Repository navigation
Verify the submission offline and add a captioned walkthrough - #15
Conversation
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 84a002d978
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| require(kind in (0, stat.S_IFREG) and not entry.flag_bits & 1 and | ||
| entry.compress_type in (zipfile.ZIP_STORED, zipfile.ZIP_DEFLATED) and | ||
| 0 <= entry.file_size <= MAX_MEMBER, 'ZIP member type, compression or size') | ||
| files = {e.filename: archive.read(e) for e in entries} |
There was a problem hiding this comment.
Bound decompression instead of trusting ZIP size metadata
When checking an untrusted ZIP, the size checks only constrain attacker-controlled file_size headers, while archive.read(e) can decompress the entire payload before truncating its output to that declared size. I reproduced this with a 921 KB archive whose LICENSE stream contains the original bytes followed by 40 MB of zeros, but whose headers declare the original length and CRC: verification reports passed while peak Python allocation reaches approximately 89 MB. Larger streams can exhaust memory despite MAX_MEMBER and MAX_TOTAL. Use bounded streaming decompression and enforce limits on actual expanded output rather than relying solely on ZIP metadata.
Useful? React with 👍 / 👎.
The submission now has a one-command offline verifier that checks the explicit ZIP against the checkout and independently recomputes the frozen 160-trial pilot. It rejects changed publication data, unsafe original ZIP names, unsupported compression, corrupt payloads and linked or oversized files without extraction or repair. Every Python CI lane runs it.
Adds an actual scripted viewer screenshot, a PNG cover, and a separately published 180-second captioned walkthrough (2.5 MB). The video is labeled as an edited sequence of browser states and result cards. The report/PDF bytes, datasets and numerical findings are unchanged. The 17-member evidence ZIP remains below 1 MB; event submission still awaits private contact details.
Validation: 320 Python tests and five viewer tests passed locally. Eleven verifier tests include test-first fixes for three findings from an independent whole-branch review: rehashed data/interpretation changes, NUL-suffixed ZIP names, and decompressor errors escaping the JSON failure boundary. All four independently constructed malicious archives now fail safely. Repository/standalone verification, planner/demo, 340 local links, PNG/scene inspection and full video decoding passed. Final-source CI run 42598265386 passed all 11 runtime lanes: seven Python lanes each ran 320 tests plus package verification, and four Node lanes each ran five viewer tests. All 11 downloaded artifacts were digest-checked; runtime/archive/lock/matrix/source identities match merge source
9aa9618b09e2d55e4aa42d319f6bbdacba5cd6e5and treefc6744147b2e4e131afbb1f2244c8ad59c86d71f. The reviewed branch head is84a002d978051336fbc0abfa7b3d3bb060d55128.No new model calls, cloud mutations or event submission. Retained evidence: task
submission-finalization-20261109; prior worktree and failed receipts preserved. Base:2ce59d34f92516cfb6666b891a4b35da2b46c9ba.