MCP tool schemas dropped after batch error array from restarted server

Summary

After an MCP server crashes and auto-restarts mid-request, Claude Code drops all cached tool schemas for that server. The restarted server is healthy and accepts connections, but the client can no longer call any of its tools — ToolSearch returns nothing. Only /mcp reconnect recovers.

Repro

1. Configure a stdio MCP server that can crash on specific inputs (e.g. ida-mcp decompiling certain functions triggers a Hex-Rays C++ segfault post-response)
2. Call a batch tool that sends multiple requests (e.g. decompile([addr1, addr2, addr3]))
3. The server crashes during cleanup after the first response — the remaining calls return error arrays from the auto-restarted instance
4. After this, all tool calls to that server fail — not because the server is down (it restarted successfully) but because the client lost the tool schemas

Expected

The client should re-fetch tool schemas (tools/list) when it detects a server restart or reconnection, so the agent can act on recovery hints in the error response (e.g. "call open_idb again").

Actual

Tool schemas are gone. The server's improved error message ("the server may have restarted — call open_idb again") is useless because the agent can't invoke open_idb without the schema.

Environment

• Claude Code (latest as of 2026-04-14)
• MCP server: ida-mcp v9.3.12 (stdio transport)
• macOS Tahoe, Apple Silicon

Workaround

User manually runs /mcp to force reconnection.