Host bundles (even org.eclipse.osgi) leak into test launch config

[akurtakov]

Steps to reproduce:

1. In a simple plugin project put the following java file

package test;

import static org.junit.jupiter.api.Assertions.*;

class Test {

	@org.junit.jupiter.api.Test
	void test() {
		fail("Not yet implemented");
	}

}

2. Set the following as target platform

<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<?pde version="3.8"?>
<target name="Eclipse" sequenceNumber="528">
<locations>
<location includeAllPlatforms="false" includeConfigurePhase="false" includeMode="planner" includeSource="true" type="InstallableUnit">
	<unit id="org.eclipse.platform.feature.group"/>
	<unit id="org.eclipse.jdt.junit5.runtime"/>
	<repository location="https://download.eclipse.org/releases/2025-09/"/>
</location>
</locations>
</target>

3. Launch the test via Run As-> JUnit plugin test

4. You will be presented with

5. If you go and validate in the launch configuration there are no problems

but launching again gives the problem again

[akurtakov]

If one puts org.eclipse.pde.feature.group in the target platform the problem is gone.

[akurtakov]

@laeubi @merks @HannesWell WDYT of akurtakov@83d4c3f ? Not making it PR as I still don't fully understand the intention of this check.

[HannesWell]

If one puts org.eclipse.pde.feature.group in the target platform the problem is gone.

I assume you are getting some dependencies from the Host and some from the target and since these are two different resolution states in the beginning you can also get dependencies twice. In most cases this doesn't happen as both states (the the TP and the host) usually are fully resolved and then self-contained and one just has to reach to the host if a whole set is missing (i.e. if all junit bundles are missing).
o.e.osgi is probably the only overlap and one where two versions are in fact always problematic as you've shown.

WDYT of akurtakov@83d4c3f ?

I think we should just always exclude the system bundle from being added.
This could for example be done by excluding dependency if it's bundle-id is org.osgi.framework.Constants.SYSTEM_BUNDLE_ID.
(I don't find a better check on short notice).

[laeubi]

I think this is a recurrence of

• Don't include more bundles if the requirement is already fulfilled #1618

(regardless of weather we should include the system bundle at all), so actually we would need some kind of a feedback-loop in DependencyManager.findRequirementsClosure that allows us to select alternatives for a requirement so it does not unnecessarily bloat the test run in the first place.

I also noticed such problem in another case where multiple singletons where selected, so it seems the resolve operation performed during launch does not resolve singleton bundles correctly or at laest we need to handle it in some way.

In the meantime I would add a simple check to not expose the equinox system bundle from the running platform, it seems not useful as we can assume that a framework is that basic that it has to be provided by whatever target platform we use to be useful at all.

[akurtakov]

With #2083 I end up with

for the same case.

[laeubi]

I have now created a completely empty workspace with your example and can confirm I see the same error now. I'll look into this, strange enough it seems there is no test failure in the whole SDK what seems strange.