Foreman runs a coding project with one lead agent and a crew of workers, in any coding agent that has skills and sub-agents, such as Claude Code or Kimi Code. The lead plans and writes a brief for each task. A worker takes one brief, opens one pull request, and stops. Every agent stays small and its cache stays warm, so your usage limit lasts the week.
Every worker, hour by hour
Each block is one agent, and its height is the context it re-read on every call. The lead is the long grey band. The workers under it are short and low, which is the point. Hold Ctrl or ⌘ and scroll to zoom, drag to pan. Projects, pull requests and tasks are anonymous.
Row height is context: 0 to 1M tokens for the lead, 0 to 600k for workers. Zoom in to see a single worker. Times are in your time zone.
Where it pays off
Who did the work
How large each agent grew
Five kinds of worker, one job each
You start the Foreman skill in a session, and that session becomes the lead. It does not write code, debug or run tests. It sends one of these, with a brief. The model column shows what these workers ran on; each profile sets a default, and a repository can override it.
| Worker | Job | Model, effort | Sent | Median peak |
|---|
Install it in your agent
Foreman is a few skills and five worker profiles in plain Markdown. Any coding agent that loads skills and sub-agents can run it. Here are two.
Claude Code
git clone https://github.com/chengluyu/foreman.git ~/.foreman
cp -R ~/.foreman/skills/* ~/.claude/skills/ && cp ~/.foreman/agents/foreman-*.md ~/.claude/agents/
{ "subagentPromptCacheTtl": "1h" }/foreman
Kimi Code
/plugins install https://github.com/chengluyu/foreman
/reload
/skill:foreman
Already cloned somewhere else? Link it instead: ln -s /path/to/your/foreman ~/.foreman. Another agent: copy the skills/ folders to wherever it loads skills, and agents/foreman-*.md to wherever it loads sub-agents. The first time in a repository, the lead asks four questions and writes the answers into the repository's handover file.