agentv 4.43.0-next.1 → 4.44.0-next.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/dist/{artifact-writer-STR7E5UO.js → artifact-writer-72MQQELS.js} +4 -4
- package/dist/{chunk-IFQXBQSY.js → chunk-2N3YSEMI.js} +6 -6
- package/dist/{chunk-RFJM2UBD.js → chunk-5DYGN547.js} +65 -12
- package/dist/chunk-5DYGN547.js.map +1 -0
- package/dist/{chunk-ROLNQBYU.js → chunk-CXHYTYHU.js} +8 -8
- package/dist/{chunk-ROLNQBYU.js.map → chunk-CXHYTYHU.js.map} +1 -1
- package/dist/{chunk-E2NL57MZ.js → chunk-KAGZRUO6.js} +2 -2
- package/dist/{chunk-5NDSGGMJ.js → chunk-NCNDA3F6.js} +2 -2
- package/dist/cli.js +5 -5
- package/dist/{dist-FTTL5IOY.js → dist-L4YCQOQ6.js} +3 -3
- package/dist/index.js +5 -5
- package/dist/{interactive-RE4DXH4A.js → interactive-DRNRKLEQ.js} +5 -5
- package/dist/skills/agentv-eval-writer/SKILL.md +1 -1
- package/dist/skills/agentv-eval-writer/references/config-schema.json +16 -0
- package/dist/{ts-eval-loader-HQDMXQQV-LBW3N3W6.js → ts-eval-loader-S7ESROX2-6CB3HB6H.js} +2 -2
- package/package.json +1 -1
- package/dist/chunk-RFJM2UBD.js.map +0 -1
- /package/dist/{artifact-writer-STR7E5UO.js.map → artifact-writer-72MQQELS.js.map} +0 -0
- /package/dist/{chunk-IFQXBQSY.js.map → chunk-2N3YSEMI.js.map} +0 -0
- /package/dist/{chunk-E2NL57MZ.js.map → chunk-KAGZRUO6.js.map} +0 -0
- /package/dist/{chunk-5NDSGGMJ.js.map → chunk-NCNDA3F6.js.map} +0 -0
- /package/dist/{dist-FTTL5IOY.js.map → dist-L4YCQOQ6.js.map} +0 -0
- /package/dist/{interactive-RE4DXH4A.js.map → interactive-DRNRKLEQ.js.map} +0 -0
- /package/dist/{ts-eval-loader-HQDMXQQV-LBW3N3W6.js.map → ts-eval-loader-S7ESROX2-6CB3HB6H.js.map} +0 -0
|
@@ -4,8 +4,8 @@ import {
|
|
|
4
4
|
buildResultIndexArtifact,
|
|
5
5
|
writeArtifactsFromResults,
|
|
6
6
|
writePerTestArtifacts
|
|
7
|
-
} from "./chunk-
|
|
8
|
-
import "./chunk-
|
|
7
|
+
} from "./chunk-KAGZRUO6.js";
|
|
8
|
+
import "./chunk-NCNDA3F6.js";
|
|
9
9
|
import "./chunk-76FOHROU.js";
|
|
10
10
|
import "./chunk-BPGJ4HBU.js";
|
|
11
11
|
import {
|
|
@@ -20,7 +20,7 @@ import {
|
|
|
20
20
|
parseJsonlResults,
|
|
21
21
|
writeArtifacts,
|
|
22
22
|
writeInitialBenchmarkArtifact
|
|
23
|
-
} from "./chunk-
|
|
23
|
+
} from "./chunk-5DYGN547.js";
|
|
24
24
|
import "./chunk-NPVGBFF6.js";
|
|
25
25
|
import "./chunk-M7BUKBAF.js";
|
|
26
26
|
import "./chunk-5H446C7X.js";
|
|
@@ -41,4 +41,4 @@ export {
|
|
|
41
41
|
writeInitialBenchmarkArtifact,
|
|
42
42
|
writePerTestArtifacts
|
|
43
43
|
};
|
|
44
|
-
//# sourceMappingURL=artifact-writer-
|
|
44
|
+
//# sourceMappingURL=artifact-writer-72MQQELS.js.map
|
|
@@ -64,12 +64,12 @@ import {
|
|
|
64
64
|
validateTargetsFile,
|
|
65
65
|
validateWorkspacePaths,
|
|
66
66
|
writeRunTags
|
|
67
|
-
} from "./chunk-
|
|
67
|
+
} from "./chunk-CXHYTYHU.js";
|
|
68
68
|
import {
|
|
69
69
|
materializeEvalBundle,
|
|
70
70
|
toSnakeCaseDeep as toSnakeCaseDeep2,
|
|
71
71
|
writeArtifactsFromResults as writeArtifactsFromResults2
|
|
72
|
-
} from "./chunk-
|
|
72
|
+
} from "./chunk-KAGZRUO6.js";
|
|
73
73
|
import {
|
|
74
74
|
DEFAULT_CATEGORY,
|
|
75
75
|
deriveCategory,
|
|
@@ -86,7 +86,7 @@ import {
|
|
|
86
86
|
syncProjects,
|
|
87
87
|
transpileEvalYamlFile,
|
|
88
88
|
trimBaselineResult
|
|
89
|
-
} from "./chunk-
|
|
89
|
+
} from "./chunk-NCNDA3F6.js";
|
|
90
90
|
import {
|
|
91
91
|
AGENTV_RESULTS_ARTIFACTS_REF,
|
|
92
92
|
DEFAULT_THRESHOLD,
|
|
@@ -145,7 +145,7 @@ import {
|
|
|
145
145
|
traceEnvelopeToTranscriptMessages,
|
|
146
146
|
traceFromTranscriptJsonLines,
|
|
147
147
|
writeArtifactsFromResults
|
|
148
|
-
} from "./chunk-
|
|
148
|
+
} from "./chunk-5DYGN547.js";
|
|
149
149
|
import {
|
|
150
150
|
__commonJS,
|
|
151
151
|
__require,
|
|
@@ -4828,7 +4828,7 @@ var evalRunCommand = command({
|
|
|
4828
4828
|
},
|
|
4829
4829
|
handler: async (args) => {
|
|
4830
4830
|
if (args.evalPaths.length === 0 && process.stdin.isTTY) {
|
|
4831
|
-
const { launchInteractiveWizard } = await import("./interactive-
|
|
4831
|
+
const { launchInteractiveWizard } = await import("./interactive-DRNRKLEQ.js");
|
|
4832
4832
|
await launchInteractiveWizard();
|
|
4833
4833
|
return;
|
|
4834
4834
|
}
|
|
@@ -18765,4 +18765,4 @@ export {
|
|
|
18765
18765
|
preprocessArgv,
|
|
18766
18766
|
runCli
|
|
18767
18767
|
};
|
|
18768
|
-
//# sourceMappingURL=chunk-
|
|
18768
|
+
//# sourceMappingURL=chunk-2N3YSEMI.js.map
|
|
@@ -4056,7 +4056,7 @@ var coerce = {
|
|
|
4056
4056
|
};
|
|
4057
4057
|
var NEVER = INVALID;
|
|
4058
4058
|
|
|
4059
|
-
// ../../packages/core/dist/chunk-
|
|
4059
|
+
// ../../packages/core/dist/chunk-FMQSS3SC.js
|
|
4060
4060
|
import { parse } from "yaml";
|
|
4061
4061
|
import os from "node:os";
|
|
4062
4062
|
import path from "node:path";
|
|
@@ -5547,6 +5547,7 @@ function resolvePiCliConfig(target, env, _evalFilePath) {
|
|
|
5547
5547
|
allowLiteral: true,
|
|
5548
5548
|
optionalEnv: true
|
|
5549
5549
|
});
|
|
5550
|
+
const piCliSubprovider = normalizePiCliSubprovider(subprovider, baseUrl);
|
|
5550
5551
|
const tools = resolveOptionalString(toolsSource, env, `${target.name} pi-cli tools`, {
|
|
5551
5552
|
allowLiteral: true,
|
|
5552
5553
|
optionalEnv: true
|
|
@@ -5570,7 +5571,7 @@ function resolvePiCliConfig(target, env, _evalFilePath) {
|
|
|
5570
5571
|
const systemPrompt = typeof systemPromptSource === "string" && systemPromptSource.trim().length > 0 ? systemPromptSource.trim() : void 0;
|
|
5571
5572
|
return {
|
|
5572
5573
|
executable,
|
|
5573
|
-
subprovider,
|
|
5574
|
+
subprovider: piCliSubprovider,
|
|
5574
5575
|
model,
|
|
5575
5576
|
apiKey,
|
|
5576
5577
|
baseUrl,
|
|
@@ -5585,6 +5586,13 @@ function resolvePiCliConfig(target, env, _evalFilePath) {
|
|
|
5585
5586
|
systemPrompt
|
|
5586
5587
|
};
|
|
5587
5588
|
}
|
|
5589
|
+
function normalizePiCliSubprovider(subprovider, baseUrl) {
|
|
5590
|
+
if (!baseUrl) return subprovider;
|
|
5591
|
+
if (!subprovider || subprovider.toLowerCase() === "openai") {
|
|
5592
|
+
return "azure";
|
|
5593
|
+
}
|
|
5594
|
+
return subprovider;
|
|
5595
|
+
}
|
|
5588
5596
|
function resolveClaudeConfig(target, env, _evalFilePath) {
|
|
5589
5597
|
const executableSource = target.executable ?? target.command ?? target.binary;
|
|
5590
5598
|
const modelSource = target.model;
|
|
@@ -6261,7 +6269,7 @@ async function expandFileReferences(tests, evalFileDir) {
|
|
|
6261
6269
|
return expanded;
|
|
6262
6270
|
}
|
|
6263
6271
|
|
|
6264
|
-
// ../../packages/core/dist/chunk-
|
|
6272
|
+
// ../../packages/core/dist/chunk-BPMHQXAX.js
|
|
6265
6273
|
import path51 from "node:path";
|
|
6266
6274
|
import { pathToFileURL as pathToFileURL2 } from "node:url";
|
|
6267
6275
|
import { existsSync as existsSync7 } from "node:fs";
|
|
@@ -18926,7 +18934,7 @@ var RequestError = class _RequestError extends Error {
|
|
|
18926
18934
|
}
|
|
18927
18935
|
};
|
|
18928
18936
|
|
|
18929
|
-
// ../../packages/core/dist/chunk-
|
|
18937
|
+
// ../../packages/core/dist/chunk-BPMHQXAX.js
|
|
18930
18938
|
import { exec as execCallback } from "node:child_process";
|
|
18931
18939
|
import { readdirSync, statSync } from "node:fs";
|
|
18932
18940
|
import { readFile as readFile32, readdir as readdir2, stat as stat2 } from "node:fs/promises";
|
|
@@ -34898,11 +34906,13 @@ var RepoManager = class {
|
|
|
34898
34906
|
progress;
|
|
34899
34907
|
heartbeatMs;
|
|
34900
34908
|
timeoutMs;
|
|
34909
|
+
projectConfigDir;
|
|
34901
34910
|
constructor(verbose = false, options = {}) {
|
|
34902
34911
|
this.verbose = verbose;
|
|
34903
34912
|
this.progress = options.progress ?? true;
|
|
34904
34913
|
this.heartbeatMs = options.heartbeatMs ?? DEFAULT_HEARTBEAT_MS;
|
|
34905
34914
|
this.timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS2;
|
|
34915
|
+
this.projectConfigDir = options.projectConfigDir;
|
|
34906
34916
|
}
|
|
34907
34917
|
async runGit(args, opts) {
|
|
34908
34918
|
const startedAt = Date.now();
|
|
@@ -35005,8 +35015,7 @@ ${output}` : ""}`
|
|
|
35005
35015
|
});
|
|
35006
35016
|
});
|
|
35007
35017
|
}
|
|
35008
|
-
|
|
35009
|
-
const filePath = configPath();
|
|
35018
|
+
loadConfiguredMirrorsFrom(filePath) {
|
|
35010
35019
|
if (!existsSync5(filePath)) return {};
|
|
35011
35020
|
try {
|
|
35012
35021
|
const parsed = parseYamlValue(readFileSync3(filePath, "utf-8"));
|
|
@@ -35027,6 +35036,50 @@ ${output}` : ""}`
|
|
|
35027
35036
|
return {};
|
|
35028
35037
|
}
|
|
35029
35038
|
}
|
|
35039
|
+
findProjectAgentvDir() {
|
|
35040
|
+
if (!this.projectConfigDir) return void 0;
|
|
35041
|
+
let current = path38.resolve(this.projectConfigDir);
|
|
35042
|
+
while (true) {
|
|
35043
|
+
const candidateDir = path38.join(current, ".agentv");
|
|
35044
|
+
if (existsSync5(path38.join(candidateDir, "config.yaml")) || existsSync5(path38.join(candidateDir, "config.override.yaml"))) {
|
|
35045
|
+
return candidateDir;
|
|
35046
|
+
}
|
|
35047
|
+
const reachedRepoRoot = existsSync5(path38.join(current, ".git"));
|
|
35048
|
+
const parent = path38.dirname(current);
|
|
35049
|
+
if (reachedRepoRoot || parent === current) return void 0;
|
|
35050
|
+
current = parent;
|
|
35051
|
+
}
|
|
35052
|
+
}
|
|
35053
|
+
mergeConfiguredMirrorLayers(layers) {
|
|
35054
|
+
const mirrorsByRepo = /* @__PURE__ */ new Map();
|
|
35055
|
+
const repoByIdentity = /* @__PURE__ */ new Map();
|
|
35056
|
+
for (const layer of layers) {
|
|
35057
|
+
for (const [repo, localPath] of Object.entries(layer)) {
|
|
35058
|
+
const repoIdentity = normalizeRepoIdentity(repo);
|
|
35059
|
+
const previousRepo = repoByIdentity.get(repoIdentity);
|
|
35060
|
+
if (previousRepo) mirrorsByRepo.delete(previousRepo);
|
|
35061
|
+
repoByIdentity.set(repoIdentity, repo);
|
|
35062
|
+
mirrorsByRepo.set(repo, localPath);
|
|
35063
|
+
}
|
|
35064
|
+
}
|
|
35065
|
+
return Object.fromEntries(mirrorsByRepo);
|
|
35066
|
+
}
|
|
35067
|
+
loadConfiguredMirrors() {
|
|
35068
|
+
const globalMirrors = this.loadConfiguredMirrorsFrom(configPath());
|
|
35069
|
+
const projectAgentvDir = this.findProjectAgentvDir();
|
|
35070
|
+
if (!projectAgentvDir) return globalMirrors;
|
|
35071
|
+
const projectMirrors = this.loadConfiguredMirrorsFrom(
|
|
35072
|
+
path38.join(projectAgentvDir, "config.yaml")
|
|
35073
|
+
);
|
|
35074
|
+
const projectOverrideMirrors = this.loadConfiguredMirrorsFrom(
|
|
35075
|
+
path38.join(projectAgentvDir, "config.override.yaml")
|
|
35076
|
+
);
|
|
35077
|
+
return this.mergeConfiguredMirrorLayers([
|
|
35078
|
+
globalMirrors,
|
|
35079
|
+
projectMirrors,
|
|
35080
|
+
projectOverrideMirrors
|
|
35081
|
+
]);
|
|
35082
|
+
}
|
|
35030
35083
|
findConfiguredMirror(repoIdentity) {
|
|
35031
35084
|
const mirrors = this.loadConfiguredMirrors();
|
|
35032
35085
|
for (const [repo, localPath] of Object.entries(mirrors)) {
|
|
@@ -35548,7 +35601,7 @@ async function prepareSharedWorkspaceSetup(options) {
|
|
|
35548
35601
|
const slotsNeeded = workers;
|
|
35549
35602
|
setupLog(`acquiring ${slotsNeeded} workspace pool slot(s) (pool capacity: ${poolMaxSlots})`);
|
|
35550
35603
|
poolManager = new WorkspacePoolManager(getWorkspacePoolRoot());
|
|
35551
|
-
const poolRepoManager = new RepoManager(verbose);
|
|
35604
|
+
const poolRepoManager = new RepoManager(verbose, { projectConfigDir: evalDir });
|
|
35552
35605
|
repoManager = poolRepoManager;
|
|
35553
35606
|
for (let i = 0; i < slotsNeeded; i++) {
|
|
35554
35607
|
const slot = await poolManager.acquireWorkspace({
|
|
@@ -35597,7 +35650,7 @@ async function prepareSharedWorkspaceSetup(options) {
|
|
|
35597
35650
|
const hasReposToMaterialize = !!suiteWorkspace?.repos?.length && !usePool && !isPerTestIsolation;
|
|
35598
35651
|
const needsRepoMaterialisation = hasReposToMaterialize && (!useStaticWorkspace || staticMaterialised);
|
|
35599
35652
|
const needsPerRepoCheck = hasReposToMaterialize && useStaticWorkspace && !staticMaterialised && isYamlConfiguredPath;
|
|
35600
|
-
repoManager = repoManager ?? (needsRepoMaterialisation || needsPerRepoCheck ? new RepoManager(verbose) : void 0);
|
|
35653
|
+
repoManager = repoManager ?? (needsRepoMaterialisation || needsPerRepoCheck ? new RepoManager(verbose, { projectConfigDir: evalDir }) : void 0);
|
|
35601
35654
|
if ((needsRepoMaterialisation || needsPerRepoCheck) && repoManager && sharedWorkspacePath && suiteWorkspace?.repos) {
|
|
35602
35655
|
try {
|
|
35603
35656
|
if (needsPerRepoCheck) {
|
|
@@ -35972,7 +36025,7 @@ async function prepareEvalCaseWorkspace(options) {
|
|
|
35972
36025
|
await mkdir17(workspacePath, { recursive: true });
|
|
35973
36026
|
}
|
|
35974
36027
|
if (evalCase.workspace?.repos?.length && workspacePath) {
|
|
35975
|
-
const perCaseRepoManager = new RepoManager(setupDebug);
|
|
36028
|
+
const perCaseRepoManager = new RepoManager(setupDebug, { projectConfigDir: evalDir });
|
|
35976
36029
|
try {
|
|
35977
36030
|
if (setupDebug) {
|
|
35978
36031
|
console.log(
|
|
@@ -39815,7 +39868,7 @@ async function loadTestSuite(evalFilePath, repoRoot, options) {
|
|
|
39815
39868
|
return { tests: await loadTestsFromAgentSkills(evalFilePath) };
|
|
39816
39869
|
}
|
|
39817
39870
|
if (format === "typescript") {
|
|
39818
|
-
const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-
|
|
39871
|
+
const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-S7ESROX2-6CB3HB6H.js");
|
|
39819
39872
|
return loadTsEvalSuite2(evalFilePath, resolveToAbsolutePath(repoRoot), options);
|
|
39820
39873
|
}
|
|
39821
39874
|
const { tests, parsed, suiteWorkspacePath } = await loadTestsFromYaml(
|
|
@@ -39844,7 +39897,7 @@ async function loadTests(evalFilePath, repoRoot, options) {
|
|
|
39844
39897
|
return loadTestsFromAgentSkills(evalFilePath);
|
|
39845
39898
|
}
|
|
39846
39899
|
if (format === "typescript") {
|
|
39847
|
-
const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-
|
|
39900
|
+
const { loadTsEvalSuite: loadTsEvalSuite2 } = await import("./ts-eval-loader-S7ESROX2-6CB3HB6H.js");
|
|
39848
39901
|
const suite = await loadTsEvalSuite2(evalFilePath, resolveToAbsolutePath(repoRoot), options);
|
|
39849
39902
|
return suite.tests;
|
|
39850
39903
|
}
|
|
@@ -46368,4 +46421,4 @@ export {
|
|
|
46368
46421
|
loadTsEvalFile,
|
|
46369
46422
|
loadTsEvalSuite
|
|
46370
46423
|
};
|
|
46371
|
-
//# sourceMappingURL=chunk-
|
|
46424
|
+
//# sourceMappingURL=chunk-5DYGN547.js.map
|