agentsclimarketplace

Ccdawn performance engineering

Skill CCDawn/codex-skills/skills/engineering/ccdawn-performance-engineering

Chinese-first Codex skill pack for intent alignment, dynamic routing, lightweight development, code review, UI design, and AI research workflows.

Install
npx -y skills add CCDawn/codex-skills --skill ccdawn-performance-engineering

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when an explicit performance goal, measurable regression, unresolved hot path, large workload, concurrency/cache/streaming decision, or latency/throughput/CPU/memory/I/O/bundle budget needs profiling and evidence-backed optimization; do not use for routine development, speculative cleanup, or an already-located N+1/local inefficiency whose batch fix and deterministic count verification are clear to the current owner.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.9 KB, as published. Nobody here has run it

CCDawn 性能工程

目标

用最小测量找到真正限制结果的瓶颈,只保留收益超过噪声与复杂度成本的优化。普通开发不经过本 skill。

BRT interface

  • Context Boundary: 已确认的性能目标或强信号、代表性负载、owning surface、现有指标/benchmark/profile、正确性契约和允许写入范围。
  • Output Contract: baseline、主要瓶颈、最小优化、同负载对比、取舍、正确性验证和必要的回归保护。
  • Allowed Action: 读取相关代码和运行证据,执行有界 benchmark/profile,修改已授权 scope 并运行相关验证;不自动安装全局工具、不压测生产、不扩大产品行为。
  • Success Evidence: 相同负载下有可比较的 before/after,收益超过测量噪声或目标已满足,且正确性验证通过。
  • Stop Condition: 没有性能强信号、无法建立安全负载、瓶颈不在可改范围、需要高风险架构取舍、收益不足以抵偿复杂度,或目标已经满足。
  • Route Out: 当前开发 owner、ccdawn-bug-reviewccdawn-pr-reviewccdawn-development-cleanupccdawn-brt 或 BLOCKED。

统一调用契约

  • 只处理 BRT interface 范围;不匹配时回最具体 owner。用户可见内容默认中文,保留指标、命令、路径、API 和工具名。
  • Route Out 仅以 BRT interface 为准;末行写 下一步建议: <一个具体动作>
  • 不把性能工程变成固定开发阶段,也不要求用户逐步确认已授权的测量、优化和验证。

三档边界

BRT 使用 FAST / CHECK / PROFILE

  • FAST:低频、低规模、机械修改,直接开发和验证,不输出性能检查。
  • CHECK:当前 owner 静默处理已定位的 N+1、循环 I/O、重复全量计算、隐藏高阶复杂度和无界增长;批量修复及查询/调用次数断言仍是确定性验证,不加载本 skill、不建 benchmark,也不因“发现问题”切到 bug owner。只有存在已观察到的故障、正确性回归或待诊断根因时才路由 ccdawn-bug-review
  • PROFILE:明确性能目标/回归,或功能进入高频热路径、大数据量、并发、缓存、批处理、流式 I/O、持久化、包体/启动关键面时,才进入本 skill。

文件多、描述长、可能“以后会扩展”或一般性最佳实践不触发 PROFILE。PR/diff 审查仍由 ccdawn-pr-review 主责,本 skill 只在需要性能测量证据时作为 support。

最小测量循环

  1. 锁定一个主指标和代表性负载:延迟分位、吞吐、CPU、内存、I/O、查询/请求次数、包体或启动时间。没有目标阈值时,以当前回归或稳定 baseline 为比较点。
  2. 优先复用现有 benchmark、trace、日志、query plan 或 profiler;选择能区分方案的最小 probe。确定性调用次数、查询数或复杂度证据足够时,不制造负载测试。
  3. 找一个主要限制点,区分算法/重复工作、数据库/网络、分配/序列化、锁/队列、缓存、下游依赖和环境噪声;不做全仓性能扫描。
  4. 实施一个最小修改。优先消除无效工作和改进算法/数据访问,再考虑批处理、缓存或并发;微优化最后。缓存、线程池和并行不能凭直觉加入。
  5. 在相同负载下复测并运行正确性检查。记录方差或至少重复结果;延迟敏感链路优先看 p95/p99,不只看平均值。
  6. 收益达到目标或明显超过噪声且复杂度合理时保留;否则只撤销本轮自己的优化。只有测量稳定、回归代价高时才增加性能 guard,避免把脆弱时间阈值塞进普通单元测试。

目标达到即停止。进一步优化进入 Deferred,不自动扩展范围。

低噪声约束

  • 不默认生成性能报告、规划文档、TASK_GRAPH、worktree 或子 Agent。
  • 不因缺少专用工具阻塞;先用项目已有能力,安装依赖或长时间压测需要明确收益和权限。
  • 不用代码审美代替测量,也不把微小 benchmark 改善包装成用户价值。
  • 性能变化破坏行为、数据、安全或可维护性时,不接受该优化。

输出

性能结果:
- 指标与负载: ...
- Baseline: ...
- 瓶颈与最小修改: ...
- After 与取舍: ...
- 正确性验证/剩余风险: ...
下一步建议: <一个具体动作>

只有维护本 skill 或核对来源许可时读取 references/sources.md;执行性能任务不加载它。

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.