model-routing
Installation
SKILL.md
Model routing
Climb a budget ladder: choose the cheapest adequate model and lowest explicit effort, then escalate only when evidence warrants it. Risk overrides surface simplicity.
Assign the worker
| Model | Route here | Starting effort |
|---|---|---|
gpt-5.6-luna |
Bounded extraction, classification, summaries, support triage, mechanical low-risk checks | low |
gpt-5.6-terra |
Documentation, structured writing, product copy, content planning | low; medium for multi-source synthesis |
gpt-5.5 |
Default general coding, routine review, coordination, workload planning, mixed everyday work | low for mechanical changes; otherwise medium |
gpt-5.6-sol |
Hard debugging, security, architecture, difficult reasoning, compatibility-sensitive code, legal or business risk review | medium; high when ambiguous or high-risk |
For these Codex models, Pi maps minimal to provider effort low; standardize on low. Thinking off merely omits an explicit reasoning setting, so low is the lowest predictable route. Numeric thinkingBudgets apply to budget-based adapters, while Codex effort is controlled with --thinking.
Use xhigh or max after high fails for a reasoning-related cause. GPT-5.5 supports through xhigh; Luna, Terra, and Sol also support max. When an output misses acceptance criteria, raise one effort rung if the model still fits the workload; switch models when capability or risk was misrouted.