Switchboard is a local Claude Code and Codex router. It assesses a new conversation's task with Jev, applies deterministic confidence rules to select the model and reasoning effort, and pins that route through follow-ups, tool calls, and resume to avoid unnecessary prompt-cache disruption.
Jev classifies the task's capability tier, whether context is sufficient, and the lowest adequate reasoning effort for the candidate model. Application code applies confidence floors, maps the result to a model and effort, and persists that pair for later turns and tool continuations.
Adds structured choices or scores to the workflow; performance and cost benefits have not been independently verified.
The cited question builder and README were checked for model-specific reasoning effort and conversation route pinning. No CLI integration or cache measurement was run here. A pinned route preserves a model choice; it does not guarantee a provider cache hit.
Select model capability and per-model reasoning effort while preserving a conversation's chosen route.
Material analysed：
buildJevQuestions asks effort relative to each eligible model. A stronger model does not automatically need greater effort, and the selected model's answer is the one policy uses.
Inspect the supporting source ↗The README keeps the initial route stable and permits native manual overrides. This avoids automatic switching churn but does not guarantee a provider cache hit.
Inspect the supporting source ↗Consider it when routing explainability and conversation stability matter. Do not equate confidence in a task category with probability that the chosen model will succeed.
Inspect the questions and Gateway metadata normalization, then use explain output to compare route reason, fallback and selected-model effort without changing a live conversation.