fix: pin all MCP server versions for reproducible benchmarks

Pin filesystem (@2025.12.18), postgres (0.3.0), and playwright (0.0.68)
versions. Also pin notion @1.9.1 in base_agent.py for consistency with
mcpmark_agent.py. GitHub (v0.15.0) and notion were already pinned in #246.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: zjwu0522

Dockerfile

@@ -79,7 +79,7 @@ RUN python3 -m playwright install chromium && \
     npx -y playwright install chromium
 
 # Layer 9: Install PostgreSQL MCP server (Python, used via `pipx run postgres-mcp`)
-RUN pipx install postgres-mcp
+RUN pipx install postgres-mcp==0.3.0
 
 # Set working directory
 WORKDIR /app

src/agents/base_agent.py

@@ -176,7 +176,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
                 raise ValueError("Notion API key required")
             return MCPStdioServer(
                 command="npx",
-                args=["-y", "@notionhq/notion-mcp-server"],
+                args=["-y", "@notionhq/notion-mcp-server@1.9.1"],
                 env={
                     "OPENAPI_MCP_HEADERS": (
                         '{"Authorization": "Bearer ' + notion_key + '", '
@@ -193,7 +193,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
                 command="npx",
                 args=[
                     "-y",
-                    "@modelcontextprotocol/server-filesystem",
+                    "@modelcontextprotocol/server-filesystem@2025.12.18",
                     str(test_directory),
                 ],
             )
@@ -204,7 +204,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
             viewport_width = self.service_config.get("viewport_width", 1280)
             viewport_height = self.service_config.get("viewport_height", 720)
 
-            args = ["-y", "@playwright/mcp@latest"]
+            args = ["-y", "@playwright/mcp@0.0.68"]
             if headless:
                 args.append("--headless")
             args.extend(
@@ -234,7 +234,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
             )
             return MCPStdioServer(
                 command="pipx",
-                args=["run", "postgres-mcp", "--access-mode=unrestricted"],
+                args=["run", "postgres-mcp==0.3.0", "--access-mode=unrestricted"],
                 env={"DATABASE_URI": database_url},
             )
 

src/agents/mcpmark_agent.py

@@ -1131,7 +1131,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
                 command="npx",
                 args=[
                     "-y",
-                    "@modelcontextprotocol/server-filesystem",
+                    "@modelcontextprotocol/server-filesystem@2025.12.18",
                     str(test_directory),
                 ],
             )
@@ -1142,7 +1142,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
             viewport_width = self.service_config.get("viewport_width", 1280)
             viewport_height = self.service_config.get("viewport_height", 720)
 
-            args = ["-y", "@playwright/mcp@latest"]
+            args = ["-y", "@playwright/mcp@0.0.68"]
             if headless:
                 args.append("--headless")
             args.extend(
@@ -1176,7 +1176,7 @@ def _create_stdio_server(self) -> MCPStdioServer:
 
             return MCPStdioServer(
                 command="pipx",
-                args=["run", "postgres-mcp", "--access-mode=unrestricted"],
+                args=["run", "postgres-mcp==0.3.0", "--access-mode=unrestricted"],
                 env={"DATABASE_URI": database_url},
             )