This caused a bug in Lua Carousel. Scenario:
- Run carousel.love
- run the abbreviations screen which creates an abbreviation called `color`
- run driver.love
Before this commit, carousel.love would silently exit, and driver.love
would then hang waiting for it to come back. The cause:
- driver.love sends MANIFEST
- in processing it, carousel.love calls color() to emit a color escape.
- the error silently quits the app
After restarting carousel.love, the error shows up in driver.love:
Error: live.lua:104: bad argument #3 to 'color' (number expected, got no value)
stack traceback:
[C]: in function 'color'
live.lua:104: in function 'receive_from_driver'
...
Scenario:
* run a whole new app (no save dir yet) as a .love file
* open the driver, make an edit
Before this commit, I was seeing errors due to the save dir not
existing. Let's depend on love.filesystem as needed. Another "good seam"
IMO (https://lobste.rs/s/idi1wt/open_source_vs_ux).
Changes inside on.initialize are minefields. Until now, if you made a
mistake when modifying on.initialize, you could end up in a situation
where the app would fail irrecoverably on the next startup. You'd have
to go dig up a text editor to fix it.
After this commit, errors in on.initialize wait for commands from
driver.love just like any other error.
Recovering from errors during initialization is a little different than
normal. I don't know how much of initialization completed successfully,
so I redo all of it.
I think this should be safe; the sorts of things we want to do on
startup tend to be idempotent just like the sorts of things we do within
an event loop with our existing error handling.
Things are still not ideal. Initialization by definition happens only
when the app starts up. When you make changes to it, you won't find out
about errors until you restart the app[1], which can be much later and a
big context switch. But at least you'll be able to fix it in the usual
way. Slightly more seamless[2].
One glitch to note: at least on Linux, an app with an initialization
error feels "sticky". I can't seem to switch focus away from it using
Alt-tab. Hitting F4 on the driver also jarringly brings the client app
back in focus when there was an initialization error. But the mouse does
work consistently. This feels similar to the issues I find when an app
goes unresponsive sometimes. The window manager really wants me to
respond to the dialog that it's unresponsive.
Still, feels like an improvement.
[1] I really need to provide that driver command to restart the app! But
there's no room in the menus! I really need a first-class command
palette like pensieve.love has!
[2] https://lobste.rs/s/idi1wt/open_source_vs_ux
The intent of hiding errors is to see if a code change I made improved
things. But many commands from the driver don't involve a code change.
It seems more useful to leave the error up in those situations.
It seems to have been introduced in commit 46889593da back in Jan, and I
was using it to summarize multiple failing tests. However, it's not used
in that scenario anymore (and anyway it seems wasteful to compute the
stack for each failing test and then throw it all away).
This is much easier to explain, and it also avoids errors like
forgetting to delete the save dir after editing sources in source dir.
One drawback: there's a little more risk of accidental data loss, either
because of bugs here or programmer mistakes. To mitigate this risk, put
your apps in a git repo and periodically `git commit`.
To find the first word we currently strip _all_ Lua comments. This is
really wasteful for large definitions. At least avoid all these
allocations if the first word is not a comment.
A comment is now in a better place.
Inlining an unnecessary function boundary removes ambiguity between the
similar names load_all and load_files_so_far. It also shrinks call
stacks on error.
Returning errors can feel warm and fuzzy, but it does no good if the
caller doesn't actually check the return value.
Scenario:
define a new function B
call it when initializing global variable A
everything works fine
restart
Before this commit you get a weird error because we silently stopped
loading, and now lots of functions are missing.
After this commit you get a better error message.
There's a protocol invariant to preserve here: If a live command sends a
response in any situation, it needs to send a response in all
situations. Otherwise the driver can't decide whether to wait or not.
Second issue introduced in commit a2451aa26. It's worth checking error
recovery in the simpler test, even though it's technically caught in the
next one. A little redundancy seems good here.
- I often find that opening up driver.love and sending any message gets
my app unwedged. Retrying should be easy. Just press a key.
- There's really no reason for the app to become unresponsive while
waiting for code changes.
This approach tries to address both. Press any key to retry, or show the
error right on the window while I whip out driver.love. You can also
copy the message to the clipboard by hitting 'c'.
Drawbacks of this approach:
- nobody's going to read the message. I myself didn't notice the
default support for copying the error message for _months_.
- more complexity for anyone reading the code to parse. More stuff
that can go wrong. Any errors within the handler will crash the app
hard. (Though now there's less code in the handler, so maybe this is
an improvement on that score.)
The best plan is still to not rely on this too much. Don't ship bugs.
1. No more version history, now we have just the contents of the current
version.
2. Editing a definition no longer changes the order in which definitions
load.
This should make repos easier to browse, and more amenable to modify.
You don't need driver.love anymore. And a stable order eliminates some
gotchas. For example:
using driver.love, define `Foo = 3` in a definition
define `Bar = Foo + 1`
edit and redefine `Foo = 4`
Before this commit, you'd get an error when you restart the app.
Definitions used to be loaded in version order, and editing a definition
would move it to the end of the load order, potentially after
definitions using it. I mostly avoided this by keeping top-level
definitions independent. It's fine to refer to any definition inside a
function body, we only need to be careful with initializers for global
variables which run immediately while loading.
After this commit you can still end up in a weird state if you modify a
definition that other later definitions use. In the above example, you
will now see Foo = 4 and Bar = 4. But when you restart, Foo = 4 and Bar
= 5. But that's no more confusing than Emacs's C-x C-e. It's still
a good idea to keep top-level definitions order-independent. It's just
confusing in a similar way to existing tools if you fail to do so. And
your tools won't tend to break as badly.
Why did I ever do my weird version history thing? I think it's my deep
aversion to risking losing any data entered. (Even though the app
currently will seem to lose data in those situations. You'd need to
leave your tools to find the data.) Now I rely on driver.love's undo to
avoid data loss, but once you shut it down you're stuck with what you
have on disk. Or in git.
I also wasn't aware for a long time of any primitives for deleting
files. This might have colored my choices a lot.
Kartik K. Agaram