/prowler-test-mcp

name: prowler-test-mcp
description: >
  Testing patterns for the Prowler MCP Server: in-memory FastMCP clients, the
  ProwlerAPIClient singleton, JSON:API model builders and mocked httpx transports.
  Trigger: When writing tests under mcp_server/tests/ (tools, models, api_client, auth, sub-servers).
license: Apache-2.0
metadata:
  author: prowler-cloud
  version: "1.1.0"
  scope: [root, mcp_server]
  auto_invoke:
    - "Writing Prowler MCP server tests"
    - "Testing MCP tools or models"
allowed-tools: Read, Edit, Write, Glob, Grep, Bash, WebFetch, WebSearch, Task

Critical Rules

• ALWAYS drive tools through an in-memory client: `async with Client(mcp_root_server)`.

Tool parameters use pydantic `Field(default=...)`, and only FastMCP's wrapper resolves those defaults. Calling a tool method directly with an argument omitted leaves it as a raw `FieldInfo` — which is truthy, so `if email:` silently builds a filter out of the `FieldInfo` repr. Direct calls MUST pass every argument.

• NEVER open a `fastmcp.Client` inside a fixture. FastMCP warns this causes

hard-to-diagnose event-loop issues; open it inline in the test.

• ALWAYS use the `mock_api_client` fixture; NEVER construct a `ProwlerAPIClient`.

Tool instances captured the singleton by reference at import time, so only an in-place patch of `.client` reaches them.

• NEVER clear `SingletonMeta._instances`. It orphans every registered tool on an

instance holding a real `httpx.AsyncClient`. Use `isolated_api_client` if you genuinely need a fresh instance.

• NEVER strip `PROWLER_API_KEY`. Tools are built at import time and a construction

failure is swallowed, so the whole `prowler_*` namespace silently drops to zero tools. It is pinned in `[tool.pytest_env]`.

• For `ProwlerAppAuth`, pass `mode=` / `base_url=` explicitly. Those are resolved in

default arguments, evaluated once at module import, so `monkeypatch.setenv` has no effect on them.

• NEVER assert an exact tool count — every future branch would have to bump it.
• Assert on `result.data` (structured output), not `result.content[0].text`.
• Tests are `test_*.py` (prefix), like the API — not the SDK's `*_test.py` suffix.
• `__init__.py` IS required in every `tests/` subdirectory here (unlike the SDK's

repo-root `tests/`), or same-named modules collide under pytest's import mode.

• Async tests need no marker (`asyncio_mode = "auto"`). Do not use `@pytest.mark.anyio`.
• Warnings are errors (`filterwarnings = ["error", ...]` in `pyproject.toml`): a new

deprecation warning fails the suite. Fix the cause; do not add a blanket filter.

• Use only obviously-fake credentials from `tests.helpers.tokens` (TruffleHog).
• One behaviour per test; keep tests self-contained and order-independent.

---

1. Layout

Mirror the source tree *below the package root* — drop the `prowler_mcp_server/` level, exactly as the SDK maps `prowler/providers/...` to `tests/providers/...`. So `prowler_mcp_server/prowler_app/tools/` is tested in `tests/prowler_app/tools/`.

mcp_server/tests/
├── conftest.py                  # all shared fixtures
├── helpers/                     # jsonapi.py, http.py, assertions.py, tokens.py
├── test_server.py               # mounted-server contract
├── test_health.py
├── prowler_app/{models,tools,utils}/
├── prowler_hub/
└── prowler_documentation/

---

2. Fixtures

| Fixture | Autouse | What it gives you | |---------|---------|-------------------| | `_pinned_environment` | yes | Deterministic env; blocks a developer's `.env` from leaking | | `_no_real_network` | yes | Any real socket connect raises `RuntimeError` | | `_singleton_registry_guard` | yes | Snapshots/restores `SingletonMeta._instances` | | `mock_router` | no | Route registry + request recorder | | `api_client` | no | The live `ProwlerAPIClient` singleton | | `mock_api_client` | no | **The workhorse** — singleton with a mocked transport | | `isolated_api_client` | no | Evicts the singleton, for construction/identity tests | | `mcp_root_server` | no | The mounted root server (session-scoped) | | `health_client` | no | Starlette `TestClient` for `/health` | | `http_request_headers` | no | Injects headers for HTTP-mode auth | | `hub_router` | no | Mocks the Hub sub-server's two sync clients | | `docs_router` | no | Mocks the docs search engine's two sync clients |

`MockRouter`

mock_router.add("GET", "/api/v1/users", json=jsonapi_collection([...]))
mock_router.add("GET", "/api/v1/tasks/t1", json=task_document("t1", "completed"))

mock_router.request_for("GET", "/api/v1/users")     # last request, for header asserts
mock_router.query_params("GET", "/api/v1/users")    # decoded query string
mock_router.paths()                                 # everything requested so far

Register a route more than once to return a sequence — the last response repeats. That is how you drive `poll_task_until_complete` (`executing`, `executing`, `completed`). An unregistered request raises, listing what *was* registered.

---

3. Patterns

**Tool test** — see `assets/mcp_tool_test.py`. Register routes, call through the in-memory client, assert on `result.data` *and* on the recorded request. When a tool chooses between endpoints, assert `mock_router.paths()` — a wrong choice is invisible in the response body.

**Model test** — see `assets/mcp_model_test.py`. Build the document with the `jsonapi` helpers, run `from_api_response()`, assert on both the model and `model_dump()`. `MinimalSerializerMixin` makes those differ, and an absent relationship (`None`) must never be conflated with an empty one (`[]`).

**Contract test** — see `assets/mcp_contract_test.py`. Namespacing and description coverage across every registered tool.

The worked example in the repo is `findings`, covered across both layers in `tests/prowler_app/{models,tools}/test_findings.py`. Read those first — they exercise every foundation capability in one feature.

Reading coverage

Coverage has a meanin