| name | ci-cd |
| description | Use when creating, debugging, or improving GitHub Actions workflows for this repo — especially the crates.io release orchestrator and the downstream CLI and SDK release workflows for PyPI, npm, GitHub Releases, Homebrew tap, and NuGet. Also covers SDK distribution README branding alignment across npm, PyPI, and the main repo. Covers cargo-workspaces publish flow, release orchestration, OpenSSL vendoring, Windows vs non-Windows platform-conditional dependencies, manylinux containers, Perl module gaps, workspace scoping, native vs QEMU arm64 runners, trusted publishing, artifact actions, toolchain parity, idempotency patterns, Homebrew formula generation, and cross-workflow chaining. |
CI/CD
Source of truth for release orchestration
| Role | Workflow file |
|---|
| Workspace crates publish entrypoint | .github/workflows/release-crate.yml |
| GitHub Release binaries | .github/workflows/release-github.yml |
| Homebrew tap update | .github/workflows/release-homebrew.yml |
| CLI npm release | .github/workflows/release-npm.yml |
| CLI PyPI release | .github/workflows/release-pypi.yml |
| CLI NuGet release | .github/workflows/release-nuget.yml |
| SDK npm release | .github/workflows/release-sdk-npm.yml |
| SDK PyPI release | .github/workflows/release-sdk-pypi.yml |
| SDK Ruby gem release | .github/workflows/release-sdk-gem.yml |
release-crate.yml is the top of the chain. It publishes Rust crates first, then fans out to the CLI distribution workflows, and then conditionally triggers SDK workflows when the published crate set says they need to run.
Source of truth for distribution channel READMEs
| File | Published as |
|---|
README.MD | GitHub repo — canonical brand voice |
npm/smbcloud-cli/README.md | npmjs.com — @smbcloud/cli |
pypi/README.md | pypi.org — smbcloud-cli |
npm/README.md.tmpl | npmjs.com — @smbcloud/cli-* (platform packages) |
When the main README.MD changes its tagline, logo, quick start, or value proposition, update all three distribution READMEs to match.
Universal workflow patterns
The repo now has two kinds of release workflows.
Trigger shape
Orchestrator workflow:
release-crate.yml is the only release workflow that should run directly from a tag push.
on:
push:
tags:
- "v*.*.*"
workflow_dispatch:
inputs:
tag:
description: "Release tag (e.g. v0.3.35)"
required: true
Downstream workflows:
Everything else in the release chain should be workflow_dispatch only. The orchestrator passes the tag in explicitly.
on:
workflow_dispatch:
inputs:
tag:
description: "Release tag (e.g. v0.3.35)"
required: true
Checkout — always include the ref override
- name: Checkout
uses: actions/checkout@v6
with:
ref: ${{ github.event.inputs.tag || github.ref }}
Without ref:, a workflow_dispatch run checks out the default branch HEAD, not the tagged commit. The source code and the version would silently diverge.
Version resolution
For simple downstream workflows, resolving RELEASE_VERSION is enough:
- name: Set the release version
shell: bash
run: |
if [[ "${GITHUB_REF_TYPE}" == "tag" ]]; then
release_version="${GITHUB_REF_NAME#v}"
else
release_version="${{ github.event.inputs.tag }}"
release_version="${release_version#v}"
fi
echo "RELEASE_VERSION=${release_version}" >> "$GITHUB_ENV"
For the orchestrator, resolve both the full tag and the stripped version:
- name: Set the release version
id: release
shell: bash
run: |
if [[ "${GITHUB_REF_TYPE}" == "tag" ]]; then
release_tag="${GITHUB_REF_NAME}"
release_version="${GITHUB_REF_NAME#v}"
else
release_tag="${{ github.event.inputs.tag }}"
release_tag="${release_tag#refs/tags/}"
release_version="${release_tag#v}"
fi
echo "RELEASE_TAG=${release_tag}" >> "$GITHUB_ENV"
echo "RELEASE_VERSION=${release_version}" >> "$GITHUB_ENV"
echo "tag=${release_tag}" >> "$GITHUB_OUTPUT"
Never read the version from Cargo.toml with sed in the else branch. That quietly turns a tag-driven release into a whatever-is-on-branch release.
Toolchain — always read from rust-toolchain.toml
- name: Read Rust toolchain
shell: bash
run: |
rust_toolchain="$(sed -n 's/^channel = "\(.*\)"/\1/p' rust-toolchain.toml | head -n 1)"
if [ -z "$rust_toolchain" ]; then
echo "Failed to read Rust toolchain from rust-toolchain.toml" >&2
exit 1
fi
echo "RUST_TOOLCHAIN=${rust_toolchain}" >> "$GITHUB_ENV"
- name: Install toolkit
uses: dtolnay/rust-toolchain@3c5f7ea28cd621ae0bf5283f0e981fb97b8a7af9
with:
toolchain: ${{ env.RUST_TOOLCHAIN }}
Never hardcode the toolchain version or use @stable. The pinned SHA on dtolnay/rust-toolchain is required because the action does not publish semver tags — its only named ref is @master. Pinning to a commit SHA makes the action immutable: it will always refer to that exact tree regardless of what the maintainer pushes later, and it cannot be silently replaced if the account were ever compromised. For first-party actions (actions/checkout, actions/upload-artifact, etc.) the @v6, @v7 tags are maintained under a strict semver contract by GitHub, so floating major-version tags are safe there.
Rust cache — use a per-target key
- name: Setup Rust cache
uses: Swatinem/rust-cache@v2
with:
key: <job-name>-${{ matrix.target }}
Without a unique key, cache entries from different targets can collide and corrupt each other.
Idempotency — check before publishing
Every publish step must be safe to rerun.
crates.io workspace release:
This repo no longer publishes a single crate with cargo publish --package ... in CI. The crates workflow uses cargo workspaces plan --skip-published --json to decide whether there is anything left to do.
- name: Build unpublished crate plan
id: plan
shell: bash
env:
CARGO_REGISTRY_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}
run: |
cargo workspaces plan --skip-published --json --token "${CARGO_REGISTRY_TOKEN}" > plan.json
plan_count="$(jq 'length' plan.json)"
echo "count=${plan_count}" >> "$GITHUB_OUTPUT"
- name: Publish workspace crates to crates.io
if: steps.plan.outputs.count != '0'
run: |
cargo workspaces publish \
--publish-as-is \
--yes \
--locked \
--allow-branch "*" \
--publish-interval 15 \
--token "${CARGO_REGISTRY_TOKEN}"
If the plan is empty, stop cleanly and let the workflow say so.
PyPI:
- name: Check whether release already exists on PyPI
id: pypi-check
shell: bash
run: |
if curl -fsS "https://pypi.org/pypi/smbcloud-cli/${RELEASE_VERSION}/json" >/dev/null 2>&1; then
echo "exists=true" >> "$GITHUB_OUTPUT"
else
echo "exists=false" >> "$GITHUB_OUTPUT"
fi
- name: Publish distribution
if: steps.pypi-check.outputs.exists != 'true'
uses: pypa/gh-action-pypi-publish@release/v1
with:
skip-existing: true
skip-existing: true is a safety net for partial re-runs where some wheels already exist but others do not.
npm: The check is inline in the publish shell script:
if npm view "${npm_package_name}@${node_version}" version >/dev/null 2>&1; then
echo "${npm_package_name}@${node_version} already exists on npm, skipping publish"
exit 0
fi
npm publish --access public
The exit 0 inside the run block is correct here because it prevents the npm publish line below it in the same script from executing. This is different from the PyPI anti-pattern where exit 0 in a separate step does not prevent the next step from running.
Version guard — match the check to the workflow
For single-package workflows, checking one package against the tag is still fine.
For the crates orchestrator, check the workspace members under /crates as a group and be honest about what you are guarding. The point is to catch a bad tag early, not to pretend this is a single-package publish flow.
Artifact actions
Use actions/upload-artifact@v7 and actions/download-artifact@v7 consistently across all workflows.
Workspace scoping — use --package for build jobs, not for the crates orchestrator
For normal build jobs, do not build the whole workspace by accident.
This workspace includes smbcloud-auth-sdk-py, a PyO3 cdylib that requires Python symbols at link time. Building the full workspace on platforms without a matching Python interpreter causes:
ld: symbol(s) not found for architecture arm64
clang: error: linker command failed with exit code 1
Scope cargo build commands like this:
cargo build --release --locked --target <target> --package smbcloud-cli
The crates release orchestrator is the exception. That workflow is supposed to operate at the workspace root with cargo workspaces plan and cargo workspaces publish.
The PyPI workflow is also exempt because maturin is already pointed at the right crate via manifest-path.
Platform matrix
Current runners and targets
| Platform | Runner | Target | Notes |
|---|
| Linux x86_64 | ubuntu-latest | x86_64-unknown-linux-gnu | Native |
| Linux arm64 | ubuntu-24.04-arm | aarch64-unknown-linux-gnu | Native — do not use QEMU |
| Windows x64 | windows-2022 | x86_64-pc-windows-msvc | Native |
| Windows arm64 | windows-2022 | aarch64-pc-windows-msvc | Do not use for PyPI wheels |
| macOS x64 | macos-15-intel | x86_64-apple-darwin | Native Intel runner |
| macOS arm64 | macos-latest | aarch64-apple-darwin | Native Apple Silicon |
For npm and GitHub Release, use native runners. For PyPI (maturin), the macOS x64 build uses macos-latest with cross-compilation because maturin handles it transparently inside its container. Windows arm64 is the exception: GitHub's hosted Windows runners only provide x64 Python, so maturin will refuse to build an aarch64-pc-windows-msvc wheel when the interpreter reports win-amd64.
Do not use QEMU for arm64 Linux
When ubuntu-latest (x86_64) targets aarch64-unknown-linux-gnu via maturin, the manylinux container runs under QEMU emulation. This is slow and causes container toolchain mismatches. Use ubuntu-24.04-arm (native arm64 runner) instead.
OpenSSL vendoring
This repo uses git2 which depends on libgit2-sys → openssl-sys. On non-Windows platforms without system OpenSSL headers (manylinux containers, macOS cross-compilation), the build fails with:
Could not find directory of OpenSSL installation
On Windows, libgit2-sys uses WinHTTP (built into Windows) for HTTPS — OpenSSL is not needed at all. Enabling vendored on Windows causes openssl-src to invoke Perl to compile OpenSSL from source, which fails because the Windows runner Perl is missing required modules.
Fix applied
Cargo.toml (workspace) — declares the vendored feature once:
openssl = { version = "0.10", features = ["vendored"] }
crates/cli/Cargo.toml — activates it only on non-Windows targets:
[target.'cfg(not(windows))'.dependencies]
openssl = { workspace = true }
This activates openssl-sys/vendored via Cargo feature unification on Linux and macOS (where it is needed), and leaves the openssl crate completely absent from the Windows dependency graph (where WinHTTP handles SSL natively).
Per-platform behaviour
| Platform | SSL backend | openssl-sys/vendored | OpenSSL compiled from source |
|---|
| Linux (gnu) | vendored OpenSSL | ✅ | ✅ |
| macOS arm64 | vendored OpenSSL | ✅ | ✅ |
| macOS x64 | vendored OpenSSL | ✅ | ✅ |
| Windows x64 | WinHTTP (OS native) | ❌ | ❌ |
| Windows arm64 | WinHTTP (OS native) | ❌ | ❌ |
manylinux containers and Perl
The vendored OpenSSL build compiles OpenSSL from source by running perl ./Configure. The manylinux2014 container (CentOS 7) has a stripped Perl installation missing core modules.
Symptom
Can't locate IPC/Cmd.pm in @INC
Can't locate Time/Piece.pm in @INC
Fix
- name: Build wheel
uses: PyO3/maturin-action@v1
with:
before-script-linux: yum install -y perl-core
perl-core is the CentOS 7 meta-package that installs all standard Perl core modules. Install it once rather than chasing individual missing modules — each new OpenSSL version can require additional ones.
before-script-linux runs inside the manylinux Docker container, not the host runner. It is ignored on macOS and Windows runners.
manylinux2014 uses yum, not apt
The manylinux2014 image is CentOS 7 based. Use yum. Using apt fails with exit code 127 ("command not found").
manylinux target per architecture
| Architecture | manylinux target | Container base | Package manager |
|---|
| x86_64 | 2014 | CentOS 7 | yum |
| aarch64 | 2014 | CentOS 7 (via native runner) | yum |
Do not use manylinux: "2014" for aarch64 with an x86_64 host — the cross-compilation container does not have yum. Use a native ubuntu-24.04-arm runner with manylinux: "2014" instead.
Trusted publishing
PyPI — uses OIDC trusted publishing
permissions:
id-token: write
contents: read
- name: Publish distribution
uses: pypa/gh-action-pypi-publish@release/v1
No PYPI_TOKEN secret needed. The action exchanges the GitHub OIDC JWT for a short-lived PyPI token automatically.
Each PyPI package × workflow combination needs its own trusted publisher. The CLI package (smbcloud-cli) is configured for release-pypi.yml, but the SDK package (smbcloud-sdk-auth) needs a separate trusted publisher pointing at release-sdk-pypi.yml. Configure it at https://pypi.org/manage/project/<package>/settings/publishing/ with:
- Owner:
smbcloudXYZ
- Repository:
smbcloud-cli
- Workflow name: the exact
.yml filename that publishes that package
- Environment: leave blank
If the trusted publisher is missing, the publish step fails with invalid-publisher: valid token, but no corresponding publisher.
npm — does NOT support trusted publishing natively
npm has no OIDC-based trusted publishing equivalent to PyPI. The id-token: write permission in the npm workflow is unused for auth.
Authentication requires a stored secret:
- name: Publish to NPM
env:
NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
setup-node with registry-url writes an .npmrc referencing ${NODE_AUTH_TOKEN}. That variable must be in the step's env block at publish time. If it is missing, npm publish silently has no auth and returns 404 for new packages (existing packages that pass the npm view idempotency check are never attempted, masking the missing token).
The 404 on npm publish does not mean the package does not exist. npm returns 404 for auth failures to avoid leaking package existence.
npm trusted publishing (future): npm does support OIDC trusted publishing via npm trust github. It requires the package to exist on npm first. Configure it per-package after the first manual publish. See smbcloud-cli-release skill for details.
crates.io — uses CARGO_REGISTRY_TOKEN
- name: Publish workspace crates to crates.io
env:
CARGO_REGISTRY_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}
run: |
cargo workspaces publish \
--publish-as-is \
--yes \
--locked \
--allow-branch "*" \
--publish-interval 15 \
--token "${CARGO_REGISTRY_TOKEN}"
This repo uses cargo-workspaces for the Rust release flow. CI is not doing the interactive version bump step. It is publishing the already-prepared tagged workspace state.
crates.io also supports OIDC trusted publishing, but that is separate from the current workflow.
Split build + release job pattern
For the GitHub Release workflow, separate the build matrix from the release step:
jobs:
build:
strategy:
matrix: ...
steps:
- name: Build binary
run: cargo build ...
- name: Upload artifact
uses: actions/upload-artifact@v7
release:
needs: build
runs-on: ubuntu-latest
steps:
- name: Download all artifacts
uses: actions/download-artifact@v7
with:
pattern: "binary-*"
merge-multiple: true
- name: Release
uses: softprops/action-gh-release@v2
with:
files: release/*
This ensures all platform binaries are collected before any are attached to the GitHub Release. If a single platform build fails, no partial release is created.
Chaining workflows — use explicit dispatches
Do not rely on GitHub events magically fanning out through the release chain. Keep the chain explicit.
GitHub does not fire downstream repository events the way people expect when the upstream action used GITHUB_TOKEN. If you want one release workflow to kick off another, dispatch it yourself.
The current release chain is:
release-crate.yml publishes workspace crates
- then dispatches CLI distribution workflows:
release-github.yml
release-npm.yml
release-nuget.yml
release-pypi.yml
- then conditionally dispatches SDK workflows based on the crate publish plan:
release-sdk-pypi.yml
release-sdk-npm.yml
release-sdk-gem.yml
release-github.yml dispatches release-homebrew.yml after the GitHub Release assets exist
This pattern needs actions: write in the upstream workflow permissions.
permissions:
contents: write
actions: write
- name: Trigger Homebrew release
uses: actions/github-script@v7
with:
script: |
await github.rest.actions.createWorkflowDispatch({
owner: context.repo.owner,
repo: context.repo.repo,
workflow_id: 'release-homebrew.yml',
ref: context.payload.repository.default_branch,
inputs: {
tag: '${{ steps.tag.outputs.tag }}'
}
})
A few practical rules:
- downstream workflows must expose
workflow_dispatch with a tag input
- dispatch from the repo default branch so the workflow definition comes from the current release lane
- still check out the requested tag inside the downstream workflow so the built code matches the release
- do not dispatch Homebrew until the GitHub Release job has uploaded the assets it needs
This is a fire-and-forget chain. If you want one green check to mean the whole release tree passed, you need a more involved callback or polling design. Right now the pattern is good enough, but it is not magic.
Homebrew tap release pattern
release-homebrew.yml is triggered by release-github.yml via workflow_dispatch (see above). It runs a single cheap ubuntu-latest job — no Rust compilation, no macOS runners.
Key principle: compute SHA256 once, at build time
The Homebrew formula's sha256 field is the checksum of the tarball that Homebrew downloads at brew install time. The tarball lives at the GitHub Release URL. CI only needs to write that checksum into cli.rb — it does not need to download the tarball again.
In release-github.yml (build job, macOS legs only):
- name: Package macOS tarball for Homebrew
if: contains(matrix.target, 'apple-darwin')
shell: bash
run: |
ARCHIVE_NAME="${PROJECT_NAME}-${{ matrix.name }}.tar.gz"
mkdir -p staging-brew
cp "./release/${PROJECT_NAME}-${{ matrix.name }}" "staging-brew/${PROJECT_NAME}"
strip "staging-brew/${PROJECT_NAME}"
tar -C staging-brew -czf "${ARCHIVE_NAME}" "${PROJECT_NAME}"
shasum -a 256 "${ARCHIVE_NAME}" | awk '{print $1}' > "${ARCHIVE_NAME}.sha256"
- name: Upload macOS Homebrew artifacts
if: contains(matrix.target, 'apple-darwin')
uses: actions/upload-artifact@v7
with:
name: homebrew-${{ matrix.name }}
path: |
${{ env.PROJECT_NAME }}-${{ matrix.name }}.tar.gz
${{ env.PROJECT_NAME }}-${{ matrix.name }}.tar.gz.sha256
Both the .tar.gz and the .sha256 file are downloaded in the release job and attached to the GitHub Release via files: release/*.
In release-homebrew.yml (single job):
- name: Read SHA256 checksums from release
id: sha
env:
GH_TOKEN: ${{ github.token }}
shell: bash
run: |
mkdir -p artifacts
for attempt in 1 2 3 4 5; do
if gh release download "${{ steps.release.outputs.tag }}" \
--repo "${{ env.REPO }}" \
--pattern "smb-macos-*.tar.gz.sha256" \
--dir artifacts/; then
break
fi
if [ "$attempt" -eq 5 ]; then
echo "Failed to download Homebrew checksum artifacts from the GitHub release." >&2
exit 1
fi
echo "Release assets are not visible yet. Waiting 10 seconds..."
sleep 10
done
gh is pre-installed on GitHub-hosted runners. Add a small retry loop here because the GitHub Release exists first and the Homebrew workflow arrives a moment later. Usually that gap is tiny. Sometimes it is not.
Formula URL convention
The formula url: references the release asset directly. Homebrew fetches it at install time — CI never downloads it:
on_macos do
on_arm do
url 'https://github.com/smbcloudXYZ/smbcloud-cli/releases/download/v0.3.36/smb-macos-arm64.tar.gz'
sha256 '<computed during build>'
end
on_intel do
url 'https://github.com/smbcloudXYZ/smbcloud-cli/releases/download/v0.3.36/smb-macos-amd64.tar.gz'
sha256 '<computed during build>'
end
end
The tarball must contain the binary named smb (not smb-macos-arm64) because the formula's install step is bin.install 'smb'.
Homebrew tap permissions
release-homebrew.yml only needs contents: read on smbcloud-cli. Writing to the tap repo (smbcloudXYZ/homebrew-tap) is authenticated via secrets.HOMEBREW_TAP_TOKEN in the actions/checkout step, not via workflow-level permissions.
fail-fast: false
Always set fail-fast: false on release build matrices:
strategy:
fail-fast: false
matrix: ...
The default fail-fast: true cancels all in-progress jobs as soon as one fails. For release builds, you want every platform to complete so you can diagnose all failures in a single run rather than fixing them one at a time.
SDK distribution README branding
Every distribution channel README must stay aligned with the main README.MD. The three elements that must always match exactly:
- Logo —
https://avatars.githubusercontent.com/u/89791739?s=200&v=4, width 128
- Tagline — "Deploy to the cloud in one command."
- Quick start —
smb login → smb init → smb deploy
Structure per README type
Wrapper package (npm/smbcloud-cli/README.md → @smbcloud/cli, pypi/README.md → smbcloud-cli):
- Logo + tagline + nav links + badges
- About section (copy from main README verbatim)
- Primary install method for that channel first (
npm install -g or pip install)
- Quick start
- All other install methods (Homebrew, npm, pip, shell, PowerShell)
- Documentation link
- Platform support table
- Source & Issues pointer to GitHub
- License
The PyPI README leads with the PyPI badge; the npm README leads with the npm badge. Both include the full badge row.
Platform binary package (npm/README.md.tmpl → @smbcloud/cli-darwin-arm64, etc.):
Keep it minimal. Show the logo and tagline for brand recognition, then immediately redirect to @smbcloud/cli. Do not repeat install instructions — the user landed on the wrong package.
Platform support table
Keep this table consistent across all READMEs:
| Platform | Architecture |
| ------------- | ------------ |
| macOS | arm64, x64 |
| Linux (glibc) | arm64, x64 |
| Windows | arm64, x64 |
Update it whenever a new platform is added to the release matrix.
PyPI-specific copy
Add this line to distinguish the pip install from npm:
This package installs the native smb executable for your platform directly — no Node.js, no Docker, no runtime dependencies.
Template variable in README.md.tmpl
The platform package template uses ${node_pkg} as a shell-style variable. This is substituted at release time by the npm publish script. Do not replace it with a hardcoded package name.
SDK trigger rules
Use the crates publish plan to decide whether an SDK release should run.
Current mappings:
- SDK PyPI runs when the plan includes any of:
smbcloud-auth-sdk-py
smbcloud-auth-sdk
smbcloud-model
smbcloud-network
- SDK npm runs when the plan includes any of:
smbcloud-auth-sdk-wasm
smbcloud-auth-sdk
smbcloud-network
smbcloud-networking
- SDK Ruby gem runs when the plan includes any of:
smbcloud-auth-sdk
smbcloud-model
smbcloud-network
These are based on the actual wrappers in sdk/:
sdk/python/pyproject.toml points maturin at crates/smbcloud-auth-sdk-py/Cargo.toml
sdk/npm/smbcloud-auth/prepare-package.mjs builds from crates/smbcloud-auth-sdk-wasm
sdk/gems/auth/ext/auth/Cargo.toml pulls published crates from crates.io
One practical warning: the Ruby gem release depends on crates.io propagation. If crates were just published, wait for them to appear before building the gem.
Common mistakes
SDK npm version out of sync with Rust crate version
sdk/npm/smbcloud-auth/package.json must have the same version as crates/smbcloud-auth-sdk-wasm/Cargo.toml. The prepare-package.mjs script compares them at build time and throws a hard error on mismatch. When bumping workspace crate versions for a release, always update the npm package version in the same commit. Forgetting this causes the check-npm-sdk CI job and the release-sdk-npm workflow to fail.
SDK Ruby gem versions out of sync with workspace crate versions
Three files must be bumped together for each gem in sdk/gems/:
lib/<gem>/version.rb — the gem version (Auth::VERSION, Model::VERSION)
ext/<gem>/Cargo.toml — the native extension crate version
ext/<gem>/Cargo.toml dependencies — the smbcloud-* version constraints (e.g. "0.3" → "0.4")
After bumping, regenerate lockfiles with cargo generate-lockfile and bundle lock inside each gem directory. The release-sdk-gem workflow checks Auth::VERSION against the tag and fails on mismatch.
Tagging a release on a feature branch instead of development
Always tag on development (the mainline branch). Tagging on a feature branch leaves development without the release commit, making git history confusing and future releases error-prone. Merge the feature branch into development first, then tag. If a tag was already placed on the wrong commit, move it with git tag -f v<version> and git push origin v<version> --force.
"Re-run all jobs" on an old workflow run rebuilds from the old commit
GitHub Actions workflow re-runs always use the commit the run was originally dispatched with. If you moved a tag after the run was created, the re-run still checks out the old commit. Use "Re-run failed jobs" only when the code on that commit is correct and only an external issue (e.g. missing trusted publisher) caused the failure. If the code itself was fixed, dispatch a fresh run instead: gh workflow run <workflow>.yml -f tag=v<version>.
Building the full workspace in release workflows
Always pass --package smbcloud-cli to cargo build and cargo publish. Omitting it builds smbcloud-auth-sdk-py (PyO3 cdylib) which fails on platforms without a matching Python interpreter.
exit 0 in a separate step does not skip the next step
exit 0 only terminates the current run shell. The following step runs regardless. Use $GITHUB_OUTPUT + if: conditions for inter-step flow control.
Hardcoded toolchain version
Never hardcode toolchain: 1.92 or use @stable. Always read from rust-toolchain.toml.
Missing ref: on checkout for workflow_dispatch
Without ref: ${{ github.event.inputs.tag || github.ref }}, a manual dispatch checks out the default branch, not the tag. The compiled binary will be from the wrong commit.
Using sed on Cargo.toml as the version fallback
This reads whatever version is currently in Cargo.toml on the checked-out branch, which may not match the tag being re-published. Use the tag input directly.
QEMU for arm64 Linux
Using ubuntu-latest (x86_64) to build aarch64-unknown-linux-gnu via QEMU inside manylinux is slow and causes container mismatch errors. Use ubuntu-24.04-arm.
Windows arm64 PyPI wheels on hosted runners
Maturin needs a matching Python interpreter for wheel generation. On GitHub-hosted Windows runners, the available interpreter reports win-amd64, so an aarch64-pc-windows-msvc wheel build gets skipped and then fails. Unless you have an arm64 Windows runner with arm64 Python installed, leave Windows arm64 out of the PyPI wheel matrix.
yum vs apt in manylinux
manylinux2014 is CentOS 7. Use yum, not apt. For manylinux_2_28 (AlmaLinux 8) use dnf.
Missing User-Agent on crates.io API requests
crates.io rejects curl requests without a User-Agent header. Always include -H "User-Agent: smbcloud-cli-release-workflow".
Enabling vendored OpenSSL unconditionally on Windows
openssl = { features = ["vendored"] } as an unconditional dependency forces openssl-src to compile OpenSSL from source on Windows. Windows builds fail because the runner Perl is missing modules required by OpenSSL's Configure script. Use [target.'cfg(not(windows))'.dependencies] so the openssl crate is never pulled into the Windows dependency graph. libgit2-sys uses WinHTTP on Windows and does not need OpenSSL.
Distribution README out of sync with main README
When the main README.MD changes its tagline, logo URL, quick start commands, or value proposition, the npm and PyPI READMEs must be updated in the same commit. Stale distribution READMEs show outdated branding on npmjs.com and pypi.org — the highest-visibility surfaces for new users discovering the CLI.
Platform support table missing new targets
When a new platform is added to any release matrix (e.g. linux-arm64), update the platform support table in npm/smbcloud-cli/README.md and pypi/README.md in the same PR.
Use workflow-level env: for compile-time secrets, not $GITHUB_ENV export steps
Compile-time secrets like CLI_CLIENT_SECRET (read by env!()) must be available to every cargo invocation across all platforms. Writing to $GITHUB_ENV in a separate step is unreliable: it does not propagate into Docker containers, and on Windows runners the default PowerShell shell does not expand "$GITHUB_ENV" correctly.
Declare the secret in the workflow-level env: block:
env:
CLI_CLIENT_SECRET: ${{ secrets.CLI_CLIENT_SECRET }}
CARGO_TERM_COLOR: always
Never add a dedicated "Export build secret" step for this purpose.
manylinux Docker containers require docker-options to forward env vars
Even with the secret in the workflow-level env: block, maturin-action's manylinux builds run inside a Docker container that does not automatically inherit the runner's environment. The secret is visible to the action process (shown as CLI_CLIENT_SECRET: *** in the step log) but is not forwarded into the container unless explicitly requested.
Fix: add docker-options: -e CLI_CLIENT_SECRET to the Build wheel step:
- name: Build wheel
uses: PyO3/maturin-action@v1
with:
target: ${{ matrix.build.TARGET }}
manylinux: ${{ matrix.build.MANYLINUX || 'off' }}
working-directory: pypi
args: --release --locked --compatibility pypi --out dist
before-script-linux: yum install -y perl-core
docker-options: -e CLI_CLIENT_SECRET
-e CLI_CLIENT_SECRET tells docker run to forward that variable from the runner process into the container. Windows and macOS legs are unaffected (they do not use Docker). Without this, env!("CLI_CLIENT_SECRET") fails only on Linux manylinux legs while all other platforms succeed — making the root cause easy to miss.
PyPI blocks short and protocol-reserved package names
PyPI maintains a blocklist of names that are too generic or conflict with well-known protocols and tools. smb is blocked because it is the name of the Windows SMB/CIFS file sharing protocol. Attempting to publish a package named smb returns 400 The name 'smb' isn't allowed. There is no workaround — the name cannot be registered regardless of who owns it. If you need uvx <name> to work without --from, the package name and binary name must match AND the package name must not be on the blocklist. For this repo, uvx --from smbcloud-cli smb is the correct form.
Multiple release workflows all listening to tag pushes
Only release-crate.yml should run directly from a release tag. The downstream workflows should be workflow_dispatch only. If they all listen to push.tags, they race each other and you lose control of the order.
on: release: types: [published] does not fire when the release is created by automation
When softprops/action-gh-release creates a GitHub Release using the implicit GITHUB_TOKEN, GitHub does not dispatch the release event to other workflows. The downstream workflow just never runs. Use workflow_dispatch via actions/github-script@v7 instead, and give the upstream workflow actions: write.
Downloading full tarballs in the Homebrew workflow to compute SHA256
The Homebrew formula only needs the SHA256 hash. Compute it during the build in release-github.yml, attach the .sha256 files to the GitHub Release, and have release-homebrew.yml download only those tiny files. Do not re-download full binaries and do not rebuild Rust code there.
Forgetting that Homebrew depends on the GitHub Release assets
Homebrew cannot run first. It needs the .sha256 files attached by release-github.yml. Keep that dependency explicit.
Starting the Ruby gem release before crates.io has caught up
sdk/gems/auth/ext/auth/Cargo.toml pulls smbcloud-auth-sdk, smbcloud-model, and smbcloud-network from crates.io. If the gem workflow starts immediately after the crates publish job, the index may not be ready yet. Add a small wait-and-retry loop against the crates.io API before the gem build starts.
Rebuilding macOS binaries in release-homebrew.yml
release-github.yml already builds macOS binaries. Running a second Rust compilation in release-homebrew.yml wastes two macOS runner slots and risks producing binaries with different checksums if the environment differs. Build once, reuse the output.
Homebrew tarball contains wrong binary name
The formula uses bin.install 'smb'. The tarball must contain the binary named smb, not smb-macos-arm64. Strip and repackage correctly in the "Package macOS tarball for Homebrew" step.