Writing Hegel Conformance Tests

Writing Hegel Conformance Tests

Conformance tests validate that hegel library implementations (hegel-go, hegel-rust, etc.) correctly implement the hegel protocol. They run compiled library binaries against the real hegel server and validate the results.

Architecture: Three Parts

1. ConformanceTest Subclass (hegel-core, src/hegel/conformance.py)

Defines what to test and how to validate:

• params_strategy() — Hypothesis strategy generating random test parameters
• validate(metrics_list, params) — Assertions on the collected metrics
• modes class var — Optional list of modes; framework iterates and adds mode to params
• extra_env() — Optional extra env vars for the binary
• default_test_cases — Override for default number of test cases

The framework (run_conformance_tests) runs each test 5 times with random params via @given.

2. Conformance Binary (library side)

A compiled executable (Go/Rust/Python) that:

1. Parses JSON params from argv[1] (including mode if the test has modes)
2. Reads CONFORMANCE_TEST_CASES and CONFORMANCE_METRICS_FILE from env
3. Connects to the hegel server (the library spawns the server internally)
4. Runs a test through the library that exercises the feature being tested
5. Writes per-test-case client metrics (one JSON line per test case) to CONFORMANCE_METRICS_FILE

Critical principle: the binary is the library's code. It uses the library's API (hegel.MustRun in Go, Hegel::new().run() in Rust). It does NOT reimplement the protocol.

3. Server Instrumentation (hegel-core, src/hegel/server.py)

The server writes per-test-case server metrics to CONFORMANCE_SERVER_METRICS_FILE on each mark_complete. These are things only the server can know (e.g., generate_call_count). The framework merges server metrics with client metrics 1:1 and passes the merged list to validate().

The server is the source of truth. Server-written metrics cannot be faked by the library binary. Use server metrics for anything that validates server-side behavior.

Metrics Flow

Library Binary              Hegel Server                  Framework
    |                            |                            |
    |-- spawns server ---------->|                            |
    |-- run_test --------------->|                            |
    |                            |                            |
    |<-- test_case event --------|                            |
    | writes to METRICS_FILE     | writes to SERVER_METRICS   |
    |  (per test case)           |  (per mark_complete)       |
    |                            |                            |
    |<-- test_done event --------|                            |
    |                            |                            |
    |                            |         reads both files   |
    |                            |         merges 1:1         |
    |                            |         calls validate()   |

The 1:1 merge requires the same number of lines in both files. Each mark_complete produces one server line. Each test function invocation produces one client line. They naturally align.

Self-Test Binaries (Python reference, tests/conformance/tests/)

Hegel-core includes Python reference binaries that use Hypothesis directly (NOT the server):

• Pass skip_server_metrics=True to skip the 1:1 server merge
• Verify the conformance validation logic is correct
• Do NOT test actual server behavior (that's what real library tests do)

Pattern: @given(strategy) generates values, writes metrics in the test function body.

How to Add a New Conformance Test

Step 1: Identify what to validate

Determine what property of a library implementation to test. Identify whether validation requires:

• Client metrics only (e.g., value constraints) — library reports what it observed
• Server metrics (e.g., generate_call_count, interesting_test_cases) — server reports independently

Step 2: Instrument the server (if needed)

If validation requires server-side data, add instrumentation in server.py. Write to CONFORMANCE_SERVER_METRICS_FILE (per-test-case) or CONFORMANCE_SERVER_RUN_METRICS_FILE (per-run). Add a unit test in tests/test_server.py verifying the server writes the expected data.

Step 3: Define the ConformanceTest subclass

In src/hegel/conformance.py, create a subclass:

class MyConformance(ConformanceTest):
    modes: ClassVar[list[str]] = ["mode_a", "mode_b"]  # optional

    def params_strategy(self) -> st.SearchStrategy[dict[str, Any]]:
        return st.just({})  # or a composite strategy

    def validate(self, metrics_list, params):
        for metrics in metrics_list:
            # Check client metrics and/or merged server metrics
            assert metrics["some_field"] == expected

Step 4: Write the self-test binary

In tests/conformance/tests/my_test.py:

#!/usr/bin/env python3
import json
import os
import sys

from hypothesis import given, settings, strategies as st


def main():
    json.loads(sys.argv[1])
    metrics_file = os.environ["CONFORMANCE_METRICS_FILE"]
    test_cases = int(os.environ["CONFORMANCE_TEST_CASES"])

    @settings(max_examples=test_cases, database=None)
    @given(st.integers(...))
    def run(value):
        with open(metrics_file, "a") as f:
            f.write(json.dumps({"value": value}) + "\n")

    run()


if __name__ == "__main__":
    main()

Step 5: Register in test_conformance.py

MyConformance(TESTS_DIR / "my_test.py", skip_server_metrics=True),

The self-test uses skip_server_metrics=True. Real library repos (hegel-go, hegel-rust) create their own binaries that go through the actual server.

Step 6: Library binaries implement the test

A hegel-go binary looks like:

func main() {
    params := parseParams()
    n := conformance.GetTestCases()
    hegel.MustRun(func(s *hegel.TestCase) {
        v := hegel.Draw(s, hegel.Integers[int](0, 100))
        conformance.WriteMetrics(map[string]any{"value": v})
    }, hegel.WithTestCases(n))
}

Common Mistakes

1. Writing a custom client in the binary. The binary should use the library's API, not reimplement the protocol. The library spawns the server.

2. Having the client write server-side metrics. If a metric represents server behavior (like interesting_test_cases), the SERVER must write it. A client can say whatever it wants.

3. Confusing self-test with real test. The Python self-test only validates the validation logic. Real testing happens when hegel-go/hegel-rust binaries run through the actual server.

4. Breaking the 1:1 line count. Client metrics lines must match server metrics lines. Each mark_complete = one server line. Each test function call = one client line.

5. Being defensive when parsing JSON values. Use direct indexing (params["mode"], metrics["value"]) rather than .get(..., default) when reading conformance params or metrics. Missing keys indicate a real bug in the test setup or the library binary, and a KeyError is a clearer signal than silently using a fallback value.

Reference Repos

• hegel-go: internal/conformance/cmd/ has Go binaries, tests/conformance/ has pytest
• hegel-rust: tests/conformance/rust/src/bin/ has Rust binaries
• Both use CONFORMANCE_METRICS_FILE, CONFORMANCE_TEST_CASES, CONFORMANCE_SERVER_METRICS_FILE