foreman

Foreman's premise is that the orchestrator, not the model, should hold the guardrails — and the implementation backs that up in concrete ways. The PreToolUse deny hook stops a worker from writing its own verification, which is the failure mode that quietly ruins most TDD-with-agents experiments. Budgets are enforced per run on turns, cost, and time, with a separate daily ceiling that hard-stops work rather than warning. Hash-sealed approvals mean a design document that changes after sign-off invalidates the approval and the run reverts, so you cannot approve an ADR and then have the build proceed against a different PRD. Operationally, the design choices are conservative in a useful way. The pipeline is fixed — plan → ADR/PRD → issues → TDD build → e2e — with human gates on every design phase, and everything is committed as human-readable files in your repository with no database, so a killed run recovers from disk. Parallel workers get separate git worktrees and a declared touches set that gates their footprint, which is the practical answer to two agents editing the same file. Configuration is a config.sample.yaml, and the CLI exposes foreman init, build, status, retro, and bench. The evals flywheel is the part competitors usually skip. Every run is outcome-labelled, foreman retro clusters recurring failures into skill and prompt patches, and those patches must pass foreman bench — which replays the eval set and reports success-rate, cost, and turn deltas — before they land. You can also exercise the whole pipeline at zero token cost with foreman --demo against a throwaway sample repo on canned stream-json, which is the cheapest way to walk every gate before spending real budget. The trade-offs are real. You bring your own authenticated claude CLI, so Foreman is tied to headless Claude Code and is not model-agnostic. Platform support is Linux/WSL2, which puts macOS and Windows users behind an emulation layer. Everything runs from a keyboard-only Textual TUI with YAML configuration, so effective use assumes terminal and git comfort. And the gated workflow is a process commitment, not a toggle — runs pause for sign-off rather than operating fully autonomously, and the orchestrator will halt work when a budget ceiling is hit. At 444 stars and 116 forks with 14 commits, this is an active early-stage open-source project, not a vendor with a support contract.