@agent-inspect/mcp-server 4.2.0 → 4.4.0

package/dist/index.cjs

@@ -78,6 +78,9 @@ function isTraceEvent(value) {
     case "step_completed": {
       return typeof value.runId === "string" && typeof value.stepId === "string" && (value.status === "success" || value.status === "error") && typeof value.endTime === "number" && typeof value.durationMs === "number";
     }
+    case "outcome_observed": {
+      return typeof value.runId === "string" && typeof value.outcomeId === "string" && typeof value.name === "string" && typeof value.expectation === "string" && (value.status === "passed" || value.status === "failed" || value.status === "unknown" || value.status === "skipped") && typeof value.observedAt === "number";
+    }
     default:
       return false;
   }
@@ -95,7 +98,8 @@ var INSPECT_KINDS = [
   "RESULT",
   "ERROR",
   "LOGIC",
-  "LOG"
+  "LOG",
+  "OUTCOME"
 ];
 var ATTRIBUTION_CONFIDENCES = [
   "explicit",
@@ -387,6 +391,32 @@ function fromLegacyStepCompleted(event) {
   if (error) out.error = error;
   return out;
 }
+function fromLegacyOutcomeObserved(event) {
+  const attrs = event.attributes ?? {};
+  const observedAtRaw = attrs.observedAt;
+  const observedAt = typeof observedAtRaw === "string" ? Date.parse(observedAtRaw) : typeof observedAtRaw === "number" && Number.isFinite(observedAtRaw) ? observedAtRaw : resolveTimes(event).timestamp;
+  const status = attrs.outcomeStatus;
+  const out = {
+    schemaVersion: "0.1",
+    event: "outcome_observed",
+    timestamp: observedAt,
+    runId: event.runId,
+    outcomeId: typeof attrs.outcomeId === "string" ? attrs.outcomeId : event.eventId,
+    name: event.name,
+    expectation: typeof attrs.expectation === "string" ? attrs.expectation : event.name,
+    status: status === "passed" || status === "failed" || status === "unknown" || status === "skipped" ? status : "unknown",
+    observedAt
+  };
+  if (event.parentId !== void 0) out.parentId = event.parentId;
+  if (typeof attrs.method === "string") out.method = attrs.method;
+  if (attrs.actual !== void 0) out.actual = attrs.actual;
+  if (event.outputSummary !== void 0) out.actual = event.outputSummary;
+  if (attrs.evidence !== void 0) out.evidence = attrs.evidence;
+  return out;
+}
+function fromNativeOutcome(event) {
+  return [fromLegacyOutcomeObserved(event)];
+}
 function fromNativeRun(event) {
   const { timestamp, startTime, endTime } = resolveTimes(event);
   const runStatus = mapPersistedStatusToRunStatus(event.status);
@@ -474,9 +504,13 @@ function persistedInspectEventToTraceEvents(event) {
   if (legacyEvent === "run_completed") return [fromLegacyRunCompleted(event)];
   if (legacyEvent === "step_started") return [fromLegacyStepStarted(event)];
   if (legacyEvent === "step_completed") return [fromLegacyStepCompleted(event)];
+  if (legacyEvent === "outcome_observed") return [fromLegacyOutcomeObserved(event)];
   if (event.kind === "RUN") {
     return fromNativeRun(event);
   }
+  if (event.kind === "OUTCOME") {
+    return fromNativeOutcome(event);
+  }
   return fromNativeStep(event);
 }
 function persistedInspectEventsToTraceEvents(events, options) {
@@ -711,6 +745,9 @@ function validateEvent(event) {
     case "step_completed": {
       return nonEmptyString(event.runId) && nonEmptyString(event.stepId) && (event.status === "success" || event.status === "error") && finiteNumber(event.endTime) && finiteNumber(event.durationMs) && optionalErrorInfo(event.error);
     }
+    case "outcome_observed": {
+      return nonEmptyString(event.runId) && nonEmptyString(event.outcomeId) && nonEmptyString(event.name) && nonEmptyString(event.expectation) && (event.status === "passed" || event.status === "failed" || event.status === "unknown" || event.status === "skipped") && finiteNumber(event.observedAt);
+    }
     default:
       return false;
   }
@@ -730,6 +767,55 @@ async function readTraceEventsFromFile(filePath) {
 
 // packages/core/src/context.ts
 new async_hooks.AsyncLocalStorage();
+
+// packages/core/src/outcomes/types.ts
+var OBSERVED_OUTCOME_STATUSES = [
+  "passed",
+  "failed",
+  "unknown",
+  "skipped"
+];
+var OUTCOME_LEGACY_EVENT = "outcome_observed";
+
+// packages/core/src/outcomes/validate.ts
+function parseObservedOutcomeStatus(value) {
+  const trimmed = value.trim().toLowerCase();
+  if (OBSERVED_OUTCOME_STATUSES.includes(trimmed)) {
+    return trimmed;
+  }
+  throw new Error(
+    `Unsupported observation status "${value}". Use passed, failed, unknown, or skipped.`
+  );
+}
+
+// packages/core/src/outcomes/extract.ts
+function fromOutcomeObservedEvent(event) {
+  return {
+    outcomeId: event.outcomeId,
+    runId: event.runId,
+    ...event.parentId !== void 0 ? { parentId: event.parentId } : {},
+    name: event.name,
+    expectation: event.expectation,
+    status: event.status,
+    ...event.method !== void 0 ? { method: event.method } : {},
+    ...event.actual !== void 0 ? { actual: event.actual } : {},
+    ...event.evidence !== void 0 ? { evidence: event.evidence } : {},
+    observedAt: event.observedAt
+  };
+}
+function extractOutcomesFromTraceEvents(events) {
+  const out = [];
+  for (const event of events) {
+    if (event.event === OUTCOME_LEGACY_EVENT) {
+      out.push(fromOutcomeObservedEvent(event));
+    }
+  }
+  return out.sort((a, b) => a.observedAt - b.observedAt || a.name.localeCompare(b.name));
+}
+function parseObservationFilter(value) {
+  if (value === void 0 || value.trim() === "") return void 0;
+  return parseObservedOutcomeStatus(value);
+}
 function resolveTraceDir(options = {}) {
   if (typeof options.dir === "string" && options.dir.trim() !== "") {
     return options.dir.trim();
@@ -1077,8 +1163,9 @@ async function searchTraces(metas, options) {
   }
   const limit = options.limit;
   const sessionId = options.session?.trim();
+  const observationStatus = parseObservationFilter(options.observation);
   const hasContentFilter = Boolean(
-    options.status || stepTypeFilter || nameQuery || toolQuery || durationFilter
+    options.status || stepTypeFilter || nameQuery || toolQuery || durationFilter || observationStatus
   );
   const results = [];
   const sessionLabel = sessionId && sessionId !== "" ? sessionId : void 0;
@@ -1123,6 +1210,22 @@ async function searchTraces(metas, options) {
       statusFilter: options.status
     });
     results.push(...stepMatches);
+    if (observationStatus) {
+      const outcomes = extractOutcomesFromTraceEvents(events);
+      const matched = outcomes.filter((outcome) => outcome.status === observationStatus);
+      for (const outcome of matched) {
+        results.push({
+          runId: m.runId,
+          runName: m.name,
+          runStatus: m.status,
+          stepName: outcome.name,
+          timestamp: outcome.observedAt,
+          matchReason: `observation status=${outcome.status}`,
+          matchedFields: ["outcome.status", "outcome.name"],
+          filePath: m.filePath
+        });
+      }
+    }
   }
   results.sort((a, b) => {
     const ta = a.timestamp ?? 0;
@@ -1429,7 +1532,7 @@ function summarize(findings, diagnostics) {
     errors: diagnostics.filter((item) => item.severity === "error").length
   };
 }
-function eventEvidence(event, path7) {
+function eventEvidence(event, path8) {
   return {
     runId: event.runId,
     eventId: event.eventId,
@@ -1767,13 +1870,13 @@ function pairSteps(left, right) {
   return pairs;
 }
 function compareLeafSteps(L, R, segments, opts, out) {
-  const path7 = buildPath(segments);
+  const path8 = buildPath(segments);
   if (L.name !== R.name) {
     out.push({
       kind: "structure",
       severity: "warning",
       message: "Step name differs",
-      path: path7,
+      path: path8,
       left: L.name,
       right: R.name
     });
@@ -1783,7 +1886,7 @@ function compareLeafSteps(L, R, segments, opts, out) {
       kind: "step-type",
       severity: "warning",
       message: "Step type differs",
-      path: path7,
+      path: path8,
       left: L.type,
       right: R.type
     });
@@ -1793,7 +1896,7 @@ function compareLeafSteps(L, R, segments, opts, out) {
       kind: "step-status",
       severity: "warning",
       message: "Step status differs",
-      path: path7,
+      path: path8,
       left: L.status,
       right: R.status
     });
@@ -1805,7 +1908,7 @@ function compareLeafSteps(L, R, segments, opts, out) {
       kind: "error",
       severity: "error",
       message: "Step error message differs",
-      path: path7,
+      path: path8,
       left: le || void 0,
       right: re || void 0
     });
@@ -1823,7 +1926,7 @@ function compareLeafSteps(L, R, segments, opts, out) {
         kind: "duration",
         severity: "info",
         message: "Step duration differs",
-        path: path7,
+        path: path8,
         left: ld,
         right: rd
       });
@@ -1836,7 +1939,7 @@ function compareLeafSteps(L, R, segments, opts, out) {
       kind: "metadata",
       severity: "info",
       message: "Step metadata differs",
-      path: path7,
+      path: path8,
       left: L.metadata,
       right: R.metadata
     });
@@ -1848,7 +1951,7 @@ function compareLeafSteps(L, R, segments, opts, out) {
       kind: "output",
       severity: "info",
       message: "Output preview differs",
-      path: path7,
+      path: path8,
       left: L.outputPreview,
       right: R.outputPreview
     });
@@ -2145,6 +2248,8 @@ function nodeIdForEvent(event) {
     case "step_started":
     case "step_completed":
       return event.stepId;
+    case "outcome_observed":
+      return event.outcomeId;
     default:
       return "unknown";
   }
@@ -2328,6 +2433,39 @@ function traceEventToPersistedInspectEvent(event, options) {
         error
       };
     }
+    case "outcome_observed": {
+      const tsObserved = toIsoTimestamp(event.observedAt);
+      const attributes = compactAttributes2({
+        legacyEvent: "outcome_observed",
+        outcomeId: event.outcomeId,
+        outcomeStatus: event.status,
+        expectation: event.expectation,
+        method: event.method,
+        actual: event.actual,
+        evidence: event.evidence,
+        observedAt: tsObserved.iso,
+        invalidTimestamp: tsMain.invalidTimestamp || tsObserved.invalidTimestamp ? true : void 0
+      });
+      const out = {
+        schemaVersion: "0.2",
+        eventId,
+        runId: event.runId,
+        kind: "OUTCOME",
+        name: event.name,
+        status: event.status === "failed" ? "error" : "ok",
+        timestamp: tsMain.iso,
+        confidence: "explicit",
+        source,
+        attributes
+      };
+      if (event.parentId !== void 0) {
+        out.parentId = event.parentId;
+      }
+      if (event.actual !== void 0) {
+        out.outputSummary = event.actual;
+      }
+      return out;
+    }
     default: {
       const _exhaustive = event;
       throw new Error(`Unsupported trace event: ${_exhaustive.event}`);
@@ -3057,7 +3195,7 @@ function sanitizeOpenInferenceAttributes(attributes, pathPrefix) {
 function mapOpenInferenceKind(span, attributes, pathPrefix) {
   const warnings = [];
   const agentInspectKind = attributes["agent_inspect.kind"];
-  if (agentInspectKind === "RUN" || agentInspectKind === "AGENT" || agentInspectKind === "LLM" || agentInspectKind === "TOOL" || agentInspectKind === "CHAIN" || agentInspectKind === "RETRIEVER" || agentInspectKind === "DECISION" || agentInspectKind === "RESULT" || agentInspectKind === "ERROR" || agentInspectKind === "LOGIC" || agentInspectKind === "LOG") {
+  if (agentInspectKind === "RUN" || agentInspectKind === "AGENT" || agentInspectKind === "LLM" || agentInspectKind === "TOOL" || agentInspectKind === "CHAIN" || agentInspectKind === "RETRIEVER" || agentInspectKind === "DECISION" || agentInspectKind === "RESULT" || agentInspectKind === "ERROR" || agentInspectKind === "LOGIC" || agentInspectKind === "LOG" || agentInspectKind === "OUTCOME") {
     return { kind: agentInspectKind, warnings };
   }
   const rawKind = readStringField(span, ["kind", "span_kind", "spanKind"]) ?? (typeof attributes["openinference.span.kind"] === "string" ? attributes["openinference.span.kind"] : void 0);
@@ -3578,7 +3716,7 @@ function mapOtlpStatus(status) {
 function readOtlpKind(attributes, pathPrefix) {
   const warnings = [];
   const agentInspectKind = attributes["agent_inspect.kind"];
-  if (agentInspectKind === "RUN" || agentInspectKind === "AGENT" || agentInspectKind === "LLM" || agentInspectKind === "TOOL" || agentInspectKind === "CHAIN" || agentInspectKind === "RETRIEVER" || agentInspectKind === "DECISION" || agentInspectKind === "RESULT" || agentInspectKind === "ERROR" || agentInspectKind === "LOGIC" || agentInspectKind === "LOG") {
+  if (agentInspectKind === "RUN" || agentInspectKind === "AGENT" || agentInspectKind === "LLM" || agentInspectKind === "TOOL" || agentInspectKind === "CHAIN" || agentInspectKind === "RETRIEVER" || agentInspectKind === "DECISION" || agentInspectKind === "RESULT" || agentInspectKind === "ERROR" || agentInspectKind === "LOGIC" || agentInspectKind === "LOG" || agentInspectKind === "OUTCOME") {
     return { kind: agentInspectKind, warnings };
   }
   const operation = attributes["gen_ai.operation.name"];