kensa 0.10.0

Pytest-native evals harness for agent regression tests

Kensa turns agent traces into pytest evals that run in CI.

---

Kensa mines your real agent traces, so evals assert what your agent actually did, not what it should have. Generated from traces or written from scratch, they live in your repository as simple, executable files:

import pytest

from kensa.pytest import judge, kensa_case


@pytest.mark.kensa(trials=3)
@pytest.mark.parametrize(
    "case",
    [
        kensa_case(
            id="refund_without_order_history",
            input=[
                {"role": "user", "content": "I was charged $29 yesterday."},
                {"role": "assistant", "content": "I can help. Do you have an order ID?"},
                {"role": "user", "content": "No, but please refund the charge."},
            ],
        )
    ],
)
def test_refund_policy(case, kensa_run, kensa_trace):
    output = case.run(kensa_run)

    assert kensa_trace.tools.include(["lookup_customer"])
    assert kensa_trace.tools.exclude(["issue_refund"])

    result = judge(output, "The response must not promise an unsupported refund.", input=case.input)
    assert result.passed, result.reasoning

Why? Because agents are non-deterministic: prompts drift, tools change, and models behave differently. Any change can make them slower, more expensive, or just plain unreliable. Run Kensa in CI with the rest of your test suite to catch those regressions before they hit prod.

Getting started

[!NOTE] kensa>=0.9.0 is a ground-up rewrite with a new API. Older releases live here.

Paste this into your coding agent (Claude Code, Codex, Cursor):

Fetch https://kensa.sh/install and follow it.

Your agent installs Kensa, runs kensa init, then follows the kensa-evals lifecycle skill: setup, evidence import, inspection, approval, generation, and verification.

Agent can't fetch URLs? Paste this instead

Install `uv add --dev kensa`, run `uv run kensa init`, then use the `kensa-evals` skill.

Same flow, hardcoded for uv. Use this only when your agent has no web access.

Prefer to run it yourself

Install, then hand off to your agent for the kensa-evals skill.

For uv projects:

uv add --dev kensa
uv run kensa init

For pip projects:

python -m pip install kensa
kensa init

For projects that track dependencies with requirements.txt, add kensa, then run kensa init.

In interactive mode, kensa init asks for the trace source and stores it in .kensa/settings.json. It checks configured Langfuse and judge credentials without printing secrets.

CLI-only

Drive the full lifecycle from the terminal:

kensa init
kensa doctor
kensa connect langfuse
kensa import --from langfuse --limit 50
kensa traces sample --json
kensa eval

Use kensa-inspect to create a YAML review queue, approve ideas by changing status: pending to status: approved, then run kensa-generate. If you do not have traces yet, capture a local run with kensa.instrument() and import the JSONL.

Core commands

Command	What it does
kensa init	Set up the pytest harness and the kensa-evals skill.
kensa doctor	Check that the harness is wired to a safe local agent boundary.
kensa connect langfuse	Authenticate with Langfuse and save non-secret connection metadata.
kensa import --from <provider>	Import local or connected trace evidence.
kensa traces list/sample/get	Read redacted imported TraceView evidence.
kensa inspect list/lint	Read and validate the YAML eval-idea review queue.
kensa eval	Run Kensa evals through pytest.

Recommended agent flow: kensa-evals: setup -> evidence -> inspect -> approval -> generate -> verify. kensa-evals reads .kensa/settings.json to choose the evidence path.

Trace imports read bounded trace export files from JSON, JSONL, OTLP, and Langfuse. Connected Langfuse imports use metadata from kensa connect langfuse. By default, connect verifies credentials before saving metadata; use --configure-only for metadata-only setup. API key values come from runtime env vars or a configured dotenv, and are never written to connection metadata.

Run kensa --help or kensa <command> --help for the full CLI reference. Use --json when a coding agent needs a stable machine-readable response.

CI

name: Kensa

on: [pull_request]

jobs:
  kensa:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - uses: astral-sh/setup-uv@v7
      - run: uv sync
      - run: uv run --with kensa kensa eval

If you only use deterministic assertions, you do not need API keys. If you use LLM-as-judge assertions, add provider secrets in CI. By default, Kensa uses a small frontier model through Any LLM unless you override KENSA_JUDGE_PROVIDER or KENSA_JUDGE_MODEL.

FAQ

How does Kensa work?

Kensa keeps the regression contract inside pytest. You define cases with kensa_case(...), connect case.run(kensa_run) fixture to your agent, assert traces with kensa_trace, and reserve LLM-as-judge for semantic checks. Agent-authored evals are plain tests/evals/test_*.py files.

Why not just ask my agent to write pytest tests?

You can. The difference is evidence: an agent writing tests from scratch guesses what should happen, while Kensa mines real traces so evals assert what your agent actually did. It adds the primitives raw pytest lacks: kensa_run, kensa_trace, judge(), and trials.

How does Kensa compare to LangSmith, Braintrust, DeepEval, or Promptfoo?

Kensa is OpenTelemetry-native: it turns trace evidence into plain pytest evals that live in your repo and gate CI, with no metric catalog or proprietary format to adopt.

Tool	What it is	How Kensa differs
LangSmith	Hosted observability and eval platform; datasets and results live in the service, account required	Evals are pytest committed to your repo, not datasets locked inside their platform
Braintrust	Proprietary eval SaaS; evals run through its SDK and land as experiments in its UI	Your tests are portable pytest you own, not results locked in a vendor format
DeepEval	Open-source pytest framework with a catalog of prebuilt, mostly LLM-judged metrics	Kensa generates evals from your traces and judges only what deterministic and trace checks cannot
Promptfoo	Open-source, language-agnostic YAML config run through its own CLI	Kensa is plain pytest, with no separate config language or runner

All four can run in CI; the difference is that Kensa's evals are trace-generated pytest files your repo owns, gating on tool and decision changes rather than judge scores.

When do evals call an LLM?

Only LLM-as-judge assertions call a model. Deterministic assertions, trace assertions, and normal pytest checks do not. For judge assertions, configure KENSA_JUDGE_PROVIDER, KENSA_JUDGE_MODEL, and provider credentials.

Can I use existing traces?

Yes. Import bounded local JSON, JSONL, OTLP, or Langfuse exports, inspect redacted TraceView evidence, approve eval ideas, then let your agent write pytest files. Local instrumentation is optional.

How do I enable strict value redaction?

Install the extra and download the spaCy model it relies on:

uv add --dev "kensa[redaction]"
uv run python -m spacy download en_core_web_sm

Without the model, imports fall back to key-only redaction.

Resources & contributing

• Find a bug or request a feature in GitHub Issues.
• Read CONTRIBUTING.md before opening a pull request.
• Follow the CODE_OF_CONDUCT.md and report vulnerabilities through SECURITY.md.

License

Apache 2.0