Skip to content

AGENTS.md

Latest commit

 

History

History
51 lines (44 loc) · 2.72 KB

AGENTS.md

File metadata and controls

51 lines (44 loc) · 2.72 KB
Raw

Repository instructions

Read .github/CONTRIBUTING.md, docs/concepts/design.md, and docs/ARCHITECTURE.md before making a non-trivial change. Preserve unrelated work in a dirty worktree and keep edits inside the requested scope.

Test policy

The repository test suite is intentionally layered because some end-to-end checks create complete workspaces and can take minutes. Do not run every test after every edit.

  1. Focused checks (during development): run the smallest relevant test node or file for the code being changed. Repeat this layer while iterating.
  2. Default suite (before handoff): uv run --frozen pytest -q runs the normal unit, behavior, and lightweight integration tests. Tests marked slow are skipped by default.
  3. Slow integration checks (only when related): run a marked test with uv run --frozen pytest -q --run-slow <test-path-or-node>. Use this layer only when the change touches the workflow that test exercises, when a release gate explicitly requires it, or when investigating a failure in that workflow. Run all slow tests only for broad cross-cutting changes.
  4. External/live checks (manual): tests that require Docker, Harbor jobs, model calls, credentials, network services, or a real evolution campaign are never routine validation. Run them only when the task explicitly needs that integration and report the environment and artifacts separately.

Choose tests by behavior, not just by filename. Typical mappings are:

  • CLI or one source module: its direct test file or exact test nodes.
  • Recipe, scaffold, or seed changes: the recipe-specific tests plus tests/test_recipe_composition.py and/or tests/test_resource_layout.py as applicable.
  • Architecture/module inventory changes: the relevant tests/test_coherence.py nodes.
  • Legacy recipe/operator/driver changes: tests/test_recipe_full_cycle.py with --run-slow on Linux (requires libseccomp and cached uv dependencies/Python). It runs nine original recipes through one generation with real operators and evaluator scripts, a deterministic Harbor substitute, and blocked network access.
  • gepa_local local-Harbor generation changes: the focused recipe tests and tests/test_gepa_local_recipe.py::test_gepa_local_full_generation_improves_champion with --run-slow.
  • Packaging/release changes: build the distribution and run tests/test_release_artifact.py with EVOLVE_RELEASE_DIST set, following the release workflow.

If a new test routinely takes tens of seconds, launches a complete evolution, or depends on expensive setup, mark it @pytest.mark.slow and document the feature/path that should trigger it. A slow test must stay deterministic and must not silently depend on live external services.