ordeal

Your tests pass. Your code still breaks.

Ordeal finds what you missed — edge cases, untested code paths, bugs that only show up in production. No test code to write. Just point and run.

Open a terminal and paste this (uvx runs Python tools without installing them):

uvx ordeal mine ordeal.demo

mine(score): 500 examples
  ALWAYS  output in [0, 1] (500/500)         ← always returns a value between 0 and 1
  ALWAYS  monotonically non-decreasing        ← bigger input = bigger output, always

mine(normalize): 500 examples
     97%  idempotent (29/30)                  ← normalizing twice should give the same result
                                                 ...but ordeal found a case where it doesn't

Now point it at your code. If you have myapp/scoring.py, the module path is myapp.scoring:

uvx ordeal mine myapp.scoring       # what do my functions actually do?
uvx ordeal audit myapp.scoring      # what are my tests missing?

Or let your AI assistant do it — open Claude Code, Cursor, or any coding assistant and paste:

"Run uvx ordeal mine and uvx ordeal audit on my main modules. Explain what it finds and fix the issues."

ordeal ships with an AGENTS.md — your AI reads it automatically and knows how to use every command.

pip install ordeal                        # or: uv tool install ordeal

30-second example

from ordeal import ChaosTest, rule, invariant, always
from ordeal.faults import timing, numerical

class MyServiceChaos(ChaosTest):
    faults = [
        timing.timeout("myapp.api.call"),         # API times out
        numerical.nan_injection("myapp.predict"),  # model returns NaN
    ]

    @rule()
    def call_service(self):
        result = self.service.process("input")
        always(result is not None, "process never returns None")

    @invariant()
    def no_corruption(self):
        for item in self.service.results:
            always(not math.isnan(item), "no NaN in output")

TestMyServiceChaos = MyServiceChaos.TestCase

pytest --chaos                    # explore fault interleavings
pytest --chaos --chaos-seed 42    # reproduce exactly
ordeal explore                    # coverage-guided, reads ordeal.toml
ordeal explore --runner compose   # long-lived Docker Compose services

You declare what can go wrong (faults), what your system does (rules), and what must stay true (assertions). Ordeal explores the combinations.

Move from functions to real services

Already have a Docker Compose application? Ordeal can keep the topology alive, send repeated HTTP requests, kill or restart workers, delay or corrupt responses, carry IDs and tokens between operations, and save the exact action sequence.

[compose]
base_url = "http://127.0.0.1:8080"
health_path = "/health"
services = ["api", "worker"]

[[compose.requests]]
name = "list-items"
path = "/items"

ordeal explore --runner compose
ordeal explore --runner compose --save-artifacts  # bound trace + durable manifest

Real service timing is not perfectly deterministic. Ordeal says so directly: the trace is exact, while replay is reported as attempted N / reproduced M. The complete service evidence loop is: explore → coverage → exact replay → bounded finding → committed regression → verify --ci → test the test.

New to those terms? Start with From “it broke once” to “it cannot come back”. Then follow the copyable Compose evidence loop or use the plain-English service overview to choose a quickstart, schema, fault model, trace reference, CI guide, or fix.

Ordeal itself runs a real Docker Compose kill-and-recovery gate before publishing. See Compose CI and operations for the exact sequence, what a green job proves, and what remains outside its scope.

Know what reliability behavior was actually tested

Line coverage can tell you that order code ran. It cannot tell you whether an API timeout happened while an order was created, or whether duplicate charging was checked afterward. Add operation and fault to existing assertions:

from ordeal import always, declare

declare(
    "eventual_commit",
    "sometimes",
    operation="create_order",
    fault="worker_restart",
)

always(
    charge_count == 1,
    "no_duplicate_charge",
    operation="create_order",
    fault="timeout",
)

The end of pytest --chaos shows the evidence directly:

operation × fault × property
create_order × timeout × no_duplicate_charge     PASS
create_order × worker_restart × eventual_commit  NOT EXERCISED
refund × stale_response × balance_conserved      FAIL

PASS means the cell was observed and held. NOT EXERCISED exposes an expected test that never happened. FAIL means Ordeal observed a violation. The same rows are available as JSON from report(), including under pytest-xdist. Labels describe a fault your test really injected; they do not inject it themselves.

Start with Reliability Coverage, then use the practical guide and CI/platform guide.

Why ordeal

Testing catches bugs you can imagine. The dangerous bugs are the ones you can't — a timeout during a retry, a NaN inside a recovery path, a permission error after the cache warmed up. These come from combinations, and the space of combinations is too large to explore by hand.

Ordeal automates this. It brings ideas from the most rigorous engineering cultures in the world to Python:

Read the full philosophy to understand why this matters.

Evidence you can act on

ordeal scan myapp.scoring gives each finding a bounded evidence card:

• the exact claim justified by this finding
• a SHA-256 binding to the inspected callable source
• the exact input witness and its canonical hash
• immediate replay counts, matching the same exception type, message, and source seam
• a pending same-witness regression to run after the fix
• explicit limits: no root-cause claim, no claim about untested behavior

A supported finding reproduced exactly in the recorded attempts. An exploratory finding is still a lead, and an expected precondition is not a defect. None of these statuses certifies the whole project. The pass or finding applies only to the inputs, states, faults, properties, and runtime actually exercised. See Finding Evidence. For a jargon-free explanation of why this becomes a permanent test, read Fix a Bug Once.

Choose the scan documentation at your depth:

The durable regression workflow is one path, not a report handoff:

ordeal scan myapp.scoring --save-artifacts  # discover, replay, minimize, save
# fix the code
ordeal verify <finding-id> --allow-unsafe-artifacts
ordeal verify --ci                          # provider-neutral CI guard

Commit tests/test_ordeal_regressions.py and tests/ordeal-regressions.json; the richer .ordeal/findings/ dossier can remain local.

Use Durable Regressions in CI for rollout and exit codes, the FAQ for common decisions, and the schema reference for integrations and proof review.

Install

# From PyPI
pip install ordeal

# With extras
pip install ordeal[atheris]    # coverage-guided fuzzing via Atheris
pip install ordeal[all]        # everything

# As a CLI tool
uv tool install ordeal         # global install
uvx ordeal explore             # ephemeral, no install

# Development
git clone https://github.com/teilomillet/ordeal
cd ordeal && uv sync --locked --extra dev
uv run pytest                  # run the full test suite

When dependency metadata changes, run uv lock and commit uv.lock in the same change. Normal development should use uv sync --locked so lock drift fails early instead of rewriting the lockfile locally.

What's in the box

Stateful chaos testing

ChaosTest extends Hypothesis's RuleBasedStateMachine. You declare faults and rules — ordeal auto-injects a nemesis that toggles faults during exploration. The nemesis is just another Hypothesis rule, so the engine explores fault schedules like it explores any other state transition. Shrinking works automatically.

from ordeal import ChaosTest, rule, invariant
from ordeal.faults import io, numerical, timing

class StorageChaos(ChaosTest):
    faults = [
        io.error_on_call("myapp.storage.save", IOError),
        timing.intermittent_crash("myapp.worker.process", every_n=3),
        numerical.nan_injection("myapp.scoring.predict"),
    ]
    swarm = True  # random fault subsets per run — better aggregate coverage

    @rule()
    def write_data(self):
        self.service.save({"key": "value"})

    @rule()
    def read_data(self):
        result = self.service.load("key")
        always(result is not None, "reads never return None after write")

Swarm mode: each run activates a random subset of faults. Over many runs, this covers more fault combinations than always-all-on. Hypothesis handles the subset selection, so shrinking isolates the exact fault combination that triggers a failure.

Property assertions

Four property types, plus an optional declaration helper for deferred checks:

from ordeal import always, declare, sometimes, reachable, unreachable

always(len(results) > 0, "never empty")          # must hold every time — fails immediately
sometimes(cache_hit, "cache is used")             # must hold at least once — checked at session end
declare("error-recovery-path", "reachable")       # declare deferred expectation up front
reachable("error-recovery-path")                  # code path must execute at least once
unreachable("silent-data-corruption")             # code path must never execute — fails immediately

always and unreachable fail instantly, triggering Hypothesis shrinking. sometimes and reachable accumulate evidence across the full session. Use declare() when you want those deferred properties to fail even if the marker was never observed.

Inline fault injection (BUGGIFY)

Place buggify() gates in your production code. They return False normally. During chaos testing, they probabilistically return True:

from ordeal.buggify import buggify, buggify_value

def process(data):
    if buggify():                                    # sometimes inject delay
        time.sleep(random.random() * 5)
    result = compute(data)
    return buggify_value(result, float('nan'))        # sometimes corrupt output

Seed-controlled. Thread-local. No-op when inactive. This is FoundationDB's BUGGIFY for Python — the code under test is the test harness.

Coverage-guided exploration

The Explorer tracks which code paths each run discovers (AFL-style edge hashing). When a run finds new coverage, it saves a checkpoint. Future runs branch from high-value checkpoints, systematically exploring the state space:

from ordeal.explore import Explorer

explorer = Explorer(
    MyServiceChaos,
    target_modules=["myapp"],
    checkpoint_strategy="energy",  # favor productive checkpoints
)
result = explorer.run(max_time=60)
print(result.summary())
# Exploration: 5000 runs, 52000 steps, 60.0s
# Coverage: 287 edges, 43 checkpoints
# Failures found: 2
#   Run 342, step 15: ValueError (3 steps after shrinking)

Failures are shrunk — delta debugging removes unnecessary steps, then fault simplification removes unnecessary faults. You get the minimal sequence that reproduces the bug.

Scale with workers — each process gets a unique seed for independent exploration, results are aggregated:

explorer = Explorer(MyServiceChaos, target_modules=["myapp"], workers=8)

Configuration

# ordeal.toml — one file, human and machine readable
[explorer]
target_modules = ["myapp"]
max_time = 60
seed = 42
checkpoint_strategy = "energy"

[[tests]]
class = "tests.test_chaos:MyServiceChaos"

[report]
format = "both"
traces = true
verbose = true

See ordeal.toml.example for the full schema with every option documented.

QuickCheck with boundary bias

@quickcheck infers strategies from type hints. It biases toward boundary values — 0, -1, empty list, max length — where implementation bugs cluster:

from ordeal.quickcheck import quickcheck

@quickcheck
def test_sort_idempotent(xs: list[int]):
    assert sorted(sorted(xs)) == sorted(xs)

@quickcheck
def test_score_bounded(x: float, y: float):
    result = score(x, y)
    assert 0 <= result <= 1

Composable invariants

from ordeal.invariants import no_nan, no_inf, bounded, finite

valid_score = finite & bounded(0, 1)
valid_score(model_output)  # raises AssertionError with clear message

Invariants compose with &. Works with scalars, sequences, and numpy arrays.

Simulation primitives

Deterministic Clock and FileSystem — no mocks, no real I/O, instant:

from ordeal.simulate import Clock, FileSystem

clock = Clock()
fs = FileSystem()

clock.advance(3600)                      # instant — no real waiting
fs.inject_fault("/data.json", "corrupt") # reads return random bytes

Differential testing

Compare two implementations on the same random inputs — catches regressions and validates refactors:

New to this? Start with Differential Testing, Without Jargon, then follow the copy-paste Differential Quickstart.

from ordeal.diff import diff

result = diff(score_v1, score_v2, rtol=1e-6)
assert result.status == "no_divergence_observed", result.summary()
# diff(score_v1, score_v2): 100 examples, NO DIVERGENCE OBSERVED

Each side receives independent inputs and receiver state. The comparison covers returns, exceptions, mutated arguments, receiver state, and selected restorable side effects. A divergence carries one minimized, replay-verified witness; sampled agreement remains bounded evidence, not equivalence.

Every divergence also exposes a source-bound JSON record in result.artifacts. Pass artifact_dir=".ordeal/divergences" to persist the same canonical record. Use State and Side Effects for mutations, receivers, logs, caches, or other external state, and Evidence and Statuses when reviewing a witness or deciding what a result proves. For a story-led explanation of the durable card itself, read Divergence Evidence, then use its workflow, troubleshooting, or exact schema.

For stateful systems, pass two zero-argument factories and one operation/fault sequence. The same diff() call compares interface, outcomes, state, selected side effects, recovery parity, and an optional separate performance budget:

from ordeal.diff import FaultEvent, Operation, diff

result = diff(
    OldStore,
    NewStore,
    sequence=[FaultEvent("timeout"), Operation("read")],
    apply_fault=apply_fault,
)

If “operation-and-fault story” is new to you, start with the plain-language mental model, then run the self-contained first system comparison. The recipes cover APIs, state, effects, and budgets; the troubleshooting guide explains every common surprising result.

For a committed refactor, run both versions in isolated worktrees and subprocesses while replaying the same inputs:

ordeal diff myapp.scoring --base-ref origin/main --candidate-ref HEAD --save-artifacts

Start with Compare Two Git Revisions for the sealed-worktree mental model and first result. Use Revision Diff Troubleshooting for refs, imports, fixtures, and inconclusive replay; the Revision Diff Schema is the machine contract for saved JSON.

Replacing a whole module? Matching the old version is not enough: a perfect copy can preserve an old bug. The migration workflow also checks explicit business rules, test strength, and the replacement on its own:

ordeal migrate oldpkg.scoring newpkg.scoring -c ordeal.toml

Start with the layman-friendly Safe Module Migrations, then follow the migration workflow.

Mutation testing

Validates that your chaos tests actually catch bugs. If you flip + to - in the code and your tests still pass, your tests have a blind spot:

from ordeal.mutations import mutate_function_and_test

result = mutate_function_and_test("myapp.scoring.compute", my_tests)
print(result.summary())
print(result.test_protection_view())
# Mutation score: 15/18 (83%)
#   SURVIVED  L42:8 + -> -
#   SURVIVED  L67:4 negate if-condition

test_protection_view() combines the score, survivors, kill attribution, and property strength into a scoped weak, inconclusive, or protective_within_measured_scope verdict. It never treats line coverage alone as proof that assertions are meaningful.

Fault library

from ordeal.faults import io, numerical, timing, network, concurrency

# I/O faults
io.error_on_call("mod.func")            # raise IOError
io.disk_full()                           # writes fail with ENOSPC
io.corrupt_output("mod.func")           # return random bytes
io.truncate_output("mod.func", 0.5)     # truncate to half

# Numerical faults
numerical.nan_injection("mod.func")      # output becomes NaN
numerical.inf_injection("mod.func")      # output becomes Inf
numerical.wrong_shape("mod.func", (1,512), (1,256))

# Timing faults
timing.timeout("mod.func")              # raise TimeoutError
timing.slow("mod.func", delay=2.0)      # add delay
timing.intermittent_crash("mod.func", every_n=3)
timing.jitter("mod.func", magnitude=0.01)

# Network faults
network.http_error("mod.client.post", status_code=503)
network.connection_reset("mod.client.post")
network.rate_limited("mod.client.get", retry_after=60)
network.dns_failure("mod.client.resolve")

# Concurrency faults
concurrency.contended_call("mod.pool.acquire", contention=0.1)
concurrency.thread_boundary("mod.cache.get")
concurrency.stale_state(service, "config", old_config)

Integrations

# Atheris — coverage-guided fuzzing steers buggify() decisions
from ordeal.integrations.atheris_engine import fuzz
fuzz(my_function, max_time=60)

# API chaos testing (built-in, no extra install)
from ordeal.integrations.openapi import chaos_api_test
chaos_api_test("http://localhost:8080/openapi.json", faults=[...])
chaos_api_test("http://localhost:8080/openapi.json", faults=[...], stateful=True)

Audit — justify adoption with data

ordeal audit myapp.scoring --test-dir tests/
ordeal audit myapp.scoring --validation-mode deep

myapp.scoring
  current suite:         33 tests | 343 lines | 98% coverage [verified]
  generated incremental: 12 tests | 130 lines | 100% coverage [verified]
  mined:    compute: output in [0, 1] (500/500, >=99% CI)
  mutation: 14/18 (78%)
  protection: WEAK: 100% line coverage but 4/18 mutation(s) survived
  suggest:
    - L42 in compute(): test when x < 0
    - L67 in normalize(): test that ValueError is raised

Every number is either [verified] (measured and cross-checked) or FAILED: reason — the audit never silently returns 0%. The protection line is mutation-first: a surviving mutant keeps the verdict weak even at 100% line coverage. The verdict applies to the generated/migrated checks; use ordeal mutate to measure the selected existing pytest tests directly.

Read Are your tests meaningful? for the plain-language model, then use the Test Protection Guide or CI Guide for the full workflow.

CLI

ordeal audit myapp.scoring              # compare existing tests vs ordeal
ordeal migrate oldpkg newpkg -c ordeal.toml  # ordered replacement validation
ordeal explore                          # run from ordeal.toml
ordeal explore -w 8                     # parallel with 8 workers
ordeal explore -c ci.toml -v            # custom config, verbose
ordeal explore --max-time 300 --seed 99 # override settings
ordeal explore --runner compose        # persistent services + real worker faults
ordeal replay trace.json                # reproduce a failure
ordeal replay --shrink trace.json       # minimize a failure trace
ordeal explore --generate-tests tests/test_gen.py  # turn traces into pytest tests

Find what you need

Every goal maps to a starting point — a command to run, a module to import, and a page to read. Nothing is hidden.

New to ordeal? Start with ordeal mine ordeal.demo to see it in action, then read Getting Started. Have existing tests? Run ordeal audit mymodule --test-dir tests/ to see how they compare. Want the full picture? Browse the full documentation.**

Architecture — code map

Every module does one thing. When you want to understand, use, or extend ordeal, this tells you where to look.

ordeal/
├── chaos.py           Your tests extend this — ChaosTest base class, nemesis, swarm mode
├── explore.py         The exploration engine — coverage tracking, checkpoints, energy scheduling
├── assertions.py      always / sometimes / reachable / unreachable — the assertion model
├── buggify.py         Inline fault gates — thread-local, seed-controlled, no-op when inactive
├── quickcheck.py      @quickcheck — type-driven strategies with boundary bias
├── simulate.py        Deterministic Clock and FileSystem — no mocks, no real I/O
├── invariants.py      Composable checks — no_nan & bounded(0, 1), works with numpy
├── mutations.py       AST mutation testing — 14 operators, count-and-apply pattern
├── auto.py            Auto-testing — scan_module, fuzz, mine, diff, chaos_for
├── finding_evidence.py Bounded claims, witness hashes, replay and fix controls
├── regression_evidence.py AST bindings for generated witness regressions
├── trace.py           Trace recording, JSON serialization, replay, delta-debugging shrink
├── compose.py         Long-lived Compose services, faults, state, exact trace and replay counts
├── system_diff.py     Stateful refactor parity, fault/recovery replay, performance budgets
├── migration.py       Ordered audit/mine/diff/regression/mutation/scan migration gate
├── config.py          ordeal.toml loader — strict validation
├── cli.py             CLI entry point — explore, replay, mine, audit
├── plugin.py          Pytest plugin — --chaos, --chaos-seed, --buggify-prob
├── strategies.py      Adversarial Hypothesis strategies for fuzzing
├── faults/            All fault types, organized by what can go wrong:
│   ├── io.py              disk_full, corrupt_output, permission_denied
│   ├── numerical.py       nan_injection, inf_injection, wrong_shape
│   ├── timing.py          timeout, slow, intermittent_crash, jitter
│   ├── network.py         http_error, connection_reset, dns_failure
│   └── concurrency.py     contended_call, thread_boundary, stale_state
└── integrations/      Optional bridges to specialized tools:
    ├── openapi.py         Built-in API chaos testing (no extra deps)
    └── atheris_engine.py  Coverage-guided fuzzing (pip install ordeal[atheris])

Want to add a new fault? Look at any function in ordeal/faults/ — they all follow the same pattern: take a dotted target path, return a PatchFault or LambdaFault. Adding a new fault type means adding a new function that follows this pattern.

Want to understand how something works? Every module is self-contained. Read the module you're interested in — the code matches the documentation.

License

Apache 2.0