Can you explain what's going on with ClientContext here? Why do you declare the Client without context (DropContextService) and then wrap it with context? What is the reason that Client was not declared directly with the context?

In theory, we can wrap any context with any context provider (for example, the one that reads\writes context from\to file, and not that just wraps the context and keeps it in-memory). Practically, It could be solved with generic types and some methods, not the whole wrapper, but since that was the part of auto-generated code, I decided to keep it this way
Also we can replace the context on specific calls, not changing the wrapped one, but I fail to see the situation where that's actually useful (and the wrapper part too)

I understand that generics are used for better flexibility. But here a specific type with all substitutions is declared. Why are additional layers needed here? Why can't it be written like this:

pub type HttpClient = ContextWrapper<
    Client<
        hyper::Client<HttpConnector, Body>,
        ClientContext,
        Basic,
    >,
    ClientContext,
>;

Unnecessary empty line

Perhaps, it is better to call it BobClusterClient, because inside it contains clients for all nodes

I see no difference between this trait and the one below. Looks like the separation based on cfg is not needed on trait declaration level

The separation of implementations occurs at different levels. Here the struct level is used, and above - the function level. It is better to unify places, where cfg is applied. It seems that it is better to make a separation at the function level here

The only difference seems to be the call to check_api, I suggest to cfg only the call and the function itself

Implementations are also look identical

Looks like async_trait can be removed here

Looks like async_trait is not needed here

What about https, can it be added? bob technically supports it

API ports can be different on each node, and they are not known to other nodes. There should be a way to override default ports. In bob-tools it's done with argument api-port, Override default api port for the node. E.g. node1:80,node2:8000. Wildcard char (*) can be used to set port for all nodes.

Comment out of sync?

What is the reason to have single main node? Why not any other node in cluster?