Git-style conversation branching for Claude Code. 中文文档

When you're deep in a conversation and hit an unfamiliar concept, open a branch to explore it. When you're done, squash it back — the parent branch gets a concise summary without the detour polluting its context.

main ●                         "Reading a PagedAttention paper"
└─ learn-mha ◈                 "How does Multi-Head Attention work?"
   └─ why-sqrt-dk ◈            "Why divide by √d_k?"

Each squash compresses an exploration into 1-3 sentences that bubble up to the parent. The parent sees only conclusions, not the journey.

git clone https://github.com/v3nividiv1ci/cit-mcp.git && cd cit-mcp
./setup.sh

setup.sh handles everything:

• Creates Python venv and installs dependencies
• Registers the MCP server in ~/.claude.json
• Registers session hooks in ~/.claude/settings.json

Restart Claude Code after setup.

Requirements: Python 3.11+, tmux (for multi-pane layout)

You're reading a paper and keep hitting unfamiliar concepts:

┌──────────┬──────────┬──────────┐
│          │          │          │
│  main    │ learn-mha│ why-sqrt │
│  (paper) │ (concept)│ (detail) │
│          │          │          │
└──────────┴──────────┴──────────┘

1. main: Reading a paper, hit "Multi-Head Attention"
2. cit branch learn-mha → new pane opens, explore MHA
3. Hit "√d_k scaling" → cit branch why-sqrt-dk → go deeper
4. Understand it → cit squash → auto-returns to learn-mha with summary
5. Finish MHA → cit squash → auto-returns to main with summary
6. cit inbox → main gets both summaries, continues the paper

Why not sub-agents? Sub-agents do "fetch me an answer." Branching is for interactive exploration — you ask, follow up, go deeper, at your own pace.

Children split right, siblings split down:

┌──────────┬──────────┐
│          │ child-A  │
│  main    ├──────────┤
│          │ child-B  │
└──────────┴──────────┘

Remove the cit entry from ~/.claude.json (mcpServers.cit) and the hook entries from ~/.claude/settings.json, then delete the repo.