* Remove completion code from help/value shells
* Tweak the `Completer` trait for nushell.
Previously, this trait was built around rustyline's completion traits, and for
`Shell` instances. Now it is built for individual completers inside of nushell
that will complete a specific location based on a partial string. For example,
for completing a partially typed command in the command position.
Previously, we weren't expanding `~`, so `std::fs::metadata` was failing. We now
make use of `PathSuggestion` to get the actual path, as represented by a
`PathBuf`.
Previously, we'd check for a `~/` prefix and then return the home directory as
the base `PathBuf`. We failed to consider pushing any of the other possible path
components into this base dir, which prevent completions more than one level
deep (for example, `~/.config/<TAB>` would fail).
Although convenient, since the user doesn't have to type the space, it could be
a little surprising to users since they may think that was the only completion
in certain completions modes (for example, `cycle`).
* Use the cursor position for the span when between locations.
This fixes a bug where completing
ls <TAB>
would result in replacing the entire line.
* Revert to the default update implementation.
Replacing the length of the elected value was intended to do replacement when
one moves inside a quote. The problem is that a long elected suggestion could
replace bits of a pipeline that are after the cursor. For example:
ls <TAB> | get name | str collect
Previously, we used rustyline's filename completer. This allowed us to make
progress on the completion engine without building all the parts at once. We now
need our own filename completer to make progress.
The primary driver to having our own filename completer is that it can better
integrate with our path constructs. For example, if we have
> ls .../<TAB>
we want to show a list of suggestions that includes all files two directories up
from the current working directory. The least jarring experience to a user would
be to maintain the three dots. The easiest way for us to do this is by building
our own completer and path constructs.
* Show directories and executable for command completion.
Previously we chose from two sets for completing the command position:
1. internal commands, and
2. executables relative to the PATH environment variable.
We now also show directories/executables that match the relative/absolute path
that has been partially typed.
* Fix for Windows
Previously, lite parse would stack up opening delimiters in vec, and if we
didn't close everything off, it would simply return an error with a partial form
that didn't include the missing closing delimiters. This commits adds those
delimiters so that `classify_block` can parse correctly.
The parser sees this as a positional argument, but when requesting completions
this could be either a filename that starts with a hyphen, or it could be a
flag. This expands the completion engine's interface to return a vec of possible
completion locations instead of an optional one, because we want to show all
possibilities instead of assuming one or the either.
The completion engine maps completion locations to spans on a line, which
indicate whther to complete a command name, flag name, argument, and so on.
Initial implementation is simplistic, with some rough edges, since it relies
heavily on the parser's interpretation. For example
du -
if asking for completions, `-` is considered a positional argument by the
parser, but the user is likely looking for a flag. These scenarios will be
addressed in a series of progressive enhancements to the engine.
