Filter events and scope memory

The bridge decides which events reach the agent and which conversation history each run sees. See Channel agent for service setup and the full option reference.

Filter incoming events

The default filter (defaultShouldHandleChannelEvent) handles non-bot messages that arrive as direct messages or mention the bot. Bot-authored events and lifecycle events such as edits and reactions never start agent runs. Action events never start new runs — they resume a paused approval or question via agent.resume() (see Approvals and interactions).

Override shouldHandle for product-specific routing:

const service = createChannelAgent({
  channel,
  agent,
  shouldHandle(event) {
    return event.conversation.kind === 'direct' || event.text.startsWith('/ask ')
  },
})

Errors thrown by the filter are reported with the filter stage and the message is skipped.

Conversation memory scope

The bridge chooses the session scope passed to every run; the agent's memory store itself is configured through @anvia/core (see Memory).

Product behavior	createSession value
Each sender has private history in a group	omit it (the default, defaultChannelAgentSession) or use channelConversationUserSession
Everyone in one channel or thread shares history	channelConversationSession
No memory for an event	custom function returning undefined
Custom tenant or user mapping	custom function returning your MemoryScope

The default is sender-isolated: history is scoped by platform, bot account, conversation, thread, and sender. If the executor exposes no memory store, the bridge runs without a session instead of throwing a session-without-memory error.

import { channelConversationSession } from '@anvia/channel-agent'

const service = createChannelAgent({ channel, agent, createSession: channelConversationSession })

channelConversationUserSession() states the default explicitly, and channelConversationKey() returns the stable conversation key the bridge uses for queueing and session IDs.

Continue with

• Prompts and attachments — build multimodal prompts without dropping files.
• Streaming and delivery — placeholders, edits, and long-message splitting.