agent-hitch 0.2.6 → 0.2.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +30 -3
- package/README.zh-CN.md +14 -2
- package/dist/bin/hitch.js +25 -3
- package/dist/bin/hitch.js.map +1 -1
- package/dist/scripts/canary-benchmark.js +54 -0
- package/dist/scripts/canary-benchmark.js.map +1 -0
- package/dist/scripts/canary-buildkit-secrets.js +166 -0
- package/dist/scripts/canary-buildkit-secrets.js.map +1 -0
- package/dist/scripts/canary-collision-domains.js +77 -0
- package/dist/scripts/canary-collision-domains.js.map +1 -0
- package/dist/scripts/canary-docker-image-observation.js +181 -0
- package/dist/scripts/canary-docker-image-observation.js.map +1 -0
- package/dist/scripts/canary-eval-scheduler-throughput.js +47 -0
- package/dist/scripts/canary-eval-scheduler-throughput.js.map +1 -0
- package/dist/scripts/canary-gpu-hardware.js +95 -0
- package/dist/scripts/canary-gpu-hardware.js.map +1 -0
- package/dist/scripts/canary-harbor-load.js +365 -0
- package/dist/scripts/canary-harbor-load.js.map +1 -0
- package/dist/scripts/canary-packaged-harness.js +75 -0
- package/dist/scripts/canary-packaged-harness.js.map +1 -0
- package/dist/scripts/canary-resource-load.js +188 -0
- package/dist/scripts/canary-resource-load.js.map +1 -0
- package/dist/scripts/check-architecture.js +10 -5
- package/dist/scripts/check-architecture.js.map +1 -1
- package/dist/scripts/docker-reaper-crash-child.js +21 -0
- package/dist/scripts/docker-reaper-crash-child.js.map +1 -0
- package/dist/src/adapters/catalog.js +3 -1
- package/dist/src/adapters/catalog.js.map +1 -1
- package/dist/src/adapters/discovery.js +1 -1
- package/dist/src/adapters/discovery.js.map +1 -1
- package/dist/src/adapters/providers/claude.js +6 -0
- package/dist/src/adapters/providers/claude.js.map +1 -1
- package/dist/src/adapters/providers/codex-auth.js +29 -0
- package/dist/src/adapters/providers/codex-auth.js.map +1 -0
- package/dist/src/adapters/providers/codex.js +29 -6
- package/dist/src/adapters/providers/codex.js.map +1 -1
- package/dist/src/adapters/providers/deepseek.js +6 -0
- package/dist/src/adapters/providers/deepseek.js.map +1 -1
- package/dist/src/adapters/providers/model-call.js +26 -0
- package/dist/src/adapters/providers/model-call.js.map +1 -0
- package/dist/src/adapters/providers/opencode.js +6 -0
- package/dist/src/adapters/providers/opencode.js.map +1 -1
- package/dist/src/adapters/providers/pi.js +6 -0
- package/dist/src/adapters/providers/pi.js.map +1 -1
- package/dist/src/artifacts/index.js +1 -0
- package/dist/src/artifacts/index.js.map +1 -1
- package/dist/src/artifacts/preparer.js +7 -2
- package/dist/src/artifacts/preparer.js.map +1 -1
- package/dist/src/backends/harbor/agent-budget.js +31 -0
- package/dist/src/backends/harbor/agent-budget.js.map +1 -0
- package/dist/src/backends/harbor/backend.js +59 -107
- package/dist/src/backends/harbor/backend.js.map +1 -1
- package/dist/src/backends/harbor/environment-config.js +92 -0
- package/dist/src/backends/harbor/environment-config.js.map +1 -0
- package/dist/src/backends/harbor/index.js +6 -1
- package/dist/src/backends/harbor/index.js.map +1 -1
- package/dist/src/backends/harbor/model-proxy-config.js +23 -0
- package/dist/src/backends/harbor/model-proxy-config.js.map +1 -0
- package/dist/src/backends/harbor/process.js +177 -0
- package/dist/src/backends/harbor/process.js.map +1 -0
- package/dist/src/backends/harbor/regrade.js +77 -0
- package/dist/src/backends/harbor/regrade.js.map +1 -0
- package/dist/src/backends/harbor/runtime-toolchain.js +30 -0
- package/dist/src/backends/harbor/runtime-toolchain.js.map +1 -0
- package/dist/src/backends/harbor/supervisor.js +74 -0
- package/dist/src/backends/harbor/supervisor.js.map +1 -0
- package/dist/src/backends/harbor/task-resources.js +206 -0
- package/dist/src/backends/harbor/task-resources.js.map +1 -0
- package/dist/src/backends/harbor/tools.js +6 -29
- package/dist/src/backends/harbor/tools.js.map +1 -1
- package/dist/src/backends/index.js +5 -1
- package/dist/src/backends/index.js.map +1 -1
- package/dist/src/benchmarks/index.js +3 -0
- package/dist/src/benchmarks/index.js.map +1 -0
- package/dist/src/benchmarks/loader.js +271 -0
- package/dist/src/benchmarks/loader.js.map +1 -0
- package/dist/src/benchmarks/metrics.js +16 -0
- package/dist/src/benchmarks/metrics.js.map +1 -0
- package/dist/src/benchmarks/toml.js +13 -0
- package/dist/src/benchmarks/toml.js.map +1 -0
- package/dist/src/benchmarks/validation.js +242 -0
- package/dist/src/benchmarks/validation.js.map +1 -0
- package/dist/src/cli/arguments.js +6 -4
- package/dist/src/cli/arguments.js.map +1 -1
- package/dist/src/cli/commands/benchmark.js +22 -0
- package/dist/src/cli/commands/benchmark.js.map +1 -0
- package/dist/src/cli/commands/capabilities.js +18 -0
- package/dist/src/cli/commands/capabilities.js.map +1 -0
- package/dist/src/cli/commands/daemon.js +135 -8
- package/dist/src/cli/commands/daemon.js.map +1 -1
- package/dist/src/cli/commands/eval.js +228 -6
- package/dist/src/cli/commands/eval.js.map +1 -1
- package/dist/src/cli/commands/images.js +40 -0
- package/dist/src/cli/commands/images.js.map +1 -0
- package/dist/src/cli/commands/run.js +42 -17
- package/dist/src/cli/commands/run.js.map +1 -1
- package/dist/src/cli/commands/runs.js +25 -3
- package/dist/src/cli/commands/runs.js.map +1 -1
- package/dist/src/cli/commands/trajectory.js +96 -5
- package/dist/src/cli/commands/trajectory.js.map +1 -1
- package/dist/src/cli/commands/verifier.js +36 -0
- package/dist/src/cli/commands/verifier.js.map +1 -0
- package/dist/src/cli/commands/worker.js +79 -0
- package/dist/src/cli/commands/worker.js.map +1 -0
- package/dist/src/cli/main.js +10 -0
- package/dist/src/cli/main.js.map +1 -1
- package/dist/src/cli/output.js +84 -4
- package/dist/src/cli/output.js.map +1 -1
- package/dist/src/control-plane/build-admission.js +69 -0
- package/dist/src/control-plane/build-admission.js.map +1 -0
- package/dist/src/control-plane/build-records.js +6 -0
- package/dist/src/control-plane/build-records.js.map +1 -0
- package/dist/src/control-plane/collisions.js +40 -0
- package/dist/src/control-plane/collisions.js.map +1 -0
- package/dist/src/control-plane/eval-control-state.js +22 -0
- package/dist/src/control-plane/eval-control-state.js.map +1 -0
- package/dist/src/control-plane/eval-control-work.js +44 -0
- package/dist/src/control-plane/eval-control-work.js.map +1 -0
- package/dist/src/control-plane/eval-image-resolution.js +52 -0
- package/dist/src/control-plane/eval-image-resolution.js.map +1 -0
- package/dist/src/control-plane/eval-image-services.js +30 -0
- package/dist/src/control-plane/eval-image-services.js.map +1 -0
- package/dist/src/control-plane/eval-records.js +252 -0
- package/dist/src/control-plane/eval-records.js.map +1 -0
- package/dist/src/control-plane/eval-recovery.js +181 -0
- package/dist/src/control-plane/eval-recovery.js.map +1 -0
- package/dist/src/control-plane/eval-scheduler.js +450 -0
- package/dist/src/control-plane/eval-scheduler.js.map +1 -0
- package/dist/src/control-plane/index.js +20 -0
- package/dist/src/control-plane/index.js.map +1 -0
- package/dist/src/control-plane/local-worker.js +49 -0
- package/dist/src/control-plane/local-worker.js.map +1 -0
- package/dist/src/control-plane/model-capture-planning.js +12 -0
- package/dist/src/control-plane/model-capture-planning.js.map +1 -0
- package/dist/src/control-plane/remote-result-transport.js +175 -0
- package/dist/src/control-plane/remote-result-transport.js.map +1 -0
- package/dist/src/control-plane/remote-work-coordinator.js +333 -0
- package/dist/src/control-plane/remote-work-coordinator.js.map +1 -0
- package/dist/src/control-plane/remote-work-inputs.js +266 -0
- package/dist/src/control-plane/remote-work-inputs.js.map +1 -0
- package/dist/src/control-plane/remote-work-item.js +62 -0
- package/dist/src/control-plane/remote-work-item.js.map +1 -0
- package/dist/src/control-plane/remote-work-recovery.js +220 -0
- package/dist/src/control-plane/remote-work-recovery.js.map +1 -0
- package/dist/src/control-plane/remote-worker-artifacts.js +135 -0
- package/dist/src/control-plane/remote-worker-artifacts.js.map +1 -0
- package/dist/src/control-plane/remote-worker-client.js +202 -0
- package/dist/src/control-plane/remote-worker-client.js.map +1 -0
- package/dist/src/control-plane/remote-worker-configuration.js +34 -0
- package/dist/src/control-plane/remote-worker-configuration.js.map +1 -0
- package/dist/src/control-plane/remote-worker-credentials.js +58 -0
- package/dist/src/control-plane/remote-worker-credentials.js.map +1 -0
- package/dist/src/control-plane/remote-worker-protocol.js +489 -0
- package/dist/src/control-plane/remote-worker-protocol.js.map +1 -0
- package/dist/src/control-plane/remote-worker-runner.js +214 -0
- package/dist/src/control-plane/remote-worker-runner.js.map +1 -0
- package/dist/src/control-plane/remote-workers.js +377 -0
- package/dist/src/control-plane/remote-workers.js.map +1 -0
- package/dist/src/control-plane/rerun-scheduler.js +448 -0
- package/dist/src/control-plane/rerun-scheduler.js.map +1 -0
- package/dist/src/control-plane/rerun-submission.js +62 -0
- package/dist/src/control-plane/rerun-submission.js.map +1 -0
- package/dist/src/control-plane/resources.js +182 -0
- package/dist/src/control-plane/resources.js.map +1 -0
- package/dist/src/control-plane/scheduler-eval-entry.js +32 -0
- package/dist/src/control-plane/scheduler-eval-entry.js.map +1 -0
- package/dist/src/control-plane/synthetic-result.js +21 -0
- package/dist/src/control-plane/synthetic-result.js.map +1 -0
- package/dist/src/control-plane/work-admission.js +23 -0
- package/dist/src/control-plane/work-admission.js.map +1 -0
- package/dist/src/control-plane/work-dispatcher.js +237 -0
- package/dist/src/control-plane/work-dispatcher.js.map +1 -0
- package/dist/src/controller-runtime/hash.js +19 -3
- package/dist/src/controller-runtime/hash.js.map +1 -1
- package/dist/src/controller-runtime/index.js +1 -1
- package/dist/src/controller-runtime/index.js.map +1 -1
- package/dist/src/controller-runtime/store.js +15 -6
- package/dist/src/controller-runtime/store.js.map +1 -1
- package/dist/src/daemon/auth.js +5 -0
- package/dist/src/daemon/auth.js.map +1 -1
- package/dist/src/daemon/client.js.map +1 -1
- package/dist/src/daemon/event-stream.js +46 -0
- package/dist/src/daemon/event-stream.js.map +1 -0
- package/dist/src/daemon/health.js +66 -0
- package/dist/src/daemon/health.js.map +1 -0
- package/dist/src/daemon/launcher.js +28 -1
- package/dist/src/daemon/launcher.js.map +1 -1
- package/dist/src/daemon/logging.js +23 -0
- package/dist/src/daemon/logging.js.map +1 -0
- package/dist/src/daemon/scheduler.js +26 -5
- package/dist/src/daemon/scheduler.js.map +1 -1
- package/dist/src/daemon/server.js +255 -98
- package/dist/src/daemon/server.js.map +1 -1
- package/dist/src/daemon/telemetry.js +163 -0
- package/dist/src/daemon/telemetry.js.map +1 -0
- package/dist/src/daemon/worker-routes.js +219 -0
- package/dist/src/daemon/worker-routes.js.map +1 -0
- package/dist/src/domain/benchmarks.js +2 -0
- package/dist/src/domain/benchmarks.js.map +1 -0
- package/dist/src/domain/bundles.js +2 -0
- package/dist/src/domain/bundles.js.map +1 -0
- package/dist/src/domain/control-plane.js +2 -0
- package/dist/src/domain/control-plane.js.map +1 -0
- package/dist/src/domain/execution-evidence.js +2 -0
- package/dist/src/domain/execution-evidence.js.map +1 -0
- package/dist/src/domain/execution-plan.js +2 -0
- package/dist/src/domain/execution-plan.js.map +1 -0
- package/dist/src/domain/images.js +2 -0
- package/dist/src/domain/images.js.map +1 -0
- package/dist/src/domain/index.js +13 -0
- package/dist/src/domain/index.js.map +1 -1
- package/dist/src/domain/interactions.js +2 -0
- package/dist/src/domain/interactions.js.map +1 -0
- package/dist/src/domain/providers.js +2 -0
- package/dist/src/domain/providers.js.map +1 -0
- package/dist/src/domain/resources.js +2 -0
- package/dist/src/domain/resources.js.map +1 -0
- package/dist/src/domain/runs.js.map +1 -1
- package/dist/src/domain/validation.js +11 -4
- package/dist/src/domain/validation.js.map +1 -1
- package/dist/src/domain/verifier-evidence-validation.js +214 -0
- package/dist/src/domain/verifier-evidence-validation.js.map +1 -0
- package/dist/src/domain/verifier-evidence.js +3 -0
- package/dist/src/domain/verifier-evidence.js.map +1 -0
- package/dist/src/domain/verifier-score-contract.js +232 -0
- package/dist/src/domain/verifier-score-contract.js.map +1 -0
- package/dist/src/domain/workers.js +2 -0
- package/dist/src/domain/workers.js.map +1 -0
- package/dist/src/evals/benchmark-adapter-manifest.js +219 -0
- package/dist/src/evals/benchmark-adapter-manifest.js.map +1 -0
- package/dist/src/evals/benchmark-candidate.js +36 -0
- package/dist/src/evals/benchmark-candidate.js.map +1 -0
- package/dist/src/evals/benchmark-run.js +166 -0
- package/dist/src/evals/benchmark-run.js.map +1 -0
- package/dist/src/evals/collect-only-rerun.js +183 -0
- package/dist/src/evals/collect-only-rerun.js.map +1 -0
- package/dist/src/evals/directory.js +30 -0
- package/dist/src/evals/directory.js.map +1 -0
- package/dist/src/evals/docker-engine-stats.js +89 -0
- package/dist/src/evals/docker-engine-stats.js.map +1 -0
- package/dist/src/evals/docker-ownership.js +51 -0
- package/dist/src/evals/docker-ownership.js.map +1 -0
- package/dist/src/evals/docker-reaper.js +144 -0
- package/dist/src/evals/docker-reaper.js.map +1 -0
- package/dist/src/evals/docker-resource-observer.js +207 -0
- package/dist/src/evals/docker-resource-observer.js.map +1 -0
- package/dist/src/evals/duration-estimator.js +126 -0
- package/dist/src/evals/duration-estimator.js.map +1 -0
- package/dist/src/evals/environment-image-planning.js +164 -0
- package/dist/src/evals/environment-image-planning.js.map +1 -0
- package/dist/src/evals/eval-artifact-planning.js +56 -0
- package/dist/src/evals/eval-artifact-planning.js.map +1 -0
- package/dist/src/evals/eval-finalization.js +10 -0
- package/dist/src/evals/eval-finalization.js.map +1 -0
- package/dist/src/evals/eval-lifecycle-events.js +29 -0
- package/dist/src/evals/eval-lifecycle-events.js.map +1 -0
- package/dist/src/evals/eval-logical-plan.js +51 -0
- package/dist/src/evals/eval-logical-plan.js.map +1 -0
- package/dist/src/evals/eval-result-builder.js +81 -0
- package/dist/src/evals/eval-result-builder.js.map +1 -0
- package/dist/src/evals/events.js +53 -5
- package/dist/src/evals/events.js.map +1 -1
- package/dist/src/evals/evolution-baseline.js +191 -0
- package/dist/src/evals/evolution-baseline.js.map +1 -0
- package/dist/src/evals/execution-evidence.js +116 -0
- package/dist/src/evals/execution-evidence.js.map +1 -0
- package/dist/src/evals/execution-leases.js +342 -0
- package/dist/src/evals/execution-leases.js.map +1 -0
- package/dist/src/evals/execution-plan-artifacts.js +72 -0
- package/dist/src/evals/execution-plan-artifacts.js.map +1 -0
- package/dist/src/evals/execution-plan-images.js +69 -0
- package/dist/src/evals/execution-plan-images.js.map +1 -0
- package/dist/src/evals/execution-plan-resources.js +174 -0
- package/dist/src/evals/execution-plan-resources.js.map +1 -0
- package/dist/src/evals/execution-plan.js +451 -0
- package/dist/src/evals/execution-plan.js.map +1 -0
- package/dist/src/evals/failure-classifier.js +61 -0
- package/dist/src/evals/failure-classifier.js.map +1 -0
- package/dist/src/evals/harbor-artifact-builder-image.js +99 -0
- package/dist/src/evals/harbor-artifact-builder-image.js.map +1 -0
- package/dist/src/evals/harbor-artifact-builder.js +358 -0
- package/dist/src/evals/harbor-artifact-builder.js.map +1 -0
- package/dist/src/evals/harbor-bridge-error.js +58 -3
- package/dist/src/evals/harbor-bridge-error.js.map +1 -1
- package/dist/src/evals/harbor-node-runtime.js +146 -0
- package/dist/src/evals/harbor-node-runtime.js.map +1 -0
- package/dist/src/evals/harbor-phase-timings.js +60 -0
- package/dist/src/evals/harbor-phase-timings.js.map +1 -0
- package/dist/src/evals/index.js +40 -2
- package/dist/src/evals/index.js.map +1 -1
- package/dist/src/evals/infrastructure-retry.js +208 -84
- package/dist/src/evals/infrastructure-retry.js.map +1 -1
- package/dist/src/evals/interaction-capture-import.js +49 -0
- package/dist/src/evals/interaction-capture-import.js.map +1 -0
- package/dist/src/evals/local-docker-provider.js +372 -0
- package/dist/src/evals/local-docker-provider.js.map +1 -0
- package/dist/src/evals/local-eval-planning.js +80 -0
- package/dist/src/evals/local-eval-planning.js.map +1 -0
- package/dist/src/evals/model-capture-plan.js +22 -0
- package/dist/src/evals/model-capture-plan.js.map +1 -0
- package/dist/src/evals/model-capture-runtime.js +114 -0
- package/dist/src/evals/model-capture-runtime.js.map +1 -0
- package/dist/src/evals/model-proxy-runtime-state.js +53 -0
- package/dist/src/evals/model-proxy-runtime-state.js.map +1 -0
- package/dist/src/evals/native-phase-evidence.js +440 -0
- package/dist/src/evals/native-phase-evidence.js.map +1 -0
- package/dist/src/evals/physical-retry-work.js +31 -0
- package/dist/src/evals/physical-retry-work.js.map +1 -0
- package/dist/src/evals/planned-execution-support.js +66 -0
- package/dist/src/evals/planned-execution-support.js.map +1 -0
- package/dist/src/evals/planned-execution.js +469 -0
- package/dist/src/evals/planned-execution.js.map +1 -0
- package/dist/src/evals/planned-progress-publisher.js +62 -0
- package/dist/src/evals/planned-progress-publisher.js.map +1 -0
- package/dist/src/evals/planned-remote-execution.js +40 -0
- package/dist/src/evals/planned-remote-execution.js.map +1 -0
- package/dist/src/evals/planned-retry-execution.js +174 -0
- package/dist/src/evals/planned-retry-execution.js.map +1 -0
- package/dist/src/evals/planned-retry-lifecycle.js +188 -0
- package/dist/src/evals/planned-retry-lifecycle.js.map +1 -0
- package/dist/src/evals/preparation-rerun.js +225 -0
- package/dist/src/evals/preparation-rerun.js.map +1 -0
- package/dist/src/evals/prepared-harness.js +76 -0
- package/dist/src/evals/prepared-harness.js.map +1 -0
- package/dist/src/evals/progress.js +52 -12
- package/dist/src/evals/progress.js.map +1 -1
- package/dist/src/evals/records.js +20 -3
- package/dist/src/evals/records.js.map +1 -1
- package/dist/src/evals/recovery.js +242 -0
- package/dist/src/evals/recovery.js.map +1 -0
- package/dist/src/evals/regrade-evidence.js +77 -0
- package/dist/src/evals/regrade-evidence.js.map +1 -0
- package/dist/src/evals/remote-infrastructure-retry.js +165 -0
- package/dist/src/evals/remote-infrastructure-retry.js.map +1 -0
- package/dist/src/evals/request.js +21 -3
- package/dist/src/evals/request.js.map +1 -1
- package/dist/src/evals/rerun-inputs.js +108 -0
- package/dist/src/evals/rerun-inputs.js.map +1 -0
- package/dist/src/evals/rerun-slots.js +4 -3
- package/dist/src/evals/rerun-slots.js.map +1 -1
- package/dist/src/evals/rerun-types.js +45 -0
- package/dist/src/evals/rerun-types.js.map +1 -0
- package/dist/src/evals/rerun.js +133 -88
- package/dist/src/evals/rerun.js.map +1 -1
- package/dist/src/evals/result-helpers.js +78 -0
- package/dist/src/evals/result-helpers.js.map +1 -0
- package/dist/src/evals/resume-state.js +36 -0
- package/dist/src/evals/resume-state.js.map +1 -0
- package/dist/src/evals/retry-backoff.js +20 -0
- package/dist/src/evals/retry-backoff.js.map +1 -0
- package/dist/src/evals/retry-state.js +195 -0
- package/dist/src/evals/retry-state.js.map +1 -0
- package/dist/src/evals/scheduler-metrics.js +95 -0
- package/dist/src/evals/scheduler-metrics.js.map +1 -0
- package/dist/src/evals/scheduler-trace-replay.js +107 -0
- package/dist/src/evals/scheduler-trace-replay.js.map +1 -0
- package/dist/src/evals/service-types.js +2 -0
- package/dist/src/evals/service-types.js.map +1 -0
- package/dist/src/evals/service.js +331 -347
- package/dist/src/evals/service.js.map +1 -1
- package/dist/src/evals/task-resources.js +209 -0
- package/dist/src/evals/task-resources.js.map +1 -0
- package/dist/src/evals/trial-environment-evidence.js +87 -0
- package/dist/src/evals/trial-environment-evidence.js.map +1 -0
- package/dist/src/evals/trial-execution-evidence.js +13 -0
- package/dist/src/evals/trial-execution-evidence.js.map +1 -0
- package/dist/src/evals/trial-import-identity.js +31 -0
- package/dist/src/evals/trial-import-identity.js.map +1 -0
- package/dist/src/evals/trial-import.js +183 -188
- package/dist/src/evals/trial-import.js.map +1 -1
- package/dist/src/evals/trial-publication-recovery.js +144 -0
- package/dist/src/evals/trial-publication-recovery.js.map +1 -0
- package/dist/src/evals/trial-publication.js +25 -0
- package/dist/src/evals/trial-publication.js.map +1 -0
- package/dist/src/evals/trial-reference-validation.js +49 -0
- package/dist/src/evals/trial-reference-validation.js.map +1 -0
- package/dist/src/evals/verifier-artifacts.js +317 -0
- package/dist/src/evals/verifier-artifacts.js.map +1 -0
- package/dist/src/evals/verifier-eligibility.js +25 -0
- package/dist/src/evals/verifier-eligibility.js.map +1 -0
- package/dist/src/evals/verifier-only-rerun.js +210 -0
- package/dist/src/evals/verifier-only-rerun.js.map +1 -0
- package/dist/src/evals/verifier-runtime.js +61 -0
- package/dist/src/evals/verifier-runtime.js.map +1 -0
- package/dist/src/evals/verifier-score-artifacts.js +119 -0
- package/dist/src/evals/verifier-score-artifacts.js.map +1 -0
- package/dist/src/evals/work-item-artifacts.js +23 -0
- package/dist/src/evals/work-item-artifacts.js.map +1 -0
- package/dist/src/foundation/config.js +17 -0
- package/dist/src/foundation/config.js.map +1 -1
- package/dist/src/foundation/contained-file.js +85 -0
- package/dist/src/foundation/contained-file.js.map +1 -0
- package/dist/src/foundation/credential-redaction.js +162 -0
- package/dist/src/foundation/credential-redaction.js.map +1 -0
- package/dist/src/foundation/environment-image-references.js +59 -0
- package/dist/src/foundation/environment-image-references.js.map +1 -0
- package/dist/src/foundation/executable.js +45 -5
- package/dist/src/foundation/executable.js.map +1 -1
- package/dist/src/foundation/fs.js +2 -1
- package/dist/src/foundation/fs.js.map +1 -1
- package/dist/src/foundation/index.js +6 -2
- package/dist/src/foundation/index.js.map +1 -1
- package/dist/src/foundation/process-identity.js +69 -0
- package/dist/src/foundation/process-identity.js.map +1 -0
- package/dist/src/images/context.js +47 -0
- package/dist/src/images/context.js.map +1 -0
- package/dist/src/images/docker-buildkit.js +137 -0
- package/dist/src/images/docker-buildkit.js.map +1 -0
- package/dist/src/images/dockerfile.js +67 -0
- package/dist/src/images/dockerfile.js.map +1 -0
- package/dist/src/images/gc.js +261 -0
- package/dist/src/images/gc.js.map +1 -0
- package/dist/src/images/index.js +10 -0
- package/dist/src/images/index.js.map +1 -0
- package/dist/src/images/manifest.js +109 -0
- package/dist/src/images/manifest.js.map +1 -0
- package/dist/src/images/ownership.js +5 -0
- package/dist/src/images/ownership.js.map +1 -0
- package/dist/src/images/records.js +32 -0
- package/dist/src/images/records.js.map +1 -0
- package/dist/src/images/registry.js +154 -0
- package/dist/src/images/registry.js.map +1 -0
- package/dist/src/images/service.js +224 -0
- package/dist/src/images/service.js.map +1 -0
- package/dist/src/model-access/capture.js +255 -0
- package/dist/src/model-access/capture.js.map +1 -0
- package/dist/src/model-access/index.js +5 -0
- package/dist/src/model-access/index.js.map +1 -0
- package/dist/src/model-access/policy.js +49 -0
- package/dist/src/model-access/policy.js.map +1 -0
- package/dist/src/model-access/proxy.js +336 -0
- package/dist/src/model-access/proxy.js.map +1 -0
- package/dist/src/model-access/records.js +167 -0
- package/dist/src/model-access/records.js.map +1 -0
- package/dist/src/runs/adapter-process.js +12 -0
- package/dist/src/runs/adapter-process.js.map +1 -0
- package/dist/src/runs/bundle.js +450 -0
- package/dist/src/runs/bundle.js.map +1 -0
- package/dist/src/runs/events.js +12 -1
- package/dist/src/runs/events.js.map +1 -1
- package/dist/src/runs/executor.js +46 -45
- package/dist/src/runs/executor.js.map +1 -1
- package/dist/src/runs/finalizer.js +17 -0
- package/dist/src/runs/finalizer.js.map +1 -1
- package/dist/src/runs/index.js +8 -1
- package/dist/src/runs/index.js.map +1 -1
- package/dist/src/runs/manifest.js +5 -5
- package/dist/src/runs/manifest.js.map +1 -1
- package/dist/src/runs/phase-bundle.js +86 -0
- package/dist/src/runs/phase-bundle.js.map +1 -0
- package/dist/src/runs/phase-cancellation.js +107 -0
- package/dist/src/runs/phase-cancellation.js.map +1 -0
- package/dist/src/runs/phase-group.js +113 -0
- package/dist/src/runs/phase-group.js.map +1 -0
- package/dist/src/runs/query.js +1 -1
- package/dist/src/runs/query.js.map +1 -1
- package/dist/src/runs/records.js +3 -0
- package/dist/src/runs/records.js.map +1 -1
- package/dist/src/runs/request.js +10 -3
- package/dist/src/runs/request.js.map +1 -1
- package/dist/src/runs/training-candidate.js +161 -0
- package/dist/src/runs/training-candidate.js.map +1 -0
- package/dist/src/runs/verifier-evidence-redaction.js +61 -0
- package/dist/src/runs/verifier-evidence-redaction.js.map +1 -0
- package/dist/src/runs/verifier-evidence.js +459 -0
- package/dist/src/runs/verifier-evidence.js.map +1 -0
- package/dist/src/runs/verifier-structured-evidence.js +83 -0
- package/dist/src/runs/verifier-structured-evidence.js.map +1 -0
- package/dist/src/trajectories/analysis.js +288 -0
- package/dist/src/trajectories/analysis.js.map +1 -0
- package/dist/src/trajectories/chunk-projection.js +148 -0
- package/dist/src/trajectories/chunk-projection.js.map +1 -0
- package/dist/src/trajectories/content-projection.js +389 -0
- package/dist/src/trajectories/content-projection.js.map +1 -0
- package/dist/src/trajectories/dsh-chunk-contract.js +176 -0
- package/dist/src/trajectories/dsh-chunk-contract.js.map +1 -0
- package/dist/src/trajectories/dsh-contract.js +284 -0
- package/dist/src/trajectories/dsh-contract.js.map +1 -0
- package/dist/src/trajectories/events-chunk-drill.js +95 -0
- package/dist/src/trajectories/events-chunk-drill.js.map +1 -0
- package/dist/src/trajectories/events-page.js +367 -0
- package/dist/src/trajectories/events-page.js.map +1 -0
- package/dist/src/trajectories/index.js +3 -0
- package/dist/src/trajectories/index.js.map +1 -1
- package/dist/src/trajectories/provider-capture.js +15 -24
- package/dist/src/trajectories/provider-capture.js.map +1 -1
- package/dist/src/trajectories/providers/deepseek.js +3 -3
- package/dist/src/trajectories/providers/deepseek.js.map +1 -1
- package/dist/src/trajectories/request-attempt.js +47 -0
- package/dist/src/trajectories/request-attempt.js.map +1 -0
- package/dist/src/trajectories/stream-reader.js +311 -0
- package/dist/src/trajectories/stream-reader.js.map +1 -0
- package/dist/src/trajectories/surface-fold.js +126 -0
- package/dist/src/trajectories/surface-fold.js.map +1 -0
- package/dist/src/workers/index.js +3 -0
- package/dist/src/workers/index.js.map +1 -0
- package/dist/src/workers/remote-harbor-work-spec.js +97 -0
- package/dist/src/workers/remote-harbor-work-spec.js.map +1 -0
- package/dist/src/workers/remote-harbor-worker.js +223 -0
- package/dist/src/workers/remote-harbor-worker.js.map +1 -0
- package/docs/schemas/benchmark-hook-request-v1.schema.json +65 -0
- package/docs/schemas/benchmark-hook-response-v1.schema.json +75 -0
- package/docs/schemas/benchmark-hook-v1.schema.json +33 -0
- package/docs/schemas/benchmark-lock-v1.schema.json +351 -0
- package/docs/schemas/benchmark-metric-v1.schema.json +37 -0
- package/docs/schemas/benchmark-package-v1.schema.json +167 -0
- package/docs/schemas/benchmark-phase-group.schema.json +41 -0
- package/docs/schemas/benchmark-profile-v1.schema.json +141 -0
- package/docs/schemas/benchmark-task-v1.schema.json +384 -0
- package/docs/schemas/benchmark-tool-result-v1.schema.json +30 -0
- package/docs/schemas/environment-build-index.schema.json +13 -0
- package/docs/schemas/environment-build-record.schema.json +25 -0
- package/docs/schemas/environment-image-manifest.schema.json +62 -0
- package/docs/schemas/error.schema.json +22 -0
- package/docs/schemas/eval-control.schema.json +36 -0
- package/docs/schemas/eval-execution-plan.schema.json +216 -0
- package/docs/schemas/eval-progress.schema.json +1 -15
- package/docs/schemas/eval-rerun-result.schema.json +58 -0
- package/docs/schemas/eval-rerun-state.schema.json +61 -0
- package/docs/schemas/eval-rerun-submission.schema.json +60 -0
- package/docs/schemas/eval-result.schema.json +32 -13
- package/docs/schemas/eval-submission-input.schema.json +14 -0
- package/docs/schemas/eval-submission.schema.json +68 -0
- package/docs/schemas/eval-trial-publication.schema.json +15 -0
- package/docs/schemas/eval-trial-reference.schema.json +129 -0
- package/docs/schemas/execution-evidence.schema.json +96 -0
- package/docs/schemas/execution-lease.schema.json +50 -0
- package/docs/schemas/execution-provider-status.schema.json +65 -0
- package/docs/schemas/execution-worker.schema.json +53 -0
- package/docs/schemas/harbor-process-exit.schema.json +14 -0
- package/docs/schemas/interaction-capture-ref.schema.json +45 -0
- package/docs/schemas/local-provider-execution.schema.json +33 -0
- package/docs/schemas/model-interaction.schema.json +41 -0
- package/docs/schemas/regrade-assessment-reference.schema.json +11 -0
- package/docs/schemas/remote-result-envelope.schema.json +32 -0
- package/docs/schemas/remote-tree-envelope.schema.json +25 -0
- package/docs/schemas/remote-work-offer.schema.json +58 -0
- package/docs/schemas/remote-work-receipt.schema.json +50 -0
- package/docs/schemas/remote-worker-artifact.schema.json +17 -0
- package/docs/schemas/remote-worker-event.schema.json +18 -0
- package/docs/schemas/remote-worker-heartbeat.schema.json +37 -0
- package/docs/schemas/remote-worker-registration.schema.json +52 -0
- package/docs/schemas/result-bundle-index.schema.json +133 -0
- package/docs/schemas/run-context.schema.json +15 -0
- package/docs/schemas/run-environment-images.schema.json +39 -0
- package/docs/schemas/run-manifest.schema.json +4 -0
- package/docs/schemas/run-request.schema.json +4 -0
- package/docs/schemas/training-data-candidate.schema.json +46 -0
- package/docs/schemas/trajectory-analysis.schema.json +258 -0
- package/docs/schemas/trajectory-events-page.schema.json +40 -0
- package/docs/schemas/verifier-assessment.schema.json +77 -0
- package/docs/schemas/verifier-evidence.schema.json +276 -0
- package/integrations/harbor/hitch_benchmark.py +252 -0
- package/integrations/harbor/hitch_candidate_recycle.py +262 -0
- package/integrations/harbor/hitch_harbor_agent.py +698 -568
- package/integrations/harbor/hitch_harbor_environment.py +469 -0
- package/integrations/harbor/hitch_harbor_task_resources.py +275 -0
- package/integrations/harbor/hitch_harbor_verifier.py +83 -2
- package/integrations/harbor/hitch_phase_supervisor.py +452 -0
- package/integrations/harbor/hitch_tool_client.mjs +91 -0
- package/integrations/model-call/cli.js +57 -0
- package/package.json +30 -6
|
@@ -0,0 +1,225 @@
|
|
|
1
|
+
import { cp, lstat, readdir, rename } from "node:fs/promises";
|
|
2
|
+
import path from "node:path";
|
|
3
|
+
import { HitchError, SCHEMA_VERSION, atomicWriteJSON, ensureDir, readJSON } from "../foundation/index.js";
|
|
4
|
+
import { logicalPlanModelCapture, readEvalLogicalPlan } from "./eval-logical-plan.js";
|
|
5
|
+
import { readEvalProgress } from "./progress.js";
|
|
6
|
+
import { resolveLocalDatasetTaskIds, validateEvalId } from "./request.js";
|
|
7
|
+
import { evalRerunSemantics } from "./rerun-types.js";
|
|
8
|
+
import { invalidTrialSlots, slotKey, sortSlots, uniqueTasks, validateProgressPlan } from "./rerun-slots.js";
|
|
9
|
+
import { runEval } from "./service.js";
|
|
10
|
+
/** Restart an eval that failed before its executable plan was fully materialized. */
|
|
11
|
+
export async function restartIncompleteEval(options, request, startedAt, rerunDirectory, statePath) {
|
|
12
|
+
const materializedFiles = ["plan.json", "execution-plan.json", "progress.json"];
|
|
13
|
+
if ((await Promise.all(materializedFiles.map((name) => exists(path.join(options.evalDirectory, name))))).every(Boolean))
|
|
14
|
+
return null;
|
|
15
|
+
const failedResult = await readJSON(path.join(options.evalDirectory, "result.json"), null);
|
|
16
|
+
if (failedResult === null)
|
|
17
|
+
return null;
|
|
18
|
+
if (options.rerunType !== "candidate-restart" || options.selector.mode !== "invalid") {
|
|
19
|
+
throw unavailable("an eval with incomplete preparation can only be restarted with candidate-restart --invalid");
|
|
20
|
+
}
|
|
21
|
+
if (failedResult?.status === "cancelled") {
|
|
22
|
+
throw new HitchError("cancelled eval cannot be rerun", { code: "eval_rerun_cancelled", exitCode: 2 });
|
|
23
|
+
}
|
|
24
|
+
if (failedResult?.status !== "failed" || !Array.isArray(failedResult.trials) || failedResult.trials.length !== 0) {
|
|
25
|
+
throw unavailable("eval preparation is incomplete but the source is not a failed pre-execution attempt");
|
|
26
|
+
}
|
|
27
|
+
const taskIds = await resolveLocalDatasetTaskIds(request.dataset);
|
|
28
|
+
if (taskIds === null || taskIds.length === 0) {
|
|
29
|
+
throw unavailable("incomplete preparation recovery requires an enumerable local task plan");
|
|
30
|
+
}
|
|
31
|
+
const logicalPlan = await readEvalLogicalPlan(options.evalDirectory, options.evalId, request).catch((error) => {
|
|
32
|
+
throw unavailable(error.message);
|
|
33
|
+
});
|
|
34
|
+
let modelCapturePlan = logicalPlanModelCapture(logicalPlan) ?? options.modelCapturePlan;
|
|
35
|
+
if (!modelCapturePlan && !await exists(path.join(options.evalDirectory, "submission.json"))) {
|
|
36
|
+
modelCapturePlan = defaultModelCapturePlan();
|
|
37
|
+
}
|
|
38
|
+
if (!modelCapturePlan) {
|
|
39
|
+
throw unavailable("incomplete preparation recovery requires the original model capture plan");
|
|
40
|
+
}
|
|
41
|
+
await writeRequest(rerunDirectory, options, startedAt, [], []);
|
|
42
|
+
await writeState(statePath, options, startedAt, "running", [], [], [], []);
|
|
43
|
+
await archiveIncompleteAttempt(options, rerunDirectory, failedResult);
|
|
44
|
+
try {
|
|
45
|
+
const restarted = await runEval({
|
|
46
|
+
evalId: validateEvalId(options.evalId),
|
|
47
|
+
root: options.root,
|
|
48
|
+
request,
|
|
49
|
+
normalizedRequest: request,
|
|
50
|
+
precreated: true,
|
|
51
|
+
replaceTerminal: true,
|
|
52
|
+
env: options.env ?? process.env,
|
|
53
|
+
...(options.maxConcurrentOverride === undefined ? {} : { maxConcurrentOverride: options.maxConcurrentOverride }),
|
|
54
|
+
...(options.harborExecutable === undefined ? {} : { harborExecutable: options.harborExecutable }),
|
|
55
|
+
...(options.signal ? { signal: options.signal } : {}),
|
|
56
|
+
...(options.trialBundleGraceMs === undefined ? {} : { trialBundleGraceMs: options.trialBundleGraceMs }),
|
|
57
|
+
...(options.executionResources ? { executionResources: options.executionResources } : {}),
|
|
58
|
+
...(options.executionResourceSource ? { executionResourceSource: options.executionResourceSource } : {}),
|
|
59
|
+
executionStrategy: options.executionStrategy
|
|
60
|
+
?? (logicalPlan?.attempt_execution === "harbor-task-slots-v1" ? "local-task-slots-v1" : "legacy-attempt-shards"),
|
|
61
|
+
...(options.environmentBuildMode ? { environmentBuildMode: options.environmentBuildMode } : {}),
|
|
62
|
+
modelCapturePlan,
|
|
63
|
+
...(options.harborArtifactBuilder ? { harborArtifactBuilder: options.harborArtifactBuilder } : {}),
|
|
64
|
+
});
|
|
65
|
+
if (restarted.status === "cancelled") {
|
|
66
|
+
throw new HitchError("eval rerun was aborted", { code: "eval_rerun_aborted", exitCode: 9 });
|
|
67
|
+
}
|
|
68
|
+
const [planValue, executionPlanValue, progress] = await Promise.all([
|
|
69
|
+
readJSON(path.join(options.evalDirectory, "plan.json"), null),
|
|
70
|
+
readJSON(path.join(options.evalDirectory, "execution-plan.json"), null),
|
|
71
|
+
readEvalProgress(options.evalDirectory),
|
|
72
|
+
]);
|
|
73
|
+
if (planValue === null || executionPlanValue === null || progress === null)
|
|
74
|
+
throw restartFailure(restarted);
|
|
75
|
+
const plan = parseRecoveredPlan(planValue, options.evalId, request);
|
|
76
|
+
validateProgressPlan(progress, plan, request, options.evalId);
|
|
77
|
+
const selectedTrials = sortSlots(plan.tasks.flatMap((taskId) => Array.from({ length: plan.attempts }, (_, index) => ({ task_id: taskId, attempt: index + 1 }))));
|
|
78
|
+
const valid = new Set(progress.trials.filter((trial) => trial.observation_status === "valid").map(slotKey));
|
|
79
|
+
const repairedTrials = selectedTrials.filter((trial) => valid.has(slotKey(trial)));
|
|
80
|
+
const remainingTrials = invalidTrialSlots(plan.tasks, plan.attempts, progress);
|
|
81
|
+
const selectedTasks = uniqueTasks(selectedTrials);
|
|
82
|
+
const repairedTasks = uniqueTasks(repairedTrials);
|
|
83
|
+
const remainingTasks = uniqueTasks(remainingTrials);
|
|
84
|
+
const completedAt = new Date().toISOString();
|
|
85
|
+
await writeRequest(rerunDirectory, options, startedAt, selectedTasks, selectedTrials);
|
|
86
|
+
const output = {
|
|
87
|
+
schema_version: "1",
|
|
88
|
+
kind: "eval-rerun",
|
|
89
|
+
rerun_id: options.rerunId,
|
|
90
|
+
rerun_type: options.rerunType,
|
|
91
|
+
semantics: evalRerunSemantics(options.rerunType),
|
|
92
|
+
eval_id: options.evalId,
|
|
93
|
+
status: "completed",
|
|
94
|
+
selected_tasks: selectedTasks,
|
|
95
|
+
repaired_tasks: repairedTasks,
|
|
96
|
+
remaining_invalid_tasks: remainingTasks,
|
|
97
|
+
selected_trials: selectedTrials,
|
|
98
|
+
repaired_trials: repairedTrials,
|
|
99
|
+
remaining_invalid_trials: remainingTrials,
|
|
100
|
+
eval_status: restarted.status === "succeeded" ? "succeeded" : "failed",
|
|
101
|
+
started_at: startedAt,
|
|
102
|
+
completed_at: completedAt,
|
|
103
|
+
};
|
|
104
|
+
await writeState(statePath, options, startedAt, "completed", selectedTasks, selectedTrials, repairedTasks, repairedTrials, {
|
|
105
|
+
completedAt,
|
|
106
|
+
evalStatus: output.eval_status,
|
|
107
|
+
remainingTasks,
|
|
108
|
+
remainingTrials,
|
|
109
|
+
});
|
|
110
|
+
return output;
|
|
111
|
+
}
|
|
112
|
+
catch (error) {
|
|
113
|
+
await writeState(statePath, options, startedAt, "failed", [], [], [], [], {
|
|
114
|
+
completedAt: new Date().toISOString(),
|
|
115
|
+
errorCode: error instanceof HitchError ? error.code : "eval_rerun_failed",
|
|
116
|
+
}).catch(() => { });
|
|
117
|
+
throw error;
|
|
118
|
+
}
|
|
119
|
+
}
|
|
120
|
+
async function archiveIncompleteAttempt(options, rerunDirectory, result) {
|
|
121
|
+
const archive = path.join(rerunDirectory, "previous-attempt");
|
|
122
|
+
await ensureDir(archive);
|
|
123
|
+
for (const name of ["request.json", "submission.json", "control.json", "result.json"]) {
|
|
124
|
+
const source = path.join(options.evalDirectory, name);
|
|
125
|
+
if (await exists(source))
|
|
126
|
+
await cp(source, path.join(archive, name), { recursive: true, errorOnExist: true });
|
|
127
|
+
}
|
|
128
|
+
const stable = new Set(["request.json", "submission.json", "control.json", "result.json", "reruns"]);
|
|
129
|
+
for (const entry of await readdir(options.evalDirectory, { withFileTypes: true })) {
|
|
130
|
+
if (stable.has(entry.name))
|
|
131
|
+
continue;
|
|
132
|
+
await rename(path.join(options.evalDirectory, entry.name), path.join(archive, entry.name));
|
|
133
|
+
}
|
|
134
|
+
const failure = result.error && typeof result.error === "object" && !Array.isArray(result.error)
|
|
135
|
+
? result.error
|
|
136
|
+
: null;
|
|
137
|
+
await atomicWriteJSON(path.join(archive, "recovery.json"), {
|
|
138
|
+
schema_version: SCHEMA_VERSION,
|
|
139
|
+
kind: "eval-preparation-restart",
|
|
140
|
+
eval_id: options.evalId,
|
|
141
|
+
rerun_id: options.rerunId,
|
|
142
|
+
...(typeof result.failure_stage === "string" ? { failure_stage: result.failure_stage } : {}),
|
|
143
|
+
...(typeof failure?.code === "string" ? { failure_code: failure.code } : {}),
|
|
144
|
+
archived_at: new Date().toISOString(),
|
|
145
|
+
});
|
|
146
|
+
}
|
|
147
|
+
function restartFailure(result) {
|
|
148
|
+
const error = result.error && typeof result.error === "object" && !Array.isArray(result.error)
|
|
149
|
+
? result.error
|
|
150
|
+
: null;
|
|
151
|
+
const code = typeof error?.code === "string" && error.code ? error.code : "eval_rerun_preparation_failed";
|
|
152
|
+
const stage = typeof result.failure_stage === "string" ? result.failure_stage : "preparing";
|
|
153
|
+
const message = typeof error?.message === "string" && error.message ? error.message : "eval restart did not produce an executable plan";
|
|
154
|
+
return new HitchError(`eval restart failed during ${stage}: ${message}`, {
|
|
155
|
+
code,
|
|
156
|
+
exitCode: Number.isSafeInteger(result.exit_code) ? result.exit_code : 12,
|
|
157
|
+
});
|
|
158
|
+
}
|
|
159
|
+
function parseRecoveredPlan(value, evalId, request) {
|
|
160
|
+
if (!value || typeof value !== "object" || Array.isArray(value))
|
|
161
|
+
throw unavailable("recovered eval plan is missing");
|
|
162
|
+
const plan = value;
|
|
163
|
+
if (plan.schema_version !== SCHEMA_VERSION || plan.eval_id !== evalId || plan.dataset !== request.dataset
|
|
164
|
+
|| plan.benchmark_id !== request.benchmark_id || plan.benchmark_revision !== request.benchmark_revision
|
|
165
|
+
|| !Number.isSafeInteger(plan.attempts) || plan.attempts !== request.attempts) {
|
|
166
|
+
throw unavailable("recovered eval plan identity changed");
|
|
167
|
+
}
|
|
168
|
+
if (!Array.isArray(plan.tasks) || plan.tasks.length === 0
|
|
169
|
+
|| plan.tasks.some((task) => typeof task !== "string" || task.length === 0)
|
|
170
|
+
|| new Set(plan.tasks).size !== plan.tasks.length)
|
|
171
|
+
throw unavailable("recovered eval task plan is invalid");
|
|
172
|
+
return { tasks: [...plan.tasks], attempts: plan.attempts };
|
|
173
|
+
}
|
|
174
|
+
async function writeRequest(directory, options, startedAt, tasks, trials) {
|
|
175
|
+
await atomicWriteJSON(path.join(directory, "request.json"), {
|
|
176
|
+
schema_version: SCHEMA_VERSION,
|
|
177
|
+
rerun_id: options.rerunId,
|
|
178
|
+
eval_id: options.evalId,
|
|
179
|
+
rerun_type: options.rerunType,
|
|
180
|
+
semantics: evalRerunSemantics(options.rerunType),
|
|
181
|
+
mode: options.selector.mode,
|
|
182
|
+
tasks: [...tasks],
|
|
183
|
+
trials: sortSlots(trials),
|
|
184
|
+
base_generation: 0,
|
|
185
|
+
created_at: startedAt,
|
|
186
|
+
});
|
|
187
|
+
}
|
|
188
|
+
async function writeState(file, options, startedAt, status, tasks, trials, repairedTasks, repairedTrials, terminal = {}) {
|
|
189
|
+
await atomicWriteJSON(file, {
|
|
190
|
+
schema_version: SCHEMA_VERSION,
|
|
191
|
+
rerun_id: options.rerunId,
|
|
192
|
+
eval_id: options.evalId,
|
|
193
|
+
rerun_type: options.rerunType,
|
|
194
|
+
semantics: evalRerunSemantics(options.rerunType),
|
|
195
|
+
status,
|
|
196
|
+
tasks: [...tasks],
|
|
197
|
+
trials: sortSlots(trials),
|
|
198
|
+
repaired_tasks: [...repairedTasks],
|
|
199
|
+
repaired_trials: sortSlots(repairedTrials),
|
|
200
|
+
started_at: startedAt,
|
|
201
|
+
...(terminal.completedAt ? { completed_at: terminal.completedAt } : {}),
|
|
202
|
+
...(terminal.evalStatus ? { eval_status: terminal.evalStatus } : {}),
|
|
203
|
+
...(terminal.remainingTasks ? { remaining_invalid_tasks: [...terminal.remainingTasks] } : {}),
|
|
204
|
+
...(terminal.remainingTrials ? { remaining_invalid_trials: sortSlots(terminal.remainingTrials) } : {}),
|
|
205
|
+
...(terminal.errorCode ? { error: { code: terminal.errorCode } } : {}),
|
|
206
|
+
});
|
|
207
|
+
}
|
|
208
|
+
async function exists(file) {
|
|
209
|
+
try {
|
|
210
|
+
await lstat(file);
|
|
211
|
+
return true;
|
|
212
|
+
}
|
|
213
|
+
catch (error) {
|
|
214
|
+
if (error.code === "ENOENT")
|
|
215
|
+
return false;
|
|
216
|
+
throw error;
|
|
217
|
+
}
|
|
218
|
+
}
|
|
219
|
+
function unavailable(message) {
|
|
220
|
+
return new HitchError(message, { code: "eval_rerun_unavailable", exitCode: 2 });
|
|
221
|
+
}
|
|
222
|
+
function defaultModelCapturePlan() {
|
|
223
|
+
return { requested_mode: "native", effective_mode: "native", required: false };
|
|
224
|
+
}
|
|
225
|
+
//# sourceMappingURL=preparation-rerun.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"preparation-rerun.js","sourceRoot":"","sources":["../../../src/evals/preparation-rerun.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,kBAAkB,CAAC;AAC9D,OAAO,IAAI,MAAM,WAAW,CAAC;AAE7B,OAAO,EAAE,UAAU,EAAE,cAAc,EAAE,eAAe,EAAE,SAAS,EAAE,QAAQ,EAAE,MAAM,wBAAwB,CAAC;AAC1G,OAAO,EAAE,uBAAuB,EAAE,mBAAmB,EAAE,MAAM,wBAAwB,CAAC;AACtF,OAAO,EAAE,gBAAgB,EAAE,MAAM,eAAe,CAAC;AACjD,OAAO,EAAE,0BAA0B,EAAE,cAAc,EAAE,MAAM,cAAc,CAAC;AAC1E,OAAO,EAAE,kBAAkB,EAAE,MAAM,kBAAkB,CAAC;AAEtD,OAAO,EAAE,iBAAiB,EAAE,OAAO,EAAE,SAAS,EAAE,WAAW,EAAE,oBAAoB,EAAE,MAAM,kBAAkB,CAAC;AAE5G,OAAO,EAAE,OAAO,EAAE,MAAM,cAAc,CAAC;AASvC,qFAAqF;AACrF,MAAM,CAAC,KAAK,UAAU,qBAAqB,CACzC,OAAwB,EACxB,OAAoB,EACpB,SAAiB,EACjB,cAAsB,EACtB,SAAiB;IAEjB,MAAM,iBAAiB,GAAG,CAAC,WAAW,EAAE,qBAAqB,EAAE,eAAe,CAAC,CAAC;IAChF,IAAI,CAAC,MAAM,OAAO,CAAC,GAAG,CAAC,iBAAiB,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC;QAAE,OAAO,IAAI,CAAC;IACrI,MAAM,YAAY,GAAG,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,aAAa,CAAC,EAAE,IAAI,CAAC,CAAC;IAC3H,IAAI,YAAY,KAAK,IAAI;QAAE,OAAO,IAAI,CAAC;IACvC,IAAI,OAAO,CAAC,SAAS,KAAK,mBAAmB,IAAI,OAAO,CAAC,QAAQ,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;QACrF,MAAM,WAAW,CAAC,4FAA4F,CAAC,CAAC;IAClH,CAAC;IACD,IAAI,YAAY,EAAE,MAAM,KAAK,WAAW,EAAE,CAAC;QACzC,MAAM,IAAI,UAAU,CAAC,gCAAgC,EAAE,EAAE,IAAI,EAAE,sBAAsB,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC,CAAC;IACxG,CAAC;IACD,IAAI,YAAY,EAAE,MAAM,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,YAAY,CAAC,MAAM,CAAC,IAAI,YAAY,CAAC,MAAM,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACjH,MAAM,WAAW,CAAC,qFAAqF,CAAC,CAAC;IAC3G,CAAC;IACD,MAAM,OAAO,GAAG,MAAM,0BAA0B,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IAClE,IAAI,OAAO,KAAK,IAAI,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC7C,MAAM,WAAW,CAAC,wEAAwE,CAAC,CAAC;IAC9F,CAAC;IAED,MAAM,WAAW,GAAG,MAAM,mBAAmB,CAAC,OAAO,CAAC,aAAa,EAAE,OAAO,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;QAC5G,MAAM,WAAW,CAAE,KAAe,CAAC,OAAO,CAAC,CAAC;IAC9C,CAAC,CAAC,CAAC;IACH,IAAI,gBAAgB,GAAG,uBAAuB,CAAC,WAAW,CAAC,IAAI,OAAO,CAAC,gBAAgB,CAAC;IACxF,IAAI,CAAC,gBAAgB,IAAI,CAAC,MAAM,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,iBAAiB,CAAC,CAAC,EAAE,CAAC;QAC5F,gBAAgB,GAAG,uBAAuB,EAAE,CAAC;IAC/C,CAAC;IACD,IAAI,CAAC,gBAAgB,EAAE,CAAC;QACtB,MAAM,WAAW,CAAC,0EAA0E,CAAC,CAAC;IAChG,CAAC;IACD,MAAM,YAAY,CAAC,cAAc,EAAE,OAAO,EAAE,SAAS,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;IAC/D,MAAM,UAAU,CAAC,SAAS,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;IAC3E,MAAM,wBAAwB,CAAC,OAAO,EAAE,cAAc,EAAE,YAAY,CAAC,CAAC;IAEtE,IAAI,CAAC;QACH,MAAM,SAAS,GAAG,MAAM,OAAO,CAAC;YAC9B,MAAM,EAAE,cAAc,CAAC,OAAO,CAAC,MAAM,CAAC;YACtC,IAAI,EAAE,OAAO,CAAC,IAAI;YAClB,OAAO;YACP,iBAAiB,EAAE,OAAO;YAC1B,UAAU,EAAE,IAAI;YAChB,eAAe,EAAE,IAAI;YACrB,GAAG,EAAE,OAAO,CAAC,GAAG,IAAI,OAAO,CAAC,GAAG;YAC/B,GAAG,CAAC,OAAO,CAAC,qBAAqB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,qBAAqB,EAAE,OAAO,CAAC,qBAAqB,EAAE,CAAC;YAChH,GAAG,CAAC,OAAO,CAAC,gBAAgB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,OAAO,CAAC,gBAAgB,EAAE,CAAC;YACjG,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACrD,GAAG,CAAC,OAAO,CAAC,kBAAkB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,OAAO,CAAC,kBAAkB,EAAE,CAAC;YACvG,GAAG,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,OAAO,CAAC,kBAAkB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACzF,GAAG,CAAC,OAAO,CAAC,uBAAuB,CAAC,CAAC,CAAC,EAAE,uBAAuB,EAAE,OAAO,CAAC,uBAAuB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACxG,iBAAiB,EAAE,OAAO,CAAC,iBAAiB;mBACvC,CAAC,WAAW,EAAE,iBAAiB,KAAK,sBAAsB,CAAC,CAAC,CAAC,qBAAqB,CAAC,CAAC,CAAC,uBAAuB,CAAC;YAClH,GAAG,CAAC,OAAO,CAAC,oBAAoB,CAAC,CAAC,CAAC,EAAE,oBAAoB,EAAE,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC/F,gBAAgB;YAChB,GAAG,CAAC,OAAO,CAAC,qBAAqB,CAAC,CAAC,CAAC,EAAE,qBAAqB,EAAE,OAAO,CAAC,qBAAqB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACnG,CAAC,CAAC;QACH,IAAI,SAAS,CAAC,MAAM,KAAK,WAAW,EAAE,CAAC;YACrC,MAAM,IAAI,UAAU,CAAC,wBAAwB,EAAE,EAAE,IAAI,EAAE,oBAAoB,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC,CAAC;QAC9F,CAAC;QACD,MAAM,CAAC,SAAS,EAAE,kBAAkB,EAAE,QAAQ,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;YAClE,QAAQ,CAAiB,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,WAAW,CAAC,EAAE,IAAI,CAAC;YAC7E,QAAQ,CAAiB,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,qBAAqB,CAAC,EAAE,IAAI,CAAC;YACvF,gBAAgB,CAAC,OAAO,CAAC,aAAa,CAAC;SACxC,CAAC,CAAC;QACH,IAAI,SAAS,KAAK,IAAI,IAAI,kBAAkB,KAAK,IAAI,IAAI,QAAQ,KAAK,IAAI;YAAE,MAAM,cAAc,CAAC,SAAS,CAAC,CAAC;QAE5G,MAAM,IAAI,GAAG,kBAAkB,CAAC,SAAS,EAAE,OAAO,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;QACpE,oBAAoB,CAAC,QAAQ,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC;QAC9D,MAAM,cAAc,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CACxE,EAAE,MAAM,EAAE,IAAI,CAAC,QAAQ,EAAE,EACzB,CAAC,CAAC,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,GAAG,CAAC,EAAE,CAAC,CACxD,CAAC,CAAC,CAAC;QACJ,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,QAAQ,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC;QAC5G,MAAM,cAAc,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QACnF,MAAM,eAAe,GAAG,iBAAiB,CAAC,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,QAAQ,EAAE,QAAQ,CAAC,CAAC;QAC/E,MAAM,aAAa,GAAG,WAAW,CAAC,cAAc,CAAC,CAAC;QAClD,MAAM,aAAa,GAAG,WAAW,CAAC,cAAc,CAAC,CAAC;QAClD,MAAM,cAAc,GAAG,WAAW,CAAC,eAAe,CAAC,CAAC;QACpD,MAAM,WAAW,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;QAC7C,MAAM,YAAY,CAAC,cAAc,EAAE,OAAO,EAAE,SAAS,EAAE,aAAa,EAAE,cAAc,CAAC,CAAC;QACtF,MAAM,MAAM,GAAoB;YAC9B,cAAc,EAAE,GAAG;YACnB,IAAI,EAAE,YAAY;YAClB,QAAQ,EAAE,OAAO,CAAC,OAAO;YACzB,UAAU,EAAE,OAAO,CAAC,SAAS;YAC7B,SAAS,EAAE,kBAAkB,CAAC,OAAO,CAAC,SAAS,CAAC;YAChD,OAAO,EAAE,OAAO,CAAC,MAAM;YACvB,MAAM,EAAE,WAAW;YACnB,cAAc,EAAE,aAAa;YAC7B,cAAc,EAAE,aAAa;YAC7B,uBAAuB,EAAE,cAAc;YACvC,eAAe,EAAE,cAAc;YAC/B,eAAe,EAAE,cAAc;YAC/B,wBAAwB,EAAE,eAAe;YACzC,WAAW,EAAE,SAAS,CAAC,MAAM,KAAK,WAAW,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ;YACtE,UAAU,EAAE,SAAS;YACrB,YAAY,EAAE,WAAW;SAC1B,CAAC;QACF,MAAM,UAAU,CAAC,SAAS,EAAE,OAAO,EAAE,SAAS,EAAE,WAAW,EAAE,aAAa,EAAE,cAAc,EAAE,aAAa,EAAE,cAAc,EAAE;YACzH,WAAW;YACX,UAAU,EAAE,MAAM,CAAC,WAAW;YAC9B,cAAc;YACd,eAAe;SAChB,CAAC,CAAC;QACH,OAAO,MAAM,CAAC;IAChB,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,MAAM,UAAU,CAAC,SAAS,EAAE,OAAO,EAAE,SAAS,EAAE,QAAQ,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE;YACxE,WAAW,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;YACrC,SAAS,EAAE,KAAK,YAAY,UAAU,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,mBAAmB;SAC1E,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;QACnB,MAAM,KAAK,CAAC;IACd,CAAC;AACH,CAAC;AAED,KAAK,UAAU,wBAAwB,CACrC,OAAwB,EACxB,cAAsB,EACtB,MAA+B;IAE/B,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,CAAC,cAAc,EAAE,kBAAkB,CAAC,CAAC;IAC9D,MAAM,SAAS,CAAC,OAAO,CAAC,CAAC;IACzB,KAAK,MAAM,IAAI,IAAI,CAAC,cAAc,EAAE,iBAAiB,EAAE,cAAc,EAAE,aAAa,CAAC,EAAE,CAAC;QACtF,MAAM,MAAM,GAAG,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,IAAI,CAAC,CAAC;QACtD,IAAI,MAAM,MAAM,CAAC,MAAM,CAAC;YAAE,MAAM,EAAE,CAAC,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,IAAI,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE,CAAC,CAAC;IAChH,CAAC;IACD,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,CAAC,cAAc,EAAE,iBAAiB,EAAE,cAAc,EAAE,aAAa,EAAE,QAAQ,CAAC,CAAC,CAAC;IACrG,KAAK,MAAM,KAAK,IAAI,MAAM,OAAO,CAAC,OAAO,CAAC,aAAa,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;QAClF,IAAI,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC;YAAE,SAAS;QACrC,MAAM,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC;IAC7F,CAAC;IACD,MAAM,OAAO,GAAG,MAAM,CAAC,KAAK,IAAI,OAAO,MAAM,CAAC,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC;QAC9F,CAAC,CAAC,MAAM,CAAC,KAAgC;QACzC,CAAC,CAAC,IAAI,CAAC;IACT,MAAM,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,eAAe,CAAC,EAAE;QACzD,cAAc,EAAE,cAAc;QAC9B,IAAI,EAAE,0BAA0B;QAChC,OAAO,EAAE,OAAO,CAAC,MAAM;QACvB,QAAQ,EAAE,OAAO,CAAC,OAAO;QACzB,GAAG,CAAC,OAAO,MAAM,CAAC,aAAa,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,aAAa,EAAE,MAAM,CAAC,aAAa,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC5F,GAAG,CAAC,OAAO,OAAO,EAAE,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC5E,WAAW,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;KACtC,CAAC,CAAC;AACL,CAAC;AAED,SAAS,cAAc,CAAC,MAA+B;IACrD,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,OAAO,MAAM,CAAC,KAAK,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC;QAC5F,CAAC,CAAC,MAAM,CAAC,KAAgC;QACzC,CAAC,CAAC,IAAI,CAAC;IACT,MAAM,IAAI,GAAG,OAAO,KAAK,EAAE,IAAI,KAAK,QAAQ,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,+BAA+B,CAAC;IAC1G,MAAM,KAAK,GAAG,OAAO,MAAM,CAAC,aAAa,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC,CAAC,CAAC,WAAW,CAAC;IAC5F,MAAM,OAAO,GAAG,OAAO,KAAK,EAAE,OAAO,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,iDAAiD,CAAC;IACxI,OAAO,IAAI,UAAU,CAAC,8BAA8B,KAAK,KAAK,OAAO,EAAE,EAAE;QACvE,IAAI;QACJ,QAAQ,EAAE,MAAM,CAAC,aAAa,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,SAAmB,CAAC,CAAC,CAAC,EAAE;KACnF,CAAC,CAAC;AACL,CAAC;AAED,SAAS,kBAAkB,CAAC,KAAc,EAAE,MAAc,EAAE,OAAoB;IAC9E,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,WAAW,CAAC,gCAAgC,CAAC,CAAC;IACrH,MAAM,IAAI,GAAG,KAAgC,CAAC;IAC9C,IAAI,IAAI,CAAC,cAAc,KAAK,cAAc,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC,OAAO,KAAK,OAAO,CAAC,OAAO;WACpG,IAAI,CAAC,YAAY,KAAK,OAAO,CAAC,YAAY,IAAI,IAAI,CAAC,kBAAkB,KAAK,OAAO,CAAC,kBAAkB;WACpG,CAAC,MAAM,CAAC,aAAa,CAAC,IAAI,CAAC,QAAQ,CAAC,IAAI,IAAI,CAAC,QAAQ,KAAK,OAAO,CAAC,QAAQ,EAAE,CAAC;QAChF,MAAM,WAAW,CAAC,sCAAsC,CAAC,CAAC;IAC5D,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC;WACpD,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC,CAAC;WACxE,IAAI,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,KAAK,CAAC,MAAM;QAAE,MAAM,WAAW,CAAC,qCAAqC,CAAC,CAAC;IAC9G,OAAO,EAAE,KAAK,EAAE,CAAC,GAAG,IAAI,CAAC,KAAiB,CAAC,EAAE,QAAQ,EAAE,IAAI,CAAC,QAAkB,EAAE,CAAC;AACnF,CAAC;AAED,KAAK,UAAU,YAAY,CACzB,SAAiB,EACjB,OAAwB,EACxB,SAAiB,EACjB,KAAwB,EACxB,MAAgC;IAEhC,MAAM,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,cAAc,CAAC,EAAE;QAC1D,cAAc,EAAE,cAAc;QAC9B,QAAQ,EAAE,OAAO,CAAC,OAAO;QACzB,OAAO,EAAE,OAAO,CAAC,MAAM;QACvB,UAAU,EAAE,OAAO,CAAC,SAAS;QAC7B,SAAS,EAAE,kBAAkB,CAAC,OAAO,CAAC,SAAS,CAAC;QAChD,IAAI,EAAE,OAAO,CAAC,QAAQ,CAAC,IAAI;QAC3B,KAAK,EAAE,CAAC,GAAG,KAAK,CAAC;QACjB,MAAM,EAAE,SAAS,CAAC,MAAM,CAAC;QACzB,eAAe,EAAE,CAAC;QAClB,UAAU,EAAE,SAAS;KACtB,CAAC,CAAC;AACL,CAAC;AAED,KAAK,UAAU,UAAU,CACvB,IAAY,EACZ,OAAwB,EACxB,SAAiB,EACjB,MAA0C,EAC1C,KAAwB,EACxB,MAAgC,EAChC,aAAgC,EAChC,cAAwC,EACxC,WAMI,EAAE;IAEN,MAAM,eAAe,CAAC,IAAI,EAAE;QAC1B,cAAc,EAAE,cAAc;QAC9B,QAAQ,EAAE,OAAO,CAAC,OAAO;QACzB,OAAO,EAAE,OAAO,CAAC,MAAM;QACvB,UAAU,EAAE,OAAO,CAAC,SAAS;QAC7B,SAAS,EAAE,kBAAkB,CAAC,OAAO,CAAC,SAAS,CAAC;QAChD,MAAM;QACN,KAAK,EAAE,CAAC,GAAG,KAAK,CAAC;QACjB,MAAM,EAAE,SAAS,CAAC,MAAM,CAAC;QACzB,cAAc,EAAE,CAAC,GAAG,aAAa,CAAC;QAClC,eAAe,EAAE,SAAS,CAAC,cAAc,CAAC;QAC1C,UAAU,EAAE,SAAS;QACrB,GAAG,CAAC,QAAQ,CAAC,WAAW,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,QAAQ,CAAC,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACvE,GAAG,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,WAAW,EAAE,QAAQ,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACpE,GAAG,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,uBAAuB,EAAE,CAAC,GAAG,QAAQ,CAAC,cAAc,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7F,GAAG,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC,CAAC,EAAE,wBAAwB,EAAE,SAAS,CAAC,QAAQ,CAAC,eAAe,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACtG,GAAG,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACvE,CAAC,CAAC;AACL,CAAC;AAED,KAAK,UAAU,MAAM,CAAC,IAAY;IAChC,IAAI,CAAC;QACH,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC;QAClB,OAAO,IAAI,CAAC;IACd,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAK,KAA+B,CAAC,IAAI,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QACrE,MAAM,KAAK,CAAC;IACd,CAAC;AACH,CAAC;AAED,SAAS,WAAW,CAAC,OAAe;IAClC,OAAO,IAAI,UAAU,CAAC,OAAO,EAAE,EAAE,IAAI,EAAE,wBAAwB,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC,CAAC;AAClF,CAAC;AAED,SAAS,uBAAuB;IAC9B,OAAO,EAAE,cAAc,EAAE,QAAQ,EAAE,cAAc,EAAE,QAAQ,EAAE,QAAQ,EAAE,KAAK,EAAE,CAAC;AACjF,CAAC"}
|
|
@@ -0,0 +1,76 @@
|
|
|
1
|
+
import { preparedArtifactDirectory } from "../artifacts/index.js";
|
|
2
|
+
import { HitchError } from "../foundation/index.js";
|
|
3
|
+
import { prepareHarness } from "../artifacts/index.js";
|
|
4
|
+
export function harborPreparedArtifact(root, artifact) {
|
|
5
|
+
if (!artifact.artifact_integrity || !artifact.entrypoint_integrity) {
|
|
6
|
+
throw new HitchError("host-prepared harness artifact has no complete integrity metadata", {
|
|
7
|
+
code: "artifact_integrity_mismatch",
|
|
8
|
+
exitCode: 5,
|
|
9
|
+
});
|
|
10
|
+
}
|
|
11
|
+
return {
|
|
12
|
+
directory: preparedArtifactDirectory(root, artifact.artifact_id),
|
|
13
|
+
artifact_id: artifact.artifact_id,
|
|
14
|
+
artifact_integrity: artifact.artifact_integrity,
|
|
15
|
+
entrypoint_integrity: artifact.entrypoint_integrity,
|
|
16
|
+
harness_id: artifact.harness_id,
|
|
17
|
+
revision_identity: artifact.revision_identity,
|
|
18
|
+
adapter_version: artifact.adapter_version,
|
|
19
|
+
recipe_version: artifact.recipe_version,
|
|
20
|
+
platform: artifact.platform,
|
|
21
|
+
node_version: artifact.toolchain.node || process.version,
|
|
22
|
+
source_type: artifact.source_type,
|
|
23
|
+
storage: "host-artifact-store-v1",
|
|
24
|
+
};
|
|
25
|
+
}
|
|
26
|
+
export function preparedHarnessEvent(artifact, cacheHit, source, builder) {
|
|
27
|
+
return {
|
|
28
|
+
type: "eval.harness-artifact.prepared",
|
|
29
|
+
harness: artifact.harness_id,
|
|
30
|
+
artifact_id: artifact.artifact_id,
|
|
31
|
+
platform: artifact.platform,
|
|
32
|
+
cache_hit: cacheHit,
|
|
33
|
+
source,
|
|
34
|
+
...(builder?.image ? { builder_image: builder.image } : {}),
|
|
35
|
+
...(builder?.imageId ? { builder_image_id: builder.imageId } : {}),
|
|
36
|
+
};
|
|
37
|
+
}
|
|
38
|
+
export async function prepareEvalHarborArtifact(input) {
|
|
39
|
+
const selector = input.requestedReference.selector;
|
|
40
|
+
const verifiedLocalGitSource = input.localTransport && selector.type === "commit" && selector.source?.explicit
|
|
41
|
+
? {
|
|
42
|
+
directory: selector.source.local_path,
|
|
43
|
+
commit: input.localTransport.manifest.commit,
|
|
44
|
+
tree: input.localTransport.manifest.tree,
|
|
45
|
+
resolutionIdentity: input.localTransport.manifest.resolution_identity,
|
|
46
|
+
payloadSha256: input.localTransport.manifest.payload_sha256,
|
|
47
|
+
}
|
|
48
|
+
: undefined;
|
|
49
|
+
return input.builder({
|
|
50
|
+
root: input.root,
|
|
51
|
+
resolvedRevision: input.resolvedRevision,
|
|
52
|
+
runtimeDirectory: input.controllerRuntime.directory,
|
|
53
|
+
runtimeId: input.controllerRuntime.runtime_id,
|
|
54
|
+
runtimeContract: input.runtimeContract,
|
|
55
|
+
...(input.localTransport ? { localTransport: input.localTransport } : {}),
|
|
56
|
+
...(verifiedLocalGitSource ? { verifiedLocalGitSource } : {}),
|
|
57
|
+
env: input.env,
|
|
58
|
+
...(input.signal ? { signal: input.signal } : {}),
|
|
59
|
+
});
|
|
60
|
+
}
|
|
61
|
+
/** In-process seam for tests whose fake package manager cannot be mounted into
|
|
62
|
+
* Docker. Production callers must use the dedicated builder. */
|
|
63
|
+
export const prepareHostHarborArtifactForTest = async (input) => {
|
|
64
|
+
const artifact = await prepareHarness(input.resolvedRevision, {
|
|
65
|
+
root: input.root,
|
|
66
|
+
env: input.env,
|
|
67
|
+
...(input.signal ? { signal: input.signal } : {}),
|
|
68
|
+
...(input.verifiedLocalGitSource ? { verifiedLocalGitSource: input.verifiedLocalGitSource } : {}),
|
|
69
|
+
});
|
|
70
|
+
return {
|
|
71
|
+
artifact: harborPreparedArtifact(input.root, artifact),
|
|
72
|
+
cacheHit: artifact.cache_hit,
|
|
73
|
+
source: "test-host",
|
|
74
|
+
};
|
|
75
|
+
};
|
|
76
|
+
//# sourceMappingURL=prepared-harness.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"prepared-harness.js","sourceRoot":"","sources":["../../../src/evals/prepared-harness.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,yBAAyB,EAAE,MAAM,uBAAuB,CAAC;AAIlE,OAAO,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AAEpD,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAKvD,MAAM,UAAU,sBAAsB,CAAC,IAAY,EAAE,QAA0B;IAC7E,IAAI,CAAC,QAAQ,CAAC,kBAAkB,IAAI,CAAC,QAAQ,CAAC,oBAAoB,EAAE,CAAC;QACnE,MAAM,IAAI,UAAU,CAAC,mEAAmE,EAAE;YACxF,IAAI,EAAE,6BAA6B;YACnC,QAAQ,EAAE,CAAC;SACZ,CAAC,CAAC;IACL,CAAC;IACD,OAAO;QACL,SAAS,EAAE,yBAAyB,CAAC,IAAI,EAAE,QAAQ,CAAC,WAAW,CAAC;QAChE,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,kBAAkB,EAAE,QAAQ,CAAC,kBAAkB;QAC/C,oBAAoB,EAAE,QAAQ,CAAC,oBAAoB;QACnD,UAAU,EAAE,QAAQ,CAAC,UAAU;QAC/B,iBAAiB,EAAE,QAAQ,CAAC,iBAAiB;QAC7C,eAAe,EAAE,QAAQ,CAAC,eAAe;QACzC,cAAc,EAAE,QAAQ,CAAC,cAAc;QACvC,QAAQ,EAAE,QAAQ,CAAC,QAAQ;QAC3B,YAAY,EAAE,QAAQ,CAAC,SAAS,CAAC,IAAI,IAAI,OAAO,CAAC,OAAO;QACxD,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,OAAO,EAAE,wBAAwB;KAClC,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,oBAAoB,CAClC,QAA2E,EAC3E,QAAiB,EACjB,MAAyC,EACzC,OAA8C;IAE9C,OAAO;QACL,IAAI,EAAE,gCAAgC;QACtC,OAAO,EAAE,QAAQ,CAAC,UAAU;QAC5B,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,QAAQ,EAAE,QAAQ,CAAC,QAAQ;QAC3B,SAAS,EAAE,QAAQ;QACnB,MAAM;QACN,GAAG,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC,EAAE,aAAa,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC3D,GAAG,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACnE,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,yBAAyB,CAAC,KAU/C;IACC,MAAM,QAAQ,GAAG,KAAK,CAAC,kBAAkB,CAAC,QAAQ,CAAC;IACnD,MAAM,sBAAsB,GAAG,KAAK,CAAC,cAAc,IAAI,QAAQ,CAAC,IAAI,KAAK,QAAQ,IAAI,QAAQ,CAAC,MAAM,EAAE,QAAQ;QAC5G,CAAC,CAAC;YACE,SAAS,EAAE,QAAQ,CAAC,MAAM,CAAC,UAAU;YACrC,MAAM,EAAE,KAAK,CAAC,cAAc,CAAC,QAAQ,CAAC,MAAM;YAC5C,IAAI,EAAE,KAAK,CAAC,cAAc,CAAC,QAAQ,CAAC,IAAI;YACxC,kBAAkB,EAAE,KAAK,CAAC,cAAc,CAAC,QAAQ,CAAC,mBAAmB;YACrE,aAAa,EAAE,KAAK,CAAC,cAAc,CAAC,QAAQ,CAAC,cAAc;SAC5D;QACH,CAAC,CAAC,SAAS,CAAC;IACd,OAAO,KAAK,CAAC,OAAO,CAAC;QACnB,IAAI,EAAE,KAAK,CAAC,IAAI;QAChB,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;QACxC,gBAAgB,EAAE,KAAK,CAAC,iBAAiB,CAAC,SAAS;QACnD,SAAS,EAAE,KAAK,CAAC,iBAAiB,CAAC,UAAU;QAC7C,eAAe,EAAE,KAAK,CAAC,eAAe;QACtC,GAAG,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,KAAK,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACzE,GAAG,CAAC,sBAAsB,CAAC,CAAC,CAAC,EAAE,sBAAsB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7D,GAAG,EAAE,KAAK,CAAC,GAAG;QACd,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAClD,CAAC,CAAC;AACL,CAAC;AAED;gEACgE;AAChE,MAAM,CAAC,MAAM,gCAAgC,GAA8B,KAAK,EAAE,KAAK,EAAE,EAAE;IACzF,MAAM,QAAQ,GAAG,MAAM,cAAc,CAAC,KAAK,CAAC,gBAAgB,EAAE;QAC5D,IAAI,EAAE,KAAK,CAAC,IAAI;QAChB,GAAG,EAAE,KAAK,CAAC,GAAG;QACd,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACjD,GAAG,CAAC,KAAK,CAAC,sBAAsB,CAAC,CAAC,CAAC,EAAE,sBAAsB,EAAE,KAAK,CAAC,sBAAsB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAClG,CAAC,CAAC;IACH,OAAO;QACL,QAAQ,EAAE,sBAAsB,CAAC,KAAK,CAAC,IAAI,EAAE,QAAQ,CAAC;QACtD,QAAQ,EAAE,QAAQ,CAAC,SAAS;QAC5B,MAAM,EAAE,WAAW;KACpB,CAAC;AACJ,CAAC,CAAC"}
|
|
@@ -32,15 +32,15 @@ export async function writeEvalProgress(evalDirectory, progress) {
|
|
|
32
32
|
await atomicWriteJSON(path.join(evalDirectory, "progress.json"), parseEvalProgress(progress));
|
|
33
33
|
}
|
|
34
34
|
export function mergeEvalProgressTrial(progress, trial, now = new Date().toISOString()) {
|
|
35
|
-
const parsed =
|
|
35
|
+
const parsed = parseEvalTrialRef(trial, "eval progress trial");
|
|
36
36
|
const byTrial = progress.trials.find((item) => item.trial_id === parsed.trial_id);
|
|
37
37
|
if (byTrial !== undefined) {
|
|
38
38
|
if (JSON.stringify(byTrial) !== JSON.stringify(parsed))
|
|
39
39
|
throw new TypeError(`eval progress trial identity conflict: ${parsed.trial_id}`);
|
|
40
40
|
return progress;
|
|
41
41
|
}
|
|
42
|
-
if (progress.trials.some((item) => item
|
|
43
|
-
throw new TypeError(`eval progress run identity conflict: ${parsed
|
|
42
|
+
if (progress.trials.some((item) => evalTrialCandidateKey(item) === evalTrialCandidateKey(parsed))) {
|
|
43
|
+
throw new TypeError(`eval progress run identity conflict: ${evalTrialCandidateKey(parsed)}`);
|
|
44
44
|
}
|
|
45
45
|
if (progress.trials.some((item) => evalTrialKey(item) === evalTrialKey(parsed))) {
|
|
46
46
|
throw new TypeError(`eval progress logical trial conflict: ${parsed.task_id} attempt ${parsed.attempt}`);
|
|
@@ -64,9 +64,12 @@ export function mergeEvalProgressTrial(progress, trial, now = new Date().toISOSt
|
|
|
64
64
|
export function evalTrialKey(trial) {
|
|
65
65
|
return `${trial.task_id}\u0000${trial.attempt}`;
|
|
66
66
|
}
|
|
67
|
+
export function evalTrialCandidateKey(trial) {
|
|
68
|
+
return trial.run_group ? trial.run_group.run_group_id : trial.run_id;
|
|
69
|
+
}
|
|
67
70
|
/** Replace one invalid logical trial, or fill a missing slot, with a valid verifier result. */
|
|
68
71
|
export function replaceInvalidEvalProgressTrial(progress, trial, now = new Date().toISOString()) {
|
|
69
|
-
const parsed =
|
|
72
|
+
const parsed = parseEvalTrialRef(trial, "eval rerun trial");
|
|
70
73
|
if (parsed.observation_status !== "valid")
|
|
71
74
|
throw new TypeError("eval rerun replacement must be valid");
|
|
72
75
|
const key = evalTrialKey(parsed);
|
|
@@ -76,8 +79,8 @@ export function replaceInvalidEvalProgressTrial(progress, trial, now = new Date(
|
|
|
76
79
|
return progress;
|
|
77
80
|
throw new TypeError(`eval rerun cannot replace valid task: ${parsed.task_id}`);
|
|
78
81
|
}
|
|
79
|
-
if (progress.trials.some((item) => item
|
|
80
|
-
throw new TypeError(`eval progress run identity conflict: ${parsed
|
|
82
|
+
if (progress.trials.some((item) => evalTrialCandidateKey(item) === evalTrialCandidateKey(parsed) && evalTrialKey(item) !== key)) {
|
|
83
|
+
throw new TypeError(`eval progress run identity conflict: ${evalTrialCandidateKey(parsed)}`);
|
|
81
84
|
}
|
|
82
85
|
if (progress.trials.some((item) => item.trial_id === parsed.trial_id && evalTrialKey(item) !== key)) {
|
|
83
86
|
throw new TypeError(`eval progress trial identity conflict: ${parsed.trial_id}`);
|
|
@@ -118,9 +121,9 @@ export function parseEvalProgress(value) {
|
|
|
118
121
|
}
|
|
119
122
|
if (!Array.isArray(record.trials))
|
|
120
123
|
throw new TypeError("eval progress trials are invalid");
|
|
121
|
-
const trials = record.trials.map((trial, index) =>
|
|
124
|
+
const trials = record.trials.map((trial, index) => parseEvalTrialRef(trial, `eval progress trial ${index}`));
|
|
122
125
|
if (new Set(trials.map((trial) => trial.trial_id)).size !== trials.length
|
|
123
|
-
|| new Set(trials.map(
|
|
126
|
+
|| new Set(trials.map(evalTrialCandidateKey)).size !== trials.length)
|
|
124
127
|
throw new TypeError("eval progress trial identities are duplicated");
|
|
125
128
|
const sorted = [...trials].sort((left, right) => left.task_id.localeCompare(right.task_id)
|
|
126
129
|
|| left.attempt - right.attempt
|
|
@@ -164,12 +167,11 @@ export function parseEvalProgress(value) {
|
|
|
164
167
|
updated_at: record.updated_at,
|
|
165
168
|
};
|
|
166
169
|
}
|
|
167
|
-
function
|
|
170
|
+
export function parseEvalTrialRef(value, label = "eval trial") {
|
|
168
171
|
if (!value || typeof value !== "object" || Array.isArray(value))
|
|
169
172
|
throw new TypeError(`${label} must be an object`);
|
|
170
173
|
const trial = value;
|
|
171
174
|
if (typeof trial.trial_id !== "string" || !trial.trial_id
|
|
172
|
-
|| typeof trial.run_id !== "string" || !RUN_ID.test(trial.run_id)
|
|
173
175
|
|| typeof trial.task_id !== "string" || !trial.task_id
|
|
174
176
|
|| !Number.isSafeInteger(trial.attempt) || trial.attempt < 1
|
|
175
177
|
|| (trial.observation_status !== "valid" && trial.observation_status !== "invalid")) {
|
|
@@ -178,21 +180,59 @@ function parseTrial(value, label) {
|
|
|
178
180
|
if (trial.observation_status === "valid" && (typeof trial.reward !== "number" || !Number.isFinite(trial.reward))) {
|
|
179
181
|
throw new TypeError(`${label} valid reward is invalid`);
|
|
180
182
|
}
|
|
183
|
+
const scores = trial.scores === undefined ? undefined : parseScores(trial.scores, label);
|
|
184
|
+
if (scores !== undefined && (trial.observation_status !== "valid" || scores.total_score !== trial.reward)) {
|
|
185
|
+
throw new TypeError(`${label} scores do not match the valid reward`);
|
|
186
|
+
}
|
|
181
187
|
if (trial.observation_status === "invalid" && (typeof trial.invalid_reason !== "string" || !trial.invalid_reason)) {
|
|
182
188
|
throw new TypeError(`${label} invalid reason is missing`);
|
|
183
189
|
}
|
|
184
190
|
if (trial.verifier_result_ref !== undefined && (typeof trial.verifier_result_ref !== "string" || !trial.verifier_result_ref)) {
|
|
185
191
|
throw new TypeError(`${label} verifier ref is invalid`);
|
|
186
192
|
}
|
|
187
|
-
|
|
193
|
+
const assessment = trial.assessment;
|
|
194
|
+
if (assessment !== undefined && (!assessment || typeof assessment.id !== "string" || !/^assessment_[a-f0-9]{32}$/.test(assessment.id)
|
|
195
|
+
|| typeof assessment.digest !== "string" || !/^sha256:[a-f0-9]{64}$/.test(assessment.digest)))
|
|
196
|
+
throw new TypeError(`${label} assessment is invalid`);
|
|
197
|
+
const group = trial.run_group;
|
|
198
|
+
if (group !== undefined) {
|
|
199
|
+
if (!group || typeof group !== "object" || Array.isArray(group) || Object.keys(group).some(key => !["run_group_id", "digest"].includes(key))
|
|
200
|
+
|| typeof group.run_group_id !== "string" || !/^run_group_[a-f0-9]{32}$/.test(group.run_group_id)
|
|
201
|
+
|| typeof group.digest !== "string" || !/^sha256:[a-f0-9]{64}$/.test(group.digest) || trial.run_id !== undefined || !assessment) {
|
|
202
|
+
throw new TypeError(`${label} phase group identity is invalid`);
|
|
203
|
+
}
|
|
204
|
+
}
|
|
205
|
+
else if (typeof trial.run_id !== "string" || !RUN_ID.test(trial.run_id))
|
|
206
|
+
throw new TypeError(`${label} run identity is invalid`);
|
|
207
|
+
const common = {
|
|
188
208
|
trial_id: trial.trial_id,
|
|
189
|
-
run_id: trial.run_id,
|
|
190
209
|
task_id: trial.task_id,
|
|
191
210
|
attempt: trial.attempt,
|
|
192
211
|
observation_status: trial.observation_status,
|
|
193
212
|
...(trial.reward === undefined ? {} : { reward: trial.reward }),
|
|
213
|
+
...(scores === undefined ? {} : { scores }),
|
|
194
214
|
...(trial.verifier_result_ref === undefined ? {} : { verifier_result_ref: trial.verifier_result_ref }),
|
|
195
215
|
...(trial.invalid_reason === undefined ? {} : { invalid_reason: trial.invalid_reason }),
|
|
196
216
|
};
|
|
217
|
+
const reference = assessment ? { id: assessment.id, digest: assessment.digest } : undefined;
|
|
218
|
+
return group ? { ...common, run_group: { run_group_id: group.run_group_id, digest: group.digest }, assessment: reference }
|
|
219
|
+
: { ...common, run_id: trial.run_id, ...(reference ? { assessment: reference } : {}) };
|
|
220
|
+
}
|
|
221
|
+
function parseScores(value, label) {
|
|
222
|
+
if (!value || typeof value !== "object" || Array.isArray(value))
|
|
223
|
+
throw new TypeError(`${label} scores are invalid`);
|
|
224
|
+
const record = value;
|
|
225
|
+
if (Object.keys(record).some((key) => !["total_score", "process_score", "normalization"].includes(key))
|
|
226
|
+
|| typeof record.total_score !== "number" || !Number.isFinite(record.total_score)
|
|
227
|
+
|| record.process_score !== undefined && (typeof record.process_score !== "number" || !Number.isFinite(record.process_score))
|
|
228
|
+
|| record.normalization !== "standard" && record.normalization !== "legacy-reward"
|
|
229
|
+
|| record.normalization === "legacy-reward" && record.process_score !== undefined) {
|
|
230
|
+
throw new TypeError(`${label} scores are invalid`);
|
|
231
|
+
}
|
|
232
|
+
return {
|
|
233
|
+
total_score: record.total_score,
|
|
234
|
+
...(record.process_score === undefined ? {} : { process_score: record.process_score }),
|
|
235
|
+
normalization: record.normalization,
|
|
236
|
+
};
|
|
197
237
|
}
|
|
198
238
|
//# sourceMappingURL=progress.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"progress.js","sourceRoot":"","sources":["../../../src/evals/progress.ts"],"names":[],"mappings":"AAAA,OAAO,IAAI,MAAM,WAAW,CAAC;AAE7B,OAAO,EAAE,eAAe,EAAE,QAAQ,EAAE,MAAM,wBAAwB,CAAC;AAEnE,MAAM,OAAO,GAAG,qBAAqB,CAAC;AACtC,MAAM,MAAM,GAAG,oBAAoB,CAAC;AAEpC,MAAM,UAAU,kBAAkB,CAAC,KAOlC;IACC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,kCAAkC,CAAC,CAAC;IACzF,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,KAAK,CAAC,iBAAiB;QAAE,MAAM,IAAI,SAAS,CAAC,6CAA6C,CAAC,CAAC;IACvH,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,qCAAqC,CAAC,CAAC;IAC9G,OAAO;QACL,cAAc,EAAE,GAAG;QACnB,OAAO,EAAE,KAAK,CAAC,MAAM;QACrB,YAAY,EAAE,KAAK,CAAC,WAAW;QAC/B,kBAAkB,EAAE,KAAK,CAAC,iBAAiB;QAC3C,MAAM,EAAE,SAAS;QACjB,UAAU,EAAE,CAAC;QACb,aAAa,EAAE,KAAK,CAAC,YAAY,IAAI,IAAI;QACzC,cAAc,EAAE,KAAK,CAAC,aAAa,IAAI,IAAI;QAC3C,MAAM,EAAE,EAAE;QACV,OAAO,EAAE,EAAE,cAAc,EAAE,CAAC,EAAE,YAAY,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,EAAE;QAClE,UAAU,EAAE,KAAK,CAAC,SAAS;QAC3B,UAAU,EAAE,KAAK,CAAC,SAAS;KAC5B,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,aAAqB;IAC1D,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAiB,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,eAAe,CAAC,EAAE,IAAI,CAAC,CAAC;IAC9F,OAAO,KAAK,KAAK,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC;AAC1D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,aAAqB,EAAE,QAAwB;IACrF,MAAM,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,eAAe,CAAC,EAAE,iBAAiB,CAAC,QAAQ,CAAC,CAAC,CAAC;AAChG,CAAC;AAED,MAAM,UAAU,sBAAsB,CAAC,QAAwB,EAAE,KAAqB,EAAE,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;IACpH,MAAM,MAAM,GAAG,UAAU,CAAC,KAAK,EAAE,qBAAqB,CAAC,CAAC;IACxD,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,KAAK,MAAM,CAAC,QAAQ,CAAC,CAAC;IAClF,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;QAC1B,IAAI,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC;YAAE,MAAM,IAAI,SAAS,CAAC,0CAA0C,MAAM,CAAC,QAAQ,EAAE,CAAC,CAAC;QACzI,OAAO,QAAQ,CAAC;IAClB,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,MAAM,KAAK,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;QAClE,MAAM,IAAI,SAAS,CAAC,wCAAwC,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;IAC/E,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,YAAY,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;QAChF,MAAM,IAAI,SAAS,CAAC,yCAAyC,MAAM,CAAC,OAAO,YAAY,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3G,CAAC;IACD,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC;WACtG,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO;WAC5B,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;IAClF,OAAO,iBAAiB,CAAC;QACvB,GAAG,QAAQ;QACX,UAAU,EAAE,QAAQ,CAAC,UAAU,GAAG,CAAC;QACnC,MAAM;QACN,OAAO,EAAE;YACP,cAAc,EAAE,MAAM,CAAC,MAAM;YAC7B,YAAY,EAAE,KAAK;YACnB,cAAc,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;SACtC;QACD,UAAU,EAAE,GAAG;KAChB,CAAC,CAAC;AACL,CAAC;AAED,MAAM,UAAU,YAAY,CAAC,KAAkD;IAC7E,OAAO,GAAG,KAAK,CAAC,OAAO,SAAS,KAAK,CAAC,OAAO,EAAE,CAAC;AAClD,CAAC;AAED,+FAA+F;AAC/F,MAAM,UAAU,+BAA+B,CAC7C,QAAwB,EACxB,KAAqB,EACrB,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;IAE9B,MAAM,MAAM,GAAG,UAAU,CAAC,KAAK,EAAE,kBAAkB,CAAC,CAAC;IACrD,IAAI,MAAM,CAAC,kBAAkB,KAAK,OAAO;QAAE,MAAM,IAAI,SAAS,CAAC,sCAAsC,CAAC,CAAC;IACvG,MAAM,GAAG,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC;IACjC,MAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC;IAC5E,IAAI,QAAQ,EAAE,kBAAkB,KAAK,OAAO,EAAE,CAAC;QAC7C,IAAI,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC;YAAE,OAAO,QAAQ,CAAC;QACzE,MAAM,IAAI,SAAS,CAAC,yCAAyC,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;IACjF,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,MAAM,KAAK,MAAM,CAAC,MAAM,IAAI,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC;QAChG,MAAM,IAAI,SAAS,CAAC,wCAAwC,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;IAC/E,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,KAAK,MAAM,CAAC,QAAQ,IAAI,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC;QACpG,MAAM,IAAI,SAAS,CAAC,0CAA0C,MAAM,CAAC,QAAQ,EAAE,CAAC,CAAC;IACnF,CAAC;IACD,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC;WACnJ,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO;WAC5B,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;IAClF,OAAO,iBAAiB,CAAC;QACvB,GAAG,QAAQ;QACX,UAAU,EAAE,QAAQ,CAAC,UAAU,GAAG,CAAC;QACnC,MAAM;QACN,OAAO,EAAE;YACP,cAAc,EAAE,MAAM,CAAC,MAAM;YAC7B,YAAY,EAAE,KAAK;YACnB,cAAc,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;SACtC;QACD,UAAU,EAAE,GAAG;KAChB,CAAC,CAAC;AACL,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,KAAc;IAC9C,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,iCAAiC,CAAC,CAAC;IACxH,MAAM,MAAM,GAAG,KAAgC,CAAC;IAChD,IAAI,MAAM,CAAC,cAAc,KAAK,GAAG,IAAI,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,EAAE,CAAC;QACzG,MAAM,IAAI,SAAS,CAAC,mCAAmC,CAAC,CAAC;IAC3D,CAAC;IACD,IAAI,OAAO,MAAM,CAAC,YAAY,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,YAAY;WAC9D,OAAO,MAAM,CAAC,kBAAkB,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,kBAAkB;WAC3E,MAAM,CAAC,MAAM,KAAK,SAAS;QAAE,MAAM,IAAI,SAAS,CAAC,2CAA2C,CAAC,CAAC;IACnG,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,MAAM,CAAC,UAAU,CAAC,IAAK,MAAM,CAAC,UAAqB,GAAG,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,qCAAqC,CAAC,CAAC;IAC9I,KAAK,MAAM,IAAI,IAAI,CAAC,eAAe,EAAE,gBAAgB,CAAU,EAAE,CAAC;QAChE,MAAM,OAAO,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC;QAC7B,IAAI,OAAO,KAAK,IAAI,IAAI,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC,OAAO,CAAC,IAAK,OAAkB,GAAG,CAAC,CAAC;YAAE,MAAM,IAAI,SAAS,CAAC,iBAAiB,IAAI,aAAa,CAAC,CAAC;IAC/I,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,kCAAkC,CAAC,CAAC;IAC3F,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,uBAAuB,KAAK,EAAE,CAAC,CAAC,CAAC;IACtG,IAAI,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,MAAM;WACpE,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,MAAM;QAAE,MAAM,IAAI,SAAS,CAAC,+CAA+C,CAAC,CAAC;IAC/I,MAAM,MAAM,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC;WACrF,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO;WAC5B,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,MAAM,CAAC,KAAK,CAAC,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,iDAAiD,CAAC,CAAC;IACnI,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;IAC/B,IAAI,CAAC,OAAO,IAAI,OAAO,OAAO,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,kCAAkC,CAAC,CAAC;IAC/H,MAAM,aAAa,GAAG,OAAkC,CAAC;IACzD,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;IACpF,IAAI,MAAM,CAAC,cAAc,KAAK,IAAI,IAAI,MAAM,CAAC,MAAM,GAAI,MAAM,CAAC,cAAyB,EAAE,CAAC;QACxF,MAAM,IAAI,SAAS,CAAC,oDAAoD,CAAC,CAAC;IAC5E,CAAC;IACD,IAAI,MAAM,CAAC,aAAa,KAAK,IAAI,IAAI,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,GAAI,MAAM,CAAC,aAAwB,EAAE,CAAC;QAC3H,MAAM,IAAI,SAAS,CAAC,mDAAmD,CAAC,CAAC;IAC3E,CAAC;IACD,IAAI,aAAa,CAAC,cAAc,KAAK,MAAM,CAAC,MAAM,IAAI,aAAa,CAAC,YAAY,KAAK,KAAK;WACrF,aAAa,CAAC,cAAc,KAAK,MAAM,CAAC,MAAM,GAAG,KAAK;QAAE,MAAM,IAAI,SAAS,CAAC,6CAA6C,CAAC,CAAC;IAChI,IAAI,OAAO,MAAM,CAAC,UAAU,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC;WACvF,OAAO,MAAM,CAAC,UAAU,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC;QAC9F,MAAM,IAAI,SAAS,CAAC,sCAAsC,CAAC,CAAC;IAC9D,CAAC;IACD,OAAO;QACL,cAAc,EAAE,GAAG;QACnB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,YAAY,EAAE,MAAM,CAAC,YAAY;QACjC,kBAAkB,EAAE,MAAM,CAAC,kBAAkB;QAC7C,MAAM,EAAE,SAAS;QACjB,UAAU,EAAE,MAAM,CAAC,UAAoB;QACvC,aAAa,EAAE,MAAM,CAAC,aAA8B;QACpD,cAAc,EAAE,MAAM,CAAC,cAA+B;QACtD,MAAM;QACN,OAAO,EAAE;YACP,cAAc,EAAE,MAAM,CAAC,MAAM;YAC7B,YAAY,EAAE,KAAK;YACnB,cAAc,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;SACtC;QACD,UAAU,EAAE,MAAM,CAAC,UAAU;QAC7B,UAAU,EAAE,MAAM,CAAC,UAAU;KAC9B,CAAC;AACJ,CAAC;AAED,SAAS,UAAU,CAAC,KAAc,EAAE,KAAa;IAC/C,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,oBAAoB,CAAC,CAAC;IACnH,MAAM,KAAK,GAAG,KAAgC,CAAC;IAC/C,IAAI,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,QAAQ;WACpD,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;WAC9D,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO;WACnD,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC,IAAK,KAAK,CAAC,OAAkB,GAAG,CAAC;WACrE,CAAC,KAAK,CAAC,kBAAkB,KAAK,OAAO,IAAI,KAAK,CAAC,kBAAkB,KAAK,SAAS,CAAC,EAAE,CAAC;QACtF,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,sBAAsB,CAAC,CAAC;IACtD,CAAC;IACD,IAAI,KAAK,CAAC,kBAAkB,KAAK,OAAO,IAAI,CAAC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;QACjH,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,0BAA0B,CAAC,CAAC;IAC1D,CAAC;IACD,IAAI,KAAK,CAAC,kBAAkB,KAAK,SAAS,IAAI,CAAC,OAAO,KAAK,CAAC,cAAc,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,cAAc,CAAC,EAAE,CAAC;QAClH,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,4BAA4B,CAAC,CAAC;IAC5D,CAAC;IACD,IAAI,KAAK,CAAC,mBAAmB,KAAK,SAAS,IAAI,CAAC,OAAO,KAAK,CAAC,mBAAmB,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,mBAAmB,CAAC,EAAE,CAAC;QAC7H,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,0BAA0B,CAAC,CAAC;IAC1D,CAAC;IACD,OAAO;QACL,QAAQ,EAAE,KAAK,CAAC,QAAQ;QACxB,MAAM,EAAE,KAAK,CAAC,MAAM;QACpB,OAAO,EAAE,KAAK,CAAC,OAAO;QACtB,OAAO,EAAE,KAAK,CAAC,OAAiB;QAChC,kBAAkB,EAAE,KAAK,CAAC,kBAAkB;QAC5C,GAAG,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC,MAAgB,EAAE,CAAC;QACzE,GAAG,CAAC,KAAK,CAAC,mBAAmB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,mBAAmB,EAAE,KAAK,CAAC,mBAA6B,EAAE,CAAC;QAChH,GAAG,CAAC,KAAK,CAAC,cAAc,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,KAAK,CAAC,cAAwB,EAAE,CAAC;KAClG,CAAC;AACJ,CAAC"}
|
|
1
|
+
{"version":3,"file":"progress.js","sourceRoot":"","sources":["../../../src/evals/progress.ts"],"names":[],"mappings":"AAAA,OAAO,IAAI,MAAM,WAAW,CAAC;AAE7B,OAAO,EAAE,eAAe,EAAE,QAAQ,EAAE,MAAM,wBAAwB,CAAC;AAEnE,MAAM,OAAO,GAAG,qBAAqB,CAAC;AACtC,MAAM,MAAM,GAAG,oBAAoB,CAAC;AAEpC,MAAM,UAAU,kBAAkB,CAAC,KAOlC;IACC,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,kCAAkC,CAAC,CAAC;IACzF,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,KAAK,CAAC,iBAAiB;QAAE,MAAM,IAAI,SAAS,CAAC,6CAA6C,CAAC,CAAC;IACvH,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,qCAAqC,CAAC,CAAC;IAC9G,OAAO;QACL,cAAc,EAAE,GAAG;QACnB,OAAO,EAAE,KAAK,CAAC,MAAM;QACrB,YAAY,EAAE,KAAK,CAAC,WAAW;QAC/B,kBAAkB,EAAE,KAAK,CAAC,iBAAiB;QAC3C,MAAM,EAAE,SAAS;QACjB,UAAU,EAAE,CAAC;QACb,aAAa,EAAE,KAAK,CAAC,YAAY,IAAI,IAAI;QACzC,cAAc,EAAE,KAAK,CAAC,aAAa,IAAI,IAAI;QAC3C,MAAM,EAAE,EAAE;QACV,OAAO,EAAE,EAAE,cAAc,EAAE,CAAC,EAAE,YAAY,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,EAAE;QAClE,UAAU,EAAE,KAAK,CAAC,SAAS;QAC3B,UAAU,EAAE,KAAK,CAAC,SAAS;KAC5B,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,aAAqB;IAC1D,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAiB,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,eAAe,CAAC,EAAE,IAAI,CAAC,CAAC;IAC9F,OAAO,KAAK,KAAK,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC;AAC1D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,aAAqB,EAAE,QAAwB;IACrF,MAAM,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,eAAe,CAAC,EAAE,iBAAiB,CAAC,QAAQ,CAAC,CAAC,CAAC;AAChG,CAAC;AAED,MAAM,UAAU,sBAAsB,CAAC,QAAwB,EAAE,KAAqB,EAAE,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;IACpH,MAAM,MAAM,GAAG,iBAAiB,CAAC,KAAK,EAAE,qBAAqB,CAAC,CAAC;IAC/D,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,KAAK,MAAM,CAAC,QAAQ,CAAC,CAAC;IAClF,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;QAC1B,IAAI,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC;YAAE,MAAM,IAAI,SAAS,CAAC,0CAA0C,MAAM,CAAC,QAAQ,EAAE,CAAC,CAAC;QACzI,OAAO,QAAQ,CAAC;IAClB,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,qBAAqB,CAAC,IAAI,CAAC,KAAK,qBAAqB,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;QAClG,MAAM,IAAI,SAAS,CAAC,wCAAwC,qBAAqB,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;IAC/F,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,YAAY,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;QAChF,MAAM,IAAI,SAAS,CAAC,yCAAyC,MAAM,CAAC,OAAO,YAAY,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3G,CAAC;IACD,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC;WACtG,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO;WAC5B,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;IAClF,OAAO,iBAAiB,CAAC;QACvB,GAAG,QAAQ;QACX,UAAU,EAAE,QAAQ,CAAC,UAAU,GAAG,CAAC;QACnC,MAAM;QACN,OAAO,EAAE;YACP,cAAc,EAAE,MAAM,CAAC,MAAM;YAC7B,YAAY,EAAE,KAAK;YACnB,cAAc,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;SACtC;QACD,UAAU,EAAE,GAAG;KAChB,CAAC,CAAC;AACL,CAAC;AAED,MAAM,UAAU,YAAY,CAAC,KAAkD;IAC7E,OAAO,GAAG,KAAK,CAAC,OAAO,SAAS,KAAK,CAAC,OAAO,EAAE,CAAC;AAClD,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,KAAqB;IACzD,OAAO,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,CAAC,SAAS,CAAC,YAAY,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC;AACvE,CAAC;AAED,+FAA+F;AAC/F,MAAM,UAAU,+BAA+B,CAC7C,QAAwB,EACxB,KAAqB,EACrB,GAAG,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;IAE9B,MAAM,MAAM,GAAG,iBAAiB,CAAC,KAAK,EAAE,kBAAkB,CAAC,CAAC;IAC5D,IAAI,MAAM,CAAC,kBAAkB,KAAK,OAAO;QAAE,MAAM,IAAI,SAAS,CAAC,sCAAsC,CAAC,CAAC;IACvG,MAAM,GAAG,GAAG,YAAY,CAAC,MAAM,CAAC,CAAC;IACjC,MAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC;IAC5E,IAAI,QAAQ,EAAE,kBAAkB,KAAK,OAAO,EAAE,CAAC;QAC7C,IAAI,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC;YAAE,OAAO,QAAQ,CAAC;QACzE,MAAM,IAAI,SAAS,CAAC,yCAAyC,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC;IACjF,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,qBAAqB,CAAC,IAAI,CAAC,KAAK,qBAAqB,CAAC,MAAM,CAAC,IAAI,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC;QAChI,MAAM,IAAI,SAAS,CAAC,wCAAwC,qBAAqB,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;IAC/F,CAAC;IACD,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,KAAK,MAAM,CAAC,QAAQ,IAAI,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC;QACpG,MAAM,IAAI,SAAS,CAAC,0CAA0C,MAAM,CAAC,QAAQ,EAAE,CAAC,CAAC;IACnF,CAAC;IACD,MAAM,MAAM,GAAG,CAAC,GAAG,QAAQ,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC;WACnJ,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO;WAC5B,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;IAClF,OAAO,iBAAiB,CAAC;QACvB,GAAG,QAAQ;QACX,UAAU,EAAE,QAAQ,CAAC,UAAU,GAAG,CAAC;QACnC,MAAM;QACN,OAAO,EAAE;YACP,cAAc,EAAE,MAAM,CAAC,MAAM;YAC7B,YAAY,EAAE,KAAK;YACnB,cAAc,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;SACtC;QACD,UAAU,EAAE,GAAG;KAChB,CAAC,CAAC;AACL,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,KAAc;IAC9C,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,iCAAiC,CAAC,CAAC;IACxH,MAAM,MAAM,GAAG,KAAgC,CAAC;IAChD,IAAI,MAAM,CAAC,cAAc,KAAK,GAAG,IAAI,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,EAAE,CAAC;QACzG,MAAM,IAAI,SAAS,CAAC,mCAAmC,CAAC,CAAC;IAC3D,CAAC;IACD,IAAI,OAAO,MAAM,CAAC,YAAY,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,YAAY;WAC9D,OAAO,MAAM,CAAC,kBAAkB,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,kBAAkB;WAC3E,MAAM,CAAC,MAAM,KAAK,SAAS;QAAE,MAAM,IAAI,SAAS,CAAC,2CAA2C,CAAC,CAAC;IACnG,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,MAAM,CAAC,UAAU,CAAC,IAAK,MAAM,CAAC,UAAqB,GAAG,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,qCAAqC,CAAC,CAAC;IAC9I,KAAK,MAAM,IAAI,IAAI,CAAC,eAAe,EAAE,gBAAgB,CAAU,EAAE,CAAC;QAChE,MAAM,OAAO,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC;QAC7B,IAAI,OAAO,KAAK,IAAI,IAAI,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC,OAAO,CAAC,IAAK,OAAkB,GAAG,CAAC,CAAC;YAAE,MAAM,IAAI,SAAS,CAAC,iBAAiB,IAAI,aAAa,CAAC,CAAC;IAC/I,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,kCAAkC,CAAC,CAAC;IAC3F,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,EAAE,CAAC,iBAAiB,CAAC,KAAK,EAAE,uBAAuB,KAAK,EAAE,CAAC,CAAC,CAAC;IAC7G,IAAI,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,MAAM;WACpE,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,qBAAqB,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,MAAM;QAAE,MAAM,IAAI,SAAS,CAAC,+CAA+C,CAAC,CAAC;IAC7I,MAAM,MAAM,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC;WACrF,IAAI,CAAC,OAAO,GAAG,KAAK,CAAC,OAAO;WAC5B,IAAI,CAAC,QAAQ,CAAC,aAAa,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC;IAClD,IAAI,MAAM,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,MAAM,CAAC,KAAK,CAAC,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,iDAAiD,CAAC,CAAC;IACnI,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;IAC/B,IAAI,CAAC,OAAO,IAAI,OAAO,OAAO,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,kCAAkC,CAAC,CAAC;IAC/H,MAAM,aAAa,GAAG,OAAkC,CAAC;IACzD,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;IACpF,IAAI,MAAM,CAAC,cAAc,KAAK,IAAI,IAAI,MAAM,CAAC,MAAM,GAAI,MAAM,CAAC,cAAyB,EAAE,CAAC;QACxF,MAAM,IAAI,SAAS,CAAC,oDAAoD,CAAC,CAAC;IAC5E,CAAC;IACD,IAAI,MAAM,CAAC,aAAa,KAAK,IAAI,IAAI,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,GAAI,MAAM,CAAC,aAAwB,EAAE,CAAC;QAC3H,MAAM,IAAI,SAAS,CAAC,mDAAmD,CAAC,CAAC;IAC3E,CAAC;IACD,IAAI,aAAa,CAAC,cAAc,KAAK,MAAM,CAAC,MAAM,IAAI,aAAa,CAAC,YAAY,KAAK,KAAK;WACrF,aAAa,CAAC,cAAc,KAAK,MAAM,CAAC,MAAM,GAAG,KAAK;QAAE,MAAM,IAAI,SAAS,CAAC,6CAA6C,CAAC,CAAC;IAChI,IAAI,OAAO,MAAM,CAAC,UAAU,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC;WACvF,OAAO,MAAM,CAAC,UAAU,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC;QAC9F,MAAM,IAAI,SAAS,CAAC,sCAAsC,CAAC,CAAC;IAC9D,CAAC;IACD,OAAO;QACL,cAAc,EAAE,GAAG;QACnB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,YAAY,EAAE,MAAM,CAAC,YAAY;QACjC,kBAAkB,EAAE,MAAM,CAAC,kBAAkB;QAC7C,MAAM,EAAE,SAAS;QACjB,UAAU,EAAE,MAAM,CAAC,UAAoB;QACvC,aAAa,EAAE,MAAM,CAAC,aAA8B;QACpD,cAAc,EAAE,MAAM,CAAC,cAA+B;QACtD,MAAM;QACN,OAAO,EAAE;YACP,cAAc,EAAE,MAAM,CAAC,MAAM;YAC7B,YAAY,EAAE,KAAK;YACnB,cAAc,EAAE,MAAM,CAAC,MAAM,GAAG,KAAK;SACtC;QACD,UAAU,EAAE,MAAM,CAAC,UAAU;QAC7B,UAAU,EAAE,MAAM,CAAC,UAAU;KAC9B,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,KAAc,EAAE,KAAK,GAAG,YAAY;IACpE,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,oBAAoB,CAAC,CAAC;IACnH,MAAM,KAAK,GAAG,KAAgC,CAAC;IAC/C,IAAI,OAAO,KAAK,CAAC,QAAQ,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,QAAQ;WACpD,OAAO,KAAK,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,OAAO;WACnD,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,OAAO,CAAC,IAAK,KAAK,CAAC,OAAkB,GAAG,CAAC;WACrE,CAAC,KAAK,CAAC,kBAAkB,KAAK,OAAO,IAAI,KAAK,CAAC,kBAAkB,KAAK,SAAS,CAAC,EAAE,CAAC;QACtF,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,sBAAsB,CAAC,CAAC;IACtD,CAAC;IACD,IAAI,KAAK,CAAC,kBAAkB,KAAK,OAAO,IAAI,CAAC,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,EAAE,CAAC;QACjH,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,0BAA0B,CAAC,CAAC;IAC1D,CAAC;IACD,MAAM,MAAM,GAAG,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC;IACzF,IAAI,MAAM,KAAK,SAAS,IAAI,CAAC,KAAK,CAAC,kBAAkB,KAAK,OAAO,IAAI,MAAM,CAAC,WAAW,KAAK,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC;QAC1G,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,uCAAuC,CAAC,CAAC;IACvE,CAAC;IACD,IAAI,KAAK,CAAC,kBAAkB,KAAK,SAAS,IAAI,CAAC,OAAO,KAAK,CAAC,cAAc,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,cAAc,CAAC,EAAE,CAAC;QAClH,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,4BAA4B,CAAC,CAAC;IAC5D,CAAC;IACD,IAAI,KAAK,CAAC,mBAAmB,KAAK,SAAS,IAAI,CAAC,OAAO,KAAK,CAAC,mBAAmB,KAAK,QAAQ,IAAI,CAAC,KAAK,CAAC,mBAAmB,CAAC,EAAE,CAAC;QAC7H,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,0BAA0B,CAAC,CAAC;IAC1D,CAAC;IACD,MAAM,UAAU,GAAG,KAAK,CAAC,UAA4D,CAAC;IACtF,IAAI,UAAU,KAAK,SAAS,IAAI,CAAC,CAAC,UAAU,IAAI,OAAO,UAAU,CAAC,EAAE,KAAK,QAAQ,IAAI,CAAC,2BAA2B,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE,CAAC;WAChI,OAAO,UAAU,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,wBAAwB,CAAC,CAAC;IACvJ,MAAM,KAAK,GAAG,KAAK,CAAC,SAAqE,CAAC;IAC1F,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACxB,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,cAAc,EAAE,QAAQ,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC;eACvI,OAAO,KAAK,CAAC,YAAY,KAAK,QAAQ,IAAI,CAAC,0BAA0B,CAAC,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC;eAC9F,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS,IAAI,CAAC,UAAU,EAAE,CAAC;YAClI,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,kCAAkC,CAAC,CAAC;QAClE,CAAC;IACH,CAAC;SAAM,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,0BAA0B,CAAC,CAAC;IACnI,MAAM,MAAM,GAAG;QACb,QAAQ,EAAE,KAAK,CAAC,QAAQ;QACxB,OAAO,EAAE,KAAK,CAAC,OAAO;QACtB,OAAO,EAAE,KAAK,CAAC,OAAiB;QAChC,kBAAkB,EAAE,KAAK,CAAC,kBAAyC;QACnE,GAAG,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC,MAAgB,EAAE,CAAC;QACzE,GAAG,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC;QAC3C,GAAG,CAAC,KAAK,CAAC,mBAAmB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,mBAAmB,EAAE,KAAK,CAAC,mBAA6B,EAAE,CAAC;QAChH,GAAG,CAAC,KAAK,CAAC,cAAc,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,KAAK,CAAC,cAAwB,EAAE,CAAC;KAClG,CAAC;IACF,MAAM,SAAS,GAAG,UAAU,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,UAAU,CAAC,EAAY,EAAE,MAAM,EAAE,UAAU,CAAC,MAAgB,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IAChH,OAAO,KAAK,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,SAAS,EAAE,EAAE,YAAY,EAAE,KAAK,CAAC,YAAsB,EAAE,MAAM,EAAE,KAAK,CAAC,MAA4B,EAAE,EAAE,UAAU,EAAE,SAAU,EAAE;QACzJ,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,KAAK,CAAC,MAAgB,EAAE,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;AACrG,CAAC;AAED,SAAS,WAAW,CAAC,KAAc,EAAE,KAAa;IAChD,IAAI,CAAC,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,qBAAqB,CAAC,CAAC;IACpH,MAAM,MAAM,GAAG,KAAgC,CAAC;IAChD,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,EAAE,eAAe,EAAE,eAAe,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC;WAClG,OAAO,MAAM,CAAC,WAAW,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,WAAW,CAAC;WAC9E,MAAM,CAAC,aAAa,KAAK,SAAS,IAAI,CAAC,OAAO,MAAM,CAAC,aAAa,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,aAAa,CAAC,CAAC;WAC1H,MAAM,CAAC,aAAa,KAAK,UAAU,IAAI,MAAM,CAAC,aAAa,KAAK,eAAe;WAC/E,MAAM,CAAC,aAAa,KAAK,eAAe,IAAI,MAAM,CAAC,aAAa,KAAK,SAAS,EAAE,CAAC;QACpF,MAAM,IAAI,SAAS,CAAC,GAAG,KAAK,qBAAqB,CAAC,CAAC;IACrD,CAAC;IACD,OAAO;QACL,WAAW,EAAE,MAAM,CAAC,WAAW;QAC/B,GAAG,CAAC,MAAM,CAAC,aAAa,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,aAAa,EAAE,MAAM,CAAC,aAAuB,EAAE,CAAC;QAChG,aAAa,EAAE,MAAM,CAAC,aAAa;KACpC,CAAC;AACJ,CAAC"}
|
|
@@ -1,7 +1,8 @@
|
|
|
1
|
-
import { readdir } from "node:fs/promises";
|
|
1
|
+
import { lstat, readdir } from "node:fs/promises";
|
|
2
2
|
import path from "node:path";
|
|
3
3
|
import { inspectEvalRuntimeKind } from "../controller-runtime/index.js";
|
|
4
4
|
import { HitchError, SCHEMA_VERSION, invalidInput, readJSON, statePaths } from "../foundation/index.js";
|
|
5
|
+
import { readExecutionLeases } from "./execution-leases.js";
|
|
5
6
|
export async function listEvals({ root }) {
|
|
6
7
|
const directory = statePaths(root).evals;
|
|
7
8
|
let entries;
|
|
@@ -19,16 +20,17 @@ export async function listEvals({ root }) {
|
|
|
19
20
|
continue;
|
|
20
21
|
const result = await readJSON(path.join(directory, entry.name, "result.json"), null).catch(() => null);
|
|
21
22
|
const request = await readJSON(path.join(directory, entry.name, "request.json"), null).catch(() => null);
|
|
23
|
+
const control = await readJSON(path.join(directory, entry.name, "control.json"), null).catch(() => null);
|
|
22
24
|
if (!result && !request)
|
|
23
25
|
continue;
|
|
24
26
|
evals.push({
|
|
25
27
|
eval_id: entry.name,
|
|
26
|
-
status: result?.status || "running",
|
|
28
|
+
status: result?.status || control?.state || "running",
|
|
27
29
|
backend: result?.backend?.name || request?.backend || null,
|
|
28
30
|
dataset: result?.dataset || request?.dataset || null,
|
|
29
31
|
harness_ref: result?.candidate?.harness_ref || request?.harness_ref || null,
|
|
30
32
|
primary_reward: result?.summary?.primary_reward ?? null,
|
|
31
|
-
started_at: result?.started_at || null,
|
|
33
|
+
started_at: result?.started_at || control?.created_at || null,
|
|
32
34
|
completed_at: result?.completed_at || null,
|
|
33
35
|
});
|
|
34
36
|
}
|
|
@@ -48,8 +50,23 @@ export async function inspectEval(evalId, { root }) {
|
|
|
48
50
|
request,
|
|
49
51
|
resolution: await readJSON(path.join(directory, "resolution.json"), null),
|
|
50
52
|
plan: await readJSON(path.join(directory, "plan.json"), null),
|
|
53
|
+
execution_plan: await readJSON(path.join(directory, "execution-plan.json"), null),
|
|
54
|
+
submission: await readJSON(path.join(directory, "submission.json"), null),
|
|
55
|
+
control: await readJSON(path.join(directory, "control.json"), null),
|
|
56
|
+
progress: await readJSON(path.join(directory, "progress.json"), null),
|
|
57
|
+
leases: await readExecutionLeases(directory),
|
|
51
58
|
result: await readJSON(path.join(directory, "result.json"), null),
|
|
52
59
|
runtime_storage: await inspectEvalRuntimeKind(directory),
|
|
53
60
|
};
|
|
54
61
|
}
|
|
62
|
+
export async function isControlPlaneEval(root, evalId) {
|
|
63
|
+
try {
|
|
64
|
+
return (await lstat(path.join(statePaths(root).evals, evalId, "submission.json"))).isFile();
|
|
65
|
+
}
|
|
66
|
+
catch (error) {
|
|
67
|
+
if (error.code === "ENOENT")
|
|
68
|
+
return false;
|
|
69
|
+
throw error;
|
|
70
|
+
}
|
|
71
|
+
}
|
|
55
72
|
//# sourceMappingURL=records.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"records.js","sourceRoot":"","sources":["../../../src/evals/records.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,OAAO,EAAE,MAAM,kBAAkB,CAAC;
|
|
1
|
+
{"version":3,"file":"records.js","sourceRoot":"","sources":["../../../src/evals/records.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,kBAAkB,CAAC;AAClD,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,EAAE,sBAAsB,EAAE,MAAM,gCAAgC,CAAC;AAExE,OAAO,EAAE,UAAU,EAAE,cAAc,EAAE,YAAY,EAAE,QAAQ,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AACxG,OAAO,EAAE,mBAAmB,EAAE,MAAM,uBAAuB,CAAC;AAa5D,MAAM,CAAC,KAAK,UAAU,SAAS,CAAC,EAAE,IAAI,EAAoB;IACxD,MAAM,SAAS,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC;IACzC,IAAI,OAAO,CAAC;IACZ,IAAI,CAAC;QACH,OAAO,GAAG,MAAM,OAAO,CAAC,SAAS,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,CAAC;IAC9D,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAK,KAA+B,EAAE,IAAI,KAAK,QAAQ;YAAE,OAAO,EAAE,CAAC;QACnE,MAAM,KAAK,CAAC;IACd,CAAC;IACD,MAAM,KAAK,GAAiB,EAAE,CAAC;IAC/B,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC5B,IAAI,CAAC,KAAK,CAAC,WAAW,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,OAAO,CAAC;YAAE,SAAS;QACtE,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,KAAK,CAAC,IAAI,EAAE,aAAa,CAAC,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QACvI,MAAM,OAAO,GAAG,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,KAAK,CAAC,IAAI,EAAE,cAAc,CAAC,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QACzI,MAAM,OAAO,GAAG,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,KAAK,CAAC,IAAI,EAAE,cAAc,CAAC,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QACzI,IAAI,CAAC,MAAM,IAAI,CAAC,OAAO;YAAE,SAAS;QAClC,KAAK,CAAC,IAAI,CAAC;YACT,OAAO,EAAE,KAAK,CAAC,IAAI;YACnB,MAAM,EAAG,MAAM,EAAE,MAAiB,IAAK,OAAO,EAAE,KAAgB,IAAI,SAAS;YAC7E,OAAO,EAAI,MAAM,EAAE,OAAmC,EAAE,IAAe,IAAK,OAAO,EAAE,OAAkB,IAAI,IAAI;YAC/G,OAAO,EAAG,MAAM,EAAE,OAAkB,IAAK,OAAO,EAAE,OAAkB,IAAI,IAAI;YAC5E,WAAW,EAAI,MAAM,EAAE,SAAqC,EAAE,WAAsB,IAAK,OAAO,EAAE,WAAsB,IAAI,IAAI;YAChI,cAAc,EAAG,MAAM,EAAE,OAAmC,EAAE,cAA+B,IAAI,IAAI;YACrG,UAAU,EAAG,MAAM,EAAE,UAAqB,IAAK,OAAO,EAAE,UAAqB,IAAI,IAAI;YACrF,YAAY,EAAG,MAAM,EAAE,YAAuB,IAAI,IAAI;SACvD,CAAC,CAAC;IACL,CAAC;IACD,OAAO,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,UAAU,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,UAAU,IAAI,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;AACvI,CAAC;AAkBD,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,MAAc,EAAE,EAAE,IAAI,EAAoB;IAC1E,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,CAAC,qBAAqB,CAAC,IAAI,CAAC,MAAM,CAAC;QAAE,MAAM,YAAY,CAAC,oBAAoB,MAAM,EAAE,CAAC,CAAC;IACxH,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC;IAC5D,MAAM,OAAO,GAAG,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,cAAc,CAAC,EAAE,IAAI,CAAC,CAAC;IAC3G,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,UAAU,CAAC,mBAAmB,MAAM,EAAE,EAAE,EAAE,IAAI,EAAE,gBAAgB,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC,CAAC;IACzG,OAAO;QACL,cAAc,EAAE,cAAc;QAC9B,OAAO,EAAE,MAAM;QACf,SAAS;QACT,OAAO;QACP,UAAU,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,iBAAiB,CAAC,EAAE,IAAI,CAAC;QACzG,IAAI,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,WAAW,CAAC,EAAE,IAAI,CAAC;QAC7F,cAAc,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,qBAAqB,CAAC,EAAE,IAAI,CAAC;QACjH,UAAU,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,iBAAiB,CAAC,EAAE,IAAI,CAAC;QACzG,OAAO,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,cAAc,CAAC,EAAE,IAAI,CAAC;QACnG,QAAQ,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,eAAe,CAAC,EAAE,IAAI,CAAC;QACrG,MAAM,EAAE,MAAM,mBAAmB,CAAC,SAAS,CAAC;QAC5C,MAAM,EAAE,MAAM,QAAQ,CAAiC,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,EAAE,IAAI,CAAC;QACjG,eAAe,EAAE,MAAM,sBAAsB,CAAC,SAAS,CAAC;KACzD,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAC,IAAY,EAAE,MAAc;IACnE,IAAI,CAAC;QACH,OAAO,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,MAAM,EAAE,iBAAiB,CAAC,CAAC,CAAC,CAAC,MAAM,EAAE,CAAC;IAC9F,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAK,KAA+B,CAAC,IAAI,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QACrE,MAAM,KAAK,CAAC;IACd,CAAC;AACH,CAAC"}
|