SEP-986: Specify Format for Tool Names

[mikekistler]

Is your feature request related to a problem? Please describe.

Implement support for SEP-986, which has been accepted for the next version of the MCP spec.

Describe the solution you'd like

Compliance with the next version of the MCP spec within two weeks of its release.

[stephentoub]

There are a lot of "SHOULD"s in that change.

@mikekistler, do we want to enforce the rules, i.e. have both client and server throw if an invalid tool name is specified? Normalize in both client and server? Or just have an analyzer that tries to flag invalid tool names specified server-side? Or something else?

[PederHP]

Also note that some of the names that are now allowed will cause errors with some APIs. Forward slash and dot especially, are an interesting choice to allow - as OpenAI-style APIs generally don't allow these. This being allowed in MCP but not in all APIs means that it becomes client-specific how they're mapped to a valid name, which is annoying, as it makes evals harder to rely on as a server author.

We need to consider if the SDK should have convenience to do this mapping. I think configuring a default mapping for these characters is a reasonable convenience to add, as having to inspect every McpClientTool and do WithName if they're invalid is annoying.

[mikekistler]

I think that anywhere we generate tool names we need to comply with the rules. In particular, we should restrict inferred names to 64 characters, which I think we currently do not do:

event: message
data: {"result":{"tools":[{"name":"an_mcp_tool_with_a_very_looooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooong_name","description":"An MCP tool with a very long name","inputSchema":{"type":"object","properties":{}}}]},"id":1,"jsonrpc":"2.0"}

Adding an analyzer would be nice to check explicitly defined names.

[stephentoub]

Ok, so must-have is just adding a length limit for when we infer a name from a method name, and raising diagnostics about other statically-detectable cases is nice-to-have?