passthrough_domains": pipelock_effective_tls_passthrough(bottle),
}
if ssrf_ip_allowlist:
cfg["ssrf"] = {"ip_allowlist": list(ssrf_ip_allowlist)}
return cfg
def pipelock_render_yaml(cfg: dict[str, object]) -> str:
"""Render a pipelock config dict (as produced by
`pipelock_build_config`) as YAML. Hand-rolled so we don't take a
YAML-parser dependency for a fixed, narrow shape."""
def _bool(b: object) -> str:
return "true" if b else "false"
lines: list[str] = []
lines.append(f"version: {cfg['version']}")
lines.append(f"mode: {cfg['mode']}")
lines.append(f"enforce: {_bool(cfg['enforce'])}")
lines.append("")
lines.append("api_allowlist:")
for h in cast(list[str], cfg["api_allowlist"]):
lines.append(f' - "{h}"')
lines.append("")
if "seed_phrase_detection" in cfg:
lines.append("seed_phrase_detection:")
spd = cast(dict[str, object], cfg["seed_phrase_detection"])
lines.append(f" enabled: {_bool(spd['enabled'])}")
lines.append("")
lines.append("forward_proxy:")
fp = cast(dict[str, object], cfg["forward_proxy"])
lines.append(f" enabled: {_bool(fp['enabled'])}")
lines.append("")
lines.append("dlp:")
dlp = cast(dict[str, object], cfg["dlp"])
lines.append(f" include_defaults: {_bool(dlp['include_defaults'])}")
lines.append(f" scan_env: {_bool(dlp['scan_env'])}")
lines.append("")
lines.append("request_body_scanning:")
rbs = cast(dict[str, object], cfg["request_body_scanning"])
lines.append(f' action: "{rbs["action"]}"')
if "tls_interception" in cfg:
lines.append("")
lines.append("tls_interception:")
tls = cast(dict[str, object], cfg["tls_interception"])
lines.append(f" enabled: {_bool(tls['enabled'])}")
lines.append(f' ca_cert: "{tls["ca_cert"]}"')
lines.append(f' ca_key: "{tls["ca_key"]}"')
passthrough = cast(list[str], tls.get("passthrough_domains", []))
if passthrough:
lines.append(" passthrough_domains:")
for d in passthrough:
lines.append(f' - "{d}"')
if "ssrf" in cfg:
lines.append("")
lines.append("ssrf:")
ssrf = cast(dict[str, object], cfg["ssrf"])
lines.append(" ip_allowlist:")
for ip in cast(list[str], ssrf["ip_allowlist"]):
lines.append(f' - "{ip}"')
return "\n".join(lines) + "\n"
# --- Proxy class -----------------------------------------------------------
@dataclass(frozen=True)
class PipelockProxyPlan:
"""Output of PipelockProxy.prepare; consumed by .start when the
sidecar needs to be brought up.
yaml_path + slug are filled in at prepare time (host-side, side-
effect-free; the YAML references the in-container CA paths
already so it doesn't need the host paths to be valid). The
remaining fields are populated by the backend's launch step
via `dataclasses.replace`: internal/egress networks once
those networks exist, the CA host paths once the one-shot
`pipelock tls init` has run, and `internal_network_cidr` once
Docker has assigned a subnet to the internal network. Empty
defaults are sentinels meaning "not yet set"; `.start` validates
that they are populated.
`internal_network_cidr` ends up on pipelock's `ssrf.ip_allowlist`
so traffic from sibling sidecars (egress-proxy → pipelock on the
upstream leg, etc.) bypasses pipelock's RFC1918 SSRF guard while
api_allowlist and body-scanning still apply."""
yaml_path: Path
slug: str
internal_network: str = ""
internal_network_cidr: str = ""
egress_network: str = ""
ca_cert_host_path: Path = Path()
ca_key_host_path: Path = Path()
class PipelockProxy(ABC):
"""The pipelock egress proxy. Encapsulates the YAML-config
generation; the sidecar's start/stop lifecycle is backend-specific
and lives on concrete subclasses.
The class-level constants `CA_CERT_IN_CONTAINER` /
`CA_KEY_IN_CONTAINER` are the in-container paths the YAML config
references — they correspond to wherever the backend's `.start`
places the CA cert and key inside the sidecar. Subclasses
override the constants."""
CA_CERT_IN_CONTAINER: str = ""
CA_KEY_IN_CONTAINER: str = ""
def prepare(
self, bottle: Bottle, slug: str, stage_dir: Path
) -> PipelockProxyPlan:
"""Write the pipelock yaml config (mode 600) under `stage_dir`
and return the plan for `.start`. Pure host-side, no docker
subprocess.
`slug` is the agent-derived identifier (lowercased,
hyphen-normalized) used as the suffix in every per-agent
resource name — the agent container, the pipelock container
(`claude-bottle-pipelock-<slug>`), the internal/egress
networks. It's stored on the returned plan so the backend's
start step can derive the sidecar's container name.
The CA paths the YAML references are the in-container paths
from the concrete subclass's class-level constants. The
host-side counterparts are generated by the launch step
(not here, so prepare stays side-effect-free on docker) and
added to the plan via `dataclasses.replace` before `.start`."""
yaml_path = stage_dir / "pipelock.yaml"
cfg = pipelock_build_config(
bottle,
ca_cert_path=self.CA_CERT_IN_CONTAINER,
ca_key_path=self.CA_KEY_IN_CONTAINER,
)
yaml_path.write_text(pipelock_render_yaml(cfg))
yaml_path.chmod(0o600)
return PipelockProxyPlan(yaml_path=yaml_path, slug=slug)
@abstractmethod
def start(self, plan: PipelockProxyPlan) -> str:
"""Bring up the pipelock sidecar according to `plan`. Returns
the proxy_target string identifying the running instance — the
same value to pass to `.stop`. Backend-specific."""
@abstractmethod
def stop(self, proxy_target: str) -> None:
"""Tear down the pipelock sidecar identified by `proxy_target`
(the value `.start` returned). Idempotent: a missing target is
success. Backend-specific."""