Switchboard 0.2.0

Switchboard 0.2.0 adds the current model lineup, interactive configuration, compatibility repair, and experimental self-hosted Laya.

Routing

• Claude Code's strong tier now uses Opus 5.5 (claude-opus-5-5).
• Codex now routes to GPT-6 Luna, GPT-6 Sol, and GPT-6 Astra. GPT-6 has no Terra
  model, so Sol serves both the balanced and strong tiers: the new
  codex-sol-balanced profile defaults to medium reasoning when effort is
  uncertain, and codex-sol keeps high. The shipped policy id is now
  default-v0.4.
• GPT-5.6 Luna, Terra, Sol, and Opus 5 stay in the catalog. Personal policies
  that reference them remain valid, and codex-terra keeps Terra selectable while
  Codex offers it. Existing conversations keep their saved model and effort.
• Jev questions describe capability requirements without naming executor
  models. Routing supports four capability tiers and retains conservative
  fallbacks when context or confidence is insufficient.

Setup and visibility

• switchboard doctor, run in a terminal, offers to fix what it finds after its
  offline report. With your consent it compares your policy with the models your
  installed Codex offers, then proposes changes. For a missing GPT-6 model it
  defaults to the previous generation until you update Codex; skipping the model
  instead names the tier it would move to. It checks that Claude Code is new
  enough for Opus 5.5 (2.1.280 or later) and otherwise offers Opus 5. It also
  asks once whether your Claude plan can use Fable and, if not, routes the
  highest Claude tier to Opus 5.5. Changes are shown and confirmed before
  saving, and the previous policy.json is backed up. doctor --fix runs the
  same checks without asking first.
• switchboard config opens a settings menu for models per tier, skipped
  models, effort limits, agents, prompt history, and the classifier connection.
  switchboard config set|get|unset <setting> changes one setting from the
  shell. Both validate before writing, so policy JSON never needs hand editing.
• doctor shows the effective automatic lineup for each enabled agent, and a
  Codex launch that fails because a routed model is missing points to doctor.
• GPT-6 Luna and Sol require a Codex build that bundles them (0.156.1 or later).
• Codex's automatic task-title request is recognized again, so it no longer
  costs an extra classifier call per conversation. Setup now points to
  switchboard config for changing agents.

Classifiers

• Experimental self-hosted Laya support. Choose Self-hosted Laya in
  switchboard init to classify with your own laya-serve (tested with Laya
  0.3.20). A server on this machine needs no key; another host requires HTTPS
  and a key. Laya receives a shorter question set that normalizes into the same
  classification, and its routing quality is not yet evaluated. Tasks longer
  than the 700 characters Laya reliably reads use the uncertain fallback.

Install or update

npm install -g @ruban24/switchboard@0.2.0
# Or:
brew update
brew upgrade ruban-24/tap/switchboard

Run switchboard doctor after upgrading. Personal policy is preserved, and existing conversations retain their saved model and effort.

Switchboard 0.1.0

Switchboard chooses a model and reasoning effort for Claude Code and Codex, using Jev and a policy you control. This first release supports interactive CLI sessions on macOS and Linux.

• One Jev request assesses the task; local policy selects the model and effort.
• The selected pair stays fixed through follow-ups, tool calls, and resume to avoid cache disruption from automatic switching.
• Personal overrides let you exclude models, change tier mappings, cap effort, and adjust confidence thresholds.
• Claude shows a route status line; Codex shows a routing notice. switchboard explain shows saved decisions.
• switchboard init detects installed agents and guides you through the Jev connection and API key setup.

Switchboard runs locally without telemetry. Extracted task text goes to your chosen Jev provider, and provider usage is billed separately. Raw-prompt logging is off by default.

Install

With npm (Node.js 22.18 or newer):

npm install -g @ruban24/switchboard

With Homebrew:

brew install ruban-24/tap/switchboard

Then:

switchboard init
switchboard doctor
switchboard claude
# Or: switchboard codex

You need an installed, signed-in coding agent and a Jev API key. You can also run npx @ruban24/switchboard init without a global installation.

Scope

Native Windows, desktop-app routing, remote/background sessions, other coding agents, and additional System One classifiers are outside 0.1.0. Jev works through TypeSafe directly, Vercel AI Gateway, OpenRouter, or an endpoint implementing a supported Jev contract. OpenRouter has documented-contract and synthetic coverage; it has not been live-tested for this release.

See the installation guide, personal policy guide, and full changelog.