libreoffice-online/loolwsd/protocol.txt

All communication consists of messages that are one line of
human-readable UTF-8 text (with no terminating newline), optionally
followed by a single newline and arbitrary (potentialy even binary)
data.

The WebSocket distinction between 'text' and 'binary' frames has no
meaning for us for messages that don't contain additional binary data;
such messages can be either 'binary' or 'text' from the WebSocket
point of view even if we require them (the single line) to be
UTF-8. In other words, an implementation is free to send such a
single-line message as a WebSocket 'binary' frame, and the receiving
implementation must treat that equally as if it was a 'text' frame.

The WebSocket protocol says that 'text' frames are to be "interpreted"
as UTF-8, so it is probably best to indeed use 'binary' frames for
messages that contain optional non-UTF-8 data.

The protocol is not a request-response one. Messages may be sent in
either direction at any time, either in response to some message, or
spontaneously. For 'tile' messages, the client may send a bunch of
tile requests without waiting for return messages. The server may send
tiles proactively (guessing what the client might need). Etc.

client -> server
================

canceltiles

    All outstanding tile messages from the client to the server are
    dropped and will not be handled. There is no guarantee of exactly
    which tile: messages might still be sent back to the client.

downloadas name=<fileName> id=<id> format=<document format> options=<SkipImages, etc>

    Exports the current document to the desired format and returns a download URL
    The id identifies the request on the client.

getchildid

    Requests the child id so that it knows where the files needs to be sent when it is
    inserted in the document

gettextselection mimetype=<mimeType>

    Request selection's content

paste mimetype=<mimeType>
<binaryPasteData>

    Paste content at the current cursor position.

insertfile name=<name> type=<type>

    Inserts the file with the name <name> into the document, we currently support type = 'graphic'

key type=<type> char=<charcode> key=<keycode>

    <type> is 'input' or 'up', <charcode> and <keycode> are numbers.

load <pathname>

    Deprecated.

load [part=<partNumber>] url=<url> [timestamp=<time>] [options=<options>]

    part is an optional parameter. <partNumber> is a number.

    timestamp is an optional parameter.  <time> is provided in microseconds
    since the Unix epoch - midnight, January 1, 1970.

    options are the whole rest of the line, not URL-encoded

loolclient <major.minor[-patch]>

    Upon connection, a client must announce the version number it supports.
    Major: an integer that must always match between client and server,
           otherwise there are no guarantees of any sensible
           compatibility. This is bumped when API changes.
    Minor: an integer is more flexible and is at the discretion of either party.
           Security fixes that do not alter the API would bump the minor version number.
    Patch: an optional string that is informational.

mouse type=<type> x=<x> y=<y> count=<count>

    <type> is 'buttondown', 'buttonup' or 'move', others are numbers.

renderfont font=<font>

    requests the rendering of the given font.
    The font parameter is URL encoded

requestloksession

    requests the initialization of a LOK process in an attempt to predict the user's
    interaction with the document

resetselection

saveas url=<url> format=<format> options=<options>

    <url> is a URL, encoded. <format> is also URL-encoded, i.e. spaces as %20 and it can be empty
    options are the whole rest of the line, not URL-encoded, and can be empty

selecttext type=<type> x=<x> y=<y>

    <type> is 'start', 'end' or 'reset', <x> and <y> are numbers.

selectgraphic type=<type> x=<x> y=<y>

    <type> is 'start' or 'end' <x> and <y> are numbers.

status

styles

tile part=<partNumber> width=<width> height=<height> tileposx=<xpos> tileposy=<ypos> tilewidth=<tileWidth> tileheight=<tileHeight> [timestamp=<time>]

    All parameters are numbers.

tilecombine <parameters>

    Accept same parameters as 'tile' message except parameters 'tileposx' and 'tileposy'
    can be a comma separated list, and number of elements in both must be same.

uno <command>

    <command> is a line of text.

partpagerectangles

    Invokes lok::Document::getPartPageRectangles().

clientvisiblearea x=<x> y=<y> width=<width> height=<height>

    Invokes lok::Document::setClientVisibleArea().

takeedit

    Request for an edit lock. If successful, client will receive an 'editlock: 1'
    message meaning editlock is granted.

    See 'editlock:' message in server -> client.

useractive

    Sent when the user regains focus or clicks within the active area to
    disable the inactive state.
    Will send invalidation and update notifications to force refreshing the screen.

    See 'userinactive'.

userinactive

    Sent when the user has switched tabs or away from the Browser altogether.
    It should throttle updates until the user is active again.

    See 'useractive'.

pong

    Sent instead of a PONG frame as reply to a PING frame. A comment
    in our code says "Technically, we should send back a PONG control
    frame. However Firefox (probably) or Node.js (possibly) doesn't
    like that and closes the socket when we do."

server -> client
================

loolserver <major.minor[-patch]>

    Upon connection, the server must announce the version number it supports.
    Major: an integer that must always match between client and server,
           otherwise there are no guarantees of any sensible
           compatibility. This is bumped when API changes.
    Minor: an integer is more flexible and is at the discretion of either party.
           Security fixes that do not alter the API would bump the minor version number.
    Patch: an optional string that is informational.

contextmenu: <json description of the context menu>

    When the user right-clicks in the document, the content of the context
    menu is sent back via this callback.

    The structure of the context menu is a JSON, and looks like:

        {
            "menu": [
                { "text": "label text1", "type": "command", "command": ".uno:Something1", "enabled": "true" },
                { "text": "label text2", "type": "command", "command": ".uno:Something2", "enabled": "false" },
                { "type": "separator" },
                { "text": "label text2", "type": "menu", "menu": [ { ... }, { ... }, ... ] },
                ...
            ]
        }

downloadas: jail=<jail directory> dir=<a tmp dir> name=<name> port=<port>

    The client should then request http://server:port/jail/dir/name in order to download
    the document

editlock: <1 or 0>

    Informs the client of any change in ownership of edit lock. A value of '1'
    means client can edit the document, and '0' means that client can only view
    the document. This message always follows the 'status:' message after a
    document is loaded, so that client has this information as soon as it loads
    the document.

    Note that only one client can have the editlock at a time and
    others can only view.

error: cmd=<command> kind=<kind> [code=<error_code>]
<freeErrorText>

    <command> is the command part of the corresponding client->server
    message that caused the error. <kind> is some single-word
    classification

    <code> (when provided) further specifies the error as forwarded from
    LibreOffice

getchildid: id=<id>

    Returns the child id

invalidate: part=<partNumber> x=<x> y=<y> width=<width> height=<height>

    All parameters are numbers. Tells the client to invalidate any
    cached tiles for the document area specified (in twips), at any
    zoom level.

    The client should handle either this message or the
    invalidatetiles: message, which has a different syntax, with
    payload directly from the LOK_CALLBACK_INVALIDATE_TILES
    callback. (The latter does not contain a part number, and as the
    protocol is asynchronous, it is unclear whether a client can be
    sure, or find out with certainty, for what part the
    invalidatetiles: message is. The invalidatetiles: message will be
    dropped soon.)

nextmessage: size=<byteSize>

    <byteSize> is the size, in bytes, of the next message, in case it
    is "large". (In practice, nextmessage: messages precede each tile:
    message). Can be ignored by clients using an API that can read
    arbitrarily large buffers from a WebSocket (like JavaScript), but
    must be handled by clients that cannot (like those using Poco
    1.6.0, like the "loadtest" program in the loolwsd sources).

status: type=<typeName> parts=<numberOfParts> current=<currentPartNumber> width=<width> height=<height> [partNames]

    <typeName> is 'text, 'spreadsheet', 'presentation', 'drawing' or 'other. Others are numbers.
    if the document has multiple parts and those have names, part names follow separated by '\n'

styles: {"styleFamily": ["styles in family"], etc. }

partpagerectangles: <payload>

    Payload format is the same as LOK_CALLBACK_TEXT_SELECTION.

textselectioncontent: <content>

    Current selection's content

tile: part=<partNumber> width=<width> height=<height> tileposx=<xpos> tileposy=<ypos> tilewidth=<tileWidth> tileheight=<tileHeight> [timestamp=<time>] [renderid=<id>]
<binaryPngImage>

    The parameters from the corresponding 'tile' command.

    Additionally, in a debug build, the renderid is either a unique
    identifier, different for each actual call to LibreOfficeKit to
    render a tile, or the string 'cached' if the tile was found in the
    cache.

Each LOK_CALLBACK_FOO_BAR callback causes a corresponding message to
the client, consisting of the FOO_BAR part in lowercase, without
underscore, followed by a colon, space and the callback payload. For
instance:

invalidatecursor:

invalidatetiles: <payload>

The communication between the parent process (the one keeping open the
Websocket connections to the clients) and a child process (handling
one document through LibreOfficeKit) uses the same protocol, with
the following additions and changes:

unocommandresult: <payload>

Callback that an UNO command has finished.
See LOK_CALLBACK_UNO_COMMAND_RESULT for details.

pong

    See above.

child -> parent
===============

child <id>

    Must be the first message sent from the child to the parent. The
    parent has passed the id (a 64-bit random number) to the child
    when starting it, so this is how the child identificates itself.

curpart: part=<partNumber>

    Sent to the parent process before certain messages that the parent
    needs to act on in addition to passing them on to the client, like
    invalidatetiles:

nextmessage: size=<upperlimit>

    each tile: message sent from the child to the parent is preceded
    by a nextmessage: message that gives an upper limit on the size of
    the tile: message that will follow. (We assume it is only tile:
    messages that can be "large".) Once we depend on Poco 1.6.1, where
    one doesn't need to use a pre-allocated buffer when receiving
    WebSocket messages, this will go away.

saveas: url=<url>

    <url> is a URL of the destination, encoded. Sent from the child to the
    parent after a saveAs() completed.