01 / Flaky-test detective2026
A serverless service that analyzes GitHub Actions history to find tests that pass and fail on the same commit, estimating true rerun costs.
100%
serverless
0
false positives
95%
Wilson interval
Author
my role
Teams know they have flaky tests, but “this test failed before” does not distinguish a flake from a fixed bug. Quarantine ingests JUnit artifacts from GitHub Actions and detects pass/nonpass reversals only when the canonical test, commit SHA, and execution fingerprint exactly match. It then ranks these conflicts by statistically cautious Wilson lower bounds and their estimated CI compute waste.
What I did
- 01Built a full-stack Cloudflare Workers application using Hono and D1 (SQLite).
- 02Implemented impact-first ranking with Wilson lower bounds and explicit insufficient-data states.
- 03Designed an at-least-once queue ingestion pipeline with D1 leases and idempotent artifact processing.
- 04Added PKCE OAuth, identity-bound AES-GCM token encryption, and CSRF checks for secure public repository onboarding.