codeguard-cuda

CUDA Codeguard Gate

Goal: turn "looks fine" into repeatable tool evidence, and honestly distinguish pass, fail, unverifiable, and planned.

Start in 30 seconds

1. At the repository root, confirm 无专用标记文件 or the target file exists.
2. Detect tools:clang-tidy --version。
3. Read-only check first:bash -c 'find . -name '"'"'*.cu'"'"' | xargs -r clang-tidy --quiet'。
4. Classify results as: code violations, missing configuration, missing tools, tool internal errors.
5. Only perform on safely fixable items:clang-format -i '{file}'。
6. Rerun lint, and run the project's existing test/build gates.

Can be directly proposed:

• "Check this CUDA project, do not modify anything yet, give me the lint failure classification."
• "Only fix CUDA formatting issues, list semantic issues first."
• "Verify CUDA lint, test, and build evidence before commit."

Capability boundaries

✅ Good at handling

1. Identify the CUDA workspace based on .cu, .cuh and project markers.
2. Execute clang-tidy --quiet '{file}' and preserve exit codes, scope, and tool versions.
3. Distinguish formatting, static rules, configuration, dependency, and environment failures.
4. Use clang-format -i '{file}' to handle reversible issues and rescan.
5. Report local results and CI/build/test evidence in layers.

⚠️ Requires conditions

1. The target tools must be installed and their versions compatible with project constraints.
2. Must be executed from the correct repository or module root.
3. If the project has custom configuration, the committed configuration prevails.
4. Whether generated code, vendor, and cache directories are scanned must follow repository rules.

❌ Not applicable

1. Do not treat formatter passing as compilation, test, or security audit passing.
2. Do not silently install tools, upgrade dependencies, or modify team rules.
3. Do not use suppress/ignore/raising thresholds to hide real issues.
4. Do not claim "checks passed" when tools or configuration are missing.

When to use

• The user requests checking or fixing CUDA code standards.
• AI modified .cu, .cuh, requiring pre-commit verification.
• CI shows failures related to clang-tidy --quiet '{file}' .
• Need to determine whether an issue can be automatically fixed or must be handled manually.

When not to use

• The main task is learning language syntax, framework design, or business modeling; use the corresponding language professional skill instead.
• The main task is dependency vulnerabilities, security auditing, or license governance; use specialized security skills.
• When the user only asks to explain one error and there is no project context, first provide a local explanation and do not claim the repository is all green.

Data and security

This skill does not collect, upload, send, or store user code and credentials. By default, it only reads local repository configuration and runs local tools; it does not install dependencies over the network, and does not record tokens, passwords, or private source code. Before any automatic fix, check the diff first; destructive or large-scale changes must be explicitly confirmed by the user.

Execution Contract

Item	Current Fact
Status	stable(Since Codeguard V0.5)
File Extensions	.cu, .cuh
Project Marker	No dedicated marker file
Integration Configuration	Based on the repository's existing configuration
Tool Preparation	CUDA Toolkit + clangd
Detection Commands	使用命令查找与项目配置检查
lint	clang-tidy --quiet '{file}'
format/fix	clang-format -i '{file}'
Project-level gate	bash -c 'find . -name '"'"'*.cu'"'"' | xargs -r clang-tidy --quiet'

Standard Workflow

Step 1: Determine the Real Scope

Confirm the repository root, module roots, generated directories, vendor directories, and user-specified files. For multi-module projects, first list the modules to be scanned; do not guess based on the current directory.

Step 2: Read the Project Contract

Read based on the repository's existing configuration and the actual commands in CI. The repository configuration takes precedence over the examples in this skill; if the two conflict, report the differences first.

Step 3: Detect Tools and Versions

Run 对应工具 --version. Missing tools, incompatible versions, or inability to load configuration are all recorded as "unable to verify", not as a pass.

Step 4: Perform Read-only Checks

Run bash -c 'find . -name '"'"'*.cu'"'"' | xargs -r clang-tidy --quiet', recording the command, working directory, exit code, tool version, and the first actionable error.

Step 5: Failure Classification

Type	Typical Signal	Handling
Formatting	Indentation, whitespace, import order	Formatter allowed, then check the diff
Static Rules	unused, complexity, naming, dangerous APIs	Minimal semantic fixes; no bulk guessing
Configuration	parser/config/schema not found	Fix the integration or report prerequisites
Environment	command not found, version conflicts	Report installation/version requirements; do not install silently
Tool Internal Errors	crash, timeout, parser exception	Preserve the original evidence and narrow the reproduction scope

Step 6: Controlled Fixes

Preferentially run clang-format -i '{file}'. Only handle issues that are certain and reversible; dependency upgrades, rule exemptions, generated files, and business semantic changes must be explained separately.

Step 7: Verification Loop

Re-run the same lint command, then run the project's existing tests and build. Only write "pass" when the command, scope, and exit code are all clear; otherwise write "not run" or "unable to verify".

Rules

1. Re-verify with the same command: after fixing, you must re-run the original command that triggered the failure.
2. Layered evidence: report lint, format, compile, test, and security separately.
3. Minimal changes: do not refactor along the way, and do not expand style fixes into business rewrites.
4. Configuration first: do not override the project's existing ignore, dialect, target, or version constraints.
5. No fabrication: tools not run, configurations not seen, and platforms not verified must be clearly marked.

Output Template

CUDA Codeguard 结果
- 范围：<仓库/模块/文件>
- 工具：<名称与版本>
- 命令：<实际命令>
- 状态：PASS / FAIL / UNVERIFIED / PLANNED
- 发现：<按格式/规则/配置/环境分类>
- 修改：<文件与原因；无修改写 none>
- 复验：<命令、退出码>
- 未验证：<测试/构建/平台差异>

Gotchas

1. Missing tools are not a pass — must output UNVERIFIED and a clear preparation method.
2. formatter is not a linter — All-green formatting cannot override static rules or compilation errors.
3. Root directory determines results — Running in the wrong module may miss issues or load the wrong configuration.
4. Placeholders are not literal arguments — {file} must be replaced with real, authorized paths.
5. Auto-fixes may expand the diff — After fixing, review the diff first, then run the regression gate.
6. Generated and vendor directories require an explicit policy — Do not scan or modify third-party content across the whole repository at will.

Domain pitfalls and fixes

1. Domain pitfalls — CUDA files are constrained by both the host compiler and nvcc; clang-tidy cannot replace a real CUDA build.

When information is insufficient

Do not just say "please provide more information." First give a safe, read-only inspection plan and list the specific items that still need confirmation:

1. Repository/module root directory;
2. Target tools and versions;
3. Project configuration files;
4. The authoritative commands in CI;
5. Whether auto-fixes are allowed.

FAQ

Q1: The tool is not installed; can we conclude the code is fine? No. The status must be UNVERIFIED。

Q2: Can ignore or suppression entries be added automatically? No. Exemptions are allowed only when the user explicitly accepts them and the reason is documented.

Q3: Why does it still fail after formatting? The formatter only covers formatting; static rules, configuration, compilation, and tests still need to be handled separately.

Q4: Should generated code be scanned? By default, follow the repository configuration; when there are no rules, exclude generated/vendor first, then explain to the user.

Q5: How can we avoid changing too much at once? Fix in batches by failure category and file, and rerun the original command and review the diff after each batch.

Q6: If it passes locally, can we say CI will pass? No. You also need to align with CI's versions, environment variables, operating system, and command scope.

Load resources on demand

• When encountering exit codes, failure classifications, or exemption decisions, read references/rules/gate-rules.md。
• When you need to check this language's registry, toolchain, and project configuration, read references/tooling/toolchain.md。
• When you need copyable task examples, read by scenario examples/, and there is no need to load all of them at once.