dsh-custom-search-fetch

English | 中文

A DeepSeek Harness plugin that runs the built-in web_search tool on a self-hosted SearXNG instance and the built-in web_fetch tool on a self-hosted Firecrawl instance.

The plugin does not register its own tools. It uses the ctx.web provider seam that @deepseek-ai/dsh-web provides:

it registers a search provider custom-searxng and a fetch provider custom-firecrawl;
its bundle patch pins @deepseek-ai/dsh-web to them (searchProvider: custom-searxng, fetchProvider: custom-firecrawl).

The shipped web_search / web_fetch tools, with their output format, citation instruction, per-agent-preset availability, and timeout policy, are unchanged. Only the backend changes.

Requires DeepSeek Harness >=0.2.2-rc.1.

Install

dsh plugin --profile <profile> add github:mercurynomercy/dsh-custom-search-fetch

You can also install it from the Plugins page / dsh-market once it is listed. Removing the plugin restores DeepSeek search and anonymous HTTP fetch.

Configure

Open Plugins → dsh-custom-search-fetch. Under the README, a settings card edits every field below, grouped into SearXNG search and Firecrawl fetch. Saving writes the values to your profile's cordis.patch.yml and applies them right away, with no restart. Each overridden field has Reset to default, and an invalid value blocks the save. The Firecrawl API key field is write-only: it starts blank, and saving it blank keeps the stored key.

You can also edit cordis.patch.yml by hand. An id-targeted row replaces the whole config, so restate every field you set:

- id: dsh-custom-search-fetch
  config:
    searxngBaseURL: http://localhost:8080
    firecrawlBaseURL: http://localhost:3002

Field	Default	Meaning
`searxngBaseURL`	`http://localhost:8080`	SearXNG base URL, without `/search`
`searxngMaxResults`	`10`	Results per search (the tool's `searchMaxResults`, 8 by default, caps it further)
`searxngMaxPages`	`2`	Most result pages fetched to reach that count
`searxngLanguage` / `searxngCategories` / `searxngEngines`	empty	Passed through to SearXNG when set
`searxngSafeSearch`	unset	`0`, `1`, or `2`; unset uses the instance's `search.safe_search`
`searxngTimeRange`	empty	`day`, `week`, `month`, or `year`
`searxngTimeoutMs`	`15000`	Timeout for one page request
`firecrawlBaseURL`	`http://localhost:3002`	Firecrawl API base, without `/v2`
`firecrawlApiVersion`	`v2`	`v2` or `v1` scrape endpoint
`firecrawlApiKey`	empty	Literal key (secret field); leave empty for a self-host without auth
`firecrawlApiKeyEnv`	`FIRECRAWL_API_KEY`	Credential or environment variable to read when no literal key is set
`firecrawlOnlyMainContent`	`true`	Strip navigation, headers, and footers
`firecrawlTimeoutMs`	`25000`	Firecrawl's own scrape timeout
`firecrawlMaxAgeMs`	`0`	v2 only: accept a cached copy up to this age; `0` always scrapes fresh
`firecrawlMaxBodyChars`	`100000`	Markdown characters returned before truncation

SearXNG prerequisites

SearXNG answers 403 unless its JSON output format is on. Enable it in settings.yml:

search:
  formats:
    - html
    - json

If the instance's limiter is on, allow the Harness host or turn the limiter off for it. Otherwise searches fail with 429.

Timeouts

web_fetch runs under tool-web's fetchTimeoutMs budget, which is 30 s by default. Keep firecrawlTimeoutMs below it. If you raise it for slow, JavaScript-heavy pages, raise fetchTimeoutMs on the tool-web row as well.

Behaviour

Search: GET /search?format=json. The plugin pages until it has enough unique URLs or reaches searxngMaxPages. SearXNG answers and infobox text become the result's content, which web_search shows above the sources. If no results come back and some engines failed, the error names those engines. If nothing was found but SearXNG has spelling suggestions, they are returned as the content.
Fetch: POST /v2/scrape (or /v1/scrape) with formats: ["markdown"]. The page Markdown is returned as a text body, with the page title as a heading. The status code and final URL come from Firecrawl's metadata. The plugin checks URLs the same way the shipped HTTP fetcher does before sending anything: http(s) only, at most 2048 characters, no embedded credentials. Firecrawl also converts PDFs and other documents to Markdown.
Errors are WebErrors: WEB_PROVIDER_ERROR names the endpoint and how to fix it, WEB_ABORTED is a cancellation, and WEB_INVALID_URL / WEB_BLOCKED_URL are rejected URLs. If you clear a base URL, the pinned provider becomes unavailable and calls fail with WEB_PROVIDER_CONFIGURED_UNAVAILABLE.

Switching one backend back

Restate the web row in your profile cordis.patch.yml. For example, to keep SearXNG search and fetch with the built-in anonymous HTTP fetcher:

- id: web
  name: '@deepseek-ai/dsh-web'
  config:
    searchProvider: custom-searxng
    fetchProvider: http

The built-in ids are deepseek-official (search), http (fetch), and deepseek-search-api / deepseek-fetch-api. Saving the Web search settings page's "Use the new Web Search API" switch also rewrites this row in your profile patch, and that row wins over this plugin.

Conflicts

Provider ids are unique within each kind. Another plugin that registers custom-searxng or custom-firecrawl makes the second registration fail with WEB_DUPLICATE_PROVIDER. Do not also insert the dsh-custom-search-fetch row by hand, because the bundle patch already inserts it.

Develop

npm install
npm test                                                      # unit tests, mocked fetch, real dsh-web + Cordis
SEARXNG_URL=http://localhost:8080 FIRECRAWL_URL=http://localhost:3002 npm run smoke -- "query" https://example.com

To try a checkout in a profile: dsh plugin --profile <profile> add link:$(pwd).

License

MIT

dsh-custom-search-fetch

安装

README