Use ANSI control characters to display text decorations matching the
VScode terminal theme, and strip them out when providing text content
for rustc diagnostics.
This adds the small `anser` library to parse the control codes, and it
also supports HTML output so it should be fairly easy to switch to a
rendered HTML/webview implementation if desired.
While VSCode [uses it's own implementation for URIs](https://github.com/microsoft/vscode-uri)
which notably doesn't have any limits of URI size, the renderer itself
relies on Web platform engine, that limits the length of the URLs and
bails out when the attribute length of an `href` inside `a` tag is too
long.
Command URIs have a form of `command:command-name?arguments`, where
`arguments` is a percent-encoded array of data we want to pass along to
the command function. For "Show References" this is a list of all file
URIs with locations of every reference, and it can get quite long.
This PR introduces another intermediary `linkToCommand` command. When
we render a command link, a reference to a command with all its arguments
is stored in a map, and instead a `linkToCommand` link is rendered
with the key to that map.
For now the map is cleaned up periodically (I've set it to every
10 minutes). In general case we'll probably need to introduce TTLs or
flags to denote ephemeral links (like these in hover popups) and
persistent links and clean those separately. But for now simply keeping
the last few links in the map should be good enough. Likewise, we could
add code to remove a target command from the map after the link is
clicked, but assuming most links in hover sheets won't be clicked anyway
this code won't change the overall memory use much.
Closes#9926
Some features of rust-analyzer requires support for custom commands on
the client side. Specifically, hover & code lens need this.
Stock LSP doesn't have a way for the server to know which client-side
commands are available. For that reason, we historically were just
sending the commands, not worrying whether the client supports then or
not.
That's not really great though, so in this PR we add infrastructure for
the client to explicitly opt-into custom commands, via `extensions`
field of the ClientCapabilities.
To preserve backwards compatability, if the client doesn't set the
field, we assume that it does support all custom commands. In the
future, we'll start treating that case as if the client doesn't support
commands.
So, if you maintain a rust-analyzer client and implement
`rust-analyzer/runSingle` and such, please also advertise this via a
capability.
