lattice-commit 0.1.0

Incremental commit for LLM code repair — breaks the multi-file wall

lattice-commit

Incremental commit for LLM code repair -- breaks the multi-file wall.

LLM code repair fails when bugs span multiple files. The standard edit-test-revert loop throws away successful fixes to file A when file B still has bugs. This 30-line mechanism fixes it: commit on test improvement, rollback to last committed state (not the original).

• 0% -> 100% solve rate at 2+ files (1,200+ trials)
• Works with any LLM and any test suite
• No dependencies beyond Python stdlib

Install

pip install lattice-commit

Python API

from lattice_commit import incremental_repair

result = incremental_repair(
    workspace="./my_project",
    test_cmd="pytest -q",
    llm_fix=my_fix_function,  # (filepath, content, error) -> fixed_content
    max_cycles=30,
)

print(f"Solved: {result.solved}, Cycles: {result.cycles}, Commits: {result.commits}")

The llm_fix function takes a filepath, its content, and the test error output, and returns the fixed file content (or None to skip).

CLI (requires Ollama)

lattice-commit --workspace ./my_project --test-cmd "pytest -q"
lattice-commit --workspace ./my_project --majority-vote 3  # noise-resistant mode

The CLI uses a local Ollama model (default: qwen2.5-coder:7b) as the LLM fixer.

How it works

for cycle in range(max_cycles):
    error = run_tests(workspace)
    if no error: return SUCCESS
    file, content = llm.propose_fix(error)
    write(file, content)
    if count_passing() > best_passing:
        best_passing = current
        commit()           # save progress
    else:
        rollback()         # to last commit, not original

That's it. The mechanism preserves partial progress across files. Fix file A, tests improve, commit. Now work on file B from a state where A is already fixed.

Key findings

• The wall is at 2 files, not any number of bugs. 4 bugs in 1 file: 100% bare. 2 bugs in 2 files: 0% bare.
• Model-invariant: tested on llama3.1:8b, qwen2.5-coder:7b, qwen2.5-coder:14b
• Temperature, prompt decomposition, and commit history context have no effect
• Test flakiness breaks it; majority-vote (3x test, median) restores 88%
• A 2,100-line substrate system performs identically to this 30-line version

License

MIT