@tangle-network/agent-eval 0.129.0 → 0.130.0

package/dist/builder-eval/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/builder-eval/builder-session.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts","../../src/builder-eval/three-layer-eval.ts"],"sourcesContent":["/**\n * BuilderSession — ties a builder-of-builders workflow together.\n *\n * Models agent-builder's shape: Project → Chat → Edit → Ship → App →\n * AppAgent. Each layer is a Run (linked via parentRunId). The\n * framework-enforced invariants:\n *\n *   - One Project → many Chats; chatId scopes runs within a project.\n *   - One Chat = one builder Run with `layer='builder'`.\n *   - One Ship = one child Run with `layer='app-build'` + SandboxHarness.\n *   - One AppScenario = one grandchild Run with `layer='app-runtime'`.\n *\n * Consumers obtain a BuilderSession, call `startChat`, drive the\n * builder agent (emitting spans), and call `ship` / `runAppScenario`\n * as the workflow progresses. The session reconstructs itself from\n * trace data via `resume(store, projectId)`.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from '../sandbox-harness'\nimport { SandboxHarness } from '../sandbox-harness'\nimport type { TestGradedRunResult, TestGradedScenario } from '../test-graded-scenario'\nimport { runTestGradedScenario } from '../test-graded-scenario'\nimport { TraceEmitter } from '../trace/emitter'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BuilderSessionInit {\n  projectId: string\n  chatId?: string\n  /** Free-form: user's task description, project name, etc. Stored on the builder Run. */\n  tags?: Record<string, string>\n}\n\nexport interface ShipOptions {\n  harness: HarnessConfig\n  driver?: SandboxDriver\n  /** scenarioId of this app-build run. Defaults to `${projectId}/build`. */\n  scenarioId?: string\n}\n\nexport interface RunAppScenarioOptions {\n  scenario: TestGradedScenario\n  /** Harness driver override; defaults to the one the session was created with. */\n  driver?: SandboxDriver\n}\n\nexport class BuilderSession {\n  private store: TraceStore\n  private builderEmitter: TraceEmitter\n  readonly projectId: string\n  readonly chatId: string\n  private builderRunId?: string\n  private lastBuildRunId?: string\n  private defaultDriver?: SandboxDriver\n\n  constructor(store: TraceStore, init: BuilderSessionInit, driver?: SandboxDriver) {\n    this.store = store\n    this.projectId = init.projectId\n    this.chatId = init.chatId ?? cryptoId()\n    this.defaultDriver = driver\n    this.builderEmitter = new TraceEmitter(store)\n  }\n\n  /** Start the builder (L0) run for this chat. Returns the runId. */\n  async startChat(scenarioId = `${this.projectId}/chat`): Promise<string> {\n    await this.builderEmitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'builder',\n    })\n    this.builderRunId = this.builderEmitter.runId\n    return this.builderRunId\n  }\n\n  /** The emitter for builder-level spans (edits, LLM calls, tool invocations). */\n  get emitter(): TraceEmitter {\n    if (!this.builderRunId) throw new Error('BuilderSession.emitter: call startChat() first')\n    return this.builderEmitter\n  }\n\n  /**\n   * Ship the project's generated app: run the sandbox harness as a child\n   * Run (`layer='app-build'`). Returns the build result + runId.\n   */\n  async ship(options: ShipOptions): Promise<{ runId: string; result: SandboxHarnessResult }> {\n    if (!this.builderRunId) throw new Error('BuilderSession.ship: call startChat() first')\n    const buildEmitter = new TraceEmitter(this.store)\n    await buildEmitter.startRun({\n      scenarioId: options.scenarioId ?? `${this.projectId}/build`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    const harness = new SandboxHarness(options.driver ?? this.defaultDriver)\n    const result = await harness.run(options.harness, buildEmitter)\n    await buildEmitter.endRun({\n      pass: result.passed,\n      score: result.score,\n      failureClass: result.passed ? 'success' : 'sandbox_failure',\n    })\n    this.lastBuildRunId = buildEmitter.runId\n    return { runId: buildEmitter.runId, result }\n  }\n\n  /**\n   * Run a domain scenario against the just-built app as a grandchild Run\n   * (`layer='app-runtime'`). The `ship` call must precede this so the\n   * parent is set correctly; if no build exists yet the session attaches\n   * directly to the builder run (useful for prototypes).\n   */\n  async runAppScenario(options: RunAppScenarioOptions): Promise<TestGradedRunResult> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error('BuilderSession.runAppScenario: call startChat() + ship() first')\n    const { scenario, driver } = options\n    const result = await runTestGradedScenario(scenario, this.store, {\n      driver: driver ?? this.defaultDriver,\n      provenance: { codeSha: undefined, promptSha: undefined, modelFingerprint: undefined },\n    })\n    // Attach to the parent chain by updating the stored Run in place.\n    await this.store.updateRun(result.runId, {\n      parentRunId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'app-runtime',\n    })\n    return result\n  }\n\n  /** Record an end-of-chat meta score (judge verdict on whether the builder\n   *  served the user's intent). Accepts a numeric score + optional rationale. */\n  async recordMetaScore(score: number, rationale?: string): Promise<void> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordMetaScore: call startChat() first')\n    await this.builderEmitter.recordJudge({\n      judgeId: 'builder-meta',\n      targetSpanId: this.builderRunId, // attach to the builder run itself\n      dimension: 'user_intent_satisfaction',\n      score,\n      rationale,\n      name: 'builder-meta',\n    })\n  }\n\n  /** Close the builder Run with a final outcome. */\n  async endChat(outcome: { pass: boolean; score?: number; notes?: string }): Promise<void> {\n    await this.builderEmitter.endRun({\n      pass: outcome.pass,\n      score: outcome.score,\n      notes: outcome.notes,\n    })\n  }\n\n  /**\n   * Inline app-runtime run — for cases where the \"scenario\" isn't a\n   * SWE-bench-style test suite but a live agent interaction (LLM chat,\n   * domain flow). Returns an emitter bound to a fresh Run in the\n   * `app-runtime` layer; caller emits spans inside and calls\n   * `.endRun()` with the final verdict.\n   */\n  async startAppRuntime(scenarioId: string): Promise<TraceEmitter> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error(\n        'BuilderSession.startAppRuntime: call startChat() + (optionally) ship() first',\n      )\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId,\n      layer: 'app-runtime',\n    })\n    return emitter\n  }\n\n  /**\n   * Lightweight \"ship marker\" — record an app-build Run with a caller-\n   * provided verdict. Use when there isn't a sandbox harness to run but\n   * you still want to mark the build state at publish time.\n   */\n  async recordShipMarker(args: {\n    pass: boolean\n    score: number\n    scenarioId?: string\n    notes?: string\n  }): Promise<string> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordShipMarker: call startChat() first')\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId: args.scenarioId ?? `${this.projectId}/ship`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    await emitter.endRun({\n      pass: args.pass,\n      score: args.score,\n      failureClass: args.pass ? 'success' : 'sandbox_failure',\n      notes: args.notes,\n    })\n    this.lastBuildRunId = emitter.runId\n    return emitter.runId\n  }\n\n  get lastBuildRunIdValue(): string | undefined {\n    return this.lastBuildRunId\n  }\n  get builderRunIdValue(): string | undefined {\n    return this.builderRunId\n  }\n}\n\n/**\n * Reconstruct the most recent BuilderSession state for a given project —\n * returns { builderRunId, lastBuildRunId, chatRuns }. For chat-first UIs\n * this is how a resumed session finds its place in the edit history.\n */\nexport async function resumeBuilderSession(\n  store: TraceStore,\n  projectId: string,\n): Promise<{\n  projectId: string\n  chatRuns: Run[]\n  lastBuilderRun?: Run\n  lastBuildRun?: Run\n  lastAppRuntimeRuns: Run[]\n}> {\n  const runs = await store.listRuns({ projectId })\n  const chatRuns = runs\n    .filter((r) => r.layer === 'builder')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const buildRuns = runs\n    .filter((r) => r.layer === 'app-build')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const appRuntimeRuns = runs\n    .filter((r) => r.layer === 'app-runtime')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  return {\n    projectId,\n    chatRuns,\n    lastBuilderRun: chatRuns[0],\n    lastBuildRun: buildRuns[0],\n    lastAppRuntimeRuns: appRuntimeRuns,\n  }\n}\n\nfunction cryptoId(): string {\n  if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n","/**\n * Meta-eval correlation — the highest-leverage signal in the framework.\n *\n * Given a corpus of three-layer project reports, compute how well each\n * pair of layers correlates. The question we care about most:\n *\n *   Does `metaScore` (what the builder thinks it did) predict\n *   `runtimeScore` (what the user actually gets)?\n *\n * If r < ~0.4, the builder's self-scoring is broken — it's optimizing\n * for something other than real-world success. If r > 0.7, meta_score\n * is a usable proxy and can drive CI gates cheaply.\n *\n * Non-parametric rank correlation (Spearman) is also reported because\n * meta scores are often ordinal-ish.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { ThreeLayerProjectReport } from './three-layer-eval'\n\nexport interface LayerCorrelation {\n  n: number\n  pearson: number\n  spearman: number\n}\n\nexport interface CorrelationReport {\n  /** Pairs present in the corpus (layers with ≥ 2 matched data points). */\n  metaVsBuild?: LayerCorrelation\n  metaVsRuntime?: LayerCorrelation\n  buildVsRuntime?: LayerCorrelation\n  /** Number of complete projects (all 3 scores present). */\n  completeProjects: number\n}\n\nexport function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport {\n  const completeProjects = reports.filter((r) => r.complete).length\n  return {\n    metaVsBuild: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.buildScore,\n    ),\n    metaVsRuntime: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.runtimeScore,\n    ),\n    buildVsRuntime: pairwise(\n      reports,\n      (r) => r.buildScore,\n      (r) => r.runtimeScore,\n    ),\n    completeProjects,\n  }\n}\n\nfunction pairwise(\n  reports: ThreeLayerProjectReport[],\n  a: (r: ThreeLayerProjectReport) => number | null,\n  b: (r: ThreeLayerProjectReport) => number | null,\n): LayerCorrelation | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const r of reports) {\n    const x = a(r)\n    const y = b(r)\n    if (x !== null && y !== null && Number.isFinite(x) && Number.isFinite(y)) {\n      xs.push(x)\n      ys.push(y)\n    }\n  }\n  if (xs.length < 2) return undefined\n  return {\n    n: xs.length,\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n  }\n}\n","/**\n * ProjectRegistry — project-level aggregation over the trace corpus.\n *\n * Thin reader over TraceStore that answers the questions a chat-first,\n * resumable UI needs:\n *   - listProjects() → project IDs with latest activity\n *   - projectTimeline(id) → chats + builds + runtime runs, chronological\n *   - projectChats(id) → chat-level summaries (turn count, outcome)\n *\n * All queries are pure reads; no state duplication.\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ProjectSummary {\n  projectId: string\n  chatCount: number\n  buildCount: number\n  appRuntimeCount: number\n  lastActivityAt: number\n  latestChatId?: string\n  latestOutcome?: { pass: boolean; score?: number }\n}\n\nexport interface ChatSummary {\n  chatId: string\n  projectId: string\n  builderRunId: string\n  startedAt: number\n  endedAt?: number\n  status: Run['status']\n  outcome?: Run['outcome']\n  /** Counts of spans emitted during the chat. */\n  llmTurns?: number\n  toolCalls?: number\n  buildRunId?: string\n  appRuntimeRunIds: string[]\n}\n\nexport interface ProjectTimelineEntry {\n  run: Run\n  layerBucket: 'chat' | 'build' | 'runtime' | 'other'\n}\n\nexport class ProjectRegistry {\n  constructor(private store: TraceStore) {}\n\n  async listProjects(): Promise<ProjectSummary[]> {\n    const runs = await this.store.listRuns()\n    const byProject = new Map<string, Run[]>()\n    for (const r of runs) {\n      if (!r.projectId) continue\n      const arr = byProject.get(r.projectId) ?? []\n      arr.push(r)\n      byProject.set(r.projectId, arr)\n    }\n    const summaries: ProjectSummary[] = []\n    for (const [projectId, projectRuns] of byProject) {\n      const sorted = projectRuns.slice().sort((a, b) => b.startedAt - a.startedAt)\n      const chats = projectRuns.filter((r) => r.layer === 'builder')\n      const builds = projectRuns.filter((r) => r.layer === 'app-build')\n      const runtimes = projectRuns.filter((r) => r.layer === 'app-runtime')\n      const latest = sorted[0]\n      if (!latest) continue\n      summaries.push({\n        projectId,\n        chatCount: chats.length,\n        buildCount: builds.length,\n        appRuntimeCount: runtimes.length,\n        lastActivityAt: latest.startedAt,\n        latestChatId: chats[0]?.chatId,\n        latestOutcome: latest.outcome\n          ? { pass: latest.outcome.pass ?? false, score: latest.outcome.score }\n          : undefined,\n      })\n    }\n    return summaries.sort((a, b) => b.lastActivityAt - a.lastActivityAt)\n  }\n\n  async projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]> {\n    const runs = await this.store.listRuns({ projectId })\n    const ordered = runs.slice().sort((a, b) => a.startedAt - b.startedAt)\n    return ordered.map((run) => ({\n      run,\n      layerBucket:\n        run.layer === 'builder'\n          ? 'chat'\n          : run.layer === 'app-build'\n            ? 'build'\n            : run.layer === 'app-runtime'\n              ? 'runtime'\n              : 'other',\n    }))\n  }\n\n  async projectChats(projectId: string): Promise<ChatSummary[]> {\n    const builderRuns = (await this.store.listRuns({ projectId, layer: 'builder' })).sort(\n      (a, b) => b.startedAt - a.startedAt,\n    )\n    const childrenFor = async (runId: string) => this.store.listRuns({ parentRunId: runId })\n    const out: ChatSummary[] = []\n    for (const run of builderRuns) {\n      const spans = await this.store.spans({ runId: run.runId })\n      const children = await childrenFor(run.runId)\n      const build = children.find((c) => c.layer === 'app-build')\n      const runtime: string[] = []\n      // Runtime runs can be grandchildren (attached to the build) or siblings\n      // when shipped skipped.\n      if (build) {\n        const grands = await childrenFor(build.runId)\n        for (const g of grands) if (g.layer === 'app-runtime') runtime.push(g.runId)\n      }\n      for (const c of children) if (c.layer === 'app-runtime') runtime.push(c.runId)\n      out.push({\n        chatId: run.chatId ?? run.runId,\n        projectId,\n        builderRunId: run.runId,\n        startedAt: run.startedAt,\n        endedAt: run.endedAt,\n        status: run.status,\n        outcome: run.outcome,\n        llmTurns: spans.filter((s) => s.kind === 'llm').length,\n        toolCalls: spans.filter((s) => s.kind === 'tool').length,\n        buildRunId: build?.runId,\n        appRuntimeRunIds: runtime,\n      })\n    }\n    return out\n  }\n}\n","/**\n * Three-layer evaluation — the canonical scoring breakdown for\n * builder-of-builders workflows.\n *\n *   meta_score:    did the builder understand + satisfy user intent?\n *                  (judge verdict attached to the builder run)\n *   build_score:   did the generated scaffold build + pass its own tests?\n *                  (outcome.score on the app-build child run)\n *   runtime_score: did the generated agent pass its domain scenarios?\n *                  (mean outcome.score over app-runtime grandchild runs)\n *\n * Returns a structured report per project. The cross-layer correlation\n * is the highest-leverage signal the framework computes — if\n * meta_score doesn't predict runtime_score, the builder's self-scoring\n * is broken.\n *\n * Scaffold-only mode: when a project has no `app-runtime` runs (e.g. a\n * scaffold-builder eval that grades compose + build without driving a\n * runtime scenario), `kind` is `'scaffold-only'` and `complete` measures\n * meta + build only. Consumers can tell the two apart without having to\n * interpret null-runtime as either \"not yet computed\" or \"N/A for this\n * project shape\".\n */\n\nimport { judgeSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport type ProjectKind = 'full' | 'scaffold-only'\n\nexport interface ThreeLayerProjectReport {\n  projectId: string\n  /**\n   * `'full'` when the project has at least one `app-runtime` run;\n   * `'scaffold-only'` when it only has meta + build layers. Lets\n   * downstream consumers treat a null runtime score as expected\n   * (scaffold-only) vs. missing (full, pipeline broke).\n   */\n  kind: ProjectKind\n  builderRunId?: string\n  /** Judge-verdict score on the builder run (0..1 after normalization). */\n  metaScore: number | null\n  buildRunId?: string\n  /** 0..1 from the sandbox harness (testsPassed / testsTotal). */\n  buildScore: number | null\n  appRuntimeRunIds: string[]\n  /** Mean of outcome.score over app-runtime runs, 0..1. Always null in scaffold-only mode. */\n  runtimeScore: number | null\n  runtimePassRate: number | null\n  /**\n   * Layer-aware completeness:\n   *   - `kind='full'`: all three layers scored\n   *   - `kind='scaffold-only'`: meta + build scored (runtime not applicable)\n   */\n  complete: boolean\n}\n\nexport async function scoreProject(\n  store: TraceStore,\n  projectId: string,\n): Promise<ThreeLayerProjectReport> {\n  const allRuns = await store.listRuns({ projectId })\n  const builder = latestByLayer(allRuns, 'builder')\n  const build = latestByLayer(allRuns, 'app-build')\n  const runtime = allRuns.filter((r) => r.layer === 'app-runtime')\n\n  const metaScore = builder ? await extractMetaScore(store, builder.runId) : null\n  const buildScore = build?.outcome?.score ?? null\n  const runtimeScores = runtime\n    .map((r) => r.outcome?.score)\n    .filter((s): s is number => typeof s === 'number')\n  const runtimeScore =\n    runtimeScores.length > 0\n      ? runtimeScores.reduce((a, b) => a + b, 0) / runtimeScores.length\n      : null\n  const runtimePassed = runtime.filter((r) => r.outcome?.pass === true).length\n  const runtimePassRate = runtime.length > 0 ? runtimePassed / runtime.length : null\n\n  const kind: ProjectKind = runtime.length === 0 ? 'scaffold-only' : 'full'\n  const complete =\n    kind === 'scaffold-only'\n      ? metaScore !== null && buildScore !== null\n      : metaScore !== null && buildScore !== null && runtimeScore !== null\n\n  return {\n    projectId,\n    kind,\n    builderRunId: builder?.runId,\n    metaScore,\n    buildRunId: build?.runId,\n    buildScore,\n    appRuntimeRunIds: runtime.map((r) => r.runId),\n    runtimeScore,\n    runtimePassRate,\n    complete,\n  }\n}\n\n/** Aggregate scoring across every project in a corpus. */\nexport async function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]> {\n  const runs = await store.listRuns()\n  const projectIds = [...new Set(runs.map((r) => r.projectId).filter((p): p is string => !!p))]\n  return Promise.all(projectIds.map((p) => scoreProject(store, p)))\n}\n\nfunction latestByLayer(runs: Run[], layer: Run['layer']): Run | undefined {\n  const filtered = runs.filter((r) => r.layer === layer).sort((a, b) => b.startedAt - a.startedAt)\n  return filtered[0]\n}\n\nasync function extractMetaScore(store: TraceStore, builderRunId: string): Promise<number | null> {\n  const js = await judgeSpans(store, builderRunId)\n  const meta = js.find(\n    (s) => s.judgeId === 'builder-meta' && s.dimension === 'user_intent_satisfaction',\n  )\n  if (!meta) return null\n  // Normalize score to 0..1. Accept 0-1 natively; 0-10 scale is also common.\n  if (meta.score >= 0 && meta.score <= 1) return meta.score\n  if (meta.score >= 0 && meta.score <= 10) return meta.score / 10\n  return null\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;AA8CO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACA;AAAA,EACC;AAAA,EACA;AAAA,EACD;AAAA,EACA;AAAA,EACA;AAAA,EAER,YAAY,OAAmB,MAA0B,QAAwB;AAC/E,SAAK,QAAQ;AACb,SAAK,YAAY,KAAK;AACtB,SAAK,SAAS,KAAK,UAAU,SAAS;AACtC,SAAK,gBAAgB;AACrB,SAAK,iBAAiB,IAAI,aAAa,KAAK;AAAA,EAC9C;AAAA;AAAA,EAGA,MAAM,UAAU,aAAa,GAAG,KAAK,SAAS,SAA0B;AACtE,UAAM,KAAK,eAAe,SAAS;AAAA,MACjC;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,OAAO;AAAA,IACT,CAAC;AACD,SAAK,eAAe,KAAK,eAAe;AACxC,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,UAAwB;AAC1B,QAAI,CAAC,KAAK,aAAc,OAAM,IAAI,MAAM,gDAAgD;AACxF,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,KAAK,SAAgF;AACzF,QAAI,CAAC,KAAK,aAAc,OAAM,IAAI,MAAM,6CAA6C;AACrF,UAAM,eAAe,IAAI,aAAa,KAAK,KAAK;AAChD,UAAM,aAAa,SAAS;AAAA,MAC1B,YAAY,QAAQ,cAAc,GAAG,KAAK,SAAS;AAAA,MACnD,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,aAAa,KAAK;AAAA,MAClB,OAAO;AAAA,IACT,CAAC;AACD,UAAM,UAAU,IAAI,eAAe,QAAQ,UAAU,KAAK,aAAa;AACvE,UAAM,SAAS,MAAM,QAAQ,IAAI,QAAQ,SAAS,YAAY;AAC9D,UAAM,aAAa,OAAO;AAAA,MACxB,MAAM,OAAO;AAAA,MACb,OAAO,OAAO;AAAA,MACd,cAAc,OAAO,SAAS,YAAY;AAAA,IAC5C,CAAC;AACD,SAAK,iBAAiB,aAAa;AACnC,WAAO,EAAE,OAAO,aAAa,OAAO,OAAO;AAAA,EAC7C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,SAA8D;AACjF,UAAM,cAAc,KAAK,kBAAkB,KAAK;AAChD,QAAI,CAAC;AACH,YAAM,IAAI,MAAM,gEAAgE;AAClF,UAAM,EAAE,UAAU,OAAO,IAAI;AAC7B,UAAM,SAAS,MAAM,sBAAsB,UAAU,KAAK,OAAO;AAAA,MAC/D,QAAQ,UAAU,KAAK;AAAA,MACvB,YAAY,EAAE,SAAS,QAAW,WAAW,QAAW,kBAAkB,OAAU;AAAA,IACtF,CAAC;AAED,UAAM,KAAK,MAAM,UAAU,OAAO,OAAO;AAAA,MACvC;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,OAAO;AAAA,IACT,CAAC;AACD,WAAO;AAAA,EACT;AAAA;AAAA;AAAA,EAIA,MAAM,gBAAgB,OAAe,WAAmC;AACtE,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,wDAAwD;AAC1E,UAAM,KAAK,eAAe,YAAY;AAAA,MACpC,SAAS;AAAA,MACT,cAAc,KAAK;AAAA;AAAA,MACnB,WAAW;AAAA,MACX;AAAA,MACA;AAAA,MACA,MAAM;AAAA,IACR,CAAC;AAAA,EACH;AAAA;AAAA,EAGA,MAAM,QAAQ,SAA2E;AACvF,UAAM,KAAK,eAAe,OAAO;AAAA,MAC/B,MAAM,QAAQ;AAAA,MACd,OAAO,QAAQ;AAAA,MACf,OAAO,QAAQ;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,gBAAgB,YAA2C;AAC/D,UAAM,cAAc,KAAK,kBAAkB,KAAK;AAChD,QAAI,CAAC;AACH,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AACF,UAAM,UAAU,IAAI,aAAa,KAAK,KAAK;AAC3C,UAAM,QAAQ,SAAS;AAAA,MACrB;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA,OAAO;AAAA,IACT,CAAC;AACD,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBAAiB,MAKH;AAClB,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,yDAAyD;AAC3E,UAAM,UAAU,IAAI,aAAa,KAAK,KAAK;AAC3C,UAAM,QAAQ,SAAS;AAAA,MACrB,YAAY,KAAK,cAAc,GAAG,KAAK,SAAS;AAAA,MAChD,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,aAAa,KAAK;AAAA,MAClB,OAAO;AAAA,IACT,CAAC;AACD,UAAM,QAAQ,OAAO;AAAA,MACnB,MAAM,KAAK;AAAA,MACX,OAAO,KAAK;AAAA,MACZ,cAAc,KAAK,OAAO,YAAY;AAAA,MACtC,OAAO,KAAK;AAAA,IACd,CAAC;AACD,SAAK,iBAAiB,QAAQ;AAC9B,WAAO,QAAQ;AAAA,EACjB;AAAA,EAEA,IAAI,sBAA0C;AAC5C,WAAO,KAAK;AAAA,EACd;AAAA,EACA,IAAI,oBAAwC;AAC1C,WAAO,KAAK;AAAA,EACd;AACF;AAOA,eAAsB,qBACpB,OACA,WAOC;AACD,QAAM,OAAO,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;AAC/C,QAAM,WAAW,KACd,OAAO,CAAC,MAAM,EAAE,UAAU,SAAS,EACnC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,QAAM,YAAY,KACf,OAAO,CAAC,MAAM,EAAE,UAAU,WAAW,EACrC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,QAAM,iBAAiB,KACpB,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa,EACvC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,gBAAgB,SAAS,CAAC;AAAA,IAC1B,cAAc,UAAU,CAAC;AAAA,IACzB,oBAAoB;AAAA,EACtB;AACF;AAEA,SAAS,WAAmB;AAC1B,MAAI,OAAO,WAAW,QAAQ,eAAe,WAAY,QAAO,WAAW,OAAO,WAAW;AAC7F,SAAO,GAAG,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,EAAE,CAAC;AAC9E;;;AC5NO,SAAS,gBAAgB,SAAuD;AACrF,QAAM,mBAAmB,QAAQ,OAAO,CAAC,MAAM,EAAE,QAAQ,EAAE;AAC3D,SAAO;AAAA,IACL,aAAa;AAAA,MACX;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA,eAAe;AAAA,MACb;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA,gBAAgB;AAAA,MACd;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,SACP,SACA,GACA,GAC8B;AAC9B,QAAM,KAAe,CAAC;AACtB,QAAM,KAAe,CAAC;AACtB,aAAW,KAAK,SAAS;AACvB,UAAM,IAAI,EAAE,CAAC;AACb,UAAM,IAAI,EAAE,CAAC;AACb,QAAI,MAAM,QAAQ,MAAM,QAAQ,OAAO,SAAS,CAAC,KAAK,OAAO,SAAS,CAAC,GAAG;AACxE,SAAG,KAAK,CAAC;AACT,SAAG,KAAK,CAAC;AAAA,IACX;AAAA,EACF;AACA,MAAI,GAAG,SAAS,EAAG,QAAO;AAC1B,SAAO;AAAA,IACL,GAAG,GAAG;AAAA,IACN,SAAS,SAAS,IAAI,EAAE;AAAA,IACxB,UAAU,UAAU,IAAI,EAAE;AAAA,EAC5B;AACF;;;ACjCO,IAAM,kBAAN,MAAsB;AAAA,EAC3B,YAAoB,OAAmB;AAAnB;AAAA,EAAoB;AAAA,EAApB;AAAA,EAEpB,MAAM,eAA0C;AAC9C,UAAM,OAAO,MAAM,KAAK,MAAM,SAAS;AACvC,UAAM,YAAY,oBAAI,IAAmB;AACzC,eAAW,KAAK,MAAM;AACpB,UAAI,CAAC,EAAE,UAAW;AAClB,YAAM,MAAM,UAAU,IAAI,EAAE,SAAS,KAAK,CAAC;AAC3C,UAAI,KAAK,CAAC;AACV,gBAAU,IAAI,EAAE,WAAW,GAAG;AAAA,IAChC;AACA,UAAM,YAA8B,CAAC;AACrC,eAAW,CAAC,WAAW,WAAW,KAAK,WAAW;AAChD,YAAM,SAAS,YAAY,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3E,YAAM,QAAQ,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,SAAS;AAC7D,YAAM,SAAS,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,WAAW;AAChE,YAAM,WAAW,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa;AACpE,YAAM,SAAS,OAAO,CAAC;AACvB,UAAI,CAAC,OAAQ;AACb,gBAAU,KAAK;AAAA,QACb;AAAA,QACA,WAAW,MAAM;AAAA,QACjB,YAAY,OAAO;AAAA,QACnB,iBAAiB,SAAS;AAAA,QAC1B,gBAAgB,OAAO;AAAA,QACvB,cAAc,MAAM,CAAC,GAAG;AAAA,QACxB,eAAe,OAAO,UAClB,EAAE,MAAM,OAAO,QAAQ,QAAQ,OAAO,OAAO,OAAO,QAAQ,MAAM,IAClE;AAAA,MACN,CAAC;AAAA,IACH;AACA,WAAO,UAAU,KAAK,CAAC,GAAG,MAAM,EAAE,iBAAiB,EAAE,cAAc;AAAA,EACrE;AAAA,EAEA,MAAM,gBAAgB,WAAoD;AACxE,UAAM,OAAO,MAAM,KAAK,MAAM,SAAS,EAAE,UAAU,CAAC;AACpD,UAAM,UAAU,KAAK,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACrE,WAAO,QAAQ,IAAI,CAAC,SAAS;AAAA,MAC3B;AAAA,MACA,aACE,IAAI,UAAU,YACV,SACA,IAAI,UAAU,cACZ,UACA,IAAI,UAAU,gBACZ,YACA;AAAA,IACZ,EAAE;AAAA,EACJ;AAAA,EAEA,MAAM,aAAa,WAA2C;AAC5D,UAAM,eAAe,MAAM,KAAK,MAAM,SAAS,EAAE,WAAW,OAAO,UAAU,CAAC,GAAG;AAAA,MAC/E,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE;AAAA,IAC5B;AACA,UAAM,cAAc,OAAO,UAAkB,KAAK,MAAM,SAAS,EAAE,aAAa,MAAM,CAAC;AACvF,UAAM,MAAqB,CAAC;AAC5B,eAAW,OAAO,aAAa;AAC7B,YAAM,QAAQ,MAAM,KAAK,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;AACzD,YAAM,WAAW,MAAM,YAAY,IAAI,KAAK;AAC5C,YAAM,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,UAAU,WAAW;AAC1D,YAAM,UAAoB,CAAC;AAG3B,UAAI,OAAO;AACT,cAAM,SAAS,MAAM,YAAY,MAAM,KAAK;AAC5C,mBAAW,KAAK,OAAQ,KAAI,EAAE,UAAU,cAAe,SAAQ,KAAK,EAAE,KAAK;AAAA,MAC7E;AACA,iBAAW,KAAK,SAAU,KAAI,EAAE,UAAU,cAAe,SAAQ,KAAK,EAAE,KAAK;AAC7E,UAAI,KAAK;AAAA,QACP,QAAQ,IAAI,UAAU,IAAI;AAAA,QAC1B;AAAA,QACA,cAAc,IAAI;AAAA,QAClB,WAAW,IAAI;AAAA,QACf,SAAS,IAAI;AAAA,QACb,QAAQ,IAAI;AAAA,QACZ,SAAS,IAAI;AAAA,QACb,UAAU,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,EAAE;AAAA,QAChD,WAAW,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE;AAAA,QAClD,YAAY,OAAO;AAAA,QACnB,kBAAkB;AAAA,MACpB,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AACF;;;ACzEA,eAAsB,aACpB,OACA,WACkC;AAClC,QAAM,UAAU,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;AAClD,QAAM,UAAU,cAAc,SAAS,SAAS;AAChD,QAAM,QAAQ,cAAc,SAAS,WAAW;AAChD,QAAM,UAAU,QAAQ,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa;AAE/D,QAAM,YAAY,UAAU,MAAM,iBAAiB,OAAO,QAAQ,KAAK,IAAI;AAC3E,QAAM,aAAa,OAAO,SAAS,SAAS;AAC5C,QAAM,gBAAgB,QACnB,IAAI,CAAC,MAAM,EAAE,SAAS,KAAK,EAC3B,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACnD,QAAM,eACJ,cAAc,SAAS,IACnB,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc,SACzD;AACN,QAAM,gBAAgB,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,SAAS,IAAI,EAAE;AACtE,QAAM,kBAAkB,QAAQ,SAAS,IAAI,gBAAgB,QAAQ,SAAS;AAE9E,QAAM,OAAoB,QAAQ,WAAW,IAAI,kBAAkB;AACnE,QAAM,WACJ,SAAS,kBACL,cAAc,QAAQ,eAAe,OACrC,cAAc,QAAQ,eAAe,QAAQ,iBAAiB;AAEpE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc,SAAS;AAAA,IACvB;AAAA,IACA,YAAY,OAAO;AAAA,IACnB;AAAA,IACA,kBAAkB,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,IAC5C;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAGA,eAAsB,iBAAiB,OAAuD;AAC5F,QAAM,OAAO,MAAM,MAAM,SAAS;AAClC,QAAM,aAAa,CAAC,GAAG,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,OAAO,CAAC,MAAmB,CAAC,CAAC,CAAC,CAAC,CAAC;AAC5F,SAAO,QAAQ,IAAI,WAAW,IAAI,CAAC,MAAM,aAAa,OAAO,CAAC,CAAC,CAAC;AAClE;AAEA,SAAS,cAAc,MAAa,OAAsC;AACxE,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,UAAU,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC/F,SAAO,SAAS,CAAC;AACnB;AAEA,eAAe,iBAAiB,OAAmB,cAA8C;AAC/F,QAAM,KAAK,MAAM,WAAW,OAAO,YAAY;AAC/C,QAAM,OAAO,GAAG;AAAA,IACd,CAAC,MAAM,EAAE,YAAY,kBAAkB,EAAE,cAAc;AAAA,EACzD;AACA,MAAI,CAAC,KAAM,QAAO;AAElB,MAAI,KAAK,SAAS,KAAK,KAAK,SAAS,EAAG,QAAO,KAAK;AACpD,MAAI,KAAK,SAAS,KAAK,KAAK,SAAS,GAAI,QAAO,KAAK,QAAQ;AAC7D,SAAO;AACT;","names":[]}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/builder-eval/builder-session.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts","../../src/builder-eval/three-layer-eval.ts"],"sourcesContent":["/**\n * BuilderSession — ties a builder-of-builders workflow together.\n *\n * Models agent-builder's shape: Project → Chat → Edit → Ship → App →\n * AppAgent. Each layer is a Run (linked via parentRunId). The\n * framework-enforced invariants:\n *\n *   - One Project → many Chats; chatId scopes runs within a project.\n *   - One Chat = one builder Run with `layer='builder'`.\n *   - One Ship = one child Run with `layer='app-build'` + SandboxHarness.\n *   - One AppScenario = one grandchild Run with `layer='app-runtime'`.\n *\n * Consumers obtain a BuilderSession, call `startChat`, drive the\n * builder agent (emitting spans), and call `ship` / `runAppScenario`\n * as the workflow progresses. The session reconstructs itself from\n * trace data via `resume(store, projectId)`.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from '../sandbox-harness'\nimport { SandboxHarness } from '../sandbox-harness'\nimport type { TestGradedRunResult, TestGradedScenario } from '../test-graded-scenario'\nimport { runTestGradedScenario } from '../test-graded-scenario'\nimport { TraceEmitter } from '../trace/emitter'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BuilderSessionInit {\n  projectId: string\n  chatId?: string\n  /** Free-form: user's task description, project name, etc. Stored on the builder Run. */\n  tags?: Record<string, string>\n}\n\nexport interface ShipOptions {\n  harness: HarnessConfig\n  driver?: SandboxDriver\n  /** scenarioId of this app-build run. Defaults to `${projectId}/build`. */\n  scenarioId?: string\n}\n\nexport interface RunAppScenarioOptions {\n  scenario: TestGradedScenario\n  /** Harness driver override; defaults to the one the session was created with. */\n  driver?: SandboxDriver\n}\n\nexport class BuilderSession {\n  private store: TraceStore\n  private builderEmitter: TraceEmitter\n  readonly projectId: string\n  readonly chatId: string\n  private builderRunId?: string\n  private lastBuildRunId?: string\n  private defaultDriver?: SandboxDriver\n\n  constructor(store: TraceStore, init: BuilderSessionInit, driver?: SandboxDriver) {\n    this.store = store\n    this.projectId = init.projectId\n    this.chatId = init.chatId ?? cryptoId()\n    this.defaultDriver = driver\n    this.builderEmitter = new TraceEmitter(store)\n  }\n\n  /** Start the builder (L0) run for this chat. Returns the runId. */\n  async startChat(scenarioId = `${this.projectId}/chat`): Promise<string> {\n    await this.builderEmitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'builder',\n    })\n    this.builderRunId = this.builderEmitter.runId\n    return this.builderRunId\n  }\n\n  /** The emitter for builder-level spans (edits, LLM calls, tool invocations). */\n  get emitter(): TraceEmitter {\n    if (!this.builderRunId) throw new Error('BuilderSession.emitter: call startChat() first')\n    return this.builderEmitter\n  }\n\n  /**\n   * Ship the project's generated app: run the sandbox harness as a child\n   * Run (`layer='app-build'`). Returns the build result + runId.\n   */\n  async ship(options: ShipOptions): Promise<{ runId: string; result: SandboxHarnessResult }> {\n    if (!this.builderRunId) throw new Error('BuilderSession.ship: call startChat() first')\n    const buildEmitter = new TraceEmitter(this.store)\n    await buildEmitter.startRun({\n      scenarioId: options.scenarioId ?? `${this.projectId}/build`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    const harness = new SandboxHarness(options.driver ?? this.defaultDriver)\n    const result = await harness.run(options.harness, buildEmitter)\n    await buildEmitter.endRun({\n      pass: result.passed,\n      score: result.score,\n      failureClass: result.passed ? 'success' : 'sandbox_failure',\n    })\n    this.lastBuildRunId = buildEmitter.runId\n    return { runId: buildEmitter.runId, result }\n  }\n\n  /**\n   * Run a domain scenario against the just-built app as a grandchild Run\n   * (`layer='app-runtime'`). The `ship` call must precede this so the\n   * parent is set correctly; if no build exists yet the session attaches\n   * directly to the builder run (useful for prototypes).\n   */\n  async runAppScenario(options: RunAppScenarioOptions): Promise<TestGradedRunResult> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error('BuilderSession.runAppScenario: call startChat() + ship() first')\n    const { scenario, driver } = options\n    const result = await runTestGradedScenario(scenario, this.store, {\n      driver: driver ?? this.defaultDriver,\n      provenance: { codeSha: undefined, promptSha: undefined, modelFingerprint: undefined },\n    })\n    // Attach to the parent chain by updating the stored Run in place.\n    await this.store.updateRun(result.runId, {\n      parentRunId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'app-runtime',\n    })\n    return result\n  }\n\n  /** Record an end-of-chat meta score (judge verdict on whether the builder\n   *  served the user's intent). Accepts a numeric score + optional rationale. */\n  async recordMetaScore(score: number, rationale?: string): Promise<void> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordMetaScore: call startChat() first')\n    await this.builderEmitter.recordJudge({\n      judgeId: 'builder-meta',\n      targetSpanId: this.builderRunId, // attach to the builder run itself\n      dimension: 'user_intent_satisfaction',\n      score,\n      rationale,\n      name: 'builder-meta',\n    })\n  }\n\n  /** Close the builder Run with a final outcome. */\n  async endChat(outcome: { pass: boolean; score?: number; notes?: string }): Promise<void> {\n    await this.builderEmitter.endRun({\n      pass: outcome.pass,\n      score: outcome.score,\n      notes: outcome.notes,\n    })\n  }\n\n  /**\n   * Inline app-runtime run — for cases where the \"scenario\" isn't a\n   * SWE-bench-style test suite but a live agent interaction (LLM chat,\n   * domain flow). Returns an emitter bound to a fresh Run in the\n   * `app-runtime` layer; caller emits spans inside and calls\n   * `.endRun()` with the final verdict.\n   */\n  async startAppRuntime(scenarioId: string): Promise<TraceEmitter> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error(\n        'BuilderSession.startAppRuntime: call startChat() + (optionally) ship() first',\n      )\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId,\n      layer: 'app-runtime',\n    })\n    return emitter\n  }\n\n  /**\n   * Lightweight \"ship marker\" — record an app-build Run with a caller-\n   * provided verdict. Use when there isn't a sandbox harness to run but\n   * you still want to mark the build state at publish time.\n   */\n  async recordShipMarker(args: {\n    pass: boolean\n    score: number\n    scenarioId?: string\n    notes?: string\n  }): Promise<string> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordShipMarker: call startChat() first')\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId: args.scenarioId ?? `${this.projectId}/ship`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    await emitter.endRun({\n      pass: args.pass,\n      score: args.score,\n      failureClass: args.pass ? 'success' : 'sandbox_failure',\n      notes: args.notes,\n    })\n    this.lastBuildRunId = emitter.runId\n    return emitter.runId\n  }\n\n  get lastBuildRunIdValue(): string | undefined {\n    return this.lastBuildRunId\n  }\n  get builderRunIdValue(): string | undefined {\n    return this.builderRunId\n  }\n}\n\n/**\n * Reconstruct the most recent BuilderSession state for a given project —\n * returns { builderRunId, lastBuildRunId, chatRuns }. For chat-first UIs\n * this is how a resumed session finds its place in the edit history.\n */\nexport async function resumeBuilderSession(\n  store: TraceStore,\n  projectId: string,\n): Promise<{\n  projectId: string\n  chatRuns: Run[]\n  lastBuilderRun?: Run\n  lastBuildRun?: Run\n  lastAppRuntimeRuns: Run[]\n}> {\n  const runs = await store.listRuns({ projectId })\n  const chatRuns = runs\n    .filter((r) => r.layer === 'builder')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const buildRuns = runs\n    .filter((r) => r.layer === 'app-build')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const appRuntimeRuns = runs\n    .filter((r) => r.layer === 'app-runtime')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  return {\n    projectId,\n    chatRuns,\n    lastBuilderRun: chatRuns[0],\n    lastBuildRun: buildRuns[0],\n    lastAppRuntimeRuns: appRuntimeRuns,\n  }\n}\n\nfunction cryptoId(): string {\n  if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n","/**\n * Meta-eval correlation — the highest-leverage signal in the framework.\n *\n * Given a corpus of three-layer project reports, compute how well each\n * pair of layers correlates. The question we care about most:\n *\n *   Does `metaScore` (what the builder thinks it did) predict\n *   `runtimeScore` (what the user actually gets)?\n *\n * If r < ~0.4, the builder's self-scoring is broken — it's optimizing\n * for something other than real-world success. If r > 0.7, meta_score\n * is a usable proxy and can drive CI gates cheaply.\n *\n * Non-parametric rank correlation (Spearman) is also reported because\n * meta scores are often ordinal-ish.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { ThreeLayerProjectReport } from './three-layer-eval'\n\nexport interface LayerCorrelation {\n  n: number\n  pearson: number\n  spearman: number\n}\n\nexport interface CorrelationReport {\n  /** Pairs present in the corpus (layers with ≥ 2 matched data points). */\n  metaVsBuild?: LayerCorrelation\n  metaVsRuntime?: LayerCorrelation\n  buildVsRuntime?: LayerCorrelation\n  /** Number of complete projects (all 3 scores present). */\n  completeProjects: number\n}\n\nexport function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport {\n  const completeProjects = reports.filter((r) => r.complete).length\n  return {\n    metaVsBuild: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.buildScore,\n    ),\n    metaVsRuntime: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.runtimeScore,\n    ),\n    buildVsRuntime: pairwise(\n      reports,\n      (r) => r.buildScore,\n      (r) => r.runtimeScore,\n    ),\n    completeProjects,\n  }\n}\n\nfunction pairwise(\n  reports: ThreeLayerProjectReport[],\n  a: (r: ThreeLayerProjectReport) => number | null,\n  b: (r: ThreeLayerProjectReport) => number | null,\n): LayerCorrelation | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const r of reports) {\n    const x = a(r)\n    const y = b(r)\n    if (x !== null && y !== null && Number.isFinite(x) && Number.isFinite(y)) {\n      xs.push(x)\n      ys.push(y)\n    }\n  }\n  if (xs.length < 2) return undefined\n  return {\n    n: xs.length,\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n  }\n}\n","/**\n * ProjectRegistry — project-level aggregation over the trace corpus.\n *\n * Thin reader over TraceStore that answers the questions a chat-first,\n * resumable UI needs:\n *   - listProjects() → project IDs with latest activity\n *   - projectTimeline(id) → chats + builds + runtime runs, chronological\n *   - projectChats(id) → chat-level summaries (turn count, outcome)\n *\n * All queries are pure reads; no state duplication.\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ProjectSummary {\n  projectId: string\n  chatCount: number\n  buildCount: number\n  appRuntimeCount: number\n  lastActivityAt: number\n  latestChatId?: string\n  latestOutcome?: { pass: boolean; score?: number }\n}\n\nexport interface ChatSummary {\n  chatId: string\n  projectId: string\n  builderRunId: string\n  startedAt: number\n  endedAt?: number\n  status: Run['status']\n  outcome?: Run['outcome']\n  /** Counts of spans emitted during the chat. */\n  llmTurns?: number\n  toolCalls?: number\n  buildRunId?: string\n  appRuntimeRunIds: string[]\n}\n\nexport interface ProjectTimelineEntry {\n  run: Run\n  layerBucket: 'chat' | 'build' | 'runtime' | 'other'\n}\n\nexport class ProjectRegistry {\n  constructor(private store: TraceStore) {}\n\n  async listProjects(): Promise<ProjectSummary[]> {\n    const runs = await this.store.listRuns()\n    const byProject = new Map<string, Run[]>()\n    for (const r of runs) {\n      if (!r.projectId) continue\n      const arr = byProject.get(r.projectId) ?? []\n      arr.push(r)\n      byProject.set(r.projectId, arr)\n    }\n    const summaries: ProjectSummary[] = []\n    for (const [projectId, projectRuns] of byProject) {\n      const sorted = projectRuns.slice().sort((a, b) => b.startedAt - a.startedAt)\n      const chats = projectRuns.filter((r) => r.layer === 'builder')\n      const builds = projectRuns.filter((r) => r.layer === 'app-build')\n      const runtimes = projectRuns.filter((r) => r.layer === 'app-runtime')\n      const latest = sorted[0]\n      if (!latest) continue\n      summaries.push({\n        projectId,\n        chatCount: chats.length,\n        buildCount: builds.length,\n        appRuntimeCount: runtimes.length,\n        lastActivityAt: latest.startedAt,\n        latestChatId: chats[0]?.chatId,\n        latestOutcome: latest.outcome\n          ? { pass: latest.outcome.pass ?? false, score: latest.outcome.score }\n          : undefined,\n      })\n    }\n    return summaries.sort((a, b) => b.lastActivityAt - a.lastActivityAt)\n  }\n\n  async projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]> {\n    const runs = await this.store.listRuns({ projectId })\n    const ordered = runs.slice().sort((a, b) => a.startedAt - b.startedAt)\n    return ordered.map((run) => ({\n      run,\n      layerBucket:\n        run.layer === 'builder'\n          ? 'chat'\n          : run.layer === 'app-build'\n            ? 'build'\n            : run.layer === 'app-runtime'\n              ? 'runtime'\n              : 'other',\n    }))\n  }\n\n  async projectChats(projectId: string): Promise<ChatSummary[]> {\n    const builderRuns = (await this.store.listRuns({ projectId, layer: 'builder' })).sort(\n      (a, b) => b.startedAt - a.startedAt,\n    )\n    const childrenFor = async (runId: string) => this.store.listRuns({ parentRunId: runId })\n    const out: ChatSummary[] = []\n    for (const run of builderRuns) {\n      const spans = await this.store.spans({ runId: run.runId })\n      const children = await childrenFor(run.runId)\n      const build = children.find((c) => c.layer === 'app-build')\n      const runtime: string[] = []\n      // Runtime runs can be grandchildren (attached to the build) or siblings\n      // when shipped skipped.\n      if (build) {\n        const grands = await childrenFor(build.runId)\n        for (const g of grands) if (g.layer === 'app-runtime') runtime.push(g.runId)\n      }\n      for (const c of children) if (c.layer === 'app-runtime') runtime.push(c.runId)\n      out.push({\n        chatId: run.chatId ?? run.runId,\n        projectId,\n        builderRunId: run.runId,\n        startedAt: run.startedAt,\n        endedAt: run.endedAt,\n        status: run.status,\n        outcome: run.outcome,\n        llmTurns: spans.filter((s) => s.kind === 'llm').length,\n        toolCalls: spans.filter((s) => s.kind === 'tool').length,\n        buildRunId: build?.runId,\n        appRuntimeRunIds: runtime,\n      })\n    }\n    return out\n  }\n}\n","/**\n * Three-layer evaluation — the canonical scoring breakdown for\n * builder-of-builders workflows.\n *\n *   meta_score:    did the builder understand + satisfy user intent?\n *                  (judge verdict attached to the builder run)\n *   build_score:   did the generated scaffold build + pass its own tests?\n *                  (outcome.score on the app-build child run)\n *   runtime_score: did the generated agent pass its domain scenarios?\n *                  (mean outcome.score over app-runtime grandchild runs)\n *\n * Returns a structured report per project. The cross-layer correlation\n * is the highest-leverage signal the framework computes — if\n * meta_score doesn't predict runtime_score, the builder's self-scoring\n * is broken.\n *\n * Scaffold-only mode: when a project has no `app-runtime` runs (e.g. a\n * scaffold-builder eval that grades compose + build without driving a\n * runtime scenario), `kind` is `'scaffold-only'` and `complete` measures\n * meta + build only. Consumers can tell the two apart without having to\n * interpret null-runtime as either \"not yet computed\" or \"N/A for this\n * project shape\".\n */\n\nimport { judgeSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport type ProjectKind = 'full' | 'scaffold-only'\n\nexport interface ThreeLayerProjectReport {\n  projectId: string\n  /**\n   * `'full'` when the project has at least one `app-runtime` run;\n   * `'scaffold-only'` when it only has meta + build layers. Lets\n   * downstream consumers treat a null runtime score as expected\n   * (scaffold-only) vs. missing (full, pipeline broke).\n   */\n  kind: ProjectKind\n  builderRunId?: string\n  /** Judge-verdict score on the builder run (0..1 after normalization). */\n  metaScore: number | null\n  buildRunId?: string\n  /** 0..1 from the sandbox harness (testsPassed / testsTotal). */\n  buildScore: number | null\n  appRuntimeRunIds: string[]\n  /** Mean of outcome.score over app-runtime runs, 0..1. Always null in scaffold-only mode. */\n  runtimeScore: number | null\n  runtimePassRate: number | null\n  /**\n   * Layer-aware completeness:\n   *   - `kind='full'`: all three layers scored\n   *   - `kind='scaffold-only'`: meta + build scored (runtime not applicable)\n   */\n  complete: boolean\n}\n\nexport async function scoreProject(\n  store: TraceStore,\n  projectId: string,\n): Promise<ThreeLayerProjectReport> {\n  const allRuns = await store.listRuns({ projectId })\n  const builder = latestByLayer(allRuns, 'builder')\n  const build = latestByLayer(allRuns, 'app-build')\n  const runtime = allRuns.filter((r) => r.layer === 'app-runtime')\n\n  const metaScore = builder ? await extractMetaScore(store, builder.runId) : null\n  const buildScore = build?.outcome?.score ?? null\n  const runtimeScores = runtime\n    .map((r) => r.outcome?.score)\n    .filter((s): s is number => typeof s === 'number')\n  const runtimeScore =\n    runtimeScores.length > 0\n      ? runtimeScores.reduce((a, b) => a + b, 0) / runtimeScores.length\n      : null\n  const runtimePassed = runtime.filter((r) => r.outcome?.pass === true).length\n  const runtimePassRate = runtime.length > 0 ? runtimePassed / runtime.length : null\n\n  const kind: ProjectKind = runtime.length === 0 ? 'scaffold-only' : 'full'\n  const complete =\n    kind === 'scaffold-only'\n      ? metaScore !== null && buildScore !== null\n      : metaScore !== null && buildScore !== null && runtimeScore !== null\n\n  return {\n    projectId,\n    kind,\n    builderRunId: builder?.runId,\n    metaScore,\n    buildRunId: build?.runId,\n    buildScore,\n    appRuntimeRunIds: runtime.map((r) => r.runId),\n    runtimeScore,\n    runtimePassRate,\n    complete,\n  }\n}\n\n/** Aggregate scoring across every project in a corpus. */\nexport async function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]> {\n  const runs = await store.listRuns()\n  const projectIds = [...new Set(runs.map((r) => r.projectId).filter((p): p is string => !!p))]\n  return Promise.all(projectIds.map((p) => scoreProject(store, p)))\n}\n\nfunction latestByLayer(runs: Run[], layer: Run['layer']): Run | undefined {\n  const filtered = runs.filter((r) => r.layer === layer).sort((a, b) => b.startedAt - a.startedAt)\n  return filtered[0]\n}\n\nasync function extractMetaScore(store: TraceStore, builderRunId: string): Promise<number | null> {\n  const js = await judgeSpans(store, builderRunId)\n  const meta = js.find(\n    (s) => s.judgeId === 'builder-meta' && s.dimension === 'user_intent_satisfaction',\n  )\n  if (!meta) return null\n  // Normalize score to 0..1. Accept 0-1 natively; 0-10 scale is also common.\n  if (meta.score >= 0 && meta.score <= 1) return meta.score\n  if (meta.score >= 0 && meta.score <= 10) return meta.score / 10\n  return null\n}\n"],"mappings":";;;;;AA8CA,IAAa,iBAAb,MAA4B;CAC1B;CACA;CACA;CACA;CACA;CACA;CACA;CAEA,YAAY,OAAmB,MAA0B,QAAwB;EAC/E,KAAK,QAAQ;EACb,KAAK,YAAY,KAAK;EACtB,KAAK,SAAS,KAAK,UAAU,SAAS;EACtC,KAAK,gBAAgB;EACrB,KAAK,iBAAiB,IAAI,aAAa,KAAK;CAC9C;;CAGA,MAAM,UAAU,aAAa,GAAG,KAAK,UAAU,QAAyB;EACtE,MAAM,KAAK,eAAe,SAAS;GACjC;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,OAAO;EACT,CAAC;EACD,KAAK,eAAe,KAAK,eAAe;EACxC,OAAO,KAAK;CACd;;CAGA,IAAI,UAAwB;EAC1B,IAAI,CAAC,KAAK,cAAc,MAAM,IAAI,MAAM,gDAAgD;EACxF,OAAO,KAAK;CACd;;;;;CAMA,MAAM,KAAK,SAAgF;EACzF,IAAI,CAAC,KAAK,cAAc,MAAM,IAAI,MAAM,6CAA6C;EACrF,MAAM,eAAe,IAAI,aAAa,KAAK,KAAK;EAChD,MAAM,aAAa,SAAS;GAC1B,YAAY,QAAQ,cAAc,GAAG,KAAK,UAAU;GACpD,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,aAAa,KAAK;GAClB,OAAO;EACT,CAAC;EAED,MAAM,SAAS,MAAM,IADD,eAAe,QAAQ,UAAU,KAAK,aAC/B,CAAC,CAAC,IAAI,QAAQ,SAAS,YAAY;EAC9D,MAAM,aAAa,OAAO;GACxB,MAAM,OAAO;GACb,OAAO,OAAO;GACd,cAAc,OAAO,SAAS,YAAY;EAC5C,CAAC;EACD,KAAK,iBAAiB,aAAa;EACnC,OAAO;GAAE,OAAO,aAAa;GAAO;EAAO;CAC7C;;;;;;;CAQA,MAAM,eAAe,SAA8D;EACjF,MAAM,cAAc,KAAK,kBAAkB,KAAK;EAChD,IAAI,CAAC,aACH,MAAM,IAAI,MAAM,gEAAgE;EAClF,MAAM,EAAE,UAAU,WAAW;EAC7B,MAAM,SAAS,MAAM,sBAAsB,UAAU,KAAK,OAAO;GAC/D,QAAQ,UAAU,KAAK;GACvB,YAAY;IAAE,SAAS,KAAA;IAAW,WAAW,KAAA;IAAW,kBAAkB,KAAA;GAAU;EACtF,CAAC;EAED,MAAM,KAAK,MAAM,UAAU,OAAO,OAAO;GACvC;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,OAAO;EACT,CAAC;EACD,OAAO;CACT;;;CAIA,MAAM,gBAAgB,OAAe,WAAmC;EACtE,IAAI,CAAC,KAAK,cACR,MAAM,IAAI,MAAM,wDAAwD;EAC1E,MAAM,KAAK,eAAe,YAAY;GACpC,SAAS;GACT,cAAc,KAAK;GACnB,WAAW;GACX;GACA;GACA,MAAM;EACR,CAAC;CACH;;CAGA,MAAM,QAAQ,SAA2E;EACvF,MAAM,KAAK,eAAe,OAAO;GAC/B,MAAM,QAAQ;GACd,OAAO,QAAQ;GACf,OAAO,QAAQ;EACjB,CAAC;CACH;;;;;;;;CASA,MAAM,gBAAgB,YAA2C;EAC/D,MAAM,cAAc,KAAK,kBAAkB,KAAK;EAChD,IAAI,CAAC,aACH,MAAM,IAAI,MACR,8EACF;EACF,MAAM,UAAU,IAAI,aAAa,KAAK,KAAK;EAC3C,MAAM,QAAQ,SAAS;GACrB;GACA,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb;GACA,OAAO;EACT,CAAC;EACD,OAAO;CACT;;;;;;CAOA,MAAM,iBAAiB,MAKH;EAClB,IAAI,CAAC,KAAK,cACR,MAAM,IAAI,MAAM,yDAAyD;EAC3E,MAAM,UAAU,IAAI,aAAa,KAAK,KAAK;EAC3C,MAAM,QAAQ,SAAS;GACrB,YAAY,KAAK,cAAc,GAAG,KAAK,UAAU;GACjD,WAAW,KAAK;GAChB,QAAQ,KAAK;GACb,aAAa,KAAK;GAClB,OAAO;EACT,CAAC;EACD,MAAM,QAAQ,OAAO;GACnB,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,cAAc,KAAK,OAAO,YAAY;GACtC,OAAO,KAAK;EACd,CAAC;EACD,KAAK,iBAAiB,QAAQ;EAC9B,OAAO,QAAQ;CACjB;CAEA,IAAI,sBAA0C;EAC5C,OAAO,KAAK;CACd;CACA,IAAI,oBAAwC;EAC1C,OAAO,KAAK;CACd;AACF;;;;;;AAOA,eAAsB,qBACpB,OACA,WAOC;CACD,MAAM,OAAO,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;CAC/C,MAAM,WAAW,KACd,QAAQ,MAAM,EAAE,UAAU,SAAS,CAAC,CACpC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,MAAM,YAAY,KACf,QAAQ,MAAM,EAAE,UAAU,WAAW,CAAC,CACtC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,MAAM,iBAAiB,KACpB,QAAQ,MAAM,EAAE,UAAU,aAAa,CAAC,CACxC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;CAC3C,OAAO;EACL;EACA;EACA,gBAAgB,SAAS;EACzB,cAAc,UAAU;EACxB,oBAAoB;CACtB;AACF;AAEA,SAAS,WAAmB;CAC1B,IAAI,OAAO,WAAW,QAAQ,eAAe,YAAY,OAAO,WAAW,OAAO,WAAW;CAC7F,OAAO,GAAG,KAAK,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,GAAG,KAAK,OAAO,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,MAAM,GAAG,EAAE;AAC7E;;;;;;;;;;;;;;;;;;;AC5NA,SAAgB,gBAAgB,SAAuD;CACrF,MAAM,mBAAmB,QAAQ,QAAQ,MAAM,EAAE,QAAQ,CAAC,CAAC;CAC3D,OAAO;EACL,aAAa,SACX,UACC,MAAM,EAAE,YACR,MAAM,EAAE,UACX;EACA,eAAe,SACb,UACC,MAAM,EAAE,YACR,MAAM,EAAE,YACX;EACA,gBAAgB,SACd,UACC,MAAM,EAAE,aACR,MAAM,EAAE,YACX;EACA;CACF;AACF;AAEA,SAAS,SACP,SACA,GACA,GAC8B;CAC9B,MAAM,KAAe,CAAC;CACtB,MAAM,KAAe,CAAC;CACtB,KAAK,MAAM,KAAK,SAAS;EACvB,MAAM,IAAI,EAAE,CAAC;EACb,MAAM,IAAI,EAAE,CAAC;EACb,IAAI,MAAM,QAAQ,MAAM,QAAQ,OAAO,SAAS,CAAC,KAAK,OAAO,SAAS,CAAC,GAAG;GACxE,GAAG,KAAK,CAAC;GACT,GAAG,KAAK,CAAC;EACX;CACF;CACA,IAAI,GAAG,SAAS,GAAG,OAAO,KAAA;CAC1B,OAAO;EACL,GAAG,GAAG;EACN,SAAS,SAAS,IAAI,EAAE;EACxB,UAAU,UAAU,IAAI,EAAE;CAC5B;AACF;;;ACjCA,IAAa,kBAAb,MAA6B;CACP;CAApB,YAAY,OAA2B;EAAnB,KAAA,QAAA;CAAoB;CAExC,MAAM,eAA0C;EAC9C,MAAM,OAAO,MAAM,KAAK,MAAM,SAAS;EACvC,MAAM,4BAAY,IAAI,IAAmB;EACzC,KAAK,MAAM,KAAK,MAAM;GACpB,IAAI,CAAC,EAAE,WAAW;GAClB,MAAM,MAAM,UAAU,IAAI,EAAE,SAAS,KAAK,CAAC;GAC3C,IAAI,KAAK,CAAC;GACV,UAAU,IAAI,EAAE,WAAW,GAAG;EAChC;EACA,MAAM,YAA8B,CAAC;EACrC,KAAK,MAAM,CAAC,WAAW,gBAAgB,WAAW;GAChD,MAAM,SAAS,YAAY,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;GAC3E,MAAM,QAAQ,YAAY,QAAQ,MAAM,EAAE,UAAU,SAAS;GAC7D,MAAM,SAAS,YAAY,QAAQ,MAAM,EAAE,UAAU,WAAW;GAChE,MAAM,WAAW,YAAY,QAAQ,MAAM,EAAE,UAAU,aAAa;GACpE,MAAM,SAAS,OAAO;GACtB,IAAI,CAAC,QAAQ;GACb,UAAU,KAAK;IACb;IACA,WAAW,MAAM;IACjB,YAAY,OAAO;IACnB,iBAAiB,SAAS;IAC1B,gBAAgB,OAAO;IACvB,cAAc,MAAM,EAAE,EAAE;IACxB,eAAe,OAAO,UAClB;KAAE,MAAM,OAAO,QAAQ,QAAQ;KAAO,OAAO,OAAO,QAAQ;IAAM,IAClE,KAAA;GACN,CAAC;EACH;EACA,OAAO,UAAU,MAAM,GAAG,MAAM,EAAE,iBAAiB,EAAE,cAAc;CACrE;CAEA,MAAM,gBAAgB,WAAoD;EAGxE,QADgB,MADG,KAAK,MAAM,SAAS,EAAE,UAAU,CAAC,EAAA,CAC/B,MAAM,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAC/C,CAAC,CAAC,KAAK,SAAS;GAC3B;GACA,aACE,IAAI,UAAU,YACV,SACA,IAAI,UAAU,cACZ,UACA,IAAI,UAAU,gBACZ,YACA;EACZ,EAAE;CACJ;CAEA,MAAM,aAAa,WAA2C;EAC5D,MAAM,eAAe,MAAM,KAAK,MAAM,SAAS;GAAE;GAAW,OAAO;EAAU,CAAC,EAAA,CAAG,MAC9E,GAAG,MAAM,EAAE,YAAY,EAAE,SAC5B;EACA,MAAM,cAAc,OAAO,UAAkB,KAAK,MAAM,SAAS,EAAE,aAAa,MAAM,CAAC;EACvF,MAAM,MAAqB,CAAC;EAC5B,KAAK,MAAM,OAAO,aAAa;GAC7B,MAAM,QAAQ,MAAM,KAAK,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;GACzD,MAAM,WAAW,MAAM,YAAY,IAAI,KAAK;GAC5C,MAAM,QAAQ,SAAS,MAAM,MAAM,EAAE,UAAU,WAAW;GAC1D,MAAM,UAAoB,CAAC;GAG3B,IAAI,OAAO;IACT,MAAM,SAAS,MAAM,YAAY,MAAM,KAAK;IAC5C,KAAK,MAAM,KAAK,QAAQ,IAAI,EAAE,UAAU,eAAe,QAAQ,KAAK,EAAE,KAAK;GAC7E;GACA,KAAK,MAAM,KAAK,UAAU,IAAI,EAAE,UAAU,eAAe,QAAQ,KAAK,EAAE,KAAK;GAC7E,IAAI,KAAK;IACP,QAAQ,IAAI,UAAU,IAAI;IAC1B;IACA,cAAc,IAAI;IAClB,WAAW,IAAI;IACf,SAAS,IAAI;IACb,QAAQ,IAAI;IACZ,SAAS,IAAI;IACb,UAAU,MAAM,QAAQ,MAAM,EAAE,SAAS,KAAK,CAAC,CAAC;IAChD,WAAW,MAAM,QAAQ,MAAM,EAAE,SAAS,MAAM,CAAC,CAAC;IAClD,YAAY,OAAO;IACnB,kBAAkB;GACpB,CAAC;EACH;EACA,OAAO;CACT;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;ACzEA,eAAsB,aACpB,OACA,WACkC;CAClC,MAAM,UAAU,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;CAClD,MAAM,UAAU,cAAc,SAAS,SAAS;CAChD,MAAM,QAAQ,cAAc,SAAS,WAAW;CAChD,MAAM,UAAU,QAAQ,QAAQ,MAAM,EAAE,UAAU,aAAa;CAE/D,MAAM,YAAY,UAAU,MAAM,iBAAiB,OAAO,QAAQ,KAAK,IAAI;CAC3E,MAAM,aAAa,OAAO,SAAS,SAAS;CAC5C,MAAM,gBAAgB,QACnB,KAAK,MAAM,EAAE,SAAS,KAAK,CAAC,CAC5B,QAAQ,MAAmB,OAAO,MAAM,QAAQ;CACnD,MAAM,eACJ,cAAc,SAAS,IACnB,cAAc,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc,SACzD;CACN,MAAM,gBAAgB,QAAQ,QAAQ,MAAM,EAAE,SAAS,SAAS,IAAI,CAAC,CAAC;CACtE,MAAM,kBAAkB,QAAQ,SAAS,IAAI,gBAAgB,QAAQ,SAAS;CAE9E,MAAM,OAAoB,QAAQ,WAAW,IAAI,kBAAkB;CACnE,MAAM,WACJ,SAAS,kBACL,cAAc,QAAQ,eAAe,OACrC,cAAc,QAAQ,eAAe,QAAQ,iBAAiB;CAEpE,OAAO;EACL;EACA;EACA,cAAc,SAAS;EACvB;EACA,YAAY,OAAO;EACnB;EACA,kBAAkB,QAAQ,KAAK,MAAM,EAAE,KAAK;EAC5C;EACA;EACA;CACF;AACF;;AAGA,eAAsB,iBAAiB,OAAuD;CAC5F,MAAM,OAAO,MAAM,MAAM,SAAS;CAClC,MAAM,aAAa,CAAC,GAAG,IAAI,IAAI,KAAK,KAAK,MAAM,EAAE,SAAS,CAAC,CAAC,QAAQ,MAAmB,CAAC,CAAC,CAAC,CAAC,CAAC;CAC5F,OAAO,QAAQ,IAAI,WAAW,KAAK,MAAM,aAAa,OAAO,CAAC,CAAC,CAAC;AAClE;AAEA,SAAS,cAAc,MAAa,OAAsC;CAExE,OADiB,KAAK,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SACxE,CAAC,CAAC;AAClB;AAEA,eAAe,iBAAiB,OAAmB,cAA8C;CAE/F,MAAM,QAAO,MADI,WAAW,OAAO,YAAY,EAAA,CAC/B,MACb,MAAM,EAAE,YAAY,kBAAkB,EAAE,cAAc,0BACzD;CACA,IAAI,CAAC,MAAM,OAAO;CAElB,IAAI,KAAK,SAAS,KAAK,KAAK,SAAS,GAAG,OAAO,KAAK;CACpD,IAAI,KAAK,SAAS,KAAK,KAAK,SAAS,IAAI,OAAO,KAAK,QAAQ;CAC7D,OAAO;AACT"}

package/dist/calibration-CNWWA6K8.js

@@ -0,0 +1,94 @@
+//#region src/meta-eval/calibration.ts
+async function calibrationCurve(traceStore, outcomeStore, evalMetric, outcomeMetric, options = {}) {
+	const runs = await traceStore.listRuns();
+	const outcomes = await outcomeStore.list();
+	const byRun = /* @__PURE__ */ new Map();
+	for (const o of outcomes) {
+		const arr = byRun.get(o.runId) ?? [];
+		arr.push(o);
+		byRun.set(o.runId, arr);
+	}
+	const extract = evalMetric.extract ?? defaultExtract(evalMetric.id);
+	const pairs = [];
+	for (const run of runs) {
+		const os = byRun.get(run.runId);
+		if (!os?.length) continue;
+		const x = await extract(run, traceStore);
+		if (x === null || !Number.isFinite(x)) continue;
+		const y = [...os].sort((a, b) => b.capturedAt - a.capturedAt)[0].metrics[outcomeMetric];
+		if (typeof y !== "number" || !Number.isFinite(y)) continue;
+		pairs.push({
+			x,
+			y
+		});
+	}
+	if (pairs.length < 2) return null;
+	return calibrationFromPairs(pairs.map((p) => ({
+		evalScore: p.x,
+		outcome: p.y
+	})), evalMetric.id, outcomeMetric, options);
+}
+function calibrationFromPairs(inputPairs, evalMetric, outcomeMetric, options = {}) {
+	const pairs = inputPairs.filter((pair) => Number.isFinite(pair.evalScore) && Number.isFinite(pair.outcome));
+	if (pairs.length < 2) return null;
+	const numBins = options.bins ?? 10;
+	const binning = options.binning ?? "equal-width";
+	const xs = pairs.map((p) => p.evalScore);
+	const lo = options.range?.lo ?? Math.min(...xs);
+	const hi = options.range?.hi ?? Math.max(...xs);
+	const bins = [];
+	if (binning === "equal-frequency") {
+		const sorted = [...pairs].sort((a, b) => a.evalScore - b.evalScore);
+		const perBin = Math.max(1, Math.floor(sorted.length / numBins));
+		for (let i = 0; i < sorted.length; i += perBin) {
+			const chunk = sorted.slice(i, i + perBin);
+			if (chunk.length === 0) continue;
+			bins.push(toBin(chunk));
+		}
+	} else {
+		const width = (hi - lo) / numBins;
+		if (width === 0) return null;
+		for (let i = 0; i < numBins; i++) {
+			const binLo = lo + i * width;
+			const binHi = i === numBins - 1 ? hi + 1e-9 : lo + (i + 1) * width;
+			const chunk = pairs.filter((p) => p.evalScore >= binLo && p.evalScore < binHi);
+			if (chunk.length === 0) continue;
+			bins.push(toBin(chunk, binLo, binHi));
+		}
+	}
+	const total = bins.reduce((a, b) => a + b.n, 0);
+	const ece = bins.reduce((a, b) => a + b.n / total * b.gap, 0);
+	const maxGap = bins.reduce((a, b) => Math.max(a, b.gap), 0);
+	return {
+		evalMetric,
+		outcomeMetric,
+		n: pairs.length,
+		bins,
+		ece,
+		maxGap
+	};
+}
+function toBin(chunk, lower, upper) {
+	const xs = chunk.map((c) => c.evalScore);
+	const ys = chunk.map((c) => c.outcome);
+	const evalMean = mean(xs);
+	const outcomeMean = mean(ys);
+	return {
+		lower: lower ?? Math.min(...xs),
+		upper: upper ?? Math.max(...xs),
+		n: chunk.length,
+		evalMean,
+		outcomeMean,
+		gap: Math.abs(outcomeMean - evalMean)
+	};
+}
+function mean(xs) {
+	return xs.reduce((a, b) => a + b, 0) / xs.length;
+}
+function defaultExtract(metric) {
+	return async (run) => run.outcome?.score ?? (metric === "pass" ? run.outcome?.pass === true ? 1 : 0 : null);
+}
+//#endregion
+export { calibrationFromPairs as n, calibrationCurve as t };
+
+//# sourceMappingURL=calibration-CNWWA6K8.js.map

package/dist/calibration-CNWWA6K8.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"calibration-CNWWA6K8.js","names":[],"sources":["../src/meta-eval/calibration.ts"],"sourcesContent":["/**\n * Calibration curve — binned \"if eval says X, what does reality show?\"\n *\n * Companion to correlationStudy. Raw correlation is a single number;\n * the calibration curve shows *where* the eval is well-calibrated vs\n * overconfident / underconfident. Buckets the eval metric, computes\n * mean outcome per bucket, reports expected-calibration-error (ECE).\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport type { EvalMetricSpec } from './correlation-study'\nimport type { DeploymentOutcome, OutcomeStore } from './outcome-store'\n\nexport interface CalibrationBin {\n  lower: number\n  upper: number\n  n: number\n  evalMean: number\n  outcomeMean: number\n  /** |outcomeMean − evalMean|; contributes to ECE weighted by n/total. */\n  gap: number\n}\n\nexport interface CalibrationReport {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  bins: CalibrationBin[]\n  /** Expected Calibration Error — Σ (n_i/N) × |outcomeMean_i − evalMean_i|. */\n  ece: number\n  /** Max bin gap — upper bound on miscalibration. */\n  maxGap: number\n}\n\nexport interface CalibrationOptions {\n  bins?: number\n  /** Equal-width (fixed bin edges) or equal-frequency (quantile bins). */\n  binning?: 'equal-width' | 'equal-frequency'\n  /** Clip eval values to [lo, hi] before binning. */\n  range?: { lo: number; hi: number }\n}\n\nexport interface CalibrationPair {\n  evalScore: number\n  outcome: number\n}\n\nexport async function calibrationCurve(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetric: EvalMetricSpec,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): Promise<CalibrationReport | null> {\n  const runs = await traceStore.listRuns()\n  const outcomes = await outcomeStore.list()\n  const byRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = byRun.get(o.runId) ?? []\n    arr.push(o)\n    byRun.set(o.runId, arr)\n  }\n\n  const extract = evalMetric.extract ?? defaultExtract(evalMetric.id)\n  const pairs: Array<{ x: number; y: number }> = []\n  for (const run of runs) {\n    const os = byRun.get(run.runId)\n    if (!os?.length) continue\n    const x = await extract(run, traceStore)\n    if (x === null || !Number.isFinite(x)) continue\n    const latest = [...os].sort((a, b) => b.capturedAt - a.capturedAt)[0]!\n    const y = latest.metrics[outcomeMetric]\n    if (typeof y !== 'number' || !Number.isFinite(y)) continue\n    pairs.push({ x, y })\n  }\n  if (pairs.length < 2) return null\n\n  return calibrationFromPairs(\n    pairs.map((p) => ({ evalScore: p.x, outcome: p.y })),\n    evalMetric.id,\n    outcomeMetric,\n    options,\n  )\n}\n\nexport function calibrationFromPairs(\n  inputPairs: CalibrationPair[],\n  evalMetric: string,\n  outcomeMetric: string,\n  options: CalibrationOptions = {},\n): CalibrationReport | null {\n  const pairs = inputPairs.filter(\n    (pair) => Number.isFinite(pair.evalScore) && Number.isFinite(pair.outcome),\n  )\n  if (pairs.length < 2) return null\n\n  const numBins = options.bins ?? 10\n  const binning = options.binning ?? 'equal-width'\n  const xs = pairs.map((p) => p.evalScore)\n  const lo = options.range?.lo ?? Math.min(...xs)\n  const hi = options.range?.hi ?? Math.max(...xs)\n\n  const bins: CalibrationBin[] = []\n  if (binning === 'equal-frequency') {\n    const sorted = [...pairs].sort((a, b) => a.evalScore - b.evalScore)\n    const perBin = Math.max(1, Math.floor(sorted.length / numBins))\n    for (let i = 0; i < sorted.length; i += perBin) {\n      const chunk = sorted.slice(i, i + perBin)\n      if (chunk.length === 0) continue\n      bins.push(toBin(chunk))\n    }\n  } else {\n    const width = (hi - lo) / numBins\n    if (width === 0) return null\n    for (let i = 0; i < numBins; i++) {\n      const binLo = lo + i * width\n      const binHi = i === numBins - 1 ? hi + 1e-9 : lo + (i + 1) * width\n      const chunk = pairs.filter((p) => p.evalScore >= binLo && p.evalScore < binHi)\n      if (chunk.length === 0) continue\n      bins.push(toBin(chunk, binLo, binHi))\n    }\n  }\n\n  const total = bins.reduce((a, b) => a + b.n, 0)\n  const ece = bins.reduce((a, b) => a + (b.n / total) * b.gap, 0)\n  const maxGap = bins.reduce((a, b) => Math.max(a, b.gap), 0)\n\n  return { evalMetric, outcomeMetric, n: pairs.length, bins, ece, maxGap }\n}\n\nfunction toBin(chunk: CalibrationPair[], lower?: number, upper?: number): CalibrationBin {\n  const xs = chunk.map((c) => c.evalScore)\n  const ys = chunk.map((c) => c.outcome)\n  const evalMean = mean(xs)\n  const outcomeMean = mean(ys)\n  return {\n    lower: lower ?? Math.min(...xs),\n    upper: upper ?? Math.max(...xs),\n    n: chunk.length,\n    evalMean,\n    outcomeMean,\n    gap: Math.abs(outcomeMean - evalMean),\n  }\n}\n\nfunction mean(xs: number[]): number {\n  return xs.reduce((a, b) => a + b, 0) / xs.length\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run) =>\n    run.outcome?.score ?? (metric === 'pass' ? (run.outcome?.pass === true ? 1 : 0) : null)\n}\n"],"mappings":";AAgDA,eAAsB,iBACpB,YACA,cACA,YACA,eACA,UAA8B,CAAC,GACI;CACnC,MAAM,OAAO,MAAM,WAAW,SAAS;CACvC,MAAM,WAAW,MAAM,aAAa,KAAK;CACzC,MAAM,wBAAQ,IAAI,IAAiC;CACnD,KAAK,MAAM,KAAK,UAAU;EACxB,MAAM,MAAM,MAAM,IAAI,EAAE,KAAK,KAAK,CAAC;EACnC,IAAI,KAAK,CAAC;EACV,MAAM,IAAI,EAAE,OAAO,GAAG;CACxB;CAEA,MAAM,UAAU,WAAW,WAAW,eAAe,WAAW,EAAE;CAClE,MAAM,QAAyC,CAAC;CAChD,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,KAAK,MAAM,IAAI,IAAI,KAAK;EAC9B,IAAI,CAAC,IAAI,QAAQ;EACjB,MAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;EACvC,IAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,GAAG;EAEvC,MAAM,IADS,CAAC,GAAG,EAAE,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,CAAC,CAAC,EACnD,CAAC,QAAQ;EACzB,IAAI,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,GAAG;EAClD,MAAM,KAAK;GAAE;GAAG;EAAE,CAAC;CACrB;CACA,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,OAAO,qBACL,MAAM,KAAK,OAAO;EAAE,WAAW,EAAE;EAAG,SAAS,EAAE;CAAE,EAAE,GACnD,WAAW,IACX,eACA,OACF;AACF;AAEA,SAAgB,qBACd,YACA,YACA,eACA,UAA8B,CAAC,GACL;CAC1B,MAAM,QAAQ,WAAW,QACtB,SAAS,OAAO,SAAS,KAAK,SAAS,KAAK,OAAO,SAAS,KAAK,OAAO,CAC3E;CACA,IAAI,MAAM,SAAS,GAAG,OAAO;CAE7B,MAAM,UAAU,QAAQ,QAAQ;CAChC,MAAM,UAAU,QAAQ,WAAW;CACnC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAC9C,MAAM,KAAK,QAAQ,OAAO,MAAM,KAAK,IAAI,GAAG,EAAE;CAE9C,MAAM,OAAyB,CAAC;CAChC,IAAI,YAAY,mBAAmB;EACjC,MAAM,SAAS,CAAC,GAAG,KAAK,CAAC,CAAC,MAAM,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;EAClE,MAAM,SAAS,KAAK,IAAI,GAAG,KAAK,MAAM,OAAO,SAAS,OAAO,CAAC;EAC9D,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK,QAAQ;GAC9C,MAAM,QAAQ,OAAO,MAAM,GAAG,IAAI,MAAM;GACxC,IAAI,MAAM,WAAW,GAAG;GACxB,KAAK,KAAK,MAAM,KAAK,CAAC;EACxB;CACF,OAAO;EACL,MAAM,SAAS,KAAK,MAAM;EAC1B,IAAI,UAAU,GAAG,OAAO;EACxB,KAAK,IAAI,IAAI,GAAG,IAAI,SAAS,KAAK;GAChC,MAAM,QAAQ,KAAK,IAAI;GACvB,MAAM,QAAQ,MAAM,UAAU,IAAI,KAAK,OAAO,MAAM,IAAI,KAAK;GAC7D,MAAM,QAAQ,MAAM,QAAQ,MAAM,EAAE,aAAa,SAAS,EAAE,YAAY,KAAK;GAC7E,IAAI,MAAM,WAAW,GAAG;GACxB,KAAK,KAAK,MAAM,OAAO,OAAO,KAAK,CAAC;EACtC;CACF;CAEA,MAAM,QAAQ,KAAK,QAAQ,GAAG,MAAM,IAAI,EAAE,GAAG,CAAC;CAC9C,MAAM,MAAM,KAAK,QAAQ,GAAG,MAAM,IAAK,EAAE,IAAI,QAAS,EAAE,KAAK,CAAC;CAC9D,MAAM,SAAS,KAAK,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,GAAG,GAAG,CAAC;CAE1D,OAAO;EAAE;EAAY;EAAe,GAAG,MAAM;EAAQ;EAAM;EAAK;CAAO;AACzE;AAEA,SAAS,MAAM,OAA0B,OAAgB,OAAgC;CACvF,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,SAAS;CACvC,MAAM,KAAK,MAAM,KAAK,MAAM,EAAE,OAAO;CACrC,MAAM,WAAW,KAAK,EAAE;CACxB,MAAM,cAAc,KAAK,EAAE;CAC3B,OAAO;EACL,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,OAAO,SAAS,KAAK,IAAI,GAAG,EAAE;EAC9B,GAAG,MAAM;EACT;EACA;EACA,KAAK,KAAK,IAAI,cAAc,QAAQ;CACtC;AACF;AAEA,SAAS,KAAK,IAAsB;CAClC,OAAO,GAAG,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAAS,eAAe,QAAyE;CAC/F,OAAO,OAAO,QACZ,IAAI,SAAS,UAAU,WAAW,SAAU,IAAI,SAAS,SAAS,OAAO,IAAI,IAAK;AACtF"}