Stoneforge is the coordination layer I wanted. A Director agent breaks goals into tasks. A dispatch daemon assigns them to workers when available, each task runs in its own git worktree. Stewards review completed tasks and squash-merge to main if everything passes inspection, otherwise they handoff the task with review comments to be picked up by a new agent. When a worker hits its context limit, it commits, writes handoff notes, and exits, so the next worker can pick up on the same branch with a fresh context window and important notes from the previous agent's work.
Some design decisions that might be interesting to this crowd:
- Fully event-sourced with a complete audit log. - Supports syncing tasks to Github or Linear, and documents to Notion, Obsidian, or a local folder. Custom providers are also supported. - JSONL as source of truth, SQLite as a disposable cache. JSONL diffs, merges across branches, and survives corruption. SQLite gives you FTS5 and indexed queries. The SQLite .db can be rebuilt on different devices in seconds. - No approval gates by default. If five agents each need confirmation for every file write, you won't be moving any faster. Review happens at the merge steward level. - Worktrees over containers. The conflict surface for coding agents is git and the file system, containers or remote instances are overkill. Worktrees create in milliseconds, share node_modules and build caches, and don't need Docker or separate servers. - You can run multiple Claude Code / Codex plans simultaneously on the same codebase.
Works with Claude Code, OpenAI Codex, and OpenCode. Apache 2.0. GitHub: https://github.com/stoneforge-ai/stoneforge
Happy to discuss the architecture or any of the tradeoffs.