Bastion — AI-Native Company Brain & Agentic OS

Bastion isn't a wrapper around an LLM. It's a five-layer operating system for commanding many autonomous agents from a single seat: a strict SDLC Factory that builds the software, an Engine that executes agent workflows, a semantic-graph Brain that stays provably current, a real-time Console to observe and steer, and a phone Surface to do it untethered. The system is deliberately not a monorepo. Ten repositories — spanning Rust, Python, and Flutter — keep their own git histories, test suites, and release cadences, and compose through a small number of versioned, one-way API seams. This repo (bastion-os) is the front door: the architecture, the Mermaid system diagram, and the links out to every module. The five layers: the Console (bastion, Rust) is the operator's seat — sessions, live cost and budget, a kill switch, and brain queries; the Engine (orchestrator in Python, engine-rs in Rust) executes DAG-validated agent workflows; the Brain (mev + pgvector) is the context layer, where mev is the single reference parser for the company's file formats and what it validates is exactly what gets embedded; the Factory (base-template) is the SDLC harness that scaffolds every spec into isolated git worktrees; and the Surface (bastion-ui, Flutter) is the untethered mobile operator, driving the whole system over Tailscale via a versioned HTTP + WebSocket API. The guiding principle is architectural, not feature-driven: a solo-run system fails not from missing features but from self-knowledge drift — hand-maintained views lagging reality, multiple parsers of the same formats, a semantic index that's never current. Bastion is designed against that failure mode first: one parser (okf-core, a shared contract crate), one writer of derived state, and a Brain where validated == embedded == derived.