Class: LLMExperiment::Auth

Inherits:
Object
  • Object
show all
Defined in:
lib/llm_experiment/auth.rb

Overview

One-time interactive step: log the agent CLIs in to the user's plan subscriptions, and keep the credentials in a directory that every later trial mounts read-write.

Nothing is baked into an image and no API key is required. Re-run this when a refresh token expires; llmx run stops with that instruction rather than burning trials against a dead login.

Claude and Codex are pointed at the credential store through CLAUDE_CONFIG_DIR and CODEX_HOME, so their logins land there directly. opencode takes no such variable: it keeps credentials in XDG_DATA_HOME next to its sessions and its database, so logging in writes a whole state directory. This points XDG_DATA_HOME at a container-local directory that dies with the container and copies out the one credential file afterwards, which gets opencode the same shape as the other two without carrying its session store along.

Constant Summary collapse

AGENTS = 
%w[claude codex opencode].freeze

Instance Method Summary collapse

Constructor Details

#initialize(container: nil, shell: Shell) ⇒ Auth

Returns a new instance of Auth.



23
24
25
26
 
# File 'lib/llm_experiment/auth.rb', line 23

def initialize(container: nil, shell: Shell)
  @container = container || Container.new
  @shell = shell
end

Instance Method Details

#login(agents: AGENTS) ⇒ Object 



28
29
30
31
32
33
34
35
36
37
38
39
 
# File 'lib/llm_experiment/auth.rb', line 28

def (agents: AGENTS)
  wanted = wanted_agents(agents)
  image = ready_image!

  print_intro(wanted)
  # Interactive: hand the terminal to the container so the device flows work.
  @shell.interactive(*run_argv(image, , interactive: true))

  puts "\nVerifying what persisted..."
  @shell.sh(*run_argv(image, verify_script), allow_failure: true)
  print_persisted
end

#verify ⇒ Object 



41
42
43
 
# File 'lib/llm_experiment/auth.rb', line 41

def verify
  @shell.sh(*run_argv(ready_image!, verify_script))
end