Skip to content

Latest commit

 

History

History
111 lines (88 loc) · 4.64 KB

PKG.md

File metadata and controls

111 lines (88 loc) · 4.64 KB

Package Dependencies

Third-party packages used across the Echoes of Dustwood project, organized by purpose.

Python Dependencies

Core Infrastructure

Package Purpose Used In
pydantic Data validation & parsing All framework packages, shared
python-dotenv Environment variable management All packages
httpx Async HTTP client shared (LLM observability)

AI Frameworks

Package Purpose Used In
pydantic-ai Pydantic AI agent framework pydantic package
agno Agno agent framework agno package
strands-agents Strands Agents framework strands package
google-adk Google AI Development Kit with MCP extensions adk package

LLM Provider SDKs

Package Purpose Used In
anthropic Anthropic Claude API agno (framework support)
openai OpenAI API agno (framework support)
google-genai Google Generative AI API agno (framework support)
ollama Local Ollama model integration agno (framework support)
litellm LLM abstraction layer (unified API) strands (model routing)

Model Context Protocol (MCP)

Package Purpose Used In
mcp MCP SDK for structured tool communication agno, strands

Visualization & Analysis

Package Purpose Used In
matplotlib Plotting and data visualization charts (benchmark visualizations)
numpy Numerical computing charts (data processing)

Internal

Package Purpose Used In
vibepascal-shared Guidance loader, LLM observability, MCP command policy All framework packages

Go Dependencies

Direct Dependencies

Package Purpose
github.com/modelcontextprotocol/go-sdk MCP server implementation (structured tool definitions, JSON-RPC transport)
golang.org/x/term Terminal I/O control (raw mode for game input)
gopkg.in/ini.v1 INI file parsing (world.ini configuration)

Indirect Dependencies

Package Purpose
github.com/google/jsonschema-go JSON schema validation (MCP tool definitions)
github.com/segmentio/asm Assembly utilities (encoding optimization)
github.com/segmentio/encoding Efficient JSON/binary encoding
github.com/stretchr/testify Testing assertions and utilities
github.com/yosida95/uritemplate/v3 URI template processing
go.etcd.io/bbolt Embedded key-value store (if used by MCP SDK)
golang.org/x/oauth2 OAuth2 support (potential provider integration)
golang.org/x/sys System-level calls (signal handling, platform-specific I/O)

Dependency Rationale

Why These LLM Providers?

Multiple provider SDKs (Anthropic, OpenAI, Google, Ollama) are included to support the framework comparison. Each framework can be configured to use different providers via environment variables.

Why Both Pydantic AI and Alternative Frameworks?

The project compares 4 different AI agent frameworks:

  1. Pydantic AI — Modern, Python-native, Pydantic integration
  2. Agno — Feature-rich with built-in multi-provider support
  3. Strands — LiteLLM-based for flexible model routing
  4. Google ADK — Google-native with integrated MCP tooling

Why MCP?

MCP (Model Context Protocol) allows frameworks to call game commands via structured tool definitions rather than text parsing, enabling:

  • Reliable command execution with JSON responses
  • Consistent agent behavior across frameworks
  • Extensibility for new game commands

Why Go for the Game Server?

The Go implementation provides:

  • Better concurrency for MCP (HTTP + stdio servers)
  • Stronger type safety than Pascal
  • Cross-platform binary distribution
  • Reliable signal handling and graceful shutdown

Version Strategy

  • Framework packages: Allow minor/patch updates (>= constraints)
  • Pydantic AI: Pinned to 2.0.0b3 (early access, API stability required)
  • Google ADK: Pinned to 2.0.0 (breaking changes expected in future)
  • Python: Requires >=3.12 across all packages
  • Go: Go 1.24.0

Related Documentation

  • Overview Index: README.md
  • Framework Packages Setup: packages/README.md — How isolated framework virtual environments are structured.
  • Client Implementation Details: packages/IMPL.md — Comparative overview of each client's dependencies and logic.