@tangle-network/agent-eval 0.59.1 → 0.60.0

package/dist/adapters/http.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-BgrxOJSf.js';
+import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-VCIXx_yo.js';
 
 /**
  * # `@tangle-network/agent-eval/adapters/http` — distributed Dispatch over HTTP.

package/dist/adapters/http.js

@@ -1,4 +1,4 @@
-import "../chunk-NSBPE2FW.js";
+import "../chunk-PZ5AY32C.js";
 
 // src/adapters/http.ts
 function resolveAuth(auth) {

package/dist/adapters/langchain.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-BgrxOJSf.js';
+import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-VCIXx_yo.js';
 
 /**
  * # `@tangle-network/agent-eval/adapters/langchain` — wrap any LangChain

package/dist/adapters/langchain.js

@@ -1,4 +1,4 @@
-import "../chunk-NSBPE2FW.js";
+import "../chunk-PZ5AY32C.js";
 
 // src/adapters/langchain.ts
 function langchainDispatch(opts) {

package/dist/adapters/otel.d.ts

@@ -1,5 +1,5 @@
-import { T as TraceSpanEvent, H as HostedClient } from '../index-D2nT6_KT.js';
-import '../types-BgrxOJSf.js';
+import { T as TraceSpanEvent, H as HostedClient } from '../index-BIkvdkSU.js';
+import '../types-VCIXx_yo.js';
 import '../summary-report-DLxh4yWk.js';
 import '../run-record-etiCMsUq.js';
 import '../errors-mje_cKOs.js';

package/dist/adapters/otel.js

@@ -1,4 +1,4 @@
-import "../chunk-NSBPE2FW.js";
+import "../chunk-PZ5AY32C.js";
 
 // src/adapters/otel.ts
 var OTEL_STATUS_UNSET = 0;

package/dist/benchmarks/index.js

@@ -2,8 +2,8 @@ import {
   BENCHMARK_SPLIT_SEED,
   deterministicSplit,
   routing_exports
-} from "../chunk-MHQPVHXU.js";
-import "../chunk-NSBPE2FW.js";
+} from "../chunk-6QDKWHLS.js";
+import "../chunk-PZ5AY32C.js";
 export {
   BENCHMARK_SPLIT_SEED,
   deterministicSplit,

package/dist/builder-eval/index.js

@@ -10,7 +10,7 @@ import {
   TraceEmitter
 } from "../chunk-TVVP3ZZQ.js";
 import "../chunk-QYJT52YW.js";
-import "../chunk-NSBPE2FW.js";
+import "../chunk-PZ5AY32C.js";
 
 // src/builder-eval/builder-session.ts
 var BuilderSession = class {

package/dist/campaign/index.d.ts

@@ -1,6 +1,6 @@
-export { C as CampaignStorage, D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, G as GepaDriverConstraints, a as GepaDriverOptions, H as HeldOutGateOptions, O as OpenAutoPrOptions, b as OpenAutoPrResult, R as RunCampaignOptions, c as RunEvalOptions, d as RunImprovementLoopOptions, e as RunImprovementLoopResult, f as RunOptimizationOptions, g as RunOptimizationResult, h as composeGate, i as countSentenceEdits, j as defaultProductionGate, k as evolutionaryDriver, l as extractH2Sections, m as fsCampaignStorage, n as gepaDriver, o as heldOutGate, p as inMemoryCampaignStorage, q as openAutoPr, r as runCampaign, s as runEval, t as runImprovementLoop, u as runOptimization, v as surfaceHash } from '../run-improvement-loop-BhfdjrMY.js';
-import { L as LabeledScenarioStore, c as LabeledScenarioWrite, d as LabeledScenarioSampleArgs, e as LabeledScenarioRecord, f as LabelTrust, C as CodeSurface } from '../types-BgrxOJSf.js';
-export { g as CampaignAggregates, h as CampaignArtifactWriter, i as CampaignCellResult, j as CampaignCostMeter, k as CampaignResult, l as CampaignTraceWriter, b as DispatchContext, D as DispatchFn, G as Gate, m as GateContext, n as GateDecision, o as GateResult, p as GenerationCandidate, q as GenerationRecord, I as ImprovementDriver, r as JudgeAggregate, a as JudgeConfig, s as JudgeDimension, J as JudgeScore, t as LabeledScenarioSource, M as MutableSurface, u as Mutator, O as OptimizerConfig, P as ProposeContext, R as RedactionStatus, S as Scenario, v as ScenarioAggregate, w as SessionScript, T as TraceSpan, x as labelTrustRank } from '../types-BgrxOJSf.js';
+export { B as BuildLoopProvenanceArgs, C as CampaignStorage, D as DefaultProductionGateOptions, E as EmitLoopProvenanceArgs, a as EmitLoopProvenanceResult, b as EvolutionaryDriverOptions, G as GepaDriverConstraints, c as GepaDriverOptions, H as HeldOutGateOptions, L as LoopProvenanceBackend, d as LoopProvenanceCandidate, e as LoopProvenanceRecord, O as OpenAutoPrOptions, f as OpenAutoPrResult, R as RunCampaignOptions, g as RunEvalOptions, h as RunImprovementLoopOptions, i as RunImprovementLoopResult, j as RunOptimizationOptions, k as RunOptimizationResult, l as buildLoopProvenanceRecord, m as composeGate, n as countSentenceEdits, o as defaultProductionGate, p as defaultRenderDiff, q as emitLoopProvenance, r as evolutionaryDriver, s as extractH2Sections, t as fsCampaignStorage, u as gepaDriver, v as heldOutGate, w as inMemoryCampaignStorage, x as loopProvenanceSpans, y as openAutoPr, z as provenanceRecordPath, A as provenanceSpansPath, F as runCampaign, I as runEval, J as runImprovementLoop, K as runOptimization, M as surfaceContentHash, N as surfaceHash } from '../provenance-BM8vmMBa.js';
+import { L as LabeledScenarioStore, c as LabeledScenarioWrite, d as LabeledScenarioSampleArgs, e as LabeledScenarioRecord, f as LabelTrust, C as CodeSurface } from '../types-VCIXx_yo.js';
+export { g as CampaignAggregates, h as CampaignArtifactWriter, i as CampaignCellResult, j as CampaignCostMeter, k as CampaignResult, l as CampaignTraceWriter, b as DispatchContext, D as DispatchFn, G as Gate, m as GateContext, n as GateDecision, o as GateResult, p as GenerationCandidate, q as GenerationRecord, I as ImprovementDriver, r as JudgeAggregate, a as JudgeConfig, s as JudgeDimension, J as JudgeScore, t as LabeledScenarioSource, M as MutableSurface, u as Mutator, O as OptimizerConfig, P as ProposeContext, v as ProposedCandidate, R as RedactionStatus, S as Scenario, w as ScenarioAggregate, x as SessionScript, T as TraceSpan, y as isProposedCandidate, z as labelTrustRank } from '../types-VCIXx_yo.js';
 import '../llm-client-BXVRUZyX.js';
 import '../errors-mje_cKOs.js';
 import '../raw-provider-sink-C46HDghv.js';
@@ -9,6 +9,10 @@ import '../dataset-BlwAtYYf.js';
 import '../store-CKUAgsJz.js';
 import '../schema-m0gsnbt3.js';
 import '../run-record-etiCMsUq.js';
+import '../index-BIkvdkSU.js';
+import '../summary-report-DLxh4yWk.js';
+import '../failure-cluster-CL7IVgkJ.js';
+import '../judge-calibration-DilmB3Ml.js';
 
 /**
  * @experimental

package/dist/campaign/index.js

@@ -1,47 +1,44 @@
 import {
+  buildLoopProvenanceRecord,
   composeGate,
   countSentenceEdits,
   defaultProductionGate,
+  defaultRenderDiff,
+  emitLoopProvenance,
   evolutionaryDriver,
   extractH2Sections,
   gepaDriver,
   heldOutGate,
+  isProposedCandidate,
+  labelTrustRank,
+  loopProvenanceSpans,
   openAutoPr,
+  provenanceRecordPath,
+  provenanceSpansPath,
   runEval,
   runImprovementLoop,
   runOptimization,
+  surfaceContentHash,
   surfaceHash
-} from "../chunk-JB4UWIM6.js";
+} from "../chunk-LBSXXH56.js";
 import {
   fsCampaignStorage,
   inMemoryCampaignStorage,
   runCampaign
-} from "../chunk-74Y2EMNH.js";
-import "../chunk-N4SBKEPJ.js";
+} from "../chunk-NOPYCRNG.js";
+import "../chunk-GBHRUAOF.js";
 import "../chunk-YV7J7X5N.js";
 import "../chunk-S3SDD56V.js";
 import "../chunk-GGE4NNQT.js";
 import "../chunk-VXNVVBZO.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-QYJT52YW.js";
-import "../chunk-NSBPE2FW.js";
+import "../chunk-PZ5AY32C.js";
 
 // src/campaign/labeled-store/fs-adapter.ts
 import { createHash } from "crypto";
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
 import { join } from "path";
-
-// src/campaign/types.ts
-var LABEL_TRUST_RANK = {
-  unverified: 0,
-  "verified-signal": 1,
-  "human-rated": 2
-};
-function labelTrustRank(trust) {
-  return LABEL_TRUST_RANK[trust ?? "unverified"];
-}
-
-// src/campaign/labeled-store/fs-adapter.ts
 var LabeledScenarioStoreError = class extends Error {
   constructor(code, message) {
     super(message);
@@ -312,9 +309,12 @@ export {
   FsLabeledScenarioStore,
   LabeledScenarioStoreError,
   WorktreeAdapterError,
+  buildLoopProvenanceRecord,
   composeGate,
   countSentenceEdits,
   defaultProductionGate,
+  defaultRenderDiff,
+  emitLoopProvenance,
   evolutionaryDriver,
   extractH2Sections,
   fsCampaignStorage,
@@ -322,13 +322,18 @@ export {
   gitWorktreeAdapter,
   heldOutGate,
   inMemoryCampaignStorage,
+  isProposedCandidate,
   labelTrustRank,
+  loopProvenanceSpans,
   openAutoPr,
+  provenanceRecordPath,
+  provenanceSpansPath,
   resolveWorktreePath,
   runCampaign,
   runEval,
   runImprovementLoop,
   runOptimization,
+  surfaceContentHash,
   surfaceHash
 };
 //# sourceMappingURL=index.js.map

package/dist/campaign/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/campaign/labeled-store/fs-adapter.ts","../../src/campaign/types.ts","../../src/campaign/worktree/index.ts"],"sourcesContent":["/**\n * @experimental\n *\n * Filesystem `LabeledScenarioStore` adapter. The default capture sink for\n * traces + eval artifacts. Production deployments typically swap for a\n * Turso/SQLite adapter (same interface).\n *\n * Records land as one JSONL file per source under `<root>/<source>.jsonl`.\n * Each line is a `LabeledScenarioRecord`. Append-only — no in-place edits.\n *\n * Safety properties enforced at write-time:\n *\n *   - **Provenance required**: writes without `source`, `sourceVersionHash`,\n *     `capturedAt`, `redactionStatus` are rejected. Closes the alignment\n *     reviewer's data-poisoning gap.\n *   - **Per-source rate limits**: optional `rateLimitBucket` + `maxWritesPerMinute`\n *     stops a single tenant/source from flooding the store.\n *\n * Safety properties enforced at sample-time:\n *\n *   - **Required split + capturedBefore**: substrate refuses to sample without\n *     an explicit `split` ('train' | 'test') AND a temporal cutoff. Eliminates\n *     accidental train/test contamination.\n *   - **Default training-source filter**: when the store is sampled with\n *     `split: 'train'`, production-trace records are EXCLUDED unless the\n *     caller passes `filter.source: 'production-trace'` explicitly. Closes\n *     the contamination-by-default gap flagged by the senior eval engineer.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type {\n  LabeledScenarioRecord,\n  LabeledScenarioSampleArgs,\n  LabeledScenarioSource,\n  LabeledScenarioStore,\n  LabeledScenarioWrite,\n  LabelTrust,\n} from '../types'\nimport { labelTrustRank } from '../types'\n\nexport interface FsLabeledScenarioStoreOptions {\n  /** Root directory for JSONL files. Created if missing. */\n  root: string\n  /** Per-source rate limit. When set, writes exceeding the cap are rejected\n   *  with a typed error. Default: no limit. */\n  maxWritesPerMinutePerBucket?: number\n  /** Test seam — override `Date.now()` for deterministic tests. */\n  now?: () => number\n}\n\nexport class LabeledScenarioStoreError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n  ) {\n    super(message)\n    this.name = 'LabeledScenarioStoreError'\n  }\n}\n\ninterface RateLimitState {\n  bucket: string\n  windowStartMs: number\n  count: number\n}\n\nexport class FsLabeledScenarioStore implements LabeledScenarioStore {\n  private readonly now: () => number\n  private readonly rateLimits = new Map<string, RateLimitState>()\n\n  constructor(private readonly options: FsLabeledScenarioStoreOptions) {\n    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true })\n    this.now = options.now ?? Date.now\n  }\n\n  async observe(write: LabeledScenarioWrite): Promise<void> {\n    this.assertProvenance(write)\n    this.assertRateLimit(write)\n    const record = this.toRecord(write)\n    const path = this.pathForSource(write.source)\n    const line = `${JSON.stringify(record)}\\n`\n    // Append atomically. For high-throughput a writev-friendly buffered\n    // implementation lands in the Turso adapter; FS adapter is for tests +\n    // local dev + small workloads.\n    appendLine(path, line)\n  }\n\n  async sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]> {\n    if (!args.split) {\n      throw new LabeledScenarioStoreError(\n        'split_required',\n        'sample() requires an explicit `split` (train | test) — substrate refuses ambiguous reads',\n      )\n    }\n    if (!args.capturedBefore) {\n      throw new LabeledScenarioStoreError(\n        'capturedBefore_required',\n        'sample() requires an explicit `capturedBefore` timestamp for temporal-split discipline',\n      )\n    }\n\n    const all: LabeledScenarioRecord[] = []\n    for (const source of ALL_SOURCES) {\n      // Default training-source filter: when sampling train, EXCLUDE\n      // production-trace records unless the caller asks for them.\n      if (args.split === 'train' && source === 'production-trace') {\n        const explicit = sourceFilterContains(args.filter?.source, 'production-trace')\n        if (!explicit) continue\n      }\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) continue\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      for (const line of lines) {\n        let record: LabeledScenarioRecord\n        try {\n          record = JSON.parse(line) as LabeledScenarioRecord\n        } catch {\n          continue\n        }\n        if (!matchesFilter(record, args, source)) continue\n        all.push(record)\n      }\n    }\n\n    // Deterministic order: by capturedAt ascending, then recordHash.\n    all.sort((a, b) => {\n      if (a.capturedAt !== b.capturedAt) return a.capturedAt.localeCompare(b.capturedAt)\n      return a.recordHash.localeCompare(b.recordHash)\n    })\n\n    return all.slice(0, args.count)\n  }\n\n  async size(): Promise<{\n    train: number\n    test: number\n    bySource: Record<string, number>\n    byTrust: Record<LabelTrust, number>\n  }> {\n    const bySource: Record<string, number> = {}\n    const byTrust: Record<LabelTrust, number> = {\n      unverified: 0,\n      'verified-signal': 0,\n      'human-rated': 0,\n    }\n    let total = 0\n    for (const source of ALL_SOURCES) {\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) {\n        bySource[source] = 0\n        continue\n      }\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      bySource[source] = lines.length\n      total += lines.length\n      for (const line of lines) {\n        let trust: LabelTrust = 'unverified'\n        try {\n          trust = (JSON.parse(line) as LabeledScenarioRecord).labelTrust ?? 'unverified'\n        } catch {\n          // A malformed line counts as unverified — never silently gold.\n        }\n        byTrust[trust] += 1\n      }\n    }\n    // FS adapter doesn't track split assignments per-record (split is\n    // computed at sample-time based on `capturedBefore`). For size(), we\n    // report `train`+`test` as the same total — split is a sampling concept.\n    return { train: total, test: total, bySource, byTrust }\n  }\n\n  private assertProvenance(write: LabeledScenarioWrite): void {\n    if (!write.source) {\n      throw new LabeledScenarioStoreError(\n        'missing_source',\n        'LabeledScenarioWrite requires `source`',\n      )\n    }\n    if (!write.sourceVersionHash || write.sourceVersionHash.length === 0) {\n      throw new LabeledScenarioStoreError(\n        'missing_source_version',\n        'LabeledScenarioWrite requires `sourceVersionHash` (git sha or substrate version)',\n      )\n    }\n    if (!write.capturedAt) {\n      throw new LabeledScenarioStoreError(\n        'missing_captured_at',\n        'LabeledScenarioWrite requires `capturedAt` ISO timestamp',\n      )\n    }\n    if (!write.redactionStatus) {\n      throw new LabeledScenarioStoreError(\n        'missing_redaction_status',\n        'LabeledScenarioWrite requires explicit `redactionStatus` — raw / redacted-pii / redacted-secrets / fully-redacted',\n      )\n    }\n    if (!ALL_SOURCES.includes(write.source)) {\n      throw new LabeledScenarioStoreError(\n        'unknown_source',\n        `LabeledScenarioWrite.source must be one of: ${ALL_SOURCES.join(', ')}`,\n      )\n    }\n  }\n\n  private assertRateLimit(write: LabeledScenarioWrite): void {\n    const cap = this.options.maxWritesPerMinutePerBucket\n    if (!cap || !write.rateLimitBucket) return\n    const now = this.now()\n    const windowMs = 60_000\n    let state = this.rateLimits.get(write.rateLimitBucket)\n    if (!state || now - state.windowStartMs >= windowMs) {\n      state = { bucket: write.rateLimitBucket, windowStartMs: now, count: 0 }\n      this.rateLimits.set(write.rateLimitBucket, state)\n    }\n    if (state.count >= cap) {\n      throw new LabeledScenarioStoreError(\n        'rate_limit_exceeded',\n        `LabeledScenarioStore: bucket ${write.rateLimitBucket} exceeded ${cap} writes/min`,\n      )\n    }\n    state.count += 1\n  }\n\n  private toRecord(write: LabeledScenarioWrite): LabeledScenarioRecord {\n    const recordHash = sha256(\n      JSON.stringify({\n        id: write.scenario.id,\n        src: write.source,\n        at: write.capturedAt,\n        ver: write.sourceVersionHash,\n      }),\n    )\n    // FS adapter assigns split at sample-time, but we cache a hint here\n    // based on capturedAt vs the world's \"now\" — sampler overrides this.\n    return {\n      ...write,\n      recordHash,\n      split: 'train',\n    }\n  }\n\n  private pathForSource(source: string): string {\n    return join(this.options.root, `${source}.jsonl`)\n  }\n}\n\nconst ALL_SOURCES: LabeledScenarioWrite['source'][] = [\n  'production-trace',\n  'eval-run',\n  'manual',\n  'red-team',\n  'synthetic',\n]\n\nfunction sourceFilterContains(\n  filter: LabeledScenarioSource | LabeledScenarioSource[] | undefined,\n  needle: LabeledScenarioSource,\n): boolean {\n  if (!filter) return false\n  if (Array.isArray(filter)) return filter.includes(needle)\n  return filter === needle\n}\n\nfunction matchesFilter(\n  record: LabeledScenarioRecord,\n  args: LabeledScenarioSampleArgs,\n  source: string,\n): boolean {\n  // Temporal cutoff — train must be capturedAt < capturedBefore.\n  if (args.split === 'train' && record.capturedAt >= args.capturedBefore) return false\n  if (args.split === 'test' && record.capturedAt < args.capturedBefore) return false\n\n  const f = args.filter\n  if (!f) return true\n  if (f.kind && record.scenario.kind !== f.kind) return false\n  if (f.source) {\n    const sources = Array.isArray(f.source) ? f.source : [f.source]\n    if (!sources.includes(source as never)) return false\n  }\n  if (f.minComposite !== undefined || f.maxComposite !== undefined) {\n    const composites = Object.values(record.judgeScores).map((s) => s.composite)\n    const max = composites.length === 0 ? 0 : Math.max(...composites)\n    if (f.minComposite !== undefined && max < f.minComposite) return false\n    if (f.maxComposite !== undefined && max > f.maxComposite) return false\n  }\n  if (f.minTrust !== undefined && labelTrustRank(record.labelTrust) < labelTrustRank(f.minTrust)) {\n    return false\n  }\n  return true\n}\n\nfunction sha256(input: string): string {\n  return createHash('sha256').update(input).digest('hex').slice(0, 16)\n}\n\nfunction appendLine(path: string, line: string): void {\n  if (existsSync(path)) {\n    const existing = readFileSync(path, 'utf8')\n    writeFileSync(path, existing + line)\n  } else {\n    writeFileSync(path, line)\n  }\n}\n","/**\n * @experimental\n *\n * Pass A substrate types — `runCampaign` is the one primitive every\n * eval flow composes from. Three contracts in this file:\n *\n *   - `Scenario`            input set\n *   - `DispatchFn`          how to run one scenario → artifact\n *   - `CampaignResult`      defined output schema (the contract downstream tools depend on)\n *\n * Three more lifted from earlier substrate work (re-exported):\n *\n *   - `JudgeConfig`         pluggable dimensional scorer (0.38)\n *   - `Mutator`             optimization-loop surface mutator\n *   - `Gate`                promotion gate (`HeldOutGate` and friends adapt to this)\n *\n * No new architecture vs 0.38 — Pass A formalizes the shapes so consumers\n * can build dashboards / CI gates / regression diffs against a stable schema.\n */\n\n/** @experimental Stable identifier + kind tag for any scenario. Consumers\n *  extend with their per-domain payload (persona, task, requirement, ...). */\nexport interface Scenario {\n  id: string\n  kind: string\n  tags?: string[]\n}\n\n/** @experimental Context handed to every dispatch invocation. Scoped — every\n *  trace/span carries the cellId, every artifact write lands under the cell's\n *  artifact root, the cost meter accumulates per cell. */\nexport interface DispatchContext {\n  cellId: string\n  rep: number\n  generation?: number\n  seed: number\n  signal: AbortSignal\n  trace: CampaignTraceWriter\n  artifacts: CampaignArtifactWriter\n  cost: CampaignCostMeter\n  /** Populated when this run is part of a multi-cycle improvement loop. */\n  cycleId?: string\n  /** Populated when the substrate resumed from a prior cache hit. */\n  resumedFrom?: string\n  /**\n   * Opaque placement key supplied by `RunCampaignOptions.cellPlacement`.\n   * The substrate forwards it through unchanged; placement-aware Dispatch\n   * implementations (e.g. `httpDispatch` from `/adapters/http`) read it to\n   * route the cell to the right worker / region / sandbox. `undefined`\n   * when no placement strategy is configured.\n   */\n  placement?: string\n}\n\n/** @experimental One function: scenario + ctx → artifact. Dispatcher chooses\n *  whether to call `runMultishot`, `runLoop`, raw `streamPrompt`, anything. */\nexport type DispatchFn<TScenario extends Scenario, TArtifact> = (\n  scenario: TScenario,\n  ctx: DispatchContext,\n) => Promise<TArtifact>\n\n// ── Sessions ──────────────────────────────────────────────────────────\n\n/** @experimental One session within a multi-session journey. Dispatch is\n *  invoked once per session in order; state from prior session's artifact\n *  is exposed via `ctx.priorSessionArtifact`. */\nexport interface SessionScript<TScenario, TArtifact> {\n  id: string\n  intent: string\n  maxTurns?: number\n  /** When true, knowledge accumulated this session persists to next. */\n  affectsKnowledge?: boolean\n  /** Optional per-session persona evolution — called after the session\n   *  resolves. Returns the persona shape used by the NEXT session. */\n  evolveAfterSession?: (artifact: TArtifact, sessionIndex: number, scenario: TScenario) => TScenario\n}\n\n// ── Judges (re-export 0.38 shape) ─────────────────────────────────────\n\nexport interface JudgeDimension {\n  /** JSON field name + score key. */\n  key: string\n  /** Description shown in the judge's user prompt. */\n  description: string\n}\n\n/** @experimental Pluggable dimensional scorer. `score` is the contract:\n *  given an artifact + scenario, return a `JudgeScore`. This is deliberately a\n *  function, not a fixed LLM-prompt shape — real consumers judge with\n *  ensembles, deterministic checks, or a single LLM call, and the substrate\n *  must not constrain that. The `llmJudge()` helper builds a `score` that does\n *  one LLM call for the common case. `appliesTo` lets a judge run only on\n *  scenarios that match (e.g. a legal-citation judge only on legal scenarios). */\nexport interface JudgeConfig<TArtifact, TScenario extends Scenario = Scenario> {\n  name: string\n  dimensions: JudgeDimension[]\n  /** Score one artifact. Throw on failure — a thrown judge is recorded as a\n   *  failed cell, never silently folded into a zero. */\n  score(input: {\n    artifact: TArtifact\n    scenario: TScenario\n    signal: AbortSignal\n  }): JudgeScore | Promise<JudgeScore>\n  appliesTo?: (scenario: TScenario) => boolean\n}\n\nexport interface JudgeScore {\n  dimensions: Record<string, number>\n  composite: number\n  notes: string\n}\n\n// ── Optimization (population + generations + mutator) ─────────────────\n\n/** @experimental A tier-4 code surface — a candidate change to the agent's\n *  IMPLEMENTATION, not its prompt. Produced by autoresearch (reads codebase +\n *  trace findings → opens a worktree). Measured by checking out `worktreeRef`\n *  and running the worker against the changed code. See the improvement-tier\n *  table in `docs/design/loop-taxonomy.md`. */\nexport interface CodeSurface {\n  kind: 'code'\n  /** Worktree path or git ref holding the candidate code change. The\n   *  consumer's `dispatchWithSurface` checks this out before running. */\n  worktreeRef: string\n  /** Base ref the change is measured against. Default: the repo's main. */\n  baseRef?: string\n  /** Human summary of what changed — rendered into the auto-PR body. */\n  summary?: string\n}\n\n/** @experimental The mutable surface a driver proposes. Tiers (see\n *  `docs/design/loop-taxonomy.md`):\n *   - `string`      — tiers 1-2: system-prompt addendum / serialized tool\n *                     config. Cheap, reversible, text-diffable.\n *   - `CodeSurface` — tier 4: an implementation change behind a worktree ref.\n *  Tier 3 (knowledge) is owned by agent-knowledge and rides its own adapter,\n *  not this type. */\nexport type MutableSurface = string | CodeSurface\n\n/** @experimental Stateless surface mutation — given findings + current\n *  surface, return N candidate surfaces. Pure transform, no generation\n *  awareness. Reflective-mutation, `runMultiShotOptimization`, `AxGEPA`\n *  conform. Wrapped by `evolutionaryDriver` to become an `ImprovementDriver`. */\nexport interface Mutator<TFindings = unknown> {\n  kind: string\n  mutate(args: {\n    findings: TFindings[]\n    currentSurface: MutableSurface\n    populationSize: number\n    signal: AbortSignal\n  }): Promise<MutableSurface[]>\n}\n\n/** @experimental Everything a driver's `propose()` may read to plan the next\n *  batch of candidates. The first six fields are always present; the rest are\n *  optional context the loop supplies when available, so cheap drivers\n *  (`evolutionaryDriver`) can ignore them while a code-tier agentic generator\n *  consumes the research report + dataset to drive a coding harness.\n *  See `docs/design/self-improvement-engine.md`. */\nexport interface ProposeContext<TFindings = unknown> {\n  currentSurface: MutableSurface\n  history: GenerationRecord[]\n  findings: TFindings[]\n  /** BREADTH: how many candidate surfaces to return this generation. */\n  populationSize: number\n  generation: number\n  signal: AbortSignal\n  /** The Phase-2 research report (analyst findings + diff), produced AFTER the\n   *  trace analysts run. Opaque to the substrate — the driver that consumes it\n   *  types it. See the phase diagram in self-improvement-engine.md. */\n  report?: unknown\n  /** Handle to all captured data — the driver samples traces / artifacts /\n   *  rewards here to ground its proposals. */\n  dataset?: LabeledScenarioStore\n  /** DEPTH: max iterations the agentic generator may take per candidate.\n   *  1 = single-shot; >1 = it may iterate on its own change before handing it\n   *  back to be measured. */\n  maxImprovementShots?: number\n}\n\n/** @experimental A surface-improvement strategy — the DRIVER of the\n *  improvement loop. Given the current best surface, the history of what's\n *  been tried + scored, and any external findings, propose the next batch of\n *  candidate surfaces to measure. Optionally decide to stop early.\n *\n *  The evolutionary mutator (`evolutionaryDriver`, here) and agent-runtime's\n *  `improvementDriver` (with reflective / agentic generators) both conform —\n *  drivers of the SAME loop, not separate loops. The loop body\n *  (`runOptimization`) and the gated promotion shell (`runImprovementLoop`)\n *  are driver-agnostic. */\nexport interface ImprovementDriver<TFindings = unknown> {\n  kind: string\n  /** Plan: propose N candidate surfaces for the next generation. */\n  propose(ctx: ProposeContext<TFindings>): Promise<MutableSurface[]>\n  /** Decide: stop early when the driver judges the search converged or\n   *  exhausted. Default (omitted) runs all `maxGenerations`. */\n  decide?(args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n}\n\nexport interface OptimizerConfig {\n  driver: ImprovementDriver\n  populationSize: number\n  maxGenerations: number\n  surfaceExtractor: (profile: unknown) => MutableSurface\n}\n\n// ── Gates ─────────────────────────────────────────────────────────────\n\n/** @experimental Five-valued verdict taxonomy (MOSS-paper alignment). */\nexport type GateDecision = 'ship' | 'hold' | 'need_more_work' | 'model_ceiling' | 'arch_ceiling'\n\nexport interface GateContext<TArtifact, TScenario extends Scenario> {\n  candidateArtifacts: Map<string, TArtifact>\n  baselineArtifacts?: Map<string, TArtifact>\n  /** Candidate (winner) judge scores, keyed by cellId. */\n  judgeScores: Map<string, Record<string, JudgeScore>>\n  /** Baseline judge scores, keyed by cellId. SEPARATE from `judgeScores` —\n   *  baseline + candidate share cellIds (same scenarios), so a single map\n   *  cannot represent both. A gate computing a holdout delta MUST read\n   *  candidate from `judgeScores` and baseline from here. */\n  baselineJudgeScores?: Map<string, Record<string, JudgeScore>>\n  scenarios: TScenario[]\n  cost: { candidate: number; baseline: number }\n  signal: AbortSignal\n}\n\nexport interface GateResult {\n  decision: GateDecision\n  reasons: string[]\n  contributingGates: Array<{ name: string; passed: boolean; detail: unknown }>\n  delta?: number\n}\n\n/** @experimental Composable promotion gate. */\nexport interface Gate<TArtifact = unknown, TScenario extends Scenario = Scenario> {\n  name: string\n  decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult>\n}\n\n// ── Tracing / artifacts / cost ────────────────────────────────────────\n\n/** @experimental Scoped trace writer handed to each dispatch — every span\n *  auto-tagged with the cellId so traces filter cleanly. */\nexport interface CampaignTraceWriter {\n  span(name: string, attributes?: Record<string, unknown>): TraceSpan\n  flush(): Promise<void>\n}\n\nexport interface TraceSpan {\n  end(attributes?: Record<string, unknown>): void\n  setAttribute(key: string, value: unknown): void\n}\n\n/** @experimental Scoped artifact writer — `write(path, content)` lands under\n *  `<runDir>/<cellId>/<path>`. */\nexport interface CampaignArtifactWriter {\n  write(path: string, content: string | Uint8Array): Promise<string>\n  writeJson(path: string, value: unknown): Promise<string>\n}\n\n/** @experimental Cell-scoped cost meter. Substrate auto-tracks LLM costs\n *  via the cost-ledger backend hooks; consumers can record additional\n *  spend (sandbox time, tool costs) via `observe`. */\nexport interface CampaignCostMeter {\n  observe(amountUsd: number, source: string): void\n  current(): number\n}\n\n// ── LabeledScenarioStore ──────────────────────────────────────────────\n\n/** @experimental Source tag — required on every store write. Used by the\n *  default training-source filter (production-trace samples NOT used as\n *  training scenarios unless explicitly opted in). */\nexport type LabeledScenarioSource =\n  | 'production-trace'\n  | 'eval-run'\n  | 'manual'\n  | 'red-team'\n  | 'synthetic'\n\nexport type RedactionStatus = 'raw' | 'redacted-pii' | 'redacted-secrets' | 'fully-redacted'\n\n/** How much a label can be trusted to evaluate against — the gold-admission\n *  gate. Strictly ordered: a record qualifies for a `minTrust` filter when its\n *  trust rank is >= the requested rank.\n *\n *   - `unverified`      — label is a heuristic (e.g. raw outcome success/fail).\n *                          Fine as corpus; MUST NOT enter a gold set that lift\n *                          numbers are computed against.\n *   - `verified-signal` — an external signal confirmed the outcome (PR merged,\n *                          tests green, user did not retry, downstream check).\n *   - `human-rated`     — a human explicitly rated or corrected the artifact.\n *\n *  Absent on a write ⇒ treated as `unverified` (fail-closed: a writer must\n *  explicitly assert trust to make a record gold-eligible — it never happens\n *  by accident). */\nexport type LabelTrust = 'unverified' | 'verified-signal' | 'human-rated'\n\nconst LABEL_TRUST_RANK: Record<LabelTrust, number> = {\n  unverified: 0,\n  'verified-signal': 1,\n  'human-rated': 2,\n}\n\n/** Ordinal rank for a label-trust tier; absent ⇒ `unverified` (rank 0). */\nexport function labelTrustRank(trust: LabelTrust | undefined): number {\n  return LABEL_TRUST_RANK[trust ?? 'unverified']\n}\n\n/** @experimental Required-provenance write. The store rejects writes that\n *  lack provenance — a default-on flywheel without provenance is the\n *  data-poisoning vector flagged in the alignment review. */\nexport interface LabeledScenarioWrite<TScenario extends Scenario = Scenario, TArtifact = unknown> {\n  scenario: TScenario\n  artifact: TArtifact\n  judgeScores: Record<string, JudgeScore>\n  source: LabeledScenarioSource\n  sourceVersionHash: string\n  capturedAt: string\n  redactionStatus: RedactionStatus\n  /** Gold-admission trust tier. Absent ⇒ `unverified` (fail-closed): the\n   *  record is corpus, never gold. A writer must explicitly assert\n   *  `verified-signal` or `human-rated` to make it eligible for a gold\n   *  sample. See {@link LabelTrust}. */\n  labelTrust?: LabelTrust\n  /** Optional per-source rate-limit bucket key (e.g., the tenant id). */\n  rateLimitBucket?: string\n}\n\nexport interface LabeledScenarioRecord<TScenario extends Scenario = Scenario, TArtifact = unknown>\n  extends LabeledScenarioWrite<TScenario, TArtifact> {\n  /** Stable hash of (scenario.id, source, capturedAt, sourceVersionHash). */\n  recordHash: string\n  /** Substrate-assigned split — train if captured before the campaign's\n   *  `temporalCutoff`, test if after. Explicit override allowed via filter. */\n  split: 'train' | 'test'\n}\n\nexport interface LabeledScenarioSampleArgs {\n  count: number\n  /** REQUIRED — substrate refuses to sample without an explicit split. */\n  split: 'train' | 'test'\n  /** REQUIRED — only records captured before this timestamp are returned.\n   *  Enforces temporal split discipline (test scenarios captured AFTER train\n   *  cannot enter the training pool). */\n  capturedBefore: string\n  filter?: {\n    kind?: string\n    source?: LabeledScenarioSource | LabeledScenarioSource[]\n    minComposite?: number\n    maxComposite?: number\n    /** Gold gate: only records whose trust rank is >= this tier are\n     *  returned. `sample({ split: 'test', minTrust: 'verified-signal' })` is\n     *  the canonical \"give me the gold set\" call. Absent ⇒ no trust gate\n     *  (corpus-level read). */\n    minTrust?: LabelTrust\n  }\n}\n\nexport interface LabeledScenarioStore {\n  observe(write: LabeledScenarioWrite): Promise<void>\n  sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]>\n  size(): Promise<{\n    train: number\n    test: number\n    bySource: Record<string, number>\n    /** Count by trust tier — tells the flywheel how much gold it has\n     *  accumulated vs. raw corpus. */\n    byTrust: Record<LabelTrust, number>\n  }>\n}\n\n// ── The CampaignResult schema (the downstream-tools contract) ─────────\n\nexport interface CampaignCellResult<TArtifact> {\n  cellId: string\n  scenarioId: string\n  rep: number\n  generation?: number\n  artifact: TArtifact\n  judgeScores: Record<string, JudgeScore>\n  costUsd: number\n  durationMs: number\n  seed: number\n  cached: boolean\n  error?: string\n}\n\nexport interface JudgeAggregate {\n  mean: number\n  stdev: number\n  ci95: [number, number]\n  n: number\n}\n\nexport interface ScenarioAggregate {\n  meanComposite: number\n  ci95: [number, number]\n  n: number\n}\n\nexport interface GenerationRecord {\n  generationIndex: number\n  candidates: GenerationCandidate[]\n  promoted: string[]\n}\n\n/** One scored candidate surface in a generation. `dimensions` + `scenarios`\n *  let a reflective `ImprovementDriver` ground its next proposal on WHICH\n *  dimensions the candidate is weakest on and WHICH scenarios it best/worst\n *  handled — the evidence a blind `Mutator` cannot see. */\nexport interface GenerationCandidate {\n  surfaceHash: string\n  composite: number\n  ci95: [number, number]\n  /** Mean score per judge dimension across all cells (scenarios × reps ×\n   *  judges that reported the dimension). */\n  dimensions: Record<string, number>\n  /** Per-scenario composite (mean over reps + judges). */\n  scenarios: Array<{ scenarioId: string; composite: number }>\n}\n\nexport interface CampaignAggregates {\n  byJudge: Record<string, JudgeAggregate>\n  byScenario: Record<string, ScenarioAggregate>\n  totalCostUsd: number\n  cellsExecuted: number\n  cellsSkipped: number\n  cellsCached: number\n  cellsFailed: number\n}\n\nexport interface CampaignResult<TArtifact = unknown, TScenario extends Scenario = Scenario> {\n  /** sha256(scenarios, judges, dispatch source ref, optimizer config, seed). Stable identity for reruns. */\n  manifestHash: string\n  seed: number\n  startedAt: string\n  endedAt: string\n  durationMs: number\n  cells: Array<CampaignCellResult<TArtifact>>\n  aggregates: CampaignAggregates\n  optimization?: {\n    generations: GenerationRecord[]\n    winnerSurfaceHash?: string\n  }\n  gate?: GateResult\n  prUrl?: string\n  runDir: string\n  artifactsByPath: Record<string, string>\n  /** Substrate strips the input scenarios to id+kind for the result manifest;\n   *  consumers needing full payload look it up via the original input. The\n   *  type parameter `TScenario` is propagated for downstream consumers that\n   *  want narrowed types when extending `CampaignResult`. */\n  scenarios: Array<Pick<TScenario, 'id' | 'kind'>>\n}\n","/**\n * @experimental\n *\n * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like\n * (multiple commits allowed). A code-tier driver's `propose()` creates a\n * worktree, an agent commits the change into it, and `finalize()` returns a\n * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker\n * against the changed code. On promotion the worktree becomes the PR branch.\n *\n * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))\n * adapter can slot in without touching driver code. Only the git adapter\n * ships today. See `docs/design/self-improvement-engine.md`.\n */\n\nimport { execFileSync } from 'node:child_process'\nimport { existsSync } from 'node:fs'\nimport { basename, isAbsolute, join } from 'node:path'\nimport type { CodeSurface } from '../types'\n\nexport interface Worktree {\n  /** Absolute path to the checked-out worktree directory. */\n  path: string\n  /** The branch the worktree is on (becomes the PR branch on promotion). */\n  branch: string\n  /** The ref the worktree was forked from. */\n  baseRef: string\n}\n\nexport interface WorktreeAdapter {\n  /** Create an isolated worktree on a fresh branch off `baseRef`. */\n  create(opts: { baseRef: string; label: string }): Promise<Worktree>\n  /** Commit any pending changes in the worktree, then return a CodeSurface\n   *  pointing at it. The agent has already written its change into\n   *  `worktree.path` by the time this is called. */\n  finalize(worktree: Worktree, summary: string): Promise<CodeSurface>\n  /** Remove the worktree (and its branch) — called for losing candidates. */\n  discard(worktree: Worktree): Promise<void>\n}\n\nexport class WorktreeAdapterError extends Error {\n  constructor(\n    message: string,\n    readonly cause?: unknown,\n  ) {\n    super(message)\n    this.name = 'WorktreeAdapterError'\n  }\n}\n\nexport interface GitWorktreeAdapterOptions {\n  /** Repo root the worktrees fork from. */\n  repoRoot: string\n  /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */\n  worktreeDir?: string\n  /** Branch-name prefix. Default: `improve`. */\n  branchPrefix?: string\n  /** Test seam — defaults to a real `git` runner. */\n  git?: (args: string[], cwd: string) => string\n}\n\nfunction defaultGit(args: string[], cwd: string): string {\n  try {\n    return execFileSync('git', args, { cwd, encoding: 'utf8' }).trim()\n  } catch (err) {\n    const stderr =\n      err && typeof err === 'object' && 'stderr' in err\n        ? String((err as { stderr: unknown }).stderr)\n        : ''\n    throw new WorktreeAdapterError(`git ${args.join(' ')} failed: ${stderr || String(err)}`, err)\n  }\n}\n\n/** Slugify a label into a branch-safe segment. */\nfunction slug(label: string): string {\n  return (\n    label\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 48) || 'candidate'\n  )\n}\n\nexport function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter {\n  const git = opts.git ?? defaultGit\n  const worktreeDir = opts.worktreeDir ?? join(opts.repoRoot, '.worktrees')\n  const branchPrefix = opts.branchPrefix ?? 'improve'\n\n  return {\n    async create({ baseRef, label }) {\n      const id = `${slug(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`\n      const branch = `${branchPrefix}/${id}`\n      const path = join(worktreeDir, id)\n      git(['worktree', 'add', '-b', branch, path, baseRef], opts.repoRoot)\n      return { path, branch, baseRef }\n    },\n\n    async finalize(worktree, summary) {\n      // Stage + commit any pending changes the agent left in the worktree.\n      // A no-op commit is refused by git, so only commit when the tree is dirty.\n      const status = git(['status', '--porcelain'], worktree.path)\n      if (status.length > 0) {\n        git(['add', '-A'], worktree.path)\n        git(['commit', '-m', summary], worktree.path)\n      }\n      return {\n        kind: 'code',\n        worktreeRef: worktree.path,\n        baseRef: worktree.baseRef,\n        summary,\n      }\n    },\n\n    async discard(worktree) {\n      // Remove the worktree, then delete its branch. Force-remove because the\n      // worktree may hold uncommitted experiment state we're discarding.\n      git(['worktree', 'remove', '--force', worktree.path], opts.repoRoot)\n      git(['branch', '-D', worktree.branch], opts.repoRoot)\n    },\n  }\n}\n\n/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can\n *  run the worker in. A path ref is returned as-is; anything else is treated\n *  as a ref under the adapter's worktree dir. */\nexport function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string {\n  if (isAbsolute(surface.worktreeRef) && existsSync(surface.worktreeRef)) return surface.worktreeRef\n  if (worktreeDir) return join(worktreeDir, basename(surface.worktreeRef))\n  return surface.worktreeRef\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;;;AC2QrB,IAAM,mBAA+C;AAAA,EACnD,YAAY;AAAA,EACZ,mBAAmB;AAAA,EACnB,eAAe;AACjB;AAGO,SAAS,eAAe,OAAuC;AACpE,SAAO,iBAAiB,SAAS,YAAY;AAC/C;;;AD/PO,IAAM,4BAAN,cAAwC,MAAM;AAAA,EACnD,YACkB,MAChB,SACA;AACA,UAAM,OAAO;AAHG;AAIhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAMpB;AAQO,IAAM,yBAAN,MAA6D;AAAA,EAIlE,YAA6B,SAAwC;AAAxC;AAC3B,QAAI,CAAC,WAAW,QAAQ,IAAI,EAAG,WAAU,QAAQ,MAAM,EAAE,WAAW,KAAK,CAAC;AAC1E,SAAK,MAAM,QAAQ,OAAO,KAAK;AAAA,EACjC;AAAA,EAH6B;AAAA,EAHZ;AAAA,EACA,aAAa,oBAAI,IAA4B;AAAA,EAO9D,MAAM,QAAQ,OAA4C;AACxD,SAAK,iBAAiB,KAAK;AAC3B,SAAK,gBAAgB,KAAK;AAC1B,UAAM,SAAS,KAAK,SAAS,KAAK;AAClC,UAAM,OAAO,KAAK,cAAc,MAAM,MAAM;AAC5C,UAAM,OAAO,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAItC,eAAW,MAAM,IAAI;AAAA,EACvB;AAAA,EAEA,MAAM,OAAO,MAAmE;AAC9E,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAEA,UAAM,MAA+B,CAAC;AACtC,eAAW,UAAU,aAAa;AAGhC,UAAI,KAAK,UAAU,WAAW,WAAW,oBAAoB;AAC3D,cAAM,WAAW,qBAAqB,KAAK,QAAQ,QAAQ,kBAAkB;AAC7E,YAAI,CAAC,SAAU;AAAA,MACjB;AACA,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,EAAG;AACvB,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,iBAAW,QAAQ,OAAO;AACxB,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,QAAQ;AACN;AAAA,QACF;AACA,YAAI,CAAC,cAAc,QAAQ,MAAM,MAAM,EAAG;AAC1C,YAAI,KAAK,MAAM;AAAA,MACjB;AAAA,IACF;AAGA,QAAI,KAAK,CAAC,GAAG,MAAM;AACjB,UAAI,EAAE,eAAe,EAAE,WAAY,QAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AACjF,aAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,IAChD,CAAC;AAED,WAAO,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,EAChC;AAAA,EAEA,MAAM,OAKH;AACD,UAAM,WAAmC,CAAC;AAC1C,UAAM,UAAsC;AAAA,MAC1C,YAAY;AAAA,MACZ,mBAAmB;AAAA,MACnB,eAAe;AAAA,IACjB;AACA,QAAI,QAAQ;AACZ,eAAW,UAAU,aAAa;AAChC,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,GAAG;AACrB,iBAAS,MAAM,IAAI;AACnB;AAAA,MACF;AACA,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,eAAS,MAAM,IAAI,MAAM;AACzB,eAAS,MAAM;AACf,iBAAW,QAAQ,OAAO;AACxB,YAAI,QAAoB;AACxB,YAAI;AACF,kBAAS,KAAK,MAAM,IAAI,EAA4B,cAAc;AAAA,QACpE,QAAQ;AAAA,QAER;AACA,gBAAQ,KAAK,KAAK;AAAA,MACpB;AAAA,IACF;AAIA,WAAO,EAAE,OAAO,OAAO,MAAM,OAAO,UAAU,QAAQ;AAAA,EACxD;AAAA,EAEQ,iBAAiB,OAAmC;AAC1D,QAAI,CAAC,MAAM,QAAQ;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,qBAAqB,MAAM,kBAAkB,WAAW,GAAG;AACpE,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,YAAY;AACrB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,iBAAiB;AAC1B,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,YAAY,SAAS,MAAM,MAAM,GAAG;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,YAAY,KAAK,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,OAAmC;AACzD,UAAM,MAAM,KAAK,QAAQ;AACzB,QAAI,CAAC,OAAO,CAAC,MAAM,gBAAiB;AACpC,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,WAAW;AACjB,QAAI,QAAQ,KAAK,WAAW,IAAI,MAAM,eAAe;AACrD,QAAI,CAAC,SAAS,MAAM,MAAM,iBAAiB,UAAU;AACnD,cAAQ,EAAE,QAAQ,MAAM,iBAAiB,eAAe,KAAK,OAAO,EAAE;AACtE,WAAK,WAAW,IAAI,MAAM,iBAAiB,KAAK;AAAA,IAClD;AACA,QAAI,MAAM,SAAS,KAAK;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA,gCAAgC,MAAM,eAAe,aAAa,GAAG;AAAA,MACvE;AAAA,IACF;AACA,UAAM,SAAS;AAAA,EACjB;AAAA,EAEQ,SAAS,OAAoD;AACnE,UAAM,aAAa;AAAA,MACjB,KAAK,UAAU;AAAA,QACb,IAAI,MAAM,SAAS;AAAA,QACnB,KAAK,MAAM;AAAA,QACX,IAAI,MAAM;AAAA,QACV,KAAK,MAAM;AAAA,MACb,CAAC;AAAA,IACH;AAGA,WAAO;AAAA,MACL,GAAG;AAAA,MACH;AAAA,MACA,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEQ,cAAc,QAAwB;AAC5C,WAAO,KAAK,KAAK,QAAQ,MAAM,GAAG,MAAM,QAAQ;AAAA,EAClD;AACF;AAEA,IAAM,cAAgD;AAAA,EACpD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,qBACP,QACA,QACS;AACT,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,EAAG,QAAO,OAAO,SAAS,MAAM;AACxD,SAAO,WAAW;AACpB;AAEA,SAAS,cACP,QACA,MACA,QACS;AAET,MAAI,KAAK,UAAU,WAAW,OAAO,cAAc,KAAK,eAAgB,QAAO;AAC/E,MAAI,KAAK,UAAU,UAAU,OAAO,aAAa,KAAK,eAAgB,QAAO;AAE7E,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,EAAG,QAAO;AACf,MAAI,EAAE,QAAQ,OAAO,SAAS,SAAS,EAAE,KAAM,QAAO;AACtD,MAAI,EAAE,QAAQ;AACZ,UAAM,UAAU,MAAM,QAAQ,EAAE,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM;AAC9D,QAAI,CAAC,QAAQ,SAAS,MAAe,EAAG,QAAO;AAAA,EACjD;AACA,MAAI,EAAE,iBAAiB,UAAa,EAAE,iBAAiB,QAAW;AAChE,UAAM,aAAa,OAAO,OAAO,OAAO,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3E,UAAM,MAAM,WAAW,WAAW,IAAI,IAAI,KAAK,IAAI,GAAG,UAAU;AAChE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AACjE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AAAA,EACnE;AACA,MAAI,EAAE,aAAa,UAAa,eAAe,OAAO,UAAU,IAAI,eAAe,EAAE,QAAQ,GAAG;AAC9F,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEA,SAAS,OAAO,OAAuB;AACrC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,MAAM,GAAG,EAAE;AACrE;AAEA,SAAS,WAAW,MAAc,MAAoB;AACpD,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,WAAW,aAAa,MAAM,MAAM;AAC1C,kBAAc,MAAM,WAAW,IAAI;AAAA,EACrC,OAAO;AACL,kBAAc,MAAM,IAAI;AAAA,EAC1B;AACF;;;AElSA,SAAS,oBAAoB;AAC7B,SAAS,cAAAA,mBAAkB;AAC3B,SAAS,UAAU,YAAY,QAAAC,aAAY;AAuBpC,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC9C,YACE,SACS,OACT;AACA,UAAM,OAAO;AAFJ;AAGT,SAAK,OAAO;AAAA,EACd;AAAA,EAJW;AAKb;AAaA,SAAS,WAAW,MAAgB,KAAqB;AACvD,MAAI;AACF,WAAO,aAAa,OAAO,MAAM,EAAE,KAAK,UAAU,OAAO,CAAC,EAAE,KAAK;AAAA,EACnE,SAAS,KAAK;AACZ,UAAM,SACJ,OAAO,OAAO,QAAQ,YAAY,YAAY,MAC1C,OAAQ,IAA4B,MAAM,IAC1C;AACN,UAAM,IAAI,qBAAqB,OAAO,KAAK,KAAK,GAAG,CAAC,YAAY,UAAU,OAAO,GAAG,CAAC,IAAI,GAAG;AAAA,EAC9F;AACF;AAGA,SAAS,KAAK,OAAuB;AACnC,SACE,MACG,YAAY,EACZ,QAAQ,eAAe,GAAG,EAC1B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEO,SAAS,mBAAmB,MAAkD;AACnF,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,cAAc,KAAK,eAAeA,MAAK,KAAK,UAAU,YAAY;AACxE,QAAM,eAAe,KAAK,gBAAgB;AAE1C,SAAO;AAAA,IACL,MAAM,OAAO,EAAE,SAAS,MAAM,GAAG;AAC/B,YAAM,KAAK,GAAG,KAAK,KAAK,CAAC,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AAC9F,YAAM,SAAS,GAAG,YAAY,IAAI,EAAE;AACpC,YAAM,OAAOA,MAAK,aAAa,EAAE;AACjC,UAAI,CAAC,YAAY,OAAO,MAAM,QAAQ,MAAM,OAAO,GAAG,KAAK,QAAQ;AACnE,aAAO,EAAE,MAAM,QAAQ,QAAQ;AAAA,IACjC;AAAA,IAEA,MAAM,SAAS,UAAU,SAAS;AAGhC,YAAM,SAAS,IAAI,CAAC,UAAU,aAAa,GAAG,SAAS,IAAI;AAC3D,UAAI,OAAO,SAAS,GAAG;AACrB,YAAI,CAAC,OAAO,IAAI,GAAG,SAAS,IAAI;AAChC,YAAI,CAAC,UAAU,MAAM,OAAO,GAAG,SAAS,IAAI;AAAA,MAC9C;AACA,aAAO;AAAA,QACL,MAAM;AAAA,QACN,aAAa,SAAS;AAAA,QACtB,SAAS,SAAS;AAAA,QAClB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,QAAQ,UAAU;AAGtB,UAAI,CAAC,YAAY,UAAU,WAAW,SAAS,IAAI,GAAG,KAAK,QAAQ;AACnE,UAAI,CAAC,UAAU,MAAM,SAAS,MAAM,GAAG,KAAK,QAAQ;AAAA,IACtD;AAAA,EACF;AACF;AAKO,SAAS,oBAAoB,SAAsB,aAA8B;AACtF,MAAI,WAAW,QAAQ,WAAW,KAAKD,YAAW,QAAQ,WAAW,EAAG,QAAO,QAAQ;AACvF,MAAI,YAAa,QAAOC,MAAK,aAAa,SAAS,QAAQ,WAAW,CAAC;AACvE,SAAO,QAAQ;AACjB;","names":["existsSync","join"]}
+{"version":3,"sources":["../../src/campaign/labeled-store/fs-adapter.ts","../../src/campaign/worktree/index.ts"],"sourcesContent":["/**\n * @experimental\n *\n * Filesystem `LabeledScenarioStore` adapter. The default capture sink for\n * traces + eval artifacts. Production deployments typically swap for a\n * Turso/SQLite adapter (same interface).\n *\n * Records land as one JSONL file per source under `<root>/<source>.jsonl`.\n * Each line is a `LabeledScenarioRecord`. Append-only — no in-place edits.\n *\n * Safety properties enforced at write-time:\n *\n *   - **Provenance required**: writes without `source`, `sourceVersionHash`,\n *     `capturedAt`, `redactionStatus` are rejected. Closes the alignment\n *     reviewer's data-poisoning gap.\n *   - **Per-source rate limits**: optional `rateLimitBucket` + `maxWritesPerMinute`\n *     stops a single tenant/source from flooding the store.\n *\n * Safety properties enforced at sample-time:\n *\n *   - **Required split + capturedBefore**: substrate refuses to sample without\n *     an explicit `split` ('train' | 'test') AND a temporal cutoff. Eliminates\n *     accidental train/test contamination.\n *   - **Default training-source filter**: when the store is sampled with\n *     `split: 'train'`, production-trace records are EXCLUDED unless the\n *     caller passes `filter.source: 'production-trace'` explicitly. Closes\n *     the contamination-by-default gap flagged by the senior eval engineer.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type {\n  LabeledScenarioRecord,\n  LabeledScenarioSampleArgs,\n  LabeledScenarioSource,\n  LabeledScenarioStore,\n  LabeledScenarioWrite,\n  LabelTrust,\n} from '../types'\nimport { labelTrustRank } from '../types'\n\nexport interface FsLabeledScenarioStoreOptions {\n  /** Root directory for JSONL files. Created if missing. */\n  root: string\n  /** Per-source rate limit. When set, writes exceeding the cap are rejected\n   *  with a typed error. Default: no limit. */\n  maxWritesPerMinutePerBucket?: number\n  /** Test seam — override `Date.now()` for deterministic tests. */\n  now?: () => number\n}\n\nexport class LabeledScenarioStoreError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n  ) {\n    super(message)\n    this.name = 'LabeledScenarioStoreError'\n  }\n}\n\ninterface RateLimitState {\n  bucket: string\n  windowStartMs: number\n  count: number\n}\n\nexport class FsLabeledScenarioStore implements LabeledScenarioStore {\n  private readonly now: () => number\n  private readonly rateLimits = new Map<string, RateLimitState>()\n\n  constructor(private readonly options: FsLabeledScenarioStoreOptions) {\n    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true })\n    this.now = options.now ?? Date.now\n  }\n\n  async observe(write: LabeledScenarioWrite): Promise<void> {\n    this.assertProvenance(write)\n    this.assertRateLimit(write)\n    const record = this.toRecord(write)\n    const path = this.pathForSource(write.source)\n    const line = `${JSON.stringify(record)}\\n`\n    // Append atomically. For high-throughput a writev-friendly buffered\n    // implementation lands in the Turso adapter; FS adapter is for tests +\n    // local dev + small workloads.\n    appendLine(path, line)\n  }\n\n  async sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]> {\n    if (!args.split) {\n      throw new LabeledScenarioStoreError(\n        'split_required',\n        'sample() requires an explicit `split` (train | test) — substrate refuses ambiguous reads',\n      )\n    }\n    if (!args.capturedBefore) {\n      throw new LabeledScenarioStoreError(\n        'capturedBefore_required',\n        'sample() requires an explicit `capturedBefore` timestamp for temporal-split discipline',\n      )\n    }\n\n    const all: LabeledScenarioRecord[] = []\n    for (const source of ALL_SOURCES) {\n      // Default training-source filter: when sampling train, EXCLUDE\n      // production-trace records unless the caller asks for them.\n      if (args.split === 'train' && source === 'production-trace') {\n        const explicit = sourceFilterContains(args.filter?.source, 'production-trace')\n        if (!explicit) continue\n      }\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) continue\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      for (const line of lines) {\n        let record: LabeledScenarioRecord\n        try {\n          record = JSON.parse(line) as LabeledScenarioRecord\n        } catch {\n          continue\n        }\n        if (!matchesFilter(record, args, source)) continue\n        all.push(record)\n      }\n    }\n\n    // Deterministic order: by capturedAt ascending, then recordHash.\n    all.sort((a, b) => {\n      if (a.capturedAt !== b.capturedAt) return a.capturedAt.localeCompare(b.capturedAt)\n      return a.recordHash.localeCompare(b.recordHash)\n    })\n\n    return all.slice(0, args.count)\n  }\n\n  async size(): Promise<{\n    train: number\n    test: number\n    bySource: Record<string, number>\n    byTrust: Record<LabelTrust, number>\n  }> {\n    const bySource: Record<string, number> = {}\n    const byTrust: Record<LabelTrust, number> = {\n      unverified: 0,\n      'verified-signal': 0,\n      'human-rated': 0,\n    }\n    let total = 0\n    for (const source of ALL_SOURCES) {\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) {\n        bySource[source] = 0\n        continue\n      }\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      bySource[source] = lines.length\n      total += lines.length\n      for (const line of lines) {\n        let trust: LabelTrust = 'unverified'\n        try {\n          trust = (JSON.parse(line) as LabeledScenarioRecord).labelTrust ?? 'unverified'\n        } catch {\n          // A malformed line counts as unverified — never silently gold.\n        }\n        byTrust[trust] += 1\n      }\n    }\n    // FS adapter doesn't track split assignments per-record (split is\n    // computed at sample-time based on `capturedBefore`). For size(), we\n    // report `train`+`test` as the same total — split is a sampling concept.\n    return { train: total, test: total, bySource, byTrust }\n  }\n\n  private assertProvenance(write: LabeledScenarioWrite): void {\n    if (!write.source) {\n      throw new LabeledScenarioStoreError(\n        'missing_source',\n        'LabeledScenarioWrite requires `source`',\n      )\n    }\n    if (!write.sourceVersionHash || write.sourceVersionHash.length === 0) {\n      throw new LabeledScenarioStoreError(\n        'missing_source_version',\n        'LabeledScenarioWrite requires `sourceVersionHash` (git sha or substrate version)',\n      )\n    }\n    if (!write.capturedAt) {\n      throw new LabeledScenarioStoreError(\n        'missing_captured_at',\n        'LabeledScenarioWrite requires `capturedAt` ISO timestamp',\n      )\n    }\n    if (!write.redactionStatus) {\n      throw new LabeledScenarioStoreError(\n        'missing_redaction_status',\n        'LabeledScenarioWrite requires explicit `redactionStatus` — raw / redacted-pii / redacted-secrets / fully-redacted',\n      )\n    }\n    if (!ALL_SOURCES.includes(write.source)) {\n      throw new LabeledScenarioStoreError(\n        'unknown_source',\n        `LabeledScenarioWrite.source must be one of: ${ALL_SOURCES.join(', ')}`,\n      )\n    }\n  }\n\n  private assertRateLimit(write: LabeledScenarioWrite): void {\n    const cap = this.options.maxWritesPerMinutePerBucket\n    if (!cap || !write.rateLimitBucket) return\n    const now = this.now()\n    const windowMs = 60_000\n    let state = this.rateLimits.get(write.rateLimitBucket)\n    if (!state || now - state.windowStartMs >= windowMs) {\n      state = { bucket: write.rateLimitBucket, windowStartMs: now, count: 0 }\n      this.rateLimits.set(write.rateLimitBucket, state)\n    }\n    if (state.count >= cap) {\n      throw new LabeledScenarioStoreError(\n        'rate_limit_exceeded',\n        `LabeledScenarioStore: bucket ${write.rateLimitBucket} exceeded ${cap} writes/min`,\n      )\n    }\n    state.count += 1\n  }\n\n  private toRecord(write: LabeledScenarioWrite): LabeledScenarioRecord {\n    const recordHash = sha256(\n      JSON.stringify({\n        id: write.scenario.id,\n        src: write.source,\n        at: write.capturedAt,\n        ver: write.sourceVersionHash,\n      }),\n    )\n    // FS adapter assigns split at sample-time, but we cache a hint here\n    // based on capturedAt vs the world's \"now\" — sampler overrides this.\n    return {\n      ...write,\n      recordHash,\n      split: 'train',\n    }\n  }\n\n  private pathForSource(source: string): string {\n    return join(this.options.root, `${source}.jsonl`)\n  }\n}\n\nconst ALL_SOURCES: LabeledScenarioWrite['source'][] = [\n  'production-trace',\n  'eval-run',\n  'manual',\n  'red-team',\n  'synthetic',\n]\n\nfunction sourceFilterContains(\n  filter: LabeledScenarioSource | LabeledScenarioSource[] | undefined,\n  needle: LabeledScenarioSource,\n): boolean {\n  if (!filter) return false\n  if (Array.isArray(filter)) return filter.includes(needle)\n  return filter === needle\n}\n\nfunction matchesFilter(\n  record: LabeledScenarioRecord,\n  args: LabeledScenarioSampleArgs,\n  source: string,\n): boolean {\n  // Temporal cutoff — train must be capturedAt < capturedBefore.\n  if (args.split === 'train' && record.capturedAt >= args.capturedBefore) return false\n  if (args.split === 'test' && record.capturedAt < args.capturedBefore) return false\n\n  const f = args.filter\n  if (!f) return true\n  if (f.kind && record.scenario.kind !== f.kind) return false\n  if (f.source) {\n    const sources = Array.isArray(f.source) ? f.source : [f.source]\n    if (!sources.includes(source as never)) return false\n  }\n  if (f.minComposite !== undefined || f.maxComposite !== undefined) {\n    const composites = Object.values(record.judgeScores).map((s) => s.composite)\n    const max = composites.length === 0 ? 0 : Math.max(...composites)\n    if (f.minComposite !== undefined && max < f.minComposite) return false\n    if (f.maxComposite !== undefined && max > f.maxComposite) return false\n  }\n  if (f.minTrust !== undefined && labelTrustRank(record.labelTrust) < labelTrustRank(f.minTrust)) {\n    return false\n  }\n  return true\n}\n\nfunction sha256(input: string): string {\n  return createHash('sha256').update(input).digest('hex').slice(0, 16)\n}\n\nfunction appendLine(path: string, line: string): void {\n  if (existsSync(path)) {\n    const existing = readFileSync(path, 'utf8')\n    writeFileSync(path, existing + line)\n  } else {\n    writeFileSync(path, line)\n  }\n}\n","/**\n * @experimental\n *\n * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like\n * (multiple commits allowed). A code-tier driver's `propose()` creates a\n * worktree, an agent commits the change into it, and `finalize()` returns a\n * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker\n * against the changed code. On promotion the worktree becomes the PR branch.\n *\n * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))\n * adapter can slot in without touching driver code. Only the git adapter\n * ships today. See `docs/design/self-improvement-engine.md`.\n */\n\nimport { execFileSync } from 'node:child_process'\nimport { existsSync } from 'node:fs'\nimport { basename, isAbsolute, join } from 'node:path'\nimport type { CodeSurface } from '../types'\n\nexport interface Worktree {\n  /** Absolute path to the checked-out worktree directory. */\n  path: string\n  /** The branch the worktree is on (becomes the PR branch on promotion). */\n  branch: string\n  /** The ref the worktree was forked from. */\n  baseRef: string\n}\n\nexport interface WorktreeAdapter {\n  /** Create an isolated worktree on a fresh branch off `baseRef`. */\n  create(opts: { baseRef: string; label: string }): Promise<Worktree>\n  /** Commit any pending changes in the worktree, then return a CodeSurface\n   *  pointing at it. The agent has already written its change into\n   *  `worktree.path` by the time this is called. */\n  finalize(worktree: Worktree, summary: string): Promise<CodeSurface>\n  /** Remove the worktree (and its branch) — called for losing candidates. */\n  discard(worktree: Worktree): Promise<void>\n}\n\nexport class WorktreeAdapterError extends Error {\n  constructor(\n    message: string,\n    readonly cause?: unknown,\n  ) {\n    super(message)\n    this.name = 'WorktreeAdapterError'\n  }\n}\n\nexport interface GitWorktreeAdapterOptions {\n  /** Repo root the worktrees fork from. */\n  repoRoot: string\n  /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */\n  worktreeDir?: string\n  /** Branch-name prefix. Default: `improve`. */\n  branchPrefix?: string\n  /** Test seam — defaults to a real `git` runner. */\n  git?: (args: string[], cwd: string) => string\n}\n\nfunction defaultGit(args: string[], cwd: string): string {\n  try {\n    return execFileSync('git', args, { cwd, encoding: 'utf8' }).trim()\n  } catch (err) {\n    const stderr =\n      err && typeof err === 'object' && 'stderr' in err\n        ? String((err as { stderr: unknown }).stderr)\n        : ''\n    throw new WorktreeAdapterError(`git ${args.join(' ')} failed: ${stderr || String(err)}`, err)\n  }\n}\n\n/** Slugify a label into a branch-safe segment. */\nfunction slug(label: string): string {\n  return (\n    label\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 48) || 'candidate'\n  )\n}\n\nexport function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter {\n  const git = opts.git ?? defaultGit\n  const worktreeDir = opts.worktreeDir ?? join(opts.repoRoot, '.worktrees')\n  const branchPrefix = opts.branchPrefix ?? 'improve'\n\n  return {\n    async create({ baseRef, label }) {\n      const id = `${slug(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`\n      const branch = `${branchPrefix}/${id}`\n      const path = join(worktreeDir, id)\n      git(['worktree', 'add', '-b', branch, path, baseRef], opts.repoRoot)\n      return { path, branch, baseRef }\n    },\n\n    async finalize(worktree, summary) {\n      // Stage + commit any pending changes the agent left in the worktree.\n      // A no-op commit is refused by git, so only commit when the tree is dirty.\n      const status = git(['status', '--porcelain'], worktree.path)\n      if (status.length > 0) {\n        git(['add', '-A'], worktree.path)\n        git(['commit', '-m', summary], worktree.path)\n      }\n      return {\n        kind: 'code',\n        worktreeRef: worktree.path,\n        baseRef: worktree.baseRef,\n        summary,\n      }\n    },\n\n    async discard(worktree) {\n      // Remove the worktree, then delete its branch. Force-remove because the\n      // worktree may hold uncommitted experiment state we're discarding.\n      git(['worktree', 'remove', '--force', worktree.path], opts.repoRoot)\n      git(['branch', '-D', worktree.branch], opts.repoRoot)\n    },\n  }\n}\n\n/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can\n *  run the worker in. A path ref is returned as-is; anything else is treated\n *  as a ref under the adapter's worktree dir. */\nexport function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string {\n  if (isAbsolute(surface.worktreeRef) && existsSync(surface.worktreeRef)) return surface.worktreeRef\n  if (worktreeDir) return join(worktreeDir, basename(surface.worktreeRef))\n  return surface.worktreeRef\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;AAqBd,IAAM,4BAAN,cAAwC,MAAM;AAAA,EACnD,YACkB,MAChB,SACA;AACA,UAAM,OAAO;AAHG;AAIhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAMpB;AAQO,IAAM,yBAAN,MAA6D;AAAA,EAIlE,YAA6B,SAAwC;AAAxC;AAC3B,QAAI,CAAC,WAAW,QAAQ,IAAI,EAAG,WAAU,QAAQ,MAAM,EAAE,WAAW,KAAK,CAAC;AAC1E,SAAK,MAAM,QAAQ,OAAO,KAAK;AAAA,EACjC;AAAA,EAH6B;AAAA,EAHZ;AAAA,EACA,aAAa,oBAAI,IAA4B;AAAA,EAO9D,MAAM,QAAQ,OAA4C;AACxD,SAAK,iBAAiB,KAAK;AAC3B,SAAK,gBAAgB,KAAK;AAC1B,UAAM,SAAS,KAAK,SAAS,KAAK;AAClC,UAAM,OAAO,KAAK,cAAc,MAAM,MAAM;AAC5C,UAAM,OAAO,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAItC,eAAW,MAAM,IAAI;AAAA,EACvB;AAAA,EAEA,MAAM,OAAO,MAAmE;AAC9E,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAEA,UAAM,MAA+B,CAAC;AACtC,eAAW,UAAU,aAAa;AAGhC,UAAI,KAAK,UAAU,WAAW,WAAW,oBAAoB;AAC3D,cAAM,WAAW,qBAAqB,KAAK,QAAQ,QAAQ,kBAAkB;AAC7E,YAAI,CAAC,SAAU;AAAA,MACjB;AACA,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,EAAG;AACvB,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,iBAAW,QAAQ,OAAO;AACxB,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,QAAQ;AACN;AAAA,QACF;AACA,YAAI,CAAC,cAAc,QAAQ,MAAM,MAAM,EAAG;AAC1C,YAAI,KAAK,MAAM;AAAA,MACjB;AAAA,IACF;AAGA,QAAI,KAAK,CAAC,GAAG,MAAM;AACjB,UAAI,EAAE,eAAe,EAAE,WAAY,QAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AACjF,aAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,IAChD,CAAC;AAED,WAAO,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,EAChC;AAAA,EAEA,MAAM,OAKH;AACD,UAAM,WAAmC,CAAC;AAC1C,UAAM,UAAsC;AAAA,MAC1C,YAAY;AAAA,MACZ,mBAAmB;AAAA,MACnB,eAAe;AAAA,IACjB;AACA,QAAI,QAAQ;AACZ,eAAW,UAAU,aAAa;AAChC,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,GAAG;AACrB,iBAAS,MAAM,IAAI;AACnB;AAAA,MACF;AACA,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,eAAS,MAAM,IAAI,MAAM;AACzB,eAAS,MAAM;AACf,iBAAW,QAAQ,OAAO;AACxB,YAAI,QAAoB;AACxB,YAAI;AACF,kBAAS,KAAK,MAAM,IAAI,EAA4B,cAAc;AAAA,QACpE,QAAQ;AAAA,QAER;AACA,gBAAQ,KAAK,KAAK;AAAA,MACpB;AAAA,IACF;AAIA,WAAO,EAAE,OAAO,OAAO,MAAM,OAAO,UAAU,QAAQ;AAAA,EACxD;AAAA,EAEQ,iBAAiB,OAAmC;AAC1D,QAAI,CAAC,MAAM,QAAQ;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,qBAAqB,MAAM,kBAAkB,WAAW,GAAG;AACpE,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,YAAY;AACrB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,iBAAiB;AAC1B,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,YAAY,SAAS,MAAM,MAAM,GAAG;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,YAAY,KAAK,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,OAAmC;AACzD,UAAM,MAAM,KAAK,QAAQ;AACzB,QAAI,CAAC,OAAO,CAAC,MAAM,gBAAiB;AACpC,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,WAAW;AACjB,QAAI,QAAQ,KAAK,WAAW,IAAI,MAAM,eAAe;AACrD,QAAI,CAAC,SAAS,MAAM,MAAM,iBAAiB,UAAU;AACnD,cAAQ,EAAE,QAAQ,MAAM,iBAAiB,eAAe,KAAK,OAAO,EAAE;AACtE,WAAK,WAAW,IAAI,MAAM,iBAAiB,KAAK;AAAA,IAClD;AACA,QAAI,MAAM,SAAS,KAAK;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA,gCAAgC,MAAM,eAAe,aAAa,GAAG;AAAA,MACvE;AAAA,IACF;AACA,UAAM,SAAS;AAAA,EACjB;AAAA,EAEQ,SAAS,OAAoD;AACnE,UAAM,aAAa;AAAA,MACjB,KAAK,UAAU;AAAA,QACb,IAAI,MAAM,SAAS;AAAA,QACnB,KAAK,MAAM;AAAA,QACX,IAAI,MAAM;AAAA,QACV,KAAK,MAAM;AAAA,MACb,CAAC;AAAA,IACH;AAGA,WAAO;AAAA,MACL,GAAG;AAAA,MACH;AAAA,MACA,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEQ,cAAc,QAAwB;AAC5C,WAAO,KAAK,KAAK,QAAQ,MAAM,GAAG,MAAM,QAAQ;AAAA,EAClD;AACF;AAEA,IAAM,cAAgD;AAAA,EACpD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,qBACP,QACA,QACS;AACT,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,EAAG,QAAO,OAAO,SAAS,MAAM;AACxD,SAAO,WAAW;AACpB;AAEA,SAAS,cACP,QACA,MACA,QACS;AAET,MAAI,KAAK,UAAU,WAAW,OAAO,cAAc,KAAK,eAAgB,QAAO;AAC/E,MAAI,KAAK,UAAU,UAAU,OAAO,aAAa,KAAK,eAAgB,QAAO;AAE7E,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,EAAG,QAAO;AACf,MAAI,EAAE,QAAQ,OAAO,SAAS,SAAS,EAAE,KAAM,QAAO;AACtD,MAAI,EAAE,QAAQ;AACZ,UAAM,UAAU,MAAM,QAAQ,EAAE,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM;AAC9D,QAAI,CAAC,QAAQ,SAAS,MAAe,EAAG,QAAO;AAAA,EACjD;AACA,MAAI,EAAE,iBAAiB,UAAa,EAAE,iBAAiB,QAAW;AAChE,UAAM,aAAa,OAAO,OAAO,OAAO,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3E,UAAM,MAAM,WAAW,WAAW,IAAI,IAAI,KAAK,IAAI,GAAG,UAAU;AAChE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AACjE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AAAA,EACnE;AACA,MAAI,EAAE,aAAa,UAAa,eAAe,OAAO,UAAU,IAAI,eAAe,EAAE,QAAQ,GAAG;AAC9F,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEA,SAAS,OAAO,OAAuB;AACrC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,MAAM,GAAG,EAAE;AACrE;AAEA,SAAS,WAAW,MAAc,MAAoB;AACpD,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,WAAW,aAAa,MAAM,MAAM;AAC1C,kBAAc,MAAM,WAAW,IAAI;AAAA,EACrC,OAAO;AACL,kBAAc,MAAM,IAAI;AAAA,EAC1B;AACF;;;AClSA,SAAS,oBAAoB;AAC7B,SAAS,cAAAA,mBAAkB;AAC3B,SAAS,UAAU,YAAY,QAAAC,aAAY;AAuBpC,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC9C,YACE,SACS,OACT;AACA,UAAM,OAAO;AAFJ;AAGT,SAAK,OAAO;AAAA,EACd;AAAA,EAJW;AAKb;AAaA,SAAS,WAAW,MAAgB,KAAqB;AACvD,MAAI;AACF,WAAO,aAAa,OAAO,MAAM,EAAE,KAAK,UAAU,OAAO,CAAC,EAAE,KAAK;AAAA,EACnE,SAAS,KAAK;AACZ,UAAM,SACJ,OAAO,OAAO,QAAQ,YAAY,YAAY,MAC1C,OAAQ,IAA4B,MAAM,IAC1C;AACN,UAAM,IAAI,qBAAqB,OAAO,KAAK,KAAK,GAAG,CAAC,YAAY,UAAU,OAAO,GAAG,CAAC,IAAI,GAAG;AAAA,EAC9F;AACF;AAGA,SAAS,KAAK,OAAuB;AACnC,SACE,MACG,YAAY,EACZ,QAAQ,eAAe,GAAG,EAC1B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEO,SAAS,mBAAmB,MAAkD;AACnF,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,cAAc,KAAK,eAAeA,MAAK,KAAK,UAAU,YAAY;AACxE,QAAM,eAAe,KAAK,gBAAgB;AAE1C,SAAO;AAAA,IACL,MAAM,OAAO,EAAE,SAAS,MAAM,GAAG;AAC/B,YAAM,KAAK,GAAG,KAAK,KAAK,CAAC,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AAC9F,YAAM,SAAS,GAAG,YAAY,IAAI,EAAE;AACpC,YAAM,OAAOA,MAAK,aAAa,EAAE;AACjC,UAAI,CAAC,YAAY,OAAO,MAAM,QAAQ,MAAM,OAAO,GAAG,KAAK,QAAQ;AACnE,aAAO,EAAE,MAAM,QAAQ,QAAQ;AAAA,IACjC;AAAA,IAEA,MAAM,SAAS,UAAU,SAAS;AAGhC,YAAM,SAAS,IAAI,CAAC,UAAU,aAAa,GAAG,SAAS,IAAI;AAC3D,UAAI,OAAO,SAAS,GAAG;AACrB,YAAI,CAAC,OAAO,IAAI,GAAG,SAAS,IAAI;AAChC,YAAI,CAAC,UAAU,MAAM,OAAO,GAAG,SAAS,IAAI;AAAA,MAC9C;AACA,aAAO;AAAA,QACL,MAAM;AAAA,QACN,aAAa,SAAS;AAAA,QACtB,SAAS,SAAS;AAAA,QAClB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,QAAQ,UAAU;AAGtB,UAAI,CAAC,YAAY,UAAU,WAAW,SAAS,IAAI,GAAG,KAAK,QAAQ;AACnE,UAAI,CAAC,UAAU,MAAM,SAAS,MAAM,GAAG,KAAK,QAAQ;AAAA,IACtD;AAAA,EACF;AACF;AAKO,SAAS,oBAAoB,SAAsB,aAA8B;AACtF,MAAI,WAAW,QAAQ,WAAW,KAAKD,YAAW,QAAQ,WAAW,EAAG,QAAO,QAAQ;AACvF,MAAI,YAAa,QAAOC,MAAK,aAAa,SAAS,QAAQ,WAAW,CAAC;AACvE,SAAO,QAAQ;AACjB;","names":["existsSync","join"]}

package/dist/{chunk-MHQPVHXU.js → chunk-6QDKWHLS.js}

@@ -1,6 +1,6 @@
 import {
   __export
-} from "./chunk-NSBPE2FW.js";
+} from "./chunk-PZ5AY32C.js";
 
 // src/benchmarks/index.ts
 var benchmarks_exports = {};
@@ -220,4 +220,4 @@ export {
   routing_exports,
   benchmarks_exports
 };
-//# sourceMappingURL=chunk-MHQPVHXU.js.map
+//# sourceMappingURL=chunk-6QDKWHLS.js.map

package/dist/{chunk-N4SBKEPJ.js → chunk-GBHRUAOF.js}

@@ -2,9 +2,111 @@ import {
   DEFAULT_REDACTION_RULES
 } from "./chunk-GGE4NNQT.js";
 import {
+  AgentEvalError,
   ValidationError
 } from "./chunk-QYJT52YW.js";
 
+// src/integrity/backend-integrity.ts
+var BackendIntegrityError = class extends AgentEvalError {
+  constructor(message, report) {
+    super("backend_integrity", message);
+    this.report = report;
+  }
+  report;
+};
+function isStubRecord(rec) {
+  return rec.tokenUsage.input === 0 && rec.tokenUsage.output === 0;
+}
+function isUncostedRecord(rec) {
+  return rec.tokenUsage.output > 0 && rec.costUsd === 0;
+}
+function summarizeBackendIntegrity(records) {
+  const totalRecords = records.length;
+  let stubRecords = 0;
+  let realRecords = 0;
+  let uncostedRecords = 0;
+  let totalInputTokens = 0;
+  let totalOutputTokens = 0;
+  let totalCostUsd = 0;
+  for (const rec of records) {
+    totalInputTokens += rec.tokenUsage.input;
+    totalOutputTokens += rec.tokenUsage.output;
+    totalCostUsd += rec.costUsd;
+    if (isStubRecord(rec)) stubRecords++;
+    else realRecords++;
+    if (isUncostedRecord(rec)) uncostedRecords++;
+  }
+  const verdict = totalRecords === 0 ? "stub" : stubRecords === totalRecords ? "stub" : stubRecords === 0 ? "real" : "mixed";
+  const diagnosis = buildDiagnosis({
+    totalRecords,
+    stubRecords,
+    realRecords,
+    uncostedRecords,
+    totalInputTokens,
+    totalOutputTokens,
+    totalCostUsd,
+    verdict
+  });
+  return {
+    totalRecords,
+    stubRecords,
+    realRecords,
+    uncostedRecords,
+    totalInputTokens,
+    totalOutputTokens,
+    totalCostUsd,
+    verdict,
+    diagnosis
+  };
+}
+function buildDiagnosis(r) {
+  if (r.totalRecords === 0) {
+    return "no records \u2014 eval produced zero runs; backend likely failed before first turn";
+  }
+  if (r.verdict === "stub") {
+    return [
+      `all ${r.totalRecords} records have zero token usage \u2014 the LLM backend was never called.`,
+      "common causes: --backend sandbox without a sandbox bridge running; stub model returning hard-coded strings;",
+      "auth misconfigured so requests were silently dropped before the LLM. Re-run with --backend tcloud and TANGLE_API_KEY set,",
+      "or boot the cli-bridge / sandbox before invoking the eval."
+    ].join(" ");
+  }
+  if (r.verdict === "mixed") {
+    const pct = (r.stubRecords / r.totalRecords * 100).toFixed(0);
+    return [
+      `${r.stubRecords}/${r.totalRecords} records (${pct}%) have zero token usage \u2014 the backend partially failed.`,
+      "common causes: rate-limit cascade (429s after the first N personas);",
+      "transient auth expiry mid-run; provider outage. Treat the affected records as missing data, not agent failures."
+    ].join(" ");
+  }
+  if (r.uncostedRecords > 0) {
+    const pct = (r.uncostedRecords / r.totalRecords * 100).toFixed(0);
+    return [
+      `${r.totalRecords} records with real LLM activity (in=${r.totalInputTokens}, out=${r.totalOutputTokens} tokens).`,
+      `${r.uncostedRecords} (${pct}%) have output tokens but costUsd=0 \u2014 cost ledger is mis-wired (no input-token`,
+      "propagation from the runtime stream into RunRecord)."
+    ].join(" ");
+  }
+  return `${r.totalRecords} records with real LLM activity (in=${r.totalInputTokens}, out=${r.totalOutputTokens} tokens, $${r.totalCostUsd.toFixed(4)}).`;
+}
+function assertRealBackend(records, opts = {}) {
+  const report = summarizeBackendIntegrity(records);
+  const allowMixed = opts.allowMixed ?? true;
+  if (report.verdict === "stub") {
+    throw new BackendIntegrityError(
+      `backend-integrity: ran against a stub or unconfigured backend \u2014 ${report.diagnosis}`,
+      report
+    );
+  }
+  if (!allowMixed && report.verdict === "mixed") {
+    throw new BackendIntegrityError(
+      `backend-integrity: partial backend failure rejected \u2014 ${report.diagnosis}`,
+      report
+    );
+  }
+  return report;
+}
+
 // src/dataset.ts
 var HoldoutLockedError = class extends ValidationError {
   constructor(datasetName) {
@@ -746,6 +848,9 @@ function parseReflectionResponse(raw, maxProposals) {
 }
 
 export {
+  BackendIntegrityError,
+  summarizeBackendIntegrity,
+  assertRealBackend,
   HoldoutLockedError,
   Dataset,
   hashScenarios,
@@ -759,4 +864,4 @@ export {
   buildReflectionPrompt,
   parseReflectionResponse
 };
-//# sourceMappingURL=chunk-N4SBKEPJ.js.map
+//# sourceMappingURL=chunk-GBHRUAOF.js.map