@tangle-network/agent-eval 0.146.0 → 0.148.0
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 +36 -0
- package/README.md +7 -0
- package/dist/analyst/index.d.ts +13 -13
- package/dist/analyst/index.js +6 -6
- package/dist/{backend-integrity-HVmWyOaX.d.ts → backend-integrity-Bz8nSrRE.d.ts} +2 -2
- package/dist/{backend-integrity-HVmWyOaX.d.ts.map → backend-integrity-Bz8nSrRE.d.ts.map} +1 -1
- package/dist/{benchmark-DTQ1RApl.d.ts → benchmark-Cn7ZMFNW.d.ts} +3 -3
- package/dist/{benchmark-DTQ1RApl.d.ts.map → benchmark-Cn7ZMFNW.d.ts.map} +1 -1
- package/dist/{benchmark-command-CxLJljS1.js → benchmark-command-Cz4fDmsk.js} +11 -11
- package/dist/{benchmark-command-CxLJljS1.js.map → benchmark-command-Cz4fDmsk.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +4 -4
- package/dist/benchmarks/index.js +3 -3
- package/dist/campaign/index.d.ts +9 -9
- package/dist/campaign/index.js +6 -6
- package/dist/{campaign-CbQXcARJ.js → campaign-BcfXzmPM.js} +7 -7
- package/dist/{campaign-CbQXcARJ.js.map → campaign-BcfXzmPM.js.map} +1 -1
- package/dist/{capture-fetch-BZiO2cEH.d.ts → capture-fetch-BAx3Gntt.d.ts} +2 -2
- package/dist/{capture-fetch-BZiO2cEH.d.ts.map → capture-fetch-BAx3Gntt.d.ts.map} +1 -1
- package/dist/{chat-client-BJmwfnjN.js → chat-client-BX8Wh3fn.js} +3 -3
- package/dist/{chat-client-BJmwfnjN.js.map → chat-client-BX8Wh3fn.js.map} +1 -1
- package/dist/cli.js +2 -2
- package/dist/{client-YDkU5QST.d.ts → client-BV7icIfy.d.ts} +2 -2
- package/dist/{client-YDkU5QST.d.ts.map → client-BV7icIfy.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +10 -10
- package/dist/contract/index.js +9 -9
- package/dist/{cost-ledger-BSe92yAV.js → cost-ledger-B1qx30B4.js} +2 -2
- package/dist/{cost-ledger-BSe92yAV.js.map → cost-ledger-B1qx30B4.js.map} +1 -1
- package/dist/{default-registry-DONui3mA.d.ts → default-registry-BUtRMRKk.d.ts} +6 -6
- package/dist/{default-registry-DONui3mA.d.ts.map → default-registry-BUtRMRKk.d.ts.map} +1 -1
- package/dist/{define-agent-eval-BGxmy4W6.js → define-agent-eval-CvZQW4u9.js} +3 -3
- package/dist/{define-agent-eval-BGxmy4W6.js.map → define-agent-eval-CvZQW4u9.js.map} +1 -1
- package/dist/{define-agent-eval-DrpgTVPp.d.ts → define-agent-eval-hXLUBKtb.d.ts} +5 -5
- package/dist/{define-agent-eval-DrpgTVPp.d.ts.map → define-agent-eval-hXLUBKtb.d.ts.map} +1 -1
- package/dist/{dspy-rlm-engine-B2nst4NR.js → dspy-rlm-engine-BmtOl_kP.js} +4 -4
- package/dist/{dspy-rlm-engine-B2nst4NR.js.map → dspy-rlm-engine-BmtOl_kP.js.map} +1 -1
- package/dist/{engine-Bc3seRPe.d.ts → engine-CIy18RTX.d.ts} +4 -4
- package/dist/{engine-Bc3seRPe.d.ts.map → engine-CIy18RTX.d.ts.map} +1 -1
- package/dist/{eval-campaign-DeGLwACc.js → eval-campaign-bmZ6NIIP.js} +2 -2
- package/dist/{eval-campaign-DeGLwACc.js.map → eval-campaign-bmZ6NIIP.js.map} +1 -1
- package/dist/{exact-types-CYFbvA5U.d.ts → exact-types-rdKFzEnK.d.ts} +2 -2
- package/dist/{exact-types-CYFbvA5U.d.ts.map → exact-types-rdKFzEnK.d.ts.map} +1 -1
- package/dist/experiment/index.d.ts +121 -5
- package/dist/experiment/index.d.ts.map +1 -1
- package/dist/experiment/index.js +219 -3
- package/dist/experiment/index.js.map +1 -1
- package/dist/{experiment-tracker-DWHZBAYL.d.ts → experiment-tracker-D9VHwRm-.d.ts} +6 -24
- package/dist/experiment-tracker-D9VHwRm-.d.ts.map +1 -0
- package/dist/{experiment-tracker-C29gXM4B.js → experiment-tracker-Ym6rEQT1.js} +15 -4
- package/dist/experiment-tracker-Ym6rEQT1.js.map +1 -0
- package/dist/{external-optimizer-contracts-CKEY98bK.d.ts → external-optimizer-contracts-CSjDLLmr.d.ts} +2 -2
- package/dist/{external-optimizer-contracts-CKEY98bK.d.ts.map → external-optimizer-contracts-CSjDLLmr.d.ts.map} +1 -1
- package/dist/{external-optimizer-process-DYNizz8W.js → external-optimizer-process-C4AyCcQd.js} +2 -2
- package/dist/{external-optimizer-process-DYNizz8W.js.map → external-optimizer-process-C4AyCcQd.js.map} +1 -1
- package/dist/{external-optimizer-subprocess-r6dP9DAj.js → external-optimizer-subprocess-DU1KM8yV.js} +3 -3
- package/dist/{external-optimizer-subprocess-r6dP9DAj.js.map → external-optimizer-subprocess-DU1KM8yV.js.map} +1 -1
- package/dist/{feedback-trajectory-Cg9EDxKB.d.ts → feedback-trajectory-niGeR6uJ.d.ts} +3 -3
- package/dist/{feedback-trajectory-Cg9EDxKB.d.ts.map → feedback-trajectory-niGeR6uJ.d.ts.map} +1 -1
- package/dist/fuzz.js +1 -1
- package/dist/hosted/index.d.ts +2 -2
- package/dist/index-CKblBxtr.d.ts +1 -0
- package/dist/{index-DGrcBMbC.d.ts → index-Dzn8Q3C2.d.ts} +18 -364
- package/dist/index-Dzn8Q3C2.d.ts.map +1 -0
- package/dist/{index-CevPFhDT.d.ts → index-gtjtfLEJ.d.ts} +7 -7
- package/dist/{index-CevPFhDT.d.ts.map → index-gtjtfLEJ.d.ts.map} +1 -1
- package/dist/index.d.ts +21 -51
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +16 -125
- package/dist/index.js.map +1 -1
- package/dist/{integrity-DDoeWibF.d.ts → integrity-CGfpTE5-.d.ts} +2 -2
- package/dist/{integrity-DDoeWibF.d.ts.map → integrity-CGfpTE5-.d.ts.map} +1 -1
- package/dist/{kind-factory-CPmSd58s.js → kind-factory-DmAa0h3K.js} +7 -7
- package/dist/kind-factory-DmAa0h3K.js.map +1 -0
- package/dist/{llm-client-_UE4fU7Z.js → llm-client-BMuxYoZy.js} +2 -2
- package/dist/{llm-client-_UE4fU7Z.js.map → llm-client-BMuxYoZy.js.map} +1 -1
- package/dist/{llm-judge-B7ffRF8w.js → llm-judge-DWq1Ptco.js} +262 -8
- package/dist/llm-judge-DWq1Ptco.js.map +1 -0
- package/dist/{matrix-su7mIfbB.d.ts → matrix-Bc111FKv.d.ts} +18 -2
- package/dist/{matrix-su7mIfbB.d.ts.map → matrix-Bc111FKv.d.ts.map} +1 -1
- package/dist/meta-eval/index.d.ts +1 -1
- package/dist/{metrics-Cl0L1KUy.js → metrics-Qv-cpptD.js} +15 -8
- package/dist/metrics-Qv-cpptD.js.map +1 -0
- package/dist/multishot/golden/index.d.ts +1 -1
- package/dist/multishot/index.d.ts +2 -2
- package/dist/multishot/index.js +17 -1
- package/dist/multishot/index.js.map +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{produced-state-D3k45S9a.js → produced-state-C0oJ4vr-.js} +4 -4
- package/dist/{produced-state-D3k45S9a.js.map → produced-state-C0oJ4vr-.js.map} +1 -1
- package/dist/{promotion-policy-BKi5OK5l.d.ts → promotion-policy-CewdjfCJ.d.ts} +2 -2
- package/dist/{promotion-policy-BKi5OK5l.d.ts.map → promotion-policy-CewdjfCJ.d.ts.map} +1 -1
- package/dist/{provenance-Bb7Zmyj2.d.ts → provenance-DA-Pmyfv.d.ts} +446 -7
- package/dist/provenance-DA-Pmyfv.d.ts.map +1 -0
- package/dist/{registry-CZgEmVXb.d.ts → registry-CaqJIDdN.d.ts} +4 -4
- package/dist/{registry-CZgEmVXb.d.ts.map → registry-CaqJIDdN.d.ts.map} +1 -1
- package/dist/reporting.d.ts +1 -1
- package/dist/reporting.js +1 -1
- package/dist/{researcher-CC305Ed-.d.ts → researcher-BySlpla_.d.ts} +3 -3
- package/dist/{researcher-CC305Ed-.d.ts.map → researcher-BySlpla_.d.ts.map} +1 -1
- package/dist/rl.d.ts +3 -3
- package/dist/rl.js +2 -2
- package/dist/{semantic-concept-judge-BsY2Q0Oq.js → semantic-concept-judge-BI7Rrl5-.js} +3 -3
- package/dist/{semantic-concept-judge-BsY2Q0Oq.js.map → semantic-concept-judge-BI7Rrl5-.js.map} +1 -1
- package/dist/{sequential-CYwq6Ff_.d.ts → sequential-BhsrMupG.d.ts} +31 -2
- package/dist/sequential-BhsrMupG.d.ts.map +1 -0
- package/dist/{sequential-Br0mAPHA.js → sequential-CzK5DarL.js} +48 -6
- package/dist/sequential-CzK5DarL.js.map +1 -0
- package/dist/series-convergence-CjO2QdRW.js.map +1 -1
- package/dist/{series-convergence-C9G-GNYK.d.ts → series-convergence-D2fsoJ2w.d.ts} +4 -7
- package/dist/{series-convergence-C9G-GNYK.d.ts.map → series-convergence-D2fsoJ2w.d.ts.map} +1 -1
- package/dist/{server-cQXve8i4.js → server-D9wQclzG.js} +3 -3
- package/dist/{server-cQXve8i4.js.map → server-D9wQclzG.js.map} +1 -1
- package/dist/{skillopt-optimization-method-C9MrMgW9.d.ts → skillopt-optimization-method-CrY0OG17.d.ts} +5 -5
- package/dist/{skillopt-optimization-method-C9MrMgW9.d.ts.map → skillopt-optimization-method-CrY0OG17.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-Cv3K4DaS.js → skillopt-optimization-method-Donmq4sq.js} +4 -4
- package/dist/{skillopt-optimization-method-Cv3K4DaS.js.map → skillopt-optimization-method-Donmq4sq.js.map} +1 -1
- package/dist/{statistical-heldout-fvxtHnKQ.d.ts → statistical-heldout-uxSpFEjm.d.ts} +2 -2
- package/dist/{statistical-heldout-fvxtHnKQ.d.ts.map → statistical-heldout-uxSpFEjm.d.ts.map} +1 -1
- package/dist/{store-otlp-CsptLYpN.js → store-otlp-CDYWW_8N.js} +2 -2
- package/dist/{store-otlp-CsptLYpN.js.map → store-otlp-CDYWW_8N.js.map} +1 -1
- package/dist/{store-tool-spans-Cq9mFd-q.js → store-tool-spans-CykkbOlv.js} +3 -3
- package/dist/{store-tool-spans-Cq9mFd-q.js.map → store-tool-spans-CykkbOlv.js.map} +1 -1
- package/dist/{store-tool-spans-CV0hRsUp.d.ts → store-tool-spans-D5FhM0_A.d.ts} +4 -4
- package/dist/{store-tool-spans-CV0hRsUp.d.ts.map → store-tool-spans-D5FhM0_A.d.ts.map} +1 -1
- package/dist/{task-failure-attributes-CpQ4y5RD.js → task-failure-attributes--ZTP3tYO.js} +2 -2
- package/dist/{task-failure-attributes-CpQ4y5RD.js.map → task-failure-attributes--ZTP3tYO.js.map} +1 -1
- package/dist/{tool-groups-CDDXNKhd.d.ts → tool-groups-zpufabP8.d.ts} +3 -3
- package/dist/tool-groups-zpufabP8.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +1 -1
- package/dist/traces.d.ts +7 -7
- package/dist/traces.js +4 -4
- package/dist/{types-oAsa-fs4.d.ts → types-BohHewKK.d.ts} +2 -2
- package/dist/{types-oAsa-fs4.d.ts.map → types-BohHewKK.d.ts.map} +1 -1
- package/dist/{types-DABZgDGV.d.ts → types-C1Bmeb8X.d.ts} +3 -3
- package/dist/{types-DABZgDGV.d.ts.map → types-C1Bmeb8X.d.ts.map} +1 -1
- package/dist/{types-CWcPguwd.d.ts → types-ColZlZtc.d.ts} +2 -2
- package/dist/{types-CWcPguwd.d.ts.map → types-ColZlZtc.d.ts.map} +1 -1
- package/dist/{types-XBRdmbj8.d.ts → types-DOhGq4S1.d.ts} +1 -7
- package/dist/{types-XBRdmbj8.d.ts.map → types-DOhGq4S1.d.ts.map} +1 -1
- package/dist/wire/index.d.ts +2 -2
- package/dist/wire/index.js +1 -1
- package/docs/multishot-golden-records.md +12 -1
- package/docs/search-history-receipts.md +139 -0
- package/package.json +3 -3
- package/dist/experiment-tracker-C29gXM4B.js.map +0 -1
- package/dist/experiment-tracker-DWHZBAYL.d.ts.map +0 -1
- package/dist/index-BgHYx1QG.d.ts +0 -1
- package/dist/index-DGrcBMbC.d.ts.map +0 -1
- package/dist/kind-factory-CPmSd58s.js.map +0 -1
- package/dist/llm-judge-B7ffRF8w.js.map +0 -1
- package/dist/metrics-Cl0L1KUy.js.map +0 -1
- package/dist/provenance-Bb7Zmyj2.d.ts.map +0 -1
- package/dist/sequential-Br0mAPHA.js.map +0 -1
- package/dist/sequential-CYwq6Ff_.d.ts.map +0 -1
- package/dist/tool-groups-CDDXNKhd.d.ts.map +0 -1
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
|
|
2
|
-
import { w as JudgeScore } from "./types-
|
|
2
|
+
import { w as JudgeScore } from "./types-C1Bmeb8X.js";
|
|
3
3
|
import { o as MatrixResult } from "./index-D_P7Ye43.js";
|
|
4
4
|
import { AgentProfile } from "@tangle-network/agent-interface";
|
|
5
5
|
//#region src/multishot/judges.d.ts
|
|
@@ -233,6 +233,22 @@ interface RunMultishotOptions<TPersona extends MultishotPersona> {
|
|
|
233
233
|
* ceiling, the judge slots, the cell composite, and the per-cell writers — so
|
|
234
234
|
* swapping the engine needs no copy of the cell body. */
|
|
235
235
|
type MultishotShot<TPersona extends MultishotPersona> = (opts: RunMultishotOptions<TPersona>) => Promise<MultishotResult>;
|
|
236
|
+
/** The package's reference conversation engine, and the referent of three
|
|
237
|
+
* things that stop working without it.
|
|
238
|
+
*
|
|
239
|
+
* 1. `runMultishotMatrix` runs it as the per-cell engine when the caller names
|
|
240
|
+
* no `runShot`, so the matrix always has an engine to run.
|
|
241
|
+
* 2. `golden.test.ts` proves the frozen records against it on every run, and
|
|
242
|
+
* perturbs it to prove the records detect a change. Records checked only
|
|
243
|
+
* against a replay of themselves prove nothing.
|
|
244
|
+
* 3. The repo's recording command names it as `--engine`. The script defaults
|
|
245
|
+
* no engine, so a version records whichever one the caller names — and a
|
|
246
|
+
* record set minted from a consumer's engine would let a product's wiring
|
|
247
|
+
* define this package's frozen contract.
|
|
248
|
+
*
|
|
249
|
+
* An alternative engine reaches the matrix through `runShot` and reaches the
|
|
250
|
+
* records through `assertMultishotGoldenScenario`. Neither seam removes the
|
|
251
|
+
* three roles above. */
|
|
236
252
|
declare function runMultishot<TPersona extends MultishotPersona>(opts: RunMultishotOptions<TPersona>): Promise<MultishotResult>;
|
|
237
253
|
//#endregion
|
|
238
254
|
//#region src/multishot/matrix.d.ts
|
|
@@ -380,4 +396,4 @@ interface RunMultishotMatrixResult {
|
|
|
380
396
|
declare function runMultishotMatrix<TPersona extends MultishotPersona>(opts: RunMultishotMatrixOptions<TPersona>): Promise<RunMultishotMatrixResult>;
|
|
381
397
|
//#endregion
|
|
382
398
|
export { JudgeConfig as A, MultishotToolExecutor as C, MultishotTransportToolCall as D, MultishotTransportResponse as E, runJudge as F, JudgeRunResult as M, renderDimensions as N, assertMultishotShotResult as O, renderJsonFooter as P, MultishotToolDefinition as S, MultishotTransportRequest as T, MultishotMessage as _, MultishotCellOutput as a, MultishotShape as b, RunMultishotMatrixResult as c, MultishotShot as d, RunMultishotOptions as f, MultishotFatalToolError as g, MultishotDriverEmptyError as h, ConversationJudgeInput as i, JudgeDimension as j, DEFAULT_JUDGE_MODEL as k, computeCellComposite as l, MultishotArtifact as m, CellCompositeInput as n, MultishotJudges as o, runMultishot as p, CellCompositeScore as r, RunMultishotMatrixOptions as s, ArtifactJudgeInput as t, runMultishotMatrix as u, MultishotPersona as v, MultishotTransport as w, MultishotShotResultError as x, MultishotResult as y };
|
|
383
|
-
//# sourceMappingURL=matrix-
|
|
399
|
+
//# sourceMappingURL=matrix-Bc111FKv.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"matrix-
|
|
1
|
+
{"version":3,"file":"matrix-Bc111FKv.d.ts","names":[],"sources":["../src/multishot/judges.ts","../src/multishot/types.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;cAyBa;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;EAEA;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;EACA;;EAEA;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAkIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UClM/B;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;EAAO;EAAgB;EAAiB,SAAS;MAC9C;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;UCpH3D,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;EAMA,iBAAiB;;;;EAIjB,kBAAkB;EAClB;EACA;EACA,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UCtFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,iBAAiB;;EAEjB,kBAAkB;;;;;;;;;;;;;;;EAelB,UAAU,cAAc;;EAExB;EACA;;;;;UAMe;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { f as Run } from "../schema-BtVldJ3T.js";
|
|
2
2
|
import { s as TraceStore } from "../store-CT9YIIve.js";
|
|
3
3
|
import { a as ContinuousCalibrationResult, n as CandidateScore, o as GoldenItem, r as ContinuousAgreement, t as CalibrationResult } from "../judge-calibration-C5CbMYce.js";
|
|
4
|
-
import { n as SeriesConvergenceResult, o as CorpusAgreementReport, t as SeriesConvergenceOptions } from "../series-convergence-
|
|
4
|
+
import { n as SeriesConvergenceResult, o as CorpusAgreementReport, t as SeriesConvergenceOptions } from "../series-convergence-D2fsoJ2w.js";
|
|
5
5
|
import { a as OutcomeFilter, i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
|
|
6
6
|
import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "../rubric-predictive-validity-C7LnNvF2.js";
|
|
7
7
|
//#region src/meta-eval/correlation-study.d.ts
|
|
@@ -24,6 +24,10 @@ const MODEL_PRICING = {
|
|
|
24
24
|
"claude-3-haiku-20240307": {
|
|
25
25
|
input: 25e-5,
|
|
26
26
|
output: .00125
|
|
27
|
+
},
|
|
28
|
+
"glm-5.3": {
|
|
29
|
+
input: .00168,
|
|
30
|
+
output: .00528
|
|
27
31
|
}
|
|
28
32
|
};
|
|
29
33
|
/** Family-level pricing fallbacks (per-1K), matched against a normalized id
|
|
@@ -88,19 +92,22 @@ const FAMILY_PRICING = [
|
|
|
88
92
|
}]
|
|
89
93
|
];
|
|
90
94
|
/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,
|
|
91
|
-
* and
|
|
92
|
-
*
|
|
93
|
-
|
|
94
|
-
|
|
95
|
+
* and return both the complete harness/provider-prefixed id and its final
|
|
96
|
+
* served-model segment (`opencode/zai-coding-plan/glm-5.3` → full id plus
|
|
97
|
+
* `glm-5.3`). Exact rows are checked against both before family fallback. */
|
|
98
|
+
function normalizedModelIds(model) {
|
|
99
|
+
const id = (model.split("@")[0] ?? model).trim().toLowerCase();
|
|
100
|
+
const finalSegment = id.slice(id.lastIndexOf("/") + 1);
|
|
101
|
+
return finalSegment === id ? [id] : [id, finalSegment];
|
|
95
102
|
}
|
|
96
103
|
/** Resolve pricing for a model id: exact table, then family fallback.
|
|
97
104
|
* Returns null when the id matches nothing (caller decides — never a
|
|
98
105
|
* silent-zero masquerading as a real $0 cost). */
|
|
99
106
|
function resolveModelPricing(model) {
|
|
100
107
|
if (MODEL_PRICING[model]) return MODEL_PRICING[model];
|
|
101
|
-
const
|
|
102
|
-
if (MODEL_PRICING[id]) return MODEL_PRICING[id];
|
|
103
|
-
for (const [pattern, price] of FAMILY_PRICING) if (pattern.test(id)) return price;
|
|
108
|
+
const ids = normalizedModelIds(model);
|
|
109
|
+
for (const id of ids) if (MODEL_PRICING[id]) return MODEL_PRICING[id];
|
|
110
|
+
for (const [pattern, price] of FAMILY_PRICING) if (ids.some((id) => pattern.test(id))) return price;
|
|
104
111
|
return null;
|
|
105
112
|
}
|
|
106
113
|
/** True when `model` has known pricing (exact or family). Lets cost-aware
|
|
@@ -130,4 +137,4 @@ function estimateCost(inputTokens, outputTokens, model) {
|
|
|
130
137
|
//#endregion
|
|
131
138
|
export { resolveModelPricing as a, isModelPriced as i, estimateCost as n, estimateTokens as r, MODEL_PRICING as t };
|
|
132
139
|
|
|
133
|
-
//# sourceMappingURL=metrics-
|
|
140
|
+
//# sourceMappingURL=metrics-Qv-cpptD.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"metrics-Qv-cpptD.js","names":[],"sources":["../src/metrics.ts"],"sourcesContent":["interface TokenPrice {\n input: number\n output: number\n}\n\n/** Per-1K token pricing for exact model ids. */\nexport const MODEL_PRICING: Record<string, TokenPrice> = {\n 'gpt-4o': { input: 0.0025, output: 0.01 },\n 'gpt-4o-mini': { input: 0.00015, output: 0.0006 },\n 'gpt-4-turbo': { input: 0.01, output: 0.03 },\n 'claude-sonnet-4-20250514': { input: 0.003, output: 0.015 },\n 'claude-opus-4-20250514': { input: 0.015, output: 0.075 },\n 'claude-3-haiku-20240307': { input: 0.00025, output: 0.00125 },\n // Router response-header rates observed for the exact served model on\n // 2026-08-15. Keep this exact row ahead of the deliberately coarse family\n // fallback so glm-5.3 campaign costs are not understated by ~2.5x.\n 'glm-5.3': { input: 0.00168, output: 0.00528 },\n}\n\n/** Family-level pricing fallbacks (per-1K), matched against a normalized id\n * after exact lookup misses. Ordered — first match wins. Covers the model\n * ids actually used through the Tangle router + cli-bridge harnesses\n * (`claude-code/sonnet`, `opencode/zai-coding-plan/glm-5.1`,\n * `kimi-code/kimi-k2.6`, `deepseek-v4-pro`, `anthropic/claude-sonnet-4-6`, …),\n * none of which appear in the exact table above — without this they priced\n * to a silent $0, blanking every cost/Pareto axis downstream. */\nconst FAMILY_PRICING: Array<[RegExp, TokenPrice]> = [\n [/claude.*opus/, { input: 0.015, output: 0.075 }],\n [/claude.*haiku/, { input: 0.0008, output: 0.004 }],\n [/claude.*sonnet|claude-code|claude-sonnet/, { input: 0.003, output: 0.015 }],\n [/gpt-4o-mini/, { input: 0.00015, output: 0.0006 }],\n [/gpt-5|gpt-4\\.1|o[134]\\b/, { input: 0.00125, output: 0.01 }],\n [/gpt-4o|gpt-4/, { input: 0.0025, output: 0.01 }],\n [/deepseek/, { input: 0.0003, output: 0.0011 }],\n [/glm|zhipu|zai/, { input: 0.0006, output: 0.0022 }],\n [/kimi|moonshot/, { input: 0.0006, output: 0.0025 }],\n [/qwen/, { input: 0.0004, output: 0.0012 }],\n [/gemini.*flash/, { input: 0.0001, output: 0.0004 }],\n [/gemini/, { input: 0.00125, output: 0.005 }],\n [/llama/, { input: 0.0002, output: 0.0006 }],\n]\n\n/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,\n * and return both the complete harness/provider-prefixed id and its final\n * served-model segment (`opencode/zai-coding-plan/glm-5.3` → full id plus\n * `glm-5.3`). Exact rows are checked against both before family fallback. */\nfunction normalizedModelIds(model: string): string[] {\n const id = (model.split('@')[0] ?? model).trim().toLowerCase()\n const finalSegment = id.slice(id.lastIndexOf('/') + 1)\n return finalSegment === id ? [id] : [id, finalSegment]\n}\n\n/** Resolve pricing for a model id: exact table, then family fallback.\n * Returns null when the id matches nothing (caller decides — never a\n * silent-zero masquerading as a real $0 cost). */\nexport function resolveModelPricing(model: string): TokenPrice | null {\n if (MODEL_PRICING[model]) return MODEL_PRICING[model]\n const ids = normalizedModelIds(model)\n for (const id of ids) {\n if (MODEL_PRICING[id]) return MODEL_PRICING[id]\n }\n for (const [pattern, price] of FAMILY_PRICING) {\n if (ids.some((id) => pattern.test(id))) return price\n }\n return null\n}\n\n/** True when `model` has known pricing (exact or family). Lets cost-aware\n * callers distinguish a real $0 from an unpriced model. */\nexport function isModelPriced(model: string): boolean {\n return resolveModelPricing(model) !== null\n}\n\nconst warnedUnpricedModels = new Set<string>()\n\n/** Estimate token count from string length (chars / 4 approximation) */\nexport function estimateTokens(text: string): number {\n return Math.ceil(text.length / 4)\n}\n\n/** Calculate cost in USD from token counts and model. Unknown models warn\n * once (not a silent zero) and return 0 so callers that ignore pricing keep\n * working; cost-sensitive callers should gate on {@link isModelPriced}. */\nexport function estimateCost(inputTokens: number, outputTokens: number, model: string): number {\n const pricing = resolveModelPricing(model)\n if (!pricing) {\n if (!warnedUnpricedModels.has(model)) {\n warnedUnpricedModels.add(model)\n console.warn(\n `estimateCost: no pricing for model \"${model}\" — returning 0; add it to ` +\n 'MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)',\n )\n }\n return 0\n }\n return (inputTokens / 1000) * pricing.input + (outputTokens / 1000) * pricing.output\n}\n\n/**\n * TokenCounter — accumulates token usage and cost across turns.\n */\nexport class TokenCounter {\n private totalInput = 0\n private totalOutput = 0\n private totalCost = 0\n private model: string\n\n constructor(model = 'gpt-4o') {\n this.model = model\n }\n\n /** Record tokens for a turn, returns per-turn cost */\n record(inputTokens: number, outputTokens: number): number {\n this.totalInput += inputTokens\n this.totalOutput += outputTokens\n const cost = estimateCost(inputTokens, outputTokens, this.model)\n this.totalCost += cost\n return cost\n }\n\n /** Estimate and record from raw text */\n recordFromText(\n inputText: string,\n outputText: string,\n ): { inputTokens: number; outputTokens: number; cost: number } {\n const inputTokens = estimateTokens(inputText)\n const outputTokens = estimateTokens(outputText)\n const cost = this.record(inputTokens, outputTokens)\n return { inputTokens, outputTokens, cost }\n }\n\n getTotalInput(): number {\n return this.totalInput\n }\n getTotalOutput(): number {\n return this.totalOutput\n }\n getTotalCost(): number {\n return this.totalCost\n }\n}\n"],"mappings":";;AAMA,MAAa,gBAA4C;CACvD,UAAU;EAAE,OAAO;EAAQ,QAAQ;CAAK;CACxC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO;CAChD,eAAe;EAAE,OAAO;EAAM,QAAQ;CAAK;CAC3C,4BAA4B;EAAE,OAAO;EAAO,QAAQ;CAAM;CAC1D,0BAA0B;EAAE,OAAO;EAAO,QAAQ;CAAM;CACxD,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAQ;CAI7D,WAAW;EAAE,OAAO;EAAS,QAAQ;CAAQ;AAC/C;;;;;;;;AASA,MAAM,iBAA8C;CAClD,CAAC,gBAAgB;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAChD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAM,CAAC;CAClD,CAAC,4CAA4C;EAAE,OAAO;EAAO,QAAQ;CAAM,CAAC;CAC5E,CAAC,eAAe;EAAE,OAAO;EAAS,QAAQ;CAAO,CAAC;CAClD,CAAC,2BAA2B;EAAE,OAAO;EAAS,QAAQ;CAAK,CAAC;CAC5D,CAAC,gBAAgB;EAAE,OAAO;EAAQ,QAAQ;CAAK,CAAC;CAChD,CAAC,YAAY;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC9C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,QAAQ;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CAC1C,CAAC,iBAAiB;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;CACnD,CAAC,UAAU;EAAE,OAAO;EAAS,QAAQ;CAAM,CAAC;CAC5C,CAAC,SAAS;EAAE,OAAO;EAAQ,QAAQ;CAAO,CAAC;AAC7C;;;;;AAMA,SAAS,mBAAmB,OAAyB;CACnD,MAAM,MAAM,MAAM,MAAM,GAAG,CAAC,CAAC,MAAM,MAAA,CAAO,KAAK,CAAC,CAAC,YAAY;CAC7D,MAAM,eAAe,GAAG,MAAM,GAAG,YAAY,GAAG,IAAI,CAAC;CACrD,OAAO,iBAAiB,KAAK,CAAC,EAAE,IAAI,CAAC,IAAI,YAAY;AACvD;;;;AAKA,SAAgB,oBAAoB,OAAkC;CACpE,IAAI,cAAc,QAAQ,OAAO,cAAc;CAC/C,MAAM,MAAM,mBAAmB,KAAK;CACpC,KAAK,MAAM,MAAM,KACf,IAAI,cAAc,KAAK,OAAO,cAAc;CAE9C,KAAK,MAAM,CAAC,SAAS,UAAU,gBAC7B,IAAI,IAAI,MAAM,OAAO,QAAQ,KAAK,EAAE,CAAC,GAAG,OAAO;CAEjD,OAAO;AACT;;;AAIA,SAAgB,cAAc,OAAwB;CACpD,OAAO,oBAAoB,KAAK,MAAM;AACxC;AAEA,MAAM,uCAAuB,IAAI,IAAY;;AAG7C,SAAgB,eAAe,MAAsB;CACnD,OAAO,KAAK,KAAK,KAAK,SAAS,CAAC;AAClC;;;;AAKA,SAAgB,aAAa,aAAqB,cAAsB,OAAuB;CAC7F,MAAM,UAAU,oBAAoB,KAAK;CACzC,IAAI,CAAC,SAAS;EACZ,IAAI,CAAC,qBAAqB,IAAI,KAAK,GAAG;GACpC,qBAAqB,IAAI,KAAK;GAC9B,QAAQ,KACN,uCAAuC,MAAM,oGAE/C;EACF;EACA,OAAO;CACT;CACA,OAAQ,cAAc,MAAQ,QAAQ,QAAS,eAAe,MAAQ,QAAQ;AAChF"}
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { S as MultishotToolDefinition, T as MultishotTransportRequest, c as RunMultishotMatrixResult, f as RunMultishotOptions, s as RunMultishotMatrixOptions, v as MultishotPersona, y as MultishotResult } from "../../matrix-
|
|
1
|
+
import { S as MultishotToolDefinition, T as MultishotTransportRequest, c as RunMultishotMatrixResult, f as RunMultishotOptions, s as RunMultishotMatrixOptions, v as MultishotPersona, y as MultishotResult } from "../../matrix-Bc111FKv.js";
|
|
2
2
|
//#region src/multishot/golden/compare.d.ts
|
|
3
3
|
interface CompareOptions {
|
|
4
4
|
/** Stop after this many mismatches. A structural divergence high in the
|
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { w as JudgeScore } from "../types-
|
|
2
|
-
import { A as JudgeConfig, C as MultishotToolExecutor, D as MultishotTransportToolCall, E as MultishotTransportResponse, F as runJudge, M as JudgeRunResult, N as renderDimensions, O as assertMultishotShotResult, P as renderJsonFooter, S as MultishotToolDefinition, T as MultishotTransportRequest, _ as MultishotMessage, a as MultishotCellOutput, b as MultishotShape, c as RunMultishotMatrixResult, d as MultishotShot, f as RunMultishotOptions, g as MultishotFatalToolError, h as MultishotDriverEmptyError, i as ConversationJudgeInput, j as JudgeDimension, k as DEFAULT_JUDGE_MODEL, l as computeCellComposite, m as MultishotArtifact, n as CellCompositeInput, o as MultishotJudges, p as runMultishot, r as CellCompositeScore, s as RunMultishotMatrixOptions, t as ArtifactJudgeInput, u as runMultishotMatrix, v as MultishotPersona, w as MultishotTransport, x as MultishotShotResultError, y as MultishotResult } from "../matrix-
|
|
1
|
+
import { w as JudgeScore } from "../types-C1Bmeb8X.js";
|
|
2
|
+
import { A as JudgeConfig, C as MultishotToolExecutor, D as MultishotTransportToolCall, E as MultishotTransportResponse, F as runJudge, M as JudgeRunResult, N as renderDimensions, O as assertMultishotShotResult, P as renderJsonFooter, S as MultishotToolDefinition, T as MultishotTransportRequest, _ as MultishotMessage, a as MultishotCellOutput, b as MultishotShape, c as RunMultishotMatrixResult, d as MultishotShot, f as RunMultishotOptions, g as MultishotFatalToolError, h as MultishotDriverEmptyError, i as ConversationJudgeInput, j as JudgeDimension, k as DEFAULT_JUDGE_MODEL, l as computeCellComposite, m as MultishotArtifact, n as CellCompositeInput, o as MultishotJudges, p as runMultishot, r as CellCompositeScore, s as RunMultishotMatrixOptions, t as ArtifactJudgeInput, u as runMultishotMatrix, v as MultishotPersona, w as MultishotTransport, x as MultishotShotResultError, y as MultishotResult } from "../matrix-Bc111FKv.js";
|
|
3
3
|
import { AgentProfile } from "@tangle-network/agent-interface";
|
|
4
4
|
//#region src/multishot/router.d.ts
|
|
5
5
|
interface RouterCompletionRequest {
|
package/dist/multishot/index.js
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { i as isModelPriced, n as estimateCost } from "../metrics-
|
|
1
|
+
import { i as isModelPriced, n as estimateCost } from "../metrics-Qv-cpptD.js";
|
|
2
2
|
import { r as withCellSpend, t as runAgentMatrix } from "../matrix-pWinRMPl.js";
|
|
3
3
|
import { i as assertMultishotShotResult, n as MultishotFatalToolError, r as MultishotShotResultError, t as MultishotDriverEmptyError } from "../types-B3jzCp0p.js";
|
|
4
4
|
import { mkdirSync, writeFileSync } from "node:fs";
|
|
@@ -366,6 +366,22 @@ function defaultShapeFromProfile(profile, shape) {
|
|
|
366
366
|
}
|
|
367
367
|
//#endregion
|
|
368
368
|
//#region src/multishot/multishot.ts
|
|
369
|
+
/** The package's reference conversation engine, and the referent of three
|
|
370
|
+
* things that stop working without it.
|
|
371
|
+
*
|
|
372
|
+
* 1. `runMultishotMatrix` runs it as the per-cell engine when the caller names
|
|
373
|
+
* no `runShot`, so the matrix always has an engine to run.
|
|
374
|
+
* 2. `golden.test.ts` proves the frozen records against it on every run, and
|
|
375
|
+
* perturbs it to prove the records detect a change. Records checked only
|
|
376
|
+
* against a replay of themselves prove nothing.
|
|
377
|
+
* 3. The repo's recording command names it as `--engine`. The script defaults
|
|
378
|
+
* no engine, so a version records whichever one the caller names — and a
|
|
379
|
+
* record set minted from a consumer's engine would let a product's wiring
|
|
380
|
+
* define this package's frozen contract.
|
|
381
|
+
*
|
|
382
|
+
* An alternative engine reaches the matrix through `runShot` and reaches the
|
|
383
|
+
* records through `assertMultishotGoldenScenario`. Neither seam removes the
|
|
384
|
+
* three roles above. */
|
|
369
385
|
async function runMultishot(opts) {
|
|
370
386
|
const meter = {
|
|
371
387
|
costUsd: 0,
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.js","names":[],"sources":["../../src/multishot/router.ts","../../src/multishot/default-tools.ts","../../src/multishot/judges.ts","../../src/multishot/shape-defaults.ts","../../src/multishot/multishot.ts","../../src/multishot/matrix.ts"],"sourcesContent":["// Router fetch helper — single source of truth for OpenAI-compat calls\n// against the Tangle router. Used by the driver, agent, judges, and the\n// default tool executors.\n\nimport type { MultishotToolDefinition } from './types'\n\nexport interface RouterCompletionRequest {\n apiKey: string\n baseUrl: string\n model: string\n messages: Array<Record<string, unknown>>\n tools?: MultishotToolDefinition[]\n temperature?: number\n maxTokens?: number\n signal?: AbortSignal\n}\n\nexport interface RouterToolCall {\n id: string\n type: 'function'\n function: { name: string; arguments: string }\n}\n\nexport interface RouterCompletionResponse {\n message: { content?: string | null; tool_calls?: RouterToolCall[] }\n usage?: { prompt_tokens?: number; completion_tokens?: number }\n /** Provider-reported spend when the endpoint supplies it. */\n costUsd?: number\n /** Model echoed by the provider, falling back to the requested model. */\n model: string\n durationMs: number\n}\n\nexport async function routerCompletion(\n req: RouterCompletionRequest,\n): Promise<RouterCompletionResponse> {\n const startedAt = Date.now()\n const body: Record<string, unknown> = {\n model: req.model,\n messages: req.messages,\n temperature: req.temperature ?? 0.7,\n max_tokens: req.maxTokens ?? 2000,\n }\n if (req.tools?.length) body.tools = req.tools\n const url = `${req.baseUrl.replace(/\\/+$/, '')}/chat/completions`\n const res = await fetch(url, {\n method: 'POST',\n headers: { Authorization: `Bearer ${req.apiKey}`, 'Content-Type': 'application/json' },\n body: JSON.stringify(body),\n signal: req.signal,\n })\n if (!res.ok) {\n const text = await res.text()\n throw new Error(`router ${res.status}: ${text.slice(0, 300)}`)\n }\n const json = (await res.json()) as {\n choices: Array<{ message: { content?: string | null; tool_calls?: RouterToolCall[] } }>\n usage?: { prompt_tokens?: number; completion_tokens?: number }\n model?: unknown\n _response_cost?: unknown\n cost_usd?: unknown\n }\n const choice = json.choices[0]\n if (!choice) throw new Error(`router returned no choices: ${JSON.stringify(json).slice(0, 200)}`)\n const rawCost = json._response_cost ?? json.cost_usd\n const costUsd =\n typeof rawCost === 'number' && Number.isFinite(rawCost) && rawCost >= 0 ? rawCost : undefined\n return {\n message: choice.message,\n usage: json.usage,\n ...(costUsd === undefined ? {} : { costUsd }),\n model: typeof json.model === 'string' && json.model ? json.model : req.model,\n durationMs: Date.now() - startedAt,\n }\n}\n\n// Rough per-model cost estimator. Used for cost-ceiling enforcement.\n// Underestimates Anthropic, overestimates oss models — fine for ceilings.\nexport function estimateRouterCost(\n model: string,\n usage?: { prompt_tokens?: number; completion_tokens?: number },\n): number {\n if (!usage) return 0\n const inputTok = usage.prompt_tokens ?? 0\n const outputTok = usage.completion_tokens ?? 0\n let inPer1k = 0.003\n let outPer1k = 0.015\n if (model.includes('gpt-4o-mini')) {\n inPer1k = 0.00015\n outPer1k = 0.0006\n } else if (model.includes('gpt-5.4') || model.includes('claude-sonnet')) {\n inPer1k = 0.003\n outPer1k = 0.015\n } else if (model.includes('kimi') || model.includes('glm') || model.includes('deepseek')) {\n inPer1k = 0.0005\n outPer1k = 0.002\n }\n return (inputTok * inPer1k + outputTok * outPer1k) / 1000\n}\n\nexport function defaultRouterBaseUrl(): string {\n return (process.env.TANGLE_ROUTER_BASE_URL ?? 'https://router.tangle.tools/v1').replace(\n /\\/+$/,\n '',\n )\n}\n\nexport function requireRouterApiKey(): string {\n const key = process.env.TANGLE_API_KEY\n if (!key) throw new Error('multishot requires TANGLE_API_KEY (router-scoped sk-tan-* key)')\n return key\n}\n","// Default delegate_research + delegate_code tools and their inline executors.\n//\n// Consumers can override either by passing their own tools + executors to\n// runMultishot. The defaults are sufficient for most domains — point the\n// researcher system prompt at your domain's citation style and the coder\n// at your preferred language.\n\nimport { estimateRouterCost, routerCompletion } from './router'\nimport type { MultishotToolDefinition, MultishotToolExecutor } from './types'\n\nexport const DEFAULT_RESEARCHER_MODEL = 'openai/gpt-4o-mini'\nexport const DEFAULT_CODER_MODEL = 'openai/gpt-4o-mini'\n\nexport interface DefaultResearcherConfig {\n /** Replace the system prompt to bias the researcher toward a domain's\n * citation style. Defaults to a generic \"cite sources by name\" prompt. */\n systemPrompt?: string\n model?: string\n}\n\nexport interface DefaultCoderConfig {\n /** Replace the system prompt to bias the coder toward a language /\n * framework / artifact style. */\n systemPrompt?: string\n model?: string\n}\n\nconst GENERIC_RESEARCHER_SYSTEM =\n 'You are a research specialist. Return a markdown brief with 3-5 findings. Each finding cites a specific source by name. Add a confidence level (high/medium/low) per finding. No fluff, no preamble.'\n\nconst GENERIC_CODER_SYSTEM =\n 'You are an expert engineer. Output ONE fenced code block containing the complete solution. Inline-comment non-obvious decisions. No explanation outside the block.'\n\nexport const DEFAULT_DELEGATE_RESEARCH_TOOL: MultishotToolDefinition = {\n type: 'function',\n function: {\n name: 'delegate_research',\n description:\n 'Research a topic deeply via specialist. Returns evidence-bearing items with citations. Use for audience research, competitive intel, regulatory landscape, market data, citation-grounded analysis.',\n parameters: {\n type: 'object',\n properties: {\n question: { type: 'string', description: 'Specific question to research' },\n scope: {\n type: 'string',\n description: 'Optional scope: time window, geography, jurisdiction, segment',\n },\n },\n required: ['question'],\n },\n },\n}\n\nexport const DEFAULT_DELEGATE_CODE_TOOL: MultishotToolDefinition = {\n type: 'function',\n function: {\n name: 'delegate_code',\n description:\n 'Generate a runnable script, template, pipeline, or tool via specialist. Returns complete working code or structured markdown. Use for content pipelines, calc snippets, dashboards, compliance checklists, deadline trackers.',\n parameters: {\n type: 'object',\n properties: {\n goal: { type: 'string', description: 'What the code must accomplish' },\n language: {\n type: 'string',\n description: 'Optional language preference (default: TypeScript)',\n },\n },\n required: ['goal'],\n },\n },\n}\n\nexport function createResearchExecutor(\n config: DefaultResearcherConfig = {},\n): MultishotToolExecutor {\n const systemPrompt = config.systemPrompt ?? GENERIC_RESEARCHER_SYSTEM\n const model = config.model ?? DEFAULT_RESEARCHER_MODEL\n return async (args, ctx) => {\n const question = String(args.question ?? '')\n const scope = args.scope ? String(args.scope) : undefined\n const { message, usage } = await routerCompletion({\n apiKey: ctx.apiKey,\n baseUrl: ctx.baseUrl,\n model,\n temperature: 0.3,\n maxTokens: 1800,\n messages: [\n { role: 'system', content: systemPrompt },\n { role: 'user', content: `Research: ${question}${scope ? `\\nScope: ${scope}` : ''}` },\n ],\n signal: ctx.signal,\n })\n return { content: message.content ?? '', costUsd: estimateRouterCost(model, usage) }\n }\n}\n\nexport function createCodeExecutor(config: DefaultCoderConfig = {}): MultishotToolExecutor {\n const systemPrompt = config.systemPrompt ?? GENERIC_CODER_SYSTEM\n const model = config.model ?? DEFAULT_CODER_MODEL\n return async (args, ctx) => {\n const goal = String(args.goal ?? '')\n const language = args.language ? String(args.language) : 'TypeScript'\n const { message, usage } = await routerCompletion({\n apiKey: ctx.apiKey,\n baseUrl: ctx.baseUrl,\n model,\n temperature: 0.2,\n maxTokens: 2000,\n messages: [\n { role: 'system', content: `${systemPrompt}\\n\\nLanguage: ${language}` },\n { role: 'user', content: `Produce: ${goal}` },\n ],\n signal: ctx.signal,\n })\n return { content: message.content ?? '', costUsd: estimateRouterCost(model, usage) }\n }\n}\n\nexport interface DefaultToolsConfig {\n research?: DefaultResearcherConfig\n code?: DefaultCoderConfig\n /** When true (default), each tool result is recorded as a typed artifact:\n * research → type='research', code → type='code'. */\n recordArtifacts?: boolean\n}\n\nexport interface DefaultToolsBundle {\n tools: MultishotToolDefinition[]\n executors: Record<string, MultishotToolExecutor>\n artifactTypeFor: (toolName: string) => string | undefined\n}\n\nexport function defaultDelegationTools(config: DefaultToolsConfig = {}): DefaultToolsBundle {\n return {\n tools: [DEFAULT_DELEGATE_RESEARCH_TOOL, DEFAULT_DELEGATE_CODE_TOOL],\n executors: {\n delegate_research: createResearchExecutor(config.research),\n delegate_code: createCodeExecutor(config.code),\n },\n artifactTypeFor: (name) =>\n name === 'delegate_research' ? 'research' : name === 'delegate_code' ? 'code' : undefined,\n }\n}\n\nexport { defaultRouterBaseUrl } from './router'\n","// Generic judge runner — domain consumers configure dimensions + prompts.\n//\n// Three judge slots are conventional for multishot eval:\n// - conversation (scores the full transcript)\n// - codeReview (scores each code artifact)\n// - contentQuality (scores each non-code artifact)\n//\n// But the runJudge primitive is fully generic: any input maps to a score plus\n// explicit observed, estimated, or uncaptured cost.\n\nimport type { JudgeScore } from '../campaign/types'\nimport type { CostProvenance } from '../cost-ledger'\nimport type { LlmCallMetadata, LlmUsage } from '../llm-client'\nimport { estimateCost, isModelPriced } from '../metrics'\nimport {\n defaultRouterBaseUrl,\n type RouterCompletionResponse,\n requireRouterApiKey,\n routerCompletion,\n} from './router'\n\n// Canonical declaration lives in campaign/types.ts. Multishot emits the same\n// shape on its producer-defined 0-10 scale.\nexport type { JudgeScore } from '../campaign/types'\n\nexport const DEFAULT_JUDGE_MODEL = 'openai/gpt-4o-mini'\n\nexport interface JudgeDimension {\n /** JSON field name + score key. */\n key: string\n /** Description shown in the judge's user prompt. */\n description: string\n}\n\nexport interface JudgeConfig<TInput> {\n /** Display name (for trace + log). */\n name: string\n /** Model used for this judge. */\n model?: string\n /** 0-10 scored dimensions. */\n dimensions: JudgeDimension[]\n /** Judge system prompt — sets persona + JSON-only constraint. */\n systemPrompt: string\n /** Build the user prompt from the typed input. Must include \"Respond with\n * ONLY this JSON: { ... }\" listing each dimension key. */\n buildPrompt: (input: TInput) => string\n /** Optional model + api overrides. */\n apiKey?: string\n baseUrl?: string\n /** Maximum output tokens for the judge response. Defaults to 1500. */\n maxTokens?: number\n}\n\nexport interface JudgeRunResult {\n /** Semantic result; failed scores remain non-throwing so matrix aggregation can exclude them. */\n score: JudgeScore\n /** Cost of the completed call, separate from diagnostic provider metadata. */\n cost: CostProvenance\n}\n\nexport async function runJudge<TInput>(\n judge: JudgeConfig<TInput>,\n input: TInput,\n): Promise<JudgeRunResult> {\n const apiKey = judge.apiKey ?? requireRouterApiKey()\n const baseUrl = judge.baseUrl ?? defaultRouterBaseUrl()\n const model = judge.model ?? process.env.JUDGE_MODEL ?? DEFAULT_JUDGE_MODEL\n const prompt = judge.buildPrompt(input)\n let raw = ''\n let llmCall: LlmCallMetadata\n let cost: CostProvenance\n const startedAt = Date.now()\n try {\n const response = await routerCompletion({\n apiKey,\n baseUrl,\n model,\n temperature: 0,\n maxTokens: judge.maxTokens ?? 1500,\n messages: [\n { role: 'system', content: judge.systemPrompt },\n { role: 'user', content: prompt },\n ],\n })\n const call = judgeCallMetadata(response)\n llmCall = call.llmCall\n cost = call.cost\n raw = (response.message.content ?? '').trim()\n } catch (err) {\n // failed:true lets consumers reading `.composite` keep working while\n // aggregators exclude this score from means instead of averaging a zero.\n return {\n score: {\n dimensions: {},\n composite: 0,\n failed: true,\n notes: `judge ${judge.name} call failed: ${err instanceof Error ? err.message : String(err)}`,\n llmCall: unknownJudgeCall(model, Date.now() - startedAt),\n },\n cost: { kind: 'uncaptured', usd: null },\n }\n }\n\n let parsed: Record<string, unknown> | null = null\n try {\n const cleaned = raw\n .replace(/^```json\\s*/i, '')\n .replace(/```\\s*$/, '')\n .trim()\n parsed = JSON.parse(cleaned) as Record<string, unknown>\n } catch {\n return {\n score: {\n dimensions: {},\n composite: 0,\n failed: true,\n notes: `judge ${judge.name} returned non-JSON: ${raw.slice(0, 200)}`,\n llmCall,\n },\n cost,\n }\n }\n\n const dimensions: Record<string, number> = {}\n let sum = 0\n for (const dim of judge.dimensions) {\n const v = Number(parsed[dim.key] ?? 0)\n const clamped = Number.isFinite(v) ? Math.max(0, Math.min(10, v)) : 0\n dimensions[dim.key] = clamped\n sum += clamped\n }\n return {\n score: {\n dimensions,\n composite: judge.dimensions.length === 0 ? 0 : sum / judge.dimensions.length,\n notes: typeof parsed.notes === 'string' ? parsed.notes : '',\n llmCall,\n },\n cost,\n }\n}\n\nfunction judgeCallMetadata(response: RouterCompletionResponse): {\n llmCall: LlmCallMetadata\n cost: CostProvenance\n} {\n const usage = canonicalUsage(response.usage)\n return {\n llmCall: {\n usage,\n costUsd: response.costUsd ?? null,\n model: response.model,\n durationMs: response.durationMs,\n },\n cost:\n response.costUsd !== undefined\n ? { kind: 'observed', usd: response.costUsd }\n : usage.captured === false || !isModelPriced(response.model)\n ? { kind: 'uncaptured', usd: null }\n : {\n kind: 'estimated',\n usd: estimateCost(usage.promptTokens, usage.completionTokens, response.model),\n },\n }\n}\n\nfunction canonicalUsage(usage: RouterCompletionResponse['usage']): LlmUsage {\n const promptTokens = tokenCount(usage?.prompt_tokens)\n const completionTokens = tokenCount(usage?.completion_tokens)\n const captured = promptTokens !== undefined && completionTokens !== undefined\n return {\n promptTokens: promptTokens ?? 0,\n completionTokens: completionTokens ?? 0,\n totalTokens: (promptTokens ?? 0) + (completionTokens ?? 0),\n captured,\n }\n}\n\nfunction unknownJudgeCall(model: string, durationMs: number): LlmCallMetadata {\n return {\n usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0, captured: false },\n costUsd: null,\n model,\n durationMs,\n }\n}\n\nfunction tokenCount(value: unknown): number | undefined {\n return typeof value === 'number' && Number.isSafeInteger(value) && value >= 0 ? value : undefined\n}\n\n/** Convenience: stringified dimension list for inclusion in a judge prompt.\n * Returns lines like `- audience_fit: Does this match what the audience cares about? (0-10)`. */\nexport function renderDimensions(dims: readonly JudgeDimension[]): string {\n return dims.map((d) => `- ${d.key}: ${d.description}`).join('\\n')\n}\n\n/** Convenience: build the \"Respond with ONLY this JSON\" footer for a judge prompt. */\nexport function renderJsonFooter(dims: readonly JudgeDimension[]): string {\n const fields = dims.map((d) => `\"${d.key}\":N`).join(',')\n return `Respond with ONLY this JSON (no markdown, no preamble):\\n{${fields},\"notes\":\"1-2 sentence critique\"}`\n}\n","// Profile-derived defaults for MultishotShape — the pure-profile path.\n//\n// `MultishotShape`'s callbacks were REQUIRED beside `profile: AgentProfile`,\n// which mixed the abstractions: the node under test was a profile, but the\n// simulated user's role could only be expressed as code. These defaults make\n// the callbacks optional — a pure-profile call works, and the derived prompts\n// are plain data (a pure function of profile + persona payload), so an\n// optimizer can sweep them (tangle-network/agent-runtime#694).\n\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport type { MultishotPersona, MultishotShape } from './types'\n\n/** Persona payload rendered as stable `- key: value` lines. `id` is identity,\n * not voice, and structured values are serialized so nothing is dropped. */\nexport function renderPersonaFacts(persona: MultishotPersona): string {\n const lines: string[] = []\n for (const [key, value] of Object.entries(persona)) {\n if (key === 'id' || value === undefined) continue\n const rendered =\n typeof value === 'string' || typeof value === 'number' || typeof value === 'boolean'\n ? String(value)\n : JSON.stringify(value)\n lines.push(`- ${key}: ${rendered}`)\n }\n return lines.join('\\n')\n}\n\n/** Default opening user message: the persona introduces itself with its whole\n * payload and asks the agent (as described by the profile) to take over. */\nexport function defaultMultishotOpener(profile: AgentProfile, persona: MultishotPersona): string {\n const facts = renderPersonaFacts(persona)\n const agentDescription = profile.description ?? profile.name ?? 'the agent'\n return [\n `Hi — I'm reaching out to ${agentDescription} for help. My situation:`,\n facts.length > 0 ? facts : `- (no details provided beyond id: ${persona.id})`,\n '',\n 'Tell me what you need from me, and get started on my case.',\n ].join('\\n')\n}\n\n/** Default driver system prompt: roleplay the persona from its payload, stay\n * in character, push back on vague answers, and never go silent. */\nexport function defaultMultishotDriverSystemPrompt(\n profile: AgentProfile,\n persona: MultishotPersona,\n): string {\n const facts = renderPersonaFacts(persona)\n const agentDescription = profile.description ?? profile.name ?? 'an AI agent'\n return [\n `You are role-playing a real person (persona \"${persona.id}\") talking to ${agentDescription}.`,\n facts.length > 0 ? `Who you are:\\n${facts}` : '',\n '',\n 'How to behave:',\n '- Stay in character: first person, realistic voice, concrete details from your situation.',\n '- Judge every reply by whether it actually moves YOUR case forward. Push back on vague, hedged, or generic answers and demand specifics.',\n \"- Answer the agent's questions with information consistent with who you are; invent plausible details rather than stalling.\",\n '- Never go silent: always produce a substantive next message.',\n '- Output ONLY your next message to the agent — no meta-commentary, no stage directions.',\n ]\n .filter((block) => block.length > 0)\n .join('\\n')\n}\n\n/** The fully-resolved shape for a profile: caller overrides win, the\n * profile-derived defaults fill the rest. */\nexport function defaultShapeFromProfile<TPersona extends MultishotPersona>(\n profile: AgentProfile,\n shape?: MultishotShape<TPersona>,\n): Required<MultishotShape<TPersona>> {\n return {\n buildOpener: shape?.buildOpener ?? ((persona) => defaultMultishotOpener(profile, persona)),\n buildDriverSystemPrompt:\n shape?.buildDriverSystemPrompt ??\n ((persona) => defaultMultishotDriverSystemPrompt(profile, persona)),\n }\n}\n","// Multi-turn driver-agent simulation with inline tool execution.\n//\n// The driver = LLM acting as the persona (reactive, non-deterministic).\n// The agent = the product agent under test (router call by default, or an\n// injected transport — with profile's systemPrompt + the configured tools).\n// Tool calls execute inline via the configured executors and feed back\n// into the agent's message log so the agent integrates the result.\n\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { withCellSpend } from '../matrix'\nimport { defaultDelegationTools } from './default-tools'\nimport {\n defaultRouterBaseUrl,\n estimateRouterCost,\n requireRouterApiKey,\n routerCompletion,\n} from './router'\nimport { defaultShapeFromProfile } from './shape-defaults'\nimport {\n type MultishotArtifact,\n MultishotDriverEmptyError,\n MultishotFatalToolError,\n type MultishotMessage,\n type MultishotPersona,\n type MultishotResult,\n type MultishotShape,\n type MultishotToolDefinition,\n type MultishotToolExecutor,\n type MultishotTransport,\n} from './types'\n\nexport interface RunMultishotOptions<TPersona extends MultishotPersona> {\n profile: AgentProfile\n persona: TPersona\n /** Persona-shaping callbacks. Optional — omitted callbacks are derived from\n * the profile + persona payload, so a pure-profile call works. */\n shape?: MultishotShape<TPersona>\n /** Tool definitions advertised to the agent. Defaults to delegate_research + delegate_code. */\n tools?: MultishotToolDefinition[]\n /** Map from tool name → executor invoked inline when the agent emits a tool_call. */\n toolExecutors?: Record<string, MultishotToolExecutor>\n /** Map from tool name → artifact type label written into MultishotArtifact.type.\n * Tools without a mapping still execute, but their results aren't surfaced as\n * typed artifacts (only as tool messages in the transcript). */\n artifactTypeFor?: (toolName: string) => string | undefined\n maxTurns?: number\n agentModel?: string\n driverModel?: string\n /** Fallback driver models tried when the primary simulated-user model returns empty twice. */\n driverFallbackModels?: string[]\n /** Maximum output tokens for the first agent call in each assistant turn. */\n agentMaxTokens?: number\n /** Maximum output tokens for agent follow-up calls after tool results. */\n toolFollowupMaxTokens?: number\n /** Maximum output tokens for each simulated-user driver response. */\n driverMaxTokens?: number\n /** Maximum tool calls the agent may dispatch inside one assistant turn. */\n maxToolDispatches?: number\n /** Execution seam for the agent leg. When provided, every agent inference\n * step goes through this function instead of the router HTTP call; the\n * string levers (agentModel, apiKey, baseUrl) stop applying to that leg.\n * apiKey/baseUrl are still resolved for tool executors and any leg\n * without an injected transport. */\n agentTransport?: MultishotTransport\n /** Execution seam for the simulated-user driver leg (symmetric to\n * agentTransport). Driver model fallback rotation still applies — the\n * transport receives each candidate model in turn. */\n driverTransport?: MultishotTransport\n apiKey?: string\n baseUrl?: string\n signal?: AbortSignal\n}\n\n/** One multishot shot — the conversation engine `runMultishotMatrix` invokes\n * once per cell. `runMultishot` is the default implementation.\n *\n * An alternative engine (a graph-backed conversation, a replay of a recorded\n * transcript, a sandbox-hosted agent) implements this exact signature and\n * reaches the matrix through `RunMultishotMatrixOptions.runShot`. The matrix\n * keeps every other cell mechanic — cell fan-out, concurrency, the cost\n * ceiling, the judge slots, the cell composite, and the per-cell writers — so\n * swapping the engine needs no copy of the cell body. */\nexport type MultishotShot<TPersona extends MultishotPersona> = (\n opts: RunMultishotOptions<TPersona>,\n) => Promise<MultishotResult>\n\n/** Running spend of one shot, readable after the shot throws. */\ninterface ShotMeter {\n costUsd: number\n startedAt: number\n /** True once a call contributed a fabricated `0` — no provider cost and no\n * usage to price. `costUsd` is a subtotal from that point on. */\n uncaptured: boolean\n}\n\nexport async function runMultishot<TPersona extends MultishotPersona>(\n opts: RunMultishotOptions<TPersona>,\n): Promise<MultishotResult> {\n const meter: ShotMeter = { costUsd: 0, startedAt: Date.now(), uncaptured: false }\n try {\n return await runShotTurns(opts, meter)\n } catch (err) {\n // The turns already paid for. Declare that spend so the matrix bills the\n // failed cell for it and the cost ceiling sees the real cumulative total.\n throw withCellSpend(err, {\n costUsd: meter.costUsd,\n durationMs: Date.now() - meter.startedAt,\n kind: meter.uncaptured ? 'uncaptured' : 'estimated',\n })\n }\n}\n\nasync function runShotTurns<TPersona extends MultishotPersona>(\n opts: RunMultishotOptions<TPersona>,\n meter: ShotMeter,\n): Promise<MultishotResult> {\n const apiKey = opts.apiKey ?? requireRouterApiKey()\n const baseUrl = opts.baseUrl ?? defaultRouterBaseUrl()\n const maxTurns = opts.maxTurns ?? 10\n const maxToolDispatches = opts.maxToolDispatches ?? 4\n const agentModel = opts.agentModel ?? 'openai/gpt-5.4'\n const driverModel = opts.driverModel ?? 'openai/gpt-4o-mini'\n const driverModels = [driverModel, ...(opts.driverFallbackModels ?? [])]\n const agentMaxTokens = opts.agentMaxTokens ?? 2500\n const toolFollowupMaxTokens = opts.toolFollowupMaxTokens ?? 2000\n const driverMaxTokens = opts.driverMaxTokens ?? 600\n\n const bundle =\n opts.tools && opts.toolExecutors\n ? {\n tools: opts.tools,\n executors: opts.toolExecutors,\n artifactTypeFor: opts.artifactTypeFor ?? (() => undefined),\n }\n : defaultDelegationTools()\n const tools = opts.tools ?? bundle.tools\n const executors = opts.toolExecutors ?? bundle.executors\n const artifactTypeFor = opts.artifactTypeFor ?? bundle.artifactTypeFor\n\n const routerTransport: MultishotTransport = (req) => routerCompletion({ apiKey, baseUrl, ...req })\n const agentTransport = opts.agentTransport ?? routerTransport\n const driverTransport = opts.driverTransport ?? routerTransport\n\n const shape = defaultShapeFromProfile(opts.profile, opts.shape)\n\n const transcript: MultishotMessage[] = []\n const artifacts: MultishotArtifact[] = []\n let toolCalls = 0\n\n const opener = shape.buildOpener(opts.persona)\n transcript.push({ role: 'user', content: opener })\n\n const systemPrompt = [opts.profile.prompt?.systemPrompt, opts.profile.prompt?.appendSystemPrompt]\n .filter((value): value is string => Boolean(value))\n .join('\\n\\n')\n const agentMessages: Array<Record<string, unknown>> = [\n { role: 'system', content: systemPrompt },\n { role: 'user', content: opener },\n ]\n\n for (let turn = 0; turn < maxTurns; turn++) {\n if (opts.signal?.aborted) throw new Error('multishot aborted')\n\n let dispatchesThisTurn = 0\n while (true) {\n const {\n message: agentMsg,\n usage: agentUsage,\n costUsd: agentCostUsd,\n } = await agentTransport({\n model: agentModel,\n messages: agentMessages,\n tools,\n temperature: 0.7,\n maxTokens: dispatchesThisTurn === 0 ? agentMaxTokens : toolFollowupMaxTokens,\n signal: opts.signal,\n })\n meter.costUsd += agentCostUsd ?? estimateRouterCost(agentModel, agentUsage)\n if (agentCostUsd === undefined && agentUsage === undefined) meter.uncaptured = true\n\n const agentText = (agentMsg.content ?? '').trim()\n const agentToolCalls = (agentMsg.tool_calls ?? []).map((tc) => ({\n id: tc.id,\n name: tc.function.name,\n args: (() => {\n try {\n return JSON.parse(tc.function.arguments) as Record<string, unknown>\n } catch {\n return {} as Record<string, unknown>\n }\n })(),\n }))\n\n agentMessages.push({\n role: 'assistant',\n content: agentText || null,\n ...(agentMsg.tool_calls?.length ? { tool_calls: agentMsg.tool_calls } : {}),\n })\n transcript.push({\n role: 'assistant',\n content: agentText,\n toolCalls: agentToolCalls.length > 0 ? agentToolCalls : undefined,\n })\n\n if (agentToolCalls.length === 0) break\n dispatchesThisTurn += agentToolCalls.length\n if (dispatchesThisTurn > maxToolDispatches) {\n throw new Error(\n `multishot: tool dispatch cap exceeded (${dispatchesThisTurn}/${maxToolDispatches}) on turn ${turn}`,\n )\n }\n\n for (const tc of agentToolCalls) {\n toolCalls++\n let toolResult = ''\n try {\n const executor = executors[tc.name]\n if (!executor) {\n toolResult = JSON.stringify({ error: `unknown tool ${tc.name}` })\n } else {\n const r = await executor(tc.args, { apiKey, baseUrl, signal: opts.signal })\n toolResult = r.content\n meter.costUsd += r.costUsd\n const artifactType = artifactTypeFor(tc.name)\n if (artifactType) {\n artifacts.push({\n type: artifactType,\n turn,\n invocation: { name: tc.name, args: tc.args },\n content: toolResult,\n })\n }\n }\n } catch (err) {\n if (err instanceof MultishotFatalToolError) throw err\n toolResult = JSON.stringify({ error: err instanceof Error ? err.message : String(err) })\n }\n agentMessages.push({ role: 'tool', tool_call_id: tc.id, content: toolResult || 'done' })\n transcript.push({ role: 'tool', content: toolResult || 'done', toolCallId: tc.id })\n }\n }\n\n if (turn < maxTurns - 1) {\n const driver = await driverTurn({\n transport: driverTransport,\n persona: opts.persona,\n shape,\n transcript,\n turn,\n models: driverModels,\n maxTokens: driverMaxTokens,\n signal: opts.signal,\n meter,\n })\n agentMessages.push({ role: 'user', content: driver.content })\n transcript.push({ role: 'user', content: driver.content })\n }\n }\n\n return {\n transcript,\n artifacts,\n toolCalls,\n durationMs: Date.now() - meter.startedAt,\n costUsd: meter.costUsd,\n // A shot that priced every call reports a complete estimate. One that had\n // a call the router priced at nothing reports a subtotal, so the cell is\n // recorded as under-counted rather than as a complete estimate.\n costProvenance: meter.uncaptured\n ? { kind: 'uncaptured', usd: null }\n : { kind: 'estimated', usd: meter.costUsd },\n }\n}\n\nasync function driverTurn<TPersona extends MultishotPersona>(opts: {\n transport: MultishotTransport\n persona: TPersona\n shape: Required<MultishotShape<TPersona>>\n transcript: MultishotMessage[]\n turn: number\n models: string[]\n maxTokens: number\n signal?: AbortSignal\n /** Charged for EVERY attempt, including the empty ones that force a retry\n * and the ones that precede `MultishotDriverEmptyError`. Those calls billed\n * the provider whether or not their content was usable. */\n meter: ShotMeter\n}): Promise<{ content: string }> {\n const driverSystem = opts.shape.buildDriverSystemPrompt(opts.persona)\n\n // Translate transcript to driver POV: agent's `assistant` messages become\n // `user` (the agent talking TO the driver); the driver's prior `user`\n // messages become `assistant` (the driver's prior responses).\n const driverMessages: Array<Record<string, unknown>> = [{ role: 'system', content: driverSystem }]\n for (const msg of opts.transcript) {\n if (msg.role === 'tool') continue\n const content = driverVisibleContent(msg)\n if (!content) continue\n if (msg.role === 'assistant') driverMessages.push({ role: 'user', content })\n else if (msg.role === 'user') driverMessages.push({ role: 'assistant', content })\n }\n\n // Driver must never go silent. Retry once on empty content; then fail loud.\n for (const model of opts.models) {\n for (let attempt = 0; attempt < 2; attempt++) {\n const { message, usage, costUsd } = await opts.transport({\n model,\n messages: driverMessages,\n temperature: 0.9,\n maxTokens: opts.maxTokens,\n signal: opts.signal,\n })\n opts.meter.costUsd += costUsd ?? estimateRouterCost(model, usage)\n if (costUsd === undefined && usage === undefined) opts.meter.uncaptured = true\n const content = (message.content ?? '').trim()\n if (content.length > 0) return { content }\n }\n }\n throw new MultishotDriverEmptyError(opts.turn)\n}\n\nfunction driverVisibleContent(msg: MultishotMessage): string | null {\n const text = msg.content.trim()\n if (text.length > 0) return text\n if (msg.role !== 'assistant' || !msg.toolCalls?.length) return null\n\n const toolNames = msg.toolCalls.map((call) => call.name.trim()).filter(Boolean)\n if (toolNames.length === 0) return 'Agent called tools.'\n return `Agent called ${toolNames.length === 1 ? 'tool' : 'tools'}: ${toolNames.join(', ')}.`\n}\n","// Multishot matrix wrapper — sweeps profiles × personas × reps, runs\n// the driver-agent loop per cell, applies up to three configured judges,\n// persists per-cell artifacts, and aggregates by axis.\n//\n// Uses runAgentMatrix from @tangle-network/agent-eval/matrix under the\n// hood so cell scheduling + concurrency + cost ceiling are unified with\n// other matrix consumers.\n\nimport { mkdirSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport type { CostProvenance } from '../cost-ledger'\nimport type { MatrixResult } from '../matrix'\nimport { runAgentMatrix, withCellSpend } from '../matrix'\nimport { type JudgeConfig, type JudgeScore, runJudge } from './judges'\nimport { type MultishotShot, runMultishot } from './multishot'\nimport {\n assertMultishotShotResult,\n type MultishotArtifact,\n type MultishotMessage,\n type MultishotPersona,\n type MultishotShape,\n type MultishotToolDefinition,\n type MultishotToolExecutor,\n type MultishotTransport,\n} from './types'\n\nexport interface ConversationJudgeInput<TPersona extends MultishotPersona> {\n transcript: MultishotMessage[]\n persona: TPersona\n}\n\nexport interface ArtifactJudgeInput<TPersona extends MultishotPersona> {\n artifact: MultishotArtifact\n persona: TPersona\n}\n\nexport interface MultishotJudges<TPersona extends MultishotPersona> {\n /** Scores the full transcript end-to-end (always runs). */\n conversation: JudgeConfig<ConversationJudgeInput<TPersona>>\n /** Scores each code-type artifact. Optional — omit when domain has no code artifacts. */\n codeReview?: JudgeConfig<ArtifactJudgeInput<TPersona>>\n /** Scores each non-code (research/content/template) artifact. Optional. */\n contentQuality?: JudgeConfig<ArtifactJudgeInput<TPersona>>\n /** Which artifact types route to codeReview. Defaults to ['code']. */\n codeArtifactTypes?: string[]\n /** Which artifact types route to contentQuality. Defaults to ['research']. */\n contentArtifactTypes?: string[]\n}\n\nexport interface CellCompositeScore {\n composite: number\n conversation: JudgeScore\n codeReview?: {\n perArtifact: Array<JudgeScore & { turn: number; type: string }>\n composite: number\n }\n contentQuality?: {\n perArtifact: Array<JudgeScore & { turn: number; type: string }>\n composite: number\n }\n}\n\nexport interface RunMultishotMatrixOptions<TPersona extends MultishotPersona> {\n /** AgentProfile axis (matrix primary). */\n profiles: Array<{ id: string; value: AgentProfile }>\n /** Persona axis. */\n personas: TPersona[]\n /** Persona-shaping callbacks. Optional — omitted callbacks are derived per\n * cell from that cell's profile + persona payload (pure-profile path). */\n shape?: MultishotShape<TPersona>\n /** Judge configurations. */\n judges: MultishotJudges<TPersona>\n /** Tool definitions advertised to the agent. Defaults to delegate_research + delegate_code. */\n tools?: MultishotToolDefinition[]\n /** Map from tool name → inline executor. Must align with `tools`. */\n toolExecutors?: Record<string, MultishotToolExecutor>\n /** Tool name → artifact type label. Defaults to research/code mapping. */\n artifactTypeFor?: (toolName: string) => string | undefined\n /** Where per-cell artifacts land. Cells write to `<runDir>/<profileId>/<personaId>/rep-N/`. */\n runDir: string\n /** Replicates per (profile, persona) cell. */\n reps?: number\n /** Max conversation turns per cell. */\n maxTurns?: number\n /** Maximum tool calls the agent may dispatch inside one assistant turn. */\n maxToolDispatches?: number\n /** Max concurrent cells. */\n maxConcurrency?: number\n /** Total $ ceiling across the matrix; cells aborted past this. */\n costCeiling?: number\n /** Upper bound on what one cell can spend. A cell whose cost is a subtotal\n * is charged this bound against `costCeiling` instead of its known amount,\n * so hidden spend cannot walk the run past its budget. */\n maxCellCostUsd?: number\n /** Agent model. */\n agentModel?: string\n /** Driver model. */\n driverModel?: string\n /** Fallback driver models tried when the primary simulated-user model returns empty twice. */\n driverFallbackModels?: string[]\n /** Maximum output tokens for the first agent call in each assistant turn. */\n agentMaxTokens?: number\n /** Maximum output tokens for agent follow-up calls after tool results. */\n toolFollowupMaxTokens?: number\n /** Maximum output tokens for each simulated-user driver response. */\n driverMaxTokens?: number\n /** Maximum output tokens for each judge response. */\n judgeMaxTokens?: number\n /** Execution seam for the agent leg of every cell — replaces the router\n * HTTP call when provided (see RunMultishotOptions.agentTransport).\n * Judges are unaffected; configure those via MultishotJudges. */\n agentTransport?: MultishotTransport\n /** Execution seam for the simulated-user driver leg of every cell. */\n driverTransport?: MultishotTransport\n /** Conversation engine for every cell. Defaults to `runMultishot`.\n *\n * The matrix owns everything around the shot — cell fan-out, concurrency,\n * the cost ceiling, the judge slots, the cell composite, the per-cell\n * artifact writers and the run summary — and forwards the whole cell input\n * to this function, so an alternative engine replaces ONLY the\n * conversation. Every option on this interface that `runMultishot` accepts\n * reaches the shot unchanged. `RunMultishotOptions.signal` has no\n * matrix-level counterpart and is not forwarded; a shot owns its own\n * cancellation.\n *\n * A shot that resolves with a value outside `MultishotResult` throws\n * `MultishotShotResultError` for that cell. The default engine is never\n * used as a fallback. */\n runShot?: MultishotShot<TPersona>\n /** Pass-thru fields. */\n apiKey?: string\n baseUrl?: string\n}\n\n/** Per-cell output the multishot matrix records in `MatrixResult.cells`.\n * A consumer that supplies its own `runShot` reads the matrix through this\n * type instead of declaring a structural copy. */\nexport interface MultishotCellOutput {\n turns: number\n toolCalls: number\n artifactCount: number\n}\n\ninterface ArtifactJudgeRun {\n score: JudgeScore & { turn: number; type: string }\n cost: CostProvenance\n}\n\n/** Mean composite over non-failed scores. `0` when the list is empty (a\n * configured judge with nothing to score contributes 0, matching the cell\n * composite's long-standing semantics); `null` when scores exist but every\n * one failed — no signal, so the slot must be EXCLUDED from the cell mean\n * rather than dragging it to zero. */\nfunction meanCompositeExcludingFailed(scores: ReadonlyArray<JudgeScore>): number | null {\n if (scores.length === 0) return 0\n const live = scores.filter((s) => !s.failed)\n if (live.length === 0) return null\n return live.reduce((sum, s) => sum + s.composite, 0) / live.length\n}\n\nexport interface CellCompositeInput {\n conversation: JudgeScore\n /** Present iff the codeReview judge is configured. */\n codeReviews?: ReadonlyArray<JudgeScore>\n /** Present iff the contentQuality judge is configured. */\n contentReviews?: ReadonlyArray<JudgeScore>\n}\n\n/** Cell composite = mean over configured judge slots, excluding failed\n * scores: a failed conversation judge or an all-failed artifact slot carries\n * no signal and is dropped from the mean. `composite` is 0 only when EVERY\n * configured slot failed (`allJudgesFailed` distinguishes that from a real\n * zero). Pure — exported for deterministic testing. */\nexport function computeCellComposite(input: CellCompositeInput): {\n composite: number\n codeComposite: number\n contentComposite: number\n allJudgesFailed: boolean\n} {\n const contributions: number[] = []\n if (!input.conversation.failed) contributions.push(input.conversation.composite)\n\n const codeMean = input.codeReviews ? meanCompositeExcludingFailed(input.codeReviews) : undefined\n if (typeof codeMean === 'number') contributions.push(codeMean)\n const contentMean = input.contentReviews\n ? meanCompositeExcludingFailed(input.contentReviews)\n : undefined\n if (typeof contentMean === 'number') contributions.push(contentMean)\n\n return {\n composite:\n contributions.length === 0\n ? 0\n : contributions.reduce((s, v) => s + v, 0) / contributions.length,\n codeComposite: codeMean ?? 0,\n contentComposite: contentMean ?? 0,\n allJudgesFailed: contributions.length === 0,\n }\n}\n\nexport interface RunMultishotMatrixResult {\n matrix: MatrixResult<MultishotCellOutput>\n}\n\nexport async function runMultishotMatrix<TPersona extends MultishotPersona>(\n opts: RunMultishotMatrixOptions<TPersona>,\n): Promise<RunMultishotMatrixResult> {\n const codeTypes = new Set(opts.judges.codeArtifactTypes ?? ['code'])\n const contentTypes = new Set(opts.judges.contentArtifactTypes ?? ['research'])\n const runShot: MultishotShot<TPersona> = opts.runShot ?? runMultishot\n mkdirSync(opts.runDir, { recursive: true })\n\n const matrix = await runAgentMatrix<MultishotCellOutput>({\n axes: [\n { name: 'profile', values: opts.profiles },\n { name: 'persona', values: opts.personas.map((p) => ({ id: p.id, value: p })) },\n ],\n reps: opts.reps ?? 1,\n maxConcurrency: opts.maxConcurrency ?? 2,\n costCeiling: opts.costCeiling,\n maxCellCostUsd: opts.maxCellCostUsd,\n async runCell(cell) {\n const cellStartedAt = Date.now()\n const profile = cell.axes.profile?.value as AgentProfile\n const persona = cell.axes.persona?.value as TPersona\n const profileId = String(cell.axes.profile?.id ?? 'unknown')\n const personaId = String(cell.axes.persona?.id ?? 'unknown')\n\n // A shot that throws declares its own spend (see `runMultishot`). Rethrow\n // it untouched: a shot that declared nothing spent an unknown amount, and\n // claiming a number here would report a fabricated total.\n const sim = await runShot({\n profile,\n persona,\n shape: opts.shape,\n tools: opts.tools,\n toolExecutors: opts.toolExecutors,\n artifactTypeFor: opts.artifactTypeFor,\n maxTurns: opts.maxTurns,\n maxToolDispatches: opts.maxToolDispatches,\n agentModel: opts.agentModel,\n driverModel: opts.driverModel,\n driverFallbackModels: opts.driverFallbackModels,\n agentMaxTokens: opts.agentMaxTokens,\n toolFollowupMaxTokens: opts.toolFollowupMaxTokens,\n driverMaxTokens: opts.driverMaxTokens,\n agentTransport: opts.agentTransport,\n driverTransport: opts.driverTransport,\n apiKey: opts.apiKey,\n baseUrl: opts.baseUrl,\n })\n // Everything from here on runs with the shot's spend already committed.\n // A throw past this line must carry it, or the money leaves the matrix's\n // cumulative sum and the cost ceiling under-counts the run.\n const shotCostUsd = shotCostSubtotal(sim)\n // A shot that says nothing about provenance is taken at its word, which\n // is how every engine written before the field behaved.\n const shotCostComplete = sim?.costProvenance?.kind !== 'uncaptured'\n let judgeCostUsd = 0\n let judgeCostComplete = false\n // Where the cell is, so a throw declares the right completeness: before\n // the judges only the shot has spent, while they are in flight their\n // spend is unknown, and after they settle their receipts decide.\n let phase: 'validate' | 'judging' | 'scoring' = 'validate'\n try {\n assertMultishotShotResult(sim)\n\n const codeArtifacts = sim.artifacts.filter((a) => codeTypes.has(a.type))\n const contentArtifacts = sim.artifacts.filter((a) => contentTypes.has(a.type))\n\n phase = 'judging'\n const [conversationRun, codeReviewRuns, contentReviewRuns] = await Promise.all([\n runJudge(withJudgeMaxTokens(opts.judges.conversation, opts.judgeMaxTokens), {\n transcript: sim.transcript,\n persona,\n }),\n opts.judges.codeReview\n ? Promise.all(\n codeArtifacts.map((artifact) =>\n runJudge(withJudgeMaxTokens(opts.judges.codeReview!, opts.judgeMaxTokens), {\n artifact,\n persona,\n }).then((result) => ({\n score: {\n ...result.score,\n turn: artifact.turn,\n type: artifact.type,\n },\n cost: result.cost,\n })),\n ),\n )\n : Promise.resolve([] as ArtifactJudgeRun[]),\n opts.judges.contentQuality\n ? Promise.all(\n contentArtifacts.map((artifact) =>\n runJudge(withJudgeMaxTokens(opts.judges.contentQuality!, opts.judgeMaxTokens), {\n artifact,\n persona,\n }).then((result) => ({\n score: {\n ...result.score,\n turn: artifact.turn,\n type: artifact.type,\n },\n cost: result.cost,\n })),\n ),\n )\n : Promise.resolve([] as ArtifactJudgeRun[]),\n ])\n const judgeRuns = [conversationRun, ...codeReviewRuns, ...contentReviewRuns]\n judgeCostUsd = judgeRuns.reduce((sum, run) => sum + (run.cost.usd ?? 0), 0)\n judgeCostComplete = judgeRuns.every((run) => run.cost.kind !== 'uncaptured')\n phase = 'scoring'\n\n const conversation = conversationRun.score\n const codeReviews = codeReviewRuns.map((run) => run.score)\n const contentReviews = contentReviewRuns.map((run) => run.score)\n\n const { composite, codeComposite, contentComposite, allJudgesFailed } =\n computeCellComposite({\n conversation,\n codeReviews: opts.judges.codeReview ? codeReviews : undefined,\n contentReviews: opts.judges.contentQuality ? contentReviews : undefined,\n })\n\n const cellScore: CellCompositeScore = { composite, conversation }\n if (opts.judges.codeReview)\n cellScore.codeReview = { perArtifact: codeReviews, composite: codeComposite }\n if (opts.judges.contentQuality)\n cellScore.contentQuality = { perArtifact: contentReviews, composite: contentComposite }\n\n const cellDir = join(opts.runDir, profileId, personaId, `rep-${cell.rep}`)\n mkdirSync(cellDir, { recursive: true })\n writeFileSync(join(cellDir, 'transcript.json'), JSON.stringify(sim.transcript, null, 2))\n writeFileSync(join(cellDir, 'artifacts.json'), JSON.stringify(sim.artifacts, null, 2))\n writeFileSync(join(cellDir, 'scores.json'), JSON.stringify(cellScore, null, 2))\n\n const notes = [`convo=${conversation.composite.toFixed(1)}`]\n if (opts.judges.codeReview) notes.push(`code=${codeComposite.toFixed(1)}`)\n if (opts.judges.contentQuality) notes.push(`content=${contentComposite.toFixed(1)}`)\n if (allJudgesFailed) notes.push('all-judges-failed')\n if (!judgeCostComplete) notes.push('judge-cost-incomplete')\n\n return {\n output: {\n turns: sim.transcript.length,\n toolCalls: sim.toolCalls,\n artifactCount: sim.artifacts.length,\n },\n verdict: { valid: composite >= 5, score: composite, notes: notes.join(' ') },\n costUsd: sim.costUsd + judgeCostUsd,\n // Either leg can leave the cell total a subtotal: a judge whose cost\n // the router never reported, or a shot that declared its own spend\n // uncaptured. The matrix counts the known part toward the ceiling and\n // reports the cell as under-counted rather than presenting the sum as\n // complete.\n costProvenance:\n judgeCostComplete && shotCostComplete\n ? { kind: 'estimated', usd: sim.costUsd + judgeCostUsd }\n : { kind: 'uncaptured', usd: null },\n durationMs: sim.durationMs,\n }\n } catch (err) {\n // No usable shot subtotal means the cell's spend is genuinely unknown;\n // rethrow untouched so the matrix records it as uncaptured instead of\n // billing a number this frame cannot support.\n if (shotCostUsd === undefined) throw err\n throw withCellSpend(err, {\n costUsd: shotCostUsd + judgeCostUsd,\n durationMs: Date.now() - cellStartedAt,\n // The judges bill before they settle, so a throw among them leaves a\n // subtotal. Before they start, and after they settle with every\n // receipt captured, the amount is the cell's whole spend — unless the\n // shot already declared its own part incomplete.\n kind:\n shotCostComplete && (phase === 'validate' || (phase === 'scoring' && judgeCostComplete))\n ? 'estimated'\n : 'uncaptured',\n })\n }\n },\n })\n\n // Persist top-level summary.\n const summary = {\n cells: matrix.summary.totalCells,\n passRate: matrix.summary.overallPassRate,\n meanScore: matrix.summary.overallMeanScore,\n totalCostUsd: matrix.summary.totalCostUsd,\n costUncapturedCells: matrix.summary.costUncapturedCells,\n ceilingChargedUsd: matrix.summary.ceilingChargedUsd,\n durationMs: matrix.summary.durationMs,\n runsExecuted: matrix.summary.runsExecuted,\n cellsSkipped: matrix.summary.cellsSkipped,\n byProfile: matrix.byAxis.profile,\n byPersona: matrix.byAxis.persona,\n }\n writeFileSync(join(opts.runDir, 'summary.json'), JSON.stringify(summary, null, 2))\n\n // A reader of the on-disk summary must not take the cost line as the run's\n // whole spend when some cells reported only a subtotal.\n const uncaptured = matrix.summary.costUncapturedCells\n const costLabel = uncaptured > 0 ? 'Cost (at least)' : 'Cost'\n\n const md: string[] = [\n `# Multishot matrix`,\n ``,\n `**Cells**: ${matrix.summary.totalCells} | **Pass rate**: ${(matrix.summary.overallPassRate * 100).toFixed(0)}% | **Mean**: ${matrix.summary.overallMeanScore.toFixed(2)} | **${costLabel}**: $${matrix.summary.totalCostUsd.toFixed(2)} | **Duration**: ${(matrix.summary.durationMs / 1000).toFixed(0)}s`,\n ``,\n ...(uncaptured > 0\n ? [\n `> ${uncaptured} of ${matrix.summary.runsExecuted} cells reported a cost subtotal, not a total. Real spend is higher than every cost figure below.`,\n ``,\n ]\n : []),\n `## By profile`,\n ``,\n '| profile | pass | mean | cost |',\n '|---|---|---|---|',\n ...Object.entries(matrix.byAxis.profile ?? {}).map(\n ([id, s]) =>\n `| ${id} | ${(s.passRate * 100).toFixed(0)}% | ${s.meanScore.toFixed(2)} | $${s.totalCostUsd.toFixed(2)} |`,\n ),\n ``,\n `## By persona`,\n ``,\n '| persona | pass | mean | cost |',\n '|---|---|---|---|',\n ...Object.entries(matrix.byAxis.persona ?? {}).map(\n ([id, s]) =>\n `| ${id} | ${(s.passRate * 100).toFixed(0)}% | ${s.meanScore.toFixed(2)} | $${s.totalCostUsd.toFixed(2)} |`,\n ),\n ``,\n ]\n writeFileSync(join(opts.runDir, 'summary.md'), md.join('\\n'))\n\n return { matrix }\n}\n\n/** The shot's own spend, when the value it resolved with reports a usable\n * amount. `undefined` when it does not — a malformed shot result is exactly\n * the case where the number cannot be trusted, and billing a wrong figure is\n * worse than recording the cell as uncaptured. */\nfunction shotCostSubtotal(\n sim: { costUsd?: unknown; costProvenance?: CostProvenance } | null | undefined,\n): number | undefined {\n const value = sim?.costUsd\n return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : undefined\n}\n\nfunction withJudgeMaxTokens<TInput>(\n judge: JudgeConfig<TInput>,\n maxTokens: number | undefined,\n): JudgeConfig<TInput> {\n if (maxTokens === undefined || judge.maxTokens !== undefined) return judge\n return { ...judge, maxTokens }\n}\n"],"mappings":";;;;;;AAiCA,eAAsB,iBACpB,KACmC;CACnC,MAAM,YAAY,KAAK,IAAI;CAC3B,MAAM,OAAgC;EACpC,OAAO,IAAI;EACX,UAAU,IAAI;EACd,aAAa,IAAI,eAAe;EAChC,YAAY,IAAI,aAAa;CAC/B;CACA,IAAI,IAAI,OAAO,QAAQ,KAAK,QAAQ,IAAI;CACxC,MAAM,MAAM,GAAG,IAAI,QAAQ,QAAQ,QAAQ,EAAE,EAAE;CAC/C,MAAM,MAAM,MAAM,MAAM,KAAK;EAC3B,QAAQ;EACR,SAAS;GAAE,eAAe,UAAU,IAAI;GAAU,gBAAgB;EAAmB;EACrF,MAAM,KAAK,UAAU,IAAI;EACzB,QAAQ,IAAI;CACd,CAAC;CACD,IAAI,CAAC,IAAI,IAAI;EACX,MAAM,OAAO,MAAM,IAAI,KAAK;EAC5B,MAAM,IAAI,MAAM,UAAU,IAAI,OAAO,IAAI,KAAK,MAAM,GAAG,GAAG,GAAG;CAC/D;CACA,MAAM,OAAQ,MAAM,IAAI,KAAK;CAO7B,MAAM,SAAS,KAAK,QAAQ;CAC5B,IAAI,CAAC,QAAQ,MAAM,IAAI,MAAM,+BAA+B,KAAK,UAAU,IAAI,CAAC,CAAC,MAAM,GAAG,GAAG,GAAG;CAChG,MAAM,UAAU,KAAK,kBAAkB,KAAK;CAC5C,MAAM,UACJ,OAAO,YAAY,YAAY,OAAO,SAAS,OAAO,KAAK,WAAW,IAAI,UAAU,KAAA;CACtF,OAAO;EACL,SAAS,OAAO;EAChB,OAAO,KAAK;EACZ,GAAI,YAAY,KAAA,IAAY,CAAC,IAAI,EAAE,QAAQ;EAC3C,OAAO,OAAO,KAAK,UAAU,YAAY,KAAK,QAAQ,KAAK,QAAQ,IAAI;EACvE,YAAY,KAAK,IAAI,IAAI;CAC3B;AACF;AAIA,SAAgB,mBACd,OACA,OACQ;CACR,IAAI,CAAC,OAAO,OAAO;CACnB,MAAM,WAAW,MAAM,iBAAiB;CACxC,MAAM,YAAY,MAAM,qBAAqB;CAC7C,IAAI,UAAU;CACd,IAAI,WAAW;CACf,IAAI,MAAM,SAAS,aAAa,GAAG;EACjC,UAAU;EACV,WAAW;CACb,OAAO,IAAI,MAAM,SAAS,SAAS,KAAK,MAAM,SAAS,eAAe,GAAG;EACvE,UAAU;EACV,WAAW;CACb,OAAO,IAAI,MAAM,SAAS,MAAM,KAAK,MAAM,SAAS,KAAK,KAAK,MAAM,SAAS,UAAU,GAAG;EACxF,UAAU;EACV,WAAW;CACb;CACA,QAAQ,WAAW,UAAU,YAAY,YAAY;AACvD;AAEA,SAAgB,uBAA+B;CAC7C,QAAQ,QAAQ,IAAI,0BAA0B,iCAAA,CAAkC,QAC9E,QACA,EACF;AACF;AAEA,SAAgB,sBAA8B;CAC5C,MAAM,MAAM,QAAQ,IAAI;CACxB,IAAI,CAAC,KAAK,MAAM,IAAI,MAAM,gEAAgE;CAC1F,OAAO;AACT;;;ACrGA,MAAa,2BAA2B;AACxC,MAAa,sBAAsB;AAgBnC,MAAM,4BACJ;AAEF,MAAM,uBACJ;AAEF,MAAa,iCAA0D;CACrE,MAAM;CACN,UAAU;EACR,MAAM;EACN,aACE;EACF,YAAY;GACV,MAAM;GACN,YAAY;IACV,UAAU;KAAE,MAAM;KAAU,aAAa;IAAgC;IACzE,OAAO;KACL,MAAM;KACN,aAAa;IACf;GACF;GACA,UAAU,CAAC,UAAU;EACvB;CACF;AACF;AAEA,MAAa,6BAAsD;CACjE,MAAM;CACN,UAAU;EACR,MAAM;EACN,aACE;EACF,YAAY;GACV,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;IAAgC;IACrE,UAAU;KACR,MAAM;KACN,aAAa;IACf;GACF;GACA,UAAU,CAAC,MAAM;EACnB;CACF;AACF;AAEA,SAAgB,uBACd,SAAkC,CAAC,GACZ;CACvB,MAAM,eAAe,OAAO,gBAAgB;CAC5C,MAAM,QAAQ,OAAO,SAAA;CACrB,OAAO,OAAO,MAAM,QAAQ;EAC1B,MAAM,WAAW,OAAO,KAAK,YAAY,EAAE;EAC3C,MAAM,QAAQ,KAAK,QAAQ,OAAO,KAAK,KAAK,IAAI,KAAA;EAChD,MAAM,EAAE,SAAS,UAAU,MAAM,iBAAiB;GAChD,QAAQ,IAAI;GACZ,SAAS,IAAI;GACb;GACA,aAAa;GACb,WAAW;GACX,UAAU,CACR;IAAE,MAAM;IAAU,SAAS;GAAa,GACxC;IAAE,MAAM;IAAQ,SAAS,aAAa,WAAW,QAAQ,YAAY,UAAU;GAAK,CACtF;GACA,QAAQ,IAAI;EACd,CAAC;EACD,OAAO;GAAE,SAAS,QAAQ,WAAW;GAAI,SAAS,mBAAmB,OAAO,KAAK;EAAE;CACrF;AACF;AAEA,SAAgB,mBAAmB,SAA6B,CAAC,GAA0B;CACzF,MAAM,eAAe,OAAO,gBAAgB;CAC5C,MAAM,QAAQ,OAAO,SAAA;CACrB,OAAO,OAAO,MAAM,QAAQ;EAC1B,MAAM,OAAO,OAAO,KAAK,QAAQ,EAAE;EACnC,MAAM,WAAW,KAAK,WAAW,OAAO,KAAK,QAAQ,IAAI;EACzD,MAAM,EAAE,SAAS,UAAU,MAAM,iBAAiB;GAChD,QAAQ,IAAI;GACZ,SAAS,IAAI;GACb;GACA,aAAa;GACb,WAAW;GACX,UAAU,CACR;IAAE,MAAM;IAAU,SAAS,GAAG,aAAa,gBAAgB;GAAW,GACtE;IAAE,MAAM;IAAQ,SAAS,YAAY;GAAO,CAC9C;GACA,QAAQ,IAAI;EACd,CAAC;EACD,OAAO;GAAE,SAAS,QAAQ,WAAW;GAAI,SAAS,mBAAmB,OAAO,KAAK;EAAE;CACrF;AACF;AAgBA,SAAgB,uBAAuB,SAA6B,CAAC,GAAuB;CAC1F,OAAO;EACL,OAAO,CAAC,gCAAgC,0BAA0B;EAClE,WAAW;GACT,mBAAmB,uBAAuB,OAAO,QAAQ;GACzD,eAAe,mBAAmB,OAAO,IAAI;EAC/C;EACA,kBAAkB,SAChB,SAAS,sBAAsB,aAAa,SAAS,kBAAkB,SAAS,KAAA;CACpF;AACF;;;ACtHA,MAAa,sBAAsB;AAmCnC,eAAsB,SACpB,OACA,OACyB;CACzB,MAAM,SAAS,MAAM,UAAU,oBAAoB;CACnD,MAAM,UAAU,MAAM,WAAW,qBAAqB;CACtD,MAAM,QAAQ,MAAM,SAAS,QAAQ,IAAI,eAAA;CACzC,MAAM,SAAS,MAAM,YAAY,KAAK;CACtC,IAAI,MAAM;CACV,IAAI;CACJ,IAAI;CACJ,MAAM,YAAY,KAAK,IAAI;CAC3B,IAAI;EACF,MAAM,WAAW,MAAM,iBAAiB;GACtC;GACA;GACA;GACA,aAAa;GACb,WAAW,MAAM,aAAa;GAC9B,UAAU,CACR;IAAE,MAAM;IAAU,SAAS,MAAM;GAAa,GAC9C;IAAE,MAAM;IAAQ,SAAS;GAAO,CAClC;EACF,CAAC;EACD,MAAM,OAAO,kBAAkB,QAAQ;EACvC,UAAU,KAAK;EACf,OAAO,KAAK;EACZ,OAAO,SAAS,QAAQ,WAAW,GAAA,CAAI,KAAK;CAC9C,SAAS,KAAK;EAGZ,OAAO;GACL,OAAO;IACL,YAAY,CAAC;IACb,WAAW;IACX,QAAQ;IACR,OAAO,SAAS,MAAM,KAAK,gBAAgB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC1F,SAAS,iBAAiB,OAAO,KAAK,IAAI,IAAI,SAAS;GACzD;GACA,MAAM;IAAE,MAAM;IAAc,KAAK;GAAK;EACxC;CACF;CAEA,IAAI,SAAyC;CAC7C,IAAI;EACF,MAAM,UAAU,IACb,QAAQ,gBAAgB,EAAE,CAAC,CAC3B,QAAQ,WAAW,EAAE,CAAC,CACtB,KAAK;EACR,SAAS,KAAK,MAAM,OAAO;CAC7B,QAAQ;EACN,OAAO;GACL,OAAO;IACL,YAAY,CAAC;IACb,WAAW;IACX,QAAQ;IACR,OAAO,SAAS,MAAM,KAAK,sBAAsB,IAAI,MAAM,GAAG,GAAG;IACjE;GACF;GACA;EACF;CACF;CAEA,MAAM,aAAqC,CAAC;CAC5C,IAAI,MAAM;CACV,KAAK,MAAM,OAAO,MAAM,YAAY;EAClC,MAAM,IAAI,OAAO,OAAO,IAAI,QAAQ,CAAC;EACrC,MAAM,UAAU,OAAO,SAAS,CAAC,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,CAAC,CAAC,IAAI;EACpE,WAAW,IAAI,OAAO;EACtB,OAAO;CACT;CACA,OAAO;EACL,OAAO;GACL;GACA,WAAW,MAAM,WAAW,WAAW,IAAI,IAAI,MAAM,MAAM,WAAW;GACtE,OAAO,OAAO,OAAO,UAAU,WAAW,OAAO,QAAQ;GACzD;EACF;EACA;CACF;AACF;AAEA,SAAS,kBAAkB,UAGzB;CACA,MAAM,QAAQ,eAAe,SAAS,KAAK;CAC3C,OAAO;EACL,SAAS;GACP;GACA,SAAS,SAAS,WAAW;GAC7B,OAAO,SAAS;GAChB,YAAY,SAAS;EACvB;EACA,MACE,SAAS,YAAY,KAAA,IACjB;GAAE,MAAM;GAAY,KAAK,SAAS;EAAQ,IAC1C,MAAM,aAAa,SAAS,CAAC,cAAc,SAAS,KAAK,IACvD;GAAE,MAAM;GAAc,KAAK;EAAK,IAChC;GACE,MAAM;GACN,KAAK,aAAa,MAAM,cAAc,MAAM,kBAAkB,SAAS,KAAK;EAC9E;CACV;AACF;AAEA,SAAS,eAAe,OAAoD;CAC1E,MAAM,eAAe,WAAW,OAAO,aAAa;CACpD,MAAM,mBAAmB,WAAW,OAAO,iBAAiB;CAC5D,MAAM,WAAW,iBAAiB,KAAA,KAAa,qBAAqB,KAAA;CACpE,OAAO;EACL,cAAc,gBAAgB;EAC9B,kBAAkB,oBAAoB;EACtC,cAAc,gBAAgB,MAAM,oBAAoB;EACxD;CACF;AACF;AAEA,SAAS,iBAAiB,OAAe,YAAqC;CAC5E,OAAO;EACL,OAAO;GAAE,cAAc;GAAG,kBAAkB;GAAG,aAAa;GAAG,UAAU;EAAM;EAC/E,SAAS;EACT;EACA;CACF;AACF;AAEA,SAAS,WAAW,OAAoC;CACtD,OAAO,OAAO,UAAU,YAAY,OAAO,cAAc,KAAK,KAAK,SAAS,IAAI,QAAQ,KAAA;AAC1F;;;AAIA,SAAgB,iBAAiB,MAAyC;CACxE,OAAO,KAAK,KAAK,MAAM,KAAK,EAAE,IAAI,IAAI,EAAE,aAAa,CAAC,CAAC,KAAK,IAAI;AAClE;;AAGA,SAAgB,iBAAiB,MAAyC;CAExE,OAAO,6DADQ,KAAK,KAAK,MAAM,IAAI,EAAE,IAAI,IAAI,CAAC,CAAC,KAAK,GACqB,EAAE;AAC7E;;;;;AC3LA,SAAgB,mBAAmB,SAAmC;CACpE,MAAM,QAAkB,CAAC;CACzB,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,OAAO,GAAG;EAClD,IAAI,QAAQ,QAAQ,UAAU,KAAA,GAAW;EACzC,MAAM,WACJ,OAAO,UAAU,YAAY,OAAO,UAAU,YAAY,OAAO,UAAU,YACvE,OAAO,KAAK,IACZ,KAAK,UAAU,KAAK;EAC1B,MAAM,KAAK,KAAK,IAAI,IAAI,UAAU;CACpC;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;;;AAIA,SAAgB,uBAAuB,SAAuB,SAAmC;CAC/F,MAAM,QAAQ,mBAAmB,OAAO;CAExC,OAAO;EACL,4BAFuB,QAAQ,eAAe,QAAQ,QAAQ,YAEjB;EAC7C,MAAM,SAAS,IAAI,QAAQ,qCAAqC,QAAQ,GAAG;EAC3E;EACA;CACF,CAAC,CAAC,KAAK,IAAI;AACb;;;AAIA,SAAgB,mCACd,SACA,SACQ;CACR,MAAM,QAAQ,mBAAmB,OAAO;CACxC,MAAM,mBAAmB,QAAQ,eAAe,QAAQ,QAAQ;CAChE,OAAO;EACL,gDAAgD,QAAQ,GAAG,gBAAgB,iBAAiB;EAC5F,MAAM,SAAS,IAAI,iBAAiB,UAAU;EAC9C;EACA;EACA;EACA;EACA;EACA;EACA;CACF,CAAC,CACE,QAAQ,UAAU,MAAM,SAAS,CAAC,CAAC,CACnC,KAAK,IAAI;AACd;;;AAIA,SAAgB,wBACd,SACA,OACoC;CACpC,OAAO;EACL,aAAa,OAAO,iBAAiB,YAAY,uBAAuB,SAAS,OAAO;EACxF,yBACE,OAAO,6BACL,YAAY,mCAAmC,SAAS,OAAO;CACrE;AACF;;;ACoBA,eAAsB,aACpB,MAC0B;CAC1B,MAAM,QAAmB;EAAE,SAAS;EAAG,WAAW,KAAK,IAAI;EAAG,YAAY;CAAM;CAChF,IAAI;EACF,OAAO,MAAM,aAAa,MAAM,KAAK;CACvC,SAAS,KAAK;EAGZ,MAAM,cAAc,KAAK;GACvB,SAAS,MAAM;GACf,YAAY,KAAK,IAAI,IAAI,MAAM;GAC/B,MAAM,MAAM,aAAa,eAAe;EAC1C,CAAC;CACH;AACF;AAEA,eAAe,aACb,MACA,OAC0B;CAC1B,MAAM,SAAS,KAAK,UAAU,oBAAoB;CAClD,MAAM,UAAU,KAAK,WAAW,qBAAqB;CACrD,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,oBAAoB,KAAK,qBAAqB;CACpD,MAAM,aAAa,KAAK,cAAc;CAEtC,MAAM,eAAe,CADD,KAAK,eAAe,sBACL,GAAI,KAAK,wBAAwB,CAAC,CAAE;CACvE,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,wBAAwB,KAAK,yBAAyB;CAC5D,MAAM,kBAAkB,KAAK,mBAAmB;CAEhD,MAAM,SACJ,KAAK,SAAS,KAAK,gBACf;EACE,OAAO,KAAK;EACZ,WAAW,KAAK;EAChB,iBAAiB,KAAK,0BAA0B,KAAA;CAClD,IACA,uBAAuB;CAC7B,MAAM,QAAQ,KAAK,SAAS,OAAO;CACnC,MAAM,YAAY,KAAK,iBAAiB,OAAO;CAC/C,MAAM,kBAAkB,KAAK,mBAAmB,OAAO;CAEvD,MAAM,mBAAuC,QAAQ,iBAAiB;EAAE;EAAQ;EAAS,GAAG;CAAI,CAAC;CACjG,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,kBAAkB,KAAK,mBAAmB;CAEhD,MAAM,QAAQ,wBAAwB,KAAK,SAAS,KAAK,KAAK;CAE9D,MAAM,aAAiC,CAAC;CACxC,MAAM,YAAiC,CAAC;CACxC,IAAI,YAAY;CAEhB,MAAM,SAAS,MAAM,YAAY,KAAK,OAAO;CAC7C,WAAW,KAAK;EAAE,MAAM;EAAQ,SAAS;CAAO,CAAC;CAKjD,MAAM,gBAAgD,CACpD;EAAE,MAAM;EAAU,SAJC,CAAC,KAAK,QAAQ,QAAQ,cAAc,KAAK,QAAQ,QAAQ,kBAAkB,CAAC,CAC9F,QAAQ,UAA2B,QAAQ,KAAK,CAAC,CAAC,CAClD,KAAK,MAEgC;CAAE,GACxC;EAAE,MAAM;EAAQ,SAAS;CAAO,CAClC;CAEA,KAAK,IAAI,OAAO,GAAG,OAAO,UAAU,QAAQ;EAC1C,IAAI,KAAK,QAAQ,SAAS,MAAM,IAAI,MAAM,mBAAmB;EAE7D,IAAI,qBAAqB;EACzB,OAAO,MAAM;GACX,MAAM,EACJ,SAAS,UACT,OAAO,YACP,SAAS,iBACP,MAAM,eAAe;IACvB,OAAO;IACP,UAAU;IACV;IACA,aAAa;IACb,WAAW,uBAAuB,IAAI,iBAAiB;IACvD,QAAQ,KAAK;GACf,CAAC;GACD,MAAM,WAAW,gBAAgB,mBAAmB,YAAY,UAAU;GAC1E,IAAI,iBAAiB,KAAA,KAAa,eAAe,KAAA,GAAW,MAAM,aAAa;GAE/E,MAAM,aAAa,SAAS,WAAW,GAAA,CAAI,KAAK;GAChD,MAAM,kBAAkB,SAAS,cAAc,CAAC,EAAA,CAAG,KAAK,QAAQ;IAC9D,IAAI,GAAG;IACP,MAAM,GAAG,SAAS;IAClB,aAAa;KACX,IAAI;MACF,OAAO,KAAK,MAAM,GAAG,SAAS,SAAS;KACzC,QAAQ;MACN,OAAO,CAAC;KACV;IACF,EAAA,CAAG;GACL,EAAE;GAEF,cAAc,KAAK;IACjB,MAAM;IACN,SAAS,aAAa;IACtB,GAAI,SAAS,YAAY,SAAS,EAAE,YAAY,SAAS,WAAW,IAAI,CAAC;GAC3E,CAAC;GACD,WAAW,KAAK;IACd,MAAM;IACN,SAAS;IACT,WAAW,eAAe,SAAS,IAAI,iBAAiB,KAAA;GAC1D,CAAC;GAED,IAAI,eAAe,WAAW,GAAG;GACjC,sBAAsB,eAAe;GACrC,IAAI,qBAAqB,mBACvB,MAAM,IAAI,MACR,0CAA0C,mBAAmB,GAAG,kBAAkB,YAAY,MAChG;GAGF,KAAK,MAAM,MAAM,gBAAgB;IAC/B;IACA,IAAI,aAAa;IACjB,IAAI;KACF,MAAM,WAAW,UAAU,GAAG;KAC9B,IAAI,CAAC,UACH,aAAa,KAAK,UAAU,EAAE,OAAO,gBAAgB,GAAG,OAAO,CAAC;UAC3D;MACL,MAAM,IAAI,MAAM,SAAS,GAAG,MAAM;OAAE;OAAQ;OAAS,QAAQ,KAAK;MAAO,CAAC;MAC1E,aAAa,EAAE;MACf,MAAM,WAAW,EAAE;MACnB,MAAM,eAAe,gBAAgB,GAAG,IAAI;MAC5C,IAAI,cACF,UAAU,KAAK;OACb,MAAM;OACN;OACA,YAAY;QAAE,MAAM,GAAG;QAAM,MAAM,GAAG;OAAK;OAC3C,SAAS;MACX,CAAC;KAEL;IACF,SAAS,KAAK;KACZ,IAAI,eAAe,yBAAyB,MAAM;KAClD,aAAa,KAAK,UAAU,EAAE,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAE,CAAC;IACzF;IACA,cAAc,KAAK;KAAE,MAAM;KAAQ,cAAc,GAAG;KAAI,SAAS,cAAc;IAAO,CAAC;IACvF,WAAW,KAAK;KAAE,MAAM;KAAQ,SAAS,cAAc;KAAQ,YAAY,GAAG;IAAG,CAAC;GACpF;EACF;EAEA,IAAI,OAAO,WAAW,GAAG;GACvB,MAAM,SAAS,MAAM,WAAW;IAC9B,WAAW;IACX,SAAS,KAAK;IACd;IACA;IACA;IACA,QAAQ;IACR,WAAW;IACX,QAAQ,KAAK;IACb;GACF,CAAC;GACD,cAAc,KAAK;IAAE,MAAM;IAAQ,SAAS,OAAO;GAAQ,CAAC;GAC5D,WAAW,KAAK;IAAE,MAAM;IAAQ,SAAS,OAAO;GAAQ,CAAC;EAC3D;CACF;CAEA,OAAO;EACL;EACA;EACA;EACA,YAAY,KAAK,IAAI,IAAI,MAAM;EAC/B,SAAS,MAAM;EAIf,gBAAgB,MAAM,aAClB;GAAE,MAAM;GAAc,KAAK;EAAK,IAChC;GAAE,MAAM;GAAa,KAAK,MAAM;EAAQ;CAC9C;AACF;AAEA,eAAe,WAA8C,MAa5B;CAM/B,MAAM,iBAAiD,CAAC;EAAE,MAAM;EAAU,SALrD,KAAK,MAAM,wBAAwB,KAAK,OAKiC;CAAE,CAAC;CACjG,KAAK,MAAM,OAAO,KAAK,YAAY;EACjC,IAAI,IAAI,SAAS,QAAQ;EACzB,MAAM,UAAU,qBAAqB,GAAG;EACxC,IAAI,CAAC,SAAS;EACd,IAAI,IAAI,SAAS,aAAa,eAAe,KAAK;GAAE,MAAM;GAAQ;EAAQ,CAAC;OACtE,IAAI,IAAI,SAAS,QAAQ,eAAe,KAAK;GAAE,MAAM;GAAa;EAAQ,CAAC;CAClF;CAGA,KAAK,MAAM,SAAS,KAAK,QACvB,KAAK,IAAI,UAAU,GAAG,UAAU,GAAG,WAAW;EAC5C,MAAM,EAAE,SAAS,OAAO,YAAY,MAAM,KAAK,UAAU;GACvD;GACA,UAAU;GACV,aAAa;GACb,WAAW,KAAK;GAChB,QAAQ,KAAK;EACf,CAAC;EACD,KAAK,MAAM,WAAW,WAAW,mBAAmB,OAAO,KAAK;EAChE,IAAI,YAAY,KAAA,KAAa,UAAU,KAAA,GAAW,KAAK,MAAM,aAAa;EAC1E,MAAM,WAAW,QAAQ,WAAW,GAAA,CAAI,KAAK;EAC7C,IAAI,QAAQ,SAAS,GAAG,OAAO,EAAE,QAAQ;CAC3C;CAEF,MAAM,IAAI,0BAA0B,KAAK,IAAI;AAC/C;AAEA,SAAS,qBAAqB,KAAsC;CAClE,MAAM,OAAO,IAAI,QAAQ,KAAK;CAC9B,IAAI,KAAK,SAAS,GAAG,OAAO;CAC5B,IAAI,IAAI,SAAS,eAAe,CAAC,IAAI,WAAW,QAAQ,OAAO;CAE/D,MAAM,YAAY,IAAI,UAAU,KAAK,SAAS,KAAK,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,OAAO;CAC9E,IAAI,UAAU,WAAW,GAAG,OAAO;CACnC,OAAO,gBAAgB,UAAU,WAAW,IAAI,SAAS,QAAQ,IAAI,UAAU,KAAK,IAAI,EAAE;AAC5F;;;;;;;;AC/KA,SAAS,6BAA6B,QAAkD;CACtF,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,OAAO,OAAO,QAAQ,MAAM,CAAC,EAAE,MAAM;CAC3C,IAAI,KAAK,WAAW,GAAG,OAAO;CAC9B,OAAO,KAAK,QAAQ,KAAK,MAAM,MAAM,EAAE,WAAW,CAAC,IAAI,KAAK;AAC9D;;;;;;AAeA,SAAgB,qBAAqB,OAKnC;CACA,MAAM,gBAA0B,CAAC;CACjC,IAAI,CAAC,MAAM,aAAa,QAAQ,cAAc,KAAK,MAAM,aAAa,SAAS;CAE/E,MAAM,WAAW,MAAM,cAAc,6BAA6B,MAAM,WAAW,IAAI,KAAA;CACvF,IAAI,OAAO,aAAa,UAAU,cAAc,KAAK,QAAQ;CAC7D,MAAM,cAAc,MAAM,iBACtB,6BAA6B,MAAM,cAAc,IACjD,KAAA;CACJ,IAAI,OAAO,gBAAgB,UAAU,cAAc,KAAK,WAAW;CAEnE,OAAO;EACL,WACE,cAAc,WAAW,IACrB,IACA,cAAc,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc;EAC/D,eAAe,YAAY;EAC3B,kBAAkB,eAAe;EACjC,iBAAiB,cAAc,WAAW;CAC5C;AACF;AAMA,eAAsB,mBACpB,MACmC;CACnC,MAAM,YAAY,IAAI,IAAI,KAAK,OAAO,qBAAqB,CAAC,MAAM,CAAC;CACnE,MAAM,eAAe,IAAI,IAAI,KAAK,OAAO,wBAAwB,CAAC,UAAU,CAAC;CAC7E,MAAM,UAAmC,KAAK,WAAW;CACzD,UAAU,KAAK,QAAQ,EAAE,WAAW,KAAK,CAAC;CAE1C,MAAM,SAAS,MAAM,eAAoC;EACvD,MAAM,CACJ;GAAE,MAAM;GAAW,QAAQ,KAAK;EAAS,GACzC;GAAE,MAAM;GAAW,QAAQ,KAAK,SAAS,KAAK,OAAO;IAAE,IAAI,EAAE;IAAI,OAAO;GAAE,EAAE;EAAE,CAChF;EACA,MAAM,KAAK,QAAQ;EACnB,gBAAgB,KAAK,kBAAkB;EACvC,aAAa,KAAK;EAClB,gBAAgB,KAAK;EACrB,MAAM,QAAQ,MAAM;GAClB,MAAM,gBAAgB,KAAK,IAAI;GAC/B,MAAM,UAAU,KAAK,KAAK,SAAS;GACnC,MAAM,UAAU,KAAK,KAAK,SAAS;GACnC,MAAM,YAAY,OAAO,KAAK,KAAK,SAAS,MAAM,SAAS;GAC3D,MAAM,YAAY,OAAO,KAAK,KAAK,SAAS,MAAM,SAAS;GAK3D,MAAM,MAAM,MAAM,QAAQ;IACxB;IACA;IACA,OAAO,KAAK;IACZ,OAAO,KAAK;IACZ,eAAe,KAAK;IACpB,iBAAiB,KAAK;IACtB,UAAU,KAAK;IACf,mBAAmB,KAAK;IACxB,YAAY,KAAK;IACjB,aAAa,KAAK;IAClB,sBAAsB,KAAK;IAC3B,gBAAgB,KAAK;IACrB,uBAAuB,KAAK;IAC5B,iBAAiB,KAAK;IACtB,gBAAgB,KAAK;IACrB,iBAAiB,KAAK;IACtB,QAAQ,KAAK;IACb,SAAS,KAAK;GAChB,CAAC;GAID,MAAM,cAAc,iBAAiB,GAAG;GAGxC,MAAM,mBAAmB,KAAK,gBAAgB,SAAS;GACvD,IAAI,eAAe;GACnB,IAAI,oBAAoB;GAIxB,IAAI,QAA4C;GAChD,IAAI;IACF,0BAA0B,GAAG;IAE7B,MAAM,gBAAgB,IAAI,UAAU,QAAQ,MAAM,UAAU,IAAI,EAAE,IAAI,CAAC;IACvE,MAAM,mBAAmB,IAAI,UAAU,QAAQ,MAAM,aAAa,IAAI,EAAE,IAAI,CAAC;IAE7E,QAAQ;IACR,MAAM,CAAC,iBAAiB,gBAAgB,qBAAqB,MAAM,QAAQ,IAAI;KAC7E,SAAS,mBAAmB,KAAK,OAAO,cAAc,KAAK,cAAc,GAAG;MAC1E,YAAY,IAAI;MAChB;KACF,CAAC;KACD,KAAK,OAAO,aACR,QAAQ,IACN,cAAc,KAAK,aACjB,SAAS,mBAAmB,KAAK,OAAO,YAAa,KAAK,cAAc,GAAG;MACzE;MACA;KACF,CAAC,CAAC,CAAC,MAAM,YAAY;MACnB,OAAO;OACL,GAAG,OAAO;OACV,MAAM,SAAS;OACf,MAAM,SAAS;MACjB;MACA,MAAM,OAAO;KACf,EAAE,CACJ,CACF,IACA,QAAQ,QAAQ,CAAC,CAAuB;KAC5C,KAAK,OAAO,iBACR,QAAQ,IACN,iBAAiB,KAAK,aACpB,SAAS,mBAAmB,KAAK,OAAO,gBAAiB,KAAK,cAAc,GAAG;MAC7E;MACA;KACF,CAAC,CAAC,CAAC,MAAM,YAAY;MACnB,OAAO;OACL,GAAG,OAAO;OACV,MAAM,SAAS;OACf,MAAM,SAAS;MACjB;MACA,MAAM,OAAO;KACf,EAAE,CACJ,CACF,IACA,QAAQ,QAAQ,CAAC,CAAuB;IAC9C,CAAC;IACD,MAAM,YAAY;KAAC;KAAiB,GAAG;KAAgB,GAAG;IAAiB;IAC3E,eAAe,UAAU,QAAQ,KAAK,QAAQ,OAAO,IAAI,KAAK,OAAO,IAAI,CAAC;IAC1E,oBAAoB,UAAU,OAAO,QAAQ,IAAI,KAAK,SAAS,YAAY;IAC3E,QAAQ;IAER,MAAM,eAAe,gBAAgB;IACrC,MAAM,cAAc,eAAe,KAAK,QAAQ,IAAI,KAAK;IACzD,MAAM,iBAAiB,kBAAkB,KAAK,QAAQ,IAAI,KAAK;IAE/D,MAAM,EAAE,WAAW,eAAe,kBAAkB,oBAClD,qBAAqB;KACnB;KACA,aAAa,KAAK,OAAO,aAAa,cAAc,KAAA;KACpD,gBAAgB,KAAK,OAAO,iBAAiB,iBAAiB,KAAA;IAChE,CAAC;IAEH,MAAM,YAAgC;KAAE;KAAW;IAAa;IAChE,IAAI,KAAK,OAAO,YACd,UAAU,aAAa;KAAE,aAAa;KAAa,WAAW;IAAc;IAC9E,IAAI,KAAK,OAAO,gBACd,UAAU,iBAAiB;KAAE,aAAa;KAAgB,WAAW;IAAiB;IAExF,MAAM,UAAU,KAAK,KAAK,QAAQ,WAAW,WAAW,OAAO,KAAK,KAAK;IACzE,UAAU,SAAS,EAAE,WAAW,KAAK,CAAC;IACtC,cAAc,KAAK,SAAS,iBAAiB,GAAG,KAAK,UAAU,IAAI,YAAY,MAAM,CAAC,CAAC;IACvF,cAAc,KAAK,SAAS,gBAAgB,GAAG,KAAK,UAAU,IAAI,WAAW,MAAM,CAAC,CAAC;IACrF,cAAc,KAAK,SAAS,aAAa,GAAG,KAAK,UAAU,WAAW,MAAM,CAAC,CAAC;IAE9E,MAAM,QAAQ,CAAC,SAAS,aAAa,UAAU,QAAQ,CAAC,GAAG;IAC3D,IAAI,KAAK,OAAO,YAAY,MAAM,KAAK,QAAQ,cAAc,QAAQ,CAAC,GAAG;IACzE,IAAI,KAAK,OAAO,gBAAgB,MAAM,KAAK,WAAW,iBAAiB,QAAQ,CAAC,GAAG;IACnF,IAAI,iBAAiB,MAAM,KAAK,mBAAmB;IACnD,IAAI,CAAC,mBAAmB,MAAM,KAAK,uBAAuB;IAE1D,OAAO;KACL,QAAQ;MACN,OAAO,IAAI,WAAW;MACtB,WAAW,IAAI;MACf,eAAe,IAAI,UAAU;KAC/B;KACA,SAAS;MAAE,OAAO,aAAa;MAAG,OAAO;MAAW,OAAO,MAAM,KAAK,GAAG;KAAE;KAC3E,SAAS,IAAI,UAAU;KAMvB,gBACE,qBAAqB,mBACjB;MAAE,MAAM;MAAa,KAAK,IAAI,UAAU;KAAa,IACrD;MAAE,MAAM;MAAc,KAAK;KAAK;KACtC,YAAY,IAAI;IAClB;GACF,SAAS,KAAK;IAIZ,IAAI,gBAAgB,KAAA,GAAW,MAAM;IACrC,MAAM,cAAc,KAAK;KACvB,SAAS,cAAc;KACvB,YAAY,KAAK,IAAI,IAAI;KAKzB,MACE,qBAAqB,UAAU,cAAe,UAAU,aAAa,qBACjE,cACA;IACR,CAAC;GACH;EACF;CACF,CAAC;CAGD,MAAM,UAAU;EACd,OAAO,OAAO,QAAQ;EACtB,UAAU,OAAO,QAAQ;EACzB,WAAW,OAAO,QAAQ;EAC1B,cAAc,OAAO,QAAQ;EAC7B,qBAAqB,OAAO,QAAQ;EACpC,mBAAmB,OAAO,QAAQ;EAClC,YAAY,OAAO,QAAQ;EAC3B,cAAc,OAAO,QAAQ;EAC7B,cAAc,OAAO,QAAQ;EAC7B,WAAW,OAAO,OAAO;EACzB,WAAW,OAAO,OAAO;CAC3B;CACA,cAAc,KAAK,KAAK,QAAQ,cAAc,GAAG,KAAK,UAAU,SAAS,MAAM,CAAC,CAAC;CAIjF,MAAM,aAAa,OAAO,QAAQ;CAClC,MAAM,YAAY,aAAa,IAAI,oBAAoB;CAEvD,MAAM,KAAe;EACnB;EACA;EACA,cAAc,OAAO,QAAQ,WAAW,qBAAqB,OAAO,QAAQ,kBAAkB,IAAA,CAAK,QAAQ,CAAC,EAAE,gBAAgB,OAAO,QAAQ,iBAAiB,QAAQ,CAAC,EAAE,OAAO,UAAU,OAAO,OAAO,QAAQ,aAAa,QAAQ,CAAC,EAAE,oBAAoB,OAAO,QAAQ,aAAa,IAAA,CAAM,QAAQ,CAAC,EAAE;EACzS;EACA,GAAI,aAAa,IACb,CACE,KAAK,WAAW,MAAM,OAAO,QAAQ,aAAa,mGAClD,EACF,IACA,CAAC;EACL;EACA;EACA;EACA;EACA,GAAG,OAAO,QAAQ,OAAO,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,KAC5C,CAAC,IAAI,OACJ,KAAK,GAAG,MAAM,EAAE,WAAW,IAAA,CAAK,QAAQ,CAAC,EAAE,MAAM,EAAE,UAAU,QAAQ,CAAC,EAAE,MAAM,EAAE,aAAa,QAAQ,CAAC,EAAE,GAC5G;EACA;EACA;EACA;EACA;EACA;EACA,GAAG,OAAO,QAAQ,OAAO,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,KAC5C,CAAC,IAAI,OACJ,KAAK,GAAG,MAAM,EAAE,WAAW,IAAA,CAAK,QAAQ,CAAC,EAAE,MAAM,EAAE,UAAU,QAAQ,CAAC,EAAE,MAAM,EAAE,aAAa,QAAQ,CAAC,EAAE,GAC5G;EACA;CACF;CACA,cAAc,KAAK,KAAK,QAAQ,YAAY,GAAG,GAAG,KAAK,IAAI,CAAC;CAE5D,OAAO,EAAE,OAAO;AAClB;;;;;AAMA,SAAS,iBACP,KACoB;CACpB,MAAM,QAAQ,KAAK;CACnB,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS,IAAI,QAAQ,KAAA;AACrF;AAEA,SAAS,mBACP,OACA,WACqB;CACrB,IAAI,cAAc,KAAA,KAAa,MAAM,cAAc,KAAA,GAAW,OAAO;CACrE,OAAO;EAAE,GAAG;EAAO;CAAU;AAC/B"}
|
|
1
|
+
{"version":3,"file":"index.js","names":[],"sources":["../../src/multishot/router.ts","../../src/multishot/default-tools.ts","../../src/multishot/judges.ts","../../src/multishot/shape-defaults.ts","../../src/multishot/multishot.ts","../../src/multishot/matrix.ts"],"sourcesContent":["// Router fetch helper — single source of truth for OpenAI-compat calls\n// against the Tangle router. Used by the driver, agent, judges, and the\n// default tool executors.\n\nimport type { MultishotToolDefinition } from './types'\n\nexport interface RouterCompletionRequest {\n apiKey: string\n baseUrl: string\n model: string\n messages: Array<Record<string, unknown>>\n tools?: MultishotToolDefinition[]\n temperature?: number\n maxTokens?: number\n signal?: AbortSignal\n}\n\nexport interface RouterToolCall {\n id: string\n type: 'function'\n function: { name: string; arguments: string }\n}\n\nexport interface RouterCompletionResponse {\n message: { content?: string | null; tool_calls?: RouterToolCall[] }\n usage?: { prompt_tokens?: number; completion_tokens?: number }\n /** Provider-reported spend when the endpoint supplies it. */\n costUsd?: number\n /** Model echoed by the provider, falling back to the requested model. */\n model: string\n durationMs: number\n}\n\nexport async function routerCompletion(\n req: RouterCompletionRequest,\n): Promise<RouterCompletionResponse> {\n const startedAt = Date.now()\n const body: Record<string, unknown> = {\n model: req.model,\n messages: req.messages,\n temperature: req.temperature ?? 0.7,\n max_tokens: req.maxTokens ?? 2000,\n }\n if (req.tools?.length) body.tools = req.tools\n const url = `${req.baseUrl.replace(/\\/+$/, '')}/chat/completions`\n const res = await fetch(url, {\n method: 'POST',\n headers: { Authorization: `Bearer ${req.apiKey}`, 'Content-Type': 'application/json' },\n body: JSON.stringify(body),\n signal: req.signal,\n })\n if (!res.ok) {\n const text = await res.text()\n throw new Error(`router ${res.status}: ${text.slice(0, 300)}`)\n }\n const json = (await res.json()) as {\n choices: Array<{ message: { content?: string | null; tool_calls?: RouterToolCall[] } }>\n usage?: { prompt_tokens?: number; completion_tokens?: number }\n model?: unknown\n _response_cost?: unknown\n cost_usd?: unknown\n }\n const choice = json.choices[0]\n if (!choice) throw new Error(`router returned no choices: ${JSON.stringify(json).slice(0, 200)}`)\n const rawCost = json._response_cost ?? json.cost_usd\n const costUsd =\n typeof rawCost === 'number' && Number.isFinite(rawCost) && rawCost >= 0 ? rawCost : undefined\n return {\n message: choice.message,\n usage: json.usage,\n ...(costUsd === undefined ? {} : { costUsd }),\n model: typeof json.model === 'string' && json.model ? json.model : req.model,\n durationMs: Date.now() - startedAt,\n }\n}\n\n// Rough per-model cost estimator. Used for cost-ceiling enforcement.\n// Underestimates Anthropic, overestimates oss models — fine for ceilings.\nexport function estimateRouterCost(\n model: string,\n usage?: { prompt_tokens?: number; completion_tokens?: number },\n): number {\n if (!usage) return 0\n const inputTok = usage.prompt_tokens ?? 0\n const outputTok = usage.completion_tokens ?? 0\n let inPer1k = 0.003\n let outPer1k = 0.015\n if (model.includes('gpt-4o-mini')) {\n inPer1k = 0.00015\n outPer1k = 0.0006\n } else if (model.includes('gpt-5.4') || model.includes('claude-sonnet')) {\n inPer1k = 0.003\n outPer1k = 0.015\n } else if (model.includes('kimi') || model.includes('glm') || model.includes('deepseek')) {\n inPer1k = 0.0005\n outPer1k = 0.002\n }\n return (inputTok * inPer1k + outputTok * outPer1k) / 1000\n}\n\nexport function defaultRouterBaseUrl(): string {\n return (process.env.TANGLE_ROUTER_BASE_URL ?? 'https://router.tangle.tools/v1').replace(\n /\\/+$/,\n '',\n )\n}\n\nexport function requireRouterApiKey(): string {\n const key = process.env.TANGLE_API_KEY\n if (!key) throw new Error('multishot requires TANGLE_API_KEY (router-scoped sk-tan-* key)')\n return key\n}\n","// Default delegate_research + delegate_code tools and their inline executors.\n//\n// Consumers can override either by passing their own tools + executors to\n// runMultishot. The defaults are sufficient for most domains — point the\n// researcher system prompt at your domain's citation style and the coder\n// at your preferred language.\n\nimport { estimateRouterCost, routerCompletion } from './router'\nimport type { MultishotToolDefinition, MultishotToolExecutor } from './types'\n\nexport const DEFAULT_RESEARCHER_MODEL = 'openai/gpt-4o-mini'\nexport const DEFAULT_CODER_MODEL = 'openai/gpt-4o-mini'\n\nexport interface DefaultResearcherConfig {\n /** Replace the system prompt to bias the researcher toward a domain's\n * citation style. Defaults to a generic \"cite sources by name\" prompt. */\n systemPrompt?: string\n model?: string\n}\n\nexport interface DefaultCoderConfig {\n /** Replace the system prompt to bias the coder toward a language /\n * framework / artifact style. */\n systemPrompt?: string\n model?: string\n}\n\nconst GENERIC_RESEARCHER_SYSTEM =\n 'You are a research specialist. Return a markdown brief with 3-5 findings. Each finding cites a specific source by name. Add a confidence level (high/medium/low) per finding. No fluff, no preamble.'\n\nconst GENERIC_CODER_SYSTEM =\n 'You are an expert engineer. Output ONE fenced code block containing the complete solution. Inline-comment non-obvious decisions. No explanation outside the block.'\n\nexport const DEFAULT_DELEGATE_RESEARCH_TOOL: MultishotToolDefinition = {\n type: 'function',\n function: {\n name: 'delegate_research',\n description:\n 'Research a topic deeply via specialist. Returns evidence-bearing items with citations. Use for audience research, competitive intel, regulatory landscape, market data, citation-grounded analysis.',\n parameters: {\n type: 'object',\n properties: {\n question: { type: 'string', description: 'Specific question to research' },\n scope: {\n type: 'string',\n description: 'Optional scope: time window, geography, jurisdiction, segment',\n },\n },\n required: ['question'],\n },\n },\n}\n\nexport const DEFAULT_DELEGATE_CODE_TOOL: MultishotToolDefinition = {\n type: 'function',\n function: {\n name: 'delegate_code',\n description:\n 'Generate a runnable script, template, pipeline, or tool via specialist. Returns complete working code or structured markdown. Use for content pipelines, calc snippets, dashboards, compliance checklists, deadline trackers.',\n parameters: {\n type: 'object',\n properties: {\n goal: { type: 'string', description: 'What the code must accomplish' },\n language: {\n type: 'string',\n description: 'Optional language preference (default: TypeScript)',\n },\n },\n required: ['goal'],\n },\n },\n}\n\nexport function createResearchExecutor(\n config: DefaultResearcherConfig = {},\n): MultishotToolExecutor {\n const systemPrompt = config.systemPrompt ?? GENERIC_RESEARCHER_SYSTEM\n const model = config.model ?? DEFAULT_RESEARCHER_MODEL\n return async (args, ctx) => {\n const question = String(args.question ?? '')\n const scope = args.scope ? String(args.scope) : undefined\n const { message, usage } = await routerCompletion({\n apiKey: ctx.apiKey,\n baseUrl: ctx.baseUrl,\n model,\n temperature: 0.3,\n maxTokens: 1800,\n messages: [\n { role: 'system', content: systemPrompt },\n { role: 'user', content: `Research: ${question}${scope ? `\\nScope: ${scope}` : ''}` },\n ],\n signal: ctx.signal,\n })\n return { content: message.content ?? '', costUsd: estimateRouterCost(model, usage) }\n }\n}\n\nexport function createCodeExecutor(config: DefaultCoderConfig = {}): MultishotToolExecutor {\n const systemPrompt = config.systemPrompt ?? GENERIC_CODER_SYSTEM\n const model = config.model ?? DEFAULT_CODER_MODEL\n return async (args, ctx) => {\n const goal = String(args.goal ?? '')\n const language = args.language ? String(args.language) : 'TypeScript'\n const { message, usage } = await routerCompletion({\n apiKey: ctx.apiKey,\n baseUrl: ctx.baseUrl,\n model,\n temperature: 0.2,\n maxTokens: 2000,\n messages: [\n { role: 'system', content: `${systemPrompt}\\n\\nLanguage: ${language}` },\n { role: 'user', content: `Produce: ${goal}` },\n ],\n signal: ctx.signal,\n })\n return { content: message.content ?? '', costUsd: estimateRouterCost(model, usage) }\n }\n}\n\nexport interface DefaultToolsConfig {\n research?: DefaultResearcherConfig\n code?: DefaultCoderConfig\n /** When true (default), each tool result is recorded as a typed artifact:\n * research → type='research', code → type='code'. */\n recordArtifacts?: boolean\n}\n\nexport interface DefaultToolsBundle {\n tools: MultishotToolDefinition[]\n executors: Record<string, MultishotToolExecutor>\n artifactTypeFor: (toolName: string) => string | undefined\n}\n\nexport function defaultDelegationTools(config: DefaultToolsConfig = {}): DefaultToolsBundle {\n return {\n tools: [DEFAULT_DELEGATE_RESEARCH_TOOL, DEFAULT_DELEGATE_CODE_TOOL],\n executors: {\n delegate_research: createResearchExecutor(config.research),\n delegate_code: createCodeExecutor(config.code),\n },\n artifactTypeFor: (name) =>\n name === 'delegate_research' ? 'research' : name === 'delegate_code' ? 'code' : undefined,\n }\n}\n\nexport { defaultRouterBaseUrl } from './router'\n","// Generic judge runner — domain consumers configure dimensions + prompts.\n//\n// Three judge slots are conventional for multishot eval:\n// - conversation (scores the full transcript)\n// - codeReview (scores each code artifact)\n// - contentQuality (scores each non-code artifact)\n//\n// But the runJudge primitive is fully generic: any input maps to a score plus\n// explicit observed, estimated, or uncaptured cost.\n\nimport type { JudgeScore } from '../campaign/types'\nimport type { CostProvenance } from '../cost-ledger'\nimport type { LlmCallMetadata, LlmUsage } from '../llm-client'\nimport { estimateCost, isModelPriced } from '../metrics'\nimport {\n defaultRouterBaseUrl,\n type RouterCompletionResponse,\n requireRouterApiKey,\n routerCompletion,\n} from './router'\n\n// Canonical declaration lives in campaign/types.ts. Multishot emits the same\n// shape on its producer-defined 0-10 scale.\nexport type { JudgeScore } from '../campaign/types'\n\nexport const DEFAULT_JUDGE_MODEL = 'openai/gpt-4o-mini'\n\nexport interface JudgeDimension {\n /** JSON field name + score key. */\n key: string\n /** Description shown in the judge's user prompt. */\n description: string\n}\n\nexport interface JudgeConfig<TInput> {\n /** Display name (for trace + log). */\n name: string\n /** Model used for this judge. */\n model?: string\n /** 0-10 scored dimensions. */\n dimensions: JudgeDimension[]\n /** Judge system prompt — sets persona + JSON-only constraint. */\n systemPrompt: string\n /** Build the user prompt from the typed input. Must include \"Respond with\n * ONLY this JSON: { ... }\" listing each dimension key. */\n buildPrompt: (input: TInput) => string\n /** Optional model + api overrides. */\n apiKey?: string\n baseUrl?: string\n /** Maximum output tokens for the judge response. Defaults to 1500. */\n maxTokens?: number\n}\n\nexport interface JudgeRunResult {\n /** Semantic result; failed scores remain non-throwing so matrix aggregation can exclude them. */\n score: JudgeScore\n /** Cost of the completed call, separate from diagnostic provider metadata. */\n cost: CostProvenance\n}\n\nexport async function runJudge<TInput>(\n judge: JudgeConfig<TInput>,\n input: TInput,\n): Promise<JudgeRunResult> {\n const apiKey = judge.apiKey ?? requireRouterApiKey()\n const baseUrl = judge.baseUrl ?? defaultRouterBaseUrl()\n const model = judge.model ?? process.env.JUDGE_MODEL ?? DEFAULT_JUDGE_MODEL\n const prompt = judge.buildPrompt(input)\n let raw = ''\n let llmCall: LlmCallMetadata\n let cost: CostProvenance\n const startedAt = Date.now()\n try {\n const response = await routerCompletion({\n apiKey,\n baseUrl,\n model,\n temperature: 0,\n maxTokens: judge.maxTokens ?? 1500,\n messages: [\n { role: 'system', content: judge.systemPrompt },\n { role: 'user', content: prompt },\n ],\n })\n const call = judgeCallMetadata(response)\n llmCall = call.llmCall\n cost = call.cost\n raw = (response.message.content ?? '').trim()\n } catch (err) {\n // failed:true lets consumers reading `.composite` keep working while\n // aggregators exclude this score from means instead of averaging a zero.\n return {\n score: {\n dimensions: {},\n composite: 0,\n failed: true,\n notes: `judge ${judge.name} call failed: ${err instanceof Error ? err.message : String(err)}`,\n llmCall: unknownJudgeCall(model, Date.now() - startedAt),\n },\n cost: { kind: 'uncaptured', usd: null },\n }\n }\n\n let parsed: Record<string, unknown> | null = null\n try {\n const cleaned = raw\n .replace(/^```json\\s*/i, '')\n .replace(/```\\s*$/, '')\n .trim()\n parsed = JSON.parse(cleaned) as Record<string, unknown>\n } catch {\n return {\n score: {\n dimensions: {},\n composite: 0,\n failed: true,\n notes: `judge ${judge.name} returned non-JSON: ${raw.slice(0, 200)}`,\n llmCall,\n },\n cost,\n }\n }\n\n const dimensions: Record<string, number> = {}\n let sum = 0\n for (const dim of judge.dimensions) {\n const v = Number(parsed[dim.key] ?? 0)\n const clamped = Number.isFinite(v) ? Math.max(0, Math.min(10, v)) : 0\n dimensions[dim.key] = clamped\n sum += clamped\n }\n return {\n score: {\n dimensions,\n composite: judge.dimensions.length === 0 ? 0 : sum / judge.dimensions.length,\n notes: typeof parsed.notes === 'string' ? parsed.notes : '',\n llmCall,\n },\n cost,\n }\n}\n\nfunction judgeCallMetadata(response: RouterCompletionResponse): {\n llmCall: LlmCallMetadata\n cost: CostProvenance\n} {\n const usage = canonicalUsage(response.usage)\n return {\n llmCall: {\n usage,\n costUsd: response.costUsd ?? null,\n model: response.model,\n durationMs: response.durationMs,\n },\n cost:\n response.costUsd !== undefined\n ? { kind: 'observed', usd: response.costUsd }\n : usage.captured === false || !isModelPriced(response.model)\n ? { kind: 'uncaptured', usd: null }\n : {\n kind: 'estimated',\n usd: estimateCost(usage.promptTokens, usage.completionTokens, response.model),\n },\n }\n}\n\nfunction canonicalUsage(usage: RouterCompletionResponse['usage']): LlmUsage {\n const promptTokens = tokenCount(usage?.prompt_tokens)\n const completionTokens = tokenCount(usage?.completion_tokens)\n const captured = promptTokens !== undefined && completionTokens !== undefined\n return {\n promptTokens: promptTokens ?? 0,\n completionTokens: completionTokens ?? 0,\n totalTokens: (promptTokens ?? 0) + (completionTokens ?? 0),\n captured,\n }\n}\n\nfunction unknownJudgeCall(model: string, durationMs: number): LlmCallMetadata {\n return {\n usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0, captured: false },\n costUsd: null,\n model,\n durationMs,\n }\n}\n\nfunction tokenCount(value: unknown): number | undefined {\n return typeof value === 'number' && Number.isSafeInteger(value) && value >= 0 ? value : undefined\n}\n\n/** Convenience: stringified dimension list for inclusion in a judge prompt.\n * Returns lines like `- audience_fit: Does this match what the audience cares about? (0-10)`. */\nexport function renderDimensions(dims: readonly JudgeDimension[]): string {\n return dims.map((d) => `- ${d.key}: ${d.description}`).join('\\n')\n}\n\n/** Convenience: build the \"Respond with ONLY this JSON\" footer for a judge prompt. */\nexport function renderJsonFooter(dims: readonly JudgeDimension[]): string {\n const fields = dims.map((d) => `\"${d.key}\":N`).join(',')\n return `Respond with ONLY this JSON (no markdown, no preamble):\\n{${fields},\"notes\":\"1-2 sentence critique\"}`\n}\n","// Profile-derived defaults for MultishotShape — the pure-profile path.\n//\n// `MultishotShape`'s callbacks were REQUIRED beside `profile: AgentProfile`,\n// which mixed the abstractions: the node under test was a profile, but the\n// simulated user's role could only be expressed as code. These defaults make\n// the callbacks optional — a pure-profile call works, and the derived prompts\n// are plain data (a pure function of profile + persona payload), so an\n// optimizer can sweep them (tangle-network/agent-runtime#694).\n\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport type { MultishotPersona, MultishotShape } from './types'\n\n/** Persona payload rendered as stable `- key: value` lines. `id` is identity,\n * not voice, and structured values are serialized so nothing is dropped. */\nexport function renderPersonaFacts(persona: MultishotPersona): string {\n const lines: string[] = []\n for (const [key, value] of Object.entries(persona)) {\n if (key === 'id' || value === undefined) continue\n const rendered =\n typeof value === 'string' || typeof value === 'number' || typeof value === 'boolean'\n ? String(value)\n : JSON.stringify(value)\n lines.push(`- ${key}: ${rendered}`)\n }\n return lines.join('\\n')\n}\n\n/** Default opening user message: the persona introduces itself with its whole\n * payload and asks the agent (as described by the profile) to take over. */\nexport function defaultMultishotOpener(profile: AgentProfile, persona: MultishotPersona): string {\n const facts = renderPersonaFacts(persona)\n const agentDescription = profile.description ?? profile.name ?? 'the agent'\n return [\n `Hi — I'm reaching out to ${agentDescription} for help. My situation:`,\n facts.length > 0 ? facts : `- (no details provided beyond id: ${persona.id})`,\n '',\n 'Tell me what you need from me, and get started on my case.',\n ].join('\\n')\n}\n\n/** Default driver system prompt: roleplay the persona from its payload, stay\n * in character, push back on vague answers, and never go silent. */\nexport function defaultMultishotDriverSystemPrompt(\n profile: AgentProfile,\n persona: MultishotPersona,\n): string {\n const facts = renderPersonaFacts(persona)\n const agentDescription = profile.description ?? profile.name ?? 'an AI agent'\n return [\n `You are role-playing a real person (persona \"${persona.id}\") talking to ${agentDescription}.`,\n facts.length > 0 ? `Who you are:\\n${facts}` : '',\n '',\n 'How to behave:',\n '- Stay in character: first person, realistic voice, concrete details from your situation.',\n '- Judge every reply by whether it actually moves YOUR case forward. Push back on vague, hedged, or generic answers and demand specifics.',\n \"- Answer the agent's questions with information consistent with who you are; invent plausible details rather than stalling.\",\n '- Never go silent: always produce a substantive next message.',\n '- Output ONLY your next message to the agent — no meta-commentary, no stage directions.',\n ]\n .filter((block) => block.length > 0)\n .join('\\n')\n}\n\n/** The fully-resolved shape for a profile: caller overrides win, the\n * profile-derived defaults fill the rest. */\nexport function defaultShapeFromProfile<TPersona extends MultishotPersona>(\n profile: AgentProfile,\n shape?: MultishotShape<TPersona>,\n): Required<MultishotShape<TPersona>> {\n return {\n buildOpener: shape?.buildOpener ?? ((persona) => defaultMultishotOpener(profile, persona)),\n buildDriverSystemPrompt:\n shape?.buildDriverSystemPrompt ??\n ((persona) => defaultMultishotDriverSystemPrompt(profile, persona)),\n }\n}\n","// Multi-turn driver-agent simulation with inline tool execution.\n//\n// The driver = LLM acting as the persona (reactive, non-deterministic).\n// The agent = the product agent under test (router call by default, or an\n// injected transport — with profile's systemPrompt + the configured tools).\n// Tool calls execute inline via the configured executors and feed back\n// into the agent's message log so the agent integrates the result.\n\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport { withCellSpend } from '../matrix'\nimport { defaultDelegationTools } from './default-tools'\nimport {\n defaultRouterBaseUrl,\n estimateRouterCost,\n requireRouterApiKey,\n routerCompletion,\n} from './router'\nimport { defaultShapeFromProfile } from './shape-defaults'\nimport {\n type MultishotArtifact,\n MultishotDriverEmptyError,\n MultishotFatalToolError,\n type MultishotMessage,\n type MultishotPersona,\n type MultishotResult,\n type MultishotShape,\n type MultishotToolDefinition,\n type MultishotToolExecutor,\n type MultishotTransport,\n} from './types'\n\nexport interface RunMultishotOptions<TPersona extends MultishotPersona> {\n profile: AgentProfile\n persona: TPersona\n /** Persona-shaping callbacks. Optional — omitted callbacks are derived from\n * the profile + persona payload, so a pure-profile call works. */\n shape?: MultishotShape<TPersona>\n /** Tool definitions advertised to the agent. Defaults to delegate_research + delegate_code. */\n tools?: MultishotToolDefinition[]\n /** Map from tool name → executor invoked inline when the agent emits a tool_call. */\n toolExecutors?: Record<string, MultishotToolExecutor>\n /** Map from tool name → artifact type label written into MultishotArtifact.type.\n * Tools without a mapping still execute, but their results aren't surfaced as\n * typed artifacts (only as tool messages in the transcript). */\n artifactTypeFor?: (toolName: string) => string | undefined\n maxTurns?: number\n agentModel?: string\n driverModel?: string\n /** Fallback driver models tried when the primary simulated-user model returns empty twice. */\n driverFallbackModels?: string[]\n /** Maximum output tokens for the first agent call in each assistant turn. */\n agentMaxTokens?: number\n /** Maximum output tokens for agent follow-up calls after tool results. */\n toolFollowupMaxTokens?: number\n /** Maximum output tokens for each simulated-user driver response. */\n driverMaxTokens?: number\n /** Maximum tool calls the agent may dispatch inside one assistant turn. */\n maxToolDispatches?: number\n /** Execution seam for the agent leg. When provided, every agent inference\n * step goes through this function instead of the router HTTP call; the\n * string levers (agentModel, apiKey, baseUrl) stop applying to that leg.\n * apiKey/baseUrl are still resolved for tool executors and any leg\n * without an injected transport. */\n agentTransport?: MultishotTransport\n /** Execution seam for the simulated-user driver leg (symmetric to\n * agentTransport). Driver model fallback rotation still applies — the\n * transport receives each candidate model in turn. */\n driverTransport?: MultishotTransport\n apiKey?: string\n baseUrl?: string\n signal?: AbortSignal\n}\n\n/** One multishot shot — the conversation engine `runMultishotMatrix` invokes\n * once per cell. `runMultishot` is the default implementation.\n *\n * An alternative engine (a graph-backed conversation, a replay of a recorded\n * transcript, a sandbox-hosted agent) implements this exact signature and\n * reaches the matrix through `RunMultishotMatrixOptions.runShot`. The matrix\n * keeps every other cell mechanic — cell fan-out, concurrency, the cost\n * ceiling, the judge slots, the cell composite, and the per-cell writers — so\n * swapping the engine needs no copy of the cell body. */\nexport type MultishotShot<TPersona extends MultishotPersona> = (\n opts: RunMultishotOptions<TPersona>,\n) => Promise<MultishotResult>\n\n/** Running spend of one shot, readable after the shot throws. */\ninterface ShotMeter {\n costUsd: number\n startedAt: number\n /** True once a call contributed a fabricated `0` — no provider cost and no\n * usage to price. `costUsd` is a subtotal from that point on. */\n uncaptured: boolean\n}\n\n/** The package's reference conversation engine, and the referent of three\n * things that stop working without it.\n *\n * 1. `runMultishotMatrix` runs it as the per-cell engine when the caller names\n * no `runShot`, so the matrix always has an engine to run.\n * 2. `golden.test.ts` proves the frozen records against it on every run, and\n * perturbs it to prove the records detect a change. Records checked only\n * against a replay of themselves prove nothing.\n * 3. The repo's recording command names it as `--engine`. The script defaults\n * no engine, so a version records whichever one the caller names — and a\n * record set minted from a consumer's engine would let a product's wiring\n * define this package's frozen contract.\n *\n * An alternative engine reaches the matrix through `runShot` and reaches the\n * records through `assertMultishotGoldenScenario`. Neither seam removes the\n * three roles above. */\nexport async function runMultishot<TPersona extends MultishotPersona>(\n opts: RunMultishotOptions<TPersona>,\n): Promise<MultishotResult> {\n const meter: ShotMeter = { costUsd: 0, startedAt: Date.now(), uncaptured: false }\n try {\n return await runShotTurns(opts, meter)\n } catch (err) {\n // The turns already paid for. Declare that spend so the matrix bills the\n // failed cell for it and the cost ceiling sees the real cumulative total.\n throw withCellSpend(err, {\n costUsd: meter.costUsd,\n durationMs: Date.now() - meter.startedAt,\n kind: meter.uncaptured ? 'uncaptured' : 'estimated',\n })\n }\n}\n\nasync function runShotTurns<TPersona extends MultishotPersona>(\n opts: RunMultishotOptions<TPersona>,\n meter: ShotMeter,\n): Promise<MultishotResult> {\n const apiKey = opts.apiKey ?? requireRouterApiKey()\n const baseUrl = opts.baseUrl ?? defaultRouterBaseUrl()\n const maxTurns = opts.maxTurns ?? 10\n const maxToolDispatches = opts.maxToolDispatches ?? 4\n const agentModel = opts.agentModel ?? 'openai/gpt-5.4'\n const driverModel = opts.driverModel ?? 'openai/gpt-4o-mini'\n const driverModels = [driverModel, ...(opts.driverFallbackModels ?? [])]\n const agentMaxTokens = opts.agentMaxTokens ?? 2500\n const toolFollowupMaxTokens = opts.toolFollowupMaxTokens ?? 2000\n const driverMaxTokens = opts.driverMaxTokens ?? 600\n\n const bundle =\n opts.tools && opts.toolExecutors\n ? {\n tools: opts.tools,\n executors: opts.toolExecutors,\n artifactTypeFor: opts.artifactTypeFor ?? (() => undefined),\n }\n : defaultDelegationTools()\n const tools = opts.tools ?? bundle.tools\n const executors = opts.toolExecutors ?? bundle.executors\n const artifactTypeFor = opts.artifactTypeFor ?? bundle.artifactTypeFor\n\n const routerTransport: MultishotTransport = (req) => routerCompletion({ apiKey, baseUrl, ...req })\n const agentTransport = opts.agentTransport ?? routerTransport\n const driverTransport = opts.driverTransport ?? routerTransport\n\n const shape = defaultShapeFromProfile(opts.profile, opts.shape)\n\n const transcript: MultishotMessage[] = []\n const artifacts: MultishotArtifact[] = []\n let toolCalls = 0\n\n const opener = shape.buildOpener(opts.persona)\n transcript.push({ role: 'user', content: opener })\n\n const systemPrompt = [opts.profile.prompt?.systemPrompt, opts.profile.prompt?.appendSystemPrompt]\n .filter((value): value is string => Boolean(value))\n .join('\\n\\n')\n const agentMessages: Array<Record<string, unknown>> = [\n { role: 'system', content: systemPrompt },\n { role: 'user', content: opener },\n ]\n\n for (let turn = 0; turn < maxTurns; turn++) {\n if (opts.signal?.aborted) throw new Error('multishot aborted')\n\n let dispatchesThisTurn = 0\n while (true) {\n const {\n message: agentMsg,\n usage: agentUsage,\n costUsd: agentCostUsd,\n } = await agentTransport({\n model: agentModel,\n messages: agentMessages,\n tools,\n temperature: 0.7,\n maxTokens: dispatchesThisTurn === 0 ? agentMaxTokens : toolFollowupMaxTokens,\n signal: opts.signal,\n })\n meter.costUsd += agentCostUsd ?? estimateRouterCost(agentModel, agentUsage)\n if (agentCostUsd === undefined && agentUsage === undefined) meter.uncaptured = true\n\n const agentText = (agentMsg.content ?? '').trim()\n const agentToolCalls = (agentMsg.tool_calls ?? []).map((tc) => ({\n id: tc.id,\n name: tc.function.name,\n args: (() => {\n try {\n return JSON.parse(tc.function.arguments) as Record<string, unknown>\n } catch {\n return {} as Record<string, unknown>\n }\n })(),\n }))\n\n agentMessages.push({\n role: 'assistant',\n content: agentText || null,\n ...(agentMsg.tool_calls?.length ? { tool_calls: agentMsg.tool_calls } : {}),\n })\n transcript.push({\n role: 'assistant',\n content: agentText,\n toolCalls: agentToolCalls.length > 0 ? agentToolCalls : undefined,\n })\n\n if (agentToolCalls.length === 0) break\n dispatchesThisTurn += agentToolCalls.length\n if (dispatchesThisTurn > maxToolDispatches) {\n throw new Error(\n `multishot: tool dispatch cap exceeded (${dispatchesThisTurn}/${maxToolDispatches}) on turn ${turn}`,\n )\n }\n\n for (const tc of agentToolCalls) {\n toolCalls++\n let toolResult = ''\n try {\n const executor = executors[tc.name]\n if (!executor) {\n toolResult = JSON.stringify({ error: `unknown tool ${tc.name}` })\n } else {\n const r = await executor(tc.args, { apiKey, baseUrl, signal: opts.signal })\n toolResult = r.content\n meter.costUsd += r.costUsd\n const artifactType = artifactTypeFor(tc.name)\n if (artifactType) {\n artifacts.push({\n type: artifactType,\n turn,\n invocation: { name: tc.name, args: tc.args },\n content: toolResult,\n })\n }\n }\n } catch (err) {\n if (err instanceof MultishotFatalToolError) throw err\n toolResult = JSON.stringify({ error: err instanceof Error ? err.message : String(err) })\n }\n agentMessages.push({ role: 'tool', tool_call_id: tc.id, content: toolResult || 'done' })\n transcript.push({ role: 'tool', content: toolResult || 'done', toolCallId: tc.id })\n }\n }\n\n if (turn < maxTurns - 1) {\n const driver = await driverTurn({\n transport: driverTransport,\n persona: opts.persona,\n shape,\n transcript,\n turn,\n models: driverModels,\n maxTokens: driverMaxTokens,\n signal: opts.signal,\n meter,\n })\n agentMessages.push({ role: 'user', content: driver.content })\n transcript.push({ role: 'user', content: driver.content })\n }\n }\n\n return {\n transcript,\n artifacts,\n toolCalls,\n durationMs: Date.now() - meter.startedAt,\n costUsd: meter.costUsd,\n // A shot that priced every call reports a complete estimate. One that had\n // a call the router priced at nothing reports a subtotal, so the cell is\n // recorded as under-counted rather than as a complete estimate.\n costProvenance: meter.uncaptured\n ? { kind: 'uncaptured', usd: null }\n : { kind: 'estimated', usd: meter.costUsd },\n }\n}\n\nasync function driverTurn<TPersona extends MultishotPersona>(opts: {\n transport: MultishotTransport\n persona: TPersona\n shape: Required<MultishotShape<TPersona>>\n transcript: MultishotMessage[]\n turn: number\n models: string[]\n maxTokens: number\n signal?: AbortSignal\n /** Charged for EVERY attempt, including the empty ones that force a retry\n * and the ones that precede `MultishotDriverEmptyError`. Those calls billed\n * the provider whether or not their content was usable. */\n meter: ShotMeter\n}): Promise<{ content: string }> {\n const driverSystem = opts.shape.buildDriverSystemPrompt(opts.persona)\n\n // Translate transcript to driver POV: agent's `assistant` messages become\n // `user` (the agent talking TO the driver); the driver's prior `user`\n // messages become `assistant` (the driver's prior responses).\n const driverMessages: Array<Record<string, unknown>> = [{ role: 'system', content: driverSystem }]\n for (const msg of opts.transcript) {\n if (msg.role === 'tool') continue\n const content = driverVisibleContent(msg)\n if (!content) continue\n if (msg.role === 'assistant') driverMessages.push({ role: 'user', content })\n else if (msg.role === 'user') driverMessages.push({ role: 'assistant', content })\n }\n\n // Driver must never go silent. Retry once on empty content; then fail loud.\n for (const model of opts.models) {\n for (let attempt = 0; attempt < 2; attempt++) {\n const { message, usage, costUsd } = await opts.transport({\n model,\n messages: driverMessages,\n temperature: 0.9,\n maxTokens: opts.maxTokens,\n signal: opts.signal,\n })\n opts.meter.costUsd += costUsd ?? estimateRouterCost(model, usage)\n if (costUsd === undefined && usage === undefined) opts.meter.uncaptured = true\n const content = (message.content ?? '').trim()\n if (content.length > 0) return { content }\n }\n }\n throw new MultishotDriverEmptyError(opts.turn)\n}\n\nfunction driverVisibleContent(msg: MultishotMessage): string | null {\n const text = msg.content.trim()\n if (text.length > 0) return text\n if (msg.role !== 'assistant' || !msg.toolCalls?.length) return null\n\n const toolNames = msg.toolCalls.map((call) => call.name.trim()).filter(Boolean)\n if (toolNames.length === 0) return 'Agent called tools.'\n return `Agent called ${toolNames.length === 1 ? 'tool' : 'tools'}: ${toolNames.join(', ')}.`\n}\n","// Multishot matrix wrapper — sweeps profiles × personas × reps, runs\n// the driver-agent loop per cell, applies up to three configured judges,\n// persists per-cell artifacts, and aggregates by axis.\n//\n// Uses runAgentMatrix from @tangle-network/agent-eval/matrix under the\n// hood so cell scheduling + concurrency + cost ceiling are unified with\n// other matrix consumers.\n\nimport { mkdirSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport type { CostProvenance } from '../cost-ledger'\nimport type { MatrixResult } from '../matrix'\nimport { runAgentMatrix, withCellSpend } from '../matrix'\nimport { type JudgeConfig, type JudgeScore, runJudge } from './judges'\nimport { type MultishotShot, runMultishot } from './multishot'\nimport {\n assertMultishotShotResult,\n type MultishotArtifact,\n type MultishotMessage,\n type MultishotPersona,\n type MultishotShape,\n type MultishotToolDefinition,\n type MultishotToolExecutor,\n type MultishotTransport,\n} from './types'\n\nexport interface ConversationJudgeInput<TPersona extends MultishotPersona> {\n transcript: MultishotMessage[]\n persona: TPersona\n}\n\nexport interface ArtifactJudgeInput<TPersona extends MultishotPersona> {\n artifact: MultishotArtifact\n persona: TPersona\n}\n\nexport interface MultishotJudges<TPersona extends MultishotPersona> {\n /** Scores the full transcript end-to-end (always runs). */\n conversation: JudgeConfig<ConversationJudgeInput<TPersona>>\n /** Scores each code-type artifact. Optional — omit when domain has no code artifacts. */\n codeReview?: JudgeConfig<ArtifactJudgeInput<TPersona>>\n /** Scores each non-code (research/content/template) artifact. Optional. */\n contentQuality?: JudgeConfig<ArtifactJudgeInput<TPersona>>\n /** Which artifact types route to codeReview. Defaults to ['code']. */\n codeArtifactTypes?: string[]\n /** Which artifact types route to contentQuality. Defaults to ['research']. */\n contentArtifactTypes?: string[]\n}\n\nexport interface CellCompositeScore {\n composite: number\n conversation: JudgeScore\n codeReview?: {\n perArtifact: Array<JudgeScore & { turn: number; type: string }>\n composite: number\n }\n contentQuality?: {\n perArtifact: Array<JudgeScore & { turn: number; type: string }>\n composite: number\n }\n}\n\nexport interface RunMultishotMatrixOptions<TPersona extends MultishotPersona> {\n /** AgentProfile axis (matrix primary). */\n profiles: Array<{ id: string; value: AgentProfile }>\n /** Persona axis. */\n personas: TPersona[]\n /** Persona-shaping callbacks. Optional — omitted callbacks are derived per\n * cell from that cell's profile + persona payload (pure-profile path). */\n shape?: MultishotShape<TPersona>\n /** Judge configurations. */\n judges: MultishotJudges<TPersona>\n /** Tool definitions advertised to the agent. Defaults to delegate_research + delegate_code. */\n tools?: MultishotToolDefinition[]\n /** Map from tool name → inline executor. Must align with `tools`. */\n toolExecutors?: Record<string, MultishotToolExecutor>\n /** Tool name → artifact type label. Defaults to research/code mapping. */\n artifactTypeFor?: (toolName: string) => string | undefined\n /** Where per-cell artifacts land. Cells write to `<runDir>/<profileId>/<personaId>/rep-N/`. */\n runDir: string\n /** Replicates per (profile, persona) cell. */\n reps?: number\n /** Max conversation turns per cell. */\n maxTurns?: number\n /** Maximum tool calls the agent may dispatch inside one assistant turn. */\n maxToolDispatches?: number\n /** Max concurrent cells. */\n maxConcurrency?: number\n /** Total $ ceiling across the matrix; cells aborted past this. */\n costCeiling?: number\n /** Upper bound on what one cell can spend. A cell whose cost is a subtotal\n * is charged this bound against `costCeiling` instead of its known amount,\n * so hidden spend cannot walk the run past its budget. */\n maxCellCostUsd?: number\n /** Agent model. */\n agentModel?: string\n /** Driver model. */\n driverModel?: string\n /** Fallback driver models tried when the primary simulated-user model returns empty twice. */\n driverFallbackModels?: string[]\n /** Maximum output tokens for the first agent call in each assistant turn. */\n agentMaxTokens?: number\n /** Maximum output tokens for agent follow-up calls after tool results. */\n toolFollowupMaxTokens?: number\n /** Maximum output tokens for each simulated-user driver response. */\n driverMaxTokens?: number\n /** Maximum output tokens for each judge response. */\n judgeMaxTokens?: number\n /** Execution seam for the agent leg of every cell — replaces the router\n * HTTP call when provided (see RunMultishotOptions.agentTransport).\n * Judges are unaffected; configure those via MultishotJudges. */\n agentTransport?: MultishotTransport\n /** Execution seam for the simulated-user driver leg of every cell. */\n driverTransport?: MultishotTransport\n /** Conversation engine for every cell. Defaults to `runMultishot`.\n *\n * The matrix owns everything around the shot — cell fan-out, concurrency,\n * the cost ceiling, the judge slots, the cell composite, the per-cell\n * artifact writers and the run summary — and forwards the whole cell input\n * to this function, so an alternative engine replaces ONLY the\n * conversation. Every option on this interface that `runMultishot` accepts\n * reaches the shot unchanged. `RunMultishotOptions.signal` has no\n * matrix-level counterpart and is not forwarded; a shot owns its own\n * cancellation.\n *\n * A shot that resolves with a value outside `MultishotResult` throws\n * `MultishotShotResultError` for that cell. The default engine is never\n * used as a fallback. */\n runShot?: MultishotShot<TPersona>\n /** Pass-thru fields. */\n apiKey?: string\n baseUrl?: string\n}\n\n/** Per-cell output the multishot matrix records in `MatrixResult.cells`.\n * A consumer that supplies its own `runShot` reads the matrix through this\n * type instead of declaring a structural copy. */\nexport interface MultishotCellOutput {\n turns: number\n toolCalls: number\n artifactCount: number\n}\n\ninterface ArtifactJudgeRun {\n score: JudgeScore & { turn: number; type: string }\n cost: CostProvenance\n}\n\n/** Mean composite over non-failed scores. `0` when the list is empty (a\n * configured judge with nothing to score contributes 0, matching the cell\n * composite's long-standing semantics); `null` when scores exist but every\n * one failed — no signal, so the slot must be EXCLUDED from the cell mean\n * rather than dragging it to zero. */\nfunction meanCompositeExcludingFailed(scores: ReadonlyArray<JudgeScore>): number | null {\n if (scores.length === 0) return 0\n const live = scores.filter((s) => !s.failed)\n if (live.length === 0) return null\n return live.reduce((sum, s) => sum + s.composite, 0) / live.length\n}\n\nexport interface CellCompositeInput {\n conversation: JudgeScore\n /** Present iff the codeReview judge is configured. */\n codeReviews?: ReadonlyArray<JudgeScore>\n /** Present iff the contentQuality judge is configured. */\n contentReviews?: ReadonlyArray<JudgeScore>\n}\n\n/** Cell composite = mean over configured judge slots, excluding failed\n * scores: a failed conversation judge or an all-failed artifact slot carries\n * no signal and is dropped from the mean. `composite` is 0 only when EVERY\n * configured slot failed (`allJudgesFailed` distinguishes that from a real\n * zero). Pure — exported for deterministic testing. */\nexport function computeCellComposite(input: CellCompositeInput): {\n composite: number\n codeComposite: number\n contentComposite: number\n allJudgesFailed: boolean\n} {\n const contributions: number[] = []\n if (!input.conversation.failed) contributions.push(input.conversation.composite)\n\n const codeMean = input.codeReviews ? meanCompositeExcludingFailed(input.codeReviews) : undefined\n if (typeof codeMean === 'number') contributions.push(codeMean)\n const contentMean = input.contentReviews\n ? meanCompositeExcludingFailed(input.contentReviews)\n : undefined\n if (typeof contentMean === 'number') contributions.push(contentMean)\n\n return {\n composite:\n contributions.length === 0\n ? 0\n : contributions.reduce((s, v) => s + v, 0) / contributions.length,\n codeComposite: codeMean ?? 0,\n contentComposite: contentMean ?? 0,\n allJudgesFailed: contributions.length === 0,\n }\n}\n\nexport interface RunMultishotMatrixResult {\n matrix: MatrixResult<MultishotCellOutput>\n}\n\nexport async function runMultishotMatrix<TPersona extends MultishotPersona>(\n opts: RunMultishotMatrixOptions<TPersona>,\n): Promise<RunMultishotMatrixResult> {\n const codeTypes = new Set(opts.judges.codeArtifactTypes ?? ['code'])\n const contentTypes = new Set(opts.judges.contentArtifactTypes ?? ['research'])\n const runShot: MultishotShot<TPersona> = opts.runShot ?? runMultishot\n mkdirSync(opts.runDir, { recursive: true })\n\n const matrix = await runAgentMatrix<MultishotCellOutput>({\n axes: [\n { name: 'profile', values: opts.profiles },\n { name: 'persona', values: opts.personas.map((p) => ({ id: p.id, value: p })) },\n ],\n reps: opts.reps ?? 1,\n maxConcurrency: opts.maxConcurrency ?? 2,\n costCeiling: opts.costCeiling,\n maxCellCostUsd: opts.maxCellCostUsd,\n async runCell(cell) {\n const cellStartedAt = Date.now()\n const profile = cell.axes.profile?.value as AgentProfile\n const persona = cell.axes.persona?.value as TPersona\n const profileId = String(cell.axes.profile?.id ?? 'unknown')\n const personaId = String(cell.axes.persona?.id ?? 'unknown')\n\n // A shot that throws declares its own spend (see `runMultishot`). Rethrow\n // it untouched: a shot that declared nothing spent an unknown amount, and\n // claiming a number here would report a fabricated total.\n const sim = await runShot({\n profile,\n persona,\n shape: opts.shape,\n tools: opts.tools,\n toolExecutors: opts.toolExecutors,\n artifactTypeFor: opts.artifactTypeFor,\n maxTurns: opts.maxTurns,\n maxToolDispatches: opts.maxToolDispatches,\n agentModel: opts.agentModel,\n driverModel: opts.driverModel,\n driverFallbackModels: opts.driverFallbackModels,\n agentMaxTokens: opts.agentMaxTokens,\n toolFollowupMaxTokens: opts.toolFollowupMaxTokens,\n driverMaxTokens: opts.driverMaxTokens,\n agentTransport: opts.agentTransport,\n driverTransport: opts.driverTransport,\n apiKey: opts.apiKey,\n baseUrl: opts.baseUrl,\n })\n // Everything from here on runs with the shot's spend already committed.\n // A throw past this line must carry it, or the money leaves the matrix's\n // cumulative sum and the cost ceiling under-counts the run.\n const shotCostUsd = shotCostSubtotal(sim)\n // A shot that says nothing about provenance is taken at its word, which\n // is how every engine written before the field behaved.\n const shotCostComplete = sim?.costProvenance?.kind !== 'uncaptured'\n let judgeCostUsd = 0\n let judgeCostComplete = false\n // Where the cell is, so a throw declares the right completeness: before\n // the judges only the shot has spent, while they are in flight their\n // spend is unknown, and after they settle their receipts decide.\n let phase: 'validate' | 'judging' | 'scoring' = 'validate'\n try {\n assertMultishotShotResult(sim)\n\n const codeArtifacts = sim.artifacts.filter((a) => codeTypes.has(a.type))\n const contentArtifacts = sim.artifacts.filter((a) => contentTypes.has(a.type))\n\n phase = 'judging'\n const [conversationRun, codeReviewRuns, contentReviewRuns] = await Promise.all([\n runJudge(withJudgeMaxTokens(opts.judges.conversation, opts.judgeMaxTokens), {\n transcript: sim.transcript,\n persona,\n }),\n opts.judges.codeReview\n ? Promise.all(\n codeArtifacts.map((artifact) =>\n runJudge(withJudgeMaxTokens(opts.judges.codeReview!, opts.judgeMaxTokens), {\n artifact,\n persona,\n }).then((result) => ({\n score: {\n ...result.score,\n turn: artifact.turn,\n type: artifact.type,\n },\n cost: result.cost,\n })),\n ),\n )\n : Promise.resolve([] as ArtifactJudgeRun[]),\n opts.judges.contentQuality\n ? Promise.all(\n contentArtifacts.map((artifact) =>\n runJudge(withJudgeMaxTokens(opts.judges.contentQuality!, opts.judgeMaxTokens), {\n artifact,\n persona,\n }).then((result) => ({\n score: {\n ...result.score,\n turn: artifact.turn,\n type: artifact.type,\n },\n cost: result.cost,\n })),\n ),\n )\n : Promise.resolve([] as ArtifactJudgeRun[]),\n ])\n const judgeRuns = [conversationRun, ...codeReviewRuns, ...contentReviewRuns]\n judgeCostUsd = judgeRuns.reduce((sum, run) => sum + (run.cost.usd ?? 0), 0)\n judgeCostComplete = judgeRuns.every((run) => run.cost.kind !== 'uncaptured')\n phase = 'scoring'\n\n const conversation = conversationRun.score\n const codeReviews = codeReviewRuns.map((run) => run.score)\n const contentReviews = contentReviewRuns.map((run) => run.score)\n\n const { composite, codeComposite, contentComposite, allJudgesFailed } =\n computeCellComposite({\n conversation,\n codeReviews: opts.judges.codeReview ? codeReviews : undefined,\n contentReviews: opts.judges.contentQuality ? contentReviews : undefined,\n })\n\n const cellScore: CellCompositeScore = { composite, conversation }\n if (opts.judges.codeReview)\n cellScore.codeReview = { perArtifact: codeReviews, composite: codeComposite }\n if (opts.judges.contentQuality)\n cellScore.contentQuality = { perArtifact: contentReviews, composite: contentComposite }\n\n const cellDir = join(opts.runDir, profileId, personaId, `rep-${cell.rep}`)\n mkdirSync(cellDir, { recursive: true })\n writeFileSync(join(cellDir, 'transcript.json'), JSON.stringify(sim.transcript, null, 2))\n writeFileSync(join(cellDir, 'artifacts.json'), JSON.stringify(sim.artifacts, null, 2))\n writeFileSync(join(cellDir, 'scores.json'), JSON.stringify(cellScore, null, 2))\n\n const notes = [`convo=${conversation.composite.toFixed(1)}`]\n if (opts.judges.codeReview) notes.push(`code=${codeComposite.toFixed(1)}`)\n if (opts.judges.contentQuality) notes.push(`content=${contentComposite.toFixed(1)}`)\n if (allJudgesFailed) notes.push('all-judges-failed')\n if (!judgeCostComplete) notes.push('judge-cost-incomplete')\n\n return {\n output: {\n turns: sim.transcript.length,\n toolCalls: sim.toolCalls,\n artifactCount: sim.artifacts.length,\n },\n verdict: { valid: composite >= 5, score: composite, notes: notes.join(' ') },\n costUsd: sim.costUsd + judgeCostUsd,\n // Either leg can leave the cell total a subtotal: a judge whose cost\n // the router never reported, or a shot that declared its own spend\n // uncaptured. The matrix counts the known part toward the ceiling and\n // reports the cell as under-counted rather than presenting the sum as\n // complete.\n costProvenance:\n judgeCostComplete && shotCostComplete\n ? { kind: 'estimated', usd: sim.costUsd + judgeCostUsd }\n : { kind: 'uncaptured', usd: null },\n durationMs: sim.durationMs,\n }\n } catch (err) {\n // No usable shot subtotal means the cell's spend is genuinely unknown;\n // rethrow untouched so the matrix records it as uncaptured instead of\n // billing a number this frame cannot support.\n if (shotCostUsd === undefined) throw err\n throw withCellSpend(err, {\n costUsd: shotCostUsd + judgeCostUsd,\n durationMs: Date.now() - cellStartedAt,\n // The judges bill before they settle, so a throw among them leaves a\n // subtotal. Before they start, and after they settle with every\n // receipt captured, the amount is the cell's whole spend — unless the\n // shot already declared its own part incomplete.\n kind:\n shotCostComplete && (phase === 'validate' || (phase === 'scoring' && judgeCostComplete))\n ? 'estimated'\n : 'uncaptured',\n })\n }\n },\n })\n\n // Persist top-level summary.\n const summary = {\n cells: matrix.summary.totalCells,\n passRate: matrix.summary.overallPassRate,\n meanScore: matrix.summary.overallMeanScore,\n totalCostUsd: matrix.summary.totalCostUsd,\n costUncapturedCells: matrix.summary.costUncapturedCells,\n ceilingChargedUsd: matrix.summary.ceilingChargedUsd,\n durationMs: matrix.summary.durationMs,\n runsExecuted: matrix.summary.runsExecuted,\n cellsSkipped: matrix.summary.cellsSkipped,\n byProfile: matrix.byAxis.profile,\n byPersona: matrix.byAxis.persona,\n }\n writeFileSync(join(opts.runDir, 'summary.json'), JSON.stringify(summary, null, 2))\n\n // A reader of the on-disk summary must not take the cost line as the run's\n // whole spend when some cells reported only a subtotal.\n const uncaptured = matrix.summary.costUncapturedCells\n const costLabel = uncaptured > 0 ? 'Cost (at least)' : 'Cost'\n\n const md: string[] = [\n `# Multishot matrix`,\n ``,\n `**Cells**: ${matrix.summary.totalCells} | **Pass rate**: ${(matrix.summary.overallPassRate * 100).toFixed(0)}% | **Mean**: ${matrix.summary.overallMeanScore.toFixed(2)} | **${costLabel}**: $${matrix.summary.totalCostUsd.toFixed(2)} | **Duration**: ${(matrix.summary.durationMs / 1000).toFixed(0)}s`,\n ``,\n ...(uncaptured > 0\n ? [\n `> ${uncaptured} of ${matrix.summary.runsExecuted} cells reported a cost subtotal, not a total. Real spend is higher than every cost figure below.`,\n ``,\n ]\n : []),\n `## By profile`,\n ``,\n '| profile | pass | mean | cost |',\n '|---|---|---|---|',\n ...Object.entries(matrix.byAxis.profile ?? {}).map(\n ([id, s]) =>\n `| ${id} | ${(s.passRate * 100).toFixed(0)}% | ${s.meanScore.toFixed(2)} | $${s.totalCostUsd.toFixed(2)} |`,\n ),\n ``,\n `## By persona`,\n ``,\n '| persona | pass | mean | cost |',\n '|---|---|---|---|',\n ...Object.entries(matrix.byAxis.persona ?? {}).map(\n ([id, s]) =>\n `| ${id} | ${(s.passRate * 100).toFixed(0)}% | ${s.meanScore.toFixed(2)} | $${s.totalCostUsd.toFixed(2)} |`,\n ),\n ``,\n ]\n writeFileSync(join(opts.runDir, 'summary.md'), md.join('\\n'))\n\n return { matrix }\n}\n\n/** The shot's own spend, when the value it resolved with reports a usable\n * amount. `undefined` when it does not — a malformed shot result is exactly\n * the case where the number cannot be trusted, and billing a wrong figure is\n * worse than recording the cell as uncaptured. */\nfunction shotCostSubtotal(\n sim: { costUsd?: unknown; costProvenance?: CostProvenance } | null | undefined,\n): number | undefined {\n const value = sim?.costUsd\n return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : undefined\n}\n\nfunction withJudgeMaxTokens<TInput>(\n judge: JudgeConfig<TInput>,\n maxTokens: number | undefined,\n): JudgeConfig<TInput> {\n if (maxTokens === undefined || judge.maxTokens !== undefined) return judge\n return { ...judge, maxTokens }\n}\n"],"mappings":";;;;;;AAiCA,eAAsB,iBACpB,KACmC;CACnC,MAAM,YAAY,KAAK,IAAI;CAC3B,MAAM,OAAgC;EACpC,OAAO,IAAI;EACX,UAAU,IAAI;EACd,aAAa,IAAI,eAAe;EAChC,YAAY,IAAI,aAAa;CAC/B;CACA,IAAI,IAAI,OAAO,QAAQ,KAAK,QAAQ,IAAI;CACxC,MAAM,MAAM,GAAG,IAAI,QAAQ,QAAQ,QAAQ,EAAE,EAAE;CAC/C,MAAM,MAAM,MAAM,MAAM,KAAK;EAC3B,QAAQ;EACR,SAAS;GAAE,eAAe,UAAU,IAAI;GAAU,gBAAgB;EAAmB;EACrF,MAAM,KAAK,UAAU,IAAI;EACzB,QAAQ,IAAI;CACd,CAAC;CACD,IAAI,CAAC,IAAI,IAAI;EACX,MAAM,OAAO,MAAM,IAAI,KAAK;EAC5B,MAAM,IAAI,MAAM,UAAU,IAAI,OAAO,IAAI,KAAK,MAAM,GAAG,GAAG,GAAG;CAC/D;CACA,MAAM,OAAQ,MAAM,IAAI,KAAK;CAO7B,MAAM,SAAS,KAAK,QAAQ;CAC5B,IAAI,CAAC,QAAQ,MAAM,IAAI,MAAM,+BAA+B,KAAK,UAAU,IAAI,CAAC,CAAC,MAAM,GAAG,GAAG,GAAG;CAChG,MAAM,UAAU,KAAK,kBAAkB,KAAK;CAC5C,MAAM,UACJ,OAAO,YAAY,YAAY,OAAO,SAAS,OAAO,KAAK,WAAW,IAAI,UAAU,KAAA;CACtF,OAAO;EACL,SAAS,OAAO;EAChB,OAAO,KAAK;EACZ,GAAI,YAAY,KAAA,IAAY,CAAC,IAAI,EAAE,QAAQ;EAC3C,OAAO,OAAO,KAAK,UAAU,YAAY,KAAK,QAAQ,KAAK,QAAQ,IAAI;EACvE,YAAY,KAAK,IAAI,IAAI;CAC3B;AACF;AAIA,SAAgB,mBACd,OACA,OACQ;CACR,IAAI,CAAC,OAAO,OAAO;CACnB,MAAM,WAAW,MAAM,iBAAiB;CACxC,MAAM,YAAY,MAAM,qBAAqB;CAC7C,IAAI,UAAU;CACd,IAAI,WAAW;CACf,IAAI,MAAM,SAAS,aAAa,GAAG;EACjC,UAAU;EACV,WAAW;CACb,OAAO,IAAI,MAAM,SAAS,SAAS,KAAK,MAAM,SAAS,eAAe,GAAG;EACvE,UAAU;EACV,WAAW;CACb,OAAO,IAAI,MAAM,SAAS,MAAM,KAAK,MAAM,SAAS,KAAK,KAAK,MAAM,SAAS,UAAU,GAAG;EACxF,UAAU;EACV,WAAW;CACb;CACA,QAAQ,WAAW,UAAU,YAAY,YAAY;AACvD;AAEA,SAAgB,uBAA+B;CAC7C,QAAQ,QAAQ,IAAI,0BAA0B,iCAAA,CAAkC,QAC9E,QACA,EACF;AACF;AAEA,SAAgB,sBAA8B;CAC5C,MAAM,MAAM,QAAQ,IAAI;CACxB,IAAI,CAAC,KAAK,MAAM,IAAI,MAAM,gEAAgE;CAC1F,OAAO;AACT;;;ACrGA,MAAa,2BAA2B;AACxC,MAAa,sBAAsB;AAgBnC,MAAM,4BACJ;AAEF,MAAM,uBACJ;AAEF,MAAa,iCAA0D;CACrE,MAAM;CACN,UAAU;EACR,MAAM;EACN,aACE;EACF,YAAY;GACV,MAAM;GACN,YAAY;IACV,UAAU;KAAE,MAAM;KAAU,aAAa;IAAgC;IACzE,OAAO;KACL,MAAM;KACN,aAAa;IACf;GACF;GACA,UAAU,CAAC,UAAU;EACvB;CACF;AACF;AAEA,MAAa,6BAAsD;CACjE,MAAM;CACN,UAAU;EACR,MAAM;EACN,aACE;EACF,YAAY;GACV,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;IAAgC;IACrE,UAAU;KACR,MAAM;KACN,aAAa;IACf;GACF;GACA,UAAU,CAAC,MAAM;EACnB;CACF;AACF;AAEA,SAAgB,uBACd,SAAkC,CAAC,GACZ;CACvB,MAAM,eAAe,OAAO,gBAAgB;CAC5C,MAAM,QAAQ,OAAO,SAAA;CACrB,OAAO,OAAO,MAAM,QAAQ;EAC1B,MAAM,WAAW,OAAO,KAAK,YAAY,EAAE;EAC3C,MAAM,QAAQ,KAAK,QAAQ,OAAO,KAAK,KAAK,IAAI,KAAA;EAChD,MAAM,EAAE,SAAS,UAAU,MAAM,iBAAiB;GAChD,QAAQ,IAAI;GACZ,SAAS,IAAI;GACb;GACA,aAAa;GACb,WAAW;GACX,UAAU,CACR;IAAE,MAAM;IAAU,SAAS;GAAa,GACxC;IAAE,MAAM;IAAQ,SAAS,aAAa,WAAW,QAAQ,YAAY,UAAU;GAAK,CACtF;GACA,QAAQ,IAAI;EACd,CAAC;EACD,OAAO;GAAE,SAAS,QAAQ,WAAW;GAAI,SAAS,mBAAmB,OAAO,KAAK;EAAE;CACrF;AACF;AAEA,SAAgB,mBAAmB,SAA6B,CAAC,GAA0B;CACzF,MAAM,eAAe,OAAO,gBAAgB;CAC5C,MAAM,QAAQ,OAAO,SAAA;CACrB,OAAO,OAAO,MAAM,QAAQ;EAC1B,MAAM,OAAO,OAAO,KAAK,QAAQ,EAAE;EACnC,MAAM,WAAW,KAAK,WAAW,OAAO,KAAK,QAAQ,IAAI;EACzD,MAAM,EAAE,SAAS,UAAU,MAAM,iBAAiB;GAChD,QAAQ,IAAI;GACZ,SAAS,IAAI;GACb;GACA,aAAa;GACb,WAAW;GACX,UAAU,CACR;IAAE,MAAM;IAAU,SAAS,GAAG,aAAa,gBAAgB;GAAW,GACtE;IAAE,MAAM;IAAQ,SAAS,YAAY;GAAO,CAC9C;GACA,QAAQ,IAAI;EACd,CAAC;EACD,OAAO;GAAE,SAAS,QAAQ,WAAW;GAAI,SAAS,mBAAmB,OAAO,KAAK;EAAE;CACrF;AACF;AAgBA,SAAgB,uBAAuB,SAA6B,CAAC,GAAuB;CAC1F,OAAO;EACL,OAAO,CAAC,gCAAgC,0BAA0B;EAClE,WAAW;GACT,mBAAmB,uBAAuB,OAAO,QAAQ;GACzD,eAAe,mBAAmB,OAAO,IAAI;EAC/C;EACA,kBAAkB,SAChB,SAAS,sBAAsB,aAAa,SAAS,kBAAkB,SAAS,KAAA;CACpF;AACF;;;ACtHA,MAAa,sBAAsB;AAmCnC,eAAsB,SACpB,OACA,OACyB;CACzB,MAAM,SAAS,MAAM,UAAU,oBAAoB;CACnD,MAAM,UAAU,MAAM,WAAW,qBAAqB;CACtD,MAAM,QAAQ,MAAM,SAAS,QAAQ,IAAI,eAAA;CACzC,MAAM,SAAS,MAAM,YAAY,KAAK;CACtC,IAAI,MAAM;CACV,IAAI;CACJ,IAAI;CACJ,MAAM,YAAY,KAAK,IAAI;CAC3B,IAAI;EACF,MAAM,WAAW,MAAM,iBAAiB;GACtC;GACA;GACA;GACA,aAAa;GACb,WAAW,MAAM,aAAa;GAC9B,UAAU,CACR;IAAE,MAAM;IAAU,SAAS,MAAM;GAAa,GAC9C;IAAE,MAAM;IAAQ,SAAS;GAAO,CAClC;EACF,CAAC;EACD,MAAM,OAAO,kBAAkB,QAAQ;EACvC,UAAU,KAAK;EACf,OAAO,KAAK;EACZ,OAAO,SAAS,QAAQ,WAAW,GAAA,CAAI,KAAK;CAC9C,SAAS,KAAK;EAGZ,OAAO;GACL,OAAO;IACL,YAAY,CAAC;IACb,WAAW;IACX,QAAQ;IACR,OAAO,SAAS,MAAM,KAAK,gBAAgB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC1F,SAAS,iBAAiB,OAAO,KAAK,IAAI,IAAI,SAAS;GACzD;GACA,MAAM;IAAE,MAAM;IAAc,KAAK;GAAK;EACxC;CACF;CAEA,IAAI,SAAyC;CAC7C,IAAI;EACF,MAAM,UAAU,IACb,QAAQ,gBAAgB,EAAE,CAAC,CAC3B,QAAQ,WAAW,EAAE,CAAC,CACtB,KAAK;EACR,SAAS,KAAK,MAAM,OAAO;CAC7B,QAAQ;EACN,OAAO;GACL,OAAO;IACL,YAAY,CAAC;IACb,WAAW;IACX,QAAQ;IACR,OAAO,SAAS,MAAM,KAAK,sBAAsB,IAAI,MAAM,GAAG,GAAG;IACjE;GACF;GACA;EACF;CACF;CAEA,MAAM,aAAqC,CAAC;CAC5C,IAAI,MAAM;CACV,KAAK,MAAM,OAAO,MAAM,YAAY;EAClC,MAAM,IAAI,OAAO,OAAO,IAAI,QAAQ,CAAC;EACrC,MAAM,UAAU,OAAO,SAAS,CAAC,IAAI,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,CAAC,CAAC,IAAI;EACpE,WAAW,IAAI,OAAO;EACtB,OAAO;CACT;CACA,OAAO;EACL,OAAO;GACL;GACA,WAAW,MAAM,WAAW,WAAW,IAAI,IAAI,MAAM,MAAM,WAAW;GACtE,OAAO,OAAO,OAAO,UAAU,WAAW,OAAO,QAAQ;GACzD;EACF;EACA;CACF;AACF;AAEA,SAAS,kBAAkB,UAGzB;CACA,MAAM,QAAQ,eAAe,SAAS,KAAK;CAC3C,OAAO;EACL,SAAS;GACP;GACA,SAAS,SAAS,WAAW;GAC7B,OAAO,SAAS;GAChB,YAAY,SAAS;EACvB;EACA,MACE,SAAS,YAAY,KAAA,IACjB;GAAE,MAAM;GAAY,KAAK,SAAS;EAAQ,IAC1C,MAAM,aAAa,SAAS,CAAC,cAAc,SAAS,KAAK,IACvD;GAAE,MAAM;GAAc,KAAK;EAAK,IAChC;GACE,MAAM;GACN,KAAK,aAAa,MAAM,cAAc,MAAM,kBAAkB,SAAS,KAAK;EAC9E;CACV;AACF;AAEA,SAAS,eAAe,OAAoD;CAC1E,MAAM,eAAe,WAAW,OAAO,aAAa;CACpD,MAAM,mBAAmB,WAAW,OAAO,iBAAiB;CAC5D,MAAM,WAAW,iBAAiB,KAAA,KAAa,qBAAqB,KAAA;CACpE,OAAO;EACL,cAAc,gBAAgB;EAC9B,kBAAkB,oBAAoB;EACtC,cAAc,gBAAgB,MAAM,oBAAoB;EACxD;CACF;AACF;AAEA,SAAS,iBAAiB,OAAe,YAAqC;CAC5E,OAAO;EACL,OAAO;GAAE,cAAc;GAAG,kBAAkB;GAAG,aAAa;GAAG,UAAU;EAAM;EAC/E,SAAS;EACT;EACA;CACF;AACF;AAEA,SAAS,WAAW,OAAoC;CACtD,OAAO,OAAO,UAAU,YAAY,OAAO,cAAc,KAAK,KAAK,SAAS,IAAI,QAAQ,KAAA;AAC1F;;;AAIA,SAAgB,iBAAiB,MAAyC;CACxE,OAAO,KAAK,KAAK,MAAM,KAAK,EAAE,IAAI,IAAI,EAAE,aAAa,CAAC,CAAC,KAAK,IAAI;AAClE;;AAGA,SAAgB,iBAAiB,MAAyC;CAExE,OAAO,6DADQ,KAAK,KAAK,MAAM,IAAI,EAAE,IAAI,IAAI,CAAC,CAAC,KAAK,GACqB,EAAE;AAC7E;;;;;AC3LA,SAAgB,mBAAmB,SAAmC;CACpE,MAAM,QAAkB,CAAC;CACzB,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,OAAO,GAAG;EAClD,IAAI,QAAQ,QAAQ,UAAU,KAAA,GAAW;EACzC,MAAM,WACJ,OAAO,UAAU,YAAY,OAAO,UAAU,YAAY,OAAO,UAAU,YACvE,OAAO,KAAK,IACZ,KAAK,UAAU,KAAK;EAC1B,MAAM,KAAK,KAAK,IAAI,IAAI,UAAU;CACpC;CACA,OAAO,MAAM,KAAK,IAAI;AACxB;;;AAIA,SAAgB,uBAAuB,SAAuB,SAAmC;CAC/F,MAAM,QAAQ,mBAAmB,OAAO;CAExC,OAAO;EACL,4BAFuB,QAAQ,eAAe,QAAQ,QAAQ,YAEjB;EAC7C,MAAM,SAAS,IAAI,QAAQ,qCAAqC,QAAQ,GAAG;EAC3E;EACA;CACF,CAAC,CAAC,KAAK,IAAI;AACb;;;AAIA,SAAgB,mCACd,SACA,SACQ;CACR,MAAM,QAAQ,mBAAmB,OAAO;CACxC,MAAM,mBAAmB,QAAQ,eAAe,QAAQ,QAAQ;CAChE,OAAO;EACL,gDAAgD,QAAQ,GAAG,gBAAgB,iBAAiB;EAC5F,MAAM,SAAS,IAAI,iBAAiB,UAAU;EAC9C;EACA;EACA;EACA;EACA;EACA;EACA;CACF,CAAC,CACE,QAAQ,UAAU,MAAM,SAAS,CAAC,CAAC,CACnC,KAAK,IAAI;AACd;;;AAIA,SAAgB,wBACd,SACA,OACoC;CACpC,OAAO;EACL,aAAa,OAAO,iBAAiB,YAAY,uBAAuB,SAAS,OAAO;EACxF,yBACE,OAAO,6BACL,YAAY,mCAAmC,SAAS,OAAO;CACrE;AACF;;;;;;;;;;;;;;;;;;;ACoCA,eAAsB,aACpB,MAC0B;CAC1B,MAAM,QAAmB;EAAE,SAAS;EAAG,WAAW,KAAK,IAAI;EAAG,YAAY;CAAM;CAChF,IAAI;EACF,OAAO,MAAM,aAAa,MAAM,KAAK;CACvC,SAAS,KAAK;EAGZ,MAAM,cAAc,KAAK;GACvB,SAAS,MAAM;GACf,YAAY,KAAK,IAAI,IAAI,MAAM;GAC/B,MAAM,MAAM,aAAa,eAAe;EAC1C,CAAC;CACH;AACF;AAEA,eAAe,aACb,MACA,OAC0B;CAC1B,MAAM,SAAS,KAAK,UAAU,oBAAoB;CAClD,MAAM,UAAU,KAAK,WAAW,qBAAqB;CACrD,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,oBAAoB,KAAK,qBAAqB;CACpD,MAAM,aAAa,KAAK,cAAc;CAEtC,MAAM,eAAe,CADD,KAAK,eAAe,sBACL,GAAI,KAAK,wBAAwB,CAAC,CAAE;CACvE,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,wBAAwB,KAAK,yBAAyB;CAC5D,MAAM,kBAAkB,KAAK,mBAAmB;CAEhD,MAAM,SACJ,KAAK,SAAS,KAAK,gBACf;EACE,OAAO,KAAK;EACZ,WAAW,KAAK;EAChB,iBAAiB,KAAK,0BAA0B,KAAA;CAClD,IACA,uBAAuB;CAC7B,MAAM,QAAQ,KAAK,SAAS,OAAO;CACnC,MAAM,YAAY,KAAK,iBAAiB,OAAO;CAC/C,MAAM,kBAAkB,KAAK,mBAAmB,OAAO;CAEvD,MAAM,mBAAuC,QAAQ,iBAAiB;EAAE;EAAQ;EAAS,GAAG;CAAI,CAAC;CACjG,MAAM,iBAAiB,KAAK,kBAAkB;CAC9C,MAAM,kBAAkB,KAAK,mBAAmB;CAEhD,MAAM,QAAQ,wBAAwB,KAAK,SAAS,KAAK,KAAK;CAE9D,MAAM,aAAiC,CAAC;CACxC,MAAM,YAAiC,CAAC;CACxC,IAAI,YAAY;CAEhB,MAAM,SAAS,MAAM,YAAY,KAAK,OAAO;CAC7C,WAAW,KAAK;EAAE,MAAM;EAAQ,SAAS;CAAO,CAAC;CAKjD,MAAM,gBAAgD,CACpD;EAAE,MAAM;EAAU,SAJC,CAAC,KAAK,QAAQ,QAAQ,cAAc,KAAK,QAAQ,QAAQ,kBAAkB,CAAC,CAC9F,QAAQ,UAA2B,QAAQ,KAAK,CAAC,CAAC,CAClD,KAAK,MAEgC;CAAE,GACxC;EAAE,MAAM;EAAQ,SAAS;CAAO,CAClC;CAEA,KAAK,IAAI,OAAO,GAAG,OAAO,UAAU,QAAQ;EAC1C,IAAI,KAAK,QAAQ,SAAS,MAAM,IAAI,MAAM,mBAAmB;EAE7D,IAAI,qBAAqB;EACzB,OAAO,MAAM;GACX,MAAM,EACJ,SAAS,UACT,OAAO,YACP,SAAS,iBACP,MAAM,eAAe;IACvB,OAAO;IACP,UAAU;IACV;IACA,aAAa;IACb,WAAW,uBAAuB,IAAI,iBAAiB;IACvD,QAAQ,KAAK;GACf,CAAC;GACD,MAAM,WAAW,gBAAgB,mBAAmB,YAAY,UAAU;GAC1E,IAAI,iBAAiB,KAAA,KAAa,eAAe,KAAA,GAAW,MAAM,aAAa;GAE/E,MAAM,aAAa,SAAS,WAAW,GAAA,CAAI,KAAK;GAChD,MAAM,kBAAkB,SAAS,cAAc,CAAC,EAAA,CAAG,KAAK,QAAQ;IAC9D,IAAI,GAAG;IACP,MAAM,GAAG,SAAS;IAClB,aAAa;KACX,IAAI;MACF,OAAO,KAAK,MAAM,GAAG,SAAS,SAAS;KACzC,QAAQ;MACN,OAAO,CAAC;KACV;IACF,EAAA,CAAG;GACL,EAAE;GAEF,cAAc,KAAK;IACjB,MAAM;IACN,SAAS,aAAa;IACtB,GAAI,SAAS,YAAY,SAAS,EAAE,YAAY,SAAS,WAAW,IAAI,CAAC;GAC3E,CAAC;GACD,WAAW,KAAK;IACd,MAAM;IACN,SAAS;IACT,WAAW,eAAe,SAAS,IAAI,iBAAiB,KAAA;GAC1D,CAAC;GAED,IAAI,eAAe,WAAW,GAAG;GACjC,sBAAsB,eAAe;GACrC,IAAI,qBAAqB,mBACvB,MAAM,IAAI,MACR,0CAA0C,mBAAmB,GAAG,kBAAkB,YAAY,MAChG;GAGF,KAAK,MAAM,MAAM,gBAAgB;IAC/B;IACA,IAAI,aAAa;IACjB,IAAI;KACF,MAAM,WAAW,UAAU,GAAG;KAC9B,IAAI,CAAC,UACH,aAAa,KAAK,UAAU,EAAE,OAAO,gBAAgB,GAAG,OAAO,CAAC;UAC3D;MACL,MAAM,IAAI,MAAM,SAAS,GAAG,MAAM;OAAE;OAAQ;OAAS,QAAQ,KAAK;MAAO,CAAC;MAC1E,aAAa,EAAE;MACf,MAAM,WAAW,EAAE;MACnB,MAAM,eAAe,gBAAgB,GAAG,IAAI;MAC5C,IAAI,cACF,UAAU,KAAK;OACb,MAAM;OACN;OACA,YAAY;QAAE,MAAM,GAAG;QAAM,MAAM,GAAG;OAAK;OAC3C,SAAS;MACX,CAAC;KAEL;IACF,SAAS,KAAK;KACZ,IAAI,eAAe,yBAAyB,MAAM;KAClD,aAAa,KAAK,UAAU,EAAE,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAE,CAAC;IACzF;IACA,cAAc,KAAK;KAAE,MAAM;KAAQ,cAAc,GAAG;KAAI,SAAS,cAAc;IAAO,CAAC;IACvF,WAAW,KAAK;KAAE,MAAM;KAAQ,SAAS,cAAc;KAAQ,YAAY,GAAG;IAAG,CAAC;GACpF;EACF;EAEA,IAAI,OAAO,WAAW,GAAG;GACvB,MAAM,SAAS,MAAM,WAAW;IAC9B,WAAW;IACX,SAAS,KAAK;IACd;IACA;IACA;IACA,QAAQ;IACR,WAAW;IACX,QAAQ,KAAK;IACb;GACF,CAAC;GACD,cAAc,KAAK;IAAE,MAAM;IAAQ,SAAS,OAAO;GAAQ,CAAC;GAC5D,WAAW,KAAK;IAAE,MAAM;IAAQ,SAAS,OAAO;GAAQ,CAAC;EAC3D;CACF;CAEA,OAAO;EACL;EACA;EACA;EACA,YAAY,KAAK,IAAI,IAAI,MAAM;EAC/B,SAAS,MAAM;EAIf,gBAAgB,MAAM,aAClB;GAAE,MAAM;GAAc,KAAK;EAAK,IAChC;GAAE,MAAM;GAAa,KAAK,MAAM;EAAQ;CAC9C;AACF;AAEA,eAAe,WAA8C,MAa5B;CAM/B,MAAM,iBAAiD,CAAC;EAAE,MAAM;EAAU,SALrD,KAAK,MAAM,wBAAwB,KAAK,OAKiC;CAAE,CAAC;CACjG,KAAK,MAAM,OAAO,KAAK,YAAY;EACjC,IAAI,IAAI,SAAS,QAAQ;EACzB,MAAM,UAAU,qBAAqB,GAAG;EACxC,IAAI,CAAC,SAAS;EACd,IAAI,IAAI,SAAS,aAAa,eAAe,KAAK;GAAE,MAAM;GAAQ;EAAQ,CAAC;OACtE,IAAI,IAAI,SAAS,QAAQ,eAAe,KAAK;GAAE,MAAM;GAAa;EAAQ,CAAC;CAClF;CAGA,KAAK,MAAM,SAAS,KAAK,QACvB,KAAK,IAAI,UAAU,GAAG,UAAU,GAAG,WAAW;EAC5C,MAAM,EAAE,SAAS,OAAO,YAAY,MAAM,KAAK,UAAU;GACvD;GACA,UAAU;GACV,aAAa;GACb,WAAW,KAAK;GAChB,QAAQ,KAAK;EACf,CAAC;EACD,KAAK,MAAM,WAAW,WAAW,mBAAmB,OAAO,KAAK;EAChE,IAAI,YAAY,KAAA,KAAa,UAAU,KAAA,GAAW,KAAK,MAAM,aAAa;EAC1E,MAAM,WAAW,QAAQ,WAAW,GAAA,CAAI,KAAK;EAC7C,IAAI,QAAQ,SAAS,GAAG,OAAO,EAAE,QAAQ;CAC3C;CAEF,MAAM,IAAI,0BAA0B,KAAK,IAAI;AAC/C;AAEA,SAAS,qBAAqB,KAAsC;CAClE,MAAM,OAAO,IAAI,QAAQ,KAAK;CAC9B,IAAI,KAAK,SAAS,GAAG,OAAO;CAC5B,IAAI,IAAI,SAAS,eAAe,CAAC,IAAI,WAAW,QAAQ,OAAO;CAE/D,MAAM,YAAY,IAAI,UAAU,KAAK,SAAS,KAAK,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,OAAO;CAC9E,IAAI,UAAU,WAAW,GAAG,OAAO;CACnC,OAAO,gBAAgB,UAAU,WAAW,IAAI,SAAS,QAAQ,IAAI,UAAU,KAAK,IAAI,EAAE;AAC5F;;;;;;;;AC/LA,SAAS,6BAA6B,QAAkD;CACtF,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,OAAO,OAAO,QAAQ,MAAM,CAAC,EAAE,MAAM;CAC3C,IAAI,KAAK,WAAW,GAAG,OAAO;CAC9B,OAAO,KAAK,QAAQ,KAAK,MAAM,MAAM,EAAE,WAAW,CAAC,IAAI,KAAK;AAC9D;;;;;;AAeA,SAAgB,qBAAqB,OAKnC;CACA,MAAM,gBAA0B,CAAC;CACjC,IAAI,CAAC,MAAM,aAAa,QAAQ,cAAc,KAAK,MAAM,aAAa,SAAS;CAE/E,MAAM,WAAW,MAAM,cAAc,6BAA6B,MAAM,WAAW,IAAI,KAAA;CACvF,IAAI,OAAO,aAAa,UAAU,cAAc,KAAK,QAAQ;CAC7D,MAAM,cAAc,MAAM,iBACtB,6BAA6B,MAAM,cAAc,IACjD,KAAA;CACJ,IAAI,OAAO,gBAAgB,UAAU,cAAc,KAAK,WAAW;CAEnE,OAAO;EACL,WACE,cAAc,WAAW,IACrB,IACA,cAAc,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc;EAC/D,eAAe,YAAY;EAC3B,kBAAkB,eAAe;EACjC,iBAAiB,cAAc,WAAW;CAC5C;AACF;AAMA,eAAsB,mBACpB,MACmC;CACnC,MAAM,YAAY,IAAI,IAAI,KAAK,OAAO,qBAAqB,CAAC,MAAM,CAAC;CACnE,MAAM,eAAe,IAAI,IAAI,KAAK,OAAO,wBAAwB,CAAC,UAAU,CAAC;CAC7E,MAAM,UAAmC,KAAK,WAAW;CACzD,UAAU,KAAK,QAAQ,EAAE,WAAW,KAAK,CAAC;CAE1C,MAAM,SAAS,MAAM,eAAoC;EACvD,MAAM,CACJ;GAAE,MAAM;GAAW,QAAQ,KAAK;EAAS,GACzC;GAAE,MAAM;GAAW,QAAQ,KAAK,SAAS,KAAK,OAAO;IAAE,IAAI,EAAE;IAAI,OAAO;GAAE,EAAE;EAAE,CAChF;EACA,MAAM,KAAK,QAAQ;EACnB,gBAAgB,KAAK,kBAAkB;EACvC,aAAa,KAAK;EAClB,gBAAgB,KAAK;EACrB,MAAM,QAAQ,MAAM;GAClB,MAAM,gBAAgB,KAAK,IAAI;GAC/B,MAAM,UAAU,KAAK,KAAK,SAAS;GACnC,MAAM,UAAU,KAAK,KAAK,SAAS;GACnC,MAAM,YAAY,OAAO,KAAK,KAAK,SAAS,MAAM,SAAS;GAC3D,MAAM,YAAY,OAAO,KAAK,KAAK,SAAS,MAAM,SAAS;GAK3D,MAAM,MAAM,MAAM,QAAQ;IACxB;IACA;IACA,OAAO,KAAK;IACZ,OAAO,KAAK;IACZ,eAAe,KAAK;IACpB,iBAAiB,KAAK;IACtB,UAAU,KAAK;IACf,mBAAmB,KAAK;IACxB,YAAY,KAAK;IACjB,aAAa,KAAK;IAClB,sBAAsB,KAAK;IAC3B,gBAAgB,KAAK;IACrB,uBAAuB,KAAK;IAC5B,iBAAiB,KAAK;IACtB,gBAAgB,KAAK;IACrB,iBAAiB,KAAK;IACtB,QAAQ,KAAK;IACb,SAAS,KAAK;GAChB,CAAC;GAID,MAAM,cAAc,iBAAiB,GAAG;GAGxC,MAAM,mBAAmB,KAAK,gBAAgB,SAAS;GACvD,IAAI,eAAe;GACnB,IAAI,oBAAoB;GAIxB,IAAI,QAA4C;GAChD,IAAI;IACF,0BAA0B,GAAG;IAE7B,MAAM,gBAAgB,IAAI,UAAU,QAAQ,MAAM,UAAU,IAAI,EAAE,IAAI,CAAC;IACvE,MAAM,mBAAmB,IAAI,UAAU,QAAQ,MAAM,aAAa,IAAI,EAAE,IAAI,CAAC;IAE7E,QAAQ;IACR,MAAM,CAAC,iBAAiB,gBAAgB,qBAAqB,MAAM,QAAQ,IAAI;KAC7E,SAAS,mBAAmB,KAAK,OAAO,cAAc,KAAK,cAAc,GAAG;MAC1E,YAAY,IAAI;MAChB;KACF,CAAC;KACD,KAAK,OAAO,aACR,QAAQ,IACN,cAAc,KAAK,aACjB,SAAS,mBAAmB,KAAK,OAAO,YAAa,KAAK,cAAc,GAAG;MACzE;MACA;KACF,CAAC,CAAC,CAAC,MAAM,YAAY;MACnB,OAAO;OACL,GAAG,OAAO;OACV,MAAM,SAAS;OACf,MAAM,SAAS;MACjB;MACA,MAAM,OAAO;KACf,EAAE,CACJ,CACF,IACA,QAAQ,QAAQ,CAAC,CAAuB;KAC5C,KAAK,OAAO,iBACR,QAAQ,IACN,iBAAiB,KAAK,aACpB,SAAS,mBAAmB,KAAK,OAAO,gBAAiB,KAAK,cAAc,GAAG;MAC7E;MACA;KACF,CAAC,CAAC,CAAC,MAAM,YAAY;MACnB,OAAO;OACL,GAAG,OAAO;OACV,MAAM,SAAS;OACf,MAAM,SAAS;MACjB;MACA,MAAM,OAAO;KACf,EAAE,CACJ,CACF,IACA,QAAQ,QAAQ,CAAC,CAAuB;IAC9C,CAAC;IACD,MAAM,YAAY;KAAC;KAAiB,GAAG;KAAgB,GAAG;IAAiB;IAC3E,eAAe,UAAU,QAAQ,KAAK,QAAQ,OAAO,IAAI,KAAK,OAAO,IAAI,CAAC;IAC1E,oBAAoB,UAAU,OAAO,QAAQ,IAAI,KAAK,SAAS,YAAY;IAC3E,QAAQ;IAER,MAAM,eAAe,gBAAgB;IACrC,MAAM,cAAc,eAAe,KAAK,QAAQ,IAAI,KAAK;IACzD,MAAM,iBAAiB,kBAAkB,KAAK,QAAQ,IAAI,KAAK;IAE/D,MAAM,EAAE,WAAW,eAAe,kBAAkB,oBAClD,qBAAqB;KACnB;KACA,aAAa,KAAK,OAAO,aAAa,cAAc,KAAA;KACpD,gBAAgB,KAAK,OAAO,iBAAiB,iBAAiB,KAAA;IAChE,CAAC;IAEH,MAAM,YAAgC;KAAE;KAAW;IAAa;IAChE,IAAI,KAAK,OAAO,YACd,UAAU,aAAa;KAAE,aAAa;KAAa,WAAW;IAAc;IAC9E,IAAI,KAAK,OAAO,gBACd,UAAU,iBAAiB;KAAE,aAAa;KAAgB,WAAW;IAAiB;IAExF,MAAM,UAAU,KAAK,KAAK,QAAQ,WAAW,WAAW,OAAO,KAAK,KAAK;IACzE,UAAU,SAAS,EAAE,WAAW,KAAK,CAAC;IACtC,cAAc,KAAK,SAAS,iBAAiB,GAAG,KAAK,UAAU,IAAI,YAAY,MAAM,CAAC,CAAC;IACvF,cAAc,KAAK,SAAS,gBAAgB,GAAG,KAAK,UAAU,IAAI,WAAW,MAAM,CAAC,CAAC;IACrF,cAAc,KAAK,SAAS,aAAa,GAAG,KAAK,UAAU,WAAW,MAAM,CAAC,CAAC;IAE9E,MAAM,QAAQ,CAAC,SAAS,aAAa,UAAU,QAAQ,CAAC,GAAG;IAC3D,IAAI,KAAK,OAAO,YAAY,MAAM,KAAK,QAAQ,cAAc,QAAQ,CAAC,GAAG;IACzE,IAAI,KAAK,OAAO,gBAAgB,MAAM,KAAK,WAAW,iBAAiB,QAAQ,CAAC,GAAG;IACnF,IAAI,iBAAiB,MAAM,KAAK,mBAAmB;IACnD,IAAI,CAAC,mBAAmB,MAAM,KAAK,uBAAuB;IAE1D,OAAO;KACL,QAAQ;MACN,OAAO,IAAI,WAAW;MACtB,WAAW,IAAI;MACf,eAAe,IAAI,UAAU;KAC/B;KACA,SAAS;MAAE,OAAO,aAAa;MAAG,OAAO;MAAW,OAAO,MAAM,KAAK,GAAG;KAAE;KAC3E,SAAS,IAAI,UAAU;KAMvB,gBACE,qBAAqB,mBACjB;MAAE,MAAM;MAAa,KAAK,IAAI,UAAU;KAAa,IACrD;MAAE,MAAM;MAAc,KAAK;KAAK;KACtC,YAAY,IAAI;IAClB;GACF,SAAS,KAAK;IAIZ,IAAI,gBAAgB,KAAA,GAAW,MAAM;IACrC,MAAM,cAAc,KAAK;KACvB,SAAS,cAAc;KACvB,YAAY,KAAK,IAAI,IAAI;KAKzB,MACE,qBAAqB,UAAU,cAAe,UAAU,aAAa,qBACjE,cACA;IACR,CAAC;GACH;EACF;CACF,CAAC;CAGD,MAAM,UAAU;EACd,OAAO,OAAO,QAAQ;EACtB,UAAU,OAAO,QAAQ;EACzB,WAAW,OAAO,QAAQ;EAC1B,cAAc,OAAO,QAAQ;EAC7B,qBAAqB,OAAO,QAAQ;EACpC,mBAAmB,OAAO,QAAQ;EAClC,YAAY,OAAO,QAAQ;EAC3B,cAAc,OAAO,QAAQ;EAC7B,cAAc,OAAO,QAAQ;EAC7B,WAAW,OAAO,OAAO;EACzB,WAAW,OAAO,OAAO;CAC3B;CACA,cAAc,KAAK,KAAK,QAAQ,cAAc,GAAG,KAAK,UAAU,SAAS,MAAM,CAAC,CAAC;CAIjF,MAAM,aAAa,OAAO,QAAQ;CAClC,MAAM,YAAY,aAAa,IAAI,oBAAoB;CAEvD,MAAM,KAAe;EACnB;EACA;EACA,cAAc,OAAO,QAAQ,WAAW,qBAAqB,OAAO,QAAQ,kBAAkB,IAAA,CAAK,QAAQ,CAAC,EAAE,gBAAgB,OAAO,QAAQ,iBAAiB,QAAQ,CAAC,EAAE,OAAO,UAAU,OAAO,OAAO,QAAQ,aAAa,QAAQ,CAAC,EAAE,oBAAoB,OAAO,QAAQ,aAAa,IAAA,CAAM,QAAQ,CAAC,EAAE;EACzS;EACA,GAAI,aAAa,IACb,CACE,KAAK,WAAW,MAAM,OAAO,QAAQ,aAAa,mGAClD,EACF,IACA,CAAC;EACL;EACA;EACA;EACA;EACA,GAAG,OAAO,QAAQ,OAAO,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,KAC5C,CAAC,IAAI,OACJ,KAAK,GAAG,MAAM,EAAE,WAAW,IAAA,CAAK,QAAQ,CAAC,EAAE,MAAM,EAAE,UAAU,QAAQ,CAAC,EAAE,MAAM,EAAE,aAAa,QAAQ,CAAC,EAAE,GAC5G;EACA;EACA;EACA;EACA;EACA;EACA,GAAG,OAAO,QAAQ,OAAO,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,KAC5C,CAAC,IAAI,OACJ,KAAK,GAAG,MAAM,EAAE,WAAW,IAAA,CAAK,QAAQ,CAAC,EAAE,MAAM,EAAE,UAAU,QAAQ,CAAC,EAAE,MAAM,EAAE,aAAa,QAAQ,CAAC,EAAE,GAC5G;EACA;CACF;CACA,cAAc,KAAK,KAAK,QAAQ,YAAY,GAAG,GAAG,KAAK,IAAI,CAAC;CAE5D,OAAO,EAAE,OAAO;AAClB;;;;;AAMA,SAAS,iBACP,KACoB;CACpB,MAAM,QAAQ,KAAK;CACnB,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,KAAK,SAAS,IAAI,QAAQ,KAAA;AACrF;AAEA,SAAS,mBACP,OACA,WACqB;CACrB,IAAI,cAAc,KAAA,KAAa,MAAM,cAAc,KAAA,GAAW,OAAO;CACrE,OAAO;EAAE,GAAG;EAAO;CAAU;AAC/B"}
|
package/dist/openapi.json
CHANGED
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
"openapi": "3.1.0",
|
|
3
3
|
"info": {
|
|
4
4
|
"title": "@tangle-network/agent-eval — wire protocol",
|
|
5
|
-
"version": "0.
|
|
5
|
+
"version": "0.148.0",
|
|
6
6
|
"description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
|
|
7
7
|
"contact": {
|
|
8
8
|
"name": "Tangle Network",
|
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
import { s as ValidationError } from "./errors-Dngq5h35.js";
|
|
2
2
|
import { t as canonicalize } from "./pre-registration-DakwTRXk.js";
|
|
3
|
-
import {
|
|
4
|
-
import { i as CostLedger } from "./cost-ledger-
|
|
3
|
+
import { W as recoverTruncatedJson, n as JudgeParseError } from "./llm-judge-DWq1Ptco.js";
|
|
4
|
+
import { i as CostLedger } from "./cost-ledger-B1qx30B4.js";
|
|
5
5
|
import { t as certificationEvidenceDigest } from "./verdict-BndeTAh_.js";
|
|
6
|
-
import { f as maximumChargeForLlmRequest, h as ModelSubstitutionError, l as costReceiptFromLlm, u as costReceiptFromLlmError, v as assertServedModel } from "./llm-client-
|
|
6
|
+
import { f as maximumChargeForLlmRequest, h as ModelSubstitutionError, l as costReceiptFromLlm, u as costReceiptFromLlmError, v as assertServedModel } from "./llm-client-BMuxYoZy.js";
|
|
7
7
|
import { createHash, randomUUID } from "node:crypto";
|
|
8
8
|
import { harnessSupportsModel } from "@tangle-network/agent-interface";
|
|
9
9
|
//#region src/agent-profile.ts
|
|
@@ -583,4 +583,4 @@ function extractProducedState(events) {
|
|
|
583
583
|
//#endregion
|
|
584
584
|
export { CODING_HARNESSES as a, agentProfileId as c, harnessAxisOf as d, verifyCompletion as i, agentProfileModelId as l, completionVerdict as n, HARNESS_NATIVE_MODEL as o, createLlmCorrectnessChecker as r, agentProfileHash as s, extractProducedState as t, expandProfileAxes as u };
|
|
585
585
|
|
|
586
|
-
//# sourceMappingURL=produced-state-
|
|
586
|
+
//# sourceMappingURL=produced-state-C0oJ4vr-.js.map
|