Files
aiturk-hermes-ide/agent/process_bootstrap.py

652 lines
24 KiB
Python

"""Process-level bootstrap helpers for ``run_agent``.
Three concerns, all tied to ``AIAgent`` boot-time / runtime IO setup:
1. **Lazy OpenAI SDK import** — ``_load_openai_cls`` + ``_OpenAIProxy``
defer the 240ms-ish ``from openai import OpenAI`` cost until first use,
while preserving ``isinstance(client, OpenAI)`` checks and
``patch("run_agent.OpenAI", ...)`` test patterns.
2. **Crash-resistant stdio** — ``_SafeWriter`` wraps stdout/stderr so
``OSError: Input/output error`` from broken pipes (systemd, Docker,
thread teardown races) cannot crash the agent. ``_install_safe_stdio``
applies the wrapper.
3. **HTTP proxy resolution** — ``_get_proxy_from_env`` reads
``HTTPS_PROXY`` / ``HTTP_PROXY`` / ``ALL_PROXY``;
``_get_proxy_for_base_url`` respects ``NO_PROXY`` for the given base URL.
4. **Codex dual-stack resilience** — the synchronous ChatGPT/Codex transport
races resolved IPv6/IPv4 addresses so a blackholed family cannot exhaust
the request watchdog before a working address is attempted.
``run_agent`` re-exports every name so existing
``from run_agent import _get_proxy_from_env`` imports keep working
unchanged.
"""
from __future__ import annotations
import errno
import os
import selectors
import socket
import sys
import threading
import time
import urllib.request
from typing import Any, Optional
from utils import base_url_hostname, normalize_proxy_url
# Cached at module level so we only pay the OpenAI SDK import cost once
# per process (after the first lazy load).
_OPENAI_CLS_CACHE = None
_HAPPY_EYEBALLS_DELAY_SECONDS = 0.25
# Process-wide pool of sync ``httpx.HTTPTransport`` objects shared by every
# keepalive client with the same (verify, proxy, happy-eyeballs) identity.
# Each delegated child AIAgent used to get its own transport = its own TLS
# pool, so a fan-out of N children held N separate socket sets to the same
# provider. Bounded: past the cap, callers get a private transport again.
_SHARED_TRANSPORTS: dict[tuple, Any] = {}
_SHARED_TRANSPORTS_LOCK = threading.Lock()
_SHARED_TRANSPORTS_MAX = 32
# ``request.extensions`` key stamped by ``_SharedTransport.handle_request``;
# the socket-abort walker in agent_runtime_helpers uses it to find only the
# owning client's in-flight connections on a shared pool.
HERMES_TRANSPORT_OWNER_EXT = "hermes_transport_owner"
def _interleave_addrinfos(addrinfos: list[tuple]) -> list[tuple]:
"""Interleave resolved address families while preserving resolver order."""
queues: dict[int, list[tuple]] = {}
family_order: list[int] = []
seen: set[tuple] = set()
for addrinfo in addrinfos:
family, socktype, proto, _canonname, sockaddr = addrinfo
marker = (family, socktype, proto, sockaddr)
if marker in seen:
continue
seen.add(marker)
if family not in queues:
queues[family] = []
family_order.append(family)
queues[family].append(addrinfo)
interleaved: list[tuple] = []
while any(queues.values()):
for family in family_order:
if queues[family]:
interleaved.append(queues[family].pop(0))
return interleaved
def _happy_eyeballs_create_connection(
address: tuple[str, int],
timeout: Optional[float],
source_address: Optional[tuple[str, int]] = None,
socket_options=(),
):
"""Connect using staggered non-blocking attempts across resolved families.
``socket.create_connection`` tries every address serially. A host with
broken-but-advertised IPv6 can therefore consume the full connect timeout
for each AAAA record before trying a working IPv4 address. This follows the
Happy Eyeballs shape from RFC 8305: retain resolver preference, interleave
families, and start the next candidate after a short delay.
"""
host, port = address
addrinfos = _interleave_addrinfos(
socket.getaddrinfo(host, port, type=socket.SOCK_STREAM)
)
if not addrinfos:
raise OSError(f"getaddrinfo returned no addresses for {host}")
selector = selectors.DefaultSelector()
active: set[socket.socket] = set()
winner = None
last_error: Optional[OSError] = None
deadline = None if timeout is None else time.monotonic() + max(timeout, 0.0)
next_launch = time.monotonic()
pending = list(addrinfos)
in_progress = {
0,
errno.EINPROGRESS,
errno.EWOULDBLOCK,
errno.EALREADY,
errno.EINTR,
getattr(errno, "WSAEWOULDBLOCK", 10035),
}
def start_attempt(addrinfo):
family, socktype, proto, _canonname, sockaddr = addrinfo
candidate = socket.socket(family, socktype, proto)
try:
if source_address is not None:
local_infos = socket.getaddrinfo(
source_address[0],
source_address[1],
family=family,
type=socktype,
)
if not local_infos:
raise OSError(
f"getaddrinfo returned no local {family} address for "
f"{source_address[0]}"
)
candidate.bind(local_infos[0][4])
candidate.setblocking(False)
result = candidate.connect_ex(sockaddr)
if result == 0 or result == errno.EISCONN:
return candidate
if result not in in_progress:
raise OSError(result, os.strerror(result))
selector.register(candidate, selectors.EVENT_WRITE)
active.add(candidate)
return None
except Exception:
candidate.close()
raise
try:
while pending or active:
now = time.monotonic()
if deadline is not None and now >= deadline:
raise socket.timeout("timed out")
if pending and now >= next_launch:
addrinfo = pending.pop(0)
try:
winner = start_attempt(addrinfo)
except OSError as exc:
last_error = exc
if not active:
next_launch = now
continue
if winner is not None:
break
next_launch = now + _HAPPY_EYEBALLS_DELAY_SECONDS
wait_timeout = None if deadline is None else max(0.0, deadline - now)
if pending:
until_launch = max(0.0, next_launch - now)
wait_timeout = (
until_launch
if wait_timeout is None
else min(wait_timeout, until_launch)
)
events = selector.select(wait_timeout)
for key, _mask in events:
candidate = key.fileobj
error_code = candidate.getsockopt(socket.SOL_SOCKET, socket.SO_ERROR)
selector.unregister(candidate)
active.discard(candidate)
if error_code == 0:
winner = candidate
break
candidate.close()
last_error = OSError(error_code, os.strerror(error_code))
if winner is not None:
break
if not active and pending:
next_launch = time.monotonic()
if winner is None:
if last_error is not None:
raise last_error
raise OSError(f"Could not connect to {host}:{port}")
try:
selector.unregister(winner)
except Exception:
pass
active.discard(winner)
winner.settimeout(timeout)
for option in socket_options or ():
winner.setsockopt(*option)
winner.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)
return winner
finally:
for candidate in active:
try:
selector.unregister(candidate)
except Exception:
pass
candidate.close()
selector.close()
class _HappyEyeballsSyncBackend:
"""httpcore sync backend with concurrent IPv6/IPv4 connection fallback."""
def __init__(self):
self._fallback = None
def _default_backend(self):
if self._fallback is None:
from httpcore import SyncBackend
self._fallback = SyncBackend()
return self._fallback
def connect_tcp(
self,
host: str,
port: int,
timeout: Optional[float] = None,
local_address: Optional[str] = None,
socket_options=None,
):
from httpcore import ConnectError, ConnectTimeout
from httpcore._backends.sync import SyncStream
source_address = None if local_address is None else (local_address, 0)
try:
sock = _happy_eyeballs_create_connection(
(host, port),
timeout,
source_address=source_address,
socket_options=socket_options or (),
)
except socket.timeout as exc:
raise ConnectTimeout(str(exc)) from exc
except OSError as exc:
raise ConnectError(str(exc)) from exc
return SyncStream(sock)
def connect_unix_socket(self, *args, **kwargs):
return self._default_backend().connect_unix_socket(*args, **kwargs)
def sleep(self, seconds: float) -> None:
self._default_backend().sleep(seconds)
def _uses_codex_cloud_transport(base_url: str) -> bool:
return (
base_url_hostname(base_url).lower() == "chatgpt.com"
and "/backend-api/codex" in str(base_url).lower()
)
def _enable_happy_eyeballs(transport) -> None:
"""Install the sync racing backend on one httpx transport, if compatible.
Reaches into httpx/httpcore private attributes (``transport._pool`` /
``pool._network_backend``); safe because httpcore is pinned (1.0.x) and
both lookups are hasattr-guarded — on an incompatible httpcore this
degrades to the default serial backend instead of crashing.
"""
pool = getattr(transport, "_pool", None)
if pool is not None and hasattr(pool, "_network_backend"):
pool._network_backend = _HappyEyeballsSyncBackend()
def enable_happy_eyeballs_on_client(client) -> None:
"""Install the sync racing backend on every direct transport of a client.
Covers a ready-built ``httpx.Client`` (its default transport plus any
mounts), for callers that construct clients inline instead of going
through :func:`build_keepalive_http_client` — e.g. the Codex OAuth token
refresh / device-login / usage-probe clients in ``hermes_cli.auth``.
Proxy-backed transports (``httpcore.HTTPProxy`` / SOCKS pools) are left
untouched: with a proxy in play the TCP connect goes to the proxy host,
which is out of scope for the direct-transport racing added in #94388.
Async clients are also left untouched — httpcore's async backend already
performs RFC 8305 racing natively via
``anyio.connect_tcp(happy_eyeballs_delay=0.25)``.
Best-effort and hasattr-guarded like ``_enable_happy_eyeballs``; on an
incompatible httpx/httpcore this silently keeps the default backend.
"""
try:
import httpcore
proxy_pool_types = tuple(
t
for t in (
getattr(httpcore, "HTTPProxy", None),
getattr(httpcore, "SOCKSProxy", None),
)
if t is not None
)
except Exception:
return
transports = [getattr(client, "_transport", None)]
transports.extend((getattr(client, "_mounts", None) or {}).values())
for transport in transports:
pool = getattr(transport, "_pool", None)
if pool is None or not hasattr(pool, "_network_backend"):
continue
if proxy_pool_types and isinstance(pool, proxy_pool_types):
continue
pool._network_backend = _HappyEyeballsSyncBackend()
def _load_openai_cls() -> type:
"""Import and cache ``openai.OpenAI``."""
global _OPENAI_CLS_CACHE
if _OPENAI_CLS_CACHE is None:
from openai import OpenAI as _cls
_OPENAI_CLS_CACHE = _cls
return _OPENAI_CLS_CACHE
class _OpenAIProxy:
"""Module-level proxy that looks like ``openai.OpenAI`` but imports lazily."""
__slots__ = ()
def __call__(self, *args, **kwargs):
return _load_openai_cls()(*args, **kwargs)
def __instancecheck__(self, obj):
return isinstance(obj, _load_openai_cls())
def __repr__(self):
return "<lazy openai.OpenAI proxy>"
class _SafeWriter:
"""Transparent stdio wrapper that catches OSError/ValueError from broken pipes.
When hermes-agent runs as a systemd service, Docker container, or headless
daemon, the stdout/stderr pipe can become unavailable (idle timeout, buffer
exhaustion, socket reset). Any print() call then raises
``OSError: [Errno 5] Input/output error``, which can crash agent setup or
run_conversation() — especially via double-fault when an except handler
also tries to print.
Additionally, when subagents run in ThreadPoolExecutor threads, the shared
stdout handle can close between thread teardown and cleanup, raising
``ValueError: I/O operation on closed file`` instead of OSError.
This wrapper delegates all writes to the underlying stream and silently
catches both OSError and ValueError. It is transparent when the wrapped
stream is healthy.
"""
__slots__ = ("_inner",)
def __init__(self, inner):
object.__setattr__(self, "_inner", inner)
def write(self, data):
try:
return self._inner.write(data)
except (OSError, ValueError):
return len(data) if isinstance(data, str) else 0
def flush(self):
try:
self._inner.flush()
except (OSError, ValueError):
pass
def fileno(self):
return self._inner.fileno()
def isatty(self):
try:
return self._inner.isatty()
except (OSError, ValueError):
return False
def __getattr__(self, name):
return getattr(self._inner, name)
def _get_proxy_from_env() -> Optional[str]:
"""Read proxy URL from environment variables.
Checks HTTPS_PROXY, HTTP_PROXY, ALL_PROXY (and lowercase variants) in order.
Returns the first valid proxy URL found, or None if no proxy is configured.
"""
for key in ("HTTPS_PROXY", "HTTP_PROXY", "ALL_PROXY",
"https_proxy", "http_proxy", "all_proxy"):
value = os.environ.get(key, "").strip()
if value:
return normalize_proxy_url(value)
return None
def _get_proxy_for_base_url(base_url: Optional[str]) -> Optional[str]:
"""Return an env-configured proxy unless NO_PROXY excludes this base URL."""
proxy = _get_proxy_from_env()
if not proxy or not base_url:
return proxy
host = base_url_hostname(base_url)
if not host:
return proxy
try:
if urllib.request.proxy_bypass_environment(host):
return None
except Exception:
pass
return proxy
def _shared_transport_cls():
"""Lazily define the per-client transport view (httpx import is deferred)."""
global _SharedTransport
if _SharedTransport is not None:
return _SharedTransport
import httpx
class _SharedTransportImpl(httpx.BaseTransport):
"""Per-client view of a process-shared ``httpx.HTTPTransport``.
``httpx.Client.close()`` closes every mounted transport. Each OpenAI
client still owns its own ``httpx.Client`` (the #10933 contract:
closing one client must never poison the next), so the object we
mount must absorb that close while the underlying connection pool
keeps serving every other client. ``handle_request`` stamps the
owning view into ``request.extensions`` so socket-abort sweeps can
target only this client's in-flight connections on the shared pool.
"""
__slots__ = ("_inner", "_closed")
def __init__(self, inner: Any) -> None:
self._inner = inner
self._closed = False
# httpx-private ``_pool`` is what our socket walkers and the
# happy-eyeballs / ssl-verify tests introspect: expose the shared one.
@property
def _pool(self) -> Any:
return getattr(self._inner, "_pool", None)
def handle_request(self, request: Any) -> Any:
if self._closed:
raise RuntimeError("Cannot send a request, as the client has been closed.")
request.extensions[HERMES_TRANSPORT_OWNER_EXT] = id(self)
return self._inner.handle_request(request)
def close(self) -> None:
# Deliberately does NOT close ``_inner``: it is shared. Idle
# connections are reaped by ``keepalive_expiry``; the pool lives
# for the process (see ``close_shared_transports``).
self._closed = True
_SharedTransportImpl.__name__ = _SharedTransportImpl.__qualname__ = "_SharedTransport"
_SharedTransport = _SharedTransportImpl
return _SharedTransport
_SharedTransport: Any = None
def _shared_transport_key(base_url: str, verify: Any, proxy: Optional[str]) -> tuple:
"""Identity under which sync direct transports are pooled process-wide."""
if verify is True or verify is False:
verify_key: Any = verify
elif isinstance(verify, str):
verify_key = ("path", verify)
else:
# An ssl.SSLContext (or custom object): share only by object identity,
# which is what a caller passing the same context twice expects.
verify_key = ("id", id(verify))
return (verify_key, proxy, _uses_codex_cloud_transport(base_url))
def _get_shared_transport(key: tuple, build) -> Any:
with _SHARED_TRANSPORTS_LOCK:
transport = _SHARED_TRANSPORTS.get(key)
if transport is None:
transport = build()
if len(_SHARED_TRANSPORTS) < _SHARED_TRANSPORTS_MAX:
_SHARED_TRANSPORTS[key] = transport
return transport
def close_shared_transports() -> int:
"""Really close every process-shared transport (test teardown / atexit)."""
with _SHARED_TRANSPORTS_LOCK:
transports = list(_SHARED_TRANSPORTS.values())
_SHARED_TRANSPORTS.clear()
for transport in transports:
try:
transport.close()
except Exception:
pass
return len(transports)
def build_keepalive_http_client(
base_url: str = "",
*,
async_mode: bool = False,
verify: Any = True,
) -> Optional[Any]:
"""Build an httpx client for OpenAI SDK calls with env-only proxy policy.
Uses explicit ``HTTPS_PROXY`` / ``NO_PROXY`` env vars via
``_get_proxy_for_base_url``. Plain no-proxy mounts disable httpx's default
``trust_env`` proxy path, so macOS system proxy settings from
``urllib.request.getproxies()`` (which omit the ExceptionsList) are not
applied. Mirrors ``AIAgent._build_keepalive_http_client``.
Connection lifecycle is managed at the HTTP pool layer
(``keepalive_expiry=20.0`` reaps idle connections before reverse proxies'
typical 30-60 s timeouts) instead of the former custom
``socket_options`` transport, which broke streaming behind reverse
proxies (#54049, #12952) and stalled TLS handshakes by stripping
``TCP_NODELAY``.
``verify`` is forwarded to httpx so auxiliary-client calls (compression,
vision, web_extract, title generation, etc.) honor the same per-provider
``ssl_ca_cert`` / ``ssl_verify`` and ``HERMES_CA_BUNDLE`` settings the main
client uses. It is passed on the client AND on the plain no-proxy mounts
(a mounted transport owns the SSL context for its scheme).
Every call returns a NEW ``httpx.Client`` (per-client close semantics are
what #10933 pins), but sync clients with the same
(verify, proxy, happy-eyeballs) identity mount the SAME underlying
``HTTPTransport`` through a :class:`_SharedTransport` view, so N delegated
children share one connection pool + SSL context instead of N. Async
clients are never shared: an httpcore async pool is bound to the event
loop that first used it.
"""
try:
import httpx
proxy = _get_proxy_for_base_url(base_url)
limits = httpx.Limits(
max_keepalive_connections=20,
max_connections=100,
keepalive_expiry=20.0,
)
# Generous read=None for SSE streaming endpoints.
timeout = httpx.Timeout(connect=15.0, read=None, write=15.0, pool=10.0)
transport_cls = httpx.AsyncHTTPTransport if async_mode else httpx.HTTPTransport
client_cls = httpx.AsyncClient if async_mode else httpx.Client
mounts = {}
if proxy is None:
happy_eyeballs = not async_mode and _uses_codex_cloud_transport(base_url)
# One pool now serves every agent in the process, so its ceiling
# must cover a whole fan-out of concurrently streaming children,
# not one client. (Note: previously the mounts silently ran on
# httpx defaults — keepalive_expiry=5s — since Client-level
# ``limits`` only reach the default transport.)
direct_limits = limits if async_mode else httpx.Limits(
max_keepalive_connections=50,
max_connections=1000,
keepalive_expiry=20.0,
)
def _build_direct():
transport = transport_cls(verify=verify, limits=direct_limits)
# Async transports need no explicit racing: httpcore's anyio
# backend already implements RFC 8305 natively
# (``anyio.connect_tcp(happy_eyeballs_delay=0.25)``), covered
# by tests/agent/test_codex_happy_eyeballs.py.
if happy_eyeballs:
_enable_happy_eyeballs(transport)
return transport
if async_mode:
mounts = {"http://": _build_direct(), "https://": _build_direct()}
else:
key = _shared_transport_key(base_url, verify, proxy)
view_cls = _shared_transport_cls()
mounts = {
f"{scheme}://": view_cls(
_get_shared_transport((scheme, *key), _build_direct)
)
for scheme in ("http", "https")
}
# Without this httpx builds a third, never-used direct
# transport (and pool + SSL context) per client.
return client_cls(
limits=limits,
timeout=timeout,
transport=mounts["https://"],
mounts=mounts,
)
return client_cls(
limits=limits,
timeout=timeout,
proxy=proxy,
mounts=mounts or None,
verify=verify,
)
except Exception:
return None
def _install_safe_stdio() -> None:
"""Wrap stdout/stderr so best-effort console output cannot crash the agent."""
for stream_name in ("stdout", "stderr"):
stream = getattr(sys, stream_name, None)
if stream is not None and not isinstance(stream, _SafeWriter):
setattr(sys, stream_name, _SafeWriter(stream))
# Module-level proxy instance — drops in for ``openai.OpenAI``. Imported as
# ``from agent.process_bootstrap import OpenAI`` (or re-exported via
# ``run_agent`` for legacy tests).
OpenAI = _OpenAIProxy()
__all__ = [
"OpenAI",
"_OpenAIProxy",
"_load_openai_cls",
"_SafeWriter",
"_install_safe_stdio",
"_get_proxy_from_env",
"_get_proxy_for_base_url",
"build_keepalive_http_client",
"close_shared_transports",
"enable_happy_eyeballs_on_client",
]