git calls in the tree/file-browser crash with ENOBUFS on large repos

[ugolafosse]

What happens

Open diffity tree (or any tour view) on a large repo and the whole UI goes down with:

Something went wrong
Failed to get tree: Error: spawnSync git ENOBUFS

/api/tree returns a 500. Since the SPA surfaces that at the top level, views that don't even need the file tree (a code tour) go down with it too.

What I found

The git calls behind the file browser run through execFileSync("git", [...], { encoding: "utf-8" }) with no maxBuffer, so they get Node's 1 MB default. Once git ls-files prints more than 1 MB of paths, the call throws ENOBUFS. In 0.9.5's dist/index.js it's the git ls-files calls in getTreeEntries and getTreeFingerprint, the git status --porcelain -u, and the git rev-parse --show-toplevel calls.

How to reproduce

Any repo whose tracked paths exceed ~1 MB of text. Mine is a monorepo that vendors a few external repos as git subtrees, so git ls-files is ~14,300 files / 1.24 MB, just over the line. A large public repo would do the same.

git ls-files | wc -c   # > 1048576 reproduces it

Environment

• diffity 0.9.5
• node (system), macOS arm64

[gustavo-depaula]

Opened #36 for the ENOBUFS half of this.

Confirmed the same diagnosis independently on a ~29k-file monorepo (git ls-files at 2.3 MB), and your list of call sites is exactly right. One thing to add: exec.ts already has a 50 MB ceiling, but only on its string-command helpers (execLarge, execWithStdin) — tree.ts simply never used them. So it is a bypassed helper rather than a missing option.

The PR keeps the argv form rather than switching those calls to execLarge, because dirPath comes from user input and repository paths contain spaces and quotes in practice. It adds execFileLarge as the argv counterpart, names the shared MAX_BUFFER, and routes all seven execFileSync calls in tree.ts through it. There is a regression test that builds a temp repo over the 1 MB line: it fails with spawnSync git ENOBUFS without the fix and passes with it.

Your second observation is not addressed there — that a /api/tree 500 takes down views which do not need the tree, a tour included. That is an error-boundary problem rather than a git one, so I left it out and used no closing keyword on the PR. Worth keeping this issue open for it.