# Releases and updates Shift's update channel is this repository. There is no update server, no object storage, and no release service to configure: the installers and a small manifest are pushed to a git branch, and a running copy reads that branch over HTTPS. ## The channel An orphan branch named `releases` holds exactly one release: ```text releases/ ├── shift-update.json ├── shift-update.sig ├── Shift-macOS-universal.dmg └── Shift-Setup-0.1.1-win-x64.exe ``` `shift-update.json` names the version and, per platform, the artifact file, its size in bytes and its SHA256: ```json { "schema": 1, "version": "0.1.1", "date": "2026-10-26", "notes": "", "artifacts": { "macos": { "file": "Shift-macOS-universal.dmg", "sha256": "57fc…", "size": 55479751 }, "windows": { "file": "Shift-Setup-0.1.1-win-x64.exe", "sha256": "0255…", "size": 24545040 } } } ``` ## Supported release targets **Shift ships two installers: macOS (universal) and Windows x64.** That is a scope decision, not a code limit, and it decides what the channel may contain: | Key | Shipped | Notes | | --- | --- | --- | | `macos` | yes | One universal bundle, Apple silicon and Intel. The DMG, because that is what the in-app updater mounts | | — | yes | `Shift-macOS-universal.pkg`, also macOS universal: the file to hand a machine installing Shift for the first time. It carries no manifest key — see "macOS packaging" | | `windows` | yes | x64, which is also what an ARM machine falls back to | | `windows-arm64` | no | Buildable — `build-windows-release.sh --arch arm64` still works — just not published | | `linux` | no | Build target, not a product; see below | Publishing an ARM64 Windows build later needs no code change: the ladder already prefers `windows-arm64` and falls back to `windows`, so adding the file to `artifacts/` and re-publishing is enough. Keys are platform, then platform and architecture where they differ. A build picks the first key its own architecture matches and never crosses operating systems: | Running on | Keys tried, in order | | --- | --- | | macOS (Apple silicon or Intel) | `macos` | | Windows x64 | `windows` | | Windows ARM64 | `windows-arm64`, then `windows` | | Linux | `linux` | The ARM64 fallback to `windows` is deliberate: Windows on ARM runs x64 binaries under emulation, so a channel that only carries x64 keeps serving ARM machines exactly as it did before the ARM build existed. `artifact_keys_for` and `select_artifact` in `src-tauri/src/update.rs` hold that ladder and are tested directly, so the rule is fixed by a test rather than by a comment. The branch is rebuilt as a fresh root commit on every publish and force-pushed, so it never accumulates gigabytes of installer history while `main` stays clean. ## macOS packaging Shift builds are **ad-hoc signed** (`signingIdentity: "-"` in `tauri.conf.json`) and are **not notarized** — notarization needs an Apple Developer ID account, which this project does not have. Gatekeeper therefore cannot vouch for any Shift build, and what a fresh machine says depends on how the binary is signed: | Signing | What a fresh machine says | | --- | --- | | Ad-hoc (`Signature=adhoc`) — **current** | "cannot be opened because it is from an unidentified developer", and the box offers a way through: Cancel, then right-click → 打开, or 允许 in 系统设置 → 隐私与安全性 | | Self-signed certificate | "cannot be opened because it is damaged" / reported as malware, with **only 移至废纸篓** — no override, no pane entry to approve | | Developer ID + notarized | Opens on double-click, no prompt | The middle row is the trap that cost a release. Signing with a locally created certificate ("Shift Local Signing") *looks* better — `codesign` on the build machine reports `valid on disk` and `satisfies its Designated Requirement`, and `spctl` even accepts it — but the certificate lives only in that one Mac's keychain, is self-issued, and is `CA:FALSE`, so on any other machine it cannot chain to a trusted root. Sequoia turns an unchainable signature on a quarantined download into a hard refusal, which is strictly worse than no certificate at all. If the identity is used again it must be installed **and trusted** on every machine, which is not something to ask of a user. So: ad-hoc, and the two containers each cover one job. | Container | First launch on a fresh machine | | --- | --- | | `.pkg`, via Installer.app or `installer` | `installer` does not propagate quarantine onto what it installs, so the app opens with **no prompt at all**. Verified: an installed copy carries `com.apple.provenance` and no `com.apple.quarantine`. The package itself is unsigned (`productsign` needs an *Installer Signing* certificate; a code-signing identity is rejected), so a browser-downloaded package may need right-click → 打开 once, before anything installs. `installer -pkg … -target /` needs nothing. | | `.dmg`, app dragged to /Applications | The app keeps the browser's quarantine flag, so the first launch needs the one approval above. Shift's own downloader sets no quarantine flag, which is why **in-app updates never prompt** — only a browser download on a new machine does. | `scripts/macos-pkg.sh ` builds the package: it stages the bundle under an `/Applications` root and runs `pkgbuild`, which neither re-signs nor strips the payload — the CDHash is identical in the built bundle, the DMG copy and the pkg copy. The `macos` manifest key stays a DMG on purpose: `update_install`'s handoff mounts the image with `hdiutil` and copies the bundle out of it, so pointing the key at a `.pkg` would make every in-app update fail to mount. Closing this for good means an Apple Developer ID: `Developer ID Application` + `notarytool` + `stapler staple`. Then both containers open on double-click with no pane to visit, and Screen Recording approval stops resetting between builds because the identity stops changing. ## Publishing ```bash pnpm tauri build --target universal-apple-darwin --bundles app,dmg scripts/macos-pkg.sh src-tauri/target/universal-apple-darwin/release/bundle/macos/Shift.app \ artifacts/Shift-macOS-universal.pkg # the first-install container printf 'Fixes the ring on external displays.\n' > /tmp/notes.md scripts/publish-release.sh --notes /tmp/notes.md # hash, commit, push ``` `scripts/publish-release.sh` hashes every file in `artifacts/`, builds the manifest, writes the tree with git plumbing (`update-index` → `write-tree` → `commit-tree` → `update-ref`, through a temporary `GIT_INDEX_FILE` so the working tree is untouched) and force-pushes `releases`. The manifest and the pushed bytes cannot disagree, because the same Python pass that walks the files produces both. | Flag | Effect | | --- | --- | | `--notes FILE` | Release notes shown in the Updates panel | | `--key FILE` | Sign the manifest with this ed25519 key (`SHIFT_RELEASE_KEY` also works) | | `--dry-run` | Print the manifest and the tree that would be pushed; write nothing | | `--no-push` | Commit locally on `releases` but leave the remote alone | | `--branch NAME` | Publish to another channel instead of `releases` | Recognised installers: `Shift-macOS-universal.dmg` (macOS), `Shift-Setup--win-x64.exe` and `Shift-Setup--win-arm64.exe` (Windows, keys `windows` and `windows-arm64`), `*.AppImage` / `*.deb` (Linux). Other files in `artifacts/` — the macOS `.pkg`, the portable ZIPs, `SHA256SUMS.txt`, the `README-Windows-.txt` notes — travel with the branch and are downloadable from it, but are not offered as installers by the manifest. The `.pkg` carries no key on purpose: the `macos` key must stay a DMG so the in-app handoff can mount it ("macOS packaging"). Dotfiles and an old manifest are skipped. Both Windows architectures come from one command, which cross-compiles each target, checks the PE machine field of the binary it just built (`0x8664` for x64, `0xAA64` for ARM64) so a wrong `--target` cannot ship an x64 EXE inside an arm64 package, and gates the installer to the matching machine: ```bash scripts/build-windows-release.sh # both architectures scripts/build-windows-release.sh --arch arm64 # one of them scripts/build-windows-release.sh --skip-build # repackage existing EXEs ``` ## Signing the channel Integrity says the download matches the manifest; authenticity says the manifest came from you. `scripts/release-key.py` closes that gap with one ed25519 key. ```bash scripts/release-key.py keygen ~/.config/shift-release/release.ed25519 # -> prints the hex public key to bake into the build export SHIFT_RELEASE_KEY=~/.config/shift-release/release.ed25519 pnpm tauri build --target universal-apple-darwin --bundles app,dmg # macOS scripts/build-windows-release.sh # both Windows architectures scripts/publish-release.sh --key "$SHIFT_RELEASE_KEY" ``` Give the release build `SHIFT_UPDATE_PUBKEY` (the hex key) and it can verify the feed; give it `SHIFT_RELEASE_KEY` (the private key) and `build-windows-release.sh` derives the public key from it, so a cross-build can never be packaged with a key that does not match the signature `publish-release.sh --key` produces. Without either, the script prints `no release key: this build will report the feed as unsigned` rather than letting it pass for a quieter log. The signature is **detached**: `shift-update.sig` holds base64 over the manifest's exact bytes, so verification needs no JSON canonicalisation rules and the manifest's formatting stays free to change. It is fetched from the directory that already holds the manifest, so a staging feed needs only its own manifest plus signature — a fork never configures a second URL. | Build | Feed | Result | | --- | --- | --- | | Key baked | signature verifies | update offered, panel says the signature was verified | | Key baked | signature missing or wrong | check fails with that reason; nothing is offered | | No key | any | integrity gate only; panel says this build cannot verify a signature | The key is the one secret in this design. Losing it means publishing a new app build with a new baked key; leaking it means an attacker can publish anything to the channel, exactly as leaking push access to `releases` already does today. Keep it outside the repository — `~/.config/shift-release/` is where the local publishing machine keeps it. ## Where a build looks for updates `src-tauri/build.rs` derives the feed from the `origin` remote at build time: | Remote | Feed URL | | --- | --- | | Gitea / Forgejo | `https://host/owner/repo/raw/branch/releases/shift-update.json` | | GitHub | `https://github.com/owner/repo/raw/releases/shift-update.json` | | GitLab | `https://gitlab.example/owner/repo/-/raw/releases/shift-update.json` | Two overrides exist, and neither needs a rebuild of the frontend: - `SHIFT_UPDATE_PUBKEY` at build time bakes in the release key; anything that is not 32 bytes of hex fails the build rather than silently producing an updater that trusts the feed anyway. A key that parses but is not a curve point is a runtime error, never a quiet downgrade to "no key". - `SHIFT_UPDATE_URL` at build time bakes a different feed into the binary. - `SHIFT_UPDATE_URL` in the environment of a running copy wins over the baked value, which is how a fork or a staging branch is tested. A source tree with no remote builds an updater that reports "no feed" rather than guessing at one. ## What the app does with it 1. Once a day, at startup (and on demand from Settings → **Updates**), the manifest is fetched with the system HTTP client — `/usr/bin/curl` on macOS and Linux, `Invoke-WebRequest` on Windows. The day of the last probe is stored in `settings.json`, so restarts cannot spend the day's check twice. 2. Versions are compared with semver rules, prereleases included: `0.2.0-alpha` sorts below `0.2.0`, so a test build does not offer to downgrade itself. 3. Download writes into a temp folder, then two gates run before anything can be installed: the byte length must equal the manifest's `size`, and the recomputed SHA256 must equal the manifest's `sha256`. A mismatch deletes the file and reports the error, so a truncated or swapped installer never reaches the installer step. 4. **Install & relaunch** applies the verified artifact and restarts Shift on it, as described in the next section; **Discard** removes the download. When an update is available the tray menu gains a row that opens Settings on the Updates tab. Manifest `schema` is checked, and a newer schema than this build understands is refused instead of being parsed optimistically. Unknown platform keys are ignored, so the manifest can name a platform an older app has never heard of. The Updates panel always states which of the two properties the current feed satisfies, so an unsigned channel never looks signed. ## Applying an update Neither operating system lets a running program replace its own executable, so the work belongs to a helper that outlives the app. `update_install` writes a script into the download folder, starts it detached, and exits about half a second later; the script waits for that pid to disappear and only then touches the installed copy. | Platform | What the helper does | | --- | --- | | macOS | Mounts the downloaded `.dmg`, `ditto`s the new bundle next to the old one, swaps them with two `mv` calls, removes the quarantine flag, detaches the image, and `open`s the result | | Windows | Waits for `shift.exe` to leave the process table, runs the installer with `/S`, then starts the `shift.exe` the installer wrote | | Linux, and a macOS binary run outside a bundle | Nothing is replaced; the artifact is opened instead, and the panel says so | Both scripts are written so a failure cannot end with Shift simply closed: every macOS failure path still `open`s the bundle it had, and the swap is `move aside, move in`, so a copy that fails halfway leaves the previous version where it was. Before the handoff starts, macOS writes a probe file next to the bundle — a root-owned copy in `/Applications` is reported as "open the downloaded installer yourself" rather than quietly quitting on an update that could not be applied. The Windows installer is the same one a user runs by hand, so the update path cannot drift from the manual one. Because the silent install comes from Shift itself, `${Silent}` skips the interactive "Shift is running, close it?" box: the app already told the user it is about to restart, and a box nobody can answer would abort the update. This is not a delta update and not a signature check on the running binary: the same SHA256 gate that guards a manual install is the only integrity gate, and the manifest signature rules in "Signing the channel" apply unchanged. Replacing the bundle gives it a new code-signing hash, and macOS keys permissions such as Screen Recording to that hash. A release signed with the same identity as the copy it replaces normally keeps its approvals; a locally signed build, where every build has its own identity, can ask again after an update. That is the same consequence a manual reinstall has, and it is why the notarized-build item in the roadmap is worth closing. ## Trust boundary Two properties, two mechanisms: - **Integrity of the installer** — byte length plus recomputed SHA256 against the manifest. Always enforced, in every build. - **Authenticity of the manifest** — a detached ed25519 signature verified against a baked public key. Enforced in every build that was compiled with `SHIFT_UPDATE_PUBKEY`. A build without a key falls back to the weaker trust it had before signing existed: the HTTPS connection to the forge plus whoever can push to `releases`. That is the only remaining gap for such a build, and the panel says so out loud. Developer ID signing plus notarization for the macOS package — which macOS requires for downloaded apps regardless — is still open and tracked in `docs/ROADMAP.md` under M5. ```bash scripts/release-key.py verify ~/.config/shift-release/release.pub \ manifest.json manifest.sig # check a channel by hand ``` ## Verifying a channel by hand ```bash SHIFT_UPDATE_URL="https://host/owner/repo/raw/branch/releases/shift-update.json" \ cargo test --lib update:: -- --ignored # from src-tauri/ ``` ```bash cargo test --lib update:: -- --ignored # unsigned build: integrity only SHIFT_UPDATE_PUBKEY= cargo test --lib update:: -- --ignored ``` The test fetches the real feed, checks the manifest signature when the build carries a key, then downloads the artifact for this platform and architecture and checks size and hash against what is on disk. A baked key that is valid but not the publisher's fails the test, which is the proof that the check is real and not a pass-through. ## State of the published channel Checked against the live `releases` branch of this repository: Version **0.1.3**, checked by downloading each artifact back from the raw endpoint rather than trusting the local copy: | Key | Artifact | Verified from the remote | | --- | --- | --- | | `macos` | `Shift-macOS-universal.dmg` (55,528,662 B) | bytes downloaded, SHA256 matches the manifest | | `windows` | `Shift-Setup-0.1.3-win-x64.exe` (24,517,835 B) | bytes downloaded, SHA256 matches the manifest | | `windows-arm64` | not published | by choice — see "Supported release targets" | | `linux` | not published | by choice: Linux is a build target here, not a shipped product | `Shift-macOS-universal.pkg` sits on the same branch with no manifest key, for first installs ("macOS packaging"); it was hashed the same way and downloaded back to check. `shift-update.sig` verifies over the manifest's exact bytes against the key baked into both artifacts. The channel's home is derived from `origin` at build time, and this forge is Gitea, so the raw form is `/raw/branch/releases/`. Both artifacts carry the key and the feed URL, which is what makes an in-app upgrade possible at all: a build whose binary predates those constants reports "this build has no feed" and can never see the channel, however new it is. The forge runs **no Actions**, so nothing on this list was built by CI: every artifact was cross-compiled on one macOS host and pushed with `scripts/publish-release.sh`. Treat "pass CI" in the contributing guide as a definition to run locally, not a gate that already ran. A Linux release would slot in under the `linux` key with no code change — `artifact_keys_for` already answers `linux` for that host — so publishing one is a build task, not a feature.