@yasserkhanorg/e2e-agents 1.9.5 → 1.10.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/cli/commands/gate.d.ts.map +1 -1
- package/dist/cli/commands/gate.js +4 -1
- package/dist/cli/commands/impact.d.ts.map +1 -1
- package/dist/cli/commands/impact.js +7 -0
- package/dist/cli/commands/llm_health.d.ts.map +1 -1
- package/dist/cli/commands/llm_health.js +57 -17
- package/dist/cli/usage.d.ts.map +1 -1
- package/dist/cli/usage.js +4 -0
- package/dist/cli.js +1 -1
- package/dist/esm/cli/commands/gate.js +4 -1
- package/dist/esm/cli/commands/impact.js +7 -0
- package/dist/esm/cli/commands/llm_health.js +57 -17
- package/dist/esm/cli/usage.js +4 -0
- package/dist/esm/cli.js +1 -1
- package/dist/esm/knowledge/api_surface.js +265 -34
- package/dist/esm/knowledge/failure_history.js +121 -0
- package/dist/esm/knowledge/route_families.js +31 -1
- package/dist/esm/pipeline/stage1_impact.js +19 -3
- package/dist/esm/pipeline/stage2_coverage.js +28 -7
- package/dist/esm/pipeline/stage3_generation.js +20 -1
- package/dist/esm/prompts/coverage.js +10 -0
- package/dist/esm/prompts/generation.js +41 -7
- package/dist/esm/validation/guardrails.js +5 -0
- package/dist/index.d.ts +1 -1
- package/dist/knowledge/api_surface.d.ts +12 -0
- package/dist/knowledge/api_surface.d.ts.map +1 -1
- package/dist/knowledge/api_surface.js +268 -34
- package/dist/knowledge/failure_history.d.ts +39 -0
- package/dist/knowledge/failure_history.d.ts.map +1 -0
- package/dist/knowledge/failure_history.js +128 -0
- package/dist/knowledge/route_families.d.ts +11 -0
- package/dist/knowledge/route_families.d.ts.map +1 -1
- package/dist/knowledge/route_families.js +32 -1
- package/dist/pipeline/stage1_impact.d.ts +1 -1
- package/dist/pipeline/stage1_impact.d.ts.map +1 -1
- package/dist/pipeline/stage1_impact.js +18 -2
- package/dist/pipeline/stage2_coverage.d.ts.map +1 -1
- package/dist/pipeline/stage2_coverage.js +28 -7
- package/dist/pipeline/stage3_generation.d.ts.map +1 -1
- package/dist/pipeline/stage3_generation.js +20 -1
- package/dist/prompts/coverage.d.ts.map +1 -1
- package/dist/prompts/coverage.js +10 -0
- package/dist/prompts/generation.d.ts +1 -1
- package/dist/prompts/generation.d.ts.map +1 -1
- package/dist/prompts/generation.js +41 -7
- package/dist/validation/guardrails.d.ts +2 -0
- package/dist/validation/guardrails.d.ts.map +1 -1
- package/dist/validation/guardrails.js +5 -0
- package/dist/validation/output_schema.d.ts +3 -0
- package/dist/validation/output_schema.d.ts.map +1 -1
- package/package.json +1 -1
|
@@ -2,39 +2,233 @@
|
|
|
2
2
|
// See LICENSE.txt for license information.
|
|
3
3
|
import { existsSync, readFileSync, readdirSync, writeFileSync } from 'fs';
|
|
4
4
|
import { join, extname } from 'path';
|
|
5
|
+
import ts from 'typescript';
|
|
6
|
+
// ── TypeScript AST-based extraction ────────────────────────
|
|
7
|
+
function extractMethodsFromAST(sourceFile, checker) {
|
|
8
|
+
const surfaces = [];
|
|
9
|
+
ts.forEachChild(sourceFile, (node) => {
|
|
10
|
+
if (!ts.isClassDeclaration(node) || !node.name) {
|
|
11
|
+
return;
|
|
12
|
+
}
|
|
13
|
+
const className = node.name.text;
|
|
14
|
+
const methods = [];
|
|
15
|
+
const seen = new Set();
|
|
16
|
+
// Get base class name if extends
|
|
17
|
+
let extendsName;
|
|
18
|
+
if (node.heritageClauses) {
|
|
19
|
+
for (const clause of node.heritageClauses) {
|
|
20
|
+
if (clause.token === ts.SyntaxKind.ExtendsKeyword && clause.types.length > 0) {
|
|
21
|
+
const expr = clause.types[0].expression;
|
|
22
|
+
if (ts.isIdentifier(expr)) {
|
|
23
|
+
extendsName = expr.text;
|
|
24
|
+
}
|
|
25
|
+
}
|
|
26
|
+
}
|
|
27
|
+
}
|
|
28
|
+
for (const member of node.members) {
|
|
29
|
+
// Skip constructor
|
|
30
|
+
if (ts.isConstructorDeclaration(member)) {
|
|
31
|
+
continue;
|
|
32
|
+
}
|
|
33
|
+
// Skip private/protected members
|
|
34
|
+
const modifiers = ts.canHaveModifiers(member) ? ts.getModifiers(member) : undefined;
|
|
35
|
+
if (modifiers?.some((m) => m.kind === ts.SyntaxKind.PrivateKeyword || m.kind === ts.SyntaxKind.ProtectedKeyword)) {
|
|
36
|
+
continue;
|
|
37
|
+
}
|
|
38
|
+
const name = member.name && ts.isIdentifier(member.name) ? member.name.text : null;
|
|
39
|
+
if (!name || name.startsWith('_') || seen.has(name)) {
|
|
40
|
+
continue;
|
|
41
|
+
}
|
|
42
|
+
seen.add(name);
|
|
43
|
+
if (ts.isMethodDeclaration(member)) {
|
|
44
|
+
const isAsync = modifiers?.some((m) => m.kind === ts.SyntaxKind.AsyncKeyword) ?? false;
|
|
45
|
+
const params = extractParams(member.parameters, checker);
|
|
46
|
+
const returnType = extractReturnType(member, checker);
|
|
47
|
+
methods.push({
|
|
48
|
+
name,
|
|
49
|
+
kind: 'method',
|
|
50
|
+
async: isAsync ? true : undefined,
|
|
51
|
+
params: params.length > 0 ? params : undefined,
|
|
52
|
+
returnType: returnType || undefined,
|
|
53
|
+
});
|
|
54
|
+
}
|
|
55
|
+
else if (ts.isGetAccessorDeclaration(member)) {
|
|
56
|
+
methods.push({ name, kind: 'getter' });
|
|
57
|
+
}
|
|
58
|
+
else if (ts.isPropertyDeclaration(member)) {
|
|
59
|
+
// Check if it's an arrow function property (e.g., name = async () => {})
|
|
60
|
+
if (member.initializer && (ts.isArrowFunction(member.initializer) || ts.isFunctionExpression(member.initializer))) {
|
|
61
|
+
const fn = member.initializer;
|
|
62
|
+
const fnModifiers = ts.canHaveModifiers(fn) ? ts.getModifiers(fn) : undefined;
|
|
63
|
+
const isAsync = fnModifiers?.some((m) => m.kind === ts.SyntaxKind.AsyncKeyword) ?? false;
|
|
64
|
+
const params = extractParams(fn.parameters, checker);
|
|
65
|
+
methods.push({
|
|
66
|
+
name,
|
|
67
|
+
kind: 'method',
|
|
68
|
+
async: isAsync ? true : undefined,
|
|
69
|
+
params: params.length > 0 ? params : undefined,
|
|
70
|
+
});
|
|
71
|
+
}
|
|
72
|
+
else {
|
|
73
|
+
methods.push({ name, kind: 'property' });
|
|
74
|
+
}
|
|
75
|
+
}
|
|
76
|
+
}
|
|
77
|
+
if (methods.length > 0) {
|
|
78
|
+
surfaces.push({
|
|
79
|
+
className,
|
|
80
|
+
file: sourceFile.fileName,
|
|
81
|
+
methods,
|
|
82
|
+
extends: extendsName,
|
|
83
|
+
});
|
|
84
|
+
}
|
|
85
|
+
});
|
|
86
|
+
return surfaces;
|
|
87
|
+
}
|
|
88
|
+
function extractParams(params, checker) {
|
|
89
|
+
return params.map((p) => {
|
|
90
|
+
const name = ts.isIdentifier(p.name) ? p.name.text : p.name.getText();
|
|
91
|
+
const optional = p.questionToken !== undefined || p.initializer !== undefined;
|
|
92
|
+
let type;
|
|
93
|
+
if (p.type) {
|
|
94
|
+
type = p.type.getText();
|
|
95
|
+
}
|
|
96
|
+
else if (checker) {
|
|
97
|
+
try {
|
|
98
|
+
const symbol = checker.getSymbolAtLocation(p.name);
|
|
99
|
+
if (symbol) {
|
|
100
|
+
const t = checker.getTypeOfSymbolAtLocation(symbol, p);
|
|
101
|
+
type = checker.typeToString(t);
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
catch {
|
|
105
|
+
// Type inference failure is non-fatal
|
|
106
|
+
}
|
|
107
|
+
}
|
|
108
|
+
return { name, type, optional: optional || undefined };
|
|
109
|
+
});
|
|
110
|
+
}
|
|
111
|
+
function extractReturnType(method, checker) {
|
|
112
|
+
if (method.type) {
|
|
113
|
+
return method.type.getText();
|
|
114
|
+
}
|
|
115
|
+
if (checker) {
|
|
116
|
+
try {
|
|
117
|
+
const signature = checker.getSignatureFromDeclaration(method);
|
|
118
|
+
if (signature) {
|
|
119
|
+
const returnType = checker.getReturnTypeOfSignature(signature);
|
|
120
|
+
const typeStr = checker.typeToString(returnType);
|
|
121
|
+
// Skip overly verbose inferred types
|
|
122
|
+
if (typeStr.length < 100) {
|
|
123
|
+
return typeStr;
|
|
124
|
+
}
|
|
125
|
+
}
|
|
126
|
+
}
|
|
127
|
+
catch {
|
|
128
|
+
// Type inference failure is non-fatal
|
|
129
|
+
}
|
|
130
|
+
}
|
|
131
|
+
return null;
|
|
132
|
+
}
|
|
133
|
+
/**
|
|
134
|
+
* Extract page objects using the TypeScript Compiler API.
|
|
135
|
+
* Falls back to regex if compilation fails.
|
|
136
|
+
*/
|
|
137
|
+
function extractWithAST(files) {
|
|
138
|
+
if (files.length === 0) {
|
|
139
|
+
return [];
|
|
140
|
+
}
|
|
141
|
+
// Find the nearest tsconfig.json
|
|
142
|
+
const firstDir = join(files[0], '..');
|
|
143
|
+
const tsconfigPath = ts.findConfigFile(firstDir, ts.sys.fileExists, 'tsconfig.json');
|
|
144
|
+
let program;
|
|
145
|
+
if (tsconfigPath) {
|
|
146
|
+
const configFile = ts.readConfigFile(tsconfigPath, ts.sys.readFile);
|
|
147
|
+
const parsedConfig = ts.parseJsonConfigFileContent(configFile.config, ts.sys, join(tsconfigPath, '..'));
|
|
148
|
+
// Only compile the files we care about, using the config's compiler options
|
|
149
|
+
program = ts.createProgram(files, parsedConfig.options);
|
|
150
|
+
}
|
|
151
|
+
else {
|
|
152
|
+
program = ts.createProgram(files, {
|
|
153
|
+
target: ts.ScriptTarget.ES2022,
|
|
154
|
+
module: ts.ModuleKind.ESNext,
|
|
155
|
+
moduleResolution: ts.ModuleResolutionKind.Bundler,
|
|
156
|
+
strict: true,
|
|
157
|
+
allowJs: false,
|
|
158
|
+
noEmit: true,
|
|
159
|
+
});
|
|
160
|
+
}
|
|
161
|
+
const checker = program.getTypeChecker();
|
|
162
|
+
const surfaces = [];
|
|
163
|
+
for (const filePath of files) {
|
|
164
|
+
const sourceFile = program.getSourceFile(filePath);
|
|
165
|
+
if (!sourceFile) {
|
|
166
|
+
continue;
|
|
167
|
+
}
|
|
168
|
+
surfaces.push(...extractMethodsFromAST(sourceFile, checker));
|
|
169
|
+
}
|
|
170
|
+
// Resolve inherited methods: if class extends another class in the catalog,
|
|
171
|
+
// merge parent methods that the child doesn't override
|
|
172
|
+
resolveInheritance(surfaces);
|
|
173
|
+
return surfaces;
|
|
174
|
+
}
|
|
175
|
+
/**
|
|
176
|
+
* Merge parent class methods into child classes that extend them.
|
|
177
|
+
*/
|
|
178
|
+
function resolveInheritance(surfaces) {
|
|
179
|
+
const byName = new Map(surfaces.map((s) => [s.className, s]));
|
|
180
|
+
for (const surface of surfaces) {
|
|
181
|
+
if (!surface.extends) {
|
|
182
|
+
continue;
|
|
183
|
+
}
|
|
184
|
+
const parent = byName.get(surface.extends);
|
|
185
|
+
if (!parent) {
|
|
186
|
+
continue;
|
|
187
|
+
}
|
|
188
|
+
const childMethodNames = new Set(surface.methods.map((m) => m.name));
|
|
189
|
+
for (const parentMethod of parent.methods) {
|
|
190
|
+
if (!childMethodNames.has(parentMethod.name)) {
|
|
191
|
+
surface.methods.push({ ...parentMethod });
|
|
192
|
+
}
|
|
193
|
+
}
|
|
194
|
+
}
|
|
195
|
+
}
|
|
196
|
+
// ── Regex-based extraction (fallback) ──────────────────────
|
|
5
197
|
const RESERVED_WORDS = new Set([
|
|
6
198
|
'private', 'protected', 'static', 'abstract', 'override',
|
|
7
199
|
'if', 'for', 'while', 'switch', 'return',
|
|
8
200
|
'const', 'let', 'var', 'import', 'export',
|
|
9
201
|
'class', 'type', 'interface', 'constructor',
|
|
10
202
|
]);
|
|
11
|
-
function
|
|
203
|
+
function extractMethodsFromRegex(content) {
|
|
12
204
|
const methods = [];
|
|
13
205
|
const seen = new Set();
|
|
14
|
-
// Match async method declarations: async methodName(
|
|
15
206
|
const asyncMethodRe = /(?:async\s+)([a-zA-Z_]\w*)\s*\(/g;
|
|
16
207
|
let match;
|
|
17
208
|
while ((match = asyncMethodRe.exec(content)) !== null) {
|
|
18
209
|
const name = match[1];
|
|
19
|
-
if (RESERVED_WORDS.has(name)) {
|
|
20
|
-
continue;
|
|
21
|
-
}
|
|
22
|
-
if (!seen.has(name)) {
|
|
210
|
+
if (!RESERVED_WORDS.has(name) && !seen.has(name)) {
|
|
23
211
|
seen.add(name);
|
|
24
|
-
methods.push({ name, kind: 'method' });
|
|
212
|
+
methods.push({ name, kind: 'method', async: true });
|
|
25
213
|
}
|
|
26
214
|
}
|
|
27
|
-
// Match non-async public method patterns
|
|
28
215
|
const methodRe = /^\s+(?:readonly\s+)?([a-zA-Z_]\w*)\s*(?:\(|=\s*(?:async\s*)?\()/gm;
|
|
29
216
|
while ((match = methodRe.exec(content)) !== null) {
|
|
30
217
|
const name = match[1];
|
|
31
|
-
if (RESERVED_WORDS.has(name)
|
|
32
|
-
|
|
218
|
+
if (!RESERVED_WORDS.has(name) && !seen.has(name)) {
|
|
219
|
+
seen.add(name);
|
|
220
|
+
const isAsync = match[0].includes('async');
|
|
221
|
+
methods.push({ name, kind: 'method', async: isAsync ? true : undefined });
|
|
222
|
+
}
|
|
223
|
+
}
|
|
224
|
+
const arrowRe = /^\s+([a-zA-Z_]\w*)\s*=\s*(async\s+)?(?:\([^)]*\)|[a-zA-Z_]\w*)\s*=>/gm;
|
|
225
|
+
while ((match = arrowRe.exec(content)) !== null) {
|
|
226
|
+
const name = match[1];
|
|
227
|
+
if (!RESERVED_WORDS.has(name) && !seen.has(name)) {
|
|
228
|
+
seen.add(name);
|
|
229
|
+
methods.push({ name, kind: 'method', async: match[2] ? true : undefined });
|
|
33
230
|
}
|
|
34
|
-
seen.add(name);
|
|
35
|
-
methods.push({ name, kind: 'method' });
|
|
36
231
|
}
|
|
37
|
-
// Match getter patterns: get propertyName()
|
|
38
232
|
const getterRe = /\bget\s+([a-zA-Z_]\w*)\s*\(\)/g;
|
|
39
233
|
while ((match = getterRe.exec(content)) !== null) {
|
|
40
234
|
const name = match[1];
|
|
@@ -43,15 +237,13 @@ function extractMethodsFromSource(content) {
|
|
|
43
237
|
methods.push({ name, kind: 'getter' });
|
|
44
238
|
}
|
|
45
239
|
}
|
|
46
|
-
// Match readonly property declarations
|
|
47
240
|
const propRe = /^\s+(?:readonly\s+)?([a-zA-Z_]\w*)\s*[:=]/gm;
|
|
48
241
|
while ((match = propRe.exec(content)) !== null) {
|
|
49
242
|
const name = match[1];
|
|
50
|
-
if (RESERVED_WORDS.has(name)
|
|
51
|
-
|
|
243
|
+
if (!RESERVED_WORDS.has(name) && !seen.has(name)) {
|
|
244
|
+
seen.add(name);
|
|
245
|
+
methods.push({ name, kind: 'property' });
|
|
52
246
|
}
|
|
53
|
-
seen.add(name);
|
|
54
|
-
methods.push({ name, kind: 'property' });
|
|
55
247
|
}
|
|
56
248
|
return methods;
|
|
57
249
|
}
|
|
@@ -59,7 +251,27 @@ function extractClassName(content) {
|
|
|
59
251
|
const match = content.match(/(?:export\s+)?class\s+(\w+)/);
|
|
60
252
|
return match ? match[1] : null;
|
|
61
253
|
}
|
|
62
|
-
|
|
254
|
+
// ── Directory scanning ─────────────────────────────────────
|
|
255
|
+
function collectTypeScriptFiles(dir) {
|
|
256
|
+
const files = [];
|
|
257
|
+
if (!existsSync(dir)) {
|
|
258
|
+
return files;
|
|
259
|
+
}
|
|
260
|
+
const entries = readdirSync(dir, { withFileTypes: true });
|
|
261
|
+
for (const entry of entries) {
|
|
262
|
+
const fullPath = join(dir, entry.name);
|
|
263
|
+
if (entry.isDirectory()) {
|
|
264
|
+
files.push(...collectTypeScriptFiles(fullPath));
|
|
265
|
+
continue;
|
|
266
|
+
}
|
|
267
|
+
const ext = extname(entry.name);
|
|
268
|
+
if (ext === '.ts' || ext === '.tsx') {
|
|
269
|
+
files.push(fullPath);
|
|
270
|
+
}
|
|
271
|
+
}
|
|
272
|
+
return files;
|
|
273
|
+
}
|
|
274
|
+
function scanDirectoryWithRegex(dir) {
|
|
63
275
|
const surfaces = [];
|
|
64
276
|
if (!existsSync(dir)) {
|
|
65
277
|
return surfaces;
|
|
@@ -68,29 +280,22 @@ function scanDirectory(dir) {
|
|
|
68
280
|
for (const entry of entries) {
|
|
69
281
|
const fullPath = join(dir, entry.name);
|
|
70
282
|
if (entry.isDirectory()) {
|
|
71
|
-
surfaces.push(...
|
|
283
|
+
surfaces.push(...scanDirectoryWithRegex(fullPath));
|
|
72
284
|
continue;
|
|
73
285
|
}
|
|
74
286
|
const ext = extname(entry.name);
|
|
75
287
|
if (ext !== '.ts' && ext !== '.tsx') {
|
|
76
288
|
continue;
|
|
77
289
|
}
|
|
78
|
-
if (entry.name === 'index.ts' || entry.name === 'index.tsx') {
|
|
79
|
-
continue;
|
|
80
|
-
}
|
|
81
290
|
try {
|
|
82
291
|
const content = readFileSync(fullPath, 'utf-8');
|
|
83
292
|
const className = extractClassName(content);
|
|
84
293
|
if (!className) {
|
|
85
294
|
continue;
|
|
86
295
|
}
|
|
87
|
-
const extractedMethods =
|
|
296
|
+
const extractedMethods = extractMethodsFromRegex(content);
|
|
88
297
|
if (extractedMethods.length > 0) {
|
|
89
|
-
surfaces.push({
|
|
90
|
-
className,
|
|
91
|
-
file: fullPath,
|
|
92
|
-
methods: extractedMethods,
|
|
93
|
-
});
|
|
298
|
+
surfaces.push({ className, file: fullPath, methods: extractedMethods });
|
|
94
299
|
}
|
|
95
300
|
}
|
|
96
301
|
catch {
|
|
@@ -99,6 +304,7 @@ function scanDirectory(dir) {
|
|
|
99
304
|
}
|
|
100
305
|
return surfaces;
|
|
101
306
|
}
|
|
307
|
+
// ── Public API ──────────────────────────────────────────────
|
|
102
308
|
export function buildApiSurface(testsRoot, config) {
|
|
103
309
|
const pageObjectsDir = config?.pageObjectsDir
|
|
104
310
|
? join(testsRoot, config.pageObjectsDir)
|
|
@@ -106,10 +312,30 @@ export function buildApiSurface(testsRoot, config) {
|
|
|
106
312
|
const componentsDir = config?.componentsDir
|
|
107
313
|
? join(testsRoot, config.componentsDir)
|
|
108
314
|
: join(testsRoot, 'lib', 'src', 'ui', 'components');
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
315
|
+
let pageObjects;
|
|
316
|
+
if (config?.useRegexFallback) {
|
|
317
|
+
pageObjects = [
|
|
318
|
+
...scanDirectoryWithRegex(pageObjectsDir),
|
|
319
|
+
...scanDirectoryWithRegex(componentsDir),
|
|
320
|
+
];
|
|
321
|
+
}
|
|
322
|
+
else {
|
|
323
|
+
// Use TypeScript AST — full type info, inheritance, params
|
|
324
|
+
const allFiles = [
|
|
325
|
+
...collectTypeScriptFiles(pageObjectsDir),
|
|
326
|
+
...collectTypeScriptFiles(componentsDir),
|
|
327
|
+
];
|
|
328
|
+
try {
|
|
329
|
+
pageObjects = extractWithAST(allFiles);
|
|
330
|
+
}
|
|
331
|
+
catch {
|
|
332
|
+
// Fall back to regex if AST extraction fails
|
|
333
|
+
pageObjects = [
|
|
334
|
+
...scanDirectoryWithRegex(pageObjectsDir),
|
|
335
|
+
...scanDirectoryWithRegex(componentsDir),
|
|
336
|
+
];
|
|
337
|
+
}
|
|
338
|
+
}
|
|
113
339
|
return {
|
|
114
340
|
pageObjects,
|
|
115
341
|
generatedAt: new Date().toISOString(),
|
|
@@ -168,7 +394,12 @@ export function formatApiSurfaceForPrompt(catalog, classNames) {
|
|
|
168
394
|
if (m.kind === 'getter') {
|
|
169
395
|
return ` get ${m.name}()`;
|
|
170
396
|
}
|
|
171
|
-
|
|
397
|
+
const prefix = m.async ? 'async ' : '';
|
|
398
|
+
const paramStr = m.params
|
|
399
|
+
? m.params.map((p) => `${p.name}${p.optional ? '?' : ''}${p.type ? `: ${p.type}` : ''}`).join(', ')
|
|
400
|
+
: '';
|
|
401
|
+
const retStr = m.returnType ? `: ${m.returnType}` : '';
|
|
402
|
+
return ` ${prefix}${m.name}(${paramStr})${retStr}`;
|
|
172
403
|
})
|
|
173
404
|
.join('\n');
|
|
174
405
|
sections.push(`${name}:\n${methodList}`);
|
|
@@ -0,0 +1,121 @@
|
|
|
1
|
+
// Copyright (c) 2015-present Mattermost, Inc. All Rights Reserved.
|
|
2
|
+
// See LICENSE.txt for license information.
|
|
3
|
+
/**
|
|
4
|
+
* Tracks historical test failure correlations: which tests fail when certain files change.
|
|
5
|
+
* Used to boost confidence in impact analysis — if a file change historically breaks a test,
|
|
6
|
+
* future changes to that file should prioritize that test.
|
|
7
|
+
*
|
|
8
|
+
* Data is stored as a JSON file at .e2e-ai-agents/failure-history.json.
|
|
9
|
+
*/
|
|
10
|
+
import { existsSync, readFileSync, writeFileSync, mkdirSync } from 'fs';
|
|
11
|
+
import { join, dirname } from 'path';
|
|
12
|
+
const DEFAULT_HISTORY = {
|
|
13
|
+
correlations: [],
|
|
14
|
+
totalRuns: 0,
|
|
15
|
+
updatedAt: new Date().toISOString(),
|
|
16
|
+
};
|
|
17
|
+
export function loadFailureHistory(testsRoot) {
|
|
18
|
+
const historyPath = join(testsRoot, '.e2e-ai-agents', 'failure-history.json');
|
|
19
|
+
if (!existsSync(historyPath)) {
|
|
20
|
+
return { ...DEFAULT_HISTORY };
|
|
21
|
+
}
|
|
22
|
+
try {
|
|
23
|
+
const raw = JSON.parse(readFileSync(historyPath, 'utf-8'));
|
|
24
|
+
if (!Array.isArray(raw.correlations)) {
|
|
25
|
+
return { ...DEFAULT_HISTORY };
|
|
26
|
+
}
|
|
27
|
+
return raw;
|
|
28
|
+
}
|
|
29
|
+
catch {
|
|
30
|
+
return { ...DEFAULT_HISTORY };
|
|
31
|
+
}
|
|
32
|
+
}
|
|
33
|
+
export function saveFailureHistory(testsRoot, history) {
|
|
34
|
+
const historyPath = join(testsRoot, '.e2e-ai-agents', 'failure-history.json');
|
|
35
|
+
try {
|
|
36
|
+
const dir = dirname(historyPath);
|
|
37
|
+
if (!existsSync(dir)) {
|
|
38
|
+
mkdirSync(dir, { recursive: true });
|
|
39
|
+
}
|
|
40
|
+
history.updatedAt = new Date().toISOString();
|
|
41
|
+
writeFileSync(historyPath, JSON.stringify(history, null, 2), 'utf-8');
|
|
42
|
+
}
|
|
43
|
+
catch {
|
|
44
|
+
// Non-fatal — history is advisory, not required
|
|
45
|
+
}
|
|
46
|
+
}
|
|
47
|
+
/**
|
|
48
|
+
* Record that a set of changed files caused a set of spec failures.
|
|
49
|
+
* Call this after a test run where failures were observed.
|
|
50
|
+
*/
|
|
51
|
+
export function recordFailures(history, changedFiles, failedSpecs) {
|
|
52
|
+
const now = new Date().toISOString();
|
|
53
|
+
const updated = { ...history, totalRuns: history.totalRuns + 1, correlations: [...history.correlations] };
|
|
54
|
+
for (const changedFile of changedFiles) {
|
|
55
|
+
for (const specFile of failedSpecs) {
|
|
56
|
+
const existing = updated.correlations.find((c) => c.changedFile === changedFile && c.specFile === specFile);
|
|
57
|
+
if (existing) {
|
|
58
|
+
existing.count++;
|
|
59
|
+
existing.lastSeen = now;
|
|
60
|
+
}
|
|
61
|
+
else {
|
|
62
|
+
updated.correlations.push({
|
|
63
|
+
changedFile,
|
|
64
|
+
specFile,
|
|
65
|
+
count: 1,
|
|
66
|
+
lastSeen: now,
|
|
67
|
+
});
|
|
68
|
+
}
|
|
69
|
+
}
|
|
70
|
+
}
|
|
71
|
+
// Prune stale correlations (not seen in 90 days)
|
|
72
|
+
const cutoff = new Date();
|
|
73
|
+
cutoff.setDate(cutoff.getDate() - 90);
|
|
74
|
+
const cutoffStr = cutoff.toISOString();
|
|
75
|
+
updated.correlations = updated.correlations.filter((c) => c.lastSeen >= cutoffStr);
|
|
76
|
+
return updated;
|
|
77
|
+
}
|
|
78
|
+
/**
|
|
79
|
+
* Get a confidence boost (0-20) for a file based on historical failure patterns.
|
|
80
|
+
* A file that historically causes test failures gets a higher confidence boost
|
|
81
|
+
* when detected as impacted, meaning the system is more confident it needs testing.
|
|
82
|
+
*/
|
|
83
|
+
export function getConfidenceBoost(history, changedFile) {
|
|
84
|
+
const correlations = history.correlations.filter((c) => c.changedFile === changedFile);
|
|
85
|
+
if (correlations.length === 0) {
|
|
86
|
+
return 0;
|
|
87
|
+
}
|
|
88
|
+
// More correlations and higher counts = more confidence
|
|
89
|
+
const totalCount = correlations.reduce((sum, c) => sum + c.count, 0);
|
|
90
|
+
const uniqueSpecs = correlations.length;
|
|
91
|
+
// Scale: 1 correlation = +5, 3+ = +10, 5+ with high counts = +15, max +20
|
|
92
|
+
if (totalCount >= 10 && uniqueSpecs >= 5)
|
|
93
|
+
return 20;
|
|
94
|
+
if (totalCount >= 5 && uniqueSpecs >= 3)
|
|
95
|
+
return 15;
|
|
96
|
+
if (totalCount >= 3)
|
|
97
|
+
return 10;
|
|
98
|
+
return 5;
|
|
99
|
+
}
|
|
100
|
+
/**
|
|
101
|
+
* Get the most likely failing specs for a set of changed files, based on history.
|
|
102
|
+
* Returns specs sorted by correlation strength (count * recency).
|
|
103
|
+
*/
|
|
104
|
+
export function getPredictedFailures(history, changedFiles, limit = 10) {
|
|
105
|
+
const specScores = new Map();
|
|
106
|
+
for (const changedFile of changedFiles) {
|
|
107
|
+
for (const c of history.correlations) {
|
|
108
|
+
if (c.changedFile !== changedFile)
|
|
109
|
+
continue;
|
|
110
|
+
// Score: count weighted by recency (days since last seen)
|
|
111
|
+
const daysSince = (Date.now() - new Date(c.lastSeen).getTime()) / (1000 * 60 * 60 * 24);
|
|
112
|
+
const recencyWeight = Math.max(0.1, 1 - daysSince / 90);
|
|
113
|
+
const score = c.count * recencyWeight;
|
|
114
|
+
specScores.set(c.specFile, (specScores.get(c.specFile) || 0) + score);
|
|
115
|
+
}
|
|
116
|
+
}
|
|
117
|
+
return Array.from(specScores.entries())
|
|
118
|
+
.map(([specFile, score]) => ({ specFile, score }))
|
|
119
|
+
.sort((a, b) => b.score - a.score)
|
|
120
|
+
.slice(0, limit);
|
|
121
|
+
}
|
|
@@ -106,8 +106,22 @@ function validateFamily(family) {
|
|
|
106
106
|
if (testType) {
|
|
107
107
|
result.testType = testType;
|
|
108
108
|
}
|
|
109
|
+
if (Array.isArray(obj.assertionPatterns)) {
|
|
110
|
+
result.assertionPatterns = validateAssertionPatterns(obj.assertionPatterns);
|
|
111
|
+
}
|
|
109
112
|
return result;
|
|
110
113
|
}
|
|
114
|
+
const VALID_ASSERTION_TYPES = [
|
|
115
|
+
'state-change', 'cross-user', 'persistence', 'negative',
|
|
116
|
+
'permission', 'data-integrity', 'error-handling',
|
|
117
|
+
];
|
|
118
|
+
function validateAssertionPatterns(patterns) {
|
|
119
|
+
return patterns
|
|
120
|
+
.filter((p) => p != null && typeof p === 'object')
|
|
121
|
+
.filter((p) => typeof p.type === 'string' && typeof p.pattern === 'string')
|
|
122
|
+
.filter((p) => VALID_ASSERTION_TYPES.includes(p.type))
|
|
123
|
+
.map((p) => ({ type: p.type, pattern: p.pattern }));
|
|
124
|
+
}
|
|
111
125
|
function validateFeature(feature) {
|
|
112
126
|
if (!feature || typeof feature !== 'object') {
|
|
113
127
|
return null;
|
|
@@ -141,6 +155,9 @@ function validateFeature(feature) {
|
|
|
141
155
|
if (Array.isArray(obj.userFlows)) {
|
|
142
156
|
result.userFlows = obj.userFlows.filter((v) => typeof v === 'string');
|
|
143
157
|
}
|
|
158
|
+
if (Array.isArray(obj.assertionPatterns)) {
|
|
159
|
+
result.assertionPatterns = validateAssertionPatterns(obj.assertionPatterns);
|
|
160
|
+
}
|
|
144
161
|
return result;
|
|
145
162
|
}
|
|
146
163
|
export function loadRouteFamilyManifest(testsRoot, config) {
|
|
@@ -294,6 +311,19 @@ export function getRoutesForBinding(manifest, binding) {
|
|
|
294
311
|
}
|
|
295
312
|
return family.routes;
|
|
296
313
|
}
|
|
314
|
+
export function getAssertionPatternsForBinding(manifest, binding) {
|
|
315
|
+
const family = getFamilyById(manifest, binding.family);
|
|
316
|
+
if (!family) {
|
|
317
|
+
return [];
|
|
318
|
+
}
|
|
319
|
+
if (binding.feature) {
|
|
320
|
+
const feature = getFeatureById(family, binding.feature);
|
|
321
|
+
if (feature?.assertionPatterns && feature.assertionPatterns.length > 0) {
|
|
322
|
+
return feature.assertionPatterns;
|
|
323
|
+
}
|
|
324
|
+
}
|
|
325
|
+
return family.assertionPatterns || [];
|
|
326
|
+
}
|
|
297
327
|
export function clearManifestCache() {
|
|
298
328
|
manifestCache.clear();
|
|
299
329
|
}
|
|
@@ -345,7 +375,7 @@ export function serializeManifest(manifest) {
|
|
|
345
375
|
const cleaned = { ...f };
|
|
346
376
|
const optionalArrays = [
|
|
347
377
|
'pageObjects', 'components', 'webappPaths', 'serverPaths',
|
|
348
|
-
'specDirs', 'cypressSpecDirs', 'tags', 'userFlows', 'features', 'apiEndpoints',
|
|
378
|
+
'specDirs', 'cypressSpecDirs', 'tags', 'userFlows', 'features', 'apiEndpoints', 'assertionPatterns',
|
|
349
379
|
];
|
|
350
380
|
for (const key of optionalArrays) {
|
|
351
381
|
if (!cleaned[key] || (Array.isArray(cleaned[key]) && cleaned[key].length === 0)) {
|
|
@@ -3,7 +3,8 @@
|
|
|
3
3
|
import { LLMProviderFactory } from '../provider_factory.js';
|
|
4
4
|
import { buildImpactPrompt, parseImpactResponse } from '../prompts/impact.js';
|
|
5
5
|
import { formatContextForPrompt } from '../knowledge/context_loader.js';
|
|
6
|
-
import { getFamilyById } from '../knowledge/route_families.js';
|
|
6
|
+
import { getFamilyById, getAssertionPatternsForBinding } from '../knowledge/route_families.js';
|
|
7
|
+
import { loadFailureHistory, getConfidenceBoost } from '../knowledge/failure_history.js';
|
|
7
8
|
import { getSpecsForFamily } from '../knowledge/spec_index.js';
|
|
8
9
|
import { computeConfidence, shouldForceCannotDetermine } from '../validation/guardrails.js';
|
|
9
10
|
function normalizePriority(value) {
|
|
@@ -18,7 +19,7 @@ async function getProvider(config) {
|
|
|
18
19
|
}
|
|
19
20
|
return LLMProviderFactory.createFromEnv();
|
|
20
21
|
}
|
|
21
|
-
export async function runImpactStage(familyGroups, manifest, specIndex, apiSurface, context, config) {
|
|
22
|
+
export async function runImpactStage(familyGroups, manifest, specIndex, apiSurface, context, config, testsRoot) {
|
|
22
23
|
const warnings = [];
|
|
23
24
|
const allDecisions = [];
|
|
24
25
|
if (familyGroups.length === 0) {
|
|
@@ -35,6 +36,8 @@ export async function runImpactStage(familyGroups, manifest, specIndex, apiSurfa
|
|
|
35
36
|
return { decisions: [], warnings, providerName: 'none' };
|
|
36
37
|
}
|
|
37
38
|
const contextBlock = formatContextForPrompt(context);
|
|
39
|
+
// Load historical failure correlations for confidence boosting
|
|
40
|
+
const failureHistory = testsRoot ? loadFailureHistory(testsRoot) : null;
|
|
38
41
|
for (const group of familyGroups) {
|
|
39
42
|
const family = manifest ? getFamilyById(manifest, group.familyId) : null;
|
|
40
43
|
if (!family) {
|
|
@@ -83,15 +86,27 @@ export async function runImpactStage(familyGroups, manifest, specIndex, apiSurfa
|
|
|
83
86
|
if (!flow.id || !flow.changedFiles || !Array.isArray(flow.changedFiles)) {
|
|
84
87
|
continue;
|
|
85
88
|
}
|
|
89
|
+
// Compute confidence with optional historical failure boost
|
|
90
|
+
const changedFilesList = Array.isArray(flow.changedFiles)
|
|
91
|
+
? flow.changedFiles.filter((f) => typeof f === 'string')
|
|
92
|
+
: [];
|
|
93
|
+
const historyBoost = failureHistory
|
|
94
|
+
? Math.max(...changedFilesList.map((f) => getConfidenceBoost(failureHistory, f)), 0)
|
|
95
|
+
: 0;
|
|
86
96
|
const confidence = typeof flow.confidence === 'number'
|
|
87
|
-
? Math.
|
|
97
|
+
? Math.min(100, Math.max(0, flow.confidence) + historyBoost)
|
|
88
98
|
: computeConfidence({
|
|
89
99
|
hasRouteFamily: true,
|
|
90
100
|
hasSpecificRoute: Boolean(flow.route),
|
|
91
101
|
hasPageObject: Boolean(flow.pageObjects && flow.pageObjects.length > 0),
|
|
92
102
|
hasUserAction: Boolean(flow.userActions && flow.userActions.length > 0),
|
|
93
103
|
hasExistingSpecCited: false,
|
|
104
|
+
historyBoost,
|
|
94
105
|
});
|
|
106
|
+
// Resolve assertion patterns from manifest for this flow's family/feature
|
|
107
|
+
const assertionPatterns = manifest
|
|
108
|
+
? getAssertionPatternsForBinding(manifest, { family: group.familyId, feature: group.featureId })
|
|
109
|
+
: [];
|
|
95
110
|
const decision = {
|
|
96
111
|
flowId: flow.id,
|
|
97
112
|
flowName: flow.name || flow.id,
|
|
@@ -107,6 +122,7 @@ export async function runImpactStage(familyGroups, manifest, specIndex, apiSurfa
|
|
|
107
122
|
blockingReason: shouldForceCannotDetermine(confidence) ? 'Confidence too low to determine action.' : undefined,
|
|
108
123
|
priority: normalizePriority(flow.priority),
|
|
109
124
|
userActions: Array.isArray(flow.userActions) ? flow.userActions.filter((a) => typeof a === 'string') : [],
|
|
125
|
+
assertionPatterns: assertionPatterns.length > 0 ? assertionPatterns : undefined,
|
|
110
126
|
};
|
|
111
127
|
allDecisions.push(decision);
|
|
112
128
|
}
|
|
@@ -43,13 +43,26 @@ export async function runCoverageStage(decisions, specIndex, context, testsRoot,
|
|
|
43
43
|
for (const [familyId, familyDecisions] of byFamily) {
|
|
44
44
|
// Gather relevant specs
|
|
45
45
|
const specs = getSpecsForFamily(specIndex, familyId);
|
|
46
|
-
|
|
47
|
-
|
|
46
|
+
// Two-tier approach: send all spec titles (compact), full content for top matches only
|
|
47
|
+
const allSpecSummaries = specs.map((s) => ({
|
|
48
|
+
relativePath: s.relativePath,
|
|
49
|
+
testTitles: s.testTitles,
|
|
50
|
+
}));
|
|
51
|
+
// Load full content with a total budget of 200K chars (~50K tokens) to avoid blowing context windows
|
|
52
|
+
const MAX_TOTAL_SPEC_CHARS = 200000;
|
|
53
|
+
let totalSpecChars = 0;
|
|
54
|
+
const specsWithContent = [];
|
|
55
|
+
for (const s of specs) {
|
|
56
|
+
if (specsWithContent.length >= 30)
|
|
57
|
+
break;
|
|
48
58
|
const content = loadSpecFileContent(testsRoot, s.relativePath, maxSpecChars);
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
59
|
+
if (!content)
|
|
60
|
+
continue;
|
|
61
|
+
if (totalSpecChars + content.length > MAX_TOTAL_SPEC_CHARS)
|
|
62
|
+
break;
|
|
63
|
+
totalSpecChars += content.length;
|
|
64
|
+
specsWithContent.push({ relativePath: s.relativePath, content, testTitles: s.testTitles });
|
|
65
|
+
}
|
|
53
66
|
if (specsWithContent.length === 0) {
|
|
54
67
|
// No specs to evaluate — mark all as create_spec
|
|
55
68
|
for (const d of familyDecisions) {
|
|
@@ -70,10 +83,18 @@ export async function runCoverageStage(decisions, specIndex, context, testsRoot,
|
|
|
70
83
|
evidence: d.evidence,
|
|
71
84
|
priority: d.priority,
|
|
72
85
|
}));
|
|
86
|
+
// Include titles-only summaries for specs beyond the content limit
|
|
87
|
+
const extraSummaries = allSpecSummaries
|
|
88
|
+
.slice(specsWithContent.length)
|
|
89
|
+
.map((s) => ` - ${s.relativePath}: ${s.testTitles.join(', ')}`)
|
|
90
|
+
.join('\n');
|
|
91
|
+
const extraContext = extraSummaries
|
|
92
|
+
? `\nADDITIONAL SPECS (titles only, no content loaded):\n${extraSummaries}\n`
|
|
93
|
+
: '';
|
|
73
94
|
const prompt = buildCoveragePrompt({
|
|
74
95
|
flows,
|
|
75
96
|
specs: specsWithContent,
|
|
76
|
-
contextBlock,
|
|
97
|
+
contextBlock: contextBlock + extraContext,
|
|
77
98
|
profile: config.profile,
|
|
78
99
|
});
|
|
79
100
|
try {
|