AgentSCM 推出 alpha 版，將 Agent 衝突偵測提前至編輯階段
AgentSCM - Catch agent merge conflicts at edit time, not merge time.
AgentSCM 開放 alpha 版供預覽及測試，會在 Agent 編輯檔案前查詢其他協作者的處理情況，並在合併前檢查分支衝突。首次重疊編輯不會套用，Agent 會收到檔案持有人及其任務資訊；合併檢查採唯讀方式，使用 git merge-tree 判斷實際衝突。
Today we're soft-launching the alpha version of AgentSCM for preview and testing.
Sub-agents and other team mates contributing to the same source git repo can cause issues not seen until merge or PR time.
There are still questions on whether the usual git workflow survives automated agentic coding long-term, but in the interim we are building AgentSCM to build a community of builders looking to work together, as native as possible embedded in their daily workspace(s), to catch conflicts between sub-agents and other contributors' agents in real-time.
- At the edit: before an agent edits a file, its hook asks a local hub who else holds it. The first overlapping edit isn't applied, and the agent is told who holds the file and their task. It can work elsewhere, ask for a handover, ask you, or repeat the edit, and the repeat always goes through.
- Before merge: agentscm collisions checks your branches and worktrees and asks git merge-tree which merges would actually conflict. Read-only.
- No migration: the biggest AI labs are already rethinking source control for agents, which is great to see. You don’t have to make that call yet: keep Git, and two-way Git sync already exists as experimental.
- Future state: full bi-directional git sync to help you decide when, or if, an agent source code manager should replace your git under scaling pressure. And more workflow tools focused on hundreds or thousands of agent edits, not just human commits/merges/PRs.
If you don't want to install but are curious to see it in action: https://agentscm.dev/demo/
To set it up for real: npx agentscm (guided, asks before each change; agentscm uninstall removes it).
FAQ
Isn't this just worktrees?
Worktrees isolate agents. They don't tell them about each other, so you still find out at merge.
Won't the agent just retry?
Yes, on purpose. A wrong refusal costs one retry, not a stuck agent. The point is the agent learns who's on the file and why, at the moment of the edit.
What agents work now?
Claude Code, Codex and Copilot, through hooks. Cursor gets a note right after an edit instead of a block. Any MCP-capable editor can connect to the hub to see who holds which files and announce its own work.
Is it a file lock?
No. Nothing is locked, and if the hub is down or slow the edit proceeds.
Telemetry?
None in the alpha yet. We plan anonymous usage stats soon with an opt-out during both the self-install process and in the settings / cli flag.
License?
FSL-1.1-ALv2
來源：DEV Community · dev.to