[eric] tls: the Node children get the OS roots, so a proxy or endpoint-tool root no longer kills every model call (ENG-408)

This commit is contained in:
ciregenz
2026-08-27 14:01:30 -07:00
parent a1e1c4224e
commit 54ab6863ab
4 changed files with 320 additions and 3 deletions
@@ -251,6 +251,17 @@ async def configure_provider_env(
p_env = options_kwargs.get("env")
if isinstance(p_env, dict):
p_env.setdefault("ENABLE_TOOL_SEARCH", "auto")
# The CLI is Node and ignores the OS trust store, so on a machine whose endpoint tool or
# corporate proxy installed a root, sign-in succeeds (ENG-407 armed truststore for THIS
# process) and the first model call still dies on TLS. Additive and fail-open: empty dict
# when nothing can be exported, which is every build before this (ENG-408).
try:
from backend.config.node_trust import node_ca_env
from backend.config.paths import DATA_ROOT
for k, v in node_ca_env(os.path.join(DATA_ROOT, "node-ca-roots.pem")).items():
p_env.setdefault(k, v)
except Exception as e:
logger.debug("node trust: skipped for the CLI (%s)", e)
# Fault-injection seam: lets a QA harness front the provider with a local proxy (mid-run 401 drills, ENG-302 family). Absent in prod, so every branch's real base URL stands.
p_base_override = os.environ.get("OPENSWARM_ANTHROPIC_BASE_OVERRIDE")
+21 -3
View File
@@ -23,7 +23,7 @@ import subprocess
import sys
import tempfile
import time
from typing import Any, List, Optional
from typing import Any, Dict, List, Optional
import httpx
@@ -71,6 +71,22 @@ P_REQUEST_LOG_MAX_BYTES = 5 * 1024 * 1024
P_NODE_HEAP_MB = 4096
def p_node_ca_env() -> Dict[str, str]:
"""OS roots for the router, because Node ignores the OS trust store (ENG-408).
Additive: NODE_EXTRA_CA_CERTS adds to Node's bundled roots rather than replacing them, so a
machine that works today cannot be broken by it. Empty dict on any failure, which is exactly the
behaviour of every build before this.
"""
try:
from backend.config.node_trust import node_ca_env
from backend.config.paths import DATA_ROOT
return node_ca_env(os.path.join(DATA_ROOT, "node-ca-roots.pem"))
except Exception as e:
logger.debug("node trust: skipped for the router (%s)", e)
return {}
def p_rotate_request_log() -> None:
"""Rotate ~/.9router/request-details.json to a single .0 backup when it grows past the cap,
BEFORE 9Router is spawned (never racing a live writer). 9Router recreates a fresh file, exactly
@@ -600,7 +616,8 @@ async def p_ensure_running_impl():
logger.info("Starting 9Router (production) on port %d...", NINE_ROUTER_PORT)
cmd = [node, f"--max-old-space-size={P_NODE_HEAP_MB}"] + (["--require", p_patch] if p_patch else []) + [standalone_server]
cwd = os.path.dirname(standalone_server)
env = {**os.environ, "PORT": str(NINE_ROUTER_PORT), "NODE_ENV": "production"}
env = {**os.environ, "PORT": str(NINE_ROUTER_PORT), "NODE_ENV": "production",
**p_node_ca_env()}
if node == os.environ.get("OPENSWARM_ELECTRON_PATH"):
env["ELECTRON_RUN_AS_NODE"] = "1"
else:
@@ -618,7 +635,8 @@ async def p_ensure_running_impl():
)
cmd = [node, f"--max-old-space-size={P_NODE_HEAP_MB}"] + (["--require", p_patch] if p_patch else []) + [cached_server]
cwd = os.path.dirname(cached_server)
env = {**os.environ, "PORT": str(NINE_ROUTER_PORT), "NODE_ENV": "production"}
env = {**os.environ, "PORT": str(NINE_ROUTER_PORT), "NODE_ENV": "production",
**p_node_ca_env()}
# Capture stdout+stderr so a failed start can tell us WHY (the old DEVNULL default made every "router never came up" a silent mystery, which is the whole reason #90 was un-diagnosable). Packaged prod (NODE_ENV=production standalone) is quiet, so one fixed temp file, truncated each start attempt, won't grow; dev keeps its chatty-Next.js DEVNULL unless debug is set.
p_cap_path = os.path.join(tempfile.gettempdir(), "openswarm-9router-start.log")
+155
View File
@@ -0,0 +1,155 @@
"""Give the Node children the OS roots, because Node ignores the OS trust store.
ENG-407 armed `truststore` so THIS process trusts what the user's browser trusts. That was only the
first wall. 9router and the claude CLI are Node, and Node verifies against its own compiled-in root
list on both Windows and macOS, so on a machine whose endpoint tool or corporate proxy installed a
root, sign-in now succeeds and the first model call still dies with the ENG-218 certificate card
(ENG-408).
`NODE_USE_SYSTEM_CA=1` would make this file unnecessary, and it is the right long-term answer, but it
landed in Node 22.15 and the bundled runtime is v20.18.1 (`scripts/build-app.sh`). So we export the
OS roots to a PEM and point `NODE_EXTRA_CA_CERTS` at it.
Two properties this must have, and both are the reason it is safe to ship:
- **Additive, never substitutive.** `NODE_EXTRA_CA_CERTS` ADDS to Node's bundled roots rather than
replacing them, so a machine that works today cannot be broken by this. Trust becomes "what Node
shipped, plus what this machine's own OS store already trusts", and nothing wider.
- **Fails to today's behaviour.** Any failure (no export, empty export, unwritable path) leaves the
variable unset, which is exactly the current build. A guard that cannot arm says so and gets out of
the way; it never guesses.
"""
import logging
import os
import platform
import subprocess
from typing import List, Optional
from typeguard import typechecked
logger = logging.getLogger(__name__)
# Linux Node already reads OpenSSL's default paths, which is the system store there.
NEEDS_EXPORT = frozenset({"Darwin", "Windows"})
# macOS keeps the shipped roots and the admin-installed ones in different keychains, and an endpoint
# tool can land in either. The user's login keychain is included because that is where a per-user
# proxy root goes.
P_MAC_KEYCHAINS = (
"/System/Library/Keychains/SystemRootCertificates.keychain",
"/Library/Keychains/System.keychain",
)
PEM_HEADER = "-----BEGIN CERTIFICATE-----"
# A store this size is a sign we read something that is not a root list; refuse rather than hand Node
# a multi-megabyte file to parse on every spawn.
MAX_CERTS = 1000
@typechecked
def p_mac_roots() -> List[str]:
"""Every PEM block in the system keychains plus the user's login keychain."""
p_out: List[str] = []
p_chains = list(P_MAC_KEYCHAINS)
p_login = os.path.expanduser("~/Library/Keychains/login.keychain-db")
if os.path.exists(p_login):
p_chains.append(p_login)
for chain in p_chains:
if not os.path.exists(chain):
continue
try:
r = subprocess.run(
["security", "find-certificate", "-a", "-p", chain],
capture_output=True, text=True, timeout=20,
)
except Exception as e:
logger.debug("node trust: could not read %s (%s)", chain, e)
continue
if r.returncode == 0 and PEM_HEADER in r.stdout:
p_out.append(r.stdout)
return p_out
@typechecked
def p_windows_roots() -> List[str]:
"""The Windows ROOT and CA stores, via stdlib; no shelling out and no extra dependency."""
import ssl
p_out: List[str] = []
# enum_certificates exists only on Windows, so it is fetched by name; a checker on any other
# platform is right that the attribute is absent, and this function never runs there.
p_enum = getattr(ssl, "enum_certificates", None)
if p_enum is None:
return p_out
for store in ("ROOT", "CA"):
try:
for cert, enc, p_trust in p_enum(store):
if enc == "x509_asn" and p_trust:
p_out.append(ssl.DER_cert_to_PEM_cert(cert))
except Exception as e:
logger.debug("node trust: could not read the %s store (%s)", store, e)
return p_out
@typechecked
def export_os_roots(dest: str) -> Optional[str]:
"""Write the OS roots to `dest` and return the path, or None when there is nothing to hand Node.
None is the safe answer everywhere: the caller leaves NODE_EXTRA_CA_CERTS unset, and Node keeps
exactly the roots it ships with, which is what every build does today.
"""
system = platform.system()
if system not in NEEDS_EXPORT:
return None
p_blocks = p_mac_roots() if system == "Darwin" else p_windows_roots()
p_pem = "\n".join(b.strip() for b in p_blocks if b.strip())
p_count = p_pem.count(PEM_HEADER)
if p_count == 0:
logger.warning(
"node trust: read no roots out of the %s store, so 9router and the CLI keep Node's "
"bundled roots only; a proxy or security-tool root will still be refused by them", system,
)
return None
if p_count > MAX_CERTS:
logger.warning("node trust: %d certificates is not a root list; refusing to export", p_count)
return None
try:
os.makedirs(os.path.dirname(dest), exist_ok=True)
with open(dest, "w", encoding="utf-8") as f:
f.write(p_pem + "\n")
except Exception as e:
logger.warning(
"node trust: could not write %s (%s); 9router and the CLI keep Node's bundled roots only",
dest, e,
)
return None
logger.info("node trust: exported %d OS roots for the Node children -> %s", p_count, dest)
return dest
# Exported once per process. Every spawn used to shell out to `security find-certificate` and rewrite
# ~250KB; measured three times in five seconds on one turn. The OS store does not change mid-session,
# and if a user installs a root while the app is open, a restart picks it up -- the same contract
# truststore already has for this process.
P_CACHED: Optional[dict] = None
@typechecked
def node_ca_env(dest: str) -> dict:
"""`{NODE_EXTRA_CA_CERTS: <path>}` when there is something to add, else `{}`.
Returning a dict rather than mutating os.environ keeps this out of the parent process: the
backend's own TLS is truststore's job, and two mechanisms for one concern is how they drift.
"""
global P_CACHED
if P_CACHED is not None:
return dict(P_CACHED)
p_path = export_os_roots(dest)
P_CACHED = {"NODE_EXTRA_CA_CERTS": p_path} if p_path else {}
return dict(P_CACHED)
@typechecked
def reset_cache_for_test() -> None:
global P_CACHED
P_CACHED = None
+133
View File
@@ -0,0 +1,133 @@
"""The Node children get the OS roots, because Node ignores the OS trust store.
ENG-407 armed truststore so the PYTHON backend trusts what the browser trusts. 9router and the claude
CLI are Node and verify against Node's compiled-in list, so on a machine with an endpoint tool's or a
corporate proxy's root, sign-in now works and the first model call still dies (ENG-408).
`NODE_USE_SYSTEM_CA=1` would make the whole file unnecessary; it landed in Node 22.15 and the bundled
runtime is v20.18.1, so we export the OS roots and point NODE_EXTRA_CA_CERTS at them.
DRILLED 2026-08-27 against a real Node TLS server signed by a throwaway root:
stock Node -> REFUSED UNABLE_TO_VERIFY_LEAF_SIGNATURE (the ThinkPad bug)
NODE_EXTRA_CA_CERTS incl. root -> OK 200 (the cure)
NODE_EXTRA_CA_CERTS OS-only -> REFUSED (trust not widened)
"""
import os
import platform
import pytest
from backend.config import node_trust
@pytest.fixture(autouse=True)
def p_no_cache():
"""The export is cached once per process; a cached answer would make every case below a no-op."""
node_trust.reset_cache_for_test()
yield
node_trust.reset_cache_for_test()
ROUTER = "backend/apps/nine_router/process.py"
CLI = "backend/apps/agents/manager/configure_provider_env.py"
def test_an_export_that_yields_nothing_sets_no_variable(tmp_path, monkeypatch):
"""Fail to today's behaviour. An unset variable is Node's bundled roots, which is every build
before this; a broken export must never be louder than that."""
monkeypatch.setattr(node_trust, "p_mac_roots", lambda: [])
monkeypatch.setattr(node_trust, "p_windows_roots", lambda: [])
monkeypatch.setattr(platform, "system", lambda: "Darwin")
assert node_trust.node_ca_env(str(tmp_path / "r.pem")) == {}
def test_an_unwritable_destination_sets_no_variable(monkeypatch):
monkeypatch.setattr(platform, "system", lambda: "Darwin")
monkeypatch.setattr(node_trust, "p_mac_roots", lambda: ["-----BEGIN CERTIFICATE-----\nx\n"])
assert node_trust.node_ca_env("/proc/nope/cannot/write.pem") == {}
def test_linux_is_left_alone(monkeypatch, tmp_path):
"""Node on Linux already reads OpenSSL's default paths, which IS the system store there."""
monkeypatch.setattr(platform, "system", lambda: "Linux")
assert node_trust.export_os_roots(str(tmp_path / "r.pem")) is None
def test_an_absurd_store_is_refused_rather_than_handed_to_node(monkeypatch, tmp_path):
monkeypatch.setattr(platform, "system", lambda: "Darwin")
monkeypatch.setattr(node_trust, "p_mac_roots",
lambda: ["-----BEGIN CERTIFICATE-----\nx\n" * (node_trust.MAX_CERTS + 1)])
assert node_trust.export_os_roots(str(tmp_path / "r.pem")) is None
def test_a_real_export_on_this_machine_produces_parseable_pem(tmp_path):
"""Skipped off the two platforms that need it; on them it must produce real certificates."""
if platform.system() not in node_trust.NEEDS_EXPORT:
return
p = node_trust.export_os_roots(str(tmp_path / "roots.pem"))
assert p and os.path.exists(p)
body = open(p).read()
assert body.count(node_trust.PEM_HEADER) > 10, "an OS root store is not this small"
assert body.count(node_trust.PEM_HEADER) == body.count("-----END CERTIFICATE-----")
def test_both_node_children_are_wired_not_just_one():
"""The recurring defect: a fix applied to one spawn site and not the other. The router and the
CLI are two separate Node processes and BOTH verify provider TLS."""
for path in (ROUTER, CLI):
src = open(path).read()
assert "node_ca_env" in src, f"{path} spawns Node without the OS roots"
def test_it_is_additive_and_never_replaces_the_env():
"""setdefault, not assignment: a lane that deliberately set its own CA path keeps it, and this
can never blank a variable someone else needed."""
src = open(CLI).read()
i = src.index("node_ca_env(")
assert "p_env.setdefault(k, v)" in src[i:i + 300]
def test_the_export_happens_once_per_process(monkeypatch, tmp_path):
"""It used to shell out to `security` and rewrite ~250KB on EVERY spawn: measured three times in
five seconds on one turn."""
calls = {"n": 0}
def p_count():
calls["n"] += 1
return ["-----BEGIN CERTIFICATE-----\nx\n-----END CERTIFICATE-----\n"]
monkeypatch.setattr(platform, "system", lambda: "Darwin")
monkeypatch.setattr(node_trust, "p_mac_roots", p_count)
dest = str(tmp_path / "r.pem")
a = node_trust.node_ca_env(dest)
b = node_trust.node_ca_env(dest)
c = node_trust.node_ca_env(dest)
assert calls["n"] == 1, f"exported {calls['n']} times"
assert a == b == c and a
def test_a_failed_export_is_cached_too_and_does_not_retry_every_spawn(monkeypatch, tmp_path):
"""The failure path is the one that would otherwise shell out forever on a machine where the
store cannot be read at all."""
calls = {"n": 0}
def p_empty():
calls["n"] += 1
return []
monkeypatch.setattr(platform, "system", lambda: "Darwin")
monkeypatch.setattr(node_trust, "p_mac_roots", p_empty)
dest = str(tmp_path / "r.pem")
assert node_trust.node_ca_env(dest) == {}
assert node_trust.node_ca_env(dest) == {}
assert calls["n"] == 1
def test_the_caller_cannot_mutate_the_cache(monkeypatch, tmp_path):
monkeypatch.setattr(platform, "system", lambda: "Darwin")
monkeypatch.setattr(node_trust, "p_mac_roots",
lambda: ["-----BEGIN CERTIFICATE-----\nx\n-----END CERTIFICATE-----\n"])
dest = str(tmp_path / "r.pem")
first = node_trust.node_ca_env(dest)
first["NODE_EXTRA_CA_CERTS"] = "/tmp/evil.pem"
assert node_trust.node_ca_env(dest)["NODE_EXTRA_CA_CERTS"] != "/tmp/evil.pem"