@tangle-network/agent-eval 0.94.0 → 0.95.0

package/dist/campaign/worktree/index.d.ts

@@ -0,0 +1,53 @@
+/**
+ * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like
+ * (multiple commits allowed). A code-tier proposer's `propose()` creates a
+ * worktree, an agent commits the change into it, and `finalize()` returns a
+ * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker
+ * against the changed code. On promotion the worktree becomes the PR branch.
+ *
+ * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))
+ * adapter can slot in without touching proposer code. Only the git adapter
+ * ships today. See `docs/design/loop-taxonomy.md`.
+ */
+import type { CodeSurface } from '../types';
+export interface Worktree {
+    /** Absolute path to the checked-out worktree directory. */
+    path: string;
+    /** The branch the worktree is on (becomes the PR branch on promotion). */
+    branch: string;
+    /** The ref the worktree was forked from. */
+    baseRef: string;
+}
+export interface WorktreeAdapter {
+    /** Create an isolated worktree on a fresh branch off `baseRef`. */
+    create(opts: {
+        baseRef: string;
+        label: string;
+    }): Promise<Worktree>;
+    /** Commit any pending changes in the worktree, then return a CodeSurface
+     *  pointing at it. The agent has already written its change into
+     *  `worktree.path` by the time this is called. */
+    finalize(worktree: Worktree, summary: string): Promise<CodeSurface>;
+    /** Remove the worktree (and its branch) — called for losing candidates. */
+    discard(worktree: Worktree): Promise<void>;
+}
+export declare class WorktreeAdapterError extends Error {
+    readonly cause?: unknown | undefined;
+    constructor(message: string, cause?: unknown | undefined);
+}
+export interface GitWorktreeAdapterOptions {
+    /** Repo root the worktrees fork from. */
+    repoRoot: string;
+    /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */
+    worktreeDir?: string;
+    /** Branch-name prefix. Default: `improve`. */
+    branchPrefix?: string;
+    /** Test seam — defaults to a real `git` runner. */
+    git?: (args: string[], cwd: string) => string;
+}
+export declare function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter;
+/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can
+ *  run the worker in. A path ref is returned as-is; anything else is treated
+ *  as a ref under the adapter's worktree dir. */
+export declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
+//# sourceMappingURL=index.d.ts.map

package/dist/campaign/worktree/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/campaign/worktree/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAKH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,UAAU,CAAA;AAE3C,MAAM,WAAW,QAAQ;IACvB,2DAA2D;IAC3D,IAAI,EAAE,MAAM,CAAA;IACZ,0EAA0E;IAC1E,MAAM,EAAE,MAAM,CAAA;IACd,4CAA4C;IAC5C,OAAO,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,eAAe;IAC9B,mEAAmE;IACnE,MAAM,CAAC,IAAI,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,GAAG,OAAO,CAAC,QAAQ,CAAC,CAAA;IACnE;;sDAEkD;IAClD,QAAQ,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,WAAW,CAAC,CAAA;IACnE,2EAA2E;IAC3E,OAAO,CAAC,QAAQ,EAAE,QAAQ,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;CAC3C;AAED,qBAAa,oBAAqB,SAAQ,KAAK;IAG3C,QAAQ,CAAC,KAAK,CAAC,EAAE,OAAO;gBADxB,OAAO,EAAE,MAAM,EACN,KAAK,CAAC,EAAE,OAAO,YAAA;CAK3B;AAED,MAAM,WAAW,yBAAyB;IACxC,yCAAyC;IACzC,QAAQ,EAAE,MAAM,CAAA;IAChB,+EAA+E;IAC/E,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,8CAA8C;IAC9C,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,mDAAmD;IACnD,GAAG,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE,GAAG,EAAE,MAAM,KAAK,MAAM,CAAA;CAC9C;AAyBD,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,yBAAyB,GAAG,eAAe,CAqCnF;AAED;;iDAEiD;AACjD,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,WAAW,EAAE,WAAW,CAAC,EAAE,MAAM,GAAG,MAAM,CAItF"}

package/dist/canary.d.ts

@@ -0,0 +1,101 @@
+/**
+ * Liveness canaries — cheap statistical checks that catch the failure
+ * modes a green test suite never sees.
+ *
+ * Three canary types in this module:
+ *
+ *   1. **Silent judge fallback** — the judge degraded to a fallback
+ *      path (rules-only / cached / heuristic) without anyone
+ *      noticing. Signature: a string of consecutive runs whose
+ *      `judgeMetadata.confidence` equals a known fallback constant
+ *      (default 0.30) OR whose `judgeMetadata.fallback` is true.
+ *
+ *   2. **Judge calibration drift** — the judge's confidence
+ *      distribution has drifted from a historical window. Two-sample
+ *      Kolmogorov-Smirnov test on the recent vs historical confidences,
+ *      with the empirical-CDF max-difference statistic.
+ *
+ *   3. **Eval-set distribution shift** — the mix of categories /
+ *      buckets in the recent runs differs significantly from the
+ *      historical mix. Chi-square test on the binned counts.
+ *
+ * Outputs are alerts. The canary does NOT fail loud the way a test
+ * does — failing tests are reserved for hard correctness violations.
+ * A canary that fires is a *signal* to investigate, not a verdict.
+ *
+ * Why this lives here rather than in `observability.ts`: that module
+ * exports already, and is a pure-fanout-to-Langfuse/Prometheus
+ * adapter. Canaries are statistical detectors, not adapters.
+ */
+import type { RunRecord } from './run-record';
+export type CanaryKind = 'silent_judge_fallback' | 'judge_calibration_drift' | 'distribution_shift';
+export type CanarySeverity = 'info' | 'warn' | 'error';
+export interface CanaryAlert {
+    kind: CanaryKind;
+    severity: CanarySeverity;
+    message: string;
+    /** Numbers that informed the decision — drop straight into a
+     *  dashboard / paper figure. */
+    evidence: Record<string, unknown>;
+}
+export interface CanaryReport {
+    alerts: CanaryAlert[];
+    /** Per-kind summary count. */
+    counts: Record<CanaryKind, number>;
+}
+export interface CanaryOptions {
+    /**
+     * Silent-fallback detection.
+     * - `constant`: confidence value treated as the fallback signal.
+     *   Default 0.30 (matches the soft-fail default in
+     *   `propose-review.ts`).
+     * - `consecutiveThreshold`: trip the alert after this many
+     *   consecutive runs at `constant` (or `fallback === true`).
+     *   Default 3.
+     */
+    silentFallback?: {
+        constant?: number;
+        consecutiveThreshold?: number;
+        /** Floating-point tolerance when comparing against `constant`. */
+        epsilon?: number;
+    };
+    /**
+     * Calibration-drift detection.
+     * - `historyWindow`: number of past runs (oldest-first) treated as
+     *   the historical baseline. Default 50.
+     * - `recentWindow`: number of recent runs (newest-first) compared
+     *   against history. Default 20.
+     * - `ksAlpha`: alpha for the KS statistic vs critical value.
+     *   Default 0.05.
+     * - `minRecent`: minimum recent runs required to even attempt the
+     *   check. Default 10.
+     */
+    calibrationDrift?: {
+        historyWindow?: number;
+        recentWindow?: number;
+        ksAlpha?: number;
+        minRecent?: number;
+    };
+    /**
+     * Distribution-shift detection.
+     * - `category`: function that maps a run to a categorical bucket.
+     *   Required to enable this canary; if omitted the chi-square check
+     *   is skipped entirely.
+     * - `chiSquareAlpha`: alpha. Default 0.05.
+     * - `historyWindow`, `recentWindow`, `minRecent`: like above.
+     */
+    distributionShift?: {
+        category: (run: RunRecord) => string | null;
+        chiSquareAlpha?: number;
+        historyWindow?: number;
+        recentWindow?: number;
+        minRecent?: number;
+    };
+}
+/**
+ * Run all configured canaries against a chronological run list.
+ * Runs MUST be sorted oldest-to-newest by the caller — the order of
+ * the input is used to define "recent" vs "historical" windows.
+ */
+export declare function runCanaries(runs: RunRecord[], opts?: CanaryOptions): CanaryReport;
+//# sourceMappingURL=canary.d.ts.map

package/dist/canary.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"canary.d.ts","sourceRoot":"","sources":["../src/canary.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAEH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,cAAc,CAAA;AAE7C,MAAM,MAAM,UAAU,GAAG,uBAAuB,GAAG,yBAAyB,GAAG,oBAAoB,CAAA;AAEnG,MAAM,MAAM,cAAc,GAAG,MAAM,GAAG,MAAM,GAAG,OAAO,CAAA;AAEtD,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,UAAU,CAAA;IAChB,QAAQ,EAAE,cAAc,CAAA;IACxB,OAAO,EAAE,MAAM,CAAA;IACf;oCACgC;IAChC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAClC;AAED,MAAM,WAAW,YAAY;IAC3B,MAAM,EAAE,WAAW,EAAE,CAAA;IACrB,8BAA8B;IAC9B,MAAM,EAAE,MAAM,CAAC,UAAU,EAAE,MAAM,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,aAAa;IAC5B;;;;;;;;OAQG;IACH,cAAc,CAAC,EAAE;QACf,QAAQ,CAAC,EAAE,MAAM,CAAA;QACjB,oBAAoB,CAAC,EAAE,MAAM,CAAA;QAC7B,kEAAkE;QAClE,OAAO,CAAC,EAAE,MAAM,CAAA;KACjB,CAAA;IAED;;;;;;;;;;OAUG;IACH,gBAAgB,CAAC,EAAE;QACjB,aAAa,CAAC,EAAE,MAAM,CAAA;QACtB,YAAY,CAAC,EAAE,MAAM,CAAA;QACrB,OAAO,CAAC,EAAE,MAAM,CAAA;QAChB,SAAS,CAAC,EAAE,MAAM,CAAA;KACnB,CAAA;IAED;;;;;;;OAOG;IACH,iBAAiB,CAAC,EAAE;QAClB,QAAQ,EAAE,CAAC,GAAG,EAAE,SAAS,KAAK,MAAM,GAAG,IAAI,CAAA;QAC3C,cAAc,CAAC,EAAE,MAAM,CAAA;QACvB,aAAa,CAAC,EAAE,MAAM,CAAA;QACtB,YAAY,CAAC,EAAE,MAAM,CAAA;QACrB,SAAS,CAAC,EAAE,MAAM,CAAA;KACnB,CAAA;CACF;AAED;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,IAAI,EAAE,SAAS,EAAE,EAAE,IAAI,GAAE,aAAkB,GAAG,YAAY,CAarF"}

package/dist/causal-attribution.d.ts

@@ -0,0 +1,45 @@
+/**
+ * Causal attribution via factorial experiments.
+ *
+ * Run every combination of {model × prompt × scenario × seed}, then
+ * decompose observed score variance into main effects + interactions.
+ * Moves from correlational "variant B is better" to causal "the model
+ * swap accounts for 42% of the lead; the prompt change accounts for 28%;
+ * interaction is 30%."
+ *
+ * Minimal implementation: 2-way factorial (two factors at a time) with
+ * main-effect + interaction decomposition via variance of cell means.
+ * Consumers run the factorial design themselves (we don't schedule
+ * runs); this module consumes the (factorLevels, observedScores)
+ * table and does the attribution math.
+ */
+export interface FactorialCell {
+    /** Map factor name → level id. e.g. { model: 'claude', prompt: 'v2' } */
+    levels: Record<string, string>;
+    /** Observed score for this cell (mean over replications if n > 1). */
+    score: number;
+    /** Number of replications averaged to produce `score`. */
+    n: number;
+}
+export interface FactorContribution {
+    factor: string;
+    /** Variance attributed to this factor's main effect, as a fraction of total. */
+    shareOfVariance: number;
+    /** Range of cell means across levels of this factor. */
+    range: number;
+}
+export interface InteractionContribution {
+    factors: [string, string];
+    shareOfVariance: number;
+}
+export interface CausalAttributionReport {
+    totalVariance: number;
+    mainEffects: FactorContribution[];
+    interactions: InteractionContribution[];
+    /** Residual = variance unexplained by main effects + modeled interactions. */
+    residualShare: number;
+    /** Sanity: shares sum to 1 (within fp). */
+    sharesSum: number;
+}
+export declare function causalAttribution(cells: FactorialCell[]): CausalAttributionReport;
+//# sourceMappingURL=causal-attribution.d.ts.map

package/dist/causal-attribution.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"causal-attribution.d.ts","sourceRoot":"","sources":["../src/causal-attribution.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,MAAM,WAAW,aAAa;IAC5B,yEAAyE;IACzE,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC9B,sEAAsE;IACtE,KAAK,EAAE,MAAM,CAAA;IACb,0DAA0D;IAC1D,CAAC,EAAE,MAAM,CAAA;CACV;AAED,MAAM,WAAW,kBAAkB;IACjC,MAAM,EAAE,MAAM,CAAA;IACd,gFAAgF;IAChF,eAAe,EAAE,MAAM,CAAA;IACvB,wDAAwD;IACxD,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,uBAAuB;IACtC,OAAO,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACzB,eAAe,EAAE,MAAM,CAAA;CACxB;AAED,MAAM,WAAW,uBAAuB;IACtC,aAAa,EAAE,MAAM,CAAA;IACrB,WAAW,EAAE,kBAAkB,EAAE,CAAA;IACjC,YAAY,EAAE,uBAAuB,EAAE,CAAA;IACvC,8EAA8E;IAC9E,aAAa,EAAE,MAAM,CAAA;IACrB,2CAA2C;IAC3C,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,aAAa,EAAE,GAAG,uBAAuB,CA8DjF"}

package/dist/{chunk-2K6UUZ7P.js → chunk-2T4EZACH.js}

@@ -538,4 +538,4 @@ export {
   inMemoryCampaignStorage,
   runCampaign
 };
-//# sourceMappingURL=chunk-2K6UUZ7P.js.map
+//# sourceMappingURL=chunk-2T4EZACH.js.map

package/dist/chunk-2T4EZACH.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/campaign/run-campaign.ts","../src/integrity/backend-integrity.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 { createHash } from 'node:crypto'\nimport { join } from 'node:path'\nimport { BackendIntegrityError, type BackendIntegrityReport } from '../integrity/backend-integrity'\nimport { confidenceInterval } from '../statistics'\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  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. */\n  runDir: 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  storage.ensureDir(opts.runDir)\n\n  const manifestHash = computeManifestHash({\n    scenarios: opts.scenarios,\n    judges: judges as unknown as JudgeConfig<unknown>[],\n    dispatchRef: opts.dispatch.name || 'anonymous',\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: Array<{ scenario: TScenario; rep: number; cellId: string; cellSeed: number }> = []\n  let cellIndex = 0\n  for (const scenario of opts.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\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 raw = storage.read(cachePath)\n    if (raw !== undefined) {\n      try {\n        const cached = JSON.parse(raw) as CampaignCellResult<TArtifact>\n        if (cached.cellId === args.slot.cellId) {\n          return { cell: { ...cached, cached: true }, artifactsByPath: {} }\n        }\n      } catch {\n        // Corrupt cache — fall through to re-run.\n      }\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: ${err instanceof Error ? err.message : String(err)}`\n        break\n      }\n    }\n  }\n\n  await trace.flush()\n\n  const cell: CampaignCellResult<TArtifact> = {\n    cellId: args.slot.cellId,\n    scenarioId: args.slot.scenario.id,\n    rep: args.slot.rep,\n    artifact: (artifact ?? null) as TArtifact,\n    judgeScores,\n    costUsd: costSoFar,\n    tokenUsage: { ...tokensSoFar },\n    durationMs: Date.now() - startMs,\n    seed: args.slot.cellSeed,\n    cached: false,\n    error: errorMessage,\n  }\n\n  if (!errorMessage && args.resumable) {\n    storage.write(cachePath, JSON.stringify(cell))\n  }\n\n  return { cell, artifactsByPath }\n}\n\n/**\n * Per-cell stub guard. A cell that produced an artifact (no error) but reported\n * `costUsd === 0` AND zero tokens means the dispatch never called `ctx.cost` —\n * i.e. it ran against a stub or silently dropped its usage. `'warn'` logs it,\n * `'assert'` throws (fail-fast), `'off'` skips. An errored/skipped cell or a\n * deterministic judge-only run that genuinely made no LLM call is not flagged.\n */\nfunction enforceCellUsage<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  mode: 'assert' | 'warn' | 'off',\n): void {\n  if (mode === 'off' || cell.error) return\n  if (cell.artifact === null || cell.artifact === undefined) return\n  const zeroTokens = cell.tokenUsage.input === 0 && cell.tokenUsage.output === 0\n  if (cell.costUsd !== 0 || !zeroTokens) return\n  const msg = `cell '${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\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  const canonical = {\n    scenarios: input.scenarios.map((s) => ({ id: s.id, kind: s.kind })),\n    judges: input.judges.map((j) => ({ name: j.name, dims: j.dimensions.map((d) => d.key) })),\n    dispatch: input.dispatchRef,\n    seed: input.seed,\n    reps: input.reps,\n  }\n  return createHash('sha256').update(JSON.stringify(canonical)).digest('hex')\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","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,kBAAkB;AAC3B,SAAS,YAAY;;;ACyCd,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;;;ACxLA,SAAS,qBAAqB;AAmCvB,SAAS,oBAAqC;AACnD,QAAM,cAAc,cAAc,YAAY,GAAG;AACjD,QAAM,EAAE,YAAY,WAAW,cAAc,cAAc,IAAI;AAAA,IAC7D;AAAA,EACF;AACA,SAAO;AAAA,IACL,UAAU,KAAK;AACb,UAAI,CAAC,WAAW,GAAG,EAAG,WAAU,KAAK,EAAE,WAAW,KAAK,CAAC;AAAA,IAC1D;AAAA,IACA,OAAO,MAAM;AACX,aAAO,WAAW,IAAI;AAAA,IACxB;AAAA,IACA,KAAK,MAAM;AACT,UAAI;AACF,eAAO,aAAa,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;;;AF8BA,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,UAAQ,UAAU,KAAK,MAAM;AAE7B,QAAM,eAAe,oBAAoB;AAAA,IACvC,WAAW,KAAK;AAAA,IAChB;AAAA,IACA,aAAa,KAAK,SAAS,QAAQ;AAAA,IACnC;AAAA,IACA;AAAA,EACF,CAAC;AAED,QAAM,YAAY,IAAI;AACtB,QAAM,QAAyC,CAAC;AAChD,QAAM,kBAA0C,CAAC;AAGjD,QAAM,WAA0F,CAAC;AACjG,MAAI,YAAY;AAChB,aAAW,YAAY,KAAK,WAAW;AACrC,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;AAGA,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,UAAU,KAAK,KAAK,KAAK,QAAQ,KAAK,KAAK,OAAO,QAAQ,mBAAmB,GAAG,CAAC;AACvF,UAAQ,UAAU,OAAO;AAGzB,QAAM,YAAY,KAAK,SAAS,oBAAoB;AACpD,MAAI,KAAK,WAAW;AAClB,UAAM,MAAM,QAAQ,KAAK,SAAS;AAClC,QAAI,QAAQ,QAAW;AACrB,UAAI;AACF,cAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,YAAI,OAAO,WAAW,KAAK,KAAK,QAAQ;AACtC,iBAAO,EAAE,MAAM,EAAE,GAAG,QAAQ,QAAQ,KAAK,GAAG,iBAAiB,CAAC,EAAE;AAAA,QAClE;AAAA,MACF,QAAQ;AAAA,MAER;AAAA,IACF;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,WAAW,KAAK,SAAS,IAAI;AACnC,cAAQ,UAAU,KAAK,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,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;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,MAAM,KAAK,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;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,QAAM,YAAY;AAAA,IAChB,WAAW,MAAM,UAAU,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,MAAM,EAAE,KAAK,EAAE;AAAA,IAClE,QAAQ,MAAM,OAAO,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,MAAM,EAAE,WAAW,IAAI,CAAC,MAAM,EAAE,GAAG,EAAE,EAAE;AAAA,IACxF,UAAU,MAAM;AAAA,IAChB,MAAM,MAAM;AAAA,IACZ,MAAM,MAAM;AAAA,EACd;AACA,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,UAAU,SAAS,CAAC,EAAE,OAAO,KAAK;AAC5E;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":[]}

package/dist/{chunk-CTBHKLEU.js → chunk-77T4STFI.js}

@@ -1,20 +1,32 @@
 import {
   analyzeTraces
-} from "./chunk-TBDR6PAI.js";
+} from "./chunk-IZCEK2HR.js";
 import {
+  LLM_CACHED_TOKEN_ATTR_KEYS,
+  LLM_COST_ATTR_KEYS,
+  LLM_COST_USD,
+  LLM_INPUT_TOKENS,
+  LLM_INPUT_TOKEN_ATTR_KEYS,
+  LLM_MODEL_ATTR_KEYS,
+  LLM_MODEL_NAME,
+  LLM_OUTPUT_TOKENS,
+  LLM_OUTPUT_TOKEN_ATTR_KEYS,
+  OPENINFERENCE_SPAN_KIND,
+  TOOL_NAME,
   compareSpanTime,
   firstNumberAttr,
   firstStringAttr,
   projectOtlpFlatLine,
-  spanEpochMillis
-} from "./chunk-QMUEXQJS.js";
+  spanEpochMillis,
+  traceSpanKindToOpenInferenceKind
+} from "./chunk-YBIGNSCZ.js";
 import {
   defaultProviderRedactor,
   providerFromBaseUrl
 } from "./chunk-PC4UYEBM.js";
 import {
   validateRunRecord
-} from "./chunk-KWRRMR3J.js";
+} from "./chunk-LO6IOIJ2.js";
 import {
   canonicalize,
   hashJson
@@ -350,14 +362,26 @@ function nanoToIso(nano) {
   return Number.isFinite(ms) ? new Date(ms).toISOString() : (/* @__PURE__ */ new Date(0)).toISOString();
 }
 function applyOpenInference(attrs) {
-  if ("llm.model" in attrs && !("llm.model_name" in attrs)) {
-    attrs["llm.model_name"] = attrs["llm.model"];
+  if ("llm.model" in attrs && !(LLM_MODEL_NAME in attrs)) {
+    attrs[LLM_MODEL_NAME] = attrs["llm.model"];
   }
-  if ("tool.name" in attrs && !("inference.tool.name" in attrs)) {
-    attrs["inference.tool.name"] = attrs["tool.name"];
+  if ("llm.input_tokens" in attrs && !(LLM_INPUT_TOKENS in attrs)) {
+    attrs[LLM_INPUT_TOKENS] = attrs["llm.input_tokens"];
   }
-  if ("span.kind" in attrs && !("openinference.span.kind" in attrs)) {
-    attrs["openinference.span.kind"] = String(attrs["span.kind"]).toUpperCase();
+  if ("inference.llm.input_tokens" in attrs && !(LLM_INPUT_TOKENS in attrs)) {
+    attrs[LLM_INPUT_TOKENS] = attrs["inference.llm.input_tokens"];
+  }
+  if ("llm.output_tokens" in attrs && !(LLM_OUTPUT_TOKENS in attrs)) {
+    attrs[LLM_OUTPUT_TOKENS] = attrs["llm.output_tokens"];
+  }
+  if ("inference.llm.output_tokens" in attrs && !(LLM_OUTPUT_TOKENS in attrs)) {
+    attrs[LLM_OUTPUT_TOKENS] = attrs["inference.llm.output_tokens"];
+  }
+  if (TOOL_NAME in attrs && !("inference.tool.name" in attrs)) {
+    attrs["inference.tool.name"] = attrs[TOOL_NAME];
+  }
+  if ("span.kind" in attrs && !(OPENINFERENCE_SPAN_KIND in attrs)) {
+    attrs[OPENINFERENCE_SPAN_KIND] = String(attrs["span.kind"]).toUpperCase();
   }
 }
 function flattenOtlpExportToNdjson(otlpExport, opts = {}) {
@@ -400,39 +424,6 @@ function flattenOtlpExportToNdjson(otlpExport, opts = {}) {
 }
 
 // src/trace-analyst/otlp-to-run-records.ts
-var INPUT_TOKEN_KEYS = [
-  "llm.input_tokens",
-  "inference.llm.input_tokens",
-  "llm.token_count.prompt",
-  "gen_ai.usage.input_tokens",
-  "gen_ai.usage.prompt_tokens"
-];
-var OUTPUT_TOKEN_KEYS = [
-  "llm.output_tokens",
-  "inference.llm.output_tokens",
-  "llm.token_count.completion",
-  "gen_ai.usage.output_tokens",
-  "gen_ai.usage.completion_tokens"
-];
-var CACHED_TOKEN_KEYS = [
-  "llm.cached_tokens",
-  "inference.llm.cached_tokens",
-  "llm.token_count.prompt_cache_hit",
-  "gen_ai.usage.cached_tokens"
-];
-var COST_KEYS = [
-  "llm.cost_usd",
-  "inference.llm.cost.total",
-  "llm.cost.total",
-  "gen_ai.usage.cost"
-];
-var MODEL_KEYS = [
-  "llm.model_name",
-  "inference.llm.model_name",
-  "llm.model",
-  "gen_ai.request.model",
-  "gen_ai.response.model"
-];
 function otlpToRunRecords(otlpJsonl, opts) {
   return otlpToTraceRunRecords(otlpJsonl, opts).map((r) => r.record);
 }
@@ -538,16 +529,16 @@ function aggregateTrace(traceId, spans, fallbackModel) {
     if (s.end_time && (!latest || compareSpanTime(s.end_time, latest) > 0)) latest = s.end_time;
     if (s.kind === "LLM") {
       llmSpanCount += 1;
-      const i = firstNumberAttr(s.attributes, INPUT_TOKEN_KEYS);
-      const o = firstNumberAttr(s.attributes, OUTPUT_TOKEN_KEYS);
-      const c = firstNumberAttr(s.attributes, CACHED_TOKEN_KEYS);
+      const i = firstNumberAttr(s.attributes, LLM_INPUT_TOKEN_ATTR_KEYS);
+      const o = firstNumberAttr(s.attributes, LLM_OUTPUT_TOKEN_ATTR_KEYS);
+      const c = firstNumberAttr(s.attributes, LLM_CACHED_TOKEN_ATTR_KEYS);
       if (i !== null) input += i;
       if (o !== null) output += o;
       if (c !== null) {
         cached += c;
         sawCached = true;
       }
-      const m = firstStringAttr(s.attributes, MODEL_KEYS) ?? s.model_name;
+      const m = firstStringAttr(s.attributes, LLM_MODEL_ATTR_KEYS) ?? s.model_name;
       if (m) modelVotes.set(m, (modelVotes.get(m) ?? 0) + 1);
     } else if (s.kind === "TOOL") {
       toolSpanCount += 1;
@@ -599,7 +590,7 @@ function resolveCost(opts, agg, spans) {
   let sawCost = false;
   for (const s of spans) {
     if (s.kind !== "LLM") continue;
-    const c = firstNumberAttr(s.attributes, COST_KEYS);
+    const c = firstNumberAttr(s.attributes, LLM_COST_ATTR_KEYS);
     if (c !== null) {
       perSpanCost += c;
       sawCost = true;
@@ -642,7 +633,7 @@ function topVote(votes) {
 }
 function firstModelAttr(spans) {
   for (const s of spans) {
-    const m = firstStringAttr(s.attributes, MODEL_KEYS) ?? s.model_name;
+    const m = firstStringAttr(s.attributes, LLM_MODEL_ATTR_KEYS) ?? s.model_name;
     if (m) return m;
   }
   return null;
@@ -1222,16 +1213,16 @@ function spanToOtlp(span, traceId, events) {
 }
 function flattenSpanAttributes(span) {
   const base = {
-    "span.kind": span.kind
+    [OPENINFERENCE_SPAN_KIND]: traceSpanKindToOpenInferenceKind(span.kind)
   };
   if (span.kind === "llm") {
-    base["llm.model"] = span.model;
-    if (span.inputTokens !== void 0) base["llm.input_tokens"] = span.inputTokens;
-    if (span.outputTokens !== void 0) base["llm.output_tokens"] = span.outputTokens;
-    if (span.costUsd !== void 0) base["llm.cost_usd"] = span.costUsd;
+    base[LLM_MODEL_NAME] = span.model;
+    if (span.inputTokens !== void 0) base[LLM_INPUT_TOKENS] = span.inputTokens;
+    if (span.outputTokens !== void 0) base[LLM_OUTPUT_TOKENS] = span.outputTokens;
+    if (span.costUsd !== void 0) base[LLM_COST_USD] = span.costUsd;
     if (span.finishReason) base["llm.finish_reason"] = span.finishReason;
   } else if (span.kind === "tool") {
-    base["tool.name"] = span.toolName;
+    base[TOOL_NAME] = span.toolName;
     if (span.latencyMs !== void 0) base["tool.latency_ms"] = span.latencyMs;
   } else if (span.kind === "retrieval") {
     base["retrieval.query"] = span.query;
@@ -1441,12 +1432,12 @@ function parseHeadersFromEnv() {
 function toOtlpSpan(span) {
   const endedAt = span.endedAt ?? span.startedAt;
   const attrs = {
-    "span.kind": span.kind
+    [OPENINFERENCE_SPAN_KIND]: traceSpanKindToOpenInferenceKind(span.kind)
   };
-  if (span.model) attrs["llm.model"] = span.model;
-  if (span.inputTokens !== void 0) attrs["llm.input_tokens"] = span.inputTokens;
-  if (span.outputTokens !== void 0) attrs["llm.output_tokens"] = span.outputTokens;
-  if (span.costUsd !== void 0) attrs["llm.cost_usd"] = span.costUsd;
+  if (span.model) attrs[LLM_MODEL_NAME] = span.model;
+  if (span.inputTokens !== void 0) attrs[LLM_INPUT_TOKENS] = span.inputTokens;
+  if (span.outputTokens !== void 0) attrs[LLM_OUTPUT_TOKENS] = span.outputTokens;
+  if (span.costUsd !== void 0) attrs[LLM_COST_USD] = span.costUsd;
   if (span.attributes) {
     for (const [k, v] of Object.entries(span.attributes)) {
       if (typeof v === "string" || typeof v === "number" || typeof v === "boolean") attrs[k] = v;
@@ -1588,7 +1579,7 @@ function projectCell(args) {
     const runEnd = msToIso(run.endedAt ?? run.startedAt);
     const runStatus = run.outcome?.failureClass && run.outcome.failureClass !== "success" ? "STATUS_CODE_ERROR" : "STATUS_CODE_OK";
     const runAttrs = {
-      "openinference.span.kind": "AGENT",
+      [OPENINFERENCE_SPAN_KIND]: "AGENT",
       "agent.name": agentName,
       "agent.workflow.name": serviceName,
       ...runAttributes(run)
@@ -1695,13 +1686,13 @@ function readMergedShards(cellDir, name, idKey) {
 }
 function spanToAttributes(span, events) {
   const attrs = {
-    "openinference.span.kind": spanKindToOpenInferenceKind(span.kind)
+    [OPENINFERENCE_SPAN_KIND]: traceSpanKindToOpenInferenceKind(span.kind)
   };
   if (span.kind === "llm") {
-    attrs["llm.model_name"] = span.model;
-    if (span.inputTokens !== void 0) attrs["llm.token_count.prompt"] = span.inputTokens;
-    if (span.outputTokens !== void 0) attrs["llm.token_count.completion"] = span.outputTokens;
-    if (span.costUsd !== void 0) attrs["llm.cost_usd"] = span.costUsd;
+    attrs[LLM_MODEL_NAME] = span.model;
+    if (span.inputTokens !== void 0) attrs[LLM_INPUT_TOKENS] = span.inputTokens;
+    if (span.outputTokens !== void 0) attrs[LLM_OUTPUT_TOKENS] = span.outputTokens;
+    if (span.costUsd !== void 0) attrs[LLM_COST_USD] = span.costUsd;
     if (span.finishReason) attrs["llm.finish_reason"] = span.finishReason;
     if (Array.isArray(span.messages)) {
       attrs["llm.input_messages"] = JSON.stringify(span.messages.slice(-6));
@@ -1710,7 +1701,7 @@ function spanToAttributes(span, events) {
       attrs["llm.output_messages"] = JSON.stringify([{ role: "assistant", content: span.output }]);
     }
   } else if (span.kind === "tool") {
-    attrs["tool.name"] = span.toolName;
+    attrs[TOOL_NAME] = span.toolName;
     if (span.latencyMs !== void 0) attrs["tool.latency_ms"] = span.latencyMs;
     attrs["input.value"] = safeStringify(span.args);
     if (span.result !== void 0) attrs["output.value"] = safeStringify(span.result);
@@ -1741,24 +1732,6 @@ function spanKindToOtlpKind(kind) {
       return "SPAN_KIND_INTERNAL";
   }
 }
-function spanKindToOpenInferenceKind(kind) {
-  switch (kind) {
-    case "llm":
-      return "LLM";
-    case "tool":
-      return "TOOL";
-    case "retrieval":
-      return "CHAIN";
-    case "judge":
-      return "GUARDRAIL";
-    case "sandbox":
-      return "CHAIN";
-    case "agent":
-      return "AGENT";
-    default:
-      return "SPAN";
-  }
-}
 function toLine(args) {
   return {
     trace_id: args.traceId,
@@ -1981,4 +1954,4 @@ export {
   createReplayFetch,
   iterateRawCalls
 };
-//# sourceMappingURL=chunk-CTBHKLEU.js.map
+//# sourceMappingURL=chunk-77T4STFI.js.map