From e275f291f58b157f261b978057ebedc29006f279 Mon Sep 17 00:00:00 2001
From: lyla <lyla@dotorim.uk>
Date: Fri, 17 Apr 2026 08:13:24 +0900
Subject: [PATCH] =?UTF-8?q?fix(mcp):=20=ED=8C=8C=EC=9D=B4=ED=94=84?=
 =?UTF-8?q?=EB=9D=BC=EC=9D=B8=20=ED=98=B8=EC=B6=9C=EC=9D=84=20=EC=A0=84?=
 =?UTF-8?q?=EC=9A=A9=20=EB=8B=A8=EC=9D=BC=20=EC=9B=8C=EC=BB=A4=20=EC=8A=A4?=
 =?UTF-8?q?=EB=A0=88=EB=93=9C=EC=97=90=20=EA=B3=A0=EC=A0=95?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`asyncio.to_thread`의 기본 executor는 여러 워커에 작업을 분배하는데,
trawl 파이프라인이 쓰는 sync_playwright의 greenlet 디스패처는 최초
호출 스레드에 핀되므로 다른 워커로 들어온 호출이 "Cannot switch to
a different thread" greenlet.error를 뱉었다.

max_workers=1인 전용 ThreadPoolExecutor로 모든 fetch_page /
profile_page 호출을 동일 스레드에 고정해 이 간헐적 실패를 제거한다.
---
 src/trawl_mcp/server.py | 45 +++++++++++++++++++++++++++++------------
 1 file changed, 32 insertions(+), 13 deletions(-)

diff --git a/src/trawl_mcp/server.py b/src/trawl_mcp/server.py
index ed1fae0..5f09f92 100644
--- a/src/trawl_mcp/server.py
+++ b/src/trawl_mcp/server.py
@@ -1,16 +1,23 @@
 """Stdio/HTTP MCP server exposing trawl's fetch_page and profile_page tools.
 
 The pipeline uses sync_playwright internally, which can't run inside an
-asyncio event loop on its own. We run each pipeline invocation in a worker
-thread via `asyncio.to_thread`, keeping the MCP server responsive.
+asyncio event loop on its own. We run every pipeline invocation on a
+single dedicated worker thread so the process-wide sync_playwright
+greenlet dispatcher — which is pinned to the thread that first called
+sync_playwright() — always sees the same thread. Using
+`asyncio.to_thread` (default executor) instead causes intermittent
+"Cannot switch to a different thread" greenlet errors whenever a call
+is dispatched to a different worker thread.
 """
 
 from __future__ import annotations
 
 import asyncio
+import functools
 import json
 import logging
 import os
+from concurrent.futures import ThreadPoolExecutor
 
 from mcp.server import Server
 from mcp.server.stdio import stdio_server
@@ -18,6 +25,10 @@
 
 from trawl import fetch_relevant, to_dict
 
+_pipeline_executor = ThreadPoolExecutor(
+    max_workers=1, thread_name_prefix="trawl-pipeline"
+)
+
 logger = logging.getLogger("trawl_mcp")
 
 server: Server = Server("trawl")
@@ -166,13 +177,17 @@ async def _call_fetch_page(arguments: dict) -> list[TextContent]:
         use_hyde,
         use_rerank,
     )
-    result = await asyncio.to_thread(
-        fetch_relevant,
-        url,
-        query,
-        k=k,
-        use_hyde=use_hyde,
-        use_rerank=use_rerank,
+    loop = asyncio.get_running_loop()
+    result = await loop.run_in_executor(
+        _pipeline_executor,
+        functools.partial(
+            fetch_relevant,
+            url,
+            query,
+            k=k,
+            use_hyde=use_hyde,
+            use_rerank=use_rerank,
+        ),
     )
     payload = to_dict(result)
     payload["ok"] = not bool(payload.get("error"))
@@ -195,10 +210,14 @@ async def _call_profile_page(arguments: dict) -> list[TextContent]:
     # when the tool is actually called.
     from trawl.profiles import generate_profile
 
-    payload = await asyncio.to_thread(
-        generate_profile,
-        url,
-        force_refresh=force_refresh,
+    loop = asyncio.get_running_loop()
+    payload = await loop.run_in_executor(
+        _pipeline_executor,
+        functools.partial(
+            generate_profile,
+            url,
+            force_refresh=force_refresh,
+        ),
     )
     return [TextContent(type="text", text=json.dumps(payload, ensure_ascii=False))]