Overnight Ralph Loop Runner

Yes — that's the Ralph loop pattern, wired to the CLI headless mode.
Here's how it works and the trap to set first.

THE MECHANISM
Instead of one-shotting a task, a shell while-loop re-invokes Claude in
headless mode (the -p flag) each iteration, feeding it the CURRENT file
state plus the remaining to-dos. Claude does a chunk, the loop re-reads
the files, and re-prompts with what's left. It converges on a finished
feature over many small passes rather than one giant one.

  while tasks remain:
    run claude -p with (prompt + current files + open to-dos)
    let it make progress and update the to-do list
    repeat

WHY IT WORKS OVERNIGHT
Each pass starts fresh from the real file state, so a mistake in pass 3
gets seen and corrected in pass 4 — the loop self-heals instead of
compounding. That's what lets it run unattended for hours.

THE GUARDRAIL TO SET BEFORE YOU SLEEP
A stop condition and a blast radius. Define "done" as a checkable
signal (tests green, to-do list empty), and scope it to a branch, not
main. An overnight loop with no exit condition and write access to main
is how you wake up to a mess.

WHAT YOU GET
Headless -p mode, the while-loop pattern, a to-do-driven convergence
loop, and a per-iteration log so you can read what happened over coffee.

Give me the feature and its "done" test, and I'll wire the loop.