A malformed WebP chunk size hangs the agent forever: RIFF size decoded with a signed shift

[Jiaaqiliu]

Environment. prime-agent 0.7.3, main at f8f0036, Node 22.17.0, macOS 15 (darwin 25.5.0).

Summary

findWebpTiffOffset decodes the RIFF chunk size with a signed shift, so a size with the high bit set becomes negative and the scan loop stops advancing. The loop is synchronous, so it pins a core and blocks the event loop: the TUI stops repainting and Ctrl-C does nothing. Because the function never returns, the try/catch in resizeImage never runs.

Any image the agent handles goes through this, including one it merely finds in the repository it is pointed at.

Reproduction

utils/exif-orientation.ts:65:

const chunkSize =
    bytes[offset + 4] | (bytes[offset + 5] << 8) | (bytes[offset + 6] << 16) | (bytes[offset + 7] << 24);
const dataStart = offset + 8;
...
offset = dataStart + chunkSize + (chunkSize % 2);

Running the function body against a minimal RIFF/WEBP container with one chunk:

chunk size bytes	value	result
08 00 00 00	8	terminates
F8 FF FF FF	-8	offset + 8 - 8 + -0 — offset never changes
F9 FF FF FF	-7	offset + 8 - 7 + -1 — offset never changes
00 00 00 80	-2147483648	walks far negative, then crawls back 8 bytes at a time

A valid WebP with eight trailing bytes — a 4-byte chunk id plus F8 FF FF FF — and a widened RIFF size is enough to trigger it. Decoders stop at the VP8 chunk, so the file still opens normally in every other viewer.

Reachability

getExifOrientation is called from utils/image-resize.ts:74 and utils/image-convert.ts:26, which are reached from cli/file-processor.ts:48 (a file passed on the command line), interactive-mode.ts:4301 (a clipboard image paste), and the image-attachment path generally.

Related, same file

readOrientationFromTiff's little-endian read32 at :17 is missing the >>> 0 that its big-endian sibling on the next line has, so ifdOffset can go negative past the upper-bound-only guard on :23. That one degrades to entryCount = 0 rather than hanging, but it is the same class of mistake.

Suggested fix

Read the size unsigned (>>> 0) and refuse to continue when the computed next offset does not move forward. The regression test has to drive the function from a child process, because no in-process timeout can interrupt a synchronous loop.

---

I have a fix with a regression test on a branch: fix/webp-exif-scan-hang. npm run check passes and the surrounding suites still pass. I opened it as a PR first and the contribution gate closed it, which is what CONTRIBUTING.md says should happen, so I am bringing it here instead. Happy to leave it as is, adjust it, or drop it entirely if you would rather fix this differently.