mix skill_kit.ralph (SkillKit v0.1.0)

Copy Markdown View Source

Run a Ralph loop against a TODO file.

# Loop on an existing TODO.md in the current directory
mix skill_kit.ralph TODO.md

# Generate TODO.md from a prompt, then loop
mix skill_kit.ralph TODO.md --prompt "Add JSON parsing to lib/foo.ex with tests"

# Choose a different agent (default: ralph)
mix skill_kit.ralph TODO.md --agent some-other-ralph

# Operate in a different working directory
mix skill_kit.ralph TODO.md --cwd path/to/project

How it works

This task is a thin driver. The Ralph contract lives in skills:

  • examples/agents/ralph/AGENT.md — agent identity, routes user requests to the right skill.
  • examples/agents/ralph/skills/plan/SKILL.md — generates a TODO file from a goal.
  • examples/agents/ralph/skills/iterate/SKILL.md — does one iteration: pick top unchecked item, do it, mark [x], commit.

The task sends a trigger message per turn — "Plan a TODO file at PATH for: GOAL" or "Iterate on PATH" — and waits for the agent to echo the skill's final word: PLANNED, CONTINUE, or DONE. The loop exits on DONE (or on an %Event.Error{}). There is no iteration cap and no per-turn timeout — pacing is handled by Anthropic.Client's 429 retry.

See guides/ralph-loop.md for the design.