Remote MCP connection timeouts lack retry/fallback; calls hang #36
Description
Description
When using networked MCP remote services (example: Notion MCP), a network hiccup that causes a connection timeout makes the tool call fail and hang. The agent has no retry or fallback and the session must be manually interrupted.
Steps to Reproduce
- Enable a remote MCP (example: Notion MCP) and start mini-agent.
- With unstable network, invoke any tool (e.g., notion-create-database).
Observe
TypeError: fetch failed [cause]: ConnectTimeoutError ...;
call/load fails and requires manual interruption.
Expected Result
- On timeout, automatically retry a few times or fail fast with a clear message; do not block the session.
- Allow configurable timeout and retry counts; optionally skip that MCP server after failure.
Suggested Improvement
- Add retry/timeout handling in mcp_loader.py connect logic; mark the server unavailable on repeated failure.
- Wrap MCP tool execute with retry or timeout to avoid hangs.
- Expose timeout/retry parameters via config.
Environment
- mini-agent: current main (Dec 2025)
- OS: macOS 14 (Apple Silicon)
- Python: 3.11 (Anaconda)
- Node/npm: system npx