@tangle-network/agent-eval 0.172.0 → 0.173.0
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 +49 -0
- package/README.md +17 -2
- package/dist/adapters/http.d.ts +2 -2
- package/dist/analyst/index.d.ts +13 -13
- package/dist/analyst/index.js +7 -6
- package/dist/analyst/index.js.map +1 -1
- package/dist/{attestation-CJBGmMVh.d.ts → attestation-c1QvaBdX.d.ts} +2 -2
- package/dist/{attestation-CJBGmMVh.d.ts.map → attestation-c1QvaBdX.d.ts.map} +1 -1
- package/dist/{backend-integrity-e79K3UPD.d.ts → backend-integrity-CeuTgqsd.d.ts} +3 -4
- package/dist/backend-integrity-CeuTgqsd.d.ts.map +1 -0
- package/dist/{benchmark-h-h4bfqj.d.ts → benchmark-BjLGkfnN.d.ts} +3 -3
- package/dist/{benchmark-h-h4bfqj.d.ts.map → benchmark-BjLGkfnN.d.ts.map} +1 -1
- package/dist/{benchmark-command-DoFcisuM.js → benchmark-command-9S20PRel.js} +9 -9
- package/dist/{benchmark-command-DoFcisuM.js.map → benchmark-command-9S20PRel.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +5 -5
- package/dist/benchmarks/index.js +3 -3
- package/dist/{bounded-process-CVOC_D3H.js → bounded-process-BBZob7vl.js} +64 -10
- package/dist/bounded-process-BBZob7vl.js.map +1 -0
- package/dist/builder-eval/index.d.ts +3 -3
- package/dist/builder-eval/index.js +2 -2
- package/dist/campaign/index.d.ts +9 -9
- package/dist/campaign/index.js +6 -6
- package/dist/{campaign-Dp35pBbS.js → campaign-pxS0wmo4.js} +8 -8
- package/dist/{campaign-Dp35pBbS.js.map → campaign-pxS0wmo4.js.map} +1 -1
- package/dist/chat-client-Db4bqYfA.js +115 -0
- package/dist/chat-client-Db4bqYfA.js.map +1 -0
- package/dist/{chat-json-call-5Jxna-aV.js → chat-json-call-C26igCih.js} +16 -6
- package/dist/chat-json-call-C26igCih.js.map +1 -0
- package/dist/cli.js +31 -17
- package/dist/cli.js.map +1 -1
- package/dist/{client-BvwNkIRN.js → client-BlLY6o2w.js} +2 -2
- package/dist/{client-BvwNkIRN.js.map → client-BlLY6o2w.js.map} +1 -1
- package/dist/{client-Df7wdslk.d.ts → client-DlqdbM7n.d.ts} +4 -4
- package/dist/{client-Df7wdslk.d.ts.map → client-DlqdbM7n.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +13 -13
- package/dist/contract/index.js +202 -10
- package/dist/contract/index.js.map +1 -1
- package/dist/{counterfactual-Bee5_BIn.d.ts → counterfactual-CLgrwhkY.d.ts} +4 -4
- package/dist/{counterfactual-Bee5_BIn.d.ts.map → counterfactual-CLgrwhkY.d.ts.map} +1 -1
- package/dist/{chat-client-DI79OPye.js → default-registry-B0bKikCb.js} +3 -39
- package/dist/default-registry-B0bKikCb.js.map +1 -0
- package/dist/{default-registry-XxedTLwu.d.ts → default-registry-BKwc8bN5.d.ts} +6 -6
- package/dist/{default-registry-XxedTLwu.d.ts.map → default-registry-BKwc8bN5.d.ts.map} +1 -1
- package/dist/{define-agent-eval-0wW7gFhr.d.ts → define-agent-eval-CY6qdlGV.d.ts} +6 -6
- package/dist/{define-agent-eval-0wW7gFhr.d.ts.map → define-agent-eval-CY6qdlGV.d.ts.map} +1 -1
- package/dist/{define-agent-eval-jS8xj_Q_.js → define-agent-eval-D_i_s69h.js} +7 -7
- package/dist/{define-agent-eval-jS8xj_Q_.js.map → define-agent-eval-D_i_s69h.js.map} +1 -1
- package/dist/{dspy-rlm-engine-CS3qcCEk.js → dspy-rlm-engine-D5byiHn9.js} +5 -6
- package/dist/dspy-rlm-engine-D5byiHn9.js.map +1 -0
- package/dist/{emitter-Bvnu0VzL.d.ts → emitter-Cs0egaFd.d.ts} +3 -3
- package/dist/{emitter-Bvnu0VzL.d.ts.map → emitter-Cs0egaFd.d.ts.map} +1 -1
- package/dist/{engine-BfRay1qD.d.ts → engine-DhFir3Ys.d.ts} +23 -8
- package/dist/{engine-BfRay1qD.d.ts.map → engine-DhFir3Ys.d.ts.map} +1 -1
- package/dist/{eval-campaign-JDTeE6Pl.js → eval-campaign-BeAjdhzC.js} +2 -2
- package/dist/{eval-campaign-JDTeE6Pl.js.map → eval-campaign-BeAjdhzC.js.map} +1 -1
- package/dist/{exact-types-BEecmnWm.d.ts → exact-types-BKOEILRP.d.ts} +2 -2
- package/dist/{exact-types-BEecmnWm.d.ts.map → exact-types-BKOEILRP.d.ts.map} +1 -1
- package/dist/experiment/index.d.ts +4 -4
- package/dist/{external-optimizer-contracts-szBJ_1vh.d.ts → external-optimizer-contracts-CQCpyrIL.d.ts} +2 -2
- package/dist/{external-optimizer-contracts-szBJ_1vh.d.ts.map → external-optimizer-contracts-CQCpyrIL.d.ts.map} +1 -1
- package/dist/{external-optimizer-process-CQxylYeG.js → external-optimizer-process-Cq_Pg15r.js} +2 -2
- package/dist/{external-optimizer-process-CQxylYeG.js.map → external-optimizer-process-Cq_Pg15r.js.map} +1 -1
- package/dist/{external-optimizer-subprocess-Cex8Da2i.js → external-optimizer-subprocess-DgNebftP.js} +2 -2
- package/dist/{external-optimizer-subprocess-Cex8Da2i.js.map → external-optimizer-subprocess-DgNebftP.js.map} +1 -1
- package/dist/failure-cluster-OldNRoAt.d.ts +154 -0
- package/dist/failure-cluster-OldNRoAt.d.ts.map +1 -0
- package/dist/{feedback-trajectory-DIqpCyF0.d.ts → feedback-trajectory-CMnv_uYs.d.ts} +6 -6
- package/dist/{feedback-trajectory-DIqpCyF0.d.ts.map → feedback-trajectory-CMnv_uYs.d.ts.map} +1 -1
- package/dist/{heldout-gate-JgNRDZwZ.d.ts → heldout-gate-Df5hsqmm.d.ts} +7 -7
- package/dist/{heldout-gate-JgNRDZwZ.d.ts.map → heldout-gate-Df5hsqmm.d.ts.map} +1 -1
- package/dist/hosted/index.d.ts +2 -2
- package/dist/hosted/index.js +1 -1
- package/dist/{index-DnglhM0A.d.ts → index-BQqOjerE.d.ts} +94 -13
- package/dist/index-BQqOjerE.d.ts.map +1 -0
- package/dist/{index-_vPrVMRX.d.ts → index-CFDffsKz.d.ts} +11 -11
- package/dist/{index-_vPrVMRX.d.ts.map → index-CFDffsKz.d.ts.map} +1 -1
- package/dist/{index-DDAPhUJJ.d.ts → index-D0Db5X-4.d.ts} +45 -9
- package/dist/index-D0Db5X-4.d.ts.map +1 -0
- package/dist/{index-DMoxLG8P.d.ts → index-e7LXeRVa.d.ts} +3 -3
- package/dist/{index-DMoxLG8P.d.ts.map → index-e7LXeRVa.d.ts.map} +1 -1
- package/dist/index.d.ts +172 -46
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +64 -41
- package/dist/index.js.map +1 -1
- package/dist/{insight-report-08F022xN.d.ts → insight-report-DETqPc_A.d.ts} +4 -4
- package/dist/{insight-report-08F022xN.d.ts.map → insight-report-DETqPc_A.d.ts.map} +1 -1
- package/dist/{integrity-B_EDELom.d.ts → integrity-BKTcA-HP.d.ts} +3 -3
- package/dist/{integrity-B_EDELom.d.ts.map → integrity-BKTcA-HP.d.ts.map} +1 -1
- package/dist/{kind-factory-DMeEoMQZ.js → kind-factory-gP6lDySe.js} +164 -149
- package/dist/kind-factory-gP6lDySe.js.map +1 -0
- package/dist/{llm-client-BFMRpmqb.js → llm-client-CxQtdtd6.js} +12 -5
- package/dist/llm-client-CxQtdtd6.js.map +1 -0
- package/dist/{llm-judge-aQHIk5_-.js → llm-judge-B2YxbAJb.js} +81 -7
- package/dist/{llm-judge-aQHIk5_-.js.map → llm-judge-B2YxbAJb.js.map} +1 -1
- package/dist/{matrix-Ch8JO1pG.d.ts → matrix-DGu8KhSs.d.ts} +2 -2
- package/dist/{matrix-Ch8JO1pG.d.ts.map → matrix-DGu8KhSs.d.ts.map} +1 -1
- package/dist/meta-eval/index.d.ts +100 -5
- package/dist/meta-eval/index.d.ts.map +1 -1
- package/dist/meta-eval/index.js +200 -2
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/{mint-DjfDUMHr.js → mint-vWOdD8Ae.js} +2 -2
- package/dist/{mint-DjfDUMHr.js.map → mint-vWOdD8Ae.js.map} +1 -1
- package/dist/multishot/golden/index.d.ts +1 -1
- package/dist/multishot/index.d.ts +2 -2
- package/dist/openapi.json +1 -1
- package/dist/pipelines/index.d.ts +5 -5
- package/dist/pipelines/index.js +3 -3
- package/dist/{pre-registration-DHz6P_6f.d.ts → pre-registration-BoI4ucR3.d.ts} +2 -2
- package/dist/{pre-registration-DHz6P_6f.d.ts.map → pre-registration-BoI4ucR3.d.ts.map} +1 -1
- package/dist/{produced-state-CxmbFxFd.js → produced-state-7VYDwtkk.js} +3 -3
- package/dist/{produced-state-CxmbFxFd.js.map → produced-state-7VYDwtkk.js.map} +1 -1
- package/dist/{promotion-policy-BBBcz5_3.d.ts → promotion-policy-CvMda3kU.d.ts} +2 -2
- package/dist/{promotion-policy-BBBcz5_3.d.ts.map → promotion-policy-CvMda3kU.d.ts.map} +1 -1
- package/dist/{provenance-Dp-vvyrU.d.ts → provenance-CRY67X50.d.ts} +39 -165
- package/dist/provenance-CRY67X50.d.ts.map +1 -0
- package/dist/{query-BPGMVlbM.js → query-D1nLIKt7.js} +2 -2
- package/dist/{query-BPGMVlbM.js.map → query-D1nLIKt7.js.map} +1 -1
- package/dist/{query-Na5gEIGd.d.ts → query-D6W6MaGx.d.ts} +3 -3
- package/dist/{query-Na5gEIGd.d.ts.map → query-D6W6MaGx.d.ts.map} +1 -1
- package/dist/{registry-xEb_xfns.d.ts → registry-7pOUBrtX.d.ts} +4 -4
- package/dist/{registry-xEb_xfns.d.ts.map → registry-7pOUBrtX.d.ts.map} +1 -1
- package/dist/{release-confidence-D6lQw_o7.d.ts → release-confidence-BAcNYOf1.d.ts} +4 -4
- package/dist/{release-confidence-D6lQw_o7.d.ts.map → release-confidence-BAcNYOf1.d.ts.map} +1 -1
- package/dist/{release-confidence-CzUHc4z4.js → release-confidence-BsGEg_xg.js} +3 -3
- package/dist/{release-confidence-CzUHc4z4.js.map → release-confidence-BsGEg_xg.js.map} +1 -1
- package/dist/reporting.d.ts +3 -3
- package/dist/reporting.js +1 -1
- package/dist/{researcher-CMUTQXD7.d.ts → researcher-jsW1X94L.d.ts} +7 -8
- package/dist/researcher-jsW1X94L.d.ts.map +1 -0
- package/dist/{reward-hacking-SkxYgT0x.js → reward-hacking-CKW4teig.js} +2 -2
- package/dist/{reward-hacking-SkxYgT0x.js.map → reward-hacking-CKW4teig.js.map} +1 -1
- package/dist/{reward-hacking-CgPRUesA.d.ts → reward-hacking-ZXEi9VCq.d.ts} +2 -2
- package/dist/{reward-hacking-CgPRUesA.d.ts.map → reward-hacking-ZXEi9VCq.d.ts.map} +1 -1
- package/dist/rl.d.ts +7 -7
- package/dist/rl.js +4 -4
- package/dist/rollout/index.d.ts +1 -1
- package/dist/rollout/index.js +2 -2
- package/dist/{rollout-Crypdx8s.js → rollout-C-znbbYg.js} +2 -2
- package/dist/{rollout-Crypdx8s.js.map → rollout-C-znbbYg.js.map} +1 -1
- package/dist/{rubric-predictive-validity-DluJLCKQ.d.ts → rubric-predictive-validity-Dl1dvKCv.d.ts} +2 -2
- package/dist/{rubric-predictive-validity-DluJLCKQ.d.ts.map → rubric-predictive-validity-Dl1dvKCv.d.ts.map} +1 -1
- package/dist/{run-record-DQjRcYwA.d.ts → run-record-DTv1MdjK.d.ts} +2 -2
- package/dist/{run-record-DQjRcYwA.d.ts.map → run-record-DTv1MdjK.d.ts.map} +1 -1
- package/dist/{run-record-DLORoL7t.js → run-record-ZIsR9Fif.js} +2 -2
- package/dist/{run-record-DLORoL7t.js.map → run-record-ZIsR9Fif.js.map} +1 -1
- package/dist/{schema-DID1Cqct.d.ts → schema-CR5cpjQ3.d.ts} +57 -2
- package/dist/{schema-DID1Cqct.d.ts.map → schema-CR5cpjQ3.d.ts.map} +1 -1
- package/dist/{schema-CdIX2aHu.js → schema-CSf6qWgZ.js} +35 -1
- package/dist/{schema-CdIX2aHu.js.map → schema-CSf6qWgZ.js.map} +1 -1
- package/dist/semantic-concept-judge-Ct3QU7t5.js +780 -0
- package/dist/semantic-concept-judge-Ct3QU7t5.js.map +1 -0
- package/dist/{series-convergence-D9WgpXGi.d.ts → series-convergence-DeG33RpC.d.ts} +2 -2
- package/dist/{series-convergence-D9WgpXGi.d.ts.map → series-convergence-DeG33RpC.d.ts.map} +1 -1
- package/dist/{server-CCEnywOR.js → server-BR6onwZB.js} +2 -2
- package/dist/{server-CCEnywOR.js.map → server-BR6onwZB.js.map} +1 -1
- package/dist/{skillopt-optimization-method-LHi02MzH.js → skillopt-optimization-method-BzdphODy.js} +6 -6
- package/dist/{skillopt-optimization-method-LHi02MzH.js.map → skillopt-optimization-method-BzdphODy.js.map} +1 -1
- package/dist/{statistical-heldout-Yldkntvy.d.ts → statistical-heldout-DTyB_6-1.d.ts} +3 -3
- package/dist/{statistical-heldout-Yldkntvy.d.ts.map → statistical-heldout-DTyB_6-1.d.ts.map} +1 -1
- package/dist/{store-Cq9oOrI1.d.ts → store-BErPvYBr.d.ts} +2 -2
- package/dist/{store-Cq9oOrI1.d.ts.map → store-BErPvYBr.d.ts.map} +1 -1
- package/dist/{store-otlp-CHjBvWQY.js → store-otlp-Dow0pk_5.js} +2 -2
- package/dist/{store-otlp-CHjBvWQY.js.map → store-otlp-Dow0pk_5.js.map} +1 -1
- package/dist/{store-tool-spans-BvdUbeOB.d.ts → store-tool-spans-CCZNsihA.d.ts} +8 -8
- package/dist/{store-tool-spans-BvdUbeOB.d.ts.map → store-tool-spans-CCZNsihA.d.ts.map} +1 -1
- package/dist/{store-tool-spans-B9o6tU8f.js → store-tool-spans-CeNj_m2L.js} +3 -3
- package/dist/{store-tool-spans-B9o6tU8f.js.map → store-tool-spans-CeNj_m2L.js.map} +1 -1
- package/dist/storyboard/index.d.ts +1 -1
- package/dist/{summary-report-DRstQNBX.d.ts → summary-report-gMrbYawB.d.ts} +3 -3
- package/dist/{summary-report-DRstQNBX.d.ts.map → summary-report-gMrbYawB.d.ts.map} +1 -1
- package/dist/{task-failure-attributes-CBGtLS_H.js → task-failure-attributes-CZjZeBsY.js} +3 -3
- package/dist/{task-failure-attributes-CBGtLS_H.js.map → task-failure-attributes-CZjZeBsY.js.map} +1 -1
- package/dist/{tool-groups-DjwlMBvW.d.ts → tool-groups-Cp4Xdzrp.d.ts} +3 -3
- package/dist/tool-groups-Cp4Xdzrp.d.ts.map +1 -0
- package/dist/{tool-waste-CwGHzBzX.js → tool-waste-B9tdWV6g.js} +253 -6
- package/dist/tool-waste-B9tdWV6g.js.map +1 -0
- package/dist/{tool-waste-BrmLKxMw.d.ts → tool-waste-D23I0zWm.d.ts} +4 -4
- package/dist/{tool-waste-BrmLKxMw.d.ts.map → tool-waste-D23I0zWm.d.ts.map} +1 -1
- package/dist/trace-repair/index.d.ts +2 -2
- package/dist/traces.d.ts +10 -10
- package/dist/traces.js +7 -7
- package/dist/{trajectory-r1bQqvBQ.d.ts → trajectory-D7qrNvaN.d.ts} +3 -3
- package/dist/{trajectory-r1bQqvBQ.d.ts.map → trajectory-D7qrNvaN.d.ts.map} +1 -1
- package/dist/trajectory-replay/index.d.ts +3 -3
- package/dist/{types-CCZ34qmV.d.ts → types-BDV4PiMR.d.ts} +3 -3
- package/dist/{types-CCZ34qmV.d.ts.map → types-BDV4PiMR.d.ts.map} +1 -1
- package/dist/{types-nokrtr7M.d.ts → types-Ba5UQyVD.d.ts} +4 -4
- package/dist/{types-nokrtr7M.d.ts.map → types-Ba5UQyVD.d.ts.map} +1 -1
- package/dist/{types-DMoNFDWi.d.ts → types-DN2WdT5S.d.ts} +3 -3
- package/dist/{types-DMoNFDWi.d.ts.map → types-DN2WdT5S.d.ts.map} +1 -1
- package/dist/types-gvRsyJLh.d.ts +831 -0
- package/dist/types-gvRsyJLh.d.ts.map +1 -0
- package/dist/wire/index.d.ts +3 -3
- package/dist/wire/index.js +1 -1
- package/docs/plants.md +69 -1
- package/docs/public-api.md +6 -4
- package/docs/trace-analysis.md +55 -8
- package/package.json +1 -1
- package/dist/backend-integrity-e79K3UPD.d.ts.map +0 -1
- package/dist/bounded-process-CVOC_D3H.js.map +0 -1
- package/dist/chat-client-DI79OPye.js.map +0 -1
- package/dist/chat-json-call-5Jxna-aV.js.map +0 -1
- package/dist/dspy-rlm-engine-CS3qcCEk.js.map +0 -1
- package/dist/failure-cluster-6YSvsKlp.d.ts +0 -58
- package/dist/failure-cluster-6YSvsKlp.d.ts.map +0 -1
- package/dist/index-DDAPhUJJ.d.ts.map +0 -1
- package/dist/index-DnglhM0A.d.ts.map +0 -1
- package/dist/kind-factory-DMeEoMQZ.js.map +0 -1
- package/dist/llm-client-BFMRpmqb.js.map +0 -1
- package/dist/provenance-Dp-vvyrU.d.ts.map +0 -1
- package/dist/raw-provider-sink-BU29Sh8h.d.ts +0 -134
- package/dist/raw-provider-sink-BU29Sh8h.d.ts.map +0 -1
- package/dist/researcher-CMUTQXD7.d.ts.map +0 -1
- package/dist/semantic-concept-judge-I36eejJx.js +0 -382
- package/dist/semantic-concept-judge-I36eejJx.js.map +0 -1
- package/dist/tool-groups-DjwlMBvW.d.ts.map +0 -1
- package/dist/tool-waste-CwGHzBzX.js.map +0 -1
- package/dist/types-Bfk0uxRj.d.ts +0 -443
- package/dist/types-Bfk0uxRj.d.ts.map +0 -1
|
@@ -0,0 +1,154 @@
|
|
|
1
|
+
import { C as TraceEvent, _ as Span, f as Run, o as FailureClass } from "./schema-CR5cpjQ3.js";
|
|
2
|
+
import { s as TraceStore } from "./store-BErPvYBr.js";
|
|
3
|
+
//#region src/failure-taxonomy.d.ts
|
|
4
|
+
/**
|
|
5
|
+
* Who a failure is charged to. This is the axis a comparison reads before it
|
|
6
|
+
* counts a loss.
|
|
7
|
+
*
|
|
8
|
+
* `machine` — the run never reached a model. The bridge refused, its lane was
|
|
9
|
+
* full, the box never provisioned, the credential was absent. An
|
|
10
|
+
* agent cannot route around this, so a run that recorded nothing
|
|
11
|
+
* behind only machine failures is not evidence about the agent.
|
|
12
|
+
* `provider` — the model was reached and the turn did not complete: an empty
|
|
13
|
+
* completion, a torn stream, a killed process. Every arm faces
|
|
14
|
+
* the same provider on the same day, so this stays charged.
|
|
15
|
+
* `agent` — the agent's own reasoning, tool use, budgeting, or scheduling
|
|
16
|
+
* produced the failure. Charged.
|
|
17
|
+
* `unknown` — the record does not say. Never voids a run: an unread reason is
|
|
18
|
+
* not a proven outage.
|
|
19
|
+
* `none` — no failure to charge.
|
|
20
|
+
*/
|
|
21
|
+
type FailureBlame = 'none' | 'machine' | 'provider' | 'agent' | 'unknown';
|
|
22
|
+
/**
|
|
23
|
+
* Blame for every canonical class. The `Record` is exhaustive on purpose: a
|
|
24
|
+
* class added to `FAILURE_CLASSES` without a blame here fails to compile, so
|
|
25
|
+
* the axis cannot silently acquire a hole that reads as `undefined` — and a
|
|
26
|
+
* missing blame would void or charge a run by accident.
|
|
27
|
+
*
|
|
28
|
+
* The assignment rule, applied member by member: a class is `machine` only
|
|
29
|
+
* when the run demonstrably never reached a model. An environment that refuses
|
|
30
|
+
* one specific action the agent chose (a denied write, an out-of-scope call,
|
|
31
|
+
* an approval hold) stays `agent`, because the agent selected the action and
|
|
32
|
+
* every arm faces the same policy.
|
|
33
|
+
*/
|
|
34
|
+
declare const FAILURE_BLAME: Readonly<Record<FailureClass, FailureBlame>>;
|
|
35
|
+
/**
|
|
36
|
+
* The blames that mean the run got no usable turn out of the machine or the
|
|
37
|
+
* provider.
|
|
38
|
+
*
|
|
39
|
+
* This answers a different question from `machine` alone, and the two must not
|
|
40
|
+
* be confused. "Is this run evidence about the agent?" is answered by `machine`
|
|
41
|
+
* and by nothing else — a `provider` death stays charged, because every arm
|
|
42
|
+
* faces the same provider on the same day. "Is this seat producing usable turns
|
|
43
|
+
* at all?" is answered by this set, because a seat that only ever returns empty
|
|
44
|
+
* completions and torn streams is as dead as one whose bridge refuses, however
|
|
45
|
+
* the individual runs are charged.
|
|
46
|
+
*
|
|
47
|
+
* So a reader counting an outage streak counts these; a reader deciding whether
|
|
48
|
+
* to void a run reads `machine`. Nothing an operator did lands here: `cancelled`
|
|
49
|
+
* is blamed `unknown` for exactly that reason.
|
|
50
|
+
*/
|
|
51
|
+
declare const INFRA_FAILURE_BLAMES: ReadonlySet<FailureBlame>;
|
|
52
|
+
/** Who the named class is charged to. */
|
|
53
|
+
declare function failureBlame(failureClass: FailureClass): FailureBlame;
|
|
54
|
+
interface FailureContext {
|
|
55
|
+
run: Run;
|
|
56
|
+
spans: Span[];
|
|
57
|
+
events: TraceEvent[];
|
|
58
|
+
}
|
|
59
|
+
interface FailureClassification {
|
|
60
|
+
failureClass: FailureClass;
|
|
61
|
+
/** Who the failure is charged to. Derived from the class unless a rule overrode it. */
|
|
62
|
+
blame: FailureBlame;
|
|
63
|
+
reason: string;
|
|
64
|
+
triggerSpanId?: string;
|
|
65
|
+
triggerEventId?: string;
|
|
66
|
+
}
|
|
67
|
+
/** Ordered rules — first match wins. */
|
|
68
|
+
interface FailureRule {
|
|
69
|
+
id: string;
|
|
70
|
+
match: (ctx: FailureContext) => {
|
|
71
|
+
failureClass: FailureClass;
|
|
72
|
+
reason: string;
|
|
73
|
+
/** Overrides the class's canonical blame. State why at the call site. */
|
|
74
|
+
blame?: FailureBlame;
|
|
75
|
+
triggerSpanId?: string;
|
|
76
|
+
triggerEventId?: string;
|
|
77
|
+
} | null;
|
|
78
|
+
}
|
|
79
|
+
/**
|
|
80
|
+
* One ordered rule over a failure reason string.
|
|
81
|
+
*
|
|
82
|
+
* Order is load-bearing, because one reason string carries two signals: a seat
|
|
83
|
+
* killed by SIGKILL while the bridge refuses connections is a bridge failure,
|
|
84
|
+
* not a process failure, so the bridge patterns are tested first.
|
|
85
|
+
*/
|
|
86
|
+
interface FailureReasonRule {
|
|
87
|
+
/** The class the first matching pattern names. */
|
|
88
|
+
failureClass: FailureClass;
|
|
89
|
+
pattern: RegExp;
|
|
90
|
+
/** Overrides the class's canonical blame. State why at the call site. */
|
|
91
|
+
blame?: FailureBlame;
|
|
92
|
+
}
|
|
93
|
+
/**
|
|
94
|
+
* The default ordered reason rules, measured against the transport death
|
|
95
|
+
* strings this stack records — runtime settle reasons, bridge errors, sandbox
|
|
96
|
+
* provisioning errors, and provider stream errors.
|
|
97
|
+
*
|
|
98
|
+
* A consumer with a different transport passes its own set to
|
|
99
|
+
* `classifyFailureReason`; nothing here is a gate.
|
|
100
|
+
*/
|
|
101
|
+
declare const DEFAULT_FAILURE_REASON_RULES: readonly FailureReasonRule[];
|
|
102
|
+
/**
|
|
103
|
+
* Name the failure one reason string describes.
|
|
104
|
+
*
|
|
105
|
+
* This is the entry point for a run that died before it produced a span: a
|
|
106
|
+
* terminal record, a settle row, a spawn journal line. An unmatched reason is
|
|
107
|
+
* never dropped — it returns `unknown` and keeps the text, so a new failure
|
|
108
|
+
* mode shows up as a named gap instead of a silent bucket. An absent reason
|
|
109
|
+
* returns `unreported`, which is a different fact: the record itself is
|
|
110
|
+
* incomplete.
|
|
111
|
+
*/
|
|
112
|
+
declare function classifyFailureReason(reason: string | null | undefined, rules?: readonly FailureReasonRule[]): FailureClassification;
|
|
113
|
+
/**
|
|
114
|
+
* Classify the failure mode of a run using an ordered rule list.
|
|
115
|
+
*
|
|
116
|
+
* Structural rules run first, in their declared order. When none matches, the
|
|
117
|
+
* reason strings the run did record are classified with
|
|
118
|
+
* `classifyFailureReason`, so a run whose only evidence is a transport error
|
|
119
|
+
* on an error event is named instead of landing in `unknown`. Only a run that
|
|
120
|
+
* would already have returned `unknown` can change class this way.
|
|
121
|
+
*/
|
|
122
|
+
declare function classifyFailure(ctx: FailureContext, rules?: FailureRule[], reasonRules?: readonly FailureReasonRule[]): FailureClassification;
|
|
123
|
+
//#endregion
|
|
124
|
+
//#region src/pipelines/failure-cluster.d.ts
|
|
125
|
+
interface FailureCluster {
|
|
126
|
+
failureClass: FailureClass;
|
|
127
|
+
/** Tool name when the trigger was a tool span, else undefined. */
|
|
128
|
+
toolName?: string;
|
|
129
|
+
/** First 16 chars of argHash — clusters similar args. */
|
|
130
|
+
argPrefix?: string;
|
|
131
|
+
/**
|
|
132
|
+
* Source dimension when the trigger was a judge span (e.g. `'format'`,
|
|
133
|
+
* `'safety'`, `'correctness'`). Lets cross-template aggregators
|
|
134
|
+
* group failures by the dimension that fired without overloading
|
|
135
|
+
* `argPrefix`. Optional — clusters without this field deserialize cleanly.
|
|
136
|
+
*/
|
|
137
|
+
dimension?: string;
|
|
138
|
+
runCount: number;
|
|
139
|
+
scenarioIds: string[];
|
|
140
|
+
exampleError?: string;
|
|
141
|
+
exampleRunId: string;
|
|
142
|
+
}
|
|
143
|
+
interface FailureClusterReport {
|
|
144
|
+
clusters: FailureCluster[];
|
|
145
|
+
totalFailures: number;
|
|
146
|
+
totalRuns: number;
|
|
147
|
+
}
|
|
148
|
+
declare function failureClusterView(store: TraceStore, options?: {
|
|
149
|
+
rules?: FailureRule[];
|
|
150
|
+
minClusterSize?: number;
|
|
151
|
+
}): Promise<FailureClusterReport>;
|
|
152
|
+
//#endregion
|
|
153
|
+
export { FAILURE_BLAME as a, FailureContext as c, INFRA_FAILURE_BLAMES as d, classifyFailure as f, DEFAULT_FAILURE_REASON_RULES as i, FailureReasonRule as l, failureBlame as m, FailureClusterReport as n, FailureBlame as o, classifyFailureReason as p, failureClusterView as r, FailureClassification as s, FailureCluster as t, FailureRule as u };
|
|
154
|
+
//# sourceMappingURL=failure-cluster-OldNRoAt.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"failure-cluster-OldNRoAt.d.ts","names":[],"sources":["../src/failure-taxonomy.ts","../src/pipelines/failure-cluster.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;KA6CY;;;;;;;;;;;;;cAcC,eAAe,SAAS,OAAO,cAAc;;;;;;;;;;;;;;;;;cAmG7C,sBAAsB,YAAY;;iBAM/B,aAAa,cAAc,eAAe;UAIzC;EACf,KAAK;EACL,OAAO;EACP,QAAQ;;UAGO;EACf,cAAc;;EAEd,OAAO;EACP;EACA;EACA;;;UAIe;EACf;EACA,QAAQ,KAAK;IACX,cAAc;IACd;;IAEA,QAAQ;IACR;IACA;;;;;;;;;;UAWa;;EAEf,cAAc;EACd,SAAS;;EAET,QAAQ;;;;;;;;;;cAWG,uCAAuC;;;;;;;;;;;iBA6DpC,sBACd,mCACA,iBAAgB,sBACf;;;;;;;;;;iBAkba,gBACd,KAAK,gBACL,QAAO,eACP,uBAAsB,sBACrB;;;UCpsBc;EACf,cAAc;;EAEd;;EAEA;;;;;;;EAOA;EACA;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;iBAGoB,mBACpB,OAAO,YACP;EAAW,QAAQ;EAAe;IACjC,QAAQ"}
|
|
@@ -1,9 +1,9 @@
|
|
|
1
|
-
import { o as FailureClass } from "./schema-
|
|
2
|
-
import { i as AnalystFinding, l as AnalystRunResult, p as EvidenceRef } from "./types-
|
|
1
|
+
import { o as FailureClass } from "./schema-CR5cpjQ3.js";
|
|
2
|
+
import { i as AnalystFinding, l as AnalystRunResult, p as EvidenceRef } from "./types-DN2WdT5S.js";
|
|
3
3
|
import { i as DatasetSplit, r as DatasetScenario } from "./dataset-DQqhOCPt.js";
|
|
4
|
-
import { s as TraceStore } from "./store-
|
|
5
|
-
import "./exact-types-
|
|
6
|
-
import { i as TraceEmitter } from "./emitter-
|
|
4
|
+
import { s as TraceStore } from "./store-BErPvYBr.js";
|
|
5
|
+
import "./exact-types-BKOEILRP.js";
|
|
6
|
+
import { i as TraceEmitter } from "./emitter-Cs0egaFd.js";
|
|
7
7
|
//#region src/control-runtime.d.ts
|
|
8
8
|
type ControlSeverity = 'info' | 'warning' | 'error' | 'critical';
|
|
9
9
|
type ControlActionFailureMode = 'continue' | 'stop';
|
|
@@ -394,4 +394,4 @@ declare function controlRunToFeedbackTrajectory<TState, TAction, TActionResult>(
|
|
|
394
394
|
}): FeedbackTrajectory;
|
|
395
395
|
//#endregion
|
|
396
396
|
export { AnalystFindingDigest as A, ControlRunResult as B, createFeedbackTrajectory as C, renderPreferenceMemoryMarkdown as D, feedbackTrajectoryToOptimizerRow as E, ControlActionOutcome as F, StopDecision as G, ControlRuntimeError as H, ControlBudget as I, subjectiveEval as J, objectiveEval as K, ControlContext as L, AnalystRunDigest as M, analystFindingDigest as N, summarizePreferenceMemory as O, analystRunDigest as P, ControlDecision as R, controlRunToFeedbackTrajectory as S, feedbackTrajectoriesToOptimizerRows as T, ControlSeverity as U, ControlRuntimeConfig as V, ControlStep as W, PreferenceMemoryEntry as _, FeedbackLabel as a, analystRunToReviewRequests as b, FeedbackOptimizerRow as c, FeedbackTask as d, FeedbackTrajectory as f, InMemoryFeedbackTrajectoryStore as g, FileSystemFeedbackTrajectoryStore as h, FeedbackAttempt as i, AnalystReviewDecision as j, withAssignedFeedbackSplit as k, FeedbackOutcome as l, FeedbackTrajectoryStore as m, AnalystReviewRequest as n, FeedbackLabelKind as o, FeedbackTrajectoryFilter as p, runAgentControlLoop as q, FeedbackArtifactType as r, FeedbackLabelSource as s, AnalystFeedbackTrajectoryOptions as t, FeedbackSplitPolicy as u, ProposedSideEffect as v, feedbackTrajectoriesToDatasetScenarios as w, assignFeedbackSplit as x, analystRunToFeedbackTrajectory as y, ControlEvalResult as z };
|
|
397
|
-
//# sourceMappingURL=feedback-trajectory-
|
|
397
|
+
//# sourceMappingURL=feedback-trajectory-CMnv_uYs.d.ts.map
|
package/dist/{feedback-trajectory-DIqpCyF0.d.ts.map → feedback-trajectory-CMnv_uYs.d.ts.map}
RENAMED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"feedback-trajectory-
|
|
1
|
+
{"version":3,"file":"feedback-trajectory-CMnv_uYs.d.ts","names":[],"sources":["../src/control-runtime.ts","../src/feedback-trajectory-review.ts","../src/feedback-trajectory.ts"],"mappings":";;;;;;;KAkBY;KACA;UAEK;;EAEf;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA,WAAW;;UAGI;EACf;EACA;EACA;;UAGe,oBAAoB,QAAQ;;;;;EAK3C;;;;;EAKA;;EAEA;;EAEA,oBAAoB,OAAO;;EAE3B,qBAAqB,QAAQ;;UAGd,eACf,QACA,SACA,eACA,cAAc,oBAAoB;EAElC;EACA,OAAO;EACP,OAAO;EACP,SAAS,YAAY,QAAQ,SAAS,eAAe;EACrD,QAAQ;EACR;EACA;EACA;EACA;EACA,aAAa;EACb,UAAU;;KAGA,gBAAgB;EAEtB;EACA,QAAQ;EACR;;EAGA;EACA;EACA;EACA;;EAEA,eAAe;;UAGJ;EACf;EACA;EACA;EACA;EACA,eAAe;;UAGA,qBAAqB;EACpC;EACA,SAAS;EACT;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe,YACf,QACA,SACA,eACA,cAAc,oBAAoB;EAElC;EACA,UAAU,gBAAgB;EAC1B,aAAa;EACb,YAAY;EACZ,aAAa;EACb,YAAY;EACZ,gBAAgB,qBAAqB;EACrC;EACA;;UAGe,iBACf,QACA,SACA,eACA,cAAc,oBAAoB;EAElC;EACA;EACA;EACA;EACA;EACA,OAAO,YAAY,QAAQ,SAAS,eAAe;EACnD,YAAY;EACZ,YAAY;EACZ;EACA;;EAEA;EACA,eAAe;EACf,eAAe;EACf;;UAGe,qBACf,QACA,SACA,eACA,cAAc,oBAAoB;EAElC;EACA,SAAS,QAAQ;EACjB,SAAS;;EAET,gBAAgB;;;;;EAKhB,oBAAoB;IAClB,QAAQ;IACR,QAAQ;IACR,OAAO;IACP,OAAO;IACP,SAAS,YAAY,QAAQ,SAAS,eAAe;;;EAIvD,UAAU;IACR,SAAS,YAAY,QAAQ,SAAS,eAAe;IACrD,aAAa;QACT,QAAQ,UAAU;;EAGxB,WAAW;IACT;IACA,OAAO;IACP,SAAS,YAAY,QAAQ,SAAS,eAAe;IACrD,aAAa;QACT,QAAQ,WAAW;;EAGzB,SACE,KAAK,eAAe,QAAQ,SAAS,eAAe,WACjD,QAAQ,gBAAgB,YAAY,gBAAgB;;EAGzD,MACE,QAAQ,SACR,KAAK,eAAe,QAAQ,SAAS,eAAe,WACjD,QAAQ,iBAAiB;;EAG9B,cACE,KAAK,eAAe,QAAQ,SAAS,eAAe,WACjD,QAAQ,gBAAgB;;EAG7B,UAAU,MAAM,YAAY,QAAQ,SAAS,eAAe,WAAW;;EAGvE,eAAe,oBAAoB,QAAQ;;EAG3C,QAAQ;EACR;EACA;EACA;;iBAQoB,oBACpB,QACA,SACA,eACA,cAAc,oBAAoB,mBAElC,QAAQ,qBAAqB,QAAQ,SAAS,eAAe,SAC5D,QAAQ,iBAAiB,QAAQ,SAAS,eAAe;iBAinB5C,cAAc,OAAO,KAAK,kCAAkC;iBAI5D,eAAe,OAAO,KAAK,kCAAkC;;;KCx1BjE,sBAAsB,QAAQ;KAC9B;KACA;UAEK;EACf;EACA;EACA,WAAW;;UAGH;EACR,WAAW;EACX,QAAQ;EACR;EACA;EACA;EACA;;KAGU,yBACP;EACC;EACA,eAAe;EACf;EACA;MAED;EACC;EACA,cAAc;EACd;EACA;;;iBA2BU,qBAAqB,SAAS,iBAAiB;;iBAK/C,iBAAiB,KAAK,mBAAmB;;;KChD7C;KAWA;KAEA;KAYA;UAEK;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA,WAAW;;UAGI;EACf;EACA,QAAQ;EACR,MAAM;EACN;EACA;EACA,WAAW;EACX;EACA,WAAW;;UAGI;EACf;EACA;EACA,cAAc;EACd;EACA;EACA,iBAAiB;EACjB,QAAQ;EACR,WAAW;EACX;EACA,WAAW;;UAGI;EACf;EACA;EACA,UAAU;EACV;EACA;EACA;EACA,WAAW;;UAGI;EACf;EACA;EACA;EACA,MAAM;EACN,UAAU;EACV,QAAQ;EACR,UAAU;EACV,QAAQ;EACR,OAAO;EACP;EACA;EACA,WAAW;;UAGI;EACf,KAAK,YAAY,qBAAqB;EACtC,IAAI,aAAa,QAAQ;EACzB,KAAK,SAAS,2BAA2B,QAAQ;EACjD,cAAc,YAAY,SAAS,kBAAkB,QAAQ;EAC7D,YAAY,YAAY,OAAO,eAAe,qBAAqB,QAAQ;;UAG5D;EACf;EACA;EACA,QAAQ;EACR;;UAGe;EACf;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,YAAY;EACZ;EACA,WAAW;;UAoBI;EACf,MAAM;EACN,kBAAkB;EAClB,0BAA0B;EAC1B,2BAA2B;EAC3B,UAAU;EACV;EACA;EACA;EACA,QAAQ;EACR,OAAO;EACP;EACA;IACE;IACA;IACA;;EAEF,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA,UAAU;EACV;EACA;;cAUW,2CAA2C;mBACrC;EAEX,KAAK,YAAY,qBAAqB;EAItC,IAAI,aAAa,QAAQ;EAKzB,KAAK,SAAQ,2BAAgC,QAAQ;EAMrD,cAAc,YAAY,SAAS,kBAAkB,QAAQ;EAa7D,YACJ,YACA,OAAO,eACP,qBACC,QAAQ;;cAsBA,6CAA6C;mBACvC;mBACA;UACT;EAER,YAAY;IAAW;;EAIjB,KAAK,YAAY,qBAAqB;EAMtC,IAAI,aAAa,QAAQ;EAKzB,KAAK,SAAQ,2BAAgC,QAAQ;EAKrD,cAAc,YAAY,SAAS,kBAAkB,QAAQ;EAO7D,YACJ,YACA,OAAO,eACP,qBACC,QAAQ;UAOG;UAWA;;iBA8BA,yBAAyB;EACvC;EACA;EACA;EACA,MAAM;EACN,WAAW;EACX,SAAS;EACT,UAAU;EACV,QAAQ;EACR,OAAO;EACP;EACA,WAAW;IACT;;iBAqBY,+BACd,KAAK,kBACL,SAAS,mCACR;;iBAsFa,2BACd,KAAK,kBACL;EAAW;IACV;iBA4Da,oBACd,YAAY,KAAK,iEACjB,SAAQ,sBACP;iBAca,0BACd,YAAY,oBACZ,SAAS,sBACR;iBAuBa,uCACd,cAAc,uBACb;iBAIa,iCACd,YAAY,qBACX;iBA6Ca,oCACd,cAAc,uBACb;iBAoDa,0BACd,cAAc,sBACd;EAAW;IACV;iBA2Ba,+BAA+B,SAAS;iBA+BxC,+BAA+B,QAAQ,SAAS,eAC9D,KAAK,iBAAiB,QAAQ,SAAS,gBACvC;EACE;EACA;EACA,eAAe;EACf,oBAAoB,MAAM,YAAY,QAAQ,SAAS;EACvD,0BACE,MAAM,YAAY,QAAQ,SAAS,mBAChC;EACL;IAED"}
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
|
-
import { a as RunRecord } from "./run-record-
|
|
3
|
-
import {
|
|
4
|
-
import {
|
|
5
|
-
import {
|
|
6
|
-
import { _ as ExternalTextCandidate, c as ExternalOptimizerModelBudget, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, m as ExternalOptimizerResumeMode, r as ExternalOptimizerCallbackLimits, v as ExternalTextEvaluationRequest } from "./external-optimizer-contracts-
|
|
7
|
-
import { t as DetectRewardHackingInput } from "./reward-hacking-
|
|
2
|
+
import { a as RunRecord } from "./run-record-DTv1MdjK.js";
|
|
3
|
+
import { it as ServedModelPolicy, p as ChatClient, w as LlmCallMetadata } from "./types-gvRsyJLh.js";
|
|
4
|
+
import { R as Scenario, S as JudgeConfig, i as CampaignCostMeter, j as MutableSurface, o as CampaignScenarioIdentity, p as Gate } from "./types-Ba5UQyVD.js";
|
|
5
|
+
import { A as CanaryOptions, F as OptimizationMethod, U as OptimizationPackageSource, b as RedTeamCase, bt as SearchRunIdentity } from "./provenance-CRY67X50.js";
|
|
6
|
+
import { _ as ExternalTextCandidate, c as ExternalOptimizerModelBudget, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, m as ExternalOptimizerResumeMode, r as ExternalOptimizerCallbackLimits, v as ExternalTextEvaluationRequest } from "./external-optimizer-contracts-CQCpyrIL.js";
|
|
7
|
+
import { t as DetectRewardHackingInput } from "./reward-hacking-ZXEi9VCq.js";
|
|
8
8
|
//#region src/campaign/coverage.d.ts
|
|
9
9
|
/** Reject campaign designs whose denominator cannot be identified exactly. */
|
|
10
10
|
declare function assertCampaignDesign<TScenario extends Scenario>(scenarios: readonly TScenario[], reps: number): void;
|
|
@@ -450,4 +450,4 @@ interface HeldOutGateOptions<TScenario extends Scenario = Scenario> {
|
|
|
450
450
|
declare function heldOutGate<TArtifact, TScenario extends Scenario>(options: HeldOutGateOptions<TScenario>): Gate<TArtifact, TScenario>;
|
|
451
451
|
//#endregion
|
|
452
452
|
export { ReferenceEquivalenceJudgeInput as A, campaignSplitDigestFromIdentities as B, ExternalTextOptimizerContext as C, decodeExternalTextCandidate as D, ExternalTextEvaluationResponse as E, runReferenceEquivalenceJudge as F, assertCampaignDesign as I, assertCampaignSplitIdentity as L, ReferenceEquivalenceJudgeResult as M, ReferenceEquivalenceScenario as N, REFERENCE_EQUIVALENCE_INPUT_LIMITS as O, createReferenceEquivalenceJudge as P, campaignScenarioIdentity as R, ExternalTextOptimizationMethodConfig as S, ExternalOptimizationExample as T, GepaRunnerCommand as _, DefaultProductionRewardHackingOptions as a, OptimizerModelBudget as b, SkillOptOptimizationMethodConfig as c, skillOptOptimizationMethod as d, GepaAdaptiveEngineRun as f, GepaOptimizationRecipe as g, GepaOptimizationMethodConfig as h, DefaultProductionGateOptions as i, ReferenceEquivalenceJudgeOptions as j, REFERENCE_EQUIVALENCE_JUDGE_VERSION as k, SkillOptRunnerCommand as l, GepaEngineRun as m, heldOutGate as n, defaultProductionGate as o, GepaEngineOptions as p, DefaultProductionGateCheck as r, composeGate as s, HeldOutGateOptions as t, SkillOptTrainerConfig as u, gepaOptimizationMethod as v, ExternalTextOptimizerResult as w, externalTextOptimizationMethod as x, OpenAICompatibleOptimizerModel as y, campaignSplitDigest as z };
|
|
453
|
-
//# sourceMappingURL=heldout-gate-
|
|
453
|
+
//# sourceMappingURL=heldout-gate-Df5hsqmm.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"heldout-gate-
|
|
1
|
+
{"version":3,"file":"heldout-gate-Df5hsqmm.d.ts","names":[],"sources":["../src/campaign/coverage.ts","../src/reference-equivalence-judge.ts","../src/campaign/external-text-evaluation.ts","../src/campaign/external-text-optimization-contract.ts","../src/campaign/external-text-optimization.ts","../src/campaign/optimizer-model.ts","../src/campaign/gepa-optimization-method.ts","../src/campaign/skillopt-optimization-method.ts","../src/campaign/gates/compose.ts","../src/campaign/gates/default-production-gate.ts","../src/campaign/gates/heldout-gate.ts"],"mappings":";;;;;;;;;iBAWgB,qBAAqB,kBAAkB,UACrD,oBAAoB,aACpB;;iBA2Bc,yBAAyB,kBAAkB,UACzD,UAAU,YACT,2BAA2B,KAAK;;iBAUnB,kCACd,oBAAoB,4BACpB;;iBAgBc,oBAAoB,kBAAkB,UACpD,oBAAoB,aACpB;;iBAOc,4BACd,oBAAoB,4BACpB,cACA;;;cC3EW;cAEA;WACX;WACA;WACA;;UAGe,qCAAqC;EACpD;EACA;;UAGe;EACf;EACA;EACA;;UAGe;;EAEf,MAAM;;EAEN;;EAEA,SAAS;;EAET,aAAa;;UAGE,wCAAwC;EACvD;EACA;EACA;EACA;;;iBA0Bc,gCACd,SAAS,mCACR,oBAAoB;;iBA2CD,6BACpB,OAAO,gCACP,SAAS,mCACR,QAAQ;;;UCrGM;EACf;EACA;;UAGe;EACf;EACA;IACE;IACA,YAAY;IACZ;IACA;;;IAGA;IACA;MACE;MACA;;;;iBAwHU,4BAA4B,WAAW,wBAAwB;;;UC3I9D;WACN;WACA;WACA;WACA;WACA;WACA,eAAe;WACf,mBAAmB;WACnB,uBAAuB;WACvB;WACA;;WAEA;WACA;WACA;WACA,QAAQ;;WAER,MAAM;WACN,WACP,SAFa,kCAGV,QAAQ;;UAGE;EACf,eAAe;EACf;EACA;IACM;;IACA;;IACA;IAAkB;;;;;;;;;UAqBT,qCACf,kBAAkB,UAClB;EAEA;EACA,QAAQ,KAAK;EACb;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA,SAAS;EACT;EACA;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,MAAM,SAAS,iCAAiC,QAAQ;;;;;;;;;;iBCN1C,+BAA+B,kBAAkB,UAAU,WACzE,QAAQ,qCAAqC,WAAW,aACvD,mBAAmB,WAAW;;;KCvErB,uBAAuB;;UAGlB;EACf;EACA,QAAQ;;EAER,MAAM;;EAEN;;;;;;;EAOA,oBAAoB;;;;;;;EAOpB;;;;;UCyCe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;EAKA,eAAe;;;UAIA,sBAAsB;;EAErC;;;KAIU,wBAAwB;;;;;;;KAQxB;EAEN;EACA,KAAK;;EAGL;EACA,eAAe;;EAGf;EACA,eAAe;;EAEf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,kBAAkB;EAClB,cAAc;EACd;;;KAIM,oBAAoB;UAEf,6BAA6B,kBAAkB,UAAU;;EAExE;;EAEA,QAAQ;;EAER;;EAEA;;EAEA;;;;;EAKA;;EAEA;;EAEA;;EAEA,2BAA2B,QAAQ;;EAEnC;;;;;;EAMA,YAAY;;;;;;EAMZ,oBAAoB,UAAU;;EAE9B,oBAAoB,UAAU,WAAW,UAAU;;EAEnD,SAAS;;;;;EAKT;EACA,SAAS;;;;;;;;;;EAUT;IAAiB,UAAU;IAAmB;;;iBA6ChC,uBAAuB,kBAAkB,UAAU,WACjE,QAAQ,6BAA6B,WAAW,aAC/C,mBAAmB,WAAW;;;UC7LhB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;EAKA,YAAY;;KAGF,wBAAwB;UAEnB,iCAAiC,kBAAkB,UAAU;EAC5E;;EAEA;EACA;;EAEA;EACA,SAAS;;;;;EAKT,WAAW;;EAEX;;EAEA;EACA;;EAEA;;EAEA,2BAA2B,QAAQ;EACnC;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,SAAS;EACT,SAAS;;;iBAIK,2BAA2B,kBAAkB,UAAU,WACrE,QAAQ,iCAAiC,WAAW,aACnD,mBAAmB,WAAW;;;;;;iBC1GjB,YAAY,qBAAqB,kBAAkB,WAAW,aACzE,OAAO,MAAM,KAAK,WAAW,cAC/B,KAAK,WAAW;;;KCaP;KAOA,wCAAwC,KAClD;EAGA,SAAS,YAAY;;UAGN;;;EAGf,kBAAkB;;;;;EAKlB;;EAEA;;EAEA;;EAEA;;;;EAIA;;;;EAIA;;;;;EAKA;;;;EAIA;;;EAGA;;;;EAIA,iBAAiB;;;EAGjB,aAAa;;EAEb,gBAAgB;;EAEhB,SAAS;;;;;EAKT,iBAAiB;;;;;iBAMH,sBAAsB,WAAW,kBAAkB,UACjE,SAAS,+BACR,KAAK,WAAW;;;UC/EF,mBAAmB,kBAAkB,WAAW;EAC/D,WAAW;;;EAGX;;EAEA;;;;EAIA;;EAEA;;EAEA;;;;;;;;iBASc,YAAY,WAAW,kBAAkB,UACvD,SAAS,mBAAmB,aAC3B,KAAK,WAAW"}
|
package/dist/hosted/index.d.ts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { o as InsightReport } from "../insight-report-
|
|
2
|
-
import { _ as UnixNanoTimestamp, a as hostedTenantFromEnv, c as EvalRunGenerationSnapshot, d as HostedIngestHeaders, f as HostedWireVersion, g as TraceSpanEvent, h as IngestTracesRequest, i as hostedClientFromEnv, l as EvalRunStatus, m as IngestResponse, n as HostedTenant, o as EvalRunCellScore, p as IngestEvalRunsRequest, r as createHostedClient, s as EvalRunEvent, t as HostedClient, u as HOSTED_WIRE_VERSION } from "../client-
|
|
1
|
+
import { o as InsightReport } from "../insight-report-DETqPc_A.js";
|
|
2
|
+
import { _ as UnixNanoTimestamp, a as hostedTenantFromEnv, c as EvalRunGenerationSnapshot, d as HostedIngestHeaders, f as HostedWireVersion, g as TraceSpanEvent, h as IngestTracesRequest, i as hostedClientFromEnv, l as EvalRunStatus, m as IngestResponse, n as HostedTenant, o as EvalRunCellScore, p as IngestEvalRunsRequest, r as createHostedClient, s as EvalRunEvent, t as HostedClient, u as HOSTED_WIRE_VERSION } from "../client-DlqdbM7n.js";
|
|
3
3
|
import { z } from "zod";
|
|
4
4
|
//#region src/hosted/schemas.d.ts
|
|
5
5
|
declare const UnixNanoTimestampSchema: z.ZodType<UnixNanoTimestamp>;
|
package/dist/hosted/index.js
CHANGED
|
@@ -1,2 +1,2 @@
|
|
|
1
|
-
import { a as IngestEvalRunsRequestSchema, c as InsightReportSchema, d as HOSTED_WIRE_VERSION, i as EvalRunEventSchema, l as TraceSpanEventSchema, n as hostedClientFromEnv, o as IngestResponseSchema, r as hostedTenantFromEnv, s as IngestTracesRequestSchema, t as createHostedClient, u as UnixNanoTimestampSchema } from "../client-
|
|
1
|
+
import { a as IngestEvalRunsRequestSchema, c as InsightReportSchema, d as HOSTED_WIRE_VERSION, i as EvalRunEventSchema, l as TraceSpanEventSchema, n as hostedClientFromEnv, o as IngestResponseSchema, r as hostedTenantFromEnv, s as IngestTracesRequestSchema, t as createHostedClient, u as UnixNanoTimestampSchema } from "../client-BlLY6o2w.js";
|
|
2
2
|
export { EvalRunEventSchema, HOSTED_WIRE_VERSION, IngestEvalRunsRequestSchema, IngestResponseSchema, IngestTracesRequestSchema, InsightReportSchema, TraceSpanEventSchema, UnixNanoTimestampSchema, createHostedClient, hostedClientFromEnv, hostedTenantFromEnv };
|
|
@@ -1,15 +1,15 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
|
-
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "./run-record-
|
|
3
|
-
import "./types-
|
|
4
|
-
import "./
|
|
5
|
-
import
|
|
6
|
-
import "./
|
|
7
|
-
import "./promotion-policy-
|
|
8
|
-
import { o as InsightReport } from "./insight-report-
|
|
9
|
-
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, o as EvalRunCellScore, s as EvalRunEvent } from "./client-
|
|
10
|
-
import "./define-agent-eval-
|
|
11
|
-
import "./default-registry-
|
|
12
|
-
import {
|
|
2
|
+
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "./run-record-DTv1MdjK.js";
|
|
3
|
+
import "./types-DN2WdT5S.js";
|
|
4
|
+
import { _ as GateDecision } from "./types-Ba5UQyVD.js";
|
|
5
|
+
import "./heldout-gate-Df5hsqmm.js";
|
|
6
|
+
import "./provenance-CRY67X50.js";
|
|
7
|
+
import "./promotion-policy-CvMda3kU.js";
|
|
8
|
+
import { o as InsightReport } from "./insight-report-DETqPc_A.js";
|
|
9
|
+
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, o as EvalRunCellScore, s as EvalRunEvent } from "./client-DlqdbM7n.js";
|
|
10
|
+
import "./define-agent-eval-CY6qdlGV.js";
|
|
11
|
+
import "./default-registry-BKwc8bN5.js";
|
|
12
|
+
import { v as AnalyzeRunsOptions } from "./engine-DhFir3Ys.js";
|
|
13
13
|
import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementCost, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
|
|
14
14
|
//#region src/contract/measured-comparison.d.ts
|
|
15
15
|
interface SealCandidateBenchmarkSuiteOptions {
|
|
@@ -612,6 +612,87 @@ declare function fromOpenCodeSession(options: CodeAgentSessionIntakeOptions): Co
|
|
|
612
612
|
declare function fromKimiCodeSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
613
613
|
declare function fromPiSession(options: CodeAgentSessionIntakeOptions): CodeAgentSessionIntakeResult;
|
|
614
614
|
//#endregion
|
|
615
|
+
//#region src/contract/intake/code-agent-store.d.ts
|
|
616
|
+
/** Which store to read, and what makes a session in it belong to this run. */
|
|
617
|
+
interface CodeAgentStoreScope {
|
|
618
|
+
/**
|
|
619
|
+
* Absolute path to the run's OWN captured session store. Required, with no default: a default
|
|
620
|
+
* would be the host store, and the host store is another run's evidence.
|
|
621
|
+
*/
|
|
622
|
+
readonly root: string;
|
|
623
|
+
/** The harness family that wrote it. */
|
|
624
|
+
readonly source: CodeAgentSessionSource;
|
|
625
|
+
/**
|
|
626
|
+
* Attribute only a session whose recorded working directory is this path or below it. A session
|
|
627
|
+
* recording no working directory is rejected while this is set.
|
|
628
|
+
*/
|
|
629
|
+
readonly workspaceRoot?: string;
|
|
630
|
+
/**
|
|
631
|
+
* Attribute only a session that started inside the run. A session recording no start is rejected
|
|
632
|
+
* while this is set.
|
|
633
|
+
*/
|
|
634
|
+
readonly window?: {
|
|
635
|
+
readonly startedAtMs: number;
|
|
636
|
+
/** Absent leaves the window open-ended, for a run still in flight. */
|
|
637
|
+
readonly finishedAtMs?: number;
|
|
638
|
+
};
|
|
639
|
+
}
|
|
640
|
+
/** Why a file under `root` was not attributed to the run. */
|
|
641
|
+
type CodeAgentStoreRejection = 'outside-workspace' | 'no-workspace-recorded' | 'before-run' | 'after-run' | 'no-start-recorded' | 'no-entries' | 'unreadable';
|
|
642
|
+
/** One attributed session, named so a caller can cite the file the numbers came from. */
|
|
643
|
+
interface CodeAgentStoreSession {
|
|
644
|
+
readonly path: string;
|
|
645
|
+
readonly sessionId: string;
|
|
646
|
+
readonly workspace?: string;
|
|
647
|
+
readonly startedAtMs?: number;
|
|
648
|
+
readonly observation: CodeAgentSessionObservation;
|
|
649
|
+
}
|
|
650
|
+
/** Every file the scan looked at and could not attribute. */
|
|
651
|
+
interface CodeAgentStoreRejections {
|
|
652
|
+
readonly total: number;
|
|
653
|
+
readonly byReason: Readonly<Partial<Record<CodeAgentStoreRejection, number>>>;
|
|
654
|
+
/** The first few rejected paths with their reason, for a caller that must explain a zero. */
|
|
655
|
+
readonly sample: ReadonlyArray<{
|
|
656
|
+
readonly path: string;
|
|
657
|
+
readonly reason: CodeAgentStoreRejection;
|
|
658
|
+
}>;
|
|
659
|
+
}
|
|
660
|
+
/**
|
|
661
|
+
* What one store scan found.
|
|
662
|
+
*
|
|
663
|
+
* The union is the point. An `unavailable` result carries NO action counts, so no caller can print
|
|
664
|
+
* a `0` — or a number from another run — as this run's measurement.
|
|
665
|
+
*/
|
|
666
|
+
type CodeAgentStoreObservation = {
|
|
667
|
+
readonly status: 'observed';
|
|
668
|
+
readonly scope: CodeAgentStoreScope;
|
|
669
|
+
readonly filesScanned: number;
|
|
670
|
+
readonly sessions: readonly CodeAgentStoreSession[];
|
|
671
|
+
/** Total actions across the attributed sessions. */
|
|
672
|
+
readonly actions: number;
|
|
673
|
+
/** Actions on the `subagent` surface — harness-native fan-out this run performed. */
|
|
674
|
+
readonly subagentActions: number;
|
|
675
|
+
/** Sessions carrying at least one `subagent` action. */
|
|
676
|
+
readonly sessionsWithSubagentActions: number;
|
|
677
|
+
readonly bySurface: Readonly<Partial<Record<CodeAgentSessionActionSurface, number>>>;
|
|
678
|
+
readonly rejections: CodeAgentStoreRejections;
|
|
679
|
+
} | {
|
|
680
|
+
readonly status: 'unavailable';
|
|
681
|
+
readonly scope: CodeAgentStoreScope;
|
|
682
|
+
readonly filesScanned: number;
|
|
683
|
+
/** Why nothing could be counted, in the terms a report can print verbatim. */
|
|
684
|
+
readonly reason: string;
|
|
685
|
+
readonly rejections: CodeAgentStoreRejections;
|
|
686
|
+
};
|
|
687
|
+
/**
|
|
688
|
+
* Read every session this run captured under `scope.root` and project them.
|
|
689
|
+
*
|
|
690
|
+
* A store directory that does not exist, or that holds no session the scope attributes, returns
|
|
691
|
+
* `status: 'unavailable'`. That is not an error: a run whose harness store was never synced back
|
|
692
|
+
* has no evidence, and saying so is the correct answer.
|
|
693
|
+
*/
|
|
694
|
+
declare function observeCodeAgentStore(scope: CodeAgentStoreScope): Promise<CodeAgentStoreObservation>;
|
|
695
|
+
//#endregion
|
|
615
696
|
//#region src/contract/intake/feedback-table.d.ts
|
|
616
697
|
interface FeedbackTableRow {
|
|
617
698
|
/** Stable id for this run — the unit a rater scored. Drives pairing
|
|
@@ -700,5 +781,5 @@ interface FromOtelSpansOptions {
|
|
|
700
781
|
}
|
|
701
782
|
declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
|
|
702
783
|
//#endregion
|
|
703
|
-
export {
|
|
704
|
-
//# sourceMappingURL=index-
|
|
784
|
+
export { diffRunBaselineToWinner as $, CodeAgentSessionActionStatus as A, measuredComparisonFromCandidateExperiment as At, AgentTraceFile as B, fromOpenCodeSession as C, EvaluatePairedMeasurementsOptions as Ct, streamCodeAgentJsonlFile as D, RunCandidateExperimentOptions as Dt, parseCodeAgentJsonlFile as E, PairedMeasurementEvaluation as Et, CodeAgentSessionTerminalStatus as F, verifyCandidateExperiment as Ft, PartitionByAuthoringModelResult as G, AgentTraceRange as H, observeCodeAgentSession as I, verifyCandidateExperimentComparison as It, EvalCellScoreDelta as J, parseAgentTrace as K, AgentTraceContributor as L, CodeAgentSessionExecutionReceipt as M, sealCandidateBenchmarkSuite as Mt, CodeAgentSessionObservation as N, sealCandidateBenchmarkTask as Nt, CodeAgentSessionAction as O, SealCandidateBenchmarkSuiteOptions as Ot, CodeAgentSessionSource as P, sealCandidateExperiment as Pt, diffGenerations as Q, AgentTraceContributorType as R, fromKimiCodeSession as S, CompareCandidateExperimentOptions as St, parseCodeAgentJsonl as T, PairedMeasurementAdapter as Tt, AgentTraceRecord as U, AgentTraceIndex as V, AuthoringProvenance as W, EvalGenerationDiff as X, EvalDimensionDelta as Y, EvalRunDiff as Z, CodeAgentSessionIntakeResult as _, sealAgentProfileImprovementSuite as _t, FromFeedbackTableOptions as a, FromRunRecordDirOptions as at, fromClaudeCodeSession as b, CandidateExperimentExecutionInput as bt, CodeAgentStoreObservation as c, fromRunRecordDir as ct, CodeAgentStoreScope as d, CompareAgentProfileImprovementExperimentOptions as dt, diffRuns as et, CodeAgentStoreSession as f, RunAgentProfileImprovementExperimentOptions as ft, CodeAgentSessionIntakeOptions as g, sealAgentProfileImprovementExperiment as gt, CodeAgentSessionDiagnostic as h, runAgentProfileImprovementExperiment as ht, FeedbackTableRow as i, evalReportingSuite as it, CodeAgentSessionActionSurface as j, runCandidateExperiment as jt, CodeAgentSessionActionKind as k, evaluatePairedMeasurements as kt, CodeAgentStoreRejection as l, AgentProfileImprovementExperimentExecutionInput as lt, CodeAgentJsonlLine as m, measuredComparisonFromAgentProfileImprovementExperiment as mt, fromOtelSpans as n, EvalReportingSuiteOptions as nt, FromFeedbackTableResult as o, FromRunRecordDirResult as ot, observeCodeAgentStore as p, SealAgentProfileImprovementSuiteOptions as pt, partitionRunsByAuthoringModel as q, FeedbackTableMeta as r, EvalReportingSuiteResult as rt, fromFeedbackTable as s, RunRecordRejection as st, FromOtelSpansOptions as t, EvalReportingSuiteInput as tt, CodeAgentStoreRejections as u, AgentProfileImprovementExperimentRun as ut, CodeAgentSessionMetrics as v, sealAgentProfileImprovementTask as vt, fromPiSession as w, PairedMeasurement as wt, fromCodexSession as x, CandidateExperimentRun as xt, ParsedCodeAgentJsonl as y, verifyAgentProfileImprovementExperimentComparison as yt, AgentTraceConversation as z };
|
|
785
|
+
//# sourceMappingURL=index-BQqOjerE.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-BQqOjerE.d.ts","names":[],"sources":["../src/contract/measured-comparison.ts","../src/contract/profile-measured-comparison.ts","../src/contract/intake/run-record-dir.ts","../src/contract/eval-reporting-suite.ts","../src/contract/diff.ts","../src/contract/intake/agent-trace.ts","../src/contract/intake/code-agent-observation.ts","../src/contract/intake/code-agent-session.ts","../src/contract/intake/code-agent-store.ts","../src/contract/intake/feedback-table.ts","../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;;;;UAiCiB;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,eAAe,KAAK,OAAO;EAC3B,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA,kBAAkB;;EAElB,kBAAkB;;;KAIR,8BAA8B,KACxC;EAGA,iBAAiB;EACjB,WAAW;EACX;;;iBAIc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;;;;;;;iBAkEK,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;;iBA8Ya,0CACd,SAAS,oCACR;;iBAuEa,oCACd,iBACC;;;UCjrBc;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;;;;;;iBAsBmB,qCACpB,SAAS,8CACR,QAAQ;;iBA0CK,wDACd,SAAS,kDACR;;iBAuEa,kDACd,iBACC;;;;UC7Oc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;;;;;;;iBAkBoB,iBACpB,cACA,UAAS,0BACR,QAAQ;;;;;KC7CC,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;;;;;;iBAUoB,mBACpB,OAAO,yBACP,UAAS,4BACR,QAAQ;;;;;;UC5DM;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;EAGA,YAAY,eAAe,eAAe;;;;UAK3B;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;EAET,SAAS;;EAET,OAAO;;EAEP;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;UAMe;EACf;EACA;EACA;EACA;EACA,oBAAoB;EACpB,mBAAmB;EACnB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA,cAAc;;;;EAId,aAAa;;;;;;;;;iBAoDC,gBACd,QAAQ,2BACR,OAAO,4BACN;;;;;;iBAqEa,SAAS,QAAQ,cAAc,OAAO,eAAe;;;;;;;iBAsCrD,wBAAwB,KAAK,eAAe;;;KC1OhD;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;KCjLS;KAEA;KAEA;KAEA;KASA;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,MAAM;EACN,SAAS;EACT;EACA,QAAQ;EACR;EACA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR;EACA;EACA;IACE,QAAQ;IACR;;EAEF,SAAS;;UAGM;EACf,QAAQ;EACR;EACA;EACA,YAAY;;;;;;;iBAeE,wBACd,SAAS,iCACR;;;UCjCc;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;;UAGe;EACf,MAAM;EACN,aAAa;EACb,SAAS;EACT,cAAc;;UAGC;EACf;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;EACA;EACA;EACA;EACA;EACA;;;EAGA,iBAAiB;;;EAGjB,YAAY;;;;KAKF;EACN;EAAe;EAAoB;;EACnC;EAAmB;;iBAcT,oBAAoB,gBAAgB;;;;;;;;;;iBAuB7B,yBAAyB,eAAe,eAAe;;;;;iBA6BxD,wBAAwB,eAAe,QAAQ;iBAUrD,iBACd,SAAS,gCACR;iBAIa,sBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,cACd,SAAS,gCACR;;;;UC/Lc;;;;;WAKN;;WAEA,QAAQ;;;;;WAKR;;;;;WAKA;aACE;;aAEA;;;;KAKD;;UAUK;WACN;WACA;WACA;WACA;WACA,aAAa;;;UAIP;WACN;WACA,UAAU,SAAS,QAAQ,OAAO;;WAElC,QAAQ;aACN;aACA,QAAQ;;;;;;;;;KAUT;WAEG;WACA,OAAO;WACP;WACA,mBAAmB;;WAEnB;;WAEA;;WAEA;WACA,WAAW,SAAS,QAAQ,OAAO;WACnC,YAAY;;WAGZ;WACA,OAAO;WACP;;WAEA;WACA,YAAY;;;;;;;;;iBAaL,sBACpB,OAAO,sBACN,QAAQ;;;UCtHM;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;;;UCvBlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}
|
|
@@ -1,17 +1,17 @@
|
|
|
1
1
|
import { c as ValidationError, t as AgentEvalError } from "./errors-DEE6u6ot.js";
|
|
2
2
|
import { T as RunPaidCallInput, p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
|
|
3
|
-
import { a as RunRecord, s as RunSplitTag } from "./run-record-
|
|
4
|
-
import { f as AnalystUsageReceipt, i as AnalystFinding, l as AnalystRunResult, w as TraceAnalysisStore } from "./types-
|
|
5
|
-
import {
|
|
6
|
-
import { _ as PairedArmsComparison } from "./pre-registration-DHz6P_6f.js";
|
|
3
|
+
import { a as RunRecord, s as RunSplitTag } from "./run-record-DTv1MdjK.js";
|
|
4
|
+
import { f as AnalystUsageReceipt, i as AnalystFinding, l as AnalystRunResult, w as TraceAnalysisStore } from "./types-DN2WdT5S.js";
|
|
5
|
+
import { _ as PairedArmsComparison } from "./pre-registration-BoI4ucR3.js";
|
|
7
6
|
import { d as PairedBootstrapResult } from "./paired-promotion-decision-CGzg0cI_.js";
|
|
8
|
-
import { A as LabeledScenarioWrite, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, p as Gate, u as ComponentSurface } from "./types-
|
|
9
|
-
import "./heldout-gate-
|
|
10
|
-
import "./
|
|
11
|
-
import
|
|
12
|
-
import {
|
|
7
|
+
import { A as LabeledScenarioWrite, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, p as Gate, u as ComponentSurface } from "./types-Ba5UQyVD.js";
|
|
8
|
+
import "./heldout-gate-Df5hsqmm.js";
|
|
9
|
+
import { In as CampaignRunPlan, Kn as CampaignStorage, Rn as PlanCampaignRunOptions } from "./provenance-CRY67X50.js";
|
|
10
|
+
import "./promotion-policy-CvMda3kU.js";
|
|
11
|
+
import { d as CompletionVerdict, f as CorrectnessChecker, h as ProducedState, n as BackendIntegrityReport, s as RuntimeEventLike, u as CompletionRequirement } from "./backend-integrity-CeuTgqsd.js";
|
|
12
|
+
import { _ as AnalystIssueExpectation, a as AnalystBenchmarkLabelState, t as AnalystBenchmarkCase } from "./benchmark-BjLGkfnN.js";
|
|
13
13
|
import { t as AgentProfile$1 } from "./agent-profile-B9_GGsG8.js";
|
|
14
|
-
import "./statistical-heldout-
|
|
14
|
+
import "./statistical-heldout-DTyB_6-1.js";
|
|
15
15
|
import { AgentProfile } from "@tangle-network/agent-interface";
|
|
16
16
|
//#region src/campaign/search-ledger-errors.d.ts
|
|
17
17
|
/** Base error for invalid search-ledger input or operations. */
|
|
@@ -1132,4 +1132,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
|
|
|
1132
1132
|
declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
|
|
1133
1133
|
//#endregion
|
|
1134
1134
|
export { makePlaybackDispatch as $, CrossSurfaceEvidenceBreakdown as $t, scoreDiscrimination as A, LoadEvalFixtureScenariosOptions as At, ProfileMatrixSegmentResult as B, CrossSurfaceAttemptCompleteness as Bt, acquireSingleRunLock as C, SearchLedgerIntegrityError as Cn, neutralizationGate as Ct, compareRankKeys as D, EvalFixtureRunPlan as Dt, campaignMeanComposite as E, EvalFixtureLoadOptions as Et, FinalizeProfileMatrixOptions as F, planEvalFixtureRun as Ft, PlaybackContext as G, CrossSurfaceCandidateEvidence as Gt, createProfileMatrixPlan as H, CrossSurfaceBootstrapPolicy as Ht, FinalizedProfileMatrixResult as I, analyzeCrossSurfaceInteractions as It, ScoreboardRenderOptions as J, CrossSurfaceComponent as Jt, PlaybackDriver as K, CrossSurfaceCandidateOutcome as Kt, ProfileMatrixCoverage as L, AnalyzeCrossSurfaceInteractionsInput as Lt, resolveRunDir as M, discoverEvalFixtures as Mt, tangleTracesRoot as N, loadEvalFixture as Nt, DiscriminationScore as O, EvalFixtureScenario as Ot, CreateProfileMatrixPlanOptions as P, loadEvalFixtureScenarios as Pt, UserStoryVerdict as Q, CrossSurfaceEligibility as Qt, ProfileMatrixPlan as R, CrossSurfaceAdditionDecision as Rt, SingleRunLockOptions as S, SearchLedgerError as Sn, NeutralizationGateOptions as St, campaignBreakdown as T, EvalFixtureFile as Tt, finalizeProfileMatrix as U, CrossSurfaceCandidate as Ut, RunProfileMatrixSegmentOptions as V, CrossSurfaceBestSingleSelection as Vt, runProfileMatrixSegment as W, CrossSurfaceCandidateComparison as Wt, ScoreboardSummary as X, CrossSurfaceCompositionStep as Xt, ScoreboardRow as Y, CrossSurfaceComponentEvidence as Yt, UserStory as Z, CrossSurfaceDistribution as Zt, componentSurfaceIdentityMaterial as _, TraceAnalystArtifact as _n, RolloutCall as _t, WorktreeAdapterError as a, CrossSurfaceInteractionTask as an, ProfileMatrixError as at, surfaceHash as b, traceAnalystQualityJudge as bn, classifyUngroundedLiterals as bt, verifyCodeSurface as c, CrossSurfacePairEvidence as cn, RunProfileMatrixResult as ct, PhoenixEvaluationResultLike as d, CrossSurfaceRankedSingle as dn, neutralizeText as dt, CrossSurfaceIneligibilityReason as en, renderScoreboardMarkdown as et, PhoenixEvaluatorLike as f, CrossSurfaceRelativeCost as fn, FsLabeledScenarioStore as ft, codeSurfaceIdentityMaterial as g, BuildTraceAnalystSurfaceDispatchOptions as gn, RolloutArgumentDiffOptions as gt, assertCodeSurfaceIdentity as h, CrossSurfaceTaskRow as hn, RolloutArgumentDiff as ht, WorktreeAdapter as i, CrossSurfaceInteractionReport as in, ProfileDispatchFn as it, selectDiscriminative as j, PlanEvalFixtureRunOptions as jt, ScenarioSignal as k, EvalFixtureValidationMode as kt, AutoevalsScoreLike as l, CrossSurfacePairIncompatibilityReason as ln, ScenarioRollup as lt, phoenixEvaluatorJudge as m, CrossSurfaceSelections as mn, LabeledScenarioStoreError as mt, GitWorktreeAdapterOptions as n, CrossSurfaceInteractionEffect as nn, scoreboardSummary as nt, gitWorktreeAdapter as o, CrossSurfaceNaiveStackSelection as on, ProfileSummary as ot, autoevalsScorerJudge as p, CrossSurfaceSelectionPolicy as pn, FsLabeledScenarioStoreOptions as pt, PlaybackStep as q, CrossSurfaceCandidateSummary as qt, Worktree as r, CrossSurfaceInteractionPath as rn, userStoryScoreboard as rt, resolveWorktreePath as s, CrossSurfacePairCompatibility as sn, RunProfileMatrixOptions as st, CodeSurfaceVerification as t, CrossSurfaceInteractionAwareSelection as tn, scoreUserStory as tt, AutoevalsScorerLike as u, CrossSurfacePairwiseEntry as un, runProfileMatrix as ut, renderSurfaceDiff as v, TraceAnalystScenario as vn, ScoredRollout as vt, CampaignBreakdown as w, EvalFixture as wt, SingleRunLock as x, SearchLedgerConflictError as xn, rolloutArgumentDiff as xt, surfaceContentHash as y, buildTraceAnalystSurfaceDispatch as yn, UngroundedLiteralReport as yt, ProfileMatrixRow as z, CrossSurfaceAdditionRejectionReason as zt };
|
|
1135
|
-
//# sourceMappingURL=index-
|
|
1135
|
+
//# sourceMappingURL=index-CFDffsKz.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-CFDffsKz.d.ts","names":[],"sources":["../src/campaign/search-ledger-errors.ts","../src/campaign/analyst-surface.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/presets/segmented-profile-matrix.ts","../src/campaign/run-dir.ts","../src/campaign/scenario-selection.ts","../src/campaign/score-utils.ts","../src/campaign/single-run-lock.ts","../src/campaign/surface-identity.ts","../src/campaign/upstream-evaluators.ts","../src/campaign/worktree/index.ts"],"mappings":";;;;;;;;;;;;;;;;;cAGa,0BAA0B;;cAG1B,mCAAmC;;cAGnC,kCAAkC;;;UCE9B,6BAA6B;EAC5C;EACA,YAAY;EACZ,YAAY;EACZ,yBAAyB;EACzB,kBAAkB;;UAGH;EACf,mBAAmB;EACnB,QAAQ;EACR,MAAM;;UAGS;EACf,QAAQ;IACN;IACA,YAAY;IACZ;IACA,QAAQ;MACN,QAAQ;;iBAGE,iCACd,SAAS,2CAET,SAAS,gBACT,UAAU,sBACV,SAAS,oBACN,QAAQ;iBAgBG,4BAA4B,YAC1C,sBACA;;;;KCtDU;;UAGK;EACf;EACA;;EAEA;;;UAIe;EACf;EACA;EACA;EACA;;;UAIe;EACf;;EAEA;;EAEA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA,cAAc;EACd;EACA;;;;;;EAMA,MAAM;EACN,mBAAmB;;EAEnB;;UAGe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA,kCAAkC;;EAElC;;UAGe,qCACf,aAAa,sBAAsB;EAEnC,qBAAqB;EACrB,qBAAqB;EACrB,eAAe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;EACA,WAAW;EACX,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;UAGe;EACf,aAAa;EACb;EACA;EACA;EACA;;KAGU;UAWK;EACf;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,WAAW;EACX,SAAS;EACT,OAAO;EACP,OAAO,eAAe;EACtB,QAAQ;EACR,QAAQ;;EAER,sBAAsB;;EAEtB,aAAa;;UAGE;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,cAAc,eAAe;;UAGd;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;UAGe;EACf,SAAS;EACT;EACA;EACA;EACA;EACA,eAAe;EACf,gBAAgB;;KAGN;UAYK;EACf;EACA,SAAS;EACT;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,4BAA4B,iCAAiC;EAC7D,wBAAwB,eAAe;EACvC,QAAQ;EACR,QAAQ;EACR,aAAa;EACb,eAAe;;UAGA;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;;EAEf;EACA;;KAGU;UAaK;EACf;EACA;EACA;EACA;EACA;EACA,uBAAuB;EACvB;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,YAAY;EACZ;;;UAIe;EACf;EACA;EACA;EACA;EACA,OAAO;;UAGQ;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA,gBAAgB;;EAEhB,OAAO;;UAGQ;EACf,YAAY;EACZ,YAAY;EACZ,kBAAkB;;UAGH,8BACf,aAAa,sBAAsB;EAEnC;EACA;EACA;EACA;;EAEA,MAAM;EACN,iBAAiB;EACjB,iBAAiB;EACjB,YAAY;EACZ,UAAU;EACV,YAAY;;;;;;;;;iBCnRE,gCAAgC,aAAa,qBAC3D,OAAO,qCAAqC,QAC3C,8BAA8B;;;KC7CrB;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,4BAA4B;EAC3C;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf,aAAa;;EAEb;;UAGe,wCAAwC;EACvD;;UAGe,0BAA0B,6BACjC,KACN,uBAAuB,qBAAqB;EAG9C;EACA,aAAa;EACb;EACA;EACA,UAAU;;KAGA,qBAAqB;EAC/B,UAAU,MAAM,KAAK;;;iBAIP,qBAAqB;;;;;iBA4BrB,gBACd,kBACA,cACA,UAAS,yBACR;;iBA4Ca,yBACd,kBACA,UAAS,kCACR;;;;;iBAuBa,mBAAmB,qBACjC,SAAS,0BAA0B,aAClC;;;UChJc,0BAA0B,kBAAkB,WAAW;EACtE,WAAW;;;;;EAKX;;;;;;iBAuBc,mBAAmB,WAAW,kBAAkB,UAC9D,SAAS,0BAA0B,aAClC,KAAK,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC9BF;WACN;WACA,MAAM,SAAS;;;UAIT;;WAEN;;WAEA;WACA,gBAAgB;;UAGV;;WAEN;;WAEA;;UAGM;;WAEN;;WAEA,eAAe;;WAEf,eAAe;;;;;;;;iBASV,oBACd,mBAAmB,iBACnB,OAAM,6BACL;UAsCc;;WAEN;;WAEA;;;;;;;;;iBAUK,2BACd,cACA,MAAM,KAAK,0DACV;;;UChFc;;EAEf;;;EAGA;;EAEA;;;cAIW,kCAAkC;WAE3B;EADlB,YACkB,cAChB;;;;;;cAiBS,kCAAkC;mBAIhB;mBAHZ;mBACA;EAEjB,YAA6B,SAAS;EAKhC,QAAQ,OAAO,uBAAuB;EAYtC,OAAO,MAAM,4BAA4B,QAAQ;EA8CjD,QAAQ;IACZ;IACA;IACA,UAAU;IACV,SAAS,OAAO;;UAkCV;UAiCA;UAmBA;UAkBA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCvNM,eAAe;;;;;;cC8ClB,2BAA2B;EACtC,YAAY;;;;;KAQF,kBAAkB,kBAAkB,UAAU,cACxD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;UAEI,wBAAwB,kBAAkB,UAAU;;EAEnE,UAAU;;EAEV,WAAW;;EAEX,UAAU,kBAAkB,WAAW;;EAEvC,SAAS,YAAY,WAAW;;;EAGhC;;;EAGA;;;EAGA;;;;EAIA;;EAEA,WAAW;;EAEX;;EAEA;;;;;;;;EAQA;;;EAGA;;EAEA;;;EAGA;;EAEA;;EAEA,eAAe;EACf,gBAAgB;;;EAGhB,UAAU;;EAEV,YAAY;;;EAGZ,aAAa,UAAU;;;;EAIvB;;;;;;;EAOA,cACE,UAAU,WACV,UAAU;IACL;IAAgB;;;;;;;EAMvB,aAAa;IAAS,SAAS;IAAc,UAAU;IAAW;;;EAElE;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA,gBAAgB;;;EAGhB,WAAW;;UAGI;EACf;EACA;;UAGe,uBAAuB,WAAW,kBAAkB;EACnE;EACA;;;;EAIA,SAAS;EACT,WAAW,eAAe;EAC1B,YAAY,eAAe;;EAE3B,YAAY,eAAe;;EAE3B,WAAW;;EAEX,WAAW,eAAe,eAAe,WAAW;;;;;iBAoOhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UC9Z5B;;EAEf;;EAEA,UAAU;;;;;;;UAQK,kBAAkB;;EAEjC;;EAEA,OAAO;;EAEP,cAAc;;;UAIC,wBAAwB;EACvC,SAAS;;;;;;;;;;;UAYM,eAAe,eAAe,YAAY;EACzD,IAAI,OAAO,QAAQ,KAAK,kBAAkB,iBAAiB;;;;;;;iBAQ7C,qBAAqB,eAAe,WAClD,QAAQ,eAAe,UACtB,kBAAkB,QAAQ;;UAQZ,yBAAyB;EACxC;;;;;;;;iBASoB,eACpB,OAAO,WACP,OAAO,eACP,kBAAkB,qBACjB,QAAQ;;UAUM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;iBAQc,oBAAoB,mBAAmB,qBAAqB;;UAkB3D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;iBAIc,kBAAkB,eAAe,kBAAkB;UAwBlD;;EAEf;;EAEA,OAAO;;EAEP;;;;;;;;;iBAkBc,yBACd,eAAe,iBACf,OAAM;;;UCpKS;EACf;EACA;EACA;EACA;EACA;;UAGe,kBAAkB,kBAAkB,UAAU;WACpD;WACA;WACA;WACA;WACA,mBAAmB;WACnB,oBAAoB;WACpB,iBAAiB,YAAY,WAAW;WACxC;WACA;WACA,UAAU,YAAY,wBAAwB,WAAW;WACzD;;WAEA;WACA,WAAW,YAAY,wBAAwB,WAAW;WAC1D;WACA;WACA,aAAa,UAAU;WACvB,cACP,UAAU,WACV,UAAU;IACL;IAAgB;;WACd,eAAe;;UAGT,+BAA+B,kBAAkB,UAAU,mBAClE,KACN,wBAAwB,WAAW;;EAgBrC;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe,+BAA+B,kBAAkB,UAAU;EAC1E,MAAM,kBAAkB,WAAW;;EAEnC;;EAEA,yBAAyB;EACzB,UAAU,kBAAkB,WAAW;EACvC;EACA,UAAU;EACV;EACA;EACA;EACA,eAAe,wBAAwB,WAAW;EAClD,gBAAgB,wBAAwB,WAAW;EACnD,YAAY;;UAGG,2BAA2B,WAAW,kBAAkB;EACvE;EACA;EACA,QAAQ,uBAAuB,WAAW;EAC1C,UAAU;;UAGK,6BAA6B,kBAAkB,UAAU;EACxE,MAAM,kBAAkB,WAAW;EACnC;EACA,UAAU;EACV;EACA;EACA,YAAY;;UAGG,6BAA6B,WAAW,kBAAkB,kBACjE,uBAAuB,WAAW;EAC1C,UAAU;;iBAGI,wBAAwB,kBAAkB,UAAU,WAClE,MAAM,+BAA+B,WAAW,aAC/C,kBAAkB,WAAW;iBA0HV,wBAAwB,kBAAkB,UAAU,WACxE,MAAM,+BAA+B,WAAW,aAC/C,QAAQ,2BAA2B,WAAW;iBAkD3B,sBAAsB,kBAAkB,UAAU,WACtE,MAAM,6BAA6B,WAAW,aAC7C,QAAQ,6BAA6B,WAAW;;;;;;;;iBCrTnC;;;;;iBAQA,cAAc,gBAAgB;;;;;;;;;;;;;;;;;UCD7B;EACf;;EAEA;;UAGe;EACf;;EAEA;;EAEA;EACA;;EAEA;;;;;;;;;;iBA4Cc,oBACd,SAAS,kBACT;EAAS;IACR;;;;;;;;;;;iBA0Ba,qBACd,SAAS,kBACT,WACA;EAAS;;;;;;;iBC3FK,sBAAsB,WAAW,kBAAkB,UACjE,UAAU,eAAe,WAAW;;;;iBA0CtB,gBAAgB,sBAAsB;UAWrC;;EAEf,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;;iBAK5D,kBAAkB,WAAW,kBAAkB,UAC7D,UAAU,eAAe,WAAW,aACnC;;;;;;;;;;;;;;;;;;UC5Dc;;WAEN;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf;;;;;;;iBAwBc,qBAAqB,MAAM,uBAAuB;;;;iBCnDlD,0BAA0B,2BAA2B,WAAW;;;;;;;;;;iBAyEhE,iCAAiC,SAAS;;;;iBAoC1C,4BAA4B,SAAS;;iBAgBrC,mBAAmB,SAAS;;iBAW5B,YAAY,SAAS;;iBA4BrB,kBACd,eAAe,gBACf,iBAAiB;;;UCvKF;EACf;EACA;EACA;;UAGe,qBAAqB,gBAAgB;EACpD;EACA;EACA;EACA,SACE,QAAQ,SACR,SAAS,4BACR,QAAQ;;UAGI;EACf;EACA;EACA,WAAW;;KAGD,oBAAoB,eAAe,4BAC7C,OAAO,QACP,SAAS,8BACN,qBAAqB,QAAQ;UAEjB;WACN,QAAQ;WACR;;KAGN,sBAAsB,WAAW,KACpC,iBAAiB;EAGjB;;UAGQ,qBAAqB,kBAAkB;EAC/C;EACA;EACA;EACA,aAAa,UAAU;;EAEvB,eAAe;;iBAGD,sBACd,gBAAgB,yBAChB,WACA,kBAAkB,WAAW,UAE7B,WAAW,qBAAqB,UAChC,SAAS,qBAAqB;EAC5B,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;EAC/D,WAAW,sBAAsB;IAElC,YAAY,WAAW;iBA8CV,qBACd,eAAe,yBACf,WACA,kBAAkB,WAAW,UAE7B,QAAQ,oBAAoB,SAC5B,SAAS,qBAAqB;EAC5B;EACA,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;;EAEzD;EAAc;;EACd;EAAa,UAAU,sBAAsB;KAEpD,YAAY,WAAW;;;KCxFrB,qBAAqB;KACrB,iBAAiB,SAAS;KAC1B,aAAa,gBAAgB,aAAa,MAAM,mBAAmB;UAcvD;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf,OAAO;IAAQ;IAAiB;MAAkB,QAAQ;;;EAG1D,SAAS,UAAU,UAAU,kBAAkB,QAAQ;;EAEvD,QAAQ,UAAU,WAAW;;;cAIlB,6BAA6B;WAG7B;EAFX,YACE,iBACS;;UAOI;;EAEf;;EAEA;;EAEA;;;;EAIA,MAAM;;UAobS;;EAEf;;EAEA;;EAEA;;;EAGA,YAAY;;;;;iBAoJE,mBAAmB,MAAM,4BAA4B;;;;iBA0FrD,kBACd,SAAS,aACT,uBACC;;;iBAOa,oBAAoB,SAAS,aAAa"}
|
|
@@ -1,11 +1,11 @@
|
|
|
1
1
|
import { b as CustomTokenPricing, c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
|
-
import { i as AnalystFinding, n as AnalystContext } from "./types-
|
|
3
|
-
import { p as ChatClient } from "./types-
|
|
4
|
-
import { f as ExternalOptimizerModelExecutionObservation, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, r as ExternalOptimizerCallbackLimits } from "./external-optimizer-contracts-
|
|
2
|
+
import { i as AnalystFinding, n as AnalystContext } from "./types-DN2WdT5S.js";
|
|
3
|
+
import { A as LlmThinkingMode, p as ChatClient } from "./types-gvRsyJLh.js";
|
|
4
|
+
import { f as ExternalOptimizerModelExecutionObservation, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, r as ExternalOptimizerCallbackLimits } from "./external-optimizer-contracts-CQCpyrIL.js";
|
|
5
5
|
import { o as Severity } from "./multi-layer-verifier-BUaQ4C17.js";
|
|
6
|
-
import { i as TraceAnalystDefinition } from "./default-registry-
|
|
7
|
-
import { l as ExactCapableAnalyst } from "./exact-types-
|
|
8
|
-
import {
|
|
6
|
+
import { i as TraceAnalystDefinition } from "./default-registry-BKwc8bN5.js";
|
|
7
|
+
import { l as ExactCapableAnalyst } from "./exact-types-BKOEILRP.js";
|
|
8
|
+
import { n as TraceAnalysisEngine } from "./engine-DhFir3Ys.js";
|
|
9
9
|
import { S as SupervisorRunTree, x as SupervisorRunSources } from "./types-CoPUTiXb.js";
|
|
10
10
|
//#region src/semantic-concept-judge.d.ts
|
|
11
11
|
/**
|
|
@@ -123,6 +123,42 @@ declare const SEMANTIC_CONCEPT_JUDGE_VERSION = "semantic-concept-judge-v1-2026-0
|
|
|
123
123
|
*/
|
|
124
124
|
declare function runSemanticConceptJudge(input: SemanticConceptJudgeInput, options: SemanticConceptJudgeOptions): Promise<SemanticConceptJudgeResult>;
|
|
125
125
|
//#endregion
|
|
126
|
+
//#region src/analyst/chat-trace-engine.d.ts
|
|
127
|
+
interface ChatTraceEngineOptions {
|
|
128
|
+
/**
|
|
129
|
+
* Caller-owned transport. Build it with `createChatClient`; agent-eval never
|
|
130
|
+
* receives the provider credential.
|
|
131
|
+
*/
|
|
132
|
+
chat: ChatClient;
|
|
133
|
+
/** Model for every call. Defaults to the client's `defaultModel`. */
|
|
134
|
+
model?: string;
|
|
135
|
+
/** Endpoint rates used when the transport reports no billed amount. */
|
|
136
|
+
pricing?: CustomTokenPricing;
|
|
137
|
+
/** Completion cap per turn. Default: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. */
|
|
138
|
+
maxOutputTokens?: number;
|
|
139
|
+
/** Sampling temperature. Omitted from the request when absent. */
|
|
140
|
+
temperature?: number;
|
|
141
|
+
/** Provider reasoning mode. Omitted when the provider default should apply. */
|
|
142
|
+
thinking?: LlmThinkingMode;
|
|
143
|
+
/** Per-call deadline handed to the transport. */
|
|
144
|
+
requestTimeoutMs?: number;
|
|
145
|
+
/**
|
|
146
|
+
* How the report turn asks for structured output.
|
|
147
|
+
*
|
|
148
|
+
* `json-object` sends JSON mode, which every OpenAI-compatible endpoint
|
|
149
|
+
* accepts, and relies on the finding grammar the analyst kind already puts
|
|
150
|
+
* in its instructions. `json-schema` sends the generated response schema as
|
|
151
|
+
* well, which stricter endpoints honour and older ones reject. Default:
|
|
152
|
+
* `json-object`.
|
|
153
|
+
*/
|
|
154
|
+
reportFormat?: 'json-object' | 'json-schema';
|
|
155
|
+
}
|
|
156
|
+
/**
|
|
157
|
+
* Run bounded recursive trace analysis in-process over a caller-owned chat
|
|
158
|
+
* transport. No Python, no subprocess, no loopback proxy.
|
|
159
|
+
*/
|
|
160
|
+
declare function createChatTraceEngine(options: ChatTraceEngineOptions): TraceAnalysisEngine;
|
|
161
|
+
//#endregion
|
|
126
162
|
//#region src/analyst/trace-tool-callback.d.ts
|
|
127
163
|
type TraceToolCallbackLimits = ExternalOptimizerCallbackLimits;
|
|
128
164
|
//#endregion
|
|
@@ -139,7 +175,7 @@ interface DspyRlmTraceEngineOptions {
|
|
|
139
175
|
pricing?: CustomTokenPricing;
|
|
140
176
|
/** Maximum provider spend for one investigation. Default: 1 USD. */
|
|
141
177
|
maxCostUsd?: number;
|
|
142
|
-
/** Controller response cap. Default:
|
|
178
|
+
/** Controller response cap. Default: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. */
|
|
143
179
|
maxOutputTokens?: number;
|
|
144
180
|
/**
|
|
145
181
|
* Thinking tokens one controller turn may bill on top of its completion.
|
|
@@ -439,5 +475,5 @@ declare const KNOWLEDGE_POISONING_KIND_SPEC: TraceAnalystDefinition;
|
|
|
439
475
|
*/
|
|
440
476
|
declare const DEFAULT_TRACE_ANALYST_KINDS: readonly TraceAnalystDefinition[];
|
|
441
477
|
//#endregion
|
|
442
|
-
export {
|
|
443
|
-
//# sourceMappingURL=index-
|
|
478
|
+
export { SemanticConceptJudgeInput as A, DspyRlmTraceEngineOptions as C, ConceptFinding as D, createChatTraceEngine as E, SemanticConceptJudgeResult as M, runSemanticConceptJudge as N, ConceptSpec as O, renderFindingSubject as S, ChatTraceEngineOptions as T, FindingSubject as _, FAILURE_MODE_KIND_SPEC as a, findingSubjectGrammarPromptFor as b, emitControlIntegrityFindings as c, FindingsStore as d, PersistedFinding as f, FINDING_SUBJECT_SYNTAX as g, FINDING_SUBJECT_KINDS as h, IMPROVEMENT_KIND_SPEC as i, SemanticConceptJudgeOptions as j, SEMANTIC_CONCEPT_JUDGE_VERSION as k, DiffPolicy as l, diffFindings as m, KNOWLEDGE_POISONING_KIND_SPEC as n, CONTROL_INTEGRITY_ANALYST as o, defaultIsMaterial as p, KNOWLEDGE_GAP_KIND_SPEC as r, ControlIntegrityAnalyst as s, DEFAULT_TRACE_ANALYST_KINDS as t, FindingsDiff as u, FindingSubjectKind as v, createDspyRlmTraceEngine as w, parseFindingSubject as x, KIND_EXPECTED_SUBJECTS as y };
|
|
479
|
+
//# sourceMappingURL=index-D0Db5X-4.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-D0Db5X-4.d.ts","names":[],"sources":["../src/semantic-concept-judge.ts","../src/analyst/chat-trace-engine.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts","../src/analyst/finding-subject.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/control-integrity.ts","../src/analyst/kinds/failure-mode.ts","../src/analyst/kinds/improvement.ts","../src/analyst/kinds/knowledge-gap.ts","../src/analyst/kinds/knowledge-poisoning.ts","../src/analyst/kinds/index.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAuDY;UAEK;EACf;;EAEA;;EAEA;;EAEA,aAAa;;UAGE;EACf;EACA;;EAEA;EACA;EACA,UAAU;;UAGK;;EAEf;;EAEA;;EAEA,aAAa;IAAQ;IAAc;;;EAEnC,kBAAkB;;EAElB;EACA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA;;EAEA;EACA;;;;;;;;KASU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,UAAU;EACV,aAAa;EACb;EACA,WAAW;EACX,SAAS;;;;;;EAMT,iBAAiB;;EAEjB,oBAAoB,QAAQ,OAAO;;cAKxB;;;;;;iBAkGS,wBACpB,OAAO,2BACP,SAAS,8BACR,QAAQ;;;UCjMM;;;;;EAKf,MAAM;;EAEN;;EAEA,UAAU;;EAEV;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;EAUA;;;;;;iBAkCc,sBAAsB,SAAS,yBAAyB;;;KCrG5D,0BAA0B;;;UCYrB;;EAEf,MAAM;;EAEN;;EAEA,kBAAkB,aAAa;EAC/B;;EAEA,UAAU;;EAEV;;EAEA;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,kBAAkB,QAAQ;;EAE1B;;;;;;;;;;EAUA;;EAEA,SAAS;;EAET;;;iBAIc,yBAAyB,SAAS,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KCtClE;EAEN;EAAwB;EAAc;;EACtC;EAAyB;;EACzB;EAAuB;;EACvB;EAAyB;;EAGzB;EAAuB;;EACvB;EAAe;;EACf;EAAkB;EAAc;;EAChC;EAAkB;;EAClB;EAAa;EAAgB;;EAC7B;EAAc;;EACd;EAAkB;;EAClB;EAAkB;;EAClB;EAAwB;;EACxB;EAAuB;;EACvB;EAAc;;EACd;EAAa;EAAgB;;EAC7B;EAAgB;;EAChB;EAAqB;;EACrB;EAAuB;;EAEvB;EAA4B;;EAC5B;EAA2B;;EAE3B;EAAiB;;KAEX,qBAAqB;cAEpB,uBAAuB,cAAc;;;;;;;;;;;;;;;;;iBAgGlC,oBAAoB,iCAAiC;;;;;;;iBAoHrD,qBAAqB,GAAG;;;;;;;;;;cA8D3B,wBAAwB,SAAS,OAAO;;;;;;;;;cAoExC,wBAAwB,eAAe,cAAc;;iBAqElD,+BAA+B;;;;;;;;UCxc9B,yBAAyB;EACxC;;cAGW;WAGiB;mBAFX;EAEjB,YAA4B;EAItB,OAAO,eAAe,UAAU,mBAAmB;;EAQzD,WAAW;;EAkBX,QAAQ,gBAAgB;;UAOT;;EAEf,UAAU;;EAEV,aAAa;;EAEb,WAAW;;;;;EAKX,SAAS;IAAQ,UAAU;IAAkB,SAAS;;;UAGvC;;;;;;;;EAQf,cAAc,UAAU,gBAAgB,SAAS;;;;;;iBAOnC,kBAAkB,GAAG,gBAAgB,GAAG;;;;;iBAWxC,aACd,UAAU,oBACV,SAAS,oBACT,SAAQ,aACP;;;;iBC7Fa,6BACd,OAAO,uBAAuB,mBAC9B,qBACC;;cA4BU,mCACA,oBAAoB,uBAAuB;WAE7C;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;;EAGI,QACJ,OAAO,uBAAuB,mBAC9B,KAAK,iBACJ,QAAQ;;cAUA,2BAAyB;;;cCHzB,wBAAwB;;;cC3BxB,uBAAuB;;;cCGvB,yBAAyB;;;cCPzB,+BAA+B;;;;;;;;;;;;cCP/B,sCAAsC"}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { a as RunRecord } from "./run-record-
|
|
2
|
-
import { s as TraceStore } from "./store-
|
|
1
|
+
import { a as RunRecord } from "./run-record-DTv1MdjK.js";
|
|
2
|
+
import { s as TraceStore } from "./store-BErPvYBr.js";
|
|
3
3
|
import { S as ToolDef, a as MintedRolloutLine, b as RolloutStep, h as RolloutOutcome, i as GatedEvidence, m as RolloutLine, t as ChatMessage, v as RolloutRole, y as RolloutSplit } from "./schema-BzWDXhOR.js";
|
|
4
4
|
import { DatabaseSync } from "node:sqlite";
|
|
5
5
|
//#region src/rollout/exporters.d.ts
|
|
@@ -931,4 +931,4 @@ declare function parseRolloutReleaseArgs(argv: string[]): RolloutReleaseCliArgs;
|
|
|
931
931
|
declare function runRolloutReleaseCli(argv: string[]): Promise<number>;
|
|
932
932
|
//#endregion
|
|
933
933
|
export { observedSplitScore as $, measureFormatGate as A, GateCheckDisposition as At, DEFAULT_CLAUDE_PROJECTS_DIR as B, SftExportOptions as Bt, FORMAT_GATE_DISPOSITION as C, relabelImportedSplit as Ct, ReleaseRowRef as D, GATE_CHECK_IDS as Dt, GateReport as E, GATE_CHECKS as Et, openOpencodeDb as F, gateErrors as Ft, MintRolloutResult as G, toRewardRows as Gt, findClaudeTranscripts as H, VerifiersRolloutOutput as Ht, readOpencodeSessionMessages as I, gatedEvidenceOf as It, unmintableReasons as J, toSftRows as Jt, RolloutScrubber as K, toRftItem as Kt, ClaudeTranscript as L, RealnessLabels as Lt, DEFAULT_OPENCODE_DB as M, GateCheckedOutcome as Mt, OpencodeSessionRow as N, GateEntryPoint as Nt, assertGateReport as O, GATE_POLICIES as Ot, findOpencodeSessionsByDirectory as P, GatePolicy as Pt, observedScore as Q, ClaudeTranscriptRef as R, RewardRow as Rt, EmittedEvidence as S, fromHarborTrajectory as St, GateDisposition as T, toHarborTrajectory as Tt, readClaudeTranscript as U, VerifiersTokenUsage as Ut, claudeProjectSlug as V, SftRow as Vt, MintRolloutOptions as W, toJsonl as Wt, ScorePreference as X, toVerifiersRolloutOutputs as Xt, ScoreOrigin as Y, toVerifiersRolloutOutput as Yt, isRealnessGated as Z, defaultRolloutScrubber as _, HarborStep as _t, buildHfDataset as a, readRolloutLedger as at, scrubRolloutLine as b, HarborToolCall as bt, runRolloutReleaseCli as c, FromHarborOptions as ct, RELEASE_FORMATS as d, HarborContentPart as dt, scoreOrigin as et, ReleaseFormat as f, HarborFinalMetrics as ft, addScrubCounts as g, HarborObservationResult as gt, ScrubRule as h, HarborObservation as ht, ScrubReport as i, readRolloutJournal as it, releaseRowRefs as j, GateCheckId as jt, gatedRolloutIds as k, GateCheck as kt, DatasetCardInputs as l, HARBOR_IMPORT_GAP as lt, ScrubCounts as m, HarborMetrics as mt, BuildSummary as n, trainingScore as nt, parseRolloutReleaseArgs as o, writeRolloutLedger as ot, buildDatasetCard as p, HarborImageSource as pt, mintRolloutRows as q, toRftItems as qt, RolloutReleaseCliArgs as r, appendRolloutLines as rt, planPushCommand as s, ATIF_SCHEMA_VERSION as st, BuildOptions as t, trainingReward as tt, FORMAT_FILES as u, HarborAgent as ut, emptyScrubCounts as v, HarborStepSource as vt, FormatGateCounts as w, toHarborTrajectories as wt, scrubText as x, HarborTrajectory as xt, scrubLines as y, HarborSubagentTrajectoryRef as yt, ClaudeUsageTotals as z, RftItem as zt };
|
|
934
|
-
//# sourceMappingURL=index-
|
|
934
|
+
//# sourceMappingURL=index-e7LXeRVa.d.ts.map
|