The `ElectronBuilder` strategy uses `YAML#safe_load` to parse YAML
content and this limits deserialization to appropriate classes. We
recently encountered a `Tried to load unspecified class: Time` error
when using the `ElectronBuilder` strategy on a `latest-mac.yml` file
containing `releaseDate: 2022-12-01T02:02:46.419Z`.
The electron-builder YAML files we usually encounter use single
quotes around the `releaseDate` value to ensure it's treated as a
string (e.g., `releaseDate: '2022-10-12T17:55:26.718Z'`) and this is
what we do in `electron_builder_spec.rb`. The aforementioned YAML
file doesn't use single quotes around the value, so it's treated as
a timestamp and apparently this makes Psych use `Time` (which
`#safe_load` doesn't allow by default).
Seeing as we can't control the YAML content and there's a chance we
may encounter other files like this in the future, this commit
modifies the related `#safe_load` call to allow `Time` (and `Date`
for good measure). This will resolve the aforementioned error and
allow the `ElectronBuilder` strategy to work as expected in this
scenario.
- This was done with `brew typecheck --update --suggest-typed` which
(as of the previous commit) uses Spoom, yet another gem. I thought I'd
see how well it works. There are no Sorbet errors after these changes!
The `Gnome` strategy's default regex uses the `+` form of the standard
regex for matching versions like 1.2.3. However, with the switch to
the new version scheme, some packages had a release that omits a
minor and patch (i.e., `40` instead of `40.0`). The default regex
fails to match versions like this but the looser `*` form will match
both. [When creating regexes, we generally start with the `+` form
and only switch to the looser `*` form when it's necessary and
contextually-appropriate.]
This also updates the default version filtering logic that's applied
to versions using the old GNOME version scheme (below version 40).
Outside of the refactoring changes, this also filters out versions
where the patch number is 90+, as these are also unstable.
The `tags_only_debian` code in livecheck's `Git` strategy was
originally introduced in Homebrew/homebrew-livecheck#131 when
livecheck was in a less mature state and relied more on internal
special-casing like this (i.e., while we worked to add appropriate
`livecheck` blocks). This logic only has the potential to be
beneficial when a formula/cask doesn't contain a `livecheck` block
but I would argue that we shouldn't be making assumptions in the
strategy around whether tags with a `debian/` prefix should be
matched or omitted. The answer depends upon the context of a given
formula/cask and should be handled with a `livecheck` block, as we
do with other situations like this.
In a past discussion, it was preferred that we use `system_command`
in livecheck's `Git` strategy instead of `Open3.capture3` but it
wasn't feasible at the time because we couldn't prevent
`#system_command` from printing certain output. I resolved the
`SystemCommand` shortcoming long ago in Homebrew/brew#10066, so this
finally migrates the `Git` strategy to `system_command`.
Making `channel` information available in the `Item` is necessary
to be able to filter out unstable items using a `strategy` block. If
an item doesn't specify a channel, then it uses the default channel
(this is what Sparkle itself uses for updates). Channels like `beta`
are something we want to avoid for stable casks and this allows for
that type of [cask-specific] filtering.
It's technically possible to automatically filter out items that
aren't using the default channel (i.e., `channel != nil`) in
`#items_from_content` but some casks use an unstable version, so we
can't do this internally. That is to say, we wouldn't be able to
override internal filtering in a `strategy` block, as any omitted
items wouldn't be provided to the block. Conversely, if we pass all
items to a `strategy` block, we can easily filter by channel there.
We haven't been filtering by channel internally and we've only found
one cask where this has been a problem, so it seems fine for now.
It's sometimes necessary to work with all the items in a Sparkle feed
to be able to correctly identify the newest version but livecheck's
`Sparkle` strategy only passes the `item` it views as newest into a
`strategy` block.
This updates the `Sparkle` strategy to optionally pass all items into
a `strategy` block, so we can manipulate them (e.g., filtering,
sorting). This is enabled by naming the first argument of the
strategy block `items` instead of `item`. `Sparkle` `strategy` blocks
where the first argument is `item` will continue to work as expected.
This necessarily updates `#item_from_content` (now
`items_from_content`) to return all items. I've decided to move the
sorting out of `#items_from_content`, so it simply returns the items
in the order they appear. If there is ever an exceptional situation
where we need the original order, this will technically allow for it.
The sorting has instead been moved into the `#versions_from_content`
method, to maintain the existing behavior. I thought about passing
the items into the `strategy` block in their original order but it
feels like sorting by default is the better approach for now (partly
from the perspective of maintaining existing behavior) and we can
always revisit this in the future if a cask ever requires the
original order.
Lastly, this expands the `Sparkle` tests to increase coverage. The
only untested parts are `#find_versions` (which currently
requires a network request) and a couple safeguard `raise` calls
when there's a `REXML::UndefinedNamespaceException` (which shouldn't
be encountered unless something is broken).
This aligns `PageMatch` with how cask strategies handle requirements
in their `#find_versions` methods. This moves related logic from
`Livecheck#latest_version` into the `PageMatch` strategy, which feels
a bit more appropriate.
There are times where a regex isn't needed in a `strategy` block and
these changes explicitly handle that situation.
This allows the Symbol Proc format used in some `Sparkle` `livecheck`
blocks (e.g., `strategy :sparkle, &:version`) to continue working
instead of failing with a "wrong number of arguments (given 1,
expected 0)" error. This error would occur because a Symbol Proc only
only expects one argument (e.g., an `Item`, not an `Item` and a
regex/nil).
There's an argument to be made for avoiding the Symbol Proc format
for `strategy` blocks but I haven't found a way of selectively
disabling the Style/SymbolProc cop only for a `strategy` DSL method
call. That is to say, if we don't use the Symbol Proc format, `brew
style` will give a "Pass &:version as an argument to strategy instead
of a block." offense.
Besides that, this also replaces the `block` type signatures in
livecheck strategies with `T.untyped`. Sorbet doesn't know how to
handle a `Proc` with a variable number of arguments and can't be
taught how (i.e., `T.any` with a `Proc` signature for each variation
doesn't work). The aforementioned changes cause Sorbet to complain
about there being both too many and too few arguments, so the only
way to win is not to play the game. Hopefully we can restore the
`block` type signatures in the future (if upstream resolves this
years-old issue) but `T.untyped` seems to be our only option for now.
This modifies cask-related livecheck strategies to allow passing a
regex into a `strategy` block, when appropriate. These strategies
were outliers that explicitly rejected a regex even if a `strategy`
block was used, forcing any regex to be inlined in the `strategy`
block (instead of being defined using `#regex`).
With these changes, all `strategy` blocks will be able to accept a
regex, further simplifying the mental model. This also helps to
better align the various `find_versions` and `versions_from_*`
methods across strategies.
Some Apache URLs use a `filename` query string parameter instead of
`path` and the `Apache` strategy isn't applied. The parameter value
is the same as `path` (i.e., a path to a file), so this updates the
strategy's `URL_MATCH_REGEX` to handle this URL format as well.
Valid `strategy` block return types currently vary between
strategies. Some only accept a string whereas others accept a string
or array of strings. [`strategy` blocks also accept a `nil` return
(to simplify early returns) but this was already standardized across
strategies.]
While some strategies only identify one version by default (where a
string is an appropriate return type), it could be that a strategy
block identifies more than one version. In this situation, the
strategy would need to be modified to accept (and work with) an
array from a `strategy` block.
Rather than waiting for this to become a problem, this modifies all
strategies to standardize on allowing `strategy` blocks to return a
string or array of strings (even if only one of these is currently
used in practice). Standardizing valid return types helps to further
simplify the mental model for `strategy` blocks and reduce cognitive
load.
This commit extracts related logic from `#find_versions` into
methods like `#versions_from_content`, which is conceptually similar
to `PageMatch#page_matches` (renamed to `#versions_from_content`
for consistency). This allows us to write tests for the related code
without having to make network requests (or stub them) at this point.
In general, this also helps to better align the structure of
strategies and how the various `#find_versions` methods work with
versions.
There's still more planned work to be done here but this is a step
in the right direction.
