Fork Graph

See every fork of your Claude Code conversations as a git-graph — and jump back to "that good state" in one click.

Claude Code sessions branch constantly: /rewind, Fork conversation from here, --fork-session, resumed chats. After a week you have dozens of sessions and no way to remember which branch had the good result. Fork Graph reads your local session history, reconstructs the full fork genealogy, and lets you annotate, search, and reopen any point in it.

Features

• Fork genealogy as a git-graph — every prompt is a node; session-internal rewinds and cross-session forks are merged into one tree, even though Claude Code rewrites session/message IDs on fork
• Memos and pins — leave subjective notes ("this was the good state") on any node; they survive restarts, reloads, and forks
• Pin inheritance — everything descending from a pinned node (the same conversation and any forked branches) is pinned automatically, colored after the original pin so you can trace its lineage
• Markdown + math rendering — prompts in the detail panel render as Markdown with TeX math (KaTeX bundled, fully offline); one click back to raw text
• Open in Claude Code — reopen the session at any node, via the official extension or a CLI terminal tab (toggle; CLI is the default)
• Search — incremental text search over prompts + memos with in-page-find style ▲▼/Enter hit navigation (3/12 counter, wraps around), plus a calendar with multi-day and time-range filters; non-matching nodes dim so you keep the genealogy context
• Keyboard navigation — walk the tree from the detail panel with ◀▲▼▶ buttons or arrow keys: parent, child, sibling branches
• Pinned-only view — show just your pinned nodes and their ancestor paths, with lanes compacted
• Project switcher — browse the session history of any project folder, newest first; sessions from other folders open via CLI with the right working directory
• Soft delete with restore preview — hide a branch from the graph (your data is never touched); before restoring, preview the tree in a separate tab to see exactly where it reconnects
• /compact markers — a double wavy line (≈) crosses the edge where the context was compacted
• Auto reload — the graph follows your conversations live as session files grow
• Deleted-session awareness — sessions hidden in the official UI show as dashed pills and can still be opened via CLI
• 4 languages — English (default), 日本語, 中文, 한국어

UI at a glance

1. Launch icon in the editor title bar — or run Fork Graph: Open from the Command Palette
2. The fork graph — every prompt is a node; each lane is a branch born from a rewind or fork
3. Detail panel — full prompt, timestamp, memo & pin, soft delete, and Open in Claude Code
4. Text search across prompts and memos — non-matching nodes dim, so the genealogy stays visible
5. Calendar search — filter by date and time range (days that have chats are marked)
6. Menu — project folder switcher, language, pinned-only filter, restore soft-deleted branches

Getting started

1. Install the extension
2. Run Fork Graph: Open from the Command Palette (or click ⑂ Fork Graph in the status bar / the fork icon in the editor title bar)
3. Click a node → read the prompt, leave a memo, pin it, or press Open in Claude Code

Requirements

• Claude Code — sessions are read from ~/.claude/projects/
• To reopen sessions: the Claude Code CLI (claude) on your PATH (recommended; the default), and/or the official Claude Code for VS Code extension

Works in VS Code, Cursor, and Antigravity. The CLI route (default) works everywhere. The "open via the official extension" route is only available in VS Code, since that extension is VS Code–only — Cursor/Antigravity users simply use the CLI route.

Privacy

Fork Graph runs entirely on your machine. It reads your local session files (~/.claude/projects/**/*.jsonl) and never writes to them, and never sends anything anywhere. Memos, pins, and settings are stored in VS Code's local storage.

How it works (short version)

Claude Code rewrites sessionId/uuid when a conversation is forked, so files can't be joined by IDs. Fork Graph canonicalizes records by timestamp + type + content fingerprint (with a secondary merge by stable prompt/request IDs for the one record that forking re-stamps), which reconstructs the cross-file fork DAG regardless of read order. Details live in the design docs.

License

MIT