agentv 3.11.1 → 3.13.0

package/dist/{chunk-OYD2NB55.js → chunk-7OHZAFND.js}

@@ -1,4 +1,7 @@
 import { createRequire } from 'node:module'; const require = createRequire(import.meta.url);
+import {
+  require_token_error
+} from "./chunk-HQDCIXVH.js";
 import {
   AISDKError,
   APICallError,
@@ -147,9 +150,6 @@ import {
   withoutTrailingSlash,
   zodSchema
 } from "./chunk-BJV6MDBE.js";
-import {
-  require_token_error
-} from "./chunk-HQDCIXVH.js";
 import {
   SpanStatusCode,
   context,
@@ -14856,12 +14856,6 @@ function parseExecutionDefaults(raw, configPath) {
   } else if (obj.verbose !== void 0) {
     logWarning(`Invalid execution.verbose in ${configPath}, expected boolean`);
   }
-  const traceFile = obj.trace_file;
-  if (typeof traceFile === "string" && traceFile.trim().length > 0) {
-    result.trace_file = traceFile.trim();
-  } else if (traceFile !== void 0) {
-    logWarning(`Invalid execution.trace_file in ${configPath}, expected non-empty string`);
-  }
   if (typeof obj.keep_workspaces === "boolean") {
     result.keep_workspaces = obj.keep_workspaces;
   } else if (obj.keep_workspaces !== void 0) {
@@ -21644,7 +21638,7 @@ async function loadSdkModules() {
     try {
       [piCodingAgentModule, piAiModule] = await Promise.all([
         import("@mariozechner/pi-coding-agent"),
-        import("./dist-BOIN5LC5-T5UWUK43.js")
+        import("@mariozechner/pi-ai")
       ]);
     } catch {
       if (await promptInstall()) {
@@ -21652,7 +21646,7 @@ async function loadSdkModules() {
         execSync("bun add @mariozechner/pi-coding-agent", { stdio: "inherit" });
         [piCodingAgentModule, piAiModule] = await Promise.all([
           import("@mariozechner/pi-coding-agent"),
-          import("./dist-BOIN5LC5-T5UWUK43.js")
+          import("@mariozechner/pi-ai")
         ]);
       } else {
         throw new Error(
@@ -23614,10 +23608,26 @@ function extractJsonBlob(text2) {
   const match = text2.match(/\{[\s\S]*\}/);
   return match?.[0];
 }
+function repairSchemaNearBooleanFields(text2) {
+  return text2.replace(
+    /("passed"\s*:\s*)(?:"([^"]+)"|([A-Za-z_][A-Za-z0-9_-]*))/gi,
+    (_match, prefix, quotedValue, bareValue) => {
+      const value = (quotedValue ?? bareValue ?? "").trim().toLowerCase();
+      if (value === "true") {
+        return `${prefix}true`;
+      }
+      if (value === "false") {
+        return `${prefix}false`;
+      }
+      return `${prefix}false`;
+    }
+  );
+}
 function parseJsonFromText(text2) {
   const cleaned = typeof text2 === "string" ? text2.replace(/```json\n?|```/g, "").trim() : "";
   const blob = extractJsonBlob(cleaned) ?? cleaned;
-  return JSON.parse(blob);
+  const repaired = repairSchemaNearBooleanFields(blob);
+  return JSON.parse(repaired);
 }
 function isNonEmptyString(value) {
   return typeof value === "string" && value.trim().length > 0;
@@ -24416,7 +24426,7 @@ ${context2.fileChanges}`;
   async evaluateWithRubrics(context2, graderProvider, rubrics) {
     if (!rubrics || rubrics.length === 0) {
       throw new Error(
-        `No rubrics found for evaluator "${context2.evaluator?.name ?? "llm-grader"}". Run "agentv generate rubrics" first.`
+        `No rubrics found for evaluator "${context2.evaluator?.name ?? "llm-grader"}". Add rubric criteria under assertions or use the agentv-eval-writer skill for authoring help.`
       );
     }
     const hasScoreRanges = rubrics.some((r) => r.score_ranges && r.score_ranges.length > 0);
@@ -28654,7 +28664,7 @@ async function runEvaluation(options) {
         if (!cliModel) {
           throw new Error('--grader-target "agentv" requires --model (e.g., "openai:gpt-5-mini")');
         }
-        const { AgentvProvider: AgentvProvider2 } = await import("./agentv-provider-MIDKLYIH-6LIYKQRP.js");
+        const { AgentvProvider: AgentvProvider2 } = await import("./agentv-provider-NFFLXG5M-TJAWCWCX.js");
         return new AgentvProvider2("agentv", { model: cliModel, temperature: 0 });
       }
       const overrideTarget = resolveTargetByName(cliGraderTarget);
@@ -30717,8 +30727,6 @@ var AgentVConfigSchema = external_exports2.object({
     agentTimeoutMs: external_exports2.number().int().min(0).optional(),
     /** Enable verbose logging */
     verbose: external_exports2.boolean().optional(),
-    /** Write human-readable trace JSONL to this path (supports {timestamp} placeholder) */
-    traceFile: external_exports2.string().optional(),
     /** Always keep temp workspaces after eval */
     keepWorkspaces: external_exports2.boolean().optional(),
     /** Write OTLP JSON trace to this path (supports {timestamp} placeholder) */
@@ -30978,11 +30986,11 @@ var OtelTraceExporter = class {
   async init() {
     try {
       const [sdkTraceNode, resourcesMod, semconvMod, api, coreMod] = await Promise.all([
-        import("./src-PXDA7QIS.js"),
-        import("./esm-UYZ3HJBU.js"),
+        import("./src-ML4D2MC2.js"),
+        import("./esm-CZAWIY6F.js"),
         import("./esm-RVQPUGWH.js"),
         import("./esm-R77SNOF5.js"),
-        import("./esm-OJ2BXJK4-YKEI3Z7E.js").catch(() => null)
+        import("./esm-5Q4BZALM-5REQWAUV.js").catch(() => null)
       ]);
       const { NodeTracerProvider: Provider, SimpleSpanProcessor } = sdkTraceNode;
       const { resourceFromAttributes } = resourcesMod;
@@ -31001,17 +31009,11 @@ var OtelTraceExporter = class {
         processors.push(new SimpleSpanProcessor(exporter));
       }
       if (this.options.otlpFilePath) {
-        const { OtlpJsonFileExporter: OtlpJsonFileExporter2 } = await import("./otlp-json-file-exporter-VN67MK3S-RQIM6EHY.js");
+        const { OtlpJsonFileExporter: OtlpJsonFileExporter2 } = await import("./otlp-json-file-exporter-77FDBRSY-EZAPHWP6.js");
         processors.push(
           new SimpleSpanProcessor(new OtlpJsonFileExporter2(this.options.otlpFilePath))
         );
       }
-      if (this.options.traceFilePath) {
-        const { SimpleTraceFileExporter: SimpleTraceFileExporter2 } = await import("./simple-trace-file-exporter-XWZTIZR2-4JKATE5G.js");
-        processors.push(
-          new SimpleSpanProcessor(new SimpleTraceFileExporter2(this.options.traceFilePath))
-        );
-      }
       if (processors.length === 0) {
         return false;
       }
@@ -31071,6 +31073,17 @@ var OtelTraceExporter = class {
         if (result.durationMs != null)
           rootSpan.setAttribute("agentv.trace.duration_ms", result.durationMs);
         if (result.costUsd != null) rootSpan.setAttribute("agentv.trace.cost_usd", result.costUsd);
+        if (result.tokenUsage) {
+          if (result.tokenUsage.input != null) {
+            rootSpan.setAttribute("agentv.trace.token_input", result.tokenUsage.input);
+          }
+          if (result.tokenUsage.output != null) {
+            rootSpan.setAttribute("agentv.trace.token_output", result.tokenUsage.output);
+          }
+          if (result.tokenUsage.cached != null) {
+            rootSpan.setAttribute("agentv.trace.token_cached", result.tokenUsage.cached);
+          }
+        }
         if (result.trace) {
           const t = result.trace;
           rootSpan.setAttribute("agentv.trace.event_count", t.eventCount);
@@ -31173,6 +31186,7 @@ var OtelTraceExporter = class {
       tracer.startActiveSpan(
         spanName,
         { startTime: startHr },
+        parentCtx,
         (span) => {
           if (isAssistant) {
             span.setAttribute("gen_ai.operation.name", "chat");
@@ -31205,6 +31219,7 @@ var OtelTraceExporter = class {
                 tracer.startActiveSpan(
                   `execute_tool ${tc.tool}`,
                   {},
+                  msgCtx,
                   (toolSpan) => {
                     toolSpan.setAttribute("gen_ai.tool.name", tc.tool);
                     if (tc.id) toolSpan.setAttribute("gen_ai.tool.call.id", tc.id);
@@ -31245,8 +31260,12 @@ var OtelStreamingObserver = class {
   rootSpan = null;
   // biome-ignore lint/suspicious/noExplicitAny: OTel context loaded dynamically
   rootCtx = null;
+  observedChildSpans = false;
+  pendingMetrics = null;
   /** Create root eval span immediately (visible in backend right away) */
   startEvalCase(testId, target, evalSet) {
+    this.pendingMetrics = null;
+    this.observedChildSpans = false;
     const ctx = this.parentCtx ?? this.api.context.active();
     this.rootSpan = this.tracer.startSpan("agentv.eval", void 0, ctx);
     this.rootSpan.setAttribute("gen_ai.operation.name", "evaluate");
@@ -31259,8 +31278,9 @@ var OtelStreamingObserver = class {
   /** Create and immediately export a tool span */
   onToolCall(name21, input, output, _durationMs, toolCallId) {
     if (!this.rootCtx) return;
+    this.observedChildSpans = true;
     this.api.context.with(this.rootCtx, () => {
-      const span = this.tracer.startSpan(`execute_tool ${name21}`);
+      const span = this.tracer.startSpan(`execute_tool ${name21}`, void 0, this.rootCtx);
       span.setAttribute("gen_ai.tool.name", name21);
       if (toolCallId) span.setAttribute("gen_ai.tool.call.id", toolCallId);
       if (this.captureContent) {
@@ -31281,8 +31301,9 @@ var OtelStreamingObserver = class {
   /** Create and immediately export an LLM span */
   onLlmCall(model, tokenUsage) {
     if (!this.rootCtx) return;
+    this.observedChildSpans = true;
     this.api.context.with(this.rootCtx, () => {
-      const span = this.tracer.startSpan(`chat ${model}`);
+      const span = this.tracer.startSpan(`chat ${model}`, void 0, this.rootCtx);
       span.setAttribute("gen_ai.operation.name", "chat");
       span.setAttribute("gen_ai.request.model", model);
       span.setAttribute("gen_ai.response.model", model);
@@ -31297,10 +31318,53 @@ var OtelStreamingObserver = class {
       span.end();
     });
   }
+  /** Record final execution metrics before the root span is finalized. */
+  recordEvalMetrics(result) {
+    this.pendingMetrics = result;
+  }
   /** Finalize root span with score/verdict after evaluation completes */
   finalizeEvalCase(score, error) {
     if (!this.rootSpan) return;
     this.rootSpan.setAttribute("agentv.score", score);
+    if (this.pendingMetrics?.durationMs != null) {
+      this.rootSpan.setAttribute("agentv.trace.duration_ms", this.pendingMetrics.durationMs);
+    }
+    if (this.pendingMetrics?.costUsd != null) {
+      this.rootSpan.setAttribute("agentv.trace.cost_usd", this.pendingMetrics.costUsd);
+    }
+    if (this.pendingMetrics?.tokenUsage) {
+      if (this.pendingMetrics.tokenUsage.input != null) {
+        this.rootSpan.setAttribute(
+          "agentv.trace.token_input",
+          this.pendingMetrics.tokenUsage.input
+        );
+      }
+      if (this.pendingMetrics.tokenUsage.output != null) {
+        this.rootSpan.setAttribute(
+          "agentv.trace.token_output",
+          this.pendingMetrics.tokenUsage.output
+        );
+      }
+      if (this.pendingMetrics.tokenUsage.cached != null) {
+        this.rootSpan.setAttribute(
+          "agentv.trace.token_cached",
+          this.pendingMetrics.tokenUsage.cached
+        );
+      }
+    }
+    if (this.pendingMetrics?.trace) {
+      this.rootSpan.setAttribute("agentv.trace.event_count", this.pendingMetrics.trace.eventCount);
+      this.rootSpan.setAttribute(
+        "agentv.trace.tool_names",
+        Object.keys(this.pendingMetrics.trace.toolCalls).sort().join(",")
+      );
+      if (this.pendingMetrics.trace.llmCallCount != null) {
+        this.rootSpan.setAttribute(
+          "agentv.trace.llm_call_count",
+          this.pendingMetrics.trace.llmCallCount
+        );
+      }
+    }
     if (error) {
       this.rootSpan.setStatus({ code: this.api.SpanStatusCode.ERROR, message: error });
     } else {
@@ -31309,6 +31373,33 @@ var OtelStreamingObserver = class {
     this.rootSpan.end();
     this.rootSpan = null;
     this.rootCtx = null;
+    this.observedChildSpans = false;
+    this.pendingMetrics = null;
+  }
+  /** Backfill child spans from the completed result when the provider emitted no live callbacks. */
+  completeFromResult(result) {
+    this.recordEvalMetrics({
+      durationMs: result.durationMs,
+      costUsd: result.costUsd,
+      tokenUsage: result.tokenUsage,
+      trace: result.trace
+    });
+    if (this.observedChildSpans || !this.rootCtx) {
+      return;
+    }
+    const model = result.output.find((msg) => msg.role === "assistant")?.metadata?.model ?? result.target ?? "unknown";
+    this.onLlmCall(String(model), result.tokenUsage);
+    for (const message of result.output) {
+      for (const toolCall of message.toolCalls ?? []) {
+        this.onToolCall(
+          toolCall.tool,
+          toolCall.input,
+          toolCall.output,
+          toolCall.durationMs ?? 0,
+          toolCall.id
+        );
+      }
+    }
   }
   /** Return the active eval span's trace ID and span ID for Braintrust trace bridging */
   getActiveSpanIds() {
@@ -31499,4 +31590,4 @@ export {
   OtelStreamingObserver,
   createAgentKernel
 };
-//# sourceMappingURL=chunk-OYD2NB55.js.map
+//# sourceMappingURL=chunk-7OHZAFND.js.map