I think the problem is that when a proxy is not in use proxy-agent
returns the global agent...which is itself since we set it globally,
causing the loop.
VS Code already covers proxies meaning we only need to do it in our own
requests so to fix this pass in the agent in the version fetch request
instead of overidding globally.
Also avoid proxy-from-env and pass in the proxy URI instead as both
http_proxy and https_proxy can be used for either http or https requests
but it does not allow that.
* Implement last opened functionality
Fixes https://github.com/cdr/code-server/issues/4619
* Fix test temp dirs not being cleaned up
* Mock logger everywhere
This suppresses all the error and debug output we generate which makes
it hard to actually find which test has failed. It also gives us a
standard way to test logging for the few places we do that.
* Use separate data directories for unit test instances
Exactly as we do for the e2e tests.
* Add integration tests for vscode route
* Make settings use --user-data-dir
Without this test instances step on each other feet and they also
clobber your own non-test settings.
* Make redirects consistent
They will preserve the trailing slash if there is one.
* Remove compilation check
If you do a regular non-watch build there are no compilation stats so
this bricks VS Code in CI when running the unit tests.
I am not sure how best to fix this for the case where you have a build
that has not been packaged yet so I just removed it for now and added a
message to check if VS Code is compiling when in dev mode.
* Update code-server update endpoint name
The problem before was that the pop() caused the open in existing
instance functionality to break because the arguments no longer
contained the file.
We could simply remove the pop() but since `workspace` and `folder` are
not CLI arguments I think it makes sense to handle them in a separate
function which can be called at the point where they are needed. This
also lets us de-duplicate some logic since we create these arguments in
two spots and lets us skip this logic when we do not need it.
The pop() is still avoided because manipulating a passed-in object
in-place seems like a risky move. If we really need to do this we
should copy the positional argument array instead.
* Add tests for relativeRoot
* Remove path.posix.join
Since this is for file system paths it feels incorrect to use it on
URL paths as they are different in many ways.
* Rewrite cookie path logic
Before we relied on the client to resolve the base given to it by the
backend against the path.
Instead have the client pass that information along so we can resolve it
on the backend. This means the client has to do less work.
* Do not remove out directory before watch
This is re-used for incremental compilation.
Also remove del since that was the only use (and we can use fs.rmdir in
the future if we need something like this).
* Remove unused function resolveBase
* Flesh out fixes to align with upstream.
* Update route handlers to better reflect fallback behavior.
* Add platform to vscode-reh-web task
Our strategy has been to build once and then recompile native modules
for individual platforms. It looks like VS Code builds from scratch for
each platform.
But we can target any platform, grab the pre-packaged folder, then
continue with own packaging.
In the future we may want to rework to match upstream.
* Fix issue where workspace args are not parsed.
* Fix issues surrounding opening files within code-server's terminal.
* Readd parent wrapper for hot reload.
* Allow more errors.
* Fix issues surrounding Coder link.
* Add dir creation and fix cli
It seems VS Code explodes when certain directories do not exist so
import the reh agent instead of the server component since it creates
the directories (require patching thus the VS Code update).
Also the CLI (for installing extensions) did not seem to be working so
point that to the same place since it also exports a function for
running that part of the CLI.
* Remove hardcoded VSCODE_DEV=1
This causes VS Code to use the development HTML file. Move this to the
watch command instead.
I deleted the other stuff before it as well since in the latest main.js
they do not have this code so I figure we should be safe to omit it.
* Fix mismatching commit between client and server
* Mostly restore command-line parity
Restore most everything and remove the added server arguments. This
will let us add and remove options after later so we can contain the
number of breaking changes.
To accomplish this a hard separation is added between the CLI arguments
and the server arguments.
The separation between user-provided arguments and arguments with
defaults is also made more clear.
The extra directory flags have been left out as they were buggy and
should be implemented upstream although I think there are better
solutions anyway. locale and install-source are unsupported with the
web remote and are left removed. It is unclear whether they were used
before anyway.
Some restored flags still need to have their behavior re-implemented.
* Fix static endpoint not emitting 404s
This fixes the last failing unit test.
Fix a missing dependency, add some generic reverse proxy support for the
protocol, and add back a missing nfpm fix.
* Import missing logError
* Fix 403 errors
* Add code-server version to about dialog
* Use user settings to disable welcome page
The workspace setting seems to be recognized but if so it is having no
effect.
* Update VS Code cache step with new build directories
Co-authored-by: Asher <ash@coder.com>
* fix(testing): revert change & fix playwright tests
* fix(constants): add type to import statement
* refactor(e2e): delete browser test
This test was originally added to ensure playwright was working.
At this point, we know it works so removing this test because it doesn't help
with anything specific to code-server and only adds unnecessary code to the
codebase plus increases the e2e test job duration.
* chore(e2e): use 1 worker for e2e test
I don't know if it's a resources issue, playwright, or code-server but it seems
like the e2e tests choke when multiple workers are used.
This change is okay because our CI runner only has 2 cores so it would only use
1 worker anyway, but by specifying it in our playwright config, we ensure more
stability in our e2e tests working correctly.
See these PRs:
- https://github.com/cdr/code-server/pull/3263
- https://github.com/cdr/code-server/pull/4310
* revert(vscode): add missing route with redirect
* chore(vscode): update to latest fork
* Touch up compilation step.
* Bump vendor.
* Fix VS Code minification step
* Move ClientConfiguration to common
Common code must not import Node code as it is imported by the browser.
* Ensure lib directory exists before curling
cURL errors now because VS Code was moved and the directory does not
exist.
* Update incorrect e2e test help output
Revert workers change as well; this can be overridden when desired.
* Add back extension compilation step
* Include missing resources in release
This includes a favicon, for example. I opted to include the entire
directory to make sure we do not miss anything. Some of the other
stuff looks potentially useful (like completions).
* Set quality property in product configuration
When httpWebWorkerExtensionHostIframe.html is fetched it uses the web
endpoint template (in which we do not include the commit) but if the
quality is not set it prepends the commit to the web endpoint instead.
The new static endpoint does not use/handle commits so this 404s.
Long-term we might want to make the new static endpoint use commits like
the old one but we will also need to update the various other static
URLs to include the commit.
For now I just fixed this by adding the quality since:
1. Probably faster than trying to find and update all static uses.
2. VS Code probably expects it anyway.
3. Gives us better control over the endpoint.
* Update VS Code
This fixes several build issues.
* Bump vscode.
* Bump.
* Bump.
* Use CLI directly.
* Update tests to reflect new upstream behavior.
* Move unit tests to after the build
Our code has new dependencies on VS Code that are pulled in when the
unit tests run. Because of this we need to build VS Code before running
the unit tests (as it only pulls built code).
* Upgrade proxy-agent dependencies
This resolves a security report with one of its dependencies (vm2).
* Symlink VS Code output directory before unit tests
This is necessary now that we import from the out directory.
* Fix issues surrounding persistent processes between tests.
* Update VS Code cache directories
These were renamed so the cached paths need to be updated. I changed
the key as well to force a rebuild.
* Move test symlink to script
This way it works for local testing as well.
I had to use out-build instead of out-vscode-server-min because Jest
throws some obscure error about a handlebars haste map.
* Fix listening on a socket
* Update VS Code
It contains fixes for missing files in the build.
* Standardize disposals
* Dispose HTTP server
Shares code with the test HTTP server. For now it is a function but
maybe we should make it a class that is extended by tests.
* Dispose app on exit
* Fix logging link errors
Unfortunately the logger currently chokes when provided with error
objects.
Also for some reason the bracketed text was not displaying...
* Update regex used by e2e to extract address
The address was recently changed to use URL which seems to add a
trailing slash when using toString, causing the regex match to fail.
* Log browser console in e2e tests
* Add base back to login page
This is used to set cookies when using a base path.
* Remove login page test
The file this was testing no longer exists.
* Use path.posix for static base
Since this is a web path and not platform-dependent.
* Add test for invalid password
Co-authored-by: Teffen Ellis <teffen@nirri.us>
Co-authored-by: Asher <ash@coder.com>
CodeQL caught a path where we were passing in req.query.path
to pathToFsPath, which may not have been a string.
So we refactored some things to ensure we only pass it a string
which also let us change the parameter type to string
instead of string | string[].
There's a chance this function can be called with a path that is not a string.
To catch that, we check if path is of a different type and throw an error if it
is.
This also adds a couple tests for this function.
This can be used to escape any special characters in a string with HTML before
sending from the server back to the client. This is important to prevent a
cross-site scripting attack.
Previously, we used argon2 to verify the hash with the password.
If the hash didn't start with a $, then it would enter the catch block.
Now we check the hash before trying to verify it and we also throw an Error if
the verify fails.
This makes the isHashMatch function more robust.
Joe Previte