Refactor PyTorch test files to decouple tests from specific hardware accelerators using three strategies: accelerator-unrelated (CPU-only standalone classes with instantiate_parametrized_tests), accelerator-agnostic (device-generic classes with instantiate_device_type_tests), and accelerator-specific (single-accelerator standalone classes). Pay special attention to tests tagged @onlyCUDA or using .cuda()/device="cuda" that are NOT actually CUDA-specific — most should be refactored to accelerator-agnostic. Use when asked to decouple, refactor, or reorganize tests to work across multiple accelerators, or when a test file imports TEST_CUDA/TEST_MPS/TEST_XPU but most tests don't require a specific device.
Review PyTorch test refactoring for correctness and completeness against the decoupling standards defined in the refactor-test-decoupling skill. Accepts a test file path (whole-file review), a PR URL, a git diff, or a branch name. Use this when asked to review a test refactoring PR, check a test decoupling change, verify a refactored test file, review a single test file for classification correctness, or when the user mentions "review" in the context of test splitting, test decoupling, test refactoring, or accelerator-agnostic test migration. Also use when the user opens a PR, diff, or Python test file and asks for a quality check.
Orchestrate PyTorch test file refactoring to decouple tests from specific hardware accelerators. Use this skill whenever the user asks to refactor, decouple, or reorganize a PyTorch test file for cross-accelerator compatibility, or when they ask to apply the test decoupling workflow to a specific test file. Triggers on phrases like "refactor test_ops.py", "decouple tests in test_file.py", "apply test decoupling to X", or "use pytorch-test-refactoring on X". Also triggers on CI monitoring phrases like "look after the ci", "monitor ci", "check ci status", "watch the pr checks", or "debug ci failures". Always invoke this skill before starting any test refactoring work that involves splitting tests into accelerator-unrelated, accelerator-agnostic, and accelerator-specific classes.
Runs an adversarial review loop where a challenger agent critically reviews an implementer's output, driving iterative improvement until the challenger issues a PASS verdict. Never triggered automatically; must be explicitly called by the user (2 persistent agents, up to 10 rounds).
Use when the pytorch-test-refactoring orchestrator emits "schedule_cron" or "need_agent" with phase "ci_debug", when a CronCreate cron job fires and you need to run orchestrator.py --ci-check, or when the user says "look after the ci", "monitor ci", "check ci status", "watch the pr", or "debug ci failures".
Use when classifying PyTorch test files as device_unrelated, device_agnostic, or device_specific, or when asked to categorize/scan test files by device dependency
Use gh (GitHub CLI) to fetch and process GitHub issues, pull requests, repositories, and other GitHub resources. Use when the user provides a GitHub URL (https://github.com/*), asks about GitHub issues/PRs, or mentions gh commands.