Orchestration
5/5Decomposes a goal into the intent — plans, specs, fan-out shapes — that the rest of the squad acts on.
Kira is the squad's head — summoned for a decision and then released. It reasons and emits intent: the plans and swarm specs the rest of the squad executes. A deterministic executor turns that intent into action, so the failure mode to manage is the hand-off from reasoning to a tool call, not raw capability.

as of 2026-10-12 22:44 UTC · brain read from the harness dispatch map at deploy; runs from the workflow log
Decomposes a goal into the intent — plans, specs, fan-out shapes — that the rest of the squad acts on.
Holds a large working context for a single head-decision rather than a sustained loop.
Uses tools, but its intent is actuated deterministically rather than depended on for the tool call itself.
The squad rebalance moved Kira's seat to GLM through the :8788 proxy — the glm-5.2 alias actually serves the full GLM-5.3. The head's wake is seconds, not minutes: the runner wakes it on terminal events and a heartbeat keeps it current. ds4, the DeepSeek seat it ran for a day, became the lab's on-demand vision engine.
The 09-02 head flip put Kira on DeepSeek-V4-Flash through ds4, the lab’s local inference engine — superseded the next day by the proxy topology.
Every Mac seat, Kira’s included, moved onto Laguna XS quantized by our own calibration pipeline (the published sbQ build) — local and unmetered. Kira left it on 09-02 when the head role called for a bigger brain.
Stood up as the orchestrator: it reasons and emits intent that a deterministic executor turns into squad work.
Energy per token to be measured once the head is resident, on the same method as the rest of the roster.