@tangle-network/agent-eval 0.93.0 → 0.95.0

package/dist/chunk-T375SUOZ.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/sandbox-harness.ts","../src/test-graded-scenario.ts"],"sourcesContent":["/**\n * SandboxHarness — executes a scenario in an isolated environment and\n * emits a rich SandboxSpan into the trace.\n *\n * Two built-in drivers:\n *   - `SubprocessSandboxDriver` — spawn in a local cwd with env vars.\n *     Fast, no dependencies, fine for unit tests and most CI gates.\n *   - `DockerSandboxDriver` — lifted from tangle-router's sandbox path;\n *     shells out to `docker run`. Stronger isolation, slower startup.\n *\n * Consumers implement `SandboxDriver` for custom backends (Firecracker,\n * Cloudflare sandbox product, etc.). The harness doesn't care which.\n */\n\nimport { ConfigError } from './errors'\nimport type { TraceEmitter } from './trace/emitter'\nimport type { SandboxSpan } from './trace/schema'\n\nexport interface HarnessConfig {\n  /** Setup command (e.g. \"pnpm install\"). Non-zero exit fails the run. */\n  setupCommand?: string\n  /** Run command (e.g. \"pnpm build\"). */\n  runCommand?: string\n  /** Test command (e.g. \"pnpm test --run\"). Drives the test count + pass count. */\n  testCommand?: string\n  /** Absolute cwd for the subprocess driver. Ignored by docker driver. */\n  cwd?: string\n  /** Max wall-clock per phase in ms. Default 10 minutes. */\n  timeoutMs?: number\n  /** Image for the docker driver. */\n  image?: string\n  /** Extra env vars (validated; shell-escaped). */\n  env?: Record<string, string>\n  /** Parser for the test output — maps stdout/stderr/exit code → pass count. */\n  testParser?: TestOutputParser\n}\n\nexport interface TestOutputParser {\n  id: string\n  parse(\n    stdout: string,\n    stderr: string,\n    exitCode: number,\n  ): { testsTotal: number; testsPassed: number } | undefined\n}\n\nexport interface SandboxResult {\n  phase: 'setup' | 'run' | 'test'\n  exitCode: number\n  stdout: string\n  stderr: string\n  wallMs: number\n  testsTotal?: number\n  testsPassed?: number\n}\n\nexport interface SandboxDriver {\n  id: string\n  exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult>\n}\n\n// ── Parsers ──────────────────────────────────────────────────────────\n\n/** Vitest default summary line: \"Tests  X passed | Y failed\". */\nexport const vitestTestParser: TestOutputParser = {\n  id: 'vitest',\n  parse(stdout) {\n    const m = stdout.match(/Tests\\s+(\\d+)\\s+(passed|failed)(?:\\s*\\|\\s*(\\d+)\\s+(passed|failed))?/i)\n    if (!m) return undefined\n    let passed = 0\n    let failed = 0\n    const a = parseInt(m[1]!, 10)\n    const aLabel = m[2]!.toLowerCase()\n    if (aLabel === 'passed') passed += a\n    else failed += a\n    if (m[3] && m[4]) {\n      const b = parseInt(m[3], 10)\n      if (m[4].toLowerCase() === 'passed') passed += b\n      else failed += b\n    }\n    return { testsTotal: passed + failed, testsPassed: passed }\n  },\n}\n\n/** Pytest default: \"collected N items\" + \" X passed, Y failed\". */\nexport const pytestTestParser: TestOutputParser = {\n  id: 'pytest',\n  parse(stdout) {\n    const total = stdout.match(/collected\\s+(\\d+)\\s+items?/i)\n    const passed = stdout.match(/(\\d+)\\s+passed/)\n    if (!total || !passed) return undefined\n    return { testsTotal: parseInt(total[1]!, 10), testsPassed: parseInt(passed[1]!, 10) }\n  },\n}\n\n/** Jest: \"Tests: X passed, Y total\" (and optional failed). */\nexport const jestTestParser: TestOutputParser = {\n  id: 'jest',\n  parse(stdout) {\n    const m = stdout.match(/Tests:\\s+(?:(\\d+)\\s+failed[^,]*,\\s*)?(\\d+)\\s+passed,\\s+(\\d+)\\s+total/i)\n    if (!m) return undefined\n    return { testsTotal: parseInt(m[3]!, 10), testsPassed: parseInt(m[2]!, 10) }\n  },\n}\n\n/** Composite parser — tries a list of parsers in order. */\nexport function composeParsers(...parsers: TestOutputParser[]): TestOutputParser {\n  return {\n    id: parsers.map((p) => p.id).join('|'),\n    parse(stdout, stderr, exitCode) {\n      for (const p of parsers) {\n        const res = p.parse(stdout, stderr, exitCode)\n        if (res) return res\n      }\n      return undefined\n    },\n  }\n}\n\n// ── Drivers ──────────────────────────────────────────────────────────\n\nexport interface SubprocessSandboxDriverOptions {\n  /**\n   * Default cwd for all `exec` calls. Used when the per-call `HarnessConfig`\n   * does not set its own `cwd`. Lets callers bind the driver to a working\n   * directory once instead of spreading cwd into every harness config —\n   * useful when the harness config is constructed far from the call site\n   * (e.g. starter-foundry's promoter passes a static HarnessConfig per\n   * family taxonomy but needs a per-run composed-scaffold cwd).\n   */\n  cwd?: string\n  /**\n   * Default env merged into every `exec` call's env (per-call `HarnessConfig.env`\n   * still wins on key collision). Same ergonomic rationale as `cwd` above.\n   */\n  env?: Record<string, string>\n}\n\nexport class SubprocessSandboxDriver implements SandboxDriver {\n  id = 'subprocess'\n  private defaultCwd?: string\n  private defaultEnv?: Record<string, string>\n\n  constructor(options: SubprocessSandboxDriverOptions = {}) {\n    this.defaultCwd = options.cwd\n    this.defaultEnv = options.env\n  }\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    const { spawn } = await import('node:child_process')\n    const start = Date.now()\n    // Per-call config wins; fall back to constructor defaults. Honoring\n    // the constructor `cwd` keeps the subprocess from inheriting Node's\n    // cwd when only the constructor arg is supplied.\n    const effectiveCwd = config.cwd ?? this.defaultCwd\n    const effectiveEnv = { ...process.env, ...(this.defaultEnv ?? {}), ...(config.env ?? {}) }\n    return await new Promise<SandboxResult>((resolve) => {\n      const child = spawn(command, {\n        shell: true,\n        cwd: effectiveCwd,\n        env: effectiveEnv,\n      })\n      let stdout = ''\n      let stderr = ''\n      child.stdout?.on('data', (d) => {\n        stdout += String(d)\n      })\n      child.stderr?.on('data', (d) => {\n        stderr += String(d)\n      })\n      const timeout = setTimeout(\n        () => {\n          try {\n            child.kill('SIGKILL')\n          } catch (err) {\n            console.warn('[sandbox-harness] SIGKILL on timeout failed:', err)\n          }\n        },\n        config.timeoutMs ?? 10 * 60_000,\n      )\n      child.on('close', (code) => {\n        clearTimeout(timeout)\n        const wallMs = Date.now() - start\n        const parsed =\n          phase === 'test' && config.testParser\n            ? config.testParser.parse(stdout, stderr, code ?? 1)\n            : undefined\n        resolve({\n          phase,\n          exitCode: code ?? 1,\n          stdout,\n          stderr,\n          wallMs,\n          testsTotal: parsed?.testsTotal,\n          testsPassed: parsed?.testsPassed,\n        })\n      })\n      child.on('error', (err) => {\n        clearTimeout(timeout)\n        const wallMs = Date.now() - start\n        resolve({ phase, exitCode: 127, stdout, stderr: stderr + String(err), wallMs })\n      })\n    })\n  }\n}\n\nexport class DockerSandboxDriver implements SandboxDriver {\n  id = 'docker'\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    if (!config.image) throw new ConfigError('DockerSandboxDriver requires config.image')\n    const sub = new SubprocessSandboxDriver()\n    const envArgs = Object.entries(config.env ?? {})\n      .map(([k, v]) => `-e ${shellQuote(k)}=${shellQuote(v)}`)\n      .join(' ')\n    const wrapped = `docker run --rm ${envArgs} ${shellQuote(config.image)} sh -c ${shellQuote(command)}`\n    return sub.exec(phase, wrapped, { ...config, env: undefined })\n  }\n}\n\nfunction shellQuote(v: string): string {\n  if (/^[A-Za-z0-9_\\-/.@:=]+$/.test(v)) return v\n  return `'${v.replace(/'/g, `'\\\\''`)}'`\n}\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  private driver: SandboxDriver\n  constructor(driver: SandboxDriver = new SubprocessSandboxDriver()) {\n    this.driver = driver\n  }\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${this.driver.id})`,\n      image: config.image,\n      command: [config.setupCommand, config.runCommand, config.testCommand]\n        .filter(Boolean)\n        .join(' && '),\n    })\n    const result: SandboxHarnessResult = { passed: false, totalWallMs: 0, score: 0 }\n    try {\n      if (config.setupCommand) {\n        result.setup = await this.driver.exec('setup', config.setupCommand, config)\n        result.totalWallMs += result.setup.wallMs\n        if (result.setup.exitCode !== 0) {\n          await handle.fail(`setup failed (exit ${result.setup.exitCode})`, {\n            exitCode: result.setup.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.runCommand) {\n        result.run = await this.driver.exec('run', config.runCommand, config)\n        result.totalWallMs += result.run.wallMs\n        if (result.run.exitCode !== 0) {\n          await handle.fail(`run failed (exit ${result.run.exitCode})`, {\n            exitCode: result.run.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.testCommand) {\n        result.test = await this.driver.exec('test', config.testCommand, config)\n        result.totalWallMs += result.test.wallMs\n        const passed = result.test.exitCode === 0\n        result.passed = passed\n        if (result.test.testsTotal !== undefined && result.test.testsTotal > 0) {\n          result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal\n        } else {\n          result.score = passed ? 1 : 0\n        }\n        await handle.end({\n          exitCode: result.test.exitCode,\n          testsTotal: result.test.testsTotal,\n          testsPassed: result.test.testsPassed,\n          wallMs: result.totalWallMs,\n          status: passed ? 'ok' : 'error',\n        } as Partial<SandboxSpan>)\n      } else {\n        result.passed = true\n        result.score = 1\n        await handle.end({ wallMs: result.totalWallMs } as Partial<SandboxSpan>)\n      }\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n    return result\n  }\n}\n","/**\n * TestGradedScenario — a scenario whose score comes from a test suite.\n *\n * This is the SWE-bench pattern generalized. The scenario ships:\n *   - fixture data (setup instructions)\n *   - a test command the harness runs\n *   - optional assertion overrides\n *\n * The runner emits a run, delegates to SandboxHarness, records the\n * outcome, and returns a structured verdict. Consumers bind their own\n * agent execution to this contract.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from './sandbox-harness'\nimport { SandboxHarness } from './sandbox-harness'\nimport { TraceEmitter } from './trace/emitter'\nimport type { FailureClass, Run } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface TestGradedScenario {\n  id: string\n  description?: string\n  harness: HarnessConfig\n  /** Optional pass threshold in 0..1 (default 1.0 = all tests must pass). */\n  passThreshold?: number\n  /** Provenance for dataset tracking. */\n  datasetVersion?: string\n  /** Free-form tags (difficulty, category, etc.). */\n  tags?: Record<string, string>\n}\n\nexport interface TestGradedRunOptions {\n  variantId?: string\n  driver?: SandboxDriver\n  /** Metadata recorded on the Run (codeSha, promptSha, modelFingerprint, seed). */\n  provenance?: Pick<Run, 'codeSha' | 'promptSha' | 'modelFingerprint' | 'seed' | 'envFingerprint'>\n}\n\nexport interface TestGradedRunResult {\n  runId: string\n  scenario: TestGradedScenario\n  harness: SandboxHarnessResult\n  pass: boolean\n  score: number\n  failureClass?: FailureClass\n}\n\nexport async function runTestGradedScenario(\n  scenario: TestGradedScenario,\n  store: TraceStore,\n  options: TestGradedRunOptions = {},\n): Promise<TestGradedRunResult> {\n  const emitter = new TraceEmitter(store)\n  await emitter.startRun({\n    scenarioId: scenario.id,\n    variantId: options.variantId,\n    datasetVersion: scenario.datasetVersion,\n    tags: scenario.tags,\n    ...options.provenance,\n  })\n  const harness = new SandboxHarness(options.driver)\n  const result = await harness.run(scenario.harness, emitter)\n  const threshold = scenario.passThreshold ?? 1.0\n  const pass = result.passed && result.score >= threshold\n  const setupFailed = result.setup !== undefined && result.setup.exitCode !== 0\n  const runFailed = result.run !== undefined && result.run.exitCode !== 0\n  const testFailed = result.test !== undefined && result.test.exitCode !== 0\n  const failureClass: FailureClass | undefined = pass\n    ? 'success'\n    : setupFailed || runFailed\n      ? 'sandbox_failure'\n      : testFailed\n        ? 'format_drift'\n        : 'unknown'\n  await emitter.endRun({\n    pass,\n    score: result.score,\n    failureClass,\n    notes: pass ? undefined : reasonForFailure(result),\n  })\n  return {\n    runId: emitter.runId,\n    scenario,\n    harness: result,\n    pass,\n    score: result.score,\n    failureClass,\n  }\n}\n\nfunction reasonForFailure(result: SandboxHarnessResult): string {\n  if (result.setup && result.setup.exitCode !== 0)\n    return `setup failed: exit ${result.setup.exitCode}`\n  if (result.run && result.run.exitCode !== 0) return `run failed: exit ${result.run.exitCode}`\n  if (result.test) {\n    if (result.test.testsTotal !== undefined) {\n      return `tests: ${result.test.testsPassed ?? 0}/${result.test.testsTotal}`\n    }\n    return `test exit ${result.test.exitCode}`\n  }\n  return 'no test command'\n}\n"],"mappings":";;;;;;;;AAoEO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,sEAAsE;AAC7F,QAAI,CAAC,EAAG,QAAO;AACf,QAAI,SAAS;AACb,QAAI,SAAS;AACb,UAAM,IAAI,SAAS,EAAE,CAAC,GAAI,EAAE;AAC5B,UAAM,SAAS,EAAE,CAAC,EAAG,YAAY;AACjC,QAAI,WAAW,SAAU,WAAU;AAAA,QAC9B,WAAU;AACf,QAAI,EAAE,CAAC,KAAK,EAAE,CAAC,GAAG;AAChB,YAAM,IAAI,SAAS,EAAE,CAAC,GAAG,EAAE;AAC3B,UAAI,EAAE,CAAC,EAAE,YAAY,MAAM,SAAU,WAAU;AAAA,UAC1C,WAAU;AAAA,IACjB;AACA,WAAO,EAAE,YAAY,SAAS,QAAQ,aAAa,OAAO;AAAA,EAC5D;AACF;AAGO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,QAAQ,OAAO,MAAM,6BAA6B;AACxD,UAAM,SAAS,OAAO,MAAM,gBAAgB;AAC5C,QAAI,CAAC,SAAS,CAAC,OAAQ,QAAO;AAC9B,WAAO,EAAE,YAAY,SAAS,MAAM,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,OAAO,CAAC,GAAI,EAAE,EAAE;AAAA,EACtF;AACF;AAGO,IAAM,iBAAmC;AAAA,EAC9C,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,uEAAuE;AAC9F,QAAI,CAAC,EAAG,QAAO;AACf,WAAO,EAAE,YAAY,SAAS,EAAE,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,EAAE,CAAC,GAAI,EAAE,EAAE;AAAA,EAC7E;AACF;AAGO,SAAS,kBAAkB,SAA+C;AAC/E,SAAO;AAAA,IACL,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK,GAAG;AAAA,IACrC,MAAM,QAAQ,QAAQ,UAAU;AAC9B,iBAAW,KAAK,SAAS;AACvB,cAAM,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ;AAC5C,YAAI,IAAK,QAAO;AAAA,MAClB;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAqBO,IAAM,0BAAN,MAAuD;AAAA,EAC5D,KAAK;AAAA,EACG;AAAA,EACA;AAAA,EAER,YAAY,UAA0C,CAAC,GAAG;AACxD,SAAK,aAAa,QAAQ;AAC1B,SAAK,aAAa,QAAQ;AAAA,EAC5B;AAAA,EAEA,MAAM,KACJ,OACA,SACA,QACwB;AACxB,UAAM,EAAE,MAAM,IAAI,MAAM,OAAO,eAAoB;AACnD,UAAM,QAAQ,KAAK,IAAI;AAIvB,UAAM,eAAe,OAAO,OAAO,KAAK;AACxC,UAAM,eAAe,EAAE,GAAG,QAAQ,KAAK,GAAI,KAAK,cAAc,CAAC,GAAI,GAAI,OAAO,OAAO,CAAC,EAAG;AACzF,WAAO,MAAM,IAAI,QAAuB,CAAC,YAAY;AACnD,YAAM,QAAQ,MAAM,SAAS;AAAA,QAC3B,OAAO;AAAA,QACP,KAAK;AAAA,QACL,KAAK;AAAA,MACP,CAAC;AACD,UAAI,SAAS;AACb,UAAI,SAAS;AACb,YAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM;AAC9B,kBAAU,OAAO,CAAC;AAAA,MACpB,CAAC;AACD,YAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM;AAC9B,kBAAU,OAAO,CAAC;AAAA,MACpB,CAAC;AACD,YAAM,UAAU;AAAA,QACd,MAAM;AACJ,cAAI;AACF,kBAAM,KAAK,SAAS;AAAA,UACtB,SAAS,KAAK;AACZ,oBAAQ,KAAK,gDAAgD,GAAG;AAAA,UAClE;AAAA,QACF;AAAA,QACA,OAAO,aAAa,KAAK;AAAA,MAC3B;AACA,YAAM,GAAG,SAAS,CAAC,SAAS;AAC1B,qBAAa,OAAO;AACpB,cAAM,SAAS,KAAK,IAAI,IAAI;AAC5B,cAAM,SACJ,UAAU,UAAU,OAAO,aACvB,OAAO,WAAW,MAAM,QAAQ,QAAQ,QAAQ,CAAC,IACjD;AACN,gBAAQ;AAAA,UACN;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB;AAAA,UACA;AAAA,UACA;AAAA,UACA,YAAY,QAAQ;AAAA,UACpB,aAAa,QAAQ;AAAA,QACvB,CAAC;AAAA,MACH,CAAC;AACD,YAAM,GAAG,SAAS,CAAC,QAAQ;AACzB,qBAAa,OAAO;AACpB,cAAM,SAAS,KAAK,IAAI,IAAI;AAC5B,gBAAQ,EAAE,OAAO,UAAU,KAAK,QAAQ,QAAQ,SAAS,OAAO,GAAG,GAAG,OAAO,CAAC;AAAA,MAChF,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AACF;AAEO,IAAM,sBAAN,MAAmD;AAAA,EACxD,KAAK;AAAA,EAEL,MAAM,KACJ,OACA,SACA,QACwB;AACxB,QAAI,CAAC,OAAO,MAAO,OAAM,IAAI,YAAY,2CAA2C;AACpF,UAAM,MAAM,IAAI,wBAAwB;AACxC,UAAM,UAAU,OAAO,QAAQ,OAAO,OAAO,CAAC,CAAC,EAC5C,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,MAAM,WAAW,CAAC,CAAC,IAAI,WAAW,CAAC,CAAC,EAAE,EACtD,KAAK,GAAG;AACX,UAAM,UAAU,mBAAmB,OAAO,IAAI,WAAW,OAAO,KAAK,CAAC,UAAU,WAAW,OAAO,CAAC;AACnG,WAAO,IAAI,KAAK,OAAO,SAAS,EAAE,GAAG,QAAQ,KAAK,OAAU,CAAC;AAAA,EAC/D;AACF;AAEA,SAAS,WAAW,GAAmB;AACrC,MAAI,yBAAyB,KAAK,CAAC,EAAG,QAAO;AAC7C,SAAO,IAAI,EAAE,QAAQ,MAAM,OAAO,CAAC;AACrC;AAcO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACR,YAAY,SAAwB,IAAI,wBAAwB,GAAG;AACjE,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,MAAM,IAAI,QAAuB,SAAsD;AACrF,UAAM,SAAS,MAAM,QAAQ,QAAQ;AAAA,MACnC,MAAM,WAAW,KAAK,OAAO,EAAE;AAAA,MAC/B,OAAO,OAAO;AAAA,MACd,SAAS,CAAC,OAAO,cAAc,OAAO,YAAY,OAAO,WAAW,EACjE,OAAO,OAAO,EACd,KAAK,MAAM;AAAA,IAChB,CAAC;AACD,UAAM,SAA+B,EAAE,QAAQ,OAAO,aAAa,GAAG,OAAO,EAAE;AAC/E,QAAI;AACF,UAAI,OAAO,cAAc;AACvB,eAAO,QAAQ,MAAM,KAAK,OAAO,KAAK,SAAS,OAAO,cAAc,MAAM;AAC1E,eAAO,eAAe,OAAO,MAAM;AACnC,YAAI,OAAO,MAAM,aAAa,GAAG;AAC/B,gBAAM,OAAO,KAAK,sBAAsB,OAAO,MAAM,QAAQ,KAAK;AAAA,YAChE,UAAU,OAAO,MAAM;AAAA,YACvB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,YAAY;AACrB,eAAO,MAAM,MAAM,KAAK,OAAO,KAAK,OAAO,OAAO,YAAY,MAAM;AACpE,eAAO,eAAe,OAAO,IAAI;AACjC,YAAI,OAAO,IAAI,aAAa,GAAG;AAC7B,gBAAM,OAAO,KAAK,oBAAoB,OAAO,IAAI,QAAQ,KAAK;AAAA,YAC5D,UAAU,OAAO,IAAI;AAAA,YACrB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,aAAa;AACtB,eAAO,OAAO,MAAM,KAAK,OAAO,KAAK,QAAQ,OAAO,aAAa,MAAM;AACvE,eAAO,eAAe,OAAO,KAAK;AAClC,cAAM,SAAS,OAAO,KAAK,aAAa;AACxC,eAAO,SAAS;AAChB,YAAI,OAAO,KAAK,eAAe,UAAa,OAAO,KAAK,aAAa,GAAG;AACtE,iBAAO,SAAS,OAAO,KAAK,eAAe,KAAK,OAAO,KAAK;AAAA,QAC9D,OAAO;AACL,iBAAO,QAAQ,SAAS,IAAI;AAAA,QAC9B;AACA,cAAM,OAAO,IAAI;AAAA,UACf,UAAU,OAAO,KAAK;AAAA,UACtB,YAAY,OAAO,KAAK;AAAA,UACxB,aAAa,OAAO,KAAK;AAAA,UACzB,QAAQ,OAAO;AAAA,UACf,QAAQ,SAAS,OAAO;AAAA,QAC1B,CAAyB;AAAA,MAC3B,OAAO;AACL,eAAO,SAAS;AAChB,eAAO,QAAQ;AACf,cAAM,OAAO,IAAI,EAAE,QAAQ,OAAO,YAAY,CAAyB;AAAA,MACzE;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;AAC1D,YAAM;AAAA,IACR;AACA,WAAO;AAAA,EACT;AACF;;;AC5QA,eAAsB,sBACpB,UACA,OACA,UAAgC,CAAC,GACH;AAC9B,QAAM,UAAU,IAAI,aAAa,KAAK;AACtC,QAAM,QAAQ,SAAS;AAAA,IACrB,YAAY,SAAS;AAAA,IACrB,WAAW,QAAQ;AAAA,IACnB,gBAAgB,SAAS;AAAA,IACzB,MAAM,SAAS;AAAA,IACf,GAAG,QAAQ;AAAA,EACb,CAAC;AACD,QAAM,UAAU,IAAI,eAAe,QAAQ,MAAM;AACjD,QAAM,SAAS,MAAM,QAAQ,IAAI,SAAS,SAAS,OAAO;AAC1D,QAAM,YAAY,SAAS,iBAAiB;AAC5C,QAAM,OAAO,OAAO,UAAU,OAAO,SAAS;AAC9C,QAAM,cAAc,OAAO,UAAU,UAAa,OAAO,MAAM,aAAa;AAC5E,QAAM,YAAY,OAAO,QAAQ,UAAa,OAAO,IAAI,aAAa;AACtE,QAAM,aAAa,OAAO,SAAS,UAAa,OAAO,KAAK,aAAa;AACzE,QAAM,eAAyC,OAC3C,YACA,eAAe,YACb,oBACA,aACE,iBACA;AACR,QAAM,QAAQ,OAAO;AAAA,IACnB;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,IACA,OAAO,OAAO,SAAY,iBAAiB,MAAM;AAAA,EACnD,CAAC;AACD,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf;AAAA,IACA,SAAS;AAAA,IACT;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,QAAsC;AAC9D,MAAI,OAAO,SAAS,OAAO,MAAM,aAAa;AAC5C,WAAO,sBAAsB,OAAO,MAAM,QAAQ;AACpD,MAAI,OAAO,OAAO,OAAO,IAAI,aAAa,EAAG,QAAO,oBAAoB,OAAO,IAAI,QAAQ;AAC3F,MAAI,OAAO,MAAM;AACf,QAAI,OAAO,KAAK,eAAe,QAAW;AACxC,aAAO,UAAU,OAAO,KAAK,eAAe,CAAC,IAAI,OAAO,KAAK,UAAU;AAAA,IACzE;AACA,WAAO,aAAa,OAAO,KAAK,QAAQ;AAAA,EAC1C;AACA,SAAO;AACT;","names":[]}

package/dist/chunk-TWS7AZEY.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/campaign/run-campaign.ts","../src/integrity/backend-integrity.ts","../src/campaign/storage.ts"],"sourcesContent":["/**\n * @experimental\n *\n * `runCampaign` — Pass A substrate primitive. ONE function that orchestrates\n * scenarios → dispatch → artifacts → judges → aggregates, with full\n * reproducibility (seed + manifest hash), cell-level resumability, bootstrap\n * CIs, and the `LabeledScenarioStore` capture flywheel.\n *\n * Improvement loops (optimizer / gate / autoOnPromote) ride on top of this\n * primitive but live in `presets/run-improvement-loop.ts`. This file keeps\n * the core orchestrator minimal — Phase 1 of the Pass A track.\n */\n\nimport { createHash } from 'node:crypto'\nimport { join } from 'node:path'\nimport { BackendIntegrityError, type BackendIntegrityReport } from '../integrity/backend-integrity'\nimport { confidenceInterval } from '../statistics'\nimport { type CampaignStorage, fsCampaignStorage } from './storage'\nimport type {\n  CampaignAggregates,\n  CampaignArtifactWriter,\n  CampaignCellResult,\n  CampaignCostMeter,\n  CampaignResult,\n  CampaignTokenUsage,\n  CampaignTraceWriter,\n  DispatchContext,\n  DispatchFn,\n  JudgeAggregate,\n  JudgeConfig,\n  JudgeScore,\n  LabeledScenarioStore,\n  Scenario,\n  ScenarioAggregate,\n  TraceSpan,\n} from './types'\n\nexport interface RunCampaignOptions<TScenario extends Scenario, TArtifact> {\n  scenarios: TScenario[]\n  dispatch: DispatchFn<TScenario, TArtifact>\n  judges?: JudgeConfig<TArtifact, TScenario>[]\n  /** Required for reproducibility. Default 42. */\n  seed?: number\n  /** Per-scenario replicates for CI bands. Default 1; raise to 5+ for\n   *  bootstrap-tight intervals on critical eval. */\n  reps?: number\n  /** When true (default), completed cells are cached by\n   *  (manifestHash, scenarioId, rep, generation). Re-runs skip cached cells. */\n  resumable?: boolean\n  /** Optional store — when present, every artifact + judge score is captured\n   *  with the configured `captureSource`. Capture is default ON; pass `'off'`\n   *  to disable. */\n  labeledStore?: LabeledScenarioStore | 'off'\n  captureSource?: 'production-trace' | 'eval-run' | 'manual' | 'red-team' | 'synthetic'\n  captureSourceVersionHash?: string\n  /** Wall-clock cost cap across all cells. Cells beyond ceiling are skipped. */\n  costCeiling?: number\n  /** Max concurrent cells. Default 2. */\n  maxConcurrency?: number\n  /**\n   * Per-cell dispatch deadline in ms. A `dispatch` that neither resolves nor\n   * rejects within this window is a hang (a stalled model request, an\n   * exhausted runtime resource, a backend that never closes its stream). When\n   * set, the cell's `ctx.signal` is aborted and the cell is recorded as a LOUD\n   * error (`dispatch exceeded <N>ms`) so the campaign proceeds and the failure\n   * is visible — instead of one wedged cell silently hanging the whole run (and\n   * every loop/CI job above it) forever. `undefined`/`0` = unbounded (legacy).\n   */\n  dispatchTimeoutMs?: number\n  /** Required: where artifacts + traces land. */\n  runDir: string\n  /** Tracing posture. Default is the substrate's `FileSystemTraceStore` rooted\n   *  at `<runDir>/traces/`. `'off'` disables capture entirely — substrate\n   *  refuses this when the caller wires `autoOnPromote !== 'none'`. */\n  tracing?: 'on' | 'off'\n  /**\n   * Per-cell usage expectation — the early, fine-grained sibling of the\n   * batch `assertRealBackend` guard. A cell that produced an artifact (no\n   * error) but reported `costUsd === 0` AND zero tokens is a stub: the\n   * dispatch never reported LLM activity via `ctx.cost`. Modes:\n   *   - `'warn'` (default) — log the offending cell loudly, keep going.\n   *   - `'assert'` — throw `BackendIntegrityError` on the first such cell\n   *     (fail-fast; recommended for CI campaigns expecting real LLM calls).\n   *   - `'off'` — no check (replay / deterministic-only / offline analysis).\n   */\n  expectUsage?: 'assert' | 'warn' | 'off'\n  /** Test seam — override the wall clock for deterministic tests. */\n  now?: () => Date\n  /** Test seam — override per-cell trace writer factory. */\n  buildTraceWriter?: (cellId: string, dir: string) => CampaignTraceWriter\n  /** Storage backend for run/cell dirs, the resumability cache, artifacts,\n   *  and trace spans. Default: the Node filesystem (`fsCampaignStorage`).\n   *  Pass `inMemoryCampaignStorage()` to run in a filesystem-less runtime\n   *  (Cloudflare Workers, Deno, edge) — the `CampaignResult` is still\n   *  produced; artifacts/traces just aren't persisted to disk. */\n  storage?: CampaignStorage\n  /**\n   * Optional per-cell placement strategy. Returns an opaque string the\n   * substrate forwards as `ctx.placement` to the Dispatch — placement-aware\n   * Dispatches (e.g. `httpDispatch` from `/adapters/http`) use it to route\n   * each cell to the right worker, region, or sandbox. When unset, every\n   * cell receives `ctx.placement = undefined` and behaves identically to\n   * the in-process case.\n   *\n   * @example\n   *   cellPlacement: ({ scenario }) => scenario.tags?.includes('eu') ? 'eu-west' : 'us-east'\n   */\n  cellPlacement?: (input: {\n    scenario: TScenario\n    rep: number\n    generation?: number\n  }) => string | undefined\n}\n\nexport async function runCampaign<TScenario extends Scenario, TArtifact>(\n  opts: RunCampaignOptions<TScenario, TArtifact>,\n): Promise<CampaignResult<TArtifact, TScenario>> {\n  const seed = opts.seed ?? 42\n  const reps = opts.reps ?? 1\n  const resumable = opts.resumable ?? true\n  const maxConcurrency = opts.maxConcurrency ?? 2\n  const now = opts.now ?? (() => new Date())\n  const judges = opts.judges ?? []\n  const storage = opts.storage ?? fsCampaignStorage()\n\n  if (typeof opts.runDir !== 'string' || opts.runDir.trim().length === 0) {\n    throw new Error('runCampaign: runDir is required and must be a non-empty string')\n  }\n  storage.ensureDir(opts.runDir)\n\n  const manifestHash = computeManifestHash({\n    scenarios: opts.scenarios,\n    judges: judges as unknown as JudgeConfig<unknown>[],\n    dispatchRef: opts.dispatch.name || 'anonymous',\n    seed,\n    reps,\n  })\n\n  const startedAt = now()\n  const cells: CampaignCellResult<TArtifact>[] = []\n  const artifactsByPath: Record<string, string> = {}\n\n  // Build the cell schedule (scenario × rep).\n  const schedule: Array<{ scenario: TScenario; rep: number; cellId: string; cellSeed: number }> = []\n  let cellIndex = 0\n  for (const scenario of opts.scenarios) {\n    for (let rep = 0; rep < reps; rep++) {\n      const cellId = `${scenario.id}:${rep}`\n      const cellSeed = seed + cellIndex\n      schedule.push({ scenario, rep, cellId, cellSeed })\n      cellIndex += 1\n    }\n  }\n\n  // Concurrency-limited execution.\n  let totalCostUsd = 0\n  let costCeilingReached = false\n  const abortController = new AbortController()\n  // Concurrency lanes that drain the cell schedule. Named \"lanes\" — not\n  // \"workers\" — to avoid clashing with the taxonomy's worker (= the agent\n  // harness in a sandbox, invoked behind `dispatch`). See loop-taxonomy.md.\n  const lanes: Promise<void>[] = []\n  let nextIdx = 0\n  const cellsRef = cells\n\n  for (let i = 0; i < maxConcurrency; i++) {\n    lanes.push(\n      (async () => {\n        while (true) {\n          const myIdx = nextIdx++\n          if (myIdx >= schedule.length) return\n          const slot = schedule[myIdx]!\n          if (costCeilingReached) {\n            cellsRef.push(skippedCell(slot, 'cost_ceiling_reached'))\n            continue\n          }\n          const result = await executeCell({\n            slot,\n            opts,\n            manifestHash,\n            resumable,\n            now,\n            storage,\n            buildTraceWriter: opts.buildTraceWriter ?? defaultBuildTraceWriter(storage),\n            signal: abortController.signal,\n            dispatchTimeoutMs: opts.dispatchTimeoutMs,\n          })\n          cellsRef.push(result.cell)\n          enforceCellUsage(result.cell, opts.expectUsage ?? 'warn')\n          totalCostUsd += result.cell.costUsd\n          Object.assign(artifactsByPath, result.artifactsByPath)\n          if (opts.costCeiling !== undefined && totalCostUsd >= opts.costCeiling) {\n            costCeilingReached = true\n          }\n          // Capture into LabeledScenarioStore unless explicitly disabled.\n          if (opts.labeledStore && opts.labeledStore !== 'off' && !result.cell.error) {\n            await captureToStore({\n              store: opts.labeledStore,\n              cell: result.cell,\n              scenario: slot.scenario,\n              opts,\n              now,\n            }).catch((err) => {\n              // Capture failures are non-fatal — log but don't crash the campaign.\n              // (Trace would normally land here.)\n              console.warn(\n                `[runCampaign] capture failed for ${result.cell.cellId}: ${err instanceof Error ? err.message : String(err)}`,\n              )\n            })\n          }\n        }\n      })(),\n    )\n  }\n  await Promise.all(lanes)\n\n  const endedAt = now()\n  cellsRef.sort((a, b) => a.cellId.localeCompare(b.cellId))\n\n  const aggregates = computeAggregates(\n    cellsRef,\n    judges as unknown as JudgeConfig<TArtifact>[],\n    seed,\n  )\n\n  return {\n    manifestHash,\n    seed,\n    startedAt: startedAt.toISOString(),\n    endedAt: endedAt.toISOString(),\n    durationMs: endedAt.getTime() - startedAt.getTime(),\n    cells: cellsRef,\n    aggregates,\n    runDir: opts.runDir,\n    artifactsByPath,\n    scenarios: opts.scenarios.map((s) => ({ id: s.id, kind: s.kind })),\n  }\n}\n\n// ── Internals ─────────────────────────────────────────────────────────\n\ninterface ExecuteCellArgs<TScenario extends Scenario, TArtifact> {\n  slot: { scenario: TScenario; rep: number; cellId: string; cellSeed: number }\n  opts: RunCampaignOptions<TScenario, TArtifact>\n  manifestHash: string\n  resumable: boolean\n  now: () => Date\n  storage: CampaignStorage\n  buildTraceWriter: (cellId: string, dir: string) => CampaignTraceWriter\n  signal: AbortSignal\n  dispatchTimeoutMs?: number\n}\n\nasync function executeCell<TScenario extends Scenario, TArtifact>(\n  args: ExecuteCellArgs<TScenario, TArtifact>,\n): Promise<{ cell: CampaignCellResult<TArtifact>; artifactsByPath: Record<string, string> }> {\n  const storage = args.storage\n  const cellDir = join(args.opts.runDir, args.slot.cellId.replace(/[^a-zA-Z0-9_-]/g, '_'))\n  storage.ensureDir(cellDir)\n\n  // Resumability: cache key = (manifestHash, scenarioId, rep)\n  const cachePath = join(cellDir, 'cached-result.json')\n  if (args.resumable) {\n    const raw = storage.read(cachePath)\n    if (raw !== undefined) {\n      try {\n        const cached = JSON.parse(raw) as CampaignCellResult<TArtifact>\n        if (cached.cellId === args.slot.cellId) {\n          return { cell: { ...cached, cached: true }, artifactsByPath: {} }\n        }\n      } catch {\n        // Corrupt cache — fall through to re-run.\n      }\n    }\n  }\n\n  const startMs = Date.now()\n  const trace = args.buildTraceWriter(args.slot.cellId, cellDir)\n  const artifactsByPath: Record<string, string> = {}\n  const artifacts: CampaignArtifactWriter = {\n    async write(path, content) {\n      const fullPath = join(cellDir, path)\n      storage.ensureDir(join(fullPath, '..'))\n      storage.write(fullPath, content)\n      artifactsByPath[`${args.slot.cellId}/${path}`] = fullPath\n      return fullPath\n    },\n    async writeJson(path, value) {\n      return artifacts.write(path, JSON.stringify(value, null, 2))\n    },\n  }\n  let costSoFar = 0\n  const tokensSoFar: CampaignTokenUsage = { input: 0, output: 0 }\n  const cost: CampaignCostMeter = {\n    observe(amount, source) {\n      costSoFar += amount\n      trace.span(`cost.${source}`, { amountUsd: amount }).end()\n    },\n    observeTokens(usage) {\n      tokensSoFar.input += usage.input\n      tokensSoFar.output += usage.output\n      if (usage.cached) tokensSoFar.cached = (tokensSoFar.cached ?? 0) + usage.cached\n    },\n    current() {\n      return costSoFar\n    },\n    tokens() {\n      return { ...tokensSoFar }\n    },\n  }\n\n  const placement = args.opts.cellPlacement?.({\n    scenario: args.slot.scenario,\n    rep: args.slot.rep,\n  })\n\n  // Per-cell abort signal, chained to the campaign signal. The dispatch sees\n  // THIS signal so a timeout (below) can abort just this cell's in-flight work\n  // without tearing down sibling cells — and a signal-honoring dispatch\n  // releases its open request instead of leaking it past the deadline.\n  const cellAbort = new AbortController()\n  const onCampaignAbort = () => cellAbort.abort((args.signal as { reason?: unknown }).reason)\n  if (args.signal.aborted) cellAbort.abort((args.signal as { reason?: unknown }).reason)\n  else args.signal.addEventListener('abort', onCampaignAbort, { once: true })\n\n  const ctx: DispatchContext = {\n    cellId: args.slot.cellId,\n    rep: args.slot.rep,\n    seed: args.slot.cellSeed,\n    signal: cellAbort.signal,\n    trace,\n    artifacts,\n    cost,\n    placement,\n  }\n\n  let artifact: TArtifact | undefined\n  let errorMessage: string | undefined\n  const timeoutMs = args.dispatchTimeoutMs\n  let timeoutTimer: ReturnType<typeof setTimeout> | undefined\n  try {\n    const dispatched = args.opts.dispatch(args.slot.scenario, ctx)\n    if (timeoutMs !== undefined && timeoutMs > 0) {\n      // A dispatch that never settles (stalled model request, exhausted runtime\n      // resource, a stream that never closes) must NOT hang the cell — and with\n      // it the lane, the campaign, the loop, the CI job — forever. Race it\n      // against the deadline; on timeout, abort the cell and fail it LOUD.\n      artifact = await Promise.race([\n        dispatched,\n        new Promise<never>((_, reject) => {\n          timeoutTimer = setTimeout(() => {\n            cellAbort.abort(new Error('dispatch timeout'))\n            reject(\n              new Error(\n                `dispatch exceeded ${timeoutMs}ms for cell '${args.slot.cellId}' — aborted and failed loud (no silent hang)`,\n              ),\n            )\n          }, timeoutMs)\n          if (typeof (timeoutTimer as { unref?: () => void }).unref === 'function')\n            (timeoutTimer as { unref: () => void }).unref()\n        }),\n      ])\n    } else {\n      artifact = await dispatched\n    }\n  } catch (err) {\n    errorMessage = err instanceof Error ? err.message : String(err)\n  } finally {\n    if (timeoutTimer) clearTimeout(timeoutTimer)\n    args.signal.removeEventListener('abort', onCampaignAbort)\n  }\n\n  // Run judges (only if we have an artifact). A judge that throws invalidates\n  // the cell — recorded as `error`, NOT folded into a fake composite:0 (a fake\n  // zero is indistinguishable from a real zero and poisons every aggregate).\n  const judgeScores: Record<string, JudgeScore> = {}\n  if (artifact !== undefined) {\n    for (const judge of args.opts.judges ?? []) {\n      if (judge.appliesTo && !judge.appliesTo(args.slot.scenario)) continue\n      try {\n        judgeScores[judge.name] = await runJudgeCell(judge, {\n          artifact,\n          scenario: args.slot.scenario,\n          signal: args.signal,\n        })\n      } catch (err) {\n        errorMessage = `judge '${judge.name}' failed: ${err instanceof Error ? err.message : String(err)}`\n        break\n      }\n    }\n  }\n\n  await trace.flush()\n\n  const cell: CampaignCellResult<TArtifact> = {\n    cellId: args.slot.cellId,\n    scenarioId: args.slot.scenario.id,\n    rep: args.slot.rep,\n    artifact: (artifact ?? null) as TArtifact,\n    judgeScores,\n    costUsd: costSoFar,\n    tokenUsage: { ...tokensSoFar },\n    durationMs: Date.now() - startMs,\n    seed: args.slot.cellSeed,\n    cached: false,\n    error: errorMessage,\n  }\n\n  if (!errorMessage && args.resumable) {\n    storage.write(cachePath, JSON.stringify(cell))\n  }\n\n  return { cell, artifactsByPath }\n}\n\n/**\n * Per-cell stub guard. A cell that produced an artifact (no error) but reported\n * `costUsd === 0` AND zero tokens means the dispatch never called `ctx.cost` —\n * i.e. it ran against a stub or silently dropped its usage. `'warn'` logs it,\n * `'assert'` throws (fail-fast), `'off'` skips. An errored/skipped cell or a\n * deterministic judge-only run that genuinely made no LLM call is not flagged.\n */\nfunction enforceCellUsage<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  mode: 'assert' | 'warn' | 'off',\n): void {\n  if (mode === 'off' || cell.error) return\n  if (cell.artifact === null || cell.artifact === undefined) return\n  const zeroTokens = cell.tokenUsage.input === 0 && cell.tokenUsage.output === 0\n  if (cell.costUsd !== 0 || !zeroTokens) return\n  const msg = `cell '${cell.cellId}' produced an artifact but reported zero cost and zero tokens — the dispatch never reported LLM usage via ctx.cost.observe/observeTokens (a stub cell)`\n  if (mode === 'assert') {\n    const report: BackendIntegrityReport = {\n      totalRecords: 1,\n      stubRecords: 1,\n      realRecords: 0,\n      uncostedRecords: 0,\n      totalInputTokens: 0,\n      totalOutputTokens: 0,\n      totalCostUsd: 0,\n      verdict: 'stub',\n      diagnosis: msg,\n    }\n    throw new BackendIntegrityError(`expectUsage: ${msg}`, report)\n  }\n  // eslint-disable-next-line no-console\n  console.warn(`[runCampaign] expectUsage: ${msg}`)\n}\n\nasync function runJudgeCell<TArtifact, TScenario extends Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  input: { artifact: TArtifact; scenario: TScenario; signal: AbortSignal },\n): Promise<JudgeScore> {\n  return judge.score(input)\n}\n\nfunction defaultBuildTraceWriter(\n  storage: CampaignStorage,\n): (cellId: string, dir: string) => CampaignTraceWriter {\n  return (cellId, dir) => {\n    const spans: Array<Record<string, unknown>> = []\n    return {\n      span(name, attributes) {\n        const startMs = Date.now()\n        const record: Record<string, unknown> = { name, cellId, startMs, ...(attributes ?? {}) }\n        const finish: TraceSpan = {\n          end(endAttrs) {\n            record.durationMs = Date.now() - startMs\n            if (endAttrs) Object.assign(record, endAttrs)\n            spans.push(record)\n          },\n          setAttribute(key, value) {\n            record[key] = value\n          },\n        }\n        return finish\n      },\n      async flush() {\n        storage.write(join(dir, 'spans.jsonl'), spans.map((s) => JSON.stringify(s)).join('\\n'))\n      },\n    }\n  }\n}\n\nfunction skippedCell<TScenario extends Scenario, TArtifact>(\n  slot: { scenario: TScenario; rep: number; cellId: string; cellSeed: number },\n  reason: string,\n): CampaignCellResult<TArtifact> {\n  return {\n    cellId: slot.cellId,\n    scenarioId: slot.scenario.id,\n    rep: slot.rep,\n    artifact: null as unknown as TArtifact,\n    judgeScores: {},\n    costUsd: 0,\n    tokenUsage: { input: 0, output: 0 },\n    durationMs: 0,\n    seed: slot.cellSeed,\n    cached: false,\n    error: `skipped: ${reason}`,\n  }\n}\n\ninterface CaptureArgs<TScenario extends Scenario, TArtifact> {\n  store: LabeledScenarioStore\n  cell: CampaignCellResult<TArtifact>\n  scenario: TScenario\n  opts: RunCampaignOptions<TScenario, TArtifact>\n  now: () => Date\n}\n\nasync function captureToStore<TScenario extends Scenario, TArtifact>(\n  args: CaptureArgs<TScenario, TArtifact>,\n): Promise<void> {\n  await args.store.observe({\n    scenario: args.scenario,\n    artifact: args.cell.artifact,\n    judgeScores: args.cell.judgeScores,\n    source: args.opts.captureSource ?? 'eval-run',\n    sourceVersionHash: args.opts.captureSourceVersionHash ?? 'unknown',\n    capturedAt: args.now().toISOString(),\n    redactionStatus: 'raw',\n  })\n}\n\n// ── Aggregates + manifest hash ────────────────────────────────────────\n\nfunction computeManifestHash(input: {\n  scenarios: Scenario[]\n  judges: JudgeConfig<unknown>[]\n  dispatchRef: string\n  seed: number\n  reps: number\n}): string {\n  const canonical = {\n    scenarios: input.scenarios.map((s) => ({ id: s.id, kind: s.kind })),\n    judges: input.judges.map((j) => ({ name: j.name, dims: j.dimensions.map((d) => d.key) })),\n    dispatch: input.dispatchRef,\n    seed: input.seed,\n    reps: input.reps,\n  }\n  return createHash('sha256').update(JSON.stringify(canonical)).digest('hex')\n}\n\nfunction computeAggregates<TArtifact>(\n  cells: CampaignCellResult<TArtifact>[],\n  judges: JudgeConfig<TArtifact>[],\n  seed: number,\n): CampaignAggregates {\n  const byJudge: Record<string, JudgeAggregate> = {}\n  for (const judge of judges) {\n    const scores: number[] = []\n    for (const cell of cells) {\n      const s = cell.judgeScores[judge.name]\n      if (s !== undefined) scores.push(s.composite)\n    }\n    byJudge[judge.name] = aggregate(scores, seed)\n  }\n  const byScenario: Record<string, ScenarioAggregate> = {}\n  const scenarioGroups = new Map<string, number[]>()\n  for (const cell of cells) {\n    const composites = Object.values(cell.judgeScores).map((s) => s.composite)\n    if (composites.length === 0) continue\n    const mean = composites.reduce((a, b) => a + b, 0) / composites.length\n    const arr = scenarioGroups.get(cell.scenarioId) ?? []\n    arr.push(mean)\n    scenarioGroups.set(cell.scenarioId, arr)\n  }\n  for (const [scenarioId, samples] of scenarioGroups) {\n    const ag = aggregate(samples, seed)\n    byScenario[scenarioId] = { meanComposite: ag.mean, ci95: ag.ci95, n: ag.n }\n  }\n  return {\n    byJudge,\n    byScenario,\n    totalCostUsd: cells.reduce((a, c) => a + c.costUsd, 0),\n    cellsExecuted: cells.filter((c) => !c.error).length,\n    cellsSkipped: cells.filter((c) => c.error?.startsWith('skipped:')).length,\n    cellsCached: cells.filter((c) => c.cached).length,\n    cellsFailed: cells.filter((c) => c.error && !c.error.startsWith('skipped:')).length,\n  }\n}\n\n// Percentile bootstrap CI95 via seeded resampling. Deterministic for a given\n// seed — same campaign re-run produces identical CI bands. Falls back to\n// degenerate intervals at n<=1 (the bootstrap is undefined there).\nfunction aggregate(samples: number[], seed: number): JudgeAggregate {\n  const n = samples.length\n  if (n === 0) return { mean: 0, stdev: 0, ci95: [0, 0], n: 0 }\n  const mean = samples.reduce((a, b) => a + b, 0) / n\n  const variance = samples.reduce((a, b) => a + (b - mean) ** 2, 0) / Math.max(1, n - 1)\n  const stdev = Math.sqrt(variance)\n  const ci = confidenceInterval(samples, 0.95, { seed, resamples: 1000 })\n  return { mean, stdev, ci95: [ci.lower, ci.upper], n }\n}\n","/**\n * Backend-integrity guard: distinguish \"agent failed\" from \"eval ran against\n * a stub / unconfigured backend.\" Without this guard a canonical eval can\n * silently report `0/N passed` and look like an agent-quality problem when\n * the LLM was never actually called — the failure mode we just hit running\n * the 4-vertical parallel eval (legal-sandbox-stub returned hard-coded 33-104\n * char strings; gtm/creative defaulted to a cli-bridge that wasn't running).\n *\n * The shape:\n *\n *   const report = summarizeBackendIntegrity(records)\n *   assertRealBackend(records)   // throws BackendIntegrityError if 100% stub\n *\n * A record is \"stub-mode\" if its `tokenUsage.input === 0 && tokenUsage.output === 0`.\n * (`costUsd` alone is unreliable — some backends successfully call LLMs but\n *  don't propagate pricing, producing real tokens with $0 cost.)\n *\n * Verdicts:\n *   - `real`   — at least one record has nonzero token usage\n *   - `stub`   — every record is stub-mode (eval ran blind)\n *   - `mixed`  — some records real, some stub (partial backend failure;\n *                often the 429-cascade or auth-half-failed case)\n */\n\nimport { AgentEvalError } from '../errors'\nimport type { RunRecord } from '../run-record'\n\nexport interface BackendIntegrityReport {\n  /** Total records inspected. */\n  totalRecords: number\n  /** Records with input=0 AND output=0 (a stub fingerprint). */\n  stubRecords: number\n  /** Records with nonzero token usage (real LLM activity). */\n  realRecords: number\n  /** Records where output>0 but costUsd=0 (real LLM, broken cost ledger). */\n  uncostedRecords: number\n  /** Sum of input tokens across all records. */\n  totalInputTokens: number\n  /** Sum of output tokens across all records. */\n  totalOutputTokens: number\n  /** Sum of costUsd across all records. */\n  totalCostUsd: number\n  /** Worst-case integrity verdict. */\n  verdict: 'real' | 'mixed' | 'stub'\n  /** Human-readable diagnosis suitable for terminal output. */\n  diagnosis: string\n}\n\n/**\n * Error thrown when an integrity assertion fails. Caller can pattern-match\n * by `code === 'AGENT_EVAL_BACKEND_STUB'` to differentiate from other\n * errors.\n */\nexport class BackendIntegrityError extends AgentEvalError {\n  constructor(\n    message: string,\n    public readonly report: BackendIntegrityReport,\n  ) {\n    super('backend_integrity', message)\n  }\n}\n\nfunction isStubRecord(rec: RunRecord): boolean {\n  return rec.tokenUsage.input === 0 && rec.tokenUsage.output === 0\n}\n\nfunction isUncostedRecord(rec: RunRecord): boolean {\n  return rec.tokenUsage.output > 0 && rec.costUsd === 0\n}\n\n/**\n * Inspect a batch of RunRecords and return an integrity report. Pure\n * function — no I/O, no logging. The caller decides what to do with the\n * verdict (print warning, throw, gate CI, etc.).\n */\nexport function summarizeBackendIntegrity(\n  records: ReadonlyArray<RunRecord>,\n): BackendIntegrityReport {\n  const totalRecords = records.length\n  let stubRecords = 0\n  let realRecords = 0\n  let uncostedRecords = 0\n  let totalInputTokens = 0\n  let totalOutputTokens = 0\n  let totalCostUsd = 0\n  for (const rec of records) {\n    totalInputTokens += rec.tokenUsage.input\n    totalOutputTokens += rec.tokenUsage.output\n    totalCostUsd += rec.costUsd\n    if (isStubRecord(rec)) stubRecords++\n    else realRecords++\n    if (isUncostedRecord(rec)) uncostedRecords++\n  }\n  const verdict: BackendIntegrityReport['verdict'] =\n    totalRecords === 0\n      ? 'stub'\n      : stubRecords === totalRecords\n        ? 'stub'\n        : stubRecords === 0\n          ? 'real'\n          : 'mixed'\n  const diagnosis = buildDiagnosis({\n    totalRecords,\n    stubRecords,\n    realRecords,\n    uncostedRecords,\n    totalInputTokens,\n    totalOutputTokens,\n    totalCostUsd,\n    verdict,\n  })\n  return {\n    totalRecords,\n    stubRecords,\n    realRecords,\n    uncostedRecords,\n    totalInputTokens,\n    totalOutputTokens,\n    totalCostUsd,\n    verdict,\n    diagnosis,\n  }\n}\n\nfunction buildDiagnosis(r: Omit<BackendIntegrityReport, 'diagnosis'>): string {\n  if (r.totalRecords === 0) {\n    return 'no records — eval produced zero runs; backend likely failed before first turn'\n  }\n  if (r.verdict === 'stub') {\n    return [\n      `all ${r.totalRecords} records have zero token usage — the LLM backend was never called.`,\n      'common causes: --backend sandbox without a sandbox bridge running; stub model returning hard-coded strings;',\n      'auth misconfigured so requests were silently dropped before the LLM. Re-run with --backend tcloud and TANGLE_API_KEY set,',\n      'or boot the cli-bridge / sandbox before invoking the eval.',\n    ].join(' ')\n  }\n  if (r.verdict === 'mixed') {\n    const pct = ((r.stubRecords / r.totalRecords) * 100).toFixed(0)\n    return [\n      `${r.stubRecords}/${r.totalRecords} records (${pct}%) have zero token usage — the backend partially failed.`,\n      'common causes: rate-limit cascade (429s after the first N personas);',\n      'transient auth expiry mid-run; provider outage. Treat the affected records as missing data, not agent failures.',\n    ].join(' ')\n  }\n  // verdict === 'real'\n  if (r.uncostedRecords > 0) {\n    const pct = ((r.uncostedRecords / r.totalRecords) * 100).toFixed(0)\n    return [\n      `${r.totalRecords} records with real LLM activity (in=${r.totalInputTokens}, out=${r.totalOutputTokens} tokens).`,\n      `${r.uncostedRecords} (${pct}%) have output tokens but costUsd=0. Two distinct roots:`,\n      '(a) cost ledger mis-wired — no usage propagation from the runtime stream into RunRecord; or',\n      '(b) the model is unpriced at the source (sandbox/router returned $0 despite real tokens).',\n      'For (b), price the measured tokens against the substrate table (estimateCost) instead of leaving $0.',\n    ].join(' ')\n  }\n  return `${r.totalRecords} records with real LLM activity (in=${r.totalInputTokens}, out=${r.totalOutputTokens} tokens, $${r.totalCostUsd.toFixed(4)}).`\n}\n\n/**\n * Throw BackendIntegrityError if the verdict is 'stub' — i.e. every record\n * shows zero LLM activity. Non-strict callers can pass `{ allowMixed: false }`\n * to also reject mixed verdicts (recommended for CI gates).\n *\n * Real backends pass through silently.\n */\nexport function assertRealBackend(\n  records: ReadonlyArray<RunRecord>,\n  opts: { allowMixed?: boolean } = {},\n): BackendIntegrityReport {\n  const report = summarizeBackendIntegrity(records)\n  const allowMixed = opts.allowMixed ?? true\n  if (report.verdict === 'stub') {\n    throw new BackendIntegrityError(\n      `backend-integrity: ran against a stub or unconfigured backend — ${report.diagnosis}`,\n      report,\n    )\n  }\n  if (!allowMixed && report.verdict === 'mixed') {\n    throw new BackendIntegrityError(\n      `backend-integrity: partial backend failure rejected — ${report.diagnosis}`,\n      report,\n    )\n  }\n  return report\n}\n","import { createRequire } from 'node:module'\n\n/**\n * @experimental\n *\n * `CampaignStorage` — the filesystem seam `runCampaign` writes through\n * (run/cell dirs, the resumability cache, per-cell artifacts, trace spans).\n *\n * The default (`fsCampaignStorage`) is the Node filesystem — identical\n * behavior to the inline `node:fs` calls it replaces, so existing CLI\n * consumers are unaffected. `inMemoryCampaignStorage` keeps everything in a\n * `Map`, so the substrate runs in environments WITHOUT a filesystem\n * (Cloudflare Workers, Deno Deploy, other edge runtimes) — the campaign\n * still produces its `CampaignResult` (cells + aggregates) in memory;\n * artifacts/traces simply aren't persisted to disk.\n *\n * Paths are opaque keys to the in-memory adapter — it does not parse them,\n * so the same `join(...)`-built paths work unchanged across both adapters.\n */\nexport interface CampaignStorage {\n  /** Ensure a directory exists (recursive). No-op for in-memory. */\n  ensureDir(dir: string): void\n  /** Does this path exist (as a written file or an ensured dir)? */\n  exists(path: string): boolean\n  /** Read a UTF-8 file; `undefined` when missing or unreadable. */\n  read(path: string): string | undefined\n  /** Write a file (string or bytes). Parent dir is assumed ensured. */\n  write(path: string, content: string | Uint8Array): void\n}\n\n/** Node-filesystem storage — the default. Lazily requires `node:fs` so the\n *  module imports cleanly in non-Node runtimes (where the caller passes\n *  `inMemoryCampaignStorage` instead and never constructs this).\n *\n *  `createRequire(import.meta.url)` is the ESM-native lazy require — a bare\n *  `require` is a ReferenceError under `\"type\": \"module\"`, which is exactly\n *  the shape this package publishes. */\nexport function fsCampaignStorage(): CampaignStorage {\n  const nodeRequire = createRequire(import.meta.url)\n  const { existsSync, mkdirSync, readFileSync, writeFileSync } = nodeRequire(\n    'node:fs',\n  ) as typeof import('node:fs')\n  return {\n    ensureDir(dir) {\n      if (!existsSync(dir)) mkdirSync(dir, { recursive: true })\n    },\n    exists(path) {\n      return existsSync(path)\n    },\n    read(path) {\n      try {\n        return readFileSync(path, 'utf8')\n      } catch {\n        return undefined\n      }\n    },\n    write(path, content) {\n      writeFileSync(path, content as Uint8Array)\n    },\n  }\n}\n\n/** In-memory storage for filesystem-less runtimes. Artifacts + trace spans\n *  live in a `Map` for the duration of the run; the `CampaignResult` is\n *  fully populated, but nothing is persisted to disk. */\nexport function inMemoryCampaignStorage(): CampaignStorage {\n  const files = new Map<string, string | Uint8Array>()\n  const dirs = new Set<string>()\n  return {\n    ensureDir(dir) {\n      dirs.add(dir)\n    },\n    exists(path) {\n      return files.has(path) || dirs.has(path)\n    },\n    read(path) {\n      const value = files.get(path)\n      if (value === undefined) return undefined\n      return typeof value === 'string' ? value : new TextDecoder().decode(value)\n    },\n    write(path, content) {\n      files.set(path, content)\n    },\n  }\n}\n"],"mappings":";;;;;;;;AAaA,SAAS,kBAAkB;AAC3B,SAAS,YAAY;;;ACuCd,IAAM,wBAAN,cAAoC,eAAe;AAAA,EACxD,YACE,SACgB,QAChB;AACA,UAAM,qBAAqB,OAAO;AAFlB;AAAA,EAGlB;AAAA,EAHkB;AAIpB;AAEA,SAAS,aAAa,KAAyB;AAC7C,SAAO,IAAI,WAAW,UAAU,KAAK,IAAI,WAAW,WAAW;AACjE;AAEA,SAAS,iBAAiB,KAAyB;AACjD,SAAO,IAAI,WAAW,SAAS,KAAK,IAAI,YAAY;AACtD;AAOO,SAAS,0BACd,SACwB;AACxB,QAAM,eAAe,QAAQ;AAC7B,MAAI,cAAc;AAClB,MAAI,cAAc;AAClB,MAAI,kBAAkB;AACtB,MAAI,mBAAmB;AACvB,MAAI,oBAAoB;AACxB,MAAI,eAAe;AACnB,aAAW,OAAO,SAAS;AACzB,wBAAoB,IAAI,WAAW;AACnC,yBAAqB,IAAI,WAAW;AACpC,oBAAgB,IAAI;AACpB,QAAI,aAAa,GAAG,EAAG;AAAA,QAClB;AACL,QAAI,iBAAiB,GAAG,EAAG;AAAA,EAC7B;AACA,QAAM,UACJ,iBAAiB,IACb,SACA,gBAAgB,eACd,SACA,gBAAgB,IACd,SACA;AACV,QAAM,YAAY,eAAe;AAAA,IAC/B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AACD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,eAAe,GAAsD;AAC5E,MAAI,EAAE,iBAAiB,GAAG;AACxB,WAAO;AAAA,EACT;AACA,MAAI,EAAE,YAAY,QAAQ;AACxB,WAAO;AAAA,MACL,OAAO,EAAE,YAAY;AAAA,MACrB;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,GAAG;AAAA,EACZ;AACA,MAAI,EAAE,YAAY,SAAS;AACzB,UAAM,OAAQ,EAAE,cAAc,EAAE,eAAgB,KAAK,QAAQ,CAAC;AAC9D,WAAO;AAAA,MACL,GAAG,EAAE,WAAW,IAAI,EAAE,YAAY,aAAa,GAAG;AAAA,MAClD;AAAA,MACA;AAAA,IACF,EAAE,KAAK,GAAG;AAAA,EACZ;AAEA,MAAI,EAAE,kBAAkB,GAAG;AACzB,UAAM,OAAQ,EAAE,kBAAkB,EAAE,eAAgB,KAAK,QAAQ,CAAC;AAClE,WAAO;AAAA,MACL,GAAG,EAAE,YAAY,uCAAuC,EAAE,gBAAgB,SAAS,EAAE,iBAAiB;AAAA,MACtG,GAAG,EAAE,eAAe,KAAK,GAAG;AAAA,MAC5B;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,GAAG;AAAA,EACZ;AACA,SAAO,GAAG,EAAE,YAAY,uCAAuC,EAAE,gBAAgB,SAAS,EAAE,iBAAiB,aAAa,EAAE,aAAa,QAAQ,CAAC,CAAC;AACrJ;AASO,SAAS,kBACd,SACA,OAAiC,CAAC,GACV;AACxB,QAAM,SAAS,0BAA0B,OAAO;AAChD,QAAM,aAAa,KAAK,cAAc;AACtC,MAAI,OAAO,YAAY,QAAQ;AAC7B,UAAM,IAAI;AAAA,MACR,wEAAmE,OAAO,SAAS;AAAA,MACnF;AAAA,IACF;AAAA,EACF;AACA,MAAI,CAAC,cAAc,OAAO,YAAY,SAAS;AAC7C,UAAM,IAAI;AAAA,MACR,8DAAyD,OAAO,SAAS;AAAA,MACzE;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;;;ACxLA,SAAS,qBAAqB;AAqCvB,SAAS,oBAAqC;AACnD,QAAM,cAAc,cAAc,YAAY,GAAG;AACjD,QAAM,EAAE,YAAY,WAAW,cAAc,cAAc,IAAI;AAAA,IAC7D;AAAA,EACF;AACA,SAAO;AAAA,IACL,UAAU,KAAK;AACb,UAAI,CAAC,WAAW,GAAG,EAAG,WAAU,KAAK,EAAE,WAAW,KAAK,CAAC;AAAA,IAC1D;AAAA,IACA,OAAO,MAAM;AACX,aAAO,WAAW,IAAI;AAAA,IACxB;AAAA,IACA,KAAK,MAAM;AACT,UAAI;AACF,eAAO,aAAa,MAAM,MAAM;AAAA,MAClC,QAAQ;AACN,eAAO;AAAA,MACT;AAAA,IACF;AAAA,IACA,MAAM,MAAM,SAAS;AACnB,oBAAc,MAAM,OAAqB;AAAA,IAC3C;AAAA,EACF;AACF;AAKO,SAAS,0BAA2C;AACzD,QAAM,QAAQ,oBAAI,IAAiC;AACnD,QAAM,OAAO,oBAAI,IAAY;AAC7B,SAAO;AAAA,IACL,UAAU,KAAK;AACb,WAAK,IAAI,GAAG;AAAA,IACd;AAAA,IACA,OAAO,MAAM;AACX,aAAO,MAAM,IAAI,IAAI,KAAK,KAAK,IAAI,IAAI;AAAA,IACzC;AAAA,IACA,KAAK,MAAM;AACT,YAAM,QAAQ,MAAM,IAAI,IAAI;AAC5B,UAAI,UAAU,OAAW,QAAO;AAChC,aAAO,OAAO,UAAU,WAAW,QAAQ,IAAI,YAAY,EAAE,OAAO,KAAK;AAAA,IAC3E;AAAA,IACA,MAAM,MAAM,SAAS;AACnB,YAAM,IAAI,MAAM,OAAO;AAAA,IACzB;AAAA,EACF;AACF;;;AF8BA,eAAsB,YACpB,MAC+C;AAC/C,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,iBAAiB,KAAK,kBAAkB;AAC9C,QAAM,MAAM,KAAK,QAAQ,MAAM,oBAAI,KAAK;AACxC,QAAM,SAAS,KAAK,UAAU,CAAC;AAC/B,QAAM,UAAU,KAAK,WAAW,kBAAkB;AAElD,MAAI,OAAO,KAAK,WAAW,YAAY,KAAK,OAAO,KAAK,EAAE,WAAW,GAAG;AACtE,UAAM,IAAI,MAAM,gEAAgE;AAAA,EAClF;AACA,UAAQ,UAAU,KAAK,MAAM;AAE7B,QAAM,eAAe,oBAAoB;AAAA,IACvC,WAAW,KAAK;AAAA,IAChB;AAAA,IACA,aAAa,KAAK,SAAS,QAAQ;AAAA,IACnC;AAAA,IACA;AAAA,EACF,CAAC;AAED,QAAM,YAAY,IAAI;AACtB,QAAM,QAAyC,CAAC;AAChD,QAAM,kBAA0C,CAAC;AAGjD,QAAM,WAA0F,CAAC;AACjG,MAAI,YAAY;AAChB,aAAW,YAAY,KAAK,WAAW;AACrC,aAAS,MAAM,GAAG,MAAM,MAAM,OAAO;AACnC,YAAM,SAAS,GAAG,SAAS,EAAE,IAAI,GAAG;AACpC,YAAM,WAAW,OAAO;AACxB,eAAS,KAAK,EAAE,UAAU,KAAK,QAAQ,SAAS,CAAC;AACjD,mBAAa;AAAA,IACf;AAAA,EACF;AAGA,MAAI,eAAe;AACnB,MAAI,qBAAqB;AACzB,QAAM,kBAAkB,IAAI,gBAAgB;AAI5C,QAAM,QAAyB,CAAC;AAChC,MAAI,UAAU;AACd,QAAM,WAAW;AAEjB,WAAS,IAAI,GAAG,IAAI,gBAAgB,KAAK;AACvC,UAAM;AAAA,OACH,YAAY;AACX,eAAO,MAAM;AACX,gBAAM,QAAQ;AACd,cAAI,SAAS,SAAS,OAAQ;AAC9B,gBAAM,OAAO,SAAS,KAAK;AAC3B,cAAI,oBAAoB;AACtB,qBAAS,KAAK,YAAY,MAAM,sBAAsB,CAAC;AACvD;AAAA,UACF;AACA,gBAAM,SAAS,MAAM,YAAY;AAAA,YAC/B;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA,kBAAkB,KAAK,oBAAoB,wBAAwB,OAAO;AAAA,YAC1E,QAAQ,gBAAgB;AAAA,YACxB,mBAAmB,KAAK;AAAA,UAC1B,CAAC;AACD,mBAAS,KAAK,OAAO,IAAI;AACzB,2BAAiB,OAAO,MAAM,KAAK,eAAe,MAAM;AACxD,0BAAgB,OAAO,KAAK;AAC5B,iBAAO,OAAO,iBAAiB,OAAO,eAAe;AACrD,cAAI,KAAK,gBAAgB,UAAa,gBAAgB,KAAK,aAAa;AACtE,iCAAqB;AAAA,UACvB;AAEA,cAAI,KAAK,gBAAgB,KAAK,iBAAiB,SAAS,CAAC,OAAO,KAAK,OAAO;AAC1E,kBAAM,eAAe;AAAA,cACnB,OAAO,KAAK;AAAA,cACZ,MAAM,OAAO;AAAA,cACb,UAAU,KAAK;AAAA,cACf;AAAA,cACA;AAAA,YACF,CAAC,EAAE,MAAM,CAAC,QAAQ;AAGhB,sBAAQ;AAAA,gBACN,oCAAoC,OAAO,KAAK,MAAM,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,cAC7G;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF,GAAG;AAAA,IACL;AAAA,EACF;AACA,QAAM,QAAQ,IAAI,KAAK;AAEvB,QAAM,UAAU,IAAI;AACpB,WAAS,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,cAAc,EAAE,MAAM,CAAC;AAExD,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,WAAW,UAAU,YAAY;AAAA,IACjC,SAAS,QAAQ,YAAY;AAAA,IAC7B,YAAY,QAAQ,QAAQ,IAAI,UAAU,QAAQ;AAAA,IAClD,OAAO;AAAA,IACP;AAAA,IACA,QAAQ,KAAK;AAAA,IACb;AAAA,IACA,WAAW,KAAK,UAAU,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,MAAM,EAAE,KAAK,EAAE;AAAA,EACnE;AACF;AAgBA,eAAe,YACb,MAC2F;AAC3F,QAAM,UAAU,KAAK;AACrB,QAAM,UAAU,KAAK,KAAK,KAAK,QAAQ,KAAK,KAAK,OAAO,QAAQ,mBAAmB,GAAG,CAAC;AACvF,UAAQ,UAAU,OAAO;AAGzB,QAAM,YAAY,KAAK,SAAS,oBAAoB;AACpD,MAAI,KAAK,WAAW;AAClB,UAAM,MAAM,QAAQ,KAAK,SAAS;AAClC,QAAI,QAAQ,QAAW;AACrB,UAAI;AACF,cAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,YAAI,OAAO,WAAW,KAAK,KAAK,QAAQ;AACtC,iBAAO,EAAE,MAAM,EAAE,GAAG,QAAQ,QAAQ,KAAK,GAAG,iBAAiB,CAAC,EAAE;AAAA,QAClE;AAAA,MACF,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF;AAEA,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,QAAQ,KAAK,iBAAiB,KAAK,KAAK,QAAQ,OAAO;AAC7D,QAAM,kBAA0C,CAAC;AACjD,QAAM,YAAoC;AAAA,IACxC,MAAM,MAAM,MAAM,SAAS;AACzB,YAAM,WAAW,KAAK,SAAS,IAAI;AACnC,cAAQ,UAAU,KAAK,UAAU,IAAI,CAAC;AACtC,cAAQ,MAAM,UAAU,OAAO;AAC/B,sBAAgB,GAAG,KAAK,KAAK,MAAM,IAAI,IAAI,EAAE,IAAI;AACjD,aAAO;AAAA,IACT;AAAA,IACA,MAAM,UAAU,MAAM,OAAO;AAC3B,aAAO,UAAU,MAAM,MAAM,KAAK,UAAU,OAAO,MAAM,CAAC,CAAC;AAAA,IAC7D;AAAA,EACF;AACA,MAAI,YAAY;AAChB,QAAM,cAAkC,EAAE,OAAO,GAAG,QAAQ,EAAE;AAC9D,QAAM,OAA0B;AAAA,IAC9B,QAAQ,QAAQ,QAAQ;AACtB,mBAAa;AACb,YAAM,KAAK,QAAQ,MAAM,IAAI,EAAE,WAAW,OAAO,CAAC,EAAE,IAAI;AAAA,IAC1D;AAAA,IACA,cAAc,OAAO;AACnB,kBAAY,SAAS,MAAM;AAC3B,kBAAY,UAAU,MAAM;AAC5B,UAAI,MAAM,OAAQ,aAAY,UAAU,YAAY,UAAU,KAAK,MAAM;AAAA,IAC3E;AAAA,IACA,UAAU;AACR,aAAO;AAAA,IACT;AAAA,IACA,SAAS;AACP,aAAO,EAAE,GAAG,YAAY;AAAA,IAC1B;AAAA,EACF;AAEA,QAAM,YAAY,KAAK,KAAK,gBAAgB;AAAA,IAC1C,UAAU,KAAK,KAAK;AAAA,IACpB,KAAK,KAAK,KAAK;AAAA,EACjB,CAAC;AAMD,QAAM,YAAY,IAAI,gBAAgB;AACtC,QAAM,kBAAkB,MAAM,UAAU,MAAO,KAAK,OAAgC,MAAM;AAC1F,MAAI,KAAK,OAAO,QAAS,WAAU,MAAO,KAAK,OAAgC,MAAM;AAAA,MAChF,MAAK,OAAO,iBAAiB,SAAS,iBAAiB,EAAE,MAAM,KAAK,CAAC;AAE1E,QAAM,MAAuB;AAAA,IAC3B,QAAQ,KAAK,KAAK;AAAA,IAClB,KAAK,KAAK,KAAK;AAAA,IACf,MAAM,KAAK,KAAK;AAAA,IAChB,QAAQ,UAAU;AAAA,IAClB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,MAAI;AACJ,MAAI;AACJ,QAAM,YAAY,KAAK;AACvB,MAAI;AACJ,MAAI;AACF,UAAM,aAAa,KAAK,KAAK,SAAS,KAAK,KAAK,UAAU,GAAG;AAC7D,QAAI,cAAc,UAAa,YAAY,GAAG;AAK5C,iBAAW,MAAM,QAAQ,KAAK;AAAA,QAC5B;AAAA,QACA,IAAI,QAAe,CAAC,GAAG,WAAW;AAChC,yBAAe,WAAW,MAAM;AAC9B,sBAAU,MAAM,IAAI,MAAM,kBAAkB,CAAC;AAC7C;AAAA,cACE,IAAI;AAAA,gBACF,qBAAqB,SAAS,gBAAgB,KAAK,KAAK,MAAM;AAAA,cAChE;AAAA,YACF;AAAA,UACF,GAAG,SAAS;AACZ,cAAI,OAAQ,aAAwC,UAAU;AAC5D,YAAC,aAAuC,MAAM;AAAA,QAClD,CAAC;AAAA,MACH,CAAC;AAAA,IACH,OAAO;AACL,iBAAW,MAAM;AAAA,IACnB;AAAA,EACF,SAAS,KAAK;AACZ,mBAAe,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,EAChE,UAAE;AACA,QAAI,aAAc,cAAa,YAAY;AAC3C,SAAK,OAAO,oBAAoB,SAAS,eAAe;AAAA,EAC1D;AAKA,QAAM,cAA0C,CAAC;AACjD,MAAI,aAAa,QAAW;AAC1B,eAAW,SAAS,KAAK,KAAK,UAAU,CAAC,GAAG;AAC1C,UAAI,MAAM,aAAa,CAAC,MAAM,UAAU,KAAK,KAAK,QAAQ,EAAG;AAC7D,UAAI;AACF,oBAAY,MAAM,IAAI,IAAI,MAAM,aAAa,OAAO;AAAA,UAClD;AAAA,UACA,UAAU,KAAK,KAAK;AAAA,UACpB,QAAQ,KAAK;AAAA,QACf,CAAC;AAAA,MACH,SAAS,KAAK;AACZ,uBAAe,UAAU,MAAM,IAAI,aAAa,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAChG;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,MAAM,MAAM;AAElB,QAAM,OAAsC;AAAA,IAC1C,QAAQ,KAAK,KAAK;AAAA,IAClB,YAAY,KAAK,KAAK,SAAS;AAAA,IAC/B,KAAK,KAAK,KAAK;AAAA,IACf,UAAW,YAAY;AAAA,IACvB;AAAA,IACA,SAAS;AAAA,IACT,YAAY,EAAE,GAAG,YAAY;AAAA,IAC7B,YAAY,KAAK,IAAI,IAAI;AAAA,IACzB,MAAM,KAAK,KAAK;AAAA,IAChB,QAAQ;AAAA,IACR,OAAO;AAAA,EACT;AAEA,MAAI,CAAC,gBAAgB,KAAK,WAAW;AACnC,YAAQ,MAAM,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,EAC/C;AAEA,SAAO,EAAE,MAAM,gBAAgB;AACjC;AASA,SAAS,iBACP,MACA,MACM;AACN,MAAI,SAAS,SAAS,KAAK,MAAO;AAClC,MAAI,KAAK,aAAa,QAAQ,KAAK,aAAa,OAAW;AAC3D,QAAM,aAAa,KAAK,WAAW,UAAU,KAAK,KAAK,WAAW,WAAW;AAC7E,MAAI,KAAK,YAAY,KAAK,CAAC,WAAY;AACvC,QAAM,MAAM,SAAS,KAAK,MAAM;AAChC,MAAI,SAAS,UAAU;AACrB,UAAM,SAAiC;AAAA,MACrC,cAAc;AAAA,MACd,aAAa;AAAA,MACb,aAAa;AAAA,MACb,iBAAiB;AAAA,MACjB,kBAAkB;AAAA,MAClB,mBAAmB;AAAA,MACnB,cAAc;AAAA,MACd,SAAS;AAAA,MACT,WAAW;AAAA,IACb;AACA,UAAM,IAAI,sBAAsB,gBAAgB,GAAG,IAAI,MAAM;AAAA,EAC/D;AAEA,UAAQ,KAAK,8BAA8B,GAAG,EAAE;AAClD;AAEA,eAAe,aACb,OACA,OACqB;AACrB,SAAO,MAAM,MAAM,KAAK;AAC1B;AAEA,SAAS,wBACP,SACsD;AACtD,SAAO,CAAC,QAAQ,QAAQ;AACtB,UAAM,QAAwC,CAAC;AAC/C,WAAO;AAAA,MACL,KAAK,MAAM,YAAY;AACrB,cAAM,UAAU,KAAK,IAAI;AACzB,cAAM,SAAkC,EAAE,MAAM,QAAQ,SAAS,GAAI,cAAc,CAAC,EAAG;AACvF,cAAM,SAAoB;AAAA,UACxB,IAAI,UAAU;AACZ,mBAAO,aAAa,KAAK,IAAI,IAAI;AACjC,gBAAI,SAAU,QAAO,OAAO,QAAQ,QAAQ;AAC5C,kBAAM,KAAK,MAAM;AAAA,UACnB;AAAA,UACA,aAAa,KAAK,OAAO;AACvB,mBAAO,GAAG,IAAI;AAAA,UAChB;AAAA,QACF;AACA,eAAO;AAAA,MACT;AAAA,MACA,MAAM,QAAQ;AACZ,gBAAQ,MAAM,KAAK,KAAK,aAAa,GAAG,MAAM,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,MACxF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,YACP,MACA,QAC+B;AAC/B,SAAO;AAAA,IACL,QAAQ,KAAK;AAAA,IACb,YAAY,KAAK,SAAS;AAAA,IAC1B,KAAK,KAAK;AAAA,IACV,UAAU;AAAA,IACV,aAAa,CAAC;AAAA,IACd,SAAS;AAAA,IACT,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,IAClC,YAAY;AAAA,IACZ,MAAM,KAAK;AAAA,IACX,QAAQ;AAAA,IACR,OAAO,YAAY,MAAM;AAAA,EAC3B;AACF;AAUA,eAAe,eACb,MACe;AACf,QAAM,KAAK,MAAM,QAAQ;AAAA,IACvB,UAAU,KAAK;AAAA,IACf,UAAU,KAAK,KAAK;AAAA,IACpB,aAAa,KAAK,KAAK;AAAA,IACvB,QAAQ,KAAK,KAAK,iBAAiB;AAAA,IACnC,mBAAmB,KAAK,KAAK,4BAA4B;AAAA,IACzD,YAAY,KAAK,IAAI,EAAE,YAAY;AAAA,IACnC,iBAAiB;AAAA,EACnB,CAAC;AACH;AAIA,SAAS,oBAAoB,OAMlB;AACT,QAAM,YAAY;AAAA,IAChB,WAAW,MAAM,UAAU,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,MAAM,EAAE,KAAK,EAAE;AAAA,IAClE,QAAQ,MAAM,OAAO,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,MAAM,EAAE,WAAW,IAAI,CAAC,MAAM,EAAE,GAAG,EAAE,EAAE;AAAA,IACxF,UAAU,MAAM;AAAA,IAChB,MAAM,MAAM;AAAA,IACZ,MAAM,MAAM;AAAA,EACd;AACA,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,UAAU,SAAS,CAAC,EAAE,OAAO,KAAK;AAC5E;AAEA,SAAS,kBACP,OACA,QACA,MACoB;AACpB,QAAM,UAA0C,CAAC;AACjD,aAAW,SAAS,QAAQ;AAC1B,UAAM,SAAmB,CAAC;AAC1B,eAAW,QAAQ,OAAO;AACxB,YAAM,IAAI,KAAK,YAAY,MAAM,IAAI;AACrC,UAAI,MAAM,OAAW,QAAO,KAAK,EAAE,SAAS;AAAA,IAC9C;AACA,YAAQ,MAAM,IAAI,IAAI,UAAU,QAAQ,IAAI;AAAA,EAC9C;AACA,QAAM,aAAgD,CAAC;AACvD,QAAM,iBAAiB,oBAAI,IAAsB;AACjD,aAAW,QAAQ,OAAO;AACxB,UAAM,aAAa,OAAO,OAAO,KAAK,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AACzE,QAAI,WAAW,WAAW,EAAG;AAC7B,UAAM,OAAO,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;AAChE,UAAM,MAAM,eAAe,IAAI,KAAK,UAAU,KAAK,CAAC;AACpD,QAAI,KAAK,IAAI;AACb,mBAAe,IAAI,KAAK,YAAY,GAAG;AAAA,EACzC;AACA,aAAW,CAAC,YAAY,OAAO,KAAK,gBAAgB;AAClD,UAAM,KAAK,UAAU,SAAS,IAAI;AAClC,eAAW,UAAU,IAAI,EAAE,eAAe,GAAG,MAAM,MAAM,GAAG,MAAM,GAAG,GAAG,EAAE;AAAA,EAC5E;AACA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,SAAS,CAAC;AAAA,IACrD,eAAe,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,KAAK,EAAE;AAAA,IAC7C,cAAc,MAAM,OAAO,CAAC,MAAM,EAAE,OAAO,WAAW,UAAU,CAAC,EAAE;AAAA,IACnE,aAAa,MAAM,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE;AAAA,IAC3C,aAAa,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC,EAAE,MAAM,WAAW,UAAU,CAAC,EAAE;AAAA,EAC/E;AACF;AAKA,SAAS,UAAU,SAAmB,MAA8B;AAClE,QAAM,IAAI,QAAQ;AAClB,MAAI,MAAM,EAAG,QAAO,EAAE,MAAM,GAAG,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC,GAAG,GAAG,EAAE;AAC5D,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAClD,QAAM,WAAW,QAAQ,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,SAAS,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AACrF,QAAM,QAAQ,KAAK,KAAK,QAAQ;AAChC,QAAM,KAAK,mBAAmB,SAAS,MAAM,EAAE,MAAM,WAAW,IAAK,CAAC;AACtE,SAAO,EAAE,MAAM,OAAO,MAAM,CAAC,GAAG,OAAO,GAAG,KAAK,GAAG,EAAE;AACtD;","names":[]}

package/dist/chunk-UMMZHCPB.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/default-registry.ts"],"sourcesContent":["/**\n * Deterministic behavioral metrics over OTLP spans — pure arithmetic, no LLM.\n *\n * These are the model-independent multiplier: the four trace-quality signals a\n * tolerant analyzer (e.g. HALO) re-derives per run inside the model — token\n * growth, output decay, tool monoculture, missing self-verification — computed\n * here once, in TypeScript, with zero model judgment. A finding that falls out\n * of arithmetic is trivially model-agnostic and cannot hallucinate the trend.\n *\n * General, not trace-specific: the detectors key off token trajectories and\n * tool usage present in any agentic OTLP trace, not any one benchmark.\n */\n\nimport type { TraceAnalystSpan } from './types'\n\nexport type SuboptimalCode =\n  | 'monotonic-input-growth'\n  | 'output-length-decay'\n  | 'single-tool-dependency'\n  | 'no-self-verification'\n\nexport interface SuboptimalSignal {\n  code: SuboptimalCode\n  severity: 'high' | 'medium' | 'low'\n  /** Human-readable claim, with the backing numbers inlined. */\n  detail: string\n  /** The exact figures the detector fired on — auditable, no model in the loop. */\n  evidence: Record<string, number | string | boolean>\n}\n\nexport interface BehavioralMetrics {\n  llmCallCount: number\n  inputTokenTrajectory: number[]\n  outputTokenTrajectory: number[]\n  toolHistogram: Record<string, number>\n  totalToolCalls: number\n  distinctTools: number\n  /** distinct/total tool calls; 1.0 when there are no tool calls. */\n  toolDiversityRatio: number\n  hasSelfVerification: boolean\n  signals: SuboptimalSignal[]\n}\n\n/** ≥ this input-token growth ratio across a run, with no compression, fires. */\nconst INPUT_GROWTH_FACTOR = 3\n/** Tool-usage signals need at least this many calls to be meaningful. */\nconst MIN_TOOL_CALLS = 3\n/** Tool names matching this are self-verification, not state mutation. */\nconst VERIFY_RE = /verif|eval|inspect|check|assert|validat|review|confirm/i\n\nfunction num(v: unknown): number | null {\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\nfunction inputTokensOf(s: TraceAnalystSpan): number | null {\n  return num(s.attributes['llm.input_tokens']) ?? num(s.attributes['llm.usage.input_tokens'])\n}\nfunction outputTokensOf(s: TraceAnalystSpan): number | null {\n  return num(s.attributes['llm.output_tokens']) ?? num(s.attributes['llm.usage.output_tokens'])\n}\nfunction stepOf(s: TraceAnalystSpan): number | null {\n  return num(s.attributes.step)\n}\nfunction toolNameOf(s: TraceAnalystSpan): string | null {\n  if (s.tool_name) return s.tool_name\n  const t = s.attributes['tool.name']\n  return typeof t === 'string' && t.length > 0 ? t : null\n}\n\n/**\n * Reduce a span list to behavioral metrics + fired suboptimality signals.\n * Pure + deterministic: same spans → same output, on any machine, no model.\n */\nexport function computeTraceMetrics(spans: readonly TraceAnalystSpan[]): BehavioralMetrics {\n  // Order by step (when present) then start_time so trajectories reflect run order.\n  const ordered = [...spans].sort((a, b) => {\n    const sa = stepOf(a)\n    const sb = stepOf(b)\n    if (sa !== null && sb !== null && sa !== sb) return sa - sb\n    return a.start_time.localeCompare(b.start_time)\n  })\n\n  const inputTokenTrajectory: number[] = []\n  const outputTokenTrajectory: number[] = []\n  const toolHistogram: Record<string, number> = {}\n  let hasSelfVerification = false\n\n  for (const s of ordered) {\n    const inT = inputTokensOf(s)\n    if (inT !== null) inputTokenTrajectory.push(inT)\n    const outT = outputTokensOf(s)\n    if (outT !== null) outputTokenTrajectory.push(outT)\n    const tool = toolNameOf(s)\n    if (tool) {\n      toolHistogram[tool] = (toolHistogram[tool] ?? 0) + 1\n      if (VERIFY_RE.test(tool)) hasSelfVerification = true\n    }\n  }\n\n  const totalToolCalls = Object.values(toolHistogram).reduce((a, b) => a + b, 0)\n  const distinctTools = Object.keys(toolHistogram).length\n  const toolDiversityRatio = totalToolCalls === 0 ? 1 : distinctTools / totalToolCalls\n\n  const signals: SuboptimalSignal[] = []\n\n  if (inputTokenTrajectory.length >= 3) {\n    const first = inputTokenTrajectory[0]!\n    const last = inputTokenTrajectory[inputTokenTrajectory.length - 1]!\n    const growth = first > 0 ? last / first : 0\n    if (last > first && growth >= INPUT_GROWTH_FACTOR) {\n      signals.push({\n        code: 'monotonic-input-growth',\n        severity: 'high',\n        detail: `LLM input tokens grew ${growth.toFixed(1)}x (${first}→${last}) across ${inputTokenTrajectory.length} calls — full history re-sent each step with no compression.`,\n        evidence: {\n          first,\n          last,\n          growth_x: Number(growth.toFixed(2)),\n          calls: inputTokenTrajectory.length,\n        },\n      })\n    }\n  }\n\n  if (outputTokenTrajectory.length >= 3) {\n    const first = outputTokenTrajectory[0]!\n    const last = outputTokenTrajectory[outputTokenTrajectory.length - 1]!\n    if (last < first) {\n      signals.push({\n        code: 'output-length-decay',\n        severity: 'medium',\n        detail: `LLM output tokens shrank ${first}→${last} over ${outputTokenTrajectory.length} calls — less planning/reasoning per step as context grows.`,\n        evidence: { first, last, calls: outputTokenTrajectory.length },\n      })\n    }\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && distinctTools === 1) {\n    const only = Object.keys(toolHistogram)[0]!\n    signals.push({\n      code: 'single-tool-dependency',\n      severity: 'medium',\n      detail: `All ${totalToolCalls} tool calls are \\`${only}\\` — no tool diversity and no fallback path.`,\n      evidence: { tool: only, calls: totalToolCalls, distinct_tools: 1 },\n    })\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && !hasSelfVerification) {\n    signals.push({\n      code: 'no-self-verification',\n      severity: 'medium',\n      detail: `${totalToolCalls} tool calls and none verify/inspect/check state — the agent never validates its own actions.`,\n      evidence: { tool_calls: totalToolCalls, verification_calls: 0 },\n    })\n  }\n\n  return {\n    llmCallCount: inputTokenTrajectory.length,\n    inputTokenTrajectory,\n    outputTokenTrajectory,\n    toolHistogram,\n    totalToolCalls,\n    distinctTools,\n    toolDiversityRatio,\n    hasSelfVerification,\n    signals,\n  }\n}\n","/**\n * `behavioralAnalyst` — a DETERMINISTIC analyst (cost.kind = 'deterministic',\n * never calls the LLM). It produces the efficiency/behavioral findings a\n * tolerant agentic analyzer (HALO) re-derives per run inside the model —\n * context bloat, output decay, tool monoculture, missing self-verification —\n * directly from arithmetic over spans (`computeTraceMetrics`).\n *\n * Why it matters: these findings are model-agnostic BY CONSTRUCTION (no model\n * in the loop), so they cannot return 0 on a weak model the way the Ax-RLM\n * does — and they are strictly more reliable than HALO, which spends tokens\n * re-deriving the same numbers and can hallucinate the trend. The agentic\n * RLM kinds remain for SEMANTIC findings that genuinely need a model; this\n * analyst owns the behavioral class.\n */\n\nimport {\n  type BehavioralMetrics,\n  computeTraceMetrics,\n  type SuboptimalCode,\n} from '../trace-analyst/behavioral-metrics'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { TraceAnalystSpan } from '../trace-analyst/types'\nimport { type Analyst, type AnalystFinding, makeFinding } from './types'\n\nconst RECOMMENDED_ACTION: Record<SuboptimalCode, string> = {\n  'monotonic-input-growth':\n    'Add a context-budget instruction: once prior context exceeds a threshold, summarize earlier steps into a short status line instead of re-sending full history.',\n  'output-length-decay':\n    'Require a minimum planning/reasoning budget per step so late steps do not degrade into terse, error-prone commands.',\n  'single-tool-dependency':\n    'Direct the agent to use the full toolset (verify / inspect / alternate actions), not a single execute call, and to plan a fallback when a call returns an unexpected result.',\n  'no-self-verification':\n    'After every state-mutating action, verify the result (eval / inspect / assert) before proceeding.',\n}\n\nconst ANALYST_ID = 'efficiency-behavioral'\n\n/**\n * Map computed signals → structured AnalystFindings. Pure: no LLM, no clock\n * dependence beyond `produced_at` (overridable for deterministic tests).\n */\nexport function deriveEfficiencyFindings(\n  metrics: BehavioralMetrics,\n  opts: { analystId?: string; producedAt?: string } = {},\n): AnalystFinding[] {\n  const analystId = opts.analystId ?? ANALYST_ID\n  return metrics.signals.map((sig) =>\n    makeFinding({\n      analyst_id: analystId,\n      area: 'efficiency',\n      subject: sig.code, // kebab — passes the cluster grammar; stable key for diffFindings\n      claim: sig.detail,\n      severity: sig.severity,\n      // Deterministic arithmetic over spans, not a model judgment → certain.\n      confidence: 1,\n      evidence_refs: [\n        {\n          kind: 'metric',\n          uri: `metric://efficiency/${sig.code}`,\n          excerpt: JSON.stringify(sig.evidence),\n        },\n      ],\n      recommended_action: RECOMMENDED_ACTION[sig.code],\n      metadata: { deterministic: true, evidence: sig.evidence },\n      ...(opts.producedAt ? { produced_at: opts.producedAt } : {}),\n    }),\n  )\n}\n\n/** The deterministic behavioral/efficiency analyst (no LLM, any-model). */\nexport function behavioralAnalyst(): Analyst<TraceAnalysisStore> {\n  return {\n    id: ANALYST_ID,\n    description:\n      'Deterministic behavioral/efficiency findings over OTLP spans — token-growth, output-decay, tool-monoculture, missing self-verification. Zero LLM; model-agnostic by construction.',\n    inputKind: 'trace-store',\n    cost: { kind: 'deterministic' },\n    version: '1.0.0',\n    async analyze(store) {\n      const overview = await store.getOverview()\n      const spans: TraceAnalystSpan[] = []\n      for (const traceId of overview.sample_trace_ids) {\n        const viewed = await store.viewTrace({ trace_id: traceId })\n        if (viewed.spans) spans.push(...viewed.spans)\n      }\n      return deriveEfficiencyFindings(computeTraceMetrics(spans))\n    },\n  }\n}\n","/**\n * `buildDefaultAnalystRegistry` — the canonical analyst suite, so consumers\n * stop hand-wiring `new AnalystRegistry()` + per-kind `createTraceAnalystKind`.\n *\n * The deterministic `behavioralAnalyst` is ALWAYS registered (it needs no\n * model and is model-agnostic by construction). The agentic RLM kinds are\n * registered only when an `ai` service is supplied — so a caller with no LLM\n * still gets the full behavioral/efficiency diagnosis, and the substrate's\n * \"any model (including no model)\" guarantee holds at the suite level.\n */\n\nimport type { AxAIService } from '@ax-llm/ax'\nimport { behavioralAnalyst } from './behavioral-analyst'\nimport { createTraceAnalystKind, type TraceAnalystKindSpec } from './kind-factory'\nimport { DEFAULT_TRACE_ANALYST_KINDS } from './kinds'\nimport { AnalystRegistry, type AnalystRegistryOptions } from './registry'\n\nexport interface DefaultAnalystRegistryOptions {\n  /** Ax service for the agentic RLM kinds. Omit → only the deterministic analyst. */\n  ai?: AxAIService\n  /** Model for the agentic kinds (falls back to the ai service default). */\n  model?: string\n  /** Which agentic kinds to register when `ai` is present. Default = the shipped suite. */\n  kinds?: readonly TraceAnalystKindSpec[]\n  /** Set false to omit the deterministic behavioral analyst (default: include). */\n  includeBehavioral?: boolean\n  /** Forwarded to the AnalystRegistry constructor (signal, tags, priorFindings). */\n  registry?: AnalystRegistryOptions\n}\n\nexport function buildDefaultAnalystRegistry(\n  opts: DefaultAnalystRegistryOptions = {},\n): AnalystRegistry {\n  const registry = new AnalystRegistry(opts.registry)\n  if (opts.includeBehavioral !== false) {\n    registry.register(behavioralAnalyst())\n  }\n  if (opts.ai) {\n    const kinds = opts.kinds ?? DEFAULT_TRACE_ANALYST_KINDS\n    for (const spec of kinds) {\n      registry.register(createTraceAnalystKind(spec, { ai: opts.ai, model: opts.model }))\n    }\n  }\n  return registry\n}\n"],"mappings":";;;;;;;;;;AA4CA,IAAM,sBAAsB;AAE5B,IAAM,iBAAiB;AAEvB,IAAM,YAAY;AAElB,SAAS,IAAI,GAA2B;AACtC,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AACA,SAAS,cAAc,GAAoC;AACzD,SAAO,IAAI,EAAE,WAAW,kBAAkB,CAAC,KAAK,IAAI,EAAE,WAAW,wBAAwB,CAAC;AAC5F;AACA,SAAS,eAAe,GAAoC;AAC1D,SAAO,IAAI,EAAE,WAAW,mBAAmB,CAAC,KAAK,IAAI,EAAE,WAAW,yBAAyB,CAAC;AAC9F;AACA,SAAS,OAAO,GAAoC;AAClD,SAAO,IAAI,EAAE,WAAW,IAAI;AAC9B;AACA,SAAS,WAAW,GAAoC;AACtD,MAAI,EAAE,UAAW,QAAO,EAAE;AAC1B,QAAM,IAAI,EAAE,WAAW,WAAW;AAClC,SAAO,OAAO,MAAM,YAAY,EAAE,SAAS,IAAI,IAAI;AACrD;AAMO,SAAS,oBAAoB,OAAuD;AAEzF,QAAM,UAAU,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM;AACxC,UAAM,KAAK,OAAO,CAAC;AACnB,UAAM,KAAK,OAAO,CAAC;AACnB,QAAI,OAAO,QAAQ,OAAO,QAAQ,OAAO,GAAI,QAAO,KAAK;AACzD,WAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,EAChD,CAAC;AAED,QAAM,uBAAiC,CAAC;AACxC,QAAM,wBAAkC,CAAC;AACzC,QAAM,gBAAwC,CAAC;AAC/C,MAAI,sBAAsB;AAE1B,aAAW,KAAK,SAAS;AACvB,UAAM,MAAM,cAAc,CAAC;AAC3B,QAAI,QAAQ,KAAM,sBAAqB,KAAK,GAAG;AAC/C,UAAM,OAAO,eAAe,CAAC;AAC7B,QAAI,SAAS,KAAM,uBAAsB,KAAK,IAAI;AAClD,UAAM,OAAO,WAAW,CAAC;AACzB,QAAI,MAAM;AACR,oBAAc,IAAI,KAAK,cAAc,IAAI,KAAK,KAAK;AACnD,UAAI,UAAU,KAAK,IAAI,EAAG,uBAAsB;AAAA,IAClD;AAAA,EACF;AAEA,QAAM,iBAAiB,OAAO,OAAO,aAAa,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC7E,QAAM,gBAAgB,OAAO,KAAK,aAAa,EAAE;AACjD,QAAM,qBAAqB,mBAAmB,IAAI,IAAI,gBAAgB;AAEtE,QAAM,UAA8B,CAAC;AAErC,MAAI,qBAAqB,UAAU,GAAG;AACpC,UAAM,QAAQ,qBAAqB,CAAC;AACpC,UAAM,OAAO,qBAAqB,qBAAqB,SAAS,CAAC;AACjE,UAAM,SAAS,QAAQ,IAAI,OAAO,QAAQ;AAC1C,QAAI,OAAO,SAAS,UAAU,qBAAqB;AACjD,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,yBAAyB,OAAO,QAAQ,CAAC,CAAC,MAAM,KAAK,SAAI,IAAI,YAAY,qBAAqB,MAAM;AAAA,QAC5G,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,UAAU,OAAO,OAAO,QAAQ,CAAC,CAAC;AAAA,UAClC,OAAO,qBAAqB;AAAA,QAC9B;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MAAI,sBAAsB,UAAU,GAAG;AACrC,UAAM,QAAQ,sBAAsB,CAAC;AACrC,UAAM,OAAO,sBAAsB,sBAAsB,SAAS,CAAC;AACnE,QAAI,OAAO,OAAO;AAChB,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,4BAA4B,KAAK,SAAI,IAAI,SAAS,sBAAsB,MAAM;AAAA,QACtF,UAAU,EAAE,OAAO,MAAM,OAAO,sBAAsB,OAAO;AAAA,MAC/D,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MAAI,kBAAkB,kBAAkB,kBAAkB,GAAG;AAC3D,UAAM,OAAO,OAAO,KAAK,aAAa,EAAE,CAAC;AACzC,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,OAAO,cAAc,qBAAqB,IAAI;AAAA,MACtD,UAAU,EAAE,MAAM,MAAM,OAAO,gBAAgB,gBAAgB,EAAE;AAAA,IACnE,CAAC;AAAA,EACH;AAEA,MAAI,kBAAkB,kBAAkB,CAAC,qBAAqB;AAC5D,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,GAAG,cAAc;AAAA,MACzB,UAAU,EAAE,YAAY,gBAAgB,oBAAoB,EAAE;AAAA,IAChE,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,cAAc,qBAAqB;AAAA,IACnC;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;;;AC9IA,IAAM,qBAAqD;AAAA,EACzD,0BACE;AAAA,EACF,uBACE;AAAA,EACF,0BACE;AAAA,EACF,wBACE;AACJ;AAEA,IAAM,aAAa;AAMZ,SAAS,yBACd,SACA,OAAoD,CAAC,GACnC;AAClB,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO,QAAQ,QAAQ;AAAA,IAAI,CAAC,QAC1B,YAAY;AAAA,MACV,YAAY;AAAA,MACZ,MAAM;AAAA,MACN,SAAS,IAAI;AAAA;AAAA,MACb,OAAO,IAAI;AAAA,MACX,UAAU,IAAI;AAAA;AAAA,MAEd,YAAY;AAAA,MACZ,eAAe;AAAA,QACb;AAAA,UACE,MAAM;AAAA,UACN,KAAK,uBAAuB,IAAI,IAAI;AAAA,UACpC,SAAS,KAAK,UAAU,IAAI,QAAQ;AAAA,QACtC;AAAA,MACF;AAAA,MACA,oBAAoB,mBAAmB,IAAI,IAAI;AAAA,MAC/C,UAAU,EAAE,eAAe,MAAM,UAAU,IAAI,SAAS;AAAA,MACxD,GAAI,KAAK,aAAa,EAAE,aAAa,KAAK,WAAW,IAAI,CAAC;AAAA,IAC5D,CAAC;AAAA,EACH;AACF;AAGO,SAAS,oBAAiD;AAC/D,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS;AAAA,IACT,MAAM,QAAQ,OAAO;AACnB,YAAM,WAAW,MAAM,MAAM,YAAY;AACzC,YAAM,QAA4B,CAAC;AACnC,iBAAW,WAAW,SAAS,kBAAkB;AAC/C,cAAM,SAAS,MAAM,MAAM,UAAU,EAAE,UAAU,QAAQ,CAAC;AAC1D,YAAI,OAAO,MAAO,OAAM,KAAK,GAAG,OAAO,KAAK;AAAA,MAC9C;AACA,aAAO,yBAAyB,oBAAoB,KAAK,CAAC;AAAA,IAC5D;AAAA,EACF;AACF;;;AC1DO,SAAS,4BACd,OAAsC,CAAC,GACtB;AACjB,QAAM,WAAW,IAAI,gBAAgB,KAAK,QAAQ;AAClD,MAAI,KAAK,sBAAsB,OAAO;AACpC,aAAS,SAAS,kBAAkB,CAAC;AAAA,EACvC;AACA,MAAI,KAAK,IAAI;AACX,UAAM,QAAQ,KAAK,SAAS;AAC5B,eAAW,QAAQ,OAAO;AACxB,eAAS,SAAS,uBAAuB,MAAM,EAAE,IAAI,KAAK,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC;AAAA,IACpF;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/chunk-XY4DDNEG.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/analyst/ax-service.ts","../src/analyst/chat-client.ts","../src/concurrency.ts","../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/skill-usage.ts","../src/run-score.ts","../src/run-critic.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["import type { AxAIService } from '@ax-llm/ax'\nimport { ai } from '@ax-llm/ax'\n\nexport interface CreateAnalystAiConfig {\n  /** OpenAI-compatible API key forwarded as `Authorization: Bearer`.\n   *  cli-bridge ignores the value on loopback but Ax requires a non-empty string. */\n  apiKey: string\n  /** OpenAI-compatible base URL — e.g. `https://router.tangle.tools/v1` or a\n   *  cli-bridge loopback. */\n  baseUrl: string\n  /** Model id forwarded to the analyst actor + responder. */\n  model: string\n  /** Ax provider name. Defaults to the OpenAI-compatible client. */\n  provider?: 'openai' | 'anthropic'\n}\n\n/**\n * Construct the `AxAIService` an analyst kind calls through\n * (`createTraceAnalystKind({ ai })`).\n *\n * Ax's `ai()` pins `config.model` to the OpenAI catalog enum, but every\n * OpenAI-compatible router an analyst points at (router.tangle.tools,\n * cli-bridge) accepts arbitrary model ids (claude-code/sonnet, openai/gpt-5.4,\n * …). Consumers were each re-rolling `ai({ name, apiKey, apiURL, config })`\n * behind an `as (a: any) => any` cast to dodge the enum; this is the one\n * canonical constructor so they don't have to — and don't take a direct\n * `@ax-llm/ax` dependency for it.\n */\nexport function createAnalystAi(config: CreateAnalystAiConfig): AxAIService {\n  return ai({\n    name: config.provider ?? 'openai',\n    apiKey: config.apiKey,\n    apiURL: config.baseUrl,\n    config: { model: config.model },\n  })\n}\n","/**\n * ChatClient — the single LLM abstraction analysts call.\n *\n * agent-eval already ships an `LlmClient` (OpenAI-compatible, retry,\n * graceful JSON-schema degrade) and judges that talk to `TCloud`. Two\n * mixed patterns force every analyst author to pick a transport, which\n * couples analyst code to runtime concerns (cli-bridge vs router vs\n * sandbox-sdk) it shouldn't know about.\n *\n * `ChatClient` is one interface every analyst takes via `AnalystContext.chat`.\n * The operator decides at the registry boundary which transport binds\n * to it. Analyst code stays transport-agnostic; swapping production\n * (sandbox-sdk) for local dev (cli-bridge) or tests (mock) is a one-\n * line factory call.\n *\n * Designed to coexist: existing `LlmClient` callers and existing\n * `TCloud`-based judges keep working untouched. New analyst code uses\n * `ChatClient`. When old call sites migrate, they pick up budgeting,\n * cancellation, and unified telemetry for free.\n */\n\nimport {\n  type LlmCallRequest,\n  type LlmCallResult,\n  LlmClient,\n  type LlmClientOptions,\n} from '../llm-client'\n\n/**\n * Unified chat interface. Mirrors LlmCallRequest/Result so the OpenAI-\n * compatible mental model stays. Two methods: a one-shot `chat()` and\n * an `streamChat()` for future agentic loops (not yet exposed).\n */\nexport interface ChatClient {\n  /** Display name of the bound transport — included in telemetry. */\n  readonly transport: ChatTransport\n  /** Default model when caller omits — operators bind this per environment. */\n  readonly defaultModel?: string\n\n  chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>\n}\n\nexport type ChatTransport =\n  | 'router' // router.tangle.tools — production paid models\n  | 'sandbox-sdk' // box.streamPrompt() — chat completion via sandbox SDK\n  | 'cli-bridge' // local cli-bridge for dev / local-only runs\n  | 'direct-provider' // direct OpenAI / Anthropic / etc. — bypass router\n  | 'mock' // test-time injection\n\nexport interface ChatRequest extends Omit<LlmCallRequest, 'model'> {\n  /** Optional — falls back to ChatClient.defaultModel. */\n  model?: string\n}\n\nexport type ChatResponse = LlmCallResult\n\nexport interface ChatCallOpts {\n  /** Cancel the in-flight request. */\n  signal?: AbortSignal\n  /** Hard USD ceiling for this single call (informational; the underlying transport may not enforce). */\n  maxCostUsd?: number\n  /** Correlation tag carried into request headers when the transport allows. */\n  correlationId?: string\n}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | RouterTransportOpts\n  | CliBridgeTransportOpts\n  | DirectProviderTransportOpts\n  | SandboxSdkTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\n}\n\nexport interface RouterTransportOpts extends BaseTransportOpts {\n  transport: 'router'\n  baseUrl?: string\n  apiKey: string\n}\n\nexport interface CliBridgeTransportOpts extends BaseTransportOpts {\n  transport: 'cli-bridge'\n  baseUrl?: string\n  bearer?: string\n}\n\nexport interface DirectProviderTransportOpts extends BaseTransportOpts {\n  transport: 'direct-provider'\n  baseUrl: string\n  apiKey: string\n}\n\n/**\n * Sandbox-SDK transport. Provided as a thin pass-through: the caller\n * supplies a callable that mimics LlmClient.chat() against an already-\n * configured Sandbox handle. We don't import the SDK here to keep\n * agent-eval dep-free of @tangle-network/sandbox.\n */\nexport interface SandboxSdkTransportOpts extends BaseTransportOpts {\n  transport: 'sandbox-sdk'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Mock transport for tests. The handler receives the request and returns\n * whatever the test wants. No retries, no JSON-schema degrade.\n */\nexport interface MockTransportOpts extends BaseTransportOpts {\n  transport: 'mock'\n  handler: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Build a ChatClient bound to a specific transport. The returned client\n * is safe to share across analysts in a single registry run.\n */\nexport function createChatClient(opts: CreateChatClientOpts): ChatClient {\n  switch (opts.transport) {\n    case 'router':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'https://router.tangle.tools/v1',\n          apiKey: opts.apiKey,\n        } as LlmClientOptions),\n      )\n    case 'cli-bridge':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'http://127.0.0.1:3344/v1',\n          apiKey: opts.bearer ?? '',\n        } as LlmClientOptions),\n      )\n    case 'direct-provider':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl,\n          apiKey: opts.apiKey,\n        } as LlmClientOptions),\n      )\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'mock':\n      return {\n        transport: 'mock',\n        defaultModel: opts.defaultModel,\n        chat: async (req, callOpts) => opts.handler(resolveModel(req, opts.defaultModel), callOpts),\n      }\n  }\n}\n\nfunction wrapLlmClient(\n  transport: ChatTransport,\n  defaultModel: string | undefined,\n  inner: LlmClient,\n): ChatClient {\n  return {\n    transport,\n    defaultModel,\n    chat: async (req, callOpts) => {\n      const resolved = resolveModel(req, defaultModel)\n      // LlmClient.call doesn't accept an external AbortSignal today (it\n      // owns its own AbortController for the per-attempt timeout). We\n      // race the response against the caller's signal so awaiting code\n      // unblocks on abort. The in-flight HTTP request still runs to its\n      // own timeoutMs — when LlmClient grows a signal parameter, wire\n      // it directly here and drop the race.\n      const call = inner.call({\n        model: resolved.model!,\n        messages: req.messages,\n        jsonMode: req.jsonMode,\n        jsonSchema: req.jsonSchema,\n        temperature: req.temperature,\n        maxTokens: req.maxTokens,\n        timeoutMs: req.timeoutMs,\n      })\n      if (!callOpts?.signal) return await call\n      return await Promise.race([call, abortAsRejection(callOpts.signal)])\n    },\n  }\n}\n\nfunction abortAsRejection(signal: AbortSignal): Promise<never> {\n  if (signal.aborted) return Promise.reject(toAbortError(signal))\n  return new Promise<never>((_, reject) => {\n    signal.addEventListener('abort', () => reject(toAbortError(signal)), { once: true })\n  })\n}\n\nfunction toAbortError(signal: AbortSignal): Error {\n  const reason = (signal as { reason?: unknown }).reason\n  if (reason instanceof Error) return reason\n  const e = new Error('ChatClient.chat: aborted')\n  e.name = 'AbortError'\n  return e\n}\n\nfunction resolveModel(req: ChatRequest, defaultModel: string | undefined): ChatRequest {\n  if (req.model) return req\n  if (!defaultModel) {\n    throw new Error(\n      'ChatClient.chat: no model on request and no defaultModel on the client. ' +\n        'Either pass req.model or bind defaultModel at createChatClient().',\n    )\n  }\n  return { ...req, model: defaultModel }\n}\n","/**\n * concurrency — small primitives the evolution loop needs.\n *\n * `Mutex` is a zero-dep async lock with FIFO fairness. The evolution loop\n * uses it to serialise checkout/build/commit sequences inside a single\n * pool slot, and to gate concurrent JSONL writers (see\n * `lockedJsonlReferenceReplayStore`).\n *\n * Deliberately minimal — no priority queue, no timeouts. If you need\n * those, swap to `async-mutex` at the call site.\n */\n\nexport class Mutex {\n  private locked = false\n  private readonly waiters: Array<() => void> = []\n\n  async acquire(): Promise<() => void> {\n    if (!this.locked) {\n      this.locked = true\n      return () => this.release()\n    }\n    return new Promise<() => void>((resolve) => {\n      this.waiters.push(() => {\n        resolve(() => this.release())\n      })\n    })\n  }\n\n  private release(): void {\n    const next = this.waiters.shift()\n    if (next) {\n      next()\n    } else {\n      this.locked = false\n    }\n  }\n\n  async runExclusive<T>(fn: () => Promise<T> | T): Promise<T> {\n    const release = await this.acquire()\n    try {\n      return await fn()\n    } finally {\n      release()\n    }\n  }\n\n  /** True iff someone holds the lock right now. Diagnostics only. */\n  get isLocked(): boolean {\n    return this.locked\n  }\n\n  /** Pending waiter count. Diagnostics only. */\n  get pending(): number {\n    return this.waiters.length\n  }\n}\n","/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Skill-usage analyst — a DETERMINISTIC `Analyst` over a Claude/Codex skill\n * library + its trace corpus. Unlike the trace-store kinds (failure-mode,\n * improvement, ...) this kind calls no LLM: it mines real usage and skill\n * structure and emits findings by rule.\n *\n * It exists because the naive \"Skill-tool invocation count\" lies low — it\n * misses orchestrated sub-dispatch (a leaf skill run BY /pursue or /governor\n * logs under the parent), slash-command entry, local-script bypass, and\n * on-disk artifacts. The 2026-05-30 skill audit found 39/53 skills at zero\n * direct invocations, yet only one was a genuine cut: the rest were\n * measurement-invisible or discovery-limited. This analyst encodes that\n * lesson as a multi-signal usage model so a cheap repeatable pass can keep\n * the library honest, and so the expensive audit workflow's verdicts can\n * GEPA-distill it toward agreement (see `gold/skill-verdicts.gold.jsonl`).\n *\n * Report-building (`buildSkillUsageReport`, an fs scan) is separated from\n * finding emission (`SkillUsageAnalyst.analyze`, pure) so the slow scan runs\n * once at the registry boundary and the rule logic stays unit-testable.\n */\n\nimport { type Dirent, existsSync, readdirSync, readFileSync, statSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { Analyst, AnalystContext, AnalystFinding, AnalystSeverity } from '../types'\nimport { computeFindingId } from '../types'\n\n// ── Input model ──────────────────────────────────────────────────────\n\nexport type SkillKind = 'public' | 'private'\n\n/** One skill's multi-signal usage + structure. All counts are deterministic. */\nexport interface SkillUsageRecord {\n  name: string\n  kind: SkillKind\n  /** Absolute path to the skill's SKILL.md. */\n  path: string\n  lines: number\n  /** `\"skill\":\"<name>\"` Skill-tool invocations across the trace corpus. */\n  directInvocations: number\n  /** `<command-name>/<name>` slash invocations across the trace corpus. */\n  slashInvocations: number\n  /** Sibling skills whose SKILL.md dispatches to this one (`/<name>`). Proxy\n   *  for orchestrated sub-dispatch the per-skill counter cannot see. */\n  inboundRefs: number\n  /** On-disk artifacts attributable to the skill (e.g. `.evolve/<name>/**`). */\n  artifactCount: number\n  /** Tangle-private reference count in the body (leak signal for public skills). */\n  tanglePrivateRefs: number\n  hasReferencesDir: boolean\n  hasEvalsDir: boolean\n  /** Body mentions `skill-runs.jsonl` (visible to /reflect + /governor). */\n  logsRuns: boolean\n  /** Description carries an explicit `Triggers:` clause / trigger phrases. */\n  hasTriggerPhrases: boolean\n}\n\nexport interface SkillUsageReport {\n  generatedFromTraces: number\n  records: SkillUsageRecord[]\n}\n\nexport interface SkillUsageScanConfig {\n  /** Dirs holding `*.jsonl` transcripts (Claude `~/.claude/projects`, Codex sessions). */\n  transcriptDirs: string[]\n  /** Skill roots to scan; each dir directly under `root` with a `SKILL.md` is a skill. */\n  skillRoots: { root: string; kind: SkillKind }[]\n  /** Roots scanned for `<root>/.evolve/<skill>` artifact dirs. */\n  artifactRoots?: string[]\n  /** Token-prefixed mappings: skill name → extra artifact subpaths under an artifactRoot\n   *  (e.g. reflect → `.evolve/reflections`). Catches non-eponymous artifact dirs. */\n  artifactAliases?: Record<string, string[]>\n  /** Cap files read per transcript dir (bounds a huge corpus); 0 = unbounded. */\n  maxTranscriptsPerDir?: number\n}\n\n// ── Deterministic thresholds ─────────────────────────────────────────\n\n/** Anthropic's authoring guidance keeps SKILL.md short; past this with no\n *  `references/` split the body burns context budget every session. */\nconst BLOAT_LINE_THRESHOLD = 300\n\nconst TANGLE_PRIVATE_RE =\n  /\\b(cli-bridge|tangletools|ops-board|drew-gtr-pro|@tangle-network\\/|~\\/company|tangle\\.tools|gtm-agent)\\b|\\bkimi\\b|\\btcloud\\b/gi\nconst TRIGGER_RE = /triggers?\\s*[:-]/i\n\n// ── Report builder (fs scan — slow, runs once at the registry boundary) ──\n\nfunction listSkillDirs(root: string): { name: string; path: string }[] {\n  if (!existsSync(root)) return []\n  const out: { name: string; path: string }[] = []\n  for (const entry of readdirSync(root, { withFileTypes: true })) {\n    if (!entry.isDirectory() && !entry.isSymbolicLink()) continue\n    const skillMd = join(root, entry.name, 'SKILL.md')\n    if (existsSync(skillMd)) out.push({ name: entry.name, path: skillMd })\n  }\n  return out\n}\n\nfunction walkJsonl(dir: string, cap: number): string[] {\n  if (!existsSync(dir)) return []\n  const files: string[] = []\n  const stack = [dir]\n  while (stack.length) {\n    const cur = stack.pop()!\n    let entries: Dirent[]\n    try {\n      entries = readdirSync(cur, { withFileTypes: true })\n    } catch {\n      continue\n    }\n    for (const e of entries) {\n      const full = join(cur, e.name)\n      if (e.isDirectory()) stack.push(full)\n      else if (e.name.endsWith('.jsonl')) {\n        files.push(full)\n        if (cap > 0 && files.length >= cap) return files\n      }\n    }\n  }\n  return files\n}\n\nfunction frontmatterDescription(body: string): string {\n  const fm = /^---\\n([\\s\\S]*?)\\n---/.exec(body)\n  const block = fm?.[1] ?? ''\n  const m = /description:\\s*(.+)/i.exec(block)\n  return m?.[1] ?? ''\n}\n\nfunction countArtifacts(roots: string[], name: string, aliases: string[]): number {\n  let n = 0\n  for (const root of roots) {\n    const candidates = [join(root, '.evolve', name), ...aliases.map((a) => join(root, a))]\n    for (const dir of candidates) {\n      if (!existsSync(dir)) continue\n      try {\n        if (statSync(dir).isDirectory()) n += readdirSync(dir).length\n        else n += 1\n      } catch {\n        /* unreadable — skip */\n      }\n    }\n  }\n  return n\n}\n\n/** Scan the corpus + skill roots into a {@link SkillUsageReport}. Deterministic. */\nexport function buildSkillUsageReport(config: SkillUsageScanConfig): SkillUsageReport {\n  const skills = config.skillRoots.flatMap(({ root, kind }) =>\n    listSkillDirs(root).map((s) => ({ ...s, kind })),\n  )\n  const names = skills.map((s) => s.name)\n\n  // One pass over the corpus accumulating direct + slash counts per skill.\n  const direct = new Map<string, number>(names.map((n) => [n, 0]))\n  const slash = new Map<string, number>(names.map((n) => [n, 0]))\n  const skillRe = /\"skill\"\\s*:\\s*\"([a-z0-9_:-]+)\"/g\n  const cmdRe = /<command-name>\\/?([a-z0-9_:-]+)<\\/command-name>/g\n  let transcripts = 0\n  for (const dir of config.transcriptDirs) {\n    for (const file of walkJsonl(dir, config.maxTranscriptsPerDir ?? 0)) {\n      transcripts += 1\n      let data: string\n      try {\n        data = readFileSync(file, 'utf8')\n      } catch {\n        continue\n      }\n      for (const m of data.matchAll(skillRe)) {\n        const g = m[1]\n        if (!g) continue\n        const n = g.split(':').pop() ?? g\n        const prev = direct.get(n)\n        if (prev !== undefined) direct.set(n, prev + 1)\n      }\n      for (const m of data.matchAll(cmdRe)) {\n        const g = m[1]\n        if (g === undefined) continue\n        const prev = slash.get(g)\n        if (prev !== undefined) slash.set(g, prev + 1)\n      }\n    }\n  }\n\n  // Read each skill body once; compute structure + inbound refs across siblings.\n  const bodies = new Map<string, string>()\n  for (const s of skills) {\n    try {\n      bodies.set(s.name, readFileSync(s.path, 'utf8'))\n    } catch {\n      bodies.set(s.name, '')\n    }\n  }\n  const inbound = new Map<string, number>(names.map((n) => [n, 0]))\n  for (const target of names) {\n    const ref = new RegExp(`/${target}\\\\b|\\\\[\\\\[${target}\\\\]\\\\]`)\n    for (const s of skills) {\n      if (s.name === target) continue\n      if (ref.test(bodies.get(s.name) ?? '')) inbound.set(target, inbound.get(target)! + 1)\n    }\n  }\n\n  const records: SkillUsageRecord[] = skills.map((s) => {\n    const body = bodies.get(s.name) ?? ''\n    const dir = s.path.replace(/\\/SKILL\\.md$/, '')\n    return {\n      name: s.name,\n      kind: s.kind,\n      path: s.path,\n      lines: body ? body.split('\\n').length : 0,\n      directInvocations: direct.get(s.name) ?? 0,\n      slashInvocations: slash.get(s.name) ?? 0,\n      inboundRefs: inbound.get(s.name) ?? 0,\n      artifactCount: countArtifacts(\n        config.artifactRoots ?? [],\n        s.name,\n        config.artifactAliases?.[s.name] ?? [],\n      ),\n      tanglePrivateRefs: (body.match(TANGLE_PRIVATE_RE) ?? []).length,\n      hasReferencesDir: existsSync(join(dir, 'references')),\n      hasEvalsDir: existsSync(join(dir, 'evals')),\n      logsRuns: body.includes('skill-runs.jsonl'),\n      hasTriggerPhrases: TRIGGER_RE.test(frontmatterDescription(body) || body.slice(0, 600)),\n    }\n  })\n  return { generatedFromTraces: transcripts, records }\n}\n\n// ── Finding emission (pure — unit-testable, no LLM, no fs) ────────────\n\nconst ANALYST_ID = 'skill-usage'\n\nfunction finding(\n  area: string,\n  subject: string,\n  claim: string,\n  severity: AnalystSeverity,\n  confidence: number,\n  producedAt: string,\n  recommended: string,\n  evidenceUri: string,\n  rationale?: string,\n): AnalystFinding {\n  return {\n    schema_version: '1.0.0',\n    finding_id: computeFindingId({ analyst_id: ANALYST_ID, area, subject, claim }),\n    analyst_id: ANALYST_ID,\n    produced_at: producedAt,\n    severity,\n    area,\n    claim,\n    rationale,\n    evidence_refs: [{ kind: 'artifact', uri: evidenceUri }],\n    recommended_action: recommended,\n    confidence,\n    subject,\n  }\n}\n\n/** Pure rule pass over a report → findings. Exported for direct/unit use. */\nexport function emitSkillUsageFindings(\n  report: SkillUsageReport,\n  producedAt: string,\n): AnalystFinding[] {\n  const out: AnalystFinding[] = []\n  for (const r of report.records) {\n    const directTotal = r.directInvocations + r.slashInvocations\n    const trueUsage = directTotal + r.inboundRefs + r.artifactCount\n\n    // 1. Dead: no usage signal of ANY kind. The only real deprecation candidate.\n    if (trueUsage === 0) {\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' has zero usage across all signals (direct, slash, inbound-refs, artifacts)`,\n          'high',\n          0.6,\n          producedAt,\n          'Confirm the skill covers a real recurring job; if not, deprecate. Zero true usage is the only deterministic deprecation candidate.',\n          r.path,\n          'No Skill-tool call, no slash invocation, no sibling dispatches to it, and no on-disk artifacts.',\n        ),\n      )\n    } else if (directTotal === 0 && r.inboundRefs + r.artifactCount > 0) {\n      // 2. Measurement-invisible: real use via orchestration/artifacts, never invoked directly.\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' shows 0 direct invocations but is used via orchestration/artifacts (inbound=${r.inboundRefs}, artifacts=${r.artifactCount})`,\n          'info',\n          0.8,\n          producedAt,\n          'Do NOT treat as unused — usage is real but logged under parent skills or on disk. Strengthen direct-invocation discovery only if direct use is desired.',\n          r.path,\n          'The Skill-tool counter undercounts orchestrated/chained leaf skills.',\n        ),\n      )\n    }\n\n    // 3. Discovery gap: low direct use AND weak trigger surface.\n    if (directTotal <= 2 && !r.hasTriggerPhrases) {\n      out.push(\n        finding(\n          'discoverability',\n          r.name,\n          `Skill '${r.name}' is rarely invoked directly and its description has no explicit trigger phrases`,\n          'medium',\n          0.7,\n          producedAt,\n          'Add a `Triggers:` clause with verbatim user phrases to the frontmatter description so the model auto-invokes it.',\n          r.path,\n        ),\n      )\n    }\n\n    // 4. Public-repo leak.\n    if (r.kind === 'public' && r.tanglePrivateRefs > 0) {\n      out.push(\n        finding(\n          'safety',\n          r.name,\n          `Public skill '${r.name}' carries ${r.tanglePrivateRefs} Tangle-private reference(s)`,\n          'high',\n          0.75,\n          producedAt,\n          'Sanitize incidental internal refs (cli-bridge/kimi/tcloud/~company/private repos) or relocate to a private repo. Verify @tangle-network/* refs are to PUBLISHED packages before treating as a leak.',\n          r.path,\n        ),\n      )\n    }\n\n    // 5. Bloat / no progressive disclosure.\n    if (r.lines > BLOAT_LINE_THRESHOLD && !r.hasReferencesDir) {\n      out.push(\n        finding(\n          'maintainability',\n          r.name,\n          `Skill '${r.name}' is ${r.lines} lines with no references/ split (progressive disclosure)`,\n          'medium',\n          0.8,\n          producedAt,\n          `Split detail into references/ loaded on demand; keep SKILL.md a short overview. ${r.lines} lines load into every session's context budget.`,\n          r.path,\n        ),\n      )\n    }\n\n    // 6. No evals (Anthropic's \">=3 evals before docs\" rule).\n    if (!r.hasEvalsDir) {\n      out.push(\n        finding(\n          'data-quality',\n          r.name,\n          `Skill '${r.name}' ships no evals/`,\n          'low',\n          0.6,\n          producedAt,\n          'Add evals/evals.json with >=3 scenarios proving the skill beats baseline; gives regression coverage.',\n          r.path,\n        ),\n      )\n    }\n\n    // 7. No run logging → invisible to /reflect and /governor.\n    if (!r.logsRuns) {\n      out.push(\n        finding(\n          'observability',\n          r.name,\n          `Skill '${r.name}' never appends to .evolve/skill-runs.jsonl`,\n          'low',\n          0.55,\n          producedAt,\n          'Append one run line to .evolve/skill-runs.jsonl on completion, or declare it a non-logging leaf, so the self-improvement loop can see it ran.',\n          r.path,\n        ),\n      )\n    }\n  }\n  return out\n}\n\n// ── The Analyst ──────────────────────────────────────────────────────\n\nexport class SkillUsageAnalyst implements Analyst<SkillUsageReport> {\n  readonly id = ANALYST_ID\n  readonly description =\n    'Deterministic multi-signal skill-usage analysis: flags dead skills, measurement-invisible (orchestrated) usage, discovery gaps, public-repo leaks, bloat, missing evals, and missing run-logging.'\n  readonly inputKind = 'custom' as const\n  readonly cost = { kind: 'deterministic' as const, est_usd_per_run: 0 }\n  readonly version = '1.0.0'\n\n  async analyze(input: SkillUsageReport, ctx: AnalystContext): Promise<AnalystFinding[]> {\n    const producedAt = ctx.tags?.producedAt ?? new Date().toISOString()\n    ctx.log?.(\n      `skill-usage: ${input.records.length} skills over ${input.generatedFromTraces} transcripts`,\n    )\n    return emitSkillUsageFindings(input, producedAt)\n  }\n}\n\nexport const SKILL_USAGE_ANALYST = new SkillUsageAnalyst()\n","export interface RunScore {\n  success: number\n  goalProgress: number\n  repoGroundedness: number\n  driftPenalty: number\n  toolUseQuality: number\n  patchQuality: number\n  testReality: number\n  finalGate: number\n  reviewerBlockers: number\n  costUsd: number\n  wallSeconds: number\n  notes?: string[]\n}\n\nexport interface RunScoreWeights {\n  success: number\n  goalProgress: number\n  repoGroundedness: number\n  driftPenalty: number\n  toolUseQuality: number\n  patchQuality: number\n  testReality: number\n  finalGate: number\n  reviewerBlockers: number\n  costUsd: number\n  wallSeconds: number\n}\n\nexport const DEFAULT_RUN_SCORE_WEIGHTS: RunScoreWeights = {\n  success: 4,\n  goalProgress: 2,\n  repoGroundedness: 1.5,\n  driftPenalty: -1.5,\n  toolUseQuality: 1,\n  patchQuality: 1.25,\n  testReality: 1.5,\n  finalGate: 3,\n  reviewerBlockers: -2,\n  costUsd: -0.2,\n  wallSeconds: -0.1,\n}\n\nexport function aggregateRunScore(score: RunScore, weights: Partial<RunScoreWeights> = {}): number {\n  const w = { ...DEFAULT_RUN_SCORE_WEIGHTS, ...weights }\n  return (\n    w.success * clamp01(score.success) +\n    w.goalProgress * clamp01(score.goalProgress) +\n    w.repoGroundedness * clamp01(score.repoGroundedness) +\n    w.driftPenalty * clamp01(score.driftPenalty) +\n    w.toolUseQuality * clamp01(score.toolUseQuality) +\n    w.patchQuality * clamp01(score.patchQuality) +\n    w.testReality * clamp01(score.testReality) +\n    w.finalGate * clamp01(score.finalGate) +\n    w.reviewerBlockers * clamp01(score.reviewerBlockers) +\n    w.costUsd * Math.max(0, finiteOrZero(score.costUsd)) +\n    w.wallSeconds * Math.max(0, finiteOrZero(score.wallSeconds) / 60)\n  )\n}\n\nexport function clamp01(value: number): number {\n  if (!Number.isFinite(value)) return 0\n  return Math.max(0, Math.min(1, value))\n}\n\nfunction finiteOrZero(value: number): number {\n  return Number.isFinite(value) ? value : 0\n}\n","import { NotFoundError } from './errors'\nimport { aggregateRunScore, clamp01, type RunScore, type RunScoreWeights } from './run-score'\nimport type { Artifact, BudgetLedgerEntry, Run, Span, TraceEvent, TraceStore } from './trace'\n\nexport interface RunTrace {\n  run: Run\n  spans: Span[]\n  events: TraceEvent[]\n  artifacts: Artifact[]\n  budget: BudgetLedgerEntry[]\n}\n\nexport interface RunCriticOptions {\n  weights?: Partial<RunScoreWeights>\n  driftPatterns?: RegExp[]\n}\n\nconst DEFAULT_DRIFT_PATTERNS = [\n  /https?:\\/\\//i,\n  /\\btitle:\\s/i,\n  /\\bsummary:\\s/i,\n  /\\burl:\\s/i,\n  /\\bnpm package usage\\b/i,\n  /\\bnews\\b/i,\n]\n\nexport class RunCritic {\n  private readonly weights?: Partial<RunScoreWeights>\n  private readonly driftPatterns: RegExp[]\n\n  constructor(options: RunCriticOptions = {}) {\n    this.weights = options.weights\n    this.driftPatterns = options.driftPatterns ?? DEFAULT_DRIFT_PATTERNS\n  }\n\n  async score(store: TraceStore, runId: string): Promise<RunScore> {\n    const run = await store.getRun(runId)\n    if (!run) throw new NotFoundError(`run ${runId} not found`)\n    const [spans, events, artifacts, budget] = await Promise.all([\n      store.spans({ runId }),\n      store.events({ runId }),\n      store.artifacts(runId),\n      store.budget(runId),\n    ])\n    return this.scoreTrace({ run, spans, events, artifacts, budget })\n  }\n\n  scoreTrace(trace: RunTrace): RunScore {\n    const notes: string[] = []\n    const llmSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'llm' }> => s.kind === 'llm',\n    )\n    const toolSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'tool' }> => s.kind === 'tool',\n    )\n    const judgeSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'judge' }> => s.kind === 'judge',\n    )\n    const sandboxSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'sandbox' }> => s.kind === 'sandbox',\n    )\n    const finalGateSpans = judgeSpans.filter(\n      (span) => span.dimension === 'final_gate' || span.attributes?.finalGate === true,\n    )\n\n    const success =\n      trace.run.outcome?.pass === true ? 1 : trace.run.status === 'completed' ? 0.5 : 0\n    if (!success) notes.push('run did not complete with pass=true')\n\n    const judgeAverage = judgeSpans.length\n      ? judgeSpans.reduce((sum, span) => sum + normalizeJudgeScore(span.score), 0) /\n        judgeSpans.length\n      : undefined\n    const outcomeScore =\n      typeof trace.run.outcome?.score === 'number'\n        ? clamp01(\n            trace.run.outcome.score > 1 ? trace.run.outcome.score / 100 : trace.run.outcome.score,\n          )\n        : undefined\n    const goalProgress = outcomeScore ?? judgeAverage ?? success\n\n    const successfulTools = toolSpans.filter((span) => span.status !== 'error').length\n    const toolUseQuality = toolSpans.length === 0 ? 0 : successfulTools / toolSpans.length\n    if (toolSpans.length === 0) notes.push('no tool spans recorded')\n\n    const patchEvidence =\n      trace.artifacts.length +\n      toolSpans.filter((span) => /write|edit|patch|apply/i.test(span.toolName)).length\n    const patchQuality = patchEvidence > 0 ? clamp01(patchEvidence / 4) : 0\n    if (!patchQuality) notes.push('no artifact or edit evidence recorded')\n\n    const sandboxTests = sandboxSpans.filter(\n      (span) => typeof span.testsTotal === 'number' && span.testsTotal > 0,\n    )\n    const testReality = sandboxTests.length\n      ? sandboxTests.reduce(\n          (sum, span) => sum + (span.testsPassed ?? 0) / Math.max(1, span.testsTotal ?? 1),\n          0,\n        ) / sandboxTests.length\n      : toolSpans.some((span) =>\n            /\\btest|vitest|pytest|jest|build|tsc\\b/i.test(JSON.stringify(span.args)),\n          )\n        ? 0.4\n        : 0\n    if (!testReality) notes.push('no real test/build evidence recorded')\n\n    const blockerSpans = judgeSpans.filter((span) => isBlockingJudge(span))\n    const finalGateBlockers = finalGateSpans.filter((span) => isBlockingJudge(span))\n    const finalGate = finalGateSpans.length ? (finalGateBlockers.length ? 0 : 1) : success\n    if (finalGateBlockers.length)\n      notes.push(`final gate blocked by ${finalGateBlockers.length} reviewer(s)`)\n    else if (!finalGateSpans.length) notes.push('no final gate judgment recorded')\n\n    const reviewerBlockers = judgeSpans.length ? blockerSpans.length / judgeSpans.length : 0\n    if (reviewerBlockers) notes.push(`detected ${blockerSpans.length} blocking reviewer signal(s)`)\n\n    const positiveGroundingSignals =\n      patchEvidence +\n      sandboxSpans.length +\n      llmSpans.filter((span) => looksRepoGrounded(span.output ?? '')).length\n    const driftSignals =\n      llmSpans.filter((span) => this.isDrift(span.output ?? '')).length +\n      trace.events.filter((event) => this.isDrift(JSON.stringify(event.payload))).length\n    const repoGroundedness =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : positiveGroundingSignals / (positiveGroundingSignals + driftSignals)\n    const driftPenalty =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : driftSignals / (positiveGroundingSignals + driftSignals)\n    if (driftSignals > 0) notes.push(`detected ${driftSignals} drift signal(s)`)\n\n    const costUsd = trace.budget.length\n      ? Math.max(\n          ...trace.budget\n            .filter((entry: BudgetLedgerEntry) => entry.dimension === 'usd')\n            .map((entry: BudgetLedgerEntry) => entry.consumed),\n          0,\n        )\n      : llmSpans.reduce((sum, span) => sum + (span.costUsd ?? 0), 0)\n    const wallSeconds =\n      trace.run.endedAt && trace.run.startedAt\n        ? Math.max(0, (trace.run.endedAt - trace.run.startedAt) / 1000)\n        : 0\n\n    return {\n      success,\n      goalProgress,\n      repoGroundedness,\n      driftPenalty,\n      toolUseQuality,\n      patchQuality,\n      testReality,\n      finalGate,\n      reviewerBlockers,\n      costUsd,\n      wallSeconds,\n      notes,\n    }\n  }\n\n  rank(score: RunScore): number {\n    return aggregateRunScore(score, this.weights)\n  }\n\n  private isDrift(text: string): boolean {\n    return this.driftPatterns.some((pattern) => pattern.test(text))\n  }\n}\n\nfunction normalizeJudgeScore(score: number): number {\n  return score > 1 ? clamp01(score / 10) : clamp01(score)\n}\n\nfunction looksRepoGrounded(text: string): boolean {\n  return /(?:src\\/|tests?\\/|package\\.json|tsconfig|\\.ts\\b|\\.tsx\\b|git status|pnpm |npm |vitest|pytest|jest)/i.test(\n    text,\n  )\n}\n\nfunction isBlockingJudge(span: Extract<Span, { kind: 'judge' }>): boolean {\n  return (\n    span.attributes?.blocking === true ||\n    span.attributes?.verdict === 'BLOCKING' ||\n    positiveNumber(span.attributes?.blockingFindings) ||\n    positiveNumber(span.attributes?.highFindings) ||\n    span.score <= 2\n  )\n}\n\nfunction positiveNumber(value: unknown): boolean {\n  return typeof value === 'number' && value > 0\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport { callLlmJson, type LlmClientOptions } from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** LlmClient config (baseUrl, apiKey, authHeader, …). */\n  llm?: LlmClientOptions\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: Required<SemanticConceptJudgeOptions>,\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions = {},\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts: Required<SemanticConceptJudgeOptions> = {\n    model: options.model ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    llm: options.llm ?? {},\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  try {\n    const { value, result } = await callLlmJson<{\n      summary: string\n      concepts: ConceptFinding[]\n    }>(\n      {\n        model: opts.model,\n        messages: [\n          {\n            role: 'system',\n            content:\n              'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n          },\n          { role: 'user', content: buildPrompt(input, opts) },\n        ],\n        jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n        temperature: 0,\n        timeoutMs: opts.timeoutMs,\n      },\n      opts.llm,\n    )\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: result.costUsd ?? null,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions = {},\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;AACA,SAAS,UAAU;AA2BZ,SAAS,gBAAgB,QAA4C;AAC1E,SAAO,GAAG;AAAA,IACR,MAAM,OAAO,YAAY;AAAA,IACzB,QAAQ,OAAO;AAAA,IACf,QAAQ,OAAO;AAAA,IACf,QAAQ,EAAE,OAAO,OAAO,MAAM;AAAA,EAChC,CAAC;AACH;;;ACqFO,SAAS,iBAAiB,MAAwC;AACvE,UAAQ,KAAK,WAAW;AAAA,IACtB,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK;AAAA,QACf,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK,UAAU;AAAA,QACzB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK;AAAA,UACd,QAAQ,KAAK;AAAA,QACf,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MACzF;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,MAAM,OAAO,KAAK,aAAa,KAAK,QAAQ,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MAC5F;AAAA,EACJ;AACF;AAEA,SAAS,cACP,WACA,cACA,OACY;AACZ,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,MAAM,OAAO,KAAK,aAAa;AAC7B,YAAM,WAAW,aAAa,KAAK,YAAY;AAO/C,YAAM,OAAO,MAAM,KAAK;AAAA,QACtB,OAAO,SAAS;AAAA,QAChB,UAAU,IAAI;AAAA,QACd,UAAU,IAAI;AAAA,QACd,YAAY,IAAI;AAAA,QAChB,aAAa,IAAI;AAAA,QACjB,WAAW,IAAI;AAAA,QACf,WAAW,IAAI;AAAA,MACjB,CAAC;AACD,UAAI,CAAC,UAAU,OAAQ,QAAO,MAAM;AACpC,aAAO,MAAM,QAAQ,KAAK,CAAC,MAAM,iBAAiB,SAAS,MAAM,CAAC,CAAC;AAAA,IACrE;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,QAAqC;AAC7D,MAAI,OAAO,QAAS,QAAO,QAAQ,OAAO,aAAa,MAAM,CAAC;AAC9D,SAAO,IAAI,QAAe,CAAC,GAAG,WAAW;AACvC,WAAO,iBAAiB,SAAS,MAAM,OAAO,aAAa,MAAM,CAAC,GAAG,EAAE,MAAM,KAAK,CAAC;AAAA,EACrF,CAAC;AACH;AAEA,SAAS,aAAa,QAA4B;AAChD,QAAM,SAAU,OAAgC;AAChD,MAAI,kBAAkB,MAAO,QAAO;AACpC,QAAM,IAAI,IAAI,MAAM,0BAA0B;AAC9C,IAAE,OAAO;AACT,SAAO;AACT;AAEA,SAAS,aAAa,KAAkB,cAA+C;AACrF,MAAI,IAAI,MAAO,QAAO;AACtB,MAAI,CAAC,cAAc;AACjB,UAAM,IAAI;AAAA,MACR;AAAA,IAEF;AAAA,EACF;AACA,SAAO,EAAE,GAAG,KAAK,OAAO,aAAa;AACvC;;;AC/MO,IAAM,QAAN,MAAY;AAAA,EACT,SAAS;AAAA,EACA,UAA6B,CAAC;AAAA,EAE/C,MAAM,UAA+B;AACnC,QAAI,CAAC,KAAK,QAAQ;AAChB,WAAK,SAAS;AACd,aAAO,MAAM,KAAK,QAAQ;AAAA,IAC5B;AACA,WAAO,IAAI,QAAoB,CAAC,YAAY;AAC1C,WAAK,QAAQ,KAAK,MAAM;AACtB,gBAAQ,MAAM,KAAK,QAAQ,CAAC;AAAA,MAC9B,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AAAA,EAEQ,UAAgB;AACtB,UAAM,OAAO,KAAK,QAAQ,MAAM;AAChC,QAAI,MAAM;AACR,WAAK;AAAA,IACP,OAAO;AACL,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEA,MAAM,aAAgB,IAAsC;AAC1D,UAAM,UAAU,MAAM,KAAK,QAAQ;AACnC,QAAI;AACF,aAAO,MAAM,GAAG;AAAA,IAClB,UAAE;AACA,cAAQ;AAAA,IACV;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,UAAkB;AACpB,WAAO,KAAK,QAAQ;AAAA,EACtB;AACF;;;AC5CA,SAAS,gBAAgB,YAAY,iBAAiB;AACtD,SAAS,eAAe;AAGxB,IAAM,UAAU,oBAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;AACrC,MAAI,IAAI,QAAQ,IAAI,IAAI;AACxB,MAAI,CAAC,GAAG;AACN,QAAI,IAAI,MAAM;AACd,YAAQ,IAAI,MAAM,CAAC;AAAA,EACrB;AACA,SAAO;AACT;AAEO,IAAM,sBAAN,MAA0B;AAAA,EAE/B,YAA4B,MAAc;AAAd;AAC1B,SAAK,QAAQ,SAAS,IAAI;AAC1B,QAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAAG;AAC9B,gBAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAAA,IAC9C;AAAA,EACF;AAAA,EAL4B;AAAA,EADX;AAAA,EAQjB,MAAM,OAAO,OAA+B;AAC1C,UAAM,OAAO,GAAG,KAAK,UAAU,KAAK,CAAC;AAAA;AACrC,UAAM,KAAK,MAAM,aAAa,MAAM;AAClC,qBAAe,KAAK,MAAM,IAAI;AAAA,IAChC,CAAC;AAAA,EACH;AACF;AAGO,SAAS,iCAAuC;AACrD,UAAQ,MAAM;AAChB;;;AC1BA,SAAS,cAAAA,aAAY,oBAAoB;AAclC,IAAM,gBAAN,MAAoB;AAAA,EAGzB,YAA4B,MAAc;AAAd;AAC1B,SAAK,WAAW,IAAI,oBAAoB,IAAI;AAAA,EAC9C;AAAA,EAF4B;AAAA,EAFX;AAAA,EAMjB,MAAM,OAAO,OAAe,UAA2C;AACrE,eAAW,KAAK,UAAU;AACxB,YAAM,MAAwB,EAAE,GAAG,GAAG,QAAQ,MAAM;AACpD,YAAM,KAAK,SAAS,OAAO,GAAG;AAAA,IAChC;AAAA,EACF;AAAA;AAAA,EAGA,UAA8B;AAC5B,QAAI,CAACC,YAAW,KAAK,IAAI,EAAG,QAAO,CAAC;AACpC,UAAM,MAAM,aAAa,KAAK,MAAM,MAAM;AAC1C,QAAI,CAAC,IAAK,QAAO,CAAC;AAClB,UAAM,MAA0B,CAAC;AACjC,eAAW,QAAQ,IAAI,MAAM,IAAI,GAAG;AAClC,UAAI,CAAC,KAAM;AACX,UAAI;AACF,YAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;AAAA,MAC/C,QAAQ;AAAA,MAGR;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,QAAQ,OAAmC;AACzC,WAAO,KAAK,QAAQ,EAAE,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK;AAAA,EACxD;AACF;AAiCO,SAAS,kBAAkB,GAAmB,GAA4B;AAC/E,MAAI,EAAE,aAAa,EAAE,SAAU,QAAO;AACtC,MAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,QAAO;AACvE,MAAI,EAAE,cAAc,WAAW,EAAE,cAAc,OAAQ,QAAO;AAC9D,SAAO;AACT;AAMO,SAAS,aACd,UACA,SACA,SAAqB,CAAC,GACR;AACd,QAAM,aAAa,OAAO,cAAc;AACxC,QAAM,WAAW,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;AAC/D,QAAM,UAAU,IAAI,IAAI,QAAQ,IAAI,CAAC,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;AAE7D,QAAM,WAA+B,CAAC;AACtC,QAAM,cAAkC,CAAC;AACzC,QAAM,YAAgC,CAAC;AACvC,QAAM,UAAmC,CAAC;AAE1C,aAAW,CAAC,IAAI,GAAG,KAAK,SAAS;AAC/B,UAAM,OAAO,SAAS,IAAI,EAAE;AAC5B,QAAI,CAAC,MAAM;AACT,eAAS,KAAK,GAAG;AACjB;AAAA,IACF;AACA,QAAI,WAAW,MAAM,GAAG,GAAG;AACzB,cAAQ,KAAK,EAAE,UAAU,MAAM,SAAS,IAAI,CAAC;AAAA,IAC/C,OAAO;AACL,gBAAU,KAAK,GAAG;AAAA,IACpB;AAAA,EACF;AACA,aAAW,CAAC,IAAI,IAAI,KAAK,UAAU;AACjC,QAAI,CAAC,QAAQ,IAAI,EAAE,EAAG,aAAY,KAAK,IAAI;AAAA,EAC7C;AACA,SAAO,EAAE,UAAU,aAAa,WAAW,QAAQ;AACrD;;;AC3HA,SAAsB,cAAAC,aAAY,aAAa,gBAAAC,eAAc,gBAAgB;AAC7E,SAAS,YAAY;AAyDrB,IAAM,uBAAuB;AAE7B,IAAM,oBACJ;AACF,IAAM,aAAa;AAInB,SAAS,cAAc,MAAgD;AACrE,MAAI,CAACC,YAAW,IAAI,EAAG,QAAO,CAAC;AAC/B,QAAM,MAAwC,CAAC;AAC/C,aAAW,SAAS,YAAY,MAAM,EAAE,eAAe,KAAK,CAAC,GAAG;AAC9D,QAAI,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,eAAe,EAAG;AACrD,UAAM,UAAU,KAAK,MAAM,MAAM,MAAM,UAAU;AACjD,QAAIA,YAAW,OAAO,EAAG,KAAI,KAAK,EAAE,MAAM,MAAM,MAAM,MAAM,QAAQ,CAAC;AAAA,EACvE;AACA,SAAO;AACT;AAEA,SAAS,UAAU,KAAa,KAAuB;AACrD,MAAI,CAACA,YAAW,GAAG,EAAG,QAAO,CAAC;AAC9B,QAAM,QAAkB,CAAC;AACzB,QAAM,QAAQ,CAAC,GAAG;AAClB,SAAO,MAAM,QAAQ;AACnB,UAAM,MAAM,MAAM,IAAI;AACtB,QAAI;AACJ,QAAI;AACF,gBAAU,YAAY,KAAK,EAAE,eAAe,KAAK,CAAC;AAAA,IACpD,QAAQ;AACN;AAAA,IACF;AACA,eAAW,KAAK,SAAS;AACvB,YAAM,OAAO,KAAK,KAAK,EAAE,IAAI;AAC7B,UAAI,EAAE,YAAY,EAAG,OAAM,KAAK,IAAI;AAAA,eAC3B,EAAE,KAAK,SAAS,QAAQ,GAAG;AAClC,cAAM,KAAK,IAAI;AACf,YAAI,MAAM,KAAK,MAAM,UAAU,IAAK,QAAO;AAAA,MAC7C;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,uBAAuB,MAAsB;AACpD,QAAM,KAAK,wBAAwB,KAAK,IAAI;AAC5C,QAAM,QAAQ,KAAK,CAAC,KAAK;AACzB,QAAM,IAAI,uBAAuB,KAAK,KAAK;AAC3C,SAAO,IAAI,CAAC,KAAK;AACnB;AAEA,SAAS,eAAe,OAAiB,MAAc,SAA2B;AAChF,MAAI,IAAI;AACR,aAAW,QAAQ,OAAO;AACxB,UAAM,aAAa,CAAC,KAAK,MAAM,WAAW,IAAI,GAAG,GAAG,QAAQ,IAAI,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC;AACrF,eAAW,OAAO,YAAY;AAC5B,UAAI,CAACA,YAAW,GAAG,EAAG;AACtB,UAAI;AACF,YAAI,SAAS,GAAG,EAAE,YAAY,EAAG,MAAK,YAAY,GAAG,EAAE;AAAA,YAClD,MAAK;AAAA,MACZ,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,sBAAsB,QAAgD;AACpF,QAAM,SAAS,OAAO,WAAW;AAAA,IAAQ,CAAC,EAAE,MAAM,KAAK,MACrD,cAAc,IAAI,EAAE,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,KAAK,EAAE;AAAA,EACjD;AACA,QAAM,QAAQ,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI;AAGtC,QAAM,SAAS,IAAI,IAAoB,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAC/D,QAAM,QAAQ,IAAI,IAAoB,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAC9D,QAAM,UAAU;AAChB,QAAM,QAAQ;AACd,MAAI,cAAc;AAClB,aAAW,OAAO,OAAO,gBAAgB;AACvC,eAAW,QAAQ,UAAU,KAAK,OAAO,wBAAwB,CAAC,GAAG;AACnE,qBAAe;AACf,UAAI;AACJ,UAAI;AACF,eAAOC,cAAa,MAAM,MAAM;AAAA,MAClC,QAAQ;AACN;AAAA,MACF;AACA,iBAAW,KAAK,KAAK,SAAS,OAAO,GAAG;AACtC,cAAM,IAAI,EAAE,CAAC;AACb,YAAI,CAAC,EAAG;AACR,cAAM,IAAI,EAAE,MAAM,GAAG,EAAE,IAAI,KAAK;AAChC,cAAM,OAAO,OAAO,IAAI,CAAC;AACzB,YAAI,SAAS,OAAW,QAAO,IAAI,GAAG,OAAO,CAAC;AAAA,MAChD;AACA,iBAAW,KAAK,KAAK,SAAS,KAAK,GAAG;AACpC,cAAM,IAAI,EAAE,CAAC;AACb,YAAI,MAAM,OAAW;AACrB,cAAM,OAAO,MAAM,IAAI,CAAC;AACxB,YAAI,SAAS,OAAW,OAAM,IAAI,GAAG,OAAO,CAAC;AAAA,MAC/C;AAAA,IACF;AAAA,EACF;AAGA,QAAM,SAAS,oBAAI,IAAoB;AACvC,aAAW,KAAK,QAAQ;AACtB,QAAI;AACF,aAAO,IAAI,EAAE,MAAMA,cAAa,EAAE,MAAM,MAAM,CAAC;AAAA,IACjD,QAAQ;AACN,aAAO,IAAI,EAAE,MAAM,EAAE;AAAA,IACvB;AAAA,EACF;AACA,QAAM,UAAU,IAAI,IAAoB,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAChE,aAAW,UAAU,OAAO;AAC1B,UAAM,MAAM,IAAI,OAAO,IAAI,MAAM,aAAa,MAAM,QAAQ;AAC5D,eAAW,KAAK,QAAQ;AACtB,UAAI,EAAE,SAAS,OAAQ;AACvB,UAAI,IAAI,KAAK,OAAO,IAAI,EAAE,IAAI,KAAK,EAAE,EAAG,SAAQ,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAK,CAAC;AAAA,IACtF;AAAA,EACF;AAEA,QAAM,UAA8B,OAAO,IAAI,CAAC,MAAM;AACpD,UAAM,OAAO,OAAO,IAAI,EAAE,IAAI,KAAK;AACnC,UAAM,MAAM,EAAE,KAAK,QAAQ,gBAAgB,EAAE;AAC7C,WAAO;AAAA,MACL,MAAM,EAAE;AAAA,MACR,MAAM,EAAE;AAAA,MACR,MAAM,EAAE;AAAA,MACR,OAAO,OAAO,KAAK,MAAM,IAAI,EAAE,SAAS;AAAA,MACxC,mBAAmB,OAAO,IAAI,EAAE,IAAI,KAAK;AAAA,MACzC,kBAAkB,MAAM,IAAI,EAAE,IAAI,KAAK;AAAA,MACvC,aAAa,QAAQ,IAAI,EAAE,IAAI,KAAK;AAAA,MACpC,eAAe;AAAA,QACb,OAAO,iBAAiB,CAAC;AAAA,QACzB,EAAE;AAAA,QACF,OAAO,kBAAkB,EAAE,IAAI,KAAK,CAAC;AAAA,MACvC;AAAA,MACA,oBAAoB,KAAK,MAAM,iBAAiB,KAAK,CAAC,GAAG;AAAA,MACzD,kBAAkBD,YAAW,KAAK,KAAK,YAAY,CAAC;AAAA,MACpD,aAAaA,YAAW,KAAK,KAAK,OAAO,CAAC;AAAA,MAC1C,UAAU,KAAK,SAAS,kBAAkB;AAAA,MAC1C,mBAAmB,WAAW,KAAK,uBAAuB,IAAI,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,IACvF;AAAA,EACF,CAAC;AACD,SAAO,EAAE,qBAAqB,aAAa,QAAQ;AACrD;AAIA,IAAM,aAAa;AAEnB,SAAS,QACP,MACA,SACA,OACA,UACA,YACA,YACA,aACA,aACA,WACgB;AAChB,SAAO;AAAA,IACL,gBAAgB;AAAA,IAChB,YAAY,iBAAiB,EAAE,YAAY,YAAY,MAAM,SAAS,MAAM,CAAC;AAAA,IAC7E,YAAY;AAAA,IACZ,aAAa;AAAA,IACb;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,YAAY,CAAC;AAAA,IACtD,oBAAoB;AAAA,IACpB;AAAA,IACA;AAAA,EACF;AACF;AAGO,SAAS,uBACd,QACA,YACkB;AAClB,QAAM,MAAwB,CAAC;AAC/B,aAAW,KAAK,OAAO,SAAS;AAC9B,UAAM,cAAc,EAAE,oBAAoB,EAAE;AAC5C,UAAM,YAAY,cAAc,EAAE,cAAc,EAAE;AAGlD,QAAI,cAAc,GAAG;AACnB,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF,WAAW,gBAAgB,KAAK,EAAE,cAAc,EAAE,gBAAgB,GAAG;AAEnE,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI,iFAAiF,EAAE,WAAW,eAAe,EAAE,aAAa;AAAA,UAC5I;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAGA,QAAI,eAAe,KAAK,CAAC,EAAE,mBAAmB;AAC5C,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,EAAE,SAAS,YAAY,EAAE,oBAAoB,GAAG;AAClD,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,iBAAiB,EAAE,IAAI,aAAa,EAAE,iBAAiB;AAAA,UACvD;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,EAAE,QAAQ,wBAAwB,CAAC,EAAE,kBAAkB;AACzD,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI,QAAQ,EAAE,KAAK;AAAA,UAC/B;AAAA,UACA;AAAA,UACA;AAAA,UACA,mFAAmF,EAAE,KAAK;AAAA,UAC1F,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,CAAC,EAAE,aAAa;AAClB,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,CAAC,EAAE,UAAU;AACf,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAIO,IAAM,oBAAN,MAA6D;AAAA,EACzD,KAAK;AAAA,EACL,cACP;AAAA,EACO,YAAY;AAAA,EACZ,OAAO,EAAE,MAAM,iBAA0B,iBAAiB,EAAE;AAAA,EAC5D,UAAU;AAAA,EAEnB,MAAM,QAAQ,OAAyB,KAAgD;AACrF,UAAM,aAAa,IAAI,MAAM,eAAc,oBAAI,KAAK,GAAE,YAAY;AAClE,QAAI;AAAA,MACF,gBAAgB,MAAM,QAAQ,MAAM,gBAAgB,MAAM,mBAAmB;AAAA,IAC/E;AACA,WAAO,uBAAuB,OAAO,UAAU;AAAA,EACjD;AACF;AAEO,IAAM,sBAAsB,IAAI,kBAAkB;;;ACtXlD,IAAM,4BAA6C;AAAA,EACxD,SAAS;AAAA,EACT,cAAc;AAAA,EACd,kBAAkB;AAAA,EAClB,cAAc;AAAA,EACd,gBAAgB;AAAA,EAChB,cAAc;AAAA,EACd,aAAa;AAAA,EACb,WAAW;AAAA,EACX,kBAAkB;AAAA,EAClB,SAAS;AAAA,EACT,aAAa;AACf;AAEO,SAAS,kBAAkB,OAAiB,UAAoC,CAAC,GAAW;AACjG,QAAM,IAAI,EAAE,GAAG,2BAA2B,GAAG,QAAQ;AACrD,SACE,EAAE,UAAU,QAAQ,MAAM,OAAO,IACjC,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,mBAAmB,QAAQ,MAAM,gBAAgB,IACnD,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,iBAAiB,QAAQ,MAAM,cAAc,IAC/C,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,cAAc,QAAQ,MAAM,WAAW,IACzC,EAAE,YAAY,QAAQ,MAAM,SAAS,IACrC,EAAE,mBAAmB,QAAQ,MAAM,gBAAgB,IACnD,EAAE,UAAU,KAAK,IAAI,GAAG,aAAa,MAAM,OAAO,CAAC,IACnD,EAAE,cAAc,KAAK,IAAI,GAAG,aAAa,MAAM,WAAW,IAAI,EAAE;AAEpE;AAEO,SAAS,QAAQ,OAAuB;AAC7C,MAAI,CAAC,OAAO,SAAS,KAAK,EAAG,QAAO;AACpC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,CAAC;AACvC;AAEA,SAAS,aAAa,OAAuB;AAC3C,SAAO,OAAO,SAAS,KAAK,IAAI,QAAQ;AAC1C;;;AClDA,IAAM,yBAAyB;AAAA,EAC7B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,IAAM,YAAN,MAAgB;AAAA,EACJ;AAAA,EACA;AAAA,EAEjB,YAAY,UAA4B,CAAC,GAAG;AAC1C,SAAK,UAAU,QAAQ;AACvB,SAAK,gBAAgB,QAAQ,iBAAiB;AAAA,EAChD;AAAA,EAEA,MAAM,MAAM,OAAmB,OAAkC;AAC/D,UAAM,MAAM,MAAM,MAAM,OAAO,KAAK;AACpC,QAAI,CAAC,IAAK,OAAM,IAAI,cAAc,OAAO,KAAK,YAAY;AAC1D,UAAM,CAAC,OAAO,QAAQ,WAAW,MAAM,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC3D,MAAM,MAAM,EAAE,MAAM,CAAC;AAAA,MACrB,MAAM,OAAO,EAAE,MAAM,CAAC;AAAA,MACtB,MAAM,UAAU,KAAK;AAAA,MACrB,MAAM,OAAO,KAAK;AAAA,IACpB,CAAC;AACD,WAAO,KAAK,WAAW,EAAE,KAAK,OAAO,QAAQ,WAAW,OAAO,CAAC;AAAA,EAClE;AAAA,EAEA,WAAW,OAA2B;AACpC,UAAM,QAAkB,CAAC;AACzB,UAAM,WAAW,MAAM,MAAM;AAAA,MAC3B,CAAC,MAA2C,EAAE,SAAS;AAAA,IACzD;AACA,UAAM,YAAY,MAAM,MAAM;AAAA,MAC5B,CAAC,MAA4C,EAAE,SAAS;AAAA,IAC1D;AACA,UAAM,aAAa,MAAM,MAAM;AAAA,MAC7B,CAAC,MAA6C,EAAE,SAAS;AAAA,IAC3D;AACA,UAAM,eAAe,MAAM,MAAM;AAAA,MAC/B,CAAC,MAA+C,EAAE,SAAS;AAAA,IAC7D;AACA,UAAM,iBAAiB,WAAW;AAAA,MAChC,CAAC,SAAS,KAAK,cAAc,gBAAgB,KAAK,YAAY,cAAc;AAAA,IAC9E;AAEA,UAAM,UACJ,MAAM,IAAI,SAAS,SAAS,OAAO,IAAI,MAAM,IAAI,WAAW,cAAc,MAAM;AAClF,QAAI,CAAC,QAAS,OAAM,KAAK,qCAAqC;AAE9D,UAAM,eAAe,WAAW,SAC5B,WAAW,OAAO,CAAC,KAAK,SAAS,MAAM,oBAAoB,KAAK,KAAK,GAAG,CAAC,IACzE,WAAW,SACX;AACJ,UAAM,eACJ,OAAO,MAAM,IAAI,SAAS,UAAU,WAChC;AAAA,MACE,MAAM,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAI,QAAQ,QAAQ,MAAM,MAAM,IAAI,QAAQ;AAAA,IAClF,IACA;AACN,UAAM,eAAe,gBAAgB,gBAAgB;AAErD,UAAM,kBAAkB,UAAU,OAAO,CAAC,SAAS,KAAK,WAAW,OAAO,EAAE;AAC5E,UAAM,iBAAiB,UAAU,WAAW,IAAI,IAAI,kBAAkB,UAAU;AAChF,QAAI,UAAU,WAAW,EAAG,OAAM,KAAK,wBAAwB;AAE/D,UAAM,gBACJ,MAAM,UAAU,SAChB,UAAU,OAAO,CAAC,SAAS,0BAA0B,KAAK,KAAK,QAAQ,CAAC,EAAE;AAC5E,UAAM,eAAe,gBAAgB,IAAI,QAAQ,gBAAgB,CAAC,IAAI;AACtE,QAAI,CAAC,aAAc,OAAM,KAAK,uCAAuC;AAErE,UAAM,eAAe,aAAa;AAAA,MAChC,CAAC,SAAS,OAAO,KAAK,eAAe,YAAY,KAAK,aAAa;AAAA,IACrE;AACA,UAAM,cAAc,aAAa,SAC7B,aAAa;AAAA,MACX,CAAC,KAAK,SAAS,OAAO,KAAK,eAAe,KAAK,KAAK,IAAI,GAAG,KAAK,cAAc,CAAC;AAAA,MAC/E;AAAA,IACF,IAAI,aAAa,SACjB,UAAU;AAAA,MAAK,CAAC,SACZ,yCAAyC,KAAK,KAAK,UAAU,KAAK,IAAI,CAAC;AAAA,IACzE,IACA,MACA;AACN,QAAI,CAAC,YAAa,OAAM,KAAK,sCAAsC;AAEnE,UAAM,eAAe,WAAW,OAAO,CAAC,SAAS,gBAAgB,IAAI,CAAC;AACtE,UAAM,oBAAoB,eAAe,OAAO,CAAC,SAAS,gBAAgB,IAAI,CAAC;AAC/E,UAAM,YAAY,eAAe,SAAU,kBAAkB,SAAS,IAAI,IAAK;AAC/E,QAAI,kBAAkB;AACpB,YAAM,KAAK,yBAAyB,kBAAkB,MAAM,cAAc;AAAA,aACnE,CAAC,eAAe,OAAQ,OAAM,KAAK,iCAAiC;AAE7E,UAAM,mBAAmB,WAAW,SAAS,aAAa,SAAS,WAAW,SAAS;AACvF,QAAI,iBAAkB,OAAM,KAAK,YAAY,aAAa,MAAM,8BAA8B;AAE9F,UAAM,2BACJ,gBACA,aAAa,SACb,SAAS,OAAO,CAAC,SAAS,kBAAkB,KAAK,UAAU,EAAE,CAAC,EAAE;AAClE,UAAM,eACJ,SAAS,OAAO,CAAC,SAAS,KAAK,QAAQ,KAAK,UAAU,EAAE,CAAC,EAAE,SAC3D,MAAM,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,EAAE;AAC9E,UAAM,mBACJ,2BAA2B,iBAAiB,IACxC,IACA,4BAA4B,2BAA2B;AAC7D,UAAM,eACJ,2BAA2B,iBAAiB,IACxC,IACA,gBAAgB,2BAA2B;AACjD,QAAI,eAAe,EAAG,OAAM,KAAK,YAAY,YAAY,kBAAkB;AAE3E,UAAM,UAAU,MAAM,OAAO,SACzB,KAAK;AAAA,MACH,GAAG,MAAM,OACN,OAAO,CAAC,UAA6B,MAAM,cAAc,KAAK,EAC9D,IAAI,CAAC,UAA6B,MAAM,QAAQ;AAAA,MACnD;AAAA,IACF,IACA,SAAS,OAAO,CAAC,KAAK,SAAS,OAAO,KAAK,WAAW,IAAI,CAAC;AAC/D,UAAM,cACJ,MAAM,IAAI,WAAW,MAAM,IAAI,YAC3B,KAAK,IAAI,IAAI,MAAM,IAAI,UAAU,MAAM,IAAI,aAAa,GAAI,IAC5D;AAEN,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,KAAK,OAAyB;AAC5B,WAAO,kBAAkB,OAAO,KAAK,OAAO;AAAA,EAC9C;AAAA,EAEQ,QAAQ,MAAuB;AACrC,WAAO,KAAK,cAAc,KAAK,CAAC,YAAY,QAAQ,KAAK,IAAI,CAAC;AAAA,EAChE;AACF;AAEA,SAAS,oBAAoB,OAAuB;AAClD,SAAO,QAAQ,IAAI,QAAQ,QAAQ,EAAE,IAAI,QAAQ,KAAK;AACxD;AAEA,SAAS,kBAAkB,MAAuB;AAChD,SAAO,qGAAqG;AAAA,IAC1G;AAAA,EACF;AACF;AAEA,SAAS,gBAAgB,MAAiD;AACxE,SACE,KAAK,YAAY,aAAa,QAC9B,KAAK,YAAY,YAAY,cAC7B,eAAe,KAAK,YAAY,gBAAgB,KAChD,eAAe,KAAK,YAAY,YAAY,KAC5C,KAAK,SAAS;AAElB;AAEA,SAAS,eAAe,OAAyB;AAC/C,SAAO,OAAO,UAAU,YAAY,QAAQ;AAC9C;;;ACvFO,IAAM,6BAAgE;AAAA,EAC3E,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,SAAS;AACX;AA2BO,IAAM,iCAAiC;AAE9C,IAAM,qBAAqB;AAC3B,IAAM,mBAAmB;AACzB,IAAM,uBAAuB;AAC7B,IAAM,kBAAkB;AACxB,IAAM,gBAAgB;AAEtB,IAAM,kBAAkB;AAAA,EACtB,MAAM;AAAA,EACN,sBAAsB;AAAA,EACtB,UAAU,CAAC,WAAW,UAAU;AAAA,EAChC,YAAY;AAAA,IACV,SAAS,EAAE,MAAM,UAAU,WAAW,IAAI,WAAW,IAAI;AAAA,IACzD,UAAU;AAAA,MACR,MAAM;AAAA,MACN,UAAU;AAAA,MACV,OAAO;AAAA,QACL,MAAM;AAAA,QACN,sBAAsB;AAAA,QACtB,UAAU,CAAC,WAAW,WAAW,SAAS,YAAY,UAAU;AAAA,QAChE,YAAY;AAAA,UACV,SAAS,EAAE,MAAM,UAAU,WAAW,GAAG,WAAW,IAAI;AAAA,UACxD,SAAS,EAAE,MAAM,UAAU;AAAA,UAC3B,OAAO,EAAE,MAAM,UAAU,SAAS,GAAG,SAAS,GAAG;AAAA,UACjD,UAAU,EAAE,MAAM,UAAU,WAAW,GAAG,WAAW,IAAI;AAAA,UACzD,UAAU,EAAE,MAAM,UAAU,MAAM,CAAC,YAAY,SAAS,SAAS,MAAM,EAAE;AAAA,QAC3E;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;AAClE,MAAI,KAAK,UAAU,IAAK,QAAO;AAC/B,SAAO,GAAG,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,oBAAkB,KAAK,SAAS,GAAG,aAAa,KAAK;AACnF;AAEA,SAAS,YACP,OACA,MACQ;AACR,QAAM,aAAa,MAAM,YACtB,OAAO,CAAC,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,EACtD,IAAI,CAAC,MAAM,aAAa,EAAE,IAAI;AAAA,EAAS,EAAE,OAAO,EAAE,EAClD,KAAK,MAAM;AAEd,QAAM,OAAO,MAAM,cAAc;AAEjC,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUP,MAAM,WAAW;AAAA;AAAA,EAEjB,MAAM,gBAAgB;AAAA,UAA+B,MAAM,aAAa;AAAA,iBAAoB,MAAM,uBAAuB,EAAE;AAAA;AAAA,IAAS,EAAE;AAAA,EACtI,MAAM,iBACL;AAAA,IACC,CAAC,GAAG,MACF,KAAK,IAAI,CAAC,MAAM,EAAE,IAAI,IAAI,EAAE,UAAU,SAAS,mBAAc,EAAE,SAAS,MAAM,GAAG,CAAC,EAAE,KAAK,KAAK,CAAC,MAAM,EAAE;AAAA,EAC3G,EACC,KAAK,IAAI,CAAC;AAAA;AAAA,EAEX,OAAO;AAAA,EAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,CAAC;AAAA;AAAA,IAAS,EAAE;AAAA,EAChH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAkBrD;AASA,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACH;AACrC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,aAAa,MAAM,iBAAiB;AAE1C,MAAI,eAAe,GAAG;AACpB,WAAO;AAAA,MACL,MAAM;AAAA,MACN,SAAS;AAAA,MACT,OAAO;AAAA,MACP,cAAc;AAAA,MACd,YAAY;AAAA,MACZ,UAAU,CAAC;AAAA,MACX,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,SAAS;AAAA,MACT,WAAW;AAAA,MACX,OAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,OAA8C;AAAA,IAClD,OAAO,QAAQ,SAAS;AAAA,IACxB,WAAW,QAAQ,aAAa;AAAA,IAChC,gBAAgB,QAAQ,kBAAkB;AAAA,IAC1C,iBAAiB,QAAQ,mBAAmB;AAAA,IAC5C,cAAc,QAAQ,gBAAgB;AAAA,IACtC,KAAK,QAAQ,OAAO,CAAC;AAAA,IACrB,gBAAgB,QAAQ,kBAAkB;AAAA,IAC1C,mBAAmB,EAAE,GAAG,4BAA4B,GAAI,QAAQ,qBAAqB,CAAC,EAAG;AAAA,EAC3F;AAMA,QAAM,mBAAmB,CAAC,SAA8B;AACtD,QAAI,KAAK,mBAAmB,OAAQ,QAAO;AAC3C,QAAI,KAAK,UAAU,KAAM,QAAO,KAAK;AACrC,QAAI,KAAK,mBAAmB,cAAc;AACxC,aAAO,KAAK,kBAAkB,KAAK,cAAc,QAAQ,KAAK;AAAA,IAChE;AACA,WAAO;AAAA,EACT;AACA,QAAM,eAAe,IAAI;AAAA,IACvB,MAAM,iBAAiB,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC;AAAA,EACjE;AAEA,MAAI;AACF,UAAM,EAAE,OAAO,OAAO,IAAI,MAAM;AAAA,MAI9B;AAAA,QACE,OAAO,KAAK;AAAA,QACZ,UAAU;AAAA,UACR;AAAA,YACE,MAAM;AAAA,YACN,SACE;AAAA,UACJ;AAAA,UACA,EAAE,MAAM,QAAQ,SAAS,YAAY,OAAO,IAAI,EAAE;AAAA,QACpD;AAAA,QACA,YAAY,EAAE,MAAM,0BAA0B,QAAQ,gBAAgB;AAAA,QACtE,aAAa;AAAA,QACb,WAAW,KAAK;AAAA,MAClB;AAAA,MACA,KAAK;AAAA,IACP;AAEA,QAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GAAG;AACtD,YAAM,IAAI,MAAM,0EAAqE;AAAA,IACvF;AAEA,UAAM,WAA6B,MAAM,SAAS,IAAI,CAAC,OAAO;AAAA,MAC5D,SAAS,OAAO,EAAE,OAAO;AAAA,MACzB,SAAS,QAAQ,EAAE,OAAO;AAAA,MAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;AAAA,MACrD,UAAU,OAAO,EAAE,YAAY,EAAE;AAAA,MACjC,UAAW,CAAC,YAAY,SAAS,SAAS,MAAM,EAAY,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;AAAA,IACN,EAAE;AAEF,UAAM,eAAe,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,EAAE;AACvE,QAAI,YAAY;AAChB,QAAI,mBAAmB;AACvB,eAAW,KAAK,UAAU;AACxB,YAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;AACzC,mBAAa;AACb,0BAAoB,IAAI,EAAE;AAAA,IAC5B;AACA,UAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;AAE7E,WAAO;AAAA,MACL,MAAM;AAAA,MACN,SAAS;AAAA,MACT,OAAO,QAAQ,WAAW,IAAI,QAAQ,CAAC,CAAC;AAAA,MACxC;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,OAAO,MAAM,WAAW,EAAE;AAAA,MACnC,YAAY,KAAK,IAAI,IAAI;AAAA,MACzB,SAAS,OAAO,WAAW;AAAA,MAC3B,WAAW;AAAA,IACb;AAAA,EACF,SAAS,KAAK;AACZ,WAAO;AAAA,MACL,MAAM;AAAA,MACN,SAAS;AAAA,MACT,OAAO;AAAA,MACP,cAAc;AAAA,MACd;AAAA,MACA,UAAU,CAAC;AAAA,MACX,SAAS;AAAA,MACT,YAAY,KAAK,IAAI,IAAI;AAAA,MACzB,SAAS;AAAA,MACT,WAAW;AAAA,MACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD;AAAA,EACF;AACF;AAMO,SAAS,2BACd,UAAuC,CAAC,GACmC;AAC3E,SAAO,CAAC,UAAU,wBAAwB,OAAO,OAAO;AAC1D;","names":["existsSync","existsSync","existsSync","readFileSync","existsSync","readFileSync"]}