Tool approval

Arivie exposes a configurable tool-approval policy through ArivieConfig.limits.requireToolApproval. When a policy is active, the agent pauses before calling a gated tool and emits a suspension that the host app can resolve.

Use cases

• Destructive SQL — require human confirmation before any execute_<source> call in production.
• Regulated environments — gate all tool calls by default, then explicitly allow safe tools like compile_metric.
• Per-user policies — vary approval by user role or permissions.

Policy shapes

import { defineArivie } from "@arivie/core";

defineArivie({
  // ...
  limits: {
    // Gate every tool call
    requireToolApproval: true,

    // Gate only these tools
    requireToolApproval: {
      tools: ["execute_postgres", "mastra_workspace_write_file"],
    },

    // Gate everything except safe tools
    requireToolApproval: {
      exceptTools: ["compile_metric"],
    },

    // Custom predicate: (toolName, args, requestContext?) => boolean
    requireToolApproval: (toolName, _args) => {
      if (toolName === "compile_metric") return false;
      return toolName.startsWith("execute_");
    },
  },
});

The function form receives the tool name, its arguments, and an optional request context, so you can vary policy by role, environment, or request metadata.

How it works

Arivie forwards the policy to Mastra’s native requireToolApproval option on the Agent. When the agent hits a gated tool, Mastra emits a suspension that the host app resolves by approving or denying the call. See the Mastra tool suspension docs for the exact runtime shape.

Other limits

The limits config also enforces hard ceilings:

Field	Effect
rowsPerQuery	Max rows returned per execute_<source> call
queryTimeoutMs	SQL statement timeout
tokensPerRequest	Max tokens per agent turn
tokensPerUserPerMonth	Monthly token budget per user
maxSteps	Max agent steps per turn

These are enforced inside @arivie/agent tools, not advisory.