[Slai.life (huong8373tt-beep)]

Problem

Some published Linux benchmark tests depend on public DNS but inherit Docker Engine's resolver configuration unchanged. In the kubernetes__kops-17519 replay, TestNvidia and many unrelated integration tests stopped before their assertions because lookups of example.com through the engine resolver (192.168.65.7) returned i/o timeout or no such host.

The candidate changes NVIDIA configuration, not DNS. The same resolver fault therefore turns an external network condition into a candidate verdict.

A local Docker control establishes that the runtime can use a reviewed explicit resolver when needed:

# Docker default resolver
nameserver 192.168.65.7

# docker run --dns 1.1.1.1 ...
nameserver 1.1.1.1
getent hosts example.com  # succeeds

Hard-coding a resolver for every task would change normal sandbox behavior. RepoLaunch currently lacks a narrow, explicit launch control for the evaluator to select a resolver only for a reviewed network-dependent task.

Change

• Keep Docker Engine DNS behavior unchanged by default.
• Add optional comma-separated SWE_BENCH_LIVE_DOCKER_DNS handling when a Linux task container is created.
• Pass only non-empty, trimmed nameserver strings to Docker's dns argument.

For a reviewed task, an evaluator can opt in explicitly:

SWE_BENCH_LIVE_DOCKER_DNS=1.1.1.1,2606:4700:4700::1111

Validation

python -m pytest -q tests/test_linux_docker_dns.py
# 2 passed

python -m py_compile launch/core/platforms/linux.py tests/test_linux_docker_dns.py
# passed

git diff --check
# passed

Tests cover preserving the default Docker resolver and forwarding a whitespace/empty-entry-separated DNS list.

This is runtime plumbing only. It does not modify benchmark metadata, test patches, candidate patches, or scoring; choosing a resolver remains an explicit evaluator decision.