forge-workflow 0.1.0-beta.4 → 0.1.0-beta.6

package/lib/commands/test.js

@@ -5,7 +5,6 @@
  *
  * Wraps test execution with smart defaults:
  * - Auto-detects package manager from lockfiles
- * - Checks Beads (Dolt) connectivity, sets BEADS_SKIP_TESTS if unavailable
  * - Supports --affected flag to run only tests for changed files
  *
  * Security: Uses execFileSync for subprocess calls (OWASP A03)
@@ -17,6 +16,8 @@ const { execFileSync: defaultExecFileSync, spawnSync: defaultSpawnSync } = requi
 const defaultFs = require('node:fs');
 const path = require('node:path');
 
+const { selectDocAssertingTests } = require('../doc-assertions');
+
 /** @type {Array<[string, string]>} Lockfile → package manager mapping (order matters) */
 const LOCKFILE_MAP = [
 	['bun.lockb', 'bun'],
@@ -26,13 +27,14 @@ const LOCKFILE_MAP = [
 ];
 
 const DEFAULT_TIMEOUT = 120000;
-const BEADS_CHECK_TIMEOUT = 3000;
 const DIRECT_TEST_CANDIDATES = Object.freeze({
 	'bin/forge.js': [
 		'test/cli-flags.test.js',
 		'test/forge-cli-registry.test.js',
-		'test/setup-github-sync.test.js',
 		'test/setup-runtime-flags.test.js',
+		'test/github-launcher.test.js',
+		'test/integration/github-account-context.test.js',
+		'test/structural/github-account-public-surface.test.js',
 	],
 	// The second CLI surface (command descriptions + shepherd wiring). Without an
 	// entry here it resolves to zero tests and pushes fall back to the full suite.
@@ -41,15 +43,150 @@ const DIRECT_TEST_CANDIDATES = Object.freeze({
 		'test/forge-cmd-shepherd.test.js',
 	],
 	'bin/forge-preflight.js': ['test/bin/forge-preflight.test.js'],
+	// The CI entry point is covered by the doc-assertion regression suite; without
+	// this entry it resolves to zero tests and forces the full suite.
+	// (lib/doc-assertions.js resolves via the lib/ -> test/ convention below.)
+	'scripts/doc-asserting-tests.js': [
+		'test/doc-assertions.test.js',
+		'test/scripts/doc-asserting-tests.test.js',
+	],
+	// The suites that assert the repo .gitignore's actual content. The path is
+	// already known to the classifier, so without candidates here it resolves
+	// zero tests and a one-line ignore addition pays the full-suite lane.
+	'.gitignore': [
+		'test/check-forge-token.test.js',
+		'test/cleanup/dropped-agent-config.test.js',
+		'test/coverage-config.test.js',
+		'test/eval/eval-history.test.js',
+		'test/freshness-token.test.js',
+		'test/mutation-config.test.js',
+		'test/test-dashboard.test.js',
+	],
 	'lib/lefthook-check.js': ['test/lefthook-check.test.js', 'test/runtime-health.test.js'],
 	'lib/runtime-health.js': ['test/runtime-health.test.js'],
 	// skill-eval hosts the accuracy-lint detectors (auditCommandDocumentation /
 	// auditRouterPrecision), so edits there must also run their detector suite.
 	'lib/skill-eval.js': ['test/skill-eval.test.js', 'test/skill-accuracy.test.js'],
+	// Schema is shared by every kernel authority; keep it on the full-suite lane.
+	'lib/kernel/schema.js': [],
+	// Migrations have consumers across the kernel; an empty mapping deliberately
+	// selects the full-suite lane, matching the shared SQLite driver.
+	'lib/kernel/migrations.js': [],
+	'lib/commands/shepherd.js': [
+		'test/commands/shepherd.test.js',
+		'test/commands/github-route-matrix.test.js',
+		'test/pr-monitor/arm-on-push.test.js',
+		'test/pr-monitor/shepherd-watch.test.js',
+		'test/pr-monitor/watch-owner-cutover.test.js',
+		'test/pr-monitor/watch-owner-launch-contention.test.js',
+		'test/pr-monitor/watch-owner-launch-gate.test.js',
+	],
+	'lib/commands/github.js': [
+		'test/commands/github.test.js',
+		'test/github-launcher.test.js',
+		'test/integration/github-account-context.test.js',
+	],
+	'lib/github-context.js': ['test/github-context.test.js', 'test/integration/github-account-context.test.js'],
+	'lib/commands/_registry.js': [
+		'test/commands/_registry.test.js',
+		'test/commands/_registry-github-context.test.js',
+	],
+	'lib/commands/pr.js': ['test/commands/pr.test.js', 'test/commands/github-route-matrix.test.js'],
+	'lib/commands/ship.js': ['test/commands/ship.test.js', 'test/commands/github-route-matrix.test.js'],
+	'lib/commands/merge.js': ['test/commands/merge.test.js', 'test/commands/github-route-matrix.test.js'],
+	'lib/commands/team.js': ['test/commands/team.test.js', 'test/commands/github-route-matrix.test.js'],
+	'lib/commands/clean.js': ['test/commands/clean.test.js', 'test/commands/github-route-matrix.test.js'],
+	'scripts/github-context-bridge.sh': ['test/commands/github-route-matrix.test.js'],
+	'lib/commands/push.js': ['test/commands/push.test.js', 'test/commands/github-indirect-routes.test.js'],
+	'lib/commands/hooks.js': ['test/commands/hooks.test.js', 'test/hooks-session-start.test.js', 'test/commands/github-indirect-routes.test.js'],
+	'lib/commands/serve.js': ['test/commands/serve.test.js', 'test/commands/github-indirect-routes.test.js'],
+	'lib/commands/skill.js': ['test/commands/skill.test.js', 'test/commands/github-indirect-routes.test.js'],
+	'scripts/lib/behavioral-eval-runtime.js': ['test/eval/behavioral-eval-runtime.test.js', 'test/commands/github-indirect-routes.test.js'],
+	'web/dashboard/generate-snapshot.mjs': ['test/commands/github-indirect-routes.test.js'],
+	// The shared SQLite driver has more consumers than a bounded direct list can
+	// safely represent; its empty mapping deliberately selects the full-suite lane.
+	'lib/kernel/sqlite-driver.js': [],
+	'lib/pr-monitor/watch-owner.js': [
+		'test/pr-monitor/watch-owner.test.js',
+		'test/kernel/watch-owner-transaction.test.js',
+	],
+	'lib/pr-monitor/reconcile.js': [
+		'test/pr-monitor/reconcile.test.js',
+		'test/pr-monitor/reconcile-owner-authority.test.js',
+	],
+	'lib/pr-monitor/reconcile-executor.js': [
+		'test/commands/github-indirect-routes.test.js',
+		'test/pr-monitor/auto-trigger-containment.test.js',
+		'test/pr-monitor/reconcile-executor-owner-authority.test.js',
+		'test/pr-monitor/reconcile-executor.test.js',
+		'test/pr-monitor/reconcile-daemon-owner-authority.test.js',
+		'test/pr-monitor/reconcile-migration-gate.test.js',
+		'test/pr-monitor/watch-owner-launch-contention.test.js',
+		'test/windows-hide-background-spawns.test.js',
+	],
+	'lib/pr-monitor/watch.js': [
+		'test/pr-monitor/watch-owner-cutover.test.js',
+		'test/pr-monitor/watch.test.js',
+	],
+	'lib/pr-monitor/watch-lifecycle.js': [
+		'test/commands/github-indirect-routes.test.js',
+		'test/pr-monitor/watch-lifecycle.test.js',
+		'test/pr-monitor/watch-owner-launch-contention.test.js',
+		'test/pr-monitor/watch-owner-launch-gate.test.js',
+	],
+	'lib/pr-monitor/shepherd-lease.js': [
+		'test/pr-monitor/shepherd-lease.test.js',
+		'test/pr-monitor/shepherd-lease-election-only.test.js',
+	],
+	'lib/pr-monitor/monitor.js': [
+		'test/pr-monitor/monitor.test.js',
+		'test/pr-monitor/monitor-owner-signal.test.js',
+	],
 	'lib/upgrade-safety.js': ['test/commands/upgrade.test.js'],
 	'scripts/test.js': ['test/scripts/test-runner.test.js'],
+	// `bun.lock` is deliberately absent: a dependency change can break anything,
+	// so it stays on the full-suite lane.
+	// lefthook.yml is the repository's hook wiring, so this must also cover the
+	// suites that assert the individual pre-commit/pre-push commands exist.
+	'lefthook.yml': [
+		'test/lefthook-check.test.js',
+		'test/lefthook-wiring.test.js',
+		'test/commands/lefthook-user-config.test.js',
+		'test/branch-protection.test.js',
+		'test/commitlint.test.js',
+		'test/cross-platform-install.test.js',
+		'test/sync-d20-audit-script.test.js',
+	],
+	'eslint.config.js': ['test/lint-script.test.js'],
+	'.coderabbit.yaml': ['test/coderabbit-config.test.js'],
+	'.claude-plugin/marketplace.json': [
+		'test/activation/plugin-scaffold.test.js',
+		'test/plugin-catalog.test.js',
+	],
 });
 
+// Shared authority modules must run the full suite even when another changed file
+// resolves a targeted test. Keep this signal separate from test candidates: an
+// empty candidate list is not enough once a mapped consumer is also changed.
+const FULL_SUITE_REQUIRED_FILES = new Set([
+	'lib/kernel/schema.js',
+	'lib/kernel/migrations.js',
+	'lib/kernel/sqlite-driver.js',
+]);
+
+// Directory-level mappings for source trees that own a dedicated suite. Without
+// these a change under any of them resolves to zero tests and forces the full
+// suite (see hasZeroResolvedTests in scripts/test.js).
+const PREFIX_TEST_TARGETS = Object.freeze([
+	['validation/', ['test/validation/risk-manifest.test.js', 'test/validation/risk-manifest-generator.test.js']],
+	// `eval/corpus/` must come first: the corpus hash and manifest checks live in
+	// immutable-corpus, and eval-sets only reads `eval/commands/`.
+	['eval/corpus/', ['test/eval/immutable-corpus.test.js']],
+	['eval/', ['test/eval/eval-sets.test.js']],
+	['rules/', ['test/rules-sync.test.js', 'test/skill-dispatch-parity.test.js']],
+	['plugin/', ['test/activation/plugin-scaffold.test.js', 'test/plugin-catalog.test.js', 'test/release-readiness.test.js']],
+]);
+
 // Skill sources (canonical `skills/**` and the committed `.agents/skills/**` mirror)
 // map to the fast skill suite. This is what keeps a skills-only PR on the targeted
 // lane instead of the full ~1500-test suite. `skills-sync-drift` guards the mirror.
@@ -84,22 +221,6 @@ function detectPackageManager(projectRoot, fs) {
 	return 'npm';
 }
 
-/**
- * Check if Beads (bd CLI) is reachable.
- *
- * @param {Function} execFileSync - Injected execFileSync
- * @returns {boolean} true if bd is available
- */
-function checkBeadsConnectivity(execFileSync) {
-	try {
-		execFileSync('bd', ['list', '--limit=1'], { timeout: BEADS_CHECK_TIMEOUT });
-		return true;
-	} catch (_e) { // NOSONAR S2486
-		/* intentional: bd not installed or unreachable */
-		return false;
-	}
-}
-
 function resolveBaseBranch(execFileSync) {
 	let baseBranch = 'main';
 	try {
@@ -186,7 +307,9 @@ function getChangedFiles(execFileSync, options = {}) {
  * @returns {string[]} Array of test file paths (e.g. ['test/foo.test.js'])
  */
 function getAffectedTestFiles(projectRoot, execFileSync, fs = defaultFs, options = {}) {
-	const changedFiles = getChangedFiles(execFileSync, options);
+	const changedFiles = Array.isArray(options.changedFiles)
+		? options.changedFiles
+		: getChangedFiles(execFileSync, options);
 	const testFiles = new Set();
 	for (const file of changedFiles) {
 		for (const candidate of getTestCandidatesForChangedFile(file)) {
@@ -196,6 +319,16 @@ function getAffectedTestFiles(projectRoot, execFileSync, fs = defaultFs, options
 		}
 	}
 
+	// Markdown changes must also run the suites that assert on that markdown's
+	// CONTENT (README size badge, AGENTS.md convention, docs consistency). The
+	// static mapping above only knows the docs-consistency suite, so a docs PR
+	// used to merge green and break master for the next code PR (issue 63556816).
+	// The set is derived from the test sources, so new markdown-reading suites are
+	// picked up without editing a list here.
+	for (const candidate of selectDocAssertingTests(changedFiles, projectRoot, fs)) {
+		testFiles.add(candidate);
+	}
+
 	return Array.from(testFiles).sort((left, right) => left.localeCompare(right));
 }
 
@@ -215,10 +348,24 @@ function getTestCandidatesForChangedFile(file) {
 		return [...SKILL_TEST_TARGETS];
 	}
 
+	// Maintainer-only contributor skills. They document the AGENTS.md product /
+	// maintainer split, so the docs-bleed gate is the suite that can catch a
+	// contradiction between them.
+	if (file.startsWith('.forge/contributor-skills/')) {
+		return ['test/structural/agents-docs-bleed.test.js'];
+	}
+
+	for (const [prefix, targets] of PREFIX_TEST_TARGETS) {
+		if (file.startsWith(prefix)) {
+			return [...targets];
+		}
+	}
+
 	if (file === 'README.md'
 		|| file === 'QUICKSTART.md'
 		|| file === 'CHANGELOG.md'
 		|| file === 'AGENTS.md'
+		|| file === 'CODING_STANDARDS.md'
 		|| file === 'DEVELOPMENT.md'
 		|| file === 'docs/INDEX.md'
 		|| file === 'docs/PROJECT_DESIGN.md'
@@ -226,7 +373,9 @@ function getTestCandidatesForChangedFile(file) {
 		|| file.startsWith('docs/guides/')
 		|| file.startsWith('docs/reference/')
 		|| file.startsWith('docs/work/')) {
-		return ['test/docs-consistency.test.js'];
+		return file === 'AGENTS.md'
+			? ['test/docs-consistency.test.js', 'test/structural/agents-docs-bleed.test.js']
+			: ['test/docs-consistency.test.js'];
 	}
 
 	if (file.startsWith('lib/') && file.endsWith('.js')) {
@@ -270,10 +419,15 @@ function getTestCandidatesForChangedFile(file) {
 	return [];
 }
 
+function isFullSuiteRequiredFile(file) {
+	return FULL_SUITE_REQUIRED_FILES.has(file);
+}
+
 module.exports = {
 	getChangedFiles,
 	getAffectedTestFiles,
 	getTestCandidatesForChangedFile,
+	isFullSuiteRequiredFile,
 	name: 'test',
 	description: 'Run tests with smart defaults (timeout, affected-only)',
 	usage: 'forge test [--affected]',
@@ -291,7 +445,7 @@ module.exports = {
 	 * @param {Object} [deps.fs] - fs module
 	 * @param {Function} [deps.execFileSync] - child_process.execFileSync
 	 * @param {Function} [deps.spawnSync] - child_process.spawnSync
-	 * @returns {Promise<{ success: boolean, exitCode: number, beadsSkipped: boolean }>}
+	 * @returns {Promise<{ success: boolean, exitCode: number }>}
 	 */
 	async handler(_args, flags, projectRoot, deps = {}) {
 		const fs = deps.fs || defaultFs;
@@ -313,32 +467,28 @@ module.exports = {
 			}
 		} catch (_e) { /* intentional: package.json missing or unreadable, use default timeout */ } // NOSONAR S2486
 
-		// 3. Check Beads connectivity
-		const beadsAvailable = checkBeadsConnectivity(execFileSync);
-		const beadsSkipped = !beadsAvailable;
-
-		const extraEnv = {};
-		if (beadsSkipped) {
-			extraEnv.BEADS_SKIP_TESTS = '1';
-		}
-
-		// 4. Build test command args
+		// 3. Build test command args
 		let testArgs = ['run', 'test'];
 
-		// 5. --affected flag: find changed test files
+		// 4. --affected flag: find changed test files
 		if (flags['--affected'] || flags.affected) {
-			const affectedTests = getAffectedTestFiles(projectRoot, execFileSync, fs, {
+			const diffOptions = {
 				sinceUpstream: flags.sinceUpstream || flags['--since-upstream'],
+			};
+			const changedFiles = getChangedFiles(execFileSync, diffOptions);
+			const affectedTests = getAffectedTestFiles(projectRoot, execFileSync, fs, {
+				...diffOptions,
+				changedFiles,
 			});
-			if (affectedTests.length > 0) {
+			if (!changedFiles.some(isFullSuiteRequiredFile) && affectedTests.length > 0) {
 				testArgs = ['run', 'test', ...affectedTests];
 			}
 			// If no affected tests found, fall back to running all tests
 		}
 
-		// 6. Run tests
+		// 5. Run tests
 		const result = spawnSync(pkgManager, testArgs, {
-			env: { ...process.env, ...extraEnv },
+			env: { ...process.env },
 			timeout,
 			stdio: 'inherit',
 			shell: process.platform === 'win32',
@@ -349,7 +499,6 @@ module.exports = {
 		return {
 			success: exitCode === 0,
 			exitCode,
-			beadsSkipped,
 		};
 	},
 };

package/lib/commands/validate.js

@@ -44,10 +44,12 @@ const CONFLICT_DIVIDER_PATTERN = /^=======$/;
 const CONFLICT_END_PATTERN = /^>>>>>>>.*$/;
 const GIT_CONFLICT_SCAN_MAX_BUFFER = 16 * 1024 * 1024;
 const VALIDATION_COMMAND_TIMEOUT_MS = 600000;
-const VALIDATION_COMMAND_TIMEOUT_MINUTES = Math.round(VALIDATION_COMMAND_TIMEOUT_MS / 60000);
+// Match the full-suite runner's measured headroom; its verbose output exceeds 1 MiB.
+const FULL_SUITE_TIMEOUT_MS = 25 * 60 * 1000;
+const FULL_SUITE_MAX_BUFFER = 16 * 1024 * 1024;
 
-function validationTimeoutMessage(label) {
-	return `${label} timed out after ${VALIDATION_COMMAND_TIMEOUT_MINUTES} minutes`;
+function validationTimeoutMessage(label, timeoutMs = VALIDATION_COMMAND_TIMEOUT_MS) {
+	return `${label} timed out after ${Math.round(timeoutMs / 60000)} minutes`;
 }
 
 function listConflictScanFiles(rootDir) {
@@ -80,8 +82,15 @@ function shouldSkipConflictScanPath(relativePath) {
 	));
 }
 
-function getExecOptions() {
-	return { encoding: 'utf8', cwd: process.cwd(), timeout: VALIDATION_COMMAND_TIMEOUT_MS };
+/** Build bounded process options, with extra headroom only for Forge's full suite. */
+function getExecOptions(rootDir = process.cwd(), fullSuite = false) {
+	const options = {
+		encoding: 'utf8',
+		cwd: rootDir,
+		timeout: fullSuite ? FULL_SUITE_TIMEOUT_MS : VALIDATION_COMMAND_TIMEOUT_MS,
+	};
+	if (fullSuite) options.maxBuffer = FULL_SUITE_MAX_BUFFER;
+	return options;
 }
 
 const ERROR_PATTERNS = {
@@ -130,7 +139,7 @@ function parseNumber(match, index = 1, defaultValue = 0) {
  */
 function getCheckStatus(check) {
 	if (!check) return null;
-	if (check.skipped) return 'SKIPPED';
+	if (check.skipped === true) return 'SKIPPED';
 	return check.success ? 'PASS' : 'FAIL';
 }
 
@@ -343,17 +352,18 @@ async function runTypeCheck() {
  * Run ESLint
  * Executes eslint . to check code quality
  *
+ * @param {typeof execFileSync} exec - Synchronous process runner
  * @returns {Promise<{success: boolean, duration: number, warnings?: number, errors?: number, message?: string}>} Lint result
  * @example
  * const result = await runLint();
  * console.log(`Warnings: ${result.warnings}, Errors: ${result.errors}`);
  */
-async function runLint() {
+async function runLint(exec = execFileSync) {
 	const startTime = Date.now();
 
 	try {
 		// Run eslint with no output (exit code determines success)
-		execFileSync('eslint', ['.'], getExecOptions());  // NOSONAR S4036 - hardcoded CLI command, no user input, developer tool context
+		exec('eslint', ['.'], getExecOptions());  // NOSONAR S4036 - hardcoded CLI command, no user input, developer tool context
 
 		return {
 			success: true,
@@ -556,7 +566,7 @@ async function runSecurityScan() { // NOSONAR S3776
 
 /**
  * Run all tests
- * Executes bun test to run the test suite
+ * Uses Forge's full-suite runner when present, with raw Bun as an external-repository fallback
  *
  * @returns {Promise<{success: boolean, duration: number, passed: number, failed: number, total: number, message?: string}>} Test execution result
  * @example
@@ -565,6 +575,22 @@ async function runSecurityScan() { // NOSONAR S3776
  */
 function parseTestCounts(output) {
 	const text = String(output || '');
+	const aggregateMatches = [...text.matchAll(
+		/Full suite aggregate: status=(PASS|FAIL|INCOMPLETE) tests=(\d+) assertions=\d+ passed=(\d+) failed=(\d+) errors=(\d+) skipped=(\d+)/g,
+	)]; // NOSONAR S5852 - fixed tokens and bounded numeric fields
+	const aggregate = aggregateMatches.at(-1);
+	if (aggregate) {
+		const total = parseNumber(aggregate, 2);
+		return {
+			status: aggregate[1],
+			passed: parseNumber(aggregate, 3),
+			failed: parseNumber(aggregate, 4),
+			errors: parseNumber(aggregate, 5),
+			skipped: parseNumber(aggregate, 6),
+			total,
+			testsFound: total > 0,
+		};
+	}
 	const passed = parseNumber(/(\d+) pass/.exec(text));  // NOSONAR S5852 - bounded \d+ pattern
 	const failed = parseNumber(/(\d+) fail/.exec(text));  // NOSONAR S5852 - bounded \d+ pattern
 	const skipped = parseNumber(/(\d+) skip/.exec(text));  // NOSONAR S5852 - bounded \d+ pattern
@@ -573,13 +599,31 @@ function parseTestCounts(output) {
 	return { passed, failed, skipped, total, testsFound: total > 0 };
 }
 
-async function runAllTests(exec = execFileSync) {
+function isForgeCheckout(rootDir) {
+	try {
+		const manifest = JSON.parse(fs.readFileSync(path.join(rootDir, 'package.json'), 'utf8'));
+		return manifest.name === 'forge-workflow'
+			&& manifest.bin?.forge === 'bin/forge.js'
+			&& manifest.scripts?.['test:full:parallel'] === 'node scripts/test-full-suite.js'
+			&& fs.existsSync(path.join(rootDir, 'scripts', 'test-full-suite.js'));
+	} catch (error) {
+		if (error instanceof SyntaxError || error.code === 'ENOENT' || error.code === 'ENOTDIR') return false;
+		throw error;
+	}
+}
+
+async function runAllTests(exec = execFileSync, rootDir = process.cwd()) {
 	const startTime = Date.now();
+	const useFullSuite = isForgeCheckout(rootDir);
+	const command = useFullSuite ? 'node' : 'bun';
+	const args = useFullSuite ? ['scripts/test-full-suite.js'] : ['test', '--timeout', '30000'];
+	const execOptions = getExecOptions(rootDir, useFullSuite);
 
 	try {
-		const result = exec('bun', ['test', '--timeout', '30000'], getExecOptions());  // NOSONAR S4036 - hardcoded CLI command, no user input, developer tool context
+		const result = exec(command, args, execOptions);  // NOSONAR S4036 - fixed CLI commands, no user input, developer tool context
 
-		const { passed, failed, skipped, total, testsFound } = parseTestCounts(result);
+		const { status, passed, failed, errors = 0, skipped, total, testsFound } = parseTestCounts(result);
+		const success = failed === 0 && errors === 0 && (status === undefined || status === 'PASS');
 
 		// B2 (N1): a bun run that executes 0 tests must NEVER report PASS — that
 		// manufactures false confidence. Surface it as an explicit SKIP instead.
@@ -597,27 +641,27 @@ async function runAllTests(exec = execFileSync) {
 		}
 
 		return {
-			success: failed === 0,
+			success,
 			testsFound: true,
 			duration: Date.now() - startTime,
 			passed,
 			failed,
 			skipped,
 			total,
-			message: failed === 0
+			message: success
 				? `All ${total} tests passed`
-				: `${failed}/${total} tests failed`,
+				: `${failed + errors}/${total} tests failed`,
 		};
 	} catch (error) {
 		// Check for timeout
-		if (error.killed && error.signal === 'SIGTERM') {
+		if (error.code === 'ETIMEDOUT' || (error.killed && error.signal === 'SIGTERM')) {
 			return {
 				success: false,
 				duration: Date.now() - startTime,
 				passed: 0,
 				failed: 0,
 				total: 0,
-				message: validationTimeoutMessage('Test execution'),
+				message: validationTimeoutMessage('Test execution', execOptions.timeout),
 			};
 		}
 
@@ -631,7 +675,7 @@ async function runAllTests(exec = execFileSync) {
 				passed: 0,
 				failed: 0,
 				total: 0,
-				message: 'Tests skipped: bun test runner not found. Install bun to run tests.',
+				message: 'Tests skipped: test runner not found. Install Node.js and Bun to run tests.',
 			};
 		}
 
@@ -641,7 +685,7 @@ async function runAllTests(exec = execFileSync) {
 		const passed = counts.passed;
 		// bun exited non-zero: if the output shows no explicit failure count, treat
 		// it as at least one failure rather than swallowing the non-zero exit.
-		const failed = counts.failed > 0 ? counts.failed : Math.max(1, counts.failed);
+		const failed = counts.failed > 0 ? counts.failed : Math.max(1, counts.errors || 0);
 		const total = counts.total > 0 ? counts.total : passed + failed + counts.skipped;
 
 		return {
@@ -753,8 +797,8 @@ async function executeValidate(options = {}) { // NOSONAR S3776
 	// 4. Tests
 	if (!skip.includes(CHECK_TYPES.TESTS)) {
 		try {
-			checks.tests = await runAllTests();
-			if (!checks.tests.success && !checks.tests.skipped) {
+			checks.tests = await runAllTests(execFileSync, rootDir);
+			if (!checks.tests.success && checks.tests.skipped !== true) {
 				failedChecks.push(CHECK_TYPES.TESTS);
 				if (!continueOnError) {
 					return buildResult(checks, failedChecks, errors, startTime);