cleartoship 0.10.1 → 0.10.3

package/README.md

@@ -98,7 +98,7 @@ runtimes, React Native, Go and shell.
 
 27 upstream rules are **superseded** where ClearToShip's own AST check is more
 precise, 6 are **withheld** as measurably noisy, 10 React Native rules are
-**skipped as inapplicable** on a project that is not React Native, 9 carry a
+**skipped as inapplicable** on a project that is not React Native, 11 carry a
 **match guard** for a
 shape their regex cannot exclude (a `"link": true` lockfile entry has no
 integrity hash by design; `querySelectorAll` is not a SQL call; `eval()` inside
@@ -218,7 +218,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v7
-      - uses: murtazaozdemir/cleartoship@v0.10.1
+      - uses: murtazaozdemir/cleartoship@v0.10.3
         with:
           fail-on: critical
           comment: true
@@ -240,7 +240,7 @@ above `fail-on`) for use in later steps. The comment is *sticky* — re-runs edi
 the same comment instead of piling up.
 
 By default the action runs the scanner version its own ref declares, so
-`@v0.10.1` runs `cleartoship@0.10.1` and pinning the ref pins the behaviour. If
+`@v0.10.3` runs `cleartoship@0.10.3` and pinning the ref pins the behaviour. If
 that version is not on the registry, it builds from its own checkout instead, so
 `uses: …@ref` works against an unpublished commit.
 
@@ -290,6 +290,12 @@ uploaded, and no database is connected to.
   excluded: one is a published CLI's entry point, the other is production schema.
   Across five dogfooded repos this moved 21 findings out of `critical` without
   hiding one of them.
+- **A CVE that only runs on a build machine is not a shipping vulnerability.**
+  CTS024 already split those when OSV answers; the vendored CVE rules — what
+  runs under `--offline` — now make the same split, dropping a match under
+  `devDependencies` (or a lockfile entry marked `"dev": true`) to `low` with the
+  reason attached. The same advisory against a dependency your users run keeps
+  its full severity.
 - **A bound parameter is not an injection.** `db.prepare(\`UPDATE ${table} SET
   csv = ? WHERE id = ?\`).bind(...)` interpolates an identifier while its values
   go through placeholders — the correct pattern, and the one a SQL-injection
@@ -297,6 +303,12 @@ uploaded, and no database is connected to.
   call chained to it, and stand down when the values are bound. They still fire
   when any interpolation reads from the request, so a query that binds one value
   and concatenates another is reported.
+- **A match in a comment is prose about code, not code.** The scanner lexes each
+  file once for strings and comments — properly, tracking quotes, so a URL
+  inside a string is not mistaken for the start of one — and a community rule
+  that matched inside a comment is dropped. Our own source found this: a comment
+  reading *"merely calling `jwt.verify(token, secret)`"* was reported as a JWT
+  vulnerability.
 - **A rule that cannot apply here is not run.** The vendored ruleset covers
   ground this project may not stand on: certificate pinning and WebView
   hardening are React Native concerns, and a browser will not let a page pin a

package/action.yml

@@ -29,7 +29,7 @@ inputs:
   version:
     description: >-
       Version of the cleartoship npm package to run. Defaults to the version this
-      action's own ref declares, so `uses: …@v0.10.1` runs cleartoship@0.10.1. Set
+      action's own ref declares, so `uses: …@v0.10.3` runs cleartoship@0.10.3. Set
       `latest` to always track the newest release, or `local` to build from the checkout.
     required: false
     default: ''
@@ -76,7 +76,7 @@ runs:
       run: |
         # With no version pinned, run the exact version this action's checkout
         # declares. That keeps the action ref and the scanner in lockstep:
-        # `uses: <owner>/cleartoship@v0.10.1` runs cleartoship@0.10.1 instead of
+        # `uses: <owner>/cleartoship@v0.10.3` runs cleartoship@0.10.3 instead of
         # whatever npm happens to tag `latest` at the time.
         ver="$INPUT_VERSION"
         if [ -z "$ver" ]; then

package/dist/scanners/community.js

@@ -1,7 +1,8 @@
 import { read, rel, lineAt, snippetAt, languagesFor } from '../utils/files.js';
 import { Suppressions } from '../utils/suppress.js';
 import { adjustForPath } from '../utils/paths.js';
-import { GUARDVIBE_RULES, GUARDVIBE_ATTRIBUTION, GUARDVIBE_REACT_NATIVE_RULE_IDS, } from '../vendor/guardvibe/index.js';
+import { commentStyleFor, lexSpans, isInside } from '../utils/spans.js';
+import { GUARDVIBE_RULES, GUARDVIBE_ATTRIBUTION, GUARDVIBE_REACT_NATIVE_RULE_IDS, GUARDVIBE_CVE_RULE_IDS, } from '../vendor/guardvibe/index.js';
 import { emptyResult } from '../types.js';
 /**
  * Upstream rules that restate a check ClearToShip already performs against the
@@ -160,6 +161,15 @@ function isParameterized(statement) {
     // '?').join(',')})`, or a constant SQL fragment interpolated beside binds.
     if (/\.\s*bind\s*\(\s*[^)\s]/.test(statement))
         return true;
+    // The other calling convention: the values follow the query.
+    // `$executeRawUnsafe(\`… VALUES ${placeholders}\`, ...params)`,
+    // `db.query(sql, [id])`. A closing backtick with an argument after it.
+    if (/`\s*,\s*[^)\s]/.test(statement))
+        return true;
+    // An interpolation that builds placeholders is parameterizing too:
+    // `chunk.map(() => "(?, ?, ?)").join(", ")`.
+    if (/\$\{[^}]*['"][^'"]*\?[^'"]*['"]/.test(statement))
+        return true;
     // Otherwise look for the placeholders themselves — with interpolations
     // dropped first, so a JavaScript ternary is not read as a `?` parameter.
     return BIND_PLACEHOLDER.test(statement.replace(/\$\{[^}]*\}/g, ' '));
@@ -171,26 +181,6 @@ function sqlGuard(match, source, index) {
         return false;
     return !isParameterized(statementAround(source, index));
 }
-/** Whether `index` falls inside a quoted string on its own line. */
-function insideStringLiteral(source, index) {
-    const lineStart = source.lastIndexOf('\n', index - 1) + 1;
-    let quote = null;
-    for (let i = lineStart; i < index; i++) {
-        const ch = source[i];
-        if (ch === '\\') {
-            i++;
-            continue;
-        }
-        if (quote) {
-            if (ch === quote)
-                quote = null;
-        }
-        else if (ch === '"' || ch === "'" || ch === '`') {
-            quote = ch;
-        }
-    }
-    return quote !== null;
-}
 /**
  * Per-rule filters for a match shape the upstream regex cannot exclude on its
  * own. Given the matched text plus where it sat, so a guard can look around it.
@@ -229,6 +219,35 @@ const MATCH_GUARDS = {
     // interpolated expression may read from the request.
     VG010: (match, source, index) => sqlGuard(match, source, index),
     VG123: (_match, source, index) => !isParameterized(statementAround(source, index)),
+    // The "base64 payload" test is a run of 20+ characters from the base64
+    // alphabet, which any long camelCase identifier satisfies:
+    // `description: \`${pct(clusteredAroundMedian, …)}\`` matched on the
+    // identifier. Interpolated expressions are code, not the description text,
+    // and real encoded content is not purely alphabetic.
+    VG881: (match) => {
+        const text = match.replace(/\$\{[^}]*\}/g, ' ');
+        if (/(?:\\x[0-9a-f]{2}){4,}|(?:\\u[0-9a-f]{4}){4,}|(?:&#\d{2,4};){4,}/i.test(text))
+            return true;
+        const run = /[A-Za-z0-9+/]{20,}={0,2}/.exec(text)?.[0];
+        // A slash is not evidence: "new/used/refurbished" is twenty characters of
+        // the base64 alphabet and a sentence. Real encoded content carries digits
+        // or padding.
+        return run !== undefined && /[0-9+=]/.test(run);
+    },
+    // `eval("require")` is the documented escape hatch for keeping a bundler from
+    // statically resolving a require — a constant the author typed, with no input
+    // reaching it. Dynamic code execution is about the dynamic part.
+    VG014: (match, source, index, spans) => {
+        if (isInside(spans, index, 'string'))
+            return false;
+        const after = source.slice(index, index + 60);
+        return !/^(?:eval|new\s+Function)\s*\(\s*(['"])[A-Za-z_$][\w$]*\1\s*\)/.test(after);
+    },
+    // `$executeRawUnsafe` is named for who builds the SQL string, not for whether
+    // values can be bound — Prisma takes positional parameters after the query.
+    // `$executeRawUnsafe(\`INSERT … VALUES ${placeholders}\`, ...params)`, where
+    // `placeholders` is `chunk.map(() => "(?, ?)").join(",")`, binds every value.
+    VG433: (_match, source, index) => !isParameterized(statementAround(source, index)),
     // SSRF is a *server* being made to fetch a URL it should not. A module marked
     // `'use client'` runs in the browser, where the request leaves the user's own
     // machine and crosses no trust boundary of yours.
@@ -245,11 +264,6 @@ const MATCH_GUARDS = {
     // query filtered to the caller's own rows does not let them. An unbounded
     // fetch of your own data is a scalability question, not a security finding.
     VG955: (match) => !/\bwhere\b[\s\S]{0,200}?\b(userId|user_id|ownerId|owner_id|orgId|org_id|organizationId|tenantId|tenant_id|workspaceId|workspace_id|accountId|account_id|teamId|team_id|shop|shopDomain|storeId|store_id)\b/i.test(match),
-    // `description: 'eval() executes arbitrary code…'` is prose about eval, not a
-    // call to it — and security tooling, which is a good deal of what gets
-    // scanned, is full of that prose. Code held in a string is not code running
-    // here; the eval that would run it is its own match, outside the quotes.
-    VG014: (_match, source, index) => !insideStringLiteral(source, index),
 };
 /** Regexes over very large files are where catastrophic backtracking bites. */
 const MAX_BYTES = 400_000;
@@ -280,6 +294,38 @@ function inapplicable(ctx) {
     }
     return { ids, why };
 }
+/**
+ * Whether a dependency-manifest match sits under `devDependencies`, or in a
+ * lockfile entry marked `"dev": true`. Both mean the package is a build-time
+ * tool that no user ever runs — the split CTS024 already makes for CVEs found
+ * through OSV, applied to the vendored CVE rules that run when offline.
+ */
+function inDevDependencies(source, index) {
+    const before = source.slice(Math.max(0, index - 4000), index);
+    if (/"dev"\s*:\s*true[\s\S]{0,600}$/.test(before))
+        return true;
+    const nearest = /"(dev|peer|optional)?[dD]ependencies"\s*:\s*\{(?![\s\S]*"[a-z]*[dD]ependencies"\s*:\s*\{)/.exec(before);
+    return nearest?.[1] === 'dev';
+}
+/**
+ * Rules whose upstream severity is right for one shape they match and wrong for
+ * another. Returning null leaves the rule's own severity alone.
+ */
+const SEVERITY_ADJUSTERS = {
+    // The rule matches two different things. Explicitly accepting `alg: none` is
+    // the critical it is named for. Merely calling `jwt.verify(token, secret)`
+    // without pinning `algorithms` is not: jsonwebtoken has rejected `none` on a
+    // keyed verify since v9, so what is left is defence against algorithm
+    // confusion — worth doing, not worth blocking a deploy over.
+    VG105: (match) => /algorithms\s*:\s*\[\s*['"]none['"]/i.test(match)
+        ? null
+        : {
+            severity: 'medium',
+            note: ' (Reported at medium: no `algorithms` option is pinned, but nothing here accepts ' +
+                '`alg: none` — a keyed `jwt.verify` rejects it. Pinning the algorithm is defence ' +
+                'against algorithm confusion, which matters most when the key could be a public key.)',
+        },
+};
 export const communityScanner = {
     name: `Community ruleset (${GUARDVIBE_RULES.length - SUPERSEDED.size - WITHHELD.size} rules)`,
     applies() {
@@ -300,6 +346,7 @@ export const communityScanner = {
                 continue;
             const relPath = rel(ctx.root, file);
             const lockfile = LOCKFILE.test(relPath);
+            const spans = lexSpans(source, commentStyleFor(languages));
             const suppress = new Suppressions(source);
             filesScanned++;
             for (const rule of active) {
@@ -320,7 +367,15 @@ export const communityScanner = {
                     }
                     // Skipping a match must not skip the non-global `break` below, or a
                     // rule without /g would rescan from zero forever.
-                    if (guard && !guard(m[0], source, m.index)) {
+                    // A rule that matched inside a comment matched prose about code, not
+                    // code. Nothing in the vendored ruleset targets comment content, and
+                    // a commented-out call is not a call.
+                    if (isInside(spans, m.index, 'comment')) {
+                        if (!re.global)
+                            break;
+                        continue;
+                    }
+                    if (guard && !guard(m[0], source, m.index, spans)) {
                         if (!re.global)
                             break;
                         continue;
@@ -329,12 +384,27 @@ export const communityScanner = {
                     const key = `${relPath}:${line}:${rule.id}`;
                     if (!seen.has(key) && !suppress.suppressed(line, rule.id)) {
                         seen.add(key);
-                        const placed = adjustForPath(severityOf(rule.severity), relPath);
+                        let adjusted = SEVERITY_ADJUSTERS[rule.id]?.(m[0], source, m.index) ?? null;
+                        // A CVE in something that only ever runs on a build machine is not
+                        // a shipping vulnerability. OSV-sourced findings are already split
+                        // this way (CTS024); this is the same split for the vendored CVE
+                        // rules, which are what runs with --offline.
+                        if (!adjusted &&
+                            GUARDVIBE_CVE_RULE_IDS.has(rule.id) &&
+                            (lockfile || /(^|\/)package\.json$/.test(relPath)) &&
+                            inDevDependencies(source, m.index)) {
+                            adjusted = {
+                                severity: 'low',
+                                note: ' (Reported at low: this version is declared under devDependencies, so it is a ' +
+                                    'build-time tool rather than something your users run.)',
+                            };
+                        }
+                        const placed = adjustForPath(adjusted?.severity ?? severityOf(rule.severity), relPath);
                         result.findings.push({
                             id: rule.id,
                             severity: placed.severity,
                             title: rule.name,
-                            detail: rule.description + placed.note,
+                            detail: rule.description + (adjusted?.note ?? '') + placed.note,
                             fix: rule.fixCode ? `${rule.fix}\n\n${rule.fixCode}` : rule.fix,
                             file: relPath,
                             line,

package/dist/utils/gitignore.d.ts

@@ -17,5 +17,11 @@ export declare function extendedAt(parent: Gitignore, dir: string, read: (path:
  * Repository-local excludes. Same syntax, same precedence as a root
  * `.gitignore`, but kept out of version control — so a machine-specific
  * scratch directory is invisible here without appearing in anyone's diff.
+ *
+ * The user's *global* ignore file (`core.excludesFile`, usually
+ * `~/.config/git/ignore`) is deliberately not read: it lives outside the scan
+ * root, which SECURITY.md promises we do not touch, and it is per-machine — a
+ * CI checkout would not have it, so honouring it would make a local scan
+ * quieter than the one that gates the merge.
  */
 export declare function repositoryExcludes(root: string, read: (path: string) => string | null): Gitignore;

package/dist/utils/gitignore.js

@@ -12,6 +12,10 @@ function translate(pattern) {
             continue;
         }
         if (ch === '*') {
+            // `****` means no more than `**` does, and collapsing it keeps the
+            // translated regex free of the nested quantifiers that backtrack badly.
+            while (pattern[i + 1] === '*' && pattern[i + 2] === '*')
+                i++;
             const doubled = pattern[i + 1] === '*';
             if (doubled) {
                 const atStart = i === 0 || pattern[i - 1] === '/';
@@ -51,11 +55,22 @@ function translate(pattern) {
     }
     return out;
 }
+/**
+ * A pattern longer than this is not a real ignore rule. The cap exists because
+ * this text comes from the repository being scanned, which for a security tool
+ * is untrusted input: it reaches a regex compiler, and a regex compiler is a
+ * place where hostile input has leverage.
+ */
+const MAX_PATTERN = 500;
+/** Likewise, a `.gitignore` with more rules than this is not one. */
+const MAX_RULES = 2000;
 function compile(line) {
     // Trailing whitespace is not part of the pattern unless it was escaped.
     let pattern = line.replace(/(?<!\\)\s+$/, '');
     if (pattern === '' || pattern.startsWith('#'))
         return null;
+    if (pattern.length > MAX_PATTERN)
+        return null;
     let negated = false;
     if (pattern.startsWith('!')) {
         negated = true;
@@ -77,8 +92,23 @@ function compile(line) {
     if (pattern.startsWith('/'))
         pattern = pattern.slice(1);
     const body = translate(pattern);
-    const source = anchored ? `^${body}(?:/.*)?$` : `(?:^|/)${body}(?:/.*)?$`;
-    return { re: new RegExp(source), negated, dirOnly };
+    // No subtree suffix: a pattern matches a path, not everything beneath it.
+    // Git prunes ignored directories during traversal instead, which is what the
+    // walk does — and only that order makes `logos/*` followed by
+    // `!logos/logos-in-app/` mean what git means. Matching the subtree here made
+    // the negation unreachable: every file under the re-included directory stayed
+    // ignored, 28 of them tracked, in one real repository.
+    const source = anchored ? `^${body}$` : `(?:^|/)${body}$`;
+    try {
+        return { re: new RegExp(source), negated, dirOnly };
+    }
+    catch {
+        // `[z-a]` is a reversed range, and one line of it used to end the scan:
+        // the throw escaped `walk()`, which runs before any scanner's error
+        // handling. A pattern git itself would treat as literal is not worth
+        // dying over — skip it and read the rest of the file.
+        return null;
+    }
 }
 export class Gitignore {
     layers;
@@ -92,6 +122,8 @@ export class Gitignore {
     extend(base, content) {
         const rules = [];
         for (const line of content.split(/\r?\n/)) {
+            if (rules.length >= MAX_RULES)
+                break;
             const rule = compile(line);
             if (rule)
                 rules.push(rule);
@@ -133,6 +165,12 @@ export function extendedAt(parent, dir, read) {
  * Repository-local excludes. Same syntax, same precedence as a root
  * `.gitignore`, but kept out of version control — so a machine-specific
  * scratch directory is invisible here without appearing in anyone's diff.
+ *
+ * The user's *global* ignore file (`core.excludesFile`, usually
+ * `~/.config/git/ignore`) is deliberately not read: it lives outside the scan
+ * root, which SECURITY.md promises we do not touch, and it is per-machine — a
+ * CI checkout would not have it, so honouring it would make a local scan
+ * quieter than the one that gates the merge.
  */
 export function repositoryExcludes(root, read) {
     const content = read(join(root, '.git', 'info', 'exclude'));

package/dist/utils/spans.d.ts

@@ -0,0 +1,32 @@
+/**
+ * Where the strings and comments are in a file.
+ *
+ * A regex ruleset has no idea whether it matched code, a sentence about code,
+ * or a URL — and the difference decides whether a finding is real. Our own
+ * source proved it: a comment reading "merely calling `jwt.verify(token,
+ * secret)`" was reported as a JWT vulnerability, and a rule *description*
+ * quoting `eval()` was reported as dynamic code execution.
+ *
+ * The naive fix — treat everything after `//` as a comment — is worse than the
+ * bug, because `"https://api.example.com"` would silence every finding on that
+ * line. So this walks the file once, tracking quotes as it goes, and both
+ * answers come out correct.
+ */
+export interface Span {
+    start: number;
+    end: number;
+    kind: 'string' | 'comment';
+}
+export interface CommentStyle {
+    /** `//` line comments and `/* *\/` blocks. */
+    slashes: boolean;
+    /** `#` line comments: shell, Python, Ruby, YAML, Terraform, Dockerfile. */
+    hash: boolean;
+    /** `--` line comments: SQL. */
+    dashes: boolean;
+}
+export declare function commentStyleFor(languages: readonly string[]): CommentStyle;
+export declare function lexSpans(source: string, style: CommentStyle): Span[];
+/** Binary search: is `index` inside a span of this kind? */
+export declare function spanAt(spans: readonly Span[], index: number): Span | null;
+export declare function isInside(spans: readonly Span[], index: number, kind: Span['kind']): boolean;

package/dist/utils/spans.js

@@ -0,0 +1,98 @@
+/**
+ * Where the strings and comments are in a file.
+ *
+ * A regex ruleset has no idea whether it matched code, a sentence about code,
+ * or a URL — and the difference decides whether a finding is real. Our own
+ * source proved it: a comment reading "merely calling `jwt.verify(token,
+ * secret)`" was reported as a JWT vulnerability, and a rule *description*
+ * quoting `eval()` was reported as dynamic code execution.
+ *
+ * The naive fix — treat everything after `//` as a comment — is worse than the
+ * bug, because `"https://api.example.com"` would silence every finding on that
+ * line. So this walks the file once, tracking quotes as it goes, and both
+ * answers come out correct.
+ */
+export function commentStyleFor(languages) {
+    const has = (l) => languages.includes(l);
+    return {
+        slashes: has('javascript') || has('typescript') || has('go') || has('php') || has('sql'),
+        hash: has('python') ||
+            has('shell') ||
+            has('ruby') ||
+            has('yaml') ||
+            has('terraform') ||
+            has('dockerfile'),
+        dashes: has('sql'),
+    };
+}
+/** Files past this size are skipped by the caller anyway; this is belt and braces. */
+const MAX_SOURCE = 1_000_000;
+export function lexSpans(source, style) {
+    const spans = [];
+    if (source.length > MAX_SOURCE)
+        return spans;
+    for (let i = 0; i < source.length; i++) {
+        const ch = source[i];
+        // Strings. A template literal is taken whole, interpolations included —
+        // nothing here needs to reason about what is inside one.
+        if (ch === '"' || ch === "'" || ch === '`') {
+            const start = i;
+            const quote = ch;
+            i++;
+            while (i < source.length) {
+                const c = source[i];
+                if (c === '\\') {
+                    i += 2;
+                    continue;
+                }
+                if (c === quote)
+                    break;
+                // An unterminated single- or double-quoted string ends at the newline,
+                // which is what an apostrophe in a comment looks like.
+                if (c === '\n' && quote !== '`')
+                    break;
+                i++;
+            }
+            spans.push({ start, end: Math.min(i, source.length - 1), kind: 'string' });
+            continue;
+        }
+        const two = source.slice(i, i + 2);
+        if (style.slashes && two === '/*') {
+            const end = source.indexOf('*/', i + 2);
+            const close = end === -1 ? source.length - 1 : end + 1;
+            spans.push({ start: i, end: close, kind: 'comment' });
+            i = close;
+            continue;
+        }
+        const lineComment = (style.slashes && two === '//') ||
+            (style.dashes && two === '--') ||
+            (style.hash && ch === '#');
+        if (lineComment) {
+            const newline = source.indexOf('\n', i);
+            const close = newline === -1 ? source.length - 1 : newline - 1;
+            spans.push({ start: i, end: close, kind: 'comment' });
+            i = close;
+            continue;
+        }
+    }
+    return spans;
+}
+/** Binary search: is `index` inside a span of this kind? */
+export function spanAt(spans, index) {
+    let low = 0;
+    let high = spans.length - 1;
+    while (low <= high) {
+        const mid = (low + high) >> 1;
+        const span = spans[mid];
+        if (index < span.start)
+            high = mid - 1;
+        else if (index > span.end)
+            low = mid + 1;
+        else
+            return span;
+    }
+    return null;
+}
+export function isInside(spans, index, kind) {
+    return spanAt(spans, index)?.kind === kind;
+}

package/examples/security.yml

@@ -22,7 +22,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v7
-      - uses: murtazaozdemir/cleartoship@v0.10.1
+      - uses: murtazaozdemir/cleartoship@v0.10.3
         with:
           fail-on: critical   # block the PR only on criticals
           comment: true        # post a summary comment on the PR

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cleartoship",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "description": "The 30-second pre-launch security clearance for AI-built & vibe-coded apps. Catches missing Server Action auth, Supabase RLS holes, hallucinated npm packages and leaked keys.",
   "keywords": [
     "security",