@tangle-network/agent-eval 0.221.7 → 0.222.0

package/dist/{systemone-review-DFGtIUcf.js.map → systemone-review-S49gsy_y.js.map}

@@ -1 +1 @@
-{"version":3,"file":"systemone-review-DFGtIUcf.js","names":[],"sources":["../src/systemone-review.ts"],"sourcesContent":["import { z } from 'zod'\nimport {\n  ANALYST_SEVERITIES,\n  type AnalystFinding,\n  type EvidenceRef,\n  makeFinding,\n} from './analyst/types'\nimport { canonicalString, hashCanonical, jsonDocument } from './ledger-core/canonical'\nimport { deepFreezeCanonicalJson } from './ledger-core/deep-freeze'\nimport type { SystemOneQuestions, SystemOneRequest, SystemOneResult } from './systemone-protocol'\nimport { parseSystemOneRequest, parseSystemOneResult } from './systemone-protocol'\n\nconst name = z.string().trim().min(1)\nconst probability = z.number().finite().gt(0).max(1)\nconst checkSchema = z\n  .object({\n    claim: name,\n    area: name,\n    subject: name,\n    severity: z.enum(ANALYST_SEVERITIES),\n    /** Choice labels or score-level keys, never scores invented from category order. */\n    supports: z.array(z.string()).min(1),\n    refutes: z.array(z.string()).min(1),\n    supportAtLeast: probability,\n    refuteAtLeast: probability,\n    /** Host-supplied coverage of the question's required evidence, not model self-assessment. */\n    coverage: z.enum(['complete', 'partial', 'missing']),\n    evidence: z.array(\n      z\n        .object({\n          kind: z.enum(['span', 'event', 'artifact', 'finding', 'metric']),\n          uri: name,\n          excerpt: z.string().optional(),\n        })\n        .strict(),\n    ),\n  })\n  .strict()\n  .refine((value) => value.supportAtLeast + value.refuteAtLeast > 1, {\n    message: 'Support and refutation thresholds must not overlap',\n  })\n\n/** Optional interpretation of native choices/scores. Policy and evidence belong to the caller. */\nexport type SystemOneReviewCheck = z.infer<typeof checkSchema>\nexport interface SystemOneReviewInput<Q extends SystemOneQuestions = SystemOneQuestions> {\n  version: string\n  request: SystemOneRequest<Q>\n  checks: { [K in keyof Q]: SystemOneReviewCheck }\n}\n\nexport interface PreparedSystemOneReview<Q extends SystemOneQuestions = SystemOneQuestions>\n  extends SystemOneReviewInput<Q> {\n  /** Binds questions, option order, context, evidence, thresholds and definition version. */\n  digest: string\n}\n\nexport interface SystemOneReviewAssessment {\n  question: string\n  status: 'supported' | 'refuted' | 'unresolved'\n  reason:\n    | 'support-threshold'\n    | 'refutation-threshold'\n    | 'missing-evidence'\n    | 'partial-negative-evidence'\n    | 'below-threshold'\n    | 'ambiguous-thresholds'\n  /** Unnormalized marginal mass of caller-designated native alternatives. Not calibrated risk. */\n  supportProbability: number\n  refutationProbability: number\n  unresolvedProbability: number\n  modelConfidence: number | null\n  check: SystemOneReviewCheck\n}\n\nexport interface SystemOneReviewReport {\n  reviewDigest: string\n  requestedModel: string\n  servedModel: string\n  assessments: SystemOneReviewAssessment[]\n}\n\nfunction alternatives(question: SystemOneQuestions[string]): string[] {\n  if (question.type === 'noul') return ['true', 'false']\n  return question.type === 'choice'\n    ? Object.keys(question.criteria)\n    : question.criteria.map((_, index) => String(index))\n}\n\n/** Validate before paid admission; make a detached, frozen snapshot without rewriting native JSON. */\nexport function prepareSystemOneReview<const Q extends SystemOneQuestions>(\n  input: SystemOneReviewInput<Q>,\n): PreparedSystemOneReview<Q> {\n  if (!input.version?.trim()) throw new TypeError('Review version is required')\n  parseSystemOneRequest(input.request)\n  // Document form drops only optional undefined object fields; canonicalization rejects loss.\n  const document = jsonDocument(input)\n  canonicalString(document)\n  // Canonical bytes are for hashing, not dispatch: option order can affect a classifier.\n  const snapshot = structuredClone(document) as SystemOneReviewInput<Q>\n  const keys = Object.keys(snapshot.request.questions)\n  if (\n    !snapshot.checks ||\n    Object.keys(snapshot.checks).length !== keys.length ||\n    keys.some((key) => !Object.hasOwn(snapshot.checks, key))\n  ) {\n    throw new TypeError('Review checks must match the exact question names')\n  }\n  for (const key of keys) {\n    const check = checkSchema.parse(snapshot.checks[key])\n    const labels = [...check.supports, ...check.refutes]\n    const allowed = alternatives(snapshot.request.questions[key]!)\n    if (\n      new Set(labels).size !== labels.length ||\n      labels.some((label) => !allowed.includes(label))\n    ) {\n      throw new TypeError('Review alternatives must exist and be disjoint')\n    }\n    if (check.coverage !== 'missing' && check.evidence.length === 0) {\n      throw new TypeError('Review evidence references are required for non-missing coverage')\n    }\n    // Validation must not normalize a label into a different native option.\n    if (canonicalString(jsonDocument(check)) !== canonicalString(snapshot.checks[key])) {\n      throw new TypeError('Review metadata must not require normalization')\n    }\n  }\n  const order = keys.map((key) => [key, alternatives(snapshot.request.questions[key]!)])\n  return deepFreezeCanonicalJson({ ...snapshot, digest: hashCanonical({ snapshot, order }) })\n}\n\n/** Re-score the same retained distribution without new inference or any authorization effect. */\nexport function assessSystemOneReview<Q extends SystemOneQuestions>(\n  review: PreparedSystemOneReview<Q>,\n  raw: unknown,\n): SystemOneReviewReport {\n  const { digest, ...input } = review\n  const prepared = prepareSystemOneReview(input)\n  if (prepared.digest !== digest) throw new TypeError('Review definition changed after preparation')\n  const result = parseSystemOneResult(raw, prepared.request)\n  const assessments = Object.entries(prepared.checks).map(([question, check]) => {\n    const answer = result.answers[question]!\n    const probabilities: Record<string, number> =\n      answer.type === 'noul' ? { true: answer.noul, false: 1 - answer.noul } : answer.probabilities\n    const sum = (labels: string[]) =>\n      labels.reduce((total, label) => total + probabilities[label]!, 0)\n    const supportProbability = sum(check.supports)\n    const refutationProbability = sum(check.refutes)\n    const used = new Set([...check.supports, ...check.refutes])\n    const unresolvedProbability = sum(\n      Object.keys(probabilities).filter((label) => !used.has(label)),\n    )\n    let status: SystemOneReviewAssessment['status'] = 'unresolved'\n    let reason: SystemOneReviewAssessment['reason'] = 'below-threshold'\n    if (supportProbability > 1 || refutationProbability > 1 || unresolvedProbability > 1) {\n      throw new TypeError('Review probability mass exceeds one; do not silently normalize it')\n    }\n    if (check.coverage === 'missing') reason = 'missing-evidence'\n    else if (\n      supportProbability >= check.supportAtLeast &&\n      refutationProbability >= check.refuteAtLeast\n    ) {\n      reason = 'ambiguous-thresholds'\n    } else if (supportProbability >= check.supportAtLeast) {\n      status = 'supported'\n      reason = 'support-threshold'\n    } else if (refutationProbability >= check.refuteAtLeast) {\n      if (check.coverage === 'complete') {\n        status = 'refuted'\n        reason = 'refutation-threshold'\n      } else reason = 'partial-negative-evidence'\n    }\n    return {\n      question,\n      status,\n      reason,\n      supportProbability,\n      refutationProbability,\n      unresolvedProbability,\n      modelConfidence: answer.type === 'noul' ? null : answer.confidence,\n      check: structuredClone(check),\n    }\n  })\n  return {\n    reviewDigest: digest,\n    requestedModel: review.request.model,\n    servedModel: result.model,\n    assessments,\n  }\n}\n\n/** Ordinary analyst findings; a negative review is not a global safety certificate. */\nexport function systemOneReviewFindings<Q extends SystemOneQuestions>(\n  review: PreparedSystemOneReview<Q>,\n  result: SystemOneResult<Q>,\n  options: { analystId: string; producedAt?: string },\n): AnalystFinding[] {\n  if (!options.analystId.trim()) throw new TypeError('Analyst id is required')\n  const report = assessSystemOneReview(review, result)\n  return report.assessments\n    .filter((item) => item.status !== 'refuted')\n    .map((item) =>\n      makeFinding({\n        analyst_id: options.analystId,\n        ...(options.producedAt === undefined ? {} : { produced_at: options.producedAt }),\n        subject: item.check.subject,\n        area: item.status === 'unresolved' ? 'assessment-coverage' : item.check.area,\n        severity: item.status === 'unresolved' ? 'info' : item.check.severity,\n        claim:\n          item.status === 'supported'\n            ? `Model-supported hypothesis: ${item.check.claim}`\n            : `Unresolved review: ${item.check.claim}`,\n        evidence_refs: structuredClone(item.check.evidence) as EvidenceRef[],\n        // For unresolved findings, confidence describes the deterministic unresolved status only.\n        confidence: item.status === 'supported' ? item.supportProbability : 1,\n        derived_from_judge: true,\n        id_basis: hashCanonical([item.question, item.status, review.version, item.check.claim]),\n        metadata: {\n          review_digest: report.reviewDigest,\n          question: item.question,\n          requested_model: report.requestedModel,\n          served_model: report.servedModel,\n          assessment: item.status,\n          assessment_reason: item.reason,\n          evidence_coverage: item.check.coverage,\n          support_probability: item.supportProbability,\n          refutation_probability: item.refutationProbability,\n          unresolved_probability: item.unresolvedProbability,\n          model_confidence: item.modelConfidence,\n          calibration: 'not-established',\n          confidence_basis:\n            item.status === 'supported' ? 'model-support-mass' : 'deterministic-status',\n          native_answer: structuredClone(result.answers[item.question]),\n        },\n      }),\n    )\n}\n"],"mappings":";;;;;;AAYA,MAAM,OAAO,EAAE,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC;AACpC,MAAM,cAAc,EAAE,OAAO,CAAC,CAAC,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC;AACnD,MAAM,cAAc,EACjB,OAAO;CACN,OAAO;CACP,MAAM;CACN,SAAS;CACT,UAAU,EAAE,KAAK,kBAAkB;;CAEnC,UAAU,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC;CACnC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC;CAClC,gBAAgB;CAChB,eAAe;;CAEf,UAAU,EAAE,KAAK;EAAC;EAAY;EAAW;CAAS,CAAC;CACnD,UAAU,EAAE,MACV,EACG,OAAO;EACN,MAAM,EAAE,KAAK;GAAC;GAAQ;GAAS;GAAY;GAAW;EAAQ,CAAC;EAC/D,KAAK;EACL,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS;CAC/B,CAAC,CAAC,CACD,OAAO,CACZ;AACF,CAAC,CAAC,CACD,OAAO,CAAC,CACR,QAAQ,UAAU,MAAM,iBAAiB,MAAM,gBAAgB,GAAG,EACjE,SAAS,qDACX,CAAC;AAyCH,SAAS,aAAa,UAAgD;CACpE,IAAI,SAAS,SAAS,QAAQ,OAAO,CAAC,QAAQ,OAAO;CACrD,OAAO,SAAS,SAAS,WACrB,OAAO,KAAK,SAAS,QAAQ,IAC7B,SAAS,SAAS,KAAK,GAAG,UAAU,OAAO,KAAK,CAAC;AACvD;;AAGA,SAAgB,uBACd,OAC4B;CAC5B,IAAI,CAAC,MAAM,SAAS,KAAK,GAAG,MAAM,IAAI,UAAU,4BAA4B;CAC5E,sBAAsB,MAAM,OAAO;CAEnC,MAAM,WAAW,aAAa,KAAK;CACnC,gBAAgB,QAAQ;CAExB,MAAM,WAAW,gBAAgB,QAAQ;CACzC,MAAM,OAAO,OAAO,KAAK,SAAS,QAAQ,SAAS;CACnD,IACE,CAAC,SAAS,UACV,OAAO,KAAK,SAAS,MAAM,CAAC,CAAC,WAAW,KAAK,UAC7C,KAAK,MAAM,QAAQ,CAAC,OAAO,OAAO,SAAS,QAAQ,GAAG,CAAC,GAEvD,MAAM,IAAI,UAAU,mDAAmD;CAEzE,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,YAAY,MAAM,SAAS,OAAO,IAAI;EACpD,MAAM,SAAS,CAAC,GAAG,MAAM,UAAU,GAAG,MAAM,OAAO;EACnD,MAAM,UAAU,aAAa,SAAS,QAAQ,UAAU,IAAK;EAC7D,IACE,IAAI,IAAI,MAAM,CAAC,CAAC,SAAS,OAAO,UAChC,OAAO,MAAM,UAAU,CAAC,QAAQ,SAAS,KAAK,CAAC,GAE/C,MAAM,IAAI,UAAU,gDAAgD;EAEtE,IAAI,MAAM,aAAa,aAAa,MAAM,SAAS,WAAW,GAC5D,MAAM,IAAI,UAAU,kEAAkE;EAGxF,IAAI,gBAAgB,aAAa,KAAK,CAAC,MAAM,gBAAgB,SAAS,OAAO,IAAI,GAC/E,MAAM,IAAI,UAAU,gDAAgD;CAExE;CACA,MAAM,QAAQ,KAAK,KAAK,QAAQ,CAAC,KAAK,aAAa,SAAS,QAAQ,UAAU,IAAK,CAAC,CAAC;CACrF,OAAO,wBAAwB;EAAE,GAAG;EAAU,QAAQ,cAAc;GAAE;GAAU;EAAM,CAAC;CAAE,CAAC;AAC5F;;AAGA,SAAgB,sBACd,QACA,KACuB;CACvB,MAAM,EAAE,QAAQ,GAAG,UAAU;CAC7B,MAAM,WAAW,uBAAuB,KAAK;CAC7C,IAAI,SAAS,WAAW,QAAQ,MAAM,IAAI,UAAU,6CAA6C;CACjG,MAAM,SAAS,qBAAqB,KAAK,SAAS,OAAO;CACzD,MAAM,cAAc,OAAO,QAAQ,SAAS,MAAM,CAAC,CAAC,KAAK,CAAC,UAAU,WAAW;EAC7E,MAAM,SAAS,OAAO,QAAQ;EAC9B,MAAM,gBACJ,OAAO,SAAS,SAAS;GAAE,MAAM,OAAO;GAAM,OAAO,IAAI,OAAO;EAAK,IAAI,OAAO;EAClF,MAAM,OAAO,WACX,OAAO,QAAQ,OAAO,UAAU,QAAQ,cAAc,QAAS,CAAC;EAClE,MAAM,qBAAqB,IAAI,MAAM,QAAQ;EAC7C,MAAM,wBAAwB,IAAI,MAAM,OAAO;EAC/C,MAAM,uBAAO,IAAI,IAAI,CAAC,GAAG,MAAM,UAAU,GAAG,MAAM,OAAO,CAAC;EAC1D,MAAM,wBAAwB,IAC5B,OAAO,KAAK,aAAa,CAAC,CAAC,QAAQ,UAAU,CAAC,KAAK,IAAI,KAAK,CAAC,CAC/D;EACA,IAAI,SAA8C;EAClD,IAAI,SAA8C;EAClD,IAAI,qBAAqB,KAAK,wBAAwB,KAAK,wBAAwB,GACjF,MAAM,IAAI,UAAU,mEAAmE;EAEzF,IAAI,MAAM,aAAa,WAAW,SAAS;OACtC,IACH,sBAAsB,MAAM,kBAC5B,yBAAyB,MAAM,eAE/B,SAAS;OACJ,IAAI,sBAAsB,MAAM,gBAAgB;GACrD,SAAS;GACT,SAAS;EACX,OAAO,IAAI,yBAAyB,MAAM,eAAe;GACvD,IAAI,MAAM,aAAa,YAAY;IACjC,SAAS;IACT,SAAS;GACX,OAAO,SAAS;EAClB;EACA,OAAO;GACL;GACA;GACA;GACA;GACA;GACA;GACA,iBAAiB,OAAO,SAAS,SAAS,OAAO,OAAO;GACxD,OAAO,gBAAgB,KAAK;EAC9B;CACF,CAAC;CACD,OAAO;EACL,cAAc;EACd,gBAAgB,OAAO,QAAQ;EAC/B,aAAa,OAAO;EACpB;CACF;AACF;;AAGA,SAAgB,wBACd,QACA,QACA,SACkB;CAClB,IAAI,CAAC,QAAQ,UAAU,KAAK,GAAG,MAAM,IAAI,UAAU,wBAAwB;CAC3E,MAAM,SAAS,sBAAsB,QAAQ,MAAM;CACnD,OAAO,OAAO,YACX,QAAQ,SAAS,KAAK,WAAW,SAAS,CAAC,CAC3C,KAAK,SACJ,YAAY;EACV,YAAY,QAAQ;EACpB,GAAI,QAAQ,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,QAAQ,WAAW;EAC9E,SAAS,KAAK,MAAM;EACpB,MAAM,KAAK,WAAW,eAAe,wBAAwB,KAAK,MAAM;EACxE,UAAU,KAAK,WAAW,eAAe,SAAS,KAAK,MAAM;EAC7D,OACE,KAAK,WAAW,cACZ,+BAA+B,KAAK,MAAM,UAC1C,sBAAsB,KAAK,MAAM;EACvC,eAAe,gBAAgB,KAAK,MAAM,QAAQ;EAElD,YAAY,KAAK,WAAW,cAAc,KAAK,qBAAqB;EACpE,oBAAoB;EACpB,UAAU,cAAc;GAAC,KAAK;GAAU,KAAK;GAAQ,OAAO;GAAS,KAAK,MAAM;EAAK,CAAC;EACtF,UAAU;GACR,eAAe,OAAO;GACtB,UAAU,KAAK;GACf,iBAAiB,OAAO;GACxB,cAAc,OAAO;GACrB,YAAY,KAAK;GACjB,mBAAmB,KAAK;GACxB,mBAAmB,KAAK,MAAM;GAC9B,qBAAqB,KAAK;GAC1B,wBAAwB,KAAK;GAC7B,wBAAwB,KAAK;GAC7B,kBAAkB,KAAK;GACvB,aAAa;GACb,kBACE,KAAK,WAAW,cAAc,uBAAuB;GACvD,eAAe,gBAAgB,OAAO,QAAQ,KAAK,SAAS;EAC9D;CACF,CAAC,CACH;AACJ"}
+{"version":3,"file":"systemone-review-S49gsy_y.js","names":[],"sources":["../src/systemone-review.ts"],"sourcesContent":["import { z } from 'zod'\nimport {\n  ANALYST_SEVERITIES,\n  type AnalystFinding,\n  type EvidenceRef,\n  makeFinding,\n} from './analyst/types'\nimport { canonicalString, hashCanonical, jsonDocument } from './ledger-core/canonical'\nimport { deepFreezeCanonicalJson } from './ledger-core/deep-freeze'\nimport type { SystemOneQuestions, SystemOneRequest, SystemOneResult } from './systemone-protocol'\nimport { parseSystemOneRequest, parseSystemOneResult } from './systemone-protocol'\n\nconst name = z.string().trim().min(1)\nconst probability = z.number().finite().gt(0).max(1)\nconst checkSchema = z\n  .object({\n    claim: name,\n    area: name,\n    subject: name,\n    severity: z.enum(ANALYST_SEVERITIES),\n    /** Choice labels or score-level keys, never scores invented from category order. */\n    supports: z.array(z.string()).min(1),\n    refutes: z.array(z.string()).min(1),\n    supportAtLeast: probability,\n    refuteAtLeast: probability,\n    /** Host-supplied coverage of the question's required evidence, not model self-assessment. */\n    coverage: z.enum(['complete', 'partial', 'missing']),\n    evidence: z.array(\n      z\n        .object({\n          kind: z.enum(['span', 'event', 'artifact', 'finding', 'metric']),\n          uri: name,\n          excerpt: z.string().optional(),\n        })\n        .strict(),\n    ),\n  })\n  .strict()\n  .refine((value) => value.supportAtLeast + value.refuteAtLeast > 1, {\n    message: 'Support and refutation thresholds must not overlap',\n  })\n\n/** Optional interpretation of native choices/scores. Policy and evidence belong to the caller. */\nexport type SystemOneReviewCheck = z.infer<typeof checkSchema>\nexport interface SystemOneReviewInput<Q extends SystemOneQuestions = SystemOneQuestions> {\n  version: string\n  request: SystemOneRequest<Q>\n  checks: { [K in keyof Q]: SystemOneReviewCheck }\n}\n\nexport interface PreparedSystemOneReview<Q extends SystemOneQuestions = SystemOneQuestions>\n  extends SystemOneReviewInput<Q> {\n  /** Binds questions, option order, context, evidence, thresholds and definition version. */\n  digest: string\n}\n\nexport interface SystemOneReviewAssessment {\n  question: string\n  status: 'supported' | 'refuted' | 'unresolved'\n  reason:\n    | 'support-threshold'\n    | 'refutation-threshold'\n    | 'missing-evidence'\n    | 'partial-negative-evidence'\n    | 'below-threshold'\n    | 'ambiguous-thresholds'\n  /** Unnormalized marginal mass of caller-designated native alternatives. Not calibrated risk. */\n  supportProbability: number\n  refutationProbability: number\n  unresolvedProbability: number\n  modelConfidence: number | null\n  check: SystemOneReviewCheck\n}\n\nexport interface SystemOneReviewReport {\n  reviewDigest: string\n  requestedModel: string\n  servedModel: string\n  assessments: SystemOneReviewAssessment[]\n}\n\nfunction alternatives(question: SystemOneQuestions[string]): string[] {\n  if (question.type === 'noul') return ['true', 'false']\n  return question.type === 'choice'\n    ? Object.keys(question.criteria)\n    : question.criteria.map((_, index) => String(index))\n}\n\n/** Validate before paid admission; make a detached, frozen snapshot without rewriting native JSON. */\nexport function prepareSystemOneReview<const Q extends SystemOneQuestions>(\n  input: SystemOneReviewInput<Q>,\n): PreparedSystemOneReview<Q> {\n  if (!input.version?.trim()) throw new TypeError('Review version is required')\n  parseSystemOneRequest(input.request)\n  // Document form drops only optional undefined object fields; canonicalization rejects loss.\n  const document = jsonDocument(input)\n  canonicalString(document)\n  // Canonical bytes are for hashing, not dispatch: option order can affect a classifier.\n  const snapshot = structuredClone(document) as SystemOneReviewInput<Q>\n  const keys = Object.keys(snapshot.request.questions)\n  if (\n    !snapshot.checks ||\n    Object.keys(snapshot.checks).length !== keys.length ||\n    keys.some((key) => !Object.hasOwn(snapshot.checks, key))\n  ) {\n    throw new TypeError('Review checks must match the exact question names')\n  }\n  for (const key of keys) {\n    const check = checkSchema.parse(snapshot.checks[key])\n    const labels = [...check.supports, ...check.refutes]\n    const allowed = alternatives(snapshot.request.questions[key]!)\n    if (\n      new Set(labels).size !== labels.length ||\n      labels.some((label) => !allowed.includes(label))\n    ) {\n      throw new TypeError('Review alternatives must exist and be disjoint')\n    }\n    if (check.coverage !== 'missing' && check.evidence.length === 0) {\n      throw new TypeError('Review evidence references are required for non-missing coverage')\n    }\n    // Validation must not normalize a label into a different native option.\n    if (canonicalString(jsonDocument(check)) !== canonicalString(snapshot.checks[key])) {\n      throw new TypeError('Review metadata must not require normalization')\n    }\n  }\n  const order = keys.map((key) => [key, alternatives(snapshot.request.questions[key]!)])\n  return deepFreezeCanonicalJson({ ...snapshot, digest: hashCanonical({ snapshot, order }) })\n}\n\n/** Re-score the same retained distribution without new inference or any authorization effect. */\nexport function assessSystemOneReview<Q extends SystemOneQuestions>(\n  review: PreparedSystemOneReview<Q>,\n  raw: unknown,\n): SystemOneReviewReport {\n  const { digest, ...input } = review\n  const prepared = prepareSystemOneReview(input)\n  if (prepared.digest !== digest) throw new TypeError('Review definition changed after preparation')\n  const result = parseSystemOneResult(raw, prepared.request)\n  const assessments = Object.entries(prepared.checks).map(([question, check]) => {\n    const answer = result.answers[question]!\n    const probabilities: Record<string, number> =\n      answer.type === 'noul' ? { true: answer.noul, false: 1 - answer.noul } : answer.probabilities\n    const sum = (labels: string[]) =>\n      labels.reduce((total, label) => total + probabilities[label]!, 0)\n    const supportProbability = sum(check.supports)\n    const refutationProbability = sum(check.refutes)\n    const used = new Set([...check.supports, ...check.refutes])\n    const unresolvedProbability = sum(\n      Object.keys(probabilities).filter((label) => !used.has(label)),\n    )\n    let status: SystemOneReviewAssessment['status'] = 'unresolved'\n    let reason: SystemOneReviewAssessment['reason'] = 'below-threshold'\n    if (supportProbability > 1 || refutationProbability > 1 || unresolvedProbability > 1) {\n      throw new TypeError('Review probability mass exceeds one; do not silently normalize it')\n    }\n    if (check.coverage === 'missing') reason = 'missing-evidence'\n    else if (\n      supportProbability >= check.supportAtLeast &&\n      refutationProbability >= check.refuteAtLeast\n    ) {\n      reason = 'ambiguous-thresholds'\n    } else if (supportProbability >= check.supportAtLeast) {\n      status = 'supported'\n      reason = 'support-threshold'\n    } else if (refutationProbability >= check.refuteAtLeast) {\n      if (check.coverage === 'complete') {\n        status = 'refuted'\n        reason = 'refutation-threshold'\n      } else reason = 'partial-negative-evidence'\n    }\n    return {\n      question,\n      status,\n      reason,\n      supportProbability,\n      refutationProbability,\n      unresolvedProbability,\n      modelConfidence: answer.type === 'noul' ? null : answer.confidence,\n      check: structuredClone(check),\n    }\n  })\n  return {\n    reviewDigest: digest,\n    requestedModel: review.request.model,\n    servedModel: result.model,\n    assessments,\n  }\n}\n\n/** Ordinary analyst findings; a negative review is not a global safety certificate. */\nexport function systemOneReviewFindings<Q extends SystemOneQuestions>(\n  review: PreparedSystemOneReview<Q>,\n  result: SystemOneResult<Q>,\n  options: { analystId: string; producedAt?: string },\n): AnalystFinding[] {\n  if (!options.analystId.trim()) throw new TypeError('Analyst id is required')\n  const report = assessSystemOneReview(review, result)\n  return report.assessments\n    .filter((item) => item.status !== 'refuted')\n    .map((item) =>\n      makeFinding({\n        analyst_id: options.analystId,\n        ...(options.producedAt === undefined ? {} : { produced_at: options.producedAt }),\n        subject: item.check.subject,\n        area: item.status === 'unresolved' ? 'assessment-coverage' : item.check.area,\n        severity: item.status === 'unresolved' ? 'info' : item.check.severity,\n        claim:\n          item.status === 'supported'\n            ? `Model-supported hypothesis: ${item.check.claim}`\n            : `Unresolved review: ${item.check.claim}`,\n        evidence_refs: structuredClone(item.check.evidence) as EvidenceRef[],\n        // For unresolved findings, confidence describes the deterministic unresolved status only.\n        confidence: item.status === 'supported' ? item.supportProbability : 1,\n        derived_from_judge: true,\n        id_basis: hashCanonical([item.question, item.status, review.version, item.check.claim]),\n        metadata: {\n          review_digest: report.reviewDigest,\n          question: item.question,\n          requested_model: report.requestedModel,\n          served_model: report.servedModel,\n          assessment: item.status,\n          assessment_reason: item.reason,\n          evidence_coverage: item.check.coverage,\n          support_probability: item.supportProbability,\n          refutation_probability: item.refutationProbability,\n          unresolved_probability: item.unresolvedProbability,\n          model_confidence: item.modelConfidence,\n          calibration: 'not-established',\n          confidence_basis:\n            item.status === 'supported' ? 'model-support-mass' : 'deterministic-status',\n          native_answer: structuredClone(result.answers[item.question]),\n        },\n      }),\n    )\n}\n"],"mappings":";;;;;;AAYA,MAAM,OAAO,EAAE,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC;AACpC,MAAM,cAAc,EAAE,OAAO,CAAC,CAAC,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC;AACnD,MAAM,cAAc,EACjB,OAAO;CACN,OAAO;CACP,MAAM;CACN,SAAS;CACT,UAAU,EAAE,KAAK,kBAAkB;;CAEnC,UAAU,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC;CACnC,SAAS,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC;CAClC,gBAAgB;CAChB,eAAe;;CAEf,UAAU,EAAE,KAAK;EAAC;EAAY;EAAW;CAAS,CAAC;CACnD,UAAU,EAAE,MACV,EACG,OAAO;EACN,MAAM,EAAE,KAAK;GAAC;GAAQ;GAAS;GAAY;GAAW;EAAQ,CAAC;EAC/D,KAAK;EACL,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS;CAC/B,CAAC,CAAC,CACD,OAAO,CACZ;AACF,CAAC,CAAC,CACD,OAAO,CAAC,CACR,QAAQ,UAAU,MAAM,iBAAiB,MAAM,gBAAgB,GAAG,EACjE,SAAS,qDACX,CAAC;AAyCH,SAAS,aAAa,UAAgD;CACpE,IAAI,SAAS,SAAS,QAAQ,OAAO,CAAC,QAAQ,OAAO;CACrD,OAAO,SAAS,SAAS,WACrB,OAAO,KAAK,SAAS,QAAQ,IAC7B,SAAS,SAAS,KAAK,GAAG,UAAU,OAAO,KAAK,CAAC;AACvD;;AAGA,SAAgB,uBACd,OAC4B;CAC5B,IAAI,CAAC,MAAM,SAAS,KAAK,GAAG,MAAM,IAAI,UAAU,4BAA4B;CAC5E,sBAAsB,MAAM,OAAO;CAEnC,MAAM,WAAW,aAAa,KAAK;CACnC,gBAAgB,QAAQ;CAExB,MAAM,WAAW,gBAAgB,QAAQ;CACzC,MAAM,OAAO,OAAO,KAAK,SAAS,QAAQ,SAAS;CACnD,IACE,CAAC,SAAS,UACV,OAAO,KAAK,SAAS,MAAM,CAAC,CAAC,WAAW,KAAK,UAC7C,KAAK,MAAM,QAAQ,CAAC,OAAO,OAAO,SAAS,QAAQ,GAAG,CAAC,GAEvD,MAAM,IAAI,UAAU,mDAAmD;CAEzE,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,YAAY,MAAM,SAAS,OAAO,IAAI;EACpD,MAAM,SAAS,CAAC,GAAG,MAAM,UAAU,GAAG,MAAM,OAAO;EACnD,MAAM,UAAU,aAAa,SAAS,QAAQ,UAAU,IAAK;EAC7D,IACE,IAAI,IAAI,MAAM,CAAC,CAAC,SAAS,OAAO,UAChC,OAAO,MAAM,UAAU,CAAC,QAAQ,SAAS,KAAK,CAAC,GAE/C,MAAM,IAAI,UAAU,gDAAgD;EAEtE,IAAI,MAAM,aAAa,aAAa,MAAM,SAAS,WAAW,GAC5D,MAAM,IAAI,UAAU,kEAAkE;EAGxF,IAAI,gBAAgB,aAAa,KAAK,CAAC,MAAM,gBAAgB,SAAS,OAAO,IAAI,GAC/E,MAAM,IAAI,UAAU,gDAAgD;CAExE;CACA,MAAM,QAAQ,KAAK,KAAK,QAAQ,CAAC,KAAK,aAAa,SAAS,QAAQ,UAAU,IAAK,CAAC,CAAC;CACrF,OAAO,wBAAwB;EAAE,GAAG;EAAU,QAAQ,cAAc;GAAE;GAAU;EAAM,CAAC;CAAE,CAAC;AAC5F;;AAGA,SAAgB,sBACd,QACA,KACuB;CACvB,MAAM,EAAE,QAAQ,GAAG,UAAU;CAC7B,MAAM,WAAW,uBAAuB,KAAK;CAC7C,IAAI,SAAS,WAAW,QAAQ,MAAM,IAAI,UAAU,6CAA6C;CACjG,MAAM,SAAS,qBAAqB,KAAK,SAAS,OAAO;CACzD,MAAM,cAAc,OAAO,QAAQ,SAAS,MAAM,CAAC,CAAC,KAAK,CAAC,UAAU,WAAW;EAC7E,MAAM,SAAS,OAAO,QAAQ;EAC9B,MAAM,gBACJ,OAAO,SAAS,SAAS;GAAE,MAAM,OAAO;GAAM,OAAO,IAAI,OAAO;EAAK,IAAI,OAAO;EAClF,MAAM,OAAO,WACX,OAAO,QAAQ,OAAO,UAAU,QAAQ,cAAc,QAAS,CAAC;EAClE,MAAM,qBAAqB,IAAI,MAAM,QAAQ;EAC7C,MAAM,wBAAwB,IAAI,MAAM,OAAO;EAC/C,MAAM,uBAAO,IAAI,IAAI,CAAC,GAAG,MAAM,UAAU,GAAG,MAAM,OAAO,CAAC;EAC1D,MAAM,wBAAwB,IAC5B,OAAO,KAAK,aAAa,CAAC,CAAC,QAAQ,UAAU,CAAC,KAAK,IAAI,KAAK,CAAC,CAC/D;EACA,IAAI,SAA8C;EAClD,IAAI,SAA8C;EAClD,IAAI,qBAAqB,KAAK,wBAAwB,KAAK,wBAAwB,GACjF,MAAM,IAAI,UAAU,mEAAmE;EAEzF,IAAI,MAAM,aAAa,WAAW,SAAS;OACtC,IACH,sBAAsB,MAAM,kBAC5B,yBAAyB,MAAM,eAE/B,SAAS;OACJ,IAAI,sBAAsB,MAAM,gBAAgB;GACrD,SAAS;GACT,SAAS;EACX,OAAO,IAAI,yBAAyB,MAAM,eAAe;GACvD,IAAI,MAAM,aAAa,YAAY;IACjC,SAAS;IACT,SAAS;GACX,OAAO,SAAS;EAClB;EACA,OAAO;GACL;GACA;GACA;GACA;GACA;GACA;GACA,iBAAiB,OAAO,SAAS,SAAS,OAAO,OAAO;GACxD,OAAO,gBAAgB,KAAK;EAC9B;CACF,CAAC;CACD,OAAO;EACL,cAAc;EACd,gBAAgB,OAAO,QAAQ;EAC/B,aAAa,OAAO;EACpB;CACF;AACF;;AAGA,SAAgB,wBACd,QACA,QACA,SACkB;CAClB,IAAI,CAAC,QAAQ,UAAU,KAAK,GAAG,MAAM,IAAI,UAAU,wBAAwB;CAC3E,MAAM,SAAS,sBAAsB,QAAQ,MAAM;CACnD,OAAO,OAAO,YACX,QAAQ,SAAS,KAAK,WAAW,SAAS,CAAC,CAC3C,KAAK,SACJ,YAAY;EACV,YAAY,QAAQ;EACpB,GAAI,QAAQ,eAAe,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,QAAQ,WAAW;EAC9E,SAAS,KAAK,MAAM;EACpB,MAAM,KAAK,WAAW,eAAe,wBAAwB,KAAK,MAAM;EACxE,UAAU,KAAK,WAAW,eAAe,SAAS,KAAK,MAAM;EAC7D,OACE,KAAK,WAAW,cACZ,+BAA+B,KAAK,MAAM,UAC1C,sBAAsB,KAAK,MAAM;EACvC,eAAe,gBAAgB,KAAK,MAAM,QAAQ;EAElD,YAAY,KAAK,WAAW,cAAc,KAAK,qBAAqB;EACpE,oBAAoB;EACpB,UAAU,cAAc;GAAC,KAAK;GAAU,KAAK;GAAQ,OAAO;GAAS,KAAK,MAAM;EAAK,CAAC;EACtF,UAAU;GACR,eAAe,OAAO;GACtB,UAAU,KAAK;GACf,iBAAiB,OAAO;GACxB,cAAc,OAAO;GACrB,YAAY,KAAK;GACjB,mBAAmB,KAAK;GACxB,mBAAmB,KAAK,MAAM;GAC9B,qBAAqB,KAAK;GAC1B,wBAAwB,KAAK;GAC7B,wBAAwB,KAAK;GAC7B,kBAAkB,KAAK;GACvB,aAAa;GACb,kBACE,KAAK,WAAW,cAAc,uBAAuB;GACvD,eAAe,gBAAgB,OAAO,QAAQ,KAAK,SAAS;EAC9D;CACF,CAAC,CACH;AACJ"}

package/dist/systemone.d.ts

@@ -1,5 +1,5 @@
-import { a as EvaluationResult, c as asJudge, i as EvaluationJudgeOptions, l as createEvaluator, n as EvaluationAnalystOptions, o as EvaluatorOptions, r as EvaluationContext, s as asAnalyst, t as CostedEvaluator } from "./evaluation-DblGquCB.js";
-import { C as systemOneUsage, S as systemOneConfidence, _ as normalizeSystemOneResult, a as SystemOneJson, b as parseSystemOneRequest, c as SystemOneRecordedQuestion, d as SystemOneRequest, f as SystemOneResponseError, g as SystemOneState, h as SystemOneScoreCriteria, i as SystemOneDescription, l as SystemOneRecordedQuestions, m as SystemOneResultProfile, n as SystemOneAnswer, o as SystemOneQuestion, p as SystemOneResult, r as SystemOneAnswerFor, s as SystemOneQuestions, u as SystemOneRecordedRequest, x as parseSystemOneResult, y as parseSystemOneRecordedRequest } from "./systemone-protocol-Bl2EI2KQ.js";
-import { C as DecisionPointResult, D as decisionQuestionDigest, E as decisionPoint, O as DecisionRecord, S as DecisionPointOptions, T as DecisionQuestions, _ as assessSystemOneDomainMatch, a as SystemOneOptions, b as DECISION_POINT_MIN_KAPPA, c as systemOneEvaluator, d as SystemOneDomainCoverage, f as SystemOneDomainEvidence, g as SystemOneDomainMatchThresholds, h as SystemOneDomainMatchReport, i as SystemOneJudgeOptions, k as DecisionRoute, l as systemOneJudge, m as SystemOneDomainMatchReason, n as SystemOneEvaluate, o as normalizedSystemOneScore, p as SystemOneDomainMatchInput, r as SystemOneEvaluatorOptions, s as systemOneAnalyst, t as SystemOneAnalystOptions, u as PreparedSystemOneDomainMatch, v as prepareSystemOneDomainMatch, w as DecisionQuestionPolicy, x as DecisionInvocation, y as systemOneChoiceMargin } from "./systemone-CHnqwmYg.js";
-import { a as SystemOneReviewReport, c as systemOneReviewFindings, i as SystemOneReviewInput, n as SystemOneReviewAssessment, o as assessSystemOneReview, r as SystemOneReviewCheck, s as prepareSystemOneReview, t as PreparedSystemOneReview } from "./systemone-review-cQ9acAsb.js";
+import { a as EvaluationResult, c as asJudge, i as EvaluationJudgeOptions, l as createEvaluator, n as EvaluationAnalystOptions, o as EvaluatorOptions, r as EvaluationContext, s as asAnalyst, t as CostedEvaluator } from "./evaluation-BPqUvZ66.js";
+import { C as systemOneUsage, S as systemOneConfidence, _ as normalizeSystemOneResult, a as SystemOneJson, b as parseSystemOneRequest, c as SystemOneRecordedQuestion, d as SystemOneRequest, f as SystemOneResponseError, g as SystemOneState, h as SystemOneScoreCriteria, i as SystemOneDescription, l as SystemOneRecordedQuestions, m as SystemOneResultProfile, n as SystemOneAnswer, o as SystemOneQuestion, p as SystemOneResult, r as SystemOneAnswerFor, s as SystemOneQuestions, u as SystemOneRecordedRequest, x as parseSystemOneResult, y as parseSystemOneRecordedRequest } from "./systemone-protocol-C4t7JH4Y.js";
+import { C as DecisionPointResult, D as decisionQuestionDigest, E as decisionPoint, O as DecisionRecord, S as DecisionPointOptions, T as DecisionQuestions, _ as assessSystemOneDomainMatch, a as SystemOneOptions, b as DECISION_POINT_MIN_KAPPA, c as systemOneEvaluator, d as SystemOneDomainCoverage, f as SystemOneDomainEvidence, g as SystemOneDomainMatchThresholds, h as SystemOneDomainMatchReport, i as SystemOneJudgeOptions, k as DecisionRoute, l as systemOneJudge, m as SystemOneDomainMatchReason, n as SystemOneEvaluate, o as normalizedSystemOneScore, p as SystemOneDomainMatchInput, r as SystemOneEvaluatorOptions, s as systemOneAnalyst, t as SystemOneAnalystOptions, u as PreparedSystemOneDomainMatch, v as prepareSystemOneDomainMatch, w as DecisionQuestionPolicy, x as DecisionInvocation, y as systemOneChoiceMargin } from "./systemone-Cw2txK5m.js";
+import { a as SystemOneReviewReport, c as systemOneReviewFindings, i as SystemOneReviewInput, n as SystemOneReviewAssessment, o as assessSystemOneReview, r as SystemOneReviewCheck, s as prepareSystemOneReview, t as PreparedSystemOneReview } from "./systemone-review-BLgxt-4Q.js";
 export { CostedEvaluator, DECISION_POINT_MIN_KAPPA, DecisionInvocation, DecisionPointOptions, DecisionPointResult, DecisionQuestionPolicy, DecisionQuestions, type DecisionRecord, type DecisionRoute, EvaluationAnalystOptions, EvaluationContext, EvaluationJudgeOptions, EvaluationResult, EvaluatorOptions, PreparedSystemOneDomainMatch, PreparedSystemOneReview, SystemOneAnalystOptions, type SystemOneAnswer, type SystemOneAnswerFor, type SystemOneDescription, SystemOneDomainCoverage, SystemOneDomainEvidence, SystemOneDomainMatchInput, SystemOneDomainMatchReason, SystemOneDomainMatchReport, SystemOneDomainMatchThresholds, SystemOneEvaluate, SystemOneEvaluatorOptions, type SystemOneJson, SystemOneJudgeOptions, SystemOneOptions, type SystemOneQuestion, type SystemOneQuestions, type SystemOneRecordedQuestion, type SystemOneRecordedQuestions, type SystemOneRecordedRequest, type SystemOneRequest, SystemOneResponseError, type SystemOneResult, type SystemOneResultProfile, SystemOneReviewAssessment, SystemOneReviewCheck, SystemOneReviewInput, SystemOneReviewReport, type SystemOneScoreCriteria, type SystemOneState, asAnalyst, asJudge, assessSystemOneDomainMatch, assessSystemOneReview, createEvaluator, decisionPoint, decisionQuestionDigest, normalizeSystemOneResult, normalizedSystemOneScore, parseSystemOneRecordedRequest, parseSystemOneRequest, parseSystemOneResult, prepareSystemOneDomainMatch, prepareSystemOneReview, systemOneAnalyst, systemOneChoiceMargin, systemOneConfidence, systemOneEvaluator, systemOneJudge, systemOneReviewFindings, systemOneUsage };

package/dist/systemone.js

@@ -1,5 +1,5 @@
 import { SystemOneResponseError, normalizeSystemOneResult, parseSystemOneRecordedRequest, parseSystemOneRequest, parseSystemOneResult, systemOneConfidence, systemOneUsage } from "./systemone-protocol.js";
-import { a as systemOneAnalyst, c as assessSystemOneDomainMatch, i as normalizedSystemOneScore, l as prepareSystemOneDomainMatch, n as decisionPoint, o as systemOneEvaluator, r as decisionQuestionDigest, s as systemOneJudge, t as DECISION_POINT_MIN_KAPPA, u as systemOneChoiceMargin } from "./decision-point-0nJubuoM.js";
-import { n as asJudge, r as createEvaluator, t as asAnalyst } from "./evaluation-wdGyZ7jV.js";
-import { n as prepareSystemOneReview, r as systemOneReviewFindings, t as assessSystemOneReview } from "./systemone-review-DFGtIUcf.js";
+import { a as systemOneAnalyst, c as assessSystemOneDomainMatch, i as normalizedSystemOneScore, l as prepareSystemOneDomainMatch, n as decisionPoint, o as systemOneEvaluator, r as decisionQuestionDigest, s as systemOneJudge, t as DECISION_POINT_MIN_KAPPA, u as systemOneChoiceMargin } from "./decision-point-lhMgODOR.js";
+import { n as asJudge, r as createEvaluator, t as asAnalyst } from "./evaluation-B8H3GYej.js";
+import { n as prepareSystemOneReview, r as systemOneReviewFindings, t as assessSystemOneReview } from "./systemone-review-S49gsy_y.js";
 export { DECISION_POINT_MIN_KAPPA, SystemOneResponseError, asAnalyst, asJudge, assessSystemOneDomainMatch, assessSystemOneReview, createEvaluator, decisionPoint, decisionQuestionDigest, normalizeSystemOneResult, normalizedSystemOneScore, parseSystemOneRecordedRequest, parseSystemOneRequest, parseSystemOneResult, prepareSystemOneDomainMatch, prepareSystemOneReview, systemOneAnalyst, systemOneChoiceMargin, systemOneConfidence, systemOneEvaluator, systemOneJudge, systemOneReviewFindings, systemOneUsage };

package/dist/{task-failure-attributes-BMsblaDh.js → task-failure-attributes-CC6VuIpk.js}

@@ -1,9 +1,9 @@
-import { s as ValidationError } from "./errors-BcOs5MKW.js";
-import { t as FAILURE_CLASSES } from "./schema-Bt2vLFm3.js";
+import { s as ValidationError } from "./errors-DeEEV9Vx.js";
+import { t as FAILURE_CLASSES } from "./schema-D-xu3MU_.js";
 import { LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_COST_ATTR_KEYS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKEN_ATTR_KEYS, RUN_COST_ATTR_KEYS } from "./trace-attributes.js";
-import "./otlp-attributes-D9ptWaIJ.js";
+import "./otlp-attributes-Boeb7gtH.js";
 import { isModelCallKind } from "@tangle-network/agent-trace-contract";
-//#region src/trace/error-classification.ts
+// src/trace/error-classification.ts
 /**
 * Classify errored spans without counting a propagated parent status as a
 * second execution failure.
@@ -48,8 +48,7 @@ function summarizeTraceErrors(signals) {
 	}
 	return summary;
 }
-//#endregion
-//#region src/trace/execution-measurements.ts
+// src/trace/execution-measurements.ts
 const TOKEN_MEASUREMENT_KEY_GROUPS = [
 	LLM_INPUT_TOKEN_ATTR_KEYS,
 	LLM_OUTPUT_TOKEN_ATTR_KEYS,
@@ -95,6 +94,7 @@ function summarizeExecutionMeasurements(source) {
 	const output = reconcileMeasurement(calls, byId, aggregateSourceIds, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKEN_ATTR_KEYS);
 	const cached = reconcileMeasurement(calls, byId, aggregateSourceIds, LLM_CACHED_TOKEN_ATTR_KEYS);
 	const cacheWrite = reconcileMeasurement(calls, byId, aggregateSourceIds, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS);
+	const tokensUnknown = calls.length > 0 && (!input.complete || !output.complete || calls.some((call) => call.attributes["tangle.supervise.tokens_known"] === false));
 	const aggregate = summarizeAggregateMeasurements(spans, byId, /* @__PURE__ */ new Set([...aggregateIds, ...untypedRunCostIds]));
 	const reasoningValue = reasoning.value ?? aggregate?.tokenUsage.reasoning;
 	const cachedValue = cached.value ?? aggregate?.tokenUsage.cached;
@@ -103,6 +103,7 @@ function summarizeExecutionMeasurements(source) {
 		tokenUsage: {
 			input: input.value ?? aggregate?.tokenUsage.input ?? 0,
 			output: Math.max(output.value ?? aggregate?.tokenUsage.output ?? 0, reasoningValue ?? 0),
+			...tokensUnknown ? { tokensKnown: false } : {},
 			...reasoningValue !== void 0 ? { reasoning: reasoningValue } : {},
 			...cachedValue !== void 0 ? { cached: cachedValue } : {},
 			...cacheWriteValue !== void 0 ? { cacheWrite: cacheWriteValue } : {}
@@ -159,9 +160,11 @@ function reconcileMeasurement(calls, byId, aggregateSourceIds, keys, fallbackKey
 	const selected = /* @__PURE__ */ new Map();
 	const callIds = new Set(calls.map((call) => call.id));
 	let reportingCalls = 0;
+	let complete = calls.length > 0;
 	for (const call of calls) {
 		const primary = nearestMeasurement(call, byId, callIds, aggregateSourceIds, keys);
 		const fallback = fallbackKeys ? nearestMeasurement(call, byId, callIds, aggregateSourceIds, fallbackKeys) : void 0;
+		if (!primary) complete = false;
 		const source = primary && fallback && primary.span.id === fallback.span.id ? {
 			span: primary.span,
 			value: Math.max(primary.value, fallback.value)
@@ -178,7 +181,7 @@ function reconcileMeasurement(calls, byId, aggregateSourceIds, keys, fallbackKey
 	return {
 		...selected.size > 0 ? { value: [...selected.values()].reduce((total, value) => total + value, 0) } : {},
 		reportingCalls,
-		complete: calls.length > 0 && reportingCalls === calls.length
+		complete: complete && reportingCalls === calls.length
 	};
 }
 function nearestMeasurement(call, byId, callIds, aggregateSourceIds, keys) {
@@ -287,8 +290,7 @@ function readNumber(attributes, keys) {
 		if (Number.isFinite(parsed) && parsed >= 0) return parsed;
 	}
 }
-//#endregion
-//#region src/trace/task-failure-attributes.ts
+// src/trace/task-failure-attributes.ts
 const TASK_FAILURE_CLASS_ATTR = "tangle.task.failure_class";
 const TASK_FAILURE_MODE_ATTR = "tangle.task.failure_mode";
 function readTaskFailureLabels(roots, context) {
@@ -314,7 +316,6 @@ function readConsistentRootString(roots, key, context) {
 	if (values.size > 1) throw new ValidationError(`${context}: conflicting ${key} values: ${[...values].sort().join(", ")}`);
 	return values.values().next().value;
 }
-//#endregion
 export { summarizeTraceErrors as i, recordAggregateMeasurements as n, summarizeExecutionMeasurements as r, readTaskFailureLabels as t };
 
-//# sourceMappingURL=task-failure-attributes-BMsblaDh.js.map
+//# sourceMappingURL=task-failure-attributes-CC6VuIpk.js.map

package/dist/task-failure-attributes-CC6VuIpk.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"task-failure-attributes-CC6VuIpk.js","names":[],"sources":["../src/trace/error-classification.ts","../src/trace/execution-measurements.ts","../src/trace/task-failure-attributes.ts"],"sourcesContent":["import type { SpanKind } from '@tangle-network/agent-trace-contract'\n\nexport type TraceErrorRole = SpanKind\n\ninterface TraceErrorSignal {\n  id: string\n  parentId?: string\n  role: TraceErrorRole\n  error: boolean\n  processRoot: boolean\n}\n\ninterface TraceErrorSummary {\n  total: number\n  execution: number\n  process: number\n  guardrail: number\n  evaluation: number\n  propagated: number\n  unclassified: number\n}\n\n/**\n * Classify errored spans without counting a propagated parent status as a\n * second execution failure.\n */\nexport function summarizeTraceErrors(signals: readonly TraceErrorSignal[]): TraceErrorSummary {\n  const byId = new Map<string, TraceErrorSignal>()\n  for (const signal of signals) {\n    if (byId.has(signal.id)) {\n      throw new Error(`summarizeTraceErrors: duplicate span id '${signal.id}'`)\n    }\n    byId.set(signal.id, signal)\n  }\n\n  const propagated = new Set<string>()\n  for (const signal of signals) {\n    if (!signal.error) continue\n    const visited = new Set<string>()\n    let parentId = signal.parentId\n    while (parentId && !visited.has(parentId)) {\n      visited.add(parentId)\n      const parent = byId.get(parentId)\n      if (!parent) break\n      if (parent.error) propagated.add(parent.id)\n      parentId = parent.parentId\n    }\n  }\n\n  const summary: TraceErrorSummary = {\n    total: 0,\n    execution: 0,\n    process: 0,\n    guardrail: 0,\n    evaluation: 0,\n    propagated: 0,\n    unclassified: 0,\n  }\n\n  for (const signal of signals) {\n    if (!signal.error) continue\n    summary.total += 1\n    if (signal.role === 'GUARDRAIL') {\n      summary.guardrail += 1\n    } else if (signal.role === 'EVALUATOR') {\n      summary.evaluation += 1\n    } else if (signal.processRoot) {\n      summary.process += 1\n    } else if (propagated.has(signal.id)) {\n      summary.propagated += 1\n    } else if (signal.role !== 'UNKNOWN') {\n      summary.execution += 1\n    } else {\n      summary.unclassified += 1\n    }\n  }\n\n  return summary\n}\n","import { isModelCallKind, type SpanKind } from '@tangle-network/agent-trace-contract'\nimport type { RunTokenUsage } from '../run-record'\nimport {\n  LLM_CACHE_WRITE_TOKEN_ATTR_KEYS,\n  LLM_CACHED_TOKEN_ATTR_KEYS,\n  LLM_COST_ATTR_KEYS,\n  LLM_INPUT_TOKEN_ATTR_KEYS,\n  LLM_OUTPUT_TOKEN_ATTR_KEYS,\n  LLM_REASONING_TOKEN_ATTR_KEYS,\n  RUN_COST_ATTR_KEYS,\n} from './otlp-attributes'\n\ninterface ExecutionMeasurementSpan {\n  id: string\n  parentId?: string\n  attributes: Record<string, unknown>\n  /**\n   * The contract kind, declared or inferred. `LLM`, `EMBEDDING` and `RERANKER`\n   * ({@link isModelCallKind}) are each a leaf model call. Any other known kind\n   * (an agent, a workflow step, a tool) is an aggregate whose tokens total the\n   * calls beneath it. `UNKNOWN` is neither: its tokens count only when no model\n   * call beneath it reports the same field.\n   */\n  kind: SpanKind\n}\n\ninterface MeasuredSpan extends ExecutionMeasurementSpan {\n  modelCall: boolean\n  aggregate: boolean\n}\n\nexport interface MeasurementCoverage {\n  value?: number\n  reportingCalls: number\n  complete: boolean\n}\n\ninterface ExecutionMeasurements {\n  tokenUsage: RunTokenUsage\n  modelCallCount: number\n  callSpanIds: string[]\n  cost: MeasurementCoverage\n  aggregate?: {\n    tokenUsage: RunTokenUsage\n    costUsd?: number\n  }\n}\n\nconst TOKEN_MEASUREMENT_KEY_GROUPS = [\n  LLM_INPUT_TOKEN_ATTR_KEYS,\n  LLM_OUTPUT_TOKEN_ATTR_KEYS,\n  LLM_REASONING_TOKEN_ATTR_KEYS,\n  LLM_CACHED_TOKEN_ATTR_KEYS,\n  LLM_CACHE_WRITE_TOKEN_ATTR_KEYS,\n] as const\n\nconst EXECUTION_MEASUREMENT_KEY_GROUPS = [\n  ...TOKEN_MEASUREMENT_KEY_GROUPS,\n  LLM_COST_ATTR_KEYS,\n] as const\n\ninterface RetainedCallSummary {\n  callCount: number\n  measurements: Array<{\n    total: number\n    reportingCalls: number\n  }>\n}\n\n/**\n * Reconcile execution measurements across nested telemetry wrappers.\n * A measured parent is used only when a descendant call does not report the\n * same field, so aggregate wrappers neither duplicate complete child data nor\n * erase complementary parent fields.\n */\nexport function summarizeExecutionMeasurements(\n  source: readonly ExecutionMeasurementSpan[],\n): ExecutionMeasurements {\n  const spans: MeasuredSpan[] = source.map((span) => ({\n    ...span,\n    // EMBEDDING and RERANKER are leaf model calls too, billed like LLM — only a\n    // wrapper kind (AGENT/CHAIN/TOOL/RETRIEVER/GUARDRAIL/EVALUATOR/PROMPT) is an\n    // aggregate whose totals restate a descendant's.\n    modelCall: isModelCallKind(span.kind),\n    aggregate: !isModelCallKind(span.kind) && span.kind !== 'UNKNOWN',\n  }))\n  const byId = new Map<string, MeasuredSpan>()\n  for (const span of spans) {\n    if (byId.has(span.id)) {\n      throw new Error(`summarizeExecutionMeasurements: duplicate span id \"${span.id}\"`)\n    }\n    byId.set(span.id, span)\n  }\n  const tokenMeasurementKeys = TOKEN_MEASUREMENT_KEY_GROUPS.flat()\n  const candidates = spans.filter(\n    (span) =>\n      span.modelCall ||\n      (!span.aggregate && readNumber(span.attributes, tokenMeasurementKeys) !== undefined),\n  )\n  const candidateIds = new Set(candidates.map((span) => span.id))\n  const candidateChildren = new Map<string, MeasuredSpan[]>()\n  for (const candidate of candidates) {\n    const parentId = nearestCandidateParent(candidate, byId, candidateIds)\n    if (!parentId) continue\n    const children = candidateChildren.get(parentId) ?? []\n    children.push(candidate)\n    candidateChildren.set(parentId, children)\n  }\n  const aggregateIds = classifyAggregateSpans(candidates, candidateChildren)\n  const untypedRunCostIds = new Set(\n    spans\n      .filter(\n        (span) =>\n          !span.modelCall &&\n          !span.aggregate &&\n          !candidateIds.has(span.id) &&\n          readNumber(span.attributes, RUN_COST_ATTR_KEYS) !== undefined,\n      )\n      .map((span) => span.id),\n  )\n  const aggregateSourceIds = new Set(\n    spans\n      .filter(\n        (span) => span.aggregate || aggregateIds.has(span.id) || untypedRunCostIds.has(span.id),\n      )\n      .map((span) => span.id),\n  )\n\n  const calls = candidates.filter((span) => !aggregateIds.has(span.id))\n  const input = reconcileMeasurement(calls, byId, aggregateSourceIds, LLM_INPUT_TOKEN_ATTR_KEYS)\n  const reasoning = reconcileMeasurement(\n    calls,\n    byId,\n    aggregateSourceIds,\n    LLM_REASONING_TOKEN_ATTR_KEYS,\n  )\n  const output = reconcileMeasurement(\n    calls,\n    byId,\n    aggregateSourceIds,\n    LLM_OUTPUT_TOKEN_ATTR_KEYS,\n    LLM_REASONING_TOKEN_ATTR_KEYS,\n  )\n  const cached = reconcileMeasurement(calls, byId, aggregateSourceIds, LLM_CACHED_TOKEN_ATTR_KEYS)\n  const cacheWrite = reconcileMeasurement(\n    calls,\n    byId,\n    aggregateSourceIds,\n    LLM_CACHE_WRITE_TOKEN_ATTR_KEYS,\n  )\n  // A parent total can supply a subtotal without proving that an explicitly\n  // unmeasured call reported its usage.\n  const tokensUnknown =\n    calls.length > 0 &&\n    (!input.complete ||\n      !output.complete ||\n      calls.some((call) => call.attributes['tangle.supervise.tokens_known'] === false))\n  const aggregate = summarizeAggregateMeasurements(\n    spans,\n    byId,\n    new Set([...aggregateIds, ...untypedRunCostIds]),\n  )\n\n  // A trace whose only token-bearing span is an aggregate (a run or tool\n  // span that reports totals but has no llm_call events beneath it) has no\n  // model-call candidates, so `input`/`output`/etc. above are all undefined.\n  // Fall back to the aggregate measurement's own field before defaulting to\n  // 0, so \"no model call reported this\" does not read the same as \"measured\n  // zero\". When neither reports a field, it really is 0/unset.\n  const reasoningValue = reasoning.value ?? aggregate?.tokenUsage.reasoning\n  const cachedValue = cached.value ?? aggregate?.tokenUsage.cached\n  const cacheWriteValue = cacheWrite.value ?? aggregate?.tokenUsage.cacheWrite\n\n  return {\n    tokenUsage: {\n      input: input.value ?? aggregate?.tokenUsage.input ?? 0,\n      output: Math.max(output.value ?? aggregate?.tokenUsage.output ?? 0, reasoningValue ?? 0),\n      ...(tokensUnknown ? { tokensKnown: false } : {}),\n      ...(reasoningValue !== undefined ? { reasoning: reasoningValue } : {}),\n      ...(cachedValue !== undefined ? { cached: cachedValue } : {}),\n      ...(cacheWriteValue !== undefined ? { cacheWrite: cacheWriteValue } : {}),\n    },\n    modelCallCount: calls.length,\n    callSpanIds: calls.map((span) => span.id),\n    cost: reconcileMeasurement(calls, byId, aggregateSourceIds, LLM_COST_ATTR_KEYS),\n    ...(aggregate ? { aggregate } : {}),\n  }\n}\n\nexport function recordAggregateMeasurements(\n  raw: Record<string, number>,\n  aggregate: ExecutionMeasurements['aggregate'],\n): void {\n  if (!aggregate) return\n  raw.aggregate_prompt_tokens = aggregate.tokenUsage.input\n  raw.aggregate_completion_tokens = aggregate.tokenUsage.output\n  if (aggregate.tokenUsage.reasoning !== undefined)\n    raw.aggregate_reasoning_tokens = aggregate.tokenUsage.reasoning\n  if (aggregate.tokenUsage.cached !== undefined)\n    raw.aggregate_cached_tokens = aggregate.tokenUsage.cached\n  if (aggregate.tokenUsage.cacheWrite !== undefined)\n    raw.aggregate_cache_write_tokens = aggregate.tokenUsage.cacheWrite\n  if (aggregate.costUsd !== undefined) raw.aggregate_cost_usd = aggregate.costUsd\n}\n\nfunction summarizeAggregateMeasurements(\n  spans: MeasuredSpan[],\n  byId: Map<string, MeasuredSpan>,\n  aggregateIds: Set<string>,\n): ExecutionMeasurements['aggregate'] {\n  const aggregates = spans.filter((span) => span.aggregate || aggregateIds.has(span.id))\n  const input = reconcileTopLevelMeasurement(aggregates, byId, LLM_INPUT_TOKEN_ATTR_KEYS)\n  const output = reconcileTopLevelMeasurement(aggregates, byId, LLM_OUTPUT_TOKEN_ATTR_KEYS)\n  const reasoning = reconcileTopLevelMeasurement(aggregates, byId, LLM_REASONING_TOKEN_ATTR_KEYS)\n  const cached = reconcileTopLevelMeasurement(aggregates, byId, LLM_CACHED_TOKEN_ATTR_KEYS)\n  const cacheWrite = reconcileTopLevelMeasurement(aggregates, byId, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS)\n  const costUsd = reconcileTopLevelMeasurement(aggregates, byId, LLM_COST_ATTR_KEYS)\n  if (\n    input === undefined &&\n    output === undefined &&\n    reasoning === undefined &&\n    cached === undefined &&\n    cacheWrite === undefined &&\n    costUsd === undefined\n  )\n    return undefined\n  return {\n    tokenUsage: {\n      input: input ?? 0,\n      output: output ?? reasoning ?? 0,\n      ...(reasoning !== undefined ? { reasoning } : {}),\n      ...(cached !== undefined ? { cached } : {}),\n      ...(cacheWrite !== undefined ? { cacheWrite } : {}),\n    },\n    ...(costUsd !== undefined ? { costUsd } : {}),\n  }\n}\n\nfunction reconcileTopLevelMeasurement(\n  spans: MeasuredSpan[],\n  byId: Map<string, MeasuredSpan>,\n  keys: readonly string[],\n): number | undefined {\n  const selected = new Map<string, number>()\n  for (const span of spans) {\n    const value = readNumber(span.attributes, keys)\n    if (value !== undefined) selected.set(span.id, value)\n  }\n  for (const spanId of [...selected.keys()]) {\n    const span = byId.get(spanId)\n    if (!span) continue\n    if (ancestorIds(span, byId).some((ancestorId) => selected.has(ancestorId))) {\n      selected.delete(spanId)\n    }\n  }\n  return selected.size > 0\n    ? [...selected.values()].reduce((total, value) => total + value, 0)\n    : undefined\n}\n\nfunction reconcileMeasurement(\n  calls: MeasuredSpan[],\n  byId: Map<string, MeasuredSpan>,\n  aggregateSourceIds: Set<string>,\n  keys: readonly string[],\n  fallbackKeys?: readonly string[],\n): MeasurementCoverage {\n  const selected = new Map<string, number>()\n  const callIds = new Set(calls.map((call) => call.id))\n  let reportingCalls = 0\n  let complete = calls.length > 0\n\n  for (const call of calls) {\n    const primary = nearestMeasurement(call, byId, callIds, aggregateSourceIds, keys)\n    const fallback = fallbackKeys\n      ? nearestMeasurement(call, byId, callIds, aggregateSourceIds, fallbackKeys)\n      : undefined\n    // A fallback retains a known lower bound, such as reasoning tokens when\n    // total output is absent; it cannot establish complete coverage.\n    if (!primary) complete = false\n    const source =\n      primary && fallback && primary.span.id === fallback.span.id\n        ? { span: primary.span, value: Math.max(primary.value, fallback.value) }\n        : (primary ?? fallback)\n    if (!source) continue\n    reportingCalls += 1\n    selected.set(source.span.id, source.value)\n  }\n\n  for (const spanId of [...selected.keys()]) {\n    const span = byId.get(spanId)\n    if (!span) continue\n    if (\n      ancestorIds(span, byId).some(\n        (ancestorId) => selected.has(ancestorId) && !callIds.has(ancestorId),\n      )\n    ) {\n      selected.delete(spanId)\n    }\n  }\n\n  return {\n    ...(selected.size > 0\n      ? { value: [...selected.values()].reduce((total, value) => total + value, 0) }\n      : {}),\n    reportingCalls,\n    complete: complete && reportingCalls === calls.length,\n  }\n}\n\nfunction nearestMeasurement(\n  call: MeasuredSpan,\n  byId: Map<string, MeasuredSpan>,\n  callIds: Set<string>,\n  aggregateSourceIds: Set<string>,\n  keys: readonly string[],\n): { span: MeasuredSpan; value: number } | undefined {\n  let current: MeasuredSpan | undefined = call\n  const seen = new Set<string>()\n  while (current && !seen.has(current.id)) {\n    seen.add(current.id)\n    const value = readNumber(current.attributes, keys)\n    if (\n      value !== undefined &&\n      (current.id === call.id || (!callIds.has(current.id) && aggregateSourceIds.has(current.id)))\n    ) {\n      return { span: current, value }\n    }\n    current = current.parentId ? byId.get(current.parentId) : undefined\n  }\n  return undefined\n}\n\nfunction classifyAggregateSpans(\n  candidates: MeasuredSpan[],\n  childrenById: Map<string, MeasuredSpan[]>,\n): Set<string> {\n  const aggregateIds = new Set<string>()\n  const summaries = new Map<string, RetainedCallSummary>()\n  const visiting = new Set<string>()\n\n  const visit = (span: MeasuredSpan): RetainedCallSummary => {\n    const cached = summaries.get(span.id)\n    if (cached) return cached\n    if (visiting.has(span.id)) return emptyRetainedCallSummary()\n    visiting.add(span.id)\n\n    const descendants = emptyRetainedCallSummary()\n    for (const child of childrenById.get(span.id) ?? []) {\n      const childDescendants = visit(child)\n      if (!aggregateIds.has(child.id)) addRetainedCall(descendants, child)\n      addRetainedCallSummary(descendants, childDescendants)\n    }\n\n    if (\n      descendants.callCount > 0 &&\n      (!span.modelCall || hasCompatibleDescendantMeasurements(span, descendants))\n    ) {\n      aggregateIds.add(span.id)\n    }\n\n    visiting.delete(span.id)\n    summaries.set(span.id, descendants)\n    return descendants\n  }\n\n  for (const candidate of candidates) visit(candidate)\n  return aggregateIds\n}\n\nfunction emptyRetainedCallSummary(): RetainedCallSummary {\n  return {\n    callCount: 0,\n    measurements: EXECUTION_MEASUREMENT_KEY_GROUPS.map(() => ({\n      total: 0,\n      reportingCalls: 0,\n    })),\n  }\n}\n\nfunction addRetainedCall(summary: RetainedCallSummary, span: MeasuredSpan): void {\n  summary.callCount += 1\n  for (let index = 0; index < EXECUTION_MEASUREMENT_KEY_GROUPS.length; index += 1) {\n    const value = readNumber(span.attributes, EXECUTION_MEASUREMENT_KEY_GROUPS[index]!)\n    if (value === undefined) continue\n    const measurement = summary.measurements[index]!\n    measurement.total += value\n    measurement.reportingCalls += 1\n  }\n}\n\nfunction addRetainedCallSummary(target: RetainedCallSummary, source: RetainedCallSummary): void {\n  target.callCount += source.callCount\n  for (let index = 0; index < target.measurements.length; index += 1) {\n    const measurement = target.measurements[index]!\n    const sourceMeasurement = source.measurements[index]!\n    measurement.total += sourceMeasurement.total\n    measurement.reportingCalls += sourceMeasurement.reportingCalls\n  }\n}\n\nfunction hasCompatibleDescendantMeasurements(\n  span: MeasuredSpan,\n  descendants: RetainedCallSummary,\n): boolean {\n  let parentMeasurements = 0\n  let descendantMeasurements = 0\n  for (let index = 0; index < EXECUTION_MEASUREMENT_KEY_GROUPS.length; index += 1) {\n    const keys = EXECUTION_MEASUREMENT_KEY_GROUPS[index]!\n    const parentValue = readNumber(span.attributes, keys)\n    const measurement = descendants.measurements[index]!\n    if (parentValue !== undefined) parentMeasurements += 1\n    if (measurement.reportingCalls > 0) descendantMeasurements += 1\n    if (parentValue === undefined || measurement.reportingCalls === 0) continue\n    if (measurement.reportingCalls !== descendants.callCount) continue\n    if (Math.abs(parentValue - measurement.total) > 1e-12) return false\n  }\n  return parentMeasurements === 0 || descendantMeasurements > 0\n}\n\nfunction ancestorIds(span: MeasuredSpan, byId: Map<string, MeasuredSpan>): string[] {\n  const ids: string[] = []\n  const seen = new Set<string>()\n  let parentId = span.parentId\n  while (parentId && !seen.has(parentId)) {\n    ids.push(parentId)\n    seen.add(parentId)\n    parentId = byId.get(parentId)?.parentId\n  }\n  return ids\n}\n\nfunction nearestCandidateParent(\n  span: MeasuredSpan,\n  byId: Map<string, MeasuredSpan>,\n  candidateIds: Set<string>,\n): string | undefined {\n  const seen = new Set<string>()\n  let parentId = span.parentId\n  while (parentId && !seen.has(parentId)) {\n    if (candidateIds.has(parentId)) return parentId\n    seen.add(parentId)\n    parentId = byId.get(parentId)?.parentId\n  }\n  return undefined\n}\n\nfunction readNumber(\n  attributes: Record<string, unknown>,\n  keys: readonly string[],\n): number | undefined {\n  for (const key of keys) {\n    const value = attributes[key]\n    const parsed =\n      typeof value === 'number'\n        ? value\n        : typeof value === 'string' && value.length > 0\n          ? Number(value)\n          : Number.NaN\n    if (Number.isFinite(parsed) && parsed >= 0) return parsed\n  }\n  return undefined\n}\n","import { ValidationError } from '../errors'\nimport { FAILURE_CLASSES, type FailureClass } from './schema'\n\nconst TASK_FAILURE_CLASS_ATTR = 'tangle.task.failure_class'\nconst TASK_FAILURE_MODE_ATTR = 'tangle.task.failure_mode'\n\ninterface AttributeCarrier {\n  attributes: Record<string, unknown>\n}\n\ntype TaskFailureLabels =\n  | { failureClass?: undefined; failureMode?: undefined }\n  | { failureClass: 'success'; failureMode?: undefined }\n  | { failureClass: Exclude<FailureClass, 'success'>; failureMode?: string }\n\nexport function readTaskFailureLabels(\n  roots: readonly AttributeCarrier[],\n  context: string,\n): TaskFailureLabels {\n  const failureClass = readConsistentRootString(roots, TASK_FAILURE_CLASS_ATTR, context)\n  const failureMode = readConsistentRootString(roots, TASK_FAILURE_MODE_ATTR, context)\n\n  if (failureClass !== undefined && !FAILURE_CLASSES.includes(failureClass as FailureClass)) {\n    throw new ValidationError(\n      `${context}: ${TASK_FAILURE_CLASS_ATTR} must be one of ${FAILURE_CLASSES.join(', ')}`,\n    )\n  }\n  if (failureMode !== undefined && (failureClass === undefined || failureClass === 'success')) {\n    throw new ValidationError(\n      `${context}: ${TASK_FAILURE_MODE_ATTR} requires a non-success ${TASK_FAILURE_CLASS_ATTR}`,\n    )\n  }\n\n  if (failureClass === undefined) return {}\n  if (failureClass === 'success') return { failureClass }\n  return {\n    failureClass: failureClass as Exclude<FailureClass, 'success'>,\n    ...(failureMode ? { failureMode } : {}),\n  }\n}\n\nfunction readConsistentRootString(\n  roots: readonly AttributeCarrier[],\n  key: string,\n  context: string,\n): string | undefined {\n  const values = new Set<string>()\n  for (const root of roots) {\n    if (!Object.hasOwn(root.attributes, key)) continue\n    const value = root.attributes[key]\n    if (typeof value !== 'string' || value.trim().length === 0) {\n      throw new ValidationError(`${context}: ${key} must be a non-empty string`)\n    }\n    values.add(value)\n  }\n\n  if (values.size > 1) {\n    throw new ValidationError(\n      `${context}: conflicting ${key} values: ${[...values].sort().join(', ')}`,\n    )\n  }\n  return values.values().next().value\n}\n"],"mappings":";;;;;;;;;;AA0BA,SAAgB,qBAAqB,SAAyD;CAC5F,MAAM,uBAAO,IAAI,IAA8B;CAC/C,KAAK,MAAM,UAAU,SAAS;EAC5B,IAAI,KAAK,IAAI,OAAO,EAAE,GACpB,MAAM,IAAI,MAAM,4CAA4C,OAAO,GAAG,EAAE;EAE1E,KAAK,IAAI,OAAO,IAAI,MAAM;CAC5B;CAEA,MAAM,6BAAa,IAAI,IAAY;CACnC,KAAK,MAAM,UAAU,SAAS;EAC5B,IAAI,CAAC,OAAO,OAAO;EACnB,MAAM,0BAAU,IAAI,IAAY;EAChC,IAAI,WAAW,OAAO;EACtB,OAAO,YAAY,CAAC,QAAQ,IAAI,QAAQ,GAAG;GACzC,QAAQ,IAAI,QAAQ;GACpB,MAAM,SAAS,KAAK,IAAI,QAAQ;GAChC,IAAI,CAAC,QAAQ;GACb,IAAI,OAAO,OAAO,WAAW,IAAI,OAAO,EAAE;GAC1C,WAAW,OAAO;EACpB;CACF;CAEA,MAAM,UAA6B;EACjC,OAAO;EACP,WAAW;EACX,SAAS;EACT,WAAW;EACX,YAAY;EACZ,YAAY;EACZ,cAAc;CAChB;CAEA,KAAK,MAAM,UAAU,SAAS;EAC5B,IAAI,CAAC,OAAO,OAAO;EACnB,QAAQ,SAAS;EACjB,IAAI,OAAO,SAAS,aAClB,QAAQ,aAAa;OAChB,IAAI,OAAO,SAAS,aACzB,QAAQ,cAAc;OACjB,IAAI,OAAO,aAChB,QAAQ,WAAW;OACd,IAAI,WAAW,IAAI,OAAO,EAAE,GACjC,QAAQ,cAAc;OACjB,IAAI,OAAO,SAAS,WACzB,QAAQ,aAAa;OAErB,QAAQ,gBAAgB;CAE5B;CAEA,OAAO;AACT;;AC9BA,MAAM,+BAA+B;CACnC;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,mCAAmC,CACvC,GAAG,8BACH,kBACF;;;;;;;AAgBA,SAAgB,+BACd,QACuB;CACvB,MAAM,QAAwB,OAAO,KAAK,UAAU;EAClD,GAAG;EAIH,WAAW,gBAAgB,KAAK,IAAI;EACpC,WAAW,CAAC,gBAAgB,KAAK,IAAI,KAAK,KAAK,SAAS;CAC1D,EAAE;CACF,MAAM,uBAAO,IAAI,IAA0B;CAC3C,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,KAAK,IAAI,KAAK,EAAE,GAClB,MAAM,IAAI,MAAM,sDAAsD,KAAK,GAAG,EAAE;EAElF,KAAK,IAAI,KAAK,IAAI,IAAI;CACxB;CACA,MAAM,uBAAuB,6BAA6B,KAAK;CAC/D,MAAM,aAAa,MAAM,QACtB,SACC,KAAK,aACJ,CAAC,KAAK,aAAa,WAAW,KAAK,YAAY,oBAAoB,MAAM,KAAA,CAC9E;CACA,MAAM,eAAe,IAAI,IAAI,WAAW,KAAK,SAAS,KAAK,EAAE,CAAC;CAC9D,MAAM,oCAAoB,IAAI,IAA4B;CAC1D,KAAK,MAAM,aAAa,YAAY;EAClC,MAAM,WAAW,uBAAuB,WAAW,MAAM,YAAY;EACrE,IAAI,CAAC,UAAU;EACf,MAAM,WAAW,kBAAkB,IAAI,QAAQ,KAAK,CAAC;EACrD,SAAS,KAAK,SAAS;EACvB,kBAAkB,IAAI,UAAU,QAAQ;CAC1C;CACA,MAAM,eAAe,uBAAuB,YAAY,iBAAiB;CACzE,MAAM,oBAAoB,IAAI,IAC5B,MACG,QACE,SACC,CAAC,KAAK,aACN,CAAC,KAAK,aACN,CAAC,aAAa,IAAI,KAAK,EAAE,KACzB,WAAW,KAAK,YAAY,kBAAkB,MAAM,KAAA,CACxD,CAAC,CACA,KAAK,SAAS,KAAK,EAAE,CAC1B;CACA,MAAM,qBAAqB,IAAI,IAC7B,MACG,QACE,SAAS,KAAK,aAAa,aAAa,IAAI,KAAK,EAAE,KAAK,kBAAkB,IAAI,KAAK,EAAE,CACxF,CAAC,CACA,KAAK,SAAS,KAAK,EAAE,CAC1B;CAEA,MAAM,QAAQ,WAAW,QAAQ,SAAS,CAAC,aAAa,IAAI,KAAK,EAAE,CAAC;CACpE,MAAM,QAAQ,qBAAqB,OAAO,MAAM,oBAAoB,yBAAyB;CAC7F,MAAM,YAAY,qBAChB,OACA,MACA,oBACA,6BACF;CACA,MAAM,SAAS,qBACb,OACA,MACA,oBACA,4BACA,6BACF;CACA,MAAM,SAAS,qBAAqB,OAAO,MAAM,oBAAoB,0BAA0B;CAC/F,MAAM,aAAa,qBACjB,OACA,MACA,oBACA,+BACF;CAGA,MAAM,gBACJ,MAAM,SAAS,MACd,CAAC,MAAM,YACN,CAAC,OAAO,YACR,MAAM,MAAM,SAAS,KAAK,WAAW,qCAAqC,KAAK;CACnF,MAAM,YAAY,+BAChB,OACA,sBACA,IAAI,IAAI,CAAC,GAAG,cAAc,GAAG,iBAAiB,CAAC,CACjD;CAQA,MAAM,iBAAiB,UAAU,SAAS,WAAW,WAAW;CAChE,MAAM,cAAc,OAAO,SAAS,WAAW,WAAW;CAC1D,MAAM,kBAAkB,WAAW,SAAS,WAAW,WAAW;CAElE,OAAO;EACL,YAAY;GACV,OAAO,MAAM,SAAS,WAAW,WAAW,SAAS;GACrD,QAAQ,KAAK,IAAI,OAAO,SAAS,WAAW,WAAW,UAAU,GAAG,kBAAkB,CAAC;GACvF,GAAI,gBAAgB,EAAE,aAAa,MAAM,IAAI,CAAC;GAC9C,GAAI,mBAAmB,KAAA,IAAY,EAAE,WAAW,eAAe,IAAI,CAAC;GACpE,GAAI,gBAAgB,KAAA,IAAY,EAAE,QAAQ,YAAY,IAAI,CAAC;GAC3D,GAAI,oBAAoB,KAAA,IAAY,EAAE,YAAY,gBAAgB,IAAI,CAAC;EACzE;EACA,gBAAgB,MAAM;EACtB,aAAa,MAAM,KAAK,SAAS,KAAK,EAAE;EACxC,MAAM,qBAAqB,OAAO,MAAM,oBAAoB,kBAAkB;EAC9E,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC;CACnC;AACF;AAEA,SAAgB,4BACd,KACA,WACM;CACN,IAAI,CAAC,WAAW;CAChB,IAAI,0BAA0B,UAAU,WAAW;CACnD,IAAI,8BAA8B,UAAU,WAAW;CACvD,IAAI,UAAU,WAAW,cAAc,KAAA,GACrC,IAAI,6BAA6B,UAAU,WAAW;CACxD,IAAI,UAAU,WAAW,WAAW,KAAA,GAClC,IAAI,0BAA0B,UAAU,WAAW;CACrD,IAAI,UAAU,WAAW,eAAe,KAAA,GACtC,IAAI,+BAA+B,UAAU,WAAW;CAC1D,IAAI,UAAU,YAAY,KAAA,GAAW,IAAI,qBAAqB,UAAU;AAC1E;AAEA,SAAS,+BACP,OACA,MACA,cACoC;CACpC,MAAM,aAAa,MAAM,QAAQ,SAAS,KAAK,aAAa,aAAa,IAAI,KAAK,EAAE,CAAC;CACrF,MAAM,QAAQ,6BAA6B,YAAY,MAAM,yBAAyB;CACtF,MAAM,SAAS,6BAA6B,YAAY,MAAM,0BAA0B;CACxF,MAAM,YAAY,6BAA6B,YAAY,MAAM,6BAA6B;CAC9F,MAAM,SAAS,6BAA6B,YAAY,MAAM,0BAA0B;CACxF,MAAM,aAAa,6BAA6B,YAAY,MAAM,+BAA+B;CACjG,MAAM,UAAU,6BAA6B,YAAY,MAAM,kBAAkB;CACjF,IACE,UAAU,KAAA,KACV,WAAW,KAAA,KACX,cAAc,KAAA,KACd,WAAW,KAAA,KACX,eAAe,KAAA,KACf,YAAY,KAAA,GAEZ,OAAO,KAAA;CACT,OAAO;EACL,YAAY;GACV,OAAO,SAAS;GAChB,QAAQ,UAAU,aAAa;GAC/B,GAAI,cAAc,KAAA,IAAY,EAAE,UAAU,IAAI,CAAC;GAC/C,GAAI,WAAW,KAAA,IAAY,EAAE,OAAO,IAAI,CAAC;GACzC,GAAI,eAAe,KAAA,IAAY,EAAE,WAAW,IAAI,CAAC;EACnD;EACA,GAAI,YAAY,KAAA,IAAY,EAAE,QAAQ,IAAI,CAAC;CAC7C;AACF;AAEA,SAAS,6BACP,OACA,MACA,MACoB;CACpB,MAAM,2BAAW,IAAI,IAAoB;CACzC,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,QAAQ,WAAW,KAAK,YAAY,IAAI;EAC9C,IAAI,UAAU,KAAA,GAAW,SAAS,IAAI,KAAK,IAAI,KAAK;CACtD;CACA,KAAK,MAAM,UAAU,CAAC,GAAG,SAAS,KAAK,CAAC,GAAG;EACzC,MAAM,OAAO,KAAK,IAAI,MAAM;EAC5B,IAAI,CAAC,MAAM;EACX,IAAI,YAAY,MAAM,IAAI,CAAC,CAAC,MAAM,eAAe,SAAS,IAAI,UAAU,CAAC,GACvE,SAAS,OAAO,MAAM;CAE1B;CACA,OAAO,SAAS,OAAO,IACnB,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAAC,QAAQ,OAAO,UAAU,QAAQ,OAAO,CAAC,IAChE,KAAA;AACN;AAEA,SAAS,qBACP,OACA,MACA,oBACA,MACA,cACqB;CACrB,MAAM,2BAAW,IAAI,IAAoB;CACzC,MAAM,UAAU,IAAI,IAAI,MAAM,KAAK,SAAS,KAAK,EAAE,CAAC;CACpD,IAAI,iBAAiB;CACrB,IAAI,WAAW,MAAM,SAAS;CAE9B,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,UAAU,mBAAmB,MAAM,MAAM,SAAS,oBAAoB,IAAI;EAChF,MAAM,WAAW,eACb,mBAAmB,MAAM,MAAM,SAAS,oBAAoB,YAAY,IACxE,KAAA;EAGJ,IAAI,CAAC,SAAS,WAAW;EACzB,MAAM,SACJ,WAAW,YAAY,QAAQ,KAAK,OAAO,SAAS,KAAK,KACrD;GAAE,MAAM,QAAQ;GAAM,OAAO,KAAK,IAAI,QAAQ,OAAO,SAAS,KAAK;EAAE,IACpE,WAAW;EAClB,IAAI,CAAC,QAAQ;EACb,kBAAkB;EAClB,SAAS,IAAI,OAAO,KAAK,IAAI,OAAO,KAAK;CAC3C;CAEA,KAAK,MAAM,UAAU,CAAC,GAAG,SAAS,KAAK,CAAC,GAAG;EACzC,MAAM,OAAO,KAAK,IAAI,MAAM;EAC5B,IAAI,CAAC,MAAM;EACX,IACE,YAAY,MAAM,IAAI,CAAC,CAAC,MACrB,eAAe,SAAS,IAAI,UAAU,KAAK,CAAC,QAAQ,IAAI,UAAU,CACrE,GAEA,SAAS,OAAO,MAAM;CAE1B;CAEA,OAAO;EACL,GAAI,SAAS,OAAO,IAChB,EAAE,OAAO,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAAC,QAAQ,OAAO,UAAU,QAAQ,OAAO,CAAC,EAAE,IAC3E,CAAC;EACL;EACA,UAAU,YAAY,mBAAmB,MAAM;CACjD;AACF;AAEA,SAAS,mBACP,MACA,MACA,SACA,oBACA,MACmD;CACnD,IAAI,UAAoC;CACxC,MAAM,uBAAO,IAAI,IAAY;CAC7B,OAAO,WAAW,CAAC,KAAK,IAAI,QAAQ,EAAE,GAAG;EACvC,KAAK,IAAI,QAAQ,EAAE;EACnB,MAAM,QAAQ,WAAW,QAAQ,YAAY,IAAI;EACjD,IACE,UAAU,KAAA,MACT,QAAQ,OAAO,KAAK,MAAO,CAAC,QAAQ,IAAI,QAAQ,EAAE,KAAK,mBAAmB,IAAI,QAAQ,EAAE,IAEzF,OAAO;GAAE,MAAM;GAAS;EAAM;EAEhC,UAAU,QAAQ,WAAW,KAAK,IAAI,QAAQ,QAAQ,IAAI,KAAA;CAC5D;AAEF;AAEA,SAAS,uBACP,YACA,cACa;CACb,MAAM,+BAAe,IAAI,IAAY;CACrC,MAAM,4BAAY,IAAI,IAAiC;CACvD,MAAM,2BAAW,IAAI,IAAY;CAEjC,MAAM,SAAS,SAA4C;EACzD,MAAM,SAAS,UAAU,IAAI,KAAK,EAAE;EACpC,IAAI,QAAQ,OAAO;EACnB,IAAI,SAAS,IAAI,KAAK,EAAE,GAAG,OAAO,yBAAyB;EAC3D,SAAS,IAAI,KAAK,EAAE;EAEpB,MAAM,cAAc,yBAAyB;EAC7C,KAAK,MAAM,SAAS,aAAa,IAAI,KAAK,EAAE,KAAK,CAAC,GAAG;GACnD,MAAM,mBAAmB,MAAM,KAAK;GACpC,IAAI,CAAC,aAAa,IAAI,MAAM,EAAE,GAAG,gBAAgB,aAAa,KAAK;GACnE,uBAAuB,aAAa,gBAAgB;EACtD;EAEA,IACE,YAAY,YAAY,MACvB,CAAC,KAAK,aAAa,oCAAoC,MAAM,WAAW,IAEzE,aAAa,IAAI,KAAK,EAAE;EAG1B,SAAS,OAAO,KAAK,EAAE;EACvB,UAAU,IAAI,KAAK,IAAI,WAAW;EAClC,OAAO;CACT;CAEA,KAAK,MAAM,aAAa,YAAY,MAAM,SAAS;CACnD,OAAO;AACT;AAEA,SAAS,2BAAgD;CACvD,OAAO;EACL,WAAW;EACX,cAAc,iCAAiC,WAAW;GACxD,OAAO;GACP,gBAAgB;EAClB,EAAE;CACJ;AACF;AAEA,SAAS,gBAAgB,SAA8B,MAA0B;CAC/E,QAAQ,aAAa;CACrB,KAAK,IAAI,QAAQ,GAAG,QAAQ,iCAAiC,QAAQ,SAAS,GAAG;EAC/E,MAAM,QAAQ,WAAW,KAAK,YAAY,iCAAiC,MAAO;EAClF,IAAI,UAAU,KAAA,GAAW;EACzB,MAAM,cAAc,QAAQ,aAAa;EACzC,YAAY,SAAS;EACrB,YAAY,kBAAkB;CAChC;AACF;AAEA,SAAS,uBAAuB,QAA6B,QAAmC;CAC9F,OAAO,aAAa,OAAO;CAC3B,KAAK,IAAI,QAAQ,GAAG,QAAQ,OAAO,aAAa,QAAQ,SAAS,GAAG;EAClE,MAAM,cAAc,OAAO,aAAa;EACxC,MAAM,oBAAoB,OAAO,aAAa;EAC9C,YAAY,SAAS,kBAAkB;EACvC,YAAY,kBAAkB,kBAAkB;CAClD;AACF;AAEA,SAAS,oCACP,MACA,aACS;CACT,IAAI,qBAAqB;CACzB,IAAI,yBAAyB;CAC7B,KAAK,IAAI,QAAQ,GAAG,QAAQ,iCAAiC,QAAQ,SAAS,GAAG;EAC/E,MAAM,OAAO,iCAAiC;EAC9C,MAAM,cAAc,WAAW,KAAK,YAAY,IAAI;EACpD,MAAM,cAAc,YAAY,aAAa;EAC7C,IAAI,gBAAgB,KAAA,GAAW,sBAAsB;EACrD,IAAI,YAAY,iBAAiB,GAAG,0BAA0B;EAC9D,IAAI,gBAAgB,KAAA,KAAa,YAAY,mBAAmB,GAAG;EACnE,IAAI,YAAY,mBAAmB,YAAY,WAAW;EAC1D,IAAI,KAAK,IAAI,cAAc,YAAY,KAAK,IAAI,OAAO,OAAO;CAChE;CACA,OAAO,uBAAuB,KAAK,yBAAyB;AAC9D;AAEA,SAAS,YAAY,MAAoB,MAA2C;CAClF,MAAM,MAAgB,CAAC;CACvB,MAAM,uBAAO,IAAI,IAAY;CAC7B,IAAI,WAAW,KAAK;CACpB,OAAO,YAAY,CAAC,KAAK,IAAI,QAAQ,GAAG;EACtC,IAAI,KAAK,QAAQ;EACjB,KAAK,IAAI,QAAQ;EACjB,WAAW,KAAK,IAAI,QAAQ,CAAC,EAAE;CACjC;CACA,OAAO;AACT;AAEA,SAAS,uBACP,MACA,MACA,cACoB;CACpB,MAAM,uBAAO,IAAI,IAAY;CAC7B,IAAI,WAAW,KAAK;CACpB,OAAO,YAAY,CAAC,KAAK,IAAI,QAAQ,GAAG;EACtC,IAAI,aAAa,IAAI,QAAQ,GAAG,OAAO;EACvC,KAAK,IAAI,QAAQ;EACjB,WAAW,KAAK,IAAI,QAAQ,CAAC,EAAE;CACjC;AAEF;AAEA,SAAS,WACP,YACA,MACoB;CACpB,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,WAAW;EACzB,MAAM,SACJ,OAAO,UAAU,WACb,QACA,OAAO,UAAU,YAAY,MAAM,SAAS,IAC1C,OAAO,KAAK,IACZ;EACR,IAAI,OAAO,SAAS,MAAM,KAAK,UAAU,GAAG,OAAO;CACrD;AAEF;;AC3cA,MAAM,0BAA0B;AAChC,MAAM,yBAAyB;AAW/B,SAAgB,sBACd,OACA,SACmB;CACnB,MAAM,eAAe,yBAAyB,OAAO,yBAAyB,OAAO;CACrF,MAAM,cAAc,yBAAyB,OAAO,wBAAwB,OAAO;CAEnF,IAAI,iBAAiB,KAAA,KAAa,CAAC,gBAAgB,SAAS,YAA4B,GACtF,MAAM,IAAI,gBACR,GAAG,QAAQ,IAAI,wBAAwB,kBAAkB,gBAAgB,KAAK,IAAI,GACpF;CAEF,IAAI,gBAAgB,KAAA,MAAc,iBAAiB,KAAA,KAAa,iBAAiB,YAC/E,MAAM,IAAI,gBACR,GAAG,QAAQ,IAAI,uBAAuB,0BAA0B,yBAClE;CAGF,IAAI,iBAAiB,KAAA,GAAW,OAAO,CAAC;CACxC,IAAI,iBAAiB,WAAW,OAAO,EAAE,aAAa;CACtD,OAAO;EACS;EACd,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;CACvC;AACF;AAEA,SAAS,yBACP,OACA,KACA,SACoB;CACpB,MAAM,yBAAS,IAAI,IAAY;CAC/B,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,CAAC,OAAO,OAAO,KAAK,YAAY,GAAG,GAAG;EAC1C,MAAM,QAAQ,KAAK,WAAW;EAC9B,IAAI,OAAO,UAAU,YAAY,MAAM,KAAK,CAAC,CAAC,WAAW,GACvD,MAAM,IAAI,gBAAgB,GAAG,QAAQ,IAAI,IAAI,4BAA4B;EAE3E,OAAO,IAAI,KAAK;CAClB;CAEA,IAAI,OAAO,OAAO,GAChB,MAAM,IAAI,gBACR,GAAG,QAAQ,gBAAgB,IAAI,WAAW,CAAC,GAAG,MAAM,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GACxE;CAEF,OAAO,OAAO,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC;AAChC"}

package/dist/{tool-groups-b3oU_D88.d.ts → tool-groups-C3K-d3gd.d.ts}

@@ -1,6 +1,5 @@
-import { J as TraceAnalysisStore } from "./types-6Lfdlshp.js";
-import { p as TraceAnalysisToolDescriptor } from "./engine-dtRVeuQs.js";
-//#region src/analyst/tool-groups.d.ts
+import { J as TraceAnalysisStore } from "./types-DzCO-S1m.js";
+import { p as TraceAnalysisToolDescriptor } from "./engine-CaHaCc3W.js";
 /** Named tool sets. Kinds pass `tools: TRACE_TOOL_GROUPS.failureForensics` etc. */
 type TraceToolGroupName =
 /** All available tools. Use for open-ended discovery kinds. */
@@ -23,6 +22,5 @@ type TraceToolGroupName =
  * silently returning all tools would defeat the cost-control point.
  */
 declare function buildTraceToolsForGroup(group: TraceToolGroupName, store: TraceAnalysisStore): TraceAnalysisToolDescriptor[];
-//#endregion
 export { buildTraceToolsForGroup as n, TraceToolGroupName as t };
-//# sourceMappingURL=tool-groups-b3oU_D88.d.ts.map
+//# sourceMappingURL=tool-groups-C3K-d3gd.d.ts.map

package/dist/tool-groups-C3K-d3gd.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"tool-groups-C3K-d3gd.d.ts","names":[],"sources":["../src/analyst/tool-groups.ts"],"mappings":";;;KAmBY;;;;;;;;;;;;;;;;;;;;iBAyDI,wBACd,OAAO,oBACP,OAAO,qBACN"}

package/dist/{tool-waste-CkjzWmeY.d.ts → tool-waste-BFG1vwTV.d.ts}

@@ -1,7 +1,6 @@
-import { S as TraceEvent, _ as Span, r as BudgetSpec, x as ToolSpan } from "./schema-D6EVKzTB.js";
-import { s as TraceStore } from "./store-B3nWRy3B.js";
-import { a as llmSpans } from "./query-Cmv_U8vf.js";
-//#region src/baseline.d.ts
+import { S as TraceEvent, _ as Span, r as BudgetSpec, x as ToolSpan } from "./schema-CkIvnmLU.js";
+import { s as TraceStore } from "./store-CMs5wGXP.js";
+import { a as llmSpans } from "./query-UD0B08Sb.js";
 interface BaselineOptions {
   /** Effect size threshold for meaningful delta (default 0.5 — medium effect). */
   effectThreshold?: number;
@@ -52,8 +51,6 @@ declare function iqr(xs: number[]): number;
  * `NaN` inferential fields, and no interval rather than fabricated certainty.
  */
 declare function welchsTTest(a: number[], b: number[]): WelchTestResult;
-//#endregion
-//#region src/tool-use-metrics.d.ts
 interface ToolUseMetrics {
   runId: string;
   totalCalls: number;
@@ -80,8 +77,6 @@ interface ToolUseOptions {
   selectionLabels?: Record<string, boolean>;
 }
 declare function computeToolUseMetrics(store: TraceStore, runId: string, options?: ToolUseOptions): Promise<ToolUseMetrics>;
-//#endregion
-//#region src/trajectory.d.ts
 interface TrajectoryStep {
   index: number;
   span: Span;
@@ -100,8 +95,6 @@ interface Trajectory {
   totalDurationMs: number;
 }
 declare function buildTrajectory(store: TraceStore, runId: string): Promise<Trajectory>;
-//#endregion
-//#region src/pipelines/budget-breach.d.ts
 interface BudgetBreachFinding {
   runId: string;
   scenarioId: string;
@@ -124,8 +117,6 @@ declare function budgetBreachView(store: TraceStore, options?: {
   scenarioId?: string;
   variantId?: string;
 }): Promise<BudgetBreachReport>;
-//#endregion
-//#region src/pipelines/judge-agreement.d.ts
 interface JudgePair {
   judgeA: string;
   judgeB: string;
@@ -141,8 +132,6 @@ interface JudgeAgreementReport {
   judgeIds: string[];
 }
 declare function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport>;
-//#endregion
-//#region src/pipelines/tool-waste.d.ts
 interface ToolWasteFinding {
   runId: string;
   wastedCalls: number;
@@ -160,6 +149,5 @@ interface ToolWasteOptions {
   }) => boolean;
 }
 declare function toolWasteView(store: TraceStore, options?: ToolWasteOptions): Promise<ToolWasteReport>;
-//#endregion
 export { welchsTTest as S, computeToolUseMetrics as _, JudgeAgreementReport as a, WelchTestStatus as b, BudgetBreachFinding as c, Trajectory as d, TrajectoryStep as f, ToolUseOptions as g, ToolUseMetrics as h, toolWasteView as i, BudgetBreachReport as l, ToolStats as m, ToolWasteOptions as n, JudgePair as o, buildTrajectory as p, ToolWasteReport as r, judgeAgreementView as s, ToolWasteFinding as t, budgetBreachView as u, BaselineOptions as v, iqr as x, WelchTestResult as y };
-//# sourceMappingURL=tool-waste-CkjzWmeY.d.ts.map
+//# sourceMappingURL=tool-waste-BFG1vwTV.d.ts.map

package/dist/tool-waste-BFG1vwTV.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"tool-waste-BFG1vwTV.d.ts","names":[],"sources":["../src/baseline.ts","../src/tool-use-metrics.ts","../src/trajectory.ts","../src/pipelines/budget-breach.ts","../src/pipelines/judge-agreement.ts","../src/pipelines/tool-waste.ts"],"mappings":";;;UAsDiB;;EAEf;;EAEA;;EAEA;;KAGU;;;;;;;;;UAUF;EACR;EACA;EACA;;KAGU,mBACP;EACC;EACA;EACA;EACA;EACA;EACA;EACA;MAED;EACC;EACA;EACA;EACA;EACA;EACA;EACA;;;iBA6EU,IAAI;;;;;;;;;iBAcJ,YAAY,aAAa,cAAc;UC9KtC;EACf;EACA;;EAEA;EACA,QAAQ,eAAe;EACvB;;EAEA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe;;EAEf,kBAAkB;;iBAGE,sBACpB,OAAO,YACP,eACA,UAAS,iBACR,QAAQ;UChCM;EACf;EACA,MAAM;;EAEN;;EAEA,QAAQ;;UAGO;EACf;EACA,OAAO;EACP;EACA;EACA;EACA;EACA;;iBAGoB,gBAAgB,OAAO,YAAY,gBAAgB,QAAQ;UCtBhE;EACf;EACA;EACA;EACA,iBAAiB;EACjB;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,aAAa;EACb,YAAY;EACZ,WAAW;EACX;EACA;;iBAGoB,iBACpB,OAAO,YACP;EAAW;EAAqB;IAC/B,QAAQ;UCnBM;EACf;EACA;EACA;;EAEA;EACA;EACA;;UAGe;EACf,OAAO;EACP;EACA;;iBAGoB,mBAAmB,OAAO,aAAa,QAAQ;UCbpD;EACf;EACA;EACA;EACA;;UAGe;EACf,OAAO;EACP;;UAGe;EACf;EACA,eAAe,MAAM,UAAU;IAAS,KAAK,QAAQ,kBAAkB;;;iBAGnD,cACpB,OAAO,YACP,UAAS,mBACR,QAAQ"}

package/dist/{tool-waste-DLsjS_p-.js → tool-waste-CLJ46AbT.js}

@@ -1,8 +1,8 @@
-import { r as pearsonR } from "./descriptive-CVNy48rR.js";
-import { r as interRaterReliability } from "./agreement-irr-BhAqjluA.js";
-import { a as llmSpans, l as toolSpans, n as groupBy, r as hasCapturedToolArgs, t as argHash } from "./query-B_RZ36an.js";
-import { n as classifyFailure, t as DEFAULT_RULES } from "./failure-taxonomy-Bx5oWqaW.js";
-//#region src/pipelines/budget-breach.ts
+import { r as pearsonR } from "./descriptive-cPwITVfy.js";
+import { r as interRaterReliability } from "./agreement-irr-BTjDN0i6.js";
+import { a as llmSpans, l as toolSpans, n as groupBy, r as hasCapturedToolArgs, t as argHash } from "./query-CGXzO9d4.js";
+import { n as classifyFailure, t as DEFAULT_RULES } from "./failure-taxonomy-9eQfXKoW.js";
+// src/pipelines/budget-breach.ts
 async function budgetBreachView(store, options = {}) {
 	const runs = await store.listRuns({
 		scenarioId: options.scenarioId,
@@ -42,8 +42,7 @@ async function budgetBreachView(store, options = {}) {
 		breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0
 	};
 }
-//#endregion
-//#region src/pipelines/failure-cluster.ts
+// src/pipelines/failure-cluster.ts
 /**
 * FailureClusterView — groups failed runs by (failureClass, triggerTool,
 * argHash-prefix) so weekly reviews can prioritize the top-N clusters.
@@ -115,8 +114,7 @@ async function failureClusterView(store, options = {}) {
 function firstErrorMessage(spans) {
 	return spans.find((s) => s.status === "error")?.error;
 }
-//#endregion
-//#region src/pipelines/judge-agreement.ts
+// src/pipelines/judge-agreement.ts
 /**
 * JudgeAgreementView — pairwise agreement between judges across the
 * corpus, grouped by dimension.
@@ -188,8 +186,7 @@ async function judgeAgreementView(store) {
 		judgeIds
 	};
 }
-//#endregion
-//#region src/tool-use-metrics.ts
+// src/tool-use-metrics.ts
 /**
 * Tool-use metrics — derived purely from trace data.
 *
@@ -266,19 +263,7 @@ async function computeToolUseMetrics(store, runId, options = {}) {
 		selectionAccuracy
 	};
 }
-//#endregion
-//#region src/pipelines/tool-waste.ts
-/**
-* ToolWasteView — fraction of tool calls whose results weren't used
-* downstream. Without a "used" signal we fall back to structural
-* proxies: error calls, duplicate calls, and tool calls followed by
-* zero subsequent LLM spans are all considered waste.
-*
-* Consumers can pass a `usageOracle` that inspects a tool span and
-* returns true iff the tool's result appears in a later LLM message,
-* artifact, or state mutation — that's the canonical definition; the
-* default heuristic is a reasonable fallback.
-*/
+// src/pipelines/tool-waste.ts
 async function toolWasteView(store, options = {}) {
 	const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId);
 	const byRun = [];
@@ -362,7 +347,6 @@ function stringify(v) {
 		return String(v);
 	}
 }
-//#endregion
 export { budgetBreachView as a, failureClusterView as i, computeToolUseMetrics as n, judgeAgreementView as r, toolWasteView as t };
 
-//# sourceMappingURL=tool-waste-DLsjS_p-.js.map
+//# sourceMappingURL=tool-waste-CLJ46AbT.js.map

package/dist/{tool-waste-DLsjS_p-.js.map → tool-waste-CLJ46AbT.js.map}

@@ -1 +1 @@
-{"version":3,"file":"tool-waste-DLsjS_p-.js","names":[],"sources":["../src/pipelines/budget-breach.ts","../src/pipelines/failure-cluster.ts","../src/pipelines/judge-agreement.ts","../src/tool-use-metrics.ts","../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, hasCapturedToolArgs, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearsonR(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n","/**\n * Tool-use metrics — derived purely from trace data.\n *\n * No scoring assumptions: consumers supply optional ground-truth tool\n * selections per turn + optional \"information used downstream\" signals.\n * Without those, we still compute descriptive metrics (error rate,\n * retry rate, duplicate-call rate) that are useful on their own.\n */\n\nimport { argHash, groupBy, hasCapturedToolArgs, toolSpans } from './trace/query'\nimport type { Span } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface ToolUseMetrics {\n  runId: string\n  totalCalls: number\n  /** Calls whose arguments were captured and can be compared for duplication. */\n  callsWithCapturedArgs: number\n  byTool: Record<string, ToolStats>\n  errorRate: number\n  /** Ratio of captured-argument calls already seen with the same tool name and arguments. */\n  duplicateRate: number\n  /** Ratio of error calls followed by ≥1 retry on same tool. */\n  retryRate: number\n  /** Optional: of the calls agent made, fraction the evaluator marked as \"correct selection\". */\n  selectionAccuracy?: number\n}\n\nexport interface ToolStats {\n  calls: number\n  callsWithCapturedArgs: number\n  errors: number\n  avgLatencyMs: number\n  duplicates: number\n}\n\nexport interface ToolUseOptions {\n  /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */\n  selectionLabels?: Record<string, boolean>\n}\n\nexport async function computeToolUseMetrics(\n  store: TraceStore,\n  runId: string,\n  options: ToolUseOptions = {},\n): Promise<ToolUseMetrics> {\n  const tools = await toolSpans(store, runId)\n  if (tools.length === 0) {\n    return {\n      runId,\n      totalCalls: 0,\n      callsWithCapturedArgs: 0,\n      byTool: {},\n      errorRate: 0,\n      duplicateRate: 0,\n      retryRate: 0,\n    }\n  }\n\n  const byTool: Record<string, ToolStats> = {}\n  let totalErrors = 0\n  let totalDuplicates = 0\n  let callsWithCapturedArgs = 0\n  const sortedTools = [...tools].sort((a, b) => a.startedAt - b.startedAt)\n  const seenSignatures = new Set<string>()\n\n  // duplicate detection + per-tool aggregation\n  for (const t of sortedTools) {\n    byTool[t.toolName] ??= {\n      calls: 0,\n      callsWithCapturedArgs: 0,\n      errors: 0,\n      avgLatencyMs: 0,\n      duplicates: 0,\n    }\n    const stat = byTool[t.toolName]!\n    stat.calls += 1\n    if (t.status === 'error') {\n      stat.errors += 1\n      totalErrors += 1\n    }\n    if (typeof t.latencyMs === 'number') stat.avgLatencyMs += t.latencyMs\n    if (hasCapturedToolArgs(t)) {\n      callsWithCapturedArgs += 1\n      stat.callsWithCapturedArgs += 1\n      const sig = `${t.toolName}|${argHash(t.args)}`\n      if (seenSignatures.has(sig)) {\n        stat.duplicates += 1\n        totalDuplicates += 1\n      }\n      seenSignatures.add(sig)\n    }\n  }\n\n  for (const stat of Object.values(byTool)) {\n    stat.avgLatencyMs = stat.calls > 0 ? stat.avgLatencyMs / stat.calls : 0\n  }\n\n  // retry detection: per-tool chronological adjacency where error → next same-tool call\n  let retryOpportunities = 0\n  let retriesFollowed = 0\n  for (const [, arr] of groupBy(sortedTools, (t) => t.toolName)) {\n    for (let i = 0; i < arr.length; i++) {\n      if (arr[i]!.status !== 'error') continue\n      retryOpportunities += 1\n      if (arr[i + 1]) retriesFollowed += 1\n    }\n  }\n  const retryRate = retryOpportunities > 0 ? retriesFollowed / retryOpportunities : 0\n\n  let selectionAccuracy: number | undefined\n  if (options.selectionLabels) {\n    const labeled = sortedTools.filter((t) => t.spanId in options.selectionLabels!)\n    if (labeled.length > 0) {\n      selectionAccuracy =\n        labeled.filter((t) => options.selectionLabels![t.spanId]).length / labeled.length\n    }\n  }\n\n  return {\n    runId,\n    totalCalls: sortedTools.length,\n    callsWithCapturedArgs,\n    byTool,\n    errorRate: totalErrors / sortedTools.length,\n    duplicateRate: callsWithCapturedArgs > 0 ? totalDuplicates / callsWithCapturedArgs : 0,\n    retryRate,\n    selectionAccuracy,\n  }\n}\n\nexport type { Span }\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;EAChC,YAAY,QAAQ;EACpB,WAAW,QAAQ;CACrB,CAAC;CACD,MAAM,WAAkC,CAAC;CACzC,MAAM,cAAsC,CAAC;CAC7C,MAAM,aAAqC,CAAC;CAC5C,MAAM,YAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;EAC5C,KAAK,MAAM,KAAK,SAAS;GACvB,IAAI,CAAC,EAAE,UAAU;GACjB,MAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;GACzD,SAAS,KAAK;IACZ,OAAO,IAAI;IACX,YAAY,IAAI;IAChB,WAAW,IAAI;IACf,WAAW,EAAE;IACb,OAAO,EAAE;IACT,UAAU,EAAE;IACZ;IACA,WAAW,EAAE;GACf,CAAC;GACD,YAAY,EAAE,cAAc,YAAY,EAAE,cAAc,KAAK;GAC7D,WAAW,IAAI,eAAe,WAAW,IAAI,eAAe,KAAK;GACjE,IAAI,IAAI,WAAW,UAAU,IAAI,cAAc,UAAU,IAAI,cAAc,KAAK;EAClF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA;EACA;EACA;EACA,WAAW,KAAK;EAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;CACxE;AACF;;;;;;;;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;CAC/B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,UAAU,QAAQ,kBAAkB;CAC1C,MAAM,OAAO,MAAM,MAAM,SAAS;CAGlC,MAAM,2BAAW,IAAI,IAAyB;CAC9C,IAAI,gBAAgB;CAEpB,KAAK,MAAM,OAAO,MAAM;EACtB,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO;EAC/D;EACA,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;EACpD,MAAM,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;EACtD,MAAM,MAAM,gBAAgB;GAAE;GAAK;GAAO;EAAO,GAAG,KAAK;EAEzD,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,IAAI,eAAe;GACrB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,aAAa;GAC7D,IAAI,MAAM,SAAS,QAAQ;IACzB,WAAW,KAAK;IAChB,IAAI,oBAAoB,IAAI,GAAG,YAAY,QAAQ,KAAK,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GAC3E,OAAO,IAAI,MAAM,SAAS,SACxB,YAAY,KAAK;EAErB;EAEA,IAAI,CAAC,UAAU;GAEb,MAAM,WAAU,MADC,UAAU,OAAO,IAAI,KAAK,EAAA,CACxB,QAAQ,MAAM,EAAE,WAAW,OAAO,CAAC,CAAC,IAAI;GAC3D,IAAI,SAAS;IACX,WAAW,QAAQ;IACnB,IAAI,oBAAoB,OAAO,GAAG,YAAY,QAAQ,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GACjF;EACF;EAIA,IAAI,CAAC,WAAW;GACd,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;GACrF,IAAI,OAAO,SAAS,SAAS,YAAY,MAAM;EACjD;EAEA,MAAM,MAAM,GAAG,IAAI,aAAa,GAAG,YAAY,GAAG,GAAG,aAAa,GAAG,GAAG,aAAa;EACrF,IAAI,UAAU,SAAS,IAAI,GAAG;EAC9B,IAAI,CAAC,SAAS;GACZ,UAAU;IACR,cAAc,IAAI;IAClB;IACA;IACA;IACA,UAAU;IACV,aAAa,CAAC;IACd,cAAc,IAAI;IAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;GAChD;GACA,SAAS,IAAI,KAAK,OAAO;EAC3B;EACA,QAAQ;EACR,IAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,GAAG,QAAQ,YAAY,KAAK,IAAI,UAAU;CAC5F;CAMA,OAAO;EAAE,UAJG,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAC/B,QAAQ,MAAM,EAAE,YAAY,OAAO,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,WAAW,EAAE,QAEZ;EAAG;EAAe,WAAW,KAAK;CAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;CAE5D,OADgB,MAAM,MAAM,MAAM,EAAE,WAAW,OAClC,CAAC,EAAE;AAClB;;;;;;;;;;;;ACrFA,eAAsB,mBAAmB,OAAkD;CACzF,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,EAAA,CAAG,QAChD,MAAsB,EAAE,SAAS,OACpC;CACA,IAAI,IAAI,WAAW,GAAG,OAAO;EAAE,OAAO,CAAC;EAAG,YAAY,CAAC;EAAG,UAAU,CAAC;CAAE;CAEvE,MAAM,8BAAc,IAAI,IAAyB;CACjD,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;EAC7C,IAAI,KAAK,CAAC;EACV,YAAY,IAAI,EAAE,WAAW,GAAG;CAClC;CAEA,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK;CAC9D,MAAM,QAAqB,CAAC;CAC5B,KAAK,MAAM,CAAC,KAAK,UAAU,aAAa;EACtC,MAAM,0BAAU,IAAI,IAAiC;EACrD,KAAK,MAAM,KAAK,OAAO;GACrB,MAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,qBAAK,IAAI,IAAoB;GAC5D,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK;GAC7B,QAAQ,IAAI,EAAE,SAAS,CAAC;EAC1B;EACA,MAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;EACrC,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KACrC,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;GAC9C,MAAM,SAAS,WAAW;GAC1B,MAAM,SAAS,WAAW;GAC1B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,SAAkC,CAAC;GACzC,KAAK,MAAM,CAAC,QAAQ,WAAW,GAAG;IAChC,MAAM,SAAS,EAAE,IAAI,MAAM;IAC3B,IAAI,WAAW,KAAA,GAAW,OAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;GACxD;GACA,IAAI,OAAO,SAAS,GAAG;GACvB,MAAM,cAAc,OAAO,KACxB,CAAC,QAAQ,YACR,CACE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,GAClE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,CACpE,CACJ;GACA,MAAM,IAAI,sBACR,YAAY,EAAE,CAAE,KAAK,GAAG,OAAO,YAAY,KAAK,SAAS,KAAK,GAAI,CAAC,CACrE;GACA,MAAM,KAAK;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,aAAa,OAAO;IACpB,SAAS,SACP,OAAO,KAAK,MAAM,EAAE,EAAE,GACtB,OAAO,KAAK,MAAM,EAAE,EAAE,CACxB;IACA,cAAc;GAChB,CAAC;EACH;CAEJ;CAEA,OAAO;EACL,OAAO,MAAM,MAAM,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;EACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;EACzC;CACF;AACF;;;;;;;;;;;ACtDA,eAAsB,sBACpB,OACA,OACA,UAA0B,CAAC,GACF;CACzB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;CAC1C,IAAI,MAAM,WAAW,GACnB,OAAO;EACL;EACA,YAAY;EACZ,uBAAuB;EACvB,QAAQ,CAAC;EACT,WAAW;EACX,eAAe;EACf,WAAW;CACb;CAGF,MAAM,SAAoC,CAAC;CAC3C,IAAI,cAAc;CAClB,IAAI,kBAAkB;CACtB,IAAI,wBAAwB;CAC5B,MAAM,cAAc,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CACvE,MAAM,iCAAiB,IAAI,IAAY;CAGvC,KAAK,MAAM,KAAK,aAAa;EAC3B,OAAO,EAAE,cAAc;GACrB,OAAO;GACP,uBAAuB;GACvB,QAAQ;GACR,cAAc;GACd,YAAY;EACd;EACA,MAAM,OAAO,OAAO,EAAE;EACtB,KAAK,SAAS;EACd,IAAI,EAAE,WAAW,SAAS;GACxB,KAAK,UAAU;GACf,eAAe;EACjB;EACA,IAAI,OAAO,EAAE,cAAc,UAAU,KAAK,gBAAgB,EAAE;EAC5D,IAAI,oBAAoB,CAAC,GAAG;GAC1B,yBAAyB;GACzB,KAAK,yBAAyB;GAC9B,MAAM,MAAM,GAAG,EAAE,SAAS,GAAG,QAAQ,EAAE,IAAI;GAC3C,IAAI,eAAe,IAAI,GAAG,GAAG;IAC3B,KAAK,cAAc;IACnB,mBAAmB;GACrB;GACA,eAAe,IAAI,GAAG;EACxB;CACF;CAEA,KAAK,MAAM,QAAQ,OAAO,OAAO,MAAM,GACrC,KAAK,eAAe,KAAK,QAAQ,IAAI,KAAK,eAAe,KAAK,QAAQ;CAIxE,IAAI,qBAAqB;CACzB,IAAI,kBAAkB;CACtB,KAAK,MAAM,GAAG,QAAQ,QAAQ,cAAc,MAAM,EAAE,QAAQ,GAC1D,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,IAAI,IAAI,EAAE,CAAE,WAAW,SAAS;EAChC,sBAAsB;EACtB,IAAI,IAAI,IAAI,IAAI,mBAAmB;CACrC;CAEF,MAAM,YAAY,qBAAqB,IAAI,kBAAkB,qBAAqB;CAElF,IAAI;CACJ,IAAI,QAAQ,iBAAiB;EAC3B,MAAM,UAAU,YAAY,QAAQ,MAAM,EAAE,UAAU,QAAQ,eAAgB;EAC9E,IAAI,QAAQ,SAAS,GACnB,oBACE,QAAQ,QAAQ,MAAM,QAAQ,gBAAiB,EAAE,OAAO,CAAC,CAAC,SAAS,QAAQ;CAEjF;CAEA,OAAO;EACL;EACA,YAAY,YAAY;EACxB;EACA;EACA,WAAW,cAAc,YAAY;EACrC,eAAe,wBAAwB,IAAI,kBAAkB,wBAAwB;EACrF;EACA;CACF;AACF;;;;;;;;;;;;;;AC/FA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,MAAM,EAAE,KAAK;CAE1F,MAAM,QAA4B,CAAC;CACnC,IAAI,aAAa;CACjB,IAAI,cAAc;CAClB,KAAK,MAAM,SAAS,MAAM;EACxB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC1C,IAAI,MAAM,WAAW,GAAG;GACtB,MAAM,KAAK;IAAE;IAAO,aAAa;IAAG,YAAY;IAAG,WAAW;GAAE,CAAC;GACjE;EACF;EAQA,MAAM,YAAY,CAAC,GAAG,MAPH,SAAS,OAAO,KAAK,CAOd,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,aAAa,UAAU,KAAK,MAAM,EAAE,SAAS;EACnD,MAAM,aAAa,gBAAgB,SAAS;EAC5C,IAAI,SAAS;EACb,KAAK,MAAM,KAAK,OAAO;GACrB,IAAI,EAAE,WAAW,SAAS;IACxB;IACA;GACF;GAEA,MAAM,SAAS,WAAW,YAAY,EAAE,SAAS;GACjD,IAAI,QAAQ,aACN;QAAA,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,GAAG;GAAA,OAC1D;IAKL,MAAM,YAAY,UAAU,EAAE,MAAM;IACpC,IAAI,cAAc,IAAI;IAGtB,IAAI,EAFa,WAAW,WAAW,GAAA,CACjB,SAAS,UAAU,MAAM,GAAG,GAAG,CAC7C,GAAG;GACb;EACF;EACA,MAAM,YAAY,SAAS,MAAM;EACjC,MAAM,KAAK;GAAE;GAAO,aAAa;GAAQ,YAAY,MAAM;GAAQ;EAAU,CAAC;EAC9E,cAAc,MAAM;EACpB,eAAe;CACjB;CACA,OAAO;EAAE;EAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa;CAAE;AAClF;;;;;;AAOA,SAAS,gBAAgB,OAA4B;CACnD,MAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;CACjD,OAAO,MAAM,UAAU;CACvB,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAAK;EAC1C,MAAM,MAAM,MAAM,EAAE,CAAE,SAAS,KAAK,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAC9C,CAAC,CAAC,KAAK,IAAI;EACX,OAAO,KAAK,GAAG,IAAI,IAAI,OAAO,IAAI;CACpC;CACA,OAAO;AACT;;AAGA,SAAS,WAAW,QAAkB,QAAwB;CAC5D,IAAI,KAAK;CACT,IAAI,KAAK,OAAO;CAChB,OAAO,KAAK,IAAI;EACd,MAAM,MAAO,KAAK,OAAQ;EAC1B,IAAI,OAAO,QAAS,QAAQ,KAAK,MAAM;OAClC,KAAK;CACZ;CACA,OAAO;AACT;AAEA,SAAS,UAAU,GAAoB;CACrC,IAAI,MAAM,QAAQ,MAAM,KAAA,GAAW,OAAO;CAC1C,IAAI,OAAO,MAAM,UAAU,OAAO;CAClC,IAAI;EACF,OAAO,KAAK,UAAU,CAAC;CACzB,QAAQ;EACN,OAAO,OAAO,CAAC;CACjB;AACF"}
+{"version":3,"file":"tool-waste-CLJ46AbT.js","names":[],"sources":["../src/pipelines/budget-breach.ts","../src/pipelines/failure-cluster.ts","../src/pipelines/judge-agreement.ts","../src/tool-use-metrics.ts","../src/pipelines/tool-waste.ts"],"sourcesContent":["/**\n * BudgetBreachView — aggregates breach events across the corpus.\n *\n * Answers: which dimensions get hit most often? Which scenarios are\n * underbudgeted? Which variants trigger the most breaches?\n */\n\nimport type { BudgetSpec } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BudgetBreachFinding {\n  runId: string\n  scenarioId: string\n  variantId?: string\n  dimension: keyof BudgetSpec\n  limit: number\n  consumed: number\n  excessRatio: number\n  timestamp: number\n}\n\nexport interface BudgetBreachReport {\n  findings: BudgetBreachFinding[]\n  byDimension: Record<string, number>\n  byScenario: Record<string, number>\n  byVariant: Record<string, number>\n  totalRuns: number\n  breachedRunRatio: number\n}\n\nexport async function budgetBreachView(\n  store: TraceStore,\n  options: { scenarioId?: string; variantId?: string } = {},\n): Promise<BudgetBreachReport> {\n  const runs = await store.listRuns({\n    scenarioId: options.scenarioId,\n    variantId: options.variantId,\n  })\n  const findings: BudgetBreachFinding[] = []\n  const byDimension: Record<string, number> = {}\n  const byScenario: Record<string, number> = {}\n  const byVariant: Record<string, number> = {}\n\n  for (const run of runs) {\n    const entries = await store.budget(run.runId)\n    for (const e of entries) {\n      if (!e.breached) continue\n      const excessRatio = e.limit > 0 ? e.consumed / e.limit : Infinity\n      findings.push({\n        runId: run.runId,\n        scenarioId: run.scenarioId,\n        variantId: run.variantId,\n        dimension: e.dimension,\n        limit: e.limit,\n        consumed: e.consumed,\n        excessRatio,\n        timestamp: e.timestamp,\n      })\n      byDimension[e.dimension] = (byDimension[e.dimension] ?? 0) + 1\n      byScenario[run.scenarioId] = (byScenario[run.scenarioId] ?? 0) + 1\n      if (run.variantId) byVariant[run.variantId] = (byVariant[run.variantId] ?? 0) + 1\n    }\n  }\n\n  const breachedRuns = new Set(findings.map((f) => f.runId))\n  return {\n    findings,\n    byDimension,\n    byScenario,\n    byVariant,\n    totalRuns: runs.length,\n    breachedRunRatio: runs.length > 0 ? breachedRuns.size / runs.length : 0,\n  }\n}\n","/**\n * FailureClusterView — groups failed runs by (failureClass, triggerTool,\n * argHash-prefix) so weekly reviews can prioritize the top-N clusters.\n *\n * Each cluster includes: N runs, scenarios affected, representative\n * error message, a proposed mitigation hint (rule → action table).\n */\n\nimport { classifyFailure, DEFAULT_RULES, type FailureRule } from '../failure-taxonomy'\nimport { argHash, hasCapturedToolArgs, toolSpans } from '../trace/query'\nimport type { FailureClass, Span } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface FailureCluster {\n  failureClass: FailureClass\n  /** Tool name when the trigger was a tool span, else undefined. */\n  toolName?: string\n  /** First 16 chars of argHash — clusters similar args. */\n  argPrefix?: string\n  /**\n   * Source dimension when the trigger was a judge span (e.g. `'format'`,\n   * `'safety'`, `'correctness'`). Lets cross-template aggregators\n   * group failures by the dimension that fired without overloading\n   * `argPrefix`. Optional — clusters without this field deserialize cleanly.\n   */\n  dimension?: string\n  runCount: number\n  scenarioIds: string[]\n  exampleError?: string\n  exampleRunId: string\n}\n\nexport interface FailureClusterReport {\n  clusters: FailureCluster[]\n  totalFailures: number\n  totalRuns: number\n}\n\nexport async function failureClusterView(\n  store: TraceStore,\n  options: { rules?: FailureRule[]; minClusterSize?: number } = {},\n): Promise<FailureClusterReport> {\n  const rules = options.rules ?? DEFAULT_RULES\n  const minSize = options.minClusterSize ?? 1\n  const runs = await store.listRuns()\n\n  type Key = string\n  const clusters = new Map<Key, FailureCluster>()\n  let totalFailures = 0\n\n  for (const run of runs) {\n    if (run.status === 'completed' && run.outcome?.pass !== false) continue\n    totalFailures++\n    const spans = await store.spans({ runId: run.runId })\n    const events = await store.events({ runId: run.runId })\n    const cls = classifyFailure({ run, spans, events }, rules)\n\n    let toolName: string | undefined\n    let argPrefix: string | undefined\n    let dimension: string | undefined\n    if (cls.triggerSpanId) {\n      const trig = spans.find((s) => s.spanId === cls.triggerSpanId)\n      if (trig?.kind === 'tool') {\n        toolName = trig.toolName\n        if (hasCapturedToolArgs(trig)) argPrefix = argHash(trig.args).slice(0, 16)\n      } else if (trig?.kind === 'judge') {\n        dimension = trig.dimension\n      }\n    }\n    // Fallback: look at the last errored tool span\n    if (!toolName) {\n      const ts = await toolSpans(store, run.runId)\n      const errored = ts.filter((t) => t.status === 'error').pop()\n      if (errored) {\n        toolName = errored.toolName\n        if (hasCapturedToolArgs(errored)) argPrefix = argHash(errored.args).slice(0, 16)\n      }\n    }\n    // Secondary signal: any judge span on the failed run carries a\n    // dimension. Useful when the rule classified by judge score but\n    // didn't surface the trigger span (or surfaced a non-judge span).\n    if (!dimension) {\n      const judge = spans.find((s) => s.kind === 'judge' && typeof s.dimension === 'string')\n      if (judge?.kind === 'judge') dimension = judge.dimension\n    }\n\n    const key = `${cls.failureClass}|${toolName ?? ''}|${argPrefix ?? ''}|${dimension ?? ''}`\n    let cluster = clusters.get(key)\n    if (!cluster) {\n      cluster = {\n        failureClass: cls.failureClass,\n        toolName,\n        argPrefix,\n        dimension,\n        runCount: 0,\n        scenarioIds: [],\n        exampleRunId: run.runId,\n        exampleError: firstErrorMessage(spans) ?? cls.reason,\n      }\n      clusters.set(key, cluster)\n    }\n    cluster.runCount++\n    if (!cluster.scenarioIds.includes(run.scenarioId)) cluster.scenarioIds.push(run.scenarioId)\n  }\n\n  const arr = [...clusters.values()]\n    .filter((c) => c.runCount >= minSize)\n    .sort((a, b) => b.runCount - a.runCount)\n\n  return { clusters: arr, totalFailures, totalRuns: runs.length }\n}\n\nfunction firstErrorMessage(spans: Span[]): string | undefined {\n  const errored = spans.find((s) => s.status === 'error')\n  return errored?.error\n}\n","/**\n * JudgeAgreementView — pairwise agreement between judges across the\n * corpus, grouped by dimension.\n *\n * Output drives two workflows:\n *   - Judge robustness audit: \"does Claude agree with GPT at κ ≥ 0.6?\"\n *   - Calibration tracking: κ vs golden human labels over time (by\n *     providing a `humanGoldenJudgeId`).\n */\n\nimport { interRaterReliability, pearsonR } from '../statistics'\nimport type { JudgeSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface JudgePair {\n  judgeA: string\n  judgeB: string\n  dimension: string\n  /** Number of (targetSpanId, dimension) tuples both judges scored. */\n  commonItems: number\n  pearson: number\n  krippendorff: number\n}\n\nexport interface JudgeAgreementReport {\n  pairs: JudgePair[]\n  dimensions: string[]\n  judgeIds: string[]\n}\n\nexport async function judgeAgreementView(store: TraceStore): Promise<JudgeAgreementReport> {\n  const all = (await store.spans({ kind: 'judge' })).filter(\n    (s): s is JudgeSpan => s.kind === 'judge',\n  )\n  if (all.length === 0) return { pairs: [], dimensions: [], judgeIds: [] }\n\n  const byDimension = new Map<string, JudgeSpan[]>()\n  for (const s of all) {\n    const arr = byDimension.get(s.dimension) ?? []\n    arr.push(s)\n    byDimension.set(s.dimension, arr)\n  }\n\n  const judgeIds = [...new Set(all.map((s) => s.judgeId))].sort()\n  const pairs: JudgePair[] = []\n  for (const [dim, spans] of byDimension) {\n    const byJudge = new Map<string, Map<string, number>>()\n    for (const s of spans) {\n      const m = byJudge.get(s.judgeId) ?? new Map<string, number>()\n      m.set(s.targetSpanId, s.score)\n      byJudge.set(s.judgeId, m)\n    }\n    const judgesHere = [...byJudge.keys()]\n    for (let i = 0; i < judgesHere.length; i++) {\n      for (let j = i + 1; j < judgesHere.length; j++) {\n        const judgeI = judgesHere[i]!\n        const judgeJ = judgesHere[j]!\n        const a = byJudge.get(judgeI)!\n        const b = byJudge.get(judgeJ)!\n        const common: Array<[number, number]> = []\n        for (const [target, scoreA] of a) {\n          const scoreB = b.get(target)\n          if (scoreB !== undefined) common.push([scoreA, scoreB])\n        }\n        if (common.length < 2) continue\n        const judgeScores = common.map(\n          ([scoreA, scoreB]) =>\n            [\n              { judgeName: judgeI, dimension: dim, score: scoreA, reasoning: '' },\n              { judgeName: judgeJ, dimension: dim, score: scoreB, reasoning: '' },\n            ] as const,\n        )\n        const k = interRaterReliability(\n          judgeScores[0]!.map((_, k2) => judgeScores.map((pair) => pair[k2]!)),\n        )\n        pairs.push({\n          judgeA: judgeI,\n          judgeB: judgeJ,\n          dimension: dim,\n          commonItems: common.length,\n          pearson: pearsonR(\n            common.map((c) => c[0]),\n            common.map((c) => c[1]),\n          ),\n          krippendorff: k,\n        })\n      }\n    }\n  }\n\n  return {\n    pairs: pairs.sort((a, b) => b.commonItems - a.commonItems),\n    dimensions: [...byDimension.keys()].sort(),\n    judgeIds,\n  }\n}\n","/**\n * Tool-use metrics — derived purely from trace data.\n *\n * No scoring assumptions: consumers supply optional ground-truth tool\n * selections per turn + optional \"information used downstream\" signals.\n * Without those, we still compute descriptive metrics (error rate,\n * retry rate, duplicate-call rate) that are useful on their own.\n */\n\nimport { argHash, groupBy, hasCapturedToolArgs, toolSpans } from './trace/query'\nimport type { Span } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface ToolUseMetrics {\n  runId: string\n  totalCalls: number\n  /** Calls whose arguments were captured and can be compared for duplication. */\n  callsWithCapturedArgs: number\n  byTool: Record<string, ToolStats>\n  errorRate: number\n  /** Ratio of captured-argument calls already seen with the same tool name and arguments. */\n  duplicateRate: number\n  /** Ratio of error calls followed by ≥1 retry on same tool. */\n  retryRate: number\n  /** Optional: of the calls agent made, fraction the evaluator marked as \"correct selection\". */\n  selectionAccuracy?: number\n}\n\nexport interface ToolStats {\n  calls: number\n  callsWithCapturedArgs: number\n  errors: number\n  avgLatencyMs: number\n  duplicates: number\n}\n\nexport interface ToolUseOptions {\n  /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */\n  selectionLabels?: Record<string, boolean>\n}\n\nexport async function computeToolUseMetrics(\n  store: TraceStore,\n  runId: string,\n  options: ToolUseOptions = {},\n): Promise<ToolUseMetrics> {\n  const tools = await toolSpans(store, runId)\n  if (tools.length === 0) {\n    return {\n      runId,\n      totalCalls: 0,\n      callsWithCapturedArgs: 0,\n      byTool: {},\n      errorRate: 0,\n      duplicateRate: 0,\n      retryRate: 0,\n    }\n  }\n\n  const byTool: Record<string, ToolStats> = {}\n  let totalErrors = 0\n  let totalDuplicates = 0\n  let callsWithCapturedArgs = 0\n  const sortedTools = [...tools].sort((a, b) => a.startedAt - b.startedAt)\n  const seenSignatures = new Set<string>()\n\n  // duplicate detection + per-tool aggregation\n  for (const t of sortedTools) {\n    byTool[t.toolName] ??= {\n      calls: 0,\n      callsWithCapturedArgs: 0,\n      errors: 0,\n      avgLatencyMs: 0,\n      duplicates: 0,\n    }\n    const stat = byTool[t.toolName]!\n    stat.calls += 1\n    if (t.status === 'error') {\n      stat.errors += 1\n      totalErrors += 1\n    }\n    if (typeof t.latencyMs === 'number') stat.avgLatencyMs += t.latencyMs\n    if (hasCapturedToolArgs(t)) {\n      callsWithCapturedArgs += 1\n      stat.callsWithCapturedArgs += 1\n      const sig = `${t.toolName}|${argHash(t.args)}`\n      if (seenSignatures.has(sig)) {\n        stat.duplicates += 1\n        totalDuplicates += 1\n      }\n      seenSignatures.add(sig)\n    }\n  }\n\n  for (const stat of Object.values(byTool)) {\n    stat.avgLatencyMs = stat.calls > 0 ? stat.avgLatencyMs / stat.calls : 0\n  }\n\n  // retry detection: per-tool chronological adjacency where error → next same-tool call\n  let retryOpportunities = 0\n  let retriesFollowed = 0\n  for (const [, arr] of groupBy(sortedTools, (t) => t.toolName)) {\n    for (let i = 0; i < arr.length; i++) {\n      if (arr[i]!.status !== 'error') continue\n      retryOpportunities += 1\n      if (arr[i + 1]) retriesFollowed += 1\n    }\n  }\n  const retryRate = retryOpportunities > 0 ? retriesFollowed / retryOpportunities : 0\n\n  let selectionAccuracy: number | undefined\n  if (options.selectionLabels) {\n    const labeled = sortedTools.filter((t) => t.spanId in options.selectionLabels!)\n    if (labeled.length > 0) {\n      selectionAccuracy =\n        labeled.filter((t) => options.selectionLabels![t.spanId]).length / labeled.length\n    }\n  }\n\n  return {\n    runId,\n    totalCalls: sortedTools.length,\n    callsWithCapturedArgs,\n    byTool,\n    errorRate: totalErrors / sortedTools.length,\n    duplicateRate: callsWithCapturedArgs > 0 ? totalDuplicates / callsWithCapturedArgs : 0,\n    retryRate,\n    selectionAccuracy,\n  }\n}\n\nexport type { Span }\n","/**\n * ToolWasteView — fraction of tool calls whose results weren't used\n * downstream. Without a \"used\" signal we fall back to structural\n * proxies: error calls, duplicate calls, and tool calls followed by\n * zero subsequent LLM spans are all considered waste.\n *\n * Consumers can pass a `usageOracle` that inspects a tool span and\n * returns true iff the tool's result appears in a later LLM message,\n * artifact, or state mutation — that's the canonical definition; the\n * default heuristic is a reasonable fallback.\n */\n\nimport { computeToolUseMetrics } from '../tool-use-metrics'\nimport { llmSpans, toolSpans } from '../trace/query'\nimport type { LlmSpan, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ToolWasteFinding {\n  runId: string\n  wastedCalls: number\n  totalCalls: number\n  wasteRate: number\n}\n\nexport interface ToolWasteReport {\n  byRun: ToolWasteFinding[]\n  overallWasteRate: number\n}\n\nexport interface ToolWasteOptions {\n  runId?: string\n  usageOracle?: (tool: ToolSpan, later: { llm: Awaited<ReturnType<typeof llmSpans>> }) => boolean\n}\n\nexport async function toolWasteView(\n  store: TraceStore,\n  options: ToolWasteOptions = {},\n): Promise<ToolWasteReport> {\n  const runs = options.runId ? [options.runId] : (await store.listRuns()).map((r) => r.runId)\n\n  const byRun: ToolWasteFinding[] = []\n  let totalCalls = 0\n  let totalWasted = 0\n  for (const runId of runs) {\n    const tools = await toolSpans(store, runId)\n    if (tools.length === 0) {\n      byRun.push({ runId, wastedCalls: 0, totalCalls: 0, wasteRate: 0 })\n      continue\n    }\n    const llms = await llmSpans(store, runId)\n    // Sort LLM spans once by start time, then build a suffix index of the\n    // concatenated message text. `suffixText[i]` is the haystack of every\n    // string message content in spans[i..]. Per tool we binary-search the\n    // first span started strictly after the tool, then test that one suffix\n    // — turning the per-tool O(llms × messages × content) scan into a single\n    // O(log llms) lookup over precomputed text.\n    const sortedLlm = [...llms].sort((a, b) => a.startedAt - b.startedAt)\n    const startTimes = sortedLlm.map((l) => l.startedAt)\n    const suffixText = buildSuffixText(sortedLlm)\n    let wasted = 0\n    for (const t of tools) {\n      if (t.status === 'error') {\n        wasted++\n        continue\n      }\n      // First LLM span started strictly after this tool (upper-bound search).\n      const cutoff = upperBound(startTimes, t.startedAt)\n      if (options.usageOracle) {\n        if (!options.usageOracle(t, { llm: sortedLlm.slice(cutoff) })) wasted++\n      } else {\n        // Default heuristic: a tool whose result is NOT mentioned in any\n        // later LLM input message is likely wasted. An empty/null result has\n        // no payload to propagate downstream — there is nothing to find in a\n        // later message, so it is not evidence of waste; skip it.\n        const resultStr = stringify(t.result)\n        if (resultStr === '') continue\n        const haystack = suffixText[cutoff] ?? ''\n        const used = haystack.includes(resultStr.slice(0, 120))\n        if (!used) wasted++\n      }\n    }\n    const wasteRate = wasted / tools.length\n    byRun.push({ runId, wastedCalls: wasted, totalCalls: tools.length, wasteRate })\n    totalCalls += tools.length\n    totalWasted += wasted\n  }\n  return { byRun, overallWasteRate: totalCalls > 0 ? totalWasted / totalCalls : 0 }\n}\n\n/**\n * Build per-position suffix haystacks: result[i] is the concatenation of every\n * string message content in spans[i..end]. Built back-to-front so each entry\n * reuses the next one — O(total message text) rather than O(spans²).\n */\nfunction buildSuffixText(spans: LlmSpan[]): string[] {\n  const result = new Array<string>(spans.length + 1)\n  result[spans.length] = ''\n  for (let i = spans.length - 1; i >= 0; i--) {\n    const own = spans[i]!.messages.map((m) =>\n      typeof m.content === 'string' ? m.content : '',\n    ).join('\\n')\n    result[i] = `${own}\\n${result[i + 1]}`\n  }\n  return result\n}\n\n/** Index of the first element strictly greater than `target` in a sorted array. */\nfunction upperBound(sorted: number[], target: number): number {\n  let lo = 0\n  let hi = sorted.length\n  while (lo < hi) {\n    const mid = (lo + hi) >>> 1\n    if (sorted[mid]! <= target) lo = mid + 1\n    else hi = mid\n  }\n  return lo\n}\n\nfunction stringify(v: unknown): string {\n  if (v === null || v === undefined) return ''\n  if (typeof v === 'string') return v\n  try {\n    return JSON.stringify(v)\n  } catch {\n    return String(v)\n  }\n}\n\n// Re-export for convenience in consumers that want both descriptive and usage metrics.\nexport { computeToolUseMetrics }\n"],"mappings":";;;;;AA8BA,eAAsB,iBACpB,OACA,UAAuD,CAAC,GAC3B;CAC7B,MAAM,OAAO,MAAM,MAAM,SAAS;EAChC,YAAY,QAAQ;EACpB,WAAW,QAAQ;CACrB,CAAC;CACD,MAAM,WAAkC,CAAC;CACzC,MAAM,cAAsC,CAAC;CAC7C,MAAM,aAAqC,CAAC;CAC5C,MAAM,YAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,UAAU,MAAM,MAAM,OAAO,IAAI,KAAK;EAC5C,KAAK,MAAM,KAAK,SAAS;GACvB,IAAI,CAAC,EAAE,UAAU;GACjB,MAAM,cAAc,EAAE,QAAQ,IAAI,EAAE,WAAW,EAAE,QAAQ;GACzD,SAAS,KAAK;IACZ,OAAO,IAAI;IACX,YAAY,IAAI;IAChB,WAAW,IAAI;IACf,WAAW,EAAE;IACb,OAAO,EAAE;IACT,UAAU,EAAE;IACZ;IACA,WAAW,EAAE;GACf,CAAC;GACD,YAAY,EAAE,cAAc,YAAY,EAAE,cAAc,KAAK;GAC7D,WAAW,IAAI,eAAe,WAAW,IAAI,eAAe,KAAK;GACjE,IAAI,IAAI,WAAW,UAAU,IAAI,cAAc,UAAU,IAAI,cAAc,KAAK;EAClF;CACF;CAEA,MAAM,eAAe,IAAI,IAAI,SAAS,KAAK,MAAM,EAAE,KAAK,CAAC;CACzD,OAAO;EACL;EACA;EACA;EACA;EACA,WAAW,KAAK;EAChB,kBAAkB,KAAK,SAAS,IAAI,aAAa,OAAO,KAAK,SAAS;CACxE;AACF;;;;;;;;;ACnCA,eAAsB,mBACpB,OACA,UAA8D,CAAC,GAChC;CAC/B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,UAAU,QAAQ,kBAAkB;CAC1C,MAAM,OAAO,MAAM,MAAM,SAAS;CAGlC,MAAM,2BAAW,IAAI,IAAyB;CAC9C,IAAI,gBAAgB;CAEpB,KAAK,MAAM,OAAO,MAAM;EACtB,IAAI,IAAI,WAAW,eAAe,IAAI,SAAS,SAAS,OAAO;EAC/D;EACA,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;EACpD,MAAM,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,IAAI,MAAM,CAAC;EACtD,MAAM,MAAM,gBAAgB;GAAE;GAAK;GAAO;EAAO,GAAG,KAAK;EAEzD,IAAI;EACJ,IAAI;EACJ,IAAI;EACJ,IAAI,IAAI,eAAe;GACrB,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,WAAW,IAAI,aAAa;GAC7D,IAAI,MAAM,SAAS,QAAQ;IACzB,WAAW,KAAK;IAChB,IAAI,oBAAoB,IAAI,GAAG,YAAY,QAAQ,KAAK,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GAC3E,OAAO,IAAI,MAAM,SAAS,SACxB,YAAY,KAAK;EAErB;EAEA,IAAI,CAAC,UAAU;GAEb,MAAM,WAAU,MADC,UAAU,OAAO,IAAI,KAAK,EAAA,CACxB,QAAQ,MAAM,EAAE,WAAW,OAAO,CAAC,CAAC,IAAI;GAC3D,IAAI,SAAS;IACX,WAAW,QAAQ;IACnB,IAAI,oBAAoB,OAAO,GAAG,YAAY,QAAQ,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,EAAE;GACjF;EACF;EAIA,IAAI,CAAC,WAAW;GACd,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,SAAS,WAAW,OAAO,EAAE,cAAc,QAAQ;GACrF,IAAI,OAAO,SAAS,SAAS,YAAY,MAAM;EACjD;EAEA,MAAM,MAAM,GAAG,IAAI,aAAa,GAAG,YAAY,GAAG,GAAG,aAAa,GAAG,GAAG,aAAa;EACrF,IAAI,UAAU,SAAS,IAAI,GAAG;EAC9B,IAAI,CAAC,SAAS;GACZ,UAAU;IACR,cAAc,IAAI;IAClB;IACA;IACA;IACA,UAAU;IACV,aAAa,CAAC;IACd,cAAc,IAAI;IAClB,cAAc,kBAAkB,KAAK,KAAK,IAAI;GAChD;GACA,SAAS,IAAI,KAAK,OAAO;EAC3B;EACA,QAAQ;EACR,IAAI,CAAC,QAAQ,YAAY,SAAS,IAAI,UAAU,GAAG,QAAQ,YAAY,KAAK,IAAI,UAAU;CAC5F;CAMA,OAAO;EAAE,UAJG,CAAC,GAAG,SAAS,OAAO,CAAC,CAAC,CAC/B,QAAQ,MAAM,EAAE,YAAY,OAAO,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,WAAW,EAAE,QAEZ;EAAG;EAAe,WAAW,KAAK;CAAO;AAChE;AAEA,SAAS,kBAAkB,OAAmC;CAE5D,OADgB,MAAM,MAAM,MAAM,EAAE,WAAW,OAClC,CAAC,EAAE;AAClB;;;;;;;;;;;ACrFA,eAAsB,mBAAmB,OAAkD;CACzF,MAAM,OAAO,MAAM,MAAM,MAAM,EAAE,MAAM,QAAQ,CAAC,EAAA,CAAG,QAChD,MAAsB,EAAE,SAAS,OACpC;CACA,IAAI,IAAI,WAAW,GAAG,OAAO;EAAE,OAAO,CAAC;EAAG,YAAY,CAAC;EAAG,UAAU,CAAC;CAAE;CAEvE,MAAM,8BAAc,IAAI,IAAyB;CACjD,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,MAAM,YAAY,IAAI,EAAE,SAAS,KAAK,CAAC;EAC7C,IAAI,KAAK,CAAC;EACV,YAAY,IAAI,EAAE,WAAW,GAAG;CAClC;CAEA,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,IAAI,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK;CAC9D,MAAM,QAAqB,CAAC;CAC5B,KAAK,MAAM,CAAC,KAAK,UAAU,aAAa;EACtC,MAAM,0BAAU,IAAI,IAAiC;EACrD,KAAK,MAAM,KAAK,OAAO;GACrB,MAAM,IAAI,QAAQ,IAAI,EAAE,OAAO,qBAAK,IAAI,IAAoB;GAC5D,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK;GAC7B,QAAQ,IAAI,EAAE,SAAS,CAAC;EAC1B;EACA,MAAM,aAAa,CAAC,GAAG,QAAQ,KAAK,CAAC;EACrC,KAAK,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KACrC,KAAK,IAAI,IAAI,IAAI,GAAG,IAAI,WAAW,QAAQ,KAAK;GAC9C,MAAM,SAAS,WAAW;GAC1B,MAAM,SAAS,WAAW;GAC1B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,IAAI,QAAQ,IAAI,MAAM;GAC5B,MAAM,SAAkC,CAAC;GACzC,KAAK,MAAM,CAAC,QAAQ,WAAW,GAAG;IAChC,MAAM,SAAS,EAAE,IAAI,MAAM;IAC3B,IAAI,WAAW,KAAA,GAAW,OAAO,KAAK,CAAC,QAAQ,MAAM,CAAC;GACxD;GACA,IAAI,OAAO,SAAS,GAAG;GACvB,MAAM,cAAc,OAAO,KACxB,CAAC,QAAQ,YACR,CACE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,GAClE;IAAE,WAAW;IAAQ,WAAW;IAAK,OAAO;IAAQ,WAAW;GAAG,CACpE,CACJ;GACA,MAAM,IAAI,sBACR,YAAY,EAAE,CAAE,KAAK,GAAG,OAAO,YAAY,KAAK,SAAS,KAAK,GAAI,CAAC,CACrE;GACA,MAAM,KAAK;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,aAAa,OAAO;IACpB,SAAS,SACP,OAAO,KAAK,MAAM,EAAE,EAAE,GACtB,OAAO,KAAK,MAAM,EAAE,EAAE,CACxB;IACA,cAAc;GAChB,CAAC;EACH;CAEJ;CAEA,OAAO;EACL,OAAO,MAAM,MAAM,GAAG,MAAM,EAAE,cAAc,EAAE,WAAW;EACzD,YAAY,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK;EACzC;CACF;AACF;;;;;;;;;;ACtDA,eAAsB,sBACpB,OACA,OACA,UAA0B,CAAC,GACF;CACzB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;CAC1C,IAAI,MAAM,WAAW,GACnB,OAAO;EACL;EACA,YAAY;EACZ,uBAAuB;EACvB,QAAQ,CAAC;EACT,WAAW;EACX,eAAe;EACf,WAAW;CACb;CAGF,MAAM,SAAoC,CAAC;CAC3C,IAAI,cAAc;CAClB,IAAI,kBAAkB;CACtB,IAAI,wBAAwB;CAC5B,MAAM,cAAc,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CACvE,MAAM,iCAAiB,IAAI,IAAY;CAGvC,KAAK,MAAM,KAAK,aAAa;EAC3B,OAAO,EAAE,cAAc;GACrB,OAAO;GACP,uBAAuB;GACvB,QAAQ;GACR,cAAc;GACd,YAAY;EACd;EACA,MAAM,OAAO,OAAO,EAAE;EACtB,KAAK,SAAS;EACd,IAAI,EAAE,WAAW,SAAS;GACxB,KAAK,UAAU;GACf,eAAe;EACjB;EACA,IAAI,OAAO,EAAE,cAAc,UAAU,KAAK,gBAAgB,EAAE;EAC5D,IAAI,oBAAoB,CAAC,GAAG;GAC1B,yBAAyB;GACzB,KAAK,yBAAyB;GAC9B,MAAM,MAAM,GAAG,EAAE,SAAS,GAAG,QAAQ,EAAE,IAAI;GAC3C,IAAI,eAAe,IAAI,GAAG,GAAG;IAC3B,KAAK,cAAc;IACnB,mBAAmB;GACrB;GACA,eAAe,IAAI,GAAG;EACxB;CACF;CAEA,KAAK,MAAM,QAAQ,OAAO,OAAO,MAAM,GACrC,KAAK,eAAe,KAAK,QAAQ,IAAI,KAAK,eAAe,KAAK,QAAQ;CAIxE,IAAI,qBAAqB;CACzB,IAAI,kBAAkB;CACtB,KAAK,MAAM,GAAG,QAAQ,QAAQ,cAAc,MAAM,EAAE,QAAQ,GAC1D,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;EACnC,IAAI,IAAI,EAAE,CAAE,WAAW,SAAS;EAChC,sBAAsB;EACtB,IAAI,IAAI,IAAI,IAAI,mBAAmB;CACrC;CAEF,MAAM,YAAY,qBAAqB,IAAI,kBAAkB,qBAAqB;CAElF,IAAI;CACJ,IAAI,QAAQ,iBAAiB;EAC3B,MAAM,UAAU,YAAY,QAAQ,MAAM,EAAE,UAAU,QAAQ,eAAgB;EAC9E,IAAI,QAAQ,SAAS,GACnB,oBACE,QAAQ,QAAQ,MAAM,QAAQ,gBAAiB,EAAE,OAAO,CAAC,CAAC,SAAS,QAAQ;CAEjF;CAEA,OAAO;EACL;EACA,YAAY,YAAY;EACxB;EACA;EACA,WAAW,cAAc,YAAY;EACrC,eAAe,wBAAwB,IAAI,kBAAkB,wBAAwB;EACrF;EACA;CACF;AACF;;AC/FA,eAAsB,cACpB,OACA,UAA4B,CAAC,GACH;CAC1B,MAAM,OAAO,QAAQ,QAAQ,CAAC,QAAQ,KAAK,KAAK,MAAM,MAAM,SAAS,EAAA,CAAG,KAAK,MAAM,EAAE,KAAK;CAE1F,MAAM,QAA4B,CAAC;CACnC,IAAI,aAAa;CACjB,IAAI,cAAc;CAClB,KAAK,MAAM,SAAS,MAAM;EACxB,MAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;EAC1C,IAAI,MAAM,WAAW,GAAG;GACtB,MAAM,KAAK;IAAE;IAAO,aAAa;IAAG,YAAY;IAAG,WAAW;GAAE,CAAC;GACjE;EACF;EAQA,MAAM,YAAY,CAAC,GAAG,MAPH,SAAS,OAAO,KAAK,CAOd,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EACpE,MAAM,aAAa,UAAU,KAAK,MAAM,EAAE,SAAS;EACnD,MAAM,aAAa,gBAAgB,SAAS;EAC5C,IAAI,SAAS;EACb,KAAK,MAAM,KAAK,OAAO;GACrB,IAAI,EAAE,WAAW,SAAS;IACxB;IACA;GACF;GAEA,MAAM,SAAS,WAAW,YAAY,EAAE,SAAS;GACjD,IAAI,QAAQ,aACN;QAAA,CAAC,QAAQ,YAAY,GAAG,EAAE,KAAK,UAAU,MAAM,MAAM,EAAE,CAAC,GAAG;GAAA,OAC1D;IAKL,MAAM,YAAY,UAAU,EAAE,MAAM;IACpC,IAAI,cAAc,IAAI;IAGtB,IAAI,EAFa,WAAW,WAAW,GAAA,CACjB,SAAS,UAAU,MAAM,GAAG,GAAG,CAC7C,GAAG;GACb;EACF;EACA,MAAM,YAAY,SAAS,MAAM;EACjC,MAAM,KAAK;GAAE;GAAO,aAAa;GAAQ,YAAY,MAAM;GAAQ;EAAU,CAAC;EAC9E,cAAc,MAAM;EACpB,eAAe;CACjB;CACA,OAAO;EAAE;EAAO,kBAAkB,aAAa,IAAI,cAAc,aAAa;CAAE;AAClF;;;;;;AAOA,SAAS,gBAAgB,OAA4B;CACnD,MAAM,SAAS,IAAI,MAAc,MAAM,SAAS,CAAC;CACjD,OAAO,MAAM,UAAU;CACvB,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,KAAK,GAAG,KAAK;EAC1C,MAAM,MAAM,MAAM,EAAE,CAAE,SAAS,KAAK,MAClC,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAC9C,CAAC,CAAC,KAAK,IAAI;EACX,OAAO,KAAK,GAAG,IAAI,IAAI,OAAO,IAAI;CACpC;CACA,OAAO;AACT;;AAGA,SAAS,WAAW,QAAkB,QAAwB;CAC5D,IAAI,KAAK;CACT,IAAI,KAAK,OAAO;CAChB,OAAO,KAAK,IAAI;EACd,MAAM,MAAO,KAAK,OAAQ;EAC1B,IAAI,OAAO,QAAS,QAAQ,KAAK,MAAM;OAClC,KAAK;CACZ;CACA,OAAO;AACT;AAEA,SAAS,UAAU,GAAoB;CACrC,IAAI,MAAM,QAAQ,MAAM,KAAA,GAAW,OAAO;CAC1C,IAAI,OAAO,MAAM,UAAU,OAAO;CAClC,IAAI;EACF,OAAO,KAAK,UAAU,CAAC;CACzB,QAAQ;EACN,OAAO,OAAO,CAAC;CACjB;AACF"}

package/dist/trace-attributes.d.ts

@@ -1,2 +1,2 @@
-import { C as TOOL_LATENCY_MS, D as asNumber, E as applyLlmSpanOtlpAttributes, O as contextInputTokens, S as TOOL_ARGS_CAPTURED, T as TOOL_NAME_ATTR_KEYS, _ as LlmSpanOtlpInput, a as LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, b as RUN_COST_ATTR_KEYS, c as LLM_COST_USD, d as LLM_MODEL_ATTR_KEYS, f as LLM_MODEL_NAME, g as LLM_REASONING_TOKEN_ATTR_KEYS, h as LLM_REASONING_TOKENS, i as LLM_CACHE_WRITE_TOKENS, k as firstNumberAttr, l as LLM_INPUT_TOKENS, m as LLM_OUTPUT_TOKEN_ATTR_KEYS, n as LLM_CACHED_TOKENS, o as LLM_CONTEXT_TOKENS, p as LLM_OUTPUT_TOKENS, r as LLM_CACHED_TOKEN_ATTR_KEYS, s as LLM_COST_ATTR_KEYS, t as INPUT_VALUE, u as LLM_INPUT_TOKEN_ATTR_KEYS, v as OPENINFERENCE_SPAN_KIND, w as TOOL_NAME, x as SPAN_KIND_ATTR_KEYS, y as OUTPUT_VALUE } from "./attribute-vocabulary-CcXqHODT.js";
+import { C as TOOL_LATENCY_MS, D as asNumber, E as applyLlmSpanOtlpAttributes, O as contextInputTokens, S as TOOL_ARGS_CAPTURED, T as TOOL_NAME_ATTR_KEYS, _ as LlmSpanOtlpInput, a as LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, b as RUN_COST_ATTR_KEYS, c as LLM_COST_USD, d as LLM_MODEL_ATTR_KEYS, f as LLM_MODEL_NAME, g as LLM_REASONING_TOKEN_ATTR_KEYS, h as LLM_REASONING_TOKENS, i as LLM_CACHE_WRITE_TOKENS, k as firstNumberAttr, l as LLM_INPUT_TOKENS, m as LLM_OUTPUT_TOKEN_ATTR_KEYS, n as LLM_CACHED_TOKENS, o as LLM_CONTEXT_TOKENS, p as LLM_OUTPUT_TOKENS, r as LLM_CACHED_TOKEN_ATTR_KEYS, s as LLM_COST_ATTR_KEYS, t as INPUT_VALUE, u as LLM_INPUT_TOKEN_ATTR_KEYS, v as OPENINFERENCE_SPAN_KIND, w as TOOL_NAME, x as SPAN_KIND_ATTR_KEYS, y as OUTPUT_VALUE } from "./attribute-vocabulary-5oo8ykIx.js";
 export { INPUT_VALUE, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, LlmSpanOtlpInput, OPENINFERENCE_SPAN_KIND, OUTPUT_VALUE, RUN_COST_ATTR_KEYS, SPAN_KIND_ATTR_KEYS, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, applyLlmSpanOtlpAttributes, asNumber, contextInputTokens, firstNumberAttr };

package/dist/trace-attributes.js

@@ -1,5 +1,5 @@
 import { ATTR, CACHE_READ_TOKEN_ATTR_KEYS, CACHE_WRITE_TOKEN_ATTR_KEYS, COST_ATTR_KEYS, INPUT_TOKEN_ATTR_KEYS, MODEL_ATTR_KEYS, OUTPUT_TOKEN_ATTR_KEYS, REASONING_TOKEN_ATTR_KEYS, SPAN_KIND_ATTR_KEYS as SPAN_KIND_ATTR_KEYS$1, TOOL_NAME_ATTR_KEYS as TOOL_NAME_ATTR_KEYS$1 } from "@tangle-network/agent-trace-contract";
-//#region src/trace/attribute-vocabulary.ts
+// src/trace/attribute-vocabulary.ts
 /**
 * Canonical OpenInference-over-OTLP attribute names used at trace boundaries.
 *
@@ -74,7 +74,6 @@ function applyLlmSpanOtlpAttributes(attributes, span) {
 	if (span.costUsd !== void 0) attributes[LLM_COST_USD] = span.costUsd;
 	if (span.finishReason) attributes["llm.finish_reason"] = span.finishReason;
 }
-//#endregion
 export { INPUT_VALUE, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, OPENINFERENCE_SPAN_KIND, OUTPUT_VALUE, RUN_COST_ATTR_KEYS, SPAN_KIND_ATTR_KEYS, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, applyLlmSpanOtlpAttributes, asNumber, contextInputTokens, firstNumberAttr };
 
 //# sourceMappingURL=trace-attributes.js.map

package/dist/{trace-questions-BmT8M6UQ.d.ts → trace-questions-ej5AEJ5X.d.ts}

@@ -1,7 +1,6 @@
-import { o as CostLedgerHandle } from "./cost-ledger-C1W3ICmP.js";
-import { J as TraceAnalysisStore } from "./types-6Lfdlshp.js";
-import { i as TraceAnalystLimits, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-dtRVeuQs.js";
-//#region src/analyst/trace-questions.d.ts
+import { o as CostLedgerHandle } from "./cost-ledger-Czro9wmo.js";
+import { J as TraceAnalysisStore } from "./types-DzCO-S1m.js";
+import { i as TraceAnalystLimits, s as RawAnalystFinding, t as TraceAnalysisEngine } from "./engine-CaHaCc3W.js";
 /** One question to ask of the store. */
 interface TraceQuestionSpec {
   /** Stable id: letters, digits, dot, underscore, hyphen. */
@@ -58,6 +57,5 @@ interface TraceQuestionOutcome {
  * question list or options.
  */
 declare function askTraceQuestions(opts: AskTraceQuestionsOptions): Promise<TraceQuestionOutcome[]>;
-//#endregion
 export { askTraceQuestions as a, VerifiedTraceFinding as i, TraceQuestionOutcome as n, TraceQuestionSpec as r, AskTraceQuestionsOptions as t };
-//# sourceMappingURL=trace-questions-BmT8M6UQ.d.ts.map
+//# sourceMappingURL=trace-questions-ej5AEJ5X.d.ts.map

package/dist/trace-questions-ej5AEJ5X.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"trace-questions-ej5AEJ5X.d.ts","names":[],"sources":["../src/analyst/trace-questions.ts"],"mappings":";;;;UA6BiB;;WAEN;WACA;;WAEA;;UAGM;WACN,oBAAoB;WACpB,OAAO;WACP,QAAQ;;WAER,YAAY;;WAEZ;;WAEA;;WAEA;;WAEA;;WAEA;WACA,SAAS,QAAQ;WACjB,SAAS;WACT,OAAO,iBAAiB,SAAS;;;UAI3B;WACN,SAAS;;WAET;;WAEA;;WAEA;;UAGM;WACN;WACA;WACA;WACA;;WAEA;WACA,mBAAmB;;WAEnB;WACA;;;;;;;iBAyBW,kBACpB,MAAM,2BACL,QAAQ"}