[andrewpark-scaleai]

Why

The goal is a slow tier that can run two tests at a time on each runner, which is the most promising remaining speedup on the same 12 runners. Draft #582 tried -n 2 once. Its only failures were three multi_env_computers_test.py tests on shard 3, which failed with Gateway container not running. When the test failed, docker ps showed no gateway at all.

The other worker on that shard had just finished local_parallel_deploy_test.py. Its local_stack_at teardown, which eight other test files share, force-removes every container named agent-env-local-*. Every local sandbox's compose project is named after its work dir, agent-env-<sandbox id>-…, so the teardown removed the other worker's gateway as it deployed.

What changes

While the stack is up, local_stack_at records the id of each LocalSandbox its own process makes. At teardown it removes only the containers named for those ids. Pytest workers are separate processes, so another worker's sandboxes are never recorded.

Verification

• With a stand-in container named for another sandbox, agent-env-local-0therw0r-abc-gateway-1, and one named for a sandbox made inside the stack, the teardown removed the stack's container and kept the other one.
• When tests run one at a time, as on main today, nothing else changes: the stack still removes every sandbox it made.
• ruff check passes.

I'll rerun #582 with this change to check whether any other test collides.

Confidence Score: 5/5

This PR appears safe to merge.

Summary

local_stack_at now records sandbox IDs in the test process and removes only matching Compose containers at teardown.

• Local test stacks remove only containers for sandboxes their process created.

_{Reviews (1) · Last reviewed commit: "test: have the local stack remove only t..." · Reviewed by Greptile}