@lp181818/web-search-advanced

Advanced web search plugin for DeepSeek Harness (DSH), supporting both native
DeepSeek search and OpenAI-compatible "Custom" search providers.

Features

• Provider selector: deepseek (native) or custom (OpenAI-compatible, e.g. OpenRouter)
• Settings section: provider choice, endpoint URL, model, API key, max searches
• Two search backends: native DeepSeek web_search_20250305 or an OpenAI-compatible web plugin
• Automatic retry (v0.1.4): transient failures (HTTP 429 / 5xx / network
  errors) are retried with exponential backoff, honoring the server
  Retry-After header (configurable maxRetries, default 3)
• Test before save (v0.1.4): the settings card probes the configured
  endpoint (/chat/completions for Custom, /messages for DeepSeek) with the
  current draft values before saving; real config errors (bad endpoint/model/key)
  block the save, transient 429/5xx only warn

DSH compatibility

The current source no longer imports the removed settingsNamespace,
installSettingsSection, or @deepseek-ai/dsh-client-runtime APIs.

Enabling this provider in DSH

Installing this package automatically routes the DSH web service to
web-search-advanced; no manual cordis.patch.yml edit is required. After
installation, restart dsh web and configure the package in the Web Search
settings section.

Only the provider, API key, and model are required. The endpoint is optional:
DeepSeek uses https://api.deepseek.com/anthropic/v1, while Custom uses
https://openrouter.ai/api/v1 when the endpoint is left blank. Inside the
web-search-advanced settings namespace:

• searchProvider: deepseek uses this plugin's DeepSeek branch.
• searchProvider: custom uses the configured OpenAI-compatible
  baseURL / model, for example OpenRouter.
• maxRetries: 3 (optional) — exponential-backoff retries for transient
  failures (429 / 5xx / network), default 3.

Settings persist in ~/.dsh/settings.yaml under the web-search-advanced
namespace and survive DSH restarts until deleted.

The package patch also selects the built-in http fetch provider so search
results can be opened without an additional profile change.

Build

This repository includes the prebuilt lib/ artifacts required by DSH skill
hub validation. Direct GitHub installs also have a prepare script that can
rebuild them.

DSH resolves @deepseek-ai/* packages as runtime peers, so local installs
should not make npm resolve them:

npm install --legacy-peer-deps
npm run build

Build output:

• lib/index.js - DSH host/plugin half
• lib/client.js - DSH web client bundle

After a build, sync into an installed DSH profile:

npm run sync

If no profile has the package installed yet, install it first:

dsh plugin --profile web add github:bug-huntter/web-search-advanced

Then restart DSH and apply the profile patch:

dsh web

The plugin registers its provider through ctx.web, registers the
web-search-advanced settings namespace, and exposes that namespace to the
web client through the configurable-provider registry.