CLAUDE.local.md

Test Generation and Execution

• It is IMPORTANT to execute your generated tests in both languages to verify correctness. DO NOT give your response until you have executed the tests in both languages and captured the output.
• When generating tests, YOU MUST only create standalone files in <target_fragment_path> as provided in the initial prompt. DO NOT create global tests in other modules/packages/directories.
• For Rust tests, make sure you execute tests using cargo test.
• For Python tests, make sure you execute tests using python -m unittest or pytest.
• For Java tests, make sure you execute tests using mvn clean test -Drat.skip.
• For JavaScript tests, make sure you execute tests using node.
• For Go tests, make sure you execute tests using go.
• For C tests, make sure you compile and run tests using gcc.
• To fix the <target_fragment_method>, YOU MUST make edits in-place in the file given in <target_fragment_path>.

MCP Servers

• Use the available function get_directory_tree(path: str, print_dirs_only: bool) -> str from DirectoryTreeExplorer server to get more context about files and directory structure. This will help you understand where files are located and where to store your generated test files.
• When calling the get_directory_tree(path: str, print_dirs_only: bool) -> str function from DirectoryTreeExplorer, make sure first you call it with print_dirs_only=True as sometimes high-level directories (e.g., directories higher in the path) can produce very large tree structures. You can call low-level directories (e.g., directories lower in th path) with print_dirs_only=False to see their files as their tree structure is relatively shorter.

TODO list for Test Generation and Repair Agent

• You must always follow the following TODO when running the test generation and repair agent:
  • Read the source file to understand context. The path is given in the initial prompt.
  • Read the target file to understand context. The path is given in the initial prompt.
  • Call the get_directory_tree(path: str, print_dirs_only: bool) -> str with a proper path to understand the directory structure and available files in both source and target projects.
  • After understanding the project structure, generate test files in both source and target projects and determine where to store them.
  • Execute tests in both source and target projects and verify results. Your tests should exercise the same behavior. Your goal is to expose bugs in target fragment, if any.
  • If tests in both source and target projects are passing, provide your final response in expected <final_response_format> with "is_equivalent" key set to "yes" as given in initial prompt and then terminate as soon as possible.
  • If tests in target project are failing, set "is_equivalent" key to "no" in <final_response_format> and make necessary fixes to target fragment if needed. Execute target tests again to verify your patch is working. Keep going until you fix the target fragment.
  • Provide your response in expected <final_response_format> given in initial prompt.
  • Terminate as soon as possible.

TODO list for Verdict Agent

• You must always follow the following TODO when running the verdict agent:
  • Consider all semantic analysis results.
  • Consider test generation and repair analysis results (give more attention to these).
  • To the best of your knowledge, determine the functional equivalency between source and target fragments.
  • Provide your response in expected <final_response_format> given in initial prompt.
  • Terminate as soon as possible.