Chief: a task loop for Claude Code
Chief
A Go terminal dashboard that runs Claude Code through the tasks in prd.md, with a fresh context and one commit per task
- 474 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Claude Code starts with `--dangerously-skip-permissions`: the agent edits files and runs commands without asking for confirmation
- The loop makes commits to the repository on its own and, on long runs, spends subscription limits or paid tokens
Install
- Works in
Manual install
brew install minicodemonkey/chief/chiefThen run `chief new` in the project root and describe the tasks.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Claude Code starts with `--dangerously-skip-permissions`: the agent edits files and runs commands without asking for confirmation
- The loop makes commits to the repository on its own and, on long runs, spends subscription limits or paid tokens
This is third-party code. Review the repository files before installing.
What it does
Chief is a Go program with a terminal dashboard. You write the work on a large feature as a list of tasks in a prd.md file, and Chief launches an agent, Claude Code by default, once per task, one task per iteration. Every iteration starts with a fresh context. The agent records what was built and which files changed in progress.md, and the next iteration reads those notes. Each finished task leaves one commit with a message like `feat: [US-001] - Title`. The git history is easy to review task by task, and a failure traces back to a specific iteration. Task priority and status (`todo`, `in-progress`, `done`) live in prd.md itself. Claude Code can be swapped for Codex CLI, OpenCode, Cursor or Gemini CLI: the provider is set in `.chief/config.yaml`, with the `--agent` flag or with the `CHIEF_AGENT` variable.
Who it is for. Developers who hand Claude Code a large feature for several hours and want a chain of readable commits instead of one huge diff.
Good fit when
- A large feature splits into independent tasks, and the agent's context overflows in one long session
- Review works better commit by commit, each commit covering exactly one task from the list
- The loop runs in the background while you do other work, and the dashboard shows which task it stopped on
- You want to swap Claude Code for Codex CLI or OpenCode and keep the process: changing `provider` in `.chief/config.yaml` is enough
Not a fit when
- There is one small task: a PRD and a commit per task only slow things down, an ordinary agent session is faster
- Requirements are not written down: the loop follows a finished task list in prd.md and does not invent tasks
- You want to control every step: iterations run one after another until the list is empty
Example request
Write registration, login and password reset as separate tasks in prd.md, then start `chief` and show which tasks are doneLimitations
An installed and signed-in agent is required: Claude Code, Codex CLI, OpenCode, Cursor or Gemini CLI. Chief itself gives no access to models.
In the Claude provider code the agent starts with the `--dangerously-skip-permissions` flag, so all its actions run without confirmation. Work in a separate branch or a sandbox.
Installation goes through `brew` or the `install.sh` script. The documentation lists no npm or apt package.
How to disable. Stop the `chief` process and delete the `.chief/` directory in the project. Remove the program the way you installed it: with `brew`, or by deleting the binary that `install.sh` placed.
FAQ
How do I install Chief for Claude Code?
Run `brew install minicodemonkey/chief/chief` or the `install.sh` script from the repository. Then call `chief new` in the project root, describe the tasks and open the dashboard with `chief`. Claude Code must be installed and signed in.
What is the Ralph Wiggum loop in Chief?
Each iteration starts the agent with a fresh context and takes the next task from prd.md. Progress between iterations is kept in progress.md, so the context does not bloat and the work is not lost.
Can Chief run with Codex or OpenCode?
Yes. Put `provider: opencode` in the `agent` block of `.chief/config.yaml`, run `chief --agent opencode` or set `CHIEF_AGENT=opencode`. Valid values are `claude`, `codex`, `opencode`, `cursor` and `gemini`.
Where does Chief keep project data?
The `.chief/prds/<name>/` directory holds prd.md with tasks and statuses, progress.md with the agent's notes and the `claude.log` journal with raw output (other providers use a different file name).
How many iterations will Chief run?
When `--max-iterations` is not set, the limit is computed from the number of remaining tasks plus a buffer. You can change it with the `+` and `-` keys in the dashboard.
Discussion
No questions or reviews yet.
Related
CLI toolsA bash loop that restarts Claude Code or Amp with a clean context until every PRD story passes its checks, plus two skills for writing the PRD
Claude Code kit with eight phase agents, 21 skills and hooks that keep the agent inside set limits.
A CLI that keeps agent instructions in one place and distributes them to the right config files for Claude Code, Cursor, Codex, Copilot and two dozen other tool
A Claude Code working configuration: a planner with per-milestone review, sub-agents and skills for analysis, refactoring and docs sync
A Rust coding agent with a native GUI, a terminal mode, an ACP mode for Zed and an MCP server
A Claude Code plugin: 39 commands for PRs, review, commits and specs, nine expert roles and hooks, in eight language versions
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Chief: a task loop for Claude Code (this card) | no | 474 | |
| Ralph TUI | no | 2,467 | a terminal dashboard that drives seven agents through a task list in a loop; Chief keeps the list in prd.md and separates iterations by a fresh context and one commit per task |
| Open Ralph Wiggum | no | 1,906 | the `ralph` command repeats one prompt until a completion marker appears; Chief hands the agent the next task from prd.md |
README in short
The project page briefly covers installation, the `chief new` and `chief` commands, the Ralph Wiggum loop and the list of supported agents. Details live on the documentation site. MIT license.