Because one instance = one world, parallel test workers must not share an instance. Each worker gets its own. The CLI provisions them; your plugin assigns them.

​
Boot N instances

prod-break run deck --workers 4
This boots 4 instances, prints 4 url/key pairs, and exposes them via two ordered, comma-separated env vars: 
PRODBREAK_WORKER_URLS=http://localhost:8801,http://localhost:8802,http://localhost:8803,http://localhost:8804
PRODBREAK_WORKER_KEYS=pbw_aaa…,pbw_bbb…,pbw_ccc…,pbw_ddd…

​
The plugin assigns each worker its pair

You don’t wire this up — the plugin reads the env vars and gives worker i the i-th pair on first use, keyed off the framework’s worker id:

Jest

JEST_WORKER_ID (1-based) → the (i−1)-th pair.

Vitest

VITEST_WORKER_ID / VITEST_POOL_ID → the matching pair.

pytest

pytest-xdist worker gw0, gw1… → the matching pair.
reset() still runs between tests within a worker, so tests stay independent at two levels: a private world per worker, and a clean slate per test.

​
Single-worker runs

Sequential runs need no env CSVs — just point at one instance: 
PRODBREAK_URL=http://localhost:8801
PRODBREAK_KEY=pbw_3f9a…
N workers means N processes — heavier than a single shared server, but it’s what guarantees true isolation today. Lifecycle (evicting a stuck worker) is the CLI’s job; a crashed worker doesn’t auto-recover mid-run. The lighter “many worlds per instance” model is on the roadmap.