feat(mcp): surface MCP server log notifications in agent.log (#57416)

Port from anomalyco/opencode#34529: MCP servers can emit
notifications/message logging notifications (RFC 5424 levels), but the
MCP SDK's default logging_callback silently discards them — server-side
warnings/errors during tool calls were invisible.

- tools/mcp_tool.py: pass a logging_callback to every ClientSession
  (stdio, SSE, streamable HTTP old+new API paths via the shared
  sampling_kwargs sites), mapping the 8 MCP log levels onto Python
  logging levels and tagging entries with [server/logger] origin.
- JSON-serialize non-string payloads, cap at 2000 chars so a chatty
  server can't flood agent.log, never raise from the handler.
- Gated on SDK support (_check_logging_callback_support) mirroring the
  existing message_handler gate for old SDK versions.
- tests/tools/test_mcp_server_log_notifications.py: 10 tests covering
  level mapping, origin tagging, JSON payloads, truncation, and the
  never-raise contract.
This commit is contained in:
Teknium 2026-07-05 02:06:39 -07:00 committed by GitHub
parent 4751af0a0b
commit edf8e0ba94
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 196 additions and 0 deletions

View file

@ -0,0 +1,126 @@
"""Tests for MCP server log notification handling (port of anomalyco/opencode#34529).
MCP servers can emit ``notifications/message`` logging notifications
(RFC 5424 syslog levels). The MCP SDK's default ``logging_callback``
silently discards them; Hermes now passes ``_make_logging_callback()``
to ``ClientSession`` so server-side diagnostics land in agent.log,
tagged with the server name.
"""
import logging
from types import SimpleNamespace
import pytest
from tools.mcp_tool import (
_MCP_LOG_LEVEL_MAP,
_MCP_LOGGING_CALLBACK_SUPPORTED,
MCPServerTask,
)
def _params(level="info", data="hello", logger_name=None):
return SimpleNamespace(level=level, data=data, logger=logger_name)
class TestLogLevelMap:
def test_all_mcp_levels_mapped(self):
# MCP spec (RFC 5424) defines these eight levels.
for lvl in ("debug", "info", "notice", "warning",
"error", "critical", "alert", "emergency"):
assert lvl in _MCP_LOG_LEVEL_MAP
def test_severity_ordering(self):
assert _MCP_LOG_LEVEL_MAP["debug"] == logging.DEBUG
assert _MCP_LOG_LEVEL_MAP["notice"] == logging.INFO
assert _MCP_LOG_LEVEL_MAP["warning"] == logging.WARNING
assert _MCP_LOG_LEVEL_MAP["emergency"] == logging.ERROR
class TestLoggingCallback:
@pytest.mark.asyncio
async def test_routes_to_hermes_logger_with_server_tag(self, caplog):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
with caplog.at_level(logging.INFO, logger="tools.mcp_tool"):
await callback(_params(level="info", data="server started"))
assert any(
"MCP server log [log_srv]: server started" in rec.getMessage()
for rec in caplog.records
)
@pytest.mark.asyncio
async def test_includes_sub_logger_name(self, caplog):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
with caplog.at_level(logging.WARNING, logger="tools.mcp_tool"):
await callback(_params(level="warning", data="rate limited",
logger_name="http"))
assert any(
"MCP server log [log_srv/http]: rate limited" in rec.getMessage()
and rec.levelno == logging.WARNING
for rec in caplog.records
)
@pytest.mark.asyncio
async def test_error_family_maps_to_error_level(self, caplog):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
with caplog.at_level(logging.ERROR, logger="tools.mcp_tool"):
for lvl in ("error", "critical", "alert", "emergency"):
await callback(_params(level=lvl, data=f"boom-{lvl}"))
errors = [r for r in caplog.records if r.levelno == logging.ERROR]
assert len(errors) == 4
@pytest.mark.asyncio
async def test_non_string_data_is_json_serialized(self, caplog):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
with caplog.at_level(logging.INFO, logger="tools.mcp_tool"):
await callback(_params(data={"event": "connect", "port": 8080}))
assert any(
'"event": "connect"' in rec.getMessage() for rec in caplog.records
)
@pytest.mark.asyncio
async def test_unknown_level_defaults_to_info(self, caplog):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
with caplog.at_level(logging.INFO, logger="tools.mcp_tool"):
await callback(_params(level="bogus", data="odd level"))
assert any(
rec.levelno == logging.INFO and "odd level" in rec.getMessage()
for rec in caplog.records
)
@pytest.mark.asyncio
async def test_oversized_payload_truncated(self, caplog):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
with caplog.at_level(logging.INFO, logger="tools.mcp_tool"):
await callback(_params(data="x" * 10_000))
msg = next(
rec.getMessage() for rec in caplog.records
if "MCP server log" in rec.getMessage()
)
assert "... [truncated]" in msg
assert len(msg) < 3000
@pytest.mark.asyncio
async def test_handler_never_raises(self):
server = MCPServerTask("log_srv")
callback = server._make_logging_callback()
# A params object missing every attribute must not blow up the
# SDK's notification dispatch loop.
await callback(object())
class TestSDKSupportGate:
def test_current_sdk_supports_logging_callback(self):
# The pinned MCP SDK in this repo supports logging_callback; if this
# starts failing after an SDK downgrade the feature silently degrades
# (by design), but we want to know.
import inspect
from mcp import ClientSession
expected = "logging_callback" in inspect.signature(ClientSession).parameters
assert _MCP_LOGGING_CALLBACK_SUPPORTED == expected

View file

@ -280,6 +280,38 @@ _MCP_MESSAGE_HANDLER_SUPPORTED = _check_message_handler_support()
if _MCP_AVAILABLE and not _MCP_MESSAGE_HANDLER_SUPPORTED:
logger.debug("MCP SDK does not support message_handler -- dynamic tool discovery disabled")
def _check_logging_callback_support() -> bool:
"""Check if ClientSession accepts the ``logging_callback`` kwarg.
Mirrors ``_check_message_handler_support`` for backward compatibility
with older MCP SDK versions. Without a logging_callback, the SDK's
default handler silently discards every ``notifications/message`` a
server emits, so server-side diagnostics never reach Hermes' logs.
"""
if not _MCP_AVAILABLE:
return False
try:
return "logging_callback" in inspect.signature(ClientSession).parameters
except (TypeError, ValueError):
return False
_MCP_LOGGING_CALLBACK_SUPPORTED = _check_logging_callback_support()
# MCP logging levels (RFC 5424 syslog severities) -> Python logging levels.
# Port of anomalyco/opencode#34529's serverLog mapping.
_MCP_LOG_LEVEL_MAP = {
"debug": logging.DEBUG,
"info": logging.INFO,
"notice": logging.INFO,
"warning": logging.WARNING,
"error": logging.ERROR,
"critical": logging.ERROR,
"alert": logging.ERROR,
"emergency": logging.ERROR,
}
# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------
@ -1539,6 +1571,40 @@ class MCPServerTask:
task.add_done_callback(self._pending_refresh_tasks.discard)
return task
def _make_logging_callback(self):
"""Build a ``logging_callback`` for ``ClientSession``.
Routes MCP ``notifications/message`` log notifications from the
server into Hermes' logging (agent.log via hermes_logging), tagged
with the server name. Without this, the SDK's default callback
silently discards them, so server-side warnings/errors during a
tool call were invisible. Port of anomalyco/opencode#34529.
"""
async def _on_log(params):
try:
level = _MCP_LOG_LEVEL_MAP.get(
str(getattr(params, "level", "info")).lower(), logging.INFO,
)
data = getattr(params, "data", None)
if not isinstance(data, str):
try:
data = json.dumps(data, ensure_ascii=False, default=str)
except (TypeError, ValueError):
data = str(data)
# Cap pathological payloads so a chatty/broken server can't
# flood agent.log with megabyte lines.
if len(data) > 2000:
data = data[:2000] + "... [truncated]"
logger_name = getattr(params, "logger", None)
origin = f"{self.name}/{logger_name}" if logger_name else self.name
logger.log(level, "MCP server log [%s]: %s", origin, data)
except Exception:
logger.debug(
"Failed to handle MCP log notification from '%s'",
self.name, exc_info=True,
)
return _on_log
def _make_message_handler(self):
"""Build a ``message_handler`` callback for ``ClientSession``.
@ -1864,6 +1930,8 @@ class MCPServerTask:
sampling_kwargs.update(self._elicitation.session_kwargs())
if _MCP_NOTIFICATION_TYPES and _MCP_MESSAGE_HANDLER_SUPPORTED:
sampling_kwargs["message_handler"] = self._make_message_handler()
if _MCP_LOGGING_CALLBACK_SUPPORTED:
sampling_kwargs["logging_callback"] = self._make_logging_callback()
# Snapshot child PIDs before spawning so we can track the new one.
pids_before = _snapshot_child_pids()
@ -2079,6 +2147,8 @@ class MCPServerTask:
sampling_kwargs.update(self._elicitation.session_kwargs())
if _MCP_NOTIFICATION_TYPES and _MCP_MESSAGE_HANDLER_SUPPORTED:
sampling_kwargs["message_handler"] = self._make_message_handler()
if _MCP_LOGGING_CALLBACK_SUPPORTED:
sampling_kwargs["logging_callback"] = self._make_logging_callback()
# SSE transport (for MCP servers that implement the SSE transport protocol
# rather than Streamable HTTP). Configure with ``transport: sse`` in the