agentv 4.41.1 → 4.41.4

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 (29) hide show
  1. package/README.md +3 -1
  2. package/dist/{artifact-writer-AMV64TWV.js → artifact-writer-GKRPI5WH.js} +4 -4
  3. package/dist/{chunk-LXXMQSAR.js → chunk-P7YRPQFD.js} +659 -286
  4. package/dist/chunk-P7YRPQFD.js.map +1 -0
  5. package/dist/{chunk-ENHX2CCS.js → chunk-QCF3LUI7.js} +269 -157
  6. package/dist/chunk-QCF3LUI7.js.map +1 -0
  7. package/dist/{chunk-A4J456KS.js → chunk-VBEWRK44.js} +429 -12
  8. package/dist/chunk-VBEWRK44.js.map +1 -0
  9. package/dist/{chunk-Z45FKRMJ.js → chunk-XDE2ZRII.js} +2 -2
  10. package/dist/{chunk-JQ6XRG7X.js → chunk-YO5PBZOH.js} +22 -20
  11. package/dist/{chunk-JQ6XRG7X.js.map → chunk-YO5PBZOH.js.map} +1 -1
  12. package/dist/cli.js +5 -5
  13. package/dist/{dist-X5P5IR65.js → dist-FDVFAGRG.js} +9 -47
  14. package/dist/index.js +5 -5
  15. package/dist/{interactive-OPSG4MPD.js → interactive-ITOKDIDQ.js} +5 -5
  16. package/dist/skills/agentv-bench/references/environment-adaptation.md +1 -1
  17. package/dist/skills/agentv-bench/references/eval-yaml-spec.md +2 -2
  18. package/dist/skills/agentv-eval-writer/SKILL.md +4 -4
  19. package/dist/skills/agentv-eval-writer/references/custom-evaluators.md +2 -2
  20. package/dist/{ts-eval-loader-ZVL6CGTE-TZYZX3QS.js → ts-eval-loader-UPFWPMQO-SKZNUXPB.js} +2 -2
  21. package/package.json +1 -1
  22. package/dist/chunk-A4J456KS.js.map +0 -1
  23. package/dist/chunk-ENHX2CCS.js.map +0 -1
  24. package/dist/chunk-LXXMQSAR.js.map +0 -1
  25. /package/dist/{artifact-writer-AMV64TWV.js.map → artifact-writer-GKRPI5WH.js.map} +0 -0
  26. /package/dist/{chunk-Z45FKRMJ.js.map → chunk-XDE2ZRII.js.map} +0 -0
  27. /package/dist/{dist-X5P5IR65.js.map → dist-FDVFAGRG.js.map} +0 -0
  28. /package/dist/{interactive-OPSG4MPD.js.map → interactive-ITOKDIDQ.js.map} +0 -0
  29. /package/dist/{ts-eval-loader-ZVL6CGTE-TZYZX3QS.js.map → ts-eval-loader-UPFWPMQO-SKZNUXPB.js.map} +0 -0
@@ -6156,7 +6156,7 @@ async function expandFileReferences(tests, evalFileDir) {
6156
6156
  return expanded;
6157
6157
  }
6158
6158
 
6159
- // ../../packages/core/dist/chunk-REU6TJT4.js
6159
+ // ../../packages/core/dist/chunk-7MDLZQ2R.js
6160
6160
  import path50 from "node:path";
6161
6161
  import { pathToFileURL as pathToFileURL2 } from "node:url";
6162
6162
  import { existsSync as existsSync7 } from "node:fs";
@@ -18820,7 +18820,7 @@ var RequestError = class _RequestError extends Error {
18820
18820
  }
18821
18821
  };
18822
18822
 
18823
- // ../../packages/core/dist/chunk-REU6TJT4.js
18823
+ // ../../packages/core/dist/chunk-7MDLZQ2R.js
18824
18824
  import { exec as execCallback } from "node:child_process";
18825
18825
  import { readdirSync, statSync } from "node:fs";
18826
18826
  import { readFile as readFile32, readdir as readdir2, stat as stat2 } from "node:fs/promises";
@@ -21409,8 +21409,7 @@ var CostGrader = class {
21409
21409
  };
21410
21410
  }
21411
21411
  };
21412
- var NORMALIZED_TRAJECTORY_SCHEMA_VERSION = "agentv.trajectory.v1";
21413
- var NORMALIZED_TRACE_SOURCE_KINDS = [
21412
+ var TRACE_SOURCE_KINDS = [
21414
21413
  "agentv_run",
21415
21414
  "otlp",
21416
21415
  "phoenix",
@@ -21419,7 +21418,7 @@ var NORMALIZED_TRACE_SOURCE_KINDS = [
21419
21418
  "imported_transcript",
21420
21419
  "compact_transcript"
21421
21420
  ];
21422
- var NORMALIZED_TRACE_EVENT_TYPES = [
21421
+ var TRACE_EVENT_TYPES = [
21423
21422
  "message",
21424
21423
  "model_turn",
21425
21424
  "tool_call",
@@ -21427,13 +21426,8 @@ var NORMALIZED_TRACE_EVENT_TYPES = [
21427
21426
  "final_response",
21428
21427
  "error"
21429
21428
  ];
21430
- var NORMALIZED_TOOL_STATUSES = ["ok", "error", "timeout", "cancelled", "unknown"];
21431
- var NORMALIZED_REDACTION_LEVELS = ["none", "partial", "full"];
21432
- var TRACE_SCHEMA_VERSION = NORMALIZED_TRAJECTORY_SCHEMA_VERSION;
21433
- var TRACE_SOURCE_KINDS = NORMALIZED_TRACE_SOURCE_KINDS;
21434
- var TRACE_EVENT_TYPES = NORMALIZED_TRACE_EVENT_TYPES;
21435
- var TRACE_TOOL_STATUSES = NORMALIZED_TOOL_STATUSES;
21436
- var TRACE_REDACTION_LEVELS = NORMALIZED_REDACTION_LEVELS;
21429
+ var TRACE_TOOL_STATUSES = ["ok", "error", "timeout", "cancelled", "unknown"];
21430
+ var TRACE_REDACTION_LEVELS = ["none", "partial", "full"];
21437
21431
  function omitUndefinedProperties(value) {
21438
21432
  return Object.fromEntries(
21439
21433
  Object.entries(value).filter(([, property]) => property !== void 0)
@@ -21446,20 +21440,20 @@ var TokenUsageWireSchema = external_exports.object({
21446
21440
  cached: external_exports.number().optional(),
21447
21441
  reasoning: external_exports.number().optional()
21448
21442
  });
21449
- var NormalizedRedactionStateWireSchema = external_exports.object({
21450
- level: external_exports.enum(NORMALIZED_REDACTION_LEVELS),
21443
+ var TraceRedactionStateWireSchema = external_exports.object({
21444
+ level: external_exports.enum(TRACE_REDACTION_LEVELS),
21451
21445
  fields: external_exports.array(external_exports.string()).optional(),
21452
21446
  reason: external_exports.string().optional()
21453
21447
  });
21454
- var NormalizedTraceErrorWireSchema = external_exports.object({
21448
+ var TraceErrorWireSchema = external_exports.object({
21455
21449
  message: external_exports.string(),
21456
21450
  name: external_exports.string().optional(),
21457
21451
  code: external_exports.string().optional(),
21458
21452
  stack: external_exports.string().optional(),
21459
21453
  metadata: MetadataWireSchema.optional()
21460
21454
  });
21461
- var NormalizedTraceSourceWireSchema = external_exports.object({
21462
- kind: external_exports.enum(NORMALIZED_TRACE_SOURCE_KINDS),
21455
+ var TraceSourceWireSchema = external_exports.object({
21456
+ kind: external_exports.enum(TRACE_SOURCE_KINDS),
21463
21457
  path: external_exports.string().optional(),
21464
21458
  url: external_exports.string().optional(),
21465
21459
  provider: external_exports.string().optional(),
@@ -21467,7 +21461,7 @@ var NormalizedTraceSourceWireSchema = external_exports.object({
21467
21461
  version: external_exports.string().optional(),
21468
21462
  metadata: MetadataWireSchema.optional()
21469
21463
  });
21470
- var NormalizedTraceSessionWireSchema = external_exports.object({
21464
+ var TraceSessionWireSchema = external_exports.object({
21471
21465
  session_id: external_exports.string().optional(),
21472
21466
  conversation_id: external_exports.string().optional(),
21473
21467
  cwd: external_exports.string().optional(),
@@ -21475,14 +21469,14 @@ var NormalizedTraceSessionWireSchema = external_exports.object({
21475
21469
  ended_at: external_exports.string().optional(),
21476
21470
  metadata: MetadataWireSchema.optional()
21477
21471
  });
21478
- var NormalizedTraceBranchWireSchema = external_exports.object({
21472
+ var TraceBranchWireSchema = external_exports.object({
21479
21473
  selected_leaf_id: external_exports.string().optional(),
21480
21474
  selected_path_ids: external_exports.array(external_exports.string()).optional(),
21481
21475
  included_event_ids: external_exports.array(external_exports.string()).optional(),
21482
21476
  omitted_event_ids: external_exports.array(external_exports.string()).optional(),
21483
21477
  selection_reason: external_exports.string().optional()
21484
21478
  });
21485
- var NormalizedTraceSourceRefWireSchema = external_exports.object({
21479
+ var TraceSourceRefWireSchema = external_exports.object({
21486
21480
  event_id: external_exports.string().optional(),
21487
21481
  message_id: external_exports.string().optional(),
21488
21482
  span_id: external_exports.string().optional(),
@@ -21492,7 +21486,7 @@ var NormalizedTraceSourceRefWireSchema = external_exports.object({
21492
21486
  line: external_exports.number().int().nonnegative().optional(),
21493
21487
  metadata: MetadataWireSchema.optional()
21494
21488
  });
21495
- var NormalizedRawEvidenceWireSchema = external_exports.object({
21489
+ var TraceRawEvidenceWireSchema = external_exports.object({
21496
21490
  kind: external_exports.string(),
21497
21491
  ref: external_exports.string().optional(),
21498
21492
  media_type: external_exports.string().optional(),
@@ -21500,55 +21494,54 @@ var NormalizedRawEvidenceWireSchema = external_exports.object({
21500
21494
  redacted: external_exports.boolean().optional(),
21501
21495
  metadata: MetadataWireSchema.optional()
21502
21496
  });
21503
- var NormalizedTraceMessageWireSchema = external_exports.object({
21497
+ var TraceMessageWireSchema = external_exports.object({
21504
21498
  role: external_exports.string(),
21505
21499
  name: external_exports.string().optional(),
21506
21500
  content: external_exports.unknown().optional(),
21507
- redaction: NormalizedRedactionStateWireSchema.optional(),
21501
+ redaction: TraceRedactionStateWireSchema.optional(),
21508
21502
  token_usage: TokenUsageWireSchema.optional(),
21509
21503
  metadata: MetadataWireSchema.optional()
21510
21504
  });
21511
- var NormalizedTraceModelWireSchema = external_exports.object({
21505
+ var TraceModelWireSchema = external_exports.object({
21512
21506
  provider: external_exports.string().optional(),
21513
21507
  name: external_exports.string().optional(),
21514
21508
  invocation_id: external_exports.string().optional(),
21515
21509
  token_usage: TokenUsageWireSchema.optional(),
21516
21510
  metadata: MetadataWireSchema.optional()
21517
21511
  });
21518
- var NormalizedTraceToolWireSchema = external_exports.object({
21512
+ var TraceToolWireSchema = external_exports.object({
21519
21513
  name: external_exports.string(),
21520
21514
  call_id: external_exports.string().optional(),
21521
21515
  input: external_exports.unknown().optional(),
21522
21516
  output: external_exports.unknown().optional(),
21523
- status: external_exports.enum(NORMALIZED_TOOL_STATUSES).optional(),
21524
- error: NormalizedTraceErrorWireSchema.optional(),
21525
- redaction: NormalizedRedactionStateWireSchema.optional(),
21517
+ status: external_exports.enum(TRACE_TOOL_STATUSES).optional(),
21518
+ error: TraceErrorWireSchema.optional(),
21519
+ redaction: TraceRedactionStateWireSchema.optional(),
21526
21520
  metadata: MetadataWireSchema.optional()
21527
21521
  });
21528
- var NormalizedTraceEventWireSchema = external_exports.object({
21522
+ var TraceEventWireSchema = external_exports.object({
21529
21523
  event_id: external_exports.string(),
21530
21524
  parent_event_id: external_exports.string().optional(),
21531
21525
  ordinal: external_exports.number().int().nonnegative(),
21532
- type: external_exports.enum(NORMALIZED_TRACE_EVENT_TYPES),
21526
+ type: external_exports.enum(TRACE_EVENT_TYPES),
21533
21527
  timestamp: external_exports.string().optional(),
21534
21528
  duration_ms: external_exports.number().nonnegative().optional(),
21535
21529
  duration_inferred: external_exports.boolean().optional(),
21536
21530
  turn_index: external_exports.number().int().nonnegative().optional(),
21537
- message: NormalizedTraceMessageWireSchema.optional(),
21538
- model: NormalizedTraceModelWireSchema.optional(),
21539
- tool: NormalizedTraceToolWireSchema.optional(),
21540
- error: NormalizedTraceErrorWireSchema.optional(),
21541
- source_ref: NormalizedTraceSourceRefWireSchema.optional(),
21542
- raw_evidence: external_exports.array(NormalizedRawEvidenceWireSchema).optional(),
21543
- redaction: NormalizedRedactionStateWireSchema.optional(),
21531
+ message: TraceMessageWireSchema.optional(),
21532
+ model: TraceModelWireSchema.optional(),
21533
+ tool: TraceToolWireSchema.optional(),
21534
+ error: TraceErrorWireSchema.optional(),
21535
+ source_ref: TraceSourceRefWireSchema.optional(),
21536
+ raw_evidence: external_exports.array(TraceRawEvidenceWireSchema).optional(),
21537
+ redaction: TraceRedactionStateWireSchema.optional(),
21544
21538
  metadata: MetadataWireSchema.optional()
21545
21539
  });
21546
- var NormalizedTrajectoryWireSchema = external_exports.object({
21547
- schema_version: external_exports.literal(NORMALIZED_TRAJECTORY_SCHEMA_VERSION),
21548
- source: NormalizedTraceSourceWireSchema,
21549
- session: NormalizedTraceSessionWireSchema,
21550
- branch: NormalizedTraceBranchWireSchema.optional(),
21551
- events: external_exports.array(NormalizedTraceEventWireSchema),
21540
+ var TraceArtifactWireSchema = external_exports.object({
21541
+ source: TraceSourceWireSchema,
21542
+ session: TraceSessionWireSchema,
21543
+ branch: TraceBranchWireSchema.optional(),
21544
+ events: external_exports.array(TraceEventWireSchema),
21552
21545
  token_usage: TokenUsageWireSchema.optional(),
21553
21546
  cost_usd: external_exports.number().optional(),
21554
21547
  duration_ms: external_exports.number().optional(),
@@ -21556,43 +21549,29 @@ var NormalizedTrajectoryWireSchema = external_exports.object({
21556
21549
  ended_at: external_exports.string().optional(),
21557
21550
  metadata: MetadataWireSchema.optional()
21558
21551
  });
21559
- var TraceRedactionStateWireSchema = NormalizedRedactionStateWireSchema;
21560
- var TraceErrorWireSchema = NormalizedTraceErrorWireSchema;
21561
- var TraceSourceWireSchema = NormalizedTraceSourceWireSchema;
21562
- var TraceSessionWireSchema = NormalizedTraceSessionWireSchema;
21563
- var TraceBranchWireSchema = NormalizedTraceBranchWireSchema;
21564
- var TraceSourceRefWireSchema = NormalizedTraceSourceRefWireSchema;
21565
- var TraceRawEvidenceWireSchema = NormalizedRawEvidenceWireSchema;
21566
- var TraceMessageWireSchema = NormalizedTraceMessageWireSchema;
21567
- var TraceModelWireSchema = NormalizedTraceModelWireSchema;
21568
- var TraceToolWireSchema = NormalizedTraceToolWireSchema;
21569
- var TraceEventWireSchema = NormalizedTraceEventWireSchema;
21570
- var TraceArtifactWireSchema = NormalizedTrajectoryWireSchema;
21571
- function toNormalizedTrajectoryWire(trajectory) {
21572
- return NormalizedTrajectoryWireSchema.parse(
21552
+ function toTraceArtifactWire(artifact) {
21553
+ return TraceArtifactWireSchema.parse(
21573
21554
  omitUndefinedProperties({
21574
- schema_version: trajectory.schemaVersion,
21575
- source: toNormalizedTraceSourceWire(trajectory.source),
21576
- session: toNormalizedTraceSessionWire(trajectory.session),
21577
- branch: trajectory.branch ? toNormalizedTraceBranchWire(trajectory.branch) : void 0,
21578
- events: trajectory.events.map(toNormalizedTraceEventWire),
21579
- token_usage: trajectory.tokenUsage,
21580
- cost_usd: trajectory.costUsd,
21581
- duration_ms: trajectory.durationMs,
21582
- started_at: trajectory.startedAt,
21583
- ended_at: trajectory.endedAt,
21584
- metadata: trajectory.metadata
21555
+ source: toTraceSourceWire(artifact.source),
21556
+ session: toTraceSessionWire(artifact.session),
21557
+ branch: artifact.branch ? toTraceBranchWire(artifact.branch) : void 0,
21558
+ events: artifact.events.map(toTraceEventWire),
21559
+ token_usage: artifact.tokenUsage,
21560
+ cost_usd: artifact.costUsd,
21561
+ duration_ms: artifact.durationMs,
21562
+ started_at: artifact.startedAt,
21563
+ ended_at: artifact.endedAt,
21564
+ metadata: artifact.metadata
21585
21565
  })
21586
21566
  );
21587
21567
  }
21588
- function fromNormalizedTrajectoryWire(input) {
21589
- const wire = NormalizedTrajectoryWireSchema.parse(input);
21568
+ function fromTraceArtifactWire(input) {
21569
+ const wire = TraceArtifactWireSchema.parse(input);
21590
21570
  return {
21591
- schemaVersion: wire.schema_version,
21592
- source: fromNormalizedTraceSourceWire(wire.source),
21593
- session: fromNormalizedTraceSessionWire(wire.session),
21594
- branch: wire.branch ? fromNormalizedTraceBranchWire(wire.branch) : void 0,
21595
- events: wire.events.map(fromNormalizedTraceEventWire),
21571
+ source: fromTraceSourceWire(wire.source),
21572
+ session: fromTraceSessionWire(wire.session),
21573
+ branch: wire.branch ? fromTraceBranchWire(wire.branch) : void 0,
21574
+ events: wire.events.map(fromTraceEventWire),
21596
21575
  tokenUsage: wire.token_usage,
21597
21576
  costUsd: wire.cost_usd,
21598
21577
  durationMs: wire.duration_ms,
@@ -21601,13 +21580,7 @@ function fromNormalizedTrajectoryWire(input) {
21601
21580
  metadata: wire.metadata
21602
21581
  };
21603
21582
  }
21604
- function toTraceArtifactWire(artifact) {
21605
- return toNormalizedTrajectoryWire(artifact);
21606
- }
21607
- function fromTraceArtifactWire(input) {
21608
- return fromNormalizedTrajectoryWire(input);
21609
- }
21610
- function toNormalizedTraceSourceWire(source) {
21583
+ function toTraceSourceWire(source) {
21611
21584
  return omitUndefinedProperties({
21612
21585
  kind: source.kind,
21613
21586
  path: source.path,
@@ -21618,7 +21591,7 @@ function toNormalizedTraceSourceWire(source) {
21618
21591
  metadata: source.metadata
21619
21592
  });
21620
21593
  }
21621
- function fromNormalizedTraceSourceWire(source) {
21594
+ function fromTraceSourceWire(source) {
21622
21595
  return {
21623
21596
  kind: source.kind,
21624
21597
  path: source.path,
@@ -21629,7 +21602,7 @@ function fromNormalizedTraceSourceWire(source) {
21629
21602
  metadata: source.metadata
21630
21603
  };
21631
21604
  }
21632
- function toNormalizedTraceSessionWire(session) {
21605
+ function toTraceSessionWire(session) {
21633
21606
  return omitUndefinedProperties({
21634
21607
  session_id: session.sessionId,
21635
21608
  conversation_id: session.conversationId,
@@ -21639,7 +21612,7 @@ function toNormalizedTraceSessionWire(session) {
21639
21612
  metadata: session.metadata
21640
21613
  });
21641
21614
  }
21642
- function fromNormalizedTraceSessionWire(session) {
21615
+ function fromTraceSessionWire(session) {
21643
21616
  return {
21644
21617
  sessionId: session.session_id,
21645
21618
  conversationId: session.conversation_id,
@@ -21649,7 +21622,7 @@ function fromNormalizedTraceSessionWire(session) {
21649
21622
  metadata: session.metadata
21650
21623
  };
21651
21624
  }
21652
- function toNormalizedTraceBranchWire(branch) {
21625
+ function toTraceBranchWire(branch) {
21653
21626
  return omitUndefinedProperties({
21654
21627
  selected_leaf_id: branch.selectedLeafId,
21655
21628
  selected_path_ids: branch.selectedPathIds,
@@ -21658,7 +21631,7 @@ function toNormalizedTraceBranchWire(branch) {
21658
21631
  selection_reason: branch.selectionReason
21659
21632
  });
21660
21633
  }
21661
- function fromNormalizedTraceBranchWire(branch) {
21634
+ function fromTraceBranchWire(branch) {
21662
21635
  return {
21663
21636
  selectedLeafId: branch.selected_leaf_id,
21664
21637
  selectedPathIds: branch.selected_path_ids,
@@ -21667,8 +21640,8 @@ function fromNormalizedTraceBranchWire(branch) {
21667
21640
  selectionReason: branch.selection_reason
21668
21641
  };
21669
21642
  }
21670
- function toNormalizedTraceEventWire(event) {
21671
- return NormalizedTraceEventWireSchema.parse(
21643
+ function toTraceEventWire(event) {
21644
+ return TraceEventWireSchema.parse(
21672
21645
  omitUndefinedProperties({
21673
21646
  event_id: event.eventId,
21674
21647
  parent_event_id: event.parentEventId,
@@ -21678,18 +21651,18 @@ function toNormalizedTraceEventWire(event) {
21678
21651
  duration_ms: event.durationMs,
21679
21652
  duration_inferred: event.durationInferred,
21680
21653
  turn_index: event.turnIndex,
21681
- message: event.message ? toNormalizedTraceMessageWire(event.message) : void 0,
21682
- model: event.model ? toNormalizedTraceModelWire(event.model) : void 0,
21683
- tool: event.tool ? toNormalizedTraceToolWire(event.tool) : void 0,
21684
- error: event.error ? toNormalizedTraceErrorWire(event.error) : void 0,
21685
- source_ref: event.sourceRef ? toNormalizedTraceSourceRefWire(event.sourceRef) : void 0,
21686
- raw_evidence: event.rawEvidence?.map(toNormalizedRawEvidenceWire),
21654
+ message: event.message ? toTraceMessageWire(event.message) : void 0,
21655
+ model: event.model ? toTraceModelWire(event.model) : void 0,
21656
+ tool: event.tool ? toTraceToolWire(event.tool) : void 0,
21657
+ error: event.error ? toTraceErrorWire(event.error) : void 0,
21658
+ source_ref: event.sourceRef ? toTraceSourceRefWire(event.sourceRef) : void 0,
21659
+ raw_evidence: event.rawEvidence?.map(toTraceRawEvidenceWire),
21687
21660
  redaction: event.redaction,
21688
21661
  metadata: event.metadata
21689
21662
  })
21690
21663
  );
21691
21664
  }
21692
- function fromNormalizedTraceEventWire(event) {
21665
+ function fromTraceEventWire(event) {
21693
21666
  return {
21694
21667
  eventId: event.event_id,
21695
21668
  parentEventId: event.parent_event_id,
@@ -21699,17 +21672,17 @@ function fromNormalizedTraceEventWire(event) {
21699
21672
  durationMs: event.duration_ms,
21700
21673
  durationInferred: event.duration_inferred,
21701
21674
  turnIndex: event.turn_index,
21702
- message: event.message ? fromNormalizedTraceMessageWire(event.message) : void 0,
21703
- model: event.model ? fromNormalizedTraceModelWire(event.model) : void 0,
21704
- tool: event.tool ? fromNormalizedTraceToolWire(event.tool) : void 0,
21705
- error: event.error ? fromNormalizedTraceErrorWire(event.error) : void 0,
21706
- sourceRef: event.source_ref ? fromNormalizedTraceSourceRefWire(event.source_ref) : void 0,
21707
- rawEvidence: event.raw_evidence?.map(fromNormalizedRawEvidenceWire),
21675
+ message: event.message ? fromTraceMessageWire(event.message) : void 0,
21676
+ model: event.model ? fromTraceModelWire(event.model) : void 0,
21677
+ tool: event.tool ? fromTraceToolWire(event.tool) : void 0,
21678
+ error: event.error ? fromTraceErrorWire(event.error) : void 0,
21679
+ sourceRef: event.source_ref ? fromTraceSourceRefWire(event.source_ref) : void 0,
21680
+ rawEvidence: event.raw_evidence?.map(fromTraceRawEvidenceWire),
21708
21681
  redaction: event.redaction,
21709
21682
  metadata: event.metadata
21710
21683
  };
21711
21684
  }
21712
- function toNormalizedTraceMessageWire(message) {
21685
+ function toTraceMessageWire(message) {
21713
21686
  return omitUndefinedProperties({
21714
21687
  role: message.role,
21715
21688
  name: message.name,
@@ -21719,7 +21692,7 @@ function toNormalizedTraceMessageWire(message) {
21719
21692
  metadata: message.metadata
21720
21693
  });
21721
21694
  }
21722
- function fromNormalizedTraceMessageWire(message) {
21695
+ function fromTraceMessageWire(message) {
21723
21696
  return {
21724
21697
  role: message.role,
21725
21698
  name: message.name,
@@ -21729,7 +21702,7 @@ function fromNormalizedTraceMessageWire(message) {
21729
21702
  metadata: message.metadata
21730
21703
  };
21731
21704
  }
21732
- function toNormalizedTraceModelWire(model) {
21705
+ function toTraceModelWire(model) {
21733
21706
  return omitUndefinedProperties({
21734
21707
  provider: model.provider,
21735
21708
  name: model.name,
@@ -21738,7 +21711,7 @@ function toNormalizedTraceModelWire(model) {
21738
21711
  metadata: model.metadata
21739
21712
  });
21740
21713
  }
21741
- function fromNormalizedTraceModelWire(model) {
21714
+ function fromTraceModelWire(model) {
21742
21715
  return {
21743
21716
  provider: model.provider,
21744
21717
  name: model.name,
@@ -21747,7 +21720,7 @@ function fromNormalizedTraceModelWire(model) {
21747
21720
  metadata: model.metadata
21748
21721
  };
21749
21722
  }
21750
- function toNormalizedTraceToolWire(tool) {
21723
+ function toTraceToolWire(tool) {
21751
21724
  return omitUndefinedProperties({
21752
21725
  name: tool.name,
21753
21726
  call_id: tool.callId,
@@ -21759,7 +21732,7 @@ function toNormalizedTraceToolWire(tool) {
21759
21732
  metadata: tool.metadata
21760
21733
  });
21761
21734
  }
21762
- function fromNormalizedTraceToolWire(tool) {
21735
+ function fromTraceToolWire(tool) {
21763
21736
  return {
21764
21737
  name: tool.name,
21765
21738
  callId: tool.call_id,
@@ -21771,7 +21744,7 @@ function fromNormalizedTraceToolWire(tool) {
21771
21744
  metadata: tool.metadata
21772
21745
  };
21773
21746
  }
21774
- function toNormalizedTraceErrorWire(error40) {
21747
+ function toTraceErrorWire(error40) {
21775
21748
  return omitUndefinedProperties({
21776
21749
  message: error40.message,
21777
21750
  name: error40.name,
@@ -21780,7 +21753,7 @@ function toNormalizedTraceErrorWire(error40) {
21780
21753
  metadata: error40.metadata
21781
21754
  });
21782
21755
  }
21783
- function fromNormalizedTraceErrorWire(error40) {
21756
+ function fromTraceErrorWire(error40) {
21784
21757
  return {
21785
21758
  message: error40.message,
21786
21759
  name: error40.name,
@@ -21789,7 +21762,7 @@ function fromNormalizedTraceErrorWire(error40) {
21789
21762
  metadata: error40.metadata
21790
21763
  };
21791
21764
  }
21792
- function toNormalizedTraceSourceRefWire(sourceRef) {
21765
+ function toTraceSourceRefWire(sourceRef) {
21793
21766
  return omitUndefinedProperties({
21794
21767
  event_id: sourceRef.eventId,
21795
21768
  message_id: sourceRef.messageId,
@@ -21801,7 +21774,7 @@ function toNormalizedTraceSourceRefWire(sourceRef) {
21801
21774
  metadata: sourceRef.metadata
21802
21775
  });
21803
21776
  }
21804
- function fromNormalizedTraceSourceRefWire(sourceRef) {
21777
+ function fromTraceSourceRefWire(sourceRef) {
21805
21778
  return {
21806
21779
  eventId: sourceRef.event_id,
21807
21780
  messageId: sourceRef.message_id,
@@ -21813,7 +21786,7 @@ function fromNormalizedTraceSourceRefWire(sourceRef) {
21813
21786
  metadata: sourceRef.metadata
21814
21787
  };
21815
21788
  }
21816
- function toNormalizedRawEvidenceWire(evidence) {
21789
+ function toTraceRawEvidenceWire(evidence) {
21817
21790
  return omitUndefinedProperties({
21818
21791
  kind: evidence.kind,
21819
21792
  ref: evidence.ref,
@@ -21823,7 +21796,7 @@ function toNormalizedRawEvidenceWire(evidence) {
21823
21796
  metadata: evidence.metadata
21824
21797
  });
21825
21798
  }
21826
- function fromNormalizedRawEvidenceWire(evidence) {
21799
+ function fromTraceRawEvidenceWire(evidence) {
21827
21800
  return {
21828
21801
  kind: evidence.kind,
21829
21802
  ref: evidence.ref,
@@ -21951,7 +21924,6 @@ function buildTraceFromMessages(options = {}) {
21951
21924
  });
21952
21925
  }
21953
21926
  return {
21954
- schemaVersion: TRACE_SCHEMA_VERSION,
21955
21927
  eventCount: summary.eventCount,
21956
21928
  toolCalls: summary.toolCalls,
21957
21929
  errorCount: summary.errorCount + (options.error ? 1 : 0),
@@ -22057,15 +22029,15 @@ function computeTraceSummary(messages) {
22057
22029
  endTime: latestEnd?.toISOString()
22058
22030
  };
22059
22031
  }
22060
- function getSelectedTrajectoryEvents(trajectory) {
22061
- if (!trajectory.branch?.includedEventIds || trajectory.branch.includedEventIds.length === 0) {
22062
- return trajectory.events;
22032
+ function getSelectedTraceArtifactEvents(artifact) {
22033
+ if (!artifact.branch?.includedEventIds || artifact.branch.includedEventIds.length === 0) {
22034
+ return artifact.events;
22063
22035
  }
22064
- const includedIds = new Set(trajectory.branch.includedEventIds);
22065
- return trajectory.events.filter((event) => includedIds.has(event.eventId));
22036
+ const includedIds = new Set(artifact.branch.includedEventIds);
22037
+ return artifact.events.filter((event) => includedIds.has(event.eventId));
22066
22038
  }
22067
- function computeTraceSummaryFromTrajectory(trajectory) {
22068
- const selectedEvents = getSelectedTrajectoryEvents(trajectory);
22039
+ function computeTraceSummaryFromTraceArtifact(artifact) {
22040
+ const selectedEvents = getSelectedTraceArtifactEvents(artifact);
22069
22041
  const hasModelTurnEvents = selectedEvents.some((event) => event.type === "model_turn");
22070
22042
  const toolCallCounts = {};
22071
22043
  const toolDurations = {};
@@ -22111,11 +22083,163 @@ function computeTraceSummaryFromTrajectory(trajectory) {
22111
22083
  llmCallCount,
22112
22084
  ...hasAnyDuration ? { toolDurations } : {}
22113
22085
  },
22114
- tokenUsage: trajectory.tokenUsage,
22115
- costUsd: trajectory.costUsd,
22116
- durationMs: trajectory.durationMs,
22117
- startTime: trajectory.startedAt ?? earliestStart?.toISOString(),
22118
- endTime: trajectory.endedAt ?? latestEnd?.toISOString()
22086
+ tokenUsage: artifact.tokenUsage,
22087
+ costUsd: artifact.costUsd,
22088
+ durationMs: artifact.durationMs,
22089
+ startTime: artifact.startedAt ?? earliestStart?.toISOString(),
22090
+ endTime: artifact.endedAt ?? latestEnd?.toISOString()
22091
+ };
22092
+ }
22093
+ function eventEndIso(event) {
22094
+ const start = parseTimestamp(event.timestamp);
22095
+ return deriveEventEnd(start, event.durationMs)?.toISOString();
22096
+ }
22097
+ function messageFromTraceArtifactEvent(event, fallbackRole = "assistant") {
22098
+ return {
22099
+ role: event.message?.role ?? fallbackRole,
22100
+ name: event.message?.name,
22101
+ content: event.message?.content,
22102
+ startTime: event.timestamp,
22103
+ endTime: eventEndIso(event),
22104
+ durationMs: event.durationMs,
22105
+ metadata: omitUndefinedProperties({
22106
+ event_id: event.eventId,
22107
+ source_ref: event.sourceRef,
22108
+ ...event.message?.metadata
22109
+ }),
22110
+ tokenUsage: event.message?.tokenUsage ?? event.model?.tokenUsage
22111
+ };
22112
+ }
22113
+ function toolCallFromTraceArtifactEvent(event) {
22114
+ const tool = event.tool;
22115
+ if (!tool) {
22116
+ return void 0;
22117
+ }
22118
+ return {
22119
+ tool: tool.name,
22120
+ id: tool.callId,
22121
+ input: tool.input,
22122
+ output: tool.output,
22123
+ startTime: event.timestamp,
22124
+ endTime: eventEndIso(event),
22125
+ durationMs: event.durationMs
22126
+ };
22127
+ }
22128
+ function attachTraceArtifactToolCall(messages, eventMessageIndex, event) {
22129
+ const toolCall = toolCallFromTraceArtifactEvent(event);
22130
+ if (!toolCall) {
22131
+ return;
22132
+ }
22133
+ const parentIndex = event.parentEventId ? eventMessageIndex.get(event.parentEventId) : void 0;
22134
+ let fallbackIndex = -1;
22135
+ for (let index = messages.length - 1; index >= 0; index--) {
22136
+ if (messages[index].role === "assistant") {
22137
+ fallbackIndex = index;
22138
+ break;
22139
+ }
22140
+ }
22141
+ const messageIndex = parentIndex ?? fallbackIndex;
22142
+ if (messageIndex >= 0) {
22143
+ const message = messages[messageIndex];
22144
+ messages[messageIndex] = {
22145
+ ...message,
22146
+ toolCalls: [...message.toolCalls ?? [], toolCall]
22147
+ };
22148
+ return;
22149
+ }
22150
+ const syntheticIndex = messages.length;
22151
+ messages.push({
22152
+ role: "assistant",
22153
+ startTime: event.timestamp,
22154
+ endTime: eventEndIso(event),
22155
+ durationMs: event.durationMs,
22156
+ metadata: { source: "trace_artifact_tool_call", event_id: event.eventId },
22157
+ toolCalls: [toolCall]
22158
+ });
22159
+ eventMessageIndex.set(event.eventId, syntheticIndex);
22160
+ }
22161
+ function mergeTraceArtifactFinalResponse(messages, eventMessageIndex, event) {
22162
+ if (!event.parentEventId || !event.message) {
22163
+ return false;
22164
+ }
22165
+ const parentIndex = eventMessageIndex.get(event.parentEventId);
22166
+ if (parentIndex === void 0) {
22167
+ return false;
22168
+ }
22169
+ const parent = messages[parentIndex];
22170
+ const finalMessage = messageFromTraceArtifactEvent(event);
22171
+ eventMessageIndex.set(event.eventId, parentIndex);
22172
+ if (sameMessageContent(parent, finalMessage)) {
22173
+ return true;
22174
+ }
22175
+ if (parent.content === void 0 && finalMessage.content !== void 0) {
22176
+ messages[parentIndex] = {
22177
+ ...parent,
22178
+ name: parent.name ?? finalMessage.name,
22179
+ content: finalMessage.content,
22180
+ endTime: parent.endTime ?? finalMessage.endTime,
22181
+ tokenUsage: parent.tokenUsage ?? finalMessage.tokenUsage,
22182
+ metadata: omitUndefinedProperties({
22183
+ ...parent.metadata,
22184
+ final_response_event_id: event.eventId,
22185
+ final_response_source_ref: event.sourceRef
22186
+ })
22187
+ };
22188
+ return true;
22189
+ }
22190
+ return false;
22191
+ }
22192
+ function traceFromTraceArtifact(artifact) {
22193
+ const events = [...getSelectedTraceArtifactEvents(artifact)].sort(
22194
+ (first, second) => first.ordinal - second.ordinal
22195
+ );
22196
+ const messages = [];
22197
+ const eventMessageIndex = /* @__PURE__ */ new Map();
22198
+ for (const event of events) {
22199
+ if (event.type === "message" && event.message) {
22200
+ eventMessageIndex.set(event.eventId, messages.length);
22201
+ messages.push(messageFromTraceArtifactEvent(event));
22202
+ continue;
22203
+ }
22204
+ if (event.type === "model_turn") {
22205
+ eventMessageIndex.set(event.eventId, messages.length);
22206
+ messages.push(messageFromTraceArtifactEvent(event, "assistant"));
22207
+ continue;
22208
+ }
22209
+ if (event.type === "final_response" && event.message) {
22210
+ if (mergeTraceArtifactFinalResponse(messages, eventMessageIndex, event)) {
22211
+ continue;
22212
+ }
22213
+ eventMessageIndex.set(event.eventId, messages.length);
22214
+ messages.push(messageFromTraceArtifactEvent(event));
22215
+ continue;
22216
+ }
22217
+ if (event.type === "tool_call") {
22218
+ attachTraceArtifactToolCall(messages, eventMessageIndex, event);
22219
+ }
22220
+ }
22221
+ const computed = computeTraceSummaryFromTraceArtifact(artifact);
22222
+ const conversationId = artifact.session.sessionId ?? artifact.session.conversationId;
22223
+ return {
22224
+ eventCount: computed.trace.eventCount,
22225
+ toolCalls: computed.trace.toolCalls,
22226
+ errorCount: computed.trace.errorCount,
22227
+ llmCallCount: computed.trace.llmCallCount,
22228
+ ...computed.trace.toolDurations ? { toolDurations: computed.trace.toolDurations } : {},
22229
+ messages,
22230
+ events,
22231
+ tokenUsage: computed.tokenUsage,
22232
+ costUsd: computed.costUsd,
22233
+ durationMs: computed.durationMs,
22234
+ startTime: computed.startTime,
22235
+ endTime: computed.endTime,
22236
+ metadata: {
22237
+ ...artifact.source.provider ? { provider: artifact.source.provider } : {},
22238
+ ...conversationId ? { provider_session_id: conversationId } : {},
22239
+ source: artifact.source,
22240
+ session: artifact.session,
22241
+ ...artifact.metadata
22242
+ }
22119
22243
  };
22120
22244
  }
22121
22245
  function parseTimestamp(timestamp) {
@@ -30608,7 +30732,6 @@ function traceEnvelopeToToolTrajectoryView(envelope) {
30608
30732
  };
30609
30733
  });
30610
30734
  return {
30611
- schemaVersion: NORMALIZED_TRAJECTORY_SCHEMA_VERSION,
30612
30735
  traceId: envelope.trace.traceId,
30613
30736
  rootSpanId: envelope.trace.rootSpanId,
30614
30737
  tools
@@ -30710,7 +30833,6 @@ function traceEnvelopeToTraceArtifact(envelope) {
30710
30833
  }
30711
30834
  const summary = traceEnvelopeToTraceSummary(envelope);
30712
30835
  return {
30713
- schemaVersion: NORMALIZED_TRAJECTORY_SCHEMA_VERSION,
30714
30836
  source: {
30715
30837
  kind: envelope.source.kind,
30716
30838
  path: envelope.source.path,
@@ -38605,7 +38727,7 @@ async function loadTestSuite(evalFilePath, repoRoot, options) {
38605
38727
  return { tests: await loadTestsFromAgentSkills(evalFilePath) };
38606
38728
  }
38607
38729
  if (format === "typescript") {
38608
- const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-ZVL6CGTE-TZYZX3QS.js");
38730
+ const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-UPFWPMQO-SKZNUXPB.js");
38609
38731
  return loadTsEvalSuite2(evalFilePath, resolveToAbsolutePath(repoRoot), options);
38610
38732
  }
38611
38733
  const { tests, parsed, suiteWorkspacePath } = await loadTestsFromYaml(
@@ -38640,7 +38762,7 @@ async function loadTests(evalFilePath, repoRoot, options) {
38640
38762
  return loadTestsFromAgentSkills(evalFilePath);
38641
38763
  }
38642
38764
  if (format === "typescript") {
38643
- const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-ZVL6CGTE-TZYZX3QS.js");
38765
+ const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-UPFWPMQO-SKZNUXPB.js");
38644
38766
  const suite = await loadTsEvalSuite2(evalFilePath, resolveToAbsolutePath(repoRoot), options);
38645
38767
  return suite.tests;
38646
38768
  }
@@ -39592,6 +39714,7 @@ async function gradePreparedEvalCase(options) {
39592
39714
  workspacePath,
39593
39715
  baselineCommit,
39594
39716
  response,
39717
+ trace: preparedTrace,
39595
39718
  verbose,
39596
39719
  threshold: caseThreshold,
39597
39720
  preparedAttempt
@@ -39629,10 +39752,10 @@ async function gradePreparedEvalCase(options) {
39629
39752
  }
39630
39753
  }
39631
39754
  }
39632
- const candidate = response ?? "";
39755
+ const candidate = response ?? extractLastAssistantContent(preparedTrace?.messages ?? []) ?? "";
39633
39756
  const input = buildResultInput(promptInputs);
39634
- const outputMessages = candidate.length > 0 ? [{ role: "assistant", content: candidate }] : [];
39635
- const resultTrace = buildTraceFromMessages({
39757
+ const outputMessages = preparedTrace?.messages ?? (candidate.length > 0 ? [{ role: "assistant", content: candidate }] : []);
39758
+ const resultTrace = preparedTrace ?? buildTraceFromMessages({
39636
39759
  input,
39637
39760
  output: outputMessages,
39638
39761
  finalOutput: candidate,
@@ -39654,6 +39777,13 @@ async function gradePreparedEvalCase(options) {
39654
39777
  promptInputs,
39655
39778
  now: gradeStartedAt,
39656
39779
  agentTimeoutMs,
39780
+ output: preparedTrace ? outputMessages : void 0,
39781
+ trace: preparedTrace ? resultTrace : void 0,
39782
+ costUsd: preparedTrace ? resultTrace.costUsd : void 0,
39783
+ durationMs: preparedTrace ? resultTrace.durationMs : void 0,
39784
+ tokenUsage: preparedTrace ? resultTrace.tokenUsage : void 0,
39785
+ startTime: preparedTrace ? resultTrace.startTime : void 0,
39786
+ endTime: preparedTrace ? resultTrace.endTime : void 0,
39657
39787
  targetResolver: runtime.targetResolver,
39658
39788
  availableTargets: runtime.availableTargets,
39659
39789
  fileChanges,
@@ -42518,6 +42648,7 @@ function toIndexPreparedAttempt(value) {
42518
42648
  prepared_dir: value.preparedDir,
42519
42649
  workspace_path: value.workspacePath,
42520
42650
  prompt_path: value.promptPath,
42651
+ trace_path: value.tracePath,
42521
42652
  target: value.target,
42522
42653
  prepared_at: value.preparedAt,
42523
42654
  setup_status: value.setupStatus,
@@ -43626,28 +43757,10 @@ export {
43626
43757
  extractImageBlocks,
43627
43758
  CompositeGrader,
43628
43759
  CostGrader,
43629
- NORMALIZED_TRAJECTORY_SCHEMA_VERSION,
43630
- NORMALIZED_TRACE_SOURCE_KINDS,
43631
- NORMALIZED_TRACE_EVENT_TYPES,
43632
- NORMALIZED_TOOL_STATUSES,
43633
- NORMALIZED_REDACTION_LEVELS,
43634
- TRACE_SCHEMA_VERSION,
43635
43760
  TRACE_SOURCE_KINDS,
43636
43761
  TRACE_EVENT_TYPES,
43637
43762
  TRACE_TOOL_STATUSES,
43638
43763
  TRACE_REDACTION_LEVELS,
43639
- NormalizedRedactionStateWireSchema,
43640
- NormalizedTraceErrorWireSchema,
43641
- NormalizedTraceSourceWireSchema,
43642
- NormalizedTraceSessionWireSchema,
43643
- NormalizedTraceBranchWireSchema,
43644
- NormalizedTraceSourceRefWireSchema,
43645
- NormalizedRawEvidenceWireSchema,
43646
- NormalizedTraceMessageWireSchema,
43647
- NormalizedTraceModelWireSchema,
43648
- NormalizedTraceToolWireSchema,
43649
- NormalizedTraceEventWireSchema,
43650
- NormalizedTrajectoryWireSchema,
43651
43764
  TraceRedactionStateWireSchema,
43652
43765
  TraceErrorWireSchema,
43653
43766
  TraceSourceWireSchema,
@@ -43660,15 +43773,14 @@ export {
43660
43773
  TraceToolWireSchema,
43661
43774
  TraceEventWireSchema,
43662
43775
  TraceArtifactWireSchema,
43663
- toNormalizedTrajectoryWire,
43664
- fromNormalizedTrajectoryWire,
43665
43776
  toTraceArtifactWire,
43666
43777
  fromTraceArtifactWire,
43667
43778
  buildTraceFromMessages,
43668
43779
  appendErrorEventToTrace,
43669
43780
  computeTraceSummary,
43670
- getSelectedTrajectoryEvents,
43671
- computeTraceSummaryFromTrajectory,
43781
+ getSelectedTraceArtifactEvents,
43782
+ computeTraceSummaryFromTraceArtifact,
43783
+ traceFromTraceArtifact,
43672
43784
  DEFAULT_EXPLORATION_TOOLS,
43673
43785
  explorationRatio,
43674
43786
  tokensPerTool,
@@ -43842,4 +43954,4 @@ export {
43842
43954
  loadTsEvalFile,
43843
43955
  loadTsEvalSuite
43844
43956
  };
43845
- //# sourceMappingURL=chunk-ENHX2CCS.js.map
43957
+ //# sourceMappingURL=chunk-QCF3LUI7.js.map