@tangle-network/agent-eval 0.102.1 → 0.103.1

package/dist/chunk-VNM52AGA.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/campaign/run-campaign.ts","../src/integrity/backend-integrity.ts","../src/verdict-cache.ts","../src/campaign/run-dir.ts","../src/campaign/storage.ts"],"sourcesContent":["/**\n * `runCampaign` — Pass A substrate primitive. ONE function that orchestrates\n * scenarios → dispatch → artifacts → judges → aggregates, with full\n * reproducibility (seed + manifest hash), cell-level resumability, bootstrap\n * CIs, and the `LabeledScenarioStore` capture flywheel.\n *\n * Improvement loops (optimizer / gate / autoOnPromote) ride on top of this\n * primitive but live in `presets/run-improvement-loop.ts`. This file keeps\n * the core orchestrator minimal — Phase 1 of the Pass A track.\n */\n\nimport { join } from 'node:path'\nimport { BackendIntegrityError, type BackendIntegrityReport } from '../integrity/backend-integrity'\nimport { confidenceInterval } from '../statistics'\nimport { contentHash } from '../verdict-cache'\nimport { resolveRunDir } from './run-dir'\nimport { type CampaignStorage, fsCampaignStorage } from './storage'\nimport type {\n  CampaignAggregates,\n  CampaignArtifactWriter,\n  CampaignCellResult,\n  CampaignCostMeter,\n  CampaignResult,\n  CampaignTokenUsage,\n  CampaignTraceWriter,\n  DispatchContext,\n  DispatchFn,\n  JudgeAggregate,\n  JudgeConfig,\n  JudgeScore,\n  LabeledScenarioStore,\n  Scenario,\n  ScenarioAggregate,\n  TraceSpan,\n} from './types'\n\nexport interface RunCampaignOptions<TScenario extends Scenario, TArtifact> {\n  scenarios: TScenario[]\n  dispatch: DispatchFn<TScenario, TArtifact>\n  /**\n   * Stable identity for the dispatch behavior, included in the manifest/cache\n   * key. Set this when the same function name can run different models,\n   * prompts, tools, or external config.\n   */\n  dispatchRef?: string\n  judges?: JudgeConfig<TArtifact, TScenario>[]\n  /** Required for reproducibility. Default 42. */\n  seed?: number\n  /** Per-scenario replicates for CI bands. Default 1; raise to 5+ for\n   *  bootstrap-tight intervals on critical eval. */\n  reps?: number\n  /** When true (default), completed cells are cached by\n   *  (manifestHash, scenarioId, rep, generation). Re-runs skip cached cells. */\n  resumable?: boolean\n  /** Optional store — when present, every artifact + judge score is captured\n   *  with the configured `captureSource`. Capture is default ON; pass `'off'`\n   *  to disable. */\n  labeledStore?: LabeledScenarioStore | 'off'\n  captureSource?: 'production-trace' | 'eval-run' | 'manual' | 'red-team' | 'synthetic'\n  captureSourceVersionHash?: string\n  /** Wall-clock cost cap across all cells. Cells beyond ceiling are skipped. */\n  costCeiling?: number\n  /** Max concurrent cells. Default 2. */\n  maxConcurrency?: number\n  /**\n   * Per-cell dispatch deadline in ms. A `dispatch` that neither resolves nor\n   * rejects within this window is a hang (a stalled model request, an\n   * exhausted runtime resource, a backend that never closes its stream). When\n   * set, the cell's `ctx.signal` is aborted and the cell is recorded as a LOUD\n   * error (`dispatch exceeded <N>ms`) so the campaign proceeds and the failure\n   * is visible — instead of one wedged cell silently hanging the whole run (and\n   * every loop/CI job above it) forever. `undefined`/`0` = unbounded (legacy).\n   */\n  dispatchTimeoutMs?: number\n  /** Required: where artifacts + traces land. A bare name (not an absolute path)\n   *  resolves to the shared `~/.tangle/traces/<repo>/runs/<name>` root so run\n   *  bundles never pollute a repo working tree. Pass an absolute path to override. */\n  runDir: string\n  /** Subject repo for the shared run-dir root (defaults to the CWD basename).\n   *  Only consulted when `runDir` is a bare name. */\n  repo?: string\n  /** Tracing posture. Default is the substrate's `FileSystemTraceStore` rooted\n   *  at `<runDir>/traces/`. `'off'` disables capture entirely — substrate\n   *  refuses this when the caller wires `autoOnPromote !== 'none'`. */\n  tracing?: 'on' | 'off'\n  /**\n   * Per-cell usage expectation — the early, fine-grained sibling of the\n   * batch `assertRealBackend` guard. A cell that produced an artifact (no\n   * error) but reported `costUsd === 0` AND zero tokens is a stub: the\n   * dispatch never reported LLM activity via `ctx.cost`. Modes:\n   *   - `'warn'` (default) — log the offending cell loudly, keep going.\n   *   - `'assert'` — throw `BackendIntegrityError` on the first such cell\n   *     (fail-fast; recommended for CI campaigns expecting real LLM calls).\n   *   - `'off'` — no check (replay / deterministic-only / offline analysis).\n   */\n  expectUsage?: 'assert' | 'warn' | 'off'\n  /** Test seam — override the wall clock for deterministic tests. */\n  now?: () => Date\n  /** Test seam — override per-cell trace writer factory. */\n  buildTraceWriter?: (cellId: string, dir: string) => CampaignTraceWriter\n  /** Storage backend for run/cell dirs, the resumability cache, artifacts,\n   *  and trace spans. Default: the Node filesystem (`fsCampaignStorage`).\n   *  Pass `inMemoryCampaignStorage()` to run in a filesystem-less runtime\n   *  (Cloudflare Workers, Deno, edge) — the `CampaignResult` is still\n   *  produced; artifacts/traces just aren't persisted to disk. */\n  storage?: CampaignStorage\n  /**\n   * Optional per-cell placement strategy. Returns an opaque string the\n   * substrate forwards as `ctx.placement` to the Dispatch — placement-aware\n   * Dispatches (e.g. `httpDispatch` from `/adapters/http`) use it to route\n   * each cell to the right worker, region, or sandbox. When unset, every\n   * cell receives `ctx.placement = undefined` and behaves identically to\n   * the in-process case.\n   *\n   * @example\n   *   cellPlacement: ({ scenario }) => scenario.tags?.includes('eu') ? 'eu-west' : 'us-east'\n   */\n  cellPlacement?: (input: {\n    scenario: TScenario\n    rep: number\n    generation?: number\n  }) => string | undefined\n}\n\n/**\n * Core campaign orchestrator: fan scenarios through dispatch, score with judges, aggregate bootstrap CIs, and persist reproducible `CampaignResult` records.\n */\nexport async function runCampaign<TScenario extends Scenario, TArtifact>(\n  opts: RunCampaignOptions<TScenario, TArtifact>,\n): Promise<CampaignResult<TArtifact, TScenario>> {\n  const seed = opts.seed ?? 42\n  const reps = opts.reps ?? 1\n  const resumable = opts.resumable ?? true\n  const maxConcurrency = opts.maxConcurrency ?? 2\n  const now = opts.now ?? (() => new Date())\n  const judges = opts.judges ?? []\n  const storage = opts.storage ?? fsCampaignStorage()\n\n  if (typeof opts.runDir !== 'string' || opts.runDir.trim().length === 0) {\n    throw new Error('runCampaign: runDir is required and must be a non-empty string')\n  }\n  opts.runDir = resolveRunDir(opts.runDir, opts.repo)\n  storage.ensureDir(opts.runDir)\n\n  const manifestHash = computeManifestHash({\n    scenarios: opts.scenarios,\n    judges: judges as unknown as JudgeConfig<unknown>[],\n    dispatchRef: dispatchRefFor(opts.dispatch, opts.dispatchRef),\n    seed,\n    reps,\n  })\n\n  const startedAt = now()\n  const cells: CampaignCellResult<TArtifact>[] = []\n  const artifactsByPath: Record<string, string> = {}\n\n  // Build the cell schedule (scenario × rep).\n  const schedule = buildCellSchedule(opts.scenarios, seed, reps)\n\n  // Concurrency-limited execution.\n  let totalCostUsd = 0\n  let costCeilingReached = false\n  const abortController = new AbortController()\n  // Concurrency lanes that drain the cell schedule. Named \"lanes\" — not\n  // \"workers\" — to avoid clashing with the taxonomy's worker (= the agent\n  // harness in a sandbox, invoked behind `dispatch`). See loop-taxonomy.md.\n  const lanes: Promise<void>[] = []\n  let nextIdx = 0\n  const cellsRef = cells\n\n  for (let i = 0; i < maxConcurrency; i++) {\n    lanes.push(\n      (async () => {\n        while (true) {\n          const myIdx = nextIdx++\n          if (myIdx >= schedule.length) return\n          const slot = schedule[myIdx]!\n          if (costCeilingReached) {\n            cellsRef.push(skippedCell(slot, 'cost_ceiling_reached'))\n            continue\n          }\n          const result = await executeCell({\n            slot,\n            opts,\n            manifestHash,\n            resumable,\n            now,\n            storage,\n            buildTraceWriter: opts.buildTraceWriter ?? defaultBuildTraceWriter(storage),\n            signal: abortController.signal,\n            dispatchTimeoutMs: opts.dispatchTimeoutMs,\n          })\n          cellsRef.push(result.cell)\n          enforceCellUsage(result.cell, opts.expectUsage ?? 'warn')\n          totalCostUsd += result.cell.costUsd\n          Object.assign(artifactsByPath, result.artifactsByPath)\n          if (opts.costCeiling !== undefined && totalCostUsd >= opts.costCeiling) {\n            costCeilingReached = true\n          }\n          // Capture into LabeledScenarioStore unless explicitly disabled.\n          if (opts.labeledStore && opts.labeledStore !== 'off' && !result.cell.error) {\n            await captureToStore({\n              store: opts.labeledStore,\n              cell: result.cell,\n              scenario: slot.scenario,\n              opts,\n              now,\n            }).catch((err) => {\n              // Capture failures are non-fatal — log but don't crash the campaign.\n              // (Trace would normally land here.)\n              console.warn(\n                `[runCampaign] capture failed for ${result.cell.cellId}: ${err instanceof Error ? err.message : String(err)}`,\n              )\n            })\n          }\n        }\n      })(),\n    )\n  }\n  await Promise.all(lanes)\n\n  const endedAt = now()\n  cellsRef.sort((a, b) => a.cellId.localeCompare(b.cellId))\n\n  const aggregates = computeAggregates(\n    cellsRef,\n    judges as unknown as JudgeConfig<TArtifact>[],\n    seed,\n  )\n\n  return {\n    manifestHash,\n    seed,\n    startedAt: startedAt.toISOString(),\n    endedAt: endedAt.toISOString(),\n    durationMs: endedAt.getTime() - startedAt.getTime(),\n    cells: cellsRef,\n    aggregates,\n    runDir: opts.runDir,\n    artifactsByPath,\n    scenarios: opts.scenarios.map((s) => ({ id: s.id, kind: s.kind })),\n  }\n}\n\n// ── Internals ─────────────────────────────────────────────────────────\n\ninterface ExecuteCellArgs<TScenario extends Scenario, TArtifact> {\n  slot: { scenario: TScenario; rep: number; cellId: string; cellSeed: number }\n  opts: RunCampaignOptions<TScenario, TArtifact>\n  manifestHash: string\n  resumable: boolean\n  now: () => Date\n  storage: CampaignStorage\n  buildTraceWriter: (cellId: string, dir: string) => CampaignTraceWriter\n  signal: AbortSignal\n  dispatchTimeoutMs?: number\n}\n\nasync function executeCell<TScenario extends Scenario, TArtifact>(\n  args: ExecuteCellArgs<TScenario, TArtifact>,\n): Promise<{ cell: CampaignCellResult<TArtifact>; artifactsByPath: Record<string, string> }> {\n  const storage = args.storage\n  const cellDir = join(args.opts.runDir, args.slot.cellId.replace(/[^a-zA-Z0-9_-]/g, '_'))\n  storage.ensureDir(cellDir)\n\n  // Resumability: cache key = (manifestHash, scenarioId, rep)\n  const cachePath = join(cellDir, 'cached-result.json')\n  if (args.resumable) {\n    const cached = readCachedCell<TArtifact>({\n      storage,\n      cachePath,\n      cellId: args.slot.cellId,\n      manifestHash: args.manifestHash,\n    })\n    if (cached.status === 'hit') {\n      return { cell: { ...cached.cell, cached: true }, artifactsByPath: {} }\n    }\n  }\n\n  const startMs = Date.now()\n  const trace = args.buildTraceWriter(args.slot.cellId, cellDir)\n  const artifactsByPath: Record<string, string> = {}\n  const artifacts: CampaignArtifactWriter = {\n    async write(path, content) {\n      const fullPath = join(cellDir, path)\n      storage.ensureDir(join(fullPath, '..'))\n      storage.write(fullPath, content)\n      artifactsByPath[`${args.slot.cellId}/${path}`] = fullPath\n      return fullPath\n    },\n    async writeJson(path, value) {\n      return artifacts.write(path, JSON.stringify(value, null, 2))\n    },\n  }\n  let costSoFar = 0\n  const tokensSoFar: CampaignTokenUsage = { input: 0, output: 0 }\n  let resolvedModel: string | undefined\n  const cost: CampaignCostMeter = {\n    observe(amount, source) {\n      costSoFar += amount\n      trace.span(`cost.${source}`, { amountUsd: amount }).end()\n    },\n    observeTokens(usage) {\n      tokensSoFar.input += usage.input\n      tokensSoFar.output += usage.output\n      if (usage.cached) tokensSoFar.cached = (tokensSoFar.cached ?? 0) + usage.cached\n    },\n    observeModel(model) {\n      const trimmed = model?.trim()\n      if (trimmed) resolvedModel = trimmed\n    },\n    current() {\n      return costSoFar\n    },\n    tokens() {\n      return { ...tokensSoFar }\n    },\n    resolvedModel() {\n      return resolvedModel\n    },\n  }\n\n  const placement = args.opts.cellPlacement?.({\n    scenario: args.slot.scenario,\n    rep: args.slot.rep,\n  })\n\n  // Per-cell abort signal, chained to the campaign signal. The dispatch sees\n  // THIS signal so a timeout (below) can abort just this cell's in-flight work\n  // without tearing down sibling cells — and a signal-honoring dispatch\n  // releases its open request instead of leaking it past the deadline.\n  const cellAbort = new AbortController()\n  const onCampaignAbort = () => cellAbort.abort((args.signal as { reason?: unknown }).reason)\n  if (args.signal.aborted) cellAbort.abort((args.signal as { reason?: unknown }).reason)\n  else args.signal.addEventListener('abort', onCampaignAbort, { once: true })\n\n  const ctx: DispatchContext = {\n    cellId: args.slot.cellId,\n    rep: args.slot.rep,\n    seed: args.slot.cellSeed,\n    signal: cellAbort.signal,\n    trace,\n    artifacts,\n    cost,\n    placement,\n  }\n\n  let artifact: TArtifact | undefined\n  let errorMessage: string | undefined\n  const timeoutMs = args.dispatchTimeoutMs\n  let timeoutTimer: ReturnType<typeof setTimeout> | undefined\n  try {\n    const dispatched = args.opts.dispatch(args.slot.scenario, ctx)\n    if (timeoutMs !== undefined && timeoutMs > 0) {\n      // A dispatch that never settles (stalled model request, exhausted runtime\n      // resource, a stream that never closes) must NOT hang the cell — and with\n      // it the lane, the campaign, the loop, the CI job — forever. Race it\n      // against the deadline; on timeout, abort the cell and fail it LOUD.\n      artifact = await Promise.race([\n        dispatched,\n        new Promise<never>((_, reject) => {\n          timeoutTimer = setTimeout(() => {\n            cellAbort.abort(new Error('dispatch timeout'))\n            reject(\n              new Error(\n                `dispatch exceeded ${timeoutMs}ms for cell '${args.slot.cellId}' — aborted and failed loud (no silent hang)`,\n              ),\n            )\n          }, timeoutMs)\n          if (typeof (timeoutTimer as { unref?: () => void }).unref === 'function')\n            (timeoutTimer as { unref: () => void }).unref()\n        }),\n      ])\n    } else {\n      artifact = await dispatched\n    }\n  } catch (err) {\n    errorMessage = err instanceof Error ? err.message : String(err)\n  } finally {\n    if (timeoutTimer) clearTimeout(timeoutTimer)\n    args.signal.removeEventListener('abort', onCampaignAbort)\n  }\n\n  // Run judges (only if we have an artifact). A judge that throws invalidates\n  // the cell — recorded as `error`, NOT folded into a fake composite:0 (a fake\n  // zero is indistinguishable from a real zero and poisons every aggregate).\n  const judgeScores: Record<string, JudgeScore> = {}\n  if (artifact !== undefined) {\n    for (const judge of args.opts.judges ?? []) {\n      if (judge.appliesTo && !judge.appliesTo(args.slot.scenario)) continue\n      try {\n        judgeScores[judge.name] = await runJudgeCell(judge, {\n          artifact,\n          scenario: args.slot.scenario,\n          signal: args.signal,\n        })\n      } catch (err) {\n        errorMessage = `judge '${judge.name}' failed: ${err instanceof Error ? err.message : String(err)}`\n        break\n      }\n    }\n  }\n\n  await trace.flush()\n\n  const cell: CampaignCellResult<TArtifact> = {\n    manifestHash: args.manifestHash,\n    cellId: args.slot.cellId,\n    scenarioId: args.slot.scenario.id,\n    rep: args.slot.rep,\n    artifact: (artifact ?? null) as TArtifact,\n    judgeScores,\n    costUsd: costSoFar,\n    tokenUsage: { ...tokensSoFar },\n    ...(resolvedModel ? { resolvedModel } : {}),\n    durationMs: Date.now() - startMs,\n    seed: args.slot.cellSeed,\n    cached: false,\n    error: errorMessage,\n  }\n\n  if (!errorMessage && args.resumable) {\n    storage.write(cachePath, JSON.stringify(cell))\n  }\n\n  return { cell, artifactsByPath }\n}\n\nexport interface CampaignRunPlanCell {\n  cellId: string\n  scenarioId: string\n  rep: number\n  seed: number\n  cachePath: string\n  status: 'cached' | 'run'\n  reason?: 'missing' | 'manifest-mismatch' | 'cell-mismatch' | 'corrupt' | 'resumable-off'\n}\n\nexport interface CampaignRunPlan {\n  manifestHash: string\n  totalCells: number\n  cellsCached: number\n  cellsToRun: number\n  cells: CampaignRunPlanCell[]\n}\n\nexport interface PlanCampaignRunOptions<TScenario extends Scenario, TArtifact> {\n  scenarios: TScenario[]\n  dispatch?: DispatchFn<TScenario, TArtifact>\n  dispatchRef?: string\n  judges?: JudgeConfig<TArtifact, TScenario>[]\n  seed?: number\n  reps?: number\n  resumable?: boolean\n  runDir: string\n  /** Subject repo for the shared run-dir root (see RunCampaignOptions.repo). */\n  repo?: string\n  storage?: CampaignStorage\n}\n\nexport function planCampaignRun<TScenario extends Scenario, TArtifact>(\n  opts: PlanCampaignRunOptions<TScenario, TArtifact>,\n): CampaignRunPlan {\n  const seed = opts.seed ?? 42\n  const reps = opts.reps ?? 1\n  const resumable = opts.resumable ?? true\n  const storage = opts.storage ?? fsCampaignStorage()\n\n  if (typeof opts.runDir !== 'string' || opts.runDir.trim().length === 0) {\n    throw new Error('planCampaignRun: runDir is required and must be a non-empty string')\n  }\n  opts.runDir = resolveRunDir(opts.runDir, opts.repo)\n\n  const manifestHash = computeManifestHash({\n    scenarios: opts.scenarios,\n    judges: (opts.judges ?? []) as unknown as JudgeConfig<unknown>[],\n    dispatchRef: dispatchRefFor(opts.dispatch, opts.dispatchRef),\n    seed,\n    reps,\n  })\n\n  const cells = buildCellSchedule(opts.scenarios, seed, reps).map((slot): CampaignRunPlanCell => {\n    const cachePath = join(\n      opts.runDir,\n      slot.cellId.replace(/[^a-zA-Z0-9_-]/g, '_'),\n      'cached-result.json',\n    )\n    if (!resumable) {\n      return {\n        cellId: slot.cellId,\n        scenarioId: slot.scenario.id,\n        rep: slot.rep,\n        seed: slot.cellSeed,\n        cachePath,\n        status: 'run',\n        reason: 'resumable-off',\n      }\n    }\n\n    const cached = readCachedCell<unknown>({\n      storage,\n      cachePath,\n      cellId: slot.cellId,\n      manifestHash,\n    })\n    if (cached.status === 'hit') {\n      return {\n        cellId: slot.cellId,\n        scenarioId: slot.scenario.id,\n        rep: slot.rep,\n        seed: slot.cellSeed,\n        cachePath,\n        status: 'cached',\n      }\n    }\n\n    return {\n      cellId: slot.cellId,\n      scenarioId: slot.scenario.id,\n      rep: slot.rep,\n      seed: slot.cellSeed,\n      cachePath,\n      status: 'run',\n      reason: cached.reason,\n    }\n  })\n\n  const cellsCached = cells.filter((cell) => cell.status === 'cached').length\n  return {\n    manifestHash,\n    totalCells: cells.length,\n    cellsCached,\n    cellsToRun: cells.length - cellsCached,\n    cells,\n  }\n}\n\n/**\n * Per-cell stub guard. A cell that produced an artifact (no error) but reported\n * `costUsd === 0` AND zero tokens means the dispatch never called `ctx.cost` —\n * i.e. it ran against a stub or silently dropped its usage. `'warn'` logs it,\n * `'assert'` throws (fail-fast), `'off'` skips. An errored/skipped cell or a\n * deterministic judge-only run that genuinely made no LLM call is not flagged.\n */\nfunction enforceCellUsage<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  mode: 'assert' | 'warn' | 'off',\n): void {\n  if (mode === 'off' || cell.error) return\n  if (cell.artifact === null || cell.artifact === undefined) return\n  const zeroTokens = cell.tokenUsage.input === 0 && cell.tokenUsage.output === 0\n  if (cell.costUsd !== 0 || !zeroTokens) return\n  const msg = `cell '${cell.cellId}' produced an artifact but reported zero cost and zero tokens — the dispatch never reported LLM usage via ctx.cost.observe/observeTokens (a stub cell)`\n  if (mode === 'assert') {\n    const report: BackendIntegrityReport = {\n      totalRecords: 1,\n      stubRecords: 1,\n      realRecords: 0,\n      uncostedRecords: 0,\n      totalInputTokens: 0,\n      totalOutputTokens: 0,\n      totalCostUsd: 0,\n      verdict: 'stub',\n      diagnosis: msg,\n    }\n    throw new BackendIntegrityError(`expectUsage: ${msg}`, report)\n  }\n  // eslint-disable-next-line no-console\n  console.warn(`[runCampaign] expectUsage: ${msg}`)\n}\n\nasync function runJudgeCell<TArtifact, TScenario extends Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  input: { artifact: TArtifact; scenario: TScenario; signal: AbortSignal },\n): Promise<JudgeScore> {\n  return judge.score(input)\n}\n\nfunction defaultBuildTraceWriter(\n  storage: CampaignStorage,\n): (cellId: string, dir: string) => CampaignTraceWriter {\n  return (cellId, dir) => {\n    const spans: Array<Record<string, unknown>> = []\n    return {\n      span(name, attributes) {\n        const startMs = Date.now()\n        const record: Record<string, unknown> = { name, cellId, startMs, ...(attributes ?? {}) }\n        const finish: TraceSpan = {\n          end(endAttrs) {\n            record.durationMs = Date.now() - startMs\n            if (endAttrs) Object.assign(record, endAttrs)\n            spans.push(record)\n          },\n          setAttribute(key, value) {\n            record[key] = value\n          },\n        }\n        return finish\n      },\n      async flush() {\n        storage.write(join(dir, 'spans.jsonl'), spans.map((s) => JSON.stringify(s)).join('\\n'))\n      },\n    }\n  }\n}\n\nfunction skippedCell<TScenario extends Scenario, TArtifact>(\n  slot: { scenario: TScenario; rep: number; cellId: string; cellSeed: number },\n  reason: string,\n): CampaignCellResult<TArtifact> {\n  return {\n    cellId: slot.cellId,\n    scenarioId: slot.scenario.id,\n    rep: slot.rep,\n    artifact: null as unknown as TArtifact,\n    judgeScores: {},\n    costUsd: 0,\n    tokenUsage: { input: 0, output: 0 },\n    durationMs: 0,\n    seed: slot.cellSeed,\n    cached: false,\n    error: `skipped: ${reason}`,\n  }\n}\n\nfunction buildCellSchedule<TScenario extends Scenario>(\n  scenarios: TScenario[],\n  seed: number,\n  reps: number,\n): Array<{ scenario: TScenario; rep: number; cellId: string; cellSeed: number }> {\n  const schedule: Array<{ scenario: TScenario; rep: number; cellId: string; cellSeed: number }> = []\n  let cellIndex = 0\n  for (const scenario of scenarios) {\n    for (let rep = 0; rep < reps; rep++) {\n      const cellId = `${scenario.id}:${rep}`\n      const cellSeed = seed + cellIndex\n      schedule.push({ scenario, rep, cellId, cellSeed })\n      cellIndex += 1\n    }\n  }\n  return schedule\n}\n\nfunction dispatchRefFor<TScenario extends Scenario, TArtifact>(\n  dispatch: DispatchFn<TScenario, TArtifact> | undefined,\n  override: string | undefined,\n): string {\n  const ref = override ?? dispatch?.name ?? 'anonymous'\n  if (typeof ref !== 'string' || ref.trim().length === 0) {\n    throw new Error('runCampaign: dispatchRef must be a non-empty string when provided')\n  }\n  return ref\n}\n\ntype CacheRead<TArtifact> =\n  | { status: 'hit'; cell: CampaignCellResult<TArtifact> }\n  | { status: 'miss'; reason: 'missing' | 'manifest-mismatch' | 'cell-mismatch' | 'corrupt' }\n\nfunction readCachedCell<TArtifact>(args: {\n  storage: CampaignStorage\n  cachePath: string\n  cellId: string\n  manifestHash: string\n}): CacheRead<TArtifact> {\n  const raw = args.storage.read(args.cachePath)\n  if (raw === undefined) return { status: 'miss', reason: 'missing' }\n\n  try {\n    const cached = JSON.parse(raw) as CampaignCellResult<TArtifact>\n    if (cached.cellId !== args.cellId) return { status: 'miss', reason: 'cell-mismatch' }\n    if (cached.manifestHash !== args.manifestHash) {\n      return { status: 'miss', reason: 'manifest-mismatch' }\n    }\n    return { status: 'hit', cell: cached }\n  } catch {\n    return { status: 'miss', reason: 'corrupt' }\n  }\n}\n\ninterface CaptureArgs<TScenario extends Scenario, TArtifact> {\n  store: LabeledScenarioStore\n  cell: CampaignCellResult<TArtifact>\n  scenario: TScenario\n  opts: RunCampaignOptions<TScenario, TArtifact>\n  now: () => Date\n}\n\nasync function captureToStore<TScenario extends Scenario, TArtifact>(\n  args: CaptureArgs<TScenario, TArtifact>,\n): Promise<void> {\n  await args.store.observe({\n    scenario: args.scenario,\n    artifact: args.cell.artifact,\n    judgeScores: args.cell.judgeScores,\n    source: args.opts.captureSource ?? 'eval-run',\n    sourceVersionHash: args.opts.captureSourceVersionHash ?? 'unknown',\n    capturedAt: args.now().toISOString(),\n    redactionStatus: 'raw',\n  })\n}\n\n// ── Aggregates + manifest hash ────────────────────────────────────────\n\nfunction computeManifestHash(input: {\n  scenarios: Scenario[]\n  judges: JudgeConfig<unknown>[]\n  dispatchRef: string\n  seed: number\n  reps: number\n}): string {\n  return contentHash({\n    scenarios: input.scenarios,\n    judges: input.judges.map((j) => ({ name: j.name, dims: j.dimensions })),\n    dispatch: input.dispatchRef,\n    seed: input.seed,\n    reps: input.reps,\n  })\n}\n\nfunction computeAggregates<TArtifact>(\n  cells: CampaignCellResult<TArtifact>[],\n  judges: JudgeConfig<TArtifact>[],\n  seed: number,\n): CampaignAggregates {\n  const byJudge: Record<string, JudgeAggregate> = {}\n  for (const judge of judges) {\n    const scores: number[] = []\n    for (const cell of cells) {\n      const s = cell.judgeScores[judge.name]\n      if (s !== undefined) scores.push(s.composite)\n    }\n    byJudge[judge.name] = aggregate(scores, seed)\n  }\n  const byScenario: Record<string, ScenarioAggregate> = {}\n  const scenarioGroups = new Map<string, number[]>()\n  for (const cell of cells) {\n    const composites = Object.values(cell.judgeScores).map((s) => s.composite)\n    if (composites.length === 0) continue\n    const mean = composites.reduce((a, b) => a + b, 0) / composites.length\n    const arr = scenarioGroups.get(cell.scenarioId) ?? []\n    arr.push(mean)\n    scenarioGroups.set(cell.scenarioId, arr)\n  }\n  for (const [scenarioId, samples] of scenarioGroups) {\n    const ag = aggregate(samples, seed)\n    byScenario[scenarioId] = { meanComposite: ag.mean, ci95: ag.ci95, n: ag.n }\n  }\n  return {\n    byJudge,\n    byScenario,\n    totalCostUsd: cells.reduce((a, c) => a + c.costUsd, 0),\n    cellsExecuted: cells.filter((c) => !c.error).length,\n    cellsSkipped: cells.filter((c) => c.error?.startsWith('skipped:')).length,\n    cellsCached: cells.filter((c) => c.cached).length,\n    cellsFailed: cells.filter((c) => c.error && !c.error.startsWith('skipped:')).length,\n  }\n}\n\n// Percentile bootstrap CI95 via seeded resampling. Deterministic for a given\n// seed — same campaign re-run produces identical CI bands. Falls back to\n// degenerate intervals at n<=1 (the bootstrap is undefined there).\nfunction aggregate(samples: number[], seed: number): JudgeAggregate {\n  const n = samples.length\n  if (n === 0) return { mean: 0, stdev: 0, ci95: [0, 0], n: 0 }\n  const mean = samples.reduce((a, b) => a + b, 0) / n\n  const variance = samples.reduce((a, b) => a + (b - mean) ** 2, 0) / Math.max(1, n - 1)\n  const stdev = Math.sqrt(variance)\n  const ci = confidenceInterval(samples, 0.95, { seed, resamples: 1000 })\n  return { mean, stdev, ci95: [ci.lower, ci.upper], n }\n}\n","/**\n * Backend-integrity guard: distinguish \"agent failed\" from \"eval ran against\n * a stub / unconfigured backend.\" Without this guard a canonical eval can\n * silently report `0/N passed` and look like an agent-quality problem when\n * the LLM was never actually called — the failure mode we just hit running\n * the 4-vertical parallel eval (legal-sandbox-stub returned hard-coded 33-104\n * char strings; gtm/creative defaulted to a cli-bridge that wasn't running).\n *\n * The shape:\n *\n *   const report = summarizeBackendIntegrity(records)\n *   assertRealBackend(records)   // throws BackendIntegrityError if 100% stub\n *\n * A record is \"stub-mode\" if its `tokenUsage.input === 0 && tokenUsage.output === 0`.\n * (`costUsd` alone is unreliable — some backends successfully call LLMs but\n *  don't propagate pricing, producing real tokens with $0 cost.)\n *\n * Verdicts:\n *   - `real`   — at least one record has nonzero token usage\n *   - `stub`   — every record is stub-mode (eval ran blind)\n *   - `mixed`  — some records real, some stub (partial backend failure;\n *                often the 429-cascade or auth-half-failed case)\n */\n\nimport { AgentEvalError } from '../errors'\nimport type { RunRecord } from '../run-record'\n\nexport interface BackendIntegrityReport {\n  /** Total records inspected. */\n  totalRecords: number\n  /** Records with input=0 AND output=0 (a stub fingerprint). */\n  stubRecords: number\n  /** Records with nonzero token usage (real LLM activity). */\n  realRecords: number\n  /** Records where output>0 but costUsd=0 (real LLM, broken cost ledger). */\n  uncostedRecords: number\n  /** Sum of input tokens across all records. */\n  totalInputTokens: number\n  /** Sum of output tokens across all records. */\n  totalOutputTokens: number\n  /** Sum of costUsd across all records. */\n  totalCostUsd: number\n  /** Worst-case integrity verdict. */\n  verdict: 'real' | 'mixed' | 'stub'\n  /** Human-readable diagnosis suitable for terminal output. */\n  diagnosis: string\n}\n\n/**\n * Error thrown when an integrity assertion fails. Caller can pattern-match\n * by `code === 'AGENT_EVAL_BACKEND_STUB'` to differentiate from other\n * errors.\n */\nexport class BackendIntegrityError extends AgentEvalError {\n  constructor(\n    message: string,\n    public readonly report: BackendIntegrityReport,\n  ) {\n    super('backend_integrity', message)\n  }\n}\n\nfunction isStubRecord(rec: RunRecord): boolean {\n  return rec.tokenUsage.input === 0 && rec.tokenUsage.output === 0\n}\n\nfunction isUncostedRecord(rec: RunRecord): boolean {\n  return rec.tokenUsage.output > 0 && rec.costUsd === 0\n}\n\n/**\n * Inspect a batch of RunRecords and return an integrity report. Pure\n * function — no I/O, no logging. The caller decides what to do with the\n * verdict (print warning, throw, gate CI, etc.).\n */\nexport function summarizeBackendIntegrity(\n  records: ReadonlyArray<RunRecord>,\n): BackendIntegrityReport {\n  const totalRecords = records.length\n  let stubRecords = 0\n  let realRecords = 0\n  let uncostedRecords = 0\n  let totalInputTokens = 0\n  let totalOutputTokens = 0\n  let totalCostUsd = 0\n  for (const rec of records) {\n    totalInputTokens += rec.tokenUsage.input\n    totalOutputTokens += rec.tokenUsage.output\n    totalCostUsd += rec.costUsd\n    if (isStubRecord(rec)) stubRecords++\n    else realRecords++\n    if (isUncostedRecord(rec)) uncostedRecords++\n  }\n  const verdict: BackendIntegrityReport['verdict'] =\n    totalRecords === 0\n      ? 'stub'\n      : stubRecords === totalRecords\n        ? 'stub'\n        : stubRecords === 0\n          ? 'real'\n          : 'mixed'\n  const diagnosis = buildDiagnosis({\n    totalRecords,\n    stubRecords,\n    realRecords,\n    uncostedRecords,\n    totalInputTokens,\n    totalOutputTokens,\n    totalCostUsd,\n    verdict,\n  })\n  return {\n    totalRecords,\n    stubRecords,\n    realRecords,\n    uncostedRecords,\n    totalInputTokens,\n    totalOutputTokens,\n    totalCostUsd,\n    verdict,\n    diagnosis,\n  }\n}\n\nfunction buildDiagnosis(r: Omit<BackendIntegrityReport, 'diagnosis'>): string {\n  if (r.totalRecords === 0) {\n    return 'no records — eval produced zero runs; backend likely failed before first turn'\n  }\n  if (r.verdict === 'stub') {\n    return [\n      `all ${r.totalRecords} records have zero token usage — the LLM backend was never called.`,\n      'common causes: --backend sandbox without a sandbox bridge running; stub model returning hard-coded strings;',\n      'auth misconfigured so requests were silently dropped before the LLM. Re-run with --backend tcloud and TANGLE_API_KEY set,',\n      'or boot the cli-bridge / sandbox before invoking the eval.',\n    ].join(' ')\n  }\n  if (r.verdict === 'mixed') {\n    const pct = ((r.stubRecords / r.totalRecords) * 100).toFixed(0)\n    return [\n      `${r.stubRecords}/${r.totalRecords} records (${pct}%) have zero token usage — the backend partially failed.`,\n      'common causes: rate-limit cascade (429s after the first N personas);',\n      'transient auth expiry mid-run; provider outage. Treat the affected records as missing data, not agent failures.',\n    ].join(' ')\n  }\n  // verdict === 'real'\n  if (r.uncostedRecords > 0) {\n    const pct = ((r.uncostedRecords / r.totalRecords) * 100).toFixed(0)\n    return [\n      `${r.totalRecords} records with real LLM activity (in=${r.totalInputTokens}, out=${r.totalOutputTokens} tokens).`,\n      `${r.uncostedRecords} (${pct}%) have output tokens but costUsd=0. Two distinct roots:`,\n      '(a) cost ledger mis-wired — no usage propagation from the runtime stream into RunRecord; or',\n      '(b) the model is unpriced at the source (sandbox/router returned $0 despite real tokens).',\n      'For (b), price the measured tokens against the substrate table (estimateCost) instead of leaving $0.',\n    ].join(' ')\n  }\n  return `${r.totalRecords} records with real LLM activity (in=${r.totalInputTokens}, out=${r.totalOutputTokens} tokens, $${r.totalCostUsd.toFixed(4)}).`\n}\n\n/**\n * Throw BackendIntegrityError if the verdict is 'stub' — i.e. every record\n * shows zero LLM activity. Non-strict callers can pass `{ allowMixed: false }`\n * to also reject mixed verdicts (recommended for CI gates).\n *\n * Real backends pass through silently.\n */\nexport function assertRealBackend(\n  records: ReadonlyArray<RunRecord>,\n  opts: { allowMixed?: boolean } = {},\n): BackendIntegrityReport {\n  const report = summarizeBackendIntegrity(records)\n  const allowMixed = opts.allowMixed ?? true\n  if (report.verdict === 'stub') {\n    throw new BackendIntegrityError(\n      `backend-integrity: ran against a stub or unconfigured backend — ${report.diagnosis}`,\n      report,\n    )\n  }\n  if (!allowMixed && report.verdict === 'mixed') {\n    throw new BackendIntegrityError(\n      `backend-integrity: partial backend failure rejected — ${report.diagnosis}`,\n      report,\n    )\n  }\n  return report\n}\n","/**\n * Content-addressed judge-verdict caching.\n *\n * LAW: cache JUDGE VERDICTS only — judging the same artifact with the same\n * judge+rubric is pure. NEVER cache agent rollouts. (A router that cached\n * identical fanout prompts silently destroyed best-of-N diversity; rollout\n * caching reintroduces that failure class. Judging has no diversity to\n * destroy — same artifact + same rubric ⇒ same verdict is the desired\n * property, not a bug.)\n *\n * The cache key is a sha-256 over the canonical JSON of everything that can\n * change a verdict: the artifact content, the scenario id, the judge name,\n * the full dimension list (key + description — the description IS the rubric\n * text shown to the judge), and a caller-supplied `judgeVersion`.\n * `judgeVersion` is REQUIRED: a judge whose prompt/model/ensemble changes\n * without a version bump would otherwise silently serve stale verdicts.\n *\n * Strict canonicalization (`canonicalJson`) throws on undefined / function /\n * symbol / non-finite numbers — an artifact that cannot be unambiguously\n * serialized cannot be content-addressed, and coercing it would let two\n * different artifacts collide on one key.\n */\n\nimport { createHash } from 'node:crypto'\nimport { appendFileSync, existsSync, readFileSync } from 'node:fs'\nimport type { JudgeConfig, JudgeScore, Scenario } from './campaign/types'\n\n// ── canonical JSON + content hash ─────────────────────────────────────────\n\nfunction canonicalizeAt(value: unknown, path: string): string {\n  if (value === null) return 'null'\n  switch (typeof value) {\n    case 'boolean':\n      return value ? 'true' : 'false'\n    case 'number':\n      if (!Number.isFinite(value)) {\n        throw new Error(\n          `canonicalJson: non-finite number (${value}) at ${path} — ambiguity is an error, not a coercion`,\n        )\n      }\n      return JSON.stringify(value)\n    case 'string':\n      return JSON.stringify(value)\n    case 'undefined':\n    case 'function':\n    case 'symbol':\n      throw new Error(\n        `canonicalJson: ${typeof value} at ${path} — ambiguity is an error, not a coercion`,\n      )\n    case 'bigint':\n      throw new Error(`canonicalJson: bigint at ${path} — not representable in JSON`)\n    case 'object':\n      break\n  }\n  const obj = value as Record<string, unknown>\n  // Honor toJSON (Date → ISO string) before structural checks — without it a\n  // Date would canonicalize to '{}' and every timestamp would collide.\n  if (typeof obj.toJSON === 'function') {\n    return canonicalizeAt((obj as { toJSON(): unknown }).toJSON(), path)\n  }\n  if (Array.isArray(obj)) {\n    return `[${obj.map((item, i) => canonicalizeAt(item, `${path}[${i}]`)).join(',')}]`\n  }\n  if (obj instanceof Map || obj instanceof Set) {\n    throw new Error(\n      `canonicalJson: ${obj instanceof Map ? 'Map' : 'Set'} at ${path} — would serialize as '{}'; convert to a plain object/array first`,\n    )\n  }\n  const keys = Object.keys(obj).sort()\n  const parts = keys.map((k) => `${JSON.stringify(k)}:${canonicalizeAt(obj[k], `${path}.${k}`)}`)\n  return `{${parts.join(',')}}`\n}\n\n/**\n * Stable JSON stringify: object keys sorted recursively, so two semantically\n * equal values produce byte-identical output regardless of key insertion\n * order. Throws on undefined / function / symbol / NaN / ±Infinity / bigint /\n * Map / Set — anything JSON.stringify would coerce or drop silently.\n *\n * Distinct from `pre-registration.ts`'s `canonicalize`/`hashJson`, which are\n * permissive (coercion allowed) and async (web-crypto). Use THIS pair when a\n * hash collision or silent coercion would corrupt a cache key or attestation.\n */\nexport function canonicalJson(value: unknown): string {\n  return canonicalizeAt(value, '$')\n}\n\n/** Hex sha-256 over `canonicalJson(value)`. The content address used by the\n *  verdict cache and report attestation. */\nexport function contentHash(value: unknown): string {\n  return createHash('sha256').update(canonicalJson(value)).digest('hex')\n}\n\n// ── store contract ─────────────────────────────────────────────────────────\n\n/** Pluggable verdict store. Sync or async on both legs — `cachedJudge`\n *  awaits the results either way. */\nexport interface VerdictCacheStore {\n  get(key: string): Promise<JudgeScore | undefined> | JudgeScore | undefined\n  set(key: string, score: JudgeScore): Promise<void> | void\n}\n\n/** Process-local Map-backed store. */\nexport function inMemoryVerdictCache(): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      entries.set(key, score)\n    },\n  }\n}\n\ninterface VerdictCacheLine {\n  key: string\n  score: JudgeScore\n}\n\nfunction parseCacheLine(line: string, path: string, lineNo: number): VerdictCacheLine {\n  let parsed: unknown\n  try {\n    parsed = JSON.parse(line)\n  } catch (err) {\n    throw new Error(\n      `fileVerdictCache: corrupt JSONL at ${path}:${lineNo} — ${err instanceof Error ? err.message : String(err)}`,\n    )\n  }\n  const rec = parsed as Partial<VerdictCacheLine>\n  if (\n    typeof rec !== 'object' ||\n    rec === null ||\n    typeof rec.key !== 'string' ||\n    typeof rec.score !== 'object' ||\n    rec.score === null ||\n    typeof rec.score.composite !== 'number' ||\n    typeof rec.score.dimensions !== 'object'\n  ) {\n    throw new Error(\n      `fileVerdictCache: invalid record shape at ${path}:${lineNo} — expected {key, score:{dimensions, composite, notes}}`,\n    )\n  }\n  return rec as VerdictCacheLine\n}\n\n/**\n * JSONL-file-backed store: the full file is loaded into an in-memory index at\n * construction; every `set` appends one line synchronously (durable before\n * the verdict is returned). A corrupt or malformed line throws at load with\n * file:line — a skipped line would silently re-judge (cost) or, worse, mask\n * a half-written file that needs operator attention.\n */\nexport function fileVerdictCache(path: string): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  if (existsSync(path)) {\n    const lines = readFileSync(path, 'utf8').split('\\n')\n    for (let i = 0; i < lines.length; i++) {\n      const line = lines[i]\n      if (line === undefined || line.trim() === '') continue\n      const rec = parseCacheLine(line, path, i + 1)\n      entries.set(rec.key, rec.score)\n    }\n  }\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      appendFileSync(path, `${JSON.stringify({ key, score })}\\n`, 'utf8')\n      entries.set(key, score)\n    },\n  }\n}\n\n// ── cached judge wrapper ───────────────────────────────────────────────────\n\nexport interface VerdictCacheStats {\n  hits: number\n  misses: number\n}\n\nexport interface CachedJudgeOptions {\n  /** REQUIRED — part of the cache key. Bump on any change to the judge's\n   *  prompt, model, ensemble, or scoring logic; silent judge upgrades must\n   *  never serve stale verdicts. */\n  judgeVersion: string\n}\n\n/** The wrapped judge: same `JudgeConfig` seam, plus hit/miss observability. */\nexport type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig<\n  TArtifact,\n  TScenario\n> & {\n  stats(): VerdictCacheStats\n}\n\n/**\n * Wrap a `JudgeConfig` so repeat judgments of the same artifact are served\n * from the store instead of re-invoking `score()`. The wrapper is generic\n * over the judge's own type parameters and preserves `appliesTo` — it is a\n * drop-in replacement anywhere a `JudgeConfig` is accepted.\n *\n * A judge that throws is NOT cached: the error propagates and the next\n * attempt re-judges (caching a failure would pin a transient outage forever).\n */\nexport function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  store: VerdictCacheStore,\n  options: CachedJudgeOptions,\n): CachedJudge<TArtifact, TScenario> {\n  if (typeof options.judgeVersion !== 'string' || options.judgeVersion.trim() === '') {\n    throw new Error('cachedJudge: judgeVersion is required and must be a non-empty string')\n  }\n  const stats: VerdictCacheStats = { hits: 0, misses: 0 }\n  const wrapped: CachedJudge<TArtifact, TScenario> = {\n    name: judge.name,\n    dimensions: judge.dimensions,\n    async score(input) {\n      const key = contentHash({\n        artifact: canonicalJson(input.artifact),\n        scenarioId: input.scenario.id,\n        judgeName: judge.name,\n        dimensions: judge.dimensions,\n        judgeVersion: options.judgeVersion,\n      })\n      const cached = await store.get(key)\n      if (cached !== undefined) {\n        stats.hits += 1\n        return cached\n      }\n      const score = await judge.score(input)\n      await store.set(key, score)\n      stats.misses += 1\n      return score\n    },\n    stats: () => ({ ...stats }),\n  }\n  if (judge.appliesTo) wrapped.appliesTo = judge.appliesTo\n  return wrapped\n}\n","import { homedir } from 'node:os'\nimport { basename, isAbsolute, join } from 'node:path'\n\n/** The shared, out-of-repo root for campaign/benchmark run bundles. Keeping run\n *  outputs here means they never land in a repo working tree (no per-repo\n *  gitignore, no clutter, no accidental commits). Layout:\n *    ~/.tangle/traces/<repo>/runs/<runName>/\n *  where <repo> disambiguates runs across repos in one place. */\nexport function tangleTracesRoot(): string {\n  return join(homedir(), '.tangle', 'traces')\n}\n\n/** Resolve a campaign `runDir`. An absolute path is honored as-is (the caller\n *  chose an explicit location). A bare name is placed under the shared home root\n *  so bundles never pollute a repo working tree — the default the harness should\n *  compute so callers pass a *name*, not a path. */\nexport function resolveRunDir(runDir: string, repo?: string): string {\n  if (isAbsolute(runDir)) return runDir\n  const r = repo && repo.trim().length > 0 ? repo : basename(process.cwd())\n  return join(tangleTracesRoot(), r, 'runs', runDir)\n}\n","import { createRequire } from 'node:module'\n\n/**\n * `CampaignStorage` — the filesystem seam `runCampaign` writes through\n * (run/cell dirs, the resumability cache, per-cell artifacts, trace spans).\n *\n * The default (`fsCampaignStorage`) is the Node filesystem — identical\n * behavior to the inline `node:fs` calls it replaces, so existing CLI\n * consumers are unaffected. `inMemoryCampaignStorage` keeps everything in a\n * `Map`, so the substrate runs in environments WITHOUT a filesystem\n * (Cloudflare Workers, Deno Deploy, other edge runtimes) — the campaign\n * still produces its `CampaignResult` (cells + aggregates) in memory;\n * artifacts/traces simply aren't persisted to disk.\n *\n * Paths are opaque keys to the in-memory adapter — it does not parse them,\n * so the same `join(...)`-built paths work unchanged across both adapters.\n */\nexport interface CampaignStorage {\n  /** Ensure a directory exists (recursive). No-op for in-memory. */\n  ensureDir(dir: string): void\n  /** Does this path exist (as a written file or an ensured dir)? */\n  exists(path: string): boolean\n  /** Read a UTF-8 file; `undefined` when missing or unreadable. */\n  read(path: string): string | undefined\n  /** Write a file (string or bytes). Parent dir is assumed ensured. */\n  write(path: string, content: string | Uint8Array): void\n}\n\n/** Node-filesystem storage — the default. Lazily requires `node:fs` so the\n *  module imports cleanly in non-Node runtimes (where the caller passes\n *  `inMemoryCampaignStorage` instead and never constructs this).\n *\n *  `createRequire(import.meta.url)` is the ESM-native lazy require — a bare\n *  `require` is a ReferenceError under `\"type\": \"module\"`, which is exactly\n *  the shape this package publishes. */\nexport function fsCampaignStorage(): CampaignStorage {\n  const nodeRequire = createRequire(import.meta.url)\n  const { existsSync, mkdirSync, readFileSync, writeFileSync } = nodeRequire(\n    'node:fs',\n  ) as typeof import('node:fs')\n  return {\n    ensureDir(dir) {\n      if (!existsSync(dir)) mkdirSync(dir, { recursive: true })\n    },\n    exists(path) {\n      return existsSync(path)\n    },\n    read(path) {\n      try {\n        return readFileSync(path, 'utf8')\n      } catch {\n        return undefined\n      }\n    },\n    write(path, content) {\n      writeFileSync(path, content as Uint8Array)\n    },\n  }\n}\n\n/** In-memory storage for filesystem-less runtimes. Artifacts + trace spans\n *  live in a `Map` for the duration of the run; the `CampaignResult` is\n *  fully populated, but nothing is persisted to disk. */\nexport function inMemoryCampaignStorage(): CampaignStorage {\n  const files = new Map<string, string | Uint8Array>()\n  const dirs = new Set<string>()\n  return {\n    ensureDir(dir) {\n      dirs.add(dir)\n    },\n    exists(path) {\n      return files.has(path) || dirs.has(path)\n    },\n    read(path) {\n      const value = files.get(path)\n      if (value === undefined) return undefined\n      return typeof value === 'string' ? value : new TextDecoder().decode(value)\n    },\n    write(path, content) {\n      files.set(path, content)\n    },\n  }\n}\n"],"mappings":";;;;;;;;AAWA,SAAS,QAAAA,aAAY;;;AC0Cd,IAAM,wBAAN,cAAoC,eAAe;AAAA,EACxD,YACE,SACgB,QAChB;AACA,UAAM,qBAAqB,OAAO;AAFlB;AAAA,EAGlB;AAAA,EAHkB;AAIpB;AAEA,SAAS,aAAa,KAAyB;AAC7C,SAAO,IAAI,WAAW,UAAU,KAAK,IAAI,WAAW,WAAW;AACjE;AAEA,SAAS,iBAAiB,KAAyB;AACjD,SAAO,IAAI,WAAW,SAAS,KAAK,IAAI,YAAY;AACtD;AAOO,SAAS,0BACd,SACwB;AACxB,QAAM,eAAe,QAAQ;AAC7B,MAAI,cAAc;AAClB,MAAI,cAAc;AAClB,MAAI,kBAAkB;AACtB,MAAI,mBAAmB;AACvB,MAAI,oBAAoB;AACxB,MAAI,eAAe;AACnB,aAAW,OAAO,SAAS;AACzB,wBAAoB,IAAI,WAAW;AACnC,yBAAqB,IAAI,WAAW;AACpC,oBAAgB,IAAI;AACpB,QAAI,aAAa,GAAG,EAAG;AAAA,QAClB;AACL,QAAI,iBAAiB,GAAG,EAAG;AAAA,EAC7B;AACA,QAAM,UACJ,iBAAiB,IACb,SACA,gBAAgB,eACd,SACA,gBAAgB,IACd,SACA;AACV,QAAM,YAAY,eAAe;AAAA,IAC/B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AACD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,eAAe,GAAsD;AAC5E,MAAI,EAAE,iBAAiB,GAAG;AACxB,WAAO;AAAA,EACT;AACA,MAAI,EAAE,YAAY,QAAQ;AACxB,WAAO;AAAA,MACL,OAAO,EAAE,YAAY;AAAA,MACrB;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,GAAG;AAAA,EACZ;AACA,MAAI,EAAE,YAAY,SAAS;AACzB,UAAM,OAAQ,EAAE,cAAc,EAAE,eAAgB,KAAK,QAAQ,CAAC;AAC9D,WAAO;AAAA,MACL,GAAG,EAAE,WAAW,IAAI,EAAE,YAAY,aAAa,GAAG;AAAA,MAClD;AAAA,MACA;AAAA,IACF,EAAE,KAAK,GAAG;AAAA,EACZ;AAEA,MAAI,EAAE,kBAAkB,GAAG;AACzB,UAAM,OAAQ,EAAE,kBAAkB,EAAE,eAAgB,KAAK,QAAQ,CAAC;AAClE,WAAO;AAAA,MACL,GAAG,EAAE,YAAY,uCAAuC,EAAE,gBAAgB,SAAS,EAAE,iBAAiB;AAAA,MACtG,GAAG,EAAE,eAAe,KAAK,GAAG;AAAA,MAC5B;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,GAAG;AAAA,EACZ;AACA,SAAO,GAAG,EAAE,YAAY,uCAAuC,EAAE,gBAAgB,SAAS,EAAE,iBAAiB,aAAa,EAAE,aAAa,QAAQ,CAAC,CAAC;AACrJ;AASO,SAAS,kBACd,SACA,OAAiC,CAAC,GACV;AACxB,QAAM,SAAS,0BAA0B,OAAO;AAChD,QAAM,aAAa,KAAK,cAAc;AACtC,MAAI,OAAO,YAAY,QAAQ;AAC7B,UAAM,IAAI;AAAA,MACR,wEAAmE,OAAO,SAAS;AAAA,MACnF;AAAA,IACF;AAAA,EACF;AACA,MAAI,CAAC,cAAc,OAAO,YAAY,SAAS;AAC7C,UAAM,IAAI;AAAA,MACR,8DAAyD,OAAO,SAAS;AAAA,MACzE;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;;;ACjKA,SAAS,kBAAkB;AAC3B,SAAS,gBAAgB,YAAY,oBAAoB;AAKzD,SAAS,eAAe,OAAgB,MAAsB;AAC5D,MAAI,UAAU,KAAM,QAAO;AAC3B,UAAQ,OAAO,OAAO;AAAA,IACpB,KAAK;AACH,aAAO,QAAQ,SAAS;AAAA,IAC1B,KAAK;AACH,UAAI,CAAC,OAAO,SAAS,KAAK,GAAG;AAC3B,cAAM,IAAI;AAAA,UACR,qCAAqC,KAAK,QAAQ,IAAI;AAAA,QACxD;AAAA,MACF;AACA,aAAO,KAAK,UAAU,KAAK;AAAA,IAC7B,KAAK;AACH,aAAO,KAAK,UAAU,KAAK;AAAA,IAC7B,KAAK;AAAA,IACL,KAAK;AAAA,IACL,KAAK;AACH,YAAM,IAAI;AAAA,QACR,kBAAkB,OAAO,KAAK,OAAO,IAAI;AAAA,MAC3C;AAAA,IACF,KAAK;AACH,YAAM,IAAI,MAAM,4BAA4B,IAAI,mCAA8B;AAAA,IAChF,KAAK;AACH;AAAA,EACJ;AACA,QAAM,MAAM;AAGZ,MAAI,OAAO,IAAI,WAAW,YAAY;AACpC,WAAO,eAAgB,IAA8B,OAAO,GAAG,IAAI;AAAA,EACrE;AACA,MAAI,MAAM,QAAQ,GAAG,GAAG;AACtB,WAAO,IAAI,IAAI,IAAI,CAAC,MAAM,MAAM,eAAe,MAAM,GAAG,IAAI,IAAI,CAAC,GAAG,CAAC,EAAE,KAAK,GAAG,CAAC;AAAA,EAClF;AACA,MAAI,eAAe,OAAO,eAAe,KAAK;AAC5C,UAAM,IAAI;AAAA,MACR,kBAAkB,eAAe,MAAM,QAAQ,KAAK,OAAO,IAAI;AAAA,IACjE;AAAA,EACF;AACA,QAAM,OAAO,OAAO,KAAK,GAAG,EAAE,KAAK;AACnC,QAAM,QAAQ,KAAK,IAAI,CAAC,MAAM,GAAG,KAAK,UAAU,CAAC,CAAC,IAAI,eAAe,IAAI,CAAC,GAAG,GAAG,IAAI,IAAI,CAAC,EAAE,CAAC,EAAE;AAC9F,SAAO,IAAI,MAAM,KAAK,GAAG,CAAC;AAC5B;AAYO,SAAS,cAAc,OAAwB;AACpD,SAAO,eAAe,OAAO,GAAG;AAClC;AAIO,SAAS,YAAY,OAAwB;AAClD,SAAO,WAAW,QAAQ,EAAE,OAAO,cAAc,KAAK,CAAC,EAAE,OAAO,KAAK;AACvE;AAYO,SAAS,uBAA0C;AACxD,QAAM,UAAU,oBAAI,IAAwB;AAC5C,SAAO;AAAA,IACL,KAAK,CAAC,QAAQ,QAAQ,IAAI,GAAG;AAAA,IAC7B,KAAK,CAAC,KAAK,UAAU;AACnB,cAAQ,IAAI,KAAK,KAAK;AAAA,IACxB;AAAA,EACF;AACF;AAOA,SAAS,eAAe,MAAc,MAAc,QAAkC;AACpF,MAAI;AACJ,MAAI;AACF,aAAS,KAAK,MAAM,IAAI;AAAA,EAC1B,SAAS,KAAK;AACZ,UAAM,IAAI;AAAA,MACR,sCAAsC,IAAI,IAAI,MAAM,WAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IAC5G;AAAA,EACF;AACA,QAAM,MAAM;AACZ,MACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAO,IAAI,QAAQ,YACnB,OAAO,IAAI,UAAU,YACrB,IAAI,UAAU,QACd,OAAO,IAAI,MAAM,cAAc,YAC/B,OAAO,IAAI,MAAM,eAAe,UAChC;AACA,UAAM,IAAI;AAAA,MACR,6CAA6C,IAAI,IAAI,MAAM;AAAA,IAC7D;AAAA,EACF;AACA,SAAO;AACT;AASO,SAAS,iBAAiB,MAAiC;AAChE,QAAM,UAAU,oBAAI,IAAwB;AAC5C,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI;AACnD,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,YAAM,OAAO,MAAM,CAAC;AACpB,UAAI,SAAS,UAAa,KAAK,KAAK,MAAM,GAAI;AAC9C,YAAM,MAAM,eAAe,MAAM,MAAM,IAAI,CAAC;AAC5C,cAAQ,IAAI,IAAI,KAAK,IAAI,KAAK;AAAA,IAChC;AAAA,EACF;AACA,SAAO;AAAA,IACL,KAAK,CAAC,QAAQ,QAAQ,IAAI,GAAG;AAAA,IAC7B,KAAK,CAAC,KAAK,UAAU;AACnB,qBAAe,MAAM,GAAG,KAAK,UAAU,EAAE,KAAK,MAAM,CAAC,CAAC;AAAA,GAAM,MAAM;AAClE,cAAQ,IAAI,KAAK,KAAK;AAAA,IACxB;AAAA,EACF;AACF;AAiCO,SAAS,YACd,OACA,OACA,SACmC;AACnC,MAAI,OAAO,QAAQ,iBAAiB,YAAY,QAAQ,aAAa,KAAK,MAAM,IAAI;AAClF,UAAM,IAAI,MAAM,sEAAsE;AAAA,EACxF;AACA,QAAM,QAA2B,EAAE,MAAM,GAAG,QAAQ,EAAE;AACtD,QAAM,UAA6C;AAAA,IACjD,MAAM,MAAM;AAAA,IACZ,YAAY,MAAM;AAAA,IAClB,MAAM,MAAM,OAAO;AACjB,YAAM,MAAM,YAAY;AAAA,QACtB,UAAU,cAAc,MAAM,QAAQ;AAAA,QACtC,YAAY,MAAM,SAAS;AAAA,QAC3B,WAAW,MAAM;AAAA,QACjB,YAAY,MAAM;AAAA,QAClB,cAAc,QAAQ;AAAA,MACxB,CAAC;AACD,YAAM,SAAS,MAAM,MAAM,IAAI,GAAG;AAClC,UAAI,WAAW,QAAW;AACxB,cAAM,QAAQ;AACd,eAAO;AAAA,MACT;AACA,YAAM,QAAQ,MAAM,MAAM,MAAM,KAAK;AACrC,YAAM,MAAM,IAAI,KAAK,KAAK;AAC1B,YAAM,UAAU;AAChB,aAAO;AAAA,IACT;AAAA,IACA,OAAO,OAAO,EAAE,GAAG,MAAM;AAAA,EAC3B;AACA,MAAI,MAAM,UAAW,SAAQ,YAAY,MAAM;AAC/C,SAAO;AACT;;;AC5OA,SAAS,eAAe;AACxB,SAAS,UAAU,YAAY,YAAY;AAOpC,SAAS,mBAA2B;AACzC,SAAO,KAAK,QAAQ,GAAG,WAAW,QAAQ;AAC5C;AAMO,SAAS,cAAc,QAAgB,MAAuB;AACnE,MAAI,WAAW,MAAM,EAAG,QAAO;AAC/B,QAAM,IAAI,QAAQ,KAAK,KAAK,EAAE,SAAS,IAAI,OAAO,SAAS,QAAQ,IAAI,CAAC;AACxE,SAAO,KAAK,iBAAiB,GAAG,GAAG,QAAQ,MAAM;AACnD;;;ACpBA,SAAS,qBAAqB;AAmCvB,SAAS,oBAAqC;AACnD,QAAM,cAAc,cAAc,YAAY,GAAG;AACjD,QAAM,EAAE,YAAAC,aAAY,WAAW,cAAAC,eAAc,cAAc,IAAI;AAAA,IAC7D;AAAA,EACF;AACA,SAAO;AAAA,IACL,UAAU,KAAK;AACb,UAAI,CAACD,YAAW,GAAG,EAAG,WAAU,KAAK,EAAE,WAAW,KAAK,CAAC;AAAA,IAC1D;AAAA,IACA,OAAO,MAAM;AACX,aAAOA,YAAW,IAAI;AAAA,IACxB;AAAA,IACA,KAAK,MAAM;AACT,UAAI;AACF,eAAOC,cAAa,MAAM,MAAM;AAAA,MAClC,QAAQ;AACN,eAAO;AAAA,MACT;AAAA,IACF;AAAA,IACA,MAAM,MAAM,SAAS;AACnB,oBAAc,MAAM,OAAqB;AAAA,IAC3C;AAAA,EACF;AACF;AAKO,SAAS,0BAA2C;AACzD,QAAM,QAAQ,oBAAI,IAAiC;AACnD,QAAM,OAAO,oBAAI,IAAY;AAC7B,SAAO;AAAA,IACL,UAAU,KAAK;AACb,WAAK,IAAI,GAAG;AAAA,IACd;AAAA,IACA,OAAO,MAAM;AACX,aAAO,MAAM,IAAI,IAAI,KAAK,KAAK,IAAI,IAAI;AAAA,IACzC;AAAA,IACA,KAAK,MAAM;AACT,YAAM,QAAQ,MAAM,IAAI,IAAI;AAC5B,UAAI,UAAU,OAAW,QAAO;AAChC,aAAO,OAAO,UAAU,WAAW,QAAQ,IAAI,YAAY,EAAE,OAAO,KAAK;AAAA,IAC3E;AAAA,IACA,MAAM,MAAM,SAAS;AACnB,YAAM,IAAI,MAAM,OAAO;AAAA,IACzB;AAAA,EACF;AACF;;;AJ6CA,eAAsB,YACpB,MAC+C;AAC/C,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,iBAAiB,KAAK,kBAAkB;AAC9C,QAAM,MAAM,KAAK,QAAQ,MAAM,oBAAI,KAAK;AACxC,QAAM,SAAS,KAAK,UAAU,CAAC;AAC/B,QAAM,UAAU,KAAK,WAAW,kBAAkB;AAElD,MAAI,OAAO,KAAK,WAAW,YAAY,KAAK,OAAO,KAAK,EAAE,WAAW,GAAG;AACtE,UAAM,IAAI,MAAM,gEAAgE;AAAA,EAClF;AACA,OAAK,SAAS,cAAc,KAAK,QAAQ,KAAK,IAAI;AAClD,UAAQ,UAAU,KAAK,MAAM;AAE7B,QAAM,eAAe,oBAAoB;AAAA,IACvC,WAAW,KAAK;AAAA,IAChB;AAAA,IACA,aAAa,eAAe,KAAK,UAAU,KAAK,WAAW;AAAA,IAC3D;AAAA,IACA;AAAA,EACF,CAAC;AAED,QAAM,YAAY,IAAI;AACtB,QAAM,QAAyC,CAAC;AAChD,QAAM,kBAA0C,CAAC;AAGjD,QAAM,WAAW,kBAAkB,KAAK,WAAW,MAAM,IAAI;AAG7D,MAAI,eAAe;AACnB,MAAI,qBAAqB;AACzB,QAAM,kBAAkB,IAAI,gBAAgB;AAI5C,QAAM,QAAyB,CAAC;AAChC,MAAI,UAAU;AACd,QAAM,WAAW;AAEjB,WAAS,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,UAAM;AAAA,OACH,YAAY;AACX,eAAO,MAAM;AACX,gBAAM,QAAQ;AACd,cAAI,SAAS,SAAS,OAAQ;AAC9B,gBAAM,OAAO,SAAS,KAAK;AAC3B,cAAI,oBAAoB;AACtB,qBAAS,KAAK,YAAY,MAAM,sBAAsB,CAAC;AACvD;AAAA,UACF;AACA,gBAAM,SAAS,MAAM,YAAY;AAAA,YAC/B;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA,kBAAkB,KAAK,oBAAoB,wBAAwB,OAAO;AAAA,YAC1E,QAAQ,gBAAgB;AAAA,YACxB,mBAAmB,KAAK;AAAA,UAC1B,CAAC;AACD,mBAAS,KAAK,OAAO,IAAI;AACzB,2BAAiB,OAAO,MAAM,KAAK,eAAe,MAAM;AACxD,0BAAgB,OAAO,KAAK;AAC5B,iBAAO,OAAO,iBAAiB,OAAO,eAAe;AACrD,cAAI,KAAK,gBAAgB,UAAa,gBAAgB,KAAK,aAAa;AACtE,iCAAqB;AAAA,UACvB;AAEA,cAAI,KAAK,gBAAgB,KAAK,iBAAiB,SAAS,CAAC,OAAO,KAAK,OAAO;AAC1E,kBAAM,eAAe;AAAA,cACnB,OAAO,KAAK;AAAA,cACZ,MAAM,OAAO;AAAA,cACb,UAAU,KAAK;AAAA,cACf;AAAA,cACA;AAAA,YACF,CAAC,EAAE,MAAM,CAAC,QAAQ;AAGhB,sBAAQ;AAAA,gBACN,oCAAoC,OAAO,KAAK,MAAM,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,cAC7G;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF,GAAG;AAAA,IACL;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,KAAK;AAEvB,QAAM,UAAU,IAAI;AACpB,WAAS,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,cAAc,EAAE,MAAM,CAAC;AAExD,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,WAAW,UAAU,YAAY;AAAA,IACjC,SAAS,QAAQ,YAAY;AAAA,IAC7B,YAAY,QAAQ,QAAQ,IAAI,UAAU,QAAQ;AAAA,IAClD,OAAO;AAAA,IACP;AAAA,IACA,QAAQ,KAAK;AAAA,IACb;AAAA,IACA,WAAW,KAAK,UAAU,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,MAAM,EAAE,KAAK,EAAE;AAAA,EACnE;AACF;AAgBA,eAAe,YACb,MAC2F;AAC3F,QAAM,UAAU,KAAK;AACrB,QAAM,UAAUC,MAAK,KAAK,KAAK,QAAQ,KAAK,KAAK,OAAO,QAAQ,mBAAmB,GAAG,CAAC;AACvF,UAAQ,UAAU,OAAO;AAGzB,QAAM,YAAYA,MAAK,SAAS,oBAAoB;AACpD,MAAI,KAAK,WAAW;AAClB,UAAM,SAAS,eAA0B;AAAA,MACvC;AAAA,MACA;AAAA,MACA,QAAQ,KAAK,KAAK;AAAA,MAClB,cAAc,KAAK;AAAA,IACrB,CAAC;AACD,QAAI,OAAO,WAAW,OAAO;AAC3B,aAAO,EAAE,MAAM,EAAE,GAAG,OAAO,MAAM,QAAQ,KAAK,GAAG,iBAAiB,CAAC,EAAE;AAAA,IACvE;AAAA,EACF;AAEA,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,QAAQ,KAAK,iBAAiB,KAAK,KAAK,QAAQ,OAAO;AAC7D,QAAM,kBAA0C,CAAC;AACjD,QAAM,YAAoC;AAAA,IACxC,MAAM,MAAM,MAAM,SAAS;AACzB,YAAM,WAAWA,MAAK,SAAS,IAAI;AACnC,cAAQ,UAAUA,MAAK,UAAU,IAAI,CAAC;AACtC,cAAQ,MAAM,UAAU,OAAO;AAC/B,sBAAgB,GAAG,KAAK,KAAK,MAAM,IAAI,IAAI,EAAE,IAAI;AACjD,aAAO;AAAA,IACT;AAAA,IACA,MAAM,UAAU,MAAM,OAAO;AAC3B,aAAO,UAAU,MAAM,MAAM,KAAK,UAAU,OAAO,MAAM,CAAC,CAAC;AAAA,IAC7D;AAAA,EACF;AACA,MAAI,YAAY;AAChB,QAAM,cAAkC,EAAE,OAAO,GAAG,QAAQ,EAAE;AAC9D,MAAI;AACJ,QAAM,OAA0B;AAAA,IAC9B,QAAQ,QAAQ,QAAQ;AACtB,mBAAa;AACb,YAAM,KAAK,QAAQ,MAAM,IAAI,EAAE,WAAW,OAAO,CAAC,EAAE,IAAI;AAAA,IAC1D;AAAA,IACA,cAAc,OAAO;AACnB,kBAAY,SAAS,MAAM;AAC3B,kBAAY,UAAU,MAAM;AAC5B,UAAI,MAAM,OAAQ,aAAY,UAAU,YAAY,UAAU,KAAK,MAAM;AAAA,IAC3E;AAAA,IACA,aAAa,OAAO;AAClB,YAAM,UAAU,OAAO,KAAK;AAC5B,UAAI,QAAS,iBAAgB;AAAA,IAC/B;AAAA,IACA,UAAU;AACR,aAAO;AAAA,IACT;AAAA,IACA,SAAS;AACP,aAAO,EAAE,GAAG,YAAY;AAAA,IAC1B;AAAA,IACA,gBAAgB;AACd,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,YAAY,KAAK,KAAK,gBAAgB;AAAA,IAC1C,UAAU,KAAK,KAAK;AAAA,IACpB,KAAK,KAAK,KAAK;AAAA,EACjB,CAAC;AAMD,QAAM,YAAY,IAAI,gBAAgB;AACtC,QAAM,kBAAkB,MAAM,UAAU,MAAO,KAAK,OAAgC,MAAM;AAC1F,MAAI,KAAK,OAAO,QAAS,WAAU,MAAO,KAAK,OAAgC,MAAM;AAAA,MAChF,MAAK,OAAO,iBAAiB,SAAS,iBAAiB,EAAE,MAAM,KAAK,CAAC;AAE1E,QAAM,MAAuB;AAAA,IAC3B,QAAQ,KAAK,KAAK;AAAA,IAClB,KAAK,KAAK,KAAK;AAAA,IACf,MAAM,KAAK,KAAK;AAAA,IAChB,QAAQ,UAAU;AAAA,IAClB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,MAAI;AACJ,MAAI;AACJ,QAAM,YAAY,KAAK;AACvB,MAAI;AACJ,MAAI;AACF,UAAM,aAAa,KAAK,KAAK,SAAS,KAAK,KAAK,UAAU,GAAG;AAC7D,QAAI,cAAc,UAAa,YAAY,GAAG;AAK5C,iBAAW,MAAM,QAAQ,KAAK;AAAA,QAC5B;AAAA,QACA,IAAI,QAAe,CAAC,GAAG,WAAW;AAChC,yBAAe,WAAW,MAAM;AAC9B,sBAAU,MAAM,IAAI,MAAM,kBAAkB,CAAC;AAC7C;AAAA,cACE,IAAI;AAAA,gBACF,qBAAqB,SAAS,gBAAgB,KAAK,KAAK,MAAM;AAAA,cAChE;AAAA,YACF;AAAA,UACF,GAAG,SAAS;AACZ,cAAI,OAAQ,aAAwC,UAAU;AAC5D,YAAC,aAAuC,MAAM;AAAA,QAClD,CAAC;AAAA,MACH,CAAC;AAAA,IACH,OAAO;AACL,iBAAW,MAAM;AAAA,IACnB;AAAA,EACF,SAAS,KAAK;AACZ,mBAAe,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,EAChE,UAAE;AACA,QAAI,aAAc,cAAa,YAAY;AAC3C,SAAK,OAAO,oBAAoB,SAAS,eAAe;AAAA,EAC1D;AAKA,QAAM,cAA0C,CAAC;AACjD,MAAI,aAAa,QAAW;AAC1B,eAAW,SAAS,KAAK,KAAK,UAAU,CAAC,GAAG;AAC1C,UAAI,MAAM,aAAa,CAAC,MAAM,UAAU,KAAK,KAAK,QAAQ,EAAG;AAC7D,UAAI;AACF,oBAAY,MAAM,IAAI,IAAI,MAAM,aAAa,OAAO;AAAA,UAClD;AAAA,UACA,UAAU,KAAK,KAAK;AAAA,UACpB,QAAQ,KAAK;AAAA,QACf,CAAC;AAAA,MACH,SAAS,KAAK;AACZ,uBAAe,UAAU,MAAM,IAAI,aAAa,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAChG;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,MAAM,MAAM;AAElB,QAAM,OAAsC;AAAA,IAC1C,cAAc,KAAK;AAAA,IACnB,QAAQ,KAAK,KAAK;AAAA,IAClB,YAAY,KAAK,KAAK,SAAS;AAAA,IAC/B,KAAK,KAAK,KAAK;AAAA,IACf,UAAW,YAAY;AAAA,IACvB;AAAA,IACA,SAAS;AAAA,IACT,YAAY,EAAE,GAAG,YAAY;AAAA,IAC7B,GAAI,gBAAgB,EAAE,cAAc,IAAI,CAAC;AAAA,IACzC,YAAY,KAAK,IAAI,IAAI;AAAA,IACzB,MAAM,KAAK,KAAK;AAAA,IAChB,QAAQ;AAAA,IACR,OAAO;AAAA,EACT;AAEA,MAAI,CAAC,gBAAgB,KAAK,WAAW;AACnC,YAAQ,MAAM,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,EAC/C;AAEA,SAAO,EAAE,MAAM,gBAAgB;AACjC;AAkCO,SAAS,gBACd,MACiB;AACjB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,UAAU,KAAK,WAAW,kBAAkB;AAElD,MAAI,OAAO,KAAK,WAAW,YAAY,KAAK,OAAO,KAAK,EAAE,WAAW,GAAG;AACtE,UAAM,IAAI,MAAM,oEAAoE;AAAA,EACtF;AACA,OAAK,SAAS,cAAc,KAAK,QAAQ,KAAK,IAAI;AAElD,QAAM,eAAe,oBAAoB;AAAA,IACvC,WAAW,KAAK;AAAA,IAChB,QAAS,KAAK,UAAU,CAAC;AAAA,IACzB,aAAa,eAAe,KAAK,UAAU,KAAK,WAAW;AAAA,IAC3D;AAAA,IACA;AAAA,EACF,CAAC;AAED,QAAM,QAAQ,kBAAkB,KAAK,WAAW,MAAM,IAAI,EAAE,IAAI,CAAC,SAA8B;AAC7F,UAAM,YAAYA;AAAA,MAChB,KAAK;AAAA,MACL,KAAK,OAAO,QAAQ,mBAAmB,GAAG;AAAA,MAC1C;AAAA,IACF;AACA,QAAI,CAAC,WAAW;AACd,aAAO;AAAA,QACL,QAAQ,KAAK;AAAA,QACb,YAAY,KAAK,SAAS;AAAA,QAC1B,KAAK,KAAK;AAAA,QACV,MAAM,KAAK;AAAA,QACX;AAAA,QACA,QAAQ;AAAA,QACR,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,UAAM,SAAS,eAAwB;AAAA,MACrC;AAAA,MACA;AAAA,MACA,QAAQ,KAAK;AAAA,MACb;AAAA,IACF,CAAC;AACD,QAAI,OAAO,WAAW,OAAO;AAC3B,aAAO;AAAA,QACL,QAAQ,KAAK;AAAA,QACb,YAAY,KAAK,SAAS;AAAA,QAC1B,KAAK,KAAK;AAAA,QACV,MAAM,KAAK;AAAA,QACX;AAAA,QACA,QAAQ;AAAA,MACV;AAAA,IACF;AAEA,WAAO;AAAA,MACL,QAAQ,KAAK;AAAA,MACb,YAAY,KAAK,SAAS;AAAA,MAC1B,KAAK,KAAK;AAAA,MACV,MAAM,KAAK;AAAA,MACX;AAAA,MACA,QAAQ;AAAA,MACR,QAAQ,OAAO;AAAA,IACjB;AAAA,EACF,CAAC;AAED,QAAM,cAAc,MAAM,OAAO,CAAC,SAAS,KAAK,WAAW,QAAQ,EAAE;AACrE,SAAO;AAAA,IACL;AAAA,IACA,YAAY,MAAM;AAAA,IAClB;AAAA,IACA,YAAY,MAAM,SAAS;AAAA,IAC3B;AAAA,EACF;AACF;AASA,SAAS,iBACP,MACA,MACM;AACN,MAAI,SAAS,SAAS,KAAK,MAAO;AAClC,MAAI,KAAK,aAAa,QAAQ,KAAK,aAAa,OAAW;AAC3D,QAAM,aAAa,KAAK,WAAW,UAAU,KAAK,KAAK,WAAW,WAAW;AAC7E,MAAI,KAAK,YAAY,KAAK,CAAC,WAAY;AACvC,QAAM,MAAM,SAAS,KAAK,MAAM;AAChC,MAAI,SAAS,UAAU;AACrB,UAAM,SAAiC;AAAA,MACrC,cAAc;AAAA,MACd,aAAa;AAAA,MACb,aAAa;AAAA,MACb,iBAAiB;AAAA,MACjB,kBAAkB;AAAA,MAClB,mBAAmB;AAAA,MACnB,cAAc;AAAA,MACd,SAAS;AAAA,MACT,WAAW;AAAA,IACb;AACA,UAAM,IAAI,sBAAsB,gBAAgB,GAAG,IAAI,MAAM;AAAA,EAC/D;AAEA,UAAQ,KAAK,8BAA8B,GAAG,EAAE;AAClD;AAEA,eAAe,aACb,OACA,OACqB;AACrB,SAAO,MAAM,MAAM,KAAK;AAC1B;AAEA,SAAS,wBACP,SACsD;AACtD,SAAO,CAAC,QAAQ,QAAQ;AACtB,UAAM,QAAwC,CAAC;AAC/C,WAAO;AAAA,MACL,KAAK,MAAM,YAAY;AACrB,cAAM,UAAU,KAAK,IAAI;AACzB,cAAM,SAAkC,EAAE,MAAM,QAAQ,SAAS,GAAI,cAAc,CAAC,EAAG;AACvF,cAAM,SAAoB;AAAA,UACxB,IAAI,UAAU;AACZ,mBAAO,aAAa,KAAK,IAAI,IAAI;AACjC,gBAAI,SAAU,QAAO,OAAO,QAAQ,QAAQ;AAC5C,kBAAM,KAAK,MAAM;AAAA,UACnB;AAAA,UACA,aAAa,KAAK,OAAO;AACvB,mBAAO,GAAG,IAAI;AAAA,UAChB;AAAA,QACF;AACA,eAAO;AAAA,MACT;AAAA,MACA,MAAM,QAAQ;AACZ,gBAAQ,MAAMA,MAAK,KAAK,aAAa,GAAG,MAAM,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,MACxF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,YACP,MACA,QAC+B;AAC/B,SAAO;AAAA,IACL,QAAQ,KAAK;AAAA,IACb,YAAY,KAAK,SAAS;AAAA,IAC1B,KAAK,KAAK;AAAA,IACV,UAAU;AAAA,IACV,aAAa,CAAC;AAAA,IACd,SAAS;AAAA,IACT,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,IAClC,YAAY;AAAA,IACZ,MAAM,KAAK;AAAA,IACX,QAAQ;AAAA,IACR,OAAO,YAAY,MAAM;AAAA,EAC3B;AACF;AAEA,SAAS,kBACP,WACA,MACA,MAC+E;AAC/E,QAAM,WAA0F,CAAC;AACjG,MAAI,YAAY;AAChB,aAAW,YAAY,WAAW;AAChC,aAAS,MAAM,GAAG,MAAM,MAAM,OAAO;AACnC,YAAM,SAAS,GAAG,SAAS,EAAE,IAAI,GAAG;AACpC,YAAM,WAAW,OAAO;AACxB,eAAS,KAAK,EAAE,UAAU,KAAK,QAAQ,SAAS,CAAC;AACjD,mBAAa;AAAA,IACf;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,eACP,UACA,UACQ;AACR,QAAM,MAAM,YAAY,UAAU,QAAQ;AAC1C,MAAI,OAAO,QAAQ,YAAY,IAAI,KAAK,EAAE,WAAW,GAAG;AACtD,UAAM,IAAI,MAAM,mEAAmE;AAAA,EACrF;AACA,SAAO;AACT;AAMA,SAAS,eAA0B,MAKV;AACvB,QAAM,MAAM,KAAK,QAAQ,KAAK,KAAK,SAAS;AAC5C,MAAI,QAAQ,OAAW,QAAO,EAAE,QAAQ,QAAQ,QAAQ,UAAU;AAElE,MAAI;AACF,UAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,QAAI,OAAO,WAAW,KAAK,OAAQ,QAAO,EAAE,QAAQ,QAAQ,QAAQ,gBAAgB;AACpF,QAAI,OAAO,iBAAiB,KAAK,cAAc;AAC7C,aAAO,EAAE,QAAQ,QAAQ,QAAQ,oBAAoB;AAAA,IACvD;AACA,WAAO,EAAE,QAAQ,OAAO,MAAM,OAAO;AAAA,EACvC,QAAQ;AACN,WAAO,EAAE,QAAQ,QAAQ,QAAQ,UAAU;AAAA,EAC7C;AACF;AAUA,eAAe,eACb,MACe;AACf,QAAM,KAAK,MAAM,QAAQ;AAAA,IACvB,UAAU,KAAK;AAAA,IACf,UAAU,KAAK,KAAK;AAAA,IACpB,aAAa,KAAK,KAAK;AAAA,IACvB,QAAQ,KAAK,KAAK,iBAAiB;AAAA,IACnC,mBAAmB,KAAK,KAAK,4BAA4B;AAAA,IACzD,YAAY,KAAK,IAAI,EAAE,YAAY;AAAA,IACnC,iBAAiB;AAAA,EACnB,CAAC;AACH;AAIA,SAAS,oBAAoB,OAMlB;AACT,SAAO,YAAY;AAAA,IACjB,WAAW,MAAM;AAAA,IACjB,QAAQ,MAAM,OAAO,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,MAAM,EAAE,WAAW,EAAE;AAAA,IACtE,UAAU,MAAM;AAAA,IAChB,MAAM,MAAM;AAAA,IACZ,MAAM,MAAM;AAAA,EACd,CAAC;AACH;AAEA,SAAS,kBACP,OACA,QACA,MACoB;AACpB,QAAM,UAA0C,CAAC;AACjD,aAAW,SAAS,QAAQ;AAC1B,UAAM,SAAmB,CAAC;AAC1B,eAAW,QAAQ,OAAO;AACxB,YAAM,IAAI,KAAK,YAAY,MAAM,IAAI;AACrC,UAAI,MAAM,OAAW,QAAO,KAAK,EAAE,SAAS;AAAA,IAC9C;AACA,YAAQ,MAAM,IAAI,IAAI,UAAU,QAAQ,IAAI;AAAA,EAC9C;AACA,QAAM,aAAgD,CAAC;AACvD,QAAM,iBAAiB,oBAAI,IAAsB;AACjD,aAAW,QAAQ,OAAO;AACxB,UAAM,aAAa,OAAO,OAAO,KAAK,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AACzE,QAAI,WAAW,WAAW,EAAG;AAC7B,UAAM,OAAO,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;AAChE,UAAM,MAAM,eAAe,IAAI,KAAK,UAAU,KAAK,CAAC;AACpD,QAAI,KAAK,IAAI;AACb,mBAAe,IAAI,KAAK,YAAY,GAAG;AAAA,EACzC;AACA,aAAW,CAAC,YAAY,OAAO,KAAK,gBAAgB;AAClD,UAAM,KAAK,UAAU,SAAS,IAAI;AAClC,eAAW,UAAU,IAAI,EAAE,eAAe,GAAG,MAAM,MAAM,GAAG,MAAM,GAAG,GAAG,EAAE;AAAA,EAC5E;AACA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,SAAS,CAAC;AAAA,IACrD,eAAe,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,KAAK,EAAE;AAAA,IAC7C,cAAc,MAAM,OAAO,CAAC,MAAM,EAAE,OAAO,WAAW,UAAU,CAAC,EAAE;AAAA,IACnE,aAAa,MAAM,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE;AAAA,IAC3C,aAAa,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC,EAAE,MAAM,WAAW,UAAU,CAAC,EAAE;AAAA,EAC/E;AACF;AAKA,SAAS,UAAU,SAAmB,MAA8B;AAClE,QAAM,IAAI,QAAQ;AAClB,MAAI,MAAM,EAAG,QAAO,EAAE,MAAM,GAAG,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC,GAAG,GAAG,EAAE;AAC5D,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAClD,QAAM,WAAW,QAAQ,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AACrF,QAAM,QAAQ,KAAK,KAAK,QAAQ;AAChC,QAAM,KAAK,mBAAmB,SAAS,MAAM,EAAE,MAAM,WAAW,IAAK,CAAC;AACtE,SAAO,EAAE,MAAM,OAAO,MAAM,CAAC,GAAG,OAAO,GAAG,KAAK,GAAG,EAAE;AACtD;","names":["join","existsSync","readFileSync","join"]}

package/dist/{chunk-CMJSTXUR.js → chunk-YFYIOSNV.js}

@@ -1,9 +1,12 @@
+import {
+  recoverTruncatedJson
+} from "./chunk-2B4HPFXN.js";
 import {
   clamp01
-} from "./chunk-IN3SHQML.js";
+} from "./chunk-LMSQ6EFA.js";
 import {
   weightedComposite
-} from "./chunk-HRGUJTER.js";
+} from "./chunk-HYGRFL7C.js";
 import {
   canonicalize
 } from "./chunk-VSMTAMNK.js";
@@ -224,20 +227,29 @@ function parseJudgeResponse(judgeName, resp) {
 }
 
 // src/completion-verifier.ts
+import { randomUUID } from "crypto";
 function completionVerdict(input) {
   if (input.requirements.length === 0) {
     throw new Error(
       `completionVerdict: task '${input.taskId}' has no requirement checks \u2014 nothing to derive a verdict from`
     );
   }
-  const satisfiedCount = input.requirements.filter((r) => r.satisfied).length;
-  const completionRate = satisfiedCount / input.requirements.length;
-  const fullyComplete = satisfiedCount === input.requirements.length;
+  const measurable = input.requirements.filter((r) => !r.unmeasured);
+  const unmeasuredCount = input.requirements.length - measurable.length;
+  if (measurable.length === 0) {
+    throw new Error(
+      `completionVerdict: task '${input.taskId}' has no measurable requirements \u2014 all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? "unknown reason"})`
+    );
+  }
+  const satisfiedCount = measurable.filter((r) => r.satisfied).length;
+  const completionRate = satisfiedCount / measurable.length;
+  const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length;
   return {
     taskId: input.taskId,
     requirements: input.requirements,
     completionRate,
     fullyComplete,
+    unmeasuredCount,
     valid: fullyComplete,
     score: completionRate
   };
@@ -382,12 +394,18 @@ async function verifyCompletion(gold, state, checkCorrectness) {
     const match = assigned.get(i);
     const evidence = [];
     let correct = null;
+    let unmeasuredReason;
     if (match) {
       evidence.push(match.evidence);
       if (match.content !== null) {
-        const r = await checkCorrectness(req, match.content);
-        correct = r.correct;
-        evidence.push(`correctness: ${r.correct ? "pass" : "fail"} \u2014 ${r.reason}`);
+        try {
+          const r = await checkCorrectness(req, match.content);
+          correct = r.correct;
+          evidence.push(`correctness: ${r.correct ? "pass" : "fail"} \u2014 ${r.reason}`);
+        } catch (err) {
+          unmeasuredReason = err instanceof JudgeParseError ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}` : `checker call failed: ${err instanceof Error ? err.message : String(err)}`;
+          evidence.push(`correctness: UNMEASURED \u2014 ${unmeasuredReason}`);
+        }
       } else {
         evidence.push("correctness: not assessed \u2014 matched item carries no content");
       }
@@ -397,34 +415,55 @@ async function verifyCompletion(gold, state, checkCorrectness) {
       evidence.push(`no produced ${kind} matched this requirement`);
     }
     const structurallyPresent = match !== void 0;
-    const satisfied = structurallyPresent && correct !== false;
+    const unmeasured = unmeasuredReason !== void 0;
+    const satisfied = structurallyPresent && !unmeasured && correct !== false;
     requirements.push({
       reqId: req.reqId,
       title: req.title,
       structurallyPresent,
       correct,
       satisfied,
+      ...unmeasured ? { unmeasured: true, unmeasuredReason } : {},
       evidence
     });
   }
   return completionVerdict({ taskId: gold.taskId, requirements });
 }
 function parseCorrectnessResponse(raw) {
+  const readVerdict = (candidate) => {
+    if (candidate === null || typeof candidate !== "object") return null;
+    const { correct, reason } = candidate;
+    if (typeof correct !== "boolean") return null;
+    return { correct, reason: typeof reason === "string" ? reason : "" };
+  };
   const match = raw.match(/\{[\s\S]*\}/);
-  if (!match) {
-    throw new Error(`correctness checker: no JSON object in model response: ${raw.slice(0, 200)}`);
+  if (match) {
+    try {
+      const strict = readVerdict(JSON.parse(match[0]));
+      if (strict) return strict;
+    } catch {
+    }
   }
-  const parsed = JSON.parse(match[0]);
-  if (typeof parsed.correct !== "boolean") {
-    throw new Error(`correctness checker: 'correct' is not a boolean in: ${match[0].slice(0, 200)}`);
+  const start = raw.indexOf("{");
+  if (start !== -1) {
+    const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)));
+    if (recovered) return recovered;
   }
-  return { correct: parsed.correct, reason: typeof parsed.reason === "string" ? parsed.reason : "" };
+  throw new JudgeParseError("correctness-checker", raw);
 }
 function createLlmCorrectnessChecker(tc, opts = {}) {
   const model = opts.model ?? "claude-sonnet-4-6";
   const maxContentChars = opts.maxContentChars ?? 8e3;
+  const maxAttempts = opts.maxAttempts ?? 2;
+  const sink = opts.rawSink;
+  const record = async (event) => {
+    try {
+      await sink?.record(event);
+    } catch {
+    }
+  };
   return async (requirement, content) => {
-    const resp = await tc.chat({
+    const request = {
       model,
       messages: [
         {
@@ -442,9 +481,57 @@ ${content.slice(0, maxContentChars)}`
       ],
       temperature: 0,
       maxTokens: 200
-    });
-    const raw = resp.choices?.[0]?.message?.content ?? "";
-    return parseCorrectnessResponse(raw);
+    };
+    let lastErr;
+    for (let attempt = 0; attempt < maxAttempts; attempt++) {
+      const started = Date.now();
+      await record({
+        eventId: randomUUID(),
+        provider: "correctness-checker",
+        model,
+        endpoint: "/chat",
+        baseUrl: "",
+        attemptIndex: attempt,
+        direction: "request",
+        timestamp: started,
+        requestBody: request,
+        redactedFields: []
+      });
+      try {
+        const resp = await tc.chat(request);
+        const raw = resp.choices?.[0]?.message?.content ?? "";
+        await record({
+          eventId: randomUUID(),
+          provider: "correctness-checker",
+          model,
+          endpoint: "/chat",
+          baseUrl: "",
+          attemptIndex: attempt,
+          direction: "response",
+          timestamp: Date.now(),
+          durationMs: Date.now() - started,
+          responseBody: resp,
+          redactedFields: []
+        });
+        return parseCorrectnessResponse(raw);
+      } catch (err) {
+        lastErr = err;
+        await record({
+          eventId: randomUUID(),
+          provider: "correctness-checker",
+          model,
+          endpoint: "/chat",
+          baseUrl: "",
+          attemptIndex: attempt,
+          direction: "error",
+          timestamp: Date.now(),
+          durationMs: Date.now() - started,
+          errorMessage: err instanceof Error ? err.message : String(err),
+          redactedFields: []
+        });
+      }
+    }
+    throw lastErr instanceof Error ? lastErr : new Error(String(lastErr));
   };
 }
 var TITLE_STOPWORDS = /* @__PURE__ */ new Set([
@@ -771,4 +858,4 @@ export {
   agentProfileModelId,
   agentProfileHash
 };
-//# sourceMappingURL=chunk-CMJSTXUR.js.map
+//# sourceMappingURL=chunk-YFYIOSNV.js.map

package/dist/chunk-YFYIOSNV.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/judges.ts","../src/completion-verifier.ts","../src/llm-judge.ts","../src/produced-state.ts","../src/agent-profile.ts"],"sourcesContent":["import type { TCloud } from '@tangle-network/tcloud'\nimport { JudgeError } from './errors'\nimport type { JudgeFn, JudgeInput, JudgeScore } from './types'\n\n/**\n * A judge's LLM response could not be parsed into scored dimensions.\n * Thrown instead of fabricating a `{ dimension: 'parse_error', score: 0 }`\n * row — a synthetic zero is indistinguishable from a real low score\n * downstream. Carries the raw response for forensics. Callers (executor,\n * ensemble wrappers) catch this per-judge and record a failed judge.\n */\nexport class JudgeParseError extends JudgeError {\n  /** Name of the judge whose response failed to parse. */\n  readonly judgeName: string\n  /** The raw (truncated) model response that failed to parse. */\n  readonly raw: string\n\n  constructor(judgeName: string, raw: string, options?: { cause?: unknown }) {\n    super(`judge '${judgeName}' returned an unparseable response: ${raw.slice(0, 200)}`, options)\n    this.judgeName = judgeName\n    this.raw = raw\n  }\n}\n\n/**\n * Create a domain expert judge with a configurable domain.\n *\n * The judge evaluates professional accuracy and depth.\n *\n * @deprecated Legacy `JudgeFn` factory tied to the fixed gpt-4o prompt shape.\n * Build judges as campaign `JudgeConfig`s (src/campaign/types.ts) — or\n * multi-model panels via `ensembleJudge` (src/judge-panel.ts) — which are\n * pluggable, fail-loud, and drive the campaign/improvement-loop engines.\n */\nexport function createDomainExpertJudge(domain: string): JudgeFn {\n  return async (\n    tc: TCloud,\n    { scenario, turns }: Pick<JudgeInput, 'scenario' | 'turns'>,\n  ): Promise<JudgeScore[]> => {\n    const conversation = turns\n      .map(\n        (t, i) =>\n          `Turn ${i + 1}:\\nUser: ${t.userMessage}\\nAgent: ${t.agentResponse.slice(0, 2000)}`,\n      )\n      .join('\\n\\n---\\n\\n')\n\n    const resp = await tc.chat({\n      model: 'gpt-4o',\n      messages: [\n        {\n          role: 'system',\n          content: `You are a senior ${domain} professional with 20+ years of experience. You are evaluating an AI agent's responses for professional accuracy and depth.\n\nScore STRICTLY. A 5 means \"a junior professional could do this.\" An 8 means \"solid mid-career work.\" A 10 means \"I would hire this agent.\"\n\nEvaluate:\n1. **domain_accuracy** (0-10): Are the technical terms correct? Are the recommendations what you'd actually do? Would this advice cause problems if followed?\n2. **professional_depth** (0-10): Does it go beyond surface-level? Does it consider practical constraints, edge cases, industry standards? Or is it generic textbook advice?\n\nRespond with JSON only: [{\"dimension\":\"domain_accuracy\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"quote from response\"},{\"dimension\":\"professional_depth\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"quote\"}]`,\n        },\n        {\n          role: 'user',\n          content: `Persona: ${scenario.persona} (${scenario.label})\\nScenario: ${scenario.thesis}\\n\\n${conversation}`,\n        },\n      ],\n      temperature: 0.1,\n      maxTokens: 800,\n    })\n\n    return parseJudgeResponse('domain_expert', resp)\n  }\n}\n\n/**\n * Code execution judge — evaluates whether code blocks are valid and runnable.\n *\n * @deprecated Legacy `JudgeFn` factory tied to the fixed gpt-4o prompt shape.\n * Build judges as campaign `JudgeConfig`s (src/campaign/types.ts) — or\n * multi-model panels via `ensembleJudge` (src/judge-panel.ts).\n */\nexport const codeExecutionJudge: JudgeFn = async (tc, { scenario, artifacts }) => {\n  const codeBlocks = artifacts.codeBlocks\n  if (codeBlocks.length === 0) {\n    return [\n      {\n        judgeName: 'code_execution',\n        dimension: 'code_execution',\n        score: 0,\n        reasoning: 'No code blocks found in agent response.',\n      },\n    ]\n  }\n\n  const codeText = codeBlocks\n    .map(\n      (b, i) =>\n        `Block ${i + 1} (${b.language}):\\n\\`\\`\\`${b.language}\\n${b.code.slice(0, 3000)}\\n\\`\\`\\``,\n    )\n    .join('\\n\\n')\n\n  const resp = await tc.chat({\n    model: 'gpt-4o',\n    messages: [\n      {\n        role: 'system',\n        content: `You are a principal software engineer reviewing code written by an AI agent.\n\nScore STRICTLY:\n1. **executability** (0-10): Would this code run without errors? Check: import errors, undefined variables, missing deps, syntax errors. A 5 means \"would run with minor fixes.\" A 10 means \"copy-paste and it works.\"\n2. **completeness** (0-10): Does it handle the FULL task, or just the happy path? A 5 means \"handles the main case.\" A 10 means \"production-ready.\"\n3. **reusability** (0-10): Could this be saved as a tool and reused? A 5 means \"works for this case.\" A 10 means \"general-purpose tool.\"\n\nRespond with JSON only: [{\"dimension\":\"executability\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"specific line/issue\"},{\"dimension\":\"completeness\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"...\"},{\"dimension\":\"reusability\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"...\"}]`,\n      },\n      {\n        role: 'user',\n        content: `Task: ${scenario.thesis}\\n\\n${codeText}`,\n      },\n    ],\n    temperature: 0.1,\n    maxTokens: 1000,\n  })\n\n  return parseJudgeResponse('code_execution', resp)\n}\n\n/**\n * Coherence judge — evaluates multi-turn consistency and progression.\n *\n * @deprecated Legacy `JudgeFn` factory tied to the fixed gpt-4o prompt shape.\n * Build judges as campaign `JudgeConfig`s (src/campaign/types.ts) — or\n * multi-model panels via `ensembleJudge` (src/judge-panel.ts).\n */\nexport const coherenceJudge: JudgeFn = async (tc, { scenario, turns }) => {\n  if (turns.length < 2) {\n    // Single-turn scenarios carry no multi-turn signal. Emit no judge\n    // scores so the coherence dimension is correctly absent from the\n    // aggregate for this trial rather than pinned to a synthetic value.\n    return []\n  }\n\n  const conversation = turns\n    .map(\n      (t, i) =>\n        `Turn ${i + 1}:\\nUser: ${t.userMessage}\\nAgent (${t.agentResponse.length} chars): ${t.agentResponse.slice(0, 1500)}`,\n    )\n    .join('\\n\\n---\\n\\n')\n\n  const resp = await tc.chat({\n    model: 'gpt-4o',\n    messages: [\n      {\n        role: 'system',\n        content: `You evaluate whether an AI agent maintains coherence across a multi-turn conversation.\n\nScore STRICTLY:\n1. **consistency** (0-10): Does the agent contradict itself across turns? Does it remember what it said/built earlier?\n2. **progression** (0-10): Does each turn BUILD on the previous? Or does it start fresh? A 5 means \"vaguely related.\" A 10 means \"each turn clearly advances the work.\"\n3. **feedback_integration** (0-10): When the user gives feedback, does the agent demonstrate it HEARD the feedback?\n\nRespond with JSON only: [{\"dimension\":\"consistency\",\"score\":N,\"reasoning\":\"...\"},{\"dimension\":\"progression\",\"score\":N,\"reasoning\":\"...\"},{\"dimension\":\"feedback_integration\",\"score\":N,\"reasoning\":\"...\"}]`,\n      },\n      {\n        role: 'user',\n        content: `Scenario: ${scenario.thesis}\\n\\n${conversation}`,\n      },\n    ],\n    temperature: 0.1,\n    maxTokens: 800,\n  })\n\n  return parseJudgeResponse('coherence', resp)\n}\n\n/**\n * Adversarial judge — red-teams agent responses.\n *\n * @deprecated Legacy `JudgeFn` factory tied to the fixed gpt-4o prompt shape.\n * Build judges as campaign `JudgeConfig`s (src/campaign/types.ts) — or\n * multi-model panels via `ensembleJudge` (src/judge-panel.ts).\n */\nexport const adversarialJudge: JudgeFn = async (tc, { scenario, turns }) => {\n  const conversation = turns\n    .map(\n      (t, i) => `Turn ${i + 1}:\\nUser: ${t.userMessage}\\nAgent: ${t.agentResponse.slice(0, 1500)}`,\n    )\n    .join('\\n\\n---\\n\\n')\n\n  const resp = await tc.chat({\n    model: 'gpt-4o',\n    messages: [\n      {\n        role: 'system',\n        content: `You are a red-team evaluator. Find the WEAKEST point in the AI agent's responses. Be adversarial.\n\n1. **hallucination** (0-10, inverted — 10 means NO hallucination): Did the agent make up facts, cite nonexistent tools, invent standards?\n2. **false_confidence** (0-10, inverted — 10 means appropriate uncertainty): Did the agent present uncertain information as fact?\n3. **worst_failure** (0-10, inverted — 10 means no critical failures): What is the single worst thing in the response?\n\nBe harsh. If everything is genuinely good, say so — but look hard first.\n\nRespond with JSON only: [{\"dimension\":\"hallucination\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"specific quote\"},{\"dimension\":\"false_confidence\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"...\"},{\"dimension\":\"worst_failure\",\"score\":N,\"reasoning\":\"...\",\"evidence\":\"...\"}]`,\n      },\n      {\n        role: 'user',\n        content: `Persona: ${scenario.persona}\\nScenario: ${scenario.thesis}\\n\\n${conversation}`,\n      },\n    ],\n    temperature: 0.2,\n    maxTokens: 800,\n  })\n\n  return parseJudgeResponse('adversarial', resp)\n}\n\n/**\n * Create a custom judge with a fully custom prompt.\n *\n * @deprecated Legacy `JudgeFn` factory tied to the fixed gpt-4o prompt shape.\n * Build judges as campaign `JudgeConfig`s (src/campaign/types.ts) — or\n * multi-model panels via `ensembleJudge` (src/judge-panel.ts).\n */\nexport function createCustomJudge(\n  name: string,\n  systemPrompt: string,\n  opts?: { model?: string; temperature?: number; maxTokens?: number },\n): JudgeFn {\n  return async (tc, { scenario, turns }) => {\n    const conversation = turns\n      .map(\n        (t, i) =>\n          `Turn ${i + 1}:\\nUser: ${t.userMessage}\\nAgent: ${t.agentResponse.slice(0, 2000)}`,\n      )\n      .join('\\n\\n---\\n\\n')\n\n    const resp = await tc.chat({\n      model: opts?.model ?? 'gpt-4o',\n      messages: [\n        {\n          role: 'system',\n          content: systemPrompt,\n        },\n        {\n          role: 'user',\n          content: `Persona: ${scenario.persona} (${scenario.label})\\nScenario: ${scenario.thesis}\\n\\n${conversation}`,\n        },\n      ],\n      temperature: opts?.temperature ?? 0.1,\n      maxTokens: opts?.maxTokens ?? 1000,\n    })\n\n    return parseJudgeResponse(name, resp)\n  }\n}\n\n/**\n * Default judge set (domain must be provided for domain expert)\n *\n * @deprecated Legacy `JudgeFn` factory tied to the fixed gpt-4o prompt shape.\n * Build judges as campaign `JudgeConfig`s (src/campaign/types.ts) — or\n * multi-model panels via `ensembleJudge` (src/judge-panel.ts).\n */\nexport function defaultJudges(domain: string): JudgeFn[] {\n  return [createDomainExpertJudge(domain), codeExecutionJudge, coherenceJudge, adversarialJudge]\n}\n\n// ── Helpers ──\n\nfunction parseJudgeResponse(judgeName: string, resp: unknown): JudgeScore[] {\n  const content =\n    (resp as { choices?: { message?: { content?: string } }[] }).choices?.[0]?.message?.content ??\n    ''\n  try {\n    let cleaned = content.replace(/```json\\n?|\\n?```/g, '').trim()\n    const arrayMatch = cleaned.match(/\\[[\\s\\S]*\\]/)\n    if (arrayMatch) cleaned = arrayMatch[0]\n    const parsed = JSON.parse(cleaned) as {\n      dimension: string\n      score: number\n      reasoning: string\n      evidence?: string\n    }[]\n    return parsed.map((p) => ({\n      judgeName,\n      dimension: p.dimension,\n      score: Math.max(0, Math.min(10, p.score)),\n      reasoning: p.reasoning ?? '',\n      evidence: p.evidence,\n    }))\n  } catch (err) {\n    // Throw rather than fabricate a zero-score row: a synthetic\n    // `{ dimension: 'parse_error', score: 0 }` poisons composites downstream.\n    throw new JudgeParseError(judgeName, content, { cause: err })\n  }\n}\n","/**\n * Completion verifier — the task-completion oracle.\n *\n * Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n *\n * Per requirement, a two-stage check:\n *   1. Structural — a produced item (vault artifact / approved proposal /\n *      tool call) of the right kind is matched against the requirement and\n *      carries non-empty content. Deterministic; no LLM.\n *   2. Correctness — only if structurally present AND the matched item\n *      carries content, one targeted check decides whether that item\n *      actually fulfils the requirement. A hallucinated artifact fails here;\n *      an absent one already failed stage 1.\n *\n * `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —\n * checker failures — are excluded from the denominator, never scored as\n * zeros). Quality dimensions are meaningless on an incomplete task — callers\n * gate on `fullyComplete` / `completionRate` before scoring quality.\n */\n\nimport { randomUUID } from 'node:crypto'\nimport type { TCloud } from '@tangle-network/tcloud'\nimport type { Artifact } from './artifact-validator'\nimport { recoverTruncatedJson } from './json-recovery'\nimport { JudgeParseError } from './judges'\nimport type { RawProviderEvent, RawProviderSink } from './trace/raw-provider-sink'\nimport type { DefaultVerdict } from './verdict'\n\n/** What kind of produced state can satisfy a requirement structurally. */\nexport type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any'\n\nexport interface CompletionRequirement {\n  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */\n  reqId: string\n  /** Human-readable description of the required deliverable. */\n  title: string\n  /** Optional kind/category hint, matched against a produced item's kind. */\n  category?: string\n  /** What produced state satisfies this requirement. Defaults to 'any'. */\n  satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n  taskId: string\n  requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n  id: string\n  title: string\n  status: 'pending' | 'approved' | 'rejected'\n  /** Optional persisted body — when present, enables a correctness check. */\n  content?: string\n}\n\n/** Everything observable about what a run actually produced. */\nexport interface ProducedState {\n  /** Persisted vault artifacts. Reuses the shared `Artifact` shape. */\n  artifacts: Artifact[]\n  /** Proposals / filings the agent created. */\n  proposals: ProducedProposal[]\n  /** Names of tools the agent invoked. */\n  toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n  reqId: string\n  title: string\n  /** A produced item of the right kind matched the requirement, non-empty. */\n  structurallyPresent: boolean\n  /**\n   * Whether the matched item actually fulfils the requirement. `null` when\n   * not structurally present, when the matched item carries no content\n   * to assess, or when the correctness check itself failed (`unmeasured`).\n   */\n  correct: boolean | null\n  /** structurallyPresent && !unmeasured && correct !== false. */\n  satisfied: boolean\n  /**\n   * Set when the correctness check itself errored (LLM call failure or an\n   * unparseable response after retry). The requirement's fulfilment is\n   * UNKNOWN — `correct` stays null, `satisfied` is false, and\n   * `completionVerdict` excludes the row from `completionRate`'s\n   * denominator. Never folded into a zero: a synthetic zero is\n   * indistinguishable from a real failure (see `JudgeParseError`).\n   */\n  unmeasured?: true\n  /** Why the correctness check could not be measured (present iff `unmeasured`). */\n  unmeasuredReason?: string\n  /** Human-readable evidence for the verdict. */\n  evidence: string[]\n}\n\n/** Extends the substrate verdict spine: `valid` = `fullyComplete` and\n *  `score` = `completionRate` — derived in `completionVerdict()`, the one\n *  place those equalities hold by construction. */\nexport interface CompletionVerdict extends DefaultVerdict {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). */\n  completionRate: number\n  /** Every measurable requirement satisfied (false when anything is unmeasured). */\n  fullyComplete: boolean\n  /** Requirements whose correctness check errored — reported, never scored as zero. */\n  unmeasuredCount: number\n}\n\n/**\n * Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n */\nexport function completionVerdict(input: {\n  taskId: string\n  requirements: RequirementCheck[]\n}): CompletionVerdict {\n  if (input.requirements.length === 0) {\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n    )\n  }\n  const measurable = input.requirements.filter((r) => !r.unmeasured)\n  const unmeasuredCount = input.requirements.length - measurable.length\n  if (measurable.length === 0) {\n    // Every check errored: this is an infrastructure failure, not a scored\n    // run. A 0-rate verdict here would be a fabricated measurement.\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no measurable requirements — all ${input.requirements.length} correctness checks failed (${input.requirements[0]?.unmeasuredReason ?? 'unknown reason'})`,\n    )\n  }\n  const satisfiedCount = measurable.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / measurable.length\n  // A run with unmeasured rows can still report a rate over what WAS\n  // measured, but must not claim full completion.\n  const fullyComplete = unmeasuredCount === 0 && satisfiedCount === measurable.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    unmeasuredCount,\n    valid: fullyComplete,\n    score: completionRate,\n  }\n}\n\n/**\n * Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n */\nexport type CorrectnessChecker = (\n  requirement: CompletionRequirement,\n  content: string,\n) => Promise<{ correct: boolean; reason: string }>\n\nconst STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'of',\n  'for',\n  'and',\n  'or',\n  'to',\n  'in',\n  'on',\n  'with',\n  'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/** artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n  'generated',\n  'generate',\n  'view',\n  'render',\n  'rendered',\n  'persisted',\n  'persist',\n  'artifact',\n  'file',\n  'document',\n  'note',\n  'proposal',\n  'deliverable',\n  'output',\n  'created',\n  'create',\n  'produce',\n  'produced',\n  'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n  return new Set(\n    s\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n  )\n}\n\n/**\n * Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n  if (req.size === 0) return 0\n  const cand = tokens(candidateText)\n  let hit = 0\n  for (const t of req) if (cand.has(t)) hit++\n  return hit / req.size\n}\n\ninterface Candidate {\n  reqIndex: number\n  /** Unique key for a produced item — each item satisfies at most one requirement. */\n  itemKey: string\n  score: number\n  evidence: string\n  /** Content to correctness-check, or null when the matched item has none. */\n  content: string | null\n}\n\nfunction artifactCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  artifacts: Artifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    // Match against the artifact CONTENT too, not just its path + kind — a\n    // generated view / note whose path is generic still satisfies a requirement\n    // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n    // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n    let score = tokenRecall(\n      reqText,\n      `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n    )\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${i}`,\n      score,\n      evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n      content: a.content ?? null,\n    })\n  })\n  return out\n}\n\nfunction proposalCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  proposals: ProducedProposal[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    // A proposal needs an assessable BODY to be a deliverable. A bare title is\n    // not completion: correctness cannot be judged on it, so a title-only match\n    // would auto-pass the oracle (structurallyPresent && correct===null →\n    // satisfied) with no verifiable content. Tool calls are the only\n    // legitimately content-less deliverable (`toolCallCandidates`).\n    const body = (p.content ?? '').trim()\n    if (body.length < MIN_CONTENT_CHARS) continue\n    // Match against the body as well as the (often short) title — a refusal /\n    // flag / analysis proposal whose title is a label still satisfies a\n    // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n    // + the requirement's distinctive tokens keep an off-topic proposal out;\n    // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n    // polarity/fulfilment, so a negation that merely contains the tokens fails.\n    // Structural and correctness must use different evidence or the two-stage\n    // check collapses to one lexical gate.\n    const score = tokenRecall(reqText, `${p.title} ${body}`)\n    if (score < MATCH_THRESHOLD) continue\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${i}`,\n      score,\n      evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n      content: null,\n    })\n  })\n  return out\n}\n\n/**\n * Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n *\n * Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n */\nexport async function verifyCompletion(\n  gold: TaskGold,\n  state: ProducedState,\n  checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n  if (gold.requirements.length === 0) {\n    throw new Error(\n      `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n    )\n  }\n\n  // Collect every above-threshold (requirement, produced-item) candidate, then\n  // assign greedily by descending score: each requirement and each produced\n  // item is used at most once. One deliverable fulfils one requirement.\n  const candidates: Candidate[] = []\n  gold.requirements.forEach((req, i) => {\n    const by = req.satisfiedBy ?? 'any'\n    if (by === 'artifact' || by === 'any') {\n      candidates.push(...artifactCandidates(req, i, state.artifacts))\n    }\n    if (by === 'proposal' || by === 'any') {\n      candidates.push(...proposalCandidates(req, i, state.proposals))\n    }\n    if (by === 'tool-call' || by === 'any') {\n      candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n    }\n  })\n  candidates.sort((a, b) => b.score - a.score)\n\n  const assigned = new Map<number, Candidate>()\n  const itemTaken = new Set<string>()\n  for (const c of candidates) {\n    if (assigned.has(c.reqIndex) || itemTaken.has(c.itemKey)) continue\n    assigned.set(c.reqIndex, c)\n    itemTaken.add(c.itemKey)\n  }\n\n  const requirements: RequirementCheck[] = []\n  for (let i = 0; i < gold.requirements.length; i++) {\n    const req = gold.requirements[i]!\n    const match = assigned.get(i)\n    const evidence: string[] = []\n    let correct: boolean | null = null\n    let unmeasuredReason: string | undefined\n\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        try {\n          const r = await checkCorrectness(req, match.content)\n          correct = r.correct\n          evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n        } catch (err) {\n          // The CHECKER failed, not the requirement. Recording this as a\n          // zero would fabricate a model failure out of an infrastructure\n          // one; the requirement is unmeasured and leaves the denominator.\n          unmeasuredReason =\n            err instanceof JudgeParseError\n              ? `checker response unparseable after retry: ${err.raw.slice(0, 200)}`\n              : `checker call failed: ${err instanceof Error ? err.message : String(err)}`\n          evidence.push(`correctness: UNMEASURED — ${unmeasuredReason}`)\n        }\n      } else {\n        evidence.push('correctness: not assessed — matched item carries no content')\n      }\n    } else {\n      const by = req.satisfiedBy ?? 'any'\n      const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n      evidence.push(`no produced ${kind} matched this requirement`)\n    }\n\n    const structurallyPresent = match !== undefined\n    const unmeasured = unmeasuredReason !== undefined\n    const satisfied = structurallyPresent && !unmeasured && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      ...(unmeasured ? { unmeasured: true as const, unmeasuredReason } : {}),\n      evidence,\n    })\n  }\n\n  return completionVerdict({ taskId: gold.taskId, requirements })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n  /**\n   * Checker LLM calls per requirement before giving up (parse failures and\n   * call errors both consume attempts). The failure then surfaces as an\n   * `unmeasured` requirement, never a zero.\n   */\n  maxAttempts?: number\n  /**\n   * Forensic capture of every checker request/response/error — without it a\n   * checker failure is unauditable (the agent-turn raws never contain the\n   * checker's own calls). Same sink contract as `LlmClient`.\n   */\n  rawSink?: RawProviderSink\n}\n\n/**\n * Parse the correctness checker's model response. Tolerates a response\n * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the\n * verdict boolean usually lands in the first few tokens, so a recovered\n * prefix with a boolean `correct` is a real measurement, not a guess.\n * Fails loud (JudgeParseError) when no boolean verdict is recoverable.\n */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const readVerdict = (candidate: unknown): { correct: boolean; reason: string } | null => {\n    if (candidate === null || typeof candidate !== 'object') return null\n    const { correct, reason } = candidate as { correct?: unknown; reason?: unknown }\n    if (typeof correct !== 'boolean') return null\n    return { correct, reason: typeof reason === 'string' ? reason : '' }\n  }\n\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (match) {\n    try {\n      const strict = readVerdict(JSON.parse(match[0]))\n      if (strict) return strict\n    } catch {\n      // fall through to truncation recovery\n    }\n  }\n  // The strict path needs a closing `}`; a cap-hit response has none. Take\n  // everything from the first `{` and auto-close it.\n  const start = raw.indexOf('{')\n  if (start !== -1) {\n    const recovered = readVerdict(recoverTruncatedJson(raw.slice(start)))\n    if (recovered) return recovered\n  }\n  throw new JudgeParseError('correctness-checker', raw)\n}\n\n/**\n * Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n */\nexport function createLlmCorrectnessChecker(\n  tc: TCloud,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  const maxAttempts = opts.maxAttempts ?? 2\n  const sink = opts.rawSink\n  const record = async (event: RawProviderEvent): Promise<void> => {\n    // Forensic capture is best-effort; the verdict is the system of record.\n    try {\n      await sink?.record(event)\n    } catch {\n      // Intentionally swallowed.\n    }\n  }\n  return async (requirement, content) => {\n    const request = {\n      model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n        },\n        {\n          role: 'user' as const,\n          content: `Requirement: ${requirement.title}\\n${\n            requirement.category ? `Category: ${requirement.category}\\n` : ''\n          }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n        },\n      ],\n      temperature: 0,\n      maxTokens: 200,\n    }\n    let lastErr: unknown\n    for (let attempt = 0; attempt < maxAttempts; attempt++) {\n      const started = Date.now()\n      await record({\n        eventId: randomUUID(),\n        provider: 'correctness-checker',\n        model,\n        endpoint: '/chat',\n        baseUrl: '',\n        attemptIndex: attempt,\n        direction: 'request',\n        timestamp: started,\n        requestBody: request,\n        redactedFields: [],\n      })\n      try {\n        const resp = await tc.chat(request)\n        const raw =\n          (resp as { choices?: { message?: { content?: string } }[] }).choices?.[0]?.message\n            ?.content ?? ''\n        await record({\n          eventId: randomUUID(),\n          provider: 'correctness-checker',\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'response',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          responseBody: resp,\n          redactedFields: [],\n        })\n        return parseCorrectnessResponse(raw)\n      } catch (err) {\n        lastErr = err\n        await record({\n          eventId: randomUUID(),\n          provider: 'correctness-checker',\n          model,\n          endpoint: '/chat',\n          baseUrl: '',\n          attemptIndex: attempt,\n          direction: 'error',\n          timestamp: Date.now(),\n          durationMs: Date.now() - started,\n          errorMessage: err instanceof Error ? err.message : String(err),\n          redactedFields: [],\n        })\n      }\n    }\n    throw lastErr instanceof Error ? lastErr : new Error(String(lastErr))\n  }\n}\n\n/** Stopwords for requirement-title tokenization — drops the imperative verbs\n *  ('review', 'update', …) common to deliverable titles so recall keys on the\n *  substantive nouns, not the boilerplate ask. */\nconst TITLE_STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'and',\n  'or',\n  'for',\n  'to',\n  'of',\n  'in',\n  'on',\n  'with',\n  'review',\n  'update',\n  'new',\n  'proposed',\n])\n\n/**\n * Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network.\n *\n * Polarity-blind: token recall credits a negation that contains the\n * requirement's tokens (\"I will NOT produce the comparison\" recalls every token\n * of \"produce the comparison\"). The structural match stage is ALSO lexical, so\n * pairing the two collapses to a single gameable gate. Use this only as an\n * opt-in structural pre-filter or for tasks whose requirements have no polarity\n * to invert; for produced-state grading the correctness checker MUST be semantic\n * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.\n */\nexport function createTokenRecallChecker(\n  opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n  const minRecall = opts.minRecall ?? 0.5\n  const minLen = opts.minContentLength ?? 120\n  return async (requirement, content) => {\n    const body = content.trim()\n    if (body.length < minLen)\n      return {\n        correct: false,\n        reason: `content too thin (${body.length} chars) to be the deliverable`,\n      }\n    const titleTokens = requirement.title\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n    if (titleTokens.length === 0)\n      return {\n        correct: true,\n        reason: 'requirement title has no significant tokens — structural match accepted',\n      }\n    const lower = body.toLowerCase()\n    const hits = titleTokens.filter((t) => lower.includes(t)).length\n    const recall = hits / titleTokens.length\n    return recall >= minRecall\n      ? {\n          correct: true,\n          reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n        }\n      : {\n          correct: false,\n          reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n        }\n  }\n}\n","/**\n * `llmJudge` — the single-LLM-call bridge that turns a rubric prompt into a\n * canonical campaign `JudgeConfig`.\n *\n * The `JudgeConfig` contract (src/campaign/types.ts) is deliberately a\n * function, not a fixed LLM-prompt shape: real consumers judge with\n * ensembles, deterministic checks, or one LLM call. `ensembleJudge`\n * (src/judge-panel.ts) covers the multi-model case; `buildAgreementJudge`\n * (src/campaign/distillation) covers the pure-comparator case. `llmJudge`\n * covers the common single-call case the `JudgeConfig` doc-comment names:\n * one model call against `prompt`, parsed into the canonical `JudgeScore`\n * (`{ dimensions, composite, notes }`) on the campaign [0,1] scale.\n *\n * Transport is injected as a `ChatClient` (src/analyst/chat-client.ts) — the\n * substrate's transport-agnostic LLM seam — so the judge stays decoupled from\n * router-vs-sandbox-vs-cli-bridge and is unit-testable with the `mock`\n * transport. The composite is computed by `weightedComposite` (the same\n * sum-normalized weighting `ensembleJudge` uses), so a lift is attributable to\n * the dimension scores, not to a bespoke reducer.\n *\n * Fail-loud throughout: an unparseable model response throws `JudgeParseError`;\n * a response missing a declared dimension throws; an out-of-range score throws.\n * A thrown judge is recorded by the campaign engine as a failed cell, never\n * folded into a silent zero.\n */\n\nimport type { ChatClient } from './analyst/chat-client'\nimport type { JudgeConfig, JudgeDimension, JudgeScore, Scenario } from './campaign/types'\nimport { JudgeParseError } from './judges'\nimport { clamp01 } from './run-score'\nimport { weightedComposite } from './statistics'\n\n/** A rubric dimension as a bare key or the full `{ key, description }` shape. A\n *  bare string uses the key as its own description. */\nexport type LlmJudgeDimension = string | JudgeDimension\n\nexport interface LlmJudgeOptions<TArtifact, TScenario extends Scenario = Scenario> {\n  /** The injected LLM transport. One `chat()` call per `score()`. Required —\n   *  there is no default route, so a misconfigured judge fails at construction,\n   *  never silently against the free-tier router. */\n  chat: ChatClient\n  /** Rubric dimensions the model scores. Each becomes a `[0,1]` field of the\n   *  returned `JudgeScore.dimensions`. Defaults to a single `quality` dimension. */\n  dimensions?: LlmJudgeDimension[]\n  /** Model id. Falls back to `chat.defaultModel`; one of the two MUST resolve. */\n  model?: string\n  temperature?: number\n  maxTokens?: number\n  /** Composite weights forwarded to `weightedComposite`: a partial map selects\n   *  AND weights exactly the named dimensions. Omit for a uniform mean. */\n  weights?: Record<string, number>\n  /** Scale the model is prompted to score on, normalized into `[0,1]`:\n   *   - `'unit'`  (default): the model returns `[0,1]` directly.\n   *   - `'ten'`:  the model returns `[0,10]`; divided by 10 here.\n   *  The prompt is annotated with the expected range either way. */\n  scale?: 'unit' | 'ten'\n  /** Run this judge only on matching scenarios (mirrors `JudgeConfig.appliesTo`). */\n  appliesTo?: (scenario: TScenario) => boolean\n  /** Render the artifact + scenario into the user message. Default:\n   *  pretty-printed JSON of `{ scenario, artifact }`. */\n  renderUser?: (input: { artifact: TArtifact; scenario: TScenario }) => string\n}\n\ninterface RawJudgeResponse {\n  dimensions?: Record<string, unknown>\n  scores?: Record<string, unknown>\n  notes?: unknown\n  rationale?: unknown\n}\n\n/**\n * Build a campaign-shaped `JudgeConfig` whose `score()` makes ONE LLM call\n * against `prompt` and reduces the model's per-dimension scores to a canonical\n * `JudgeScore` in `[0,1]`.\n *\n * The model is instructed to return JSON `{ \"dimensions\": { <key>: <number>, … },\n * \"notes\": \"…\" }`; the helper strips fenced JSON, validates every declared\n * dimension is present and in range, normalizes by `scale`, and composites via\n * `weightedComposite`.\n */\nexport function llmJudge<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  name: string,\n  prompt: string,\n  opts: LlmJudgeOptions<TArtifact, TScenario>,\n): JudgeConfig<TArtifact, TScenario> {\n  if (!name.trim()) {\n    throw new Error('llmJudge: name must be non-empty')\n  }\n  if (!prompt.trim()) {\n    throw new Error(`llmJudge '${name}': prompt must be non-empty`)\n  }\n  const model = opts.model ?? opts.chat.defaultModel\n  if (!model) {\n    throw new Error(\n      `llmJudge '${name}': no model on opts and no defaultModel on the ChatClient — ` +\n        'pass opts.model or bind defaultModel at createChatClient().',\n    )\n  }\n\n  const dimensions = normalizeDimensions(opts.dimensions, name)\n  const scale = opts.scale ?? 'unit'\n  const divisor = scale === 'ten' ? 10 : 1\n  const renderUser =\n    opts.renderUser ??\n    ((input: { artifact: TArtifact; scenario: TScenario }) =>\n      JSON.stringify({ scenario: input.scenario, artifact: input.artifact }, null, 2))\n\n  if (opts.weights) {\n    for (const key of Object.keys(opts.weights)) {\n      if (!dimensions.some((d) => d.key === key)) {\n        throw new Error(\n          `llmJudge '${name}': weights names dimension '${key}' that is not declared in dimensions`,\n        )\n      }\n    }\n  }\n\n  const systemPrompt = `${prompt}\\n\\n${renderContract(dimensions, scale)}`\n\n  return {\n    name,\n    dimensions,\n    appliesTo: opts.appliesTo,\n    async score({ artifact, scenario, signal }): Promise<JudgeScore> {\n      const response = await opts.chat.chat(\n        {\n          model,\n          messages: [\n            { role: 'system', content: systemPrompt },\n            { role: 'user', content: renderUser({ artifact, scenario }) },\n          ],\n          jsonMode: true,\n          temperature: opts.temperature ?? 0.1,\n          maxTokens: opts.maxTokens ?? 800,\n        },\n        { signal },\n      )\n\n      const parsed = parseResponse(name, response.content)\n      const rawDims = parsed.dimensions ?? parsed.scores\n      if (!rawDims || typeof rawDims !== 'object') {\n        throw new JudgeParseError(name, response.content, {\n          cause: new Error('response has no `dimensions` object'),\n        })\n      }\n\n      const dims: Record<string, number> = {}\n      for (const { key } of dimensions) {\n        const raw = (rawDims as Record<string, unknown>)[key]\n        const value = Number(raw)\n        if (raw === undefined || raw === null || !Number.isFinite(value)) {\n          throw new JudgeParseError(name, response.content, {\n            cause: new Error(\n              `dimension '${key}' missing or non-numeric (got ${JSON.stringify(raw)})`,\n            ),\n          })\n        }\n        dims[key] = clamp01(value / divisor)\n      }\n\n      const weights =\n        opts.weights ?? Object.fromEntries(dimensions.map((d) => [d.key, 1 / dimensions.length]))\n      const { composite } = weightedComposite({ dims, weights })\n\n      const notes =\n        firstString(parsed.notes) ??\n        firstString(parsed.rationale) ??\n        `${name}: composite ${composite.toFixed(3)} over ${dimensions.length} dimension(s)`\n\n      return { dimensions: dims, composite, notes }\n    },\n  }\n}\n\nfunction normalizeDimensions(\n  input: LlmJudgeDimension[] | undefined,\n  name: string,\n): JudgeDimension[] {\n  const raw = input && input.length > 0 ? input : ['quality']\n  const out: JudgeDimension[] = []\n  const seen = new Set<string>()\n  for (const d of raw) {\n    const dim = typeof d === 'string' ? { key: d, description: d } : d\n    if (!dim.key.trim()) {\n      throw new Error(`llmJudge '${name}': dimension key must be non-empty`)\n    }\n    if (seen.has(dim.key)) {\n      throw new Error(`llmJudge '${name}': duplicate dimension key '${dim.key}'`)\n    }\n    seen.add(dim.key)\n    out.push(dim)\n  }\n  return out\n}\n\nfunction renderContract(dimensions: JudgeDimension[], scale: 'unit' | 'ten'): string {\n  const range = scale === 'ten' ? '0 to 10' : '0.0 to 1.0'\n  const lines = dimensions.map((d) => `  - \"${d.key}\": ${d.description} (score ${range})`)\n  const example = `{\"dimensions\": {${dimensions\n    .map((d) => `\"${d.key}\": <number>`)\n    .join(', ')}}, \"notes\": \"<one-line rationale>\"}`\n  return [\n    'Score the artifact on EACH of these dimensions:',\n    ...lines,\n    '',\n    `Respond with JSON ONLY, no prose. Every dimension is a number in [${range}]:`,\n    example,\n  ].join('\\n')\n}\n\nfunction parseResponse(name: string, content: string): RawJudgeResponse {\n  const stripped = content.replace(/```json\\n?|\\n?```/g, '').trim()\n  const objMatch = stripped.match(/\\{[\\s\\S]*\\}/)\n  const payload = objMatch ? objMatch[0] : stripped\n  try {\n    const parsed = JSON.parse(payload) as RawJudgeResponse\n    if (typeof parsed !== 'object' || parsed === null) {\n      throw new Error('parsed value is not an object')\n    }\n    return parsed\n  } catch (err) {\n    throw new JudgeParseError(name, content, { cause: err })\n  }\n}\n\nfunction firstString(value: unknown): string | undefined {\n  return typeof value === 'string' && value.trim() ? value : undefined\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n  // body of the proposal (e.g. a submit_proposal `description`). When present,\n  // the completion oracle correctness-checks it like artifact content; absent,\n  // the proposal is graded presence-only.\n  content?: string\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize a run's runtime event stream into `ProducedState`.\n *\n * Pure and total — unrecognized event types are skipped. `toolCalls` is\n * deduplicated by name in first-seen order (completion cares about a tool's\n * presence, not its call count). An artifact with neither a name nor a uri\n * still yields an entry keyed by its `artifactId` so it is never silently\n * dropped; an artifact with no `content` yields empty content, which the\n * completion oracle's structural check then rejects on its own.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts: Artifact[] = []\n  const proposals: ProducedProposal[] = []\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      artifacts.push({\n        kind: artifactKind(a.mimeType),\n        path: a.name ?? a.uri ?? a.artifactId,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.push({\n        id: p.proposalId,\n        title: p.title,\n        status: p.status ?? 'pending',\n        ...(p.content !== undefined ? { content: p.content } : {}),\n      })\n    }\n  }\n\n  return { artifacts, proposals, toolCalls }\n}\n","import { createHash } from 'node:crypto'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { type HarnessType, harnessSupportsModel } from '@tangle-network/agent-interface'\nimport { ValidationError } from './errors'\nimport { canonicalize } from './pre-registration'\n\nexport type { AgentProfile, HarnessType } from '@tangle-network/agent-interface'\n\n/**\n * The agentic coding harnesses an eval sweeps by default — the ones we care about\n * ranking. This is the SINGLE source of that list; consumers import it instead of\n * re-declaring their own (a re-declared list is how the fleet drifts). Pass an\n * explicit `harnesses` (e.g. `harnessTypeSchema.options` for literally every known\n * harness) to widen beyond these.\n */\nexport const CODING_HARNESSES: readonly HarnessType[] = [\n  'opencode',\n  'claude-code',\n  'codex',\n  'kimi-code',\n]\n\nexport interface ProfileAxisSpec {\n  /** The domain profile to sweep. Its prompt/tools/skills are held fixed; only the\n   *  harness and model vary. `model.default` is the fallback model. */\n  base: AgentProfile\n  /** Harnesses to cross. Default: {@link CODING_HARNESSES}. */\n  harnesses?: readonly HarnessType[]\n  /** Models to cross. Default: `[base.model.default]` — one model, i.e. today's\n   *  single-model behaviour, so omitting this never changes an existing run. */\n  models?: readonly string[]\n  /** Force every (harness, model) pair verbatim, even ones the harness can't run —\n   *  for deliberately testing failure modes. Default (false): SNAP instead — a\n   *  vendor-locked harness runs only the swept models in its family, or its native\n   *  default when it supports none, so no harness is dropped and none gets a\n   *  guaranteed-failing foreign-model cell. */\n  keepIncompatible?: boolean\n}\n\n/** Model sentinel for a vendor-locked harness that supports none of the swept models:\n *  it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness\n *  resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi\n *  model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship\n *  table that would rot as router catalogs change. */\nexport const HARNESS_NATIVE_MODEL = 'default'\n\n/**\n * Expand a base profile across the harness × model matrix into the `AgentProfile[]`\n * that `runProfileMatrix` / `selfImprove` score — the ONE place \"which harnesses ×\n * which models do we evaluate\" lives, so no product hand-rolls its own harness list\n * or column→profile mapping (the pattern that let those copies drift and silently\n * break the harness pivot).\n *\n * Each cell clones `base`, sets `model.default`, and stamps `metadata.harness` +\n * `metadata.harnessModel` (both hash-bearing, so every cell gets a distinct\n * `agentProfileId` row and results join back by harness/model via {@link harnessAxisOf}\n * with no hand-recomputed key). A vendor-locked harness snaps to its family's swept\n * models — or its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none —\n * so every requested harness runs; `keepIncompatible` forces every pair verbatim.\n *\n * Omit `harnesses`/`models` to sweep the full default set — the \"turn it on for\n * everything we care about\" switch, identical in shape whether one harness or all.\n */\nexport function expandProfileAxes(spec: ProfileAxisSpec): AgentProfile[] {\n  const harnesses = spec.harnesses ?? CODING_HARNESSES\n  if (harnesses.length === 0) throw new ValidationError('expandProfileAxes: no harnesses to sweep')\n  const baseModel = spec.base.model?.default\n  const models = spec.models ?? (baseModel ? [baseModel] : [])\n  if (models.length === 0) {\n    throw new ValidationError(\n      'expandProfileAxes: no models to sweep — base profile has no model.default and none were supplied',\n    )\n  }\n  const out: AgentProfile[] = []\n  const seen = new Set<string>()\n  for (const harness of harnesses) {\n    // A universal (router-backed) harness — opencode/pi/claudish — runs every swept\n    // model. A vendor-locked harness — codex/claude-code/kimi-code — runs only the\n    // swept models in its own family; when it supports NONE of them it snaps to its\n    // native default (the `HARNESS_NATIVE_MODEL` sentinel it resolves at runtime)\n    // rather than being dropped, so every requested harness still appears in the\n    // sweep on a model it can actually run — e.g. sweeping `deepseek/x` puts opencode\n    // on deepseek and kimi-code on its own Kimi model, a real head-to-head.\n    // `keepIncompatible` forces every (harness, model) pair verbatim (failure-mode runs).\n    const supported = spec.keepIncompatible\n      ? models\n      : models.filter((model) => harnessSupportsModel(harness, model))\n    const effective = supported.length > 0 ? supported : [HARNESS_NATIVE_MODEL]\n    for (const model of effective) {\n      const profile: AgentProfile = {\n        ...spec.base,\n        name: `${spec.base.name ?? 'agent'}/${harness}/${model}`,\n        model: { ...spec.base.model, default: model },\n        metadata: { ...(spec.base.metadata ?? {}), harness, harnessModel: model },\n      }\n      const id = agentProfileId(profile)\n      if (seen.has(id)) continue\n      seen.add(id)\n      out.push(profile)\n    }\n  }\n  if (out.length === 0) {\n    // Unreachable in normal use — snapping guarantees ≥1 cell per harness — but keep a\n    // fail-closed guard so a future refactor can't silently produce an empty sweep.\n    throw new ValidationError(\n      `expandProfileAxes: produced no profiles (harnesses=[${harnesses.join(', ')}], models=[${models.join(', ')}]).`,\n    )\n  }\n  return out\n}\n\n/**\n * Read the (harness, model) a matrix cell ran under, off a profile or a result row's\n * profile — the join-back for a `byHarness` pivot. Returns undefined when the profile\n * wasn't produced by {@link expandProfileAxes}. Callers group `result.byProfile` by\n * this instead of recomputing an id (recomputing the wrong key is what broke the pivot\n * in the hand-rolled copies).\n */\nexport function harnessAxisOf(\n  profile: Pick<AgentProfile, 'metadata'>,\n): { harness: HarnessType; model: string } | undefined {\n  const m = profile.metadata as Record<string, unknown> | undefined\n  const harness = m?.harness\n  const model = m?.harnessModel\n  if (typeof harness === 'string' && typeof model === 'string') {\n    return { harness: harness as HarnessType, model }\n  }\n  return undefined\n}\n\n/**\n * Collision-resistant, path-safe, human-readable profile id for eval artifacts.\n * Scorecard joins still use `agentProfileHash`; this id is for run ids, matrix\n * keys, and directory names where two profiles must not collapse onto one row.\n * The suffix is the first 64 bits of the behaviour hash, enough for ordinary\n * eval matrices while keeping filenames readable.\n */\nexport function agentProfileId(profile: AgentProfile): string {\n  const label = pathSafeProfileLabel(agentProfileDisplayLabel(profile)) ?? 'profile'\n  return `${label}-${agentProfileHash(profile).slice(0, 16)}`\n}\n\n/**\n * Model snapshot used for `RunRecord.model`. Eval surfaces require a concrete\n * model id because run records reject bare/missing model aliases.\n */\nexport function agentProfileModelId(profile: AgentProfile): string {\n  const model = profile.model?.default?.trim()\n  if (!model) {\n    const label = agentProfileDisplayLabel(profile) ?? 'unnamed profile'\n    throw new ValidationError(\n      `AgentProfile \"${label}\" has no model.default — cannot record eval run`,\n    )\n  }\n  return model\n}\n\nfunction agentProfileDisplayLabel(profile: AgentProfile): string | undefined {\n  return profile.name?.trim() || profile.version?.trim() || undefined\n}\n\nfunction pathSafeProfileLabel(label: string | undefined): string | undefined {\n  const safe = label\n    ?.trim()\n    .replace(/[^A-Za-z0-9._-]+/g, '-')\n    .replace(/-+/g, '-')\n    .replace(/^-|-$/g, '')\n  return safe || undefined\n}\n\nfunction compact<T extends Record<string, unknown>>(input: T): Partial<T> {\n  const out: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(input)) {\n    if (value !== undefined) out[key] = value\n  }\n  return out as Partial<T>\n}\n\n/**\n * Deterministic behaviour identity for the canonical\n * `@tangle-network/agent-interface` AgentProfile.\n *\n * `name` and `description` are labels and do not affect the hash. Profile\n * `version`, prompt, model hints, tools, resources, hooks, modes, permissions,\n * and extensions do affect the hash. Resource array order is hash-bearing\n * because mount order can change agent behaviour. Undefined fields are treated\n * as absent; explicit `null` fields remain hash-bearing.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  const model = agentProfileModelId(profile)\n  const behaviour = {\n    ...profile,\n    name: undefined,\n    description: undefined,\n    tags: profile.tags ? [...profile.tags].sort() : undefined,\n    model: compact({ ...profile.model, default: model }),\n  }\n  return createHash('sha256')\n    .update(JSON.stringify(canonicalize(behaviour)))\n    .digest('hex')\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAWO,IAAM,kBAAN,cAA8B,WAAW;AAAA;AAAA,EAErC;AAAA;AAAA,EAEA;AAAA,EAET,YAAY,WAAmB,KAAa,SAA+B;AACzE,UAAM,UAAU,SAAS,uCAAuC,IAAI,MAAM,GAAG,GAAG,CAAC,IAAI,OAAO;AAC5F,SAAK,YAAY;AACjB,SAAK,MAAM;AAAA,EACb;AACF;AAYO,SAAS,wBAAwB,QAAyB;AAC/D,SAAO,OACL,IACA,EAAE,UAAU,MAAM,MACQ;AAC1B,UAAM,eAAe,MAClB;AAAA,MACC,CAAC,GAAG,MACF,QAAQ,IAAI,CAAC;AAAA,QAAY,EAAE,WAAW;AAAA,SAAY,EAAE,cAAc,MAAM,GAAG,GAAI,CAAC;AAAA,IACpF,EACC,KAAK,aAAa;AAErB,UAAM,OAAO,MAAM,GAAG,KAAK;AAAA,MACzB,OAAO;AAAA,MACP,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS,oBAAoB,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,QASrC;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,YAAY,SAAS,OAAO,KAAK,SAAS,KAAK;AAAA,YAAgB,SAAS,MAAM;AAAA;AAAA,EAAO,YAAY;AAAA,QAC5G;AAAA,MACF;AAAA,MACA,aAAa;AAAA,MACb,WAAW;AAAA,IACb,CAAC;AAED,WAAO,mBAAmB,iBAAiB,IAAI;AAAA,EACjD;AACF;AASO,IAAM,qBAA8B,OAAO,IAAI,EAAE,UAAU,UAAU,MAAM;AAChF,QAAM,aAAa,UAAU;AAC7B,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO;AAAA,MACL;AAAA,QACE,WAAW;AAAA,QACX,WAAW;AAAA,QACX,OAAO;AAAA,QACP,WAAW;AAAA,MACb;AAAA,IACF;AAAA,EACF;AAEA,QAAM,WAAW,WACd;AAAA,IACC,CAAC,GAAG,MACF,SAAS,IAAI,CAAC,KAAK,EAAE,QAAQ;AAAA,QAAa,EAAE,QAAQ;AAAA,EAAK,EAAE,KAAK,MAAM,GAAG,GAAI,CAAC;AAAA;AAAA,EAClF,EACC,KAAK,MAAM;AAEd,QAAM,OAAO,MAAM,GAAG,KAAK;AAAA,IACzB,OAAO;AAAA,IACP,UAAU;AAAA,MACR;AAAA,QACE,MAAM;AAAA,QACN,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,MAQX;AAAA,MACA;AAAA,QACE,MAAM;AAAA,QACN,SAAS,SAAS,SAAS,MAAM;AAAA;AAAA,EAAO,QAAQ;AAAA,MAClD;AAAA,IACF;AAAA,IACA,aAAa;AAAA,IACb,WAAW;AAAA,EACb,CAAC;AAED,SAAO,mBAAmB,kBAAkB,IAAI;AAClD;AASO,IAAM,iBAA0B,OAAO,IAAI,EAAE,UAAU,MAAM,MAAM;AACxE,MAAI,MAAM,SAAS,GAAG;AAIpB,WAAO,CAAC;AAAA,EACV;AAEA,QAAM,eAAe,MAClB;AAAA,IACC,CAAC,GAAG,MACF,QAAQ,IAAI,CAAC;AAAA,QAAY,EAAE,WAAW;AAAA,SAAY,EAAE,cAAc,MAAM,YAAY,EAAE,cAAc,MAAM,GAAG,IAAI,CAAC;AAAA,EACtH,EACC,KAAK,aAAa;AAErB,QAAM,OAAO,MAAM,GAAG,KAAK;AAAA,IACzB,OAAO;AAAA,IACP,UAAU;AAAA,MACR;AAAA,QACE,MAAM;AAAA,QACN,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,MAQX;AAAA,MACA;AAAA,QACE,MAAM;AAAA,QACN,SAAS,aAAa,SAAS,MAAM;AAAA;AAAA,EAAO,YAAY;AAAA,MAC1D;AAAA,IACF;AAAA,IACA,aAAa;AAAA,IACb,WAAW;AAAA,EACb,CAAC;AAED,SAAO,mBAAmB,aAAa,IAAI;AAC7C;AASO,IAAM,mBAA4B,OAAO,IAAI,EAAE,UAAU,MAAM,MAAM;AAC1E,QAAM,eAAe,MAClB;AAAA,IACC,CAAC,GAAG,MAAM,QAAQ,IAAI,CAAC;AAAA,QAAY,EAAE,WAAW;AAAA,SAAY,EAAE,cAAc,MAAM,GAAG,IAAI,CAAC;AAAA,EAC5F,EACC,KAAK,aAAa;AAErB,QAAM,OAAO,MAAM,GAAG,KAAK;AAAA,IACzB,OAAO;AAAA,IACP,UAAU;AAAA,MACR;AAAA,QACE,MAAM;AAAA,QACN,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,MASX;AAAA,MACA;AAAA,QACE,MAAM;AAAA,QACN,SAAS,YAAY,SAAS,OAAO;AAAA,YAAe,SAAS,MAAM;AAAA;AAAA,EAAO,YAAY;AAAA,MACxF;AAAA,IACF;AAAA,IACA,aAAa;AAAA,IACb,WAAW;AAAA,EACb,CAAC;AAED,SAAO,mBAAmB,eAAe,IAAI;AAC/C;AASO,SAAS,kBACd,MACA,cACA,MACS;AACT,SAAO,OAAO,IAAI,EAAE,UAAU,MAAM,MAAM;AACxC,UAAM,eAAe,MAClB;AAAA,MACC,CAAC,GAAG,MACF,QAAQ,IAAI,CAAC;AAAA,QAAY,EAAE,WAAW;AAAA,SAAY,EAAE,cAAc,MAAM,GAAG,GAAI,CAAC;AAAA,IACpF,EACC,KAAK,aAAa;AAErB,UAAM,OAAO,MAAM,GAAG,KAAK;AAAA,MACzB,OAAO,MAAM,SAAS;AAAA,MACtB,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS;AAAA,QACX;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,YAAY,SAAS,OAAO,KAAK,SAAS,KAAK;AAAA,YAAgB,SAAS,MAAM;AAAA;AAAA,EAAO,YAAY;AAAA,QAC5G;AAAA,MACF;AAAA,MACA,aAAa,MAAM,eAAe;AAAA,MAClC,WAAW,MAAM,aAAa;AAAA,IAChC,CAAC;AAED,WAAO,mBAAmB,MAAM,IAAI;AAAA,EACtC;AACF;AASO,SAAS,cAAc,QAA2B;AACvD,SAAO,CAAC,wBAAwB,MAAM,GAAG,oBAAoB,gBAAgB,gBAAgB;AAC/F;AAIA,SAAS,mBAAmB,WAAmB,MAA6B;AAC1E,QAAM,UACH,KAA4D,UAAU,CAAC,GAAG,SAAS,WACpF;AACF,MAAI;AACF,QAAI,UAAU,QAAQ,QAAQ,sBAAsB,EAAE,EAAE,KAAK;AAC7D,UAAM,aAAa,QAAQ,MAAM,aAAa;AAC9C,QAAI,WAAY,WAAU,WAAW,CAAC;AACtC,UAAM,SAAS,KAAK,MAAM,OAAO;AAMjC,WAAO,OAAO,IAAI,CAAC,OAAO;AAAA,MACxB;AAAA,MACA,WAAW,EAAE;AAAA,MACb,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,EAAE,KAAK,CAAC;AAAA,MACxC,WAAW,EAAE,aAAa;AAAA,MAC1B,UAAU,EAAE;AAAA,IACd,EAAE;AAAA,EACJ,SAAS,KAAK;AAGZ,UAAM,IAAI,gBAAgB,WAAW,SAAS,EAAE,OAAO,IAAI,CAAC;AAAA,EAC9D;AACF;;;AChRA,SAAS,kBAAkB;AA8FpB,SAAS,kBAAkB,OAGZ;AACpB,MAAI,MAAM,aAAa,WAAW,GAAG;AACnC,UAAM,IAAI;AAAA,MACR,4BAA4B,MAAM,MAAM;AAAA,IAC1C;AAAA,EACF;AACA,QAAM,aAAa,MAAM,aAAa,OAAO,CAAC,MAAM,CAAC,EAAE,UAAU;AACjE,QAAM,kBAAkB,MAAM,aAAa,SAAS,WAAW;AAC/D,MAAI,WAAW,WAAW,GAAG;AAG3B,UAAM,IAAI;AAAA,MACR,4BAA4B,MAAM,MAAM,+CAA0C,MAAM,aAAa,MAAM,+BAA+B,MAAM,aAAa,CAAC,GAAG,oBAAoB,gBAAgB;AAAA,IACvM;AAAA,EACF;AACA,QAAM,iBAAiB,WAAW,OAAO,CAAC,MAAM,EAAE,SAAS,EAAE;AAC7D,QAAM,iBAAiB,iBAAiB,WAAW;AAGnD,QAAM,gBAAgB,oBAAoB,KAAK,mBAAmB,WAAW;AAC7E,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd,cAAc,MAAM;AAAA,IACpB;AAAA,IACA;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP,OAAO;AAAA,EACT;AACF;AAYA,IAAM,YAAY,oBAAI,IAAI;AAAA,EACxB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUD,IAAM,6BAA6B,oBAAI,IAAI;AAAA,EACzC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;AAC/D,SAAO,IAAI;AAAA,IACT,EACG,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC;AAAA,EAC1E;AACF;AAQA,SAAS,YAAY,iBAAyB,eAA+B;AAC3E,QAAM,MAAM,OAAO,iBAAiB,0BAA0B;AAC9D,MAAI,IAAI,SAAS,EAAG,QAAO;AAC3B,QAAM,OAAO,OAAO,aAAa;AACjC,MAAI,MAAM;AACV,aAAW,KAAK,IAAK,KAAI,KAAK,IAAI,CAAC,EAAG;AACtC,SAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,GAAG,MAAM;AAC1B,SAAK,EAAE,WAAW,IAAI,KAAK,EAAE,SAAS,kBAAmB;AAKzD,QAAI,QAAQ;AAAA,MACV;AAAA,MACA,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,IAAI,KAAK,EAAE,WAAW,IAAI,MAAM,GAAG,GAAI,CAAC;AAAA,IAC/D;AACA,QAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAAG;AACjF,cAAQ,KAAK,IAAI,OAAO,CAAC;AAAA,IAC3B;AACA,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,CAAC;AAAA,MACtB;AAAA,MACA,UAAU,aAAa,EAAE,QAAQ,EAAE,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,EAAE,WAAW;AAAA,IACxB,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,aAAW,KAAK,WAAW;AAEzB,QAAI,EAAE,WAAW,WAAY;AAM7B,UAAM,QAAQ,EAAE,WAAW,IAAI,KAAK;AACpC,QAAI,KAAK,SAAS,kBAAmB;AASrC,UAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,KAAK,IAAI,IAAI,EAAE;AACvD,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,EAAE,EAAE;AAAA,MACzB;AAAA,MACA,UAAU,sBAAsB,EAAE,KAAK,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,MAAM,MAAM;AAC7B,UAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;AACzC,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,QAAQ,CAAC;AAAA,MAClB;AAAA,MACA,UAAU,cAAc,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MACvE,SAAS;AAAA,IACX,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;AAC5B,MAAI,KAAK,aAAa,WAAW,GAAG;AAClC,UAAM,IAAI;AAAA,MACR,2BAA2B,KAAK,MAAM;AAAA,IACxC;AAAA,EACF;AAKA,QAAM,aAA0B,CAAC;AACjC,OAAK,aAAa,QAAQ,CAAC,KAAK,MAAM;AACpC,UAAM,KAAK,IAAI,eAAe;AAC9B,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,eAAe,OAAO,OAAO;AACtC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AAAA,EACF,CAAC;AACD,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAE3C,QAAM,WAAW,oBAAI,IAAuB;AAC5C,QAAM,YAAY,oBAAI,IAAY;AAClC,aAAW,KAAK,YAAY;AAC1B,QAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,EAAG;AAC1D,aAAS,IAAI,EAAE,UAAU,CAAC;AAC1B,cAAU,IAAI,EAAE,OAAO;AAAA,EACzB;AAEA,QAAM,eAAmC,CAAC;AAC1C,WAAS,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;AACjD,UAAM,MAAM,KAAK,aAAa,CAAC;AAC/B,UAAM,QAAQ,SAAS,IAAI,CAAC;AAC5B,UAAM,WAAqB,CAAC;AAC5B,QAAI,UAA0B;AAC9B,QAAI;AAEJ,QAAI,OAAO;AACT,eAAS,KAAK,MAAM,QAAQ;AAC5B,UAAI,MAAM,YAAY,MAAM;AAC1B,YAAI;AACF,gBAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;AACnD,oBAAU,EAAE;AACZ,mBAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,MAAM,WAAM,EAAE,MAAM,EAAE;AAAA,QAC3E,SAAS,KAAK;AAIZ,6BACE,eAAe,kBACX,6CAA6C,IAAI,IAAI,MAAM,GAAG,GAAG,CAAC,KAClE,wBAAwB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAC9E,mBAAS,KAAK,kCAA6B,gBAAgB,EAAE;AAAA,QAC/D;AAAA,MACF,OAAO;AACL,iBAAS,KAAK,kEAA6D;AAAA,MAC7E;AAAA,IACF,OAAO;AACL,YAAM,KAAK,IAAI,eAAe;AAC9B,YAAM,OAAO,OAAO,QAAQ,gCAAgC;AAC5D,eAAS,KAAK,eAAe,IAAI,2BAA2B;AAAA,IAC9D;AAEA,UAAM,sBAAsB,UAAU;AACtC,UAAM,aAAa,qBAAqB;AACxC,UAAM,YAAY,uBAAuB,CAAC,cAAc,YAAY;AACpE,iBAAa,KAAK;AAAA,MAChB,OAAO,IAAI;AAAA,MACX,OAAO,IAAI;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA,GAAI,aAAa,EAAE,YAAY,MAAe,iBAAiB,IAAI,CAAC;AAAA,MACpE;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO,kBAAkB,EAAE,QAAQ,KAAK,QAAQ,aAAa,CAAC;AAChE;AA2BO,SAAS,yBAAyB,KAAmD;AAC1F,QAAM,cAAc,CAAC,cAAoE;AACvF,QAAI,cAAc,QAAQ,OAAO,cAAc,SAAU,QAAO;AAChE,UAAM,EAAE,SAAS,OAAO,IAAI;AAC5B,QAAI,OAAO,YAAY,UAAW,QAAO;AACzC,WAAO,EAAE,SAAS,QAAQ,OAAO,WAAW,WAAW,SAAS,GAAG;AAAA,EACrE;AAEA,QAAM,QAAQ,IAAI,MAAM,aAAa;AACrC,MAAI,OAAO;AACT,QAAI;AACF,YAAM,SAAS,YAAY,KAAK,MAAM,MAAM,CAAC,CAAC,CAAC;AAC/C,UAAI,OAAQ,QAAO;AAAA,IACrB,QAAQ;AAAA,IAER;AAAA,EACF;AAGA,QAAM,QAAQ,IAAI,QAAQ,GAAG;AAC7B,MAAI,UAAU,IAAI;AAChB,UAAM,YAAY,YAAY,qBAAqB,IAAI,MAAM,KAAK,CAAC,CAAC;AACpE,QAAI,UAAW,QAAO;AAAA,EACxB;AACA,QAAM,IAAI,gBAAgB,uBAAuB,GAAG;AACtD;AAQO,SAAS,4BACd,IACA,OAAkC,CAAC,GACf;AACpB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,QAAM,cAAc,KAAK,eAAe;AACxC,QAAM,OAAO,KAAK;AAClB,QAAM,SAAS,OAAO,UAA2C;AAE/D,QAAI;AACF,YAAM,MAAM,OAAO,KAAK;AAAA,IAC1B,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,UAAU;AAAA,MACd;AAAA,MACA,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SACE;AAAA,QACJ;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,gBAAgB,YAAY,KAAK;AAAA,EACxC,YAAY,WAAW,aAAa,YAAY,QAAQ;AAAA,IAAO,EACjE;AAAA;AAAA,EAAyB,QAAQ,MAAM,GAAG,eAAe,CAAC;AAAA,QAC5D;AAAA,MACF;AAAA,MACA,aAAa;AAAA,MACb,WAAW;AAAA,IACb;AACA,QAAI;AACJ,aAAS,UAAU,GAAG,UAAU,aAAa,WAAW;AACtD,YAAM,UAAU,KAAK,IAAI;AACzB,YAAM,OAAO;AAAA,QACX,SAAS,WAAW;AAAA,QACpB,UAAU;AAAA,QACV;AAAA,QACA,UAAU;AAAA,QACV,SAAS;AAAA,QACT,cAAc;AAAA,QACd,WAAW;AAAA,QACX,WAAW;AAAA,QACX,aAAa;AAAA,QACb,gBAAgB,CAAC;AAAA,MACnB,CAAC;AACD,UAAI;AACF,cAAM,OAAO,MAAM,GAAG,KAAK,OAAO;AAClC,cAAM,MACH,KAA4D,UAAU,CAAC,GAAG,SACvE,WAAW;AACjB,cAAM,OAAO;AAAA,UACX,SAAS,WAAW;AAAA,UACpB,UAAU;AAAA,UACV;AAAA,UACA,UAAU;AAAA,UACV,SAAS;AAAA,UACT,cAAc;AAAA,UACd,WAAW;AAAA,UACX,WAAW,KAAK,IAAI;AAAA,UACpB,YAAY,KAAK,IAAI,IAAI;AAAA,UACzB,cAAc;AAAA,UACd,gBAAgB,CAAC;AAAA,QACnB,CAAC;AACD,eAAO,yBAAyB,GAAG;AAAA,MACrC,SAAS,KAAK;AACZ,kBAAU;AACV,cAAM,OAAO;AAAA,UACX,SAAS,WAAW;AAAA,UACpB,UAAU;AAAA,UACV;AAAA,UACA,UAAU;AAAA,UACV,SAAS;AAAA,UACT,cAAc;AAAA,UACd,WAAW;AAAA,UACX,WAAW,KAAK,IAAI;AAAA,UACpB,YAAY,KAAK,IAAI,IAAI;AAAA,UACzB,cAAc,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,UAC7D,gBAAgB,CAAC;AAAA,QACnB,CAAC;AAAA,MACH;AAAA,IACF;AACA,UAAM,mBAAmB,QAAQ,UAAU,IAAI,MAAM,OAAO,OAAO,CAAC;AAAA,EACtE;AACF;AAKA,IAAM,kBAAkB,oBAAI,IAAI;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAgBM,SAAS,yBACd,OAA0D,CAAC,GACvC;AACpB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,SAAS,KAAK,oBAAoB;AACxC,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,QAAQ,KAAK;AAC1B,QAAI,KAAK,SAAS;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ,qBAAqB,KAAK,MAAM;AAAA,MAC1C;AACF,UAAM,cAAc,YAAY,MAC7B,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,gBAAgB,IAAI,CAAC,CAAC;AACxD,QAAI,YAAY,WAAW;AACzB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ;AAAA,MACV;AACF,UAAM,QAAQ,KAAK,YAAY;AAC/B,UAAM,OAAO,YAAY,OAAO,CAAC,MAAM,MAAM,SAAS,CAAC,CAAC,EAAE;AAC1D,UAAM,SAAS,OAAO,YAAY;AAClC,WAAO,UAAU,YACb;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,mBAAmB,IAAI,IAAI,YAAY,MAAM;AAAA,IACvD,IACA;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,wBAAwB,IAAI,IAAI,YAAY,MAAM;AAAA,IAC5D;AAAA,EACN;AACF;;;ACrjBO,SAAS,SACd,MACA,QACA,MACmC;AACnC,MAAI,CAAC,KAAK,KAAK,GAAG;AAChB,UAAM,IAAI,MAAM,kCAAkC;AAAA,EACpD;AACA,MAAI,CAAC,OAAO,KAAK,GAAG;AAClB,UAAM,IAAI,MAAM,aAAa,IAAI,6BAA6B;AAAA,EAChE;AACA,QAAM,QAAQ,KAAK,SAAS,KAAK,KAAK;AACtC,MAAI,CAAC,OAAO;AACV,UAAM,IAAI;AAAA,MACR,aAAa,IAAI;AAAA,IAEnB;AAAA,EACF;AAEA,QAAM,aAAa,oBAAoB,KAAK,YAAY,IAAI;AAC5D,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,UAAU,UAAU,QAAQ,KAAK;AACvC,QAAM,aACJ,KAAK,eACJ,CAAC,UACA,KAAK,UAAU,EAAE,UAAU,MAAM,UAAU,UAAU,MAAM,SAAS,GAAG,MAAM,CAAC;AAElF,MAAI,KAAK,SAAS;AAChB,eAAW,OAAO,OAAO,KAAK,KAAK,OAAO,GAAG;AAC3C,UAAI,CAAC,WAAW,KAAK,CAAC,MAAM,EAAE,QAAQ,GAAG,GAAG;AAC1C,cAAM,IAAI;AAAA,UACR,aAAa,IAAI,+BAA+B,GAAG;AAAA,QACrD;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,eAAe,GAAG,MAAM;AAAA;AAAA,EAAO,eAAe,YAAY,KAAK,CAAC;AAEtE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,WAAW,KAAK;AAAA,IAChB,MAAM,MAAM,EAAE,UAAU,UAAU,OAAO,GAAwB;AAC/D,YAAM,WAAW,MAAM,KAAK,KAAK;AAAA,QAC/B;AAAA,UACE;AAAA,UACA,UAAU;AAAA,YACR,EAAE,MAAM,UAAU,SAAS,aAAa;AAAA,YACxC,EAAE,MAAM,QAAQ,SAAS,WAAW,EAAE,UAAU,SAAS,CAAC,EAAE;AAAA,UAC9D;AAAA,UACA,UAAU;AAAA,UACV,aAAa,KAAK,eAAe;AAAA,UACjC,WAAW,KAAK,aAAa;AAAA,QAC/B;AAAA,QACA,EAAE,OAAO;AAAA,MACX;AAEA,YAAM,SAAS,cAAc,MAAM,SAAS,OAAO;AACnD,YAAM,UAAU,OAAO,cAAc,OAAO;AAC5C,UAAI,CAAC,WAAW,OAAO,YAAY,UAAU;AAC3C,cAAM,IAAI,gBAAgB,MAAM,SAAS,SAAS;AAAA,UAChD,OAAO,IAAI,MAAM,qCAAqC;AAAA,QACxD,CAAC;AAAA,MACH;AAEA,YAAM,OAA+B,CAAC;AACtC,iBAAW,EAAE,IAAI,KAAK,YAAY;AAChC,cAAM,MAAO,QAAoC,GAAG;AACpD,cAAM,QAAQ,OAAO,GAAG;AACxB,YAAI,QAAQ,UAAa,QAAQ,QAAQ,CAAC,OAAO,SAAS,KAAK,GAAG;AAChE,gBAAM,IAAI,gBAAgB,MAAM,SAAS,SAAS;AAAA,YAChD,OAAO,IAAI;AAAA,cACT,cAAc,GAAG,iCAAiC,KAAK,UAAU,GAAG,CAAC;AAAA,YACvE;AAAA,UACF,CAAC;AAAA,QACH;AACA,aAAK,GAAG,IAAI,QAAQ,QAAQ,OAAO;AAAA,MACrC;AAEA,YAAM,UACJ,KAAK,WAAW,OAAO,YAAY,WAAW,IAAI,CAAC,MAAM,CAAC,EAAE,KAAK,IAAI,WAAW,MAAM,CAAC,CAAC;AAC1F,YAAM,EAAE,UAAU,IAAI,kBAAkB,EAAE,MAAM,QAAQ,CAAC;AAEzD,YAAM,QACJ,YAAY,OAAO,KAAK,KACxB,YAAY,OAAO,SAAS,KAC5B,GAAG,IAAI,eAAe,UAAU,QAAQ,CAAC,CAAC,SAAS,WAAW,MAAM;AAEtE,aAAO,EAAE,YAAY,MAAM,WAAW,MAAM;AAAA,IAC9C;AAAA,EACF;AACF;AAEA,SAAS,oBACP,OACA,MACkB;AAClB,QAAM,MAAM,SAAS,MAAM,SAAS,IAAI,QAAQ,CAAC,SAAS;AAC1D,QAAM,MAAwB,CAAC;AAC/B,QAAM,OAAO,oBAAI,IAAY;AAC7B,aAAW,KAAK,KAAK;AACnB,UAAM,MAAM,OAAO,MAAM,WAAW,EAAE,KAAK,GAAG,aAAa,EAAE,IAAI;AACjE,QAAI,CAAC,IAAI,IAAI,KAAK,GAAG;AACnB,YAAM,IAAI,MAAM,aAAa,IAAI,oCAAoC;AAAA,IACvE;AACA,QAAI,KAAK,IAAI,IAAI,GAAG,GAAG;AACrB,YAAM,IAAI,MAAM,aAAa,IAAI,+BAA+B,IAAI,GAAG,GAAG;AAAA,IAC5E;AACA,SAAK,IAAI,IAAI,GAAG;AAChB,QAAI,KAAK,GAAG;AAAA,EACd;AACA,SAAO;AACT;AAEA,SAAS,eAAe,YAA8B,OAA+B;AACnF,QAAM,QAAQ,UAAU,QAAQ,YAAY;AAC5C,QAAM,QAAQ,WAAW,IAAI,CAAC,MAAM,QAAQ,EAAE,GAAG,MAAM,EAAE,WAAW,WAAW,KAAK,GAAG;AACvF,QAAM,UAAU,mBAAmB,WAChC,IAAI,CAAC,MAAM,IAAI,EAAE,GAAG,aAAa,EACjC,KAAK,IAAI,CAAC;AACb,SAAO;AAAA,IACL;AAAA,IACA,GAAG;AAAA,IACH;AAAA,IACA,qEAAqE,KAAK;AAAA,IAC1E;AAAA,EACF,EAAE,KAAK,IAAI;AACb;AAEA,SAAS,cAAc,MAAc,SAAmC;AACtE,QAAM,WAAW,QAAQ,QAAQ,sBAAsB,EAAE,EAAE,KAAK;AAChE,QAAM,WAAW,SAAS,MAAM,aAAa;AAC7C,QAAM,UAAU,WAAW,SAAS,CAAC,IAAI;AACzC,MAAI;AACF,UAAM,SAAS,KAAK,MAAM,OAAO;AACjC,QAAI,OAAO,WAAW,YAAY,WAAW,MAAM;AACjD,YAAM,IAAI,MAAM,+BAA+B;AAAA,IACjD;AACA,WAAO;AAAA,EACT,SAAS,KAAK;AACZ,UAAM,IAAI,gBAAgB,MAAM,SAAS,EAAE,OAAO,IAAI,CAAC;AAAA,EACzD;AACF;AAEA,SAAS,YAAY,OAAoC;AACvD,SAAO,OAAO,UAAU,YAAY,MAAM,KAAK,IAAI,QAAQ;AAC7D;;;ACpKA,SAAS,aAAa,UAAsC;AAC1D,MAAI,CAAC,SAAU,QAAO;AACtB,MAAI,SAAS,SAAS,MAAM,EAAG,QAAO;AACtC,MAAI,SAAS,WAAW,OAAO,EAAG,QAAO;AACzC,SAAO;AACT;AAYO,SAAS,qBAAqB,QAAoD;AACvF,QAAM,YAAwB,CAAC;AAC/B,QAAM,YAAgC,CAAC;AACvC,QAAM,YAAsB,CAAC;AAC7B,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,MAAM,QAAQ;AACvB,QAAI,GAAG,SAAS,aAAa;AAC3B,YAAM,OAAQ,GAAyB;AACvC,UAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;AAChC,kBAAU,IAAI,IAAI;AAClB,kBAAU,KAAK,IAAI;AAAA,MACrB;AAAA,IACF,WAAW,GAAG,SAAS,YAAY;AACjC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,MAAM,aAAa,EAAE,QAAQ;AAAA,QAC7B,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE;AAAA,QAC3B,SAAS,EAAE,WAAW;AAAA,MACxB,CAAC;AAAA,IACH,WAAW,GAAG,SAAS,oBAAoB;AACzC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,IAAI,EAAE;AAAA,QACN,OAAO,EAAE;AAAA,QACT,QAAQ,EAAE,UAAU;AAAA,QACpB,GAAI,EAAE,YAAY,SAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;AAAA,MAC1D,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO,EAAE,WAAW,WAAW,UAAU;AAC3C;;;AChHA,SAAS,kBAAkB;AAE3B,SAA2B,4BAA4B;AAahD,IAAM,mBAA2C;AAAA,EACtD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAwBO,IAAM,uBAAuB;AAmB7B,SAAS,kBAAkB,MAAuC;AACvE,QAAM,YAAY,KAAK,aAAa;AACpC,MAAI,UAAU,WAAW,EAAG,OAAM,IAAI,gBAAgB,0CAA0C;AAChG,QAAM,YAAY,KAAK,KAAK,OAAO;AACnC,QAAM,SAAS,KAAK,WAAW,YAAY,CAAC,SAAS,IAAI,CAAC;AAC1D,MAAI,OAAO,WAAW,GAAG;AACvB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,MAAsB,CAAC;AAC7B,QAAM,OAAO,oBAAI,IAAY;AAC7B,aAAW,WAAW,WAAW;AAS/B,UAAM,YAAY,KAAK,mBACnB,SACA,OAAO,OAAO,CAAC,UAAU,qBAAqB,SAAS,KAAK,CAAC;AACjE,UAAM,YAAY,UAAU,SAAS,IAAI,YAAY,CAAC,oBAAoB;AAC1E,eAAW,SAAS,WAAW;AAC7B,YAAM,UAAwB;AAAA,QAC5B,GAAG,KAAK;AAAA,QACR,MAAM,GAAG,KAAK,KAAK,QAAQ,OAAO,IAAI,OAAO,IAAI,KAAK;AAAA,QACtD,OAAO,EAAE,GAAG,KAAK,KAAK,OAAO,SAAS,MAAM;AAAA,QAC5C,UAAU,EAAE,GAAI,KAAK,KAAK,YAAY,CAAC,GAAI,SAAS,cAAc,MAAM;AAAA,MAC1E;AACA,YAAM,KAAK,eAAe,OAAO;AACjC,UAAI,KAAK,IAAI,EAAE,EAAG;AAClB,WAAK,IAAI,EAAE;AACX,UAAI,KAAK,OAAO;AAAA,IAClB;AAAA,EACF;AACA,MAAI,IAAI,WAAW,GAAG;AAGpB,UAAM,IAAI;AAAA,MACR,uDAAuD,UAAU,KAAK,IAAI,CAAC,cAAc,OAAO,KAAK,IAAI,CAAC;AAAA,IAC5G;AAAA,EACF;AACA,SAAO;AACT;AASO,SAAS,cACd,SACqD;AACrD,QAAM,IAAI,QAAQ;AAClB,QAAM,UAAU,GAAG;AACnB,QAAM,QAAQ,GAAG;AACjB,MAAI,OAAO,YAAY,YAAY,OAAO,UAAU,UAAU;AAC5D,WAAO,EAAE,SAAiC,MAAM;AAAA,EAClD;AACA,SAAO;AACT;AASO,SAAS,eAAe,SAA+B;AAC5D,QAAM,QAAQ,qBAAqB,yBAAyB,OAAO,CAAC,KAAK;AACzE,SAAO,GAAG,KAAK,IAAI,iBAAiB,OAAO,EAAE,MAAM,GAAG,EAAE,CAAC;AAC3D;AAMO,SAAS,oBAAoB,SAA+B;AACjE,QAAM,QAAQ,QAAQ,OAAO,SAAS,KAAK;AAC3C,MAAI,CAAC,OAAO;AACV,UAAM,QAAQ,yBAAyB,OAAO,KAAK;AACnD,UAAM,IAAI;AAAA,MACR,iBAAiB,KAAK;AAAA,IACxB;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,yBAAyB,SAA2C;AAC3E,SAAO,QAAQ,MAAM,KAAK,KAAK,QAAQ,SAAS,KAAK,KAAK;AAC5D;AAEA,SAAS,qBAAqB,OAA+C;AAC3E,QAAM,OAAO,OACT,KAAK,EACN,QAAQ,qBAAqB,GAAG,EAChC,QAAQ,OAAO,GAAG,EAClB,QAAQ,UAAU,EAAE;AACvB,SAAO,QAAQ;AACjB;AAEA,SAAS,QAA2C,OAAsB;AACxE,QAAM,MAA+B,CAAC;AACtC,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,KAAK,GAAG;AAChD,QAAI,UAAU,OAAW,KAAI,GAAG,IAAI;AAAA,EACtC;AACA,SAAO;AACT;AAYO,SAAS,iBAAiB,SAA+B;AAC9D,QAAM,QAAQ,oBAAoB,OAAO;AACzC,QAAM,YAAY;AAAA,IAChB,GAAG;AAAA,IACH,MAAM;AAAA,IACN,aAAa;AAAA,IACb,MAAM,QAAQ,OAAO,CAAC,GAAG,QAAQ,IAAI,EAAE,KAAK,IAAI;AAAA,IAChD,OAAO,QAAQ,EAAE,GAAG,QAAQ,OAAO,SAAS,MAAM,CAAC;AAAA,EACrD;AACA,SAAO,WAAW,QAAQ,EACvB,OAAO,KAAK,UAAU,aAAa,SAAS,CAAC,CAAC,EAC9C,OAAO,KAAK;AACjB;","names":[]}

package/dist/{code-agent-session-Ce-9u7YM.d.ts → code-agent-session-Cen-qD2y.d.ts}

@@ -1,4 +1,4 @@
-import { b as RunSplitTag, R as RunRecord } from './run-record-MRdJ-Kq2.js';
+import { b as RunSplitTag, R as RunRecord } from './run-record-CPfd1ARZ.js';
 
 type CodeAgentSessionSource = 'codex' | 'claude-code' | 'opencode' | 'kimi-code' | 'pi';
 interface ParsedCodeAgentJsonl {