dsh-crawshrimp-scripts
用于可移植 Crawshrimp 页面脚本的独立 DeepSeek Harness 工具
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:howtimeschange/dsh-crawshrimp-scripts#42a70294f58d6588ec39417dbb1da363d9fbf2f0Crawshrimp Scripts 以 dsh-crawshrimp-scripts 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
dsh-crawshrimp-scripts adds two standalone tools to DeepSeek Harness:
crawshrimp_scripts_list discovers portable Crawshrimp adapters.crawshrimp_scripts_run opens the declared URL in a fresh isolated Playwright browser context, evaluates the declared page script, and saves the returned JSON rows as a uniquely named artifact.The plugin has no Crawshrimp backend dependency. It supports the small portable adapter subset described below and does not claim login emulation, existing-page control, scheduling, Excel export, or the full Crawshrimp runtime.
For a released GitHub package:
dsh plugin --profile demo add -w github:howtimeschange/dsh-crawshrimp-scripts#v0.1.0
For a local checkout, clone the repository, install the lockfile dependencies, and add the absolute checkout path:
git clone https://github.com/howtimeschange/dsh-crawshrimp-scripts.git
cd dsh-crawshrimp-scripts
npm ci
dsh plugin --profile demo add -w /absolute/path/to/dsh-crawshrimp-scripts
The package declares dsh.bundle.patch metadata through package.json's dsh.bundle.patch entry. cordis.patch.yml registers the plugin and is included in the package bundle.
Set the plugin config in the DSH profile. Use absolute adapter roots in a real deployment:
adapterRoots:
- /absolute/path/to/adapters
workspaceDir: /absolute/path/to/dsh-script-workspace
browserChannel: chrome
storageStatePath: /absolute/path/to/private/storage-state.json
timeoutMs: 120000
maxScriptBytes: 524288
maxOutputBytes: 2097152
The default browser is the installed Chrome channel. Set CHROME_PATH or executablePath when the executable is installed in a nonstandard location. cdpEndpoint is optional; when supplied, every run still requests a new browser context and fails with an isolation error if the endpoint cannot provide one. The plugin never selects or mutates existing user pages.
Every script run requires native DSH approval before browser startup. A missing approval service, missing agent, rejected request, cancelled request, or the DSH never deny policy fails closed. Only a one-shot allowed-once outcome can start the browser.
Each configured root contains adapter directories. Every adapter directory contains a strict manifest.yaml:
id: dsh-demo-table
name: DSH demo table
version: 0.1.0
entry_url: https://example.com
tasks:
- id: read-table
name: Read page title and table
script: read-table.js
Supported manifest fields are id, name, version, optional description and author, entry_url, and tasks. Each task supports id, name, optional description, and a relative .js script. Adapter and task IDs must be lowercase slugs; task IDs must be unique. Symlinked manifests and scripts are resolved with realpath and are rejected when they escape their configured adapter directory.
The task script is an async IIFE evaluated in the page after navigation. It receives a JSON-cloned object at window.__PARAMS__ and must resolve to:
{
success: true,
data: [{ /* JSON row */ }],
meta: { /* JSON metadata */ },
}
The result is bounded by maxOutputBytes and is saved under workspaceDir/artifacts/<adapter>/<task>/ using exclusive file creation, so an existing artifact is never overwritten. The tool returns the artifact path relative to workspaceDir.
For sites that require a login, set storageStatePath to an explicitly selected Playwright storage-state JSON file. The file is read-only, bounded to 5 MiB, and loaded into each newly created context; the default is empty. Create one interactively with node examples/save-storage-state.js <url> <absolute-output-path> using a headful Chrome window. The helper never overwrites an existing state file. Prefer a filename ending in .storage-state.json; it is ignored by this package's .gitignore. Keep this file private and out of source control.
Copy only the portable page task into an adapter directory and remove backend-only manifest fields such as auth, trigger, params, output, task-specific entry overrides, and scheduling declarations. Rename window.__CRAWSHRIMP_PARAMS__ to window.__PARAMS__, and end each task with the { success, data, meta } result contract. Put any caller values in the params object passed to crawshrimp_scripts_run.
This plugin does not provide the Crawshrimp backend token/API bridge, login persistence, current-tab matching, task scheduling, pagination orchestration, download/export helpers, or Excel/notification sinks. A migrated script should perform only the page work it can complete inside one isolated browser context and return JSON rows.
See examples/demo-adapter for a synthetic title/table adapter.