fix(mcp): stability fix pack — reload timeout, shutdown cleanup, event loop handler, OAuth non-blocking (#4757)

Four fixes for MCP server stability issues reported by community member
(terminal lockup, zombie processes, escape sequence pollution, startup hang):

1. MCP reload timeout guard (cli.py): _check_config_mcp_changes now runs
   _reload_mcp in a separate daemon thread with a 30s hard timeout. Previously,
   a hung MCP server could block the process_loop thread indefinitely, freezing
   the entire TUI (user can type but nothing happens, only Ctrl+D/Ctrl+\ work).

2. MCP stdio subprocess PID tracking (mcp_tool.py): Tracks child PIDs spawned
   by stdio_client via before/after snapshots of /proc children. On shutdown,
   _stop_mcp_loop force-kills any tracked PIDs that survived the SDK's graceful
   SIGTERM→SIGKILL cleanup. Prevents zombie MCP server processes from
   accumulating across sessions.

3. MCP event loop exception handler (mcp_tool.py): Installs
   _mcp_loop_exception_handler on the MCP background event loop — same pattern
   as the existing _suppress_closed_loop_errors on prompt_toolkit's loop.
   Suppresses benign 'Event loop is closed' RuntimeError from httpx transport
   __del__ during MCP shutdown. Salvaged from PR #2538 (acsezen).

4. MCP OAuth non-blocking (mcp_oauth.py): Replaces blocking input() call in
   _wait_for_callback with OAuthNonInteractiveError raise. Adds _is_interactive()
   TTY detection. In non-interactive environments, build_oauth_auth() still
   returns a provider (cached tokens + refresh work), but the callback handler
   raises immediately instead of blocking the MCP event loop for 120s. Re-raises
   OAuth setup failures in _run_http so failed servers are reported cleanly
   without blocking others. Salvaged from PRs #4521 (voidborne-d) and #4465
   (heathley).

Closes #2537, closes #4462
Related: #4128, #3436
This commit is contained in:
Teknium
2026-04-03 02:29:20 -07:00
committed by GitHub
parent f374ae4c61
commit cc54818d26
5 changed files with 431 additions and 12 deletions
+83 -6
View File
@@ -5,6 +5,12 @@ Wraps the MCP SDK's built-in ``OAuthClientProvider`` (which implements
authorization. The SDK handles all of the heavy lifting: PKCE generation,
metadata discovery, dynamic client registration, token exchange, and refresh.
Startup safety:
The callback handler never calls blocking ``input()`` on the event loop.
In non-interactive environments (no TTY, SSH, headless), the OAuth flow
raises ``OAuthNonInteractiveError`` instead of blocking, so that the
server degrades gracefully and other MCP servers are not affected.
Usage in mcp_tool.py::
from tools.mcp_oauth import build_oauth_auth
@@ -19,6 +25,7 @@ import json
import logging
import os
import socket
import sys
import threading
import webbrowser
from http.server import BaseHTTPRequestHandler, HTTPServer
@@ -28,6 +35,11 @@ from urllib.parse import parse_qs, urlparse
logger = logging.getLogger(__name__)
class OAuthNonInteractiveError(RuntimeError):
"""Raised when OAuth requires user interaction but the environment is non-interactive."""
pass
_TOKEN_DIR_NAME = "mcp-tokens"
@@ -164,7 +176,13 @@ async def _redirect_to_browser(auth_url: str) -> None:
async def _wait_for_callback() -> tuple[str, str | None]:
"""Start a local HTTP server on the pre-registered port and wait for the OAuth redirect."""
"""Start a local HTTP server on the pre-registered port and wait for the OAuth redirect.
If the callback times out, raises ``OAuthNonInteractiveError`` instead of
calling blocking ``input()`` — the old ``input()`` call would block the
entire MCP asyncio event loop, preventing all other MCP servers from
connecting and potentially hanging Hermes startup indefinitely.
"""
global _oauth_port
port = _oauth_port or _find_free_port()
HandlerClass, result = _make_callback_handler()
@@ -186,8 +204,10 @@ async def _wait_for_callback() -> tuple[str, str | None]:
code = result["auth_code"] or ""
state = result["state"]
if not code:
print(" Browser callback timed out. Paste the authorization code manually:")
code = input(" Code: ").strip()
raise OAuthNonInteractiveError(
"OAuth browser callback timed out after 120 seconds. "
"Run 'hermes mcp auth <server-name>' to authorize interactively."
)
return code, state
@@ -199,6 +219,17 @@ def _can_open_browser() -> bool:
return True
def _is_interactive() -> bool:
"""Check if the current environment can support interactive OAuth flows.
Returns False in headless/daemon/container environments where no user
can interact with a browser or paste an auth code.
"""
if not hasattr(sys.stdin, "isatty") or not sys.stdin.isatty():
return False
return True
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
@@ -209,6 +240,11 @@ def build_oauth_auth(server_name: str, server_url: str):
Uses the MCP SDK's ``OAuthClientProvider`` which handles discovery,
registration, PKCE, token exchange, and refresh automatically.
In non-interactive environments (no TTY), this still returns a provider
so that **cached tokens and refresh flows work**. Only the interactive
authorization-code grant will fail fast with a clear error instead of
blocking the event loop.
Returns an ``OAuthClientProvider`` instance (implements ``httpx.Auth``),
or ``None`` if the MCP SDK auth module is not available.
"""
@@ -219,6 +255,25 @@ def build_oauth_auth(server_name: str, server_url: str):
logger.warning("MCP SDK auth module not available — OAuth disabled")
return None
storage = HermesTokenStorage(server_name)
interactive = _is_interactive()
if not interactive:
# Check whether cached tokens exist. If they do, the SDK can still
# use them (and refresh them) without any user interaction. If not,
# we still build the provider — the callback_handler will raise
# OAuthNonInteractiveError if a fresh authorization is actually
# needed, which surfaces as a clean connection failure for this
# server only (other MCP servers are unaffected).
has_cached = storage._read_json(storage._tokens_path()) is not None
if not has_cached:
logger.warning(
"MCP server '%s' requires OAuth but no cached tokens found "
"and environment is non-interactive. The server will fail to "
"connect. Run 'hermes mcp auth %s' to authorize interactively.",
server_name, server_name,
)
global _oauth_port
_oauth_port = _find_free_port()
redirect_uri = f"http://127.0.0.1:{_oauth_port}/callback"
@@ -232,14 +287,36 @@ def build_oauth_auth(server_name: str, server_url: str):
token_endpoint_auth_method="none",
)
storage = HermesTokenStorage(server_name)
# In non-interactive mode, the redirect handler logs the URL and the
# callback handler raises immediately — no blocking, no input().
redirect_handler = _redirect_to_browser
callback_handler = _wait_for_callback
if not interactive:
async def _noninteractive_redirect(auth_url: str) -> None:
logger.warning(
"MCP server '%s' needs OAuth authorization (non-interactive, "
"cannot open browser). URL: %s",
server_name, auth_url,
)
async def _noninteractive_callback() -> tuple[str, str | None]:
raise OAuthNonInteractiveError(
f"MCP server '{server_name}' requires interactive OAuth "
f"authorization but the environment is non-interactive "
f"(no TTY). Run 'hermes mcp auth {server_name}' to "
f"authorize, then restart."
)
redirect_handler = _noninteractive_redirect
callback_handler = _noninteractive_callback
return OAuthClientProvider(
server_url=server_url,
client_metadata=client_metadata,
storage=storage,
redirect_handler=_redirect_to_browser,
callback_handler=_wait_for_callback,
redirect_handler=redirect_handler,
callback_handler=callback_handler,
timeout=120.0,
)