ghostrun 0.1.5

CI-native LLM evals for real applications.

ghostrun turns real LLM app behavior into deterministic pytest evals: record API calls once, replay them in CI, and catch semantic regressions before they ship.

Is this for you? · Install · Quickstart · Documentation · Roadmap

[!TIP] LLM evals often live outside the product: separate datasets, dashboards, YAML files, and one-off scripts. ghostrun keeps evals inside normal pytest, runs them against the real code path your users hit, records the live LLM call once, and replays it deterministically in CI.

The problem

reply = generate_reply("Where is my refund?")
assert reply == "I'm sorry for the delay..."   # fails tomorrow: LLM never says the same thing twice

LLM apps need evals, but live model calls make test suites painful: they cost money, need API keys in CI, run slowly, and fail for reasons unrelated to your code. Generic eval platforms are powerful, but they often test a prompt, dataset, or trace outside your application. ghostrun focuses on app-native LLM regression testing: the exact Python code path your users exercise.

How ghostrun fixes it

1. Deterministic replay — the first run records real LLM HTTP calls to a local .ghostrun_cache/; every run after replays them instantly from disk. Zero API cost, zero latency, zero flakiness, no key needed in CI.
2. Semantic assertions — assert on meaning, not exact text:

   ghostrun.expect(reply).contains_intent("apology")
   ghostrun.expect(reply).tone_is("empathetic")
   

   Graded by a local Ollama model by default — your prompts and data never leave your machine. That grading verdict gets cached too, so it's also free and deterministic after the first run.

No cloud dashboard. No custom CLI to learn. No dataset to author by hand. Just pytest.

Search terms this project is built for: LLM evals in CI, LLM regression testing, pytest LLM evals, deterministic LLM tests, prompt regression testing, semantic assertions for LLM apps, and testing OpenAI or Anthropic applications with pytest.

Install

pip install ghostrun

For the default (local, free, private) judge, install Ollama and pull a small model:

ollama pull llama3.2:3b

If anything doesn't work, ghostrun doctor diagnoses the setup — see Configuration.

Fastest start: ghostrun init scaffolds a working first test against whatever LLM SDK it finds in your project (OpenAI, Anthropic, or a generic HTTP fallback) plus a .ghostrun.yaml — no config to author by hand:

ghostrun init
pytest test_ghostrun_example.py

Quickstart

# test_customer_support.py
import ghostrun
from my_app import generate_reply

@ghostrun.record(model="gpt-4o-mini")
def test_reply_generation():
    reply = generate_reply("Where is my refund?")

    ghostrun.expect(reply).contains_intent("apology")
    ghostrun.expect(reply).contains_intent("refund policy")
    ghostrun.expect(reply).does_not_contain_intent("arguing")
    ghostrun.expect(reply).tone_is("empathetic")

$ pytest test_customer_support.py
================================ test session starts ================================
collected 1 item

test_customer_support.py .                                                     [100%]
================================ 1 passed in 0.04s =================================

The 0.04s is the whole point — after the first record, calls replay from disk.

Note on the API: the assertion entry point is ghostrun.expect(...), not ghostrun.assert(...) — assert is a reserved Python keyword and cannot be a function name.

Record/replay alone needs no Ollama at all — the judge is only touched when you call a judge-backed assertion (contains_intent, tone_is, matches). Deterministic assertions (contains, is_valid_json) and tool-call assertions never invoke it.

Is this for you?

Use ghostrun if you're writing pytest tests around code that calls an LLM (directly or via the OpenAI/Anthropic SDKs) and want that suite to run offline, free, and deterministically after the first recording.

Skip it if you need a hosted dashboard/observability platform for production traffic (see Langfuse/LangSmith/Braintrust instead), you're building a red-team/adversarial test suite (see Giskard), or you want 50+ pre-built judge metrics out of the box today (see DeepEval — more mature, more metrics, but doesn't intercept your app's own HTTP calls the way ghostrun does). See doc/comparison.md for the full, researched breakdown of where ghostrun is ahead and where it's duplicating existing work.

Documentation

Start here, in order:

A hosted, searchable version of this documentation is planned at ghostrun.parthmax.in (config in mkdocs.yml, builds via .github/workflows/docs.yml).

Roadmap

• Deterministic HTTP record/replay (@ghostrun.record, .ghostrun_cache/)
• Semantic assertions (contains_intent, tone_is, matches) via local Ollama or an offline echo stub
• Judge-verdict caching, including majority-vote grading (judge.votes) with a benchmarked reliability tradeoff
• Tool/function-call assertions (expect_tool_calls)
• Prompt regression tracking — ghostrun diff, PR-comment and JUnit CI output
• ghostrun init / ghostrun doctor — scaffolding and setup diagnostics in one command
• Secret redaction so the cache is safe to commit
• 16-provider HTTP coverage (OpenAI, Anthropic, Gemini, Bedrock, and more)
• Published to PyPI with trusted-publishing (OIDC) releases
• Hosted, searchable documentation site (mkdocs.yml ready; not yet deployed)
• Broader provider/framework integration guides

Contributing

See CONTRIBUTING.md — setup, test requirements, and where things live in the codebase.

Development

pip install -e ".[dev]"
pytest            # runs fully offline using the echo judge

License

MIT