Claude/eager cray 2n42az - #4
Open
matthewcodergamer wants to merge 6 commits into
Open
matthewcodergamer wants to merge 6 commits into
matthewcodergamer wants to merge 6 commits into
Conversation
The web build could not start on an iPhone, and even where it started there was no way to move: iOS has no keyboard and no Pointer Lock. Startup blockers: - INITIAL_MEMORY=2047mb is a single upfront WebAssembly.Memory reservation that mobile Safari refuses outright. Start at 512mb and grow to the same 2047mb ceiling instead. Memory limits and the thread pool size are now overridable via EM_* environment variables. - PTHREAD_POOL_SIZE=8 with POOL_SIZE_STRICT=2 hard-aborts on a device with fewer cores. Size the pool from navigator.hardwareConcurrency and downgrade strictness to a warning so extra threads spawn on demand. - The page must be cross-origin isolated for SharedArrayBuffer, and nothing in the tree provided those headers. Add serve.py (a dev server that sends COOP/COEP and the right MIME types) and a _headers file for static hosts, plus per-server config in README-hosting.md. Input and presentation: - shell.html had no viewport meta at all, so iPhone Safari laid the page out at 980px and every touch coordinate reaching the engine was wrong. Add the viewport and web-app meta tags, suppress pinch/double-tap zoom, rubber-band scrolling and the long-press callout, handle safe-area insets and orientation changes, and size the canvas to a pixel budget rather than an iPhone's native 3x grid. - Gate startup behind a "Tap to play" overlay. iOS only allows starting an AudioContext, entering fullscreen or locking orientation inside a real user gesture, so main() is held on a run dependency until then. - Detect touch devices and pass +touch_enable 1, so the engine's existing on-screen controls come up. They default to off everywhere but Android, which is why there was no way to move. Also pass a device-appropriate -w/-h and +mat_picmip 2. - Report failures on screen. A phone has no console, so unsupported browsers, missing cross-origin isolation, out-of-memory aborts and lost GL contexts now explain themselves instead of leaving a black page. - iPhone Safari has no Element.requestFullscreen; hide the button there and point at Add to Home Screen, which does give a fullscreen window. Bugs found along the way: - A failed map chunk download left the engine thread parked forever in memory.atomic.wait32, because only the success path stored and notified the lock. Release it on failure too. A 404 was worse than a hang: it fired onload, and the error page was parsed as chunk data. Check the status and bounds-check every entry. - touch.cpp built a zero-quad mesh from texture id 0 whenever no touch texture made it into the atlas, which is exactly what happens when the packed game data has no vgui/touch materials. Skip the atlas pass. - touch.cpp cleared an stbrp_rect array using sizeof(stbrp_node), leaving each entry half initialised, and kept a stale touchTextureID after deleting the texture. - An Error() call was missing the argument for its %s. - Only request pointer lock where the browser actually supports it. - repackage.js missed loose materials/vgui/touch/*.vtf, so touch buttons drew with the missing-material texture even when the files were there. Note: not build-tested. No emsdk in this environment, so the C++ changes are unverified by a compiler; the JS/HTML was exercised against a DOM harness covering iPhone, Android, desktop and each failure mode. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qy5DKdwTqGQqScWvJyNcft
The CI artifact was the engine and nothing else, so "download release.zip and open it" could never work: there is no game data in it, and opening the page over file:// or any host that does not send COOP/COEP leaves it without SharedArrayBuffer. - Add fetch_chunks.py, which downloads the packed map chunks into <bundle>/chunks. It reads the map list out of pre.js rather than keeping a second copy that can drift. - Ship serve.py, _headers, fetch_chunks.py and pre.js inside the bundle so the artifact is self-contained. - Generate an index.html that points at hl2_launcher.html, so opening the bundle root works. - Document the actual path from a push to a running game: download the artifact, fetch the chunks, serve it cross-origin isolated. Call out that GitHub Pages cannot host this, since it cannot set headers. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qy5DKdwTqGQqScWvJyNcft
… worker GitHub Pages cannot send Cross-Origin-Opener-Policy or Cross-Origin-Embedder-Policy, so it could not host a threaded build at all: no SharedArrayBuffer, no startup. A service worker can add those headers to responses as they come in, which makes the page cross-origin isolated on a host that has no header control. - Add coi-serviceworker.js and load it from shell.html's <head>. The first visit registers it and reloads once; after that the page is isolated. It also sets Cross-Origin-Resource-Policy on what it passes through, so game data served from another origin is not rejected. - Refuse to reload unless the attempt can be recorded in sessionStorage. Reading it throws in Safari private browsing, and a reload that cannot be remembered repeats forever. Both reload paths now go through one guarded helper, and the page reports why it gave up. - Teach the capability gate the difference between "not isolated" and "isolation is one reload away", so a pending service worker shows progress instead of a fatal error, and a failed one says why. - Add a Deploy to GitHub Pages workflow. It builds, optionally pulls the chunks when a CHUNKS_BASE_URL variable is set, and fails with a clear message if the site exceeds the 1GB Pages limit instead of letting the deploy fail on its own. - Make the chunk location configurable with ?chunks=<url> or window.CHUNK_BASE_URL, since GitHub blocks files over 100MB and caps a Pages site at 1GB, which the full Portal data does not fit under. The engine can now be on Pages with the data hosted anywhere. - Drop a .nojekyll into the bundle so Pages does not strip _headers. Verified with a harness covering both halves of the worker: header injection preserves body and content-type, only-if-cached passes through, and every reload path stops after one attempt. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qy5DKdwTqGQqScWvJyNcft
`zip -r out.zip *` skips dotfiles, so .nojekyll never made it into release.zip. Anyone unzipping the artifact to deploy GitHub Pages by hand would get a site where Jekyll strips files beginning with an underscore. Zipping "." instead picks them up. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qy5DKdwTqGQqScWvJyNcft
iPhone Safari kills the page while compiling the largest side module, so the size of each .so is the number that matters. Print them in CI so changes to the link flags can be compared without a device. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qy5DKdwTqGQqScWvJyNcft
SIDE_MODULE=1 exports every symbol and therefore keeps every function in each of the 24 side modules. The tree already compiles with -fvisibility=hidden and marks its genuine exports DLL_EXPORT, which is exactly the contract SIDE_MODULE=2 honours, so wasm-ld can now garbage-collect unreferenced code per module. iPhone Safari kills the page while compiling the largest module, so the amount of code it has to compile is the constraint. CI prints the per-module sizes to measure the effect. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Qy5DKdwTqGQqScWvJyNcft
matthewcodergamer
force-pushed
the
claude/eager-cray-2n42az
branch
from
September 22, 2026 18:55
b027018 to
c5769e9
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.