What the experiment actually found
Our original packaged implementation passed only two of five acceptance cases. After a worker exception it could repeat a side effect; a crashed process left a stale lock; and a conflicting concurrent request could receive another payload’s cached response.
The corrected package passed all five cases. Five additional tests cover failed writes, corrupt state and input identity. Installation, the tests and the demo passed on Linux and macOS with Python 3.11 and 3.14 in GitHub CI.
Read the method, results and limitations · Inspect the CI run
Who should try it?
Python developers maintaining local or legacy job handlers who want to inspect retry behavior. Start with synthetic or authorized nonproduction data. The five-minute example needs no sign-up, API key or hosted service.
Where it stops
This is single-host POSIX software with trusted local storage. It serializes work within a store. Deleting request history can permit repeated effects. Ambiguous work needs manual reconciliation. It does not prove exactly-once remote effects, withstand every storage failure, or replace a production queue.
Existing tools already cover idempotency. We have not shown an advantage over existing idempotency libraries, durable workflow systems or your current solution, and we have not validated customer demand.
Help us find the useful part
Tell us the workflow you tried, what currently handles retries, what failed, and whether you would use this again. Please keep credentials and customer data out of public reports.
The core is free under MIT. Future integration or support work would need evidence that it creates value. There is no payment requirement or hosted subscription.