D-CSIL LabRats

LabRats is the lab's production coding and research agent program — a gateway-first runtime where every UI is just a client. It consolidated and replaced the earlier Hermes and OpenClaw agents, importing their Telegram and email identities. Its live session database was last written the day of this inventory.

Remy and Justin are identities assigned to running agents, not separate codebases. Remy is the lead agent and runs against a cloud model. Justin is the LabRat Assistant, working from a second instance against a local model.

The split is the point: the expensive reasoning tier is spent where judgment is actually needed, while a local model carries assistant work at no marginal cost and with nothing leaving the machine. It is the same preference the sub-agent router encodes as prefer_local, applied one level up — to the agents themselves rather than to the workers they spawn.

Policy is deterministic, not model-driven: an allow/deny engine gates every tool call across ask/plan/edit/auto modes. Every file edit is a reversible git-backed checkpoint with an inspectable tool timeline. Their tools span filesystem, shell, git, web search and HTTP, email, and native Cortex memory calls.

spawn_agents launches concurrent sub-agents, each with its own model, routing strategy (prefer_local, cheapest, fastest, best…), capability requirements, fallback chain, and token/cost budget. The router live-probes Ollama, LM Studio, MLX-LM, llama.cpp, vLLM, and LocalAI alongside cloud providers. Verified live: five sub-agents across four providers ran concurrently, a dead endpoint triggered automatic failover, and a $0.02 cloud budget correctly forced a reroute to local. 35/35 router tests pass.

The Soul package gives Remy a persistent identity, mood, and 'whisker nudges'; a Cortex project named remy-brain is pulled into every system prompt, so the rat remembers who it is.