Is this a type system bug or am I missing somethin...

# getting-started

[Stephan Schröder]

Is this a type system bug or am I missing something? (kotlin version: 2.2.21)
So this code works: (notice that OpenRouterModels is an object, so OpenRouterModels can denote a type or an instance)

import ai.koog.prompt.llm.LLModel
import ai.koog.prompt.executor.clients.openrouter.OpenRouterModels
import kotlin.reflect.full.memberProperties

private val openRouterModelsByName: Map<String, LLModel> = OpenRouterModels::class.memberProperties
    .filter { it.returnType.classifier == LLModel::class }
    .associate { it.name to it.get(OpenRouterModels) as LLModel }

and I would expect that the following code is equivalent (it's another way of doing the filtering):

private val openRouterModelsByName: Map<String, LLModel> = OpenRouterModels::class.memberProperties
    .mapNotNull {
        val name: String = it.name
        val model: LLModel? = it.get(OpenRouterModels) as? LLModel
        model?.let { name to model }
    }.toMap()

alas the it.get(OpenRouterModels) fails 🤯with Caused by: java.lang.IllegalArgumentException: Callable expects 0 arguments, but 1 were provided. even though IntelliJ shows no problem with the code.
OpenRouterModels is defined in "ai.koogkoog agents0.5.3" but I don't expect this issue to be linked to koog.

[CLOVIS]

These two code samples are only equivalent if `it.get(OpenRouterModels)`never throws an exception, but that's apparently what's going on.

Try:

private val openRouterModelsByName: Map<String, LLModel> = OpenRouterModels::class.memberProperties
    .mapNotNull {
        if (it.returnType.classifier != LLModel::class)
            return@mapNotNull null

        val name: String = it.name
        val model: LLModel? = it.get(OpenRouterModels) as? LLModel
        model?.let { name to model }
    }.toMap()

[CLOVIS]

Maybe log the name of the field that complains and see what its signature is?

[Youssef Shoaib [MOD]]

Yeah it seems like that some memberProperty is somehow not actually a member... I think you could check if the underlying java function is static or not.
The first debugging step I'd do is just print all the member properties

[Stephan Schröder]

The code fails for the property:

private val additionalCapabilities: List<LLMCapability> = listOf(
        LLMCapability.Schema.JSON.Standard,
        LLMCapability.ToolChoice
    )

which is the first private property that is iterated over, but that isn't an explaination to way get shouldn't have a single parameter (the instance that this property is to be extracted from) !?

[Stephan Schröder]

@Youssef Shoaib [MOD] memberProperty is defined in KClasses.kt:

@SinceKotlin("1.1")
val <T : Any> KClass<T>.memberProperties: Collection<KProperty1<T, *>>
    get() = (this as KClassImpl<T>).data.value.allNonStaticMembers.filter { it.isNotExtension && it is KProperty1<*, *> } as Collection<KProperty1<T, *>>

[Youssef Shoaib [MOD]]

That's really strange. I mean, this code shouldn't work for private properties anyways since you haven't marked their field accessible. But somehow, for some bizarre reason, this 0-arg error is thrown instead of complaining about access violation.
Maybe see where exactly the error is thrown? That might help explain why it's behaving so weirdly

[Stephan Schröder]

gradle completely swallows the StackTrace, it only shows errors in Gradle. I tried to print the stacktrace myself via:

val OpenRouterModels.modelsByName: Map<String, LLModel>
    get() {
        return runCatching {
            openRouterModelsByName
        }.getOrElse{e->
            e.printStackTrace(System.out)
            throw e
        }
    }

after making the computation lazy

private val openRouterModelsByName: Map<String, LLModel> by lazy {
    OpenRouterModels::class.memberProperties
        .mapNotNull {
            val name: String = it.name
            println("name: $name")
            val model: LLModel? = it.get(OpenRouterModels) as? LLModel
            model?.let { name to model }
        }.toMap().also { map ->
            require(map.isNotEmpty()) { "No LLModels were found in OpenRouterModels." }
        }
}

but still no stacktrace is printed 🤯🤷‍♂️

[Youssef Shoaib [MOD]]

Looks related to https://youtrack.jetbrains.com/projects/KT/issues/KT-23267/Reflection-IllegalArgumentException-Callable-expects-0-arguments-but-1-were-provided-on-retrieving-value-of-interfaces-companion, although there's no const property here

[Youssef Shoaib [MOD]]

Is Koog's additionalCapabilities marked @JvmField? If so, then it's https://youtrack.jetbrains.com/issue/KT-55872/Reflection-redundant-instanceParameter-on-property-annotated-with-JvmField
As a workaround, I'd add a try catch around the get call that checks for an IAE with that specific message, and then recover by doing a call with no arguments

[Stephan Schröder]

I pushed my code to Codeberg.

[Youssef Shoaib [MOD]]

runCatching { it.get(OpenRoutersModels) }.getOrElse { if (it is IllegalArgumentException && it.message == "Callable expects 0 arguments, but 1 were provided.") it.get() else throw it }

[Stephan Schröder]

why would printing this specific error work when printing any error doesn't?

[Stephan Schröder]

I can break on that error:

[Youssef Shoaib [MOD]]

Not printing. It's a workaround for the bug in Kotlin reflection. The linked issue shows that it messes up and shows some properties as KProperty1, that then error when called. It seems to happen with const and @JvmField properties, but some mockk issues mention something similar for private properties. Either way, it's a pretty solid workaround

[CLOVIS]

As a workaround, I'd add a try catch around the get call that checks for an IAE with that specific message, and then recover by doing a call with no arguments

Honestly, in this situation, I think keeping the initial code is better. Testing if the declared output is a given type is much cheaper than calling all the properties (which would trigger lazy initialization etc) when you're only interested by the result of a few of them

[Youssef Shoaib [MOD]]

Well, if OpenRoutersModels adds a private or @JvmField property in the future, it'll retrigger the bug...

[Stephan Schröder]

I've already got a workaround, filtering by return type first:

private val openRouterModelsByName: Map<String, LLModel> by lazy {
    OpenRouterModels::class.memberProperties
        .filter { it.returnType.classifier == LLModel::class }
        .associate { it.name to it.get(OpenRouterModels) as LLModel }
}

I've added a link to my example to the bug report you mentioned 👍

[Youssef Shoaib [MOD]]

Again, this could theoretically break if OpenRouterModels gains a new private or @JvmField property. At the very least, I'd filter out private properties, since you likely don't want them anyway (and they'll fail, I think, since you'll be trying to access a JVM Field without calling setAccessible on it first)

[Stephan Schröder]

so isAccessible is correct for that, isn't it?

.filter { it.isAccessible && it.returnType.classifier == LLModel::class }

[Stephan Schröder]

or

.filter { it.visibility == KVisibility.PUBLIC && it.returnType.classifier == LLModel::class }

[Stephan Schröder]

ok, i checked and

private val openRouterModelsByName: Map<String, LLModel> = OpenRouterModels::class.memberProperties
    .mapNotNull {
        if (it.visibility == KVisibility.PUBLIC) {
            val name: String = it.name
            val model: LLModel? = it.get(OpenRouterModels) as? LLModel
            model?.let { name to model }
       } else null
    }.toMap()

now works as well 👍 even though I prefer the version with associate .