在之前发布的SKILL，我们已经能把接口测试的 AI 流水线跑通了，但一直留着一个缺口：登录、加购、下单各自都测得很好，拼到一起却不一定通——单接口全绿，业务照样可能卡死。
这次新增的两个 Skill 就是补这块的：把「注册→登录→搜索→加购→下单→支付→查单」这样一条完整业务链路，当成一个测试对象来搞。
🔹 api-scenario-modeler（场景建模） 🔹 api-scenario-runner（链路执行引擎）
两个 Skill 一个管「测什么」，一个管「怎么跑」，和前面篇章的执行、诊断、清理、报告技能无缝衔接，单接口和业务场景两条线从此在同一个体系里跑。
上一篇，我们已经介绍了api-scenario-modeler（场景建模），今天这篇来来带大家拆解一下api-scenario-runner（链路执行引擎）这个Skill。
实战项目准备
本篇教程的实战案例，依旧延续 shop-lab 全栈电商项目，这次业务场景链路skill，会横跨项目的 5 个业务模块：
先确保后端服务已正常启动，Swagger UI 可访问：http://localhost:8080/swagger-ui.html，
业务场景接口测试执行引擎：api-scenario-runner
api-scenario-runner这个Skill的作用，就是把 api-scenario-modeler（或人工）产出的 scenario_xxx.yaml 编排文件真实执行起来：按序发请求、自动变量提取与注入、鉴权头注入、三级断言（步骤断言 + db 终态断言）、每步执行快照留存、失败即停保留现场，还支持把编排文件编译为 pytest 或任意指定框架链路脚本工程。
适用场景
- 执行场景编排文件、跑链路、端到端流程验证
- 把场景 yaml 编译成 pytest 链路脚本，纳入 api-test-executor 统一调度
- 失败后保留完整现场，交给 api-failure-diagnoser 深度诊断
使用方式
在 Claude Code 中提供编排文件路径，或输入：
/api-scenario-runner
两种执行模式
核心能力
- 场景上下文：${var} 自动提取与注入（支持整串替换与拼接），auth: true 步骤自动注入 Bearer token
- 三级断言：步骤断言（status_code → 业务码 → jsonpath）→ 终态 db 断言（expect_value 直接比对 / expect_expr 变量表达式求值）
- 失败即停 + 现场留存：每步快照写入 steps_snapshot.json（注入后真实请求、响应、变量视图，敏感值脱敏）
- 归因初判五类：VARIABLE_MISSING / ASSERTION_FAILED / API_ERROR / CONNECTION_ERROR / DB_ERROR
- 安全机制：db 断言仅允许 SELECT、SQL 参数化执行、编排文件禁止写死连接信息、配置缺失 fail-fast
输出： scenario_report.md（执行报告）+ steps_snapshot.json（诊断快照）+ compile 产物（pytest 工程）
实操演示
输入下方提示词，如图所示：
skill运行期间，会自行对比与修正各接口字段名称
并对MySQL数据库、redis缓存数据一致性进行校验。
执行完成，SC-order-001 全链路通过（全链路共包括了14个接口调用步骤）
同时，scenario_shopping_flow.yaml和modeling_notes.md这两份场景资产已同步为实测最新版：
最后，还会自动生成一份md格式的场景测试执行报告 （默认会存放在：接口测试项目目录/shop-lab-api-test/scenario_run_results/xxx/scenario_report.md）
融入既有体系（compile + 流水线）：
/api-scenario-runner 编译场景编排文件 /Users/xxx/scenarios/scenario_order_flow.yaml 为 pytest 脚本，输出到 shop-lab-api-test 项目
编译产物检查两点：目录结构符合 007 篇的分层规范（testcases/ + data/ ），如上图所示，在testcases/目录下，生成了包含 pytest 链路脚本：test_scenario_order_001.py
在data/目录下，生成业务场景关联脚本测试数据scenario_order_001.yaml
之后的一切都是熟悉的：用 009 篇的 api-test-executor 按标签（scene:business-flow）调度执行，或交给 api-pipeline-scheduler 纳入"执行 → 清理 → 报告"流水线。
比如输入如下提示词：
调用api-test-executor技能，跑一下~/ai_project/shop-lab-api-test/testcases/test_scenario_order_001.py这个业务场景测试脚本
执行后，链路 14 步 + 终态 DB 校验 4 条全部通过，期间还顺手修复了一个编译器的Bug。
执行完成后，顺带调用api-report-generator技能，让其生成测试报告
会自动生成好两份测试报告，一份是基于Allure 原生报告（在线服务）
另一份，是一份自研定制的HTML 测试报告（单文件，可独立打开）
幂等与重复执行
第四轮解决"昨天能跑今天跑不过"的问题：不做任何修改，紧接着再执行一遍同一条链路。
合格标准非常明确：第二遍同样全绿。因为 username 是 timestamp 策略动态生成的，第二遍注册的是新用户，不存在 009 篇 newuser_001 那种重复注册冲突；链路数据回收按 cleanup 配置联动 api-testdata-cleaner，两遍执行之间环境是干净的。如果第二遍失败，大概率暴露两类问题——唯一性字段漏了动态策略，或 cleanup 配置漏了某个数据对象，回改编排文件即可。
这一轮直接决定链路能不能进 CI 流水线天天跑，是五轮里最不能省的一轮。
至此，业务场景测试完全并入前面篇章建立的自动化体系，单接口与业务场景两条线，在同一个体系里跑了起来。
至此，接口自动化测试的 Agent Skill 体系补上了最后一块拼图，完整版图如下：
如何获取Skill
所有实战演练创建生成的 skill，会统一上传到 GitHub中，可按需自取。
同时，目前所有 Skill 均已统一上传至网站 testfather.cn Skill 技能商店中。
本篇新增的两个 Skill，安装方式与之前一致，将 Skill 目录复制或克隆到你的 Claude Code 或 WorkBuddy 技能目录下即可使用：
## 在线安装
npx skills add xxx/api-scenario-modeler
npx skills add xxx/api-scenario-runner
# 克隆仓库
git clone git@github.com:xxx/skills.git
# 复制到 Claude Code 技能目录
cp -r skills/api-scenario-modeler ~/.claude/skills/
cp -r skills/api-scenario-runner ~/.claude/skills/
# 或复制到 WorkBuddy 技能目录
cp -r skills/api-scenario-modeler /Users/xxx/.workbuddy/skills/
cp -r skills/api-scenario-runner /Users/xxx/.workbuddy/skills/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The SKILL of the previous publish, we could have run through the AI water line that had been tested for the interface, but we kept a gap: login, additions, orders, each of them, were well measured, but one did not necessarily fit -- single interfaces were green and business could be stuck.
The two new Skills this time are filled with this piece: "Registering log-in search and purchase orders to pay."
api-scenario-moder (scene modelling) api-scenario-runner (link implementation engine)
Two Skills, one on "what to measure," one on "how to run" and one on "how to run" are seamlessly connected to the implementation, diagnosis, clean-up, reporting skills of the previous chapters, and the single interface and business scene run from now on in the same system.
In the last one, we've already introduced api-scenario-moder, and today it's here to take you down this Skyll.
Operational project readiness
The field case of this course continues with the "shop-lab All-Trending Project ", a business scene-skill that cuts across the five business modules of the project:
To ensure that back-end services are properly launched, Swagger UI can be accessed at http://localhost:808080/swagger-ui.html.
Operational scene interface test implementation engine: api-scenario-runner
The role of Skill, api-scenario-runner, is to authenticate the scenario xx.yaml presentation of api-scenario-moder (or manual) output: sequential requests, automatic variable extraction and injection, brain injections, three-tier assertions (step assertion + db final assertion), the retention of a snapshot every step, the closure of the site if the site fails, and to support the translation of the organization into pytest or the arbitrary designation of a framework link script.
Apply scene
- Execute scene-formulation files, run links, end-to-end process validation
- Translation of scene Yaml into pytest link scripts for integration of api-test-executor
- Keep full site after failure and submit to api-failure-diagnoser for in-depth diagnosis
Usage
Provides a layout file path in Claude Code, or enter:
/api-scenario-runner
Two modes of implementation
Core competencies
- scene context: ${var} Auto-extract and inject (support full-string replacement and spell), auth: true step automatically injects Bearer token
- Level three assertions: step assertion (status code →jsonpath) > final db assertion (expect value directly against / extect expr expression)
- Failed to stop + on-site retention: steps snapshot.json per step
- under five categories: VARIABLE MISSING / ASSERTION FAILED / API ERRO / CONNECTION ERRO / DB ERRO
- Security mechanism: db asserts that only SELECT, SQL parameterization is allowed, the file is banned from writing dead connection information, the configuration is missing
Output: scenario report.md (implementation report)+steps snapshot.json (diagnostic snapshot)+compile product (pytest Project)
Practical demonstration
Enter the following prompt, as shown:
During skill's operation, you will compare the names of the interfaces with the names of field.
The MySQL database, the redis cache data consistency was also verified.
Implementation completed, full-link passage of SC-order-001 (total 14 interface call steps included in the chain)
Meanwhile, two scenario assets, senario shopping flow.yaml and modeling notes.md, have been synchronized into the latest reality:
Finally, a scenario test implementation report in md format will be automatically generated (the default will be stored in the interface test project directory/shop-lab-api-test/scenario run results/xx/scenario report.md)
Integration of existing systems (compile + streaming lines):
/api-scenario-runner compiles scenario layout files /Users/xx/scenarios/scenario order low.yaml for pytest scripts, output to shop-lab-api-test
Compilers check two points: the directory structure corresponds to the 007 stratification (testcases/ + data/) and, as shown in the above graph, under the testcases/ directory, a pytest link script was produced: test scenario order 001.py
Generate business scenario correlation script testing data ssenario order 001.yaml under data/directory
After that, everything was familiar: api-test-executor, by scene:business-flow, or hand over to api-pipeline-scheduler for "executing clean-up" reporting.
For example, enter the following prompt:
Call api-test-executor skills and run the script ~/ai project/shop-lab-api-test/testcases/test scenario order 001.py
After implementation, the 14-step link + final DB verification 4 passed, during which a compiler's Bug was repaired.
Upon completion of implementation, call the api-report-generator skills to generate test reports
It automatically produces two test reports, one based on Allure's original report (online service).
The other is a self-developed HTML test report (single file available independently).
幂 etc. and duplicate execution
Round four to solve the problem of being able to run today: Without any modification, the same link is implemented again immediately.
The eligibility criteria are clear: the same green for the second time. Because username is the dynamic of the timestamp policy, the second registration is for new users, and there is no double registration conflict of 009 newuser 001; api-testdata-cleaner is connected by clairup for chain data recovery, and the environment between the two is clean. If the second one fails, it probably exposes two types of problems - single field missing a dynamic strategy, or the cleanup configuration missing a data object, reprogramming the file.
This round directly determines whether the chain can run into the C.I. flow line, which is the most inescapable round of five.
As a result, the business scenario test was fully integrated into the automated system established in the previous chapter, with a single interface and business scenario line running in the same system.
As a result, the Agent Skill system, which has been tested for interface automation, has been completed with the last piece of puzzle, which is complete as follows:
How to get Skill
All active exercises create skills, which are uploaded centrally to GitHub and are available on demand.
At the same time, all Skills are now centrally uploaded on the website testfather.cn Skill.
Two new Skills, installed in the same manner as before, are copied or cloned into your Claude Code or WorkBuddy skills directory:
# Online installation
npx skills add xxx/api-scenario-modeler
npx skills add xxx/api-scenario-runner
# Cloning warehouse
git clone git@github.com:xxx/skills.git
# Copy to Claude Code Skills Directory
cp -r skills/api-scenario-modeler ~/.claude/skills/
cp -r skills/api-scenario-runner ~/.claude/skills/
# Or copy to WorkBuddy Skills Directory
cp -r skills/api-scenario-modeler /Users/xxx/.workbuddy/skills/
cp -r skills/api-scenario-runner /Users/xxx/.workbuddy/skills/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.