reeval 0.11.0 → 0.13.0
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/README.md +115 -257
- package/assets/package.json +1 -1
- package/assets/pnpm-lock.yaml +41 -30
- package/bin/mcp-timer.cjs +224 -0
- package/dist/{adapter-Bwy3QxAK.js → adapter-IKKo4sZj.js} +4 -4
- package/dist/{adapter-Bwy3QxAK.js.map → adapter-IKKo4sZj.js.map} +1 -1
- package/dist/{agent-DpE80WPz.js → agent-ClE-g9VL.js} +22 -7
- package/dist/{agent-DpE80WPz.js.map → agent-ClE-g9VL.js.map} +1 -1
- package/dist/{analysis-Bxub9LMT.js → analysis-B0nOfAzA.js} +3 -3
- package/dist/{analysis-Bxub9LMT.js.map → analysis-B0nOfAzA.js.map} +1 -1
- package/dist/{assemble-D9GWr-HS.js → assemble-BO93VAk1.js} +22 -20
- package/dist/assemble-BO93VAk1.js.map +1 -0
- package/dist/assemble-VQxyoDxK.js +1 -0
- package/dist/{attempt-CGPV3BZY.js → attempt-jI0OWxlz.js} +2 -2
- package/dist/{attempt-CGPV3BZY.js.map → attempt-jI0OWxlz.js.map} +1 -1
- package/dist/bridge/index.mjs +2 -1
- package/dist/{capture-BQlGHiMC.js → capture-dn2hjSLj.js} +29 -21
- package/dist/{capture-BQlGHiMC.js.map → capture-dn2hjSLj.js.map} +1 -1
- package/dist/capture-nIJ-7kUE.js +1 -0
- package/dist/{cli-ClpPT8kn.js → cli-BCE1xlm4.js} +176 -120
- package/dist/cli-BCE1xlm4.js.map +1 -0
- package/dist/{cli-spec-qmEnQ25_.js → cli-spec-BXIBUlaK.js} +34 -15
- package/dist/cli-spec-BXIBUlaK.js.map +1 -0
- package/dist/cli.js +2 -2
- package/dist/{command-BlM0y8jN.js → command-Bh3yhN_b.js} +8 -8
- package/dist/{command-BlM0y8jN.js.map → command-Bh3yhN_b.js.map} +1 -1
- package/dist/command-C6vwmjS7.js +91 -0
- package/dist/command-C6vwmjS7.js.map +1 -0
- package/dist/{command-DbAe1Rzb.js → command-CANun0m0.js} +16 -19
- package/dist/command-CANun0m0.js.map +1 -0
- package/dist/command-CBaHukyl.js +335 -0
- package/dist/command-CBaHukyl.js.map +1 -0
- package/dist/{command-TgbzPizb.js → command-D_kEnYcX.js} +17 -16
- package/dist/command-D_kEnYcX.js.map +1 -0
- package/dist/commands-7tH3EsVu.js +666 -0
- package/dist/commands-7tH3EsVu.js.map +1 -0
- package/dist/{commands-CthqwkAn.js → commands-BKGzBMgb.js} +1 -1
- package/dist/{config-DrVSEVN0.js → config-D7TFvUpL.js} +40 -55
- package/dist/config-D7TFvUpL.js.map +1 -0
- package/dist/{context-C1mxXh80.js → context-Dgn48BqT.js} +13 -2
- package/dist/context-Dgn48BqT.js.map +1 -0
- package/dist/{contrasts-5zsj32gk.js → contrasts-B4nOhn0Z.js} +3 -3
- package/dist/{contrasts-5zsj32gk.js.map → contrasts-B4nOhn0Z.js.map} +1 -1
- package/dist/db-DJUuFoWQ.js +602 -0
- package/dist/db-DJUuFoWQ.js.map +1 -0
- package/dist/db-W9EFlXnn.js +2 -0
- package/dist/{dist-CuY4ehEk.js → dist-Dp_AJgYN.js} +3 -3
- package/dist/{dist-CuY4ehEk.js.map → dist-Dp_AJgYN.js.map} +1 -1
- package/dist/{docker-D1HZDS6q.js → docker-CQxucRPe.js} +16 -15
- package/dist/docker-CQxucRPe.js.map +1 -0
- package/dist/{docker-DNmMPOBb.js → docker-Dac6Zalz.js} +1 -1
- package/dist/{docker-DOvjji9I.js → docker-DuHcQidD.js} +3 -3
- package/dist/{docker-DOvjji9I.js.map → docker-DuHcQidD.js.map} +1 -1
- package/dist/{doctor-D2MbNO8L.js → doctor-De_QlOTM.js} +72 -38
- package/dist/doctor-De_QlOTM.js.map +1 -0
- package/dist/{egress-2-qzPvGi.js → egress-DEAvq8Fl.js} +453 -31
- package/dist/egress-DEAvq8Fl.js.map +1 -0
- package/dist/env-BQqj93gu.js +21 -0
- package/dist/env-BQqj93gu.js.map +1 -0
- package/dist/{environment-Cca6a4wB.js → environment-B4IuURK1.js} +247 -14
- package/dist/environment-B4IuURK1.js.map +1 -0
- package/dist/{evaluate-DMyaGQEj.d.ts → evaluate-C9-ritA5.d.ts} +2 -2
- package/dist/{evaluate-DKIl54Hb.js → evaluate-Djk1YPFL.js} +540 -560
- package/dist/evaluate-Djk1YPFL.js.map +1 -0
- package/dist/facts-Cfx1H70y.js +898 -0
- package/dist/facts-Cfx1H70y.js.map +1 -0
- package/dist/files-Bs6H3HiG.js +65 -0
- package/dist/files-Bs6H3HiG.js.map +1 -0
- package/dist/{files-BpUYBwGu.js → files-CN1QfenC.js} +31 -8
- package/dist/files-CN1QfenC.js.map +1 -0
- package/dist/files-CtjEXWXm.js +1 -0
- package/dist/{files-BXHZFWmx.js → files-D27nUk0K.js} +6 -17
- package/dist/files-D27nUk0K.js.map +1 -0
- package/dist/files-D3gJa91B.js +1 -0
- package/dist/generate-BnrvBAzF.js +1646 -0
- package/dist/generate-BnrvBAzF.js.map +1 -0
- package/dist/{git-Crl3b87o.js → git-rPfl7AD1.js} +3 -3
- package/dist/{git-Crl3b87o.js.map → git-rPfl7AD1.js.map} +1 -1
- package/dist/{harness-CEuntrla.js → harness-1OkumOAm.js} +3 -3
- package/dist/{harness-CEuntrla.js.map → harness-1OkumOAm.js.map} +1 -1
- package/dist/harness.d.ts +1 -1
- package/dist/harness.js +1 -1
- package/dist/{header-DHQv_ZrF.js → header-BVzbRnNv.js} +10 -3
- package/dist/{header-DHQv_ZrF.js.map → header-BVzbRnNv.js.map} +1 -1
- package/dist/{images-Bi3aGxb9.js → images-m784ueGK.js} +61 -13
- package/dist/images-m784ueGK.js.map +1 -0
- package/dist/index.d.ts +3 -3
- package/dist/index.js +4 -4
- package/dist/init-hiTkiM4L.js +543 -0
- package/dist/init-hiTkiM4L.js.map +1 -0
- package/dist/init-kMkTgTZc.js +1 -0
- package/dist/insights-B2Xmhz9o.js +64 -0
- package/dist/insights-B2Xmhz9o.js.map +1 -0
- package/dist/{invocations-BYDPe3fV.js → invocations-D_ePcwyL.js} +4 -4
- package/dist/{invocations-BYDPe3fV.js.map → invocations-D_ePcwyL.js.map} +1 -1
- package/dist/{judge-DW_GDOmB.js → judge-CEDIfRiU.js} +50 -15
- package/dist/judge-CEDIfRiU.js.map +1 -0
- package/dist/judge.d.ts +7 -6
- package/dist/judge.js +8 -7
- package/dist/judge.js.map +1 -1
- package/dist/{live-jlfszAbb.js → live-DYd2QEyR.js} +2 -2
- package/dist/{live-jlfszAbb.js.map → live-DYd2QEyR.js.map} +1 -1
- package/dist/model-B8H3gL69.js +1 -0
- package/dist/{model-DaVCY3rb.js → model-CTiuzkWw.js} +9 -7
- package/dist/model-CTiuzkWw.js.map +1 -0
- package/dist/model-DcIbmggK.js +1 -0
- package/dist/{model-D0IcOQao.js → model-DpkclE4r.js} +60 -25
- package/dist/model-DpkclE4r.js.map +1 -0
- package/dist/{model-DaX4bNLZ.js → model-UgHDMhI1.js} +144 -46
- package/dist/model-UgHDMhI1.js.map +1 -0
- package/dist/{node-B3ydHo6t.js → node-CGZYh4Fj.js} +4 -4
- package/dist/{node-B3ydHo6t.js.map → node-CGZYh4Fj.js.map} +1 -1
- package/dist/{opencode-providers-J9colP0-.js → opencode-providers-snfeWd6Z.js} +86 -47
- package/dist/opencode-providers-snfeWd6Z.js.map +1 -0
- package/dist/{options-C0d0AzeR.js → options-CyAyLjgJ.js} +22 -11
- package/dist/options-CyAyLjgJ.js.map +1 -0
- package/dist/{pairs-CnInRu-G.js → pairs-DP6XOJby.js} +2 -2
- package/dist/{pairs-CnInRu-G.js.map → pairs-DP6XOJby.js.map} +1 -1
- package/dist/{patch-B1bnfhN-.js → patch-uMq9rCl_.js} +2 -2
- package/dist/{patch-B1bnfhN-.js.map → patch-uMq9rCl_.js.map} +1 -1
- package/dist/{plan-DgQxp9yy.js → plan-CZYgXnZh.js} +21 -8
- package/dist/plan-CZYgXnZh.js.map +1 -0
- package/dist/plugins-P-uh54VF.js +654 -0
- package/dist/plugins-P-uh54VF.js.map +1 -0
- package/dist/{preflight-DlQvbySb.js → preflight-BQlPTHVu.js} +1 -1
- package/dist/{preflight-BRzD30L5.js → preflight-D3HvKeAx.js} +14 -10
- package/dist/preflight-D3HvKeAx.js.map +1 -0
- package/dist/prepare-DCwM2KKD.js +2 -0
- package/dist/{prepare-B_y-AFMv.js → prepare-Dx7dOnJB.js} +70 -23
- package/dist/prepare-Dx7dOnJB.js.map +1 -0
- package/dist/{provenance-DQ0AXn6c.js → provenance-C_-8y-x3.js} +4 -4
- package/dist/{provenance-DQ0AXn6c.js.map → provenance-C_-8y-x3.js.map} +1 -1
- package/dist/{providers-DoTq1jHW.js → providers-CF5UEaP8.js} +24 -10
- package/dist/providers-CF5UEaP8.js.map +1 -0
- package/dist/providers-CTBYvMxu.js +1 -0
- package/dist/{query-BO-ux7RB.js → query-DXthKmq0.js} +3 -3
- package/dist/{query-BO-ux7RB.js.map → query-DXthKmq0.js.map} +1 -1
- package/dist/query-D_1ARx5q.js +1 -0
- package/dist/{questions-C2dym2ce.js → questions-KENpWnfV.js} +8 -9
- package/dist/questions-KENpWnfV.js.map +1 -0
- package/dist/{record-BKmMvZ6O.js → record-kk9M1hj-.js} +8 -4
- package/dist/record-kk9M1hj-.js.map +1 -0
- package/dist/{redact-DJ9JCShu.js → redact-CRGdfV2s.js} +4 -2
- package/dist/redact-CRGdfV2s.js.map +1 -0
- package/dist/remarks-1521zwfW.js +38 -0
- package/dist/remarks-1521zwfW.js.map +1 -0
- package/dist/remarks-BhAV1H4V.js +1 -0
- package/dist/remarks-Ml6bcSdV.js +278 -0
- package/dist/remarks-Ml6bcSdV.js.map +1 -0
- package/dist/{remote-CPPW86Q9.js → remote-Dkeghk8k.js} +3 -3
- package/dist/{remote-CPPW86Q9.js.map → remote-Dkeghk8k.js.map} +1 -1
- package/dist/resolve-CMmhACr6.js +1 -0
- package/dist/resolve-DzdWcIL6.js +724 -0
- package/dist/resolve-DzdWcIL6.js.map +1 -0
- package/dist/{results-BE2f-1XS.js → results-D-naai_Q.js} +33 -8
- package/dist/results-D-naai_Q.js.map +1 -0
- package/dist/{resume-Ca1c8vix.js → resume-8-nnW8iO.js} +4 -4
- package/dist/{resume-Ca1c8vix.js.map → resume-8-nnW8iO.js.map} +1 -1
- package/dist/{root-DICDHCtp.js → root-DNuLhaA5.js} +2 -2
- package/dist/root-DNuLhaA5.js.map +1 -0
- package/dist/rules-CcV3zLDR.js +318 -0
- package/dist/rules-CcV3zLDR.js.map +1 -0
- package/dist/run-Bel7Vde4.js +88 -0
- package/dist/run-Bel7Vde4.js.map +1 -0
- package/dist/{sandbox-DLrsvu7U.js → sandbox-YLbk9mO7.js} +2 -2
- package/dist/{sandbox-DLrsvu7U.js.map → sandbox-YLbk9mO7.js.map} +1 -1
- package/dist/sandbox.d.ts +1 -1
- package/dist/sandbox.js +1 -1
- package/dist/server-C-ysqkLv.js +1 -0
- package/dist/{server-TUY5xA_A.js → server-De8oH3bt.js} +10 -10
- package/dist/{server-TUY5xA_A.js.map → server-De8oH3bt.js.map} +1 -1
- package/dist/{source-DJGS0a0u.js → source-C686eZpY.js} +2 -2
- package/dist/{source-DJGS0a0u.js.map → source-C686eZpY.js.map} +1 -1
- package/dist/source-DjgypUHW.js +1 -0
- package/dist/src-3mP821Pt.js +2 -0
- package/dist/src-B3KCLZVk.js +3 -0
- package/dist/src-BAyg6CtQ.js +2 -0
- package/dist/{src-CF17Ug08.js → src-BYLM3hEz.js} +22 -6
- package/dist/src-BYLM3hEz.js.map +1 -0
- package/dist/src-BezgGX4X.js +2 -0
- package/dist/{src-DvCg0NBs.js → src-C7W6kg81.js} +40 -10
- package/dist/{src-DvCg0NBs.js.map → src-C7W6kg81.js.map} +1 -1
- package/dist/{src-ftu5wPSt.js → src-CIpKL45U.js} +782 -169
- package/dist/src-CIpKL45U.js.map +1 -0
- package/dist/src-ChBhk06Q.js +2 -0
- package/dist/{src-BxfCKA3_.js → src-CoqZYspK.js} +1 -1
- package/dist/{src-CFp86GtV.js → src-CzQrAcF8.js} +49 -49
- package/dist/src-CzQrAcF8.js.map +1 -0
- package/dist/src-D9fSTxlc.js +1 -0
- package/dist/{src-FGXhkEkH.js → src-DeY52Gyf.js} +182 -23
- package/dist/src-DeY52Gyf.js.map +1 -0
- package/dist/src-DrJSnoMR.js +2 -0
- package/dist/{src-DvK0EIZx.js → src-Xj_BZKbq.js} +112 -13
- package/dist/src-Xj_BZKbq.js.map +1 -0
- package/dist/{src-u8_ixn_F.js → src-cbCb-MPy.js} +20 -4
- package/dist/src-cbCb-MPy.js.map +1 -0
- package/dist/src-dhSTRN50.js +2 -0
- package/dist/{src-C7_1SqDP.js → src-djNNYIdP.js} +2 -2
- package/dist/{src-C7_1SqDP.js.map → src-djNNYIdP.js.map} +1 -1
- package/dist/state-BGWGAwdo.js +1 -0
- package/dist/{state-DfpHox6b.js → state-C0i_ujQZ.js} +19 -17
- package/dist/{state-DfpHox6b.js.map → state-C0i_ujQZ.js.map} +1 -1
- package/dist/studio-app/assets/{index-CBkhNAv8.js → index-Bb_cid9k.js} +9 -9
- package/dist/studio-app/index.html +1 -1
- package/dist/{suite-Cw_7G2RM.js → suite-DmuF_ApO.js} +80 -30
- package/dist/suite-DmuF_ApO.js.map +1 -0
- package/dist/suite-mn9g4veR.js +2 -0
- package/dist/{summary-BUKfU-9l.js → summary-BFMHMTz2.js} +273 -47
- package/dist/summary-BFMHMTz2.js.map +1 -0
- package/dist/testing.d.ts +2 -2
- package/dist/testing.js +2 -2
- package/dist/{text-gEcKjAyp.js → text-CrVy5ioL.js} +18 -2
- package/dist/text-CrVy5ioL.js.map +1 -0
- package/dist/{transcript-BNNANU8A.js → transcript-BTxJpW4R.js} +2 -2
- package/dist/{transcript-BNNANU8A.js.map → transcript-BTxJpW4R.js.map} +1 -1
- package/dist/{types-BYdlwMwo.d.ts → types-B-9b1MDV.d.ts} +850 -34
- package/dist/types-C5vI3g1C.js.map +1 -1
- package/dist/usage-BDo2q_yu.js +1 -0
- package/dist/{command-Bj39y-Wm.js → validate-DWjkshl1.js} +20 -377
- package/dist/validate-DWjkshl1.js.map +1 -0
- package/dist/view/index.d.ts +31 -263
- package/dist/view/index.js +8 -7
- package/dist/views-DE2A1jEX.js +2 -0
- package/dist/{views-B8w2J-gD.js → views-Dk8MCQKr.js} +8 -98
- package/dist/views-Dk8MCQKr.js.map +1 -0
- package/dist/{worker-entry-DNnveLqx.js → worker-entry-CqrzGPF4.js} +2 -2
- package/dist/{worker-entry-DNnveLqx.js.map → worker-entry-CqrzGPF4.js.map} +1 -1
- package/dist/workspace-B1eGO46M.js +1 -0
- package/dist/{workspace-C4RA3srV.js → workspace-CLeOybkr.js} +45 -11
- package/dist/workspace-CLeOybkr.js.map +1 -0
- package/package.json +5 -1
- package/skills/eval-rubric/SKILL.md +130 -0
- package/dist/agent-D483ru7D.js +0 -334
- package/dist/agent-D483ru7D.js.map +0 -1
- package/dist/assemble-D5KGrwnJ.js +0 -1
- package/dist/assemble-D9GWr-HS.js.map +0 -1
- package/dist/capture-CYGLpWxR.js +0 -1
- package/dist/cli-ClpPT8kn.js.map +0 -1
- package/dist/cli-spec-qmEnQ25_.js.map +0 -1
- package/dist/command-Bj39y-Wm.js.map +0 -1
- package/dist/command-C-cVjlny.js +0 -1
- package/dist/command-DbAe1Rzb.js.map +0 -1
- package/dist/command-TgbzPizb.js.map +0 -1
- package/dist/commands-YUwlyTbr.js +0 -228
- package/dist/commands-YUwlyTbr.js.map +0 -1
- package/dist/config-DrVSEVN0.js.map +0 -1
- package/dist/context-C1mxXh80.js.map +0 -1
- package/dist/db-CKg5NtiW.js +0 -2
- package/dist/db-DfhipHEF.js +0 -1416
- package/dist/db-DfhipHEF.js.map +0 -1
- package/dist/docker-D1HZDS6q.js.map +0 -1
- package/dist/doctor-D2MbNO8L.js.map +0 -1
- package/dist/egress-2-qzPvGi.js.map +0 -1
- package/dist/environment-Cca6a4wB.js.map +0 -1
- package/dist/evaluate-DKIl54Hb.js.map +0 -1
- package/dist/files-BXHZFWmx.js.map +0 -1
- package/dist/files-BpUYBwGu.js.map +0 -1
- package/dist/files-Bw8lWEjg.js +0 -1
- package/dist/files-D6J0iSdx.js +0 -1
- package/dist/images-Bi3aGxb9.js.map +0 -1
- package/dist/init-Cl3f3NBT.js +0 -360
- package/dist/init-Cl3f3NBT.js.map +0 -1
- package/dist/init-fhOXED8F.js +0 -1
- package/dist/judge-DW_GDOmB.js.map +0 -1
- package/dist/model-CXhrziBe.js +0 -1
- package/dist/model-D0IcOQao.js.map +0 -1
- package/dist/model-DaVCY3rb.js.map +0 -1
- package/dist/model-DaX4bNLZ.js.map +0 -1
- package/dist/model-_bxNrk2h.js +0 -1
- package/dist/opencode-providers-J9colP0-.js.map +0 -1
- package/dist/options-C0d0AzeR.js.map +0 -1
- package/dist/plan-DgQxp9yy.js.map +0 -1
- package/dist/preflight-BRzD30L5.js.map +0 -1
- package/dist/prepare-B_y-AFMv.js.map +0 -1
- package/dist/prepare-DCKTLHN3.js +0 -2
- package/dist/providers-B9yTdIqF.js +0 -1
- package/dist/providers-DoTq1jHW.js.map +0 -1
- package/dist/query-CRQfsw0e.js +0 -1
- package/dist/questions-C2dym2ce.js.map +0 -1
- package/dist/record-BKmMvZ6O.js.map +0 -1
- package/dist/redact-DJ9JCShu.js.map +0 -1
- package/dist/results-BE2f-1XS.js.map +0 -1
- package/dist/root-DICDHCtp.js.map +0 -1
- package/dist/server-Dp8eOeK5.js +0 -1
- package/dist/source-DdbLGR1M.js +0 -1
- package/dist/src-CF17Ug08.js.map +0 -1
- package/dist/src-CFp86GtV.js.map +0 -1
- package/dist/src-CLoOMVWI.js +0 -3
- package/dist/src-CcrQbpG8.js +0 -2
- package/dist/src-DCFtJRp0.js +0 -2
- package/dist/src-DSYYPEOj.js +0 -2
- package/dist/src-DYO3DdnQ.js +0 -1
- package/dist/src-DvK0EIZx.js.map +0 -1
- package/dist/src-FGXhkEkH.js.map +0 -1
- package/dist/src-WD-zgQYZ.js +0 -2
- package/dist/src-_oI9AePb.js +0 -2
- package/dist/src-ftu5wPSt.js.map +0 -1
- package/dist/src-qsKCqv7a.js +0 -2
- package/dist/src-u8_ixn_F.js.map +0 -1
- package/dist/state-DpAxYyED.js +0 -1
- package/dist/suite-Cw_7G2RM.js.map +0 -1
- package/dist/suite-DamAEnNF.js +0 -1
- package/dist/summary-BUKfU-9l.js.map +0 -1
- package/dist/summary-Baa4rTHS.js +0 -192
- package/dist/summary-Baa4rTHS.js.map +0 -1
- package/dist/text-gEcKjAyp.js.map +0 -1
- package/dist/tools-DLHgetpx.js +0 -249
- package/dist/tools-DLHgetpx.js.map +0 -1
- package/dist/views-B8w2J-gD.js.map +0 -1
- package/dist/views-DblnI3kA.js +0 -2
- package/dist/workspace-C4RA3srV.js.map +0 -1
- package/dist/workspace-C_X0tdv6.js +0 -1
package/dist/view/index.d.ts
CHANGED
|
@@ -1,9 +1,19 @@
|
|
|
1
|
-
import { $t as
|
|
1
|
+
import { $n as cellSamples, $t as fromModel, An as FamilyRate, Bn as DIRECTIONAL_MIN, Cn as messageLine, Dn as CellRecord, En as AttemptRecord, Fn as JudgementRecord, Ft as TokenUsage, Gn as EffortSample, Gt as INSIGHTS_FILE, Hn as EFFORT_MIN_RUNS, In as Lift, It as TranscriptEvent, Jn as TOKENS_UNREPORTED, Jt as NO_MODEL, Kn as EffortVerdict, Kt as Job, Ln as RunRecord, Mn as FileRef, Nn as HistoryCell, On as EvalEntry, Pn as HistoryEntry, Qn as cellEffort, Qt as Sentence, Rn as StackSummary, Sn as factSegment, Tn as ArtifactRef, Un as EffortContrast, Vn as EFFORT_ALPHA, Wn as EffortMetric, Wt as EvalInsight, Xn as bothUnderATenth, Xt as Remark, Yn as VERDICT_TEXT, Yt as REMARK_KINDS, Zn as cellContrast, Zt as RunInsights, _n as cancelledOf, _t as Outcome, an as CheckResult, ar as effortSample, bn as effectOf, cn as EvalResult, cr as formatCount, dn as PooledArm, dr as tokenValues, en as Across, er as contrastEffort, fn as QuestionResult, gn as SuiteTally, hn as SuiteResult, in as CheckOutcome, ir as effortOf, jn as FamilyRecord, kn as FamilyLift, kt as SandboxWarning, ln as Interval, lr as formatRatio, mn as ResultOutcome, nn as CellError, nr as effortChangeLabel, on as Effect, or as effortSummary, pn as ResultFile, qn as Spread, qt as MODEL_REMARK_KINDS, rn as CellResult, rr as effortLabel, sn as EffortCounters, sr as effortVerdict, tn as CaseResult, tr as directional, un as JudgeResult, ur as formatSeconds, vn as checkEffect, wn as suiteResult, xn as evalAcross, yn as countedOf, zn as CellEffort, zt as TurnStatus } from "../types-B-9b1MDV.js";
|
|
2
|
+
//#region src/stats.d.ts
|
|
3
|
+
/**
|
|
4
|
+
* Counted attempts a side before a pass-rate lift can be called: below it, every view says
|
|
5
|
+
* `insufficient` (`can't tell yet`, `unclear`), whatever the interval. A product floor that keeps
|
|
6
|
+
* `reeval run`, `reeval query`, the judge-key lift and `reeval analyze` agreeing, not one that
|
|
7
|
+
* makes the stated 95% exact: at 4 a side, a 4/4 against 0/4 split happens by chance about 7% of
|
|
8
|
+
* the time when nothing changed.
|
|
9
|
+
*/
|
|
10
|
+
export declare const EFFECT_MIN_COUNTED = 4;
|
|
11
|
+
//#endregion
|
|
2
12
|
//#region src/view/analysis.d.ts
|
|
3
13
|
/** Bumped when the system prompt or output schema changes: an older analysis is superseded. */
|
|
4
14
|
export declare const ANALYSIS_PROMPT_VERSION = "analysis-v10";
|
|
5
15
|
/** Bumped when the digest's shape or selection rules change. */
|
|
6
|
-
export declare const DIGEST_VERSION =
|
|
16
|
+
export declare const DIGEST_VERSION = 11;
|
|
7
17
|
export type Confidence = 'high' | 'medium' | 'low';
|
|
8
18
|
export type AnalysisClaim = {
|
|
9
19
|
text: string;
|
|
@@ -105,14 +115,14 @@ export type AnalysisDoc = {
|
|
|
105
115
|
/** Each cited fact or evidence id, summarised in one line. */
|
|
106
116
|
cited: Record<string, string>;
|
|
107
117
|
/**
|
|
108
|
-
* Why
|
|
109
|
-
* a
|
|
118
|
+
* Why the analysis failed; its `analysis` is then empty. `reeval analyze` treats a cached one as
|
|
119
|
+
* a miss and calls the model again.
|
|
110
120
|
*/
|
|
111
121
|
error?: string;
|
|
112
122
|
};
|
|
113
123
|
/**
|
|
114
|
-
*
|
|
115
|
-
*
|
|
124
|
+
* A note on one eval and harness of a run, which the studio shows:
|
|
125
|
+
* `.reeval/analysis/<runId>/notes/<eval slug>__<harness slug>.json`.
|
|
116
126
|
*/
|
|
117
127
|
export type AnalysisNote = {
|
|
118
128
|
kind: 'reeval-analysis-note';
|
|
@@ -144,7 +154,7 @@ export type AnalysisEntry = {
|
|
|
144
154
|
export type AnalysisIndex = {
|
|
145
155
|
runId: string;
|
|
146
156
|
entries: AnalysisEntry[];
|
|
147
|
-
/**
|
|
157
|
+
/** The run's notes (`.reeval/analysis/<run>/notes/`), by eval and harness; absent when there are none. */
|
|
148
158
|
notes?: AnalysisNote[];
|
|
149
159
|
};
|
|
150
160
|
export declare const analysisDataPath: (runId: string) => string;
|
|
@@ -174,6 +184,11 @@ export type CompareRow = {
|
|
|
174
184
|
};
|
|
175
185
|
/** Cells in both runs, with each side's lift (variants) and effort. */
|
|
176
186
|
perCell?: CompareCell[];
|
|
187
|
+
/**
|
|
188
|
+
* The sides whose `fail` is only attempts the run's interrupt cancelled, which `reeval show`
|
|
189
|
+
* reads as `cancelled`. Such an eval is never `regressed` or `fixed`: it's `drift`.
|
|
190
|
+
*/
|
|
191
|
+
cancelled?: Array<'base' | 'head'>;
|
|
177
192
|
};
|
|
178
193
|
export type CompareCell = {
|
|
179
194
|
key: string;
|
|
@@ -215,6 +230,8 @@ type Comparable = Pick<HistoryEntry, 'stack'> & {
|
|
|
215
230
|
export declare function sourceChanges(before: ComparableEval['sources'], after: ComparableEval['sources']): string[];
|
|
216
231
|
/** Why two runs of an eval aren't comparable: the source files and fingerprint parts that differ. */
|
|
217
232
|
export declare function whyChanged(base: ComparableEval, head: ComparableEval): string[];
|
|
233
|
+
/** An eval only head has. */
|
|
234
|
+
export declare const addedRow: (h: ComparableEval) => CompareRow;
|
|
218
235
|
export declare function compare(base: Comparable, head: Comparable): CompareRow[];
|
|
219
236
|
//#endregion
|
|
220
237
|
//#region src/view/debug.d.ts
|
|
@@ -326,11 +343,6 @@ export declare const familyCases: (family: Pick<FamilyRecord, 'id'>, evals: read
|
|
|
326
343
|
export declare function runHeader(record: RunRecord, recorded: boolean): HistoryEntry;
|
|
327
344
|
//#endregion
|
|
328
345
|
//#region src/view/judgements.d.ts
|
|
329
|
-
/**
|
|
330
|
-
* Fewest counted attempts on each side for a lift that can exclude 0 (2/2 vs 0/2).
|
|
331
|
-
* Same floor as analyse's insufficient lift.
|
|
332
|
-
*/
|
|
333
|
-
export declare const LIFT_MIN_COUNTED = 2;
|
|
334
346
|
/** Short name for a judge question, e.g. `login` from `login?`. */
|
|
335
347
|
export declare function judgeLabel(question: string): string;
|
|
336
348
|
/**
|
|
@@ -532,8 +544,9 @@ export declare const shortSha: (sha: string | null | undefined) => string;
|
|
|
532
544
|
*/
|
|
533
545
|
export declare function activityLabel(a: Pick<AttemptRecord['activity'], 'tools' | 'fileEdits'>): string;
|
|
534
546
|
/**
|
|
535
|
-
* A run's outcome,
|
|
536
|
-
* the run, while failures in ungated cells (controls, by
|
|
547
|
+
* A run's outcome, which `reeval run`'s exit code is read from: an error outside any attempt, a
|
|
548
|
+
* failed gated cell or a budget stop fails the run, while failures in ungated cells (controls, by
|
|
549
|
+
* default) don't. An interrupted run is `skipped`, whatever ran.
|
|
537
550
|
*/
|
|
538
551
|
export declare function runOutcome(run: {
|
|
539
552
|
status: RunRecord['status'];
|
|
@@ -543,6 +556,7 @@ export declare function runOutcome(run: {
|
|
|
543
556
|
families?: ReadonlyArray<{
|
|
544
557
|
outcome: 'pass' | 'fail';
|
|
545
558
|
}>;
|
|
559
|
+
errors?: readonly string[];
|
|
546
560
|
}): Outcome;
|
|
547
561
|
/** Distinct judge labels that failed on a counted attempt, in first-seen order. */
|
|
548
562
|
export declare function failedJudgeLabels(record: Pick<RunRecord, 'evals'>): string[];
|
|
@@ -832,254 +846,8 @@ export declare function readableRecord(record: RunRecord, what?: string): RunRec
|
|
|
832
846
|
export declare const readableTranscriptLine: (line: TranscriptLine) => TranscriptLine;
|
|
833
847
|
/** An analysis document as every reader sees it; one from a newer reeval throws. */
|
|
834
848
|
export declare const readableAnalysis: (doc: AnalysisDoc, what?: string) => AnalysisDoc;
|
|
849
|
+
/** A run's insights as every reader sees them; a file from a newer reeval throws. */
|
|
850
|
+
export declare const readableInsights: (doc: RunInsights, what?: string) => RunInsights;
|
|
835
851
|
//#endregion
|
|
836
|
-
|
|
837
|
-
/**
|
|
838
|
-
* A variant against its baseline: `better` or `worse` when the stored 95% interval on the
|
|
839
|
-
* difference excludes 0, `same` when both pass at the same rate, and `unclear` otherwise,
|
|
840
|
-
* including whenever either side has fewer than `EFFECT_MIN_COUNTED` counted attempts.
|
|
841
|
-
*/
|
|
842
|
-
export type Effect = 'better' | 'worse' | 'same' | 'unclear';
|
|
843
|
-
/**
|
|
844
|
-
* What an eval's cells compare: `change` when it has variants (or several environments),
|
|
845
|
-
* `harness` when it has several harnesses, `both`, or `none`.
|
|
846
|
-
*/
|
|
847
|
-
export type Across = 'none' | 'change' | 'harness' | 'both';
|
|
848
|
-
/** The fewest counted attempts a side for an effect to be anything but `unclear`. */
|
|
849
|
-
export declare const EFFECT_MIN_COUNTED = 4;
|
|
850
|
-
/**
|
|
851
|
-
* One finished run's facts. Every node has a `path`, a stable id to cite it by, and `paths`
|
|
852
|
-
* lists them all in the order the nodes appear.
|
|
853
|
-
*
|
|
854
|
-
* A path is `/`-separated segments, each with `%`, `/` and `+` percent-encoded:
|
|
855
|
-
*
|
|
856
|
-
* - `<eval>`: the eval's id (`evals/a.eval.ts › group › name`) or a family's
|
|
857
|
-
* (`evals/a.eval.ts › fixes $id`);
|
|
858
|
-
* - `<eval>/<case>`: a family case, by its label (`GH-101` for `fixes GH-101`);
|
|
859
|
-
* - `<eval>[/<case>]/<harness>+<environment>+<sandbox>`: a cell;
|
|
860
|
-
* - `<cell>/checks/<label>`: a check, by its assertion label;
|
|
861
|
-
* - `<cell>/effort`: the cell's effort, only when it was tested;
|
|
862
|
-
* - `<eval>/<harness>+<environment>+<sandbox>`: on a family, a variant's pooled effect.
|
|
863
|
-
*
|
|
864
|
-
* A case segment never has a raw `+` and a cell segment always has two, so the two can't be
|
|
865
|
-
* confused. Ids, labels and cell keys are what runs are matched on, so a path names the same
|
|
866
|
-
* fact in every run that has it.
|
|
867
|
-
*/
|
|
868
|
-
export type SuiteResult = {
|
|
869
|
-
runId: string;
|
|
870
|
-
status: RunRecord['status'];
|
|
871
|
-
/**
|
|
872
|
-
* The run's outcome, as `reeval run`'s exit code has it. An error thrown outside any attempt
|
|
873
|
-
* isn't in the record, so `reeval run` exits 1 on one whatever this says.
|
|
874
|
-
*/
|
|
875
|
-
outcome: Outcome;
|
|
876
|
-
/**
|
|
877
|
-
* Outcomes counted per block: each family case is one, each eval with no cases is one, and a
|
|
878
|
-
* family whose pooled gate failed is one more fail.
|
|
879
|
-
*/
|
|
880
|
-
tally: {
|
|
881
|
-
pass: number;
|
|
882
|
-
fail: number;
|
|
883
|
-
infraError: number;
|
|
884
|
-
skipped: number;
|
|
885
|
-
};
|
|
886
|
-
/** Sorted by path. */
|
|
887
|
-
files: ResultFile[];
|
|
888
|
-
/** Every node's `path`, in the order they appear. */
|
|
889
|
-
paths: string[];
|
|
890
|
-
};
|
|
891
|
-
export type ResultFile = {
|
|
892
|
-
path: string;
|
|
893
|
-
evals: EvalResult[];
|
|
894
|
-
};
|
|
895
|
-
export type EvalResult = {
|
|
896
|
-
path: string;
|
|
897
|
-
/** The eval's id, or the family's. */
|
|
898
|
-
id: string;
|
|
899
|
-
group: string[];
|
|
900
|
-
/** The eval's name; on a family, its template. */
|
|
901
|
-
name: string;
|
|
902
|
-
/** A family's name template (`fixes $id`); null on a single eval. */
|
|
903
|
-
template: string | null;
|
|
904
|
-
/** Source order in the file. */
|
|
905
|
-
order: number;
|
|
906
|
-
across: Across;
|
|
907
|
-
/**
|
|
908
|
-
* A pooled `evaluate.each` with two or more cases: `cases` is set and `cells` is null. An
|
|
909
|
-
* unpooled one (`{ pool: false }`) is recorded as separate evals, each a single eval.
|
|
910
|
-
*/
|
|
911
|
-
family: boolean;
|
|
912
|
-
/** A placeholder (`evaluate.todo`) that never ran. */
|
|
913
|
-
todo: boolean;
|
|
914
|
-
/** On a family: `fail` when a case failed or its pooled gate did. */
|
|
915
|
-
outcome: Outcome;
|
|
916
|
-
cells: CellResult[] | null;
|
|
917
|
-
cases: CaseResult[] | null;
|
|
918
|
-
/** A family's variants against their baseline, pooled over the cases; null on a single eval. */
|
|
919
|
-
pooled: PooledArm[] | null;
|
|
920
|
-
/** Its compare questions' answers; on a family, the first case's. */
|
|
921
|
-
questions: QuestionResult[];
|
|
922
|
-
};
|
|
923
|
-
/** A compare question: each side's wins over the pairs the model decided. Never gates. */
|
|
924
|
-
export type QuestionResult = {
|
|
925
|
-
question: string;
|
|
926
|
-
/** Pairs voted, each counted once per order it was asked in. */
|
|
927
|
-
votes: number;
|
|
928
|
-
rates: Array<{
|
|
929
|
-
cut: 'within' | 'across';
|
|
930
|
-
name: string;
|
|
931
|
-
wins: number;
|
|
932
|
-
n: number;
|
|
933
|
-
}>;
|
|
934
|
-
/** Why some or all of its votes are missing: the model call failed. */
|
|
935
|
-
error: string | null;
|
|
936
|
-
};
|
|
937
|
-
export type CaseResult = {
|
|
938
|
-
path: string;
|
|
939
|
-
/** The case's own eval id. */
|
|
940
|
-
id: string;
|
|
941
|
-
/** Its name less the template's literal text: `GH-101` for `fixes GH-101`. */
|
|
942
|
-
label: string;
|
|
943
|
-
name: string;
|
|
944
|
-
/** Its place in the case list. */
|
|
945
|
-
order: number;
|
|
946
|
-
outcome: Outcome;
|
|
947
|
-
cells: CellResult[];
|
|
948
|
-
};
|
|
949
|
-
export type CellResult = {
|
|
950
|
-
path: string;
|
|
951
|
-
key: string;
|
|
952
|
-
harness: string;
|
|
953
|
-
/** `-` when no environment is set. */
|
|
954
|
-
environment: string;
|
|
955
|
-
sandbox: string;
|
|
956
|
-
/** The variant's name; null on a control. */
|
|
957
|
-
variant: string | null;
|
|
958
|
-
/** The control of an eval that has variants: the arm the variants are measured against. */
|
|
959
|
-
baseline: boolean;
|
|
960
|
-
/** On a variant, its baseline cell's key; else null. */
|
|
961
|
-
vs: string | null;
|
|
962
|
-
/** The cell's place in the eval's cells. */
|
|
963
|
-
order: number;
|
|
964
|
-
/** Its failures fail the eval (its gate isn't `'none'`). */
|
|
965
|
-
gated: boolean;
|
|
966
|
-
outcome: Outcome;
|
|
967
|
-
/** Every attempt: `passed + failed + infraError + skipped`. */
|
|
968
|
-
attempts: number;
|
|
969
|
-
passed: number;
|
|
970
|
-
failed: number;
|
|
971
|
-
/** Infrastructure errors, which are neither passes nor fails. */
|
|
972
|
-
infraError: number;
|
|
973
|
-
skipped: number;
|
|
974
|
-
fingerprint: string;
|
|
975
|
-
/** The skill sources and files the cell's environment placed, as written. */
|
|
976
|
-
context: string[];
|
|
977
|
-
checks: CheckResult[];
|
|
978
|
-
/** On a variant, against its baseline; null on a control. */
|
|
979
|
-
effect: Effect | null;
|
|
980
|
-
/** The stored 95% interval on the variant's difference from its baseline. Not for display. */
|
|
981
|
-
interval: Interval | null;
|
|
982
|
-
/** On a variant, the compare's `minLift`, a fraction like the interval; null without one. */
|
|
983
|
-
minLift: number | null;
|
|
984
|
-
/** Medians over the counted attempts; null below `EFFECT_MIN_COUNTED` of them. */
|
|
985
|
-
effort: EffortCounters | null;
|
|
986
|
-
/** On a variant, its tool calls, agent time or tokens differ from its baseline's by the rank test. */
|
|
987
|
-
effortDiffers: boolean;
|
|
988
|
-
/** Why attempts failed or errored where no check says: a timeout, an infra error, a throw. */
|
|
989
|
-
errors: CellError[];
|
|
990
|
-
/** Web searches and fetches per counted attempt; null where the harness can't report them. */
|
|
991
|
-
searches: Array<number | null>;
|
|
992
|
-
fetches: Array<number | null>;
|
|
993
|
-
};
|
|
994
|
-
export type CheckResult = {
|
|
995
|
-
path: string;
|
|
996
|
-
label: string;
|
|
997
|
-
/** First-seen order across the eval's attempts. */
|
|
998
|
-
order: number;
|
|
999
|
-
outcome: CheckOutcome;
|
|
1000
|
-
/**
|
|
1001
|
-
* On a variant, this check against the baseline's (`unclear` when the baseline never ran it);
|
|
1002
|
-
* null on a control or without one.
|
|
1003
|
-
*/
|
|
1004
|
-
effect: Effect | null;
|
|
1005
|
-
/** Set when the check is a judge verdict. */
|
|
1006
|
-
judge: JudgeResult | null;
|
|
1007
|
-
};
|
|
1008
|
-
/**
|
|
1009
|
-
* `not run`: no counted attempt reached the check, because a hard `expect` before it failed.
|
|
1010
|
-
* That is not a failure, and is never `0 of N`. `ran`: counts over the counted attempts that
|
|
1011
|
-
* reached it (`errored` ones couldn't answer), with `notRun` for those that didn't, and the
|
|
1012
|
-
* first failure's message.
|
|
1013
|
-
*/
|
|
1014
|
-
export type CheckOutcome = {
|
|
1015
|
-
status: 'not run';
|
|
1016
|
-
} | {
|
|
1017
|
-
status: 'ran';
|
|
1018
|
-
passed: number;
|
|
1019
|
-
failed: number;
|
|
1020
|
-
errored: number;
|
|
1021
|
-
notRun: number;
|
|
1022
|
-
assertion: string | null;
|
|
1023
|
-
};
|
|
1024
|
-
/**
|
|
1025
|
-
* One distinct error over a cell's attempts: its kind (`turn` for a turn the harness ended with
|
|
1026
|
-
* an error), its message's first line, and how often.
|
|
1027
|
-
*/
|
|
1028
|
-
export type CellError = {
|
|
1029
|
-
kind: ErrorKind | 'turn';
|
|
1030
|
-
message: string;
|
|
1031
|
-
attempts: number;
|
|
1032
|
-
};
|
|
1033
|
-
export type JudgeResult = {
|
|
1034
|
-
question: string;
|
|
1035
|
-
threshold: number;
|
|
1036
|
-
/** One per counted attempt that judged it. */
|
|
1037
|
-
scores: number[];
|
|
1038
|
-
};
|
|
1039
|
-
export type Interval = {
|
|
1040
|
-
low: number;
|
|
1041
|
-
high: number;
|
|
1042
|
-
};
|
|
1043
|
-
/** A null counter is one the harness didn't record; a recorded zero stays zero. */
|
|
1044
|
-
export type EffortCounters = {
|
|
1045
|
-
tools: number | null;
|
|
1046
|
-
agentMs: number | null;
|
|
1047
|
-
tokens: number | null;
|
|
1048
|
-
};
|
|
1049
|
-
export type PooledArm = {
|
|
1050
|
-
path: string;
|
|
1051
|
-
/** The variant cell's key, the same in every case. */
|
|
1052
|
-
key: string;
|
|
1053
|
-
vs: string;
|
|
1054
|
-
effect: Effect;
|
|
1055
|
-
/** Cases with counted attempts on both sides. */
|
|
1056
|
-
pairs: number;
|
|
1057
|
-
interval: Interval;
|
|
1058
|
-
/** The compare's `minLift`, a fraction like the interval, which gates its low end; null without one. */
|
|
1059
|
-
minLift: number | null;
|
|
1060
|
-
outcome: 'pass' | 'fail';
|
|
1061
|
-
};
|
|
1062
|
-
type Side = {
|
|
1063
|
-
pass: number;
|
|
1064
|
-
counted: number;
|
|
1065
|
-
};
|
|
1066
|
-
/**
|
|
1067
|
-
* The effect of `variant` against `control`, from the interval on their difference. A side
|
|
1068
|
-
* with fewer than `EFFECT_MIN_COUNTED` counted attempts is `unclear`, whatever the interval.
|
|
1069
|
-
*/
|
|
1070
|
-
export declare function effectOf(variant: Side, control: Side, interval: Interval | null): Effect;
|
|
1071
|
-
/** The check's effect, on the same Newcombe interval the record stores for a cell's lift. */
|
|
1072
|
-
export declare function checkEffect(variant: CheckTally, control: CheckTally): Effect;
|
|
1073
|
-
/** One path segment: `%`, `/` and `+` percent-encoded, so segments split back unambiguously. */
|
|
1074
|
-
export declare const factSegment: (s: string) => string;
|
|
1075
|
-
export declare function evalAcross(cells: readonly Pick<CellRecord, 'harness' | 'environment' | 'role'>[]): Across;
|
|
1076
|
-
/**
|
|
1077
|
-
* An error's one line: its first line with any letter or digit in it, joined with the next one
|
|
1078
|
-
* when it ends with `:`, since `Docker failed:` alone says nothing. A log line's `| ` is dropped.
|
|
1079
|
-
*/
|
|
1080
|
-
export declare function messageLine(text: string): string;
|
|
1081
|
-
/** The facts of a run, from its record. Pure: the same record always gives the same result. */
|
|
1082
|
-
export declare function suiteResult(record: RunRecord): SuiteResult;
|
|
1083
|
-
//#endregion
|
|
1084
|
-
export { type ArtifactRef, type AttemptRecord, CellEffort, type CellRecord, DIRECTIONAL_MIN, EFFORT_ALPHA, EFFORT_MIN_RUNS, EffortContrast, EffortMetric, EffortSample, EffortVerdict, type EvalEntry, type FamilyLift, type FamilyRate, type FamilyRecord, type FileRef, type HistoryCell, type HistoryEntry, type JudgementRecord, type Outcome, type RunRecord, type SandboxWarning, Spread, type StackSummary, TOKENS_UNREPORTED, type TranscriptEvent, VERDICT_TEXT, bothUnderATenth, cellContrast, cellEffort, cellSamples, contrastEffort, directional, effortChangeLabel, effortLabel, effortOf, effortSample, effortSummary, effortVerdict, formatCount, formatRatio, formatSeconds, tokenValues };
|
|
852
|
+
export { Across, type ArtifactRef, type AttemptRecord, CaseResult, CellEffort, CellError, type CellRecord, CellResult, CheckOutcome, CheckResult, DIRECTIONAL_MIN, EFFORT_ALPHA, EFFORT_MIN_RUNS, Effect, EffortContrast, EffortCounters, EffortMetric, EffortSample, EffortVerdict, type EvalEntry, EvalInsight, EvalResult, type FamilyLift, type FamilyRate, type FamilyRecord, type FileRef, type HistoryCell, type HistoryEntry, INSIGHTS_FILE, Interval, Job, JudgeResult, type JudgementRecord, MODEL_REMARK_KINDS, NO_MODEL, type Outcome, PooledArm, QuestionResult, REMARK_KINDS, Remark, ResultFile, ResultOutcome, RunInsights, type RunRecord, type SandboxWarning, Sentence, Spread, type StackSummary, SuiteResult, SuiteTally, TOKENS_UNREPORTED, type TranscriptEvent, VERDICT_TEXT, bothUnderATenth, cancelledOf, cellContrast, cellEffort, cellSamples, checkEffect, contrastEffort, countedOf, directional, effectOf, effortChangeLabel, effortLabel, effortOf, effortSample, effortSummary, effortVerdict, evalAcross, factSegment, formatCount, formatRatio, formatSeconds, fromModel, messageLine, suiteResult, tokenValues };
|
|
1085
853
|
//# sourceMappingURL=index.d.ts.map
|
package/dist/view/index.js
CHANGED
|
@@ -1,9 +1,10 @@
|
|
|
1
1
|
import { a as parseDataPath, c as shardPath, i as manifestPath, l as studioPath, n as artifactPath, o as recordPath, r as artifactType, s as shardOf, t as RUN_ID } from "../paths-BmR-7tx3.js";
|
|
2
2
|
import { a as debugApiPath, c as formatBytes, d as requestLine, f as traceApiPath, i as TRACE_FILE, l as parseDebugLog, n as DEBUG_PHASES, o as debugLogLine, r as REQUESTS_FILE, s as debugTime, t as DEBUG_LOG, u as parseNdjson } from "../debug-D24NOQH1.js";
|
|
3
|
-
import { a as pinnedAnalysis, i as isSuperseded, n as DIGEST_VERSION, o as sortAnalyses, r as analysisDataPath, t as ANALYSIS_PROMPT_VERSION } from "../analysis-
|
|
4
|
-
import { $ as
|
|
5
|
-
import { a as
|
|
6
|
-
import { t as runHeader } from "../header-
|
|
7
|
-
import {
|
|
8
|
-
import {
|
|
9
|
-
|
|
3
|
+
import { a as pinnedAnalysis, i as isSuperseded, n as DIGEST_VERSION, o as sortAnalyses, r as analysisDataPath, t as ANALYSIS_PROMPT_VERSION } from "../analysis-B0nOfAzA.js";
|
|
4
|
+
import { $ as directional, A as splitPatch, B as labelledJudgements, C as parseTranscript, Ct as sourceChanges, D as runtimeLabel, E as runRows, F as collapsedJudgeLabel, G as EFFORT_MIN_RUNS, H as uniqueJudgeLabels, I as collapsedJudges, J as bothUnderATenth, K as TOKENS_UNREPORTED, L as commandLine, M as cellJudgeRates, N as cellUsage, O as shortSha, P as checkLabels, Q as contrastEffort, R as judgeLabel, S as languageLabel, St as compare, T as runOutcome, Tt as whyChanged, U as DIRECTIONAL_MIN, V as summaryJudgeRates, W as EFFORT_ALPHA, X as cellEffort, Y as cellContrast, Z as cellSamples, _ as harnessLabel, a as cellRows, at as effortVerdict, b as isMatrix, bt as addedRow, c as evalHistory, ct as formatSeconds, d as failedJudgeLabels, et as effortChangeLabel, f as findAttempt, g as formatUsage, h as formatPercent, i as buildIndex, it as effortSummary, j as transcriptItems, k as splitEvalId, l as evalMarkOutcome, lt as tokenValues, m as formatDuration, n as activityLabel, nt as effortOf, o as compareView, ot as formatCount, p as findEval, q as VERDICT_TEXT, r as answerRows, rt as effortSample, s as defaultAttempt, st as formatRatio, t as COMPARE_ORDER, tt as effortLabel, u as evalTree, ut as EFFECT_MIN_COUNTED, v as hasUngatedCells, w as runLead, wt as stackChanges, x as isUngatedMiss, xt as cellsChanged, y as indexFiles, z as judgeRatesLabel } from "../model-UgHDMhI1.js";
|
|
5
|
+
import { a as evalAcross, c as suiteResult, d as familyEffortLabel, f as familyHeadLabel, i as effectOf, l as caseLabel, m as familyRateLabel, n as checkEffect, o as factSegment, p as familyLiftLabel, r as countedOf, s as messageLine, t as cancelledOf, u as familyCases } from "../suite-DmuF_ApO.js";
|
|
6
|
+
import { t as runHeader } from "../header-BVzbRnNv.js";
|
|
7
|
+
import { a as fromModel, i as REMARK_KINDS, n as MODEL_REMARK_KINDS, r as NO_MODEL, t as INSIGHTS_FILE } from "../insights-B2Xmhz9o.js";
|
|
8
|
+
import { c as readableRecord, l as readableTranscriptLine, r as readableAnalysis, s as readableInsights, t as NewerSchemaError } from "../record-kk9M1hj-.js";
|
|
9
|
+
import { a as liveApiPath, c as readableLiveProgress, i as isFinished, n as LIVE_FILE, o as liveCounts, r as LIVE_SCHEMA_VERSION, s as liveTranscriptPath, t as FINISHED } from "../live-DYd2QEyR.js";
|
|
10
|
+
export { ANALYSIS_PROMPT_VERSION, COMPARE_ORDER, DEBUG_LOG, DEBUG_PHASES, DIGEST_VERSION, DIRECTIONAL_MIN, EFFECT_MIN_COUNTED, EFFORT_ALPHA, EFFORT_MIN_RUNS, FINISHED, INSIGHTS_FILE, LIVE_FILE, LIVE_SCHEMA_VERSION, MODEL_REMARK_KINDS, NO_MODEL, NewerSchemaError, REMARK_KINDS, REQUESTS_FILE, RUN_ID, TOKENS_UNREPORTED, TRACE_FILE, VERDICT_TEXT, activityLabel, addedRow, analysisDataPath, answerRows, artifactPath, artifactType, bothUnderATenth, buildIndex, cancelledOf, caseLabel, cellContrast, cellEffort, cellJudgeRates, cellRows, cellSamples, cellUsage, cellsChanged, checkEffect, checkLabels, collapsedJudgeLabel, collapsedJudges, commandLine, compare, compareView, contrastEffort, countedOf, debugApiPath, debugLogLine, debugTime, defaultAttempt, directional, effectOf, effortChangeLabel, effortLabel, effortOf, effortSample, effortSummary, effortVerdict, evalAcross, evalHistory, evalMarkOutcome, evalTree, factSegment, failedJudgeLabels, familyCases, familyEffortLabel, familyHeadLabel, familyLiftLabel, familyRateLabel, findAttempt, findEval, formatBytes, formatCount, formatDuration, formatPercent, formatRatio, formatSeconds, formatUsage, fromModel, harnessLabel, hasUngatedCells, indexFiles, isFinished, isMatrix, isSuperseded, isUngatedMiss, judgeLabel, judgeRatesLabel, labelledJudgements, languageLabel, liveApiPath, liveCounts, liveTranscriptPath, manifestPath, messageLine, parseDataPath, parseDebugLog, parseNdjson, parseTranscript, pinnedAnalysis, readableAnalysis, readableInsights, readableLiveProgress, readableRecord, readableTranscriptLine, recordPath, requestLine, runHeader, runLead, runOutcome, runRows, runtimeLabel, shardOf, shardPath, shortSha, sortAnalyses, sourceChanges, splitEvalId, splitPatch, stackChanges, studioPath, suiteResult, summaryJudgeRates, tokenValues, traceApiPath, transcriptItems, uniqueJudgeLabels, whyChanged };
|
|
@@ -1,8 +1,8 @@
|
|
|
1
|
-
import {
|
|
2
|
-
import {
|
|
3
|
-
import {
|
|
4
|
-
import { o as warnings } from "./state-
|
|
5
|
-
import {
|
|
1
|
+
import { i as formatDuration } from "./text-CrVy5ioL.js";
|
|
2
|
+
import { xt as cellsChanged } from "./model-UgHDMhI1.js";
|
|
3
|
+
import { c as suiteResult } from "./suite-DmuF_ApO.js";
|
|
4
|
+
import { o as warnings } from "./state-C0i_ujQZ.js";
|
|
5
|
+
import { i as stateRows, p as render, r as runTally, t as OUTCOME_MARK } from "./summary-BFMHMTz2.js";
|
|
6
6
|
//#region src/history/views.ts
|
|
7
7
|
var RefError = class extends Error {
|
|
8
8
|
name = "RefError";
|
|
@@ -101,96 +101,6 @@ function formatHistory({ runs, rows, orphaned }) {
|
|
|
101
101
|
if (orphaned.length > 0) lines.push("", `${orphaned.length} run${orphaned.length === 1 ? " was" : "s were"} recorded on commits that aren't in HEAD's history (rebased or dropped?):`, ...orphaned.map((id) => ` ${id}`));
|
|
102
102
|
return lines.join("\n");
|
|
103
103
|
}
|
|
104
|
-
const STATUS_LABEL = {
|
|
105
|
-
regressed: "regressed",
|
|
106
|
-
fixed: "fixed",
|
|
107
|
-
same: "same",
|
|
108
|
-
drift: "changed (not comparable)",
|
|
109
|
-
added: "new",
|
|
110
|
-
removed: "removed",
|
|
111
|
-
infra: "infra error"
|
|
112
|
-
};
|
|
113
|
-
/**
|
|
114
|
-
* `: evals/helpers.ts changed, harness config changed` for a row that isn't comparable, and
|
|
115
|
-
* ` · cells removed: …` when the runs ran different cells of the eval.
|
|
116
|
-
*/
|
|
117
|
-
const why = (r) => [
|
|
118
|
-
r.changes?.length ? `: ${r.changes.join(", ")}` : "",
|
|
119
|
-
r.cells?.added.length ? ` · cells added: ${r.cells.added.join(", ")}` : "",
|
|
120
|
-
r.cells?.removed.length ? ` · cells removed: ${r.cells.removed.join(", ")}` : ""
|
|
121
|
-
].join("");
|
|
122
|
-
const ORDER = [
|
|
123
|
-
"regressed",
|
|
124
|
-
"fixed",
|
|
125
|
-
"drift",
|
|
126
|
-
"infra",
|
|
127
|
-
"added",
|
|
128
|
-
"removed",
|
|
129
|
-
"same"
|
|
130
|
-
];
|
|
131
|
-
const describeRun = (e, reach) => `${e.runId} (${shortSha(e)}${e.recorded ? "" : ", not recorded"}${unreachable(e, reach) ? ", commit not in HEAD's history" : ""})`;
|
|
132
|
-
function formatCompare(base, head, rows, md, reach = null) {
|
|
133
|
-
const sorted = [...rows].sort((a, b) => ORDER.indexOf(a.status) - ORDER.indexOf(b.status));
|
|
134
|
-
const mark = (o) => o ? OUTCOME_MARK[o] : "·";
|
|
135
|
-
const count = (s) => rows.filter((r) => r.status === s).length;
|
|
136
|
-
const tally = `${count("regressed")} regressed · ${count("fixed")} fixed · ${count("drift")} not comparable · ${count("same")} same`;
|
|
137
|
-
const changes = stackChanges(base, head);
|
|
138
|
-
const changeText = changes.map((c) => `${c.label} ${c.base} → ${c.head}`).join("; ");
|
|
139
|
-
const esc = (s) => s.replace(/\|/g, "\\|");
|
|
140
|
-
const cells = sorted.flatMap((r) => (r.perCell ?? []).map((c) => ({
|
|
141
|
-
row: r,
|
|
142
|
-
cell: c
|
|
143
|
-
})));
|
|
144
|
-
if (md) return [
|
|
145
|
-
`### reeval: ${head.runId} vs ${base.runId}`,
|
|
146
|
-
"",
|
|
147
|
-
`Base ${describeRun(base, reach)} → head ${describeRun(head, reach)}. ${tally}.`,
|
|
148
|
-
"",
|
|
149
|
-
...changes.length ? [`**Stack changed:** ${changeText}.`, ""] : [],
|
|
150
|
-
"| Eval | Base | Head | Change |",
|
|
151
|
-
"|---|---|---|---|",
|
|
152
|
-
...sorted.map((r) => `| ${esc(r.id)} | ${mark(r.base)} | ${mark(r.head)} | ${STATUS_LABEL[r.status]}${why(r)} |`),
|
|
153
|
-
"",
|
|
154
|
-
...cells.length ? [
|
|
155
|
-
"| Eval | Cell | Lift, base → head | Effort, base → head |",
|
|
156
|
-
"|---|---|---|---|",
|
|
157
|
-
...cells.map(({ row, cell }) => `| ${esc(row.id)} | ${esc(cell.key)} | ${liftChange(cell)} | ${effortChange(cell)} |`),
|
|
158
|
-
""
|
|
159
|
-
] : []
|
|
160
|
-
].join("\n");
|
|
161
|
-
return [
|
|
162
|
-
`base ${describeRun(base, reach)} → head ${describeRun(head, reach)}`,
|
|
163
|
-
tally,
|
|
164
|
-
...changes.length ? [`stack changed: ${changeText}`] : [],
|
|
165
|
-
"",
|
|
166
|
-
...sorted.flatMap((r) => [` ${mark(r.base)} → ${mark(r.head)} ${STATUS_LABEL[r.status].padEnd(24)} ${r.id}${why(r)}`, ...(r.perCell ?? []).map((c) => ` ${c.key} ${[c.base.lift || c.head.lift ? `lift ${liftChange(c)}` : "", effortChange(c)].filter(Boolean).join(" · ")}`)])
|
|
167
|
-
].join("\n");
|
|
168
|
-
}
|
|
169
|
-
/** `reeval compare` when head has no earlier recorded run: every eval is new. */
|
|
170
|
-
function formatNoBaseline(head, rows, md) {
|
|
171
|
-
const failing = rows.filter((r) => r.head === "fail").length;
|
|
172
|
-
const note = `No earlier recorded run to compare ${head.runId} with: every eval is new${failing ? `, and ${failing} fail${failing === 1 ? "s" : ""}` : ""}.`;
|
|
173
|
-
const mark = (o) => o ? OUTCOME_MARK[o] : "·";
|
|
174
|
-
if (md) return [
|
|
175
|
-
`### reeval: ${head.runId}`,
|
|
176
|
-
"",
|
|
177
|
-
note,
|
|
178
|
-
"",
|
|
179
|
-
"| Eval | Head |",
|
|
180
|
-
"|---|---|",
|
|
181
|
-
...rows.map((r) => `| ${r.id.replace(/\|/g, "\\|")} | ${mark(r.head)} |`),
|
|
182
|
-
""
|
|
183
|
-
].join("\n");
|
|
184
|
-
return [
|
|
185
|
-
note,
|
|
186
|
-
"",
|
|
187
|
-
...rows.map((r) => ` ${mark(r.head)} new ${r.id}`)
|
|
188
|
-
].join("\n");
|
|
189
|
-
}
|
|
190
|
-
const liftShort = (l) => l ? `${signed(l.points)} (${signed(l.lo)}…${signed(l.hi)})${l.withinNoise ? " noise" : ""}` : "–";
|
|
191
|
-
const signed = (x) => `${x >= 0 ? "+" : "−"}${Math.round(Math.abs(x) * 100)}`;
|
|
192
|
-
const liftChange = (c) => `${liftShort(c.base.lift)} → ${liftShort(c.head.lift)}`;
|
|
193
|
-
const effortChange = (c) => c.base.effort && c.head.effort ? effortChangeLabel(c.base.effort, c.head.effort) : "effort –";
|
|
194
104
|
/** `reeval show`: the run's header and state, then the result blocks `reeval run` ended with. */
|
|
195
105
|
function formatShow(run, opts) {
|
|
196
106
|
const rows = stateRows(run);
|
|
@@ -202,7 +112,7 @@ function formatShow(run, opts) {
|
|
|
202
112
|
...rows.map(([k, v]) => ` ${k.padEnd(width)} ${v}`),
|
|
203
113
|
...warnings(run).map((w) => ` ! ${w}`),
|
|
204
114
|
"",
|
|
205
|
-
...render(suiteResult(run), null, {
|
|
115
|
+
...render(suiteResult(run), opts.insights ?? null, {
|
|
206
116
|
color: false,
|
|
207
117
|
verbose: false,
|
|
208
118
|
hint: opts.hint
|
|
@@ -210,6 +120,6 @@ function formatShow(run, opts) {
|
|
|
210
120
|
].join("\n");
|
|
211
121
|
}
|
|
212
122
|
//#endregion
|
|
213
|
-
export {
|
|
123
|
+
export { pickBaseline as a, history as i, formatHistory as n, resolveRun as o, formatShow as r, unreachable as s, RefError as t };
|
|
214
124
|
|
|
215
|
-
//# sourceMappingURL=views-
|
|
125
|
+
//# sourceMappingURL=views-Dk8MCQKr.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"views-Dk8MCQKr.js","names":[],"sources":["../src/history/views.ts"],"mappings":";;;;;;AAiBA,IAAa,WAAb,cAA8B,MAAM;CAClC,OAAgB;AAClB;;;;;AAMA,SAAgB,WAAW,SAAyB,MAAM,UAAwB;CAChF,IAAI,QAAQ,WAAW,GAAG,MAAM,IAAI,SAAS,8BAA8B;CAC3E,MAAM,OAAO,sBAAsB,KAAK,GAAG;CAC3C,IAAI,MAAM;EACR,MAAM,QAAQ,QAAQ,QAAQ,SAAS,IAAI,OAAO,KAAK,MAAM,CAAC;EAC9D,IAAI,CAAC,OACH,MAAM,IAAI,SACR,SAAS,QAAQ,WAAW,IAAI,kBAAkB,YAAY,QAAQ,OAAO,OAAO,OAAO,IAAI,wCACjG;EAEF,OAAO;CACT;CACA,MAAM,QAAQ,QAAQ,MAAM,MAAM,EAAE,UAAU,GAAG;CACjD,IAAI,OAAO,OAAO;CAClB,MAAM,OAAO,QAAQ,QAAQ,MAAM,EAAE,MAAM,WAAW,GAAG,CAAC;CAC1D,IAAI,KAAK,WAAW,KAAK,KAAK,IAAI,OAAO,KAAK;CAC9C,IAAI,KAAK,SAAS,GAChB,MAAM,IAAI,SAAS,IAAI,IAAI,YAAY,KAAK,OAAO,0BAA0B;CAC/E,IAAI,oBAAoB,KAAK,GAAG,GAAG;EACjC,MAAM,QAAQ,QAAQ,QAAQ,MAAM,EAAE,KAAK,IAAI,WAAW,IAAI,YAAY,CAAC,CAAC;EAC5E,MAAM,SAAS,MAAM,MAAM,SAAS;EACpC,IAAI,QAAQ,OAAO;CACrB;CACA,MAAM,IAAI,SACR,mBAAmB,IAAI,4EACzB;AACF;;;;;;;;AASA,SAAgB,aACd,SACA,MACA,UACqB;CACrB,MAAM,UAAU,QAAQ,QAAQ,MAAM,EAAE,QAAQ,KAAK,SAAS,EAAE,WAAW,UAAU;CACrF,IAAI,YAAY,SAAS,SAAS,GAAG;EACnC,MAAM,OAAO,IAAI,IAAI,SAAS,KAAK,KAAK,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;EACvD,IAAI,OAA4B;EAChC,IAAI,WAAW,OAAO;EACtB,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,IAAI,EAAE,MAAM,KAAK,IAAI,EAAE,IAAI,GAAG,IAAI,KAAA;GACxC,IAAI,MAAM,KAAA,MAAc,IAAI,YAAa,MAAM,YAAY,QAAQ,EAAE,QAAQ,KAAK,QAAS;IACzF,OAAO;IACP,WAAW;GACb;EACF;EACA,IAAI,MAAM,OAAO;CACnB;CACA,OAAO,QAAQ,QAAQ,SAAS,MAAM;AACxC;AAEA,MAAM,YAAY,MAAqB,EAAE,MAAM,EAAE,IAAI,IAAI,MAAM,GAAG,CAAC,IAAI;AAkBvE,MAAa,eAAe,OAAqB,UAC/C,CAAC,CAAC,SAAS,CAAC,CAAC,MAAM,OAAO,CAAC,MAAM,IAAI,MAAM,IAAI,GAAG;AAEpD,SAAgB,QACd,SACA,UAAqE,CAAC,GACtE;CACA,MAAM,OAAO,QAAQ,MAAM,EAAE,QAAQ,SAAS,GAAG;CACjD,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,OAAO,CAAC,GAAG,IAAI,CAAC,CAAC,QAAQ,GAClC,KAAK,MAAM,KAAK,IAAI,OAAO,IAAI,CAAC,IAAI,SAAS,EAAE,EAAE,GAAG,IAAI,KAAK,EAAE,EAAE;CAEnE,MAAM,SAAS,QAAQ,QAAQ,YAAY;CAiB3C,OAAO;EAAE;EAAM,MAhBY,IACxB,QAAQ,OAAO,CAAC,UAAU,GAAG,YAAY,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC,CAC5D,KAAK,OAAO;GACX,MAAM,MAAkB;IAAE;IAAI,UAAU,CAAC;IAAG,OAAO,CAAC;IAAG,MAAM;IAAG,SAAS;GAAE;GAC3E,MAAM,uBAAO,IAAI,IAAoB;GACrC,KAAK,SAAS,KAAK,MAAM;IACvB,MAAM,IAAI,IAAI,MAAM,MAAM,MAAM,EAAE,OAAO,EAAE;IAC3C,IAAI,SAAS,KAAK,GAAG,WAAW,IAAI;IACpC,IAAI,CAAC,GAAG;IACR,IAAI,aAAa,MAAM,IAAI,EAAE,KAAK,GAAG,IAAI,MAAM,KAAK,CAAC;IACrD,IAAI,EAAE,YAAY,UAAU,EAAE,YAAY,QAAQ,IAAI;IACtD,IAAI,EAAE,YAAY,QAAQ,IAAI;GAChC,CAAC;GACD,OAAO;EACT,CAEgB;EAAG,UADJ,KAAK,QAAQ,MAAM,YAAY,GAAG,QAAQ,SAAS,IAAI,CAAC,CAAC,CAAC,KAAK,MAAM,EAAE,KAC5D;CAAE;AAChC;AAEA,SAAgB,cAAc,EAAE,MAAM,MAAM,YAAgD;CAC1F,IAAI,KAAK,WAAW,GAClB,OAAO;CACT,MAAM,QAAQ,KAAK;CACnB,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,MAAM,QAAQ,CACZ,GAAG,KAAK,OAAO,MAAM,KAAK,WAAW,IAAI,KAAK,IAAI,qBAAqB,MAAM,MAAM,IAAI,SAAS,KAAK,EAAE,MAAM,KAAK,MAAM,IAAI,SAAS,IAAI,EAAE,IAC3I,EACF;CACA,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,MAAM,EAAE,SAAS,SAAS,EAAE,MAAM,MAAM,GAAG,CAAC;CAChF,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,IAAI,SACf,KAAK,GAAG,MAAM,GAAG,IAAI,MAAM,SAAS,CAAC,IAAI,MAAM,KAAK,IAAI,aAAa,KAAK,KAAK,CAAC,CAChF,KAAK,EAAE;EACV,MAAM,OAAO,IAAI,UAAU,GAAG,IAAI,KAAK,GAAG,IAAI,YAAY;EAC1D,MAAM,KAAK,KAAK,MAAM,OAAO,KAAK,EAAE,IAAI,KAAK,SAAS,CAAC,EAAE,IAAI,IAAI,IAAI;CACvE;CACA,MAAM,KAAK,IAAI,8EAA8E;CAC7F,IAAI,SAAS,SAAS,GACpB,MAAM,KACJ,IACA,GAAG,SAAS,OAAO,MAAM,SAAS,WAAW,IAAI,SAAS,SAAS,4EACnE,GAAG,SAAS,KAAK,OAAO,KAAK,IAAI,CACnC;CAEF,OAAO,MAAM,KAAK,IAAI;AACxB;;AAGA,SAAgB,WACd,KACA,MACQ;CACR,MAAM,OAAO,UAAU,GAAG;CAC1B,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC;CACrD,OAAO;EACL,OAAO,IAAI,MAAM,IAAI,SAAS,GAAG;EACjC,aAAa,IAAI,UAAU,KAAK,eAAe,IAAI,UAAU,EAAE,gBAAgB,IAAI;EACnF;EACA,GAAG,KAAK,KAAK,CAAC,GAAG,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,IAAI,GAAG;EACpD,GAAG,SAAS,GAAG,CAAC,CAAC,KAAK,MAAM,OAAO,GAAG;EACtC;EACA,GAAG,OAAO,YAAY,GAAG,GAAG,KAAK,YAAY,MAAM;GACjD,OAAO;GACP,SAAS;GACT,MAAM,KAAK;EACb,CAAC;CACH,CAAC,CAAC,KAAK,IAAI;AACb"}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { n as UsageError, t as TimeoutError } from "./errors-B6BSMs03.js";
|
|
2
|
-
import { i as originalStack, t as adapterForFile } from "./adapter-
|
|
2
|
+
import { i as originalStack, t as adapterForFile } from "./adapter-IKKo4sZj.js";
|
|
3
3
|
import { a as useEngine, c as createExpect, f as rawAssertions, m as withAttemptExpect, p as unawaitedError, s as attemptExpectError, u as extendExpect } from "./runtime-PX_oXiCS.js";
|
|
4
4
|
import { n as useSchedulerChannel, t as acquire } from "./client-fcwgwOAD.js";
|
|
5
5
|
import { t as budgetMarker } from "./budget-DBUrH4jX.js";
|
|
@@ -365,4 +365,4 @@ process.on("uncaughtException", (error) => worker.unhandled(error));
|
|
|
365
365
|
//#endregion
|
|
366
366
|
export {};
|
|
367
367
|
|
|
368
|
-
//# sourceMappingURL=worker-entry-
|
|
368
|
+
//# sourceMappingURL=worker-entry-CqrzGPF4.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"worker-entry-
|
|
1
|
+
{"version":3,"file":"worker-entry-CqrzGPF4.js","names":[],"sources":["../src/language-runner/load.ts","../src/engine/collector.ts","../src/engine/run.ts","../src/engine/worker.ts","../src/engine/worker-entry.ts"],"mappings":";;;;;;;;;;;;AAGA,MAAa,kBAAkB,MAAc,SAC3C,eAAe,IAAI,CAAC,CAAC,KAAK,MAAM,IAAI;;;AC2BtC,IAAI;AAEJ,MAAa,uBAA2C,YAAY;AAEpE,SAAgB,cAAc,MAAc,MAAgB,SAA2B;CACrF,QAAQ,CAAC,CAAC,MAAM,KAAK;EAAE,MAAM;EAAS;EAAM;EAAM;EAAS,OAAO,CAAC;EAAG,cAAc;CAAM,CAAC;AAC7F;AAEA,SAAgB,aACd,MACA,SACA,MACM;CACN,QAAQ,CAAC,CAAC,MAAM,KAAK;EACnB,MAAM;EACN,OAAO;EACP;EACA,QAAQ,CAAC;EACT,MAAM,QAAQ;EACd;EACA;CACF,CAAC;AACH;AAEA,SAAS,UAAiB;CACxB,IAAI,CAAC,YACH,MAAM,IAAI,WACR,wGACF;CACF,OAAO,WAAW;AACpB;;AAGA,eAAsB,QACpB,MACA,MAC0B;CAC1B,MAAM,OAAc;EAAE,MAAM;EAAS,MAAM;EAAI,MAAM;EAAO,OAAO,CAAC;EAAG,cAAc;CAAM;CAC3F,aAAa;EAAE;EAAM,SAAS;CAAK;CACnC,IAAI;EACF,MAAM,KAAK;EACX,MAAM,aAAa,IAAI;CACzB,UAAU;EACR,aAAa,KAAA;CACf;CACA,eAAe,MAAM,KAAK,cAAc,KAAK;CAC7C,MAAM,QAAyB,CAAC;CAChC,MAAM,QAAQ,OAAc,WAAqB;EAC/C,KAAK,MAAM,QAAQ,MAAM,OACvB,IAAI,KAAK,SAAS,SAAS,KAAK,MAAM,CAAC,GAAG,QAAQ,KAAK,IAAI,CAAC;OACvD,MAAM,KAAK,OAAO,OAAO,MAAM;GAAE,OAAO,MAAM;GAAQ;EAAO,CAAC,CAAC;CAExE;CACA,KAAK,MAAM,CAAC,CAAC;CACb,OAAO;AACT;AAEA,eAAe,aAAa,OAA6B;CACvD,MAAM,QAAQ;CACd,IAAI,MAAM,SAAS;EACjB,MAAM,QAAQ,MAAM;EACpB,MAAM,UAAU;EAChB,IAAI;GACF,MAAM,MAAM,QAAQ;EACtB,UAAU;GACR,MAAM,UAAU;EAClB;CACF;CACA,KAAK,MAAM,QAAQ,MAAM,OAAO,IAAI,KAAK,SAAS,SAAS,MAAM,aAAa,IAAI;CAClF,MAAM,eAAe,MAAM,MAAM,MAC9B,MAAM,EAAE,SAAS,UAAW,EAAE,SAAS,WAAW,EAAE,YACvD;AACF;;AAGA,SAAS,eAAe,OAAc,UAAmB,cAA6B;CACpF,MAAM,cAAc,gBAAgB,MAAM,SAAS;CACnD,MAAM,WAAW,YAAY,MAAM;CACnC,KAAK,MAAM,KAAK,MAAM,OAAO;EAC3B,MAAM,UAAU,WACZ,EAAE,SAAS,UAAW,EAAE,SAAS,WAAW,EAAE,eAC9C,eAAe,EAAE,SAAS;EAC9B,IAAI,UAAU;GACZ,IAAI,EAAE,SAAS,YAAY,WAAW,EAAE,eAClC;QAAA,EAAE,SAAS,QAAQ,EAAE,OAAO;GAAA,OAC3B,IAAI,EAAE,SAAS,SAAS,CAAC,SAAS,EAAE,OAAO;QAC7C,IAAI,EAAE,SAAS,QAAQ,EAAE,OAAO;EACvC;EACA,IAAI,EAAE,SAAS,SAAS;GACtB,IAAI,EAAE,SAAS,UAAU,EAAE,SAAS,QAAQ,QAAQ,GAAG,EAAE,IAAI;QACxD,eAAe,GAAG,UAAU,OAAO;EAC1C;CACF;AACF;AAEA,SAAS,QAAQ,OAAc,MAA6B;CAC1D,KAAK,MAAM,KAAK,MAAM,OAAO;EAC3B,IAAI,EAAE,SAAS,OAAO;EACtB,EAAE,OAAO;EACT,IAAI,EAAE,SAAS,SAAS,QAAQ,GAAG,IAAI;CACzC;AACF;;;;AC9FA,SAAgB,eAAe,OAAiC;CAC9D,IAAI,iBAAiB,SAAS,OAAO,MAAM,UAAU,UACnD,IAAI;EACF,MAAM,QAAQ,cAAc,MAAM,KAAK;CACzC,QAAQ,CAAC;CAEX,OAAO,aAAa,KAAK;AAC3B;AAEA,MAAM,UAAU;CACd,KAAK;CACL,MAAM;CACN,OAAO;CACP,MAAM;CACN,OAAO;AACT;AAEA,MAAM,iBAAiB,IAAI,kBAA0B;;AAGrD,SAAgB,aAAa,UAA0B;CACrD,MAAM,SAAS,WAAW;CAC1B,KAAK,MAAM,CAAC,QAAQ,WAAW,OAAO,QAAQ,OAAO,GAInD,OAAO,WAAW,GAAG,SACnB,SAAS,IAAI,eAAe,SAAS,GAAG,KAAK,IAAI,GAAG,QAAQ,GAAG,OAAO,GAAG,IAAI,EAAE,GAAG;AAExF;;;;;;AAiBA,eAAsB,SACpB,OACA,QACA,SACe;CACf,MAAM,UAA8B,CAAC;CACrC,MAAM,yBAAS,IAAI,IAAmB;CACtC,KAAK,MAAM,QAAQ,OAAO;EACxB,IAAI,CAAC,OAAO,IAAI,KAAK,KAAK,KAAK,KAAK,SAAS,OAAO;GAClD,QAAQ,SAAS,OAAO,KAAK,OAAO;IAAE,OAAO;IAAW,QAAQ,CAAC;IAAG,YAAY;GAAE,CAAC;GACnF;EACF;EACA,QAAQ,KAAK,QAAQ,MAAM,SAAS,MAAM,CAAC;CAC7C;CACA,QAAQ,QAAQ;CAChB,MAAM,QAAQ,IAAI,OAAO;CACzB,MAAM,QAAQ,WAAW,MAAM;AACjC;AAEA,eAAe,QACb,MACA,EAAE,MAAM,YAAY,QAAQ,YAC5B,QACe;CACf,MAAM,EAAE,OAAO,YAAY;CAC3B,MAAM,WAAW,KAAK,IAAI;CAC1B,MAAM,aACJ,SAAS,OAAO,OAAO;EAAE,OAAO;EAAW,QAAQ,CAAC;EAAG,YAAY,KAAK,IAAI,IAAI;CAAS,CAAC;CAC5F,IAAI,OAAO,SAAS,OAAO,KAAK;CAChC,IAAI;CACJ,IAAI,QAAQ,MAAM;EAChB,MAAM,EAAE,QAAQ,SAAS,QAAQ;EACjC,IAAI;GACF,QAAQ,MAAM,QAAQ,QAAQ;IAAE;IAAM;GAAO,CAAC;EAChD,SAAS,OAAO;GACd,IAAI,OAAO,SAAS,OAAO,KAAK;GAChC,OAAO,KAAK,KAAK;EACnB;CACF;CACA,IAAI,WAAW,aAAa,UAAU,CAAC,GAAG;EACxC,OAAO,QAAQ;EACf,OAAO,KAAK;CACd;CACA,MAAM,YAAY,KAAK,IAAI;CAC3B,SAAS,MAAM,OAAO,SAAS;CAC/B,MAAM,OAAmB;EACvB,MAAM;EACN,MAAM,KAAK;EACX,cAAc,CAAC,GAAG,KAAK,QAAQ,KAAK,IAAI,CAAC,CAAC,KAAK,KAAK;CACtD;CACA,MAAM,MAAM,aAAa,IAAI;CAC7B,MAAM,UAAqB,CAAC;CAC5B,MAAM,UAAU,IAAI,gBAAgB;CACpC,IAAI,WAAW;CACf,IAAI;CACJ,MAAM,OAAO,eAAe,IAAI,aAC9B,kBAAkB,KAAK,eACrB,KAAK,KAAK;EACR;EACA,QAAQ,YAAY,IAAI,CAAC,QAAQ,QAAQ,MAAM,CAAC;EAChD,UAAU,SAAS;GACjB,OAAO;GACP,IAAI,UAAU,SAAS,KAAK,OAAO,IAAI;EACzC;EACA,WAAW,aAAa,SAAS,SAAS,OAAO,QAAQ;EACzD,QAAQ,SAAS,SAAS,MAAM,OAAO,IAAI;EAC3C,iBAAiB,eAAe,OAAO;EACvC,mBAAmB,mBAAmB,GAAG;EACzC,kBAAkB,cAAc,IAAI;CACtC,CAAC,CACH,CACF;CACA,OAAO,IAAI,IAAI;CACf,MAAM,gBAAgB,OAAO,QAAQ;CACrC,KAAK,KAAK,SAAS,OAAO;CAC1B,IAAI;EACF,OAAO,QAAQ,UAAU,YAAY,MAAM,QAAQ,SAAS,OAAO,IAAI;EACvE,OAAO;GAAE,OAAO;GAAU,QAAQ,CAAC;EAAE,CAAC;CACxC,SAAS,OAAO;EACd,OAAO;GAAE,OAAO;GAAU,QAAQ,CAAC,eAAe,KAAK,CAAC;EAAE,CAAC;CAC7D;CAEA,SAAS,OAAO,QAAwC;EACtD,WAAW;EACX,SAAS,OAAO,OAAO;GACrB,GAAG;GACH,YAAY,KAAK,IAAI,IAAI;GACzB,GAAI,OAAO,EAAE,KAAK,IAAI,CAAC;EACzB,CAAC;CACH;CAEA,SAAS,KAAK,OAAgB;EAC5B,SAAS,OAAO,OAAO;GACrB,OAAO;GACP,QAAQ,CAAC,eAAe,KAAK,CAAC;GAC9B,YAAY,KAAK,IAAI,IAAI;EAC3B,CAAC;CACH;AACF;;AAGA,SAAS,YAAY,SAAwB,IAAY,OAAuC;CAC9F,IAAI;CACJ,MAAM,WAAW,IAAI,SAAgB,GAAG,WAAW;EACjD,QAAQ,iBAAiB;GACvB,MAAM,QAAQ,IAAI,aAAa,wBAAwB,KAAK,IAAK,EAAE;GACnE,MAAM,MAAM,KAAK;GACjB,OAAO,KAAK;EACd,GAAG,EAAE;CACP,CAAC;CACD,OAAO,QAAQ,KAAK,CAAC,SAAS,QAAQ,CAAC,CAAC,CAAC,cAAc,aAAa,KAAK,CAAC;AAC5E;;;ACvKA,SAAgB,YAAY,SAA0B;CACpD,MAAM,SAAS,IAAI,gBAAgB;CACnC,IAAI;CACJ,IAAI,UAAyB,QAAQ,QAAQ;CAC7C,MAAM,aACJ,OAAO,MAAM,OAAO,uBAAO,IAAI,MAAM,uBAAuB,GAAG,EAAE,MAAM,aAAa,CAAC,CAAC;CACxF,QAAQ,WAAW,YAAY;EAC7B,IAAI,QAAQ,MAAM,QAAQ,UAAU,KAAK,QAAQ,KAAK,QAAQ,MAAM,QAAQ,MAAM;OAC7E,IAAI,QAAQ,MAAM,SAAS,QAAQ,QAAQ,EAAE;OAC7C,KAAK;CACZ,CAAC;CAED,QAAQ,mBAAmB;EACzB,KAAK;EACL,QAAa,cAAc,QAAQ,KAAK,CAAC,CAAC;CAC5C,CAAC;CAED,MAAM,WAAqB;EACzB,QAAQ,OAAO,OAAO,QAAQ,KAAK;GAAE,GAAG;GAAS;GAAO;EAAG,CAAC;EAC5D,WAAW,OAAO,aAAa,QAAQ,KAAK;GAAE,GAAG;GAAY;GAAO;EAAS,CAAC;EAC9E,QAAQ,OAAO,SAAS,QAAQ,KAAK;GAAE,GAAG;GAAS;GAAO;EAAK,CAAC;EAChE,MAAM,OAAO,IAAI,QAAQ,SACvB,QAAQ,KAAK;GAAE,GAAG;GAAO,GAAI,UAAU,KAAA,IAAY,CAAC,IAAI,EAAE,MAAM;GAAI;GAAI;GAAQ;EAAK,CAAC;EACxF,SAAS,OAAO,WAAW,QAAQ,KAAK;GAAE,GAAG;GAAU;GAAO;EAAO,CAAC;EACtE,OAAO,OAAO,SAAS,QAAQ,KAAK;GAAE,GAAG;GAAQ;GAAO;EAAK,CAAC;CAChE;CAEA,eAAe,KAAK,KAAiB,MAAc,QAAwC;EACzF,UAAU;GAAE;GAAc;GAAe,qBAAqB;GAAK;EAAe,CAAC;EACnF,aAAa,YAAY;EACzB,aAAa,QAAQ;EACrB,IAAI;GACF,MAAM,QAAQ,MAAM,QAAQ,YAAY,eAAe,MAAM,IAAI,OAAO,IAAI,CAAC;GAC7E,QAAQ,KAAK;IACX,GAAG;IACH,OAAO,MAAM,KAAK,EAAE,OAAO,MAAM,QAAQ,MAAM,eAAe;KAC5D;KACA;KACA;KACA;KACA,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;IACxD,EAAE;GACJ,CAAC;GACD,IAAI,QAAQ;IACV,QAAQ,qBAAqB,YAAY,QAAQ,KAAK,OAAqB,CAAC;IAC5E,MAAM,SAAS,OAAO,IAAI,IAAI,MAAM,GAAG;KACrC;KACA,YAAY,IAAI;KAChB,QAAQ,OAAO;KACf;KACA,eAAe,QAAQ,KAAK,EAAE,GAAG,QAAQ,CAAC;IAC5C,CAAC;GACH;EACF,SAAS,OAAO;GACd,QAAQ,KAAK;IAAE,GAAG;IAAc,OAAO,eAAe,KAAK;GAAE,CAAC;EAChE;EACA,QAAQ,KAAK,EAAE,GAAG,OAAO,SAAS,QAAQ,KAAK,CAAC,CAAC;CACnD;CAEA,QAAQ,KAAK;EAAE,GAAG;EAAS,UAAA;CAAmB,CAAC;CAC/C,OAAO,EAAE,YAAY,UAAU,QAAQ,KAAK;EAAE,GAAG;EAAa,OAAO,eAAe,KAAK;CAAE,CAAC,EAAE;AAChG;;;ACtFA,IAAI,CAAC,QAAQ,MAAM;CACjB,QAAQ,MACN,sFACF;CACA,QAAQ,KAAK,CAAC;AAChB;AAEA,QAAQ,GAAG,gBAAgB,CAAC,CAAC;AAC7B,MAAM,SAAS,YAAY;CACzB,OAAO,SAAS,SAAS,QAAQ,OAAO,SAAS,KAAA,GAAW,CAAC,SAAS,OAAO,CAAC;CAC9E,YAAY,aAAa,QAAQ,GAAG,WAAW,QAAQ;CACvD,eAAe,aAAa,QAAQ,GAAG,cAAc,QAAQ;CAC7D,OAAO,SAAS,QAAQ,KAAK,IAAI;AACnC,CAAC;AACD,QAAQ,GAAG,uBAAuB,UAAU,OAAO,UAAU,KAAK,CAAC;AACnE,QAAQ,GAAG,sBAAsB,UAAU,OAAO,UAAU,KAAK,CAAC"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
export { r as prepareWorkspaces, a as removePreparedWorkspaces, o as sameTreeArms, s as workspaceKey } from "./workspace-CLeOybkr.js";
|