bump openapi spec to latest

braintrust-api/build.gradle

@@ -18,7 +18,7 @@
 
 plugins {
     id 'java'
-    id 'org.openapi.generator' version '7.14.0'
+    id 'org.openapi.generator' version '7.27.0'
 }
 
 java {
@@ -70,11 +70,16 @@ openApiGenerate {
     // Custom templates under src/generator/ (not src/main/resources, so they don't
     // end up on the runtime classpath or in the published JAR — they're only needed
     // at code-generation time). See src/generator/templates for the diffs from upstream.
-    templateDir = "${projectDir}/src/generator/templates/Java"
+    templateDir = layout.projectDirectory.dir('src/generator/templates/Java')
 
-    inputSpec = normalizedSpecFile.map { it.asFile.absolutePath }
+    inputSpec = normalizedSpecFile
 
-    outputDir = generatedSourcesDir.map { it.asFile.absolutePath }
+    outputDir = generatedSourcesDir
+    // Remove models and endpoints no longer present in the pinned spec.
+    cleanupOutput = true
+
+    // Adapt OpenAPI 3.1 schema types to the generator's internal representation.
+    openapiNormalizer = [NORMALIZE_31SPEC: 'true']
 
     apiPackage = 'dev.braintrust.openapi.api'
     modelPackage = 'dev.braintrust.openapi.model'
@@ -144,7 +149,7 @@ openApiGenerate {
 // referenced schema, so a genuinely-narrowing allOf is left alone.
 
 tasks.register('normalizeOpenApiSpec') {
-    description = 'Collapses redundant allOf-over-$ref-plus-composition nodes before generation.'
+    description = 'Normalizes license metadata and redundant allOf compositions before generation.'
     group = 'Build'
 
     def rawSpec = openApiSpecFile
@@ -156,6 +161,15 @@ tasks.register('normalizeOpenApiSpec') {
         def yaml = new org.yaml.snakeyaml.Yaml()
         def spec = yaml.load(rawSpec.get().asFile.text)
 
+        // Swagger Parser requires identifier or url for OAS 3.1 licenses, although both are
+        // optional. Supply the SPDX identifier for the spec's known license without replacing
+        // upstream metadata or setting identifier alongside its mutually exclusive url field.
+        def license = spec?.info?.license
+        if (spec?.openapi?.startsWith('3.1.') && license?.name == 'Apache 2.0'
+                && !license.containsKey('identifier') && !license.containsKey('url')) {
+            license.identifier = 'Apache-2.0'
+        }
+
         // The branches of a component schema's own anyOf/oneOf, or null if it isn't a composition.
         def declaredBranches = { String refName ->
             def target = spec?.components?.schemas?.get(refName)

braintrust-api/src/generator/templates/PATCHES.md

@@ -1,7 +1,8 @@
 # Custom Mustache Template Patches
 
-These templates are forked from **openapi-generator v7.14.0** and contain targeted
-patches to fix bugs or add functionality not present upstream. They live under
+These templates contain targeted patches to fix bugs or add functionality not present
+upstream. Their base versions are recorded below; both were compared against v7.27.0.
+The generator runs v7.27.0 with Jackson 2. They live under
 `src/generator/` (not `src/main/resources`) so they are only available at
 code-generation time and do not appear on the runtime classpath or in the published JAR.
 
@@ -15,6 +16,10 @@ When upgrading the generator version, diff each file against the new upstream te
 
 **Upstream:** `Java/libraries/native/anyof_model.mustache` @ v7.14.0
 
+Retained for the explicit variant API below. Upstream v7.27.0 includes container
+`TypeReference` deserialization, but not our `SchemaType`, explicit constructor, or
+named variant getters. Its Jackson 3 branches are not needed by this build.
+
 ### Patch 1 — TypeReference for container variants in deserializer
 The upstream deserializer calls `readValueAs(List.class)` / `readValueAs(Map.class)` for
 container-typed anyOf variants, using the raw erased type. This causes silent
@@ -64,7 +69,7 @@ parameter names or code that does not compile.
 
 ## api.mustache
 
-**Upstream:** `Java/libraries/native/api.mustache` @ v7.14.0
+**Upstream:** `Java/libraries/native/api.mustache` @ v7.27.0
 
 ### Patch 1 — anyOf/oneOf model query parameters are serialized via `parameterToPairs`
 In the `isExplode`/`!hasVars`/`isModel` branch, the upstream template serializes the parameter

braintrust-api/src/test/java/dev/braintrust/openapi/model/AnyOfModelTemplateTest.java

@@ -189,9 +189,8 @@ void projectScoreCategories_getVariantType_weighted_viaConstructor() {
 
     @Test
     void projectScoreCategories_getVariantType_minimum_viaConstructor() {
-        // Explicitly declare the List<String> variant as Minimum — this is the key fix.
-        // Without the (SchemaType, Object) constructor, getVariantType() would incorrectly
-        // return Categorical for both List variants due to erasure.
+        // Explicit variants distinguish List<String> from List<ProjectScoreCategory>
+        // despite their identical raw type after erasure.
         ProjectScoreCategories c =
                 new ProjectScoreCategories(
                         ProjectScoreCategories.SchemaType.Minimum, List.of("pass", "fail"));
@@ -200,9 +199,7 @@ void projectScoreCategories_getVariantType_minimum_viaConstructor() {
 
     @Test
     void projectScoreCategories_getVariantType_viaDeserialization_minimum() throws Exception {
-        // Uses TypeReference<List<String>> so Jackson fails fast on object elements,
-        // correctly distinguishing Minimum (List<String>) from Categorical
-        // (List<ProjectScoreCategory>)
+        // TypeReference distinguishes string elements from ProjectScoreCategory objects.
         ProjectScoreCategories c =
                 mapper.readValue("[\"pass\",\"fail\"]", ProjectScoreCategories.class);
         assertEquals(ProjectScoreCategories.SchemaType.Minimum, c.getVariantType());
@@ -250,15 +247,4 @@ void projectScoreCategories_typedGetter_weighted() {
                 new ProjectScoreCategories(ProjectScoreCategories.SchemaType.Weighted, weights);
         assertEquals(weights, c.getWeightedInstance());
     }
-
-    @Test
-    void projectScoreCategories_schemaType_dataType_field() {
-        // Each enum constant carries the full generic type as .dataType
-        assertEquals(
-                "List<ProjectScoreCategory>",
-                ProjectScoreCategories.SchemaType.Categorical.dataType);
-        assertEquals(
-                "Map<String, BigDecimal>", ProjectScoreCategories.SchemaType.Weighted.dataType);
-        assertEquals("List<String>", ProjectScoreCategories.SchemaType.Minimum.dataType);
-    }
 }

braintrust-sdk/instrumentation/anthropic_2_2_0/src/test/java/dev/braintrust/instrumentation/anthropic/v2_2_0/BraintrustAnthropicPromptCachingTest.java

@@ -25,23 +25,25 @@
  * Tests that prompt caching metrics (per-TTL breakdown) are correctly extracted from Anthropic API
  * responses and attached to spans.
  *
- * <p>Uses a cache-buster nonce in the system prompt to guarantee cache misses, ensuring {@code
- * cache_creation_input_tokens} is always positive. In VCR replay mode the nonce is a fixed string
- * so cassette matching still works.
+ * <p>Uses a random nonce for live requests and a client-specific stable nonce for recording and
+ * replay. Recording requires a cold cache; repeated recordings must wait for the cache TTLs.
  */
 public class BraintrustAnthropicPromptCachingTest {
     private static final String TEST_MODEL = "claude-sonnet-4-5-20250929";
     private static final ObjectMapper JSON_MAPPER = new ObjectMapper();
 
     /**
-     * Nonce injected into system prompts to bust Anthropic's server-side prompt cache. Random UUID
-     * when running live ({@code VCR_MODE=off}), fixed string otherwise so VCR cassettes match.
+     * Nonce injected into system prompts to isolate each client's Anthropic prompt cache. Random
+     * UUID in {@code VCR_MODE=off}, otherwise the versioned test class name for cassette matching.
      */
     private static final String VCR_NONCE;
 
     static {
         String vcrMode = System.getenv().getOrDefault("VCR_MODE", "replay").toUpperCase();
-        VCR_NONCE = "OFF".equals(vcrMode) ? UUID.randomUUID().toString() : "vcr-mode";
+        VCR_NONCE =
+                "OFF".equals(vcrMode)
+                        ? UUID.randomUUID().toString()
+                        : BraintrustAnthropicPromptCachingTest.class.getName();
     }
 
     @BeforeAll

braintrust-sdk/instrumentation/anthropic_3_0_0/src/test/java/dev/braintrust/instrumentation/anthropic/v3_0_0/BraintrustAnthropicPromptCachingTest.java

@@ -25,23 +25,25 @@
  * Tests that prompt caching metrics (per-TTL breakdown) are correctly extracted from Anthropic API
  * responses and attached to spans.
  *
- * <p>Uses a cache-buster nonce in the system prompt to guarantee cache misses, ensuring {@code
- * cache_creation_input_tokens} is always positive. In VCR replay mode the nonce is a fixed string
- * so cassette matching still works.
+ * <p>Uses a random nonce for live requests and a client-specific stable nonce for recording and
+ * replay. Recording requires a cold cache; repeated recordings must wait for the cache TTLs.
  */
 public class BraintrustAnthropicPromptCachingTest {
     private static final String TEST_MODEL = "claude-sonnet-4-5-20250929";
     private static final ObjectMapper JSON_MAPPER = new ObjectMapper();
 
     /**
-     * Nonce injected into system prompts to bust Anthropic's server-side prompt cache. Random UUID
-     * when running live ({@code VCR_MODE=off}), fixed string otherwise so VCR cassettes match.
+     * Nonce injected into system prompts to isolate each client's Anthropic prompt cache. Random
+     * UUID in {@code VCR_MODE=off}, otherwise the versioned test class name for cassette matching.
      */
     private static final String VCR_NONCE;
 
     static {
         String vcrMode = System.getenv().getOrDefault("VCR_MODE", "replay").toUpperCase();
-        VCR_NONCE = "OFF".equals(vcrMode) ? UUID.randomUUID().toString() : "vcr-mode";
+        VCR_NONCE =
+                "OFF".equals(vcrMode)
+                        ? UUID.randomUUID().toString()
+                        : BraintrustAnthropicPromptCachingTest.class.getName();
     }
 
     @BeforeAll

braintrust-sdk/instrumentation/openai_2_15_0/src/test/java/dev/braintrust/instrumentation/openai/v2_15_0/BraintrustOpenAITest.java

@@ -127,7 +127,7 @@ void testImageGeneration() {
                     client.images()
                             .generate(
                                     ImageGenerateParams.builder()
-                                            .model(ImageModel.GPT_IMAGE_1)
+                                            .model(ImageModel.of("gpt-image-2.5-flare"))
                                             .prompt(
                                                     "A small blue circle on a plain white"
                                                             + " background.")
@@ -148,7 +148,7 @@ void testImageGeneration() {
         } finally {
             parent.end();
         }
-        assertSingleMediaSpan(parent, "images/generations", "gpt-image-1");
+        assertSingleMediaSpan(parent, "images/generations", "gpt-image-2.5-flare");
     }
 
     @Test
@@ -491,7 +491,7 @@ void testResponses() {
 
         var request =
                 ResponseCreateParams.builder()
-                        .model("o4-mini")
+                        .model("gpt-6-luna")
                         .reasoning(
                                 Reasoning.builder()
                                         .effort(ReasoningEffort.LOW)

braintrust-sdk/src/main/java/dev/braintrust/eval/DatasetBrainstoreImpl.java

@@ -193,15 +193,15 @@ public Optional<DatasetCase<INPUT, OUTPUT>> next() {
             OUTPUT expected = outputConverter.apply(event.getExpected());
 
             var metadataObj = event.getMetadata();
-            // InsertProjectLogsEventMetadata extends HashMap<String,Object>. Jackson stores
-            // unknown fields in the HashMap base (not in additionalProperties) for Map subclasses,
-            // so copy from both sources defensively.
             Map<String, Object> metadata;
             if (metadataObj != null) {
-                metadata = new HashMap<>(metadataObj);
+                metadata = new HashMap<>();
                 if (metadataObj.getAdditionalProperties() != null) {
                     metadata.putAll(metadataObj.getAdditionalProperties());
                 }
+                if (metadataObj.getModel_JsonNullable().isPresent()) {
+                    metadata.put("model", metadataObj.getModel());
+                }
             } else {
                 metadata = Map.of();
             }

braintrust-sdk/src/test/java/dev/braintrust/eval/DatasetBrainstoreImplTest.java

@@ -242,6 +242,7 @@ void testMetadataPopulatedFromDatasetRow() {
                                       "input": "test input",
                                       "expected": "test output",
                                       "metadata": {
+                                        "model": "gpt-4o",
                                         "performance_strategy": "TURBO",
                                         "userId": "user123"
                                       }
@@ -260,9 +261,9 @@ void testMetadataPopulatedFromDatasetRow() {
 
         assertEquals(1, cases.size());
         Map<String, Object> metadata = cases.get(0).metadata();
-        assertFalse(metadata.isEmpty(), "metadata should not be empty");
-        assertEquals("TURBO", metadata.get("performance_strategy"));
-        assertEquals("user123", metadata.get("userId"));
+        assertEquals(
+                Map.of("model", "gpt-4o", "performance_strategy", "TURBO", "userId", "user123"),
+                metadata);
     }
 
     @Test

btx/src/test/java/dev/braintrust/sdkspecimpl/clients/LangChain114OpenAiSpecClient.java

@@ -209,7 +209,7 @@ private void executeResponses(LlmSpanSpec spec, SpecClientContext ctx) throws Ex
         // seed it from the first request.
         String defaultModel =
                 spec.requests().isEmpty()
-                        ? "o4-mini"
+                        ? "gpt-6-luna"
                         : String.valueOf(spec.requests().get(0).get("model"));
         OpenAiResponsesChatModel model =
                 BraintrustLangchain.wrap(

gradle.properties

@@ -8,10 +8,10 @@ org.gradle.daemon=true
 org.gradle.warning.mode=summary
 
 # braintrust-spec git ref (SHA or tag) used by btx tests
-braintrustSpecRef=a751e325b60c1d9e84c5aaaf49f990496d2fe7da
+braintrustSpecRef=c3c80aa77cc722164dc6748400853b9a5b2512be
 
 # braintrust-openapi commit SHA used by braintrust-api
-braintrustOpenApiRef=8d6e24ce78d8e33cbf9ba5f35d2fbc71749306d7
+braintrustOpenApiRef=4cc0615694512d6e77fce6bff2c70d1531208daa
 
 # Let Gradle locate local JDKs and download one if needed
 org.gradle.java.installations.auto-detect=true

test-harness/build.gradle

@@ -22,6 +22,10 @@ ext {
 }
 
 dependencies {
+  testImplementation 'org.wiremock:wiremock:3.13.1'
+  testImplementation "org.junit.jupiter:junit-jupiter:${junitVersion}"
+  testRuntimeOnly 'org.junit.platform:junit-platform-launcher'
+
   // testFixtures dependencies — everything lives in testFixtures source set
   testFixturesImplementation project(":braintrust-api")
   testFixturesImplementation project(":braintrust-sdk")
@@ -48,3 +52,7 @@ dependencies {
   testFixturesCompileOnly 'software.amazon.awssdk:bedrockruntime:2.30.0'
   testFixturesCompileOnly 'software.amazon.awssdk:netty-nio-client:2.30.0'
 }
+
+test {
+  useJUnitPlatform()
+}