Group Chats: Mentions, Debounce & Reply Context

Control when the agent responds in group conversations — mention gating, group allowlists, inbound message debouncing, and reply/quote context.

An agent in a busy group chat shouldn't answer every message. Sophon gates group traffic per channel: you choose a group policy, the gate runs before any session or agent turn is created, and messages that don't pass are dropped silently — the group never sees an error, and the drop is logged with a reason (no mention, not allowlisted, policy disabled) for diagnostics.

This page covers the four pieces of group-chat behavior: group policies, mention triggers, inbound debouncing, and reply/quote context.

Group policies

Each channel carries a groupBehavior setting in its routing configuration:

Policy	Behavior
`mention-only`	Respond only when a mention trigger fires. Default.
`all-messages`	Respond to every group message on this channel
`disabled`	Never respond in groups; DMs are unaffected
`allowlist`	Respond only in groups listed in `allowedGroups`; mention triggers still apply within those groups

Under the allowlist policy, allowedGroups is a list of group/channel IDs. An empty list means no groups are allowed.

Personal-tier instances bypass group gating entirely and respond to all group messages, preserving the original single-user behavior. Pro and Enterprise tiers enforce mention-only by default as a secure-by-default posture.

Configuring policies

Four ways to set the policy and allowlist:

Dashboard channel wizard — step 2 (Configure) when adding a new channel
Edit Routing modal — on the Channels list page for existing channels
CLI — sophon channels routing <channel-id> --group-behavior <policy> --allow <group-id> ...
REST API — PUT /api/channels/{id}/routing with groupBehavior and allowedGroups; partial updates keep current values

Routing changes require channel owner or Admin permission.

{
  "routing": {
    "agentId": "default",
    "dmBehavior": "main-session",
    "groupBehavior": "allowlist",
    "allowedGroups": ["-1001234567890", "C0123ABCDEF"]
  }
}

Mention triggers

Under mention-only (and within allowlisted groups), four triggers wake the agent:

Channel	Explicit @mention	Reply to the bot	Quote of the bot	Thread participation
Telegram	Yes	Yes	Yes	Yes
Discord	Yes	Yes	Yes	Yes
Slack	Yes	Yes	Yes	Yes
Microsoft Teams	Yes	Yes	Yes	Yes
Matrix	Yes	Yes	Yes	Yes
Mattermost	Yes	Yes	Yes	Yes

On platforms where quoting and replying are the same gesture (Telegram, Discord, Matrix), the reply and quote triggers are equivalent. Thread participation means the agent keeps responding inside a thread it has already been pulled into — you mention it once, then converse naturally. See Threads & Sessions for how those threads map to sessions.

Inbound debouncing

People type in bursts. Instead of triggering a separate agent turn for each fragment, Sophon coalesces rapid-fire messages into a single turn: the agent sees the whole burst at once and replies once, with full context.

A buffered burst flushes when the first of three conditions is met — a quiet window elapses after the last message, a hard deadline passes since the first message, or the message count cap is reached.

Key	Default	Description
`Sophon:InboundDebounce:Enabled`	`true`	Global switch for inbound debouncing
`Sophon:InboundDebounce:QuietWindowMs`	`1500`	Milliseconds of silence after the last message before flushing
`Sophon:InboundDebounce:MaxWaitMs`	`10000`	Hard cap from the first buffered message — flush regardless of quiet window
`Sophon:InboundDebounce:MaxMessages`	`10`	Flush immediately once this many messages buffer

Two per-channel overrides live in the channel's routing configuration:

Setting	Default	Description
`debounceEnabled`	`null` (inherit global)	Set `false` to disable debouncing for this channel
`debounceQuietWindowMs`	`null` (inherit global)	Override the quiet window for this channel

Coalescing is keyed per sender and per thread — two people typing simultaneously, or one person active in two threads, never get merged into one turn.

Reply & quote context

When someone replies to or quotes an earlier message, Sophon captures the quoted author and text and injects them ahead of the new message as a fenced context block:

Replying to [Author]: [quoted text]

The agent sees what's being referenced without you re-pasting it. Reply context is also preserved in session history, so it stays visible on replay.

Support	Channels	Notes
In-payload	Telegram, Discord	Quoted content arrives with the message — no extra latency
Fetched server-side	Slack, Mattermost, Matrix	One extra platform API call adds a small amount of latency
Partial	Microsoft Teams	Thread binding shipped; fetching the replied-to message body is on the roadmap
Planned	All other channels	Debouncing works everywhere; reply context capture is on the roadmap

Two behaviors worth knowing:

Quoted content is untrusted. It is wrapped inside the same external-content boundary as any other user-generated input, so instructions hidden in a quoted message do not gain authority. See Prompt Injection Defense.
Fetch failures degrade gracefully. If a server-side lookup fails (network, permissions, rate limit), the message is processed normally without the quote — never blocked.

Where to go next

Threads & Sessions — per-thread isolated sessions and their lifecycle
Channels Overview — supported channels and routing basics
Prompt Injection Defense — how external content is contained

Group Chats: Mentions, Debounce & Reply ContextNEW