indygreg / clitest-rs

Proposal for a command line testing tool implemented in Rust

clitest-rs

clitest-rs aims to be the following:

• An expressive / literate way to define tests of CLI applications.
• A standalone executable that can evaluate CLI tests. Just point it at an executable you want to test and some test files and it does the rest.
• A testing library that can be embedded into Rust code bases. Rust projects can execute integration tests via cargo test. Behavior of the test runner can be customized using the crate's Rust API to tailor to the specific use cases of individual projects.

This project is still in its design phase. We're working on the test file format definition and execution semantics before writing a lot of code because as we learned from deficiencies in prior art mistakes here are difficult to impossible to correct. So we want to get it right from the beginning.

Example Test File

Here's a high-level example showing the proposed syntax and some high-level features.

This is a test file. This line is ignored by the test runner.

The HTML processing instruction says this file and its tests should only
run on Windows.
<?clitest require=windows ?>

The code fence below defines commands to execute.

```
$ myapp help
MyApp Version 1.0

Usage: myapp <action>
```

Demonstrate that we can redirect output to a file and use common
coreutils programs `cut` and `hexdump` to effectively create a snapshot
test of content.

```
$ myapp hello > hello
$ cut -z -b 1-32 hello | hexdump -C
00000000  23 20 63 6c 69 74 65 73  74 2d 72 73 0a 0a 63 6c  |# clitest-rs..cl|
00000010  69 74 65 73 74 2d 72 73  20 61 69 6d 73 20 74 6f  |itest-rs aims to|
00000020  00                                                |.|
00000021

```

```
$ myapp process
< input to stderr of spawned process
>1 expected stdout output
>2 expected stderr output
```

```ignore
This code fence is ignored by the test runner. It does not define any commands.
```

.clitest Syntax

A .clitest is an expressive / literate file format to define tests that run executables and verify their behavior.

The syntax of .clitest file leverages Markdown syntax conventions and many .clitest files can also be parsed / rendered as Markdown. But well-formed Markdown is not a hard requirement. (Leveraging Markdown as the file format has the added benefit that a project's Markdown based documentation could potentially also be executed as tests to ensure the documentation is accurate.)

The most important part of the file format are test cases. These describe commands to execute and their expected behavior.

Test cases live between code fences, which are pairs of 3 backticks or tildes. e.g.

```
Code fence
```

~~~
Another code fence
~~~

Unlike the CommonMark specification, we do not (yet) support optionally prefixing the code fence with up to 3 spaces.

All lines outside of code fences are ignored with the exception of lines starting with a processing instruction with the clitest tag. e.g.

```
<?clitest this metadata is recognized ?>
```

```
<?xml this is ignored since the tag is xml not clitest ?>
```

```
  <?clitest ignored since it isn't at the start of a line ?>
```

Test Case Code Fences

Each code fence is parsed into a single test case. Each case can define 0 to N command invocations to run. The execution environment for each test case is shared among its invocations, allowing invoked commands to interact with each other.

The opening line of the code fence can contain an optional info block after the backticks or tildes, per the CommonMark specification. This info block content is made available to the test runner so it can influence behavior.

```info block content
...
```

Within each code fence we define a custom syntax for declaring commands to invoke and their expected output.

The initial content of each line can denote special meaning:

• $ denotes a command string to run.
• < define stdin to feed to the spawned command.
• [N] defines the expected exit code of the command. MUST be the final line for a command definition.
• >1 and >2 define expected output for stdout and stderr, respectively. If not used, stdout and stderr are merged. This syntax allows verification that output is written to a specific stdio file descriptor.
• All other lines are treated as expected output from the command.

Command Strings

Lines beginning with $ declare a command string. A command string is a POSIX shell inspired grammar that defines how to invoke processes.

Often, a command string is just the name of the program under test and its arguments. e.g.

$ myapp arg0 arg1

Redirection

Process file descriptors can be controlled using the redirection grammar [n]operation word.

[n] is an optional file descriptor. 0 is stdin, 1 is stdout, 2 is stderr.

word describes where to read/write from.

The < operation reads content from the file word and sends it to stdin.

The > operation redirects content from a file descriptor (default 1 / stdout) and writes it to the file word. The file is truncated after opening.

The >> operation behaves like > but appends to the destination file instead of truncating.

<& duplicates an input file descriptor. Default of 0 / stderr. word of - closes the file descriptor.

>& duplicates an output file descriptor. Default of 1 / stdout. word of - closes the file descriptor.

The test runner proxies the following virtual files or emulates them if they don't exist:

• /dev/null - A write only file that does nothing with data written to it. Typically used for redirecting output so it isn't display. e.g. >/dev/null to redirect stdout to nothing.

Examples:

Execute `myapp` and write its stdout to the file `stdout`.
```
$ myapp > stdout
```

Execute `myapp` and write its stderr to the file `stderr`.
```
$ myapp 2> stderr
```

Execute `myapp` and merge its stdout and stderr to the file `combined`.
```
$ myapp 1>&2 > combined
```

Execute `myapp` with stdout closed.
```
$ myapp 1>&-
```

Pipelines

Pipelines are sequences of 2 or more commands separated by the operator |.

The grammar is command1 | command2.

The stdout of command1 is connected to the stdin of command2 via a pipe.

The processes are spawned in the order they are defined by the pipeline.

The exit status of the pipeline is the exit status of the last command.

Examples:

Send stdout to stdin of another:

```
$ myapp | grep expected-output
```

Test Case Info Block Directives

Optional code fence info blocks may pass directives to the test runner to influence execution of that code fence. Directives are intended to be single words. But the library API allows custom parsing to be performed.

The following single word directives are recognized by default:

• ignore says to ignore this code fence. It will not be parsed as a test case.

Processing Instructions

HTML processing instructions (<?clitest ... ?>) are used to convey instructions to the test runner outside the context of a single test case.

The default test runner exposes various functionality via processing instructions. But behavior can be customized. Library customers may implement their own processing instructions.

Test Case Execution Semantics

By default each test case is executed in a temporary directory in order to try to ensure a reproducible test environment. The temporary directory is populated with the executable under test along with copies of additional files that have been registered.

Environment variables exposed to spawned processes are normalized by default to try to ensure a reproducible test environment. The following environment variables are set by default:

• PWD Current directory the test case is executing in.
• USER The value clitest, which may not exist on the current system.
• ...

Principles

Simplicity and intuitiveness

The test format should be simple, easy to read, and easy to understand.

We base the test format on Markdown, which is well understood among the target developer community. We leverage features of markdown - code fences and HTML processing instructions - which are generally known or familiar to many developers.

Embrace Coreutils

We embed a copy of the uutils Rust implementation of GNU coreutils. This allows users to harness as little or much power from these common utilities (like grep, sed, and awk) for additional text processing / analysis as they want.

By leaning on these common utilities we defer the onus of having to reinvent this complexity in our test format and test harness. Our test harness can focus on executing processes and not complex text stream processing.

By using the Rust coreutils implementation, test execution uses a deterministic version of these tools, not whatever version or variant you have on the system. On Windows you don't need to install Cygwin, msys, or run in WSL to get access to these utilities.

Commitment to Backwards Compatibility

We strive to not require end-users to rewrite test files when upgrading. If you need to spend hours updating tests when upgrading the version of this project you are using, we've failed.

The test file format should evolve to be backwards and forwards compatible.

The test execution semantics should also be highly backwards compatible.

When you upgrade the version of this project you are using hopefully the worst thing that happens is you need to run a command to migrate to slightly different file syntax or accept machine proposed changes to expectations defined in the test files.

Embrace the Concept of the Shell Without an Actual Shell

While it could be useful to execute tests in a shell (like bash, zsh, or even nushell), we reject the existence of an explicit shell during test execution (unless end-users opt into it).

Practical experience with Mercurial and cram demonstrated there are too many sharp edges with actual shells. At the least you need a shell binary on every machine under test. And the shell version/features needs to behave the same on all systems. This can be difficult, especially on Windows.

A fully embeddable shell (like Nushell) could be a viable alternative. But now we're externalizing knowledge of a specific shell onto our end-users.

End-users are opinionated about shells. Bash. Zsh. Powershell. We don't want to pick sides in this debate.

But shells are useful. Variables. Pipes. Redirection. Control flow and loops. Builtin commands (like cd and echo). Background process execution.

We want to expose some of the features of shells without using an actual shell. Make the user think they have a shell for common functionality to test command invocations but don't let them juggle chainsaws.

TODO

Here are some ideas for the test format and execution semantics that we'd like to flush out more. Many are potential future features we could add to the test runner. We should focus on defining an extensible test format so future features don't require new syntax.

• Define syntax for regex matching of single lines.
• Define syntax for eliding multiple lines of expected output.
• Define escaping / alternative directive encoding mechanism so expected process output colliding with our special syntax can be worked around and allow expression of all outputs in the test format.
• Define syntax for matching common / repeated patterns. Logically allow expansion of tokens in expected output. We'll want to define some tokens by default, such as the current executable name and working directory.
• Define mechanism to verifying binary output. Maybe a directive / info block to escape output as hex or base64? Consider humans wanting to add arbitrary line breaks so 1 byte insertions/removals don't invalidate all following lines.
• Define additional minimal environment variables. (e.g. TZ, PATH.) Stuff that is in most environments.
• Syntax for setting additional environment variables.
• Consider % and possibly other line-leading tokens inside code fences to influence operations. e.g. % cd foo to change the CWD for future command invocations.
• Consider shell-like file redirection syntax for sending output of a command to a file, including to /dev/null (or its equivalent).
• Consider some magical commands or syntax (like $ cat <path>) to allow verification of produced file content.
• Mechanism for conditional execution. e.g. only execute on Windows or POSIX. Only execute if allowed to access a network. Maybe allow tags to be specified to the test runner so tests can be filtered. This can allow expression and skipping of slow tests. See also requires and keywords mechanisms in Mercurial's test harness.
• Command / test timeout support. How do you express that?
• Support for delivering a signal to a process. e.g. simulate a ^C to test error handling.
• Consider semantics for automatic drift overwrites when the test runner executes. We want to expose a turnkey mechanism where new/different/drifted test output can be recorded in the original test file without humans having to edit the file. The more custom syntax we add to the file format the harder this becomes. e.g. if you support regular expression matching of lines, how do you preserve those expressions or tokens when overwriting expected output? This is a known wort in Mercurial's and cram's more expressive output processing world.
• Define semantics around stdio file descriptor buffers. Could be nice to test differences when stdout is buffered/unbuffered.
• Consider ability to implement an interactive execution mode where you can step through executions. This allows developers to attach debuggers or inspect the filesystem sandbox when they want. It could also allow selectively accepting/rejecting/modifying output drift.
• Consider mechanism to define stdin/stdout in separate files. Useful for large content, like verifying behavior over input data sets. Can also be useful for sharing inputs across files and test cases.
• Consider mechanism for reusing a series of common commands. e.g. common test fixture setup code. How can we enable developers to minimize the overhead of authoring tests by minimizing DRY violations.
  • Idea: directories containing seeds for filesystem sandboxes. Test files or cases can specify directories to use to populate content of the filesystem sandbox.
  • Idea: Directive so test files or cases can specify another .clitest file whose single test case can be prepended to cases.
  • Idea: code fence info block to name a test case so it can be included in another one.
  • Idea: code fence info block to denote the prepending of content to all test cases.
• Consider allowing test cases to execute within a shell. Maybe a shell implemented in pure Rust for portability.

Unresolved Grammar / Parsing Issues

Assuming that command strings are denoted by lines beginning with $ and expected process output is following lines, there are some ambiguities with parsing the expected output.

How do we know we've hit the end of process output? In Mercurial and cram, lines indented by 2 spaces constituted output. So you could just look for the first non-indented line to find EOF. But we don't propose indenting lines. (It is annoying - if you have blank lines you need to insert empty lines with leading whitespace, which editors like to strip. Some editors may not clearly display the whitespace only lines. Behavior is not intuitive and not developer friendly!)

You can't say an empty line indicates end of output because there could be empty lines in process output. You also don't want to strip trailing lines because that whitespace could be relevant and you want to test it is there! (Don't you want to know if your program starts printing an extra trailing newline?)

We also have to consider the scenario that a command could print output that conflicts with our own file syntax.

The following scenarios are all ambiguous:

Is the actual output "foo" "foo\n" or "foo\n\n"?
```
$ myapp
foo

$ irrelevant
```

How do we validate output beginning with "$"?
Does the first command print nothing or "$ hello"?
```
$ echo '$ hello'
$ hello
```

What about commands that print code fence delimiters?
```
$ echo '```'
```
```

Mercurial solves the no trailing newline problem by annotating the line with a (no-eol) annotation. e.g.

```
$ echo -n hello
hello (no-eol)
```

For output that conflicts with our own syntax, the obvious solution is escaping. e.g.

```
$ echo '$ hello'
\$ hello

$ echo '```'
\`\`\`
```

Experience with Mercurial tells us that escaping can be annoying. Especially if your program domain contains a lot of output conflicting with our syntax. Imagine testing a program that emits Markdown using our tool! Or imagine us testing this tool using its own syntax! All that escaping could be pretty annoying.

One potential solution is alternative delimiters. For example a processing instruction or code fence info block could denote an alternative delimiter.

<%clitest command_delimiter=@ %>

```
@ echo '$ hello'
$ hello
```

Then end-users could select delimiters that make sense for their domain.

This doesn't solve the problem of code fence delimiters though!

Maybe we could leverage a heredoc style syntax for explicitly delimiting output?

```
$ echo '```' >< EOF
```
EOF
```

This could work. But it does make the file parser more complicated. Without this syntax you can scan the file and pair up code fence tokens to isolate all the lines defining code fences. Then you could parse each code fence later. But if you use a heredoc style syntax and have lines within the heredoc sharing the syntax as code fences, now your parser has to recognize the heredoc syntax so it knows to ignore a code fence token within a code fence. Not very desirable! Maybe code fence tokens deserve a one-off solution or escaping?

Project History

The goal of this project is to facilitate literate testing of command line applications. Command line applications are complex and are often designed with user interaction in mind. Their behavior is important to test and understand. Changes/differences in behavior are important to detect and capture to ensure consistent user experiences.

We believe that literate testing of CLI applications results in higher quality CLIs by increasing the surface area under test and forcing developers to confront their CLI UX. This belief is grounded in the experience of project contributors, namely around the use of the practice in the Mercurial version control tool. We want to lower the barrier to literate CLI testing in the Rust ecosystem and beyond to encourage its broader use.

At the time of this project's inception (mid 2023), literate CLI testing in the Rust ecosystem was not very mature. Popular CLI testing crates relied on spawning processes and inspecting their behavior from Rust code. See assert_cmd, insta-cmd, and rexpect. The trycmd seemed to be the lone attempt at higher-order literate CLI testing in Rust. But it had various limitations (see below).

This project is inspired by Mercurial's custom test harness. Mercurial inspired cram. And cram inspired trycmd.

This project came about due to various limitations with all of the above tools. We wanted to invent a modern, powerful literate command testing tool that exceeded capabilities of tools before while hopefully avoiding many pitfalls of earlier tools.

Here are some limitations with Mercurial and cram:

• Format was completely custom. Developers had to learn a new DSL for writing tests. It was relatively intuitive. But IDEs lacked parsing / highlighting.
• Indent of 2 spaces was kinda annoying to manually paste output differences. You were constantly doing multi-line indents in your editor.
• Line output processing features (like (re), (glob), (esc) (no-eol)) often got swallowed when outputs changed. The interaction of these directives often had unexpected consequences. You often needed to define these on several lines, which was annoying.
• Test files were effectively normalized to an auto-generated shell script which was executed on the current system. Commands were literal shell expressions. You were constantly fighting portability issues. e.g. on Windows you needed to install cygwin or msys to provide a shell and other common commands. You could have all the common problems with shell programming, including esoteric escaping and variable referencing issues. State leaking between commands. Poor debugging story. The full power of the shell could be useful but it was difficult to wield.
• Implemented in Python. Test harness was relatively slow. Required a Python interpreter to execute. Standalone binaries much easier for end-users.

And limitations with trycmd:

• Public library API does not facilitate advanced use. You can't construct test cases via the Rust API. See issue 217.
• Extensions to file format not possible. Related to above. But the low-level test format parser doesn't seem designed with an extensibility mechanism in place. See issue 218.
• Currently lacking a lot of features that Mercurial and cram have. See their enhancements list for them all. Examples include: lack of piping, lack of output redirection, lack of dynamic capture/matching, turnkey support for coreutils commands, filesystem sandboxing for .trycmd tests, handling escaping when there is a conflict between command output and trycmd test file syntax.
• No standalone test runner/binary. Scope seems limited to embedding a test runner in Rust crates.

Trycmd feels like the most modern command testing tool in the Rust space right now since it supports a literate test format and is relatively easy to use (e.g. no Python dependency). But it is still a far cry from the features of Mercurial and cram. The crate maintainer doesn't seem interested in opening up the crate's public API to support additional customization. This means we'll need to build something ourselves. We should hopefully be able to leverage some of trycmd's prior art, such as the snapbox crate.