HN Showcase
Back to browse
Stoneforge – Open-source orchestration for parallel AI coding agents

Stoneforge – Open-source orchestration for parallel AI coding agents

by adamjking3·Mar 5, 2026·1 point·1 comment
Visit ProjectView on HN

AI Analysis

●●●BangerZero to OneWizardryRabbit Hole

Multi-agent coordination with git worktrees prevents Claude conflicts and context re-hashing.

Strengths
  • •Worktree isolation + handoff notes solve a real scaling pain (context switching tax grows with agent count)
  • •Director → Workers → Stewards workflow with context-aware squash-merge is architecturally coherent, not ad-hoc glue
Weaknesses
  • •Niche audience: only developers already running 3+ parallel agents on same repo; growth limited to emerging multi-agent coding workflow
  • •Heavy git/tooling complexity; unclear if DX survives first contact with real CI/CD pipelines and large teams
Target Audience

Developers running multiple Claude/LLM coding agents on the same codebase simultaneously

Similar To

LangGraph · AutoGen

Post Description

I built this because I was running 3-5 Claude Code instances on the same repo and burning out from constantly context switching between terminal windows. Carefully making sure agents' work didn't overlap, preventing context windows from degrading, manually enforcing documentation/memory policies, and re-explaining decisions across sessions.

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.

Similar Projects