Foreman

Foreman is a local terminal app for developers who want Claude Code agents to carry a feature through planning, implementation and verification. It runs against your repository, with human approval of design documents and the issue queue before autonomous coding begins. The Python project uses the MIT license and is tested on Linux and WSL2. Real agent work requires an authenticated Claude Code CLI.

Review gates are its main distinction. Foreman withdraws approval when an approved document changes, and unresolved design questions block approval. It breaks accepted requirements into issues with dependencies, then assigns parallel workers to separate Git worktrees while preventing overlapping file changes.

Before merging, Foreman checks acceptance criteria, runs tests, lint and type checks, and rejects regressions. A separate read-only evaluator reviews each diff; optional code and security reviewers can also block it. Workers can't edit their own verification records. Repeated disagreements and uncertainty return to a human, while time, turn and spending limits constrain agent runs.

The keyboard-driven interface shows progress, live worker logs, costs and requests for human input. Workflow state stays in readable files inside the target repository, so interrupted builds can resume from disk without a database or server.

Run metrics expose recurring failures. Foreman can propose changes to agent skills and prompts, but those changes need human approval and a benchmark report before acceptance. Its demo and test suite use a mocked backend and work offline.