LingXi Router for XiaoYi Claw

Problem

In XiaoYi Claw / OpenClaw environments, users often need to switch between the built-in model, DeepSeek Flash/Pro, or external API models such as OpenAI, and tune parameters such as contextWindow, maxTokens, thinking, and compaction. Manual configuration edits are error-prone: the wrong file may be changed, settings may be lost after restart, or the global default model may conflict with session-level overrides. When multiple providers coexist, routing, parameters, and restart state can affect one another.

How It Works

The skill wraps model onboarding, switching, parameter inspection, and modification into a dialogue workflow. It first reads existing models.providers and agents.defaults, then validates model names, parameter names, and value ranges. Changes are persisted to openclaw.json, synced to the .last-good backup, and applied through a restart safety protocol. It supports provider-level tuning of context, output length, reasoning, and sampling parameters, as well as global defaults, sub-agent defaults, and contextPruning / compaction settings. It also distinguishes session-level switching from global default changes, so one switch does not affect every new session. When errors occur, it can run self-check repair or roll back from .last-good.

Boundaries

It is best for engineers who already have an external API key and are comfortable with command-line/JSON configuration. It does not change model capability or bypass API limits; high-risk parameters should be backed up first, and production settings should match the selected model tier for thinking and reserveTokens.