dsh-plugin
DayuanJiang/next-ai-draw-iodsh 的 draw.io 绘图插件:通过 next-ai-drawio MCP 服务器作图,点回复里的链接在右侧栏打开实时预览,可以手动修改图,AI 的每次修改都是可以恢复的版本。需 Node.js 18+,首次使用由 npx 下载 MCP 服务器。
OtterMind SQLX 数据库工具:列出已保存的数据源、查看或测试连接、执行 SQL、打开本地结果页,缺少 sqlx CLI 时首次调用自动安装。
Native SQLX tools for DeepSeek Harness. The plugin registers six defineTool tools and calls the
sqlx CLI for each one, so encrypted credentials, TLS policy, worker downloads and local result
pages stay CLI responsibilities.
dsh plugin --profile <profile> add @ottermind/sqlx-dsh
cordis.patch.yml is declared as dsh.bundle.patch, so installing the package appends the plugin
row to the profile's bundle stack; no profile file edits are needed. Remove it with
dsh plugin --profile <profile> remove @ottermind/sqlx-dsh.
@deepseek-ai/dsh-tools (peer dependency) — bundled with dsh, and the profile installer runs
pnpm with autoInstallPeers: false, so nothing extra is installed. The peer range covers
>=0.1.5-rc.1 <0.2.0, >=0.1.6-alpha.1 <0.2.0-0 and >=0.2.0-rc.1 <0.3.0-0; the plugin
uses only defineTool and ctx.tools.register, which are unchanged in 0.2.sqlx CLI 0.1.15 or later. The plugin looks
for SQLX_BIN, then sqlx on PATH, then the official user-level installation
(~/.local/bin/sqlx, %LOCALAPPDATA%\Programs\SQLX\sqlx.exe, or SQLX_INSTALL_DIR), and when
none of them exists it installs the CLI itself with
npx -y @ottermind/sqlx@latest --target dsh. Adding the plugin is therefore the only setup step.| Tool | Purpose | May modify data |
|---|---|---|
sqlx_datasource_list | List saved datasources (no secrets) | no |
sqlx_datasource_show | Show one datasource by UUID or unique name | no |
sqlx_datasource_test | Test connectivity; executes no SQL | no |
sqlx_datasource_setup | Start adding a datasource the user finishes in a browser page; returns that page's URL | adds a connection |
sqlx_datasource_setup_status | Report the outcome of a browser setup; read-only | no |
sqlx_sql_execute | Execute statements, return the table-shaped result with a preview (full returns every row) | yes |
sqlx_results_rows | Read one page of a stored result; read-only | no |
sqlx_sql_view | Execute once, return a local result-page URL | yes |
sqlx_prefetch | Download components ahead of a first query (all fetches every one) | downloads only |
The two execution tools describe the authorization rule themselves: run them only after the user authorized that exact operation and scope. Results are never replayed automatically.
Unlike the MCP server, these definitions carry no approval metadata: defineTool has no annotation
field, so the rule above is a description the model is asked to follow rather than something dsh
enforces. A deployment that needs an enforced prompt can gate the two execution tools from a
tools/pre-execute handler.
Save credentials in the SQLX data directory (sqlx datasource add --connection-stdin, or the local
sqlx ui page). A datasource created with --username-env/--password-env reads environment
variables at execution time, and a harness started from a GUI usually cannot see your shell
environment.