@dg-kit/tools

LLM-facing tool definitions for DG-Lab Coyote control.

13 tools, all generating JSON Schema for direct injection into OpenAI / Anthropic / MCP tool-call APIs: five Coyote tools (shock_start, shock_stop, shock_adjust, shock_change_wave, shock_burst), five Opossum tools (vibrate_start, vibrate_stop, vibrate_adjust, vibrate_change_pattern, vibrate_burst), design_wave, set_indicator_color, and timer. Pre-1.9.0 Coyote names (start, stop, adjust_strength, change_wave, burst) still resolve via alias — listDefinitions() just no longer advertises them. Each tool validates its inputs (zod) and returns a ToolExecutionPlan — a DeviceCommand/OpossumCommand for the device tools, an inline string for design_wave after saving, a setIndicatorColor plan, or a TimerCommand for timer.

Rate-limit policy

Tools that hint at "per-turn" caps (shock_adjust, shock_burst, vibrate_adjust, vibrate_burst, design_wave) accept an injectable RateLimitPolicy. Consumers pick the implementation that matches their runtime model:

• DG-Agent (turn-based): createTurnRateLimitPolicy() — counters reset on each new user turn.
• DG-MCP (stateless RPC): createSlidingWindowRateLimitPolicy({ windowMs }) — keeps a timestamp ring and rejects calls that exceed the window cap.
• No-op: createNoOpRateLimitPolicy() — always allows; useful in tests.

import { createDefaultToolRegistry } from '@dg-kit/tools';

const registry = createDefaultToolRegistry({
  waveformLibrary,
  rateLimitPolicy: createSlidingWindowRateLimitPolicy({ windowMs: 5000 }),
});

const tools = await registry.listDefinitions(); // → ToolDefinition[]
const plan = await registry.resolve(toolCall); // → ToolExecutionPlan