菜单
Run the per-module coverage gate for an AlphaGalerkin package the way CI enforces it. Use before opening a PR to confirm a changed package still meets its --cov-fail-under threshold (branch coverage), matching the gates in .github/workflows/ci.yml and the CLAUDE.md Regression Surface.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Audit a package's @abstractmethod and Protocol members for missing call sites — dead abstractions that every subclass implements but nothing invokes. Use before adding a new ABC/Protocol or reviewing a diff that touches one, to catch the F1-class defect (an abstract method spec'd in a docstring that never actually runs).
Standards for authoring C4 architecture documentation in Mermaid — Context, Container, and Component levels, dynamic views for key flows, and diagram hygiene rules. Use when creating or updating architecture docs — trigger phrases include "C4 diagram", "architecture doc", "document the architecture", "add a component diagram", "system context", or "sequence diagram for this flow".
Prescriptive Python house standards covering Protocol-based dependency injection, Pydantic v2 configuration schemas, asyncio discipline, structlog structured logging, and strict typing/lint gates. Use when writing or reviewing any Python module, class, config, or async code — trigger phrases include "new module", "add a config", "wire a dependency", "add logging", "make this async", "fix mypy errors", or "does this meet standards".
House testing standards covering the test pyramid, per-module coverage gates, Hypothesis property-based testing, subprocess contract tests for CLIs, fixtures, deterministic tests, and marker-gated GPU/external tests. Use when writing tests, adding a module, reviewing test coverage, or deciding what kind of test a change needs — trigger phrases include "add tests", "write a test for", "coverage gate", "is this tested", "test this CLI", or "flaky test".
Add a new PDE operator to AlphaGalerkin end-to-end. Use when introducing a PDE family (e.g. a new elliptic/parabolic/vector operator) so it registers in the operator registry, the PDEType enum, the canonical PDE_TYPE_MAP, and every dependent Literal enum — following the documented checklist so scenarios auto-pick it up.
Run the correct AlphaGalerkin regression-surface test command block for a changed code path. Use after editing solver/evaluator/PDE/scenario/agent/codec code to run exactly the guarding test suites the CLAUDE.md Regression Surface table prescribes, instead of guessing which tests to run.
| name | coverage-gate |
| description | Run the per-module coverage gate for an AlphaGalerkin package the way CI enforces it. Use before opening a PR to confirm a changed package still meets its --cov-fail-under threshold (branch coverage), matching the gates in .github/workflows/ci.yml and the CLAUDE.md Regression Surface. |
AlphaGalerkin gates coverage globally (85%) and per-module. This skill runs the correct gate for the package you changed.
.github/workflows/ci.ymlThere is no threshold table here on purpose: two copies of the same numbers drift (that
duplication is the mechanism by which src/pde/game.py's docstring became a lie). Read the gate
straight from CI:
# List every per-module coverage gate and its threshold, from ci.yml
grep -nE "cov=src/|cov-fail-under=" .github/workflows/ci.yml
Each Per-module coverage gate step in ci.yml pairs a --cov=src/<pkg> with its
--cov-fail-under=<N>. Scenario / integration packages (src/poc/scenarios/*,
src/integrations/*, src/agents/*) are gated at 85 branch via their dedicated ci.yml steps and
the CLAUDE.md Regression-Surface rows. A new package's gate is added to ci.yml in the same PR
as the package.
ci.yml (command above). For a new scenario or
agent, use 85 (branch coverage).pytest tests/<pkg>/ -m "not gpu_required" \
--cov=src/<pkg> --cov-branch --cov-fail-under=<N>
tests/poc/test_scaling_law_scenario.py).Certain nightly PyTorch CPU wheels crash on import torch while coverage's default C tracer
is active (SystemError: ... bad call flags), so pytest --cov collects no data. When that
happens, run coverage with the pure-Python tracer via the native runner (not the pytest-cov
plugin):
COVERAGE_CORE=pytrace python -m coverage run --branch \
--include="*/<pkg>/*.py" -m pytest tests/<pkg>/ -m "not gpu_required" -q -p no:cov
COVERAGE_CORE=pytrace python -m coverage report -m
CI uses a stable wheel where the C tracer works, so this fallback is only needed locally.