Hive

Hive is the workshop the rest of these projects come out of. It runs entirely on my own hardware on local models — no per-token API bill, no data leaving the house — and its job is to turn an intent into shipped, tested code with as little of my attention as possible.

The spine is a crew board: a single-owner task pipeline that flows Proposed → Backlog → Ready → In Progress → Review → Done. It auto-discovers the git repos under my projects folder, decomposes a request into dependency-ordered tasks, and assigns each one to a runner. Agents make two honest attempts; on repeated failure the task escalates to a heavier model instead of silently rotting.

A turn is a loop, not a single prompt. A planner decomposes intent into structured delegations; specialist helpers — librarian, coder, critic — run in parallel under VRAM-aware concurrency that falls back to CPU before it ever downgrades model quality; a critic gate blocks risky actions before they happen; a synthesizer emits the reply and the side effects. A hallucination guard strips any sentence whose numbers are not grounded in a helper's actual output.

Nothing is trusted because an agent said so. The verifier enforces a four-gate acceptance — owner approval, the project's real test command passing, an independent verify pass, and a visible diff — before a task is allowed to call itself done. That gate is the whole reason the output is usable. (It is also the gate that, when it was weaker, let a game get marked shipped that never actually shipped — which is its own story.)

Memory is layered: verbatim recent turns, a mid-tier fact store, and a vault with full-text plus vector search over every past conversation and a hand-curated knowledge base. An hourly auditor scans its own behavior for hallucinations, dropped requests, and security gaps; an idle-time groomer hunts duplicate and contradictory notes. Python and FastAPI at the core, Ollama for the models, a Flutter companion app and a web crew board on top. Private infrastructure, single operator.