The shooter checkout could not install its locked Three.js download
Claude Opus 5.5 Demo stopped at dependency installation after 283 seconds in our sandbox. The failing download came from bnpm.byted.org, a registry host recorded in the shooter project's lockfile. That is the first thing a developer should know before treating this collection as a reusable game starter. The repository offers source for browser games, but the documented clone-and-build route did not work for the project our lab selected.
The README is primarily Chinese and provides no separate English guide. It presents a pelican cycling game, a CrossFire-inspired transport-ship shooter, and a QQ Speed-inspired racing game as outputs of Claude Code sessions. The author's claim is that each began with a short prompt and received no manual code edits. Read that as a description of the author's process. The public source lets you study the output without proving the generation history.
What happened when we ran it
Our run of commit 54adc1f failed installation with exit code 1 after 283 seconds inside cf-transport-ship/. The October 9, 2026 sandbox used the lab-node:22 image, 3 CPUs, 8 GB of RAM, no secrets, and no elevated privileges. The checkout contained 61 files, roughly 18,461 source lines, and occupied 0.9 MB. The scan found 0 CI workflow files, no Dockerfile, and no tests directory.
The log tail reports npm's ENOTFOUND error from getaddrinfo while fetching https://bnpm.byted.org/three/-/three-0.186.0.tgz. That means the hostname could not be resolved in this run. The log also contains npm's warning that --force disables recommended protections. It does not establish a proxy misconfiguration, a broken Three.js package, or a game-code defect. The supplied results contain no completed build or test measurement, and they cover the shooter directory rather than every game.
The lockfile carries the same hostname named in the failure
The 283-second failure has a concrete source counterpart: the shooter lockfile points its Three.js package at the exact URL in the npm error. Setting a different default registry may still leave a recorded tarball URL in play. For a fork, inspect and repair the locked resolution using a registry you can access, then check package identity and integrity before rerunning the documented commands.
That is a proposed repair, not a result we measured. The useful boundary is that installation failed before the lab could establish a working build. You should budget a dependency-provenance check before any rendering or gameplay work. A live hosted copy, if it works for you, can answer whether the game is enjoyable. It cannot establish that the current source and lockfile reproduce that hosted copy on a fresh machine.
The build script packages a static page without a model service
The 0.9 MB checkout contains ordinary JavaScript projects, rather than a service that calls Claude during play. In cf-transport-ship/, build.mjs bundles the entry point with esbuild and inserts JavaScript and CSS into dist/index.html. Local gameplay needs no model API key. Cloudflare credentials belong to publishing your own deployment, and downloading dependencies still requires network access before a local artifact exists.
The README describes procedural geometry, textures, and synthesized audio, which makes the source useful for studying how a browser scene can avoid a separate asset pipeline. The shooter separates its map, bots, player, and weapons into modules. You can follow a gameplay change through those files instead of starting in a generated bundle. That organization earns it a place as reading material even while installation remains unresolved.
Open shooter reports concern aiming and unfinished matches
Our checkout's 18,461 source lines do not imply that the games have reliable controls or complete matches. Issue 11, updated October 11, reports that the shooter view sometimes fails to follow mouse movement. Issue 4 says everyone can run out of ammunition and the match cannot end. Both refer to the transport-ship game, so they are relevant to the directory our lab attempted.
These are user reports, not failures observed during our run. The source contains pointer-lock handling and match-ending logic, but the reports deserve specific reproduction steps before a fork goes public. After repairing dependencies, check entering and leaving pointer lock, resuming play, and completing a match under depleted ammunition. Watching the scene render would leave those problems unanswered. The examples should earn confidence through actual play, not the number of modules they contain.
A September push sits beside October user activity
The repository was last pushed on September 23, 2026. GitHub showed 1,059 stars and 10 open issues and PRs on October 12, including an open racing-related pull request. The latest-release endpoint returned no release. Fresh user reports show continuing interest, while the older push date provides no evidence that those reports have been fixed. Neither a missing release tag nor a growing discussion count settles maintenance quality by itself.
Our scan's 0 CI workflows means there is no checked-in automated release gate visible through that signal. Choose Three.js examples when you want a smaller rendering lesson, or Babylon.js or PlayCanvas when you need an engine to build around. This collection is worth inspecting for procedural game ideas. For reuse, start by getting the shooter dependencies from an accessible source, then reproduce the reported control and match problems before trusting the generated page.