agent-inspect 6.31.9 → 6.31.11

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (61) hide show
  1. package/CHANGELOG.md +12 -0
  2. package/README.md +1 -1
  3. package/docs/NESTJS.md +56 -18
  4. package/package.json +1 -1
  5. package/packages/cli/dist/{chunk-SWUZRRDT.mjs → chunk-MQYHJL3W.mjs} +2216 -1938
  6. package/packages/cli/dist/chunk-MQYHJL3W.mjs.map +1 -0
  7. package/packages/cli/dist/index.cjs +975 -671
  8. package/packages/cli/dist/index.cjs.map +1 -1
  9. package/packages/cli/dist/index.mjs +32 -9
  10. package/packages/cli/dist/index.mjs.map +1 -1
  11. package/packages/cli/dist/{src-XFIPYODW.mjs → src-76R6E2LQ.mjs} +3 -3
  12. package/packages/cli/dist/{src-XFIPYODW.mjs.map → src-76R6E2LQ.mjs.map} +1 -1
  13. package/packages/core/dist/advanced.cjs +2079 -1879
  14. package/packages/core/dist/advanced.cjs.map +1 -1
  15. package/packages/core/dist/advanced.d.cts +56 -4
  16. package/packages/core/dist/advanced.d.ts +56 -4
  17. package/packages/core/dist/advanced.mjs +188 -23
  18. package/packages/core/dist/advanced.mjs.map +1 -1
  19. package/packages/core/dist/checks.cjs +126 -9
  20. package/packages/core/dist/checks.cjs.map +1 -1
  21. package/packages/core/dist/checks.d.cts +3 -3
  22. package/packages/core/dist/checks.d.ts +3 -3
  23. package/packages/core/dist/checks.mjs +3 -2
  24. package/packages/core/dist/{chunk-L5WLVBW4.mjs → chunk-AHUZNHFI.mjs} +4 -55
  25. package/packages/core/dist/chunk-AHUZNHFI.mjs.map +1 -0
  26. package/packages/core/dist/{chunk-2CCRUI42.mjs → chunk-CCYKNAOD.mjs} +119 -12
  27. package/packages/core/dist/chunk-CCYKNAOD.mjs.map +1 -0
  28. package/packages/core/dist/{chunk-7YYS742B.mjs → chunk-M3RJD3BP.mjs} +5 -5
  29. package/packages/core/dist/{chunk-7YYS742B.mjs.map → chunk-M3RJD3BP.mjs.map} +1 -1
  30. package/packages/core/dist/{chunk-2IRRUNIN.mjs → chunk-OAUNS5CS.mjs} +3 -3
  31. package/packages/core/dist/{chunk-2IRRUNIN.mjs.map → chunk-OAUNS5CS.mjs.map} +1 -1
  32. package/packages/core/dist/{chunk-IJJSPJXV.mjs → chunk-QOC3PJO3.mjs} +4 -4
  33. package/packages/core/dist/{chunk-IJJSPJXV.mjs.map → chunk-QOC3PJO3.mjs.map} +1 -1
  34. package/packages/core/dist/chunk-S6LQBLX5.mjs +67 -0
  35. package/packages/core/dist/chunk-S6LQBLX5.mjs.map +1 -0
  36. package/packages/core/dist/{chunk-3SPBELJB.mjs → chunk-SSG3ICI4.mjs} +16 -2
  37. package/packages/core/dist/chunk-SSG3ICI4.mjs.map +1 -0
  38. package/packages/core/dist/exporters.cjs +3 -3
  39. package/packages/core/dist/exporters.cjs.map +1 -1
  40. package/packages/core/dist/exporters.mjs +1 -1
  41. package/packages/core/dist/{index-EQihlEei.d.cts → index-DACU4Dy1.d.cts} +12 -0
  42. package/packages/core/dist/{index-CgOAGuM5.d.ts → index-b4tIGrzq.d.ts} +12 -0
  43. package/packages/core/dist/index.cjs +77 -0
  44. package/packages/core/dist/index.cjs.map +1 -1
  45. package/packages/core/dist/index.mjs +3 -2
  46. package/packages/core/dist/index.mjs.map +1 -1
  47. package/packages/core/dist/persisted.cjs +11 -0
  48. package/packages/core/dist/persisted.cjs.map +1 -1
  49. package/packages/core/dist/persisted.mjs +4 -3
  50. package/packages/core/dist/persisted.mjs.map +1 -1
  51. package/packages/core/dist/{programmatic-gZTWLObB.d.cts → programmatic-BLlC7Ytq.d.cts} +1 -1
  52. package/packages/core/dist/{programmatic-gZTWLObB.d.ts → programmatic-BLlC7Ytq.d.ts} +1 -1
  53. package/packages/core/dist/readers.cjs +11 -0
  54. package/packages/core/dist/readers.cjs.map +1 -1
  55. package/packages/core/dist/readers.d.cts +1 -1
  56. package/packages/core/dist/readers.d.ts +1 -1
  57. package/packages/core/dist/readers.mjs +4 -3
  58. package/packages/cli/dist/chunk-SWUZRRDT.mjs.map +0 -1
  59. package/packages/core/dist/chunk-2CCRUI42.mjs.map +0 -1
  60. package/packages/core/dist/chunk-3SPBELJB.mjs.map +0 -1
  61. package/packages/core/dist/chunk-L5WLVBW4.mjs.map +0 -1
@@ -2868,7 +2868,7 @@ function stepMixLine(summary) {
2868
2868
  if (summary.logicSteps > 0) parts.push(`${summary.logicSteps} logic`);
2869
2869
  return parts.length > 0 ? parts.join(", ") : "none";
2870
2870
  }
2871
- function outcomeLine(summary) {
2871
+ function executionLine(summary) {
2872
2872
  if (summary.status === "success") {
2873
2873
  return summary.errorSteps > 0 ? "Completed with step errors recorded." : "Completed successfully.";
2874
2874
  }
@@ -2886,7 +2886,7 @@ function outcomeLine(summary) {
2886
2886
  if (summary.status === "running") {
2887
2887
  return "Run is still in progress (no run_completed).";
2888
2888
  }
2889
- return "Outcome unknown \u2014 inspect events may be incomplete.";
2889
+ return "Execution status unknown \u2014 inspect events may be incomplete.";
2890
2890
  }
2891
2891
  function buildRunWhatSummary(events) {
2892
2892
  const base = buildRunSummary(events);
@@ -2977,7 +2977,7 @@ function renderRunWhat(summary, options = {}) {
2977
2977
  lines.push(`Correlation: ${parts.join(", ")}`);
2978
2978
  }
2979
2979
  }
2980
- lines.push(`Outcome: ${outcomeLine(summary)}`);
2980
+ lines.push(`Execution: ${executionLine(summary)}`);
2981
2981
  if (summary.longestStep && summary.totalSteps > 0) {
2982
2982
  lines.push(
2983
2983
  `Slowest: ${summary.longestStep.name} (${formatDuration2(summary.longestStep.durationMs)}, ${summary.longestStep.type})`
@@ -3003,163 +3003,570 @@ var init_causal_failure = __esm({
3003
3003
  }
3004
3004
  });
3005
3005
 
3006
- // packages/core/src/explain.ts
3007
- function toDisplayStatus(status) {
3008
- if (status === "ok") return "success";
3009
- if (status === "error") return "error";
3010
- if (status === "running") return "running";
3011
- return "unknown";
3006
+ // packages/core/src/diagnostics/programmatic.ts
3007
+ function formatProgrammaticDiagnostic(code, detail) {
3008
+ const spec = PROGRAMMATIC_DIAGNOSTIC_SPECS[code];
3009
+ const summary = detail?.trim() ? detail.trim() : spec.summary;
3010
+ return `${code}: ${summary} Remediation: ${spec.remediation}`;
3012
3011
  }
3013
- function flatten(nodes, out = []) {
3014
- for (const node of nodes) {
3015
- out.push({ node, index: out.length + 1 });
3016
- flatten(node.children, out);
3012
+ var PROGRAMMATIC_DIAGNOSTIC_SPECS;
3013
+ var init_programmatic = __esm({
3014
+ "packages/core/src/diagnostics/programmatic.ts"() {
3015
+ PROGRAMMATIC_DIAGNOSTIC_SPECS = Object.freeze({
3016
+ AI_TRACE_INPUT_INVALID: {
3017
+ code: "AI_TRACE_INPUT_INVALID",
3018
+ summary: 'Expected { type: "file", path }, { type: "directory", path }, { type: "string", content }, { type: "buffer", content }, or { type: "stdin" }.',
3019
+ remediation: "For a file path, use openTraceFile(path).",
3020
+ relatedCodes: ["invalid_input"]
3021
+ },
3022
+ AI_TRACE_FORMAT_UNSUPPORTED: {
3023
+ code: "AI_TRACE_FORMAT_UNSUPPORTED",
3024
+ summary: "No trace reader could detect the input format.",
3025
+ remediation: "Pass an AgentInspect JSONL file via openTraceFile, or set options.format to a registered reader.",
3026
+ relatedCodes: ["unsupported_format"]
3027
+ },
3028
+ AI_TRACE_FORMAT_AMBIGUOUS: {
3029
+ code: "AI_TRACE_FORMAT_AMBIGUOUS",
3030
+ summary: "Multiple trace readers matched the input with equal confidence.",
3031
+ remediation: "Set options.format explicitly to disambiguate the reader.",
3032
+ relatedCodes: ["ambiguous_format"]
3033
+ },
3034
+ AI_TRACE_FACTS_INPUT_NOT_NORMALIZED: {
3035
+ code: "AI_TRACE_FACTS_INPUT_NOT_NORMALIZED",
3036
+ summary: "TraceFacts requires TraceReadResult or PersistedInspectEvent[].",
3037
+ remediation: "Use openTraceFile() to normalize a JSONL trace first."
3038
+ },
3039
+ AI_TRACE_CONTRACT_RUN_SELECTION_REQUIRED: {
3040
+ code: "AI_TRACE_CONTRACT_RUN_SELECTION_REQUIRED",
3041
+ summary: "Multiple runs are available; select a run before executing checks.",
3042
+ remediation: "Pass options.runId or TraceCheckInput.selectedRun.",
3043
+ relatedCodes: ["AI_CHECK_RUN_SELECTION_REQUIRED"]
3044
+ },
3045
+ AI_TRACE_RELATIONSHIP_SELF_PARENT: {
3046
+ code: "AI_TRACE_RELATIONSHIP_SELF_PARENT",
3047
+ summary: "A parentId equals its own event/step id.",
3048
+ remediation: "Reject at capture (AI_LANGGRAPH_SELF_PARENT_REJECTED) or drop via logical projection; do not invent replacement parents.",
3049
+ relatedCodes: [
3050
+ "AI_LANGGRAPH_SELF_PARENT_REJECTED",
3051
+ "AI_LOGICAL_SELF_PARENT_REMOVED"
3052
+ ]
3053
+ },
3054
+ AI_TRACE_RELATIONSHIP_CYCLE: {
3055
+ code: "AI_TRACE_RELATIONSHIP_CYCLE",
3056
+ summary: "Trace contains a parentId cycle.",
3057
+ remediation: "Use visibility-first tree linking for legacy fixtures; prefer acyclic capture for new adapter output.",
3058
+ relatedCodes: ["structure.cycle"]
3059
+ },
3060
+ AI_TRACE_CONTRACT_INPUT_INVALID: {
3061
+ code: "AI_TRACE_CONTRACT_INPUT_INVALID",
3062
+ summary: "evaluateTraceContract expects ({ read }, contract) or use evaluateTraceContractRead(read, contract).",
3063
+ remediation: "Pass TraceCheckInput as the first argument: evaluateTraceContract({ read }, contract).",
3064
+ relatedCodes: ["AI_CHECK_INVALID_ARGUMENTS"]
3065
+ },
3066
+ AI_INSPECT_RUN_UNSUPPORTED_OPTION: {
3067
+ code: "AI_INSPECT_RUN_UNSUPPORTED_OPTION",
3068
+ summary: 'inspectRun does not accept a "writer" option.',
3069
+ remediation: "Use createInspector({ writer }) then inspector.run(...). inspectRun persists via the global JSONL path (schema 0.1)."
3070
+ }
3071
+ });
3017
3072
  }
3018
- return out;
3073
+ });
3074
+
3075
+ // packages/core/src/checks/logical-events.ts
3076
+ function isRecord6(value) {
3077
+ return typeof value === "object" && value !== null && !Array.isArray(value);
3019
3078
  }
3020
- function redactValue(redactor, key, value) {
3021
- return redactor.redactValue(key, value);
3079
+ function legacyEvent(event) {
3080
+ const value = event.attributes?.legacyEvent;
3081
+ return typeof value === "string" ? value : void 0;
3022
3082
  }
3023
- function fact(id, label, value, redactor) {
3083
+ function stepIdOf(event) {
3084
+ const value = event.attributes?.stepId;
3085
+ if (typeof value === "string" && value.trim() !== "") return value;
3086
+ return void 0;
3087
+ }
3088
+ function cloneEvent(event) {
3024
3089
  return {
3025
- id,
3026
- label,
3027
- value: redactValue(redactor, id.split(".").at(-1) ?? id, value),
3028
- source: "trace",
3029
- confidence: "observed"
3090
+ ...event,
3091
+ ...event.attributes !== void 0 ? { attributes: { ...event.attributes } } : {},
3092
+ ...event.error !== void 0 ? { error: { ...event.error } } : {},
3093
+ ...event.tokenUsage !== void 0 ? { tokenUsage: { ...event.tokenUsage } } : {},
3094
+ ...event.source !== void 0 ? { source: { ...event.source } } : {}
3030
3095
  };
3031
3096
  }
3032
- function topKinds(run) {
3033
- return Object.entries(run.metadata.kinds).filter(([, count]) => count > 0).sort((a, b) => {
3034
- if (b[1] !== a[1]) return b[1] - a[1];
3035
- return a[0].localeCompare(b[0]);
3036
- }).slice(0, 5).map(([kind, count]) => `${kind}:${count}`);
3037
- }
3038
- function countErrorNodes(nodes) {
3039
- return nodes.filter((entry) => entry.node.event.status === "error").length;
3040
- }
3041
- function slowestNode(nodes) {
3042
- return nodes.filter(
3043
- (entry) => entry.node.event.durationMs !== void 0 && // The RUN boundary spans the whole run, so its duration always dominates
3044
- // and shadows every real step. Rank actual work nodes only, matching the
3045
- // stats slowest-step ranking, which never counts the run envelope.
3046
- entry.node.event.kind !== "RUN"
3047
- ).sort((a, b) => {
3048
- const delta = (b.node.event.durationMs ?? 0) - (a.node.event.durationMs ?? 0);
3049
- return delta !== 0 ? delta : a.index - b.index;
3050
- })[0];
3051
- }
3052
- function attributeFacts(nodes, redactor) {
3053
- const facts = [];
3054
- for (const entry of nodes) {
3055
- const attrs = entry.node.event.attributes;
3056
- if (attrs === void 0) continue;
3057
- for (const key of Object.keys(attrs).sort()) {
3058
- facts.push({
3059
- id: `node.${entry.index}.attributes.${key}`,
3060
- label: `${entry.node.event.name} attribute ${key}`,
3061
- value: redactValue(redactor, key, attrs[key]),
3062
- source: "trace",
3063
- confidence: "observed"
3064
- });
3065
- if (facts.length >= 8) return facts;
3066
- }
3067
- }
3068
- return facts;
3069
- }
3070
- function buildFacts(run, redactor) {
3071
- const nodes = flatten(run.children);
3072
- const facts = [
3073
- fact("run.id", "Run id", run.runId, redactor),
3074
- fact("run.name", "Run name", run.name ?? run.runId, redactor),
3075
- fact("run.status", "Run status", toDisplayStatus(run.status), redactor),
3076
- fact("run.totalEvents", "Total events", run.metadata.totalEvents, redactor),
3077
- fact("run.stepCount", "Top-level step count", run.children.length, redactor),
3078
- fact("run.nodeCount", "Total node count", nodes.length, redactor),
3079
- fact("run.errorNodeCount", "Error node count", countErrorNodes(nodes), redactor),
3080
- fact("run.kinds", "Observed kind mix", topKinds(run), redactor)
3081
- ];
3082
- if (run.durationMs !== void 0) {
3083
- facts.push(fact("run.durationMs", "Run duration milliseconds", run.durationMs, redactor));
3084
- }
3085
- const slowest = slowestNode(nodes);
3086
- if (slowest !== void 0) {
3087
- facts.push(
3088
- fact("run.slowestNode", "Slowest observed node", {
3089
- name: slowest.node.event.name,
3090
- kind: slowest.node.event.kind,
3091
- durationMs: slowest.node.event.durationMs
3092
- }, redactor)
3093
- );
3097
+ function mergeAttributes(start, complete) {
3098
+ const merged = {
3099
+ ...isRecord6(complete.attributes) ? complete.attributes : {},
3100
+ ...isRecord6(start.attributes) ? start.attributes : {}
3101
+ };
3102
+ merged.legacyEvent = start.attributes?.legacyEvent ?? complete.attributes?.legacyEvent;
3103
+ merged.legacyCompleteEvent = complete.attributes?.legacyEvent;
3104
+ if (complete.attributes?.errorStack !== void 0) {
3105
+ merged.errorStack = complete.attributes.errorStack;
3094
3106
  }
3095
- facts.push(...attributeFacts(nodes, redactor));
3096
- return facts;
3107
+ return Object.keys(merged).length > 0 ? merged : void 0;
3097
3108
  }
3098
- function buildInferences(run, facts) {
3099
- const inferences = [];
3100
- const errorFact = facts.find((item) => item.id === "run.errorNodeCount");
3101
- const kindFact = facts.find((item) => item.id === "run.kinds");
3102
- const durationFact = facts.find((item) => item.id === "run.durationMs");
3103
- const errorNodeCount = typeof errorFact?.value === "number" ? errorFact.value : 0;
3104
- if (run.status === "error" || errorNodeCount > 0) {
3105
- inferences.push({
3106
- id: "outcome.error",
3107
- label: "Outcome",
3108
- text: "The run recorded an error status or at least one error node.",
3109
- evidence: ["run.status", "run.errorNodeCount"],
3110
- confidence: "deterministic"
3111
- });
3112
- } else if (run.status === "ok") {
3113
- inferences.push({
3114
- id: "outcome.success",
3115
- label: "Outcome",
3116
- text: "The run completed without observed error nodes.",
3117
- evidence: ["run.status", "run.errorNodeCount"],
3118
- confidence: "deterministic"
3119
- });
3120
- }
3121
- if (kindFact !== void 0) {
3122
- inferences.push({
3123
- id: "shape.kind-mix",
3124
- label: "Trace shape",
3125
- text: "The explanation is based on the observed event kind mix, not generated content.",
3126
- evidence: [kindFact.id],
3127
- confidence: "deterministic"
3128
- });
3129
- }
3130
- if (durationFact !== void 0) {
3131
- inferences.push({
3132
- id: "timing.duration",
3133
- label: "Timing",
3134
- text: "Timing claims are limited to persisted duration fields in the trace.",
3135
- evidence: [durationFact.id],
3136
- confidence: "deterministic"
3137
- });
3138
- }
3139
- return inferences;
3109
+ function pairStartComplete(start, complete) {
3110
+ const attributes = mergeAttributes(start, complete);
3111
+ const paired = {
3112
+ ...cloneEvent(start),
3113
+ status: complete.status,
3114
+ timestamp: complete.timestamp ?? start.timestamp,
3115
+ ...complete.endedAt !== void 0 ? { endedAt: complete.endedAt } : {},
3116
+ ...complete.durationMs !== void 0 ? { durationMs: complete.durationMs } : {},
3117
+ ...complete.error !== void 0 ? { error: { ...complete.error } } : {},
3118
+ ...complete.tokenUsage !== void 0 && start.tokenUsage === void 0 ? { tokenUsage: { ...complete.tokenUsage } } : {},
3119
+ ...attributes !== void 0 ? { attributes } : {}
3120
+ };
3121
+ return {
3122
+ ...paired,
3123
+ sourceEventIds: Object.freeze([start.eventId, complete.eventId]),
3124
+ projection: {
3125
+ paired: true,
3126
+ absorbedEventIds: Object.freeze([complete.eventId]),
3127
+ parentNormalized: false,
3128
+ ...start.parentId !== void 0 ? { originalParentId: start.parentId } : {}
3129
+ }
3130
+ };
3140
3131
  }
3141
- function buildLocalExplanation(run, options = {}) {
3142
- const redactionProfile = options.redactionProfile ?? "local";
3143
- const resolved = resolveRedactionProfile(redactionProfile);
3144
- const redactor = new Redactor({ extraKeys: resolved.extraKeys });
3145
- const mode = options.mode ?? "local";
3146
- const facts = buildFacts(run, redactor);
3132
+ function asLogical(event, extras) {
3147
3133
  return {
3148
- mode,
3149
- runId: String(redactValue(redactor, "runId", run.runId)),
3150
- ...run.name !== void 0 ? { name: String(redactValue(redactor, "name", run.name)) } : {},
3151
- ...run.status !== void 0 ? { status: toDisplayStatus(run.status) } : {},
3152
- redactionProfile,
3153
- facts,
3154
- inferences: mode === "dry-run" ? [] : buildInferences(run, facts),
3155
- notes: [
3156
- "Generated locally without provider or network calls.",
3157
- "Facts are observed from normalized trace data; inferences are deterministic labels."
3158
- ]
3134
+ ...cloneEvent(event),
3135
+ sourceEventIds: Object.freeze([event.eventId]),
3136
+ projection: {
3137
+ paired: false,
3138
+ absorbedEventIds: Object.freeze([]),
3139
+ parentNormalized: extras?.parentNormalized === true,
3140
+ ...{}
3141
+ }
3159
3142
  };
3160
3143
  }
3161
- var init_explain = __esm({
3162
- "packages/core/src/explain.ts"() {
3144
+ function projectLogicalEvents(events) {
3145
+ const diagnostics = [];
3146
+ const byRun = /* @__PURE__ */ new Map();
3147
+ for (const event of events) {
3148
+ const list2 = byRun.get(event.runId) ?? [];
3149
+ list2.push(event);
3150
+ byRun.set(event.runId, list2);
3151
+ }
3152
+ const absorbedIds = /* @__PURE__ */ new Set();
3153
+ const logicalByRawId = /* @__PURE__ */ new Map();
3154
+ const stepIdToLogicalId = /* @__PURE__ */ new Map();
3155
+ const logical = [];
3156
+ const orderedRuns = [...byRun.keys()].sort((a, b) => a.localeCompare(b));
3157
+ for (const runId of orderedRuns) {
3158
+ const runEvents = byRun.get(runId) ?? [];
3159
+ const starts = [];
3160
+ const completes = [];
3161
+ const others = [];
3162
+ for (const event of runEvents) {
3163
+ const legacy = legacyEvent(event);
3164
+ if (legacy === "step_started" || legacy === "run_started" && event.status === "running") {
3165
+ starts.push(event);
3166
+ } else if (legacy === "step_completed" || legacy === "run_completed") {
3167
+ completes.push(event);
3168
+ } else {
3169
+ others.push(event);
3170
+ }
3171
+ }
3172
+ const usedCompletes = /* @__PURE__ */ new Set();
3173
+ for (const start of starts) {
3174
+ const stepId = stepIdOf(start);
3175
+ let match;
3176
+ if (legacyEvent(start) === "run_started") {
3177
+ const candidates = completes.filter(
3178
+ (c) => !usedCompletes.has(c.eventId) && legacyEvent(c) === "run_completed"
3179
+ );
3180
+ if (candidates.length > 1) {
3181
+ diagnostics.push({
3182
+ code: "AI_LOGICAL_PAIR_AMBIGUOUS",
3183
+ message: `Multiple run_completed rows for run ${runId}; using first by eventId.`,
3184
+ eventIds: candidates.map((c) => c.eventId)
3185
+ });
3186
+ candidates.sort((a, b) => a.eventId.localeCompare(b.eventId));
3187
+ }
3188
+ match = candidates[0];
3189
+ } else if (stepId) {
3190
+ const candidates = completes.filter(
3191
+ (c) => !usedCompletes.has(c.eventId) && legacyEvent(c) === "step_completed" && stepIdOf(c) === stepId
3192
+ );
3193
+ if (candidates.length > 1) {
3194
+ diagnostics.push({
3195
+ code: "AI_LOGICAL_PAIR_AMBIGUOUS",
3196
+ message: `Multiple step_completed rows for stepId ${stepId}; using first by eventId.`,
3197
+ eventIds: candidates.map((c) => c.eventId)
3198
+ });
3199
+ candidates.sort((a, b) => a.eventId.localeCompare(b.eventId));
3200
+ }
3201
+ match = candidates[0];
3202
+ }
3203
+ if (match) {
3204
+ usedCompletes.add(match.eventId);
3205
+ absorbedIds.add(match.eventId);
3206
+ const paired = pairStartComplete(start, match);
3207
+ logical.push(paired);
3208
+ logicalByRawId.set(start.eventId, paired);
3209
+ logicalByRawId.set(match.eventId, paired);
3210
+ if (stepId) stepIdToLogicalId.set(`${runId}:${stepId}`, paired.eventId);
3211
+ } else {
3212
+ diagnostics.push({
3213
+ code: "AI_LOGICAL_PAIR_UNMATCHED_START",
3214
+ message: `No matching complete for start ${start.eventId}.`,
3215
+ eventIds: [start.eventId]
3216
+ });
3217
+ const alone = asLogical(start);
3218
+ logical.push(alone);
3219
+ logicalByRawId.set(start.eventId, alone);
3220
+ if (stepId) stepIdToLogicalId.set(`${runId}:${stepId}`, alone.eventId);
3221
+ }
3222
+ }
3223
+ for (const complete of completes) {
3224
+ if (usedCompletes.has(complete.eventId)) continue;
3225
+ diagnostics.push({
3226
+ code: "AI_LOGICAL_PAIR_UNMATCHED_COMPLETE",
3227
+ message: `No matching start for complete ${complete.eventId}.`,
3228
+ eventIds: [complete.eventId]
3229
+ });
3230
+ const alone = asLogical(complete);
3231
+ logical.push(alone);
3232
+ logicalByRawId.set(complete.eventId, alone);
3233
+ const stepId = stepIdOf(complete);
3234
+ if (stepId && !stepIdToLogicalId.has(`${runId}:${stepId}`)) {
3235
+ stepIdToLogicalId.set(`${runId}:${stepId}`, alone.eventId);
3236
+ }
3237
+ }
3238
+ for (const event of others) {
3239
+ const alone = asLogical(event);
3240
+ logical.push(alone);
3241
+ logicalByRawId.set(event.eventId, alone);
3242
+ const stepId = stepIdOf(event);
3243
+ if (stepId && !stepIdToLogicalId.has(`${runId}:${stepId}`)) {
3244
+ stepIdToLogicalId.set(`${runId}:${stepId}`, alone.eventId);
3245
+ }
3246
+ }
3247
+ }
3248
+ const logicalById = new Map(logical.map((e) => [e.eventId, e]));
3249
+ const normalized = [];
3250
+ for (const event of logical) {
3251
+ const originalParentId = event.parentId;
3252
+ if (!originalParentId) {
3253
+ normalized.push(event);
3254
+ continue;
3255
+ }
3256
+ let nextParent = originalParentId;
3257
+ let remapped = false;
3258
+ const viaAbsorbed = logicalByRawId.get(originalParentId);
3259
+ if (viaAbsorbed && viaAbsorbed.eventId !== originalParentId) {
3260
+ nextParent = viaAbsorbed.eventId;
3261
+ remapped = true;
3262
+ } else if (!logicalById.has(originalParentId)) {
3263
+ const viaStep = stepIdToLogicalId.get(`${event.runId}:${originalParentId}`);
3264
+ if (viaStep) {
3265
+ nextParent = viaStep;
3266
+ remapped = true;
3267
+ }
3268
+ }
3269
+ if (!remapped) {
3270
+ if (originalParentId === event.eventId) {
3271
+ diagnostics.push({
3272
+ code: "AI_LOGICAL_SELF_PARENT_REMOVED",
3273
+ message: formatProgrammaticDiagnostic(
3274
+ "AI_TRACE_RELATIONSHIP_SELF_PARENT",
3275
+ `Removed self-parent edge on ${event.eventId}.`
3276
+ ),
3277
+ eventIds: [event.eventId]
3278
+ });
3279
+ const { parentId: _drop, ...rest } = event;
3280
+ normalized.push({
3281
+ ...rest,
3282
+ projection: {
3283
+ ...event.projection,
3284
+ parentNormalized: true,
3285
+ originalParentId
3286
+ }
3287
+ });
3288
+ continue;
3289
+ }
3290
+ if (!logicalById.has(originalParentId) && !logicalByRawId.has(originalParentId)) {
3291
+ const mapping = event.attributes?.parentMapping;
3292
+ const unresolved = mapping === "unresolved" || event.attributes?.parentUnresolved === true || event.attributes?.unresolvedParent === true || // LangGraph/framework scaffolding sentinel labels are not event ids.
3293
+ /^LangGraph$/i.test(originalParentId) || originalParentId.startsWith("unresolved:");
3294
+ if (!unresolved) {
3295
+ diagnostics.push({
3296
+ code: "AI_LOGICAL_PARENT_UNRESOLVED",
3297
+ message: `Parent ${originalParentId} unresolved for ${event.eventId}.`,
3298
+ eventIds: [event.eventId]
3299
+ });
3300
+ }
3301
+ }
3302
+ normalized.push(event);
3303
+ continue;
3304
+ }
3305
+ if (nextParent === event.eventId) {
3306
+ diagnostics.push({
3307
+ code: "AI_LOGICAL_SELF_PARENT_REMOVED",
3308
+ message: formatProgrammaticDiagnostic(
3309
+ "AI_TRACE_RELATIONSHIP_SELF_PARENT",
3310
+ `Removed self-parent edge on ${event.eventId} (was ${originalParentId}).`
3311
+ ),
3312
+ eventIds: [event.eventId]
3313
+ });
3314
+ const { parentId: _drop, ...rest } = event;
3315
+ normalized.push({
3316
+ ...rest,
3317
+ projection: {
3318
+ ...event.projection,
3319
+ parentNormalized: true,
3320
+ originalParentId
3321
+ }
3322
+ });
3323
+ continue;
3324
+ }
3325
+ diagnostics.push({
3326
+ code: "AI_LOGICAL_PARENT_REMAPPED",
3327
+ message: `Remapped parent ${originalParentId} \u2192 ${nextParent} for ${event.eventId}.`,
3328
+ eventIds: [event.eventId]
3329
+ });
3330
+ normalized.push({
3331
+ ...event,
3332
+ parentId: nextParent,
3333
+ projection: {
3334
+ ...event.projection,
3335
+ parentNormalized: true,
3336
+ originalParentId
3337
+ }
3338
+ });
3339
+ }
3340
+ const rawIndex = new Map(events.map((e, i) => [e.eventId, i]));
3341
+ normalized.sort((a, b) => {
3342
+ const ai = rawIndex.get(a.sourceEventIds[0]) ?? 0;
3343
+ const bi = rawIndex.get(b.sourceEventIds[0]) ?? 0;
3344
+ return ai - bi || a.eventId.localeCompare(b.eventId);
3345
+ });
3346
+ return {
3347
+ logicalEvents: Object.freeze(normalized),
3348
+ diagnostics: Object.freeze(diagnostics)
3349
+ };
3350
+ }
3351
+ function resolveCanonicalToolName(event) {
3352
+ const attrs = event.attributes;
3353
+ const direct = pickString(attrs, ["toolName", "tool"]);
3354
+ if (direct) return direct;
3355
+ const metadata = attrs?.metadata;
3356
+ if (isRecord6(metadata)) {
3357
+ const nested = pickString(metadata, ["toolName", "tool"]);
3358
+ if (nested) return nested;
3359
+ }
3360
+ for (const prefix of ["tool:", "function:", "mcp-tools:"]) {
3361
+ if (event.name.startsWith(prefix)) return event.name.slice(prefix.length);
3362
+ }
3363
+ return event.name;
3364
+ }
3365
+ function pickString(record, keys) {
3366
+ if (!record) return void 0;
3367
+ for (const key of keys) {
3368
+ const value = record[key];
3369
+ if (typeof value === "string" && value.trim() !== "") return value.trim();
3370
+ }
3371
+ return void 0;
3372
+ }
3373
+ var init_logical_events = __esm({
3374
+ "packages/core/src/checks/logical-events.ts"() {
3375
+ init_programmatic();
3376
+ }
3377
+ });
3378
+
3379
+ // packages/core/src/explain.ts
3380
+ function toDisplayStatus(status) {
3381
+ if (status === "ok") return "success";
3382
+ if (status === "error") return "error";
3383
+ if (status === "running") return "running";
3384
+ return "unknown";
3385
+ }
3386
+ function flatten(nodes, out = []) {
3387
+ for (const node of nodes) {
3388
+ out.push({ node, index: out.length + 1 });
3389
+ flatten(node.children, out);
3390
+ }
3391
+ return out;
3392
+ }
3393
+ function redactValue(redactor, key, value) {
3394
+ return redactor.redactValue(key, value);
3395
+ }
3396
+ function fact(id, label, value, redactor) {
3397
+ return {
3398
+ id,
3399
+ label,
3400
+ value: redactValue(redactor, id.split(".").at(-1) ?? id, value),
3401
+ source: "trace",
3402
+ confidence: "observed"
3403
+ };
3404
+ }
3405
+ function topKinds(run) {
3406
+ return Object.entries(run.metadata.kinds).filter(([, count]) => count > 0).sort((a, b) => {
3407
+ if (b[1] !== a[1]) return b[1] - a[1];
3408
+ return a[0].localeCompare(b[0]);
3409
+ }).slice(0, 5).map(([kind, count]) => `${kind}:${count}`);
3410
+ }
3411
+ function countErrorNodes(nodes) {
3412
+ return nodes.filter((entry) => entry.node.event.status === "error").length;
3413
+ }
3414
+ function slowestNode(nodes) {
3415
+ return nodes.filter(
3416
+ (entry) => entry.node.event.durationMs !== void 0 && // The RUN boundary spans the whole run, so its duration always dominates
3417
+ // and shadows every real step. Rank actual work nodes only, matching the
3418
+ // stats slowest-step ranking, which never counts the run envelope.
3419
+ entry.node.event.kind !== "RUN"
3420
+ ).sort((a, b) => {
3421
+ const delta = (b.node.event.durationMs ?? 0) - (a.node.event.durationMs ?? 0);
3422
+ return delta !== 0 ? delta : a.index - b.index;
3423
+ })[0];
3424
+ }
3425
+ function attributeFacts(nodes, redactor) {
3426
+ const facts = [];
3427
+ for (const entry of nodes) {
3428
+ const attrs = entry.node.event.attributes;
3429
+ if (attrs === void 0) continue;
3430
+ for (const key of Object.keys(attrs).sort()) {
3431
+ facts.push({
3432
+ id: `node.${entry.index}.attributes.${key}`,
3433
+ label: `${entry.node.event.name} attribute ${key}`,
3434
+ value: redactValue(redactor, key, attrs[key]),
3435
+ source: "trace",
3436
+ confidence: "observed"
3437
+ });
3438
+ if (facts.length >= 8) return facts;
3439
+ }
3440
+ }
3441
+ return facts;
3442
+ }
3443
+ function buildFacts(run, redactor, events) {
3444
+ const nodes = flatten(run.children);
3445
+ const facts = [
3446
+ fact("run.id", "Run id", run.runId, redactor),
3447
+ fact("run.name", "Run name", run.name ?? run.runId, redactor),
3448
+ fact("run.status", "Run status", toDisplayStatus(run.status), redactor),
3449
+ fact("run.totalEvents", "Total events", run.metadata.totalEvents, redactor),
3450
+ fact("run.stepCount", "Top-level step count", run.children.length, redactor),
3451
+ fact("run.nodeCount", "Total node count", nodes.length, redactor),
3452
+ fact("run.errorNodeCount", "Error node count", countErrorNodes(nodes), redactor),
3453
+ fact("run.kinds", "Observed kind mix", topKinds(run), redactor)
3454
+ ];
3455
+ if (run.durationMs !== void 0) {
3456
+ facts.push(fact("run.durationMs", "Run duration milliseconds", run.durationMs, redactor));
3457
+ }
3458
+ const slowest = slowestNode(nodes);
3459
+ if (slowest !== void 0) {
3460
+ facts.push(
3461
+ fact("run.slowestNode", "Slowest observed node", {
3462
+ name: slowest.node.event.name,
3463
+ kind: slowest.node.event.kind,
3464
+ durationMs: slowest.node.event.durationMs
3465
+ }, redactor)
3466
+ );
3467
+ }
3468
+ if (events !== void 0) {
3469
+ const scoped = events.filter((event) => event.runId === run.runId);
3470
+ const { logicalEvents: logical } = projectLogicalEvents(scoped);
3471
+ const logicalSteps = logical.filter((event) => event.kind !== "RUN");
3472
+ facts.push(
3473
+ fact("run.rawEventCount", "Raw persisted event count", scoped.length, redactor),
3474
+ fact(
3475
+ "run.logicalStepCount",
3476
+ "Logical lifecycle step count",
3477
+ logicalSteps.length,
3478
+ redactor
3479
+ ),
3480
+ fact(
3481
+ "run.logicalEventCount",
3482
+ "Logical lifecycle event count (including RUN)",
3483
+ logical.length,
3484
+ redactor
3485
+ )
3486
+ );
3487
+ const slowestLogical = [...logicalSteps].filter((event) => typeof event.durationMs === "number").sort((a, b) => {
3488
+ const delta = (b.durationMs ?? 0) - (a.durationMs ?? 0);
3489
+ return delta !== 0 ? delta : a.eventId.localeCompare(b.eventId);
3490
+ })[0];
3491
+ if (slowestLogical !== void 0) {
3492
+ facts.push(
3493
+ fact("run.slowestLogicalStep", "Slowest logical lifecycle step", {
3494
+ name: slowestLogical.name,
3495
+ kind: slowestLogical.kind,
3496
+ durationMs: slowestLogical.durationMs
3497
+ }, redactor)
3498
+ );
3499
+ }
3500
+ }
3501
+ facts.push(...attributeFacts(nodes, redactor));
3502
+ return facts;
3503
+ }
3504
+ function buildInferences(run, facts) {
3505
+ const inferences = [];
3506
+ const errorFact = facts.find((item) => item.id === "run.errorNodeCount");
3507
+ const kindFact = facts.find((item) => item.id === "run.kinds");
3508
+ const durationFact = facts.find((item) => item.id === "run.durationMs");
3509
+ const errorNodeCount = typeof errorFact?.value === "number" ? errorFact.value : 0;
3510
+ if (run.status === "error" || errorNodeCount > 0) {
3511
+ inferences.push({
3512
+ id: "outcome.error",
3513
+ label: "Outcome",
3514
+ text: "The run recorded an error status or at least one error node.",
3515
+ evidence: ["run.status", "run.errorNodeCount"],
3516
+ confidence: "deterministic"
3517
+ });
3518
+ } else if (run.status === "ok") {
3519
+ inferences.push({
3520
+ id: "outcome.success",
3521
+ label: "Outcome",
3522
+ text: "The run completed without observed error nodes.",
3523
+ evidence: ["run.status", "run.errorNodeCount"],
3524
+ confidence: "deterministic"
3525
+ });
3526
+ }
3527
+ if (kindFact !== void 0) {
3528
+ inferences.push({
3529
+ id: "shape.kind-mix",
3530
+ label: "Trace shape",
3531
+ text: "The explanation is based on the observed event kind mix, not generated content.",
3532
+ evidence: [kindFact.id],
3533
+ confidence: "deterministic"
3534
+ });
3535
+ }
3536
+ if (durationFact !== void 0) {
3537
+ inferences.push({
3538
+ id: "timing.duration",
3539
+ label: "Timing",
3540
+ text: "Timing claims are limited to persisted duration fields in the trace.",
3541
+ evidence: [durationFact.id],
3542
+ confidence: "deterministic"
3543
+ });
3544
+ }
3545
+ return inferences;
3546
+ }
3547
+ function buildLocalExplanation(run, options = {}) {
3548
+ const redactionProfile = options.redactionProfile ?? "local";
3549
+ const resolved = resolveRedactionProfile(redactionProfile);
3550
+ const redactor = new Redactor({ extraKeys: resolved.extraKeys });
3551
+ const mode = options.mode ?? "local";
3552
+ const facts = buildFacts(run, redactor, options.events);
3553
+ return {
3554
+ mode,
3555
+ runId: String(redactValue(redactor, "runId", run.runId)),
3556
+ ...run.name !== void 0 ? { name: String(redactValue(redactor, "name", run.name)) } : {},
3557
+ ...run.status !== void 0 ? { status: toDisplayStatus(run.status) } : {},
3558
+ redactionProfile,
3559
+ facts,
3560
+ inferences: mode === "dry-run" ? [] : buildInferences(run, facts),
3561
+ notes: [
3562
+ "Generated locally without provider or network calls.",
3563
+ "Facts are observed from normalized trace data; inferences are deterministic labels."
3564
+ ]
3565
+ };
3566
+ }
3567
+ var init_explain = __esm({
3568
+ "packages/core/src/explain.ts"() {
3569
+ init_logical_events();
3163
3570
  init_redactor();
3164
3571
  init_redaction_profiles();
3165
3572
  }
@@ -3849,7 +4256,17 @@ var init_cohort = __esm({
3849
4256
 
3850
4257
  // packages/core/src/sessions/checks.ts
3851
4258
  function emptySummary() {
3852
- return { passed: 0, failed: 0, warnings: 0, errors: 0, rulesEvaluated: 0 };
4259
+ return {
4260
+ passed: 0,
4261
+ failed: 0,
4262
+ warnings: 0,
4263
+ errors: 0,
4264
+ rulesEvaluated: 0,
4265
+ rulesPassed: 0,
4266
+ rulesWarning: 0,
4267
+ rulesFailed: 0,
4268
+ rulesError: 0
4269
+ };
3853
4270
  }
3854
4271
  function mergeSummary(target, source) {
3855
4272
  return {
@@ -3857,7 +4274,11 @@ function mergeSummary(target, source) {
3857
4274
  failed: target.failed + source.failed,
3858
4275
  warnings: target.warnings + source.warnings,
3859
4276
  errors: target.errors + source.errors,
3860
- rulesEvaluated: target.rulesEvaluated + source.rulesEvaluated
4277
+ rulesEvaluated: target.rulesEvaluated + source.rulesEvaluated,
4278
+ rulesPassed: target.rulesPassed + source.rulesPassed,
4279
+ rulesWarning: target.rulesWarning + source.rulesWarning,
4280
+ rulesFailed: target.rulesFailed + source.rulesFailed,
4281
+ rulesError: target.rulesError + source.rulesError
3861
4282
  };
3862
4283
  }
3863
4284
  function sessionDiagnostic(code, message) {
@@ -4202,7 +4623,7 @@ var init_sessions = __esm({
4202
4623
  init_checks();
4203
4624
  }
4204
4625
  });
4205
- function isRecord6(value) {
4626
+ function isRecord7(value) {
4206
4627
  return typeof value === "object" && value !== null && !Array.isArray(value);
4207
4628
  }
4208
4629
  function safeParse(line) {
@@ -4224,7 +4645,7 @@ async function isAgentInspectTrace(filePath) {
4224
4645
  if (trimmed === "") continue;
4225
4646
  const parsed = safeParse(trimmed);
4226
4647
  if (!parsed) continue;
4227
- if (!isRecord6(parsed)) continue;
4648
+ if (!isRecord7(parsed)) continue;
4228
4649
  checked += 1;
4229
4650
  if (isTraceEvent(parsed)) return true;
4230
4651
  const ev = parsed.event;
@@ -4590,518 +5011,157 @@ function inferEvidenceFileRole(relativePath) {
4590
5011
  return "redaction-report";
4591
5012
  }
4592
5013
  if (base === "summary.md") {
4593
- return "summary";
4594
- }
4595
- return "other";
4596
- }
4597
- function buildEvidenceFileEntries(files) {
4598
- const byPath = /* @__PURE__ */ new Map();
4599
- for (const file of files) {
4600
- const relativePath = assertEvidenceRelativePath(file.path);
4601
- if (relativePath === EVIDENCE_MANIFEST_FILENAME) {
4602
- throw new Error(
4603
- `Do not include ${EVIDENCE_MANIFEST_FILENAME} in packaged file hashes (self-hash is undefined).`
4604
- );
4605
- }
4606
- if (byPath.has(relativePath)) {
4607
- throw new Error(`Duplicate evidence file path: ${relativePath}`);
4608
- }
4609
- byPath.set(relativePath, {
4610
- path: relativePath,
4611
- sha256: sha256Hex(file.content),
4612
- role: file.role ?? inferEvidenceFileRole(relativePath)
4613
- });
4614
- }
4615
- return [...byPath.values()].sort((a, b) => a.path.localeCompare(b.path));
4616
- }
4617
- function buildEvidenceManifest(parts) {
4618
- const runIds = [...parts.runIds];
4619
- if (runIds.length === 0) {
4620
- throw new Error("Evidence manifest requires at least one run id.");
4621
- }
4622
- for (const item of parts.sourceHashes) {
4623
- if (!runIds.includes(item.runId)) {
4624
- throw new Error(`sourceHashes runId "${item.runId}" is not listed in source.runIds.`);
4625
- }
4626
- if (item.algorithm !== "sha256") {
4627
- throw new Error(`Unsupported source hash algorithm: ${item.algorithm}`);
4628
- }
4629
- }
4630
- const assessment = {
4631
- status: parts.assessmentStatus,
4632
- note: parts.note ?? EVIDENCE_ASSESSMENT_NOTE
4633
- };
4634
- if (parts.sourceStatus !== void 0) {
4635
- assessment.sourceStatus = parts.sourceStatus;
4636
- }
4637
- return {
4638
- evidenceFormatVersion: EVIDENCE_FORMAT_VERSION,
4639
- generator: {
4640
- name: parts.generatorName ?? "agent-inspect",
4641
- version: parts.generatorVersion
4642
- },
4643
- createdAt: parts.createdAt ?? (/* @__PURE__ */ new Date()).toISOString(),
4644
- source: {
4645
- runIds,
4646
- traceSchemaVersions: [...parts.traceSchemaVersions].sort((a, b) => a.localeCompare(b)),
4647
- sourceHashes: [...parts.sourceHashes].sort((a, b) => a.runId.localeCompare(b.runId))
4648
- },
4649
- policy: {
4650
- redactionProfile: parts.redactionProfile,
4651
- verificationPolicy: parts.verificationPolicy ?? parts.redactionProfile
4652
- },
4653
- assessment,
4654
- ...parts.semantics !== void 0 ? { semantics: { ...parts.semantics } } : {},
4655
- ...parts.contract !== void 0 ? { contract: { ...parts.contract } } : {},
4656
- files: buildEvidenceFileEntries(parts.files)
4657
- };
4658
- }
4659
- function validateEvidenceManifest(value) {
4660
- if (value === null || typeof value !== "object" || Array.isArray(value)) {
4661
- throw new Error("Evidence manifest must be a JSON object.");
4662
- }
4663
- const record = value;
4664
- if (record.evidenceFormatVersion !== EVIDENCE_FORMAT_VERSION) {
4665
- throw new Error(
4666
- `Unsupported evidenceFormatVersion: ${String(record.evidenceFormatVersion)}`
4667
- );
4668
- }
4669
- const generator = record.generator;
4670
- if (generator === null || typeof generator !== "object" || Array.isArray(generator) || typeof generator.name !== "string" || typeof generator.version !== "string") {
4671
- throw new Error("Evidence manifest requires generator.name and generator.version.");
4672
- }
4673
- const source = record.source;
4674
- if (source === null || typeof source !== "object" || Array.isArray(source)) {
4675
- throw new Error("Evidence manifest requires source.");
4676
- }
4677
- const sourceRecord = source;
4678
- if (!Array.isArray(sourceRecord.runIds) || sourceRecord.runIds.length === 0) {
4679
- throw new Error("Evidence manifest source.runIds must be a non-empty array.");
4680
- }
4681
- if (!Array.isArray(sourceRecord.traceSchemaVersions)) {
4682
- throw new Error("Evidence manifest source.traceSchemaVersions must be an array.");
4683
- }
4684
- if (!Array.isArray(sourceRecord.sourceHashes)) {
4685
- throw new Error("Evidence manifest source.sourceHashes must be an array.");
4686
- }
4687
- const policy = record.policy;
4688
- if (policy === null || typeof policy !== "object" || Array.isArray(policy)) {
4689
- throw new Error("Evidence manifest requires policy.");
4690
- }
4691
- const assessment = record.assessment;
4692
- if (assessment === null || typeof assessment !== "object" || Array.isArray(assessment) || typeof assessment.status !== "string") {
4693
- throw new Error("Evidence manifest requires assessment.status.");
4694
- }
4695
- if (!Array.isArray(record.files) || record.files.length === 0) {
4696
- throw new Error("Evidence manifest files must be a non-empty array.");
4697
- }
4698
- for (const file of record.files) {
4699
- if (file === null || typeof file !== "object" || Array.isArray(file)) {
4700
- throw new Error("Evidence manifest file entries must be objects.");
4701
- }
4702
- const entry = file;
4703
- if (typeof entry.path !== "string") {
4704
- throw new Error("Evidence file entry requires path.");
4705
- }
4706
- assertEvidenceRelativePath(entry.path);
4707
- if (typeof entry.sha256 !== "string" || !/^[a-f0-9]{64}$/i.test(entry.sha256)) {
4708
- throw new Error(`Evidence file entry requires sha256 hex for ${entry.path}.`);
4709
- }
4710
- }
4711
- return value;
4712
- }
4713
- function parseEvidenceManifestJson(text) {
4714
- let parsed;
4715
- try {
4716
- parsed = JSON.parse(text);
4717
- } catch (error) {
4718
- const message = error instanceof Error ? error.message : String(error);
4719
- throw new Error(`Evidence manifest is not valid JSON: ${message}`);
4720
- }
4721
- return validateEvidenceManifest(parsed);
4722
- }
4723
- function collectTraceSchemaVersions(jsonl) {
4724
- const versions = /* @__PURE__ */ new Set();
4725
- for (const line of jsonl.split(/\r?\n/)) {
4726
- const trimmed = line.trim();
4727
- if (trimmed === "") continue;
4728
- try {
4729
- const row = JSON.parse(trimmed);
4730
- if (typeof row.schemaVersion === "string" && row.schemaVersion.trim() !== "") {
4731
- versions.add(row.schemaVersion.trim());
4732
- }
4733
- } catch {
4734
- }
4735
- }
4736
- return [...versions].sort((a, b) => a.localeCompare(b));
4737
- }
4738
- var init_manifest2 = __esm({
4739
- "packages/core/src/evidence/manifest.ts"() {
4740
- init_hash();
4741
- init_paths2();
4742
- init_types4();
4743
- }
4744
- });
4745
-
4746
- // packages/core/src/diagnostics/programmatic.ts
4747
- function formatProgrammaticDiagnostic(code, detail) {
4748
- const spec = PROGRAMMATIC_DIAGNOSTIC_SPECS[code];
4749
- const summary = detail?.trim() ? detail.trim() : spec.summary;
4750
- return `${code}: ${summary} Remediation: ${spec.remediation}`;
4751
- }
4752
- var PROGRAMMATIC_DIAGNOSTIC_SPECS;
4753
- var init_programmatic = __esm({
4754
- "packages/core/src/diagnostics/programmatic.ts"() {
4755
- PROGRAMMATIC_DIAGNOSTIC_SPECS = Object.freeze({
4756
- AI_TRACE_INPUT_INVALID: {
4757
- code: "AI_TRACE_INPUT_INVALID",
4758
- summary: 'Expected { type: "file", path }, { type: "directory", path }, { type: "string", content }, { type: "buffer", content }, or { type: "stdin" }.',
4759
- remediation: "For a file path, use openTraceFile(path).",
4760
- relatedCodes: ["invalid_input"]
4761
- },
4762
- AI_TRACE_FORMAT_UNSUPPORTED: {
4763
- code: "AI_TRACE_FORMAT_UNSUPPORTED",
4764
- summary: "No trace reader could detect the input format.",
4765
- remediation: "Pass an AgentInspect JSONL file via openTraceFile, or set options.format to a registered reader.",
4766
- relatedCodes: ["unsupported_format"]
4767
- },
4768
- AI_TRACE_FORMAT_AMBIGUOUS: {
4769
- code: "AI_TRACE_FORMAT_AMBIGUOUS",
4770
- summary: "Multiple trace readers matched the input with equal confidence.",
4771
- remediation: "Set options.format explicitly to disambiguate the reader.",
4772
- relatedCodes: ["ambiguous_format"]
4773
- },
4774
- AI_TRACE_FACTS_INPUT_NOT_NORMALIZED: {
4775
- code: "AI_TRACE_FACTS_INPUT_NOT_NORMALIZED",
4776
- summary: "TraceFacts requires TraceReadResult or PersistedInspectEvent[].",
4777
- remediation: "Use openTraceFile() to normalize a JSONL trace first."
4778
- },
4779
- AI_TRACE_CONTRACT_RUN_SELECTION_REQUIRED: {
4780
- code: "AI_TRACE_CONTRACT_RUN_SELECTION_REQUIRED",
4781
- summary: "Multiple runs are available; select a run before executing checks.",
4782
- remediation: "Pass options.runId or TraceCheckInput.selectedRun.",
4783
- relatedCodes: ["AI_CHECK_RUN_SELECTION_REQUIRED"]
4784
- },
4785
- AI_TRACE_RELATIONSHIP_SELF_PARENT: {
4786
- code: "AI_TRACE_RELATIONSHIP_SELF_PARENT",
4787
- summary: "A parentId equals its own event/step id.",
4788
- remediation: "Reject at capture (AI_LANGGRAPH_SELF_PARENT_REJECTED) or drop via logical projection; do not invent replacement parents.",
4789
- relatedCodes: [
4790
- "AI_LANGGRAPH_SELF_PARENT_REJECTED",
4791
- "AI_LOGICAL_SELF_PARENT_REMOVED"
4792
- ]
4793
- },
4794
- AI_TRACE_RELATIONSHIP_CYCLE: {
4795
- code: "AI_TRACE_RELATIONSHIP_CYCLE",
4796
- summary: "Trace contains a parentId cycle.",
4797
- remediation: "Use visibility-first tree linking for legacy fixtures; prefer acyclic capture for new adapter output.",
4798
- relatedCodes: ["structure.cycle"]
4799
- }
4800
- });
4801
- }
4802
- });
4803
-
4804
- // packages/core/src/checks/logical-events.ts
4805
- function isRecord7(value) {
4806
- return typeof value === "object" && value !== null && !Array.isArray(value);
4807
- }
4808
- function legacyEvent(event) {
4809
- const value = event.attributes?.legacyEvent;
4810
- return typeof value === "string" ? value : void 0;
4811
- }
4812
- function stepIdOf(event) {
4813
- const value = event.attributes?.stepId;
4814
- if (typeof value === "string" && value.trim() !== "") return value;
4815
- return void 0;
4816
- }
4817
- function cloneEvent(event) {
4818
- return {
4819
- ...event,
4820
- ...event.attributes !== void 0 ? { attributes: { ...event.attributes } } : {},
4821
- ...event.error !== void 0 ? { error: { ...event.error } } : {},
4822
- ...event.tokenUsage !== void 0 ? { tokenUsage: { ...event.tokenUsage } } : {},
4823
- ...event.source !== void 0 ? { source: { ...event.source } } : {}
4824
- };
4825
- }
4826
- function mergeAttributes(start, complete) {
4827
- const merged = {
4828
- ...isRecord7(complete.attributes) ? complete.attributes : {},
4829
- ...isRecord7(start.attributes) ? start.attributes : {}
4830
- };
4831
- merged.legacyEvent = start.attributes?.legacyEvent ?? complete.attributes?.legacyEvent;
4832
- merged.legacyCompleteEvent = complete.attributes?.legacyEvent;
4833
- if (complete.attributes?.errorStack !== void 0) {
4834
- merged.errorStack = complete.attributes.errorStack;
5014
+ return "summary";
4835
5015
  }
4836
- return Object.keys(merged).length > 0 ? merged : void 0;
5016
+ return "other";
4837
5017
  }
4838
- function pairStartComplete(start, complete) {
4839
- const attributes = mergeAttributes(start, complete);
4840
- const paired = {
4841
- ...cloneEvent(start),
4842
- status: complete.status,
4843
- timestamp: complete.timestamp ?? start.timestamp,
4844
- ...complete.endedAt !== void 0 ? { endedAt: complete.endedAt } : {},
4845
- ...complete.durationMs !== void 0 ? { durationMs: complete.durationMs } : {},
4846
- ...complete.error !== void 0 ? { error: { ...complete.error } } : {},
4847
- ...complete.tokenUsage !== void 0 && start.tokenUsage === void 0 ? { tokenUsage: { ...complete.tokenUsage } } : {},
4848
- ...attributes !== void 0 ? { attributes } : {}
4849
- };
4850
- return {
4851
- ...paired,
4852
- sourceEventIds: Object.freeze([start.eventId, complete.eventId]),
4853
- projection: {
4854
- paired: true,
4855
- absorbedEventIds: Object.freeze([complete.eventId]),
4856
- parentNormalized: false,
4857
- ...start.parentId !== void 0 ? { originalParentId: start.parentId } : {}
5018
+ function buildEvidenceFileEntries(files) {
5019
+ const byPath = /* @__PURE__ */ new Map();
5020
+ for (const file of files) {
5021
+ const relativePath = assertEvidenceRelativePath(file.path);
5022
+ if (relativePath === EVIDENCE_MANIFEST_FILENAME) {
5023
+ throw new Error(
5024
+ `Do not include ${EVIDENCE_MANIFEST_FILENAME} in packaged file hashes (self-hash is undefined).`
5025
+ );
4858
5026
  }
4859
- };
4860
- }
4861
- function asLogical(event, extras) {
4862
- return {
4863
- ...cloneEvent(event),
4864
- sourceEventIds: Object.freeze([event.eventId]),
4865
- projection: {
4866
- paired: false,
4867
- absorbedEventIds: Object.freeze([]),
4868
- parentNormalized: extras?.parentNormalized === true,
4869
- ...{}
5027
+ if (byPath.has(relativePath)) {
5028
+ throw new Error(`Duplicate evidence file path: ${relativePath}`);
4870
5029
  }
4871
- };
5030
+ byPath.set(relativePath, {
5031
+ path: relativePath,
5032
+ sha256: sha256Hex(file.content),
5033
+ role: file.role ?? inferEvidenceFileRole(relativePath)
5034
+ });
5035
+ }
5036
+ return [...byPath.values()].sort((a, b) => a.path.localeCompare(b.path));
4872
5037
  }
4873
- function projectLogicalEvents(events) {
4874
- const diagnostics = [];
4875
- const byRun = /* @__PURE__ */ new Map();
4876
- for (const event of events) {
4877
- const list2 = byRun.get(event.runId) ?? [];
4878
- list2.push(event);
4879
- byRun.set(event.runId, list2);
5038
+ function buildEvidenceManifest(parts) {
5039
+ const runIds = [...parts.runIds];
5040
+ if (runIds.length === 0) {
5041
+ throw new Error("Evidence manifest requires at least one run id.");
4880
5042
  }
4881
- const absorbedIds = /* @__PURE__ */ new Set();
4882
- const logicalByRawId = /* @__PURE__ */ new Map();
4883
- const stepIdToLogicalId = /* @__PURE__ */ new Map();
4884
- const logical = [];
4885
- const orderedRuns = [...byRun.keys()].sort((a, b) => a.localeCompare(b));
4886
- for (const runId of orderedRuns) {
4887
- const runEvents = byRun.get(runId) ?? [];
4888
- const starts = [];
4889
- const completes = [];
4890
- const others = [];
4891
- for (const event of runEvents) {
4892
- const legacy = legacyEvent(event);
4893
- if (legacy === "step_started" || legacy === "run_started" && event.status === "running") {
4894
- starts.push(event);
4895
- } else if (legacy === "step_completed" || legacy === "run_completed") {
4896
- completes.push(event);
4897
- } else {
4898
- others.push(event);
4899
- }
4900
- }
4901
- const usedCompletes = /* @__PURE__ */ new Set();
4902
- for (const start of starts) {
4903
- const stepId = stepIdOf(start);
4904
- let match;
4905
- if (legacyEvent(start) === "run_started") {
4906
- const candidates = completes.filter(
4907
- (c) => !usedCompletes.has(c.eventId) && legacyEvent(c) === "run_completed"
4908
- );
4909
- if (candidates.length > 1) {
4910
- diagnostics.push({
4911
- code: "AI_LOGICAL_PAIR_AMBIGUOUS",
4912
- message: `Multiple run_completed rows for run ${runId}; using first by eventId.`,
4913
- eventIds: candidates.map((c) => c.eventId)
4914
- });
4915
- candidates.sort((a, b) => a.eventId.localeCompare(b.eventId));
4916
- }
4917
- match = candidates[0];
4918
- } else if (stepId) {
4919
- const candidates = completes.filter(
4920
- (c) => !usedCompletes.has(c.eventId) && legacyEvent(c) === "step_completed" && stepIdOf(c) === stepId
4921
- );
4922
- if (candidates.length > 1) {
4923
- diagnostics.push({
4924
- code: "AI_LOGICAL_PAIR_AMBIGUOUS",
4925
- message: `Multiple step_completed rows for stepId ${stepId}; using first by eventId.`,
4926
- eventIds: candidates.map((c) => c.eventId)
4927
- });
4928
- candidates.sort((a, b) => a.eventId.localeCompare(b.eventId));
4929
- }
4930
- match = candidates[0];
4931
- }
4932
- if (match) {
4933
- usedCompletes.add(match.eventId);
4934
- absorbedIds.add(match.eventId);
4935
- const paired = pairStartComplete(start, match);
4936
- logical.push(paired);
4937
- logicalByRawId.set(start.eventId, paired);
4938
- logicalByRawId.set(match.eventId, paired);
4939
- if (stepId) stepIdToLogicalId.set(`${runId}:${stepId}`, paired.eventId);
4940
- } else {
4941
- diagnostics.push({
4942
- code: "AI_LOGICAL_PAIR_UNMATCHED_START",
4943
- message: `No matching complete for start ${start.eventId}.`,
4944
- eventIds: [start.eventId]
4945
- });
4946
- const alone = asLogical(start);
4947
- logical.push(alone);
4948
- logicalByRawId.set(start.eventId, alone);
4949
- if (stepId) stepIdToLogicalId.set(`${runId}:${stepId}`, alone.eventId);
4950
- }
4951
- }
4952
- for (const complete of completes) {
4953
- if (usedCompletes.has(complete.eventId)) continue;
4954
- diagnostics.push({
4955
- code: "AI_LOGICAL_PAIR_UNMATCHED_COMPLETE",
4956
- message: `No matching start for complete ${complete.eventId}.`,
4957
- eventIds: [complete.eventId]
4958
- });
4959
- const alone = asLogical(complete);
4960
- logical.push(alone);
4961
- logicalByRawId.set(complete.eventId, alone);
4962
- const stepId = stepIdOf(complete);
4963
- if (stepId && !stepIdToLogicalId.has(`${runId}:${stepId}`)) {
4964
- stepIdToLogicalId.set(`${runId}:${stepId}`, alone.eventId);
4965
- }
5043
+ for (const item of parts.sourceHashes) {
5044
+ if (!runIds.includes(item.runId)) {
5045
+ throw new Error(`sourceHashes runId "${item.runId}" is not listed in source.runIds.`);
4966
5046
  }
4967
- for (const event of others) {
4968
- const alone = asLogical(event);
4969
- logical.push(alone);
4970
- logicalByRawId.set(event.eventId, alone);
4971
- const stepId = stepIdOf(event);
4972
- if (stepId && !stepIdToLogicalId.has(`${runId}:${stepId}`)) {
4973
- stepIdToLogicalId.set(`${runId}:${stepId}`, alone.eventId);
4974
- }
5047
+ if (item.algorithm !== "sha256") {
5048
+ throw new Error(`Unsupported source hash algorithm: ${item.algorithm}`);
4975
5049
  }
4976
5050
  }
4977
- const logicalById = new Map(logical.map((e) => [e.eventId, e]));
4978
- const normalized = [];
4979
- for (const event of logical) {
4980
- const originalParentId = event.parentId;
4981
- if (!originalParentId) {
4982
- normalized.push(event);
4983
- continue;
4984
- }
4985
- let nextParent = originalParentId;
4986
- let remapped = false;
4987
- const viaAbsorbed = logicalByRawId.get(originalParentId);
4988
- if (viaAbsorbed && viaAbsorbed.eventId !== originalParentId) {
4989
- nextParent = viaAbsorbed.eventId;
4990
- remapped = true;
4991
- } else if (!logicalById.has(originalParentId)) {
4992
- const viaStep = stepIdToLogicalId.get(`${event.runId}:${originalParentId}`);
4993
- if (viaStep) {
4994
- nextParent = viaStep;
4995
- remapped = true;
4996
- }
4997
- }
4998
- if (!remapped) {
4999
- if (originalParentId === event.eventId) {
5000
- diagnostics.push({
5001
- code: "AI_LOGICAL_SELF_PARENT_REMOVED",
5002
- message: formatProgrammaticDiagnostic(
5003
- "AI_TRACE_RELATIONSHIP_SELF_PARENT",
5004
- `Removed self-parent edge on ${event.eventId}.`
5005
- ),
5006
- eventIds: [event.eventId]
5007
- });
5008
- const { parentId: _drop, ...rest } = event;
5009
- normalized.push({
5010
- ...rest,
5011
- projection: {
5012
- ...event.projection,
5013
- parentNormalized: true,
5014
- originalParentId
5015
- }
5016
- });
5017
- continue;
5018
- }
5019
- if (!logicalById.has(originalParentId) && !logicalByRawId.has(originalParentId)) {
5020
- const mapping = event.attributes?.parentMapping;
5021
- const unresolved = mapping === "unresolved" || event.attributes?.parentUnresolved === true || event.attributes?.unresolvedParent === true || // LangGraph/framework scaffolding sentinel labels are not event ids.
5022
- /^LangGraph$/i.test(originalParentId) || originalParentId.startsWith("unresolved:");
5023
- if (!unresolved) {
5024
- diagnostics.push({
5025
- code: "AI_LOGICAL_PARENT_UNRESOLVED",
5026
- message: `Parent ${originalParentId} unresolved for ${event.eventId}.`,
5027
- eventIds: [event.eventId]
5028
- });
5029
- }
5030
- }
5031
- normalized.push(event);
5032
- continue;
5033
- }
5034
- if (nextParent === event.eventId) {
5035
- diagnostics.push({
5036
- code: "AI_LOGICAL_SELF_PARENT_REMOVED",
5037
- message: formatProgrammaticDiagnostic(
5038
- "AI_TRACE_RELATIONSHIP_SELF_PARENT",
5039
- `Removed self-parent edge on ${event.eventId} (was ${originalParentId}).`
5040
- ),
5041
- eventIds: [event.eventId]
5042
- });
5043
- const { parentId: _drop, ...rest } = event;
5044
- normalized.push({
5045
- ...rest,
5046
- projection: {
5047
- ...event.projection,
5048
- parentNormalized: true,
5049
- originalParentId
5050
- }
5051
- });
5052
- continue;
5053
- }
5054
- diagnostics.push({
5055
- code: "AI_LOGICAL_PARENT_REMAPPED",
5056
- message: `Remapped parent ${originalParentId} \u2192 ${nextParent} for ${event.eventId}.`,
5057
- eventIds: [event.eventId]
5058
- });
5059
- normalized.push({
5060
- ...event,
5061
- parentId: nextParent,
5062
- projection: {
5063
- ...event.projection,
5064
- parentNormalized: true,
5065
- originalParentId
5066
- }
5067
- });
5051
+ const assessment = {
5052
+ status: parts.assessmentStatus,
5053
+ note: parts.note ?? EVIDENCE_ASSESSMENT_NOTE
5054
+ };
5055
+ if (parts.sourceStatus !== void 0) {
5056
+ assessment.sourceStatus = parts.sourceStatus;
5068
5057
  }
5069
- const rawIndex = new Map(events.map((e, i) => [e.eventId, i]));
5070
- normalized.sort((a, b) => {
5071
- const ai = rawIndex.get(a.sourceEventIds[0]) ?? 0;
5072
- const bi = rawIndex.get(b.sourceEventIds[0]) ?? 0;
5073
- return ai - bi || a.eventId.localeCompare(b.eventId);
5074
- });
5075
5058
  return {
5076
- logicalEvents: Object.freeze(normalized),
5077
- diagnostics: Object.freeze(diagnostics)
5059
+ evidenceFormatVersion: EVIDENCE_FORMAT_VERSION,
5060
+ generator: {
5061
+ name: parts.generatorName ?? "agent-inspect",
5062
+ version: parts.generatorVersion
5063
+ },
5064
+ createdAt: parts.createdAt ?? (/* @__PURE__ */ new Date()).toISOString(),
5065
+ source: {
5066
+ runIds,
5067
+ traceSchemaVersions: [...parts.traceSchemaVersions].sort((a, b) => a.localeCompare(b)),
5068
+ sourceHashes: [...parts.sourceHashes].sort((a, b) => a.runId.localeCompare(b.runId))
5069
+ },
5070
+ policy: {
5071
+ redactionProfile: parts.redactionProfile,
5072
+ verificationPolicy: parts.verificationPolicy ?? parts.redactionProfile
5073
+ },
5074
+ assessment,
5075
+ ...parts.semantics !== void 0 ? { semantics: { ...parts.semantics } } : {},
5076
+ ...parts.contract !== void 0 ? { contract: { ...parts.contract } } : {},
5077
+ ...parts.inputs !== void 0 ? { inputs: { ...parts.inputs } } : {},
5078
+ files: buildEvidenceFileEntries(parts.files)
5078
5079
  };
5079
5080
  }
5080
- function resolveCanonicalToolName(event) {
5081
- const attrs = event.attributes;
5082
- const direct = pickString(attrs, ["toolName", "tool"]);
5083
- if (direct) return direct;
5084
- const metadata = attrs?.metadata;
5085
- if (isRecord7(metadata)) {
5086
- const nested = pickString(metadata, ["toolName", "tool"]);
5087
- if (nested) return nested;
5081
+ function validateEvidenceManifest(value) {
5082
+ if (value === null || typeof value !== "object" || Array.isArray(value)) {
5083
+ throw new Error("Evidence manifest must be a JSON object.");
5088
5084
  }
5089
- for (const prefix of ["tool:", "function:", "mcp-tools:"]) {
5090
- if (event.name.startsWith(prefix)) return event.name.slice(prefix.length);
5085
+ const record = value;
5086
+ if (record.evidenceFormatVersion !== EVIDENCE_FORMAT_VERSION) {
5087
+ throw new Error(
5088
+ `Unsupported evidenceFormatVersion: ${String(record.evidenceFormatVersion)}`
5089
+ );
5091
5090
  }
5092
- return event.name;
5091
+ const generator = record.generator;
5092
+ if (generator === null || typeof generator !== "object" || Array.isArray(generator) || typeof generator.name !== "string" || typeof generator.version !== "string") {
5093
+ throw new Error("Evidence manifest requires generator.name and generator.version.");
5094
+ }
5095
+ const source = record.source;
5096
+ if (source === null || typeof source !== "object" || Array.isArray(source)) {
5097
+ throw new Error("Evidence manifest requires source.");
5098
+ }
5099
+ const sourceRecord = source;
5100
+ if (!Array.isArray(sourceRecord.runIds) || sourceRecord.runIds.length === 0) {
5101
+ throw new Error("Evidence manifest source.runIds must be a non-empty array.");
5102
+ }
5103
+ if (!Array.isArray(sourceRecord.traceSchemaVersions)) {
5104
+ throw new Error("Evidence manifest source.traceSchemaVersions must be an array.");
5105
+ }
5106
+ if (!Array.isArray(sourceRecord.sourceHashes)) {
5107
+ throw new Error("Evidence manifest source.sourceHashes must be an array.");
5108
+ }
5109
+ const policy = record.policy;
5110
+ if (policy === null || typeof policy !== "object" || Array.isArray(policy)) {
5111
+ throw new Error("Evidence manifest requires policy.");
5112
+ }
5113
+ const assessment = record.assessment;
5114
+ if (assessment === null || typeof assessment !== "object" || Array.isArray(assessment) || typeof assessment.status !== "string") {
5115
+ throw new Error("Evidence manifest requires assessment.status.");
5116
+ }
5117
+ if (!Array.isArray(record.files) || record.files.length === 0) {
5118
+ throw new Error("Evidence manifest files must be a non-empty array.");
5119
+ }
5120
+ for (const file of record.files) {
5121
+ if (file === null || typeof file !== "object" || Array.isArray(file)) {
5122
+ throw new Error("Evidence manifest file entries must be objects.");
5123
+ }
5124
+ const entry = file;
5125
+ if (typeof entry.path !== "string") {
5126
+ throw new Error("Evidence file entry requires path.");
5127
+ }
5128
+ assertEvidenceRelativePath(entry.path);
5129
+ if (typeof entry.sha256 !== "string" || !/^[a-f0-9]{64}$/i.test(entry.sha256)) {
5130
+ throw new Error(`Evidence file entry requires sha256 hex for ${entry.path}.`);
5131
+ }
5132
+ }
5133
+ return value;
5093
5134
  }
5094
- function pickString(record, keys) {
5095
- if (!record) return void 0;
5096
- for (const key of keys) {
5097
- const value = record[key];
5098
- if (typeof value === "string" && value.trim() !== "") return value.trim();
5135
+ function parseEvidenceManifestJson(text) {
5136
+ let parsed;
5137
+ try {
5138
+ parsed = JSON.parse(text);
5139
+ } catch (error) {
5140
+ const message = error instanceof Error ? error.message : String(error);
5141
+ throw new Error(`Evidence manifest is not valid JSON: ${message}`);
5099
5142
  }
5100
- return void 0;
5143
+ return validateEvidenceManifest(parsed);
5101
5144
  }
5102
- var init_logical_events = __esm({
5103
- "packages/core/src/checks/logical-events.ts"() {
5104
- init_programmatic();
5145
+ function collectTraceSchemaVersions(jsonl) {
5146
+ const versions = /* @__PURE__ */ new Set();
5147
+ for (const line of jsonl.split(/\r?\n/)) {
5148
+ const trimmed = line.trim();
5149
+ if (trimmed === "") continue;
5150
+ try {
5151
+ const row = JSON.parse(trimmed);
5152
+ if (typeof row.schemaVersion === "string" && row.schemaVersion.trim() !== "") {
5153
+ versions.add(row.schemaVersion.trim());
5154
+ }
5155
+ } catch {
5156
+ }
5157
+ }
5158
+ return [...versions].sort((a, b) => a.localeCompare(b));
5159
+ }
5160
+ var init_manifest2 = __esm({
5161
+ "packages/core/src/evidence/manifest.ts"() {
5162
+ init_hash();
5163
+ init_paths2();
5164
+ init_types4();
5105
5165
  }
5106
5166
  });
5107
5167
 
@@ -5493,7 +5553,11 @@ function emptySummary2() {
5493
5553
  failed: 0,
5494
5554
  warnings: 0,
5495
5555
  errors: 0,
5496
- rulesEvaluated: 0
5556
+ rulesEvaluated: 0,
5557
+ rulesPassed: 0,
5558
+ rulesWarning: 0,
5559
+ rulesFailed: 0,
5560
+ rulesError: 0
5497
5561
  };
5498
5562
  }
5499
5563
  function errorResult(input3, diagnostics, selectedRun, ruleExecutions = []) {
@@ -5505,7 +5569,11 @@ function errorResult(input3, diagnostics, selectedRun, ruleExecutions = []) {
5505
5569
  summary: {
5506
5570
  ...emptySummary2(),
5507
5571
  errors: diagnostics.filter((item) => item.severity === "error").length,
5508
- rulesEvaluated: ruleExecutions.length
5572
+ rulesEvaluated: ruleExecutions.length,
5573
+ rulesPassed: ruleExecutions.filter((item) => item.status === "pass").length,
5574
+ rulesWarning: ruleExecutions.filter((item) => item.status === "warning").length,
5575
+ rulesFailed: ruleExecutions.filter((item) => item.status === "fail").length,
5576
+ rulesError: ruleExecutions.filter((item) => item.status === "error").length
5509
5577
  },
5510
5578
  findings: [],
5511
5579
  diagnostics: [...diagnostics],
@@ -5665,7 +5733,7 @@ function normalizeFinding(rule, finding) {
5665
5733
  ...finding.action !== void 0 ? { action: finding.action } : {}
5666
5734
  };
5667
5735
  }
5668
- function summarize(findings, diagnostics, rulesEvaluated) {
5736
+ function summarize(findings, diagnostics, ruleExecutions) {
5669
5737
  return {
5670
5738
  passed: findings.filter((finding) => finding.status === "pass").length,
5671
5739
  failed: findings.filter(
@@ -5675,7 +5743,11 @@ function summarize(findings, diagnostics, rulesEvaluated) {
5675
5743
  (finding) => finding.status === "warning" || finding.severity === "warning"
5676
5744
  ).length,
5677
5745
  errors: diagnostics.filter((item) => item.severity === "error").length,
5678
- rulesEvaluated
5746
+ rulesEvaluated: ruleExecutions.length,
5747
+ rulesPassed: ruleExecutions.filter((item) => item.status === "pass").length,
5748
+ rulesWarning: ruleExecutions.filter((item) => item.status === "warning").length,
5749
+ rulesFailed: ruleExecutions.filter((item) => item.status === "fail").length,
5750
+ rulesError: ruleExecutions.filter((item) => item.status === "error").length
5679
5751
  };
5680
5752
  }
5681
5753
  function classifyRuleExecution(findings, threw) {
@@ -7341,7 +7413,7 @@ function runTraceChecks(input3, options = {}) {
7341
7413
  }
7342
7414
  const eventById = new Map(input3.read.events.map((event) => [event.eventId, event]));
7343
7415
  const sortedFindings = findings.sort(compareFindings(eventById));
7344
- const summary = summarize(sortedFindings, diagnostics, ruleExecutions.length);
7416
+ const summary = summarize(sortedFindings, diagnostics, ruleExecutions);
7345
7417
  const status = summary.failed > 0 ? "fail" : "pass";
7346
7418
  return {
7347
7419
  ok: status === "pass",
@@ -9480,7 +9552,17 @@ function emptyPassResult(input3, runId) {
9480
9552
  status: "pass",
9481
9553
  format: input3.read.format,
9482
9554
  ...runId !== void 0 ? { runId } : {},
9483
- summary: { passed: 0, failed: 0, warnings: 0, errors: 0, rulesEvaluated: 0 },
9555
+ summary: {
9556
+ passed: 0,
9557
+ failed: 0,
9558
+ warnings: 0,
9559
+ errors: 0,
9560
+ rulesEvaluated: 0,
9561
+ rulesPassed: 0,
9562
+ rulesWarning: 0,
9563
+ rulesFailed: 0,
9564
+ rulesError: 0
9565
+ },
9484
9566
  findings: [],
9485
9567
  diagnostics: [],
9486
9568
  ruleExecutions: []
@@ -9508,7 +9590,11 @@ function mergeContractResults(input3, parts, extraFindings = []) {
9508
9590
  (finding) => finding.status === "warning" || finding.severity === "warning"
9509
9591
  ).length,
9510
9592
  errors,
9511
- rulesEvaluated: ruleExecutions.length
9593
+ rulesEvaluated: ruleExecutions.length,
9594
+ rulesPassed: ruleExecutions.filter((item) => item.status === "pass").length,
9595
+ rulesWarning: ruleExecutions.filter((item) => item.status === "warning").length,
9596
+ rulesFailed: ruleExecutions.filter((item) => item.status === "fail").length,
9597
+ rulesError: ruleExecutions.filter((item) => item.status === "error").length
9512
9598
  },
9513
9599
  findings,
9514
9600
  diagnostics,
@@ -9556,7 +9642,76 @@ function defineTraceContract(input3) {
9556
9642
  ...alternatives ? { alternatives } : {}
9557
9643
  };
9558
9644
  }
9645
+ function isTraceReadLike(value) {
9646
+ if (typeof value !== "object" || value === null) return false;
9647
+ const candidate = value;
9648
+ return typeof candidate.format === "string" && Array.isArray(candidate.runs) && Array.isArray(candidate.events);
9649
+ }
9650
+ function isTraceCheckInputLike(value) {
9651
+ if (typeof value !== "object" || value === null || Array.isArray(value)) {
9652
+ return false;
9653
+ }
9654
+ return isTraceReadLike(value.read);
9655
+ }
9656
+ function looksLikeTraceContract(value) {
9657
+ if (typeof value !== "object" || value === null || Array.isArray(value)) {
9658
+ return false;
9659
+ }
9660
+ if (isTraceReadLike(value) || isTraceCheckInputLike(value)) {
9661
+ return false;
9662
+ }
9663
+ return true;
9664
+ }
9665
+ function invalidContractInputResult() {
9666
+ return {
9667
+ ok: false,
9668
+ status: "error",
9669
+ format: "unknown",
9670
+ summary: {
9671
+ passed: 0,
9672
+ failed: 0,
9673
+ warnings: 0,
9674
+ errors: 1,
9675
+ rulesEvaluated: 0,
9676
+ rulesPassed: 0,
9677
+ rulesWarning: 0,
9678
+ rulesFailed: 0,
9679
+ rulesError: 0
9680
+ },
9681
+ findings: [],
9682
+ diagnostics: [
9683
+ {
9684
+ code: "AI_CHECK_INVALID_ARGUMENTS",
9685
+ message: formatProgrammaticDiagnostic("AI_TRACE_CONTRACT_INPUT_INVALID"),
9686
+ severity: "error"
9687
+ }
9688
+ ],
9689
+ ruleExecutions: []
9690
+ };
9691
+ }
9692
+ function validateEvaluateTraceContractArgs(input3, contract) {
9693
+ if (isTraceCheckInputLike(input3) && looksLikeTraceContract(contract)) {
9694
+ return void 0;
9695
+ }
9696
+ if (looksLikeTraceContract(input3) && (isTraceCheckInputLike(contract) || isTraceReadLike(contract))) {
9697
+ return invalidContractInputResult();
9698
+ }
9699
+ if (isTraceReadLike(input3)) {
9700
+ return invalidContractInputResult();
9701
+ }
9702
+ if (!isTraceCheckInputLike(input3)) {
9703
+ return invalidContractInputResult();
9704
+ }
9705
+ if (!looksLikeTraceContract(contract)) {
9706
+ return invalidContractInputResult();
9707
+ }
9708
+ return void 0;
9709
+ }
9559
9710
  function evaluateTraceContract(input3, contract, options = {}) {
9711
+ const inputGuard = validateEvaluateTraceContractArgs(input3, contract);
9712
+ if (inputGuard !== void 0) {
9713
+ return inputGuard;
9714
+ }
9560
9715
  const shapeErrors = [
9561
9716
  ...validateAllowedStatusesShape(contract.run?.allowedStatuses),
9562
9717
  ...validateAlternativesShape(contract.alternatives),
@@ -9574,7 +9729,11 @@ function evaluateTraceContract(input3, contract, options = {}) {
9574
9729
  failed: 0,
9575
9730
  warnings: 0,
9576
9731
  errors: shapeErrors.length,
9577
- rulesEvaluated: 0
9732
+ rulesEvaluated: 0,
9733
+ rulesPassed: 0,
9734
+ rulesWarning: 0,
9735
+ rulesFailed: 0,
9736
+ rulesError: 0
9578
9737
  },
9579
9738
  findings: [],
9580
9739
  diagnostics: shapeErrors.map((item) => ({
@@ -9599,7 +9758,11 @@ function evaluateTraceContract(input3, contract, options = {}) {
9599
9758
  failed: 0,
9600
9759
  warnings: 0,
9601
9760
  errors: scoped.diagnostics.length,
9602
- rulesEvaluated: 0
9761
+ rulesEvaluated: 0,
9762
+ rulesPassed: 0,
9763
+ rulesWarning: 0,
9764
+ rulesFailed: 0,
9765
+ rulesError: 0
9603
9766
  },
9604
9767
  findings: [],
9605
9768
  diagnostics: scoped.diagnostics,
@@ -9705,6 +9868,7 @@ var init_contract = __esm({
9705
9868
  init_retry_safety();
9706
9869
  init_tool_arguments();
9707
9870
  init_outcomes();
9871
+ init_programmatic();
9708
9872
  init_contract_scope();
9709
9873
  init_tool_arguments();
9710
9874
  ALLOWED_STATUS_ALIASES = {
@@ -11758,6 +11922,28 @@ function asPositiveNumber(value, label) {
11758
11922
  }
11759
11923
  return value;
11760
11924
  }
11925
+ function asExpectConfig(value, label) {
11926
+ if (value === void 0) return void 0;
11927
+ if (typeof value !== "object" || value === null || Array.isArray(value)) {
11928
+ throw new Error(`${label} must be an object.`);
11929
+ }
11930
+ const raw = value;
11931
+ const checkStatus = raw.checkStatus;
11932
+ if (checkStatus !== "pass" && checkStatus !== "fail" && checkStatus !== "error") {
11933
+ throw new Error(`${label}.checkStatus must be "pass", "fail", or "error".`);
11934
+ }
11935
+ const findingRuleIds = asStringArray(raw.findingRuleIds, `${label}.findingRuleIds`);
11936
+ const minAssertions = asPositiveNumber(raw.minAssertions, `${label}.minAssertions`);
11937
+ const rejectUnexpectedFindings = raw.rejectUnexpectedFindings === void 0 ? void 0 : typeof raw.rejectUnexpectedFindings === "boolean" ? raw.rejectUnexpectedFindings : (() => {
11938
+ throw new Error(`${label}.rejectUnexpectedFindings must be a boolean.`);
11939
+ })();
11940
+ return {
11941
+ checkStatus,
11942
+ ...findingRuleIds !== void 0 ? { findingRuleIds } : {},
11943
+ ...rejectUnexpectedFindings !== void 0 ? { rejectUnexpectedFindings } : {},
11944
+ ...minAssertions !== void 0 ? { minAssertions } : {}
11945
+ };
11946
+ }
11761
11947
  function validateCaseConfig(value, index) {
11762
11948
  const diagnostics = [];
11763
11949
  if (typeof value !== "object" || value === null || Array.isArray(value)) {
@@ -11800,7 +11986,8 @@ function validateCaseConfig(value, index) {
11800
11986
  raw.expectedObservations,
11801
11987
  `cases[${index}].expectedObservations`
11802
11988
  )
11803
- } : {}
11989
+ } : {},
11990
+ ...asExpectConfig(raw.expect, `cases[${index}].expect`) !== void 0 ? { expect: asExpectConfig(raw.expect, `cases[${index}].expect`) } : {}
11804
11991
  },
11805
11992
  diagnostics
11806
11993
  };
@@ -14480,6 +14667,78 @@ function validateExpectedObservations(suiteCase, read) {
14480
14667
  }
14481
14668
  return { ok: diagnostics.length === 0, diagnostics };
14482
14669
  }
14670
+ function evaluateCaseExpect(suiteCase, checkResult, observationCount) {
14671
+ const expectation = suiteCase.expect;
14672
+ if (expectation === void 0) {
14673
+ return { ok: true, diagnostics: [] };
14674
+ }
14675
+ const diagnostics = [];
14676
+ if (checkResult.status !== expectation.checkStatus) {
14677
+ diagnostics.push(
14678
+ diagnostic5(
14679
+ "AI_SUITE_EXPECT_MISMATCH",
14680
+ `Expected check status "${expectation.checkStatus}", got "${checkResult.status}".`,
14681
+ "error",
14682
+ suiteCase.id
14683
+ )
14684
+ );
14685
+ }
14686
+ const failedFindings = checkResult.findings.filter(
14687
+ (finding) => finding.status === "fail" && finding.severity === "error"
14688
+ );
14689
+ const requiredIds = expectation.findingRuleIds ?? [];
14690
+ for (const ruleId of requiredIds) {
14691
+ if (!failedFindings.some((finding) => finding.ruleId === ruleId)) {
14692
+ diagnostics.push(
14693
+ diagnostic5(
14694
+ "AI_SUITE_EXPECT_MISMATCH",
14695
+ `Expected failing finding for rule "${ruleId}" was not present.`,
14696
+ "error",
14697
+ suiteCase.id
14698
+ )
14699
+ );
14700
+ }
14701
+ }
14702
+ const rejectUnexpected = expectation.rejectUnexpectedFindings !== false;
14703
+ if (rejectUnexpected && requiredIds.length > 0) {
14704
+ for (const finding of failedFindings) {
14705
+ if (!requiredIds.includes(finding.ruleId)) {
14706
+ diagnostics.push(
14707
+ diagnostic5(
14708
+ "AI_SUITE_EXPECT_MISMATCH",
14709
+ `Unexpected failing finding for rule "${finding.ruleId}".`,
14710
+ "error",
14711
+ suiteCase.id
14712
+ )
14713
+ );
14714
+ }
14715
+ }
14716
+ }
14717
+ if (expectation.minAssertions !== void 0) {
14718
+ const assertionCount = checkResult.ruleExecutions.length + observationCount;
14719
+ if (assertionCount < expectation.minAssertions) {
14720
+ diagnostics.push(
14721
+ diagnostic5(
14722
+ "AI_SUITE_EXPECT_MISMATCH",
14723
+ `Expected at least ${expectation.minAssertions} assertion(s), got ${assertionCount}.`,
14724
+ "error",
14725
+ suiteCase.id
14726
+ )
14727
+ );
14728
+ }
14729
+ }
14730
+ if (expectation.checkStatus === "fail" && checkResult.diagnostics.some((item) => item.severity === "error") && checkResult.status === "error") {
14731
+ diagnostics.push(
14732
+ diagnostic5(
14733
+ "AI_SUITE_EXPECT_MISMATCH",
14734
+ 'Check ended in evaluator/config error; semantic failure expectation requires status "fail".',
14735
+ "error",
14736
+ suiteCase.id
14737
+ )
14738
+ );
14739
+ }
14740
+ return { ok: diagnostics.length === 0, diagnostics };
14741
+ }
14483
14742
  async function runSuiteCase(suiteCase, config, options) {
14484
14743
  const resolved = await resolveSuiteCaseTrace(suiteCase, options);
14485
14744
  if (resolved.missing || resolved.tracePath === void 0) {
@@ -14532,9 +14791,21 @@ async function runSuiteCase(suiteCase, config, options) {
14532
14791
  status: "pass",
14533
14792
  format: read.format,
14534
14793
  findings: [],
14535
- diagnostics: []};
14794
+ diagnostics: [],
14795
+ ruleExecutions: []
14796
+ };
14536
14797
  const observationResult = validateExpectedObservations(suiteCase, read);
14537
- const diagnostics = [
14798
+ const checkOk = checkResult.ok;
14799
+ const observationsOk = observationResult.ok;
14800
+ const evalOk = config.eval?.requireSuccess === true ? checkResult.findings.every(
14801
+ (finding) => finding.ruleId !== "run.status" || finding.status !== "fail"
14802
+ ) && checkResult.diagnostics.every((item) => item.severity !== "error") : void 0;
14803
+ const expectResult = evaluateCaseExpect(
14804
+ suiteCase,
14805
+ checkResult,
14806
+ compiled.observationCount
14807
+ );
14808
+ const checkFailureDiagnostics = [
14538
14809
  ...checkResult.diagnostics.map(
14539
14810
  (item) => diagnostic5(
14540
14811
  "AI_SUITE_CASE_CHECK_FAILED",
@@ -14550,15 +14821,24 @@ async function runSuiteCase(suiteCase, config, options) {
14550
14821
  finding.severity,
14551
14822
  suiteCase.id
14552
14823
  )
14553
- ),
14554
- ...observationResult.diagnostics
14824
+ )
14555
14825
  ];
14556
- const checkOk = checkResult.ok;
14557
- const observationsOk = observationResult.ok;
14558
- const evalOk = config.eval?.requireSuccess === true ? checkResult.findings.every(
14559
- (finding) => finding.ruleId !== "run.status" || finding.status !== "fail"
14560
- ) && checkResult.diagnostics.every((item) => item.severity !== "error") : void 0;
14561
- const ok = checkOk && observationsOk;
14826
+ const expectMatched = suiteCase.expect !== void 0 && expectResult.ok && observationsOk;
14827
+ const diagnostics = expectMatched ? [
14828
+ ...observationResult.diagnostics,
14829
+ ...expectResult.diagnostics,
14830
+ diagnostic5(
14831
+ "AI_SUITE_CASE_CHECK_FAILED",
14832
+ `Expected semantic failure matched (check.status=${checkResult.status}).`,
14833
+ "info",
14834
+ suiteCase.id
14835
+ )
14836
+ ] : [
14837
+ ...checkFailureDiagnostics,
14838
+ ...observationResult.diagnostics,
14839
+ ...expectResult.diagnostics
14840
+ ];
14841
+ const ok = suiteCase.expect !== void 0 ? expectResult.ok && observationsOk : checkOk && observationsOk;
14562
14842
  const status = ok ? "pass" : checkResult.status === "error" ? "error" : "fail";
14563
14843
  return {
14564
14844
  id: suiteCase.id,
@@ -15858,6 +16138,7 @@ var init_inspect_run = __esm({
15858
16138
  "packages/core/src/inspect-run.ts"() {
15859
16139
  init_correlation_metadata();
15860
16140
  init_context();
16141
+ init_programmatic();
15861
16142
  init_storage();
15862
16143
  init_terminal();
15863
16144
  init_trace_directory();
@@ -16375,7 +16656,7 @@ var init_src = __esm({
16375
16656
  });
16376
16657
 
16377
16658
  // package.json
16378
- var version = "6.31.9";
16659
+ var version = "6.31.11";
16379
16660
 
16380
16661
  // packages/cli/src/list.ts
16381
16662
  init_advanced();
@@ -21240,6 +21521,15 @@ function actionForRule(rule) {
21240
21521
  if (rule.strategy === "full") return "replace";
21241
21522
  return rule.strategy;
21242
21523
  }
21524
+ function isCompleteRedactionMarker(value, replacement = "[REDACTED]") {
21525
+ if (typeof value !== "string") return false;
21526
+ const trimmed = value.trim();
21527
+ if (trimmed.length === 0) return false;
21528
+ if (trimmed === replacement) return true;
21529
+ if (/^\[HASH:[A-Za-z0-9_-]+\]$/.test(trimmed)) return true;
21530
+ if (/^\[REDACTED:[^\]]*\]$/.test(trimmed)) return true;
21531
+ return false;
21532
+ }
21243
21533
  function applyRule(rule, value, replacement) {
21244
21534
  if (rule.strategy === "full") return replacement;
21245
21535
  const asString2 = stringifyScalar(value);
@@ -21354,6 +21644,9 @@ var Redactor2 = class {
21354
21644
  if (key !== void 0) {
21355
21645
  const rule = findCompiledKeyRule(key, this.#rules);
21356
21646
  if (rule) {
21647
+ if (isCompleteRedactionMarker(value, this.#replacement)) {
21648
+ return value;
21649
+ }
21357
21650
  this.#recordFinding(
21358
21651
  state,
21359
21652
  makeFinding(path44, `key.${rule.key}`, actionForRule(rule), "key", "warning")
@@ -22416,7 +22709,8 @@ async function explainCommand(target, options = {}, stdin = process.stdin) {
22416
22709
  const mode = options.dryRun ? "dry-run" : "local";
22417
22710
  const explanation = buildLocalExplanation(selected, {
22418
22711
  mode,
22419
- redactionProfile
22712
+ redactionProfile,
22713
+ events: read.events
22420
22714
  });
22421
22715
  if (options.json) {
22422
22716
  writeJson({
@@ -23733,7 +24027,11 @@ function mergeSafetyExtensions(result, read, options) {
23733
24027
  failed,
23734
24028
  warnings,
23735
24029
  errors,
23736
- rulesEvaluated: ruleExecutions.length
24030
+ rulesEvaluated: ruleExecutions.length,
24031
+ rulesPassed: ruleExecutions.filter((item) => item.status === "pass").length,
24032
+ rulesWarning: ruleExecutions.filter((item) => item.status === "warning").length,
24033
+ rulesFailed: ruleExecutions.filter((item) => item.status === "fail").length,
24034
+ rulesError: ruleExecutions.filter((item) => item.status === "error").length
23737
24035
  },
23738
24036
  findings,
23739
24037
  ruleExecutions
@@ -23885,7 +24183,11 @@ function errorResult2(code, message, format = "unknown") {
23885
24183
  failed: 0,
23886
24184
  warnings: 0,
23887
24185
  errors: 1,
23888
- rulesEvaluated: 0
24186
+ rulesEvaluated: 0,
24187
+ rulesPassed: 0,
24188
+ rulesWarning: 0,
24189
+ rulesFailed: 0,
24190
+ rulesError: 0
23889
24191
  },
23890
24192
  findings: [],
23891
24193
  diagnostics,
@@ -25289,9 +25591,11 @@ function printHuman2(result, options = {}) {
25289
25591
  console.log(`Format: ${result.format}`);
25290
25592
  if (result.runId !== void 0) console.log(`Run: ${result.runId}`);
25291
25593
  console.log(
25292
- `Summary: ${result.summary.failed} failed, ${result.summary.warnings} warning(s), ${result.summary.errors} error(s)`
25594
+ `Findings: ${result.summary.failed} failed, ${result.summary.warnings} warning(s), ${result.summary.errors} error(s), ${result.summary.passed} passed`
25595
+ );
25596
+ console.log(
25597
+ `Rules: ${result.summary.rulesEvaluated ?? result.ruleExecutions?.length ?? 0} evaluated (${result.summary.rulesPassed ?? 0} passed, ${result.summary.rulesWarning ?? 0} warning, ${result.summary.rulesFailed ?? 0} failed, ${result.summary.rulesError ?? 0} error)`
25293
25598
  );
25294
- console.log(`Rules evaluated: ${result.summary.rulesEvaluated ?? result.ruleExecutions?.length ?? 0}`);
25295
25599
  for (const diagnostic8 of result.diagnostics) {
25296
25600
  console.log(`- ${diagnostic8.code}: ${diagnostic8.message}`);
25297
25601
  }