Releases: EvanZhang008/open-walnut
Release list
v0.2.0 — Notes vault, diff review & live workflow visualization
The first major update since the initial release — 503 commits that turn Open Walnut from a task-and-session dashboard into a full AI-native workspace.
Highlights
- 📝 Notes vault — a Notion/Obsidian-class multi-file notes system with a TipTap WYSIWYG editor, wiki-links, tables, image paste, attachments, slash commands, and hybrid (semantic + keyword) search. The agent reads and writes your notes as first-class context.
- 🔍 Per-session diff review — a GitHub-style "Changed" view for every session: review exactly what the agent changed, leave line-range comments, and reply inline.
- 🌊 Live workflow visualization — dynamic multi-agent workflows render as a real-time phase flow-graph with per-subagent drill-in and full transcripts; reconstructed on reload.
- 🛰️ Resilient remote-session daemon — a new transport architecture (bun binary + SSH tunnel + FIFO) keeps remote Claude Code sessions alive across tunnel/daemon crashes, with chunked auto-deploy that survives corporate SSH proxies.
- ⚡ Zero-config onboarding — a unified credential resolver auto-detects Bedrock/Anthropic credentials from config,
settings.json, env, and~/.aws; three setup paths get you running out of the box. - 💬 Multi-conversation agents — each agent now has multiple independent conversations (fresh context per tab) with automatic knowledge distillation into agent memory.
What's new
Notes & Knowledge — WYSIWYG editor (wiki-links, tables, image paste, per-line Tab indent, paste-URL-as-hyperlink, attachments), Obsidian-like folder tree with drag-to-move + breadcrumbs, hybrid search with relevance bands, /task slash references, notes injected into the main agent + context inspector, repository memory layer, and a working-memory scratchpad.
Sessions — per-session Changed view (diff, line-range comments, three compare modes), dynamic-workflow visualization (live flow-graph, drill-in, full-screen, reload reconstruction), resilient remote daemon, embedded terminal (persisted via dtach), VS Code-style file explorer + file intelligence, Quick Start panel, instant session switching, session retry, and the /session command with fuzzy path picker + live SSH auto-complete.
Tasks & Focus — 3-tier pinned system (Focus / Next / Satellite + Wait), virtual task groups, Fork-in-Walnut with multi-level child nesting, HUMAN_VERIFIED / POST_WORK_COMPLETED phases, sprint as a first-class citizen, Quick Add, date pills/pickers, and cross-source task migration.
Agent & Models — multi-conversation per agent + on-demand creation + memory distillation, Opus 4.8 added and set as the default, catalog-driven model resolution with adaptive thinking, new tools (ask_question, pin_task, files_glob/files_grep, unified files_*), Execution / Plan mode toggle, Claude Code Teams, and a Skills page (local + remote discovery).
Onboarding, Providers & Voice — zero-config onboarding, AI Providers settings with provider adapters + Ollama dynamic models + Tavily web search, auto-save settings, Speech-to-Text on all inputs, system-audio capture, and a whisper-server daemon engine with VAD + prompt biasing.
Observability — forensic wide-event recorder, invariant engine, auto-incidents, session self-report, and an in-app notification store + toaster.
Performance & stability
- Real-token compaction gate fixes the "context never converges / hits 1M" failure.
- Lazy-load subagent content (fixes 40s session loading); zero-remount full-screen.
- Event-bus interest set fixes event-loop starvation; read cache + slimmer payloads turn 15s timeouts into tens of milliseconds.
- 246 fixes spanning session-status correctness (false-zombie kills, fake
session:error, mid-turn QUEUED stalls, premature idle completion), notes flicker/drag, daemon reconnection/replay, and STT mic detection.
See CHANGELOG.md for the full list.
Getting Started
git clone https://github.com/EvanZhang008/open-walnut.git
cd open-walnut
npm install
npm run build
cd web && npx vite build && cd ..
walnut webOpen http://localhost:3456 and start chatting with your AI butler. See the Getting Started Guide.
Requirements
- Node.js >= 22
- Claude API access (AWS Bedrock or Anthropic API)
- Optional: Ollama with BGE-M3 for semantic search
Full changelog: v0.1.0...v0.2.0
v0.1.0 — Initial Release
First public release of Walnut — a personal AI butler powered by Claude.
Highlights
- Claude Code Web UI — spawn, monitor, and chat with Claude Code sessions from a real-time web dashboard
- 4-layer task hierarchy — Category → Project → Task → Subtask with 7-phase lifecycle
- AI agent with 30+ tools — task management, memory, sessions, search, cron, coding, and more
- Persistent memory system — per-project memory, daily logs, session summaries, full-text + semantic search (SQLite FTS5 + BGE-M3 embeddings)
- Multi-session orchestration — run up to 7 local + 20 remote Claude Code sessions simultaneously
- Local-first — all data stored in
~/.walnut/as JSON, Markdown, and SQLite - Self-hosted — single
npm install && npm startto run everything - Heartbeat system — periodic AI self-check with configurable schedules
- Cron jobs — schedule recurring agent tasks with cron expressions or intervals
- Plugin system — Microsoft To-Do two-way sync, Slack notifications, extensible plugin architecture
- Git-sync — automatic version-controlled backup of all task and memory data
- CLI + Web — full CLI for terminal users, React SPA for the browser
Getting Started
git clone https://github.com/EvanZhang008/walnut.git
cd walnut
npm install
npm run build
cd web && npx vite build && cd ..
walnut webOpen http://localhost:3456 and start chatting with your AI butler.
Requirements
- Node.js >= 22
- Claude API access (via AWS Bedrock)
- Optional: Ollama with BGE-M3 for semantic search