@tangle-network/agent-eval 0.95.0 → 0.96.0
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 +17 -0
- package/dist/adapters/http.d.ts +17 -10
- package/dist/adapters/langchain.d.ts +14 -7
- package/dist/adapters/otel.d.ts +25 -13
- package/dist/{rl/adversarial.d.ts → adversarial-DIVcDoI_.d.ts} +7 -6
- package/dist/analyst/index.d.ts +236 -28
- package/dist/analyst/index.js +1 -1
- package/dist/{trace-analyst/analyst.d.ts → analyst-C8HHvfJp.d.ts} +14 -12
- package/dist/{contract/analyze-runs.d.ts → analyze-runs-DtT6F_6T.d.ts} +10 -9
- package/dist/authenticity/index.d.ts +16 -15
- package/dist/{baseline.d.ts → baseline-Bbid3WoO.d.ts} +44 -8
- package/dist/belief-state/index.d.ts +605 -14
- package/dist/benchmarks/index.d.ts +5 -23
- package/dist/builder-eval/index.d.ts +250 -5
- package/dist/calibration-BPmzuVPk.d.ts +101 -0
- package/dist/campaign/index.d.ts +1452 -38
- package/dist/{chunk-AQ5WQAIV.js → chunk-3NHEO6ZC.js} +2 -2
- package/dist/chunk-3NHEO6ZC.js.map +1 -0
- package/dist/cli.d.ts +0 -2
- package/dist/{contract/intake/code-agent-session.d.ts → code-agent-session-CPHRCb4-.d.ts} +17 -15
- package/dist/contract/index.d.ts +824 -107
- package/dist/contract/index.js +145 -1
- package/dist/contract/index.js.map +1 -1
- package/dist/control-Doncu-B_.d.ts +259 -0
- package/dist/{control-runtime.d.ts → control-runtime-Acf9CGhw.d.ts} +26 -23
- package/dist/control.d.ts +10 -11
- package/dist/corpus-D4YW9UoJ.d.ts +560 -0
- package/dist/{cost-ledger.d.ts → cost-ledger-DuSqlw5B.d.ts} +11 -10
- package/dist/{counterfactual.d.ts → counterfactual-DlOz8PBx.d.ts} +13 -12
- package/dist/{dataset.d.ts → dataset-BbGkaN2I.d.ts} +14 -11
- package/dist/{analyst/registry.d.ts → default-registry-GyE8X5SP.d.ts} +37 -8
- package/dist/diagnose.d.ts +252 -1
- package/dist/{trace/emitter.d.ts → emitter-C2rqGH_l.d.ts} +12 -9
- package/dist/{errors.d.ts → errors-CzMUYo7b.d.ts} +11 -10
- package/dist/failure-cluster-DH9Flgcf.d.ts +76 -0
- package/dist/{feedback-trajectory.d.ts → feedback-trajectory-BxY0cKfs.d.ts} +38 -36
- package/dist/fuzz.d.ts +547 -1
- package/dist/gepa-Dprxvz8r.d.ts +414 -0
- package/dist/governance/index.d.ts +135 -5
- package/dist/harness-optimizer-mOl9XX_O.d.ts +106 -0
- package/dist/hosted/index.d.ts +239 -10
- package/dist/index-_Y4oNOOb.d.ts +159 -0
- package/dist/index.d.ts +5660 -277
- package/dist/index.js +1 -1
- package/dist/{contract/insight-report.d.ts → insight-report-BnRjTibG.d.ts} +19 -16
- package/dist/{trace/integrity.d.ts → integrity-D2t12mMw.d.ts} +14 -11
- package/dist/{judge-calibration.d.ts → judge-calibration-0p2QcWNE.d.ts} +17 -16
- package/dist/{analyst/kind-factory.d.ts → kind-factory-X3eDYbKn.d.ts} +61 -10
- package/dist/knowledge/index.d.ts +103 -3
- package/dist/{llm-client.d.ts → llm-client-Bj7g0rqu.d.ts} +23 -21
- package/dist/matrix/index.d.ts +30 -12
- package/dist/meta-eval/index.d.ts +182 -6
- package/dist/{multi-layer-verifier.d.ts → multi-layer-verifier-DUZXrPDA.d.ts} +15 -12
- package/dist/multishot/index.d.ts +290 -7
- package/dist/{rl/off-policy.d.ts → off-policy-DiwuKKg7.d.ts} +9 -8
- package/dist/openapi.json +1 -1
- package/dist/{meta-eval/outcome-store.d.ts → outcome-store-rnXLEqSn.d.ts} +8 -7
- package/dist/{pareto.d.ts → pareto-E-pembql.d.ts} +10 -9
- package/dist/perf/index.d.ts +119 -13
- package/dist/pipelines/index.d.ts +173 -8
- package/dist/pre-registration-nfUdc9EQ.d.ts +483 -0
- package/dist/prm/index.d.ts +104 -5
- package/dist/provenance-CIxfBnkl.d.ts +426 -0
- package/dist/query-B7GGjRox.d.ts +32 -0
- package/dist/{trace/raw-provider-sink.d.ts → raw-provider-sink-C46HDghv.d.ts} +14 -13
- package/dist/{red-team.d.ts → red-team-BWdoyleI.d.ts} +16 -14
- package/dist/{trace/redact.d.ts → redact-B40YG2M_.d.ts} +8 -7
- package/dist/release-report-pidWUMZ2.d.ts +233 -0
- package/dist/reporting.d.ts +16 -15
- package/dist/{eval-campaign.d.ts → researcher-Jr8ME1dZ.d.ts} +156 -34
- package/dist/rl.d.ts +1193 -1
- package/dist/{prm/rubric.d.ts → rubric-Cc6UHvUb.d.ts} +13 -10
- package/dist/{meta-eval/rubric-predictive-validity.d.ts → rubric-predictive-validity-C2hDKM8Z.d.ts} +11 -8
- package/dist/run-critic-CmMf05uV.d.ts +56 -0
- package/dist/{run-record.d.ts → run-record-CP2ObebC.d.ts} +117 -15
- package/dist/runtime-trajectory-BOUUjI0y.d.ts +49 -0
- package/dist/{trace/schema.d.ts → schema-m0gsnbt3.d.ts} +30 -29
- package/dist/semantic-concept-judge-DSBB2Cfp.d.ts +624 -0
- package/dist/{sequential.d.ts → sequential-5iSVfzl2.d.ts} +10 -9
- package/dist/{series-convergence.d.ts → series-convergence-D5OWMBg6.d.ts} +5 -4
- package/dist/sink-fetch-B1Yg4Til.d.ts +101 -0
- package/dist/{statistics.d.ts → statistics-CCJpTGOS.d.ts} +48 -46
- package/dist/{trace/store.d.ts → store-BcFXE6LG.d.ts} +12 -21
- package/dist/{trace-analyst/types.d.ts → store-C1YxJDEK.d.ts} +74 -18
- package/dist/storyboard/index.d.ts +81 -16
- package/dist/{summary-report.d.ts → summary-report-CInXwsza.d.ts} +160 -23
- package/dist/telemetry/{sink-file.d.ts → file.d.ts} +7 -5
- package/dist/telemetry/index.d.ts +35 -17
- package/dist/{sandbox-harness.d.ts → test-graded-scenario-DeODGLra.d.ts} +60 -15
- package/dist/{locked-jsonl-appender.d.ts → testing-C21CHsq2.d.ts} +4 -3
- package/dist/testing.d.ts +1 -5
- package/dist/traces.d.ts +976 -4
- package/dist/{trajectory.d.ts → trajectory-2TkpSEVh.d.ts} +9 -8
- package/dist/{analyst/types.d.ts → types-B5x54y6n.d.ts} +112 -19
- package/dist/{campaign/types.d.ts → types-BMahhhio.d.ts} +47 -44
- package/dist/{matrix/types.d.ts → types-BUxNaJ8c.d.ts} +11 -9
- package/dist/{types.d.ts → types-C7DGg5ex.d.ts} +33 -31
- package/dist/{verdict.d.ts → verdict-C9MlYujm.d.ts} +3 -2
- package/dist/wire/index.d.ts +570 -13
- package/dist/workflow/index.d.ts +496 -22
- package/package.json +2 -2
- package/dist/action-policy.d.ts +0 -24
- package/dist/action-policy.d.ts.map +0 -1
- package/dist/action-policy.test.d.ts +0 -2
- package/dist/action-policy.test.d.ts.map +0 -1
- package/dist/active-learning.d.ts +0 -41
- package/dist/active-learning.d.ts.map +0 -1
- package/dist/adapters/http.d.ts.map +0 -1
- package/dist/adapters/langchain.d.ts.map +0 -1
- package/dist/adapters/otel.d.ts.map +0 -1
- package/dist/agent-profile-cell.d.ts +0 -101
- package/dist/agent-profile-cell.d.ts.map +0 -1
- package/dist/agent-profile.d.ts +0 -27
- package/dist/agent-profile.d.ts.map +0 -1
- package/dist/agent-profile.test.d.ts +0 -2
- package/dist/agent-profile.test.d.ts.map +0 -1
- package/dist/analyst/adapters.d.ts +0 -62
- package/dist/analyst/adapters.d.ts.map +0 -1
- package/dist/analyst/analyst.test.d.ts +0 -2
- package/dist/analyst/analyst.test.d.ts.map +0 -1
- package/dist/analyst/ax-service.d.ts +0 -27
- package/dist/analyst/ax-service.d.ts.map +0 -1
- package/dist/analyst/behavioral-analyst.d.ts +0 -28
- package/dist/analyst/behavioral-analyst.d.ts.map +0 -1
- package/dist/analyst/chat-client.d.ts +0 -91
- package/dist/analyst/chat-client.d.ts.map +0 -1
- package/dist/analyst/default-registry.d.ts +0 -27
- package/dist/analyst/default-registry.d.ts.map +0 -1
- package/dist/analyst/default-registry.test.d.ts +0 -2
- package/dist/analyst/default-registry.test.d.ts.map +0 -1
- package/dist/analyst/finding-signature.d.ts +0 -48
- package/dist/analyst/finding-signature.d.ts.map +0 -1
- package/dist/analyst/finding-subject.d.ts +0 -146
- package/dist/analyst/finding-subject.d.ts.map +0 -1
- package/dist/analyst/finding-subject.test.d.ts +0 -2
- package/dist/analyst/finding-subject.test.d.ts.map +0 -1
- package/dist/analyst/findings-store.d.ts +0 -75
- package/dist/analyst/findings-store.d.ts.map +0 -1
- package/dist/analyst/index.d.ts.map +0 -1
- package/dist/analyst/kind-factory.d.ts.map +0 -1
- package/dist/analyst/kinds/failure-mode.d.ts +0 -19
- package/dist/analyst/kinds/failure-mode.d.ts.map +0 -1
- package/dist/analyst/kinds/improvement.d.ts +0 -23
- package/dist/analyst/kinds/improvement.d.ts.map +0 -1
- package/dist/analyst/kinds/index.d.ts +0 -22
- package/dist/analyst/kinds/index.d.ts.map +0 -1
- package/dist/analyst/kinds/kinds.test.d.ts +0 -2
- package/dist/analyst/kinds/kinds.test.d.ts.map +0 -1
- package/dist/analyst/kinds/knowledge-gap.d.ts +0 -28
- package/dist/analyst/kinds/knowledge-gap.d.ts.map +0 -1
- package/dist/analyst/kinds/knowledge-poisoning.d.ts +0 -22
- package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +0 -1
- package/dist/analyst/kinds/skill-usage.d.ts +0 -84
- package/dist/analyst/kinds/skill-usage.d.ts.map +0 -1
- package/dist/analyst/kinds/skill-usage.test.d.ts +0 -2
- package/dist/analyst/kinds/skill-usage.test.d.ts.map +0 -1
- package/dist/analyst/parse-tolerant.d.ts +0 -26
- package/dist/analyst/parse-tolerant.d.ts.map +0 -1
- package/dist/analyst/parse-tolerant.test.d.ts +0 -2
- package/dist/analyst/parse-tolerant.test.d.ts.map +0 -1
- package/dist/analyst/registry.budget.test.d.ts +0 -2
- package/dist/analyst/registry.budget.test.d.ts.map +0 -1
- package/dist/analyst/registry.d.ts.map +0 -1
- package/dist/analyst/steer-firewall.d.ts +0 -35
- package/dist/analyst/steer-firewall.d.ts.map +0 -1
- package/dist/analyst/steer-firewall.test.d.ts +0 -2
- package/dist/analyst/steer-firewall.test.d.ts.map +0 -1
- package/dist/analyst/structure-findings.d.ts +0 -37
- package/dist/analyst/structure-findings.d.ts.map +0 -1
- package/dist/analyst/structure-findings.test.d.ts +0 -2
- package/dist/analyst/structure-findings.test.d.ts.map +0 -1
- package/dist/analyst/tool-groups.d.ts +0 -34
- package/dist/analyst/tool-groups.d.ts.map +0 -1
- package/dist/analyst/types.d.ts.map +0 -1
- package/dist/anti-slop.d.ts +0 -59
- package/dist/anti-slop.d.ts.map +0 -1
- package/dist/artifact-validator.d.ts +0 -74
- package/dist/artifact-validator.d.ts.map +0 -1
- package/dist/attestation.d.ts +0 -63
- package/dist/attestation.d.ts.map +0 -1
- package/dist/attestation.test.d.ts +0 -2
- package/dist/attestation.test.d.ts.map +0 -1
- package/dist/authenticity/index.d.ts.map +0 -1
- package/dist/authenticity/index.test.d.ts +0 -2
- package/dist/authenticity/index.test.d.ts.map +0 -1
- package/dist/auto-pr.d.ts +0 -120
- package/dist/auto-pr.d.ts.map +0 -1
- package/dist/baseline.d.ts.map +0 -1
- package/dist/behavior-dsl.d.ts +0 -73
- package/dist/behavior-dsl.d.ts.map +0 -1
- package/dist/belief-state/calibration.d.ts +0 -10
- package/dist/belief-state/calibration.d.ts.map +0 -1
- package/dist/belief-state/calibration.test.d.ts +0 -2
- package/dist/belief-state/calibration.test.d.ts.map +0 -1
- package/dist/belief-state/code-agent-corpus.d.ts +0 -66
- package/dist/belief-state/code-agent-corpus.d.ts.map +0 -1
- package/dist/belief-state/code-agent-corpus.test.d.ts +0 -2
- package/dist/belief-state/code-agent-corpus.test.d.ts.map +0 -1
- package/dist/belief-state/code-agent-evidence.d.ts +0 -22
- package/dist/belief-state/code-agent-evidence.d.ts.map +0 -1
- package/dist/belief-state/code-agent-evidence.test.d.ts +0 -2
- package/dist/belief-state/code-agent-evidence.test.d.ts.map +0 -1
- package/dist/belief-state/extract.d.ts +0 -7
- package/dist/belief-state/extract.d.ts.map +0 -1
- package/dist/belief-state/extract.test.d.ts +0 -2
- package/dist/belief-state/extract.test.d.ts.map +0 -1
- package/dist/belief-state/index.d.ts.map +0 -1
- package/dist/belief-state/ope.d.ts +0 -17
- package/dist/belief-state/ope.d.ts.map +0 -1
- package/dist/belief-state/ope.test.d.ts +0 -2
- package/dist/belief-state/ope.test.d.ts.map +0 -1
- package/dist/belief-state/phase0-measurement.d.ts +0 -55
- package/dist/belief-state/phase0-measurement.d.ts.map +0 -1
- package/dist/belief-state/report.d.ts +0 -17
- package/dist/belief-state/report.d.ts.map +0 -1
- package/dist/belief-state/report.test.d.ts +0 -2
- package/dist/belief-state/report.test.d.ts.map +0 -1
- package/dist/belief-state/research-evidence.d.ts +0 -23
- package/dist/belief-state/research-evidence.d.ts.map +0 -1
- package/dist/belief-state/research-evidence.test.d.ts +0 -2
- package/dist/belief-state/research-evidence.test.d.ts.map +0 -1
- package/dist/belief-state/runtime-benchmark-corpus.d.ts +0 -32
- package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +0 -1
- package/dist/belief-state/runtime-hooks.d.ts +0 -87
- package/dist/belief-state/runtime-hooks.d.ts.map +0 -1
- package/dist/belief-state/runtime-hooks.test.d.ts +0 -2
- package/dist/belief-state/runtime-hooks.test.d.ts.map +0 -1
- package/dist/belief-state/selective.d.ts +0 -15
- package/dist/belief-state/selective.d.ts.map +0 -1
- package/dist/belief-state/selective.test.d.ts +0 -2
- package/dist/belief-state/selective.test.d.ts.map +0 -1
- package/dist/belief-state/shadow-probe.d.ts +0 -80
- package/dist/belief-state/shadow-probe.d.ts.map +0 -1
- package/dist/belief-state/shadow-probe.test.d.ts +0 -2
- package/dist/belief-state/shadow-probe.test.d.ts.map +0 -1
- package/dist/belief-state/types.d.ts +0 -195
- package/dist/belief-state/types.d.ts.map +0 -1
- package/dist/belief-state/types.test.d.ts +0 -2
- package/dist/belief-state/types.test.d.ts.map +0 -1
- package/dist/benchmark.d.ts +0 -14
- package/dist/benchmark.d.ts.map +0 -1
- package/dist/benchmarks/index.d.ts.map +0 -1
- package/dist/benchmarks/routing/dataset.d.ts +0 -34
- package/dist/benchmarks/routing/dataset.d.ts.map +0 -1
- package/dist/benchmarks/routing/index.d.ts +0 -34
- package/dist/benchmarks/routing/index.d.ts.map +0 -1
- package/dist/benchmarks/types.d.ts +0 -49
- package/dist/benchmarks/types.d.ts.map +0 -1
- package/dist/bisector.d.ts +0 -81
- package/dist/bisector.d.ts.map +0 -1
- package/dist/budget-guard.d.ts +0 -31
- package/dist/budget-guard.d.ts.map +0 -1
- package/dist/builder-eval/builder-session.d.ts +0 -111
- package/dist/builder-eval/builder-session.d.ts.map +0 -1
- package/dist/builder-eval/correlation.d.ts +0 -32
- package/dist/builder-eval/correlation.d.ts.map +0 -1
- package/dist/builder-eval/index.d.ts.map +0 -1
- package/dist/builder-eval/project-registry.d.ts +0 -51
- package/dist/builder-eval/project-registry.d.ts.map +0 -1
- package/dist/builder-eval/three-layer-eval.d.ts +0 -55
- package/dist/builder-eval/three-layer-eval.d.ts.map +0 -1
- package/dist/campaign/analyst-surface.d.ts +0 -108
- package/dist/campaign/analyst-surface.d.ts.map +0 -1
- package/dist/campaign/analyst-surface.test.d.ts +0 -2
- package/dist/campaign/analyst-surface.test.d.ts.map +0 -1
- package/dist/campaign/auto-pr.d.ts +0 -46
- package/dist/campaign/auto-pr.d.ts.map +0 -1
- package/dist/campaign/distillation/agreement-judge.d.ts +0 -69
- package/dist/campaign/distillation/agreement-judge.d.ts.map +0 -1
- package/dist/campaign/distillation/cli.d.ts +0 -35
- package/dist/campaign/distillation/cli.d.ts.map +0 -1
- package/dist/campaign/distillation/distillation.test.d.ts +0 -2
- package/dist/campaign/distillation/distillation.test.d.ts.map +0 -1
- package/dist/campaign/distillation/gold-scenarios.d.ts +0 -54
- package/dist/campaign/distillation/gold-scenarios.d.ts.map +0 -1
- package/dist/campaign/distillation/run-distillation.d.ts +0 -119
- package/dist/campaign/distillation/run-distillation.d.ts.map +0 -1
- package/dist/campaign/gates/compose.d.ts +0 -12
- package/dist/campaign/gates/compose.d.ts.map +0 -1
- package/dist/campaign/gates/default-production-gate.d.ts +0 -58
- package/dist/campaign/gates/default-production-gate.d.ts.map +0 -1
- package/dist/campaign/gates/heldout-gate.d.ts +0 -12
- package/dist/campaign/gates/heldout-gate.d.ts.map +0 -1
- package/dist/campaign/gates/promotion-policy.d.ts +0 -125
- package/dist/campaign/gates/promotion-policy.d.ts.map +0 -1
- package/dist/campaign/gates/promotion-policy.test.d.ts +0 -2
- package/dist/campaign/gates/promotion-policy.test.d.ts.map +0 -1
- package/dist/campaign/gates/sequential.d.ts +0 -146
- package/dist/campaign/gates/sequential.d.ts.map +0 -1
- package/dist/campaign/gates/sequential.test.d.ts +0 -2
- package/dist/campaign/gates/sequential.test.d.ts.map +0 -1
- package/dist/campaign/gates/statistical-heldout.d.ts +0 -99
- package/dist/campaign/gates/statistical-heldout.d.ts.map +0 -1
- package/dist/campaign/gates/statistical-heldout.test.d.ts +0 -2
- package/dist/campaign/gates/statistical-heldout.test.d.ts.map +0 -1
- package/dist/campaign/index.d.ts.map +0 -1
- package/dist/campaign/labeled-store/fs-adapter.d.ts +0 -59
- package/dist/campaign/labeled-store/fs-adapter.d.ts.map +0 -1
- package/dist/campaign/presets/compare-proposers.d.ts +0 -146
- package/dist/campaign/presets/compare-proposers.d.ts.map +0 -1
- package/dist/campaign/presets/playback.d.ts +0 -120
- package/dist/campaign/presets/playback.d.ts.map +0 -1
- package/dist/campaign/presets/playback.test.d.ts +0 -2
- package/dist/campaign/presets/playback.test.d.ts.map +0 -1
- package/dist/campaign/presets/run-eval.d.ts +0 -14
- package/dist/campaign/presets/run-eval.d.ts.map +0 -1
- package/dist/campaign/presets/run-improvement-loop.d.ts +0 -63
- package/dist/campaign/presets/run-improvement-loop.d.ts.map +0 -1
- package/dist/campaign/presets/run-improvement-loop.test.d.ts +0 -2
- package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +0 -1
- package/dist/campaign/presets/run-optimization.d.ts +0 -92
- package/dist/campaign/presets/run-optimization.d.ts.map +0 -1
- package/dist/campaign/presets/run-profile-matrix.d.ts +0 -151
- package/dist/campaign/presets/run-profile-matrix.d.ts.map +0 -1
- package/dist/campaign/presets/run-skill-opt.d.ts +0 -96
- package/dist/campaign/presets/run-skill-opt.d.ts.map +0 -1
- package/dist/campaign/proposers/_findings-text.d.ts +0 -22
- package/dist/campaign/proposers/_findings-text.d.ts.map +0 -1
- package/dist/campaign/proposers/ace.d.ts +0 -33
- package/dist/campaign/proposers/ace.d.ts.map +0 -1
- package/dist/campaign/proposers/ace.test.d.ts +0 -2
- package/dist/campaign/proposers/ace.test.d.ts.map +0 -1
- package/dist/campaign/proposers/analysis-edit.d.ts +0 -32
- package/dist/campaign/proposers/analysis-edit.d.ts.map +0 -1
- package/dist/campaign/proposers/evolutionary.d.ts +0 -20
- package/dist/campaign/proposers/evolutionary.d.ts.map +0 -1
- package/dist/campaign/proposers/fapo.d.ts +0 -120
- package/dist/campaign/proposers/fapo.d.ts.map +0 -1
- package/dist/campaign/proposers/gepa.d.ts +0 -86
- package/dist/campaign/proposers/gepa.d.ts.map +0 -1
- package/dist/campaign/proposers/halo.d.ts +0 -44
- package/dist/campaign/proposers/halo.d.ts.map +0 -1
- package/dist/campaign/proposers/halo.test.d.ts +0 -2
- package/dist/campaign/proposers/halo.test.d.ts.map +0 -1
- package/dist/campaign/proposers/memory.d.ts +0 -47
- package/dist/campaign/proposers/memory.d.ts.map +0 -1
- package/dist/campaign/proposers/memory.test.d.ts +0 -2
- package/dist/campaign/proposers/memory.test.d.ts.map +0 -1
- package/dist/campaign/proposers/skill-opt.d.ts +0 -88
- package/dist/campaign/proposers/skill-opt.d.ts.map +0 -1
- package/dist/campaign/proposers/trace-analyst.d.ts +0 -48
- package/dist/campaign/proposers/trace-analyst.d.ts.map +0 -1
- package/dist/campaign/proposers/trace-analyst.test.d.ts +0 -2
- package/dist/campaign/proposers/trace-analyst.test.d.ts.map +0 -1
- package/dist/campaign/provenance.d.ts +0 -185
- package/dist/campaign/provenance.d.ts.map +0 -1
- package/dist/campaign/run-campaign.d.ts +0 -90
- package/dist/campaign/run-campaign.d.ts.map +0 -1
- package/dist/campaign/score-utils.d.ts +0 -26
- package/dist/campaign/score-utils.d.ts.map +0 -1
- package/dist/campaign/skill-patch.d.ts +0 -62
- package/dist/campaign/skill-patch.d.ts.map +0 -1
- package/dist/campaign/storage.d.ts +0 -38
- package/dist/campaign/storage.d.ts.map +0 -1
- package/dist/campaign/types.d.ts.map +0 -1
- package/dist/campaign/worktree/index.d.ts +0 -53
- package/dist/campaign/worktree/index.d.ts.map +0 -1
- package/dist/canary.d.ts +0 -101
- package/dist/canary.d.ts.map +0 -1
- package/dist/causal-attribution.d.ts +0 -45
- package/dist/causal-attribution.d.ts.map +0 -1
- package/dist/chunk-AQ5WQAIV.js.map +0 -1
- package/dist/ci-gate.d.ts +0 -44
- package/dist/ci-gate.d.ts.map +0 -1
- package/dist/cli.d.ts.map +0 -1
- package/dist/client.d.ts +0 -77
- package/dist/client.d.ts.map +0 -1
- package/dist/client.test.d.ts +0 -2
- package/dist/client.test.d.ts.map +0 -1
- package/dist/command-runner.d.ts +0 -74
- package/dist/command-runner.d.ts.map +0 -1
- package/dist/command-runner.test.d.ts +0 -2
- package/dist/command-runner.test.d.ts.map +0 -1
- package/dist/completion-verifier.d.ts +0 -147
- package/dist/completion-verifier.d.ts.map +0 -1
- package/dist/completion-verifier.test.d.ts +0 -9
- package/dist/completion-verifier.test.d.ts.map +0 -1
- package/dist/concurrency.d.ts +0 -23
- package/dist/concurrency.d.ts.map +0 -1
- package/dist/contamination-guard.d.ts +0 -81
- package/dist/contamination-guard.d.ts.map +0 -1
- package/dist/contract/analyze-runs.d.ts.map +0 -1
- package/dist/contract/define-agent-eval.d.ts +0 -52
- package/dist/contract/define-agent-eval.d.ts.map +0 -1
- package/dist/contract/diff.d.ts +0 -114
- package/dist/contract/diff.d.ts.map +0 -1
- package/dist/contract/index.d.ts.map +0 -1
- package/dist/contract/insight-report.d.ts.map +0 -1
- package/dist/contract/insight-types-fwd.d.ts +0 -7
- package/dist/contract/insight-types-fwd.d.ts.map +0 -1
- package/dist/contract/intake/agent-trace.d.ts +0 -97
- package/dist/contract/intake/agent-trace.d.ts.map +0 -1
- package/dist/contract/intake/code-agent-session.d.ts.map +0 -1
- package/dist/contract/intake/feedback-table.d.ts +0 -87
- package/dist/contract/intake/feedback-table.d.ts.map +0 -1
- package/dist/contract/intake/index.d.ts +0 -22
- package/dist/contract/intake/index.d.ts.map +0 -1
- package/dist/contract/intake/otel-spans.d.ts +0 -33
- package/dist/contract/intake/otel-spans.d.ts.map +0 -1
- package/dist/contract/self-improve.d.ts +0 -284
- package/dist/contract/self-improve.d.ts.map +0 -1
- package/dist/control-runtime.d.ts.map +0 -1
- package/dist/control-runtime.test.d.ts +0 -2
- package/dist/control-runtime.test.d.ts.map +0 -1
- package/dist/control.d.ts.map +0 -1
- package/dist/convergence.d.ts +0 -29
- package/dist/convergence.d.ts.map +0 -1
- package/dist/cost-ledger.d.ts.map +0 -1
- package/dist/cost-ledger.test.d.ts +0 -2
- package/dist/cost-ledger.test.d.ts.map +0 -1
- package/dist/cost-report.d.ts +0 -43
- package/dist/cost-report.d.ts.map +0 -1
- package/dist/cost-report.test.d.ts +0 -2
- package/dist/cost-report.test.d.ts.map +0 -1
- package/dist/cost-tracker.d.ts +0 -76
- package/dist/cost-tracker.d.ts.map +0 -1
- package/dist/counterfactual.d.ts.map +0 -1
- package/dist/cross-trace-diff.d.ts +0 -56
- package/dist/cross-trace-diff.d.ts.map +0 -1
- package/dist/dataset.d.ts.map +0 -1
- package/dist/deploy-gate-layer.d.ts +0 -125
- package/dist/deploy-gate-layer.d.ts.map +0 -1
- package/dist/deploy-gate-layer.test.d.ts +0 -2
- package/dist/deploy-gate-layer.test.d.ts.map +0 -1
- package/dist/description-length-gate.d.ts +0 -119
- package/dist/description-length-gate.d.ts.map +0 -1
- package/dist/detectors/edge.test.d.ts +0 -2
- package/dist/detectors/edge.test.d.ts.map +0 -1
- package/dist/detectors/index.d.ts +0 -81
- package/dist/detectors/index.d.ts.map +0 -1
- package/dist/detectors/index.test.d.ts +0 -2
- package/dist/detectors/index.test.d.ts.map +0 -1
- package/dist/diagnose/causal-sweep.d.ts +0 -100
- package/dist/diagnose/causal-sweep.d.ts.map +0 -1
- package/dist/diagnose/index.d.ts +0 -36
- package/dist/diagnose/index.d.ts.map +0 -1
- package/dist/diagnose/remediation.d.ts +0 -68
- package/dist/diagnose/remediation.d.ts.map +0 -1
- package/dist/diagnose/repair.d.ts +0 -77
- package/dist/diagnose/repair.d.ts.map +0 -1
- package/dist/discover-personas.d.ts +0 -35
- package/dist/discover-personas.d.ts.map +0 -1
- package/dist/driver.d.ts +0 -95
- package/dist/driver.d.ts.map +0 -1
- package/dist/driver.test.d.ts +0 -8
- package/dist/driver.test.d.ts.map +0 -1
- package/dist/dual-agent-bench.d.ts +0 -81
- package/dist/dual-agent-bench.d.ts.map +0 -1
- package/dist/error-count-extractor.d.ts +0 -47
- package/dist/error-count-extractor.d.ts.map +0 -1
- package/dist/error-count-extractor.test.d.ts +0 -2
- package/dist/error-count-extractor.test.d.ts.map +0 -1
- package/dist/errors.d.ts.map +0 -1
- package/dist/eval-campaign.d.ts.map +0 -1
- package/dist/eval-campaign.test.d.ts +0 -2
- package/dist/eval-campaign.test.d.ts.map +0 -1
- package/dist/eval-tools.d.ts +0 -55
- package/dist/eval-tools.d.ts.map +0 -1
- package/dist/eval-trace-store.d.ts +0 -107
- package/dist/eval-trace-store.d.ts.map +0 -1
- package/dist/eval-trace-store.test.d.ts +0 -2
- package/dist/eval-trace-store.test.d.ts.map +0 -1
- package/dist/executor.d.ts +0 -38
- package/dist/executor.d.ts.map +0 -1
- package/dist/executor.test.d.ts +0 -10
- package/dist/executor.test.d.ts.map +0 -1
- package/dist/experiment-tracker.d.ts +0 -178
- package/dist/experiment-tracker.d.ts.map +0 -1
- package/dist/experiment-tracker.test.d.ts +0 -2
- package/dist/experiment-tracker.test.d.ts.map +0 -1
- package/dist/failure-taxonomy.d.ts +0 -38
- package/dist/failure-taxonomy.d.ts.map +0 -1
- package/dist/feedback-trajectory.d.ts.map +0 -1
- package/dist/feedback-trajectory.test.d.ts +0 -2
- package/dist/feedback-trajectory.test.d.ts.map +0 -1
- package/dist/flow-layer.d.ts +0 -90
- package/dist/flow-layer.d.ts.map +0 -1
- package/dist/flow-layer.test.d.ts +0 -2
- package/dist/flow-layer.test.d.ts.map +0 -1
- package/dist/fuzz/capsule.d.ts +0 -46
- package/dist/fuzz/capsule.d.ts.map +0 -1
- package/dist/fuzz/cube.d.ts +0 -36
- package/dist/fuzz/cube.d.ts.map +0 -1
- package/dist/fuzz/explorer-cost.test.d.ts +0 -2
- package/dist/fuzz/explorer-cost.test.d.ts.map +0 -1
- package/dist/fuzz/explorer.d.ts +0 -64
- package/dist/fuzz/explorer.d.ts.map +0 -1
- package/dist/fuzz/fuzz-agent.d.ts +0 -16
- package/dist/fuzz/fuzz-agent.d.ts.map +0 -1
- package/dist/fuzz/fuzz-agent.test.d.ts +0 -2
- package/dist/fuzz/fuzz-agent.test.d.ts.map +0 -1
- package/dist/fuzz/gates.d.ts +0 -33
- package/dist/fuzz/gates.d.ts.map +0 -1
- package/dist/fuzz/index.d.ts +0 -26
- package/dist/fuzz/index.d.ts.map +0 -1
- package/dist/fuzz/policies.d.ts +0 -28
- package/dist/fuzz/policies.d.ts.map +0 -1
- package/dist/fuzz/tools.d.ts +0 -20
- package/dist/fuzz/tools.d.ts.map +0 -1
- package/dist/fuzz/types.d.ts +0 -307
- package/dist/fuzz/types.d.ts.map +0 -1
- package/dist/golden-matcher.d.ts +0 -71
- package/dist/golden-matcher.d.ts.map +0 -1
- package/dist/governance/eu-ai-act.d.ts +0 -37
- package/dist/governance/eu-ai-act.d.ts.map +0 -1
- package/dist/governance/index.d.ts.map +0 -1
- package/dist/governance/nist-ai-rmf.d.ts +0 -15
- package/dist/governance/nist-ai-rmf.d.ts.map +0 -1
- package/dist/governance/soc2.d.ts +0 -12
- package/dist/governance/soc2.d.ts.map +0 -1
- package/dist/governance/types.d.ts +0 -66
- package/dist/governance/types.d.ts.map +0 -1
- package/dist/harness-optimizer.d.ts +0 -82
- package/dist/harness-optimizer.d.ts.map +0 -1
- package/dist/held-out-gate.d.ts +0 -135
- package/dist/held-out-gate.d.ts.map +0 -1
- package/dist/hosted/client.d.ts +0 -73
- package/dist/hosted/client.d.ts.map +0 -1
- package/dist/hosted/from-env.test.d.ts +0 -8
- package/dist/hosted/from-env.test.d.ts.map +0 -1
- package/dist/hosted/index.d.ts.map +0 -1
- package/dist/hosted/types.d.ts +0 -159
- package/dist/hosted/types.d.ts.map +0 -1
- package/dist/index.d.ts.map +0 -1
- package/dist/integrity/backend-integrity.d.ts +0 -71
- package/dist/integrity/backend-integrity.d.ts.map +0 -1
- package/dist/integrity/preflight.d.ts +0 -72
- package/dist/integrity/preflight.d.ts.map +0 -1
- package/dist/integrity/preflight.test.d.ts +0 -2
- package/dist/integrity/preflight.test.d.ts.map +0 -1
- package/dist/integrity/single-backend.d.ts +0 -67
- package/dist/integrity/single-backend.d.ts.map +0 -1
- package/dist/intent-match-judge.d.ts +0 -69
- package/dist/intent-match-judge.d.ts.map +0 -1
- package/dist/intent-match-judge.test.d.ts +0 -2
- package/dist/intent-match-judge.test.d.ts.map +0 -1
- package/dist/judge-calibration.d.ts.map +0 -1
- package/dist/judge-ensemble.d.ts +0 -66
- package/dist/judge-ensemble.d.ts.map +0 -1
- package/dist/judge-ensemble.test.d.ts +0 -8
- package/dist/judge-ensemble.test.d.ts.map +0 -1
- package/dist/judge-families.d.ts +0 -38
- package/dist/judge-families.d.ts.map +0 -1
- package/dist/judge-panel.d.ts +0 -65
- package/dist/judge-panel.d.ts.map +0 -1
- package/dist/judge-retry.d.ts +0 -70
- package/dist/judge-retry.d.ts.map +0 -1
- package/dist/judge-runner.d.ts +0 -36
- package/dist/judge-runner.d.ts.map +0 -1
- package/dist/judge-runner.test.d.ts +0 -2
- package/dist/judge-runner.test.d.ts.map +0 -1
- package/dist/judges.d.ts +0 -74
- package/dist/judges.d.ts.map +0 -1
- package/dist/keyword-coverage-judge.d.ts +0 -89
- package/dist/keyword-coverage-judge.d.ts.map +0 -1
- package/dist/keyword-coverage-judge.test.d.ts +0 -2
- package/dist/keyword-coverage-judge.test.d.ts.map +0 -1
- package/dist/knowledge/index.d.ts.map +0 -1
- package/dist/knowledge/readiness.d.ts +0 -26
- package/dist/knowledge/readiness.d.ts.map +0 -1
- package/dist/knowledge/types.d.ts +0 -75
- package/dist/knowledge/types.d.ts.map +0 -1
- package/dist/live-proof.d.ts +0 -62
- package/dist/live-proof.d.ts.map +0 -1
- package/dist/llm-client.d.ts.map +0 -1
- package/dist/llm-client.test.d.ts +0 -2
- package/dist/llm-client.test.d.ts.map +0 -1
- package/dist/locked-jsonl-appender.d.ts.map +0 -1
- package/dist/matrix/aggregation.d.ts +0 -16
- package/dist/matrix/aggregation.d.ts.map +0 -1
- package/dist/matrix/index.d.ts.map +0 -1
- package/dist/matrix/runner.d.ts +0 -15
- package/dist/matrix/runner.d.ts.map +0 -1
- package/dist/matrix/types.d.ts.map +0 -1
- package/dist/meta-eval/calibration.d.ts +0 -47
- package/dist/meta-eval/calibration.d.ts.map +0 -1
- package/dist/meta-eval/correlation-study.d.ts +0 -53
- package/dist/meta-eval/correlation-study.d.ts.map +0 -1
- package/dist/meta-eval/index.d.ts.map +0 -1
- package/dist/meta-eval/outcome-store.d.ts.map +0 -1
- package/dist/meta-eval/rubric-predictive-validity.d.ts.map +0 -1
- package/dist/meta-eval/sentinel.d.ts +0 -169
- package/dist/meta-eval/sentinel.d.ts.map +0 -1
- package/dist/metrics.d.ts +0 -63
- package/dist/metrics.d.ts.map +0 -1
- package/dist/model-seats.d.ts +0 -71
- package/dist/model-seats.d.ts.map +0 -1
- package/dist/model-seats.test.d.ts +0 -2
- package/dist/model-seats.test.d.ts.map +0 -1
- package/dist/muffled-gate-scanner.d.ts +0 -102
- package/dist/muffled-gate-scanner.d.ts.map +0 -1
- package/dist/multi-layer-verifier.d.ts.map +0 -1
- package/dist/multi-layer-verifier.test.d.ts +0 -2
- package/dist/multi-layer-verifier.test.d.ts.map +0 -1
- package/dist/multi-toolchain-layer.d.ts +0 -80
- package/dist/multi-toolchain-layer.d.ts.map +0 -1
- package/dist/multi-toolchain-layer.test.d.ts +0 -2
- package/dist/multi-toolchain-layer.test.d.ts.map +0 -1
- package/dist/multishot/default-tools.d.ts +0 -34
- package/dist/multishot/default-tools.d.ts.map +0 -1
- package/dist/multishot/index.d.ts.map +0 -1
- package/dist/multishot/judges.d.ts +0 -32
- package/dist/multishot/judges.d.ts.map +0 -1
- package/dist/multishot/matrix.d.ts +0 -107
- package/dist/multishot/matrix.d.ts.map +0 -1
- package/dist/multishot/multishot.d.ts +0 -23
- package/dist/multishot/multishot.d.ts.map +0 -1
- package/dist/multishot/router.d.ts +0 -37
- package/dist/multishot/router.d.ts.map +0 -1
- package/dist/multishot/types.d.ts +0 -60
- package/dist/multishot/types.d.ts.map +0 -1
- package/dist/observability.d.ts +0 -71
- package/dist/observability.d.ts.map +0 -1
- package/dist/oracle.d.ts +0 -55
- package/dist/oracle.d.ts.map +0 -1
- package/dist/orthogonality.d.ts +0 -35
- package/dist/orthogonality.d.ts.map +0 -1
- package/dist/otel-pipeline.d.ts +0 -31
- package/dist/otel-pipeline.d.ts.map +0 -1
- package/dist/paraphrase.d.ts +0 -107
- package/dist/paraphrase.d.ts.map +0 -1
- package/dist/pareto.d.ts.map +0 -1
- package/dist/partition-held-out.d.ts +0 -70
- package/dist/partition-held-out.d.ts.map +0 -1
- package/dist/partition-held-out.test.d.ts +0 -2
- package/dist/partition-held-out.test.d.ts.map +0 -1
- package/dist/perf/index.d.ts.map +0 -1
- package/dist/perf/integrity.d.ts +0 -30
- package/dist/perf/integrity.d.ts.map +0 -1
- package/dist/perf/journey.d.ts +0 -45
- package/dist/perf/journey.d.ts.map +0 -1
- package/dist/perf/ratchet.d.ts +0 -47
- package/dist/perf/ratchet.d.ts.map +0 -1
- package/dist/pipelines/budget-breach.d.ts +0 -31
- package/dist/pipelines/budget-breach.d.ts.map +0 -1
- package/dist/pipelines/budget-breach.test.d.ts +0 -2
- package/dist/pipelines/budget-breach.test.d.ts.map +0 -1
- package/dist/pipelines/failure-cluster.d.ts +0 -38
- package/dist/pipelines/failure-cluster.d.ts.map +0 -1
- package/dist/pipelines/failure-cluster.test.d.ts +0 -2
- package/dist/pipelines/failure-cluster.test.d.ts.map +0 -1
- package/dist/pipelines/first-divergence.d.ts +0 -26
- package/dist/pipelines/first-divergence.d.ts.map +0 -1
- package/dist/pipelines/first-divergence.test.d.ts +0 -2
- package/dist/pipelines/first-divergence.test.d.ts.map +0 -1
- package/dist/pipelines/index.d.ts.map +0 -1
- package/dist/pipelines/judge-agreement.d.ts +0 -26
- package/dist/pipelines/judge-agreement.d.ts.map +0 -1
- package/dist/pipelines/judge-agreement.test.d.ts +0 -2
- package/dist/pipelines/judge-agreement.test.d.ts.map +0 -1
- package/dist/pipelines/regression.d.ts +0 -23
- package/dist/pipelines/regression.d.ts.map +0 -1
- package/dist/pipelines/regression.test.d.ts +0 -2
- package/dist/pipelines/regression.test.d.ts.map +0 -1
- package/dist/pipelines/stuck-loop.d.ts +0 -32
- package/dist/pipelines/stuck-loop.d.ts.map +0 -1
- package/dist/pipelines/stuck-loop.test.d.ts +0 -2
- package/dist/pipelines/stuck-loop.test.d.ts.map +0 -1
- package/dist/pipelines/tool-waste.d.ts +0 -34
- package/dist/pipelines/tool-waste.d.ts.map +0 -1
- package/dist/pipelines/tool-waste.test.d.ts +0 -2
- package/dist/pipelines/tool-waste.test.d.ts.map +0 -1
- package/dist/playbook.d.ts +0 -16
- package/dist/playbook.d.ts.map +0 -1
- package/dist/pr-review-benchmark.d.ts +0 -88
- package/dist/pr-review-benchmark.d.ts.map +0 -1
- package/dist/pr-review-benchmark.test.d.ts +0 -2
- package/dist/pr-review-benchmark.test.d.ts.map +0 -1
- package/dist/pre-registration.d.ts +0 -125
- package/dist/pre-registration.d.ts.map +0 -1
- package/dist/prm/builtin-rubrics.d.ts +0 -33
- package/dist/prm/builtin-rubrics.d.ts.map +0 -1
- package/dist/prm/index.d.ts.map +0 -1
- package/dist/prm/inference.d.ts +0 -29
- package/dist/prm/inference.d.ts.map +0 -1
- package/dist/prm/inference.test.d.ts +0 -2
- package/dist/prm/inference.test.d.ts.map +0 -1
- package/dist/prm/rubric.d.ts.map +0 -1
- package/dist/prm/training-export.d.ts +0 -38
- package/dist/prm/training-export.d.ts.map +0 -1
- package/dist/produced-state.d.ts +0 -63
- package/dist/produced-state.d.ts.map +0 -1
- package/dist/produced-state.test.d.ts +0 -8
- package/dist/produced-state.test.d.ts.map +0 -1
- package/dist/profile/baselines.d.ts +0 -37
- package/dist/profile/baselines.d.ts.map +0 -1
- package/dist/profile/index.d.ts +0 -105
- package/dist/profile/index.d.ts.map +0 -1
- package/dist/promotion-gate.d.ts +0 -93
- package/dist/promotion-gate.d.ts.map +0 -1
- package/dist/prompt-registry.d.ts +0 -41
- package/dist/prompt-registry.d.ts.map +0 -1
- package/dist/propose-review-control.d.ts +0 -49
- package/dist/propose-review-control.d.ts.map +0 -1
- package/dist/propose-review-control.test.d.ts +0 -2
- package/dist/propose-review-control.test.d.ts.map +0 -1
- package/dist/propose-review.d.ts +0 -155
- package/dist/propose-review.d.ts.map +0 -1
- package/dist/red-team.d.ts.map +0 -1
- package/dist/reference-replay-steering.d.ts +0 -11
- package/dist/reference-replay-steering.d.ts.map +0 -1
- package/dist/reference-replay.d.ts +0 -176
- package/dist/reference-replay.d.ts.map +0 -1
- package/dist/reflective-mutation.d.ts +0 -79
- package/dist/reflective-mutation.d.ts.map +0 -1
- package/dist/registry.d.ts +0 -31
- package/dist/registry.d.ts.map +0 -1
- package/dist/release-confidence.d.ts +0 -128
- package/dist/release-confidence.d.ts.map +0 -1
- package/dist/release-report.d.ts +0 -11
- package/dist/release-report.d.ts.map +0 -1
- package/dist/replay.d.ts +0 -120
- package/dist/replay.d.ts.map +0 -1
- package/dist/reporter.d.ts +0 -14
- package/dist/reporter.d.ts.map +0 -1
- package/dist/reporting.d.ts.map +0 -1
- package/dist/researcher.d.ts +0 -140
- package/dist/researcher.d.ts.map +0 -1
- package/dist/reviewer.d.ts +0 -118
- package/dist/reviewer.d.ts.map +0 -1
- package/dist/reviewer.test.d.ts +0 -2
- package/dist/reviewer.test.d.ts.map +0 -1
- package/dist/reward-model-export.d.ts +0 -60
- package/dist/reward-model-export.d.ts.map +0 -1
- package/dist/rl/active-curriculum.d.ts +0 -110
- package/dist/rl/active-curriculum.d.ts.map +0 -1
- package/dist/rl/adaptation-eval.d.ts +0 -109
- package/dist/rl/adaptation-eval.d.ts.map +0 -1
- package/dist/rl/adversarial.d.ts.map +0 -1
- package/dist/rl/compute-curves.d.ts +0 -127
- package/dist/rl/compute-curves.d.ts.map +0 -1
- package/dist/rl/contamination.d.ts +0 -117
- package/dist/rl/contamination.d.ts.map +0 -1
- package/dist/rl/corpus.d.ts +0 -55
- package/dist/rl/corpus.d.ts.map +0 -1
- package/dist/rl/corpus.test.d.ts +0 -2
- package/dist/rl/corpus.test.d.ts.map +0 -1
- package/dist/rl/dataset.d.ts +0 -102
- package/dist/rl/dataset.d.ts.map +0 -1
- package/dist/rl/dataset.test.d.ts +0 -2
- package/dist/rl/dataset.test.d.ts.map +0 -1
- package/dist/rl/exporters.d.ts +0 -141
- package/dist/rl/exporters.d.ts.map +0 -1
- package/dist/rl/index.d.ts +0 -49
- package/dist/rl/index.d.ts.map +0 -1
- package/dist/rl/off-policy.d.ts.map +0 -1
- package/dist/rl/predictive-validity-researcher.d.ts +0 -69
- package/dist/rl/predictive-validity-researcher.d.ts.map +0 -1
- package/dist/rl/preferences.d.ts +0 -141
- package/dist/rl/preferences.d.ts.map +0 -1
- package/dist/rl/process-reward.d.ts +0 -122
- package/dist/rl/process-reward.d.ts.map +0 -1
- package/dist/rl/reward-hacking.d.ts +0 -104
- package/dist/rl/reward-hacking.d.ts.map +0 -1
- package/dist/rl/rl-campaign.d.ts +0 -85
- package/dist/rl/rl-campaign.d.ts.map +0 -1
- package/dist/rl/run-record-adapters.d.ts +0 -56
- package/dist/rl/run-record-adapters.d.ts.map +0 -1
- package/dist/rl/sim-fidelity.d.ts +0 -166
- package/dist/rl/sim-fidelity.d.ts.map +0 -1
- package/dist/rl/sim-fidelity.test.d.ts +0 -2
- package/dist/rl/sim-fidelity.test.d.ts.map +0 -1
- package/dist/rl/tournament.d.ts +0 -115
- package/dist/rl/tournament.d.ts.map +0 -1
- package/dist/rl/verifiable-reward.d.ts +0 -124
- package/dist/rl/verifiable-reward.d.ts.map +0 -1
- package/dist/run-critic.d.ts +0 -23
- package/dist/run-critic.d.ts.map +0 -1
- package/dist/run-evidence.d.ts +0 -32
- package/dist/run-evidence.d.ts.map +0 -1
- package/dist/run-record.d.ts.map +0 -1
- package/dist/run-record.test.d.ts +0 -2
- package/dist/run-record.test.d.ts.map +0 -1
- package/dist/run-score.d.ts +0 -31
- package/dist/run-score.d.ts.map +0 -1
- package/dist/runtime-trajectory.d.ts +0 -47
- package/dist/runtime-trajectory.d.ts.map +0 -1
- package/dist/sandbox-harness.d.ts.map +0 -1
- package/dist/sandbox-harness.test.d.ts +0 -2
- package/dist/sandbox-harness.test.d.ts.map +0 -1
- package/dist/sandbox-pool.d.ts +0 -74
- package/dist/sandbox-pool.d.ts.map +0 -1
- package/dist/sandbox-pool.test.d.ts +0 -2
- package/dist/sandbox-pool.test.d.ts.map +0 -1
- package/dist/scorecard.d.ts +0 -133
- package/dist/scorecard.d.ts.map +0 -1
- package/dist/scorecard.test.d.ts +0 -2
- package/dist/scorecard.test.d.ts.map +0 -1
- package/dist/self-play.d.ts +0 -69
- package/dist/self-play.d.ts.map +0 -1
- package/dist/semantic-concept-judge.d.ts +0 -135
- package/dist/semantic-concept-judge.d.ts.map +0 -1
- package/dist/semantic-concept-judge.test.d.ts +0 -2
- package/dist/semantic-concept-judge.test.d.ts.map +0 -1
- package/dist/sequential.d.ts.map +0 -1
- package/dist/series-convergence.d.ts.map +0 -1
- package/dist/slo.d.ts +0 -48
- package/dist/slo.d.ts.map +0 -1
- package/dist/state-continuity.d.ts +0 -47
- package/dist/state-continuity.d.ts.map +0 -1
- package/dist/statistics.d.ts.map +0 -1
- package/dist/statistics.test.d.ts +0 -2
- package/dist/statistics.test.d.ts.map +0 -1
- package/dist/steering-optimizer.d.ts +0 -58
- package/dist/steering-optimizer.d.ts.map +0 -1
- package/dist/steering.d.ts +0 -24
- package/dist/steering.d.ts.map +0 -1
- package/dist/storyboard/code-edit.d.ts +0 -64
- package/dist/storyboard/code-edit.d.ts.map +0 -1
- package/dist/storyboard/code-edit.test.d.ts +0 -2
- package/dist/storyboard/code-edit.test.d.ts.map +0 -1
- package/dist/storyboard/index.d.ts.map +0 -1
- package/dist/storyboard/index.test.d.ts +0 -2
- package/dist/storyboard/index.test.d.ts.map +0 -1
- package/dist/summary-report.d.ts.map +0 -1
- package/dist/telemetry/client.d.ts +0 -35
- package/dist/telemetry/client.d.ts.map +0 -1
- package/dist/telemetry/index.d.ts.map +0 -1
- package/dist/telemetry/schema.d.ts +0 -61
- package/dist/telemetry/schema.d.ts.map +0 -1
- package/dist/telemetry/sink-fetch.d.ts +0 -39
- package/dist/telemetry/sink-fetch.d.ts.map +0 -1
- package/dist/telemetry/sink-file.d.ts.map +0 -1
- package/dist/test-graded-scenario.d.ts +0 -42
- package/dist/test-graded-scenario.d.ts.map +0 -1
- package/dist/testing.d.ts.map +0 -1
- package/dist/tool-use-metrics.d.ts +0 -35
- package/dist/tool-use-metrics.d.ts.map +0 -1
- package/dist/trace/capture-fetch.d.ts +0 -48
- package/dist/trace/capture-fetch.d.ts.map +0 -1
- package/dist/trace/capture-fetch.test.d.ts +0 -2
- package/dist/trace/capture-fetch.test.d.ts.map +0 -1
- package/dist/trace/emitter.d.ts.map +0 -1
- package/dist/trace/extract-usage.d.ts +0 -46
- package/dist/trace/extract-usage.d.ts.map +0 -1
- package/dist/trace/extract-usage.test.d.ts +0 -2
- package/dist/trace/extract-usage.test.d.ts.map +0 -1
- package/dist/trace/index.d.ts +0 -15
- package/dist/trace/index.d.ts.map +0 -1
- package/dist/trace/integrity.d.ts.map +0 -1
- package/dist/trace/otel-bridge.d.ts +0 -29
- package/dist/trace/otel-bridge.d.ts.map +0 -1
- package/dist/trace/otel-export.d.ts +0 -52
- package/dist/trace/otel-export.d.ts.map +0 -1
- package/dist/trace/otel.d.ts +0 -57
- package/dist/trace/otel.d.ts.map +0 -1
- package/dist/trace/otlp-attributes.d.ts +0 -17
- package/dist/trace/otlp-attributes.d.ts.map +0 -1
- package/dist/trace/query.d.ts +0 -29
- package/dist/trace/query.d.ts.map +0 -1
- package/dist/trace/query.test.d.ts +0 -2
- package/dist/trace/query.test.d.ts.map +0 -1
- package/dist/trace/raw-provider-sink.d.ts.map +0 -1
- package/dist/trace/redact.d.ts.map +0 -1
- package/dist/trace/schema.d.ts.map +0 -1
- package/dist/trace/store-to-otlp.d.ts +0 -72
- package/dist/trace/store-to-otlp.d.ts.map +0 -1
- package/dist/trace/store-to-otlp.test.d.ts +0 -2
- package/dist/trace/store-to-otlp.test.d.ts.map +0 -1
- package/dist/trace/store.d.ts.map +0 -1
- package/dist/trace/store.test.d.ts +0 -2
- package/dist/trace/store.test.d.ts.map +0 -1
- package/dist/trace-analyst/analyst.d.ts.map +0 -1
- package/dist/trace-analyst/analyst.test.d.ts +0 -2
- package/dist/trace-analyst/analyst.test.d.ts.map +0 -1
- package/dist/trace-analyst/behavioral-metrics.d.ts +0 -40
- package/dist/trace-analyst/behavioral-metrics.d.ts.map +0 -1
- package/dist/trace-analyst/behavioral-metrics.test.d.ts +0 -2
- package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +0 -1
- package/dist/trace-analyst/hook.d.ts +0 -55
- package/dist/trace-analyst/hook.d.ts.map +0 -1
- package/dist/trace-analyst/index.d.ts +0 -18
- package/dist/trace-analyst/index.d.ts.map +0 -1
- package/dist/trace-analyst/insights.d.ts +0 -71
- package/dist/trace-analyst/insights.d.ts.map +0 -1
- package/dist/trace-analyst/insights.test.d.ts +0 -2
- package/dist/trace-analyst/insights.test.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-flatten.d.ts +0 -42
- package/dist/trace-analyst/otlp-flatten.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-span.d.ts +0 -85
- package/dist/trace-analyst/otlp-span.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-span.test.d.ts +0 -2
- package/dist/trace-analyst/otlp-span.test.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-to-run-records.d.ts +0 -115
- package/dist/trace-analyst/otlp-to-run-records.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts +0 -2
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +0 -2
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +0 -1
- package/dist/trace-analyst/prompts.d.ts +0 -6
- package/dist/trace-analyst/prompts.d.ts.map +0 -1
- package/dist/trace-analyst/store-otlp.d.ts +0 -126
- package/dist/trace-analyst/store-otlp.d.ts.map +0 -1
- package/dist/trace-analyst/store-otlp.test.d.ts +0 -8
- package/dist/trace-analyst/store-otlp.test.d.ts.map +0 -1
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +0 -2
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +0 -1
- package/dist/trace-analyst/store.d.ts +0 -63
- package/dist/trace-analyst/store.d.ts.map +0 -1
- package/dist/trace-analyst/tools.d.ts +0 -44
- package/dist/trace-analyst/tools.d.ts.map +0 -1
- package/dist/trace-analyst/tools.test.d.ts +0 -10
- package/dist/trace-analyst/tools.test.d.ts.map +0 -1
- package/dist/trace-analyst/types.d.ts.map +0 -1
- package/dist/trace-contracts.d.ts +0 -180
- package/dist/trace-contracts.d.ts.map +0 -1
- package/dist/traced-analyst.d.ts +0 -26
- package/dist/traced-analyst.d.ts.map +0 -1
- package/dist/traced-judges.d.ts +0 -27
- package/dist/traced-judges.d.ts.map +0 -1
- package/dist/traces.d.ts.map +0 -1
- package/dist/trajectory.d.ts.map +0 -1
- package/dist/types.d.ts.map +0 -1
- package/dist/ui-finding.d.ts +0 -104
- package/dist/ui-finding.d.ts.map +0 -1
- package/dist/verdict-cache.d.ts +0 -78
- package/dist/verdict-cache.d.ts.map +0 -1
- package/dist/verdict-cache.test.d.ts +0 -2
- package/dist/verdict-cache.test.d.ts.map +0 -1
- package/dist/verdict.d.ts.map +0 -1
- package/dist/visual-diff.d.ts +0 -32
- package/dist/visual-diff.d.ts.map +0 -1
- package/dist/wire/handlers.d.ts +0 -54
- package/dist/wire/handlers.d.ts.map +0 -1
- package/dist/wire/index.d.ts.map +0 -1
- package/dist/wire/openapi.d.ts +0 -3
- package/dist/wire/openapi.d.ts.map +0 -1
- package/dist/wire/rpc.d.ts +0 -21
- package/dist/wire/rpc.d.ts.map +0 -1
- package/dist/wire/rubrics.d.ts +0 -34
- package/dist/wire/rubrics.d.ts.map +0 -1
- package/dist/wire/schemas.d.ts +0 -410
- package/dist/wire/schemas.d.ts.map +0 -1
- package/dist/wire/server.d.ts +0 -60
- package/dist/wire/server.d.ts.map +0 -1
- package/dist/workflow/event-schema.d.ts +0 -5
- package/dist/workflow/event-schema.d.ts.map +0 -1
- package/dist/workflow/feedback-pack.d.ts +0 -99
- package/dist/workflow/feedback-pack.d.ts.map +0 -1
- package/dist/workflow/index.d.ts.map +0 -1
- package/dist/workflow/intelligence-export.d.ts +0 -62
- package/dist/workflow/intelligence-export.d.ts.map +0 -1
- package/dist/workflow/partner-report.d.ts +0 -49
- package/dist/workflow/partner-report.d.ts.map +0 -1
- package/dist/workflow/phase-graph.d.ts +0 -43
- package/dist/workflow/phase-graph.d.ts.map +0 -1
- package/dist/workflow/promotion-gate.d.ts +0 -61
- package/dist/workflow/promotion-gate.d.ts.map +0 -1
- package/dist/workflow/run-record.d.ts +0 -12
- package/dist/workflow/run-record.d.ts.map +0 -1
- package/dist/workflow/runtime-adapter.d.ts +0 -20
- package/dist/workflow/runtime-adapter.d.ts.map +0 -1
- package/dist/workflow/sanitize.d.ts +0 -21
- package/dist/workflow/sanitize.d.ts.map +0 -1
- package/dist/workflow/schema.d.ts +0 -5
- package/dist/workflow/schema.d.ts.map +0 -1
- package/dist/workflow/summary.d.ts +0 -43
- package/dist/workflow/summary.d.ts.map +0 -1
- package/dist/workflow/trace-event-fields.d.ts +0 -6
- package/dist/workflow/trace-event-fields.d.ts.map +0 -1
- package/dist/workflow/trajectory.d.ts +0 -15
- package/dist/workflow/trajectory.d.ts.map +0 -1
- package/dist/workflow/types.d.ts +0 -68
- package/dist/workflow/types.d.ts.map +0 -1
- package/dist/workspace-inspector.d.ts +0 -67
- package/dist/workspace-inspector.d.ts.map +0 -1
- package/dist/wrangler-deploy-runner.test.d.ts +0 -2
- package/dist/wrangler-deploy-runner.test.d.ts.map +0 -1
|
@@ -1,75 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* FindingsStore — durable persistence for AnalystFinding rows + a diff
|
|
3
|
-
* helper so we can answer "what changed since the last run?" without
|
|
4
|
-
* recomputing analysts.
|
|
5
|
-
*
|
|
6
|
-
* On-disk shape is JSONL: one finding per line, append-only, locked via
|
|
7
|
-
* LockedJsonlAppender. Operators get crash-safety (no partial JSON),
|
|
8
|
-
* cheap reads (sequential parse), and trivial backup (rsync the file).
|
|
9
|
-
*
|
|
10
|
-
* Reads are non-locking: a reader sees a consistent snapshot of all
|
|
11
|
-
* fully-written lines and skips an incomplete trailing line if the
|
|
12
|
-
* writer is mid-append. Cross-process locking is intentionally out of
|
|
13
|
-
* scope (see locked-jsonl-appender.ts).
|
|
14
|
-
*
|
|
15
|
-
* The store is run-scoped: callers pass `runId` on append and on load,
|
|
16
|
-
* which keeps multi-run files cleanly partitioned. The `diffFindings`
|
|
17
|
-
* helper compares two run-id sets using stable `finding_id` semantics —
|
|
18
|
-
* the diff is the cross-run signal the regression dashboard renders.
|
|
19
|
-
*/
|
|
20
|
-
import type { AnalystFinding } from './types';
|
|
21
|
-
/**
|
|
22
|
-
* One persisted row. We attach `run_id` on disk so a single file can
|
|
23
|
-
* hold multiple runs and the diff helper can query without re-walking
|
|
24
|
-
* separate files.
|
|
25
|
-
*/
|
|
26
|
-
export interface PersistedFinding extends AnalystFinding {
|
|
27
|
-
run_id: string;
|
|
28
|
-
}
|
|
29
|
-
export declare class FindingsStore {
|
|
30
|
-
readonly path: string;
|
|
31
|
-
private readonly appender;
|
|
32
|
-
constructor(path: string);
|
|
33
|
-
append(runId: string, findings: AnalystFinding[]): Promise<void>;
|
|
34
|
-
/** Load every persisted finding. Discards malformed trailing lines silently. */
|
|
35
|
-
loadAll(): PersistedFinding[];
|
|
36
|
-
/** Filter to a single run. */
|
|
37
|
-
loadRun(runId: string): PersistedFinding[];
|
|
38
|
-
}
|
|
39
|
-
export interface FindingsDiff {
|
|
40
|
-
/** New finding ids in `current` that weren't in `previous`. */
|
|
41
|
-
appeared: PersistedFinding[];
|
|
42
|
-
/** Finding ids in `previous` that aren't in `current`. */
|
|
43
|
-
disappeared: PersistedFinding[];
|
|
44
|
-
/** Same finding id present in both runs and unchanged per the materiality test. */
|
|
45
|
-
persisted: PersistedFinding[];
|
|
46
|
-
/**
|
|
47
|
-
* Same finding id in both runs but at least one non-identity field
|
|
48
|
-
* shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].
|
|
49
|
-
*/
|
|
50
|
-
changed: Array<{
|
|
51
|
-
previous: PersistedFinding;
|
|
52
|
-
current: PersistedFinding;
|
|
53
|
-
}>;
|
|
54
|
-
}
|
|
55
|
-
export interface DiffPolicy {
|
|
56
|
-
/**
|
|
57
|
-
* Predicate that decides whether two findings (same finding_id) count
|
|
58
|
-
* as a material change. Defaults to {@link defaultIsMaterial}: severity
|
|
59
|
-
* shift, confidence Δ > 0.05, or evidence count change. Compliance /
|
|
60
|
-
* perf consumers MAY supply a stricter predicate (e.g. rationale text
|
|
61
|
-
* diff, metric Δ thresholds).
|
|
62
|
-
*/
|
|
63
|
-
isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean;
|
|
64
|
-
}
|
|
65
|
-
/**
|
|
66
|
-
* Default materiality test. Deliberately narrow so LLM-reword churn
|
|
67
|
-
* doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.
|
|
68
|
-
*/
|
|
69
|
-
export declare function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean;
|
|
70
|
-
/**
|
|
71
|
-
* Diff two findings sets by stable finding_id. Callers typically load
|
|
72
|
-
* the two run-id slices from the same store and pass them in.
|
|
73
|
-
*/
|
|
74
|
-
export declare function diffFindings(previous: PersistedFinding[], current: PersistedFinding[], policy?: DiffPolicy): FindingsDiff;
|
|
75
|
-
//# sourceMappingURL=findings-store.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"findings-store.d.ts","sourceRoot":"","sources":["../../src/analyst/findings-store.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAKH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,SAAS,CAAA;AAE7C;;;;GAIG;AACH,MAAM,WAAW,gBAAiB,SAAQ,cAAc;IACtD,MAAM,EAAE,MAAM,CAAA;CACf;AAED,qBAAa,aAAa;aAGI,IAAI,EAAE,MAAM;IAFxC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAqB;gBAElB,IAAI,EAAE,MAAM;IAIlC,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,cAAc,EAAE,GAAG,OAAO,CAAC,IAAI,CAAC;IAOtE,gFAAgF;IAChF,OAAO,IAAI,gBAAgB,EAAE;IAiB7B,8BAA8B;IAC9B,OAAO,CAAC,KAAK,EAAE,MAAM,GAAG,gBAAgB,EAAE;CAG3C;AAID,MAAM,WAAW,YAAY;IAC3B,+DAA+D;IAC/D,QAAQ,EAAE,gBAAgB,EAAE,CAAA;IAC5B,0DAA0D;IAC1D,WAAW,EAAE,gBAAgB,EAAE,CAAA;IAC/B,mFAAmF;IACnF,SAAS,EAAE,gBAAgB,EAAE,CAAA;IAC7B;;;OAGG;IACH,OAAO,EAAE,KAAK,CAAC;QAAE,QAAQ,EAAE,gBAAgB,CAAC;QAAC,OAAO,EAAE,gBAAgB,CAAA;KAAE,CAAC,CAAA;CAC1E;AAED,MAAM,WAAW,UAAU;IACzB;;;;;;OAMG;IACH,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,cAAc,EAAE,OAAO,EAAE,cAAc,KAAK,OAAO,CAAA;CAC5E;AAED;;;GAGG;AACH,wBAAgB,iBAAiB,CAAC,CAAC,EAAE,cAAc,EAAE,CAAC,EAAE,cAAc,GAAG,OAAO,CAK/E;AAED;;;GAGG;AACH,wBAAgB,YAAY,CAC1B,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,OAAO,EAAE,gBAAgB,EAAE,EAC3B,MAAM,GAAE,UAAe,GACtB,YAAY,CA0Bd"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/analyst/index.ts"],"names":[],"mappings":"AAQA,YAAY,EACV,gBAAgB,EAChB,oBAAoB,EACpB,+BAA+B,EAC/B,mBAAmB,GACpB,MAAM,YAAY,CAAA;AACnB,OAAO,EACL,kBAAkB,EAClB,sBAAsB,EACtB,iCAAiC,EACjC,qBAAqB,EACrB,YAAY,GACb,MAAM,YAAY,CAAA;AACnB,OAAO,EAAE,KAAK,qBAAqB,EAAE,eAAe,EAAE,MAAM,cAAc,CAAA;AAC1E,OAAO,EAAE,iBAAiB,EAAE,wBAAwB,EAAE,MAAM,sBAAsB,CAAA;AAClF,YAAY,EACV,YAAY,EACZ,UAAU,EACV,WAAW,EACX,YAAY,EACZ,aAAa,EACb,sBAAsB,EACtB,oBAAoB,EACpB,2BAA2B,EAC3B,iBAAiB,EACjB,mBAAmB,EACnB,uBAAuB,GACxB,MAAM,eAAe,CAAA;AACtB,OAAO,EAAE,gBAAgB,EAAE,MAAM,eAAe,CAAA;AAChD,OAAO,EACL,2BAA2B,EAC3B,KAAK,6BAA6B,GACnC,MAAM,oBAAoB,CAAA;AAC3B,YAAY,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAA;AAC5D,OAAO,EACL,kBAAkB,EAClB,eAAe,EACf,yBAAyB,EACzB,uBAAuB,GACxB,MAAM,qBAAqB,CAAA;AAC5B,YAAY,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAA;AAC3E,OAAO,EACL,8BAA8B,EAC9B,qBAAqB,EACrB,0BAA0B,EAC1B,sBAAsB,EACtB,mBAAmB,EACnB,oBAAoB,GACrB,MAAM,mBAAmB,CAAA;AAC1B,YAAY,EAAE,UAAU,EAAE,YAAY,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AAClF,OAAO,EAAE,iBAAiB,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAA;AACjF,YAAY,EACV,0BAA0B,EAC1B,kBAAkB,EAClB,oBAAoB,GACrB,MAAM,gBAAgB,CAAA;AACvB,OAAO,EAAE,sBAAsB,EAAE,mBAAmB,EAAE,MAAM,gBAAgB,CAAA;AAC5E,OAAO,EACL,2BAA2B,EAC3B,sBAAsB,EACtB,qBAAqB,EACrB,uBAAuB,EACvB,6BAA6B,GAC9B,MAAM,SAAS,CAAA;AAChB,YAAY,EACV,gBAAgB,EAChB,gBAAgB,EAChB,oBAAoB,GACrB,MAAM,qBAAqB,CAAA;AAC5B,OAAO,EACL,qBAAqB,EACrB,sBAAsB,EACtB,mBAAmB,EACnB,iBAAiB,GAClB,MAAM,qBAAqB,CAAA;AAC5B,OAAO,EAAE,UAAU,EAAE,mBAAmB,EAAE,eAAe,EAAE,MAAM,kBAAkB,CAAA;AACnF,YAAY,EACV,YAAY,EACZ,sBAAsB,EACtB,YAAY,EACZ,eAAe,GAChB,MAAM,YAAY,CAAA;AACnB,OAAO,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAC5C,OAAO,EAAE,oBAAoB,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAA;AAC1F,OAAO,EACL,KAAK,wBAAwB,EAC7B,KAAK,uBAAuB,EAC5B,iBAAiB,GAClB,MAAM,sBAAsB,CAAA;AAC7B,YAAY,EAAE,kBAAkB,EAAE,MAAM,eAAe,CAAA;AACvD,OAAO,EAAE,uBAAuB,EAAE,MAAM,eAAe,CAAA;AACvD,YAAY,EACV,OAAO,EACP,cAAc,EACd,WAAW,EACX,cAAc,EACd,gBAAgB,EAChB,mBAAmB,EACnB,eAAe,EACf,gBAAgB,EAChB,gBAAgB,EAChB,iBAAiB,EACjB,eAAe,EACf,WAAW,GACZ,MAAM,SAAS,CAAA;AAChB,OAAO,EAAE,gBAAgB,EAAE,WAAW,EAAE,MAAM,SAAS,CAAA"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"kind-factory.d.ts","sourceRoot":"","sources":["../../src/analyst/kind-factory.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,UAAU,EAAE,MAAM,YAAY,CAAA;AAEzD,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,wBAAwB,CAAA;AAEhE,OAAO,EAGL,KAAK,iBAAiB,EACvB,MAAM,qBAAqB,CAAA;AAG5B,OAAO,KAAK,EAAE,OAAO,EAAE,cAAc,EAAE,WAAW,EAAkB,MAAM,SAAS,CAAA;AAGnF;;;GAGG;AACH,MAAM,WAAW,oBAAoB;IACnC,4EAA4E;IAC5E,EAAE,EAAE,MAAM,CAAA;IACV,2DAA2D;IAC3D,WAAW,EAAE,MAAM,CAAA;IACnB,qGAAqG;IACrG,IAAI,EAAE,MAAM,CAAA;IACZ,wEAAwE;IACxE,OAAO,EAAE,MAAM,CAAA;IACf,oFAAoF;IACpF,gBAAgB,EAAE,MAAM,CAAA;IACxB,0FAA0F;IAC1F,oBAAoB,CAAC,EAAE,MAAM,CAAA;IAC7B,wFAAwF;IACxF,UAAU,EAAE,CAAC,KAAK,EAAE,kBAAkB,KAAK,UAAU,EAAE,CAAA;IACvD,0DAA0D;IAC1D,SAAS,CAAC,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,oBAAoB,CAAC,EAAE,MAAM,CAAA;KAAE,CAAA;IAC/D,kCAAkC;IAClC,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,sCAAsC;IACtC,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,gFAAgF;IAChF,IAAI,EAAE,WAAW,CAAA;IACjB,wEAAwE;IACxE,WAAW,CAAC,EAAE,CAAC,GAAG,EAAE,iBAAiB,EAAE,GAAG,EAAE,cAAc,KAAK,iBAAiB,GAAG,IAAI,CAAA;IACvF,wGAAwG;IACxG,OAAO,CAAC,EAAE,kBAAkB,EAAE,CAAA;CAC/B;AAED;;;;;GAKG;AACH,MAAM,WAAW,kBAAkB;IACjC,QAAQ,EAAE,MAAM,CAAA;IAChB,QAAQ,EAAE,aAAa,CAAC,IAAI,CAAC,iBAAiB,EAAE,YAAY,CAAC,CAAC,CAAA;CAC/D;AAED,MAAM,WAAW,0BAA0B;IACzC,8CAA8C;IAC9C,EAAE,EAAE,WAAW,CAAA;IACf,uEAAuE;IACvE,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,sFAAsF;IACtF,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;;;;;OAMG;IACH,QAAQ,CAAC,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,OAAO,KAAK,CAAA;KAAE,CAAA;CAC1F;AAED;;;;;;;GAOG;AACH,wBAAgB,sBAAsB,CACpC,IAAI,EAAE,oBAAoB,EAC1B,IAAI,EAAE,0BAA0B,GAC/B,OAAO,CAAC,kBAAkB,CAAC,CAoK7B;AA0CD;;;;;;;;;;;;;GAaG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,cAAc,CAAC,eAAe,CAAC,GAAG,MAAM,CAsBlF"}
|
|
@@ -1,19 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Failure-mode analyst — classifies what went wrong and why.
|
|
3
|
-
*
|
|
4
|
-
* Brief: read the trace dataset, identify the top failure modes across
|
|
5
|
-
* runs, classify each with severity + evidence, and surface them as
|
|
6
|
-
* findings. The actor's job is *taxonomy + evidence*, not fix-design —
|
|
7
|
-
* that's the improvement-analyst's job.
|
|
8
|
-
*
|
|
9
|
-
* Recursion is deep (`maxDepth: 3`) because real failure-mode
|
|
10
|
-
* discovery is genuinely tree-shaped: the actor splits the dataset
|
|
11
|
-
* into candidate clusters, each cluster spawns a focused investigator
|
|
12
|
-
* that drills into representative traces, and a deeply-recursed
|
|
13
|
-
* investigator may itself split a confounded mode into two sub-modes.
|
|
14
|
-
* Each level fans out 4-way, so the analyst can investigate up to
|
|
15
|
-
* ~16 leaf clusters before hitting the depth ceiling.
|
|
16
|
-
*/
|
|
17
|
-
import type { TraceAnalystKindSpec } from '../kind-factory';
|
|
18
|
-
export declare const FAILURE_MODE_KIND_SPEC: TraceAnalystKindSpec;
|
|
19
|
-
//# sourceMappingURL=failure-mode.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"failure-mode.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/failure-mode.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAA;AAkC3D,eAAO,MAAM,sBAAsB,EAAE,oBAWpC,CAAA"}
|
|
@@ -1,23 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Improvement analyst — actionable, recursive self-improvement findings.
|
|
3
|
-
*
|
|
4
|
-
* Brief: read findings from upstream analysts (failure-mode,
|
|
5
|
-
* knowledge-gap, knowledge-poisoning) AND the trace dataset itself,
|
|
6
|
-
* then propose **concrete edits** to the agent's runtime: prompt
|
|
7
|
-
* additions, RAG documents to ingest, tool descriptions to rewrite,
|
|
8
|
-
* scaffolding changes to make, memory entries to invalidate. Each
|
|
9
|
-
* finding is one proposed edit with the locus, the diff, and the
|
|
10
|
-
* expected effect.
|
|
11
|
-
*
|
|
12
|
-
* This is the recursive-self-improvement loop's last mile: the prior
|
|
13
|
-
* kinds describe *what's wrong*; this kind describes *what to change*.
|
|
14
|
-
*
|
|
15
|
-
* Recursion is deep (`maxDepth: 3`) because real improvement proposals
|
|
16
|
-
* are competitive: for each failure-mode there are usually 2-3 viable
|
|
17
|
-
* fix directions (tighten prompt vs add tool vs adjust scaffolding),
|
|
18
|
-
* and the actor should explore each with a focused subagent before
|
|
19
|
-
* picking the highest-leverage one to recommend.
|
|
20
|
-
*/
|
|
21
|
-
import type { TraceAnalystKindSpec } from '../kind-factory';
|
|
22
|
-
export declare const IMPROVEMENT_KIND_SPEC: TraceAnalystKindSpec;
|
|
23
|
-
//# sourceMappingURL=improvement.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"improvement.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/improvement.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAA;AAyC3D,eAAO,MAAM,qBAAqB,EAAE,oBAYnC,CAAA"}
|
|
@@ -1,22 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Default analyst kinds focused on agent failure + recursive
|
|
3
|
-
* self-improvement.
|
|
4
|
-
*
|
|
5
|
-
* The four kinds chain: failure-mode classifies; knowledge-gap and
|
|
6
|
-
* knowledge-poisoning explain *why* in two orthogonal ways; improvement
|
|
7
|
-
* proposes concrete edits. Register all four against the same trace
|
|
8
|
-
* store and the registry runs them in dependency order if the operator
|
|
9
|
-
* pipes findings between them.
|
|
10
|
-
*/
|
|
11
|
-
export { FAILURE_MODE_KIND_SPEC } from './failure-mode';
|
|
12
|
-
export { IMPROVEMENT_KIND_SPEC } from './improvement';
|
|
13
|
-
export { KNOWLEDGE_GAP_KIND_SPEC } from './knowledge-gap';
|
|
14
|
-
export { KNOWLEDGE_POISONING_KIND_SPEC } from './knowledge-poisoning';
|
|
15
|
-
import type { TraceAnalystKindSpec } from '../kind-factory';
|
|
16
|
-
/**
|
|
17
|
-
* The default kind suite. Order is the run order operators should
|
|
18
|
-
* use: failure-mode first (no upstream deps), gap + poisoning next
|
|
19
|
-
* (both depend on failures), improvement last (chains all three).
|
|
20
|
-
*/
|
|
21
|
-
export declare const DEFAULT_TRACE_ANALYST_KINDS: readonly TraceAnalystKindSpec[];
|
|
22
|
-
//# sourceMappingURL=index.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,sBAAsB,EAAE,MAAM,gBAAgB,CAAA;AACvD,OAAO,EAAE,qBAAqB,EAAE,MAAM,eAAe,CAAA;AACrD,OAAO,EAAE,uBAAuB,EAAE,MAAM,iBAAiB,CAAA;AACzD,OAAO,EAAE,6BAA6B,EAAE,MAAM,uBAAuB,CAAA;AAErE,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAA;AAM3D;;;;GAIG;AACH,eAAO,MAAM,2BAA2B,EAAE,SAAS,oBAAoB,EAK7D,CAAA"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"kinds.test.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/kinds.test.ts"],"names":[],"mappings":""}
|
|
@@ -1,28 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Knowledge-gap analyst — what did the agent NOT know that it needed?
|
|
3
|
-
*
|
|
4
|
-
* Brief: find moments in the trace where the agent had to guess, ask
|
|
5
|
-
* the user to fill in context, recover from a wrong assumption, or
|
|
6
|
-
* loop on a retrieval. Each finding names a *missing or outdated piece
|
|
7
|
-
* of knowledge* the agent's curated knowledge base should have held —
|
|
8
|
-
* or a downstream lookup (web, docs, tool description) that surfaced
|
|
9
|
-
* stale or outdated information.
|
|
10
|
-
*
|
|
11
|
-
* The primary expected store is `@tangle-network/agent-knowledge`: a
|
|
12
|
-
* Karpathy-style wiki the agent maintains with raw ↔ curated pages,
|
|
13
|
-
* source anchors, and claim/relation triples. A gap is anything the
|
|
14
|
-
* agent had to discover at run-time that should already have lived
|
|
15
|
-
* there. Secondary loci: web-search results that returned outdated
|
|
16
|
-
* pages, tool descriptions that omitted critical behavior, system-
|
|
17
|
-
* prompt sections that didn't cover the case.
|
|
18
|
-
*
|
|
19
|
-
* Distinct from failure-mode: failure-mode classifies *how* it broke;
|
|
20
|
-
* knowledge-gap names the *information* whose absence (or staleness)
|
|
21
|
-
* caused the break. One failure-mode often maps to several gaps.
|
|
22
|
-
*
|
|
23
|
-
* Recursion (`maxDepth: 2`) is enough to fan out one subagent per
|
|
24
|
-
* candidate gap-source layer; each subagent runs a focused detection.
|
|
25
|
-
*/
|
|
26
|
-
import type { TraceAnalystKindSpec } from '../kind-factory';
|
|
27
|
-
export declare const KNOWLEDGE_GAP_KIND_SPEC: TraceAnalystKindSpec;
|
|
28
|
-
//# sourceMappingURL=knowledge-gap.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"knowledge-gap.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/knowledge-gap.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAEH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAA;AA+C3D,eAAO,MAAM,uBAAuB,EAAE,oBAWrC,CAAA"}
|
|
@@ -1,22 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Knowledge-poisoning analyst — what FALSE information misled the agent?
|
|
3
|
-
*
|
|
4
|
-
* Brief: find moments where the agent acted on information that was
|
|
5
|
-
* *wrong* — stale memory, RAG documents that contradicted ground truth,
|
|
6
|
-
* tool descriptions that lied about return shapes, system-prompt
|
|
7
|
-
* instructions that no longer matched reality, prior-run summaries that
|
|
8
|
-
* cached a wrong decision.
|
|
9
|
-
*
|
|
10
|
-
* Distinct from knowledge-gap: a gap is "the agent didn't know X"; a
|
|
11
|
-
* poisoning is "the agent confidently used X, but X was wrong." Gaps
|
|
12
|
-
* surface as questions / self-correction; poisonings surface as
|
|
13
|
-
* confident-but-wrong actions that downstream evidence contradicts.
|
|
14
|
-
*
|
|
15
|
-
* Recursion is moderate (`maxDepth: 2`) because each candidate
|
|
16
|
-
* poisoning typically needs two sub-investigations: one to confirm
|
|
17
|
-
* the agent acted on the false belief, one to confirm the belief
|
|
18
|
-
* itself is actually false in ground truth.
|
|
19
|
-
*/
|
|
20
|
-
import type { TraceAnalystKindSpec } from '../kind-factory';
|
|
21
|
-
export declare const KNOWLEDGE_POISONING_KIND_SPEC: TraceAnalystKindSpec;
|
|
22
|
-
//# sourceMappingURL=knowledge-poisoning.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"knowledge-poisoning.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/knowledge-poisoning.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAA;AAuC3D,eAAO,MAAM,6BAA6B,EAAE,oBAW3C,CAAA"}
|
|
@@ -1,84 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Skill-usage analyst — a DETERMINISTIC `Analyst` over a Claude/Codex skill
|
|
3
|
-
* library + its trace corpus. Unlike the trace-store kinds (failure-mode,
|
|
4
|
-
* improvement, ...) this kind calls no LLM: it mines real usage and skill
|
|
5
|
-
* structure and emits findings by rule.
|
|
6
|
-
*
|
|
7
|
-
* It exists because the naive "Skill-tool invocation count" lies low — it
|
|
8
|
-
* misses orchestrated sub-dispatch (a leaf skill run BY /pursue or /governor
|
|
9
|
-
* logs under the parent), slash-command entry, local-script bypass, and
|
|
10
|
-
* on-disk artifacts. The 2026-05-30 skill audit found 39/53 skills at zero
|
|
11
|
-
* direct invocations, yet only one was a genuine cut: the rest were
|
|
12
|
-
* measurement-invisible or discovery-limited. This analyst encodes that
|
|
13
|
-
* lesson as a multi-signal usage model so a cheap repeatable pass can keep
|
|
14
|
-
* the library honest, and so the expensive audit workflow's verdicts can
|
|
15
|
-
* GEPA-distill it toward agreement (see `gold/skill-verdicts.gold.jsonl`).
|
|
16
|
-
*
|
|
17
|
-
* Report-building (`buildSkillUsageReport`, an fs scan) is separated from
|
|
18
|
-
* finding emission (`SkillUsageAnalyst.analyze`, pure) so the slow scan runs
|
|
19
|
-
* once at the registry boundary and the rule logic stays unit-testable.
|
|
20
|
-
*/
|
|
21
|
-
import type { Analyst, AnalystContext, AnalystFinding } from '../types';
|
|
22
|
-
export type SkillKind = 'public' | 'private';
|
|
23
|
-
/** One skill's multi-signal usage + structure. All counts are deterministic. */
|
|
24
|
-
export interface SkillUsageRecord {
|
|
25
|
-
name: string;
|
|
26
|
-
kind: SkillKind;
|
|
27
|
-
/** Absolute path to the skill's SKILL.md. */
|
|
28
|
-
path: string;
|
|
29
|
-
lines: number;
|
|
30
|
-
/** `"skill":"<name>"` Skill-tool invocations across the trace corpus. */
|
|
31
|
-
directInvocations: number;
|
|
32
|
-
/** `<command-name>/<name>` slash invocations across the trace corpus. */
|
|
33
|
-
slashInvocations: number;
|
|
34
|
-
/** Sibling skills whose SKILL.md dispatches to this one (`/<name>`). Proxy
|
|
35
|
-
* for orchestrated sub-dispatch the per-skill counter cannot see. */
|
|
36
|
-
inboundRefs: number;
|
|
37
|
-
/** On-disk artifacts attributable to the skill (e.g. `.evolve/<name>/**`). */
|
|
38
|
-
artifactCount: number;
|
|
39
|
-
/** Tangle-private reference count in the body (leak signal for public skills). */
|
|
40
|
-
tanglePrivateRefs: number;
|
|
41
|
-
hasReferencesDir: boolean;
|
|
42
|
-
hasEvalsDir: boolean;
|
|
43
|
-
/** Body mentions `skill-runs.jsonl` (visible to /reflect + /governor). */
|
|
44
|
-
logsRuns: boolean;
|
|
45
|
-
/** Description carries an explicit `Triggers:` clause / trigger phrases. */
|
|
46
|
-
hasTriggerPhrases: boolean;
|
|
47
|
-
}
|
|
48
|
-
export interface SkillUsageReport {
|
|
49
|
-
generatedFromTraces: number;
|
|
50
|
-
records: SkillUsageRecord[];
|
|
51
|
-
}
|
|
52
|
-
export interface SkillUsageScanConfig {
|
|
53
|
-
/** Dirs holding `*.jsonl` transcripts (Claude `~/.claude/projects`, Codex sessions). */
|
|
54
|
-
transcriptDirs: string[];
|
|
55
|
-
/** Skill roots to scan; each dir directly under `root` with a `SKILL.md` is a skill. */
|
|
56
|
-
skillRoots: {
|
|
57
|
-
root: string;
|
|
58
|
-
kind: SkillKind;
|
|
59
|
-
}[];
|
|
60
|
-
/** Roots scanned for `<root>/.evolve/<skill>` artifact dirs. */
|
|
61
|
-
artifactRoots?: string[];
|
|
62
|
-
/** Token-prefixed mappings: skill name → extra artifact subpaths under an artifactRoot
|
|
63
|
-
* (e.g. reflect → `.evolve/reflections`). Catches non-eponymous artifact dirs. */
|
|
64
|
-
artifactAliases?: Record<string, string[]>;
|
|
65
|
-
/** Cap files read per transcript dir (bounds a huge corpus); 0 = unbounded. */
|
|
66
|
-
maxTranscriptsPerDir?: number;
|
|
67
|
-
}
|
|
68
|
-
/** Scan the corpus + skill roots into a {@link SkillUsageReport}. Deterministic. */
|
|
69
|
-
export declare function buildSkillUsageReport(config: SkillUsageScanConfig): SkillUsageReport;
|
|
70
|
-
/** Pure rule pass over a report → findings. Exported for direct/unit use. */
|
|
71
|
-
export declare function emitSkillUsageFindings(report: SkillUsageReport, producedAt: string): AnalystFinding[];
|
|
72
|
-
export declare class SkillUsageAnalyst implements Analyst<SkillUsageReport> {
|
|
73
|
-
readonly id = "skill-usage";
|
|
74
|
-
readonly description = "Deterministic multi-signal skill-usage analysis: flags dead skills, measurement-invisible (orchestrated) usage, discovery gaps, public-repo leaks, bloat, missing evals, and missing run-logging.";
|
|
75
|
-
readonly inputKind: "custom";
|
|
76
|
-
readonly cost: {
|
|
77
|
-
kind: "deterministic";
|
|
78
|
-
est_usd_per_run: number;
|
|
79
|
-
};
|
|
80
|
-
readonly version = "1.0.0";
|
|
81
|
-
analyze(input: SkillUsageReport, ctx: AnalystContext): Promise<AnalystFinding[]>;
|
|
82
|
-
}
|
|
83
|
-
export declare const SKILL_USAGE_ANALYST: SkillUsageAnalyst;
|
|
84
|
-
//# sourceMappingURL=skill-usage.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"skill-usage.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/skill-usage.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAIH,OAAO,KAAK,EAAE,OAAO,EAAE,cAAc,EAAE,cAAc,EAAmB,MAAM,UAAU,CAAA;AAKxF,MAAM,MAAM,SAAS,GAAG,QAAQ,GAAG,SAAS,CAAA;AAE5C,gFAAgF;AAChF,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,SAAS,CAAA;IACf,6CAA6C;IAC7C,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,MAAM,CAAA;IACb,yEAAyE;IACzE,iBAAiB,EAAE,MAAM,CAAA;IACzB,yEAAyE;IACzE,gBAAgB,EAAE,MAAM,CAAA;IACxB;0EACsE;IACtE,WAAW,EAAE,MAAM,CAAA;IACnB,8EAA8E;IAC9E,aAAa,EAAE,MAAM,CAAA;IACrB,kFAAkF;IAClF,iBAAiB,EAAE,MAAM,CAAA;IACzB,gBAAgB,EAAE,OAAO,CAAA;IACzB,WAAW,EAAE,OAAO,CAAA;IACpB,0EAA0E;IAC1E,QAAQ,EAAE,OAAO,CAAA;IACjB,4EAA4E;IAC5E,iBAAiB,EAAE,OAAO,CAAA;CAC3B;AAED,MAAM,WAAW,gBAAgB;IAC/B,mBAAmB,EAAE,MAAM,CAAA;IAC3B,OAAO,EAAE,gBAAgB,EAAE,CAAA;CAC5B;AAED,MAAM,WAAW,oBAAoB;IACnC,wFAAwF;IACxF,cAAc,EAAE,MAAM,EAAE,CAAA;IACxB,wFAAwF;IACxF,UAAU,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,SAAS,CAAA;KAAE,EAAE,CAAA;IAC/C,gEAAgE;IAChE,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;IACxB;uFACmF;IACnF,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC,CAAA;IAC1C,+EAA+E;IAC/E,oBAAoB,CAAC,EAAE,MAAM,CAAA;CAC9B;AAyED,oFAAoF;AACpF,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,oBAAoB,GAAG,gBAAgB,CA+EpF;AAiCD,6EAA6E;AAC7E,wBAAgB,sBAAsB,CACpC,MAAM,EAAE,gBAAgB,EACxB,UAAU,EAAE,MAAM,GACjB,cAAc,EAAE,CAuHlB;AAID,qBAAa,iBAAkB,YAAW,OAAO,CAAC,gBAAgB,CAAC;IACjE,QAAQ,CAAC,EAAE,iBAAa;IACxB,QAAQ,CAAC,WAAW,uMACiL;IACrM,QAAQ,CAAC,SAAS,EAAG,QAAQ,CAAS;IACtC,QAAQ,CAAC,IAAI;;;MAAyD;IACtE,QAAQ,CAAC,OAAO,WAAU;IAEpB,OAAO,CAAC,KAAK,EAAE,gBAAgB,EAAE,GAAG,EAAE,cAAc,GAAG,OAAO,CAAC,cAAc,EAAE,CAAC;CAOvF;AAED,eAAO,MAAM,mBAAmB,mBAA0B,CAAA"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"skill-usage.test.d.ts","sourceRoot":"","sources":["../../../src/analyst/kinds/skill-usage.test.ts"],"names":[],"mappings":""}
|
|
@@ -1,26 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Forgiving pre-parse for analyst findings. Weak models routinely emit
|
|
3
|
-
* schema-correct content in an unusable wrapper — fenced ```json blocks, a
|
|
4
|
-
* single object where an array is expected, trailing commas. Measured: GPT-4o
|
|
5
|
-
* drops to 0% usable output purely from markdown-fence wrapping
|
|
6
|
-
* (arXiv:2605.02363). A five-line de-fence recovers most of it. This module is
|
|
7
|
-
* the de-fence/coerce step that runs BEFORE Zod, so a recoverable finding is
|
|
8
|
-
* repaired, not dropped.
|
|
9
|
-
*
|
|
10
|
-
* Pure + deterministic. No model, no network.
|
|
11
|
-
*/
|
|
12
|
-
/** Strip a ```lang ... ``` (or bare ``` ... ```) code fence, if the string is one. */
|
|
13
|
-
export declare function stripCodeFences(text: string): string;
|
|
14
|
-
/**
|
|
15
|
-
* Best-effort parse of a string into JSON. De-fences, drops trailing commas,
|
|
16
|
-
* then `JSON.parse`. Returns `undefined` (never throws) when unrecoverable.
|
|
17
|
-
*/
|
|
18
|
-
export declare function coerceJson(text: string): unknown;
|
|
19
|
-
/**
|
|
20
|
-
* Coerce arbitrary actor/structurer output into an array of candidate finding
|
|
21
|
-
* rows: a JSON string → parse; a single object → 1-element array; an array →
|
|
22
|
-
* as-is; anything else → []. Callers still run each row through Zod
|
|
23
|
-
* (`parseRawFinding`) — this only fixes the SHAPE, never invents fields.
|
|
24
|
-
*/
|
|
25
|
-
export declare function coerceToFindingRows(raw: unknown): unknown[];
|
|
26
|
-
//# sourceMappingURL=parse-tolerant.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"parse-tolerant.d.ts","sourceRoot":"","sources":["../../src/analyst/parse-tolerant.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,sFAAsF;AACtF,wBAAgB,eAAe,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAKpD;AAOD;;;GAGG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAOhD;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,GAAG,EAAE,OAAO,GAAG,OAAO,EAAE,CAe3D"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"parse-tolerant.test.d.ts","sourceRoot":"","sources":["../../src/analyst/parse-tolerant.test.ts"],"names":[],"mappings":""}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"registry.budget.test.d.ts","sourceRoot":"","sources":["../../src/analyst/registry.budget.test.ts"],"names":[],"mappings":""}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"registry.d.ts","sourceRoot":"","sources":["../../src/analyst/registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAGH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAC/C,OAAO,KAAK,EACV,OAAO,EACP,cAAc,EACd,cAAc,EACd,eAAe,EACf,gBAAgB,EAChB,gBAAgB,EAChB,iBAAiB,EAClB,MAAM,SAAS,CAAA;AAIhB,MAAM,WAAW,YAAY;IAC3B,wFAAwF;IACxF,eAAe,CAAC,CAAC,IAAI,EAAE;QACrB,OAAO,EAAE,OAAO,CAAA;QAChB,GAAG,EAAE,cAAc,CAAA;QACnB,KAAK,EAAE,MAAM,CAAA;KACd,GAAG,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;IACxB,2FAA2F;IAC3F,cAAc,CAAC,CAAC,IAAI,EAAE;QACpB,OAAO,EAAE,OAAO,CAAA;QAChB,OAAO,EAAE,iBAAiB,CAAA;QAC1B,QAAQ,EAAE,cAAc,EAAE,CAAA;QAC1B,KAAK,EAAE,MAAM,CAAA;KACd,GAAG,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;IACxB;;;;OAIG;IACH,OAAO,CAAC,CAAC,IAAI,EAAE;QACb,OAAO,EAAE,OAAO,CAAA;QAChB,KAAK,EAAE,KAAK,CAAA;QACZ,KAAK,EAAE,MAAM,CAAA;KACd,GAAG,cAAc,EAAE,GAAG,SAAS,GAAG,OAAO,CAAC,cAAc,EAAE,GAAG,SAAS,CAAC,CAAA;IACxE,mFAAmF;IACnF,UAAU,CAAC,CAAC,IAAI,EAAE;QAAE,MAAM,EAAE,gBAAgB,CAAA;KAAE,GAAG,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;CACtE;AAED,MAAM,WAAW,YAAY;IAC3B,iDAAiD;IACjD,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,8EAA8E;IAC9E,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAChC;;;;OAIG;IACH,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE;QAChB,OAAO,EAAE,OAAO,CAAA;QAChB,QAAQ,EAAE,MAAM,GAAG,SAAS,CAAA;QAC5B,YAAY,EAAE,MAAM,GAAG,SAAS,CAAA;QAChC,YAAY,EAAE,MAAM,CAAA;KACrB,KAAK,MAAM,GAAG,SAAS,CAAA;CACzB;AAED,MAAM,WAAW,sBAAsB;IACrC,yEAAyE;IACzE,IAAI,CAAC,EAAE,UAAU,CAAA;IACjB,4CAA4C;IAC5C,GAAG,CAAC,EAAE,CAAC,GAAG,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAAK,IAAI,CAAA;IAC7D,2EAA2E;IAC3E,KAAK,CAAC,EAAE,YAAY,CAAA;IACpB,kDAAkD;IAClD,aAAa,CAAC,EAAE,YAAY,CAAA;CAC7B;AAED,MAAM,WAAW,eAAe;IAC9B,yDAAyD;IACzD,IAAI,CAAC,EAAE,MAAM,EAAE,CAAA;IACf,0EAA0E;IAC1E,IAAI,CAAC,EAAE,MAAM,EAAE,CAAA;IACf,kGAAkG;IAClG,MAAM,CAAC,EAAE,YAAY,CAAA;IACrB,8DAA8D;IAC9D,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,6DAA6D;IAC7D,MAAM,CAAC,EAAE,WAAW,CAAA;IACpB,kGAAkG;IAClG,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC7B;;;;;;;OAOG;IACH,aAAa,CAAC,EAAE,aAAa,CAAC,cAAc,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,cAAc,CAAC,CAAC,CAAA;CAC9F;AAED,qBAAa,eAAe;IAC1B,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAA6B;IACtD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAwB;gBAEpC,OAAO,GAAE,sBAA2B;IAIhD,QAAQ,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAWhC,IAAI,IAAI,aAAa,CAAC;QACpB,EAAE,EAAE,MAAM,CAAA;QACV,WAAW,EAAE,MAAM,CAAA;QACnB,OAAO,EAAE,MAAM,CAAA;QACf,IAAI,EAAE,OAAO,CAAC,MAAM,CAAC,CAAA;KACtB,CAAC;IASI,GAAG,CACP,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,gBAAgB,EACxB,OAAO,GAAE,eAAoB,GAC5B,OAAO,CAAC,gBAAgB,CAAC;IAS5B;;;;;;;;;;OAUG;IACI,SAAS,CACd,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,gBAAgB,EACxB,OAAO,GAAE,eAAoB,GAC5B,cAAc,CAAC,eAAe,EAAE,IAAI,EAAE,IAAI,CAAC;IAyI9C,OAAO,CAAC,cAAc;IAatB,OAAO,CAAC,UAAU;CAyBnB"}
|
|
@@ -1,35 +0,0 @@
|
|
|
1
|
-
import type { AnalystFinding } from './types';
|
|
2
|
-
/** DESCRIPTIVE predicate: does the finding cite at least one observable
|
|
3
|
-
* (span/event/artifact) evidence ref. Useful for ranking evidence quality or
|
|
4
|
-
* rendering — it is NOT the steer gate. Evidence presence is the WRONG
|
|
5
|
-
* discriminator for steering: a legitimate trace-analyst observation may cite
|
|
6
|
-
* nothing (it would be wrongly rejected), and a judge verdict may cite an
|
|
7
|
-
* artifact (it would be wrongly admitted). Use `assertNoJudgeVerdict` to gate
|
|
8
|
-
* steering; use this only where "is this grounded in observable evidence" is the
|
|
9
|
-
* literal question. */
|
|
10
|
-
export declare function isTraceObservable(finding: AnalystFinding): boolean;
|
|
11
|
-
/** True iff the finding is a JUDGE VERDICT (an acceptance score lifted into a
|
|
12
|
-
* finding), identified by provenance set at the lift site — independent of
|
|
13
|
-
* whatever evidence it cites. */
|
|
14
|
-
export declare function isJudgeVerdict(finding: AnalystFinding): boolean;
|
|
15
|
-
/**
|
|
16
|
-
* THE steer firewall. Fail-loud guard for any path that admits analyst findings
|
|
17
|
-
* as STEERING input (the `f(trace)` role): rejects — naming the offenders — any
|
|
18
|
-
* finding whose provenance is a judge verdict, rather than let `J` leak into the
|
|
19
|
-
* loop. Returns the findings unchanged for chaining.
|
|
20
|
-
*
|
|
21
|
-
* Call this at the chokepoint where a detector that ALSO scores/gates has its
|
|
22
|
-
* findings turned into a steer (the judge-and-steer dual-role case). It keys on
|
|
23
|
-
* provenance, so it correctly admits evidence-less trace-analyst observations and
|
|
24
|
-
* correctly rejects an artifact-citing judge verdict — the cases an evidence
|
|
25
|
-
* check gets backwards.
|
|
26
|
-
*
|
|
27
|
-
* It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge
|
|
28
|
-
* whose output is laundered through a hand-built finding with no provenance flag
|
|
29
|
-
* is out of its reach — provenance must be honestly set at every judge→finding
|
|
30
|
-
* lift (today: createJudgeAdapter). That is why the integrity rule lives at the
|
|
31
|
-
* lift site, and why ProposeContext.judgeScores?: never is the complementary
|
|
32
|
-
* compile-time tripwire on the obvious direct channel.
|
|
33
|
-
*/
|
|
34
|
-
export declare function assertNoJudgeVerdict(findings: ReadonlyArray<AnalystFinding>, context?: string): ReadonlyArray<AnalystFinding>;
|
|
35
|
-
//# sourceMappingURL=steer-firewall.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"steer-firewall.d.ts","sourceRoot":"","sources":["../../src/analyst/steer-firewall.ts"],"names":[],"mappings":"AAoBA,OAAO,KAAK,EAAE,cAAc,EAAe,MAAM,SAAS,CAAA;AAU1D;;;;;;;wBAOwB;AACxB,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,cAAc,GAAG,OAAO,CAElE;AAED;;kCAEkC;AAClC,wBAAgB,cAAc,CAAC,OAAO,EAAE,cAAc,GAAG,OAAO,CAE/D;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,aAAa,CAAC,cAAc,CAAC,EACvC,OAAO,SAAU,GAChB,aAAa,CAAC,cAAc,CAAC,CAW/B"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"steer-firewall.test.d.ts","sourceRoot":"","sources":["../../src/analyst/steer-firewall.test.ts"],"names":[],"mappings":""}
|
|
@@ -1,37 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* `structureFindings` — the deferred structuring pass (DSPy TwoStepAdapter /
|
|
3
|
-
* HALO `synthesize_traces` analog). The agentic actor reasons FREE-FORM and
|
|
4
|
-
* emits a prose `report` (which any model does reliably); this separate, cheap
|
|
5
|
-
* call's ONLY job is to turn that report into `AnalystFinding[]`. Decoupling
|
|
6
|
-
* reasoning from structuring is what makes the SEMANTIC findings model-agnostic
|
|
7
|
-
* — the reasoning model never has to satisfy a strict typed-array contract
|
|
8
|
-
* while it diagnoses.
|
|
9
|
-
*
|
|
10
|
-
* Forgiving: the response runs through `coerceToFindingRows` (de-fence, lift
|
|
11
|
-
* single→array) before Zod, and on a zero-finding extraction from a substantive
|
|
12
|
-
* report it reasks ONCE with the schema restated. Returns a typed outcome so a
|
|
13
|
-
* legitimate "nothing to report" is distinguishable from a failed extraction
|
|
14
|
-
* (no silent empty).
|
|
15
|
-
*/
|
|
16
|
-
import { type LlmClientOptions } from '../llm-client';
|
|
17
|
-
import { type AnalystFinding } from './types';
|
|
18
|
-
export interface StructureFindingsOptions {
|
|
19
|
-
/** The actor's free-form diagnosis prose. */
|
|
20
|
-
report: string;
|
|
21
|
-
analystId: string;
|
|
22
|
-
/** Coarse classification stamped on every extracted finding. */
|
|
23
|
-
area: string;
|
|
24
|
-
model: string;
|
|
25
|
-
baseUrl: string;
|
|
26
|
-
apiKey?: string;
|
|
27
|
-
/** Max reask attempts after a zero/invalid extraction. Default 1. */
|
|
28
|
-
maxReasks?: number;
|
|
29
|
-
/** Test seam: inject a fetch (no network in unit tests). */
|
|
30
|
-
fetchImpl?: LlmClientOptions['fetch'];
|
|
31
|
-
}
|
|
32
|
-
export interface StructureFindingsResult {
|
|
33
|
-
findings: AnalystFinding[];
|
|
34
|
-
outcome: 'ok' | 'extraction_failed';
|
|
35
|
-
}
|
|
36
|
-
export declare function structureFindings(opts: StructureFindingsOptions): Promise<StructureFindingsResult>;
|
|
37
|
-
//# sourceMappingURL=structure-findings.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"structure-findings.d.ts","sourceRoot":"","sources":["../../src/analyst/structure-findings.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAAW,KAAK,gBAAgB,EAAE,MAAM,eAAe,CAAA;AAG9D,OAAO,EAAE,KAAK,cAAc,EAAe,MAAM,SAAS,CAAA;AAE1D,MAAM,WAAW,wBAAwB;IACvC,6CAA6C;IAC7C,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,EAAE,MAAM,CAAA;IACjB,gEAAgE;IAChE,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,qEAAqE;IACrE,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,4DAA4D;IAC5D,SAAS,CAAC,EAAE,gBAAgB,CAAC,OAAO,CAAC,CAAA;CACtC;AAED,MAAM,WAAW,uBAAuB;IACtC,QAAQ,EAAE,cAAc,EAAE,CAAA;IAC1B,OAAO,EAAE,IAAI,GAAG,mBAAmB,CAAA;CACpC;AAmDD,wBAAsB,iBAAiB,CACrC,IAAI,EAAE,wBAAwB,GAC7B,OAAO,CAAC,uBAAuB,CAAC,CAyBlC"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"structure-findings.test.d.ts","sourceRoot":"","sources":["../../src/analyst/structure-findings.test.ts"],"names":[],"mappings":""}
|
|
@@ -1,34 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Pre-curated tool subsets for analyst kinds.
|
|
3
|
-
*
|
|
4
|
-
* The full trace-analyst tool set is seven functions. Most kinds only
|
|
5
|
-
* need three or four. Picking from named groups instead of importing
|
|
6
|
-
* the whole bundle keeps every kind's actor-context budget tight and
|
|
7
|
-
* makes "what can this analyst see?" obvious at registration time.
|
|
8
|
-
*
|
|
9
|
-
* Each function in the group keeps its full `name`/`description` from
|
|
10
|
-
* `buildTraceAnalystTools` — we filter, we don't re-implement.
|
|
11
|
-
*/
|
|
12
|
-
import type { AxFunction } from '@ax-llm/ax';
|
|
13
|
-
import type { TraceAnalysisStore } from '../trace-analyst/store';
|
|
14
|
-
/** Named tool sets. Kinds pass `tools: TRACE_TOOL_GROUPS.failureForensics` etc. */
|
|
15
|
-
export type TraceToolGroupName =
|
|
16
|
-
/** All seven tools. Use for open-ended discovery kinds. */
|
|
17
|
-
'all'
|
|
18
|
-
/** Overview + paginated query + count. No deep reads. Cheap. */
|
|
19
|
-
| 'discovery'
|
|
20
|
-
/** Discovery + viewTrace + viewSpans. Deep-read but no regex search. */
|
|
21
|
-
| 'discoveryAndRead'
|
|
22
|
-
/** Discovery + search tools. For pattern-matching across many traces. */
|
|
23
|
-
| 'discoveryAndSearch'
|
|
24
|
-
/** Discovery + viewSpans + searchSpan. Targeted-span work after another kind narrows down. */
|
|
25
|
-
| 'targeted';
|
|
26
|
-
/**
|
|
27
|
-
* Build the tool set for a named group bound to a specific trace store.
|
|
28
|
-
*
|
|
29
|
-
* `all` returns every tool. Other groups filter `buildTraceAnalystTools`
|
|
30
|
-
* by name to the documented subset. An unrecognised group name throws —
|
|
31
|
-
* silently returning all tools would defeat the cost-control point.
|
|
32
|
-
*/
|
|
33
|
-
export declare function buildTraceToolsForGroup(group: TraceToolGroupName, store: TraceAnalysisStore): AxFunction[];
|
|
34
|
-
//# sourceMappingURL=tool-groups.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"tool-groups.d.ts","sourceRoot":"","sources":["../../src/analyst/tool-groups.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,YAAY,CAAA;AAC5C,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,wBAAwB,CAAA;AAGhE,mFAAmF;AACnF,MAAM,MAAM,kBAAkB;AAC5B,2DAA2D;AACzD,KAAK;AACP,gEAAgE;GAC9D,WAAW;AACb,wEAAwE;GACtE,kBAAkB;AACpB,yEAAyE;GACvE,oBAAoB;AACtB,8FAA8F;GAC5F,UAAU,CAAA;AAsBd;;;;;;GAMG;AACH,wBAAgB,uBAAuB,CACrC,KAAK,EAAE,kBAAkB,EACzB,KAAK,EAAE,kBAAkB,GACxB,UAAU,EAAE,CAMd"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/analyst/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAGH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,eAAe,CAAA;AAC9C,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,wBAAwB,CAAA;AAChE,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,UAAU,CAAA;AAC1C,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAE/C;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC7B,cAAc,EAAE,OAAO,CAAA;IACvB;;;;;OAKG;IACH,UAAU,EAAE,MAAM,CAAA;IAClB,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,MAAM,CAAA;IACnB,QAAQ,EAAE,eAAe,CAAA;IACzB;;;;;OAKG;IACH,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,MAAM,CAAA;IACb,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,aAAa,EAAE,WAAW,EAAE,CAAA;IAC5B,kBAAkB,CAAC,EAAE,MAAM,CAAA;IAC3B,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,2EAA2E;IAC3E,UAAU,EAAE,MAAM,CAAA;IAClB;;;;OAIG;IACH,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB;;;;;;;6DAOyD;IACzD,kBAAkB,CAAC,EAAE,OAAO,CAAA;IAC5B,6EAA6E;IAC7E,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,MAAM,eAAe,GAAG,UAAU,GAAG,MAAM,GAAG,QAAQ,GAAG,KAAK,GAAG,MAAM,CAAA;AAE7E,MAAM,WAAW,WAAW;IAC1B;;;;;OAKG;IACH,IAAI,EAAE,MAAM,GAAG,OAAO,GAAG,UAAU,GAAG,SAAS,GAAG,QAAQ,CAAA;IAC1D,GAAG,EAAE,MAAM,CAAA;IACX,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAID;;;;;GAKG;AACH,MAAM,MAAM,gBAAgB,GACxB,aAAa,GACb,cAAc,GACd,YAAY,GACZ,aAAa,GACb,QAAQ,CAAA;AAEZ,MAAM,WAAW,WAAW;IAC1B,sDAAsD;IACtD,IAAI,EAAE,eAAe,GAAG,KAAK,CAAA;IAC7B,wEAAwE;IACxE,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,yDAAyD;IACzD,MAAM,CAAC,EAAE,MAAM,EAAE,CAAA;CAClB;AAED,MAAM,WAAW,mBAAmB;IAClC,yEAAyE;IACzE,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,sEAAsE;IACtE,YAAY,CAAC,EAAE,MAAM,EAAE,CAAA;CACxB;AAED;;;GAGG;AACH,MAAM,WAAW,gBAAgB;IAC/B,UAAU,CAAC,EAAE,kBAAkB,CAAA;IAC/B,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,SAAS,CAAC,EAAE,SAAS,CAAA;IACrB,UAAU,CAAC,EAAE,UAAU,CAAA;IACvB,iFAAiF;IACjF,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACjC;AAED,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,MAAM,CAAA;IACb,8EAA8E;IAC9E,aAAa,EAAE,MAAM,CAAA;IACrB,iFAAiF;IACjF,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,2EAA2E;IAC3E,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB;;;;;OAKG;IACH,IAAI,CAAC,EAAE,UAAU,CAAA;IACjB;;;;;;;;OAQG;IACH,aAAa,CAAC,EAAE,aAAa,CAAC,cAAc,CAAC,CAAA;IAC7C,qFAAqF;IACrF,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC7B,oFAAoF;IACpF,GAAG,CAAC,EAAE,CAAC,GAAG,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAAK,IAAI,CAAA;IAC7D,2EAA2E;IAC3E,MAAM,CAAC,EAAE,WAAW,CAAA;CACrB;AAED;;;;;GAKG;AACH,MAAM,WAAW,OAAO,CAAC,MAAM,GAAG,OAAO;IACvC,0FAA0F;IAC1F,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAA;IACnB,oCAAoC;IACpC,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAA;IAC5B,QAAQ,CAAC,SAAS,EAAE,gBAAgB,CAAA;IACpC,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAA;IAC1B,QAAQ,CAAC,QAAQ,CAAC,EAAE,mBAAmB,CAAA;IACvC,0FAA0F;IAC1F,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;IACxB,OAAO,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,cAAc,GAAG,OAAO,CAAC,cAAc,EAAE,CAAC,CAAA;CACvE;AAID;;;;;;;GAOG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE;IACtC,UAAU,EAAE,MAAM,CAAA;IAClB,IAAI,EAAE,MAAM,CAAA;IACZ,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,KAAK,EAAE,MAAM,CAAA;IACb,2FAA2F;IAC3F,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB,GAAG,MAAM,CAQT;AAYD;;;GAGG;AACH,wBAAgB,WAAW,CACzB,IAAI,EAAE,IAAI,CAAC,cAAc,EAAE,gBAAgB,GAAG,YAAY,GAAG,aAAa,CAAC,GAAG;IAC5E,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,WAAW,CAAC,EAAE,MAAM,CAAA;CACrB,GACA,cAAc,CAchB;AAID,MAAM,WAAW,iBAAiB;IAChC,UAAU,EAAE,MAAM,CAAA;IAClB,MAAM,EAAE,IAAI,GAAG,SAAS,GAAG,QAAQ,CAAA;IACnC,sEAAsE;IACtE,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,cAAc,EAAE,MAAM,CAAA;IACtB,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,MAAM,CAAA;IAChB,+EAA+E;IAC/E,KAAK,CAAC,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAA;CAC3C;AAED,MAAM,WAAW,gBAAgB;IAC/B,MAAM,EAAE,MAAM,CAAA;IACd,cAAc,EAAE,MAAM,CAAA;IACtB,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,MAAM,CAAA;IAChB,QAAQ,EAAE,cAAc,EAAE,CAAA;IAC1B,WAAW,EAAE,iBAAiB,EAAE,CAAA;IAChC,wEAAwE;IACxE,cAAc,EAAE,MAAM,CAAA;CACvB;AAID;;;;;;;;;;;;GAYG;AACH,MAAM,MAAM,eAAe,GACvB;IACE,IAAI,EAAE,aAAa,CAAA;IACnB,MAAM,EAAE,MAAM,CAAA;IACd,cAAc,EAAE,MAAM,CAAA;IACtB,UAAU,EAAE,MAAM,CAAA;IAClB,6DAA6D;IAC7D,WAAW,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;CACnC,GACD;IACE,IAAI,EAAE,iBAAiB,CAAA;IACvB,OAAO,EAAE,iBAAiB,CAAA;CAC3B,GACD;IACE,IAAI,EAAE,iBAAiB,CAAA;IACvB,UAAU,EAAE,MAAM,CAAA;IAClB,UAAU,EAAE,MAAM,CAAA;CACnB,GACD;IACE,IAAI,EAAE,mBAAmB,CAAA;IACzB,yFAAyF;IACzF,OAAO,EAAE,iBAAiB,CAAA;IAC1B,QAAQ,EAAE,aAAa,CAAC,cAAc,CAAC,CAAA;CACxC,GACD;IACE,IAAI,EAAE,eAAe,CAAA;IACrB,MAAM,EAAE,gBAAgB,CAAA;CACzB,CAAA"}
|
package/dist/anti-slop.d.ts
DELETED
|
@@ -1,59 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Anti-slop quality judge.
|
|
3
|
-
*
|
|
4
|
-
* Deterministic pattern-based quality check — no LLM call. Catches the
|
|
5
|
-
* 80% of AI slop that every production agent leaks:
|
|
6
|
-
* - Banned phrases (voice-specific: "delve", "it's worth noting", etc.)
|
|
7
|
-
* - N-gram repetition (same phrase over and over)
|
|
8
|
-
* - Hedging overuse ("I could be wrong, but...")
|
|
9
|
-
* - Apology padding ("I'm so sorry for the confusion...")
|
|
10
|
-
* - Unused opening formulas ("Great question!")
|
|
11
|
-
* - Length bounds (too short to be useful, too long to be read)
|
|
12
|
-
*
|
|
13
|
-
* Produces a JudgeScore in the same shape as LLM judges so it composes into
|
|
14
|
-
* `BenchmarkRunner`'s judge array transparently.
|
|
15
|
-
*/
|
|
16
|
-
import type { JudgeFn } from './types';
|
|
17
|
-
export interface AntiSlopConfig {
|
|
18
|
-
/** Domain label — appears in the JudgeScore output */
|
|
19
|
-
domain?: string;
|
|
20
|
-
/** Case-insensitive substrings that must not appear. Each occurrence = penalty. */
|
|
21
|
-
bannedPhrases?: string[];
|
|
22
|
-
/** Regexes matching opening formulas to penalize (e.g. /^great question/i). */
|
|
23
|
-
bannedOpenings?: RegExp[];
|
|
24
|
-
/** Regexes matching hedges (e.g. /i could be wrong/i). Ratio of hedged sentences drives score. */
|
|
25
|
-
hedgingPatterns?: RegExp[];
|
|
26
|
-
/** Regexes matching apology padding. */
|
|
27
|
-
apologyPatterns?: RegExp[];
|
|
28
|
-
/** Fraction of sentences that can be duplicates before penalty (default 0.15 = 15%). */
|
|
29
|
-
repetitionThreshold?: number;
|
|
30
|
-
/** Min output length in chars; below this the turn is deemed too terse. */
|
|
31
|
-
minLength?: number;
|
|
32
|
-
/** Max output length in chars; above this the turn is deemed too verbose. */
|
|
33
|
-
maxLength?: number;
|
|
34
|
-
/** How heavily each violation class reduces the score (default 1). */
|
|
35
|
-
penaltyWeights?: Partial<Record<SlopCategory, number>>;
|
|
36
|
-
}
|
|
37
|
-
export type SlopCategory = 'banned_phrase' | 'banned_opening' | 'hedging' | 'apology' | 'repetition' | 'length';
|
|
38
|
-
/** Create a reusable Judge function from an anti-slop config. */
|
|
39
|
-
export declare function createAntiSlopJudge(config?: AntiSlopConfig): JudgeFn;
|
|
40
|
-
export interface AntiSlopIssue {
|
|
41
|
-
category: SlopCategory;
|
|
42
|
-
detail: string;
|
|
43
|
-
example?: string;
|
|
44
|
-
}
|
|
45
|
-
export interface AntiSlopReport {
|
|
46
|
-
/** 0–10 score; 10 is clean, lower values mean more slop. */
|
|
47
|
-
score: number;
|
|
48
|
-
issues: AntiSlopIssue[];
|
|
49
|
-
/** Count of each category for programmatic aggregation. */
|
|
50
|
-
counts: Record<SlopCategory, number>;
|
|
51
|
-
}
|
|
52
|
-
/**
|
|
53
|
-
* Pure function — analyze one or more outputs against the config. Exposed
|
|
54
|
-
* separately so consumers can build their own reporters on top.
|
|
55
|
-
*/
|
|
56
|
-
export declare function analyzeAntiSlop(outputs: string[], config: Omit<Required<AntiSlopConfig>, 'domain'> & {
|
|
57
|
-
penaltyWeights: Record<SlopCategory, number>;
|
|
58
|
-
}): AntiSlopReport;
|
|
59
|
-
//# sourceMappingURL=anti-slop.d.ts.map
|
package/dist/anti-slop.d.ts.map
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"anti-slop.d.ts","sourceRoot":"","sources":["../src/anti-slop.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,KAAK,EAAE,OAAO,EAA0B,MAAM,SAAS,CAAA;AAE9D,MAAM,WAAW,cAAc;IAC7B,sDAAsD;IACtD,MAAM,CAAC,EAAE,MAAM,CAAA;IAEf,mFAAmF;IACnF,aAAa,CAAC,EAAE,MAAM,EAAE,CAAA;IAExB,+EAA+E;IAC/E,cAAc,CAAC,EAAE,MAAM,EAAE,CAAA;IAEzB,kGAAkG;IAClG,eAAe,CAAC,EAAE,MAAM,EAAE,CAAA;IAE1B,wCAAwC;IACxC,eAAe,CAAC,EAAE,MAAM,EAAE,CAAA;IAE1B,wFAAwF;IACxF,mBAAmB,CAAC,EAAE,MAAM,CAAA;IAE5B,2EAA2E;IAC3E,SAAS,CAAC,EAAE,MAAM,CAAA;IAElB,6EAA6E;IAC7E,SAAS,CAAC,EAAE,MAAM,CAAA;IAElB,sEAAsE;IACtE,cAAc,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAA;CACvD;AAED,MAAM,MAAM,YAAY,GACpB,eAAe,GACf,gBAAgB,GAChB,SAAS,GACT,SAAS,GACT,YAAY,GACZ,QAAQ,CAAA;AAeZ,iEAAiE;AACjE,wBAAgB,mBAAmB,CAAC,MAAM,GAAE,cAAmB,GAAG,OAAO,CA0CxE;AAED,MAAM,WAAW,aAAa;IAC5B,QAAQ,EAAE,YAAY,CAAA;IACtB,MAAM,EAAE,MAAM,CAAA;IACd,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAED,MAAM,WAAW,cAAc;IAC7B,4DAA4D;IAC5D,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,aAAa,EAAE,CAAA;IACvB,2DAA2D;IAC3D,MAAM,EAAE,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAA;CACrC;AAED;;;GAGG;AACH,wBAAgB,eAAe,CAC7B,OAAO,EAAE,MAAM,EAAE,EACjB,MAAM,EAAE,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC,EAAE,QAAQ,CAAC,GAAG;IACjD,cAAc,EAAE,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAA;CAC7C,GACA,cAAc,CAgHhB"}
|