Litmus: a 450-line stdlib-only eval harness to catch LLM prompt regressions in CI
A developer grew frustrated with LLM app regressions caused by small prompt edits and built Litmus, a single Python file with no dependencies. It runs an app against a JSONL dataset of inputs and expected content, reporting pass rates and latency percentiles. A compare command flags cases that previously passed but now fail and exits non-zero so GitHub Actions can block a merge. The tool works with any app via stdin piping or any OpenAI-compatible endpoint, including local Ollama models. It is sold as a one-time $12 purchase that includes a field guide and CI workflow.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.



Discussion (0)
Log in to join the discussion and vote.
Log in