metabase/skills/metabase-data-app-actions/SKILL.md

name, description

name	description
metabase-data-app-actions	Use when a Metabase data app needs to trigger a write or mutation — submitting a form, updating a row, deleting an entry, running a saved action, or any "do something" interaction. Covers invoking an existing action via `useAction`, parameter typing, response handling, and the critical post-action refresh of any UI data the action may have changed.

Triggering actions from a Metabase data app

A Metabase action is a saved, parameterized SQL write against the data warehouse — an INSERT, UPDATE, or DELETE. Actions are configured ahead of time on the Metabase instance, with their parameters and permissions already set. A data app's job is to invoke an action with the right parameters when the user does something — clicking a button, submitting a form, confirming a destructive prompt.

The mental model

• Every action the app can run appears in the schema as schema.actions.<actionName>, with type: "query".
• Each action publishes a parameters list. Each parameter has a slug (the key the Data App sends), a jsType ("string" / "number" / "Date" / "boolean" / "unknown"), and an optional required flag.
• Use action.parameters to know which fields to render and submit. The result reports result["rows-affected"]; use it for lightweight confirmation, then refresh the table/question/query data the page already shows.

What's in the schema (and what isn't)

Action entries are only generated when the typed schema includes actions. Before writing action-invoking code, make sure the schema was generated with include-actions=true; with database=<name-or-id>&include-actions=true, Metabase includes the actions for that database only.

Before writing any action-invoking code, enumerate what's available under schema.actions. The schema is your complete catalog of the actions the app can run. Anything not present doesn't exist as far as the Data App is concerned: when the app needs a write the schema lacks, ask the user to create it as a query action without a model (POST /api/action with type: "query" and no model_id; actions created in the Metabase UI belong to a model and are never listed), on a database with actions enabled, then regenerate the schema.

The hook

import { useAction } from "@metabase/embedding-sdk-react/data-app";

const { execute, isExecuting, result, error, reset } = useAction(MyAction);

• Import it from @metabase/embedding-sdk-react/data-app. That entry's useAction is the data-app form: it accepts a defineAction(...) export or null and nothing else. The main entry's useAction is the general SDK hook, which also takes plain objects and raw ids, so it cannot enforce the definition; a data app that imports it from the main entry compiles with an unsynchronized action and fails in production.
• The argument is the defineAction(...) export itself, declared in the app's root-level actions/ directory (one <topic>.action.ts file per topic, beside package.json, never under src/) — pass MyAction, not MyAction.copiedActionId. The hook accepts nothing else in a data app: an inline { action: ... } object, the schema entry, or a spread copy of a definition fails to compile with Property 'definedWithDefineAction' is missing. Fix that by adding the export to actions/, not with a cast and not by calling defineAction(...) at the hook, which compiles but is never synchronized. A production build then runs the synchronized copy in the app's own collection; the dev preview keeps running the authored action, so an app works before it has ever been synchronized. Do not pass schema.actions.<action> or its .id: the entry is a compile error, and the raw id names the authored action, which the app's users cannot read, so the call fails on permissions in production.
• Don't write the generics. The definition carries its schema entry, so the hook infers both the parameters object and the discriminated result from it: parameters[] becomes a keyed object (required: true entries are required keys, each value typed from its jsType), and type: "query" makes the kind "sql". The raw-id form exists only on the SDK's useAction from the main entry, where useAction<TParameters, TKind>(42) needs them spelled out since an id describes nothing; a data app never names an action by id.
• execute(parameters) — triggers the action. Parameters object is keyed by parameter slug; parameters declared required: true are required keys, everything else optional. Returns the response body on success AND throws on failure (the error is also written to error state for render-time consumers). Resolves to null (without making a request) when actionId is null or the SDK is not yet initialized — guard the call site if those cases are reachable.
• No enabled / options argument. The hook only ever runs when execute(...) is called, so a gate option would be redundant. Skip the action by branching in the event handler:

  const onClick = async () => {
    if (!user.canEdit) return;
    await execute({ id: orderId, discount });
  };
  

• isExecuting — true between the call and its resolution. Drive button disabled from this so the user can't double-click into duplicate requests.
• result — the response body, discriminated by TKind (or the AnyActionResult union when TKind is omitted). null before the first call and after reset(). Use it for lightweight confirmation (result?.["rows-affected"]), then refresh surrounding data — see After an action runs.
• error — the last thrown error, typed ActionExecuteError | null. Read fields directly with no cast: error?.data?.message, error?.status, error?.isCancelled.
• reset() — clears result and error back to null. Useful after the user acknowledges success or dismisses an error.

The hook does NOT auto-fire on mount. Actions only run when the Data App calls execute(...) explicitly.

Canonical usage — a form that creates a row

import { useAction } from "@metabase/embedding-sdk-react/data-app";

import { CreatePerson } from "../../actions/people.action";

function AddPersonForm({ onCreated }: { onCreated: () => void }) {
  const { useState } = React;
  const { execute, isExecuting, error, reset } = useAction(CreatePerson);

  const [name, setName] = useState("");
  const [email, setEmail] = useState("");

  async function onSubmit(e: React.FormEvent) {
    e.preventDefault();
    try {
      await execute({ name, email });   // typed: keys match parameter slugs
      setName("");
      setEmail("");
      onCreated();                      // ← let the parent refresh dependent data
    } catch {
      // error is captured into hook state for render-time display
    }
  }

  return (
    <form onSubmit={onSubmit}>
      <input value={name} onChange={(e) => setName(e.target.value)} />
      <input value={email} onChange={(e) => setEmail(e.target.value)} />
      <button type="submit" disabled={isExecuting || !name || !email}>
        {isExecuting ? "Saving…" : "Add"}
      </button>
    </form>
  );
}

Showing the error message

When execute(...) fails, surface both error layers. The hook's error is typed ActionExecuteError | null — shape { status?, data: { message?, errors? }, isCancelled }. error.data.message is the whole-request failure. error.data.errors is a per-field validation map keyed by parameter slug ({ <slug>: <message> }), or {} for whole-request failures — which is every failure of the SQL itself. Read both directly, no cast:

const fieldErrors = error?.data.errors ?? {};

{error ? (
  <pre style={{ whiteSpace: "pre-wrap", margin: 0 }}>
    {error.data.message ?? "Action failed."}
  </pre>
) : null}

{action.parameters.map((parameter) => {
  const fieldError = fieldErrors[parameter.slug];
  return (
    <label key={parameter.slug}>
      {parameter.displayName}
      <input
        aria-invalid={Boolean(fieldError)}
        style={{ borderColor: fieldError ? "#dc2626" : undefined }}
      />
      {fieldError ? <div role="alert">{fieldError}</div> : null}
    </label>
  );
})}

Use <pre> (or any element with white-space: pre-wrap) — the messages contain newlines that matter (driver errors include the SQL on its own line). A <span> collapses them into one wall of text.

Example: submitting a 9-character value into a CHARACTER(2) column produces

Value too long for column "STATE CHARACTER(2)": "'dadasdasd' (9)";
SQL statement:
UPDATE "PUBLIC"."PEOPLE" SET … WHERE "PUBLIC"."PEOPLE"."ID" = 1 [22001-214]

That whole string is error.data.message. Render it as-is.

Don't:

• Render "Failed" / "Something went wrong" / String(error) instead of error.data.message — the user loses the only fix-it info they have.
• Render only error.data.message for validation failures — error.data.errors[parameter.slug] is often the actionable "which field and why" detail.
• Paraphrase the message into a "friendlier" version. The raw H2/Postgres/SQL error is more actionable than any rewrite.

After an action runs — keeping the UI honest

When an action succeeds, the data the user sees may be stale. Without an explicit refresh, the list still shows the old rows; the stat tile still shows the old count; the row the user just edited still shows its old values. The action worked, but the UI lies — and there is no warning.

The rule is simple and absolute: after an action resolves successfully, every piece of data on the screen that the action could have changed must be refreshed.

Mapping parameters

Each parameters[] entry on a schema action exposes slug, displayName, jsType, and optionally required. The Data App's job:

• Object keys in execute({ … }) match the parameter slug strings. Use them as literal string keys — the shape the hook derives from the definition rejects typos at compile time, but only when the argument is the defineAction(...) export (not a bare id).

• Pick the right input type from jsType so the browser provides the correct UX and built-in coercion. The agent never invents the input type from the slug name (a slug called "phone" is still jsType: "string" → type="text").

  jsType	Input element
  "string"	<input type="text" …>
  "number"	<input type="number" …>
  "boolean"	<input type="checkbox" …> (or a Mantine Switch/Checkbox)
  "Date"	<input type="date" …> (or "datetime-local" if a time component is expected)
  "unknown"	<input type="text" …> — best effort, coerce at the call site

• Value types in execute({ … }) match jsType. A jsType: "number" parameter wants a number, not a string. With type="number" the input emits a number for .valueAsNumber, but <input>.value is still a string — coerce at the call site (Number(input)) if you're reading the latter.

• required: true parameters cannot be omitted. Reflected in the TS type: required keys are required.

• displayName is for labels; never use it as a key.

Slugs are whatever the action's author named the SQL parameters in Metabase; the schema is the source of truth.

Form-side validation — match the main app

The schema exposes parameter.required (and jsType for type coercion). That's the full validation contract today, matching Metabase's built-in action-execute form. No length checks, no min/max, no format checks — anything more granular comes back as a BE error after submit (see Showing the error message).

Schema is the ONLY source of truth. Read parameter.required and wire it — nothing else. Don't add required because a field is named "email", don't cap a "name" at 100 chars on a hunch, don't set type="email" from the slug. If the schema is silent, the field is unconstrained.

Disable the submit button while the form is invalid. Drive it from the browser's native verdict so required is the only rule in play:

const [isFormValid, setIsFormValid] = useState(false);
const onFormChange = (e: React.FormEvent<HTMLFormElement>) =>
  setIsFormValid(e.currentTarget.checkValidity());

<form onSubmit={...} onChange={onFormChange}>
  <input required={parameter.required} />
  <button type="submit" disabled={isExecuting || !isFormValid}>Create</button>
</form>

No hand-rolled !name || !email checks.

Debugging Checklist

When an action appears to succeed but the screen doesn't update, or a call fails with a 400:

1. Log result, error, and isExecuting after await execute(...) to confirm the request actually went through and succeeded.
2. Confirm useAction was called with the defineAction(...) export imported from actions/. Passing the schema entry, an inline object, or a spread copy is a compile error; passing its .id compiles but leaves execute untyped and 403s in production, since the authored action is not the one an app's users can run. Calling defineAction(...) inside the component compiles too, but sync-resources scans only actions/, so that action is never copied and also 403s.
3. Log the object passed to execute({ ... }). Every key must match a parameter slug from schema.actions.<action>.parameters; every value must match its declared jsType.
4. List every data view on the screen that reads the rows the action writes. Confirm each one's data hook is mounted ABOVE the action trigger so its refresh callback can be passed down.
5. Confirm the refresh callback is called AFTER await execute(...) AND that the refresh itself is awaited. When multiple refreshes apply, confirm they're awaited together (Promise.all).
6. If the schema doesn't list an action you expect, the action isn't defined on the instance or the schema file is stale. Regenerate the schema.

Common Mistakes

• Importing useAction from @metabase/embedding-sdk-react instead of @metabase/embedding-sdk-react/data-app. The main entry's hook is the general SDK one and accepts anything, so the definition check never runs.
• Declaring the action anywhere but root-level actions/: inline at the hook, wrapped in defineAction(...) inside a component, or under src/actions/. Only actions/ is synchronized.
• Passing schema.actions.<action>.id to useAction instead of the defineAction(...) export. execute then accepts any Record<string, unknown>, typos like { wrongKey: 1 } slip through, and production 403s because the authored action is not the one an app's users can run.
• Forgetting to refresh after a successful action. The UI keeps rendering stale data with no error or warning.
• Rendering "Failed" / "Something went wrong" / String(error) instead of the real backend message. Always extract error.data.message / error.data.errors (the diagnostic the user needs is in there) and render it verbatim — see Showing the error message.
• Calling the refresh callback without awaiting it. The modal dismisses or the form clears before fresh data arrives, leaving the user staring at the stale view for a beat.
• Loading data INSIDE the same component that triggers the action. The hook is below the trigger, so its refresh callback can't be wired up. Lift the data hook to a parent and hand its refresh callback down to the trigger.
• Passing raw <input> strings as parameter values when the slug's jsType is "number" or "boolean". Coerce at the call site.
• Letting the trigger fire while a previous request is in flight. Drive disabled={isExecuting} from the hook.
• Prepending the action's result directly to a local list to skip the refresh. Server-filled defaults (auto-IDs, timestamps, computed columns, derived join columns) diverge from any client-side guess.
• Inventing an action because the user asked for behavior the instance doesn't expose. The schema is the catalog of what exists; surface the gap, don't fake the call.
• Reaching for fetch("/api/action/...") directly. The sandbox blocks raw network calls to the Metabase origin (raw fetch/XHR work only for external allowed_hosts declared in data_app.yaml); the only path to actions is useAction.