llm-switcher（LLM switcher 文档）

Model Switcher

Change the global default LLM model in OpenClaw config, with a session-level availability check first.

Workflow

1) Read config and show available models first

Run:

openclaw config get agents.defaults.models
openclaw config get agents.defaults.model.primary

Extract and show:

• model key (for example openai-codex/gpt-5.3-codex)
• alias (for example codex, opus) when present
• current default model

Do not switch anything before showing options.

2) Ask user to choose model

Ask for the model name they want to set as the global default. Accept:

• alias (preferred): codex, opus
• full provider/model id when configured

If the name is not in configured options, reject and ask again with the valid list.

3) Test selected model in current session first (availability/conflict check)

Before changing default config, apply a session-only override and run a live reply test.

In OpenClaw agent/tooling context:

• Use session_status with model=<selected-model>.
• Confirm the override succeeds (no error about unknown/unavailable model).
• Then send a short test prompt to the model and require an actual response, for example:

- "What model/version are you currently running? Return provider/model id and a one-line status."

• Treat this as pass only if the model returns a normal reply.

If override or live reply test fails:

• Do not update default config.
• Report the error clearly.
• Explain that the model appears unavailable/misconfigured.
• Ask user to choose a different configured model.

If override and live reply test both succeed:

• Continue to update default config.

4) Apply default-config change

Update default model in config:

openclaw config set agents.defaults.model.primary '"<selected-model>"'

Then verify:

openclaw config get agents.defaults.model.primary

Never claim success without verification output.

5) Ask before restarting gateway

After updating config, explicitly ask:

• "Do you want me to restart the gateway now so the change takes effect?"

Only restart after a clear yes.

6) Restart gateway on confirmation

Run:

openclaw gateway restart

Then verify status:

openclaw gateway status

If restart fails, report the error and suggest:

openclaw gateway stop
openclaw gateway start
openclaw gateway status

Response style

• Be short and direct.
• Always show current default and valid model choices before asking for selection.
• Do not ask scope questions (no session-only final path in this skill).
• Always run a session-level availability check before writing default config.
• Always require explicit confirmation before restarting gateway.
• Never claim success without verification output.