@tangle-network/agent-eval 0.136.0 → 0.137.0

package/dist/{run-record-BIwU2wdV.js.map → run-record-vRgqWmJw.js.map}

@@ -1 +1 @@
-{"version":3,"file":"run-record-BIwU2wdV.js","names":[],"sources":["../src/run-record.ts"],"sourcesContent":["/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag. A task score is optional because execution-only records\n * must preserve missing labels instead of converting errors into zero quality.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n *   const rec = validateRunRecord(rawJson)         // throws on missing\n *   const ok  = isRunRecord(rawJson)               // boolean check\n *   const rec = parseRunRecordSafe(rawJson)        // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport type { CostProvenance } from './cost-ledger'\nimport { ValidationError } from './errors'\n// Value import of a leaf module that itself imports only this file's TYPES —\n// no runtime cycle. It keeps the raw split-score derivation spelled in exactly\n// one place (see `rollout/score-derivation-guard`).\nimport { observedScore } from './rollout/reward'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n *  combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\n/**\n * Explicit execution-lifecycle result for a run.\n *\n * This is separate from task quality (`outcome`) and failure classification.\n * Producers set it only from root-run or process evidence.\n */\nexport type RunTerminalOutcome = 'succeeded' | 'failed' | 'cancelled' | 'incomplete' | 'unknown'\n\nexport interface RunTokenUsage {\n  input: number\n  /** All generated tokens charged as output, including reasoning tokens. */\n  output: number\n  /** Reasoning-token subset of `output`, when the provider reports it. */\n  reasoning?: number\n  /** Prompt tokens served from a provider cache. */\n  cached?: number\n  /** Prompt tokens written into a provider cache. */\n  cacheWrite?: number\n}\n\n/** How a run's USD amount was obtained. */\nexport type RunCostProvenance = CostProvenance\n\nexport interface RunJudgeMetadata {\n  model: string\n  promptVersion: string\n  /** [0,1] confidence the judge declared. Constant judge confidence\n   *  across many runs is a fallback signal (see `canary.ts`). */\n  confidence: number\n  /** True if the judge degraded to a fallback path (rules-only,\n   *  prior-call cache, etc.). The canary uses this to alert. */\n  fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n  /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n  perJudge: Record<string, Record<string, number>>\n  /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n  perDimMean: Record<string, number>\n  /** Composite mean across successful judges. Mirrors the task score only\n   *  when `failedJudges` is empty. */\n  composite: number\n  /** Judges that errored or returned an unparseable verdict. Recorded\n   *  by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n   *  not inferred from missing keys in `perJudge`. */\n  failedJudges?: string[]\n  /** Free-form notes the judges emitted (joined across judges or\n   *  first-judge only — consumer's choice). */\n  notes?: string\n}\n\nexport interface RunOutcome {\n  /** Score on the search/optimization split. Optional for holdout-only and\n   *  execution-only records. */\n  searchScore?: number\n  /** Score on the held-out split. Optional for search-only and execution-only\n   *  records. When both scores are absent, the run is explicitly unlabeled. */\n  holdoutScore?: number\n  /** Bag of any other metric the run produced — judge dimensions,\n   *  pass/fail counters, latency stats, etc. Numeric only — keeps\n   *  reporters honest. */\n  raw: Record<string, number>\n  /** Per-judge / per-dim breakdown. Consumers writing ensemble\n   *  judgements populate this; substrate primitives like\n   *  `interRaterReliability` and `corpusInterRaterAgreement` accept\n   *  these records as input. Optional — single-judge or scalar-only\n   *  runs leave it unset. */\n  judgeScores?: JudgeScoresRecord\n  /** Authenticity / realness verdict — did the run build the REAL thing on the\n   *  intended infra, or fake it (see `./authenticity`)? Optional: only domains\n   *  with an authenticity config populate it. Carried in the corpus so the\n   *  flywheel / off-policy learning can optimize for real completion, not gamed\n   *  pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n   *  must not count as a real success regardless of `score`. */\n  realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n *   - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n *     model (after any steering bundle merge).\n *   - `configHash` is the sha256 of the effective run config (model,\n *     temperature, tools, judges, splits). The pair (promptHash,\n *     configHash) uniquely identifies an experiment cell.\n *\n * Model snapshot discipline:\n *   - `model` MUST encode a snapshot version. Bare aliases like\n *     `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n *     Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n */\nexport interface RunRecord {\n  /** UUID for the run. */\n  runId: string\n  /** Logical experiment grouping (a treatment vs a baseline within\n   *  the same sweep should share `experimentId`). */\n  experimentId: string\n  /** Stable identifier for the candidate (variant) being run. The\n   *  promotion gate compares two `candidateId`s on matched items. */\n  candidateId: string\n  /** RNG seed for the run. Always recorded — silent re-seeding is\n   *  the most common cause of non-reproducible numbers. */\n  seed: number\n  /** Model identifier WITH snapshot version. */\n  model: string\n  /** sha256 of the effective prompt (post-steering). */\n  promptHash: string\n  /** sha256 of the effective config. */\n  configHash: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** End-to-end wall-clock duration in milliseconds. */\n  wallMs: number\n  /** Time spent queued before execution started, if known. */\n  queueMs?: number\n  /** Total USD cost, or null when the producer could not capture one. */\n  costUsd: number | null\n  /** Whether `costUsd` came from billing data, a price calculation, or is unavailable. */\n  costProvenance: RunCostProvenance\n  /** Token usage breakdown. */\n  tokenUsage: RunTokenUsage\n  /** Root-run or process terminal result. Never inferred from a child span. */\n  terminalOutcome: RunTerminalOutcome\n  /** Root-run or process failure reason. Valid only for a failed, cancelled,\n   *  or incomplete terminal result; never populated from a child span. */\n  terminalFailureReason?: string\n  /** Judge-side metadata, if a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /** Per-split scores + raw bag. */\n  outcome: RunOutcome\n  /** Canonical task-failure class drawn from the shared\n   *  `FAILURE_CLASSES` taxonomy. Producers set it only from task-result\n   *  evidence. Execution errors belong in\n   *  `outcome.raw.execution_error_count`. */\n  failureClass?: FailureClass\n  /** Free-form task-failure detail scoped under a non-success\n   *  `failureClass`. It is invalid without that class. */\n  failureMode?: string\n  /** Which split this run was drawn from. */\n  splitTag: RunSplitTag\n  /**\n   * Stable scenario identifier the run observed or was scored against.\n   * Comparison primitives match this identity rather than input order.\n   */\n  scenarioId: string\n  /**\n   * Canonical identity for the agent profile cell that produced this row:\n   * profile artifact hash plus optional harness/model/prompt/reporting\n   * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n   * longitudinal reports by the complete source profile, not by a loose\n   * candidate label or opaque config hash.\n   */\n  agentProfile?: AgentProfileCell\n}\n\n/**\n * Canonical task-result classification.\n *\n * A producer may omit classification, record explicit success, or attach\n * domain-specific detail to a non-success class. Detail can never stand alone.\n * Execution errors belong in `outcome.raw.execution_error_count`.\n */\nexport type RunTaskFailure =\n  | { failureClass?: undefined; failureMode?: undefined }\n  | { failureClass: 'success'; failureMode?: undefined }\n  | {\n      failureClass: Exclude<FailureClass, 'success'>\n      failureMode?: string\n    }\n\n/**\n * Return task quality, preferring held-out evidence when both scores exist.\n *\n * RAW: no realness protection is applied. Built on `observedScore` rather\n * than repeating the split derivation, so only `rollout/reward.ts` reads the\n * raw fields. Anything that becomes training data must use `trainingScore` or\n * `trainingReward` instead.\n */\nexport function runTaskScore(record: RunRecord): number | undefined {\n  const score = observedScore(record)\n  return typeof score === 'number' && Number.isFinite(score) ? score : undefined\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n  'runId',\n  'experimentId',\n  'candidateId',\n  'seed',\n  'model',\n  'promptHash',\n  'configHash',\n  'commitSha',\n  'wallMs',\n  'costUsd',\n  'costProvenance',\n  'tokenUsage',\n  'terminalOutcome',\n  'outcome',\n  'splitTag',\n  'scenarioId',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\nconst TERMINAL_OUTCOMES: ReadonlyArray<RunTerminalOutcome> = [\n  'succeeded',\n  'failed',\n  'cancelled',\n  'incomplete',\n  'unknown',\n]\n\nexport class RunRecordValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n\n  for (const key of MANDATORY_TOP_LEVEL) {\n    if (!(key in obj)) {\n      throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n    }\n  }\n\n  expectString(obj.runId, 'runId')\n  expectString(obj.experimentId, 'experimentId')\n  expectString(obj.candidateId, 'candidateId')\n  expectFiniteNumber(obj.seed, 'seed')\n  expectString(obj.model, 'model')\n  expectString(obj.promptHash, 'promptHash')\n  expectString(obj.configHash, 'configHash')\n  expectString(obj.commitSha, 'commitSha')\n  expectNonNegativeNumber(obj.wallMs, 'wallMs')\n  if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n  validateCost(obj.costUsd, obj.costProvenance)\n\n  // Snapshot discipline: bare model aliases are not paper-grade.\n  if (!modelHasSnapshot(obj.model as string)) {\n    throw new RunRecordValidationError(\n      `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,\n      'model',\n    )\n  }\n\n  // Token usage.\n  const tu = obj.tokenUsage\n  if (tu === null || typeof tu !== 'object') {\n    throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n  }\n  const tuRec = tu as Record<string, unknown>\n  expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n  expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n  if (tuRec.reasoning !== undefined) {\n    expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n    if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n      throw new RunRecordValidationError(\n        'reasoning tokens must be a subset of output tokens',\n        'tokenUsage.reasoning',\n      )\n    }\n  }\n  if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n  if (tuRec.cacheWrite !== undefined) {\n    expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n  }\n\n  // Judge metadata, optional.\n  if (obj.judgeMetadata !== undefined) {\n    const jm = obj.judgeMetadata\n    if (jm === null || typeof jm !== 'object') {\n      throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n    }\n    const jmRec = jm as Record<string, unknown>\n    expectString(jmRec.model, 'judgeMetadata.model')\n    expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n    expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n    if (typeof jmRec.fallback !== 'boolean') {\n      throw new RunRecordValidationError(\n        'judgeMetadata.fallback must be boolean',\n        'judgeMetadata.fallback',\n      )\n    }\n  }\n\n  // Outcome.\n  const out = obj.outcome\n  if (out === null || typeof out !== 'object') {\n    throw new RunRecordValidationError('outcome must be an object', 'outcome')\n  }\n  const outRec = out as Record<string, unknown>\n  if (outRec.searchScore !== undefined)\n    expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n  if (outRec.holdoutScore !== undefined)\n    expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n  const raw = outRec.raw\n  if (raw === null || typeof raw !== 'object') {\n    throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n  }\n  for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n    expectFiniteNumber(v, `outcome.raw.${k}`)\n  }\n  // Realness verdict, optional.\n  if (outRec.realness !== undefined) {\n    const r = outRec.realness\n    if (r === null || typeof r !== 'object') {\n      throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n    }\n    const rr = r as Record<string, unknown>\n    expectFiniteNumber(rr.score, 'outcome.realness.score')\n    if (typeof rr.gated !== 'boolean') {\n      throw new RunRecordValidationError(\n        'outcome.realness.gated must be a boolean',\n        'outcome.realness.gated',\n      )\n    }\n  }\n\n  // Per-judge / per-dim breakdown, optional.\n  if (outRec.judgeScores !== undefined) {\n    validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n  }\n\n  // Failure mode optional.\n  if (\n    obj.failureClass !== undefined &&\n    (typeof obj.failureClass !== 'string' ||\n      !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n  ) {\n    throw new RunRecordValidationError(\n      `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n      'failureClass',\n    )\n  }\n  if (obj.failureMode !== undefined) {\n    expectString(obj.failureMode, 'failureMode')\n    if (obj.failureClass === undefined || obj.failureClass === 'success') {\n      throw new RunRecordValidationError(\n        'failureMode requires a non-success failureClass',\n        'failureMode',\n      )\n    }\n  }\n\n  if (\n    typeof obj.terminalOutcome !== 'string' ||\n    !TERMINAL_OUTCOMES.includes(obj.terminalOutcome as RunTerminalOutcome)\n  ) {\n    throw new RunRecordValidationError(\n      `terminalOutcome must be one of ${TERMINAL_OUTCOMES.join(', ')}`,\n      'terminalOutcome',\n    )\n  }\n  if (obj.terminalFailureReason !== undefined) {\n    expectString(obj.terminalFailureReason, 'terminalFailureReason')\n    if (\n      obj.terminalOutcome !== 'failed' &&\n      obj.terminalOutcome !== 'cancelled' &&\n      obj.terminalOutcome !== 'incomplete'\n    ) {\n      throw new RunRecordValidationError(\n        'terminalFailureReason requires terminalOutcome failed, cancelled, or incomplete',\n        'terminalFailureReason',\n      )\n    }\n  }\n\n  if (obj.agentProfile !== undefined) {\n    try {\n      const profile = validateAgentProfileCell(obj.agentProfile)\n      if (profile.model !== undefined && profile.model !== obj.model) {\n        throw new RunRecordValidationError(\n          `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n          'agentProfile.model',\n        )\n      }\n      if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n        throw new RunRecordValidationError(\n          `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n          'agentProfile.promptHash',\n        )\n      }\n    } catch (error) {\n      if (error instanceof RunRecordValidationError) throw error\n      if (error instanceof Error) {\n        throw new RunRecordValidationError(error.message, 'agentProfile')\n      }\n      throw error\n    }\n  }\n\n  expectString(obj.scenarioId, 'scenarioId')\n\n  // Split tag.\n  if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n    throw new RunRecordValidationError(\n      `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n      'splitTag',\n    )\n  }\n\n  return input as RunRecord\n}\n\nfunction validateCost(costUsd: unknown, provenance: unknown): void {\n  if (provenance === null || typeof provenance !== 'object') {\n    throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n  }\n  const value = provenance as Record<string, unknown>\n  if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n    throw new RunRecordValidationError(\n      'costProvenance.kind must be observed, estimated, or uncaptured',\n      'costProvenance.kind',\n    )\n  }\n  if (value.kind === 'uncaptured') {\n    if (value.usd !== null) {\n      throw new RunRecordValidationError(\n        'uncaptured costProvenance.usd must be null',\n        'costProvenance.usd',\n      )\n    }\n    if (costUsd !== null) {\n      throw new RunRecordValidationError('uncaptured cost requires costUsd to be null', 'costUsd')\n    }\n    return\n  }\n  expectNonNegativeNumber(costUsd, 'costUsd')\n  expectNonNegativeNumber(value.usd, 'costProvenance.usd')\n  if (value.usd !== costUsd) {\n    throw new RunRecordValidationError(\n      'costProvenance.usd must equal costUsd',\n      'costProvenance.usd',\n    )\n  }\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n  try {\n    validateRunRecord(input)\n    return true\n  } catch {\n    return false\n  }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n  input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n  try {\n    return { ok: true, value: validateRunRecord(input) }\n  } catch (e) {\n    if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n    throw e\n  }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n  const json = JSON.stringify(record)\n  return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new RunRecordValidationError(`expected non-empty string`, path)\n  }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value)) {\n    throw new RunRecordValidationError(`expected finite number`, path)\n  }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n  expectFiniteNumber(value, path)\n  if ((value as number) < 0) {\n    throw new RunRecordValidationError('expected non-negative number', path)\n  }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object') {\n    throw new RunRecordValidationError('judgeScores must be an object', path)\n  }\n  const rec = value as Record<string, unknown>\n\n  const perJudge = rec.perJudge\n  if (perJudge === null || typeof perJudge !== 'object') {\n    throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n  }\n  for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n    if (dims === null || typeof dims !== 'object') {\n      throw new RunRecordValidationError(\n        'per-judge entry must be an object of dimension scores',\n        `${path}.perJudge.${judgeId}`,\n      )\n    }\n    for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n      expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n    }\n  }\n\n  const perDimMean = rec.perDimMean\n  if (perDimMean === null || typeof perDimMean !== 'object') {\n    throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n  }\n  for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n    expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n  }\n\n  expectFiniteNumber(rec.composite, `${path}.composite`)\n\n  if (rec.failedJudges !== undefined) {\n    if (!Array.isArray(rec.failedJudges)) {\n      throw new RunRecordValidationError(\n        'failedJudges must be an array of strings',\n        `${path}.failedJudges`,\n      )\n    }\n    for (let i = 0; i < rec.failedJudges.length; i++) {\n      const id = rec.failedJudges[i]\n      if (typeof id !== 'string' || id.length === 0) {\n        throw new RunRecordValidationError(\n          'failedJudges entry must be a non-empty string',\n          `${path}.failedJudges[${i}]`,\n        )\n      }\n    }\n  }\n\n  if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n    throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n  }\n}\n\n/**\n * Heuristic snapshot check. Accepts:\n *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)\n *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)\n *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)\n *   - explicit `:date-...` Vertex-style tags\n *\n * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap\n * silently as providers ship new snapshots.\n */\nexport function modelHasSnapshot(model: string): boolean {\n  if (model.includes('@')) return true\n  if (/-\\d{8}$/.test(model)) return true\n  if (/-\\d{4}-\\d{2}-\\d{2}$/.test(model)) return true\n  if (/:date-/.test(model)) return true\n  return false\n}\n"],"mappings":";;;;;;;;;;;;;AA+OA,SAAgB,aAAa,QAAuC;CAClE,MAAM,QAAQ,cAAc,MAAM;CAClC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ,KAAA;AACvE;AAIA,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,aAAyC;CAAC;CAAU;CAAO;AAAS;AAC1E,MAAM,oBAAuD;CAC3D;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,2BAAb,cAA8C,gBAAgB;CAC5D;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;;;;;;AAOA,SAAgB,kBAAkB,OAA2B;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iBAAiB;CAEtD,MAAM,MAAM;CAEZ,KAAK,MAAM,OAAO,qBAChB,IAAI,EAAE,OAAO,MACX,MAAM,IAAI,yBAAyB,4BAA4B,IAAI,EAAE;CAIzE,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,cAAc,cAAc;CAC7C,aAAa,IAAI,aAAa,aAAa;CAC3C,mBAAmB,IAAI,MAAM,MAAM;CACnC,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,WAAW,WAAW;CACvC,wBAAwB,IAAI,QAAQ,QAAQ;CAC5C,IAAI,IAAI,YAAY,KAAA,GAAW,wBAAwB,IAAI,SAAS,SAAS;CAC7E,aAAa,IAAI,SAAS,IAAI,cAAc;CAG5C,IAAI,CAAC,iBAAiB,IAAI,KAAe,GACvC,MAAM,IAAI,yBACR,UAAU,IAAI,MAAM,wEACpB,OACF;CAIF,MAAM,KAAK,IAAI;CACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,gCAAgC,YAAY;CAEjF,MAAM,QAAQ;CACd,wBAAwB,MAAM,OAAO,kBAAkB;CACvD,wBAAwB,MAAM,QAAQ,mBAAmB;CACzD,IAAI,MAAM,cAAc,KAAA,GAAW;EACjC,wBAAwB,MAAM,WAAW,sBAAsB;EAC/D,IAAK,MAAM,YAAwB,MAAM,QACvC,MAAM,IAAI,yBACR,sDACA,sBACF;CAEJ;CACA,IAAI,MAAM,WAAW,KAAA,GAAW,wBAAwB,MAAM,QAAQ,mBAAmB;CACzF,IAAI,MAAM,eAAe,KAAA,GACvB,wBAAwB,MAAM,YAAY,uBAAuB;CAInE,IAAI,IAAI,kBAAkB,KAAA,GAAW;EACnC,MAAM,KAAK,IAAI;EACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,mCAAmC,eAAe;EAEvF,MAAM,QAAQ;EACd,aAAa,MAAM,OAAO,qBAAqB;EAC/C,aAAa,MAAM,eAAe,6BAA6B;EAC/D,mBAAmB,MAAM,YAAY,0BAA0B;EAC/D,IAAI,OAAO,MAAM,aAAa,WAC5B,MAAM,IAAI,yBACR,0CACA,wBACF;CAEJ;CAGA,MAAM,MAAM,IAAI;CAChB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,6BAA6B,SAAS;CAE3E,MAAM,SAAS;CACf,IAAI,OAAO,gBAAgB,KAAA,GACzB,mBAAmB,OAAO,aAAa,qBAAqB;CAC9D,IAAI,OAAO,iBAAiB,KAAA,GAC1B,mBAAmB,OAAO,cAAc,sBAAsB;CAChE,MAAM,MAAM,OAAO;CACnB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,iCAAiC,aAAa;CAEnF,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,GAA8B,GAChE,mBAAmB,GAAG,eAAe,GAAG;CAG1C,IAAI,OAAO,aAAa,KAAA,GAAW;EACjC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,QAAQ,OAAO,MAAM,UAC7B,MAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;EAE7F,MAAM,KAAK;EACX,mBAAmB,GAAG,OAAO,wBAAwB;EACrD,IAAI,OAAO,GAAG,UAAU,WACtB,MAAM,IAAI,yBACR,4CACA,wBACF;CAEJ;CAGA,IAAI,OAAO,gBAAgB,KAAA,GACzB,oBAAoB,OAAO,aAAa,qBAAqB;CAI/D,IACE,IAAI,iBAAiB,KAAA,MACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAE5D,MAAM,IAAI,yBACR,+BAA+B,gBAAgB,KAAK,IAAI,KACxD,cACF;CAEF,IAAI,IAAI,gBAAgB,KAAA,GAAW;EACjC,aAAa,IAAI,aAAa,aAAa;EAC3C,IAAI,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,WACzD,MAAM,IAAI,yBACR,mDACA,aACF;CAEJ;CAEA,IACE,OAAO,IAAI,oBAAoB,YAC/B,CAAC,kBAAkB,SAAS,IAAI,eAAqC,GAErE,MAAM,IAAI,yBACR,kCAAkC,kBAAkB,KAAK,IAAI,KAC7D,iBACF;CAEF,IAAI,IAAI,0BAA0B,KAAA,GAAW;EAC3C,aAAa,IAAI,uBAAuB,uBAAuB;EAC/D,IACE,IAAI,oBAAoB,YACxB,IAAI,oBAAoB,eACxB,IAAI,oBAAoB,cAExB,MAAM,IAAI,yBACR,mFACA,uBACF;CAEJ;CAEA,IAAI,IAAI,iBAAiB,KAAA,GACvB,IAAI;EACF,MAAM,UAAU,yBAAyB,IAAI,YAAY;EACzD,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,IAAI,OACvD,MAAM,IAAI,yBACR,uBAAuB,QAAQ,MAAM,0BAA0B,IAAI,MAAM,IACzE,oBACF;EAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,IAAI,YACjE,MAAM,IAAI,yBACR,4BAA4B,QAAQ,WAAW,+BAA+B,IAAI,WAAW,IAC7F,yBACF;CAEJ,SAAS,OAAO;EACd,IAAI,iBAAiB,0BAA0B,MAAM;EACrD,IAAI,iBAAiB,OACnB,MAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;EAElE,MAAM;CACR;CAGF,aAAa,IAAI,YAAY,YAAY;CAGzC,IAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GACtF,MAAM,IAAI,yBACR,2BAA2B,WAAW,KAAK,IAAI,EAAE,QAAQ,OAAO,IAAI,QAAQ,KAC5E,UACF;CAGF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,YAA2B;CACjE,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;CAEzF,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,yBACR,kEACA,qBACF;CAEF,IAAI,MAAM,SAAS,cAAc;EAC/B,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,8CACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBAAyB,+CAA+C,SAAS;EAE7F;CACF;CACA,wBAAwB,SAAS,SAAS;CAC1C,wBAAwB,MAAM,KAAK,oBAAoB;CACvD,IAAI,MAAM,QAAQ,SAChB,MAAM,IAAI,yBACR,yCACA,oBACF;AAEJ;;AAGA,SAAgB,YAAY,OAAoC;CAC9D,IAAI;EACF,kBAAkB,KAAK;EACvB,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;;AAGA,SAAgB,mBACd,OACiF;CACjF,IAAI;EACF,OAAO;GAAE,IAAI;GAAM,OAAO,kBAAkB,KAAK;EAAE;CACrD,SAAS,GAAG;EACV,IAAI,aAAa,0BAA0B,OAAO;GAAE,IAAI;GAAO,OAAO;EAAE;EACxE,MAAM;CACR;AACF;;AAGA,SAAgB,mBAAmB,QAA8B;CAC/D,MAAM,OAAO,KAAK,UAAU,MAAM;CAClC,OAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAExE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAErE;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;CACnE,mBAAmB,OAAO,IAAI;CAC9B,IAAK,QAAmB,GACtB,MAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAE3E;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iCAAiC,IAAI;CAE1E,MAAM,MAAM;CAEZ,MAAM,WAAW,IAAI;CACrB,IAAI,aAAa,QAAQ,OAAO,aAAa,UAC3C,MAAM,IAAI,yBAAyB,8BAA8B,GAAG,KAAK,UAAU;CAErF,KAAK,MAAM,CAAC,SAAS,SAAS,OAAO,QAAQ,QAAmC,GAAG;EACjF,IAAI,SAAS,QAAQ,OAAO,SAAS,UACnC,MAAM,IAAI,yBACR,yDACA,GAAG,KAAK,YAAY,SACtB;EAEF,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAA+B,GACvE,mBAAmB,OAAO,GAAG,KAAK,YAAY,QAAQ,GAAG,KAAK;CAElE;CAEA,MAAM,aAAa,IAAI;CACvB,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,gCAAgC,GAAG,KAAK,YAAY;CAEzF,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,UAAqC,GAC5E,mBAAmB,MAAM,GAAG,KAAK,cAAc,KAAK;CAGtD,mBAAmB,IAAI,WAAW,GAAG,KAAK,WAAW;CAErD,IAAI,IAAI,iBAAiB,KAAA,GAAW;EAClC,IAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GACjC,MAAM,IAAI,yBACR,4CACA,GAAG,KAAK,cACV;EAEF,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;GAChD,MAAM,KAAK,IAAI,aAAa;GAC5B,IAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAC1C,MAAM,IAAI,yBACR,iDACA,GAAG,KAAK,gBAAgB,EAAE,EAC5B;EAEJ;CACF;CAEA,IAAI,IAAI,UAAU,KAAA,KAAa,OAAO,IAAI,UAAU,UAClD,MAAM,IAAI,yBAAyB,0BAA0B,GAAG,KAAK,OAAO;AAEhF;;;;;;;;;;;AAYA,SAAgB,iBAAiB,OAAwB;CACvD,IAAI,MAAM,SAAS,GAAG,GAAG,OAAO;CAChC,IAAI,UAAU,KAAK,KAAK,GAAG,OAAO;CAClC,IAAI,sBAAsB,KAAK,KAAK,GAAG,OAAO;CAC9C,IAAI,SAAS,KAAK,KAAK,GAAG,OAAO;CACjC,OAAO;AACT"}
+{"version":3,"file":"run-record-vRgqWmJw.js","names":[],"sources":["../src/run-record.ts"],"sourcesContent":["/**\n * Paper-grade RunRecord schema + runtime validator.\n *\n * Every run that participates in a promotion gate, paper table, or\n * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory\n * fields are exactly those the paper \"Two Loops, Three Roles\" requires\n * for reproducibility: who/what/when/cost/seed/hash, plus the search vs\n * holdout split tag. A task score is optional because execution-only records\n * must preserve missing labels instead of converting errors into zero quality.\n *\n * This is intentionally NOT a replacement for the rich `Run` /\n * `ProposeReviewReport` / `ScenarioResult` types already in the\n * package. Those are runtime structures with full provenance. A\n * `RunRecord` is the analysis-time projection — the JSON-friendly\n * row you'd put in a parquet file or paste into a notebook.\n *\n * Validate at the boundary:\n *\n *   const rec = validateRunRecord(rawJson)         // throws on missing\n *   const ok  = isRunRecord(rawJson)               // boolean check\n *   const rec = parseRunRecordSafe(rawJson)        // { ok, value | error }\n *\n * The validator runs in pure TS — zod is intentionally NOT a\n * dependency. Round-trip tested in `tests/run-record.test.ts`.\n */\n\nimport type { AgentProfileCell } from './agent-profile-cell'\nimport { validateAgentProfileCell } from './agent-profile-cell'\nimport type { CostProvenance } from './cost-ledger'\nimport { ValidationError } from './errors'\n// Value import of a leaf module that itself imports only this file's TYPES —\n// no runtime cycle. It keeps the raw split-score derivation spelled in exactly\n// one place (see `rollout/score-derivation-guard`).\nimport { observedScore } from './rollout/reward'\nimport { FAILURE_CLASSES, type FailureClass } from './trace/schema'\n\n/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the\n *  combined train+test pool that the optimizer is allowed to read. */\nexport type RunSplitTag = 'search' | 'dev' | 'holdout'\n\n/**\n * Explicit execution-lifecycle result for a run.\n *\n * This is separate from task quality (`outcome`) and failure classification.\n * Producers set it only from root-run or process evidence.\n */\nexport type RunTerminalOutcome = 'succeeded' | 'failed' | 'cancelled' | 'incomplete' | 'unknown'\n\nexport interface RunTokenUsage {\n  input: number\n  /** All generated tokens charged as output, including reasoning tokens. */\n  output: number\n  /** Reasoning-token subset of `output`, when the provider reports it. */\n  reasoning?: number\n  /** Prompt tokens served from a provider cache. */\n  cached?: number\n  /** Prompt tokens written into a provider cache. */\n  cacheWrite?: number\n}\n\n/** How a run's USD amount was obtained. */\nexport type RunCostProvenance = CostProvenance\n\nexport interface RunJudgeMetadata {\n  model: string\n  promptVersion: string\n  /** [0,1] confidence the judge declared. Constant judge confidence\n   *  across many runs is a fallback signal (see `canary.ts`). */\n  confidence: number\n  /** True if the judge degraded to a fallback path (rules-only,\n   *  prior-call cache, etc.). The canary uses this to alert. */\n  fallback: boolean\n}\n\n/**\n * Per-judge / per-dimension breakdown for runs scored by an ensemble of\n * judges over a multi-dimensional rubric.\n *\n * The collapsed `outcome.searchScore` / `holdoutScore` carries the\n * composite the gate uses. The full breakdown belongs here so consumers\n * can answer \"which judge disagreed?\", \"which dimension dragged the\n * composite down?\", and \"did half the panel fail?\" without re-running.\n *\n * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and\n * `composite` are convenience projections — derivable but precomputed so\n * downstream IRR primitives (`interRaterReliability`,\n * `corpusInterRaterAgreement`) and reporters don't pay the same\n * aggregation twice.\n *\n * Fail-loud discipline: judges that errored out land in `failedJudges`\n * by id. A missing key in `perJudge` is ambiguous (silent zero vs not\n * run); the explicit list makes a partial-failure recorded as such.\n */\nexport interface JudgeScoresRecord {\n  /** Per-judge per-dimension scores. `{ \"kimi-k2.6\": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */\n  perJudge: Record<string, Record<string, number>>\n  /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */\n  perDimMean: Record<string, number>\n  /** Composite mean across successful judges. Mirrors the task score only\n   *  when `failedJudges` is empty. */\n  composite: number\n  /** Judges that errored or returned an unparseable verdict. Recorded\n   *  by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,\n   *  not inferred from missing keys in `perJudge`. */\n  failedJudges?: string[]\n  /** Free-form notes the judges emitted (joined across judges or\n   *  first-judge only — consumer's choice). */\n  notes?: string\n}\n\nexport interface RunOutcome {\n  /** Score on the search/optimization split. Optional for holdout-only and\n   *  execution-only records. */\n  searchScore?: number\n  /** Score on the held-out split. Optional for search-only and execution-only\n   *  records. When both scores are absent, the run is explicitly unlabeled. */\n  holdoutScore?: number\n  /** Bag of any other metric the run produced — judge dimensions,\n   *  pass/fail counters, latency stats, etc. Numeric only — keeps\n   *  reporters honest. */\n  raw: Record<string, number>\n  /** Per-judge / per-dim breakdown. Consumers writing ensemble\n   *  judgements populate this; substrate primitives like\n   *  `interRaterReliability` and `corpusInterRaterAgreement` accept\n   *  these records as input. Optional — single-judge or scalar-only\n   *  runs leave it unset. */\n  judgeScores?: JudgeScoresRecord\n  /** Authenticity / realness verdict — did the run build the REAL thing on the\n   *  intended infra, or fake it (see `./authenticity`)? Optional: only domains\n   *  with an authenticity config populate it. Carried in the corpus so the\n   *  flywheel / off-policy learning can optimize for real completion, not gamed\n   *  pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run\n   *  must not count as a real success regardless of `score`. */\n  realness?: { score: number; gated: boolean; reason?: string }\n}\n\n/**\n * Mandatory paper-grade fields for a single evaluation run. Optional\n * fields are extension points; mandatory fields throw if missing.\n *\n * Hash discipline:\n *   - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the\n *     model (after any steering bundle merge).\n *   - `configHash` is the sha256 of the effective run config (model,\n *     temperature, tools, judges, splits). The pair (promptHash,\n *     configHash) uniquely identifies an experiment cell.\n *\n * Model snapshot discipline:\n *   - `model` MUST encode a snapshot version. Bare aliases like\n *     `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.\n *     Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.\n */\nexport interface RunRecord {\n  /** UUID for the run. */\n  runId: string\n  /** Logical experiment grouping (a treatment vs a baseline within\n   *  the same sweep should share `experimentId`). */\n  experimentId: string\n  /** Stable identifier for the candidate (variant) being run. The\n   *  promotion gate compares two `candidateId`s on matched items. */\n  candidateId: string\n  /** RNG seed for the run. Always recorded — silent re-seeding is\n   *  the most common cause of non-reproducible numbers. */\n  seed: number\n  /** Model identifier WITH snapshot version. */\n  model: string\n  /** sha256 of the effective prompt (post-steering). */\n  promptHash: string\n  /** sha256 of the effective config. */\n  configHash: string\n  /** Git SHA the harness was run from. */\n  commitSha: string\n  /** End-to-end wall-clock duration in milliseconds. */\n  wallMs: number\n  /** Time spent queued before execution started, if known. */\n  queueMs?: number\n  /** Total USD cost, or null when the producer could not capture one. */\n  costUsd: number | null\n  /** Whether `costUsd` came from billing data, a price calculation, or is unavailable. */\n  costProvenance: RunCostProvenance\n  /** Token usage breakdown. */\n  tokenUsage: RunTokenUsage\n  /** Root-run or process terminal result. Never inferred from a child span. */\n  terminalOutcome: RunTerminalOutcome\n  /** Root-run or process failure reason. Valid only for a failed, cancelled,\n   *  or incomplete terminal result; never populated from a child span. */\n  terminalFailureReason?: string\n  /** Judge-side metadata, if a judge was used. */\n  judgeMetadata?: RunJudgeMetadata\n  /** Per-split scores + raw bag. */\n  outcome: RunOutcome\n  /** Canonical task-failure class drawn from the shared\n   *  `FAILURE_CLASSES` taxonomy. Producers set it only from task-result\n   *  evidence. Execution errors belong in\n   *  `outcome.raw.execution_error_count`. */\n  failureClass?: FailureClass\n  /** Free-form task-failure detail scoped under a non-success\n   *  `failureClass`. It is invalid without that class. */\n  failureMode?: string\n  /** Which split this run was drawn from. */\n  splitTag: RunSplitTag\n  /**\n   * Stable scenario identifier the run observed or was scored against.\n   * Comparison primitives match this identity rather than input order.\n   */\n  scenarioId: string\n  /**\n   * Canonical identity for the agent profile cell that produced this row:\n   * profile artifact hash plus optional harness/model/prompt/reporting\n   * dimensions. Use `agentProfile.cellId` to group persona sweeps and\n   * longitudinal reports by the complete source profile, not by a loose\n   * candidate label or opaque config hash.\n   */\n  agentProfile?: AgentProfileCell\n}\n\n/**\n * Canonical task-result classification.\n *\n * A producer may omit classification, record explicit success, or attach\n * domain-specific detail to a non-success class. Detail can never stand alone.\n * Execution errors belong in `outcome.raw.execution_error_count`.\n */\nexport type RunTaskFailure =\n  | { failureClass?: undefined; failureMode?: undefined }\n  | { failureClass: 'success'; failureMode?: undefined }\n  | {\n      failureClass: Exclude<FailureClass, 'success'>\n      failureMode?: string\n    }\n\n/**\n * Return task quality, preferring held-out evidence when both scores exist.\n *\n * RAW: no realness protection is applied. Built on `observedScore` rather\n * than repeating the split derivation, so only `rollout/reward.ts` reads the\n * raw fields. Anything that becomes training data must use `trainingScore` or\n * `trainingReward` instead.\n */\nexport function runTaskScore(record: RunRecord): number | undefined {\n  const score = observedScore(record)\n  return typeof score === 'number' && Number.isFinite(score) ? score : undefined\n}\n\n// ── Validation ───────────────────────────────────────────────────────\n\nconst MANDATORY_TOP_LEVEL = [\n  'runId',\n  'experimentId',\n  'candidateId',\n  'seed',\n  'model',\n  'promptHash',\n  'configHash',\n  'commitSha',\n  'wallMs',\n  'costUsd',\n  'costProvenance',\n  'tokenUsage',\n  'terminalOutcome',\n  'outcome',\n  'splitTag',\n  'scenarioId',\n] as const\n\nconst SPLIT_TAGS: ReadonlyArray<RunSplitTag> = ['search', 'dev', 'holdout']\nconst TERMINAL_OUTCOMES: ReadonlyArray<RunTerminalOutcome> = [\n  'succeeded',\n  'failed',\n  'cancelled',\n  'incomplete',\n  'unknown',\n]\n\nexport class RunRecordValidationError extends ValidationError {\n  readonly path: string\n  constructor(message: string, path = '') {\n    super(path ? `${message} (at ${path})` : message)\n    this.path = path\n  }\n}\n\n/**\n * Strict validator. Throws `RunRecordValidationError` on the first\n * missing or wrongly-typed field. Returns the input cast to\n * `RunRecord` on success — the validator does not coerce.\n */\nexport function validateRunRecord(input: unknown): RunRecord {\n  if (input === null || typeof input !== 'object') {\n    throw new RunRecordValidationError('expected object')\n  }\n  const obj = input as Record<string, unknown>\n\n  for (const key of MANDATORY_TOP_LEVEL) {\n    if (!(key in obj)) {\n      throw new RunRecordValidationError(`missing mandatory field \"${key}\"`)\n    }\n  }\n\n  expectString(obj.runId, 'runId')\n  expectString(obj.experimentId, 'experimentId')\n  expectString(obj.candidateId, 'candidateId')\n  expectFiniteNumber(obj.seed, 'seed')\n  expectString(obj.model, 'model')\n  expectString(obj.promptHash, 'promptHash')\n  expectString(obj.configHash, 'configHash')\n  expectString(obj.commitSha, 'commitSha')\n  expectNonNegativeNumber(obj.wallMs, 'wallMs')\n  if (obj.queueMs !== undefined) expectNonNegativeNumber(obj.queueMs, 'queueMs')\n  validateCost(obj.costUsd, obj.costProvenance)\n\n  // Snapshot discipline: bare model aliases are not paper-grade.\n  if (!modelHasSnapshot(obj.model as string)) {\n    throw new RunRecordValidationError(\n      `model \"${obj.model}\" lacks a snapshot version (use 'name@YYYY-MM-DD' or 'name-YYYYMMDD')`,\n      'model',\n    )\n  }\n\n  // Token usage.\n  const tu = obj.tokenUsage\n  if (tu === null || typeof tu !== 'object') {\n    throw new RunRecordValidationError('tokenUsage must be an object', 'tokenUsage')\n  }\n  const tuRec = tu as Record<string, unknown>\n  expectNonNegativeNumber(tuRec.input, 'tokenUsage.input')\n  expectNonNegativeNumber(tuRec.output, 'tokenUsage.output')\n  if (tuRec.reasoning !== undefined) {\n    expectNonNegativeNumber(tuRec.reasoning, 'tokenUsage.reasoning')\n    if ((tuRec.reasoning as number) > (tuRec.output as number)) {\n      throw new RunRecordValidationError(\n        'reasoning tokens must be a subset of output tokens',\n        'tokenUsage.reasoning',\n      )\n    }\n  }\n  if (tuRec.cached !== undefined) expectNonNegativeNumber(tuRec.cached, 'tokenUsage.cached')\n  if (tuRec.cacheWrite !== undefined) {\n    expectNonNegativeNumber(tuRec.cacheWrite, 'tokenUsage.cacheWrite')\n  }\n\n  // Judge metadata, optional.\n  if (obj.judgeMetadata !== undefined) {\n    const jm = obj.judgeMetadata\n    if (jm === null || typeof jm !== 'object') {\n      throw new RunRecordValidationError('judgeMetadata must be an object', 'judgeMetadata')\n    }\n    const jmRec = jm as Record<string, unknown>\n    expectString(jmRec.model, 'judgeMetadata.model')\n    expectString(jmRec.promptVersion, 'judgeMetadata.promptVersion')\n    expectFiniteNumber(jmRec.confidence, 'judgeMetadata.confidence')\n    if (typeof jmRec.fallback !== 'boolean') {\n      throw new RunRecordValidationError(\n        'judgeMetadata.fallback must be boolean',\n        'judgeMetadata.fallback',\n      )\n    }\n  }\n\n  // Outcome.\n  const out = obj.outcome\n  if (out === null || typeof out !== 'object') {\n    throw new RunRecordValidationError('outcome must be an object', 'outcome')\n  }\n  const outRec = out as Record<string, unknown>\n  if (outRec.searchScore !== undefined)\n    expectFiniteNumber(outRec.searchScore, 'outcome.searchScore')\n  if (outRec.holdoutScore !== undefined)\n    expectFiniteNumber(outRec.holdoutScore, 'outcome.holdoutScore')\n  const raw = outRec.raw\n  if (raw === null || typeof raw !== 'object') {\n    throw new RunRecordValidationError('outcome.raw must be an object', 'outcome.raw')\n  }\n  for (const [k, v] of Object.entries(raw as Record<string, unknown>)) {\n    expectFiniteNumber(v, `outcome.raw.${k}`)\n  }\n  // Realness verdict, optional.\n  if (outRec.realness !== undefined) {\n    const r = outRec.realness\n    if (r === null || typeof r !== 'object') {\n      throw new RunRecordValidationError('outcome.realness must be an object', 'outcome.realness')\n    }\n    const rr = r as Record<string, unknown>\n    expectFiniteNumber(rr.score, 'outcome.realness.score')\n    if (typeof rr.gated !== 'boolean') {\n      throw new RunRecordValidationError(\n        'outcome.realness.gated must be a boolean',\n        'outcome.realness.gated',\n      )\n    }\n  }\n\n  // Per-judge / per-dim breakdown, optional.\n  if (outRec.judgeScores !== undefined) {\n    validateJudgeScores(outRec.judgeScores, 'outcome.judgeScores')\n  }\n\n  // Failure mode optional.\n  if (\n    obj.failureClass !== undefined &&\n    (typeof obj.failureClass !== 'string' ||\n      !FAILURE_CLASSES.includes(obj.failureClass as FailureClass))\n  ) {\n    throw new RunRecordValidationError(\n      `failureClass must be one of ${FAILURE_CLASSES.join(', ')}`,\n      'failureClass',\n    )\n  }\n  if (obj.failureMode !== undefined) {\n    expectString(obj.failureMode, 'failureMode')\n    if (obj.failureClass === undefined || obj.failureClass === 'success') {\n      throw new RunRecordValidationError(\n        'failureMode requires a non-success failureClass',\n        'failureMode',\n      )\n    }\n  }\n\n  if (\n    typeof obj.terminalOutcome !== 'string' ||\n    !TERMINAL_OUTCOMES.includes(obj.terminalOutcome as RunTerminalOutcome)\n  ) {\n    throw new RunRecordValidationError(\n      `terminalOutcome must be one of ${TERMINAL_OUTCOMES.join(', ')}`,\n      'terminalOutcome',\n    )\n  }\n  if (obj.terminalFailureReason !== undefined) {\n    expectString(obj.terminalFailureReason, 'terminalFailureReason')\n    if (\n      obj.terminalOutcome !== 'failed' &&\n      obj.terminalOutcome !== 'cancelled' &&\n      obj.terminalOutcome !== 'incomplete'\n    ) {\n      throw new RunRecordValidationError(\n        'terminalFailureReason requires terminalOutcome failed, cancelled, or incomplete',\n        'terminalFailureReason',\n      )\n    }\n  }\n\n  if (obj.agentProfile !== undefined) {\n    try {\n      const profile = validateAgentProfileCell(obj.agentProfile)\n      if (profile.model !== undefined && profile.model !== obj.model) {\n        throw new RunRecordValidationError(\n          `agentProfile.model \"${profile.model}\" does not match model \"${obj.model}\"`,\n          'agentProfile.model',\n        )\n      }\n      if (profile.promptHash !== undefined && profile.promptHash !== obj.promptHash) {\n        throw new RunRecordValidationError(\n          `agentProfile.promptHash \"${profile.promptHash}\" does not match promptHash \"${obj.promptHash}\"`,\n          'agentProfile.promptHash',\n        )\n      }\n    } catch (error) {\n      if (error instanceof RunRecordValidationError) throw error\n      if (error instanceof Error) {\n        throw new RunRecordValidationError(error.message, 'agentProfile')\n      }\n      throw error\n    }\n  }\n\n  expectString(obj.scenarioId, 'scenarioId')\n\n  // Split tag.\n  if (typeof obj.splitTag !== 'string' || !SPLIT_TAGS.includes(obj.splitTag as RunSplitTag)) {\n    throw new RunRecordValidationError(\n      `splitTag must be one of ${SPLIT_TAGS.join(', ')}, got ${String(obj.splitTag)}`,\n      'splitTag',\n    )\n  }\n\n  return input as RunRecord\n}\n\nfunction validateCost(costUsd: unknown, provenance: unknown): void {\n  if (provenance === null || typeof provenance !== 'object') {\n    throw new RunRecordValidationError('costProvenance must be an object', 'costProvenance')\n  }\n  const value = provenance as Record<string, unknown>\n  if (value.kind !== 'observed' && value.kind !== 'estimated' && value.kind !== 'uncaptured') {\n    throw new RunRecordValidationError(\n      'costProvenance.kind must be observed, estimated, or uncaptured',\n      'costProvenance.kind',\n    )\n  }\n  if (value.kind === 'uncaptured') {\n    if (value.usd !== null) {\n      throw new RunRecordValidationError(\n        'uncaptured costProvenance.usd must be null',\n        'costProvenance.usd',\n      )\n    }\n    if (costUsd !== null) {\n      throw new RunRecordValidationError('uncaptured cost requires costUsd to be null', 'costUsd')\n    }\n    return\n  }\n  expectNonNegativeNumber(costUsd, 'costUsd')\n  expectNonNegativeNumber(value.usd, 'costProvenance.usd')\n  if (value.usd !== costUsd) {\n    throw new RunRecordValidationError(\n      'costProvenance.usd must equal costUsd',\n      'costProvenance.usd',\n    )\n  }\n}\n\n/** Boolean validator — convenience for filtering arrays. */\nexport function isRunRecord(input: unknown): input is RunRecord {\n  try {\n    validateRunRecord(input)\n    return true\n  } catch {\n    return false\n  }\n}\n\n/** Non-throwing validator — returns a discriminated union. */\nexport function parseRunRecordSafe(\n  input: unknown,\n): { ok: true; value: RunRecord } | { ok: false; error: RunRecordValidationError } {\n  try {\n    return { ok: true, value: validateRunRecord(input) }\n  } catch (e) {\n    if (e instanceof RunRecordValidationError) return { ok: false, error: e }\n    throw e\n  }\n}\n\n/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */\nexport function roundTripRunRecord(record: RunRecord): RunRecord {\n  const json = JSON.stringify(record)\n  return validateRunRecord(JSON.parse(json))\n}\n\n// ── Internals ────────────────────────────────────────────────────────\n\nfunction expectString(value: unknown, path: string): void {\n  if (typeof value !== 'string' || value.length === 0) {\n    throw new RunRecordValidationError(`expected non-empty string`, path)\n  }\n}\n\nfunction expectFiniteNumber(value: unknown, path: string): void {\n  if (typeof value !== 'number' || !Number.isFinite(value)) {\n    throw new RunRecordValidationError(`expected finite number`, path)\n  }\n}\n\nfunction expectNonNegativeNumber(value: unknown, path: string): void {\n  expectFiniteNumber(value, path)\n  if ((value as number) < 0) {\n    throw new RunRecordValidationError('expected non-negative number', path)\n  }\n}\n\nfunction validateJudgeScores(value: unknown, path: string): void {\n  if (value === null || typeof value !== 'object') {\n    throw new RunRecordValidationError('judgeScores must be an object', path)\n  }\n  const rec = value as Record<string, unknown>\n\n  const perJudge = rec.perJudge\n  if (perJudge === null || typeof perJudge !== 'object') {\n    throw new RunRecordValidationError('perJudge must be an object', `${path}.perJudge`)\n  }\n  for (const [judgeId, dims] of Object.entries(perJudge as Record<string, unknown>)) {\n    if (dims === null || typeof dims !== 'object') {\n      throw new RunRecordValidationError(\n        'per-judge entry must be an object of dimension scores',\n        `${path}.perJudge.${judgeId}`,\n      )\n    }\n    for (const [dim, score] of Object.entries(dims as Record<string, unknown>)) {\n      expectFiniteNumber(score, `${path}.perJudge.${judgeId}.${dim}`)\n    }\n  }\n\n  const perDimMean = rec.perDimMean\n  if (perDimMean === null || typeof perDimMean !== 'object') {\n    throw new RunRecordValidationError('perDimMean must be an object', `${path}.perDimMean`)\n  }\n  for (const [dim, mean] of Object.entries(perDimMean as Record<string, unknown>)) {\n    expectFiniteNumber(mean, `${path}.perDimMean.${dim}`)\n  }\n\n  expectFiniteNumber(rec.composite, `${path}.composite`)\n\n  if (rec.failedJudges !== undefined) {\n    if (!Array.isArray(rec.failedJudges)) {\n      throw new RunRecordValidationError(\n        'failedJudges must be an array of strings',\n        `${path}.failedJudges`,\n      )\n    }\n    for (let i = 0; i < rec.failedJudges.length; i++) {\n      const id = rec.failedJudges[i]\n      if (typeof id !== 'string' || id.length === 0) {\n        throw new RunRecordValidationError(\n          'failedJudges entry must be a non-empty string',\n          `${path}.failedJudges[${i}]`,\n        )\n      }\n    }\n  }\n\n  if (rec.notes !== undefined && typeof rec.notes !== 'string') {\n    throw new RunRecordValidationError('notes must be a string', `${path}.notes`)\n  }\n}\n\n/**\n * Heuristic snapshot check. Accepts:\n *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)\n *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)\n *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)\n *   - explicit `:date-...` Vertex-style tags\n *\n * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap\n * silently as providers ship new snapshots.\n */\nexport function modelHasSnapshot(model: string): boolean {\n  if (model.includes('@')) return true\n  if (/-\\d{8}$/.test(model)) return true\n  if (/-\\d{4}-\\d{2}-\\d{2}$/.test(model)) return true\n  if (/:date-/.test(model)) return true\n  return false\n}\n"],"mappings":";;;;;;;;;;;;;AA+OA,SAAgB,aAAa,QAAuC;CAClE,MAAM,QAAQ,cAAc,MAAM;CAClC,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ,KAAA;AACvE;AAIA,MAAM,sBAAsB;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,MAAM,aAAyC;CAAC;CAAU;CAAO;AAAS;AAC1E,MAAM,oBAAuD;CAC3D;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,2BAAb,cAA8C,gBAAgB;CAC5D;CACA,YAAY,SAAiB,OAAO,IAAI;EACtC,MAAM,OAAO,GAAG,QAAQ,OAAO,KAAK,KAAK,OAAO;EAChD,KAAK,OAAO;CACd;AACF;;;;;;AAOA,SAAgB,kBAAkB,OAA2B;CAC3D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iBAAiB;CAEtD,MAAM,MAAM;CAEZ,KAAK,MAAM,OAAO,qBAChB,IAAI,EAAE,OAAO,MACX,MAAM,IAAI,yBAAyB,4BAA4B,IAAI,EAAE;CAIzE,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,cAAc,cAAc;CAC7C,aAAa,IAAI,aAAa,aAAa;CAC3C,mBAAmB,IAAI,MAAM,MAAM;CACnC,aAAa,IAAI,OAAO,OAAO;CAC/B,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,YAAY,YAAY;CACzC,aAAa,IAAI,WAAW,WAAW;CACvC,wBAAwB,IAAI,QAAQ,QAAQ;CAC5C,IAAI,IAAI,YAAY,KAAA,GAAW,wBAAwB,IAAI,SAAS,SAAS;CAC7E,aAAa,IAAI,SAAS,IAAI,cAAc;CAG5C,IAAI,CAAC,iBAAiB,IAAI,KAAe,GACvC,MAAM,IAAI,yBACR,UAAU,IAAI,MAAM,wEACpB,OACF;CAIF,MAAM,KAAK,IAAI;CACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,gCAAgC,YAAY;CAEjF,MAAM,QAAQ;CACd,wBAAwB,MAAM,OAAO,kBAAkB;CACvD,wBAAwB,MAAM,QAAQ,mBAAmB;CACzD,IAAI,MAAM,cAAc,KAAA,GAAW;EACjC,wBAAwB,MAAM,WAAW,sBAAsB;EAC/D,IAAK,MAAM,YAAwB,MAAM,QACvC,MAAM,IAAI,yBACR,sDACA,sBACF;CAEJ;CACA,IAAI,MAAM,WAAW,KAAA,GAAW,wBAAwB,MAAM,QAAQ,mBAAmB;CACzF,IAAI,MAAM,eAAe,KAAA,GACvB,wBAAwB,MAAM,YAAY,uBAAuB;CAInE,IAAI,IAAI,kBAAkB,KAAA,GAAW;EACnC,MAAM,KAAK,IAAI;EACf,IAAI,OAAO,QAAQ,OAAO,OAAO,UAC/B,MAAM,IAAI,yBAAyB,mCAAmC,eAAe;EAEvF,MAAM,QAAQ;EACd,aAAa,MAAM,OAAO,qBAAqB;EAC/C,aAAa,MAAM,eAAe,6BAA6B;EAC/D,mBAAmB,MAAM,YAAY,0BAA0B;EAC/D,IAAI,OAAO,MAAM,aAAa,WAC5B,MAAM,IAAI,yBACR,0CACA,wBACF;CAEJ;CAGA,MAAM,MAAM,IAAI;CAChB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,6BAA6B,SAAS;CAE3E,MAAM,SAAS;CACf,IAAI,OAAO,gBAAgB,KAAA,GACzB,mBAAmB,OAAO,aAAa,qBAAqB;CAC9D,IAAI,OAAO,iBAAiB,KAAA,GAC1B,mBAAmB,OAAO,cAAc,sBAAsB;CAChE,MAAM,MAAM,OAAO;CACnB,IAAI,QAAQ,QAAQ,OAAO,QAAQ,UACjC,MAAM,IAAI,yBAAyB,iCAAiC,aAAa;CAEnF,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,GAA8B,GAChE,mBAAmB,GAAG,eAAe,GAAG;CAG1C,IAAI,OAAO,aAAa,KAAA,GAAW;EACjC,MAAM,IAAI,OAAO;EACjB,IAAI,MAAM,QAAQ,OAAO,MAAM,UAC7B,MAAM,IAAI,yBAAyB,sCAAsC,kBAAkB;EAE7F,MAAM,KAAK;EACX,mBAAmB,GAAG,OAAO,wBAAwB;EACrD,IAAI,OAAO,GAAG,UAAU,WACtB,MAAM,IAAI,yBACR,4CACA,wBACF;CAEJ;CAGA,IAAI,OAAO,gBAAgB,KAAA,GACzB,oBAAoB,OAAO,aAAa,qBAAqB;CAI/D,IACE,IAAI,iBAAiB,KAAA,MACpB,OAAO,IAAI,iBAAiB,YAC3B,CAAC,gBAAgB,SAAS,IAAI,YAA4B,IAE5D,MAAM,IAAI,yBACR,+BAA+B,gBAAgB,KAAK,IAAI,KACxD,cACF;CAEF,IAAI,IAAI,gBAAgB,KAAA,GAAW;EACjC,aAAa,IAAI,aAAa,aAAa;EAC3C,IAAI,IAAI,iBAAiB,KAAA,KAAa,IAAI,iBAAiB,WACzD,MAAM,IAAI,yBACR,mDACA,aACF;CAEJ;CAEA,IACE,OAAO,IAAI,oBAAoB,YAC/B,CAAC,kBAAkB,SAAS,IAAI,eAAqC,GAErE,MAAM,IAAI,yBACR,kCAAkC,kBAAkB,KAAK,IAAI,KAC7D,iBACF;CAEF,IAAI,IAAI,0BAA0B,KAAA,GAAW;EAC3C,aAAa,IAAI,uBAAuB,uBAAuB;EAC/D,IACE,IAAI,oBAAoB,YACxB,IAAI,oBAAoB,eACxB,IAAI,oBAAoB,cAExB,MAAM,IAAI,yBACR,mFACA,uBACF;CAEJ;CAEA,IAAI,IAAI,iBAAiB,KAAA,GACvB,IAAI;EACF,MAAM,UAAU,yBAAyB,IAAI,YAAY;EACzD,IAAI,QAAQ,UAAU,KAAA,KAAa,QAAQ,UAAU,IAAI,OACvD,MAAM,IAAI,yBACR,uBAAuB,QAAQ,MAAM,0BAA0B,IAAI,MAAM,IACzE,oBACF;EAEF,IAAI,QAAQ,eAAe,KAAA,KAAa,QAAQ,eAAe,IAAI,YACjE,MAAM,IAAI,yBACR,4BAA4B,QAAQ,WAAW,+BAA+B,IAAI,WAAW,IAC7F,yBACF;CAEJ,SAAS,OAAO;EACd,IAAI,iBAAiB,0BAA0B,MAAM;EACrD,IAAI,iBAAiB,OACnB,MAAM,IAAI,yBAAyB,MAAM,SAAS,cAAc;EAElE,MAAM;CACR;CAGF,aAAa,IAAI,YAAY,YAAY;CAGzC,IAAI,OAAO,IAAI,aAAa,YAAY,CAAC,WAAW,SAAS,IAAI,QAAuB,GACtF,MAAM,IAAI,yBACR,2BAA2B,WAAW,KAAK,IAAI,EAAE,QAAQ,OAAO,IAAI,QAAQ,KAC5E,UACF;CAGF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,YAA2B;CACjE,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,oCAAoC,gBAAgB;CAEzF,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS,cAAc,MAAM,SAAS,eAAe,MAAM,SAAS,cAC5E,MAAM,IAAI,yBACR,kEACA,qBACF;CAEF,IAAI,MAAM,SAAS,cAAc;EAC/B,IAAI,MAAM,QAAQ,MAChB,MAAM,IAAI,yBACR,8CACA,oBACF;EAEF,IAAI,YAAY,MACd,MAAM,IAAI,yBAAyB,+CAA+C,SAAS;EAE7F;CACF;CACA,wBAAwB,SAAS,SAAS;CAC1C,wBAAwB,MAAM,KAAK,oBAAoB;CACvD,IAAI,MAAM,QAAQ,SAChB,MAAM,IAAI,yBACR,yCACA,oBACF;AAEJ;;AAGA,SAAgB,YAAY,OAAoC;CAC9D,IAAI;EACF,kBAAkB,KAAK;EACvB,OAAO;CACT,QAAQ;EACN,OAAO;CACT;AACF;;AAGA,SAAgB,mBACd,OACiF;CACjF,IAAI;EACF,OAAO;GAAE,IAAI;GAAM,OAAO,kBAAkB,KAAK;EAAE;CACrD,SAAS,GAAG;EACV,IAAI,aAAa,0BAA0B,OAAO;GAAE,IAAI;GAAO,OAAO;EAAE;EACxE,MAAM;CACR;AACF;;AAGA,SAAgB,mBAAmB,QAA8B;CAC/D,MAAM,OAAO,KAAK,UAAU,MAAM;CAClC,OAAO,kBAAkB,KAAK,MAAM,IAAI,CAAC;AAC3C;AAIA,SAAS,aAAa,OAAgB,MAAoB;CACxD,IAAI,OAAO,UAAU,YAAY,MAAM,WAAW,GAChD,MAAM,IAAI,yBAAyB,6BAA6B,IAAI;AAExE;AAEA,SAAS,mBAAmB,OAAgB,MAAoB;CAC9D,IAAI,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GACrD,MAAM,IAAI,yBAAyB,0BAA0B,IAAI;AAErE;AAEA,SAAS,wBAAwB,OAAgB,MAAoB;CACnE,mBAAmB,OAAO,IAAI;CAC9B,IAAK,QAAmB,GACtB,MAAM,IAAI,yBAAyB,gCAAgC,IAAI;AAE3E;AAEA,SAAS,oBAAoB,OAAgB,MAAoB;CAC/D,IAAI,UAAU,QAAQ,OAAO,UAAU,UACrC,MAAM,IAAI,yBAAyB,iCAAiC,IAAI;CAE1E,MAAM,MAAM;CAEZ,MAAM,WAAW,IAAI;CACrB,IAAI,aAAa,QAAQ,OAAO,aAAa,UAC3C,MAAM,IAAI,yBAAyB,8BAA8B,GAAG,KAAK,UAAU;CAErF,KAAK,MAAM,CAAC,SAAS,SAAS,OAAO,QAAQ,QAAmC,GAAG;EACjF,IAAI,SAAS,QAAQ,OAAO,SAAS,UACnC,MAAM,IAAI,yBACR,yDACA,GAAG,KAAK,YAAY,SACtB;EAEF,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,IAA+B,GACvE,mBAAmB,OAAO,GAAG,KAAK,YAAY,QAAQ,GAAG,KAAK;CAElE;CAEA,MAAM,aAAa,IAAI;CACvB,IAAI,eAAe,QAAQ,OAAO,eAAe,UAC/C,MAAM,IAAI,yBAAyB,gCAAgC,GAAG,KAAK,YAAY;CAEzF,KAAK,MAAM,CAAC,KAAK,SAAS,OAAO,QAAQ,UAAqC,GAC5E,mBAAmB,MAAM,GAAG,KAAK,cAAc,KAAK;CAGtD,mBAAmB,IAAI,WAAW,GAAG,KAAK,WAAW;CAErD,IAAI,IAAI,iBAAiB,KAAA,GAAW;EAClC,IAAI,CAAC,MAAM,QAAQ,IAAI,YAAY,GACjC,MAAM,IAAI,yBACR,4CACA,GAAG,KAAK,cACV;EAEF,KAAK,IAAI,IAAI,GAAG,IAAI,IAAI,aAAa,QAAQ,KAAK;GAChD,MAAM,KAAK,IAAI,aAAa;GAC5B,IAAI,OAAO,OAAO,YAAY,GAAG,WAAW,GAC1C,MAAM,IAAI,yBACR,iDACA,GAAG,KAAK,gBAAgB,EAAE,EAC5B;EAEJ;CACF;CAEA,IAAI,IAAI,UAAU,KAAA,KAAa,OAAO,IAAI,UAAU,UAClD,MAAM,IAAI,yBAAyB,0BAA0B,GAAG,KAAK,OAAO;AAEhF;;;;;;;;;;;AAYA,SAAgB,iBAAiB,OAAwB;CACvD,IAAI,MAAM,SAAS,GAAG,GAAG,OAAO;CAChC,IAAI,UAAU,KAAK,KAAK,GAAG,OAAO;CAClC,IAAI,sBAAsB,KAAK,KAAK,GAAG,OAAO;CAC9C,IAAI,SAAS,KAAK,KAAK,GAAG,OAAO;CACjC,OAAO;AACT"}

package/dist/{semantic-concept-judge-Btozx3Vc.js → semantic-concept-judge-Bz64IckK.js}

@@ -1,6 +1,6 @@
-import { a as NotFoundError } from "./errors-8YnH8WlF.js";
-import { i as CostLedger } from "./cost-ledger-DHAjwNj7.js";
-import { c as callLlmJson, f as maximumChargeForLlmRequest, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-DHx8pzyJ.js";
+import { o as NotFoundError } from "./errors-D-LKuDhb.js";
+import { i as CostLedger } from "./cost-ledger-D-5_-dhi.js";
+import { c as callLlmJson, f as maximumChargeForLlmRequest, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-DKB25jV8.js";
 import { a as clamp01, i as aggregateRunScore, s as computeFindingId } from "./proposal-findings-DCawte-y.js";
 import { t as Mutex } from "./concurrency-MUjT7VjM.js";
 import { appendFileSync, existsSync, mkdirSync, readFileSync, readdirSync, statSync } from "node:fs";
@@ -721,4 +721,4 @@ function createSemanticConceptJudge(options = {}) {
 //#endregion
 export { RunCritic as a, buildSkillUsageReport as c, defaultIsMaterial as d, diffFindings as f, runSemanticConceptJudge as i, emitSkillUsageFindings as l, SEMANTIC_CONCEPT_JUDGE_VERSION as n, SKILL_USAGE_ANALYST as o, LockedJsonlAppender as p, createSemanticConceptJudge as r, SkillUsageAnalyst as s, DEFAULT_COMPLEXITY_WEIGHTS as t, FindingsStore as u };
 
-//# sourceMappingURL=semantic-concept-judge-Btozx3Vc.js.map
+//# sourceMappingURL=semantic-concept-judge-Bz64IckK.js.map

package/dist/{semantic-concept-judge-Btozx3Vc.js.map → semantic-concept-judge-Bz64IckK.js.map}

@@ -1 +1 @@
-{"version":3,"file":"semantic-concept-judge-Btozx3Vc.js","names":[],"sources":["../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/skill-usage.ts","../src/run-critic.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Skill-usage analyst — a DETERMINISTIC `Analyst` over a Claude/Codex skill\n * library + its trace corpus. Unlike the trace-store kinds (failure-mode,\n * improvement, ...) this kind calls no LLM: it mines real usage and skill\n * structure and emits findings by rule.\n *\n * It exists because the naive \"Skill-tool invocation count\" lies low — it\n * misses orchestrated sub-dispatch (a leaf skill run BY /pursue or /governor\n * logs under the parent), slash-command entry, local-script bypass, and\n * on-disk artifacts. The 2026-05-30 skill audit found 39/53 skills at zero\n * direct invocations, yet only one was a genuine cut: the rest were\n * measurement-invisible or discovery-limited. This analyst encodes that\n * lesson as a multi-signal usage model so a cheap repeatable pass can keep\n * the library honest, and so the expensive audit workflow's verdicts can\n * GEPA-distill it toward agreement (see `gold/skill-verdicts.gold.jsonl`).\n *\n * Report-building (`buildSkillUsageReport`, an fs scan) is separated from\n * finding emission (`SkillUsageAnalyst.analyze`, pure) so the slow scan runs\n * once at the registry boundary and the rule logic stays unit-testable.\n */\n\nimport { type Dirent, existsSync, readdirSync, readFileSync, statSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { Analyst, AnalystContext, AnalystFinding, AnalystSeverity } from '../types'\nimport { computeFindingId } from '../types'\n\n// ── Input model ──────────────────────────────────────────────────────\n\nexport type SkillKind = 'public' | 'private'\n\n/** One skill's multi-signal usage + structure. All counts are deterministic. */\nexport interface SkillUsageRecord {\n  name: string\n  kind: SkillKind\n  /** Absolute path to the skill's SKILL.md. */\n  path: string\n  lines: number\n  /** `\"skill\":\"<name>\"` Skill-tool invocations across the trace corpus. */\n  directInvocations: number\n  /** `<command-name>/<name>` slash invocations across the trace corpus. */\n  slashInvocations: number\n  /** Sibling skills whose SKILL.md dispatches to this one (`/<name>`). Proxy\n   *  for orchestrated sub-dispatch the per-skill counter cannot see. */\n  inboundRefs: number\n  /** On-disk artifacts attributable to the skill (e.g. `.evolve/<name>/**`). */\n  artifactCount: number\n  /** Tangle-private reference count in the body (leak signal for public skills). */\n  tanglePrivateRefs: number\n  hasReferencesDir: boolean\n  hasEvalsDir: boolean\n  /** Body mentions `skill-runs.jsonl` (visible to /reflect + /governor). */\n  logsRuns: boolean\n  /** Description carries an explicit `Triggers:` clause / trigger phrases. */\n  hasTriggerPhrases: boolean\n}\n\nexport interface SkillUsageReport {\n  generatedFromTraces: number\n  records: SkillUsageRecord[]\n}\n\nexport interface SkillUsageScanConfig {\n  /** Dirs holding `*.jsonl` transcripts (Claude `~/.claude/projects`, Codex sessions). */\n  transcriptDirs: string[]\n  /** Skill roots to scan; each dir directly under `root` with a `SKILL.md` is a skill. */\n  skillRoots: { root: string; kind: SkillKind }[]\n  /** Roots scanned for `<root>/.evolve/<skill>` artifact dirs. */\n  artifactRoots?: string[]\n  /** Token-prefixed mappings: skill name → extra artifact subpaths under an artifactRoot\n   *  (e.g. reflect → `.evolve/reflections`). Catches non-eponymous artifact dirs. */\n  artifactAliases?: Record<string, string[]>\n  /** Cap files read per transcript dir (bounds a huge corpus); 0 = unbounded. */\n  maxTranscriptsPerDir?: number\n}\n\n// ── Deterministic thresholds ─────────────────────────────────────────\n\n/** Anthropic's authoring guidance keeps SKILL.md short; past this with no\n *  `references/` split the body burns context budget every session. */\nconst BLOAT_LINE_THRESHOLD = 300\n\nconst TANGLE_PRIVATE_RE =\n  /\\b(cli-bridge|tangletools|ops-board|drew-gtr-pro|@tangle-network\\/|~\\/company|tangle\\.tools|gtm-agent)\\b|\\bkimi\\b|\\btcloud\\b/gi\nconst TRIGGER_RE = /triggers?\\s*[:-]/i\n\n// ── Report builder (fs scan — slow, runs once at the registry boundary) ──\n\nfunction listSkillDirs(root: string): { name: string; path: string }[] {\n  if (!existsSync(root)) return []\n  const out: { name: string; path: string }[] = []\n  for (const entry of readdirSync(root, { withFileTypes: true })) {\n    if (!entry.isDirectory() && !entry.isSymbolicLink()) continue\n    const skillMd = join(root, entry.name, 'SKILL.md')\n    if (existsSync(skillMd)) out.push({ name: entry.name, path: skillMd })\n  }\n  return out\n}\n\nfunction walkJsonl(dir: string, cap: number): string[] {\n  if (!existsSync(dir)) return []\n  const files: string[] = []\n  const stack = [dir]\n  while (stack.length) {\n    const cur = stack.pop()!\n    let entries: Dirent[]\n    try {\n      entries = readdirSync(cur, { withFileTypes: true })\n    } catch {\n      continue\n    }\n    for (const e of entries) {\n      const full = join(cur, e.name)\n      if (e.isDirectory()) stack.push(full)\n      else if (e.name.endsWith('.jsonl')) {\n        files.push(full)\n        if (cap > 0 && files.length >= cap) return files\n      }\n    }\n  }\n  return files\n}\n\nfunction frontmatterDescription(body: string): string {\n  const fm = /^---\\n([\\s\\S]*?)\\n---/.exec(body)\n  const block = fm?.[1] ?? ''\n  const m = /description:\\s*(.+)/i.exec(block)\n  return m?.[1] ?? ''\n}\n\nfunction countArtifacts(roots: string[], name: string, aliases: string[]): number {\n  let n = 0\n  for (const root of roots) {\n    const candidates = [join(root, '.evolve', name), ...aliases.map((a) => join(root, a))]\n    for (const dir of candidates) {\n      if (!existsSync(dir)) continue\n      try {\n        if (statSync(dir).isDirectory()) n += readdirSync(dir).length\n        else n += 1\n      } catch {\n        /* unreadable — skip */\n      }\n    }\n  }\n  return n\n}\n\n/** Scan the corpus + skill roots into a {@link SkillUsageReport}. Deterministic. */\nexport function buildSkillUsageReport(config: SkillUsageScanConfig): SkillUsageReport {\n  const skills = config.skillRoots.flatMap(({ root, kind }) =>\n    listSkillDirs(root).map((s) => ({ ...s, kind })),\n  )\n  const names = skills.map((s) => s.name)\n\n  // One pass over the corpus accumulating direct + slash counts per skill.\n  const direct = new Map<string, number>(names.map((n) => [n, 0]))\n  const slash = new Map<string, number>(names.map((n) => [n, 0]))\n  const skillRe = /\"skill\"\\s*:\\s*\"([a-z0-9_:-]+)\"/g\n  const cmdRe = /<command-name>\\/?([a-z0-9_:-]+)<\\/command-name>/g\n  let transcripts = 0\n  for (const dir of config.transcriptDirs) {\n    for (const file of walkJsonl(dir, config.maxTranscriptsPerDir ?? 0)) {\n      transcripts += 1\n      let data: string\n      try {\n        data = readFileSync(file, 'utf8')\n      } catch {\n        continue\n      }\n      for (const m of data.matchAll(skillRe)) {\n        const g = m[1]\n        if (!g) continue\n        const n = g.split(':').pop() ?? g\n        const prev = direct.get(n)\n        if (prev !== undefined) direct.set(n, prev + 1)\n      }\n      for (const m of data.matchAll(cmdRe)) {\n        const g = m[1]\n        if (g === undefined) continue\n        const prev = slash.get(g)\n        if (prev !== undefined) slash.set(g, prev + 1)\n      }\n    }\n  }\n\n  // Read each skill body once; compute structure + inbound refs across siblings.\n  const bodies = new Map<string, string>()\n  for (const s of skills) {\n    try {\n      bodies.set(s.name, readFileSync(s.path, 'utf8'))\n    } catch {\n      bodies.set(s.name, '')\n    }\n  }\n  const inbound = new Map<string, number>(names.map((n) => [n, 0]))\n  for (const target of names) {\n    const ref = new RegExp(`/${target}\\\\b|\\\\[\\\\[${target}\\\\]\\\\]`)\n    for (const s of skills) {\n      if (s.name === target) continue\n      if (ref.test(bodies.get(s.name) ?? '')) inbound.set(target, inbound.get(target)! + 1)\n    }\n  }\n\n  const records: SkillUsageRecord[] = skills.map((s) => {\n    const body = bodies.get(s.name) ?? ''\n    const dir = s.path.replace(/\\/SKILL\\.md$/, '')\n    return {\n      name: s.name,\n      kind: s.kind,\n      path: s.path,\n      lines: body ? body.split('\\n').length : 0,\n      directInvocations: direct.get(s.name) ?? 0,\n      slashInvocations: slash.get(s.name) ?? 0,\n      inboundRefs: inbound.get(s.name) ?? 0,\n      artifactCount: countArtifacts(\n        config.artifactRoots ?? [],\n        s.name,\n        config.artifactAliases?.[s.name] ?? [],\n      ),\n      tanglePrivateRefs: (body.match(TANGLE_PRIVATE_RE) ?? []).length,\n      hasReferencesDir: existsSync(join(dir, 'references')),\n      hasEvalsDir: existsSync(join(dir, 'evals')),\n      logsRuns: body.includes('skill-runs.jsonl'),\n      hasTriggerPhrases: TRIGGER_RE.test(frontmatterDescription(body) || body.slice(0, 600)),\n    }\n  })\n  return { generatedFromTraces: transcripts, records }\n}\n\n// ── Finding emission (pure — unit-testable, no LLM, no fs) ────────────\n\nconst ANALYST_ID = 'skill-usage'\n\nfunction finding(\n  area: string,\n  subject: string,\n  claim: string,\n  severity: AnalystSeverity,\n  confidence: number,\n  producedAt: string,\n  recommended: string,\n  evidenceUri: string,\n  rationale?: string,\n): AnalystFinding {\n  return {\n    schema_version: '1.0.0',\n    finding_id: computeFindingId({ analyst_id: ANALYST_ID, area, subject, claim }),\n    analyst_id: ANALYST_ID,\n    produced_at: producedAt,\n    severity,\n    area,\n    claim,\n    rationale,\n    evidence_refs: [{ kind: 'artifact', uri: evidenceUri }],\n    recommended_action: recommended,\n    confidence,\n    subject,\n  }\n}\n\n/** Pure rule pass over a report → findings. Exported for direct/unit use. */\nexport function emitSkillUsageFindings(\n  report: SkillUsageReport,\n  producedAt: string,\n): AnalystFinding[] {\n  const out: AnalystFinding[] = []\n  for (const r of report.records) {\n    const directTotal = r.directInvocations + r.slashInvocations\n    const trueUsage = directTotal + r.inboundRefs + r.artifactCount\n\n    // 1. Dead: no usage signal of ANY kind. The only real deprecation candidate.\n    if (trueUsage === 0) {\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' has zero usage across all signals (direct, slash, inbound-refs, artifacts)`,\n          'high',\n          0.6,\n          producedAt,\n          'Confirm the skill covers a real recurring job; if not, deprecate. Zero true usage is the only deterministic deprecation candidate.',\n          r.path,\n          'No Skill-tool call, no slash invocation, no sibling dispatches to it, and no on-disk artifacts.',\n        ),\n      )\n    } else if (directTotal === 0 && r.inboundRefs + r.artifactCount > 0) {\n      // 2. Measurement-invisible: real use via orchestration/artifacts, never invoked directly.\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' shows 0 direct invocations but is used via orchestration/artifacts (inbound=${r.inboundRefs}, artifacts=${r.artifactCount})`,\n          'info',\n          0.8,\n          producedAt,\n          'Do NOT treat as unused — usage is real but logged under parent skills or on disk. Strengthen direct-invocation discovery only if direct use is desired.',\n          r.path,\n          'The Skill-tool counter undercounts orchestrated/chained leaf skills.',\n        ),\n      )\n    }\n\n    // 3. Discovery gap: low direct use AND weak trigger surface.\n    if (directTotal <= 2 && !r.hasTriggerPhrases) {\n      out.push(\n        finding(\n          'discoverability',\n          r.name,\n          `Skill '${r.name}' is rarely invoked directly and its description has no explicit trigger phrases`,\n          'medium',\n          0.7,\n          producedAt,\n          'Add a `Triggers:` clause with verbatim user phrases to the frontmatter description so the model auto-invokes it.',\n          r.path,\n        ),\n      )\n    }\n\n    // 4. Public-repo leak.\n    if (r.kind === 'public' && r.tanglePrivateRefs > 0) {\n      out.push(\n        finding(\n          'safety',\n          r.name,\n          `Public skill '${r.name}' carries ${r.tanglePrivateRefs} Tangle-private reference(s)`,\n          'high',\n          0.75,\n          producedAt,\n          'Sanitize incidental internal refs (cli-bridge/kimi/tcloud/~company/private repos) or relocate to a private repo. Verify @tangle-network/* refs are to PUBLISHED packages before treating as a leak.',\n          r.path,\n        ),\n      )\n    }\n\n    // 5. Bloat / no progressive disclosure.\n    if (r.lines > BLOAT_LINE_THRESHOLD && !r.hasReferencesDir) {\n      out.push(\n        finding(\n          'maintainability',\n          r.name,\n          `Skill '${r.name}' is ${r.lines} lines with no references/ split (progressive disclosure)`,\n          'medium',\n          0.8,\n          producedAt,\n          `Split detail into references/ loaded on demand; keep SKILL.md a short overview. ${r.lines} lines load into every session's context budget.`,\n          r.path,\n        ),\n      )\n    }\n\n    // 6. No evals (Anthropic's \">=3 evals before docs\" rule).\n    if (!r.hasEvalsDir) {\n      out.push(\n        finding(\n          'data-quality',\n          r.name,\n          `Skill '${r.name}' ships no evals/`,\n          'low',\n          0.6,\n          producedAt,\n          'Add evals/evals.json with >=3 scenarios proving the skill beats baseline; gives regression coverage.',\n          r.path,\n        ),\n      )\n    }\n\n    // 7. No run logging → invisible to /reflect and /governor.\n    if (!r.logsRuns) {\n      out.push(\n        finding(\n          'observability',\n          r.name,\n          `Skill '${r.name}' never appends to .evolve/skill-runs.jsonl`,\n          'low',\n          0.55,\n          producedAt,\n          'Append one run line to .evolve/skill-runs.jsonl on completion, or declare it a non-logging leaf, so the self-improvement loop can see it ran.',\n          r.path,\n        ),\n      )\n    }\n  }\n  return out\n}\n\n// ── The Analyst ──────────────────────────────────────────────────────\n\nexport class SkillUsageAnalyst implements Analyst<SkillUsageReport> {\n  readonly id = ANALYST_ID\n  readonly description =\n    'Deterministic multi-signal skill-usage analysis: flags dead skills, measurement-invisible (orchestrated) usage, discovery gaps, public-repo leaks, bloat, missing evals, and missing run-logging.'\n  readonly inputKind = 'custom' as const\n  readonly cost = { kind: 'deterministic' as const, est_usd_per_run: 0 }\n  readonly version = '1.0.0'\n\n  async analyze(input: SkillUsageReport, ctx: AnalystContext): Promise<AnalystFinding[]> {\n    const producedAt = ctx.tags?.producedAt ?? new Date().toISOString()\n    ctx.log?.(\n      `skill-usage: ${input.records.length} skills over ${input.generatedFromTraces} transcripts`,\n    )\n    return emitSkillUsageFindings(input, producedAt)\n  }\n}\n\nexport const SKILL_USAGE_ANALYST = new SkillUsageAnalyst()\n","import { NotFoundError } from './errors'\nimport { aggregateRunScore, clamp01, type RunScore, type RunScoreWeights } from './run-score'\nimport type { Artifact, BudgetLedgerEntry, Run, Span, TraceEvent, TraceStore } from './trace'\n\nexport interface RunTrace {\n  run: Run\n  spans: Span[]\n  events: TraceEvent[]\n  artifacts: Artifact[]\n  budget: BudgetLedgerEntry[]\n}\n\nexport interface RunCriticOptions {\n  weights?: Partial<RunScoreWeights>\n  driftPatterns?: RegExp[]\n}\n\nconst DEFAULT_DRIFT_PATTERNS = [\n  /https?:\\/\\//i,\n  /\\btitle:\\s/i,\n  /\\bsummary:\\s/i,\n  /\\burl:\\s/i,\n  /\\bnpm package usage\\b/i,\n  /\\bnews\\b/i,\n]\n\nexport class RunCritic {\n  private readonly weights?: Partial<RunScoreWeights>\n  private readonly driftPatterns: RegExp[]\n\n  constructor(options: RunCriticOptions = {}) {\n    this.weights = options.weights\n    this.driftPatterns = options.driftPatterns ?? DEFAULT_DRIFT_PATTERNS\n  }\n\n  async score(store: TraceStore, runId: string): Promise<RunScore> {\n    const run = await store.getRun(runId)\n    if (!run) throw new NotFoundError(`run ${runId} not found`)\n    const [spans, events, artifacts, budget] = await Promise.all([\n      store.spans({ runId }),\n      store.events({ runId }),\n      store.artifacts(runId),\n      store.budget(runId),\n    ])\n    return this.scoreTrace({ run, spans, events, artifacts, budget })\n  }\n\n  scoreTrace(trace: RunTrace): RunScore {\n    const notes: string[] = []\n    const llmSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'llm' }> => s.kind === 'llm',\n    )\n    const toolSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'tool' }> => s.kind === 'tool',\n    )\n    const judgeSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'judge' }> => s.kind === 'judge',\n    )\n    const sandboxSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'sandbox' }> => s.kind === 'sandbox',\n    )\n    const finalGateSpans = judgeSpans.filter(\n      (span) => span.dimension === 'final_gate' || span.attributes?.finalGate === true,\n    )\n\n    const success =\n      trace.run.outcome?.pass === true ? 1 : trace.run.status === 'completed' ? 0.5 : 0\n    if (!success) notes.push('run did not complete with pass=true')\n\n    const judgeAverage = judgeSpans.length\n      ? judgeSpans.reduce((sum, span) => sum + normalizeJudgeScore(span.score), 0) /\n        judgeSpans.length\n      : undefined\n    const outcomeScore =\n      typeof trace.run.outcome?.score === 'number'\n        ? clamp01(\n            trace.run.outcome.score > 1 ? trace.run.outcome.score / 100 : trace.run.outcome.score,\n          )\n        : undefined\n    const goalProgress = outcomeScore ?? judgeAverage ?? success\n\n    const successfulTools = toolSpans.filter((span) => span.status !== 'error').length\n    const toolUseQuality = toolSpans.length === 0 ? 0 : successfulTools / toolSpans.length\n    if (toolSpans.length === 0) notes.push('no tool spans recorded')\n\n    const patchEvidence =\n      trace.artifacts.length +\n      toolSpans.filter((span) => /write|edit|patch|apply/i.test(span.toolName)).length\n    const patchQuality = patchEvidence > 0 ? clamp01(patchEvidence / 4) : 0\n    if (!patchQuality) notes.push('no artifact or edit evidence recorded')\n\n    const sandboxTests = sandboxSpans.filter(\n      (span) => typeof span.testsTotal === 'number' && span.testsTotal > 0,\n    )\n    const testReality = sandboxTests.length\n      ? sandboxTests.reduce(\n          (sum, span) => sum + (span.testsPassed ?? 0) / Math.max(1, span.testsTotal ?? 1),\n          0,\n        ) / sandboxTests.length\n      : toolSpans.some((span) =>\n            /\\btest|vitest|pytest|jest|build|tsc\\b/i.test(JSON.stringify(span.args)),\n          )\n        ? 0.4\n        : 0\n    if (!testReality) notes.push('no real test/build evidence recorded')\n\n    const blockerSpans = judgeSpans.filter((span) => isBlockingJudge(span))\n    const finalGateBlockers = finalGateSpans.filter((span) => isBlockingJudge(span))\n    const finalGate = finalGateSpans.length ? (finalGateBlockers.length ? 0 : 1) : success\n    if (finalGateBlockers.length)\n      notes.push(`final gate blocked by ${finalGateBlockers.length} reviewer(s)`)\n    else if (!finalGateSpans.length) notes.push('no final gate judgment recorded')\n\n    const reviewerBlockers = judgeSpans.length ? blockerSpans.length / judgeSpans.length : 0\n    if (reviewerBlockers) notes.push(`detected ${blockerSpans.length} blocking reviewer signal(s)`)\n\n    const positiveGroundingSignals =\n      patchEvidence +\n      sandboxSpans.length +\n      llmSpans.filter((span) => looksRepoGrounded(span.output ?? '')).length\n    const driftSignals =\n      llmSpans.filter((span) => this.isDrift(span.output ?? '')).length +\n      trace.events.filter((event) => this.isDrift(JSON.stringify(event.payload))).length\n    const repoGroundedness =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : positiveGroundingSignals / (positiveGroundingSignals + driftSignals)\n    const driftPenalty =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : driftSignals / (positiveGroundingSignals + driftSignals)\n    if (driftSignals > 0) notes.push(`detected ${driftSignals} drift signal(s)`)\n\n    const costUsd = trace.budget.length\n      ? Math.max(\n          ...trace.budget\n            .filter((entry: BudgetLedgerEntry) => entry.dimension === 'usd')\n            .map((entry: BudgetLedgerEntry) => entry.consumed),\n          0,\n        )\n      : llmSpans.reduce((sum, span) => sum + (span.costUsd ?? 0), 0)\n    const wallSeconds =\n      trace.run.endedAt && trace.run.startedAt\n        ? Math.max(0, (trace.run.endedAt - trace.run.startedAt) / 1000)\n        : 0\n\n    return {\n      success,\n      goalProgress,\n      repoGroundedness,\n      driftPenalty,\n      toolUseQuality,\n      patchQuality,\n      testReality,\n      finalGate,\n      reviewerBlockers,\n      costUsd,\n      wallSeconds,\n      notes,\n    }\n  }\n\n  rank(score: RunScore): number {\n    return aggregateRunScore(score, this.weights)\n  }\n\n  private isDrift(text: string): boolean {\n    return this.driftPatterns.some((pattern) => pattern.test(text))\n  }\n}\n\nfunction normalizeJudgeScore(score: number): number {\n  return score > 1 ? clamp01(score / 10) : clamp01(score)\n}\n\nfunction looksRepoGrounded(text: string): boolean {\n  return /(?:src\\/|tests?\\/|package\\.json|tsconfig|\\.ts\\b|\\.tsx\\b|git status|pnpm |npm |vitest|pytest|jest)/i.test(\n    text,\n  )\n}\n\nfunction isBlockingJudge(span: Extract<Span, { kind: 'judge' }>): boolean {\n  return (\n    span.attributes?.blocking === true ||\n    span.attributes?.verdict === 'BLOCKING' ||\n    positiveNumber(span.attributes?.blockingFindings) ||\n    positiveNumber(span.attributes?.highFindings) ||\n    span.score <= 2\n  )\n}\n\nfunction positiveNumber(value: unknown): boolean {\n  return typeof value === 'number' && value > 0\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport { CostLedger, type CostLedgerHandle, type CostReceipt } from './cost-ledger'\nimport {\n  callLlmJson,\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  type LlmCallRequest,\n  type LlmClientOptions,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Provider-enforced output limit. Default 16000. */\n  maxTokens?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** LlmClient config (baseUrl, apiKey, authHeader, …). */\n  llm?: LlmClientOptions\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MAX_TOKENS = 16_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: Required<SemanticConceptJudgeOptions>,\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions = {},\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts: Required<SemanticConceptJudgeOptions> = {\n    model: options.model ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    llm: options.llm ?? {},\n    costLedger: options.costLedger ?? new CostLedger(),\n    costPhase: options.costPhase ?? 'judge.semantic-concept',\n    costTags: options.costTags ?? {},\n    signal: options.signal ?? new AbortController().signal,\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  let receipt: CostReceipt | undefined\n  try {\n    const request = {\n      model: opts.model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n        },\n        { role: 'user' as const, content: buildPrompt(input, opts) },\n      ],\n      jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n      temperature: 0,\n      maxTokens: opts.maxTokens,\n      timeoutMs: opts.timeoutMs,\n    } satisfies LlmCallRequest\n    const paid = await opts.costLedger.runPaidCall({\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      model: opts.model,\n      ...(Object.keys(opts.costTags).length > 0 ? { tags: opts.costTags } : {}),\n      maximumCharge: maximumChargeForLlmRequest(request, opts.llm),\n      signal: opts.signal,\n      execute: (signal, callId) =>\n        callLlmJson<{ summary: string; concepts: ConceptFinding[] }>(request, {\n          ...opts.llm,\n          signal,\n          idempotencyKey: callId,\n        }),\n      receipt: ({ result }) => costReceiptFromLlm(result),\n      receiptFromError: costReceiptFromLlmError,\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid.value\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: paid.receipt.costUnknown ? null : paid.receipt.costUsd,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: receipt && !receipt.costUnknown ? receipt.costUsd : null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions = {},\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAeA,MAAM,0BAAU,IAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;CACrC,IAAI,IAAI,QAAQ,IAAI,IAAI;CACxB,IAAI,CAAC,GAAG;EACN,IAAI,IAAI,MAAM;EACd,QAAQ,IAAI,MAAM,CAAC;CACrB;CACA,OAAO;AACT;AAEA,IAAa,sBAAb,MAAiC;CAEH;CAD5B;CACA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,QAAQ,SAAS,IAAI;EAC1B,IAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAC3B,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAEhD;CAEA,MAAM,OAAO,OAA+B;EAC1C,MAAM,OAAO,GAAG,KAAK,UAAU,KAAK,EAAE;EACtC,MAAM,KAAK,MAAM,mBAAmB;GAClC,eAAe,KAAK,MAAM,IAAI;EAChC,CAAC;CACH;AACF;;;;;;;;;;;;;;;;;;;;;;ACPA,IAAa,gBAAb,MAA2B;CAGG;CAF5B;CAEA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,WAAW,IAAI,oBAAoB,IAAI;CAC9C;CAEA,MAAM,OAAO,OAAe,UAA2C;EACrE,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,MAAwB;IAAE,GAAG;IAAG,QAAQ;GAAM;GACpD,MAAM,KAAK,SAAS,OAAO,GAAG;EAChC;CACF;;CAGA,UAA8B;EAC5B,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,MAAM,aAAa,KAAK,MAAM,MAAM;EAC1C,IAAI,CAAC,KAAK,OAAO,CAAC;EAClB,MAAM,MAA0B,CAAC;EACjC,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;GAClC,IAAI,CAAC,MAAM;GACX,IAAI;IACF,IAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;GAC/C,QAAQ,CAGR;EACF;EACA,OAAO;CACT;;CAGA,QAAQ,OAAmC;EACzC,OAAO,KAAK,QAAQ,CAAC,CAAC,QAAQ,MAAM,EAAE,WAAW,KAAK;CACxD;AACF;;;;;AAiCA,SAAgB,kBAAkB,GAAmB,GAA4B;CAC/E,IAAI,EAAE,aAAa,EAAE,UAAU,OAAO;CACtC,IAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,OAAO;CACvE,IAAI,EAAE,cAAc,WAAW,EAAE,cAAc,QAAQ,OAAO;CAC9D,OAAO;AACT;;;;;AAMA,SAAgB,aACd,UACA,SACA,SAAqB,CAAC,GACR;CACd,MAAM,aAAa,OAAO,cAAc;CACxC,MAAM,WAAW,IAAI,IAAI,SAAS,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAC/D,MAAM,UAAU,IAAI,IAAI,QAAQ,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAE7D,MAAM,WAA+B,CAAC;CACtC,MAAM,cAAkC,CAAC;CACzC,MAAM,YAAgC,CAAC;CACvC,MAAM,UAAmC,CAAC;CAE1C,KAAK,MAAM,CAAC,IAAI,QAAQ,SAAS;EAC/B,MAAM,OAAO,SAAS,IAAI,EAAE;EAC5B,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,GAAG;GACjB;EACF;EACA,IAAI,WAAW,MAAM,GAAG,GACtB,QAAQ,KAAK;GAAE,UAAU;GAAM,SAAS;EAAI,CAAC;OAE7C,UAAU,KAAK,GAAG;CAEtB;CACA,KAAK,MAAM,CAAC,IAAI,SAAS,UACvB,IAAI,CAAC,QAAQ,IAAI,EAAE,GAAG,YAAY,KAAK,IAAI;CAE7C,OAAO;EAAE;EAAU;EAAa;EAAW;CAAQ;AACrD;;;;;;;;;;;;;;;;;;;;;;;;;ACjEA,MAAM,uBAAuB;AAE7B,MAAM,oBACJ;AACF,MAAM,aAAa;AAInB,SAAS,cAAc,MAAgD;CACrE,IAAI,CAAC,WAAW,IAAI,GAAG,OAAO,CAAC;CAC/B,MAAM,MAAwC,CAAC;CAC/C,KAAK,MAAM,SAAS,YAAY,MAAM,EAAE,eAAe,KAAK,CAAC,GAAG;EAC9D,IAAI,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,eAAe,GAAG;EACrD,MAAM,UAAU,KAAK,MAAM,MAAM,MAAM,UAAU;EACjD,IAAI,WAAW,OAAO,GAAG,IAAI,KAAK;GAAE,MAAM,MAAM;GAAM,MAAM;EAAQ,CAAC;CACvE;CACA,OAAO;AACT;AAEA,SAAS,UAAU,KAAa,KAAuB;CACrD,IAAI,CAAC,WAAW,GAAG,GAAG,OAAO,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,QAAQ,CAAC,GAAG;CAClB,OAAO,MAAM,QAAQ;EACnB,MAAM,MAAM,MAAM,IAAI;EACtB,IAAI;EACJ,IAAI;GACF,UAAU,YAAY,KAAK,EAAE,eAAe,KAAK,CAAC;EACpD,QAAQ;GACN;EACF;EACA,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,OAAO,KAAK,KAAK,EAAE,IAAI;GAC7B,IAAI,EAAE,YAAY,GAAG,MAAM,KAAK,IAAI;QAC/B,IAAI,EAAE,KAAK,SAAS,QAAQ,GAAG;IAClC,MAAM,KAAK,IAAI;IACf,IAAI,MAAM,KAAK,MAAM,UAAU,KAAK,OAAO;GAC7C;EACF;CACF;CACA,OAAO;AACT;AAEA,SAAS,uBAAuB,MAAsB;CAEpD,MAAM,QADK,wBAAwB,KAAK,IACzB,CAAC,GAAG,MAAM;CAEzB,OADU,uBAAuB,KAAK,KAC/B,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,eAAe,OAAiB,MAAc,SAA2B;CAChF,IAAI,IAAI;CACR,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,aAAa,CAAC,KAAK,MAAM,WAAW,IAAI,GAAG,GAAG,QAAQ,KAAK,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC;EACrF,KAAK,MAAM,OAAO,YAAY;GAC5B,IAAI,CAAC,WAAW,GAAG,GAAG;GACtB,IAAI;IACF,IAAI,SAAS,GAAG,CAAC,CAAC,YAAY,GAAG,KAAK,YAAY,GAAG,CAAC,CAAC;SAClD,KAAK;GACZ,QAAQ,CAER;EACF;CACF;CACA,OAAO;AACT;;AAGA,SAAgB,sBAAsB,QAAgD;CACpF,MAAM,SAAS,OAAO,WAAW,SAAS,EAAE,MAAM,WAChD,cAAc,IAAI,CAAC,CAAC,KAAK,OAAO;EAAE,GAAG;EAAG;CAAK,EAAE,CACjD;CACA,MAAM,QAAQ,OAAO,KAAK,MAAM,EAAE,IAAI;CAGtC,MAAM,SAAS,IAAI,IAAoB,MAAM,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;CAC/D,MAAM,QAAQ,IAAI,IAAoB,MAAM,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;CAC9D,MAAM,UAAU;CAChB,MAAM,QAAQ;CACd,IAAI,cAAc;CAClB,KAAK,MAAM,OAAO,OAAO,gBACvB,KAAK,MAAM,QAAQ,UAAU,KAAK,OAAO,wBAAwB,CAAC,GAAG;EACnE,eAAe;EACf,IAAI;EACJ,IAAI;GACF,OAAO,aAAa,MAAM,MAAM;EAClC,QAAQ;GACN;EACF;EACA,KAAK,MAAM,KAAK,KAAK,SAAS,OAAO,GAAG;GACtC,MAAM,IAAI,EAAE;GACZ,IAAI,CAAC,GAAG;GACR,MAAM,IAAI,EAAE,MAAM,GAAG,CAAC,CAAC,IAAI,KAAK;GAChC,MAAM,OAAO,OAAO,IAAI,CAAC;GACzB,IAAI,SAAS,KAAA,GAAW,OAAO,IAAI,GAAG,OAAO,CAAC;EAChD;EACA,KAAK,MAAM,KAAK,KAAK,SAAS,KAAK,GAAG;GACpC,MAAM,IAAI,EAAE;GACZ,IAAI,MAAM,KAAA,GAAW;GACrB,MAAM,OAAO,MAAM,IAAI,CAAC;GACxB,IAAI,SAAS,KAAA,GAAW,MAAM,IAAI,GAAG,OAAO,CAAC;EAC/C;CACF;CAIF,MAAM,yBAAS,IAAI,IAAoB;CACvC,KAAK,MAAM,KAAK,QACd,IAAI;EACF,OAAO,IAAI,EAAE,MAAM,aAAa,EAAE,MAAM,MAAM,CAAC;CACjD,QAAQ;EACN,OAAO,IAAI,EAAE,MAAM,EAAE;CACvB;CAEF,MAAM,UAAU,IAAI,IAAoB,MAAM,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;CAChE,KAAK,MAAM,UAAU,OAAO;EAC1B,MAAM,MAAM,IAAI,OAAO,IAAI,OAAO,YAAY,OAAO,OAAO;EAC5D,KAAK,MAAM,KAAK,QAAQ;GACtB,IAAI,EAAE,SAAS,QAAQ;GACvB,IAAI,IAAI,KAAK,OAAO,IAAI,EAAE,IAAI,KAAK,EAAE,GAAG,QAAQ,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAK,CAAC;EACtF;CACF;CAEA,MAAM,UAA8B,OAAO,KAAK,MAAM;EACpD,MAAM,OAAO,OAAO,IAAI,EAAE,IAAI,KAAK;EACnC,MAAM,MAAM,EAAE,KAAK,QAAQ,gBAAgB,EAAE;EAC7C,OAAO;GACL,MAAM,EAAE;GACR,MAAM,EAAE;GACR,MAAM,EAAE;GACR,OAAO,OAAO,KAAK,MAAM,IAAI,CAAC,CAAC,SAAS;GACxC,mBAAmB,OAAO,IAAI,EAAE,IAAI,KAAK;GACzC,kBAAkB,MAAM,IAAI,EAAE,IAAI,KAAK;GACvC,aAAa,QAAQ,IAAI,EAAE,IAAI,KAAK;GACpC,eAAe,eACb,OAAO,iBAAiB,CAAC,GACzB,EAAE,MACF,OAAO,kBAAkB,EAAE,SAAS,CAAC,CACvC;GACA,oBAAoB,KAAK,MAAM,iBAAiB,KAAK,CAAC,EAAA,CAAG;GACzD,kBAAkB,WAAW,KAAK,KAAK,YAAY,CAAC;GACpD,aAAa,WAAW,KAAK,KAAK,OAAO,CAAC;GAC1C,UAAU,KAAK,SAAS,kBAAkB;GAC1C,mBAAmB,WAAW,KAAK,uBAAuB,IAAI,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;EACvF;CACF,CAAC;CACD,OAAO;EAAE,qBAAqB;EAAa;CAAQ;AACrD;AAIA,MAAM,aAAa;AAEnB,SAAS,QACP,MACA,SACA,OACA,UACA,YACA,YACA,aACA,aACA,WACgB;CAChB,OAAO;EACL,gBAAgB;EAChB,YAAY,iBAAiB;GAAE,YAAY;GAAY;GAAM;GAAS;EAAM,CAAC;EAC7E,YAAY;EACZ,aAAa;EACb;EACA;EACA;EACA;EACA,eAAe,CAAC;GAAE,MAAM;GAAY,KAAK;EAAY,CAAC;EACtD,oBAAoB;EACpB;EACA;CACF;AACF;;AAGA,SAAgB,uBACd,QACA,YACkB;CAClB,MAAM,MAAwB,CAAC;CAC/B,KAAK,MAAM,KAAK,OAAO,SAAS;EAC9B,MAAM,cAAc,EAAE,oBAAoB,EAAE;EAI5C,IAHkB,cAAc,EAAE,cAAc,EAAE,kBAGhC,GAChB,IAAI,KACF,QACE,eACA,EAAE,MACF,UAAU,EAAE,KAAK,+EACjB,QACA,IACA,YACA,sIACA,EAAE,MACF,iGACF,CACF;OACK,IAAI,gBAAgB,KAAK,EAAE,cAAc,EAAE,gBAAgB,GAEhE,IAAI,KACF,QACE,eACA,EAAE,MACF,UAAU,EAAE,KAAK,gFAAgF,EAAE,YAAY,cAAc,EAAE,cAAc,IAC7I,QACA,IACA,YACA,2JACA,EAAE,MACF,sEACF,CACF;EAIF,IAAI,eAAe,KAAK,CAAC,EAAE,mBACzB,IAAI,KACF,QACE,mBACA,EAAE,MACF,UAAU,EAAE,KAAK,mFACjB,UACA,IACA,YACA,oHACA,EAAE,IACJ,CACF;EAIF,IAAI,EAAE,SAAS,YAAY,EAAE,oBAAoB,GAC/C,IAAI,KACF,QACE,UACA,EAAE,MACF,iBAAiB,EAAE,KAAK,YAAY,EAAE,kBAAkB,+BACxD,QACA,KACA,YACA,uMACA,EAAE,IACJ,CACF;EAIF,IAAI,EAAE,QAAQ,wBAAwB,CAAC,EAAE,kBACvC,IAAI,KACF,QACE,mBACA,EAAE,MACF,UAAU,EAAE,KAAK,OAAO,EAAE,MAAM,4DAChC,UACA,IACA,YACA,mFAAmF,EAAE,MAAM,mDAC3F,EAAE,IACJ,CACF;EAIF,IAAI,CAAC,EAAE,aACL,IAAI,KACF,QACE,gBACA,EAAE,MACF,UAAU,EAAE,KAAK,oBACjB,OACA,IACA,YACA,wGACA,EAAE,IACJ,CACF;EAIF,IAAI,CAAC,EAAE,UACL,IAAI,KACF,QACE,iBACA,EAAE,MACF,UAAU,EAAE,KAAK,8CACjB,OACA,KACA,YACA,iJACA,EAAE,IACJ,CACF;CAEJ;CACA,OAAO;AACT;AAIA,IAAa,oBAAb,MAAoE;CAClE,KAAc;CACd,cACE;CACF,YAAqB;CACrB,OAAgB;EAAE,MAAM;EAA0B,iBAAiB;CAAE;CACrE,UAAmB;CAEnB,MAAM,QAAQ,OAAyB,KAAgD;EACrF,MAAM,aAAa,IAAI,MAAM,+BAAc,IAAI,KAAK,EAAA,CAAE,YAAY;EAClE,IAAI,MACF,gBAAgB,MAAM,QAAQ,OAAO,eAAe,MAAM,oBAAoB,aAChF;EACA,OAAO,uBAAuB,OAAO,UAAU;CACjD;AACF;AAEA,MAAa,sBAAsB,IAAI,kBAAkB;;;AClYzD,MAAM,yBAAyB;CAC7B;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,YAAb,MAAuB;CACrB;CACA;CAEA,YAAY,UAA4B,CAAC,GAAG;EAC1C,KAAK,UAAU,QAAQ;EACvB,KAAK,gBAAgB,QAAQ,iBAAiB;CAChD;CAEA,MAAM,MAAM,OAAmB,OAAkC;EAC/D,MAAM,MAAM,MAAM,MAAM,OAAO,KAAK;EACpC,IAAI,CAAC,KAAK,MAAM,IAAI,cAAc,OAAO,MAAM,WAAW;EAC1D,MAAM,CAAC,OAAO,QAAQ,WAAW,UAAU,MAAM,QAAQ,IAAI;GAC3D,MAAM,MAAM,EAAE,MAAM,CAAC;GACrB,MAAM,OAAO,EAAE,MAAM,CAAC;GACtB,MAAM,UAAU,KAAK;GACrB,MAAM,OAAO,KAAK;EACpB,CAAC;EACD,OAAO,KAAK,WAAW;GAAE;GAAK;GAAO;GAAQ;GAAW;EAAO,CAAC;CAClE;CAEA,WAAW,OAA2B;EACpC,MAAM,QAAkB,CAAC;EACzB,MAAM,WAAW,MAAM,MAAM,QAC1B,MAA2C,EAAE,SAAS,KACzD;EACA,MAAM,YAAY,MAAM,MAAM,QAC3B,MAA4C,EAAE,SAAS,MAC1D;EACA,MAAM,aAAa,MAAM,MAAM,QAC5B,MAA6C,EAAE,SAAS,OAC3D;EACA,MAAM,eAAe,MAAM,MAAM,QAC9B,MAA+C,EAAE,SAAS,SAC7D;EACA,MAAM,iBAAiB,WAAW,QAC/B,SAAS,KAAK,cAAc,gBAAgB,KAAK,YAAY,cAAc,IAC9E;EAEA,MAAM,UACJ,MAAM,IAAI,SAAS,SAAS,OAAO,IAAI,MAAM,IAAI,WAAW,cAAc,KAAM;EAClF,IAAI,CAAC,SAAS,MAAM,KAAK,qCAAqC;EAE9D,MAAM,eAAe,WAAW,SAC5B,WAAW,QAAQ,KAAK,SAAS,MAAM,oBAAoB,KAAK,KAAK,GAAG,CAAC,IACzE,WAAW,SACX,KAAA;EAOJ,MAAM,gBALJ,OAAO,MAAM,IAAI,SAAS,UAAU,WAChC,QACE,MAAM,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAI,QAAQ,QAAQ,MAAM,MAAM,IAAI,QAAQ,KAClF,IACA,KAAA,MAC+B,gBAAgB;EAErD,MAAM,kBAAkB,UAAU,QAAQ,SAAS,KAAK,WAAW,OAAO,CAAC,CAAC;EAC5E,MAAM,iBAAiB,UAAU,WAAW,IAAI,IAAI,kBAAkB,UAAU;EAChF,IAAI,UAAU,WAAW,GAAG,MAAM,KAAK,wBAAwB;EAE/D,MAAM,gBACJ,MAAM,UAAU,SAChB,UAAU,QAAQ,SAAS,0BAA0B,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC;EAC5E,MAAM,eAAe,gBAAgB,IAAI,QAAQ,gBAAgB,CAAC,IAAI;EACtE,IAAI,CAAC,cAAc,MAAM,KAAK,uCAAuC;EAErE,MAAM,eAAe,aAAa,QAC/B,SAAS,OAAO,KAAK,eAAe,YAAY,KAAK,aAAa,CACrE;EACA,MAAM,cAAc,aAAa,SAC7B,aAAa,QACV,KAAK,SAAS,OAAO,KAAK,eAAe,KAAK,KAAK,IAAI,GAAG,KAAK,cAAc,CAAC,GAC/E,CACF,IAAI,aAAa,SACjB,UAAU,MAAM,SACZ,yCAAyC,KAAK,KAAK,UAAU,KAAK,IAAI,CAAC,CACzE,IACA,KACA;EACN,IAAI,CAAC,aAAa,MAAM,KAAK,sCAAsC;EAEnE,MAAM,eAAe,WAAW,QAAQ,SAAS,gBAAgB,IAAI,CAAC;EACtE,MAAM,oBAAoB,eAAe,QAAQ,SAAS,gBAAgB,IAAI,CAAC;EAC/E,MAAM,YAAY,eAAe,SAAU,kBAAkB,SAAS,IAAI,IAAK;EAC/E,IAAI,kBAAkB,QACpB,MAAM,KAAK,yBAAyB,kBAAkB,OAAO,aAAa;OACvE,IAAI,CAAC,eAAe,QAAQ,MAAM,KAAK,iCAAiC;EAE7E,MAAM,mBAAmB,WAAW,SAAS,aAAa,SAAS,WAAW,SAAS;EACvF,IAAI,kBAAkB,MAAM,KAAK,YAAY,aAAa,OAAO,6BAA6B;EAE9F,MAAM,2BACJ,gBACA,aAAa,SACb,SAAS,QAAQ,SAAS,kBAAkB,KAAK,UAAU,EAAE,CAAC,CAAC,CAAC;EAClE,MAAM,eACJ,SAAS,QAAQ,SAAS,KAAK,QAAQ,KAAK,UAAU,EAAE,CAAC,CAAC,CAAC,SAC3D,MAAM,OAAO,QAAQ,UAAU,KAAK,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,CAAC,CAAC;EAC9E,MAAM,mBACJ,2BAA2B,iBAAiB,IACxC,IACA,4BAA4B,2BAA2B;EAC7D,MAAM,eACJ,2BAA2B,iBAAiB,IACxC,IACA,gBAAgB,2BAA2B;EACjD,IAAI,eAAe,GAAG,MAAM,KAAK,YAAY,aAAa,iBAAiB;EAe3E,OAAO;GACL;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA,SAvBc,MAAM,OAAO,SACzB,KAAK,IACH,GAAG,MAAM,OACN,QAAQ,UAA6B,MAAM,cAAc,KAAK,CAAC,CAC/D,KAAK,UAA6B,MAAM,QAAQ,GACnD,CACF,IACA,SAAS,QAAQ,KAAK,SAAS,OAAO,KAAK,WAAW,IAAI,CAAC;GAiB7D,aAfA,MAAM,IAAI,WAAW,MAAM,IAAI,YAC3B,KAAK,IAAI,IAAI,MAAM,IAAI,UAAU,MAAM,IAAI,aAAa,GAAI,IAC5D;GAcJ;EACF;CACF;CAEA,KAAK,OAAyB;EAC5B,OAAO,kBAAkB,OAAO,KAAK,OAAO;CAC9C;CAEA,QAAgB,MAAuB;EACrC,OAAO,KAAK,cAAc,MAAM,YAAY,QAAQ,KAAK,IAAI,CAAC;CAChE;AACF;AAEA,SAAS,oBAAoB,OAAuB;CAClD,OAAO,QAAQ,IAAI,QAAQ,QAAQ,EAAE,IAAI,QAAQ,KAAK;AACxD;AAEA,SAAS,kBAAkB,MAAuB;CAChD,OAAO,qGAAqG,KAC1G,IACF;AACF;AAEA,SAAS,gBAAgB,MAAiD;CACxE,OACE,KAAK,YAAY,aAAa,QAC9B,KAAK,YAAY,YAAY,cAC7B,eAAe,KAAK,YAAY,gBAAgB,KAChD,eAAe,KAAK,YAAY,YAAY,KAC5C,KAAK,SAAS;AAElB;AAEA,SAAS,eAAe,OAAyB;CAC/C,OAAO,OAAO,UAAU,YAAY,QAAQ;AAC9C;;;;;;;;;;;;;;;;;;;;;;;AC/EA,MAAa,6BAAgE;CAC3E,QAAQ;CACR,WAAW;CACX,SAAS;AACX;AAiCA,MAAa,iCAAiC;AAE9C,MAAM,qBAAqB;AAC3B,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,kBAAkB;AACxB,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;AAEtB,MAAM,kBAAkB;CACtB,MAAM;CACN,sBAAsB;CACtB,UAAU,CAAC,WAAW,UAAU;CAChC,YAAY;EACV,SAAS;GAAE,MAAM;GAAU,WAAW;GAAI,WAAW;EAAI;EACzD,UAAU;GACR,MAAM;GACN,UAAU;GACV,OAAO;IACL,MAAM;IACN,sBAAsB;IACtB,UAAU;KAAC;KAAW;KAAW;KAAS;KAAY;IAAU;IAChE,YAAY;KACV,SAAS;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACxD,SAAS,EAAE,MAAM,UAAU;KAC3B,OAAO;MAAE,MAAM;MAAU,SAAS;MAAG,SAAS;KAAG;KACjD,UAAU;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACzD,UAAU;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAS;OAAS;MAAM;KAAE;IAC3E;GACF;EACF;CACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;CAClE,IAAI,KAAK,UAAU,KAAK,OAAO;CAC/B,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,iBAAiB,KAAK,SAAS,IAAI,YAAY,MAAM;AACpF;AAEA,SAAS,YACP,OACA,MACQ;CACR,MAAM,aAAa,MAAM,YACtB,QAAQ,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,CAAC,CACvD,KAAK,MAAM,aAAa,EAAE,KAAK,QAAQ,EAAE,SAAS,CAAC,CACnD,KAAK,MAAM;CAEd,MAAM,OAAO,MAAM,cAAc;CAEjC,OAAO;;;;;;;;;;EAUP,MAAM,YAAY;;EAElB,MAAM,gBAAgB,+BAA+B,MAAM,cAAc,mBAAmB,MAAM,uBAAuB,GAAG,QAAQ,GAAG;EACvI,MAAM,iBACL,KACE,GAAG,MACF,KAAK,IAAI,EAAE,KAAK,EAAE,KAAK,GAAG,EAAE,UAAU,SAAS,cAAc,EAAE,SAAS,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,EAAE,KAAK,IACzG,CAAC,CACA,KAAK,IAAI,EAAE;;EAEZ,OAAO,qDAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,EAAE,QAAQ,GAAG;EACjH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,EAAE;;;;;;;;;;;;;;;;;;AAkBtD;;;;;;AASA,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACH;CACrC,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,aAAa,MAAM,iBAAiB;CAE1C,IAAI,eAAe,GACjB,OAAO;EACL,MAAM;EACN,SAAS;EACT,OAAO;EACP,cAAc;EACd,YAAY;EACZ,UAAU,CAAC;EACX,SAAS;EACT,YAAY;EACZ,SAAS;EACT,WAAW;EACX,OAAO;CACT;CAGF,MAAM,OAA8C;EAClD,OAAO,QAAQ,SAAS;EACxB,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,KAAK,QAAQ,OAAO,CAAC;EACrB,YAAY,QAAQ,cAAc,IAAI,WAAW;EACjD,WAAW,QAAQ,aAAa;EAChC,UAAU,QAAQ,YAAY,CAAC;EAC/B,QAAQ,QAAQ,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAChD,gBAAgB,QAAQ,kBAAkB;EAC1C,mBAAmB;GAAE,GAAG;GAA4B,GAAI,QAAQ,qBAAqB,CAAC;EAAG;CAC3F;CAMA,MAAM,oBAAoB,SAA8B;EACtD,IAAI,KAAK,mBAAmB,QAAQ,OAAO;EAC3C,IAAI,KAAK,UAAU,MAAM,OAAO,KAAK;EACrC,IAAI,KAAK,mBAAmB,cAC1B,OAAO,KAAK,kBAAkB,KAAK,cAAc,aAAa;EAEhE,OAAO;CACT;CACA,MAAM,eAAe,IAAI,IACvB,MAAM,iBAAiB,KAAK,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC,CACjE;CAEA,IAAI;CACJ,IAAI;EACF,MAAM,UAAU;GACd,OAAO,KAAK;GACZ,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IAAE,MAAM;IAAiB,SAAS,YAAY,OAAO,IAAI;GAAE,CAC7D;GACA,YAAY;IAAE,MAAM;IAA0B,QAAQ;GAAgB;GACtE,aAAa;GACb,WAAW,KAAK;GAChB,WAAW,KAAK;EAClB;EACA,MAAM,OAAO,MAAM,KAAK,WAAW,YAAY;GAC7C,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,OAAO,KAAK;GACZ,GAAI,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,SAAS,IAAI,EAAE,MAAM,KAAK,SAAS,IAAI,CAAC;GACvE,eAAe,2BAA2B,SAAS,KAAK,GAAG;GAC3D,QAAQ,KAAK;GACb,UAAU,QAAQ,WAChB,YAA6D,SAAS;IACpE,GAAG,KAAK;IACR;IACA,gBAAgB;GAClB,CAAC;GACH,UAAU,EAAE,aAAa,mBAAmB,MAAM;GAClD,kBAAkB;EACpB,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU,KAAK;EAEvB,IAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GACnD,MAAM,IAAI,MAAM,uEAAqE;EAGvF,MAAM,WAA6B,MAAM,SAAS,KAAK,OAAO;GAC5D,SAAS,OAAO,EAAE,OAAO;GACzB,SAAS,QAAQ,EAAE,OAAO;GAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;GACrD,UAAU,OAAO,EAAE,YAAY,EAAE;GACjC,UAAW;IAAC;IAAY;IAAS;IAAS;GAAM,CAAC,CAAW,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;EACN,EAAE;EAEF,MAAM,eAAe,SAAS,QAAQ,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;EACvE,IAAI,YAAY;EAChB,IAAI,mBAAmB;EACvB,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;GACzC,aAAa;GACb,oBAAoB,IAAI,EAAE;EAC5B;EACA,MAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;EAE7E,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO,QAAQ,WAAW,GAAA,CAAI,QAAQ,CAAC,CAAC;GACxC;GACA;GACA;GACA,SAAS,OAAO,MAAM,WAAW,EAAE;GACnC,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,KAAK,QAAQ,cAAc,OAAO,KAAK,QAAQ;GACxD,WAAW;EACb;CACF,SAAS,KAAK;EACZ,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO;GACP,cAAc;GACd;GACA,UAAU,CAAC;GACX,SAAS;GACT,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,WAAW,CAAC,QAAQ,cAAc,QAAQ,UAAU;GAC7D,WAAW;GACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;EACxD;CACF;AACF;;;;;AAMA,SAAgB,2BACd,UAAuC,CAAC,GACmC;CAC3E,QAAQ,UAAU,wBAAwB,OAAO,OAAO;AAC1D"}
+{"version":3,"file":"semantic-concept-judge-Bz64IckK.js","names":[],"sources":["../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/skill-usage.ts","../src/run-critic.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Skill-usage analyst — a DETERMINISTIC `Analyst` over a Claude/Codex skill\n * library + its trace corpus. Unlike the trace-store kinds (failure-mode,\n * improvement, ...) this kind calls no LLM: it mines real usage and skill\n * structure and emits findings by rule.\n *\n * It exists because the naive \"Skill-tool invocation count\" lies low — it\n * misses orchestrated sub-dispatch (a leaf skill run BY /pursue or /governor\n * logs under the parent), slash-command entry, local-script bypass, and\n * on-disk artifacts. The 2026-05-30 skill audit found 39/53 skills at zero\n * direct invocations, yet only one was a genuine cut: the rest were\n * measurement-invisible or discovery-limited. This analyst encodes that\n * lesson as a multi-signal usage model so a cheap repeatable pass can keep\n * the library honest, and so the expensive audit workflow's verdicts can\n * GEPA-distill it toward agreement (see `gold/skill-verdicts.gold.jsonl`).\n *\n * Report-building (`buildSkillUsageReport`, an fs scan) is separated from\n * finding emission (`SkillUsageAnalyst.analyze`, pure) so the slow scan runs\n * once at the registry boundary and the rule logic stays unit-testable.\n */\n\nimport { type Dirent, existsSync, readdirSync, readFileSync, statSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { Analyst, AnalystContext, AnalystFinding, AnalystSeverity } from '../types'\nimport { computeFindingId } from '../types'\n\n// ── Input model ──────────────────────────────────────────────────────\n\nexport type SkillKind = 'public' | 'private'\n\n/** One skill's multi-signal usage + structure. All counts are deterministic. */\nexport interface SkillUsageRecord {\n  name: string\n  kind: SkillKind\n  /** Absolute path to the skill's SKILL.md. */\n  path: string\n  lines: number\n  /** `\"skill\":\"<name>\"` Skill-tool invocations across the trace corpus. */\n  directInvocations: number\n  /** `<command-name>/<name>` slash invocations across the trace corpus. */\n  slashInvocations: number\n  /** Sibling skills whose SKILL.md dispatches to this one (`/<name>`). Proxy\n   *  for orchestrated sub-dispatch the per-skill counter cannot see. */\n  inboundRefs: number\n  /** On-disk artifacts attributable to the skill (e.g. `.evolve/<name>/**`). */\n  artifactCount: number\n  /** Tangle-private reference count in the body (leak signal for public skills). */\n  tanglePrivateRefs: number\n  hasReferencesDir: boolean\n  hasEvalsDir: boolean\n  /** Body mentions `skill-runs.jsonl` (visible to /reflect + /governor). */\n  logsRuns: boolean\n  /** Description carries an explicit `Triggers:` clause / trigger phrases. */\n  hasTriggerPhrases: boolean\n}\n\nexport interface SkillUsageReport {\n  generatedFromTraces: number\n  records: SkillUsageRecord[]\n}\n\nexport interface SkillUsageScanConfig {\n  /** Dirs holding `*.jsonl` transcripts (Claude `~/.claude/projects`, Codex sessions). */\n  transcriptDirs: string[]\n  /** Skill roots to scan; each dir directly under `root` with a `SKILL.md` is a skill. */\n  skillRoots: { root: string; kind: SkillKind }[]\n  /** Roots scanned for `<root>/.evolve/<skill>` artifact dirs. */\n  artifactRoots?: string[]\n  /** Token-prefixed mappings: skill name → extra artifact subpaths under an artifactRoot\n   *  (e.g. reflect → `.evolve/reflections`). Catches non-eponymous artifact dirs. */\n  artifactAliases?: Record<string, string[]>\n  /** Cap files read per transcript dir (bounds a huge corpus); 0 = unbounded. */\n  maxTranscriptsPerDir?: number\n}\n\n// ── Deterministic thresholds ─────────────────────────────────────────\n\n/** Anthropic's authoring guidance keeps SKILL.md short; past this with no\n *  `references/` split the body burns context budget every session. */\nconst BLOAT_LINE_THRESHOLD = 300\n\nconst TANGLE_PRIVATE_RE =\n  /\\b(cli-bridge|tangletools|ops-board|drew-gtr-pro|@tangle-network\\/|~\\/company|tangle\\.tools|gtm-agent)\\b|\\bkimi\\b|\\btcloud\\b/gi\nconst TRIGGER_RE = /triggers?\\s*[:-]/i\n\n// ── Report builder (fs scan — slow, runs once at the registry boundary) ──\n\nfunction listSkillDirs(root: string): { name: string; path: string }[] {\n  if (!existsSync(root)) return []\n  const out: { name: string; path: string }[] = []\n  for (const entry of readdirSync(root, { withFileTypes: true })) {\n    if (!entry.isDirectory() && !entry.isSymbolicLink()) continue\n    const skillMd = join(root, entry.name, 'SKILL.md')\n    if (existsSync(skillMd)) out.push({ name: entry.name, path: skillMd })\n  }\n  return out\n}\n\nfunction walkJsonl(dir: string, cap: number): string[] {\n  if (!existsSync(dir)) return []\n  const files: string[] = []\n  const stack = [dir]\n  while (stack.length) {\n    const cur = stack.pop()!\n    let entries: Dirent[]\n    try {\n      entries = readdirSync(cur, { withFileTypes: true })\n    } catch {\n      continue\n    }\n    for (const e of entries) {\n      const full = join(cur, e.name)\n      if (e.isDirectory()) stack.push(full)\n      else if (e.name.endsWith('.jsonl')) {\n        files.push(full)\n        if (cap > 0 && files.length >= cap) return files\n      }\n    }\n  }\n  return files\n}\n\nfunction frontmatterDescription(body: string): string {\n  const fm = /^---\\n([\\s\\S]*?)\\n---/.exec(body)\n  const block = fm?.[1] ?? ''\n  const m = /description:\\s*(.+)/i.exec(block)\n  return m?.[1] ?? ''\n}\n\nfunction countArtifacts(roots: string[], name: string, aliases: string[]): number {\n  let n = 0\n  for (const root of roots) {\n    const candidates = [join(root, '.evolve', name), ...aliases.map((a) => join(root, a))]\n    for (const dir of candidates) {\n      if (!existsSync(dir)) continue\n      try {\n        if (statSync(dir).isDirectory()) n += readdirSync(dir).length\n        else n += 1\n      } catch {\n        /* unreadable — skip */\n      }\n    }\n  }\n  return n\n}\n\n/** Scan the corpus + skill roots into a {@link SkillUsageReport}. Deterministic. */\nexport function buildSkillUsageReport(config: SkillUsageScanConfig): SkillUsageReport {\n  const skills = config.skillRoots.flatMap(({ root, kind }) =>\n    listSkillDirs(root).map((s) => ({ ...s, kind })),\n  )\n  const names = skills.map((s) => s.name)\n\n  // One pass over the corpus accumulating direct + slash counts per skill.\n  const direct = new Map<string, number>(names.map((n) => [n, 0]))\n  const slash = new Map<string, number>(names.map((n) => [n, 0]))\n  const skillRe = /\"skill\"\\s*:\\s*\"([a-z0-9_:-]+)\"/g\n  const cmdRe = /<command-name>\\/?([a-z0-9_:-]+)<\\/command-name>/g\n  let transcripts = 0\n  for (const dir of config.transcriptDirs) {\n    for (const file of walkJsonl(dir, config.maxTranscriptsPerDir ?? 0)) {\n      transcripts += 1\n      let data: string\n      try {\n        data = readFileSync(file, 'utf8')\n      } catch {\n        continue\n      }\n      for (const m of data.matchAll(skillRe)) {\n        const g = m[1]\n        if (!g) continue\n        const n = g.split(':').pop() ?? g\n        const prev = direct.get(n)\n        if (prev !== undefined) direct.set(n, prev + 1)\n      }\n      for (const m of data.matchAll(cmdRe)) {\n        const g = m[1]\n        if (g === undefined) continue\n        const prev = slash.get(g)\n        if (prev !== undefined) slash.set(g, prev + 1)\n      }\n    }\n  }\n\n  // Read each skill body once; compute structure + inbound refs across siblings.\n  const bodies = new Map<string, string>()\n  for (const s of skills) {\n    try {\n      bodies.set(s.name, readFileSync(s.path, 'utf8'))\n    } catch {\n      bodies.set(s.name, '')\n    }\n  }\n  const inbound = new Map<string, number>(names.map((n) => [n, 0]))\n  for (const target of names) {\n    const ref = new RegExp(`/${target}\\\\b|\\\\[\\\\[${target}\\\\]\\\\]`)\n    for (const s of skills) {\n      if (s.name === target) continue\n      if (ref.test(bodies.get(s.name) ?? '')) inbound.set(target, inbound.get(target)! + 1)\n    }\n  }\n\n  const records: SkillUsageRecord[] = skills.map((s) => {\n    const body = bodies.get(s.name) ?? ''\n    const dir = s.path.replace(/\\/SKILL\\.md$/, '')\n    return {\n      name: s.name,\n      kind: s.kind,\n      path: s.path,\n      lines: body ? body.split('\\n').length : 0,\n      directInvocations: direct.get(s.name) ?? 0,\n      slashInvocations: slash.get(s.name) ?? 0,\n      inboundRefs: inbound.get(s.name) ?? 0,\n      artifactCount: countArtifacts(\n        config.artifactRoots ?? [],\n        s.name,\n        config.artifactAliases?.[s.name] ?? [],\n      ),\n      tanglePrivateRefs: (body.match(TANGLE_PRIVATE_RE) ?? []).length,\n      hasReferencesDir: existsSync(join(dir, 'references')),\n      hasEvalsDir: existsSync(join(dir, 'evals')),\n      logsRuns: body.includes('skill-runs.jsonl'),\n      hasTriggerPhrases: TRIGGER_RE.test(frontmatterDescription(body) || body.slice(0, 600)),\n    }\n  })\n  return { generatedFromTraces: transcripts, records }\n}\n\n// ── Finding emission (pure — unit-testable, no LLM, no fs) ────────────\n\nconst ANALYST_ID = 'skill-usage'\n\nfunction finding(\n  area: string,\n  subject: string,\n  claim: string,\n  severity: AnalystSeverity,\n  confidence: number,\n  producedAt: string,\n  recommended: string,\n  evidenceUri: string,\n  rationale?: string,\n): AnalystFinding {\n  return {\n    schema_version: '1.0.0',\n    finding_id: computeFindingId({ analyst_id: ANALYST_ID, area, subject, claim }),\n    analyst_id: ANALYST_ID,\n    produced_at: producedAt,\n    severity,\n    area,\n    claim,\n    rationale,\n    evidence_refs: [{ kind: 'artifact', uri: evidenceUri }],\n    recommended_action: recommended,\n    confidence,\n    subject,\n  }\n}\n\n/** Pure rule pass over a report → findings. Exported for direct/unit use. */\nexport function emitSkillUsageFindings(\n  report: SkillUsageReport,\n  producedAt: string,\n): AnalystFinding[] {\n  const out: AnalystFinding[] = []\n  for (const r of report.records) {\n    const directTotal = r.directInvocations + r.slashInvocations\n    const trueUsage = directTotal + r.inboundRefs + r.artifactCount\n\n    // 1. Dead: no usage signal of ANY kind. The only real deprecation candidate.\n    if (trueUsage === 0) {\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' has zero usage across all signals (direct, slash, inbound-refs, artifacts)`,\n          'high',\n          0.6,\n          producedAt,\n          'Confirm the skill covers a real recurring job; if not, deprecate. Zero true usage is the only deterministic deprecation candidate.',\n          r.path,\n          'No Skill-tool call, no slash invocation, no sibling dispatches to it, and no on-disk artifacts.',\n        ),\n      )\n    } else if (directTotal === 0 && r.inboundRefs + r.artifactCount > 0) {\n      // 2. Measurement-invisible: real use via orchestration/artifacts, never invoked directly.\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' shows 0 direct invocations but is used via orchestration/artifacts (inbound=${r.inboundRefs}, artifacts=${r.artifactCount})`,\n          'info',\n          0.8,\n          producedAt,\n          'Do NOT treat as unused — usage is real but logged under parent skills or on disk. Strengthen direct-invocation discovery only if direct use is desired.',\n          r.path,\n          'The Skill-tool counter undercounts orchestrated/chained leaf skills.',\n        ),\n      )\n    }\n\n    // 3. Discovery gap: low direct use AND weak trigger surface.\n    if (directTotal <= 2 && !r.hasTriggerPhrases) {\n      out.push(\n        finding(\n          'discoverability',\n          r.name,\n          `Skill '${r.name}' is rarely invoked directly and its description has no explicit trigger phrases`,\n          'medium',\n          0.7,\n          producedAt,\n          'Add a `Triggers:` clause with verbatim user phrases to the frontmatter description so the model auto-invokes it.',\n          r.path,\n        ),\n      )\n    }\n\n    // 4. Public-repo leak.\n    if (r.kind === 'public' && r.tanglePrivateRefs > 0) {\n      out.push(\n        finding(\n          'safety',\n          r.name,\n          `Public skill '${r.name}' carries ${r.tanglePrivateRefs} Tangle-private reference(s)`,\n          'high',\n          0.75,\n          producedAt,\n          'Sanitize incidental internal refs (cli-bridge/kimi/tcloud/~company/private repos) or relocate to a private repo. Verify @tangle-network/* refs are to PUBLISHED packages before treating as a leak.',\n          r.path,\n        ),\n      )\n    }\n\n    // 5. Bloat / no progressive disclosure.\n    if (r.lines > BLOAT_LINE_THRESHOLD && !r.hasReferencesDir) {\n      out.push(\n        finding(\n          'maintainability',\n          r.name,\n          `Skill '${r.name}' is ${r.lines} lines with no references/ split (progressive disclosure)`,\n          'medium',\n          0.8,\n          producedAt,\n          `Split detail into references/ loaded on demand; keep SKILL.md a short overview. ${r.lines} lines load into every session's context budget.`,\n          r.path,\n        ),\n      )\n    }\n\n    // 6. No evals (Anthropic's \">=3 evals before docs\" rule).\n    if (!r.hasEvalsDir) {\n      out.push(\n        finding(\n          'data-quality',\n          r.name,\n          `Skill '${r.name}' ships no evals/`,\n          'low',\n          0.6,\n          producedAt,\n          'Add evals/evals.json with >=3 scenarios proving the skill beats baseline; gives regression coverage.',\n          r.path,\n        ),\n      )\n    }\n\n    // 7. No run logging → invisible to /reflect and /governor.\n    if (!r.logsRuns) {\n      out.push(\n        finding(\n          'observability',\n          r.name,\n          `Skill '${r.name}' never appends to .evolve/skill-runs.jsonl`,\n          'low',\n          0.55,\n          producedAt,\n          'Append one run line to .evolve/skill-runs.jsonl on completion, or declare it a non-logging leaf, so the self-improvement loop can see it ran.',\n          r.path,\n        ),\n      )\n    }\n  }\n  return out\n}\n\n// ── The Analyst ──────────────────────────────────────────────────────\n\nexport class SkillUsageAnalyst implements Analyst<SkillUsageReport> {\n  readonly id = ANALYST_ID\n  readonly description =\n    'Deterministic multi-signal skill-usage analysis: flags dead skills, measurement-invisible (orchestrated) usage, discovery gaps, public-repo leaks, bloat, missing evals, and missing run-logging.'\n  readonly inputKind = 'custom' as const\n  readonly cost = { kind: 'deterministic' as const, est_usd_per_run: 0 }\n  readonly version = '1.0.0'\n\n  async analyze(input: SkillUsageReport, ctx: AnalystContext): Promise<AnalystFinding[]> {\n    const producedAt = ctx.tags?.producedAt ?? new Date().toISOString()\n    ctx.log?.(\n      `skill-usage: ${input.records.length} skills over ${input.generatedFromTraces} transcripts`,\n    )\n    return emitSkillUsageFindings(input, producedAt)\n  }\n}\n\nexport const SKILL_USAGE_ANALYST = new SkillUsageAnalyst()\n","import { NotFoundError } from './errors'\nimport { aggregateRunScore, clamp01, type RunScore, type RunScoreWeights } from './run-score'\nimport type { Artifact, BudgetLedgerEntry, Run, Span, TraceEvent, TraceStore } from './trace'\n\nexport interface RunTrace {\n  run: Run\n  spans: Span[]\n  events: TraceEvent[]\n  artifacts: Artifact[]\n  budget: BudgetLedgerEntry[]\n}\n\nexport interface RunCriticOptions {\n  weights?: Partial<RunScoreWeights>\n  driftPatterns?: RegExp[]\n}\n\nconst DEFAULT_DRIFT_PATTERNS = [\n  /https?:\\/\\//i,\n  /\\btitle:\\s/i,\n  /\\bsummary:\\s/i,\n  /\\burl:\\s/i,\n  /\\bnpm package usage\\b/i,\n  /\\bnews\\b/i,\n]\n\nexport class RunCritic {\n  private readonly weights?: Partial<RunScoreWeights>\n  private readonly driftPatterns: RegExp[]\n\n  constructor(options: RunCriticOptions = {}) {\n    this.weights = options.weights\n    this.driftPatterns = options.driftPatterns ?? DEFAULT_DRIFT_PATTERNS\n  }\n\n  async score(store: TraceStore, runId: string): Promise<RunScore> {\n    const run = await store.getRun(runId)\n    if (!run) throw new NotFoundError(`run ${runId} not found`)\n    const [spans, events, artifacts, budget] = await Promise.all([\n      store.spans({ runId }),\n      store.events({ runId }),\n      store.artifacts(runId),\n      store.budget(runId),\n    ])\n    return this.scoreTrace({ run, spans, events, artifacts, budget })\n  }\n\n  scoreTrace(trace: RunTrace): RunScore {\n    const notes: string[] = []\n    const llmSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'llm' }> => s.kind === 'llm',\n    )\n    const toolSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'tool' }> => s.kind === 'tool',\n    )\n    const judgeSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'judge' }> => s.kind === 'judge',\n    )\n    const sandboxSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'sandbox' }> => s.kind === 'sandbox',\n    )\n    const finalGateSpans = judgeSpans.filter(\n      (span) => span.dimension === 'final_gate' || span.attributes?.finalGate === true,\n    )\n\n    const success =\n      trace.run.outcome?.pass === true ? 1 : trace.run.status === 'completed' ? 0.5 : 0\n    if (!success) notes.push('run did not complete with pass=true')\n\n    const judgeAverage = judgeSpans.length\n      ? judgeSpans.reduce((sum, span) => sum + normalizeJudgeScore(span.score), 0) /\n        judgeSpans.length\n      : undefined\n    const outcomeScore =\n      typeof trace.run.outcome?.score === 'number'\n        ? clamp01(\n            trace.run.outcome.score > 1 ? trace.run.outcome.score / 100 : trace.run.outcome.score,\n          )\n        : undefined\n    const goalProgress = outcomeScore ?? judgeAverage ?? success\n\n    const successfulTools = toolSpans.filter((span) => span.status !== 'error').length\n    const toolUseQuality = toolSpans.length === 0 ? 0 : successfulTools / toolSpans.length\n    if (toolSpans.length === 0) notes.push('no tool spans recorded')\n\n    const patchEvidence =\n      trace.artifacts.length +\n      toolSpans.filter((span) => /write|edit|patch|apply/i.test(span.toolName)).length\n    const patchQuality = patchEvidence > 0 ? clamp01(patchEvidence / 4) : 0\n    if (!patchQuality) notes.push('no artifact or edit evidence recorded')\n\n    const sandboxTests = sandboxSpans.filter(\n      (span) => typeof span.testsTotal === 'number' && span.testsTotal > 0,\n    )\n    const testReality = sandboxTests.length\n      ? sandboxTests.reduce(\n          (sum, span) => sum + (span.testsPassed ?? 0) / Math.max(1, span.testsTotal ?? 1),\n          0,\n        ) / sandboxTests.length\n      : toolSpans.some((span) =>\n            /\\btest|vitest|pytest|jest|build|tsc\\b/i.test(JSON.stringify(span.args)),\n          )\n        ? 0.4\n        : 0\n    if (!testReality) notes.push('no real test/build evidence recorded')\n\n    const blockerSpans = judgeSpans.filter((span) => isBlockingJudge(span))\n    const finalGateBlockers = finalGateSpans.filter((span) => isBlockingJudge(span))\n    const finalGate = finalGateSpans.length ? (finalGateBlockers.length ? 0 : 1) : success\n    if (finalGateBlockers.length)\n      notes.push(`final gate blocked by ${finalGateBlockers.length} reviewer(s)`)\n    else if (!finalGateSpans.length) notes.push('no final gate judgment recorded')\n\n    const reviewerBlockers = judgeSpans.length ? blockerSpans.length / judgeSpans.length : 0\n    if (reviewerBlockers) notes.push(`detected ${blockerSpans.length} blocking reviewer signal(s)`)\n\n    const positiveGroundingSignals =\n      patchEvidence +\n      sandboxSpans.length +\n      llmSpans.filter((span) => looksRepoGrounded(span.output ?? '')).length\n    const driftSignals =\n      llmSpans.filter((span) => this.isDrift(span.output ?? '')).length +\n      trace.events.filter((event) => this.isDrift(JSON.stringify(event.payload))).length\n    const repoGroundedness =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : positiveGroundingSignals / (positiveGroundingSignals + driftSignals)\n    const driftPenalty =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : driftSignals / (positiveGroundingSignals + driftSignals)\n    if (driftSignals > 0) notes.push(`detected ${driftSignals} drift signal(s)`)\n\n    const costUsd = trace.budget.length\n      ? Math.max(\n          ...trace.budget\n            .filter((entry: BudgetLedgerEntry) => entry.dimension === 'usd')\n            .map((entry: BudgetLedgerEntry) => entry.consumed),\n          0,\n        )\n      : llmSpans.reduce((sum, span) => sum + (span.costUsd ?? 0), 0)\n    const wallSeconds =\n      trace.run.endedAt && trace.run.startedAt\n        ? Math.max(0, (trace.run.endedAt - trace.run.startedAt) / 1000)\n        : 0\n\n    return {\n      success,\n      goalProgress,\n      repoGroundedness,\n      driftPenalty,\n      toolUseQuality,\n      patchQuality,\n      testReality,\n      finalGate,\n      reviewerBlockers,\n      costUsd,\n      wallSeconds,\n      notes,\n    }\n  }\n\n  rank(score: RunScore): number {\n    return aggregateRunScore(score, this.weights)\n  }\n\n  private isDrift(text: string): boolean {\n    return this.driftPatterns.some((pattern) => pattern.test(text))\n  }\n}\n\nfunction normalizeJudgeScore(score: number): number {\n  return score > 1 ? clamp01(score / 10) : clamp01(score)\n}\n\nfunction looksRepoGrounded(text: string): boolean {\n  return /(?:src\\/|tests?\\/|package\\.json|tsconfig|\\.ts\\b|\\.tsx\\b|git status|pnpm |npm |vitest|pytest|jest)/i.test(\n    text,\n  )\n}\n\nfunction isBlockingJudge(span: Extract<Span, { kind: 'judge' }>): boolean {\n  return (\n    span.attributes?.blocking === true ||\n    span.attributes?.verdict === 'BLOCKING' ||\n    positiveNumber(span.attributes?.blockingFindings) ||\n    positiveNumber(span.attributes?.highFindings) ||\n    span.score <= 2\n  )\n}\n\nfunction positiveNumber(value: unknown): boolean {\n  return typeof value === 'number' && value > 0\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport { CostLedger, type CostLedgerHandle, type CostReceipt } from './cost-ledger'\nimport {\n  callLlmJson,\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  type LlmCallRequest,\n  type LlmClientOptions,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Provider-enforced output limit. Default 16000. */\n  maxTokens?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** LlmClient config (baseUrl, apiKey, authHeader, …). */\n  llm?: LlmClientOptions\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MAX_TOKENS = 16_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: Required<SemanticConceptJudgeOptions>,\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions = {},\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts: Required<SemanticConceptJudgeOptions> = {\n    model: options.model ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    llm: options.llm ?? {},\n    costLedger: options.costLedger ?? new CostLedger(),\n    costPhase: options.costPhase ?? 'judge.semantic-concept',\n    costTags: options.costTags ?? {},\n    signal: options.signal ?? new AbortController().signal,\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  let receipt: CostReceipt | undefined\n  try {\n    const request = {\n      model: opts.model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n        },\n        { role: 'user' as const, content: buildPrompt(input, opts) },\n      ],\n      jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n      temperature: 0,\n      maxTokens: opts.maxTokens,\n      timeoutMs: opts.timeoutMs,\n    } satisfies LlmCallRequest\n    const paid = await opts.costLedger.runPaidCall({\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      model: opts.model,\n      ...(Object.keys(opts.costTags).length > 0 ? { tags: opts.costTags } : {}),\n      maximumCharge: maximumChargeForLlmRequest(request, opts.llm),\n      signal: opts.signal,\n      execute: (signal, callId) =>\n        callLlmJson<{ summary: string; concepts: ConceptFinding[] }>(request, {\n          ...opts.llm,\n          signal,\n          idempotencyKey: callId,\n        }),\n      receipt: ({ result }) => costReceiptFromLlm(result),\n      receiptFromError: costReceiptFromLlmError,\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid.value\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: paid.receipt.costUnknown ? null : paid.receipt.costUsd,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: receipt && !receipt.costUnknown ? receipt.costUsd : null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions = {},\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAeA,MAAM,0BAAU,IAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;CACrC,IAAI,IAAI,QAAQ,IAAI,IAAI;CACxB,IAAI,CAAC,GAAG;EACN,IAAI,IAAI,MAAM;EACd,QAAQ,IAAI,MAAM,CAAC;CACrB;CACA,OAAO;AACT;AAEA,IAAa,sBAAb,MAAiC;CAEH;CAD5B;CACA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,QAAQ,SAAS,IAAI;EAC1B,IAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAC3B,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAEhD;CAEA,MAAM,OAAO,OAA+B;EAC1C,MAAM,OAAO,GAAG,KAAK,UAAU,KAAK,EAAE;EACtC,MAAM,KAAK,MAAM,mBAAmB;GAClC,eAAe,KAAK,MAAM,IAAI;EAChC,CAAC;CACH;AACF;;;;;;;;;;;;;;;;;;;;;;ACPA,IAAa,gBAAb,MAA2B;CAGG;CAF5B;CAEA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,WAAW,IAAI,oBAAoB,IAAI;CAC9C;CAEA,MAAM,OAAO,OAAe,UAA2C;EACrE,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,MAAwB;IAAE,GAAG;IAAG,QAAQ;GAAM;GACpD,MAAM,KAAK,SAAS,OAAO,GAAG;EAChC;CACF;;CAGA,UAA8B;EAC5B,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,MAAM,aAAa,KAAK,MAAM,MAAM;EAC1C,IAAI,CAAC,KAAK,OAAO,CAAC;EAClB,MAAM,MAA0B,CAAC;EACjC,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;GAClC,IAAI,CAAC,MAAM;GACX,IAAI;IACF,IAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;GAC/C,QAAQ,CAGR;EACF;EACA,OAAO;CACT;;CAGA,QAAQ,OAAmC;EACzC,OAAO,KAAK,QAAQ,CAAC,CAAC,QAAQ,MAAM,EAAE,WAAW,KAAK;CACxD;AACF;;;;;AAiCA,SAAgB,kBAAkB,GAAmB,GAA4B;CAC/E,IAAI,EAAE,aAAa,EAAE,UAAU,OAAO;CACtC,IAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,OAAO;CACvE,IAAI,EAAE,cAAc,WAAW,EAAE,cAAc,QAAQ,OAAO;CAC9D,OAAO;AACT;;;;;AAMA,SAAgB,aACd,UACA,SACA,SAAqB,CAAC,GACR;CACd,MAAM,aAAa,OAAO,cAAc;CACxC,MAAM,WAAW,IAAI,IAAI,SAAS,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAC/D,MAAM,UAAU,IAAI,IAAI,QAAQ,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAE7D,MAAM,WAA+B,CAAC;CACtC,MAAM,cAAkC,CAAC;CACzC,MAAM,YAAgC,CAAC;CACvC,MAAM,UAAmC,CAAC;CAE1C,KAAK,MAAM,CAAC,IAAI,QAAQ,SAAS;EAC/B,MAAM,OAAO,SAAS,IAAI,EAAE;EAC5B,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,GAAG;GACjB;EACF;EACA,IAAI,WAAW,MAAM,GAAG,GACtB,QAAQ,KAAK;GAAE,UAAU;GAAM,SAAS;EAAI,CAAC;OAE7C,UAAU,KAAK,GAAG;CAEtB;CACA,KAAK,MAAM,CAAC,IAAI,SAAS,UACvB,IAAI,CAAC,QAAQ,IAAI,EAAE,GAAG,YAAY,KAAK,IAAI;CAE7C,OAAO;EAAE;EAAU;EAAa;EAAW;CAAQ;AACrD;;;;;;;;;;;;;;;;;;;;;;;;;ACjEA,MAAM,uBAAuB;AAE7B,MAAM,oBACJ;AACF,MAAM,aAAa;AAInB,SAAS,cAAc,MAAgD;CACrE,IAAI,CAAC,WAAW,IAAI,GAAG,OAAO,CAAC;CAC/B,MAAM,MAAwC,CAAC;CAC/C,KAAK,MAAM,SAAS,YAAY,MAAM,EAAE,eAAe,KAAK,CAAC,GAAG;EAC9D,IAAI,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,eAAe,GAAG;EACrD,MAAM,UAAU,KAAK,MAAM,MAAM,MAAM,UAAU;EACjD,IAAI,WAAW,OAAO,GAAG,IAAI,KAAK;GAAE,MAAM,MAAM;GAAM,MAAM;EAAQ,CAAC;CACvE;CACA,OAAO;AACT;AAEA,SAAS,UAAU,KAAa,KAAuB;CACrD,IAAI,CAAC,WAAW,GAAG,GAAG,OAAO,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,QAAQ,CAAC,GAAG;CAClB,OAAO,MAAM,QAAQ;EACnB,MAAM,MAAM,MAAM,IAAI;EACtB,IAAI;EACJ,IAAI;GACF,UAAU,YAAY,KAAK,EAAE,eAAe,KAAK,CAAC;EACpD,QAAQ;GACN;EACF;EACA,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,OAAO,KAAK,KAAK,EAAE,IAAI;GAC7B,IAAI,EAAE,YAAY,GAAG,MAAM,KAAK,IAAI;QAC/B,IAAI,EAAE,KAAK,SAAS,QAAQ,GAAG;IAClC,MAAM,KAAK,IAAI;IACf,IAAI,MAAM,KAAK,MAAM,UAAU,KAAK,OAAO;GAC7C;EACF;CACF;CACA,OAAO;AACT;AAEA,SAAS,uBAAuB,MAAsB;CAEpD,MAAM,QADK,wBAAwB,KAAK,IACzB,CAAC,GAAG,MAAM;CAEzB,OADU,uBAAuB,KAAK,KAC/B,CAAC,GAAG,MAAM;AACnB;AAEA,SAAS,eAAe,OAAiB,MAAc,SAA2B;CAChF,IAAI,IAAI;CACR,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,aAAa,CAAC,KAAK,MAAM,WAAW,IAAI,GAAG,GAAG,QAAQ,KAAK,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC;EACrF,KAAK,MAAM,OAAO,YAAY;GAC5B,IAAI,CAAC,WAAW,GAAG,GAAG;GACtB,IAAI;IACF,IAAI,SAAS,GAAG,CAAC,CAAC,YAAY,GAAG,KAAK,YAAY,GAAG,CAAC,CAAC;SAClD,KAAK;GACZ,QAAQ,CAER;EACF;CACF;CACA,OAAO;AACT;;AAGA,SAAgB,sBAAsB,QAAgD;CACpF,MAAM,SAAS,OAAO,WAAW,SAAS,EAAE,MAAM,WAChD,cAAc,IAAI,CAAC,CAAC,KAAK,OAAO;EAAE,GAAG;EAAG;CAAK,EAAE,CACjD;CACA,MAAM,QAAQ,OAAO,KAAK,MAAM,EAAE,IAAI;CAGtC,MAAM,SAAS,IAAI,IAAoB,MAAM,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;CAC/D,MAAM,QAAQ,IAAI,IAAoB,MAAM,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;CAC9D,MAAM,UAAU;CAChB,MAAM,QAAQ;CACd,IAAI,cAAc;CAClB,KAAK,MAAM,OAAO,OAAO,gBACvB,KAAK,MAAM,QAAQ,UAAU,KAAK,OAAO,wBAAwB,CAAC,GAAG;EACnE,eAAe;EACf,IAAI;EACJ,IAAI;GACF,OAAO,aAAa,MAAM,MAAM;EAClC,QAAQ;GACN;EACF;EACA,KAAK,MAAM,KAAK,KAAK,SAAS,OAAO,GAAG;GACtC,MAAM,IAAI,EAAE;GACZ,IAAI,CAAC,GAAG;GACR,MAAM,IAAI,EAAE,MAAM,GAAG,CAAC,CAAC,IAAI,KAAK;GAChC,MAAM,OAAO,OAAO,IAAI,CAAC;GACzB,IAAI,SAAS,KAAA,GAAW,OAAO,IAAI,GAAG,OAAO,CAAC;EAChD;EACA,KAAK,MAAM,KAAK,KAAK,SAAS,KAAK,GAAG;GACpC,MAAM,IAAI,EAAE;GACZ,IAAI,MAAM,KAAA,GAAW;GACrB,MAAM,OAAO,MAAM,IAAI,CAAC;GACxB,IAAI,SAAS,KAAA,GAAW,MAAM,IAAI,GAAG,OAAO,CAAC;EAC/C;CACF;CAIF,MAAM,yBAAS,IAAI,IAAoB;CACvC,KAAK,MAAM,KAAK,QACd,IAAI;EACF,OAAO,IAAI,EAAE,MAAM,aAAa,EAAE,MAAM,MAAM,CAAC;CACjD,QAAQ;EACN,OAAO,IAAI,EAAE,MAAM,EAAE;CACvB;CAEF,MAAM,UAAU,IAAI,IAAoB,MAAM,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;CAChE,KAAK,MAAM,UAAU,OAAO;EAC1B,MAAM,MAAM,IAAI,OAAO,IAAI,OAAO,YAAY,OAAO,OAAO;EAC5D,KAAK,MAAM,KAAK,QAAQ;GACtB,IAAI,EAAE,SAAS,QAAQ;GACvB,IAAI,IAAI,KAAK,OAAO,IAAI,EAAE,IAAI,KAAK,EAAE,GAAG,QAAQ,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAK,CAAC;EACtF;CACF;CAEA,MAAM,UAA8B,OAAO,KAAK,MAAM;EACpD,MAAM,OAAO,OAAO,IAAI,EAAE,IAAI,KAAK;EACnC,MAAM,MAAM,EAAE,KAAK,QAAQ,gBAAgB,EAAE;EAC7C,OAAO;GACL,MAAM,EAAE;GACR,MAAM,EAAE;GACR,MAAM,EAAE;GACR,OAAO,OAAO,KAAK,MAAM,IAAI,CAAC,CAAC,SAAS;GACxC,mBAAmB,OAAO,IAAI,EAAE,IAAI,KAAK;GACzC,kBAAkB,MAAM,IAAI,EAAE,IAAI,KAAK;GACvC,aAAa,QAAQ,IAAI,EAAE,IAAI,KAAK;GACpC,eAAe,eACb,OAAO,iBAAiB,CAAC,GACzB,EAAE,MACF,OAAO,kBAAkB,EAAE,SAAS,CAAC,CACvC;GACA,oBAAoB,KAAK,MAAM,iBAAiB,KAAK,CAAC,EAAA,CAAG;GACzD,kBAAkB,WAAW,KAAK,KAAK,YAAY,CAAC;GACpD,aAAa,WAAW,KAAK,KAAK,OAAO,CAAC;GAC1C,UAAU,KAAK,SAAS,kBAAkB;GAC1C,mBAAmB,WAAW,KAAK,uBAAuB,IAAI,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;EACvF;CACF,CAAC;CACD,OAAO;EAAE,qBAAqB;EAAa;CAAQ;AACrD;AAIA,MAAM,aAAa;AAEnB,SAAS,QACP,MACA,SACA,OACA,UACA,YACA,YACA,aACA,aACA,WACgB;CAChB,OAAO;EACL,gBAAgB;EAChB,YAAY,iBAAiB;GAAE,YAAY;GAAY;GAAM;GAAS;EAAM,CAAC;EAC7E,YAAY;EACZ,aAAa;EACb;EACA;EACA;EACA;EACA,eAAe,CAAC;GAAE,MAAM;GAAY,KAAK;EAAY,CAAC;EACtD,oBAAoB;EACpB;EACA;CACF;AACF;;AAGA,SAAgB,uBACd,QACA,YACkB;CAClB,MAAM,MAAwB,CAAC;CAC/B,KAAK,MAAM,KAAK,OAAO,SAAS;EAC9B,MAAM,cAAc,EAAE,oBAAoB,EAAE;EAI5C,IAHkB,cAAc,EAAE,cAAc,EAAE,kBAGhC,GAChB,IAAI,KACF,QACE,eACA,EAAE,MACF,UAAU,EAAE,KAAK,+EACjB,QACA,IACA,YACA,sIACA,EAAE,MACF,iGACF,CACF;OACK,IAAI,gBAAgB,KAAK,EAAE,cAAc,EAAE,gBAAgB,GAEhE,IAAI,KACF,QACE,eACA,EAAE,MACF,UAAU,EAAE,KAAK,gFAAgF,EAAE,YAAY,cAAc,EAAE,cAAc,IAC7I,QACA,IACA,YACA,2JACA,EAAE,MACF,sEACF,CACF;EAIF,IAAI,eAAe,KAAK,CAAC,EAAE,mBACzB,IAAI,KACF,QACE,mBACA,EAAE,MACF,UAAU,EAAE,KAAK,mFACjB,UACA,IACA,YACA,oHACA,EAAE,IACJ,CACF;EAIF,IAAI,EAAE,SAAS,YAAY,EAAE,oBAAoB,GAC/C,IAAI,KACF,QACE,UACA,EAAE,MACF,iBAAiB,EAAE,KAAK,YAAY,EAAE,kBAAkB,+BACxD,QACA,KACA,YACA,uMACA,EAAE,IACJ,CACF;EAIF,IAAI,EAAE,QAAQ,wBAAwB,CAAC,EAAE,kBACvC,IAAI,KACF,QACE,mBACA,EAAE,MACF,UAAU,EAAE,KAAK,OAAO,EAAE,MAAM,4DAChC,UACA,IACA,YACA,mFAAmF,EAAE,MAAM,mDAC3F,EAAE,IACJ,CACF;EAIF,IAAI,CAAC,EAAE,aACL,IAAI,KACF,QACE,gBACA,EAAE,MACF,UAAU,EAAE,KAAK,oBACjB,OACA,IACA,YACA,wGACA,EAAE,IACJ,CACF;EAIF,IAAI,CAAC,EAAE,UACL,IAAI,KACF,QACE,iBACA,EAAE,MACF,UAAU,EAAE,KAAK,8CACjB,OACA,KACA,YACA,iJACA,EAAE,IACJ,CACF;CAEJ;CACA,OAAO;AACT;AAIA,IAAa,oBAAb,MAAoE;CAClE,KAAc;CACd,cACE;CACF,YAAqB;CACrB,OAAgB;EAAE,MAAM;EAA0B,iBAAiB;CAAE;CACrE,UAAmB;CAEnB,MAAM,QAAQ,OAAyB,KAAgD;EACrF,MAAM,aAAa,IAAI,MAAM,+BAAc,IAAI,KAAK,EAAA,CAAE,YAAY;EAClE,IAAI,MACF,gBAAgB,MAAM,QAAQ,OAAO,eAAe,MAAM,oBAAoB,aAChF;EACA,OAAO,uBAAuB,OAAO,UAAU;CACjD;AACF;AAEA,MAAa,sBAAsB,IAAI,kBAAkB;;;AClYzD,MAAM,yBAAyB;CAC7B;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,IAAa,YAAb,MAAuB;CACrB;CACA;CAEA,YAAY,UAA4B,CAAC,GAAG;EAC1C,KAAK,UAAU,QAAQ;EACvB,KAAK,gBAAgB,QAAQ,iBAAiB;CAChD;CAEA,MAAM,MAAM,OAAmB,OAAkC;EAC/D,MAAM,MAAM,MAAM,MAAM,OAAO,KAAK;EACpC,IAAI,CAAC,KAAK,MAAM,IAAI,cAAc,OAAO,MAAM,WAAW;EAC1D,MAAM,CAAC,OAAO,QAAQ,WAAW,UAAU,MAAM,QAAQ,IAAI;GAC3D,MAAM,MAAM,EAAE,MAAM,CAAC;GACrB,MAAM,OAAO,EAAE,MAAM,CAAC;GACtB,MAAM,UAAU,KAAK;GACrB,MAAM,OAAO,KAAK;EACpB,CAAC;EACD,OAAO,KAAK,WAAW;GAAE;GAAK;GAAO;GAAQ;GAAW;EAAO,CAAC;CAClE;CAEA,WAAW,OAA2B;EACpC,MAAM,QAAkB,CAAC;EACzB,MAAM,WAAW,MAAM,MAAM,QAC1B,MAA2C,EAAE,SAAS,KACzD;EACA,MAAM,YAAY,MAAM,MAAM,QAC3B,MAA4C,EAAE,SAAS,MAC1D;EACA,MAAM,aAAa,MAAM,MAAM,QAC5B,MAA6C,EAAE,SAAS,OAC3D;EACA,MAAM,eAAe,MAAM,MAAM,QAC9B,MAA+C,EAAE,SAAS,SAC7D;EACA,MAAM,iBAAiB,WAAW,QAC/B,SAAS,KAAK,cAAc,gBAAgB,KAAK,YAAY,cAAc,IAC9E;EAEA,MAAM,UACJ,MAAM,IAAI,SAAS,SAAS,OAAO,IAAI,MAAM,IAAI,WAAW,cAAc,KAAM;EAClF,IAAI,CAAC,SAAS,MAAM,KAAK,qCAAqC;EAE9D,MAAM,eAAe,WAAW,SAC5B,WAAW,QAAQ,KAAK,SAAS,MAAM,oBAAoB,KAAK,KAAK,GAAG,CAAC,IACzE,WAAW,SACX,KAAA;EAOJ,MAAM,gBALJ,OAAO,MAAM,IAAI,SAAS,UAAU,WAChC,QACE,MAAM,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAI,QAAQ,QAAQ,MAAM,MAAM,IAAI,QAAQ,KAClF,IACA,KAAA,MAC+B,gBAAgB;EAErD,MAAM,kBAAkB,UAAU,QAAQ,SAAS,KAAK,WAAW,OAAO,CAAC,CAAC;EAC5E,MAAM,iBAAiB,UAAU,WAAW,IAAI,IAAI,kBAAkB,UAAU;EAChF,IAAI,UAAU,WAAW,GAAG,MAAM,KAAK,wBAAwB;EAE/D,MAAM,gBACJ,MAAM,UAAU,SAChB,UAAU,QAAQ,SAAS,0BAA0B,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC;EAC5E,MAAM,eAAe,gBAAgB,IAAI,QAAQ,gBAAgB,CAAC,IAAI;EACtE,IAAI,CAAC,cAAc,MAAM,KAAK,uCAAuC;EAErE,MAAM,eAAe,aAAa,QAC/B,SAAS,OAAO,KAAK,eAAe,YAAY,KAAK,aAAa,CACrE;EACA,MAAM,cAAc,aAAa,SAC7B,aAAa,QACV,KAAK,SAAS,OAAO,KAAK,eAAe,KAAK,KAAK,IAAI,GAAG,KAAK,cAAc,CAAC,GAC/E,CACF,IAAI,aAAa,SACjB,UAAU,MAAM,SACZ,yCAAyC,KAAK,KAAK,UAAU,KAAK,IAAI,CAAC,CACzE,IACA,KACA;EACN,IAAI,CAAC,aAAa,MAAM,KAAK,sCAAsC;EAEnE,MAAM,eAAe,WAAW,QAAQ,SAAS,gBAAgB,IAAI,CAAC;EACtE,MAAM,oBAAoB,eAAe,QAAQ,SAAS,gBAAgB,IAAI,CAAC;EAC/E,MAAM,YAAY,eAAe,SAAU,kBAAkB,SAAS,IAAI,IAAK;EAC/E,IAAI,kBAAkB,QACpB,MAAM,KAAK,yBAAyB,kBAAkB,OAAO,aAAa;OACvE,IAAI,CAAC,eAAe,QAAQ,MAAM,KAAK,iCAAiC;EAE7E,MAAM,mBAAmB,WAAW,SAAS,aAAa,SAAS,WAAW,SAAS;EACvF,IAAI,kBAAkB,MAAM,KAAK,YAAY,aAAa,OAAO,6BAA6B;EAE9F,MAAM,2BACJ,gBACA,aAAa,SACb,SAAS,QAAQ,SAAS,kBAAkB,KAAK,UAAU,EAAE,CAAC,CAAC,CAAC;EAClE,MAAM,eACJ,SAAS,QAAQ,SAAS,KAAK,QAAQ,KAAK,UAAU,EAAE,CAAC,CAAC,CAAC,SAC3D,MAAM,OAAO,QAAQ,UAAU,KAAK,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,CAAC,CAAC;EAC9E,MAAM,mBACJ,2BAA2B,iBAAiB,IACxC,IACA,4BAA4B,2BAA2B;EAC7D,MAAM,eACJ,2BAA2B,iBAAiB,IACxC,IACA,gBAAgB,2BAA2B;EACjD,IAAI,eAAe,GAAG,MAAM,KAAK,YAAY,aAAa,iBAAiB;EAe3E,OAAO;GACL;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA,SAvBc,MAAM,OAAO,SACzB,KAAK,IACH,GAAG,MAAM,OACN,QAAQ,UAA6B,MAAM,cAAc,KAAK,CAAC,CAC/D,KAAK,UAA6B,MAAM,QAAQ,GACnD,CACF,IACA,SAAS,QAAQ,KAAK,SAAS,OAAO,KAAK,WAAW,IAAI,CAAC;GAiB7D,aAfA,MAAM,IAAI,WAAW,MAAM,IAAI,YAC3B,KAAK,IAAI,IAAI,MAAM,IAAI,UAAU,MAAM,IAAI,aAAa,GAAI,IAC5D;GAcJ;EACF;CACF;CAEA,KAAK,OAAyB;EAC5B,OAAO,kBAAkB,OAAO,KAAK,OAAO;CAC9C;CAEA,QAAgB,MAAuB;EACrC,OAAO,KAAK,cAAc,MAAM,YAAY,QAAQ,KAAK,IAAI,CAAC;CAChE;AACF;AAEA,SAAS,oBAAoB,OAAuB;CAClD,OAAO,QAAQ,IAAI,QAAQ,QAAQ,EAAE,IAAI,QAAQ,KAAK;AACxD;AAEA,SAAS,kBAAkB,MAAuB;CAChD,OAAO,qGAAqG,KAC1G,IACF;AACF;AAEA,SAAS,gBAAgB,MAAiD;CACxE,OACE,KAAK,YAAY,aAAa,QAC9B,KAAK,YAAY,YAAY,cAC7B,eAAe,KAAK,YAAY,gBAAgB,KAChD,eAAe,KAAK,YAAY,YAAY,KAC5C,KAAK,SAAS;AAElB;AAEA,SAAS,eAAe,OAAyB;CAC/C,OAAO,OAAO,UAAU,YAAY,QAAQ;AAC9C;;;;;;;;;;;;;;;;;;;;;;;AC/EA,MAAa,6BAAgE;CAC3E,QAAQ;CACR,WAAW;CACX,SAAS;AACX;AAiCA,MAAa,iCAAiC;AAE9C,MAAM,qBAAqB;AAC3B,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,kBAAkB;AACxB,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;AAEtB,MAAM,kBAAkB;CACtB,MAAM;CACN,sBAAsB;CACtB,UAAU,CAAC,WAAW,UAAU;CAChC,YAAY;EACV,SAAS;GAAE,MAAM;GAAU,WAAW;GAAI,WAAW;EAAI;EACzD,UAAU;GACR,MAAM;GACN,UAAU;GACV,OAAO;IACL,MAAM;IACN,sBAAsB;IACtB,UAAU;KAAC;KAAW;KAAW;KAAS;KAAY;IAAU;IAChE,YAAY;KACV,SAAS;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACxD,SAAS,EAAE,MAAM,UAAU;KAC3B,OAAO;MAAE,MAAM;MAAU,SAAS;MAAG,SAAS;KAAG;KACjD,UAAU;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACzD,UAAU;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAS;OAAS;MAAM;KAAE;IAC3E;GACF;EACF;CACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;CAClE,IAAI,KAAK,UAAU,KAAK,OAAO;CAC/B,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,iBAAiB,KAAK,SAAS,IAAI,YAAY,MAAM;AACpF;AAEA,SAAS,YACP,OACA,MACQ;CACR,MAAM,aAAa,MAAM,YACtB,QAAQ,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,CAAC,CACvD,KAAK,MAAM,aAAa,EAAE,KAAK,QAAQ,EAAE,SAAS,CAAC,CACnD,KAAK,MAAM;CAEd,MAAM,OAAO,MAAM,cAAc;CAEjC,OAAO;;;;;;;;;;EAUP,MAAM,YAAY;;EAElB,MAAM,gBAAgB,+BAA+B,MAAM,cAAc,mBAAmB,MAAM,uBAAuB,GAAG,QAAQ,GAAG;EACvI,MAAM,iBACL,KACE,GAAG,MACF,KAAK,IAAI,EAAE,KAAK,EAAE,KAAK,GAAG,EAAE,UAAU,SAAS,cAAc,EAAE,SAAS,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,EAAE,KAAK,IACzG,CAAC,CACA,KAAK,IAAI,EAAE;;EAEZ,OAAO,qDAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,EAAE,QAAQ,GAAG;EACjH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,EAAE;;;;;;;;;;;;;;;;;;AAkBtD;;;;;;AASA,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACH;CACrC,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,aAAa,MAAM,iBAAiB;CAE1C,IAAI,eAAe,GACjB,OAAO;EACL,MAAM;EACN,SAAS;EACT,OAAO;EACP,cAAc;EACd,YAAY;EACZ,UAAU,CAAC;EACX,SAAS;EACT,YAAY;EACZ,SAAS;EACT,WAAW;EACX,OAAO;CACT;CAGF,MAAM,OAA8C;EAClD,OAAO,QAAQ,SAAS;EACxB,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,KAAK,QAAQ,OAAO,CAAC;EACrB,YAAY,QAAQ,cAAc,IAAI,WAAW;EACjD,WAAW,QAAQ,aAAa;EAChC,UAAU,QAAQ,YAAY,CAAC;EAC/B,QAAQ,QAAQ,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAChD,gBAAgB,QAAQ,kBAAkB;EAC1C,mBAAmB;GAAE,GAAG;GAA4B,GAAI,QAAQ,qBAAqB,CAAC;EAAG;CAC3F;CAMA,MAAM,oBAAoB,SAA8B;EACtD,IAAI,KAAK,mBAAmB,QAAQ,OAAO;EAC3C,IAAI,KAAK,UAAU,MAAM,OAAO,KAAK;EACrC,IAAI,KAAK,mBAAmB,cAC1B,OAAO,KAAK,kBAAkB,KAAK,cAAc,aAAa;EAEhE,OAAO;CACT;CACA,MAAM,eAAe,IAAI,IACvB,MAAM,iBAAiB,KAAK,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC,CACjE;CAEA,IAAI;CACJ,IAAI;EACF,MAAM,UAAU;GACd,OAAO,KAAK;GACZ,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IAAE,MAAM;IAAiB,SAAS,YAAY,OAAO,IAAI;GAAE,CAC7D;GACA,YAAY;IAAE,MAAM;IAA0B,QAAQ;GAAgB;GACtE,aAAa;GACb,WAAW,KAAK;GAChB,WAAW,KAAK;EAClB;EACA,MAAM,OAAO,MAAM,KAAK,WAAW,YAAY;GAC7C,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,OAAO,KAAK;GACZ,GAAI,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,SAAS,IAAI,EAAE,MAAM,KAAK,SAAS,IAAI,CAAC;GACvE,eAAe,2BAA2B,SAAS,KAAK,GAAG;GAC3D,QAAQ,KAAK;GACb,UAAU,QAAQ,WAChB,YAA6D,SAAS;IACpE,GAAG,KAAK;IACR;IACA,gBAAgB;GAClB,CAAC;GACH,UAAU,EAAE,aAAa,mBAAmB,MAAM;GAClD,kBAAkB;EACpB,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU,KAAK;EAEvB,IAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GACnD,MAAM,IAAI,MAAM,uEAAqE;EAGvF,MAAM,WAA6B,MAAM,SAAS,KAAK,OAAO;GAC5D,SAAS,OAAO,EAAE,OAAO;GACzB,SAAS,QAAQ,EAAE,OAAO;GAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;GACrD,UAAU,OAAO,EAAE,YAAY,EAAE;GACjC,UAAW;IAAC;IAAY;IAAS;IAAS;GAAM,CAAC,CAAW,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;EACN,EAAE;EAEF,MAAM,eAAe,SAAS,QAAQ,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;EACvE,IAAI,YAAY;EAChB,IAAI,mBAAmB;EACvB,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;GACzC,aAAa;GACb,oBAAoB,IAAI,EAAE;EAC5B;EACA,MAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;EAE7E,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO,QAAQ,WAAW,GAAA,CAAI,QAAQ,CAAC,CAAC;GACxC;GACA;GACA;GACA,SAAS,OAAO,MAAM,WAAW,EAAE;GACnC,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,KAAK,QAAQ,cAAc,OAAO,KAAK,QAAQ;GACxD,WAAW;EACb;CACF,SAAS,KAAK;EACZ,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO;GACP,cAAc;GACd;GACA,UAAU,CAAC;GACX,SAAS;GACT,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,WAAW,CAAC,QAAQ,cAAc,QAAQ,UAAU;GAC7D,WAAW;GACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;EACxD;CACF;AACF;;;;;AAMA,SAAgB,2BACd,UAAuC,CAAC,GACmC;CAC3E,QAAQ,UAAU,wBAAwB,OAAO,OAAO;AAC1D"}

package/dist/{server-Bz3WQJs6.js → server-KjXZZUDX.js}

@@ -1,5 +1,5 @@
-import { i as CostLedger } from "./cost-ledger-DHAjwNj7.js";
-import { a as assertLlmRoute, c as callLlmJson, f as maximumChargeForLlmRequest, i as LlmRouteAssertionError, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-DHx8pzyJ.js";
+import { i as CostLedger } from "./cost-ledger-D-5_-dhi.js";
+import { a as assertLlmRoute, c as callLlmJson, f as maximumChargeForLlmRequest, i as LlmRouteAssertionError, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-DKB25jV8.js";
 import { z } from "zod";
 import { readFileSync } from "node:fs";
 import { dirname, resolve } from "node:path";
@@ -1037,4 +1037,4 @@ function startServerAsync(opts = {}) {
 //#endregion
 export { TraceEventSchema as A, HealthResponseSchema as C, RubricDimensionSchema as D, ListRubricsResponseSchema as E, hashRubric as F, TracesIngestResponseSchema as M, VersionResponseSchema as N, RubricInfoSchema as O, WIRE_VERSION as P, FeedbackTrajectorySchema as S, JudgeResultSchema as T, ErrorResponseSchema as _, runRpcBatch as a, FeedbackIngestResponseSchema as b, WireError as c, handleListRubrics as d, handleTracesIngest as f, listBuiltinRubrics as g, getBuiltinRubric as h, dispatchRpc as i, TracesIngestRequestSchema as j, RubricSchema as k, handleFeedbackIngest as l, BUILTIN_RUBRICS as m, startServer as n, runRpcOnce as o, handleVersion as p, startServerAsync as r, buildOpenApi as s, createApp as t, handleJudge as u, FailureModeSchema as v, JudgeRequestSchema as w, FeedbackLabelSchema as x, FeedbackAttemptSchema as y };
 
-//# sourceMappingURL=server-Bz3WQJs6.js.map
+//# sourceMappingURL=server-KjXZZUDX.js.map