Ensure Test Coverage

Analyze a single evaluation's test coverage against repository testing conventions. This skill can both review existing tests and create missing ones.

Expected Arguments

When invoked, this skill expects:

• eval_name (required): The name of one or more evaluations to check (e.g., gpqa, healthbench)

If not provided, ask the user.

Workflow

Phase 1: Discover Testable Components

Analyze the eval's source code in src/inspect_evals/<eval_name>/ to build a complete inventory of what needs testing.

1.1 Scan for decorated functions

Use AST analysis or grep to find all functions decorated with:

• @task - Task entry points (and their parameters/variants)
• @solver - Custom solvers
• @scorer - Custom scorers
• @tool - Custom tools

For each, record: function name, file path, line number, parameters.

1.2 Scan for dataset patterns

Look for:

• record_to_sample functions (named exactly or as a lambda/callable passed to hf_dataset/csv_dataset/json_dataset)
• hf_dataset() or load_dataset() calls (indicates HuggingFace datasets)
• csv_dataset() or json_dataset() calls
• Dataset URLs or paths (for understanding data sources)

1.3 Scan for sandbox usage

Look for:

• compose.yaml or docker-compose.yaml files in the eval's data/ directory
• sandbox() calls in source code
• SandboxEnvironmentSpec references

1.4 Scan for utility functions

Look for non-trivial helper functions that contain meaningful logic (not just thin wrappers). Functions with:

• Complex branching (if/elif chains)
• Data transformations
• String parsing/formatting
• Score calculations
• File I/O operations

1.5 Identify task variants

Check if the eval has multiple @task functions or a single task with parameters that create meaningfully different variants (e.g., different subsets, different Docker environments, different scorers).

Phase 2: Run Autolint and Measure Coverage

2.1 Run autolint

Run all autolint checks in one command to get a baseline of structural test coverage:

uv run inspect-evals-lint <eval_name>

This checks: tests_exist, tests_init, e2e_test, record_to_sample_test, custom_solver_tests, custom_scorer_tests, custom_tool_tests (plus non-test checks). You do not need to report to the user specific details of this beyond what the autolint tool itself tells them.

If autolint reports failures, those become Priority 1 items to fix in Phase 4.

2.2 Measure line coverage with pytest-cov

Run pytest-cov to get line-level coverage percentages. Include --runslow to run slow tests — they often exercise significantly more code:

uv run pytest tests/<eval_name>/ --cov=src/inspect_evals/<eval_name> --cov-report=term-missing -q --runslow

This shows per-file coverage and which specific lines are missed. Record:

• Overall coverage percentage
• Per-file coverage percentages
• Which lines are missed (the Missing column)

Read the source code for missed lines to understand what they do. Categorize them as:

• Testable: Logic that can be exercised with unit tests or mocked E2E tests (branches, calculations, parsing, error handling)
• Requires LLM/sandbox: Lines that only run with a real model response or Docker container (acceptable to miss in fast tests)
• Dead code or error handlers: Defensive code that's hard to trigger (lower priority)

There is no fixed coverage threshold. Instead, use judgment: all major branches and behaviors should have tests. Focus on whether important logic paths are exercised, not on hitting a number. When evaluating gaps, ask: "If this code path broke, would we catch it?" If the answer is no and the path contains meaningful logic (not just framework glue), it's a gap worth addressing. It is entirely plausible that some decently large gaps include the same code that is checked elsewhere, but if so, this should be justified to the user. Do not assume a gap is fine because testing it would be difficult - that likely indicates a gap in our ability to mock relevant entities for testing. Let the user decide what's worth doing.

Reference points from existing evals as of 13/02/2026:

• gpqa (simple eval, 17 statements, no custom scorer/solver): 100% coverage
• healthbench (complex eval, 317 statements, LLM-graded scorer): 91% overall with --runslow. The remaining gaps are the LLM judge call path — testing it would require mocking the judge model, which is valuable but not trivial. This likely requires a few new tests.
• gdm_self_reasoning (complex eval, 945 statements, Docker-heavy): 51% overall with --runslow. Most missed code is in solver files that only execute inside Docker sandboxes with real model interactions — these should be mocked with our utility functions, and if this is not possible that indicates a problem. This should be tested more thoroughly.

When evaluating gaps, ask: "If this code path broke, would we catch it?" If the answer is no and the path contains meaningful logic (not just framework glue or duplicated code that is exercised elsewhere), it's a gap worth addressing. If the gap involves duplicated code that is addressed elsewhere, consider whether it can be refactored away without harming readability / maintainability too much. If so, recommend this to the user.

Worked example: healthbench scorer.py at 87% coverage:

Name                                               Stmts   Miss  Cover   Missing
--------------------------------------------------------------------------------
src/inspect_evals/healthbench/scorer.py              195     26    87%   37, 276, 285, 417-433, 491-493, 498-501, 586, 599, 638

Not every missed line is a problem. Here are three real examples from healthbench to illustrate the judgment needed:

(a) Line 37 — defensive guard, not a problem:

def macro_f1(y_true: list[bool], y_pred: list[bool]) -> float:
    from sklearn.metrics import f1_score

    if not y_true or not y_pred or len(y_true) != len(y_pred):
        return 0.0 # Line 37

    return float(f1_score(y_true, y_pred, labels=[False, True], average="macro"))

This is a simple input validation guard returning a default. Missing coverage here is fine. Adding tests here would just add unnecessary bloat to the test package.

(b) Lines 417-433 — untested branch with real logic, worth testing:

        if include_subset_scores:
            category_scores = calculate_category_scores(scores, ...)

            # Add flattened axis scores
            for axis, data in category_scores["axis_scores"].items():
                prefix = f"axis_{axis}"
                result[f"{prefix}_score"] = data["score"]
                result[f"{prefix}_std"] = data["std_error"]
                result[f"{prefix}_n_samples"] = data["n_samples"]
                result[f"{prefix}_criteria_count"] = data["criteria_count"]

            # Add flattened theme scores
            for theme, data in category_scores["theme_scores"].items():
                prefix = f"theme_{theme}"
                result[f"{prefix}_score"] = data["score"]
                result[f"{prefix}_std"] = data["std_error"]
                result[f"{prefix}_n_samples"] = data["n_samples"]
                result[f"{prefix}_criteria_count"] = data["criteria_count"]

No single line here is complex, but this is an entire code path — the include_subset_scores=True branch — that is never exercised. There is sufficient complexity here that it is possible for a detail to be missed which a test can catch.

(c) Lines 491-493, 498-501 — borderline, but examining the code suggests worth testing:

    grading_response_dict = {}
    max_retries = 3
    for attempt in range(max_retries):
        try:
            judge_response = await judge.generate(input=grader_prompt, ...)
            grading_response = judge_response.completion
            grading_response_dict = parse_json_to_dict(grading_response)

            if "criteria_met" in grading_response_dict:
                label = grading_response_dict["criteria_met"]  # line 491
                if label is True or label is False:
                    break
            logger.warning(
                f"Grading failed due to bad JSON output, retrying... (attempt {attempt + 1})"
            )

        except Exception as e:  # line 498
            logger.warning(f"Grading attempt {attempt + 1} failed: {str(e)}")
            if attempt == max_retries - 1:
                grading_response_dict = {
                    "criteria_met": False,
                    "explanation": f"Failed to parse judge response: {str(e)}",
                }

    criteria_met = grading_response_dict.get("criteria_met", False)

These lines individually look like fairly simple validation guards, but represent an important check to ensure the judge doesn't get caught in an infinite loop. A single test should be able to check both of these, and does seem worth doing.

2.3 Evaluate test quality (beyond autolint and coverage)

Autolint checks are shallow (function name present in test files). Coverage measures quantity, not quality. This step evaluates whether tests actually test the right things. Read each test file and check:

Assertion quality and task wiring:

• Every non-trivial test should name the realistic regression it would catch in its docstring. If no plausible source change would make the test fail, do not add it.
• Do not mock or substitute the unit under test. Mock external boundaries such as network calls, Docker, Hugging Face, and grader models instead.
• Expected values should be independently specified where that is what makes the test discriminating. Prefer hand-computed literals for behavioural outputs. Production constants may be appropriate for stable identifiers such as revision hashes or registry names when the contract under test is that the task passes that named constant through; use a literal instead when the test is intended to freeze the exact identifier itself.
• Reject assertions that cannot meaningfully fail: bounds true by construction, prefix-only checks that ignore the meaningful payload, assertions after fixtures that already enforce the same condition, and output-equals-input round trips.
• Characterisation tests may be used to demonstrate non-obvious behaviour of library or framework code, or to document a repository-level compatibility contract. Start the docstring with Characterisation: and state the library/framework behaviour being pinned. Eval tests should otherwise exercise custom eval behaviour.
• For every non-trivial @task, verify benchmark-critical wiring that can change benchmark comparability, such as scorer identity, epochs and reducer, pinned dataset revision, subset-to-mapper pairing, and eval-specific configuration flags. Assert solver/attempt settings only when the eval deliberately treats them as benchmark-defining. Do not pin prompts, tool lists, or Inspect framework defaults just because they are visible on the constructed task. Use assert_task_structure where appropriate.
• Treat unusually high test-to-source LOC as a review signal, not a quality target.
• See references/test-patterns.md for the task-wiring recipe and examples of the banned low-signal patterns.

Scorer tests - the most common quality gap:

• Non-trivial scorers must test actual scoring logic with real inputs, not just isinstance(scorer, Scorer) type checks
• Tests should verify both CORRECT and INCORRECT outcomes
• Failure paths should return INCORRECT with an explanatory message, not crash
• Trivial wrappers around Inspect primitives (e.g., match(), includes()) only need type-check
• Sandbox-based scorers (using sandbox().exec()) must mock the sandbox and test with ExecResult objects — not just isinstance. See the "Sandbox-Based Scorer Test" pattern in Phase 4.
• LLM-graded scorers (using get_model()) must mock the model and verify parsing of judge responses. See the "LLM-Graded Scorer Test" pattern in Phase 4.
• Multi-scorer reducers: When a scorer uses multi_scorer() with a custom reducer, the reducer must be tested separately with constructed Score objects. See the "Multi-Scorer Reducer Test" pattern in Phase 4.
• Dispatch scorers: When a scorer dispatches to different sub-scorers based on metadata, mock the dispatch key and verify the correct branch is taken.
• Parametrize test cases: Use @pytest.mark.parametrize instead of writing many near-identical test functions that differ only in input/expected values. If a test would need four or more parametrize arguments, prefer individual named tests for readability.

Solver tests - similar quality gap to scorers:

• Thin wrappers (just assembling chain([system_message(...), generate()]) or wrapping react()) only need isinstance. If already exercised by E2E tests, the isinstance test can be omitted entirely.
• Non-trivial solvers with branching logic (e.g., dispatching based on metadata, constructing different prompts, modifying state.tools) must test the actual branches — create a TaskState, run the solver with a mock generate, and assert state changes.
• If a solver is already exercised by an E2E test that covers its logic paths, a standalone isinstance test is redundant and should be removed.
• Solvers that use sandbox().exec() should be tested with mocked sandbox, same pattern as sandbox-based scorers.

Tool tests:

• Error cases should use pytest.raises(ToolError), not just test the happy path
• Sandbox tools should use the @solver + state.metadata["test_passed"] pattern (see Phase 4)

Dataset/record_to_sample tests:

• Tests should use a real example from the actual dataset, not fabricated data
• Tests should verify sample.id, sample.input, sample.target, sample.metadata

E2E tests:

• Each meaningfully different task variant should have its own E2E test
• "Meaningfully different" = different Docker environment, different scorer, different solver pipeline (NOT just different questions from the same dataset)

Pytest markers (commonly missing or wrong):

• @pytest.mark.dataset_download if test instantiates a dataset (including E2E tests)
• @pytest.mark.huggingface for HF datasets -- do NOT implement manual have_hf_token() checks, pytest-inspect-evals handles it
• @pytest.mark.docker for sandbox tests (NOT @pytest.mark.skip)
• @pytest.mark.slow(<seconds>) with actual observed duration for tests >10s

Phase 3: Report Coverage Gaps

Present findings organized by priority:

Priority 1 - Autolint failures

These will block CI/review. List each failing autolint check.

Priority 2 - Important untested logic

Files with significant testable missed lines. For each, list the file, coverage %, what the missed lines do, and whether they contain important logic (branches, calculations, parsing).

Priority 3 - Quality gaps

Tests exist but don't meet quality standards:

• Scorer tests that only check type (not CORRECT/INCORRECT logic)
• Missing error path tests for tools
• record_to_sample tests using fabricated data instead of real examples
• Missing or wrong pytest markers
• Missing E2E variant coverage
• Large or important gaps in Pytest coverage

Phase 4: Create or Fix Tests (if requested)

If the user asks to create or fix tests (not just review), proceed with the following.

Ask the user what they want:

• Create all missing tests
• Fix specific coverage gaps
• Only fix Priority 1 items (autolint failures)

4.1 Test file organization

tests/<eval_name>/
├── __init__.py                 # Always required
├── test_<eval_name>.py         # For simple evals (single file is fine)
├── test_e2e.py                 # For larger evals: E2E tests
├── test_scorer.py              # For larger evals: scorer tests
├── test_tools.py               # For larger evals: tool tests
└── test_dataset.py             # For larger evals: dataset/record_to_sample tests

Use a single file for simple evals. Split into multiple files for evals with 3+ testable component types.

4.2 Test patterns

For concrete test templates for each component type, see references/test-patterns.md. It covers:

• E2E tests — basic and parameterized variants
• Scorer tests — parametrized CORRECT/INCORRECT assertions
• Non-sandbox tool tests — happy path + error handling
• Sandbox tool tests — using the pytest_inspect_evals.sandbox shared utilities
• Dataset / record_to_sample tests — using real dataset examples
• HuggingFace dataset validation — using pytest_inspect_evals.huggingface assertions
• Solver tests — type-checks for thin wrappers (reviewers reject over-testing; PR #1009, #1008)
• Mocking get_model() — for evals that call get_model() at import time

4.3 Validation after creating tests

After creating tests, run them and verify:

# Run the new tests
uv run pytest tests/<eval_name>/

# Run linting
uv run ruff check tests/<eval_name>/
uv run ruff format tests/<eval_name>/

# Re-run autolint to verify all checks pass
uv run inspect-evals-lint <eval_name>

# Re-measure coverage
uv run pytest tests/<eval_name>/ --cov=src/inspect_evals/<eval_name> --cov-report=term-missing -q --runslow

Fix any failures before presenting results.

Phase 5: Present Results

Show the user a summary:

1. Component inventory: What testable components were found
2. Autolint results: Pass/fail for each check (from inspect-evals-lint output)
3. Coverage: Overall and per-file percentages, with missed line analysis
4. Quality gaps: Issues found in Phase 2.3
5. Actions taken: If tests were created/modified, list what was done
6. Remaining items: Any gaps that require manual attention or user decision

Things NOT to Do

• Don't re-check what autolint checks: Run autolint and report its results. Don't manually verify test existence, __init__.py, function name presence, etc.
• Don't over-test thin wrappers: isinstance(solver, Solver) (or isinstance(agent, Agent)) is sufficient for solvers that just wrap react() or assemble Inspect built-ins. Reviewers explicitly reject over-testing these (PR #1009, #1008).
• Don't treat non-trivial components as wrappers: Any solvers/scorers/tools with custom logic beyond wrapping Inspect built-ins do require testing for those portions. See the expanded scorer quality checklist in Phase 2.3 for specific guidance on sandbox-based, LLM-graded, dispatch, and reducer scorers.
• Don't use bare strings for TaskState model arg: Always use ModelName("mockllm/model"), never model="mockllm/model" — the latter causes mypy arg-type errors.
• Don't access score attributes without None checks: Always assert score is not None before score.value, and assert score.explanation is not None before "text" in score.explanation. Otherwise mypy reports union-attr errors.
• Don't add task API parameters for testability: Mock get_model() instead. This approach has been explicitly rejected in review (PR #998).
• Don't test Inspect framework internals by default: Test the eval's custom logic, not Inspect's built-in behaviour. The exception is an intentional characterisation/compatibility test; start its docstring with Characterisation: and state the library/framework contract being pinned.
• Don't fabricate dataset examples: Use real examples from the actual dataset for record_to_sample tests. This serves as both a test and documentation.
• Don't implement manual HF token checks: pytest-inspect-evals handles @pytest.mark.huggingface skip/retry logic. Do NOT add have_hf_token() or can_access_gated_dataset() functions (PR #987, #993).
• Don't guess pytest.mark.slow durations: Derive from actual CI run data (ceiling of observed max). If you can't measure, add a TODO for the user to fill in after running locally.
• Don't run full evaluations for testing: Only run unit tests. Full evals are slow and unnecessary for coverage checking.