Sluice: one queue for every build, test and heavy command your agents run

Описание и применение

Sluice объединяет проверки AI‑агентов в одну очередь, кэширует результаты и управляет памятью, экономя до 28 часов работы и предотвращая OOM‑события.

Задача

Multiple AI agents on a single Linux machine ran expensive checks (tests, linters, type checks) independently, causing uncontrolled concurrency, redundancy, and excessive memory usage, leading to OOM kills and slow pushes.

Результат

Sluice reduced redundant verification runs by 38–52%, saved 28 h of test and linter time in a week, and prevented OOM kills by enforcing memory admission and isolation.

Рабочий процесс

• Install Sluice binary via curl script
• Configure agent to use Sluice MCP server or install Claude Code plugin
• Run sluice init --dry-run to detect hooks, then sluice init to rewrite hooks
• Define gates in .sluice.toml with commands, weight, and hooks
• Agents submit gates to Sluice; Sluice queues, runs each gate once per tree, caches results
• On cache hit, Sluice returns cached result instantly; on miss, runs gate and stores result
• Sluice coalesces concurrent identical requests
• Sluice enforces memory limits and isolates jobs in cgroups

Доказательства

Sluice runs each gate only once per git tree, under a common CPU and memory budget, and stores the result.

Источник

Of all the checks requested, 52% on the development machine (1,426 of 2,742) and 38% on the second machine (1,142 of 3,016) were on code that was already verified.

Источник

Sluice recorded 15 interventions by the kernel or systemd-oomd on agent sessions, and a single gate reached 18 GB.

Источник