Self-improving Codex loop

Three models, two role files, one driver. DeepSeek runs the task from prompt.md, Luna grades each run, Sol rewrites prompt.md, and the cycle repeats until the task is done. The prompt file is the product, everything else is plumbing.

Layout

Path	What it is
loop/	The driver (loop.sh) plus watcher.md, rewriter.md, prompt.md
skills/	Slash-command skills: /loop, /flash, /deepseek, /luna, /sol, /build, /utility
demo/	Runnable demo: a threaded counter bug fixed by the loop, with evidence
install.sh	Copies the skills into ~/.codex/skills

Quick start

1. Install the slash commands:

   bash install.sh

2. Restart Codex.

3. Type /utility to see the menu, or /loop <task> to start the loop.

Run the demo

cd demo
bash ../loop/loop.sh

The loop stops when the runner reports DONE:, after 6 cycles, or after two consecutive cycles leave prompt.md unchanged. --every 600 gives a 10-minute cadence. See loop/README.md for all options and the portable setup recipe for machines without the router.

Slash commands

Command	Model / role
/loop <task>	DeepSeek runs, Luna watches, Sol rewrites
/flash <task>	opencode-go/deepseek-v4-flash
/deepseek <task>	opencode-go/deepseek-v4-pro
/luna <task>	gpt-5.6-luna
/sol <task>	gpt-5.6-sol
/build <task>	opencode-go/deepseek-v4-pro with verification
/utility	Prints this mapping

The model slugs assume the local Codex Router catalog. On other machines, replace them with any model codex exec -m can reach.