testo-flaky-tests

name: testo-flaky-tests description: Stabilize flaky Testo tests with #[Retry] or stress-verify with #[Repeat]. Use when the user mentions "flaky test", "intermittent failure", "retry", "rerun", or asks to "verify a fix sticks" by running the test many times. Also use to mark a test flaky for reporting.

Flaky and repeated tests in Testo

Two attributes, two different jobs. Don't mix them up.

Both should be a last resort — first investigate the root cause (shared global state, time/timezone, ordering, randomness, network). Surface this to the user before reaching for #[Retry].

Fetch https://php-testo.github.io/llms.txt for the current attribute namespaces and parameters.

#[Retry] — make a known-flaky test green-ish

use Testo\Retry;
use Testo\Test;

#[Test]
#[Retry(maxAttempts: 3)]
public function pollsExternalService(): void
{
    $response = $this->api->fetch();
    Assert::same($response->status, 200);
}

Constructor (verified against plugin/retry/Retry.php):

public function __construct(
    public int $maxAttempts = 3,
    public bool $markFlaky  = true,   // ← default ON
) {}

• maxAttempts is the total number of attempts (3 = first run + up to 2 retries).
• markFlaky is on by default — when a retry was needed, the run is reported flaky even though it eventually passed. Do not disable this unless the user explicitly asks: silent retries are how flakiness rots a suite.
• Valid targets: method, function, class (TARGET_CLASS is allowed). Apply at the class level only when every test in it is independently flaky for the same external reason — that's rare; usually it's a smell.

#[Repeat] — run a test N times unconditionally

use Testo\Repeat;

#[Test]
#[Repeat(times: 50)]
public function concurrentInsertNeverDeadlocks(): void
{
    $this->runConcurrentInsert();
}

Constructor (verified against plugin/repeat/Repeat.php):

public function __construct(
    public int  $times       = 2,    // total runs, NOT additional repetitions
    public int  $maxFailures = 0,    // failures tolerated before the whole loop fails
    public bool $markFlaky   = true, // ← default ON, reports flaky if any run failed but stayed within maxFailures
) {}

• times is the total number of runs. #[Repeat(times: 3)] runs the test three times (mirrors Kotlin's repeat(n) and JUnit's @RepeatedTest(n)). It is not "additional repetitions on top of one run".
• maxFailures defaults to 0 — any single failure fails the whole loop.
• Combining with #[Retry]: Repeat runs inside Retry — each retry attempt re-runs the full repeat cycle. Possible, but the semantics are subtle; surface it to the user before suggesting both.

Use cases:

• Verifying a fix for a flaky test actually sticks (#[Repeat(times: 100)], run locally, remove before merging).
• Probabilistic / concurrency / randomness tests where one run is not enough evidence.

Don't ship #[Repeat(times: 50)] long-term on a fast suite — CI cost adds up. Remove or scale down once the fix has been validated.

Decision flow

1. Is the test failing intermittently in CI?
   • Yes → investigate root cause first. If genuinely external (DNS, third-party API) → #[Retry(maxAttempts: 3)] (markFlaky is already on by default).
   • No, but I want to verify a fix → #[Repeat(times: N)], run locally, remove before merging.
2. Is the test deterministic but slow / probabilistic by nature (sampling, fuzz)?
   • Use #[Repeat], never #[Retry].
3. Is the flakiness from shared state inside the suite (ordering)?
   • Don't reach for either attribute. Fix isolation (lifecycle hooks, fresh fixtures).

Pitfalls

• Don't disable markFlaky on #[Retry] / #[Repeat] (it defaults to true). Setting markFlaky: false is silent rot — a flaky test that retries to green hides the underlying defect. Only flip it off when the user explicitly asks.
• Repeat(times: N) is total runs, not extra runs. Repeat(times: 1) runs the test once. People coming from older PHPUnit @Repeat semantics expect "additional" — they're wrong here.
• Combining #[Retry] with #[Repeat] is allowed: Repeat runs inside Retry (each retry attempt re-runs the full repeat cycle). Only suggest both when the user genuinely wants that nesting.
• A test with Expect::exception(...) and #[Retry] is almost always wrong — expected exceptions are deterministic by design.
• Don't use retries to paper over network calls in unit tests — replace the dependency with a fake instead.
• Throwing SkipTest / CancelTest from the body short-circuits both #[Retry] and #[Repeat] — the loop stops immediately and the result keeps the Skipped / Cancelled status. That's intentional (skipping isn't a failure to retry against), but worth knowing when a "flaky" test is actually skipping on some runs.