pytest-strategies 🧪

Powerful, constrained-randomized test generation for pytest.

pytest-strategies extends pytest with a robust framework for defining test strategies that combine random generation, directed edge cases, and constraints. It bridges the gap between simple parametrization and property-based testing, giving you full control over your test data.

---

🚀 Key Features

• Hybrid Generation: mix randomly generated rows with directed (hardcoded) edge cases and test vectors that run on request.
• Stable Test IDs: rows are named, not spelled out by their values: test_write[directed-zeros], test_write[rand-3]. The IDs are the same for every seed, so -k zeros, --deselect, CI history and node-ID reruns keep working.
• Reproducible Failures: a failed run prints the seed and the command that reruns each failed row alone (--strategy-details adds each row's values under its traceback), and --lf reruns the failed rows with the values they failed with.
• Row-Stable Random Streams: a row's values depend on the seed, the test and the row, not on --nsamples, the other rows, the collection order or pytest-xdist.
• Sequence Testing: use Series for deterministic, ordered value sequences (Cartesian product) or RNGSequence for a randomized permutation of the same values.
• Type-Safe RNG: built-in generators for Integers, Floats, Booleans, Strings, Choices, Sequences, and Enums.
• Weighted Probabilities: define custom distributions for Enums, Integers, and Floats.
• Named Constraints: filter values with predicates and rows with named constraints that read the row by argument name (v.addr); turn one off for a run with --strategy-constraint-off.
• Testbench Context: build vectors from configuration that is only known when the session runs, with a context per folder, the same object in your fixtures, and a fingerprint that tells two runs' configurations apart.
• Fixture Integration: works out-of-the-box with standard pytest fixtures (custom, parametrized, or built-in), and a test can take the whole row as one dataclass.
• CLI Control: filter vectors, change generation modes, or increase sample sizes directly from the command line.
• Per-Strategy Sample Count: let a strategy declare its own vector count as a soft default that an integer --nsamples can still override.
• Reusable Strategies: build a strategy on another with Parameter.extend(): replace or add arguments, and add, replace or remove vectors and constraints by name.
• Folder-Scoped Names: strategy names are looked up like fixtures, nearest folder first, and strategy files are imported only for the folders being tested.
• Metadata: item.stash[VECTOR_KEY] describes the row of each test, and export_strategies() writes every strategy as versioned JSON.
• Typed: the package ships type hints (py.typed), and the decorators keep the decorated function's type.
• Agent Skill: pytest-strategies skill install gives Claude Code, Codex and other coding agents a guide to writing strategies.

📦 Installation

pytest-strategies needs Python 3.11 or later and pytest 8.4.2 or later. Install it from PyPI:

pip install pytest-strategies

or from a local clone with pip install -e ..

Upgrading from 3.x? Test IDs and random values change, and the APIs that 3.0 deprecated are gone: see Upgrading to 4.0.

Agent skill

The package ships a skill that teaches coding agents (Claude Code, Codex and others that read SKILL.md files) how to write strategies and tests with this plugin. Install it into your project:

pytest-strategies skill install            # .claude/skills/ and .agents/skills/
pytest-strategies skill install --claude   # only .claude/skills/
pytest-strategies skill install --agents   # only .agents/skills/ (--generic works too)
pytest-strategies skill install --global   # in your home folder instead

python -m pytest_strategy skill install does the same. A global Claude install honors CLAUDE_CONFIG_DIR. Run the command again after upgrading the package to update the skill.

⚡ Quick Start

Define a strategy and apply it to your test:

from pytest_strategy import Parameter, RNGChoice, RNGInteger, TestArg, register, strategy

# 1. Register a strategy
@register("user_age_strategy")
def create_user_strategy() -> Parameter:
    return Parameter(
        # Randomly generate ages between 0 and 100
        TestArg("age", rng_type=RNGInteger(0, 100)),
        # Randomly choose a user type
        TestArg("user_type", rng_type=RNGChoice(["admin", "user", "guest"])),
        # Always include these specific edge cases, by position or by name
        directed_vectors={
            "newborn": (0, "guest"),
            "centenarian": (100, "user"),
            "admin_edge": {"age": 18, "user_type": "admin"},
        },
    )

# 2. Use the strategy in your test
@strategy("user_age_strategy")
def test_user_validation(age, user_type):
    assert 0 <= age <= 100
    assert user_type in ["admin", "user", "guest"]

Run it:

pytest test_users.py -v

pytest-strategies: RNG seed = 1763926297314361000
...
test_users.py::test_user_validation[directed-newborn] PASSED
test_users.py::test_user_validation[directed-centenarian] PASSED
test_users.py::test_user_validation[directed-admin_edge] PASSED
test_users.py::test_user_validation[rand-0] PASSED
...
test_users.py::test_user_validation[rand-9] PASSED

The directed vectors run first, then 10 random rows (--nsamples=50 asks for 50). Each row has a name, so pytest test_users.py -k newborn runs one edge case, and pytest "test_users.py::test_user_validation[rand-3]" --rng-seed=1763926297314361000 runs one random row again with the values it had (see Test IDs and Selecting Rows).

Strategies can be registered in the test module that uses them, as here, or in separate strategy files that the plugin imports for you (see Strategy Files and Scoped Names). A test can also take the factory itself, @strategy(create_user_strategy), which needs no name lookup.

register and strategy are also available as Strategy.register and Strategy.strategy, the 2.x spelling, which keeps working.

📖 Core Concepts

1. Strategies and Factories

A strategy is a factory function that returns a Parameter object. The Parameter defines the shape of your test data using TestArg definitions.

@register("math_ops")
def math_strategy():
    return Parameter(
        TestArg("x", rng_type=RNGInteger(0, 100)),
        TestArg("y", rng_type=RNGInteger(1, 100)), # Avoid 0 for division
        directed_vectors={
            "identity": (1, 1),
            "large_nums": (100, 100)
        }
    )

The factory is called once for each test that uses the strategy, when pytest collects that test. Like a test with fixtures, it receives by name the inputs its parameters ask for, in any order, and nothing else:

• nsamples: the run's sample count, an integer (the --nsamples value, or 10 when the option is not given), or the string "auto" under --nsamples=auto. A factory that returns a Parameter does not need it, because the plugin generates the vectors itself.
• ctx: the testbench context of the test's folder (see Configuration-Dependent Strategies).
• rng: the plugin's random.Random (RNG.generator()), for draws the RNG types do not cover. Its draws follow --rng-seed (see Reproducibility).
• options: a frozen StrategyOptions, exported from pytest_strategy, with the run's options for this strategy:

Field	Value
strategy	The name the strategy was resolved under, as in the -v summary
nsamples	The --nsamples value, "auto", or 10 without the option
nsamples_source	"--nsamples" or "default"
mode	The --vector-mode value ("all" by default)
vector_name	The --vector-name value, or None
vector_index	The --vector-index value, or None. options.filtered is true when either is set
constraints_off	The constraint names --strategy-constraint-off turns off for this strategy, a frozenset

import random

from pytest_strategy import Parameter, RNGChoice, RNGInteger, StrategyOptions, TestArg, register

from vectors import load_vectors  # your project's code


@register("alu")
def alu(rng: random.Random, options: StrategyOptions) -> Parameter:
    # The factory's own draws follow --rng-seed too
    ops = rng.sample(["add", "sub", "mul", "div", "shl", "shr"], k=3)
    # Reading the golden vectors is only worth it when they run
    golden = {} if options.mode == "random_only" else load_vectors("alu.yaml")
    return Parameter(
        TestArg("op", rng_type=RNGChoice(ops)),
        TestArg("a", rng_type=RNGInteger(0, 255)),
        directed_vectors=golden,
    )

Any other parameter needs a default, which it keeps; *args and **kwargs receive nothing. A parameter without a default that is not one of these names fails collection with a message naming it, and suggests the closest input (def factory(n) no longer receives nsamples by position). base, config and request are reserved names, an error even with a default. A factory decorated without functools.wraps is called with no arguments, and mock.patch mocks must be the first parameters. Later 4.x releases add inputs only as new StrategyOptions fields or under the reserved names, so a factory written for 4.0 keeps receiving exactly what it receives now.

A factory must return a Parameter. Anything else fails the collection of the tests that use the strategy, with a message that names the strategy and what it returned.

Strategy names are scoped by folder, like fixtures in conftest.py files (see Strategy Files and Scoped Names). A name registered twice in the same folder is an error.

PytestStrategiesWarning is a UserWarning subclass that you can import from pytest_strategy. The plugin emits it for problems that do not stop a run, such as a Series combination skipped because the constraints reject it. To turn it into an error, add error::pytest_strategy.PytestStrategiesWarning to your filterwarnings setting.

2. RNG Types

pytest-strategies provides rich, type-safe generators:

Type	Description	Example
RNGInteger	Integers in range	RNGInteger(0, 10)
RNGFloat	Floats in range	RNGFloat(0.0, 1.0)
RNGBoolean	Booleans with probability	RNGBoolean(true_probability=0.8)
RNGString	Random strings	RNGString(min_length=5, max_length=10)
Series	Ordered, deterministic sequence	Series([1, 2, 3])
RNGSequence	Randomized sequence (permutation)	RNGSequence([1, 2, 3])
RNGChoice	Choice from list	RNGChoice(["a", "b", "c"])
RNGEnum	Python Enum members	RNGEnum(MyEnum)
RNGWeighted*	Weighted ranges	RNGWeightedInteger({(0,10): 0.9, (11,100): 0.1})

RNG types check their arguments when they are constructed. A misconfigured strategy fails at collection with an RNGValueError (from pytest_strategy.rng) instead of drawing wrong values later. These cases are rejected:

• RNGInteger or RNGFloat with min > max. This includes a single bound that crosses the other bound's default: RNGFloat(min=5.0) fails because max defaults to 1.0.
• RNGFloat or RNGWeightedFloat bounds that are infinite or NaN.
• Weights of RNGWeightedInteger, RNGWeightedFloat or RNGEnum that are empty, negative, not finite or all zero, or whose total overflows. Some zero weights are allowed.
• A range of RNGWeightedInteger or RNGWeightedFloat that is not a (min, max) tuple with min <= max.
• RNGEnum given something that is not an Enum class, an Enum with no members, or a predicate that no member satisfies.
• RNGString with an empty charset (unless the length is 0), a negative length, or min_length > max_length.
• A set or frozenset passed to Series or RNGSequence. Their iteration order is not reproducible, so pass sorted(...) or a list instead.
• An empty Series or RNGSequence, or one whose predicate rejects every value, unless it has skip_if_empty (see Skipping when a sequence is empty).

TestArg's rng_type must be an RNG type (or an object with a generate() method): a bare lambda, or a class instead of an instance (RNGBoolean for RNGBoolean()), raises TypeError. The options after rng_type are keyword-only: TestArg(name, rng_type=None, *, value=None, validator=None, description=""). A TestArg with value= passes that value in every random row.

Parameter raises ValueError when nsamples is not None, "auto" or an integer >= 0, when max_retries is not an integer >= 1, when max_exhaustive is not None or an integer >= 1, when per_sequence_samples is not a bool, when ids is not None, "names", "values" or a callable, or when two of its arguments have the same name. An argument name must be a valid Python identifier that is not a keyword and does not start with _, because it names a field of the row (see Rows and Constraints).

Custom RNG types. Subclass RNGType and implement generate() and the python_type property. generate() is called once per draw, while RNG.generator() is that argument's own random stream for the row, so draw there, from RNG.generator() or the RNG.* helpers, and keep no state between calls:

from pytest_strategy import RNG, RNGType


class RNGAligned(RNGType[int]):
    """An address below `top`, aligned to `align` bytes."""

    def __init__(self, top: int, align: int) -> None:
        self.top = top
        self.align = align

    def generate(self) -> int:
        # Drawn here, from this argument's stream for this row
        return RNG.integer(0, self.top // self.align - 1) * self.align

    @property
    def python_type(self) -> type[int]:
        return int

A type that draws from a generator kept from earlier (such as the factory's rng), or that counts its calls (a walking-ones pattern), makes a row's values depend on the rows drawn before it; see Reproducibility.

3. Enums & Weighted Generation

The RNGEnum class supports standard Python Enums, including weighted selection and predicates.

from enum import Enum
class Status(Enum):
    SUCCESS = "success"
    FAILED = "failed"
    PENDING = "pending"

# Weights: 70% SUCCESS, 10% FAILED, 20% PENDING.
# The predicate excludes PENDING, so SUCCESS and FAILED are drawn 7:1.
arg = TestArg("status", rng_type=RNGEnum(
    Status,
    weights={Status.SUCCESS: 0.7, Status.FAILED: 0.1, Status.PENDING: 0.2},
    predicate=lambda s: s != Status.PENDING
))

With a predicate, each draw picks among the members that the predicate accepts, and they keep their relative weights.

4. Rows and Constraints

Every row a Parameter generates is a Vector: a namedtuple, exported from pytest_strategy, whose fields are the strategy's argument names in declaration order. v.addr and v[0] both work, so do unpacking and len(v), and a Vector compares and hashes like the plain tuple of its values (v == (0, 1)), although type(v) is tuple is false. A misspelled field raises AttributeError: Vector has no argument 'lenght'; its arguments are addr, len. Fields named count or index hide the tuple methods of the same name: use tuple.index(v, x) then. Parameter.vector_type returns the strategy's Vector class, and generate_vector(), generate_vectors(), generate_exhaustive() and the get_*_vector() methods return Vectors. Tests receive the arguments one by one (or as a record, see Record Parameters), never the Vector itself.

You can enforce rules on generated data:

Per-Argument Predicates:

# Only even numbers
RNGInteger(0, 100, predicate=lambda x: x % 2 == 0)

Cross-Argument Constraints:

Parameter(
    TestArg("min", rng_type=RNGInteger(0, 10)),
    TestArg("max", rng_type=RNGInteger(0, 10)),
    vector_constraints={
        "ordered": lambda v: v.min < v.max,  # Ensure min < max
    }
)

A constraint receives the row as a Vector and returns whether to keep it. Constraints are named: by the keys of a dict, or, in a list, by each function's name (constraint_<i> for a lambda, a functools.partial or a callable object, i being its position). A name is a non-empty string without whitespace, :, , or =. Two constraints with one name (typically two closures made by one helper) fail when the Parameter is built: pass a dict to name them. param.vector_constraints is a read-only mapping of names to functions, in the order they run; change it with add_constraint(fn, *, name=None), which returns the name, remove_constraint(name) and clear_constraints().

Constraints run in order, and the first falsy result rejects the row, so a constraint may rely on the ones before it. A constraint should only read the row: one that draws random values gets a warning (see Reproducibility). A constraint that raises fails collection with an error that names it and the row, followed by its own traceback. Directed and test vectors are never checked by constraints.

A random row that fails a constraint is drawn again, up to max_retries times (a Parameter argument, default 100). If no valid row turns up, collecting the test fails, and the message counts the rejected draws by the name of the first constraint that rejected each one, with the first row each constraint rejected, so you can tell which one is too strict:

Could not generate random row 3 after max_retries=100 draws. Rejected by (first failing constraint per draw): aligned=97, no_4k_cross=3. First rows rejected: aligned: Vector(addr=4097, len=16); no_4k_cross: Vector(addr=4090, len=16). Raise Parameter(max_retries=...), relax a constraint, or turn one off for this run with --strategy-constraint-off=dma_burst:aligned.

With -v, the Strategy Summary lists the rejections per constraint for each strategy.

Turning a constraint off for one run. Name it with --strategy-constraint-off: --strategy-constraint-off=ordered turns off every constraint named ordered, and --strategy-constraint-off=my_strategy:ordered only the one in that strategy (the name the strategy was resolved under, as in the -v summary). Items are separated by commas and the option can be repeated, so --strategy-constraint-off=my_strategy:ordered,small turns ordered off in my_strategy and small in every strategy. The Parameter keeps its constraints, and the rows that the constraint never rejected keep their values. The report header lists what is off, and -v adds off: ordered to the strategy's summary line. A factory sees the names in options.constraints_off.

An item that matches no constraint of any strategy the run resolved stops a run of the whole suite with a usage error that suggests the closest name and lists the constraints by strategy. A run narrowed to some tests (paths or node IDs, --lf, --sw, --ignore, or no path from a folder below the rootdir, where pytest collects only that folder) resolves only some strategies, and so does a run in which a test module is skipped or fails to collect: it prints that message in red and goes on, so the option can be kept in addopts.

5. Directed and Test Vectors

Directed vectors are the fixed rows a strategy always tests: edge cases, known bugs. Test vectors are scenarios that run only with --vector-mode=test, to verify them on their own:

from pytest_strategy import Parameter, RNGInteger, TestArg, register

@register("api_test")
def api_test_strategy():
    return Parameter(
        TestArg("status_code", rng_type=RNGInteger(200, 500)),
        directed_vectors={"ok": (200,)},
        # Test vectors: specific scenarios to verify
        test_vectors={
            "success": (200,),
            "not_found": (404,),
            "server_error": (500,)
        }
    )

--vector-mode chooses which rows run:

Mode	Rows
all (the default)	The directed vectors, then the random rows
random_only	The random rows
directed_only	The directed vectors
mixed	The random rows, and the directed vectors before them when Parameter(always_include_directed=True) (the default)
test	The test vectors

--vector-name=NAME and --vector-index=I run a single directed vector, by its name or its position in directed_vectors (from 0), and take precedence over --vector-mode and --nsamples. A strategy without the requested directed vector yields no vectors, so the tests that use it are skipped ("got empty parameter set"). If no strategy in the run has the vector, for example because of a typo or an index that is out of range everywhere, pytest stops with a usage error that lists each strategy's directed vectors. It does not skip every test.

A directed or test vector holds one value per argument:

• a tuple or a list (for example read from a YAML file), in argument order: (404,) or [404];
• a dict of argument names to values, in any order: {"status_code": 404}. Its keys must be the argument names, and a missing or unknown key fails with a message that names it (Directed vector 'zeros' has unknown argument 'lenght' (did you mean 'len'?)). Tuple and dict vectors can be mixed in one Parameter;
• a namedtuple, placed by its field names, which must be the argument names;
• pytest.param(..., marks=...), to mark a row, with the values spread or one dict: pytest.param(0, 1, marks=pytest.mark.xfail) or pytest.param({"addr": 0, "len": 1}, marks=...). A tuple, list or namedtuple inside it is one value, so pytest.param((404,), marks=...) gives a one-argument strategy the value (404,); write pytest.param(404, marks=...). Its values are checked like the others. It cannot have an id=, because the vector's name is its ID.

A bare value such as 404 or "a" is an error, with a hint: for a one-argument strategy write (404,) or {"status_code": 404}. A dict is always a named vector, so a dict value for a one-argument strategy is written ({"a": 1},) or {"cfg": {"a": 1}}, and in a pytest.param only pytest.param({"cfg": {"a": 1}}, marks=...). A dataclass or pydantic instance as a vector is not supported yet; use a dict. Vector names are non-empty strings; identifier-like names are easiest to select with -k (see Test IDs and Selecting Rows).

directed_vectors and test_vectors are read-only mappings of names to Vectors (a pytest.param vector stays a pytest.param whose values are a Vector): change them with add_directed_vector(), remove_directed_vector() and the test equivalents.

6. Sequence Testing & Exhaustive Generation

There are two sequence types. Both walk a fixed set of values, but they differ in ordering:

Type	--nsamples=auto	--nsamples=K (finite)
Series	Values in declaration order (Cartesian product)	Cycles through the product in order (K >= len) or takes the first K (K < len), skipping combinations that the constraints reject
RNGSequence	A random permutation (each value once)	Random picks, like RNGChoice

Rule of thumb: the RNG prefix means random. Series is deterministic and ordered; RNGSequence is randomized.

Deterministic Strategy (Series):

from pytest_strategy import Series

@register("matrix_test")
def matrix_strategy():
    return Parameter(
        TestArg("x", rng_type=Series([1, 2, 3])),
        TestArg("y", rng_type=Series(["a", "b"]))
    )

Running with pytest --nsamples=auto generates 6 tests in order: (1, 'a'), (1, 'b'), (2, 'a'), (2, 'b'), (3, 'a'), (3, 'b'), named x=1-y=a, x=1-y=b and so on. The leftmost arg is the slowest counter.

Randomized Strategy (RNGSequence):

from pytest_strategy import RNGSequence

TestArg("x", rng_type=RNGSequence([1, 2, 3]))

Under --nsamples=auto, this yields a permutation of [1, 2, 3] (each value exactly once, random order). Under a finite --nsamples=K, it picks K random values.

Mixed Mode (Sequence + Random): If you mix a sequence type with random types (e.g., RNGInteger), the random values are regenerated for each sequence combination.

@register("mixed_test")
def mixed_strategy():
    return Parameter(
        # Deterministic: iterate through all user roles in order
        TestArg("role", rng_type=Series(["admin", "user", "guest"])),
        # Random: generate a fresh random ID for each role
        TestArg("id", rng_type=RNGInteger(1, 1000))
    )

Running pytest --nsamples=auto generates 3 tests (one per role, role=admin, role=user and role=guest), each with a random ID.

Filtering with Predicates: Both sequence types accept a predicate argument to exclude values before generation.

# Iterate only even numbers from 0-9, in order
TestArg("evens", rng_type=Series(range(10), predicate=lambda x: x % 2 == 0))

What --nsamples=auto does: auto replaces a strategy's random samples with the combinations of its sequence arguments. Everything else works as it does with a finite count:

• Only Series and RNGSequence arguments are enumerated. Every other argument (RNGInteger, RNGEnum, RNGChoice, RNGBoolean, ...) gets a fresh random value for each combination.
• Combinations that vector_constraints reject are left out. Random arguments are redrawn up to max_retries times before a combination is dropped. If the constraints reject every combination, collecting the test fails.
• Directed vectors are placed before the combinations in the default all mode. In mixed mode they are included when always_include_directed is set (the default). random_only gives only the combinations.
• --vector-mode=test, --vector-mode=directed_only, --vector-name and --vector-index take precedence: they select only those vectors, and no combinations are generated.
• A strategy with no Series or RNGSequence argument has nothing to enumerate. It falls back to its own nsamples (see Per-Strategy Sample Count), or 10 random samples.
• The number of combinations is checked before any is generated. More than 100,000 rows fails collection with a message that shows the product (a=1,000 x b=1,000). Raise the limit for one strategy with Parameter(max_exhaustive=...), or for the project with the strategies_max_exhaustive ini option. The same limit applies to per_sequence_samples=True, where each combination counts K rows.

Constraints on Series in finite mode: With a finite --nsamples, a Series combination that the constraints reject is skipped, and the cycle continues with the next combination. If the strategy also has random arguments, they are redrawn up to max_retries times before the combination is skipped. Each such skip emits a PytestStrategiesWarning: raise max_retries, or relax the constraint if that combination should be tested. If a whole cycle of combinations yields no valid vector, collecting the test fails.

n samples per sequence combination (per_sequence_samples): By default a finite --nsamples=K gives K rows in total, shared among the sequence values. Pass per_sequence_samples=True to get K random rows for each combination of the Series/RNGSequence arguments instead:

@register("per_device")
def per_device():
    return Parameter(
        TestArg("device", rng_type=Series(["devA", "devB"])),
        TestArg("width", rng_type=RNGInteger(1, 64)),
        per_sequence_samples=True,
    )

With the default of 10 samples this runs 20 tests: 10 for devA (device=devA-rand-0 to device=devA-rand-9), then 10 for devB, each with a fresh random width. Several sequence arguments multiply (2 devices × 3 modes × K), and a value listed twice counts twice. Combinations follow declaration order for RNGSequence too, and directed vectors are still placed first. Under --nsamples=auto the flag has no effect: every combination runs once. A combination whose random arguments the constraints reject max_retries times in a row gets fewer rows and a PytestStrategiesWarning; collecting the test fails only if no combination yields a row. The warning names the strategy and the test.

Skipping when a sequence is empty

When the values come from configuration, there may be none: a testbench without any Esm peripheral, for example. An empty Series or RNGSequence normally fails collection. Give it a skip_if_empty reason instead (a keyword argument), and the strategy contributes a single skipped row with that reason, so each test that uses it is skipped:

from pytest_strategy import Parameter, RNGInteger, RNGSequence, TestArg, register, strategy

# From your testbench configuration, e.g.
# [p.channel for p in config.peripherals.values() if p.type == "Esm"]
ESM_CHANNELS = []

@register("esm_rw")
def esm_rw():
    return Parameter(
        TestArg("channel", rng_type=RNGSequence(ESM_CHANNELS, skip_if_empty="no Esm peripheral in this testbench config")),
        TestArg("wdata", rng_type=RNGInteger(min=0, max=255)),
        per_sequence_samples=True,
    )

@strategy("esm_rw")
def test_rw(channel, wdata):
    ...

With channels, the option changes nothing. Without any, pytest -rs shows SKIPPED [1] test_esm.py:<line>: no Esm peripheral in this testbench config, and the test's ID is test_rw[skipped]. A test with other parametrization (a stacked @pytest.mark.parametrize, a parametrized fixture) is skipped once per combination of it. The skip applies in every --vector-mode and with --nsamples=auto, and directed and test vectors are skipped too. A --vector-name or --vector-index that names one of the strategy's directed vectors also gives the skipped test. The reason must be a non-empty string. It also applies when the predicate rejects every value. The test's signature (or record parameter) is still checked against the strategy, so a mismatch fails collection even on a configuration without values. skip_if_empty works for strategies that return a Parameter.

7. Per-Strategy Sample Count

By default the number of generated vectors is controlled globally by --nsamples (10 when unset). A strategy can declare its own count by passing nsamples to its Parameter:

@register("edge_heavy")
def edge_heavy_strategy():
    return Parameter(
        TestArg("x", rng_type=RNGInteger(0, 100)),
        nsamples=25,  # this strategy generates 25 vectors by default
    )

This is a soft default: an explicit integer --nsamples on the command line still wins, so you can always scale a whole run from the CLI.

--nsamples (CLI)	Parameter(nsamples=...)	Random vectors generated
not passed	25	25 (strategy value)
not passed	unset	10 (global default)
--nsamples=5	25	5 (CLI overrides)
--nsamples=auto	any	Every Series/RNGSequence combination. A strategy without such arguments uses its own value (here 25), or 10 when unset
not passed	"auto"	Same as --nsamples=auto

Directed vectors are added on top of these, according to --vector-mode. With Parameter(per_sequence_samples=True), a finite count applies to each Series/RNGSequence combination instead of the whole strategy (see per_sequence_samples). A row keeps its values when the count grows: --nsamples=50 gives the rows of --nsamples=10, with the same values, and 40 more.

8. Record Parameters

Instead of one test parameter per strategy argument, a test can take a single record, a dataclass whose fields are the strategy's arguments:

from dataclasses import dataclass

import pytest

from pytest_strategy import Parameter, RNGInteger, TestArg, register, strategy


@dataclass
class Point:
    x: int
    y: int


@register("points")
def points_strategy():
    return Parameter(
        TestArg("x", rng_type=RNGInteger(0, 10)),
        TestArg("y", rng_type=RNGInteger(0, 10)),
    )


@pytest.fixture
def scale():
    return 2


class TestPoints:
    @strategy("points")
    def test_scaled(self, point: Point, scale):
        assert 0 <= point.x * scale <= 20

The rule: a test receives the row as one record when (1) neither the test nor any fixture it uses asks for one of the strategy's argument names, and (2) exactly one test parameter is annotated with a record type whose fields are exactly those names. Otherwise the strategy passes its arguments by name, one test parameter each.

• "Asks for" covers the test's parameters, usefixtures, autouse fixtures and what those fixtures ask for in turn. So a fixture that takes x and y and builds the object itself receives the arguments, with or without validate_signature, which does not change the choice.
• A test parameter is one that pytest fills: no default, not *args or **kwargs, and not self, cls or a built-in fixture. The other parameters are left alone, in any position.
• The record types are dataclasses, pydantic dataclasses included, once Annotated[...] and generic arguments are stripped (Pair[int] is Pair). Their fields are the init=True fields: fields with init=False are not counted (nor shown in test IDs), and kw_only fields work. A strategy with one argument can use a record too.
• NamedTuple, TypedDict and pydantic models are recognized but not supported yet: a test whose record would be one fails collection with "not supported yet". Their fields are _fields, the required and optional keys, and the model_fields names (not aliases). Because they are recognized now, supporting them later cannot change which parameter an existing test uses.
• Optional[...] and other unions, pydantic v1 models, attrs classes, Vector and annotations that cannot be resolved are not record types. String annotations (from __future__ import annotations or quoted names) work, as long as the class is defined at module level.
• Two parameters that both match fail collection, naming both. If a single parameter has a record type but its fields do not match, collection fails with a message that lists the missing and extra fields.

9. Strategy Files and Scoped Names

Strategies do not have to be registered in the test module. A strategy file is a file named strategies.py, strategy.py, *_strategies.py or *_strategy.py that registers a strategy with a decorator: @register("name"), @Strategy.register(...), or an alias such as @S.register("name"). The file's text is checked before it is imported, so write the name as a string literal in the @register(...) forms. A plain call such as register("x")(fn), or a name held in a variable (@register(NAME)), is not recognized, and a file that only registers that way is not imported.

Loading. Before pytest imports a test module, the plugin imports the strategy files in the module's folder and in each folder above it, up to the rootdir (or the testpaths entry or command-line directory that contains the module). Each file is imported once per session, closest folder first and in name order within a folder. Strategy files in folders that no collected test module is in or below are not imported, unless a test asks for a name that is not registered on its path (see below) or you run pytest --list-strategies.

Names are scoped by folder. A strategy name is looked up like a fixture in a conftest.py file: in the test's folder first, then in each folder above it, and the nearest registration wins. A strategy registered in a test module or a conftest.py belongs to that file's folder. So with this layout, the tests under tests/esm/ get the default strategy of tests/esm/strategies.py, and every other test gets the one of tests/strategies.py:

tests/
├── strategies.py          # @register("default")
├── test_dma.py            # @strategy("default") -> tests/strategies.py
└── esm/
    ├── strategies.py      # @register("default")
    └── test_esm.py        # @strategy("default") -> tests/esm/strategies.py

When no folder on the test's path registers the name, the plugin imports every strategy file it can find and uses a registration from another place if there is exactly one: a sibling folder, or a module installed outside the rootdir (a registration outside the rootdir wins over the ones inside it). This keeps layouts like a shared tests/strategies/ folder working. If several folders register the name, collecting the test fails with an error that lists them; register the strategy in a folder above the test, or pass the factory itself.

Passing the factory. @strategy(my_factory) uses that function, with no name lookup, so it cannot pick up another folder's strategy. The factory does not need to be registered. The test's random streams are keyed by a name that the factory's own file, the module that defines it, registers it under when it is imported, the first in alphabetical order when it has several: the test gets the same vectors and IDs as @strategy("name") where that name finds this factory. Such a register call is at the module level of that file, or in a helper function that its import calls. A factory that its own file does not register this way is keyed by its qualified name (my_factory). A registration made by any other file does not name it: an alias in another folder's strategy file or conftest.py, in a test module or in a plugin, though @strategy("alias") finds it. Neither does a call made in a hook such as pytest_configure or in a fixture, not even in the factory's own file, nor another factory registered under the same name nearer to the test. A functools.partial or a functools.wraps wrapper counts as the function it wraps, a bound method as its function, and an object as its class: one that its file does not register itself (reading obj.method again gives the same factory) gets the first name its file registers one of that function or class under. The test holds the factory, so the module that defines it has run, and with it that module's register calls: the name is the same in every run that collects the test, its printed rerun command and --lf included, whichever other files the run imported, and wherever the factory's file is (in another folder, or in a package, installed or checked out).

A name registered twice in one folder (in one file or in two) is an error: the plugin emits a PytestStrategiesWarning naming both functions, and pytest stops with a usage error after collection. A second registration made while pytest starts, by a plugin or by a conftest.py that pytest loads before it collects (such as the rootdir's), only warns. Running the same registration again (the same file executed twice) is silent. Decorated factories count as the function they decorate. Factories built by one shared function or class (closures, functools.partial objects, instances) count as that function or class, so registering two of them under one name is not reported.

Where files are searched. The search for all strategy files covers the testpaths directories from your pytest configuration, expanding glob patterns such as pkgs/*/tests, or the rootdir when testpaths is not set, and the directory of each path given on the command line. Below these directories it skips what pytest's collection skips: hidden directories (names starting with .), __pycache__, directories matching norecursedirs (by default these include build, dist, venv and node_modules; a pattern with a /, such as tests/data, is matched against the path), and virtual environments (a directory containing pyvenv.cfg, or a conda environment containing conda-meta/history). Like pytest, it follows symlinked directories.

Imports. A strategy file is imported the way pytest imports a test module in its folder, following --import-mode. In the default prepend mode its folder (or the root of its package) is put on sys.path, so it can import modules next to it. A test module or conftest.py that imports a strategy file (from strategies import Mode, for example to use an Enum it defines) gets the module the plugin loaded, so the file is not executed a second time and the tests compare against the same classes the strategy uses. When two folders without __init__.py both have a strategies.py, the second one is imported under a unique module name instead. A plain from strategies import Mode in that second folder still returns the first folder's module (Python caches modules by name), so its tests cannot import from it: give such files distinct names (esm_strategies.py), or add __init__.py files and use from .strategies import Mode.

Values that a strategy file draws when it is imported come from a random stream of their own, derived from the seed and the file's path relative to the rootdir, so --rng-seed reproduces them whatever else was collected first. A file outside the rootdir is keyed by its relative path too (../shared/strategies.py), so two checkouts in different folders get the same values. An RNG.seed() call in a strategy file changes only the rest of that file's draws.

Failures. A strategy file that fails to import does not stop the run. The plugin prints pytest-strategies: Warning - Failed to load <path>: <error>. A file that calls pytest.skip(..., allow_module_level=True) or pytest.importorskip() at module level is skipped, and that is reported with -v. A "Strategy 'name' not found" error names the test, suggests close names, and lists the files that failed to load or were skipped, and the files with a strategy file name that mention register but were not imported because they have no registration decorator. pytest -vv prints each loaded file, and pytest --list-strategies lists the registered strategy names (with the file of each registration when a name is registered in several folders) and exits. It collects the tests as usual first, which calls their factories (and the context hook, for those with ctx), so a factory that fails shows as a collection error and the listing exits with code 2; a factory that no collected test uses is listed without being called.

10. Configuration-Dependent Strategies

Some vectors depend on configuration that is only known when the session runs, such as a testbench description whose file is named on the command line. Strategy factories run before any fixture exists, so they cannot use one. Instead, implement the pytest_strategies_context hook in a conftest.py (or a plugin): what it returns is passed as ctx to every factory with a ctx parameter that a test in that conftest.py's folder, or below it, uses.

# conftest.py
import pytest

from testbench import Testbench  # your project's code

def pytest_addoption(parser):
    parser.addoption("--tb-config", default="testbench.yaml")

@pytest.hookimpl(optionalhook=True)
def pytest_strategies_context(config):
    return Testbench.parse_config(config.getoption("--tb-config"))

optionalhook=True keeps the conftest.py usable when pytest-strategies is not loaded (not installed, or -p no:pytest_strategy): without it, pytest stops with "unknown hook".

# esm_strategies.py
from pytest_strategy import Parameter, RNGInteger, RNGSequence, TestArg, register

@register("esm_rw")
def esm_rw(ctx):
    channels = [p.channel for p in ctx.peripherals.values() if p.type == "Esm"]
    return Parameter(
        TestArg("channel", rng_type=RNGSequence(channels, skip_if_empty="no Esm peripheral")),
        TestArg("wdata", rng_type=RNGInteger(min=0, max=255)),
        per_sequence_samples=True,
    )

Tests use the strategy as usual, with @strategy("esm_rw"). With two Esm channels in the configuration they run 10 writes per channel; with none they are skipped with the reason.

A context per folder

• Each test gets the context of its own folder, so one test tree can hold two testbench configurations. The plugin calls the implementations it can see from there in this order: tryfirst ones, then the conftest.py files from the test's folder upward, then the other plugins (last registered first), then trylast ones. The first that returns something other than None answers: the nearest conftest.py wins, one that returns None defers to the folder above, and a plugin answers only where no conftest.py does, whatever order they were registered in.
• A wrapper=True implementation runs around the one that answered and can change its answer by returning a new object ({**ctx, "extra": 1}); a child folder whose conftest.py returns None shares its parent's wrapped object. A wrapper must leave the object it receives as it is, because the folders that do not see the wrapper get that object too: when the object's fingerprint (below) changed while the wrappers ran, the tests of the folders that see the wrapper fail with an error that says so. A wrapper that returns the very object it received (one that only checks or logs it) gives its folders the context of the folders without it, with the same label. Each implementation runs once, for the first folder that asks, so a wrapper's code before its yield runs after the implementations it wraps.
• A factory registered in another folder gets the context of the test's folder, the way a fixture override follows the test that requests it.
• Each implementation that is not a wrapper is called at most once per session, the first time a factory with a ctx parameter (or strategies_ctx or get_context(), below) needs it, and its result is reused: tests in folders that end at the same implementation share one object. A wrapper runs once for each implementation that answers under it, and again for each other set of wrappers it is in. Factories without ctx are called as before and never trigger it.
• When no implementation returns a value, a ctx parameter keeps its default (or a value bound with functools.partial), and is None without one.
• If an implementation raises, each test module that uses a factory with ctx in a folder that consults it fails collection with Strategy factory '<name>' has a 'ctx' parameter, but the pytest_strategies_context hook raised <error>, and its other tests do not run. For a method of a test class, the class fails collection instead, and the module's other tests are still collected (pytest runs them with --continue-on-collection-errors; by default a collection error stops the run). pytest.fail() in the hook is reported as it is. The hook can also call pytest.skip(..., allow_module_level=True) to skip those modules or classes.
• When a factory fails with ctx set to None while a conftest.py in another folder implements the hook, the error adds which folders implement it, so that the hook can be moved to a common parent conftest.py.
• Random draws in the hook come from a stream of their own, derived from the seed and started anew for each implementation, so they are reproduced by --rng-seed, do not depend on which tests run, and do not change any test's vectors.
• Under pytest-xdist every worker calls the hook, so it must return the same configuration in each: a run whose workers built different contexts fails (see pytest-xdist below).
• export_strategies(), which has no test, gives a factory the context of the folder its file is in, as a test there gets it, or the rootdir's for a file outside the rootdir or in an installed package inside it (a site-packages or dist-packages folder below the rootdir). When pytest has not loaded a conftest.py of that folder, or of a folder between it and the rootdir, because the run collected no test there (pytest tests/a leaves tests/b/conftest.py out), that context cannot be known: a factory with ctx there is not called, and its entry reads {"unavailable": "tests/b/conftest.py was not loaded in this session"}.

The context in fixtures: strategies_ctx and get_context()

The hook only has to describe the configuration the vectors depend on. Your testbench fixture can build on the same object instead of parsing the configuration a second time:

# conftest.py
import pytest

from testbench import Testbench  # your project's code

@pytest.fixture(scope="session")
def tb(strategies_ctx):
    return Testbench(strategies_ctx)

# tests/tb_a/conftest.py: a folder with its own pytest_strategies_context
import pytest

from pytest_strategy import get_context
from testbench import Testbench  # your project's code

@pytest.fixture(scope="session")
def tb_a(request):
    return Testbench(get_context(request.config, __file__))

• strategies_ctx is a session fixture of the plugin. It returns the object the factories of the tests that use it received, and computes it when no factory needed it yet.
• A session fixture has one value, so the tests that use it must share one context. When they are in folders whose contexts come from different implementations, each of them fails with strategies_ctx is a session fixture, but the tests that use it have different contexts (conftest.py: ...; tests/tb_a/conftest.py: ...), which names each context's first test. Deselecting one folder's tests makes the run pass; a folder with its own implementation uses get_context() in its conftest.py instead. The tests that use it are those that request it, directly or through their fixtures, except the tests of a folder whose conftest.py defines its own strategies_ctx fixture, which does not request the plugin's. When none does and a test asks for it through request.getfixturevalue(), every test of the run counts. When some do, a test that asks for it that way gets their context, whichever test asks first, and fails with the message after its setup or its body when its own folder's context is another one. So does a test that gets a fixture whose setup asked for it that way, directly or through other fixtures, when the test gets the value (or the error) that setup cached, whichever test it ran for; a function-scoped fixture that asks only for some tests leaves the others alone. An error the test raised keeps its traceback, and the message follows it in a pytest-strategies section of that test's report only, also when other tests get the same error from a fixture's cache.
• pytest_strategy.get_context(config, path) returns the context of the folder of path (a file or a folder): the object a test there gets. A folder whose conftest.py pytest has not loaded, because no test there was collected, gets the context of the nearest loaded conftest.py above it. Call it from fixtures or hooks: in pytest_configure, or while conftest.py files are imported, it sees only those loaded so far. It raises RuntimeError for a config that no running session uses.
• Both raise what the implementation raised, as it is: a pytest.skip in the hook skips the tests that use them.

The context fingerprint

After the collection, the plugin prints a fingerprint of the contexts computed so far, so you can tell whether two runs (a CI job and your rerun) built the same configuration. Several contexts are named by the conftest.py (or plugin) that answered:

pytest-strategies: context 976bcfdf
pytest-strategies: contexts conftest.py 976bcfdf, tests/tb_a/conftest.py b1e1b237

• The fingerprint is the first 8 hex characters of the SHA-256 of a canonical encoding of the object. It is computed when the hook returns the object, before any factory or fixture receives it, so a factory or a test that changes the object later changes no fingerprint. The line lists the contexts computed by the end of the collection, for factories with ctx or by get_context() calls made by then; a context that strategies_ctx or a fixture computes first appears only in the -v Contexts block. Nothing is printed when no context was computed by then, or when every context is None; the line appears with -q and --collect-only too. Under pytest-xdist the controller collects no tests, so it prints the line the workers printed after their collection at the end of the run, in its terminal summary.
• The encoding depends only on what the object holds. Sets are sorted, so the hash order of PYTHONHASHSEED does not matter. Paths inside the rootdir are written relative to it, so two checkouts agree. A pydantic v2 model is written as its model_dump(), which leaves out Field(exclude=True) fields and keeps a SecretStr masked. Dataclasses, attrs classes and NamedTuples are written field by field, SimpleNamespace and argparse.Namespace objects by their attributes, mappings as their pairs in order, floats, dates, Decimal and UUID as text, and classes and Enum members by their qualified names, never their modules. Anything else is written as its repr, without memory addresses ( at 0x7f... inside a <...> repr, and a mock's id='140...'; an address a repr of its own shows, such as Periph('uart0' at 0x40001000), is kept), and with the items of the sets it shows as Python does ({'b', 'a'}) sorted. An object that keeps the default repr (<Plain object at 0x...>) is in the fingerprint by its type alone, and the line says so: context 976bcfdf (partial: Plain). A repr that shows a set in another form (",".join(tags)) should sort it. An object that cannot be encoded (its repr raises) gives unavailable, and never fails the run.
• Leave volatile values out of the context (temporary paths, process IDs, times), or mark them Field(exclude=True) in a pydantic model, so that the fingerprint stays the same from one run to the next.
• When tests fail, the line that says how to reproduce the run ends with the contexts their factories received: pytest-strategies: reproduce with --rng-seed=S (context 976bcfdf), or (contexts conftest.py 976bcfdf, tests/tb_a/conftest.py b1e1b237). Under pytest-xdist the workers send them to the controller. Tests whose setup or call failed count; an error in a test's teardown alone adds no context. With --strategy-details, the pytest-strategies section of each failed row shows the fingerprint of its factory's context in its context line (see Reproducing a failure).
• -v adds a "Contexts" block to the Strategy Summary, with each context's label, fingerprint and number of tests whose factories received it. item.stash[VECTOR_KEY].context holds the fingerprint for the rows of a factory that received ctx, and None for the others (see Per-Test Metadata).

11. Metadata Export

export_strategies() returns every registered strategy as a JSON document, for analysis or integration with other tools:

from pytest_strategy import export_strategies

# "json" is the only format; the argument is keyword-only
json_data = export_strategies(format="json")
print(json_data)

For a tests/dma/strategies.py that registers burst with two arguments, a directed vector and a named constraint, and a run with --rng-seed=1 (shortened):

{
  "schema": 1,
  "kind": "strategies",
  "generator": {"name": "pytest-strategies", "version": "4.1.2"},
  "seed": 1,
  "nsamples": 10,
  "strategies": [
    {
      "name": "burst",
      "origin": {"folder": "tests/dma", "file": "tests/dma/strategies.py", "qualname": "burst", "line": 4},
      "context": null,
      "parameter": {
        "schema": 1,
        "arguments": [
          {"name": "addr", "description": "", "python_type": "int", "validator": false,
           "source": "rng", "rng": {"type": "RNGInteger", "min": 0, "max": 4095, "predicate": false}},
          {"name": "len", "description": "", "python_type": "int", "validator": false,
           "source": "rng", "rng": {"type": "RNGInteger", "min": 1, "max": 64, "predicate": false}}
        ],
        "directed_vectors": [{"name": "zeros", "id": "directed-zeros", "values": {"addr": 0, "len": 1}}],
        "test_vectors": [],
        "constraints": [{"name": "no_4k_cross", "enabled": true}],
        "always_include_directed": true,
        "max_retries": 100,
        "nsamples": null,
        "per_sequence_samples": false,
        "max_exhaustive": null,
        "skip_reason": null
      }
    }
  ]
}

• seed and nsamples are the run's seed and --nsamples value (10 without the option).
• strategies holds one entry per registration, a name registered in several folders once per folder, sorted by name and folder. An entry has the strategy's name, its origin (folder and file, relative to the rootdir in posix form or absolute outside it, qualname and line), its context (the fingerprint of the context its factory received, or null) and one of: parameter, the factory's Parameter.to_dict(); error, such as {"type": "RuntimeError", "message": "boom"} for a factory that raised; or unavailable (see A context per folder).
• In Parameter.to_dict(), each argument has its source: "rng" with rng, its RNG type's to_dict() (a custom type lists its public attributes), or "value" with its value. Each constraint has enabled, false for one that --strategy-constraint-off turns off in this run.
• Values keep their type: JSON for None, bools, ints, strings and finite floats, {"$float": "nan"}, {"$enum": "Color", "member": "RED"}, or {"$repr": "b'\\x00'", "$type": "bytes"} for anything else, so the text has no NaN and a strict json.loads reads it. VectorInfo.to_dict() (see Per-Test Metadata) uses the same encoding.
• How the schema evolves. A reader should ignore the keys it does not know, and the values of kind and source it does not know, and read an unknown $-tagged object like $repr: 4.x releases may add them within schema 1. Removing, renaming or retyping a key changes the schema number.

In a pytest session every strategy file is imported first. Each factory gets the inputs it declares, as at collection, with the session's options: nsamples is the --nsamples value, "auto", or 10 without the option, and ctx the context of the factory's own folder (see A context per folder).

Its random draws come from a stream derived from the seed, the strategy name and the factory's module name, or the folder of its file for a factory in a strategy file, a test module or a conftest.py, so every call in a session, and in every environment and checkout, exports the same values. A file counts as a strategy file, a test module or a conftest.py when it has such a name and is inside the rootdir, below a testpaths entry, or imported by its path in the session (a test module collected from a folder named on the command line) and, when consider_namespace_packages is off (pytest's default), not a module of a regular package that Python holds under its package name (acme.strategies for acme/strategies.py next to acme/__init__.py), the name pytest then imports it under in every import mode. A module named like one elsewhere, such as a library's extacme/strategies.py on sys.path, is keyed by its module's name, so every checkout folder exports the same values.

A factory in a package module named like one inside the rootdir (src/acme/strategies.py, src/acme/test_utils.py) is therefore keyed by its folder in a checkout or an editable install and by its module's name when installed, so the two export different values; rename the module (acme/catalog.py) to avoid that. Outside the rootdir and the testpaths, these are keyed by their folder in a run that collects that folder (one that names it on the command line) and by their module's name in a run that does not: (a) a helper named like one, not in a regular package, imported by its module's name; (b) a regular package's module that Python holds only under a longer namespace-package name (ns.acme.strategies); (c) with consider_namespace_packages = true, also a regular package's module imported by its name. List such a folder in testpaths to key them by their folder in every run.

12. Reusing and Extending Strategies

A strategy is reused by its name from any folder below the one that registers it, or by passing its factory (see Strategy Files and Scoped Names). To build a new strategy on an existing one, call its factory and extend() the Parameter it returns. extend() returns a new Parameter; the one it is called on does not change.

# tests/strategies.py
from pytest_strategy import Parameter, RNGChoice, RNGInteger, TestArg, register

@register("dma_burst")
def dma_burst():
    return Parameter(
        TestArg("addr", rng_type=RNGInteger(0, 0xFFFF)),
        TestArg("length", rng_type=RNGInteger(1, 256)),
        directed_vectors={"zeros": {"addr": 0, "length": 1}},
        vector_constraints={"aligned": lambda v: v.addr % 4 == 0},
    )

@register("dma_short")
def dma_short():
    # Same vectors and constraints, a narrower length
    return dma_burst().extend(TestArg("length", rng_type=RNGInteger(1, 8)), nsamples=5)

@register("dma_prio")
def dma_prio():
    return dma_burst().extend(
        TestArg("prio", rng_type=RNGChoice([0, 1, 2])),        # a new argument, added last
        defaults={"prio": 0},                                   # its value in the kept vectors
        directed_vectors={"urgent": {"addr": 0, "length": 4, "prio": 2}},
        vector_constraints={"aligned": None},                   # None removes one
    )

• Arguments. A TestArg with a name the base has replaces that argument in its place, and the vectors keep their values. A new name is added after the others; every kept vector then needs a value for it, from defaults or the TestArg's fixed value=, or the call fails and names the vector.
• Vectors and constraints are given as dicts by name: a new name is added, a known name is replaced in its place, and None removes it (a name the base does not have fails). A list of constraints is added after the others. Vectors are written for the new argument list, in any form the constructor takes, and a pytest.param(...) vector keeps its marks.
• Settings (nsamples, max_retries, per_sequence_samples, max_exhaustive, ids, always_include_directed) are the base's unless given.
• A factory that takes inputs passes them on: def esm_burst(ctx): return esm_rw(ctx).extend(...). The new strategy has a name of its own, so its random rows differ from the base's for the same seed.

🔌 Fixture Integration

Strategies work seamlessly with standard pytest fixtures. You don't need any special configuration; just add the fixture to your test signature.

@pytest.fixture
def database():
    return MockDB()

@strategy("user_strategy")
def test_db_insert(username, age, database): # 'database' is a fixture
    # 'username' and 'age' come from the strategy
    user = database.create_user(username, age)
    assert user.id is not None

A fixture can take strategy arguments too: with a fixture def user(username, age), a test def test_login(user) decorated with @strategy("user_strategy") gets a user built from each row. The testbench context reaches fixtures through strategies_ctx and get_context() (see The context in fixtures). A fixture's random draws follow --rng-seed (see Draws outside the rows).

🔖 Test IDs and Selecting Rows

Each row's test ID names the row instead of showing its values, so a row keeps its node ID for every seed:

Row	ID
Directed vector zeros	test_write[directed-zeros]
Test vector max	test_write[test-max]
Random row 3	test_write[rand-3]
Random row 1 of a finite run with the Series value 2	test_esm[ch=2-rand-1]
A row of --nsamples=auto	test_esm[ch=0-dev=b]
The row of an empty skip_if_empty sequence	test_esm[skipped]

• Random rows are numbered from 0. An argument's value is in the ID exactly when the row enumerates it: the Series arguments in finite mode, the Series and RNGSequence arguments with per_sequence_samples=True, and every Series and RNGSequence argument under --nsamples=auto. Several of them are joined with - in declaration order. Rows of --nsamples=auto have no rand- part, because there is one per combination.
• An enumerated value is written ARG=TEXT. The text of a bool or None is str(v), of an Enum member its name, of an int or a float its repr, of a class or function its __name__, and of a string the string itself, when it is non-empty, at most 40 characters, printable, and has no whitespace, =, ~, [ or ]. A value listed twice gets ~1 the second time (Series([1, 1, 2]) gives ch=1, ch=1~1 and ch=2). A value without text (a tuple, bytes, another string) is written by its position, ARG<position>: cfg0, cfg1. So are two values that would get the same text (1 and "1").
• Under --nsamples=auto, every seed gives the same set of IDs as long as the constraints drop no combination; only the order of RNGSequence rows changes. A Series-only strategy with more rows than combinations repeats values, as ch=0-rand-0 and ch=0-rand-1.
• A test with a stacked @pytest.mark.parametrize, or several @strategy decorators, gets IDs joined with - in node-ID order, the decorator closest to the def first: test_two[directed-zeros-fast].

Selecting rows with -k. pytest matches -k against the parts of the ID:

• -k zeros selects the directed vector zeros, -k directed- every directed row, and -k "not rand-" leaves out the random rows. Keep the -: -k matches substrings of the whole node name, so a test, class, module or vector whose name contains the word is matched too (-k directed also selects test_directed_mode, and -k "not rand" also leaves out test_random_access and the directed vector operand_max).
• -k "rand-3" also matches rand-30 to rand-39. To select one random row, use -k "test_write[rand-3]" or its node ID.
• -k cannot contain = (pytest's grammar), so select the rows of a sequence value by node ID: pytest "tests/test_esm.py::test_esm[ch=2-rand-1]".
• Directed and test vector names can be any non-empty string, but a name in -k can hold only letters, digits and _ - . : / + \ [ ] (not whitespace, =, parentheses, commas or quotes), and a non-ASCII name needs pytest's escaped form (-k 'caf\xe9'): identifier-like names are easiest to select.

Rerunning one row. pytest "tests/test_dma.py::test_write[rand-3]" --rng-seed=S runs that row with the values it had in the run with seed S, for any --nsamples that still generates it: rows keep their values when there are more rows or fewer. A row the run does not generate (rand-12 with the default of 10 rows) gives pytest's own "not found" error, so give the same --nsamples as the run. When a test fails, the plugin prints this command for each failed row (see Reproducing a failure), and --lf reruns the failed rows with their seed (see Rerunning failures with --lf and --sw).

strategies_ids = values. The ini option strategies_ids sets the format of the IDs: names, the default, or values, the 3.0 IDs built from the values (addr=0,len=1), which change with the seed. A run can switch with -o strategies_ids=values. Any other value is a usage error.

Parameter(ids=...) sets the test IDs of one strategy: "names" or "values" overrides the ini option, and a function builds them. The function is called once per row while the tests are collected, with the row's VectorInfo (see Per-Test Metadata), whose id is the row's ID in the ini option's format. It returns the ID to use, or None to keep that one:

def by_mode(info):
    # (8, "le") has no text of its own: the default ID is mode0-rand-1
    if info.kind == "random":
        width, order = info.values.mode
        return f"w{width}{order}-rand-{info.index}"
    return None  # keep directed-zeros

Parameter(
    TestArg("mode", rng_type=Series([(8, "le"), (16, "be")])),
    TestArg("addr", rng_type=RNGInteger(0, 4095)),
    directed_vectors={"zeros": ((8, "le"), 0)},
    ids=by_mode,  # test_write[directed-zeros], test_write[w8le-rand-0], test_write[w16be-rand-0], ...
)

Build IDs from what does not depend on the seed: the row's kind, name and index, and the values of its enumerated arguments. An ID built from a drawn value changes with the seed, as in 3.0, so -k expressions, --deselect lists and CI history lose track of the row from one run to the next (the plugin's rerun commands, which carry the seed, still work).

Rows that get the same ID are suffixed the way pytest suffixes duplicate IDs (len=17_0 and len=17_1, odd0 and odd1), so they also pass under pytest's strict_parametrization_ids, and item.stash[VECTOR_KEY].id is the suffixed ID. Anything but a non-empty str or None, or an exception, fails the collection of the test: In test_write: Strategy 'burst': ids= returned 42 for row rand-3; return a str or None. The skipped row of an empty skip_if_empty sequence keeps skipped without a call. A pytest.param vector cannot have an id=: the vector's name is its ID, so rename the vector or use ids=.

ids= and strategies_ids never change the generated rows. The RNG draws of a test's setup, body and teardown, and of its function-scoped fixtures, come from streams of the test's node ID (see Draws outside the rows), which contains the ID, so they change with it: reproduce a run with the same IDs.

🧾 Per-Test Metadata

The item of each strategy row carries a VectorInfo that describes the row, in item.stash[VECTOR_KEY]. VectorInfo, VECTOR_KEY and VECTORS_KEY are exported from pytest_strategy. The plugin stores it while the tests are collected, before any pytest_collection_modifyitems hook runs, so your hooks, fixtures and reports can read it:

# conftest.py
import pytest

from pytest_strategy import VECTOR_KEY


def pytest_collection_modifyitems(items):
    # Run every directed vector before the random rows
    def rank(item):
        info = item.stash.get(VECTOR_KEY, None)
        return 0 if info is not None and info.kind == "directed" else 1

    items.sort(key=rank)


@pytest.fixture(autouse=True)
def _log_row(request):
    info = request.node.stash.get(VECTOR_KEY, None)
    if info is not None:
        print(f"{info.strategy} {info.id}: {info.values}")

Field	Value
strategy	The strategy's resolved name, as in the -v summary
origin	Where the factory is defined, "tests/dma/strategies.py:12" (relative to the rootdir), or None
kind	"directed", "test", "random", "exhaustive" (a row of --nsamples=auto) or "skipped"
name	The directed or test vector's name, else None
index	The vector's position in directed_vectors (what --vector-index takes) or test_vectors; the row's number within its combination for a random row; its position among the combinations for an exhaustive row; None for "skipped"
enumerated	The arguments the row enumerates, whose values are in the ID
values	The row, a Vector (info.values.addr); Nones for "skipped"
id	This strategy's part of the test ID, after ids= and duplicate suffixes, before pytest escapes it
seed	The run's seed
context	The fingerprint of the context the factory received, or None when it declares no ctx
constraints_off	The constraints turned off in this strategy
streams	The version of the random streams the values come from (1)

• info.to_dict() gives the row as JSON-ready data with "schema": 1, its values in the export's encoding (see Metadata Export).
• A test with several @strategy decorators has one VectorInfo per strategy in item.stash[VECTORS_KEY], in the order of the node ID; VECTOR_KEY holds the first.
• Items without a strategy, and the item pytest makes for an empty parameter set, have none: use item.stash.get(VECTOR_KEY, None).
• The infos also travel as the argument of a strategy mark on each row, which adds no -k keyword and no marker. item.iter_markers("strategy") yields the test's own @strategy mark first and the row's marks after it, and item.get_closest_marker("strategy") returns the test's own mark: read the stash instead.
• VectorInfo is frozen, and you read it rather than build it. Fields that later releases add come with defaults.

🎛️ Options

Control test generation directly from the command line:

Option	Description	Example
--nsamples	Number of random samples per strategy (default 10), or per Series/RNGSequence combination for a strategy with per_sequence_samples=True. An integer overrides a strategy's own nsamples. auto enumerates the Series/RNGSequence arguments (see What --nsamples=auto does). Any other value is a usage error.	pytest --nsamples=50 or --nsamples=auto
--vector-mode	Generation mode: all, random_only, directed_only, mixed, test (see Directed and Test Vectors)	pytest --vector-mode=test
--vector-name	Run only the directed vector with this name	pytest --vector-name=edge_case_1
--vector-index	Run only the directed vector at this index (0-based, in definition order)	pytest --vector-index=0
--rng-seed	Set seed for reproducibility. Without it, --lf and --sw reuse the seed of the failed run (see Reproducibility)	pytest --rng-seed=42
--strategy-constraint-off	Turn named constraints off for this run: NAME in every strategy, STRATEGY:NAME in one; comma-separated, repeatable (see Rows and Constraints)	pytest --strategy-constraint-off=dma_burst:aligned
--strategy-details	Show a pytest-strategies section under each failed strategy row's traceback: its values, seed and rerun command (off by default; see Reproducing a failure)	pytest --strategy-details
--list-strategies	List the registered strategy names and exit	pytest --list-strategies

Ini options go in your pytest configuration, and -o NAME=VALUE sets one for a single run:

Ini option	Default	Description
strategies_ids	names	The format of the test IDs: names (rand-3) or values (addr=0,len=1, the 3.0 format); see Test IDs and Selecting Rows. Any other value is a usage error
strategies_max_exhaustive	100000	The most rows --nsamples=auto or per_sequence_samples=True may generate for one strategy; Parameter(max_exhaustive=...) overrides it

In a native TOML configuration (pytest 9's pytest.toml, or [tool.pytest] in pyproject.toml), write the values as strings: strategies_max_exhaustive = "500000".

The plugin also provides the session fixture strategies_ctx (see The context in fixtures) and the strategy marker that @strategy adds. Command-line options added in later releases start with --strategy- and ini options with strategies_; there are no command-line copies of ini options, since -o sets those.

With -v, a "Strategy Summary" section lists each strategy with the number of tests that use it, their directed and random rows (and their test rows, the exhaustive rows of --nsamples=auto and the skipped row of an empty skip_if_empty sequence, when there are some), and where the sample count came from (--nsamples, Parameter(nsamples=) or the default), the rejections per constraint and the constraints turned off, followed by the contexts' fingerprints.

🔄 Reproducibility

The RNG seed of each run is printed in the pytest report header:

pytest-strategies: RNG seed = 1763926297314361000

Reproducing a failure

When tests fail, the plugin also prints how to rerun them with the same vectors, after the failure tracebacks and before the short test summary, even with -q:

pytest-strategies: reproduce with --rng-seed=1763926297314361000
pytest-strategies: failed rows:
  pytest 'tests/dma/test_write.py::test_write[rand-3]' --rng-seed=1763926297314361000  # burst random 3
  pytest 'tests/dma/test_write.py::test_write[directed-zeros]' --rng-seed=1763926297314361000  # burst directed zeros

When the factories of the failed tests received a context, the first line ends with its fingerprint, such as (context 976bcfdf) (see The context fingerprint). Then comes, for each strategy row whose setup or call failed, the command that runs that row alone with the same values, from the folder pytest was started in, and what the row is. Below -v at most 10 rows are listed, followed by ... and 4 more, and a last line points to --strategy-details; -qq prints only the first line. An error in a test's teardown alone lists nothing. However many rows fail, this list stays short: the plugin adds nothing else to a failure's report by default.

With --strategy-details (on the command line, or in addopts to keep it on), each failed strategy row also gets a pytest-strategies section under its traceback that says what the row is, its values and how to run it again. One block per failed row: a run where 5,000 rows fail prints 5,000 blocks, so turn it on for the failures you are looking into, such as pytest --lf --strategy-details -x.

------------------------------ pytest-strategies -------------------------------
strategy  burst (tests/dma/strategies.py:12)
vector    rand-3 (random row 3)
values    addr=4096
          len=17
seed      1763926297314361000
context   3f2a9c1e
rerun     pytest 'tests/dma/test_write.py::test_write[rand-3]' --rng-seed=1763926297314361000

• strategy names the strategy and where its factory is defined, and vector the row: directed-zeros (directed vector 'zeros', #0), test-max (test vector 'max', #1), ch=2 (exhaustive row 5). values shows each argument's value by its repr (by its type's name when the repr shows a memory address, and as <Reg: repr() raised RuntimeError> when it raises), cut at 4,000 characters below -vv. The context line appears only when the factory received ctx. A test with several @strategy decorators gets one block per strategy and one rerun line.
• The command adds every option that decides which rows exist and what they hold: --nsamples, --vector-mode, --vector-name and --vector-index when the run had them (from the command line, addopts or a conftest), the -o strategies_* overrides, -c and --rootdir as given, and the constraints the run turned off in the row's own strategies, as --strategy-constraint-off=STRATEGY:NAME items, so that rerunning one module never names a constraint only another module has. It also adds the run's -o empty_parameter_set_mark (the last one, as pytest applies it): the rerun collects the row's whole module, where a strategy without the vector --vector-name selects gets an empty parameter set, which an ini file's empty_parameter_set_mark = fail_at_collect turns into a collection error. Arguments are quoted for the platform's shell. What shapes the context (an environment variable, an option of your own) is not added: compare the context lines.
• pytest names a test file outside the rootdir (pytest -c ci/pytest.ini with the tests in tests/) by the path the run started from (given on the command line, or the folder pytest runs in), so a rerun of its node ID would give the row another node ID and other values. Its command starts from the run's own paths instead, with the run's --ignore and --ignore-glob options, and selects the row with -k: pytest . --rng-seed=S -c ci/pytest.ini -k 'test_write[rand-3]', with the module's name too (test_dma.py and test_write[rand-3]) when another test has the same name. When no -k expression selects only that row (a name with =, or two modules of the same name), the command names the node ID, its section ends with a note line, and its row in the list with (outside the rootdir): run with a --rootdir that contains the tests, such as --rootdir=., to get a command that reproduces them.
• A character the terminal cannot encode, such as a Greek letter in a Windows CI log written in cp1252, is written as a backslash escape (\u03a9), as pytest writes test IDs, in the section and in the list of failed rows. pytest would otherwise write the whole section escaped, on one line, with the backslashes of the escapes in its node ID doubled. So the commands run as printed, unless such a character is in a path or in a test module's class or function name (pytest escapes only parametrize IDs, so the node ID has it raw), in --vector-name, in a strategy or constraint name of a --strategy-constraint-off item, or in an -o value: that argument then reaches pytest escaped, and a constraint it names stays on.
• A failure that pytest reports as plain text, such as an XPASS of a strict xfail, is listed but gets no section. An error reported without a traceback, such as a missing fixture, gets the section below its report. Under pytest-xdist the workers send the rows to the controller.

A --junitxml report holds the same information, under pytest-xdist too:

• With --strategy-details, the failure (or error) text of each failed strategy row ends with its pytest-strategies section.
• The test suite gets the property pytest_strategies.seed, and pytest_strategies.failed.0, pytest_strategies.failed.1 and so on, the commands of the list of failed rows, in the same order:

  <properties>
    <property name="pytest_strategies.seed" value="1763926297314361000"/>
    <property name="pytest_strategies.failed.0" value="pytest 'tests/dma/test_write.py::test_write[rand-3]' --rng-seed=1763926297314361000"/>
  </properties>
  

• With junit_family = xunit1 or legacy, each failed strategy row's test case also gets string properties: pytest_strategies.strategy, .kind, .name (for a directed or test vector), .index, .id, .value.<argument> for each value (as in the section), .seed, .context and .constraints_off when set, and .command, the command to run from the rootdir. With several @strategy decorators they are numbered per strategy, in the order of the node ID: pytest_strategies.0.strategy, pytest_strategies.1.strategy. pytest's default family, xunit2, has no properties per test case in its schema, so there the test cases get none.

Rerunning failures with --lf and --sw

--lf and --sw rerun the failed rows with the values they failed with. The plugin records the seed of each failed strategy row, with the options its rerun command adds, in pytest's cache (pytest-strategies/failed-seeds). A run with --lf, --sw or --sw-skip and no --rng-seed reuses the seed of the newest of the failed rows it reruns, among those whose file still exists and that the run collects: those the paths and node IDs on the command line select or, without any, those in the testpaths or in the folder pytest runs in. A test file outside the rootdir (-c ci/pytest.ini with the tests in tests/) counts by the name pytest gives it, its path from the path the run started from, so run --lf from the same paths as the failed run. -k and -m apply once the tests are collected, after the seed is chosen, so they do not choose it: give a path or a node ID instead. A line under the seed line says so:

pytest-strategies: RNG seed = 1763926297314361000
pytest-strategies: seed reused from the failed run for --lf (--rng-seed overrides)

• The recorded options are not applied. When they differ from the run's, the line ends with them (; recorded with --nsamples=13, or without --vector-mode=test): add them to get the same rows. A recorded -c or --rootdir is kept relative to the rootdir, so it names the same file from any folder (an ini file outside the rootdir, as in -c /dev/null, by its absolute path), and agrees with a run that uses that ini file or rootdir without giving it (pytest --lf in ci after pytest -c ci/pytest.ini). A --rootdir the run gives agrees with rows recorded without it. A -o empty_parameter_set_mark is not compared, as it decides no row's values, but the commands below keep it, and the line names a recorded one with the other recorded options it names: a recorded --vector-name may need it to collect the rows' modules.
• Failed rows recorded under another seed are deselected, so pytest keeps them in its last-failed set, and the end of the run gives the command that reruns them, one per seed (and per set of options), even with -qq. It names their files, or their node IDs when a file also holds other failed tests, so it reruns no other failed test with values it did not fail with:

  pytest-strategies: deselected 2 failed rows recorded under another seed; run them with:
    pytest --lf --rng-seed=1763926297314360000 tests/dma/test_write.py  # 2 rows
  

  For rows in files outside the rootdir, whose node IDs a path would change, the command starts from the run's own paths instead (pytest --lf --rng-seed=S -c ci/pytest.ini .), and leaves out the other failed tests these paths collect with --deselect, or with --ignore for a file inside the rootdir: pytest's --lf skips the files outside the rootdir once it collected such a file.
• When the run collects none, or not all, of the failed rows it reused the seed of (their test was deleted or renamed, or the run's options do not generate them), the end of the run says so and gives the command that reruns them, with their recorded options:

  pytest-strategies: 1 failed row recorded under the reused seed was not collected; unless deleted or renamed, run it with:
    pytest --lf --rng-seed=1763926297314360000 --nsamples=13 tests/dma/test_write.py  # 1 row
  

• --rng-seed always wins and deselects nothing, and so does a config.option.rng_seed that a conftest.py sets in its pytest_configure. An RNG.seed() call there does not change the reused seed, as it does not change --rng-seed. --ff and --nf draw a new seed, because they run every test, and so does --sw-reset.
• A row leaves the record once it passes under its recorded seed and options (-o empty_parameter_set_mark aside), its teardown included: a row whose call passes and whose teardown fails keeps its entry, as pytest keeps it in its last-failed set, so the next --lf reruns it with the same values. An error in a test's teardown alone is not recorded. Under pytest-xdist the controller reads and writes the record and sends the seed to the workers. Without pytest's cache plugin (-p no:cacheprovider) there is no --lf or --sw, and nothing is recorded.

Random streams

Each strategy and test pair draws from its own random streams ("streams v1"). Each argument of a random row draws from a stream derived from the seed, the strategy name, the test's node ID without its parameters (its file path relative to the rootdir, its class and its name), the row (the values it enumerates and its number) and the argument's name. The factory draws from a stream derived from the seed, the strategy name and the test's node ID without its parameters, so its draws do not change the rows. A random row that a constraint rejects is drawn again from the same streams, so the row's values are those of the first draw every constraint accepts. The derivation is versioned (VectorInfo.streams) and changes only in a major release.

For a test that passes the factory itself (@strategy(burst)), the strategy name is a name that the factory's own file registers it under when it is imported, the first in alphabetical order, or else its qualified name (see Passing the factory). It never depends on which strategy files, test modules or plugins a run loaded, nor on where the factory's file is.

What this gives you:

• A test gets the same vectors and node IDs whether you run the whole suite, one file or one test, in any collection order, with any --import-mode, on any OS, with any PYTHONHASHSEED, and under pytest-xdist.
• Two tests that use the same strategy get different random vectors, and so does a test method that two classes inherit.
• A row keeps its values when there are more rows (--nsamples=50 keeps the rows of --nsamples=10), when it runs alone by its node ID, under any --vector-mode, --vector-name or --vector-index, and when directed or test vectors are added. Adding, removing or changing another drawn argument leaves this argument's values as they are in every row that the constraints accept or reject as before. A row whose draws a constraint now accepts or rejects differently, because it reads the changed argument, stops at another draw, which gives this argument a new value in that row too. Reordering the enumerated arguments keeps every row's values. A constraint changes only the rows it rejects, so turning one off keeps the rows it never rejected, and appending a Series value keeps the other combinations' values.
• Values that a strategy factory draws itself are reproduced too, through rng or RNG.generator(), the generator the RNG types use (also for other random operations, such as shuffle or gauss).

What does not hold:

• Renaming an argument changes its values. Renaming the strategy, the test, its class or its file, or changing the rootdir, changes every row: keep the same rootdir, because the test's path relative to the rootdir is part of the stream (pytest uses the directory of your ini file, such as pytest.ini, as the rootdir when there is one).
• Adding or removing an enumerated argument, or changing whether an argument is enumerated (a Series against an RNGSequence in finite mode, per_sequence_samples, --nsamples=auto), changes every row.
• For Series and per_sequence_samples, which rows exist depends on the count and the sequence lengths, and on the seed when a combination runs out of retries; only each row's values are stable.
• Inserting a Series value that has no text of its own (a tuple, say) shifts the positions of the later values, and adding an RNGSequence value reorders its rows under --nsamples=auto.
• A new Python minor version may change what randint, choice and sample return for a seed, because Python only guarantees random().

These hold when a row's random values come from its RNG types:

• A custom RNG type (a subclass of RNGType) should draw inside its generate(), from RNG.generator() or the RNG.* helpers called there, and keep no state between calls: a counter that walks a pattern makes row k depend on the rows drawn before it.
• A constraint that draws (RNG.integer()), or an RNG type that draws from a generator kept from the factory (its rng), draws from the plugin's generator outside the rows' streams: those values repeat for the same seed, tests and options, but change when other rows or tests change. The plugin then emits one PytestStrategiesWarning per test, naming the strategy and the test; under filterwarnings = error it fails the collection.
• A factory wrapped in functools.cache runs once, for the first test that uses it, so its draws come from that test's stream: it should not draw, because a test run alone could get other values.

Draws outside the rows

The plugin draws from a random.Random instance of its own and never seeds Python's global random module. So your own use of random does not change the generated vectors, and --rng-seed does not reproduce what random.random() returns in your code, including in a factory: draw from rng or RNG.generator() there.

The seed also reproduces the RNG draws made while the tests run. Each phase of a test (setup, call and teardown) draws from a stream derived from the seed, the test's node ID and the phase, and each fixture's setup from a stream derived from the seed, the node ID of the test, class, module or package the fixture is set up for (none for a session fixture), its name, its parameter index and where it is defined (its module, or its file for a conftest.py or a test module, its function, and the folder, module or class it is registered for), so that a fixture overriding another of the same name, or one fixture function imported into two folders' conftest.py files, gets values of its own. So a test body's RNG.integer() and a fixture's draws are the same whether the test runs alone, in the suite, in another order or under pytest-xdist, also when a module- or session-scoped fixture is set up during that test's setup. A fixture's teardown (the code after its yield, or a finalizer) runs in the teardown of the test that ends the fixture's scope, and draws from that test's teardown stream. An RNG.seed() call in a test, a fixture, a factory or a strategy file changes only the rest of that stream's draws, and RNG.refresh_seed(key=...) gives a stream derived from the seed and the key, whatever ran before it.

To make plain random draws in test bodies follow the seed too, seed it for each test in your conftest.py:

# conftest.py
import random

import pytest

from pytest_strategy import RNG


@pytest.fixture(autouse=True)
def _seed_random(request):
    # Each test's random draws follow --rng-seed, whatever ran before it
    random.seed(f"{RNG.get_seed()}:{request.node.nodeid}")

Values that a strategy file draws when it is imported are reproduced by the seed (see Strategy Files and Scoped Names). A draw at module level in a test module (for example BASE = RNG.integer(0, 1000)) comes from a stream derived from the seed and the module's path, so running one file alone gives it the same value.

Draws made when a conftest.py is imported do not follow the seed: pytest imports the conftest files of the rootdir and of the folders on the command line before the seed is set, so a node-ID rerun would import them at another time. Move such draws into the context hook, a fixture or a strategy file. For the same reason, a strategy file that a conftest.py imports at its top is executed with that conftest.py, not on its own stream, and its import-time draws do not follow --rng-seed either; import it inside the fixture or hook that needs it. A helper module that test modules or strategy files import (a helpers.py that is not a test module, a conftest.py or a strategy file) runs on the stream of the first module that imports it, so its import-time draws, and the draws that this module makes after importing it, change with what was collected before: running one file alone can give them other values. Move such draws into a fixture, the context hook or a strategy file as well.

pytest-xdist

Runs with -n work with or without --rng-seed. The controller sends its seed to the workers, so they all generate the same tests, as long as the strategy factories and the context hook give the same result in every worker.

Test IDs name the rows instead of showing their values, so pytest-xdist does not notice workers that generate different values under the same IDs. The plugin checks it: when its session ends, each worker sends the fingerprint of each context it computed (by its label) and a digest of each strategy's values (the first 8 hex characters of the SHA-256 of its tests' node IDs and values, in the fingerprint's encoding, taken after the collection). When two workers differ on a context or a strategy, the controller names them in red at the end of the run, before the line that says how to reproduce it, and a run that would have passed (or collected no tests) fails with exit code 4:

pytest-strategies: the xdist workers generated different vectors:
  context tests/tb_a/conftest.py: gw0 1a2b3c4d, gw1 9f8e7d6c
  values of strategy dma_burst: gw0 5e6f7a8b, gw1 0c1d2e3f
Make pytest_strategies_context and the strategy factories give the same result in every worker:
no temporary paths, process IDs, times, unseeded random values or lists built from sets
(leave them out, or use pydantic Field(exclude=True) in a context).

• Typical causes are a factory that draws from Python's global random (draw from rng or RNG.generator()), or that builds a list from a set of strings, whose order follows each worker's PYTHONHASHSEED (sort it), and a context holding a temporary path, a process ID or the time.
• A context differs even when the values agree, because tests can read it through strategies_ctx. The fingerprint is taken when the hook returns, so a test that changes the object on one worker does not count. Nor does the context of a folder whose wrapper=True implementation built it on that worker from an object a test had already changed (a folder first asked while the tests run): its fingerprint shows the change, so it is not compared.
• A context or a strategy that only one worker computed is not compared: a test that runs on one worker only can compute its folder's context through strategies_ctx or get_context(). A worker that crashed sent nothing and is left out.
• When the values change the test IDs (a Series built from the context), pytest-xdist itself stops with "Different tests were collected", and the plugin's message follows it, naming the cause.

🧭 Upgrading to 4.0

4.0.0 changes test IDs once and random values once more, and removes what 3.0 deprecated. The CHANGELOG's 4.0.0 section starts with "Migrating from 3.x", a checklist with what you see for each change and what to do. In short:

• Test IDs name the rows (test_write[rand-3]) instead of showing their values. Update -k expressions, --deselect lists and anything else that names value IDs, or keep the 3.0 IDs for a while with strategies_ids = values.
• Random values. For the same seed, 4.0.0 gives random rows, and the values that factories, strategy files and the context hook draw themselves, other values than 3.x, so a seed recorded with 3.x does not reproduce that run's random rows. Directed and test vectors and Series values are the same, unless a factory draws them.
• Factory inputs by name. A factory parameter named anything but nsamples, ctx, rng or options needs a default: rename def factory(n) to def factory(nsamples), or drop the parameter.
• Tuple factories. A factory that returns an (argnames, samples) tuple fails the collection of the tests that use it: return a Parameter, with one TestArg per argument and the fixed rows as directed_vectors. A fixed table with no random arguments fits @pytest.mark.parametrize better. A fixed table built from the context (a register map) goes in a pytest_generate_tests hook of the test module or of a conftest.py, which calls metafunc.parametrize("name,offset", get_context(metafunc.config, metafunc.definition.path).registers) for that test; it then runs exactly the table's rows, whatever --nsamples and --vector-mode say. A Parameter whose arguments all have a value= still gets nsamples random rows with those placeholder values: nsamples=0 on the Parameter leaves them out of the default run, but an integer --nsamples=N brings back N of them, and under --vector-mode=random_only or --vector-mode=test the test gets no rows and is skipped.
• Removed APIs. TestArg(directed_values=..., test_values=..., always_include_directed=...) raises TypeError: give the Parameter directed_vectors=, test_vectors= and always_include_directed= instead. The TestArg properties directed_values, test_values and has_directed_values are gone. RNG.set_max_retries() is gone: use Parameter(max_retries=...). configure() and Strategy.set_config() did nothing: delete the calls.
• Keyword-only options. TestArg's options after rng_type, strategy(..., validate_signature=...), export_strategies(format=...) and the options of generate_vectors() must be passed by keyword.
• Rows are Vectors, tuples whose type(v) is tuple is false, and a namedtuple directed or test vector is placed by its field names, no longer by position. A dict vector gives the test its values by key, where 3.0 passed the key strings. A bare value such as 404 as a vector is an error, and so are an id= on a pytest.param vector and a pytest.param with more or fewer values than arguments.
• Record parameters. A fixture that asks for a strategy argument's name makes the strategy pass its arguments by name, so a test that also takes a record of them fails collection; with validate_signature=False, a record parameter whose same-named fixture consumes the arguments now gets that fixture's value.
• Export. export_strategies() returns schema 1, with a strategies list instead of a mapping by name.

📝 License

MIT License. See LICENSE for details.