Incident Room

Incident Room gives Reddit moderators one shared command post during a live incident. It groups evidence, lets mods claim work, asks AI for a bounded briefing, and keeps every action human-reviewed

Comment

Inspiration

Fast subreddit incidents rarely arrive as one clean report. A sports final gets spoiled, a creator AMA is brigaded, or a suspicious domain spreads through comments while five moderators are checking queues, chat, modmail, and links in separate tabs. The risk is not only a slow removal. It is duplicated review, inconsistent action, and lost reasoning while the community is watching.

What it does

Incident Room gives moderators one shared command post inside Reddit. A moderator declares an incident from a Devvit custom post, sees scored evidence, claims the next item, opens a bounded AI briefing, previews an action pack, confirms the pack, and leaves an after-action record for the next shift.

The AI layer is intentionally constrained. Deterministic rules score watch terms, repeated patterns, fresh accounts, report velocity, and watched domains. Step 3.6 then writes a compact briefing with likely pattern, recommended action pack, and moderator caveats. The model cannot remove, lock, ban, approve, or mutate Reddit content. Rules score the room; AI explains the room; humans decide.

How we built it

Incident Room is built with Devvit Web, React, TypeScript, Hono, Devvit Redis, Reddit API permissions, and Playwright. The Devvit server stores the room state, evidence, claims, action packs, briefing state, timeline, and after-action metrics in Redis. The browser receives only the room state and never receives the AI provider key.

For AI, the app uses OpenAI-compatible server-side settings. The submitted review app is configured for StepFun at https://api.stepfun.com/v1 with step-3.6. A local smoke test verifies the real model path and returns a structured briefing.

Challenges we ran into

The hard part was keeping AI useful without letting it outrun moderation policy. A generic chat surface would be easy to build but wrong for moderators. Incident Room keeps scoring deterministic, keeps every enforcement action in preview/confirm state, and treats the model as an explanation layer rather than an actor.

Accomplishments

We built and uploaded a real Devvit app, submitted version 0.0.3 for review, wired a real Step 3.6 AI smoke path, and covered the core moderator flow with Playwright on desktop and mobile. The demo video is cut to under one minute and shows the actual declare, claim, briefing, preview, confirm, and after-action flow.

What's next

Next, Incident Room should add subreddit-specific rule templates, saved post-incident reports, and a smoother queue handoff for large moderator teams that rotate during live events.

Built With

Share this project:

Updates