fix(ci): pass the sign command with --config and prove the uninstaller was signed
Secret Scan / scan (push) Successful in 8s
Build App (Preview) / compute-version (pull_request) Successful in 12s
Secret Scan / scan (pull_request) Successful in 5s
Build App (Preview) / create-release (pull_request) Successful in 1s
Build App (Preview) / build-macos (pull_request) Successful in 2m48s
Build App (Preview) / test (pull_request) Successful in 3m41s
Build App (Preview) / build-linux (pull_request) Successful in 7m8s
Build App (Preview) / build-windows (pull_request) Failing after 8m53s
Build App (Preview) / prune-previews (pull_request) Skipped
Secret Scan / scan (push) Successful in 8s
Build App (Preview) / compute-version (pull_request) Successful in 12s
Secret Scan / scan (pull_request) Successful in 5s
Build App (Preview) / create-release (pull_request) Successful in 1s
Build App (Preview) / build-macos (pull_request) Successful in 2m48s
Build App (Preview) / test (pull_request) Successful in 3m41s
Build App (Preview) / build-linux (pull_request) Successful in 7m8s
Build App (Preview) / build-windows (pull_request) Failing after 8m53s
Build App (Preview) / prune-previews (pull_request) Skipped
The first signing run built nothing signed, and "Verify signatures" failed it as intended. The Tauri 2 CLI never reads TAURI_CONFIG. It only sets that variable for tauri-build, so the sign command was dropped silently, just as the inline beforeBuildCommand override had been for as long as the Windows jobs have set it. The setup now writes a config file, and the build passes it with `cargo tauri build --config`. Review follow-ups: - The NSIS uninstaller is written to %TEMP% and signed from 32-bit makensis. SYSTEM's %TEMP% sits under System32, which WOW64 redirects for makensis but not for the x64 signtool, so they would disagree about where the file is. %TEMP% and %TMP% now point into the workspace. makensis ignores the sign command's exit code for the uninstaller, so windows-sign.ps1 logs every file it signs, and the verify step requires a logged signature under that temp directory. - The signing client is pinned by SHA-512, so the pin can be checked against nuget.org's published packageHash. - tauri-cli on Windows is pinned to =2.11.0 --locked, the @tauri-apps/cli version the Linux and macOS jobs run from the lockfile, instead of "^2". Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -845,8 +845,9 @@ unsigned or untimestamped, so an unsigned installer cannot ship quietly.
|
||||
|
||||
- `scripts/windows-signing-setup.ps1` runs once per job. It downloads the signing client
|
||||
(`Microsoft.ArtifactSigning.Client`) and a .NET runtime into `.code-signing/` in the workspace,
|
||||
**each pinned by version and hash**, writes the dlib's `metadata.json`, and exports
|
||||
`TAURI_CONFIG` with `bundle.windows.signCommand`. Nothing is installed on the build VM. To bump
|
||||
**each pinned by version and hash**, writes the dlib's `metadata.json`, and writes a Tauri
|
||||
config file with `bundle.windows.signCommand` that the build passes as
|
||||
`cargo tauri build --config`. Nothing is installed on the build VM. To bump
|
||||
a pin, take the hash from nuget.org / the .NET `releases.json`, never from your own download.
|
||||
- `scripts/windows-sign.ps1` is the sign command: `signtool sign /dlib` with SHA-256 and the
|
||||
Microsoft timestamp server, retried. Credentials never reach a command line — the dlib reads
|
||||
@@ -859,14 +860,17 @@ Secrets (repository): the three `AZURE_*` above plus `ARTIFACT_SIGNING_ENDPOINT`
|
||||
- **`metadata.json` excludes every credential but `EnvironmentCredential`.** The dlib uses
|
||||
`DefaultAzureCredential`, whose chain ends in `InteractiveBrowserCredential`; the runners run as
|
||||
SYSTEM, where that waits forever for a browser.
|
||||
- **The "Build Tauri app" steps must not set `TAURI_CONFIG`.** A step-level value overrides the
|
||||
one the setup exported and drops the sign command without an error — "Verify signatures" is
|
||||
what would catch it.
|
||||
- **The signing files live in the workspace, not `%TEMP%`.** The NSIS uninstaller is signed from
|
||||
inside 32-bit `makensis`, which runs the sign command under 32-bit PowerShell; WOW64 redirects
|
||||
SYSTEM's `%TEMP%` (under System32) for that process. The workspace is under
|
||||
`systemprofile\.cache`, which the VM junctions for exactly this (see the comment on
|
||||
`build-app.yml`'s "Build Tauri app").
|
||||
- **The sign command goes in through `--config`, never `TAURI_CONFIG`.** The v2 CLI does not read
|
||||
that variable — it only sets it, for tauri-build — so a config put there is dropped without an
|
||||
error. The Windows jobs set an inline `TAURI_CONFIG` for years and it never applied;
|
||||
"Verify signatures" is what exposed it, and it is what would catch a regression.
|
||||
- **The signing files and the job's `%TEMP%` live in the workspace.** The NSIS uninstaller is
|
||||
written to `%TEMP%` and signed from inside 32-bit `makensis`, under 32-bit PowerShell; WOW64
|
||||
redirects SYSTEM's own `%TEMP%` (under System32) for those processes but not for the x64
|
||||
signtool, so they would disagree about where the file is. The workspace is under
|
||||
`systemprofile\.cache`, which the VM junctions so both views resolve. makensis also ignores
|
||||
the sign command's exit code for the uninstaller, so `windows-sign.ps1` logs every file it
|
||||
signs and the verify step requires a logged signature under that temp directory.
|
||||
- **The build VM is `WindowsBuilder` (VM 110 on the Proxmox host `pve4`)**, carrying both the
|
||||
`winvm-builder` and `virtual-builder` runners in host mode. It has the Windows SDK's
|
||||
`signtool` (10.0.26100) but no .NET — hence the job-local runtime.
|
||||
|
||||
Reference in New Issue
Block a user