Foreman
Runs software missions without you in the loop, and shows you everything when you come back. A director agent plans, delegates to workers, verifies the result on its own terms, and reports. Any model provider, cost that is never invented, a default for every question, and your phone for the ones that matter.


What it is
You link a project folder and say what you want. A director agent talks the brief through with you, or skips the talk if you tell it to, writes the plan into a mission doc with a DONE WHEN checklist, delegates the work to worker sessions, watches them, and verifies the result before reporting. One dashboard shows what needs you, what is running with its crew and meter, and what finished.
The agents are real Claude Agent SDK sessions, but the models behind them can be Anthropic’s or anything that speaks OpenAI’s wire: a local Ollama, OpenRouter, vLLM, a Codex login. The director and the workers can run on different providers. Foreman never holds a credential it did not find already on your machine, and it never guesses a price: a run is priced only when the endpoint that sends the bill also published the rate. Otherwise the meter shows tokens and turns, and says so.
It is built to run while you are away. Every ask has a deadline and a default, so an unanswered approval is denied with a pointer and an unanswered question goes back to the director with “decide and record”. Watchdogs bound a silent worker, a looping director, or a run past its wall clock. Link the Telegram bot and approvals arrive on your phone with buttons; over Tailscale, a dev server the crew exposed is one tap away.
Every run persists as an append-only event log, replayed through the same reducer that renders live events. A refresh shows exactly what a watcher saw, and an interrupted run restores the director’s session and re-verifies before it continues.
npx @amenophis1er/foreman