Any Markdown file can be a test suite.

In order for it to be run as one, red_knot_test::run must be called with its path; see crates/red_knot_python_semantic/tests/mdtest.rs for an example that treats all Markdown files under a certain directory as test suites.

A Markdown test suite can contain any number of tests. A test consists of one or more embedded "files", each defined by a triple-backticks fenced code block. The code block must have a tag string specifying its language; currently only py (Python files) and pyi (type stub files) are supported.

The simplest possible test suite consists of just a single test, with a single embedded file:

```py
reveal_type(1)  # revealed: Literal[1]
```

When running this test, the mdtest framework will write a file with these contents to the default file path (/src/test.py) in its in-memory file system, run a type check on that file, and then match the resulting diagnostics with the assertions in the test. Assertions are in the form of Python comments. If all diagnostics and all assertions are matched, the test passes; otherwise, it fails.

See actual example mdtest suites in crates/red_knot_python_semantic/resources/mdtest.

ℹ️ Note: If you use dir-test, rstest or similar to generate a separate test for all Markdown files in a certain directory, as with the example in crates/red_knot_python_semantic/tests/mdtest.rs, you will likely want to also make sure that the crate the tests are in is rebuilt every time a Markdown file is added or removed from the directory. See crates/red_knot_python_semantic/build.rs for an example of how to do this.

This is because these macros generate their tests at build time rather than at runtime. Without the build.rs file to force a rebuild when a Markdown file is added or removed, a new Markdown test suite might not be run unless some other change in the crate caused a rebuild following the addition of the new test file.

Two kinds of assertions are supported: # revealed: (shown above) and # error:.

A # revealed: assertion should always be paired with a call to the reveal_type utility, which reveals (via a diagnostic) the inferred type of its argument (which can be any expression). The text after # revealed: must match exactly with the displayed form of the revealed type of that expression.

The reveal_type function can be imported from the typing standard library module (or, for older Python versions, from the typing_extensions pseudo-standard-library module¹):

from typing import reveal_type

reveal_type("foo")  # revealed: Literal["foo"]

For convenience, type checkers also pretend that reveal_type is a built-in, so that this import is not required. Using reveal_type without importing it issues a diagnostic warning that it was used without importing it, in addition to the diagnostic revealing the type of the expression.

The # revealed: assertion must always match a revealed-type diagnostic, and will also match the undefined-reveal diagnostic, if present, so it's safe to use reveal_type in tests either with or without importing it. (Style preference is to not import it in tests, unless specifically testing something about the behavior of importing it.)

A comment beginning with # error: is an assertion that a type checker diagnostic will be emitted, with text span starting on that line. The matching can be narrowed in three ways:

• # error: [invalid-assignment] requires that the matched diagnostic have the rule code invalid-assignment. (The square brackets are required.)
• # error: "Some text" requires that the diagnostic's full message contain the text Some text. (The double quotes are required in the assertion comment; they are not part of the matched text.)
• # error: 8 [rule-code] or # error: 8 "Some text" additionally requires that the matched diagnostic's text span begins on column 8 (one-indexed) of this line.

Assertions must contain either a rule code or a contains-text, or both, and may optionally also include a column number. They must come in order: first column, if present; then rule code, if present; then contains-text, if present. For example, an assertion using all three would look like # error: 8 [invalid-assignment] "Some text".

Error assertions in tests intended to test type checker semantics should primarily use rule-code assertions, with occasional contains-text assertions where needed to disambiguate or validate some details of the diagnostic message.

An assertion comment may be a line-trailing comment, in which case it applies to the line it is on:

x: str = 1  # error: [invalid-assignment]

Or it may be a comment on its own line, in which case it applies to the next line that does not contain an assertion comment:

# error: [invalid-assignment]
x: str = 1

Multiple assertions applying to the same line may be stacked:

# error: [invalid-assignment]
# revealed: Literal[1]
x: str = reveal_type(1)

Intervening empty lines or non-assertion comments are not allowed; an assertion stack must be one assertion per line, immediately following each other, with the line immediately following the last assertion as the line of source code on which the matched diagnostics are emitted.

Some tests require multiple files, with imports from one file into another. Multiple fenced code blocks represent multiple embedded files. Since files must have unique names, at most one file can use the default name of /src/test.py. Other files must explicitly specify their file name:

```py
from b import C
reveal_type(C)  # revealed: Literal[C]
```

```py path=b.py
class C: pass
```

Relative file names are always relative to the "workspace root", which is also an import root (that is, the equivalent of a runtime entry on sys.path).

The default workspace root is /src/. Currently it is not possible to customize this in a test, but this is a feature we will want to add in the future.

So the above test creates two files, /src/test.py and /src/b.py, and sets the workspace root to /src/, allowing test.py to import from b.py using the module name b.

A single test suite (Markdown file) can contain multiple tests, by demarcating them using Markdown header lines:

# Same-file invalid assignment

```py
x: int = "foo"  # error: [invalid-assignment]
```

# Cross-file invalid assignment

```py
from b import y
x: int = y  # error: [invalid-assignment]
```

```py path=b.py
y = "foo"
```

This test suite contains two tests, one named "Same-file invalid assignment" and the other named "Cross-file invalid assignment". The first test involves only a single embedded file, and the second test involves two embedded files.

The tests are run independently, in independent in-memory file systems and with new red-knot Salsa databases. This means that each is a from-scratch run of the type checker, with no data persisting from any previous test.

Due to cargo test limitations, an entire test suite (Markdown file) is run as a single Rust test, so it's not possible to select individual tests within it to run.

Markdown headers can also be used to group related tests within a suite:

# Literals

## Numbers

### Integer

```py
reveal_type(1)  # revealed: Literal[1]
```

### Float

```py
reveal_type(1.0)  # revealed: float
```

## Strings

```py
reveal_type("foo")  # revealed: Literal["foo"]
```

This test suite contains three tests, named "Literals - Numbers - Integer", "Literals - Numbers - Float", and "Literals - Strings".

A header-demarcated section must either be a test or a grouping header; it cannot be both. That is, a header section can either contain embedded files (making it a test), or it can contain more deeply-nested headers (headers with more #), but it cannot contain both.

Arbitrary Markdown syntax (including of course normal prose paragraphs) is permitted (and ignored by the test framework) between fenced code blocks. This permits natural documentation of why a test exists, and what it intends to assert:

Assigning a string to a variable annotated as `int` is not permitted:

```py
x: int = "foo"  # error: [invalid-assignment]
```

There are some designed features that we intend for the test framework to have, but have not yet implemented:

We may want to be able to assert that a diagnostic spans multiple lines, and to assert the columns it begins and/or ends on. The planned syntax for this will use <<< and >>> to mark the start and end lines for an assertion:

(3  # error: 2 [unsupported-operands] <<<
  +
 "foo")  # error: 6 >>>

The column assertion 6 on the ending line should be optional.

In cases of overlapping such assertions, resolve ambiguity using more angle brackets: <<<< begins an assertion ended by >>>>, etc.

Some tests may need to specify non-Python embedded files: typeshed stdlib/VERSIONS, pth files, py.typed files, pyvenv.cfg files...

We will allow specifying any of these using the text language in the code block tag string:

```text path=/third-party/foo/py.typed
partial
```

We may want to also support testing Jupyter notebooks as embedded files; exact syntax for this is yet to be determined.

Of course, red-knot is only run directly on py and pyi files, and assertion comments are only possible in these files.

A fenced code block with no language will always be an error.

We will add the ability to specify non-default red-knot configurations to use in tests, by including a TOML code block:

```toml
[tool.knot]
warn-on-any = true
```

```py
from typing import Any

def f(x: Any):  # error: [use-of-any]
    pass
```

It should be possible to include a TOML code block in a single test (as shown), or in a grouping section, in which case it applies to all nested tests within that grouping section. Configurations at multiple level are allowed and merged, with the most-nested (closest to the test) taking precedence.

Having each test in a suite always run as a distinct Rust test would require writing our own test runner or code-generating tests in a build script; neither of these is planned.

We could still allow running just a single test from a suite, for debugging purposes, either via some "focus" syntax that could be easily temporarily added to a test, or via an environment variable.

The red-knot TOML configuration format hasn't been designed yet, and we may want to implement support in the test framework for configuring search paths before it is designed. If so, we can define some configuration options for now under the [tool.knot.tests] namespace. In the future, perhaps some of these can be replaced by real red-knot configuration options; some or all may also be kept long-term as test-specific options.

Some configuration options we will want to provide:

• We should be able to configure the default workspace root to something other than /src/ using a workspace-root configuration option.

• We should be able to add a third-party root using the third-party-root configuration option.

• We may want to add additional configuration options for setting additional search path kinds.

Paths for workspace-root and third-party-root must be absolute.

Relative embedded-file paths are relative to the workspace root, even if it is explicitly set to a non-default value using the workspace-root config.

Some tests will need to override the default typeshed with custom files. The [tool.knot.tests] configuration option typeshed-root should be usable for this:

```toml
[tool.knot.tests]
typeshed-root = "/typeshed"
```

This file is importable as part of our custom typeshed, because it is within `/typeshed`, which we
configured above as our custom typeshed root:

```py path=/typeshed/stdlib/builtins.pyi
I_AM_THE_ONLY_BUILTIN = 1
```

This file is written to `/src/test.py`, because the default workspace root is `/src/ and the default
file path is `test.py`:

```py
reveal_type(I_AM_THE_ONLY_BUILTIN)  # revealed: Literal[1]
```

A fenced code block with language text can be used to provide a stdlib/VERSIONS file in the custom typeshed root. If no such file is created explicitly, one should be created implicitly including entries enabling all specified <typeshed-root>/stdlib files for all supported Python versions.

We could use an error= configuration option in the tag string to make an embedded file cause an I/O error on read.

The inline comment diagnostic assertions are useful for making quick, readable assertions about diagnostics in a particular location. But sometimes we will want to assert on the full diagnostic output of checking an embedded Python file. Or sometimes (see “incremental tests” below) we will want to assert on diagnostics in a file, without impacting the contents of that file by changing a comment in it. In these cases, a Python code block in a test could be followed by a fenced code block with language output; this would contain the full diagnostic output for the preceding test file:

# full output

```py
x = 1
reveal_type(x)
```

This is just an example, not a proposal that red-knot would ever actually output diagnostics in
precisely this format:

```output
test.py, line 1, col 1: revealed type is 'Literal[1]'
```

We will want to build tooling to automatically capture and update these “full diagnostic output” blocks, when tests are run in an update-output mode (probably specified by an environment variable.)

By default, an output block will specify diagnostic output for the file <workspace-root>/test.py. An output block can have a path= option, to explicitly specify the Python file for which it asserts diagnostic output, and a stage= option, to specify which stage of an incremental test it specifies diagnostic output at. (See “incremental tests” below.)

It is an error for an output block to exist, if there is no py or python block in the same test for the same file path.

Some tests should validate incremental checking, by initially creating some files, checking them, and then modifying/adding/deleting files and checking again.

We should add the capability to create an incremental test by using the stage= option on some fenced code blocks in the test:

# Incremental

## modify a file

Initial version of `test.py` and `b.py`:

```py
from b import x
reveal_type(x)
```

```py path=b.py
x = 1
```

Initial expected output for `test.py`:

```output
/src/test.py, line 1, col 1: revealed type is 'Literal[1]'
```

Now in our first incremental stage, modify the contents of `b.py`:

```py path=b.py stage=1
# b.py
x = 2
```

And this is our updated expected output for `test.py` at stage 1:

```output stage=1
/src/test.py, line 1, col 1: revealed type is 'Literal[2]'
```

(One reason to use full-diagnostic-output blocks in this test is that updating
inline-comment diagnostic assertions for `test.py` would require specifying new
contents for `test.py` in stage 1, which we don't want to do in this test.)

It will be possible to provide any number of stages in an incremental test. If a stage re-specifies a filename that was specified in a previous stage (or the initial stage), that file is modified. A new filename appearing for the first time in a new stage will create a new file. To delete a previously created file, specify that file with the tag delete in its tag string (in this case, it is an error to provide non-empty contents). Any previously-created files that are not re-specified in a later stage continue to exist with their previously-specified contents, and are not "touched".

All stages should be run in order, incrementally, and then the final state should also be re-checked cold, to validate equivalence of cold and incremental check results.