Analysis updated 2026-05-18
Automate multi-step coding tasks where one model plans and reviews while another writes the code.
Keep a goal-shaped task looping until tests pass, with an independent model judging each attempt.
Get a second AI model's code review before merging a change.
| tomascupr/sous-chef | nikvpn-iran/nikvpn-codespace | filipedeschamps/dotfiles | |
|---|---|---|---|
| Stars | 65 | 65 | 67 |
| Language | Shell | Shell | Shell |
| Last pushed | — | — | 2025-04-23 |
| Maintenance | — | — | Stale |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | developer | general | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires Codex CLI 0.134 or higher, signed in with a ChatGPT subscription.
sous-chef is a Claude Code plugin that splits coding work between two AI models, the way a kitchen splits work between a head chef and a line cook. One model, the head chef, plans the task, writes out what needs to be done, and reviews every change line by line. Another model, the worker, does the actual implementation and has no say in what ships. The plugin is built around Claude models for planning and review, while Codex CLI (a separate coding tool made by OpenAI) or GLM models handle the implementation. The plugin offers two main commands. The first, called serve, is for tasks that are clearly scoped: it writes a ticket describing the work, sends it to the implementer, reviews the resulting change, fixes any problems found, and verifies the result, all within a limited number of implementation runs. The second, called simmer, is meant for open-ended goals, such as making all the tests pass. It repeats the implement-and-check cycle until the goal is met, with limits on how many attempts it will make and a check for when no more progress is happening. There are also more manual commands for people who want finer control: one to write a ticket and run a single implementation pass, one to run a review without making any changes, one to turn confirmed review findings into a scoped fix, and one for first-time setup, which checks that Codex CLI is installed and logged in. To use the plugin, a person needs Codex CLI version 0.134 or higher, signed in with a ChatGPT subscription rather than a separate API key. Installation happens through the Claude Code plugin marketplace command, followed by a one-time setup command run inside the target repository. The project explains its design choices by pointing to specific incidents, including a case where a polling loop against a background job used up a large share of a weekly usage quota without producing results, and a test run where several AI-generated reviews turned out to be wrong. The license is MIT, so the code can be used, modified, and shared freely, including in commercial projects.
A Claude Code plugin where one AI model plans and reviews coding tasks while a second model, like Codex CLI or GLM, does the actual implementation.
Mainly Shell. The stack also includes Shell, Claude Code, Codex CLI.
Use, modify, and distribute freely, including for commercial projects, as long as you keep the copyright notice (MIT license).
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.