ChaseSunstrom/foreman

The discipline layer for coding agents.
Claude Code, Codex, Gemini CLI, opencode, or its own fm code on any model.
It plans before it edits, proves its work with commands it really ran, keeps to one task at a time
and works through the queue, while a guard stops the commands you'd never want run in bypass mode.
Install ·
How it works ·
The guard ·
Features ·
Any model ·
Docs ·
Site
A real session in a copy of this repo: one typed request, then about two minutes of Claude working through Foreman (sped up while the model thinks). Its fix is in Foreman now (T-1213). MP4
## Why
Coding agents are fast. Left alone, they also edit before they understand, say "tests pass" without running them, chase new ideas mid-task, lose their place after a compaction, and in bypass mode will run the one command you'd never type.
Foreman is the foreman on that job site, whichever agent works it: Claude Code, Codex, Gemini CLI, opencode, or fm code on a model of your choice.
| Foreman | So that |
| Plans first | Every request becomes a brief before any edit: type, size, what "done" means and the command that proves it. No task, no edit. |
| Proves it | Evidence is a command fm ran and recorded, never a sentence. A fix needs a test that failed first. Nothing closes without its audits. |
| Guards bypass mode | A fail-closed guard stops wide deletes, force-pushes, credential reads and curl \| sh, and saves uncommitted work before git throws it away. |
| Stays on task | One active task. New ideas go to the inbox instead of being chased; NOW: switches on purpose. |
| Never loses its place | State lives on disk, so after /compact or in a new session Claude resumes at the exact step. A revert reopens the step it undid. |
| Keeps going | Drive mode works the whole queue, then finds its own next work. fm run and fm serve keep it going headless. |
| Learns | Your corrections and Foreman's own friction become lessons, tripwires and fixes. |
## How it works
Every request, however terse, goes round the same loop:
Understand. Intake turns your words into a brief: its type, its size (S, M or L), what done means with the command that proves it, and the steps.
Build. Edits wait for the brief, and every command passes the guard.
Prove. A task closes only with fresh evidence and its audits.
Learn. Lessons and decisions are recorded, then the next queued task starts by itself.
## Quick start
bash
curl -fsSL https://raw.githubusercontent.com/ChaseSunstrom/foreman/main/install.sh | bash
Or from inside Claude Code:
/plugin marketplace add ChaseSunstrom/foreman
/plugin install foreman@foreman
| Then | How |
| Ask for something | fix the login timeout on slow wifi, in plain words. That's the whole interface. |
| Watch it first | fm demo runs the real fm and guard hook on a seeded bug in a throwaway repo: no model calls, nothing left behind. |
| See where things stand | Bare fm in a project says what's next; fm help answers "can it…?" in plain words. |
Starting a page, a 3D scene or an fm code panel? Start from a template that already passes the check that keeps it strict:
bash runnable
fm help can foreman scaffold an mcp server # the commands whose help answers it, or a plain no