@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-SZLVEKMJ.js
DELETED
|
@@ -1,1446 +0,0 @@
|
|
|
1
|
-
import {
|
|
2
|
-
estimateCost,
|
|
3
|
-
isModelPriced
|
|
4
|
-
} from "./chunk-VI2UW6B6.js";
|
|
5
|
-
import {
|
|
6
|
-
extractUsage
|
|
7
|
-
} from "./chunk-PXE2VKMX.js";
|
|
8
|
-
|
|
9
|
-
// src/contract/intake/code-agent-observation.ts
|
|
10
|
-
function observeCodeAgentSession(options) {
|
|
11
|
-
const entries = options.entries.filter(isRecord);
|
|
12
|
-
const projection = projectionFor(options.source, entries);
|
|
13
|
-
const terminal = terminalWithExecution(projection, options.execution);
|
|
14
|
-
return {
|
|
15
|
-
source: options.source,
|
|
16
|
-
sessionId: sessionIdFromEntries(options.source, entries) ?? fallbackSessionId(options.source, options.sourcePath),
|
|
17
|
-
finalText: projection.finalText,
|
|
18
|
-
terminal,
|
|
19
|
-
actions: projection.actions
|
|
20
|
-
};
|
|
21
|
-
}
|
|
22
|
-
function projectionFor(source, entries) {
|
|
23
|
-
switch (source) {
|
|
24
|
-
case "codex":
|
|
25
|
-
return codexProjection(entries);
|
|
26
|
-
case "claude-code":
|
|
27
|
-
return claudeProjection(entries);
|
|
28
|
-
case "opencode":
|
|
29
|
-
return openCodeProjection(entries);
|
|
30
|
-
case "kimi-code":
|
|
31
|
-
return kimiProjection(entries);
|
|
32
|
-
case "pi":
|
|
33
|
-
return piProjection(entries);
|
|
34
|
-
}
|
|
35
|
-
}
|
|
36
|
-
function codexProjection(entries) {
|
|
37
|
-
const actions = [];
|
|
38
|
-
const calls = /* @__PURE__ */ new Map();
|
|
39
|
-
let finalText = null;
|
|
40
|
-
let terminal = "unknown";
|
|
41
|
-
let explicitTerminal = false;
|
|
42
|
-
for (const entry of entries) {
|
|
43
|
-
const entryType = stringField(entry, "type");
|
|
44
|
-
const payload = record(entry.payload) ?? {};
|
|
45
|
-
const payloadType = stringField(payload, "type");
|
|
46
|
-
const timestampMs2 = timestamp(entry.timestamp);
|
|
47
|
-
const item = record(entry.item);
|
|
48
|
-
if (entryType === "item.started" && item) {
|
|
49
|
-
const itemType = stringField(item, "type");
|
|
50
|
-
if (isCodexActionItem(itemType)) {
|
|
51
|
-
const id = stringField(item, "id") ?? `item-${actions.length}`;
|
|
52
|
-
const action = actionFor({
|
|
53
|
-
id,
|
|
54
|
-
stepIndex: actions.length,
|
|
55
|
-
kind: itemType === "file_change" ? "patch" : "tool",
|
|
56
|
-
surface: codexItemSurface(itemType, item),
|
|
57
|
-
name: codexItemName(itemType, item),
|
|
58
|
-
status: "started",
|
|
59
|
-
timestampMs: timestampMs2,
|
|
60
|
-
metadata: compactMetadata({ sourceEventType: entryType, itemType })
|
|
61
|
-
});
|
|
62
|
-
calls.set(id, action);
|
|
63
|
-
actions.push(action);
|
|
64
|
-
}
|
|
65
|
-
}
|
|
66
|
-
if (entryType === "item.completed" && item) {
|
|
67
|
-
const itemType = stringField(item, "type");
|
|
68
|
-
if (itemType === "agent_message" || itemType === "message") {
|
|
69
|
-
finalText = nonEmpty(stringField(item, "text")) ?? finalText;
|
|
70
|
-
}
|
|
71
|
-
if (isCodexActionItem(itemType)) {
|
|
72
|
-
const id = stringField(item, "id") ?? `item-${actions.length}`;
|
|
73
|
-
const status = codexItemStatus(itemType, item);
|
|
74
|
-
const existing = calls.get(id);
|
|
75
|
-
if (existing) {
|
|
76
|
-
existing.status = status;
|
|
77
|
-
} else {
|
|
78
|
-
const action = actionFor({
|
|
79
|
-
id,
|
|
80
|
-
stepIndex: actions.length,
|
|
81
|
-
kind: itemType === "file_change" ? "patch" : "tool",
|
|
82
|
-
surface: codexItemSurface(itemType, item),
|
|
83
|
-
name: codexItemName(itemType, item),
|
|
84
|
-
status,
|
|
85
|
-
timestampMs: timestampMs2,
|
|
86
|
-
metadata: compactMetadata({ sourceEventType: entryType, itemType })
|
|
87
|
-
});
|
|
88
|
-
calls.set(id, action);
|
|
89
|
-
actions.push(action);
|
|
90
|
-
}
|
|
91
|
-
}
|
|
92
|
-
}
|
|
93
|
-
if (entryType === "response_item") {
|
|
94
|
-
if (payloadType === "message" || payloadType === "agent_message") {
|
|
95
|
-
finalText = textFromMessagePayload(payload) ?? finalText;
|
|
96
|
-
}
|
|
97
|
-
if (payloadType === "function_call" || payloadType === "custom_tool_call") {
|
|
98
|
-
const id = stringField(payload, "call_id") ?? stringField(payload, "id") ?? `call-${actions.length}`;
|
|
99
|
-
const name = stringField(payload, "name") ?? payloadType;
|
|
100
|
-
const action = actionFor({
|
|
101
|
-
id,
|
|
102
|
-
stepIndex: actions.length,
|
|
103
|
-
kind: "tool",
|
|
104
|
-
surface: surfaceForTool(name),
|
|
105
|
-
name,
|
|
106
|
-
status: "started",
|
|
107
|
-
timestampMs: timestampMs2,
|
|
108
|
-
metadata: compactMetadata({ sourceEventType: entryType, payloadType })
|
|
109
|
-
});
|
|
110
|
-
calls.set(id, action);
|
|
111
|
-
actions.push(action);
|
|
112
|
-
}
|
|
113
|
-
if (payloadType === "function_call_output" || payloadType === "custom_tool_call_output") {
|
|
114
|
-
const id = stringField(payload, "call_id") ?? stringField(payload, "id");
|
|
115
|
-
const existing = id ? calls.get(id) : void 0;
|
|
116
|
-
if (existing) existing.status = looksLikeError(payload.output) ? "failed" : "completed";
|
|
117
|
-
}
|
|
118
|
-
}
|
|
119
|
-
if (entryType === "event_msg") {
|
|
120
|
-
if (payloadType === "agent_message") {
|
|
121
|
-
finalText = nonEmpty(stringField(payload, "message")) ?? nonEmpty(stringField(payload, "text")) ?? finalText;
|
|
122
|
-
}
|
|
123
|
-
if (payloadType === "patch_apply_end") {
|
|
124
|
-
const patchStatus = typeof payload.success === "boolean" ? payload.success ? "completed" : "failed" : "unknown";
|
|
125
|
-
actions.push(
|
|
126
|
-
actionFor({
|
|
127
|
-
id: stringField(payload, "call_id") ?? `patch-${actions.length}`,
|
|
128
|
-
stepIndex: actions.length,
|
|
129
|
-
kind: "patch",
|
|
130
|
-
surface: "code",
|
|
131
|
-
name: "patch",
|
|
132
|
-
status: patchStatus,
|
|
133
|
-
timestampMs: timestampMs2,
|
|
134
|
-
metadata: compactMetadata({ sourceEventType: entryType, payloadType })
|
|
135
|
-
})
|
|
136
|
-
);
|
|
137
|
-
}
|
|
138
|
-
if (payloadType === "sub_agent_activity") {
|
|
139
|
-
const name = stringField(payload, "kind") ?? "subagent";
|
|
140
|
-
actions.push(
|
|
141
|
-
actionFor({
|
|
142
|
-
id: stringField(payload, "event_id") ?? `subagent-${actions.length}`,
|
|
143
|
-
stepIndex: actions.length,
|
|
144
|
-
kind: "tool",
|
|
145
|
-
surface: "subagent",
|
|
146
|
-
name,
|
|
147
|
-
status: statusFrom(payload),
|
|
148
|
-
timestampMs: timestampMs2,
|
|
149
|
-
metadata: compactMetadata({ sourceEventType: entryType, payloadType })
|
|
150
|
-
})
|
|
151
|
-
);
|
|
152
|
-
}
|
|
153
|
-
if (payloadType === "task_complete" || payloadType === "turn_aborted") {
|
|
154
|
-
const completed = payloadType === "task_complete";
|
|
155
|
-
terminal = completed ? "completed" : "failed";
|
|
156
|
-
explicitTerminal = true;
|
|
157
|
-
actions.push(terminalAction(actions.length, payloadType, completed, timestampMs2));
|
|
158
|
-
}
|
|
159
|
-
}
|
|
160
|
-
if (entryType === "turn.completed" || entryType === "exec_done") {
|
|
161
|
-
const completed = entry.success !== false;
|
|
162
|
-
terminal = completed ? "completed" : "failed";
|
|
163
|
-
explicitTerminal = true;
|
|
164
|
-
finalText = nonEmpty(stringField(entry, "message")) ?? finalText;
|
|
165
|
-
actions.push(terminalAction(actions.length, entryType, completed, timestampMs2));
|
|
166
|
-
}
|
|
167
|
-
if (entryType === "turn.failed" || entryType === "error") {
|
|
168
|
-
terminal = "failed";
|
|
169
|
-
explicitTerminal = true;
|
|
170
|
-
actions.push(terminalAction(actions.length, entryType, false, timestampMs2));
|
|
171
|
-
}
|
|
172
|
-
}
|
|
173
|
-
return { finalText, terminal, explicitTerminal, actions };
|
|
174
|
-
}
|
|
175
|
-
function claudeProjection(entries) {
|
|
176
|
-
const actions = [];
|
|
177
|
-
const calls = /* @__PURE__ */ new Map();
|
|
178
|
-
let finalText = null;
|
|
179
|
-
let resultText = null;
|
|
180
|
-
let terminal = "unknown";
|
|
181
|
-
let explicitTerminal = false;
|
|
182
|
-
for (const entry of entries) {
|
|
183
|
-
const entryType = stringField(entry, "type");
|
|
184
|
-
const timestampMs2 = timestamp(entry.timestamp);
|
|
185
|
-
const message = record(entry.message);
|
|
186
|
-
const content = Array.isArray(message?.content) ? message.content : [];
|
|
187
|
-
if (entryType === "assistant") {
|
|
188
|
-
const text = textBlocks(content);
|
|
189
|
-
if (text !== null) finalText = text;
|
|
190
|
-
for (const value of content) {
|
|
191
|
-
const part = record(value);
|
|
192
|
-
if (!part) continue;
|
|
193
|
-
const partType = stringField(part, "type");
|
|
194
|
-
if (partType === "tool_use") {
|
|
195
|
-
const id = stringField(part, "id") ?? `tool-${actions.length}`;
|
|
196
|
-
const name = stringField(part, "name") ?? "tool";
|
|
197
|
-
const action = actionFor({
|
|
198
|
-
id,
|
|
199
|
-
stepIndex: actions.length,
|
|
200
|
-
kind: "tool",
|
|
201
|
-
surface: surfaceForTool(name),
|
|
202
|
-
name,
|
|
203
|
-
status: "started",
|
|
204
|
-
timestampMs: timestampMs2,
|
|
205
|
-
metadata: compactMetadata({ sourceEventType: entryType, partType })
|
|
206
|
-
});
|
|
207
|
-
calls.set(id, action);
|
|
208
|
-
actions.push(action);
|
|
209
|
-
}
|
|
210
|
-
}
|
|
211
|
-
completeClaudeTools(content, calls);
|
|
212
|
-
const stopReason = stringField(message ?? {}, "stop_reason");
|
|
213
|
-
if (stopReason === "end_turn") {
|
|
214
|
-
terminal = "completed";
|
|
215
|
-
explicitTerminal = true;
|
|
216
|
-
actions.push(terminalAction(actions.length, stopReason, true, timestampMs2));
|
|
217
|
-
}
|
|
218
|
-
}
|
|
219
|
-
if (entryType === "user") {
|
|
220
|
-
completeClaudeTools(content, calls);
|
|
221
|
-
}
|
|
222
|
-
if (entryType === "tool_use") {
|
|
223
|
-
const id = stringField(entry, "id") ?? `tool-${actions.length}`;
|
|
224
|
-
const name = stringField(entry, "name") ?? "tool";
|
|
225
|
-
const action = actionFor({
|
|
226
|
-
id,
|
|
227
|
-
stepIndex: actions.length,
|
|
228
|
-
kind: "tool",
|
|
229
|
-
surface: surfaceForTool(name),
|
|
230
|
-
name,
|
|
231
|
-
status: "started",
|
|
232
|
-
timestampMs: timestampMs2,
|
|
233
|
-
metadata: { sourceEventType: entryType }
|
|
234
|
-
});
|
|
235
|
-
calls.set(id, action);
|
|
236
|
-
actions.push(action);
|
|
237
|
-
}
|
|
238
|
-
if (entryType === "tool_result") {
|
|
239
|
-
const id = stringField(entry, "tool_use_id");
|
|
240
|
-
const action = id ? calls.get(id) : void 0;
|
|
241
|
-
if (action) action.status = entry.is_error === true ? "failed" : "completed";
|
|
242
|
-
}
|
|
243
|
-
if (entryType === "system") {
|
|
244
|
-
const subtype = stringField(entry, "subtype");
|
|
245
|
-
const hookName = stringField(entry, "hook_name") ?? stringField(entry, "hookName");
|
|
246
|
-
if (subtype?.toLowerCase().includes("hook") || hookName) {
|
|
247
|
-
actions.push(
|
|
248
|
-
actionFor({
|
|
249
|
-
id: stringField(entry, "uuid") ?? `hook-${actions.length}`,
|
|
250
|
-
stepIndex: actions.length,
|
|
251
|
-
kind: "tool",
|
|
252
|
-
surface: "hook",
|
|
253
|
-
name: hookName ?? subtype ?? "hook",
|
|
254
|
-
status: statusFrom(entry),
|
|
255
|
-
timestampMs: timestampMs2,
|
|
256
|
-
metadata: compactMetadata({ sourceEventType: entryType, subtype })
|
|
257
|
-
})
|
|
258
|
-
);
|
|
259
|
-
}
|
|
260
|
-
}
|
|
261
|
-
if (entryType === "pr-link") {
|
|
262
|
-
actions.push(terminalAction(actions.length, entryType, true, timestampMs2));
|
|
263
|
-
}
|
|
264
|
-
if (entryType === "result") {
|
|
265
|
-
resultText = nonEmpty(stringField(entry, "result")) ?? resultText;
|
|
266
|
-
const subtype = stringField(entry, "subtype");
|
|
267
|
-
const completed = entry.is_error !== true && !subtype?.toLowerCase().includes("error");
|
|
268
|
-
terminal = completed ? "completed" : "failed";
|
|
269
|
-
explicitTerminal = true;
|
|
270
|
-
actions.push(terminalAction(actions.length, subtype ?? entryType, completed, timestampMs2));
|
|
271
|
-
}
|
|
272
|
-
if (entryType === "error") {
|
|
273
|
-
terminal = "failed";
|
|
274
|
-
explicitTerminal = true;
|
|
275
|
-
actions.push(terminalAction(actions.length, entryType, false, timestampMs2));
|
|
276
|
-
}
|
|
277
|
-
}
|
|
278
|
-
return {
|
|
279
|
-
finalText: resultText ?? finalText,
|
|
280
|
-
terminal,
|
|
281
|
-
explicitTerminal,
|
|
282
|
-
actions
|
|
283
|
-
};
|
|
284
|
-
}
|
|
285
|
-
function openCodeProjection(entries) {
|
|
286
|
-
const actions = [];
|
|
287
|
-
const assistantMessageIds = new Set(
|
|
288
|
-
entries.filter((entry) => stringField(entry, "role") === "assistant").map((entry) => stringField(entry, "id")).filter((id) => id !== void 0)
|
|
289
|
-
);
|
|
290
|
-
let finalText = null;
|
|
291
|
-
let terminal = "unknown";
|
|
292
|
-
let explicitTerminal = false;
|
|
293
|
-
for (const entry of entries) {
|
|
294
|
-
const entryType = stringField(entry, "type");
|
|
295
|
-
const part = record(entry.part) ?? entry;
|
|
296
|
-
const partType = stringField(part, "type");
|
|
297
|
-
const timestampMs2 = timestamp(entry.timestamp) ?? timestamp(record(entry.time)?.created);
|
|
298
|
-
if (entryType === "text" && record(entry.part)) {
|
|
299
|
-
finalText = nonEmpty(stringField(part, "text")) ?? finalText;
|
|
300
|
-
} else if (partType === "text" && (assistantMessageIds.size === 0 || assistantMessageIds.has(stringField(part, "messageID") ?? ""))) {
|
|
301
|
-
finalText = nonEmpty(stringField(part, "text")) ?? finalText;
|
|
302
|
-
}
|
|
303
|
-
if (entryType === "tool_use" || partType === "tool") {
|
|
304
|
-
const state = record(part.state) ?? {};
|
|
305
|
-
const name = stringField(part, "tool") ?? "tool";
|
|
306
|
-
actions.push(
|
|
307
|
-
actionFor({
|
|
308
|
-
id: stringField(part, "callID") ?? stringField(part, "id") ?? `tool-${actions.length}`,
|
|
309
|
-
stepIndex: actions.length,
|
|
310
|
-
kind: "tool",
|
|
311
|
-
surface: surfaceForTool(name),
|
|
312
|
-
name,
|
|
313
|
-
status: statusFrom(state),
|
|
314
|
-
timestampMs: timestampMs2,
|
|
315
|
-
metadata: compactMetadata({ sourceEventType: entryType, partType })
|
|
316
|
-
})
|
|
317
|
-
);
|
|
318
|
-
}
|
|
319
|
-
if (partType === "patch") {
|
|
320
|
-
actions.push(
|
|
321
|
-
actionFor({
|
|
322
|
-
id: stringField(part, "id") ?? `patch-${actions.length}`,
|
|
323
|
-
stepIndex: actions.length,
|
|
324
|
-
kind: "patch",
|
|
325
|
-
surface: "code",
|
|
326
|
-
name: "patch",
|
|
327
|
-
status: "completed",
|
|
328
|
-
timestampMs: timestampMs2,
|
|
329
|
-
metadata: compactMetadata({ sourceEventType: entryType, partType })
|
|
330
|
-
})
|
|
331
|
-
);
|
|
332
|
-
}
|
|
333
|
-
if (stringField(entry, "role") === "assistant") {
|
|
334
|
-
const finish = stringField(entry, "finish");
|
|
335
|
-
if (finish === "stop" || finish === "error") {
|
|
336
|
-
const completed = finish === "stop";
|
|
337
|
-
terminal = completed ? "completed" : "failed";
|
|
338
|
-
explicitTerminal = true;
|
|
339
|
-
actions.push(terminalAction(actions.length, finish, completed, timestampMs2));
|
|
340
|
-
}
|
|
341
|
-
}
|
|
342
|
-
if (entryType === "error") {
|
|
343
|
-
terminal = "failed";
|
|
344
|
-
explicitTerminal = true;
|
|
345
|
-
actions.push(terminalAction(actions.length, entryType, false, timestampMs2));
|
|
346
|
-
}
|
|
347
|
-
}
|
|
348
|
-
return { finalText, terminal, explicitTerminal, actions };
|
|
349
|
-
}
|
|
350
|
-
function kimiProjection(entries) {
|
|
351
|
-
const actions = [];
|
|
352
|
-
const calls = /* @__PURE__ */ new Map();
|
|
353
|
-
let finalText = null;
|
|
354
|
-
let terminal = "unknown";
|
|
355
|
-
let explicitTerminal = false;
|
|
356
|
-
for (const entry of entries) {
|
|
357
|
-
const timestampMs2 = timestamp(entry.timestamp);
|
|
358
|
-
const message = record(entry.message);
|
|
359
|
-
const messageType = stringField(message ?? {}, "type");
|
|
360
|
-
const payload = record(message?.payload) ?? {};
|
|
361
|
-
if (messageType === "ContentPart" && stringField(payload, "type") === "text") {
|
|
362
|
-
finalText = nonEmpty(stringField(payload, "text")) ?? nonEmpty(stringField(payload, "content")) ?? finalText;
|
|
363
|
-
}
|
|
364
|
-
if (messageType === "ToolCall") {
|
|
365
|
-
const fn = record(payload.function) ?? {};
|
|
366
|
-
const id = stringField(payload, "id") ?? `tool-${actions.length}`;
|
|
367
|
-
const name = stringField(fn, "name") ?? "tool";
|
|
368
|
-
const action = actionFor({
|
|
369
|
-
id,
|
|
370
|
-
stepIndex: actions.length,
|
|
371
|
-
kind: "tool",
|
|
372
|
-
surface: surfaceForTool(name),
|
|
373
|
-
name,
|
|
374
|
-
status: "started",
|
|
375
|
-
timestampMs: timestampMs2,
|
|
376
|
-
metadata: { sourceEventType: messageType }
|
|
377
|
-
});
|
|
378
|
-
calls.set(id, action);
|
|
379
|
-
actions.push(action);
|
|
380
|
-
}
|
|
381
|
-
if (messageType === "ToolResult") {
|
|
382
|
-
const id = stringField(payload, "tool_call_id");
|
|
383
|
-
const action = id ? calls.get(id) : void 0;
|
|
384
|
-
if (action)
|
|
385
|
-
action.status = record(payload.return_value)?.is_error === true ? "failed" : "completed";
|
|
386
|
-
}
|
|
387
|
-
if (messageType === "TurnEnd" || messageType === "StepInterrupted") {
|
|
388
|
-
const completed = messageType === "TurnEnd";
|
|
389
|
-
terminal = completed ? "completed" : "failed";
|
|
390
|
-
explicitTerminal = true;
|
|
391
|
-
actions.push(terminalAction(actions.length, messageType, completed, timestampMs2));
|
|
392
|
-
}
|
|
393
|
-
}
|
|
394
|
-
return { finalText, terminal, explicitTerminal, actions };
|
|
395
|
-
}
|
|
396
|
-
function piProjection(entries) {
|
|
397
|
-
const actions = [];
|
|
398
|
-
let terminal = "unknown";
|
|
399
|
-
let explicitTerminal = false;
|
|
400
|
-
for (const entry of entries) {
|
|
401
|
-
const nodes = Array.isArray(entry.nodes) ? entry.nodes : [];
|
|
402
|
-
for (const value of nodes) {
|
|
403
|
-
const node = record(value);
|
|
404
|
-
const ir = record(node?.ir) ?? node;
|
|
405
|
-
const kind = stringField(ir ?? {}, "kind");
|
|
406
|
-
if (kind === "ToolInvocation") {
|
|
407
|
-
actions.push(
|
|
408
|
-
actionFor({
|
|
409
|
-
id: stringField(ir ?? {}, "id") ?? stringField(node ?? {}, "id") ?? `tool-${actions.length}`,
|
|
410
|
-
stepIndex: actions.length,
|
|
411
|
-
kind: "tool",
|
|
412
|
-
surface: "tool",
|
|
413
|
-
name: stringField(ir ?? {}, "name") ?? "graph-tool",
|
|
414
|
-
status: "started",
|
|
415
|
-
timestampMs: timestamp(ir?.createdAt),
|
|
416
|
-
metadata: { sourceEventType: "graph-node", graphKind: kind }
|
|
417
|
-
})
|
|
418
|
-
);
|
|
419
|
-
}
|
|
420
|
-
if (kind === "ToolResult") {
|
|
421
|
-
const prior = [...actions].reverse().find((action) => action.kind === "tool");
|
|
422
|
-
if (prior) prior.status = "completed";
|
|
423
|
-
}
|
|
424
|
-
if (kind === "CompletionDecision") {
|
|
425
|
-
terminal = "completed";
|
|
426
|
-
explicitTerminal = true;
|
|
427
|
-
actions.push(
|
|
428
|
-
actionFor({
|
|
429
|
-
id: stringField(ir ?? {}, "id") ?? stringField(node ?? {}, "id") ?? `completion-${actions.length}`,
|
|
430
|
-
stepIndex: actions.length,
|
|
431
|
-
kind: "graph-completion",
|
|
432
|
-
surface: "tool",
|
|
433
|
-
name: "complete",
|
|
434
|
-
status: "completed",
|
|
435
|
-
timestampMs: timestamp(ir?.createdAt),
|
|
436
|
-
metadata: { sourceEventType: "graph-node", graphKind: kind }
|
|
437
|
-
})
|
|
438
|
-
);
|
|
439
|
-
}
|
|
440
|
-
}
|
|
441
|
-
}
|
|
442
|
-
return { finalText: null, terminal, explicitTerminal, actions };
|
|
443
|
-
}
|
|
444
|
-
function completeClaudeTools(content, calls) {
|
|
445
|
-
for (const value of content) {
|
|
446
|
-
const part = record(value);
|
|
447
|
-
if (!part || stringField(part, "type") !== "tool_result") continue;
|
|
448
|
-
const id = stringField(part, "tool_use_id");
|
|
449
|
-
const action = id ? calls.get(id) : void 0;
|
|
450
|
-
if (action) action.status = part.is_error === true ? "failed" : "completed";
|
|
451
|
-
}
|
|
452
|
-
}
|
|
453
|
-
function isCodexActionItem(itemType) {
|
|
454
|
-
return itemType === "command_execution" || itemType === "mcp_tool_call" || itemType === "collab_tool_call" || itemType === "web_search" || itemType === "file_change";
|
|
455
|
-
}
|
|
456
|
-
function codexItemSurface(itemType, item) {
|
|
457
|
-
if (itemType === "mcp_tool_call") return "mcp";
|
|
458
|
-
if (itemType === "collab_tool_call") return "subagent";
|
|
459
|
-
if (itemType === "web_search") return "web";
|
|
460
|
-
if (itemType === "file_change") return "code";
|
|
461
|
-
return surfaceForTool(stringField(item, "name") ?? itemType ?? "tool");
|
|
462
|
-
}
|
|
463
|
-
function codexItemName(itemType, item) {
|
|
464
|
-
if (itemType === "mcp_tool_call") {
|
|
465
|
-
const server = stringField(item, "server");
|
|
466
|
-
const tool = stringField(item, "tool");
|
|
467
|
-
return [server, tool].filter(Boolean).join("/") || "mcp";
|
|
468
|
-
}
|
|
469
|
-
if (itemType === "collab_tool_call") return stringField(item, "tool") ?? "subagent";
|
|
470
|
-
if (itemType === "web_search") return "web_search";
|
|
471
|
-
if (itemType === "file_change") return "file_change";
|
|
472
|
-
return stringField(item, "name") ?? itemType ?? "tool";
|
|
473
|
-
}
|
|
474
|
-
function codexItemStatus(itemType, item) {
|
|
475
|
-
const status = statusFrom(item);
|
|
476
|
-
if (status === "failed") return status;
|
|
477
|
-
if (itemType === "command_execution") {
|
|
478
|
-
const exitCode = numberField(item, "exit_code");
|
|
479
|
-
if (exitCode !== void 0) return exitCode === 0 ? "completed" : "failed";
|
|
480
|
-
}
|
|
481
|
-
if (itemType === "mcp_tool_call" && record(item.error)) return "failed";
|
|
482
|
-
return status === "unknown" || status === "started" ? "completed" : status;
|
|
483
|
-
}
|
|
484
|
-
function textFromMessagePayload(payload) {
|
|
485
|
-
const direct = nonEmpty(stringField(payload, "text"));
|
|
486
|
-
if (direct) return direct;
|
|
487
|
-
const content = payload.content;
|
|
488
|
-
if (typeof content === "string") return nonEmpty(content);
|
|
489
|
-
if (Array.isArray(content)) return textBlocks(content);
|
|
490
|
-
return null;
|
|
491
|
-
}
|
|
492
|
-
function textBlocks(values) {
|
|
493
|
-
const text = values.map((value) => record(value)).filter((value) => value !== null).filter((value) => stringField(value, "type") === "text").map((value) => stringField(value, "text")).filter((value) => value !== void 0).join("\n");
|
|
494
|
-
return nonEmpty(text);
|
|
495
|
-
}
|
|
496
|
-
function terminalWithExecution(projection, execution) {
|
|
497
|
-
if (!execution) {
|
|
498
|
-
return { status: projection.terminal, explicit: projection.explicitTerminal };
|
|
499
|
-
}
|
|
500
|
-
assertExecutionReceipt(execution);
|
|
501
|
-
return {
|
|
502
|
-
status: execution.exitCode !== 0 || projection.terminal === "failed" ? "failed" : "completed",
|
|
503
|
-
explicit: true
|
|
504
|
-
};
|
|
505
|
-
}
|
|
506
|
-
function assertExecutionReceipt(receipt) {
|
|
507
|
-
if (!Number.isSafeInteger(receipt.exitCode)) {
|
|
508
|
-
throw new Error("code-agent execution exitCode must be a safe integer");
|
|
509
|
-
}
|
|
510
|
-
for (const [name, value] of [
|
|
511
|
-
["startedAtMs", receipt.startedAtMs],
|
|
512
|
-
["completedAtMs", receipt.completedAtMs]
|
|
513
|
-
]) {
|
|
514
|
-
if (value !== void 0 && (!Number.isFinite(value) || value < 0)) {
|
|
515
|
-
throw new Error(`code-agent execution ${name} must be a non-negative finite number`);
|
|
516
|
-
}
|
|
517
|
-
}
|
|
518
|
-
if (receipt.startedAtMs !== void 0 && receipt.completedAtMs !== void 0 && receipt.completedAtMs < receipt.startedAtMs) {
|
|
519
|
-
throw new Error("code-agent execution completedAtMs precedes startedAtMs");
|
|
520
|
-
}
|
|
521
|
-
}
|
|
522
|
-
function terminalAction(stepIndex, name, completed, timestampMs2) {
|
|
523
|
-
return actionFor({
|
|
524
|
-
id: `terminal-${stepIndex}`,
|
|
525
|
-
stepIndex,
|
|
526
|
-
kind: "terminal",
|
|
527
|
-
surface: "tool",
|
|
528
|
-
name,
|
|
529
|
-
status: completed ? "completed" : "failed",
|
|
530
|
-
timestampMs: timestampMs2,
|
|
531
|
-
metadata: { sourceEventType: name }
|
|
532
|
-
});
|
|
533
|
-
}
|
|
534
|
-
function actionFor(input) {
|
|
535
|
-
return input;
|
|
536
|
-
}
|
|
537
|
-
function surfaceForTool(name) {
|
|
538
|
-
const normalized = name.toLowerCase();
|
|
539
|
-
if (normalized.startsWith("mcp__") || normalized.includes("mcp_tool")) return "mcp";
|
|
540
|
-
if (normalized === "task" || normalized === "agent" || normalized.includes("subagent") || normalized.includes("spawn_agent") || normalized.includes("collab") || normalized.startsWith("multi_agent")) {
|
|
541
|
-
return "subagent";
|
|
542
|
-
}
|
|
543
|
-
if (normalized === "skill" || normalized.includes("skill")) return "skill";
|
|
544
|
-
if (normalized.includes("hook")) return "hook";
|
|
545
|
-
if (normalized.includes("web_search") || normalized.includes("webfetch") || normalized === "web") {
|
|
546
|
-
return "web";
|
|
547
|
-
}
|
|
548
|
-
if (normalized === "edit" || normalized === "write" || normalized.includes("patch") || normalized.includes("file_change")) {
|
|
549
|
-
return "code";
|
|
550
|
-
}
|
|
551
|
-
return "tool";
|
|
552
|
-
}
|
|
553
|
-
function statusFrom(value) {
|
|
554
|
-
const status = (stringField(value, "status") ?? "").toLowerCase();
|
|
555
|
-
if (status === "completed" || status === "success" || status === "succeeded") return "completed";
|
|
556
|
-
if (status === "error" || status === "failed" || status === "declined") return "failed";
|
|
557
|
-
if (status === "running" || status === "started" || status === "in_progress") return "started";
|
|
558
|
-
return "unknown";
|
|
559
|
-
}
|
|
560
|
-
function compactMetadata(input) {
|
|
561
|
-
return Object.fromEntries(Object.entries(input).filter((entry) => entry[1] !== void 0));
|
|
562
|
-
}
|
|
563
|
-
function sessionIdFromEntries(source, entries) {
|
|
564
|
-
for (const entry of entries) {
|
|
565
|
-
if (source === "codex") {
|
|
566
|
-
const threadId = stringField(entry, "thread_id");
|
|
567
|
-
if (threadId) return threadId;
|
|
568
|
-
const payload2 = record(entry.payload);
|
|
569
|
-
const id = payload2 ? stringField(payload2, "id") : void 0;
|
|
570
|
-
if (id) return id;
|
|
571
|
-
}
|
|
572
|
-
const message = record(entry.message);
|
|
573
|
-
const payload = record(message?.payload);
|
|
574
|
-
const sessionId = stringField(entry, "sessionID") ?? stringField(entry, "sessionId") ?? stringField(entry, "session_id") ?? stringField(payload ?? {}, "session_id");
|
|
575
|
-
if (sessionId) return sessionId;
|
|
576
|
-
}
|
|
577
|
-
return void 0;
|
|
578
|
-
}
|
|
579
|
-
function fallbackSessionId(source, sourcePath) {
|
|
580
|
-
return `${source}:${sourcePath ?? "unknown-session"}`;
|
|
581
|
-
}
|
|
582
|
-
function looksLikeError(value) {
|
|
583
|
-
if (isRecord(value)) {
|
|
584
|
-
if (value.success === false || value.is_error === true) return true;
|
|
585
|
-
const exitCode = numberField(value, "exit_code");
|
|
586
|
-
if (exitCode !== void 0 && exitCode !== 0) return true;
|
|
587
|
-
}
|
|
588
|
-
const text = typeof value === "string" ? value : JSON.stringify(value ?? "");
|
|
589
|
-
return /\b(?:error|failed|failure)\b|exit(?:_| )code["\s:]+[1-9]/i.test(text);
|
|
590
|
-
}
|
|
591
|
-
function nonEmpty(value) {
|
|
592
|
-
return value !== void 0 && value.length > 0 ? value : null;
|
|
593
|
-
}
|
|
594
|
-
function isRecord(value) {
|
|
595
|
-
return value !== null && typeof value === "object" && !Array.isArray(value);
|
|
596
|
-
}
|
|
597
|
-
function record(value) {
|
|
598
|
-
return isRecord(value) ? value : null;
|
|
599
|
-
}
|
|
600
|
-
function stringField(value, key) {
|
|
601
|
-
const field = value[key];
|
|
602
|
-
return typeof field === "string" && field.length > 0 ? field : void 0;
|
|
603
|
-
}
|
|
604
|
-
function numberField(value, key) {
|
|
605
|
-
const field = value[key];
|
|
606
|
-
return typeof field === "number" && Number.isFinite(field) ? field : void 0;
|
|
607
|
-
}
|
|
608
|
-
function timestamp(value) {
|
|
609
|
-
if (typeof value === "number" && Number.isFinite(value)) {
|
|
610
|
-
return value > 1e12 ? value : value * 1e3;
|
|
611
|
-
}
|
|
612
|
-
if (typeof value === "string" && value.length > 0) {
|
|
613
|
-
const parsed = Date.parse(value);
|
|
614
|
-
return Number.isFinite(parsed) ? parsed : void 0;
|
|
615
|
-
}
|
|
616
|
-
return void 0;
|
|
617
|
-
}
|
|
618
|
-
|
|
619
|
-
// src/contract/intake/code-agent-session.ts
|
|
620
|
-
import { createHash } from "crypto";
|
|
621
|
-
function parseCodeAgentJsonl(jsonl) {
|
|
622
|
-
const entries = [];
|
|
623
|
-
let malformedLines = 0;
|
|
624
|
-
for (const line of jsonl.split("\n")) {
|
|
625
|
-
const trimmed = line.trim();
|
|
626
|
-
if (!trimmed) continue;
|
|
627
|
-
try {
|
|
628
|
-
entries.push(JSON.parse(trimmed));
|
|
629
|
-
} catch {
|
|
630
|
-
malformedLines += 1;
|
|
631
|
-
}
|
|
632
|
-
}
|
|
633
|
-
return { entries, malformedLines };
|
|
634
|
-
}
|
|
635
|
-
function fromCodexSession(options) {
|
|
636
|
-
return fromCodeAgentSession("codex", options);
|
|
637
|
-
}
|
|
638
|
-
function fromClaudeCodeSession(options) {
|
|
639
|
-
return fromCodeAgentSession("claude-code", options);
|
|
640
|
-
}
|
|
641
|
-
function fromOpenCodeSession(options) {
|
|
642
|
-
return fromCodeAgentSession("opencode", options);
|
|
643
|
-
}
|
|
644
|
-
function fromKimiCodeSession(options) {
|
|
645
|
-
return fromCodeAgentSession("kimi-code", options);
|
|
646
|
-
}
|
|
647
|
-
function fromPiSession(options) {
|
|
648
|
-
return fromCodeAgentSession("pi", options);
|
|
649
|
-
}
|
|
650
|
-
var fromPigraphSession = fromPiSession;
|
|
651
|
-
function fromCodeAgentSession(source, options) {
|
|
652
|
-
const entries = options.entries.filter(isRecord2);
|
|
653
|
-
if (entries.length === 0) {
|
|
654
|
-
return {
|
|
655
|
-
runs: [],
|
|
656
|
-
diagnostics: [
|
|
657
|
-
{
|
|
658
|
-
source,
|
|
659
|
-
sessionId: fallbackSessionId2(source, options.sourcePath),
|
|
660
|
-
sourcePath: options.sourcePath,
|
|
661
|
-
entries: 0,
|
|
662
|
-
malformedLines: options.malformedLines ?? 0,
|
|
663
|
-
hasExplicitTerminalSignal: false,
|
|
664
|
-
hasFinalOutput: false,
|
|
665
|
-
hasQualityLabel: false,
|
|
666
|
-
hasTokenUsage: false,
|
|
667
|
-
hasCost: false,
|
|
668
|
-
costKind: "uncaptured",
|
|
669
|
-
warnings: ["no parseable session entries"]
|
|
670
|
-
}
|
|
671
|
-
],
|
|
672
|
-
metrics: [],
|
|
673
|
-
observations: []
|
|
674
|
-
};
|
|
675
|
-
}
|
|
676
|
-
const metrics = metricsFor(source, entries);
|
|
677
|
-
const sessionId = sessionIdFromEntries2(source, entries) ?? fallbackSessionId2(source, options.sourcePath);
|
|
678
|
-
const observation = observeCodeAgentSession({
|
|
679
|
-
source,
|
|
680
|
-
entries,
|
|
681
|
-
sourcePath: options.sourcePath,
|
|
682
|
-
execution: options.execution
|
|
683
|
-
});
|
|
684
|
-
applyObservedSurfaceMetrics(metrics, observation);
|
|
685
|
-
const model = withSnapshot(options.model ?? modelFromEntries(source, entries) ?? source);
|
|
686
|
-
const tokenUsage = {
|
|
687
|
-
input: metrics.inputTokens,
|
|
688
|
-
output: metrics.outputTokens,
|
|
689
|
-
...metrics.reasoningTokens > 0 ? { reasoning: metrics.reasoningTokens } : {},
|
|
690
|
-
...metrics.cachedTokens > 0 ? { cached: metrics.cachedTokens } : {},
|
|
691
|
-
...metrics.cacheWriteTokens > 0 ? { cacheWrite: metrics.cacheWriteTokens } : {}
|
|
692
|
-
};
|
|
693
|
-
const { costUsd, costProvenance } = resolveSessionCost(options.costProvenance, metrics, model);
|
|
694
|
-
const qualityScore = finiteScore(options.score);
|
|
695
|
-
const invalidQualityScore = options.score !== void 0 && qualityScore === void 0;
|
|
696
|
-
const promptHash = options.promptHash ?? hashString(`prompt:${source}:${firstUserText(entries) ?? ""}`);
|
|
697
|
-
const configHash = options.configHash ?? hashJson({
|
|
698
|
-
source,
|
|
699
|
-
model,
|
|
700
|
-
sourcePath: options.sourcePath,
|
|
701
|
-
cwd: cwdFromEntries(entries),
|
|
702
|
-
entryCount: entries.length
|
|
703
|
-
});
|
|
704
|
-
const explicitTerminal = observation.terminal.explicit;
|
|
705
|
-
const warnings = diagnosticsFor(metrics, {
|
|
706
|
-
model,
|
|
707
|
-
explicitTerminal,
|
|
708
|
-
malformedLines: options.malformedLines ?? 0,
|
|
709
|
-
hasQualityLabel: qualityScore !== void 0,
|
|
710
|
-
invalidQualityScore,
|
|
711
|
-
costKind: costProvenance.kind
|
|
712
|
-
});
|
|
713
|
-
const terminalOutcome = terminalOutcomeFromSession(observation.terminal);
|
|
714
|
-
const processExitCode = options.execution?.exitCode;
|
|
715
|
-
const splitTag = options.splitTag ?? "holdout";
|
|
716
|
-
const run = {
|
|
717
|
-
runId: `${source}:${sessionId}`,
|
|
718
|
-
experimentId: options.experimentId ?? `${source}-local-sessions`,
|
|
719
|
-
candidateId: options.candidateId ?? model,
|
|
720
|
-
seed: options.seed ?? stableSeed(sessionId),
|
|
721
|
-
model,
|
|
722
|
-
promptHash,
|
|
723
|
-
configHash,
|
|
724
|
-
commitSha: options.commitSha ?? "local-session",
|
|
725
|
-
wallMs: metrics.wallMs,
|
|
726
|
-
costUsd,
|
|
727
|
-
costProvenance,
|
|
728
|
-
tokenUsage,
|
|
729
|
-
terminalOutcome,
|
|
730
|
-
...terminalOutcome === "failed" ? {
|
|
731
|
-
terminalFailureReason: processExitCode !== void 0 && processExitCode !== 0 ? `process exited with code ${processExitCode}` : metrics.turnsAborted > 0 ? "turn aborted" : "session reported terminal failure"
|
|
732
|
-
} : {},
|
|
733
|
-
outcome: {
|
|
734
|
-
...qualityScore === void 0 ? {} : splitTag === "holdout" ? { holdoutScore: qualityScore } : { searchScore: qualityScore },
|
|
735
|
-
raw: {
|
|
736
|
-
entries: metrics.entries,
|
|
737
|
-
user_messages: metrics.userMessages,
|
|
738
|
-
assistant_messages: metrics.assistantMessages,
|
|
739
|
-
reasoning_items: metrics.reasoningItems,
|
|
740
|
-
tool_calls: metrics.toolCalls,
|
|
741
|
-
tool_outputs: metrics.toolOutputs,
|
|
742
|
-
tool_errors: metrics.toolErrors,
|
|
743
|
-
execution_error_count: metrics.toolErrors,
|
|
744
|
-
process_error_count: terminalOutcome === "failed" ? 1 : 0,
|
|
745
|
-
unclassified_error_count: metrics.unclassifiedErrors,
|
|
746
|
-
...processExitCode !== void 0 ? { process_exit_code: processExitCode } : {},
|
|
747
|
-
patch_attempts: metrics.patchAttempts,
|
|
748
|
-
patch_successes: metrics.patchSuccesses,
|
|
749
|
-
patch_failures: metrics.patchFailures,
|
|
750
|
-
turns_started: metrics.turnsStarted,
|
|
751
|
-
turns_completed: metrics.turnsCompleted,
|
|
752
|
-
turns_aborted: metrics.turnsAborted,
|
|
753
|
-
context_compactions: metrics.contextCompactions,
|
|
754
|
-
mcp_calls: metrics.mcpCalls,
|
|
755
|
-
subagent_calls: metrics.subagentCalls,
|
|
756
|
-
skill_calls: metrics.skillCalls,
|
|
757
|
-
hook_calls: metrics.hookCalls,
|
|
758
|
-
web_calls: metrics.webCalls,
|
|
759
|
-
code_actions: metrics.codeActions,
|
|
760
|
-
pr_links: metrics.prLinks,
|
|
761
|
-
file_snapshots: metrics.fileSnapshots,
|
|
762
|
-
graph_nodes: metrics.graphNodes,
|
|
763
|
-
graph_edges: metrics.graphEdges,
|
|
764
|
-
action_candidates: metrics.actionCandidates,
|
|
765
|
-
verification_reports: metrics.verificationReports,
|
|
766
|
-
completion_decisions: metrics.completionDecisions,
|
|
767
|
-
reliability_rows: metrics.reliabilityRows,
|
|
768
|
-
reliability_lift: metrics.reliabilityLift,
|
|
769
|
-
input_tokens: metrics.inputTokens,
|
|
770
|
-
output_tokens: metrics.outputTokens,
|
|
771
|
-
reasoning_tokens: metrics.reasoningTokens,
|
|
772
|
-
cached_tokens: metrics.cachedTokens,
|
|
773
|
-
cache_write_tokens: metrics.cacheWriteTokens,
|
|
774
|
-
observed_cost_usd: metrics.observedCostUsd,
|
|
775
|
-
observed_cost_captured: metrics.observedCostCaptured ? 1 : 0,
|
|
776
|
-
estimated_cost_usd: costProvenance.kind === "estimated" ? costProvenance.usd : 0,
|
|
777
|
-
process_score: metrics.processScore,
|
|
778
|
-
explicit_terminal_signal: explicitTerminal ? 1 : 0,
|
|
779
|
-
quality_label_present: qualityScore !== void 0 ? 1 : 0,
|
|
780
|
-
cost_observed: costProvenance.kind === "observed" ? 1 : 0,
|
|
781
|
-
cost_estimated: costProvenance.kind === "estimated" ? 1 : 0,
|
|
782
|
-
cost_uncaptured: costProvenance.kind === "uncaptured" ? 1 : 0,
|
|
783
|
-
cost_unknown: costProvenance.kind === "uncaptured" ? 1 : 0
|
|
784
|
-
}
|
|
785
|
-
},
|
|
786
|
-
splitTag,
|
|
787
|
-
scenarioId: options.scenarioId ?? sessionId
|
|
788
|
-
};
|
|
789
|
-
return {
|
|
790
|
-
runs: [run],
|
|
791
|
-
diagnostics: [
|
|
792
|
-
{
|
|
793
|
-
source,
|
|
794
|
-
sessionId,
|
|
795
|
-
sourcePath: options.sourcePath,
|
|
796
|
-
entries: entries.length,
|
|
797
|
-
malformedLines: options.malformedLines ?? 0,
|
|
798
|
-
hasExplicitTerminalSignal: explicitTerminal,
|
|
799
|
-
hasFinalOutput: observation.finalText !== null,
|
|
800
|
-
hasQualityLabel: qualityScore !== void 0,
|
|
801
|
-
hasTokenUsage: metrics.inputTokens > 0 || metrics.outputTokens > 0 || metrics.reasoningTokens > 0 || metrics.cachedTokens > 0 || metrics.cacheWriteTokens > 0,
|
|
802
|
-
hasCost: costProvenance.kind !== "uncaptured",
|
|
803
|
-
costKind: costProvenance.kind,
|
|
804
|
-
warnings
|
|
805
|
-
}
|
|
806
|
-
],
|
|
807
|
-
metrics: [metrics],
|
|
808
|
-
observations: [{ ...observation, sessionId }]
|
|
809
|
-
};
|
|
810
|
-
}
|
|
811
|
-
function terminalOutcomeFromSession(terminal) {
|
|
812
|
-
if (!terminal.explicit) return "unknown";
|
|
813
|
-
if (terminal.status === "completed") return "succeeded";
|
|
814
|
-
if (terminal.status === "failed") return "failed";
|
|
815
|
-
return "unknown";
|
|
816
|
-
}
|
|
817
|
-
function metricsFor(source, entries) {
|
|
818
|
-
switch (source) {
|
|
819
|
-
case "codex":
|
|
820
|
-
return codexMetrics(entries);
|
|
821
|
-
case "claude-code":
|
|
822
|
-
return claudeCodeMetrics(entries);
|
|
823
|
-
case "opencode":
|
|
824
|
-
return openCodeMetrics(entries);
|
|
825
|
-
case "kimi-code":
|
|
826
|
-
return kimiCodeMetrics(entries);
|
|
827
|
-
case "pi":
|
|
828
|
-
return piMetrics(entries);
|
|
829
|
-
}
|
|
830
|
-
}
|
|
831
|
-
function codexMetrics(entries) {
|
|
832
|
-
const metrics = emptyMetrics(entries.length);
|
|
833
|
-
const startedToolIds = /* @__PURE__ */ new Set();
|
|
834
|
-
let startedAt;
|
|
835
|
-
let completedAt;
|
|
836
|
-
for (const entry of entries) {
|
|
837
|
-
const payload = record2(entry.payload) ?? {};
|
|
838
|
-
const entryType = stringField2(entry, "type");
|
|
839
|
-
const payloadType = stringField2(payload, "type");
|
|
840
|
-
const timestamp2 = timestampMs(entry.timestamp);
|
|
841
|
-
if (timestamp2 !== void 0) {
|
|
842
|
-
startedAt = startedAt === void 0 ? timestamp2 : Math.min(startedAt, timestamp2);
|
|
843
|
-
completedAt = completedAt === void 0 ? timestamp2 : Math.max(completedAt, timestamp2);
|
|
844
|
-
}
|
|
845
|
-
if (entryType === "turn.started") metrics.turnsStarted += 1;
|
|
846
|
-
if (entryType === "turn.completed") {
|
|
847
|
-
metrics.turnsCompleted += 1;
|
|
848
|
-
addCodexExecUsage(metrics, record2(entry.usage));
|
|
849
|
-
}
|
|
850
|
-
if (entryType === "turn.failed") metrics.turnsAborted += 1;
|
|
851
|
-
if (entryType === "error") metrics.unclassifiedErrors += 1;
|
|
852
|
-
const item = record2(entry.item);
|
|
853
|
-
if (item && (entryType === "item.started" || entryType === "item.completed")) {
|
|
854
|
-
addCodexExecItem(metrics, startedToolIds, entryType, item);
|
|
855
|
-
}
|
|
856
|
-
if (entryType === "response_item") {
|
|
857
|
-
if (payloadType === "function_call" || payloadType === "custom_tool_call")
|
|
858
|
-
metrics.toolCalls += 1;
|
|
859
|
-
if (payloadType === "function_call_output" || payloadType === "custom_tool_call_output")
|
|
860
|
-
metrics.toolOutputs += 1;
|
|
861
|
-
if (payloadType === "reasoning") metrics.reasoningItems += 1;
|
|
862
|
-
if (payloadType === "message" && stringField2(payload, "role") === "assistant")
|
|
863
|
-
metrics.assistantMessages += 1;
|
|
864
|
-
}
|
|
865
|
-
if (entryType === "event_msg") {
|
|
866
|
-
if (payloadType === "user_message") metrics.userMessages += 1;
|
|
867
|
-
if (payloadType === "agent_message") metrics.assistantMessages += 1;
|
|
868
|
-
if (payloadType === "task_started") {
|
|
869
|
-
metrics.turnsStarted += 1;
|
|
870
|
-
const started = timestampMs(payload.started_at);
|
|
871
|
-
if (started !== void 0) {
|
|
872
|
-
startedAt = startedAt === void 0 ? started : Math.min(startedAt, started);
|
|
873
|
-
}
|
|
874
|
-
}
|
|
875
|
-
if (payloadType === "task_complete") {
|
|
876
|
-
metrics.turnsCompleted += 1;
|
|
877
|
-
const completed = timestampMs(payload.completed_at);
|
|
878
|
-
if (completed !== void 0) {
|
|
879
|
-
completedAt = completedAt === void 0 ? completed : Math.max(completedAt, completed);
|
|
880
|
-
}
|
|
881
|
-
const duration = numberField2(payload, "duration_ms");
|
|
882
|
-
if (duration !== void 0) metrics.wallMs = Math.max(metrics.wallMs, duration);
|
|
883
|
-
}
|
|
884
|
-
if (payloadType === "turn_aborted") {
|
|
885
|
-
metrics.turnsAborted += 1;
|
|
886
|
-
const completed = timestampMs(payload.completed_at);
|
|
887
|
-
if (completed !== void 0) {
|
|
888
|
-
completedAt = completedAt === void 0 ? completed : Math.max(completedAt, completed);
|
|
889
|
-
}
|
|
890
|
-
}
|
|
891
|
-
if (payloadType === "context_compacted") metrics.contextCompactions += 1;
|
|
892
|
-
if (payloadType === "patch_apply_end") {
|
|
893
|
-
metrics.patchAttempts += 1;
|
|
894
|
-
if (payload.success === true) metrics.patchSuccesses += 1;
|
|
895
|
-
else {
|
|
896
|
-
metrics.patchFailures += 1;
|
|
897
|
-
metrics.toolErrors += 1;
|
|
898
|
-
}
|
|
899
|
-
}
|
|
900
|
-
if (payloadType === "token_count")
|
|
901
|
-
setCodexCumulativeUsage(metrics, record2(record2(payload.info)?.total_token_usage));
|
|
902
|
-
const result = record2(payload.result);
|
|
903
|
-
if (result && "Err" in result) metrics.unclassifiedErrors += 1;
|
|
904
|
-
}
|
|
905
|
-
}
|
|
906
|
-
if (metrics.wallMs === 0 && startedAt !== void 0 && completedAt !== void 0) {
|
|
907
|
-
metrics.wallMs = Math.max(0, completedAt - startedAt);
|
|
908
|
-
}
|
|
909
|
-
metrics.processScore = codexProcessScore(metrics);
|
|
910
|
-
return metrics;
|
|
911
|
-
}
|
|
912
|
-
function addCodexExecItem(metrics, startedToolIds, eventType, item) {
|
|
913
|
-
const itemType = stringField2(item, "type");
|
|
914
|
-
const itemId = stringField2(item, "id");
|
|
915
|
-
const isTool = itemType === "command_execution" || itemType === "mcp_tool_call" || itemType === "collab_tool_call" || itemType === "web_search";
|
|
916
|
-
if (eventType === "item.started" && isTool) {
|
|
917
|
-
metrics.toolCalls += 1;
|
|
918
|
-
if (itemId) startedToolIds.add(itemId);
|
|
919
|
-
}
|
|
920
|
-
if (eventType === "item.completed" && isTool) {
|
|
921
|
-
if (!itemId || !startedToolIds.has(itemId)) metrics.toolCalls += 1;
|
|
922
|
-
metrics.toolOutputs += 1;
|
|
923
|
-
}
|
|
924
|
-
if (eventType !== "item.completed") return;
|
|
925
|
-
if (itemType === "agent_message" || itemType === "message") metrics.assistantMessages += 1;
|
|
926
|
-
if (itemType === "reasoning") metrics.reasoningItems += 1;
|
|
927
|
-
if (itemType === "file_change") {
|
|
928
|
-
metrics.patchAttempts += 1;
|
|
929
|
-
const status = stringField2(item, "status");
|
|
930
|
-
if (status === "completed") metrics.patchSuccesses += 1;
|
|
931
|
-
if (status === "failed") {
|
|
932
|
-
metrics.patchFailures += 1;
|
|
933
|
-
metrics.toolErrors += 1;
|
|
934
|
-
}
|
|
935
|
-
}
|
|
936
|
-
if (itemType === "error") metrics.unclassifiedErrors += 1;
|
|
937
|
-
else if (codexExecToolFailed(itemType, item)) metrics.toolErrors += 1;
|
|
938
|
-
}
|
|
939
|
-
function codexExecToolFailed(itemType, item) {
|
|
940
|
-
const status = stringField2(item, "status");
|
|
941
|
-
if (itemType === "command_execution") {
|
|
942
|
-
const exitCode = numberField2(item, "exit_code");
|
|
943
|
-
return status === "failed" || status === "declined" || exitCode !== void 0 && exitCode !== 0;
|
|
944
|
-
}
|
|
945
|
-
if (itemType === "mcp_tool_call") return status === "failed" || record2(item.error) !== null;
|
|
946
|
-
return itemType === "collab_tool_call" && status === "failed";
|
|
947
|
-
}
|
|
948
|
-
function addCodexExecUsage(metrics, usage) {
|
|
949
|
-
const parsed = readUsage(usage);
|
|
950
|
-
metrics.inputTokens += parsed.input;
|
|
951
|
-
metrics.outputTokens += parsed.output;
|
|
952
|
-
metrics.reasoningTokens += parsed.reasoning;
|
|
953
|
-
metrics.cachedTokens += parsed.cached;
|
|
954
|
-
metrics.cacheWriteTokens += parsed.cacheWrite;
|
|
955
|
-
}
|
|
956
|
-
function setCodexCumulativeUsage(metrics, usage) {
|
|
957
|
-
const parsed = readUsage(usage);
|
|
958
|
-
metrics.inputTokens = Math.max(metrics.inputTokens, parsed.input);
|
|
959
|
-
metrics.outputTokens = Math.max(metrics.outputTokens, parsed.output);
|
|
960
|
-
metrics.reasoningTokens = Math.max(metrics.reasoningTokens, parsed.reasoning);
|
|
961
|
-
metrics.cachedTokens = Math.max(metrics.cachedTokens, parsed.cached);
|
|
962
|
-
metrics.cacheWriteTokens = Math.max(metrics.cacheWriteTokens, parsed.cacheWrite);
|
|
963
|
-
}
|
|
964
|
-
function claudeCodeMetrics(entries) {
|
|
965
|
-
const metrics = emptyMetrics(entries.length);
|
|
966
|
-
let startedAt;
|
|
967
|
-
let completedAt;
|
|
968
|
-
for (const entry of entries) {
|
|
969
|
-
const type = stringField2(entry, "type");
|
|
970
|
-
const timestamp2 = timestampMs(entry.timestamp);
|
|
971
|
-
if (timestamp2 !== void 0) {
|
|
972
|
-
startedAt = startedAt === void 0 ? timestamp2 : Math.min(startedAt, timestamp2);
|
|
973
|
-
completedAt = completedAt === void 0 ? timestamp2 : Math.max(completedAt, timestamp2);
|
|
974
|
-
}
|
|
975
|
-
if (type === "user") metrics.userMessages += 1;
|
|
976
|
-
if (type === "assistant") metrics.assistantMessages += 1;
|
|
977
|
-
if (type === "pr-link") metrics.prLinks += 1;
|
|
978
|
-
if (type === "file-history-snapshot") metrics.fileSnapshots += 1;
|
|
979
|
-
const message = record2(entry.message);
|
|
980
|
-
if (message) addUsage(metrics, record2(message.usage));
|
|
981
|
-
const content = Array.isArray(message?.content) ? message.content : [];
|
|
982
|
-
for (const item of content) {
|
|
983
|
-
const part = record2(item);
|
|
984
|
-
if (!part) continue;
|
|
985
|
-
const partType = stringField2(part, "type");
|
|
986
|
-
if (partType === "thinking") metrics.reasoningItems += 1;
|
|
987
|
-
if (partType === "tool_use") metrics.toolCalls += 1;
|
|
988
|
-
if (partType === "tool_result") {
|
|
989
|
-
metrics.toolOutputs += 1;
|
|
990
|
-
if (part.is_error === true) metrics.toolErrors += 1;
|
|
991
|
-
}
|
|
992
|
-
}
|
|
993
|
-
}
|
|
994
|
-
if (startedAt !== void 0 && completedAt !== void 0)
|
|
995
|
-
metrics.wallMs = Math.max(0, completedAt - startedAt);
|
|
996
|
-
metrics.processScore = claudeProcessScore(metrics);
|
|
997
|
-
return metrics;
|
|
998
|
-
}
|
|
999
|
-
function openCodeMetrics(entries) {
|
|
1000
|
-
const metrics = emptyMetrics(entries.length);
|
|
1001
|
-
const messageUsage = emptyTokenTotals();
|
|
1002
|
-
const partUsage = emptyTokenTotals();
|
|
1003
|
-
let messageCost = 0;
|
|
1004
|
-
let partCost = 0;
|
|
1005
|
-
let messageCostCaptured = false;
|
|
1006
|
-
let partCostCaptured = false;
|
|
1007
|
-
let startedAt;
|
|
1008
|
-
let completedAt;
|
|
1009
|
-
for (const entry of entries) {
|
|
1010
|
-
const time = record2(entry.time);
|
|
1011
|
-
const created = timestampMs(time?.created);
|
|
1012
|
-
const completed = timestampMs(time?.completed);
|
|
1013
|
-
if (created !== void 0)
|
|
1014
|
-
startedAt = startedAt === void 0 ? created : Math.min(startedAt, created);
|
|
1015
|
-
if (completed !== void 0)
|
|
1016
|
-
completedAt = completedAt === void 0 ? completed : Math.max(completedAt, completed);
|
|
1017
|
-
const role = stringField2(entry, "role");
|
|
1018
|
-
if (role === "user") metrics.userMessages += 1;
|
|
1019
|
-
if (role === "assistant") {
|
|
1020
|
-
metrics.assistantMessages += 1;
|
|
1021
|
-
const finish = stringField2(entry, "finish");
|
|
1022
|
-
if (finish === "stop") metrics.turnsCompleted += 1;
|
|
1023
|
-
if (finish === "error") metrics.turnsAborted += 1;
|
|
1024
|
-
}
|
|
1025
|
-
const type = stringField2(entry, "type");
|
|
1026
|
-
if (type === "reasoning") metrics.reasoningItems += 1;
|
|
1027
|
-
if (type === "tool") {
|
|
1028
|
-
metrics.toolCalls += 1;
|
|
1029
|
-
const status = stringField2(record2(entry.state) ?? {}, "status");
|
|
1030
|
-
if (status === "completed") metrics.toolOutputs += 1;
|
|
1031
|
-
if (status === "error") {
|
|
1032
|
-
metrics.toolOutputs += 1;
|
|
1033
|
-
metrics.toolErrors += 1;
|
|
1034
|
-
}
|
|
1035
|
-
}
|
|
1036
|
-
if (type === "patch") {
|
|
1037
|
-
metrics.patchAttempts += 1;
|
|
1038
|
-
metrics.patchSuccesses += 1;
|
|
1039
|
-
}
|
|
1040
|
-
const cost = numberField2(entry, "cost");
|
|
1041
|
-
if (record2(entry.tokens) && role) {
|
|
1042
|
-
addUsageTo(messageUsage, record2(entry.tokens));
|
|
1043
|
-
if (cost !== void 0) {
|
|
1044
|
-
messageCost += cost;
|
|
1045
|
-
messageCostCaptured = true;
|
|
1046
|
-
}
|
|
1047
|
-
} else if (record2(entry.tokens)) {
|
|
1048
|
-
addUsageTo(partUsage, record2(entry.tokens));
|
|
1049
|
-
if (cost !== void 0) {
|
|
1050
|
-
partCost += cost;
|
|
1051
|
-
partCostCaptured = true;
|
|
1052
|
-
}
|
|
1053
|
-
}
|
|
1054
|
-
}
|
|
1055
|
-
const usage = hasTokenTotals(messageUsage) ? messageUsage : partUsage;
|
|
1056
|
-
metrics.inputTokens = usage.input;
|
|
1057
|
-
metrics.outputTokens = usage.output;
|
|
1058
|
-
metrics.reasoningTokens = usage.reasoning;
|
|
1059
|
-
metrics.cachedTokens = usage.cached;
|
|
1060
|
-
metrics.cacheWriteTokens = usage.cacheWrite;
|
|
1061
|
-
metrics.observedCostUsd = messageCostCaptured ? messageCost : partCost;
|
|
1062
|
-
metrics.observedCostCaptured = messageCostCaptured || partCostCaptured;
|
|
1063
|
-
if (metrics.wallMs === 0 && startedAt !== void 0 && completedAt !== void 0)
|
|
1064
|
-
metrics.wallMs = Math.max(0, completedAt - startedAt);
|
|
1065
|
-
metrics.processScore = terminalProcessScore(metrics);
|
|
1066
|
-
return metrics;
|
|
1067
|
-
}
|
|
1068
|
-
function kimiCodeMetrics(entries) {
|
|
1069
|
-
const metrics = emptyMetrics(entries.length);
|
|
1070
|
-
let startedAt;
|
|
1071
|
-
let completedAt;
|
|
1072
|
-
for (const entry of entries) {
|
|
1073
|
-
const timestamp2 = timestampMs(entry.timestamp);
|
|
1074
|
-
if (timestamp2 !== void 0) {
|
|
1075
|
-
startedAt = startedAt === void 0 ? timestamp2 : Math.min(startedAt, timestamp2);
|
|
1076
|
-
completedAt = completedAt === void 0 ? timestamp2 : Math.max(completedAt, timestamp2);
|
|
1077
|
-
}
|
|
1078
|
-
const role = stringField2(entry, "role");
|
|
1079
|
-
if (role === "user") metrics.userMessages += 1;
|
|
1080
|
-
if (role === "assistant") metrics.assistantMessages += 1;
|
|
1081
|
-
if (role === "_usage") addUsage(metrics, entry);
|
|
1082
|
-
const message = record2(entry.message);
|
|
1083
|
-
const messageType = message ? stringField2(message, "type") : void 0;
|
|
1084
|
-
const payload = record2(message?.payload) ?? {};
|
|
1085
|
-
if (messageType === "TurnBegin") {
|
|
1086
|
-
metrics.userMessages += 1;
|
|
1087
|
-
metrics.turnsStarted += 1;
|
|
1088
|
-
}
|
|
1089
|
-
if (messageType === "TurnEnd") metrics.turnsCompleted += 1;
|
|
1090
|
-
if (messageType === "StepInterrupted") metrics.turnsAborted += 1;
|
|
1091
|
-
if (messageType === "StepBegin") metrics.reasoningItems += 1;
|
|
1092
|
-
if (messageType === "ContentPart") {
|
|
1093
|
-
const payloadType = stringField2(payload, "type");
|
|
1094
|
-
if (payloadType === "think") metrics.reasoningItems += 1;
|
|
1095
|
-
if (payloadType === "text") metrics.assistantMessages += 1;
|
|
1096
|
-
}
|
|
1097
|
-
if (messageType === "ToolCall") metrics.toolCalls += 1;
|
|
1098
|
-
if (messageType === "ToolResult") {
|
|
1099
|
-
metrics.toolOutputs += 1;
|
|
1100
|
-
if (record2(payload.return_value)?.is_error === true) metrics.toolErrors += 1;
|
|
1101
|
-
}
|
|
1102
|
-
if (messageType === "StatusUpdate") addUsage(metrics, record2(payload.token_usage));
|
|
1103
|
-
if (messageType === "Notification") {
|
|
1104
|
-
const notification = record2(payload.payload);
|
|
1105
|
-
const exitCode = numberField2(notification ?? {}, "exit_code");
|
|
1106
|
-
if (notification?.timed_out === true || exitCode !== void 0 && exitCode !== 0) {
|
|
1107
|
-
metrics.toolErrors += 1;
|
|
1108
|
-
}
|
|
1109
|
-
}
|
|
1110
|
-
}
|
|
1111
|
-
if (startedAt !== void 0 && completedAt !== void 0)
|
|
1112
|
-
metrics.wallMs = Math.max(0, completedAt - startedAt);
|
|
1113
|
-
metrics.processScore = terminalProcessScore(metrics);
|
|
1114
|
-
return metrics;
|
|
1115
|
-
}
|
|
1116
|
-
function piMetrics(entries) {
|
|
1117
|
-
const metrics = emptyMetrics(entries.length);
|
|
1118
|
-
let bestReliabilityScore;
|
|
1119
|
-
for (const entry of entries) {
|
|
1120
|
-
const nodes = Array.isArray(entry.nodes) ? entry.nodes : [];
|
|
1121
|
-
const edges = Array.isArray(entry.edges) ? entry.edges : [];
|
|
1122
|
-
metrics.graphNodes += nodes.length;
|
|
1123
|
-
metrics.graphEdges += edges.length;
|
|
1124
|
-
for (const node of nodes) {
|
|
1125
|
-
const obj = record2(node);
|
|
1126
|
-
const ir = record2(obj?.ir) ?? obj;
|
|
1127
|
-
const kind = stringField2(ir ?? {}, "kind");
|
|
1128
|
-
if (kind === "ActionCandidate") metrics.actionCandidates += 1;
|
|
1129
|
-
if (kind === "ToolInvocation") metrics.toolCalls += 1;
|
|
1130
|
-
if (kind === "ToolResult") metrics.toolOutputs += 1;
|
|
1131
|
-
if (kind === "VerificationReport") metrics.verificationReports += 1;
|
|
1132
|
-
if (kind === "CompletionDecision") {
|
|
1133
|
-
metrics.completionDecisions += 1;
|
|
1134
|
-
metrics.turnsCompleted += 1;
|
|
1135
|
-
}
|
|
1136
|
-
}
|
|
1137
|
-
const averageReliability = numberField2(entry, "averageNodeReliability");
|
|
1138
|
-
if (averageReliability !== void 0)
|
|
1139
|
-
bestReliabilityScore = maxOptional(bestReliabilityScore, averageReliability);
|
|
1140
|
-
const pessimisticPath = numberField2(entry, "pessimisticPathEstimate");
|
|
1141
|
-
if (pessimisticPath !== void 0)
|
|
1142
|
-
bestReliabilityScore = maxOptional(bestReliabilityScore, pessimisticPath);
|
|
1143
|
-
const rows = Array.isArray(entry.rows) ? entry.rows : [];
|
|
1144
|
-
metrics.reliabilityRows += rows.length;
|
|
1145
|
-
for (const row of rows) {
|
|
1146
|
-
const obj = record2(row);
|
|
1147
|
-
if (!obj) continue;
|
|
1148
|
-
const lift = numberField2(obj, "lift");
|
|
1149
|
-
if (lift !== void 0) metrics.reliabilityLift = Math.max(metrics.reliabilityLift, lift);
|
|
1150
|
-
const validated = numberField2(obj, "validatedSuccessEstimate");
|
|
1151
|
-
if (validated !== void 0)
|
|
1152
|
-
bestReliabilityScore = maxOptional(bestReliabilityScore, validated);
|
|
1153
|
-
}
|
|
1154
|
-
}
|
|
1155
|
-
metrics.processScore = bestReliabilityScore !== void 0 ? clamp01(bestReliabilityScore) : metrics.completionDecisions > 0 ? 1 : 0;
|
|
1156
|
-
return metrics;
|
|
1157
|
-
}
|
|
1158
|
-
function emptyMetrics(entries) {
|
|
1159
|
-
return {
|
|
1160
|
-
entries,
|
|
1161
|
-
userMessages: 0,
|
|
1162
|
-
assistantMessages: 0,
|
|
1163
|
-
reasoningItems: 0,
|
|
1164
|
-
toolCalls: 0,
|
|
1165
|
-
toolOutputs: 0,
|
|
1166
|
-
toolErrors: 0,
|
|
1167
|
-
unclassifiedErrors: 0,
|
|
1168
|
-
patchAttempts: 0,
|
|
1169
|
-
patchSuccesses: 0,
|
|
1170
|
-
patchFailures: 0,
|
|
1171
|
-
turnsStarted: 0,
|
|
1172
|
-
turnsCompleted: 0,
|
|
1173
|
-
turnsAborted: 0,
|
|
1174
|
-
contextCompactions: 0,
|
|
1175
|
-
mcpCalls: 0,
|
|
1176
|
-
subagentCalls: 0,
|
|
1177
|
-
skillCalls: 0,
|
|
1178
|
-
hookCalls: 0,
|
|
1179
|
-
webCalls: 0,
|
|
1180
|
-
codeActions: 0,
|
|
1181
|
-
prLinks: 0,
|
|
1182
|
-
fileSnapshots: 0,
|
|
1183
|
-
graphNodes: 0,
|
|
1184
|
-
graphEdges: 0,
|
|
1185
|
-
actionCandidates: 0,
|
|
1186
|
-
verificationReports: 0,
|
|
1187
|
-
completionDecisions: 0,
|
|
1188
|
-
reliabilityRows: 0,
|
|
1189
|
-
reliabilityLift: 0,
|
|
1190
|
-
inputTokens: 0,
|
|
1191
|
-
outputTokens: 0,
|
|
1192
|
-
reasoningTokens: 0,
|
|
1193
|
-
cachedTokens: 0,
|
|
1194
|
-
cacheWriteTokens: 0,
|
|
1195
|
-
observedCostUsd: 0,
|
|
1196
|
-
observedCostCaptured: false,
|
|
1197
|
-
wallMs: 0,
|
|
1198
|
-
processScore: 0
|
|
1199
|
-
};
|
|
1200
|
-
}
|
|
1201
|
-
function codexProcessScore(metrics) {
|
|
1202
|
-
const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted;
|
|
1203
|
-
const base = terminalTurns > 0 ? metrics.turnsCompleted / terminalTurns : metrics.assistantMessages > 0 ? 0.5 : 0;
|
|
1204
|
-
return penalizeErrors(base, metrics);
|
|
1205
|
-
}
|
|
1206
|
-
function claudeProcessScore(metrics) {
|
|
1207
|
-
const base = metrics.prLinks > 0 ? 1 : metrics.assistantMessages > 0 ? 0.75 : 0;
|
|
1208
|
-
return penalizeErrors(base, metrics);
|
|
1209
|
-
}
|
|
1210
|
-
function terminalProcessScore(metrics) {
|
|
1211
|
-
const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted;
|
|
1212
|
-
const base = terminalTurns > 0 ? metrics.turnsCompleted / terminalTurns : metrics.assistantMessages > 0 ? 0.75 : metrics.toolCalls > 0 ? 0.5 : 0;
|
|
1213
|
-
return penalizeErrors(base, metrics);
|
|
1214
|
-
}
|
|
1215
|
-
function penalizeErrors(base, metrics) {
|
|
1216
|
-
const operations = Math.max(1, metrics.toolCalls + metrics.patchAttempts);
|
|
1217
|
-
const errorRate = Math.min(1, (metrics.toolErrors + metrics.patchFailures) / operations);
|
|
1218
|
-
return clamp01(base * (1 - 0.5 * errorRate));
|
|
1219
|
-
}
|
|
1220
|
-
function emptyTokenTotals() {
|
|
1221
|
-
return { input: 0, output: 0, reasoning: 0, cached: 0, cacheWrite: 0 };
|
|
1222
|
-
}
|
|
1223
|
-
function hasTokenTotals(usage) {
|
|
1224
|
-
return usage.input > 0 || usage.output > 0 || usage.reasoning > 0 || usage.cached > 0 || usage.cacheWrite > 0;
|
|
1225
|
-
}
|
|
1226
|
-
function addUsage(metrics, usage) {
|
|
1227
|
-
const parsed = readUsage(usage);
|
|
1228
|
-
metrics.inputTokens += parsed.input;
|
|
1229
|
-
metrics.outputTokens += parsed.output;
|
|
1230
|
-
metrics.reasoningTokens += parsed.reasoning;
|
|
1231
|
-
metrics.cachedTokens += parsed.cached;
|
|
1232
|
-
metrics.cacheWriteTokens += parsed.cacheWrite;
|
|
1233
|
-
}
|
|
1234
|
-
function addUsageTo(totals, usage) {
|
|
1235
|
-
const parsed = readUsage(usage);
|
|
1236
|
-
totals.input += parsed.input;
|
|
1237
|
-
totals.output += parsed.output;
|
|
1238
|
-
totals.reasoning += parsed.reasoning;
|
|
1239
|
-
totals.cached += parsed.cached;
|
|
1240
|
-
totals.cacheWrite += parsed.cacheWrite;
|
|
1241
|
-
}
|
|
1242
|
-
function readUsage(usage) {
|
|
1243
|
-
const parsed = extractUsage(usage);
|
|
1244
|
-
return {
|
|
1245
|
-
input: parsed?.input ?? 0,
|
|
1246
|
-
output: parsed?.output ?? 0,
|
|
1247
|
-
reasoning: parsed?.reasoning ?? 0,
|
|
1248
|
-
cached: parsed?.cached ?? 0,
|
|
1249
|
-
cacheWrite: parsed?.cacheWrite ?? 0
|
|
1250
|
-
};
|
|
1251
|
-
}
|
|
1252
|
-
function resolveSessionCost(explicit, metrics, model) {
|
|
1253
|
-
if (explicit) {
|
|
1254
|
-
if (explicit.kind === "uncaptured") {
|
|
1255
|
-
if (explicit.usd !== null) throw new Error("uncaptured cost must have usd: null");
|
|
1256
|
-
return { costUsd: null, costProvenance: explicit };
|
|
1257
|
-
}
|
|
1258
|
-
if (!Number.isFinite(explicit.usd) || explicit.usd < 0) {
|
|
1259
|
-
throw new Error(`${explicit.kind} cost must be a finite, non-negative USD amount`);
|
|
1260
|
-
}
|
|
1261
|
-
return { costUsd: explicit.usd, costProvenance: explicit };
|
|
1262
|
-
}
|
|
1263
|
-
if (metrics.observedCostCaptured) {
|
|
1264
|
-
const costProvenance = { kind: "observed", usd: metrics.observedCostUsd };
|
|
1265
|
-
return { costUsd: metrics.observedCostUsd, costProvenance };
|
|
1266
|
-
}
|
|
1267
|
-
const hasUsage = metrics.inputTokens > 0 || metrics.outputTokens > 0;
|
|
1268
|
-
if (hasUsage && isModelPriced(model)) {
|
|
1269
|
-
const costUsd = estimateCost(metrics.inputTokens, metrics.outputTokens, model);
|
|
1270
|
-
return { costUsd, costProvenance: { kind: "estimated", usd: costUsd } };
|
|
1271
|
-
}
|
|
1272
|
-
return { costUsd: null, costProvenance: { kind: "uncaptured", usd: null } };
|
|
1273
|
-
}
|
|
1274
|
-
function diagnosticsFor(metrics, options) {
|
|
1275
|
-
const warnings = [];
|
|
1276
|
-
if (options.invalidQualityScore) warnings.push("non-finite task quality score omitted");
|
|
1277
|
-
else if (!options.hasQualityLabel) warnings.push("task quality score not supplied");
|
|
1278
|
-
if (!options.explicitTerminal) warnings.push("no explicit terminal success/failure signal");
|
|
1279
|
-
if (metrics.inputTokens === 0 && metrics.outputTokens === 0 && metrics.reasoningTokens === 0 && metrics.cachedTokens === 0 && metrics.cacheWriteTokens === 0) {
|
|
1280
|
-
warnings.push("missing token usage");
|
|
1281
|
-
}
|
|
1282
|
-
if (options.costKind === "estimated")
|
|
1283
|
-
warnings.push("USD cost estimated from token usage and model pricing");
|
|
1284
|
-
if (options.costKind === "uncaptured") warnings.push("USD cost uncaptured");
|
|
1285
|
-
if (options.costKind === "uncaptured" && !isModelPriced(options.model))
|
|
1286
|
-
warnings.push("model pricing unknown");
|
|
1287
|
-
if (options.malformedLines > 0)
|
|
1288
|
-
warnings.push(`${options.malformedLines} malformed JSONL lines skipped`);
|
|
1289
|
-
return warnings;
|
|
1290
|
-
}
|
|
1291
|
-
function applyObservedSurfaceMetrics(metrics, observation) {
|
|
1292
|
-
const actions = observation.actions.filter(
|
|
1293
|
-
(action) => action.kind === "tool" || action.kind === "patch"
|
|
1294
|
-
);
|
|
1295
|
-
metrics.mcpCalls = actions.filter((action) => action.surface === "mcp").length;
|
|
1296
|
-
metrics.subagentCalls = actions.filter((action) => action.surface === "subagent").length;
|
|
1297
|
-
metrics.skillCalls = actions.filter((action) => action.surface === "skill").length;
|
|
1298
|
-
metrics.hookCalls = actions.filter((action) => action.surface === "hook").length;
|
|
1299
|
-
metrics.webCalls = actions.filter((action) => action.surface === "web").length;
|
|
1300
|
-
metrics.codeActions = actions.filter((action) => action.surface === "code").length;
|
|
1301
|
-
}
|
|
1302
|
-
function sessionIdFromEntries2(source, entries) {
|
|
1303
|
-
for (const entry of entries) {
|
|
1304
|
-
if (source === "codex") {
|
|
1305
|
-
const threadId = stringField2(entry, "thread_id");
|
|
1306
|
-
if (threadId) return threadId;
|
|
1307
|
-
const payload2 = record2(entry.payload);
|
|
1308
|
-
const id = payload2 ? stringField2(payload2, "id") : void 0;
|
|
1309
|
-
if (id) return id;
|
|
1310
|
-
}
|
|
1311
|
-
const message = record2(entry.message);
|
|
1312
|
-
const payload = record2(message?.payload);
|
|
1313
|
-
const sessionId = stringField2(entry, "sessionID") ?? stringField2(entry, "sessionId") ?? stringField2(entry, "session_id") ?? stringField2(payload ?? {}, "session_id");
|
|
1314
|
-
if (sessionId) return sessionId;
|
|
1315
|
-
}
|
|
1316
|
-
return void 0;
|
|
1317
|
-
}
|
|
1318
|
-
function modelFromEntries(source, entries) {
|
|
1319
|
-
let providerModel;
|
|
1320
|
-
for (const entry of entries) {
|
|
1321
|
-
if (source === "codex") {
|
|
1322
|
-
const payload = record2(entry.payload);
|
|
1323
|
-
const model2 = payload ? stringField2(payload, "model") : void 0;
|
|
1324
|
-
if (model2) return model2;
|
|
1325
|
-
providerModel = providerModel ?? (payload ? stringField2(payload, "model_provider") : void 0);
|
|
1326
|
-
}
|
|
1327
|
-
const openCodeModel = stringField2(entry, "modelID");
|
|
1328
|
-
if (openCodeModel) {
|
|
1329
|
-
const provider = stringField2(entry, "providerID");
|
|
1330
|
-
return provider ? `${provider}/${openCodeModel}` : openCodeModel;
|
|
1331
|
-
}
|
|
1332
|
-
const modelObject = record2(entry.model);
|
|
1333
|
-
const nestedModel = stringField2(modelObject ?? {}, "modelID");
|
|
1334
|
-
if (nestedModel) {
|
|
1335
|
-
const provider = stringField2(modelObject ?? {}, "providerID");
|
|
1336
|
-
return provider ? `${provider}/${nestedModel}` : nestedModel;
|
|
1337
|
-
}
|
|
1338
|
-
const message = record2(entry.message);
|
|
1339
|
-
const model = message ? stringField2(message, "model") : void 0;
|
|
1340
|
-
if (model) return model;
|
|
1341
|
-
}
|
|
1342
|
-
return providerModel;
|
|
1343
|
-
}
|
|
1344
|
-
function cwdFromEntries(entries) {
|
|
1345
|
-
for (const entry of entries) {
|
|
1346
|
-
const path = record2(entry.path);
|
|
1347
|
-
const cwd = stringField2(entry, "cwd") ?? stringField2(path ?? {}, "cwd") ?? stringField2(path ?? {}, "root") ?? (record2(entry.payload) ? stringField2(record2(entry.payload), "cwd") : void 0);
|
|
1348
|
-
if (cwd) return cwd;
|
|
1349
|
-
}
|
|
1350
|
-
return void 0;
|
|
1351
|
-
}
|
|
1352
|
-
function firstUserText(entries) {
|
|
1353
|
-
for (const entry of entries) {
|
|
1354
|
-
const payload = record2(entry.payload);
|
|
1355
|
-
const payloadType = payload ? stringField2(payload, "type") : void 0;
|
|
1356
|
-
if (payloadType === "user_message") return stringField2(payload, "message");
|
|
1357
|
-
if (stringField2(entry, "role") === "user") {
|
|
1358
|
-
const content = entry.content;
|
|
1359
|
-
if (typeof content === "string") return content;
|
|
1360
|
-
}
|
|
1361
|
-
const wirePayload = record2(record2(entry.message)?.payload);
|
|
1362
|
-
const userInput = wirePayload ? stringField2(wirePayload, "user_input") : void 0;
|
|
1363
|
-
if (userInput) return userInput;
|
|
1364
|
-
const message = record2(entry.message);
|
|
1365
|
-
if (message && stringField2(message, "role") === "user") {
|
|
1366
|
-
const content = message.content;
|
|
1367
|
-
if (typeof content === "string") return content;
|
|
1368
|
-
if (Array.isArray(content)) {
|
|
1369
|
-
const text = content.map((part) => {
|
|
1370
|
-
const obj = record2(part);
|
|
1371
|
-
return obj ? stringField2(obj, "text") : void 0;
|
|
1372
|
-
}).filter((part) => part !== void 0).join("\n");
|
|
1373
|
-
if (text) return text;
|
|
1374
|
-
}
|
|
1375
|
-
}
|
|
1376
|
-
}
|
|
1377
|
-
return void 0;
|
|
1378
|
-
}
|
|
1379
|
-
function fallbackSessionId2(source, sourcePath) {
|
|
1380
|
-
return hashString(`${source}:${sourcePath ?? "unknown-session"}`).slice(
|
|
1381
|
-
"sha256:".length,
|
|
1382
|
-
"sha256:".length + 20
|
|
1383
|
-
);
|
|
1384
|
-
}
|
|
1385
|
-
function withSnapshot(model) {
|
|
1386
|
-
const trimmed = model.trim() || "unknown";
|
|
1387
|
-
if (trimmed.includes("@") || /-\d{8}$/.test(trimmed) || /-\d{4}-\d{2}-\d{2}$/.test(trimmed)) {
|
|
1388
|
-
return trimmed;
|
|
1389
|
-
}
|
|
1390
|
-
return `${trimmed}@observed-local`;
|
|
1391
|
-
}
|
|
1392
|
-
function hashJson(value) {
|
|
1393
|
-
return hashString(JSON.stringify(value));
|
|
1394
|
-
}
|
|
1395
|
-
function hashString(value) {
|
|
1396
|
-
return `sha256:${createHash("sha256").update(value).digest("hex")}`;
|
|
1397
|
-
}
|
|
1398
|
-
function stableSeed(value) {
|
|
1399
|
-
return createHash("sha256").update(value).digest().readUInt32BE(0);
|
|
1400
|
-
}
|
|
1401
|
-
function maxOptional(current, next) {
|
|
1402
|
-
return current === void 0 ? next : Math.max(current, next);
|
|
1403
|
-
}
|
|
1404
|
-
function clamp01(value) {
|
|
1405
|
-
if (!Number.isFinite(value)) return 0;
|
|
1406
|
-
return Math.max(0, Math.min(1, value));
|
|
1407
|
-
}
|
|
1408
|
-
function finiteScore(value) {
|
|
1409
|
-
return value !== void 0 && Number.isFinite(value) ? clamp01(value) : void 0;
|
|
1410
|
-
}
|
|
1411
|
-
function isRecord2(value) {
|
|
1412
|
-
return value !== null && typeof value === "object" && !Array.isArray(value);
|
|
1413
|
-
}
|
|
1414
|
-
function record2(value) {
|
|
1415
|
-
return isRecord2(value) ? value : null;
|
|
1416
|
-
}
|
|
1417
|
-
function stringField2(obj, key) {
|
|
1418
|
-
const value = obj[key];
|
|
1419
|
-
return typeof value === "string" && value.length > 0 ? value : void 0;
|
|
1420
|
-
}
|
|
1421
|
-
function numberField2(obj, key) {
|
|
1422
|
-
const value = obj[key];
|
|
1423
|
-
return typeof value === "number" && Number.isFinite(value) ? value : void 0;
|
|
1424
|
-
}
|
|
1425
|
-
function timestampMs(value) {
|
|
1426
|
-
if (typeof value === "number" && Number.isFinite(value)) {
|
|
1427
|
-
return value > 1e12 ? value : value * 1e3;
|
|
1428
|
-
}
|
|
1429
|
-
if (typeof value === "string" && value.length > 0) {
|
|
1430
|
-
const parsed = Date.parse(value);
|
|
1431
|
-
return Number.isFinite(parsed) ? parsed : void 0;
|
|
1432
|
-
}
|
|
1433
|
-
return void 0;
|
|
1434
|
-
}
|
|
1435
|
-
|
|
1436
|
-
export {
|
|
1437
|
-
observeCodeAgentSession,
|
|
1438
|
-
parseCodeAgentJsonl,
|
|
1439
|
-
fromCodexSession,
|
|
1440
|
-
fromClaudeCodeSession,
|
|
1441
|
-
fromOpenCodeSession,
|
|
1442
|
-
fromKimiCodeSession,
|
|
1443
|
-
fromPiSession,
|
|
1444
|
-
fromPigraphSession
|
|
1445
|
-
};
|
|
1446
|
-
//# sourceMappingURL=chunk-SZLVEKMJ.js.map
|