Open-source status page, incident response and uptime monitoring
TypeScript alertingincident-managementmcpmcp-servermonitoringmonitoring-as-codeobservabilityon-callopen-sourcesrestatus-pagestatuspagesynthetic-monitoringterraformtinybirdtursouptimeuptime-checkeruptime-monitor
import type { ServiceContext } from "@openstatus/services";
import {
type AgentToolRegistry,
type AnyAgentTool,
isAgentToolAvailable,
} from "@openstatus/services/agent-tools";
import { type Tool, tool } from "ai";
/**
* Convert a registry entry to the AI SDK's `Tool` shape.
*
* Read tools execute server-side as soon as the model emits them.
*
* Destructive tools opt into the SDK's *approval flow* via
* `needsApproval: true`. The SDK pauses with `state: "approval-requested"`
* before the tool executes; the client renders a confirm card and
* resolves via `addToolApprovalResponse({ id, approved, reason })`. On
* approve the SDK runs `execute` itself on the next stream; on deny
* the part transitions to `output-denied` (a first-class cancel state)
* and `execute` is never invoked. This replaces the older "omit
* `execute` + custom `/api/chat/confirm`" pattern, which forced us to
* reinvent cancel semantics in user-land.
*/
export function toAiSdkTools(
registry: AgentToolRegistry,
ctx: ServiceContext,
): Record<string, Tool> {
const out: Record<string, Tool> = {};
for (const name in registry) {
const t = registry[name];
if (!t || !isAgentToolAvailable(t, ctx.workspace)) continue;
out[name] = toolToAiSdkTool(t, ctx);
}
return out;
}
function toolToAiSdkTool(t: AnyAgentTool, ctx: ServiceContext): Tool {
return tool({
description: t.description,
inputSchema: t.inputSchema,
needsApproval: t.destructive,
execute: async (input: unknown) => {
return t.run({ ctx, input: t.inputSchema.parse(input) });
},
});
}