pytest-skill-engineering 0.0.2

The testing framework for skill engineering. Test tool descriptions, prompt templates, agent skills, and custom agents with real LLMs. AI analyzes results and tells you what to fix.

pytest-skill-engineering

Skill Engineering. Test-driven. AI-analyzed.

A pytest plugin for skill engineering — test your MCP server tools, prompt templates, agent skills, and .agent.md instruction files with real LLMs. Red/Green/Refactor for the skill stack. Let AI analysis tell you what to fix.

Why?

Modern AI systems are built on skill engineering — the discipline of designing modular, reliable, callable capabilities that an LLM can discover, invoke, and orchestrate to perform real tasks. Skills are what separate "text generator" from "coding agent that actually does things."

An MCP server is the runtime for those skills. It doesn't ship alone — it comes bundled with the full skill engineering stack: tools (callable functions), prompt templates (server-side reasoning starters), agent skills (domain knowledge and behavioral guidelines), and .agent.md instruction files (specialist sub-agent definitions in VS Code / Claude Code format). Users layer on their own prompt files (slash commands like /review) on top.

Your unit tests cover the server code. Nothing covers the skill stack. And the skill stack is what the LLM actually sees.

Skill engineering breaks in ways code tests can't catch:

• The tool description is too vague — the LLM picks the wrong tool or passes garbage parameters
• The prompt template renders correctly but the assembled message confuses the LLM
• A prompt file's slash command produces garbage output because the instructions are ambiguous
• The skill has the right facts but is structured so poorly the LLM skips it
• The .agent.md file has the right tools listed but the description is too vague to trigger subagent dispatch

And when you're improving it — how do you know version A is better than version B?

Skill engineering is iterative — prompt tuning, tool description refinement, .agent.md instructions, skill structure. You need A/B testing built in. Run both versions, same prompts, and let the leaderboard tell you which one wins on pass rate and cost.

That's what pytest-skill-engineering does: test the full skill engineering stack, compare variants, and get AI analysis that tells you exactly what to fix.

How It Works

Write tests as natural language prompts — you assert on what happened. If a test fails, your tool descriptions or skill need work, not your code:

1. Write a test — a prompt that describes what a user would say
2. Run it — an LLM tries to use your tools and fails
3. Fix the skill stack — improve tool descriptions, schemas, prompts, or .agent.md instructions until it passes
4. AI analysis tells you what else to optimize — cost, redundant calls, unused tools

pytest-skill-engineering ships two test harnesses:

	Eval + eval_run	CopilotEval + copilot_eval
Runs the LLM	Pydantic AI synthetic loop	Real GitHub Copilot (CLI SDK)
Model	Any provider (Azure, OpenAI, Copilot)	Copilot's active model only
MCP auth	You supply tokens / env vars	Copilot handles OAuth automatically
Introspection	Full per-call (tool name, args, timing)	Summary (tool names, final response)
Cost tracking	USD per test (via litellm pricing)	Premium requests (Copilot billing)
Setup	API keys + model config	gh auth login (Copilot subscription)

Eval + eval_run — bring your own model

You configure the model, wire up MCP servers directly, and get full per-call introspection. Best for iterating on tool descriptions, A/B testing model variants, and cheap CI runs:

from pytest_skill_engineering import Eval, Provider, MCPServer

async def test_balance_query(eval_run):
    agent = Eval(
        provider=Provider(model="azure/gpt-5-mini"),
        mcp_servers=[MCPServer(command=["python", "-m", "my_banking_server"])],
    )
    result = await eval_run(agent, "What's my checking balance?")
    assert result.success
    assert result.tool_was_called("get_balance")

CopilotEval + copilot_eval — use the real coding agent

Runs the actual GitHub Copilot coding agent — the same one your users have. No model setup, no API keys. Best for end-to-end testing: OAuth handled automatically, skills and custom agents loaded natively:

from pytest_skill_engineering.copilot import CopilotEval

async def test_skill(copilot_eval):
    agent = CopilotEval(skill_directories=["skills/my-skill"])
    result = await copilot_eval(agent, "What can you help me with?")
    assert result.success

→ Choosing a Test Harness — full trade-off guide

AI Analysis

AI analyzes your results and tells you what to fix: which model to deploy, how to improve tool descriptions, where to cut costs. See a sample report →

Quick Start

Using GitHub Copilot? Zero setup:

uv add pytest-skill-engineering[copilot]
gh auth login  # one-time
pytest tests/

Using your own model (Azure, OpenAI, Anthropic…):

uv add pytest-skill-engineering
export AZURE_API_BASE=https://your-resource.openai.azure.com/
az login
pytest tests/

AI Analysis judge model (optional but recommended)

The AI analysis report needs a model to generate insights. Configure it in pyproject.toml:

GitHub Copilot:

[tool.pytest.ini_options]
addopts = "--aitest-summary-model=copilot/gpt-5-mini"

Azure OpenAI:

[tool.pytest.ini_options]
addopts = "--aitest-summary-model=azure/gpt-5.2-chat"

Features

• MCP Server Testing — Real models against real tool interfaces and bundled prompt templates
• Prompt File Testing — Test VS Code .prompt.md and Claude Code command files (slash commands) with load_prompt_file() / load_prompt_files()
• CLI Server Testing — Wrap CLIs as testable tool servers
• Real Coding Agent Testing — CopilotEval + copilot_eval runs the actual Copilot coding agent (native OAuth, skill loading, custom agent dispatch, exact user experience)
• .agent.md Testing — Load .agent.md files with Eval.from_agent_file() to test instructions with any model, or use load_custom_agent() + CopilotEval to test real custom agent dispatch
• Eval Comparison — Compare models, skills, .agent.md versions, and server configurations
• Eval Leaderboard — Auto-ranked by pass rate and cost
• Multi-Turn Sessions — Test conversations that build on context
• AI Analysis — Actionable feedback on tool descriptions, prompts, and costs
• Multi-Provider — Any model via Pydantic AI (OpenAI, Anthropic, Gemini, Azure, Bedrock, Mistral, and more)
• Copilot SDK Provider — Use copilot/gpt-5-mini for all LLM calls (judge, insights, scoring) — zero additional setup with pytest-skill-engineering[copilot]
• Clarification Detection — Catch evals that ask questions instead of acting
• Semantic Assertions — Built-in llm_assert fixture powered by pydantic-evals LLM judge
• Multi-Dimension Scoring — llm_score fixture for granular quality measurement across named dimensions
• Image Assertions — llm_assert_image for AI-graded visual evaluation of screenshots and charts
• Cost Estimation — Automatic per-test cost tracking with pricing from litellm + custom overrides

Who This Is For

• MCP server authors — Validate that LLMs can actually use your tools
• Copilot skill authors — Test skills and .agent.md instructions exactly as users experience them
• Eval builders — Compare models, prompts, and skills to find the best configuration
• Teams shipping AI systems — Catch LLM-facing regressions in CI/CD

Documentation

📚 Full Documentation

Requirements

• Python 3.11+
• pytest 9.0+
• An LLM provider (Azure, OpenAI, Anthropic, etc.) or a GitHub Copilot subscription (pytest-skill-engineering[copilot])

Acknowledgments

Inspired by agent-benchmark.

License

MIT