YukiZygisk

A new kernel-level Zygisk implementation designed to explore better, cleaner, and more flexible ways of injecting.

The standalone kernel sources live in kernel. A single test KMI can be built through DDK:

./build.sh kernel -k android15-6.6

The output is KMI-tagged as build/out/lkm/android15-6.6_yukizygisk.ko. Use --all-kmis to build all supported GKI targets locally. CI builds the seven supported targets as a matrix and assembles one release module package containing every KO.

This is a buildable extraction checkpoint, not a complete runtime replacement for the YukiSU-integrated module yet. See the source inventory for the remaining daemon mediation, payload staging, and host-backend runtime validation work. The kernel-side setresuid tracepoint monitor, SELinux policy adapter, and mount cleanup adapter are present in the standalone LKM, but still need device-side validation before they can be treated as runtime parity.

The standalone control path no longer creates /dev/yukizygisk. The LKM arms a one-shot prctl bootstrap when loaded with a per-boot cookie; zygiskd claims an anonymous control fd immediately after startup and then reuses the YZ_IOCTL_* command surface on that fd. Root control clients use the separate yzctl binary, which requests an independently authenticated anonymous control fd and talks to the same kernel ioctl surface without going through zygiskd.

If zygiskd never claims the bootstrap fd, the kernel guard checks for zygote service sockets after a short delay. Once service startup is visible, the guard clears the bootstrap cookie, disables new control sessions, and fails closed. The later boot-completed health check performs the external module unload.

The standalone design remains root-implementation agnostic at its internal boundaries, but its current admission policy is deliberately narrow. Module initialization accepts exactly one KernelSU/YukiSU backend (redirect or non-redirect) or one KernelPatch/APatch backend with a readable denylist. Magisk-only, multi-root, and no-root environments fail closed. Kernel code outside the host adapter calls YukiZygisk-owned yz_* and yz_host_* interfaces; detection and denylist routing reuse the Kasumi implementation.

The LSM interception point is now extracted into the standalone host layer: selinux_bprm_committed_creds is patched through a versioned adapter that uses the 6.12+ static_calls_table path or the older security_hook_heads path.

Standalone mount cleanup is also owned by YukiZygisk now. YZ_IOCTL_UMOUNT_PID schedules target-context task work, scans that app's /proc/self/mountinfo, and detaches KSU/Magisk/APatch/YukiZygisk tagged mounts plus /data/adb module mounts itself. It must not depend on KSU's kernel_umount feature being enabled.

The standalone control ABI is YZ_IOCTL_* with ioctl magic 'Y' only. It does not accept the integrated YukiSU/YukiZygisk KSU_IOCTL_YZ_*/'K' ABI.

The default package is a normal module containing zygiskd64, zygiskd32, the arm64 yzctl control client, paired libzygisk64.so/libzygisk32.so, libyukilinker64.so/libyukilinker32.so, and libyukizncore64.so/libyukizncore32.so payloads, plus a KMI-specific LKM directory. A local test package may contain one lkm/<kmi>_yukizygisk.ko:

./build.sh package -k android15-6.6

A release package contains all supported KMIs and is produced with ./build.sh package --all-kmis (or by CI's parallel matrix). During install and post-fs-data, the module derives the exact GKI KMI from uname -r and loads only the matching KO. Unknown releases and missing matches fail closed. The script then starts the daemon with the same bootstrap cookie. This gives up early-native injection by default; that capability can remain a future optional host backend rather than the baseline standalone path.

The packaged WebUI has three pages: device/injection status, configuration, and about/credits. It reads kernel-owned runtime state and requests reloads through yzctl; zygiskd is not a manager or user control interface. The WebUI does not own a separately configured denylist. The preferred path asks the accepted KernelSU or KernelPatch backend through a CFI-safe kernel callable. If that callable cannot be resolved, the kernel asks zygiskd for a refresh over netlink: zygiskd uses KernelSU's userspace ioctl/prctl policy API or parses APatch's package_config, then atomically hands a bounded snapshot back through a sealed memfd on the authenticated anonymous control fd. The WebUI only selects whether matching processes skip injection or keep injection before mount cleanup.