@tangle-network/agent-eval 0.221.8 → 0.222.0
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 +82 -0
- package/dist/adapters/http.d.ts +3 -5
- package/dist/adapters/http.d.ts.map +1 -1
- package/dist/adapters/http.js +1 -2
- package/dist/{adversarial-smnADNFS.d.ts → adversarial-DnQQFqTR.d.ts} +1 -3
- package/dist/adversarial-DnQQFqTR.d.ts.map +1 -0
- package/dist/{agent-profile-CNmwfaqq.d.ts → agent-profile-CoD1Dotf.d.ts} +4 -10
- package/dist/agent-profile-CoD1Dotf.d.ts.map +1 -0
- package/dist/agent-profile-Dngk3LZ0.js +153 -0
- package/dist/agent-profile-Dngk3LZ0.js.map +1 -0
- package/dist/{agent-profile-cell-BcpFbNd5.d.ts → agent-profile-cell-3kYxUsK2.d.ts} +3 -5
- package/dist/agent-profile-cell-3kYxUsK2.d.ts.map +1 -0
- package/dist/{agent-profile-cell-B1V1gUZs.js → agent-profile-cell-CT5JYueM.js} +5 -7
- package/dist/{agent-profile-cell-B1V1gUZs.js.map → agent-profile-cell-CT5JYueM.js.map} +1 -1
- package/dist/{agreement-irr-BhAqjluA.js → agreement-irr-BTjDN0i6.js} +4 -5
- package/dist/{agreement-irr-BhAqjluA.js.map → agreement-irr-BTjDN0i6.js.map} +1 -1
- package/dist/analyst/index.d.ts +95 -159
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +28 -39
- package/dist/analyst/index.js.map +1 -1
- package/dist/{analyze-runs-DfX6dHP0.d.ts → analyze-runs-C-pP1sK9.d.ts} +7 -13
- package/dist/analyze-runs-C-pP1sK9.d.ts.map +1 -0
- package/dist/{analyze-runs-CHmsykNJ.js → analyze-runs-CTheASYV.js} +23 -32
- package/dist/{analyze-runs-CHmsykNJ.js.map → analyze-runs-CTheASYV.js.map} +1 -1
- package/dist/{attribute-vocabulary-CcXqHODT.d.ts → attribute-vocabulary-5oo8ykIx.d.ts} +2 -4
- package/dist/attribute-vocabulary-5oo8ykIx.d.ts.map +1 -0
- package/dist/authenticity/index.d.ts +0 -2
- package/dist/authenticity/index.d.ts.map +1 -1
- package/dist/authenticity/index.js +1 -2
- package/dist/{benchmark-command-Cnjfo0Zh.js → benchmark-command-jYWR2gya.js} +61 -94
- package/dist/benchmark-command-jYWR2gya.js.map +1 -0
- package/dist/benchmarks/index.d.ts +5 -15
- package/dist/benchmarks/index.d.ts.map +1 -1
- package/dist/benchmarks/index.js +27 -38
- package/dist/benchmarks/index.js.map +1 -1
- package/dist/{bounded-process-Cj9YjXbE.js → bounded-process-BZrWXIKW.js} +2 -3
- package/dist/{bounded-process-Cj9YjXbE.js.map → bounded-process-BZrWXIKW.js.map} +1 -1
- package/dist/builder-eval/index.d.ts +5 -14
- package/dist/builder-eval/index.d.ts.map +1 -1
- package/dist/builder-eval/index.js +11 -17
- package/dist/builder-eval/index.js.map +1 -1
- package/dist/{calibration-CTwxpJ-t.js → calibration-6usfVYYo.js} +4 -5
- package/dist/{calibration-CTwxpJ-t.js.map → calibration-6usfVYYo.js.map} +1 -1
- package/dist/campaign/index.d.ts +9 -9
- package/dist/campaign/index.js +15 -14
- package/dist/{campaign-uddTlFg5.js → campaign-ddr1N3yD.js} +61 -58
- package/dist/campaign-ddr1N3yD.js.map +1 -0
- package/dist/{campaign-evidence-DunzcJGv.js → campaign-evidence-DN8rWuBj.js} +20 -322
- package/dist/campaign-evidence-DN8rWuBj.js.map +1 -0
- package/dist/{canonical-Dq2hOQ8L.js → canonical-CicJlINA.js} +3 -4
- package/dist/{canonical-Dq2hOQ8L.js.map → canonical-CicJlINA.js.map} +1 -1
- package/dist/{canonical-CFpojCN5.d.ts → canonical-gLeHiqzd.d.ts} +2 -4
- package/dist/canonical-gLeHiqzd.d.ts.map +1 -0
- package/dist/{cell-spend-DEWR3nj-.js → cell-spend-zxmzg4GH.js} +3 -4
- package/dist/{cell-spend-DEWR3nj-.js.map → cell-spend-zxmzg4GH.js.map} +1 -1
- package/dist/{chat-client-BNATB3rJ.js → chat-client-CO3uB33i.js} +3 -4
- package/dist/{chat-client-BNATB3rJ.js.map → chat-client-CO3uB33i.js.map} +1 -1
- package/dist/{chat-json-call-BzWYp18J.js → chat-json-call-8yUv4LFO.js} +4 -5
- package/dist/{chat-json-call-BzWYp18J.js.map → chat-json-call-8yUv4LFO.js.map} +1 -1
- package/dist/{claim-BPVuyHuy.js → claim-Dod_mjk3.js} +5 -6
- package/dist/{claim-BPVuyHuy.js.map → claim-Dod_mjk3.js.map} +1 -1
- package/dist/cli.js +179 -23
- package/dist/cli.js.map +1 -1
- package/dist/{client-hwo3cJjR.d.ts → client-BIFwUg7I.d.ts} +6 -12
- package/dist/client-BIFwUg7I.d.ts.map +1 -0
- package/dist/{cluster-bootstrap-DEwwhl0C.js → cluster-bootstrap-SbaeM44s.js} +4 -5
- package/dist/{cluster-bootstrap-DEwwhl0C.js.map → cluster-bootstrap-SbaeM44s.js.map} +1 -1
- package/dist/completion-verifier-7KEfroqx.d.ts +168 -0
- package/dist/completion-verifier-7KEfroqx.d.ts.map +1 -0
- package/dist/{completion-verifier-asaV1nSj.js → completion-verifier-rqbMh5-4.js} +9 -160
- package/dist/completion-verifier-rqbMh5-4.js.map +1 -0
- package/dist/contract/index.d.ts +13 -13
- package/dist/contract/index.js +41 -60
- package/dist/contract/index.js.map +1 -1
- package/dist/{cost-ledger-BGZB6yY2.js → cost-ledger-CS8R3IZj.js} +10 -17
- package/dist/{cost-ledger-BGZB6yY2.js.map → cost-ledger-CS8R3IZj.js.map} +1 -1
- package/dist/{cost-ledger-C1W3ICmP.d.ts → cost-ledger-Czro9wmo.d.ts} +1 -5
- package/dist/cost-ledger-Czro9wmo.d.ts.map +1 -0
- package/dist/{dataset-DQqhOCPt.d.ts → dataset-DThC1S6G.d.ts} +1 -3
- package/dist/dataset-DThC1S6G.d.ts.map +1 -0
- package/dist/{decision-point-0nJubuoM.js → decision-point-lhMgODOR.js} +14 -18
- package/dist/{decision-point-0nJubuoM.js.map → decision-point-lhMgODOR.js.map} +1 -1
- package/dist/{deep-freeze-CurcOQg3.js → deep-freeze-CK8-2W7p.js} +2 -3
- package/dist/{deep-freeze-CurcOQg3.js.map → deep-freeze-CK8-2W7p.js.map} +1 -1
- package/dist/{default-registry-CLG5YqGS.d.ts → default-registry-B1DstPYg.d.ts} +53 -61
- package/dist/default-registry-B1DstPYg.d.ts.map +1 -0
- package/dist/{default-registry-BBTDgkDM.js → default-registry-C1tPEjZ4.js} +29 -45
- package/dist/{default-registry-BBTDgkDM.js.map → default-registry-C1tPEjZ4.js.map} +1 -1
- package/dist/{define-agent-eval-CCbT7Kfr.js → define-agent-eval-BLoP4Lf9.js} +15 -18
- package/dist/{define-agent-eval-CCbT7Kfr.js.map → define-agent-eval-BLoP4Lf9.js.map} +1 -1
- package/dist/{define-agent-eval-JKt3gMQr.d.ts → define-agent-eval-CAffiFbN.d.ts} +19 -27
- package/dist/define-agent-eval-CAffiFbN.d.ts.map +1 -0
- package/dist/{descriptive-CVNy48rR.js → descriptive-cPwITVfy.js} +30 -8
- package/dist/descriptive-cPwITVfy.js.map +1 -0
- package/dist/diagnosis/index.d.ts +5 -15
- package/dist/diagnosis/index.d.ts.map +1 -1
- package/dist/diagnosis/index.js +13 -19
- package/dist/diagnosis/index.js.map +1 -1
- package/dist/{dspy-rlm-engine-DN1p_5jx.js → dspy-rlm-engine-jDiWkul6.js} +7 -10
- package/dist/{dspy-rlm-engine-DN1p_5jx.js.map → dspy-rlm-engine-jDiWkul6.js.map} +1 -1
- package/dist/{emitter-DbMXUIVP.js → emitter-DhQ65mza.js} +3 -4
- package/dist/{emitter-DbMXUIVP.js.map → emitter-DhQ65mza.js.map} +1 -1
- package/dist/{emitter-DABYYTcH.d.ts → emitter-mAXC5PSN.d.ts} +3 -5
- package/dist/emitter-mAXC5PSN.d.ts.map +1 -0
- package/dist/{engine-dtRVeuQs.d.ts → engine-CaHaCc3W.d.ts} +10 -18
- package/dist/engine-CaHaCc3W.d.ts.map +1 -0
- package/dist/{equal-terms-DbPGcVtI.js → equal-terms-K11VHmR3.js} +3 -4
- package/dist/{equal-terms-DbPGcVtI.js.map → equal-terms-K11VHmR3.js.map} +1 -1
- package/dist/{errors-BcOs5MKW.js → errors-DeEEV9Vx.js} +2 -3
- package/dist/{errors-BcOs5MKW.js.map → errors-DeEEV9Vx.js.map} +1 -1
- package/dist/{errors-DEE6u6ot.d.ts → errors-E9I2hldP.d.ts} +1 -3
- package/dist/errors-E9I2hldP.d.ts.map +1 -0
- package/dist/{eval-campaign-DoX5vz3E.js → eval-campaign-Cgixrvf7.js} +8 -9
- package/dist/{eval-campaign-DoX5vz3E.js.map → eval-campaign-Cgixrvf7.js.map} +1 -1
- package/dist/{evaluation-wdGyZ7jV.js → evaluation-B8H3GYej.js} +4 -6
- package/dist/{evaluation-wdGyZ7jV.js.map → evaluation-B8H3GYej.js.map} +1 -1
- package/dist/{evaluation-DblGquCB.d.ts → evaluation-BPqUvZ66.d.ts} +4 -6
- package/dist/evaluation-BPqUvZ66.d.ts.map +1 -0
- package/dist/evaluation.d.ts +1 -1
- package/dist/evaluation.js +1 -2
- package/dist/{exact-types-Do63KoKn.d.ts → exact-types-D98BKpxo.d.ts} +8 -10
- package/dist/exact-types-D98BKpxo.d.ts.map +1 -0
- package/dist/{exec-CpbtOXVt.js → exec-C0_AkLUv.js} +3 -5
- package/dist/{exec-CpbtOXVt.js.map → exec-C0_AkLUv.js.map} +1 -1
- package/dist/{execution-tracks-CpgFPpS5.js → execution-tracks-DaDXYwAq.js} +2 -3
- package/dist/{execution-tracks-CpgFPpS5.js.map → execution-tracks-DaDXYwAq.js.map} +1 -1
- package/dist/experiment/index.d.ts +9 -37
- package/dist/experiment/index.d.ts.map +1 -1
- package/dist/experiment/index.js +29 -40
- package/dist/experiment/index.js.map +1 -1
- package/dist/{exporters-Dk3oEbK_.d.ts → exporters-B5Ze_RKh.d.ts} +4 -4
- package/dist/exporters-B5Ze_RKh.d.ts.map +1 -0
- package/dist/{exporters-DWvM16LE.js → exporters-ByLcILYU.js} +23 -5
- package/dist/exporters-ByLcILYU.js.map +1 -0
- package/dist/{external-optimizer-process-DEzJ2CGX.js → external-optimizer-process-BRVmaDFe.js} +5 -6
- package/dist/{external-optimizer-process-DEzJ2CGX.js.map → external-optimizer-process-BRVmaDFe.js.map} +1 -1
- package/dist/{external-optimizer-subprocess-B_ukPtRD.js → external-optimizer-subprocess-U616hV_l.js} +24 -37
- package/dist/{external-optimizer-subprocess-B_ukPtRD.js.map → external-optimizer-subprocess-U616hV_l.js.map} +1 -1
- package/dist/{extract-usage-BZLU01kR.js → extract-usage-ISFqhfDz.js} +2 -3
- package/dist/{extract-usage-BZLU01kR.js.map → extract-usage-ISFqhfDz.js.map} +1 -1
- package/dist/{failure-cluster-DN0t63Z0.d.ts → failure-cluster-DOoa7CBg.d.ts} +3 -7
- package/dist/failure-cluster-DOoa7CBg.d.ts.map +1 -0
- package/dist/{failure-taxonomy-Bx5oWqaW.js → failure-taxonomy-9eQfXKoW.js} +2 -3
- package/dist/{failure-taxonomy-Bx5oWqaW.js.map → failure-taxonomy-9eQfXKoW.js.map} +1 -1
- package/dist/{feedback-trajectory-CYVJXolG.d.ts → feedback-trajectory-SBGncrCP.d.ts} +7 -13
- package/dist/feedback-trajectory-SBGncrCP.d.ts.map +1 -0
- package/dist/{finding-composition-8n0fFgVI.js → finding-composition-DlvE1Qul.js} +4 -6
- package/dist/{finding-composition-8n0fFgVI.js.map → finding-composition-DlvE1Qul.js.map} +1 -1
- package/dist/{findings-store-Bg4sdS-e.js → findings-store-B75Dspaq.js} +11 -14
- package/dist/{findings-store-Bg4sdS-e.js.map → findings-store-B75Dspaq.js.map} +1 -1
- package/dist/fuzz.d.ts +3 -17
- package/dist/fuzz.d.ts.map +1 -1
- package/dist/fuzz.js +30 -29
- package/dist/fuzz.js.map +1 -1
- package/dist/{hf-dataset-CKxIJQba.js → hf-dataset-DPVQXW5n.js} +362 -21
- package/dist/hf-dataset-DPVQXW5n.js.map +1 -0
- package/dist/hosted/index.d.ts +2 -6
- package/dist/hosted/index.d.ts.map +1 -1
- package/dist/hosted/index.js +1 -2
- package/dist/{index-CalAsXMR.d.ts → index-B1Bm9oJA.d.ts} +3 -15
- package/dist/index-B1Bm9oJA.d.ts.map +1 -0
- package/dist/{index-CdEVr8-O.d.ts → index-B7v2QHlD.d.ts} +24 -299
- package/dist/index-B7v2QHlD.d.ts.map +1 -0
- package/dist/{index-CTIRz-AB.d.ts → index-CUjeTDPv.d.ts} +14 -36
- package/dist/index-CUjeTDPv.d.ts.map +1 -0
- package/dist/{index-CfJULcW5.d.ts → index-DAk2quYu.d.ts} +18 -38
- package/dist/index-DAk2quYu.d.ts.map +1 -0
- package/dist/{index-D7fmC-1K.d.ts → index-DUn0GNiZ.d.ts} +183 -36
- package/dist/index-DUn0GNiZ.d.ts.map +1 -0
- package/dist/{index-CHxwfW5g.d.ts → index-DwUGehqX.d.ts} +3 -9
- package/dist/index-DwUGehqX.d.ts.map +1 -0
- package/dist/{index-jnuFLxOs.d.ts → index-aDPVYBIM.d.ts} +11 -39
- package/dist/index-aDPVYBIM.d.ts.map +1 -0
- package/dist/index.d.ts +55 -188
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +122 -195
- package/dist/index.js.map +1 -1
- package/dist/{integrity-D-s2iHiT.js → integrity-B1V6klyO.js} +12 -363
- package/dist/integrity-B1V6klyO.js.map +1 -0
- package/dist/{integrity-D9lBnv_K.d.ts → integrity-BAdVzT8u.d.ts} +4 -6
- package/dist/integrity-BAdVzT8u.d.ts.map +1 -0
- package/dist/{integrity-W4CADcFc.js → integrity-DOHt1mwU.js} +3 -4
- package/dist/{integrity-W4CADcFc.js.map → integrity-DOHt1mwU.js.map} +1 -1
- package/dist/{internal-BTK91Bqt.js → internal-v6XfOeJW.js} +6 -8
- package/dist/{internal-BTK91Bqt.js.map → internal-v6XfOeJW.js.map} +1 -1
- package/dist/{interval-Dq9gt7jx.js → interval-Bd2f4CRT.js} +6 -8
- package/dist/{interval-Dq9gt7jx.js.map → interval-Bd2f4CRT.js.map} +1 -1
- package/dist/{interval-ohuOJurX.d.ts → interval-CWyxqDpE.d.ts} +1 -5
- package/dist/interval-CWyxqDpE.d.ts.map +1 -0
- package/dist/{journal-C7om8Gum.js → journal-C1uLJi0w.js} +115 -121
- package/dist/{journal-C7om8Gum.js.map → journal-C1uLJi0w.js.map} +1 -1
- package/dist/{json-BhdsJaOQ.js → json-DtvZNAU9.js} +4 -6
- package/dist/{json-BhdsJaOQ.js.map → json-DtvZNAU9.js.map} +1 -1
- package/dist/json-recovery-DK4MDm5r.js +296 -0
- package/dist/json-recovery-DK4MDm5r.js.map +1 -0
- package/dist/{judge-calibration-BkXMIWA4.js → judge-calibration-DGqlfWW8.js} +5 -6
- package/dist/{judge-calibration-BkXMIWA4.js.map → judge-calibration-DGqlfWW8.js.map} +1 -1
- package/dist/{judge-calibration-DFtEMlde.d.ts → judge-calibration-DnrbQaXI.d.ts} +1 -3
- package/dist/judge-calibration-DnrbQaXI.d.ts.map +1 -0
- package/dist/{judge-gate-BVvmP00P.js → judge-gate-C2BCOt-M.js} +8 -12
- package/dist/{judge-gate-BVvmP00P.js.map → judge-gate-C2BCOt-M.js.map} +1 -1
- package/dist/judges-BLxg5p8E.js +20 -0
- package/dist/judges-BLxg5p8E.js.map +1 -0
- package/dist/{kind-factory-Dl0PmNQL.js → kind-factory-VKysp1ds.js} +23 -39
- package/dist/{kind-factory-Dl0PmNQL.js.map → kind-factory-VKysp1ds.js.map} +1 -1
- package/dist/ledger-core/index.d.ts +2 -2
- package/dist/ledger-core/index.js +3 -3
- package/dist/ledger-core/lock-holder.d.ts +1 -1
- package/dist/ledger-core/lock-holder.js +1 -2
- package/dist/{ledger-core-CAxUdDk0.js → ledger-core-CfT-FzPf.js} +5 -6
- package/dist/{ledger-core-CAxUdDk0.js.map → ledger-core-CfT-FzPf.js.map} +1 -1
- package/dist/{llm-client-b1ysZVIU.js → llm-client-kJ9EOqjt.js} +22 -30
- package/dist/{llm-client-b1ysZVIU.js.map → llm-client-kJ9EOqjt.js.map} +1 -1
- package/dist/{llm-judge-Cybq29bL.js → llm-judge-cgbhTKQp.js} +108 -164
- package/dist/llm-judge-cgbhTKQp.js.map +1 -0
- package/dist/{lock-holder-DaJbfpci.d.ts → lock-holder-DoxlZFkQ.d.ts} +1 -3
- package/dist/lock-holder-DoxlZFkQ.d.ts.map +1 -0
- package/dist/matrix/index.d.ts +3 -3
- package/dist/matrix/index.js +2 -2
- package/dist/{matrix-DtkZLD_3.d.ts → matrix-CBUzIZs4.d.ts} +4 -12
- package/dist/matrix-CBUzIZs4.d.ts.map +1 -0
- package/dist/{matrix-B8l-Nsxd.js → matrix-CgpX8wdm.js} +8 -13
- package/dist/matrix-CgpX8wdm.js.map +1 -0
- package/dist/{mean-BbA_JCzW.js → mean-CdDYG-us.js} +6 -9
- package/dist/{mean-BbA_JCzW.js.map → mean-CdDYG-us.js.map} +1 -1
- package/dist/meta-eval/index.d.ts +3 -3
- package/dist/meta-eval/index.js +17 -21
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/{metrics-BKaPBtIG.js → metrics-DNCAUtuB.js} +2 -3
- package/dist/{metrics-BKaPBtIG.js.map → metrics-DNCAUtuB.js.map} +1 -1
- package/dist/{mint-CbdwjTPW.js → mint-D9ZysYtY.js} +7 -8
- package/dist/{mint-CbdwjTPW.js.map → mint-D9ZysYtY.js.map} +1 -1
- package/dist/{multi-layer-verifier-DEgMUk6v.d.ts → multi-layer-verifier-lM4aVN78.d.ts} +2 -4
- package/dist/multi-layer-verifier-lM4aVN78.d.ts.map +1 -0
- package/dist/{multiplicity-Sg0U7h8j.js → multiplicity-DAd4xaDC.js} +3 -4
- package/dist/{multiplicity-Sg0U7h8j.js.map → multiplicity-DAd4xaDC.js.map} +1 -1
- package/dist/{multiplicity-DIWHvysC.d.ts → multiplicity-D_SEsc5j.d.ts} +1 -3
- package/dist/multiplicity-D_SEsc5j.d.ts.map +1 -0
- package/dist/multishot/golden/index.d.ts +1 -17
- package/dist/multishot/golden/index.d.ts.map +1 -1
- package/dist/multishot/golden/index.js +10 -17
- package/dist/multishot/golden/index.js.map +1 -1
- package/dist/multishot/index.d.ts +2 -10
- package/dist/multishot/index.d.ts.map +1 -1
- package/dist/multishot/index.js +11 -18
- package/dist/multishot/index.js.map +1 -1
- package/dist/{nonnegative-number-DvdMxg_f.js → nonnegative-number-lEYbOsuh.js} +2 -3
- package/dist/{nonnegative-number-DvdMxg_f.js.map → nonnegative-number-lEYbOsuh.js.map} +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{otel-spans-64hmQqON.js → otel-spans-BsGzTWIK.js} +5 -6
- package/dist/{otel-spans-64hmQqON.js.map → otel-spans-BsGzTWIK.js.map} +1 -1
- package/dist/{otlp-attributes-D9ptWaIJ.js → otlp-attributes-Boeb7gtH.js} +2 -3
- package/dist/{otlp-attributes-D9ptWaIJ.js.map → otlp-attributes-Boeb7gtH.js.map} +1 -1
- package/dist/{outcome-observations-FLmRpgYa.js → outcome-observations-D4snOtUb.js} +4 -5
- package/dist/{outcome-observations-FLmRpgYa.js.map → outcome-observations-D4snOtUb.js.map} +1 -1
- package/dist/{outcome-store-DUcMj-3D.js → outcome-store-TarrcfHG.js} +2 -3
- package/dist/{outcome-store-DUcMj-3D.js.map → outcome-store-TarrcfHG.js.map} +1 -1
- package/dist/{package-version-D7lQHt_-.js → package-version-CJl_pY9W.js} +2 -3
- package/dist/{package-version-D7lQHt_-.js.map → package-version-CJl_pY9W.js.map} +1 -1
- package/dist/{paired-arms-B2ttLc0_.js → paired-arms-D-54hO-k.js} +10 -14
- package/dist/{paired-arms-B2ttLc0_.js.map → paired-arms-D-54hO-k.js.map} +1 -1
- package/dist/{paired-binary-B_nejA44.js → paired-binary-BR0nwHOA.js} +3 -4
- package/dist/{paired-binary-B_nejA44.js.map → paired-binary-BR0nwHOA.js.map} +1 -1
- package/dist/{paired-promotion-decision-CQMJ5xGN.d.ts → paired-promotion-decision-BDnqepZ1.d.ts} +1 -5
- package/dist/paired-promotion-decision-BDnqepZ1.d.ts.map +1 -0
- package/dist/{paired-promotion-decision-D0SHb-ZL.js → paired-promotion-decision-FGvHA4p_.js} +5 -7
- package/dist/{paired-promotion-decision-D0SHb-ZL.js.map → paired-promotion-decision-FGvHA4p_.js.map} +1 -1
- package/dist/{pareto-BiBnQXAE.js → pareto-B9udAt2i.js} +28 -50
- package/dist/pareto-B9udAt2i.js.map +1 -0
- package/dist/pipelines/index.d.ts +5 -11
- package/dist/pipelines/index.d.ts.map +1 -1
- package/dist/pipelines/index.js +6 -8
- package/dist/pipelines/index.js.map +1 -1
- package/dist/{power-and-mde-CQX-0weG.js → power-and-mde-ClI6RWw4.js} +4 -5
- package/dist/{power-and-mde-CQX-0weG.js.map → power-and-mde-ClI6RWw4.js.map} +1 -1
- package/dist/{predicate-FdKmElA_.js → predicate-BYHA4MBn.js} +2 -3
- package/dist/{predicate-FdKmElA_.js.map → predicate-BYHA4MBn.js.map} +1 -1
- package/dist/{prime-bridge-transport-6feEglLf.d.ts → prime-bridge-transport-Cfqir_Im.d.ts} +1 -3
- package/dist/prime-bridge-transport-Cfqir_Im.d.ts.map +1 -0
- package/dist/{prime-bridge-transport-Z3cpKGoV.js → prime-bridge-transport-Csd5SQg7.js} +2 -3
- package/dist/{prime-bridge-transport-Z3cpKGoV.js.map → prime-bridge-transport-Csd5SQg7.js.map} +1 -1
- package/dist/{prime-protocol-BGxjfhM4.d.ts → prime-protocol-COFkxas6.d.ts} +4 -8
- package/dist/prime-protocol-COFkxas6.d.ts.map +1 -0
- package/dist/{prime-protocol-BFgGrEMP.js → prime-protocol-CrhpXRz9.js} +4 -6
- package/dist/{prime-protocol-BFgGrEMP.js.map → prime-protocol-CrhpXRz9.js.map} +1 -1
- package/dist/profile-cell.d.ts +1 -1
- package/dist/profile-cell.js +1 -2
- package/dist/{promotion-gate-CkVZHagC.d.ts → promotion-gate-CixshGI8.d.ts} +5 -11
- package/dist/promotion-gate-CixshGI8.d.ts.map +1 -0
- package/dist/{promotion-policy-6A8osGr_.js → promotion-policy-CBYeD4zi.js} +5 -6
- package/dist/{promotion-policy-6A8osGr_.js.map → promotion-policy-CBYeD4zi.js.map} +1 -1
- package/dist/{proposal-findings-DqLUWsAG.js → proposal-findings-WFQrxnfe.js} +3 -4
- package/dist/{proposal-findings-DqLUWsAG.js.map → proposal-findings-WFQrxnfe.js.map} +1 -1
- package/dist/{quantile--0yb07pD.js → quantile-CncXP-Ss.js} +2 -3
- package/dist/{quantile--0yb07pD.js.map → quantile-CncXP-Ss.js.map} +1 -1
- package/dist/{query-B_RZ36an.js → query-CGXzO9d4.js} +5 -6
- package/dist/{query-B_RZ36an.js.map → query-CGXzO9d4.js.map} +1 -1
- package/dist/{query-Cmv_U8vf.d.ts → query-UD0B08Sb.d.ts} +4 -6
- package/dist/query-UD0B08Sb.d.ts.map +1 -0
- package/dist/{random-DGEyr3zG.d.ts → random-BN150Q_X.d.ts} +1 -5
- package/dist/random-BN150Q_X.d.ts.map +1 -0
- package/dist/{random-DDDcC7YU.js → random-DrzcgC7F.js} +3 -4
- package/dist/{random-DDDcC7YU.js.map → random-DrzcgC7F.js.map} +1 -1
- package/dist/{raw-provider-sink-75w4FWBj.js → raw-provider-sink-DDLkP_eF.js} +3 -4
- package/dist/{raw-provider-sink-75w4FWBj.js.map → raw-provider-sink-DDLkP_eF.js.map} +1 -1
- package/dist/{record-id-DUgsK5qp.js → record-id-CWZXFh58.js} +2 -3
- package/dist/{record-id-DUgsK5qp.js.map → record-id-CWZXFh58.js.map} +1 -1
- package/dist/{redact-DXdlFb_U.d.ts → redact-C_wzC4c_.d.ts} +1 -3
- package/dist/redact-C_wzC4c_.d.ts.map +1 -0
- package/dist/{redact-mtkTiJVk.js → redact-D6APJS1d.js} +3 -5
- package/dist/{redact-mtkTiJVk.js.map → redact-D6APJS1d.js.map} +1 -1
- package/dist/{registry-BcdJmyoU.d.ts → registry-dhdDoAV8.d.ts} +5 -11
- package/dist/registry-dhdDoAV8.d.ts.map +1 -0
- package/dist/{release-confidence-CYIFH-fp.js → release-confidence-CTUJz5mM.js} +10 -16
- package/dist/release-confidence-CTUJz5mM.js.map +1 -0
- package/dist/{render-BhJqXoBH.js → render-pfAZ7IR4.js} +266 -33
- package/dist/render-pfAZ7IR4.js.map +1 -0
- package/dist/{reporting-html-Kvlc8oMt.js → reporting-html-DMXIjATy.js} +2 -3
- package/dist/{reporting-html-Kvlc8oMt.js.map → reporting-html-DMXIjATy.js.map} +1 -1
- package/dist/reporting.d.ts +6 -6
- package/dist/reporting.js +7 -7
- package/dist/{researcher-DKWtoMWk.d.ts → researcher-CLnNJ7Or.d.ts} +8 -12
- package/dist/researcher-CLnNJ7Or.d.ts.map +1 -0
- package/dist/{reward-nw2xZGZG.js → reward-P0VD_Aft.js} +2 -3
- package/dist/{reward-nw2xZGZG.js.map → reward-P0VD_Aft.js.map} +1 -1
- package/dist/{reward-BeUgVVC_.d.ts → reward-ST4PXNXF.d.ts} +2 -4
- package/dist/reward-ST4PXNXF.d.ts.map +1 -0
- package/dist/{reward-hacking-WTxUbtbb.d.ts → reward-hacking-C9hPVT0Q.d.ts} +4 -8
- package/dist/reward-hacking-C9hPVT0Q.d.ts.map +1 -0
- package/dist/{reward-hacking-ByuETaIb.js → reward-hacking-DZzHZlS1.js} +5 -7
- package/dist/{reward-hacking-ByuETaIb.js.map → reward-hacking-DZzHZlS1.js.map} +1 -1
- package/dist/rl.d.ts +26 -63
- package/dist/rl.d.ts.map +1 -1
- package/dist/rl.js +35 -52
- package/dist/rl.js.map +1 -1
- package/dist/{rolldown-runtime-8H4AJuhK.js → rolldown-runtime-BJohB4LM.js} +1 -2
- package/dist/rollout/index.d.ts +6 -6
- package/dist/rollout/index.js +7 -7
- package/dist/{rollout-Dd1ucT_M.js → rollout-BhkMpAJT.js} +279 -16
- package/dist/rollout-BhkMpAJT.js.map +1 -0
- package/dist/{terminal-record-CB9BFtjS.js → rollout-nodes-BHR9Q5jP.js} +351 -11
- package/dist/rollout-nodes-BHR9Q5jP.js.map +1 -0
- package/dist/{round-decimal-COrKglMh.js → round-decimal-ZmjAQPo1.js} +2 -3
- package/dist/{round-decimal-COrKglMh.js.map → round-decimal-ZmjAQPo1.js.map} +1 -1
- package/dist/{rubric-predictive-validity-Cd5SCtAf.d.ts → rubric-predictive-validity-B3RzbwOx.d.ts} +2 -8
- package/dist/rubric-predictive-validity-B3RzbwOx.d.ts.map +1 -0
- package/dist/{rubric-predictive-validity-CeKW2GQN.js → rubric-predictive-validity-CQp_M0vy.js} +3 -4
- package/dist/{rubric-predictive-validity-CeKW2GQN.js.map → rubric-predictive-validity-CQp_M0vy.js.map} +1 -1
- package/dist/{run-record-CysKLB93.js → run-record-CZy4-TnZ.js} +4 -5
- package/dist/{run-record-CysKLB93.js.map → run-record-CZy4-TnZ.js.map} +1 -1
- package/dist/{run-record-C3awiXPf.d.ts → run-record-D2BFMPGl.d.ts} +5 -7
- package/dist/run-record-D2BFMPGl.d.ts.map +1 -0
- package/dist/{run-record-DhEJgoPz.js → run-record-DEjx43cI.js} +6 -7
- package/dist/{run-record-DhEJgoPz.js.map → run-record-DEjx43cI.js.map} +1 -1
- package/dist/{sandbox-harness-BlSOu4LX.d.ts → sandbox-harness-B3Kx1fOA.d.ts} +1 -3
- package/dist/sandbox-harness-B3Kx1fOA.d.ts.map +1 -0
- package/dist/{schema-D6EVKzTB.d.ts → schema-CkIvnmLU.d.ts} +2 -43
- package/dist/schema-CkIvnmLU.d.ts.map +1 -0
- package/dist/{schema-Bt2vLFm3.js → schema-D-xu3MU_.js} +2 -3
- package/dist/{schema-Bt2vLFm3.js.map → schema-D-xu3MU_.js.map} +1 -1
- package/dist/{schema-C-ziyJD7.d.ts → schema-DkFJr3Y8.d.ts} +111 -48
- package/dist/schema-DkFJr3Y8.d.ts.map +1 -0
- package/dist/{schema-CvhfLjtZ.js → schema-YUIXT8iW.js} +37 -7
- package/dist/schema-YUIXT8iW.js.map +1 -0
- package/dist/search/index.d.ts +3 -35
- package/dist/search/index.d.ts.map +1 -1
- package/dist/search/index.js +4 -5
- package/dist/{search-kernel-CctpiDd2.d.ts → search-kernel-CBdwFZ05.d.ts} +10 -36
- package/dist/search-kernel-CBdwFZ05.d.ts.map +1 -0
- package/dist/{search-ledger-CZRqtFbW.js → search-ledger-BsVi-Ngv.js} +17 -26
- package/dist/{search-ledger-CZRqtFbW.js.map → search-ledger-BsVi-Ngv.js.map} +1 -1
- package/dist/{search-ledger-file-Bd5Tioq6.js → search-ledger-file-DTBD6X-y.js} +6 -8
- package/dist/{search-ledger-file-Bd5Tioq6.js.map → search-ledger-file-DTBD6X-y.js.map} +1 -1
- package/dist/{search-ledger-types-p2xKfiDJ.d.ts → search-ledger-types-qiYRUo_N.d.ts} +5 -9
- package/dist/search-ledger-types-qiYRUo_N.d.ts.map +1 -0
- package/dist/{search-shipper-Dn3NLsuH.js → search-shipper-pdg4viK7.js} +8 -13
- package/dist/{search-shipper-Dn3NLsuH.js.map → search-shipper-pdg4viK7.js.map} +1 -1
- package/dist/{search-summary-D4bwYnAE.js → search-summary-Ct7SrmVc.js} +6 -7
- package/dist/{search-summary-D4bwYnAE.js.map → search-summary-Ct7SrmVc.js.map} +1 -1
- package/dist/{sequential-CimQrbjv.js → sequential-B1I0tRfL.js} +6 -7
- package/dist/{sequential-CimQrbjv.js.map → sequential-B1I0tRfL.js.map} +1 -1
- package/dist/{sequential-KwGspfeR.d.ts → sequential-Bb_pdTqG.d.ts} +1 -3
- package/dist/sequential-Bb_pdTqG.d.ts.map +1 -0
- package/dist/{sequential-BVv1sjsw.js → sequential-Bjs5roiK.js} +2 -3
- package/dist/{sequential-BVv1sjsw.js.map → sequential-Bjs5roiK.js.map} +1 -1
- package/dist/{sequential-eprocess-DodI6RoF.js → sequential-eprocess-C1YLwFaz.js} +3 -4
- package/dist/{sequential-eprocess-DodI6RoF.js.map → sequential-eprocess-C1YLwFaz.js.map} +1 -1
- package/dist/{server-ChhXPCgX.js → server-DA6h3co5.js} +11 -17
- package/dist/{server-ChhXPCgX.js.map → server-DA6h3co5.js.map} +1 -1
- package/dist/{single-run-lock-Dekd8_Wv.js → single-run-lock-BGlW3OSk.js} +3 -4
- package/dist/{single-run-lock-Dekd8_Wv.js.map → single-run-lock-BGlW3OSk.js.map} +1 -1
- package/dist/{skillopt-optimization-method-sB_-yi1G.js → skillopt-optimization-method-qB1M554K.js} +38 -57
- package/dist/{skillopt-optimization-method-sB_-yi1G.js.map → skillopt-optimization-method-qB1M554K.js.map} +1 -1
- package/dist/{spans-CVngGij3.d.ts → spans-vZoW__3q.d.ts} +2 -4
- package/dist/spans-vZoW__3q.d.ts.map +1 -0
- package/dist/{statistical-heldout-BjzSOmLH.d.ts → statistical-heldout-Cei036nE.d.ts} +9 -37
- package/dist/statistical-heldout-Cei036nE.d.ts.map +1 -0
- package/dist/{statistical-heldout-Bksv92Th.js → statistical-heldout-TP6ndL0B.js} +4 -5
- package/dist/{statistical-heldout-Bksv92Th.js.map → statistical-heldout-TP6ndL0B.js.map} +1 -1
- package/dist/{steps-DlepJPwD.d.ts → steps-B8fYY_Vo.d.ts} +1 -3
- package/dist/steps-B8fYY_Vo.d.ts.map +1 -0
- package/dist/{store-B3nWRy3B.d.ts → store-CMs5wGXP.d.ts} +2 -4
- package/dist/store-CMs5wGXP.d.ts.map +1 -0
- package/dist/{store-BhNuLdJw.js → store-ZykIyQmy.js} +2 -3
- package/dist/{store-BhNuLdJw.js.map → store-ZykIyQmy.js.map} +1 -1
- package/dist/{store-otlp-CFf27E8_.js → store-otlp-DmLm65XW.js} +8 -13
- package/dist/{store-otlp-CFf27E8_.js.map → store-otlp-DmLm65XW.js.map} +1 -1
- package/dist/{store-tool-spans-DxsFwtdr.d.ts → store-tool-spans-B3i3fl0A.d.ts} +11 -29
- package/dist/store-tool-spans-B3i3fl0A.d.ts.map +1 -0
- package/dist/{store-tool-spans-BK_8FjGG.js → store-tool-spans-C2g8UYm-.js} +17 -41
- package/dist/{store-tool-spans-BK_8FjGG.js.map → store-tool-spans-C2g8UYm-.js.map} +1 -1
- package/dist/storyboard/index.d.ts +1 -5
- package/dist/storyboard/index.d.ts.map +1 -1
- package/dist/storyboard/index.js +3 -5
- package/dist/storyboard/index.js.map +1 -1
- package/dist/{summary-report-C8V49Jt3.d.ts → summary-report-0La5vejY.d.ts} +5 -13
- package/dist/summary-report-0La5vejY.d.ts.map +1 -0
- package/dist/{summary-report-DHDWxe6p.js → summary-report-CPfsrLac.js} +21 -30
- package/dist/summary-report-CPfsrLac.js.map +1 -0
- package/dist/supervisor-run/index.d.ts +30 -29
- package/dist/supervisor-run/index.d.ts.map +1 -1
- package/dist/supervisor-run/index.js +6 -7
- package/dist/{systemone-CHnqwmYg.d.ts → systemone-Cw2txK5m.d.ts} +12 -24
- package/dist/systemone-Cw2txK5m.d.ts.map +1 -0
- package/dist/{systemone-protocol-Bl2EI2KQ.d.ts → systemone-protocol-C4t7JH4Y.d.ts} +1 -3
- package/dist/systemone-protocol-C4t7JH4Y.d.ts.map +1 -0
- package/dist/systemone-protocol.d.ts +1 -1
- package/dist/systemone-protocol.js +2 -3
- package/dist/{systemone-review-cQ9acAsb.d.ts → systemone-review-BLgxt-4Q.d.ts} +8 -10
- package/dist/systemone-review-BLgxt-4Q.d.ts.map +1 -0
- package/dist/{systemone-review-DFGtIUcf.js → systemone-review-S49gsy_y.js} +5 -6
- package/dist/{systemone-review-DFGtIUcf.js.map → systemone-review-S49gsy_y.js.map} +1 -1
- package/dist/systemone.d.ts +4 -4
- package/dist/systemone.js +3 -3
- package/dist/{task-failure-attributes-oIK3BhCR.js → task-failure-attributes-CC6VuIpk.js} +7 -10
- package/dist/{task-failure-attributes-oIK3BhCR.js.map → task-failure-attributes-CC6VuIpk.js.map} +1 -1
- package/dist/{tool-groups-b3oU_D88.d.ts → tool-groups-C3K-d3gd.d.ts} +3 -5
- package/dist/tool-groups-C3K-d3gd.d.ts.map +1 -0
- package/dist/{tool-waste-CkjzWmeY.d.ts → tool-waste-BFG1vwTV.d.ts} +4 -16
- package/dist/tool-waste-BFG1vwTV.d.ts.map +1 -0
- package/dist/{tool-waste-DLsjS_p-.js → tool-waste-CLJ46AbT.js} +10 -26
- package/dist/{tool-waste-DLsjS_p-.js.map → tool-waste-CLJ46AbT.js.map} +1 -1
- package/dist/trace-attributes.d.ts +1 -1
- package/dist/trace-attributes.js +1 -2
- package/dist/{trace-questions-BmT8M6UQ.d.ts → trace-questions-ej5AEJ5X.d.ts} +4 -6
- package/dist/trace-questions-ej5AEJ5X.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +46 -92
- package/dist/trace-repair/index.d.ts.map +1 -1
- package/dist/trace-repair/index.js +32 -54
- package/dist/trace-repair/index.js.map +1 -1
- package/dist/traces.d.ts +13 -29
- package/dist/traces.d.ts.map +1 -1
- package/dist/traces.js +18 -22
- package/dist/traces.js.map +1 -1
- package/dist/{trajectory-D_7rLrvE.js → trajectory-O3GDHmev.js} +2 -3
- package/dist/{trajectory-D_7rLrvE.js.map → trajectory-O3GDHmev.js.map} +1 -1
- package/dist/trajectory-replay/index.d.ts +2 -22
- package/dist/trajectory-replay/index.d.ts.map +1 -1
- package/dist/trajectory-replay/index.js +19 -30
- package/dist/trajectory-replay/index.js.map +1 -1
- package/dist/{tree-DSqwJpL9.js → tree-DrIChF4a.js} +39 -68
- package/dist/{tree-DSqwJpL9.js.map → tree-DrIChF4a.js.map} +1 -1
- package/dist/{types-B_IFcKp5.d.ts → types-0KglnQ9o.d.ts} +3 -15
- package/dist/types-0KglnQ9o.d.ts.map +1 -0
- package/dist/{types-DZ0mYVw9.d.ts → types-0oMg7zee.d.ts} +3 -5
- package/dist/types-0oMg7zee.d.ts.map +1 -0
- package/dist/{types-CnWmj8NW.js → types-BK0blWhe.js} +3 -4
- package/dist/{types-CnWmj8NW.js.map → types-BK0blWhe.js.map} +1 -1
- package/dist/{types-CmXY0PQf.d.ts → types-C-aOqgt2.d.ts} +2 -5
- package/dist/types-C-aOqgt2.d.ts.map +1 -0
- package/dist/{types-M6eF9m-r.d.ts → types-C54Uk_a1.d.ts} +84 -24
- package/dist/types-C54Uk_a1.d.ts.map +1 -0
- package/dist/{types-BXgASBCL.js → types-DK4vNv5q.js} +2 -3
- package/dist/{types-BXgASBCL.js.map → types-DK4vNv5q.js.map} +1 -1
- package/dist/{types-6Lfdlshp.d.ts → types-DzCO-S1m.d.ts} +50 -62
- package/dist/types-DzCO-S1m.d.ts.map +1 -0
- package/dist/{verdict-BGQmxQ8A.js → verdict-C8TcdwfD.js} +3 -4
- package/dist/{verdict-BGQmxQ8A.js.map → verdict-C8TcdwfD.js.map} +1 -1
- package/dist/{verdict-C24NELtN.d.ts → verdict-D_qiBQ6_.d.ts} +1 -5
- package/dist/verdict-D_qiBQ6_.d.ts.map +1 -0
- package/dist/{verdict-cache-CL_7tePD.js → verdict-cache-BVH-fl18.js} +3 -4
- package/dist/{verdict-cache-CL_7tePD.js.map → verdict-cache-BVH-fl18.js.map} +1 -1
- package/dist/wire/index.d.ts +43 -51
- package/dist/wire/index.d.ts.map +1 -1
- package/dist/wire/index.js +1 -2
- package/docs/eval-surface-map.md +15 -0
- package/docs/rollout.md +132 -6
- package/package.json +18 -18
- package/dist/adversarial-smnADNFS.d.ts.map +0 -1
- package/dist/agent-profile-CNmwfaqq.d.ts.map +0 -1
- package/dist/agent-profile-cell-BcpFbNd5.d.ts.map +0 -1
- package/dist/analyze-runs-DfX6dHP0.d.ts.map +0 -1
- package/dist/attribute-vocabulary-CcXqHODT.d.ts.map +0 -1
- package/dist/benchmark-command-Cnjfo0Zh.js.map +0 -1
- package/dist/campaign-evidence-DunzcJGv.js.map +0 -1
- package/dist/campaign-uddTlFg5.js.map +0 -1
- package/dist/canonical-CFpojCN5.d.ts.map +0 -1
- package/dist/client-hwo3cJjR.d.ts.map +0 -1
- package/dist/completion-verifier-asaV1nSj.js.map +0 -1
- package/dist/cost-ledger-C1W3ICmP.d.ts.map +0 -1
- package/dist/dataset-DQqhOCPt.d.ts.map +0 -1
- package/dist/default-registry-CLG5YqGS.d.ts.map +0 -1
- package/dist/define-agent-eval-JKt3gMQr.d.ts.map +0 -1
- package/dist/descriptive-B2iPaT9J.d.ts +0 -89
- package/dist/descriptive-B2iPaT9J.d.ts.map +0 -1
- package/dist/descriptive-CVNy48rR.js.map +0 -1
- package/dist/emitter-DABYYTcH.d.ts.map +0 -1
- package/dist/engine-dtRVeuQs.d.ts.map +0 -1
- package/dist/errors-DEE6u6ot.d.ts.map +0 -1
- package/dist/evaluation-DblGquCB.d.ts.map +0 -1
- package/dist/exact-types-Do63KoKn.d.ts.map +0 -1
- package/dist/exporters-DWvM16LE.js.map +0 -1
- package/dist/exporters-Dk3oEbK_.d.ts.map +0 -1
- package/dist/failure-cluster-DN0t63Z0.d.ts.map +0 -1
- package/dist/feedback-trajectory-CYVJXolG.d.ts.map +0 -1
- package/dist/hf-dataset-CKxIJQba.js.map +0 -1
- package/dist/index-CHxwfW5g.d.ts.map +0 -1
- package/dist/index-CTIRz-AB.d.ts.map +0 -1
- package/dist/index-CalAsXMR.d.ts.map +0 -1
- package/dist/index-CdEVr8-O.d.ts.map +0 -1
- package/dist/index-CfJULcW5.d.ts.map +0 -1
- package/dist/index-D7fmC-1K.d.ts.map +0 -1
- package/dist/index-jnuFLxOs.d.ts.map +0 -1
- package/dist/integrity-D-s2iHiT.js.map +0 -1
- package/dist/integrity-D9lBnv_K.d.ts.map +0 -1
- package/dist/interval-ohuOJurX.d.ts.map +0 -1
- package/dist/judge-calibration-DFtEMlde.d.ts.map +0 -1
- package/dist/llm-judge-Cybq29bL.js.map +0 -1
- package/dist/lock-holder-DaJbfpci.d.ts.map +0 -1
- package/dist/matrix-B8l-Nsxd.js.map +0 -1
- package/dist/matrix-DtkZLD_3.d.ts.map +0 -1
- package/dist/mint-CkzpVPdz.d.ts +0 -81
- package/dist/mint-CkzpVPdz.d.ts.map +0 -1
- package/dist/multi-layer-verifier-DEgMUk6v.d.ts.map +0 -1
- package/dist/multiplicity-DIWHvysC.d.ts.map +0 -1
- package/dist/paired-promotion-decision-CQMJ5xGN.d.ts.map +0 -1
- package/dist/pareto-BiBnQXAE.js.map +0 -1
- package/dist/prime-bridge-transport-6feEglLf.d.ts.map +0 -1
- package/dist/prime-protocol-BGxjfhM4.d.ts.map +0 -1
- package/dist/promotion-gate-CkVZHagC.d.ts.map +0 -1
- package/dist/query-Cmv_U8vf.d.ts.map +0 -1
- package/dist/random-DGEyr3zG.d.ts.map +0 -1
- package/dist/redact-DXdlFb_U.d.ts.map +0 -1
- package/dist/registry-BcdJmyoU.d.ts.map +0 -1
- package/dist/release-confidence-CYIFH-fp.js.map +0 -1
- package/dist/render-BhJqXoBH.js.map +0 -1
- package/dist/researcher-DKWtoMWk.d.ts.map +0 -1
- package/dist/reward-BeUgVVC_.d.ts.map +0 -1
- package/dist/reward-hacking-WTxUbtbb.d.ts.map +0 -1
- package/dist/rollout-Dd1ucT_M.js.map +0 -1
- package/dist/rubric-predictive-validity-Cd5SCtAf.d.ts.map +0 -1
- package/dist/run-record-C3awiXPf.d.ts.map +0 -1
- package/dist/sandbox-harness-BlSOu4LX.d.ts.map +0 -1
- package/dist/schema-C-ziyJD7.d.ts.map +0 -1
- package/dist/schema-CvhfLjtZ.js.map +0 -1
- package/dist/schema-D6EVKzTB.d.ts.map +0 -1
- package/dist/search-kernel-CctpiDd2.d.ts.map +0 -1
- package/dist/search-ledger-types-p2xKfiDJ.d.ts.map +0 -1
- package/dist/sequential-KwGspfeR.d.ts.map +0 -1
- package/dist/spans-CVngGij3.d.ts.map +0 -1
- package/dist/statistical-heldout-BjzSOmLH.d.ts.map +0 -1
- package/dist/steps-DlepJPwD.d.ts.map +0 -1
- package/dist/store-B3nWRy3B.d.ts.map +0 -1
- package/dist/store-tool-spans-DxsFwtdr.d.ts.map +0 -1
- package/dist/summary-report-C8V49Jt3.d.ts.map +0 -1
- package/dist/summary-report-DHDWxe6p.js.map +0 -1
- package/dist/systemone-CHnqwmYg.d.ts.map +0 -1
- package/dist/systemone-protocol-Bl2EI2KQ.d.ts.map +0 -1
- package/dist/systemone-review-cQ9acAsb.d.ts.map +0 -1
- package/dist/terminal-record-CB9BFtjS.js.map +0 -1
- package/dist/tool-groups-b3oU_D88.d.ts.map +0 -1
- package/dist/tool-waste-CkjzWmeY.d.ts.map +0 -1
- package/dist/trace-questions-BmT8M6UQ.d.ts.map +0 -1
- package/dist/types-6Lfdlshp.d.ts.map +0 -1
- package/dist/types-B_IFcKp5.d.ts.map +0 -1
- package/dist/types-CmXY0PQf.d.ts.map +0 -1
- package/dist/types-DZ0mYVw9.d.ts.map +0 -1
- package/dist/types-M6eF9m-r.d.ts.map +0 -1
- package/dist/verdict-C24NELtN.d.ts.map +0 -1
|
@@ -1,7 +1,9 @@
|
|
|
1
|
-
import { a as MintedRolloutLine, b as RolloutSplit, g as RolloutOutcome, h as RolloutLine, i as GatedEvidence, t as ChatMessage, w as ToolDef } from "./schema-
|
|
2
|
-
import "./
|
|
3
|
-
import "./
|
|
4
|
-
|
|
1
|
+
import { a as MintedRolloutLine, b as RolloutSplit, g as RolloutOutcome, h as RolloutLine, i as GatedEvidence, t as ChatMessage, w as ToolDef } from "./schema-DkFJr3Y8.js";
|
|
2
|
+
import { l as CodeSurface } from "./types-C54Uk_a1.js";
|
|
3
|
+
import "./reward-ST4PXNXF.js";
|
|
4
|
+
import { n as CompletionVerdict } from "./completion-verifier-7KEfroqx.js";
|
|
5
|
+
import { o as TrainingExportOptions } from "./exporters-B5Ze_RKh.js";
|
|
6
|
+
import { HarnessSession, ReadOptions, SessionInputInit, SessionRef } from "@tangle-network/harness-sessions";
|
|
5
7
|
/**
|
|
6
8
|
* Every gate check in the package, in the order they are applied.
|
|
7
9
|
*
|
|
@@ -9,7 +11,7 @@ import "./mint-CkzpVPdz.js";
|
|
|
9
11
|
* trips two checks reports the earlier one, and `reward-relationship` is first
|
|
10
12
|
* because it is the invariant the other three protect.
|
|
11
13
|
*/
|
|
12
|
-
declare const GATE_CHECK_IDS: readonly [
|
|
14
|
+
declare const GATE_CHECK_IDS: readonly ["reward-relationship", "gated-evidence", "undeclared-step-payload", "unscreened-reward"];
|
|
13
15
|
type GateCheckId = (typeof GATE_CHECK_IDS)[number];
|
|
14
16
|
/**
|
|
15
17
|
* An outcome as it reaches a check.
|
|
@@ -117,7 +119,7 @@ declare const GATE_POLICIES: {
|
|
|
117
119
|
*/
|
|
118
120
|
readonly validateRolloutLine: {
|
|
119
121
|
readonly 'reward-relationship': {
|
|
120
|
-
readonly kind:
|
|
122
|
+
readonly kind: "enforce";
|
|
121
123
|
};
|
|
122
124
|
readonly 'gated-evidence': GateCheckDisposition;
|
|
123
125
|
readonly 'undeclared-step-payload': GateCheckDisposition;
|
|
@@ -130,12 +132,12 @@ declare const GATE_POLICIES: {
|
|
|
130
132
|
*/
|
|
131
133
|
readonly assertMinted: {
|
|
132
134
|
readonly 'reward-relationship': {
|
|
133
|
-
readonly kind:
|
|
135
|
+
readonly kind: "enforce";
|
|
134
136
|
};
|
|
135
137
|
readonly 'gated-evidence': GateCheckDisposition;
|
|
136
138
|
readonly 'undeclared-step-payload': GateCheckDisposition;
|
|
137
139
|
readonly 'unscreened-reward': {
|
|
138
|
-
readonly kind:
|
|
140
|
+
readonly kind: "enforce";
|
|
139
141
|
};
|
|
140
142
|
};
|
|
141
143
|
/**
|
|
@@ -146,16 +148,16 @@ declare const GATE_POLICIES: {
|
|
|
146
148
|
*/
|
|
147
149
|
readonly assertRewardGate: {
|
|
148
150
|
readonly 'reward-relationship': {
|
|
149
|
-
readonly kind:
|
|
151
|
+
readonly kind: "enforce";
|
|
150
152
|
};
|
|
151
153
|
readonly 'gated-evidence': {
|
|
152
|
-
readonly kind:
|
|
154
|
+
readonly kind: "enforce";
|
|
153
155
|
};
|
|
154
156
|
readonly 'undeclared-step-payload': {
|
|
155
|
-
readonly kind:
|
|
157
|
+
readonly kind: "enforce";
|
|
156
158
|
};
|
|
157
159
|
readonly 'unscreened-reward': {
|
|
158
|
-
readonly kind:
|
|
160
|
+
readonly kind: "enforce";
|
|
159
161
|
};
|
|
160
162
|
};
|
|
161
163
|
/**
|
|
@@ -166,16 +168,16 @@ declare const GATE_POLICIES: {
|
|
|
166
168
|
*/
|
|
167
169
|
readonly assertGateReport: {
|
|
168
170
|
readonly 'reward-relationship': {
|
|
169
|
-
readonly kind:
|
|
171
|
+
readonly kind: "enforce";
|
|
170
172
|
};
|
|
171
173
|
readonly 'gated-evidence': {
|
|
172
|
-
readonly kind:
|
|
174
|
+
readonly kind: "enforce";
|
|
173
175
|
};
|
|
174
176
|
readonly 'undeclared-step-payload': {
|
|
175
|
-
readonly kind:
|
|
177
|
+
readonly kind: "enforce";
|
|
176
178
|
};
|
|
177
179
|
readonly 'unscreened-reward': {
|
|
178
|
-
readonly kind:
|
|
180
|
+
readonly kind: "enforce";
|
|
179
181
|
};
|
|
180
182
|
};
|
|
181
183
|
};
|
|
@@ -187,8 +189,64 @@ type GateEntryPoint = keyof typeof GATE_POLICIES;
|
|
|
187
189
|
* exists to remove.
|
|
188
190
|
*/
|
|
189
191
|
declare function gateErrors(subject: GateSubject, policy: GatePolicy): string[];
|
|
190
|
-
|
|
191
|
-
|
|
192
|
+
/** Existing completion checks required to admit a verified episode as a runnable task. */
|
|
193
|
+
declare const HARBOR_TASK_CHECKS: readonly ["base-bundle", "acceptance", "production-delta", "dangling-references", "merge-followup"];
|
|
194
|
+
interface HarborTaskFile {
|
|
195
|
+
/** Relative path in the published task. */
|
|
196
|
+
readonly path: string;
|
|
197
|
+
/** Actual staged path; task.toml is not published until verification succeeds. */
|
|
198
|
+
readonly stagedPath: string;
|
|
199
|
+
readonly sha256: string;
|
|
200
|
+
readonly byteLength: number;
|
|
201
|
+
}
|
|
202
|
+
interface HarborTaskVerification {
|
|
203
|
+
/** Digest binding this episode, immutable source identity, merge and every task file. */
|
|
204
|
+
readonly taskId: string;
|
|
205
|
+
readonly rolloutId: string;
|
|
206
|
+
readonly episodeSha256: string;
|
|
207
|
+
readonly task: MintedRolloutLine['task'];
|
|
208
|
+
readonly source: Omit<CodeSurface, 'worktreeRef' | 'summary' | 'baseRef'>;
|
|
209
|
+
readonly mergeCommit: string;
|
|
210
|
+
readonly files: readonly HarborTaskFile[];
|
|
211
|
+
}
|
|
212
|
+
interface WriteHarborTaskOptions {
|
|
213
|
+
/** Fresh task directory, normally <dataset>/tasks/<task>. Existing paths are refused. */
|
|
214
|
+
readonly directory: string;
|
|
215
|
+
readonly episode: MintedRolloutLine;
|
|
216
|
+
readonly surface: CodeSurface;
|
|
217
|
+
readonly mergeCommit: string;
|
|
218
|
+
/** Harbor's org/name task name. */
|
|
219
|
+
readonly name: string;
|
|
220
|
+
readonly instruction: string;
|
|
221
|
+
/** Caller image with Git/dependencies and a configured user able to clone into /workspace. */
|
|
222
|
+
readonly image: string;
|
|
223
|
+
/** Git bundle containing the base and its ancestors only, never the answer/head/merge. */
|
|
224
|
+
readonly baseBundle: Uint8Array;
|
|
225
|
+
readonly commands: {
|
|
226
|
+
readonly acceptance: string;
|
|
227
|
+
readonly productionDelta: string;
|
|
228
|
+
readonly danglingReferences: string;
|
|
229
|
+
};
|
|
230
|
+
readonly agentTimeoutSec: number;
|
|
231
|
+
readonly verifierTimeoutSec: number;
|
|
232
|
+
/**
|
|
233
|
+
* Trusted caller execution boundary. Inspect the actual staged bundle and file hashes.
|
|
234
|
+
* base-bundle proves exact base/tree, base-only closure and no answer material; the
|
|
235
|
+
* three verifier checks prove the supplied commands actually passed on this candidate
|
|
236
|
+
* (strict production-line decrease, zero dangling refs); merge-followup proves this
|
|
237
|
+
* candidate merged with base=merge^1 and a closed, explicitly checked no-fix-forward
|
|
238
|
+
* followup window. Every check must cite receipts bound to taskId, not a title/keyword.
|
|
239
|
+
* Unknown, absent, open followup and provider eligibility are not affirmative evidence.
|
|
240
|
+
*/
|
|
241
|
+
readonly verify: (request: HarborTaskVerification, signal?: AbortSignal) => Promise<CompletionVerdict>;
|
|
242
|
+
readonly signal?: AbortSignal;
|
|
243
|
+
}
|
|
244
|
+
/**
|
|
245
|
+
* Write a Harbor directory from an admitted canonical episode, without executing it.
|
|
246
|
+
* Failed admission leaves diagnostic files but no task.toml, so the directory is not
|
|
247
|
+
* loadable as a task. The caller owns execution, receipt truth and artifact retention.
|
|
248
|
+
*/
|
|
249
|
+
declare function writeHarborTask(options: WriteHarborTaskOptions): Promise<string>;
|
|
192
250
|
type HarborStepSource = 'system' | 'user' | 'agent';
|
|
193
251
|
interface HarborImageSource {
|
|
194
252
|
media_type: string;
|
|
@@ -325,8 +383,6 @@ declare function fromHarborTrajectory(trajectory: HarborTrajectory, options?: Fr
|
|
|
325
383
|
* minting a reward.
|
|
326
384
|
*/
|
|
327
385
|
declare function relabelImportedSplit(lines: readonly RolloutLine[], split: RolloutSplit): RolloutLine[];
|
|
328
|
-
//#endregion
|
|
329
|
-
//#region src/rollout/ledger.d.ts
|
|
330
386
|
/** Replace the ledger file with exactly `lines`. */
|
|
331
387
|
declare function writeRolloutLedger(path: string, lines: RolloutLine[]): Promise<void>;
|
|
332
388
|
/** Append `lines` to the ledger file (created if absent). */
|
|
@@ -342,8 +398,12 @@ declare function appendRolloutLines(path: string, lines: RolloutLine[]): Promise
|
|
|
342
398
|
* line is refused rather than exported.
|
|
343
399
|
*/
|
|
344
400
|
declare function readRolloutLedger(path: string): Promise<MintedRolloutLine[]>;
|
|
345
|
-
|
|
346
|
-
|
|
401
|
+
/**
|
|
402
|
+
* Read under the write policy: imported and supervision rows whose positive
|
|
403
|
+
* rewards no realness screen has seen stay readable, as they were writable.
|
|
404
|
+
* Not training input until a screen mints them.
|
|
405
|
+
*/
|
|
406
|
+
declare function readRolloutJournal(path: string): Promise<RolloutLine[]>;
|
|
347
407
|
declare const DEFAULT_CLAUDE_PROJECTS_DIR: string;
|
|
348
408
|
interface ClaudeTranscriptRef {
|
|
349
409
|
sessionId: string;
|
|
@@ -379,13 +439,41 @@ interface ReadClaudeTranscriptOptions {
|
|
|
379
439
|
declare function readClaudeTranscript(path: string, options?: ReadClaudeTranscriptOptions): Promise<ClaudeTranscript>;
|
|
380
440
|
/** Parse retained JSONL text without another file or store; source gaps remain explicit. */
|
|
381
441
|
declare function parseClaudeTranscript(raw: string, options?: ReadClaudeTranscriptOptions): ClaudeTranscript;
|
|
382
|
-
|
|
383
|
-
|
|
442
|
+
/** Normalized source evidence, with the existing canonical message projection. */
|
|
443
|
+
type CodexRollout = Pick<HarnessSession, 'nativeSessionId' | 'parentNativeSessionId' | 'startedAt' | 'endedAt' | 'usage' | 'ending' | 'integrity' | 'modelSource'> & {
|
|
444
|
+
messages: ChatMessage[];
|
|
445
|
+
/** Null when only turn_context names a model; that is configuration, not a response. */
|
|
446
|
+
model: string | null;
|
|
447
|
+
requestedModel: string | null;
|
|
448
|
+
};
|
|
449
|
+
/** Read one already-resolved native file; strict/recover behavior belongs to the SDK. */
|
|
450
|
+
declare function readCodexRollout(path: string, options?: ReadOptions): Promise<CodexRollout>;
|
|
451
|
+
/** Parse retained text through the same SDK fold, retaining source hashes and gaps. */
|
|
452
|
+
declare function parseCodexRollout(raw: string, source?: SessionInputInit, options?: Pick<ReadOptions, 'corruption'>): CodexRollout;
|
|
453
|
+
declare function gateReceiptRolloutLines(path: string, capturedAt?: string): Promise<{
|
|
454
|
+
lines: RolloutLine[];
|
|
455
|
+
skipped: number;
|
|
456
|
+
}>;
|
|
457
|
+
interface HarnessSessionImportOptions {
|
|
458
|
+
/** Read only this harness (`opencode`, `kimi-code`, `claude-code`, `codex`, …). */
|
|
459
|
+
harness?: string;
|
|
460
|
+
/** Only sessions whose store file changed at or after this time (epoch ms). */
|
|
461
|
+
sinceMs?: number;
|
|
462
|
+
capturedAt?: string;
|
|
463
|
+
}
|
|
464
|
+
/** Session refs under a HOME or store directory, or in one session file / OpenCode store. */
|
|
465
|
+
declare function harnessSessionRefs(target: string, options?: HarnessSessionImportOptions): Promise<SessionRef[]>;
|
|
466
|
+
/** One normalized session as a rollout row. */
|
|
467
|
+
declare function harnessSessionRolloutLine(session: HarnessSession, ref: SessionRef, capturedAt?: string): RolloutLine;
|
|
468
|
+
/** Read each ref in turn; a session its reader refuses is reported, never guessed. */
|
|
469
|
+
declare function harnessSessionRolloutLines(refs: readonly SessionRef[], capturedAt?: string): AsyncGenerator<{
|
|
470
|
+
ref: SessionRef;
|
|
471
|
+
line?: RolloutLine;
|
|
472
|
+
error?: string;
|
|
473
|
+
}>;
|
|
384
474
|
declare const DEFAULT_OPENCODE_DB: string;
|
|
385
475
|
/** One session's canonical messages: each model step, then the results of its tool calls. */
|
|
386
476
|
declare function readOpencodeSessionMessages(sessionId: string, db?: string): Promise<ChatMessage[]>;
|
|
387
|
-
//#endregion
|
|
388
|
-
//#region src/rollout/release/gate-report.d.ts
|
|
389
477
|
/** What a format does with a line the realness gate flagged. */
|
|
390
478
|
type GateDisposition = 'exclude' | 'zero-and-flag';
|
|
391
479
|
/** What the gate accounting reads off an emitted row, per format. */
|
|
@@ -504,8 +592,6 @@ interface GateReport {
|
|
|
504
592
|
* all: it iterates `GATE_CHECK_IDS` under its own declared policy.
|
|
505
593
|
*/
|
|
506
594
|
declare function assertGateReport(report: GateReport): void;
|
|
507
|
-
//#endregion
|
|
508
|
-
//#region src/rollout/release/scrub.d.ts
|
|
509
595
|
interface ScrubRule {
|
|
510
596
|
name: string;
|
|
511
597
|
pattern: RegExp;
|
|
@@ -527,9 +613,14 @@ declare function scrubLines(lines: MintedRolloutLine[]): {
|
|
|
527
613
|
* is safe and counted as zero.
|
|
528
614
|
*/
|
|
529
615
|
declare function defaultRolloutScrubber(text: string): string;
|
|
530
|
-
|
|
531
|
-
|
|
532
|
-
|
|
616
|
+
declare const PRIVACY_RULE_NAMES: readonly string[];
|
|
617
|
+
/**
|
|
618
|
+
* The first privacy rule any string in `value` hits, or `null` when clean.
|
|
619
|
+
* `denylist` holds literal terms (names, addresses, numbers) matched without
|
|
620
|
+
* case; it is supplied at run time and never committed.
|
|
621
|
+
*/
|
|
622
|
+
declare function privacyHit(value: unknown, denylist?: readonly string[]): string | null;
|
|
623
|
+
declare const RELEASE_FORMATS: readonly ["sft", "verifiers", "rft", "raw"];
|
|
533
624
|
type ReleaseFormat = (typeof RELEASE_FORMATS)[number];
|
|
534
625
|
interface DatasetCardInputs {
|
|
535
626
|
/** Scrubbed, release-filtered lines (what actually ships). */
|
|
@@ -542,6 +633,7 @@ interface DatasetCardInputs {
|
|
|
542
633
|
excluded: {
|
|
543
634
|
proposers: number;
|
|
544
635
|
nonTrain: number;
|
|
636
|
+
privacy: number;
|
|
545
637
|
};
|
|
546
638
|
formatCounts: Partial<Record<ReleaseFormat, number>>;
|
|
547
639
|
/**
|
|
@@ -553,12 +645,12 @@ interface DatasetCardInputs {
|
|
|
553
645
|
gate: GateReport;
|
|
554
646
|
}
|
|
555
647
|
declare function buildDatasetCard(inputs: DatasetCardInputs): string;
|
|
556
|
-
//#endregion
|
|
557
|
-
//#region src/rollout/release/hf-dataset.d.ts
|
|
558
648
|
interface BuildOptions {
|
|
559
649
|
out: string;
|
|
560
650
|
formats: ReleaseFormat[];
|
|
561
651
|
includeProposers: boolean;
|
|
652
|
+
/** Literal terms whose presence drops a line (case-insensitive). */
|
|
653
|
+
denylist?: string[];
|
|
562
654
|
}
|
|
563
655
|
interface ScrubReport {
|
|
564
656
|
/** Input ledger path → rule → rewrite count (only shipped lines are scrubbed). */
|
|
@@ -567,7 +659,10 @@ interface ScrubReport {
|
|
|
567
659
|
excluded: {
|
|
568
660
|
proposers: number;
|
|
569
661
|
nonTrain: number;
|
|
662
|
+
privacy: number;
|
|
570
663
|
};
|
|
664
|
+
/** Privacy rule → lines dropped whole on its first hit. */
|
|
665
|
+
dropped: Record<string, number>;
|
|
571
666
|
}
|
|
572
667
|
interface BuildSummary {
|
|
573
668
|
inputs: string[];
|
|
@@ -583,7 +678,59 @@ declare function buildHfDataset(inputs: string[], options: BuildOptions): Promis
|
|
|
583
678
|
interface RolloutReleaseCliArgs extends BuildOptions {
|
|
584
679
|
inputs: string[];
|
|
585
680
|
push: string | null;
|
|
681
|
+
denylistPath: string | null;
|
|
682
|
+
}
|
|
683
|
+
/** JSON wire projection of Tinker's text-only Datum; the consumer constructs SDK objects. */
|
|
684
|
+
interface TinkerDatum {
|
|
685
|
+
model_input: {
|
|
686
|
+
chunks: Array<{
|
|
687
|
+
type: 'encoded_text';
|
|
688
|
+
tokens: number[];
|
|
689
|
+
}>;
|
|
690
|
+
};
|
|
691
|
+
loss_fn_inputs: {
|
|
692
|
+
target_tokens: {
|
|
693
|
+
data: number[];
|
|
694
|
+
dtype: 'int64';
|
|
695
|
+
shape: [number];
|
|
696
|
+
};
|
|
697
|
+
logprobs: {
|
|
698
|
+
data: number[];
|
|
699
|
+
dtype: 'float32';
|
|
700
|
+
shape: [number];
|
|
701
|
+
};
|
|
702
|
+
advantages: {
|
|
703
|
+
data: number[];
|
|
704
|
+
dtype: 'float32';
|
|
705
|
+
shape: [number];
|
|
706
|
+
};
|
|
707
|
+
};
|
|
708
|
+
}
|
|
709
|
+
interface TinkerDatumRow {
|
|
710
|
+
/** Only this object goes to the Tinker SDK. Attribution stays with the caller. */
|
|
711
|
+
datum: TinkerDatum;
|
|
712
|
+
rollout_id: string;
|
|
713
|
+
run_id: string;
|
|
714
|
+
instance_id: string;
|
|
715
|
+
step_index: number;
|
|
716
|
+
reward: number;
|
|
717
|
+
realness_gated: false;
|
|
718
|
+
realness_screened: boolean | null;
|
|
586
719
|
}
|
|
587
|
-
|
|
588
|
-
|
|
589
|
-
|
|
720
|
+
interface TinkerExportOptions extends Pick<TrainingExportOptions, 'allowHeldOutTrainingData'> {
|
|
721
|
+
/** Exact model identity of the captured tokenizer/sampling policy. */
|
|
722
|
+
model: string;
|
|
723
|
+
/** Offline decoder for that model. Never re-tokenize or fabricate missing sampling data. */
|
|
724
|
+
decodeTokens: (tokens: readonly number[]) => string;
|
|
725
|
+
/** Caller owns group baselines; a reward is not silently treated as an advantage. */
|
|
726
|
+
advantageOf: (line: MintedRolloutLine, stepIndex: number) => number;
|
|
727
|
+
}
|
|
728
|
+
/**
|
|
729
|
+
* Project known-reward completed search rollouts, including honest zero/negative rewards.
|
|
730
|
+
* Unknown rewards, gated/lossy/incomplete runs and ineligible splits do not become training data.
|
|
731
|
+
* Missing capture on an otherwise eligible model call refuses the batch, never invents tokens.
|
|
732
|
+
* The caller retains the canonical ledger and owns Tinker sampling/forward_backward/optim_step.
|
|
733
|
+
*/
|
|
734
|
+
declare function toTinkerDatumRows(lines: readonly MintedRolloutLine[], options: TinkerExportOptions): TinkerDatumRow[];
|
|
735
|
+
export { HarborMetrics as $, harnessSessionRolloutLine as A, ReadClaudeTranscriptOptions as B, GateReport as C, GateEntryPoint as Ct, readOpencodeSessionMessages as D, DEFAULT_OPENCODE_DB as E, gatedEvidenceOf as Et, readCodexRollout as F, readRolloutJournal as G, parseClaudeTranscript as H, ClaudeTranscript as I, FromHarborOptions as J, readRolloutLedger as K, ClaudeTranscriptRef as L, gateReceiptRolloutLines as M, CodexRollout as N, HarnessSessionImportOptions as O, parseCodexRollout as P, HarborImageSource as Q, ClaudeUsageTotals as R, GateDisposition as S, GateCheckedOutcome as St, assertGateReport as T, gateErrors as Tt, readClaudeTranscript as U, findClaudeTranscripts as V, appendRolloutLines as W, HarborContentPart as X, HarborAgent as Y, HarborFinalMetrics as Z, privacyHit as _, GATE_CHECK_IDS as _t, BuildOptions as a, HarborToolCall as at, EmittedEvidence as b, GateCheckDisposition as bt, ScrubReport as c, relabelImportedSplit as ct, ReleaseFormat as d, HARBOR_TASK_CHECKS as dt, HarborObservation as et, buildDatasetCard as f, HarborTaskFile as ft, defaultRolloutScrubber as g, GATE_CHECKS as gt, ScrubRule as h, writeHarborTask as ht, toTinkerDatumRows as i, HarborSubagentTrajectoryRef as it, harnessSessionRolloutLines as j, harnessSessionRefs as k, buildHfDataset as l, toHarborTrajectories as lt, ScrubCounts as m, WriteHarborTaskOptions as mt, TinkerDatumRow as n, HarborStep as nt, BuildSummary as o, HarborTrajectory as ot, PRIVACY_RULE_NAMES as p, HarborTaskVerification as pt, writeRolloutLedger as q, TinkerExportOptions as r, HarborStepSource as rt, RolloutReleaseCliArgs as s, fromHarborTrajectory as st, TinkerDatum as t, HarborObservationResult as tt, DatasetCardInputs as u, toHarborTrajectory as ut, scrubLines as v, GATE_POLICIES as vt, ReleaseRowRef as w, GatePolicy as wt, FormatGateCounts as x, GateCheckId as xt, scrubText as y, GateCheck as yt, DEFAULT_CLAUDE_PROJECTS_DIR as z };
|
|
736
|
+
//# sourceMappingURL=index-DUn0GNiZ.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-DUn0GNiZ.d.ts","names":[],"sources":["../src/rollout/gate-checks.ts","../src/rollout/harbor-task.ts","../src/rollout/interchange/harbor.ts","../src/rollout/ledger.ts","../src/rollout/readers/claude-jsonl.ts","../src/rollout/readers/codex-rollout.ts","../src/rollout/readers/gate-receipts.ts","../src/rollout/readers/harness-session-lines.ts","../src/rollout/readers/opencode-sqlite.ts","../src/rollout/release/gate-report.ts","../src/rollout/release/scrub.ts","../src/rollout/release/card.ts","../src/rollout/release/hf-dataset.ts","../src/rollout/tinker.ts"],"mappings":";;;;;;;;;;;;;cA2Ca;KAOD,sBAAsB;;;;;;;;;KAUtB,qBAAqB,QAAQ,kBAAkB,SAAS;;;;;;;;;;;;;;;;;UAkB1D;EACR,SAAS;;EAET;;UAOe;EACf,IAAI;;EAEJ;;EAEA,SAAS,SAAS;;;;;;;;;;;;EAYlB,WAAW;;;;;;;;;;;;;iBA6LG,gBAAgB,SAAS,cAAc;;;;;;cA6I1C,yBAAyB,KAAK,cAAc;;;;;;;;;KAe7C;WACG;;;;WAEA;WAAyB;;;;WAEzB;WAAuB;;;KAU1B,yBAAyB,KAAK,cAAc;;;;;;;;;cAU3C;;;;;;;;eAxBQ;;;;;;;;;;;;;eAAA;;;;;eAAA;;;;;;;;;;;eAAA;;;eAAA;;;eAAA;;;eAAA;;;;;;;;;;;eAAA;;;eAAA;;;eAAA;;;eAAA;;;;;KA+FT,8BAA8B;;;;;;iBAO1B,WAAW,SAAS,aAAa,QAAQ;;cCphB5C;UAQI;;WAEN;;WAEA;WACA;WACA;;UAGM;;WAEN;WACA;WACA;WACA,MAAM;WACN,QAAQ,KAAK;WACb;WACA,gBAAgB;;UAGV;;WAEN;WACA,SAAS;WACT,SAAS;WACT;;WAEA;WACA;;WAEA;;WAEA,YAAY;WACZ;aACE;aACA;aACA;;WAEF;WACA;;;;;;;;;;WAUA,SACP,SAAS,wBACT,SAAS,gBACN,QAAQ;WACJ,SAAS;;;;;;;iBAgDE,gBAAgB,SAAS,yBAAyB;KCrB5D;UAEK;EACf;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;EACf;EACA;;EAEA,WAAW;EACX,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,QAAQ;;UAGO;EACf;EACA,mBAAmB;EACnB,0BAA0B;EAC1B,QAAQ;;UAGO;EACf,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;;EAEf;EACA;EACA,QAAQ;EACR;EACA;EACA,kBAAkB;EAClB;EACA,aAAa;EACb,cAAc;EACd,UAAU;EACV;EACA;EACA,QAAQ;;UAGO;EACf;EACA;EACA;;EAEA,mBAAmB;EACnB,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;;UAGO;EACf;EACA;;EAEA;EACA,OAAO;EACP,OAAO;EACP;EACA,gBAAgB;EAChB;EACA,wBAAwB;EACxB,QAAQ;;;;;;;;;;;;;;;;;;iBAgXM,mBAAmB,OAAO,gBAAgB;;iBAa1C,qBAAqB,OAAO,gBAAgB;UA2O3C;;EAEf,YAAY;;;;;;;;;;;iBAwGE,qBACd,YAAY,kBACZ,UAAS,oBACR;;;;;;;;;;;;;;;iBAiCa,qBACd,gBAAgB,eAChB,OAAO,eACN;;iBC76BmB,mBAAmB,cAAc,OAAO,gBAAgB;;iBAOxD,mBAAmB,cAAc,OAAO,gBAAgB;;;;;;;;;;;iBAiBxD,kBAAkB,eAAe,QAAQ;;;;;;iBASzC,mBAAmB,eAAe,QAAQ;cCnCnD;UAOI;EACf;EACA;;;iBAIoB,sBACpB,aACA,uBACC,QAAQ;UASM;EACf;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV,OAAO;;EAEP;EACA;;EAEA;;EAEA;;UAGe;;;;;WAKN;;;iBAqDW,qBACpB,cACA,UAAS,8BACR,QAAQ;;iBAQK,sBACd,aACA,UAAS,8BACR;;KC3HS,eAAe,KACzB;EAUA,UAAU;;EAEV;EACA;;;iBA2BoB,iBACpB,cACA,UAAS,cACR,QAAQ;;iBAKK,kBACd,aACA,SAAQ,kBACR,UAAS,KAAK,6BACb;iBCxDmB,wBACpB,cACA,sBACC;EAAU,OAAO;EAAe;;UCSlB;;EAEf;;EAEA;EACA;;;iBAWoB,mBACpB,gBACA,UAAS,8BACR,QAAQ;;iBAwCK,0BACd,SAAS,gBACT,KAAK,YACL,sBACC;;iBA4EoB,2BACrB,eAAe,cACf,sBACC;EAAiB,KAAK;EAAY,OAAO;EAAa;;cCtJ5C;;iBAeS,4BACpB,mBACA,cACC,QAAQ;;KCoBC;;UAUK;EACf;EACA;;;;;;;;;;;;EAYA;;;;;;;;EAQA;;;;;;;;;;;;EAYA;;;UAIe;;EAEf;EACA;;UAGe;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;;;;;;EAcA,oBAAoB;;;;;;;;;;;;EAYpB;;EAEA;;;;;;;;;;EAUA,wBAAwB;;UAGT;;EAEf;EACA,UAAU,QAAQ,OAAO,eAAe;;;;;;;;;;;;;;;;;iBA6K1B,iBAAiB,QAAQ;UCrUxB;EACf;EACA,SAAS;;EAET,UAAU,eAAe;;;KAqEf,cAAc;iBAWV,UAAU,cAAc,QAAQ;iBAkChC,WAAW,OAAO;EAChC,OAAO;EACP,QAAQ;;;;;;;;;iBAaM,uBAAuB;cAsF1B;;;;;;iBAkBG,WAAW,gBAAgB;cC7O9B;KACD,wBAAwB;UAkBnB;;EAEf,OAAO;EACP,SAAS;EACT;;EAEA;EACA,aAAa;EACb;IAAY;IAAmB;IAAkB;;EACjD,cAAc,QAAQ,OAAO;;;;;;;EAO7B,MAAM;;iBA4GQ,iBAAiB,QAAQ;UClHxB;EACf;EACA,SAAS;EACT;;EAEA;;UAGe;;EAEf,OAAO,eAAe;EACtB,QAAQ;EACR;IAAY;IAAmB;IAAkB;;;EAEjD,SAAS;;UAGM;EACf;EACA;EACA;EACA,OAAO;EACP,cAAc,QAAQ,OAAO;;EAE7B,MAAM;EACN;;iBAGoB,eACpB,kBACA,SAAS,eACR,QAAQ;UAkIM,8BAA8B;EAC7C;EACA;EACA;;;UCxMe;EACf;IAAe,QAAQ;MAAQ;MAAsB;;;EACrD;IACE;MAAiB;MAAgB;MAAgB;;IACjD;MAAY;MAAgB;MAAkB;;IAC9C;MAAc;MAAgB;MAAkB;;;;UAInC;;EAEf,OAAO;EACP;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe,4BACP,KAAK;;EAEb;;EAEA,eAAe;;EAEf,cAAc,MAAM,mBAAmB;;;;;;;;iBAsFzB,kBACd,gBAAgB,qBAChB,SAAS,sBACR"}
|
|
@@ -1,6 +1,5 @@
|
|
|
1
|
-
import { u as CostProvenance } from "./cost-ledger-
|
|
2
|
-
import { t as DefaultVerdict } from "./verdict-
|
|
3
|
-
//#region src/matrix/cell-spend.d.ts
|
|
1
|
+
import { u as CostProvenance } from "./cost-ledger-Czro9wmo.js";
|
|
2
|
+
import { t as DefaultVerdict } from "./verdict-D_qiBQ6_.js";
|
|
4
3
|
/**
|
|
5
4
|
* Spend carrier for a cell that fails after it has already spent money.
|
|
6
5
|
*
|
|
@@ -43,8 +42,6 @@ interface CellSpend {
|
|
|
43
42
|
* because `NaN >= ceiling` is false.
|
|
44
43
|
*/
|
|
45
44
|
declare function withCellSpend(error: unknown, spend: CellSpend): unknown;
|
|
46
|
-
//#endregion
|
|
47
|
-
//#region src/matrix/types.d.ts
|
|
48
45
|
/** One axis = one dimension to iterate. `V` is the value type — pass any
|
|
49
46
|
* substrate type (AgentProfile, Driver, Validator, rubric record). */
|
|
50
47
|
interface MatrixAxis<V> {
|
|
@@ -186,9 +183,6 @@ interface RunAgentMatrixOptions<Output> {
|
|
|
186
183
|
* and suppresses scheduling of new ones. */
|
|
187
184
|
signal?: AbortSignal;
|
|
188
185
|
}
|
|
189
|
-
//#endregion
|
|
190
|
-
//#region src/matrix/runner.d.ts
|
|
191
186
|
declare function runAgentMatrix<Output>(opts: RunAgentMatrixOptions<Output>): Promise<MatrixResult<Output>>;
|
|
192
|
-
//#endregion
|
|
193
187
|
export { MatrixCell as a, CellSpend as c, MatrixAxis as i, withCellSpend as l, AxisSummary as n, MatrixResult as o, CellResult as r, RunAgentMatrixOptions as s, runAgentMatrix as t };
|
|
194
|
-
//# sourceMappingURL=index-
|
|
188
|
+
//# sourceMappingURL=index-DwUGehqX.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-DwUGehqX.d.ts","names":[],"sources":["../src/matrix/cell-spend.ts","../src/matrix/types.ts","../src/matrix/runner.ts"],"mappings":";;;;;;;;;;;;;;;;;UAsBiB;;;;EAIf;;EAEA;;;;EAIA;;;;;;;;;;;;;;;;iBAiBc,cAAc,gBAAgB,OAAO;;;UC/BpC,WAAW;;EAE1B;;EAEA,QAAQ;IAAQ;IAAY,OAAO;;;;EAGnC,SAAS,OAAO,GAAG;;;UAIJ;EACf,MAAM;IAAiB;IAAY;;;EAEnC;;EAEA;;UAGe,WAAW;EAC1B,QAAQ;EACR,SAAS;;;;EAIT;EACA;EACA;;;;;;;;;;;EAWA,iBAAiB;;;;EAIjB;IAAU;IAAiB;;;UAGZ;EACf;EACA;EACA;EACA;EACA;EACA;EACA;;;EAGA;;EAEA;EACA;;UAGe,aAAa;EAC5B,OAAO;IAAQ,MAAM;IAAY,MAAM,WAAW;;;;EAGlD,QAAQ,eAAe,eAAe;EACtC;IACE;IACA;;;IAGA;IACA;IACA;;;;IAIA;;;;IAIA;;;;IAIA;IACA;;;EAGF;;UAGe,sBAAsB;EACrC,MAAM;;;;;;EAMN,UAAU,MAAM,eAAe,QAAQ,WAAW;;EAElD;;EAEA,UAAU,MAAM,KAAK;;EAErB;;EAEA;;;;;EAKA;;;;;;;;;;;;;EAaA;;EAEA,kBAAkB,MAAM,YAAY,QAAQ,WAAW;;;EAGvD,SAAS;;iBC1BW,eAAe,QACnC,MAAM,sBAAsB,UAC3B,QAAQ,aAAa"}
|
|
@@ -1,15 +1,14 @@
|
|
|
1
|
-
import { o as CostLedgerHandle } from "./cost-ledger-
|
|
2
|
-
import { a as RunRecord, c as RunSplitTag, n as RunCostProvenance } from "./run-record-
|
|
3
|
-
import "./types-
|
|
4
|
-
import "./statistical-heldout-
|
|
5
|
-
import "./index-
|
|
6
|
-
import "./types-
|
|
7
|
-
import { f as TraceSpanEvent } from "./client-
|
|
8
|
-
import { d as InsightReport, t as AnalyzeRunsOptions } from "./analyze-runs-
|
|
9
|
-
import "./define-agent-eval-
|
|
10
|
-
import "./default-registry-
|
|
1
|
+
import { o as CostLedgerHandle } from "./cost-ledger-Czro9wmo.js";
|
|
2
|
+
import { a as RunRecord, c as RunSplitTag, n as RunCostProvenance } from "./run-record-D2BFMPGl.js";
|
|
3
|
+
import "./types-DzCO-S1m.js";
|
|
4
|
+
import "./statistical-heldout-Cei036nE.js";
|
|
5
|
+
import "./index-B7v2QHlD.js";
|
|
6
|
+
import "./types-C54Uk_a1.js";
|
|
7
|
+
import { f as TraceSpanEvent } from "./client-BIFwUg7I.js";
|
|
8
|
+
import { d as InsightReport, t as AnalyzeRunsOptions } from "./analyze-runs-C-pP1sK9.js";
|
|
9
|
+
import "./define-agent-eval-CAffiFbN.js";
|
|
10
|
+
import "./default-registry-B1DstPYg.js";
|
|
11
11
|
import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementCost, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
|
|
12
|
-
//#region src/contract/measured-comparison-types.d.ts
|
|
13
12
|
interface SealCandidateBenchmarkSuiteOptions {
|
|
14
13
|
tasks: [AgentCandidateBenchmarkTask, ...AgentCandidateBenchmarkTask[]];
|
|
15
14
|
reps: number;
|
|
@@ -89,8 +88,6 @@ type PairedMeasurementEvaluation = Pick<AgentImprovementMeasuredComparison, 'ove
|
|
|
89
88
|
totalCost: AgentImprovementCost;
|
|
90
89
|
measurementWorkDurationMs: number;
|
|
91
90
|
};
|
|
92
|
-
//#endregion
|
|
93
|
-
//#region src/contract/candidate-experiment.d.ts
|
|
94
91
|
/** Content-address one task before any measured execution can see it. */
|
|
95
92
|
declare function sealCandidateBenchmarkTask(material: AgentCandidateBenchmarkTaskMaterial): AgentCandidateBenchmarkTask;
|
|
96
93
|
/** Freeze task order, repetitions, and every seed before either arm runs. */
|
|
@@ -104,8 +101,6 @@ declare function runCandidateExperiment(options: RunCandidateExperimentOptions):
|
|
|
104
101
|
declare function measuredComparisonFromCandidateExperiment(options: CompareCandidateExperimentOptions): AgentImprovementMeasuredComparison;
|
|
105
102
|
/** Recompute every statistic and decision from the signed experiment receipts. */
|
|
106
103
|
declare function verifyCandidateExperimentComparison(input: unknown): AgentImprovementMeasuredComparison;
|
|
107
|
-
//#endregion
|
|
108
|
-
//#region src/contract/paired-measurement-evaluation.d.ts
|
|
109
104
|
/**
|
|
110
105
|
* Calculate the shared paired decision from any complete receipt shape.
|
|
111
106
|
*
|
|
@@ -114,8 +109,6 @@ declare function verifyCandidateExperimentComparison(input: unknown): AgentImpro
|
|
|
114
109
|
* only validates the projected measurements and derives their shared decision.
|
|
115
110
|
*/
|
|
116
111
|
declare function evaluatePairedMeasurements<TRun>(options: EvaluatePairedMeasurementsOptions<TRun>): PairedMeasurementEvaluation;
|
|
117
|
-
//#endregion
|
|
118
|
-
//#region src/contract/profile-measured-comparison.d.ts
|
|
119
112
|
interface SealAgentProfileImprovementSuiteOptions {
|
|
120
113
|
splitDigest: Sha256Digest;
|
|
121
114
|
tasks: [AgentProfileImprovementTask, ...AgentProfileImprovementTask[]];
|
|
@@ -176,8 +169,6 @@ declare function runAgentProfileImprovementExperiment(options: RunAgentProfileIm
|
|
|
176
169
|
declare function measuredComparisonFromAgentProfileImprovementExperiment(options: CompareAgentProfileImprovementExperimentOptions): AgentProfileImprovementMeasuredComparison;
|
|
177
170
|
/** Recompute a profile comparison from the exact sealed experiment and receipts. */
|
|
178
171
|
declare function verifyAgentProfileImprovementExperimentComparison(input: unknown): AgentProfileImprovementMeasuredComparison;
|
|
179
|
-
//#endregion
|
|
180
|
-
//#region src/contract/intake/run-record-dir.d.ts
|
|
181
172
|
/** A record that failed boundary validation, with enough context to fix it. */
|
|
182
173
|
interface RunRecordRejection {
|
|
183
174
|
/** Absolute or caller-relative path to the file the record came from. */
|
|
@@ -216,8 +207,6 @@ interface FromRunRecordDirResult {
|
|
|
216
207
|
/** The files that were read, in the order they were processed. */
|
|
217
208
|
files: string[];
|
|
218
209
|
}
|
|
219
|
-
//#endregion
|
|
220
|
-
//#region src/contract/eval-reporting-suite.d.ts
|
|
221
210
|
/** Either records in hand or a path to a `.json` / `.jsonl` file or a
|
|
222
211
|
* directory of them. */
|
|
223
212
|
type EvalReportingSuiteInput = RunRecord[] | string;
|
|
@@ -264,8 +253,6 @@ interface EvalReportingSuiteResult {
|
|
|
264
253
|
/** The path `analysis.json` was written to, or null when `write` was unset. */
|
|
265
254
|
writtenTo: string | null;
|
|
266
255
|
}
|
|
267
|
-
//#endregion
|
|
268
|
-
//#region src/contract/intake/agent-trace.d.ts
|
|
269
256
|
type AgentTraceContributorType = 'human' | 'ai' | 'mixed' | 'unknown';
|
|
270
257
|
interface AgentTraceContributor {
|
|
271
258
|
type: AgentTraceContributorType;
|
|
@@ -341,8 +328,6 @@ interface PartitionByAuthoringModelResult {
|
|
|
341
328
|
* on "model A's code vs model B's code".
|
|
342
329
|
*/
|
|
343
330
|
declare function partitionRunsByAuthoringModel(runs: RunRecord[], index: AgentTraceIndex): PartitionByAuthoringModelResult;
|
|
344
|
-
//#endregion
|
|
345
|
-
//#region src/contract/intake/code-agent-observation.d.ts
|
|
346
331
|
/**
|
|
347
332
|
* Where a session came from. `pi` is a native Pi session (`~/.pi/agent/sessions/*.jsonl`), read
|
|
348
333
|
* through @tangle-network/harness-sessions; `pi-graph` is the graph IR a Pi graph run exports
|
|
@@ -391,8 +376,6 @@ interface ObserveCodeAgentSessionOptions {
|
|
|
391
376
|
* stay out of this projection; callers retain the source JSONL as evidence.
|
|
392
377
|
*/
|
|
393
378
|
declare function observeCodeAgentSession(options: ObserveCodeAgentSessionOptions): CodeAgentSessionObservation;
|
|
394
|
-
//#endregion
|
|
395
|
-
//#region src/contract/intake/code-agent-session-types.d.ts
|
|
396
379
|
interface ParsedCodeAgentJsonl {
|
|
397
380
|
entries: unknown[];
|
|
398
381
|
malformedLines: number;
|
|
@@ -489,8 +472,6 @@ type CodeAgentJsonlLine = {
|
|
|
489
472
|
kind: 'malformed';
|
|
490
473
|
lineNumber: number;
|
|
491
474
|
};
|
|
492
|
-
//#endregion
|
|
493
|
-
//#region src/contract/intake/code-agent-jsonl.d.ts
|
|
494
475
|
declare function parseCodeAgentJsonl(jsonl: string): ParsedCodeAgentJsonl;
|
|
495
476
|
/** Reads a transcript one line at a time and never holds the file as a single
|
|
496
477
|
* string. `parseCodeAgentJsonl` needs the whole file in one string, so a
|
|
@@ -507,16 +488,12 @@ declare function streamCodeAgentJsonlFile(path: string): AsyncGenerator<CodeAgen
|
|
|
507
488
|
* above the string ceiling. The entry array still grows with the transcript;
|
|
508
489
|
* consume `streamCodeAgentJsonlFile` directly when memory must stay flat. */
|
|
509
490
|
declare function parseCodeAgentJsonlFile(path: string): Promise<ParsedCodeAgentJsonl>;
|
|
510
|
-
//#endregion
|
|
511
|
-
//#region src/contract/intake/code-agent-session.d.ts
|
|
512
491
|
declare function fromCodexSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
513
492
|
declare function fromClaudeCodeSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
514
493
|
declare function fromOpenCodeSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
515
494
|
declare function fromKimiCodeSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
516
495
|
/** A native Pi session (`~/.pi/agent/sessions/<cwd>/<ts>_<id>.jsonl`), parsed entries in order. */
|
|
517
496
|
declare function fromPiSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
518
|
-
//#endregion
|
|
519
|
-
//#region src/contract/intake/code-agent-store.d.ts
|
|
520
497
|
/** Which store to read, and what makes a session in it belong to this run. */
|
|
521
498
|
interface CodeAgentStoreScope {
|
|
522
499
|
/**
|
|
@@ -588,8 +565,6 @@ type CodeAgentStoreObservation = {
|
|
|
588
565
|
readonly reason: string;
|
|
589
566
|
readonly rejections: CodeAgentStoreRejections;
|
|
590
567
|
};
|
|
591
|
-
//#endregion
|
|
592
|
-
//#region src/contract/intake/feedback-table.d.ts
|
|
593
568
|
interface FeedbackTableRow {
|
|
594
569
|
/** Stable id for this run — the unit a rater scored. Drives pairing
|
|
595
570
|
* across analysis primitives. */
|
|
@@ -660,8 +635,6 @@ interface FromFeedbackTableResult {
|
|
|
660
635
|
}>;
|
|
661
636
|
}
|
|
662
637
|
declare function fromFeedbackTable(opts: FromFeedbackTableOptions): FromFeedbackTableResult;
|
|
663
|
-
//#endregion
|
|
664
|
-
//#region src/contract/intake/otel-spans.d.ts
|
|
665
638
|
interface FromOtelSpansOptions {
|
|
666
639
|
spans: TraceSpanEvent[];
|
|
667
640
|
/** Default split tag for synthesized records. Defaults to `'holdout'`. */
|
|
@@ -676,6 +649,5 @@ interface FromOtelSpansOptions {
|
|
|
676
649
|
scoreForRun?: (runId: string, spans: readonly TraceSpanEvent[]) => number | undefined;
|
|
677
650
|
}
|
|
678
651
|
declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
|
|
679
|
-
//#endregion
|
|
680
652
|
export { AgentProfileImprovementExperimentExecutionInput as $, CodeAgentSessionActionSurface as A, AgentTraceIndex as B, CodeAgentSessionIntakeOptions as C, PairedMeasurementAdapter as Ct, CodeAgentSessionAction as D, ParsedCodeAgentJsonl as E, SealCandidateBenchmarkSuiteOptions as Et, observeCodeAgentSession as F, parseAgentTrace as G, AgentTraceRecord as H, AgentTraceContributor as I, EvalReportingSuiteOptions as J, partitionRunsByAuthoringModel as K, AgentTraceContributorType as L, CodeAgentSessionObservation as M, CodeAgentSessionSource as N, CodeAgentSessionActionKind as O, CodeAgentSessionTerminalStatus as P, RunRecordRejection as Q, AgentTraceConversation as R, CodeAgentSessionDiagnostic as S, PairedMeasurement as St, CodeAgentSessionMetrics as T, RunCandidateExperimentOptions as Tt, AuthoringProvenance as U, AgentTraceRange as V, PartitionByAuthoringModelResult as W, FromRunRecordDirOptions as X, EvalReportingSuiteResult as Y, FromRunRecordDirResult as Z, fromPiSession as _, verifyCandidateExperimentComparison as _t, FromFeedbackTableOptions as a, runAgentProfileImprovementExperiment as at, streamCodeAgentJsonlFile as b, CompareCandidateExperimentOptions as bt, CodeAgentStoreObservation as c, sealAgentProfileImprovementTask as ct, CodeAgentStoreScope as d, measuredComparisonFromCandidateExperiment as dt, AgentProfileImprovementExperimentRun as et, CodeAgentStoreSession as f, runCandidateExperiment as ft, fromOpenCodeSession as g, verifyCandidateExperiment as gt, fromKimiCodeSession as h, sealCandidateExperiment as ht, FeedbackTableRow as i, measuredComparisonFromAgentProfileImprovementExperiment as it, CodeAgentSessionExecutionReceipt as j, CodeAgentSessionActionStatus as k, CodeAgentStoreRejection as l, verifyAgentProfileImprovementExperimentComparison as lt, fromCodexSession as m, sealCandidateBenchmarkTask as mt, fromOtelSpans as n, RunAgentProfileImprovementExperimentOptions as nt, FromFeedbackTableResult as o, sealAgentProfileImprovementExperiment as ot, fromClaudeCodeSession as p, sealCandidateBenchmarkSuite as pt, EvalReportingSuiteInput as q, FeedbackTableMeta as r, SealAgentProfileImprovementSuiteOptions as rt, fromFeedbackTable as s, sealAgentProfileImprovementSuite as st, FromOtelSpansOptions as t, CompareAgentProfileImprovementExperimentOptions as tt, CodeAgentStoreRejections as u, evaluatePairedMeasurements as ut, parseCodeAgentJsonl as v, CandidateExperimentExecutionInput as vt, CodeAgentSessionIntakeResult as w, PairedMeasurementEvaluation as wt, CodeAgentJsonlLine as x, EvaluatePairedMeasurementsOptions as xt, parseCodeAgentJsonlFile as y, CandidateExperimentRun as yt, AgentTraceFile as z };
|
|
681
|
-
//# sourceMappingURL=index-
|
|
653
|
+
//# sourceMappingURL=index-aDPVYBIM.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-aDPVYBIM.d.ts","names":[],"sources":["../src/contract/measured-comparison-types.ts","../src/contract/candidate-experiment.ts","../src/contract/paired-measurement-evaluation.ts","../src/contract/profile-measured-comparison.ts","../src/contract/intake/run-record-dir.ts","../src/contract/eval-reporting-suite.ts","../src/contract/intake/agent-trace.ts","../src/contract/intake/code-agent-observation.ts","../src/contract/intake/code-agent-session-types.ts","../src/contract/intake/code-agent-jsonl.ts","../src/contract/intake/code-agent-session.ts","../src/contract/intake/code-agent-store.ts","../src/contract/intake/feedback-table.ts","../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;UAeiB;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,eAAe,KAAK,OAAO;EAC3B,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA,kBAAkB;;EAElB,kBAAkB;;;KAIR,8BAA8B,KACxC;EAGA,iBAAiB;EACjB,WAAW;EACX;;;iBC9Dc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;iBA4DK,0CACd,SAAS,oCACR;;iBA2Da,oCACd,iBACC;;;;;;;;iBC3La,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;UCDc;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;;;;;;iBAsBmB,qCACpB,SAAS,8CACR,QAAQ;;iBA0CK,wDACd,SAAS,kDACR;;iBA2Da,kDACd,iBACC;;UClOc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;KCxBU,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;KCnDU;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;;;;KCjKS;KAQA;KAEA;KAEA;KASA;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,MAAM;EACN,SAAS;EACT;EACA,QAAQ;EACR;EACA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR;EACA;EACA;IACE,QAAQ;IACR;;EAEF,SAAS;;UAGD;EACR,QAAQ;EACR;EACA;EACA,YAAY;;;;;;;iBAeE,wBACd,SAAS,iCACR;UClFc;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;;UAGe;EACf,MAAM;EACN,aAAa;EACb,SAAS;EACT,cAAc;;UAGC;EACf;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;EACA;EACA;EACA;EACA;EACA;;;EAGA,iBAAiB;;;EAGjB,YAAY;;;;KAKF;EACN;EAAe;EAAoB;;EACnC;EAAmB;;iBCvFT,oBAAoB,gBAAgB;;;;;;;;;;iBAuB7B,yBAAyB,eAAe,eAAe;;;;;iBA6BxD,wBAAwB,eAAe,QAAQ;iBCDrD,iBACd,SAAS,gCACR;iBAIa,sBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;;iBAKa,cACd,SAAS,gCACR;;UCrDc;;;;;WAKN;;WAEA,QAAQ;;;;;WAKR;;;;;WAKA;aACE;;aAEA;;;;KAKD;;UAUK;WACN;WACA;WACA;WACA;WACA,aAAa;;;UAIP;WACN;WACA,UAAU,SAAS,QAAQ,OAAO;;WAElC,QAAQ;aACN;aACA,QAAQ;;;;;;;;;KAUT;WAEG;WACA,OAAO;WACP;WACA,mBAAmB;;WAEnB;;WAEA;;WAEA;WACA,WAAW,SAAS,QAAQ,OAAO;WACnC,YAAY;;WAGZ;WACA,OAAO;WACP;;WAEA;WACA,YAAY;;UCvGV;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;UC3BlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}
|