18 KiB
app — Development Guide
This guide is for development-mode agents editing this app's source code. For app operations and tools, see AGENTS.md.
Tech Stack
- Framework: @agent-native/core + React Router v8 (framework mode)
- Frontend: React 19, Vite, TailwindCSS, shadcn/ui
- Routing: File-based via
flatRoutes()— SSR shell + client rendering - Backend: Nitro (via @agent-native/core) — file-based API routing, server plugins, deploy-anywhere presets
- State: SQL-backed (SSE for real-time updates)
Commands
- Dev:
pnpm dev(Vite dev server with both React Router + Nitro plugins) - Build:
pnpm build(React Router build — client + SSR + Nitro server) - Start:
node .output/server/index.mjs(production)
Directory Structure
app/ # React frontend
root.tsx # HTML shell + global providers
entry.client.tsx # Client hydration entry
routes.ts # Route config — flatRoutes()
routes/ # File-based page routes (auto-discovered)
_index.tsx # / (chat page)
components/ # UI components
hooks/ # React hooks
lib/ # Utilities (cn, etc)
server/ # Nitro API server
routes/
api/ # Route-only endpoints (uploads, webhooks, OAuth, streaming)
[...page].get.ts # SSR catch-all (delegates to React Router)
plugins/ # Server plugins (startup logic)
lib/ # Shared server modules
shared/ # Isomorphic code (imported by both client & server)
actions/ # Shared app operations (defineAction; UI uses action hooks)
run.ts # Script dispatcher
*.ts # Individual actions (pnpm action <name>)
data/pglite/ # Local PGlite data directory
react-router.config.ts # React Router framework config
.agents/skills/ # Agent skills — detailed guidance for each rule
Framework Basics
SSR-first framework, CSR-by-default content: This app uses React Router v8 framework mode with ssr: true. But virtually every route renders only an SSR shell (loading spinner + meta tags). Normal app data fetching happens on the client via action hooks. Server-side data fetching is the exception — only used for public pages that need SEO/OG tags.
Chat-First Shape
Agent-Native is the application framework and execution platform. AgentKit is its agent interaction and experience layer. Toolkit supplies the semantic composer, design-system primitives, and workspace UI used around the conversation.
The / route is the app's primary AgentKit surface. Its default integration is:
createAgentNativeAgentKitTransport()from@agent-native/toolkit/app/chat/agentkit-chatfor the production Agent-Native runtime.CoreAgentKitRootfrom@agent-native/toolkit/app/chat/agentkit-chatfor one managed controller, thread context, and the default action-widget renderer.AgentKitChatfor the reference transcript, composer, queue, approvals, activities, and suggestions.CoreComposerRuntimeProviderfor Agent-Native composer capabilities.
The surrounding Core app shell owns the full-height canvas, global navigation,
durable thread routing, and route-level chrome. AgentKitChat owns the
conversation column. Pass route controls through its toolbar prop. Use
slots and registry on AgentKitRoot for presentation overrides without
forking conversation state.
Action UI is opt-in on each action through chatUI.renderer. CoreAgentKitRoot
provides Core's AgentKitActionWidget in slots.widget; it resolves the
declared React renderer from the stored tool input and result, then keeps the
widget with its assistant message after activity rollup and thread reload. App
renderers can register with registerActionChatRenderer() from
@agent-native/toolkit/app/chat/agentkit-chat using the same renderer id declared by
the action. Built-in Core renderers need no app registration.
This template is also the canonical AgentKit reference surface. Develop and
verify shared chat UX here first, then use domain apps such as Dispatch as
consumer smoke tests. Queue behavior can be exercised by submitting follow-ups
while a run is active. The shared composer remains mounted throughout these
states. The route does not provide fallback suggestions: contextual next
actions appear only when the agent injects them through AgentKit's structured
suggestions input.
For a headless app that later needs UI, this template is the intended landing
zone: bring the existing actions over, keep their names stable, and let the chat
call them before adding extra screens. For a custom agent backend, keep the app
shell and implement AgentTransport, or adapt an existing Core
AgentChatRuntime with createAgentKitProtocolAdapter() from
@agent-native/core/client/agent-chat.
Package and scaffold path
The Chat template depends on @agent-native/core,
@agent-native/agentkit, and @agent-native/toolkit through workspace
specifiers in this repository. The CLI resolves them for the generated shape:
- A normal standalone scaffold receives published package ranges.
- A local framework-development scaffold packages the local Core, AgentKit, and Toolkit implementations. Missing compiled package exports are built before the tarballs are created.
- A workspacified Chat app inherits a concrete package version already pinned by the workspace root. Otherwise it uses the compatible range supplied by the running CLI.
Do not replace these dependencies with direct dist paths or deep source
imports. Install the generated manifest as written.
Migrate an older Core chat surface
Core's AssistantChat remains supported as the AgentKit-backed surface, as do
AgentPanel, AgentSidebar, AgentChatSurface, AgentChatHome, and
MultiTabAssistantChat. Apps that only mount these components do not need to
replace their chat UI.
If your app used the removed AssistantChat.createAdapter, adapter helpers, or
assistant-ui transcript/runtime APIs, follow the version-matched guide at
node_modules/@agent-native/core/docs/migrations/agentkit-chat.md. It maps each
removed API to runtime, createTransport, AgentKit slots, or a custom
AgentTransport. Keep one conversation controller and stream owner.
Widgets invoke stable action ids through the transport. Map them to the same
defineAction surface used by the app and agent. Smart objects and client
effects request host navigation or context changes. They do not bypass action
validation, application-state helpers, request context, or ownable-data access
checks.
Adding a Page
Create a file in app/routes/. The filename determines the URL path:
app/routes/_index.tsx → /
app/routes/settings.tsx → /settings
app/routes/inbox.tsx → /inbox
app/routes/inbox.$threadId.tsx → /inbox/:threadId
app/routes/$id.tsx → /:id (dynamic param)
Each route file exports a default component, optional meta(), and optional HydrateFallback():
import MyPage from "@/pages/MyPage";
export function meta() {
return [{ title: "My Page" }];
}
export function HydrateFallback() {
return (
<div className="flex items-center justify-center h-screen">
<div className="animate-spin rounded-full h-8 w-8 border-b-2 border-foreground" />
</div>
);
}
export default function MyPageRoute() {
return <MyPage />;
}
Do NOT fetch data server-side in route loaders unless the page genuinely needs SEO/OG content. The standard pattern is: SSR renders the shell, client hydrates, and React reads/writes normal app data through actions with useActionQuery / useActionMutation.
Adding App Data
Normal app data starts as an action, not a custom route. Add actions/<verb>-<resource>.ts with defineAction, mark reads with http: { method: "GET" }, and call reads/writes from React with useActionQuery / useActionMutation from @agent-native/core/client. This keeps the UI and agent on one contract and lets mutating actions refresh action-backed queries automatically.
Adding a Route-Only Endpoint
Use server/routes/api/ only for protocols that cannot be modeled as JSON actions: multipart uploads, streaming/SSE/WebSocket, webhooks, OAuth callbacks/redirects, public SEO/OG endpoints, or binary/static asset serving. Do not add /api/* routes for normal CRUD, data queries, or pass-through wrappers around actions; the action endpoint already exists at /_agent-native/actions/:name.
Each route-only endpoint still exports a default defineEventHandler, but keep shared app logic in actions or server libraries so agent and UI behavior do not fork.
Adding a Server Plugin
Startup logic (auth, SSE, etc.) lives in server/plugins/. Use defineNitroPlugin from core:
import { defineNitroPlugin } from "@agent-native/core";
export default defineNitroPlugin(async (nitroApp) => {
// Runs once at server startup
});
Key Imports from @agent-native/core
| Import | Purpose |
|---|---|
defineNitroPlugin |
Define a server plugin (re-exported from Nitro) |
createDefaultSSEHandler |
Create SSE endpoint for DB change events (server) |
readAppState, writeAppState |
Read/write application state (from @agent-native/core/application-state) |
readSetting, writeSetting |
Read/write settings (from @agent-native/core/settings) |
defineEventHandler, readBody, getQuery |
H3 route handler utilities (re-exported) |
sendToAgentChat |
Send messages to agent from UI (client-side) |
agentChat |
Send messages to agent from scripts (server-side) |
Adding an Action
Create actions/<verb>-<resource>.ts with defineAction. Run with pnpm action <name> --id value; React callers should use useActionQuery for GET actions and useActionMutation for mutating actions, not a matching /api/* wrapper.
Sending to agent chat from UI:
import { sendToAgentChat } from "@agent-native/core/client";
sendToAgentChat({
message: "Generate something",
context: "...",
submit: true,
});
Sending to agent chat from scripts:
import { agentChat } from "@agent-native/core";
agentChat.submit("Generate something");
Database & Environment Variables
Local development uses PGlite at data/pglite. For production and shared environments, set DATABASE_URL to a persistent hosted PostgreSQL database.
Real credential values belong only in local .env files, deployment configuration, or registered secrets/settings UI. Never commit, document, log, return, paste, or include real keys, tokens, webhook URLs, signing secrets, or private data in examples; use empty values or obvious placeholders.
When adding app data, define tables with Drizzle's drizzle-orm/pg-core
exports and use Drizzle's query builder for reads/writes. Import framework
sharing helpers such as ownableColumns() and createSharesTable() from
@agent-native/core/db/schema. Keep SQL PostgreSQL-compatible and reserve raw
SQL for additive migrations, health checks, or carefully scoped maintenance.
| Variable | Required | Description |
|---|---|---|
DATABASE_URL |
Production yes, local dev no | PostgreSQL or PGlite database URL (local dev default: pglite:./data/pglite) |
AUTH_DISABLED |
Optional | Set to true or 1 to skip login/signup (local dev/preview only) |
Extensions (Framework Feature)
The framework provides Extensions — mini sandboxed Alpine.js apps that run inside iframes. Extensions let users (or the agent) create interactive widgets, dashboards, and utilities without modifying the app's source code. They appear in the sidebar under an "Extensions" section. (Distinct from LLM tools — the function-calling primitives the agent invokes.)
- Creating extensions: Via the sidebar "+" button, agent chat, or
POST /_agent-native/extensions - API calls: Extensions use
extensionFetch()(legacy aliastoolFetch) which proxies requests through the server with${keys.NAME}secret injection - Styling: Extensions inherit the main app's Tailwind v4 theme automatically
- Sharing: Private by default, shareable with org or specific users (same model as other ownable resources)
- Security: Iframe sandbox + CSP + SSRF protection on the proxy
See the extensions skill in .agents/skills/extensions/SKILL.md for full implementation details.
Preserve existing work
This starter ships as a blank Agent-Native app canvas — that describes its
initial state, not necessarily its current one. Before assuming no UI or
brand exists, check app/routes/_index.tsx and app/global.css: if they
already contain real content, that content is the current product and its
established brand. Build additively, preserve existing tokens/routes/palette,
and do not re-derive a new visual direction or overwrite shipped UI unless
the user explicitly asks for a redesign. Only treat the canvas as blank when
the files actually show the starter's placeholder content.
If DECISIONS.md exists at the repo root, read it before starting work — it
holds pre-build choices as current requirements. If the user changes one
later in chat, update DECISIONS.md to match.
README.md starts as a generic "My App" placeholder. Once the app has real
features, rewrite its title and opening paragraph to name the app and say what
it does, and keep the run instructions accurate when you change scripts or
required env vars.
Do not add internationalization or changelog support unless the user
explicitly asks for them. This starter ships English-only UI copy inline —
no app/i18n/, LanguagePicker, CHANGELOG.md, or What's New surfaces. If the
user requests i18n or changelogs, load the matching skill and add only what
they asked for.
Data & actions (read these first)
Add persistence or auth only when data must survive reload or be shared
between users. A pure UI, copy, or layout change needs no schema, action, or
auth — build it directly and do not read the security or storing-data
skills for UI-only work. Only when you actually add an action, route, or schema
that handles user input or persistence, read security and storing-data
first.
Choose the post-auth landing route from the app
Before enabling authentication, inspect app/routes, the app's navigation,
and any existing app.homePath in server/plugins/config.ts. Preserve an
existing homePath when it still points to a valid route. Otherwise choose
the actual primary authenticated landing page from the product already built,
such as /, /dashboard, or /tasks. Never assume /home unless that
route really exists.
The starter ships homePath: "/" only so sign-in lands on the blank canvas.
Once / is a public page and the signed-in app lives elsewhere, replace it
with that route.
Do not choose an auth route, API route, public-only marketing page, wildcard route, or parameterized route that cannot open without an id. If multiple authenticated landing routes are equally plausible and the product intent is unclear, ask the user instead of inventing one.
Write the selected route to the existing defineAppConfig object in
server/plugins/config.ts while preserving its plugins.disabled settings:
export default defineAppConfig({
app: { homePath: "/dashboard" },
plugins: {
disabled: ["integrations", "observational-memory", "sentry", "terminal"],
},
});
After enabling or changing auth, verify both signup and login land on an existing page rather than a 404.
When adding SQL-backed features, do not start with find / cat over
node_modules. Read these two files first:
drizzle/START_HERE.md— table map, migrate commands, path mapdrizzle/crud-action-example.ts— copy-paste list/create/update/delete
Then use getDb / schema from server/db.ts. After a batch of related
schema/action edits: one smoke test, one pnpm typecheck (see Verification).
- Guarded verification: run
pnpm agent-native:doctor; fix findings before done.
Verification
Run checks once at the end of a batch of related edits, not after every
file, action, or small UI tweak. Dev already runs route/action typegen while
pnpm dev is up — do not treat that as a reason to also run full typecheck
after each write.
| Change shape | Verify with |
|---|---|
| UI / copy / layout only | Formatter if the app has one; preview if something looks wrong. Skip full typecheck unless the edit touched types or imports. |
| New/changed actions, schema, server, or shared types | One pnpm typecheck (and lint if the app has it) after the batch. |
| New DB-backed CRUD | One smoke path only (e.g. create + list via pnpm action … or a single HTTP call). Do not CLI-test every action method. |
Do not re-run typecheck to "confirm" after a clean pass. If typecheck fails on unrelated pre-existing errors, fix or note them — do not thrash with repeated full runs and greps.
Configuration is code, not env vars
Configure framework behavior in agent-native.config.ts via
defineAgentNativeConfig({ ... }). App metadata consumed by getAppConfig(),
including app.homePath, belongs in the existing defineAppConfig object in
server/plugins/config.ts. Do not reach for process.env to drive app
behavior, feature flags, or framework options. Environment variables are only
for deploy-level secrets and host settings (see secrets); never add a
process.env fallback to configure a feature.