@tangle-network/agent-eval 0.101.1 → 0.102.0

package/dist/chunk-LSCBODPQ.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\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}/${path}`] = fullPath\n      return fullPath\n    },\n    async writeJson(path, value) {\n      return artifacts.write(path, JSON.stringify(value, null, 2))\n    },\n  }\n  let costSoFar = 0\n  const tokensSoFar: CampaignTokenUsage = { input: 0, output: 0 }\n  const cost: CampaignCostMeter = {\n    observe(amount, source) {\n      costSoFar += amount\n      trace.span(`cost.${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    current() {\n      return costSoFar\n    },\n    tokens() {\n      return { ...tokensSoFar }\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    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,QAAQ,MAAM,YAAY;AACvC,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;;;AJ0CA,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,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,UAAU;AACR,aAAO;AAAA,IACT;AAAA,IACA,SAAS;AACP,aAAO,EAAE,GAAG,YAAY;AAAA,IAC1B;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,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-CHIFZIQD.js → chunk-QUCGGMYM.js}

@@ -1,10 +1,10 @@
 import {
   runCanaries,
   scoreRedTeamOutput
-} from "./chunk-NK77GPUH.js";
+} from "./chunk-52CCCXU3.js";
 import {
   runCampaign
-} from "./chunk-XIOQHCHU.js";
+} from "./chunk-LSCBODPQ.js";
 import {
   detectRewardHacking
 } from "./chunk-AIGWQEME.js";
@@ -432,4 +432,4 @@ export {
   runEval,
   evolutionaryProposer
 };
-//# sourceMappingURL=chunk-CHIFZIQD.js.map
+//# sourceMappingURL=chunk-QUCGGMYM.js.map

package/dist/contract/index.d.ts

@@ -1,9 +1,9 @@
 import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, f as SurfaceProposer, g as Gate, L as LabeledScenarioStore, C as CampaignResult, j as GateDecision } from '../types-fWqEJm7h.js';
 export { k as CampaignAggregates, l as CampaignArtifactWriter, m as CampaignCellResult, n as CampaignCostMeter, d as CampaignTraceWriter, o as CodeSurface, D as Dispatch, h as GateContext, G as GateResult, p as GenerationCandidate, e as GenerationRecord, c as JudgeDimension, J as JudgeScore, i as Mutator, O as OptimizationProposer, q as OptimizerConfig, r as SessionScript } from '../types-fWqEJm7h.js';
-import { L as LoopProvenanceRecord, R as RunEvalOptions } from '../provenance-DdfmVfqR.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, e as buildEvidenceVector, f as composeGate, g as defaultProductionGate, h as evolutionaryProposer, i as heldOutGate, p as paretoPolicy, j as paretoSignificanceGate, r as runEval } from '../provenance-DdfmVfqR.js';
-import { C as CampaignStorage, a as RunOptimizationOptions, b as RunImprovementLoopResult } from '../gepa-CEy1AIWp.js';
-export { G as GepaProposerOptions, R as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../gepa-CEy1AIWp.js';
+import { L as LoopProvenanceRecord, R as RunEvalOptions } from '../provenance-BEITkFII.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, e as buildEvidenceVector, f as composeGate, g as defaultProductionGate, h as evolutionaryProposer, i as heldOutGate, p as paretoPolicy, j as paretoSignificanceGate, r as runEval } from '../provenance-BEITkFII.js';
+import { C as CampaignStorage, a as RunOptimizationOptions, b as RunImprovementLoopResult } from '../gepa-bxuDoaO9.js';
+export { G as GepaProposerOptions, R as RunCampaignOptions, c as RunImprovementLoopOptions, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, r as runCampaign, d as runImprovementLoop } from '../gepa-bxuDoaO9.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
 import { HostedTenant, EvalRunCellScore, EvalRunGenerationSnapshot, EvalRunEvent, TraceSpanEvent } from '../hosted/index.js';
 import { R as RunRecord, b as RunSplitTag } from '../run-record-DEwidcqn.js';

package/dist/contract/index.js

@@ -18,22 +18,22 @@ import {
   paretoPolicy,
   paretoSignificanceGate,
   runEval
-} from "../chunk-CHIFZIQD.js";
+} from "../chunk-QUCGGMYM.js";
 import {
   analyzeRuns
-} from "../chunk-X5OUZB4T.js";
+} from "../chunk-JCUREYF5.js";
 import {
   emitLoopProvenance,
   gepaProposer,
   heldOutGate,
   runImprovementLoop,
   surfaceContentHash
-} from "../chunk-NK77GPUH.js";
+} from "../chunk-52CCCXU3.js";
 import {
   fsCampaignStorage,
   inMemoryCampaignStorage,
   runCampaign
-} from "../chunk-XIOQHCHU.js";
+} from "../chunk-LSCBODPQ.js";
 import "../chunk-VI2UW6B6.js";
 import {
   buildDefaultAnalystRegistry

package/dist/{gepa-CEy1AIWp.d.ts → gepa-bxuDoaO9.d.ts}

@@ -134,8 +134,13 @@ interface RunCampaignOptions<TScenario extends Scenario, TArtifact> {
      * every loop/CI job above it) forever. `undefined`/`0` = unbounded (legacy).
      */
     dispatchTimeoutMs?: number;
-    /** Required: where artifacts + traces land. */
+    /** Required: where artifacts + traces land. A bare name (not an absolute path)
+     *  resolves to the shared `~/.tangle/traces/<repo>/runs/<name>` root so run
+     *  bundles never pollute a repo working tree. Pass an absolute path to override. */
     runDir: string;
+    /** Subject repo for the shared run-dir root (defaults to the CWD basename).
+     *  Only consulted when `runDir` is a bare name. */
+    repo?: string;
     /** Tracing posture. Default is the substrate's `FileSystemTraceStore` rooted
      *  at `<runDir>/traces/`. `'off'` disables capture entirely — substrate
      *  refuses this when the caller wires `autoOnPromote !== 'none'`. */
@@ -204,6 +209,8 @@ interface PlanCampaignRunOptions<TScenario extends Scenario, TArtifact> {
     reps?: number;
     resumable?: boolean;
     runDir: string;
+    /** Subject repo for the shared run-dir root (see RunCampaignOptions.repo). */
+    repo?: string;
     storage?: CampaignStorage;
 }
 declare function planCampaignRun<TScenario extends Scenario, TArtifact>(opts: PlanCampaignRunOptions<TScenario, TArtifact>): CampaignRunPlan;

package/dist/index.d.ts

@@ -75,7 +75,7 @@ export { C as CallbackResearcher, d as CallbackResearcherOptions, e as CampaignF
 export { G as GainDistributionBin, a as GainDistributionFigureSpec, b as GainDistributionOptions, m as GateDecision, n as GateEvidence, H as HeldOutGate, o as HeldOutGateConfig, q as HeldOutGateRejectionCode, P as ParetoFigureSpec, c as ParetoPoint, R as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as ResearchReport, e as ResearchReportCandidate, f as ResearchReportDecision, g as ResearchReportMethodology, h as ResearchReportOptions, i as ResearchReportRecommendation, S as SummaryTable, j as SummaryTableOptions, k as SummaryTableRow, l as gainHistogram, p as paretoChart, r as researchReport, s as summaryTable } from './summary-report-C4uzRWh8.js';
 export { L as LockedJsonlAppender } from './testing-C21CHsq2.js';
 export { I as InterimReleaseConfidence, a as InterimReleaseConfidenceInput, P as PairedEvalueOptions, b as PairedEvalueSequence, c as PairedEvalueStep, S as SequentialDecision, e as evaluateInterimReleaseConfidence, p as pairedEvalueSequence } from './sequential-5iSVfzl2.js';
-import { j as GepaProposerConstraints, b as RunImprovementLoopResult } from './gepa-CEy1AIWp.js';
+import { j as GepaProposerConstraints, b as RunImprovementLoopResult } from './gepa-bxuDoaO9.js';
 export { IntegrityResult, IntegrityViolation, JourneySpec, PerfBaseline, PerfGateResult, PerfRegression, PerfScenario, PerfStat, ScenarioAxes, assertRecordIntegrity, checkRecordIntegrity, expandMatrix, gatePerf, scenarioKey, summarizeRecords } from './perf/index.js';
 export { AgentProfileRuntimeReceipt, ProductBenchmarkArm, ProductBenchmarkArtifactPaths, ProductBenchmarkBudgets, ProductBenchmarkManifest, ProductBenchmarkProfileRef, ProductBenchmarkRecord, ProductBenchmarkRepoRef, ProductBenchmarkRunInput, ProductBenchmarkScenario, ProductBenchmarkSplit, ProductBenchmarkSubstrateVersions, ProductBenchmarkValidationReport, RuntimeResolution, findProductBenchmarkArtifacts, productBenchmarkIntegrityFailures, productBenchmarkSplits, readProductBenchmarkManifest, readProductBenchmarkRecords, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun } from './product-benchmark/index.js';
 import '@ax-llm/ax';
@@ -1987,11 +1987,19 @@ interface ProfileAxisSpec {
     /** Models to cross. Default: `[base.model.default]` — one model, i.e. today's
      *  single-model behaviour, so omitting this never changes an existing run. */
     models?: readonly string[];
-    /** Keep (harness, model) pairs the harness can't run instead of dropping them.
-     *  Default: drop (via `harnessSupportsModel`), so a vendor-locked harness paired
-     *  with a foreign model doesn't become a guaranteed-failing cell. */
+    /** Force every (harness, model) pair verbatim, even ones the harness can't run —
+     *  for deliberately testing failure modes. Default (false): SNAP instead — a
+     *  vendor-locked harness runs only the swept models in its family, or its native
+     *  default when it supports none, so no harness is dropped and none gets a
+     *  guaranteed-failing foreign-model cell. */
     keepIncompatible?: boolean;
 }
+/** Model sentinel for a vendor-locked harness that supports none of the swept models:
+ *  it carries no provider prefix, so `harnessSupportsModel` accepts it and the harness
+ *  resolves it to its own native default model at runtime (e.g. kimi-code → its Kimi
+ *  model). Lets `expandProfileAxes` snap-instead-of-drop without a per-harness flagship
+ *  table that would rot as router catalogs change. */
+declare const HARNESS_NATIVE_MODEL = "default";
 /**
  * Expand a base profile across the harness × model matrix into the `AgentProfile[]`
  * that `runProfileMatrix` / `selfImprove` score — the ONE place "which harnesses ×
@@ -2002,7 +2010,9 @@ interface ProfileAxisSpec {
  * Each cell clones `base`, sets `model.default`, and stamps `metadata.harness` +
  * `metadata.harnessModel` (both hash-bearing, so every cell gets a distinct
  * `agentProfileId` row and results join back by harness/model via {@link harnessAxisOf}
- * with no hand-recomputed key). Incompatible pairs are dropped unless `keepIncompatible`.
+ * with no hand-recomputed key). A vendor-locked harness snaps to its family's swept
+ * models — or its native default ({@link HARNESS_NATIVE_MODEL}) when it supports none —
+ * so every requested harness runs; `keepIncompatible` forces every pair verbatim.
  *
  * Omit `harnesses`/`models` to sweep the full default set — the "turn it on for
  * everything we care about" switch, identical in shape whether one harness or all.
@@ -6069,4 +6079,4 @@ type CachedJudge<TArtifact, TScenario extends Scenario$1 = Scenario$1> = JudgeCo
  */
 declare function cachedJudge<TArtifact, TScenario extends Scenario$1 = Scenario$1>(judge: JudgeConfig<TArtifact, TScenario>, store: VerdictCacheStore, options: CachedJudgeOptions): CachedJudge<TArtifact, TScenario>;
 
-export { ATTESTATION_ALGORITHM, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, type AgreementResult, type AlignmentOp, AnalyzeTracesInput, AnalyzeTracesOptions, AnalyzeTracesResult, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type AssertCrossFamilyOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, type BackendDescriptor, BaselineReport, BehaviorAssertion, BehavioralMetrics, BenchmarkReport, BenchmarkRunner, BenchmarkRunnerConfig, type BisectOptions, type BisectResult, type BisectStep, type BlendWeights, BudgetBreachError, BudgetGuard, BudgetLedgerEntry, BudgetSpec, type BuildAgreementJudgeOptions, CODING_HARNESSES, type CachedJudge, type CachedJudgeOptions, CallExpectation, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateComparison, type CandidateScenario, type CausalAttributionReport, type CellVerdict, ChannelRollup, ChatRequest, CheckResult, CollectedArtifacts, type CommandRunner, type CompareLabels, CompletionCriterion, ConfigError, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractCheckResult, type ContractJudgeOptions, type ContractMetric, type ContractReport, type ContractRule, type ContractRuleKind, type ContractSpan, type ContractVerdict, type ContractViolation, ConvergenceTracker, CorrectnessChecker, type CostEntry, CostLedger, type CostReport, type CostSummary, CostTracker, CreateChatClientOpts, type CreateDefaultReviewerOptions, type CreateExperimentInput, type CreateSandboxPoolOpts, CrossFamilyError, type CrossTraceDiff, type CrossTraceDiffOptions, DEFAULT_AGENT_SLOS, DEFAULT_FINDERS, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, DatasetScenario, type DecideNextUserTurnOpts, DefaultVerdict, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DescriptionLengthCandidate, type DescriptionLengthConfig, type DescriptionLengthDecision, type DescriptionLengthEvidence, DescriptionLengthGate, type DescriptionLengthRejectionCode, type DetectorEvent, type DetectorSeverity, type DetectorSignal, type DiffScorecardOptions, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, DriverResult, DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type ErrorCountPattern, type ErrorStreakOptions, type EvalToolDef, EvalTraceStore, type EvolutionRound, type ExecutorConfig, type Expectation, type Experiment, type ExperimentProvenance, type ExperimentRep, type ExperimentStats, type ExperimentStore, ExperimentTracker, type ExperimentTrackerOptions, type ExperimentVerdict, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type FactorContribution, type FactorialCell, FailureClass, FeedbackLabel, FeedbackTrajectory, FeedbackTrajectoryStore, type FieldAgreementSpec, type FieldDestination, type FileChange, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GhCliClientOptions, type GoldScenario, type GoldSplit, type GoldenSeverity, type GoldenSpec, HarnessConfig, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, HoldoutAuditor, type HttpGithubClientOptions, INTENT_MATCH_JUDGE_VERSION, type ImageData, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, JudgeError, type JudgeFamily, type JudgeFleetOptions, JudgeFn, JudgeParseError, type JudgeReplayResult, type JudgeRetryOutcome, type JudgeRetryPolicy, JudgeRunner, type JudgeScoreInput, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, Layer, LayerResult, type LeaderboardOptions, type LeaderboardRow, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmClientOptions, LlmSpan, MODEL_PRICING, type MakeEvalToolsConfig, type MatchResult, type MatcherResult, type MergeOptions, MetricsCollector, type ModelCostRollup, type ModelPreflight, type ModelSeats, ModelsUnreachableError, type MuffledFinder, type MuffledFinding, type MultiToolchainLayerConfig, type Mutator, Mutex, type NoLeakOptions, type NoProgressOptions, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, OtelExportConfig, OtelExporter, type OtelPipelineHandle, type OtelPipelineOptions, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, type ParseStudentLabel, type PartitionHeldOutOptions, PersonaConfig, type Playbook, type PlaybookEntry, type PoolSlot, type PrReviewAuditCase, type PrReviewBenchmarkSummary, type PrReviewComment, type PrReviewMatchedFinding, type PrReviewOutcome, type PrReviewReferenceFinding, type PrReviewScore, type PrReviewScoreWeights, type PrReviewSeverity, type PrReviewSource, type PreflightModelsOptions, type PreflightOutcome, ProductClient, ProductClientConfig, type ProfileAxisSpec, type PromptHandle, PromptRegistry, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ProvenanceReader, type RecordRunsOptions, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, ReleaseConfidenceScorecard, ReleaseConfidenceThresholds, type RenderStudentPrompt, type RepeatedActionOptions, type RepoRef, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, type RoutedField, Run, type RunCommandInput, type RunCommandResult, type RunDistillationOptions, type RunDistillationResult, RunFilter, RunRecord, type RunRecordBackend, type RunRecordFilter, RunScore, RunScoreWeights, RunSplitTag, SandboxDriver, SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type ScanOptions, Scenario, type ScenarioCost, ScenarioFile, ScenarioRegistry, ScenarioResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type ScoredTarget, type SeatName, type SeatPresetName, SeatUnsetError, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SerializedRegex, Severity, type SingleBackendDivergence, SingleBackendError, type SingleBackendField, type SingleBackendReport, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, type SpanPredicate, type SplitGoldOptions, SteeringBundle, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type StepAttribution, type StreamingDetector, type SynthesisReason, type SynthesisTarget, TestResult, type TextMatcher, type ThresholdContract, TokenCounter, type TokenSpec, type ToolMatcher, ToolSpan, TraceAnalystSpan, type TraceContract, TraceContractBuilder, TraceEmitter, TraceStore, type TracedAnalystOptions, type TracedJudgeOptions, Trajectory, TrajectoryStep, type TreatmentClass, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, TurnMetrics, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, type UiFinding, type UiFindingScreenshot, type UiFindingSeverity, type UiLens, type VerdictCacheStats, type VerdictCacheStore, VerifyContext, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkerDriverContext, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, adversarialJudge, agentProfileHash, agentProfileId, agentProfileModelId, agentVisibleFields, aggregateJudgeVerdicts, aggregatePrReviewScore, analyzeAntiSlop, appendScorecard, assertCrossFamily, assertModelsServed, assertNoHiddenLeak, assertSingleBackend, assignHeldOutTag, attachCostToReport, attest, bisect, blendHeldout, buildAgreementJudge, buildDriverSystemPrompt, buildReflectionPrompt, buildReviewerPrompt, buildWorkerDriverSystemPrompt, cachedJudge, canaryLeakView, canonicalJson, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, classifyTreatment, codeExecutionJudge, coherenceJudge, collectionPreserved, commentsForSource, commitBisect, compareReferenceReplay, compilerJudge, computeExperimentStats, contentHash, contractJudge, costReport, createAntiSlopJudge, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createSandboxPool, crossTraceDiff, dataDescriptionBits, decideNextUserTurn, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultBlendWeights, defaultJudges, defaultParseStudentLabel, defaultReferenceReplayMatcher, defaultRenderStudentPrompt, deployGateLayer, diffScorecard, discoverPersonas, distillPlaybook, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateContract, evaluateOracles, evaluateTraceContract, executeScenario, expandProfileAxes, expectAgent, exportRewardModel, extractAssetUrls, extractErrorCount, fieldAgreement, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, gateTreatmentApplied, gateTreatmentFromMetrics, gateTreatmentFromSpans, gateTreatmentFromToolSpans, ghCliClient, gitProvenanceReader, precision as goldenPrecision, gradeOnHidden, harnessAxisOf, hashContent, hashToUnit, hiddenGrade, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryRunRecordBackend, inMemoryVerdictCache, isHiddenDestination, isModelPriced, isOtelConfigured, jsonShape, jsonlReferenceReplayStore, jsonlRunRecordBackend, judgeFamily, keyPreserved, leaderboard, linterJudge, loadGoldScenarios, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, matchGoldens, matchSpan, mergeLayerResults, modelDescriptionBits, multiToolchainLayer, noProgressDetector, notBlocked, observeAll, paraphraseRobustness, paraphraseRobustnessScenarios, parseGoldJsonl, parseReflectionResponse, partitionHeldOut, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, preflightModels, printDriverSummary, index as profile, promptBisect, proposeSynthesisTargets, recordRuns, recordRunsToScorecard, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, repeatedActionDetector, replayScorerOverCorpus, replayTraceThroughJudge, resolveModelPricing, resolveSeat, routeFields, rowCount, rowWhere, runAssertions, runBehavioralCanaries, runCanaries, runDistillation, runE2EWorkflow, runExpectations, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runReferenceReplay, runScore, runSelfPlay, scanForMuffledGates, scoreContinuity, scorePrReviewComments, scorePrReviewSource, scoreReferenceReplay, seatPresets, securityJudge, sentenceReorderMutator, splitGold, statusAdvanced, summarizePrReviewBenchmark, testJudge, textInSnapshot, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, traceContract, traceJudge, traceJudgeEnsemble, tracedAnalyzeTraces, typoMutator, urlContains, verifyAttestation, visualDiff, viteDeployRunner, weightedRecall, whitespaceCollapseMutator, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner };
+export { ATTESTATION_ALGORITHM, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, type AgreementResult, type AlignmentOp, AnalyzeTracesInput, AnalyzeTracesOptions, AnalyzeTracesResult, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type AssertCrossFamilyOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, type BackendDescriptor, BaselineReport, BehaviorAssertion, BehavioralMetrics, BenchmarkReport, BenchmarkRunner, BenchmarkRunnerConfig, type BisectOptions, type BisectResult, type BisectStep, type BlendWeights, BudgetBreachError, BudgetGuard, BudgetLedgerEntry, BudgetSpec, type BuildAgreementJudgeOptions, CODING_HARNESSES, type CachedJudge, type CachedJudgeOptions, CallExpectation, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateComparison, type CandidateScenario, type CausalAttributionReport, type CellVerdict, ChannelRollup, ChatRequest, CheckResult, CollectedArtifacts, type CommandRunner, type CompareLabels, CompletionCriterion, ConfigError, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractCheckResult, type ContractJudgeOptions, type ContractMetric, type ContractReport, type ContractRule, type ContractRuleKind, type ContractSpan, type ContractVerdict, type ContractViolation, ConvergenceTracker, CorrectnessChecker, type CostEntry, CostLedger, type CostReport, type CostSummary, CostTracker, CreateChatClientOpts, type CreateDefaultReviewerOptions, type CreateExperimentInput, type CreateSandboxPoolOpts, CrossFamilyError, type CrossTraceDiff, type CrossTraceDiffOptions, DEFAULT_AGENT_SLOS, DEFAULT_FINDERS, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, DatasetScenario, type DecideNextUserTurnOpts, DefaultVerdict, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DescriptionLengthCandidate, type DescriptionLengthConfig, type DescriptionLengthDecision, type DescriptionLengthEvidence, DescriptionLengthGate, type DescriptionLengthRejectionCode, type DetectorEvent, type DetectorSeverity, type DetectorSignal, type DiffScorecardOptions, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, DriverResult, DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type ErrorCountPattern, type ErrorStreakOptions, type EvalToolDef, EvalTraceStore, type EvolutionRound, type ExecutorConfig, type Expectation, type Experiment, type ExperimentProvenance, type ExperimentRep, type ExperimentStats, type ExperimentStore, ExperimentTracker, type ExperimentTrackerOptions, type ExperimentVerdict, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type FactorContribution, type FactorialCell, FailureClass, FeedbackLabel, FeedbackTrajectory, FeedbackTrajectoryStore, type FieldAgreementSpec, type FieldDestination, type FileChange, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GhCliClientOptions, type GoldScenario, type GoldSplit, type GoldenSeverity, type GoldenSpec, HARNESS_NATIVE_MODEL, HarnessConfig, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, HoldoutAuditor, type HttpGithubClientOptions, INTENT_MATCH_JUDGE_VERSION, type ImageData, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, JudgeError, type JudgeFamily, type JudgeFleetOptions, JudgeFn, JudgeParseError, type JudgeReplayResult, type JudgeRetryOutcome, type JudgeRetryPolicy, JudgeRunner, type JudgeScoreInput, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, Layer, LayerResult, type LeaderboardOptions, type LeaderboardRow, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmClientOptions, LlmSpan, MODEL_PRICING, type MakeEvalToolsConfig, type MatchResult, type MatcherResult, type MergeOptions, MetricsCollector, type ModelCostRollup, type ModelPreflight, type ModelSeats, ModelsUnreachableError, type MuffledFinder, type MuffledFinding, type MultiToolchainLayerConfig, type Mutator, Mutex, type NoLeakOptions, type NoProgressOptions, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, OtelExportConfig, OtelExporter, type OtelPipelineHandle, type OtelPipelineOptions, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, type ParseStudentLabel, type PartitionHeldOutOptions, PersonaConfig, type Playbook, type PlaybookEntry, type PoolSlot, type PrReviewAuditCase, type PrReviewBenchmarkSummary, type PrReviewComment, type PrReviewMatchedFinding, type PrReviewOutcome, type PrReviewReferenceFinding, type PrReviewScore, type PrReviewScoreWeights, type PrReviewSeverity, type PrReviewSource, type PreflightModelsOptions, type PreflightOutcome, ProductClient, ProductClientConfig, type ProfileAxisSpec, type PromptHandle, PromptRegistry, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ProvenanceReader, type RecordRunsOptions, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, ReleaseConfidenceScorecard, ReleaseConfidenceThresholds, type RenderStudentPrompt, type RepeatedActionOptions, type RepoRef, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, type RoutedField, Run, type RunCommandInput, type RunCommandResult, type RunDistillationOptions, type RunDistillationResult, RunFilter, RunRecord, type RunRecordBackend, type RunRecordFilter, RunScore, RunScoreWeights, RunSplitTag, SandboxDriver, SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type ScanOptions, Scenario, type ScenarioCost, ScenarioFile, ScenarioRegistry, ScenarioResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type ScoredTarget, type SeatName, type SeatPresetName, SeatUnsetError, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SerializedRegex, Severity, type SingleBackendDivergence, SingleBackendError, type SingleBackendField, type SingleBackendReport, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, type SpanPredicate, type SplitGoldOptions, SteeringBundle, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type StepAttribution, type StreamingDetector, type SynthesisReason, type SynthesisTarget, TestResult, type TextMatcher, type ThresholdContract, TokenCounter, type TokenSpec, type ToolMatcher, ToolSpan, TraceAnalystSpan, type TraceContract, TraceContractBuilder, TraceEmitter, TraceStore, type TracedAnalystOptions, type TracedJudgeOptions, Trajectory, TrajectoryStep, type TreatmentClass, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, TurnMetrics, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, type UiFinding, type UiFindingScreenshot, type UiFindingSeverity, type UiLens, type VerdictCacheStats, type VerdictCacheStore, VerifyContext, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkerDriverContext, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, adversarialJudge, agentProfileHash, agentProfileId, agentProfileModelId, agentVisibleFields, aggregateJudgeVerdicts, aggregatePrReviewScore, analyzeAntiSlop, appendScorecard, assertCrossFamily, assertModelsServed, assertNoHiddenLeak, assertSingleBackend, assignHeldOutTag, attachCostToReport, attest, bisect, blendHeldout, buildAgreementJudge, buildDriverSystemPrompt, buildReflectionPrompt, buildReviewerPrompt, buildWorkerDriverSystemPrompt, cachedJudge, canaryLeakView, canonicalJson, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, classifyTreatment, codeExecutionJudge, coherenceJudge, collectionPreserved, commentsForSource, commitBisect, compareReferenceReplay, compilerJudge, computeExperimentStats, contentHash, contractJudge, costReport, createAntiSlopJudge, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createSandboxPool, crossTraceDiff, dataDescriptionBits, decideNextUserTurn, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultBlendWeights, defaultJudges, defaultParseStudentLabel, defaultReferenceReplayMatcher, defaultRenderStudentPrompt, deployGateLayer, diffScorecard, discoverPersonas, distillPlaybook, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateContract, evaluateOracles, evaluateTraceContract, executeScenario, expandProfileAxes, expectAgent, exportRewardModel, extractAssetUrls, extractErrorCount, fieldAgreement, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, gateTreatmentApplied, gateTreatmentFromMetrics, gateTreatmentFromSpans, gateTreatmentFromToolSpans, ghCliClient, gitProvenanceReader, precision as goldenPrecision, gradeOnHidden, harnessAxisOf, hashContent, hashToUnit, hiddenGrade, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryRunRecordBackend, inMemoryVerdictCache, isHiddenDestination, isModelPriced, isOtelConfigured, jsonShape, jsonlReferenceReplayStore, jsonlRunRecordBackend, judgeFamily, keyPreserved, leaderboard, linterJudge, loadGoldScenarios, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, matchGoldens, matchSpan, mergeLayerResults, modelDescriptionBits, multiToolchainLayer, noProgressDetector, notBlocked, observeAll, paraphraseRobustness, paraphraseRobustnessScenarios, parseGoldJsonl, parseReflectionResponse, partitionHeldOut, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, preflightModels, printDriverSummary, index as profile, promptBisect, proposeSynthesisTargets, recordRuns, recordRunsToScorecard, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, repeatedActionDetector, replayScorerOverCorpus, replayTraceThroughJudge, resolveModelPricing, resolveSeat, routeFields, rowCount, rowWhere, runAssertions, runBehavioralCanaries, runCanaries, runDistillation, runE2EWorkflow, runExpectations, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runReferenceReplay, runScore, runSelfPlay, scanForMuffledGates, scoreContinuity, scorePrReviewComments, scorePrReviewSource, scoreReferenceReplay, seatPresets, securityJudge, sentenceReorderMutator, splitGold, statusAdvanced, summarizePrReviewBenchmark, testJudge, textInSnapshot, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, traceContract, traceJudge, traceJudgeEnsemble, tracedAnalyzeTraces, typoMutator, urlContains, verifyAttestation, visualDiff, viteDeployRunner, weightedRecall, whitespaceCollapseMutator, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner };

package/dist/index.js

@@ -9,7 +9,7 @@ import {
   checkBehavioralCanary,
   checkCanaries,
   runBehavioralCanaries
-} from "./chunk-X5OUZB4T.js";
+} from "./chunk-JCUREYF5.js";
 import {
   classifyEuAiRisk,
   euAiActReport,
@@ -45,6 +45,7 @@ import {
 } from "./chunk-63MBSQTX.js";
 import {
   CODING_HARNESSES,
+  HARNESS_NATIVE_MODEL,
   JudgeParseError,
   adversarialJudge,
   agentProfileHash,
@@ -64,7 +65,7 @@ import {
   llmJudge,
   parseCorrectnessResponse,
   verifyCompletion
-} from "./chunk-GUII3E73.js";
+} from "./chunk-CMJSTXUR.js";
 import {
   DEFAULT_MUTATION_PRIMITIVES,
   DEFAULT_RED_TEAM_CORPUS,
@@ -88,7 +89,7 @@ import {
   scoreRedTeamOutput,
   surfaceContentHash,
   toolNamesForRun
-} from "./chunk-NK77GPUH.js";
+} from "./chunk-52CCCXU3.js";
 import {
   BackendIntegrityError,
   assertRealBackend,
@@ -98,7 +99,7 @@ import {
   fileVerdictCache,
   inMemoryVerdictCache,
   summarizeBackendIntegrity
-} from "./chunk-XIOQHCHU.js";
+} from "./chunk-LSCBODPQ.js";
 import {
   MODEL_PRICING,
   MetricsCollector,
@@ -10035,6 +10036,7 @@ export {
   FileSystemRawProviderSink,
   FileSystemTraceStore,
   FindingsStore,
+  HARNESS_NATIVE_MODEL,
   HeldOutGate,
   HoldoutAuditor,
   HoldoutLockedError,