Dev chains 3 AI models into a self-reviewing coding pipeline

A developer shared a workflow that combines Anthropic's newer model Fable, codex cli, and Claude Code to automate most of the coding process. Fable itself barely writes any code anymore; it acts purely as the orchestrator directing the whole process. The flow works like this: Fable first drafts a plan, another model called 5.6 sol reviews that plan in a loop until it's approved, then 5.6 luna does the actual implementation.

Fable then reads the entire diff, fixes anything it doesn't like, runs the tests, and hands it back to 5.6 sol to check the code against the original plan, repeating until approved. Finally, Fable handles the release chores itself — writing the changelog, tagging the version, and merging. The whole setup is just bash scripts calling codex cli with persistent conversation threads, invoked through skills, with no extra framework, no MCP, and no multi-agent swarm involved.

The workflow was pushed to GitHub, though the author warns most of it is bash and recommends having codex or Claude Code review the scripts before trusting them.