forge-orchestrator

About

FORGE addresses the failure mode of running several coding agents at once: duplicated work, conflicting interfaces, and diffs nobody reviewed. A planner reads the codebase, asks clarifying questions, and produces a task DAG the user edits and approves; binding API/type contracts are generated before coding begins; one agent per git worktree executes in parallel with a health monitor watching for stuck tasks. Every change passes five gates (build, lint, test, LLM review, contract check) before auto-rebase, merge, and a single pull request via the GitHub CLI. Lessons captured from failed-then-retried tasks persist across pipelines, per-stage model routing balances cost and quality, and a Textual TUI plus Next.js dashboard expose state. Solo maintainers and small teams using Claude Code as their main harness are the audience.