niwa(庭)

Async conflict-aware spec/planning, for me, you and your LLM/agent

# Recommended: installs CLI globally
pipx install niwa
# or
uv tool install niwa

Niwa (庭, "garden") is a CLI tool that enables multiple LLM agents to collaboratively edit markdown documents with automatic conflict detection and resolution. Like a zen garden where gravel is raked into patterns, Niwa helps agents weave their edits together harmoniously.

Built on LMDB for high-performance concurrent access, with full GitHub Flavored Markdown support via markdown-it-py.

Features

• Concurrent Editing: Multiple agents can read/edit simultaneously
• Conflict Detection: Automatic version tracking detects when agents edit the same section
• Smart Merging: Auto-merge suggestions for compatible changes, detailed diffs for conflicts
• Sub-Agent Support: Stored conflicts survive context switches, status commands for fresh context
• Version History: Full audit trail with rollback capability
• Search: Find content by keyword when you don't know the node ID
• Full Markdown Support: GFM tables, task lists, footnotes, frontmatter, and more
• LLM-Friendly: Comprehensive error messages with usage guides
• Claude Code Integration: Hooks to inject usage context on session start and compaction

Installation

# Recommended: install as global CLI tool
pipx install niwa
# or with uv
uv tool install niwa

# Alternative: install in a virtual environment
uv venv && source .venv/bin/activate
uv pip install niwa

# Or from source
git clone https://github.com/secemp9/niwa
cd niwa
pip install .

# For development (editable install)
pip install -e ".[dev]"

Dependencies

Niwa automatically installs:

• lmdb - Lightning Memory-Mapped Database for storage
• markdown-it-py - CommonMark-compliant markdown parser
• mdit-py-plugins - Frontmatter, footnotes, definition lists, task lists
• linkify-it-py - Automatic URL detection

Markdown Support

Niwa uses markdown-it-py with the GFM-like preset for robust markdown parsing. Unlike regex-based parsers, this properly handles:

Feature	Support
ATX Headings (# H1)	Full
Setext Headings (underline)	Full
Fenced Code Blocks	Full (with language hints)
Indented Code Blocks	Full
GFM Tables	Full (with alignment)
Task Lists (- [ ])	Full
Strikethrough (~~text~~)	Full
Autolinks	Full
Footnotes ([^1])	Full
Definition Lists	Full
YAML/TOML Frontmatter	Full
HTML Blocks	Preserved
Nested Lists/Quotes	Full

Key benefit: # characters inside code blocks, HTML comments, or inline code are correctly ignored - they won't be mistaken for headings.

Quick Start

# 1. Initialize database
niwa init

# 2. Load a markdown file
niwa load document.md

# 3. View structure
niwa tree

# 4. Read a section (as an agent)
niwa read h2_3 --agent claude_1

# 5. Edit
niwa edit h2_3 "New content" --agent claude_1 --summary "Updated section"

# 6. Export back to markdown
niwa export > updated.md

Commands

Setup

Command	Description
init	Initialize a new database
load <file>	Load markdown file into database
check	Verify database health
setup claude	Set up Claude Code hooks integration
setup claude --remove	Remove Claude Code hooks

Browse

Command	Description
tree	Show document structure with node IDs, token counts, and AST summary
search <query>	Find content by keyword

Edit

Command	Description
read <id> --agent <name>	Read for editing; large nodes show structural overview first (use --all for full content, --section N for a section, --lines M-N for a line range)
edit <id> <content> --agent <name>	Edit a node
resolve <id> <resolution> --agent <name>	Resolve a conflict
title <id> <title>	Update node title
summarize <id> <summary>	Add summary to node

Agent Status

Command	Description
status --agent <name>	Check pending reads, conflicts, recent edits
conflicts --agent <name>	List unresolved conflicts
agents	List all agents who've used this DB
whoami	Get suggested unique agent name

History & Undo

Command	Description
history <id>	View version history
rollback <id> <version> --agent <name>	Restore to previous version
dry-run <id> <content> --agent <name>	Preview edit without applying
cleanup	Remove stale pending reads/conflicts

Output

Command	Description
export	Export database to markdown
help	Show full usage guide

Multi-Agent Workflow

Agent A: reads section (v1)          Agent B: reads section (v1)
    │                                     │
    ▼                                     ▼
Agent A: edits                       Agent B: edits
    │                                     │
    ▼                                     ▼
SUCCESS (v1→v2)                      CONFLICT! (expected v1, found v2)
                                          │
                                          ▼
                                     Agent B resolves with MANUAL_MERGE
                                          │
                                          ▼
                                     SUCCESS (v2→v3) - both changes preserved!

Conflict Resolution

When a conflict is detected, you get 4 options:

# Use your version (discard theirs)
niwa resolve h2_3 ACCEPT_YOURS --agent me

# Use their version (discard yours)
niwa resolve h2_3 ACCEPT_THEIRS --agent me

# Manual merge (recommended!)
niwa resolve h2_3 MANUAL_MERGE "merged content" --agent me
niwa resolve h2_3 MANUAL_MERGE --file merged.md --agent me

Complex Content

For content with quotes, newlines, or special characters, use --file or --stdin:

# Edit from file
niwa edit h2_3 --file content.txt --agent me

# Edit from stdin
cat content.md | niwa edit h2_3 --stdin --agent me

# Resolve from file
niwa resolve h2_3 MANUAL_MERGE --file merged.md --agent me

Sub-Agents / Fresh Context

If you're a sub-agent or have fresh context, run these first:

# Get a unique agent name
niwa whoami

# Check your state
niwa status --agent <your_name>

# Check for pending conflicts
niwa conflicts --agent <your_name>

Key rules:

• Use a unique agent name (check with agents command)
• Use the same name for all your reads/edits
• Conflicts are stored and survive context switches

Node IDs

Nodes follow the pattern h{level}_{index}:

[root] v1 "Document"
  [h1_0] v1 "Main Title"
    [h2_1] v3 "Section 1" (by agent_A)     ← node_id is "h2_1"
    [h2_2] v1 "Section 2"
      [h3_3] v2 "Subsection" (by agent_B)  ← node_id is "h3_3"

Use tree to find node IDs, or search to find by content.

Examples

Basic editing workflow

niwa init
niwa load README.md
niwa tree
niwa read h2_1 --agent editor
niwa edit h2_1 "# Updated Heading\n\nNew content here." \
    --agent editor --summary "Rewrote introduction"
niwa export > README_updated.md

Handling a conflict

# You tried to edit but got a conflict
niwa edit h2_3 "My changes" --agent me
# Output: CONFLICT DETECTED!

# View the diff, then resolve
niwa resolve h2_3 MANUAL_MERGE \
    "Combined content from both versions" --agent me

Finding and editing content

# Don't know the node ID? Search for it
niwa search "authentication"
# Output: [h2_5] "Authentication" - 3 matches

# Read and edit
niwa read h2_5 --agent me
niwa edit h2_5 --file auth_rewrite.md --agent me

Rollback a bad edit

# Check history
niwa history h2_3
# Output: v3, v2, v1...

# Rollback to v2
niwa rollback h2_3 2 --agent me

Claude Code Integration

Niwa integrates with Claude Code via hooks for automatic context awareness:

# Set up Claude Code hooks
niwa setup claude

# Remove hooks later
niwa setup claude --remove

What the hooks do

Hook	Trigger	Action
SessionStart	Session begins	Injects full Niwa usage guide + current status
PreCompact	Before /compact	Preserves Niwa context so Claude remembers after compaction
PreToolUse	Before Write/Edit	Warns if there are unresolved conflicts
PostToolUse	After Write/Edit	Hints to sync markdown changes to database
Stop	Session ending	Reminds about unresolved conflicts

The SessionStart and PreCompact hooks ensure Claude always knows how to use Niwa, even after context compaction.

How it works

1. Run setup claude in your project directory
2. Creates .claude/settings.json with hook configuration
3. Hooks automatically call niwa hook --hook-event <event>
4. Claude receives context about your markdown database state

Example workflow with Claude Code

# 1. Initialize and set up hooks
niwa init
niwa load design_doc.md
niwa setup claude

# 2. Start Claude Code session
# Claude will see: "[Niwa Status] Database active with 15 nodes..."

# 3. When Claude edits markdown files
# Claude will see: "[Niwa] Markdown file modified... Consider syncing"

# 4. If conflicts exist when stopping
# Claude will see: "[Niwa Reminder] There are 2 unresolved conflict(s)..."

Architecture

┌─────────────────────────────────────────────────────────────┐
│                        Niwa CLI                             │
├─────────────────────────────────────────────────────────────┤
│  Markdown Parser (markdown-it-py)                           │
│  ├── GFM-like preset (tables, strikethrough, autolinks)     │
│  ├── front_matter_plugin (YAML/TOML)                        │
│  ├── footnote_plugin                                        │
│  ├── deflist_plugin                                         │
│  └── tasklists_plugin                                       │
├─────────────────────────────────────────────────────────────┤
│  Document Model                                             │
│  ├── Nodes (headings → content hierarchy)                   │
│  ├── Versions (each edit increments version)                │
│  ├── Agents (track who read/edited what)                    │
│  └── Conflicts (stored for resolution)                      │
├─────────────────────────────────────────────────────────────┤
│  Storage (LMDB)                                             │
│  ├── nodes_db      - Document nodes                         │
│  ├── pending_db    - Pending reads by agent                 │
│  └── conflicts_db  - Unresolved conflicts                   │
├─────────────────────────────────────────────────────────────┤
│  Claude Code Hooks                                          │
│  └── SessionStart, PreCompact, PreToolUse, PostToolUse, Stop│
└─────────────────────────────────────────────────────────────┘

Key design decisions:

• AST-based parsing: Uses markdown-it-py token line mapping to extract original source text, preserving exact formatting
• LMDB storage: Memory-mapped for fast concurrent access, ACID transactions
• Version vectors: Each node tracks version independently for fine-grained conflict detection
• Agent isolation: Each agent's pending reads/conflicts are isolated

Testing

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=niwa --cov-report=html

84 tests covering:

• Basic parsing (headings, nesting, hierarchy)
• Code blocks (fenced, indented, with # inside)
• GFM features (tables, task lists, strikethrough)
• Frontmatter (YAML, TOML)
• Edge cases (BOM, line endings, unicode, emoji)
• Adversarial inputs (unclosed fences, HTML comments)
• Round-trip export

See Also

jari(砂利) — Task/issue tracker for LLM agent workflows. Jari uses the same LMDB + conflict detection patterns as Niwa, and supports linking todos directly to Niwa node IDs. Use them together: Niwa for collaborative document editing, Jari for tracking tasks and issues.

pipx install jari

Name

Niwa (庭) means "garden" in Japanese. Like a zen garden where gravel (砂利, jari) is raked into patterns, Niwa helps organize your plans and specs with deliberate structure. Multiple agents can collaboratively tend the same garden without disturbing each other's patterns.

License

MIT