React applications
Build agent conversations and finite-run interfaces with the public Fabric client protocol.
@fabric-harness/react provides FabricProvider, useFabricClient, useFabricAgent, and
useFabricJob. The package talks only to the public authenticated HTTP protocol, so the same UI can
target local development, a Node artifact, Databricks Apps, or another compatible deployment.
pnpm add @fabric-harness/react @fabric-harness/client reactConfigure the provider
Create one client for the application. Authentication headers stay in the application transport and are not placed in agent messages or model context.
import { FabricProvider } from '@fabric-harness/react';
export function Root({ children }: { children: React.ReactNode }) {
return (
<FabricProvider options={{
baseUrl: '/api',
headers: { authorization: `Bearer ${window.fabricToken}` },
}}>
{children}
</FabricProvider>
);
}For server-rendered frameworks, pass a request-scoped client through client. Hooks use stable
external-store server snapshots and start network observation only after the component mounts.
Agent conversations
useFabricAgent catches up from offset zero, streams text, reasoning, and tool-argument deltas over
SSE, replaces transient parts with the canonical durable assistant message, deduplicates reconnect
replay, resets stale projections when the stream incarnation changes, and applies truncation
records. Render the stable messages projection; transcript remains available
for diagnostics and compatibility. Optimistic messages remain separate in pending until their
submission-correlated input is observed.
isWorking and activeSubmissionIds track admitted submissions and observed durable work until
submission_settled, independently of status (the connection/send lifecycle). Replayed entries
rebuild this projection. They do not discover queued work that has not appeared in this client's
receipt or conversation stream; use the server submission status interface for that case.
import { useFabricAgent } from '@fabric-harness/react';
function SupportChat({ customerId }: { customerId: string }) {
const chat = useFabricAgent({ agent: 'support', id: customerId, session: 'default' });
async function send(text: string) {
const submissionId = await chat.send(text);
const settled = await chat.wait(submissionId, { timeoutMs: 30_000 });
console.log(settled.outcome);
}
return (
<>
{chat.messages
.filter((message) => message.display === 'visible')
.map((message) => (
<p key={message.id}>
{message.parts
.filter((part) => part.type === 'text')
.map((part) => part.text)
.join('')}
</p>
))}
{chat.pending.map((message) => (
<button key={message.id} disabled={message.status !== 'failed'} onClick={() => chat.retry(message.id)}>
{message.status}: retry
</button>
))}
<button onClick={() => send('Check account status')}>Send</button>
<button onClick={() => chat.refresh()}>Refresh</button>
<button onClick={() => chat.abort()}>Abort</button>
</>
);
}While a tool call is still streaming, its message part has
{ type: 'dynamic-tool', state: 'input-streaming', inputText }. Treat inputText as untrusted,
partial JSON; only input-available contains the provider's completed parsed input.
Send image attachments with the normal delivered-message shape.
await chat.send({
kind: 'user',
body: 'Inspect this screenshot',
attachments: [{ type: 'image', mimeType: 'image/png', data: base64 }],
});Retry-safe agent admission
sendMessage(message, options) returns the complete SendReceipt immediately after
durable admission: submissionId, uid, acceptedAt, conversation URLs, offset, and
deduplicated when a key was supplied. It doesn't wait for generation. Existing
send() still returns only the submission ID.
// Create and persist this key once per logical user action, before the first attempt.
const key = crypto.randomUUID();
const receipt = await chat.sendMessage('Check account status', {
idempotencyKey: key,
signal: controller.signal,
});
const settled = await chat.wait(receipt.submissionId, { timeoutMs: 30_000 });Keys contain 1–256 visible ASCII characters; use opaque random IDs, never account numbers or secrets. Reuse the same key and unchanged message, creation data, and incarnation precondition after an uncertain network outcome. The server scopes the key to the authenticated actor, tenant, agent, instance, and named session. Exact retries return the original submission and pre-admission offset; conflicting reuse returns HTTP 409. Current authentication/instance checks still run on replay. Keys are hashed before durable storage. Deduplication lasts while the submission record is retained; instance deletion or store cleanup ends that protection.
retry() retains the original message/options, including the key, but starts a fresh
HTTP attempt without the previous AbortSignal. Caller mutation cannot change this
saved request. Keys are opt-in: without one, retry can admit duplicate work. Persist
your key outside the hook if retries must survive a page reload. Cancelling admission
HTTP does not cancel work that the server already accepted; use status/wait and
the explicit agent abort operation to resolve the outcome.
Call dismissFailed(id) when abandoning one failed message, or dismissFailed() for
all failures, to release retained messages and retry data. Cache eviction also clears
private retry payloads; retrying an evicted store fails closed and requires the original
request/key from application storage.
Use the matching current Node or Cloudflare runtime for keyed sends. Custom mounted
backends must implement atomic keyed admission and declare
supportsIdempotentAdmission: true; the router otherwise returns 501 for a keyed
request. Older servers may ignore the header, so upgrading only the browser package
does not provide deduplication. Durable restart protection requires durable persistence
(for example SQLite or Postgres), not the dev server's default in-memory store.
Generated Cloudflare routes without a durable runtime return 501 for keyed sends.
The built-in Cloudflare token authenticates a shared service principal. Applications
requiring per-user isolation must use their own authenticated host boundary and pass a
distinct trusted actor to the Durable Object host; sharing a service token doesn't
create separate end-user identities. A conversation metadata read failure stops
admission so the server cannot issue a misleading receipt cursor.
Finite runs
useFabricJob admits an asynchronous run, follows its offset events, publishes the terminal run
envelope, and exposes retry and abort. Starting a newer invocation supersedes the previous
observation, so a late terminal response from the old run cannot replace the new run's state. Use
invokeSync only for short request/response work.
const report = useFabricJob<{ accountId: string }, { url: string }>('account-report');
await report.invoke(
{ accountId: 'account-42' },
{ idempotencyKey: 'account-report:account-42' },
);
console.log(report.status, report.run?.output?.url);
await report.abort();Agent observation reconnects from the latest durable offset with bounded exponential backoff and
resets the delay after receiving an update. Set live: 'poll' in createFabricAgentStore when SSE
is intentionally unavailable; the default is automatic SSE with polling fallback. Stores with no
subscribers or in-flight work are evicted, which keeps long-lived
applications with dynamic agent and job addresses bounded.
Runnable application
examples/react-chat
starts fh dev --mock and Vite with one pnpm dev command. It includes multiple persistent agent
instances, image attachments, optimistic retry, abort, canonical transcript streaming, and a finite
job panel. Its UI renders the public conversation-message projection, including streaming text and
tool states. The example consumes @fabricorg/ui for shared Fabric semantic tokens and operator
controls; this UI package remains an application dependency and is not required by the Harness SDK
or runtime. Its automated browser check captures desktop/mobile light and dark layouts and fails on
browser errors or mobile horizontal overflow.