Persona-Selector Skill

Goal

Pick the best persona(s) for each task instead of relying only on the fixed per-phase assignments in loki-mode.skill. The phase defaults remain the baseline; this skill lets the orchestrator adapt — choosing a more relevant persona, tiering by cost, delegating between personas under a contract, voting on high-stakes decisions, and letting personas refine themselves within bounds. Everything here is cost-gated by GABBE_AUTONOMY + budget.

Selection (semantic routing — markdown-heuristic by default)

Rank candidate personas for the task using the persona registry (agents/personas/00-index.md) — no new embedding infra required:

1. Relevance — match task keywords/intent against each persona's role + primary output + swarm. (Optional v2: embed persona profiles and retrieve top-k; for now reason over the index text.)
2. Scope fit — does the task fall inside this persona's mandate and phase boundary?
3. Past success — if AUDIT_LOG/memory shows a persona handled similar tasks well, weight it up (and vice-versa).
4. Cost tier (SLM/LLM tiering) — route to the cheapest persona + model tier that clears the task's complexity bar. Keep the persona, drop the model tier, for simple work; reserve SOTA models for genuinely hard/critical tasks. This reconciles with each skill's context_cost and the brain/cost-benefit-router.skill. Biggest cost lever in multi-persona work.

Output the top 1–3 with a one-line justification each; pick one unless the task benefits from a panel (see Voting).