diff --git a/src/acp/connection.py b/src/acp/connection.py
index 09e5a0e..41cdebc 100644
--- a/src/acp/connection.py
+++ b/src/acp/connection.py
@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 import asyncio
-import contextlib
 import copy
 import inspect
 import json
@@ -255,6 +254,11 @@ async def _run_request(self, message: dict[str, Any]) -> Any:
                 self._notify_observers(StreamDirection.OUTGOING, payload)
                 raise err from None
             except Exception as exc:
+                logging.exception(
+                    "Unhandled error while handling request method=%s",
+                    method,
+                    exc_info=exc,
+                )
                 try:
                     data = json.loads(str(exc))
                 except Exception:
@@ -267,8 +271,15 @@ async def _run_request(self, message: dict[str, Any]) -> Any:
 
     async def _run_notification(self, message: dict[str, Any]) -> None:
         method = message["method"]
-        with span_context("acp.notification", attributes={"method": method}), contextlib.suppress(Exception):
-            await self._handler(method, message.get("params"), True)
+        with span_context("acp.notification", attributes={"method": method}):
+            try:
+                await self._handler(method, message.get("params"), True)
+            except Exception as exc:
+                logging.exception(
+                    "Unhandled error while handling notification method=%s",
+                    method,
+                    exc_info=exc,
+                )
 
     async def _handle_response(self, message: dict[str, Any]) -> None:
         request_id = message["id"]
diff --git a/tests/test_request_error_logging.py b/tests/test_request_error_logging.py
new file mode 100644
index 0000000..3581804
--- /dev/null
+++ b/tests/test_request_error_logging.py
@@ -0,0 +1,108 @@
+"""Unhandled RPC handler exceptions must be logged instead of silently swallowed.
+
+Requests already returned a JSON-RPC -32603 error but discarded the original
+traceback; notifications suppressed the exception entirely. Both now log the
+underlying exception so integrators (e.g. Sentry via its logging integration)
+can see server-side handler crashes.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from acp.connection import Connection, MethodHandler
+from acp.exceptions import RequestError
+
+
+class _RecordingSender:
+    """Duck-typed MessageSender that records outgoing frames instead of writing them."""
+
+    def __init__(self, writer: asyncio.StreamWriter, supervisor: Any) -> None:
+        self.sent: list[dict[str, Any]] = []
+
+    async def send(self, payload: dict[str, Any]) -> None:
+        self.sent.append(payload)
+
+    async def close(self) -> None:
+        pass
+
+
+def _make_connection(handler: MethodHandler) -> tuple[Connection, _RecordingSender]:
+    captured: dict[str, _RecordingSender] = {}
+
+    def sender_factory(writer: asyncio.StreamWriter, supervisor: Any) -> _RecordingSender:
+        captured["sender"] = _RecordingSender(writer, supervisor)
+        return captured["sender"]
+
+    conn = Connection(handler, MagicMock(), MagicMock(), sender_factory=sender_factory, listening=False)
+    return conn, captured["sender"]
+
+
+async def _raising_handler(method: str, params: Any, is_notification: bool) -> Any:
+    raise RuntimeError("kaboom")
+
+
+def _assert_logged_runtime_error(caplog: pytest.LogCaptureFixture, method: str) -> None:
+    records = [
+        record
+        for record in caplog.records
+        if record.levelno == logging.ERROR and record.exc_info and f"method={method}" in record.getMessage()
+    ]
+    assert len(records) == 1, f"expected exactly one logged error for method={method}"
+    exc_info = records[0].exc_info
+    assert exc_info is not None
+    logged = exc_info[1]
+    assert isinstance(logged, RuntimeError)
+    assert str(logged) == "kaboom"
+
+
+@pytest.mark.asyncio
+async def test_run_request_unhandled_exception_is_logged_and_returned_as_internal_error(caplog):
+    conn, sender = _make_connection(_raising_handler)
+    request = {"jsonrpc": "2.0", "id": 7, "method": "explode", "params": None}
+
+    try:
+        with caplog.at_level(logging.ERROR), pytest.raises(RequestError) as exc_info:
+            await conn._run_request(request)
+    finally:
+        await conn.close()
+
+    # The handler exception is re-raised as a JSON-RPC internal error...
+    raised = exc_info.value
+    assert isinstance(raised, RequestError)
+    assert raised.code == -32603
+    assert raised.data == {"details": "kaboom"}
+
+    # ...and exactly one error frame carrying the handler's message is written to the peer.
+    assert len(sender.sent) == 1
+    response = sender.sent[0]
+    assert response["id"] == 7
+    assert "result" not in response
+    assert response["error"] == {"code": -32603, "message": "Internal error", "data": {"details": "kaboom"}}
+
+    # The original exception is logged, not discarded by `raise err from None`.
+    _assert_logged_runtime_error(caplog, "explode")
+
+
+@pytest.mark.asyncio
+async def test_run_notification_unhandled_exception_is_logged_and_not_answered(caplog):
+    conn, sender = _make_connection(_raising_handler)
+    notification = {"jsonrpc": "2.0", "method": "session/cancel", "params": {"sessionId": "s1"}}
+
+    try:
+        with caplog.at_level(logging.ERROR):
+            result = await conn._run_notification(notification)
+    finally:
+        await conn.close()
+
+    # A notification has no response: the error is neither raised nor written to the wire.
+    assert result is None
+    assert sender.sent == []
+
+    # It must still be logged — previously contextlib.suppress dropped it silently.
+    _assert_logged_runtime_error(caplog, "session/cancel")