| Age | Commit message (Collapse) | Author |
|---|
|
Composer::getLoop() answers, so a plugin reaches the graph's own downloader
and executor through the route Composer's own docblocks point plugin authors
at, rather than through an instance of its own.
wait() drains the promises it is given and rethrows the first rejection once
the group is done, which is what React\Promise\all() hands PHP. abortJobs()
has nothing to cancel while every request settles before the call that
started it returns. A non-null $progress is an explicit error: a ProgressBar
is a symfony/console object each world runs its own implementation of, so
there is none to hand across.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
add() and addCopy() answer with a promise, and enableAsync(), wait() and
countActiveJobs() answer alongside them. The Rust future runs to completion
before the promise is handed over, so requests a plugin starts together run
one after another rather than overlapping; overlapping them needs a promise
representation that crosses the boundary unresolved.
Everything else the surface does is preserved. add() still refuses a
downloader outside a Loop, and it does so by throwing out of the call the way
PHP does, where a failed request instead arrives as a rejection the caller
handles — __shirabe_rejected_promise is the failure half of the resolved-
promise helper. wait() and countActiveJobs() answer for a downloader with no
outstanding job, which, once every request settles before its call returns,
it never has.
This is where HttpDownloader parts company with ProcessExecutor, whose async
surface stays an explicit error: executeAsync() resolves its promise with a
Symfony Process, whose state is the proc_open() resource of whichever process
called start(), where a request resolves its promise with a Response.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
A response is built for one request and the graph never retains it, so there
is no entity for a handle to point at. The child holds a real instance
instead, revived from the object record the wire carries, and collect() frees
the copy each world holds — which is what that method is for. The
value-object rule rejects the class only because collect() assigns to $this,
so the category comes from an overrides.list entry.
HttpDownloader::get() and copy() answer with one.
Two gaps stay: decodeJson() reaches Composer\Json\JsonFile, which a guard
shadows, and Composer answers a curl request with the CurlResponse subclass
where this port flattens the value into a Response.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
A downloader carries its own options, TLS defaults and request backends, and
what it shares with the graph is the IO it collects authentication into and
the config it reads. Plugin code writing `new HttpDownloader($io, $config)`
therefore allocates a Rust-side entity of its own rather than a second
downloader the graph knows nothing about, and the guard that shadowed the
class in the worker is gone.
The request surface is not served yet: get() and copy() have no wire
representation for the Response they return, and the async surface resolves
its promises with one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
A Rust-side failure reached plugin code as a RuntimeException whose message
carried the name of the call that failed, so `catch (TransportException $e)`
never matched and the status code the plugin branches on was gone.
The Throw frame now names the class the exception was thrown as and carries
the state that class declares beyond message and code.
\Shirabe\MaterializedThrowable rebuilds it in the child: `new $class($message,
$code)` for a class whose constructor has \Exception's shape, then the
properties by reflection. A class the child cannot build that way keeps the
RuntimeException shape.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Composer reaches this class two ways: the object graph hands one out through
Composer::getLoop()->getProcessExecutor(), and plugins write
`new ProcessExecutor($io)` freely. Both bind to a Rust-side entity, so the
timeout the run shares -- seeded from process-timeout and rewritten while the
run is in flight -- has one value instead of one per world, and the executor
can still be passed to the classes that take one (`new Filesystem($process)`).
Three things the stub generator was missing came with it:
- By-ref parameters. The call carries their positions and the answer carries
what each holds afterwards; a position the answer omits was never assigned
to, which is what PHP does with an untouched by-ref parameter.
ProcessExecutor::execute is the only one on a proxied class.
- Argument arity, reproduced where the real body reads func_num_args().
execute($cmd) forwards the child's output and execute($cmd, $out) captures
it, and nothing but the argument count separates the two.
- Static methods that cannot run in the worker. One that reads a static
property the Rust side owns, or that reaches a guarded class, forwards
through __shirabeCallStatic instead of being materialized. That also fixes
Filesystem::isLocalPath and getPlatformPath, whose materialized bodies
called the guarded Composer\Util\Platform.
The async surface stays an explicit error. executeAsync resolves its promise
with a Symfony Process, whose proc_open() resource and pipes belong to
whichever process called start(), so a Rust-side spawn has none to hand back;
running the real start() in the worker needs a promise representation that
crosses the boundary unresolved.
The fixture project drives the whole synchronous surface from plugin code and
compares the trace against upstream Composer byte for byte.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Comments that pointed at design notes kept outside the repository are dead
ends for anyone reading only the tree, so what each of them explained now
lives in a tagged TODO at the site it applies to. Several of those sites
also stated something the implementation does not do, and the TODOs record
the actual gap instead: the two halves of the codec recognize handle
descriptors by different rules, the scripts Command path drops the exception
class and collects output in a BufferedOutput that cannot carry an
interactive command, find_shortest_path panics where PHP throws, and the
package dispatch hand-rolls the variant selection AnyPackage should own.
The classifier document likewise described rust-snapshot,
plugin-constructible and several of the open questions as designed rather
than as built.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The R-table dispatchers hand-rolled argument decoding and return encoding per
method, spread over 24 helpers. `FromPluginArg` and `ToPluginValue` replace
them: blanket impls for `Option<T>`, `Vec<T>` and `IndexMap<String, T>` compose
over a handful of leaf types, so a composite like `Vec<IndexMap<String, String>>`
needs no helper of its own, and the lossy UTF-8 conversion of a PHP string lives
in one impl instead of a dozen call sites.
Argument errors now read `{method} expects {expected} at position {position}`
rather than a message written per call site. `arg_or` takes its default for an
explicit null too, not only for an omitted argument; every parameter it serves
is declared non-nullable in the proxy stubs, so PHP never passes one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
A PHP script listener that walks the event it receives (Laravel's
Illuminate\Foundation\ComposerScripts::postAutoloadDump asks for
$event->getComposer()->getConfig()->get('vendor-dir')) aborted the run
with `unknown Rust handle 2`.
The event's getComposer/getIO answers register an entity in the R table
and hand back its rhandle, but ScriptRpcDispatcher resolved only rhandle
0 and the one event handle of the call in flight, so it could not serve a
method on a handle it had just minted itself. The R-table lookup
PluginRpcDispatcher already does is now dispatch_r_table_method, shared
by both.
The stubs that graph hands out extend and implement the real Composer
contracts (Composer\Package\PackageInterface and the rest), which live in
the Composer PHP runtime and not among the generated stubs or guards, so
execute_event_php_script loads that runtime the way the plugin and
command-class paths do.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
PackageInterface::getRequires() and friends return the array of Link objects;
in PHP that is a copy-on-write array of object references, so a caller pays
nothing to look at it. The port returned IndexMap<String, Link> by value, so
every call deep-cloned the whole map, keys and constraints included. Pool
building calls these accessors once per package per candidate, which put
IndexMap::clone at 13.5% of `require laravel/laravel`.
Store the maps as Rc<IndexMap<String, Link>> and return a handle. Callers that
mutate the map clone it explicitly at the point of mutation, matching where PHP
would separate the array.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Options and arguments were stored and passed as PhpMixed even though
Symfony only ever puts a string, a bool, a list of strings or null in
one. get_option already narrowed to InputOptionValue at the boundary;
this widens that enum into InputValue and pushes it through
InputInterface, InputOption/InputArgument defaults, the Input storage,
ArgvInput/ArrayInput/StringInput/CompletionInput, Command::add_option
and add_argument, and the Composer-side wrappers.
Two neighbouring string|int unions get types of their own:
InputDefinition::{get_argument,has_argument} take an ArgumentName, and
ArrayInput keys its parameters by ParameterName. has_parameter_option
and get_parameter_option take the values they look for as &[&str],
which is what PHP's `(array) $values` cast produced anyway.
Two behaviours change along the way. Input::set_option on a negated
option now negates with PHP's loose bool cast rather than treating a
non-bool as false, matching `!$value`. ArrayInput::parse now resolves
an integer key to an argument position instead of looking up an
argument literally named "0".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The class was shadowed by a guard, so a plugin doing `new Filesystem()`
got an explicit error. It is classified rust-proxy and plugin-constructible
and the Rust port is complete, so listing it as a stub target and answering
its public surface from the entity is all it takes.
The constructor rejects a caller-supplied ProcessExecutor: that class has no
proxy stub, so the argument could only be a second instance the Rust side
never sees. findShortestPath re-checks its arguments at the boundary because
the port panics where PHP throws, and a panic would take the process down
instead of reaching the plugin's catch block.
phpstan/extension-installer matches upstream Composer byte for byte again.
|
|
Porting mapped every PHP `protected` member onto `pub(crate)`, which is
wider than nearly all of them need. Each item demoted here is reached
only from the module that defines it, so the crate-wide visibility
conveyed nothing.
Every `pub(crate)` that survives has at least one reader in another
module of the same crate.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
`Application::do_run` registers a `composer.json` script whose value names a
`Symfony\Component\Console\Command\Command` subclass as a live command. The
class checks and `new $dummy($script)` need a real PHP runtime, so they run in
the worker: the command object lives there and this side keeps a metadata
mirror for `list`/`help`, forwarding a run to the worker-side console
application it is added to. The name and description fixups are applied to the
worker-side object, so both sides carry the same values.
Loading the Composer PHP runtime into the worker is gated on the Rust-side
`ClassLoader`s resolving the class to a file, keeping that load out of every
run whose scripts are plain shell commands.
The worker-side console application handoff now accepts commands registered
after it was published, since the scripts scan runs after plugin commands are
collected. Whether it was published is tracked per application: the handoff is
process-wide, so a second application must replace it rather than extend it.
`shirabe_php_shim::is_subclass_of` has no callers left.
|
|
Every parent module in the symfony-* crates already re-exported its leaf
modules with `pub use`, so each item was reachable by two paths. Make the
leaf modules private and route all callers through the single re-exported
path.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
PluginManager::getPluginCapability hands the plugin itself to the
capability constructor. Only a PHP-implemented plugin had an entity the
child could receive, so a Rust-implemented one bailed out; it now crosses
as a handle to an R-table entity behind Shirabe\RustPluginStub, or its
Capable flavour, since `$plugin instanceof Capable` is what decides
whether Composer asks a plugin for capabilities at all.
This is what the two capability tests of PluginInstallerTest were waiting
on: the mocked Capable plugin is Rust-side, and one of them asserts the
identity of the plugin read back out of $capability->args.
|
|
shirabe-symfony-console crate
Move `Symfony\Component\Console` out of shirabe-external-packages and
into its own crate, so the path is
`shirabe_symfony_console::application::Application` instead of
`shirabe_external_packages::symfony::console::application::Application`.
The `delegate_to_inner!` and `delegate_command_trait_impls_to_inner!`
macros move with it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Ported exceptions were flat structs reached with `downcast_ref`, so
Composer's `catch (\RuntimeException $e)` only matched the exact leaf
type and `get_class($e)` had nothing to report. Each exception now
embeds an instance of the class it extends and travels inside an
`AnyThrowable`; `Catch::catch`/`catch_mut` walk that chain, and
`PhpClass::php_class_name` yields the PHP FQCN.
Dropping the `std::error::Error` impls from the exception types leaves
`AnyThrowable` as the only route into an `anyhow::Error`, so the walk
cannot be bypassed. A `no_exception_downcast` linter catches the
`downcast::<X>()` calls that would now silently answer `None`.
Three sites change behavior as a result: the `TransportException`
exit-code override reaches `MaxFileSizeExceededException`, the
`catch (\LogicException)` in findSimilar() reaches its subclasses, and
rendered exception titles carry the real class name rather than a
guess. `get_class_err()` is no longer a `todo!()`, which re-enables
FilesystemRepositoryTest::testCorruptedRepositoryFile.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
InstallationManager left both PRE_PACKAGE_* and POST_PACKAGE_* as empty
stubs, so a subscriber never ran at all and the difference from upstream was
silent rather than an explicit error.
Operations cross the boundary as R-table entities with generated proxy
stubs. Materializing them the way a Link crosses is not possible: a
materialized value is revived by unserialize() on the child side, so its
properties never pass through the wire decoder and a nested handle
descriptor would not come back as a stub — and an operation always holds a
PackageInterface. execute() now shares one Rc per operation through the
whole batch pipeline, so a plugin sees one object for both the pre- and the
post-event of an operation, as it does in PHP.
POST_PACKAGE_* also moves out of the operation's promise chain into the
post-exec callback list PHP runs after waitOnPromises().
The stub generator materializes non-public class constants verbatim now,
which the operation classes need for their `protected const TYPE`: a
constant has no entity behind it, so a copy in the worker cannot diverge,
and keeping the declared visibility exposes nothing the real class hides.
The E2E fixture added here compares the recorded events against upstream
Composer. It also surfaced that upstream starts an operation's chain where
it is built (a null prepare() becomes an already-fulfilled React promise
whose handlers run through the immediately drained queue) while this port
only drives its futures in wait_on_promises, so the repository state a
pre-event observes differs; that half of the comparison is a separate
`#[ignore]`d test.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
A proxy stub declares none of the real class's instance properties, and the
`__get`/`__set` forwarders were emitted only for classes that declare a public
one. Every other property access therefore got PHP's own answer for an
undeclared property — null on a read, a dynamic property on a write — so plugin
code reading state the entity holds ran on with null and failed somewhere else
entirely, or not at all.
Emit the forwarders on every root stub, add `__isset`/`__unset` alongside them
so `isset()` cannot answer false silently either, and serve all four from one
dispatcher that answers the state the Rust-side entity exposes and rejects every
other name.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
AliasPackage, CompleteAliasPackage and RootAliasPackage now have generated
proxy stubs, so a package handed to a plugin no longer has to be a real
package: it crosses as the stub matching its concrete variant, answers
getAliasOf / setRootPackageAlias / isRootPackageAlias /
hasSelfVersionRequires, and can be constructed from plugin code.
The setters RootPackageInterface declares are routed through that interface
for every root package instead of through the base Package state. Only the
alias variant needs it -- RootAliasPackage overrides all nine to write
through to the package it aliases -- but a real RootPackage delegates to the
same base state either way, so both take one path.
An alias of an alias has no representation here, so narrowing the
constructor argument to a real package is an explicit error rather than a
silent demotion.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The link getters and setters and the release date accessors were explicit
errors for every non-empty value, because an immutable value has no entity
to point a handle at. They now cross as materialized values: the descriptor
names the real class and the constructor arguments, and each side builds a
genuine instance of its own.
The semver constraint a link holds is encoded structurally rather than
re-parsed from its string form, so the pretty strings and the conjunctive
flag survive the crossing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
An installer that extends LibraryInstaller reaches for the Composer object
graph in ways the proxy did not answer: the download manager and the config,
the writable repository methods, a React promise as its own return value, and
`new Package(...)` from its supports() path.
* `Composer::getConfig`/`getDownloadManager` are dispatched, and both classes
become generated proxy stubs. Their reads and mutators answer from the Rust
entity; the surfaces needing stubs of their own (ConfigSourceInterface,
DownloaderInterface) stay explicit errors.
* The download manager's futures are driven to completion and handed back as
already-settled React promises, since PHP declares a non-nullable
PromiseInterface there. A promise a plugin returns is drained the same way:
settled yields its value, rejected re-raises, pending is an explicit error.
* Proxy stubs now carry the real class's constructor and ask the Rust side to
allocate the entity; reviving a stub for an existing entity binds the handle
without running it. Classes Rust cannot build name themselves in the error.
* The package proxy covers `Package`'s own setters, `CompletePackage`'s
metadata, `RootPackage`'s root-only state, and `BasePackage::$id`.
Link values and release dates still have no wire image, so the methods
carrying them remain explicit errors.
|
|
A plugin can now hand an InstallerInterface implementation to
InstallationManager::addInstaller across the wire, and a legacy
composer-installer package is loaded as one; both are backed by a
PhpInstallerProxy forwarding the whole installer contract to the entity in
the PHP worker. An installer returning a real promise is an explicit error
until promises can cross the boundary.
InstallationManager takes installers as shared handles instead of boxes, so
the object identity removeInstaller and PluginManager's registeredPlugins
compare against survives registration, and holds them in a RefCell:
Installer::run keeps a shared borrow of the manager for the whole run, and a
plugin activated inside it registers its installer from there. The type
cache keys on the installer itself, like upstream, so re-entrant
registration cannot leave a stale index behind. InstallerInterface::supports
is fallible for the same reason getCapabilities and getCommands are: it
answers over RPC.
Cloning a proxy stub clones the Rust-side entity and rebinds the copy to the
fresh handle. Previously only the classes declaring __clone got a throwing
body, and the rest let two stubs share (and twice release) one handle.
Package entities answer with AnyPackage::dup, which already carries
BasePackage::__clone and the RootAliasPackage override; the others are an
explicit error.
The package proxy covers the whole PackageInterface surface; only the link
maps and the release date still lack a wire image for their value objects.
PluginManager gains a test-only seam for the reported Plugin API version,
and the three PluginInstallerTest cases that need it are ported.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
A same-FQCN Composer\Console\Application, hand-written under the new
php/runtime/ tree, hosts CommandProvider commands inside the PHP worker:
PhpCommandProxy overrides run() and forwards the stringified input, so the
real Symfony machinery binds, validates and executes against the live
command object, while help/list render Rust-side from a definition read
back at construction. Reverse \Shirabe\RustCommandStub rows let a plugin
command invoke built-in commands back in the Rust process, keeping every
command on the side whose helper set it was written for.
Composer\EventDispatcher\Event moves from a generated stub to a dual-mode
runtime class: the real BaseCommand::initialize constructs a
PreCommandRunEvent natively in the worker, which a proxy-only constructor
guard rejected. Its PRE_COMMAND_RUN dispatch reaches a new EventDispatcher
stub whose dispatch supports the observably-no-op no-listener case and
fails explicitly otherwise. The stub generator now accepts runtime-provided
classes as stub bases (never as targets) and cross-checks the Application
handoff property table against the real class.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Port Application::getPluginCommands: resolve the local composer with
plugins force-enabled, fall back to Factory::createGlobal, and collect
commands from CommandProvider capability adapters. PhpCommandProxy now
mirrors name/description/aliases/hidden over RPC so `list` output
matches upstream; input definitions remain TODO(plugin).
PhpClass::php_class_name returns an owned String because PHP-backed
proxies only know their class at runtime (the override-skip warning
prints get_class). CommandProvider::getCommands hands out shared
Rc<RefCell> handles since the commands are stored in the application.
Factory::createGlobal now propagates createConfig errors instead of
swallowing them, and Application::getComposer only catches the
exception classes upstream catches, so a ParsingException reaches
doRun's GithubActionError path as in Composer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
getPluginCapability was a no-op returning None. Now it runs the real
flow: class_exists in the worker, new $capabilityClass($ctorArgs) with
the plugin's own phandle spliced in as $ctorArgs['plugin'], both
instanceof checks answered by is_a in the child, and a per-interface
Rust adapter over the resulting entity (CommandProvider and the plain
Capability marker; anything else is an explicit error).
getPluginCapabilities now propagates errors instead of swallowing them.
Capable::get_capabilities widens from IndexMap<String, String> to
IndexMap<String, PhpMixed>: upstream casts the return with (array) and
validates only the queried key, so the narrow type both rejected maps
Composer accepts and made the invalidImplementationClassNames data
provider unrepresentable. The old code also returned ' 0 ' (trimmed to
the falsy '0') as a valid class name where upstream throws; the
rewritten validation follows upstream's empty/is_string/trim sequence.
CommandProvider::get_commands returns BaseCommand adapters whose name is
read back over RPC after the PHP constructor ran configure(); executing
one still needs the PHP-side Symfony Application and stays an explicit
error. The is_array / instanceof BaseCommand checks that upstream's
Application::getPluginCommands performs on the raw getCommands value
live in the adapter, because Vec<Box<dyn BaseCommand>> asserts every
element up front.
Ports testCommandProviderCapability (plugin-v8 end to end against the
real worker) and testQueryingWithInvalidCapabilityClassNameThrows (all
eight provider cases); the two tests that pass a PHPUnit mock plugin
into PHP stay ignored — a Rust-native mock has no PHP-side entity to
cross the boundary as $ctorArgs['plugin'].
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Widen the R table beyond Composer/IO with InstallationManager,
RepositoryManager, repository and package entities, interned by pointer
identity through a shared register_entity. Script events answer
getComposer/getIO, Composer hands out its graph getters, repositories list
their packages as per-variant stubs, package getters cover the
extension-installer surface (getRequires only while empty: the Link
snapshot encoding does not exist yet), and getInstallPath resolves its
package argument back through the R table. Everything else stays an
explicit error.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Wires the addPlugin subscriber branch end to end: EventSubscriberInterface
and Capable become fallible and dyn-compatible (their sole implementor is
the PHP plugin proxy, which answers getSubscribedEvents over RPC), listeners
register as Callable::PhpMethod and are invoked with a per-call event handle,
and the R table now drops entries when a child-side stub destructs. The R
table keeps its IndexMap with monotonically increasing handles, so released
handles are never reused and no generation counter is needed.
Upstream has no subscriber-plugin test, so the path is covered by a
Shirabe-owned fixture exercising all three getSubscribedEvents shapes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Replace the hand-written proxy stubs under crates/shirabe-php-rpc/php/stubs
with output of scripts/plugin-stub-generator, a deterministic emitter that
derives every stub from the Composer checkout and the classifier report.
Anything it cannot faithfully proxy (by-ref/variadic parameters, magic
methods, public properties, diverging omitted overrides, stale stub files,
a STUB_FILES entry missing in lib.rs) fails generation instead of degrading
silently, so future Composer releases surface new members as explicit
errors rather than silent gaps.
Regenerating the stubs also normalizes the hand-written inconsistencies
(uniform guarded constructors, import-based type spellings) and fixes real
gaps the review of the generated diff uncovered: the BaseIO authentication
methods now carry the real class's untyped signatures, and the previously
missing ConsoleIO::sanitize is materialized together with its private
static helper. A cargo test runs generate-stubs --check to keep the
committed stubs, the generator and the embedded list from drifting apart.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
InstallationManager::execute now takes &self (the mock recorder moved
into a RefCell) and its callers hold only shared borrows: plugin
registration inside a batch re-enters the same manager handle through
Composer::getInstallationManager()->getInstallPath(), which panicked
on the RefCell re-borrow under the &mut shape — the same re-entrancy
the repository side already fixed, unreachable from the ported tests
because they call PluginInstaller::install directly like PHPUnit does.
The worker-side InstalledVersions mirror now matches the full tail of
FilesystemRepository::write: unconditional reload plus the
reflection-based selfDir/installedIsLocalDir restore. The previous
class_exists(false) guard rested on a lazy-load assumption that does
not hold in the worker (its real ClassLoader only knows the Composer
checkout's vendor dir, so a later lazy load would read the checkout's
installed.php, not the project's); the mirror is now skipped only when
the class is not autoloadable at all, i.e. no plugin runtime and hence
no observer code. Boot-time seeding stays TODO(plugin).
Also from the review: registered_plugins entries are removed only
after the deactivate/uninstall loop (PHP unsets last, and a throw must
leave the entry observable); extra.class keeps associative-array
values and fails loudly on non-strings instead of silently dropping
them; the two discarded write() results now propagate (they carry the
reload-push failure); register_package's allow-plugins skip message is
DEBUG like the addPlugin side; the loader-eviction divergence of
REGISTERED_LOADERS and the lossy UTF-8 spots carry searchable
markers; the test-only proxy downcast follows the __ naming rule; the
R-table dispatch clones the entity out instead of holding the table
borrow across the handler; the IO/PartialComposer stubs turn a
plugin-side `new NullIO()` into an explicit error instead of an
ArgumentCountError; and the empty() emulation covers float 0.0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Implement the remainder of PluginManager::registerPackage: the plugin
autoload map is built by the ported createLoader/parseAutoloads and
served to the worker over the existing reverse-RPC autoloader, files
entries go through a composerRequire-equivalent glue call, and
already-defined classes take the upstream _composer_tmp rename/eval
path. Instantiation uses the new NewObject/CallPhpMethod lanes backed
by a P table in the worker; PhpPluginProxy adapts the resulting handle
to PluginInterface, with $composer/$io exposed to plugin callbacks via
an R table (unsupported methods stay explicit errors). Hand-written
proxy stubs cover Composer, PartialComposer and the IO hierarchy, and
the stub autoloader is re-prepended after loading the Composer PHP
runtime so its vendor autoloader cannot shadow proxied FQCNs.
FilesystemRepository::write now mirrors InstalledVersions::reload into
a running worker (class_exists-guarded, so an unloaded class keeps its
upstream lazy-load behavior), removing the previously undefined
observation window.
The installer pipeline passes the installed repository as a shared
handle instead of a long-lived `&mut dyn`: plugin registration runs
inside InstallationManager::execute and re-enters the same local
repository through the RepositoryManager, which would panic on the
RefCell re-borrow under the old shape.
PluginInterface lifecycle methods now take an owned ComposerHandle
(plugins retain $composer past the call) and return anyhow::Result
(PHP plugin code may throw); the plugin list uses shared ownership so
the identity comparison of removePlugin survives the dual storage in
registeredPlugins, matching PHP reference semantics.
Ports the activate/upgrade/uninstall tests of PluginInstallerTest,
serialized across the shared worker process whose persistent class
table is exactly what exercises the rename path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
