Sub-Agents

You: How is error handling done in this codebase?

Loom: Let me search the codebase for error handling patterns...
      [spawns sub-agent]
      
      Found consistent error handling:
      - Controllers use {:ok, _} | {:error, _} tuples
      - Contexts return Ecto.Changeset for validation errors
      - Phoenix.Controller.action_fallback handles errors
      - Custom MyAppWeb.ErrorHelpers formats error messages

You: Use a sub-agent to find all files that import the Repo module.

Loom: [spawns sub-agent scoped to lib/]
      
      Found 23 files importing Repo:
      - lib/myapp/accounts.ex
      - lib/myapp/accounts/user.ex
      - lib/myapp/content.ex
      ...

context = %{project_path: "/path/to/project"}
params = %{
  task: "Find all modules that use GenServer",
  scope: "lib/myapp"  # Optional: defaults to project root
}

{:ok, %{result: findings}} = Loom.Tools.SubAgent.run(params, context)
IO.puts(findings)

%{file_path: "lib/myapp/accounts/user.ex"}

%{pattern: "**/*_controller.ex"}

%{
  pattern: "use GenServer",
  file_pattern: "**/*.ex"  # Optional: filter files
}

%{path: "lib/myapp/accounts"}

# In lib/loom/tools/sub_agent.ex
@max_iterations 10

You: How does the notification system work in this app?

Loom: I'll search for notification-related code.
      [spawns sub-agent with task: "Find and explain notification system"]

Loom: The notification system is implemented in lib/myapp/notifications.ex.
      It provides email notifications via Bamboo and in-app notifications
      via Phoenix PubSub. Preferences are stored in users.notification_settings.
      
      Key entry points:
      - Notifications.send_email/2
      - Notifications.broadcast/2
      
      Would you like me to show you the implementation of any specific part?

defmodule Loom.Tools.SubAgent do
  use Jido.Action,
    name: "sub_agent",
    description: "Spawn a read-only search sub-agent",
    schema: [
      task: [type: :string, required: true],
      scope: [type: :string]
    ]
  
  @max_iterations 10
  
  @read_only_tools [
    Loom.Tools.FileRead,
    Loom.Tools.FileSearch,
    Loom.Tools.ContentSearch,
    Loom.Tools.DirectoryList
  ]
  
  def run(params, context) do
    task = param!(params, :task)
    scope = param(params, :scope, context.project_path)
    model = weak_model()
    
    # Build system prompt
    system_prompt = """
    You are a search assistant working in #{scope}.
    Find information about: #{task}
    Use available tools to search files and code.
    When done, respond with your findings.
    """
    
    # Run search loop
    case run_sub_loop(messages, tool_defs, model, scope, 0) do
      {:ok, answer} -> {:ok, %{result: answer}}
      {:error, reason} -> {:error, reason}
    end
  end
end

# ❌ Too broad
%{task: "Find authentication logic"}  # Searches entire project

# ✅ Appropriately scoped
%{task: "Find authentication logic", scope: "lib/myapp/accounts"}

1. Sub-agent: "Find all places where User.changeset is called"
2. Main agent: "Update all those call sites to use the new validate_user/1 function"

Task.async_stream(queries, fn query ->
  Loom.Tools.SubAgent.run(%{task: query}, context)
end)