Adding A Feature

Adding a Feature — The Four-Area Checklist

Rule

Every new feature MUST update all four areas. Skipping any one breaks the agent-native contract — the agent and UI must always be equal partners.

Scope check first: this is the checklist for a feature that adds app data or a new capability. A restyle, a copy change, or a screen with no new data model and no new agent-visible operation only needs the areas it actually touches — do not manufacture actions, state keys, or a skill for it.

Why

Agent-Native apps are defined by parity: everything the UI can do, the agent can do, and vice versa. A feature that only has UI is invisible to the agent. A feature that only has scripts is invisible to the user. A feature without app-state sync means the agent is blind to what the user is doing.

The Checklist

When you add a new feature, work through these four areas in order:

1. UI Component

Build the user-facing interface — a page, component, dialog, or route. Use useActionQuery and useActionMutation from @agent-native/core/client to call actions for data fetching and mutations. Do not create a custom REST endpoint just so React can call action-backed data; the action endpoint already exists.

Auto-refresh on agent writes is non-negotiable — when the agent mutates data, the UI must reflect the change without a manual refresh. There are two paths, and you must pick the right one:

• useActionQuery / useActionMutation — covered automatically. The framework's useDbSync invalidates ["action"] on every change event, so every useActionQuery hook refetches on agent activity. No extra wiring required. Prefer this path.

• Raw useQuery with custom keys — needs explicit wiring. Fold useChangeVersions([<source>, "action"]) from @agent-native/core/client into the queryKey and set placeholderData: (prev) => prev. The action source is the reliable signal (the agent runner emits it after every successful tool call); the resource-specific source ("dashboards", "analyses", "settings", etc.) is bonus when emitted. Without this wiring, agent writes will be invisible until manual refresh — that breaks the framework's #1 promise.

  import { useChangeVersions } from "@agent-native/core/client/hooks";
  import { useQuery } from "@tanstack/react-query";
  
  const v = useChangeVersions(["dashboards", "action"]);
  useQuery({
    queryKey: ["dashboard", id, v],
    queryFn: () => fetchDashboard(id),
    placeholderData: (prev) => prev, // no flicker on refetch
  });
  

  See the real-time-sync skill for the full pattern and source catalog.

2. Action

Create an action in actions/ using defineAction. This serves double duty: the agent calls it as a tool, and the UI calls it through useActionQuery / useActionMutation while the framework owns the HTTP transport. Set http: { method: "GET" } for read actions, leave default for writes, or set http: false for agent-only actions like navigate and view-screen.

Before adding a new route or endpoint, inspect the existing actions. Reuse an action if it already covers the business operation, extend it if the shared contract is incomplete, or create a new defineAction if the agent and UI both need the capability. Do not add pass-through /api/* routes that re-export actions. If client code needs a new framework/app route, expose a named helper or hook first and use that helper from components and docs.

For provider-backed analysis/query/reporting integrations, do not turn every provider endpoint or filter into a rigid action. Prefer the shared provider-api-catalog / provider-api-docs / provider-api-request pattern from @agent-native/core/provider-api, then add narrow convenience actions only for workflows that truly deserve a first-class shortcut. Treat this as a capability requirement, not a nice-to-have: convenience actions must not become the ceiling of what the agent can ask the provider to do. Pair broad provider access with staging or sandboxed code execution when responses may be too large for chat context. If provider credentials live on resource/share rows, add the scoped resolver first so broad access preserves the same ownership boundary as the app UI.

If the feature needs credentials, design the credential path in the same change. Never hardcode API keys, tokens, webhook URLs, signing secrets, private Builder/internal data, or customer data in the action, UI, seed data, fixtures, docs, prompts, or generated extension/app content. Register required secrets, use OAuth helpers, or read scoped values from the vault/credential store.

Before writing custom setup or credential UI, perform the shared-primitive preflight from agent-native-toolkit: inspect the workspace/provider connection catalog first, then settings, secrets/vault, OAuth, onboarding, and provider API support. Reuse an existing connection's app grant and scoped credential resolver; only use app-local vault or OAuth primitives when no reusable connection exists. Keep any wizard thin and provider-specific - it may explain prerequisites, ordering, or readiness, but must not become a second vault or custom credential transport. Do not mark each field required by reflex; model onboarding around the logical connection outcome.

If the feature involves attachments, images, recordings, screenshots, exports, or other file-like payloads, design the upload path in the same change: provider upload first, then URL/id/blob handle in SQL. Do not add base64/binary columns or stuff files into application_state.

If the feature adds a server-side dependency carrying a native binary or heavy runtime — headless browser, ffmpeg, image/video processing, ML runtime — read the performance skill §9 before adding it. A deployed app has one /* page function that every visitor's cache miss wakes, and it ships whatever the server bundle depends on. Run that work from a background function or a job so the page function never carries the weight.

If the action produces or lists a navigable resource, add a link builder that returns { url: buildDeepLink({ app, view, params }), label }. External coding agents and MCP hosts (Claude / ChatGPT / Claude Code / Cowork / Codex, over MCP/A2A) then surface an "Open in … →" deep link that drops the user back into the running UI focused on the record — for free. If a compatible MCP host should render an inline review/edit surface, also add mcpApp with embedApp() so the action embeds the real React app route instead of a one-off HTML UI. The link builder and mcpApp metadata must be pure and synchronous (no I/O). Any external-agent read/ingest action must be http: { method: "GET" } + readOnly: true + publicAgent: { expose: true, readOnly: true, requiresAuth: true }. See the external-agents skill.

Scaffolding a new app of a known type (slides, design, content, forms)? Start from that template's AGENTS.md and .agents/skills/, not a blank app.

3. Skills / Instructions

Update AGENTS.md and/or create a skill in .agents/skills/ if the feature introduces patterns the agent needs to know. At minimum, add the new actions to the action table in the template's AGENTS.md.

Reusable actions are part of the app contract, not just implementation detail. When an action is useful outside one screen, update agent instructions in the same change so app agents know when to call it, which arguments matter, and what output to preserve. If the capability is workflow-heavy, cross-app, provider-backed, or has a non-obvious sequence of actions, add or update a skill instead of burying the behavior in one long AGENTS.md paragraph.

Instruction examples may name secret keys like SLACK_WEBHOOK, but must use placeholders such as ${keys.SLACK_WEBHOOK} or <SLACK_WEBHOOK>. Do not paste real keys, internal data, or customer data into instructions as examples.

If the feature adds or changes visible UI copy, prompts, toasts, labels, empty states, or formatting, update the English source copy. Read the optional internationalization skill and update additional catalogs only when translations.locales in agent-native.config.ts includes them.

For app-backed skills, declare skill visibility in the app-skill manifest:

• internal — only the app's own agents should use it.
• exported — marketplace installs receive it, but the app does not need it loaded internally.
• both — shared between the app's internal agents and exported marketplace bundles.

4. Application State Sync

Expose navigation and selection state so the agent knows what the user is looking at. Write to the navigation app-state key on route changes. Update the view-screen action to fetch relevant data for the new feature. Add a navigate command if the agent needs to open the new view.

Examples

Adding "compose email" to a mail app

Adding "create form" to a forms app

Adding "chart type" to an analytics app

Adding a new route

Templates are single-page apps with client-side routing. The app shell (AgentSidebar + top-level nav) MUST persist across navigation — it is mounted once, either in root.tsx around <Outlet /> or via a pathless _app.tsx layout route that all authed routes nest under.

Never wrap each new route in its own <AppLayout> / <Layout>. That causes React to unmount the entire app shell on every navigation, reloading the agent sidebar and destroying in-progress work.

• If the template has <AppLayout> in root.tsx — just render page content in your new route file, nothing else.
• If the template has app/routes/_app.tsx (pathless layout) — name your new route _app.<segment>.tsx to inherit the shell, or bare <segment>.tsx for public routes that should NOT have the shell.
• If a page needs per-route data (e.g. highlighting the active item in the sidebar), read it in the layout from useParams() / useLocation(). Don't pass it as a prop through every route file.

See the "Client-Side Routing" section in the root CLAUDE.md for full details.

Anti-Patterns

• Per-route <AppLayout> wrappers — Every route file wraps its content in <AppLayout> or <Layout>. React sees a different component at the outlet on each nav and unmounts the whole shell, causing the agent sidebar to reload on every click. Mount the shell once above <Outlet /> (root.tsx or _app.tsx pathless layout).
• UI without actions — The user can create forms but the agent cannot. The agent says "I don't have access to that" when it should be able to do it.
• Actions without AGENTS.md — The actions exist but the agent doesn't know about them because they're not documented. The agent reinvents solutions instead of using the actions.
• Duplicate API routes — Creating /api/ routes for operations that actions already handle, including pass-through routes that just call or repackage an action. Use useActionQuery/useActionMutation instead.
• Raw client route calls — Teaching or adding fetch("/_agent-native/..."), fetch(agentNativePath(...)), or template /api/* calls in components for normal app work. Add a named client helper/hook and call that instead.
• Features without app-state — The agent cannot see that the user is looking at a specific form, email, or chart. It asks "which one?" instead of acting on the current selection.
• Actions without UI — The agent can do something the user cannot. This is less common but still breaks parity.

Verification

After completing all four areas, verify:

1. Can the user perform the operation from the UI?
2. Can the agent perform the same operation via actions?
3. Does pnpm action view-screen show the relevant state when the user is using the feature?
4. Can the agent navigate to the feature view via the navigate action?
5. Is the feature documented in AGENTS.md with action names and args?
6. Are credentials and sensitive data supplied only through approved runtime channels, with no hardcoded real keys, tokens, webhook URLs, Builder/internal data, or customer data?

One more area — sharing

If the feature stores user-authored resources (documents, dashboards, forms, decks, etc.), make them ownable so they get private-by-default semantics and a share dialog for free. See the sharing skill.

TL;DR: spread ownableColumns() into the resource table, pair it with createSharesTable(...), call registerShareableResource(...), wrap list/read queries with accessFilter, guard writes with assertAccess, and drop <ShareButton> in the resource header. The share-resource, unshare-resource, list-resource-shares, and set-resource-visibility actions are auto-mounted framework-wide.

One more area — who inside the app may do it

If the feature is one only some teammates should be able to perform inside this app (an admin-only import, a settings reset), that is a per-app role — not a new organization, and not a share grant. See the authentication skill.

TL;DR: declare the vocabulary once with defineAppRoles({ appId, roles, defaultRole }), guard the action with authorize: appAccess.requireAny("..."), and render the picker with <TeamPage appRoles={descriptor} />. defaultRole is display only — it never satisfies a guard — and authorize gates the operation while accessFilter / assertAccess still scope the rows.

Related Skills

• authentication — Per-app member roles (defineAppRoles) when only some teammates may use a feature
• sharing — How to make a new resource ownable (private by default, share with users/orgs/public)
• context-awareness — How to expose UI state to the agent (area 4 in detail)
• actions — How to create actions with defineAction and the http option (area 2 in detail)
• external-agents — Add a link builder so external agents (MCP/A2A) get an "Open in … →" deep link
• create-skill — How to create skills for new patterns (area 3 in detail)
• storing-data — Where to store the feature's data
• real-time-sync — How the UI stays in sync when the agent writes data
• performance — Query/load cost, and (§9) cold-start artifact size when a feature adds a server-side dependency
• internationalization — How to update localized UI copy and catalogs