Forge — LLM Batch & Realtime Orchestration

LLM workloads need both batch and realtime execution, provider flexibility, retry safety, cost awareness, validation, and operational visibility. A queue-based architecture adds infrastructure and coordination complexity.

Forge explores a Postgres-centered architecture where workers coordinate through the database alone — no queue, no IPC, no shared worker processes.

Anvil (the builder) claims eligible pending rows and shapes them into provider-compatible batches. Hammer (the executor) advances those batches through worker loops that submit, poll, collect, execute realtime jobs, recover stale leases, archive cooled batches, stamp updates, recover quota, and send notifications.

OpenAI and Anthropic ship through provider ports; the core execution model stays provider-agnostic. The two daemons share only Postgres — all coordination lives in the database, with no separate queue or message bus.

• - Batch and realtime LLM execution paths in one architecture.
• - Worker coordination through Postgres without a separate queue service.
• - Lease expiry for crash recovery and stale worker reclamation.
• - OpenAI and Anthropic via provider ports.
• - Provider outputs revalidated against processor response schemas before success.
• - Operator-facing health, metrics, archive, credentials, costs, architecture surfaces.
• - Horizontal scaling: every daemon and worker is safe to run as multiple copies.