Register Action

Register Action

Contract

By the end of this skill the consumer has:

• A call to usePilotAction inside a `` subtree.
• A Zod schema for the parameters, precise enough that a malicious or

confused LLM can't produce inputs that break the handler.

• An explicit decision on the mutating flag.
• A handler that either returns a JSON-serializable result (so the

assistant can narrate the outcome) or throws. Keep in mind that the error message gets streamed to the model verbatim.

Iron Law: validate destructive actions with mutating: true

Any action that deletes data, spends money, sends an email, or otherwise produces a side effect the user would want to reverse MUST set mutating: true. The provider intercepts these and shows a window.confirm() prompt before the handler fires (see packages/agentickit/src/components/pilot-provider.tsx lines 376-394). Omitting the flag on a destructive action means the AI can call it silently. If in doubt, set the flag. A spurious confirm dialog costs one click; a wrongful delete costs trust.

Phases

Phase 1: decide the name and shape

Tool names must be stable (the LLM reasons across turns using the name) and snake_case or kebab-case (compatible with every major provider's tool-naming rules, noted in packages/agentickit/src/hooks/use-pilot-action.ts lines 20-22).

add_todo
archive_card
send_invoice
apply_discount

Avoid overly generic names (do_thing, update). Avoid names that collide with auto-generated ones: update_ is reserved for usePilotState setters.

Phase 2: write the parameters schema

Every field the handler needs, with .describe() where the LLM needs disambiguation:

import { z } from "zod";

const params = z.object({
  cardId: z.string().describe("The id of the kanban card to archive."),
  reason: z
    .enum(["done", "wont-fix", "stale"])
    .optional()
    .describe("Why it was archived. Omit when unknown."),
});

The schema is parsed at execution time via action.parameters.parse(toolCall.input) (pilot-provider.tsx line 397). Parse failures are surfaced to the model as an output-error so it can self-correct.

Phase 3: call the hook

import { usePilotAction } from "@hec-ovi/agentickit";
import { z } from "zod";

usePilotAction({
  name: "archive_card",
  description: "Archive a kanban card by its id.",
  parameters: z.object({
    cardId: z.string().describe("The id of the kanban card to archive."),
  }),
  handler: async ({ cardId }) => {
    await api.archive(cardId);
    return { ok: true, cardId };
  },
  mutating: true,
});

The signature (verified against use-pilot-action.ts lines 18-54):

interface UsePilotActionOptions {
  name: string;
  description: string;
  parameters: z.ZodType;
  handler: (params: TParams) => Promise | TResult;
  mutating?: boolean;
}

The hook returns void. Cleanup is handled on unmount; see lines 86-88.

Phase 4: pick mutating

Set mutating: true when any of these is true:

• The handler writes to a persistent store (DB, localStorage, remote API).
• The handler sends an email, SMS, or webhook.
• The handler spends money or incurs cost.
• The handler removes data.
• The handler triggers a workflow the user can't cancel.

Set mutating: false (the default; just omit) when the handler only reads, queries, computes, or derives. Examples: summarize_board, find_card_by_title, count_overdue_todos.

Phase 5: return a narratable result

The assistant sees the handler's return value on its next step (it's serialized into the tool output; see pilot-provider.tsx lines 398-403). Good returns:

return { ok: true, id: created.id };
return { ok: false, reason: "No matching card." };
return { summary: "Archived 3 cards from 'Done'." };

Avoid returning entire objects loaded with secrets or unrelated fields; everything you return goes to the LLM provider. If the action is pure side-effect, return { ok: true } is enough.

Phase 6: handle failures explicitly

Throwing is fine. The provider catches and reports the error text to the model (pilot-provider.tsx lines 404-410). But if you know the common failure, return a structured result instead so the assistant can narrate it usefully:

handler: ({ id }) => {
  const todo = store.find(id);
  if (!todo) return { ok: false, reason: "No todo with that id." };
  // ... proceed
}

Anti-Patterns

• z.any() / z.unknown() on parameters. Parsing becomes a no-op; the AI

can inject arbitrary shapes.

• Re-creating the Zod schema inline on every render without the guard of

use-pilot-action.ts line 61-62. The hook handles it, but it's still cheaper and clearer to hoist the schema.

• Registering a read-only query with mutating: true. Every call now

pops a confirm dialog; users quickly learn to click "allow" reflexively, and the flag loses its signal.

• Registering the same name from two components. The latter wins (see

registerAction duplicate-name warning in pilot-provider.tsx lines 92-100), but you've just created a confusing override.

• Returning the literal undefined. The AI SDK will coerce it, but

downstream providers differ in how they handle undefined tool outputs. Return an object with at least { ok: true }.

Output Format

After registering, report:

• The action name and one-line description.
• The parameter shape (inline Zod or a hoisted schema name).
• mutating: true|false and why.
• Expected return shape.

Tools Used

• Edit the component file to add the usePilotAction call.
• Read packages/agentickit/src/hooks/use-pilot-action.ts to confirm the

options shape.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hec-ovi
• Source: hec-ovi/agentickit
• License: MIT

Install and usage instructions live in the source repository linked above.