Class: Brute::Tools::SkillLoad

Inherits:
RubyLLM::Tool
  • Object
show all
Defined in:
lib/brute/tools/skill_load.rb

Overview

The skill tool — stages 2 (activation) and 3 (execution) of the Agent Skills progressive-disclosure lifecycle (https://agentskills.io).

Stage 1 (discovery) happens in the system prompt: Brute::Prompts::Skills lists each skill's name + description only (~100 tokens each). When a task matches, the model calls this tool with the skill name to pull the full SKILL.md body into the conversation (stage 2).

The output also reports the skill's base directory and a capped listing of bundled files. That is stage 3: skills bundle scripts/, references/, and assets/ that the model runs or reads through the agent's existing tools (shell, read) by relative path. There is no separate skill runtime.

The tool must scan the same directories as Prompts::Skills, or it would advertise skills it cannot load. Both default to Dir.pwd; agents that point the prompt at a custom root (e.g. ctx.merge(cwd: dir)) should build the tool with the matching cwd: Brute::Tools::SkillLoad.new(cwd: dir).

Reference: opencode's tool/skill.ts.

Constant Summary collapse

FILE_LIMIT = 
10

Instance Method Summary collapse

Constructor Details

#initialize(cwd: Dir.pwd) ⇒ SkillLoad

Returns a new instance of SkillLoad.



45
46
47
48
 
# File 'lib/brute/tools/skill_load.rb', line 45

def initialize(cwd: Dir.pwd)
  super()
  @cwd = cwd
end

Instance Method Details

#execute(name:) ⇒ Object 



50
51
52
53
54
55
56
57
58
 
# File 'lib/brute/tools/skill_load.rb', line 50

def execute(name:)
  skill = Brute::Skill.get(name, cwd: @cwd)
  return unknown_skill(name) unless skill

  directory = File.dirname(skill.location)
  files = bundled_files(directory)

  render(skill, directory, files)
end

#name ⇒ Object 



43
 
# File 'lib/brute/tools/skill_load.rb', line 43

def name; "skill"; end