fix: Windows CRLF crash + npm shim / native binary detection

[kurushimee]

Summary

Two Windows-specific fixes that together resolve the TypeError: Expected CommonJS module to have a function wrapper crash after --apply on Windows with CC 2.1.118.

• CRLF line ending handling in system prompt escaping — On Windows, .md files have \r\n line endings. The newline escaping for " and ' JS string contexts only replaced \n, leaving bare \r (a JS line terminator) in the patched output. All 273 prompt files are affected (4,875 stray \r chars). Changed /\n/g to /\r\n|\r|\n/g for quote contexts; added \r normalization in the backtick branch too.
• npm shim resolution + native binary search paths — On Windows, which('claude') returns an npm shim (sh/cmd wrapper), not the actual binary. New resolveNpmShimTarget() follows the shim to the real executable. Also adds bin/claude.exe search paths for CC 2.1.116+ which ships as a native binary with no cli.js.

Verification

• node --check on the patched JS after stripping \r from the pre-fix output: exit 0 (confirms \r is the sole syntax cause)
• Unit test: old escaping turns CRLF into \r\n (invalid); new escaping turns it into \n (valid)
• pnpm run lint clean; pnpm test: 228 passed, 4 skipped (baseline)
• End-to-end: --restore → --apply → claude launches without error on Windows 11, CC 2.1.118 (npm install)

Test plan

• On Windows: node dist/index.mjs --apply then claude — should launch without TypeError
• On Windows: node --check "%USERPROFILE%\.tweakcc\native-claudejs-patched.js" — should exit 0
• On macOS/Linux: verify no regression (LF-only files should behave identically)

🤖 Generated with Claude Code

[BenIsLegit]

Read through this. Looks mostly good, but I have two notes

1. resolveNpmShimTarget regex grabs the first "$basedir/..." match

For a native-binary shim (CC 2.1.116+) that's the only such reference, so it works as you intend. But JS-style npm shims look more like:

if [ -x "$basedir/node" ]; then
  exec "$basedir/node"  "$basedir/node_modules/.../cli.js" "$@"

The first match is "$basedir/node". doesFileExist returns true on that, resolvePathToInstallationType doesn't recognize node as a Claude install, and the function returns null. So the new code path silently no-ops on JS shims rather than resolving them. Doesn't break anything since the fallback path handles those, but it does mean resolveNpmShimTarget only ever fires for native-binary shims today.

Drop-in fix — switch to matchAll and require node_modules in the candidate path:

async function resolveNpmShimTarget(shimPath: string): Promise<string | null> {
  try {
    const content = await fs.readFile(shimPath, 'utf8');

    // npm sh shim: exec "$basedir/node_modules/.../claude.exe"  "$@"
    // Multiple "$basedir/..." references can appear (e.g. the "$basedir/node"
    // probe at the top of npm-cli-shim sh scripts); enumerate all and pick
    // the first node_modules path that exists on disk.
    for (const m of content.matchAll(/"\$basedir\/([^"]+)"/g)) {
      if (!m[1].includes('node_modules')) continue;
      const target = path.resolve(path.dirname(shimPath), m[1]);
      if (await doesFileExist(target)) {
        debug(`resolveNpmShimTarget: resolved sh shim -> ${target}`);
        return target;
      }
    }

    // npm .cmd shim: "%dp0%\node_modules\...\claude.exe"  %*
    for (const m of content.matchAll(/"%dp0%\\([^"]+)"/g)) {
      if (!m[1].includes('node_modules')) continue;
      const target = path.resolve(
        path.dirname(shimPath),
        m[1].replace(/\\/g, '/')
      );
      if (await doesFileExist(target)) {
        debug(`resolveNpmShimTarget: resolved cmd shim -> ${target}`);
        return target;
      }
    }

    return null;
  } catch {
    return null;
  }
}

2. No committed regression test for the CRLF fix

The PR body mentions a unit test confirming old/new behavior, but src/patches/systemPrompts.test.ts isn't touched in the diff. Bare \r is invisible in source, so a regression here only surfaces for Windows users running --apply — reverting /\r\n|\r|\n/g back to /\n/g six months from now wouldn't trip CI.

Drop-in tests that fit the existing fixture style — add these next to the 'should convert newlines to \\n for double-quoted string literals' test in src/patches/systemPrompts.test.ts:

it('should convert CRLF line endings to \\n for double-quoted string literals (Windows)', async () => {
  const mockPromptData = buildMockPromptData({
    prompt: { content: 'Hello\r\nWorld' },
    regex: 'Hello(?:\n|\\\\n)World',
    getInterpolatedContent: () => 'Hello\r\nWorld',
    pieces: ['Hello\r\nWorld'],
  });

  setupMocks(mockPromptData);

  const cliContent = 'description:"Hello\\nWorld"';

  const result = await applySystemPrompts(cliContent, '1.0.0', false);

  expect(result.newContent).toBe('description:"Hello\\nWorld"');
  expect(result.newContent).not.toMatch(/\r/);
});

it('should convert CRLF line endings to \\n for single-quoted string literals (Windows)', async () => {
  const mockPromptData = buildMockPromptData({
    prompt: { content: 'Hello\r\nWorld' },
    regex: 'Hello(?:\n|\\\\n)World',
    getInterpolatedContent: () => 'Hello\r\nWorld',
    pieces: ['Hello\r\nWorld'],
  });

  setupMocks(mockPromptData);

  const cliContent = "msg:'Hello\\nWorld'";

  const result = await applySystemPrompts(cliContent, '1.0.0', false);

  expect(result.newContent).toBe("msg:'Hello\\nWorld'");
  expect(result.newContent).not.toMatch(/\r/);
});

it('should normalize CRLF to LF for backtick template literals (Windows)', async () => {
  const mockPromptData = buildMockPromptData({
    prompt: { content: 'Hello\r\nWorld' },
    regex: 'Hello(?:\n|\\\\n)World',
    getInterpolatedContent: () => 'Hello\r\nWorld',
    pieces: ['Hello\r\nWorld'],
  });

  setupMocks(mockPromptData);

  const cliContent = 'description:`Hello\nWorld`';

  const result = await applySystemPrompts(cliContent, '1.0.0', false);

  expect(result.newContent).toBe('description:`Hello\nWorld`');
  expect(result.newContent).not.toMatch(/\r/);
});

Three tests, one per delimiter branch. The not.toMatch(/\r/) assertion is the key one — it's what would have caught the original bug