Keep a dated record of important architecture choices and why they were made.
- Context: the stack needed a single mission orchestrator with a readable audit trail.
- Options considered: Flowise/Langflow, Sim Studio, Paperclip, OpenClaw.
- Final choice: keep OpenClaw as the runtime center and use the wiki for human visibility.
- Impact: the rest of the stack can specialize without becoming the source of truth.
- Context: agent work needs reusable memory, session state and hierarchical retrieval.
- Options considered: wiki pages only, vector-only memory, OpenViking.
- Final choice: use OpenViking for machine-readable context and keep Wiki.js for human memory.
- Impact: long-running agents can retain operational context without polluting the wiki.
- Context: we need an interactive operator layer that can do real work but still be interrupted.
- Options considered: autonomous execution only, builder-only tools, Agent Zero.
- Final choice: use Agent Zero as the supervised executor.
- Impact: tasks can run 24/7 with human checkpoints for sensitive actions.