The Pack is an event-sourced (CQRS) multi-agent research orchestrator with a transactional outbox. Six diagrams tell the whole story, in one consistent style: monochrome, hairline connectors, big type, and short asides that explain why, not just what. Solid lines are the synchronous write path; dashed lines are the asynchronous read / projection path; an inverted (black) node is a single source of truth. The one exception is the deployment board, where black marks the three places Alibaba Cloud does the work.
The narrated walkthrough in plain words is The-Pack-Architecture.pdf.
Where it actually runs. The compute is one Alibaba Cloud ECS server holding three Docker containers (nginx, the Python engine, the Rust gateway) behind HTTPS. Every agent's reasoning is a Qwen call to Alibaba Model Studio, and every generated file is stored in Alibaba Cloud OSS — three Alibaba touchpoints, marked in black. Postgres (Neon) and Redis (Upstash) are managed, over TLS.
The signature loop at a glance: a command comes in, becomes exactly one validated event, one write to Postgres (the truth); a relay projects it to Redis; the read-only Rust gateway fans it back over WebSocket; the browser rebuilds all state from the event stream. Follow the numbers, down and back up.
Every layer as real modules: the FastAPI API surface, the engine (Supervisor, strategies, the wolf pack — Alpha, Beta, Scouts, Tracker, Sentinel, Howler, Elder, Warden — plus Boundary, Emitter, Forge, and the tools), the providers it calls out to (Qwen, and DuckDuckGo + Jina readers for the web), persistence, and the async projection / read path.
React 18 + Vite SPA: entry and providers, then routing, then the feature modules (organised by feature, not file type), then the three shared layers they draw on (state, UI, data access), then the seam to the backend — REST + SSE to the engine, WebSocket to the Rust gateway.
The ten domain tables in Postgres, hub-and-spoke around hunts. events.(hunt_id, seq) is the
append-only spine (the composite key the database enforces); hunts.parent_hunt_id threads follow-up
hunts to their origin; documents and instincts stand on their own; memory is a soft link (no FK)
so a lesson survives its hunt.
A hunt end to end: create, subscribe, the projection loop, plan (plan_proposed), approve, run the pack
(wolf_spawned → tool_called/tool_result → artifact_created → hunt_completed), then the final
artifact — plus the SSE chat side-channel that streams straight from the engine and skips the gateway.
