Skills DirectorySkills Directory
SkillsLearnSecurityCategoriesDocsCommunityBlog
Sign InSubmit Skill
Skills Directory

Security-tested agent skills for Claude, coding agents, and AI workflows.

Directory

  • Browse Skills
  • All Skills A–Z
  • Claude Skills
  • Claude Code Skills
  • Agent Skills
  • Categories
  • Authors
  • Submit a Skill

Learn

  • Learn Hub
  • Install Claude Skills
  • Write SKILL.md
  • Skills vs MCP
  • Directories Compared

Security

  • Security
  • Methodology
  • Secure Claude Skills
  • Security Badges

Company

  • About
  • Community
  • Blog
  • API Docs
  • Advertise

2026 Skills Directory. All rights reserved.

ProTermsPrivacyRefunds
Back to skills

Msagent Profiler Breakdown

ASecurity

DB格式Profiling数据场景化性能拆解:自动完成场景确认(框架/任务类型判定)、阶段拆解与单次执行边界拆解,并给出报告。Use when 需要对昇腾 Profiling DB 做阶段拆解、step 归因或生成性能报告, or when the user says "帮我按任务阶段拆解这个Profiling数据", "多层次性能耗时拆解", "分析这个 vllm/sglang/verl 数据,主要耗时都花费在哪里了"。

31 stars
0 votes
0 copies
0 views
Added 9/23/2026
datapythongosqlapi

Works with

cliapi

Security Analysis

A100/100

Scanned 9/23/2026

Install to Claude Code

$npx -y skills add kali20gakki/msAgent --skill msagent-profiler-breakdown --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Msagent Profiler Breakdown?

Add the live security badge to your README — it updates automatically with every re-scan.

Security grade badge for Msagent Profiler Breakdown
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/kali20gakki-msagent-profiler-breakdown/badge)](https://www.skillsdirectory.com/skills/kali20gakki-msagent-profiler-breakdown)

More formats (shields.io, HTML) on the badges page.

Download with Pro
Files
SKILL.md
---
name: msagent-profiler-breakdown
description: DB格式Profiling数据场景化性能拆解:自动完成场景确认(框架/任务类型判定)、阶段拆解与单次执行边界拆解,并给出报告。Use when 需要对昇腾 Profiling DB 做阶段拆解、step 归因或生成性能报告, or when the user says "帮我按任务阶段拆解这个Profiling数据", "多层次性能耗时拆解", "分析这个 vllm/sglang/verl 数据,主要耗时都花费在哪里了"。
metadata:
  author: MindStudio Agent Team
  category: profiler
  version: 0.1.0
---

# msagent-profiler-breakdown —— 场景化拆解(一步到位)

> 定位:合并原 stage-decompose / step-decompose / breakdown-report 三份 skill,一份搞定
> **阶段归因(level=阶段)→ 单次执行边界(level=单次执行)→ 报告**。
> 场景差异(推理 / 训练 / RL,vllm / sglang / verl / slime…)下沉到场景资源文件
> `resources/scenarios/<scenario>/<framework>.json`,主 skill 不做框架分支。
> 用户只与本文档交互;`scripts/decompose.py` 一步产出 breakdown.json(统一拆解树)/ HTML 报告 / findings。

## 触发与元信息

| 维度    | 内容                                                                                                                                                          |
| ----- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 触发关键词 | 模型按层拆解、完整性能分析、一键分析、端到端到层内拆解、框架/并行自动识别、多卡 TP/EP/PP 分析                                                                                                        |
| 外部依赖  | 昇腾 Profiling DB(sqlite)与 `analysis.db`;Python 3 标准库;场景资源 `resources/scenarios/`;拆解树契约 `references/breakdown-schema.md`;报告模板 `references/report-template.md` |
| 权限要求  | 只读打开 Profiling DB(`mode=ro`,保留 WAL 恢复);在工作区读写中间产物与报告                                                                                                        |
| 预期输出  | breakdown.json(统一拆解树:端到端 → 阶段 + 单次执行分支)+ stage.json(同内容别名,场景信息供 layer-decompose 继承)+ HTML 报告 + findings(8 字段)+ 明细 CSV                                       |

## 一句话调用

```
python scripts/decompose.py --db <ascend_pytorch_profiler_{rank_id}.db> \
    [--framework vllm|sglang|verl|slime|megatron|mindspeed|fsdp] \
    [--task-type inference|train|rl] [--anchor "forward"] [--inspect-index 0]
```

**最少必给只有** **`--db`**;framework / task-type 可省,走自动探测 / 兜底默认 / 由编排层询问用户。

## 三步流程

### 第 1 步:场景确认(framework / task-type 判定)

判定优先级(framework 与 task-type 同构):

1. **CLI / 编排层显式指定**(`--framework` / `--task-type`)→ high。用户可直接说
   「请帮我分析下这个 vllm 的 profiling 数据」,此时 framework 由上层输入。
2. **自动探测**(从 profiling 数据分析):`STRING_IDS.value` 含 `vllm%` 前缀 → vllm;
   存在 verl worker 侧 MSTX 打点名(`actor_update` / `actor_compute_log_prob` /
   `ref_compute_log_prob` / `train_batch`)→ verl。框架映射任务类型:
   vllm/sglang→inference、megatron/mindspeed/fsdp→train、verl/slime→rl。→ high
3. **识别不了 → 询问用户**(编排层用 AskUserQuestion 回显候选 + 依据让用户确认/修正);
   仍缺则默认 vllm/inference + 显式告警(`task_type_source=default`),报告低置信度清单列出。

探测工具:`scripts/db_query.py`(子命令 `string` / `mstx` / `pytorch` / `cann` / `api`),
先查打点再定规则。判定结果(`framework` / `framework_source` / `task_type` /
`task_type_source`)写入 `stage.json`,**供下游 layer-decompose 经** **`--framework`** **继承,
无需重新探测**。

### 第 2 步:执行拆解(按场景资源规则)

引擎 `decompose.py` 框架无关,按场景资源文件声明式规则执行,主 skill 不过多呈现框架分支:

- 阶段归因:读取 `resources/scenarios/<scenario>/<framework>.json` 的
  `stage_decompose.stages`(name + scopes + backing),直连 DB 统计每个 stage 的
  count / 累计 / 均值 / 最大 / 最小 / 首个。

- 单次执行边界:读取 `step_decompose.anchor`(默认 `forward`),锚点 scope 每次出现
  记一个 step,按位置(前 2 步 graph\_capture/compile)+ 阈值(wall\_time > factor×median)
  判热身并写 `warmup_reason`;`--skip-first-step` 另标首个 step 为疑似 prefill;
  剔除后算稳态统计(count/mean/P50/P99/min/max/variance/CV)。

- 锚点优先级:CLI `--anchor` 显式覆盖 > 场景资源 `step_decompose.anchor`;锚点为空
  (如 RL/训练待核验)→ 跳过 step 拆解并告警。

新增框架 = 在场景目录放一个 `<framework>.json`(支持 `extends` 继承父规则),不改引擎,
详见 `references/frameworks/README.md`。

### 第 3 步:生成报告(模板即资源,无独立 report skill)

按 `references/report-template.md` 组织结论先行 HTML(自包含、无 JS 依赖):
一句话结论 → 逐级耗时(仅列出现过的级别)→ 阶段统计表 → TOP-N 下钻链 → per-step 明细 →
稳态统计 → 异常/漂移 → 低置信度清单 → 建议下一步 → findings(8 字段)。
场景信息(框架/任务类型,用户已知)与守恒自检不做独立章节,守恒缺口告警保留在 findings 中。
同时落盘 findings JSON 与明细 CSV;多 run A/B 对比用 `scripts/compare_runs.py`。

## 输入

| 输入              | 必填 | 来源                                                          |
| --------------- | -- | ----------------------------------------------------------- |
| 昇腾 Profiling DB | 是  | 用户给定路径(`ascend_pytorch_profiler_{rank}.db` / `analysis.db`) |
| framework       | 否  | 场景确认(CLI > 自动探测 > 询问 > 默认 vllm + 告警)                        |
| task-type       | 否  | 场景确认(CLI > 框架映射 > 默认 inference + 告警)                        |
| anchor          | 否  | 覆盖 step 边界锚点(默认取场景资源 `step_decompose.anchor`)               |

## 输出(workspace 落盘)

| 产物                                             | 内容                                                                          |
| ---------------------------------------------- | --------------------------------------------------------------------------- |
| `breakdown.json`                               | 统一拆解树:端到端(会话)→ children(阶段... + 单次执行分支含 per-step 序列与稳态统计)+ 框架/任务类型来源 + 守恒自检 |
| `stage.json`                                   | 与 breakdown.json 同内容别名;场景信息供 layer-decompose `--framework` 继承               |
| `<prefix>.html`                                | 结论先行报告(模板见 `references/report-template.md`)                                 |
| `<prefix>_findings.json`                       | 结构化结论(8 字段:问题/证据/影响/根因/优化动作/预期收益/验证路径/置信度)                                  |
| `<prefix>_detail.csv` / `<prefix>_handoff.csv` | per-step 明细 / 层拆解交接清单(暂缓期间为占位)                                              |

## 涉及的 skill

| skill                      | 职责                         | 状态                       |
| -------------------------- | -------------------------- | ------------------------ |
| msagent-profiler-breakdown | 场景确认 + 阶段/单次执行拆解 + 报告,一步到位 | 在链路内                     |
| layer-decompose            | 单层归属 + 层内归因 + 结构校验(直连 DB)  | 暂缓保留(经 --framework 继承场景) |

Attribution

kali20gakkikali20gakki
View sourceMore from kali20gakki →
SSkills DirectorySkills Directory

Ship a skill? Prove it's safe.

Free 120-pattern security scan, letter grade, and an embeddable README badge.

Submit a skill

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments (0)

No comments yet. Be the first to comment!

SSkills DirectorySkills Directory

Ship a skill? Prove it's safe.

Free 120-pattern security scan, letter grade, and an embeddable README badge.

Submit a skill

Related Skills

Rank Tracker

This skill helps you track, analyze, and report on keyword ranking positions over time. It monitors both traditional SERP rankings and AI/GEO visibility to provide comprehensive search performance insights.

1821 votes

Youtube Competitor Analyzer

Find and analyze YouTube competitor channels using YouTube Data API v3. Discover competitors through keyword search, category matching, content similarity, and related channel discovery. Compare metrics, content strategies, and market positioning. Use when users want to (1) Find competitors for their YouTube channel, (2) Analyze competitor performance metrics, (3) Compare their channel against competitors, (4) Identify content gaps and opportunities, (5) Benchmark against similar creators, (6...

31 votes

Twitter Algorithm Optimizer

Analyze and optimize tweets for maximum reach using Twitter's open-source algorithm insights. Rewrite and edit user tweets to improve engagement and visibility based on how the recommendation system ranks content.

742580 votes

Weather Fetcher

Instructions for fetching current weather temperature data for Karachi, Pakistan from wttr.in API

662880 votes

Weather

Get current weather and forecasts (no API key required).

484900 votes
View all in data →