@tangle-network/agent-eval 0.12.0 → 0.13.0

package/dist/chunk-OZPRSK4A.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/wire/schemas.ts","../src/wire/rubrics.ts","../src/wire/handlers.ts","../src/wire/openapi.ts","../src/wire/server.ts","../src/wire/rpc.ts"],"sourcesContent":["/**\n * Wire-protocol schemas.\n *\n * These Zod schemas are the contract between the agent-eval runtime and\n * any non-TypeScript client (Python, Rust, Go, …). They get rendered to\n * OpenAPI by `wire/openapi.ts` and code-generators consume that spec to\n * produce typed clients in other languages.\n *\n * Rule: if it's not in this file, it isn't on the wire. Keep names and\n * shapes self-explanatory — every field has a `.describe()` so the\n * generated docs are useful without reading the source.\n */\nimport { extendZodWithOpenApi } from '@asteasolutions/zod-to-openapi'\nimport { z } from 'zod'\n\nextendZodWithOpenApi(z)\n\n// ── Building blocks ─────────────────────────────────────────────────\n\nexport const RubricDimensionSchema = z\n  .object({\n    id: z\n      .string()\n      .min(1)\n      .describe('Short stable id like \"buyer_quality\" — used as the key in scoring output.'),\n    description: z\n      .string()\n      .min(1)\n      .describe('One-line plain-English meaning. Read by humans reviewing low scores.'),\n    weight: z\n      .number()\n      .min(0)\n      .default(1)\n      .describe('Relative weight in the composite score. Default 1; 0 disables.'),\n    min: z.number().default(0).describe('Lower bound of valid score for this dimension.'),\n    max: z.number().default(1).describe('Upper bound of valid score for this dimension.'),\n  })\n  .openapi('RubricDimension')\n\nexport const FailureModeSchema = z\n  .object({\n    id: z.string().min(1).describe('Short stable id like \"ai-cadence\" — used in detection lists.'),\n    description: z.string().min(1).describe('Plain-English description of the failure pattern.'),\n  })\n  .openapi('FailureMode')\n\n// ── Rubric ──────────────────────────────────────────────────────────\n\nexport const RubricSchema = z\n  .object({\n    name: z\n      .string()\n      .min(1)\n      .describe('Stable name like \"anti-slop\" — used by clients to invoke this rubric.'),\n    description: z\n      .string()\n      .min(1)\n      .describe('What this rubric measures. Shown in /v1/rubrics listing.'),\n    systemPrompt: z\n      .string()\n      .min(1)\n      .describe(\n        'Instructs the judging LLM. Should explain the persona (e.g. \"senior engineer reviewing voice\"), what to score on, and what to return.',\n      ),\n    dimensions: z\n      .array(RubricDimensionSchema)\n      .min(1)\n      .describe('Scoring axes. The composite score is a weighted sum of these.'),\n    failureModes: z\n      .array(FailureModeSchema)\n      .default([])\n      .describe('Patterns to detect; each detected mode appears in the result.failureModes list.'),\n    wins: z\n      .array(FailureModeSchema)\n      .default([])\n      .describe('Positive patterns; each detected one appears in the result.wins list.'),\n  })\n  .openapi('Rubric')\n\n// ── Judge call ──────────────────────────────────────────────────────\n\nexport const JudgeRequestSchema = z\n  .object({\n    rubricName: z\n      .string()\n      .optional()\n      .describe('Use a built-in rubric by name. Mutually exclusive with `rubric`.'),\n    rubric: RubricSchema.optional().describe(\n      'Inline rubric definition. Mutually exclusive with `rubricName`.',\n    ),\n    content: z\n      .string()\n      .min(1)\n      .describe('The text being judged — a tweet, a blog post, a code snippet, anything stringly.'),\n    context: z\n      .record(z.string(), z.unknown())\n      .optional()\n      .describe(\n        'Free-form metadata for the rubric to use — analytics, source URL, author, etc. Surfaced to the LLM.',\n      ),\n    model: z\n      .string()\n      .optional()\n      .describe('Override the judge model (default routes via tcloud). e.g. \"claude-opus-4-7\".'),\n  })\n  .refine((v) => Boolean(v.rubricName) !== Boolean(v.rubric), {\n    message: 'Provide exactly one of `rubricName` or `rubric`.',\n  })\n  .openapi('JudgeRequest')\n\nexport const JudgeResultSchema = z\n  .object({\n    composite: z\n      .number()\n      .min(0)\n      .max(1)\n      .describe('Weighted combination of dimension scores in 0..1. The single number to gate on.'),\n    dimensions: z\n      .record(z.string(), z.number())\n      .describe('Per-dimension score, keyed by RubricDimension.id.'),\n    failureModes: z\n      .array(z.string())\n      .default([])\n      .describe('Failure-mode ids detected in the content (subset of rubric.failureModes ids).'),\n    wins: z\n      .array(z.string())\n      .default([])\n      .describe('Win ids detected in the content (subset of rubric.wins ids).'),\n    rationale: z\n      .string()\n      .describe('Plain-English explanation of the score. Surfaced to the human reviewer.'),\n    rubricVersion: z\n      .string()\n      .describe(\n        'Stable hash of the rubric used. Scores are only comparable across runs when this matches.',\n      ),\n    model: z.string().describe('Model that produced the judgement, for reproducibility.'),\n    durationMs: z.number().int().nonnegative().describe('End-to-end wall time for this call.'),\n  })\n  .openapi('JudgeResult')\n\n// ── Rubric listing ──────────────────────────────────────────────────\n\nexport const RubricInfoSchema = z\n  .object({\n    name: z.string().describe('Pass this to /v1/judge as `rubricName`.'),\n    description: z.string().describe('What this rubric measures.'),\n    dimensions: z\n      .array(z.object({ id: z.string(), description: z.string(), weight: z.number() }))\n      .describe('The scoring axes this rubric uses, with weights.'),\n    failureModes: z.array(z.string()).default([]).describe('Failure-mode ids this rubric detects.'),\n    rubricVersion: z.string().describe('Stable hash — match this to compare scores across runs.'),\n  })\n  .openapi('RubricInfo')\n\nexport const ListRubricsResponseSchema = z\n  .object({\n    rubrics: z.array(RubricInfoSchema),\n  })\n  .openapi('ListRubricsResponse')\n\n// ── Version / health ────────────────────────────────────────────────\n\nexport const VersionResponseSchema = z\n  .object({\n    package: z.string().describe('Package name (always \"@tangle-network/agent-eval\").'),\n    version: z.string().describe('Semver of the running server. Match your client to this.'),\n    wireVersion: z\n      .string()\n      .describe(\n        'Wire-protocol semver. Bumps separately from package version when the schema changes.',\n      ),\n    apiSurface: z.array(z.string()).describe('List of supported method names.'),\n  })\n  .openapi('VersionResponse')\n\nexport const HealthResponseSchema = z\n  .object({\n    status: z.literal('ok'),\n    uptimeSec: z.number(),\n  })\n  .openapi('HealthResponse')\n\n// ── Errors ──────────────────────────────────────────────────────────\n\nexport const ErrorResponseSchema = z\n  .object({\n    error: z\n      .object({\n        code: z\n          .string()\n          .describe('Machine-readable code: \"validation_error\", \"rubric_not_found\", \"judge_error\".'),\n        message: z.string().describe('Human-readable message.'),\n        details: z.unknown().optional().describe('Optional structured detail.'),\n      })\n      .describe('Errors are always wrapped in this shape across all endpoints.'),\n  })\n  .openapi('ErrorResponse')\n\n// ── Type exports for callers in the same package ────────────────────\n\nexport type RubricDimension = z.infer<typeof RubricDimensionSchema>\nexport type FailureMode = z.infer<typeof FailureModeSchema>\nexport type Rubric = z.infer<typeof RubricSchema>\nexport type JudgeRequest = z.infer<typeof JudgeRequestSchema>\nexport type JudgeResult = z.infer<typeof JudgeResultSchema>\nexport type RubricInfo = z.infer<typeof RubricInfoSchema>\nexport type ListRubricsResponse = z.infer<typeof ListRubricsResponseSchema>\nexport type VersionResponse = z.infer<typeof VersionResponseSchema>\nexport type ErrorResponse = z.infer<typeof ErrorResponseSchema>\n\n// ── Wire-protocol version ───────────────────────────────────────────\n\n/**\n * Bump on any breaking change to a request/response schema.\n * Non-breaking (additive) changes don't require a bump.\n */\nexport const WIRE_VERSION = '1.0.0'\n\n/**\n * Stable hash of a rubric. Used to make scores comparable across runs:\n * if the rubricVersion matches, the rubric was identical.\n */\nexport function hashRubric(rubric: Rubric): string {\n  // deterministic stringify (keys sorted) for stable hashing\n  const stable = JSON.stringify(rubric, Object.keys(rubric).sort())\n  let h = 5381\n  for (let i = 0; i < stable.length; i++) {\n    h = (h * 33) ^ stable.charCodeAt(i)\n  }\n  // Unsigned 32-bit hex, prefixed with rubric name + version slot\n  return `${rubric.name}@${(h >>> 0).toString(16).padStart(8, '0')}`\n}\n","/**\n * Built-in rubrics shipped with agent-eval.\n *\n * A rubric is a set of scoring axes plus a system prompt that tells the\n * judging LLM how to grade against those axes. Built-in rubrics are\n * curated for use cases that recur across Tangle projects — call them\n * by name from any client.\n *\n * Adding a rubric:\n *   1. Define the Rubric object below with a clear `description` and\n *      named `dimensions`.\n *   2. Register it in `BUILTIN_RUBRICS` at the bottom.\n *   3. Add a test in `tests/wire/rubrics.test.ts`.\n *\n * Custom rubrics: callers pass `rubric` inline to /v1/judge instead of\n * `rubricName` — see schemas.ts.\n */\nimport type { Rubric } from './schemas'\nimport { hashRubric } from './schemas'\n\n// ── anti-slop ───────────────────────────────────────────────────────\n// Voice/style judge tuned for technical-buyer audiences. Used by the\n// Postiz autoresearch loop and any content-quality gate.\n\nconst ANTI_SLOP: Rubric = {\n  name: 'anti-slop',\n  description:\n    'Voice and signal quality for content aimed at senior engineers. Catches AI cadence, marketing tone, and engagement-bait shapes.',\n  systemPrompt: `You are evaluating a piece of content written for senior engineers and technical founders.\n\nYou score three things:\n- buyer_quality (0..1): would a senior engineer in the target ICP find this worth their attention? High = specific, earned, technically interesting. Low = generic, hyped, off-target.\n- voice (0..1): does it read like a person who built the thing, or like AI/marketing copy?\n- signal (0..1): does it contain a non-obvious detail, constraint, or claim a reader couldn't get from the public docs?\n\nDetect failure modes (return ids matching):\n- ai-cadence: rule-of-three openings, em-dash flourish, \"Let me explain\", \"Here's the thing\", AI rhythm\n- marketing-tone: \"We're excited to announce\", \"thrilled\", \"delighted\", \"game-changer\", buzzword stack\n- vague-claim: technical claim without a specific component, file, or measurement\n- no-hook: opening doesn't earn attention from the target reader\n- engagement-bait: \"agree?\", \"thoughts?\", listicles, controversy-fishing, hook-detail-pitch\n- off-icp: content shape would attract motivational/grift/hype audiences instead of buyers\n- stale-claim: repeats a positioning line we've used many times this month\n\nDetect wins (return ids matching):\n- specific-component: names a real file, component, or measurement\n- earned-detail: shares a non-obvious detail not derivable from public docs\n- constraint-articulated: names a real tradeoff and the side chosen\n- honest-failure: describes a real failure mode and what was done about it\n\nReturn ONLY JSON matching the response schema. Be conservative — most content has 0-1 wins and 1-2 failure modes, not many of each.`,\n  dimensions: [\n    {\n      id: 'buyer_quality',\n      description: 'Would the target buyer find this worth attention?',\n      weight: 0.5,\n      min: 0,\n      max: 1,\n    },\n    {\n      id: 'voice',\n      description: 'Does it sound like a builder, not AI or marketing?',\n      weight: 0.3,\n      min: 0,\n      max: 1,\n    },\n    {\n      id: 'signal',\n      description: 'Non-obvious detail, constraint, or claim?',\n      weight: 0.2,\n      min: 0,\n      max: 1,\n    },\n  ],\n  failureModes: [\n    { id: 'ai-cadence', description: 'AI-rhythm openings and transitions' },\n    { id: 'marketing-tone', description: 'Buzzwords, hype, corporate-PR voice' },\n    { id: 'vague-claim', description: 'Technical claim without specifics' },\n    { id: 'no-hook', description: 'Opening fails to earn attention' },\n    { id: 'engagement-bait', description: 'Listicle/controversy/agree-pattern' },\n    { id: 'off-icp', description: 'Voice attracts the wrong audience' },\n    { id: 'stale-claim', description: 'Reuses an over-used positioning line' },\n  ],\n  wins: [\n    { id: 'specific-component', description: 'Names a real file/component/number' },\n    { id: 'earned-detail', description: 'Detail not in public docs' },\n    { id: 'constraint-articulated', description: 'Names a real tradeoff' },\n    { id: 'honest-failure', description: 'Describes a real failure honestly' },\n  ],\n}\n\n// ── Registry ────────────────────────────────────────────────────────\n\nexport const BUILTIN_RUBRICS: Record<string, Rubric> = {\n  'anti-slop': ANTI_SLOP,\n}\n\n/** Get a built-in rubric by name, or undefined. */\nexport function getBuiltinRubric(name: string): Rubric | undefined {\n  return BUILTIN_RUBRICS[name]\n}\n\n/** List built-in rubrics with their stable versions. */\nexport function listBuiltinRubrics() {\n  return Object.values(BUILTIN_RUBRICS).map((r) => ({\n    name: r.name,\n    description: r.description,\n    dimensions: r.dimensions.map((d) => ({\n      id: d.id,\n      description: d.description,\n      weight: d.weight,\n    })),\n    failureModes: r.failureModes.map((f) => f.id),\n    rubricVersion: hashRubric(r),\n  }))\n}\n","/**\n * Pure handler functions — the \"business logic\" behind every wire-protocol\n * method. The HTTP server (`server.ts`) and the stdio RPC (`rpc.ts`) both\n * call these. Tests call these directly without spinning a server.\n *\n * Each handler:\n *   - Takes a parsed request (already Zod-validated by the transport).\n *   - Returns a result that matches the response schema.\n *   - Throws `WireError` for caller-fixable errors (404, 400, 422).\n *   - Lets unexpected errors bubble — the transport maps them to 500.\n */\nimport { callLlmJson } from '../llm-client'\nimport { getBuiltinRubric, listBuiltinRubrics } from './rubrics'\nimport {\n  hashRubric,\n  WIRE_VERSION,\n  type JudgeRequest,\n  type JudgeResult,\n  type ListRubricsResponse,\n  type Rubric,\n  type VersionResponse,\n} from './schemas'\n\n/** Caller-fixable error. The transport renders this to 4xx + ErrorResponse. */\nexport class WireError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n    public readonly status: number = 400,\n    public readonly details?: unknown,\n  ) {\n    super(message)\n    this.name = 'WireError'\n  }\n}\n\n// ── judge ───────────────────────────────────────────────────────────\n\n/** The JSON schema we ask the judging LLM to fill in. */\nfunction judgeOutputSchema(rubric: Rubric) {\n  return {\n    name: 'JudgeOutput',\n    schema: {\n      type: 'object',\n      additionalProperties: false,\n      properties: {\n        dimensions: {\n          type: 'object',\n          additionalProperties: false,\n          properties: Object.fromEntries(\n            rubric.dimensions.map((d) => [\n              d.id,\n              { type: 'number', minimum: d.min, maximum: d.max },\n            ]),\n          ),\n          required: rubric.dimensions.map((d) => d.id),\n        },\n        failureModes: {\n          type: 'array',\n          items: { type: 'string', enum: rubric.failureModes.map((f) => f.id) },\n        },\n        wins: {\n          type: 'array',\n          items: { type: 'string', enum: rubric.wins.map((w) => w.id) },\n        },\n        rationale: { type: 'string' },\n      },\n      required: ['dimensions', 'rationale'],\n    } as Record<string, unknown>,\n  }\n}\n\ninterface JudgeOutput {\n  dimensions: Record<string, number>\n  failureModes?: string[]\n  wins?: string[]\n  rationale: string\n}\n\nfunction compositeScore(dimensions: Record<string, number>, rubric: Rubric): number {\n  let weighted = 0\n  let totalWeight = 0\n  for (const dim of rubric.dimensions) {\n    const raw = dimensions[dim.id] ?? 0\n    const range = dim.max - dim.min || 1\n    const normalized = Math.max(0, Math.min(1, (raw - dim.min) / range))\n    weighted += normalized * dim.weight\n    totalWeight += dim.weight\n  }\n  return totalWeight > 0 ? weighted / totalWeight : 0\n}\n\nfunction buildJudgePrompt(content: string, context: unknown): string {\n  const ctx = context && Object.keys(context as object).length ? JSON.stringify(context) : ''\n  return [\n    `CONTENT TO JUDGE:`,\n    content,\n    '',\n    ctx ? `CONTEXT (metadata, analytics, etc.):` : '',\n    ctx ? ctx : '',\n  ]\n    .filter(Boolean)\n    .join('\\n')\n}\n\nconst DEFAULT_JUDGE_MODEL = 'claude-sonnet-4-6'\n\nexport async function handleJudge(req: JudgeRequest): Promise<JudgeResult> {\n  // Resolve rubric\n  let rubric: Rubric\n  if (req.rubricName) {\n    const found = getBuiltinRubric(req.rubricName)\n    if (!found) {\n      throw new WireError('rubric_not_found', `No built-in rubric named \"${req.rubricName}\".`, 404)\n    }\n    rubric = found\n  } else if (req.rubric) {\n    rubric = req.rubric\n  } else {\n    // refine() in the schema should already have caught this — defense in depth\n    throw new WireError('validation_error', 'Provide either `rubricName` or `rubric`.', 422)\n  }\n\n  const startedAt = Date.now()\n  const model = req.model ?? DEFAULT_JUDGE_MODEL\n\n  const { value, result } = await callLlmJson<JudgeOutput>({\n    model,\n    messages: [\n      { role: 'system', content: rubric.systemPrompt },\n      { role: 'user', content: buildJudgePrompt(req.content, req.context) },\n    ],\n    jsonSchema: judgeOutputSchema(rubric),\n    temperature: 0.0,\n    timeoutMs: 60_000,\n  })\n\n  // Defensive: ensure dimensions object isn't malformed\n  if (!value || typeof value !== 'object' || !value.dimensions) {\n    throw new WireError('judge_error', 'Judge returned malformed output.', 500, value)\n  }\n\n  const composite = compositeScore(value.dimensions, rubric)\n  const durationMs = Date.now() - startedAt\n\n  return {\n    composite,\n    dimensions: value.dimensions,\n    failureModes: value.failureModes ?? [],\n    wins: value.wins ?? [],\n    rationale: value.rationale,\n    rubricVersion: hashRubric(rubric),\n    model: result.model,\n    durationMs,\n  }\n}\n\n// ── listRubrics ─────────────────────────────────────────────────────\n\nexport function handleListRubrics(): ListRubricsResponse {\n  return { rubrics: listBuiltinRubrics() }\n}\n\n// ── version ─────────────────────────────────────────────────────────\n\nimport { readFileSync } from 'node:fs'\nimport { dirname, resolve } from 'node:path'\nimport { fileURLToPath } from 'node:url'\n\nlet CACHED_VERSION: string | undefined\n\nfunction readPackageVersion(): string {\n  if (CACHED_VERSION) return CACHED_VERSION\n  // Walk up from this file looking for the nearest package.json.\n  // In dist/ this is dist/.., in src/wire/ this is ../../package.json.\n  const here = dirname(fileURLToPath(import.meta.url))\n  const candidates = [\n    resolve(here, '..', '..', 'package.json'), // src/wire → repo root\n    resolve(here, '..', 'package.json'), // dist → repo root\n  ]\n  for (const path of candidates) {\n    try {\n      const pkg = JSON.parse(readFileSync(path, 'utf-8')) as { version?: string }\n      if (pkg.version) {\n        CACHED_VERSION = pkg.version\n        return pkg.version\n      }\n    } catch {\n      // try next\n    }\n  }\n  return '0.0.0-unknown'\n}\n\nexport function handleVersion(): VersionResponse {\n  return {\n    package: '@tangle-network/agent-eval',\n    version: readPackageVersion(),\n    wireVersion: WIRE_VERSION,\n    apiSurface: ['judge', 'listRubrics', 'version'],\n  }\n}\n","/**\n * Build an OpenAPI spec from the wire schemas.\n *\n * The spec is the contract that other-language clients (Python, Rust,\n * Go, …) generate from. There is no hand-written client — clients are\n * derived artifacts of this file plus `schemas.ts`.\n *\n * Run `pnpm openapi` (defined in package.json) to write the spec to\n * `dist/openapi.json`. CI uses that file to regenerate the Python\n * client and gate the dual-publish workflow.\n */\nimport { OpenApiGeneratorV31, OpenAPIRegistry } from '@asteasolutions/zod-to-openapi'\nimport type { OpenAPIObject } from 'openapi3-ts/oas31'\n\nimport {\n  ErrorResponseSchema,\n  HealthResponseSchema,\n  JudgeRequestSchema,\n  JudgeResultSchema,\n  ListRubricsResponseSchema,\n  VersionResponseSchema,\n  WIRE_VERSION,\n} from './schemas'\n\nexport function buildOpenApi(packageVersion: string): OpenAPIObject {\n  const registry = new OpenAPIRegistry()\n\n  // Components — each schema becomes a $ref-able component\n  registry.register('JudgeRequest', JudgeRequestSchema)\n  registry.register('JudgeResult', JudgeResultSchema)\n  registry.register('ListRubricsResponse', ListRubricsResponseSchema)\n  registry.register('VersionResponse', VersionResponseSchema)\n  registry.register('HealthResponse', HealthResponseSchema)\n  registry.register('ErrorResponse', ErrorResponseSchema)\n\n  // Routes\n  registry.registerPath({\n    method: 'post',\n    path: '/v1/judge',\n    summary: 'Score a piece of content against a rubric',\n    description:\n      'Runs the judging LLM with the named (or inline) rubric and returns dimension scores, detected failure modes, wins, and a composite score in 0..1.',\n    request: {\n      body: {\n        content: {\n          'application/json': { schema: JudgeRequestSchema },\n        },\n      },\n    },\n    responses: {\n      200: {\n        description: 'Successful judgement',\n        content: { 'application/json': { schema: JudgeResultSchema } },\n      },\n      400: {\n        description: 'Validation error',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      404: {\n        description: 'Rubric not found',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n      500: {\n        description: 'Judge error',\n        content: { 'application/json': { schema: ErrorResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'get',\n    path: '/v1/rubrics',\n    summary: 'List built-in rubrics',\n    description:\n      'Returns every rubric registered server-side, with their dimensions and stable rubricVersion hash.',\n    responses: {\n      200: {\n        description: 'Listing',\n        content: { 'application/json': { schema: ListRubricsResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'get',\n    path: '/v1/version',\n    summary: 'Server and wire-protocol version',\n    description: 'Match your client version to `version`; check `wireVersion` for compatibility.',\n    responses: {\n      200: {\n        description: 'Version info',\n        content: { 'application/json': { schema: VersionResponseSchema } },\n      },\n    },\n  })\n\n  registry.registerPath({\n    method: 'get',\n    path: '/healthz',\n    summary: 'Liveness check',\n    responses: {\n      200: {\n        description: 'OK',\n        content: { 'application/json': { schema: HealthResponseSchema } },\n      },\n    },\n  })\n\n  const generator = new OpenApiGeneratorV31(registry.definitions)\n  return generator.generateDocument({\n    openapi: '3.1.0',\n    info: {\n      title: '@tangle-network/agent-eval — wire protocol',\n      version: packageVersion,\n      description: `HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: ${WIRE_VERSION}. Bumps on breaking changes to request/response schemas.`,\n      contact: { name: 'Tangle Network', url: 'https://github.com/tangle-network/agent-eval' },\n      license: { name: 'MIT' },\n    },\n    servers: [{ url: 'http://localhost:5005', description: 'Local agent-eval serve' }],\n  })\n}\n","/**\n * HTTP transport for the wire protocol.\n *\n * Hono + @hono/node-server. Every endpoint:\n *   1. Validates the request against its Zod schema.\n *   2. Calls the matching handler in `handlers.ts`.\n *   3. Renders 4xx for `WireError` with structured body, 500 for unexpected.\n *\n * The server has no internal state besides the handler imports — restart\n * costs nothing. Run via `agent-eval serve --port 5005`.\n */\nimport { serve, type ServerType } from '@hono/node-server'\nimport { Hono } from 'hono'\nimport { cors } from 'hono/cors'\n\nimport {\n  handleJudge,\n  handleListRubrics,\n  handleVersion,\n  WireError,\n} from './handlers'\nimport { buildOpenApi } from './openapi'\nimport { JudgeRequestSchema } from './schemas'\n\nconst STARTED_AT = Date.now()\n\nexport function createApp() {\n  const app = new Hono()\n\n  app.use('*', cors())\n\n  app.onError((err, c) => {\n    if (err instanceof WireError) {\n      return c.json(\n        { error: { code: err.code, message: err.message, details: err.details } },\n        err.status as 400 | 404 | 422 | 500,\n      )\n    }\n    // Unexpected — log and return generic 500 without leaking internals.\n    console.error('[agent-eval] unhandled error:', err)\n    return c.json(\n      { error: { code: 'internal_error', message: 'Internal server error.' } },\n      500,\n    )\n  })\n\n  // ── Health ──\n  app.get('/healthz', (c) =>\n    c.json({ status: 'ok' as const, uptimeSec: (Date.now() - STARTED_AT) / 1000 }),\n  )\n\n  // ── Version ──\n  app.get('/v1/version', (c) => c.json(handleVersion()))\n\n  // ── Rubrics ──\n  app.get('/v1/rubrics', (c) => c.json(handleListRubrics()))\n\n  // ── Judge ──\n  app.post('/v1/judge', async (c) => {\n    const raw = await c.req.json().catch(() => null)\n    if (raw == null) {\n      throw new WireError('validation_error', 'Request body must be JSON.', 400)\n    }\n    const parsed = JudgeRequestSchema.safeParse(raw)\n    if (!parsed.success) {\n      throw new WireError(\n        'validation_error',\n        'Request did not match JudgeRequest schema.',\n        400,\n        parsed.error.issues,\n      )\n    }\n    const result = await handleJudge(parsed.data)\n    return c.json(result)\n  })\n\n  // ── OpenAPI spec ──\n  app.get('/openapi.json', (c) => c.json(buildOpenApi(handleVersion().version)))\n\n  return app\n}\n\nexport interface ServeOptions {\n  /** Default 5005. */\n  port?: number\n  /** Default '127.0.0.1'. Set to '0.0.0.0' to listen on all interfaces. */\n  host?: string\n}\n\nexport function startServer(opts: ServeOptions = {}): ServerType {\n  const app = createApp()\n  const port = opts.port ?? 5005\n  const host = opts.host ?? '127.0.0.1'\n  return serve({ fetch: app.fetch, port, hostname: host }, ({ address, port: actualPort }) => {\n    // eslint-disable-next-line no-console\n    console.log(`[agent-eval] serving on http://${address}:${actualPort}`)\n  })\n}\n","/**\n * stdio RPC transport.\n *\n * For batch / cron use without a running server. The Python client falls\n * back to this when no server is reachable.\n *\n * Protocol (line-delimited JSON over stdin/stdout):\n *   IN:  one JSON object on stdin: {\"method\":\"judge\",\"params\":{...}}\n *   OUT: one JSON object on stdout: {\"result\":{...}} or {\"error\":{...}}\n *\n * One request per process invocation. To pipeline many calls, the client\n * writes JSONL to stdin and reads JSONL from stdout — see batch mode below.\n */\nimport { handleJudge, handleListRubrics, handleVersion, WireError } from './handlers'\nimport { JudgeRequestSchema } from './schemas'\n\ninterface RpcRequest {\n  method: 'judge' | 'listRubrics' | 'version'\n  params?: unknown\n}\n\ninterface RpcSuccess {\n  result: unknown\n}\n\ninterface RpcError {\n  error: { code: string; message: string; details?: unknown }\n}\n\nexport async function dispatchRpc(req: RpcRequest): Promise<RpcSuccess | RpcError> {\n  try {\n    switch (req.method) {\n      case 'judge': {\n        const parsed = JudgeRequestSchema.safeParse(req.params)\n        if (!parsed.success) {\n          return {\n            error: {\n              code: 'validation_error',\n              message: 'params did not match JudgeRequest schema.',\n              details: parsed.error.issues,\n            },\n          }\n        }\n        return { result: await handleJudge(parsed.data) }\n      }\n      case 'listRubrics':\n        return { result: handleListRubrics() }\n      case 'version':\n        return { result: handleVersion() }\n      default:\n        return {\n          error: {\n            code: 'unknown_method',\n            message: `No such method: ${(req as { method: string }).method}`,\n          },\n        }\n    }\n  } catch (err) {\n    if (err instanceof WireError) {\n      return { error: { code: err.code, message: err.message, details: err.details } }\n    }\n    const message = err instanceof Error ? err.message : String(err)\n    return { error: { code: 'internal_error', message } }\n  }\n}\n\n// ── stdin/stdout driver ─────────────────────────────────────────────\n\nasync function readAll(stream: NodeJS.ReadableStream): Promise<string> {\n  const chunks: Buffer[] = []\n  for await (const chunk of stream) {\n    chunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk as string))\n  }\n  return Buffer.concat(chunks).toString('utf-8')\n}\n\n/** Read one JSON request from stdin, write one JSON response to stdout. */\nexport async function runRpcOnce(method?: string): Promise<number> {\n  const raw = await readAll(process.stdin)\n  let req: RpcRequest\n  try {\n    const body = JSON.parse(raw)\n    req = method ? { method: method as RpcRequest['method'], params: body } : (body as RpcRequest)\n  } catch (err) {\n    process.stdout.write(\n      JSON.stringify({\n        error: {\n          code: 'parse_error',\n          message: `stdin was not valid JSON: ${err instanceof Error ? err.message : String(err)}`,\n        },\n      }) + '\\n',\n    )\n    return 1\n  }\n  const out = await dispatchRpc(req)\n  process.stdout.write(JSON.stringify(out) + '\\n')\n  return 'error' in out ? 1 : 0\n}\n\n/** Read JSONL requests from stdin, write JSONL responses to stdout. */\nexport async function runRpcBatch(method?: string): Promise<number> {\n  const raw = await readAll(process.stdin)\n  const lines = raw.split('\\n').filter((l) => l.trim().length > 0)\n  let exitCode = 0\n  for (const line of lines) {\n    let req: RpcRequest\n    try {\n      const body = JSON.parse(line)\n      req = method ? { method: method as RpcRequest['method'], params: body } : (body as RpcRequest)\n    } catch (err) {\n      process.stdout.write(\n        JSON.stringify({\n          error: {\n            code: 'parse_error',\n            message: `line was not valid JSON: ${err instanceof Error ? err.message : String(err)}`,\n          },\n        }) + '\\n',\n      )\n      exitCode = 1\n      continue\n    }\n    const out = await dispatchRpc(req)\n    process.stdout.write(JSON.stringify(out) + '\\n')\n    if ('error' in out) exitCode = 1\n  }\n  return exitCode\n}\n"],"mappings":";;;;;AAYA,SAAS,4BAA4B;AACrC,SAAS,SAAS;AAElB,qBAAqB,CAAC;AAIf,IAAM,wBAAwB,EAClC,OAAO;AAAA,EACN,IAAI,EACD,OAAO,EACP,IAAI,CAAC,EACL,SAAS,gFAA2E;AAAA,EACvF,aAAa,EACV,OAAO,EACP,IAAI,CAAC,EACL,SAAS,sEAAsE;AAAA,EAClF,QAAQ,EACL,OAAO,EACP,IAAI,CAAC,EACL,QAAQ,CAAC,EACT,SAAS,gEAAgE;AAAA,EAC5E,KAAK,EAAE,OAAO,EAAE,QAAQ,CAAC,EAAE,SAAS,gDAAgD;AAAA,EACpF,KAAK,EAAE,OAAO,EAAE,QAAQ,CAAC,EAAE,SAAS,gDAAgD;AACtF,CAAC,EACA,QAAQ,iBAAiB;AAErB,IAAM,oBAAoB,EAC9B,OAAO;AAAA,EACN,IAAI,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,mEAA8D;AAAA,EAC7F,aAAa,EAAE,OAAO,EAAE,IAAI,CAAC,EAAE,SAAS,mDAAmD;AAC7F,CAAC,EACA,QAAQ,aAAa;AAIjB,IAAM,eAAe,EACzB,OAAO;AAAA,EACN,MAAM,EACH,OAAO,EACP,IAAI,CAAC,EACL,SAAS,4EAAuE;AAAA,EACnF,aAAa,EACV,OAAO,EACP,IAAI,CAAC,EACL,SAAS,0DAA0D;AAAA,EACtE,cAAc,EACX,OAAO,EACP,IAAI,CAAC,EACL;AAAA,IACC;AAAA,EACF;AAAA,EACF,YAAY,EACT,MAAM,qBAAqB,EAC3B,IAAI,CAAC,EACL,SAAS,+DAA+D;AAAA,EAC3E,cAAc,EACX,MAAM,iBAAiB,EACvB,QAAQ,CAAC,CAAC,EACV,SAAS,iFAAiF;AAAA,EAC7F,MAAM,EACH,MAAM,iBAAiB,EACvB,QAAQ,CAAC,CAAC,EACV,SAAS,uEAAuE;AACrF,CAAC,EACA,QAAQ,QAAQ;AAIZ,IAAM,qBAAqB,EAC/B,OAAO;AAAA,EACN,YAAY,EACT,OAAO,EACP,SAAS,EACT,SAAS,kEAAkE;AAAA,EAC9E,QAAQ,aAAa,SAAS,EAAE;AAAA,IAC9B;AAAA,EACF;AAAA,EACA,SAAS,EACN,OAAO,EACP,IAAI,CAAC,EACL,SAAS,uFAAkF;AAAA,EAC9F,SAAS,EACN,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAC9B,SAAS,EACT;AAAA,IACC;AAAA,EACF;AAAA,EACF,OAAO,EACJ,OAAO,EACP,SAAS,EACT,SAAS,+EAA+E;AAC7F,CAAC,EACA,OAAO,CAAC,MAAM,QAAQ,EAAE,UAAU,MAAM,QAAQ,EAAE,MAAM,GAAG;AAAA,EAC1D,SAAS;AACX,CAAC,EACA,QAAQ,cAAc;AAElB,IAAM,oBAAoB,EAC9B,OAAO;AAAA,EACN,WAAW,EACR,OAAO,EACP,IAAI,CAAC,EACL,IAAI,CAAC,EACL,SAAS,iFAAiF;AAAA,EAC7F,YAAY,EACT,OAAO,EAAE,OAAO,GAAG,EAAE,OAAO,CAAC,EAC7B,SAAS,mDAAmD;AAAA,EAC/D,cAAc,EACX,MAAM,EAAE,OAAO,CAAC,EAChB,QAAQ,CAAC,CAAC,EACV,SAAS,+EAA+E;AAAA,EAC3F,MAAM,EACH,MAAM,EAAE,OAAO,CAAC,EAChB,QAAQ,CAAC,CAAC,EACV,SAAS,8DAA8D;AAAA,EAC1E,WAAW,EACR,OAAO,EACP,SAAS,yEAAyE;AAAA,EACrF,eAAe,EACZ,OAAO,EACP;AAAA,IACC;AAAA,EACF;AAAA,EACF,OAAO,EAAE,OAAO,EAAE,SAAS,yDAAyD;AAAA,EACpF,YAAY,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,SAAS,qCAAqC;AAC3F,CAAC,EACA,QAAQ,aAAa;AAIjB,IAAM,mBAAmB,EAC7B,OAAO;AAAA,EACN,MAAM,EAAE,OAAO,EAAE,SAAS,yCAAyC;AAAA,EACnE,aAAa,EAAE,OAAO,EAAE,SAAS,4BAA4B;AAAA,EAC7D,YAAY,EACT,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,OAAO,GAAG,aAAa,EAAE,OAAO,GAAG,QAAQ,EAAE,OAAO,EAAE,CAAC,CAAC,EAC/E,SAAS,kDAAkD;AAAA,EAC9D,cAAc,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,QAAQ,CAAC,CAAC,EAAE,SAAS,uCAAuC;AAAA,EAC9F,eAAe,EAAE,OAAO,EAAE,SAAS,8DAAyD;AAC9F,CAAC,EACA,QAAQ,YAAY;AAEhB,IAAM,4BAA4B,EACtC,OAAO;AAAA,EACN,SAAS,EAAE,MAAM,gBAAgB;AACnC,CAAC,EACA,QAAQ,qBAAqB;AAIzB,IAAM,wBAAwB,EAClC,OAAO;AAAA,EACN,SAAS,EAAE,OAAO,EAAE,SAAS,qDAAqD;AAAA,EAClF,SAAS,EAAE,OAAO,EAAE,SAAS,0DAA0D;AAAA,EACvF,aAAa,EACV,OAAO,EACP;AAAA,IACC;AAAA,EACF;AAAA,EACF,YAAY,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,SAAS,iCAAiC;AAC5E,CAAC,EACA,QAAQ,iBAAiB;AAErB,IAAM,uBAAuB,EACjC,OAAO;AAAA,EACN,QAAQ,EAAE,QAAQ,IAAI;AAAA,EACtB,WAAW,EAAE,OAAO;AACtB,CAAC,EACA,QAAQ,gBAAgB;AAIpB,IAAM,sBAAsB,EAChC,OAAO;AAAA,EACN,OAAO,EACJ,OAAO;AAAA,IACN,MAAM,EACH,OAAO,EACP,SAAS,+EAA+E;AAAA,IAC3F,SAAS,EAAE,OAAO,EAAE,SAAS,yBAAyB;AAAA,IACtD,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,6BAA6B;AAAA,EACxE,CAAC,EACA,SAAS,+DAA+D;AAC7E,CAAC,EACA,QAAQ,eAAe;AAoBnB,IAAM,eAAe;AAMrB,SAAS,WAAW,QAAwB;AAEjD,QAAM,SAAS,KAAK,UAAU,QAAQ,OAAO,KAAK,MAAM,EAAE,KAAK,CAAC;AAChE,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,QAAK,IAAI,KAAM,OAAO,WAAW,CAAC;AAAA,EACpC;AAEA,SAAO,GAAG,OAAO,IAAI,KAAK,MAAM,GAAG,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG,CAAC;AAClE;;;AChNA,IAAM,YAAoB;AAAA,EACxB,MAAM;AAAA,EACN,aACE;AAAA,EACF,cAAc;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAuBd,YAAY;AAAA,IACV;AAAA,MACE,IAAI;AAAA,MACJ,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,IACP;AAAA,EACF;AAAA,EACA,cAAc;AAAA,IACZ,EAAE,IAAI,cAAc,aAAa,qCAAqC;AAAA,IACtE,EAAE,IAAI,kBAAkB,aAAa,sCAAsC;AAAA,IAC3E,EAAE,IAAI,eAAe,aAAa,oCAAoC;AAAA,IACtE,EAAE,IAAI,WAAW,aAAa,kCAAkC;AAAA,IAChE,EAAE,IAAI,mBAAmB,aAAa,qCAAqC;AAAA,IAC3E,EAAE,IAAI,WAAW,aAAa,oCAAoC;AAAA,IAClE,EAAE,IAAI,eAAe,aAAa,uCAAuC;AAAA,EAC3E;AAAA,EACA,MAAM;AAAA,IACJ,EAAE,IAAI,sBAAsB,aAAa,qCAAqC;AAAA,IAC9E,EAAE,IAAI,iBAAiB,aAAa,4BAA4B;AAAA,IAChE,EAAE,IAAI,0BAA0B,aAAa,wBAAwB;AAAA,IACrE,EAAE,IAAI,kBAAkB,aAAa,oCAAoC;AAAA,EAC3E;AACF;AAIO,IAAM,kBAA0C;AAAA,EACrD,aAAa;AACf;AAGO,SAAS,iBAAiB,MAAkC;AACjE,SAAO,gBAAgB,IAAI;AAC7B;AAGO,SAAS,qBAAqB;AACnC,SAAO,OAAO,OAAO,eAAe,EAAE,IAAI,CAAC,OAAO;AAAA,IAChD,MAAM,EAAE;AAAA,IACR,aAAa,EAAE;AAAA,IACf,YAAY,EAAE,WAAW,IAAI,CAAC,OAAO;AAAA,MACnC,IAAI,EAAE;AAAA,MACN,aAAa,EAAE;AAAA,MACf,QAAQ,EAAE;AAAA,IACZ,EAAE;AAAA,IACF,cAAc,EAAE,aAAa,IAAI,CAAC,MAAM,EAAE,EAAE;AAAA,IAC5C,eAAe,WAAW,CAAC;AAAA,EAC7B,EAAE;AACJ;;;ACkDA,SAAS,oBAAoB;AAC7B,SAAS,SAAS,eAAe;AACjC,SAAS,qBAAqB;AA/IvB,IAAM,YAAN,cAAwB,MAAM;AAAA,EACnC,YACkB,MAChB,SACgB,SAAiB,KACjB,SAChB;AACA,UAAM,OAAO;AALG;AAEA;AACA;AAGhB,SAAK,OAAO;AAAA,EACd;AAAA,EAPkB;AAAA,EAEA;AAAA,EACA;AAKpB;AAKA,SAAS,kBAAkB,QAAgB;AACzC,SAAO;AAAA,IACL,MAAM;AAAA,IACN,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,sBAAsB;AAAA,MACtB,YAAY;AAAA,QACV,YAAY;AAAA,UACV,MAAM;AAAA,UACN,sBAAsB;AAAA,UACtB,YAAY,OAAO;AAAA,YACjB,OAAO,WAAW,IAAI,CAAC,MAAM;AAAA,cAC3B,EAAE;AAAA,cACF,EAAE,MAAM,UAAU,SAAS,EAAE,KAAK,SAAS,EAAE,IAAI;AAAA,YACnD,CAAC;AAAA,UACH;AAAA,UACA,UAAU,OAAO,WAAW,IAAI,CAAC,MAAM,EAAE,EAAE;AAAA,QAC7C;AAAA,QACA,cAAc;AAAA,UACZ,MAAM;AAAA,UACN,OAAO,EAAE,MAAM,UAAU,MAAM,OAAO,aAAa,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE;AAAA,QACtE;AAAA,QACA,MAAM;AAAA,UACJ,MAAM;AAAA,UACN,OAAO,EAAE,MAAM,UAAU,MAAM,OAAO,KAAK,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE;AAAA,QAC9D;AAAA,QACA,WAAW,EAAE,MAAM,SAAS;AAAA,MAC9B;AAAA,MACA,UAAU,CAAC,cAAc,WAAW;AAAA,IACtC;AAAA,EACF;AACF;AASA,SAAS,eAAe,YAAoC,QAAwB;AAClF,MAAI,WAAW;AACf,MAAI,cAAc;AAClB,aAAW,OAAO,OAAO,YAAY;AACnC,UAAM,MAAM,WAAW,IAAI,EAAE,KAAK;AAClC,UAAM,QAAQ,IAAI,MAAM,IAAI,OAAO;AACnC,UAAM,aAAa,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,MAAM,IAAI,OAAO,KAAK,CAAC;AACnE,gBAAY,aAAa,IAAI;AAC7B,mBAAe,IAAI;AAAA,EACrB;AACA,SAAO,cAAc,IAAI,WAAW,cAAc;AACpD;AAEA,SAAS,iBAAiB,SAAiB,SAA0B;AACnE,QAAM,MAAM,WAAW,OAAO,KAAK,OAAiB,EAAE,SAAS,KAAK,UAAU,OAAO,IAAI;AACzF,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM,yCAAyC;AAAA,IAC/C,MAAM,MAAM;AAAA,EACd,EACG,OAAO,OAAO,EACd,KAAK,IAAI;AACd;AAEA,IAAM,sBAAsB;AAE5B,eAAsB,YAAY,KAAyC;AAEzE,MAAI;AACJ,MAAI,IAAI,YAAY;AAClB,UAAM,QAAQ,iBAAiB,IAAI,UAAU;AAC7C,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,UAAU,oBAAoB,6BAA6B,IAAI,UAAU,MAAM,GAAG;AAAA,IAC9F;AACA,aAAS;AAAA,EACX,WAAW,IAAI,QAAQ;AACrB,aAAS,IAAI;AAAA,EACf,OAAO;AAEL,UAAM,IAAI,UAAU,oBAAoB,4CAA4C,GAAG;AAAA,EACzF;AAEA,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,QAAQ,IAAI,SAAS;AAE3B,QAAM,EAAE,OAAO,OAAO,IAAI,MAAM,YAAyB;AAAA,IACvD;AAAA,IACA,UAAU;AAAA,MACR,EAAE,MAAM,UAAU,SAAS,OAAO,aAAa;AAAA,MAC/C,EAAE,MAAM,QAAQ,SAAS,iBAAiB,IAAI,SAAS,IAAI,OAAO,EAAE;AAAA,IACtE;AAAA,IACA,YAAY,kBAAkB,MAAM;AAAA,IACpC,aAAa;AAAA,IACb,WAAW;AAAA,EACb,CAAC;AAGD,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,CAAC,MAAM,YAAY;AAC5D,UAAM,IAAI,UAAU,eAAe,oCAAoC,KAAK,KAAK;AAAA,EACnF;AAEA,QAAM,YAAY,eAAe,MAAM,YAAY,MAAM;AACzD,QAAM,aAAa,KAAK,IAAI,IAAI;AAEhC,SAAO;AAAA,IACL;AAAA,IACA,YAAY,MAAM;AAAA,IAClB,cAAc,MAAM,gBAAgB,CAAC;AAAA,IACrC,MAAM,MAAM,QAAQ,CAAC;AAAA,IACrB,WAAW,MAAM;AAAA,IACjB,eAAe,WAAW,MAAM;AAAA,IAChC,OAAO,OAAO;AAAA,IACd;AAAA,EACF;AACF;AAIO,SAAS,oBAAyC;AACvD,SAAO,EAAE,SAAS,mBAAmB,EAAE;AACzC;AAQA,IAAI;AAEJ,SAAS,qBAA6B;AACpC,MAAI,eAAgB,QAAO;AAG3B,QAAM,OAAO,QAAQ,cAAc,YAAY,GAAG,CAAC;AACnD,QAAM,aAAa;AAAA,IACjB,QAAQ,MAAM,MAAM,MAAM,cAAc;AAAA;AAAA,IACxC,QAAQ,MAAM,MAAM,cAAc;AAAA;AAAA,EACpC;AACA,aAAW,QAAQ,YAAY;AAC7B,QAAI;AACF,YAAM,MAAM,KAAK,MAAM,aAAa,MAAM,OAAO,CAAC;AAClD,UAAI,IAAI,SAAS;AACf,yBAAiB,IAAI;AACrB,eAAO,IAAI;AAAA,MACb;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,gBAAiC;AAC/C,SAAO;AAAA,IACL,SAAS;AAAA,IACT,SAAS,mBAAmB;AAAA,IAC5B,aAAa;AAAA,IACb,YAAY,CAAC,SAAS,eAAe,SAAS;AAAA,EAChD;AACF;;;AC9LA,SAAS,qBAAqB,uBAAuB;AAa9C,SAAS,aAAa,gBAAuC;AAClE,QAAM,WAAW,IAAI,gBAAgB;AAGrC,WAAS,SAAS,gBAAgB,kBAAkB;AACpD,WAAS,SAAS,eAAe,iBAAiB;AAClD,WAAS,SAAS,uBAAuB,yBAAyB;AAClE,WAAS,SAAS,mBAAmB,qBAAqB;AAC1D,WAAS,SAAS,kBAAkB,oBAAoB;AACxD,WAAS,SAAS,iBAAiB,mBAAmB;AAGtD,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aACE;AAAA,IACF,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,SAAS;AAAA,UACP,oBAAoB,EAAE,QAAQ,mBAAmB;AAAA,QACnD;AAAA,MACF;AAAA,IACF;AAAA,IACA,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,kBAAkB,EAAE;AAAA,MAC/D;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,MACA,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,oBAAoB,EAAE;AAAA,MACjE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aACE;AAAA,IACF,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,0BAA0B,EAAE;AAAA,MACvE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aAAa;AAAA,IACb,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,sBAAsB,EAAE;AAAA,MACnE;AAAA,IACF;AAAA,EACF,CAAC;AAED,WAAS,aAAa;AAAA,IACpB,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,SAAS;AAAA,IACT,WAAW;AAAA,MACT,KAAK;AAAA,QACH,aAAa;AAAA,QACb,SAAS,EAAE,oBAAoB,EAAE,QAAQ,qBAAqB,EAAE;AAAA,MAClE;AAAA,IACF;AAAA,EACF,CAAC;AAED,QAAM,YAAY,IAAI,oBAAoB,SAAS,WAAW;AAC9D,SAAO,UAAU,iBAAiB;AAAA,IAChC,SAAS;AAAA,IACT,MAAM;AAAA,MACJ,OAAO;AAAA,MACP,SAAS;AAAA,MACT,aAAa;AAAA;AAAA,yBAEM,YAAY;AAAA,MAC/B,SAAS,EAAE,MAAM,kBAAkB,KAAK,+CAA+C;AAAA,MACvF,SAAS,EAAE,MAAM,MAAM;AAAA,IACzB;AAAA,IACA,SAAS,CAAC,EAAE,KAAK,yBAAyB,aAAa,yBAAyB,CAAC;AAAA,EACnF,CAAC;AACH;;;AC/GA,SAAS,aAA8B;AACvC,SAAS,YAAY;AACrB,SAAS,YAAY;AAWrB,IAAM,aAAa,KAAK,IAAI;AAErB,SAAS,YAAY;AAC1B,QAAM,MAAM,IAAI,KAAK;AAErB,MAAI,IAAI,KAAK,KAAK,CAAC;AAEnB,MAAI,QAAQ,CAAC,KAAK,MAAM;AACtB,QAAI,eAAe,WAAW;AAC5B,aAAO,EAAE;AAAA,QACP,EAAE,OAAO,EAAE,MAAM,IAAI,MAAM,SAAS,IAAI,SAAS,SAAS,IAAI,QAAQ,EAAE;AAAA,QACxE,IAAI;AAAA,MACN;AAAA,IACF;AAEA,YAAQ,MAAM,iCAAiC,GAAG;AAClD,WAAO,EAAE;AAAA,MACP,EAAE,OAAO,EAAE,MAAM,kBAAkB,SAAS,yBAAyB,EAAE;AAAA,MACvE;AAAA,IACF;AAAA,EACF,CAAC;AAGD,MAAI;AAAA,IAAI;AAAA,IAAY,CAAC,MACnB,EAAE,KAAK,EAAE,QAAQ,MAAe,YAAY,KAAK,IAAI,IAAI,cAAc,IAAK,CAAC;AAAA,EAC/E;AAGA,MAAI,IAAI,eAAe,CAAC,MAAM,EAAE,KAAK,cAAc,CAAC,CAAC;AAGrD,MAAI,IAAI,eAAe,CAAC,MAAM,EAAE,KAAK,kBAAkB,CAAC,CAAC;AAGzD,MAAI,KAAK,aAAa,OAAO,MAAM;AACjC,UAAM,MAAM,MAAM,EAAE,IAAI,KAAK,EAAE,MAAM,MAAM,IAAI;AAC/C,QAAI,OAAO,MAAM;AACf,YAAM,IAAI,UAAU,oBAAoB,8BAA8B,GAAG;AAAA,IAC3E;AACA,UAAM,SAAS,mBAAmB,UAAU,GAAG;AAC/C,QAAI,CAAC,OAAO,SAAS;AACnB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,QACA;AAAA,QACA,OAAO,MAAM;AAAA,MACf;AAAA,IACF;AACA,UAAM,SAAS,MAAM,YAAY,OAAO,IAAI;AAC5C,WAAO,EAAE,KAAK,MAAM;AAAA,EACtB,CAAC;AAGD,MAAI,IAAI,iBAAiB,CAAC,MAAM,EAAE,KAAK,aAAa,cAAc,EAAE,OAAO,CAAC,CAAC;AAE7E,SAAO;AACT;AASO,SAAS,YAAY,OAAqB,CAAC,GAAe;AAC/D,QAAM,MAAM,UAAU;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO,MAAM,EAAE,OAAO,IAAI,OAAO,MAAM,UAAU,KAAK,GAAG,CAAC,EAAE,SAAS,MAAM,WAAW,MAAM;AAE1F,YAAQ,IAAI,kCAAkC,OAAO,IAAI,UAAU,EAAE;AAAA,EACvE,CAAC;AACH;;;ACpEA,eAAsB,YAAY,KAAiD;AACjF,MAAI;AACF,YAAQ,IAAI,QAAQ;AAAA,MAClB,KAAK,SAAS;AACZ,cAAM,SAAS,mBAAmB,UAAU,IAAI,MAAM;AACtD,YAAI,CAAC,OAAO,SAAS;AACnB,iBAAO;AAAA,YACL,OAAO;AAAA,cACL,MAAM;AAAA,cACN,SAAS;AAAA,cACT,SAAS,OAAO,MAAM;AAAA,YACxB;AAAA,UACF;AAAA,QACF;AACA,eAAO,EAAE,QAAQ,MAAM,YAAY,OAAO,IAAI,EAAE;AAAA,MAClD;AAAA,MACA,KAAK;AACH,eAAO,EAAE,QAAQ,kBAAkB,EAAE;AAAA,MACvC,KAAK;AACH,eAAO,EAAE,QAAQ,cAAc,EAAE;AAAA,MACnC;AACE,eAAO;AAAA,UACL,OAAO;AAAA,YACL,MAAM;AAAA,YACN,SAAS,mBAAoB,IAA2B,MAAM;AAAA,UAChE;AAAA,QACF;AAAA,IACJ;AAAA,EACF,SAAS,KAAK;AACZ,QAAI,eAAe,WAAW;AAC5B,aAAO,EAAE,OAAO,EAAE,MAAM,IAAI,MAAM,SAAS,IAAI,SAAS,SAAS,IAAI,QAAQ,EAAE;AAAA,IACjF;AACA,UAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,WAAO,EAAE,OAAO,EAAE,MAAM,kBAAkB,QAAQ,EAAE;AAAA,EACtD;AACF;AAIA,eAAe,QAAQ,QAAgD;AACrE,QAAM,SAAmB,CAAC;AAC1B,mBAAiB,SAAS,QAAQ;AAChC,WAAO,KAAK,OAAO,SAAS,KAAK,IAAI,QAAQ,OAAO,KAAK,KAAe,CAAC;AAAA,EAC3E;AACA,SAAO,OAAO,OAAO,MAAM,EAAE,SAAS,OAAO;AAC/C;AAGA,eAAsB,WAAW,QAAkC;AACjE,QAAM,MAAM,MAAM,QAAQ,QAAQ,KAAK;AACvC,MAAI;AACJ,MAAI;AACF,UAAM,OAAO,KAAK,MAAM,GAAG;AAC3B,UAAM,SAAS,EAAE,QAAwC,QAAQ,KAAK,IAAK;AAAA,EAC7E,SAAS,KAAK;AACZ,YAAQ,OAAO;AAAA,MACb,KAAK,UAAU;AAAA,QACb,OAAO;AAAA,UACL,MAAM;AAAA,UACN,SAAS,6BAA6B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACxF;AAAA,MACF,CAAC,IAAI;AAAA,IACP;AACA,WAAO;AAAA,EACT;AACA,QAAM,MAAM,MAAM,YAAY,GAAG;AACjC,UAAQ,OAAO,MAAM,KAAK,UAAU,GAAG,IAAI,IAAI;AAC/C,SAAO,WAAW,MAAM,IAAI;AAC9B;AAGA,eAAsB,YAAY,QAAkC;AAClE,QAAM,MAAM,MAAM,QAAQ,QAAQ,KAAK;AACvC,QAAM,QAAQ,IAAI,MAAM,IAAI,EAAE,OAAO,CAAC,MAAM,EAAE,KAAK,EAAE,SAAS,CAAC;AAC/D,MAAI,WAAW;AACf,aAAW,QAAQ,OAAO;AACxB,QAAI;AACJ,QAAI;AACF,YAAM,OAAO,KAAK,MAAM,IAAI;AAC5B,YAAM,SAAS,EAAE,QAAwC,QAAQ,KAAK,IAAK;AAAA,IAC7E,SAAS,KAAK;AACZ,cAAQ,OAAO;AAAA,QACb,KAAK,UAAU;AAAA,UACb,OAAO;AAAA,YACL,MAAM;AAAA,YACN,SAAS,4BAA4B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UACvF;AAAA,QACF,CAAC,IAAI;AAAA,MACP;AACA,iBAAW;AACX;AAAA,IACF;AACA,UAAM,MAAM,MAAM,YAAY,GAAG;AACjC,YAAQ,OAAO,MAAM,KAAK,UAAU,GAAG,IAAI,IAAI;AAC/C,QAAI,WAAW,IAAK,YAAW;AAAA,EACjC;AACA,SAAO;AACT;","names":[]}

package/dist/cli.d.ts

@@ -0,0 +1 @@
+#!/usr/bin/env node

package/dist/cli.js

@@ -0,0 +1,104 @@
+#!/usr/bin/env node
+import {
+  buildOpenApi,
+  handleVersion,
+  runRpcBatch,
+  runRpcOnce,
+  startServer
+} from "./chunk-OZPRSK4A.js";
+import "./chunk-ITN4YOZY.js";
+
+// src/cli.ts
+import { writeFileSync } from "fs";
+function parseArgs(argv) {
+  const [command, ...rest] = argv;
+  const positional = [];
+  const flags = {};
+  for (let i = 0; i < rest.length; i++) {
+    const tok = rest[i];
+    if (tok.startsWith("--")) {
+      const key = tok.slice(2);
+      const next = rest[i + 1];
+      if (next != null && !next.startsWith("--")) {
+        flags[key] = next;
+        i++;
+      } else {
+        flags[key] = "true";
+      }
+    } else {
+      positional.push(tok);
+    }
+  }
+  return { command: command ?? "help", positional, flags };
+}
+var HELP = `agent-eval \u2014 wire-protocol entry point.
+
+Commands:
+  serve [--port 5005] [--host 127.0.0.1]
+        Start the HTTP server. POST /v1/judge, GET /v1/rubrics, GET /v1/version, GET /openapi.json.
+  rpc <method>
+        Read one JSON object from stdin (the params for <method>), write one
+        JSON object to stdout. Method \u2208 {judge, listRubrics, version}.
+  rpc-batch <method>
+        Like 'rpc' but JSONL in / JSONL out.
+  openapi [--out openapi.json]
+        Write the OpenAPI 3.1 spec.
+  version
+        Print server + wire-protocol version JSON.
+
+Without arguments, prints this help.`;
+async function main() {
+  const { command, positional, flags } = parseArgs(process.argv.slice(2));
+  switch (command) {
+    case "serve": {
+      const port = Number(flags.port ?? 5005);
+      const host = flags.host ?? "127.0.0.1";
+      const server = startServer({ port, host });
+      const shutdown = (sig) => {
+        console.log(`[agent-eval] received ${sig}, shutting down`);
+        server.close(() => process.exit(0));
+        setTimeout(() => process.exit(1), 5e3).unref();
+      };
+      process.on("SIGINT", () => shutdown("SIGINT"));
+      process.on("SIGTERM", () => shutdown("SIGTERM"));
+      await new Promise(() => {
+      });
+      return 0;
+    }
+    case "rpc": {
+      const [method] = positional;
+      return await runRpcOnce(method);
+    }
+    case "rpc-batch": {
+      const [method] = positional;
+      return await runRpcBatch(method);
+    }
+    case "openapi": {
+      const out = flags.out ?? "openapi.json";
+      const spec = buildOpenApi(handleVersion().version);
+      writeFileSync(out, JSON.stringify(spec, null, 2) + "\n", "utf-8");
+      console.log(`[agent-eval] wrote OpenAPI 3.1 spec to ${out}`);
+      return 0;
+    }
+    case "version": {
+      process.stdout.write(JSON.stringify(handleVersion(), null, 2) + "\n");
+      return 0;
+    }
+    case "help":
+    case "--help":
+    case "-h":
+    case "":
+      process.stdout.write(HELP + "\n");
+      return 0;
+    default:
+      process.stderr.write(`unknown command: ${command}
+${HELP}
+`);
+      return 1;
+  }
+}
+main().then((code) => process.exit(code)).catch((err) => {
+  console.error("[agent-eval] cli error:", err);
+  process.exit(1);
+});
+//# sourceMappingURL=cli.js.map

package/dist/cli.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/cli.ts"],"sourcesContent":["#!/usr/bin/env node\n/**\n * agent-eval CLI.\n *\n *   agent-eval serve [--port 5005] [--host 127.0.0.1]\n *   agent-eval rpc <method>          # one request from stdin → one response on stdout\n *   agent-eval rpc-batch <method>    # JSONL stdin → JSONL stdout\n *   agent-eval openapi [--out path]  # write OpenAPI spec\n *   agent-eval version\n *\n * <method> is one of: judge, listRubrics, version. When omitted, the\n * stdin payload must be a full {method, params} envelope.\n */\nimport { writeFileSync } from 'node:fs'\n\nimport { buildOpenApi } from './wire/openapi'\nimport { handleVersion } from './wire/handlers'\nimport { runRpcBatch, runRpcOnce } from './wire/rpc'\nimport { startServer } from './wire/server'\n\ninterface Args {\n  command: string\n  positional: string[]\n  flags: Record<string, string>\n}\n\nfunction parseArgs(argv: string[]): Args {\n  const [command, ...rest] = argv\n  const positional: string[] = []\n  const flags: Record<string, string> = {}\n  for (let i = 0; i < rest.length; i++) {\n    const tok = rest[i]\n    if (tok.startsWith('--')) {\n      const key = tok.slice(2)\n      const next = rest[i + 1]\n      if (next != null && !next.startsWith('--')) {\n        flags[key] = next\n        i++\n      } else {\n        flags[key] = 'true'\n      }\n    } else {\n      positional.push(tok)\n    }\n  }\n  return { command: command ?? 'help', positional, flags }\n}\n\nconst HELP = `agent-eval — wire-protocol entry point.\n\nCommands:\n  serve [--port 5005] [--host 127.0.0.1]\n        Start the HTTP server. POST /v1/judge, GET /v1/rubrics, GET /v1/version, GET /openapi.json.\n  rpc <method>\n        Read one JSON object from stdin (the params for <method>), write one\n        JSON object to stdout. Method ∈ {judge, listRubrics, version}.\n  rpc-batch <method>\n        Like 'rpc' but JSONL in / JSONL out.\n  openapi [--out openapi.json]\n        Write the OpenAPI 3.1 spec.\n  version\n        Print server + wire-protocol version JSON.\n\nWithout arguments, prints this help.`\n\nasync function main(): Promise<number> {\n  const { command, positional, flags } = parseArgs(process.argv.slice(2))\n\n  switch (command) {\n    case 'serve': {\n      const port = Number(flags.port ?? 5005)\n      const host = flags.host ?? '127.0.0.1'\n      const server = startServer({ port, host })\n      // Keep process alive on SIGINT/SIGTERM\n      const shutdown = (sig: string) => {\n        // eslint-disable-next-line no-console\n        console.log(`[agent-eval] received ${sig}, shutting down`)\n        server.close(() => process.exit(0))\n        // Force exit after 5s if close hangs\n        setTimeout(() => process.exit(1), 5000).unref()\n      }\n      process.on('SIGINT', () => shutdown('SIGINT'))\n      process.on('SIGTERM', () => shutdown('SIGTERM'))\n      // Block forever\n      await new Promise(() => {})\n      return 0\n    }\n    case 'rpc': {\n      const [method] = positional\n      return await runRpcOnce(method)\n    }\n    case 'rpc-batch': {\n      const [method] = positional\n      return await runRpcBatch(method)\n    }\n    case 'openapi': {\n      const out = flags.out ?? 'openapi.json'\n      const spec = buildOpenApi(handleVersion().version)\n      writeFileSync(out, JSON.stringify(spec, null, 2) + '\\n', 'utf-8')\n      // eslint-disable-next-line no-console\n      console.log(`[agent-eval] wrote OpenAPI 3.1 spec to ${out}`)\n      return 0\n    }\n    case 'version': {\n      process.stdout.write(JSON.stringify(handleVersion(), null, 2) + '\\n')\n      return 0\n    }\n    case 'help':\n    case '--help':\n    case '-h':\n    case '':\n      process.stdout.write(HELP + '\\n')\n      return 0\n    default:\n      process.stderr.write(`unknown command: ${command}\\n${HELP}\\n`)\n      return 1\n  }\n}\n\nmain()\n  .then((code) => process.exit(code))\n  .catch((err) => {\n    // eslint-disable-next-line no-console\n    console.error('[agent-eval] cli error:', err)\n    process.exit(1)\n  })\n"],"mappings":";;;;;;;;;;;AAaA,SAAS,qBAAqB;AAa9B,SAAS,UAAU,MAAsB;AACvC,QAAM,CAAC,SAAS,GAAG,IAAI,IAAI;AAC3B,QAAM,aAAuB,CAAC;AAC9B,QAAM,QAAgC,CAAC;AACvC,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,IAAI,WAAW,IAAI,GAAG;AACxB,YAAM,MAAM,IAAI,MAAM,CAAC;AACvB,YAAM,OAAO,KAAK,IAAI,CAAC;AACvB,UAAI,QAAQ,QAAQ,CAAC,KAAK,WAAW,IAAI,GAAG;AAC1C,cAAM,GAAG,IAAI;AACb;AAAA,MACF,OAAO;AACL,cAAM,GAAG,IAAI;AAAA,MACf;AAAA,IACF,OAAO;AACL,iBAAW,KAAK,GAAG;AAAA,IACrB;AAAA,EACF;AACA,SAAO,EAAE,SAAS,WAAW,QAAQ,YAAY,MAAM;AACzD;AAEA,IAAM,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAiBb,eAAe,OAAwB;AACrC,QAAM,EAAE,SAAS,YAAY,MAAM,IAAI,UAAU,QAAQ,KAAK,MAAM,CAAC,CAAC;AAEtE,UAAQ,SAAS;AAAA,IACf,KAAK,SAAS;AACZ,YAAM,OAAO,OAAO,MAAM,QAAQ,IAAI;AACtC,YAAM,OAAO,MAAM,QAAQ;AAC3B,YAAM,SAAS,YAAY,EAAE,MAAM,KAAK,CAAC;AAEzC,YAAM,WAAW,CAAC,QAAgB;AAEhC,gBAAQ,IAAI,yBAAyB,GAAG,iBAAiB;AACzD,eAAO,MAAM,MAAM,QAAQ,KAAK,CAAC,CAAC;AAElC,mBAAW,MAAM,QAAQ,KAAK,CAAC,GAAG,GAAI,EAAE,MAAM;AAAA,MAChD;AACA,cAAQ,GAAG,UAAU,MAAM,SAAS,QAAQ,CAAC;AAC7C,cAAQ,GAAG,WAAW,MAAM,SAAS,SAAS,CAAC;AAE/C,YAAM,IAAI,QAAQ,MAAM;AAAA,MAAC,CAAC;AAC1B,aAAO;AAAA,IACT;AAAA,IACA,KAAK,OAAO;AACV,YAAM,CAAC,MAAM,IAAI;AACjB,aAAO,MAAM,WAAW,MAAM;AAAA,IAChC;AAAA,IACA,KAAK,aAAa;AAChB,YAAM,CAAC,MAAM,IAAI;AACjB,aAAO,MAAM,YAAY,MAAM;AAAA,IACjC;AAAA,IACA,KAAK,WAAW;AACd,YAAM,MAAM,MAAM,OAAO;AACzB,YAAM,OAAO,aAAa,cAAc,EAAE,OAAO;AACjD,oBAAc,KAAK,KAAK,UAAU,MAAM,MAAM,CAAC,IAAI,MAAM,OAAO;AAEhE,cAAQ,IAAI,0CAA0C,GAAG,EAAE;AAC3D,aAAO;AAAA,IACT;AAAA,IACA,KAAK,WAAW;AACd,cAAQ,OAAO,MAAM,KAAK,UAAU,cAAc,GAAG,MAAM,CAAC,IAAI,IAAI;AACpE,aAAO;AAAA,IACT;AAAA,IACA,KAAK;AAAA,IACL,KAAK;AAAA,IACL,KAAK;AAAA,IACL,KAAK;AACH,cAAQ,OAAO,MAAM,OAAO,IAAI;AAChC,aAAO;AAAA,IACT;AACE,cAAQ,OAAO,MAAM,oBAAoB,OAAO;AAAA,EAAK,IAAI;AAAA,CAAI;AAC7D,aAAO;AAAA,EACX;AACF;AAEA,KAAK,EACF,KAAK,CAAC,SAAS,QAAQ,KAAK,IAAI,CAAC,EACjC,MAAM,CAAC,QAAQ;AAEd,UAAQ,MAAM,2BAA2B,GAAG;AAC5C,UAAQ,KAAK,CAAC;AAChB,CAAC;","names":[]}

package/dist/index.d.ts

@@ -944,6 +944,114 @@ interface RunDiff {
     }>;
 }
 
+/**
+ * FileSystemExperimentStore — NDJSON-backed `ExperimentStore` for local + CI.
+ *
+ * Mirrors the file layout of `FileSystemTraceStore`: two append-only NDJSON
+ * files (`experiments.ndjson` + `runs.ndjson`) under one directory, with size-
+ * based rollover. Writes are append-only so the file log doubles as an audit
+ * trail of every state transition the tracker ever wrote.
+ *
+ * Reads lazy-load every NDJSON file in the directory (including rolled-over
+ * archives), latest-write-wins per `id`. Subsequent writes update the
+ * in-memory index in place so reads after writes are O(1).
+ *
+ * Node-only — imports `node:fs/promises`. Don't import this from a Worker;
+ * use the in-memory store or the D1 store from `./experiment-tracker-d1`.
+ */
+
+interface FileSystemExperimentStoreOptions {
+    /** Directory the NDJSON files live in. Created on first write. */
+    dir: string;
+    /** Bytes after which a file is rolled over. Default 32 MB (matches FileSystemTraceStore). */
+    maxBytes?: number;
+}
+declare class FileSystemExperimentStore implements ExperimentStore {
+    private readonly dir;
+    private readonly maxBytes;
+    private index?;
+    private loaded;
+    constructor(options: FileSystemExperimentStoreOptions);
+    saveExperiment(exp: Experiment): Promise<void>;
+    getExperiment(id: string): Promise<Experiment | null>;
+    listExperiments(): Promise<Experiment[]>;
+    saveRun(run: Run$1): Promise<void>;
+    getRun(id: string): Promise<Run$1 | null>;
+    listRuns(experimentId: string): Promise<Run$1[]>;
+    private ensureDir;
+    private append;
+    private load;
+}
+
+/**
+ * D1ExperimentStore — Cloudflare D1-backed `ExperimentStore`.
+ *
+ * Workers-safe (uses only the `D1Database` binding the runtime injects). Two
+ * tables, no joins, no migrations beyond `ensureSchema()`. Schema designed so
+ * a Worker route can both write the row at run start and update it at run end
+ * without losing the original config — the row's lifecycle mirrors the
+ * `Run.status` field one-to-one.
+ *
+ * Why this lives next to `InMemoryExperimentStore`:
+ *   - bad-app, legal-agent, gtm-agent, film-agent all run as Workers
+ *   - Workers cannot use `node:fs`, so `FileSystemExperimentStore` doesn't apply
+ *   - Hand-rolling D1 SQL in every consumer is exactly the duplication this
+ *     module exists to prevent
+ *
+ * Schema versioning: the `meta` table records `schema_version` so a future
+ * column addition can be detected and migrated additively. Today's schema is
+ * v1; bump only on breaking shape changes.
+ */
+
+/**
+ * Minimal `D1Database` shape we depend on. Avoids pulling in
+ * `@cloudflare/workers-types` as a hard dep — consumers that already have
+ * those types installed can pass the binding directly.
+ */
+interface D1Like {
+    prepare(query: string): D1PreparedStatementLike;
+    batch?(statements: D1PreparedStatementLike[]): Promise<unknown[]>;
+    exec(query: string): Promise<unknown>;
+}
+interface D1PreparedStatementLike {
+    bind(...values: unknown[]): D1PreparedStatementLike;
+    first<T = Record<string, unknown>>(): Promise<T | null>;
+    all<T = Record<string, unknown>>(): Promise<{
+        results: T[];
+    }>;
+    run(): Promise<unknown>;
+}
+interface D1ExperimentStoreOptions {
+    /** D1 binding from `env`. */
+    db: D1Like;
+    /**
+     * Optional table-name prefix so multiple ExperimentStores can share a DB
+     * without colliding (e.g. `tax_eval_experiments` vs `legal_eval_experiments`).
+     * Default: `agent_eval_`.
+     */
+    tablePrefix?: string;
+}
+declare class D1ExperimentStore implements ExperimentStore {
+    private readonly db;
+    private readonly experimentsTable;
+    private readonly runsTable;
+    private readonly metaTable;
+    private schemaReady;
+    constructor(options: D1ExperimentStoreOptions);
+    /**
+     * Idempotent schema setup. Safe to call before every operation; the second
+     * call short-circuits via `schemaReady`. Most consumers will call it once
+     * during Worker bootstrap.
+     */
+    ensureSchema(): Promise<void>;
+    saveExperiment(exp: Experiment): Promise<void>;
+    getExperiment(id: string): Promise<Experiment | null>;
+    listExperiments(): Promise<Experiment[]>;
+    saveRun(run: Run$1): Promise<void>;
+    getRun(id: string): Promise<Run$1 | null>;
+    listRuns(experimentId: string): Promise<Run$1[]>;
+}
+
 /**
  * Prompt optimizer — A/B test prompt variants with statistical rigor.
  *
@@ -6564,4 +6672,4 @@ interface ReflectionProposal {
  */
 declare function parseReflectionResponse(raw: string, maxProposals?: number): ReflectionProposal[];
 
-export { type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, type AlignmentOp, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type Artifact, type ArtifactCheck, type Artifact$1 as ArtifactCheckArtifact, type ArtifactResult, type ArtifactValidator, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, type BaselineOptions, type BaselineReport, BehaviorAssertion, type BenchmarkReport, BenchmarkRunner, type BenchmarkRunnerConfig, type BestOfNResult, type BisectOptions, type BisectResult, type BisectStep, type BootstrapOptions, type BootstrapResult, BudgetBreachError, type BudgetBreachFinding, type BudgetBreachReport, BudgetGuard, type BudgetLedgerEntry, type BudgetSpec, BuilderSession, type BuilderSessionInit, type CalibrationBin, type CalibrationOptions, type CalibrationReport, type CalibrationResult, CallExpectation, type CanaryLeak, type CandidateScenario, type CandidateScore, type CausalAttributionReport, type ChatSummary, type CheckResult, type CollectedArtifacts, type CommandRunner, type CompletionCriterion, type ConceptComplexity, type ConceptFinding, type ConceptSpec, type ConceptWeightStrategy, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractMetric, type ContractReport, ConvergenceTracker, type CorrelationReport, type CorrelationResult, type CorrelationStudyOptions, type CorrelationStudyResult, type CostEntry, type CostSummary, CostTracker, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateDefaultReviewerOptions, type CrossTraceDiff, type CrossTraceDiffOptions, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_RULES as DEFAULT_FAILURE_RULES, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, type DatasetDifficulty, type DatasetManifest, type DatasetProvenance, type DatasetScenario, type DatasetSplit, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DeploymentOutcome, type DirEntry, type Direction, type DivergenceOptions, type DivergenceReport, DockerSandboxDriver, type DriverResult, type DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type ErrorCountPattern, type EuRiskClass, type EvalMetricSpec, type EvalResult, type EventFilter, type EventKind, type EvolutionRound, type PromptVariant as EvolvableVariant, type ExecutorConfig, type Expectation, type Experiment, type Run$1 as ExperimentRun, type ExperimentStore, ExperimentTracker, type ExportedRewardModel, type ExtractOptions, type ExtractResult, FAILURE_CLASSES, type FactorContribution, type FactorialCell, type FailureClass, type FailureClassification, type FailureCluster, type FailureClusterReport, type FailureContext, type FailureRule, type FeedbackPattern, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, FileSystemTraceStore, type FileSystemTraceStoreOptions, type Finding, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GenerationReport, type GenericSpan, type GoldenItem, type GoldenSeverity, type GoldenSpec, type GovernanceContext, type GovernanceFinding, type GovernanceReport, type GradedStep, type HarnessAdapter, type HarnessConfig, type HarnessExperimentConfig, type HarnessExperimentResult, type HarnessIntervention, type HarnessRunRequest, type HarnessRunResult, type HarnessScenario, type HarnessSelection, type HarnessVariant, type HarnessVariantReport, HoldoutAuditor, HoldoutLockedError, type HostedJudgeConfig, type HostedJudgeDimension, type HostedJudgeRequest, type HostedJudgeResponse, type HostedRunCriticConfig, type HostedRunScoreRequest, type HostedRunScoreResponse, type HypothesisManifest, type HypothesisResult, INTENT_MATCH_JUDGE_VERSION, type ImageData, InMemoryExperimentStore, InMemoryOutcomeStore, InMemoryTraceStore, InMemoryTrialCache, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, type JudgeAgreementReport, type JudgeConfig, type JudgeFleetOptions, type JudgeFn, type JudgeInput, type JudgePair, type JudgeReplayGateArgs, type JudgeReplayResult, type JudgeRubric, JudgeRunner, type JudgeScore, type JudgeSpan, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, type Layer, type LayerCorrelation, type LayerResult, type LayerStatus, LlmCallError, type LlmCallRequest, type LlmCallResult, LlmClient, type LlmClientOptions, type LlmJsonCall, type LlmMessage, type LlmReviewerConfig, type LlmSpan, type LlmUsage, MODEL_PRICING, type MatchResult, type MatcherResult, type MeasurementPolicy, type MergeOptions, type Message, type MetricSamples, type MetricVerdict, MetricsCollector, type MuffledFinder, type MuffledFinding, MultiLayerVerifier, type MultiToolchainLayerConfig, type MutateAdapter, type Mutator, OTEL_AGENT_EVAL_SCOPE, type Objective, type OptimizationConfig, type OptimizationExample, OptimizationLoop, type OptimizationLoopConfig, type OptimizationLoopResult, type OptimizationResult, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, type OtlpExport, type OtlpResourceSpans, type OtlpSpan, type OutcomeFilter, type OutcomePair, type OutcomeStore, type PairwiseComparison, PairwiseSteeringOptimizer, type ParetoResult, type PersonaConfig, type Playbook, type PlaybookEntry, type PositionalBiasResult, type PrmGradedTrace, PrmGrader, type PrmTrainingSample, ProductClient, type ProductClientConfig, type ProjectKind, ProjectRegistry, type ProjectSummary, type ProjectTimelineEntry, type PromptEvolutionConfig, type PromptEvolutionEvent, type PromptEvolutionResult, type PromptHandle, PromptOptimizer, PromptRegistry, type TrialResult as PromptTrialResult, type PromptVariant$1 as PromptVariant, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewReport, type ProposeReviewShot, REDACTION_VERSION, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamPayload, type RedTeamReport, type RedactionReport, type RedactionRule, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, type RegressionOptions, type RegressionSpec, type RetrievalSpan, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, type RouteMap, type RubricDimension, type Run, type RunAppScenarioOptions, type RunCommandInput, type RunCommandResult, type RunConfig, RunCritic, type RunCriticOptions, type RunDiff, type RunFilter, type RunLayer, type RunOutcome, type RunScore, type RunScoreWeights, type RunStatus, type RunTrace, SEMANTIC_CONCEPT_JUDGE_VERSION, type SandboxDriver, SandboxHarness, type SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxResult, type SandboxSpan, type ScanOptions, type Scenario, type ScenarioAggregate, type ScenarioCost, type ScenarioFile, ScenarioRegistry, type ScenarioResult, type ScoreAdapter, type ScoredTarget, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SelfPreferenceResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SeriesConvergenceOptions, type SeriesConvergenceResult, type Severity, type ShipOptions, type SignedManifest, type SliceOptions, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type Span, type SpanBase, type SpanFilter, type SpanHandle, type SpanKind, type SpanStatus, type SteeringBundle, type SteeringDelta, type SteeringEvaluation, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type SteeringRolePrompt, type SteeringVariantReport, type StepAttribution, type StepContext, type StepRubric, type StuckLoopFinding, type StuckLoopOptions, type StuckLoopReport, SubprocessSandboxDriver, type SubprocessSandboxDriverOptions, type SynthesisReason, type SynthesisTarget, TRACE_SCHEMA_VERSION, type TestGradedRunOptions, type TestGradedRunResult, type TestGradedScenario, type TestOutputParser, type TestResult, type ThreeLayerProjectReport, type ThresholdContract, TokenCounter, type TokenSpec, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type ToolWasteFinding, type ToolWasteOptions, type ToolWasteReport, TraceEmitter, type TraceEmitterOptions, type TraceEvent, type TraceStore, type Trajectory, type TrajectoryStep, type TrialCache, type TrialTrace, type Turn, type TurnMetrics, type TurnResult, UNIVERSAL_FINDERS, type UseCaseSignals, type ValidationContext, type ValidationIssue, type ValidationResult, type VariantAggregate, type VariantScore, type VerbosityBiasResult, type Verdict, type Verification, type VerificationReport, type VerifyContext, type VerifyFn, type VerifyOptions, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkflowTopology, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, adversarialJudge, aggregateLlm, aggregateRunScore, analyzeAntiSlop, analyzeSeries, argHash, attributeCounterfactuals, benjaminiHochberg, bisect, bonferroni, bootstrapCi, budgetBreachView, buildReflectionPrompt, buildReviewerPrompt, buildTrajectory, byteLengthRange, calibrateJudge, calibrationCurve, callLlm, callLlmJson, canaryLeakView, causalAttribution, checkCanaries, checkSlos, clamp01, classifyEuAiRisk, classifyFailure, codeExecutionJudge, cohensD, coherenceJudge, collectionPreserved, commitBisect, compareReferenceReplay, compareToBaseline, compilerJudge, composeParsers, composeValidators, computeToolUseMetrics, confidenceInterval, containsAll, correlateLayers, correlationStudy, createAntiSlopJudge, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createLlmReviewer, createSemanticConceptJudge, crossTraceDiff, crowdingDistance, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultJudges, defaultReferenceReplayMatcher, deployGateLayer, distillPlaybook, dominates, estimateCost, estimateTokens, euAiActReport, evaluateContract, evaluateHypothesis, evaluateOracles, executeScenario, expectAgent, exportRewardModel, exportRunAsOtlp, exportTrainingData, extractAssetUrls, extractErrorCount, failureClusterView, fileContains, fileExists, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, firstDivergenceView, flowLayer, formatBenchmarkReport, formatDriverReport, formatFindings, precision as goldenPrecision, gradeSemanticStatus, groupBy, hashContent, hashScenarios, htmlContainsElement, inMemoryReferenceReplayStore, inMemoryReviewStore, interRaterReliability, iqr, isJudgeSpan, isLlmSpan, isPrmVerdict, isRetrievalSpan, isSandboxSpan, isToolSpan, jestTestParser, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, jsonlReviewStore, judgeAgreementView, judgeReplayGate, judgeSpans, keyPreserved, linterJudge, llmSpanFromProvider, llmSpans, loadScorerFromGrader, localCommandRunner, lowercaseMutator, mannWhitneyU, matchGoldens, mergeLayerResults, mergeSteeringBundle, multiToolchainLayer, nistAiRmfReport, nonRefusalRubric, normalizeScores, notBlocked, outputLengthRubric, pairedTTest, paraphraseRobustness, paretoFrontier, paretoFrontierWithCrowding, parseReflectionResponse, partialCredit, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, positionalBias, printDriverSummary, prmBestOfN, prmEnsembleBestOfN, probeLlm, promptBisect, proposeSynthesisTargets, pytestTestParser, redTeamDataset, redTeamReport, redactString, redactValue, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, regressionView, renderMarkdown, renderMarkdownReport, renderPlaybookMarkdown, renderSteeringText, replayScorerOverCorpus, replayTraceThroughJudge, requiredSampleSize, resumeBuilderSession, rowCount, rowWhere, runAssertions, runCounterfactual, runE2EWorkflow, runExpectations, runFailureClass, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runPromptEvolution, runProposeReview, runReferenceReplay, runSelfPlay, runSemanticConceptJudge, runTestGradedScenario, runsForScenario, scalarScore, scanForMuffledGates, scoreAllProjects, scoreContinuity, scoreProject, scoreRedTeamOutput, scoreReferenceReplay, securityJudge, selectHarnessVariant, selfPreference, sentenceReorderMutator, signManifest, soc2Report, statusAdvanced, stripFencedJson, stuckLoopView, summarize, summarizeHarnessResults, testJudge, textInSnapshot, toLangfuseEnvelope, toNdjson, toPrometheusText, toolIntentAlignmentRubric, toolNamesForRun, toolNonRedundantRubric, toolSpans, toolSuccessRubric, toolWasteView, typoMutator, urlContains, verbosityBias, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank };
+export { type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, type AlignmentOp, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type Artifact, type ArtifactCheck, type Artifact$1 as ArtifactCheckArtifact, type ArtifactResult, type ArtifactValidator, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, type BaselineOptions, type BaselineReport, BehaviorAssertion, type BenchmarkReport, BenchmarkRunner, type BenchmarkRunnerConfig, type BestOfNResult, type BisectOptions, type BisectResult, type BisectStep, type BootstrapOptions, type BootstrapResult, BudgetBreachError, type BudgetBreachFinding, type BudgetBreachReport, BudgetGuard, type BudgetLedgerEntry, type BudgetSpec, BuilderSession, type BuilderSessionInit, type CalibrationBin, type CalibrationOptions, type CalibrationReport, type CalibrationResult, CallExpectation, type CanaryLeak, type CandidateScenario, type CandidateScore, type CausalAttributionReport, type ChatSummary, type CheckResult, type CollectedArtifacts, type CommandRunner, type CompletionCriterion, type ConceptComplexity, type ConceptFinding, type ConceptSpec, type ConceptWeightStrategy, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractMetric, type ContractReport, ConvergenceTracker, type CorrelationReport, type CorrelationResult, type CorrelationStudyOptions, type CorrelationStudyResult, type CostEntry, type CostSummary, CostTracker, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateDefaultReviewerOptions, type CrossTraceDiff, type CrossTraceDiffOptions, D1ExperimentStore, type D1ExperimentStoreOptions, type D1Like, type D1PreparedStatementLike, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_RULES as DEFAULT_FAILURE_RULES, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, type DatasetDifficulty, type DatasetManifest, type DatasetProvenance, type DatasetScenario, type DatasetSplit, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DeploymentOutcome, type DirEntry, type Direction, type DivergenceOptions, type DivergenceReport, DockerSandboxDriver, type DriverResult, type DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type ErrorCountPattern, type EuRiskClass, type EvalMetricSpec, type EvalResult, type EventFilter, type EventKind, type EvolutionRound, type PromptVariant as EvolvableVariant, type ExecutorConfig, type Expectation, type Experiment, type Run$1 as ExperimentRun, type ExperimentStore, ExperimentTracker, type ExportedRewardModel, type ExtractOptions, type ExtractResult, FAILURE_CLASSES, type FactorContribution, type FactorialCell, type FailureClass, type FailureClassification, type FailureCluster, type FailureClusterReport, type FailureContext, type FailureRule, type FeedbackPattern, FileSystemExperimentStore, type FileSystemExperimentStoreOptions, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, FileSystemTraceStore, type FileSystemTraceStoreOptions, type Finding, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GenerationReport, type GenericSpan, type GoldenItem, type GoldenSeverity, type GoldenSpec, type GovernanceContext, type GovernanceFinding, type GovernanceReport, type GradedStep, type HarnessAdapter, type HarnessConfig, type HarnessExperimentConfig, type HarnessExperimentResult, type HarnessIntervention, type HarnessRunRequest, type HarnessRunResult, type HarnessScenario, type HarnessSelection, type HarnessVariant, type HarnessVariantReport, HoldoutAuditor, HoldoutLockedError, type HostedJudgeConfig, type HostedJudgeDimension, type HostedJudgeRequest, type HostedJudgeResponse, type HostedRunCriticConfig, type HostedRunScoreRequest, type HostedRunScoreResponse, type HypothesisManifest, type HypothesisResult, INTENT_MATCH_JUDGE_VERSION, type ImageData, InMemoryExperimentStore, InMemoryOutcomeStore, InMemoryTraceStore, InMemoryTrialCache, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, type JudgeAgreementReport, type JudgeConfig, type JudgeFleetOptions, type JudgeFn, type JudgeInput, type JudgePair, type JudgeReplayGateArgs, type JudgeReplayResult, type JudgeRubric, JudgeRunner, type JudgeScore, type JudgeSpan, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, type Layer, type LayerCorrelation, type LayerResult, type LayerStatus, LlmCallError, type LlmCallRequest, type LlmCallResult, LlmClient, type LlmClientOptions, type LlmJsonCall, type LlmMessage, type LlmReviewerConfig, type LlmSpan, type LlmUsage, MODEL_PRICING, type MatchResult, type MatcherResult, type MeasurementPolicy, type MergeOptions, type Message, type MetricSamples, type MetricVerdict, MetricsCollector, type MuffledFinder, type MuffledFinding, MultiLayerVerifier, type MultiToolchainLayerConfig, type MutateAdapter, type Mutator, OTEL_AGENT_EVAL_SCOPE, type Objective, type OptimizationConfig, type OptimizationExample, OptimizationLoop, type OptimizationLoopConfig, type OptimizationLoopResult, type OptimizationResult, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, type OtlpExport, type OtlpResourceSpans, type OtlpSpan, type OutcomeFilter, type OutcomePair, type OutcomeStore, type PairwiseComparison, PairwiseSteeringOptimizer, type ParetoResult, type PersonaConfig, type Playbook, type PlaybookEntry, type PositionalBiasResult, type PrmGradedTrace, PrmGrader, type PrmTrainingSample, ProductClient, type ProductClientConfig, type ProjectKind, ProjectRegistry, type ProjectSummary, type ProjectTimelineEntry, type PromptEvolutionConfig, type PromptEvolutionEvent, type PromptEvolutionResult, type PromptHandle, PromptOptimizer, PromptRegistry, type TrialResult as PromptTrialResult, type PromptVariant$1 as PromptVariant, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewReport, type ProposeReviewShot, REDACTION_VERSION, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamPayload, type RedTeamReport, type RedactionReport, type RedactionRule, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, type RegressionOptions, type RegressionSpec, type RetrievalSpan, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, type RouteMap, type RubricDimension, type Run, type RunAppScenarioOptions, type RunCommandInput, type RunCommandResult, type RunConfig, RunCritic, type RunCriticOptions, type RunDiff, type RunFilter, type RunLayer, type RunOutcome, type RunScore, type RunScoreWeights, type RunStatus, type RunTrace, SEMANTIC_CONCEPT_JUDGE_VERSION, type SandboxDriver, SandboxHarness, type SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxResult, type SandboxSpan, type ScanOptions, type Scenario, type ScenarioAggregate, type ScenarioCost, type ScenarioFile, ScenarioRegistry, type ScenarioResult, type ScoreAdapter, type ScoredTarget, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SelfPreferenceResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SeriesConvergenceOptions, type SeriesConvergenceResult, type Severity, type ShipOptions, type SignedManifest, type SliceOptions, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type Span, type SpanBase, type SpanFilter, type SpanHandle, type SpanKind, type SpanStatus, type SteeringBundle, type SteeringDelta, type SteeringEvaluation, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type SteeringRolePrompt, type SteeringVariantReport, type StepAttribution, type StepContext, type StepRubric, type StuckLoopFinding, type StuckLoopOptions, type StuckLoopReport, SubprocessSandboxDriver, type SubprocessSandboxDriverOptions, type SynthesisReason, type SynthesisTarget, TRACE_SCHEMA_VERSION, type TestGradedRunOptions, type TestGradedRunResult, type TestGradedScenario, type TestOutputParser, type TestResult, type ThreeLayerProjectReport, type ThresholdContract, TokenCounter, type TokenSpec, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type ToolWasteFinding, type ToolWasteOptions, type ToolWasteReport, TraceEmitter, type TraceEmitterOptions, type TraceEvent, type TraceStore, type Trajectory, type TrajectoryStep, type TrialCache, type TrialTrace, type Turn, type TurnMetrics, type TurnResult, UNIVERSAL_FINDERS, type UseCaseSignals, type ValidationContext, type ValidationIssue, type ValidationResult, type VariantAggregate, type VariantScore, type VerbosityBiasResult, type Verdict, type Verification, type VerificationReport, type VerifyContext, type VerifyFn, type VerifyOptions, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkflowTopology, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, adversarialJudge, aggregateLlm, aggregateRunScore, analyzeAntiSlop, analyzeSeries, argHash, attributeCounterfactuals, benjaminiHochberg, bisect, bonferroni, bootstrapCi, budgetBreachView, buildReflectionPrompt, buildReviewerPrompt, buildTrajectory, byteLengthRange, calibrateJudge, calibrationCurve, callLlm, callLlmJson, canaryLeakView, causalAttribution, checkCanaries, checkSlos, clamp01, classifyEuAiRisk, classifyFailure, codeExecutionJudge, cohensD, coherenceJudge, collectionPreserved, commitBisect, compareReferenceReplay, compareToBaseline, compilerJudge, composeParsers, composeValidators, computeToolUseMetrics, confidenceInterval, containsAll, correlateLayers, correlationStudy, createAntiSlopJudge, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createLlmReviewer, createSemanticConceptJudge, crossTraceDiff, crowdingDistance, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultJudges, defaultReferenceReplayMatcher, deployGateLayer, distillPlaybook, dominates, estimateCost, estimateTokens, euAiActReport, evaluateContract, evaluateHypothesis, evaluateOracles, executeScenario, expectAgent, exportRewardModel, exportRunAsOtlp, exportTrainingData, extractAssetUrls, extractErrorCount, failureClusterView, fileContains, fileExists, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, firstDivergenceView, flowLayer, formatBenchmarkReport, formatDriverReport, formatFindings, precision as goldenPrecision, gradeSemanticStatus, groupBy, hashContent, hashScenarios, htmlContainsElement, inMemoryReferenceReplayStore, inMemoryReviewStore, interRaterReliability, iqr, isJudgeSpan, isLlmSpan, isPrmVerdict, isRetrievalSpan, isSandboxSpan, isToolSpan, jestTestParser, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, jsonlReviewStore, judgeAgreementView, judgeReplayGate, judgeSpans, keyPreserved, linterJudge, llmSpanFromProvider, llmSpans, loadScorerFromGrader, localCommandRunner, lowercaseMutator, mannWhitneyU, matchGoldens, mergeLayerResults, mergeSteeringBundle, multiToolchainLayer, nistAiRmfReport, nonRefusalRubric, normalizeScores, notBlocked, outputLengthRubric, pairedTTest, paraphraseRobustness, paretoFrontier, paretoFrontierWithCrowding, parseReflectionResponse, partialCredit, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, positionalBias, printDriverSummary, prmBestOfN, prmEnsembleBestOfN, probeLlm, promptBisect, proposeSynthesisTargets, pytestTestParser, redTeamDataset, redTeamReport, redactString, redactValue, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, regressionView, renderMarkdown, renderMarkdownReport, renderPlaybookMarkdown, renderSteeringText, replayScorerOverCorpus, replayTraceThroughJudge, requiredSampleSize, resumeBuilderSession, rowCount, rowWhere, runAssertions, runCounterfactual, runE2EWorkflow, runExpectations, runFailureClass, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runPromptEvolution, runProposeReview, runReferenceReplay, runSelfPlay, runSemanticConceptJudge, runTestGradedScenario, runsForScenario, scalarScore, scanForMuffledGates, scoreAllProjects, scoreContinuity, scoreProject, scoreRedTeamOutput, scoreReferenceReplay, securityJudge, selectHarnessVariant, selfPreference, sentenceReorderMutator, signManifest, soc2Report, statusAdvanced, stripFencedJson, stuckLoopView, summarize, summarizeHarnessResults, testJudge, textInSnapshot, toLangfuseEnvelope, toNdjson, toPrometheusText, toolIntentAlignmentRubric, toolNamesForRun, toolNonRedundantRubric, toolSpans, toolSuccessRubric, toolWasteView, typoMutator, urlContains, verbosityBias, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank };