ralph-loop

Start Ralph Reviewed Loop

Initialize an iterative development loop with Codex review gates.

Parse Arguments

Arguments: $ARGUMENTS

Parse the following from arguments:

• PROMPT: Everything before the first -- flag (the task description)
• --max-iterations: Number (default: 30)
• --max-reviews: Number (optional, defaults to --max-iterations if not specified)
• --no-review: Boolean flag (default: false)
• --debug: Boolean flag (default: false)

Setup

1. Ensure the rl CLI is installed. If not on PATH, build from source:

   command -v rl >/dev/null 2>&1 || (
     echo "Installing rl CLI..." &&
     git clone https://github.com/0xbigboss/rl /tmp/rl-build 2>/dev/null &&
     cd /tmp/rl-build &&
     bun install --frozen-lockfile &&
     mkdir -p ~/.local/bin &&
     bun build src/cli.ts --compile --outfile ~/.local/bin/rl &&
     rm -rf /tmp/rl-build &&
     echo "rl installed to ~/.local/bin/rl"
   )
   

   Verify it's available:

   rl --version
   

2. Initialize the loop (creates .rl/ with state.json, prompt.md, and .rl/rl wrapper):

   rl init "{PROMPT}" --max-iterations {MAX_ITERATIONS} --max-reviews {MAX_REVIEWS} {--no-review if set} {--debug if set}
   

3. Verify setup:

   .rl/rl status
   

All subsequent rl calls use .rl/rl (wrapper created by init).

Completion and Escape

• Done: run .rl/rl done — triggers Codex review on next stop.
• Blocked: run .rl/rl done --blocked — terminates without review.

Working Guidelines

Pacing. Each iteration should produce thoughtful work. Researching, loading skills, and studying patterns IS productive — don't rush to .rl/rl done.

Churn breaker. If a reviewer flags the same area twice, your next iteration must be research — load skills, read docs, study the codebase. No code fix until you understand why the previous fix was wrong.

Depth before breadth. Complete each phase fully before starting the next.