@tangle-network/agent-eval 0.93.0 → 0.95.0
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/CHANGELOG.md +15 -0
- package/README.md +44 -30
- package/dist/action-policy.d.ts +24 -0
- package/dist/action-policy.d.ts.map +1 -0
- package/dist/action-policy.test.d.ts +2 -0
- package/dist/action-policy.test.d.ts.map +1 -0
- package/dist/active-learning.d.ts +41 -0
- package/dist/active-learning.d.ts.map +1 -0
- package/dist/adapters/http.d.ts +15 -21
- package/dist/adapters/http.d.ts.map +1 -0
- package/dist/adapters/http.js.map +1 -1
- package/dist/adapters/langchain.d.ts +7 -13
- package/dist/adapters/langchain.d.ts.map +1 -0
- package/dist/adapters/otel.d.ts +13 -24
- package/dist/adapters/otel.d.ts.map +1 -0
- package/dist/agent-profile-cell.d.ts +101 -0
- package/dist/agent-profile-cell.d.ts.map +1 -0
- package/dist/agent-profile.d.ts +27 -0
- package/dist/agent-profile.d.ts.map +1 -0
- package/dist/agent-profile.test.d.ts +2 -0
- package/dist/agent-profile.test.d.ts.map +1 -0
- package/dist/analyst/adapters.d.ts +62 -0
- package/dist/analyst/adapters.d.ts.map +1 -0
- package/dist/analyst/analyst.test.d.ts +2 -0
- package/dist/analyst/analyst.test.d.ts.map +1 -0
- package/dist/analyst/ax-service.d.ts +27 -0
- package/dist/analyst/ax-service.d.ts.map +1 -0
- package/dist/analyst/behavioral-analyst.d.ts +28 -0
- package/dist/analyst/behavioral-analyst.d.ts.map +1 -0
- package/dist/analyst/chat-client.d.ts +91 -0
- package/dist/analyst/chat-client.d.ts.map +1 -0
- package/dist/analyst/default-registry.d.ts +27 -0
- package/dist/analyst/default-registry.d.ts.map +1 -0
- package/dist/analyst/default-registry.test.d.ts +2 -0
- package/dist/analyst/default-registry.test.d.ts.map +1 -0
- package/dist/analyst/finding-signature.d.ts +48 -0
- package/dist/analyst/finding-signature.d.ts.map +1 -0
- package/dist/analyst/finding-subject.d.ts +146 -0
- package/dist/analyst/finding-subject.d.ts.map +1 -0
- package/dist/analyst/finding-subject.test.d.ts +2 -0
- package/dist/analyst/finding-subject.test.d.ts.map +1 -0
- package/dist/analyst/findings-store.d.ts +75 -0
- package/dist/analyst/findings-store.d.ts.map +1 -0
- package/dist/analyst/index.d.ts +28 -256
- package/dist/analyst/index.d.ts.map +1 -0
- package/dist/analyst/index.js +6 -66
- package/dist/analyst/index.js.map +1 -1
- package/dist/{kind-factory-5b7xXXOr.d.ts → analyst/kind-factory.d.ts} +11 -63
- package/dist/analyst/kind-factory.d.ts.map +1 -0
- package/dist/analyst/kinds/failure-mode.d.ts +19 -0
- package/dist/analyst/kinds/failure-mode.d.ts.map +1 -0
- package/dist/analyst/kinds/improvement.d.ts +23 -0
- package/dist/analyst/kinds/improvement.d.ts.map +1 -0
- package/dist/analyst/kinds/index.d.ts +22 -0
- package/dist/analyst/kinds/index.d.ts.map +1 -0
- package/dist/analyst/kinds/kinds.test.d.ts +2 -0
- package/dist/analyst/kinds/kinds.test.d.ts.map +1 -0
- package/dist/analyst/kinds/knowledge-gap.d.ts +28 -0
- package/dist/analyst/kinds/knowledge-gap.d.ts.map +1 -0
- package/dist/analyst/kinds/knowledge-poisoning.d.ts +22 -0
- package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +1 -0
- package/dist/analyst/kinds/skill-usage.d.ts +84 -0
- package/dist/analyst/kinds/skill-usage.d.ts.map +1 -0
- package/dist/analyst/kinds/skill-usage.test.d.ts +2 -0
- package/dist/analyst/kinds/skill-usage.test.d.ts.map +1 -0
- package/dist/analyst/parse-tolerant.d.ts +26 -0
- package/dist/analyst/parse-tolerant.d.ts.map +1 -0
- package/dist/analyst/parse-tolerant.test.d.ts +2 -0
- package/dist/analyst/parse-tolerant.test.d.ts.map +1 -0
- package/dist/analyst/registry.budget.test.d.ts +2 -0
- package/dist/analyst/registry.budget.test.d.ts.map +1 -0
- package/dist/{default-registry-zoGHUQEH.d.ts → analyst/registry.d.ts} +8 -37
- package/dist/analyst/registry.d.ts.map +1 -0
- package/dist/analyst/steer-firewall.d.ts +35 -0
- package/dist/analyst/steer-firewall.d.ts.map +1 -0
- package/dist/analyst/steer-firewall.test.d.ts +2 -0
- package/dist/analyst/steer-firewall.test.d.ts.map +1 -0
- package/dist/analyst/structure-findings.d.ts +37 -0
- package/dist/analyst/structure-findings.d.ts.map +1 -0
- package/dist/analyst/structure-findings.test.d.ts +2 -0
- package/dist/analyst/structure-findings.test.d.ts.map +1 -0
- package/dist/analyst/tool-groups.d.ts +34 -0
- package/dist/analyst/tool-groups.d.ts.map +1 -0
- package/dist/{types-2VVIL04s.d.ts → analyst/types.d.ts} +19 -112
- package/dist/analyst/types.d.ts.map +1 -0
- package/dist/anti-slop.d.ts +59 -0
- package/dist/anti-slop.d.ts.map +1 -0
- package/dist/artifact-validator.d.ts +74 -0
- package/dist/artifact-validator.d.ts.map +1 -0
- package/dist/attestation.d.ts +63 -0
- package/dist/attestation.d.ts.map +1 -0
- package/dist/attestation.test.d.ts +2 -0
- package/dist/attestation.test.d.ts.map +1 -0
- package/dist/authenticity/index.d.ts +15 -16
- package/dist/authenticity/index.d.ts.map +1 -0
- package/dist/authenticity/index.test.d.ts +2 -0
- package/dist/authenticity/index.test.d.ts.map +1 -0
- package/dist/auto-pr.d.ts +120 -0
- package/dist/auto-pr.d.ts.map +1 -0
- package/dist/{baseline-DE36-Np7.d.ts → baseline.d.ts} +8 -44
- package/dist/baseline.d.ts.map +1 -0
- package/dist/behavior-dsl.d.ts +73 -0
- package/dist/behavior-dsl.d.ts.map +1 -0
- package/dist/belief-state/calibration.d.ts +10 -0
- package/dist/belief-state/calibration.d.ts.map +1 -0
- package/dist/belief-state/calibration.test.d.ts +2 -0
- package/dist/belief-state/calibration.test.d.ts.map +1 -0
- package/dist/belief-state/code-agent-corpus.d.ts +66 -0
- package/dist/belief-state/code-agent-corpus.d.ts.map +1 -0
- package/dist/belief-state/code-agent-corpus.test.d.ts +2 -0
- package/dist/belief-state/code-agent-corpus.test.d.ts.map +1 -0
- package/dist/belief-state/code-agent-evidence.d.ts +22 -0
- package/dist/belief-state/code-agent-evidence.d.ts.map +1 -0
- package/dist/belief-state/code-agent-evidence.test.d.ts +2 -0
- package/dist/belief-state/code-agent-evidence.test.d.ts.map +1 -0
- package/dist/belief-state/extract.d.ts +7 -0
- package/dist/belief-state/extract.d.ts.map +1 -0
- package/dist/belief-state/extract.test.d.ts +2 -0
- package/dist/belief-state/extract.test.d.ts.map +1 -0
- package/dist/belief-state/index.d.ts +14 -604
- package/dist/belief-state/index.d.ts.map +1 -0
- package/dist/belief-state/index.js +1 -1
- package/dist/belief-state/ope.d.ts +17 -0
- package/dist/belief-state/ope.d.ts.map +1 -0
- package/dist/belief-state/ope.test.d.ts +2 -0
- package/dist/belief-state/ope.test.d.ts.map +1 -0
- package/dist/belief-state/phase0-measurement.d.ts +55 -0
- package/dist/belief-state/phase0-measurement.d.ts.map +1 -0
- package/dist/belief-state/report.d.ts +17 -0
- package/dist/belief-state/report.d.ts.map +1 -0
- package/dist/belief-state/report.test.d.ts +2 -0
- package/dist/belief-state/report.test.d.ts.map +1 -0
- package/dist/belief-state/research-evidence.d.ts +23 -0
- package/dist/belief-state/research-evidence.d.ts.map +1 -0
- package/dist/belief-state/research-evidence.test.d.ts +2 -0
- package/dist/belief-state/research-evidence.test.d.ts.map +1 -0
- package/dist/belief-state/runtime-benchmark-corpus.d.ts +32 -0
- package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +1 -0
- package/dist/belief-state/runtime-hooks.d.ts +87 -0
- package/dist/belief-state/runtime-hooks.d.ts.map +1 -0
- package/dist/belief-state/runtime-hooks.test.d.ts +2 -0
- package/dist/belief-state/runtime-hooks.test.d.ts.map +1 -0
- package/dist/belief-state/selective.d.ts +15 -0
- package/dist/belief-state/selective.d.ts.map +1 -0
- package/dist/belief-state/selective.test.d.ts +2 -0
- package/dist/belief-state/selective.test.d.ts.map +1 -0
- package/dist/belief-state/shadow-probe.d.ts +80 -0
- package/dist/belief-state/shadow-probe.d.ts.map +1 -0
- package/dist/belief-state/shadow-probe.test.d.ts +2 -0
- package/dist/belief-state/shadow-probe.test.d.ts.map +1 -0
- package/dist/belief-state/types.d.ts +195 -0
- package/dist/belief-state/types.d.ts.map +1 -0
- package/dist/belief-state/types.test.d.ts +2 -0
- package/dist/belief-state/types.test.d.ts.map +1 -0
- package/dist/benchmark.d.ts +14 -0
- package/dist/benchmark.d.ts.map +1 -0
- package/dist/benchmarks/index.d.ts +23 -4
- package/dist/benchmarks/index.d.ts.map +1 -0
- package/dist/benchmarks/routing/dataset.d.ts +34 -0
- package/dist/benchmarks/routing/dataset.d.ts.map +1 -0
- package/dist/benchmarks/routing/index.d.ts +34 -0
- package/dist/benchmarks/routing/index.d.ts.map +1 -0
- package/dist/benchmarks/types.d.ts +49 -0
- package/dist/benchmarks/types.d.ts.map +1 -0
- package/dist/bisector.d.ts +81 -0
- package/dist/bisector.d.ts.map +1 -0
- package/dist/budget-guard.d.ts +31 -0
- package/dist/budget-guard.d.ts.map +1 -0
- package/dist/builder-eval/builder-session.d.ts +111 -0
- package/dist/builder-eval/builder-session.d.ts.map +1 -0
- package/dist/builder-eval/correlation.d.ts +32 -0
- package/dist/builder-eval/correlation.d.ts.map +1 -0
- package/dist/builder-eval/index.d.ts +5 -250
- package/dist/builder-eval/index.d.ts.map +1 -0
- package/dist/builder-eval/index.js +2 -2
- package/dist/builder-eval/project-registry.d.ts +51 -0
- package/dist/builder-eval/project-registry.d.ts.map +1 -0
- package/dist/builder-eval/three-layer-eval.d.ts +55 -0
- package/dist/builder-eval/three-layer-eval.d.ts.map +1 -0
- package/dist/campaign/analyst-surface.d.ts +108 -0
- package/dist/campaign/analyst-surface.d.ts.map +1 -0
- package/dist/campaign/analyst-surface.test.d.ts +2 -0
- package/dist/campaign/analyst-surface.test.d.ts.map +1 -0
- package/dist/campaign/auto-pr.d.ts +46 -0
- package/dist/campaign/auto-pr.d.ts.map +1 -0
- package/dist/campaign/distillation/agreement-judge.d.ts +69 -0
- package/dist/campaign/distillation/agreement-judge.d.ts.map +1 -0
- package/dist/campaign/distillation/cli.d.ts +35 -0
- package/dist/campaign/distillation/cli.d.ts.map +1 -0
- package/dist/campaign/distillation/distillation.test.d.ts +2 -0
- package/dist/campaign/distillation/distillation.test.d.ts.map +1 -0
- package/dist/campaign/distillation/gold-scenarios.d.ts +54 -0
- package/dist/campaign/distillation/gold-scenarios.d.ts.map +1 -0
- package/dist/campaign/distillation/run-distillation.d.ts +119 -0
- package/dist/campaign/distillation/run-distillation.d.ts.map +1 -0
- package/dist/campaign/gates/compose.d.ts +12 -0
- package/dist/campaign/gates/compose.d.ts.map +1 -0
- package/dist/campaign/gates/default-production-gate.d.ts +58 -0
- package/dist/campaign/gates/default-production-gate.d.ts.map +1 -0
- package/dist/campaign/gates/heldout-gate.d.ts +12 -0
- package/dist/campaign/gates/heldout-gate.d.ts.map +1 -0
- package/dist/campaign/gates/promotion-policy.d.ts +125 -0
- package/dist/campaign/gates/promotion-policy.d.ts.map +1 -0
- package/dist/campaign/gates/promotion-policy.test.d.ts +2 -0
- package/dist/campaign/gates/promotion-policy.test.d.ts.map +1 -0
- package/dist/campaign/gates/sequential.d.ts +146 -0
- package/dist/campaign/gates/sequential.d.ts.map +1 -0
- package/dist/campaign/gates/sequential.test.d.ts +2 -0
- package/dist/campaign/gates/sequential.test.d.ts.map +1 -0
- package/dist/campaign/gates/statistical-heldout.d.ts +99 -0
- package/dist/campaign/gates/statistical-heldout.d.ts.map +1 -0
- package/dist/campaign/gates/statistical-heldout.test.d.ts +2 -0
- package/dist/campaign/gates/statistical-heldout.test.d.ts.map +1 -0
- package/dist/campaign/index.d.ts +38 -1341
- package/dist/campaign/index.d.ts.map +1 -0
- package/dist/campaign/index.js +1865 -1318
- package/dist/campaign/index.js.map +1 -1
- package/dist/campaign/labeled-store/fs-adapter.d.ts +59 -0
- package/dist/campaign/labeled-store/fs-adapter.d.ts.map +1 -0
- package/dist/campaign/presets/compare-proposers.d.ts +146 -0
- package/dist/campaign/presets/compare-proposers.d.ts.map +1 -0
- package/dist/campaign/presets/playback.d.ts +120 -0
- package/dist/campaign/presets/playback.d.ts.map +1 -0
- package/dist/campaign/presets/playback.test.d.ts +2 -0
- package/dist/campaign/presets/playback.test.d.ts.map +1 -0
- package/dist/campaign/presets/run-eval.d.ts +14 -0
- package/dist/campaign/presets/run-eval.d.ts.map +1 -0
- package/dist/campaign/presets/run-improvement-loop.d.ts +63 -0
- package/dist/campaign/presets/run-improvement-loop.d.ts.map +1 -0
- package/dist/campaign/presets/run-improvement-loop.test.d.ts +2 -0
- package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +1 -0
- package/dist/campaign/presets/run-optimization.d.ts +92 -0
- package/dist/campaign/presets/run-optimization.d.ts.map +1 -0
- package/dist/campaign/presets/run-profile-matrix.d.ts +151 -0
- package/dist/campaign/presets/run-profile-matrix.d.ts.map +1 -0
- package/dist/campaign/presets/run-skill-opt.d.ts +96 -0
- package/dist/campaign/presets/run-skill-opt.d.ts.map +1 -0
- package/dist/campaign/proposers/_findings-text.d.ts +22 -0
- package/dist/campaign/proposers/_findings-text.d.ts.map +1 -0
- package/dist/campaign/proposers/ace.d.ts +33 -0
- package/dist/campaign/proposers/ace.d.ts.map +1 -0
- package/dist/campaign/proposers/ace.test.d.ts +2 -0
- package/dist/campaign/proposers/ace.test.d.ts.map +1 -0
- package/dist/campaign/proposers/analysis-edit.d.ts +32 -0
- package/dist/campaign/proposers/analysis-edit.d.ts.map +1 -0
- package/dist/campaign/proposers/evolutionary.d.ts +20 -0
- package/dist/campaign/proposers/evolutionary.d.ts.map +1 -0
- package/dist/campaign/proposers/fapo.d.ts +120 -0
- package/dist/campaign/proposers/fapo.d.ts.map +1 -0
- package/dist/campaign/proposers/gepa.d.ts +86 -0
- package/dist/campaign/proposers/gepa.d.ts.map +1 -0
- package/dist/campaign/proposers/halo.d.ts +44 -0
- package/dist/campaign/proposers/halo.d.ts.map +1 -0
- package/dist/campaign/proposers/halo.test.d.ts +2 -0
- package/dist/campaign/proposers/halo.test.d.ts.map +1 -0
- package/dist/campaign/proposers/memory.d.ts +47 -0
- package/dist/campaign/proposers/memory.d.ts.map +1 -0
- package/dist/campaign/proposers/memory.test.d.ts +2 -0
- package/dist/campaign/proposers/memory.test.d.ts.map +1 -0
- package/dist/campaign/proposers/skill-opt.d.ts +88 -0
- package/dist/campaign/proposers/skill-opt.d.ts.map +1 -0
- package/dist/campaign/proposers/trace-analyst.d.ts +48 -0
- package/dist/campaign/proposers/trace-analyst.d.ts.map +1 -0
- package/dist/campaign/proposers/trace-analyst.test.d.ts +2 -0
- package/dist/campaign/proposers/trace-analyst.test.d.ts.map +1 -0
- package/dist/campaign/provenance.d.ts +185 -0
- package/dist/campaign/provenance.d.ts.map +1 -0
- package/dist/campaign/run-campaign.d.ts +90 -0
- package/dist/campaign/run-campaign.d.ts.map +1 -0
- package/dist/campaign/score-utils.d.ts +26 -0
- package/dist/campaign/score-utils.d.ts.map +1 -0
- package/dist/campaign/skill-patch.d.ts +62 -0
- package/dist/campaign/skill-patch.d.ts.map +1 -0
- package/dist/campaign/storage.d.ts +38 -0
- package/dist/campaign/storage.d.ts.map +1 -0
- package/dist/{types-BU-7W85F.d.ts → campaign/types.d.ts} +98 -99
- package/dist/campaign/types.d.ts.map +1 -0
- package/dist/campaign/worktree/index.d.ts +53 -0
- package/dist/campaign/worktree/index.d.ts.map +1 -0
- package/dist/canary.d.ts +101 -0
- package/dist/canary.d.ts.map +1 -0
- package/dist/causal-attribution.d.ts +45 -0
- package/dist/causal-attribution.d.ts.map +1 -0
- package/dist/{chunk-TWS7AZEY.js → chunk-2T4EZACH.js} +2 -2
- package/dist/chunk-2T4EZACH.js.map +1 -0
- package/dist/{chunk-LMZQ2Z4U.js → chunk-56GC6VYO.js} +126 -1
- package/dist/chunk-56GC6VYO.js.map +1 -0
- package/dist/{chunk-QG2OVF2D.js → chunk-77T4STFI.js} +154 -112
- package/dist/chunk-77T4STFI.js.map +1 -0
- package/dist/{chunk-ZFIBGEOL.js → chunk-7QTQKIDD.js} +179 -178
- package/dist/chunk-7QTQKIDD.js.map +1 -0
- package/dist/{chunk-UMMZHCPB.js → chunk-AQ5WQAIV.js} +26 -9
- package/dist/chunk-AQ5WQAIV.js.map +1 -0
- package/dist/{chunk-CVVHBFGN.js → chunk-CWNP4DV4.js} +53 -5
- package/dist/chunk-CWNP4DV4.js.map +1 -0
- package/dist/{chunk-QS3RBQPI.js → chunk-D5KBVULY.js} +2 -2
- package/dist/chunk-DJWX3GVS.js +81 -0
- package/dist/chunk-DJWX3GVS.js.map +1 -0
- package/dist/{chunk-S6OZEZQK.js → chunk-HMA63UEO.js} +37 -9
- package/dist/{chunk-S6OZEZQK.js.map → chunk-HMA63UEO.js.map} +1 -1
- package/dist/{chunk-UHMJT4T7.js → chunk-IZCEK2HR.js} +2 -2
- package/dist/{chunk-6SOJM3VR.js → chunk-KKWJD5E6.js} +21 -21
- package/dist/chunk-KKWJD5E6.js.map +1 -0
- package/dist/{chunk-KWRRMR3J.js → chunk-LO6IOIJ2.js} +10 -10
- package/dist/chunk-LO6IOIJ2.js.map +1 -0
- package/dist/{chunk-L3JOU6XM.js → chunk-NACM5RS7.js} +3 -3
- package/dist/{chunk-E4GH6USR.js → chunk-NZEQVRH5.js} +2 -2
- package/dist/chunk-NZEQVRH5.js.map +1 -0
- package/dist/{chunk-XY4DDNEG.js → chunk-PSWWQXHF.js} +14 -89
- package/dist/chunk-PSWWQXHF.js.map +1 -0
- package/dist/{chunk-4FBZZIYD.js → chunk-QTMYW64F.js} +2 -2
- package/dist/{chunk-D3V5B42D.js → chunk-S4SYLDFX.js} +7 -4
- package/dist/chunk-S4SYLDFX.js.map +1 -0
- package/dist/{chunk-47X6LRCE.js → chunk-UFSG7ACU.js} +10 -7
- package/dist/chunk-UFSG7ACU.js.map +1 -0
- package/dist/{chunk-CY6U5S3X.js → chunk-URWVKRCS.js} +2 -2
- package/dist/{chunk-GSH6QNNS.js → chunk-X74V6ESX.js} +102 -79
- package/dist/chunk-X74V6ESX.js.map +1 -0
- package/dist/{chunk-QAY5UIJO.js → chunk-YBIGNSCZ.js} +178 -59
- package/dist/chunk-YBIGNSCZ.js.map +1 -0
- package/dist/{chunk-Y47J2LJ3.js → chunk-Z6L6YSU6.js} +5 -5
- package/dist/{chunk-T375SUOZ.js → chunk-ZOPLCJSY.js} +86 -31
- package/dist/chunk-ZOPLCJSY.js.map +1 -0
- package/dist/ci-gate.d.ts +44 -0
- package/dist/ci-gate.d.ts.map +1 -0
- package/dist/cli.d.ts +2 -0
- package/dist/cli.d.ts.map +1 -0
- package/dist/cli.js +2 -2
- package/dist/client.d.ts +77 -0
- package/dist/client.d.ts.map +1 -0
- package/dist/client.test.d.ts +2 -0
- package/dist/client.test.d.ts.map +1 -0
- package/dist/command-runner.d.ts +74 -0
- package/dist/command-runner.d.ts.map +1 -0
- package/dist/command-runner.test.d.ts +2 -0
- package/dist/command-runner.test.d.ts.map +1 -0
- package/dist/completion-verifier.d.ts +147 -0
- package/dist/completion-verifier.d.ts.map +1 -0
- package/dist/completion-verifier.test.d.ts +9 -0
- package/dist/completion-verifier.test.d.ts.map +1 -0
- package/dist/concurrency.d.ts +23 -0
- package/dist/concurrency.d.ts.map +1 -0
- package/dist/contamination-guard.d.ts +81 -0
- package/dist/contamination-guard.d.ts.map +1 -0
- package/dist/{analyze-runs-DwCEkpO_.d.ts → contract/analyze-runs.d.ts} +9 -10
- package/dist/contract/analyze-runs.d.ts.map +1 -0
- package/dist/contract/define-agent-eval.d.ts +52 -0
- package/dist/contract/define-agent-eval.d.ts.map +1 -0
- package/dist/contract/diff.d.ts +114 -0
- package/dist/contract/diff.d.ts.map +1 -0
- package/dist/contract/index.d.ts +106 -640
- package/dist/contract/index.d.ts.map +1 -0
- package/dist/contract/index.js +131 -21
- package/dist/contract/index.js.map +1 -1
- package/dist/{insight-report-BBwvOh6x.d.ts → contract/insight-report.d.ts} +16 -19
- package/dist/contract/insight-report.d.ts.map +1 -0
- package/dist/contract/insight-types-fwd.d.ts +7 -0
- package/dist/contract/insight-types-fwd.d.ts.map +1 -0
- package/dist/contract/intake/agent-trace.d.ts +97 -0
- package/dist/contract/intake/agent-trace.d.ts.map +1 -0
- package/dist/{code-agent-session-BO8nCnv3.d.ts → contract/intake/code-agent-session.d.ts} +15 -17
- package/dist/contract/intake/code-agent-session.d.ts.map +1 -0
- package/dist/contract/intake/feedback-table.d.ts +87 -0
- package/dist/contract/intake/feedback-table.d.ts.map +1 -0
- package/dist/contract/intake/index.d.ts +22 -0
- package/dist/contract/intake/index.d.ts.map +1 -0
- package/dist/contract/intake/otel-spans.d.ts +33 -0
- package/dist/contract/intake/otel-spans.d.ts.map +1 -0
- package/dist/contract/self-improve.d.ts +284 -0
- package/dist/contract/self-improve.d.ts.map +1 -0
- package/dist/{control-runtime-DuFBYg7A.d.ts → control-runtime.d.ts} +23 -26
- package/dist/control-runtime.d.ts.map +1 -0
- package/dist/control-runtime.test.d.ts +2 -0
- package/dist/control-runtime.test.d.ts.map +1 -0
- package/dist/control.d.ts +11 -9
- package/dist/control.d.ts.map +1 -0
- package/dist/control.js +2 -2
- package/dist/convergence.d.ts +29 -0
- package/dist/convergence.d.ts.map +1 -0
- package/dist/{cost-ledger-DuSqlw5B.d.ts → cost-ledger.d.ts} +10 -11
- package/dist/cost-ledger.d.ts.map +1 -0
- package/dist/cost-ledger.test.d.ts +2 -0
- package/dist/cost-ledger.test.d.ts.map +1 -0
- package/dist/cost-report.d.ts +43 -0
- package/dist/cost-report.d.ts.map +1 -0
- package/dist/cost-report.test.d.ts +2 -0
- package/dist/cost-report.test.d.ts.map +1 -0
- package/dist/cost-tracker.d.ts +76 -0
- package/dist/cost-tracker.d.ts.map +1 -0
- package/dist/{counterfactual-Dwibr5IW.d.ts → counterfactual.d.ts} +12 -13
- package/dist/counterfactual.d.ts.map +1 -0
- package/dist/cross-trace-diff.d.ts +56 -0
- package/dist/cross-trace-diff.d.ts.map +1 -0
- package/dist/{dataset-BbGkaN2I.d.ts → dataset.d.ts} +11 -14
- package/dist/dataset.d.ts.map +1 -0
- package/dist/deploy-gate-layer.d.ts +125 -0
- package/dist/deploy-gate-layer.d.ts.map +1 -0
- package/dist/deploy-gate-layer.test.d.ts +2 -0
- package/dist/deploy-gate-layer.test.d.ts.map +1 -0
- package/dist/description-length-gate.d.ts +119 -0
- package/dist/description-length-gate.d.ts.map +1 -0
- package/dist/detectors/edge.test.d.ts +2 -0
- package/dist/detectors/edge.test.d.ts.map +1 -0
- package/dist/detectors/index.d.ts +81 -0
- package/dist/detectors/index.d.ts.map +1 -0
- package/dist/detectors/index.test.d.ts +2 -0
- package/dist/detectors/index.test.d.ts.map +1 -0
- package/dist/diagnose/causal-sweep.d.ts +100 -0
- package/dist/diagnose/causal-sweep.d.ts.map +1 -0
- package/dist/diagnose/index.d.ts +36 -0
- package/dist/diagnose/index.d.ts.map +1 -0
- package/dist/diagnose/remediation.d.ts +68 -0
- package/dist/diagnose/remediation.d.ts.map +1 -0
- package/dist/diagnose/repair.d.ts +77 -0
- package/dist/diagnose/repair.d.ts.map +1 -0
- package/dist/diagnose.d.ts +1 -251
- package/dist/diagnose.js +2 -2
- package/dist/discover-personas.d.ts +35 -0
- package/dist/discover-personas.d.ts.map +1 -0
- package/dist/driver.d.ts +95 -0
- package/dist/driver.d.ts.map +1 -0
- package/dist/driver.test.d.ts +8 -0
- package/dist/driver.test.d.ts.map +1 -0
- package/dist/dual-agent-bench.d.ts +81 -0
- package/dist/dual-agent-bench.d.ts.map +1 -0
- package/dist/error-count-extractor.d.ts +47 -0
- package/dist/error-count-extractor.d.ts.map +1 -0
- package/dist/error-count-extractor.test.d.ts +2 -0
- package/dist/error-count-extractor.test.d.ts.map +1 -0
- package/dist/{errors-CzMUYo7b.d.ts → errors.d.ts} +10 -11
- package/dist/errors.d.ts.map +1 -0
- package/dist/{researcher-DE6Gpnb4.d.ts → eval-campaign.d.ts} +34 -156
- package/dist/eval-campaign.d.ts.map +1 -0
- package/dist/eval-campaign.test.d.ts +2 -0
- package/dist/eval-campaign.test.d.ts.map +1 -0
- package/dist/eval-tools.d.ts +55 -0
- package/dist/eval-tools.d.ts.map +1 -0
- package/dist/eval-trace-store.d.ts +107 -0
- package/dist/eval-trace-store.d.ts.map +1 -0
- package/dist/eval-trace-store.test.d.ts +2 -0
- package/dist/eval-trace-store.test.d.ts.map +1 -0
- package/dist/executor.d.ts +38 -0
- package/dist/executor.d.ts.map +1 -0
- package/dist/executor.test.d.ts +10 -0
- package/dist/executor.test.d.ts.map +1 -0
- package/dist/experiment-tracker.d.ts +178 -0
- package/dist/experiment-tracker.d.ts.map +1 -0
- package/dist/experiment-tracker.test.d.ts +2 -0
- package/dist/experiment-tracker.test.d.ts.map +1 -0
- package/dist/failure-taxonomy.d.ts +38 -0
- package/dist/failure-taxonomy.d.ts.map +1 -0
- package/dist/{feedback-trajectory-D9OVLrg9.d.ts → feedback-trajectory.d.ts} +36 -38
- package/dist/feedback-trajectory.d.ts.map +1 -0
- package/dist/feedback-trajectory.test.d.ts +2 -0
- package/dist/feedback-trajectory.test.d.ts.map +1 -0
- package/dist/flow-layer.d.ts +90 -0
- package/dist/flow-layer.d.ts.map +1 -0
- package/dist/flow-layer.test.d.ts +2 -0
- package/dist/flow-layer.test.d.ts.map +1 -0
- package/dist/fuzz/capsule.d.ts +46 -0
- package/dist/fuzz/capsule.d.ts.map +1 -0
- package/dist/fuzz/cube.d.ts +36 -0
- package/dist/fuzz/cube.d.ts.map +1 -0
- package/dist/fuzz/explorer-cost.test.d.ts +2 -0
- package/dist/fuzz/explorer-cost.test.d.ts.map +1 -0
- package/dist/fuzz/explorer.d.ts +64 -0
- package/dist/fuzz/explorer.d.ts.map +1 -0
- package/dist/fuzz/fuzz-agent.d.ts +16 -0
- package/dist/fuzz/fuzz-agent.d.ts.map +1 -0
- package/dist/fuzz/fuzz-agent.test.d.ts +2 -0
- package/dist/fuzz/fuzz-agent.test.d.ts.map +1 -0
- package/dist/fuzz/gates.d.ts +33 -0
- package/dist/fuzz/gates.d.ts.map +1 -0
- package/dist/fuzz/index.d.ts +26 -0
- package/dist/fuzz/index.d.ts.map +1 -0
- package/dist/fuzz/policies.d.ts +28 -0
- package/dist/fuzz/policies.d.ts.map +1 -0
- package/dist/fuzz/tools.d.ts +20 -0
- package/dist/fuzz/tools.d.ts.map +1 -0
- package/dist/fuzz/types.d.ts +307 -0
- package/dist/fuzz/types.d.ts.map +1 -0
- package/dist/fuzz.d.ts +1 -547
- package/dist/golden-matcher.d.ts +71 -0
- package/dist/golden-matcher.d.ts.map +1 -0
- package/dist/governance/eu-ai-act.d.ts +37 -0
- package/dist/governance/eu-ai-act.d.ts.map +1 -0
- package/dist/governance/index.d.ts +5 -135
- package/dist/governance/index.d.ts.map +1 -0
- package/dist/governance/nist-ai-rmf.d.ts +15 -0
- package/dist/governance/nist-ai-rmf.d.ts.map +1 -0
- package/dist/governance/soc2.d.ts +12 -0
- package/dist/governance/soc2.d.ts.map +1 -0
- package/dist/governance/types.d.ts +66 -0
- package/dist/governance/types.d.ts.map +1 -0
- package/dist/harness-optimizer.d.ts +82 -0
- package/dist/harness-optimizer.d.ts.map +1 -0
- package/dist/held-out-gate.d.ts +135 -0
- package/dist/held-out-gate.d.ts.map +1 -0
- package/dist/hosted/client.d.ts +73 -0
- package/dist/hosted/client.d.ts.map +1 -0
- package/dist/hosted/from-env.test.d.ts +8 -0
- package/dist/hosted/from-env.test.d.ts.map +1 -0
- package/dist/hosted/index.d.ts +10 -238
- package/dist/hosted/index.d.ts.map +1 -0
- package/dist/hosted/types.d.ts +159 -0
- package/dist/hosted/types.d.ts.map +1 -0
- package/dist/index.d.ts +277 -5643
- package/dist/index.d.ts.map +1 -0
- package/dist/index.js +317 -108
- package/dist/index.js.map +1 -1
- package/dist/integrity/backend-integrity.d.ts +71 -0
- package/dist/integrity/backend-integrity.d.ts.map +1 -0
- package/dist/integrity/preflight.d.ts +72 -0
- package/dist/integrity/preflight.d.ts.map +1 -0
- package/dist/integrity/preflight.test.d.ts +2 -0
- package/dist/integrity/preflight.test.d.ts.map +1 -0
- package/dist/integrity/single-backend.d.ts +67 -0
- package/dist/integrity/single-backend.d.ts.map +1 -0
- package/dist/intent-match-judge.d.ts +69 -0
- package/dist/intent-match-judge.d.ts.map +1 -0
- package/dist/intent-match-judge.test.d.ts +2 -0
- package/dist/intent-match-judge.test.d.ts.map +1 -0
- package/dist/{judge-calibration-0p2QcWNE.d.ts → judge-calibration.d.ts} +16 -17
- package/dist/judge-calibration.d.ts.map +1 -0
- package/dist/judge-ensemble.d.ts +66 -0
- package/dist/judge-ensemble.d.ts.map +1 -0
- package/dist/judge-ensemble.test.d.ts +8 -0
- package/dist/judge-ensemble.test.d.ts.map +1 -0
- package/dist/judge-families.d.ts +38 -0
- package/dist/judge-families.d.ts.map +1 -0
- package/dist/judge-panel.d.ts +65 -0
- package/dist/judge-panel.d.ts.map +1 -0
- package/dist/judge-retry.d.ts +70 -0
- package/dist/judge-retry.d.ts.map +1 -0
- package/dist/judge-runner.d.ts +36 -0
- package/dist/judge-runner.d.ts.map +1 -0
- package/dist/judge-runner.test.d.ts +2 -0
- package/dist/judge-runner.test.d.ts.map +1 -0
- package/dist/judges.d.ts +74 -0
- package/dist/judges.d.ts.map +1 -0
- package/dist/keyword-coverage-judge.d.ts +89 -0
- package/dist/keyword-coverage-judge.d.ts.map +1 -0
- package/dist/keyword-coverage-judge.test.d.ts +2 -0
- package/dist/keyword-coverage-judge.test.d.ts.map +1 -0
- package/dist/knowledge/index.d.ts +3 -103
- package/dist/knowledge/index.d.ts.map +1 -0
- package/dist/knowledge/readiness.d.ts +26 -0
- package/dist/knowledge/readiness.d.ts.map +1 -0
- package/dist/knowledge/types.d.ts +75 -0
- package/dist/knowledge/types.d.ts.map +1 -0
- package/dist/live-proof.d.ts +62 -0
- package/dist/live-proof.d.ts.map +1 -0
- package/dist/{llm-client-BeEcAokY.d.ts → llm-client.d.ts} +52 -23
- package/dist/llm-client.d.ts.map +1 -0
- package/dist/llm-client.test.d.ts +2 -0
- package/dist/llm-client.test.d.ts.map +1 -0
- package/dist/locked-jsonl-appender.d.ts +19 -0
- package/dist/locked-jsonl-appender.d.ts.map +1 -0
- package/dist/matrix/aggregation.d.ts +16 -0
- package/dist/matrix/aggregation.d.ts.map +1 -0
- package/dist/matrix/index.d.ts +12 -30
- package/dist/matrix/index.d.ts.map +1 -0
- package/dist/matrix/runner.d.ts +15 -0
- package/dist/matrix/runner.d.ts.map +1 -0
- package/dist/{types-mn5Aqk7x.d.ts → matrix/types.d.ts} +11 -15
- package/dist/matrix/types.d.ts.map +1 -0
- package/dist/meta-eval/calibration.d.ts +47 -0
- package/dist/meta-eval/calibration.d.ts.map +1 -0
- package/dist/meta-eval/correlation-study.d.ts +53 -0
- package/dist/meta-eval/correlation-study.d.ts.map +1 -0
- package/dist/meta-eval/index.d.ts +6 -181
- package/dist/meta-eval/index.d.ts.map +1 -0
- package/dist/meta-eval/index.js +1 -1
- package/dist/{outcome-store-rnXLEqSn.d.ts → meta-eval/outcome-store.d.ts} +7 -8
- package/dist/meta-eval/outcome-store.d.ts.map +1 -0
- package/dist/{rubric-predictive-validity-Cy_W-hWZ.d.ts → meta-eval/rubric-predictive-validity.d.ts} +8 -11
- package/dist/meta-eval/rubric-predictive-validity.d.ts.map +1 -0
- package/dist/meta-eval/sentinel.d.ts +169 -0
- package/dist/meta-eval/sentinel.d.ts.map +1 -0
- package/dist/metrics.d.ts +63 -0
- package/dist/metrics.d.ts.map +1 -0
- package/dist/model-seats.d.ts +71 -0
- package/dist/model-seats.d.ts.map +1 -0
- package/dist/model-seats.test.d.ts +2 -0
- package/dist/model-seats.test.d.ts.map +1 -0
- package/dist/muffled-gate-scanner.d.ts +102 -0
- package/dist/muffled-gate-scanner.d.ts.map +1 -0
- package/dist/{multi-layer-verifier-DUZXrPDA.d.ts → multi-layer-verifier.d.ts} +12 -15
- package/dist/multi-layer-verifier.d.ts.map +1 -0
- package/dist/multi-layer-verifier.test.d.ts +2 -0
- package/dist/multi-layer-verifier.test.d.ts.map +1 -0
- package/dist/multi-toolchain-layer.d.ts +80 -0
- package/dist/multi-toolchain-layer.d.ts.map +1 -0
- package/dist/multi-toolchain-layer.test.d.ts +2 -0
- package/dist/multi-toolchain-layer.test.d.ts.map +1 -0
- package/dist/multishot/default-tools.d.ts +34 -0
- package/dist/multishot/default-tools.d.ts.map +1 -0
- package/dist/multishot/index.d.ts +7 -290
- package/dist/multishot/index.d.ts.map +1 -0
- package/dist/multishot/index.js.map +1 -1
- package/dist/multishot/judges.d.ts +32 -0
- package/dist/multishot/judges.d.ts.map +1 -0
- package/dist/multishot/matrix.d.ts +107 -0
- package/dist/multishot/matrix.d.ts.map +1 -0
- package/dist/multishot/multishot.d.ts +23 -0
- package/dist/multishot/multishot.d.ts.map +1 -0
- package/dist/multishot/router.d.ts +37 -0
- package/dist/multishot/router.d.ts.map +1 -0
- package/dist/multishot/types.d.ts +60 -0
- package/dist/multishot/types.d.ts.map +1 -0
- package/dist/observability.d.ts +71 -0
- package/dist/observability.d.ts.map +1 -0
- package/dist/openapi.json +1 -1
- package/dist/oracle.d.ts +55 -0
- package/dist/oracle.d.ts.map +1 -0
- package/dist/orthogonality.d.ts +35 -0
- package/dist/orthogonality.d.ts.map +1 -0
- package/dist/otel-pipeline.d.ts +31 -0
- package/dist/otel-pipeline.d.ts.map +1 -0
- package/dist/paraphrase.d.ts +107 -0
- package/dist/paraphrase.d.ts.map +1 -0
- package/dist/{pareto-E-pembql.d.ts → pareto.d.ts} +9 -10
- package/dist/pareto.d.ts.map +1 -0
- package/dist/partition-held-out.d.ts +70 -0
- package/dist/partition-held-out.d.ts.map +1 -0
- package/dist/partition-held-out.test.d.ts +2 -0
- package/dist/partition-held-out.test.d.ts.map +1 -0
- package/dist/perf/index.d.ts +13 -119
- package/dist/perf/index.d.ts.map +1 -0
- package/dist/perf/integrity.d.ts +30 -0
- package/dist/perf/integrity.d.ts.map +1 -0
- package/dist/perf/journey.d.ts +45 -0
- package/dist/perf/journey.d.ts.map +1 -0
- package/dist/perf/ratchet.d.ts +47 -0
- package/dist/perf/ratchet.d.ts.map +1 -0
- package/dist/pipelines/budget-breach.d.ts +31 -0
- package/dist/pipelines/budget-breach.d.ts.map +1 -0
- package/dist/pipelines/budget-breach.test.d.ts +2 -0
- package/dist/pipelines/budget-breach.test.d.ts.map +1 -0
- package/dist/pipelines/failure-cluster.d.ts +38 -0
- package/dist/pipelines/failure-cluster.d.ts.map +1 -0
- package/dist/pipelines/failure-cluster.test.d.ts +2 -0
- package/dist/pipelines/failure-cluster.test.d.ts.map +1 -0
- package/dist/pipelines/first-divergence.d.ts +26 -0
- package/dist/pipelines/first-divergence.d.ts.map +1 -0
- package/dist/pipelines/first-divergence.test.d.ts +2 -0
- package/dist/pipelines/first-divergence.test.d.ts.map +1 -0
- package/dist/pipelines/index.d.ts +8 -173
- package/dist/pipelines/index.d.ts.map +1 -0
- package/dist/pipelines/index.js +40 -15
- package/dist/pipelines/index.js.map +1 -1
- package/dist/pipelines/judge-agreement.d.ts +26 -0
- package/dist/pipelines/judge-agreement.d.ts.map +1 -0
- package/dist/pipelines/judge-agreement.test.d.ts +2 -0
- package/dist/pipelines/judge-agreement.test.d.ts.map +1 -0
- package/dist/pipelines/regression.d.ts +23 -0
- package/dist/pipelines/regression.d.ts.map +1 -0
- package/dist/pipelines/regression.test.d.ts +2 -0
- package/dist/pipelines/regression.test.d.ts.map +1 -0
- package/dist/pipelines/stuck-loop.d.ts +32 -0
- package/dist/pipelines/stuck-loop.d.ts.map +1 -0
- package/dist/pipelines/stuck-loop.test.d.ts +2 -0
- package/dist/pipelines/stuck-loop.test.d.ts.map +1 -0
- package/dist/pipelines/tool-waste.d.ts +34 -0
- package/dist/pipelines/tool-waste.d.ts.map +1 -0
- package/dist/pipelines/tool-waste.test.d.ts +2 -0
- package/dist/pipelines/tool-waste.test.d.ts.map +1 -0
- package/dist/playbook.d.ts +16 -0
- package/dist/playbook.d.ts.map +1 -0
- package/dist/pr-review-benchmark.d.ts +88 -0
- package/dist/pr-review-benchmark.d.ts.map +1 -0
- package/dist/pr-review-benchmark.test.d.ts +2 -0
- package/dist/pr-review-benchmark.test.d.ts.map +1 -0
- package/dist/pre-registration.d.ts +125 -0
- package/dist/pre-registration.d.ts.map +1 -0
- package/dist/prm/builtin-rubrics.d.ts +33 -0
- package/dist/prm/builtin-rubrics.d.ts.map +1 -0
- package/dist/prm/index.d.ts +5 -100
- package/dist/prm/index.d.ts.map +1 -0
- package/dist/prm/index.js +55 -12
- package/dist/prm/index.js.map +1 -1
- package/dist/prm/inference.d.ts +29 -0
- package/dist/prm/inference.d.ts.map +1 -0
- package/dist/prm/inference.test.d.ts +2 -0
- package/dist/prm/inference.test.d.ts.map +1 -0
- package/dist/{rubric-BOfxn4ja.d.ts → prm/rubric.d.ts} +10 -13
- package/dist/prm/rubric.d.ts.map +1 -0
- package/dist/prm/training-export.d.ts +38 -0
- package/dist/prm/training-export.d.ts.map +1 -0
- package/dist/produced-state.d.ts +63 -0
- package/dist/produced-state.d.ts.map +1 -0
- package/dist/produced-state.test.d.ts +8 -0
- package/dist/produced-state.test.d.ts.map +1 -0
- package/dist/profile/baselines.d.ts +37 -0
- package/dist/profile/baselines.d.ts.map +1 -0
- package/dist/profile/index.d.ts +105 -0
- package/dist/profile/index.d.ts.map +1 -0
- package/dist/promotion-gate.d.ts +93 -0
- package/dist/promotion-gate.d.ts.map +1 -0
- package/dist/prompt-registry.d.ts +41 -0
- package/dist/prompt-registry.d.ts.map +1 -0
- package/dist/propose-review-control.d.ts +49 -0
- package/dist/propose-review-control.d.ts.map +1 -0
- package/dist/propose-review-control.test.d.ts +2 -0
- package/dist/propose-review-control.test.d.ts.map +1 -0
- package/dist/propose-review.d.ts +155 -0
- package/dist/propose-review.d.ts.map +1 -0
- package/dist/{red-team-BXHil6c8.d.ts → red-team.d.ts} +14 -16
- package/dist/red-team.d.ts.map +1 -0
- package/dist/reference-replay-steering.d.ts +11 -0
- package/dist/reference-replay-steering.d.ts.map +1 -0
- package/dist/reference-replay.d.ts +176 -0
- package/dist/reference-replay.d.ts.map +1 -0
- package/dist/reflective-mutation.d.ts +79 -0
- package/dist/reflective-mutation.d.ts.map +1 -0
- package/dist/registry.d.ts +31 -0
- package/dist/registry.d.ts.map +1 -0
- package/dist/release-confidence.d.ts +128 -0
- package/dist/release-confidence.d.ts.map +1 -0
- package/dist/release-report.d.ts +11 -0
- package/dist/release-report.d.ts.map +1 -0
- package/dist/replay.d.ts +120 -0
- package/dist/replay.d.ts.map +1 -0
- package/dist/reporter.d.ts +14 -0
- package/dist/reporter.d.ts.map +1 -0
- package/dist/reporting.d.ts +15 -15
- package/dist/reporting.d.ts.map +1 -0
- package/dist/reporting.js +3 -3
- package/dist/researcher.d.ts +140 -0
- package/dist/researcher.d.ts.map +1 -0
- package/dist/reviewer.d.ts +118 -0
- package/dist/reviewer.d.ts.map +1 -0
- package/dist/reviewer.test.d.ts +2 -0
- package/dist/reviewer.test.d.ts.map +1 -0
- package/dist/reward-model-export.d.ts +60 -0
- package/dist/reward-model-export.d.ts.map +1 -0
- package/dist/rl/active-curriculum.d.ts +110 -0
- package/dist/rl/active-curriculum.d.ts.map +1 -0
- package/dist/rl/adaptation-eval.d.ts +109 -0
- package/dist/rl/adaptation-eval.d.ts.map +1 -0
- package/dist/{adversarial-DIVcDoI_.d.ts → rl/adversarial.d.ts} +6 -7
- package/dist/rl/adversarial.d.ts.map +1 -0
- package/dist/rl/compute-curves.d.ts +127 -0
- package/dist/rl/compute-curves.d.ts.map +1 -0
- package/dist/rl/contamination.d.ts +117 -0
- package/dist/rl/contamination.d.ts.map +1 -0
- package/dist/rl/corpus.d.ts +55 -0
- package/dist/rl/corpus.d.ts.map +1 -0
- package/dist/rl/corpus.test.d.ts +2 -0
- package/dist/rl/corpus.test.d.ts.map +1 -0
- package/dist/rl/dataset.d.ts +102 -0
- package/dist/rl/dataset.d.ts.map +1 -0
- package/dist/rl/dataset.test.d.ts +2 -0
- package/dist/rl/dataset.test.d.ts.map +1 -0
- package/dist/rl/exporters.d.ts +141 -0
- package/dist/rl/exporters.d.ts.map +1 -0
- package/dist/rl/index.d.ts +49 -0
- package/dist/rl/index.d.ts.map +1 -0
- package/dist/{off-policy-DiwuKKg7.d.ts → rl/off-policy.d.ts} +8 -9
- package/dist/rl/off-policy.d.ts.map +1 -0
- package/dist/rl/predictive-validity-researcher.d.ts +69 -0
- package/dist/rl/predictive-validity-researcher.d.ts.map +1 -0
- package/dist/rl/preferences.d.ts +141 -0
- package/dist/rl/preferences.d.ts.map +1 -0
- package/dist/rl/process-reward.d.ts +122 -0
- package/dist/rl/process-reward.d.ts.map +1 -0
- package/dist/rl/reward-hacking.d.ts +104 -0
- package/dist/rl/reward-hacking.d.ts.map +1 -0
- package/dist/rl/rl-campaign.d.ts +85 -0
- package/dist/rl/rl-campaign.d.ts.map +1 -0
- package/dist/rl/run-record-adapters.d.ts +56 -0
- package/dist/rl/run-record-adapters.d.ts.map +1 -0
- package/dist/rl/sim-fidelity.d.ts +166 -0
- package/dist/rl/sim-fidelity.d.ts.map +1 -0
- package/dist/rl/sim-fidelity.test.d.ts +2 -0
- package/dist/rl/sim-fidelity.test.d.ts.map +1 -0
- package/dist/rl/tournament.d.ts +115 -0
- package/dist/rl/tournament.d.ts.map +1 -0
- package/dist/rl/verifiable-reward.d.ts +124 -0
- package/dist/rl/verifiable-reward.d.ts.map +1 -0
- package/dist/rl.d.ts +1 -1192
- package/dist/rl.js +615 -615
- package/dist/rl.js.map +1 -1
- package/dist/{run-campaign-RDGAM5KJ.js → run-campaign-WXY7KI67.js} +3 -3
- package/dist/run-critic.d.ts +23 -0
- package/dist/run-critic.d.ts.map +1 -0
- package/dist/run-evidence.d.ts +32 -0
- package/dist/run-evidence.d.ts.map +1 -0
- package/dist/{run-record-e7vj1uZQ.d.ts → run-record.d.ts} +16 -122
- package/dist/run-record.d.ts.map +1 -0
- package/dist/run-record.test.d.ts +2 -0
- package/dist/run-record.test.d.ts.map +1 -0
- package/dist/run-score.d.ts +31 -0
- package/dist/run-score.d.ts.map +1 -0
- package/dist/runtime-trajectory.d.ts +47 -0
- package/dist/runtime-trajectory.d.ts.map +1 -0
- package/dist/{test-graded-scenario-BdVaPyHT.d.ts → sandbox-harness.d.ts} +34 -60
- package/dist/sandbox-harness.d.ts.map +1 -0
- package/dist/sandbox-harness.test.d.ts +2 -0
- package/dist/sandbox-harness.test.d.ts.map +1 -0
- package/dist/sandbox-pool.d.ts +74 -0
- package/dist/sandbox-pool.d.ts.map +1 -0
- package/dist/sandbox-pool.test.d.ts +2 -0
- package/dist/sandbox-pool.test.d.ts.map +1 -0
- package/dist/scorecard.d.ts +133 -0
- package/dist/scorecard.d.ts.map +1 -0
- package/dist/scorecard.test.d.ts +2 -0
- package/dist/scorecard.test.d.ts.map +1 -0
- package/dist/self-play.d.ts +69 -0
- package/dist/self-play.d.ts.map +1 -0
- package/dist/semantic-concept-judge.d.ts +135 -0
- package/dist/semantic-concept-judge.d.ts.map +1 -0
- package/dist/semantic-concept-judge.test.d.ts +2 -0
- package/dist/semantic-concept-judge.test.d.ts.map +1 -0
- package/dist/{sequential-5iSVfzl2.d.ts → sequential.d.ts} +9 -10
- package/dist/sequential.d.ts.map +1 -0
- package/dist/{series-convergence-D5OWMBg6.d.ts → series-convergence.d.ts} +4 -5
- package/dist/series-convergence.d.ts.map +1 -0
- package/dist/slo.d.ts +48 -0
- package/dist/slo.d.ts.map +1 -0
- package/dist/state-continuity.d.ts +47 -0
- package/dist/state-continuity.d.ts.map +1 -0
- package/dist/{statistics-C7PozGrZ.d.ts → statistics.d.ts} +152 -39
- package/dist/statistics.d.ts.map +1 -0
- package/dist/statistics.test.d.ts +2 -0
- package/dist/statistics.test.d.ts.map +1 -0
- package/dist/steering-optimizer.d.ts +58 -0
- package/dist/steering-optimizer.d.ts.map +1 -0
- package/dist/steering.d.ts +24 -0
- package/dist/steering.d.ts.map +1 -0
- package/dist/storyboard/code-edit.d.ts +64 -0
- package/dist/storyboard/code-edit.d.ts.map +1 -0
- package/dist/storyboard/code-edit.test.d.ts +2 -0
- package/dist/storyboard/code-edit.test.d.ts.map +1 -0
- package/dist/storyboard/index.d.ts +16 -81
- package/dist/storyboard/index.d.ts.map +1 -0
- package/dist/storyboard/index.test.d.ts +2 -0
- package/dist/storyboard/index.test.d.ts.map +1 -0
- package/dist/{summary-report-DGmUucwQ.d.ts → summary-report.d.ts} +23 -160
- package/dist/summary-report.d.ts.map +1 -0
- package/dist/telemetry/client.d.ts +35 -0
- package/dist/telemetry/client.d.ts.map +1 -0
- package/dist/telemetry/index.d.ts +17 -35
- package/dist/telemetry/index.d.ts.map +1 -0
- package/dist/telemetry/schema.d.ts +61 -0
- package/dist/telemetry/schema.d.ts.map +1 -0
- package/dist/telemetry/sink-fetch.d.ts +39 -0
- package/dist/telemetry/sink-fetch.d.ts.map +1 -0
- package/dist/telemetry/{file.d.ts → sink-file.d.ts} +5 -7
- package/dist/telemetry/sink-file.d.ts.map +1 -0
- package/dist/test-graded-scenario.d.ts +42 -0
- package/dist/test-graded-scenario.d.ts.map +1 -0
- package/dist/testing.d.ts +5 -0
- package/dist/testing.d.ts.map +1 -0
- package/dist/testing.js +8 -0
- package/dist/testing.js.map +1 -0
- package/dist/tool-use-metrics.d.ts +35 -0
- package/dist/tool-use-metrics.d.ts.map +1 -0
- package/dist/trace/capture-fetch.d.ts +48 -0
- package/dist/trace/capture-fetch.d.ts.map +1 -0
- package/dist/trace/capture-fetch.test.d.ts +2 -0
- package/dist/trace/capture-fetch.test.d.ts.map +1 -0
- package/dist/{emitter-DEZwY14K.d.ts → trace/emitter.d.ts} +9 -12
- package/dist/trace/emitter.d.ts.map +1 -0
- package/dist/trace/extract-usage.d.ts +46 -0
- package/dist/trace/extract-usage.d.ts.map +1 -0
- package/dist/trace/extract-usage.test.d.ts +2 -0
- package/dist/trace/extract-usage.test.d.ts.map +1 -0
- package/dist/trace/index.d.ts +15 -0
- package/dist/trace/index.d.ts.map +1 -0
- package/dist/{integrity-VJ9A7aST.d.ts → trace/integrity.d.ts} +11 -14
- package/dist/trace/integrity.d.ts.map +1 -0
- package/dist/trace/otel-bridge.d.ts +29 -0
- package/dist/trace/otel-bridge.d.ts.map +1 -0
- package/dist/trace/otel-export.d.ts +52 -0
- package/dist/trace/otel-export.d.ts.map +1 -0
- package/dist/trace/otel.d.ts +57 -0
- package/dist/trace/otel.d.ts.map +1 -0
- package/dist/trace/otlp-attributes.d.ts +17 -0
- package/dist/trace/otlp-attributes.d.ts.map +1 -0
- package/dist/trace/query.d.ts +29 -0
- package/dist/trace/query.d.ts.map +1 -0
- package/dist/trace/query.test.d.ts +2 -0
- package/dist/trace/query.test.d.ts.map +1 -0
- package/dist/{raw-provider-sink-C46HDghv.d.ts → trace/raw-provider-sink.d.ts} +13 -14
- package/dist/trace/raw-provider-sink.d.ts.map +1 -0
- package/dist/{redact-B40YG2M_.d.ts → trace/redact.d.ts} +7 -8
- package/dist/trace/redact.d.ts.map +1 -0
- package/dist/{schema-m0gsnbt3.d.ts → trace/schema.d.ts} +29 -30
- package/dist/trace/schema.d.ts.map +1 -0
- package/dist/trace/store-to-otlp.d.ts +72 -0
- package/dist/trace/store-to-otlp.d.ts.map +1 -0
- package/dist/trace/store-to-otlp.test.d.ts +2 -0
- package/dist/trace/store-to-otlp.test.d.ts.map +1 -0
- package/dist/{store-CKUAgsJz.d.ts → trace/store.d.ts} +37 -13
- package/dist/trace/store.d.ts.map +1 -0
- package/dist/trace/store.test.d.ts +2 -0
- package/dist/trace/store.test.d.ts.map +1 -0
- package/dist/{analyst-C8HHvfJp.d.ts → trace-analyst/analyst.d.ts} +12 -14
- package/dist/trace-analyst/analyst.d.ts.map +1 -0
- package/dist/trace-analyst/analyst.test.d.ts +2 -0
- package/dist/trace-analyst/analyst.test.d.ts.map +1 -0
- package/dist/trace-analyst/behavioral-metrics.d.ts +40 -0
- package/dist/trace-analyst/behavioral-metrics.d.ts.map +1 -0
- package/dist/trace-analyst/behavioral-metrics.test.d.ts +2 -0
- package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +1 -0
- package/dist/trace-analyst/hook.d.ts +55 -0
- package/dist/trace-analyst/hook.d.ts.map +1 -0
- package/dist/trace-analyst/index.d.ts +18 -0
- package/dist/trace-analyst/index.d.ts.map +1 -0
- package/dist/trace-analyst/insights.d.ts +71 -0
- package/dist/trace-analyst/insights.d.ts.map +1 -0
- package/dist/trace-analyst/insights.test.d.ts +2 -0
- package/dist/trace-analyst/insights.test.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-flatten.d.ts +42 -0
- package/dist/trace-analyst/otlp-flatten.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-span.d.ts +85 -0
- package/dist/trace-analyst/otlp-span.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-span.test.d.ts +2 -0
- package/dist/trace-analyst/otlp-span.test.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-to-run-records.d.ts +115 -0
- package/dist/trace-analyst/otlp-to-run-records.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts +2 -0
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +2 -0
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +1 -0
- package/dist/trace-analyst/prompts.d.ts +6 -0
- package/dist/trace-analyst/prompts.d.ts.map +1 -0
- package/dist/trace-analyst/store-otlp.d.ts +126 -0
- package/dist/trace-analyst/store-otlp.d.ts.map +1 -0
- package/dist/trace-analyst/store-otlp.test.d.ts +8 -0
- package/dist/trace-analyst/store-otlp.test.d.ts.map +1 -0
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +2 -0
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +1 -0
- package/dist/trace-analyst/store.d.ts +63 -0
- package/dist/trace-analyst/store.d.ts.map +1 -0
- package/dist/trace-analyst/tools.d.ts +44 -0
- package/dist/trace-analyst/tools.d.ts.map +1 -0
- package/dist/trace-analyst/tools.test.d.ts +10 -0
- package/dist/trace-analyst/tools.test.d.ts.map +1 -0
- package/dist/{store-C1YxJDEK.d.ts → trace-analyst/types.d.ts} +18 -74
- package/dist/trace-analyst/types.d.ts.map +1 -0
- package/dist/trace-contracts.d.ts +180 -0
- package/dist/trace-contracts.d.ts.map +1 -0
- package/dist/traced-analyst.d.ts +26 -0
- package/dist/traced-analyst.d.ts.map +1 -0
- package/dist/traced-judges.d.ts +27 -0
- package/dist/traced-judges.d.ts.map +1 -0
- package/dist/traces.d.ts +4 -947
- package/dist/traces.d.ts.map +1 -0
- package/dist/traces.js +42 -12
- package/dist/{trajectory-GEdXJCL5.d.ts → trajectory.d.ts} +8 -9
- package/dist/trajectory.d.ts.map +1 -0
- package/dist/{types-Croy5h7V.d.ts → types.d.ts} +33 -33
- package/dist/types.d.ts.map +1 -0
- package/dist/ui-finding.d.ts +104 -0
- package/dist/ui-finding.d.ts.map +1 -0
- package/dist/verdict-cache.d.ts +78 -0
- package/dist/verdict-cache.d.ts.map +1 -0
- package/dist/verdict-cache.test.d.ts +2 -0
- package/dist/verdict-cache.test.d.ts.map +1 -0
- package/dist/{verdict-C9MlYujm.d.ts → verdict.d.ts} +2 -3
- package/dist/verdict.d.ts.map +1 -0
- package/dist/visual-diff.d.ts +32 -0
- package/dist/visual-diff.d.ts.map +1 -0
- package/dist/wire/handlers.d.ts +54 -0
- package/dist/wire/handlers.d.ts.map +1 -0
- package/dist/wire/index.d.ts +13 -570
- package/dist/wire/index.d.ts.map +1 -0
- package/dist/wire/index.js +2 -2
- package/dist/wire/openapi.d.ts +3 -0
- package/dist/wire/openapi.d.ts.map +1 -0
- package/dist/wire/rpc.d.ts +21 -0
- package/dist/wire/rpc.d.ts.map +1 -0
- package/dist/wire/rubrics.d.ts +34 -0
- package/dist/wire/rubrics.d.ts.map +1 -0
- package/dist/wire/schemas.d.ts +410 -0
- package/dist/wire/schemas.d.ts.map +1 -0
- package/dist/wire/server.d.ts +60 -0
- package/dist/wire/server.d.ts.map +1 -0
- package/dist/workflow/event-schema.d.ts +5 -0
- package/dist/workflow/event-schema.d.ts.map +1 -0
- package/dist/workflow/feedback-pack.d.ts +99 -0
- package/dist/workflow/feedback-pack.d.ts.map +1 -0
- package/dist/workflow/index.d.ts +22 -495
- package/dist/workflow/index.d.ts.map +1 -0
- package/dist/workflow/index.js +2 -2
- package/dist/workflow/intelligence-export.d.ts +62 -0
- package/dist/workflow/intelligence-export.d.ts.map +1 -0
- package/dist/workflow/partner-report.d.ts +49 -0
- package/dist/workflow/partner-report.d.ts.map +1 -0
- package/dist/workflow/phase-graph.d.ts +43 -0
- package/dist/workflow/phase-graph.d.ts.map +1 -0
- package/dist/workflow/promotion-gate.d.ts +61 -0
- package/dist/workflow/promotion-gate.d.ts.map +1 -0
- package/dist/workflow/run-record.d.ts +12 -0
- package/dist/workflow/run-record.d.ts.map +1 -0
- package/dist/workflow/runtime-adapter.d.ts +20 -0
- package/dist/workflow/runtime-adapter.d.ts.map +1 -0
- package/dist/workflow/sanitize.d.ts +21 -0
- package/dist/workflow/sanitize.d.ts.map +1 -0
- package/dist/workflow/schema.d.ts +5 -0
- package/dist/workflow/schema.d.ts.map +1 -0
- package/dist/workflow/summary.d.ts +43 -0
- package/dist/workflow/summary.d.ts.map +1 -0
- package/dist/workflow/trace-event-fields.d.ts +6 -0
- package/dist/workflow/trace-event-fields.d.ts.map +1 -0
- package/dist/workflow/trajectory.d.ts +15 -0
- package/dist/workflow/trajectory.d.ts.map +1 -0
- package/dist/workflow/types.d.ts +68 -0
- package/dist/workflow/types.d.ts.map +1 -0
- package/dist/workspace-inspector.d.ts +67 -0
- package/dist/workspace-inspector.d.ts.map +1 -0
- package/dist/wrangler-deploy-runner.test.d.ts +2 -0
- package/dist/wrangler-deploy-runner.test.d.ts.map +1 -0
- package/docs/campaign-proposers.md +170 -0
- package/docs/concepts.md +8 -4
- package/docs/customer-journeys.md +15 -13
- package/docs/design/loop-taxonomy.md +34 -66
- package/docs/distributed-driver.md +14 -14
- package/docs/feature-guide.md +1 -1
- package/docs/hosted-ingest-spec.md +2 -3
- package/docs/multi-shot-optimization.md +8 -8
- package/docs/product-eval-adoption.md +1 -1
- package/docs/self-improvement-map.md +33 -29
- package/package.json +9 -15
- package/dist/calibration-Cpr3WaX3.d.ts +0 -101
- package/dist/chunk-47X6LRCE.js.map +0 -1
- package/dist/chunk-6SOJM3VR.js.map +0 -1
- package/dist/chunk-CVVHBFGN.js.map +0 -1
- package/dist/chunk-D3V5B42D.js.map +0 -1
- package/dist/chunk-E4GH6USR.js.map +0 -1
- package/dist/chunk-GSH6QNNS.js.map +0 -1
- package/dist/chunk-KWRRMR3J.js.map +0 -1
- package/dist/chunk-LMZQ2Z4U.js.map +0 -1
- package/dist/chunk-QAY5UIJO.js.map +0 -1
- package/dist/chunk-QG2OVF2D.js.map +0 -1
- package/dist/chunk-T375SUOZ.js.map +0 -1
- package/dist/chunk-TWS7AZEY.js.map +0 -1
- package/dist/chunk-UMMZHCPB.js.map +0 -1
- package/dist/chunk-XY4DDNEG.js.map +0 -1
- package/dist/chunk-ZFIBGEOL.js.map +0 -1
- package/dist/control-_Qb7skHX.d.ts +0 -259
- package/dist/corpus-BoR-041R.d.ts +0 -560
- package/dist/failure-cluster-CL7IVgkJ.d.ts +0 -76
- package/dist/harness-optimizer-EnEnQPsr.d.ts +0 -106
- package/dist/index-Bx3gZ8xl.d.ts +0 -159
- package/dist/pre-registration-mAnCugl9.d.ts +0 -523
- package/dist/provenance-LnqRT0sS.d.ts +0 -441
- package/dist/query-CqTxMwDw.d.ts +0 -31
- package/dist/release-report-euXIV_Sk.d.ts +0 -233
- package/dist/run-critic-BAIjX99r.d.ts +0 -56
- package/dist/run-improvement-loop-5z_l5zDz.d.ts +0 -427
- package/dist/runtime-trajectory-BDgfGZSr.d.ts +0 -49
- package/dist/semantic-concept-judge-Dn8Z6KEG.d.ts +0 -624
- package/dist/sink-fetch-B1Yg4Til.d.ts +0 -101
- package/docs/design/external-agent-wedge.md +0 -89
- package/docs/design/phase-d-rfc.md +0 -125
- package/docs/design/phase4-consumer-migration.md +0 -70
- package/docs/design/primitives-integration-spec.md +0 -393
- package/docs/design/product-self-improvement-loop.md +0 -146
- package/docs/design/self-improvement-engine.md +0 -140
- package/docs/design/self-improvement-protocol.md +0 -223
- package/docs/design/self-improvement-roadmap.md +0 -106
- package/docs/design/substrate-gaps.md +0 -118
- package/docs/phase-b-pairing-kit.md +0 -188
- package/docs/phase-b-runbook.md +0 -176
- package/docs/pilot/README.md +0 -62
- package/docs/pilot/customer-checklist.md +0 -90
- package/docs/pilot/integration-foreign-stack.md +0 -296
- package/docs/pilot/integration-tangle-stack.md +0 -248
- package/docs/pilot/one-pager.md +0 -161
- package/docs/pilot/sample-insight-report.json +0 -172
- package/docs/quickstart-external.md +0 -229
- package/docs/research/belief-state-agent-eval-roadmap.md +0 -593
- package/docs/research/research-roadmap.md +0 -205
- package/docs/specs/driver-honest-spec.md +0 -251
- package/docs/specs/hermes-self-improvement-audit.md +0 -93
- package/docs/specs/profile-versioning.md +0 -291
- package/docs/three-package-architecture.md +0 -168
- /package/dist/{chunk-QS3RBQPI.js.map → chunk-D5KBVULY.js.map} +0 -0
- /package/dist/{chunk-UHMJT4T7.js.map → chunk-IZCEK2HR.js.map} +0 -0
- /package/dist/{chunk-L3JOU6XM.js.map → chunk-NACM5RS7.js.map} +0 -0
- /package/dist/{chunk-4FBZZIYD.js.map → chunk-QTMYW64F.js.map} +0 -0
- /package/dist/{chunk-CY6U5S3X.js.map → chunk-URWVKRCS.js.map} +0 -0
- /package/dist/{chunk-Y47J2LJ3.js.map → chunk-Z6L6YSU6.js.map} +0 -0
- /package/dist/{run-campaign-RDGAM5KJ.js.map → run-campaign-WXY7KI67.js.map} +0 -0
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"traces.d.ts","sourceRoot":"","sources":["../src/traces.ts"],"names":[],"mappings":"AAAA,cAAc,UAAU,CAAA;AACxB,cAAc,SAAS,CAAA;AACvB,cAAc,iBAAiB,CAAA"}
|
package/dist/traces.js
CHANGED
|
@@ -28,7 +28,13 @@ import {
|
|
|
28
28
|
scoreTraceInsightReadiness,
|
|
29
29
|
tokenizeDomainWords,
|
|
30
30
|
traceAnalystOnRunComplete
|
|
31
|
-
} from "./chunk-
|
|
31
|
+
} from "./chunk-77T4STFI.js";
|
|
32
|
+
import {
|
|
33
|
+
TRACE_ANALYST_ACTOR_DESCRIPTION,
|
|
34
|
+
TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION,
|
|
35
|
+
TRACE_ANALYST_SUBAGENT_DESCRIPTION,
|
|
36
|
+
analyzeTraces
|
|
37
|
+
} from "./chunk-IZCEK2HR.js";
|
|
32
38
|
import {
|
|
33
39
|
DEFAULT_REDACTION_RULES,
|
|
34
40
|
REDACTION_VERSION,
|
|
@@ -44,7 +50,7 @@ import {
|
|
|
44
50
|
runFailureClass,
|
|
45
51
|
runsForScenario,
|
|
46
52
|
toolSpans
|
|
47
|
-
} from "./chunk-
|
|
53
|
+
} from "./chunk-UFSG7ACU.js";
|
|
48
54
|
import {
|
|
49
55
|
FAILURE_CLASSES,
|
|
50
56
|
TRACE_SCHEMA_VERSION,
|
|
@@ -54,16 +60,24 @@ import {
|
|
|
54
60
|
isSandboxSpan,
|
|
55
61
|
isToolSpan
|
|
56
62
|
} from "./chunk-5BKGXME7.js";
|
|
57
|
-
import {
|
|
58
|
-
TRACE_ANALYST_ACTOR_DESCRIPTION,
|
|
59
|
-
TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION,
|
|
60
|
-
TRACE_ANALYST_SUBAGENT_DESCRIPTION,
|
|
61
|
-
analyzeTraces
|
|
62
|
-
} from "./chunk-UHMJT4T7.js";
|
|
63
63
|
import {
|
|
64
64
|
DEFAULT_TRACE_ANALYST_BUDGETS,
|
|
65
|
+
LLM_CACHED_TOKENS,
|
|
66
|
+
LLM_CACHED_TOKEN_ATTR_KEYS,
|
|
67
|
+
LLM_COST_ATTR_KEYS,
|
|
68
|
+
LLM_COST_USD,
|
|
69
|
+
LLM_INPUT_TOKENS,
|
|
70
|
+
LLM_INPUT_TOKEN_ATTR_KEYS,
|
|
71
|
+
LLM_MODEL_ATTR_KEYS,
|
|
72
|
+
LLM_MODEL_NAME,
|
|
73
|
+
LLM_OUTPUT_TOKENS,
|
|
74
|
+
LLM_OUTPUT_TOKEN_ATTR_KEYS,
|
|
75
|
+
OPENINFERENCE_SPAN_KIND,
|
|
65
76
|
OtlpFileTraceStore,
|
|
77
|
+
SPAN_KIND_ATTR_KEYS,
|
|
66
78
|
SpanNotFoundError,
|
|
79
|
+
TOOL_NAME,
|
|
80
|
+
TOOL_NAME_ATTR_KEYS,
|
|
67
81
|
TRACE_ANALYST_TRUNCATION_MARKER_PREFIX,
|
|
68
82
|
TraceFileMissingError,
|
|
69
83
|
TraceNotFoundError,
|
|
@@ -77,8 +91,9 @@ import {
|
|
|
77
91
|
projectOtlpFlatLine,
|
|
78
92
|
readOtlpStatus,
|
|
79
93
|
stringField,
|
|
80
|
-
traceAnalystFunctionGroup
|
|
81
|
-
|
|
94
|
+
traceAnalystFunctionGroup,
|
|
95
|
+
traceSpanKindToOpenInferenceKind
|
|
96
|
+
} from "./chunk-YBIGNSCZ.js";
|
|
82
97
|
import {
|
|
83
98
|
RunIntegrityError,
|
|
84
99
|
assertRunCaptured,
|
|
@@ -91,7 +106,7 @@ import {
|
|
|
91
106
|
defaultProviderRedactor,
|
|
92
107
|
providerFromBaseUrl
|
|
93
108
|
} from "./chunk-PC4UYEBM.js";
|
|
94
|
-
import "./chunk-
|
|
109
|
+
import "./chunk-LO6IOIJ2.js";
|
|
95
110
|
import {
|
|
96
111
|
TraceEmitter,
|
|
97
112
|
llmSpanFromProvider
|
|
@@ -107,14 +122,28 @@ export {
|
|
|
107
122
|
FileSystemTraceStore,
|
|
108
123
|
InMemoryRawProviderSink,
|
|
109
124
|
InMemoryTraceStore,
|
|
125
|
+
LLM_CACHED_TOKENS,
|
|
126
|
+
LLM_CACHED_TOKEN_ATTR_KEYS,
|
|
127
|
+
LLM_COST_ATTR_KEYS,
|
|
128
|
+
LLM_COST_USD,
|
|
129
|
+
LLM_INPUT_TOKENS,
|
|
130
|
+
LLM_INPUT_TOKEN_ATTR_KEYS,
|
|
131
|
+
LLM_MODEL_ATTR_KEYS,
|
|
132
|
+
LLM_MODEL_NAME,
|
|
133
|
+
LLM_OUTPUT_TOKENS,
|
|
134
|
+
LLM_OUTPUT_TOKEN_ATTR_KEYS,
|
|
110
135
|
NoopRawProviderSink,
|
|
136
|
+
OPENINFERENCE_SPAN_KIND,
|
|
111
137
|
OTEL_AGENT_EVAL_SCOPE,
|
|
112
138
|
OtlpFileTraceStore,
|
|
113
139
|
REDACTION_VERSION,
|
|
114
140
|
ReplayCache,
|
|
115
141
|
ReplayCacheMissError,
|
|
116
142
|
RunIntegrityError,
|
|
143
|
+
SPAN_KIND_ATTR_KEYS,
|
|
117
144
|
SpanNotFoundError,
|
|
145
|
+
TOOL_NAME,
|
|
146
|
+
TOOL_NAME_ATTR_KEYS,
|
|
118
147
|
TRACE_ANALYST_ACTOR_DESCRIPTION,
|
|
119
148
|
TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION,
|
|
120
149
|
TRACE_ANALYST_SUBAGENT_DESCRIPTION,
|
|
@@ -178,6 +207,7 @@ export {
|
|
|
178
207
|
tokenizeDomainWords,
|
|
179
208
|
toolSpans,
|
|
180
209
|
traceAnalystFunctionGroup,
|
|
181
|
-
traceAnalystOnRunComplete
|
|
210
|
+
traceAnalystOnRunComplete,
|
|
211
|
+
traceSpanKindToOpenInferenceKind
|
|
182
212
|
};
|
|
183
213
|
//# sourceMappingURL=traces.js.map
|
|
@@ -1,6 +1,3 @@
|
|
|
1
|
-
import { S as Span, a as TraceEvent } from './schema-m0gsnbt3.js';
|
|
2
|
-
import { T as TraceStore } from './store-CKUAgsJz.js';
|
|
3
|
-
|
|
4
1
|
/**
|
|
5
2
|
* Trajectory — ordered, structured view over a run's spans.
|
|
6
3
|
*
|
|
@@ -10,8 +7,9 @@ import { T as TraceStore } from './store-CKUAgsJz.js';
|
|
|
10
7
|
* analyzers (stuck-loop detection, waste ratio) live in
|
|
11
8
|
* `pipelines/` and consume the trajectory.
|
|
12
9
|
*/
|
|
13
|
-
|
|
14
|
-
|
|
10
|
+
import type { Span, TraceEvent } from './trace/schema';
|
|
11
|
+
import type { TraceStore } from './trace/store';
|
|
12
|
+
export interface TrajectoryStep {
|
|
15
13
|
index: number;
|
|
16
14
|
span: Span;
|
|
17
15
|
/** Depth in the span tree from the root. 0 = top-level. */
|
|
@@ -19,7 +17,7 @@ interface TrajectoryStep {
|
|
|
19
17
|
/** Events attached to this span. */
|
|
20
18
|
events: TraceEvent[];
|
|
21
19
|
}
|
|
22
|
-
interface Trajectory {
|
|
20
|
+
export interface Trajectory {
|
|
23
21
|
runId: string;
|
|
24
22
|
steps: TrajectoryStep[];
|
|
25
23
|
llmTurns: number;
|
|
@@ -28,6 +26,7 @@ interface Trajectory {
|
|
|
28
26
|
retrievals: number;
|
|
29
27
|
totalDurationMs: number;
|
|
30
28
|
}
|
|
31
|
-
declare function buildTrajectory(store: TraceStore, runId: string): Promise<Trajectory>;
|
|
32
|
-
|
|
33
|
-
export
|
|
29
|
+
export declare function buildTrajectory(store: TraceStore, runId: string): Promise<Trajectory>;
|
|
30
|
+
export type { Span, TraceEvent } from './trace/schema';
|
|
31
|
+
export type { TraceStore } from './trace/store';
|
|
32
|
+
//# sourceMappingURL=trajectory.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"trajectory.d.ts","sourceRoot":"","sources":["../src/trajectory.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,gBAAgB,CAAA;AACtD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAE/C,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,MAAM,CAAA;IACb,IAAI,EAAE,IAAI,CAAA;IACV,2DAA2D;IAC3D,KAAK,EAAE,MAAM,CAAA;IACb,oCAAoC;IACpC,MAAM,EAAE,UAAU,EAAE,CAAA;CACrB;AAED,MAAM,WAAW,UAAU;IACzB,KAAK,EAAE,MAAM,CAAA;IACb,KAAK,EAAE,cAAc,EAAE,CAAA;IACvB,QAAQ,EAAE,MAAM,CAAA;IAChB,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;IACrB,UAAU,EAAE,MAAM,CAAA;IAClB,eAAe,EAAE,MAAM,CAAA;CACxB;AAED,wBAAsB,eAAe,CAAC,KAAK,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,UAAU,CAAC,CAgD3F;AAGD,YAAY,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,gBAAgB,CAAA;AACtD,YAAY,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA"}
|
|
@@ -1,6 +1,4 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
3
|
-
interface Scenario {
|
|
1
|
+
export interface Scenario {
|
|
4
2
|
id: string;
|
|
5
3
|
persona: string;
|
|
6
4
|
label: string;
|
|
@@ -10,37 +8,37 @@ interface Scenario {
|
|
|
10
8
|
artifactChecks: ArtifactCheck[];
|
|
11
9
|
systemPromptAppend?: string;
|
|
12
10
|
}
|
|
13
|
-
interface Turn {
|
|
11
|
+
export interface Turn {
|
|
14
12
|
user: string;
|
|
15
13
|
expectedBehaviors: string[];
|
|
16
14
|
adversarial?: boolean;
|
|
17
15
|
feedbackType?: 'correction' | 'rejection' | 'vague' | 'contradictory' | 'escalation';
|
|
18
16
|
}
|
|
19
|
-
interface ArtifactCheck {
|
|
17
|
+
export interface ArtifactCheck {
|
|
20
18
|
type: 'vault_file_exists' | 'vault_file_contains' | 'block_extracted' | 'code_valid' | 'generation_produced' | 'tool_created' | string;
|
|
21
19
|
target: string;
|
|
22
20
|
contains?: string;
|
|
23
21
|
minCount?: number;
|
|
24
22
|
description: string;
|
|
25
23
|
}
|
|
26
|
-
interface JudgeConfig {
|
|
24
|
+
export interface JudgeConfig {
|
|
27
25
|
model: string;
|
|
28
26
|
temperature: number;
|
|
29
27
|
rubric: JudgeRubric;
|
|
30
28
|
}
|
|
31
|
-
interface JudgeRubric {
|
|
29
|
+
export interface JudgeRubric {
|
|
32
30
|
name: string;
|
|
33
31
|
description: string;
|
|
34
32
|
dimensions: RubricDimension[];
|
|
35
33
|
}
|
|
36
|
-
interface RubricDimension {
|
|
34
|
+
export interface RubricDimension {
|
|
37
35
|
name: string;
|
|
38
36
|
description: string;
|
|
39
37
|
anchor_low: string;
|
|
40
38
|
anchor_high: string;
|
|
41
39
|
weight: number;
|
|
42
40
|
}
|
|
43
|
-
interface ScenarioResult {
|
|
41
|
+
export interface ScenarioResult {
|
|
44
42
|
scenarioId: string;
|
|
45
43
|
persona: string;
|
|
46
44
|
turns: TurnResult[];
|
|
@@ -51,7 +49,7 @@ interface ScenarioResult {
|
|
|
51
49
|
totalDurationMs: number;
|
|
52
50
|
artifacts: CollectedArtifacts;
|
|
53
51
|
}
|
|
54
|
-
interface TurnResult {
|
|
52
|
+
export interface TurnResult {
|
|
55
53
|
turnIndex: number;
|
|
56
54
|
userMessage: string;
|
|
57
55
|
agentResponse: string;
|
|
@@ -63,19 +61,19 @@ interface TurnResult {
|
|
|
63
61
|
containsCode: boolean;
|
|
64
62
|
containsToolCall: boolean;
|
|
65
63
|
}
|
|
66
|
-
interface ArtifactResult {
|
|
64
|
+
export interface ArtifactResult {
|
|
67
65
|
check: ArtifactCheck;
|
|
68
66
|
passed: boolean;
|
|
69
67
|
detail?: string;
|
|
70
68
|
}
|
|
71
|
-
interface JudgeScore {
|
|
69
|
+
export interface JudgeScore {
|
|
72
70
|
judgeName: string;
|
|
73
71
|
dimension: string;
|
|
74
72
|
score: number;
|
|
75
73
|
reasoning: string;
|
|
76
74
|
evidence?: string;
|
|
77
75
|
}
|
|
78
|
-
interface CollectedArtifacts {
|
|
76
|
+
export interface CollectedArtifacts {
|
|
79
77
|
vaultFiles: {
|
|
80
78
|
path: string;
|
|
81
79
|
content: string;
|
|
@@ -90,7 +88,7 @@ interface CollectedArtifacts {
|
|
|
90
88
|
}[];
|
|
91
89
|
toolCalls: string[];
|
|
92
90
|
}
|
|
93
|
-
interface BenchmarkReport {
|
|
91
|
+
export interface BenchmarkReport {
|
|
94
92
|
timestamp: string;
|
|
95
93
|
generation: number;
|
|
96
94
|
promptVersion: string;
|
|
@@ -119,7 +117,7 @@ interface BenchmarkReport {
|
|
|
119
117
|
}[];
|
|
120
118
|
};
|
|
121
119
|
}
|
|
122
|
-
interface RouteMap {
|
|
120
|
+
export interface RouteMap {
|
|
123
121
|
signup?: string;
|
|
124
122
|
login?: string;
|
|
125
123
|
workspaces?: string;
|
|
@@ -132,11 +130,13 @@ interface RouteMap {
|
|
|
132
130
|
generations?: string;
|
|
133
131
|
[key: string]: string | undefined;
|
|
134
132
|
}
|
|
135
|
-
interface ProductClientConfig {
|
|
133
|
+
export interface ProductClientConfig {
|
|
136
134
|
baseUrl: string;
|
|
137
135
|
routes: RouteMap;
|
|
136
|
+
/** Per-request timeout in ms before the request is aborted. Default 30s. */
|
|
137
|
+
timeoutMs?: number;
|
|
138
138
|
}
|
|
139
|
-
interface ScenarioFile {
|
|
139
|
+
export interface ScenarioFile {
|
|
140
140
|
id: string;
|
|
141
141
|
category: string;
|
|
142
142
|
persona: string;
|
|
@@ -153,12 +153,12 @@ interface ScenarioFile {
|
|
|
153
153
|
turns: Turn[];
|
|
154
154
|
artifactChecks: ArtifactCheck[];
|
|
155
155
|
}
|
|
156
|
-
interface CompletionCriterion {
|
|
156
|
+
export interface CompletionCriterion {
|
|
157
157
|
name: string;
|
|
158
158
|
check: (state: DriverState) => boolean;
|
|
159
159
|
progress?: (state: DriverState) => number;
|
|
160
160
|
}
|
|
161
|
-
interface FeedbackPattern {
|
|
161
|
+
export interface FeedbackPattern {
|
|
162
162
|
trigger: string;
|
|
163
163
|
response: string;
|
|
164
164
|
}
|
|
@@ -170,8 +170,8 @@ interface FeedbackPattern {
|
|
|
170
170
|
* relentless — senior partner reviewing for a client who will litigate;
|
|
171
171
|
* interrogates every claim, accepts nothing undefended.
|
|
172
172
|
*/
|
|
173
|
-
type PersonaRigor = 'cooperative' | 'demanding' | 'relentless';
|
|
174
|
-
interface PersonaConfig {
|
|
173
|
+
export type PersonaRigor = 'cooperative' | 'demanding' | 'relentless';
|
|
174
|
+
export interface PersonaConfig {
|
|
175
175
|
id: string;
|
|
176
176
|
role: string;
|
|
177
177
|
goal: string;
|
|
@@ -202,7 +202,7 @@ interface PersonaConfig {
|
|
|
202
202
|
*/
|
|
203
203
|
curveballs?: string[];
|
|
204
204
|
}
|
|
205
|
-
interface DriverState {
|
|
205
|
+
export interface DriverState {
|
|
206
206
|
tasks: number;
|
|
207
207
|
events: number;
|
|
208
208
|
proposals: {
|
|
@@ -214,7 +214,7 @@ interface DriverState {
|
|
|
214
214
|
codeBlocks: number;
|
|
215
215
|
generations: number;
|
|
216
216
|
}
|
|
217
|
-
interface TurnMetrics {
|
|
217
|
+
export interface TurnMetrics {
|
|
218
218
|
turn: number;
|
|
219
219
|
timestamp: string;
|
|
220
220
|
tasks: number;
|
|
@@ -236,7 +236,7 @@ interface TurnMetrics {
|
|
|
236
236
|
totalCostUsd: number;
|
|
237
237
|
completionPercent: number;
|
|
238
238
|
}
|
|
239
|
-
interface DriverResult {
|
|
239
|
+
export interface DriverResult {
|
|
240
240
|
personaId: string;
|
|
241
241
|
/** True when the simulated user professionally signed off (driver said DONE). */
|
|
242
242
|
completed: boolean;
|
|
@@ -255,7 +255,7 @@ interface DriverResult {
|
|
|
255
255
|
totalCostUsd: number;
|
|
256
256
|
finalQualityScore: number | null;
|
|
257
257
|
}
|
|
258
|
-
interface BenchmarkRunnerConfig {
|
|
258
|
+
export interface BenchmarkRunnerConfig {
|
|
259
259
|
scenarios: Scenario[];
|
|
260
260
|
judges: JudgeFn[];
|
|
261
261
|
systemPrompt: string;
|
|
@@ -265,32 +265,32 @@ interface BenchmarkRunnerConfig {
|
|
|
265
265
|
generation?: number;
|
|
266
266
|
promptVersion?: string;
|
|
267
267
|
}
|
|
268
|
-
interface JudgeInput {
|
|
268
|
+
export interface JudgeInput {
|
|
269
269
|
scenario: Scenario;
|
|
270
270
|
turns: TurnResult[];
|
|
271
271
|
artifacts: CollectedArtifacts;
|
|
272
272
|
}
|
|
273
|
-
type JudgeFn = (tc: TCloud, input: JudgeInput) => Promise<JudgeScore[]>;
|
|
274
|
-
|
|
275
|
-
|
|
273
|
+
export type JudgeFn = (tc: TCloud, input: JudgeInput) => Promise<JudgeScore[]>;
|
|
274
|
+
import type { TCloud } from '@tangle-network/tcloud';
|
|
275
|
+
export type { TCloud };
|
|
276
|
+
export interface TestResult {
|
|
276
277
|
name: string;
|
|
277
278
|
passed: boolean;
|
|
278
279
|
duration: number;
|
|
279
280
|
detail?: string;
|
|
280
281
|
checks: CheckResult[];
|
|
281
282
|
}
|
|
282
|
-
interface CheckResult {
|
|
283
|
+
export interface CheckResult {
|
|
283
284
|
name: string;
|
|
284
285
|
passed: boolean;
|
|
285
286
|
expected: string;
|
|
286
287
|
actual: string;
|
|
287
288
|
}
|
|
288
|
-
interface EvalResult {
|
|
289
|
+
export interface EvalResult {
|
|
289
290
|
scenario: string;
|
|
290
291
|
status: 'pass' | 'fail' | 'skip';
|
|
291
292
|
duration: number;
|
|
292
293
|
detail?: string;
|
|
293
294
|
artifact?: string;
|
|
294
295
|
}
|
|
295
|
-
|
|
296
|
-
export type { ArtifactCheck as A, BenchmarkRunnerConfig as B, CheckResult as C, DriverResult as D, EvalResult as E, FeedbackPattern as F, JudgeScore as J, ProductClientConfig as P, RouteMap as R, Scenario as S, TestResult as T, JudgeInput as a, JudgeFn as b, BenchmarkReport as c, PersonaConfig as d, DriverState as e, CollectedArtifacts as f, ScenarioResult as g, TurnMetrics as h, ScenarioFile as i, CompletionCriterion as j, ArtifactResult as k, JudgeConfig as l, JudgeRubric as m, PersonaRigor as n, RubricDimension as o, Turn as p, TurnResult as q };
|
|
296
|
+
//# sourceMappingURL=types.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAEA,MAAM,WAAW,QAAQ;IACvB,EAAE,EAAE,MAAM,CAAA;IACV,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,UAAU,EAAE,MAAM,EAAE,CAAA;IACpB,KAAK,EAAE,IAAI,EAAE,CAAA;IACb,cAAc,EAAE,aAAa,EAAE,CAAA;IAC/B,kBAAkB,CAAC,EAAE,MAAM,CAAA;CAC5B;AAED,MAAM,WAAW,IAAI;IACnB,IAAI,EAAE,MAAM,CAAA;IACZ,iBAAiB,EAAE,MAAM,EAAE,CAAA;IAC3B,WAAW,CAAC,EAAE,OAAO,CAAA;IACrB,YAAY,CAAC,EAAE,YAAY,GAAG,WAAW,GAAG,OAAO,GAAG,eAAe,GAAG,YAAY,CAAA;CACrF;AAID,MAAM,WAAW,aAAa;IAC5B,IAAI,EACA,mBAAmB,GACnB,qBAAqB,GACrB,iBAAiB,GACjB,YAAY,GACZ,qBAAqB,GACrB,cAAc,GACd,MAAM,CAAA;IACV,MAAM,EAAE,MAAM,CAAA;IACd,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,WAAW,EAAE,MAAM,CAAA;CACpB;AAID,MAAM,WAAW,WAAW;IAC1B,KAAK,EAAE,MAAM,CAAA;IACb,WAAW,EAAE,MAAM,CAAA;IACnB,MAAM,EAAE,WAAW,CAAA;CACpB;AAED,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,eAAe,EAAE,CAAA;CAC9B;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAA;IACZ,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,MAAM,CAAA;IACnB,MAAM,EAAE,MAAM,CAAA;CACf;AAID,MAAM,WAAW,cAAc;IAC7B,UAAU,EAAE,MAAM,CAAA;IAClB,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,UAAU,EAAE,CAAA;IACnB,eAAe,EAAE,cAAc,EAAE,CAAA;IACjC,WAAW,EAAE,UAAU,EAAE,CAAA;IACzB,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,eAAe,EAAE,MAAM,CAAA;IACvB,SAAS,EAAE,kBAAkB,CAAA;CAC9B;AAED,MAAM,WAAW,UAAU;IACzB,SAAS,EAAE,MAAM,CAAA;IACjB,WAAW,EAAE,MAAM,CAAA;IACnB,aAAa,EAAE,MAAM,CAAA;IACrB,UAAU,EAAE,MAAM,CAAA;IAClB,eAAe,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,EAAE,CAAA;IAClD,YAAY,EAAE,OAAO,CAAA;IACrB,gBAAgB,EAAE,OAAO,CAAA;CAC1B;AAED,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,aAAa,CAAA;IACpB,MAAM,EAAE,OAAO,CAAA;IACf,MAAM,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,UAAU;IACzB,SAAS,EAAE,MAAM,CAAA;IACjB,SAAS,EAAE,MAAM,CAAA;IACjB,KAAK,EAAE,MAAM,CAAA;IACb,SAAS,EAAE,MAAM,CAAA;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB;IACjC,UAAU,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,EAAE,CAAA;IAC/C,eAAe,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAAE,EAAE,CAAA;IACnE,UAAU,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,EAAE,CAAA;IAChD,SAAS,EAAE,MAAM,EAAE,CAAA;CACpB;AAID,MAAM,WAAW,eAAe;IAC9B,SAAS,EAAE,MAAM,CAAA;IACjB,UAAU,EAAE,MAAM,CAAA;IAClB,aAAa,EAAE,MAAM,CAAA;IACrB,aAAa,EAAE,MAAM,CAAA;IACrB,OAAO,EAAE,cAAc,EAAE,CAAA;IACzB,OAAO,EAAE;QACP,UAAU,EAAE,MAAM,CAAA;QAClB,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE;YAAE,GAAG,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAA;SAAE,CAAC,CAAA;QACzE,WAAW,EAAE,MAAM,CAAC,MAAM,EAAE;YAAE,GAAG,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,EAAE,CAAA;SAAE,CAAC,CAAA;QAC9D,OAAO,EAAE;YAAE,QAAQ,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAA;SAAE,EAAE,CAAA;QAC9D,SAAS,EAAE;YAAE,QAAQ,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAA;SAAE,EAAE,CAAA;KACjE,CAAA;CACF;AAID,MAAM,WAAW,QAAQ;IACvB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAAA;CAClC;AAED,MAAM,WAAW,mBAAmB;IAClC,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,QAAQ,CAAA;IAChB,4EAA4E;IAC5E,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAID,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAA;IACV,QAAQ,EAAE,MAAM,CAAA;IAChB,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,CAAC,EAAE,OAAO,CAAA;IACnB,MAAM,CAAC,EAAE;QACP,UAAU,EAAE;YACV,IAAI,EAAE,MAAM,CAAA;YACZ,WAAW,EAAE,MAAM,CAAA;YACnB,MAAM,EAAE,MAAM,CAAA;SACf,EAAE,CAAA;KACJ,CAAA;IACD,KAAK,EAAE,IAAI,EAAE,CAAA;IACb,cAAc,EAAE,aAAa,EAAE,CAAA;CAChC;AAID,MAAM,WAAW,mBAAmB;IAClC,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,CAAC,KAAK,EAAE,WAAW,KAAK,OAAO,CAAA;IACtC,QAAQ,CAAC,EAAE,CAAC,KAAK,EAAE,WAAW,KAAK,MAAM,CAAA;CAC1C;AAED,MAAM,WAAW,eAAe;IAC9B,OAAO,EAAE,MAAM,CAAA;IACf,QAAQ,EAAE,MAAM,CAAA;CACjB;AAED;;;;;;;GAOG;AACH,MAAM,MAAM,YAAY,GAAG,aAAa,GAAG,WAAW,GAAG,YAAY,CAAA;AAErE,MAAM,WAAW,aAAa;IAC5B,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,MAAM,CAAA;IACZ,kBAAkB,EAAE,mBAAmB,EAAE,CAAA;IACzC,gBAAgB,CAAC,EAAE,eAAe,EAAE,CAAA;IACpC,QAAQ,EAAE,MAAM,CAAA;IAChB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,sEAAsE;IACtE,KAAK,CAAC,EAAE,YAAY,CAAA;IACpB;;;;;OAKG;IACH,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB;;;;;OAKG;IACH,cAAc,CAAC,EAAE,MAAM,EAAE,CAAA;IACzB;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,EAAE,CAAA;CACtB;AAED,MAAM,WAAW,WAAW;IAC1B,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAA;IAClE,UAAU,EAAE,MAAM,EAAE,CAAA;IACpB,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,MAAM,CAAA;IACjB,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAA;IAClE,UAAU,EAAE,MAAM,CAAA;IAClB,iBAAiB,EAAE,MAAM,CAAA;IACzB,aAAa,EAAE,MAAM,CAAA;IACrB,kBAAkB,EAAE,MAAM,CAAA;IAC1B,eAAe,EAAE,MAAM,CAAA;IACvB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,gBAAgB,EAAE,MAAM,CAAA;IACxB,YAAY,EAAE,MAAM,CAAA;IACpB,iBAAiB,EAAE,MAAM,CAAA;CAC1B;AAED,MAAM,WAAW,YAAY;IAC3B,SAAS,EAAE,MAAM,CAAA;IACjB,iFAAiF;IACjF,SAAS,EAAE,OAAO,CAAA;IAClB,4EAA4E;IAC5E,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAA;IAChC;;;;OAIG;IACH,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAA;IAChC,UAAU,EAAE,MAAM,CAAA;IAClB,OAAO,EAAE,WAAW,EAAE,CAAA;IACtB,UAAU,EAAE,WAAW,CAAA;IACvB,gBAAgB,EAAE,MAAM,EAAE,CAAA;IAC1B,YAAY,EAAE,MAAM,CAAA;IACpB,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAA;CACjC;AAID,MAAM,WAAW,qBAAqB;IACpC,SAAS,EAAE,QAAQ,EAAE,CAAA;IACrB,MAAM,EAAE,OAAO,EAAE,CAAA;IACjB,YAAY,EAAE,MAAM,CAAA;IACpB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,aAAa,CAAC,EAAE,MAAM,CAAA;CACvB;AAED,MAAM,WAAW,UAAU;IACzB,QAAQ,EAAE,QAAQ,CAAA;IAClB,KAAK,EAAE,UAAU,EAAE,CAAA;IACnB,SAAS,EAAE,kBAAkB,CAAA;CAC9B;AAED,MAAM,MAAM,OAAO,GAAG,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,UAAU,KAAK,OAAO,CAAC,UAAU,EAAE,CAAC,CAAA;AAG9E,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,wBAAwB,CAAA;AAEpD,YAAY,EAAE,MAAM,EAAE,CAAA;AAItB,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,OAAO,CAAA;IACf,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,WAAW,EAAE,CAAA;CACtB;AAED,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,OAAO,CAAA;IACf,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,UAAU;IACzB,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAAA;IAChC,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB"}
|
|
@@ -0,0 +1,104 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* UI audit finding — substrate primitive for "what is wrong with the UI?"
|
|
3
|
+
*
|
|
4
|
+
* Used by:
|
|
5
|
+
* - `@tangle-network/agent-runtime` (ui-auditor profile + delegate) —
|
|
6
|
+
* produced as the canonical output of an audit iteration, persisted to
|
|
7
|
+
* disk as GitHub-issue Markdown, surfaced over MCP.
|
|
8
|
+
* - Downstream ship gates / dashboards / analyst consumers — load and
|
|
9
|
+
* transform findings without depending on the runtime.
|
|
10
|
+
*
|
|
11
|
+
* Repo layering: agent-eval is the substrate (no upward deps). Consumers
|
|
12
|
+
* read this type from here; the reverse is forbidden. See CLAUDE.md
|
|
13
|
+
* "Repo layering" for the rule. A UI finding makes sense WITHOUT a running
|
|
14
|
+
* agent loop (you can load a saved finding, ship-gate against a set of
|
|
15
|
+
* them, render them in a dashboard), which puts it firmly in substrate.
|
|
16
|
+
*
|
|
17
|
+
* The shape is intentionally minimal — runtime-shaped state (capture
|
|
18
|
+
* timestamps, OTel trace IDs, sandbox placement) lives on auxiliary
|
|
19
|
+
* runtime types in `agent-runtime`, not on the finding itself.
|
|
20
|
+
*/
|
|
21
|
+
/**
|
|
22
|
+
* Canonical audit lenses. Each lens scopes a finding to a single class of
|
|
23
|
+
* problem so a single audit pass can iterate them without pile-on findings
|
|
24
|
+
* under a generic label.
|
|
25
|
+
*
|
|
26
|
+
* Naming is fixed for cross-package wire compatibility. Treat additions as
|
|
27
|
+
* a substrate-level decision — analysts, gates, and writers all branch on
|
|
28
|
+
* the lens.
|
|
29
|
+
*/
|
|
30
|
+
export type UiLens = 'consistency' | 'hierarchy' | 'layout' | 'ux-flow' | 'duplication' | 'accessibility' | 'responsive' | 'states' | 'content' | 'interaction' | 'performance-perceived' | 'other';
|
|
31
|
+
/** Frozen tuple of lenses for validation + iteration. */
|
|
32
|
+
export declare const UI_LENSES: readonly UiLens[];
|
|
33
|
+
/**
|
|
34
|
+
* Severity scale — intentionally narrow.
|
|
35
|
+
*
|
|
36
|
+
* - `critical` — blocks a core task or is an accessibility blocker.
|
|
37
|
+
* - `high` — confusing, broken-looking, or noticeable friction.
|
|
38
|
+
* - `med` — visible polish issue, would be caught in code review.
|
|
39
|
+
* - `low` — nitpick worth fixing eventually.
|
|
40
|
+
*/
|
|
41
|
+
export type UiFindingSeverity = 'low' | 'med' | 'high' | 'critical';
|
|
42
|
+
/** Frozen severity tuple, ordered worst → least bad for sort/report. */
|
|
43
|
+
export declare const UI_FINDING_SEVERITIES: readonly UiFindingSeverity[];
|
|
44
|
+
/**
|
|
45
|
+
* Pointer to a screenshot referenced by the finding. The path is
|
|
46
|
+
* intentionally a relative string (relative to the audit workspace root)
|
|
47
|
+
* so findings remain portable across machines and into GitHub issues.
|
|
48
|
+
*/
|
|
49
|
+
export interface UiFindingScreenshot {
|
|
50
|
+
/** Workspace-relative path to the screenshot file (e.g. `screenshots/home--1280x800--...png`). */
|
|
51
|
+
path: string;
|
|
52
|
+
/** Optional viewport the screenshot was taken at, e.g. `1280x800`. */
|
|
53
|
+
viewport?: string;
|
|
54
|
+
/** Optional short label that disambiguates multiple captures of the same surface (e.g. `t0`, `step-1`). */
|
|
55
|
+
label?: string;
|
|
56
|
+
}
|
|
57
|
+
/**
|
|
58
|
+
* A single UI audit finding — the unit of work a contributor can act on.
|
|
59
|
+
*
|
|
60
|
+
* Every field except the documented optionals is required. The shape is
|
|
61
|
+
* deliberately constraining: a finding without a screenshot, a lens, a
|
|
62
|
+
* concrete title, and a suggested fix is not actionable, and the auditor
|
|
63
|
+
* validator hard-fails on those gaps.
|
|
64
|
+
*/
|
|
65
|
+
export interface UiFinding {
|
|
66
|
+
/**
|
|
67
|
+
* Stable identifier within a single audit workspace. Monotonically
|
|
68
|
+
* increasing integer (1, 2, …) assigned by the writer when persisting.
|
|
69
|
+
* Optional in transit (before persistence) — undefined on freshly minted
|
|
70
|
+
* findings emitted from a loop iteration.
|
|
71
|
+
*/
|
|
72
|
+
id?: number;
|
|
73
|
+
/** Concrete title — names the offending element AND what's wrong. */
|
|
74
|
+
title: string;
|
|
75
|
+
/** Lens this finding belongs to. */
|
|
76
|
+
lens: UiLens;
|
|
77
|
+
/** Severity. */
|
|
78
|
+
severity: UiFindingSeverity;
|
|
79
|
+
/** Logical route the finding was observed on (e.g. `home`, `checkout-step-2`). */
|
|
80
|
+
route: string;
|
|
81
|
+
/** Fully qualified URL the finding was observed at. */
|
|
82
|
+
url?: string;
|
|
83
|
+
/** Viewport string the offending capture was taken at (e.g. `1280x800`). */
|
|
84
|
+
viewport?: string;
|
|
85
|
+
/** CSS selector pinning the offending element, when one can be identified. */
|
|
86
|
+
selector?: string;
|
|
87
|
+
/** 1–3 sentences describing what the screenshot shows that is wrong. */
|
|
88
|
+
observation: string;
|
|
89
|
+
/** Who is affected and how. Concrete user impact. */
|
|
90
|
+
impact: string;
|
|
91
|
+
/** A specific change a contributor could apply without asking back. */
|
|
92
|
+
suggestedFix: string;
|
|
93
|
+
/** Optional explicit reproduction steps. Writer synthesizes from route/url/selector when omitted. */
|
|
94
|
+
reproSteps?: string;
|
|
95
|
+
/** Free-form tags. */
|
|
96
|
+
tags?: readonly string[];
|
|
97
|
+
/** Screenshot references — required to be non-empty for actionable findings. */
|
|
98
|
+
screenshots: readonly UiFindingScreenshot[];
|
|
99
|
+
/** Cross-references to similar findings already on file, by id. */
|
|
100
|
+
similarTo?: readonly number[];
|
|
101
|
+
/** ISO-8601 creation timestamp set by the writer when persisted. */
|
|
102
|
+
createdAt?: string;
|
|
103
|
+
}
|
|
104
|
+
//# sourceMappingURL=ui-finding.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"ui-finding.d.ts","sourceRoot":"","sources":["../src/ui-finding.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH;;;;;;;;GAQG;AACH,MAAM,MAAM,MAAM,GACd,aAAa,GACb,WAAW,GACX,QAAQ,GACR,SAAS,GACT,aAAa,GACb,eAAe,GACf,YAAY,GACZ,QAAQ,GACR,SAAS,GACT,aAAa,GACb,uBAAuB,GACvB,OAAO,CAAA;AAEX,yDAAyD;AACzD,eAAO,MAAM,SAAS,EAAE,SAAS,MAAM,EAa7B,CAAA;AAEV;;;;;;;GAOG;AACH,MAAM,MAAM,iBAAiB,GAAG,KAAK,GAAG,KAAK,GAAG,MAAM,GAAG,UAAU,CAAA;AAEnE,wEAAwE;AACxE,eAAO,MAAM,qBAAqB,EAAE,SAAS,iBAAiB,EAKpD,CAAA;AAEV;;;;GAIG;AACH,MAAM,WAAW,mBAAmB;IAClC,kGAAkG;IAClG,IAAI,EAAE,MAAM,CAAA;IACZ,sEAAsE;IACtE,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,2GAA2G;IAC3G,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,SAAS;IACxB;;;;;OAKG;IACH,EAAE,CAAC,EAAE,MAAM,CAAA;IACX,qEAAqE;IACrE,KAAK,EAAE,MAAM,CAAA;IACb,oCAAoC;IACpC,IAAI,EAAE,MAAM,CAAA;IACZ,gBAAgB;IAChB,QAAQ,EAAE,iBAAiB,CAAA;IAC3B,kFAAkF;IAClF,KAAK,EAAE,MAAM,CAAA;IACb,uDAAuD;IACvD,GAAG,CAAC,EAAE,MAAM,CAAA;IACZ,4EAA4E;IAC5E,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,8EAA8E;IAC9E,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,wEAAwE;IACxE,WAAW,EAAE,MAAM,CAAA;IACnB,qDAAqD;IACrD,MAAM,EAAE,MAAM,CAAA;IACd,uEAAuE;IACvE,YAAY,EAAE,MAAM,CAAA;IACpB,qGAAqG;IACrG,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,sBAAsB;IACtB,IAAI,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;IACxB,gFAAgF;IAChF,WAAW,EAAE,SAAS,mBAAmB,EAAE,CAAA;IAC3C,mEAAmE;IACnE,SAAS,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;IAC7B,oEAAoE;IACpE,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB"}
|
|
@@ -0,0 +1,78 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Content-addressed judge-verdict caching.
|
|
3
|
+
*
|
|
4
|
+
* LAW: cache JUDGE VERDICTS only — judging the same artifact with the same
|
|
5
|
+
* judge+rubric is pure. NEVER cache agent rollouts. (A router that cached
|
|
6
|
+
* identical fanout prompts silently destroyed best-of-N diversity; rollout
|
|
7
|
+
* caching reintroduces that failure class. Judging has no diversity to
|
|
8
|
+
* destroy — same artifact + same rubric ⇒ same verdict is the desired
|
|
9
|
+
* property, not a bug.)
|
|
10
|
+
*
|
|
11
|
+
* The cache key is a sha-256 over the canonical JSON of everything that can
|
|
12
|
+
* change a verdict: the artifact content, the scenario id, the judge name,
|
|
13
|
+
* the full dimension list (key + description — the description IS the rubric
|
|
14
|
+
* text shown to the judge), and a caller-supplied `judgeVersion`.
|
|
15
|
+
* `judgeVersion` is REQUIRED: a judge whose prompt/model/ensemble changes
|
|
16
|
+
* without a version bump would otherwise silently serve stale verdicts.
|
|
17
|
+
*
|
|
18
|
+
* Strict canonicalization (`canonicalJson`) throws on undefined / function /
|
|
19
|
+
* symbol / non-finite numbers — an artifact that cannot be unambiguously
|
|
20
|
+
* serialized cannot be content-addressed, and coercing it would let two
|
|
21
|
+
* different artifacts collide on one key.
|
|
22
|
+
*/
|
|
23
|
+
import type { JudgeConfig, JudgeScore, Scenario } from './campaign/types';
|
|
24
|
+
/**
|
|
25
|
+
* Stable JSON stringify: object keys sorted recursively, so two semantically
|
|
26
|
+
* equal values produce byte-identical output regardless of key insertion
|
|
27
|
+
* order. Throws on undefined / function / symbol / NaN / ±Infinity / bigint /
|
|
28
|
+
* Map / Set — anything JSON.stringify would coerce or drop silently.
|
|
29
|
+
*
|
|
30
|
+
* Distinct from `pre-registration.ts`'s `canonicalize`/`hashJson`, which are
|
|
31
|
+
* permissive (coercion allowed) and async (web-crypto). Use THIS pair when a
|
|
32
|
+
* hash collision or silent coercion would corrupt a cache key or attestation.
|
|
33
|
+
*/
|
|
34
|
+
export declare function canonicalJson(value: unknown): string;
|
|
35
|
+
/** Hex sha-256 over `canonicalJson(value)`. The content address used by the
|
|
36
|
+
* verdict cache and report attestation. */
|
|
37
|
+
export declare function contentHash(value: unknown): string;
|
|
38
|
+
/** Pluggable verdict store. Sync or async on both legs — `cachedJudge`
|
|
39
|
+
* awaits the results either way. */
|
|
40
|
+
export interface VerdictCacheStore {
|
|
41
|
+
get(key: string): Promise<JudgeScore | undefined> | JudgeScore | undefined;
|
|
42
|
+
set(key: string, score: JudgeScore): Promise<void> | void;
|
|
43
|
+
}
|
|
44
|
+
/** Process-local Map-backed store. */
|
|
45
|
+
export declare function inMemoryVerdictCache(): VerdictCacheStore;
|
|
46
|
+
/**
|
|
47
|
+
* JSONL-file-backed store: the full file is loaded into an in-memory index at
|
|
48
|
+
* construction; every `set` appends one line synchronously (durable before
|
|
49
|
+
* the verdict is returned). A corrupt or malformed line throws at load with
|
|
50
|
+
* file:line — a skipped line would silently re-judge (cost) or, worse, mask
|
|
51
|
+
* a half-written file that needs operator attention.
|
|
52
|
+
*/
|
|
53
|
+
export declare function fileVerdictCache(path: string): VerdictCacheStore;
|
|
54
|
+
export interface VerdictCacheStats {
|
|
55
|
+
hits: number;
|
|
56
|
+
misses: number;
|
|
57
|
+
}
|
|
58
|
+
export interface CachedJudgeOptions {
|
|
59
|
+
/** REQUIRED — part of the cache key. Bump on any change to the judge's
|
|
60
|
+
* prompt, model, ensemble, or scoring logic; silent judge upgrades must
|
|
61
|
+
* never serve stale verdicts. */
|
|
62
|
+
judgeVersion: string;
|
|
63
|
+
}
|
|
64
|
+
/** The wrapped judge: same `JudgeConfig` seam, plus hit/miss observability. */
|
|
65
|
+
export type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig<TArtifact, TScenario> & {
|
|
66
|
+
stats(): VerdictCacheStats;
|
|
67
|
+
};
|
|
68
|
+
/**
|
|
69
|
+
* Wrap a `JudgeConfig` so repeat judgments of the same artifact are served
|
|
70
|
+
* from the store instead of re-invoking `score()`. The wrapper is generic
|
|
71
|
+
* over the judge's own type parameters and preserves `appliesTo` — it is a
|
|
72
|
+
* drop-in replacement anywhere a `JudgeConfig` is accepted.
|
|
73
|
+
*
|
|
74
|
+
* A judge that throws is NOT cached: the error propagates and the next
|
|
75
|
+
* attempt re-judges (caching a failure would pin a transient outage forever).
|
|
76
|
+
*/
|
|
77
|
+
export declare function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(judge: JudgeConfig<TArtifact, TScenario>, store: VerdictCacheStore, options: CachedJudgeOptions): CachedJudge<TArtifact, TScenario>;
|
|
78
|
+
//# sourceMappingURL=verdict-cache.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"verdict-cache.d.ts","sourceRoot":"","sources":["../src/verdict-cache.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAIH,OAAO,KAAK,EAAE,WAAW,EAAE,UAAU,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAA;AAgDzE;;;;;;;;;GASG;AACH,wBAAgB,aAAa,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAEpD;AAED;4CAC4C;AAC5C,wBAAgB,WAAW,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAElD;AAID;qCACqC;AACrC,MAAM,WAAW,iBAAiB;IAChC,GAAG,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC,UAAU,GAAG,SAAS,CAAC,GAAG,UAAU,GAAG,SAAS,CAAA;IAC1E,GAAG,CAAC,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,IAAI,CAAC,GAAG,IAAI,CAAA;CAC1D;AAED,sCAAsC;AACtC,wBAAgB,oBAAoB,IAAI,iBAAiB,CAQxD;AAiCD;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,GAAG,iBAAiB,CAkBhE;AAID,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,kBAAkB;IACjC;;sCAEkC;IAClC,YAAY,EAAE,MAAM,CAAA;CACrB;AAED,+EAA+E;AAC/E,MAAM,MAAM,WAAW,CAAC,SAAS,EAAE,SAAS,SAAS,QAAQ,GAAG,QAAQ,IAAI,WAAW,CACrF,SAAS,EACT,SAAS,CACV,GAAG;IACF,KAAK,IAAI,iBAAiB,CAAA;CAC3B,CAAA;AAED;;;;;;;;GAQG;AACH,wBAAgB,WAAW,CAAC,SAAS,EAAE,SAAS,SAAS,QAAQ,GAAG,QAAQ,EAC1E,KAAK,EAAE,WAAW,CAAC,SAAS,EAAE,SAAS,CAAC,EACxC,KAAK,EAAE,iBAAiB,EACxB,OAAO,EAAE,kBAAkB,GAC1B,WAAW,CAAC,SAAS,EAAE,SAAS,CAAC,CA8BnC"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"verdict-cache.test.d.ts","sourceRoot":"","sources":["../src/verdict-cache.test.ts"],"names":[],"mappings":""}
|
|
@@ -21,7 +21,7 @@
|
|
|
21
21
|
* Need structured extras? Extend DefaultVerdict with typed fields — never
|
|
22
22
|
* serialize extras into `notes`.
|
|
23
23
|
*/
|
|
24
|
-
interface DefaultVerdict {
|
|
24
|
+
export interface DefaultVerdict {
|
|
25
25
|
/** Whether the output meets the validator's pass criteria. */
|
|
26
26
|
valid: boolean;
|
|
27
27
|
/** Aggregate score in [0, 1]. Drivers use this for winner selection. */
|
|
@@ -31,5 +31,4 @@ interface DefaultVerdict {
|
|
|
31
31
|
/** Human-readable rationale; surfaces in trace + final-result `winner.verdict`. */
|
|
32
32
|
notes?: string;
|
|
33
33
|
}
|
|
34
|
-
|
|
35
|
-
export type { DefaultVerdict as D };
|
|
34
|
+
//# sourceMappingURL=verdict.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"verdict.d.ts","sourceRoot":"","sources":["../src/verdict.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH;;;;;;;GAOG;AACH,MAAM,WAAW,cAAc;IAC7B,8DAA8D;IAC9D,KAAK,EAAE,OAAO,CAAA;IACd,wEAAwE;IACxE,KAAK,EAAE,MAAM,CAAA;IACb,+EAA+E;IAC/E,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC/B,mFAAmF;IACnF,KAAK,CAAC,EAAE,MAAM,CAAA;CACf"}
|
|
@@ -0,0 +1,32 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Visual diff — pixel-delta scoring for UI / visual outputs.
|
|
3
|
+
*
|
|
4
|
+
* Minimal dependency-free implementation: accepts two PNGs as byte
|
|
5
|
+
* arrays + width/height and returns a Δ ratio + per-channel histogram.
|
|
6
|
+
* Consumers supply the decoded pixel arrays (we don't pull a PNG
|
|
7
|
+
* decoder into the core — use `sharp`, `@napi-rs/canvas`, or Playwright
|
|
8
|
+
* in the driving test and pass the result here).
|
|
9
|
+
*/
|
|
10
|
+
export interface ImageData {
|
|
11
|
+
width: number;
|
|
12
|
+
height: number;
|
|
13
|
+
/** Pixel data in RGBA order, 4 bytes per pixel. */
|
|
14
|
+
data: Uint8Array | Uint8ClampedArray;
|
|
15
|
+
}
|
|
16
|
+
export interface VisualDiffResult {
|
|
17
|
+
/** Ratio of pixels differing beyond `tolerance` (0..1). */
|
|
18
|
+
diffRatio: number;
|
|
19
|
+
differingPixels: number;
|
|
20
|
+
totalPixels: number;
|
|
21
|
+
maxChannelDelta: number;
|
|
22
|
+
/** Status for dashboards: unchanged (< 0.1%), changed, or severely-changed (> 5%). */
|
|
23
|
+
status: 'unchanged' | 'changed' | 'severely-changed';
|
|
24
|
+
}
|
|
25
|
+
export interface VisualDiffOptions {
|
|
26
|
+
/** Pixels whose max-channel delta is ≤ this are considered unchanged. Default 8/255. */
|
|
27
|
+
tolerance?: number;
|
|
28
|
+
}
|
|
29
|
+
export declare function visualDiff(a: ImageData, b: ImageData, options?: VisualDiffOptions): VisualDiffResult;
|
|
30
|
+
/** Convenience: diffs two byte-identical-dim RGBA arrays, returns just the ratio. */
|
|
31
|
+
export declare function pixelDeltaRatio(a: Uint8Array, b: Uint8Array, width: number, height: number, tolerance?: number): number;
|
|
32
|
+
//# sourceMappingURL=visual-diff.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"visual-diff.d.ts","sourceRoot":"","sources":["../src/visual-diff.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAIH,MAAM,WAAW,SAAS;IACxB,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,mDAAmD;IACnD,IAAI,EAAE,UAAU,GAAG,iBAAiB,CAAA;CACrC;AAED,MAAM,WAAW,gBAAgB;IAC/B,2DAA2D;IAC3D,SAAS,EAAE,MAAM,CAAA;IACjB,eAAe,EAAE,MAAM,CAAA;IACvB,WAAW,EAAE,MAAM,CAAA;IACnB,eAAe,EAAE,MAAM,CAAA;IACvB,sFAAsF;IACtF,MAAM,EAAE,WAAW,GAAG,SAAS,GAAG,kBAAkB,CAAA;CACrD;AAED,MAAM,WAAW,iBAAiB;IAChC,wFAAwF;IACxF,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAED,wBAAgB,UAAU,CACxB,CAAC,EAAE,SAAS,EACZ,CAAC,EAAE,SAAS,EACZ,OAAO,GAAE,iBAAsB,GAC9B,gBAAgB,CAyBlB;AAED,qFAAqF;AACrF,wBAAgB,eAAe,CAC7B,CAAC,EAAE,UAAU,EACb,CAAC,EAAE,UAAU,EACb,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,EACd,SAAS,SAAI,GACZ,MAAM,CAER"}
|