Questions: Multiple agents / hooks

[gmarcus]

Interesting project.
I have some questions...

1. If I am already using checkpoints with Claude Code, is there a different benefit that comes from your MCP?

2. And what happens when running parallel agents in the same worktree vs. in separate worktrees - does each work tree get it's own shadowgit folder?

3. Instead of an autosave timer, would it be better to use hooks to coordinate with AI events?

[blade47]

Hey @gmarcus, apologies for my late reply!
Answers below:

1. Claude Code checkpoints are restore points, useful for rollback, but the AI can't search or reason over them. ShadowGit captures continuous history AND exposes it via MCP, so Claude can answer questions like 'what did I change in the auth module in the last hour?' or 'show me the diff before I broke the tests.'
2. Currently each project root gets its own .shadowgit/ folder. For worktrees, the cleanest approach would be a shared ShadowGit instance at the main repo level that tracks all worktrees, with branch-aware filtering in the UI and MCP queries. Does that match how you're using worktrees?
3. I explored both. Git hooks only fire on commits, useless here since the whole point is capturing changes before you commit. Filesystem watchers (fsevents/inotify) sound elegant but in practice they're heavy on resources and unreliable: they miss events under load, flood you during bulk operations, and behave inconsistently across OS/editors. A simple 15-second timer is boring but I found being bulletproof.

[gmarcus]

Hi @blade47 No worries. I forgot about this project and my questions :). The firehose of AI tools/announcements is hard to keep up with.

I am still trying to sort out best practices for my workflow. I am trying to get to Agentic first - isolation, parallelism with local and remote sessions. Instead of babysitting conversations, default to delegating work.

Ideally, I want to be able to kickoff local or remote Claude Code sessions.

• Each session would live in isolation (local using git work trees, remote is already isolated by a fresh clone of current repo).
• Within an isolated session, all changes tracked and committed to git. I don't need point-in-time autosaves per se, but am thinking every turn of work done. Ideally, the commit message would include the user prompt and the assistant response (when it has completed work)

Assuming there could be 5-10 sessions running in parallel (some local, some remote), it would be using worktree/branches, PRs and merges that allows changes to propagate.

How would you imagine shadowgit-mcp fitting into a setup like this?