origami

Origami

Context folding for Claude Code. Replaces lossy compaction with variable-resolution context: every conversation section keeps an always-visible self-compressed summary, with full detail stored on disk and expandable on demand.

How It Works

Normal conversation flow
         │
         ▼
   Auto-compact triggers
         │
         ▼
   PreCompact hook ──► Save transcript sections as folds
         │              Run Haiku librarian for relevance scoring
         │              Return fold-aware compact instructions
         ▼
   Claude Code compaction (produces fold-aware summary)
         │
         ▼
   SessionStart hook ─► Inject fold index + relevant unfolded sections
         │
         ▼
   Conversation resumes with folded context
         │
   Opus calls unfold_section MCP tool as needed

Core Concepts

• Fold — a logical section of conversation stored on disk with full detail
• Summary — self-compressed, always in context regardless of fold state
• Fold State — folded (summary only) or unfolded (full detail in context)
• Librarian — Haiku evaluates relevance and decides fold/unfold state each compaction cycle
• Unfold Tool — MCP tool Opus calls to expand any fold on demand

Installation

Prerequisites

• Node.js >= 18
• Python >= 3.9
• ANTHROPIC_API_KEY environment variable (for the Haiku librarian)

Setup

git clone https://github.com/dullfig/origami.git
cd origami
pip install -r requirements.txt

The MCP server is self-contained (no npm install needed — zero Node.js dependencies).

Then add the plugin to your Claude Code project (copy or symlink into your project, or configure via Claude Code settings).

Usage

The plugin works automatically during compaction. You can also use these commands:

Command	Description
/fold-status	Show all folds with their state and summaries
/unfold <id>	Manually expand a folded section
/fold-reset	Clear all fold state and start fresh

MCP Tools (called by the model)

Tool	Description
unfold_section(fold_id)	Expand a fold to full detail
fold_section(fold_id)	Collapse back to summary-only
list_folds()	Show fold index
write_summary(fold_id, summary)	Update a fold's self-compressed summary

Data Storage

Fold state is stored at .claude/context-folding/ in the project root:

.claude/context-folding/
├── state.json          # Fold index with summaries, scores, metadata
└── folds/
    ├── fold-001.md     # Full detail of section 1
    ├── fold-002.md     # Full detail of section 2
    └── ...

Token Budget

Aggressive folding by default. Research (Lost in the Middle, context length vs. performance) shows LLM performance degrades well before context is exhausted. The system targets:

• 20% of context window max for injected content (40k tokens for a 200k window)
• Max 3 sections unfolded simultaneously
• 0.7 relevance threshold to unfold (only highly relevant sections)

Summaries are always included regardless of budget. The model can still unfold on demand via the MCP tool when it needs specific detail.

Architecture

• Hooks (Python): PreCompact parses transcripts and saves folds; SessionStart injects folded context
• MCP Server (Node.js): Stdio server providing unfold/fold/list/write_summary tools
• Librarian (Python + Haiku API): Relevance scoring for intelligent fold/unfold decisions
• Skill: Teaches the model how to use the folding system effectively

License

MIT