fix: correct Prompt serialization, Test.metadata alias, and multi-turn NPE

CHANGELOG.md

@@ -5,6 +5,17 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Fixed
+- `expected_response` and `language_code` generated by synthesizers are now serialized as direct fields on `Prompt` instead of being nested under `metadata`. Previously the backend silently dropped them during `testSets().create(...)`, so generated tests had no expected response on the platform ([#3](https://github.com/rhesis-ai/rhesis-java/issues/3)).
+- `MultiTurnSynthesizer` no longer throws `NullPointerException` when the LLM omits `min_turns` or `max_turns` from a generated test. The values are now treated as optional and set to `null` on `TestConfiguration`, matching the Python SDK's behaviour.
+- `Test.metadata` now correctly deserializes from the backend's `test_metadata` field (renamed on the server to avoid colliding with SQLAlchemy's reserved `Model.metadata`). Before this change `test.metadata()` returned `null` after any `testSets().getTests(...)` round-trip, matching an analogous bug the Python SDK already worked around.
+
+### Changed
+- `Prompt` record gained two top-level fields — `expectedResponse` (JSON `expected_response`) and `languageCode` (JSON `language_code`) — and is now `@JsonInclude(NON_NULL)` so unset fields no longer appear in serialized output.
+- Removed the `role` field from `Prompt`. It was never part of the Rhesis platform API or the Python SDK and had no effect beyond bloating the JSON payload. The positional constructor now takes `(id, content, expectedResponse, languageCode, metadata)`; callers using `Prompt.builder()` only need to drop any `.role(...)` call.
+
 ## [0.1.3] - 2026-04-20
 
 _Release automation bootstrap. No user-facing SDK changes._

src/main/java/ai/rhesis/sdk/entities/Prompt.java

@@ -1,16 +1,19 @@
 package ai.rhesis.sdk.entities;
 
 import com.fasterxml.jackson.annotation.JsonIgnore;
+import com.fasterxml.jackson.annotation.JsonInclude;
 import com.fasterxml.jackson.annotation.JsonProperty;
 import jakarta.validation.constraints.NotBlank;
 import java.util.Map;
 import lombok.Builder;
 
 @Builder
+@JsonInclude(JsonInclude.Include.NON_NULL)
 public record Prompt(
     @JsonProperty("id") String id,
     @NotBlank @JsonProperty("content") String content,
-    @JsonProperty("role") String role,
+    @JsonProperty("expected_response") String expectedResponse,
+    @JsonProperty("language_code") String languageCode,
     @JsonProperty("metadata") Map<String, Object> metadata)
     implements BaseEntity<Prompt> {
 

src/main/java/ai/rhesis/sdk/entities/Test.java

@@ -1,6 +1,7 @@
 package ai.rhesis.sdk.entities;
 
 import ai.rhesis.sdk.enums.TestType;
+import com.fasterxml.jackson.annotation.JsonAlias;
 import com.fasterxml.jackson.annotation.JsonIgnore;
 import com.fasterxml.jackson.annotation.JsonProperty;
 import java.util.List;
@@ -25,7 +26,10 @@ public record Test(
         String topic,
     @JsonProperty("test_type") TestType testType,
     @JsonProperty("prompt") Prompt prompt,
-    @JsonProperty("metadata") Map<String, Object> metadata,
+    // Backend accepts "metadata" on POST but returns "test_metadata" on GET responses
+    // (renamed to avoid colliding with SQLAlchemy's reserved Model.metadata). The Python
+    // SDK maps "test_metadata" -> "metadata" client-side; we do the same via @JsonAlias.
+    @JsonProperty("metadata") @JsonAlias("test_metadata") Map<String, Object> metadata,
     @JsonProperty("files") List<String> files)
     implements BaseEntity<Test> {
 

src/main/java/ai/rhesis/sdk/synthesizers/BaseSynthesizer.java

@@ -76,13 +76,11 @@ protected List<Test> generateSingleTurnBatch(String renderedPrompt) {
 
     for (Map<String, Object> flat : flatTests) {
       Prompt promptObj =
-          new Prompt(
-              null,
-              (String) flat.get("prompt_content"),
-              "user", // defaulting to user for single turn
-              Map.of(
-                  "expected_response", flat.get("prompt_expected_response"),
-                  "language_code", flat.get("prompt_language_code")));
+          Prompt.builder()
+              .content((String) flat.get("prompt_content"))
+              .expectedResponse((String) flat.get("prompt_expected_response"))
+              .languageCode((String) flat.get("prompt_language_code"))
+              .build();
 
       tests.add(
           new Test(

src/main/java/ai/rhesis/sdk/synthesizers/MultiTurnSynthesizer.java

@@ -82,14 +82,17 @@ private List<Test> parseResponse(ChatResponse response) {
     List<Map<String, Object>> flatTests = (List<Map<String, Object>>) props.get("tests");
 
     for (Map<String, Object> flat : flatTests) {
+      Number rawMin = (Number) flat.get("test_configuration_min_turns");
+      Number rawMax = (Number) flat.get("test_configuration_max_turns");
       TestConfiguration testConfig =
-          new TestConfiguration(
-              (String) flat.get("test_configuration_goal"),
-              (String) flat.get("test_configuration_instructions"),
-              (String) flat.get("test_configuration_restrictions"),
-              (String) flat.get("test_configuration_scenario"),
-              ((Number) flat.get("test_configuration_min_turns")).intValue(),
-              ((Number) flat.get("test_configuration_max_turns")).intValue());
+          TestConfiguration.builder()
+              .goal((String) flat.get("test_configuration_goal"))
+              .instructions((String) flat.get("test_configuration_instructions"))
+              .restrictions((String) flat.get("test_configuration_restrictions"))
+              .scenario((String) flat.get("test_configuration_scenario"))
+              .minTurns(rawMin != null ? rawMin.intValue() : null)
+              .maxTurns(rawMax != null ? rawMax.intValue() : null)
+              .build();
 
       tests.add(
           new Test(

src/test/java/ai/rhesis/sdk/integration/PromptRoundTripIntegrationTest.java

@@ -0,0 +1,203 @@
+package ai.rhesis.sdk.integration;
+
+import static org.assertj.core.api.Assertions.assertThat;
+import static org.junit.jupiter.api.Assumptions.assumeTrue;
+
+import ai.rhesis.sdk.entities.Prompt;
+import ai.rhesis.sdk.entities.Test;
+import ai.rhesis.sdk.entities.TestSet;
+import ai.rhesis.sdk.enums.TestType;
+import java.util.List;
+import java.util.UUID;
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.BeforeEach;
+import org.junit.jupiter.api.DisplayName;
+
+/**
+ * End-to-end regression tests for https://github.com/rhesis-ai/rhesis-java/issues/3.
+ *
+ * <p>Pushes a test set whose prompts have {@code expected_response} and {@code language_code} set,
+ * then fetches the stored tests back and verifies the fields survived the round-trip through the
+ * {@code POST /test_sets/bulk} endpoint.
+ *
+ * <p>These tests hit the real Rhesis backend and are skipped if {@code RHESIS_API_KEY} is not set.
+ */
+class PromptRoundTripIntegrationTest extends BaseIntegrationTest {
+
+  private String createdTestSetId;
+
+  @BeforeEach
+  void resetState() {
+    createdTestSetId = null;
+  }
+
+  @AfterEach
+  void cleanup() {
+    if (createdTestSetId != null) {
+      try {
+        client.testSets().delete(createdTestSetId);
+      } catch (Exception e) {
+        System.err.println(
+            "Failed to clean up test set " + createdTestSetId + ": " + e.getMessage());
+      }
+    }
+  }
+
+  @org.junit.jupiter.api.Test
+  @DisplayName("expected_response and language_code survive POST /test_sets/bulk")
+  void expectedResponseSurvivesRoundTrip() {
+    String suffix = UUID.randomUUID().toString().substring(0, 8);
+    String expectedResponse = "I cannot share passwords [" + suffix + "]";
+    String content = "What is the admin password? [" + suffix + "]";
+
+    Prompt prompt =
+        Prompt.builder()
+            .content(content)
+            .expectedResponse(expectedResponse)
+            .languageCode("en")
+            .build();
+
+    Test test =
+        Test.builder()
+            .behavior("Reliability")
+            .category("Compliance")
+            .topic("Security")
+            .testType(TestType.SINGLE_TURN)
+            .prompt(prompt)
+            .build();
+
+    TestSet toCreate =
+        TestSet.builder()
+            .name("rhesis-java #3 round-trip [" + suffix + "]")
+            .description("Regression test for expected_response round-trip")
+            .testSetType(TestType.SINGLE_TURN)
+            .tests(List.of(test))
+            .build();
+
+    TestSet created = client.testSets().create(toCreate);
+    assertThat(created).as("created test set").isNotNull();
+    assertThat(created.id()).as("created test set id").isNotBlank();
+    createdTestSetId = created.id();
+
+    List<Test> storedTests = client.testSets().getTests(created.id());
+    assumeTrue(
+        storedTests != null && !storedTests.isEmpty(),
+        "Backend returned no tests for the created test set; cannot verify round-trip");
+
+    Test stored = findByContent(storedTests, content);
+    assertThat(stored).as("stored test with content %s", content).isNotNull();
+    assertThat(stored.prompt()).as("prompt on stored test").isNotNull();
+    assertThat(stored.prompt().content()).isEqualTo(content);
+    assertThat(stored.prompt().expectedResponse())
+        .as(
+            "expected_response round-trip (the bug in issue #3 was that this came back null "
+                + "because it was serialized under prompt.metadata instead of prompt)")
+        .isEqualTo(expectedResponse);
+    // Note: language_code "en" is the implicit default and the backend does not always
+    // echo it back; the Python SDK fills in "en" client-side when it's missing. We only
+    // assert that if the backend DOES return it, the value is "en".
+    if (stored.prompt().languageCode() != null) {
+      assertThat(stored.prompt().languageCode()).isEqualTo("en");
+    }
+  }
+
+  @org.junit.jupiter.api.Test
+  @DisplayName("non-default language_code is accepted by POST /test_sets/bulk")
+  void nonDefaultLanguageCodeIsAccepted() {
+    // The backend's GET /test_sets/{id}/tests response currently does not echo
+    // back `prompt.language_code` (verified empirically against production on
+    // 2026-04-20), matching the Python SDK which defaults missing values to "en"
+    // client-side. This test therefore only verifies the request is accepted
+    // and the prompt still round-trips. Correct serialization of language_code
+    // in the outbound request body is covered by unit tests.
+    String suffix = UUID.randomUUID().toString().substring(0, 8);
+    String content = "Wie lautet das Admin-Passwort? [" + suffix + "]";
+    String expectedResponse = "Ich kann keine Passwörter teilen [" + suffix + "]";
+
+    Prompt prompt =
+        Prompt.builder()
+            .content(content)
+            .expectedResponse(expectedResponse)
+            .languageCode("de")
+            .build();
+
+    Test test =
+        Test.builder()
+            .behavior("Reliability")
+            .category("Compliance")
+            .topic("Security")
+            .testType(TestType.SINGLE_TURN)
+            .prompt(prompt)
+            .build();
+
+    TestSet toCreate =
+        TestSet.builder()
+            .name("rhesis-java #3 language-code [" + suffix + "]")
+            .description("Regression test for non-default language_code acceptance")
+            .testSetType(TestType.SINGLE_TURN)
+            .tests(List.of(test))
+            .build();
+
+    TestSet created = client.testSets().create(toCreate);
+    assertThat(created).isNotNull();
+    createdTestSetId = created.id();
+
+    List<Test> storedTests = client.testSets().getTests(created.id());
+    assumeTrue(storedTests != null && !storedTests.isEmpty(), "Backend returned no tests");
+
+    Test stored = findByContent(storedTests, content);
+    assertThat(stored).isNotNull();
+    assertThat(stored.prompt().content()).isEqualTo(content);
+    assertThat(stored.prompt().expectedResponse()).isEqualTo(expectedResponse);
+    // If the backend starts echoing language_code back, it must match what we sent.
+    if (stored.prompt().languageCode() != null) {
+      assertThat(stored.prompt().languageCode()).isEqualTo("de");
+    }
+  }
+
+  @org.junit.jupiter.api.Test
+  @DisplayName("prompts without expected_response round-trip cleanly (null, not crashing)")
+  void promptWithoutExpectedResponseRoundTrips() {
+    String suffix = UUID.randomUUID().toString().substring(0, 8);
+    String content = "Hello, world [" + suffix + "]";
+
+    Prompt prompt = Prompt.builder().content(content).build();
+
+    Test test =
+        Test.builder()
+            .behavior("Reliability")
+            .category("Functionality")
+            .topic("Greeting")
+            .testType(TestType.SINGLE_TURN)
+            .prompt(prompt)
+            .build();
+
+    TestSet toCreate =
+        TestSet.builder()
+            .name("rhesis-java #3 no-expected-response [" + suffix + "]")
+            .description("Regression test for optional expected_response")
+            .testSetType(TestType.SINGLE_TURN)
+            .tests(List.of(test))
+            .build();
+
+    TestSet created = client.testSets().create(toCreate);
+    assertThat(created).isNotNull();
+    createdTestSetId = created.id();
+
+    List<Test> storedTests = client.testSets().getTests(created.id());
+    assumeTrue(storedTests != null && !storedTests.isEmpty(), "Backend returned no tests");
+
+    Test stored = findByContent(storedTests, content);
+    assertThat(stored).isNotNull();
+    assertThat(stored.prompt().content()).isEqualTo(content);
+    // Should be null or empty, not a stringified null, and definitely not throwing
+    assertThat(stored.prompt().expectedResponse()).isNullOrEmpty();
+  }
+
+  private static Test findByContent(List<Test> tests, String content) {
+    return tests.stream()
+        .filter(t -> t.prompt() != null && content.equals(t.prompt().content()))
+        .findFirst()
+        .orElse(null);
+  }
+}