@vitest-evals/github-reporter 0.11.0 → 0.12.0
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/dist/cli.js +72 -249
- package/dist/cli.js.map +1 -1
- package/dist/cli.mjs +75 -247
- package/dist/cli.mjs.map +1 -1
- package/dist/index.d.mts +7 -52
- package/dist/index.d.ts +7 -52
- package/dist/index.js +65 -120
- package/dist/index.js.map +1 -1
- package/dist/index.mjs +67 -120
- package/dist/index.mjs.map +1 -1
- package/package.json +4 -1
package/dist/index.mjs
CHANGED
|
@@ -1,8 +1,10 @@
|
|
|
1
|
+
// src/collect.ts
|
|
2
|
+
import {
|
|
3
|
+
collectReportWorkspace
|
|
4
|
+
} from "@vitest-evals/core";
|
|
5
|
+
|
|
1
6
|
// src/utils.ts
|
|
2
7
|
import { posix } from "path";
|
|
3
|
-
function isRecord(value) {
|
|
4
|
-
return Boolean(value) && typeof value === "object" && !Array.isArray(value);
|
|
5
|
-
}
|
|
6
8
|
function compactLine(value, maxLength) {
|
|
7
9
|
const line = value.split(/\r?\n/).map((part) => part.trim()).find((part) => part.length > 0);
|
|
8
10
|
if (!line) {
|
|
@@ -65,17 +67,6 @@ function formatLocation(file, location) {
|
|
|
65
67
|
}
|
|
66
68
|
return `${file}:${location.line}`;
|
|
67
69
|
}
|
|
68
|
-
function normalizePathForGitHub(path, workspace) {
|
|
69
|
-
const normalized = path.replace(/\\/g, "/");
|
|
70
|
-
if (!workspace) {
|
|
71
|
-
return normalized;
|
|
72
|
-
}
|
|
73
|
-
const workspacePath = workspace.replace(/\\/g, "/").replace(/\/+$/, "");
|
|
74
|
-
if (normalized !== workspacePath && !normalized.startsWith(`${workspacePath}/`)) {
|
|
75
|
-
return normalized;
|
|
76
|
-
}
|
|
77
|
-
return posix.relative(workspacePath, normalized);
|
|
78
|
-
}
|
|
79
70
|
function escapeFence(value) {
|
|
80
71
|
return value.replace(/```/g, "'''");
|
|
81
72
|
}
|
|
@@ -91,16 +82,19 @@ function escapeCommandProperty(value) {
|
|
|
91
82
|
|
|
92
83
|
// src/collect.ts
|
|
93
84
|
function collectEvalReport(input, options = {}) {
|
|
94
|
-
const
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
85
|
+
const workspace = collectReportWorkspace(
|
|
86
|
+
{
|
|
87
|
+
report: input
|
|
88
|
+
},
|
|
89
|
+
{
|
|
90
|
+
workspace: options.workspace
|
|
91
|
+
}
|
|
99
92
|
);
|
|
93
|
+
const cases = workspace.cases.map(collectEvalCase);
|
|
100
94
|
const failures = cases.filter((testCase) => testCase.status === "failed");
|
|
101
95
|
const evalScores = cases.map((testCase) => testCase.eval?.avgScore).filter((score) => isFiniteNumber(score));
|
|
102
96
|
const usage = sumUsage(cases);
|
|
103
|
-
const durationMs =
|
|
97
|
+
const durationMs = workspace.runs[0]?.durationMs;
|
|
104
98
|
return {
|
|
105
99
|
status: input.success && failures.length === 0 ? "passed" : "failed",
|
|
106
100
|
startedAt: input.startTime,
|
|
@@ -123,35 +117,29 @@ function collectEvalReport(input, options = {}) {
|
|
|
123
117
|
failures
|
|
124
118
|
};
|
|
125
119
|
}
|
|
126
|
-
function collectEvalCase(
|
|
127
|
-
const
|
|
128
|
-
const
|
|
129
|
-
const
|
|
130
|
-
if (!evalMeta && !harnessMeta) {
|
|
131
|
-
return null;
|
|
132
|
-
}
|
|
133
|
-
const displayFile = normalizePathForGitHub(file.name, options.workspace);
|
|
134
|
-
const scores = evalMeta?.scores ?? [];
|
|
135
|
-
const harnessRun = harnessMeta?.run;
|
|
136
|
-
const toolCalls = collectToolCalls(harnessRun?.session);
|
|
120
|
+
function collectEvalCase(reportCase) {
|
|
121
|
+
const scores = (reportCase.eval?.scores ?? []).map(normalizeScore);
|
|
122
|
+
const harnessRun = reportCase.harness?.run;
|
|
123
|
+
const toolCalls = collectToolCalls(reportCase);
|
|
137
124
|
const evalCase = {
|
|
138
|
-
id:
|
|
139
|
-
file: file
|
|
140
|
-
displayFile,
|
|
141
|
-
title:
|
|
142
|
-
displayName:
|
|
143
|
-
status:
|
|
144
|
-
durationMs:
|
|
145
|
-
location:
|
|
146
|
-
failureMessages:
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
|
|
125
|
+
id: reportCase.id,
|
|
126
|
+
file: reportCase.file,
|
|
127
|
+
displayFile: reportCase.displayFile,
|
|
128
|
+
title: reportCase.title,
|
|
129
|
+
displayName: reportCase.displayName,
|
|
130
|
+
status: reportCase.status,
|
|
131
|
+
durationMs: numberField(reportCase.durationMs),
|
|
132
|
+
location: reportCase.location,
|
|
133
|
+
failureMessages: reportCase.failureMessages,
|
|
134
|
+
toolCalls,
|
|
135
|
+
eval: reportCase.eval ? {
|
|
136
|
+
avgScore: numberField(reportCase.eval.avgScore),
|
|
137
|
+
thresholdFailed: reportCase.eval.thresholdFailed,
|
|
138
|
+
output: reportCase.eval.output,
|
|
151
139
|
scores
|
|
152
140
|
} : void 0,
|
|
153
|
-
harness:
|
|
154
|
-
name:
|
|
141
|
+
harness: reportCase.harness ? {
|
|
142
|
+
name: reportCase.harness.name,
|
|
155
143
|
output: harnessRun?.output,
|
|
156
144
|
usage: harnessRun?.usage,
|
|
157
145
|
timingMs: harnessRun?.timings?.totalMs,
|
|
@@ -162,54 +150,10 @@ function collectEvalCase(file, assertion, options) {
|
|
|
162
150
|
evalCase.primaryFailure = getPrimaryFailure(evalCase);
|
|
163
151
|
return evalCase;
|
|
164
152
|
}
|
|
165
|
-
function getEvalMeta(value) {
|
|
166
|
-
if (!isRecord(value)) {
|
|
167
|
-
return void 0;
|
|
168
|
-
}
|
|
169
|
-
return {
|
|
170
|
-
scores: Array.isArray(value.scores) ? value.scores.filter(isRecord).map(normalizeScore) : void 0,
|
|
171
|
-
avgScore: numberField(value.avgScore),
|
|
172
|
-
output: value.output,
|
|
173
|
-
thresholdFailed: typeof value.thresholdFailed === "boolean" ? value.thresholdFailed : void 0
|
|
174
|
-
};
|
|
175
|
-
}
|
|
176
153
|
function normalizeScore(score) {
|
|
177
|
-
const metadata = isRecord(score.metadata) ? score.metadata : void 0;
|
|
178
154
|
return {
|
|
179
|
-
|
|
180
|
-
score: numberField(score.score) ?? null
|
|
181
|
-
metadata
|
|
182
|
-
};
|
|
183
|
-
}
|
|
184
|
-
function getHarnessMeta(value) {
|
|
185
|
-
if (!isRecord(value)) {
|
|
186
|
-
return void 0;
|
|
187
|
-
}
|
|
188
|
-
const run = isRecord(value.run) ? value.run : void 0;
|
|
189
|
-
const usage = isRecord(run?.usage) ? getUsage(run.usage) : void 0;
|
|
190
|
-
const timings = isRecord(run?.timings) ? run.timings : void 0;
|
|
191
|
-
return {
|
|
192
|
-
name: typeof value.name === "string" ? value.name : void 0,
|
|
193
|
-
run: run ? {
|
|
194
|
-
output: run.output,
|
|
195
|
-
usage,
|
|
196
|
-
timings: {
|
|
197
|
-
totalMs: typeof timings?.totalMs === "number" ? timings.totalMs : void 0
|
|
198
|
-
},
|
|
199
|
-
session: isRecord(run.session) ? {
|
|
200
|
-
messages: Array.isArray(run.session.messages) ? run.session.messages : void 0
|
|
201
|
-
} : void 0,
|
|
202
|
-
errors: Array.isArray(run.errors) ? run.errors : void 0
|
|
203
|
-
} : void 0
|
|
204
|
-
};
|
|
205
|
-
}
|
|
206
|
-
function getUsage(value) {
|
|
207
|
-
return {
|
|
208
|
-
inputTokens: numberField(value.inputTokens),
|
|
209
|
-
outputTokens: numberField(value.outputTokens),
|
|
210
|
-
reasoningTokens: numberField(value.reasoningTokens),
|
|
211
|
-
totalTokens: numberField(value.totalTokens),
|
|
212
|
-
toolCalls: numberField(value.toolCalls)
|
|
155
|
+
...score,
|
|
156
|
+
score: numberField(score.score) ?? null
|
|
213
157
|
};
|
|
214
158
|
}
|
|
215
159
|
function numberField(value) {
|
|
@@ -218,7 +162,13 @@ function numberField(value) {
|
|
|
218
162
|
function isFiniteNumber(value) {
|
|
219
163
|
return typeof value === "number" && Number.isFinite(value);
|
|
220
164
|
}
|
|
221
|
-
function collectToolCalls(
|
|
165
|
+
function collectToolCalls(reportCase) {
|
|
166
|
+
const harnessCalls = collectSessionToolCalls(
|
|
167
|
+
reportCase.harness?.run?.session
|
|
168
|
+
);
|
|
169
|
+
return harnessCalls.length > 0 ? harnessCalls : (reportCase.eval?.toolCalls ?? []).map(toToolCallSummary);
|
|
170
|
+
}
|
|
171
|
+
function collectSessionToolCalls(session) {
|
|
222
172
|
const messages = session?.messages ?? [];
|
|
223
173
|
const toolCalls = [];
|
|
224
174
|
for (const message of messages) {
|
|
@@ -226,21 +176,22 @@ function collectToolCalls(session) {
|
|
|
226
176
|
continue;
|
|
227
177
|
}
|
|
228
178
|
for (const call of message.toolCalls) {
|
|
229
|
-
|
|
230
|
-
continue;
|
|
231
|
-
}
|
|
232
|
-
toolCalls.push({
|
|
233
|
-
name: call.name,
|
|
234
|
-
error: getToolCallError(call.error),
|
|
235
|
-
durationMs: numberField(call.durationMs)
|
|
236
|
-
});
|
|
179
|
+
toolCalls.push(toToolCallSummary(call));
|
|
237
180
|
}
|
|
238
181
|
}
|
|
239
182
|
return toolCalls;
|
|
240
183
|
}
|
|
184
|
+
function toToolCallSummary(call) {
|
|
185
|
+
return {
|
|
186
|
+
name: call.name,
|
|
187
|
+
error: getToolCallError(call.error),
|
|
188
|
+
durationMs: numberField(call.durationMs)
|
|
189
|
+
};
|
|
190
|
+
}
|
|
241
191
|
function getToolCallError(value) {
|
|
242
|
-
|
|
243
|
-
|
|
192
|
+
const error = value;
|
|
193
|
+
if (value && typeof value === "object" && !Array.isArray(value) && typeof error.message === "string") {
|
|
194
|
+
return error.message;
|
|
244
195
|
}
|
|
245
196
|
if (value !== void 0) {
|
|
246
197
|
return stringifyValue(value, 240);
|
|
@@ -269,9 +220,6 @@ function stringifyReason(value) {
|
|
|
269
220
|
}
|
|
270
221
|
return typeof value === "string" ? value : stringifyValue(value, 4e3);
|
|
271
222
|
}
|
|
272
|
-
function formatDisplayName(assertion) {
|
|
273
|
-
return [...assertion.ancestorTitles, assertion.title].filter((part) => part.length > 0).join(" > ");
|
|
274
|
-
}
|
|
275
223
|
function sumUsage(cases) {
|
|
276
224
|
const usage = {
|
|
277
225
|
inputTokens: 0,
|
|
@@ -286,17 +234,16 @@ function sumUsage(cases) {
|
|
|
286
234
|
usage.outputTokens += caseUsage?.outputTokens ?? 0;
|
|
287
235
|
usage.reasoningTokens += caseUsage?.reasoningTokens ?? 0;
|
|
288
236
|
usage.totalTokens += caseUsage?.totalTokens ?? (caseUsage?.inputTokens ?? 0) + (caseUsage?.outputTokens ?? 0) + (caseUsage?.reasoningTokens ?? 0);
|
|
289
|
-
usage.toolCalls +=
|
|
237
|
+
usage.toolCalls += toolCallCount(testCase);
|
|
290
238
|
}
|
|
291
239
|
return usage;
|
|
292
240
|
}
|
|
293
|
-
function
|
|
294
|
-
const
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
return void 0;
|
|
241
|
+
function toolCallCount(testCase) {
|
|
242
|
+
const usageToolCalls = testCase.harness?.usage?.toolCalls;
|
|
243
|
+
if (usageToolCalls !== void 0) {
|
|
244
|
+
return Math.max(usageToolCalls, testCase.toolCalls.length);
|
|
298
245
|
}
|
|
299
|
-
return
|
|
246
|
+
return testCase.toolCalls.length;
|
|
300
247
|
}
|
|
301
248
|
|
|
302
249
|
// src/annotations.ts
|
|
@@ -372,9 +319,9 @@ function formatRawDetails(testCase) {
|
|
|
372
319
|
if (finalOutput !== void 0) {
|
|
373
320
|
lines.push("", "Final:", stringifyValue(finalOutput, 8e3));
|
|
374
321
|
}
|
|
375
|
-
if (testCase.
|
|
322
|
+
if (testCase.toolCalls.length) {
|
|
376
323
|
lines.push("", "Tools:");
|
|
377
|
-
for (const toolCall of testCase.
|
|
324
|
+
for (const toolCall of testCase.toolCalls) {
|
|
378
325
|
lines.push(
|
|
379
326
|
`- ${toolCall.name}: ${toolCall.error ? `error: ${toolCall.error}` : "ok"}`
|
|
380
327
|
);
|
|
@@ -595,8 +542,8 @@ function renderFailureBlock(testCase, {
|
|
|
595
542
|
""
|
|
596
543
|
);
|
|
597
544
|
}
|
|
598
|
-
if (testCase.
|
|
599
|
-
const toolCalls = testCase.
|
|
545
|
+
if (testCase.toolCalls.length) {
|
|
546
|
+
const toolCalls = testCase.toolCalls.slice(0, maxToolCalls);
|
|
600
547
|
lines.push(
|
|
601
548
|
...renderAsciiTable(
|
|
602
549
|
["Tool", "Status", "Duration"],
|
|
@@ -607,9 +554,9 @@ function renderFailureBlock(testCase, {
|
|
|
607
554
|
])
|
|
608
555
|
)
|
|
609
556
|
);
|
|
610
|
-
if (testCase.
|
|
557
|
+
if (testCase.toolCalls.length > maxToolCalls) {
|
|
611
558
|
lines.push(
|
|
612
|
-
`${testCase.
|
|
559
|
+
`${testCase.toolCalls.length - maxToolCalls} more tool calls omitted`
|
|
613
560
|
);
|
|
614
561
|
}
|
|
615
562
|
lines.push("");
|
|
@@ -652,7 +599,7 @@ function formatCaseUsage(testCase) {
|
|
|
652
599
|
const usage = testCase.harness?.usage;
|
|
653
600
|
const parts = [];
|
|
654
601
|
const totalTokens = usage?.totalTokens ?? (usage?.inputTokens ?? 0) + (usage?.outputTokens ?? 0) + (usage?.reasoningTokens ?? 0);
|
|
655
|
-
const toolCalls = usage?.toolCalls
|
|
602
|
+
const toolCalls = usage?.toolCalls !== void 0 ? Math.max(usage.toolCalls, testCase.toolCalls.length) : testCase.toolCalls.length;
|
|
656
603
|
if (totalTokens > 0) {
|
|
657
604
|
parts.push(`${formatNumber(totalTokens)} tokens`);
|
|
658
605
|
}
|
package/dist/index.mjs.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"sources":["../src/utils.ts","../src/collect.ts","../src/annotations.ts","../src/summary.ts","../src/github.ts"],"sourcesContent":["import { posix } from \"node:path\";\n\nexport function isRecord(value: unknown): value is Record<string, unknown> {\n return Boolean(value) && typeof value === \"object\" && !Array.isArray(value);\n}\n\nexport function compactLine(value: string, maxLength: number) {\n const line = value\n .split(/\\r?\\n/)\n .map((part) => part.trim())\n .find((part) => part.length > 0);\n if (!line) {\n return \"\";\n }\n return truncate(line, maxLength);\n}\n\nexport function truncate(value: string, maxLength: number) {\n if (maxLength <= 0) {\n return \"\";\n }\n if (value.length <= maxLength) {\n return value;\n }\n return `${value.slice(0, Math.max(0, maxLength - 15)).trimEnd()}... [truncated]`;\n}\n\nexport function stringifyValue(value: unknown, maxLength: number) {\n if (value === undefined) {\n return \"\";\n }\n if (typeof value === \"string\") {\n return truncate(value, maxLength);\n }\n try {\n return truncate(JSON.stringify(value, null, 2), maxLength);\n } catch {\n return truncate(String(value), maxLength);\n }\n}\n\nexport function formatNumber(value: number) {\n return new Intl.NumberFormat(\"en-US\", {\n maximumFractionDigits: 0,\n }).format(value);\n}\n\nexport function formatScore(value: number | null | undefined) {\n if (value == null || !Number.isFinite(value)) {\n return \"n/a\";\n }\n return value.toFixed(2);\n}\n\nexport function formatDuration(ms: number | undefined) {\n if (ms === undefined || !Number.isFinite(ms)) {\n return \"n/a\";\n }\n if (ms < 1000) {\n return `${Math.round(ms)}ms`;\n }\n const seconds = ms / 1000;\n if (seconds < 60) {\n return `${seconds.toFixed(seconds < 10 ? 1 : 0)}s`;\n }\n const totalSeconds = Math.round(seconds);\n const minutes = Math.floor(totalSeconds / 60);\n const remainingSeconds = totalSeconds % 60;\n return `${minutes}m ${remainingSeconds}s`;\n}\n\nexport function formatLocation(\n file: string,\n location?: { line: number; column: number },\n) {\n if (!location) {\n return file;\n }\n return `${file}:${location.line}`;\n}\n\nexport function normalizePathForGitHub(path: string, workspace?: string) {\n const normalized = path.replace(/\\\\/g, \"/\");\n if (!workspace) {\n return normalized;\n }\n\n const workspacePath = workspace.replace(/\\\\/g, \"/\").replace(/\\/+$/, \"\");\n if (\n normalized !== workspacePath &&\n !normalized.startsWith(`${workspacePath}/`)\n ) {\n return normalized;\n }\n\n return posix.relative(workspacePath, normalized);\n}\n\nexport function escapeFence(value: string) {\n return value.replace(/```/g, \"'''\");\n}\n\nexport function escapeHtml(value: string) {\n return value\n .replace(/&/g, \"&\")\n .replace(/</g, \"<\")\n .replace(/>/g, \">\")\n .replace(/\"/g, \""\");\n}\n\nexport function escapeCommandData(value: string) {\n return value.replace(/%/g, \"%25\").replace(/\\r/g, \"%0D\").replace(/\\n/g, \"%0A\");\n}\n\nexport function escapeCommandProperty(value: string) {\n return escapeCommandData(value).replace(/:/g, \"%3A\").replace(/,/g, \"%2C\");\n}\n","import type {\n CollectOptions,\n EvalCase,\n EvalFailure,\n EvalReport,\n EvalScore,\n ToolCallSummary,\n UsageSummary,\n VitestJsonAssertion,\n VitestJsonFile,\n VitestJsonReport,\n} from \"./types\";\nimport {\n compactLine,\n isRecord,\n normalizePathForGitHub,\n stringifyValue,\n} from \"./utils\";\n\ntype EvalMeta = {\n scores?: EvalScore[];\n avgScore?: number;\n output?: unknown;\n thresholdFailed?: boolean;\n};\n\ntype HarnessMeta = {\n name?: string;\n run?: {\n output?: unknown;\n usage?: UsageSummary;\n timings?: {\n totalMs?: number;\n };\n session?: {\n messages?: Array<{\n toolCalls?: unknown[];\n }>;\n };\n errors?: unknown[];\n };\n};\n\ntype HarnessRunMeta = NonNullable<HarnessMeta[\"run\"]>;\n\n/** Converts a Vitest JSON report into the compact eval report model. */\nexport function collectEvalReport(\n input: VitestJsonReport,\n options: CollectOptions = {},\n): EvalReport {\n const cases = input.testResults.flatMap((file) =>\n file.assertionResults.flatMap((assertion) => {\n const evalCase = collectEvalCase(file, assertion, options);\n return evalCase ? [evalCase] : [];\n }),\n );\n const failures = cases.filter((testCase) => testCase.status === \"failed\");\n const evalScores = cases\n .map((testCase) => testCase.eval?.avgScore)\n .filter((score): score is number => isFiniteNumber(score));\n const usage = sumUsage(cases);\n const durationMs = resolveRunDuration(input);\n\n return {\n status: input.success && failures.length === 0 ? \"passed\" : \"failed\",\n startedAt: input.startTime,\n durationMs,\n totals: {\n total: input.numTotalTests,\n passed: input.numPassedTests,\n failed: input.numFailedTests,\n skipped: input.numPendingTests + input.numTodoTests,\n evalTotal: cases.length,\n evalPassed: cases.filter((testCase) => testCase.status === \"passed\")\n .length,\n evalFailed: failures.length,\n },\n score:\n evalScores.length > 0\n ? {\n average:\n evalScores.reduce((total, score) => total + score, 0) /\n evalScores.length,\n minimum: Math.min(...evalScores),\n }\n : undefined,\n usage,\n cases,\n failures,\n };\n}\n\nfunction collectEvalCase(\n file: VitestJsonFile,\n assertion: VitestJsonAssertion,\n options: CollectOptions,\n): EvalCase | null {\n const meta = isRecord(assertion.meta) ? assertion.meta : {};\n const evalMeta = getEvalMeta(meta.eval);\n const harnessMeta = getHarnessMeta(meta.harness);\n\n if (!evalMeta && !harnessMeta) {\n return null;\n }\n\n const displayFile = normalizePathForGitHub(file.name, options.workspace);\n const scores = evalMeta?.scores ?? [];\n const harnessRun = harnessMeta?.run;\n const toolCalls = collectToolCalls(harnessRun?.session);\n const evalCase: EvalCase = {\n id: `${file.name}:${assertion.location?.line ?? 0}:${assertion.fullName}`,\n file: file.name,\n displayFile,\n title: assertion.title,\n displayName: formatDisplayName(assertion),\n status: assertion.status,\n durationMs:\n typeof assertion.duration === \"number\" ? assertion.duration : undefined,\n location: assertion.location ?? undefined,\n failureMessages: assertion.failureMessages ?? [],\n eval: evalMeta\n ? {\n avgScore: evalMeta.avgScore,\n thresholdFailed: evalMeta.thresholdFailed,\n output: evalMeta.output,\n scores,\n }\n : undefined,\n harness: harnessMeta\n ? {\n name: harnessMeta.name,\n output: harnessRun?.output,\n usage: harnessRun?.usage,\n timingMs: harnessRun?.timings?.totalMs,\n toolCalls,\n errors: harnessRun?.errors ?? [],\n }\n : undefined,\n };\n\n evalCase.primaryFailure = getPrimaryFailure(evalCase);\n return evalCase;\n}\n\nfunction getEvalMeta(value: unknown): EvalMeta | undefined {\n if (!isRecord(value)) {\n return undefined;\n }\n\n return {\n scores: Array.isArray(value.scores)\n ? value.scores.filter(isRecord).map(normalizeScore)\n : undefined,\n avgScore: numberField(value.avgScore),\n output: value.output,\n thresholdFailed:\n typeof value.thresholdFailed === \"boolean\"\n ? value.thresholdFailed\n : undefined,\n };\n}\n\nfunction normalizeScore(score: Record<string, unknown>): EvalScore {\n const metadata = isRecord(score.metadata) ? score.metadata : undefined;\n return {\n name: typeof score.name === \"string\" ? score.name : undefined,\n score: numberField(score.score) ?? null,\n metadata,\n };\n}\n\nfunction getHarnessMeta(value: unknown): HarnessMeta | undefined {\n if (!isRecord(value)) {\n return undefined;\n }\n const run = isRecord(value.run) ? value.run : undefined;\n const usage = isRecord(run?.usage) ? getUsage(run.usage) : undefined;\n const timings = isRecord(run?.timings) ? run.timings : undefined;\n\n return {\n name: typeof value.name === \"string\" ? value.name : undefined,\n run: run\n ? {\n output: run.output,\n usage,\n timings: {\n totalMs:\n typeof timings?.totalMs === \"number\"\n ? timings.totalMs\n : undefined,\n },\n session: isRecord(run.session)\n ? {\n messages: Array.isArray(run.session.messages)\n ? (run.session.messages as Array<{ toolCalls?: unknown[] }>)\n : undefined,\n }\n : undefined,\n errors: Array.isArray(run.errors) ? run.errors : undefined,\n }\n : undefined,\n };\n}\n\nfunction getUsage(value: Record<string, unknown>): UsageSummary {\n return {\n inputTokens: numberField(value.inputTokens),\n outputTokens: numberField(value.outputTokens),\n reasoningTokens: numberField(value.reasoningTokens),\n totalTokens: numberField(value.totalTokens),\n toolCalls: numberField(value.toolCalls),\n };\n}\n\nfunction numberField(value: unknown) {\n return isFiniteNumber(value) ? value : undefined;\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n return typeof value === \"number\" && Number.isFinite(value);\n}\n\nfunction collectToolCalls(session: HarnessRunMeta[\"session\"] | undefined) {\n const messages = session?.messages ?? [];\n const toolCalls: ToolCallSummary[] = [];\n\n for (const message of messages) {\n if (!Array.isArray(message.toolCalls)) {\n continue;\n }\n\n for (const call of message.toolCalls) {\n if (!isRecord(call) || typeof call.name !== \"string\") {\n continue;\n }\n\n toolCalls.push({\n name: call.name,\n error: getToolCallError(call.error),\n durationMs: numberField(call.durationMs),\n });\n }\n }\n\n return toolCalls;\n}\n\nfunction getToolCallError(value: unknown) {\n if (isRecord(value) && typeof value.message === \"string\") {\n return value.message;\n }\n if (value !== undefined) {\n return stringifyValue(value, 240);\n }\n return undefined;\n}\n\nfunction getPrimaryFailure(testCase: EvalCase): EvalFailure | undefined {\n const failingScores = [...(testCase.eval?.scores ?? [])]\n .filter(\n (score) =>\n (score.score ?? 0) < 1 ||\n score.metadata?.rationale !== undefined ||\n score.metadata?.output !== undefined,\n )\n .sort((left, right) => (left.score ?? 0) - (right.score ?? 0));\n const primary = failingScores[0];\n const score =\n typeof primary?.score === \"number\"\n ? primary.score\n : testCase.eval?.avgScore;\n const reason =\n stringifyReason(primary?.metadata?.rationale) ??\n compactLine(testCase.failureMessages.join(\"\\n\"), 500);\n\n if (!primary && !reason && score === undefined) {\n return undefined;\n }\n\n return {\n judgeName: primary?.name,\n score,\n reason: reason || undefined,\n };\n}\n\nfunction stringifyReason(value: unknown) {\n if (value === undefined) {\n return undefined;\n }\n return typeof value === \"string\" ? value : stringifyValue(value, 4000);\n}\n\nfunction formatDisplayName(assertion: VitestJsonAssertion) {\n return [...assertion.ancestorTitles, assertion.title]\n .filter((part) => part.length > 0)\n .join(\" > \");\n}\n\nfunction sumUsage(cases: EvalCase[]) {\n const usage: Required<UsageSummary> = {\n inputTokens: 0,\n outputTokens: 0,\n reasoningTokens: 0,\n totalTokens: 0,\n toolCalls: 0,\n };\n\n for (const testCase of cases) {\n const caseUsage = testCase.harness?.usage;\n usage.inputTokens += caseUsage?.inputTokens ?? 0;\n usage.outputTokens += caseUsage?.outputTokens ?? 0;\n usage.reasoningTokens += caseUsage?.reasoningTokens ?? 0;\n usage.totalTokens +=\n caseUsage?.totalTokens ??\n (caseUsage?.inputTokens ?? 0) +\n (caseUsage?.outputTokens ?? 0) +\n (caseUsage?.reasoningTokens ?? 0);\n usage.toolCalls +=\n caseUsage?.toolCalls ?? testCase.harness?.toolCalls.length ?? 0;\n }\n\n return usage;\n}\n\nfunction resolveRunDuration(input: VitestJsonReport) {\n const startTimes = input.testResults\n .map((file) => file.startTime)\n .filter((time) => Number.isFinite(time));\n const endTimes = input.testResults\n .map((file) => file.endTime)\n .filter((time) => Number.isFinite(time));\n if (startTimes.length === 0 || endTimes.length === 0) {\n return undefined;\n }\n return Math.max(...endTimes) - Math.min(...startTimes);\n}\n","import type { EvalCase, EvalReport } from \"./types\";\nimport {\n compactLine,\n escapeCommandData,\n escapeCommandProperty,\n formatScore,\n stringifyValue,\n truncate,\n} from \"./utils\";\n\n/** Options for limiting rendered GitHub annotations. */\nexport type AnnotationOptions = {\n maxAnnotations?: number;\n};\n\n/** GitHub Check Run annotation payload. */\nexport type CheckAnnotation = {\n path: string;\n start_line: number;\n end_line: number;\n annotation_level: \"failure\" | \"warning\" | \"notice\";\n message: string;\n title?: string;\n raw_details?: string;\n};\n\nconst DEFAULT_MAX_WORKFLOW_ANNOTATIONS = 10;\nconst DEFAULT_MAX_CHECK_ANNOTATIONS = 50;\nconst MAX_CHECK_FIELD_LENGTH = 64_000;\n\ntype AnnotatedEvalCase = EvalCase & {\n location: { line: number; column: number };\n};\n\n/** Renders GitHub workflow-command annotations for failed eval cases. */\nexport function renderWorkflowCommands(\n report: EvalReport,\n options: AnnotationOptions = {},\n) {\n const maxAnnotations =\n options.maxAnnotations ?? DEFAULT_MAX_WORKFLOW_ANNOTATIONS;\n\n return report.failures\n .filter(hasAnnotationLocation)\n .slice(0, maxAnnotations)\n .map((testCase) =>\n formatWorkflowCommand({\n command: \"error\",\n properties: {\n file: testCase.displayFile,\n line: String(testCase.location.line),\n col: String(testCase.location.column),\n title: \"vitest-evals\",\n },\n message: formatWorkflowMessage(testCase),\n }),\n );\n}\n\n/** Builds Check Run annotations for failed eval cases. */\nexport function buildCheckAnnotations(\n report: EvalReport,\n options: AnnotationOptions = {},\n): CheckAnnotation[] {\n const maxAnnotations = Math.min(\n options.maxAnnotations ?? DEFAULT_MAX_CHECK_ANNOTATIONS,\n DEFAULT_MAX_CHECK_ANNOTATIONS,\n );\n\n return report.failures\n .filter(hasAnnotationLocation)\n .slice(0, maxAnnotations)\n .map((testCase) => ({\n path: testCase.displayFile,\n start_line: testCase.location.line,\n end_line: testCase.location.line,\n annotation_level: \"failure\",\n title: truncate(\n `${testCase.primaryFailure?.judgeName ?? \"vitest-evals\"} - ${testCase.displayName}`,\n 255,\n ),\n message: truncate(\n formatWorkflowMessage(testCase),\n MAX_CHECK_FIELD_LENGTH,\n ),\n raw_details: truncate(formatRawDetails(testCase), MAX_CHECK_FIELD_LENGTH),\n }));\n}\n\nfunction hasAnnotationLocation(\n testCase: EvalCase,\n): testCase is AnnotatedEvalCase {\n return Boolean(testCase.location);\n}\n\nfunction formatWorkflowMessage(testCase: EvalCase) {\n const failure = testCase.primaryFailure;\n const parts = [\n testCase.displayName,\n `score ${formatScore(failure?.score ?? testCase.eval?.avgScore)}`,\n ];\n\n if (failure?.judgeName) {\n parts.push(failure.judgeName);\n }\n\n const reason = compactLine(failure?.reason ?? \"\", 320);\n if (reason) {\n parts.push(reason);\n }\n\n return parts.join(\" - \");\n}\n\nfunction formatRawDetails(testCase: AnnotatedEvalCase) {\n const lines = [\n `Test: ${testCase.displayName}`,\n `Location: ${testCase.displayFile}:${testCase.location.line}`,\n `Harness: ${testCase.harness?.name ?? \"n/a\"}`,\n `Score: ${formatScore(testCase.primaryFailure?.score ?? testCase.eval?.avgScore)}`,\n `Judge: ${testCase.primaryFailure?.judgeName ?? \"n/a\"}`,\n \"\",\n \"Reason:\",\n testCase.primaryFailure?.reason ?? \"n/a\",\n ];\n\n const finalOutput = testCase.eval?.output ?? testCase.harness?.output;\n if (finalOutput !== undefined) {\n lines.push(\"\", \"Final:\", stringifyValue(finalOutput, 8000));\n }\n\n if (testCase.harness?.toolCalls.length) {\n lines.push(\"\", \"Tools:\");\n for (const toolCall of testCase.harness.toolCalls) {\n lines.push(\n `- ${toolCall.name}: ${toolCall.error ? `error: ${toolCall.error}` : \"ok\"}`,\n );\n }\n }\n\n return lines.join(\"\\n\");\n}\n\nfunction formatWorkflowCommand({\n command,\n properties,\n message,\n}: {\n command: \"error\" | \"warning\" | \"notice\";\n properties: Record<string, string>;\n message: string;\n}) {\n const renderedProperties = Object.entries(properties)\n .map(([key, value]) => `${key}=${escapeCommandProperty(value)}`)\n .join(\",\");\n return `::${command} ${renderedProperties}::${escapeCommandData(message)}`;\n}\n","import type { EvalCase, EvalReport } from \"./types\";\nimport {\n compactLine,\n escapeFence,\n escapeHtml,\n formatDuration,\n formatLocation,\n formatNumber,\n formatScore,\n stringifyValue,\n truncate,\n} from \"./utils\";\n\n/** Options for rendering GitHub Actions summary markdown. */\nexport type SummaryOptions = {\n maxFailures?: number;\n maxReasonChars?: number;\n maxOutputChars?: number;\n maxToolCalls?: number;\n};\n\nconst DEFAULT_MAX_FAILURES = 20;\nconst DEFAULT_MAX_REASON_CHARS = 8000;\nconst DEFAULT_MAX_OUTPUT_CHARS = 4000;\nconst DEFAULT_MAX_TOOL_CALLS = 20;\nconst SCORE_DISTRIBUTION_BUCKETS = [\n \"0-19%\",\n \"20-39%\",\n \"40-59%\",\n \"60-79%\",\n \"80-100%\",\n] as const;\nconst SCORE_DISTRIBUTION_BAR_WIDTH = 20;\n\n/** Renders the GitHub Actions job summary markdown for an eval report. */\nexport function renderJobSummary(\n report: EvalReport,\n options: SummaryOptions = {},\n) {\n const maxFailures = options.maxFailures ?? DEFAULT_MAX_FAILURES;\n const failures = report.failures.slice(0, maxFailures);\n const nonEvalFailures = report.totals.failed - report.totals.evalFailed;\n const lines: string[] = [\n \"# vitest-evals\",\n \"\",\n ...renderSummaryTable(report, nonEvalFailures),\n \"\",\n ...renderScoreDistribution(report),\n \"## Results\",\n \"\",\n ];\n\n if (report.failures.length > 0) {\n lines.push(\"### Failures\", \"\");\n failures.forEach((testCase, index) => {\n lines.push(...renderFailureDetails(testCase, index + 1, options), \"\");\n });\n\n if (report.failures.length > failures.length) {\n lines.push(\n `${report.failures.length - failures.length} more failures omitted from this summary.`,\n \"\",\n );\n }\n } else if (report.totals.evalTotal > 0) {\n lines.push(\"### Failures\", \"\", \"No eval failures.\", \"\");\n }\n\n if (report.totals.evalTotal === 0) {\n lines.push(\"No eval metadata was found in the Vitest JSON report.\", \"\");\n }\n\n return `${lines.join(\"\\n\")}\\n`;\n}\n\nfunction formatCountLine(passed: number, failed: number, total: number) {\n return `${formatNumber(passed)} passed, ${formatNumber(failed)} failed, ${formatNumber(total)} total`;\n}\n\nfunction renderSummaryTable(report: EvalReport, nonEvalFailures: number) {\n const rows: Array<[string, string]> = [\n [\"Status\", report.status],\n [\n \"Evals\",\n formatCountLine(\n report.totals.evalPassed,\n report.totals.evalFailed,\n report.totals.evalTotal,\n ),\n ],\n ];\n\n if (report.score) {\n rows.push([\"Score\", formatScoreSummary(report.score)]);\n }\n\n if (nonEvalFailures > 0) {\n rows.push([\n \"Other Failures\",\n `${formatNumber(nonEvalFailures)} non-eval test failure${\n nonEvalFailures === 1 ? \"\" : \"s\"\n }`,\n ]);\n }\n\n rows.push([\"Duration\", formatDuration(report.durationMs)]);\n\n return [\n \"| Metric | Value |\",\n \"| --- | --- |\",\n ...rows.map(\n ([metric, value]) =>\n `| ${escapeTableCell(metric)} | ${escapeTableCell(value)} |`,\n ),\n ];\n}\n\nfunction formatScoreSummary(score: NonNullable<EvalReport[\"score\"]>) {\n return `avg ${formatScore(score.average)}${\n score.minimum === undefined ? \"\" : `, min ${formatScore(score.minimum)}`\n }`;\n}\n\nfunction escapeTableCell(value: string) {\n return value\n .replace(/\\r?\\n/g, \" \")\n .replace(/\\\\/g, \"\\\\\\\\\")\n .replace(/\\|/g, \"\\\\|\");\n}\n\nfunction renderScoreDistribution(report: EvalReport) {\n const scores = report.cases\n .map((testCase) => testCase.eval?.avgScore)\n .filter(\n (score): score is number =>\n typeof score === \"number\" && Number.isFinite(score),\n );\n if (scores.length === 0) {\n return [];\n }\n\n const counts = SCORE_DISTRIBUTION_BUCKETS.map(() => 0);\n for (const score of scores) {\n const bucket = Math.min(\n SCORE_DISTRIBUTION_BUCKETS.length - 1,\n Math.max(0, Math.floor(score * SCORE_DISTRIBUTION_BUCKETS.length)),\n );\n counts[bucket] = (counts[bucket] ?? 0) + 1;\n }\n\n const maxCount = Math.max(...counts);\n return [\n \"## Scores\",\n \"\",\n \"```text\",\n ...SCORE_DISTRIBUTION_BUCKETS.map((label, index) =>\n formatScoreDistributionBucket(label, counts[index] ?? 0, maxCount),\n ),\n \"```\",\n \"\",\n ];\n}\n\nfunction formatScoreDistributionBucket(\n label: string,\n count: number,\n maxCount: number,\n) {\n const barLength =\n count === 0\n ? 0\n : Math.max(\n 1,\n Math.round((count / maxCount) * SCORE_DISTRIBUTION_BAR_WIDTH),\n );\n const bar = \"#\".repeat(barLength).padEnd(SCORE_DISTRIBUTION_BAR_WIDTH, \" \");\n return `${label.padEnd(7)} | ${bar} ${formatNumber(count)}`;\n}\n\nfunction renderFailureDetails(\n testCase: EvalCase,\n number: number,\n options: SummaryOptions,\n) {\n const failure = testCase.primaryFailure;\n const maxReasonChars = options.maxReasonChars ?? DEFAULT_MAX_REASON_CHARS;\n const maxOutputChars = options.maxOutputChars ?? DEFAULT_MAX_OUTPUT_CHARS;\n const maxToolCalls = options.maxToolCalls ?? DEFAULT_MAX_TOOL_CALLS;\n const usage = formatCaseUsage(testCase);\n const finalOutput = testCase.eval?.output ?? testCase.harness?.output;\n const summary = `${number}. ${testCase.displayName} - ${\n failure?.judgeName ?? \"failure\"\n } - ${formatScore(failure?.score ?? testCase.eval?.avgScore)}`;\n const lines = [\n \"<details>\",\n `<summary>${escapeHtml(summary)}</summary>`,\n \"\",\n \"```text\",\n ...renderFailureBlock(testCase, {\n finalOutput,\n maxOutputChars,\n maxReasonChars,\n maxToolCalls,\n number,\n usage,\n }).map(escapeFence),\n \"```\",\n \"\",\n \"</details>\",\n ];\n return lines;\n}\n\nfunction renderFailureBlock(\n testCase: EvalCase,\n {\n finalOutput,\n maxOutputChars,\n maxReasonChars,\n maxToolCalls,\n number,\n usage,\n }: {\n finalOutput: unknown;\n maxOutputChars: number;\n maxReasonChars: number;\n maxToolCalls: number;\n number: number;\n usage: string;\n },\n) {\n const failure = testCase.primaryFailure;\n const overviewRows: Array<[string, string]> = [\n [\"Case\", `${number}. ${testCase.displayName}`],\n [\"Status\", testCase.status],\n [\"Location\", formatLocation(testCase.displayFile, testCase.location)],\n [\"Harness\", testCase.harness?.name ?? \"n/a\"],\n [\"Score\", formatScore(failure?.score ?? testCase.eval?.avgScore)],\n [\"Judge\", failure?.judgeName ?? \"n/a\"],\n ];\n if (usage) {\n overviewRows.push([\"Usage\", usage]);\n }\n if (testCase.durationMs !== undefined) {\n overviewRows.push([\"Duration\", formatDuration(testCase.durationMs)]);\n }\n\n const lines = [\n ...renderAsciiSection(\"Result\", renderKeyValues(overviewRows)),\n \"\",\n ];\n\n if (failure?.reason) {\n lines.push(\n ...renderAsciiSection(\n \"Reason\",\n truncate(failure.reason, maxReasonChars).split(/\\r?\\n/),\n ),\n \"\",\n );\n }\n\n if (testCase.eval?.scores.length) {\n lines.push(\n ...renderAsciiTable(\n [\"Judge\", \"Score\"],\n testCase.eval.scores.map((score) => [\n score.name ?? \"Unknown\",\n formatScore(score.score),\n ]),\n ),\n \"\",\n );\n }\n\n if (finalOutput !== undefined) {\n lines.push(\n ...renderAsciiSection(\n \"Final Output\",\n stringifyValue(finalOutput, maxOutputChars).split(/\\r?\\n/),\n ),\n \"\",\n );\n }\n\n if (testCase.harness?.toolCalls.length) {\n const toolCalls = testCase.harness.toolCalls.slice(0, maxToolCalls);\n lines.push(\n ...renderAsciiTable(\n [\"Tool\", \"Status\", \"Duration\"],\n toolCalls.map((toolCall) => [\n toolCall.name,\n toolCall.error ? `error: ${compactLine(toolCall.error, 120)}` : \"ok\",\n toolCall.durationMs === undefined\n ? \"n/a\"\n : formatDuration(toolCall.durationMs),\n ]),\n ),\n );\n if (testCase.harness.toolCalls.length > maxToolCalls) {\n lines.push(\n `${testCase.harness.toolCalls.length - maxToolCalls} more tool calls omitted`,\n );\n }\n lines.push(\"\");\n }\n\n if (testCase.harness?.errors.length) {\n lines.push(\n ...renderAsciiSection(\n \"Harness Errors\",\n stringifyValue(testCase.harness.errors, maxReasonChars).split(/\\r?\\n/),\n ),\n \"\",\n );\n }\n\n while (lines[lines.length - 1] === \"\") {\n lines.pop();\n }\n return lines;\n}\n\nfunction renderAsciiSection(title: string, content: string[]) {\n return [title, \"-\".repeat(title.length), ...content];\n}\n\nfunction renderKeyValues(rows: Array<[string, string]>) {\n const labelWidth = Math.max(...rows.map(([label]) => label.length));\n return rows.map(\n ([label, value]) =>\n `${label.padEnd(labelWidth)} ${compactLine(value, 500)}`,\n );\n}\n\nfunction renderAsciiTable(headers: string[], rows: string[][]) {\n const widths = headers.map((header, index) =>\n Math.max(header.length, ...rows.map((row) => row[index]?.length ?? 0)),\n );\n const renderRow = (row: string[]) =>\n row\n .map((cell, index) => cell.padEnd(widths[index] ?? cell.length))\n .join(\" \")\n .trimEnd();\n\n return [\n renderRow(headers),\n widths.map((width) => \"-\".repeat(width)).join(\" \"),\n ...rows.map(renderRow),\n ];\n}\n\nfunction formatCaseUsage(testCase: EvalCase) {\n const usage = testCase.harness?.usage;\n const parts: string[] = [];\n const totalTokens =\n usage?.totalTokens ??\n (usage?.inputTokens ?? 0) +\n (usage?.outputTokens ?? 0) +\n (usage?.reasoningTokens ?? 0);\n const toolCalls = usage?.toolCalls ?? testCase.harness?.toolCalls.length ?? 0;\n\n if (totalTokens > 0) {\n parts.push(`${formatNumber(totalTokens)} tokens`);\n }\n if (toolCalls > 0) {\n parts.push(`${formatNumber(toolCalls)} tool${toolCalls === 1 ? \"\" : \"s\"}`);\n }\n if (testCase.harness?.timingMs !== undefined) {\n parts.push(formatDuration(testCase.harness.timingMs));\n }\n\n return parts.join(\", \");\n}\n","import { buildCheckAnnotations } from \"./annotations\";\nimport { renderJobSummary, type SummaryOptions } from \"./summary\";\nimport type { EvalReport } from \"./types\";\n\n/** Options for creating or updating a GitHub Check Run. */\nexport type PublishCheckRunOptions = SummaryOptions & {\n token?: string;\n repository?: string;\n sha?: string;\n name?: string;\n apiUrl?: string;\n checkRunId?: number;\n maxAnnotations?: number;\n};\n\n/** Result of attempting to publish a GitHub Check Run. */\nexport type PublishCheckRunResult =\n | {\n status: \"skipped\";\n reason: string;\n }\n | {\n status: \"created\" | \"updated\";\n id?: number;\n htmlUrl?: string;\n };\n\nconst DEFAULT_CHECK_NAME = \"vitest-evals\";\nconst MAX_CHECK_SUMMARY_LENGTH = 64_000;\nconst CHECK_SUMMARY_TRUNCATION_SUFFIX =\n \"\\n\\n[truncated for GitHub Check Run]\\n\";\n\n/** Publishes the eval report to a GitHub Check Run when configuration allows it. */\nexport async function publishCheckRun(\n report: EvalReport,\n options: PublishCheckRunOptions = {},\n): Promise<PublishCheckRunResult> {\n const token = options.token ?? process.env.GITHUB_TOKEN;\n const repository = options.repository ?? process.env.GITHUB_REPOSITORY;\n const sha = options.sha ?? process.env.GITHUB_SHA;\n\n if (!token) {\n return { status: \"skipped\", reason: \"missing GITHUB_TOKEN\" };\n }\n if (!repository) {\n return { status: \"skipped\", reason: \"missing GITHUB_REPOSITORY\" };\n }\n if (!sha && options.checkRunId === undefined) {\n return { status: \"skipped\", reason: \"missing GITHUB_SHA\" };\n }\n\n const [owner, repo] = repository.split(\"/\");\n if (!owner || !repo) {\n return {\n status: \"skipped\",\n reason: `invalid GitHub repository: ${repository}`,\n };\n }\n\n const payload = buildCheckRunPayload(report, options);\n const apiUrl =\n options.apiUrl ?? process.env.GITHUB_API_URL ?? \"https://api.github.com\";\n const requestUrl =\n options.checkRunId === undefined\n ? `${apiUrl}/repos/${owner}/${repo}/check-runs`\n : `${apiUrl}/repos/${owner}/${repo}/check-runs/${options.checkRunId}`;\n const response = await fetch(requestUrl, {\n method: options.checkRunId === undefined ? \"POST\" : \"PATCH\",\n headers: {\n accept: \"application/vnd.github+json\",\n authorization: `Bearer ${token}`,\n \"content-type\": \"application/json\",\n \"x-github-api-version\": \"2022-11-28\",\n },\n body: JSON.stringify(\n options.checkRunId === undefined\n ? {\n name: options.name ?? DEFAULT_CHECK_NAME,\n head_sha: sha,\n ...payload,\n }\n : payload,\n ),\n });\n\n if (!response.ok) {\n const text = await response.text();\n throw new Error(\n `GitHub Check Run request failed: ${response.status} ${response.statusText} ${text}`.trim(),\n );\n }\n\n const data = (await response.json()) as {\n id?: number;\n html_url?: string;\n };\n\n return {\n status: options.checkRunId === undefined ? \"created\" : \"updated\",\n id: data.id,\n htmlUrl: data.html_url,\n };\n}\n\nfunction buildCheckRunPayload(\n report: EvalReport,\n options: PublishCheckRunOptions,\n) {\n const annotations = buildCheckAnnotations(report, {\n maxAnnotations: options.maxAnnotations,\n });\n const title =\n report.failures.length === 0 && report.status === \"passed\"\n ? \"No eval failures\"\n : report.failures.length === 0\n ? \"Vitest run failed\"\n : `${report.failures.length} eval failure${\n report.failures.length === 1 ? \"\" : \"s\"\n }`;\n\n return {\n status: \"completed\",\n conclusion: report.status === \"passed\" ? \"success\" : \"failure\",\n completed_at: new Date().toISOString(),\n output: {\n title,\n summary: truncateCheckSummary(\n renderJobSummary(report, {\n ...options,\n maxFailures: options.maxFailures ?? 5,\n maxReasonChars: options.maxReasonChars ?? 4000,\n maxOutputChars: options.maxOutputChars ?? 2000,\n maxToolCalls: options.maxToolCalls ?? 10,\n }),\n ),\n annotations,\n },\n };\n}\n\nfunction truncateCheckSummary(summary: string) {\n if (summary.length <= MAX_CHECK_SUMMARY_LENGTH) {\n return summary;\n }\n\n return `${summary\n .slice(0, MAX_CHECK_SUMMARY_LENGTH - CHECK_SUMMARY_TRUNCATION_SUFFIX.length)\n .trimEnd()}${CHECK_SUMMARY_TRUNCATION_SUFFIX}`;\n}\n"],"mappings":";AAAA,SAAS,aAAa;AAEf,SAAS,SAAS,OAAkD;AACzE,SAAO,QAAQ,KAAK,KAAK,OAAO,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAEO,SAAS,YAAY,OAAe,WAAmB;AAC5D,QAAM,OAAO,MACV,MAAM,OAAO,EACb,IAAI,CAAC,SAAS,KAAK,KAAK,CAAC,EACzB,KAAK,CAAC,SAAS,KAAK,SAAS,CAAC;AACjC,MAAI,CAAC,MAAM;AACT,WAAO;AAAA,EACT;AACA,SAAO,SAAS,MAAM,SAAS;AACjC;AAEO,SAAS,SAAS,OAAe,WAAmB;AACzD,MAAI,aAAa,GAAG;AAClB,WAAO;AAAA,EACT;AACA,MAAI,MAAM,UAAU,WAAW;AAC7B,WAAO;AAAA,EACT;AACA,SAAO,GAAG,MAAM,MAAM,GAAG,KAAK,IAAI,GAAG,YAAY,EAAE,CAAC,EAAE,QAAQ,CAAC;AACjE;AAEO,SAAS,eAAe,OAAgB,WAAmB;AAChE,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AACA,MAAI,OAAO,UAAU,UAAU;AAC7B,WAAO,SAAS,OAAO,SAAS;AAAA,EAClC;AACA,MAAI;AACF,WAAO,SAAS,KAAK,UAAU,OAAO,MAAM,CAAC,GAAG,SAAS;AAAA,EAC3D,QAAQ;AACN,WAAO,SAAS,OAAO,KAAK,GAAG,SAAS;AAAA,EAC1C;AACF;AAEO,SAAS,aAAa,OAAe;AAC1C,SAAO,IAAI,KAAK,aAAa,SAAS;AAAA,IACpC,uBAAuB;AAAA,EACzB,CAAC,EAAE,OAAO,KAAK;AACjB;AAEO,SAAS,YAAY,OAAkC;AAC5D,MAAI,SAAS,QAAQ,CAAC,OAAO,SAAS,KAAK,GAAG;AAC5C,WAAO;AAAA,EACT;AACA,SAAO,MAAM,QAAQ,CAAC;AACxB;AAEO,SAAS,eAAe,IAAwB;AACrD,MAAI,OAAO,UAAa,CAAC,OAAO,SAAS,EAAE,GAAG;AAC5C,WAAO;AAAA,EACT;AACA,MAAI,KAAK,KAAM;AACb,WAAO,GAAG,KAAK,MAAM,EAAE,CAAC;AAAA,EAC1B;AACA,QAAM,UAAU,KAAK;AACrB,MAAI,UAAU,IAAI;AAChB,WAAO,GAAG,QAAQ,QAAQ,UAAU,KAAK,IAAI,CAAC,CAAC;AAAA,EACjD;AACA,QAAM,eAAe,KAAK,MAAM,OAAO;AACvC,QAAM,UAAU,KAAK,MAAM,eAAe,EAAE;AAC5C,QAAM,mBAAmB,eAAe;AACxC,SAAO,GAAG,OAAO,KAAK,gBAAgB;AACxC;AAEO,SAAS,eACd,MACA,UACA;AACA,MAAI,CAAC,UAAU;AACb,WAAO;AAAA,EACT;AACA,SAAO,GAAG,IAAI,IAAI,SAAS,IAAI;AACjC;AAEO,SAAS,uBAAuB,MAAc,WAAoB;AACvE,QAAM,aAAa,KAAK,QAAQ,OAAO,GAAG;AAC1C,MAAI,CAAC,WAAW;AACd,WAAO;AAAA,EACT;AAEA,QAAM,gBAAgB,UAAU,QAAQ,OAAO,GAAG,EAAE,QAAQ,QAAQ,EAAE;AACtE,MACE,eAAe,iBACf,CAAC,WAAW,WAAW,GAAG,aAAa,GAAG,GAC1C;AACA,WAAO;AAAA,EACT;AAEA,SAAO,MAAM,SAAS,eAAe,UAAU;AACjD;AAEO,SAAS,YAAY,OAAe;AACzC,SAAO,MAAM,QAAQ,QAAQ,KAAK;AACpC;AAEO,SAAS,WAAW,OAAe;AACxC,SAAO,MACJ,QAAQ,MAAM,OAAO,EACrB,QAAQ,MAAM,MAAM,EACpB,QAAQ,MAAM,MAAM,EACpB,QAAQ,MAAM,QAAQ;AAC3B;AAEO,SAAS,kBAAkB,OAAe;AAC/C,SAAO,MAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,OAAO,KAAK,EAAE,QAAQ,OAAO,KAAK;AAC9E;AAEO,SAAS,sBAAsB,OAAe;AACnD,SAAO,kBAAkB,KAAK,EAAE,QAAQ,MAAM,KAAK,EAAE,QAAQ,MAAM,KAAK;AAC1E;;;ACtEO,SAAS,kBACd,OACA,UAA0B,CAAC,GACf;AACZ,QAAM,QAAQ,MAAM,YAAY;AAAA,IAAQ,CAAC,SACvC,KAAK,iBAAiB,QAAQ,CAAC,cAAc;AAC3C,YAAM,WAAW,gBAAgB,MAAM,WAAW,OAAO;AACzD,aAAO,WAAW,CAAC,QAAQ,IAAI,CAAC;AAAA,IAClC,CAAC;AAAA,EACH;AACA,QAAM,WAAW,MAAM,OAAO,CAAC,aAAa,SAAS,WAAW,QAAQ;AACxE,QAAM,aAAa,MAChB,IAAI,CAAC,aAAa,SAAS,MAAM,QAAQ,EACzC,OAAO,CAAC,UAA2B,eAAe,KAAK,CAAC;AAC3D,QAAM,QAAQ,SAAS,KAAK;AAC5B,QAAM,aAAa,mBAAmB,KAAK;AAE3C,SAAO;AAAA,IACL,QAAQ,MAAM,WAAW,SAAS,WAAW,IAAI,WAAW;AAAA,IAC5D,WAAW,MAAM;AAAA,IACjB;AAAA,IACA,QAAQ;AAAA,MACN,OAAO,MAAM;AAAA,MACb,QAAQ,MAAM;AAAA,MACd,QAAQ,MAAM;AAAA,MACd,SAAS,MAAM,kBAAkB,MAAM;AAAA,MACvC,WAAW,MAAM;AAAA,MACjB,YAAY,MAAM,OAAO,CAAC,aAAa,SAAS,WAAW,QAAQ,EAChE;AAAA,MACH,YAAY,SAAS;AAAA,IACvB;AAAA,IACA,OACE,WAAW,SAAS,IAChB;AAAA,MACE,SACE,WAAW,OAAO,CAAC,OAAO,UAAU,QAAQ,OAAO,CAAC,IACpD,WAAW;AAAA,MACb,SAAS,KAAK,IAAI,GAAG,UAAU;AAAA,IACjC,IACA;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,gBACP,MACA,WACA,SACiB;AACjB,QAAM,OAAO,SAAS,UAAU,IAAI,IAAI,UAAU,OAAO,CAAC;AAC1D,QAAM,WAAW,YAAY,KAAK,IAAI;AACtC,QAAM,cAAc,eAAe,KAAK,OAAO;AAE/C,MAAI,CAAC,YAAY,CAAC,aAAa;AAC7B,WAAO;AAAA,EACT;AAEA,QAAM,cAAc,uBAAuB,KAAK,MAAM,QAAQ,SAAS;AACvE,QAAM,SAAS,UAAU,UAAU,CAAC;AACpC,QAAM,aAAa,aAAa;AAChC,QAAM,YAAY,iBAAiB,YAAY,OAAO;AACtD,QAAM,WAAqB;AAAA,IACzB,IAAI,GAAG,KAAK,IAAI,IAAI,UAAU,UAAU,QAAQ,CAAC,IAAI,UAAU,QAAQ;AAAA,IACvE,MAAM,KAAK;AAAA,IACX;AAAA,IACA,OAAO,UAAU;AAAA,IACjB,aAAa,kBAAkB,SAAS;AAAA,IACxC,QAAQ,UAAU;AAAA,IAClB,YACE,OAAO,UAAU,aAAa,WAAW,UAAU,WAAW;AAAA,IAChE,UAAU,UAAU,YAAY;AAAA,IAChC,iBAAiB,UAAU,mBAAmB,CAAC;AAAA,IAC/C,MAAM,WACF;AAAA,MACE,UAAU,SAAS;AAAA,MACnB,iBAAiB,SAAS;AAAA,MAC1B,QAAQ,SAAS;AAAA,MACjB;AAAA,IACF,IACA;AAAA,IACJ,SAAS,cACL;AAAA,MACE,MAAM,YAAY;AAAA,MAClB,QAAQ,YAAY;AAAA,MACpB,OAAO,YAAY;AAAA,MACnB,UAAU,YAAY,SAAS;AAAA,MAC/B;AAAA,MACA,QAAQ,YAAY,UAAU,CAAC;AAAA,IACjC,IACA;AAAA,EACN;AAEA,WAAS,iBAAiB,kBAAkB,QAAQ;AACpD,SAAO;AACT;AAEA,SAAS,YAAY,OAAsC;AACzD,MAAI,CAAC,SAAS,KAAK,GAAG;AACpB,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,QAAQ,MAAM,QAAQ,MAAM,MAAM,IAC9B,MAAM,OAAO,OAAO,QAAQ,EAAE,IAAI,cAAc,IAChD;AAAA,IACJ,UAAU,YAAY,MAAM,QAAQ;AAAA,IACpC,QAAQ,MAAM;AAAA,IACd,iBACE,OAAO,MAAM,oBAAoB,YAC7B,MAAM,kBACN;AAAA,EACR;AACF;AAEA,SAAS,eAAe,OAA2C;AACjE,QAAM,WAAW,SAAS,MAAM,QAAQ,IAAI,MAAM,WAAW;AAC7D,SAAO;AAAA,IACL,MAAM,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO;AAAA,IACpD,OAAO,YAAY,MAAM,KAAK,KAAK;AAAA,IACnC;AAAA,EACF;AACF;AAEA,SAAS,eAAe,OAAyC;AAC/D,MAAI,CAAC,SAAS,KAAK,GAAG;AACpB,WAAO;AAAA,EACT;AACA,QAAM,MAAM,SAAS,MAAM,GAAG,IAAI,MAAM,MAAM;AAC9C,QAAM,QAAQ,SAAS,KAAK,KAAK,IAAI,SAAS,IAAI,KAAK,IAAI;AAC3D,QAAM,UAAU,SAAS,KAAK,OAAO,IAAI,IAAI,UAAU;AAEvD,SAAO;AAAA,IACL,MAAM,OAAO,MAAM,SAAS,WAAW,MAAM,OAAO;AAAA,IACpD,KAAK,MACD;AAAA,MACE,QAAQ,IAAI;AAAA,MACZ;AAAA,MACA,SAAS;AAAA,QACP,SACE,OAAO,SAAS,YAAY,WACxB,QAAQ,UACR;AAAA,MACR;AAAA,MACA,SAAS,SAAS,IAAI,OAAO,IACzB;AAAA,QACE,UAAU,MAAM,QAAQ,IAAI,QAAQ,QAAQ,IACvC,IAAI,QAAQ,WACb;AAAA,MACN,IACA;AAAA,MACJ,QAAQ,MAAM,QAAQ,IAAI,MAAM,IAAI,IAAI,SAAS;AAAA,IACnD,IACA;AAAA,EACN;AACF;AAEA,SAAS,SAAS,OAA8C;AAC9D,SAAO;AAAA,IACL,aAAa,YAAY,MAAM,WAAW;AAAA,IAC1C,cAAc,YAAY,MAAM,YAAY;AAAA,IAC5C,iBAAiB,YAAY,MAAM,eAAe;AAAA,IAClD,aAAa,YAAY,MAAM,WAAW;AAAA,IAC1C,WAAW,YAAY,MAAM,SAAS;AAAA,EACxC;AACF;AAEA,SAAS,YAAY,OAAgB;AACnC,SAAO,eAAe,KAAK,IAAI,QAAQ;AACzC;AAEA,SAAS,eAAe,OAAiC;AACvD,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,iBAAiB,SAAgD;AACxE,QAAM,WAAW,SAAS,YAAY,CAAC;AACvC,QAAM,YAA+B,CAAC;AAEtC,aAAW,WAAW,UAAU;AAC9B,QAAI,CAAC,MAAM,QAAQ,QAAQ,SAAS,GAAG;AACrC;AAAA,IACF;AAEA,eAAW,QAAQ,QAAQ,WAAW;AACpC,UAAI,CAAC,SAAS,IAAI,KAAK,OAAO,KAAK,SAAS,UAAU;AACpD;AAAA,MACF;AAEA,gBAAU,KAAK;AAAA,QACb,MAAM,KAAK;AAAA,QACX,OAAO,iBAAiB,KAAK,KAAK;AAAA,QAClC,YAAY,YAAY,KAAK,UAAU;AAAA,MACzC,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,iBAAiB,OAAgB;AACxC,MAAI,SAAS,KAAK,KAAK,OAAO,MAAM,YAAY,UAAU;AACxD,WAAO,MAAM;AAAA,EACf;AACA,MAAI,UAAU,QAAW;AACvB,WAAO,eAAe,OAAO,GAAG;AAAA,EAClC;AACA,SAAO;AACT;AAEA,SAAS,kBAAkB,UAA6C;AACtE,QAAM,gBAAgB,CAAC,GAAI,SAAS,MAAM,UAAU,CAAC,CAAE,EACpD;AAAA,IACC,CAACA,YACEA,OAAM,SAAS,KAAK,KACrBA,OAAM,UAAU,cAAc,UAC9BA,OAAM,UAAU,WAAW;AAAA,EAC/B,EACC,KAAK,CAAC,MAAM,WAAW,KAAK,SAAS,MAAM,MAAM,SAAS,EAAE;AAC/D,QAAM,UAAU,cAAc,CAAC;AAC/B,QAAM,QACJ,OAAO,SAAS,UAAU,WACtB,QAAQ,QACR,SAAS,MAAM;AACrB,QAAM,SACJ,gBAAgB,SAAS,UAAU,SAAS,KAC5C,YAAY,SAAS,gBAAgB,KAAK,IAAI,GAAG,GAAG;AAEtD,MAAI,CAAC,WAAW,CAAC,UAAU,UAAU,QAAW;AAC9C,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,WAAW,SAAS;AAAA,IACpB;AAAA,IACA,QAAQ,UAAU;AAAA,EACpB;AACF;AAEA,SAAS,gBAAgB,OAAgB;AACvC,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AACA,SAAO,OAAO,UAAU,WAAW,QAAQ,eAAe,OAAO,GAAI;AACvE;AAEA,SAAS,kBAAkB,WAAgC;AACzD,SAAO,CAAC,GAAG,UAAU,gBAAgB,UAAU,KAAK,EACjD,OAAO,CAAC,SAAS,KAAK,SAAS,CAAC,EAChC,KAAK,KAAK;AACf;AAEA,SAAS,SAAS,OAAmB;AACnC,QAAM,QAAgC;AAAA,IACpC,aAAa;AAAA,IACb,cAAc;AAAA,IACd,iBAAiB;AAAA,IACjB,aAAa;AAAA,IACb,WAAW;AAAA,EACb;AAEA,aAAW,YAAY,OAAO;AAC5B,UAAM,YAAY,SAAS,SAAS;AACpC,UAAM,eAAe,WAAW,eAAe;AAC/C,UAAM,gBAAgB,WAAW,gBAAgB;AACjD,UAAM,mBAAmB,WAAW,mBAAmB;AACvD,UAAM,eACJ,WAAW,gBACV,WAAW,eAAe,MACxB,WAAW,gBAAgB,MAC3B,WAAW,mBAAmB;AACnC,UAAM,aACJ,WAAW,aAAa,SAAS,SAAS,UAAU,UAAU;AAAA,EAClE;AAEA,SAAO;AACT;AAEA,SAAS,mBAAmB,OAAyB;AACnD,QAAM,aAAa,MAAM,YACtB,IAAI,CAAC,SAAS,KAAK,SAAS,EAC5B,OAAO,CAAC,SAAS,OAAO,SAAS,IAAI,CAAC;AACzC,QAAM,WAAW,MAAM,YACpB,IAAI,CAAC,SAAS,KAAK,OAAO,EAC1B,OAAO,CAAC,SAAS,OAAO,SAAS,IAAI,CAAC;AACzC,MAAI,WAAW,WAAW,KAAK,SAAS,WAAW,GAAG;AACpD,WAAO;AAAA,EACT;AACA,SAAO,KAAK,IAAI,GAAG,QAAQ,IAAI,KAAK,IAAI,GAAG,UAAU;AACvD;;;ACtTA,IAAM,mCAAmC;AACzC,IAAM,gCAAgC;AACtC,IAAM,yBAAyB;AAOxB,SAAS,uBACd,QACA,UAA6B,CAAC,GAC9B;AACA,QAAM,iBACJ,QAAQ,kBAAkB;AAE5B,SAAO,OAAO,SACX,OAAO,qBAAqB,EAC5B,MAAM,GAAG,cAAc,EACvB;AAAA,IAAI,CAAC,aACJ,sBAAsB;AAAA,MACpB,SAAS;AAAA,MACT,YAAY;AAAA,QACV,MAAM,SAAS;AAAA,QACf,MAAM,OAAO,SAAS,SAAS,IAAI;AAAA,QACnC,KAAK,OAAO,SAAS,SAAS,MAAM;AAAA,QACpC,OAAO;AAAA,MACT;AAAA,MACA,SAAS,sBAAsB,QAAQ;AAAA,IACzC,CAAC;AAAA,EACH;AACJ;AAGO,SAAS,sBACd,QACA,UAA6B,CAAC,GACX;AACnB,QAAM,iBAAiB,KAAK;AAAA,IAC1B,QAAQ,kBAAkB;AAAA,IAC1B;AAAA,EACF;AAEA,SAAO,OAAO,SACX,OAAO,qBAAqB,EAC5B,MAAM,GAAG,cAAc,EACvB,IAAI,CAAC,cAAc;AAAA,IAClB,MAAM,SAAS;AAAA,IACf,YAAY,SAAS,SAAS;AAAA,IAC9B,UAAU,SAAS,SAAS;AAAA,IAC5B,kBAAkB;AAAA,IAClB,OAAO;AAAA,MACL,GAAG,SAAS,gBAAgB,aAAa,cAAc,MAAM,SAAS,WAAW;AAAA,MACjF;AAAA,IACF;AAAA,IACA,SAAS;AAAA,MACP,sBAAsB,QAAQ;AAAA,MAC9B;AAAA,IACF;AAAA,IACA,aAAa,SAAS,iBAAiB,QAAQ,GAAG,sBAAsB;AAAA,EAC1E,EAAE;AACN;AAEA,SAAS,sBACP,UAC+B;AAC/B,SAAO,QAAQ,SAAS,QAAQ;AAClC;AAEA,SAAS,sBAAsB,UAAoB;AACjD,QAAM,UAAU,SAAS;AACzB,QAAM,QAAQ;AAAA,IACZ,SAAS;AAAA,IACT,SAAS,YAAY,SAAS,SAAS,SAAS,MAAM,QAAQ,CAAC;AAAA,EACjE;AAEA,MAAI,SAAS,WAAW;AACtB,UAAM,KAAK,QAAQ,SAAS;AAAA,EAC9B;AAEA,QAAM,SAAS,YAAY,SAAS,UAAU,IAAI,GAAG;AACrD,MAAI,QAAQ;AACV,UAAM,KAAK,MAAM;AAAA,EACnB;AAEA,SAAO,MAAM,KAAK,KAAK;AACzB;AAEA,SAAS,iBAAiB,UAA6B;AACrD,QAAM,QAAQ;AAAA,IACZ,SAAS,SAAS,WAAW;AAAA,IAC7B,aAAa,SAAS,WAAW,IAAI,SAAS,SAAS,IAAI;AAAA,IAC3D,YAAY,SAAS,SAAS,QAAQ,KAAK;AAAA,IAC3C,UAAU,YAAY,SAAS,gBAAgB,SAAS,SAAS,MAAM,QAAQ,CAAC;AAAA,IAChF,UAAU,SAAS,gBAAgB,aAAa,KAAK;AAAA,IACrD;AAAA,IACA;AAAA,IACA,SAAS,gBAAgB,UAAU;AAAA,EACrC;AAEA,QAAM,cAAc,SAAS,MAAM,UAAU,SAAS,SAAS;AAC/D,MAAI,gBAAgB,QAAW;AAC7B,UAAM,KAAK,IAAI,UAAU,eAAe,aAAa,GAAI,CAAC;AAAA,EAC5D;AAEA,MAAI,SAAS,SAAS,UAAU,QAAQ;AACtC,UAAM,KAAK,IAAI,QAAQ;AACvB,eAAW,YAAY,SAAS,QAAQ,WAAW;AACjD,YAAM;AAAA,QACJ,KAAK,SAAS,IAAI,KAAK,SAAS,QAAQ,UAAU,SAAS,KAAK,KAAK,IAAI;AAAA,MAC3E;AAAA,IACF;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,sBAAsB;AAAA,EAC7B;AAAA,EACA;AAAA,EACA;AACF,GAIG;AACD,QAAM,qBAAqB,OAAO,QAAQ,UAAU,EACjD,IAAI,CAAC,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,IAAI,sBAAsB,KAAK,CAAC,EAAE,EAC9D,KAAK,GAAG;AACX,SAAO,KAAK,OAAO,IAAI,kBAAkB,KAAK,kBAAkB,OAAO,CAAC;AAC1E;;;ACvIA,IAAM,uBAAuB;AAC7B,IAAM,2BAA2B;AACjC,IAAM,2BAA2B;AACjC,IAAM,yBAAyB;AAC/B,IAAM,6BAA6B;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AACA,IAAM,+BAA+B;AAG9B,SAAS,iBACd,QACA,UAA0B,CAAC,GAC3B;AACA,QAAM,cAAc,QAAQ,eAAe;AAC3C,QAAM,WAAW,OAAO,SAAS,MAAM,GAAG,WAAW;AACrD,QAAM,kBAAkB,OAAO,OAAO,SAAS,OAAO,OAAO;AAC7D,QAAM,QAAkB;AAAA,IACtB;AAAA,IACA;AAAA,IACA,GAAG,mBAAmB,QAAQ,eAAe;AAAA,IAC7C;AAAA,IACA,GAAG,wBAAwB,MAAM;AAAA,IACjC;AAAA,IACA;AAAA,EACF;AAEA,MAAI,OAAO,SAAS,SAAS,GAAG;AAC9B,UAAM,KAAK,gBAAgB,EAAE;AAC7B,aAAS,QAAQ,CAAC,UAAU,UAAU;AACpC,YAAM,KAAK,GAAG,qBAAqB,UAAU,QAAQ,GAAG,OAAO,GAAG,EAAE;AAAA,IACtE,CAAC;AAED,QAAI,OAAO,SAAS,SAAS,SAAS,QAAQ;AAC5C,YAAM;AAAA,QACJ,GAAG,OAAO,SAAS,SAAS,SAAS,MAAM;AAAA,QAC3C;AAAA,MACF;AAAA,IACF;AAAA,EACF,WAAW,OAAO,OAAO,YAAY,GAAG;AACtC,UAAM,KAAK,gBAAgB,IAAI,qBAAqB,EAAE;AAAA,EACxD;AAEA,MAAI,OAAO,OAAO,cAAc,GAAG;AACjC,UAAM,KAAK,yDAAyD,EAAE;AAAA,EACxE;AAEA,SAAO,GAAG,MAAM,KAAK,IAAI,CAAC;AAAA;AAC5B;AAEA,SAAS,gBAAgB,QAAgB,QAAgB,OAAe;AACtE,SAAO,GAAG,aAAa,MAAM,CAAC,YAAY,aAAa,MAAM,CAAC,YAAY,aAAa,KAAK,CAAC;AAC/F;AAEA,SAAS,mBAAmB,QAAoB,iBAAyB;AACvE,QAAM,OAAgC;AAAA,IACpC,CAAC,UAAU,OAAO,MAAM;AAAA,IACxB;AAAA,MACE;AAAA,MACA;AAAA,QACE,OAAO,OAAO;AAAA,QACd,OAAO,OAAO;AAAA,QACd,OAAO,OAAO;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AAEA,MAAI,OAAO,OAAO;AAChB,SAAK,KAAK,CAAC,SAAS,mBAAmB,OAAO,KAAK,CAAC,CAAC;AAAA,EACvD;AAEA,MAAI,kBAAkB,GAAG;AACvB,SAAK,KAAK;AAAA,MACR;AAAA,MACA,GAAG,aAAa,eAAe,CAAC,yBAC9B,oBAAoB,IAAI,KAAK,GAC/B;AAAA,IACF,CAAC;AAAA,EACH;AAEA,OAAK,KAAK,CAAC,YAAY,eAAe,OAAO,UAAU,CAAC,CAAC;AAEzD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,GAAG,KAAK;AAAA,MACN,CAAC,CAAC,QAAQ,KAAK,MACb,KAAK,gBAAgB,MAAM,CAAC,MAAM,gBAAgB,KAAK,CAAC;AAAA,IAC5D;AAAA,EACF;AACF;AAEA,SAAS,mBAAmB,OAAyC;AACnE,SAAO,OAAO,YAAY,MAAM,OAAO,CAAC,GACtC,MAAM,YAAY,SAAY,KAAK,SAAS,YAAY,MAAM,OAAO,CAAC,EACxE;AACF;AAEA,SAAS,gBAAgB,OAAe;AACtC,SAAO,MACJ,QAAQ,UAAU,GAAG,EACrB,QAAQ,OAAO,MAAM,EACrB,QAAQ,OAAO,KAAK;AACzB;AAEA,SAAS,wBAAwB,QAAoB;AACnD,QAAM,SAAS,OAAO,MACnB,IAAI,CAAC,aAAa,SAAS,MAAM,QAAQ,EACzC;AAAA,IACC,CAAC,UACC,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAAA,EACtD;AACF,MAAI,OAAO,WAAW,GAAG;AACvB,WAAO,CAAC;AAAA,EACV;AAEA,QAAM,SAAS,2BAA2B,IAAI,MAAM,CAAC;AACrD,aAAW,SAAS,QAAQ;AAC1B,UAAM,SAAS,KAAK;AAAA,MAClB,2BAA2B,SAAS;AAAA,MACpC,KAAK,IAAI,GAAG,KAAK,MAAM,QAAQ,2BAA2B,MAAM,CAAC;AAAA,IACnE;AACA,WAAO,MAAM,KAAK,OAAO,MAAM,KAAK,KAAK;AAAA,EAC3C;AAEA,QAAM,WAAW,KAAK,IAAI,GAAG,MAAM;AACnC,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG,2BAA2B;AAAA,MAAI,CAAC,OAAO,UACxC,8BAA8B,OAAO,OAAO,KAAK,KAAK,GAAG,QAAQ;AAAA,IACnE;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,8BACP,OACA,OACA,UACA;AACA,QAAM,YACJ,UAAU,IACN,IACA,KAAK;AAAA,IACH;AAAA,IACA,KAAK,MAAO,QAAQ,WAAY,4BAA4B;AAAA,EAC9D;AACN,QAAM,MAAM,IAAI,OAAO,SAAS,EAAE,OAAO,8BAA8B,GAAG;AAC1E,SAAO,GAAG,MAAM,OAAO,CAAC,CAAC,MAAM,GAAG,IAAI,aAAa,KAAK,CAAC;AAC3D;AAEA,SAAS,qBACP,UACA,QACA,SACA;AACA,QAAM,UAAU,SAAS;AACzB,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,eAAe,QAAQ,gBAAgB;AAC7C,QAAM,QAAQ,gBAAgB,QAAQ;AACtC,QAAM,cAAc,SAAS,MAAM,UAAU,SAAS,SAAS;AAC/D,QAAM,UAAU,GAAG,MAAM,KAAK,SAAS,WAAW,MAChD,SAAS,aAAa,SACxB,MAAM,YAAY,SAAS,SAAS,SAAS,MAAM,QAAQ,CAAC;AAC5D,QAAM,QAAQ;AAAA,IACZ;AAAA,IACA,YAAY,WAAW,OAAO,CAAC;AAAA,IAC/B;AAAA,IACA;AAAA,IACA,GAAG,mBAAmB,UAAU;AAAA,MAC9B;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC,EAAE,IAAI,WAAW;AAAA,IAClB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,mBACP,UACA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAQA;AACA,QAAM,UAAU,SAAS;AACzB,QAAM,eAAwC;AAAA,IAC5C,CAAC,QAAQ,GAAG,MAAM,KAAK,SAAS,WAAW,EAAE;AAAA,IAC7C,CAAC,UAAU,SAAS,MAAM;AAAA,IAC1B,CAAC,YAAY,eAAe,SAAS,aAAa,SAAS,QAAQ,CAAC;AAAA,IACpE,CAAC,WAAW,SAAS,SAAS,QAAQ,KAAK;AAAA,IAC3C,CAAC,SAAS,YAAY,SAAS,SAAS,SAAS,MAAM,QAAQ,CAAC;AAAA,IAChE,CAAC,SAAS,SAAS,aAAa,KAAK;AAAA,EACvC;AACA,MAAI,OAAO;AACT,iBAAa,KAAK,CAAC,SAAS,KAAK,CAAC;AAAA,EACpC;AACA,MAAI,SAAS,eAAe,QAAW;AACrC,iBAAa,KAAK,CAAC,YAAY,eAAe,SAAS,UAAU,CAAC,CAAC;AAAA,EACrE;AAEA,QAAM,QAAQ;AAAA,IACZ,GAAG,mBAAmB,UAAU,gBAAgB,YAAY,CAAC;AAAA,IAC7D;AAAA,EACF;AAEA,MAAI,SAAS,QAAQ;AACnB,UAAM;AAAA,MACJ,GAAG;AAAA,QACD;AAAA,QACA,SAAS,QAAQ,QAAQ,cAAc,EAAE,MAAM,OAAO;AAAA,MACxD;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,MAAI,SAAS,MAAM,OAAO,QAAQ;AAChC,UAAM;AAAA,MACJ,GAAG;AAAA,QACD,CAAC,SAAS,OAAO;AAAA,QACjB,SAAS,KAAK,OAAO,IAAI,CAAC,UAAU;AAAA,UAClC,MAAM,QAAQ;AAAA,UACd,YAAY,MAAM,KAAK;AAAA,QACzB,CAAC;AAAA,MACH;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,MAAI,gBAAgB,QAAW;AAC7B,UAAM;AAAA,MACJ,GAAG;AAAA,QACD;AAAA,QACA,eAAe,aAAa,cAAc,EAAE,MAAM,OAAO;AAAA,MAC3D;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,MAAI,SAAS,SAAS,UAAU,QAAQ;AACtC,UAAM,YAAY,SAAS,QAAQ,UAAU,MAAM,GAAG,YAAY;AAClE,UAAM;AAAA,MACJ,GAAG;AAAA,QACD,CAAC,QAAQ,UAAU,UAAU;AAAA,QAC7B,UAAU,IAAI,CAAC,aAAa;AAAA,UAC1B,SAAS;AAAA,UACT,SAAS,QAAQ,UAAU,YAAY,SAAS,OAAO,GAAG,CAAC,KAAK;AAAA,UAChE,SAAS,eAAe,SACpB,QACA,eAAe,SAAS,UAAU;AAAA,QACxC,CAAC;AAAA,MACH;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU,SAAS,cAAc;AACpD,YAAM;AAAA,QACJ,GAAG,SAAS,QAAQ,UAAU,SAAS,YAAY;AAAA,MACrD;AAAA,IACF;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,SAAS,SAAS,OAAO,QAAQ;AACnC,UAAM;AAAA,MACJ,GAAG;AAAA,QACD;AAAA,QACA,eAAe,SAAS,QAAQ,QAAQ,cAAc,EAAE,MAAM,OAAO;AAAA,MACvE;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,SAAO,MAAM,MAAM,SAAS,CAAC,MAAM,IAAI;AACrC,UAAM,IAAI;AAAA,EACZ;AACA,SAAO;AACT;AAEA,SAAS,mBAAmB,OAAe,SAAmB;AAC5D,SAAO,CAAC,OAAO,IAAI,OAAO,MAAM,MAAM,GAAG,GAAG,OAAO;AACrD;AAEA,SAAS,gBAAgB,MAA+B;AACtD,QAAM,aAAa,KAAK,IAAI,GAAG,KAAK,IAAI,CAAC,CAAC,KAAK,MAAM,MAAM,MAAM,CAAC;AAClE,SAAO,KAAK;AAAA,IACV,CAAC,CAAC,OAAO,KAAK,MACZ,GAAG,MAAM,OAAO,UAAU,CAAC,KAAK,YAAY,OAAO,GAAG,CAAC;AAAA,EAC3D;AACF;AAEA,SAAS,iBAAiB,SAAmB,MAAkB;AAC7D,QAAM,SAAS,QAAQ;AAAA,IAAI,CAAC,QAAQ,UAClC,KAAK,IAAI,OAAO,QAAQ,GAAG,KAAK,IAAI,CAAC,QAAQ,IAAI,KAAK,GAAG,UAAU,CAAC,CAAC;AAAA,EACvE;AACA,QAAM,YAAY,CAAC,QACjB,IACG,IAAI,CAAC,MAAM,UAAU,KAAK,OAAO,OAAO,KAAK,KAAK,KAAK,MAAM,CAAC,EAC9D,KAAK,IAAI,EACT,QAAQ;AAEb,SAAO;AAAA,IACL,UAAU,OAAO;AAAA,IACjB,OAAO,IAAI,CAAC,UAAU,IAAI,OAAO,KAAK,CAAC,EAAE,KAAK,IAAI;AAAA,IAClD,GAAG,KAAK,IAAI,SAAS;AAAA,EACvB;AACF;AAEA,SAAS,gBAAgB,UAAoB;AAC3C,QAAM,QAAQ,SAAS,SAAS;AAChC,QAAM,QAAkB,CAAC;AACzB,QAAM,cACJ,OAAO,gBACN,OAAO,eAAe,MACpB,OAAO,gBAAgB,MACvB,OAAO,mBAAmB;AAC/B,QAAM,YAAY,OAAO,aAAa,SAAS,SAAS,UAAU,UAAU;AAE5E,MAAI,cAAc,GAAG;AACnB,UAAM,KAAK,GAAG,aAAa,WAAW,CAAC,SAAS;AAAA,EAClD;AACA,MAAI,YAAY,GAAG;AACjB,UAAM,KAAK,GAAG,aAAa,SAAS,CAAC,QAAQ,cAAc,IAAI,KAAK,GAAG,EAAE;AAAA,EAC3E;AACA,MAAI,SAAS,SAAS,aAAa,QAAW;AAC5C,UAAM,KAAK,eAAe,SAAS,QAAQ,QAAQ,CAAC;AAAA,EACtD;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;;;AC1VA,IAAM,qBAAqB;AAC3B,IAAM,2BAA2B;AACjC,IAAM,kCACJ;AAGF,eAAsB,gBACpB,QACA,UAAkC,CAAC,GACH;AAChC,QAAM,QAAQ,QAAQ,SAAS,QAAQ,IAAI;AAC3C,QAAM,aAAa,QAAQ,cAAc,QAAQ,IAAI;AACrD,QAAM,MAAM,QAAQ,OAAO,QAAQ,IAAI;AAEvC,MAAI,CAAC,OAAO;AACV,WAAO,EAAE,QAAQ,WAAW,QAAQ,uBAAuB;AAAA,EAC7D;AACA,MAAI,CAAC,YAAY;AACf,WAAO,EAAE,QAAQ,WAAW,QAAQ,4BAA4B;AAAA,EAClE;AACA,MAAI,CAAC,OAAO,QAAQ,eAAe,QAAW;AAC5C,WAAO,EAAE,QAAQ,WAAW,QAAQ,qBAAqB;AAAA,EAC3D;AAEA,QAAM,CAAC,OAAO,IAAI,IAAI,WAAW,MAAM,GAAG;AAC1C,MAAI,CAAC,SAAS,CAAC,MAAM;AACnB,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,QAAQ,8BAA8B,UAAU;AAAA,IAClD;AAAA,EACF;AAEA,QAAM,UAAU,qBAAqB,QAAQ,OAAO;AACpD,QAAM,SACJ,QAAQ,UAAU,QAAQ,IAAI,kBAAkB;AAClD,QAAM,aACJ,QAAQ,eAAe,SACnB,GAAG,MAAM,UAAU,KAAK,IAAI,IAAI,gBAChC,GAAG,MAAM,UAAU,KAAK,IAAI,IAAI,eAAe,QAAQ,UAAU;AACvE,QAAM,WAAW,MAAM,MAAM,YAAY;AAAA,IACvC,QAAQ,QAAQ,eAAe,SAAY,SAAS;AAAA,IACpD,SAAS;AAAA,MACP,QAAQ;AAAA,MACR,eAAe,UAAU,KAAK;AAAA,MAC9B,gBAAgB;AAAA,MAChB,wBAAwB;AAAA,IAC1B;AAAA,IACA,MAAM,KAAK;AAAA,MACT,QAAQ,eAAe,SACnB;AAAA,QACE,MAAM,QAAQ,QAAQ;AAAA,QACtB,UAAU;AAAA,QACV,GAAG;AAAA,MACL,IACA;AAAA,IACN;AAAA,EACF,CAAC;AAED,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,OAAO,MAAM,SAAS,KAAK;AACjC,UAAM,IAAI;AAAA,MACR,oCAAoC,SAAS,MAAM,IAAI,SAAS,UAAU,IAAI,IAAI,GAAG,KAAK;AAAA,IAC5F;AAAA,EACF;AAEA,QAAM,OAAQ,MAAM,SAAS,KAAK;AAKlC,SAAO;AAAA,IACL,QAAQ,QAAQ,eAAe,SAAY,YAAY;AAAA,IACvD,IAAI,KAAK;AAAA,IACT,SAAS,KAAK;AAAA,EAChB;AACF;AAEA,SAAS,qBACP,QACA,SACA;AACA,QAAM,cAAc,sBAAsB,QAAQ;AAAA,IAChD,gBAAgB,QAAQ;AAAA,EAC1B,CAAC;AACD,QAAM,QACJ,OAAO,SAAS,WAAW,KAAK,OAAO,WAAW,WAC9C,qBACA,OAAO,SAAS,WAAW,IACzB,sBACA,GAAG,OAAO,SAAS,MAAM,gBACvB,OAAO,SAAS,WAAW,IAAI,KAAK,GACtC;AAER,SAAO;AAAA,IACL,QAAQ;AAAA,IACR,YAAY,OAAO,WAAW,WAAW,YAAY;AAAA,IACrD,eAAc,oBAAI,KAAK,GAAE,YAAY;AAAA,IACrC,QAAQ;AAAA,MACN;AAAA,MACA,SAAS;AAAA,QACP,iBAAiB,QAAQ;AAAA,UACvB,GAAG;AAAA,UACH,aAAa,QAAQ,eAAe;AAAA,UACpC,gBAAgB,QAAQ,kBAAkB;AAAA,UAC1C,gBAAgB,QAAQ,kBAAkB;AAAA,UAC1C,cAAc,QAAQ,gBAAgB;AAAA,QACxC,CAAC;AAAA,MACH;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,qBAAqB,SAAiB;AAC7C,MAAI,QAAQ,UAAU,0BAA0B;AAC9C,WAAO;AAAA,EACT;AAEA,SAAO,GAAG,QACP,MAAM,GAAG,2BAA2B,gCAAgC,MAAM,EAC1E,QAAQ,CAAC,GAAG,+BAA+B;AAChD;","names":["score"]}
|
|
1
|
+
{"version":3,"sources":["../src/collect.ts","../src/utils.ts","../src/annotations.ts","../src/summary.ts","../src/github.ts"],"sourcesContent":["import {\n collectReportWorkspace,\n type HarnessRun,\n type ReportCase,\n type ToolCallRecord,\n} from \"@vitest-evals/core\";\nimport type {\n CollectOptions,\n EvalCase,\n EvalFailure,\n EvalReport,\n EvalScore,\n ToolCallSummary,\n UsageSummary,\n VitestJsonReport,\n} from \"./types\";\nimport { compactLine, stringifyValue } from \"./utils\";\n\n/** Converts a Vitest JSON report into the compact eval report model. */\nexport function collectEvalReport(\n input: VitestJsonReport,\n options: CollectOptions = {},\n): EvalReport {\n const workspace = collectReportWorkspace(\n {\n report: input,\n },\n {\n workspace: options.workspace,\n },\n );\n const cases = workspace.cases.map(collectEvalCase);\n const failures = cases.filter((testCase) => testCase.status === \"failed\");\n const evalScores = cases\n .map((testCase) => testCase.eval?.avgScore)\n .filter((score): score is number => isFiniteNumber(score));\n const usage = sumUsage(cases);\n const durationMs = workspace.runs[0]?.durationMs;\n\n return {\n status: input.success && failures.length === 0 ? \"passed\" : \"failed\",\n startedAt: input.startTime,\n durationMs,\n totals: {\n total: input.numTotalTests,\n passed: input.numPassedTests,\n failed: input.numFailedTests,\n skipped: input.numPendingTests + input.numTodoTests,\n evalTotal: cases.length,\n evalPassed: cases.filter((testCase) => testCase.status === \"passed\")\n .length,\n evalFailed: failures.length,\n },\n score:\n evalScores.length > 0\n ? {\n average:\n evalScores.reduce((total, score) => total + score, 0) /\n evalScores.length,\n minimum: Math.min(...evalScores),\n }\n : undefined,\n usage,\n cases,\n failures,\n };\n}\n\nfunction collectEvalCase(reportCase: ReportCase): EvalCase {\n const scores = (reportCase.eval?.scores ?? []).map(normalizeScore);\n const harnessRun = reportCase.harness?.run;\n const toolCalls = collectToolCalls(reportCase);\n const evalCase: EvalCase = {\n id: reportCase.id,\n file: reportCase.file,\n displayFile: reportCase.displayFile,\n title: reportCase.title,\n displayName: reportCase.displayName,\n status: reportCase.status,\n durationMs: numberField(reportCase.durationMs),\n location: reportCase.location,\n failureMessages: reportCase.failureMessages,\n toolCalls,\n eval: reportCase.eval\n ? {\n avgScore: numberField(reportCase.eval.avgScore),\n thresholdFailed: reportCase.eval.thresholdFailed,\n output: reportCase.eval.output,\n scores,\n }\n : undefined,\n harness: reportCase.harness\n ? {\n name: reportCase.harness.name,\n output: harnessRun?.output,\n usage: harnessRun?.usage,\n timingMs: harnessRun?.timings?.totalMs,\n toolCalls,\n errors: harnessRun?.errors ?? [],\n }\n : undefined,\n };\n\n evalCase.primaryFailure = getPrimaryFailure(evalCase);\n return evalCase;\n}\n\nfunction normalizeScore(score: EvalScore): EvalScore {\n return {\n ...score,\n score: numberField(score.score) ?? null,\n };\n}\n\nfunction numberField(value: unknown) {\n return isFiniteNumber(value) ? value : undefined;\n}\n\nfunction isFiniteNumber(value: unknown): value is number {\n return typeof value === \"number\" && Number.isFinite(value);\n}\n\nfunction collectToolCalls(reportCase: ReportCase) {\n const harnessCalls = collectSessionToolCalls(\n reportCase.harness?.run?.session,\n );\n return harnessCalls.length > 0\n ? harnessCalls\n : (reportCase.eval?.toolCalls ?? []).map(toToolCallSummary);\n}\n\nfunction collectSessionToolCalls(session: HarnessRun[\"session\"] | undefined) {\n const messages = session?.messages ?? [];\n const toolCalls: ToolCallSummary[] = [];\n\n for (const message of messages) {\n if (!Array.isArray(message.toolCalls)) {\n continue;\n }\n\n for (const call of message.toolCalls) {\n toolCalls.push(toToolCallSummary(call));\n }\n }\n\n return toolCalls;\n}\n\nfunction toToolCallSummary(call: ToolCallRecord): ToolCallSummary {\n return {\n name: call.name,\n error: getToolCallError(call.error),\n durationMs: numberField(call.durationMs),\n };\n}\n\nfunction getToolCallError(value: unknown) {\n const error = value as { message?: unknown };\n if (\n value &&\n typeof value === \"object\" &&\n !Array.isArray(value) &&\n typeof error.message === \"string\"\n ) {\n return error.message;\n }\n if (value !== undefined) {\n return stringifyValue(value, 240);\n }\n return undefined;\n}\n\nfunction getPrimaryFailure(testCase: EvalCase): EvalFailure | undefined {\n const failingScores = [...(testCase.eval?.scores ?? [])]\n .filter(\n (score) =>\n (score.score ?? 0) < 1 ||\n score.metadata?.rationale !== undefined ||\n score.metadata?.output !== undefined,\n )\n .sort((left, right) => (left.score ?? 0) - (right.score ?? 0));\n const primary = failingScores[0];\n const score =\n typeof primary?.score === \"number\"\n ? primary.score\n : testCase.eval?.avgScore;\n const reason =\n stringifyReason(primary?.metadata?.rationale) ??\n compactLine(testCase.failureMessages.join(\"\\n\"), 500);\n\n if (!primary && !reason && score === undefined) {\n return undefined;\n }\n\n return {\n judgeName: primary?.name,\n score,\n reason: reason || undefined,\n };\n}\n\nfunction stringifyReason(value: unknown) {\n if (value === undefined) {\n return undefined;\n }\n return typeof value === \"string\" ? value : stringifyValue(value, 4000);\n}\n\nfunction sumUsage(cases: EvalCase[]) {\n const usage: Required<UsageSummary> = {\n inputTokens: 0,\n outputTokens: 0,\n reasoningTokens: 0,\n totalTokens: 0,\n toolCalls: 0,\n };\n\n for (const testCase of cases) {\n const caseUsage = testCase.harness?.usage;\n usage.inputTokens += caseUsage?.inputTokens ?? 0;\n usage.outputTokens += caseUsage?.outputTokens ?? 0;\n usage.reasoningTokens += caseUsage?.reasoningTokens ?? 0;\n usage.totalTokens +=\n caseUsage?.totalTokens ??\n (caseUsage?.inputTokens ?? 0) +\n (caseUsage?.outputTokens ?? 0) +\n (caseUsage?.reasoningTokens ?? 0);\n usage.toolCalls += toolCallCount(testCase);\n }\n\n return usage;\n}\n\nfunction toolCallCount(testCase: EvalCase) {\n const usageToolCalls = testCase.harness?.usage?.toolCalls;\n if (usageToolCalls !== undefined) {\n return Math.max(usageToolCalls, testCase.toolCalls.length);\n }\n\n return testCase.toolCalls.length;\n}\n","import { posix } from \"node:path\";\n\nexport function isRecord(value: unknown): value is Record<string, unknown> {\n return Boolean(value) && typeof value === \"object\" && !Array.isArray(value);\n}\n\nexport function compactLine(value: string, maxLength: number) {\n const line = value\n .split(/\\r?\\n/)\n .map((part) => part.trim())\n .find((part) => part.length > 0);\n if (!line) {\n return \"\";\n }\n return truncate(line, maxLength);\n}\n\nexport function truncate(value: string, maxLength: number) {\n if (maxLength <= 0) {\n return \"\";\n }\n if (value.length <= maxLength) {\n return value;\n }\n return `${value.slice(0, Math.max(0, maxLength - 15)).trimEnd()}... [truncated]`;\n}\n\nexport function stringifyValue(value: unknown, maxLength: number) {\n if (value === undefined) {\n return \"\";\n }\n if (typeof value === \"string\") {\n return truncate(value, maxLength);\n }\n try {\n return truncate(JSON.stringify(value, null, 2), maxLength);\n } catch {\n return truncate(String(value), maxLength);\n }\n}\n\nexport function formatNumber(value: number) {\n return new Intl.NumberFormat(\"en-US\", {\n maximumFractionDigits: 0,\n }).format(value);\n}\n\nexport function formatScore(value: number | null | undefined) {\n if (value == null || !Number.isFinite(value)) {\n return \"n/a\";\n }\n return value.toFixed(2);\n}\n\nexport function formatDuration(ms: number | undefined) {\n if (ms === undefined || !Number.isFinite(ms)) {\n return \"n/a\";\n }\n if (ms < 1000) {\n return `${Math.round(ms)}ms`;\n }\n const seconds = ms / 1000;\n if (seconds < 60) {\n return `${seconds.toFixed(seconds < 10 ? 1 : 0)}s`;\n }\n const totalSeconds = Math.round(seconds);\n const minutes = Math.floor(totalSeconds / 60);\n const remainingSeconds = totalSeconds % 60;\n return `${minutes}m ${remainingSeconds}s`;\n}\n\nexport function formatLocation(\n file: string,\n location?: { line: number; column: number },\n) {\n if (!location) {\n return file;\n }\n return `${file}:${location.line}`;\n}\n\nexport function normalizePathForGitHub(path: string, workspace?: string) {\n const normalized = path.replace(/\\\\/g, \"/\");\n if (!workspace) {\n return normalized;\n }\n\n const workspacePath = workspace.replace(/\\\\/g, \"/\").replace(/\\/+$/, \"\");\n if (\n normalized !== workspacePath &&\n !normalized.startsWith(`${workspacePath}/`)\n ) {\n return normalized;\n }\n\n return posix.relative(workspacePath, normalized);\n}\n\nexport function escapeFence(value: string) {\n return value.replace(/```/g, \"'''\");\n}\n\nexport function escapeHtml(value: string) {\n return value\n .replace(/&/g, \"&\")\n .replace(/</g, \"<\")\n .replace(/>/g, \">\")\n .replace(/\"/g, \""\");\n}\n\nexport function escapeCommandData(value: string) {\n return value.replace(/%/g, \"%25\").replace(/\\r/g, \"%0D\").replace(/\\n/g, \"%0A\");\n}\n\nexport function escapeCommandProperty(value: string) {\n return escapeCommandData(value).replace(/:/g, \"%3A\").replace(/,/g, \"%2C\");\n}\n","import type { EvalCase, EvalReport } from \"./types\";\nimport {\n compactLine,\n escapeCommandData,\n escapeCommandProperty,\n formatScore,\n stringifyValue,\n truncate,\n} from \"./utils\";\n\n/** Options for limiting rendered GitHub annotations. */\nexport type AnnotationOptions = {\n maxAnnotations?: number;\n};\n\n/** GitHub Check Run annotation payload. */\nexport type CheckAnnotation = {\n path: string;\n start_line: number;\n end_line: number;\n annotation_level: \"failure\" | \"warning\" | \"notice\";\n message: string;\n title?: string;\n raw_details?: string;\n};\n\nconst DEFAULT_MAX_WORKFLOW_ANNOTATIONS = 10;\nconst DEFAULT_MAX_CHECK_ANNOTATIONS = 50;\nconst MAX_CHECK_FIELD_LENGTH = 64_000;\n\ntype AnnotatedEvalCase = EvalCase & {\n location: { line: number; column: number };\n};\n\n/** Renders GitHub workflow-command annotations for failed eval cases. */\nexport function renderWorkflowCommands(\n report: EvalReport,\n options: AnnotationOptions = {},\n) {\n const maxAnnotations =\n options.maxAnnotations ?? DEFAULT_MAX_WORKFLOW_ANNOTATIONS;\n\n return report.failures\n .filter(hasAnnotationLocation)\n .slice(0, maxAnnotations)\n .map((testCase) =>\n formatWorkflowCommand({\n command: \"error\",\n properties: {\n file: testCase.displayFile,\n line: String(testCase.location.line),\n col: String(testCase.location.column),\n title: \"vitest-evals\",\n },\n message: formatWorkflowMessage(testCase),\n }),\n );\n}\n\n/** Builds Check Run annotations for failed eval cases. */\nexport function buildCheckAnnotations(\n report: EvalReport,\n options: AnnotationOptions = {},\n): CheckAnnotation[] {\n const maxAnnotations = Math.min(\n options.maxAnnotations ?? DEFAULT_MAX_CHECK_ANNOTATIONS,\n DEFAULT_MAX_CHECK_ANNOTATIONS,\n );\n\n return report.failures\n .filter(hasAnnotationLocation)\n .slice(0, maxAnnotations)\n .map((testCase) => ({\n path: testCase.displayFile,\n start_line: testCase.location.line,\n end_line: testCase.location.line,\n annotation_level: \"failure\",\n title: truncate(\n `${testCase.primaryFailure?.judgeName ?? \"vitest-evals\"} - ${testCase.displayName}`,\n 255,\n ),\n message: truncate(\n formatWorkflowMessage(testCase),\n MAX_CHECK_FIELD_LENGTH,\n ),\n raw_details: truncate(formatRawDetails(testCase), MAX_CHECK_FIELD_LENGTH),\n }));\n}\n\nfunction hasAnnotationLocation(\n testCase: EvalCase,\n): testCase is AnnotatedEvalCase {\n return Boolean(testCase.location);\n}\n\nfunction formatWorkflowMessage(testCase: EvalCase) {\n const failure = testCase.primaryFailure;\n const parts = [\n testCase.displayName,\n `score ${formatScore(failure?.score ?? testCase.eval?.avgScore)}`,\n ];\n\n if (failure?.judgeName) {\n parts.push(failure.judgeName);\n }\n\n const reason = compactLine(failure?.reason ?? \"\", 320);\n if (reason) {\n parts.push(reason);\n }\n\n return parts.join(\" - \");\n}\n\nfunction formatRawDetails(testCase: AnnotatedEvalCase) {\n const lines = [\n `Test: ${testCase.displayName}`,\n `Location: ${testCase.displayFile}:${testCase.location.line}`,\n `Harness: ${testCase.harness?.name ?? \"n/a\"}`,\n `Score: ${formatScore(testCase.primaryFailure?.score ?? testCase.eval?.avgScore)}`,\n `Judge: ${testCase.primaryFailure?.judgeName ?? \"n/a\"}`,\n \"\",\n \"Reason:\",\n testCase.primaryFailure?.reason ?? \"n/a\",\n ];\n\n const finalOutput = testCase.eval?.output ?? testCase.harness?.output;\n if (finalOutput !== undefined) {\n lines.push(\"\", \"Final:\", stringifyValue(finalOutput, 8000));\n }\n\n if (testCase.toolCalls.length) {\n lines.push(\"\", \"Tools:\");\n for (const toolCall of testCase.toolCalls) {\n lines.push(\n `- ${toolCall.name}: ${toolCall.error ? `error: ${toolCall.error}` : \"ok\"}`,\n );\n }\n }\n\n return lines.join(\"\\n\");\n}\n\nfunction formatWorkflowCommand({\n command,\n properties,\n message,\n}: {\n command: \"error\" | \"warning\" | \"notice\";\n properties: Record<string, string>;\n message: string;\n}) {\n const renderedProperties = Object.entries(properties)\n .map(([key, value]) => `${key}=${escapeCommandProperty(value)}`)\n .join(\",\");\n return `::${command} ${renderedProperties}::${escapeCommandData(message)}`;\n}\n","import type { EvalCase, EvalReport } from \"./types\";\nimport {\n compactLine,\n escapeFence,\n escapeHtml,\n formatDuration,\n formatLocation,\n formatNumber,\n formatScore,\n stringifyValue,\n truncate,\n} from \"./utils\";\n\n/** Options for rendering GitHub Actions summary markdown. */\nexport type SummaryOptions = {\n maxFailures?: number;\n maxReasonChars?: number;\n maxOutputChars?: number;\n maxToolCalls?: number;\n};\n\nconst DEFAULT_MAX_FAILURES = 20;\nconst DEFAULT_MAX_REASON_CHARS = 8000;\nconst DEFAULT_MAX_OUTPUT_CHARS = 4000;\nconst DEFAULT_MAX_TOOL_CALLS = 20;\nconst SCORE_DISTRIBUTION_BUCKETS = [\n \"0-19%\",\n \"20-39%\",\n \"40-59%\",\n \"60-79%\",\n \"80-100%\",\n] as const;\nconst SCORE_DISTRIBUTION_BAR_WIDTH = 20;\n\n/** Renders the GitHub Actions job summary markdown for an eval report. */\nexport function renderJobSummary(\n report: EvalReport,\n options: SummaryOptions = {},\n) {\n const maxFailures = options.maxFailures ?? DEFAULT_MAX_FAILURES;\n const failures = report.failures.slice(0, maxFailures);\n const nonEvalFailures = report.totals.failed - report.totals.evalFailed;\n const lines: string[] = [\n \"# vitest-evals\",\n \"\",\n ...renderSummaryTable(report, nonEvalFailures),\n \"\",\n ...renderScoreDistribution(report),\n \"## Results\",\n \"\",\n ];\n\n if (report.failures.length > 0) {\n lines.push(\"### Failures\", \"\");\n failures.forEach((testCase, index) => {\n lines.push(...renderFailureDetails(testCase, index + 1, options), \"\");\n });\n\n if (report.failures.length > failures.length) {\n lines.push(\n `${report.failures.length - failures.length} more failures omitted from this summary.`,\n \"\",\n );\n }\n } else if (report.totals.evalTotal > 0) {\n lines.push(\"### Failures\", \"\", \"No eval failures.\", \"\");\n }\n\n if (report.totals.evalTotal === 0) {\n lines.push(\"No eval metadata was found in the Vitest JSON report.\", \"\");\n }\n\n return `${lines.join(\"\\n\")}\\n`;\n}\n\nfunction formatCountLine(passed: number, failed: number, total: number) {\n return `${formatNumber(passed)} passed, ${formatNumber(failed)} failed, ${formatNumber(total)} total`;\n}\n\nfunction renderSummaryTable(report: EvalReport, nonEvalFailures: number) {\n const rows: Array<[string, string]> = [\n [\"Status\", report.status],\n [\n \"Evals\",\n formatCountLine(\n report.totals.evalPassed,\n report.totals.evalFailed,\n report.totals.evalTotal,\n ),\n ],\n ];\n\n if (report.score) {\n rows.push([\"Score\", formatScoreSummary(report.score)]);\n }\n\n if (nonEvalFailures > 0) {\n rows.push([\n \"Other Failures\",\n `${formatNumber(nonEvalFailures)} non-eval test failure${\n nonEvalFailures === 1 ? \"\" : \"s\"\n }`,\n ]);\n }\n\n rows.push([\"Duration\", formatDuration(report.durationMs)]);\n\n return [\n \"| Metric | Value |\",\n \"| --- | --- |\",\n ...rows.map(\n ([metric, value]) =>\n `| ${escapeTableCell(metric)} | ${escapeTableCell(value)} |`,\n ),\n ];\n}\n\nfunction formatScoreSummary(score: NonNullable<EvalReport[\"score\"]>) {\n return `avg ${formatScore(score.average)}${\n score.minimum === undefined ? \"\" : `, min ${formatScore(score.minimum)}`\n }`;\n}\n\nfunction escapeTableCell(value: string) {\n return value\n .replace(/\\r?\\n/g, \" \")\n .replace(/\\\\/g, \"\\\\\\\\\")\n .replace(/\\|/g, \"\\\\|\");\n}\n\nfunction renderScoreDistribution(report: EvalReport) {\n const scores = report.cases\n .map((testCase) => testCase.eval?.avgScore)\n .filter(\n (score): score is number =>\n typeof score === \"number\" && Number.isFinite(score),\n );\n if (scores.length === 0) {\n return [];\n }\n\n const counts = SCORE_DISTRIBUTION_BUCKETS.map(() => 0);\n for (const score of scores) {\n const bucket = Math.min(\n SCORE_DISTRIBUTION_BUCKETS.length - 1,\n Math.max(0, Math.floor(score * SCORE_DISTRIBUTION_BUCKETS.length)),\n );\n counts[bucket] = (counts[bucket] ?? 0) + 1;\n }\n\n const maxCount = Math.max(...counts);\n return [\n \"## Scores\",\n \"\",\n \"```text\",\n ...SCORE_DISTRIBUTION_BUCKETS.map((label, index) =>\n formatScoreDistributionBucket(label, counts[index] ?? 0, maxCount),\n ),\n \"```\",\n \"\",\n ];\n}\n\nfunction formatScoreDistributionBucket(\n label: string,\n count: number,\n maxCount: number,\n) {\n const barLength =\n count === 0\n ? 0\n : Math.max(\n 1,\n Math.round((count / maxCount) * SCORE_DISTRIBUTION_BAR_WIDTH),\n );\n const bar = \"#\".repeat(barLength).padEnd(SCORE_DISTRIBUTION_BAR_WIDTH, \" \");\n return `${label.padEnd(7)} | ${bar} ${formatNumber(count)}`;\n}\n\nfunction renderFailureDetails(\n testCase: EvalCase,\n number: number,\n options: SummaryOptions,\n) {\n const failure = testCase.primaryFailure;\n const maxReasonChars = options.maxReasonChars ?? DEFAULT_MAX_REASON_CHARS;\n const maxOutputChars = options.maxOutputChars ?? DEFAULT_MAX_OUTPUT_CHARS;\n const maxToolCalls = options.maxToolCalls ?? DEFAULT_MAX_TOOL_CALLS;\n const usage = formatCaseUsage(testCase);\n const finalOutput = testCase.eval?.output ?? testCase.harness?.output;\n const summary = `${number}. ${testCase.displayName} - ${\n failure?.judgeName ?? \"failure\"\n } - ${formatScore(failure?.score ?? testCase.eval?.avgScore)}`;\n const lines = [\n \"<details>\",\n `<summary>${escapeHtml(summary)}</summary>`,\n \"\",\n \"```text\",\n ...renderFailureBlock(testCase, {\n finalOutput,\n maxOutputChars,\n maxReasonChars,\n maxToolCalls,\n number,\n usage,\n }).map(escapeFence),\n \"```\",\n \"\",\n \"</details>\",\n ];\n return lines;\n}\n\nfunction renderFailureBlock(\n testCase: EvalCase,\n {\n finalOutput,\n maxOutputChars,\n maxReasonChars,\n maxToolCalls,\n number,\n usage,\n }: {\n finalOutput: unknown;\n maxOutputChars: number;\n maxReasonChars: number;\n maxToolCalls: number;\n number: number;\n usage: string;\n },\n) {\n const failure = testCase.primaryFailure;\n const overviewRows: Array<[string, string]> = [\n [\"Case\", `${number}. ${testCase.displayName}`],\n [\"Status\", testCase.status],\n [\"Location\", formatLocation(testCase.displayFile, testCase.location)],\n [\"Harness\", testCase.harness?.name ?? \"n/a\"],\n [\"Score\", formatScore(failure?.score ?? testCase.eval?.avgScore)],\n [\"Judge\", failure?.judgeName ?? \"n/a\"],\n ];\n if (usage) {\n overviewRows.push([\"Usage\", usage]);\n }\n if (testCase.durationMs !== undefined) {\n overviewRows.push([\"Duration\", formatDuration(testCase.durationMs)]);\n }\n\n const lines = [\n ...renderAsciiSection(\"Result\", renderKeyValues(overviewRows)),\n \"\",\n ];\n\n if (failure?.reason) {\n lines.push(\n ...renderAsciiSection(\n \"Reason\",\n truncate(failure.reason, maxReasonChars).split(/\\r?\\n/),\n ),\n \"\",\n );\n }\n\n if (testCase.eval?.scores.length) {\n lines.push(\n ...renderAsciiTable(\n [\"Judge\", \"Score\"],\n testCase.eval.scores.map((score) => [\n score.name ?? \"Unknown\",\n formatScore(score.score),\n ]),\n ),\n \"\",\n );\n }\n\n if (finalOutput !== undefined) {\n lines.push(\n ...renderAsciiSection(\n \"Final Output\",\n stringifyValue(finalOutput, maxOutputChars).split(/\\r?\\n/),\n ),\n \"\",\n );\n }\n\n if (testCase.toolCalls.length) {\n const toolCalls = testCase.toolCalls.slice(0, maxToolCalls);\n lines.push(\n ...renderAsciiTable(\n [\"Tool\", \"Status\", \"Duration\"],\n toolCalls.map((toolCall) => [\n toolCall.name,\n toolCall.error ? `error: ${compactLine(toolCall.error, 120)}` : \"ok\",\n toolCall.durationMs === undefined\n ? \"n/a\"\n : formatDuration(toolCall.durationMs),\n ]),\n ),\n );\n if (testCase.toolCalls.length > maxToolCalls) {\n lines.push(\n `${testCase.toolCalls.length - maxToolCalls} more tool calls omitted`,\n );\n }\n lines.push(\"\");\n }\n\n if (testCase.harness?.errors.length) {\n lines.push(\n ...renderAsciiSection(\n \"Harness Errors\",\n stringifyValue(testCase.harness.errors, maxReasonChars).split(/\\r?\\n/),\n ),\n \"\",\n );\n }\n\n while (lines[lines.length - 1] === \"\") {\n lines.pop();\n }\n return lines;\n}\n\nfunction renderAsciiSection(title: string, content: string[]) {\n return [title, \"-\".repeat(title.length), ...content];\n}\n\nfunction renderKeyValues(rows: Array<[string, string]>) {\n const labelWidth = Math.max(...rows.map(([label]) => label.length));\n return rows.map(\n ([label, value]) =>\n `${label.padEnd(labelWidth)} ${compactLine(value, 500)}`,\n );\n}\n\nfunction renderAsciiTable(headers: string[], rows: string[][]) {\n const widths = headers.map((header, index) =>\n Math.max(header.length, ...rows.map((row) => row[index]?.length ?? 0)),\n );\n const renderRow = (row: string[]) =>\n row\n .map((cell, index) => cell.padEnd(widths[index] ?? cell.length))\n .join(\" \")\n .trimEnd();\n\n return [\n renderRow(headers),\n widths.map((width) => \"-\".repeat(width)).join(\" \"),\n ...rows.map(renderRow),\n ];\n}\n\nfunction formatCaseUsage(testCase: EvalCase) {\n const usage = testCase.harness?.usage;\n const parts: string[] = [];\n const totalTokens =\n usage?.totalTokens ??\n (usage?.inputTokens ?? 0) +\n (usage?.outputTokens ?? 0) +\n (usage?.reasoningTokens ?? 0);\n const toolCalls =\n usage?.toolCalls !== undefined\n ? Math.max(usage.toolCalls, testCase.toolCalls.length)\n : testCase.toolCalls.length;\n\n if (totalTokens > 0) {\n parts.push(`${formatNumber(totalTokens)} tokens`);\n }\n if (toolCalls > 0) {\n parts.push(`${formatNumber(toolCalls)} tool${toolCalls === 1 ? \"\" : \"s\"}`);\n }\n if (testCase.harness?.timingMs !== undefined) {\n parts.push(formatDuration(testCase.harness.timingMs));\n }\n\n return parts.join(\", \");\n}\n","import { buildCheckAnnotations } from \"./annotations\";\nimport { renderJobSummary, type SummaryOptions } from \"./summary\";\nimport type { EvalReport } from \"./types\";\n\n/** Options for creating or updating a GitHub Check Run. */\nexport type PublishCheckRunOptions = SummaryOptions & {\n token?: string;\n repository?: string;\n sha?: string;\n name?: string;\n apiUrl?: string;\n checkRunId?: number;\n maxAnnotations?: number;\n};\n\n/** Result of attempting to publish a GitHub Check Run. */\nexport type PublishCheckRunResult =\n | {\n status: \"skipped\";\n reason: string;\n }\n | {\n status: \"created\" | \"updated\";\n id?: number;\n htmlUrl?: string;\n };\n\nconst DEFAULT_CHECK_NAME = \"vitest-evals\";\nconst MAX_CHECK_SUMMARY_LENGTH = 64_000;\nconst CHECK_SUMMARY_TRUNCATION_SUFFIX =\n \"\\n\\n[truncated for GitHub Check Run]\\n\";\n\n/** Publishes the eval report to a GitHub Check Run when configuration allows it. */\nexport async function publishCheckRun(\n report: EvalReport,\n options: PublishCheckRunOptions = {},\n): Promise<PublishCheckRunResult> {\n const token = options.token ?? process.env.GITHUB_TOKEN;\n const repository = options.repository ?? process.env.GITHUB_REPOSITORY;\n const sha = options.sha ?? process.env.GITHUB_SHA;\n\n if (!token) {\n return { status: \"skipped\", reason: \"missing GITHUB_TOKEN\" };\n }\n if (!repository) {\n return { status: \"skipped\", reason: \"missing GITHUB_REPOSITORY\" };\n }\n if (!sha && options.checkRunId === undefined) {\n return { status: \"skipped\", reason: \"missing GITHUB_SHA\" };\n }\n\n const [owner, repo] = repository.split(\"/\");\n if (!owner || !repo) {\n return {\n status: \"skipped\",\n reason: `invalid GitHub repository: ${repository}`,\n };\n }\n\n const payload = buildCheckRunPayload(report, options);\n const apiUrl =\n options.apiUrl ?? process.env.GITHUB_API_URL ?? \"https://api.github.com\";\n const requestUrl =\n options.checkRunId === undefined\n ? `${apiUrl}/repos/${owner}/${repo}/check-runs`\n : `${apiUrl}/repos/${owner}/${repo}/check-runs/${options.checkRunId}`;\n const response = await fetch(requestUrl, {\n method: options.checkRunId === undefined ? \"POST\" : \"PATCH\",\n headers: {\n accept: \"application/vnd.github+json\",\n authorization: `Bearer ${token}`,\n \"content-type\": \"application/json\",\n \"x-github-api-version\": \"2022-11-28\",\n },\n body: JSON.stringify(\n options.checkRunId === undefined\n ? {\n name: options.name ?? DEFAULT_CHECK_NAME,\n head_sha: sha,\n ...payload,\n }\n : payload,\n ),\n });\n\n if (!response.ok) {\n const text = await response.text();\n throw new Error(\n `GitHub Check Run request failed: ${response.status} ${response.statusText} ${text}`.trim(),\n );\n }\n\n const data = (await response.json()) as {\n id?: number;\n html_url?: string;\n };\n\n return {\n status: options.checkRunId === undefined ? \"created\" : \"updated\",\n id: data.id,\n htmlUrl: data.html_url,\n };\n}\n\nfunction buildCheckRunPayload(\n report: EvalReport,\n options: PublishCheckRunOptions,\n) {\n const annotations = buildCheckAnnotations(report, {\n maxAnnotations: options.maxAnnotations,\n });\n const title =\n report.failures.length === 0 && report.status === \"passed\"\n ? \"No eval failures\"\n : report.failures.length === 0\n ? \"Vitest run failed\"\n : `${report.failures.length} eval failure${\n report.failures.length === 1 ? \"\" : \"s\"\n }`;\n\n return {\n status: \"completed\",\n conclusion: report.status === \"passed\" ? \"success\" : \"failure\",\n completed_at: new Date().toISOString(),\n output: {\n title,\n summary: truncateCheckSummary(\n renderJobSummary(report, {\n ...options,\n maxFailures: options.maxFailures ?? 5,\n maxReasonChars: options.maxReasonChars ?? 4000,\n maxOutputChars: options.maxOutputChars ?? 2000,\n maxToolCalls: options.maxToolCalls ?? 10,\n }),\n ),\n annotations,\n },\n };\n}\n\nfunction truncateCheckSummary(summary: string) {\n if (summary.length <= MAX_CHECK_SUMMARY_LENGTH) {\n return summary;\n }\n\n return `${summary\n .slice(0, MAX_CHECK_SUMMARY_LENGTH - CHECK_SUMMARY_TRUNCATION_SUFFIX.length)\n .trimEnd()}${CHECK_SUMMARY_TRUNCATION_SUFFIX}`;\n}\n"],"mappings":";AAAA;AAAA,EACE;AAAA,OAIK;;;ACLP,SAAS,aAAa;AAMf,SAAS,YAAY,OAAe,WAAmB;AAC5D,QAAM,OAAO,MACV,MAAM,OAAO,EACb,IAAI,CAAC,SAAS,KAAK,KAAK,CAAC,EACzB,KAAK,CAAC,SAAS,KAAK,SAAS,CAAC;AACjC,MAAI,CAAC,MAAM;AACT,WAAO;AAAA,EACT;AACA,SAAO,SAAS,MAAM,SAAS;AACjC;AAEO,SAAS,SAAS,OAAe,WAAmB;AACzD,MAAI,aAAa,GAAG;AAClB,WAAO;AAAA,EACT;AACA,MAAI,MAAM,UAAU,WAAW;AAC7B,WAAO;AAAA,EACT;AACA,SAAO,GAAG,MAAM,MAAM,GAAG,KAAK,IAAI,GAAG,YAAY,EAAE,CAAC,EAAE,QAAQ,CAAC;AACjE;AAEO,SAAS,eAAe,OAAgB,WAAmB;AAChE,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AACA,MAAI,OAAO,UAAU,UAAU;AAC7B,WAAO,SAAS,OAAO,SAAS;AAAA,EAClC;AACA,MAAI;AACF,WAAO,SAAS,KAAK,UAAU,OAAO,MAAM,CAAC,GAAG,SAAS;AAAA,EAC3D,QAAQ;AACN,WAAO,SAAS,OAAO,KAAK,GAAG,SAAS;AAAA,EAC1C;AACF;AAEO,SAAS,aAAa,OAAe;AAC1C,SAAO,IAAI,KAAK,aAAa,SAAS;AAAA,IACpC,uBAAuB;AAAA,EACzB,CAAC,EAAE,OAAO,KAAK;AACjB;AAEO,SAAS,YAAY,OAAkC;AAC5D,MAAI,SAAS,QAAQ,CAAC,OAAO,SAAS,KAAK,GAAG;AAC5C,WAAO;AAAA,EACT;AACA,SAAO,MAAM,QAAQ,CAAC;AACxB;AAEO,SAAS,eAAe,IAAwB;AACrD,MAAI,OAAO,UAAa,CAAC,OAAO,SAAS,EAAE,GAAG;AAC5C,WAAO;AAAA,EACT;AACA,MAAI,KAAK,KAAM;AACb,WAAO,GAAG,KAAK,MAAM,EAAE,CAAC;AAAA,EAC1B;AACA,QAAM,UAAU,KAAK;AACrB,MAAI,UAAU,IAAI;AAChB,WAAO,GAAG,QAAQ,QAAQ,UAAU,KAAK,IAAI,CAAC,CAAC;AAAA,EACjD;AACA,QAAM,eAAe,KAAK,MAAM,OAAO;AACvC,QAAM,UAAU,KAAK,MAAM,eAAe,EAAE;AAC5C,QAAM,mBAAmB,eAAe;AACxC,SAAO,GAAG,OAAO,KAAK,gBAAgB;AACxC;AAEO,SAAS,eACd,MACA,UACA;AACA,MAAI,CAAC,UAAU;AACb,WAAO;AAAA,EACT;AACA,SAAO,GAAG,IAAI,IAAI,SAAS,IAAI;AACjC;AAmBO,SAAS,YAAY,OAAe;AACzC,SAAO,MAAM,QAAQ,QAAQ,KAAK;AACpC;AAEO,SAAS,WAAW,OAAe;AACxC,SAAO,MACJ,QAAQ,MAAM,OAAO,EACrB,QAAQ,MAAM,MAAM,EACpB,QAAQ,MAAM,MAAM,EACpB,QAAQ,MAAM,QAAQ;AAC3B;AAEO,SAAS,kBAAkB,OAAe;AAC/C,SAAO,MAAM,QAAQ,MAAM,KAAK,EAAE,QAAQ,OAAO,KAAK,EAAE,QAAQ,OAAO,KAAK;AAC9E;AAEO,SAAS,sBAAsB,OAAe;AACnD,SAAO,kBAAkB,KAAK,EAAE,QAAQ,MAAM,KAAK,EAAE,QAAQ,MAAM,KAAK;AAC1E;;;ADjGO,SAAS,kBACd,OACA,UAA0B,CAAC,GACf;AACZ,QAAM,YAAY;AAAA,IAChB;AAAA,MACE,QAAQ;AAAA,IACV;AAAA,IACA;AAAA,MACE,WAAW,QAAQ;AAAA,IACrB;AAAA,EACF;AACA,QAAM,QAAQ,UAAU,MAAM,IAAI,eAAe;AACjD,QAAM,WAAW,MAAM,OAAO,CAAC,aAAa,SAAS,WAAW,QAAQ;AACxE,QAAM,aAAa,MAChB,IAAI,CAAC,aAAa,SAAS,MAAM,QAAQ,EACzC,OAAO,CAAC,UAA2B,eAAe,KAAK,CAAC;AAC3D,QAAM,QAAQ,SAAS,KAAK;AAC5B,QAAM,aAAa,UAAU,KAAK,CAAC,GAAG;AAEtC,SAAO;AAAA,IACL,QAAQ,MAAM,WAAW,SAAS,WAAW,IAAI,WAAW;AAAA,IAC5D,WAAW,MAAM;AAAA,IACjB;AAAA,IACA,QAAQ;AAAA,MACN,OAAO,MAAM;AAAA,MACb,QAAQ,MAAM;AAAA,MACd,QAAQ,MAAM;AAAA,MACd,SAAS,MAAM,kBAAkB,MAAM;AAAA,MACvC,WAAW,MAAM;AAAA,MACjB,YAAY,MAAM,OAAO,CAAC,aAAa,SAAS,WAAW,QAAQ,EAChE;AAAA,MACH,YAAY,SAAS;AAAA,IACvB;AAAA,IACA,OACE,WAAW,SAAS,IAChB;AAAA,MACE,SACE,WAAW,OAAO,CAAC,OAAO,UAAU,QAAQ,OAAO,CAAC,IACpD,WAAW;AAAA,MACb,SAAS,KAAK,IAAI,GAAG,UAAU;AAAA,IACjC,IACA;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,gBAAgB,YAAkC;AACzD,QAAM,UAAU,WAAW,MAAM,UAAU,CAAC,GAAG,IAAI,cAAc;AACjE,QAAM,aAAa,WAAW,SAAS;AACvC,QAAM,YAAY,iBAAiB,UAAU;AAC7C,QAAM,WAAqB;AAAA,IACzB,IAAI,WAAW;AAAA,IACf,MAAM,WAAW;AAAA,IACjB,aAAa,WAAW;AAAA,IACxB,OAAO,WAAW;AAAA,IAClB,aAAa,WAAW;AAAA,IACxB,QAAQ,WAAW;AAAA,IACnB,YAAY,YAAY,WAAW,UAAU;AAAA,IAC7C,UAAU,WAAW;AAAA,IACrB,iBAAiB,WAAW;AAAA,IAC5B;AAAA,IACA,MAAM,WAAW,OACb;AAAA,MACE,UAAU,YAAY,WAAW,KAAK,QAAQ;AAAA,MAC9C,iBAAiB,WAAW,KAAK;AAAA,MACjC,QAAQ,WAAW,KAAK;AAAA,MACxB;AAAA,IACF,IACA;AAAA,IACJ,SAAS,WAAW,UAChB;AAAA,MACE,MAAM,WAAW,QAAQ;AAAA,MACzB,QAAQ,YAAY;AAAA,MACpB,OAAO,YAAY;AAAA,MACnB,UAAU,YAAY,SAAS;AAAA,MAC/B;AAAA,MACA,QAAQ,YAAY,UAAU,CAAC;AAAA,IACjC,IACA;AAAA,EACN;AAEA,WAAS,iBAAiB,kBAAkB,QAAQ;AACpD,SAAO;AACT;AAEA,SAAS,eAAe,OAA6B;AACnD,SAAO;AAAA,IACL,GAAG;AAAA,IACH,OAAO,YAAY,MAAM,KAAK,KAAK;AAAA,EACrC;AACF;AAEA,SAAS,YAAY,OAAgB;AACnC,SAAO,eAAe,KAAK,IAAI,QAAQ;AACzC;AAEA,SAAS,eAAe,OAAiC;AACvD,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,iBAAiB,YAAwB;AAChD,QAAM,eAAe;AAAA,IACnB,WAAW,SAAS,KAAK;AAAA,EAC3B;AACA,SAAO,aAAa,SAAS,IACzB,gBACC,WAAW,MAAM,aAAa,CAAC,GAAG,IAAI,iBAAiB;AAC9D;AAEA,SAAS,wBAAwB,SAA4C;AAC3E,QAAM,WAAW,SAAS,YAAY,CAAC;AACvC,QAAM,YAA+B,CAAC;AAEtC,aAAW,WAAW,UAAU;AAC9B,QAAI,CAAC,MAAM,QAAQ,QAAQ,SAAS,GAAG;AACrC;AAAA,IACF;AAEA,eAAW,QAAQ,QAAQ,WAAW;AACpC,gBAAU,KAAK,kBAAkB,IAAI,CAAC;AAAA,IACxC;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,kBAAkB,MAAuC;AAChE,SAAO;AAAA,IACL,MAAM,KAAK;AAAA,IACX,OAAO,iBAAiB,KAAK,KAAK;AAAA,IAClC,YAAY,YAAY,KAAK,UAAU;AAAA,EACzC;AACF;AAEA,SAAS,iBAAiB,OAAgB;AACxC,QAAM,QAAQ;AACd,MACE,SACA,OAAO,UAAU,YACjB,CAAC,MAAM,QAAQ,KAAK,KACpB,OAAO,MAAM,YAAY,UACzB;AACA,WAAO,MAAM;AAAA,EACf;AACA,MAAI,UAAU,QAAW;AACvB,WAAO,eAAe,OAAO,GAAG;AAAA,EAClC;AACA,SAAO;AACT;AAEA,SAAS,kBAAkB,UAA6C;AACtE,QAAM,gBAAgB,CAAC,GAAI,SAAS,MAAM,UAAU,CAAC,CAAE,EACpD;AAAA,IACC,CAACA,YACEA,OAAM,SAAS,KAAK,KACrBA,OAAM,UAAU,cAAc,UAC9BA,OAAM,UAAU,WAAW;AAAA,EAC/B,EACC,KAAK,CAAC,MAAM,WAAW,KAAK,SAAS,MAAM,MAAM,SAAS,EAAE;AAC/D,QAAM,UAAU,cAAc,CAAC;AAC/B,QAAM,QACJ,OAAO,SAAS,UAAU,WACtB,QAAQ,QACR,SAAS,MAAM;AACrB,QAAM,SACJ,gBAAgB,SAAS,UAAU,SAAS,KAC5C,YAAY,SAAS,gBAAgB,KAAK,IAAI,GAAG,GAAG;AAEtD,MAAI,CAAC,WAAW,CAAC,UAAU,UAAU,QAAW;AAC9C,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,WAAW,SAAS;AAAA,IACpB;AAAA,IACA,QAAQ,UAAU;AAAA,EACpB;AACF;AAEA,SAAS,gBAAgB,OAAgB;AACvC,MAAI,UAAU,QAAW;AACvB,WAAO;AAAA,EACT;AACA,SAAO,OAAO,UAAU,WAAW,QAAQ,eAAe,OAAO,GAAI;AACvE;AAEA,SAAS,SAAS,OAAmB;AACnC,QAAM,QAAgC;AAAA,IACpC,aAAa;AAAA,IACb,cAAc;AAAA,IACd,iBAAiB;AAAA,IACjB,aAAa;AAAA,IACb,WAAW;AAAA,EACb;AAEA,aAAW,YAAY,OAAO;AAC5B,UAAM,YAAY,SAAS,SAAS;AACpC,UAAM,eAAe,WAAW,eAAe;AAC/C,UAAM,gBAAgB,WAAW,gBAAgB;AACjD,UAAM,mBAAmB,WAAW,mBAAmB;AACvD,UAAM,eACJ,WAAW,gBACV,WAAW,eAAe,MACxB,WAAW,gBAAgB,MAC3B,WAAW,mBAAmB;AACnC,UAAM,aAAa,cAAc,QAAQ;AAAA,EAC3C;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,UAAoB;AACzC,QAAM,iBAAiB,SAAS,SAAS,OAAO;AAChD,MAAI,mBAAmB,QAAW;AAChC,WAAO,KAAK,IAAI,gBAAgB,SAAS,UAAU,MAAM;AAAA,EAC3D;AAEA,SAAO,SAAS,UAAU;AAC5B;;;AEtNA,IAAM,mCAAmC;AACzC,IAAM,gCAAgC;AACtC,IAAM,yBAAyB;AAOxB,SAAS,uBACd,QACA,UAA6B,CAAC,GAC9B;AACA,QAAM,iBACJ,QAAQ,kBAAkB;AAE5B,SAAO,OAAO,SACX,OAAO,qBAAqB,EAC5B,MAAM,GAAG,cAAc,EACvB;AAAA,IAAI,CAAC,aACJ,sBAAsB;AAAA,MACpB,SAAS;AAAA,MACT,YAAY;AAAA,QACV,MAAM,SAAS;AAAA,QACf,MAAM,OAAO,SAAS,SAAS,IAAI;AAAA,QACnC,KAAK,OAAO,SAAS,SAAS,MAAM;AAAA,QACpC,OAAO;AAAA,MACT;AAAA,MACA,SAAS,sBAAsB,QAAQ;AAAA,IACzC,CAAC;AAAA,EACH;AACJ;AAGO,SAAS,sBACd,QACA,UAA6B,CAAC,GACX;AACnB,QAAM,iBAAiB,KAAK;AAAA,IAC1B,QAAQ,kBAAkB;AAAA,IAC1B;AAAA,EACF;AAEA,SAAO,OAAO,SACX,OAAO,qBAAqB,EAC5B,MAAM,GAAG,cAAc,EACvB,IAAI,CAAC,cAAc;AAAA,IAClB,MAAM,SAAS;AAAA,IACf,YAAY,SAAS,SAAS;AAAA,IAC9B,UAAU,SAAS,SAAS;AAAA,IAC5B,kBAAkB;AAAA,IAClB,OAAO;AAAA,MACL,GAAG,SAAS,gBAAgB,aAAa,cAAc,MAAM,SAAS,WAAW;AAAA,MACjF;AAAA,IACF;AAAA,IACA,SAAS;AAAA,MACP,sBAAsB,QAAQ;AAAA,MAC9B;AAAA,IACF;AAAA,IACA,aAAa,SAAS,iBAAiB,QAAQ,GAAG,sBAAsB;AAAA,EAC1E,EAAE;AACN;AAEA,SAAS,sBACP,UAC+B;AAC/B,SAAO,QAAQ,SAAS,QAAQ;AAClC;AAEA,SAAS,sBAAsB,UAAoB;AACjD,QAAM,UAAU,SAAS;AACzB,QAAM,QAAQ;AAAA,IACZ,SAAS;AAAA,IACT,SAAS,YAAY,SAAS,SAAS,SAAS,MAAM,QAAQ,CAAC;AAAA,EACjE;AAEA,MAAI,SAAS,WAAW;AACtB,UAAM,KAAK,QAAQ,SAAS;AAAA,EAC9B;AAEA,QAAM,SAAS,YAAY,SAAS,UAAU,IAAI,GAAG;AACrD,MAAI,QAAQ;AACV,UAAM,KAAK,MAAM;AAAA,EACnB;AAEA,SAAO,MAAM,KAAK,KAAK;AACzB;AAEA,SAAS,iBAAiB,UAA6B;AACrD,QAAM,QAAQ;AAAA,IACZ,SAAS,SAAS,WAAW;AAAA,IAC7B,aAAa,SAAS,WAAW,IAAI,SAAS,SAAS,IAAI;AAAA,IAC3D,YAAY,SAAS,SAAS,QAAQ,KAAK;AAAA,IAC3C,UAAU,YAAY,SAAS,gBAAgB,SAAS,SAAS,MAAM,QAAQ,CAAC;AAAA,IAChF,UAAU,SAAS,gBAAgB,aAAa,KAAK;AAAA,IACrD;AAAA,IACA;AAAA,IACA,SAAS,gBAAgB,UAAU;AAAA,EACrC;AAEA,QAAM,cAAc,SAAS,MAAM,UAAU,SAAS,SAAS;AAC/D,MAAI,gBAAgB,QAAW;AAC7B,UAAM,KAAK,IAAI,UAAU,eAAe,aAAa,GAAI,CAAC;AAAA,EAC5D;AAEA,MAAI,SAAS,UAAU,QAAQ;AAC7B,UAAM,KAAK,IAAI,QAAQ;AACvB,eAAW,YAAY,SAAS,WAAW;AACzC,YAAM;AAAA,QACJ,KAAK,SAAS,IAAI,KAAK,SAAS,QAAQ,UAAU,SAAS,KAAK,KAAK,IAAI;AAAA,MAC3E;AAAA,IACF;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,sBAAsB;AAAA,EAC7B;AAAA,EACA;AAAA,EACA;AACF,GAIG;AACD,QAAM,qBAAqB,OAAO,QAAQ,UAAU,EACjD,IAAI,CAAC,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,IAAI,sBAAsB,KAAK,CAAC,EAAE,EAC9D,KAAK,GAAG;AACX,SAAO,KAAK,OAAO,IAAI,kBAAkB,KAAK,kBAAkB,OAAO,CAAC;AAC1E;;;ACvIA,IAAM,uBAAuB;AAC7B,IAAM,2BAA2B;AACjC,IAAM,2BAA2B;AACjC,IAAM,yBAAyB;AAC/B,IAAM,6BAA6B;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AACA,IAAM,+BAA+B;AAG9B,SAAS,iBACd,QACA,UAA0B,CAAC,GAC3B;AACA,QAAM,cAAc,QAAQ,eAAe;AAC3C,QAAM,WAAW,OAAO,SAAS,MAAM,GAAG,WAAW;AACrD,QAAM,kBAAkB,OAAO,OAAO,SAAS,OAAO,OAAO;AAC7D,QAAM,QAAkB;AAAA,IACtB;AAAA,IACA;AAAA,IACA,GAAG,mBAAmB,QAAQ,eAAe;AAAA,IAC7C;AAAA,IACA,GAAG,wBAAwB,MAAM;AAAA,IACjC;AAAA,IACA;AAAA,EACF;AAEA,MAAI,OAAO,SAAS,SAAS,GAAG;AAC9B,UAAM,KAAK,gBAAgB,EAAE;AAC7B,aAAS,QAAQ,CAAC,UAAU,UAAU;AACpC,YAAM,KAAK,GAAG,qBAAqB,UAAU,QAAQ,GAAG,OAAO,GAAG,EAAE;AAAA,IACtE,CAAC;AAED,QAAI,OAAO,SAAS,SAAS,SAAS,QAAQ;AAC5C,YAAM;AAAA,QACJ,GAAG,OAAO,SAAS,SAAS,SAAS,MAAM;AAAA,QAC3C;AAAA,MACF;AAAA,IACF;AAAA,EACF,WAAW,OAAO,OAAO,YAAY,GAAG;AACtC,UAAM,KAAK,gBAAgB,IAAI,qBAAqB,EAAE;AAAA,EACxD;AAEA,MAAI,OAAO,OAAO,cAAc,GAAG;AACjC,UAAM,KAAK,yDAAyD,EAAE;AAAA,EACxE;AAEA,SAAO,GAAG,MAAM,KAAK,IAAI,CAAC;AAAA;AAC5B;AAEA,SAAS,gBAAgB,QAAgB,QAAgB,OAAe;AACtE,SAAO,GAAG,aAAa,MAAM,CAAC,YAAY,aAAa,MAAM,CAAC,YAAY,aAAa,KAAK,CAAC;AAC/F;AAEA,SAAS,mBAAmB,QAAoB,iBAAyB;AACvE,QAAM,OAAgC;AAAA,IACpC,CAAC,UAAU,OAAO,MAAM;AAAA,IACxB;AAAA,MACE;AAAA,MACA;AAAA,QACE,OAAO,OAAO;AAAA,QACd,OAAO,OAAO;AAAA,QACd,OAAO,OAAO;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AAEA,MAAI,OAAO,OAAO;AAChB,SAAK,KAAK,CAAC,SAAS,mBAAmB,OAAO,KAAK,CAAC,CAAC;AAAA,EACvD;AAEA,MAAI,kBAAkB,GAAG;AACvB,SAAK,KAAK;AAAA,MACR;AAAA,MACA,GAAG,aAAa,eAAe,CAAC,yBAC9B,oBAAoB,IAAI,KAAK,GAC/B;AAAA,IACF,CAAC;AAAA,EACH;AAEA,OAAK,KAAK,CAAC,YAAY,eAAe,OAAO,UAAU,CAAC,CAAC;AAEzD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,GAAG,KAAK;AAAA,MACN,CAAC,CAAC,QAAQ,KAAK,MACb,KAAK,gBAAgB,MAAM,CAAC,MAAM,gBAAgB,KAAK,CAAC;AAAA,IAC5D;AAAA,EACF;AACF;AAEA,SAAS,mBAAmB,OAAyC;AACnE,SAAO,OAAO,YAAY,MAAM,OAAO,CAAC,GACtC,MAAM,YAAY,SAAY,KAAK,SAAS,YAAY,MAAM,OAAO,CAAC,EACxE;AACF;AAEA,SAAS,gBAAgB,OAAe;AACtC,SAAO,MACJ,QAAQ,UAAU,GAAG,EACrB,QAAQ,OAAO,MAAM,EACrB,QAAQ,OAAO,KAAK;AACzB;AAEA,SAAS,wBAAwB,QAAoB;AACnD,QAAM,SAAS,OAAO,MACnB,IAAI,CAAC,aAAa,SAAS,MAAM,QAAQ,EACzC;AAAA,IACC,CAAC,UACC,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAAA,EACtD;AACF,MAAI,OAAO,WAAW,GAAG;AACvB,WAAO,CAAC;AAAA,EACV;AAEA,QAAM,SAAS,2BAA2B,IAAI,MAAM,CAAC;AACrD,aAAW,SAAS,QAAQ;AAC1B,UAAM,SAAS,KAAK;AAAA,MAClB,2BAA2B,SAAS;AAAA,MACpC,KAAK,IAAI,GAAG,KAAK,MAAM,QAAQ,2BAA2B,MAAM,CAAC;AAAA,IACnE;AACA,WAAO,MAAM,KAAK,OAAO,MAAM,KAAK,KAAK;AAAA,EAC3C;AAEA,QAAM,WAAW,KAAK,IAAI,GAAG,MAAM;AACnC,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG,2BAA2B;AAAA,MAAI,CAAC,OAAO,UACxC,8BAA8B,OAAO,OAAO,KAAK,KAAK,GAAG,QAAQ;AAAA,IACnE;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,8BACP,OACA,OACA,UACA;AACA,QAAM,YACJ,UAAU,IACN,IACA,KAAK;AAAA,IACH;AAAA,IACA,KAAK,MAAO,QAAQ,WAAY,4BAA4B;AAAA,EAC9D;AACN,QAAM,MAAM,IAAI,OAAO,SAAS,EAAE,OAAO,8BAA8B,GAAG;AAC1E,SAAO,GAAG,MAAM,OAAO,CAAC,CAAC,MAAM,GAAG,IAAI,aAAa,KAAK,CAAC;AAC3D;AAEA,SAAS,qBACP,UACA,QACA,SACA;AACA,QAAM,UAAU,SAAS;AACzB,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,eAAe,QAAQ,gBAAgB;AAC7C,QAAM,QAAQ,gBAAgB,QAAQ;AACtC,QAAM,cAAc,SAAS,MAAM,UAAU,SAAS,SAAS;AAC/D,QAAM,UAAU,GAAG,MAAM,KAAK,SAAS,WAAW,MAChD,SAAS,aAAa,SACxB,MAAM,YAAY,SAAS,SAAS,SAAS,MAAM,QAAQ,CAAC;AAC5D,QAAM,QAAQ;AAAA,IACZ;AAAA,IACA,YAAY,WAAW,OAAO,CAAC;AAAA,IAC/B;AAAA,IACA;AAAA,IACA,GAAG,mBAAmB,UAAU;AAAA,MAC9B;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC,EAAE,IAAI,WAAW;AAAA,IAClB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,mBACP,UACA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAQA;AACA,QAAM,UAAU,SAAS;AACzB,QAAM,eAAwC;AAAA,IAC5C,CAAC,QAAQ,GAAG,MAAM,KAAK,SAAS,WAAW,EAAE;AAAA,IAC7C,CAAC,UAAU,SAAS,MAAM;AAAA,IAC1B,CAAC,YAAY,eAAe,SAAS,aAAa,SAAS,QAAQ,CAAC;AAAA,IACpE,CAAC,WAAW,SAAS,SAAS,QAAQ,KAAK;AAAA,IAC3C,CAAC,SAAS,YAAY,SAAS,SAAS,SAAS,MAAM,QAAQ,CAAC;AAAA,IAChE,CAAC,SAAS,SAAS,aAAa,KAAK;AAAA,EACvC;AACA,MAAI,OAAO;AACT,iBAAa,KAAK,CAAC,SAAS,KAAK,CAAC;AAAA,EACpC;AACA,MAAI,SAAS,eAAe,QAAW;AACrC,iBAAa,KAAK,CAAC,YAAY,eAAe,SAAS,UAAU,CAAC,CAAC;AAAA,EACrE;AAEA,QAAM,QAAQ;AAAA,IACZ,GAAG,mBAAmB,UAAU,gBAAgB,YAAY,CAAC;AAAA,IAC7D;AAAA,EACF;AAEA,MAAI,SAAS,QAAQ;AACnB,UAAM;AAAA,MACJ,GAAG;AAAA,QACD;AAAA,QACA,SAAS,QAAQ,QAAQ,cAAc,EAAE,MAAM,OAAO;AAAA,MACxD;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,MAAI,SAAS,MAAM,OAAO,QAAQ;AAChC,UAAM;AAAA,MACJ,GAAG;AAAA,QACD,CAAC,SAAS,OAAO;AAAA,QACjB,SAAS,KAAK,OAAO,IAAI,CAAC,UAAU;AAAA,UAClC,MAAM,QAAQ;AAAA,UACd,YAAY,MAAM,KAAK;AAAA,QACzB,CAAC;AAAA,MACH;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,MAAI,gBAAgB,QAAW;AAC7B,UAAM;AAAA,MACJ,GAAG;AAAA,QACD;AAAA,QACA,eAAe,aAAa,cAAc,EAAE,MAAM,OAAO;AAAA,MAC3D;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,MAAI,SAAS,UAAU,QAAQ;AAC7B,UAAM,YAAY,SAAS,UAAU,MAAM,GAAG,YAAY;AAC1D,UAAM;AAAA,MACJ,GAAG;AAAA,QACD,CAAC,QAAQ,UAAU,UAAU;AAAA,QAC7B,UAAU,IAAI,CAAC,aAAa;AAAA,UAC1B,SAAS;AAAA,UACT,SAAS,QAAQ,UAAU,YAAY,SAAS,OAAO,GAAG,CAAC,KAAK;AAAA,UAChE,SAAS,eAAe,SACpB,QACA,eAAe,SAAS,UAAU;AAAA,QACxC,CAAC;AAAA,MACH;AAAA,IACF;AACA,QAAI,SAAS,UAAU,SAAS,cAAc;AAC5C,YAAM;AAAA,QACJ,GAAG,SAAS,UAAU,SAAS,YAAY;AAAA,MAC7C;AAAA,IACF;AACA,UAAM,KAAK,EAAE;AAAA,EACf;AAEA,MAAI,SAAS,SAAS,OAAO,QAAQ;AACnC,UAAM;AAAA,MACJ,GAAG;AAAA,QACD;AAAA,QACA,eAAe,SAAS,QAAQ,QAAQ,cAAc,EAAE,MAAM,OAAO;AAAA,MACvE;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAEA,SAAO,MAAM,MAAM,SAAS,CAAC,MAAM,IAAI;AACrC,UAAM,IAAI;AAAA,EACZ;AACA,SAAO;AACT;AAEA,SAAS,mBAAmB,OAAe,SAAmB;AAC5D,SAAO,CAAC,OAAO,IAAI,OAAO,MAAM,MAAM,GAAG,GAAG,OAAO;AACrD;AAEA,SAAS,gBAAgB,MAA+B;AACtD,QAAM,aAAa,KAAK,IAAI,GAAG,KAAK,IAAI,CAAC,CAAC,KAAK,MAAM,MAAM,MAAM,CAAC;AAClE,SAAO,KAAK;AAAA,IACV,CAAC,CAAC,OAAO,KAAK,MACZ,GAAG,MAAM,OAAO,UAAU,CAAC,KAAK,YAAY,OAAO,GAAG,CAAC;AAAA,EAC3D;AACF;AAEA,SAAS,iBAAiB,SAAmB,MAAkB;AAC7D,QAAM,SAAS,QAAQ;AAAA,IAAI,CAAC,QAAQ,UAClC,KAAK,IAAI,OAAO,QAAQ,GAAG,KAAK,IAAI,CAAC,QAAQ,IAAI,KAAK,GAAG,UAAU,CAAC,CAAC;AAAA,EACvE;AACA,QAAM,YAAY,CAAC,QACjB,IACG,IAAI,CAAC,MAAM,UAAU,KAAK,OAAO,OAAO,KAAK,KAAK,KAAK,MAAM,CAAC,EAC9D,KAAK,IAAI,EACT,QAAQ;AAEb,SAAO;AAAA,IACL,UAAU,OAAO;AAAA,IACjB,OAAO,IAAI,CAAC,UAAU,IAAI,OAAO,KAAK,CAAC,EAAE,KAAK,IAAI;AAAA,IAClD,GAAG,KAAK,IAAI,SAAS;AAAA,EACvB;AACF;AAEA,SAAS,gBAAgB,UAAoB;AAC3C,QAAM,QAAQ,SAAS,SAAS;AAChC,QAAM,QAAkB,CAAC;AACzB,QAAM,cACJ,OAAO,gBACN,OAAO,eAAe,MACpB,OAAO,gBAAgB,MACvB,OAAO,mBAAmB;AAC/B,QAAM,YACJ,OAAO,cAAc,SACjB,KAAK,IAAI,MAAM,WAAW,SAAS,UAAU,MAAM,IACnD,SAAS,UAAU;AAEzB,MAAI,cAAc,GAAG;AACnB,UAAM,KAAK,GAAG,aAAa,WAAW,CAAC,SAAS;AAAA,EAClD;AACA,MAAI,YAAY,GAAG;AACjB,UAAM,KAAK,GAAG,aAAa,SAAS,CAAC,QAAQ,cAAc,IAAI,KAAK,GAAG,EAAE;AAAA,EAC3E;AACA,MAAI,SAAS,SAAS,aAAa,QAAW;AAC5C,UAAM,KAAK,eAAe,SAAS,QAAQ,QAAQ,CAAC;AAAA,EACtD;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;;;AC7VA,IAAM,qBAAqB;AAC3B,IAAM,2BAA2B;AACjC,IAAM,kCACJ;AAGF,eAAsB,gBACpB,QACA,UAAkC,CAAC,GACH;AAChC,QAAM,QAAQ,QAAQ,SAAS,QAAQ,IAAI;AAC3C,QAAM,aAAa,QAAQ,cAAc,QAAQ,IAAI;AACrD,QAAM,MAAM,QAAQ,OAAO,QAAQ,IAAI;AAEvC,MAAI,CAAC,OAAO;AACV,WAAO,EAAE,QAAQ,WAAW,QAAQ,uBAAuB;AAAA,EAC7D;AACA,MAAI,CAAC,YAAY;AACf,WAAO,EAAE,QAAQ,WAAW,QAAQ,4BAA4B;AAAA,EAClE;AACA,MAAI,CAAC,OAAO,QAAQ,eAAe,QAAW;AAC5C,WAAO,EAAE,QAAQ,WAAW,QAAQ,qBAAqB;AAAA,EAC3D;AAEA,QAAM,CAAC,OAAO,IAAI,IAAI,WAAW,MAAM,GAAG;AAC1C,MAAI,CAAC,SAAS,CAAC,MAAM;AACnB,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,QAAQ,8BAA8B,UAAU;AAAA,IAClD;AAAA,EACF;AAEA,QAAM,UAAU,qBAAqB,QAAQ,OAAO;AACpD,QAAM,SACJ,QAAQ,UAAU,QAAQ,IAAI,kBAAkB;AAClD,QAAM,aACJ,QAAQ,eAAe,SACnB,GAAG,MAAM,UAAU,KAAK,IAAI,IAAI,gBAChC,GAAG,MAAM,UAAU,KAAK,IAAI,IAAI,eAAe,QAAQ,UAAU;AACvE,QAAM,WAAW,MAAM,MAAM,YAAY;AAAA,IACvC,QAAQ,QAAQ,eAAe,SAAY,SAAS;AAAA,IACpD,SAAS;AAAA,MACP,QAAQ;AAAA,MACR,eAAe,UAAU,KAAK;AAAA,MAC9B,gBAAgB;AAAA,MAChB,wBAAwB;AAAA,IAC1B;AAAA,IACA,MAAM,KAAK;AAAA,MACT,QAAQ,eAAe,SACnB;AAAA,QACE,MAAM,QAAQ,QAAQ;AAAA,QACtB,UAAU;AAAA,QACV,GAAG;AAAA,MACL,IACA;AAAA,IACN;AAAA,EACF,CAAC;AAED,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,OAAO,MAAM,SAAS,KAAK;AACjC,UAAM,IAAI;AAAA,MACR,oCAAoC,SAAS,MAAM,IAAI,SAAS,UAAU,IAAI,IAAI,GAAG,KAAK;AAAA,IAC5F;AAAA,EACF;AAEA,QAAM,OAAQ,MAAM,SAAS,KAAK;AAKlC,SAAO;AAAA,IACL,QAAQ,QAAQ,eAAe,SAAY,YAAY;AAAA,IACvD,IAAI,KAAK;AAAA,IACT,SAAS,KAAK;AAAA,EAChB;AACF;AAEA,SAAS,qBACP,QACA,SACA;AACA,QAAM,cAAc,sBAAsB,QAAQ;AAAA,IAChD,gBAAgB,QAAQ;AAAA,EAC1B,CAAC;AACD,QAAM,QACJ,OAAO,SAAS,WAAW,KAAK,OAAO,WAAW,WAC9C,qBACA,OAAO,SAAS,WAAW,IACzB,sBACA,GAAG,OAAO,SAAS,MAAM,gBACvB,OAAO,SAAS,WAAW,IAAI,KAAK,GACtC;AAER,SAAO;AAAA,IACL,QAAQ;AAAA,IACR,YAAY,OAAO,WAAW,WAAW,YAAY;AAAA,IACrD,eAAc,oBAAI,KAAK,GAAE,YAAY;AAAA,IACrC,QAAQ;AAAA,MACN;AAAA,MACA,SAAS;AAAA,QACP,iBAAiB,QAAQ;AAAA,UACvB,GAAG;AAAA,UACH,aAAa,QAAQ,eAAe;AAAA,UACpC,gBAAgB,QAAQ,kBAAkB;AAAA,UAC1C,gBAAgB,QAAQ,kBAAkB;AAAA,UAC1C,cAAc,QAAQ,gBAAgB;AAAA,QACxC,CAAC;AAAA,MACH;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,qBAAqB,SAAiB;AAC7C,MAAI,QAAQ,UAAU,0BAA0B;AAC9C,WAAO;AAAA,EACT;AAEA,SAAO,GAAG,QACP,MAAM,GAAG,2BAA2B,gCAAgC,MAAM,EAC1E,QAAQ,CAAC,GAAG,+BAA+B;AAChD;","names":["score"]}
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@vitest-evals/github-reporter",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.12.0",
|
|
4
4
|
"description": "GitHub Actions reporting internals for vitest-evals runs.",
|
|
5
5
|
"repository": {
|
|
6
6
|
"type": "git",
|
|
@@ -34,6 +34,9 @@
|
|
|
34
34
|
"import": "./dist/index.mjs"
|
|
35
35
|
}
|
|
36
36
|
},
|
|
37
|
+
"dependencies": {
|
|
38
|
+
"@vitest-evals/core": "0.12.0"
|
|
39
|
+
},
|
|
37
40
|
"peerDependencies": {
|
|
38
41
|
"vitest": ">=4 <5"
|
|
39
42
|
},
|