Amplifier Skills Context Module

Warning

DEPRECATED - No Longer Maintained

This module is deprecated and no longer being developed. The amplifier-module-tool-skills works perfectly well standalone without needing context-level integration.

Use amplifier-module-tool-skills instead - it provides all the skills functionality you need with simpler setup and better agent experience.

Context manager with progressive skills loading and metadata injection.

What Are Skills?

Skills are folders of instructions, scripts, and resources that agents load dynamically to improve performance on specialized tasks (see Anthropic Skills).

This module brings Anthropic Skills support to Amplifier's context layer, enabling automatic discovery and metadata injection with progressive disclosure.

Note: This module is designed to work with amplifier-module-tool-skills. See amplifier-module-tool-skills/examples/ for complete working profiles.

Prerequisites

• Python 3.11+
• UV - Fast Python package manager

Installing UV

# macOS/Linux/WSL
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Purpose

Extends context management with automatic skills discovery and metadata injection. Provides agents with awareness of available skills without loading full content until needed (progressive disclosure).

Key benefit: Agents see what skills are available (~100 tokens) without loading full content (~1-5k tokens per skill) until explicitly needed.

Contract

Module Type: Context Mount Point: context Entry Point: amplifier_module_context_skills:mount

Behavior

• Discovers skills from .amplifier/skills/ directory on initialization
• Optionally injects skills metadata into system instruction (~100 tokens total)
• Tracks loaded skills to prevent redundant loading
• Wraps any base context manager (context-simple, context-persistent, etc.)
• Delegates all standard context operations to base

Configuration

session:
  context:
    module: context-skills
    config:
      base_context: context-simple
      skills_dirs:  # Configure skills directories here
        - ~/anthropic-skills  # Cloned from github.com/anthropics/skills
        - .amplifier/skills   # Project-specific skills
      auto_inject_metadata: true

Default: If not configured, uses .amplifier/skills

Capability sharing: Registers skills.registry and skills.directories capabilities for tool-skills to use (no duplicate configuration needed).

See complete example profile for working configuration.

How It Works

Session Start

1. Discovery: Scans skills_dir for */SKILL.md files
2. Parsing: Extracts YAML frontmatter (name, description, metadata)
3. Indexing: Builds skills registry
4. Injection (if enabled): Adds metadata to system instruction

Token cost: ~100 tokens for 3 skills metadata

During Execution

1. Agent sees: Skills list in system instruction
2. Agent calls: load_skill(skill_name="...") via tool
3. Context tracks: Marks skill as loaded
4. Prevents: Redundant loading of same skill

Progressive Disclosure

Startup: 100 tokens (metadata for all skills)
On-demand: +2000 tokens (one skill loaded)
As-needed: +0 tokens (references read directly)

Total: 2100 tokens vs 6000 tokens (60% savings)

Integration with tool-skills

Designed to work with amplifier-module-tool-skills via capability registry:

session:
  context:
    module: context-skills
    config:
      skills_dirs:  # Single configuration
        - ~/anthropic-skills
        - .amplifier/skills

tools:
  - module: tool-skills  # No config - reads from context capability

How it works: Context registers skills.registry and skills.directories capabilities. Tool reads from these capabilities to avoid duplicate discovery.

See complete example.

Complete workflow:

1. Context: Injects skills list into system instruction
2. Agent: Sees available skills automatically
3. Tool: Agent explicitly loads when needed
4. Context: Tracks loaded skills

Context Wrapper Pattern

Skills-context wraps any base context:

# Wraps context-simple
session:
  context: context-skills
  config:
    base_context: context-simple

# Wraps context-persistent
session:
  context: context-skills
  config:
    base_context: context-persistent

Delegation: All standard context methods delegate to base:

• add_message() → base.add_message()
• get_messages() → base.get_messages()
• should_compact() → base.should_compact()
• compact() → base.compact()
• clear() → base.clear() + clear loaded skills

Usage Examples

In Profile Configuration

---
profile:
  name: with-skills
  extends: dev

session:
  context: context-skills
  config:
    base_context: context-simple
    skills_dir: .amplifier/skills
    auto_inject_metadata: true

tools:
  - module: tool-filesystem
  - module: tool-bash
  - module: tool-skills
    config:
      skills_dir: .amplifier/skills
---

Skills Metadata Injection

When auto_inject_metadata: true:

System instruction includes:

## Available Skills

**amplifier-philosophy**: Design philosophy using Linux kernel metaphor...
**python-standards**: Python coding standards including type hints...
**module-development**: Guide for creating new modules...

To access full content, use the load_skill tool.

Token cost: ~100 tokens (vs 6000+ if all content loaded)

Agent Experience

Agent starts session:
- Sees skills list in system instruction (automatic)
- Knows what expertise is available
- Can reference by name: "following python-standards skill"

Agent needs details:
- Calls load_skill(skill_name="python-standards")
- Receives full content
- Context marks as loaded
- Second load returns "already loaded" (prevents waste)

API for Skills Awareness

Context provides skills-specific methods:

# Get list of available skills
skills: list[str] = context.get_available_skills()
# Returns: ["amplifier-philosophy", "python-standards", ...]

# Check if skill is loaded
loaded: bool = context.is_skill_loaded("python-standards")
# Returns: True if already loaded

# Mark skill as loaded (called by tool)
context.mark_skill_loaded("python-standards")

# Get formatted metadata
metadata: str = context.get_skills_metadata()
# Returns: Formatted string for system instruction

Skills Directory Locations

Standard locations (future multi-source support):

1. .amplifier/skills/          # Project skills (git-shared)
2. ~/.amplifier/skills/         # User skills (personal)
3. {app}/data/skills/          # Bundled skills (built-in)

Current MVP: Single directory from config

Testing

# Run unit tests
make test

# Run specific test
uv run pytest tests/test_context.py::test_skills_context_initialization -v

Development

# Install dependencies
make install

# Format and check code
make check

# Run all tests
make test

Dependencies

• amplifier-core - Core protocols and types
• pyyaml>=6.0 - YAML parsing

Contributing

See main Amplifier repository for contribution guidelines.

License

MIT