MCP Chained Tool Routing Failure

MCP_Chained_Tool_Routing_Bug_Report.md

Note, ChatGPT recommended I report this error both here, and at https://github.com/openai/openai-apps-sdk-examples/issues/new, where it is logged as issue #207.

Please see summary of problem experienced when trying to test MCP server using Developer Mode in ChatGPT.

Basically, it seems if the pattern initialize, initialized, "call tool" is followed when making calls to the MCP server, all is good and there never is a failure.

However if two different tools are called in succession without any call to establish a new MCP session-ID, then we see failure on subsequent tool calls, and ChatGPT reports that it gets a "Resource not found" error.

Note that sometimes there is a pair of calls to initialize before the call to initialized. Call to the tools will use the same MCP session-ID for which initialized was called.

I asked ChatGPT to produce the attached MD-file.

The important thing to note is that when there is a failure, nothing is actually sent to the MCP server, as evidenced in ngrok's Traffic Inspector as well as there being no message actually received by the MCP server itself.

While I am using Codex on the coding side of things, I am actually testing the MCP server inside ChatGPT right now, where the issue has been observed and repeatedly reproduced. I have not yet added the MCP Server inside Codex, as yet.