dsh-resource-guard

Language: English · Česky

DeepSeek Harness plugin that intelligently guards local-model memory. It acts as a semaphore + memory budget so image generation, voice synthesis and local LLM tasks can never exhaust RAM on the same machine — for example, it refuses a new local image job while a local LLM is already loaded.

Features

• resourceGuard service — other plugins call ctx.get("resourceGuard") to acquire()/release() reservations before starting a local model.
• dsh_resource tool — model-callable: status, models, reserve, release.
• Settings section — configure the policy and memory budget from the Harness settings UI, with a live RAM/reservation overview.
• Three policies — auto (one local model within budget), strict (one at a time), off (no guarding).

Install

dsh plugin --profile web add dsh-resource-guard

Then add dsh-resource-guard to dsh.profile.bundles and dependencies in ~/.dsh/profiles/web/package.json, and run dsh plugin --profile web install.

Usage

Ask the agent:

• dsh_resource action=status — RAM, reservations, running Ollama models.
• dsh_resource action=models — list local Ollama models with sizes.
• dsh_resource action=reserve label="image job" estGb=6 — reserve memory.
• dsh_resource action=release reservationId=rsv-1 — free a reservation.

Settings

| Field | Type | Default | Meaning | | --- | --- | --- | --- | | policy | auto|strict|off | auto | Reservation policy. | | memoryBudgetGb | number | 0 | Memory budget; 0 = 85% of total RAM. |

Settings are stored in the resource-guard namespace of the Harness settings document (~/.dsh/settings.yaml).

Architecture

• Host (lib/index.js) — provides the resourceGuard service, registers the dsh_resource tool, publishes a status snapshot into the read-only resource-guard-status namespace.
• Client (lib/client.js) — a settings.section slot showing policy/budget controls and the live status.

License

MIT