Testing

cursorpipe has three test suites that run in order of increasing requirements.

Prerequisites

Install dev dependencies:

pip install -e ".[dev]"

For integration tests you also need:

• Cursor Agent CLI installed (installation guide)
• Authentication configured via agent login or CURSORPIPE_API_KEY

Test suites

Preflight checks

Validates that your environment is ready — agent binary found, authentication configured, API reachable. Run this first to get clear error messages instead of cryptic failures deeper in the suite.

pytest tests/test_preflight.py -v

Check	What it verifies
Agent on PATH	The agent binary is discoverable
Agent is executable	The resolved binary can actually run
Agent version	agent --version exits cleanly
Authentication	API key, auth token, or login session exists
Connectivity	A minimal prompt returns a response
Model discovery	agent --list-models returns at least one model

Unit tests

Fast, offline tests for internal components — config loading, Pydantic models, error hierarchy, and the NDJSON stream parser. No agent binary or network access required.

pytest tests/test_unit.py -v

Integration tests

End-to-end tests that make real API calls through the Cursor agent. These are slow and consume API quota.

pytest tests/test_integration.py -v -m integration

By default all integration tests use the auto model (Cursor picks the best available model for your account). Override this with the CURSORPIPE_TEST_MODEL environment variable to target a specific model:

# bash / macOS / Linux / WSL
export CURSORPIPE_TEST_MODEL=claude-4.5-sonnet-thinking
pytest tests/test_integration.py -v -m integration

# Windows (PowerShell)
$env:CURSORPIPE_TEST_MODEL = "claude-4.5-sonnet-thinking"
pytest tests/test_integration.py -v -m integration

Test	What it verifies
Subprocess generate	Single-call via subprocess transport
Subprocess streaming	Streaming via subprocess transport
ACP generate	Single-call via persistent ACP transport
ACP streaming	Streaming via ACP transport
ACP consecutive calls	Two consecutive ACP calls succeed (ACP uses Cursor's auto-selected model; it cannot switch models per-call)
AUTO specific-model routing	Specific model name routes to subprocess; "auto" stays on ACP
ACP multi-turn session	Session memory across turns
ACP session streaming	session.stream_prompt() yields chunks
Model discovery	list_models() returns available models
Auto fallback	AUTO strategy falls back from ACP to subprocess

Running everything

pytest -v

This runs all three suites. Integration tests are skipped automatically if the agent binary or authentication is not available.

Skipping integration tests

If you only want the fast offline tests:

pytest -m "not integration" -v

---

Testing the HTTP server

Quick smoke test

Start the server and hit the health endpoint:

# Terminal 1: start the server
cursorpipe-server

# Terminal 2: test health
curl http://localhost:8080/health
# {"status":"ok"}

Test completions

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-4.5-sonnet-thinking","messages":[{"role":"user","content":"Reply with OK"}]}'

Test streaming

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-4.5-sonnet-thinking","messages":[{"role":"user","content":"Count to 5"}],"stream":true}'

Test model listing

curl http://localhost:8080/v1/models

Test with OpenAI SDK

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="unused")
r = client.chat.completions.create(
    model="claude-4.5-sonnet-thinking",
    messages=[{"role": "user", "content": "Reply with OK"}],
)
assert "OK" in r.choices[0].message.content
print("OpenAI SDK test passed")

Test with Docker

docker compose up -d
curl http://localhost:8080/health
# {"status":"ok"}

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-4.5-sonnet-thinking","messages":[{"role":"user","content":"Reply with OK"}]}'

docker compose down