agent-inspect 6.31.7 → 6.31.8

package/packages/cli/dist/index.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { ATTRIBUTION_CONFIDENCES, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createSafetyOversizedAttributeRule, resolveTraceDir, parseDuration, buildSessionIndex, resolveBundleRunIds, getTraceFilePath, openTrace, sha256Hex, collectTraceSchemaVersions, aggregateBundleSafeStatus, bundleFailsOnSafety, bundleRunAssetRelativePath, buildPlaceholderArtifact, EVIDENCE_HTML_FILENAME, EVIDENCE_MANIFEST_FILENAME, buildBundleMetadata, buildBundleSummaryMarkdown, buildEvidenceHtmlShell, buildEvidenceProvenanceViewHtml, EVIDENCE_FORMAT_VERSION, inferEvidenceFileRole, buildEvidenceSafetyViewHtml, buildEvidenceDiffViewHtml, buildEvidenceCircuitViewHtml, buildEvidenceOutcomesViewHtml, buildEvidenceContractsViewHtml, buildEvidenceToolsLlmViewHtml, buildEvidenceCausalFailureViewHtml, buildEvidenceTimelineViewHtml, buildEvidenceTreeViewHtml, buildEvidenceManifest, serializeEvidenceManifest, buildZipArchive, assertBundlePathContained, enrichSessionRunRecord, normalizeBundleOutputPath, sanitizeBundleRunId, defaultBundleOutputPath, runTraceChecks, assertEvidenceRelativePath, TraceDirectory, loadTraceMetadataList, loadSessionRunRecords, TraceReadError, resolveRedactionProfile, applyProfileMetadataCaps, Redactor, flattenTree, escapeHtml, safeString, compactAttributes, stableJson, escapeMarkdown, DEFAULT_REDACT_KEYS, truncateStringForProfile, stringContainsHighConfidenceCredential, isCredentialSensitiveKey, extractMetadata, filterTraces, truncateName, buildRunSummary, formatDuration, formatTimestamp, renderErrorLine, isAgentInspectTrace, persistedInspectEventsToRunTrees, defineTraceContract, filterMetasBySessionScope, evaluateTraceContract, aggregateSessionCheckResults, createBaselineRegressionRule, summarizeSemanticParity, buildEvidenceCiPackage, verifyEvidenceDirectory, diffTraceEvents, renderRunDiff, buildRunTimeline, renderTimeline, buildTraceStats, renderTraceStats, parseDurationFilter, searchTraces, formatStepLabel, buildActivitySummary, renderActivitySummaryHuman, buildRunWhatSummary, renderRunWhat, buildLocalExplanation, resolveSuiteTemplate, defaultSuiteConfigTemplate, loadSuiteConfig, validateSuiteConfig, runSuite, renderSuiteReport, analyzeCohort, renderCohortReport, gateHasThresholds, runGate, renderGateReport, persistedInspectEventsToTraceEvents, renderStepLine, getIndent, TreeBuilder, validateEvent, traceEventToPersistedInspectEvent, isPersistedInspectEvent, createRunStatusRule, createStructureOrphanRule, createStructureCycleRule, createRunDurationRule, createMaxStepDurationRule, createRequireCompletedRule, createStallDetectionRule, createObservedOutcomeRule, createRunDepthRule, createToolUsageRule, createLlmUsageRule, createStructureRelationshipRule, createStructureParallelWidthRule, projectLogicalEvents, summarizeObservedOutcomes, extractOutcomesFromTraceEvents, renderObservedOutcomesMarkdown, renderObservedOutcomesHtml, DEFAULT_SUITE_ARTIFACTS_DIR, parseCohortMetricList, COHORT_METRIC_IDS, parseGateList, nanoid, isAttributionConfidence, zeroKinds, diffRuns, manualTraceEventsToComparableRun } from './chunk-QV73NM25.mjs';
+import { ATTRIBUTION_CONFIDENCES, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createSafetyOversizedAttributeRule, resolveTraceDir, parseDuration, buildSessionIndex, resolveBundleRunIds, getTraceFilePath, openTrace, sha256Hex, collectTraceSchemaVersions, aggregateBundleSafeStatus, bundleFailsOnSafety, bundleRunAssetRelativePath, buildPlaceholderArtifact, EVIDENCE_HTML_FILENAME, EVIDENCE_MANIFEST_FILENAME, buildBundleMetadata, buildBundleSummaryMarkdown, buildEvidenceHtmlShell, buildEvidenceProvenanceViewHtml, EVIDENCE_FORMAT_VERSION, inferEvidenceFileRole, buildEvidenceSafetyViewHtml, buildEvidenceDiffViewHtml, buildEvidenceCircuitViewHtml, buildEvidenceOutcomesViewHtml, buildEvidenceContractsViewHtml, buildEvidenceToolsLlmViewHtml, buildEvidenceCausalFailureViewHtml, buildEvidenceTimelineViewHtml, buildEvidenceTreeViewHtml, buildEvidenceManifest, serializeEvidenceManifest, buildZipArchive, assertBundlePathContained, enrichSessionRunRecord, normalizeBundleOutputPath, sanitizeBundleRunId, defaultBundleOutputPath, runTraceChecks, assertEvidenceRelativePath, TraceDirectory, loadTraceMetadataList, loadSessionRunRecords, TraceReadError, resolveRedactionProfile, applyProfileMetadataCaps, Redactor, flattenTree, escapeHtml, safeString, compactAttributes, stableJson, escapeMarkdown, DEFAULT_REDACT_KEYS, truncateStringForProfile, stringContainsHighConfidenceCredential, isCredentialSensitiveKey, extractMetadata, filterTraces, truncateName, buildRunSummary, formatDuration, formatTimestamp, renderErrorLine, isAgentInspectTrace, persistedInspectEventsToRunTrees, defineTraceContract, filterMetasBySessionScope, evaluateTraceContract, aggregateSessionCheckResults, createBaselineRegressionRule, summarizeSemanticParity, buildEvidenceCiPackage, verifyEvidenceDirectory, diffTraceEvents, renderRunDiff, buildRunTimeline, renderTimeline, buildTraceStats, renderTraceStats, parseDurationFilter, searchTraces, formatStepLabel, buildActivitySummary, renderActivitySummaryHuman, buildRunWhatSummary, renderRunWhat, buildLocalExplanation, resolveSuiteTemplate, defaultSuiteConfigTemplate, loadSuiteConfig, validateSuiteConfig, runSuite, renderSuiteReport, analyzeCohort, renderCohortReport, gateHasThresholds, runGate, renderGateReport, persistedInspectEventsToTraceEvents, renderStepLine, getIndent, TreeBuilder, validateEvent, traceEventToPersistedInspectEvent, isPersistedInspectEvent, createRunStatusRule, createStructureOrphanRule, createStructureCycleRule, createRunDurationRule, createMaxStepDurationRule, createRequireCompletedRule, createStallDetectionRule, createObservedOutcomeRule, createRunDepthRule, createToolUsageRule, createLlmUsageRule, createStructureRelationshipRule, createStructureParallelWidthRule, projectLogicalEvents, summarizeObservedOutcomes, extractOutcomesFromTraceEvents, renderObservedOutcomesMarkdown, renderObservedOutcomesHtml, DEFAULT_SUITE_ARTIFACTS_DIR, parseCohortMetricList, COHORT_METRIC_IDS, parseGateList, nanoid, isAttributionConfidence, zeroKinds, diffRuns, manualTraceEventsToComparableRun } from './chunk-6FE47XSF.mjs';
 import { realpathSync, constants as constants$1, existsSync, readFileSync } from 'fs';
 import path19 from 'path';
 import { fileURLToPath, pathToFileURL } from 'url';
@@ -14,7 +14,7 @@ import { createServer } from 'http';
 import { createRequire } from 'module';
 
 // package.json
-var version = "6.31.7";
+var version = "6.31.8";
 
 // packages/cli/src/trace-dir-scale.ts
 var TRACE_COUNT_WARN = 1e3;
@@ -2496,11 +2496,13 @@ function exportOpenInference(tree, options) {
       endNs = startNs + unixNano(ev.durationMs);
     }
     const { openInferenceKind } = mapInspectKindToOI(ev.kind, warnings);
+    const meta = ev.attributes;
+    const originalSource = typeof meta?.originalSourceType === "string" && meta.originalSourceType.trim() !== "" ? meta.originalSourceType.trim() : ev.source.type;
     const attrs = {
       "openinference.span.kind": openInferenceKind,
       "agent_inspect.kind": ev.kind,
       "agent_inspect.confidence": ev.confidence,
-      "agent_inspect.source.type": ev.source.type,
+      "agent_inspect.source.type": originalSource,
       "agent_inspect.run_id": tree.runId,
       "agent_inspect.event_id": ev.eventId,
       "agent_inspect.status": ev.status ?? "unset"
@@ -2508,9 +2510,9 @@ function exportOpenInference(tree, options) {
     if (ev.durationMs !== void 0) {
       attrs["agent_inspect.duration_ms"] = ev.durationMs;
     }
-    const meta = ev.attributes;
-    if (meta?.model !== void 0 && typeof meta.model === "string") {
-      attrs["llm.model_name"] = meta.model;
+    const requestModel = (typeof meta?.model === "string" && meta.model.trim() !== "" ? meta.model : void 0) ?? (typeof meta?.modelId === "string" && meta.modelId.trim() !== "" ? meta.modelId : void 0);
+    if (requestModel !== void 0) {
+      attrs["llm.model_name"] = requestModel;
     }
     const tokens = meta?.tokens;
     if (tokens && typeof tokens === "object" && tokens !== null) {
@@ -2520,8 +2522,9 @@ function exportOpenInference(tree, options) {
       if (typeof outp === "number") attrs["llm.token_count.completion"] = outp;
     }
     if (includeAttributes && meta && typeof meta === "object") {
+      const skipKeys = /* @__PURE__ */ new Set(["tokens", "model", "modelId", "originalSourceType"]);
       for (const [k, v] of Object.entries(meta)) {
-        if (k === "tokens" || k === "model") continue;
+        if (skipKeys.has(k)) continue;
         if (v !== void 0 && v !== null && typeof v !== "object") {
           attrs[`agent_inspect.preview.${k}`] = typeof v === "string" ? v.slice(0, maxLen) : v;
         }
@@ -2639,24 +2642,33 @@ function exportOtlpJson(tree, options) {
     const spanId = hexFrom2(`${tree.runId}:${ev.eventId}`, 8);
     const parentSpanId = ev.parentId ? hexFrom2(`${tree.runId}:${ev.parentId}`, 8) : void 0;
     const { startNs, endNs } = resolveSpanTimes(ev);
+    const meta = ev.attributes;
+    const originalSource = typeof meta?.originalSourceType === "string" && meta.originalSourceType.trim() !== "" ? meta.originalSourceType.trim() : ev.source.type;
     const attrs = [
       stringAttr("agent_inspect.kind", ev.kind),
       stringAttr("agent_inspect.confidence", ev.confidence),
-      stringAttr("agent_inspect.source.type", ev.source.type),
+      stringAttr("agent_inspect.source.type", originalSource),
       stringAttr("agent_inspect.run_id", tree.runId),
       stringAttr("agent_inspect.event_id", ev.eventId),
       stringAttr("agent_inspect.status", ev.status ?? "unset")
     ];
     if (ev.durationMs !== void 0) {
-      attrs.push(intAttr("agent_inspect.duration_ms", ev.durationMs));
+      attrs.push(numberAttr("agent_inspect.duration_ms", ev.durationMs));
     }
     const op = genAiOperationName(ev.kind);
     if (op !== void 0) {
       attrs.push(stringAttr("gen_ai.operation.name", op));
     }
-    const meta = ev.attributes;
-    if (meta?.model !== void 0 && typeof meta.model === "string") {
-      attrs.push(stringAttr("gen_ai.request.model", meta.model.slice(0, maxLen)));
+    const requestModel = (typeof meta?.model === "string" && meta.model.trim() !== "" ? meta.model : void 0) ?? (typeof meta?.modelId === "string" && meta.modelId.trim() !== "" ? meta.modelId : void 0);
+    if (requestModel !== void 0) {
+      attrs.push(stringAttr("gen_ai.request.model", requestModel.slice(0, maxLen)));
+    }
+    const responseModel = typeof meta?.responseModelId === "string" && meta.responseModelId.trim() !== "" ? meta.responseModelId : typeof meta?.responseModel === "string" && meta.responseModel.trim() !== "" ? meta.responseModel : void 0;
+    if (responseModel !== void 0) {
+      attrs.push(stringAttr("gen_ai.response.model", responseModel.slice(0, maxLen)));
+    }
+    if (typeof meta?.provider === "string" && meta.provider.trim() !== "") {
+      attrs.push(stringAttr("gen_ai.provider.name", meta.provider.slice(0, maxLen)));
     }
     const tokens = meta?.tokens;
     if (tokens && typeof tokens === "object" && tokens !== null) {
@@ -2666,8 +2678,17 @@ function exportOtlpJson(tree, options) {
       if (typeof outp === "number") attrs.push(intAttr("gen_ai.usage.output_tokens", outp));
     }
     if (includeAttributes && meta && typeof meta === "object") {
+      const skipKeys = /* @__PURE__ */ new Set([
+        "tokens",
+        "model",
+        "modelId",
+        "responseModel",
+        "responseModelId",
+        "provider",
+        "originalSourceType"
+      ]);
       for (const [k, v] of Object.entries(meta)) {
-        if (k === "tokens" || k === "model") continue;
+        if (skipKeys.has(k)) continue;
         if (typeof v === "string") {
           attrs.push(stringAttr(`agent_inspect.preview.${k}`, v.slice(0, maxLen)));
         } else if (typeof v === "number" && Number.isFinite(v)) {
@@ -2737,12 +2758,42 @@ function pushPath(errors, path30, message) {
 function isRecord5(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
+function isHexId(value, byteLen) {
+  return typeof value === "string" && new RegExp(`^[0-9a-fA-F]{${byteLen * 2}}$`).test(value);
+}
+function validateAnyValue(errors, path30, value) {
+  if (!isRecord5(value)) {
+    pushPath(errors, path30, "must be an AnyValue object");
+    return;
+  }
+  const variants = [
+    "stringValue",
+    "boolValue",
+    "intValue",
+    "doubleValue",
+    "bytesValue",
+    "arrayValue",
+    "kvlistValue"
+  ].filter((key) => value[key] !== void 0);
+  if (variants.length !== 1) {
+    pushPath(
+      errors,
+      path30,
+      `AnyValue must set exactly one variant (found ${variants.length}: ${variants.join(",")})`
+    );
+  }
+  if (value.intValue !== void 0) {
+    if (typeof value.intValue !== "string" || !/^-?\d+$/.test(value.intValue)) {
+      pushPath(errors, `${path30}.intValue`, "must be a decimal integer string");
+    }
+  }
+}
 function validateOtlpJson(content) {
   const errors = [];
   const warnings = [
     EXPERIMENTAL,
     "OTLP JSON mapping uses OTel GenAI-aligned attributes where applicable; collectors may require transformation.",
-    "protocol-valid \u2260 profile-complete \u2260 fixture-complete \u2014 nested path diagnostics report protocol shape only."
+    "producer-profile checks IDs, timing, and AnyValue shape; historical string enums are reader-compat warnings only."
   ];
   let parsed;
   try {
@@ -2785,11 +2836,26 @@ function validateOtlpJson(content) {
           pushPath(errors, spanPath, "must be an object");
           return;
         }
-        if (typeof span.traceId !== "string" || span.traceId.length === 0) {
-          pushPath(errors, `${spanPath}.traceId`, "must be a non-empty string");
+        if (!isHexId(span.traceId, 16)) {
+          pushPath(
+            errors,
+            `${spanPath}.traceId`,
+            "must be a 32-char hex trace id (16 bytes)"
+          );
         }
-        if (typeof span.spanId !== "string" || span.spanId.length === 0) {
-          pushPath(errors, `${spanPath}.spanId`, "must be a non-empty string");
+        if (!isHexId(span.spanId, 8)) {
+          pushPath(
+            errors,
+            `${spanPath}.spanId`,
+            "must be a 16-char hex span id (8 bytes)"
+          );
+        }
+        if (span.parentSpanId !== void 0 && !isHexId(span.parentSpanId, 8)) {
+          pushPath(
+            errors,
+            `${spanPath}.parentSpanId`,
+            "must be a 16-char hex span id when present"
+          );
         }
         if (typeof span.name !== "string") {
           pushPath(errors, `${spanPath}.name`, "must be a string");
@@ -2808,13 +2874,39 @@ function validateOtlpJson(content) {
             "must be a decimal-string 64-bit integer when present"
           );
         }
+        if (typeof span.startTimeUnixNano === "string" && /^\d+$/.test(span.startTimeUnixNano) && typeof span.endTimeUnixNano === "string" && /^\d+$/.test(span.endTimeUnixNano)) {
+          try {
+            if (BigInt(span.endTimeUnixNano) < BigInt(span.startTimeUnixNano)) {
+              pushPath(
+                errors,
+                `${spanPath}.endTimeUnixNano`,
+                "must be >= startTimeUnixNano"
+              );
+            }
+          } catch {
+            pushPath(errors, `${spanPath}.endTimeUnixNano`, "out of integer range");
+          }
+        }
+        if (Array.isArray(span.attributes)) {
+          span.attributes.forEach((attr, ai) => {
+            const attrPath = `${spanPath}.attributes[${ai}]`;
+            if (!isRecord5(attr)) {
+              pushPath(errors, attrPath, "must be an object");
+              return;
+            }
+            if (typeof attr.key !== "string") {
+              pushPath(errors, `${attrPath}.key`, "must be a string");
+            }
+            validateAnyValue(errors, `${attrPath}.value`, attr.value);
+          });
+        }
         if (typeof span.kind === "number") {
           if (!Number.isInteger(span.kind)) {
             pushPath(errors, `${spanPath}.kind`, "must be an integer SpanKind enum");
           }
         } else if (typeof span.kind === "string") {
           warnings.push(
-            `${spanPath}.kind: string SpanKind enum is legacy; prefer numeric (INTERNAL=1)`
+            `${spanPath}.kind: string SpanKind is historical reader-compat only; producer wire format uses numeric enums`
           );
         } else if (span.kind !== void 0) {
           pushPath(
@@ -2845,7 +2937,7 @@ function validateOtlpJson(content) {
               );
             } else {
               warnings.push(
-                `${spanPath}.status.code: string StatusCode is legacy; prefer numeric 0/1/2`
+                `${spanPath}.status.code: string StatusCode is historical reader-compat only; producer wire format uses numeric 0/1/2`
               );
             }
           } else {
@@ -3315,7 +3407,9 @@ Trace directory: ${traceDir}`);
   }
   let tree;
   try {
-    const trees = persistedInspectEventsToRunTrees(events);
+    const trees = persistedInspectEventsToRunTrees(events, {
+      coalesceLifecycle: true
+    });
     tree = trees.find((candidate) => candidate.runId === id) ?? trees[0];
     if (tree === void 0) {
       throw new Error("No run tree could be built from persisted events");
@@ -3721,7 +3815,7 @@ async function newestTraceMtimeMs(traceDir) {
 }
 async function loadSessionRuns(traceDir) {
   try {
-    const mod = await import('./src-KFOBXFPN.mjs');
+    const mod = await import('./src-NXAYHG34.mjs');
     const dbPath = mod.resolveIndexDbPath(traceDir);
     const status = mod.indexStatus(dbPath);
     if (!status.healthy) {
@@ -12552,13 +12646,15 @@ import { generateText, stepCountIs, tool } from "ai";
 import { MockLanguageModelV3 } from "ai/test";
 import { agentInspect } from "@agent-inspect/ai-sdk";
 import { z } from "zod";
-import { readdirSync, readFileSync } from "node:fs";
+import { mkdirSync, readdirSync, readFileSync, rmSync } from "node:fs";
 import path from "node:path";
+import { randomUUID } from "node:crypto";
 
-const TRACE_DIR = ".agent-inspect";
+const SESSION = randomUUID().slice(0, 8);
+const TRACE_DIR = path.join(".agent-inspect", \`demo-\${SESSION}\`);
 const ANSWER = "You have 2 orders";
-const CORRECT_RUN = "ai-sdk-demo-correct";
-const WRONG_RUN = "ai-sdk-demo-wrong";
+const CORRECT_RUN = \`ai-sdk-demo-correct-\${SESSION}\`;
+const WRONG_RUN = \`ai-sdk-demo-wrong-\${SESSION}\`;
 
 const usage = {
   inputTokens: { total: 4, noCache: 4, cacheRead: 0, cacheWrite: 0 },
@@ -12598,7 +12694,7 @@ async function runPath(runName, toolName) {
                   type: "tool-call",
                   toolCallId: \`call-\${toolName}\`,
                   toolName,
-                  input: { userId: "u1" },
+                  input: JSON.stringify({ userId: "u1" }),
                 },
               ],
               finishReason: { unified: "tool-calls", raw: "tool-calls" },
@@ -12650,15 +12746,39 @@ function selectRunIdByName(runName) {
   return matches[0];
 }
 
+function assertToolExecuted(runId, requiredTool) {
+  const text = readFileSync(path.join(TRACE_DIR, \`\${runId}.jsonl\`), "utf8");
+  const toolHits = text
+    .split("\\n")
+    .filter(Boolean)
+    .map((line) => JSON.parse(line))
+    .filter(
+      (ev) =>
+        (ev.kind === "TOOL" || ev.attributes?.stepType === "tool") &&
+        (ev.name === requiredTool || ev.attributes?.toolName === requiredTool),
+    );
+  if (toolHits.length === 0) {
+    throw new Error(\`expected TOOL events for \${requiredTool} in \${runId}\`);
+  }
+}
+
 async function main() {
-  await runPath(CORRECT_RUN, "lookup_orders");
-  await runPath(WRONG_RUN, "delete_orders");
-  const correctId = selectRunIdByName(CORRECT_RUN);
-  const wrongId = selectRunIdByName(WRONG_RUN);
-  console.log("Trace written to .agent-inspect/");
-  console.log(\`Correct path runId: \${correctId}\`);
-  console.log(\`Wrong path runId:   \${wrongId}\`);
-  console.log(\`Next: npx agent-inspect check \${correctId} --required-tool lookup_orders\`);
+  mkdirSync(TRACE_DIR, { recursive: true });
+  try {
+    await runPath(CORRECT_RUN, "lookup_orders");
+    await runPath(WRONG_RUN, "delete_orders");
+    const correctId = selectRunIdByName(CORRECT_RUN);
+    const wrongId = selectRunIdByName(WRONG_RUN);
+    assertToolExecuted(correctId, "lookup_orders");
+    assertToolExecuted(wrongId, "delete_orders");
+    console.log(\`Trace written to \${TRACE_DIR}/\`);
+    console.log(\`Correct path runId: \${correctId}\`);
+    console.log(\`Wrong path runId:   \${wrongId}\`);
+    console.log(\`Next: npx agent-inspect check \${correctId} --dir \${TRACE_DIR} --required-tool lookup_orders\`);
+  } catch (error) {
+    rmSync(TRACE_DIR, { recursive: true, force: true });
+    throw error;
+  }
 }
 
 main().catch((error) => {
@@ -13651,7 +13771,7 @@ function isModuleNotFound4(e) {
 }
 async function loadIndexSqlite() {
   try {
-    return await import('./src-KFOBXFPN.mjs');
+    return await import('./src-NXAYHG34.mjs');
   } catch (e) {
     if (isModuleNotFound4(e)) {
       console.error(