Harness, not prompts. Invariants live in code that always runs — mandatory .gitignore, build-artifact stripping, secret redaction. The model is never trusted to remember them.

Built for long runs. An epic runs as a multi-wave DAG — waves execute in dependency order, so a full run spans hours. Per-story windows are configurable (default 30 min, up to 2 h) with stuck/hang detection and a doom-loop detector (regression vs stagnation) that stops the run and escalates to a human.

Runs two ways. Auto-routed between in-process local-tmux and a remote compute-node fleet (claim + slots, wave-scheduled). Live output streams to the dashboard over tmux + ttyd; dropped completions persist and replay on reboot.

4-layer eval + human gate. Deterministic exit-code gates → differential security scan → LLM-as-judge semantic check → OWASP-scoped review. Then every story lands in review for a human to merge.

Multi-agent orchestration. Generator + independent verifier (different models, no confirmation bias); a Sonnet planner consults an Opus advisor before the coder runs.

The Vault is the brain. A git-backed markdown knowledge base — definitions, epics, stories, ADRs, QA matrices — that every agent reads for context (vault_grep) and writes back to. Versioned, access-scoped, and the reason long runs stay grounded instead of hallucinating.