Integrate Arcjet Guard Google Adk

Integrate Arcjet Guard into Google ADK JS

@arcjet/guard's Google ADK v2 namespace wraps the agent's existing Arcjet client. It never talks to the Arcjet API itself. Two surfaces, one decision rule:

• Tool calls → guardPlugin(). A Runner BasePlugin whose beforeToolCallback is the run-wide gate. DENY is a dictionary (ArcjetDenialResult) so ADK skips runAsync and the model sees the payload. undefined lets the tool execute. Do not throw from the callback — PluginManager treats a throw as a plugin error, not skip. Tools already branded by a sibling guardTool are skipped so Guard is not double-called. Inbound guard() before Runner.runAsync does not brand tools and does not skip this gate. The plugin does not screen inbound (onUserMessageCallback / beforeModelCallback are no-ops) so a preceding guard() does not double-call.
• Correlation → googleAdkContext() reads a caller-owned id from helper options (guardPlugin({ sessionId })) or a wrap (googleAdkContext({ context: appContext })). ADK Context has no nested context field. Durable session state loses to helper options. It never mints a new id. It never reads invocationId (ADK always generates it). It never reads traceId. It never reads toolContext.sessionId / session.id (session auto-ids).

There is no guardTool. Skip is the plugin return, not throw-from-execute. There is no guardInbound, no guardApproval, and this namespace does not use ADK SecurityPlugin as the Arcjet policy gate.

This namespace is Google ADK JS Runner + BasePlugin.beforeToolCallback. Not @google/genai. Not the Python SDK. Do not also wrap with @arcjet/guard/vercel-ai/v7 or @arcjet/guard/claude-managed-agents/v0.

Docs live at docs.arcjet.com/guards/google-adk/. Do not overwrite any other /guards/... slug.

Put Arcjet first in new Runner({ plugins }). PluginManager is first-win; if another plugin (including SecurityPlugin) returns a value first, Guard never runs.

Screen inbound before Runner.runAsync — there is no inbound hook.

There is no first-class inbound deny-dict channel, so there is no guardInbound. Put prompt-injection (and other inbound rules) in the application before runner.runAsync(). Call guard() directly. guard() fails open — callers must check hasFailedOpen(). onUserMessageCallback replaces the user message; it is not this policy gate.

requireConfirmation / requestConfirmation is HITL, not a policy gate.

requireConfirmation / toolContext.requestConfirmation / SecurityPlugin CONFIRM is human-in-the-loop. After a human yes, Guard still runs on the tool call. Same trap as Mastra requireApproval, Claude canUseTool, LangGraph interrupt(), Genkit toolApproval, OpenAI Agents needsApproval, LangChain humanInTheLoopMiddleware, and TanStack needsApproval. There is no guardApproval.

Questions to ask the human first

Ask only what you cannot infer from the code; suggest defaults.

1. Which tools are risky (external side effects, irreversible, spends money, sends messages)? Those are gated by guardPlugin.
2. What limits? (e.g. "10 lookups/min per order" → tokenBucket.)
3. Who is the user for metadata — an opaque user/tenant ID (never PII)? Default: none. Pass it via metadata on the policy. Put the conversation / session id you already have on helper options (guardPlugin({ sessionId })). That id is the correlation id and wins over durable session state. Do not use invocationId or toolContext.sessionId. ADK Context has no nested context field.
4. Is an Arcjet outage unacceptable? Every helper defaults to onGuardError: "deny". Ask explicitly about inbound screening before Runner.runAsync: failing closed there means the run does not start for the duration of the outage, so "allow" is a routine and legitimate choice at that one call site. guard() itself still fails open — check hasFailedOpen().

The six things readers get wrong

1. There is no guardInbound. Screen prompt injection before Runner.runAsync with guard(). Check hasFailedOpen(). onUserMessageCallback is not Guard.
2. requireConfirmation / requestConfirmation / SecurityPlugin CONFIRM is not a policy gate. It is HITL. After a human yes, Guard still runs. Do not use SecurityPlugin as the Arcjet gate.
3. The import path is versioned and there is no alias. @arcjet/guard/google-adk/v2. @arcjet/guard/google-adk does not resolve. Docs are /guards/google-adk/.
4. Correlation is read, never minted. Do not call createAgentContext inside a plugin callback — that generates a second id and splits the Sequence. Put the id you already chose on guardPlugin({ sessionId }). That wins over durable state. Do not read invocationId, traceId, or toolContext.sessionId. Do not expect toolContext.context — ADK Context has no such field.
5. Put Arcjet first. PluginManager is first-win. If another plugin returns a dict first, Guard never runs.
6. Do not add guardTool and do not double-wrap with @arcjet/guard/vercel-ai/v7 or @arcjet/guard/claude-managed-agents/v0. Google ADK JS is not the Vercel AI SDK. Skip is the plugin return.

Step 1: Install and find the guard client

Install @arcjet/guard (required), plus @google/adk (optional peer, needed for @arcjet/guard/google-adk/v2). Always use the versioned path: @arcjet/guard/google-adk/v2 resolves; @arcjet/guard/google-adk throws ERR_PACKAGE_PATH_NOT_EXPORTED. The peer range is >=2 <3. Node 22+ — do not bump Node for this adapter.

npm install @arcjet/guard @google/adk

If the agent has no guard client yet, launch one once at module scope:

import { launchArcjet } from "@arcjet/guard";

export const arcjet = launchArcjet({ key: process.env.ARCJET_KEY! });

Step 2: Gate tool calls — Arcjet first

import { Runner } from "@google/adk";
import { guardPlugin } from "@arcjet/guard/google-adk/v2";
import { tokenBucket, localDetectSensitiveInfo, policyInput } from "@arcjet/guard";

import { arcjet } from "./arcjet.js";

const lookupLimit = tokenBucket({
  refillRate: 10,
  intervalSeconds: 60,
  maxTokens: 10,
});
const detectPii = localDetectSensitiveInfo();

const runner = new Runner({
  appName: "my_app",
  agent,
  sessionService,
  plugins: [
    guardPlugin(arcjet, {
      action: ({ toolName }) => `${toolName}.invoked`,
      actor: conversationId,
      inputs: ({ toolName }) => ({
        tool: policyInput.server.string(toolName),
      }),
      rules: ({ toolName, input }) => {
        const note =
          typeof input === "object" && input !== null && "note" in input
            ? String((input as { note?: unknown }).note ?? "")
            : "";
        return [
          lookupLimit({ key: toolName, requested: 1 }),
          ...(note.length > 0 ? [detectPii(note)] : []),
        ];
      },
      sessionId: conversationId,
    }),
  ],
});

• Omit rules to submit none. The guard call still happens.
• Optional actor and inputs (static, or a resolver over this adapter's native call — parsed input plus trusted runtime/context) are forwarded on the guard call so a remote policy that declares those names can evaluate. Build each input with policyInput.
• On DENY the original runAsync never runs. Delivery is { arcjetDenied: true, reason, message, retryable } — the dict ADK treats as skip.
• Default onGuardError: "deny" blocks the tool if Arcjet is unreachable. A Guard error ALWAYS returns a deny dict, never undefined.
• ALLOW captures outcome: "success" when the policy lets the tool run, not when runAsync finishes. beforeToolCallback cannot wrap the tool; a later tool throw does not flip that capture.
• Tools already branded by a sibling guardTool skip the plugin so Guard is not double-called. This namespace has no guardTool. Inbound guard() before Runner.runAsync does not stamp that brand.

Step 3: Screen inbound before Runner.run

import { detectPromptInjection } from "@arcjet/guard";
import { googleAdkContext } from "@arcjet/guard/google-adk/v2";

import { arcjet } from "./arcjet.js";

const inbound = detectPromptInjection();
const decision = await arcjet.guard({
  label: "message.received",
  rules: [inbound(userText)],
  ...googleAdkContext({ context: { sessionId: conversationId } }),
});

if (decision.conclusion === "DENY") {
  throw new Error("message blocked");
}
if (decision.hasFailedOpen()) {
  throw new Error("inbound screening failed open");
}

for await (const event of runner.runAsync({
  userId,
  sessionId: conversationId,
  newMessage: { parts: [{ text: userText }] },
})) {
  void event;
}

There is no guardInbound. guard() fails open — always check hasFailedOpen(). The plugin does not implement inbound screening, so this call does not double-call Guard.

Step 4: Correlation

Put the id you already have on helper options:

guardPlugin(arcjet, { sessionId: conversationId });

Preference order: caller wrap context.correlationId / context.sessionId / context.conversationId (only for googleAdkContext({ context: appContext }) — ADK Context has no such field), then init.sessionId / init.correlationId (helper options; wins over durable state), then the same keys on session state. If none is a valid 1–256 printable-ASCII string, the call is uncorrelated rather than joined to a generated id nobody has.

A toolContext that has invocationId is ADK's Context envelope, so top-level sessionId on that object is ignored.

Never mint a new id. Never read invocationId (ADK always generates it). Never read traceId. Never read toolContext.sessionId / session.id. requireConfirmation resumes after a human yes — Guard still runs on the tool call. Do not treat the confirmation or its resume value as correlation.

Verify the integration

1. npm run typecheck passes.
2. Exercise inbound PI (before Runner.runAsync, including hasFailedOpen()), a plugin deny-dict skip, undefined execute, first-plugin short-circuit (Arcjet first), no-throw, never-mint, and fail-closed (an unreachable guard → deny dict, never undefined). Confirm the denial is a dict and the run is not a confirmation / HITL pause.
3. Confirm in the Arcjet dashboard that decisions share the caller-owned session id as their correlation id — not invocationId or an ephemeral session id.
4. Manual E2E with a real ARCJET_KEY is still-to-verify until you run it.

A full working demo will land in arcjet/examples google-adk-agent as a later follow-up. Do not add an example under examples/ in the JS SDK repo.

Note: capture events are fire-and-forget and batched, so events can lag the decisions they accompany by a few seconds. A dropped event is diagnosed, never thrown.