parseltongue-claude: Instructions file for Claude Code

Parseltongue Claude - Project Instructions

MCP server wrapping Parseltongue's symbolic verification engine for Claude Code integration.

Architecture

This project inverts the typical Parseltongue flow:

• Original: Parseltongue calls LLM API to generate DSL
• This: Claude Code generates DSL, calls Parseltongue to verify

The MCP server exposes Parseltongue's core engine as tools. Sessions are stored in memory.

Key Files

src/parseltongue_claude/
├── __init__.py     # Package version
└── server.py       # MCP server with all tools

Development

# Install
uv pip install -e ".[dev]"

# Test the server manually
python -m parseltongue_claude.server

# Run tests
pytest

Adding New Tools

1. Add tool definition to list_tools() in server.py
2. Add case to _handle_tool() match statement
3. Implement handler function _your_tool_name()

Parseltongue DSL

When generating DSL for verification:

1. Facts must have :evidence blocks with verbatim :quote from registered documents
2. Derives use :wff (well-formed formula) and :using to cite dependencies
3. Diffs cross-check values with :expect formulas

If a quote doesn't match the source document exactly, Parseltongue flags it as unverified (hallucination).

Integration Testing

Test with the actual parseltongue library:

from parseltongue.core import System, load_source

system = System(overridable=True)
system.register_document("test", "The revenue was $42 million.")

dsl = '''
fact revenue
  :value 42000000
  :evidence
    :doc "test"
    :quote "The revenue was $42 million."
'''
load_source(system, dsl)
print(list(system.facts()))