#1815 - Fix Gemini OpenAI-compatible guardrail IT sends gpt-5-mini instead of selected Gemini model

embabel-agent-anthropic/src/test/kotlin/com/embabel/agent/anthropic/AnthropicModelBindingTest.kt

@@ -0,0 +1,81 @@
+/*
+ * Copyright 2024-2026 Embabel Pty Ltd.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package com.embabel.agent.anthropic
+
+import com.embabel.agent.api.models.AnthropicModels
+import com.embabel.agent.spi.support.springai.SpringAiLlmService
+import com.embabel.chat.UserMessage
+import com.embabel.common.ai.model.LlmOptions
+import io.mockk.every
+import io.mockk.mockk
+import io.mockk.slot
+import org.assertj.core.api.Assertions.assertThat
+import org.junit.jupiter.api.Test
+import org.springframework.ai.anthropic.AnthropicChatOptions
+import org.springframework.ai.chat.messages.AssistantMessage as SpringAiAssistantMessage
+import org.springframework.ai.chat.metadata.ChatResponseMetadata
+import org.springframework.ai.chat.model.ChatModel
+import org.springframework.ai.chat.model.ChatResponse
+import org.springframework.ai.chat.model.Generation
+import org.springframework.ai.chat.prompt.Prompt
+
+/**
+ * End-to-end regression for the Anthropic half of the Spring AI 2.0 model-binding bug
+ * (see #1815 for the OpenAI/Gemini equivalent).
+ *
+ * [AnthropicOptionsConverter] never sets a model, so its [AnthropicChatOptions] carry the
+ * SDK's baked-in `DEFAULT_MODEL` (`claude-haiku-4-5`). Because Embabel always passes these
+ * per-request options and Spring AI 2.0 no longer merges the model bean's configured model,
+ * every call would silently target Haiku. [SpringAiLlmService] now binds the configured
+ * model generically; this test proves the selected model reaches the wire while the concrete
+ * [AnthropicChatOptions] type and other fields are preserved.
+ */
+class AnthropicModelBindingTest {
+
+    @Test
+    fun `sends the selected Anthropic model rather than the converter default`() {
+        // Prepare: bean configured with Sonnet; converter (unchanged) bakes in Haiku default.
+        val configuredOptions = AnthropicChatOptions.builder().model("claude-sonnet-4-5").build()
+        val capturedPrompt = slot<Prompt>()
+        val generation = Generation(SpringAiAssistantMessage("done"))
+        val chatResponse = mockk<ChatResponse> {
+            every { result } returns generation
+            every { results } returns listOf(generation)
+            every { metadata } returns mockk<ChatResponseMetadata>(relaxed = true) { every { usage } returns null }
+        }
+        val chatModel = mockk<ChatModel> {
+            every { options } returns configuredOptions
+            every { call(capture(capturedPrompt)) } returns chatResponse
+        }
+        val service = SpringAiLlmService(
+            name = "claude-sonnet-4-5",
+            provider = AnthropicModels.PROVIDER,
+            chatModel = chatModel,
+            optionsConverter = AnthropicOptionsConverter,
+            thinkingSupported = true,
+        )
+
+        // Execute
+        service.createMessageSender(LlmOptions().withMaxTokens(500))
+            .call(messages = listOf(UserMessage("Hi")), tools = emptyList())
+
+        // Verify: concrete Anthropic options carry the selected model, maxTokens preserved.
+        val sentOptions = capturedPrompt.captured.options
+        assertThat(sentOptions).isInstanceOf(AnthropicChatOptions::class.java)
+        assertThat(sentOptions.model).isEqualTo("claude-sonnet-4-5")
+        assertThat(sentOptions.maxTokens).isEqualTo(500)
+    }
+}

embabel-agent-anthropic/src/test/kotlin/com/embabel/agent/anthropic/AnthropicServedModelBindingIT.kt

@@ -0,0 +1,78 @@
+/*
+ * Copyright 2024-2026 Embabel Pty Ltd.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package com.embabel.agent.anthropic
+
+import com.embabel.agent.api.models.AnthropicModels
+import com.embabel.chat.UserMessage
+import com.embabel.common.ai.model.LlmOptions
+import org.assertj.core.api.Assertions.assertThat
+import org.junit.jupiter.api.Test
+import org.junit.jupiter.api.condition.EnabledIfEnvironmentVariable
+import org.opentest4j.TestAbortedException
+
+/**
+ * Live regression for the Anthropic same-provider silent variant of #1815: select Sonnet and
+ * assert Anthropic reports a sonnet model back. If binding regressed, the request would run on
+ * the AnthropicChatOptions default ("claude-haiku-4-5") and the served model would be a haiku.
+ *
+ * Requires ANTHROPIC_API_KEY. Skipped (aborted) if the key lacks access to the chosen model.
+ */
+class AnthropicServedModelBindingIT {
+
+    @Test
+    @EnabledIfEnvironmentVariable(named = "ANTHROPIC_API_KEY", matches = ".+")
+    fun `real Anthropic serves the selected model, not the default`() {
+        // Prepare
+        val model = AnthropicModels.CLAUDE_SONNET_4_5
+        val llm = AnthropicModelFactory(apiKey = System.getenv("ANTHROPIC_API_KEY")).build(model)
+
+        // Execute
+        val response = try {
+            llm.createMessageSender(LlmOptions())
+                .call(listOf(UserMessage("Reply with exactly the word READY.")), emptyList())
+        } catch (ex: Exception) {
+            if (isModelAccessError(ex)) {
+                throw TestAbortedException("ANTHROPIC_API_KEY is set but lacks access to $model", ex)
+            }
+            throw ex
+        }
+
+        // Verify: Anthropic served the selected Sonnet, not the claude-haiku-4-5 default.
+        assertThat(response.textContent).isNotBlank()
+        assertThat(response.model)
+            .withFailMessage("Served model was '%s', expected a sonnet model", response.model)
+            .isNotNull()
+            .contains("sonnet")
+        assertThat(response.model).doesNotContain("haiku")
+    }
+
+    /** Broad check for a provider "no access to this model" error, so the IT aborts rather than fails. */
+    private fun isModelAccessError(ex: Throwable): Boolean {
+        val message = generateSequence<Throwable>(ex) { it.cause }
+            .mapNotNull { it.message }
+            .joinToString(" | ")
+            .lowercase()
+        return listOf(
+            "not_found",
+            "does not exist",
+            "not found",
+            "permission",
+            "invalid model",
+            "unsupported model",
+            "does not have access",
+        ).any { message.contains(it) }
+    }
+}

embabel-agent-api/src/main/kotlin/com/embabel/agent/spi/loop/LlmMessageSender.kt

@@ -68,11 +68,14 @@ data class LlmMessageRequest @JvmOverloads constructor(
  * @param message The full message object from the LLM
  * @param textContent The text content of the message
  * @param usage Optional usage information (tokens, etc.)
+ * @param model The model that actually served the request, as reported by the provider
+ *        (from the response metadata). Null/blank if the provider does not report it.
  */
 data class LlmMessageResponse(
     val message: Message,
     val textContent: String,
     val usage: Usage? = null,
+    val model: String? = null,
 )
 
 /**

embabel-agent-api/src/main/kotlin/com/embabel/agent/spi/loop/UnfoldingToolInjectionStrategy.kt

@@ -139,4 +139,3 @@ class UnfoldingToolInjectionStrategy : ToolInjectionStrategy {
         val INSTANCE = UnfoldingToolInjectionStrategy()
     }
 }
-

embabel-agent-api/src/main/kotlin/com/embabel/agent/spi/support/springai/SpringAiLlmMessageSender.kt

@@ -111,6 +111,7 @@
             message = embabelMessage,
             textContent = assistantMessage.text ?: "",
             usage = usage,
+            model = response.metadata?.model,
         )
     }
 

embabel-agent-api/src/main/kotlin/com/embabel/agent/spi/support/springai/SpringAiLlmService.kt

@@ -28,6 +28,7 @@
 import org.springframework.ai.chat.messages.UserMessage
 import org.springframework.ai.chat.model.ChatModel
 import org.springframework.ai.chat.model.ChatResponse
+import org.springframework.ai.chat.prompt.ChatOptions
 import org.springframework.ai.chat.prompt.Prompt
 import reactor.core.publisher.Flux
 import java.time.Duration
@@ -123,8 +124,22 @@
      */
     override val model: ChatModel get() = chatModel
 
+    /**
+     * Binds the model configured on the underlying [ChatModel] onto per-request options.
+     *
+     * Spring AI 2.0 no longer merges a model's configured options into a prompt that already
+     * carries options (OpenAiChatModel/AnthropicChatModel `buildRequestPrompt` returns the
+     * prompt unchanged when `getOptions() != null`). Because Embabel always supplies
+     * per-request options — and provider option types coerce a null model to a hard-coded
+     * default in their constructor (e.g. `gpt-5-mini`, `claude-haiku-4-5`) — the selected
+     * model would otherwise be ignored on every call. Binding the [ChatModel]'s own
+     * configured model restores the pre-2.0 merge semantics for every provider at once.
+     */
+    private fun bindConfiguredModel(chatOptions: ChatOptions): ChatOptions =
+        bindModel(chatOptions, chatModel.options?.model)
+
     override fun createMessageSender(options: LlmOptions): LlmMessageSender {
-        val chatOptions = optionsConverter.convertOptions(options)
+        val chatOptions = bindConfiguredModel(optionsConverter.convertOptions(options))
         return SpringAiLlmMessageSender(
             chatModel = chatModel,
             chatOptions = chatOptions,
@@ -136,7 +151,7 @@
     }
 
     override fun createMessageStreamer(options: LlmOptions): LlmMessageStreamer {
-        val chatOptions = optionsConverter.convertOptions(options)
+        val chatOptions = bindConfiguredModel(optionsConverter.convertOptions(options))
         val chatClient = ChatClient.create(chatModel)
         return SpringAiLlmMessageStreamer(chatClient, chatOptions)
     }
@@ -160,3 +175,13 @@
     fun withOptionsConverter(converter: OptionsConverter<*>): SpringAiLlmService =
         copy(optionsConverter = converter)
 }
+
+/**
+ * Binds [model] onto [options], preserving the concrete provider option type and all other
+ * fields via each provider's overridden `mutate()` (dynamic dispatch keeps e.g.
+ * `OpenAiChatOptions`/`AnthropicChatOptions` intact). Returns [options] unchanged when
+ * [model] is null or blank, so models that expose no configured default (bare test doubles,
+ * exotic providers) keep the converter's output.
+ */
+internal fun bindModel(options: ChatOptions, model: String?): ChatOptions =
+    if (model.isNullOrBlank()) options else options.mutate().model(model).build()

embabel-agent-api/src/test/kotlin/com/embabel/agent/spi/support/springai/SpringAiLlmMessageSenderTest.kt

@@ -51,7 +51,7 @@ class SpringAiLlmMessageSenderTest {
             val configuredOptions = testChatOptions()
             val capturedPrompt = slot<Prompt>()
             val generation = Generation(SpringAiAssistantMessage("done"))
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns null
             }
             val chatResponse = mockk<ChatResponse> {
@@ -106,7 +106,7 @@ class SpringAiLlmMessageSenderTest {
             val configuredOptions = testChatOptions()
             val capturedPrompt = slot<Prompt>()
             val generation = Generation(SpringAiAssistantMessage("done"))
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns null
             }
             val chatResponse = mockk<ChatResponse> {
@@ -157,7 +157,7 @@ class SpringAiLlmMessageSenderTest {
             val originalOptions = testChatOptions()
             val capturedPrompt = slot<Prompt>()
             val generation = Generation(SpringAiAssistantMessage("done"))
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns null
             }
             val chatResponse = mockk<ChatResponse> {
@@ -240,7 +240,7 @@ class SpringAiLlmMessageSenderTest {
                     .build()
             )
 
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns mockk(relaxed = true)
             }
             val chatResponse = mockk<ChatResponse> {
@@ -313,7 +313,7 @@ class SpringAiLlmMessageSenderTest {
                     .build()
             )
 
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns mockk(relaxed = true)
             }
             val chatResponse = mockk<ChatResponse> {
@@ -374,7 +374,7 @@ class SpringAiLlmMessageSenderTest {
                     .build()
             )
 
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns mockk(relaxed = true)
             }
             val chatResponse = mockk<ChatResponse> {
@@ -439,7 +439,7 @@ class SpringAiLlmMessageSenderTest {
                     .build()
             )
 
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns mockk(relaxed = true)
             }
             val chatResponse = mockk<ChatResponse> {
@@ -503,7 +503,7 @@ class SpringAiLlmMessageSenderTest {
                     .build()
             )
 
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns mockk(relaxed = true)
             }
             val chatResponse = mockk<ChatResponse> {
@@ -581,7 +581,7 @@ class SpringAiLlmMessageSenderTest {
                     .build()
             )
 
-            val mockMetadata = mockk<ChatResponseMetadata> {
+            val mockMetadata = mockk<ChatResponseMetadata>(relaxed = true) {
                 every { usage } returns null
             }
             val chatResponse = mockk<ChatResponse> {