agentv 3.4.0 → 3.6.0

package/dist/{chunk-RE5I3U2S.js → chunk-UU5N43YS.js}

@@ -25,12 +25,12 @@ import {
   subscribeToCopilotCliLogEntries,
   subscribeToCopilotSdkLogEntries,
   subscribeToPiLogEntries
-} from "./chunk-GOZV2HN2.js";
+} from "./chunk-K4RXLQWV.js";
 
 // package.json
 var package_default = {
   name: "agentv",
-  version: "3.4.0",
+  version: "3.6.0",
   description: "CLI entry point for AgentV",
   type: "module",
   repository: {
@@ -321,33 +321,11 @@ function parseWorkspaceChanges(fileChanges) {
   };
 }
 function buildAssertions(result) {
-  const assertions = [];
-  if (result.scores && result.scores.length > 0) {
-    for (const evaluator of result.scores) {
-      for (const hit of evaluator.hits) {
-        assertions.push({
-          text: hit,
-          passed: true,
-          evidence: evaluator.reasoning ?? ""
-        });
-      }
-      for (const miss of evaluator.misses) {
-        assertions.push({
-          text: miss,
-          passed: false,
-          evidence: evaluator.reasoning ?? ""
-        });
-      }
-    }
-  } else {
-    for (const hit of result.hits) {
-      assertions.push({ text: hit, passed: true, evidence: result.reasoning ?? "" });
-    }
-    for (const miss of result.misses) {
-      assertions.push({ text: miss, passed: false, evidence: result.reasoning ?? "" });
-    }
-  }
-  return assertions;
+  return result.assertions.map((a) => ({
+    text: a.text,
+    passed: a.passed,
+    evidence: a.evidence ?? ""
+  }));
 }
 function buildEvaluators(scores) {
   if (!scores || scores.length === 0) {
@@ -357,11 +335,10 @@ function buildEvaluators(scores) {
     name: s.name,
     type: s.type,
     score: s.score,
-    reasoning: s.reasoning ?? "",
+    reasoning: "",
     weight: s.weight,
     verdict: s.verdict,
-    hits: s.hits,
-    misses: s.misses,
+    assertions: s.assertions,
     details: s.details
   }));
 }
@@ -1351,29 +1328,33 @@ var SCRIPT = `
     if(r.input!=null){
       h+='<div class="detail-block"><h4>Input</h4><pre class="detail-pre">'+esc(typeof r.input==="string"?r.input:JSON.stringify(r.input,null,2))+"</pre></div>";
     }
-    h+='<div class="detail-block"><h4>Output</h4><pre class="detail-pre">'+esc(r.answer||"")+"</pre></div>";
+    h+='<div class="detail-block"><h4>Output</h4><pre class="detail-pre">'+esc(r.outputText||"")+"</pre></div>";
     h+="</div>";
 
     /* evaluator results */
     if(r.scores&&r.scores.length>0){
       h+="<h4>Evaluator Results</h4>";
-      h+='<table class="eval-table"><thead><tr><th>Evaluator</th><th>Score</th><th>Status</th><th>Reasoning</th></tr></thead><tbody>';
+      h+='<table class="eval-table"><thead><tr><th>Evaluator</th><th>Score</th><th>Status</th><th>Assertions</th></tr></thead><tbody>';
       for(var i=0;i<r.scores.length;i++){
         var ev=r.scores[i],evS=ev.score>=0.5?"pass":"fail";
-        h+="<tr><td class=\\"fw-medium\\">"+esc(ev.name)+'</td><td class="'+sCls(ev.score)+'">'+fmtPct(ev.score)+"</td><td>"+sIcon(evS)+'</td><td class="reasoning-cell">'+esc(ev.reasoning||"")+"</td></tr>";
+        var evAssertions=ev.assertions||[];
+        var evSummary=evAssertions.map(function(a){return (a.passed?"\u2713 ":"\u2717 ")+a.text;}).join("; ");
+        h+="<tr><td class=\\"fw-medium\\">"+esc(ev.name)+'</td><td class="'+sCls(ev.score)+'">'+fmtPct(ev.score)+"</td><td>"+sIcon(evS)+'</td><td class="reasoning-cell">'+esc(evSummary)+"</td></tr>";
       }
       h+="</tbody></table>";
     }
 
-    /* hits / misses */
-    if(r.hits&&r.hits.length>0){
-      h+='<h4>Passed Expectations</h4><ul class="expect-list pass">';
-      for(var i=0;i<r.hits.length;i++)h+="<li>"+esc(r.hits[i])+"</li>";
+    /* assertions */
+    var passedA=r.assertions?r.assertions.filter(function(a){return a.passed;}):[];
+    var failedA=r.assertions?r.assertions.filter(function(a){return !a.passed;}):[];
+    if(passedA.length>0){
+      h+='<h4>Passed Assertions</h4><ul class="expect-list pass">';
+      for(var i=0;i<passedA.length;i++)h+="<li>"+esc(passedA[i].text)+(passedA[i].evidence?" <span class=\\"reasoning-cell\\">("+esc(passedA[i].evidence)+")</span>":"")+"</li>";
       h+="</ul>";
     }
-    if(r.misses&&r.misses.length>0){
-      h+='<h4>Failed Expectations</h4><ul class="expect-list fail">';
-      for(var i=0;i<r.misses.length;i++)h+="<li>"+esc(r.misses[i])+"</li>";
+    if(failedA.length>0){
+      h+='<h4>Failed Assertions</h4><ul class="expect-list fail">';
+      for(var i=0;i<failedA.length;i++)h+="<li>"+esc(failedA[i].text)+(failedA[i].evidence?" <span class=\\"reasoning-cell\\">("+esc(failedA[i].evidence)+")</span>":"")+"</li>";
       h+="</ul>";
     }
 
@@ -1562,10 +1543,10 @@ var JunitWriter = class _JunitWriter {
     `;
         } else if (r.score < 0.5) {
           const message = `score=${r.score.toFixed(3)}`;
+          const failedAssertions = r.assertions.filter((a) => !a.passed);
           const detail = [
             `Score: ${r.score.toFixed(3)}`,
-            r.reasoning ? `Reasoning: ${r.reasoning}` : "",
-            r.misses.length > 0 ? `Misses: ${r.misses.join(", ")}` : ""
+            failedAssertions.length > 0 ? `Failed: ${failedAssertions.map((a) => a.text).join(", ")}` : ""
           ].filter(Boolean).join("\n");
           inner = `
       <failure message="${escapeXml(message)}">${escapeXml(detail)}</failure>
@@ -4063,7 +4044,7 @@ async function runEvalCommand(input) {
   const useFileExport = !!(options.otelFile || options.traceFile);
   if (options.exportOtel || useFileExport) {
     try {
-      const { OtelTraceExporter, OTEL_BACKEND_PRESETS } = await import("./dist-AFDYFH6Y.js");
+      const { OtelTraceExporter, OTEL_BACKEND_PRESETS } = await import("./dist-VWEFBDZ5.js");
       let endpoint = process.env.OTEL_EXPORTER_OTLP_ENDPOINT;
       let headers = {};
       if (options.otelBackend) {
@@ -4405,4 +4386,4 @@ export {
   selectTarget,
   runEvalCommand
 };
-//# sourceMappingURL=chunk-RE5I3U2S.js.map
+//# sourceMappingURL=chunk-UU5N43YS.js.map