neovim/runtime/doc/msgpack_rpc.txt

*msgpack_rpc.txt*	Nvim


		 NVIM REFERENCE MANUAL    by Thiago de Arruda


RPC API for Nvim				     *RPC* *rpc* *msgpack-rpc*

				      Type |gO| to see the table of contents.

==============================================================================
1. Introduction						            *rpc-intro*

The primary way to control Nvim programmatically is the RPC API, which speaks
MessagePack-RPC ("msgpack-rpc"), a messaging protocol that uses the
MessagePack serialization format:
  https://github.com/msgpack/msgpack/blob/0b8f5ac/spec.md

All kinds of Nvim "clients" use the RPC API: user interfaces (GUIs), remote
plugins, scripts like "nvr" (https://github.com/mhinz/neovim-remote), and even
`nvim` itself can control other `nvim` instances. By connecting to the RPC API
programs can:

  - Call any API function
  - Listen for events
  - Receive remote calls from Nvim

The RPC API is like a more powerful version of Vim's `clientserver` feature.

==============================================================================
2. API mapping							    *rpc-api*

The Nvim C |API| is automatically exposed to the RPC API by the build system,
which parses headers at src/nvim/api/*. A dispatch function is generated which
matches RPC API method names with public API functions, converting/validating
arguments and return values back to msgpack.

Client libraries (|api-client|s) normally provide wrappers that hide
msgpack-rpc details from application developers. The wrappers can be
automatically generated by reading bundled API metadata from a compiled Nvim
instance.

There are three ways to obtain API metadata:

  1. Connect to a running Nvim instance and call `nvim_get_api_info` via
     msgpack-rpc. This is best for clients written in dynamic languages which
     can define functions at runtime.

  2. Start Nvim with the |--api-info| option. Useful for clients written in
     statically-compiled languages.

  3. Use the |api_info()| vimscript function.

To get a human-readable list of API functions: >
    :new|put =map(api_info().functions, 'v:val.name')
<
To get a formatted dump of the API using python (requires the `pyyaml` and
`msgpack-python` packages): >
    nvim --api-info | python -c 'import msgpack, sys, yaml; print yaml.dump(msgpack.unpackb(sys.stdin.read()))'
<
==============================================================================
3. Connecting						      *rpc-connecting*

There are several ways to open a msgpack-rpc channel to an Nvim instance:

  1. Through stdin/stdout when `nvim` is started with `--embed`. This is how
     applications can embed Nvim.

  2. Through stdin/stdout of some other process spawned by |jobstart()|.
     Set the "rpc" key to |v:true| in the options dict to use the job's stdin
     and stdout as a single msgpack channel that is processed directly by
     Nvim.  Then it is not possible to process raw data to or from the
     process's stdin and stdout. stderr can still be used, though.

  3. Through the socket automatically created with each instance. The socket
     location is stored in |v:servername|.

  4. Through a TCP/IP socket. To make Nvim listen on a TCP/IP socket, set the
     |$NVIM_LISTEN_ADDRESS| environment variable before starting Nvim: >
        NVIM_LISTEN_ADDRESS=127.0.0.1:6666 nvim
<
Connecting to the socket is the easiest way a programmer can test the API,
which can be done through any msgpack-rpc client library or full-featured
|api-client|. Here's a Ruby script that prints 'hello world!' in the current
Nvim instance:
>
    #!/usr/bin/env ruby
    # Requires msgpack-rpc: gem install msgpack-rpc
    #
    # To run this script, execute it from a running Nvim instance (notice the
    # trailing '&' which is required since Nvim won't process events while
    # running a blocking command):
    #
    #	:!./hello.rb &
    #
    # Or from another shell by setting NVIM_LISTEN_ADDRESS:
    # $ NVIM_LISTEN_ADDRESS=[address] ./hello.rb

    require 'msgpack/rpc'
    require 'msgpack/rpc/transport/unix'

    nvim = MessagePack::RPC::Client.new(MessagePack::RPC::UNIXTransport.new, ENV['NVIM_LISTEN_ADDRESS'])
    result = nvim.call(:nvim_command, 'echo "hello world!"')
<
A better way is to use the Python REPL with the `neovim` package, where API
functions can be called interactively:
>
    >>> from neovim import attach
    >>> nvim = attach('socket', path='[address]')
    >>> nvim.command('echo "hello world!"')
<
You can also embed an Nvim instance via |jobstart()|, and communicate using
|rpcrequest()| and |rpcnotify()|:
>
    let nvim = jobstart(['nvim', '--embed'], {'rpc': v:true})
    echo rpcrequest(nvim, 'nvim_eval', '"Hello " . "world!"')
    call jobstop(nvim)
<
==============================================================================
4. Implementing API clients			*rpc-api-client* *api-client*

"API clients" wrap the Nvim API to provide idiomatic "SDKs" for their
respective platforms (see |dev-jargon|). You can build a new API client for
your favorite platform or programming language.

Existing API clients are listed here:
    https://github.com/neovim/neovim/wiki/Related-projects#api-clients

The Python client is the reference implementation for API clients. It is
always up-to-date with the Nvim API, so its source code and test suite are
authoritative references.
    https://github.com/neovim/python-client

API client implementation guidelines ~

  - Separate the transport layer from the rest of the library. See
    |rpc-connecting| for details on how clients can connect to Nvim.
  - Use a MessagePack library that implements at least version 5 of the
    MessagePack spec, which supports the `bin` and `ext` types used by Nvim.
  - Read API metadata in order to create client-side wrappers for all
    msgpack-rpc methods.
  - Use a single-threaded event loop library/pattern.
  - Use a fiber/coroutine library for the language being used for implementing
    a client. These greatly simplify concurrency and allow the library to
    expose a blocking API on top of a non-blocking event loop without the
    complexity that comes with preemptive multitasking.
  - Don't assume anything about the order that responses to msgpack-rpc
    requests will arrive.
  - Clients should expect msgpack-rpc requests, which need to be handled
    immediately because Nvim is blocked while waiting for the client response.
  - Clients should expect to receive msgpack-rpc notifications, but these
    don't need to be handled immediately because they won't block Nvim
    (although they should probably be handled immediately anyway).

Note: Most of the complexity could be handled by a msgpack-rpc library that
supports server to client requests and notifications, but it's not clear if
this is part of the msgpack-rpc spec. At least the Ruby msgpack-rpc library
does not seem to support it:
https://github.com/msgpack-rpc/msgpack-rpc-ruby/blob/master/lib/msgpack/rpc/transport/tcp.rb#L150-L158

API metadata object ~

API clients exist to hide msgpack-rpc details. The API metadata object
contains information that makes this task easier (see also |rpc-types|):

  - The "version" key contains the Nvim version, API level, and API
    backwards-compatibility level.
  - The "functions" key contains a list of metadata objects for individual
    functions.
  - Each function metadata object has |rpc-types| information about the return
    value and parameters. These can be used for generating strongly-typed APIs
    in static languages.
  - Container types may be decorated with type/size constraints, e.g.
    ArrayOf(Buffer) or ArrayOf(Integer, 2). This can be useful to generate
    even more strongly-typed APIs.
  - Functions that are considered to be methods that operate on instances of
    Nvim special types (msgpack EXT) will have the `"method"` attribute set to
    `true`. The receiver type is the type of the first argument. The method
    names are prefixed with `nvim_` plus a shortened type name, e.g.
    `nvim_buf_get_lines` represents the `get_lines` method of a Buffer instance.
    - Global functions have `"method"` set to `false` and are prefixed with just
    `nvim_`, e.g. `nvim_get_buffers`.

So for an object-oriented language, an API client contains the classes
representing Nvim special types, and the methods of each class could be
defined by stripping the prefix for the type as defined in the `types` metadata
(this will always be the first two "_"-separated parts of the function name).
There could also be a singleton Vim class with methods where the `nvim_`
prefix is stripped off.

==============================================================================
5. Types							    *rpc-types*

The Nvim C API uses custom types for all functions. |api-types|
At the RPC layer, the types can be split into two groups:

  - Basic types that map natively to msgpack (and probably have a default
    representation in msgpack-supported programming languages)
  - Special Nvim types that map to msgpack EXT with custom type codes.

Basic types ~

  Nil                               -> msgpack nil
  Boolean                           -> msgpack boolean
  Integer (signed 64-bit integer)   -> msgpack integer
  Float (IEEE 754 double precision) -> msgpack float
  String                            -> msgpack string
  Array                             -> msgpack array
  Dictionary                        -> msgpack map

Special types (msgpack EXT) ~

  Buffer                            -> enum value kObjectTypeBuffer
  Window                            -> enum value kObjectTypeWindow
  Tabpage                           -> enum value kObjectTypeTabpage

API functions expecting one of the special EXT types may be passed an integer
instead, but not another EXT type. E.g. Buffer may be passed as an integer but
not as a Window or Tabpage. The EXT object data is the object id encoded as
a msgpack integer: For buffers this is the |bufnr()| and for windows the
|window-ID|. For tabpages the id is an internal handle, not the tabpage
number.

To determine the type codes of the special EXT types, inspect the `types` key
of the |api-metadata| at runtime. Example JSON representation: >

  "types": {
    "Buffer": {
      "id": 0,
      "prefix": "nvim_buf_"
    },
    "Window": {
      "id": 1,
      "prefix": "nvim_win_"
    },
    "Tabpage": {
      "id": 2,
      "prefix": "nvim_tabpage_"
    }
  }

Even for statically compiled clients it is good practice to avoid hardcoding
the type codes, because a client may be built against one Nvim version but
connect to another with different type codes.