pytest-fakellm

Pytest fixtures for fakellm, the mock OpenAI/Anthropic server. Spin up a server, get a clean state per test, and assert on what your code sent — with zero boilerplate.

pip install pytest-fakellm

Once installed, the fixtures are available automatically — no imports, no conftest.py setup.

The point

Without the plugin, using fakellm in a test means starting the server, wiring a client to its URL, resetting state, and tearing it all down yourself, in every test. With the plugin, that becomes:

def test_agent_handles_search(fakellm):
    fakellm.set_config_text("""
    version: 1
    rules:
      - name: summarize
        when: { messages_contain: "research" }
        respond: { content: "Based on the search, I found what you were looking for." }
    """)
    result = run_my_agent(fakellm.openai_client(), prompt="Please research fakellm")
    assert "found what you were looking for" in result
    fakellm.assert_request_count(1)

The server starts once per session, state is reset before each test, and everything is torn down at the end. You never touch a port number or a subprocess.

Fixtures

Fixture	What you get
fakellm	A FakellmServer handle with fresh conversation state for the test.
fakellm_openai	A ready openai.OpenAI client pointed at the (reset) server.
fakellm_anthropic	A ready anthropic.Anthropic client pointed at the (reset) server.
fakellm_logs	Opt-in. Dumps the server's output into the failure report only if the test fails — handy for debugging without cluttering passing runs.

FakellmServer handle

Clients and URLs:

• openai_client(**kwargs) / anthropic_client(**kwargs) — clients pointed at the server.
• openai_base_url / anthropic_base_url — raw URLs if you build your own client.

Configuring rules:

• set_config_text(yaml) — write rules inline and reload.
• load_rules(path) — load rules from a file and reload.
• reset() — clear conversation state (done for you between tests).
• reload() — re-read the config from disk.

Inspecting what happened:

• stats() / conversations() — the admin JSON, for assertions.
• request_count — absolute session total of requests seen.
• requests_since_reset — requests made during the current test (per-test count).
• tool_results_seen() — total tool results the server observed across all conversations.

Assertions (raise AssertionError with a readable message on failure):

• assert_request_count(expected) — exactly expected requests were seen.
• assert_rule_matched(rule_name, min_times=1) — a named config rule matched at least min_times.
• assert_tool_results_seen(min_results=1) — at least min_results tool results were fed back.

Error injection:

• set_error_simulation(status, error_message="...", *, when=None, name="...") — make the server return an HTTP error for matching requests.

See Assertions and error simulation for details.

Assertions and error simulation

Asserting on traffic

After your code runs, assert on what the server saw:

def test_agent_makes_one_call(fakellm):
    fakellm.set_config_text("""
    version: 1
    rules:
      - name: answer
        when: { messages_contain: "weather" }
        respond: { content: "It is sunny." }
    """)

    run_my_agent(fakellm.openai_client(), prompt="what is the weather?")

    fakellm.assert_request_count(1)
    fakellm.assert_rule_matched("answer")

assert_rule_matched reads the per-rule match counts the server keeps in stats()["by_rule"]. Requests that matched no rule are counted under "<fallthrough>", so you can assert on those too.

Both assert_request_count and assert_rule_matched count only what happened during the current test. fakellm's stats are cumulative for the whole server process (a reset() clears conversations but not stats), so the fakellm fixture records a baseline at the start of each test and these helpers measure the delta from it. If you want the raw numbers, request_count is the absolute session total and requests_since_reset is the per-test count.

Tool results

If your agent calls a tool and feeds the result back to the model, the server counts those tool results:

def test_agent_used_a_tool(fakellm):
    run_my_tool_using_agent(fakellm.openai_client(), prompt="search for X")
    fakellm.assert_tool_results_seen(1)

A deliberate limitation worth knowing: fakellm records only a count of tool results per conversation — it does not retain or expose tool names. So you can confirm that a tool result came back, but not which tool produced it. There is intentionally no assert_tool_called("search"), because the server transmits no data to implement it against. If you need to assert on a specific tool, match on it in a rule (when: { tools_include: "search" }) and then use assert_rule_matched on that rule's name.

Simulating errors

To exercise your retry/back-off and error-handling paths, make the server return an HTTP error:

import openai

def test_agent_retries_on_rate_limit(fakellm):
    fakellm.set_error_simulation(429, "slow down")
    client = fakellm.openai_client()

    with pytest.raises(openai.RateLimitError):
        client.chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": "hello"}],
        )

set_error_simulation works for both the OpenAI and Anthropic endpoints, emitting the error in each API's native shape. status must be >= 400 (fakellm only treats those as errors). Pass a when= matcher dict to scope the error to specific requests, e.g. set_error_simulation(503, "down", when={"messages_contain": "search"}); omit it to fail every request. The error message is YAML-serialized safely, so quotes, colons, and newlines in the message won't corrupt the config.

Surfacing server logs on failure

Add the fakellm_logs fixture to a test and, if that test fails, the server's output for that test is attached to the failure report. Passing tests stay quiet:

def test_something_tricky(fakellm, fakellm_logs):
    ...
    assert result == expected   # on failure, server logs appear in the report

Configuration

Set a starting config file via the command line:

pytest --fakellm-config=tests/fixtures/rules.yaml

or in pyproject.toml / pytest.ini:

[tool.pytest.ini_options]
fakellm_config = "tests/fixtures/rules.yaml"

If you don't set one, a temporary empty config is created so set_config_text and load_rules work immediately.

--fakellm-startup-timeout (default 10.0) controls how long the fixture waits for the server to come up.

Client extras

openai_client() and anthropic_client() require the respective SDKs. Install what you need:

pip install "pytest-fakellm[openai]"      # adds openai
pip install "pytest-fakellm[anthropic]"   # adds anthropic

License

MIT