Improving Python Coverage Agent Skill

Improving Python coverage

Increase Python unit test coverage by ~0.2% through meaningful tests that add real value.

Be fully autonomous — Do NOT stop or pause to ask for confirmation. Keep iterating (analyze → implement → verify) until the 0.2% coverage target is reached. If you encounter ambiguities about what to test, make a reasonable choice and proceed.

Workflow

Step 1: Run tests with coverage

make python-tests  # ~3 min, creates .coverage file

Generate JSON report for analysis:

uv run coverage json -o coverage.json

The JSON contains per-file missing_lines arrays showing uncovered line numbers.

Step 2: Analyze and prioritize

Read coverage.json to find files with:

• Large size + below-average percent_covered (high impact)
• Core modules in lib/streamlit/elements/ or lib/streamlit/runtime/
• Pure utility functions

Skip: >97% coverage, proto/*, vendor/*, static/*, test files.

Step 3: Implement tests (in subagent)

Launch a subagent to implement tests for each prioritized file. Provide the subagent with:

• The target file path and its missing_lines from coverage
• Instructions to read the source, existing tests, and write new tests
• The test selection guidelines below

The subagent should:

• Read source and existing tests at lib/tests/streamlit/<path>/<module>_test.py
• Write tests for: conditional branches, error handling, edge cases, exception paths
• Follow lib/tests/AGENTS.md: prefer pytest-style standalone functions over unittest.TestCase classes, use @pytest.mark.parametrize to consolidate tests that only differ in inputs/expected outputs, add numpydoc docstrings and type annotations
• Run the new tests to verify they pass: uv run pytest lib/tests/streamlit/path/to/module_test.py -v

Step 4: Verify and iterate

uv run pytest lib/tests/streamlit/path/to/module_test.py -v  # Run new tests
make python-tests                                             # Measure progress

Repeat steps 2-4 until coverage improves by ≥0.2%, then run make check.

Step 5: Simplify, review, and address feedback

Once all tests pass and coverage target is met:

• Run the simplifying-local-changes subagent to clean up and simplify the code changes. Wait for completion.
• Run the reviewing-local-changes subagent to review the changes. Wait for completion and read the review output.
• Address the review feedback: for each recommendation, implement it if valid and improves code quality; skip with brief reasoning if not applicable or would over-engineer.
• Run /checking-changes to verify everything still passes after changes.

Test selection

DO test: Conditional logic, error handling, edge cases (None, empty, zero, max), public API functions, complex branches.

DON'T test: Simple accessors, protobufs, implementation details, already well-covered code.

Coverage exclusions: Use # pragma: no cover sparingly for code that genuinely doesn't need testing. Always include a reason (e.g., # pragma: no cover - defensive):

• Platform-specific branches that can't run in CI (# pragma: no cover - platform-specific)
• Defensive code that should never execute (# pragma: no cover - defensive)
• Abstract method stubs or protocol definitions (# pragma: no cover - abstract)

Integration dependencies: Packages listed under [dependency-groups] integration in pyproject.toml (e.g., pydantic, sympy, polars, sqlalchemy) are only installed for integration tests, not regular unit tests. When writing tests that use these packages:

• Import them inside the test function, not at module top-level
• Add @pytest.mark.require_integration marker to the test
• This ensures tests gracefully skip when run outside the integration test environment

Test file location

lib/tests/streamlit/<package>/<module>_test.py mirrors lib/streamlit/<package>/<module>.py

Notes

• Quality > coverage numbers - skip tests that don't catch real bugs
• Target is 95%+ coverage per lib/tests/AGENTS.md
• Use /checking-changes after implementing tests
• Some code paths involving external libraries (e.g., database connectors, optional dependencies) are already covered by integration tests marked with pytest.mark.require_integration. These integration tests are not included in the coverage numbers from make python-tests. When analyzing missing lines, check whether the uncovered code is exercised by integration tests before adding unit tests or # pragma: no cover annotations.
• Local vs CI coverage differences: Code that is version-specific (Python version, library version) or uses integration dependencies may appear uncovered locally but is tested and covered in CI. Examples:
• Python version-specific branches (e.g., if sys.version_info >= (3, 14)) run only on matching CI jobs
• Library version-specific code (e.g., pandas 2.x vs 3.x behavior) is covered across CI matrix
• Integration dependency tests (@pytest.mark.require_integration) run in separate CI jobs with those packages installed

Before adding tests or # pragma: no cover for such code, verify whether it's already exercised in CI.