Repository navigation
SwitchWaker 0.1.4
Pre-release
Pre-release
SwitchWaker runs The Wind Waker (GameCube, GZLE01 revision 0) natively on a Switch with custom firmware: the decompiled game, on Aurora. No emulation.
No game inside, no binaries. Every build contains code generated from your own disc, so you build it on your computer with one script and copy the result to your SD card. Do not share what you build.
Install
- Download
SwitchWaker-v0.1.4.zipbelow and unzip it. - Follow INSTALL.md: with Docker, git and Python installed, one command builds everything from your disc (an uncompressed
.iso):The first build takes an hour or more; later ones minutes.scripts/switch/make_sd.sh --disc /path/to/GZLE01.iso
- Copy
build/sd/to the root of the SD card and start the Homebrew Menu in title mode (hold R while starting any game). The first start prepares the shaders once (8 to 10 minutes).
Updating from 0.1.x: build with --no-disc and copy switchwaker.nro and initial_pipeline_cache.db over the old ones. Saves, settings and prepared shaders stay. An old native/env.txt is moved into settings.ini on the first start and kept as env.txt.old.
What's new since 0.1.3
- Every option lives in the options menu (
settings.ini); the Switch no longer readsenv.txt. Developer variables without a menu row go in a[dev]section ofsettings.ini. - One log per session in
sdmc:/switch/switchwaker/native/logs/(switchwaker_<date>_<time>.log, the newest is the last run); the 10 most recent are kept, so logs cannot fill the card. Include the newest one in a report. - Options menu → Depuración: hitch-line threshold and HD texture statistics (both off by default on the Switch).
- Network debug server for developers, off by default (Depuración → Servidor de depuración, at the next start): deploy a build, follow the log, take screenshots and warp from the computer. The game opens no network port unless you turn it on. Same server as SwitchWakerHD's.
- Internal: finished A/B test switches removed (no change for players); the Dawn patch set now applies cleanly on a second configure.
Credits, licenses and how AI was used: see the README.