Metadata-Version: 2.4
Name: claude-sage
Version: 4.2.2
Summary: Semantic checkpointing for Claude Code
License-Expression: MIT
Keywords: claude,claude-code,checkpointing,memory,mcp,ai
Classifier: Development Status :: 5 - Production/Stable
Classifier: Programming Language :: Python :: 3
Requires-Python: >=3.12
Description-Content-Type: text/markdown
Requires-Dist: anthropic>=0.40.0
Requires-Dist: click>=8.0.0
Requires-Dist: rich>=13.0.0
Requires-Dist: pyyaml>=6.0.0
Requires-Dist: prompt-toolkit>=3.0.0
Requires-Dist: numpy>=1.24.0
Requires-Dist: sentence-transformers>=2.2.0
Requires-Dist: jinja2>=3.0.0
Provides-Extra: dev
Requires-Dist: pytest>=7.0.0; extra == "dev"
Requires-Dist: pytest-asyncio>=0.21.0; extra == "dev"
Requires-Dist: black>=23.0.0; extra == "dev"
Requires-Dist: ruff>=0.1.0; extra == "dev"
Provides-Extra: mcp
Requires-Dist: mcp>=1.2.0; extra == "mcp"
Requires-Dist: fastmcp>=2.0; extra == "mcp"
Provides-Extra: code
Requires-Dist: lancedb>=0.4.0; extra == "code"
Requires-Dist: tree-sitter-languages<1.10,>=1.9.0; extra == "code"
Requires-Dist: tree-sitter<0.22,>=0.21.0; extra == "code"

# Sage

**Memory for Claude Code.** Research → checkpoint → compaction → auto-restore.

```
┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│  Research   │───▶│ Checkpoint  │───▶│  Compaction │
│  with Claude│    │  (auto)     │    │  happens    │
└─────────────┘    └─────────────┘    └──────┬──────┘
                                             │
┌─────────────┐    ┌─────────────┐           │
│  Continue   │◀───│ Auto-inject │◀──────────┘
│  seamlessly │    │  context    │
└─────────────┘    └─────────────┘
```

**v4.0 — Invisible Context Hydration**: System folders, failure memory, MCP resources, knowledge linking.

## Quick Start

### Option A: Claude Code Plugin (Recommended)

```bash
# 1. Add the marketplace (one-time)
/plugin marketplace add b17z/sage

# 2. Install the plugin
/plugin install sage@sage
```

Or run `/plugin` and use the interactive UI to browse and install.

### Option B: Manual Install

```bash
# 1. Install
pip install claude-sage[mcp]

# 2. Setup (adds MCP server + installs methodology skills)
sage mcp install
sage skills install

# 3. Use Claude - Sage handles the rest
claude
```

That's it. Claude now has memory across sessions.

## How It Works

**The problem:** You're 2 hours into research. Context fills up, auto-compacts, nuanced findings gone. Tomorrow you start from scratch.

**The solution:** Sage checkpoints at meaningful moments—not when tokens run out, but when something worth remembering happens:

| Trigger | Example |
|---------|---------|
| Synthesis | "Therefore, the answer is..." |
| Branch point | "We could either X or Y..." |
| Constraint | "This won't work because..." |
| Topic shift | Conversation changes direction |
| Manual | You say "checkpoint this" |

Each checkpoint captures your **thesis**, **confidence**, **open questions**, **sources**, and **tensions** (where experts disagree).

## What Gets Saved

```markdown
# Where do stablecoins win vs traditional rails?

## Thesis (75% confidence)
Integrate, don't replace. Stablecoins win middle-mile,
not POS checkout.

## Open Questions
- Timeline for Stripe's full stack?

## Tensions
- sheel_mohnot vs sam_broner: merchant profitability — unresolved
```

Checkpoints are Markdown files (Obsidian-compatible) in `~/.sage/checkpoints/` or `.sage/checkpoints/` (project-local).

## The Three Layers

```
┌────────────────────────────────────────────────┐
│  Skills (methodology)                          │
│  sage-memory, sage-research, sage-session      │
│  Load on-demand when context matches           │
├────────────────────────────────────────────────┤
│  MCP Server (tools + resources)                │
│  sage_save_checkpoint, sage_recall_knowledge   │
│  @sage://system/objective.md (v4.0)            │
├────────────────────────────────────────────────┤
│  Storage                                       │
│  ~/.sage/checkpoints/, ~/.sage/knowledge/      │
│  .sage/system/, .sage/failures/ (v4.0)         │
└────────────────────────────────────────────────┘
```

- **Skills** teach Claude *when* and *how* to checkpoint
- **MCP** gives Claude the *tools* to save/load, *resources* for direct access
- **Storage** persists everything as readable Markdown

## CLI Basics

```bash
sage checkpoint list          # See your checkpoints
sage checkpoint show <id>     # View one
sage knowledge list           # See stored knowledge
sage knowledge match "query"  # Test what would recall
sage skills list              # Check installed skills
sage watcher start            # Auto-detect compaction

# Configuration
sage config list              # View current settings
sage config set checkpoint_max_age_days 30  # Customize storage
sage config set failure_memory_enabled true # Enable failure memory (v4.0)
```

## Visual Interface

```bash
sage ui              # Local web UI at localhost:5555
sage ui --api-only   # REST API for custom frontends
```

Or use any of these:
- **Obsidian** — Open `~/.sage/` as vault (it's just Markdown)
- **Custom** — Build on the REST API

See [docs/ui.md](docs/ui.md) for details.

## Learn More

- **[Features](docs/FEATURES.md)** — Complete feature reference
- **[What's New in v4.0](docs/features/README.md#invisible-context-hydration-v40)** — System folder, failures, resources
- **[Architecture](docs/ARCHITECTURE.md)** — System design
- **[Security](docs/security.md)** — Security measures and threat model
- **[Skills](docs/skills.md)** — How methodology skills work
- **[Continuity](docs/continuity.md)** — Session persistence deep-dive
- **[Maintenance](docs/maintenance.md)** — Storage maintenance and caching
- **[UI Options](docs/ui.md)** — Web UI, API, Obsidian, CoWork plugin

## Known Issues

### Plugin MCP Tool Naming

When installed as a Claude Code plugin, MCP tools follow the format `mcp__plugin_<plugin>_<server>__<tool>`. Since both the plugin and MCP server are named "sage", tools appear as `mcp__plugin_sage_sage__save_checkpoint` rather than the expected `mcp__plugin_sage__save_checkpoint`.

This is [documented Claude Code behavior](https://github.com/anthropics/claude-code/blob/main/docs/plugins.md) — the official MCP integration docs show the same pattern with `asana_asana` as an example. No official plugins currently use MCP servers, so there's no precedent for cleaner naming.

**Workaround:** The plugin works correctly despite the redundant naming. If this bothers you, install via pip instead (`pip install claude-sage[mcp]`).

## Requirements

- Python 3.12+
- [Claude Code](https://claude.ai/code) CLI

## Development

```bash
pip install -e ".[dev,mcp]"
pytest tests/ -v  # 1624 tests
```

## Acknowledgments

Output formatting inspired by [TOON](https://toon-format.org) — a token-efficient notation format for LLMs by [@mixeden](https://github.com/mixeden).

## License

MIT
