Interactive messages and approvals
Render portable Slack and Teams UI, handle actions, and resume interrupted agent runs without blocking a managed delivery.
Channels JSX renders portable components as native channel UI. Use it for buttons, summaries, and human approval without maintaining provider payloads by hand.
On Slack, Channels JSX renders as Block Kit. Button clicks return as
block_actions events through the signed managed webhook.
Use the current Slack manifest
Newly generated Slack manifests enable Interactivity and configure the managed callback URL. If the app came from an older manifest, do these steps before you test buttons:
- Open the Channel in Intelligence.
- Copy the generated manifest from its Slack setup instructions.
- Apply the manifest to the app and reinstall the app.
- Paste the current Bot User OAuth token into Intelligence.
Enable Channels JSX
The package is ESM-only. Keep "type": "module" in package.json and add:
{
"compilerOptions": {
"target": "ES2022",
"module": "NodeNext",
"moduleResolution": "NodeNext",
"jsx": "react-jsx",
"jsxImportSource": "@copilotkit/channels",
"strict": true,
"noEmit": true
},
"include": ["*.ts", "*.tsx"]
}Post a native action
Define a named component and register it on the Channel:
import {
Actions,
Button,
Header,
Message,
Section,
} from "@copilotkit/channels/ui";
export function ApprovalCard({ summary }: { summary: string }) {
return (
<Message>
<Header>Approval required</Header>
<Section>{summary}</Section>
<Actions>
<Button
style="primary"
value={{ approved: true }}
onClick={async ({ thread, message, action }) => {
const value = action.value;
if (value === undefined) {
await thread.post(
"The approval value was missing. Please try again.",
);
return;
}
if (message.ref.id) {
try {
await thread.update(
message.ref,
<Message>
<Header>Approved</Header>
<Section>{summary}</Section>
</Message>,
);
} catch {
// Updating the card is best-effort; always resume the agent.
}
}
await thread.resume(value);
}}
>
Approve
</Button>
<Button
style="danger"
value={{ approved: false }}
onClick={async ({ thread, message, action }) => {
const value = action.value;
if (value === undefined) {
await thread.post(
"The approval value was missing. Please try again.",
);
return;
}
if (message.ref.id) {
try {
await thread.update(
message.ref,
<Message>
<Header>Rejected</Header>
<Section>{summary}</Section>
</Message>,
);
} catch {
// Updating the card is best-effort; always resume the agent.
}
}
await thread.resume(value);
}}
>
Reject
</Button>
</Actions>
</Message>
);
}import { createChannel } from "@copilotkit/channels";
import { ApprovalCard } from "./approval-card.js";
import { makeAgent } from "./agent.js";
function required(name: string): string {
const value = process.env[name];
if (!value) throw new Error(`Missing ${name}`);
return value;
}
const channel = createChannel({
name: required("CHANNEL_CODE"),
identifyUser: "platform",
agent: makeAgent,
components: [ApprovalCard],
});
channel.onInterrupt<{ summary: string }>(
"on_interrupt",
async ({ payload, thread }) => {
await thread.post(<ApprovalCard summary={payload.summary} />);
},
);Registration lets the action registry rebuild a named component when a later click arrives. Keep component props JSON-serializable.
Treat an interaction message as updateable only when message.ref.id is
non-empty. The card update in this example is best-effort: a missing ref or a
failed update must not prevent thread.resume(value) from continuing the
approval.
Use post-and-resume for managed approvals
Managed deliveries cannot wait inside thread.awaitChoice(). The click arrives
as a separate claimed delivery, so blocking the first delivery would prevent the
approval from being processed.
For an agent that emits the SDK-supported on_interrupt event, the safe flow
is:
- The agent emits an
on_interruptevent. channel.onInterruptposts the registered component and returns.- Intelligence acknowledges the original turn.
- A later click updates the card when possible and calls
thread.resume(value). - The SDK re-enters the agent with the approval result.
To check the flow, ask the agent for an action that needs approval. The
Approval required card appears. Click Approve. The card changes to
Approved, and the agent continues with { approved: true }.
The selected agent framework decides how it emits and consumes the interrupt.
Follow its human-in-the-loop guide for the agent-side
interrupt; keep the Channel handler limited to portable presentation and
resume.
Framework interrupt support is not universal
Only a DeepAgent tool-driven approval is tested on the managed path. It emits
on_interrupt, returns from delivery, then resumes from a later interaction.
For every other backend, follow its human-in-the-loop guide. Then run a real
provider test and make sure that:
- the interrupt reaches
onInterrupt - the handler returns
- the later click calls
resume - the agent consumes the value
The default sanitizeAgentEvents behavior repairs a known event-validation
issue. It does not add framework interrupt support.
Restart durability for callbacks
On the managed connection, the SDK stores action snapshots in Intelligence by
default. A click on a card from a registered component still runs after a
restart. Pass createChannel({ store: { adapter } }) only to use your own
backend. See Persistence and scaling.
Capability differences
Text and sections are the portable baseline. Interactive behavior also depends on the selected provider's app configuration:
- Slack can deliver buttons, selects, and inputs through Block Kit. Current generated manifests enable Interactivity; for an older app, open the Channel in Intelligence, apply the generated manifest from its Slack setup instructions, and reinstall it. See Configure the Channel in Intelligence.
Other provider-specific features remain capability-gated:
- Incoming reactions can be handled when Intelligence delivers them, and the managed adapter supports adding or removing documented portable reactions.
- Modal open/submit callbacks are direct-adapter capabilities, not part of the managed Slack or Teams realtime path.
- Ephemeral messages, conversation titles, and suggested prompts may return
{ ok: false }; check the result when your workflow depends on them.
Use the JSX callback reference for handler arguments and the Thread API for post, update, and resume signatures.
How is this guide?