dsh-web-search-searxng

A SearXNG-backed WebSearchProvider for the
DeepSeek Harness web capability seam (ctx.web).

SearXNG is a self-hosted metasearch engine. One search here is a plain retrieval call against the
instance's /search?format=json endpoint, so unlike the shipped DeepSeek provider it needs no API
key and costs no model turn — the shipped provider issues a full Messages request with the
native web_search server tool, paying latency and generated tokens for every search.

This is an implementation package: it registers a provider into ctx.web and does not
register a model-facing tool. @deepseek-ai/dsh-tool-web owns web_search, its schema, its prompt
guidance, and the result card. Installing this package makes that existing tool work against your
own instance.

Install

From npm

dsh plugin --profile web add dsh-web-search-searxng

From this repository

dsh plugin --profile web add github:chinng-inta/dsh-web-search-searxng

A git dependency ships no lib/, so the package builds itself through its prepare script — and
pnpm blocks install scripts until you allow them. The first run therefore fails and prints the exact
key to allow. Add it to your profile's pnpm-workspace.yaml and re-run:

allowBuilds:
  "dsh-web-search-searxng@https://codeload.github.com/chinng-inta/dsh-web-search-searxng/tar.gz/<commit-sha>": true

The key pins a commit, so it changes whenever you install a newer revision. Building from source also
needs the dev toolchain fetched (~20 s versus ~2 s for the registry). Prefer npm unless you are
tracking unreleased changes.

Either way

The package declares dsh.bundle, so a single command also activates it: the bundle patch
inserts the provider row and selects it on the web row. Point it at your instance and restart:

export SEARXNG_URL=http://searxng.internal:8888

Verify the composition before booting:

dsh --profile web --dump-config | grep -A3 'id: web'

Your instance must serve JSON

SearXNG does not enable the JSON API by default. In the instance's settings.yml:

search:
  formats:
    - html
    - json

Without it the endpoint answers with the HTML result page, and this provider fails with a message
naming the fix rather than a parse error.

A public instance is a poor backend: most refuse programmatic access (HTTP 403 from a bot filter)
or rate-limit aggressively. Run your own.

Configuration

The plugin owns the web-search-searxng settings namespace, so its section resolves in the harness's
own layering:

schema defaults  →  the plugin row's `config` (composition base)  →  the user layer in the settings document

baseURL additionally falls back to $SEARXNG_URL when no layer sets it. Nothing else is needed to
get started: export the variable and the provider is configured.

The section is projected per search, so an edit to the settings document reaches the next search
without a restart — and clearing baseURL falls back to the environment again rather than stranding
the provider on a value it can no longer see. Registration itself never moves, so provider selection
does not flicker when configuration changes.

A deployment without a settings provider mounted keeps working: the source falls back to the
composition entry, exactly as composed.

All keys are optional.

As the plugin row's composition base:

- id: web-search-searxng
  name: 'dsh-web-search-searxng'
  config:
    baseURL: http://searxng.internal:8888
    language: ja
    categories:
      - general
      - news

…or as the user layer in the harness settings document ($DSH_HOME/settings.yaml by default),
which wins over the row above and is what a configuration surface writes:

web-search-searxng:
  language: ja
  maxSnippetChars: 300

Every search-shaping knob is a deployment setting, not a model argument. The seam's
WebSearchRequest is deliberately just query + maxResults; provider-neutral controls (recency,
domain filters, search depth) are named deferred work upstream. Keeping them in config is what makes
this provider substitutable for the shipped ones.

timeoutMs is a resource backstop, not the model-facing tool-call budget —
@deepseek-ai/dsh-tool-call-timeout-policy owns that via tool-web's searchTimeoutMs. Leave this
below the tool budget so a slow instance surfaces as a provider failure rather than a tool timeout.

Provider selection

The bundle patch sets web.searchProvider: searxng. This is required, not opinionated.

The seam auto-selects only when exactly one registered provider is usable, and
@deepseek-ai/dsh-web-search-deepseek reports usable whenever a credential resolver exists — which
its own apply() always supplies — so it answers available() === true on a stock composition even
with no key configured. Registering a second provider without naming one would make every search fail
with WEB_PROVIDER_AMBIGUOUS.

Bundle layers apply before your profile's cordis.patch.yml, the home patch, and any --patch
overlay, so you can always override the choice. But note that a patch replaces the targeted row's
whole config: if you patch the web row yourself for anything else, restate
searchProvider: searxng there too.

Mapping

Sources are deduplicated by URL, because a metasearch merges engines that routinely return the same
page. Blank strings are treated as absent rather than emitted as empty fields — the seam's optional
fields exist so an adapter never has to invent them.

truncated is always false from this provider: the seam owns maxResults enforcement, and
reporting our own truncation would misattribute whose bound cut the list.

Errors

Failures are WebErrors the tool layer turns into a readable tool result.

available() is a cheap synchronous check — a parseable http(s) base URL — as the seam requires;
it never touches the network.

Redirects are refused (redirect: 'error'). A self-hosted instance has no reason to redirect a
search, and following one would send the query to a host the deployment never configured.

Known limitations

• maxResults is not pushed down. SearXNG exposes no result-count parameter, so the instance
  returns its full first page and the seam truncates. This bounds tokens, not the instance's work.
• publishedDate is usually absent. General web engines rarely date results; news engines
  usually do. Filter with categories: ['news'] if you need dates.
• Infoboxes are not surfaced. They are structured entity cards rather than an answer to the
  query, so flattening them into content would present them as one.
• No per-engine failure reporting. SearXNG reports unresponsive_engines[] on partial failures;
  the seam's result shape has nowhere to put it, so a degraded search looks like a thin one.

Compatibility

0.3.3 is not backward compatible, and neither is anything before it forward compatible.
Installing the wrong pair does not merely lose this card — it fails the whole Web UI's plugin
load, because both breakages surface as an unanswerable client require or an invalid slot
registration (#1):

• rc.8 dropped @deepseek-ai/dsh-client-web-react from the loader's seed table. Up to 0.3.2
  this package imported bindSnapshotSelector from it. 0.3.3 hands its store to the renderer
  through the slot inject face's hooks compartment instead, which is where rc.8 synthesizes
  selector hooks.
• rc.8 turned settings.plugin.item from a list slot (id + order) into a keyed one
  (key = the settings namespace the card edits). 0.3.3 registers the keyed way, which rc.6
  rejects.

0.3.0 shipped the settings card without its stylesheet — it works, but renders with browser
defaults. Use 0.3.1 or later.

The harness is a developer preview with breaking changes between release candidates, and its
packages publish the active line under the next dist-tag (latest still points at the older
0.0.1-rc.1). Pin your harness version.

License

MIT