Merge branch 'next' into fix/2046-config-unset-null

This commit is contained in:
Tom Boucher
2026-07-07 13:26:16 -04:00
committed by GitHub
50 changed files with 924 additions and 214 deletions

View File

@@ -178,7 +178,10 @@ function timeAgo(date: Date): string {
function extractCanonicalPlanId(filename: string): string {
const base = filename.replace(/-PLAN\.md$/i, '').replace(/-SUMMARY\.md$/i, '').replace(/\.md$/i, '');
const parts = base.split('-').filter(Boolean);
const tokenRe = /^\d+[A-Z]?(?:\.\d+)*$/i;
// #2043: a phase/plan token component is either a zero-padded number (≥2 digits)
// or a single-digit-plus-letter id ("3A"); a *bare* single digit is a slug word,
// so "46-6-rs-…" is not paired into a "46-6" id while "3A-01" stays intact.
const tokenRe = /^(?:\d{2,}[A-Z]?|\d[A-Z])(?:\.\d+)*$/i;
const phaseIdx = parts.findIndex(p => tokenRe.test(p));
if (phaseIdx >= 0 && phaseIdx + 1 < parts.length && tokenRe.test(parts[phaseIdx + 1])) {
return `${parts[phaseIdx]}-${parts[phaseIdx + 1]}`;

View File

@@ -0,0 +1,198 @@
/**
* Test-command normalizer (#1857).
*
* A GSD verification gate resolves a project's test command and runs it. When
* that command is a watch/dev-mode runner (vitest defaults to WATCH in an
* interactive TTY — which is exactly a user running `gsd-execute-phase` in their
* terminal — and jest with `--watch`/`--watchAll`), the process never exits and
* the orchestrator waits forever.
*
* `normalizeTestCommand` rewrites a resolved command to a best-effort one-shot
* form so a gate cannot hang on watch mode. It is intentionally conservative:
* - it NEVER double-adds flags (already-one-shot commands are returned verbatim),
* - it only touches commands it positively recognises as a watch runner,
* - anything it cannot classify is returned unchanged.
* The gate's wall-clock `timeout` is the ultimate guarantee for anything this
* best-effort pass cannot defeat (e.g. an explicit `--watch` baked into a
* project's `test` script, which even `CI=1` cannot override per vitest docs).
*
* Bounded by design (#1857 security review): the input is length-capped, all
* scanning is linear-time (no super-linear regex backtracking), and package.json
* is only read when it is a regular file — so normalization itself can never
* hang or take super-linear time on an adversarial `workflow.test_command`.
*
* Single source of truth: all four test-command gates (regression, post-merge,
* audit-fix, verify-phase) route their resolved command through this helper so
* the paths cannot drift.
*
* Leaf module — depends only on node:fs / node:path.
*/
import fs from 'node:fs';
import path from 'node:path';
// A resolved test command is never realistically this long; anything larger is
// not a real runner invocation. We skip normalization above this bound (the
// gate's own timeout still bounds the actual run) so no regex ever scans an
// adversarial multi-KB string.
const MAX_COMMAND_LENGTH = 4096;
// Markers proving a command is already one-shot / non-watch. If any is present
// we return the command unchanged so we never double-add flags or fight an
// explicit user choice.
const ONE_SHOT_MARKERS: RegExp[] = [
/(?:^|\s)vitest\s+run\b/, // vitest run …
/(?:^|\s)--run\b/, // vitest --run
/(?:^|\s)--no-watch\b/, // vitest --no-watch
/(?:^|\s)--watchAll=false\b/, // jest --watchAll=false
/(?:^|\s)--watch=false\b/,
/(?:^|\s)--ci\b/, // jest --ci
/^\s*CI=/, // already forced into CI/run mode via env
];
/** True if the command already runs one-shot (so normalization is a no-op). */
function isAlreadyOneShot(cmd: string): boolean {
return ONE_SHOT_MARKERS.some((re) => re.test(cmd));
}
// Match a runner as a standalone command TOKEN (whitespace-delimited), NOT as a
// substring — so "vitest.config.js" / "jest-environment" / "node vitest-x.js"
// are not treated as the runner and are never mangled.
const VITEST_TOKEN = /(?:^|\s)vitest(?=\s|$)/;
const JEST_TOKEN = /(?:^|\s)jest(?=\s|$)/;
/**
* Linear-time detection of a package-manager `test` script invocation
* (`npm test`, `pnpm run test`, `pnpm --dir app test`, `yarn test`, …). We split
* on shell separators FIRST (linear), then test each bounded segment with simple
* anchored regexes — no tempered-greedy scan, so no super-linear backtracking on
* adversarial input.
*/
function isScriptInvocation(cmd: string): boolean {
return cmd.split(/&&|\|\||;/).some((seg) => {
const s = seg.trim();
return /^(?:npm|pnpm|yarn|bun)\b/.test(s) && /\btest\b/.test(s);
});
}
/**
* Resolve the package.json directory for a command that may target a different
* working dir via `--dir <p>` (pnpm), `-C <p>` (pnpm), or `--prefix <p>` (npm).
*/
function resolvePackageDir(cmd: string, cwd: string): string {
const m = cmd.match(/(?:--dir|--prefix|-C)[=\s]+(\S+)/);
if (m && m[1]) {
const p = m[1].replace(/^['"]|['"]$/g, '');
return path.isAbsolute(p) ? p : path.join(cwd, p);
}
return cwd;
}
type WatchRunner = 'vitest' | 'jest' | null;
/**
* Inspect package.json `scripts.test` for the given command's target dir and
* report whether it resolves to a watch-by-default / explicitly-watching runner.
* Only ever reads `scripts.test` as a STRING for classification — it is never
* executed or spliced into the output.
*/
function scriptTestRunner(cmd: string, cwd: string): WatchRunner {
try {
const pkgPath = path.join(resolvePackageDir(cmd, cwd), 'package.json');
// Only read a REGULAR file — never block on a FIFO/socket/dir named
// "package.json" reachable via --dir (#1857 security review).
let stat: fs.Stats;
try {
stat = fs.statSync(pkgPath);
} catch {
return null;
}
if (!stat.isFile()) return null;
const pkg = JSON.parse(fs.readFileSync(pkgPath, 'utf-8')) as {
scripts?: Record<string, string>;
};
const testScript = pkg.scripts?.test ?? '';
if (!testScript) return null;
// vitest watches by default unless the script itself is already one-shot.
if (VITEST_TOKEN.test(testScript) && !isAlreadyOneShot(testScript)) return 'vitest';
// jest only watches when explicitly asked to.
if (
JEST_TOKEN.test(testScript) &&
/(?:^|\s)--watch(?:All)?\b/.test(testScript) &&
!/--watch(?:All)?=false\b/.test(testScript)
) {
return 'jest';
}
return null;
} catch {
return null;
}
}
/** Strip a bare/explicit-true `--watch` / `--watchAll` (never `=false`). */
function stripWatchFlags(cmd: string): string {
return cmd
.replace(/(?:^|\s)--watch(?:All)?(?:=true)?(?=\s|$)/g, ' ')
.replace(/\s{2,}/g, ' ')
.trim();
}
/**
* Normalize a resolved test command to a best-effort one-shot form.
*
* @param rawCmd The resolved test command (e.g. from `config-get
* workflow.test_command` or the gate's runner sniff).
* @param cwd Directory used to locate package.json for script invocations.
* @returns A one-shot command, or `rawCmd` unchanged when it is already
* one-shot / not a recognised watch runner / too long to be a real runner.
*/
function normalizeTestCommand(rawCmd: string, cwd: string): string {
const cmd = (rawCmd ?? '').trim();
if (!cmd || cmd === 'true') return rawCmd;
if (cmd.length > MAX_COMMAND_LENGTH) return rawCmd; // bound all downstream scanning
if (isAlreadyOneShot(cmd)) return rawCmd;
const isScript = isScriptInvocation(cmd);
// 1. Direct vitest invocation (`vitest`, `npx vitest`, `pnpm exec vitest`, …):
// force the explicit one-shot `run` command and drop any bare --watch.
if (VITEST_TOKEN.test(cmd) && !isScript) {
return stripWatchFlags(cmd).replace(/(^|\s)vitest(?=\s|$)/, '$1vitest run');
}
// 2. Direct jest invocation with an explicit watch flag: make it one-shot.
if (JEST_TOKEN.test(cmd) && !isScript) {
if (/(?:^|\s)--watch(?:All)?\b/.test(cmd)) {
return `${stripWatchFlags(cmd)} --watchAll=false`;
}
return rawCmd; // jest without --watch already runs once
}
// 3. Package-manager `test` script invocation: the runner is inside
// package.json. If it resolves to a watch runner, force CI/run mode via
// the CI env — robust across --dir and pnpm/yarn `--` propagation quirks
// (vitest & jest both switch to run/non-interactive mode when CI is set).
if (isScript) {
const runner = scriptTestRunner(cmd, cwd);
if (runner === 'vitest' || runner === 'jest') {
return `CI=true ${cmd}`;
}
return rawCmd;
}
return rawCmd;
}
/**
* CLI handler for `gsd-tools query normalize-test-command <raw-cmd>`: prints the
* normalized one-shot command to stdout (the gates capture it via `$(…)`).
*/
function cmdNormalizeTestCommand(cwd: string, rawCmd: string | undefined): void {
process.stdout.write(normalizeTestCommand(rawCmd ?? '', cwd));
}
export = {
normalizeTestCommand,
cmdNormalizeTestCommand,
isAlreadyOneShot,
};

View File

@@ -205,9 +205,29 @@ function extractPhaseToken(dirName: string): string {
const segments = rest.split('-');
const tokenSegments: string[] = [];
// #2043: distinguish a real (zero-padded, ≥2-digit) phase/sub-phase segment
// from a single-digit slug word. A pure-numeric leading segment ("46") only
// continues with ≥2-digit segments, so "46-6-rs-…" yields "46" (the "6" is the
// slug's first word), not "46-6". Milestone-prefixed ids like "M1-2" reach here
// with "M1-" already stripped as a project-code prefix (see
// PROJECT_CODE_PREFIX_CAPTURE_RE_I), so "2" is the leading segment and the same
// pure-numeric rule applies (M1-46-6-rs → "M1-46"). The firstLetterPrefixed
// carve-out covers letter+digit leading segments that survive prefix stripping
// because of punctuation (e.g. "P0.3-2"), whose single-digit continuation is
// intentionally preserved (unchanged from prior behaviour).
let firstLetterPrefixed = false;
for (let i = 0; i < segments.length; i++) {
const seg = segments[i];
if (/^\d/.test(seg) || (i === 0 && /^[A-Za-z]{1,3}\d/.test(seg))) {
if (i === 0) {
if (/^\d/.test(seg)) {
tokenSegments.push(seg);
} else if (/^[A-Za-z]{1,3}\d/.test(seg)) {
tokenSegments.push(seg);
firstLetterPrefixed = true;
} else {
break;
}
} else if (/^\d{2,}/.test(seg) || (firstLetterPrefixed && /^\d/.test(seg))) {
tokenSegments.push(seg);
} else {
break;

View File

@@ -103,7 +103,10 @@ function extractCanonicalPlanId(filename: string): string {
.replace(/-SUMMARY\.md$/i, '')
.replace(/\.md$/i, '');
const parts = base.split('-').filter(Boolean);
const tokenRe = /^\d+[A-Z]?(?:\.\d+)*$/i;
// #2043: a phase/plan token component is either a zero-padded number (≥2 digits)
// or a single-digit-plus-letter id ("3A"); a *bare* single digit is a slug word,
// so "46-6-rs-…" is not paired into a "46-6" id while "3A-01" stays intact.
const tokenRe = /^(?:\d{2,}[A-Z]?|\d[A-Z])(?:\.\d+)*$/i;
const phaseIdx = parts.findIndex((p) => tokenRe.test(p));
if (phaseIdx >= 0 && phaseIdx + 1 < parts.length && tokenRe.test(parts[phaseIdx + 1])) {
return `${parts[phaseIdx]}-${parts[phaseIdx + 1]}`;

View File

@@ -469,8 +469,12 @@ function getMilestonePhaseFilter(cwd: string, versionOverride?: string | null, p
}
const roadmapUsesHyphenedIds = [...normalized].some(n => n.includes('-'));
// #2043: milestone-prefixed sub-phase components must be zero-padded (≥2 digits)
// — "-\d{2,}" instead of "-0*\d+" — so a single-digit slug word after the phase
// number (e.g. dir "46-6-rs-…") captures "46" and is not silently excluded from
// the milestone as a bogus "46-6" id.
const numericRe = roadmapUsesHyphenedIds
? /^0*(\d+(?:-0*\d+)*[A-Za-z]?(?:\.\d+)*)/
? /^0*(\d+(?:-\d{2,})*[A-Za-z]?(?:\.\d+)*)/
: /^0*(\d+[A-Za-z]?(?:\.\d+)*)/;
function isDirInMilestone(dirName: string): boolean {

View File

@@ -44,16 +44,25 @@ export const phaseDirNameRe = new RegExp(
);
// Extracts the full phase token from a directory name, including milestone-prefixed
// multi-segment tokens like "02-01" from "02-01-setup" or "GSD-02-01-setup".
// Greedily captures all leading all-digit segments before the first letter-start segment.
// #2043: a *continuation* sub-phase segment must be zero-padded (≥2 digits), so a
// single-digit slug word after a phase number (e.g. "46-6-rs-…", slug "6 Rs …") is
// NOT absorbed — it captures "46", not "46-6". The first component stays "\d+"
// (with the "[A-Z]?" suffix) so single-digit letter-suffixed phase ids ("1A") and
// milestone-prefixed single-digit sub-phases ("M1-2" → prefix "M1-" stripped, then
// "2") still match. The trailing boundary "(?:-|$)" (was "(?:-[a-z]|$)") lets a slug
// that starts with a digit terminate the token.
export const PHASE_TOKEN_FROM_DIR_RE = new RegExp(
`^${OPTIONAL_PROJECT_CODE_PREFIX_SOURCE}(\\d+(?:-\\d+)*[A-Z]?(?:\\.\\d+)*)(?:-[a-z]|$)`,
`^${OPTIONAL_PROJECT_CODE_PREFIX_SOURCE}(\\d+(?:-\\d{2,})*[A-Z]?(?:\\.\\d+)*)(?:-|$)`,
'i',
);
export const MILESTONE_ARCHIVE_DIR_RE = /^v\d+.*-phases$/i;
// ── Issue #26: I001 canonicalization ────────────────────────────────────────
export function canonicalPlanStem(stem: string): string {
const m = stem.match(/^(\d+[A-Z]?(?:\.\d+)*-\d+)/i);
// #2043: the plan component (after the phase number) must be zero-padded
// (≥2 digits), so a digit-leading slug word (e.g. "46-6-rs-…") is not mistaken
// for a "46-6" phase/plan pair.
const m = stem.match(/^(\d+[A-Z]?(?:\.\d+)*-\d{2,})/i);
return m ? m[1] : stem;
}