@tangle-network/agent-bench 0.1.0 → 0.3.6

package/src/index.ts

@@ -11,9 +11,82 @@
  * absent, so importing the registry is cheap; running a specific benchmark pulls only its deps.
  */
 export { ADAPTERS, resolveAdapter } from './adapters'
+export { createCragAdapter } from './benchmarks/crag'
+export { createNoMiraclAdapter } from './benchmarks/nomiracl'
+export { createOpenRagBenchAdapter } from './benchmarks/open-rag-bench'
+export { createRagBenchAdapter } from './benchmarks/ragbench'
+export {
+  FINAL_ANSWER_SENTINEL,
+  answerScoreToBenchScore,
+  contextBlock,
+  contextsFrom,
+  normalizeAnswer,
+  parseCitations,
+  parseFinalAnswer,
+  ragAnswerOutput,
+  scoreAnswerArtifact,
+  tokenF1,
+  type RagAnswerScore,
+  type RagContext,
+} from './benchmarks/rag-shared'
+export { createT2RagBenchAdapter } from './benchmarks/t2-ragbench'
+export {
+  createSweBenchAdapter,
+  scoreSweReport,
+  sweEvaluationArgv,
+  swePatchOutput,
+  type SweBenchAdapterOptions,
+  type SweBenchArtifactCaptureContext,
+  type SweBenchCacheLevel,
+} from './benchmarks/swe-bench'
+export {
+  runStagedJudge,
+  StagedJudgeError,
+  type StagedRunCaptureSpec,
+  type StagedRunSpec,
+} from './benchmarks/_harness'
 export type {
   BenchmarkAdapter,
   BenchScore,
   BenchTask,
+  JudgeArtifactFileReceipt,
+  JudgeArtifactReceipt,
   LoadOptions,
 } from './benchmarks/types'
+
+/**
+ * The unifier: run a subset of the registry over a matrix of agent cells (harness × model ×
+ * persona), each scored by the benchmark's own judge, and rank them.
+ *
+ *   import { runBenchmarks, printBenchmarksReport } from '@tangle-network/agent-bench'
+ */
+export {
+  runBenchmarks,
+  printBenchmarksReport,
+  type BenchCell,
+  type BenchShot,
+  type BenchCellTaskResult,
+  type BenchLeaderboardRow,
+  type RunBenchmarksOptions,
+  type RunBenchmarksReport,
+} from './run-benchmarks'
+export {
+  createPierCandidateRecoveryExecutor,
+  executePreparedPierCandidate,
+  type ExecutePreparedPierCandidateOptions,
+  type PierCandidateGraderPort,
+  type PierCandidateOfficialResult,
+  type PierCandidateTerminationAcknowledgement,
+  type PierCandidateTrialController,
+  type PierCandidateTrialHandle,
+  type PierCandidateTrialIdentity,
+  type PierCandidateTrialResult,
+  type StagedPierCandidateExecution,
+} from './pier-agent'
+export {
+  FilePierCandidateTrialController,
+  type FilePierCandidateTrialControllerOptions,
+  type PierCandidateProcessSpec,
+  type PierDockerConnection,
+} from './pier-trial-controller'
+export { createPierResultGrader } from './pier-result-grader'