Model Switcher Skill

Core Rule

When the user asks to switch the current session model, resolve the target dynamically from the active OpenClaw configuration first, then switch the current session directly. Do not hardcode provider/model values inside the skill.

Required Behavior

Switch current session immediately

When the user says things like:

• switch to gpt
• switch to claude
• use qianwen
• change to minimax
• switch to qwen3.5-flash

The agent should:

• read models.providers from the current OpenClaw configuration
• use skills/openclaw-session-model-switcher/model-aliases.json for provider-level alias resolution
• dynamically match against configured models when the user provides a concrete model name or fragment
• directly execute /model <provider/model> in the current session after resolving a unique target

Recommended execution path:

• prefer scripts/switch-model.sh <selection> to resolve the target model
• once the script returns a unique model, execute the returned command in the current session
• if the environment supports a direct session-level model override tool, prefer that
• do not merely print /model ... as plain advice when direct execution is possible

Restore default

When the user says things like:

• restore default model
• reset model switch
• use the default model again

The agent should directly execute:

• /model default

Show current session model

When the user says things like:

• what model am I using now
• show current model
• model status

The agent should directly query the current session model state instead of only explaining how to check it.

List configured models

When the user says things like:

• what models are available
• list current models
• show configured models
• what qwen models do I have

The agent should read models.providers from the current OpenClaw configuration and list switchable models. If the user specifies a provider, only list models for that provider.

Resolution Rules

• trust only the current OpenClaw configuration in models.providers
• keep model-aliases.json limited to provider aliases and default semantics
• if a provider has only one configured model, use it directly
• if a provider has multiple models, prefer model-name matching from the user input
• if multiple candidates still remain, list the candidates and ask the user to choose
• if nothing matches, clearly say that no configured model matches the request

Multi-Agent Rule

• this skill affects only the current persona in the current session
• do not modify global configuration
• do not affect other personas or sessions
• do not restart the Gateway

Do

• switch the current session directly
• tell the user the current session model was switched
• mention that the change affects only the current session when helpful
• return candidate lists for ambiguous matches
• list current available models when the user is unsure
• consume script output at the upper layer instead of exposing raw stderr or internal errors

Do Not

• do not edit openclaw.json
• do not restart the Gateway
• do not hardcode model IDs inside the skill
• do not maintain a static model list separate from the active configuration
• do not accidentally implement this as a global default model switch
• do not guess in multi-candidate cases
• do not dump raw JSON, stderr, or internal tool errors to normal users

Suggested Script Interface

• scripts/switch-model.sh <selection>: resolve user input and return JSON
• scripts/list-models.sh [provider]: list all configured models or only one provider
• scripts/model-status.sh: output /model status

Upper-Layer Interaction Flow

For switch requests

When the user asks to switch models, the upper layer should:

1. call scripts/switch-model.sh <selection>
2. parse the JSON output
3. branch on status

• status: ok

  • execute the returned command
  • confirm which model was selected
  • mention that it only affects the current session when useful

• status: ambiguous

  • do not surface raw script output
  • convert options into a readable candidate list
  • ask the user to reply with a model name or number
  • do not guess

• status: error

  • explain naturally that no configured model matched
  • suggest listing available models when appropriate

For list requests

When the user asks for available models or models from a specific provider:

1. call scripts/list-models.sh [provider]
2. present the result as a readable provider-grouped list
3. do not mix in models that are not present in the active configuration

For status requests

When the user asks which model is currently active:

1. call scripts/model-status.sh
2. execute the session-level status query
3. present the result in natural language instead of only teaching the command

Standard Reply Templates

Switch success

• Switched the current session to <provider/model>.
• Switched the current session to <provider/model>; this affects only the current session.

Ambiguous match

• You can switch to these models:
• 1. <provider/model-a>
• 2. <provider/model-b>
• Reply with the model name or a number.

If the user specified a provider:

• You can switch to these <provider> models:

No match

• No configured model matches "<selection>".
• I can list the models currently available in your configuration.

List models

• These models are currently available in the active configuration:
• list them grouped by provider

Fallback Rule

Only if the current environment truly cannot execute native session commands should the agent fall back to replying with:

• the exact command to run
• why it cannot execute it directly in the current environment

That fallback is a degraded path, not the default path.