reeval 0.2.0 → 0.4.0
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/THIRD_PARTY_NOTICES.md +9 -9
- package/assets/package.json +1 -1
- package/assets/pnpm-lock.yaml +84 -33
- package/dist/agent-C9dLMCM7.js +334 -0
- package/dist/agent-C9dLMCM7.js.map +1 -0
- package/dist/{agent-D-6z0Z3F.js → agent-M70OWyfG.js} +5 -5
- package/dist/{agent-D-6z0Z3F.js.map → agent-M70OWyfG.js.map} +1 -1
- package/dist/analysis-CxSCt1YY.js.map +1 -1
- package/dist/{assemble-DtouO0iQ.js → assemble-BmVT26kf.js} +25 -17
- package/dist/assemble-BmVT26kf.js.map +1 -0
- package/dist/assemble-DZiN1p_E.js +2 -0
- package/dist/{attempt-BdRQ9jKX.js → attempt-CGPV3BZY.js} +2 -2
- package/dist/{attempt-BdRQ9jKX.js.map → attempt-CGPV3BZY.js.map} +1 -1
- package/dist/bridge/index.mjs +1620 -1793
- package/dist/bridge/package.json +2 -2
- package/dist/bridge/pnpm-lock.yaml +75 -75
- package/dist/bridge/pnpm-workspace.yaml +1 -1
- package/dist/{cli-CGuUdOPK.js → cli-BdxFLEIc.js} +166 -55
- package/dist/cli-BdxFLEIc.js.map +1 -0
- package/dist/{cli-spec-DKAu09Gi.js → cli-spec-DUeyuSIQ.js} +11 -4
- package/dist/cli-spec-DUeyuSIQ.js.map +1 -0
- package/dist/cli.js +2 -2
- package/dist/{command-CCuYRGsI.js → command-BtcBitmY.js} +12 -12
- package/dist/{command-CCuYRGsI.js.map → command-BtcBitmY.js.map} +1 -1
- package/dist/{command-BekzrvZI.js → command-ByqQ-v-j.js} +13 -13
- package/dist/{command-BekzrvZI.js.map → command-ByqQ-v-j.js.map} +1 -1
- package/dist/{command-CLKuXn6Z.js → command-CIsO5xdR.js} +11 -9
- package/dist/command-CIsO5xdR.js.map +1 -0
- package/dist/{command-CEm8PZ1N.js → command-CQiFfn-C.js} +169 -280
- package/dist/command-CQiFfn-C.js.map +1 -0
- package/dist/command-DllipyoK.js +2 -0
- package/dist/{commands-CF8mOfiu.js → commands-CPrrZ2mI.js} +29 -13
- package/dist/commands-CPrrZ2mI.js.map +1 -0
- package/dist/{commands-A4udfyKZ.js → commands-D3Dxgh60.js} +1 -1
- package/dist/{config-DqJvSqYo.js → config-B93Csktg.js} +43 -15
- package/dist/{config-DqJvSqYo.js.map → config-B93Csktg.js.map} +1 -1
- package/dist/db-DIE0bsNO.js +2 -0
- package/dist/{db-DxDzMhtB.js → db-SpvY8l1g.js} +4 -4
- package/dist/{db-DxDzMhtB.js.map → db-SpvY8l1g.js.map} +1 -1
- package/dist/{dist-BZXYHN2X.js → dist-C2k9IvL3.js} +107 -55
- package/dist/dist-C2k9IvL3.js.map +1 -0
- package/dist/{docker-Buu3a60z.js → docker-Bv1ZC6V7.js} +1 -1
- package/dist/{docker-DQDZWXF9.js → docker-C2YwD_dh.js} +3 -3
- package/dist/{docker-DQDZWXF9.js.map → docker-C2YwD_dh.js.map} +1 -1
- package/dist/{docker-w0CSdQq0.js → docker-uNaAGE2C.js} +5 -5
- package/dist/{docker-w0CSdQq0.js.map → docker-uNaAGE2C.js.map} +1 -1
- package/dist/{doctor-Bs1_pG9_.js → doctor-COH-levb.js} +88 -20
- package/dist/doctor-COH-levb.js.map +1 -0
- package/dist/{egress-QJMzVtbw.js → egress-C7Obb71D.js} +35 -19
- package/dist/egress-C7Obb71D.js.map +1 -0
- package/dist/{environment-9WHSWLxH.js → environment-lDFQY2BF.js} +116 -4
- package/dist/environment-lDFQY2BF.js.map +1 -0
- package/dist/{files-D2-qn_Dl.js → files-BO9P6msM.js} +114 -14
- package/dist/files-BO9P6msM.js.map +1 -0
- package/dist/files-BRWmQJ-9.js +2 -0
- package/dist/files-Bwr8UW4b.js +2 -0
- package/dist/{files-bF2fPuT9.js → files-CMoTBEul.js} +40 -7
- package/dist/files-CMoTBEul.js.map +1 -0
- package/dist/{git-D6hMeieS.js → git-LJp55bkf.js} +3 -3
- package/dist/{git-D6hMeieS.js.map → git-LJp55bkf.js.map} +1 -1
- package/dist/{harness-ls5aX1UO.js → harness-BbMShUC-.js} +4 -4
- package/dist/{harness-ls5aX1UO.js.map → harness-BbMShUC-.js.map} +1 -1
- package/dist/harness.d.ts +2 -2
- package/dist/harness.js +1 -1
- package/dist/{header-DzFDoMCI.js → header-sPQ9Lq2S.js} +2 -2
- package/dist/{header-DzFDoMCI.js.map → header-sPQ9Lq2S.js.map} +1 -1
- package/dist/{images-2m2U-UQc.js → images-Pph_XJ68.js} +6 -6
- package/dist/{images-2m2U-UQc.js.map → images-Pph_XJ68.js.map} +1 -1
- package/dist/index.d.ts +11 -4
- package/dist/index.js +173 -65
- package/dist/index.js.map +1 -1
- package/dist/{init-CH9UY7Bt.js → init-BcYpzHXs.js} +10 -10
- package/dist/{init-CH9UY7Bt.js.map → init-BcYpzHXs.js.map} +1 -1
- package/dist/init-YS5qBRwV.js +2 -0
- package/dist/{invocations-rXktgSAn.js → invocations-DHC946b-.js} +4 -4
- package/dist/{invocations-rXktgSAn.js.map → invocations-DHC946b-.js.map} +1 -1
- package/dist/{jiti-BOLXpKyG.js → jiti-DWsGEgxD.js} +4 -4
- package/dist/{jiti-BOLXpKyG.js.map → jiti-DWsGEgxD.js.map} +1 -1
- package/dist/{judge-D8Uf4dA9.js → judge-D6wqgTDp.js} +17 -4
- package/dist/judge-D6wqgTDp.js.map +1 -0
- package/dist/judge.d.ts +12 -8
- package/dist/judge.js +13 -5
- package/dist/judge.js.map +1 -1
- package/dist/{live-Bs3LReMV.js → live-DZkCtmNP.js} +2 -2
- package/dist/{live-Bs3LReMV.js.map → live-DZkCtmNP.js.map} +1 -1
- package/dist/{matchers-DyLvm7EG.js → matchers-DbFPZQgA.js} +9 -4
- package/dist/{matchers-DyLvm7EG.js.map → matchers-DbFPZQgA.js.map} +1 -1
- package/dist/model-CJ6USX6C.js +2 -0
- package/dist/model-CNoc8t2Y.js +2 -0
- package/dist/model-DUZtNaYR.js +228 -0
- package/dist/model-DUZtNaYR.js.map +1 -0
- package/dist/{model-CDL98xue.js → model-ORtLm2UB.js} +41 -3
- package/dist/{model-CDL98xue.js.map → model-ORtLm2UB.js.map} +1 -1
- package/dist/{model-BdO79gWz.js → model-f0wprGPm.js} +3 -3
- package/dist/{model-BdO79gWz.js.map → model-f0wprGPm.js.map} +1 -1
- package/dist/{node-kG64XxdE.js → node-OXPxBD2N.js} +3 -3
- package/dist/{node-kG64XxdE.js.map → node-OXPxBD2N.js.map} +1 -1
- package/dist/{opencode-providers-DZuD_oYt.js → opencode-providers-DJVocFOU.js} +31 -3
- package/dist/opencode-providers-DJVocFOU.js.map +1 -0
- package/dist/{options-k25fjaER.js → options-Ftb2gE01.js} +85 -26
- package/dist/options-Ftb2gE01.js.map +1 -0
- package/dist/patch-BxqggH-T.js +118 -0
- package/dist/patch-BxqggH-T.js.map +1 -0
- package/dist/{plan-gwk6dJ8c.js → plan-BkhLKrsN.js} +23 -6
- package/dist/plan-BkhLKrsN.js.map +1 -0
- package/dist/{preflight-B6hfAXhh.js → preflight-CuvY0o3B.js} +3 -3
- package/dist/{preflight-B6hfAXhh.js.map → preflight-CuvY0o3B.js.map} +1 -1
- package/dist/{preflight-D2Hce5Jp.js → preflight-D3cJETui.js} +1 -1
- package/dist/{prepare-NuWsT11j.js → prepare-BI9j2-H-.js} +5 -5
- package/dist/{prepare-NuWsT11j.js.map → prepare-BI9j2-H-.js.map} +1 -1
- package/dist/{prepare-C3fefNnV.js → prepare-CVuTk5F9.js} +2 -2
- package/dist/{provenance-Dfb8YY5Y.js → provenance-6n49wska.js} +22 -4
- package/dist/provenance-6n49wska.js.map +1 -0
- package/dist/{providers-Bfx5Mnk9.js → providers-BFapaM2D.js} +3 -3
- package/dist/{providers-Bfx5Mnk9.js.map → providers-BFapaM2D.js.map} +1 -1
- package/dist/{providers-BMIGzPPG.js → providers-Cay8pI0y.js} +1 -1
- package/dist/query-Br77sQYj.js +2 -0
- package/dist/{query-CFATHZmV.js → query-CnstTDt8.js} +2 -2
- package/dist/{query-CFATHZmV.js.map → query-CnstTDt8.js.map} +1 -1
- package/dist/questions-Cm321-Mo.js +325 -0
- package/dist/questions-Cm321-Mo.js.map +1 -0
- package/dist/{record-DQ8wzyco.js → record-CbuQH4l1.js} +7 -3
- package/dist/record-CbuQH4l1.js.map +1 -0
- package/dist/{redact-C4fL_bZP.js → redact-DJ9JCShu.js} +99 -28
- package/dist/redact-DJ9JCShu.js.map +1 -0
- package/dist/{results-BtvBJAp-.js → results-D9GgBYw5.js} +4 -4
- package/dist/{results-BtvBJAp-.js.map → results-D9GgBYw5.js.map} +1 -1
- package/dist/{root-CFDjpWuc.js → root-Zeokqt3-.js} +2 -2
- package/dist/{root-CFDjpWuc.js.map → root-Zeokqt3-.js.map} +1 -1
- package/dist/{sandbox-BJPWuQE9.js → sandbox-CO__l-i1.js} +3 -3
- package/dist/{sandbox-BJPWuQE9.js.map → sandbox-CO__l-i1.js.map} +1 -1
- package/dist/sandbox.d.ts +2 -2
- package/dist/sandbox.js +1 -1
- package/dist/server-D5Zq3I5N.js +2 -0
- package/dist/{server-5DWmKKfV.js → server-DsmIrBzi.js} +12 -10
- package/dist/{server-5DWmKKfV.js.map → server-DsmIrBzi.js.map} +1 -1
- package/dist/source-BgqURlel.js +2 -0
- package/dist/{source-Bsfa3LZH.js → source-DJ1c8jwe.js} +2 -2
- package/dist/{source-Bsfa3LZH.js.map → source-DJ1c8jwe.js.map} +1 -1
- package/dist/{src-DXx_29pq.js → src-B1OT_Zhf.js} +46 -46
- package/dist/{src-DXx_29pq.js.map → src-B1OT_Zhf.js.map} +1 -1
- package/dist/{src-BQyMiWri.js → src-BAiYD4eC.js} +48 -7
- package/dist/{src-BQyMiWri.js.map → src-BAiYD4eC.js.map} +1 -1
- package/dist/{src-N7E_mfcZ.js → src-BC5FFzE8.js} +1 -1
- package/dist/src-Bi2Y3EkE.js +3 -0
- package/dist/{src-D8KYzxBX.js → src-BmBFlEIe.js} +12 -7
- package/dist/{src-D8KYzxBX.js.map → src-BmBFlEIe.js.map} +1 -1
- package/dist/src-C8WNyopo.js +3 -0
- package/dist/src-CYz-TreZ.js +3 -0
- package/dist/src-C_91Ln3Z.js +4 -0
- package/dist/{src-Cmp-7YrD.js → src-DWk0e7ix.js} +5 -5
- package/dist/{src-Cmp-7YrD.js.map → src-DWk0e7ix.js.map} +1 -1
- package/dist/{src-AEwxd1UP.js → src-DfEAlC9s.js} +4 -4
- package/dist/{src-AEwxd1UP.js.map → src-DfEAlC9s.js.map} +1 -1
- package/dist/{src-DwwiVLqv.js → src-V-iU5X9S.js} +24 -21
- package/dist/{src-DwwiVLqv.js.map → src-V-iU5X9S.js.map} +1 -1
- package/dist/{src-C-LHyTSV.js → src-gPJFzHcK.js} +1 -1
- package/dist/src-gVzsO4_f.js +3 -0
- package/dist/{state-DZ7eqb1F.js → state-DFbJrVi3.js} +15 -15
- package/dist/{state-DZ7eqb1F.js.map → state-DFbJrVi3.js.map} +1 -1
- package/dist/state-Joc2b4qk.js +2 -0
- package/dist/studio-app/assets/index-B1Pf3LAT.css +2 -0
- package/dist/studio-app/assets/index-BhorHbrj.js +58 -0
- package/dist/studio-app/index.html +2 -2
- package/dist/{summary-D7Ttky0z.js → summary-CEMYhzXm.js} +33 -5
- package/dist/summary-CEMYhzXm.js.map +1 -0
- package/dist/summary-WR5qLj_z.js +192 -0
- package/dist/summary-WR5qLj_z.js.map +1 -0
- package/dist/tools-DN-cJcfi.js +249 -0
- package/dist/tools-DN-cJcfi.js.map +1 -0
- package/dist/{types-CYJldf5N.d.ts → types-B_CqVrOS.d.ts} +210 -30
- package/dist/view/index.d.ts +63 -1
- package/dist/view/index.js +5 -5
- package/dist/{views-BrHOv6wg.js → views-DNbUdW9d.js} +4 -4
- package/dist/{views-BrHOv6wg.js.map → views-DNbUdW9d.js.map} +1 -1
- package/dist/views-DnYbVkKg.js +3 -0
- package/dist/{worker-entry-DnYkANPG.js → worker-entry-CJmG4hhE.js} +3 -3
- package/dist/{worker-entry-DnYkANPG.js.map → worker-entry-CJmG4hhE.js.map} +1 -1
- package/dist/{workspace-BVgNhC9h.js → workspace-CmSp2MXO.js} +1 -1
- package/dist/{workspace-uPapQj16.js → workspace-QtlJtRut.js} +5 -5
- package/dist/{workspace-uPapQj16.js.map → workspace-QtlJtRut.js.map} +1 -1
- package/docker/Dockerfile +2 -0
- package/docker/reeval-path.sh +21 -0
- package/package.json +13 -12
- package/dist/assemble-Dm4gD1TU.js +0 -2
- package/dist/assemble-DtouO0iQ.js.map +0 -1
- package/dist/cli-CGuUdOPK.js.map +0 -1
- package/dist/cli-spec-DKAu09Gi.js.map +0 -1
- package/dist/command-CEm8PZ1N.js.map +0 -1
- package/dist/command-CLKuXn6Z.js.map +0 -1
- package/dist/commands-CF8mOfiu.js.map +0 -1
- package/dist/db-Cjj2ZsfQ.js +0 -2
- package/dist/dist-BZXYHN2X.js.map +0 -1
- package/dist/doctor-Bs1_pG9_.js.map +0 -1
- package/dist/egress-QJMzVtbw.js.map +0 -1
- package/dist/environment-9WHSWLxH.js.map +0 -1
- package/dist/files--6CEUXm7.js +0 -2
- package/dist/files-Bd25OAQH.js +0 -2
- package/dist/files-D2-qn_Dl.js.map +0 -1
- package/dist/files-bF2fPuT9.js.map +0 -1
- package/dist/init-DJuxvYCW.js +0 -2
- package/dist/judge-D8Uf4dA9.js.map +0 -1
- package/dist/model-v-0_Ik7S.js +0 -2
- package/dist/opencode-providers-DZuD_oYt.js.map +0 -1
- package/dist/options-k25fjaER.js.map +0 -1
- package/dist/patch-BkKje4cp.js +0 -69
- package/dist/patch-BkKje4cp.js.map +0 -1
- package/dist/plan-gwk6dJ8c.js.map +0 -1
- package/dist/provenance-Dfb8YY5Y.js.map +0 -1
- package/dist/query-DR4m6CXf.js +0 -2
- package/dist/record-DQ8wzyco.js.map +0 -1
- package/dist/redact-C4fL_bZP.js.map +0 -1
- package/dist/server-buV3adNS.js +0 -2
- package/dist/source-BXzrmof5.js +0 -2
- package/dist/src-55j1zNHH.js +0 -3
- package/dist/src-BDwSuyIo.js +0 -3
- package/dist/src-DEgff9Z7.js +0 -4
- package/dist/src-W9g-y5X5.js +0 -3
- package/dist/src-daS0wKH2.js +0 -3
- package/dist/state-Cyw3QTsN.js +0 -2
- package/dist/studio-app/assets/index-Da1XDzx8.js +0 -58
- package/dist/studio-app/assets/index-xwzWdcE_.css +0 -2
- package/dist/summary-D7Ttky0z.js.map +0 -1
- package/dist/views-CwCWLWqo.js +0 -3
|
@@ -91,7 +91,10 @@ type GitInfo = {
|
|
|
91
91
|
diffHash: string | null;
|
|
92
92
|
/** Those changes as `worktree.patch` in the artifact dir; apply on `sha` to reproduce. */
|
|
93
93
|
worktreePatch: FileRef | null;
|
|
94
|
-
/**
|
|
94
|
+
/**
|
|
95
|
+
* Why a dirty run has no `worktree.patch`: too large, artifacts: 'none', or it looked like it held
|
|
96
|
+
* a secret (then where, and which rule, never the value).
|
|
97
|
+
*/
|
|
95
98
|
worktreePatchNote: string | null;
|
|
96
99
|
commitTime: string;
|
|
97
100
|
shallow: boolean;
|
|
@@ -174,7 +177,7 @@ type AttemptRecord = {
|
|
|
174
177
|
* The models the harness's own API calls were served by, as the egress gateway saw them: named
|
|
175
178
|
* in the request (e.g. `gemini-3.8-flash` for the id `gemini-3.8-flash-low`, plus any the agent
|
|
176
179
|
* uses on the side) or, for Codex, in the response, with its effort
|
|
177
|
-
* (`gpt-5.
|
|
180
|
+
* (`gpt-5.6-luna[effort=medium]`); null when neither names one (Cursor) or nothing
|
|
178
181
|
* recorded them.
|
|
179
182
|
*/
|
|
180
183
|
servedModels: string[] | null;
|
|
@@ -255,6 +258,11 @@ type EvalEntry = {
|
|
|
255
258
|
passRate: number | null;
|
|
256
259
|
cells: CellRecord[];
|
|
257
260
|
attempts: AttemptRecord[];
|
|
261
|
+
/**
|
|
262
|
+
* The compare's questions, answered after its attempts by pairwise votes, one per question;
|
|
263
|
+
* unset without any. They annotate: `outcome` and `passRate` never depend on them.
|
|
264
|
+
*/
|
|
265
|
+
answers?: CompareAnswer[];
|
|
258
266
|
};
|
|
259
267
|
/** A variant cell's pass rate against the control's, on the same harness and sandbox. */
|
|
260
268
|
type Lift = {
|
|
@@ -423,7 +431,24 @@ type RunRecord = {
|
|
|
423
431
|
/** What ran and graded the agents: the machine, images, harnesses and judge. */
|
|
424
432
|
type Stack = {
|
|
425
433
|
es: string;
|
|
434
|
+
/** `process.versions.node`: under Bun or Deno, the Node version they're compatible with. */
|
|
426
435
|
node: string;
|
|
436
|
+
/**
|
|
437
|
+
* The runtime reeval ran on, `node`, `bun`, `deno` or another, and its own version. Absent in
|
|
438
|
+
* runs recorded before it, which ran on Node.
|
|
439
|
+
*/
|
|
440
|
+
runtime?: {
|
|
441
|
+
name: string;
|
|
442
|
+
version: string;
|
|
443
|
+
};
|
|
444
|
+
/**
|
|
445
|
+
* The language the evals are written in, `typescript` for `*.eval.ts`, and the version of its
|
|
446
|
+
* toolchain when the runner reports one. Absent in runs recorded before it.
|
|
447
|
+
*/
|
|
448
|
+
language?: {
|
|
449
|
+
name: string;
|
|
450
|
+
version?: string;
|
|
451
|
+
};
|
|
427
452
|
os: string;
|
|
428
453
|
arch: string;
|
|
429
454
|
/** Every image an attempt started from. `id` is content-addressed; `repoDigests` only exist once pushed. */
|
|
@@ -449,8 +474,12 @@ type Stack = {
|
|
|
449
474
|
*/
|
|
450
475
|
resolvedModels: string[];
|
|
451
476
|
}>;
|
|
477
|
+
/**
|
|
478
|
+
* The run's judge: the config's, with what the profile's `judge` and `defaults.judge` set over it.
|
|
479
|
+
* `config.judgeModel` and `config.threshold` name the same one.
|
|
480
|
+
*/
|
|
452
481
|
judge: {
|
|
453
|
-
/**
|
|
482
|
+
/** Its model; groups, evals and calls may use others (see each verdict). */
|
|
454
483
|
model: string | null;
|
|
455
484
|
/** Model ids the provider reported serving, across every judge call in the run. */
|
|
456
485
|
resolvedModels: string[];
|
|
@@ -611,7 +640,7 @@ type TokenUsage = {
|
|
|
611
640
|
cacheWrite: number;
|
|
612
641
|
reasoning: number | null;
|
|
613
642
|
};
|
|
614
|
-
/** One agent turn. Immutable once `
|
|
643
|
+
/** One agent turn. Immutable once `c.send()` resolves. */
|
|
615
644
|
/** A turn as data, as a record or another language's runner carries it: `Turn` without its methods. */
|
|
616
645
|
type TurnData = {
|
|
617
646
|
prompt: string;
|
|
@@ -639,17 +668,22 @@ type TurnData = {
|
|
|
639
668
|
diff: Diff;
|
|
640
669
|
};
|
|
641
670
|
type Turn = TurnData & {
|
|
642
|
-
/** Runs in the eval's sandbox. Same as `
|
|
671
|
+
/** Runs in the eval's sandbox. Same as `c.exec`. */
|
|
643
672
|
exec(command: string): Promise<ExecResult>;
|
|
644
673
|
readFile(path: string): Promise<string | null>;
|
|
645
674
|
};
|
|
646
675
|
type JSONValue = null | boolean | number | string | JSONValue[] | {
|
|
647
676
|
[key: string]: JSONValue;
|
|
648
677
|
};
|
|
649
|
-
/**
|
|
678
|
+
/**
|
|
679
|
+
* A canned judge for tests and dry runs; see `fakeJudge()` in `reeval/judge`. `score` as a
|
|
680
|
+
* function of the state the judge would see works per group, eval or call, not in the config.
|
|
681
|
+
*/
|
|
650
682
|
type FakeJudgeModel = {
|
|
651
683
|
kind: 'fake-judge';
|
|
652
|
-
score: number
|
|
684
|
+
score: number | ((state: {
|
|
685
|
+
[key: string]: JSONValue;
|
|
686
|
+
}) => number);
|
|
653
687
|
};
|
|
654
688
|
/**
|
|
655
689
|
* `'provider/model'` (e.g. `'typesafe-ai/jev-1.13.0'`, `'google/gemini-3.8-flash'`), an AI SDK
|
|
@@ -666,6 +700,13 @@ type JudgeOptions = {
|
|
|
666
700
|
providerOptions?: Record<string, Record<string, JSONValue>>;
|
|
667
701
|
/** The most judge calls to this model's provider at once, across the run. */
|
|
668
702
|
maxConcurrent?: number;
|
|
703
|
+
/**
|
|
704
|
+
* Which changed files the judge's diff shows: gitignore-style patterns over workspace paths,
|
|
705
|
+
* where the last matching pattern wins and a `!` pattern excludes. Default: every changed file.
|
|
706
|
+
*/
|
|
707
|
+
include?: string[];
|
|
708
|
+
/** Workspace files the judge sees whole, as they are now, whether or not the agent changed them. */
|
|
709
|
+
files?: string[];
|
|
669
710
|
};
|
|
670
711
|
type JudgeCallOptions = JudgeOptions & {
|
|
671
712
|
/** What to grade instead of the latest turn: another turn, or a value used as the output. */
|
|
@@ -681,19 +722,28 @@ type Verdict = {
|
|
|
681
722
|
/** The model id the provider says served the call, when it reports one. */
|
|
682
723
|
resolvedModel?: string;
|
|
683
724
|
};
|
|
684
|
-
|
|
725
|
+
/**
|
|
726
|
+
* What `c.judge()` returns when the call failed (the provider's error, after any retries): the attempt fails as a judge error once the eval's body has finished,
|
|
727
|
+
* so the judges and checks after it still run.
|
|
728
|
+
*/
|
|
729
|
+
type JudgeFailure = Omit<Verdict, 'verdict' | 'score' | 'resolvedModel'> & {
|
|
730
|
+
verdict: 'error';
|
|
731
|
+
score: null;
|
|
732
|
+
error: string;
|
|
733
|
+
};
|
|
734
|
+
type AttemptContext = {
|
|
685
735
|
/** Sends the next turn of this eval's agent session, starting the sandbox on first use. */
|
|
686
736
|
send(prompt: string): Promise<Turn>;
|
|
687
737
|
/** Asks the judge a yes/no question about the latest turn (or `options.on`). */
|
|
688
|
-
judge(question: string, options?: JudgeCallOptions): Promise<Verdict>;
|
|
738
|
+
judge(question: string, options?: JudgeCallOptions): Promise<Verdict | JudgeFailure>;
|
|
689
739
|
exec(command: string): Promise<ExecResult>;
|
|
690
740
|
/** Workspace-relative, or `~/` for the home directory; `null` when there's no such file. */
|
|
691
741
|
readFile(path: string): Promise<string | null>;
|
|
692
742
|
/**
|
|
693
743
|
* Starts the sandbox and the agent session without a turn, for checks on what the environment
|
|
694
|
-
* set up. `
|
|
744
|
+
* set up. `c.send()` starts them too; calling both starts them once.
|
|
695
745
|
*/
|
|
696
|
-
sandbox(): Promise<Pick<
|
|
746
|
+
sandbox(): Promise<Pick<AttemptContext, 'exec' | 'readFile'>>;
|
|
697
747
|
/** Every turn so far, oldest first. */
|
|
698
748
|
readonly turns: readonly Turn[];
|
|
699
749
|
/** Aborted on timeout or Ctrl-C. */
|
|
@@ -705,7 +755,7 @@ type EvalContext = {
|
|
|
705
755
|
n: number;
|
|
706
756
|
};
|
|
707
757
|
};
|
|
708
|
-
type EvalFn = (
|
|
758
|
+
type EvalFn = (c: AttemptContext) => Promise<void>;
|
|
709
759
|
type FakeHarnessDescriptor = {
|
|
710
760
|
kind: 'fake';
|
|
711
761
|
/** Cell id; defaults to `fake`. */
|
|
@@ -743,7 +793,7 @@ type FakeHarnessDescriptor = {
|
|
|
743
793
|
/** Chance (0–1) that an attempt writes no files and says so, for flaky pass rates. */
|
|
744
794
|
flakeRate?: number;
|
|
745
795
|
/**
|
|
746
|
-
* Real sandbox actions, one list per `
|
|
796
|
+
* Real sandbox actions, one list per `c.send` (the last list repeats), run in order before
|
|
747
797
|
* `files`. `exec` runs as the agent: with the sandbox's own variables, none of the checks'.
|
|
748
798
|
*/
|
|
749
799
|
script?: FakeScriptStep[][];
|
|
@@ -876,7 +926,8 @@ type GooseHarnessDescriptor = {
|
|
|
876
926
|
};
|
|
877
927
|
/**
|
|
878
928
|
* pi through `@reeval/harness-pi` (`pi --mode json` in the sandbox), made by `pi()`. It runs
|
|
879
|
-
* Google models on a Gemini API key and reports each model call's tokens. It
|
|
929
|
+
* Google models on a Gemini API key and reports each model call's tokens. It takes stdio and
|
|
930
|
+
* HTTP MCP servers, not SSE ones.
|
|
880
931
|
*/
|
|
881
932
|
type PiHarnessDescriptor = {
|
|
882
933
|
kind: 'pi';
|
|
@@ -898,7 +949,7 @@ type PiHarnessDescriptor = {
|
|
|
898
949
|
*/
|
|
899
950
|
type CodexHarnessDescriptor = {
|
|
900
951
|
kind: 'codex';
|
|
901
|
-
/** An OpenAI model id Codex runs on the Responses API, e.g. `gpt-5.
|
|
952
|
+
/** An OpenAI model id Codex runs on the Responses API, e.g. `gpt-5.6-luna`. */
|
|
902
953
|
model: string;
|
|
903
954
|
/** Cell id; defaults to `codex:<model>`. */
|
|
904
955
|
id?: string;
|
|
@@ -940,7 +991,7 @@ type DockerSandboxDescriptor = {
|
|
|
940
991
|
maxConcurrent?: number;
|
|
941
992
|
/**
|
|
942
993
|
* Run the agent, setup steps and checks as root, with the capabilities `apt-get` needs, instead
|
|
943
|
-
* of the unprivileged `
|
|
994
|
+
* of the unprivileged `reeval` user. For evals that install system packages at run time.
|
|
944
995
|
*/
|
|
945
996
|
root?: boolean;
|
|
946
997
|
/** Linux capabilities to add back, e.g. `['NET_RAW']`; every other one is dropped. */
|
|
@@ -963,7 +1014,7 @@ type Secret = {
|
|
|
963
1014
|
from?: string;
|
|
964
1015
|
/** When the host variable is unset, skip it (the record says so) instead of failing. */
|
|
965
1016
|
optional?: boolean;
|
|
966
|
-
/** `agent` (default): the container's environment. `checks`: only `
|
|
1017
|
+
/** `agent` (default): the container's environment. `checks`: only `c.exec` and `setup` steps. */
|
|
967
1018
|
to?: 'agent' | 'checks';
|
|
968
1019
|
/**
|
|
969
1020
|
* Brokers the secret: the sandbox gets a per-attempt placeholder, and the Docker egress gateway
|
|
@@ -993,6 +1044,21 @@ type HostFile = {
|
|
|
993
1044
|
kind: 'file';
|
|
994
1045
|
path: string;
|
|
995
1046
|
};
|
|
1047
|
+
/** Options for the `directory` helper. */
|
|
1048
|
+
type DirectoryOptions = {
|
|
1049
|
+
/** A workspace-relative directory (or `~/…`) to place the files under, e.g. `'app'`. */
|
|
1050
|
+
into?: string;
|
|
1051
|
+
/**
|
|
1052
|
+
* Ordered gitignore-style patterns over paths relative to the directory. A plain pattern
|
|
1053
|
+
* includes and a `!pattern` excludes; the last one that matches a file decides. With no plain
|
|
1054
|
+
* pattern, every file starts included. Matched with picomatch (`*`, `?`, `**`, `[…]`, `{a,b}`):
|
|
1055
|
+
* a pattern with a `/` in it (other than at the end) is anchored to the directory, one without
|
|
1056
|
+
* matches a name at any depth, and a trailing `/` matches a directory and everything in it.
|
|
1057
|
+
*/
|
|
1058
|
+
glob?: string[];
|
|
1059
|
+
/** `false` walks into `DEFAULT_DIRECTORY_IGNORES` too. `.git` is always skipped. */
|
|
1060
|
+
defaultIgnores?: boolean;
|
|
1061
|
+
};
|
|
996
1062
|
type McpServer = {
|
|
997
1063
|
kind: 'mcp';
|
|
998
1064
|
transport: 'stdio';
|
|
@@ -1125,6 +1191,11 @@ type EnvironmentHelpers = {
|
|
|
1125
1191
|
mcp: McpHelper;
|
|
1126
1192
|
/** A host file, relative to the file that calls it (or an absolute path or `URL`). */
|
|
1127
1193
|
file: (path: string | URL) => HostFile;
|
|
1194
|
+
/**
|
|
1195
|
+
* Every file in a host directory, as `files` entries: workspace-relative paths mapped to the
|
|
1196
|
+
* same sources `file()` returns, so spreading and overriding work. Walked when it's called.
|
|
1197
|
+
*/
|
|
1198
|
+
directory: (path: string | URL, options?: DirectoryOptions) => Record<string, HostFile>;
|
|
1128
1199
|
/** A git repository for `workspace`: a URL, `github:owner/repo`, or a local path. */
|
|
1129
1200
|
git: GitHelper;
|
|
1130
1201
|
};
|
|
@@ -1210,8 +1281,11 @@ type CompareGate = 'variants' | 'all' | 'none';
|
|
|
1210
1281
|
* `false` is not a value here: a profile turns compare off with `limit.compare`.
|
|
1211
1282
|
*/
|
|
1212
1283
|
type CompareOptions<E = Environment> = {
|
|
1213
|
-
/**
|
|
1214
|
-
|
|
1284
|
+
/**
|
|
1285
|
+
* The control. Not layered under the variants. Without it, the control is the environment the
|
|
1286
|
+
* eval would otherwise run in (the cascaded `environment`), else an empty one named `baseline`.
|
|
1287
|
+
*/
|
|
1288
|
+
baseline?: E;
|
|
1215
1289
|
/** Other environments, each planned and run as itself. */
|
|
1216
1290
|
variants: E[];
|
|
1217
1291
|
/** Which cells decide pass/fail: `variants` (default), `all`, or `none` (report only). */
|
|
@@ -1220,6 +1294,57 @@ type CompareOptions<E = Environment> = {
|
|
|
1220
1294
|
minLift?: number;
|
|
1221
1295
|
/** Harnesses (ids or descriptors) to compare on; the others run the control only. */
|
|
1222
1296
|
harness?: Array<string | HarnessDescriptor>;
|
|
1297
|
+
/**
|
|
1298
|
+
* Asked of pairs of attempts after the run, always a list. They annotate the compare with
|
|
1299
|
+
* win rates and never change pass, fail or the exit code.
|
|
1300
|
+
*/
|
|
1301
|
+
questions?: CompareQuestion[];
|
|
1302
|
+
};
|
|
1303
|
+
/**
|
|
1304
|
+
* A question for pairwise judgement: its text, or the text and the `'provider/model'` that
|
|
1305
|
+
* answers it. Without a model, `analysis.model` answers, else the judge's when it's a language model.
|
|
1306
|
+
*/
|
|
1307
|
+
type CompareQuestion = string | {
|
|
1308
|
+
question: string;
|
|
1309
|
+
model?: string;
|
|
1310
|
+
};
|
|
1311
|
+
/** Which pairs a vote is over: two arms on one harness, or two harnesses on one arm. */
|
|
1312
|
+
type CompareCut = 'within' | 'across';
|
|
1313
|
+
/** One judgement of one pair of attempts, in one order. */
|
|
1314
|
+
type PairVote = {
|
|
1315
|
+
question: string;
|
|
1316
|
+
cut: CompareCut;
|
|
1317
|
+
/** Set on a within-harness pair: the harness both arms ran on. */
|
|
1318
|
+
harness?: string;
|
|
1319
|
+
/** Set on a cross-harness pair: the arm (environment) both harnesses ran. */
|
|
1320
|
+
arm?: string;
|
|
1321
|
+
/** The run number both attempts share, from 1. */
|
|
1322
|
+
run: number;
|
|
1323
|
+
/** The two sides by name: arms within a harness, harness ids across. */
|
|
1324
|
+
a: string;
|
|
1325
|
+
b: string;
|
|
1326
|
+
/** `a`, `b` or `'tie'`, mapped back from the blinded labels. */
|
|
1327
|
+
winner: string | 'tie';
|
|
1328
|
+
/** True for the repeat that showed `b` first. */
|
|
1329
|
+
swapped: boolean;
|
|
1330
|
+
};
|
|
1331
|
+
/** A compare question's answer: every vote, and each side's wins over decided votes per cut. */
|
|
1332
|
+
type CompareAnswer = {
|
|
1333
|
+
question: string;
|
|
1334
|
+
/** `'provider/model'`, or `fake` for `fakeAnalysis()`. */
|
|
1335
|
+
model: string;
|
|
1336
|
+
/** The model id the provider says served the calls, when it reports one. */
|
|
1337
|
+
resolvedModel?: string;
|
|
1338
|
+
votes: PairVote[];
|
|
1339
|
+
/** Wins over decided pairs, per named side, per cut. */
|
|
1340
|
+
rates: Array<{
|
|
1341
|
+
cut: CompareCut;
|
|
1342
|
+
name: string;
|
|
1343
|
+
wins: number;
|
|
1344
|
+
n: number;
|
|
1345
|
+
}>;
|
|
1346
|
+
/** Why some or all of its votes are missing: the model call failed. Outcomes never change. */
|
|
1347
|
+
error?: string;
|
|
1223
1348
|
};
|
|
1224
1349
|
type PinnableKey = 'harness' | 'runs' | 'compare';
|
|
1225
1350
|
/**
|
|
@@ -1257,18 +1382,27 @@ type EvalOptionFields = {
|
|
|
1257
1382
|
pin?: PinnableKey[];
|
|
1258
1383
|
};
|
|
1259
1384
|
/**
|
|
1260
|
-
*
|
|
1261
|
-
*
|
|
1262
|
-
*
|
|
1385
|
+
* `compare: false` is not an option here: a profile's `limit.compare: false` strips variants.
|
|
1386
|
+
* A compare with a `baseline` names every arm, so the cascaded `environment` is not applied, and
|
|
1387
|
+
* the same object can't set both. Without a `baseline`, the cascaded `environment` (this object's
|
|
1388
|
+
* own, or an inherited one) is the control.
|
|
1263
1389
|
*/
|
|
1264
1390
|
type EnvironmentOrCompare<E> = {
|
|
1265
|
-
/** What this group or eval runs in
|
|
1391
|
+
/** What this group or eval runs in, and the control of a compare without a `baseline`. */
|
|
1266
1392
|
environment?: E;
|
|
1267
1393
|
compare?: never;
|
|
1268
1394
|
} | {
|
|
1269
|
-
/** Forbidden beside `compare
|
|
1395
|
+
/** Forbidden beside `compare.baseline`, which names the control. */
|
|
1270
1396
|
environment?: never;
|
|
1271
|
-
compare: CompareSpec<E
|
|
1397
|
+
compare: CompareSpec<E> & {
|
|
1398
|
+
baseline: E;
|
|
1399
|
+
};
|
|
1400
|
+
} | {
|
|
1401
|
+
/** The control: the compare has no `baseline`. */
|
|
1402
|
+
environment?: E;
|
|
1403
|
+
compare: CompareSpec<E> & {
|
|
1404
|
+
baseline?: undefined;
|
|
1405
|
+
};
|
|
1272
1406
|
};
|
|
1273
1407
|
type EvalOptions = EvalOptionFields & EnvironmentOrCompare<Environment>;
|
|
1274
1408
|
type GroupOptions = EvalOptions;
|
|
@@ -1320,16 +1454,56 @@ type Profile = {
|
|
|
1320
1454
|
/** Tag filters for this profile, like `--tag` (`'!slow'` excludes). */
|
|
1321
1455
|
tags?: string[];
|
|
1322
1456
|
};
|
|
1457
|
+
/** What a fake note is written from: the eval, the harness, and what its tools returned. */
|
|
1458
|
+
type FakeNoteInput = {
|
|
1459
|
+
eval: string;
|
|
1460
|
+
harness: string;
|
|
1461
|
+
attempts: Array<{
|
|
1462
|
+
id: string;
|
|
1463
|
+
outcome: string;
|
|
1464
|
+
}>;
|
|
1465
|
+
/** Each failed attempt the fake opened beside a passing one, and the step where they diverged. */
|
|
1466
|
+
compared: Array<{
|
|
1467
|
+
failed: string;
|
|
1468
|
+
passed: string;
|
|
1469
|
+
divergedAt: number | null;
|
|
1470
|
+
}>;
|
|
1471
|
+
};
|
|
1323
1472
|
/** A canned analysis for tests; see `fakeAnalysis()` in `reeval/judge`. */
|
|
1324
1473
|
type FakeAnalysisModel = {
|
|
1325
1474
|
kind: 'fake-analysis';
|
|
1326
1475
|
/** The model's structured answer, or a function of the digest it was given. */
|
|
1327
|
-
output
|
|
1476
|
+
output?: unknown | ((digest: unknown) => unknown);
|
|
1477
|
+
/** `analysis.agent`'s note on each eval and harness: text, a function of its tools' results, or an error to throw. */
|
|
1478
|
+
note?: string | ((input: FakeNoteInput) => string) | {
|
|
1479
|
+
error: string;
|
|
1480
|
+
};
|
|
1481
|
+
/**
|
|
1482
|
+
* How it answers compare questions: a fixed pick, a function of the blinded pair, or
|
|
1483
|
+
* `{ error }` to fail every call. Default `'tie'`.
|
|
1484
|
+
*/
|
|
1485
|
+
vote?: FakeVote | ((pair: BlindPair) => FakeVote) | {
|
|
1486
|
+
error: string;
|
|
1487
|
+
};
|
|
1488
|
+
};
|
|
1489
|
+
/** A blinded label's pick: the attempt shown first (`A`), second (`B`), or neither. */
|
|
1490
|
+
type FakeVote = 'A' | 'B' | 'tie';
|
|
1491
|
+
/** What a compare question's model is shown: the prompt, and the two diffs in slot order. */
|
|
1492
|
+
type BlindPair = {
|
|
1493
|
+
question: string;
|
|
1494
|
+
prompt: string;
|
|
1495
|
+
a: string;
|
|
1496
|
+
b: string;
|
|
1328
1497
|
};
|
|
1329
1498
|
/** `reeval analyze`: a language model (not an evaluation model) and what matters in this project. */
|
|
1330
1499
|
type AnalysisConfig = {
|
|
1331
1500
|
/** `'provider/model'`, e.g. `'google/gemini-3.8-flash'`; default: the judge's, if it's one. */
|
|
1332
1501
|
model?: string | FakeAnalysisModel;
|
|
1502
|
+
/**
|
|
1503
|
+
* `reeval run` writes a note on each eval and harness once its attempts finish, with the analysis
|
|
1504
|
+
* model (default false: one call per eval and harness, every run).
|
|
1505
|
+
*/
|
|
1506
|
+
agent?: boolean;
|
|
1333
1507
|
/** Added to the prompt, e.g. "Fewer tool calls count as a win. The skill under test is reeval-good." */
|
|
1334
1508
|
instructions?: string;
|
|
1335
1509
|
/** The digest's cap; evidence is dropped, lowest priority first, to fit (default 20000). */
|
|
@@ -1420,7 +1594,8 @@ type PhaseTimings = {
|
|
|
1420
1594
|
type CellRole = 'control' | 'variant';
|
|
1421
1595
|
/**
|
|
1422
1596
|
* One cell of an eval: a harness and an environment (`-` for none) on the eval's sandbox. With
|
|
1423
|
-
* `compare`, the control is `baseline`
|
|
1597
|
+
* `compare`, the control is `baseline` (or the cascaded environment, or an empty `baseline`) and
|
|
1598
|
+
* each variant is that environment on its own.
|
|
1424
1599
|
*/
|
|
1425
1600
|
type Cell = {
|
|
1426
1601
|
key: string;
|
|
@@ -1543,9 +1718,14 @@ type StartRetry = {
|
|
|
1543
1718
|
/** Which environment ran as the control, so the studio can show it. */
|
|
1544
1719
|
type EvalCompareMeta = {
|
|
1545
1720
|
baseline: string;
|
|
1721
|
+
/**
|
|
1722
|
+
* Set when `compare.baseline` was omitted: the control is the eval's cascaded `environment`, or
|
|
1723
|
+
* an empty environment named `baseline` when there was none.
|
|
1724
|
+
*/
|
|
1725
|
+
from?: 'environment' | 'empty';
|
|
1546
1726
|
};
|
|
1547
1727
|
/**
|
|
1548
|
-
* A sandbox with more privilege than the default `
|
|
1728
|
+
* A sandbox with more privilege than the default `reeval` user: `docker({ root: true })`, or
|
|
1549
1729
|
* `docker({ capabilities })`. One per level that set such a sandbox, however many evals inherit it.
|
|
1550
1730
|
*/
|
|
1551
1731
|
type SandboxWarning = {
|
|
@@ -1562,5 +1742,5 @@ type SandboxWarning = {
|
|
|
1562
1742
|
message: string;
|
|
1563
1743
|
};
|
|
1564
1744
|
//#endregion
|
|
1565
|
-
export {
|
|
1566
|
-
//# sourceMappingURL=types-
|
|
1745
|
+
export { OpenCodeHarnessDescriptor as $, EvalFn as A, EFFORT_ALPHA as At, GooseHarnessDescriptor as B, effortChangeLabel as Bt, EgressRecord as C, FileRef as Ct, EnvironmentInput as D, RunRecord as Dt, EnvironmentHelpers as E, JudgementRecord as Et, FakeJudgeModel as F, VERDICT_TEXT as Ft, HarnessOption as G, effortVerdict as Gt, HarnessData as H, effortOf as Ht, FakeNoteInput as I, cellContrast as It, JudgeCallOptions as J, formatSeconds as Jt, HistoryConfig as K, formatCount as Kt, FakeSandboxDescriptor as L, cellEffort as Lt, ExecResult as M, EffortSample as Mt, FakeAnalysisModel as N, EffortVerdict as Nt, EnvironmentSpec as O, StackSummary as Ot, FakeHarnessDescriptor as P, Spread as Pt, McpServer as Q, FakeScriptStep as R, cellSamples as Rt, EgressPolicy as S, EvalEntry as St, Environment as T, HistoryEntry as Tt, HarnessDescriptor as U, effortSample as Ut, GroupOptions as V, effortLabel as Vt, HarnessEdit as W, effortSummary as Wt, JudgeModel as X, JudgeFailure as Y, JudgeOptions as Z, CursorHarnessDescriptor as _, TurnStatus as _t, BlindPair as a, ProfileDefaults as at, DockerSandboxDescriptor as b, AttemptRecord as bt, CellRole as c, SandboxWarning as ct, CompareAnswer as d, SetupStep as dt, Outcome as et, CompareCut as f, SkillSource as ft, ConfigJudgeOptions as g, TurnData as gt, ConfigHelpers as h, Turn as ht, AttemptContext as i, Profile as it, EvalOptions as j, EffortContrast as jt, EsConfig as k, CellEffort as kt, ClaudeCodeHarnessDescriptor as l, Secret as lt, CompareQuestion as m, TranscriptEvent as mt, AnalysisConfig as n, PiHarnessDescriptor as nt, Budget as o, ProfileLimit as ot, CompareOptions as p, TokenUsage as pt, HostFile as q, formatRatio as qt, AntigravityHarnessDescriptor as r, PinnableKey as rt, Cell as s, SandboxDescriptor as st, AgentHarnessDescriptor as t, PairVote as tt, CodexHarnessDescriptor as u, SecretTemplate as ut, Diff as v, Verdict as vt, EnvVars as w, HistoryCell as wt, Duration as x, CellRecord as xt, DirectoryOptions as y, WebEventType as yt, FakeVote as z, contrastEffort as zt };
|
|
1746
|
+
//# sourceMappingURL=types-B_CqVrOS.d.ts.map
|
package/dist/view/index.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import {
|
|
1
|
+
import { At as EFFORT_ALPHA, Bt as effortChangeLabel, Ct as FileRef, Dt as RunRecord, Et as JudgementRecord, Ft as VERDICT_TEXT, Gt as effortVerdict, Ht as effortOf, It as cellContrast, Jt as formatSeconds, Kt as formatCount, Lt as cellEffort, Mt as EffortSample, Nt as EffortVerdict, Ot as StackSummary, Pt as Spread, Rt as cellSamples, St as EvalEntry, Tt as HistoryEntry, Ut as effortSample, Vt as effortLabel, Wt as effortSummary, _t as TurnStatus, bt as AttemptRecord, ct as SandboxWarning, et as Outcome, jt as EffortContrast, kt as CellEffort, mt as TranscriptEvent, pt as TokenUsage, qt as formatRatio, wt as HistoryCell, xt as CellRecord, zt as contrastEffort } from "../types-B_CqVrOS.js";
|
|
2
2
|
//#region src/view/analysis.d.ts
|
|
3
3
|
/** Bumped when the system prompt or output schema changes: an older analysis is superseded. */
|
|
4
4
|
export declare const ANALYSIS_PROMPT_VERSION = "analysis-v2";
|
|
@@ -102,6 +102,36 @@ export type AnalysisDoc = {
|
|
|
102
102
|
}>;
|
|
103
103
|
/** Each cited fact or evidence id, summarised in one line. */
|
|
104
104
|
cited: Record<string, string>;
|
|
105
|
+
/**
|
|
106
|
+
* Why `analysis.agent`'s run summary failed; its `analysis` is then empty. `reeval analyze` treats
|
|
107
|
+
* a cached one as a miss and calls the model again.
|
|
108
|
+
*/
|
|
109
|
+
error?: string;
|
|
110
|
+
};
|
|
111
|
+
/**
|
|
112
|
+
* `analysis.agent`'s note on one eval and harness of a run, written once that harness's attempts
|
|
113
|
+
* finish: `.reeval/analysis/<runId>/notes/<eval slug>__<harness slug>.json`.
|
|
114
|
+
*/
|
|
115
|
+
export type AnalysisNote = {
|
|
116
|
+
kind: 'reeval-analysis-note';
|
|
117
|
+
schemaVersion: 1;
|
|
118
|
+
runId: string;
|
|
119
|
+
evalId: string;
|
|
120
|
+
/** The harness id, as in the cell key. */
|
|
121
|
+
harness: string;
|
|
122
|
+
model: string;
|
|
123
|
+
resolvedModel: string | null;
|
|
124
|
+
createdAt: string;
|
|
125
|
+
/** The attempt ids (`<cell key>#<n>`) the note covers. */
|
|
126
|
+
attempts: string[];
|
|
127
|
+
/** Empty when the call failed. */
|
|
128
|
+
text: string;
|
|
129
|
+
/** Why the call failed; the run's outcomes don't change. */
|
|
130
|
+
error?: string;
|
|
131
|
+
usage: {
|
|
132
|
+
inputTokens: number | null;
|
|
133
|
+
outputTokens: number | null;
|
|
134
|
+
};
|
|
105
135
|
};
|
|
106
136
|
/** One analysis file the studio can show: saved in the run's folder, or in the local cache. */
|
|
107
137
|
export type AnalysisEntry = {
|
|
@@ -112,6 +142,8 @@ export type AnalysisEntry = {
|
|
|
112
142
|
export type AnalysisIndex = {
|
|
113
143
|
runId: string;
|
|
114
144
|
entries: AnalysisEntry[];
|
|
145
|
+
/** `analysis.agent`'s notes on the run, by eval and harness; absent when there are none. */
|
|
146
|
+
notes?: AnalysisNote[];
|
|
115
147
|
};
|
|
116
148
|
export declare const analysisDataPath: (runId: string) => string;
|
|
117
149
|
/** Newest first, saved before cached at the same time. */
|
|
@@ -383,6 +415,13 @@ export declare const formatPercent: (rate: number | null) => string;
|
|
|
383
415
|
* the configured model, then the variant the harness says it resolved.
|
|
384
416
|
*/
|
|
385
417
|
export declare function harnessLabel(h: RunRecord['stack']['harnesses'][number]): string;
|
|
418
|
+
/**
|
|
419
|
+
* The runtime reeval ran on, as `node 22.14.0`, or `bun 1.4.2 (node 26.3.0)` with the Node version
|
|
420
|
+
* it's compatible with. A run recorded before `stack.runtime` ran on Node.
|
|
421
|
+
*/
|
|
422
|
+
export declare function runtimeLabel(stack: Pick<RunRecord['stack'], 'node' | 'runtime'>): string;
|
|
423
|
+
/** The evals' language as `typescript evals` or `dart 3.13.5 evals`; null when not recorded. */
|
|
424
|
+
export declare function languageLabel(stack: Pick<RunRecord['stack'], 'language'>): string | null;
|
|
386
425
|
/** Harness-reported tokens as `127k in (98k cached) · 776 out`; null when none were reported. */
|
|
387
426
|
export declare function formatUsage(usage: TokenUsage | null | undefined): string | null;
|
|
388
427
|
export declare const shortSha: (sha: string | null | undefined) => string;
|
|
@@ -475,6 +514,29 @@ export type CellRow = {
|
|
|
475
514
|
};
|
|
476
515
|
/** One row per cell: controls and variants, with each variant's recorded lift. */
|
|
477
516
|
export declare function cellRows(e: EvalEntry): CellRow[];
|
|
517
|
+
/** One side's wins over the decided votes it was in; `rate` is null when none decided. */
|
|
518
|
+
export type AnswerRate = {
|
|
519
|
+
name: string;
|
|
520
|
+
wins: number;
|
|
521
|
+
n: number;
|
|
522
|
+
rate: number | null;
|
|
523
|
+
};
|
|
524
|
+
export type AnswerRow = {
|
|
525
|
+
question: string;
|
|
526
|
+
/** The model that answered: as served, when the provider said, else as configured. */
|
|
527
|
+
model: string;
|
|
528
|
+
votes: number;
|
|
529
|
+
ties: number;
|
|
530
|
+
/** Arms against each other on one harness, then harnesses against each other on one arm. */
|
|
531
|
+
within: AnswerRate[];
|
|
532
|
+
across: AnswerRate[];
|
|
533
|
+
/** Why some or all of its votes are missing: the model call failed. */
|
|
534
|
+
error: string | null;
|
|
535
|
+
/** No votes and no error: no pair had both sides pass their checks. */
|
|
536
|
+
noPairs: boolean;
|
|
537
|
+
};
|
|
538
|
+
/** One row per compare question, with each cut's win rates in the record's order. */
|
|
539
|
+
export declare function answerRows(e: Pick<EvalEntry, 'answers'>): AnswerRow[];
|
|
478
540
|
/** Whether an eval needs the cells table: several cells, or several runs of one. */
|
|
479
541
|
export declare const isMatrix: (e: Pick<EvalEntry, 'cells'>) => boolean;
|
|
480
542
|
export type EvalHistoryPoint = {
|
package/dist/view/index.js
CHANGED
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { a as parseDataPath, c as shardPath, i as manifestPath, l as studioPath, n as artifactPath, o as recordPath, r as artifactType, s as shardOf, t as RUN_ID } from "../paths-CBvm6pap.js";
|
|
2
2
|
import { a as debugApiPath, c as formatBytes, d as requestLine, f as traceApiPath, i as TRACE_FILE, l as parseDebugLog, n as DEBUG_PHASES, o as debugLogLine, r as REQUESTS_FILE, s as debugTime, t as DEBUG_LOG, u as parseNdjson } from "../debug-D24NOQH1.js";
|
|
3
3
|
import { a as sortAnalyses, i as isSuperseded, n as DIGEST_VERSION, r as analysisDataPath, t as ANALYSIS_PROMPT_VERSION } from "../analysis-CxSCt1YY.js";
|
|
4
|
-
import { A as
|
|
5
|
-
import { t as runHeader } from "../header-
|
|
6
|
-
import {
|
|
7
|
-
import { a as liveApiPath, c as readableLiveProgress, i as isFinished, n as LIVE_FILE, o as liveCounts, r as LIVE_SCHEMA_VERSION, s as liveTranscriptPath, t as FINISHED } from "../live-
|
|
8
|
-
export { ANALYSIS_PROMPT_VERSION, COMPARE_ORDER, DEBUG_LOG, DEBUG_PHASES, DIGEST_VERSION, EFFORT_ALPHA, FINISHED, LIVE_FILE, LIVE_SCHEMA_VERSION, NewerSchemaError, REQUESTS_FILE, RUN_ID, TRACE_FILE, VERDICT_TEXT, activityLabel, analysisDataPath, artifactPath, artifactType, buildIndex, cellContrast, cellEffort, cellRows, cellSamples, cellsChanged, compare, compareView, contrastEffort, debugApiPath, debugLogLine, debugTime, defaultAttempt, effortChangeLabel, effortLabel, effortOf, effortSample, effortSummary, effortVerdict, evalHistory, evalTree, findAttempt, findEval, formatBytes, formatCount, formatDuration, formatPercent, formatRatio, formatSeconds, formatUsage, harnessLabel, hasUngatedCells, indexFiles, isFinished, isMatrix, isSuperseded, isUngatedMiss, liveApiPath, liveCounts, liveTranscriptPath, manifestPath, parseDataPath, parseDebugLog, parseNdjson, parseTranscript, readableAnalysis, readableLiveProgress, readableRecord, readableTranscriptLine, recordPath, requestLine, runHeader, runOutcome, runRows, shardOf, shardPath, shortSha, sortAnalyses, sourceChanges, splitEvalId, splitPatch, stackChanges, studioPath, traceApiPath, transcriptItems, whyChanged };
|
|
4
|
+
import { A as VERDICT_TEXT, B as effortVerdict, C as runRows, D as splitPatch, E as splitEvalId, F as effortChangeLabel, G as compare, H as formatRatio, I as effortLabel, J as whyChanged, K as sourceChanges, L as effortOf, M as cellEffort, N as cellSamples, O as transcriptItems, P as contrastEffort, R as effortSample, S as runOutcome, T as shortSha, U as formatSeconds, V as formatCount, W as cellsChanged, _ as indexFiles, a as cellRows, b as languageLabel, c as evalHistory, d as findEval, f as formatDuration, g as hasUngatedCells, h as harnessLabel, i as buildIndex, j as cellContrast, k as EFFORT_ALPHA, l as evalTree, m as formatUsage, n as activityLabel, o as compareView, p as formatPercent, q as stackChanges, r as answerRows, s as defaultAttempt, t as COMPARE_ORDER, u as findAttempt, v as isMatrix, w as runtimeLabel, x as parseTranscript, y as isUngatedMiss, z as effortSummary } from "../model-ORtLm2UB.js";
|
|
5
|
+
import { t as runHeader } from "../header-sPQ9Lq2S.js";
|
|
6
|
+
import { c as readableTranscriptLine, r as readableAnalysis, s as readableRecord, t as NewerSchemaError } from "../record-CbuQH4l1.js";
|
|
7
|
+
import { a as liveApiPath, c as readableLiveProgress, i as isFinished, n as LIVE_FILE, o as liveCounts, r as LIVE_SCHEMA_VERSION, s as liveTranscriptPath, t as FINISHED } from "../live-DZkCtmNP.js";
|
|
8
|
+
export { ANALYSIS_PROMPT_VERSION, COMPARE_ORDER, DEBUG_LOG, DEBUG_PHASES, DIGEST_VERSION, EFFORT_ALPHA, FINISHED, LIVE_FILE, LIVE_SCHEMA_VERSION, NewerSchemaError, REQUESTS_FILE, RUN_ID, TRACE_FILE, VERDICT_TEXT, activityLabel, analysisDataPath, answerRows, artifactPath, artifactType, buildIndex, cellContrast, cellEffort, cellRows, cellSamples, cellsChanged, compare, compareView, contrastEffort, debugApiPath, debugLogLine, debugTime, defaultAttempt, effortChangeLabel, effortLabel, effortOf, effortSample, effortSummary, effortVerdict, evalHistory, evalTree, findAttempt, findEval, formatBytes, formatCount, formatDuration, formatPercent, formatRatio, formatSeconds, formatUsage, harnessLabel, hasUngatedCells, indexFiles, isFinished, isMatrix, isSuperseded, isUngatedMiss, languageLabel, liveApiPath, liveCounts, liveTranscriptPath, manifestPath, parseDataPath, parseDebugLog, parseNdjson, parseTranscript, readableAnalysis, readableLiveProgress, readableRecord, readableTranscriptLine, recordPath, requestLine, runHeader, runOutcome, runRows, runtimeLabel, shardOf, shardPath, shortSha, sortAnalyses, sourceChanges, splitEvalId, splitPatch, stackChanges, studioPath, traceApiPath, transcriptItems, whyChanged };
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { r as formatDuration } from "./text-gEcKjAyp.js";
|
|
2
|
-
import {
|
|
3
|
-
import { o as warnings } from "./state-
|
|
4
|
-
import { c as
|
|
2
|
+
import { F as effortChangeLabel, W as cellsChanged, q as stackChanges } from "./model-ORtLm2UB.js";
|
|
3
|
+
import { o as warnings } from "./state-DFbJrVi3.js";
|
|
4
|
+
import { c as stateRows, l as summaryCounts, r as cellCounts, s as repeatsVerdict, t as OUTCOME_MARK } from "./summary-CEMYhzXm.js";
|
|
5
5
|
//#region src/history/views.ts
|
|
6
6
|
var RefError = class extends Error {
|
|
7
7
|
name = "RefError";
|
|
@@ -217,4 +217,4 @@ function formatShow(run) {
|
|
|
217
217
|
//#endregion
|
|
218
218
|
export { formatShow as a, resolveRun as c, formatNoBaseline as i, unreachable as l, formatCompare as n, history as o, formatHistory as r, pickBaseline as s, RefError as t };
|
|
219
219
|
|
|
220
|
-
//# sourceMappingURL=views-
|
|
220
|
+
//# sourceMappingURL=views-DNbUdW9d.js.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"views-
|
|
1
|
+
{"version":3,"file":"views-DNbUdW9d.js","names":[],"sources":["../src/history/views.ts"],"mappings":";;;;;AAsBA,IAAa,WAAb,cAA8B,MAAM;CAClC,OAAgB;AAClB;;;;;AAMA,SAAgB,WAAW,SAAyB,MAAM,UAAwB;CAChF,IAAI,QAAQ,WAAW,GAAG,MAAM,IAAI,SAAS,8BAA8B;CAC3E,MAAM,OAAO,sBAAsB,KAAK,GAAG;CAC3C,IAAI,MAAM;EACR,MAAM,QAAQ,QAAQ,QAAQ,SAAS,IAAI,OAAO,KAAK,MAAM,CAAC;EAC9D,IAAI,CAAC,OAAO,MAAM,IAAI,SAAS,kBAAkB,QAAQ,OAAO,MAAM;EACtE,OAAO;CACT;CACA,MAAM,QAAQ,QAAQ,MAAM,MAAM,EAAE,UAAU,GAAG;CACjD,IAAI,OAAO,OAAO;CAClB,MAAM,OAAO,QAAQ,QAAQ,MAAM,EAAE,MAAM,WAAW,GAAG,CAAC;CAC1D,IAAI,KAAK,WAAW,KAAK,KAAK,IAAI,OAAO,KAAK;CAC9C,IAAI,KAAK,SAAS,GAChB,MAAM,IAAI,SAAS,IAAI,IAAI,YAAY,KAAK,OAAO,0BAA0B;CAC/E,IAAI,oBAAoB,KAAK,GAAG,GAAG;EACjC,MAAM,QAAQ,QAAQ,QAAQ,MAAM,EAAE,KAAK,IAAI,WAAW,IAAI,YAAY,CAAC,CAAC;EAC5E,MAAM,SAAS,MAAM,MAAM,SAAS;EACpC,IAAI,QAAQ,OAAO;CACrB;CACA,MAAM,IAAI,SAAS,mBAAmB,IAAI,EAAE;AAC9C;;;;;;;;AASA,SAAgB,aACd,SACA,MACA,UACqB;CACrB,MAAM,UAAU,QAAQ,QAAQ,MAAM,EAAE,QAAQ,KAAK,SAAS,EAAE,WAAW,UAAU;CACrF,IAAI,YAAY,SAAS,SAAS,GAAG;EACnC,MAAM,OAAO,IAAI,IAAI,SAAS,KAAK,KAAK,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;EACvD,IAAI,OAA4B;EAChC,IAAI,WAAW,OAAO;EACtB,KAAK,MAAM,KAAK,SAAS;GACvB,MAAM,IAAI,EAAE,MAAM,KAAK,IAAI,EAAE,IAAI,GAAG,IAAI,KAAA;GACxC,IAAI,MAAM,KAAA,MAAc,IAAI,YAAa,MAAM,YAAY,QAAQ,EAAE,QAAQ,KAAK,QAAS;IACzF,OAAO;IACP,WAAW;GACb;EACF;EACA,IAAI,MAAM,OAAO;CACnB;CACA,OAAO,QAAQ,QAAQ,SAAS,MAAM;AACxC;AAEA,MAAM,YAAY,MAAqB,EAAE,MAAM,EAAE,IAAI,IAAI,MAAM,GAAG,CAAC,IAAI;AAkBvE,MAAa,eAAe,OAAqB,UAC/C,CAAC,CAAC,SAAS,CAAC,CAAC,MAAM,OAAO,CAAC,MAAM,IAAI,MAAM,IAAI,GAAG;AAEpD,SAAgB,QACd,SACA,UAAqE,CAAC,GACtE;CACA,MAAM,OAAO,QAAQ,MAAM,EAAE,QAAQ,SAAS,GAAG;CACjD,MAAM,MAAgB,CAAC;CACvB,KAAK,MAAM,OAAO,CAAC,GAAG,IAAI,CAAC,CAAC,QAAQ,GAClC,KAAK,MAAM,KAAK,IAAI,OAAO,IAAI,CAAC,IAAI,SAAS,EAAE,EAAE,GAAG,IAAI,KAAK,EAAE,EAAE;CAEnE,MAAM,SAAS,QAAQ,QAAQ,YAAY;CAiB3C,OAAO;EAAE;EAAM,MAhBY,IACxB,QAAQ,OAAO,CAAC,UAAU,GAAG,YAAY,CAAC,CAAC,SAAS,MAAM,CAAC,CAAC,CAC5D,KAAK,OAAO;GACX,MAAM,MAAkB;IAAE;IAAI,UAAU,CAAC;IAAG,OAAO,CAAC;IAAG,MAAM;IAAG,SAAS;GAAE;GAC3E,MAAM,uBAAO,IAAI,IAAoB;GACrC,KAAK,SAAS,KAAK,MAAM;IACvB,MAAM,IAAI,IAAI,MAAM,MAAM,MAAM,EAAE,OAAO,EAAE;IAC3C,IAAI,SAAS,KAAK,GAAG,WAAW,IAAI;IACpC,IAAI,CAAC,GAAG;IACR,IAAI,aAAa,MAAM,IAAI,EAAE,KAAK,GAAG,IAAI,MAAM,KAAK,CAAC;IACrD,IAAI,EAAE,YAAY,UAAU,EAAE,YAAY,QAAQ,IAAI;IACtD,IAAI,EAAE,YAAY,QAAQ,IAAI;GAChC,CAAC;GACD,OAAO;EACT,CAEgB;EAAG,UADJ,KAAK,QAAQ,MAAM,YAAY,GAAG,QAAQ,SAAS,IAAI,CAAC,CAAC,CAAC,KAAK,MAAM,EAAE,KAC5D;CAAE;AAChC;AAEA,SAAgB,cAAc,EAAE,MAAM,MAAM,YAAgD;CAC1F,IAAI,KAAK,WAAW,GAClB,OAAO;CACT,MAAM,QAAQ,KAAK;CACnB,MAAM,OAAO,KAAK,KAAK,SAAS;CAChC,MAAM,QAAQ,CACZ,GAAG,KAAK,OAAO,MAAM,KAAK,WAAW,IAAI,KAAK,IAAI,qBAAqB,MAAM,MAAM,IAAI,SAAS,KAAK,EAAE,MAAM,KAAK,MAAM,IAAI,SAAS,IAAI,EAAE,IAC3I,EACF;CACA,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,MAAM,EAAE,SAAS,SAAS,EAAE,MAAM,MAAM,GAAG,CAAC;CAChF,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,IAAI,SACf,KAAK,GAAG,MAAM,GAAG,IAAI,MAAM,SAAS,CAAC,IAAI,MAAM,KAAK,IAAI,aAAa,KAAK,KAAK,CAAC,CAChF,KAAK,EAAE;EACV,MAAM,OAAO,IAAI,UAAU,GAAG,IAAI,KAAK,GAAG,IAAI,YAAY;EAC1D,MAAM,KAAK,KAAK,MAAM,OAAO,KAAK,EAAE,IAAI,KAAK,SAAS,CAAC,EAAE,IAAI,IAAI,IAAI;CACvE;CACA,MAAM,KAAK,IAAI,8EAA8E;CAC7F,IAAI,SAAS,SAAS,GACpB,MAAM,KACJ,IACA,GAAG,SAAS,OAAO,MAAM,SAAS,WAAW,IAAI,SAAS,SAAS,4EACnE,GAAG,SAAS,KAAK,OAAO,KAAK,IAAI,CACnC;CAEF,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,MAAM,eAA8C;CAClD,WAAW;CACX,OAAO;CACP,MAAM;CACN,OAAO;CACP,OAAO;CACP,SAAS;CACT,OAAO;AACT;;;;;AAMA,MAAM,OAAO,MACX;CACE,EAAE,SAAS,SAAS,KAAK,EAAE,QAAQ,KAAK,IAAI,MAAM;CAClD,EAAE,OAAO,MAAM,SAAS,mBAAmB,EAAE,MAAM,MAAM,KAAK,IAAI,MAAM;CACxE,EAAE,OAAO,QAAQ,SAAS,qBAAqB,EAAE,MAAM,QAAQ,KAAK,IAAI,MAAM;AAChF,CAAC,CAAC,KAAK,EAAE;AAEX,MAAM,QAAyB;CAAC;CAAa;CAAS;CAAS;CAAS;CAAS;CAAW;AAAM;AAElG,MAAM,eAAe,GAAiB,UACpC,GAAG,EAAE,MAAM,IAAI,SAAS,CAAC,IAAI,EAAE,WAAW,KAAK,mBAAmB,YAAY,GAAG,KAAK,IAAI,mCAAmC,GAAG;AAElI,SAAgB,cACd,MACA,MACA,MACA,IACA,QAAsB,MACtB;CACA,MAAM,SAAS,CAAC,GAAG,IAAI,CAAC,CAAC,MAAM,GAAG,MAAM,MAAM,QAAQ,EAAE,MAAM,IAAI,MAAM,QAAQ,EAAE,MAAM,CAAC;CACzF,MAAM,QAAQ,MAAuB,IAAI,aAAa,KAAK;CAC3D,MAAM,SAAS,MAAqB,KAAK,QAAQ,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC;CACvE,MAAM,QAAQ,GAAG,MAAM,WAAW,EAAE,eAAe,MAAM,OAAO,EAAE,WAAW,MAAM,OAAO,EAAE,oBAAoB,MAAM,MAAM,EAAE;CAC9H,MAAM,UAAU,aAAa,MAAM,IAAI;CACvC,MAAM,aAAa,QAAQ,KAAK,MAAM,GAAG,EAAE,MAAM,GAAG,EAAE,KAAK,KAAK,EAAE,MAAM,CAAC,CAAC,KAAK,IAAI;CACnF,MAAM,OAAO,MAAc,EAAE,QAAQ,OAAO,KAAK;CACjD,MAAM,QAAQ,OAAO,SAAS,OAAO,EAAE,WAAW,CAAC,EAAA,CAAG,KAAK,OAAO;EAAE,KAAK;EAAG,MAAM;CAAE,EAAE,CAAC;CACvF,IAAI,IACF,OAAO;EACL,eAAe,KAAK,MAAM,MAAM,KAAK;EACrC;EACA,QAAQ,YAAY,MAAM,KAAK,EAAE,UAAU,YAAY,MAAM,KAAK,EAAE,IAAI,MAAM;EAC9E;EACA,GAAI,QAAQ,SAAS,CAAC,sBAAsB,WAAW,IAAI,EAAE,IAAI,CAAC;EAClE;EACA;EACA,GAAG,OAAO,KACP,MACC,KAAK,IAAI,EAAE,EAAE,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,aAAa,EAAE,UAAU,IAAI,CAAC,EAAE,GAC5F;EACA;EACA,GAAI,MAAM,SACN;GACE;GACA;GACA,GAAG,MAAM,KACN,EAAE,KAAK,WACN,KAAK,IAAI,IAAI,EAAE,EAAE,KAAK,IAAI,KAAK,GAAG,EAAE,KAAK,WAAW,IAAI,EAAE,KAAK,aAAa,IAAI,EAAE,GACtF;GACA;EACF,IACA,CAAC;CACP,CAAC,CAAC,KAAK,IAAI;CAEb,OAAO;EACL,QAAQ,YAAY,MAAM,KAAK,EAAE,UAAU,YAAY,MAAM,KAAK;EAClE;EACA,GAAI,QAAQ,SAAS,CAAC,kBAAkB,YAAY,IAAI,CAAC;EACzD;EACA,GAAG,OAAO,SAAS,MAAM,CACvB,KAAK,KAAK,EAAE,IAAI,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,IAAI,aAAa,EAAE,OAAO,CAAC,OAAO,EAAE,EAAE,IAAI,EAAE,KAAK,IAAI,CAAC,KAC1F,IAAI,EAAE,WAAW,CAAC,EAAA,CAAG,KAClB,MACC,SAAS,EAAE,IAAI,IAAI,CAAC,EAAE,KAAK,QAAQ,EAAE,KAAK,OAAO,QAAQ,WAAW,CAAC,MAAM,IAAI,aAAa,CAAC,CAAC,CAAC,CAAC,OAAO,OAAO,CAAC,CAAC,KAAK,KAAK,GAC9H,CACF,CAAC;CACH,CAAC,CAAC,KAAK,IAAI;AACb;;AAGA,SAAgB,iBAAiB,MAAoB,MAAoB,IAAqB;CAC5F,MAAM,UAAU,KAAK,QAAQ,MAAM,EAAE,SAAS,MAAM,CAAC,CAAC;CACtD,MAAM,OAAO,sCAAsC,KAAK,MAAM,0BAA0B,UAAU,SAAS,QAAQ,OAAO,YAAY,IAAI,MAAM,OAAO,GAAG;CAC1J,MAAM,QAAQ,MAAuB,IAAI,aAAa,KAAK;CAC3D,IAAI,IACF,OAAO;EACL,eAAe,KAAK;EACpB;EACA;EACA;EACA;EACA;EACA,GAAG,KAAK,KAAK,MAAM,KAAK,EAAE,GAAG,QAAQ,OAAO,KAAK,EAAE,KAAK,KAAK,EAAE,IAAI,EAAE,GAAG;EACxE;CACF,CAAC,CAAC,KAAK,IAAI;CAEb,OAAO;EAAC;EAAM;EAAI,GAAG,KAAK,KAAK,MAAM,KAAK,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI;CAAC,CAAC,CAAC,KAAK,IAAI;AACpF;AAEA,MAAM,aAAa,MACjB,IAAI,GAAG,OAAO,EAAE,MAAM,EAAE,IAAI,OAAO,EAAE,EAAE,EAAE,GAAG,OAAO,EAAE,EAAE,EAAE,GAAG,EAAE,cAAc,WAAW,OAAO;AAChG,MAAM,UAAU,MAAc,GAAG,KAAK,IAAI,MAAM,MAAM,KAAK,MAAM,KAAK,IAAI,CAAC,IAAI,GAAG;AAClF,MAAM,cAAc,MAAmB,GAAG,UAAU,EAAE,KAAK,IAAI,EAAE,KAAK,UAAU,EAAE,KAAK,IAAI;AAC3F,MAAM,gBAAgB,MACpB,EAAE,KAAK,UAAU,EAAE,KAAK,SAAS,kBAAkB,EAAE,KAAK,QAAQ,EAAE,KAAK,MAAM,IAAI;AAErF,SAAgB,WAAW,KAAwB;CACjD,MAAM,OAAO,UAAU,GAAG;CAC1B,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC;CACrD,MAAM,QAAQ;EACZ,OAAO,IAAI,MAAM,IAAI,cAAc,IAAI,OAAO,IAAI,IAAI,WAAW,gBAAgB,oBAAoB;EACrG,aAAa,IAAI,UAAU,KAAK,eAAe,IAAI,UAAU,EAAE,gBAAgB,IAAI;EACnF;EACA,GAAG,KAAK,KAAK,CAAC,GAAG,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,IAAI,GAAG;EACpD,GAAG,SAAS,GAAG,CAAC,CAAC,KAAK,MAAM,OAAO,GAAG;EACtC;CACF;CACA,KAAK,MAAM,KAAK,IAAI,OAAO;EACzB,MAAM,SAAS,EAAE,SAAS,WAAW;EACrC,MAAM,QAAQ,EAAE,SAAS;EACzB,MAAM,KACJ,KAAK,aAAa,EAAE,SAAS,GAAG,EAAE,KAAK,UAAU,QAAQ,KAAK,eAAe,MAAM,UAAU,MAAM,KAAK,WAAW,CAAC,KACtH;EACA,KAAK,MAAM,KAAK,EAAE,UAAU;GAC1B,MAAM,SAAS,SAAS,WAAW;GACnC,IAAI,CAAC,QACH,MAAM,KAAK,SAAS,aAAa,EAAE,SAAS,GAAG,EAAE,GAAG,IAAI,eAAe,EAAE,UAAU,GAAG;GAExF,KAAK,MAAM,KAAK,EAAE,YAChB,MAAM,KACJ,GAAG,OAAO,QAAQ,EAAE,QAAQ,GAAG,EAAE,MAAM,QAAQ,CAAC,EAAE,MAAM,EAAE,UAAU,KAAK,EAAE,UAC7E;GAEF,IAAI,EAAE,SAAS,CAAC,eAAe,CAAC,GAC9B,MAAM,KAAK,GAAG,SAAS,EAAE,MAAM,KAAK,IAAI,EAAE,MAAM,QAAQ,MAAM,IAAI,CAAC,CAAC,IAAI;EAE5E;CACF;CACA,OAAO,MAAM,KAAK,IAAI;AACxB"}
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { t as UsageError } from "./errors-CSxfJ4qu.js";
|
|
2
2
|
import { c as relativeFile } from "./ids-Bp6ut7ZD.js";
|
|
3
|
-
import { t as esJiti } from "./jiti-
|
|
4
|
-
import { d as attemptExpectError, f as createExpect, g as withAttemptExpect, h as unawaitedError, i as useSchedulerChannel, m as extendExpect, n as matchers, r as acquire, u as useEngine } from "./matchers-
|
|
3
|
+
import { t as esJiti } from "./jiti-DWsGEgxD.js";
|
|
4
|
+
import { d as attemptExpectError, f as createExpect, g as withAttemptExpect, h as unawaitedError, i as useSchedulerChannel, m as extendExpect, n as matchers, r as acquire, u as useEngine } from "./matchers-DbFPZQgA.js";
|
|
5
5
|
import { t as budgetMarker } from "./budget-DBUrH4jX.js";
|
|
6
6
|
import { t as codeFrame } from "./errors-C9w_ud7g.js";
|
|
7
7
|
import { existsSync, readFileSync } from "node:fs";
|
|
@@ -503,4 +503,4 @@ process.on("uncaughtException", (error) => worker.unhandled(error));
|
|
|
503
503
|
//#endregion
|
|
504
504
|
export {};
|
|
505
505
|
|
|
506
|
-
//# sourceMappingURL=worker-entry-
|
|
506
|
+
//# sourceMappingURL=worker-entry-CJmG4hhE.js.map
|