One assistant, two brains

Why does the shipped assistant run two models instead of one? Because answering a person and working a problem are two jobs with two budgets, and two cells put that choice in files a mutation can change. One model doing both would put the same decision inside a prompt, where nothing can measure it and no mutation can move it.

Two templates, two model slots

templates/talky holds the conversation surface: a hive referencing a session keeper, a collector and a dispatcher, plus one llm brain, with every internal edge pre-wired. It owns the session and the window and it speaks to the person.

templates/cogny is the same shape without a channel. No session keeper, no summarizer, no proxy, because a core without a channel has no sessions and no night. Its conversation is the errands the surface sends it, and it declares that errand as a tool schema of its own.

Both require ctx.model at instantiation, and both stand alone elsewhere. The level that references both is the one place the two models can differ, so templates/assistant declares a strict second key, ctx.model_surface, and spends it in the ref marker for the surface (templates/assistant/talky/config.json, trimmed):

{ "cell": { "type": "ref", "template": "talky@5.1.0" },
  "override_params": { "brain": { "model": "${ctx.model_surface}" } } }

The ref marker for the core carries no such override, so it reads ctx.model. Before that key existed, one flat ctx reached both occupants, both resolved model, and the surface ran the reasoning model (GH #516).

What the split cost while it was wired loosely

There is no latency measurement for the split itself, because nobody has run the comparison. What is measured is a failure it caused. A thinking model puts a sentence beside its tool bundle, the dispatcher used to send that sentence on the answer lane, and for the core that lane is the advice lane of the voice that asked. On one colony 11 of 26 answers on that lane were interim sentences, and a single user turn produced thirteen messages. The repair was one knob: the core ships with interim off (templates/cogny/dispatcher/config.json, GH #539).

Two brains cost two provider accounts, two failure modes and one hop more on every hard question. For a colony that fields small talk that is overhead, and a talky alone is a whole agent.

The level that wires the two, and the levels above it, are on meclaw-os.