@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
|
@@ -1,121 +1,20 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
3
|
-
interface SpanBase {
|
|
4
|
-
spanId: string;
|
|
5
|
-
parentSpanId?: string;
|
|
6
|
-
runId: string;
|
|
7
|
-
kind: SpanKind;
|
|
8
|
-
name: string;
|
|
9
|
-
startedAt: number;
|
|
10
|
-
endedAt?: number;
|
|
11
|
-
status?: SpanStatus;
|
|
12
|
-
error?: string;
|
|
13
|
-
/** Anything not covered by typed fields. Kept deliberately free-form. */
|
|
14
|
-
attributes?: Record<string, unknown>;
|
|
15
|
-
}
|
|
16
|
-
interface Message {
|
|
17
|
-
role: 'system' | 'user' | 'assistant' | 'tool';
|
|
18
|
-
content: string;
|
|
19
|
-
tokens?: number;
|
|
20
|
-
/** Multi-modal content descriptors; blobs themselves live in Artifacts. */
|
|
21
|
-
images?: Array<{
|
|
22
|
-
artifactId?: string;
|
|
23
|
-
url?: string;
|
|
24
|
-
mime?: string;
|
|
25
|
-
}>;
|
|
26
|
-
}
|
|
27
|
-
interface LlmSpan extends SpanBase {
|
|
28
|
-
kind: 'llm';
|
|
29
|
-
model: string;
|
|
30
|
-
messages: Message[];
|
|
31
|
-
output?: string;
|
|
32
|
-
inputTokens?: number;
|
|
33
|
-
/** All generated tokens, including the reasoning subset when present. */
|
|
34
|
-
outputTokens?: number;
|
|
35
|
-
cachedTokens?: number;
|
|
36
|
-
cacheWriteTokens?: number;
|
|
37
|
-
/** Reasoning-token subset of `outputTokens`. */
|
|
38
|
-
reasoningTokens?: number;
|
|
39
|
-
costUsd?: number;
|
|
40
|
-
finishReason?: string;
|
|
41
|
-
}
|
|
42
|
-
interface ToolSpan extends SpanBase {
|
|
43
|
-
kind: 'tool';
|
|
44
|
-
toolName: string;
|
|
45
|
-
args: unknown;
|
|
46
|
-
/** False when the source observed the call but did not capture its arguments. */
|
|
47
|
-
argsCaptured?: boolean;
|
|
48
|
-
result?: unknown;
|
|
49
|
-
latencyMs?: number;
|
|
50
|
-
}
|
|
51
|
-
interface RetrievalSpan extends SpanBase {
|
|
52
|
-
kind: 'retrieval';
|
|
53
|
-
query: string;
|
|
54
|
-
hits: Array<{
|
|
55
|
-
docId: string;
|
|
56
|
-
score: number;
|
|
57
|
-
content?: string;
|
|
58
|
-
}>;
|
|
59
|
-
}
|
|
60
|
-
interface JudgeSpan extends SpanBase {
|
|
61
|
-
kind: 'judge';
|
|
62
|
-
judgeId: string;
|
|
63
|
-
/** Span this judgment applies to. */
|
|
64
|
-
targetSpanId: string;
|
|
65
|
-
dimension: string;
|
|
66
|
-
/** Numeric score (free-range; interpretation up to the judge). */
|
|
67
|
-
score: number;
|
|
68
|
-
rationale?: string;
|
|
69
|
-
evidence?: string;
|
|
70
|
-
}
|
|
71
|
-
interface SandboxSpan extends SpanBase {
|
|
72
|
-
kind: 'sandbox';
|
|
73
|
-
image?: string;
|
|
74
|
-
command?: string;
|
|
75
|
-
exitCode?: number;
|
|
76
|
-
testsTotal?: number;
|
|
77
|
-
testsPassed?: number;
|
|
78
|
-
stdoutHash?: string;
|
|
79
|
-
stderrHash?: string;
|
|
80
|
-
/** Duration in ms; the harness fills this explicitly (endedAt - startedAt may miss setup). */
|
|
81
|
-
wallMs?: number;
|
|
82
|
-
}
|
|
83
|
-
interface GenericSpan extends SpanBase {
|
|
84
|
-
kind: 'agent' | 'custom';
|
|
85
|
-
}
|
|
86
|
-
type Span = LlmSpan | ToolSpan | RetrievalSpan | JudgeSpan | SandboxSpan | GenericSpan;
|
|
87
|
-
|
|
88
|
-
/**
|
|
89
|
-
* Code-edit extraction + the "watch the agent write code" animation.
|
|
90
|
-
*
|
|
91
|
-
* The base storyboard (./index) compresses a run into scenes but keeps scenes
|
|
92
|
-
* payload-free (title + narration + evidence span ids). To actually SHOW the
|
|
93
|
-
* code an agent wrote, this module pulls the concrete edit (path, diff,
|
|
94
|
-
* before/after, language) out of the raw edit-tool spans and renders a
|
|
95
|
-
* self-contained HTML replay that types each file out, character by character,
|
|
96
|
-
* like an editor recording.
|
|
97
|
-
*
|
|
98
|
-
* Additive on purpose: it reuses the canonical `Span` (trace/schema) and the
|
|
99
|
-
* `Storyboard` IR (./index) without modifying the reducer/compiler — so it
|
|
100
|
-
* composes with the existing pipeline and a Remotion consumer can read the same
|
|
101
|
-
* `CodeEdit[]` to render an MP4.
|
|
102
|
-
*/
|
|
103
|
-
|
|
1
|
+
import { _ as Span } from "../schema-BtVldJ3T.js";
|
|
2
|
+
//#region src/storyboard/code-edit.d.ts
|
|
104
3
|
/** A concrete code change lifted from a tool span. */
|
|
105
4
|
interface CodeEdit {
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
5
|
+
/** Span the edit came from — back-reference into the trace. */
|
|
6
|
+
spanId: string;
|
|
7
|
+
path: string;
|
|
8
|
+
language?: string;
|
|
9
|
+
/** Unified diff, when the tool carried one. */
|
|
10
|
+
diff?: string;
|
|
11
|
+
/** Full file body before / after, when present — enables true keystroke
|
|
12
|
+
* animation rather than only flashing a diff. */
|
|
13
|
+
before?: string;
|
|
14
|
+
after?: string;
|
|
15
|
+
additions: number;
|
|
16
|
+
deletions: number;
|
|
17
|
+
startedAt: number;
|
|
119
18
|
}
|
|
120
19
|
/** Extract a CodeEdit from a single span, or undefined if it is not an edit. */
|
|
121
20
|
declare function codeEditFromSpan(span: Span): CodeEdit | undefined;
|
|
@@ -124,20 +23,20 @@ declare function extractCodeEdits(spans: readonly Span[]): CodeEdit[];
|
|
|
124
23
|
/** Pair each code-edit (`diff`) scene with its concrete edit, matched via the
|
|
125
24
|
* scene's evidence span ids. Scenes without a recoverable edit are dropped. */
|
|
126
25
|
declare function codeEditsForStoryboard(storyboard: Storyboard, spans: readonly Span[]): Array<{
|
|
127
|
-
|
|
128
|
-
|
|
26
|
+
scene: Scene;
|
|
27
|
+
edit: CodeEdit;
|
|
129
28
|
}>;
|
|
130
29
|
/** The text the animation types for an edit: prefer the full new file body,
|
|
131
30
|
* then the added lines of a diff, then a minimal placeholder so the scene is
|
|
132
31
|
* never blank. */
|
|
133
32
|
declare function editAnimationText(edit: CodeEdit): string;
|
|
134
33
|
interface CodeAnimationOptions {
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
34
|
+
/** Page title. Default "Agent writing code". */
|
|
35
|
+
title?: string;
|
|
36
|
+
/** Characters typed per animation tick. Higher = faster. Default 4. */
|
|
37
|
+
charsPerTick?: number;
|
|
38
|
+
/** Pause between files, ms. Default 900. */
|
|
39
|
+
pauseBetweenMs?: number;
|
|
141
40
|
}
|
|
142
41
|
/**
|
|
143
42
|
* Render a self-contained HTML page that animates the agent writing each file,
|
|
@@ -147,83 +46,64 @@ interface CodeAnimationOptions {
|
|
|
147
46
|
* package; this is the dependency-free baseline.
|
|
148
47
|
*/
|
|
149
48
|
declare function renderCodeAnimationHtml(source: readonly Span[] | readonly CodeEdit[], opts?: CodeAnimationOptions): string;
|
|
150
|
-
|
|
151
|
-
|
|
152
|
-
* Storyboard — compile a raw agent trace into a small set of meaningful scenes,
|
|
153
|
-
* then render them as a shareable, watchable replay.
|
|
154
|
-
*
|
|
155
|
-
* Span[] → SemanticEvent[] → Storyboard{scenes} → { markdown | html }
|
|
156
|
-
*
|
|
157
|
-
* The trace is the source of truth; the storyboard is the IR; markdown/html
|
|
158
|
-
* (and, in a consumer package, Remotion/MP4) are compiled targets. Everything
|
|
159
|
-
* here is PURE — same spans in, same bytes out (no clock/random) — so a run's
|
|
160
|
-
* replay is deterministic and diffable.
|
|
161
|
-
*
|
|
162
|
-
* Crucially, every scene carries a modality-typed VISUAL extracted from the
|
|
163
|
-
* span — a browser/computer screenshot, a code diff, a terminal command +
|
|
164
|
-
* output, file content, an API request/response, or reasoning prose — so no
|
|
165
|
-
* matter what the agent did, the replay SHOWS it, not just narrates it. The
|
|
166
|
-
* reducer is rules-based; an LLM pass can refine titles/summaries later, but
|
|
167
|
-
* the structure + visuals do not depend on one. Renderers that need a browser
|
|
168
|
-
* (Remotion, React) live in consumers; this module only emits strings.
|
|
169
|
-
*/
|
|
170
|
-
|
|
49
|
+
//#endregion
|
|
50
|
+
//#region src/storyboard/index.d.ts
|
|
171
51
|
/** What the agent meaningfully did — the compressed vocabulary. */
|
|
172
52
|
type SemanticKind = 'understood_task' | 'user_message' | 'agent_reply' | 'reasoned' | 'ran_command' | 'read_file' | 'edited_code' | 'searched' | 'used_browser' | 'used_computer' | 'called_api' | 'called_tool' | 'evaluated' | 'observed_failure' | 'completed';
|
|
173
53
|
/** The modality-typed payload a scene shows. This is what makes any agent
|
|
174
54
|
* action — screen, browser, shell, code, API — actually viewable. */
|
|
175
55
|
type SceneVisual = {
|
|
176
|
-
|
|
177
|
-
|
|
178
|
-
|
|
56
|
+
type: 'screenshot';
|
|
57
|
+
src: string;
|
|
58
|
+
caption?: string;
|
|
179
59
|
} | {
|
|
180
|
-
|
|
181
|
-
|
|
182
|
-
|
|
183
|
-
|
|
60
|
+
type: 'browser';
|
|
61
|
+
url?: string;
|
|
62
|
+
action?: string;
|
|
63
|
+
screenshot?: string;
|
|
184
64
|
} | {
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
|
|
65
|
+
type: 'diff';
|
|
66
|
+
path: string;
|
|
67
|
+
patch: string;
|
|
188
68
|
} | {
|
|
189
|
-
|
|
190
|
-
|
|
191
|
-
|
|
69
|
+
type: 'code';
|
|
70
|
+
path: string;
|
|
71
|
+
content: string;
|
|
192
72
|
} | {
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
|
|
73
|
+
type: 'terminal';
|
|
74
|
+
command: string;
|
|
75
|
+
output?: string;
|
|
196
76
|
} | {
|
|
197
|
-
|
|
198
|
-
|
|
199
|
-
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
77
|
+
type: 'api';
|
|
78
|
+
method?: string;
|
|
79
|
+
url: string;
|
|
80
|
+
status?: number;
|
|
81
|
+
request?: string;
|
|
82
|
+
response?: string;
|
|
203
83
|
} | {
|
|
204
|
-
|
|
205
|
-
|
|
84
|
+
type: 'prose';
|
|
85
|
+
text: string;
|
|
206
86
|
} | {
|
|
207
|
-
|
|
87
|
+
type: 'none';
|
|
208
88
|
};
|
|
209
89
|
interface SemanticEvent {
|
|
210
|
-
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
214
|
-
|
|
215
|
-
|
|
216
|
-
|
|
217
|
-
|
|
218
|
-
|
|
219
|
-
|
|
220
|
-
|
|
221
|
-
|
|
222
|
-
|
|
90
|
+
kind: SemanticKind;
|
|
91
|
+
/** One-line headline (the ticket/scene title). */
|
|
92
|
+
title: string;
|
|
93
|
+
/** Short human summary. */
|
|
94
|
+
summary: string;
|
|
95
|
+
/** The modality payload to show for this moment. */
|
|
96
|
+
visual: SceneVisual;
|
|
97
|
+
/** Span ids backing this moment — the evidence trail. */
|
|
98
|
+
evidenceSpanIds: string[];
|
|
99
|
+
/** 1 (noise) … 5 (pivotal: failures, edits). Drives selection + duration. */
|
|
100
|
+
importance: 1 | 2 | 3 | 4 | 5;
|
|
101
|
+
startTs: number;
|
|
102
|
+
endTs: number;
|
|
223
103
|
}
|
|
224
104
|
interface ReduceOptions {
|
|
225
|
-
|
|
226
|
-
|
|
105
|
+
/** Collapse adjacent same-kind moments into one. Default true. */
|
|
106
|
+
collapseAdjacent?: boolean;
|
|
227
107
|
}
|
|
228
108
|
/**
|
|
229
109
|
* Reduce a span tree into the meaningful moments a viewer cares about. The
|
|
@@ -236,24 +116,24 @@ interface ReduceOptions {
|
|
|
236
116
|
declare function reduceToSemanticEvents(spans: readonly Span[], opts?: ReduceOptions): SemanticEvent[];
|
|
237
117
|
type SceneType = 'title_card' | 'prompt' | 'reply' | 'reasoning' | 'terminal' | 'file' | 'diff' | 'search' | 'browser' | 'computer' | 'api' | 'tool' | 'eval' | 'error' | 'summary';
|
|
238
118
|
interface Scene {
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
|
|
242
|
-
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
|
|
119
|
+
sceneType: SceneType;
|
|
120
|
+
title: string;
|
|
121
|
+
narration: string;
|
|
122
|
+
/** The modality payload the renderer shows. */
|
|
123
|
+
visual: SceneVisual;
|
|
124
|
+
durationMs: number;
|
|
125
|
+
evidenceSpanIds: string[];
|
|
246
126
|
}
|
|
247
127
|
interface Storyboard {
|
|
248
|
-
|
|
249
|
-
|
|
250
|
-
|
|
128
|
+
title: string;
|
|
129
|
+
totalMs: number;
|
|
130
|
+
scenes: Scene[];
|
|
251
131
|
}
|
|
252
132
|
interface CompileOptions {
|
|
253
|
-
|
|
254
|
-
|
|
255
|
-
|
|
256
|
-
|
|
133
|
+
/** Headline for the title card. Default "Agent run". */
|
|
134
|
+
title?: string;
|
|
135
|
+
/** Max action scenes between the title + summary cards. Default 16. */
|
|
136
|
+
maxScenes?: number;
|
|
257
137
|
}
|
|
258
138
|
/**
|
|
259
139
|
* Select the moments worth showing and lay them out as scenes. Every failure
|
|
@@ -264,8 +144,8 @@ declare function compileStoryboard(events: readonly SemanticEvent[], opts?: Comp
|
|
|
264
144
|
/** Render the storyboard as a Markdown timeline — the human-readable shot list. */
|
|
265
145
|
declare function renderStoryboardMarkdown(storyboard: Storyboard): string;
|
|
266
146
|
interface HtmlRenderOptions {
|
|
267
|
-
|
|
268
|
-
|
|
147
|
+
/** Document title + on-page heading. Defaults to the storyboard title. */
|
|
148
|
+
title?: string;
|
|
269
149
|
}
|
|
270
150
|
/**
|
|
271
151
|
* Render the storyboard as a self-contained, auto-playing HTML replay — one
|
|
@@ -275,5 +155,6 @@ interface HtmlRenderOptions {
|
|
|
275
155
|
* run produces; an MP4 is the same storyboard fed to Remotion in a consumer.
|
|
276
156
|
*/
|
|
277
157
|
declare function renderStoryboardHtml(storyboard: Storyboard, opts?: HtmlRenderOptions): string;
|
|
278
|
-
|
|
279
|
-
export {
|
|
158
|
+
//#endregion
|
|
159
|
+
export { CodeAnimationOptions, CodeEdit, CompileOptions, HtmlRenderOptions, ReduceOptions, Scene, SceneType, SceneVisual, SemanticEvent, SemanticKind, Storyboard, codeEditFromSpan, codeEditsForStoryboard, compileStoryboard, editAnimationText, extractCodeEdits, reduceToSemanticEvents, renderCodeAnimationHtml, renderStoryboardHtml, renderStoryboardMarkdown };
|
|
160
|
+
//# sourceMappingURL=index.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/storyboard/code-edit.ts","../../src/storyboard/index.ts"],"mappings":";;;UAuBiB;;EAEf;EACA;EACA;;EAEA;;;EAGA;EACA;EACA;EACA;EACA;;;iBA6Dc,iBAAiB,MAAM,OAAO;;iBAwB9B,iBAAiB,gBAAgB,SAAS;;;iBAS1C,uBACd,YAAY,YACZ,gBAAgB,SACf;EAAQ,OAAO;EAAO,MAAM;;;;;iBAgBf,kBAAkB,MAAM;UAiBvB;;EAEf;;EAEA;;EAEA;;;;;;;;;iBAUc,wBACd,iBAAiB,kBAAkB,YACnC,OAAM;;;;KCjKI;;;KAmBA;EACN;EAAoB;EAAa;;EACjC;EAAiB;EAAc;EAAiB;;EAChD;EAAc;EAAc;;EAC5B;EAAc;EAAc;;EAC5B;EAAkB;EAAiB;;EAEnC;EACA;EACA;EACA;EACA;EACA;;EAEA;EAAe;;EACf;;UAEW;EACf,MAAM;;EAEN;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;EACA;EACA;;UAGe;;EAEf;;;;;;;;;;iBAiTc,uBACd,gBAAgB,QAChB,OAAM,gBACL;KA8CS;UAiBK;EACf,WAAW;EACX;EACA;;EAEA,QAAQ;EACR;EACA;;UAGe;EACf;EACA;EACA,QAAQ;;UAGO;;EAEf;;EAEA;;;;;;;iBAkCc,kBACd,iBAAiB,iBACjB,OAAM,iBACL;;iBA4Ja,yBAAyB,YAAY;UAwBpC;;EAEf;;;;;;;;;iBAUc,qBAAqB,YAAY,YAAY,OAAM"}
|