Absolute Deflake Agent Skill

> Start your first response with the 🧪 emoji.

Absolute Deflake

Find tests that pass and fail nondeterministically, diagnose the root cause of each,

and fix it — not by retrying or skipping, but by removing the source of nondeterminism.

Output is evidence (failure rate per test) → cause → fix, verified by repeated runs.

Runs the shared engine in references/health-engine.md — read it for the

DETECT → SCAN → TRIAGE → FIX → VERIFY → REPORT loop and the safety contract. This file

covers only what's specific to flaky tests.

When to use

• "Our CI is flaky", "this test fails randomly", "fix the intermittent failures".
• A test passes locally but fails in CI (or vice versa), or fails ~1 in N runs.
• Burning down a backlog of retry/skip-marked tests that mask real flakiness.

Not for tests that fail *deterministically* — that's a real bug or a real regression

(/absolute work for a fix, or just fix it). deflake targets *nondeterministic* failures.

What it scans

Establish flakiness empirically — a test isn't flaky because someone said so. Use

preferences.health.deflakeRuns from config as the default N for repeat-runs (else 20):

| Ecosystem | Repeat-run / detect |

|---|---|

| Jest/Vitest | run suite N× (--run loop), randomize order (--shuffle / testSequencer) |

| pytest | pytest-randomly + pytest --count=N (pytest-repeat); -p no:randomly to A/B |

| Go | go test -count=N -shuffle=on ./..., -race |

Also mine signals: existing retry/flaky/skip annotations, CI history if reachable, and

run the suite both in isolation and in full/parallel — order- and concurrency-

dependent failures only show one way. Record a failure rate per suspect test.

Common root causes (diagnose, don't guess)

| Cause | Tell | Fix |

|---|---|---|

| Test-order / shared state | passes alone, fails in suite (or vice versa) | isolate state; reset/teardown between tests |

| Time / clock | fails near midnight, DST, or under load | fake timers / inject clock; no real sleep |

| Async race / missing await | fails under parallelism or slow CI | await the actual condition; no fixed timeouts |

| Randomness | fails ~X% with no pattern | seed the RNG; fix the seed in tests |

| Network / external I/O | fails offline or on slow links | mock/stub the boundary |

| Unordered collections | fails on map/set iteration order | sort before asserting |

| Resource leak / port reuse | fails on repeat or parallel runs | unique resources; clean up |

Risk ranking (TRIAGE)

| Wave | Class | Default |

|---|---|---|

| 1 | clear, isolated cause (seed, await, fake clock, sort) | fix now |

| 2 | shared-state / ordering — needs fixture refactor | fix this pass, per test |

| 3 | flakiness pointing at a real product race, not just the test | gated — surface; may be a genuine bug to fix in code |

A flaky test sometimes means the *code* has a race, not the test. Don't "stabilize" the test

into hiding a real concurrency bug — flag wave-3 cases for a real fix.

Fix & verify

• Fix the cause. Then prove it: re-run the test many times (and shuffled / parallel /

with -race) — green once is not deflaked; green across N randomized runs is.

• Remove the retry/skip/flaky annotation that was masking it once the cause is fixed.
• Never "fix" by adding retries, raising timeouts blindly, sleep, or skipping the test —

that hides flakiness, doesn't remove it.

• Re-run the full suite to confirm the fix didn't destabilize neighbors.

Gotchas

• Retry/skip as a fix. Masks the flake, ships the nondeterminism. Forbidden here.
• sleep to dodge a race. Slows the suite and still flakes under load. Await the condition.
• One green run = done. Flakes are probabilistic — verify with many randomized runs.
• Stabilizing a real product race. If the *code* races, fix the code, not just the assertion.
• Ignoring order/parallel dimension. Run isolated *and* in-suite; the bug hides in whichever you skip.

Companion commands

• /absolute upgrade — a flaky suite makes upgrade verification unreliable; deflake first.
• /absolute debt — flaky-test annotations are test debt; this clears them at the root.
• /absolute work — when the flake is a genuine product-code race needing real design.