When a host pushes a save to a stream that is gone, or that timed out
before anyone downloaded it, the controller answers 404 or 500 without
reading the body. Node then reads and throws away the rest of the body
to keep the connection open, so a proxy in front of the controller that
keeps passing the upload on made the whole save get uploaded for
nothing. Close the connection if the body hasn't ended a second after
the response.
Document turning off request buffering in reverse proxies, as a
buffering proxy holds the upload back until all of it has arrived.
* Answer the disconnect prepare in the close old connections test
The test's first connector had no Link, so the controller's disconnect
prepare went unanswered and it waited the full 10 s timeout.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Start the controller and host CLIs faster
Like ctl, the controller and host now compile schema validators on first
use instead of at import, which cost about 120 ms per invocation. The
Controller and Host classes are only imported by the commands that
start them, so config, bootstrap and the other commands no longer load
express, ws and the rest of the servers. clusteriocontroller config set
goes from 413 ms to 231 ms and clusteriohost config set from 412 ms to
225 ms.
The integration tests run the bins, like npx does, instead of the
package directories.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Splitting the test jobs doubled them to 14, over GitHub's concurrent
job limit together with the other jobs. Every Factorio version now runs
on the latest Node.js only, and the other Node.js versions only with
the latest Factorio, which is 10 test jobs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The clusterioctl, plugin system and instance integration tests take
about half of the test time, so they run on one runner and the rest of
the tests on another, each half with its own controller and host.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Take FieldData from antd instead of rc-field-form
web_ui depended on rc-field-form 1.x for a single type import, while
antd 5 uses rc-field-form 2.x. Derive the type from antd's FormProps
and drop the dependency.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Move typescript to web_ui devDependencies
web_ui only needs typescript to build its own types. The runtime
dependency is a leftover from webpack's ts-loader, Vite doesn't use it,
and it pulls in a 30 MB native compiler for anyone installing web_ui.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Use native FormData in controller route tests
Node's FormData and Blob do what the form-data package did here, and
fetch sets the multipart Content-Type header itself.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Remove leftover typedoc.json files
TypeDoc was dropped in #1147, but these configs still ship in the
published packages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
jimp pulls in every image format and plugin, about 60 packages and
26 MB that only export.ts uses. Build the Jimp class from @jimp/core
with the PNG format and the crop and resize plugins instead. The
generated spritesheet is byte-identical to the one from jimp.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Node's FormData and Blob do what the form-data package did here, and
fetch sets the multipart Content-Type header itself.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
web_ui only needs typescript to build its own types. The runtime
dependency is a leftover from webpack's ts-loader, Vite doesn't use it,
and it pulls in a 30 MB native compiler for anyone installing web_ui.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
web_ui depended on rc-field-form 1.x for a single type import, while
antd 5 uses rc-field-form 2.x. Derive the type from antd's FormProps
and drop the dependency.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Like ctl, the controller and host now compile schema validators on first
use instead of at import, which cost about 120 ms per invocation. The
Controller and Host classes are only imported by the commands that
start them, so config, bootstrap and the other commands no longer load
express, ws and the rest of the servers. clusteriocontroller config set
goes from 413 ms to 231 ms and clusteriohost config set from 412 ms to
225 ms.
The integration tests run the bins, like npx does, instead of the
package directories.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The test's first connector had no Link, so the controller's disconnect
prepare went unanswered and it waited the full 10 s timeout.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pnpm/action-setup installs a bootstrap pnpm and then switches to the
packageManager version, about 10 s per job; npm installs it in 3 s.
setup-node took 4 s probing package managers for a cache it never
used. The pnpm store cache took 4 s to restore and 5 s to save, more
than downloading the packages costs, and the build cache no longer pays
off with TypeScript 7: pnpm install is 13-23 s without either cache
against 17-31 s with them.
The Codecov action searched the workspace and uploaded the raw V8
coverage in coverage/tmp next to lcov.info, 72 MB taking 11 s.
The version compatibility job ran tsc --build again after pnpm install
had already built everything.
Cancel superseded runs of branches and pull requests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
loaded, enabled, pendingInstall, removed, noEntrypoint and nameTaken
allowed far more combinations than can happen, and every consumer had
to know which took priority. PluginDetails now has one state, with
loaded, loadsOnRestart and removed as getters over it.
2.0.3 web UIs and ctl send PluginListRequest straight to hosts and
require loaded and enabled, so those are still sent next to the
optional state. Plugins listed by 2.0.3 hosts get their state from
loaded and enabled.
The web UI planner gets the same treatment: MachinePlugin has one
state in place of disabled, noEntrypoint, dependencyTaken and
nameTaken. ctl host plugin list shows the state.
Closes#1134
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
While the host has no controller session, events between the instance
and the controller are dropped in both directions, and research_sync
only synced technologies on instance start. A technology finished during
the outage stayed unresearched on the controller and on every other
instance until the instance restarted.
On a new session the instance now runs the same technology sync as on
start, before taking the cluster queue. Contributions made while the
controller is unreachable are summed per technology and sent on connect
or resume, since progress from several instances can't be rebuilt from a
state sync that keeps the highest progress. A finish made while the
connection was dropped triggers the sync on resume too.
sync_technologies now runs while the game is running, so it skips
writing progress to a technology that finished after the dump.
Fixes#1123
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user page set the role selector once and never updated it. The role
page reset its whole form on every role update, wiping unsaved edits.
Both pages now follow the live data and keep local edits on top. The role
page stores only the edited fields and drops them once they match the role
or when the viewer loses core.role.update. The user page syncs the
selector while it is not dirty and drops deleted roles from a dirty
selection. A role deleted while open now says so instead of "not found".
Fixes#775, fixes#776
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The button offered Stop while the instance was starting whenever the
start came from somewhere else, such as Load save, another tab or ctl.
It now shows the action in progress as loading until the status
settles. Kill stays available from the dropdown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A stop request sent while the instance was starting ran alongside the
start. During save preparation it did nothing and the instance came up
running anyway. Once Factorio was launched it put the server into the
stopping state, the start then failed with "Expected state running but
state is stopping" and its cleanup killed the server mid shutdown.
Stop and restart requests now wait for any start, stop or save in
progress to finish first, then stop the instance if it is running.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since TypeScript 6 the types option defaults to an empty list, so
@types/node was only loaded because winston and undici-types reference
it. Name it explicitly. The emitted files do not change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TypeScript 7.1 reports TS5088 for it: the inferred type of the static
Response property references itself through the declared constructor.
The other inline Response classes do not declare it either.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TypeScript 7 is the native port of the compiler. The repository builds
with it unchanged: no new type errors, the emitted JavaScript differs
only in the order of react/jsx-runtime imports, and the declarations
only in union and property order, which TypeScript 5.9 consumers read
the same way. The plugin template moves to it too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The generated API docs were never published or built in CI, and
typedoc's peer range stops at TypeScript 6.0. TSDoc comments stay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Entries now explain why end users are not or only minimally affected,
and the docs say which version to key by and point to the breaking
changes decision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The version-compat job fails every pull request once master breaks
compatibility with the latest release on purpose. List those breaks in
scripts/version-compat-breaks.json keyed by the published version they
break against, each with a pr number and reason. Listed messages warn
instead of failing, connect skips the cluster test, entries for other
versions are ignored, and a listed message that did not change fails.
Add the two src changes from #1129 against 2.0.3 so master passes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Keep auto paused instances paused
Every RCON command makes an auto paused Factorio server run one tick, so
subspace_storage's 1 s inventory update and 5 s ping, plus a
statistics_exporter gather on every Prometheus scrape, kept a server with
nobody online running at about 1.3 ticks per second.
The instance now records the auto_pause server setting it was started
with and exposes autoPaused, true while that setting is on and no players
are online, along with an autoPauseChanged hook. subspace_storage skips
its timers while paused and sends a ping and inventory as soon as a player
joins. statistics_exporter gathers once after the game pauses and keeps
serving those values until it resumes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Check connection and status in subspace_storage timers
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Every RCON command makes an auto paused Factorio server run one tick, so
subspace_storage's 1 s inventory update and 5 s ping, plus a
statistics_exporter gather on every Prometheus scrape, kept a server with
nobody online running at about 1.3 ticks per second.
The instance now records the auto_pause server setting it was started
with and exposes autoPaused, true while that setting is on and no players
are online, along with an autoPauseChanged hook. subspace_storage skips
its timers while paused and sends a ping and inventory as soon as a player
joins. statistics_exporter gathers once after the game pauses and keeps
serving those values until it resumes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Users don't need the web UI load state apart from the plugin's own state now
that the web UI reloads itself. A web load error now shows under the plugin's
versions and on its view page, only for enabled plugins.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds the last three checks from #342 to test_plugin: an InstanceEcho
request handled on a running instance, a Lua module with a remote
interface that the request calls over RCON, and log tagging checks for
the plugin field and host/instance ids on the controller, host and
instance log lines.
test_plugin now has a module, so the create_scenario and patch unit
tests that relied on it having none use global_chat instead.
Closes#342
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The template asked for TypeBox ^0.30.6 while lib uses ^0.34, so a new
plugin failed tsc on the first message class. TypeBox is now provided
to plugin web code through the import map, so the template should be on
the version the web interface ships.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Controller.viteServer becomes devServer, matching the devMiddleware
name it replaced. The static_url limitation is described as one of the
web build rather than webpack.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Vite serves any file under fs.allow through /@fs/, and the default
allow list is the workspace root. In a clusterio checkout that is also
the usual data directory, so with --dev any client that could reach the
controller could fetch config-controller.json, config-ctl.json and the
database without logging in. webpack-dev-middleware only served the
compiled bundles.
Allow the controller, web_ui, lib and dev plugin directories and the
node_modules directories Node.js would resolve their imports from.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Nothing else from Vite+ is used, and Oxlint reads the same rules from
.oxlintrc.json. This drops vite-plus and with it Vitest, a second copy
of Vite and about 130 packages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Danielv123