A store ships the tool that opens it
The Context
casync-llvm-mingw holds the llvm-mingw toolchain as a casync chunk store, as 20260828-prebuilt-toolchains-as-casync-stores describes. The chunks are reachable only through desync, so the first line of its readme told the reader to run go install github.com/folbricht/desync/cmd/desync@latest.
That line asks somebody to install a Go toolchain before they can download a clang toolchain. On a fresh CI runner it is another provisioning step, and on a workstation without Go it is an unrelated language runtime installed to open one archive.
The store cannot hold desync itself. Everything in it is addressed by chunk, and the only thing that can turn chunks back into files is the tool being fetched.
The Problem Statement
The one path a first-time reader takes was the path with an unstated prerequisite on it. Nothing else about the repository needs anything installed: the indexes are files, the store is files, the scripts are sh.
The same gap ran through CI, which installed Go and built desync from source on every run. What the gate exercised was therefore never the binary a reader would be running.
Design
The binaries live in the repository, under desync/<os>_<arch>/, taken from upstream’s release rather than built here. desync v1.1.0 covers linux, darwin and windows for both amd64 and arm64, which is 215 MB in a checkout and roughly 72 MB packed. Upstream’s BSD 3-Clause licence sits beside each one.
Each binary was checked against upstream’s published checksums.txt when it was vendored, and its own hash is recorded in desync/sha256sums.txt. check_store.sh verifies those hashes on every run, with a negative control that alters a byte and requires the check to fail.
build.sh, extract.sh and check_store.sh all source find_desync.sh, which picks the binary for the host and falls back to one on PATH. CI installs no toolchain and runs the same binary the readme hands out, so the gate now measures what a reader actually executes.
The readme gives one curl line per platform. A reader downloads a binary, marks it executable on Unix, and fetches a toolchain.
CRIS Score
| Factor | Score | Evidence |
|---|---|---|
| Complexity | 9 | A directory of binaries, a twenty-line picker, and one check in an existing gate. |
| Reach | 7 | Every reader of the store, and every CI job that fetches from it. |
| Impediment | 7 | A Go toolchain installed to open one archive, or a reader who stops there. |
| Stakeholder | 6 | Build and CI, plus anyone the store is handed to outside V-Sekai. |
| Total | 7.25 | Schedule soon. |
The Downsides
215 MB of third-party binaries sit in every checkout, carried by everyone who syncs the manifest, most of whom will use one of the six. Git grows monotonically, so each refresh adds another 72 MB or so that cannot be removed without rewriting history.
V-Sekai now redistributes binaries it did not build. The recorded hashes prove these are the bytes upstream published on 2026-08-23 and prove nothing about how upstream built them. Anyone who needs more than that should build desync from source and set DESYNC.
The vendored copy goes stale on its own. A new desync release changes nothing here until somebody refreshes it by hand, and no check reports the gap, because a checked-in binary has no upstream to compare against without a network call the gate deliberately does not make.
The Road Not Taken
go install is what this replaces, and it remains in the readme for anyone who prefers it.
Package managers reach many machines and not all of them, and the versions they carry lag. Guidance that works for some readers and silently fails for others is the shape of problem being removed here.
Linking upstream’s release archives from the readme costs no repository space. It puts a download and an unpack in front of the reader, leaves nothing to check the bytes against, and breaks whenever upstream retires a release, which is the failure this decision is meant to make impossible.
Building in CI only, and leaving readers to arrange their own tool, is the arrangement that produced the readme this decision rewrote.
Status
Status: Proposed
Implemented in casync-llvm-mingw at desync v1.1.0. No other repository ships its extractor this way.
Decision Makers
- K. S. Ernest (iFire) Lee