reeval 0.10.0 → 0.11.0
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 +62 -39
- package/THIRD_PARTY_NOTICES.md +75 -106
- package/dist/{adapter-DWlpOtRn.js → adapter-Bwy3QxAK.js} +2 -2
- package/dist/{adapter-DWlpOtRn.js.map → adapter-Bwy3QxAK.js.map} +1 -1
- package/dist/{agent-CURAwx_q.js → agent-D483ru7D.js} +5 -5
- package/dist/{agent-CURAwx_q.js.map → agent-D483ru7D.js.map} +1 -1
- package/dist/{agent-CaS_2Sce.js → agent-DpE80WPz.js} +4 -4
- package/dist/{agent-CaS_2Sce.js.map → agent-DpE80WPz.js.map} +1 -1
- package/dist/assemble-D5KGrwnJ.js +1 -0
- package/dist/{assemble-B9Alds-z.js → assemble-D9GWr-HS.js} +8 -8
- package/dist/{assemble-B9Alds-z.js.map → assemble-D9GWr-HS.js.map} +1 -1
- package/dist/bridge/index.mjs +13 -6
- package/dist/{capture-CNb9SuU5.js → capture-BQlGHiMC.js} +5 -5
- package/dist/{capture-CNb9SuU5.js.map → capture-BQlGHiMC.js.map} +1 -1
- package/dist/capture-CYGLpWxR.js +1 -0
- package/dist/{cli-D303Vj5N.js → cli-ClpPT8kn.js} +166 -492
- package/dist/cli-ClpPT8kn.js.map +1 -0
- package/dist/{cli-spec-BPZbQG2p.js → cli-spec-qmEnQ25_.js} +15 -6
- package/dist/cli-spec-qmEnQ25_.js.map +1 -0
- package/dist/cli.js +2 -2
- package/dist/{command-Bge8F7Ko.js → command-Bj39y-Wm.js} +33 -24
- package/dist/command-Bj39y-Wm.js.map +1 -0
- package/dist/{command-CMJtHKRa.js → command-BlM0y8jN.js} +6 -6
- package/dist/{command-CMJtHKRa.js.map → command-BlM0y8jN.js.map} +1 -1
- package/dist/command-C-cVjlny.js +1 -0
- package/dist/{command-BLfT7vnQ.js → command-DbAe1Rzb.js} +10 -11
- package/dist/command-DbAe1Rzb.js.map +1 -0
- package/dist/{command-BT9a37ZR.js → command-TgbzPizb.js} +15 -13
- package/dist/command-TgbzPizb.js.map +1 -0
- package/dist/{commands-1p-bIWhN.js → commands-CthqwkAn.js} +1 -1
- package/dist/{commands-BFx16WB8.js → commands-YUwlyTbr.js} +12 -9
- package/dist/commands-YUwlyTbr.js.map +1 -0
- package/dist/{config-DEgnUiU-.js → config-DrVSEVN0.js} +29 -11
- package/dist/config-DrVSEVN0.js.map +1 -0
- package/dist/contrasts-5zsj32gk.js +101 -0
- package/dist/contrasts-5zsj32gk.js.map +1 -0
- package/dist/db-CKg5NtiW.js +2 -0
- package/dist/{db-CdHizc4_.js → db-DfhipHEF.js} +17 -7
- package/dist/{db-CdHizc4_.js.map → db-DfhipHEF.js.map} +1 -1
- package/dist/{dist-B8hJImYt.js → dist-CuY4ehEk.js} +81 -14
- package/dist/dist-CuY4ehEk.js.map +1 -0
- package/dist/{docker-7b_2SaH7.js → docker-D1HZDS6q.js} +4 -4
- package/dist/{docker-7b_2SaH7.js.map → docker-D1HZDS6q.js.map} +1 -1
- package/dist/{docker-yauK-9X8.js → docker-DNmMPOBb.js} +1 -1
- package/dist/{docker-sZrNrr82.js → docker-DOvjji9I.js} +2 -2
- package/dist/{docker-sZrNrr82.js.map → docker-DOvjji9I.js.map} +1 -1
- package/dist/{doctor-BYNok4UO.js → doctor-D2MbNO8L.js} +14 -15
- package/dist/doctor-D2MbNO8L.js.map +1 -0
- package/dist/{egress-C4PZTSSe.js → egress-2-qzPvGi.js} +12 -13
- package/dist/egress-2-qzPvGi.js.map +1 -0
- package/dist/{environment-DNjRk0nU.js → environment-Cca6a4wB.js} +2 -3
- package/dist/environment-Cca6a4wB.js.map +1 -0
- package/dist/{evaluate-D4zeRGBF.js → evaluate-DKIl54Hb.js} +14 -17
- package/dist/evaluate-DKIl54Hb.js.map +1 -0
- package/dist/{evaluate-C1UVL21l.d.ts → evaluate-DMyaGQEj.d.ts} +2 -2
- package/dist/{files-DQmHE7pb.js → files-BXHZFWmx.js} +2 -2
- package/dist/{files-DQmHE7pb.js.map → files-BXHZFWmx.js.map} +1 -1
- package/dist/{files-BXUdY8tr.js → files-BpUYBwGu.js} +5 -5
- package/dist/{files-BXUdY8tr.js.map → files-BpUYBwGu.js.map} +1 -1
- package/dist/files-Bw8lWEjg.js +1 -0
- package/dist/files-D6J0iSdx.js +1 -0
- package/dist/{git-DfG-YJ1u.js → git-Crl3b87o.js} +2 -2
- package/dist/{git-DfG-YJ1u.js.map → git-Crl3b87o.js.map} +1 -1
- package/dist/{harness-B8QC9u0O.js → harness-CEuntrla.js} +2 -2
- package/dist/{harness-B8QC9u0O.js.map → harness-CEuntrla.js.map} +1 -1
- package/dist/harness.d.ts +1 -1
- package/dist/harness.js +1 -1
- package/dist/{header-hK-Z_Po9.js → header-DHQv_ZrF.js} +2 -2
- package/dist/{header-hK-Z_Po9.js.map → header-DHQv_ZrF.js.map} +1 -1
- package/dist/{images-0On0cztu.js → images-Bi3aGxb9.js} +6 -6
- package/dist/{images-0On0cztu.js.map → images-Bi3aGxb9.js.map} +1 -1
- package/dist/index.d.ts +2 -2
- package/dist/index.js +4 -4
- package/dist/{init-B77Rpcqc.js → init-Cl3f3NBT.js} +4 -4
- package/dist/{init-B77Rpcqc.js.map → init-Cl3f3NBT.js.map} +1 -1
- package/dist/init-fhOXED8F.js +1 -0
- package/dist/{invocations-DbGfBAA_.js → invocations-BYDPe3fV.js} +2 -2
- package/dist/{invocations-DbGfBAA_.js.map → invocations-BYDPe3fV.js.map} +1 -1
- package/dist/judge.d.ts +1 -1
- package/dist/{model-NM3ID0uR.js → model-CXhrziBe.js} +1 -1
- package/dist/{model-h_cDtcRr.js → model-D0IcOQao.js} +5 -4
- package/dist/model-D0IcOQao.js.map +1 -0
- package/dist/{model-oqcDplJG.js → model-DaX4bNLZ.js} +17 -8
- package/dist/model-DaX4bNLZ.js.map +1 -0
- package/dist/model-_bxNrk2h.js +1 -0
- package/dist/{options-Bh2W63dK.js → options-C0d0AzeR.js} +6 -7
- package/dist/options-C0d0AzeR.js.map +1 -0
- package/dist/{patch-Zac5aIRM.js → patch-B1bnfhN-.js} +2 -2
- package/dist/{patch-Zac5aIRM.js.map → patch-B1bnfhN-.js.map} +1 -1
- package/dist/{plan-C9Cy27Ay.js → plan-DgQxp9yy.js} +13 -7
- package/dist/plan-DgQxp9yy.js.map +1 -0
- package/dist/{preflight-D7iFPKeK.js → preflight-BRzD30L5.js} +3 -3
- package/dist/{preflight-D7iFPKeK.js.map → preflight-BRzD30L5.js.map} +1 -1
- package/dist/{preflight-CPWpQeBI.js → preflight-DlQvbySb.js} +1 -1
- package/dist/{prepare-doGxaAGR.js → prepare-B_y-AFMv.js} +6 -6
- package/dist/{prepare-doGxaAGR.js.map → prepare-B_y-AFMv.js.map} +1 -1
- package/dist/{prepare-BGjePFQl.js → prepare-DCKTLHN3.js} +2 -2
- package/dist/{provenance-DmE9uPVt.js → provenance-DQ0AXn6c.js} +3 -3
- package/dist/{provenance-DmE9uPVt.js.map → provenance-DQ0AXn6c.js.map} +1 -1
- package/dist/{providers-gDMko8fl.js → providers-B9yTdIqF.js} +1 -1
- package/dist/{providers-tAOD1NQN.js → providers-DoTq1jHW.js} +4 -4
- package/dist/{providers-tAOD1NQN.js.map → providers-DoTq1jHW.js.map} +1 -1
- package/dist/query-BO-ux7RB.js +139 -0
- package/dist/query-BO-ux7RB.js.map +1 -0
- package/dist/query-CRQfsw0e.js +1 -0
- package/dist/{questions-DjAwMxZ3.js → questions-C2dym2ce.js} +3 -3
- package/dist/{questions-DjAwMxZ3.js.map → questions-C2dym2ce.js.map} +1 -1
- package/dist/{remote-HDRt16ew.js → remote-CPPW86Q9.js} +3 -3
- package/dist/{remote-HDRt16ew.js.map → remote-CPPW86Q9.js.map} +1 -1
- package/dist/{results-DuauQI3Z.js → results-BE2f-1XS.js} +6 -5
- package/dist/results-BE2f-1XS.js.map +1 -0
- package/dist/{resume-BRfa8E_s.js → resume-Ca1c8vix.js} +3 -3
- package/dist/{resume-BRfa8E_s.js.map → resume-Ca1c8vix.js.map} +1 -1
- package/dist/{sandbox-D3A8oUb7.js → sandbox-DLrsvu7U.js} +2 -2
- package/dist/{sandbox-D3A8oUb7.js.map → sandbox-DLrsvu7U.js.map} +1 -1
- package/dist/sandbox.d.ts +1 -1
- package/dist/sandbox.js +1 -1
- package/dist/server-Dp8eOeK5.js +1 -0
- package/dist/{server-XDleKaF_.js → server-TUY5xA_A.js} +6 -6
- package/dist/{server-XDleKaF_.js.map → server-TUY5xA_A.js.map} +1 -1
- package/dist/{shared-DYpIb1BI.js → shared-CA61BuPJ.js} +25 -3
- package/dist/shared-CA61BuPJ.js.map +1 -0
- package/dist/{source-C8xumIWj.js → source-DJGS0a0u.js} +2 -2
- package/dist/{source-C8xumIWj.js.map → source-DJGS0a0u.js.map} +1 -1
- package/dist/source-DdbLGR1M.js +1 -0
- package/dist/{src-C_lPtowT.js → src-BxfCKA3_.js} +1 -1
- package/dist/{src-ofUaYbuY.js → src-CLoOMVWI.js} +1 -1
- package/dist/{src-Du8H85Qq.js → src-CcrQbpG8.js} +1 -1
- package/dist/{src-Dj9PAroc.js → src-DCFtJRp0.js} +1 -1
- package/dist/{src-jzagbuqH.js → src-DSYYPEOj.js} +1 -1
- package/dist/src-DYO3DdnQ.js +1 -0
- package/dist/{src-DTbrY1hi.js → src-WD-zgQYZ.js} +1 -1
- package/dist/{src-DNnG4e4R.js → src-_oI9AePb.js} +1 -1
- package/dist/{src-B7treWa8.js → src-ftu5wPSt.js} +127 -15
- package/dist/src-ftu5wPSt.js.map +1 -0
- package/dist/{src-DdyJEh2v.js → src-qsKCqv7a.js} +1 -1
- package/dist/{state-C__9swoG.js → state-DfpHox6b.js} +3 -3
- package/dist/{state-C__9swoG.js.map → state-DfpHox6b.js.map} +1 -1
- package/dist/state-DpAxYyED.js +1 -0
- package/dist/studio-app/assets/index-CBkhNAv8.js +58 -0
- package/dist/studio-app/index.html +1 -1
- package/dist/suite-Cw_7G2RM.js +429 -0
- package/dist/suite-Cw_7G2RM.js.map +1 -0
- package/dist/suite-DamAEnNF.js +1 -0
- package/dist/summary-BUKfU-9l.js +611 -0
- package/dist/summary-BUKfU-9l.js.map +1 -0
- package/dist/{summary-DgyjUh_l.js → summary-Baa4rTHS.js} +6 -6
- package/dist/{summary-DgyjUh_l.js.map → summary-Baa4rTHS.js.map} +1 -1
- package/dist/testing.d.ts +2 -2
- package/dist/testing.js +2 -2
- package/dist/{tools-GCyNYSc1.js → tools-DLHgetpx.js} +3 -3
- package/dist/{tools-GCyNYSc1.js.map → tools-DLHgetpx.js.map} +1 -1
- package/dist/{types-g17bRAzt.d.ts → types-BYdlwMwo.d.ts} +12 -7
- package/dist/view/index.d.ts +250 -2
- package/dist/view/index.js +4 -4
- package/dist/{views-BmQF4UCI.js → views-B8w2J-gD.js} +16 -26
- package/dist/views-B8w2J-gD.js.map +1 -0
- package/dist/views-DblnI3kA.js +2 -0
- package/dist/{worker-entry-yOchGYYv.js → worker-entry-DNnveLqx.js} +2 -2
- package/dist/{worker-entry-yOchGYYv.js.map → worker-entry-DNnveLqx.js.map} +1 -1
- package/dist/{workspace-CgyHEM0l.js → workspace-C4RA3srV.js} +4 -4
- package/dist/{workspace-CgyHEM0l.js.map → workspace-C4RA3srV.js.map} +1 -1
- package/dist/{workspace-D8sSgpiy.js → workspace-C_X0tdv6.js} +1 -1
- package/package.json +8 -8
- package/dist/assemble-J5MukZFY.js +0 -1
- package/dist/capture-jquXZBKK.js +0 -1
- package/dist/cli-D303Vj5N.js.map +0 -1
- package/dist/cli-spec-BPZbQG2p.js.map +0 -1
- package/dist/command-BLfT7vnQ.js.map +0 -1
- package/dist/command-BT9a37ZR.js.map +0 -1
- package/dist/command-Bge8F7Ko.js.map +0 -1
- package/dist/command-BkFjijN2.js +0 -1
- package/dist/commands-BFx16WB8.js.map +0 -1
- package/dist/config-DEgnUiU-.js.map +0 -1
- package/dist/db-Dm1cRJJP.js +0 -2
- package/dist/dist-B8hJImYt.js.map +0 -1
- package/dist/doctor-BYNok4UO.js.map +0 -1
- package/dist/duration-CEOQww97.js +0 -26
- package/dist/duration-CEOQww97.js.map +0 -1
- package/dist/egress-C4PZTSSe.js.map +0 -1
- package/dist/env-BQqj93gu.js +0 -21
- package/dist/env-BQqj93gu.js.map +0 -1
- package/dist/environment-DNjRk0nU.js.map +0 -1
- package/dist/evaluate-D4zeRGBF.js.map +0 -1
- package/dist/family-D2b7WD10.js +0 -67
- package/dist/family-D2b7WD10.js.map +0 -1
- package/dist/files-BhyltlDB.js +0 -1
- package/dist/files-ZZPqfAo6.js +0 -1
- package/dist/init-BrOWin3W.js +0 -1
- package/dist/model-h_cDtcRr.js.map +0 -1
- package/dist/model-nwWJvaUR.js +0 -1
- package/dist/model-oqcDplJG.js.map +0 -1
- package/dist/options-Bh2W63dK.js.map +0 -1
- package/dist/plan-C9Cy27Ay.js.map +0 -1
- package/dist/process-CSVO_jsc.js +0 -119
- package/dist/process-CSVO_jsc.js.map +0 -1
- package/dist/query-Bes9Kh6Y.js +0 -91
- package/dist/query-Bes9Kh6Y.js.map +0 -1
- package/dist/query-C_G_svLU.js +0 -1
- package/dist/results-DuauQI3Z.js.map +0 -1
- package/dist/server-C7sg1ndo.js +0 -1
- package/dist/shared-DYpIb1BI.js.map +0 -1
- package/dist/source-fGd0SaKP.js +0 -1
- package/dist/src-B7treWa8.js.map +0 -1
- package/dist/src-pJZOrYR9.js +0 -1
- package/dist/state-CgjQo29K.js +0 -1
- package/dist/studio-app/assets/index-DrQc2Yxd.js +0 -58
- package/dist/summary-BLAVLySU.js +0 -385
- package/dist/summary-BLAVLySU.js.map +0 -1
- package/dist/views-BW6oCDUQ.js +0 -2
- package/dist/views-BmQF4UCI.js.map +0 -1
|
@@ -1,8 +1,8 @@
|
|
|
1
|
-
import { a as providerModel } from "./model-
|
|
2
|
-
import { l as writeCached } from "./files-
|
|
3
|
-
import { r as fromRecorded, t as attemptSummary } from "./tools-
|
|
4
|
-
import { n as allRuns } from "./commands-
|
|
5
|
-
import { a as failedAnalysis, i as defaultBase, o as prepareAnalysis, s as promptOf, t as analyseDigest } from "./command-
|
|
1
|
+
import { a as providerModel } from "./model-D0IcOQao.js";
|
|
2
|
+
import { l as writeCached } from "./files-BXHZFWmx.js";
|
|
3
|
+
import { r as fromRecorded, t as attemptSummary } from "./tools-DLHgetpx.js";
|
|
4
|
+
import { n as allRuns } from "./commands-YUwlyTbr.js";
|
|
5
|
+
import { a as failedAnalysis, i as defaultBase, o as prepareAnalysis, s as promptOf, t as analyseDigest } from "./command-Bj39y-Wm.js";
|
|
6
6
|
import { join } from "node:path";
|
|
7
7
|
import { z } from "zod";
|
|
8
8
|
import { generateText, stepCountIs, tool } from "ai";
|
|
@@ -189,4 +189,4 @@ function fakeSummaryModel(p) {
|
|
|
189
189
|
//#endregion
|
|
190
190
|
export { writeRunSummary };
|
|
191
191
|
|
|
192
|
-
//# sourceMappingURL=summary-
|
|
192
|
+
//# sourceMappingURL=summary-Baa4rTHS.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"summary-
|
|
1
|
+
{"version":3,"file":"summary-Baa4rTHS.js","names":[],"sources":["../src/analysis/agent/summary.ts"],"mappings":";;;;;;;;;AA4BA,MAAM,aAAa;AAEnB,MAAa,uBAAuB;;;;;;AAOpC,SAAgB,aAAa,GAAqB,UAAkB;CAClE,IAAI,OAAO;CACX,MAAM,SAAY,UAAoC;EACpD,MAAM,OAAO,KAAK,UAAU,KAAK,CAAC,CAAC;EACnC,IAAI,OAAO,MACT,OAAO,EACL,OACE,8FACJ;EAEF,QAAQ;EACR,OAAO;CACT;CACA,MAAM,OAAO,CAAC,EAAE,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS,MAAO,IAAI,CAAC,CAAC,IAAI,CAAC,CAAE;CACjE,OAAO;EACL,eAAe,KAAK;GAClB,aACE;GACF,aAAa,EAAE,OAAO,EACpB,IAAI,EAAE,OAAO,CAAC,CAAC,SAAS,0CAA+C,EACzE,CAAC;GACD,SAAS,OAAO,EAAE,SAAS;IACzB,MAAM,OAAO,EAAE,QAAQ,MAAM,MAAM,EAAE,OAAO,EAAE;IAC9C,IAAI,MAAM;KACR,MAAM,EAAE,UAAU,WAAW,GAAG,SAAS;KACzC,OAAO,MAAM,IAAI;IACnB;IACA,OAAO,EACL,OAAO,EAAE,OAAO,SAAS,MAAM,MAAM,EAAE,OAAO,EAAE,IAC5C,GAAG,GAAG,8BACN,uBAAuB,GAAG,aAAa,EAAE,OAAO,QAAQ,KAAK,IAAI,KAAK,OAAO,GACnF;GACF;EACF,CAAC;EACD,cAAc,KAAK;GACjB,aACE;GACF,aAAa,EAAE,OAAO,EACpB,IAAI,EAAE,OAAO,CAAC,CAAC,SAAS,gDAAqD,EAC/E,CAAC;GACD,SAAS,OAAO,EAAE,SAAS;IACzB,MAAM,IAAI,EAAE,OAAO,SAAS;IAC5B,MAAM,MAAM,KAAK,KAAK,MAAM,MAAM,EAAE,OAAO,UAAU,EAAE,KAAK;IAC5D,MAAM,SAAS,KAAK,OAAO,MAAM,SAAS,MAAM,EAAE,QAAQ,CAAC,CAAC,MAAM,MAAM,EAAE,QAAQ,GAAG,GAAG;IACxF,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,QACjB,OAAO,EACL,OAAO,cAAc,GAAG,qBAAqB,OAAO,KAAK,EAAE,OAAO,QAAQ,CAAC,CAAC,KAAK,IAAI,EAAE,GACzF;IAEF,MAAM,OAAO,MAAM,aAAa,KAAK,IAAI,KAAK,IAAI,OAAO,KAAK,GAAG,MAAM;IACvE,OAAO,MAAM;KAAE,GAAG,eAAe,IAAI;KAAG;IAAG,CAAC;GAC9C;EACF,CAAC;CACH;AACF;AAIA,MAAM,WAAW,MAAe,OAAO,MAAM,YAAY,MAAM,QAAQ,WAAW;;AAGlF,SAAgB,UAAU,OAA8B;CACtD,MAAM,KAAK,MAAM,QAAQ,MAAM,CAAC,QAAQ,EAAE,MAAM,CAAC;CACjD,MAAM,WAAW,GACd,QAAQ,MAAM,EAAE,SAAS,eAAe,CAAC,CACzC,KAAK,OAAO;EAAE,GAAI,EAAE;EAA6C,UAAU;CAAE,EAAE;CAClF,IAAI,CAAC,GAAG,QAAQ,OAAO;EAAE;EAAU,MAAM;CAAG;CAC5C,OAAO;EACL;EACA,MAAM,uGAAuG,GAC1G,KAAK,MAAM,GAAG,EAAE,KAAK,GAAG,KAAK,UAAU,EAAE,KAAK,EAAE,IAAI,KAAK,UAAU,EAAE,MAAM,GAAG,CAAC,CAC/E,KAAK,IAAI;CACd;AACF;;AAGA,eAAsB,YACpB,GACA,OACA,QACsF;CACtF,MAAM,SAAS,MAAM,aAAa;EAChC,OAAO,MAAM,OAAO,iBAAiB,CAAC,IAAI,MAAM,cAAc,KAAK;EACnE,QAAQ;EACR,QAAQ,SAAS,EAAE,MAAM;EACzB,OAAO,aAAa,GAAG,OAAO,SAAS,iBAAiB,CAAC;EACzD,UAAU,YAAA,CAAyB;EACnC,cAAc,EAAE,iBACd,eAAe,IAAoB,EAAE,YAAY,OAAgB,IAAI,KAAA;EACvE,aAAa;EACb,iBAAiB,OAAO,SAAS;EACjC,YAAY;CACd,CAAC;CACD,OAAO;EACL,OAAO,OAAO,MAAM,SAAS,MAC3B,EAAE,YAAY,KAAK,OAAO;GAAE,MAAM,EAAE;GAAU,OAAO,EAAE;GAAO,QAAQ,EAAE;EAAO,EAAE,CACnF;EACA,OAAO;GACL,aAAa,OAAO,YAAY,eAAe;GAC/C,cAAc,OAAO,YAAY,gBAAgB;EACnD;CACF;AACF;;;;;AAQA,eAAsB,gBAAgB,SAId;CACtB,MAAM,EAAE,QAAQ,OAAO,UAAU;CACjC,MAAM,UAAU,MAAM,QAAQ,MAAM;CACpC,MAAM,OAAO,QAAQ,MAAM,MAAM,EAAE,UAAU,KAAK;CAClD,IAAI,CAAC,MAAM,MAAM,IAAI,MAAM,GAAG,MAAM,uBAAuB;CAC3D,MAAM,WAAW,MAAM,gBACrB,QACA,MACA,MAAM,YAAY,QAAQ,SAAS,IAAI,GACvC,IACF;CACA,MAAM,UAAU,KAAK,IAAI;CACzB,IAAI;CACJ,IAAI;EACF,MAAM,EAAE,OAAO,UAAU,MAAM,YAAY,UAAU,OAAO,MAAM;EAClE,MAAM,MAAM,cAAc,QAAQ,UAAU,OAAO,UAAU,KAAK,CAAC;EACnE,MAAM,QAAQ,GAAkB,MAAe,MAAM,OAAO,OAAO,IAAI;EACvE,IAAI,QAAQ;GACV,aAAa,KAAK,IAAI,MAAM,aAAa,MAAM,WAAW;GAC1D,cAAc,KAAK,IAAI,MAAM,cAAc,MAAM,YAAY;EAC/D;EACA,IAAI,YAAY,KAAK,IAAI,IAAI;CAC/B,SAAS,OAAO;EACd,MAAM,UAAU,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;EACrE,MAAM,MAAM,eAAe,QAAQ,UAAU,OAAO,SAAS,KAAK,IAAI,IAAI,OAAO;CACnF;CACA,MAAM,OAAO,MAAM,YAAY,OAAO,MAAM,SAAS,KAAK,GAAG;CAC7D,OAAO;EAAE;EAAK,KAAK,SAAS;EAAK;CAAK;AACxC;;;;;;AASA,SAAS,iBAAiB,GAAoC;CAC5D,MAAM,QAAQ;EACZ,aAAa;GAAE,OAAO;GAAG,SAAS;GAAG,WAAW;GAAG,YAAY;EAAE;EACjE,cAAc;GAAE,OAAO;GAAG,MAAM;GAAG,WAAW;EAAE;CAClD;CACA,MAAM,UAAU,SAAoB,YAAmC;EACrE;EACA,cAAc;GAAE,SAAS;GAAQ,KAAK,WAAW,SAAS,SAAS;EAAa;EAChF;EACA,UAAU,CAAC;EACX,UAAU,EAAE,SAAS,gBAAgB;CACvC;CACA,OAAO;EACL,sBAAsB;EACtB,UAAU;EACV,SAAS;EACT,eAAe,CAAC;EAChB,YAAY,OAAO,YAAmE;GACpF,MAAM,WAAW,QAAQ,OAAO,MAC7B,MACC,EAAE,SAAS,UACX,MAAM,QAAQ,EAAE,OAAO,KACtB,EAAE,QAAyB,MAAM,MAAM,EAAE,SAAS,aAAa,CACpE;GACA,MAAM,QAAQ,CACZ,GAAG,EAAE,OAAO,QACT,MAAM,GAAG,UAAU,CAAC,CACpB,KAAK,QAAQ;IAAE,MAAM;IAAiB,OAAO,EAAE,GAAG;GAAE,EAAE,GACzD,GAAG,OAAO,QAAQ,EAAE,OAAO,QAAQ,CAAC,CACjC,QAAQ,GAAG,OAAO,EAAE,YAAY,MAAM,CAAC,CACvC,MAAM,GAAG,UAAU,CAAC,CACpB,KAAK,CAAC,SAAS;IAAE,MAAM;IAAgB,OAAO,EAAE,GAAG;GAAE,EAAE,CAC5D;GACA,IAAI,YAAY,CAAC,MAAM,QACrB,OAAO,OACL,CAAC;IAAE,MAAM;IAAQ,MAAM,QAAQ,WAAW,kBAAkB,UAAU;GAAG,CAAC,GAC1E,MACF;GAEF,OAAO,OACL,MAAM,KAAK,GAAG,OAAO;IACnB,MAAM;IACN,YAAY,QAAQ,EAAE,KAAK,GAAG;IAC9B,UAAU,EAAE;IACZ,OAAO,KAAK,UAAU,EAAE,KAAK;GAC/B,EAAE,GACF,YACF;EACF;EACA,UAAU,YAAY;GACpB,MAAM,IAAI,MAAM,yCAAyC;EAC3D;CACF;AACF"}
|
package/dist/testing.d.ts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import {
|
|
2
|
-
import { a as DeclarationTree, n as declareTree, o as DeclaredNode } from "./evaluate-
|
|
1
|
+
import { Bt as Wait, p as CaptureUseAction, u as CaptureInstruction } from "./types-BYdlwMwo.js";
|
|
2
|
+
import { a as DeclarationTree, n as declareTree, o as DeclaredNode } from "./evaluate-DMyaGQEj.js";
|
|
3
3
|
//#region src/capture/fake-instructions.d.ts
|
|
4
4
|
declare const FAKE_INSTRUCTIONS: unique symbol;
|
|
5
5
|
type FakeInstructionStep = CaptureUseAction[] | {
|
package/dist/testing.js
CHANGED
|
@@ -1,3 +1,3 @@
|
|
|
1
|
-
import { t as declareTree } from "./evaluate-
|
|
2
|
-
import { f as fakeInstructions } from "./capture-
|
|
1
|
+
import { t as declareTree } from "./evaluate-DKIl54Hb.js";
|
|
2
|
+
import { f as fakeInstructions } from "./capture-BQlGHiMC.js";
|
|
3
3
|
export { declareTree, fakeInstructions };
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { t as isObjectLike } from "./object-CLf2-tNq.js";
|
|
2
2
|
import { s as recordFileName } from "./ids-Bp6ut7ZD.js";
|
|
3
3
|
import { n as clip } from "./text-gEcKjAyp.js";
|
|
4
|
-
import { C as parseTranscript } from "./model-
|
|
5
|
-
import { s as mergeOutcome } from "./assemble-
|
|
4
|
+
import { C as parseTranscript } from "./model-DaX4bNLZ.js";
|
|
5
|
+
import { s as mergeOutcome } from "./assemble-D9GWr-HS.js";
|
|
6
6
|
import { join } from "node:path";
|
|
7
7
|
import { readFile } from "node:fs/promises";
|
|
8
8
|
//#region src/analysis/agent/tools.ts
|
|
@@ -246,4 +246,4 @@ function divergence(failed, passed, window = 6) {
|
|
|
246
246
|
//#endregion
|
|
247
247
|
export { stagedAttempts as a, listOutcomes as i, divergence as n, fromRecorded as r, attemptSummary as t };
|
|
248
248
|
|
|
249
|
-
//# sourceMappingURL=tools-
|
|
249
|
+
//# sourceMappingURL=tools-DLHgetpx.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"tools-
|
|
1
|
+
{"version":3,"file":"tools-DLHgetpx.js","names":[],"sources":["../src/analysis/agent/tools.ts"],"mappings":";;;;;;;;AA+BA,MAAM,aAAa;CACjB;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;AAGA,SAAS,aAAa,OAAwB;CAC5C,IAAI,OAAO,UAAU,UAAU,OAAO,KAAK,MAAM,QAAQ,QAAQ,GAAG,GAAG,GAAG;CAC1E,IAAI,CAAC,aAAa,KAAK,GAAG,OAAO;CACjC,KAAK,MAAM,OAAO,YAAY;EAC5B,MAAM,QAAQ,MAAM;EACpB,IAAI,OAAO,UAAU,YAAY,OAAO,OAAO,KAAK,MAAM,QAAQ,QAAQ,GAAG,GAAG,GAAG;EACnF,IAAI,MAAM,QAAQ,KAAK,KAAK,MAAM,OAAO,MAAM,OAAO,MAAM,QAAQ,GAClE,OAAO,KAAK,MAAM,KAAK,GAAG,GAAG,GAAG;CACpC;CACA,OAAO,KAAK,KAAK,UAAU,KAAK,GAAG,GAAG;AACxC;;AAGA,SAAgB,OAAO,OAAuC;CAC5D,QAAQ,MAAM,MAAd;EACE,KAAK,QAAQ;GACX,MAAM,QAAQ,aAAa,MAAM,KAAK;GACtC,OAAO,GAAG,MAAM,OAAO,QAAQ,KAAK,UAAU,KAAK,MAAM,KAAK,KAAK;EACrE;EACA,KAAK,aACH,OAAO,SAAS,MAAM,KAAK,KAAK,MAAM;EACxC,KAAK,cACH,OAAO,eAAe,KAAK,MAAM,OAAO,GAAG,IAAI,MAAM,OAAO,QAAQ,cAAc;EACpF,KAAK,aACH,OAAO,cAAc,MAAM,MAAM,MAAM,KAAK,KAAK;EACnD,KAAK,SACH,OAAO,UAAU,KAAK,MAAM,SAAS,GAAG;EAC1C,KAAK,SACH,OAAO,UAAU,KAAK,MAAM,SAAS,GAAG;EAC1C,SACE,OAAO;CACX;AACF;AAEA,MAAM,SAAS,WAA8B,OAAO,SAAS,MAAM,OAAO,CAAC,KAAK,CAAC,CAAC;AAElF,MAAM,QAAQ,MAAc,QAAiB,KAAK,UAAU,MAAM,OAAO,IAAI,KAAK,MAAM,CAAC,GAAG;AAC5F,MAAM,QAAQ,MAAc,QAC1B,KAAK,UAAU,MAAM,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,MAAM,KAAK,SAAS,IAAI;;AAG5E,SAAgB,WAAW,QAAoB,QAA+C;CAC5F,MAAM,UAAU,OAAO;CACvB,IAAI,CAAC,SAAS,OAAO;CACrB,MAAM,EAAE,SAAS,UAAU,aAAa,QAAQ,MAAM;CACtD,OAAO;EACL,IAAI,GAAG,QAAQ,KAAK,IAAI,GAAG,QAAQ;EACnC,MAAM,QAAQ,KAAK;EACnB,GAAG,QAAQ;EACX;EACA,OAAO,QAAQ;GAAE,MAAM,MAAM;GAAM,SAAS,MAAM;EAAQ,IAAI;EAC9D,SAAS,QAAQ,UAAU,CAAC,EAAA,CAAG,KAAK,OAAO;GACzC,SAAS,EAAE;GACX,UAAU,EAAE;GACZ,QAAQ,KAAK,GAAG,EAAE,OAAO,IAAI,EAAE,SAAS,KAAK,GAAG,GAAG;EACrD,EAAE;EACF,aAAa,QAAQ,cAAc,CAAC,EAAA,CAAG,KAAK,OAAO;GACjD,UAAU,EAAE;GACZ,SAAS,EAAE;GACX,OAAO,EAAE;EACX,EAAE;EACF,QAAQ,QAAQ,SAAS,CAAC,EAAA,CAAG,KAAK,OAAO;GACvC,QAAQ,EAAE;GACV,MAAM,EAAE;GACR,QAAQ,EAAE;GACV,OAAO,MAAM,EAAE,UAAU;EAC3B,EAAE;EACF,QAAQ,QAAQ,MAAM,SAAS,CAAC,EAAA,CAAG,KAAK,MAAM,EAAE,IAAI;EACpD,OAAO,QAAQ,MAAM,SAAS;CAChC;AACF;;AAGA,eAAsB,eACpB,YACA,SACwB;CACxB,MAAM,QAAuB,CAAC;CAC9B,KAAK,MAAM,UAAU,SAAS;EAC5B,MAAM,OAAO,KAAK,YAAY,SAAS,eAAe,OAAO,GAAG,CAAC;EACjE,IAAI,SAA4B;EAChC,IAAI;GACF,SAAS,KAAK,MAAM,MAAM,SAAS,MAAM,MAAM,CAAC;EAClD,QAAQ,CAER;EACA,MAAM,OAAO,WAAW,QAAQ,MAAM;EACtC,IAAI,MAAM,MAAM,KAAK,IAAI;CAC3B;CACA,OAAO;AACT;;AAGA,eAAsB,aAAa,QAAgB,SAA8C;CAC/F,MAAM,QAAQ,SACZ,OAAO,SAAS,KAAK,QAAQ,IAAI,GAAG,MAAM,CAAC,CAAC,YAAY,EAAE,IAAI,QAAQ,QAAQ,EAAE;CAClF,MAAM,QAAQ,gBAAgB,MAAM,KAAK,QAAQ,MAAM,YAAY,IAAI,CAAC;CACxE,OAAO;EACL,IAAI,QAAQ;EACZ,MAAM,QAAQ;EACd,GAAG,QAAQ;EACX,SAAS,QAAQ;EACjB,OAAO,QAAQ,QAAQ;GAAE,MAAM,QAAQ,MAAM;GAAM,SAAS,QAAQ,MAAM;EAAQ,IAAI;EACtF,QAAQ,MAAM,QAAQ,IACpB,QAAQ,OAAO,IAAI,OAAO,OAAO;GAC/B,SAAS,EAAE;GACX,UAAU,EAAE;GACZ,QAAQ,MAAM,MAAM,KAAK,EAAE,KAAK,IAAI,EAAA,CAAG,KAAK,GAAG,GAAG;EACpD,EAAE,CACJ;EACA,YAAY,QAAQ,WAAW,KAAK,OAAO;GACzC,UAAU,EAAE;GACZ,SAAS,EAAE;GACX,OAAO,EAAE;EACX,EAAE;EACF,OAAO,QAAQ,MAAM,KAAK,GAAG,OAAO;GAClC,QAAQ,EAAE;GACV,MAAM,EAAE;GACR,QAAQ,EAAE;GACV,OAAO,MACL,MAAM,QACH,MACC,EAAE,SAAS,KAAK,EAAE,SAAS,YAAY,EAAE,SAAS,UACtD,CACF;EACF,EAAE;EACF,OAAO,QAAQ,aAAa,KAAK,MAAM,EAAE,IAAI;EAC7C,OAAO,MAAM,KAAK,QAAQ,MAAM,MAAM,IAAI;CAC5C;AACF;;AAMA,SAAgB,eAAe,UAAuC;CACpE,MAAM,SAAS,SAAS,QAAQ,MAAM,EAAE,YAAY,MAAM;CAC1D,MAAM,QAAQ,OAAoB,GAAgB,MAChD,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC,IAAI,KAAK,IAAI,EAAE,IAAI,EAAE,CAAC;CAC1C,OAAO,SACJ,QAAQ,MAAM,EAAE,YAAY,MAAM,CAAC,CACnC,SAAS,MAAM;EACd,MAAM,WAAW,OAAO,QAAQ,MAAM,EAAE,SAAS,EAAE,IAAI,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC;EACvE,MAAM,QAAQ,OAAO,QAAQ,MAAM,EAAE,SAAS,EAAE,IAAI,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC;EACpE,MAAM,QAAQ,YAAY;EAC1B,OAAO,QACH,CACE;GACE,QAAQ,EAAE;GACV,QAAQ,MAAM;GACd,OAAO,WAAY,SAAoB;EACzC,CACF,IACA,CAAC;CACP,CAAC;AACL;AAEA,MAAM,aAAa,MAAmB,EAAE,MAAM,QAAQ,GAAG,MAAM,IAAI,EAAE,MAAM,QAAQ,CAAC;;AAGpF,SAAgB,aAAa,UAAyB;CACpD,OAAO;EACL,UAAU,SAAS,KAAK,OAAO;GAC7B,IAAI,EAAE;GACN,SAAS,EAAE;GACX,OAAO,EAAE,QAAQ,GAAG,EAAE,MAAM,KAAK,IAAI,KAAK,EAAE,MAAM,SAAS,GAAG,MAAM;GACpE,cAAc,EAAE,OAAO,QAAQ,MAAM,EAAE,aAAa,CAAC,CAAC,CAAC,KAAK,MAAM,EAAE,OAAO;GAC3E,YAAY,EAAE,WAAW,KACtB,MAAM,GAAG,EAAE,QAAQ,GAAG,EAAE,MAAM,QAAQ,CAAC,EAAE,IAAI,KAAK,EAAE,UAAU,GAAG,GACpE;GACA,OAAO,EAAE,MAAM;GACf,OAAO,UAAU,CAAC;GAClB,OAAO,EAAE,MAAM;EACjB,EAAE;EACF,OAAO,eAAe,QAAQ;CAChC;AACF;;AAGA,SAAgB,eAAe,GAAgB;CAC7C,OAAO;EACL,IAAI,EAAE;EACN,SAAS,EAAE;EACX,OAAO,EAAE,QAAQ;GAAE,MAAM,EAAE,MAAM;GAAM,SAAS,KAAK,EAAE,MAAM,SAAS,GAAI;EAAE,IAAI;EAChF,OAAO,EAAE,MAAM,KAAK,GAAG,OAAO;GAC5B,MAAM,IAAI;GACV,QAAQ,KAAK,EAAE,QAAQ,GAAG;GAC1B,QAAQ,EAAE;GACV,OACE,EAAE,MAAM,SAAS,KACb;IAAC,GAAG,EAAE,MAAM,MAAM,GAAG,EAAE;IAAG,KAAK,EAAE,MAAM,SAAS,GAAG;IAAW,GAAG,EAAE,MAAM,MAAM,GAAG;GAAC,IACnF,EAAE;GACR,MAAM,KAAK,EAAE,MAAM,GAAG;EACxB,EAAE;EACF,QAAQ,EAAE,OAAO,KAAK,OAAO;GAC3B,SAAS,EAAE;GACX,UAAU,EAAE;GACZ,GAAI,EAAE,aAAa,IAAI,EAAE,QAAQ,EAAE,OAAO,IAAI,CAAC;EACjD,EAAE;EACF,YAAY,EAAE;EACd,OAAO,EAAE,MAAM,MAAM,GAAG,EAAE;EAC1B,MAAM,KAAK,EAAE,OAAO,GAAI;CAC1B;AACF;;;;;AAMA,SAAgB,WAAW,QAAqB,QAAqB,SAAS,GAAG;CAC/E,MAAM,IAAI,OAAO,MAAM,SAAS,MAAM,EAAE,KAAK;CAC7C,MAAM,IAAI,OAAO,MAAM,SAAS,MAAM,EAAE,KAAK;CAC7C,IAAI,OAAO;CACX,OAAO,OAAO,EAAE,UAAU,OAAO,EAAE,UAAU,EAAE,UAAU,EAAE,OAAO;CAClE,MAAM,WAAW,OAAO,EAAE,UAAU,OAAO,EAAE;CAC7C,MAAM,OAAO,KAAK,IAAI,GAAG,OAAO,CAAC;CACjC,MAAM,WAAW,CAAC,GAAG,IAAI,IAAI,CAAC,GAAG,OAAO,QAAQ,GAAG,OAAO,MAAM,CAAC,CAAC,KAAK,MAAM,EAAE,OAAO,CAAC,CAAC;CACxF,MAAM,QAAQ,GAAgB,YAC5B,EAAE,OAAO,MAAM,MAAM,EAAE,YAAY,OAAO,CAAC,EAAE,YAAY;CAC3D,MAAM,WAAW,GAAgB,aAC/B,EAAE,WAAW,MAAM,MAAM,EAAE,aAAa,QAAQ,CAAC,EAAE,WAAW;CAChE,MAAM,YAAY,CAChB,GAAG,IAAI,IAAI,CAAC,GAAG,OAAO,YAAY,GAAG,OAAO,UAAU,CAAC,CAAC,KAAK,MAAM,EAAE,QAAQ,CAAC,CAChF;CACA,MAAM,QAAQ,MAAmB,KAAK,EAAE,MAAM,GAAG,EAAE,CAAC,EAAE,QAAQ,IAAI,GAAG;CACrE,OAAO;EACL,QAAQ,OAAO;EACf,QAAQ,OAAO;EACf,OAAO,OAAO,QAAQ,GAAG,OAAO,MAAM,KAAK,IAAI,KAAK,OAAO,MAAM,SAAS,GAAG,MAAM;EACnF,eAAe;;EAEf,YAAY,WAAW,OAAO,IAAI;EAClC,OAAO;GAAE,QAAQ,EAAE;GAAQ,QAAQ,EAAE;EAAO;EAC5C,aAAa,EAAE,MAAM,MAAM,OAAO,MAAM;EACxC,aAAa,EAAE,MAAM,MAAM,OAAO,MAAM;EACxC,QAAQ,SACL,QAAQ,MAAM,KAAK,QAAQ,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,CAClD,KAAK,OAAO;GACX,SAAS;GACT,QAAQ,KAAK,QAAQ,CAAC;GACtB,QAAQ,KAAK,QAAQ,CAAC;GACtB,cAAc,OAAO,OAAO,MAAM,MAAM,EAAE,YAAY,CAAC,CAAC,EAAE,UAAU;EACtE,EAAE;EACJ,YAAY,UACT,QAAQ,MAAM,QAAQ,QAAQ,CAAC,MAAM,QAAQ,QAAQ,CAAC,CAAC,CAAC,CACxD,KAAK,OAAO;GACX,UAAU,KAAK,GAAG,GAAG;GACrB,QAAQ,QAAQ,QAAQ,CAAC;GACzB,QAAQ,QAAQ,QAAQ,CAAC;EAC3B,EAAE;EACJ,OAAO;GACL,YAAY,OAAO,MAAM,QAAQ,MAAM,CAAC,OAAO,MAAM,SAAS,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,EAAE;GAC7E,YAAY,OAAO,MAAM,QAAQ,MAAM,CAAC,OAAO,MAAM,SAAS,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,EAAE;EAC/E;EACA,WAAW;GAAE,QAAQ,KAAK,MAAM;GAAG,QAAQ,KAAK,MAAM;EAAE;EACxD,MAAM;GAAE,QAAQ,KAAK,OAAO,OAAO,IAAI;GAAG,QAAQ,KAAK,OAAO,OAAO,IAAI;EAAE;CAC7E;AACF"}
|
|
@@ -95,14 +95,17 @@ declare function cellContrast(e: Pick<EvalEntry, 'attempts' | 'cells'>, cellKey:
|
|
|
95
95
|
/** `4`, `4.5`: medians of counts are whole or halves. */
|
|
96
96
|
declare const formatCount: (n: number) => string;
|
|
97
97
|
declare const formatRatio: (r: number | null) => string;
|
|
98
|
-
/**
|
|
98
|
+
/** Two agent times that both print `<0.1s` have no ratio worth printing. */
|
|
99
|
+
declare const bothUnderATenth: (a: number, b: number) => boolean;
|
|
100
|
+
/** Seconds for agent time: `<0.1s`, `0.4s`, `21s`, `1m 05s`. */
|
|
99
101
|
declare function formatSeconds(ms: number): string;
|
|
100
102
|
declare const VERDICT_TEXT: Record<EffortVerdict, (what: string) => string>;
|
|
101
103
|
/**
|
|
102
104
|
* `tools 5 vs 43 (×0.11), agent 21s vs 46s (×0.47): too few attempts to test` for a variant
|
|
103
105
|
* against its control; with enough attempts, `fewer tool calls (p = 0.029)`. Tokens fold in
|
|
104
106
|
* when both sides have usage samples; they are omitted (not `insufficient`) when either does not.
|
|
105
|
-
* A metric too small to test whose ranges don't overlap says so: `(×0.37, no overlap)`.
|
|
107
|
+
* A metric too small to test whose ranges don't overlap says so: `(×0.37, no overlap)`. Agent
|
|
108
|
+
* times both under a tenth of a second have no ratio: `agent <0.1s vs <0.1s`.
|
|
106
109
|
*/
|
|
107
110
|
declare function effortLabel(c: EffortContrast): string;
|
|
108
111
|
/** One cell's effort on its own: `tools 43 · agent 46s`, plus tokens when any attempt reported them. */
|
|
@@ -557,9 +560,10 @@ type CheckTally = {
|
|
|
557
560
|
notRun: number;
|
|
558
561
|
};
|
|
559
562
|
/**
|
|
560
|
-
* One check over a cell's counted attempts. `verdict` (variant cells)
|
|
561
|
-
* passes on every
|
|
562
|
-
* when it does the same on both sides, so it can't tell them
|
|
563
|
+
* One check over a cell's counted attempts. `verdict` (variant cells), over the attempts that
|
|
564
|
+
* reached the check: `discriminates` when it passes on every one of one side and none of the
|
|
565
|
+
* other; `both-pass` or `both-fail` when it does the same on both sides, so it can't tell them
|
|
566
|
+
* apart; `mixed` otherwise. No verdict when a side never reached it: not run is not a fail.
|
|
563
567
|
*/
|
|
564
568
|
type CellCheck = CheckTally & {
|
|
565
569
|
label: string;
|
|
@@ -2455,6 +2459,7 @@ type CellRole = 'control' | 'variant';
|
|
|
2455
2459
|
type Cell = {
|
|
2456
2460
|
key: string;
|
|
2457
2461
|
harness: string;
|
|
2462
|
+
/** The environment's own name, also on a variant cell (`with-house-greeting`, with no prefix). */
|
|
2458
2463
|
environment: string;
|
|
2459
2464
|
sandbox: string;
|
|
2460
2465
|
role: CellRole;
|
|
@@ -2616,5 +2621,5 @@ type SandboxWarning = {
|
|
|
2616
2621
|
message: string;
|
|
2617
2622
|
};
|
|
2618
2623
|
//#endregion
|
|
2619
|
-
export {
|
|
2620
|
-
//# sourceMappingURL=types-
|
|
2624
|
+
export { GooseHarnessDescriptor as $, JudgementRecord as $t, DirectoryOptions as A, ServeHandle as At, ErrorKind as B, Wait as Bt, CompareCut as C, effortSample as Cn, ProfileDefaults as Ct, ConfigJudgeOptions as D, formatRatio as Dn, SandboxWarning as Dt, ConfigHelpers as E, formatCount as En, SandboxDescriptor as Et, EnvVars as F, Turn as Ft, FakeAnalysisModel as G, CheckTally as Gt, EvalOptions as H, ArtifactRef as Ht, Environment as I, TurnData as It, FakeNoteInput as J, FamilyRate as Jt, FakeHarnessDescriptor as K, EvalEntry as Kt, EnvironmentHelpers as L, TurnStatus as Lt, Duration as M, SkillSource as Mt, EgressPolicy as N, TokenUsage as Nt, CursorHarnessDescriptor as O, formatSeconds as On, Secret as Ot, EgressRecord as P, TranscriptEvent as Pt, FxHarnessDescriptor as Q, HistoryEntry as Qt, EnvironmentInput as R, Verdict as Rt, CompareAnswer as S, effortOf as Sn, Profile as St, CompareQuestion as T, effortVerdict as Tn, ReevalConfig as Tt, EvalOptionsData as U, AttemptRecord as Ut, EvalFn as V, WebEventType as Vt, ExecResult as W, CellRecord as Wt, FakeScriptStep as X, FileRef as Xt, FakeSandboxDescriptor as Y, FamilyRecord as Yt, FakeVote as Z, HistoryCell as Zt, Cell as _, cellSamples as _n, Outcome as _t, BackgroundHandle as a, EFFORT_ALPHA as an, HarnessOption as at, ClaudeCodeHarnessDescriptor as b, effortChangeLabel as bn, PiHarnessDescriptor as bt, Capture as c, EffortMetric as cn, JudgeArtifactRef as ct, CaptureOptions as d, Spread as dn, JudgeModel as dt, Lift as en, GrokBuildHarnessDescriptor as et, CaptureUse as f, TOKENS_UNREPORTED as fn, JudgeOptions as ft, CaptureWindow as g, cellEffort as gn, OpenCodeHarnessDescriptor as gt, CaptureWait as h, cellContrast as hn, McpServer as ht, AttemptContext as i, DIRECTIONAL_MIN as in, HarnessEdit as it, DockerSandboxDescriptor as j, SetupStep as jt, Diff as k, tokenValues as kn, SecretTemplate as kt, CaptureCall as l, EffortSample as ln, JudgeCallOptions as lt, CaptureUseScript as m, bothUnderATenth as mn, JudgeStateOptions as mt, AnalysisConfig as n, StackSummary as nn, HarnessData as nt, BlindPair as o, EFFORT_MIN_RUNS as on, HistoryConfig as ot, CaptureUseAction as p, VERDICT_TEXT as pn, JudgeState as pt, FakeJudgeModel as q, FamilyLift as qt, AntigravityHarnessDescriptor as r, CellEffort as rn, HarnessDescriptor as rt, Budget as s, EffortContrast as sn, HostFile as st, AgentHarnessDescriptor as t, RunRecord as tn, GroupOptions as tt, CaptureInstruction as u, EffortVerdict as un, JudgeFailure as ut, CellRole as v, contrastEffort as vn, PairFactor as vt, CompareOptions as w, effortSummary as wn, ProfileLimit as wt, CodexHarnessDescriptor as x, effortLabel as xn, PinnableKey as xt, ClaudeCodeEffort as y, directional as yn, PairVote as yt, EnvironmentSpec as z, Viewport as zt };
|
|
2625
|
+
//# sourceMappingURL=types-BYdlwMwo.d.ts.map
|
package/dist/view/index.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { $t as
|
|
1
|
+
import { $t as JudgementRecord, B as ErrorKind, Cn as effortSample, Dn as formatRatio, Dt as SandboxWarning, En as formatCount, Gt as CheckTally, Ht as ArtifactRef, Jt as FamilyRate, Kt as EvalEntry, Lt as TurnStatus, Nt as TokenUsage, On as formatSeconds, Pt as TranscriptEvent, Qt as HistoryEntry, Sn as effortOf, Tn as effortVerdict, Ut as AttemptRecord, Wt as CellRecord, Xt as FileRef, Yt as FamilyRecord, Zt as HistoryCell, _n as cellSamples, _t as Outcome, an as EFFORT_ALPHA, bn as effortChangeLabel, cn as EffortMetric, dn as Spread, en as Lift, fn as TOKENS_UNREPORTED, gn as cellEffort, hn as cellContrast, in as DIRECTIONAL_MIN, kn as tokenValues, ln as EffortSample, mn as bothUnderATenth, nn as StackSummary, on as EFFORT_MIN_RUNS, pn as VERDICT_TEXT, qt as FamilyLift, rn as CellEffort, sn as EffortContrast, tn as RunRecord, un as EffortVerdict, vn as contrastEffort, wn as effortSummary, xn as effortLabel, yn as directional } from "../types-BYdlwMwo.js";
|
|
2
2
|
//#region src/view/analysis.d.ts
|
|
3
3
|
/** Bumped when the system prompt or output schema changes: an older analysis is superseded. */
|
|
4
4
|
export declare const ANALYSIS_PROMPT_VERSION = "analysis-v10";
|
|
@@ -833,5 +833,253 @@ export declare const readableTranscriptLine: (line: TranscriptLine) => Transcrip
|
|
|
833
833
|
/** An analysis document as every reader sees it; one from a newer reeval throws. */
|
|
834
834
|
export declare const readableAnalysis: (doc: AnalysisDoc, what?: string) => AnalysisDoc;
|
|
835
835
|
//#endregion
|
|
836
|
-
|
|
836
|
+
//#region src/view/suite.d.ts
|
|
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 };
|
|
837
1085
|
//# sourceMappingURL=index.d.ts.map
|
package/dist/view/index.js
CHANGED
|
@@ -1,9 +1,9 @@
|
|
|
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
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-Bxub9LMT.js";
|
|
4
|
-
import { $ as
|
|
5
|
-
import { a as familyLiftLabel, i as
|
|
6
|
-
import { t as runHeader } from "../header-
|
|
4
|
+
import { $ as contrastEffort, A as splitPatch, B as judgeRatesLabel, C as parseTranscript, Ct as whyChanged, D as runtimeLabel, E as runRows, F as checkLabels, G as EFFORT_ALPHA, H as summaryJudgeRates, I as collapsedJudgeLabel, J as VERDICT_TEXT, K as EFFORT_MIN_RUNS, L as collapsedJudges, M as LIFT_MIN_COUNTED, N as cellJudgeRates, O as shortSha, P as cellUsage, Q as cellSamples, R as commandLine, S as languageLabel, St as stackChanges, T as runOutcome, U as uniqueJudgeLabels, V as labelledJudgements, W as DIRECTIONAL_MIN, X as cellContrast, Y as bothUnderATenth, Z as cellEffort, _ as harnessLabel, a as cellRows, at as effortSummary, b as isMatrix, bt as compare, c as evalHistory, ct as formatRatio, d as failedJudgeLabels, et as directional, f as findAttempt, g as formatUsage, h as formatPercent, i as buildIndex, it as effortSample, j as transcriptItems, k as splitEvalId, l as evalMarkOutcome, lt as formatSeconds, m as formatDuration, n as activityLabel, nt as effortLabel, o as compareView, ot as effortVerdict, p as findEval, q as TOKENS_UNREPORTED, r as answerRows, rt as effortOf, s as defaultAttempt, st as formatCount, t as COMPARE_ORDER, tt as effortChangeLabel, u as evalTree, ut as tokenValues, v as hasUngatedCells, w as runLead, x as isUngatedMiss, xt as sourceChanges, y as indexFiles, yt as cellsChanged, z as judgeLabel } from "../model-DaX4bNLZ.js";
|
|
5
|
+
import { a as factSegment, c as caseLabel, d as familyHeadLabel, f as familyLiftLabel, i as evalAcross, l as familyCases, n as checkEffect, o as messageLine, p as familyRateLabel, r as effectOf, s as suiteResult, t as EFFECT_MIN_COUNTED, u as familyEffortLabel } from "../suite-Cw_7G2RM.js";
|
|
6
|
+
import { t as runHeader } from "../header-DHQv_ZrF.js";
|
|
7
7
|
import { c as readableTranscriptLine, r as readableAnalysis, s as readableRecord, t as NewerSchemaError } from "../record-BKmMvZ6O.js";
|
|
8
8
|
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-jlfszAbb.js";
|
|
9
|
-
export { ANALYSIS_PROMPT_VERSION, COMPARE_ORDER, DEBUG_LOG, DEBUG_PHASES, DIGEST_VERSION, DIRECTIONAL_MIN, EFFORT_ALPHA, EFFORT_MIN_RUNS, FINISHED, LIFT_MIN_COUNTED, LIVE_FILE, LIVE_SCHEMA_VERSION, NewerSchemaError, REQUESTS_FILE, RUN_ID, TOKENS_UNREPORTED, TRACE_FILE, VERDICT_TEXT, activityLabel, analysisDataPath, answerRows, artifactPath, artifactType, buildIndex, caseLabel, cellContrast, cellEffort, cellJudgeRates, cellRows, cellSamples, cellUsage, cellsChanged, checkLabels, collapsedJudgeLabel, collapsedJudges, commandLine, compare, compareView, contrastEffort, debugApiPath, debugLogLine, debugTime, defaultAttempt, directional, effortChangeLabel, effortLabel, effortOf, effortSample, effortSummary, effortVerdict, evalHistory, evalMarkOutcome, evalTree, failedJudgeLabels, familyCases, familyEffortLabel, familyHeadLabel, familyLiftLabel, familyRateLabel, findAttempt, findEval, formatBytes, formatCount, formatDuration, formatPercent, formatRatio, formatSeconds, formatUsage, harnessLabel, hasUngatedCells, indexFiles, isFinished, isMatrix, isSuperseded, isUngatedMiss, judgeLabel, judgeRatesLabel, labelledJudgements, languageLabel, liveApiPath, liveCounts, liveTranscriptPath, manifestPath, parseDataPath, parseDebugLog, parseNdjson, parseTranscript, pinnedAnalysis, readableAnalysis, readableLiveProgress, readableRecord, readableTranscriptLine, recordPath, requestLine, runHeader, runLead, runOutcome, runRows, runtimeLabel, shardOf, shardPath, shortSha, sortAnalyses, sourceChanges, splitEvalId, splitPatch, stackChanges, studioPath, summaryJudgeRates, tokenValues, traceApiPath, transcriptItems, uniqueJudgeLabels, whyChanged };
|
|
9
|
+
export { ANALYSIS_PROMPT_VERSION, COMPARE_ORDER, DEBUG_LOG, DEBUG_PHASES, DIGEST_VERSION, DIRECTIONAL_MIN, EFFECT_MIN_COUNTED, EFFORT_ALPHA, EFFORT_MIN_RUNS, FINISHED, LIFT_MIN_COUNTED, LIVE_FILE, LIVE_SCHEMA_VERSION, NewerSchemaError, REQUESTS_FILE, RUN_ID, TOKENS_UNREPORTED, TRACE_FILE, VERDICT_TEXT, activityLabel, analysisDataPath, answerRows, artifactPath, artifactType, bothUnderATenth, buildIndex, caseLabel, cellContrast, cellEffort, cellJudgeRates, cellRows, cellSamples, cellUsage, cellsChanged, checkEffect, checkLabels, collapsedJudgeLabel, collapsedJudges, commandLine, compare, compareView, contrastEffort, 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, harnessLabel, hasUngatedCells, indexFiles, isFinished, isMatrix, isSuperseded, isUngatedMiss, judgeLabel, judgeRatesLabel, labelledJudgements, languageLabel, liveApiPath, liveCounts, liveTranscriptPath, manifestPath, messageLine, parseDataPath, parseDebugLog, parseNdjson, parseTranscript, pinnedAnalysis, readableAnalysis, 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 };
|
|
@@ -1,7 +1,8 @@
|
|
|
1
1
|
import { r as formatDuration } from "./text-gEcKjAyp.js";
|
|
2
|
-
import {
|
|
3
|
-
import {
|
|
4
|
-
import {
|
|
2
|
+
import { St as stackChanges, tt as effortChangeLabel, yt as cellsChanged } from "./model-DaX4bNLZ.js";
|
|
3
|
+
import { s as suiteResult } from "./suite-Cw_7G2RM.js";
|
|
4
|
+
import { o as warnings } from "./state-DfpHox6b.js";
|
|
5
|
+
import { a as render, i as stateRows, r as runTally, t as OUTCOME_MARK } from "./summary-BUKfU-9l.js";
|
|
5
6
|
//#region src/history/views.ts
|
|
6
7
|
var RefError = class extends Error {
|
|
7
8
|
name = "RefError";
|
|
@@ -190,36 +191,25 @@ const liftShort = (l) => l ? `${signed(l.points)} (${signed(l.lo)}…${signed(l.
|
|
|
190
191
|
const signed = (x) => `${x >= 0 ? "+" : "−"}${Math.round(Math.abs(x) * 100)}`;
|
|
191
192
|
const liftChange = (c) => `${liftShort(c.base.lift)} → ${liftShort(c.head.lift)}`;
|
|
192
193
|
const effortChange = (c) => c.base.effort && c.head.effort ? effortChangeLabel(c.base.effort, c.head.effort) : "effort –";
|
|
193
|
-
|
|
194
|
+
/** `reeval show`: the run's header and state, then the result blocks `reeval run` ended with. */
|
|
195
|
+
function formatShow(run, opts) {
|
|
194
196
|
const rows = stateRows(run);
|
|
195
197
|
const width = Math.max(...rows.map(([k]) => k.length));
|
|
196
|
-
|
|
197
|
-
`run ${run.runId} ${
|
|
198
|
+
return [
|
|
199
|
+
`run ${run.runId} ${runTally(run)}`,
|
|
198
200
|
` started ${run.startedAt} (${formatDuration(run.durationMs)}) · artifacts ${run.artifacts}`,
|
|
199
201
|
"",
|
|
200
202
|
...rows.map(([k, v]) => ` ${k.padEnd(width)} ${v}`),
|
|
201
203
|
...warnings(run).map((w) => ` ! ${w}`),
|
|
202
|
-
""
|
|
203
|
-
|
|
204
|
-
|
|
205
|
-
|
|
206
|
-
|
|
207
|
-
|
|
208
|
-
|
|
209
|
-
if (contrasts.length) {
|
|
210
|
-
lines.push(" contrasts (each pair of arms, same harness)");
|
|
211
|
-
for (const line of contrasts) lines.push(` ${line}`);
|
|
212
|
-
}
|
|
213
|
-
for (const t of e.attempts) {
|
|
214
|
-
const indent = single ? " " : " ";
|
|
215
|
-
if (!single) lines.push(` ${OUTCOME_MARK[t.outcome]} ${t.id} ${formatDuration(t.durationMs)}`);
|
|
216
|
-
for (const { judgement: j, label, labelled } of labelledJudgements(t)) lines.push(`${indent}judge ${j.verdict} ${j.score.toFixed(2)} (≥ ${j.threshold}) ${labelled ? `${label}: ` : ""}${j.question}`);
|
|
217
|
-
if (t.error && !repeatsVerdict(t)) lines.push(`${indent}${t.error.kind}: ${t.error.message.split("\n")[0]}`);
|
|
218
|
-
}
|
|
219
|
-
}
|
|
220
|
-
return lines.join("\n");
|
|
204
|
+
"",
|
|
205
|
+
...render(suiteResult(run), null, {
|
|
206
|
+
color: false,
|
|
207
|
+
verbose: false,
|
|
208
|
+
hint: opts.hint
|
|
209
|
+
})
|
|
210
|
+
].join("\n");
|
|
221
211
|
}
|
|
222
212
|
//#endregion
|
|
223
213
|
export { formatShow as a, resolveRun as c, formatNoBaseline as i, unreachable as l, formatCompare as n, history as o, formatHistory as r, pickBaseline as s, RefError as t };
|
|
224
214
|
|
|
225
|
-
//# sourceMappingURL=views-
|
|
215
|
+
//# sourceMappingURL=views-B8w2J-gD.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"views-B8w2J-gD.js","names":[],"sources":["../src/history/views.ts"],"mappings":";;;;;;AAuBA,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;AAEA,MAAM,eAA8C;CAClD,WAAW;CACX,OAAO;CACP,MAAM;CACN,OAAO;CACP,OAAO;CACP,SAAS;CACT,OAAO;AACT;;;;;AAMA,MAAM,OAAO,MACX;CACE,EAAE,SAAS,SAAS,KAAK,EAAE,QAAQ,KAAK,IAAI,MAAM;CAClD,EAAE,OAAO,MAAM,SAAS,mBAAmB,EAAE,MAAM,MAAM,KAAK,IAAI,MAAM;CACxE,EAAE,OAAO,QAAQ,SAAS,qBAAqB,EAAE,MAAM,QAAQ,KAAK,IAAI,MAAM;AAChF,CAAC,CAAC,KAAK,EAAE;AAEX,MAAM,QAAyB;CAAC;CAAa;CAAS;CAAS;CAAS;CAAS;CAAW;AAAM;AAElG,MAAM,eAAe,GAAiB,UACpC,GAAG,EAAE,MAAM,IAAI,SAAS,CAAC,IAAI,EAAE,WAAW,KAAK,mBAAmB,YAAY,GAAG,KAAK,IAAI,mCAAmC,GAAG;AAElI,SAAgB,cACd,MACA,MACA,MACA,IACA,QAAsB,MACtB;CACA,MAAM,SAAS,CAAC,GAAG,IAAI,CAAC,CAAC,MAAM,GAAG,MAAM,MAAM,QAAQ,EAAE,MAAM,IAAI,MAAM,QAAQ,EAAE,MAAM,CAAC;CACzF,MAAM,QAAQ,MAAuB,IAAI,aAAa,KAAK;CAC3D,MAAM,SAAS,MAAqB,KAAK,QAAQ,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC;CACvE,MAAM,QAAQ,GAAG,MAAM,WAAW,EAAE,eAAe,MAAM,OAAO,EAAE,WAAW,MAAM,OAAO,EAAE,oBAAoB,MAAM,MAAM,EAAE;CAC9H,MAAM,UAAU,aAAa,MAAM,IAAI;CACvC,MAAM,aAAa,QAAQ,KAAK,MAAM,GAAG,EAAE,MAAM,GAAG,EAAE,KAAK,KAAK,EAAE,MAAM,CAAC,CAAC,KAAK,IAAI;CACnF,MAAM,OAAO,MAAc,EAAE,QAAQ,OAAO,KAAK;CACjD,MAAM,QAAQ,OAAO,SAAS,OAAO,EAAE,WAAW,CAAC,EAAA,CAAG,KAAK,OAAO;EAAE,KAAK;EAAG,MAAM;CAAE,EAAE,CAAC;CACvF,IAAI,IACF,OAAO;EACL,eAAe,KAAK,MAAM,MAAM,KAAK;EACrC;EACA,QAAQ,YAAY,MAAM,KAAK,EAAE,UAAU,YAAY,MAAM,KAAK,EAAE,IAAI,MAAM;EAC9E;EACA,GAAI,QAAQ,SAAS,CAAC,sBAAsB,WAAW,IAAI,EAAE,IAAI,CAAC;EAClE;EACA;EACA,GAAG,OAAO,KACP,MACC,KAAK,IAAI,EAAE,EAAE,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,aAAa,EAAE,UAAU,IAAI,CAAC,EAAE,GAC5F;EACA;EACA,GAAI,MAAM,SACN;GACE;GACA;GACA,GAAG,MAAM,KACN,EAAE,KAAK,WACN,KAAK,IAAI,IAAI,EAAE,EAAE,KAAK,IAAI,KAAK,GAAG,EAAE,KAAK,WAAW,IAAI,EAAE,KAAK,aAAa,IAAI,EAAE,GACtF;GACA;EACF,IACA,CAAC;CACP,CAAC,CAAC,KAAK,IAAI;CAEb,OAAO;EACL,QAAQ,YAAY,MAAM,KAAK,EAAE,UAAU,YAAY,MAAM,KAAK;EAClE;EACA,GAAI,QAAQ,SAAS,CAAC,kBAAkB,YAAY,IAAI,CAAC;EACzD;EACA,GAAG,OAAO,SAAS,MAAM,CACvB,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,IAAI,aAAa,EAAE,OAAO,CAAC,OAAO,EAAE,EAAE,IAAI,EAAE,KAAK,IAAI,CAAC,KAC1F,IAAI,EAAE,WAAW,CAAC,EAAA,CAAG,KAClB,MACC,SAAS,EAAE,IAAI,IAAI,CAAC,EAAE,KAAK,QAAQ,EAAE,KAAK,OAAO,QAAQ,WAAW,CAAC,MAAM,IAAI,aAAa,CAAC,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK,KAAK,GAC9H,CACF,CAAC;CACH,CAAC,CAAC,KAAK,IAAI;AACb;;AAGA,SAAgB,iBAAiB,MAAoB,MAAoB,IAAqB;CAC5F,MAAM,UAAU,KAAK,QAAQ,MAAM,EAAE,SAAS,MAAM,CAAC,CAAC;CACtD,MAAM,OAAO,sCAAsC,KAAK,MAAM,0BAA0B,UAAU,SAAS,QAAQ,OAAO,YAAY,IAAI,MAAM,OAAO,GAAG;CAC1J,MAAM,QAAQ,MAAuB,IAAI,aAAa,KAAK;CAC3D,IAAI,IACF,OAAO;EACL,eAAe,KAAK;EACpB;EACA;EACA;EACA;EACA;EACA,GAAG,KAAK,KAAK,MAAM,KAAK,EAAE,GAAG,QAAQ,OAAO,KAAK,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,GAAG;EACxE;CACF,CAAC,CAAC,KAAK,IAAI;CAEb,OAAO;EAAC;EAAM;EAAI,GAAG,KAAK,KAAK,MAAM,KAAK,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI;CAAC,CAAC,CAAC,KAAK,IAAI;AACpF;AAEA,MAAM,aAAa,MACjB,IAAI,GAAG,OAAO,EAAE,MAAM,EAAE,IAAI,OAAO,EAAE,EAAE,EAAE,GAAG,OAAO,EAAE,EAAE,EAAE,GAAG,EAAE,cAAc,WAAW,OAAO;AAChG,MAAM,UAAU,MAAc,GAAG,KAAK,IAAI,MAAM,MAAM,KAAK,MAAM,KAAK,IAAI,CAAC,IAAI,GAAG;AAClF,MAAM,cAAc,MAAmB,GAAG,UAAU,EAAE,KAAK,IAAI,EAAE,KAAK,UAAU,EAAE,KAAK,IAAI;AAC3F,MAAM,gBAAgB,MACpB,EAAE,KAAK,UAAU,EAAE,KAAK,SAAS,kBAAkB,EAAE,KAAK,QAAQ,EAAE,KAAK,MAAM,IAAI;;AAGrF,SAAgB,WAAW,KAAgB,MAAiC;CAC1E,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,MAAM;GAAE,OAAO;GAAO,SAAS;GAAO,MAAM,KAAK;EAAK,CAAC;CACrF,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-Bwy3QxAK.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-DNnveLqx.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"worker-entry-
|
|
1
|
+
{"version":3,"file":"worker-entry-DNnveLqx.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"}
|
|
@@ -1,8 +1,8 @@
|
|
|
1
|
-
import { A as workspaceLabel } from "./environment-
|
|
1
|
+
import { A as workspaceLabel } from "./environment-Cca6a4wB.js";
|
|
2
2
|
import { l as sha256 } from "./ids-Bp6ut7ZD.js";
|
|
3
3
|
import { s as registerSecret } from "./redact-DJ9JCShu.js";
|
|
4
|
-
import {
|
|
5
|
-
import { t as git } from "./git-
|
|
4
|
+
import { it as RETRY_DELAYS } from "./src-ftu5wPSt.js";
|
|
5
|
+
import { t as git } from "./git-Crl3b87o.js";
|
|
6
6
|
import { t as PrepareError } from "./provider-CmAYTQQs.js";
|
|
7
7
|
import { existsSync, mkdirSync, readFileSync, renameSync, rmSync, writeFileSync } from "node:fs";
|
|
8
8
|
import { join, posix } from "node:path";
|
|
@@ -324,4 +324,4 @@ function preparedWorkspace(root, runId, w) {
|
|
|
324
324
|
//#endregion
|
|
325
325
|
export { removePreparedWorkspaces as a, preparedWorkspace as i, classifyGitError as n, workspaceKey as o, prepareWorkspaces as r, WorkspaceError as t };
|
|
326
326
|
|
|
327
|
-
//# sourceMappingURL=workspace-
|
|
327
|
+
//# sourceMappingURL=workspace-C4RA3srV.js.map
|