Agentic Context Engineering

A simplified implementation of Agentic Context Engineering (ACE) for Claude Code that automatically learns and accumulates key points from reasoning trajectories.

Features

• Automatic Key Point Extraction: Learns from reasoning trajectories and extracts valuable insights
• Score-Based Filtering: Evaluates key points across trajectories and removes unhelpful ones
• Context Injection: Automatically injects accumulated knowledge at the start of new sessions
• Multiple Triggers: Works on session end, manual clear (/clear), and context compaction

Installation

Prerequisites

• Python 3.8+
• Claude Code
• anthropic Python SDK
• Node.js and npm

Setup

1. Clone and install:

git clone https://github.com/bluenoah1991/agentic_context_engineering.git
cd agentic_context_engineering
npm install

2. Install required Python package:

pip3 install anthropic

3. Set environment variables for the LLM API:

Environment Variable	Description	Required
AGENTIC_CONTEXT_MODEL	Model name for key point extraction (fallback: ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, claude-sonnet-4-5-20250929)	Optional
AGENTIC_CONTEXT_API_KEY	API key (fallback: ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY)	Optional
AGENTIC_CONTEXT_BASE_URL	API base URL (fallback: ANTHROPIC_BASE_URL)	Optional

4. Restart Claude Code - hooks will be active across all your projects

How It Works

Hooks

The system uses three types of hooks:

1. UserPromptSubmit: Injects accumulated key points at the start of each new session
2. SessionEnd: Extracts key points when a session ends
3. PreCompact: Extracts key points before context compaction

Key Point Lifecycle

1. Extraction: At the end of each session, the system analyzes the reasoning trajectories and extracts new key points
2. Evaluation: Existing key points are evaluated based on the reasoning trajectories and rated as helpful/harmful/neutral
3. Scoring:
   • Helpful: +1 point
   • Harmful: -3 points
   • Neutral: -1 point
4. Pruning: Key points with score ≤ -5 are automatically removed
5. Injection: Surviving key points are injected into new sessions

Configuration

Diagnostic Mode

To enable detailed logging of LLM interactions:

touch .claude/diagnostic_mode

Diagnostic logs will be saved to .claude/diagnostic/ with timestamped filenames.

To disable:

rm .claude/diagnostic_mode

/exit Command Behavior

By default, the system does not update the playbook when using /exit. You can enable this behavior by setting playbook_update_on_exit to true in your ~/.claude/settings.json:

{
  "playbook_update_on_exit": true
}

/clear Command Behavior

By default, the system does not update the playbook when using /clear. You can enable this behavior by setting playbook_update_on_clear to true in your ~/.claude/settings.json:

{
  "playbook_update_on_clear": true
}

Customizing Prompts

Prompts are located in ~/.claude/prompts/:

• reflection.txt: Template for key point extraction from reasoning trajectories
• playbook.txt: Template for injecting key points into sessions

File Structure

.
├── install.js                 # Installation script
├── package.json               # npm package configuration
├── src/
│   ├── hooks/
│   │   ├── common.py           # Shared utilities
│   │   ├── session_end.py      # SessionEnd hook
│   │   ├── precompact.py       # PreCompact hook
│   │   └── user_prompt_inject.py  # UserPromptSubmit hook
│   ├── prompts/
│   │   ├── reflection.txt      # Key point extraction template
│   │   └── playbook.txt        # Injection template
│   └── settings.json           # Hook configuration template
└── README.md

License

MIT