@tangle-network/agent-bench 0.4.0 → 0.4.1
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 +6 -0
- package/dist/adapters.js +23 -23
- package/dist/benchmarks/_harness.d.ts +4 -1
- package/dist/benchmarks/_harness.js +3 -1
- package/dist/benchmarks/aec-bench.js +2 -2
- package/dist/benchmarks/agentbench.js +2 -2
- package/dist/benchmarks/appworld.js +2 -2
- package/dist/benchmarks/bfcl.js +2 -2
- package/dist/benchmarks/commit0.js +2 -2
- package/dist/benchmarks/crag.js +2 -2
- package/dist/benchmarks/dabstep.js +2 -2
- package/dist/benchmarks/enterpriseops-gym.js +2 -2
- package/dist/benchmarks/finresearchbench.js +2 -2
- package/dist/benchmarks/nomiracl.js +2 -2
- package/dist/benchmarks/open-rag-bench.js +2 -2
- package/dist/benchmarks/programbench.js +2 -2
- package/dist/benchmarks/ragbench.js +2 -2
- package/dist/benchmarks/swe-bench.js +2 -2
- package/dist/benchmarks/t2-ragbench.js +2 -2
- package/dist/benchmarks/tau-bench-shared.js +2 -2
- package/dist/benchmarks/tau2-bench.js +3 -3
- package/dist/benchmarks/tau3-banking.js +3 -3
- package/dist/benchmarks/terminal-bench.js +2 -2
- package/dist/benchmarks/toollm.js +2 -2
- package/dist/benchmarks/webarena-verified.js +2 -2
- package/dist/{chunk-YSMEKBTD.js → chunk-5FEQDSCT.js} +2 -2
- package/dist/{chunk-PB64GYIG.js → chunk-67ACKDCX.js} +2 -2
- package/dist/{chunk-NQG5XDSB.js → chunk-7FKBWOQT.js} +2 -2
- package/dist/{chunk-K3BQGZCT.js → chunk-BEN6IF2X.js} +2 -2
- package/dist/{chunk-SHYIRB7I.js → chunk-BZY5QARD.js} +2 -2
- package/dist/{chunk-KP5KD6EN.js → chunk-CLIKAXKH.js} +2 -2
- package/dist/{chunk-XKEFIFIC.js → chunk-CWIOBFSP.js} +2 -2
- package/dist/{chunk-IA2FBTWC.js → chunk-CXDUTWQE.js} +2 -2
- package/dist/{chunk-RH5F53JT.js → chunk-DWALFME7.js} +2 -2
- package/dist/{chunk-V7AEBY6U.js → chunk-EEOC6QPJ.js} +21 -21
- package/dist/{chunk-WSKWVEQB.js → chunk-EIETHPD5.js} +7 -3
- package/dist/chunk-EIETHPD5.js.map +1 -0
- package/dist/{chunk-SHM6MRRF.js → chunk-GC2EPS6L.js} +2 -2
- package/dist/{chunk-IZ5M6OAC.js → chunk-GCHL6XPM.js} +2 -2
- package/dist/{chunk-IFVINJ4B.js → chunk-HQ5HCCKF.js} +2 -2
- package/dist/{chunk-7GRVHU22.js → chunk-HVW25KSX.js} +2 -2
- package/dist/{chunk-3U5TXJZS.js → chunk-J6BU3NTM.js} +2 -2
- package/dist/{chunk-XYA4XSNU.js → chunk-JSQOUKXS.js} +4 -3
- package/dist/chunk-JSQOUKXS.js.map +1 -0
- package/dist/{chunk-SVR2LKYI.js → chunk-NRMGT25X.js} +2 -2
- package/dist/{chunk-MQMRLGOG.js → chunk-QZZEAHWJ.js} +2 -2
- package/dist/{chunk-RCYQEFNX.js → chunk-WG7TM7UV.js} +3 -3
- package/dist/{chunk-HWST3SED.js → chunk-XXFF3RRD.js} +2 -2
- package/dist/{chunk-SFLA7OH3.js → chunk-ZFNOM7WR.js} +3 -3
- package/dist/{chunk-Z4TZ76N7.js → chunk-ZNCCYTFG.js} +2 -2
- package/dist/index.js +23 -23
- package/package.json +1 -1
- package/pier_agents/candidate_contract.py +238 -24
- package/pier_agents/tangle_candidate.py +75 -5
- package/scripts/verify-packed-consumer.mjs +65 -1
- package/scripts/verify-pier-agent.mts +6 -4
- package/src/benchmarks/_harness.test.mts +16 -1
- package/src/benchmarks/_harness.ts +9 -2
- package/src/benchmarks/terminal-bench.ts +2 -1
- package/dist/chunk-WSKWVEQB.js.map +0 -1
- package/dist/chunk-XYA4XSNU.js.map +0 -1
- /package/dist/{chunk-YSMEKBTD.js.map → chunk-5FEQDSCT.js.map} +0 -0
- /package/dist/{chunk-PB64GYIG.js.map → chunk-67ACKDCX.js.map} +0 -0
- /package/dist/{chunk-NQG5XDSB.js.map → chunk-7FKBWOQT.js.map} +0 -0
- /package/dist/{chunk-K3BQGZCT.js.map → chunk-BEN6IF2X.js.map} +0 -0
- /package/dist/{chunk-SHYIRB7I.js.map → chunk-BZY5QARD.js.map} +0 -0
- /package/dist/{chunk-KP5KD6EN.js.map → chunk-CLIKAXKH.js.map} +0 -0
- /package/dist/{chunk-XKEFIFIC.js.map → chunk-CWIOBFSP.js.map} +0 -0
- /package/dist/{chunk-IA2FBTWC.js.map → chunk-CXDUTWQE.js.map} +0 -0
- /package/dist/{chunk-RH5F53JT.js.map → chunk-DWALFME7.js.map} +0 -0
- /package/dist/{chunk-V7AEBY6U.js.map → chunk-EEOC6QPJ.js.map} +0 -0
- /package/dist/{chunk-SHM6MRRF.js.map → chunk-GC2EPS6L.js.map} +0 -0
- /package/dist/{chunk-IZ5M6OAC.js.map → chunk-GCHL6XPM.js.map} +0 -0
- /package/dist/{chunk-IFVINJ4B.js.map → chunk-HQ5HCCKF.js.map} +0 -0
- /package/dist/{chunk-7GRVHU22.js.map → chunk-HVW25KSX.js.map} +0 -0
- /package/dist/{chunk-3U5TXJZS.js.map → chunk-J6BU3NTM.js.map} +0 -0
- /package/dist/{chunk-SVR2LKYI.js.map → chunk-NRMGT25X.js.map} +0 -0
- /package/dist/{chunk-MQMRLGOG.js.map → chunk-QZZEAHWJ.js.map} +0 -0
- /package/dist/{chunk-RCYQEFNX.js.map → chunk-WG7TM7UV.js.map} +0 -0
- /package/dist/{chunk-HWST3SED.js.map → chunk-XXFF3RRD.js.map} +0 -0
- /package/dist/{chunk-SFLA7OH3.js.map → chunk-ZFNOM7WR.js.map} +0 -0
- /package/dist/{chunk-Z4TZ76N7.js.map → chunk-ZNCCYTFG.js.map} +0 -0
package/CHANGELOG.md
CHANGED
|
@@ -1,5 +1,11 @@
|
|
|
1
1
|
# Changelog
|
|
2
2
|
|
|
3
|
+
## 0.4.1
|
|
4
|
+
|
|
5
|
+
- Read Runtime 0.105 candidate plans from their signed run cell, benchmark records, and profile activation.
|
|
6
|
+
- Allow only the fixed public executable path signed by the Runtime plan.
|
|
7
|
+
- Restore the real Pier failure/success proof against the current receipt layout.
|
|
8
|
+
|
|
3
9
|
## 0.4.0
|
|
4
10
|
|
|
5
11
|
- Add the resumable SWE improvement loop backed by the official GEPA engine.
|
package/dist/adapters.js
CHANGED
|
@@ -1,40 +1,40 @@
|
|
|
1
1
|
import {
|
|
2
2
|
ADAPTERS,
|
|
3
3
|
resolveAdapter
|
|
4
|
-
} from "./chunk-
|
|
5
|
-
import "./chunk-
|
|
4
|
+
} from "./chunk-EEOC6QPJ.js";
|
|
5
|
+
import "./chunk-CWIOBFSP.js";
|
|
6
6
|
import "./chunk-VQRS7VUC.js";
|
|
7
|
-
import "./chunk-
|
|
8
|
-
import "./chunk-
|
|
7
|
+
import "./chunk-XXFF3RRD.js";
|
|
8
|
+
import "./chunk-HQ5HCCKF.js";
|
|
9
9
|
import "./chunk-SEVJPLZC.js";
|
|
10
|
-
import "./chunk-
|
|
11
|
-
import "./chunk-
|
|
12
|
-
import "./chunk-
|
|
13
|
-
import "./chunk-
|
|
14
|
-
import "./chunk-
|
|
15
|
-
import "./chunk-
|
|
10
|
+
import "./chunk-J6BU3NTM.js";
|
|
11
|
+
import "./chunk-GC2EPS6L.js";
|
|
12
|
+
import "./chunk-WG7TM7UV.js";
|
|
13
|
+
import "./chunk-ZFNOM7WR.js";
|
|
14
|
+
import "./chunk-GCHL6XPM.js";
|
|
15
|
+
import "./chunk-JSQOUKXS.js";
|
|
16
16
|
import "./chunk-INNOYXCP.js";
|
|
17
17
|
import "./chunk-PA2ZKHJC.js";
|
|
18
18
|
import "./chunk-5H5XV76F.js";
|
|
19
19
|
import "./chunk-TBKU5XQI.js";
|
|
20
|
-
import "./chunk-
|
|
21
|
-
import "./chunk-
|
|
22
|
-
import "./chunk-
|
|
20
|
+
import "./chunk-7FKBWOQT.js";
|
|
21
|
+
import "./chunk-BZY5QARD.js";
|
|
22
|
+
import "./chunk-QZZEAHWJ.js";
|
|
23
23
|
import "./chunk-PUIRNYI7.js";
|
|
24
24
|
import "./chunk-C7T7WEK2.js";
|
|
25
25
|
import "./chunk-JTHWEDEW.js";
|
|
26
|
-
import "./chunk-
|
|
27
|
-
import "./chunk-
|
|
26
|
+
import "./chunk-ZNCCYTFG.js";
|
|
27
|
+
import "./chunk-NRMGT25X.js";
|
|
28
28
|
import "./chunk-X3BTXCJ4.js";
|
|
29
|
-
import "./chunk-
|
|
30
|
-
import "./chunk-
|
|
31
|
-
import "./chunk-
|
|
29
|
+
import "./chunk-DWALFME7.js";
|
|
30
|
+
import "./chunk-5FEQDSCT.js";
|
|
31
|
+
import "./chunk-HVW25KSX.js";
|
|
32
32
|
import "./chunk-UPAMRDX4.js";
|
|
33
|
-
import "./chunk-
|
|
34
|
-
import "./chunk-
|
|
35
|
-
import "./chunk-
|
|
36
|
-
import "./chunk-
|
|
37
|
-
import "./chunk-
|
|
33
|
+
import "./chunk-BEN6IF2X.js";
|
|
34
|
+
import "./chunk-67ACKDCX.js";
|
|
35
|
+
import "./chunk-CXDUTWQE.js";
|
|
36
|
+
import "./chunk-CLIKAXKH.js";
|
|
37
|
+
import "./chunk-EIETHPD5.js";
|
|
38
38
|
import "./chunk-53UPUNBZ.js";
|
|
39
39
|
export {
|
|
40
40
|
ADAPTERS,
|
|
@@ -31,6 +31,9 @@ declare const venvPython: string;
|
|
|
31
31
|
* deps conflict with the shared `.venv` (commit0 downgrades pydantic/sqlalchemy)
|
|
32
32
|
* get their own venv and pass its python explicitly — keeping the shared one clean. */
|
|
33
33
|
declare const venvPythonAt: (venvDir: string) => string;
|
|
34
|
+
/** Resolve an executable in an isolated venv. Relative venv paths are package-owned;
|
|
35
|
+
* absolute paths allow installed consumers to keep large environments elsewhere. */
|
|
36
|
+
declare function venvBinAt(venvDir: string, name: string): string;
|
|
34
37
|
/** Report/transcript reads are large; 256 MiB matches the SWE harness budget. */
|
|
35
38
|
declare const bigBuffer: number;
|
|
36
39
|
/** Path to a named executable inside the bench venv (e.g. `venvBin('tb')`). */
|
|
@@ -122,4 +125,4 @@ declare function stageFile(path: string, content: string): Promise<void>;
|
|
|
122
125
|
/** Sanitize an instance id into a filesystem/run-id-safe token. */
|
|
123
126
|
declare function safeRunId(prefix: string, id: string): string;
|
|
124
127
|
|
|
125
|
-
export { StagedJudgeError, type StagedRunCaptureSpec, type StagedRunSpec, benchRoot, bigBuffer, preflightVenvImports, readJsonReport, resolveBenchPython, runStagedJudge, runVenvPython, runVenvScriptStdin, safeRunId, stageFile, venvBin, venvPython, venvPythonAt };
|
|
128
|
+
export { StagedJudgeError, type StagedRunCaptureSpec, type StagedRunSpec, benchRoot, bigBuffer, preflightVenvImports, readJsonReport, resolveBenchPython, runStagedJudge, runVenvPython, runVenvScriptStdin, safeRunId, stageFile, venvBin, venvBinAt, venvPython, venvPythonAt };
|
|
@@ -11,9 +11,10 @@ import {
|
|
|
11
11
|
safeRunId,
|
|
12
12
|
stageFile,
|
|
13
13
|
venvBin,
|
|
14
|
+
venvBinAt,
|
|
14
15
|
venvPython,
|
|
15
16
|
venvPythonAt
|
|
16
|
-
} from "../chunk-
|
|
17
|
+
} from "../chunk-EIETHPD5.js";
|
|
17
18
|
export {
|
|
18
19
|
StagedJudgeError,
|
|
19
20
|
benchRoot,
|
|
@@ -27,6 +28,7 @@ export {
|
|
|
27
28
|
safeRunId,
|
|
28
29
|
stageFile,
|
|
29
30
|
venvBin,
|
|
31
|
+
venvBinAt,
|
|
30
32
|
venvPython,
|
|
31
33
|
venvPythonAt
|
|
32
34
|
};
|
|
@@ -3,8 +3,8 @@ import {
|
|
|
3
3
|
appworldToolLoopClient,
|
|
4
4
|
createAppWorldAdapter,
|
|
5
5
|
createAppWorldReactAdapter
|
|
6
|
-
} from "../chunk-
|
|
7
|
-
import "../chunk-
|
|
6
|
+
} from "../chunk-CXDUTWQE.js";
|
|
7
|
+
import "../chunk-EIETHPD5.js";
|
|
8
8
|
export {
|
|
9
9
|
appworldSolutionOutput,
|
|
10
10
|
appworldToolLoopClient,
|
package/dist/benchmarks/bfcl.js
CHANGED
package/dist/benchmarks/crag.js
CHANGED
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import {
|
|
2
2
|
createEnterpriseOpsGymAdapter,
|
|
3
3
|
enterpriseOpsTranscriptOutput
|
|
4
|
-
} from "../chunk-
|
|
5
|
-
import "../chunk-
|
|
4
|
+
} from "../chunk-5FEQDSCT.js";
|
|
5
|
+
import "../chunk-EIETHPD5.js";
|
|
6
6
|
export {
|
|
7
7
|
createEnterpriseOpsGymAdapter,
|
|
8
8
|
enterpriseOpsTranscriptOutput
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import {
|
|
2
2
|
createProgrambenchAdapter,
|
|
3
3
|
programbenchSubmissionOutput
|
|
4
|
-
} from "../chunk-
|
|
5
|
-
import "../chunk-
|
|
4
|
+
} from "../chunk-QZZEAHWJ.js";
|
|
5
|
+
import "../chunk-EIETHPD5.js";
|
|
6
6
|
export {
|
|
7
7
|
createProgrambenchAdapter,
|
|
8
8
|
programbenchSubmissionOutput
|
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
import {
|
|
2
2
|
createTau2BenchAdapter,
|
|
3
3
|
tau2ResultsOutput
|
|
4
|
-
} from "../chunk-
|
|
5
|
-
import "../chunk-
|
|
6
|
-
import "../chunk-
|
|
4
|
+
} from "../chunk-WG7TM7UV.js";
|
|
5
|
+
import "../chunk-GCHL6XPM.js";
|
|
6
|
+
import "../chunk-EIETHPD5.js";
|
|
7
7
|
export {
|
|
8
8
|
createTau2BenchAdapter,
|
|
9
9
|
tau2ResultsOutput
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import {
|
|
2
2
|
createTau3BankingAdapter
|
|
3
|
-
} from "../chunk-
|
|
4
|
-
import "../chunk-
|
|
5
|
-
import "../chunk-
|
|
3
|
+
} from "../chunk-ZFNOM7WR.js";
|
|
4
|
+
import "../chunk-GCHL6XPM.js";
|
|
5
|
+
import "../chunk-EIETHPD5.js";
|
|
6
6
|
export {
|
|
7
7
|
createTau3BankingAdapter
|
|
8
8
|
};
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import {
|
|
2
2
|
createWebArenaVerifiedAdapter,
|
|
3
3
|
webarenaOutputDirOutput
|
|
4
|
-
} from "../chunk-
|
|
5
|
-
import "../chunk-
|
|
4
|
+
} from "../chunk-XXFF3RRD.js";
|
|
5
|
+
import "../chunk-EIETHPD5.js";
|
|
6
6
|
export {
|
|
7
7
|
createWebArenaVerifiedAdapter,
|
|
8
8
|
webarenaOutputDirOutput
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import {
|
|
2
2
|
benchRoot,
|
|
3
3
|
runVenvScriptStdin
|
|
4
|
-
} from "./chunk-
|
|
4
|
+
} from "./chunk-EIETHPD5.js";
|
|
5
5
|
|
|
6
6
|
// src/benchmarks/enterpriseops-gym.ts
|
|
7
7
|
import { mkdir, readFile, rm, writeFile } from "fs/promises";
|
|
@@ -208,4 +208,4 @@ export {
|
|
|
208
208
|
enterpriseOpsTranscriptOutput,
|
|
209
209
|
createEnterpriseOpsGymAdapter
|
|
210
210
|
};
|
|
211
|
-
//# sourceMappingURL=chunk-
|
|
211
|
+
//# sourceMappingURL=chunk-5FEQDSCT.js.map
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import {
|
|
2
2
|
benchRoot
|
|
3
|
-
} from "./chunk-
|
|
3
|
+
} from "./chunk-EIETHPD5.js";
|
|
4
4
|
|
|
5
5
|
// src/benchmarks/agentbench.ts
|
|
6
6
|
import { readFile } from "fs/promises";
|
|
@@ -115,4 +115,4 @@ export {
|
|
|
115
115
|
agentbenchAnswerOutput,
|
|
116
116
|
createAgentBenchAdapter
|
|
117
117
|
};
|
|
118
|
-
//# sourceMappingURL=chunk-
|
|
118
|
+
//# sourceMappingURL=chunk-67ACKDCX.js.map
|
|
@@ -10,7 +10,7 @@ import {
|
|
|
10
10
|
} from "./chunk-X3BTXCJ4.js";
|
|
11
11
|
import {
|
|
12
12
|
benchRoot
|
|
13
|
-
} from "./chunk-
|
|
13
|
+
} from "./chunk-EIETHPD5.js";
|
|
14
14
|
|
|
15
15
|
// src/benchmarks/nomiracl.ts
|
|
16
16
|
import { readFile } from "fs/promises";
|
|
@@ -144,4 +144,4 @@ function createNoMiraclAdapter() {
|
|
|
144
144
|
export {
|
|
145
145
|
createNoMiraclAdapter
|
|
146
146
|
};
|
|
147
|
-
//# sourceMappingURL=chunk-
|
|
147
|
+
//# sourceMappingURL=chunk-7FKBWOQT.js.map
|
|
@@ -5,7 +5,7 @@ import {
|
|
|
5
5
|
runStagedJudge,
|
|
6
6
|
stageFile,
|
|
7
7
|
venvPython
|
|
8
|
-
} from "./chunk-
|
|
8
|
+
} from "./chunk-EIETHPD5.js";
|
|
9
9
|
|
|
10
10
|
// src/benchmarks/aec-bench.ts
|
|
11
11
|
import { readFile } from "fs/promises";
|
|
@@ -218,4 +218,4 @@ Fix: ensure network access to raw.githubusercontent.com, or set AEC_FIXTURES=1 t
|
|
|
218
218
|
export {
|
|
219
219
|
createAecBenchAdapter
|
|
220
220
|
};
|
|
221
|
-
//# sourceMappingURL=chunk-
|
|
221
|
+
//# sourceMappingURL=chunk-BEN6IF2X.js.map
|
|
@@ -14,7 +14,7 @@ import {
|
|
|
14
14
|
} from "./chunk-X3BTXCJ4.js";
|
|
15
15
|
import {
|
|
16
16
|
benchRoot
|
|
17
|
-
} from "./chunk-
|
|
17
|
+
} from "./chunk-EIETHPD5.js";
|
|
18
18
|
|
|
19
19
|
// src/benchmarks/open-rag-bench.ts
|
|
20
20
|
import { readFile } from "fs/promises";
|
|
@@ -117,4 +117,4 @@ function createOpenRagBenchAdapter() {
|
|
|
117
117
|
export {
|
|
118
118
|
createOpenRagBenchAdapter
|
|
119
119
|
};
|
|
120
|
-
//# sourceMappingURL=chunk-
|
|
120
|
+
//# sourceMappingURL=chunk-BZY5QARD.js.map
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import {
|
|
2
2
|
benchRoot
|
|
3
|
-
} from "./chunk-
|
|
3
|
+
} from "./chunk-EIETHPD5.js";
|
|
4
4
|
|
|
5
5
|
// src/benchmarks/bfcl.ts
|
|
6
6
|
import { readFile, stat } from "fs/promises";
|
|
@@ -273,4 +273,4 @@ export {
|
|
|
273
273
|
bfclOutput,
|
|
274
274
|
createBfclAdapter
|
|
275
275
|
};
|
|
276
|
-
//# sourceMappingURL=chunk-
|
|
276
|
+
//# sourceMappingURL=chunk-CLIKAXKH.js.map
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import {
|
|
2
2
|
benchRoot
|
|
3
|
-
} from "./chunk-
|
|
3
|
+
} from "./chunk-EIETHPD5.js";
|
|
4
4
|
|
|
5
5
|
// src/benchmarks/toollm.ts
|
|
6
6
|
import { readFile } from "fs/promises";
|
|
@@ -194,4 +194,4 @@ export {
|
|
|
194
194
|
toollmOutput,
|
|
195
195
|
createToolLlmAdapter
|
|
196
196
|
};
|
|
197
|
-
//# sourceMappingURL=chunk-
|
|
197
|
+
//# sourceMappingURL=chunk-CWIOBFSP.js.map
|
|
@@ -3,7 +3,7 @@ import {
|
|
|
3
3
|
preflightVenvImports,
|
|
4
4
|
runVenvScriptStdin,
|
|
5
5
|
venvPython
|
|
6
|
-
} from "./chunk-
|
|
6
|
+
} from "./chunk-EIETHPD5.js";
|
|
7
7
|
|
|
8
8
|
// src/benchmarks/appworld.ts
|
|
9
9
|
import { spawn } from "child_process";
|
|
@@ -315,4 +315,4 @@ export {
|
|
|
315
315
|
appworldToolLoopClient,
|
|
316
316
|
createAppWorldReactAdapter
|
|
317
317
|
};
|
|
318
|
-
//# sourceMappingURL=chunk-
|
|
318
|
+
//# sourceMappingURL=chunk-CXDUTWQE.js.map
|
|
@@ -2,7 +2,7 @@ import {
|
|
|
2
2
|
benchRoot,
|
|
3
3
|
runVenvPython,
|
|
4
4
|
runVenvScriptStdin
|
|
5
|
-
} from "./chunk-
|
|
5
|
+
} from "./chunk-EIETHPD5.js";
|
|
6
6
|
|
|
7
7
|
// src/benchmarks/dabstep.ts
|
|
8
8
|
import { join } from "path";
|
|
@@ -179,4 +179,4 @@ export {
|
|
|
179
179
|
dabstepAnswerOutput,
|
|
180
180
|
createDabstepAdapter
|
|
181
181
|
};
|
|
182
|
-
//# sourceMappingURL=chunk-
|
|
182
|
+
//# sourceMappingURL=chunk-DWALFME7.js.map
|
|
@@ -1,33 +1,33 @@
|
|
|
1
1
|
import {
|
|
2
2
|
createToolLlmAdapter
|
|
3
|
-
} from "./chunk-
|
|
3
|
+
} from "./chunk-CWIOBFSP.js";
|
|
4
4
|
import {
|
|
5
5
|
createTrataHedgeAdapter
|
|
6
6
|
} from "./chunk-VQRS7VUC.js";
|
|
7
7
|
import {
|
|
8
8
|
createWebArenaVerifiedAdapter
|
|
9
|
-
} from "./chunk-
|
|
9
|
+
} from "./chunk-XXFF3RRD.js";
|
|
10
10
|
import {
|
|
11
11
|
createRagBenchAdapter
|
|
12
|
-
} from "./chunk-
|
|
12
|
+
} from "./chunk-HQ5HCCKF.js";
|
|
13
13
|
import {
|
|
14
14
|
createSimpleQaAdapter
|
|
15
15
|
} from "./chunk-SEVJPLZC.js";
|
|
16
16
|
import {
|
|
17
17
|
createSweBenchAdapter
|
|
18
|
-
} from "./chunk-
|
|
18
|
+
} from "./chunk-J6BU3NTM.js";
|
|
19
19
|
import {
|
|
20
20
|
createT2RagBenchAdapter
|
|
21
|
-
} from "./chunk-
|
|
21
|
+
} from "./chunk-GC2EPS6L.js";
|
|
22
22
|
import {
|
|
23
23
|
createTau2BenchAdapter
|
|
24
|
-
} from "./chunk-
|
|
24
|
+
} from "./chunk-WG7TM7UV.js";
|
|
25
25
|
import {
|
|
26
26
|
createTau3BankingAdapter
|
|
27
|
-
} from "./chunk-
|
|
27
|
+
} from "./chunk-ZFNOM7WR.js";
|
|
28
28
|
import {
|
|
29
29
|
createTerminalBenchAdapter
|
|
30
|
-
} from "./chunk-
|
|
30
|
+
} from "./chunk-JSQOUKXS.js";
|
|
31
31
|
import {
|
|
32
32
|
createFramesAdapter
|
|
33
33
|
} from "./chunk-INNOYXCP.js";
|
|
@@ -42,13 +42,13 @@ import {
|
|
|
42
42
|
} from "./chunk-TBKU5XQI.js";
|
|
43
43
|
import {
|
|
44
44
|
createNoMiraclAdapter
|
|
45
|
-
} from "./chunk-
|
|
45
|
+
} from "./chunk-7FKBWOQT.js";
|
|
46
46
|
import {
|
|
47
47
|
createOpenRagBenchAdapter
|
|
48
|
-
} from "./chunk-
|
|
48
|
+
} from "./chunk-BZY5QARD.js";
|
|
49
49
|
import {
|
|
50
50
|
createProgrambenchAdapter
|
|
51
|
-
} from "./chunk-
|
|
51
|
+
} from "./chunk-QZZEAHWJ.js";
|
|
52
52
|
import {
|
|
53
53
|
createCadBenchAdapter
|
|
54
54
|
} from "./chunk-PUIRNYI7.js";
|
|
@@ -57,35 +57,35 @@ import {
|
|
|
57
57
|
} from "./chunk-C7T7WEK2.js";
|
|
58
58
|
import {
|
|
59
59
|
createCommit0Adapter
|
|
60
|
-
} from "./chunk-
|
|
60
|
+
} from "./chunk-ZNCCYTFG.js";
|
|
61
61
|
import {
|
|
62
62
|
createCragAdapter
|
|
63
|
-
} from "./chunk-
|
|
63
|
+
} from "./chunk-NRMGT25X.js";
|
|
64
64
|
import {
|
|
65
65
|
createDabstepAdapter
|
|
66
|
-
} from "./chunk-
|
|
66
|
+
} from "./chunk-DWALFME7.js";
|
|
67
67
|
import {
|
|
68
68
|
createEnterpriseOpsGymAdapter
|
|
69
|
-
} from "./chunk-
|
|
69
|
+
} from "./chunk-5FEQDSCT.js";
|
|
70
70
|
import {
|
|
71
71
|
createFinResearchBenchAdapter
|
|
72
|
-
} from "./chunk-
|
|
72
|
+
} from "./chunk-HVW25KSX.js";
|
|
73
73
|
import {
|
|
74
74
|
createFinsearchcompAdapter
|
|
75
75
|
} from "./chunk-UPAMRDX4.js";
|
|
76
76
|
import {
|
|
77
77
|
createAecBenchAdapter
|
|
78
|
-
} from "./chunk-
|
|
78
|
+
} from "./chunk-BEN6IF2X.js";
|
|
79
79
|
import {
|
|
80
80
|
createAgentBenchAdapter
|
|
81
|
-
} from "./chunk-
|
|
81
|
+
} from "./chunk-67ACKDCX.js";
|
|
82
82
|
import {
|
|
83
83
|
createAppWorldAdapter,
|
|
84
84
|
createAppWorldReactAdapter
|
|
85
|
-
} from "./chunk-
|
|
85
|
+
} from "./chunk-CXDUTWQE.js";
|
|
86
86
|
import {
|
|
87
87
|
createBfclAdapter
|
|
88
|
-
} from "./chunk-
|
|
88
|
+
} from "./chunk-CLIKAXKH.js";
|
|
89
89
|
import {
|
|
90
90
|
createCadDesignAdapter
|
|
91
91
|
} from "./chunk-53UPUNBZ.js";
|
|
@@ -141,4 +141,4 @@ export {
|
|
|
141
141
|
ADAPTERS,
|
|
142
142
|
resolveAdapter
|
|
143
143
|
};
|
|
144
|
-
//# sourceMappingURL=chunk-
|
|
144
|
+
//# sourceMappingURL=chunk-EEOC6QPJ.js.map
|
|
@@ -44,10 +44,13 @@ function resolveBenchPython(env = process.env, root = benchRoot) {
|
|
|
44
44
|
return configured;
|
|
45
45
|
}
|
|
46
46
|
var venvPython = resolveBenchPython();
|
|
47
|
-
var venvPythonAt = (venvDir) =>
|
|
47
|
+
var venvPythonAt = (venvDir) => venvBinAt(venvDir, "python");
|
|
48
|
+
function venvBinAt(venvDir, name) {
|
|
49
|
+
return join(resolve(benchRoot, venvDir), "bin", name);
|
|
50
|
+
}
|
|
48
51
|
var bigBuffer = 1024 * 1024 * 256;
|
|
49
52
|
function venvBin(name) {
|
|
50
|
-
return
|
|
53
|
+
return venvBinAt(".venv", name);
|
|
51
54
|
}
|
|
52
55
|
async function runVenvPython(script, args = [], timeoutMs = 0, python = venvPython) {
|
|
53
56
|
const { stdout } = await execFileAsync(python, ["-c", script, ...args], {
|
|
@@ -303,6 +306,7 @@ export {
|
|
|
303
306
|
resolveBenchPython,
|
|
304
307
|
venvPython,
|
|
305
308
|
venvPythonAt,
|
|
309
|
+
venvBinAt,
|
|
306
310
|
bigBuffer,
|
|
307
311
|
venvBin,
|
|
308
312
|
runVenvPython,
|
|
@@ -314,4 +318,4 @@ export {
|
|
|
314
318
|
stageFile,
|
|
315
319
|
safeRunId
|
|
316
320
|
};
|
|
317
|
-
//# sourceMappingURL=chunk-
|
|
321
|
+
//# sourceMappingURL=chunk-EIETHPD5.js.map
|