开源AI工作流

Install dependencies and sync environment

uv sync

Install Playwright browsers (Required for high-fidelity documentation fetching)

uv run playwright install chromium ```

Review a full PR URL (requires gh cli)

compounding review https://github.com/user/repo/pull/123

Installation

Via Pip (For standard Python setups)

If you prefer using pip, you can install our lightweight wrapper which will bootstrap the uv environment under the hood:

pip install dspyce-install
dspyce-install

For Contributors (Source Installation)

```bash

Vector Database Setup (Qdrant)

This project uses Qdrant for semantic search. A Docker Compose configuration is provided.

1. Start Qdrant:

   docker compose up -d qdrant
   

This will start Qdrant on localhost:6333.

2. Configure Embeddings: Ensure your .env has the correct EMBEDDING_BASE_URL if you are using a non-standard provider (like OpenRouter). See .env.example.

Note: If Qdrant is not running, the system will automatically fall back to keyword-based search using local JSON files.

Example run

Quick example using the temporary runner uvx to see the generate-agent help:

./uvx generate-agent -h

Expected excerpt:

Usage: compounding generate-agent [OPTIONS] DESCRIPTION

Generate a new Review Agent from a natural language description.

Options:
  --dry-run  -n   Show what would be created without writing files
  --help     -h   Show this message and exit.

You can run other commands similarly, for example:

./uvx -h
./uvx review --project
./uvx work p1 --env-file test.env

Usage

[!TIP] If you have installed the tool via uv tool install --from ., you can use the compounding command directly. Otherwise, use uv run python cli.py.

Configure environment

cp .env.example .env

Global Options

The tool supports several global options that can be used with any command:

• -e, --env-file PATH: Explicitly specify a .env file to load.
• -h, --help: Show help for any command.

Configuration Priority

The tool loads configuration from multiple sources in the following priority order:

1. Explicit Flag: --env-file / -e on the command line.
2. Environment Variable: COMPOUNDING_ENV pointing to a .env file path.
3. Local Override: .env in the current working directory (CWD).
4. Tool-Specific Global: ~/.config/compounding/.env.
5. System Fallback: ~/.env in the user's home directory.

[!TIP] This priority allows you to have a global ~/.config/compounding/.env with your API keys while using local .env files for project-specific model selections.

Configuration

Edit .env to configure your LLM provider:

```bash

Context Limits (Optional)

CONTEXT_WINDOW_LIMIT=128000 CONTEXT_OUTPUT_RESERVE=4096 DOCS_MAX_TOKENS=32768 # Limit for documentation fetching (default: 32k) ```

Multi-Source Configuration

As a tool meant to be used across multiple repositories, configuration can be managed at different levels:

• Global: Store your API keys in ~/.config/compounding/.env so they are available everywhere.
• Local: Add a .env in your project root to override models or settings for that specific project.
• On-the-fly: Use --env-file path/to/.env to quickly switch between different environments (e.g., testing vs. production models).

The tool will warn you if multiple conflicting configuration files are detected.

1. Interface Layer (CLI)

• Entry Point: cli.py uses Typer to provide a robust command-line interface.
• Commands: Maps user intents (e.g., review, work) to specific workflows.

MCP Server Integration (Claude Desktop)

compounding can be attached to any Model Context Protocol (MCP) client, such as Claude Desktop, to expose its capabilities as native tools to the assistant.

To configure Claude Desktop to use the DSPy Compounding Engineering FastMCP server:

1. Open your Claude Desktop configuration file (typically at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS).
2. Add the following to the mcpServers section:

{
  "mcpServers": {
    "dspy-compounding": {
      "command": "uvx",
      "args": ["compounding-mcp"]
    }
  }
}

1. Restart Claude Desktop. The agent will now have access to compounding_review, compounding_plan, compounding_work, compounding_triage, and compounding_sync directly from your chat!

2. Orchestration Layer (Workflows)

• Logic: Python scripts in workflows/ orchestrate complex multi-step processes.
• Responsibility: Manages state, handles user interaction, and coordinates agents.
• Key Workflows:
• Unified Work: Combines planning and execution using ReAct loops.
• Review Pipeline: Parallelizes multiple specialized review agents.
• Triage System: Manages findings and prioritizes work.

Comparison with Original Plugin