Add Claude Code LM integration for DSPy

[richardwhiteii]

Summary

• Adds ClaudeCodeLM client for direct LLM access via Claude Code CLI
• Provides OpenAI-compatible response format for seamless DSPy integration
• Includes comprehensive example with multiple usage patterns

Features

• Direct LLM inference through Claude CLI subprocess calls
• Support for different Claude models (sonnet, opus, etc.)
• Multi-turn agentic conversations via --max-turns parameter
• JSON output parsing with robust error handling
• No API keys required - uses Claude Code CLI authentication

Usage

import dspy
from dspy.clients.claude_code_lm import create_claude_code_lm

# Create Claude Code LM instance
lm = create_claude_code_lm(claude_model="sonnet", max_turns=3)

# Use with DSPy
dspy.configure(lm=lm)
predictor = dspy.Predict("question -> answer")
result = predictor(question="What is machine learning?")

Prerequisites

• Claude Code CLI: npm install -g @anthropic-ai/claude-code
• Authentication: claude auth login

Test plan

• Verify Claude Code CLI integration works
• Test OpenAI-compatible response format
• Validate multi-turn conversations
• Check error handling for missing CLI
• Run comprehensive examples

Author: Richard White

[beapirate]

(not a dspy mainainer) Integrations against provider specific client APIs were removed in the 2.6 release in favor of litellm.

https://github.com/stanfordnlp/dspy/blob/main/docs/docs/roadmap.md#on-lms
0f4258f

[richardwhiteii]

This is an alternate way at reusing access to a model in this case Claude without going through the api. This works for someone prototyping on their local development environment before using something like litellm against their API token count.

[okhat]

Thank you @richardwhiteii , @beapirate ! @richardwhiteii Maybe this could be a separate helper package?