DreamForge Dashboard

What is the Switchboard?

The AI Switchboard controls which AI providers handle each type of generation in DreamForge. Each category (images, meshes, video, music, SFX, LLMs) has a priority-ordered list of providers. The 1 provider is the primary — it handles all requests first. If it fails, the system automatically falls back to 2, then 3, and so on down the chain.

Managing Providers

• Reorder: Drag and drop tiles to change priority. The leftmost provider (1) is always tried first.
• Add: Click the + button at the end of a row to add an available provider as a new fallback.
• Remove: Click the × on a tile to remove it from that category.
• Skip on refusal: Double-click a tile and tick Skip on content refusal to drop a provider out of any run whose prompt was already refused on content grounds. Position is unchanged — it is the same fallback as before for outages and timeouts. Marked tiles show a dashed amber border and a SKIP ON REFUSAL tag.
• Save: Changes are local until you press Save. All running game clients will pick up the new config on their next switchboard refresh.

Fallback & Circuit Breaker

When a provider fails, the switchboard retries once, then moves to the next fallback. After 3 consecutive failures, a provider enters a 5-minute cooldown and is automatically skipped. If all providers are in cooldown, the least-recently-failed one is tried as an emergency fallback. Providers recover automatically once they succeed again.

Content refusals are treated differently. Once two providers refuse the same prompt on content-moderation grounds, the prompt is the problem, not the provider: the chain stops, the prompt is rewritten, and the whole chain runs again with the clean text. A single refusal is already enough to pass over any provider marked skip on refusal for the rest of that run. Two things earn the mark: being a second route to a model already tried (Runware and OpenAI both reach gpt-image-2, so the refusal is the same refusal, billed twice), and being expensive enough that a prompt in trouble is not worth spending it on. The rewritten pass starts with a clean slate, so a marked provider is only out for the run that was refused.

Active Queries

Each tile shows how many calls to that provider are currently in flight. Category headers show the sum across their providers. The “Generation Queue” panel above the grid shows the real backend job queue: every queued and in-progress generation with progress and elapsed time, plus jobs finished in the last hour.

Costs

Each tile shows an estimated cost label. For media generation it’s per generation. For TTS it’s per 1k characters. For LLMs it’s the input token rate per 1M tokens.