@tangle-network/agent-eval 0.128.2 → 0.130.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +279 -0
- package/README.md +19 -0
- package/dist/active-curriculum-C4mk67HP.js +214 -0
- package/dist/active-curriculum-C4mk67HP.js.map +1 -0
- package/dist/adversarial-smnADNFS.d.ts +21 -0
- package/dist/adversarial-smnADNFS.d.ts.map +1 -0
- package/dist/analyst/index.d.ts +83 -2932
- package/dist/analyst/index.d.ts.map +1 -0
- package/dist/analyst/index.js +319 -364
- package/dist/analyst/index.js.map +1 -1
- package/dist/analyst-BkTS3C58.d.ts +89 -0
- package/dist/analyst-BkTS3C58.d.ts.map +1 -0
- package/dist/analyst-LsnNpSkm.js +152 -0
- package/dist/analyst-LsnNpSkm.js.map +1 -0
- package/dist/analyze-runs-C1CavBMk.js +1067 -0
- package/dist/analyze-runs-C1CavBMk.js.map +1 -0
- package/dist/analyze-runs-FsgYCinh.d.ts +72 -0
- package/dist/analyze-runs-FsgYCinh.d.ts.map +1 -0
- package/dist/attribute-vocabulary-DLJ6303h.d.ts +54 -0
- package/dist/attribute-vocabulary-DLJ6303h.d.ts.map +1 -0
- package/dist/authenticity/index.d.ts +81 -79
- package/dist/authenticity/index.d.ts.map +1 -0
- package/dist/authenticity/index.js +209 -193
- package/dist/authenticity/index.js.map +1 -1
- package/dist/baseline-HsBvw_dk.js +550 -0
- package/dist/baseline-HsBvw_dk.js.map +1 -0
- package/dist/baseline-hG3K85h4.d.ts +125 -0
- package/dist/baseline-hG3K85h4.d.ts.map +1 -0
- package/dist/belief-state/index.d.ts +448 -1205
- package/dist/belief-state/index.d.ts.map +1 -0
- package/dist/belief-state/index.js +1617 -1710
- package/dist/belief-state/index.js.map +1 -1
- package/dist/benchmarks/index.d.ts +2 -894
- package/dist/benchmarks/index.js +2 -59
- package/dist/benchmarks-DviOvUNr.js +754 -0
- package/dist/benchmarks-DviOvUNr.js.map +1 -0
- package/dist/builder-eval/index.d.ts +150 -662
- package/dist/builder-eval/index.d.ts.map +1 -0
- package/dist/builder-eval/index.js +356 -345
- package/dist/builder-eval/index.js.map +1 -1
- package/dist/calibration-CNWWA6K8.js +94 -0
- package/dist/calibration-CNWWA6K8.js.map +1 -0
- package/dist/campaign/index.d.ts +5 -6390
- package/dist/campaign/index.js +3 -212
- package/dist/campaign-CBKZvQ1H.js +3885 -0
- package/dist/campaign-CBKZvQ1H.js.map +1 -0
- package/dist/cli.d.ts +1 -1
- package/dist/cli.js +164 -174
- package/dist/cli.js.map +1 -1
- package/dist/client-C97NMzqi.d.ts +581 -0
- package/dist/client-C97NMzqi.d.ts.map +1 -0
- package/dist/client-CYzbdJOZ.js +637 -0
- package/dist/client-CYzbdJOZ.js.map +1 -0
- package/dist/code-agent-session-BjkMTQ7H.js +1390 -0
- package/dist/code-agent-session-BjkMTQ7H.js.map +1 -0
- package/dist/code-agent-session-DqqgOJaz.d.ts +143 -0
- package/dist/code-agent-session-DqqgOJaz.d.ts.map +1 -0
- package/dist/concurrency-DIxRZF_J.js +85 -0
- package/dist/concurrency-DIxRZF_J.js.map +1 -0
- package/dist/contract/index.d.ts +645 -5605
- package/dist/contract/index.d.ts.map +1 -0
- package/dist/contract/index.js +1730 -1937
- package/dist/contract/index.js.map +1 -1
- package/dist/control.d.ts +3 -1030
- package/dist/control.js +2 -32
- package/dist/cost-ledger-DIgQUFZZ.js +801 -0
- package/dist/cost-ledger-DIgQUFZZ.js.map +1 -0
- package/dist/cost-ledger-Dye6jCgg.d.ts +236 -0
- package/dist/cost-ledger-Dye6jCgg.d.ts.map +1 -0
- package/dist/dataset-BvtnC8Dc.d.ts +115 -0
- package/dist/dataset-BvtnC8Dc.d.ts.map +1 -0
- package/dist/default-registry-C-vFCSEc.js +2579 -0
- package/dist/default-registry-C-vFCSEc.js.map +1 -0
- package/dist/default-registry-CNPo-Vsb.d.ts +540 -0
- package/dist/default-registry-CNPo-Vsb.d.ts.map +1 -0
- package/dist/emitter-CPBAhxum.js +266 -0
- package/dist/emitter-CPBAhxum.js.map +1 -0
- package/dist/emitter-DGQGoLyj.d.ts +113 -0
- package/dist/emitter-DGQGoLyj.d.ts.map +1 -0
- package/dist/errors-8YnH8WlF.js +64 -0
- package/dist/errors-8YnH8WlF.js.map +1 -0
- package/dist/errors-CEk209JS.d.ts +76 -0
- package/dist/errors-CEk209JS.d.ts.map +1 -0
- package/dist/eval-campaign-DEm6c8ru.js +349 -0
- package/dist/eval-campaign-DEm6c8ru.js.map +1 -0
- package/dist/execution-tracks-CpgFPpS5.js +93 -0
- package/dist/execution-tracks-CpgFPpS5.js.map +1 -0
- package/dist/exporters-q9iL-2Jf.js +148 -0
- package/dist/exporters-q9iL-2Jf.js.map +1 -0
- package/dist/extract-usage-BrQ8mCLX.js +155 -0
- package/dist/extract-usage-BrQ8mCLX.js.map +1 -0
- package/dist/failure-cluster-CqcvCcdR.d.ts +59 -0
- package/dist/failure-cluster-CqcvCcdR.d.ts.map +1 -0
- package/dist/feedback-trajectory-CVaeREXV.d.ts +340 -0
- package/dist/feedback-trajectory-CVaeREXV.d.ts.map +1 -0
- package/dist/fuzz.d.ts +320 -646
- package/dist/fuzz.d.ts.map +1 -0
- package/dist/fuzz.js +670 -617
- package/dist/fuzz.js.map +1 -1
- package/dist/hf-dataset-DBJXXoY1.js +763 -0
- package/dist/hf-dataset-DBJXXoY1.js.map +1 -0
- package/dist/hosted/index.d.ts +11 -831
- package/dist/hosted/index.d.ts.map +1 -0
- package/dist/hosted/index.js +2 -37
- package/dist/index-2JJSA6-r2.d.ts +926 -0
- package/dist/index-2JJSA6-r2.d.ts.map +1 -0
- package/dist/index-6N0aYmpW.d.ts +217 -0
- package/dist/index-6N0aYmpW.d.ts.map +1 -0
- package/dist/index-BAvgST_9.d.ts +131 -0
- package/dist/index-BAvgST_9.d.ts.map +1 -0
- package/dist/index-BvnJuTGD.d.ts +68 -0
- package/dist/index-BvnJuTGD.d.ts.map +1 -0
- package/dist/index-C61Wi7yg.d.ts +547 -0
- package/dist/index-C61Wi7yg.d.ts.map +1 -0
- package/dist/index-CAPUUKaM.d.ts +335 -0
- package/dist/index-CAPUUKaM.d.ts.map +1 -0
- package/dist/index-DE5fb3EC.d.ts +2244 -0
- package/dist/index-DE5fb3EC.d.ts.map +1 -0
- package/dist/index-DSC51roc.d.ts +102 -0
- package/dist/index-DSC51roc.d.ts.map +1 -0
- package/dist/index.d.ts +3776 -15120
- package/dist/index.d.ts.map +1 -0
- package/dist/index.js +11185 -11191
- package/dist/index.js.map +1 -1
- package/dist/integrity-BzRbCHzi.js +424 -0
- package/dist/integrity-BzRbCHzi.js.map +1 -0
- package/dist/integrity-rmVhXWA7.d.ts +61 -0
- package/dist/integrity-rmVhXWA7.d.ts.map +1 -0
- package/dist/judge-calibration-DFtEMlde.d.ts +146 -0
- package/dist/judge-calibration-DFtEMlde.d.ts.map +1 -0
- package/dist/ledger-core/index.d.ts +2 -0
- package/dist/ledger-core/index.js +2 -0
- package/dist/ledger-core-DtZz1RG0.js +388 -0
- package/dist/ledger-core-DtZz1RG0.js.map +1 -0
- package/dist/llm-client--GR4JbZE.js +687 -0
- package/dist/llm-client--GR4JbZE.js.map +1 -0
- package/dist/llm-client-B_nIBlYo.d.ts +290 -0
- package/dist/llm-client-B_nIBlYo.d.ts.map +1 -0
- package/dist/matrix/index.d.ts +3 -155
- package/dist/matrix/index.js +2 -8
- package/dist/matrix-BzQnu2S6.js +270 -0
- package/dist/matrix-BzQnu2S6.js.map +1 -0
- package/dist/meta-eval/index.d.ts +4 -1027
- package/dist/meta-eval/index.js +393 -390
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/metrics-C9YY1OcL.js +239 -0
- package/dist/metrics-C9YY1OcL.js.map +1 -0
- package/dist/mint-yN2M2eh0.js +201 -0
- package/dist/mint-yN2M2eh0.js.map +1 -0
- package/dist/multi-layer-verifier-BHY1gWAc.d.ts +138 -0
- package/dist/multi-layer-verifier-BHY1gWAc.d.ts.map +1 -0
- package/dist/multishot/index.d.ts +273 -481
- package/dist/multishot/index.d.ts.map +1 -0
- package/dist/multishot/index.js +595 -548
- package/dist/multishot/index.js.map +1 -1
- package/dist/off-policy-DvgzvtIx.js +220 -0
- package/dist/off-policy-DvgzvtIx.js.map +1 -0
- package/dist/off-policy-mskQw8Mb.d.ts +153 -0
- package/dist/off-policy-mskQw8Mb.d.ts.map +1 -0
- package/dist/openapi.json +1 -1
- package/dist/opencode-sqlite-BGrHeDu3.js +318 -0
- package/dist/opencode-sqlite-BGrHeDu3.js.map +1 -0
- package/dist/outcome-store-BYHIuO0e.d.ts +65 -0
- package/dist/outcome-store-BYHIuO0e.d.ts.map +1 -0
- package/dist/outcome-store-ChBKlTd_.js +75 -0
- package/dist/outcome-store-ChBKlTd_.js.map +1 -0
- package/dist/paired-arms-D9D0wXj2.js +260 -0
- package/dist/paired-arms-D9D0wXj2.js.map +1 -0
- package/dist/pipelines/index.d.ts +95 -532
- package/dist/pipelines/index.d.ts.map +1 -0
- package/dist/pipelines/index.js +497 -478
- package/dist/pipelines/index.js.map +1 -1
- package/dist/pre-registration-DakwTRXk.js +96 -0
- package/dist/pre-registration-DakwTRXk.js.map +1 -0
- package/dist/propose-review-control-LhIWGzHl.js +1458 -0
- package/dist/propose-review-control-LhIWGzHl.js.map +1 -0
- package/dist/query-CJ_DX8vl.d.ts +28 -0
- package/dist/query-CJ_DX8vl.d.ts.map +1 -0
- package/dist/query-Di7eEQ79.js +83 -0
- package/dist/query-Di7eEQ79.js.map +1 -0
- package/dist/raw-provider-sink-BQd7mzyT.js +201 -0
- package/dist/raw-provider-sink-BQd7mzyT.js.map +1 -0
- package/dist/raw-provider-sink-BU29Sh8h.d.ts +134 -0
- package/dist/raw-provider-sink-BU29Sh8h.d.ts.map +1 -0
- package/dist/redact-7Aq1ukl-.js +107 -0
- package/dist/redact-7Aq1ukl-.js.map +1 -0
- package/dist/release-report-Cz9NKH39.js +603 -0
- package/dist/release-report-Cz9NKH39.js.map +1 -0
- package/dist/release-report-mvB2G4_J.d.ts +244 -0
- package/dist/release-report-mvB2G4_J.d.ts.map +1 -0
- package/dist/replay-GnyotH0J.js +1741 -0
- package/dist/replay-GnyotH0J.js.map +1 -0
- package/dist/replay-OoidtG1E.d.ts +749 -0
- package/dist/replay-OoidtG1E.d.ts.map +1 -0
- package/dist/reporting.d.ts +6 -1298
- package/dist/reporting.js +6 -50
- package/dist/researcher-CwTdwXG1.d.ts +314 -0
- package/dist/researcher-CwTdwXG1.d.ts.map +1 -0
- package/dist/reward-hacking-eAnOsynk.d.ts +224 -0
- package/dist/reward-hacking-eAnOsynk.d.ts.map +1 -0
- package/dist/reward-hacking-qipEpKvY.js +596 -0
- package/dist/reward-hacking-qipEpKvY.js.map +1 -0
- package/dist/reward-nw2xZGZG.js +137 -0
- package/dist/reward-nw2xZGZG.js.map +1 -0
- package/dist/rl.d.ts +916 -3596
- package/dist/rl.d.ts.map +1 -0
- package/dist/rl.js +2362 -1751
- package/dist/rl.js.map +1 -1
- package/dist/rolldown-runtime-8H4AJuhK.js +14 -0
- package/dist/rollout/index.d.ts +3 -1048
- package/dist/rollout/index.js +8 -110
- package/dist/rollout-BOYjemfR.js +624 -0
- package/dist/rollout-BOYjemfR.js.map +1 -0
- package/dist/rubric-predictive-validity-B3xmbmS1.js +141 -0
- package/dist/rubric-predictive-validity-B3xmbmS1.js.map +1 -0
- package/dist/rubric-predictive-validity-Ku_clp_1.d.ts +74 -0
- package/dist/rubric-predictive-validity-Ku_clp_1.d.ts.map +1 -0
- package/dist/run-evidence-ClRX_8A9.d.ts +225 -0
- package/dist/run-evidence-ClRX_8A9.d.ts.map +1 -0
- package/dist/run-record-BuoE80Dq.js +467 -0
- package/dist/run-record-BuoE80Dq.js.map +1 -0
- package/dist/run-record-CnZu_gjl.d.ts +357 -0
- package/dist/run-record-CnZu_gjl.d.ts.map +1 -0
- package/dist/run-score-iEEAWiBY.js +41 -0
- package/dist/run-score-iEEAWiBY.js.map +1 -0
- package/dist/runtime-trajectory-1gyaTOoC.js +93 -0
- package/dist/runtime-trajectory-1gyaTOoC.js.map +1 -0
- package/dist/runtime-trajectory-BvSZcCHD.d.ts +50 -0
- package/dist/runtime-trajectory-BvSZcCHD.d.ts.map +1 -0
- package/dist/schema-BtVldJ3T.d.ts +206 -0
- package/dist/schema-BtVldJ3T.d.ts.map +1 -0
- package/dist/schema-C6DW4ZHR.js +821 -0
- package/dist/schema-C6DW4ZHR.js.map +1 -0
- package/dist/schema-CRhEY1SO.js +69 -0
- package/dist/schema-CRhEY1SO.js.map +1 -0
- package/dist/schema-Cef2cFmb.d.ts +408 -0
- package/dist/schema-Cef2cFmb.d.ts.map +1 -0
- package/dist/semantic-concept-judge-B6cWNJ2K.js +725 -0
- package/dist/semantic-concept-judge-B6cWNJ2K.js.map +1 -0
- package/dist/sequential-Br0mAPHA.js +148 -0
- package/dist/sequential-Br0mAPHA.js.map +1 -0
- package/dist/sequential-CYwq6Ff_.d.ts +141 -0
- package/dist/sequential-CYwq6Ff_.d.ts.map +1 -0
- package/dist/series-convergence-CjO2QdRW.js +43 -0
- package/dist/series-convergence-CjO2QdRW.js.map +1 -0
- package/dist/series-convergence-ofsqPWhs.d.ts +35 -0
- package/dist/series-convergence-ofsqPWhs.d.ts.map +1 -0
- package/dist/server-m5D9cvnG.js +1040 -0
- package/dist/server-m5D9cvnG.js.map +1 -0
- package/dist/skill-usage-C_pXm7lP.d.ts +534 -0
- package/dist/skill-usage-C_pXm7lP.d.ts.map +1 -0
- package/dist/skillopt-optimization-method-B7wX7XkF.d.ts +1739 -0
- package/dist/skillopt-optimization-method-B7wX7XkF.d.ts.map +1 -0
- package/dist/skillopt-optimization-method-D4ODwFVV.js +7712 -0
- package/dist/skillopt-optimization-method-D4ODwFVV.js.map +1 -0
- package/dist/statistics-Cmj6nynr.d.ts +514 -0
- package/dist/statistics-Cmj6nynr.d.ts.map +1 -0
- package/dist/statistics-CnnxdpOg.js +1437 -0
- package/dist/statistics-CnnxdpOg.js.map +1 -0
- package/dist/store-CT9YIIve.d.ts +117 -0
- package/dist/store-CT9YIIve.d.ts.map +1 -0
- package/dist/store-CxJry_cs.d.ts +229 -0
- package/dist/store-CxJry_cs.d.ts.map +1 -0
- package/dist/storyboard/index.d.ts +84 -203
- package/dist/storyboard/index.d.ts.map +1 -0
- package/dist/storyboard/index.js +609 -542
- package/dist/storyboard/index.js.map +1 -1
- package/dist/summary-report-BNs5nmXI.js +862 -0
- package/dist/summary-report-BNs5nmXI.js.map +1 -0
- package/dist/summary-report-Cj9gdw4i.d.ts +407 -0
- package/dist/summary-report-Cj9gdw4i.d.ts.map +1 -0
- package/dist/supervisor-run/index.d.ts +2 -849
- package/dist/supervisor-run/index.js +2 -64
- package/dist/supervisor-run-_lnTLM3z.js +1679 -0
- package/dist/supervisor-run-_lnTLM3z.js.map +1 -0
- package/dist/task-failure-attributes-CQZlB3et.js +311 -0
- package/dist/task-failure-attributes-CQZlB3et.js.map +1 -0
- package/dist/test-graded-scenario-BsqWLmPt.js +318 -0
- package/dist/test-graded-scenario-BsqWLmPt.js.map +1 -0
- package/dist/test-graded-scenario-D1TaI2va.d.ts +141 -0
- package/dist/test-graded-scenario-D1TaI2va.d.ts.map +1 -0
- package/dist/tools-BmuN627J.js +1085 -0
- package/dist/tools-BmuN627J.js.map +1 -0
- package/dist/trace-attributes.d.ts +2 -52
- package/dist/trace-attributes.js +131 -61
- package/dist/trace-attributes.js.map +1 -1
- package/dist/traces.d.ts +12 -2365
- package/dist/traces.js +12 -251
- package/dist/trajectory-D_7rLrvE.js +56 -0
- package/dist/trajectory-D_7rLrvE.js.map +1 -0
- package/dist/types-DGsxbAEd.d.ts +387 -0
- package/dist/types-DGsxbAEd.d.ts.map +1 -0
- package/dist/types-k9tZGKUg.d.ts +640 -0
- package/dist/types-k9tZGKUg.d.ts.map +1 -0
- package/dist/verdict-Dps8_okt.d.ts +37 -0
- package/dist/verdict-Dps8_okt.d.ts.map +1 -0
- package/dist/wire/index.d.ts +702 -1174
- package/dist/wire/index.d.ts.map +1 -0
- package/dist/wire/index.js +2 -81
- package/docs/feature-guide.md +1 -1
- package/docs/rollout.md +116 -2
- package/package.json +18 -10
- package/dist/benchmarks/index.js.map +0 -1
- package/dist/campaign/index.js.map +0 -1
- package/dist/chunk-2JX3CFMB.js +0 -695
- package/dist/chunk-2JX3CFMB.js.map +0 -1
- package/dist/chunk-2MKQIFS4.js +0 -183
- package/dist/chunk-2MKQIFS4.js.map +0 -1
- package/dist/chunk-3RF76KTD.js +0 -84
- package/dist/chunk-3RF76KTD.js.map +0 -1
- package/dist/chunk-5DTSBUL2.js +0 -159
- package/dist/chunk-5DTSBUL2.js.map +0 -1
- package/dist/chunk-7ZZMD7UK.js +0 -386
- package/dist/chunk-7ZZMD7UK.js.map +0 -1
- package/dist/chunk-BOD4O7OF.js +0 -40
- package/dist/chunk-BOD4O7OF.js.map +0 -1
- package/dist/chunk-BYT7ELPS.js +0 -1553
- package/dist/chunk-BYT7ELPS.js.map +0 -1
- package/dist/chunk-DJKY2TSY.js +0 -2428
- package/dist/chunk-DJKY2TSY.js.map +0 -1
- package/dist/chunk-DPUHNQLN.js +0 -232
- package/dist/chunk-DPUHNQLN.js.map +0 -1
- package/dist/chunk-DRYIUNWY.js +0 -622
- package/dist/chunk-DRYIUNWY.js.map +0 -1
- package/dist/chunk-EJGRPCO3.js +0 -617
- package/dist/chunk-EJGRPCO3.js.map +0 -1
- package/dist/chunk-EOSZT7PL.js +0 -2001
- package/dist/chunk-EOSZT7PL.js.map +0 -1
- package/dist/chunk-EZJEIH2R.js +0 -1559
- package/dist/chunk-EZJEIH2R.js.map +0 -1
- package/dist/chunk-GGE4NNQT.js +0 -65
- package/dist/chunk-GGE4NNQT.js.map +0 -1
- package/dist/chunk-HHWE3POT.js +0 -94
- package/dist/chunk-HHWE3POT.js.map +0 -1
- package/dist/chunk-IHQDPH7D.js +0 -171
- package/dist/chunk-IHQDPH7D.js.map +0 -1
- package/dist/chunk-JHCHEVET.js +0 -274
- package/dist/chunk-JHCHEVET.js.map +0 -1
- package/dist/chunk-K4DBDHLK.js +0 -158
- package/dist/chunk-K4DBDHLK.js.map +0 -1
- package/dist/chunk-K6N6XJJX.js +0 -306
- package/dist/chunk-K6N6XJJX.js.map +0 -1
- package/dist/chunk-MA6HLL3S.js +0 -65
- package/dist/chunk-MA6HLL3S.js.map +0 -1
- package/dist/chunk-MAZ26DC7.js +0 -99
- package/dist/chunk-MAZ26DC7.js.map +0 -1
- package/dist/chunk-MHELPNRP.js +0 -1212
- package/dist/chunk-MHELPNRP.js.map +0 -1
- package/dist/chunk-NACAGYSY.js +0 -1040
- package/dist/chunk-NACAGYSY.js.map +0 -1
- package/dist/chunk-NKAGIDE2.js +0 -7633
- package/dist/chunk-NKAGIDE2.js.map +0 -1
- package/dist/chunk-NPCTHQIO.js +0 -91
- package/dist/chunk-NPCTHQIO.js.map +0 -1
- package/dist/chunk-NYLOYM6N.js +0 -332
- package/dist/chunk-NYLOYM6N.js.map +0 -1
- package/dist/chunk-ONWEPEDO.js +0 -57
- package/dist/chunk-ONWEPEDO.js.map +0 -1
- package/dist/chunk-P5W7RQKK.js +0 -576
- package/dist/chunk-P5W7RQKK.js.map +0 -1
- package/dist/chunk-P6FYH6K4.js +0 -1161
- package/dist/chunk-P6FYH6K4.js.map +0 -1
- package/dist/chunk-PBE2LOSS.js +0 -669
- package/dist/chunk-PBE2LOSS.js.map +0 -1
- package/dist/chunk-PC4UYEBM.js +0 -166
- package/dist/chunk-PC4UYEBM.js.map +0 -1
- package/dist/chunk-PXE2VKMX.js +0 -140
- package/dist/chunk-PXE2VKMX.js.map +0 -1
- package/dist/chunk-PZ5AY32C.js +0 -10
- package/dist/chunk-PZ5AY32C.js.map +0 -1
- package/dist/chunk-RZTMDUO7.js +0 -49
- package/dist/chunk-RZTMDUO7.js.map +0 -1
- package/dist/chunk-S5YLIBFX.js +0 -136
- package/dist/chunk-S5YLIBFX.js.map +0 -1
- package/dist/chunk-SZLVEKMJ.js +0 -1446
- package/dist/chunk-SZLVEKMJ.js.map +0 -1
- package/dist/chunk-T4SQEITX.js +0 -95
- package/dist/chunk-T4SQEITX.js.map +0 -1
- package/dist/chunk-TBL77AUT.js +0 -355
- package/dist/chunk-TBL77AUT.js.map +0 -1
- package/dist/chunk-TSN7JT6D.js +0 -1646
- package/dist/chunk-TSN7JT6D.js.map +0 -1
- package/dist/chunk-TT4KNT67.js +0 -124
- package/dist/chunk-TT4KNT67.js.map +0 -1
- package/dist/chunk-UB2LOJ6Q.js +0 -4461
- package/dist/chunk-UB2LOJ6Q.js.map +0 -1
- package/dist/chunk-UWZZKKU7.js +0 -237
- package/dist/chunk-UWZZKKU7.js.map +0 -1
- package/dist/chunk-VBQ3CRKH.js +0 -291
- package/dist/chunk-VBQ3CRKH.js.map +0 -1
- package/dist/chunk-VGRCHJON.js +0 -163
- package/dist/chunk-VGRCHJON.js.map +0 -1
- package/dist/chunk-VI2UW6B6.js +0 -162
- package/dist/chunk-VI2UW6B6.js.map +0 -1
- package/dist/chunk-VLOATJQ2.js +0 -908
- package/dist/chunk-VLOATJQ2.js.map +0 -1
- package/dist/chunk-VQMK5FMP.js +0 -247
- package/dist/chunk-VQMK5FMP.js.map +0 -1
- package/dist/chunk-VZSRQ272.js +0 -149
- package/dist/chunk-VZSRQ272.js.map +0 -1
- package/dist/chunk-WGXIEX7P.js +0 -116
- package/dist/chunk-WGXIEX7P.js.map +0 -1
- package/dist/chunk-WS3NZZQQ.js +0 -929
- package/dist/chunk-WS3NZZQQ.js.map +0 -1
- package/dist/chunk-XDWDC2MP.js +0 -695
- package/dist/chunk-XDWDC2MP.js.map +0 -1
- package/dist/chunk-XPRT64IE.js +0 -766
- package/dist/chunk-XPRT64IE.js.map +0 -1
- package/dist/chunk-YJBNWCAA.js +0 -1056
- package/dist/chunk-YJBNWCAA.js.map +0 -1
- package/dist/chunk-ZET2UAYW.js +0 -89
- package/dist/chunk-ZET2UAYW.js.map +0 -1
- package/dist/chunk-ZUUWPZCV.js +0 -752
- package/dist/chunk-ZUUWPZCV.js.map +0 -1
- package/dist/control.js.map +0 -1
- package/dist/hosted/index.js.map +0 -1
- package/dist/matrix/index.js.map +0 -1
- package/dist/reporting.js.map +0 -1
- package/dist/rollout/index.js.map +0 -1
- package/dist/run-campaign-ISHFZ7FJ.js +0 -17
- package/dist/run-campaign-ISHFZ7FJ.js.map +0 -1
- package/dist/supervisor-run/index.js.map +0 -1
- package/dist/traces.js.map +0 -1
- package/dist/wire/index.js.map +0 -1
package/dist/chunk-DRYIUNWY.js
DELETED
|
@@ -1,622 +0,0 @@
|
|
|
1
|
-
import {
|
|
2
|
-
FAILURE_CLASSES
|
|
3
|
-
} from "./chunk-MA6HLL3S.js";
|
|
4
|
-
|
|
5
|
-
// src/hosted/types.ts
|
|
6
|
-
var HOSTED_WIRE_VERSION = "2026-07-24.v1";
|
|
7
|
-
|
|
8
|
-
// src/hosted/schemas.ts
|
|
9
|
-
import { z } from "zod";
|
|
10
|
-
var finiteNumber = z.number().finite();
|
|
11
|
-
var nonNegativeNumber = finiteNumber.nonnegative();
|
|
12
|
-
var nonNegativeInteger = z.number().int().nonnegative();
|
|
13
|
-
var nonEmptyString = z.string().min(1).refine((value) => value.trim().length > 0, "must not be blank");
|
|
14
|
-
var attributeValue = z.union([z.string(), finiteNumber, z.boolean()]);
|
|
15
|
-
var attributes = z.record(z.string(), attributeValue);
|
|
16
|
-
var FailureClassSchema = z.custom(
|
|
17
|
-
(value) => typeof value === "string" && FAILURE_CLASSES.includes(value),
|
|
18
|
-
`expected one of ${FAILURE_CLASSES.join(", ")}`
|
|
19
|
-
);
|
|
20
|
-
var UINT64_MAX = 18446744073709551615n;
|
|
21
|
-
var UnixNanoTimestampSchema = z.string().regex(/^(0|[1-9][0-9]*)$/, "expected an unsigned base-10 integer string").refine((value) => BigInt(value) <= UINT64_MAX, "must fit in an unsigned 64-bit integer");
|
|
22
|
-
var GainDistributionBinSchema = z.object({
|
|
23
|
-
lo: finiteNumber,
|
|
24
|
-
hi: finiteNumber,
|
|
25
|
-
count: nonNegativeInteger
|
|
26
|
-
}).strict();
|
|
27
|
-
var ScalarDistributionSchema = z.object({
|
|
28
|
-
n: nonNegativeInteger,
|
|
29
|
-
mean: finiteNumber.nullable(),
|
|
30
|
-
p50: finiteNumber.nullable(),
|
|
31
|
-
p95: finiteNumber.nullable(),
|
|
32
|
-
stddev: nonNegativeNumber.nullable(),
|
|
33
|
-
min: finiteNumber.nullable(),
|
|
34
|
-
max: finiteNumber.nullable(),
|
|
35
|
-
histogram: z.array(GainDistributionBinSchema),
|
|
36
|
-
tailRuns: z.array(
|
|
37
|
-
z.object({
|
|
38
|
-
runId: nonEmptyString,
|
|
39
|
-
score: finiteNumber
|
|
40
|
-
}).strict()
|
|
41
|
-
).optional()
|
|
42
|
-
}).strict().superRefine((distribution, ctx) => {
|
|
43
|
-
const values = [
|
|
44
|
-
distribution.mean,
|
|
45
|
-
distribution.p50,
|
|
46
|
-
distribution.p95,
|
|
47
|
-
distribution.stddev,
|
|
48
|
-
distribution.min,
|
|
49
|
-
distribution.max
|
|
50
|
-
];
|
|
51
|
-
if (distribution.n === 0 && values.some((value) => value !== null)) {
|
|
52
|
-
ctx.addIssue({
|
|
53
|
-
code: "custom",
|
|
54
|
-
message: "distribution values must be null when n is 0"
|
|
55
|
-
});
|
|
56
|
-
}
|
|
57
|
-
if (distribution.n > 0 && values.some((value) => value === null)) {
|
|
58
|
-
ctx.addIssue({
|
|
59
|
-
code: "custom",
|
|
60
|
-
message: "distribution values must be numbers when n is greater than 0"
|
|
61
|
-
});
|
|
62
|
-
}
|
|
63
|
-
if (distribution.min !== null && distribution.max !== null && distribution.min > distribution.max) {
|
|
64
|
-
ctx.addIssue({ code: "custom", path: ["min"], message: "min must not exceed max" });
|
|
65
|
-
}
|
|
66
|
-
});
|
|
67
|
-
var TokenUsageInsightSchema = z.object({
|
|
68
|
-
input: ScalarDistributionSchema,
|
|
69
|
-
output: ScalarDistributionSchema,
|
|
70
|
-
reasoning: ScalarDistributionSchema,
|
|
71
|
-
cached: ScalarDistributionSchema,
|
|
72
|
-
cacheWrite: ScalarDistributionSchema,
|
|
73
|
-
totals: z.object({
|
|
74
|
-
input: nonNegativeNumber,
|
|
75
|
-
output: nonNegativeNumber,
|
|
76
|
-
reasoning: nonNegativeNumber,
|
|
77
|
-
cached: nonNegativeNumber,
|
|
78
|
-
cacheWrite: nonNegativeNumber
|
|
79
|
-
}).strict()
|
|
80
|
-
}).strict();
|
|
81
|
-
var ExecutionErrorOutcomeCellSchema = z.object({
|
|
82
|
-
withErrors: nonNegativeInteger,
|
|
83
|
-
withoutErrors: nonNegativeInteger,
|
|
84
|
-
unreported: nonNegativeInteger
|
|
85
|
-
}).strict();
|
|
86
|
-
var ExecutionInsightSchema = z.object({
|
|
87
|
-
durationMs: ScalarDistributionSchema,
|
|
88
|
-
queueMs: ScalarDistributionSchema,
|
|
89
|
-
tokenUsage: TokenUsageInsightSchema,
|
|
90
|
-
aggregateUsage: z.object({
|
|
91
|
-
runs: nonNegativeInteger,
|
|
92
|
-
tokenUsage: TokenUsageInsightSchema,
|
|
93
|
-
costUsd: ScalarDistributionSchema,
|
|
94
|
-
totalCostUsd: nonNegativeNumber
|
|
95
|
-
}).strict(),
|
|
96
|
-
models: z.array(
|
|
97
|
-
z.object({
|
|
98
|
-
model: nonEmptyString,
|
|
99
|
-
runs: nonNegativeInteger
|
|
100
|
-
}).strict()
|
|
101
|
-
),
|
|
102
|
-
modelCalls: z.object({
|
|
103
|
-
runs: nonNegativeInteger,
|
|
104
|
-
events: nonNegativeInteger,
|
|
105
|
-
reportingRuns: nonNegativeInteger
|
|
106
|
-
}).strict(),
|
|
107
|
-
executionErrors: z.object({
|
|
108
|
-
runs: nonNegativeInteger,
|
|
109
|
-
fraction: finiteNumber.min(0).max(1).nullable(),
|
|
110
|
-
events: nonNegativeInteger,
|
|
111
|
-
reportingRuns: nonNegativeInteger,
|
|
112
|
-
errorSpanEvents: nonNegativeInteger,
|
|
113
|
-
errorSpanReportingRuns: nonNegativeInteger,
|
|
114
|
-
byTerminalOutcome: z.object({
|
|
115
|
-
succeeded: ExecutionErrorOutcomeCellSchema,
|
|
116
|
-
failed: ExecutionErrorOutcomeCellSchema,
|
|
117
|
-
cancelled: ExecutionErrorOutcomeCellSchema,
|
|
118
|
-
incomplete: ExecutionErrorOutcomeCellSchema,
|
|
119
|
-
unknown: ExecutionErrorOutcomeCellSchema
|
|
120
|
-
}).strict()
|
|
121
|
-
}).strict(),
|
|
122
|
-
terminalOutcomes: z.object({
|
|
123
|
-
succeeded: nonNegativeInteger,
|
|
124
|
-
failed: nonNegativeInteger,
|
|
125
|
-
cancelled: nonNegativeInteger,
|
|
126
|
-
incomplete: nonNegativeInteger,
|
|
127
|
-
unknown: nonNegativeInteger
|
|
128
|
-
}).strict()
|
|
129
|
-
}).strict();
|
|
130
|
-
var CostProvenanceSummarySchema = z.object({
|
|
131
|
-
observed: z.object({
|
|
132
|
-
n: nonNegativeInteger,
|
|
133
|
-
totalUsd: nonNegativeNumber
|
|
134
|
-
}).strict(),
|
|
135
|
-
estimated: z.object({
|
|
136
|
-
n: nonNegativeInteger,
|
|
137
|
-
totalUsd: nonNegativeNumber
|
|
138
|
-
}).strict(),
|
|
139
|
-
uncaptured: z.object({ n: nonNegativeInteger }).strict(),
|
|
140
|
-
knownFraction: finiteNumber.min(0).max(1)
|
|
141
|
-
}).strict();
|
|
142
|
-
var ParetoFigureSpecSchema = z.object({
|
|
143
|
-
kind: z.literal("pareto-cost-quality"),
|
|
144
|
-
split: z.enum(["search", "holdout"]),
|
|
145
|
-
points: z.array(
|
|
146
|
-
z.object({
|
|
147
|
-
candidateId: nonEmptyString,
|
|
148
|
-
cost: nonNegativeNumber,
|
|
149
|
-
quality: finiteNumber,
|
|
150
|
-
n: nonNegativeInteger,
|
|
151
|
-
onFrontier: z.boolean(),
|
|
152
|
-
gate: z.enum(["promote", "reject"]).optional()
|
|
153
|
-
}).strict()
|
|
154
|
-
),
|
|
155
|
-
axes: z.object({ x: z.literal("costUsd"), y: z.literal("score") }).strict()
|
|
156
|
-
}).strict();
|
|
157
|
-
var ContinuousAgreementSchema = z.object({
|
|
158
|
-
weightedKappa: finiteNumber,
|
|
159
|
-
icc: finiteNumber,
|
|
160
|
-
pearson: finiteNumber,
|
|
161
|
-
spearman: finiteNumber,
|
|
162
|
-
ci: z.object({
|
|
163
|
-
icc: z.tuple([finiteNumber, finiteNumber]),
|
|
164
|
-
weightedKappa: z.tuple([finiteNumber, finiteNumber])
|
|
165
|
-
}).strict(),
|
|
166
|
-
n: nonNegativeInteger,
|
|
167
|
-
raters: nonNegativeInteger
|
|
168
|
-
}).strict();
|
|
169
|
-
var JudgeInsightSchema = z.object({
|
|
170
|
-
n: nonNegativeInteger,
|
|
171
|
-
meanScore: finiteNumber,
|
|
172
|
-
calibration: ContinuousAgreementSchema.optional(),
|
|
173
|
-
positionalBias: finiteNumber.optional(),
|
|
174
|
-
selfPreference: finiteNumber.optional(),
|
|
175
|
-
verbosityBias: finiteNumber.optional()
|
|
176
|
-
}).strict();
|
|
177
|
-
var InterRaterInsightSchema = z.object({
|
|
178
|
-
raters: nonNegativeInteger,
|
|
179
|
-
jointlyRated: nonNegativeInteger,
|
|
180
|
-
kappa: finiteNumber,
|
|
181
|
-
icc: finiteNumber,
|
|
182
|
-
pearson: finiteNumber,
|
|
183
|
-
spearman: finiteNumber,
|
|
184
|
-
perPair: z.record(z.string(), finiteNumber),
|
|
185
|
-
disagreementCases: z.array(
|
|
186
|
-
z.object({
|
|
187
|
-
runId: nonEmptyString,
|
|
188
|
-
ratings: z.array(
|
|
189
|
-
z.object({
|
|
190
|
-
rater: nonEmptyString,
|
|
191
|
-
score: finiteNumber
|
|
192
|
-
}).strict()
|
|
193
|
-
),
|
|
194
|
-
range: nonNegativeNumber
|
|
195
|
-
}).strict()
|
|
196
|
-
)
|
|
197
|
-
}).strict();
|
|
198
|
-
var LiftInsightSchema = z.object({
|
|
199
|
-
baselineMean: finiteNumber,
|
|
200
|
-
candidateMean: finiteNumber,
|
|
201
|
-
delta: finiteNumber,
|
|
202
|
-
ci95: z.tuple([finiteNumber, finiteNumber]),
|
|
203
|
-
pValue: finiteNumber.min(0).max(1),
|
|
204
|
-
n: nonNegativeInteger,
|
|
205
|
-
unpairedBaseline: nonNegativeInteger,
|
|
206
|
-
unpairedCandidate: nonNegativeInteger,
|
|
207
|
-
cohensD: finiteNumber.nullable(),
|
|
208
|
-
mde: nonNegativeNumber,
|
|
209
|
-
requiredN: nonNegativeInteger.nullable()
|
|
210
|
-
}).strict();
|
|
211
|
-
var FailureClusterInsightSchema = z.object({
|
|
212
|
-
clusters: z.array(
|
|
213
|
-
z.object({
|
|
214
|
-
id: nonEmptyString,
|
|
215
|
-
name: nonEmptyString,
|
|
216
|
-
share: finiteNumber.min(0).max(1),
|
|
217
|
-
exemplars: z.array(nonEmptyString).max(5),
|
|
218
|
-
suggestedFix: nonEmptyString.optional()
|
|
219
|
-
}).strict()
|
|
220
|
-
),
|
|
221
|
-
totalFailures: nonNegativeInteger
|
|
222
|
-
}).strict();
|
|
223
|
-
var ContaminationInsightSchema = z.object({
|
|
224
|
-
leaks: nonNegativeInteger,
|
|
225
|
-
holdoutAuditPassed: z.boolean(),
|
|
226
|
-
details: z.array(
|
|
227
|
-
z.object({
|
|
228
|
-
runId: nonEmptyString,
|
|
229
|
-
canary: nonEmptyString,
|
|
230
|
-
matched: nonEmptyString
|
|
231
|
-
}).strict()
|
|
232
|
-
).optional()
|
|
233
|
-
}).strict();
|
|
234
|
-
var OutcomeCorrelationInsightSchema = z.object({
|
|
235
|
-
metric: nonEmptyString,
|
|
236
|
-
n: nonNegativeInteger,
|
|
237
|
-
pearson: finiteNumber,
|
|
238
|
-
spearman: finiteNumber,
|
|
239
|
-
rewardModel: z.object({
|
|
240
|
-
intercept: finiteNumber,
|
|
241
|
-
slope: finiteNumber,
|
|
242
|
-
r2: finiteNumber
|
|
243
|
-
}).strict().optional()
|
|
244
|
-
}).strict();
|
|
245
|
-
var ReleaseSummarySchema = z.object({
|
|
246
|
-
status: z.enum(["pass", "warn", "fail"]),
|
|
247
|
-
axes: z.array(
|
|
248
|
-
z.object({
|
|
249
|
-
name: z.enum(["quality-lift", "contamination", "composite-distribution"]),
|
|
250
|
-
status: z.enum(["pass", "warn", "fail", "not_evaluated"]),
|
|
251
|
-
detail: nonEmptyString
|
|
252
|
-
}).strict()
|
|
253
|
-
),
|
|
254
|
-
issues: z.array(z.string())
|
|
255
|
-
}).strict();
|
|
256
|
-
var MetricDeltaSchema = z.object({
|
|
257
|
-
current: finiteNumber,
|
|
258
|
-
baseline: finiteNumber,
|
|
259
|
-
delta: finiteNumber,
|
|
260
|
-
ci95: z.tuple([finiteNumber, finiteNumber]),
|
|
261
|
-
pValue: finiteNumber.min(0).max(1),
|
|
262
|
-
cohensD: finiteNumber,
|
|
263
|
-
baselineN: nonNegativeInteger,
|
|
264
|
-
currentN: nonNegativeInteger,
|
|
265
|
-
significant: z.boolean()
|
|
266
|
-
}).strict();
|
|
267
|
-
var PriorPeriodComparisonSchema = z.object({
|
|
268
|
-
baselineN: nonNegativeInteger,
|
|
269
|
-
currentN: nonNegativeInteger,
|
|
270
|
-
windowLabel: nonEmptyString.optional(),
|
|
271
|
-
metrics: z.record(z.string(), MetricDeltaSchema),
|
|
272
|
-
regressedMetrics: z.array(nonEmptyString),
|
|
273
|
-
improvedMetrics: z.array(nonEmptyString)
|
|
274
|
-
}).strict();
|
|
275
|
-
var RecommendationSchema = z.object({
|
|
276
|
-
priority: z.enum(["critical", "high", "medium", "low"]),
|
|
277
|
-
kind: z.enum(["ship", "hold", "investigate", "fix", "recalibrate", "expand-corpus"]),
|
|
278
|
-
title: nonEmptyString,
|
|
279
|
-
detail: nonEmptyString,
|
|
280
|
-
evidencePath: nonEmptyString.optional()
|
|
281
|
-
}).strict();
|
|
282
|
-
var InsightReportSchema = z.object({
|
|
283
|
-
n: nonNegativeInteger,
|
|
284
|
-
execution: ExecutionInsightSchema,
|
|
285
|
-
composite: ScalarDistributionSchema,
|
|
286
|
-
perDimension: z.record(z.string(), ScalarDistributionSchema),
|
|
287
|
-
costQuality: z.object({
|
|
288
|
-
cost: ScalarDistributionSchema,
|
|
289
|
-
pareto: ParetoFigureSpecSchema,
|
|
290
|
-
provenance: CostProvenanceSummarySchema.optional(),
|
|
291
|
-
degraded: z.object({
|
|
292
|
-
cost: nonEmptyString.optional(),
|
|
293
|
-
pareto: nonEmptyString.optional()
|
|
294
|
-
}).strict().optional()
|
|
295
|
-
}).strict(),
|
|
296
|
-
judges: z.record(z.string(), JudgeInsightSchema),
|
|
297
|
-
interRater: InterRaterInsightSchema.optional(),
|
|
298
|
-
lift: LiftInsightSchema.optional(),
|
|
299
|
-
failureClusters: FailureClusterInsightSchema.optional(),
|
|
300
|
-
contamination: ContaminationInsightSchema.optional(),
|
|
301
|
-
outcomeCorrelation: OutcomeCorrelationInsightSchema.optional(),
|
|
302
|
-
release: ReleaseSummarySchema,
|
|
303
|
-
priorPeriodComparison: PriorPeriodComparisonSchema.optional(),
|
|
304
|
-
failureClasses: z.array(
|
|
305
|
-
z.object({
|
|
306
|
-
failureClass: FailureClassSchema,
|
|
307
|
-
count: nonNegativeInteger,
|
|
308
|
-
share: finiteNumber.min(0).max(1)
|
|
309
|
-
}).strict()
|
|
310
|
-
).optional(),
|
|
311
|
-
recommendations: z.array(RecommendationSchema)
|
|
312
|
-
}).strict();
|
|
313
|
-
var sha256Digest = z.custom(
|
|
314
|
-
(value) => typeof value === "string" && /^sha256:[0-9a-f]{64}$/.test(value),
|
|
315
|
-
"expected sha256:<64 lowercase hex characters>"
|
|
316
|
-
);
|
|
317
|
-
var MutableSurfaceSchema = z.union([
|
|
318
|
-
z.string(),
|
|
319
|
-
z.object({
|
|
320
|
-
kind: z.literal("components"),
|
|
321
|
-
components: z.record(z.string(), z.string())
|
|
322
|
-
}).strict(),
|
|
323
|
-
z.object({
|
|
324
|
-
kind: z.literal("code"),
|
|
325
|
-
worktreeRef: nonEmptyString,
|
|
326
|
-
baseRef: nonEmptyString,
|
|
327
|
-
baseCommit: nonEmptyString,
|
|
328
|
-
baseTree: nonEmptyString,
|
|
329
|
-
candidateCommit: nonEmptyString,
|
|
330
|
-
candidateTree: nonEmptyString,
|
|
331
|
-
patch: z.object({
|
|
332
|
-
format: z.literal("git-diff-binary"),
|
|
333
|
-
sha256: sha256Digest,
|
|
334
|
-
byteLength: nonNegativeInteger
|
|
335
|
-
}).strict(),
|
|
336
|
-
summary: z.string().optional()
|
|
337
|
-
}).strict()
|
|
338
|
-
]);
|
|
339
|
-
var RunTerminalOutcomeSchema = z.enum([
|
|
340
|
-
"succeeded",
|
|
341
|
-
"failed",
|
|
342
|
-
"cancelled",
|
|
343
|
-
"incomplete",
|
|
344
|
-
"unknown"
|
|
345
|
-
]);
|
|
346
|
-
var EvalRunCellScoreSchema = z.object({
|
|
347
|
-
scenarioId: nonEmptyString,
|
|
348
|
-
rep: nonNegativeInteger,
|
|
349
|
-
compositeMean: finiteNumber.nullable(),
|
|
350
|
-
dimensions: z.record(z.string(), z.record(z.string(), finiteNumber)),
|
|
351
|
-
terminalOutcome: RunTerminalOutcomeSchema,
|
|
352
|
-
executionErrorCount: nonNegativeInteger.nullable(),
|
|
353
|
-
errorMessage: z.string().optional()
|
|
354
|
-
}).strict();
|
|
355
|
-
var EvalRunGenerationSnapshotSchema = z.object({
|
|
356
|
-
index: nonNegativeInteger,
|
|
357
|
-
surfaceHash: nonEmptyString,
|
|
358
|
-
surface: MutableSurfaceSchema.optional(),
|
|
359
|
-
cells: z.array(EvalRunCellScoreSchema),
|
|
360
|
-
compositeMean: finiteNumber.nullable(),
|
|
361
|
-
costUsd: nonNegativeNumber,
|
|
362
|
-
durationMs: nonNegativeNumber
|
|
363
|
-
}).strict();
|
|
364
|
-
var EvalRunStatusSchema = z.enum([
|
|
365
|
-
"started",
|
|
366
|
-
"baseline-complete",
|
|
367
|
-
"generation-complete",
|
|
368
|
-
"gate-decided",
|
|
369
|
-
"finished",
|
|
370
|
-
"errored"
|
|
371
|
-
]);
|
|
372
|
-
var GateDecisionSchema = z.enum([
|
|
373
|
-
"ship",
|
|
374
|
-
"hold",
|
|
375
|
-
"need_more_work",
|
|
376
|
-
"model_ceiling",
|
|
377
|
-
"arch_ceiling"
|
|
378
|
-
]);
|
|
379
|
-
var EvalRunEventSchema = z.object({
|
|
380
|
-
runId: nonEmptyString,
|
|
381
|
-
runDir: nonEmptyString,
|
|
382
|
-
timestamp: z.string().datetime({ offset: true }),
|
|
383
|
-
status: EvalRunStatusSchema,
|
|
384
|
-
labels: z.record(z.string(), z.string()),
|
|
385
|
-
baseline: EvalRunGenerationSnapshotSchema.optional(),
|
|
386
|
-
generations: z.array(EvalRunGenerationSnapshotSchema),
|
|
387
|
-
gateDecision: GateDecisionSchema.optional(),
|
|
388
|
-
holdoutLift: finiteNumber.optional(),
|
|
389
|
-
totalCostUsd: nonNegativeNumber,
|
|
390
|
-
totalDurationMs: nonNegativeNumber,
|
|
391
|
-
errorMessage: z.string().optional(),
|
|
392
|
-
insightReport: InsightReportSchema.optional()
|
|
393
|
-
}).strict().superRefine((event, ctx) => {
|
|
394
|
-
if (event.baseline && event.baseline.index !== 0) {
|
|
395
|
-
ctx.addIssue({
|
|
396
|
-
code: "custom",
|
|
397
|
-
path: ["baseline", "index"],
|
|
398
|
-
message: "baseline index must be 0"
|
|
399
|
-
});
|
|
400
|
-
}
|
|
401
|
-
const seen = /* @__PURE__ */ new Set();
|
|
402
|
-
for (let i = 0; i < event.generations.length; i++) {
|
|
403
|
-
const index = event.generations[i].index;
|
|
404
|
-
if (seen.has(index)) {
|
|
405
|
-
ctx.addIssue({
|
|
406
|
-
code: "custom",
|
|
407
|
-
path: ["generations", i, "index"],
|
|
408
|
-
message: `duplicate generation index ${index}`
|
|
409
|
-
});
|
|
410
|
-
}
|
|
411
|
-
seen.add(index);
|
|
412
|
-
}
|
|
413
|
-
if (event.status === "errored" && !event.errorMessage?.trim()) {
|
|
414
|
-
ctx.addIssue({
|
|
415
|
-
code: "custom",
|
|
416
|
-
path: ["errorMessage"],
|
|
417
|
-
message: "errorMessage is required when status is errored"
|
|
418
|
-
});
|
|
419
|
-
}
|
|
420
|
-
});
|
|
421
|
-
var TraceSpanEventEntrySchema = z.object({
|
|
422
|
-
timeUnixNano: UnixNanoTimestampSchema,
|
|
423
|
-
name: nonEmptyString,
|
|
424
|
-
attributes: attributes.optional()
|
|
425
|
-
}).strict();
|
|
426
|
-
var TraceSpanEventSchema = z.object({
|
|
427
|
-
traceId: nonEmptyString,
|
|
428
|
-
spanId: nonEmptyString,
|
|
429
|
-
parentSpanId: nonEmptyString.optional(),
|
|
430
|
-
name: nonEmptyString,
|
|
431
|
-
startTimeUnixNano: UnixNanoTimestampSchema,
|
|
432
|
-
endTimeUnixNano: UnixNanoTimestampSchema,
|
|
433
|
-
attributes,
|
|
434
|
-
events: z.array(TraceSpanEventEntrySchema).optional(),
|
|
435
|
-
status: z.object({
|
|
436
|
-
code: z.enum(["OK", "ERROR", "UNSET"]),
|
|
437
|
-
message: z.string().optional()
|
|
438
|
-
}).strict().optional(),
|
|
439
|
-
"tangle.runId": nonEmptyString.optional(),
|
|
440
|
-
"tangle.generation": nonNegativeInteger.optional(),
|
|
441
|
-
"tangle.cellId": nonEmptyString.optional(),
|
|
442
|
-
"tangle.scenarioId": nonEmptyString.optional()
|
|
443
|
-
}).strict().refine((span) => BigInt(span.endTimeUnixNano) >= BigInt(span.startTimeUnixNano), {
|
|
444
|
-
path: ["endTimeUnixNano"],
|
|
445
|
-
message: "endTimeUnixNano must be greater than or equal to startTimeUnixNano"
|
|
446
|
-
});
|
|
447
|
-
var IngestEvalRunsEnvelopeSchema = z.object({
|
|
448
|
-
wireVersion: z.literal(HOSTED_WIRE_VERSION),
|
|
449
|
-
events: z.array(z.unknown())
|
|
450
|
-
}).strict();
|
|
451
|
-
var IngestTracesEnvelopeSchema = z.object({
|
|
452
|
-
wireVersion: z.literal(HOSTED_WIRE_VERSION),
|
|
453
|
-
spans: z.array(z.unknown())
|
|
454
|
-
}).strict();
|
|
455
|
-
var IngestEvalRunsRequestSchema = z.object({
|
|
456
|
-
wireVersion: z.literal(HOSTED_WIRE_VERSION),
|
|
457
|
-
events: z.array(EvalRunEventSchema)
|
|
458
|
-
}).strict();
|
|
459
|
-
var IngestTracesRequestSchema = z.object({
|
|
460
|
-
wireVersion: z.literal(HOSTED_WIRE_VERSION),
|
|
461
|
-
spans: z.array(TraceSpanEventSchema)
|
|
462
|
-
}).strict();
|
|
463
|
-
var IngestResponseSchema = z.object({
|
|
464
|
-
accepted: nonNegativeInteger,
|
|
465
|
-
rejected: z.array(
|
|
466
|
-
z.object({
|
|
467
|
-
index: nonNegativeInteger,
|
|
468
|
-
reason: nonEmptyString
|
|
469
|
-
}).strict()
|
|
470
|
-
)
|
|
471
|
-
}).strict();
|
|
472
|
-
|
|
473
|
-
// src/hosted/client.ts
|
|
474
|
-
var MAX_IDEMPOTENCY_KEY_LENGTH = 256;
|
|
475
|
-
function sleep(ms) {
|
|
476
|
-
return new Promise((resolve) => setTimeout(resolve, ms));
|
|
477
|
-
}
|
|
478
|
-
function normalizeHostedBase(endpoint) {
|
|
479
|
-
return endpoint.trim().replace(/\/+$/, "").replace(/\/v1$/, "");
|
|
480
|
-
}
|
|
481
|
-
function resolveIdempotencyKey(key) {
|
|
482
|
-
const resolved = key ?? globalThis.crypto.randomUUID();
|
|
483
|
-
if (resolved.trim().length === 0) throw new Error("idempotency key must not be blank");
|
|
484
|
-
if (resolved.length > MAX_IDEMPOTENCY_KEY_LENGTH) {
|
|
485
|
-
throw new Error(`idempotency key must be at most ${MAX_IDEMPOTENCY_KEY_LENGTH} characters`);
|
|
486
|
-
}
|
|
487
|
-
return resolved;
|
|
488
|
-
}
|
|
489
|
-
function responseValidationReason(error) {
|
|
490
|
-
return error.issues.map(
|
|
491
|
-
(issue) => `${issue.path.length > 0 ? issue.path.map(String).join(".") : "value"}: ${issue.message}`
|
|
492
|
-
).join("; ");
|
|
493
|
-
}
|
|
494
|
-
async function post(tenant, path, body, opts = {}) {
|
|
495
|
-
const timeoutMs = tenant.timeoutMs ?? 3e4;
|
|
496
|
-
const maxRetries = tenant.retries ?? 2;
|
|
497
|
-
const f = tenant.fetchImpl ?? ((...args) => fetch(...args));
|
|
498
|
-
const base = normalizeHostedBase(tenant.endpoint);
|
|
499
|
-
const url = `${base}${path}`;
|
|
500
|
-
const idempotencyKey = resolveIdempotencyKey(opts.idempotencyKey);
|
|
501
|
-
const headers = {
|
|
502
|
-
"content-type": "application/json",
|
|
503
|
-
authorization: `Bearer ${tenant.apiKey}`,
|
|
504
|
-
"x-tangle-tenant-id": tenant.tenantId,
|
|
505
|
-
"x-tangle-wire-version": HOSTED_WIRE_VERSION,
|
|
506
|
-
"idempotency-key": idempotencyKey
|
|
507
|
-
};
|
|
508
|
-
let lastError;
|
|
509
|
-
for (let attempt = 0; attempt <= maxRetries; attempt++) {
|
|
510
|
-
const ourTimeout = AbortSignal.timeout(timeoutMs);
|
|
511
|
-
const combinedSignal = opts.signal ? AbortSignal.any([opts.signal, ourTimeout]) : ourTimeout;
|
|
512
|
-
let res;
|
|
513
|
-
try {
|
|
514
|
-
res = await f(url, {
|
|
515
|
-
method: "POST",
|
|
516
|
-
headers,
|
|
517
|
-
body: JSON.stringify(body),
|
|
518
|
-
signal: combinedSignal
|
|
519
|
-
});
|
|
520
|
-
} catch (err) {
|
|
521
|
-
if (opts.signal?.aborted) throw err;
|
|
522
|
-
lastError = err;
|
|
523
|
-
if (attempt === maxRetries) throw err;
|
|
524
|
-
await sleep(2 ** attempt * 200 + Math.random() * 200);
|
|
525
|
-
continue;
|
|
526
|
-
}
|
|
527
|
-
if (!res.ok) {
|
|
528
|
-
const text = await res.text().catch(() => "");
|
|
529
|
-
const error = new Error(`hosted ingest ${url} failed (${res.status}): ${text.slice(0, 500)}`);
|
|
530
|
-
const retryable = res.status >= 500 || res.status === 408 || res.status === 429;
|
|
531
|
-
if (!retryable || attempt === maxRetries) throw error;
|
|
532
|
-
lastError = error;
|
|
533
|
-
await sleep(2 ** attempt * 200 + Math.random() * 200);
|
|
534
|
-
continue;
|
|
535
|
-
}
|
|
536
|
-
let rawResponse;
|
|
537
|
-
try {
|
|
538
|
-
rawResponse = await res.json();
|
|
539
|
-
} catch (error) {
|
|
540
|
-
throw new Error(`hosted ingest ${url} returned invalid JSON`, { cause: error });
|
|
541
|
-
}
|
|
542
|
-
const parsed = IngestResponseSchema.safeParse(rawResponse);
|
|
543
|
-
if (!parsed.success) {
|
|
544
|
-
throw new Error(
|
|
545
|
-
`hosted ingest ${url} returned an invalid response: ${responseValidationReason(parsed.error)}`
|
|
546
|
-
);
|
|
547
|
-
}
|
|
548
|
-
return parsed.data;
|
|
549
|
-
}
|
|
550
|
-
throw lastError ?? new Error("hosted ingest exhausted retries");
|
|
551
|
-
}
|
|
552
|
-
function createHostedClient(tenant) {
|
|
553
|
-
if (normalizeHostedBase(tenant.endpoint).length === 0) throw new Error("endpoint is required");
|
|
554
|
-
if (tenant.apiKey.trim().length === 0) throw new Error("apiKey is required");
|
|
555
|
-
if (tenant.tenantId.trim().length === 0) throw new Error("tenantId is required");
|
|
556
|
-
if (tenant.timeoutMs !== void 0 && (!Number.isFinite(tenant.timeoutMs) || tenant.timeoutMs <= 0)) {
|
|
557
|
-
throw new Error("timeoutMs must be greater than 0");
|
|
558
|
-
}
|
|
559
|
-
if (tenant.retries !== void 0 && (!Number.isInteger(tenant.retries) || tenant.retries < 0)) {
|
|
560
|
-
throw new Error("retries must be a non-negative integer");
|
|
561
|
-
}
|
|
562
|
-
return {
|
|
563
|
-
tenant,
|
|
564
|
-
wireVersion: HOSTED_WIRE_VERSION,
|
|
565
|
-
async ingestEvalRun(event, idempotencyKey) {
|
|
566
|
-
return this.ingestEvalRuns([event], idempotencyKey);
|
|
567
|
-
},
|
|
568
|
-
async ingestEvalRuns(events, idempotencyKey) {
|
|
569
|
-
const body = IngestEvalRunsRequestSchema.parse({
|
|
570
|
-
wireVersion: HOSTED_WIRE_VERSION,
|
|
571
|
-
events
|
|
572
|
-
});
|
|
573
|
-
return post(tenant, "/v1/ingest/eval-runs", body, {
|
|
574
|
-
idempotencyKey
|
|
575
|
-
});
|
|
576
|
-
},
|
|
577
|
-
async ingestTraces(spans, idempotencyKey) {
|
|
578
|
-
const body = IngestTracesRequestSchema.parse({
|
|
579
|
-
wireVersion: HOSTED_WIRE_VERSION,
|
|
580
|
-
spans
|
|
581
|
-
});
|
|
582
|
-
return post(tenant, "/v1/ingest/traces", body, {
|
|
583
|
-
idempotencyKey
|
|
584
|
-
});
|
|
585
|
-
}
|
|
586
|
-
};
|
|
587
|
-
}
|
|
588
|
-
function hostedTenantFromEnv(overrides = {}) {
|
|
589
|
-
const env = overrides.env ?? process.env;
|
|
590
|
-
const endpoint = (overrides.endpoint ?? env.TANGLE_INGEST_URL ?? env.TANGLE_ORCHESTRATOR_URL)?.trim();
|
|
591
|
-
const apiKey = (overrides.apiKey ?? env.TANGLE_INGEST_API_KEY ?? env.TANGLE_API_KEY)?.trim();
|
|
592
|
-
const tenantId = (overrides.tenantId ?? env.TANGLE_TENANT_ID)?.trim();
|
|
593
|
-
if (!endpoint || !apiKey || !tenantId) return void 0;
|
|
594
|
-
const tenant = { endpoint: endpoint.replace(/\/+$/, ""), apiKey, tenantId };
|
|
595
|
-
if (overrides.fetchImpl) tenant.fetchImpl = overrides.fetchImpl;
|
|
596
|
-
if (overrides.timeoutMs !== void 0) tenant.timeoutMs = overrides.timeoutMs;
|
|
597
|
-
if (overrides.retries !== void 0) tenant.retries = overrides.retries;
|
|
598
|
-
return tenant;
|
|
599
|
-
}
|
|
600
|
-
function hostedClientFromEnv(overrides = {}) {
|
|
601
|
-
const tenant = hostedTenantFromEnv(overrides);
|
|
602
|
-
return tenant ? createHostedClient(tenant) : void 0;
|
|
603
|
-
}
|
|
604
|
-
|
|
605
|
-
export {
|
|
606
|
-
HOSTED_WIRE_VERSION,
|
|
607
|
-
UnixNanoTimestampSchema,
|
|
608
|
-
InsightReportSchema,
|
|
609
|
-
MutableSurfaceSchema,
|
|
610
|
-
RunTerminalOutcomeSchema,
|
|
611
|
-
EvalRunCellScoreSchema,
|
|
612
|
-
EvalRunGenerationSnapshotSchema,
|
|
613
|
-
EvalRunEventSchema,
|
|
614
|
-
TraceSpanEventSchema,
|
|
615
|
-
IngestEvalRunsRequestSchema,
|
|
616
|
-
IngestTracesRequestSchema,
|
|
617
|
-
IngestResponseSchema,
|
|
618
|
-
createHostedClient,
|
|
619
|
-
hostedTenantFromEnv,
|
|
620
|
-
hostedClientFromEnv
|
|
621
|
-
};
|
|
622
|
-
//# sourceMappingURL=chunk-DRYIUNWY.js.map
|