balena-cli/CONTRIBUTING.md

Contributing

The balena CLI is an open source project and your contribution is welcome!

• Install the dependencies listed in the NPM Installation section section of the installation instructions. Check the section Additional Dependencies too.
• Clone the balena-cli repository (or a forked repo, if you are not in the balena team), cd to it and run npm install.
• Build the CLI with npm run build or npm test, and execute it with ./bin/balena (on a Windows command prompt, you may need to run node .\bin\balena).

In order to ease development:

• npm run build:fast skips some of the build steps for interactive testing, or
• npm run test:source skips testing the standalone zip packages (which is rather slow)
• ./bin/balena-dev uses ts-node/register to transpile on the fly.

Before opening a PR, test your changes with npm test. Keep compatibility in mind, as the CLI is meant to run on Linux, macOS and Windows. balena CI will run test code on all three platforms, but this will only help if you add some test cases for your new code!

Semantic versioning, commit messages and the ChangeLog

When a pull request is merged, Balena's versionbot / Continuous Integration system takes care of automatically creating a new CLI release on both the npm registry and the GitHub releases page. The release version numbering adheres to the Semantic Versioning's concept of patch, minor and major releases. Generally, bug fixes and documentation changes are classed as patch changes, while new features are classed as minor changes. If a change breaks backwards compatibility, it is a major change.

A new version entry is also automatically added to the CHANGELOG.md file when a pull request is merged. Each pull request corresponds to a single version / release. Each commit in the pull request becomes a bullet point entry in the Changelog. The Changelog file should not be manually edited.

To support this automation, a commit message should be structured as follows:

The first line becomes a bullet point in the CHANGELOG file

Optionally, a more detailed description in one or more paragraphs.
The detailed description can be seen with `git log`, but it is not copied
to the CHANGELOG file.

Change-type: patch|minor|major

Only the first line of the commit message is copied to the Changelog file. The Change-type footer must be preceded by a blank line, and indicates the commit's semver change type. When a PR consists of multiple commits, the commits may have different change type values. As a whole, the PR will produce a release of the "highest" change type. For example, two commits mixing patch and minor change types will produce a minor CLI release, while two commits mixing minor and major change types will produce a major CLI release.

The commit message is parsed / checked by versionbot with the resin-commit-lint package.

Because of the way that the Changelog file is automatically updated from commit messages, which become the source of "what's new" for CLI end users, we advocate "meaningful commits" and user-focused commit messages. A meaningful commit is one that, in isolation, introduces a fix or feature (or part of a fix or feature) that makes sense at the Changelog level, and which leaves the CLI in a non-broken state. Sometimes, in the course of preparing a single pull request, a developer creates several commits as a way of saving their "work in progress", which may even fail to build (e.g. npm run build fails), and which is then fixed or undone by further commits in the same PR. In this situation, the recommendation is to "squash" or "fixup" the work-in-progress commits into fewer, meaningful commits. Interactive rebase is a good tool to achieve this: blog, docs.

Mixing multiple distinct features or bug fixes in a single commit is discouraged, because the description will likely not fit in the single-line Changelog bullet point and also because it makes it harder to review the pull request (especially a large one) and harder to isolate and revert individual changes in case a bug is found later on. Create a separate commit for each feature / bug fix, or even separate pull requests.

If you need to catch up with changes to the master branch while working on a pull request, use rebase instead of merge: docs.

If package.json is updated for dependencies listed in the repo.yml file (like balena-sdk), the commit message body should also include a line in the following format:

Update balena-sdk from 12.0.0 to 12.1.0

This allows versionbot to produce nested Changelog entries (with expandable arrows), pulling in commit messages from the upstream repositories. The following npm script can be used to automatically produce a commit with a suitable commit message:

npm run update balena-sdk ^12.1.0

The script will create a new branch (only if master is currently checked out), run npm update with the given target version and commit the package.json and npm-shrinkwrap.json files. The script by default will set the Change-type to patch or minor, depending on the semver change of the updated dependency. A major change type can specified as an extra argument:

npm run update balena-sdk ^12.14.0 patch
npm run update balena-sdk ^13.0.0 major

Editing documentation files (README, INSTALL, Reference website...)

The doc/cli.markdown file is automatically generated by running npm run build:doc (which also runs as part of npm run build). That file is then pulled by scripts in the balena-io/docs GitHub repo for publishing at the CLI Documentation page.

The content sources for the auto generation of doc/cli.markdown are:

• Selected sections of the README file.
• The CLI's command documentation in source code (lib/commands/ folder), for example:
  • lib/commands/push.ts
  • lib/commands/env/add.ts

The README file is manually edited, but subsections are automatically extracted for inclusion in doc/cli.markdown by the getCapitanoDoc() function in automation/capitanodoc/capitanodoc.ts.

The INSTALL*.md and TROUBLESHOOTING.md files are also manually edited.

Windows

Besides the regular npm installation dependencies, the npm run build:installer script that produces the .exe graphical installer on Windows also requires NSIS and MSYS2 to be installed. Be sure to add C:\Program Files (x86)\NSIS to the PATH, so that makensis is available. MSYS2 is recommended when developing the balena CLI on Windows.

If changes are made to npm scripts in package.json, don't assume that a Unix shell like bash is available. For example, some Windows shells don't have the cp and rm commands, which is why you'll often find ncp and rimraf used in package.json scripts.

Updating the 'npm-shrinkwrap.json' file

The npm-shrinkwrap.json file is used to control package dependencies, as documented at https://docs.npmjs.com/files/shrinkwrap.json.

Changes to npm-shrinkwrap.json can be automatically merged by git during operations like rebase, pull and cherry-pick, but in some cases this results in suboptimal dependency resolution (the node_modules folder may end up larger than necessary, with consequences to CLI load time too). For this reason, the recommended way to update npm-shrinkwrap.json is to run npm install, possibly alongside npm dedupe as well. The following commands can be used to fix shrinkwrap issues and optimize the dependencies:

git checkout master -- npm-shrinkwrap.json
rm -rf node_modules
npm install  # update npm-shrinkwrap.json to satisfy changes to package.json
npm dedupe   # deduplicate dependencies from npm-shrinkwrap.json
npm install  # re-add optional dependencies removed by dedupe
git add npm-shrinkwrap.json  # add it for committing (solve merge errors)

Note that npm dedupe should always be followed by npm install, as shown above, even if npm install had already been executed before npm dedupe.

Optionally, these steps may be automated by installing the npm-merge-driver:

npx npm-merge-driver install -g

fast-boot and npm link - modifying the node_modules folder

During development or debugging, it is sometimes useful to temporarily modify the node_modules folder (with or without making the respective changes to the npm-shrinkwrap.json file), replacing dependencies with different versions. This can be achieved with the npm link command, or by manually editing or copying files to the node_modules folder.

Unexpected behavior may then be observed because of the CLI's use of the fast-boot2 package that caches module resolution. fast-boot2 is configured in lib/fast-boot.ts to automatically invalidate the cache if changes are made to the package.json or npm-shrinkwrap.json files, but the cache won't be automatically invalidated if npm link is used or if manual modifications are made to the node_modules folder. In this situation:

• Manually delete the module cache file (typically ~/.balena/cli-module-cache.json), or
• Use the bin/balena-dev entry point (instead of bin/balena) as it does not activate fast-boot2.

TypeScript and oclif

The CLI currently contains a mix of plain JavaScript and TypeScript code. The goal is to have all code written in Typescript, in order to take advantage of static typing and formal programming interfaces. The migration towards Typescript is taking place gradually, as part of maintenance work or the implementation of new features.

Of historical interest, the CLI was originally written in CoffeeScript and used the Capitano framework. All CoffeeScript code was migrated to either Javascript or Typescript, and Capitano was replaced with oclif. A few file or variable names still refer to this legacy, for example automation/capitanodoc/capitanodoc.ts.

Programming style

npm run build also runs balena-lint, which automatically reformats the code. Beyond that, we have a preference for Javascript promises over callbacks, and for async/await over .then().

Common gotchas

One thing that most CLI bugs have in common is the absence of test cases exercising the broken code, so writing some test code is a great idea. Having said that, there are also some common gotchas to bear in mind:

• Forward slashes ('/') vs. backslashes ('') in file paths. The Node.js path.sep variable stores a platform-specific path separator character: the backslash on Windows and the forward slash on Linux and macOS. The path.join function builds paths using such platform-specific path separator. However:

  • Note that Windows (kernel, cmd.exe, PowerShell, many applications) accepts both forward slashes and backslashes as path separators (including mixing them in a path string), so code like mypath.split(path.sep) may fail on Windows if mypath contains forward slashes. The path.parse function understands both forward slashes and backslashes on Windows, and the path.normalize function will replace forward slashes with backslashes.
  • In tar streams sent to the Docker daemon and to balenaCloud, the forward slash is the only acceptable path separator, regardless of the OS where the CLI is running. Therefore, path.sep and path.join should never be used when handling paths in tar streams! path.posix.join may be used instead of path.join.

• Avoid using the system shell to execute external commands, for example:
  child_process.exec('ssh "arg1" "arg2"');
child_process.spawn('ssh "arg1" "arg2"', { shell: true });
  Besides the usual security concerns of unsanitized strings, another problem is to get argument escaping right because of the differences between the Windows 'cmd.exe' shell and the Unix '/bin/sh'. For example, 'cmd.exe' doesn't recognize single quotes like '/bin/sh', and uses the caret (^) instead of the backslash as the escape character. Bug territory! Most of the time, it is possible to avoid relying on the shell altogether by providing a Javascript array of arguments:
  spawn('ssh', ['arg1', 'arg2'], { shell: false});
  To allow for logging and debugging, the which package may be used to get the full path of a command before executing it, without relying on any shell:
  const fullPath = await which('ssh');
console.log(fullPath); # 'C:\WINDOWS\System32\OpenSSH\ssh.EXE'
spawn(fullPath, ['arg1', 'arg2'], { shell: false });

• Avoid the instanceof operator when testing against classes/types from external packages (including base classes), because npm install may result in multiple versions of the same package being installed (to satisfy declared dependencies) and a false negative may result when comparing an object instance from one package version with a class of another package version (even if the implementations are identical in both packages). For example, once we fixed a bug where the test:
  error instanceof BalenaApplicationNotFound
  changed from true to false because npm install added an additional copy of the balena-errors package to satisfy a minor balena-sdk version update:
  $ find node_modules -name balena-errors
node_modules/balena-errors
node_modules/balena-sdk/node_modules/balena-errors
  In the case of subclasses of TypedError, a string comparison may be used instead:
  error.name === 'BalenaApplicationNotFound'

Further debugging notes

• If you need to selectively run specific tests, it.only will not work in cases when authorization is required as part of the test cycle. In order to target specific tests, control execution via .mocharc.js instead. Here is an example of targeting the deploy tests.

  replace: spec: 'tests/**/*.spec.ts',

  with: spec: ['tests/auth/*.spec.ts', 'tests/**/deploy.spec.ts'],