@tangle-network/agent-eval 0.150.1 → 0.161.0
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 +178 -1
- package/README.md +7 -3
- package/dist/{active-curriculum-C4mk67HP.js → active-curriculum-CD5TU2yW.js} +3 -13
- package/dist/active-curriculum-CD5TU2yW.js.map +1 -0
- package/dist/{agent-profile-cell-BkcRDikH.d.ts → agent-profile-cell-CTOZJUuE.d.ts} +4 -2
- package/dist/agent-profile-cell-CTOZJUuE.d.ts.map +1 -0
- package/dist/analyst/index.d.ts +19 -36
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +8 -8
- package/dist/analyst/index.js.map +1 -1
- package/dist/{backend-integrity-DOCa_QrR.d.ts → backend-integrity-DxuQCu_A.d.ts} +4 -3
- package/dist/backend-integrity-DxuQCu_A.d.ts.map +1 -0
- package/dist/{benchmark-BtAWA8nT.d.ts → benchmark-CGPp-kDC.d.ts} +3 -3
- package/dist/{benchmark-BtAWA8nT.d.ts.map → benchmark-CGPp-kDC.d.ts.map} +1 -1
- package/dist/{benchmark-command-CAFwbH0L.js → benchmark-command-BVtaq_ve.js} +26 -31
- package/dist/benchmark-command-BVtaq_ve.js.map +1 -0
- package/dist/benchmarks/index.d.ts +6 -19
- package/dist/benchmarks/index.d.ts.map +1 -1
- package/dist/benchmarks/index.js +4 -4
- package/dist/benchmarks/index.js.map +1 -1
- package/dist/builder-eval/index.d.ts +3 -3
- package/dist/builder-eval/index.js +3 -3
- package/dist/campaign/index.d.ts +9 -9
- package/dist/campaign/index.js +7 -7
- package/dist/{campaign-CN_7xJdV.js → campaign-BSmOwskD.js} +77 -795
- package/dist/campaign-BSmOwskD.js.map +1 -0
- package/dist/{canonical-D-XsTQ6_.js → canonical-IL-Bu-14.js} +26 -2
- package/dist/canonical-IL-Bu-14.js.map +1 -0
- package/dist/{capture-fetch-BBVFzhkk.d.ts → capture-fetch-CqwsJkkG.d.ts} +3 -3
- package/dist/{capture-fetch-BBVFzhkk.d.ts.map → capture-fetch-CqwsJkkG.d.ts.map} +1 -1
- package/dist/{chat-client-2bVfrzhN.js → chat-client-DlMlAeYI.js} +5 -55
- package/dist/{chat-client-2bVfrzhN.js.map → chat-client-DlMlAeYI.js.map} +1 -1
- package/dist/chat-json-call-6g5sJobJ.js +53 -0
- package/dist/chat-json-call-6g5sJobJ.js.map +1 -0
- package/dist/cli.js +54 -19
- package/dist/cli.js.map +1 -1
- package/dist/{client-LIuo-KPv.js → client-CX7KqIdB.js} +3 -3
- package/dist/client-CX7KqIdB.js.map +1 -0
- package/dist/{client-kPQYT_56.d.ts → client-L9VVPkim.d.ts} +4 -4
- package/dist/{client-kPQYT_56.d.ts.map → client-L9VVPkim.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +13 -27
- package/dist/contract/index.d.ts.map +1 -1
- package/dist/contract/index.js +14 -17
- package/dist/contract/index.js.map +1 -1
- package/dist/{counterfactual--bpysZF0.d.ts → counterfactual-BaFUWK3H.d.ts} +4 -4
- package/dist/{counterfactual--bpysZF0.d.ts.map → counterfactual-BaFUWK3H.d.ts.map} +1 -1
- package/dist/{counterfactual-lDfCx0Uz.js → counterfactual-D_VWavVm.js} +2 -2
- package/dist/{counterfactual-lDfCx0Uz.js.map → counterfactual-D_VWavVm.js.map} +1 -1
- package/dist/{dataset-CJjKqQfA.d.ts → dataset-DQqhOCPt.d.ts} +5 -4
- package/dist/{dataset-CJjKqQfA.d.ts.map → dataset-DQqhOCPt.d.ts.map} +1 -1
- package/dist/{default-registry-Cw0Ohdoj.d.ts → default-registry-G9CKMNkc.d.ts} +7 -8
- package/dist/{default-registry-Cw0Ohdoj.d.ts.map → default-registry-G9CKMNkc.d.ts.map} +1 -1
- package/dist/{define-agent-eval-CEQWL9Hy.d.ts → define-agent-eval-Dx1JnPEa.d.ts} +26 -6
- package/dist/define-agent-eval-Dx1JnPEa.d.ts.map +1 -0
- package/dist/{define-agent-eval-rqNyVhVV.js → define-agent-eval-h-s-sI-v.js} +17 -11
- package/dist/define-agent-eval-h-s-sI-v.js.map +1 -0
- package/dist/{descriptive-B5MwKfbf.js → descriptive-jDOuI6mz.js} +22 -2
- package/dist/descriptive-jDOuI6mz.js.map +1 -0
- package/dist/{dspy-rlm-engine-DHI0WrUU.js → dspy-rlm-engine-DptEII26.js} +95 -15
- package/dist/dspy-rlm-engine-DptEII26.js.map +1 -0
- package/dist/{emitter-CPBAhxum.js → emitter-BpYFQPj4.js} +2 -18
- package/dist/emitter-BpYFQPj4.js.map +1 -0
- package/dist/{emitter-DGQGoLyj.d.ts → emitter-D_jYSGRd.d.ts} +4 -20
- package/dist/{emitter-DGQGoLyj.d.ts.map → emitter-D_jYSGRd.d.ts.map} +1 -1
- package/dist/{engine-BLzhNzoY.d.ts → engine-Cu5qD5Fc.d.ts} +9 -11
- package/dist/{engine-BLzhNzoY.d.ts.map → engine-Cu5qD5Fc.d.ts.map} +1 -1
- package/dist/{eval-campaign-C4jmuM-b.js → eval-campaign-BsXWL2-2.js} +17 -28
- package/dist/eval-campaign-BsXWL2-2.js.map +1 -0
- package/dist/{exact-types-ccQAyut1.d.ts → exact-types-qnexxJ1Z.d.ts} +2 -2
- package/dist/{exact-types-ccQAyut1.d.ts.map → exact-types-qnexxJ1Z.d.ts.map} +1 -1
- package/dist/{exec-y-DCLqK7.js → exec-D9WpA2p-.js} +2 -2
- package/dist/exec-D9WpA2p-.js.map +1 -0
- package/dist/experiment/index.d.ts +45 -11
- package/dist/experiment/index.d.ts.map +1 -1
- package/dist/experiment/index.js +30 -11
- package/dist/experiment/index.js.map +1 -1
- package/dist/{experiment-tracker-0MhuPArU.d.ts → experiment-tracker-DCO6Cz4s.d.ts} +2 -2
- package/dist/{experiment-tracker-0MhuPArU.d.ts.map → experiment-tracker-DCO6Cz4s.d.ts.map} +1 -1
- package/dist/{exporters-q9iL-2Jf.js → exporters-Df7TgHFv.js} +3 -3
- package/dist/exporters-Df7TgHFv.js.map +1 -0
- package/dist/{external-optimizer-contracts-DbLsm4Po.d.ts → external-optimizer-contracts-szBJ_1vh.d.ts} +2 -2
- package/dist/{external-optimizer-contracts-DbLsm4Po.d.ts.map → external-optimizer-contracts-szBJ_1vh.d.ts.map} +1 -1
- package/dist/{external-optimizer-process-Bhmzngf-.js → external-optimizer-process-WosTBChy.js} +4 -4
- package/dist/{external-optimizer-process-Bhmzngf-.js.map → external-optimizer-process-WosTBChy.js.map} +1 -1
- package/dist/{external-optimizer-subprocess-Dn90UqN2.js → external-optimizer-subprocess-BIWbHpgD.js} +10 -6
- package/dist/external-optimizer-subprocess-BIWbHpgD.js.map +1 -0
- package/dist/{failure-cluster-BLURuWG4.d.ts → failure-cluster-CXL8NbEw.d.ts} +3 -3
- package/dist/{failure-cluster-BLURuWG4.d.ts.map → failure-cluster-CXL8NbEw.d.ts.map} +1 -1
- package/dist/{feedback-trajectory-DpTTjo0q.d.ts → feedback-trajectory-B3ZHaHV_.d.ts} +7 -7
- package/dist/{feedback-trajectory-DpTTjo0q.d.ts.map → feedback-trajectory-B3ZHaHV_.d.ts.map} +1 -1
- package/dist/fuzz.js +1 -1
- package/dist/{hf-dataset-XggBupCr.js → hf-dataset-D8_RNIis.js} +4 -4
- package/dist/{hf-dataset-XggBupCr.js.map → hf-dataset-D8_RNIis.js.map} +1 -1
- package/dist/hosted/index.d.ts +3 -14
- package/dist/hosted/index.d.ts.map +1 -1
- package/dist/hosted/index.js +2 -2
- package/dist/{index-CTKpu9ry.d.ts → index-CGtH1piv.d.ts} +48 -26
- package/dist/index-CGtH1piv.d.ts.map +1 -0
- package/dist/{index-BNPtkBPf.d.ts → index-D-IiQIBB.d.ts} +5 -10
- package/dist/index-D-IiQIBB.d.ts.map +1 -0
- package/dist/{index-IQccV3Ou.d.ts → index-D-V8gCs_.d.ts} +13 -90
- package/dist/index-D-V8gCs_.d.ts.map +1 -0
- package/dist/{index-B8Ui1mr1.d.ts → index-lfaSeKSD.d.ts} +18 -2
- package/dist/index-lfaSeKSD.d.ts.map +1 -0
- package/dist/index-vrJugRal.d.ts +1 -0
- package/dist/index.d.ts +68 -125
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +61 -112
- package/dist/index.js.map +1 -1
- package/dist/{insight-report-BeT8KCgI.d.ts → insight-report-DRe8LB6d.d.ts} +4 -4
- package/dist/{insight-report-BeT8KCgI.d.ts.map → insight-report-DRe8LB6d.d.ts.map} +1 -1
- package/dist/{integrity-DysDBWDu.js → integrity-CyWSSoQS.js} +17 -4
- package/dist/integrity-CyWSSoQS.js.map +1 -0
- package/dist/{integrity-B0dZ96EO.d.ts → integrity-DUNX9Fao.d.ts} +3 -3
- package/dist/{integrity-B0dZ96EO.d.ts.map → integrity-DUNX9Fao.d.ts.map} +1 -1
- package/dist/{judge-calibration-DZkWrm5H.js → judge-calibration-zZjLz8hr.js} +2 -2
- package/dist/{judge-calibration-DZkWrm5H.js.map → judge-calibration-zZjLz8hr.js.map} +1 -1
- package/dist/{kind-factory-DmAa0h3K.js → kind-factory-DY8FdoXf.js} +3 -71
- package/dist/kind-factory-DY8FdoXf.js.map +1 -0
- package/dist/ledger-core/index.d.ts +2 -2
- package/dist/ledger-core/index.js +3 -3
- package/dist/{ledger-core-DTae9rv_.js → ledger-core-BOzlRygb.js} +2 -2
- package/dist/{ledger-core-DTae9rv_.js.map → ledger-core-BOzlRygb.js.map} +1 -1
- package/dist/{llm-client-Bg32RW0j.js → llm-client-hgDieDNN.js} +53 -98
- package/dist/llm-client-hgDieDNN.js.map +1 -0
- package/dist/{llm-judge-CVq33oz1.js → llm-judge-BhasIPFT.js} +1136 -78
- package/dist/llm-judge-BhasIPFT.js.map +1 -0
- package/dist/{matrix-DrVnRp4G.d.ts → matrix-eXKRMHnL.d.ts} +74 -72
- package/dist/matrix-eXKRMHnL.d.ts.map +1 -0
- package/dist/meta-eval/index.d.ts +8 -6
- package/dist/meta-eval/index.d.ts.map +1 -1
- package/dist/meta-eval/index.js +9 -7
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/{mint-BV6tLVWl.js → mint-DfODW1KW.js} +3 -3
- package/dist/{mint-BV6tLVWl.js.map → mint-DfODW1KW.js.map} +1 -1
- package/dist/multishot/golden/index.d.ts +2 -8
- package/dist/multishot/golden/index.d.ts.map +1 -1
- package/dist/multishot/golden/index.js +56 -86
- package/dist/multishot/golden/index.js.map +1 -1
- package/dist/multishot/index.d.ts +11 -46
- package/dist/multishot/index.d.ts.map +1 -1
- package/dist/multishot/index.js +30 -83
- package/dist/multishot/index.js.map +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{opencode-sqlite-DJWAXLms.js → opencode-sqlite-eK6HW6dr.js} +2 -6
- package/dist/{opencode-sqlite-DJWAXLms.js.map → opencode-sqlite-eK6HW6dr.js.map} +1 -1
- package/dist/pipelines/index.d.ts +5 -5
- package/dist/pipelines/index.js +3 -3
- package/dist/{pre-registration-zFSLEiFU.d.ts → pre-registration-CzFCcwYk.d.ts} +55 -40
- package/dist/pre-registration-CzFCcwYk.d.ts.map +1 -0
- package/dist/pre-registration-KN9jkh58.js +110 -0
- package/dist/pre-registration-KN9jkh58.js.map +1 -0
- package/dist/{produced-state-jfk8Du3b.js → produced-state-DZ89riy5.js} +8 -8
- package/dist/produced-state-DZ89riy5.js.map +1 -0
- package/dist/profile-cell.d.ts +1 -1
- package/dist/profile-cell.js +31 -5
- package/dist/profile-cell.js.map +1 -1
- package/dist/{promotion-policy-DLOUkYhI.d.ts → promotion-policy-DtnOIZvk.d.ts} +2 -2
- package/dist/{promotion-policy-DLOUkYhI.d.ts.map → promotion-policy-DtnOIZvk.d.ts.map} +1 -1
- package/dist/{query-Di7eEQ79.js → query-CHmMP42p.js} +20 -11
- package/dist/query-CHmMP42p.js.map +1 -0
- package/dist/{query-CJ_DX8vl.d.ts → query-DxPYqpmT.d.ts} +10 -4
- package/dist/query-DxPYqpmT.d.ts.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/{registry-BQwrSYpC.d.ts → registry-8You7OK1.d.ts} +5 -7
- package/dist/{registry-BQwrSYpC.d.ts.map → registry-8You7OK1.d.ts.map} +1 -1
- package/dist/{release-confidence-BknrpBnO.js → release-confidence-DKfD2RYU.js} +28 -14
- package/dist/release-confidence-DKfD2RYU.js.map +1 -0
- package/dist/{release-confidence-4XrqlpFD.d.ts → release-confidence-Dqt0NFep.d.ts} +7 -6
- package/dist/release-confidence-Dqt0NFep.d.ts.map +1 -0
- package/dist/reporting.d.ts +3 -3
- package/dist/reporting.js +3 -3
- package/dist/{researcher-DJnoUE8c.d.ts → researcher-Cz565b7D.d.ts} +34 -21
- package/dist/researcher-Cz565b7D.d.ts.map +1 -0
- package/dist/{reward-hacking-62tojkQd.d.ts → reward-hacking-MBf7qpSB.d.ts} +2 -2
- package/dist/{reward-hacking-62tojkQd.d.ts.map → reward-hacking-MBf7qpSB.d.ts.map} +1 -1
- package/dist/{reward-hacking-DKI9T52l.js → reward-hacking-t4lB1yt8.js} +3 -3
- package/dist/{reward-hacking-DKI9T52l.js.map → reward-hacking-t4lB1yt8.js.map} +1 -1
- package/dist/rl.d.ts +11 -42
- package/dist/rl.d.ts.map +1 -1
- package/dist/rl.js +41 -24
- package/dist/rl.js.map +1 -1
- package/dist/rollout/index.d.ts +3 -3
- package/dist/rollout/index.js +7 -7
- package/dist/{rollout-ytVQ7WT8.js → rollout-Dm2tSdiQ.js} +6 -6
- package/dist/{rollout-ytVQ7WT8.js.map → rollout-Dm2tSdiQ.js.map} +1 -1
- package/dist/{rubric-predictive-validity-Cwwyd7ah.js → rubric-predictive-validity-CK8SCOg-.js} +6 -16
- package/dist/rubric-predictive-validity-CK8SCOg-.js.map +1 -0
- package/dist/{rubric-predictive-validity-C7LnNvF2.d.ts → rubric-predictive-validity-CxycqzX5.d.ts} +4 -3
- package/dist/rubric-predictive-validity-CxycqzX5.d.ts.map +1 -0
- package/dist/{run-record-D2lDdSAz.js → run-record-BC0ebuRP.js} +2 -2
- package/dist/{run-record-D2lDdSAz.js.map → run-record-BC0ebuRP.js.map} +1 -1
- package/dist/{run-record-DVV82Gwh.d.ts → run-record-VVy4T9OW.d.ts} +3 -3
- package/dist/{run-record-DVV82Gwh.d.ts.map → run-record-VVy4T9OW.d.ts.map} +1 -1
- package/dist/{schema-BtVldJ3T.d.ts → schema-Bjgdsn73.d.ts} +2 -4
- package/dist/{schema-BtVldJ3T.d.ts.map → schema-Bjgdsn73.d.ts.map} +1 -1
- package/dist/{schema-Cef2cFmb.d.ts → schema-BzWDXhOR.d.ts} +2 -5
- package/dist/schema-BzWDXhOR.d.ts.map +1 -0
- package/dist/{schema-C6DW4ZHR.js → schema-C1aaAxTf.js} +2 -2
- package/dist/schema-C1aaAxTf.js.map +1 -0
- package/dist/{schema-CRhEY1SO.js → schema-k6ZBftVv.js} +2 -8
- package/dist/{schema-CRhEY1SO.js.map → schema-k6ZBftVv.js.map} +1 -1
- package/dist/{semantic-concept-judge-laMCnTLn.js → semantic-concept-judge-BSkKKHeq.js} +14 -38
- package/dist/semantic-concept-judge-BSkKKHeq.js.map +1 -0
- package/dist/{sequential-eprocess-CbUt2htw.js → sequential-eprocess-D1jKoihe.js} +49 -2
- package/dist/sequential-eprocess-D1jKoihe.js.map +1 -0
- package/dist/{sequential-C458DXNf.js → sequential-rYW-Ophm.js} +41 -16
- package/dist/sequential-rYW-Ophm.js.map +1 -0
- package/dist/{series-convergence-BxKEgBwA.d.ts → series-convergence-D9WgpXGi.d.ts} +2 -2
- package/dist/{series-convergence-BxKEgBwA.d.ts.map → series-convergence-D9WgpXGi.d.ts.map} +1 -1
- package/dist/{server-dIWwF3j_.js → server-BtFd4uzB.js} +19 -42
- package/dist/server-BtFd4uzB.js.map +1 -0
- package/dist/{skillopt-optimization-method-DLeUcK-K.js → skillopt-optimization-method-DbaekMcn.js} +794 -8
- package/dist/skillopt-optimization-method-DbaekMcn.js.map +1 -0
- package/dist/{skillopt-optimization-method-BDD_o1xE.d.ts → skillopt-optimization-method-x7TTF23P.d.ts} +20 -7
- package/dist/{skillopt-optimization-method-BDD_o1xE.d.ts.map → skillopt-optimization-method-x7TTF23P.d.ts.map} +1 -1
- package/dist/{statistical-heldout-_woZ9q9j.d.ts → statistical-heldout-Cy3EhjlC.d.ts} +21 -8
- package/dist/statistical-heldout-Cy3EhjlC.d.ts.map +1 -0
- package/dist/{steps-AmkT-GIM.d.ts → steps-CiNVJry_.d.ts} +2 -17
- package/dist/steps-CiNVJry_.d.ts.map +1 -0
- package/dist/{store-CT9YIIve.d.ts → store-B06JdC56.d.ts} +2 -2
- package/dist/{store-CT9YIIve.d.ts.map → store-B06JdC56.d.ts.map} +1 -1
- package/dist/{store-otlp-CDYWW_8N.js → store-otlp-C_Rq5I4D.js} +2 -2
- package/dist/{store-otlp-CDYWW_8N.js.map → store-otlp-C_Rq5I4D.js.map} +1 -1
- package/dist/{store-tool-spans-Br2_IUhm.d.ts → store-tool-spans-DPUG7UUY.d.ts} +6 -6
- package/dist/{store-tool-spans-Br2_IUhm.d.ts.map → store-tool-spans-DPUG7UUY.d.ts.map} +1 -1
- package/dist/{store-tool-spans-CykkbOlv.js → store-tool-spans-Dlh9vkFK.js} +3 -3
- package/dist/{store-tool-spans-CykkbOlv.js.map → store-tool-spans-Dlh9vkFK.js.map} +1 -1
- package/dist/storyboard/index.d.ts +1 -1
- package/dist/{summary-report-Blysd6Z2.js → summary-report-BI5hUtvK.js} +7 -17
- package/dist/summary-report-BI5hUtvK.js.map +1 -0
- package/dist/{summary-report-B__Y5ub3.d.ts → summary-report-CC07PhEL.d.ts} +6 -5
- package/dist/summary-report-CC07PhEL.d.ts.map +1 -0
- package/dist/supervisor-run/index.d.ts +27 -18
- package/dist/supervisor-run/index.d.ts.map +1 -1
- package/dist/supervisor-run/index.js +104 -26
- package/dist/supervisor-run/index.js.map +1 -1
- package/dist/{task-failure-attributes--ZTP3tYO.js → task-failure-attributes-DTl-7-Kw.js} +3 -3
- package/dist/{task-failure-attributes--ZTP3tYO.js.map → task-failure-attributes-DTl-7-Kw.js.map} +1 -1
- package/dist/{tool-groups-B4tqh8jB.d.ts → tool-groups-Ci8i9ErB.d.ts} +3 -3
- package/dist/tool-groups-Ci8i9ErB.d.ts.map +1 -0
- package/dist/{tool-waste-BDdBZG1F.js → tool-waste-BqzmVdJk.js} +4 -4
- package/dist/{tool-waste-BDdBZG1F.js.map → tool-waste-BqzmVdJk.js.map} +1 -1
- package/dist/{tool-waste-DjRDEsuI.d.ts → tool-waste-Dro0gJi3.d.ts} +4 -4
- package/dist/{tool-waste-DjRDEsuI.d.ts.map → tool-waste-Dro0gJi3.d.ts.map} +1 -1
- package/dist/trace-repair/index.d.ts +4 -77
- package/dist/trace-repair/index.d.ts.map +1 -1
- package/dist/trace-repair/index.js +5 -15
- package/dist/trace-repair/index.js.map +1 -1
- package/dist/traces.d.ts +13 -23
- package/dist/traces.d.ts.map +1 -1
- package/dist/traces.js +9 -20
- package/dist/traces.js.map +1 -1
- package/dist/{trajectory-YC15QDYQ.d.ts → trajectory-Bi157Gun.d.ts} +3 -3
- package/dist/{trajectory-YC15QDYQ.d.ts.map → trajectory-Bi157Gun.d.ts.map} +1 -1
- package/dist/trajectory-replay/index.d.ts +5 -5
- package/dist/trajectory-replay/index.js +5 -5
- package/dist/{provenance-oA4-zUqm.d.ts → transient-failure-DKF5Mofa.d.ts} +468 -13
- package/dist/transient-failure-DKF5Mofa.d.ts.map +1 -0
- package/dist/types-B3jzCp0p.js.map +1 -1
- package/dist/{types-CLAwnY-L.d.ts → types-BPb2Kf_C2.d.ts} +3 -3
- package/dist/types-BPb2Kf_C2.d.ts.map +1 -0
- package/dist/types-Bfk0uxRj.d.ts +443 -0
- package/dist/types-Bfk0uxRj.d.ts.map +1 -0
- package/dist/{types-DdFNuyxQ.d.ts → types-D4s7Z6nq.d.ts} +30 -6
- package/dist/types-D4s7Z6nq.d.ts.map +1 -0
- package/dist/{types-B2NsbrNy.d.ts → types-D9ssmxKL.d.ts} +3 -3
- package/dist/{types-B2NsbrNy.d.ts.map → types-D9ssmxKL.d.ts.map} +1 -1
- package/dist/{types-yLK8gXE9.d.ts → types-DeIUdzNd.d.ts} +160 -10
- package/dist/types-DeIUdzNd.d.ts.map +1 -0
- package/dist/{verdict-BndeTAh_.js → verdict-B0xltqu6.js} +2 -2
- package/dist/{verdict-BndeTAh_.js.map → verdict-B0xltqu6.js.map} +1 -1
- package/dist/verdict-cache-CdVVTVmn.js +88 -0
- package/dist/verdict-cache-CdVVTVmn.js.map +1 -0
- package/dist/wire/index.d.ts +21 -111
- package/dist/wire/index.d.ts.map +1 -1
- package/dist/wire/index.js +2 -2
- package/docs/adapters-observability.md +9 -23
- package/docs/building-doctrine.md +3 -3
- package/docs/campaign-proposers.md +54 -15
- package/docs/concepts.md +3 -4
- package/docs/design/statistics-decisions.md +89 -1
- package/docs/eval-surface-map.md +14 -0
- package/docs/experiment.md +19 -2
- package/docs/feedback-trajectories.md +1 -1
- package/docs/multishot-golden-records.md +4 -4
- package/docs/public-api.md +1616 -0
- package/docs/research-report-methodology.md +1 -1
- package/docs/search-history-receipts.md +39 -1
- package/docs/trace-analysis.md +1 -1
- package/docs/trace-repair-admission.md +1 -1
- package/docs/trace-repair-continuation.md +1 -1
- package/docs/verdicts.md +24 -0
- package/docs/wire-protocol.md +1 -1
- package/package.json +6 -2
- package/dist/active-curriculum-C4mk67HP.js.map +0 -1
- package/dist/agent-profile-cell-BkcRDikH.d.ts.map +0 -1
- package/dist/backend-integrity-DOCa_QrR.d.ts.map +0 -1
- package/dist/benchmark-command-CAFwbH0L.js.map +0 -1
- package/dist/campaign-CN_7xJdV.js.map +0 -1
- package/dist/canonical-D-XsTQ6_.js.map +0 -1
- package/dist/client-LIuo-KPv.js.map +0 -1
- package/dist/define-agent-eval-CEQWL9Hy.d.ts.map +0 -1
- package/dist/define-agent-eval-rqNyVhVV.js.map +0 -1
- package/dist/descriptive-B5MwKfbf.js.map +0 -1
- package/dist/dspy-rlm-engine-DHI0WrUU.js.map +0 -1
- package/dist/emitter-CPBAhxum.js.map +0 -1
- package/dist/eval-campaign-C4jmuM-b.js.map +0 -1
- package/dist/exec-y-DCLqK7.js.map +0 -1
- package/dist/exporters-q9iL-2Jf.js.map +0 -1
- package/dist/external-optimizer-subprocess-Dn90UqN2.js.map +0 -1
- package/dist/index-B8Ui1mr1.d.ts.map +0 -1
- package/dist/index-BNPtkBPf.d.ts.map +0 -1
- package/dist/index-C1ravkGA.d.ts +0 -1
- package/dist/index-CTKpu9ry.d.ts.map +0 -1
- package/dist/index-IQccV3Ou.d.ts.map +0 -1
- package/dist/integrity-DysDBWDu.js.map +0 -1
- package/dist/kind-factory-DmAa0h3K.js.map +0 -1
- package/dist/llm-client-Bg32RW0j.js.map +0 -1
- package/dist/llm-judge-CVq33oz1.js.map +0 -1
- package/dist/matrix-DrVnRp4G.d.ts.map +0 -1
- package/dist/pre-registration-DakwTRXk.js +0 -96
- package/dist/pre-registration-DakwTRXk.js.map +0 -1
- package/dist/pre-registration-zFSLEiFU.d.ts.map +0 -1
- package/dist/produced-state-jfk8Du3b.js.map +0 -1
- package/dist/provenance-oA4-zUqm.d.ts.map +0 -1
- package/dist/query-CJ_DX8vl.d.ts.map +0 -1
- package/dist/query-Di7eEQ79.js.map +0 -1
- package/dist/release-confidence-4XrqlpFD.d.ts.map +0 -1
- package/dist/release-confidence-BknrpBnO.js.map +0 -1
- package/dist/researcher-DJnoUE8c.d.ts.map +0 -1
- package/dist/rubric-predictive-validity-C7LnNvF2.d.ts.map +0 -1
- package/dist/rubric-predictive-validity-Cwwyd7ah.js.map +0 -1
- package/dist/schema-C6DW4ZHR.js.map +0 -1
- package/dist/schema-Cef2cFmb.d.ts.map +0 -1
- package/dist/semantic-concept-judge-laMCnTLn.js.map +0 -1
- package/dist/sequential-C458DXNf.js.map +0 -1
- package/dist/sequential-eprocess-CbUt2htw.js.map +0 -1
- package/dist/server-dIWwF3j_.js.map +0 -1
- package/dist/skillopt-optimization-method-DLeUcK-K.js.map +0 -1
- package/dist/statistical-heldout-_woZ9q9j.d.ts.map +0 -1
- package/dist/steps-AmkT-GIM.d.ts.map +0 -1
- package/dist/summary-report-B__Y5ub3.d.ts.map +0 -1
- package/dist/summary-report-Blysd6Z2.js.map +0 -1
- package/dist/tool-groups-B4tqh8jB.d.ts.map +0 -1
- package/dist/types-CLAwnY-L.d.ts.map +0 -1
- package/dist/types-DdFNuyxQ.d.ts.map +0 -1
- package/dist/types-jUBXJ7Iz.d.ts +0 -884
- package/dist/types-jUBXJ7Iz.d.ts.map +0 -1
- package/dist/types-yLK8gXE9.d.ts.map +0 -1
- package/dist/verdict-cache-mZf5FEiY.js +0 -107
- package/dist/verdict-cache-mZf5FEiY.js.map +0 -1
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { b as
|
|
2
|
-
import {
|
|
1
|
+
import { A as unavailable, C as SupervisorRunTreeGap, D as WorkerLogSource, E as WallDistribution, O as isUnavailable, S as SupervisorRunTree, T as Unavailable, _ as SupervisorRunNodeRole, a as OrchestrationMetrics, b as SupervisorRunRollup, c as PerWorkerRow, d as SUPERVISOR_RUN_ROLLUP_SCHEMA, f as SUPERVISOR_RUN_SCHEMA, g as SteerBreakdown, h as SpendMeasurements, i as NO_SOURCE_LIMITS, k as showMeasured, l as RoleSpend, m as SpendMeasurement, n as EconomicsMetrics, o as OutcomeMetrics, p as SourceLimits, r as Measured, s as PatchStats, t as DecisionMetrics, u as RollupCellRow, v as SupervisorRunReader, w as SupervisorRunTreeGapCode, x as SupervisorRunSources, y as SupervisorRunReport } from "../types-DeIUdzNd.js";
|
|
2
|
+
import { y as RolloutSplit } from "../schema-BzWDXhOR.js";
|
|
3
3
|
//#region src/supervisor-run/source-facts.d.ts
|
|
4
4
|
interface Tokens {
|
|
5
5
|
input: number;
|
|
@@ -149,6 +149,13 @@ interface SupervisorTreeFacts {
|
|
|
149
149
|
readonly state: Record<string, unknown> | null;
|
|
150
150
|
readonly startedAt: number | null;
|
|
151
151
|
readonly completedAt: number | null;
|
|
152
|
+
/**
|
|
153
|
+
* First and last stamped tree events (spawned / settled / cancelled /
|
|
154
|
+
* metered) in the journal — the bounds of observed activity. They derive a
|
|
155
|
+
* lower-bound wall for a store that writes no completion stamp.
|
|
156
|
+
*/
|
|
157
|
+
readonly firstEventAt: number | null;
|
|
158
|
+
readonly lastEventAt: number | null;
|
|
152
159
|
}
|
|
153
160
|
/**
|
|
154
161
|
* How a journal's rows were shaped on disk.
|
|
@@ -185,12 +192,6 @@ declare function parsePatch(text: string): PatchStats;
|
|
|
185
192
|
declare function rollupSupervisorRuns(reports: readonly SupervisorRunReport[]): SupervisorRunRollup;
|
|
186
193
|
//#endregion
|
|
187
194
|
//#region src/supervisor-run/claude-code-reader.d.ts
|
|
188
|
-
/** Tool names that spawn a child agent. `Task` is the older name for `Agent`. */
|
|
189
|
-
declare const DEFAULT_SPAWN_TOOLS: readonly ['Agent', 'Task'];
|
|
190
|
-
/** Tool names that deliver a message to an ALREADY-RUNNING child agent. */
|
|
191
|
-
declare const DEFAULT_STEER_TOOLS: readonly ['SendMessage'];
|
|
192
|
-
/** Tool names that stop a running child agent. */
|
|
193
|
-
declare const DEFAULT_CANCEL_TOOLS: readonly ['TaskStop', 'KillAgent'];
|
|
194
195
|
interface ClaudeCodeReaderOptions {
|
|
195
196
|
/** The main session transcript: `~/.claude/projects/<slug>/<sessionId>.jsonl`. */
|
|
196
197
|
readonly transcriptPath: string;
|
|
@@ -302,13 +303,6 @@ interface WriteSupervisorRunOptions extends LoopsReaderOptions {
|
|
|
302
303
|
* every metric says why.
|
|
303
304
|
*/
|
|
304
305
|
declare function writeSupervisorRunReport(runDir: string, opts?: WriteSupervisorRunOptions): Promise<SupervisorRunReport>;
|
|
305
|
-
/**
|
|
306
|
-
* File stem for out-of-tree reports. Built from the run path's identifying
|
|
307
|
-
* segments — candidate tag (the segment under `arm-runs/`), rep, instance, arm
|
|
308
|
-
* — so two runs of the same instance from different candidates/reps never
|
|
309
|
-
* overwrite each other.
|
|
310
|
-
*/
|
|
311
|
-
declare function supervisorReportStem(runDir: string): string;
|
|
312
306
|
/**
|
|
313
307
|
* Best-effort wrapper for a hot path: a reporting failure must never kill a run
|
|
314
308
|
* that already produced real work. Returns null and logs the reason instead.
|
|
@@ -374,18 +368,33 @@ interface SupervisorRolloutOptions {
|
|
|
374
368
|
declare function supervisorRunRolloutLines(src: SupervisorRunSources, opts?: SupervisorRolloutOptions): SupervisorRunTree;
|
|
375
369
|
//#endregion
|
|
376
370
|
//#region src/supervisor-run/runtime-reader.d.ts
|
|
371
|
+
interface RuntimeReaderOptions {
|
|
372
|
+
/**
|
|
373
|
+
* Throw on a missing spawn journal instead of returning absent-shaped
|
|
374
|
+
* sources. The default (false) models a journal-less run dir — a
|
|
375
|
+
* pre-supervise death, a backfilled zombie — as a readable absence.
|
|
376
|
+
*/
|
|
377
|
+
readonly strict?: boolean;
|
|
378
|
+
}
|
|
377
379
|
/**
|
|
378
380
|
* Read one agent-runtime `createFileRunContext(dir)` directory.
|
|
379
381
|
*
|
|
382
|
+
* A run dir without `spawn-journal.jsonl` returns the same absent-shaped
|
|
383
|
+
* sources `readLoopsSupervisorRun` returns for a missing store: `journal` and
|
|
384
|
+
* `workers` null, each with its reason, so every dependent metric reads
|
|
385
|
+
* `unavailable` — never 0 and never a throw. Pass `strict: true` to throw on
|
|
386
|
+
* the missing journal instead. A journal that exists but cannot be parsed
|
|
387
|
+
* always throws: a corrupt journal is a defect, not an absence.
|
|
388
|
+
*
|
|
380
389
|
* The reader translates storage envelopes only. It does not assign research
|
|
381
390
|
* roles, interpret artifacts, or turn process completion into a quality
|
|
382
391
|
* verdict.
|
|
383
392
|
*/
|
|
384
|
-
declare function readRuntimeSupervisorRun(runDir: string): Promise<SupervisorRunSources>;
|
|
393
|
+
declare function readRuntimeSupervisorRun(runDir: string, opts?: RuntimeReaderOptions): Promise<SupervisorRunSources>;
|
|
385
394
|
/** The agent-runtime file-backed layout as a `SupervisorRunReader`. */
|
|
386
|
-
declare function runtimeSupervisorRunReader(runDir: string): SupervisorRunReader;
|
|
395
|
+
declare function runtimeSupervisorRunReader(runDir: string, opts?: RuntimeReaderOptions): SupervisorRunReader;
|
|
387
396
|
/** True when a directory contains Runtime's canonical file-backed journal. */
|
|
388
397
|
declare function isRuntimeSupervisorRunDir(runDir: string): Promise<boolean>;
|
|
389
398
|
//#endregion
|
|
390
|
-
export { type ClaudeCodeReaderOptions, type CloseRow,
|
|
399
|
+
export { type ClaudeCodeReaderOptions, type CloseRow, type DecisionMetrics, type EconomicsMetrics, type LoopsReaderOptions, type Measured, NO_SOURCE_LIMITS, type OrchestrationMetrics, type OutcomeMetrics, type PatchStats, type PerWorkerRow, type RoleSpend, type RollupCellRow, type RuntimeReaderOptions, SUPERVISOR_RUN_INTEGRITY_SCHEMA, SUPERVISOR_RUN_ROLLUP_SCHEMA, SUPERVISOR_RUN_SCHEMA, type SourceLimits, type SpawnRow, type SpendMeasurement, type SpendMeasurements, type SteerAcknowledgementFact, type SteerBreakdown, type SteerRequestFact, type SupervisorJournalDialect, type SupervisorRolloutOptions, type SupervisorRunIntegrityEvidence, type SupervisorRunIntegrityIssue, type SupervisorRunIntegrityIssueCode, type SupervisorRunIntegrityOptions, type SupervisorRunIntegrityReport, type SupervisorRunIntegritySeverity, type SupervisorRunNodeRole, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunRollup, type SupervisorRunSourceOnlyCheckCode, type SupervisorRunSources, type SupervisorRunTree, type SupervisorRunTreeGap, type SupervisorRunTreeGapCode, type SupervisorTreeFacts, type Unavailable, type WallDistribution, type WorkerLogFacts, type WorkerLogSource, type WriteSupervisorRunOptions, analyzeSupervisorRun, analyzeSupervisorRunIntegrity, analyzeSupervisorRunSources, claudeCodeSupervisorRunReader, findSupervisorRunDirIn, findSupervisorRunDirs, isRuntimeSupervisorRunDir, isUnavailable, loopsSupervisorRunReader, parsePatch, parseSupervisorTree, readClaudeCodeSupervisorRun, readLoopsSupervisorRun, readRuntimeSupervisorRun, renderSupervisorRollupMarkdown, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, reportSupervisorRound, rollupSupervisorRuns, runtimeSupervisorRunReader, showMeasured, supervisorRunRolloutLines, unavailable, writeSupervisorRunReport, writeSupervisorRunReportSafe };
|
|
391
400
|
//# sourceMappingURL=index.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/supervisor-run/source-facts.ts","../../src/supervisor-run/analyze.ts","../../src/supervisor-run/claude-code-reader.ts","../../src/supervisor-run/integrity-types.ts","../../src/supervisor-run/integrity.ts","../../src/supervisor-run/loops-reader.ts","../../src/supervisor-run/render.ts","../../src/supervisor-run/rollout-nodes.ts","../../src/supervisor-run/runtime-reader.ts"],"mappings":";;;UAwBU;EACR;EACA;EACA;EACA;;EAEA;;UAGQ;EACR,QAAQ;EACR;;UAgFe;;EAEf;EACA;EACA;EACA;;EAEA,MAAM;;EAEN;;EAEA;EACA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,OAAO;;EAEP;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA;EACA;EACA,eAAe;EACf,uBAAuB;;UAGR;;EAEf;EACA;;UAGe;;EAEf;;EAEA;EACA;;;;;;UAOe;WACN;WACA,iBAAiB;WACjB,iBAAiB;WACjB,uBAAuB;WACvB,uBAAuB;WACvB;IACP;IACA;IACA;IACA;;IAEA;IACA;;IAEA;;WAEO,YAAY,oBAAoB;;WAEhC,wBAAwB;;WAExB;;;;;;;;;WASA;;WAEA;;;;;;WAMA,0BAA0B,SAAS;;;;;;;WAOnC;;WAEA,gBAAgB;;WAEhB,OAAO;WACP;WACA;;;;;;;;;;;;;;;;;;KAmDC;iBAiFI,oBAAoB,KAAK,uBAAuB;;;;;;;
|
|
1
|
+
{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/supervisor-run/source-facts.ts","../../src/supervisor-run/analyze.ts","../../src/supervisor-run/claude-code-reader.ts","../../src/supervisor-run/integrity-types.ts","../../src/supervisor-run/integrity.ts","../../src/supervisor-run/loops-reader.ts","../../src/supervisor-run/render.ts","../../src/supervisor-run/rollout-nodes.ts","../../src/supervisor-run/runtime-reader.ts"],"mappings":";;;UAwBU;EACR;EACA;EACA;EACA;;EAEA;;UAGQ;EACR,QAAQ;EACR;;UAgFe;;EAEf;EACA;EACA;EACA;;EAEA,MAAM;;EAEN;;EAEA;EACA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,OAAO;;EAEP;;UAGe;;EAEf;EACA;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA;EACA;EACA;EACA;EACA,eAAe;EACf,uBAAuB;;UAGR;;EAEf;EACA;;UAGe;;EAEf;;EAEA;EACA;;;;;;UAOe;WACN;WACA,iBAAiB;WACjB,iBAAiB;WACjB,uBAAuB;WACvB,uBAAuB;WACvB;IACP;IACA;IACA;IACA;;IAEA;IACA;;IAEA;;WAEO,YAAY,oBAAoB;;WAEhC,wBAAwB;;WAExB;;;;;;;;;WASA;;WAEA;;;;;;WAMA,0BAA0B,SAAS;;;;;;;WAOnC;;WAEA,gBAAgB;;WAEhB,OAAO;WACP;WACA;;;;;;WAMA;WACA;;;;;;;;;;;;;;;;;;KAmDC;iBAiFI,oBAAoB,KAAK,uBAAuB;;;;;;;iBC5UhD,4BACd,KAAK,sBACL,qBACC;;iBAktBa,WAAW,eAAe;;;;;;iBA+C1B,qBAAqB,kBAAkB,wBAAwB;;;UChvB9D;;WAEN;;;;;WAKA;WACA;WACA;WACA;WACA;WACA;WACA;;;;;;iBAmOW,4BACpB,MAAM,0BACL,QAAQ;;iBAiUK,8BAA8B,MAAM,0BAA0B;;;cCnoBjE;KAED;KAEA;KAKA;UAmCK;WACN;WACA;;UAGM;WACN,MAAM;WACN;WACA,UAAU;WACV;WACA;WACA;WACA,mBAAmB;WACnB;WACA,WAAW,SAAS;;UAGd;WACN,eAAe;WACf;WACA;WACA,MAAM;WACN,iBAAiB;;UAGX;;WAEN;;;;;;;;iBC5BK,8BACd,OAAO,uBAAuB,mBAC9B,UAAS,gCACR;;;;;;;iBCDmB,uBAAuB,aAAa;UAUzC;;WAEN;;WAEA;;WAEA;;WAEA;;;;;;;iBAQW,uBACpB,gBACA,OAAM,qBACL,QAAQ;;iBAmKK,yBACd,gBACA,OAAM,qBACL;;;;;;iBAkCmB,qBACpB,gBAAgB,sBAAsB,sBACtC,OAAM,qBACL,QAAQ;UAgBM,kCAAkC;;WAExC;;WAEA;;;;;;WAMA;;;;;;;;iBASW,yBACpB,gBACA,OAAM,4BACL,QAAQ;;;;;iBA0CW,6BACpB,gBACA,OAAM,4BACL,QAAQ;;;;;;iBAgBW,sBACpB,gBACA,OAAM;EAA8B;IACnC,QAAQ;;;;;;;iBA+BW,sBAAsB,eAAe;;;;;;;iBCzY3C,4BAA4B,GAAG;iBAiC/B,4BAA4B,GAAG;iBA+K/B,+BACd,QAAQ,qBACR;;;UC9Le;;WAEN;;WAEA,QAAQ;;WAER;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;WACA;WACA;WACA;WACA;;WAEA;;;;;;;iBAwBK,0BACd,KAAK,sBACL,OAAM,2BACL;;;UCsUc;;;;;;WAMN;;;;;;;;;;;;;;;;iBAsDW,yBACpB,gBACA,OAAM,uBACL,QAAQ;;iBAkEK,2BACd,gBACA,OAAM,uBACL;;iBAKmB,0BAA0B,iBAAiB"}
|
|
@@ -1,5 +1,6 @@
|
|
|
1
|
-
import {
|
|
2
|
-
import {
|
|
1
|
+
import { o as summarizeNumberSeries } from "../descriptive-jDOuI6mz.js";
|
|
2
|
+
import { c as parseClaudeEntries, n as findOpencodeSessionsByDirectory, r as openOpencodeDb, t as DEFAULT_OPENCODE_DB, u as transcriptFromEntries } from "../opencode-sqlite-eK6HW6dr.js";
|
|
3
|
+
import { a as parseJsonl, c as SUPERVISOR_RUN_INTEGRITY_SCHEMA, i as parseJson, n as supervisorRunRolloutLines, o as parseSupervisorTree, r as asRecord, s as workerSourceKey, t as analyzeSupervisorRunIntegrity } from "../integrity-CyWSSoQS.js";
|
|
3
4
|
import { basename, dirname, join } from "node:path";
|
|
4
5
|
import { appendFile, mkdir, readFile, readdir, stat, writeFile } from "node:fs/promises";
|
|
5
6
|
//#region src/supervisor-run/types.ts
|
|
@@ -44,7 +45,7 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
44
45
|
gaps.push(`${what}: ${reason}`);
|
|
45
46
|
return unavailable(reason);
|
|
46
47
|
};
|
|
47
|
-
const journalMissing = src.supRunDir === null ? "no supervisor run dir under <ws>/.agent/supervisor (or legacy <ws>/.loops/supervisor)" : "journal.jsonl absent";
|
|
48
|
+
const journalMissing = src.journalMissingReason ?? (src.supRunDir === null ? "no supervisor run dir under <ws>/.agent/supervisor (or legacy <ws>/.loops/supervisor)" : "journal.jsonl absent");
|
|
48
49
|
const haveJournal = src.journal !== null;
|
|
49
50
|
const tree = parseSupervisorTree(src);
|
|
50
51
|
const state = tree.state;
|
|
@@ -68,7 +69,22 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
68
69
|
const matches = spawnsForSource(worker);
|
|
69
70
|
return matches.length === 1 ? matches[0] ?? null : null;
|
|
70
71
|
};
|
|
71
|
-
const
|
|
72
|
+
const wallSpanStart = startedAt ?? tree.firstEventAt;
|
|
73
|
+
const wallSpanEnd = tree.lastEventAt;
|
|
74
|
+
let supervisorWallMs;
|
|
75
|
+
let supervisorWallSource;
|
|
76
|
+
if (startedAt !== null && completedAt !== null && completedAt >= startedAt) {
|
|
77
|
+
supervisorWallMs = completedAt - startedAt;
|
|
78
|
+
supervisorWallSource = "stamps";
|
|
79
|
+
} else if (completedAt === null && wallSpanStart !== null && wallSpanEnd !== null && wallSpanEnd >= wallSpanStart) {
|
|
80
|
+
supervisorWallMs = wallSpanEnd - wallSpanStart;
|
|
81
|
+
supervisorWallSource = "journal-span";
|
|
82
|
+
} else {
|
|
83
|
+
const reason = !haveJournal ? journalMissing : "no parseable start/complete timestamps in state.json or journal";
|
|
84
|
+
supervisorWallMs = gap("supervisorWallMs", reason);
|
|
85
|
+
supervisorWallSource = unavailable(reason);
|
|
86
|
+
}
|
|
87
|
+
const wallEndAt = completedAt ?? (supervisorWallSource === "journal-span" ? wallSpanEnd : null);
|
|
72
88
|
const steerRows = [];
|
|
73
89
|
let steerQueuedTotal = 0;
|
|
74
90
|
let steerDeliveredTotal = 0;
|
|
@@ -135,8 +151,8 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
135
151
|
if (live > maxConcurrency) maxConcurrency = live;
|
|
136
152
|
prev = step.at;
|
|
137
153
|
}
|
|
138
|
-
if (prev !== null &&
|
|
139
|
-
const span =
|
|
154
|
+
if (prev !== null && wallEndAt !== null && wallEndAt >= prev) {
|
|
155
|
+
const span = wallEndAt - prev;
|
|
140
156
|
if (live === 0) idleMs += span;
|
|
141
157
|
sumWorkerWallMs += span * live;
|
|
142
158
|
}
|
|
@@ -203,6 +219,7 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
203
219
|
delegationDepth: haveJournal ? delegationDepth : unavailable(journalMissing),
|
|
204
220
|
timeToFirstSpawnMs: startedAt !== null && firstWorkerSpawnAt !== null ? firstWorkerSpawnAt - startedAt : haveJournal ? unavailable("no worker spawn timestamps") : unavailable(journalMissing),
|
|
205
221
|
supervisorWallMs,
|
|
222
|
+
supervisorWallSource,
|
|
206
223
|
idleMs: isUnavailable(supervisorWallMs) ? unavailable(supervisorWallMs.unavailable) : idleMs,
|
|
207
224
|
idlePct: isUnavailable(supervisorWallMs) || supervisorWallMs === 0 ? isUnavailable(supervisorWallMs) ? unavailable(supervisorWallMs.unavailable) : unavailable("supervisor wall is 0ms") : round(idleMs / supervisorWallMs * 100, 1),
|
|
208
225
|
workerUtilization: isUnavailable(supervisorWallMs) || supervisorWallMs === 0 ? isUnavailable(supervisorWallMs) ? unavailable(supervisorWallMs.unavailable) : unavailable("supervisor wall is 0ms") : round(sumWorkerWallMs / supervisorWallMs, 3)
|
|
@@ -267,8 +284,24 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
267
284
|
const workerOut = workerTokenLimit !== null ? unavailable(workerTokenLimit) : sq !== null ? journalWorkerOut + sq.output : haveJournal ? journalWorkerOut : unavailable(harnessGapReason);
|
|
268
285
|
const stateResult = asRecord(state?.result);
|
|
269
286
|
const stateUsd = typeof stateResult.spentUsd === "number" ? stateResult.spentUsd : null;
|
|
287
|
+
const resultSpentTotal = asRecord(result?.spentTotal);
|
|
288
|
+
const resultCloseUsd = typeof resultSpentTotal.usd === "number" && Number.isFinite(resultSpentTotal.usd) ? resultSpentTotal.usd : null;
|
|
289
|
+
const closeUsd = stateUsd ?? resultCloseUsd;
|
|
270
290
|
const usdLimit = src.limits.spendUsd;
|
|
271
291
|
const totalUsd = usdLimit !== null ? gap("totalUsd", usdLimit) : stateUsd !== null ? round(stateUsd, 6) : haveJournal ? round(tree.brain.usd + journalWorkerUsd, 6) : gap("totalUsd", journalMissing);
|
|
292
|
+
const journalSpendRecords = tree.brain.meteredCount + rootChildCloses.filter((close) => close.hasSpend).length;
|
|
293
|
+
const journalDerivedAvailable = usdLimit === null && haveJournal;
|
|
294
|
+
const closeRecordAvailable = usdLimit === null && closeUsd !== null;
|
|
295
|
+
const spend = {
|
|
296
|
+
journalDerived: {
|
|
297
|
+
usd: journalDerivedAvailable ? round(tree.brain.usd + journalWorkerUsd, 6) : unavailable(usdLimit ?? journalMissing),
|
|
298
|
+
records: journalDerivedAvailable ? journalSpendRecords : 0
|
|
299
|
+
},
|
|
300
|
+
closeRecord: {
|
|
301
|
+
usd: closeRecordAvailable ? round(closeUsd, 6) : unavailable(usdLimit ?? "no close record: neither state.json result.spentUsd nor result.json spentTotal.usd is present"),
|
|
302
|
+
records: closeRecordAvailable ? 1 : 0
|
|
303
|
+
}
|
|
304
|
+
};
|
|
272
305
|
const perWorker = (src.workers ?? []).map((w) => {
|
|
273
306
|
const f = tree.workerLogs.get(workerSourceKey(w));
|
|
274
307
|
const matchingSpawns = spawnsForSource(w);
|
|
@@ -297,7 +330,7 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
297
330
|
score: close?.score ?? f?.score ?? null
|
|
298
331
|
};
|
|
299
332
|
});
|
|
300
|
-
const
|
|
333
|
+
const wallDistribution = summarizeNumberSeries(perWorker.map((w) => w.wallMs).filter((w) => w !== null));
|
|
301
334
|
const brainCalls = parseJsonl(src.brainLog);
|
|
302
335
|
const managerTokenLimit = src.limits.managerTokens;
|
|
303
336
|
const economics = {
|
|
@@ -318,17 +351,11 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
|
|
|
318
351
|
usd: usdLimit !== null ? unavailable(usdLimit) : haveJournal ? round(journalWorkerUsd, 6) : unavailable(journalMissing),
|
|
319
352
|
source: workerTokenLimit !== null ? workerTokenLimit : sq !== null ? `journal settled spend + ${sq.store} sessions (n=${sq.sessions})` : `journal settled spend only — ${src.harnessMissingReason ?? "harness session store unavailable"}`
|
|
320
353
|
},
|
|
354
|
+
spend,
|
|
321
355
|
totalUsd,
|
|
322
356
|
totalUsdSource: usdLimit !== null ? usdLimit : stateUsd !== null ? `state.json result.spentUsd${rootChildCloses.length > 0 && journalWorkerUsd === 0 ? " — brain-priced only; worker CLI inference is unpriced (see worker token counts)" : ""}` : haveJournal ? "journal metered + settled usd" : journalMissing,
|
|
323
357
|
costPerAcceptedPatchUsd: isUnavailable(totalUsd) ? unavailable(totalUsd.unavailable) : isUnavailable(decision.accepted) ? unavailable(decision.accepted.unavailable) : decision.accepted === 0 ? unavailable("no accepted worker patch (cost has no denominator)") : round(totalUsd / decision.accepted, 6),
|
|
324
|
-
workerWallMsDistribution:
|
|
325
|
-
n: walls.length,
|
|
326
|
-
min: walls[0],
|
|
327
|
-
p50: quantile(walls, .5),
|
|
328
|
-
p90: quantile(walls, .9),
|
|
329
|
-
max: walls[walls.length - 1],
|
|
330
|
-
sum: walls.reduce((a, b) => a + b, 0)
|
|
331
|
-
},
|
|
358
|
+
workerWallMsDistribution: wallDistribution === null ? unavailable(src.workers === null ? workersGapReason : "no worker start/finish pairs captured") : wallDistribution,
|
|
332
359
|
perWorker: src.workers === null ? unavailable(workersGapReason) : perWorker
|
|
333
360
|
};
|
|
334
361
|
const patchStats = src.patch === null ? gap("patch", src.limits.deliverables ?? "delivered patch file absent") : parsePatch(src.patch);
|
|
@@ -386,10 +413,6 @@ function round(v, digits) {
|
|
|
386
413
|
const f = 10 ** digits;
|
|
387
414
|
return Math.round(v * f) / f;
|
|
388
415
|
}
|
|
389
|
-
function quantile(sorted, q) {
|
|
390
|
-
if (sorted.length === 0) return 0;
|
|
391
|
-
return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(q * sorted.length) - 1))];
|
|
392
|
-
}
|
|
393
416
|
/** Unified-diff stats. Counts `+++ b/<path>` targets, body +/- lines, and test-file touches. */
|
|
394
417
|
function parsePatch(text) {
|
|
395
418
|
const files = /* @__PURE__ */ new Set();
|
|
@@ -435,6 +458,8 @@ function rollupSupervisorRuns(reports) {
|
|
|
435
458
|
const spawnVals = known(reports.map((r) => r.orchestration.workersSpawned));
|
|
436
459
|
const acceptVals = known(reports.map((r) => r.decision.accepted));
|
|
437
460
|
const usdVals = known(reports.map((r) => r.economics.totalUsd));
|
|
461
|
+
const journalSpendVals = known(reports.map((r) => r.economics.spend.journalDerived.usd));
|
|
462
|
+
const closeSpendVals = known(reports.map((r) => r.economics.spend.closeRecord.usd));
|
|
438
463
|
const resolvedVals = known(reports.map((r) => r.outcome.judgeResolved));
|
|
439
464
|
const sum = (xs) => xs.reduce((a, b) => a + b, 0);
|
|
440
465
|
const mean = (xs) => xs.length === 0 ? unavailable("no cell reported this metric") : round(sum(xs) / xs.length, 3);
|
|
@@ -461,6 +486,16 @@ function rollupSupervisorRuns(reports) {
|
|
|
461
486
|
workersSpawnedTotal: spawnVals.length === 0 ? unavailable("no cell reported spawns") : sum(spawnVals),
|
|
462
487
|
acceptedTotal: acceptVals.length === 0 ? unavailable("no cell reported acceptance") : sum(acceptVals),
|
|
463
488
|
usdTotal: usdVals.length === 0 ? unavailable("no cell reported spend") : round(sum(usdVals), 6),
|
|
489
|
+
spendUsd: {
|
|
490
|
+
journalDerived: {
|
|
491
|
+
value: journalSpendVals.length === 0 ? unavailable("no cell measured journal-derived spend") : round(sum(journalSpendVals), 6),
|
|
492
|
+
runs: journalSpendVals.length
|
|
493
|
+
},
|
|
494
|
+
closeRecord: {
|
|
495
|
+
value: closeSpendVals.length === 0 ? unavailable("no cell carried a close record") : round(sum(closeSpendVals), 6),
|
|
496
|
+
runs: closeSpendVals.length
|
|
497
|
+
}
|
|
498
|
+
},
|
|
464
499
|
resolvedCount: resolvedVals.length === 0 ? unavailable("no cell reported a judge verdict") : resolvedVals.filter((v) => v === true).length,
|
|
465
500
|
perCell
|
|
466
501
|
};
|
|
@@ -961,7 +996,7 @@ function renderSupervisorRunHeadline(r) {
|
|
|
961
996
|
`RUN-REPORT ${r.instanceId ?? "?"} [${r.arm ?? "?"}]`,
|
|
962
997
|
` steers=${steerNote}`,
|
|
963
998
|
` waves=${showMeasured(o.waves)} sizes=${isUnavailable(o.waveSizes) ? `unavailable — ${o.waveSizes.unavailable}` : `[${o.waveSizes.join(",")}]`} workers=${showMeasured(o.workersSpawned)} settled=${showMeasured(o.workersSettled)} cancelled=${showMeasured(o.workersCancelled)}`,
|
|
964
|
-
` concurrency max=${showMeasured(o.maxConcurrency)} utilization=${showMeasured(o.workerUtilization)} idle=${fmtMs(o.idleMs)} (${showMeasured(o.idlePct)}%) wall=${fmtMs(o.supervisorWallMs)}`,
|
|
999
|
+
` concurrency max=${showMeasured(o.maxConcurrency)} utilization=${showMeasured(o.workerUtilization)} idle=${fmtMs(o.idleMs)} (${showMeasured(o.idlePct)}%) wall=${fmtMs(o.supervisorWallMs)}${o.supervisorWallSource === "journal-span" ? " (journal-span lower bound)" : ""}`,
|
|
965
1000
|
` respawns=${showMeasured(o.respawns)} evidence→respawn=${showMeasured(r.decision.observeThenRespawn)} blind-respawn=${showMeasured(r.decision.respawnWithoutEvidence)} depth=${showMeasured(o.delegationDepth)}`,
|
|
966
1001
|
` accepted=${showMeasured(r.decision.accepted)} rejected=${showMeasured(r.decision.rejected)} empty-pass=${showMeasured(r.decision.emptyPass)}`,
|
|
967
1002
|
` brain=$${showMeasured(r.economics.brain.usd)} total=$${showMeasured(r.economics.totalUsd)} judge.resolved=${showMeasured(r.outcome.judgeResolved)} score=${showMeasured(r.outcome.judgeScore)} verify=${showMeasured(r.outcome.verifyPass)}`,
|
|
@@ -1008,7 +1043,7 @@ function renderSupervisorRunMarkdown(r) {
|
|
|
1008
1043
|
out.push(`| Repeated labels | ${isUnavailable(o.repeatedLabels) ? showMeasured(o.repeatedLabels) : o.repeatedLabels.length === 0 ? "none" : o.repeatedLabels.join(", ")} |`);
|
|
1009
1044
|
out.push(`| Delegation depth | ${showMeasured(o.delegationDepth)} |`);
|
|
1010
1045
|
out.push(`| Time to first spawn | ${fmtMs(o.timeToFirstSpawnMs)} |`);
|
|
1011
|
-
out.push(`| Supervisor wall | ${fmtMs(o.supervisorWallMs)} |`);
|
|
1046
|
+
out.push(`| Supervisor wall | ${fmtMs(o.supervisorWallMs)}${isUnavailable(o.supervisorWallSource) ? "" : ` (source: ${o.supervisorWallSource}${o.supervisorWallSource === "journal-span" ? ", lower bound" : ""})`} |`);
|
|
1012
1047
|
out.push(`| Idle (zero live workers) | ${fmtMs(o.idleMs)} (${showMeasured(o.idlePct)}%) |`);
|
|
1013
1048
|
out.push(`| Worker utilization (Σ worker wall ÷ supervisor wall) | ${showMeasured(o.workerUtilization)} |`);
|
|
1014
1049
|
out.push("");
|
|
@@ -1044,6 +1079,7 @@ function renderSupervisorRunMarkdown(r) {
|
|
|
1044
1079
|
if (!isUnavailable(e.brainTruncations) && e.brainTruncations > 0) out.push(`- **BRAIN OUTPUT TRUNCATED: ${e.brainTruncations} completion(s) hit \`finish_reason: "length"\`** — the supervisor acted on a half-written plan. Its output ceiling is too low; see \`brain.jsonl\` for the per-call \`req_max_tokens\`.`);
|
|
1045
1080
|
else out.push(`- Brain completions truncated (finish_reason=length): ${showMeasured(e.brainTruncations)}`);
|
|
1046
1081
|
out.push(`- Total USD: ${showMeasured(e.totalUsd)} (source: ${e.totalUsdSource})`);
|
|
1082
|
+
out.push(`- Spend measured two ways: journal-derived $${showMeasured(e.spend.journalDerived.usd)} over ${e.spend.journalDerived.records} journal record(s) (execution accounting) · close-record $${showMeasured(e.spend.closeRecord.usd)} over ${e.spend.closeRecord.records} close record(s) (billing-shaped). Divergence is a signal, not an error.`);
|
|
1047
1083
|
out.push(`- Cost per accepted patch: ${showMeasured(e.costPerAcceptedPatchUsd)}`);
|
|
1048
1084
|
if (isUnavailable(e.workerWallMsDistribution)) out.push(`- Worker wall distribution: unavailable — ${e.workerWallMsDistribution.unavailable}`);
|
|
1049
1085
|
else {
|
|
@@ -1101,6 +1137,7 @@ function renderSupervisorRollupMarkdown(rollup, title = "Round rollup") {
|
|
|
1101
1137
|
out.push(`- Idle share (mean): ${showMeasured(rollup.idlePctMean)}%`);
|
|
1102
1138
|
out.push(`- Workers spawned: ${showMeasured(rollup.workersSpawnedTotal)} · accepted: ${showMeasured(rollup.acceptedTotal)}`);
|
|
1103
1139
|
out.push(`- Spend: $${showMeasured(rollup.usdTotal)} · judged resolved: ${showMeasured(rollup.resolvedCount)}/${rollup.cells}`);
|
|
1140
|
+
out.push(`- Spend measured two ways: journal-derived $${showMeasured(rollup.spendUsd.journalDerived.value)} over ${rollup.spendUsd.journalDerived.runs}/${rollup.cells} runs (execution accounting) · close-record $${showMeasured(rollup.spendUsd.closeRecord.value)} over ${rollup.spendUsd.closeRecord.runs}/${rollup.cells} runs (billing-shaped)`);
|
|
1104
1141
|
out.push("");
|
|
1105
1142
|
out.push("| Instance | Arm | Steers | Waves | Utilization | Idle % | Resolved | USD |");
|
|
1106
1143
|
out.push("|---|---|---:|---:|---:|---:|---|---:|");
|
|
@@ -1330,15 +1367,56 @@ function runtimeState(root, startedAt, result, trajectory, resultPath, trajector
|
|
|
1330
1367
|
});
|
|
1331
1368
|
}
|
|
1332
1369
|
/**
|
|
1370
|
+
* Sources for a run dir whose spawn journal does not exist. Mirrors the
|
|
1371
|
+
* absent shape `readLoopsSupervisorRun` returns for a missing store: every
|
|
1372
|
+
* journal-dependent metric downstream reads `unavailable`, never 0.
|
|
1373
|
+
*/
|
|
1374
|
+
function absentRuntimeSupervisorRun(runDir, resultText) {
|
|
1375
|
+
const reason = `no Runtime spawn journal (${JOURNAL_FILE}) under ${runDir}`;
|
|
1376
|
+
return {
|
|
1377
|
+
runRef: runDir,
|
|
1378
|
+
instanceId: null,
|
|
1379
|
+
arm: null,
|
|
1380
|
+
supRunDir: null,
|
|
1381
|
+
journal: null,
|
|
1382
|
+
journalMissingReason: reason,
|
|
1383
|
+
brainLog: null,
|
|
1384
|
+
brainLogMissingReason: "Runtime FileRunContext records spend but not model completion finish reasons",
|
|
1385
|
+
state: null,
|
|
1386
|
+
progress: null,
|
|
1387
|
+
workers: null,
|
|
1388
|
+
workersMissingReason: reason,
|
|
1389
|
+
result: resultText,
|
|
1390
|
+
judge: null,
|
|
1391
|
+
judgeSource: null,
|
|
1392
|
+
patch: null,
|
|
1393
|
+
driverLog: null,
|
|
1394
|
+
harnessWorkerTokens: null,
|
|
1395
|
+
harnessMissingReason: "Runtime FileRunContext has no external worker-token join",
|
|
1396
|
+
limits: NO_SOURCE_LIMITS,
|
|
1397
|
+
rootTranscriptRef: null,
|
|
1398
|
+
traceCommand: "unavailable — Runtime FileRunContext records no provider-session trace identity"
|
|
1399
|
+
};
|
|
1400
|
+
}
|
|
1401
|
+
/**
|
|
1333
1402
|
* Read one agent-runtime `createFileRunContext(dir)` directory.
|
|
1334
1403
|
*
|
|
1404
|
+
* A run dir without `spawn-journal.jsonl` returns the same absent-shaped
|
|
1405
|
+
* sources `readLoopsSupervisorRun` returns for a missing store: `journal` and
|
|
1406
|
+
* `workers` null, each with its reason, so every dependent metric reads
|
|
1407
|
+
* `unavailable` — never 0 and never a throw. Pass `strict: true` to throw on
|
|
1408
|
+
* the missing journal instead. A journal that exists but cannot be parsed
|
|
1409
|
+
* always throws: a corrupt journal is a defect, not an absence.
|
|
1410
|
+
*
|
|
1335
1411
|
* The reader translates storage envelopes only. It does not assign research
|
|
1336
1412
|
* roles, interpret artifacts, or turn process completion into a quality
|
|
1337
1413
|
* verdict.
|
|
1338
1414
|
*/
|
|
1339
|
-
async function readRuntimeSupervisorRun(runDir) {
|
|
1415
|
+
async function readRuntimeSupervisorRun(runDir, opts = {}) {
|
|
1340
1416
|
const journalPath = join(runDir, JOURNAL_FILE);
|
|
1341
|
-
const
|
|
1417
|
+
const rawJournal = opts.strict === true ? await readFile(journalPath, "utf8") : await readMaybe$1(journalPath);
|
|
1418
|
+
if (rawJournal === null) return absentRuntimeSupervisorRun(runDir, await readMaybe$1(join(runDir, RESULT_FILE)));
|
|
1419
|
+
const normalized = parseEnvelopeJournal(rawJournal, journalPath);
|
|
1342
1420
|
const resultText = await readMaybe$1(join(runDir, RESULT_FILE));
|
|
1343
1421
|
const trajectoryText = await readMaybe$1(join(runDir, TRAJECTORY_FILE));
|
|
1344
1422
|
const result = parseOptionalRecord(resultText, join(runDir, RESULT_FILE));
|
|
@@ -1381,10 +1459,10 @@ async function readRuntimeSupervisorRun(runDir) {
|
|
|
1381
1459
|
};
|
|
1382
1460
|
}
|
|
1383
1461
|
/** The agent-runtime file-backed layout as a `SupervisorRunReader`. */
|
|
1384
|
-
function runtimeSupervisorRunReader(runDir) {
|
|
1462
|
+
function runtimeSupervisorRunReader(runDir, opts = {}) {
|
|
1385
1463
|
return {
|
|
1386
1464
|
runRef: runDir,
|
|
1387
|
-
read: () => readRuntimeSupervisorRun(runDir)
|
|
1465
|
+
read: () => readRuntimeSupervisorRun(runDir, opts)
|
|
1388
1466
|
};
|
|
1389
1467
|
}
|
|
1390
1468
|
/** True when a directory contains Runtime's canonical file-backed journal. */
|
|
@@ -1685,6 +1763,6 @@ async function findSupervisorRunDirs(root) {
|
|
|
1685
1763
|
return found.sort();
|
|
1686
1764
|
}
|
|
1687
1765
|
//#endregion
|
|
1688
|
-
export {
|
|
1766
|
+
export { NO_SOURCE_LIMITS, SUPERVISOR_RUN_INTEGRITY_SCHEMA, SUPERVISOR_RUN_ROLLUP_SCHEMA, SUPERVISOR_RUN_SCHEMA, analyzeSupervisorRun, analyzeSupervisorRunIntegrity, analyzeSupervisorRunSources, claudeCodeSupervisorRunReader, findSupervisorRunDirIn, findSupervisorRunDirs, isRuntimeSupervisorRunDir, isUnavailable, loopsSupervisorRunReader, parsePatch, parseSupervisorTree, readClaudeCodeSupervisorRun, readLoopsSupervisorRun, readRuntimeSupervisorRun, renderSupervisorRollupMarkdown, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, reportSupervisorRound, rollupSupervisorRuns, runtimeSupervisorRunReader, showMeasured, supervisorRunRolloutLines, unavailable, writeSupervisorRunReport, writeSupervisorRunReportSafe };
|
|
1689
1767
|
|
|
1690
1768
|
//# sourceMappingURL=index.js.map
|