Moderation

Cariosan moderates message content automatically — no manual report/review queue. Filtering is non-AI (wordlists + patterns + counters), runs in the normal server at zero extra cost, and is off by default (a workspace opts in). AI toxicity classification is a separate, optional upgrade on the roadmap; it never gates the core.

How it works

Every send is evaluated before it persists. If a rule matches, the workspace's configured action is applied:

A clean message, or a workspace that hasn't enabled moderation, passes straight through unchanged.

Rules

• profanity — a built-in Indonesian + English wordlist, matched as whole words (so "pantai" never trips "tai"). Extend it per workspace with a custom blocklist.
• link — explicit URLs (http(s)://, www.).
• phone — contact-number leakage (Indonesian 08…/+62… mobiles and generic long digit runs) — the classic marketplace "DM me at 08…" pattern.
• flood — the same message repeated by one user in a channel more than N times within a short window.

profanity and flood are on by default once moderation is enabled; link and phone are opt-in (they're higher-false-positive and use-case dependent).

Blocking & muting

Two complementary controls sit alongside the automatic filter:

• Block — an end-user capability. client.blockUser(externalId) hides that user's messages from your channel history and search, workspace-wide; unblockUser reverses it. listBlocked() returns your list — use it to also filter the live WebSocket stream client-side (history/search are filtered server-side). Blocking yourself is rejected.
• Mute — a moderation/admin control. A user is silenced in a channel: their sends are rejected (403 MUTED) until the mute expires. Written automatically by the auto_mute action above (and by channel admins later). Muting isn't an end-user action — to stop seeing someone, use block.

await client.blockUser("user_spammer");
const blocked = await client.listBlocked(); // [{ user_id, external_id, name }]
await client.unblockUser("user_spammer");

Audit trail

Every automatic action is recorded in a purpose-built moderation_actions log — rule, action, offending user, the message (when one was created), and a redacted excerpt — powering the dashboard feed and the message.moderated webhook. Nothing requires a human in the loop.