Testing

Overview

GAIA has two test suites:

API (apps/api/tests/) — pytest + pytest-asyncio, ~918 tests, runs in ~15 seconds
Bots (apps/bots/__tests__/) — vitest, covers Discord and Slack adapters

All tests run without external services — databases, Redis, and LLM APIs are mocked at their boundaries.

Running Tests

With mise (recommended)

mise test

Without mise

cd apps/api
uv run pytest

By default, e2e and composio tests are deselected. They require live external services and real API credentials. Run the default command for the full fast suite.

Test Layout

API

apps/api/tests/
├── conftest.py                   # Root fixtures: test app, auth, fake users
├── factories.py                  # Data factories for common models
├── helpers.py                    # Shared test utilities
│
├── api/                          # HTTP endpoint tests (route + auth + response)
│   ├── test_conversations.py
│   ├── test_health.py
│   ├── test_integrations.py
│   ├── test_payments.py
│   ├── test_todos.py
│   └── test_user.py
│
├── unit/                         # Pure unit tests — real logic, mocked I/O
│   ├── agents/                   # Agent routing, state, graph manager
│   ├── middleware/               # Rate limiter, executor
│   ├── models/                   # Schema validation
│   ├── services/                 # Chat, conversation, mail, memory, user, workflow
│   ├── skills/                   # Skills registry
│   ├── tools/                    # Tool registry
│   ├── utils/                    # Command parsing, markdown frontmatter
│   └── workers/                  # ARQ background tasks
│
├── integration/                  # Integration tests — compiled graphs, real app
│   ├── api/                      # Full FastAPI lifecycle with mocked services
│   │   ├── test_chat_endpoints.py
│   │   ├── test_conversation_endpoints.py
│   │   ├── test_health_endpoints.py
│   │   ├── test_integration_endpoints.py
│   │   ├── test_mcp_endpoints.py
│   │   └── test_tools_endpoints.py
│   ├── agents/                   # Compiled LangGraph agent tests
│   ├── db/                       # ChromaDB, Redis, lazy loader
│   └── mcp/                      # MCP connection and token management
│
├── e2e/                          # End-to-end flows (deselected by default)
│   ├── test_create_todo_flow.py
│   ├── test_multi_tool_scenario.py
│   ├── test_send_email_flow.py
│   └── test_workflow_execution.py
│
├── composio/                     # Live Composio integration tests
│   ├── test_calendar.py
│   ├── test_gmail.py
│   └── ...
│
├── services/                     # Additional service-layer tests
│   ├── test_conversation_service.py
│   ├── test_mcp_tools_store.py
│   └── test_user_service.py
│
└── agents/                       # Tool infrastructure runtime
    └── test_tool_infra_runtime.py

Bots

apps/bots/__tests__/
├── discord/
│   ├── adapter.test.ts           # Discord message adapter
│   └── embed.test.ts             # Discord embed formatting
├── shared/
│   ├── adapter/
│   │   └── rich-renderer.test.ts # Rich text rendering
│   └── utils/
│       ├── commands.test.ts      # Command parsing
│       ├── formatters.test.ts    # Message formatters
│       └── text-utils.test.ts    # Text utilities
└── slack/
    ├── adapter.test.ts           # Slack message adapter
    └── mention.test.ts           # Mention handling

Test Markers

Marker	What it covers	External deps?
`unit`	Individual functions and classes with mocked I/O	None
`integration`	Real FastAPI app lifecycle or compiled graphs, mocked services	None
`e2e`	Full agent runs with near-real services	Redis, MongoDB
`composio`	Live Composio API calls	Composio credentials

The default pytest.ini config runs everything except e2e and composio.

What Gets Tested

API Endpoints (auth, routing, response contracts)

Tests in tests/api/ and tests/integration/api/ run the real FastAPI app via httpx.AsyncClient. They verify:

Correct HTTP status codes (200, 401, 403, 422)
Auth enforcement on every protected route
Response body shape and required fields
SSE content type (text/event-stream) and headers (x-stream-id, cache-control)
Error paths: Redis unavailable → [STREAM_ERROR] in SSE body
Pagination parameter validation

Services are mocked at the boundary (patch("app.api.v1.endpoints.conversations.create_conversation_service")), so the routing and response logic is exercised without hitting a database.

Service Layer (core business logic)

Tests in tests/unit/services/ import production functions directly and test real logic with mocked databases and LLM clients. Covered services:

chat_service: run_chat_stream_background, _initialize_new_conversation, _save_conversation_async, extract_tool_data, _extract_response_text
conversation_service: CRUD operations, pagination, read/unread state
user_service: User creation, lookup, preference management
memory_service: Memory extraction and persistence
mail_service: Email parsing, send logic
workflow_service: Workflow creation and trigger evaluation

Agent Routing and Graph Construction

Tests in tests/unit/agents/ and tests/integration/agents/ call the real create_agent factory and build_comms_graph / build_executor_graph builders. Verified behaviors:

Conditional edge "agent" is registered with routing targets
"tools" is always reachable from the agent node
"select_tools" appears only when retrieve_tools is enabled
"end_graph_hooks" appears only when hooks are provided
Plain text response → no ToolMessages (routes to END / end_graph_hooks)
Tool call response → ToolMessage produced with correct tool_call_id
Multiple tool calls → all produce ToolMessages
State accumulates across turns via InMemorySaver checkpointing
add_memory and search_memory are wired into the comms agent tool registry

Workers and Background Tasks

Tests in tests/unit/workers/ cover the ARQ background task functions:

cleanup_tasks: old conversation pruning, orphan cleanup
memory_tasks: background memory extraction scheduling
reminder_tasks: reminder triggering and delivery
user_tasks: user lifecycle operations
workflow_tasks: cron trigger evaluation

MCP and Tool Registry

Tests cover the MCP tool store, connection flows, and token management:

ChromaStore indexing with namespace metadata
Redis cache hit/miss paths
MCP server connection lifecycle
Token refresh and expiry handling

Bot Adapters (Discord and Slack)

Tests in apps/bots/__tests__/ use vitest and run against production adapter code. They verify:

Discord message adapter formatting and embed rendering
Slack message adapter formatting and mention handling
Shared rich-text renderer output
Command parsing utilities
Text formatting helpers

Run with mise test:bots or cd apps/bots && pnpm vitest run.

Test Infrastructure

Root conftest.py

The root conftest.py (at tests/conftest.py) sets up the test environment before any app modules load:

# Prevents connections to real external services
os.environ.setdefault("MONGO_DB", "mongodb://localhost:27017/gaia_test?...")
os.environ.setdefault("REDIS_URL", "redis://localhost:6379/0")

# Patches that persist across all tests
_patches = [
    patch("app.config.secrets.inject_infisical_secrets", return_value=None),
    patch("app.db.mongodb.mongodb.MongoDB.ping", return_value=None),
    patch("app.decorators.rate_limiting.tiered_limiter.check_and_increment", ...),
]

The test app is created once per session with a no-op lifespan (_noop_lifespan) so database connections are never attempted. Auth is bypassed via app.dependency_overrides[get_current_user] = lambda: FAKE_USER.

Key fixtures

Fixture	Scope	What it provides
`test_app`	session	FastAPI app with no-op lifespan and auth override
`client`	function	`AsyncClient` bound to the test app, authenticated
`unauthed_client`	function	`AsyncClient` without auth override (gets 401)
`fake_user`	function	Dict with test user data
`mock_mongodb`	function	`AsyncMock()` for MongoDB operations

Writing New Tests

The Golden Rule

If you deleted the production function this test targets, would the test still fail?

If the answer is “no”, the test is worthless. Always import from app. directly.

# WRONG — tests nothing about GAIA
from langgraph.graph import StateGraph
graph = StateGraph(SimpleState)  # your own graph, not GAIA's

# RIGHT — tests GAIA's graph factory
from app.override.langgraph_bigtool.create_agent import create_agent
builder = create_agent(llm=mock_llm, tool_registry=registry, ...)

Mock at the boundary

Mock external I/O (databases, HTTP, Redis), never the logic under test.

# WRONG — mocks the function being tested
with patch("app.services.chat_service.run_chat_stream_background") as mock:
    mock.return_value = "result"
    result = run_chat_stream_background(...)  # calls the mock, tests nothing

# RIGHT — mocks the LLM dependency, tests real service logic
with patch("app.services.chat_service.agent.ainvoke") as mock_agent:
    mock_agent.return_value = {"messages": [AIMessage(content="hello")]}
    result = await run_chat_stream_background(body, user, stream_id)
    assert result["status"] == "completed"

Assert on behavior, not mock calls

# WRONG — only proves your mock setup works
mock_service.create.assert_called_once_with(data)

# RIGHT — proves the endpoint returns what it should
resp = await client.post("/api/v1/conversations", json=data)
assert resp.status_code == 200
assert resp.json()["id"] == data["conversation_id"]

Cover error paths

Production bugs cluster in error handling. Always test what happens when a dependency fails.

async def test_handles_redis_unavailable(self, test_client):
    """When Redis is down, SSE body should contain [STREAM_ERROR]."""
    with patch("app.api.v1.endpoints.chat.redis_cache") as mock_redis:
        mock_redis.redis = None  # simulate Redis down
        response = await test_client.post("/api/v1/chat-stream", json=body)

    assert response.status_code == 200
    assert "[STREAM_ERROR]" in response.text

Coverage Configuration

Coverage is configured in pytest.ini:

[coverage:run]
source = app
omit =
    app/config/*
    app/patches.py
    app/static/*

[coverage:report]
show_missing = true
fail_under = 3

The fail_under = 3 threshold is intentionally low. Run coverage locally to track actual coverage:

cd apps/api
uv run pytest --cov=app --cov-report=html
open htmlcov/index.html

Get Started

CLI

Self-Hosting

Developers

Messaging Platforms (Bots)

Extra Developer Config

Overview

Running Tests

With mise (recommended)

Without mise

Test Layout

API

Bots

Test Markers

What Gets Tested

Test Infrastructure

Root conftest.py

Key fixtures

Writing New Tests

The Golden Rule

Mock at the boundary

Assert on behavior, not mock calls

Cover error paths

Coverage Configuration

Get Started

CLI

Self-Hosting

Developers

Messaging Platforms (Bots)

Extra Developer Config

​Overview

​Running Tests

​With mise (recommended)

​Without mise

​Test Layout

​API

​Bots

​Test Markers

​What Gets Tested

​Test Infrastructure

​Root conftest.py

​Key fixtures

​Writing New Tests

​The Golden Rule

​Mock at the boundary

​Assert on behavior, not mock calls

​Cover error paths

​Coverage Configuration

Overview

Running Tests

With mise (recommended)

Without mise

Test Layout

API

Bots

Test Markers

What Gets Tested

Test Infrastructure

Root conftest.py

Key fixtures

Writing New Tests

The Golden Rule

Mock at the boundary

Assert on behavior, not mock calls

Cover error paths

Coverage Configuration