this is a very good basic conception of making managed rpc calls.
for an application as an API to use, it may be implemented on top
of existing service APIs. some comments:

 which language do you aim at? it looks very much like C. (sorry
 if i missed that). can you give examples in other languages?

 why is GRPC_NO_ERROR not called GRPC_SUCCESS? 

 p. 3, 'General Workflow' typo in the first word: Defininf

 p.5 on function_handle_init, the advice to implementors. if i
 understand correctly, the problem here is that this should be a
 string that is accepted maybe also by the service discovery that the
 given service may use. it would make this interface much more
 practical if such a string could be standardised on. is there any
 effort in ggf to which this specification could point to?

 probe - wouldn't status be a better word?

 the spec does not say anything about how long a given session_id
 needs to be retained by the implementation. how long do errors
 need to be kept? if the error call was issued once already, can
 the buffer be purged?

 how are service specific errors propagated out (other than through
 the string)?