dsh-websearch-custom

dsh-websearch-custom registers a configurable ctx.web provider that calls any OpenAI Responses-compatible endpoint with its native web-search tool. Point it at a gateway that exposes a model with Responses API web search; endpoint, model, key reference, and limits are all configured by the user — the package ships no vendor-specific defaults.

It does not scrape the web, call ordinary chat completions, or turn uncited model prose into search evidence. A response must contain native URL citations or search sources; otherwise the provider returns WEB_PROVIDER_ERROR.

Install

Install from a checkout or a Git repository into the running DSH profile:

dsh plugin --profile web add /absolute/path/to/dsh-websearch-custom

For a Git-hosted package, pin a reviewed commit. Git installs need this package's prepare script to build dist/; pnpm may require adding dsh-websearch-custom: true to the profile's pnpm-workspace.yaml allowBuilds section.

Restart the DSH Host after installation.

Settings UI

The bundle ships a browser settings page (Settings → WebSearch Custom) that
edits the same configuration at runtime — no restart or composition edit is
needed after saving:

• Provider ID — the stable id the web composition row selects
  (web.searchProvider).
• API Key — optional; stored as a settings secret and never echoed back.
• API Key Env — environment variable or DSH credential reference resolved
  per search. No default; leave empty when the API Key is entered directly.
• Base URL — OpenAI Responses-compatible endpoint base, without
  /responses. Required.
• Model — Responses model expected to support native web search. Required.
• Search tool type — web_search or web_search_preview (which tool your
  gateway supports; check its documentation).
• Search context size — low / medium / high.
• Max output tokens — Responses output-token ceiling.
• Allowed / Blocked domains — comma-separated native search filters.

Values are persisted to the websearch-custom settings namespace and are read
on every search, so a saved change applies to the next web_search call.

Configure

The bundle only registers custom-openai; it intentionally does not replace
DSH's default deepseek-official provider. In a user-owned Host/profile
overlay, point the web row at it:

- id: web
  name: '@deepseek-ai/dsh-web'
  config:
    searchProvider: custom-openai

Mount the provider row. Without explicit values the endpoint, model, and key
reference must be supplied through the Settings UI before the first search:

- id: web-search-custom
  name: dsh-websearch-custom

To pin values in composition (as the base layer the UI edits), replace the
row with explicit configuration — the example below uses neutral placeholders:

- id: web-search-custom
  name: dsh-websearch-custom
  config:
    providerId: custom-openai
    apiKeyEnv: MY_PROVIDER_API_KEY
    baseURL: https://api.example.com/v1
    model: my-search-enabled-model
    searchToolType: web_search
    searchContextSize: low
    maxOutputTokens: 1024
    allowedDomains:
      - arxiv.org
      - openreview.net

apiKey is also accepted but should be avoided because it would place a secret in composition. apiKeyEnv is resolved through DSH credentials first, then the launch environment.

Supported configuration (no vendor defaults — endpoint, model, and key reference are user-supplied):

Compatibility and limits

OpenAI-compatible endpoints differ substantially. Before making this the default provider, confirm that the exact endpoint and model accept POST /responses with a native web_search_preview or web_search tool and return URL citations. The package deliberately fails rather than silently producing synthetic citations when an endpoint supports Responses text but not native search.

The request uses an OpenAI Responses shape with input, tools, and max_output_tokens. Some gateways accept only web_search, not web_search_preview; change searchToolType accordingly. Some gateways use different filter field names, in which case leave domain filters unset or adapt the provider to that gateway.

Development

pnpm install
pnpm test
pnpm run typecheck
pnpm run build

The test suite mocks fetch; it does not call a real search endpoint.