Add model_picker toolset for dynamic model switching

[dgageot]

Introduce a new 'model_picker' toolset that lets agents switch between LLM models mid-conversation. This enables agents to pick the best model for each sub-task (e.g. a fast model for simple questions, a powerful one for complex reasoning) and revert when done.

Changes:

• Add ModelPickerTool with change_model and revert_model tools
• Register model_picker in the toolset registry
• Wire runtime handlers (handleChangeModel, handleRevertModel) that validate the requested model against an allowed list, update the agent's model, and emit AgentInfo events so the UI reflects changes
• Simplify registerDefaultTools to a plain handler map instead of pre-resolving tool definitions
• Add config validation: 'models' field is required and exclusive to model_picker toolsets
• Update agent-schema.json with model_picker type and models field
• Add example configuration (examples/model_picker.yaml)
• Add comprehensive unit tests for ModelPickerTool

[docker-agent]

Review Summary

Assessment: 🟡 NEEDS ATTENTION

The model_picker implementation is well-structured overall, but there's a validation gap that could allow invalid model configurations to pass through to runtime.

Found 1 issue requiring attention:

• 1 MEDIUM severity issue in config validation

Key Finding

The validation logic checks that the models list is non-empty for model_picker toolsets, but it doesn't validate the contents of that list. This means configurations with empty strings (e.g., models: [""]) would pass validation but cause runtime errors when the agent attempts to switch models.

Recommendation

Add validation to iterate over t.Models and ensure each entry is a non-empty string with valid model identifier format.

[docker-agent]

MEDIUM: Missing validation for individual model strings

The validation checks that len(t.Models) > 0, but it doesn't validate whether each string in the Models slice is non-empty or valid. A configuration like this would pass validation:

model_picker:
  models:
    - ""
    - "gpt-4"

But would cause runtime errors when the agent tries to switch to the empty model string. Consider adding:

for i, model := range t.Models {
    if strings.TrimSpace(model) == "" {
        return fmt.Errorf("toolset %q: models[%d] cannot be empty", t.Name, i)
    }
}