Search Enhance for DeepSeek Harness

English | 简体中文

dsh-search-enhance is a search extension for DeepSeek Harness. It uses a Grok-compatible Search API for primary web answers and can optionally use Context7, Exa, Tavily, and Firecrawl for documentation lookup, supplementary discovery, page extraction, and site mapping.

The plugin handles search, source retention, and page retrieval as separate steps. web_search and docs_search return an answer or documentation snippets with visible sources; the complete source record can be stored under a source_ref and paged later. Important pages can then be retrieved with web_extract, so search snippets remain distinct from fetched page content.

Bring your own endpoints and credentials. The plugin ships no API keys. A Grok-compatible endpoint is required for web_search; Context7, Exa, Tavily, and Firecrawl are optional.

Key characteristics

• web_search uses the Grok-compatible endpoint for the main answer and runs all user-selected supplemental Providers (Exa, Tavily, Firecrawl) in parallel, with an independent per-Provider source limit.
• Sources are normalized, de-duplicated, and reordered using source category, requested version, and publication-time signals before they are shown.
• source_ref keeps the complete source record in private durable storage, allowing the Agent to paginate beyond the links included in the initial result.
• docs_search uses Context7 only with an explicit library_name or library_id; requests without a library identity use Exa discovery.
• web_extract requires one explicit provider from the user-enabled set: tavily_extract, firecrawl_scrape, smart_direct, or direct. It reports the route, evidence level, and available page metadata; failures never switch Providers.
• Source pagination, Context7 detail operations, site mapping, research planning, and diagnostics are disclosed on demand through search_tools and search_call.
• Native Tool Mode and Code Mode use the same fixed tool surface and canonical outputs. DSH Settings, Credentials, Agent Presets, guards, and lifecycle cleanup continue to apply.
• Supplemental search defaults to no Providers selected. Unselected Providers are never requested; missing credentials and failures of selected Providers remain visible as warnings.

For the complete routing, evidence, and progressive-disclosure flow, see Search workflow architecture.

Quick start

1. Install

Supports only DSH 0.1.5-rc.2. DSH dependencies are pinned to this exact release; older versions and other prereleases are not supported.

If migrating from the scoped @kkkneko/dsh-search-enhance package, first run dsh plugin --profile web remove @kkkneko/dsh-search-enhance to avoid loading both bundles.

Install the published npm bundle into the DSH web profile:

dsh plugin --profile web add dsh-search-enhance@latest

2. Start DSH Web

dsh web

3. Configure search

Open:

Settings → Plugins → Plugin configuration → dsh-search-enhance

In the settings card:

1. Set the xAI endpoint or an explicit Grok-compatible gateway and its key under Credentials → Grok search.
2. Choose the matching completions or responses protocol and model under Grok search backend.
3. Use the Supplemental search providers and Available extraction providers dropdowns to select multiple services. Each service's endpoint, timeout, and credential reference are grouped in its credential card.

The default credential reference is SEARCH_API_KEY. Credential values are stored through DSH Credentials and are not exposed as model parameters.

Official Grok endpoints automatically enable native search without another protocol or toggle:

• Version paths may be omitted: https://api.x.ai / https://us.api.x.ai gain /v1; https://openrouter.ai gains /api/v1 (or /api gains /v1). This applies to search and model-list requests, without duplicating complete paths or changing custom gateways.
• xAI: https://api.x.ai/v1 (also https://us.api.x.ai/v1), with a model such as grok-4.6. Requests use Responses with web_search + x_search, even when completions is selected; saved settings are not rewritten.
• OpenRouter: https://openrouter.ai/api/v1, with a model such as x-ai/grok-4.6. The selected protocol is retained and openrouter:web_search with engine: native is sent; no :online suffix is needed. Model support and workspace policy still govern OpenRouter's actual engine; to prohibit its fallback, allow only native in the workspace.
• Detection is restricted to those official HTTPS hosts and Grok models. Other models and compatible proxies are unchanged. Native adapters map minimal to low and max to xhigh; off still omits the reasoning parameter.
• Structured citations enter the source list. If the API reports no completed search, structured citations, or search usage, the result explicitly warns that native search is unconfirmed; prose links alone are not execution evidence.

Save the settings, restart DSH, and ask a current-information question. A successful run shows a Search tool row, an answer, and source links.

Example requests

Use normal language; the plugin gives the Agent routing guidance.

• “Find the most important React 19 user-visible changes. Prefer official release notes and include source links.”
• “Look up the current FastAPI JWT authentication API and show a minimal example from the official documentation.”
• “Read and summarize https://example.com/article, separating what the page states from your inference.”

Ask explicitly when you need complete source pagination, site discovery, a research plan, or Provider diagnostics.

Providers

Configure only the routes you need.

In the plugin settings, select any combination of supplemental search Providers. Each selected Provider runs alongside the main search, using supplementalSearch.maxSourcesPerProvider (default 5, maximum 100). This selection applies to all search profiles; there is no implicit Exa enhancement or shared 60/40 budget.

Separately, select the extraction Providers available to the Agent. Each web_extract call must choose exactly one enabled Provider, for example { "url": "https://example.com/", "provider": "direct", "format": "markdown" }. Disabled, unconfigured, unsupported, or failed Providers do not trigger automatic fallback. Extraction Providers are enabled by default; disable any you do not want the Agent to use.

Configuration change: replace the former extraDiscoverySources profile budgets with supplementalSearch selections. Enable exa, tavily, and/or firecrawl explicitly; existing API credentials alone do not enable supplemental search. Existing extraction enabled settings remain the allowlist. Save settings and restart DSH. fallbackMode now applies only to the unchanged documentation-search routing.

For docs_search, Context7 requires an explicit library_name or library_id. Without one, provider: "auto" uses Exa instead of guessing a package name from the full question.

Tool disclosure

The model-facing surface remains five tools: web_search, docs_search, web_extract, search_tools, and search_call. Advanced operations are disclosed through manifests rather than registered as additional model tools.

The default progressive mode activates a newly disclosed capability on the next model step. In all mode, deferred operations are active immediately. Native Tool Mode and Code Mode use the same schemas, execution policy, and canonical outputs.

When web_search or docs_search returns a source_ref, the plugin automatically activates source pagination and appends its real operation manifest.

Update and uninstall

To update, run the installation command above again. To remove the plugin:

dsh plugin --profile web remove dsh-search-enhance

Restart DSH after updating or removing the bundle.