agent-inspect 3.2.0 → 3.4.0

package/packages/core/dist/checks.d.cts

@@ -171,6 +171,19 @@ interface RunStatusRuleOptions {
 interface RunDurationRuleOptions {
     maxDurationMs: number;
 }
+/**
+ * Experimental options for the built-in max step duration rule.
+ */
+interface MaxStepDurationRuleOptions {
+    maxDurationMs: number;
+}
+/**
+ * Experimental options for stall detection (running / incomplete events).
+ */
+interface StallDetectionRuleOptions {
+    /** When true, events with startedAt but no endedAt also fail. */
+    requireEndedAt?: boolean;
+}
 /**
  * Experimental options for the built-in event count rule.
  *
@@ -434,6 +447,18 @@ declare function createRunStatusRule(options?: RunStatusRuleOptions): TraceCheck
  * evolve during the v1.x experimental period.
  */
 declare function createRunDurationRule(options: RunDurationRuleOptions): TraceCheckRule;
+/**
+ * Fail when any step duration exceeds the configured maximum.
+ */
+declare function createMaxStepDurationRule(options: MaxStepDurationRuleOptions): TraceCheckRule;
+/**
+ * Detect stalled runs: running events and optionally started-but-unended events.
+ */
+declare function createStallDetectionRule(options?: StallDetectionRuleOptions): TraceCheckRule;
+/**
+ * Require a completed run with no running events.
+ */
+declare function createRequireCompletedRule(): TraceCheckRule;
 /**
  * Create the experimental built-in event count rule.
  *
@@ -582,4 +607,4 @@ declare function createBaselineRegressionRule(options: BaselineRegressionRuleOpt
  */
 declare function runTraceChecks(input: TraceCheckInput, options?: RunTraceChecksOptions): TraceCheckResult;
 
-export { type BaselineRegressionRuleOptions, type DecisionRuleOptions, type GuardrailRuleOptions, type LlmUsageRuleOptions, type RetrievalRuleOptions, type RunDepthRuleOptions, type RunDurationRuleOptions, type RunEventCountRuleOptions, type RunStatusRuleOptions, type RunTraceChecksOptions, type SafetyOversizedAttributeRuleOptions, type SafetyRawContentRuleOptions, type SafetyRedactionRuleOptions, type SafetySecretPattern, type SafetySecretPatternRuleOptions, type StructureIncompleteRuleOptions, type StructureOrphanRuleOptions, type StructureParallelWidthRuleOptions, type StructureRelationshipRuleOptions, type ToolFailureRuleOptions, type ToolOrderingRuleOptions, type ToolUsageRuleOptions, type TraceCheckContext, type TraceCheckDiagnostic, type TraceCheckDiagnosticCode, type TraceCheckEvidence, type TraceCheckFacts, type TraceCheckFinding, type TraceCheckFindingStatus, type TraceCheckInput, type TraceCheckResult, type TraceCheckRule, type TraceCheckRuleCategory, type TraceCheckSeverity, type TraceCheckStatus, type TraceCheckSummary, type TraceSignalRuleOptions, createBaselineRegressionRule, createDecisionRule, createGuardrailRule, createLlmUsageRule, createRetrievalRule, createRunDepthRule, createRunDurationRule, createRunEventCountRule, createRunStatusRule, createSafetyOversizedAttributeRule, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createStructureCycleRule, createStructureIncompleteRule, createStructureOrphanRule, createStructureParallelWidthRule, createStructureRelationshipRule, createToolFailureRule, createToolOrderingRule, createToolUsageRule, runTraceChecks };
+export { type BaselineRegressionRuleOptions, type DecisionRuleOptions, type GuardrailRuleOptions, type LlmUsageRuleOptions, type MaxStepDurationRuleOptions, type RetrievalRuleOptions, type RunDepthRuleOptions, type RunDurationRuleOptions, type RunEventCountRuleOptions, type RunStatusRuleOptions, type RunTraceChecksOptions, type SafetyOversizedAttributeRuleOptions, type SafetyRawContentRuleOptions, type SafetyRedactionRuleOptions, type SafetySecretPattern, type SafetySecretPatternRuleOptions, type StallDetectionRuleOptions, type StructureIncompleteRuleOptions, type StructureOrphanRuleOptions, type StructureParallelWidthRuleOptions, type StructureRelationshipRuleOptions, type ToolFailureRuleOptions, type ToolOrderingRuleOptions, type ToolUsageRuleOptions, type TraceCheckContext, type TraceCheckDiagnostic, type TraceCheckDiagnosticCode, type TraceCheckEvidence, type TraceCheckFacts, type TraceCheckFinding, type TraceCheckFindingStatus, type TraceCheckInput, type TraceCheckResult, type TraceCheckRule, type TraceCheckRuleCategory, type TraceCheckSeverity, type TraceCheckStatus, type TraceCheckSummary, type TraceSignalRuleOptions, createBaselineRegressionRule, createDecisionRule, createGuardrailRule, createLlmUsageRule, createMaxStepDurationRule, createRequireCompletedRule, createRetrievalRule, createRunDepthRule, createRunDurationRule, createRunEventCountRule, createRunStatusRule, createSafetyOversizedAttributeRule, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createStallDetectionRule, createStructureCycleRule, createStructureIncompleteRule, createStructureOrphanRule, createStructureParallelWidthRule, createStructureRelationshipRule, createToolFailureRule, createToolOrderingRule, createToolUsageRule, runTraceChecks };

package/packages/core/dist/checks.d.ts

@@ -171,6 +171,19 @@ interface RunStatusRuleOptions {
 interface RunDurationRuleOptions {
     maxDurationMs: number;
 }
+/**
+ * Experimental options for the built-in max step duration rule.
+ */
+interface MaxStepDurationRuleOptions {
+    maxDurationMs: number;
+}
+/**
+ * Experimental options for stall detection (running / incomplete events).
+ */
+interface StallDetectionRuleOptions {
+    /** When true, events with startedAt but no endedAt also fail. */
+    requireEndedAt?: boolean;
+}
 /**
  * Experimental options for the built-in event count rule.
  *
@@ -434,6 +447,18 @@ declare function createRunStatusRule(options?: RunStatusRuleOptions): TraceCheck
  * evolve during the v1.x experimental period.
  */
 declare function createRunDurationRule(options: RunDurationRuleOptions): TraceCheckRule;
+/**
+ * Fail when any step duration exceeds the configured maximum.
+ */
+declare function createMaxStepDurationRule(options: MaxStepDurationRuleOptions): TraceCheckRule;
+/**
+ * Detect stalled runs: running events and optionally started-but-unended events.
+ */
+declare function createStallDetectionRule(options?: StallDetectionRuleOptions): TraceCheckRule;
+/**
+ * Require a completed run with no running events.
+ */
+declare function createRequireCompletedRule(): TraceCheckRule;
 /**
  * Create the experimental built-in event count rule.
  *
@@ -582,4 +607,4 @@ declare function createBaselineRegressionRule(options: BaselineRegressionRuleOpt
  */
 declare function runTraceChecks(input: TraceCheckInput, options?: RunTraceChecksOptions): TraceCheckResult;
 
-export { type BaselineRegressionRuleOptions, type DecisionRuleOptions, type GuardrailRuleOptions, type LlmUsageRuleOptions, type RetrievalRuleOptions, type RunDepthRuleOptions, type RunDurationRuleOptions, type RunEventCountRuleOptions, type RunStatusRuleOptions, type RunTraceChecksOptions, type SafetyOversizedAttributeRuleOptions, type SafetyRawContentRuleOptions, type SafetyRedactionRuleOptions, type SafetySecretPattern, type SafetySecretPatternRuleOptions, type StructureIncompleteRuleOptions, type StructureOrphanRuleOptions, type StructureParallelWidthRuleOptions, type StructureRelationshipRuleOptions, type ToolFailureRuleOptions, type ToolOrderingRuleOptions, type ToolUsageRuleOptions, type TraceCheckContext, type TraceCheckDiagnostic, type TraceCheckDiagnosticCode, type TraceCheckEvidence, type TraceCheckFacts, type TraceCheckFinding, type TraceCheckFindingStatus, type TraceCheckInput, type TraceCheckResult, type TraceCheckRule, type TraceCheckRuleCategory, type TraceCheckSeverity, type TraceCheckStatus, type TraceCheckSummary, type TraceSignalRuleOptions, createBaselineRegressionRule, createDecisionRule, createGuardrailRule, createLlmUsageRule, createRetrievalRule, createRunDepthRule, createRunDurationRule, createRunEventCountRule, createRunStatusRule, createSafetyOversizedAttributeRule, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createStructureCycleRule, createStructureIncompleteRule, createStructureOrphanRule, createStructureParallelWidthRule, createStructureRelationshipRule, createToolFailureRule, createToolOrderingRule, createToolUsageRule, runTraceChecks };
+export { type BaselineRegressionRuleOptions, type DecisionRuleOptions, type GuardrailRuleOptions, type LlmUsageRuleOptions, type MaxStepDurationRuleOptions, type RetrievalRuleOptions, type RunDepthRuleOptions, type RunDurationRuleOptions, type RunEventCountRuleOptions, type RunStatusRuleOptions, type RunTraceChecksOptions, type SafetyOversizedAttributeRuleOptions, type SafetyRawContentRuleOptions, type SafetyRedactionRuleOptions, type SafetySecretPattern, type SafetySecretPatternRuleOptions, type StallDetectionRuleOptions, type StructureIncompleteRuleOptions, type StructureOrphanRuleOptions, type StructureParallelWidthRuleOptions, type StructureRelationshipRuleOptions, type ToolFailureRuleOptions, type ToolOrderingRuleOptions, type ToolUsageRuleOptions, type TraceCheckContext, type TraceCheckDiagnostic, type TraceCheckDiagnosticCode, type TraceCheckEvidence, type TraceCheckFacts, type TraceCheckFinding, type TraceCheckFindingStatus, type TraceCheckInput, type TraceCheckResult, type TraceCheckRule, type TraceCheckRuleCategory, type TraceCheckSeverity, type TraceCheckStatus, type TraceCheckSummary, type TraceSignalRuleOptions, createBaselineRegressionRule, createDecisionRule, createGuardrailRule, createLlmUsageRule, createMaxStepDurationRule, createRequireCompletedRule, createRetrievalRule, createRunDepthRule, createRunDurationRule, createRunEventCountRule, createRunStatusRule, createSafetyOversizedAttributeRule, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createStallDetectionRule, createStructureCycleRule, createStructureIncompleteRule, createStructureOrphanRule, createStructureParallelWidthRule, createStructureRelationshipRule, createToolFailureRule, createToolOrderingRule, createToolUsageRule, runTraceChecks };

package/packages/core/dist/checks.mjs

@@ -578,6 +578,108 @@ function createRunDurationRule(options) {
     }
   };
 }
+function createMaxStepDurationRule(options) {
+  return {
+    id: "run.maxStepDuration",
+    category: "run",
+    defaultSeverity: "error",
+    evaluate(context) {
+      const over = context.events.filter((event) => {
+        const duration = eventDurationMs(event);
+        return duration !== void 0 && duration > options.maxDurationMs;
+      });
+      if (over.length === 0) return [];
+      return [
+        failFinding(
+          "run.maxStepDuration",
+          `${over.length} step(s) exceeded max duration ${options.maxDurationMs}ms.`,
+          over.map((event) => eventEvidence(event, "durationMs")),
+          { maxDurationMs: options.maxDurationMs },
+          over.map((event) => ({
+            eventId: event.eventId,
+            name: event.name,
+            durationMs: eventDurationMs(event)
+          }))
+        )
+      ];
+    }
+  };
+}
+function createStallDetectionRule(options = {}) {
+  const requireEndedAt = options.requireEndedAt === true;
+  return {
+    id: "run.stall",
+    category: "run",
+    defaultSeverity: "warning",
+    evaluate(context) {
+      const findings = [];
+      const running = context.events.filter((event) => event.status === "running");
+      if (running.length > 0) {
+        findings.push(
+          failFinding(
+            "run.stall",
+            `Found ${running.length} event(s) still running (possible stall).`,
+            running.map((event) => eventEvidence(event, "status")),
+            "no running events",
+            running.length
+          )
+        );
+      }
+      if (requireEndedAt) {
+        const incomplete = context.events.filter(
+          (event) => event.startedAt !== void 0 && event.endedAt === void 0 && event.status !== "running"
+        );
+        if (incomplete.length > 0) {
+          findings.push(
+            failFinding(
+              "run.stall",
+              `Found ${incomplete.length} started event(s) without endedAt.`,
+              incomplete.map((event) => eventEvidence(event, "endedAt")),
+              "endedAt for started events",
+              incomplete.length
+            )
+          );
+        }
+      }
+      return findings;
+    }
+  };
+}
+function createRequireCompletedRule() {
+  return {
+    id: "run.requireCompleted",
+    category: "run",
+    defaultSeverity: "error",
+    evaluate(context) {
+      const findings = [];
+      const runStatus = context.selectedRun?.status;
+      if (runStatus === "running") {
+        findings.push(
+          failFinding(
+            "run.requireCompleted",
+            "Run is still running.",
+            runEvidence(context.selectedRun),
+            "completed run",
+            runStatus
+          )
+        );
+      }
+      const running = context.events.filter((event) => event.status === "running");
+      if (running.length > 0) {
+        findings.push(
+          failFinding(
+            "run.requireCompleted",
+            `Run has ${running.length} incomplete running event(s).`,
+            running.map((event) => eventEvidence(event, "status")),
+            "no running events",
+            running.length
+          )
+        );
+      }
+      return findings;
+    }
+  };
+}
 function createRunEventCountRule(options) {
   return {
     id: "run.eventCount",
@@ -1507,6 +1609,6 @@ function runTraceChecks(input, options = {}) {
   };
 }
 
-export { createBaselineRegressionRule, createDecisionRule, createGuardrailRule, createLlmUsageRule, createRetrievalRule, createRunDepthRule, createRunDurationRule, createRunEventCountRule, createRunStatusRule, createSafetyOversizedAttributeRule, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createStructureCycleRule, createStructureIncompleteRule, createStructureOrphanRule, createStructureParallelWidthRule, createStructureRelationshipRule, createToolFailureRule, createToolOrderingRule, createToolUsageRule, runTraceChecks };
+export { createBaselineRegressionRule, createDecisionRule, createGuardrailRule, createLlmUsageRule, createMaxStepDurationRule, createRequireCompletedRule, createRetrievalRule, createRunDepthRule, createRunDurationRule, createRunEventCountRule, createRunStatusRule, createSafetyOversizedAttributeRule, createSafetyRawContentRule, createSafetyRedactionRule, createSafetySecretPatternRule, createStallDetectionRule, createStructureCycleRule, createStructureIncompleteRule, createStructureOrphanRule, createStructureParallelWidthRule, createStructureRelationshipRule, createToolFailureRule, createToolOrderingRule, createToolUsageRule, runTraceChecks };
 //# sourceMappingURL=checks.mjs.map
 //# sourceMappingURL=checks.mjs.map