braintrust 3.30.0 → 3.31.0

package/dev/dist/index.d.mts

@@ -520,6 +520,9 @@ interface Score {
      */
     error?: unknown;
 }
+type SingleScore = Omit<Score, "name"> & {
+    name?: string;
+};
 
 declare const spanComponentsV4Schema: z.ZodIntersection<z.ZodIntersection<z.ZodObject<{
     object_type: z.ZodNativeEnum<typeof SpanObjectTypeV3>;
@@ -15436,7 +15439,7 @@ type EvalScorerArgs<Input, Output, Expected, Metadata extends BaseMetadata = Def
     output: Output;
     trace?: Trace;
 };
-type OneOrMoreScores = Score | number | null | Array<Score>;
+type OneOrMoreScores = SingleScore | number | null | Array<Score>;
 type EvalScorer<Input, Output, Expected, Metadata extends BaseMetadata = DefaultMetadataType> = (args: EvalScorerArgs<Input, Output, Expected, Metadata>) => OneOrMoreScores | Promise<OneOrMoreScores>;
 type OneOrMoreClassifications = Classification | Classification[] | null;
 type EvalClassifier<Input, Output, Expected, Metadata extends BaseMetadata = DefaultMetadataType> = (args: EvalScorerArgs<Input, Output, Expected, Metadata>) => OneOrMoreClassifications | Promise<OneOrMoreClassifications>;

package/dev/dist/index.d.ts

@@ -520,6 +520,9 @@ interface Score {
      */
     error?: unknown;
 }
+type SingleScore = Omit<Score, "name"> & {
+    name?: string;
+};
 
 declare const spanComponentsV4Schema: z.ZodIntersection<z.ZodIntersection<z.ZodObject<{
     object_type: z.ZodNativeEnum<typeof SpanObjectTypeV3>;
@@ -15436,7 +15439,7 @@ type EvalScorerArgs<Input, Output, Expected, Metadata extends BaseMetadata = Def
     output: Output;
     trace?: Trace;
 };
-type OneOrMoreScores = Score | number | null | Array<Score>;
+type OneOrMoreScores = SingleScore | number | null | Array<Score>;
 type EvalScorer<Input, Output, Expected, Metadata extends BaseMetadata = DefaultMetadataType> = (args: EvalScorerArgs<Input, Output, Expected, Metadata>) => OneOrMoreScores | Promise<OneOrMoreScores>;
 type OneOrMoreClassifications = Classification | Classification[] | null;
 type EvalClassifier<Input, Output, Expected, Metadata extends BaseMetadata = DefaultMetadataType> = (args: EvalScorerArgs<Input, Output, Expected, Metadata>) => OneOrMoreClassifications | Promise<OneOrMoreClassifications>;

package/dev/dist/index.js

@@ -4118,6 +4118,7 @@ var ProjectSettings = _v3.z.union([
     ]),
     disable_realtime_queries: _v3.z.union([_v3.z.boolean(), _v3.z.null()]),
     monitor_charts_use_metrics_start: _v3.z.union([_v3.z.boolean(), _v3.z.null()]),
+    blind_reviews: _v3.z.union([_v3.z.boolean(), _v3.z.null()]),
     default_preprocessor: NullableSavedFunctionId
   }).partial(),
   _v3.z.null()
@@ -4179,6 +4180,7 @@ var WindowedAutomationConfig = _v3.z.object({
     auto_approve_tools: _v3.z.array(_v3.z.string().min(1)).optional().default([]),
     harness: _v3.z.enum(["native", "codex", "claude-code"]).optional(),
     model: _v3.z.string().min(1).optional(),
+    endpoint_name: _v3.z.string().min(1).optional(),
     reasoning_effort: _v3.z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max"]).optional()
   }).optional(),
   actions: _v3.z.array(
@@ -5972,7 +5974,7 @@ var INSTRUMENTATION_NAMES = {
 var INTERNAL_SPAN_INSTRUMENTATION_NAME = /* @__PURE__ */ Symbol.for(
   "braintrust.spanInstrumentationName"
 );
-var SDK_VERSION = true ? "3.30.0" : "0.0.0";
+var SDK_VERSION = true ? "3.31.0" : "0.0.0";
 function withSpanInstrumentationName(args, instrumentationName) {
   return {
     ...args,
@@ -19197,6 +19199,12 @@ function prepareAISDKChildTracing(params, self, parentSpan, denyOutputPaths, aiS
       if (!activeEntry) {
         return result;
       }
+      const executionOptions = args[1];
+      const toolCallId = isObject(executionOptions) ? executionOptions.toolCallId : void 0;
+      const toolInput = {
+        input: serializeToolExecutionInput(args),
+        ...typeof toolCallId === "string" ? { metadata: { toolCallId } } : {}
+      };
       if (isAsyncGenerator(result)) {
         return (async function* () {
           const span = activeEntry.parentSpan.startSpan(
@@ -19210,7 +19218,7 @@ function prepareAISDKChildTracing(params, self, parentSpan, denyOutputPaths, aiS
               INSTRUMENTATION_NAMES.AI_SDK
             )
           );
-          span.log({ input: serializeToolExecutionInput(args) });
+          span.log(toolInput);
           try {
             let lastValue;
             for await (const value of result) {
@@ -19228,7 +19236,7 @@ function prepareAISDKChildTracing(params, self, parentSpan, denyOutputPaths, aiS
       }
       return activeEntry.parentSpan.traced(
         async (span) => {
-          span.log({ input: serializeToolExecutionInput(args) });
+          span.log(toolInput);
           const awaitedResult = await result;
           span.log({ output: awaitedResult });
           return awaitedResult;
@@ -33350,14 +33358,17 @@ function getMetricsFromResponse(response) {
     if (!isRecord(usageMetadata)) {
       continue;
     }
-    const inputTokenDetails = usageMetadata.input_token_details;
+    const inputTokenDetails = isRecord(usageMetadata.input_token_details) ? usageMetadata.input_token_details : {};
     const outputTokenDetails = usageMetadata.output_token_details;
     return normalizeTokenMetrics({
       total_tokens: usageMetadata.total_tokens,
       prompt_tokens: usageMetadata.input_tokens,
       completion_tokens: usageMetadata.output_tokens,
-      prompt_cache_creation_tokens: isRecord(inputTokenDetails) ? inputTokenDetails.cache_creation : void 0,
-      prompt_cached_tokens: isRecord(inputTokenDetails) ? inputTokenDetails.cache_read : void 0,
+      // Prefer TTL-specific cache writes over the aggregate when available.
+      prompt_cache_creation_tokens: inputTokenDetails.ephemeral_5m_input_tokens == null && inputTokenDetails.ephemeral_1h_input_tokens == null ? inputTokenDetails.cache_creation : void 0,
+      prompt_cache_creation_5m_tokens: inputTokenDetails.ephemeral_5m_input_tokens,
+      prompt_cache_creation_1h_tokens: inputTokenDetails.ephemeral_1h_input_tokens,
+      prompt_cached_tokens: inputTokenDetails.cache_read,
       completion_reasoning_tokens: isRecord(outputTokenDetails) ? outputTokenDetails.reasoning : void 0
     });
   }
@@ -39371,19 +39382,29 @@ function buildSpanScores(results) {
 function _internalPrepareEvaluatorScore(scoreValue, name) {
   if (scoreValue === null) return { results: null };
   if (Array.isArray(scoreValue)) {
+    const names = /* @__PURE__ */ new Set();
     for (const score of scoreValue) {
       if (!(typeof score === "object" && !isEmpty2(score))) {
         throw new Error(
           `When returning an array of scores, each score must be a non-empty object. Got: ${JSON.stringify(score)}`
         );
       }
+      if (typeof score.name !== "string") {
+        throw new Error(
+          `When returning an array of scores, each score must have a name. Got: ${JSON.stringify(score)}`
+        );
+      }
+      if (names.has(score.name)) {
+        throw new Error(`Duplicate score name '${score.name}' in score array`);
+      }
+      names.add(score.name);
     }
   }
   let results;
   if (Array.isArray(scoreValue)) {
     results = scoreValue;
   } else if (typeof scoreValue === "object" && !isEmpty2(scoreValue)) {
-    results = [scoreValue];
+    results = [{ ...scoreValue, name: _nullishCoalesce(scoreValue.name, () => ( name)) }];
   } else {
     results = [{ name, score: scoreValue }];
   }

package/dev/dist/index.mjs

@@ -4118,6 +4118,7 @@ var ProjectSettings = z6.union([
     ]),
     disable_realtime_queries: z6.union([z6.boolean(), z6.null()]),
     monitor_charts_use_metrics_start: z6.union([z6.boolean(), z6.null()]),
+    blind_reviews: z6.union([z6.boolean(), z6.null()]),
     default_preprocessor: NullableSavedFunctionId
   }).partial(),
   z6.null()
@@ -4179,6 +4180,7 @@ var WindowedAutomationConfig = z6.object({
     auto_approve_tools: z6.array(z6.string().min(1)).optional().default([]),
     harness: z6.enum(["native", "codex", "claude-code"]).optional(),
     model: z6.string().min(1).optional(),
+    endpoint_name: z6.string().min(1).optional(),
     reasoning_effort: z6.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max"]).optional()
   }).optional(),
   actions: z6.array(
@@ -5972,7 +5974,7 @@ var INSTRUMENTATION_NAMES = {
 var INTERNAL_SPAN_INSTRUMENTATION_NAME = /* @__PURE__ */ Symbol.for(
   "braintrust.spanInstrumentationName"
 );
-var SDK_VERSION = true ? "3.30.0" : "0.0.0";
+var SDK_VERSION = true ? "3.31.0" : "0.0.0";
 function withSpanInstrumentationName(args, instrumentationName) {
   return {
     ...args,
@@ -19197,6 +19199,12 @@ function prepareAISDKChildTracing(params, self, parentSpan, denyOutputPaths, aiS
       if (!activeEntry) {
         return result;
       }
+      const executionOptions = args[1];
+      const toolCallId = isObject(executionOptions) ? executionOptions.toolCallId : void 0;
+      const toolInput = {
+        input: serializeToolExecutionInput(args),
+        ...typeof toolCallId === "string" ? { metadata: { toolCallId } } : {}
+      };
       if (isAsyncGenerator(result)) {
         return (async function* () {
           const span = activeEntry.parentSpan.startSpan(
@@ -19210,7 +19218,7 @@ function prepareAISDKChildTracing(params, self, parentSpan, denyOutputPaths, aiS
               INSTRUMENTATION_NAMES.AI_SDK
             )
           );
-          span.log({ input: serializeToolExecutionInput(args) });
+          span.log(toolInput);
           try {
             let lastValue;
             for await (const value of result) {
@@ -19228,7 +19236,7 @@ function prepareAISDKChildTracing(params, self, parentSpan, denyOutputPaths, aiS
       }
       return activeEntry.parentSpan.traced(
         async (span) => {
-          span.log({ input: serializeToolExecutionInput(args) });
+          span.log(toolInput);
           const awaitedResult = await result;
           span.log({ output: awaitedResult });
           return awaitedResult;
@@ -33350,14 +33358,17 @@ function getMetricsFromResponse(response) {
     if (!isRecord(usageMetadata)) {
       continue;
     }
-    const inputTokenDetails = usageMetadata.input_token_details;
+    const inputTokenDetails = isRecord(usageMetadata.input_token_details) ? usageMetadata.input_token_details : {};
     const outputTokenDetails = usageMetadata.output_token_details;
     return normalizeTokenMetrics({
       total_tokens: usageMetadata.total_tokens,
       prompt_tokens: usageMetadata.input_tokens,
       completion_tokens: usageMetadata.output_tokens,
-      prompt_cache_creation_tokens: isRecord(inputTokenDetails) ? inputTokenDetails.cache_creation : void 0,
-      prompt_cached_tokens: isRecord(inputTokenDetails) ? inputTokenDetails.cache_read : void 0,
+      // Prefer TTL-specific cache writes over the aggregate when available.
+      prompt_cache_creation_tokens: inputTokenDetails.ephemeral_5m_input_tokens == null && inputTokenDetails.ephemeral_1h_input_tokens == null ? inputTokenDetails.cache_creation : void 0,
+      prompt_cache_creation_5m_tokens: inputTokenDetails.ephemeral_5m_input_tokens,
+      prompt_cache_creation_1h_tokens: inputTokenDetails.ephemeral_1h_input_tokens,
+      prompt_cached_tokens: inputTokenDetails.cache_read,
       completion_reasoning_tokens: isRecord(outputTokenDetails) ? outputTokenDetails.reasoning : void 0
     });
   }
@@ -39371,19 +39382,29 @@ function buildSpanScores(results) {
 function _internalPrepareEvaluatorScore(scoreValue, name) {
   if (scoreValue === null) return { results: null };
   if (Array.isArray(scoreValue)) {
+    const names = /* @__PURE__ */ new Set();
     for (const score of scoreValue) {
       if (!(typeof score === "object" && !isEmpty2(score))) {
         throw new Error(
           `When returning an array of scores, each score must be a non-empty object. Got: ${JSON.stringify(score)}`
         );
       }
+      if (typeof score.name !== "string") {
+        throw new Error(
+          `When returning an array of scores, each score must have a name. Got: ${JSON.stringify(score)}`
+        );
+      }
+      if (names.has(score.name)) {
+        throw new Error(`Duplicate score name '${score.name}' in score array`);
+      }
+      names.add(score.name);
     }
   }
   let results;
   if (Array.isArray(scoreValue)) {
     results = scoreValue;
   } else if (typeof scoreValue === "object" && !isEmpty2(scoreValue)) {
-    results = [scoreValue];
+    results = [{ ...scoreValue, name: scoreValue.name ?? name }];
   } else {
     results = [{ name, score: scoreValue }];
   }