Improve tool/resource type handling

[domarmstrong]

Goal

While looking at MCP apps, the apps need to be able to consume and call tools. Currently the tools typing (inputs/outputs) is not available to access, so you'd need to redefine all the types, or restructure the code and manually cast the inputs/outputs. This PR demonstrates a potential approach that allows accessing the tools typing by inference without requiring manual definitions.

Design

Create a new TypesafeTool class that handles the type inference (it could likely also be extended with code that currently lives in server which would make more sense being handled by the tools).

const myTool = new TypesafeTool(
   {
     title: "My Tool",
     summary: "Does something useful",
     inputSchema: z.object({ id: z.string() }),
   },
   (client: MyClient) => async (args) => {
     const result = await client.doSomething(args.id); // args has the correct type without any explicit handling
     return {
       content: [{ type: "text", text: JSON.stringify(result) }],
       structuredContent: result,
     };
   }
 );

 // Access inferred types:
 type Name = typeof myTool.name;     // "my_tool" // inferred from the title
 type Input = typeof myTool.input;   // { id: string } // inferred from the schema
 type Output = typeof myTool.output; // output can be inferred from structuredContent

This could later allow defining some hooks that could be used with mcp apps such that they are typesafe.

const { result } = useToolResult<typeof listProjects>();

await callServerTool<typeof listProjects>({
  name: 'bugsnag_list_projects',
  arguments: { apiKey: '...' },
);

Following the pattern from zephyr and reflect, use a file per tool structure which is much nicer and also gives access to the tools.

Changeset

No functional change, just refactoring

Testing

Unit tests

[sazap10]

This change seems a good improvement, makes sense to wait for the bugsnag tool refactoring to be merged then applying this change across the board