LLMParams and Responses API

integration-tests/src/jvmTest/kotlin/ai/koog/integration/tests/utils/Models.kt

@@ -22,10 +22,12 @@ object Models {
         return Stream.of(
             OpenAIModels.Chat.GPT4o,
             OpenAIModels.Chat.GPT4_1,
+            OpenAIModels.Chat.GPT5,
+            OpenAIModels.Chat.GPT5Mini,
+            OpenAIModels.Chat.GPT5Nano,
 
             OpenAIModels.Reasoning.O4Mini,
             OpenAIModels.Reasoning.O3Mini,
-            OpenAIModels.Reasoning.O1Mini,
             OpenAIModels.Reasoning.O3,
             OpenAIModels.Reasoning.O1,
 

koog-ktor/src/commonMain/kotlin/ai/koog/ktor/utils/LLMModelParser.kt

@@ -182,12 +182,14 @@ private fun openAI(parts: List<String>, identifier: String): LLModel? {
 private val OPENAI_MODELS_MAP = mapOf(
     "chat" to mapOf(
         "gpt4o" to OpenAIModels.Chat.GPT4o,
-        "gpt4_1" to OpenAIModels.Chat.GPT4_1
+        "gpt4_1" to OpenAIModels.Chat.GPT4_1,
+        "gpt5" to OpenAIModels.Chat.GPT5,
+        "gpt5mini" to OpenAIModels.Chat.GPT5Mini,
+        "gpt5nano" to OpenAIModels.Chat.GPT5Nano,
     ),
     "reasoning" to mapOf(
         "o4mini" to OpenAIModels.Reasoning.O4Mini,
         "o3mini" to OpenAIModels.Reasoning.O3Mini,
-        "o1mini" to OpenAIModels.Reasoning.O1Mini,
         "o3" to OpenAIModels.Reasoning.O3,
         "o1" to OpenAIModels.Reasoning.O1
     ),
@@ -196,7 +198,6 @@ private val OPENAI_MODELS_MAP = mapOf(
         "gpt4_1nano" to OpenAIModels.CostOptimized.GPT4_1Nano,
         "gpt4_1mini" to OpenAIModels.CostOptimized.GPT4_1Mini,
         "gpt4omini" to OpenAIModels.CostOptimized.GPT4oMini,
-        "o1mini" to OpenAIModels.CostOptimized.O1Mini,
         "o3mini" to OpenAIModels.CostOptimized.O3Mini
     ),
     "audio" to mapOf(
@@ -209,7 +210,6 @@ private val OPENAI_MODELS_MAP = mapOf(
         "textembeddingada002" to OpenAIModels.Embeddings.TextEmbeddingAda002
     ),
     "moderation" to mapOf(
-        "text" to OpenAIModels.Moderation.Text,
         "omni" to OpenAIModels.Moderation.Omni
     )
 )
@@ -238,6 +238,10 @@ private val OPENROUTER_MODELS_MAP = mapOf(
     "claude3haiku" to OpenRouterModels.Claude3Haiku,
     "gpt4" to OpenRouterModels.GPT4,
     "gpt4o" to OpenRouterModels.GPT4o,
+    "gpt5" to OpenRouterModels.GPT5,
+    "gpt5mini" to OpenRouterModels.GPT5Mini,
+    "gpt5nano" to OpenRouterModels.GPT5Nano,
+    "gptoss120b" to OpenRouterModels.GPT_OSS_120b,
     "gpt4turbo" to OpenRouterModels.GPT4Turbo,
     "gpt35turbo" to OpenRouterModels.GPT35Turbo
 )

koog-ktor/src/commonTest/kotlin/ai/koog/ktor/ModelIdentifierParsingTest.kt

@@ -45,12 +45,6 @@ class ModelIdentifierParsingTest {
         assertEquals(LLMProvider.OpenAI, o3Mini.provider)
         assertEquals(OpenAIModels.Reasoning.O3Mini, o3Mini)
 
-        // Test O1 Mini
-        val o1Mini = getModelFromIdentifier("openai.reasoning.o1mini")
-        assertNotNull(o1Mini)
-        assertEquals(LLMProvider.OpenAI, o1Mini.provider)
-        assertEquals(OpenAIModels.Reasoning.O1Mini, o1Mini)
-
         // Test O3
         val o3 = getModelFromIdentifier("openai.reasoning.o3")
         assertNotNull(o3)
@@ -123,12 +117,6 @@ class ModelIdentifierParsingTest {
 
     @Test
     fun testOpenAIModerationModels() = runTest {
-        // Test Text Moderation
-        val textModeration = getModelFromIdentifier("openai.moderation.text")
-        assertNotNull(textModeration)
-        assertEquals(LLMProvider.OpenAI, textModeration.provider)
-        assertEquals(OpenAIModels.Moderation.Text, textModeration)
-
         // Test Omni Moderation
         val omniModeration = getModelFromIdentifier("openai.moderation.omni")
         assertNotNull(omniModeration)

prompt/prompt-executor/prompt-executor-clients/prompt-executor-deepseek-client/src/commonMain/kotlin/ai/koog/prompt/executor/clients/deepseek/DeepSeekLLMClient.kt

@@ -4,9 +4,17 @@
 import ai.koog.prompt.dsl.Prompt
 import ai.koog.prompt.executor.clients.ConnectionTimeoutConfig
 import ai.koog.prompt.executor.clients.LLMClient
+import ai.koog.prompt.executor.clients.deepseek.models.DeepSeekChatCompletionRequest
+import ai.koog.prompt.executor.clients.deepseek.models.DeepSeekChatCompletionResponse
+import ai.koog.prompt.executor.clients.deepseek.models.DeepSeekChatCompletionStreamResponse
 import ai.koog.prompt.executor.clients.openai.AbstractOpenAILLMClient
 import ai.koog.prompt.executor.clients.openai.OpenAIBasedSettings
+import ai.koog.prompt.executor.clients.openai.models.OpenAIMessage
+import ai.koog.prompt.executor.clients.openai.models.OpenAITool
+import ai.koog.prompt.executor.clients.openai.models.OpenAIToolChoice
+import ai.koog.prompt.executor.model.LLMChoice
 import ai.koog.prompt.llm.LLModel
+import ai.koog.prompt.params.LLMParams
 import io.github.oshai.kotlinlogging.KLogger
 import io.github.oshai.kotlinlogging.KotlinLogging
 import io.ktor.client.HttpClient
@@ -32,29 +40,70 @@
  * defaults to "https://api.deepseek.com" and 900s
  * @param clock Clock instance used for tracking response metadata timestamps.
  */
 public class DeepSeekLLMClient(
     apiKey: String,
     private val settings: DeepSeekClientSettings = DeepSeekClientSettings(),
     baseClient: HttpClient = HttpClient(),
     clock: Clock = Clock.System
-) : AbstractOpenAILLMClient(apiKey, settings, baseClient, clock) {
+) : AbstractOpenAILLMClient<DeepSeekChatCompletionResponse, DeepSeekChatCompletionStreamResponse>(
+    apiKey,
+    settings,
+    baseClient,
+    clock
+) {
 
     private companion object {
         private val staticLogger = KotlinLogging.logger { }
     }
 
     override val logger: KLogger = staticLogger
 
-    /**
-     * Executes a moderation action on the given prompt using the specified language model.
-     * This method is not supported by the DeepSeek API and will always throw an `UnsupportedOperationException`.
-     *
-     * @param prompt The [Prompt] object to be moderated, containing the messages and respective context.
-     * @param model The [LLModel] to be used for the moderation process.
-     * @return This method does not return a valid result as it always throws an exception.
-     * @throws UnsupportedOperationException Always thrown because moderation is not supported by the DeepSeek API.
-     */
+    override fun serializeProviderChatRequest(
+        messages: List<OpenAIMessage>,
+        model: LLModel,
+        tools: List<OpenAITool>?,
+        toolChoice: OpenAIToolChoice?,
+        params: LLMParams,
+        stream: Boolean
+    ): String {
+        val deepSeekParams = params.toDeepSeekParams()
+        val responseFormat = createResponseFormat(params.schema, model)
+
+        val request = DeepSeekChatCompletionRequest(
+            messages = messages,
+            model = model.id,
+            frequencyPenalty = deepSeekParams.frequencyPenalty,
+            logprobs = deepSeekParams.logprobs,
+            maxTokens = deepSeekParams.maxTokens,
+            presencePenalty = deepSeekParams.presencePenalty,
+            responseFormat = responseFormat,
+            stop = deepSeekParams.stop,
+            stream = stream,
+            temperature = deepSeekParams.temperature,
+            toolChoice = deepSeekParams.toolChoice?.toOpenAIToolChoice(),
+            tools = tools,
+            topLogprobs = deepSeekParams.topLogprobs,
+            topP = deepSeekParams.topP,
+        )
+
+        return json.encodeToString(request)
+    }
+
+    override fun processProviderChatResponse(response: DeepSeekChatCompletionResponse): List<LLMChoice> {
+        require(response.choices.isNotEmpty()) { "Empty choices in response" }
+        return response.choices.map { it.toMessageResponses(createMetaInfo(response.usage)) }
+    }
+
+    override fun decodeStreamingResponse(data: String): DeepSeekChatCompletionStreamResponse =
+        json.decodeFromString(data)
+
+    override fun decodeResponse(data: String): DeepSeekChatCompletionResponse =
+        json.decodeFromString(data)
+
+    override fun processStreamingChunk(chunk: DeepSeekChatCompletionStreamResponse): String? =
+        chunk.choices.firstOrNull()?.delta?.content
+
     public override suspend fun moderate(prompt: Prompt, model: LLModel): ModerationResult {
         logger.warn { "Moderation is not supported by DeepSeek API" }
         throw UnsupportedOperationException("Moderation is not supported by DeepSeek API.")
     }

prompt/prompt-executor/prompt-executor-clients/prompt-executor-deepseek-client/src/commonMain/kotlin/ai/koog/prompt/executor/clients/deepseek/DeepSeekParams.kt

@@ -0,0 +1,174 @@
+package ai.koog.prompt.executor.clients.deepseek
+
+import ai.koog.prompt.params.LLMParams
+
+internal fun LLMParams.toDeepSeekParams(): DeepSeekParams {
+    if (this is DeepSeekParams) return this
+    return DeepSeekParams(
+        temperature = temperature,
+        maxTokens = maxTokens,
+        numberOfChoices = numberOfChoices,
+        speculation = speculation,
+        schema = schema,
+        toolChoice = toolChoice,
+        user = user,
+        includeThoughts = includeThoughts,
+    )
+}
+
+/**
+ * DeepSeek chat-completions parameters layered on top of [LLMParams].
+ *
+ * @property temperature Sampling temperature in [0.0, 2.0]. Higher ⇒ more random;
+ *   lower ⇒ more deterministic. Adjust this **or** [topP], not both.
+ * @property maxTokens Maximum number of tokens the model may generate for this response.
+ * @property numberOfChoices Number of completions to generate for the prompt (cost scales with N).
+ * @property speculation Provider-specific control for speculative decoding / draft acceleration.
+ * @property schema JSON Schema to constrain model output (validated when supported).
+ * @property toolChoice Controls if/which tool must be called (`none`/`auto`/`required`/specific).
+ * @property user not used for DeepSeek
+ * @property includeThoughts Request inclusion of model “thoughts”/reasoning traces (model-dependent).
+ * @property thinkingBudget Soft cap on tokens spent on internal reasoning (reasoning models).
+ * @property frequencyPenalty Number in [-2.0, 2.0]—penalizes frequent tokens to reduce repetition.
+ * @property presencePenalty Number in [-2.0, 2.0]—encourages introduction of new tokens/topics.
+ * @property logprobs Whether to include log-probabilities for output tokens.
+ * @property stop Stop sequences (0–4 items); generation halts before any of these.
+ * @property topLogprobs Number of top alternatives per position (0–20). Requires [logprobs] = true.
+ * @property topP Nucleus sampling in (0.0, 1.0]; use **instead of** [temperature].
+ */
+public class DeepSeekParams(
+    temperature: Double? = null,
+    maxTokens: Int? = null,
+    numberOfChoices: Int? = null,
+    speculation: String? = null,
+    schema: Schema? = null,
+    toolChoice: ToolChoice? = null,
+    user: String? = null,
+    includeThoughts: Boolean? = null,
+    thinkingBudget: Int? = null,
+    public val frequencyPenalty: Double? = null,
+    public val presencePenalty: Double? = null,
+    public val logprobs: Boolean? = null,
+    public val stop: List<String>? = null,
+    public val topLogprobs: Int? = null,
+    public val topP: Double? = null,
+) : LLMParams(
+    temperature, maxTokens, numberOfChoices,
+    speculation, schema, toolChoice,
+    user, includeThoughts, thinkingBudget
+) {
+    init {
+        require(topP == null || topP in 0.0..1.0) {
+            "topP must be in (0.0, 1.0], but was $topP"
+        }
+        if (topLogprobs != null) {
+            require(logprobs == true) {
+                "`topLogprobs` requires `logprobs=true`."
+            }
+            require(topLogprobs in 0..20) {
+                "topLogprobs must be in [0, 20], but was $topLogprobs"
+            }
+        }
+        require(frequencyPenalty == null || frequencyPenalty in -2.0..2.0) {
+            "frequencyPenalty must be in [-2.0, 2.0], but was $frequencyPenalty"
+        }
+        require(presencePenalty == null || presencePenalty in -2.0..2.0) {
+            "presencePenalty must be in [-2.0, 2.0], but was $presencePenalty"
+        }
+
+        // --- Stop sequences ---
+        if (stop != null) {
+            require(stop.isNotEmpty()) { "stop must not be empty when provided." }
+            require(stop.size <= 4) { "stop supports at most 4 sequences, but was ${stop.size}" }
+            require(stop.all { it.isNotBlank() }) { "stop sequences must not be blank." }
+        }
+    }
+
+    /**
+     * Creates a copy of this instance with the ability to modify any of its properties.
+     */
+    public fun copy(
+        temperature: Double? = this.temperature,
+        maxTokens: Int? = this.maxTokens,
+        numberOfChoices: Int? = this.numberOfChoices,
+        speculation: String? = this.speculation,
+        schema: Schema? = this.schema,
+        toolChoice: ToolChoice? = this.toolChoice,
+        user: String? = this.user,
+        includeThoughts: Boolean? = this.includeThoughts,
+        thinkingBudget: Int? = this.thinkingBudget,
+        frequencyPenalty: Double? = this.frequencyPenalty,
+        presencePenalty: Double? = this.presencePenalty,
+        logprobs: Boolean? = this.logprobs,
+        stop: List<String>? = this.stop,
+        topLogprobs: Int? = this.topLogprobs,
+        topP: Double? = this.topP,
+    ): DeepSeekParams = DeepSeekParams(
+        temperature = temperature,
+        maxTokens = maxTokens,
+        numberOfChoices = numberOfChoices,
+        speculation = speculation,
+        schema = schema,
+        toolChoice = toolChoice,
+        user = user,
+        includeThoughts = includeThoughts,
+        thinkingBudget = thinkingBudget,
+        frequencyPenalty = frequencyPenalty,
+        presencePenalty = presencePenalty,
+        logprobs = logprobs,
+        stop = stop,
+        topLogprobs = topLogprobs,
+        topP = topP,
+    )
+
+    override fun equals(other: Any?): Boolean = when {
+        this === other -> true
+        other !is DeepSeekParams -> false
+        else ->
+            temperature == other.temperature &&
+                maxTokens == other.maxTokens &&
+                numberOfChoices == other.numberOfChoices &&
+                speculation == other.speculation &&
+                schema == other.schema &&
+                toolChoice == other.toolChoice &&
+                user == other.user &&
+                includeThoughts == other.includeThoughts &&
+                thinkingBudget == other.thinkingBudget &&
+                frequencyPenalty == other.frequencyPenalty &&
+                presencePenalty == other.presencePenalty &&
+                logprobs == other.logprobs &&
+                stop == other.stop &&
+                topLogprobs == other.topLogprobs &&
+                topP == other.topP
+    }
+
+    override fun hashCode(): Int = listOf(
+        temperature, maxTokens, numberOfChoices,
+        speculation, schema, toolChoice,
+        user, includeThoughts, thinkingBudget,
+        frequencyPenalty, presencePenalty,
+        logprobs, stop, topLogprobs, topP,
+    ).fold(0) { acc, element ->
+        31 * acc + (element?.hashCode() ?: 0)
+    }
+
+    override fun toString(): String = buildString {
+        append("DeepSeekParams(")
+        append("temperature=$temperature")
+        append(", maxTokens=$maxTokens")
+        append(", numberOfChoices=$numberOfChoices")
+        append(", speculation=$speculation")
+        append(", schema=$schema")
+        append(", toolChoice=$toolChoice")
+        append(", user=$user")
+        append(", includeThoughts=$includeThoughts")
+        append(", thinkingBudget=$thinkingBudget")
+        append(", frequencyPenalty=$frequencyPenalty")
+        append(", presencePenalty=$presencePenalty")
+        append(", logprobs=$logprobs")
+        append(", stop=$stop")
+        append(", topLogprobs=$topLogprobs")
+        append(", topP=$topP")
+        append(")")
+    }
+}