测试工程师 大肥鱼 · Test Engineer Big Fat Fish

面向 DeepSeek Harness 的黑盒功能测试插件：接口、页面、前后端一致性，用对话驱动。

中文 · English

---

中文说明

它是什么

dsh-plugin-big-fat-fish 是一个 DeepSeek Harness（DSH）Host 组合包：注册 3 个执行工具 + 1 个方法论 skill。

类型	名称	作用
工具	ft_http_run	批量发 HTTP 请求（套件或内联用例），断言 status / jsonPath / headers / latencyMs，提取变量，落盘证据
工具	ft_ui_run	用 playwright-core 驱动本机 Chrome/Edge：页面流程、选择器、等待、断言、截图、收集控制台报错与失败请求
工具	ft_report	从已落盘的 run 重新生成人读报告（单个或汇总，支持标记已知失败）
Skill	functional-testing	方法论：需求 → 接口/UI 入口摸底 → 用例设计 → 数据确认 → 执行 → 归因 → 每轮档案与报告

Skill 负责「测什么、期望是什么、失败算谁的」——是产品缺陷、用例缺陷还是环境问题；工具只负责执行和留证据，不猜。

它擅长：

• 通过真实接口、真实页面做版本/模块验收与回归（不 mock、不改项目源码）；
• 每项目复用登录档案：登录接口探明一次后，之后每轮默认「登录 + 本轮功能」联合测试；
• 每轮留下可复现的档案：目标 / 覆盖范围 / 测试数据 / 异常复现方式；
• 前后端一致性：同一份数据经接口与页面分别读取，逐字段比对。

安装

需要带插件管理器的 DSH（Host bundle 支持）与 Node 22+；UI 测试还需要本机装有 Chrome 或 Edge。

从 GitHub 安装（建议固定 tag 或 commit）：

dsh plugin add github:Utaoki-henji/dsh-plugin-big-fat-fish#v0.1.1

桌面端：打开 设置 → 插件 → 安装，填上面的 spec（或 https://github.com/Utaoki-henji/dsh-plugin-big-fat-fish），然后重启 Harness。

离线 / 对话安装：拿到发布包 dsh-plugin-big-fat-fish-v<version>.zip，让助手安装即可——发布包自带 node_modules/playwright-core，安装时不需要联网。

安装后插件显示为 测试工程师 大肥鱼，并带 functional-testing skill。

用法

在对话里说「开始测试」/「开始测试接口」/「开始测试页面」，或英文 "start testing"。

skill 会按顺序确认：项目 → 测试类型（接口 / 页面 / 模块）→ 入口地址或路由 → 登录档案（需要令牌时不猜路径）→ 只读代码得出契约与分支清单 → 多接口执行逻辑（顺序/分支/并行 + 具体顺序）→ 测试数据确认卡 → 先冒烟再全量 → 每轮档案与报告。

工作区状态

所有状态都在当前工作区的 .ft/ 下（不写你的项目源码）：

.ft/
  config.json              # 接口/UI 基址、允许的主机、写操作放行、响应体上限
  credentials.local.json   # 本地凭据库（被 gitignore，环境变量优先）
  logins/<project>.json    # 登录档案：接口、令牌路径、鉴权头、凭据环境变量名
  suites/                  # 可复用测试套件（凭据保留 ${env:...} 占位符）
  runs/<runId>/            # results.jsonl、证据、截图、报告

ensureWorkspaceState 会写入 .ft/.gitignore，避免 run、凭据、鉴权产物进入你的仓库。

环境要求

• 带插件管理器的 DSH（Host bundle 支持）；
• Node.js 22+（ESM）；
• ft_ui_run 需要本机 Chrome 或 Edge（直接驱动已装浏览器，不下载浏览器内核）；
• 能访问你要测的目标地址。

安全与隐私

• 对项目只读：不改源码、不插桩、不跑你的构建；
• 只写工作区 .ft/；
• 凭据来自环境变量或 .ft/credentials.local.json，日志/证据/套件按敏感键脱敏，落盘保留 ${env:...} 占位符；
• 无遥测：只访问你给的目标地址；
• 写操作需显式授权：非 GET 用例必须标 destructive: true 且被 .ft/config.json 放行；
• UI 自动化使用独立浏览器上下文，不碰你日常浏览器的 profile。

边界

• 执行层只有黑盒：HTTP 请求 + 真实浏览器 UI；没有单测执行、没有覆盖率、没有断点插桩；
• 白盒只读静态：第 5 步允许只读 Java/Spring Controller、DTO、@Valid、Service 分支、Vue 路由与 API 层来推导期望，不执行任何代码；
• concurrency 固定为 1，并行请求按确认顺序串行；
• ft_ui_run 使用本机 Chrome/Edge，支持无头模式与登录态注入（storageState / cookies）。

许可证

MIT，见 LICENSE。

---

English

Black-box functional testing for DeepSeek Harness — HTTP APIs, Vue/Web pages, and front-end/back-end consistency, driven by conversation.

What it is

dsh-plugin-big-fat-fish is a DSH Host bundle. It adds three executor tools and one methodology skill:

Kind	Name	What it does
Tool	ft_http_run	Sends HTTP requests (suite of cases or inline cases), evaluates status / jsonPath / headers / latencyMs assertions, extracts variables, persists evidence
Tool	ft_ui_run	Drives local Chrome/Edge via playwright-core: page flows, selectors, waits, assertions, screenshots, console and failed-request capture
Tool	ft_report	Rebuilds a human-readable report from stored runs (single or aggregated, with known-failure marking)
Skill	functional-testing	The methodology: requirements → interface/UI entry review → test-case design → data confirmation → execution → attribution → per-round record and report

The skill owns what to test, what "expected" means, and whether a failure is a product bug, a test-case bug, or an environment problem. The tools only execute and keep evidence — they never guess.

What it is good at:

• Acceptance/regression testing of a version or module through its real interfaces and real pages (no mocks, no source changes).
• Reusing a login profile per project: after a login interface is probed once, every later round runs login + feature together.
• Writing per-round records: goal, coverage, test data, anomaly reproduction — so a report can be reproduced by another person.
• Front-end/back-end consistency checks: the same data read through the API and rendered on the page, compared field by field.

Installation

Requires a DSH build with the plugin manager and Node 22+. UI tests additionally need local Chrome or Edge.

Install from GitHub (pin a tag or commit for reproducibility):

dsh plugin add github:Utaoki-henji/dsh-plugin-big-fat-fish#v0.1.1

Desktop app: open Settings → Plugins → Install, and enter the spec above (or https://github.com/Utaoki-henji/dsh-plugin-big-fat-fish). Then restart the Harness.

Offline / conversation install: download the release bundle (dsh-plugin-big-fat-fish-v<version>.zip) and ask the assistant to install it — the bundle ships its own node_modules/playwright-core and needs no network during install.

After installing, the plugin appears as 测试工程师 大肥鱼 with the functional-testing skill.

Usage

Start a conversation turn with one of:

• 「开始测试」 / 「开始测试接口」 / 「开始测试页面」
• "start testing", "test this API", "test this page"

The skill then walks these steps, asking you for what it cannot know:

1. Confirm the project (name / repository / service).
2. Choose the test kind: interface only · UI only · module (API + UI + front-end/back-end consistency).
3. Give the entry: API base URL and path (or Controller/API doc) or the UI route; environment and credential source.
4. Login profile: if the interface needs a token, it reads .ft/logins/<project>.json; without a profile it asks you for the login interface rather than guessing a path.
5. Read the code (read-only) to derive the contract and branch checklist — never modified, never instrumented.
6. Confirm execution logic for multiple interfaces: sequential / branching / parallel, and the concrete order (it will not infer call chains from code structure).
7. Prepare test data and show you a data-confirmation card before writing anything.
8. Execute: 3–5 smoke cases first, then the full set; one run per topic.
9. Record and report: per-round goal, coverage, test data, anomaly reproduction, plus a combined report.

Workspace state

Everything the plugin keeps lives under .ft/ in the current workspace (never in your project source):

.ft/
  config.json              # api/ui base URLs, allowed hosts, destructive rules, size limits
  credentials.local.json   # local credential store (gitignored; env vars override it)
  logins/<project>.json    # login profile: endpoint, token path, auth header, env var names
  suites/                  # reusable test suites (credentials stay as ${env:...} placeholders)
  runs/<runId>/            # results.jsonl, evidence, screenshots, report

ensureWorkspaceState writes a .ft/.gitignore so runs, credentials and auth artifacts never leak into your repository.

Requirements

• DSH with plugin manager (Host bundle support).
• Node.js 22+ (ESM).
• Local Chrome or Edge for ft_ui_run (it drives your installed browser; no browser download required).
• Network access to the targets you test.

Safety and privacy

• Read-only on your project: the plugin never edits project source, never injects instrumentation, never runs your build.
• Writes are confined to the workspace .ft/ state directory.
• Secrets: credentials come from environment variables or .ft/credentials.local.json. Logs, evidence and persisted suites redact secret keys (authorization, cookie, password, token, …) and keep ${env:...} placeholders instead of values.
• No telemetry: the tools talk only to the base URLs you provide.
• Destructive calls are opt-in: non-GET calls must be marked destructive: true in the case and allowed by .ft/config.json.
• UI automation launches an isolated browser context; it does not touch your everyday browser profile.

Scope and limitations

• Execution is black-box only: HTTP requests plus real-browser UI flows. There is no unit-test execution, no code coverage, no breakpoints or instrumentation.
• White-box is static and read-only: during step 5 the assistant may read Java/Spring controllers, DTOs, validation and service branches, or Vue routes and API layers, to derive expectations. Nothing is executed from that reading.
• concurrency is fixed at 1; parallel requests run sequentially in the confirmed order.
• ft_ui_run uses the local Chrome/Edge channel. Headless mode and login state (storageState / cookies) are supported.

License

MIT — see LICENSE.