dsh-free-web-search

Keyless web search for DeepSeek Harness (dsh). No API key, no billing, no proxy — three public endpoints behind one search provider with language-aware failover.

中文说明

---

Why

dsh ships a search provider that needs a DeepSeek API key. That is a problem in two common setups:

• No official key. The user runs DSH against a third-party model gateway, so web_search cannot authenticate and simply fails.
• Official key, third-party model. The search still goes through DeepSeek's own endpoint, so the account is billed even though the chat runs elsewhere.

This plugin replaces the search side of ctx.web with providers that need no credentials at all. It does not touch web_fetch — fetching keeps whatever provider you already use.

How it searches

One provider, id free. Per request it walks a short chain and returns the first non-empty result set. Which backend leads depends on the shape of the query, because they fail in different ways:

Query	Order	Why
Single token, non-CJK (vllm)	Bing → Firecrawl → Exa	Bing answers in ~0.3 s and a one-token query has nothing to collapse
Everything else	Firecrawl → Exa → Bing	Bing collapses multi-token queries to their first token — see below

Backend	Endpoint	Notes
Bing	https://www.bing.com/search (HTML)	Fastest, but only safe for one-token queries
Firecrawl	https://api.firecrawl.dev/v2/search	Most reliable overall, 1–6 s, handles Chinese and multi-word queries
Exa	https://mcp.exa.ai/mcp (anonymous MCP)	Best snippets, but the anonymous quota dies hard and often

A backend that throws is parked for 3 minutes, in memory and in ~/.dsh/web-search-free.state.json, so a rate-limited endpoint is not paid for on every search. If every backend is parked, the parking lot is cleared and one real round is attempted — search never fails on stale bookkeeping.

When a fallback happened, the result carries (free search fell back to <backend>) so you can tell which endpoint actually served.

Install

Requires dsh with the ctx.web seam (@deepseek-ai/dsh-web). Node 20+ for the CLI.

From a local checkout:

git clone https://github.com/ywy-cyber/dsh-free-web-search.git
dsh plugin --profile <your-profile> add ./dsh-free-web-search

Or, inside DSH, install the directory through Plugins → Add plugin.

Restart DSH afterwards. The provider is loaded as an ES module and is not hot-replaced.

What it changes in your profile

The bundle's patch inserts one row and points the web seam at it:

- insert:
    - id: web-search-free
      name: 'dsh-free-web-search'
- id: web
  config:
    searchProvider: free
    fetchProvider: http

The - id: web override replaces the whole config of that row, which is why both fields are listed. fetchProvider: http is the shipped default — search changes, fetching does not.

To go back to the stock provider, remove this bundle; searchProvider becomes unset and dsh falls back to its default.

If you also run another plugin that registers a fetch provider, set fetchProvider to that plugin's id here instead of http, otherwise this patch will override its choice.

CLI

node cli.mjs probe            # connectivity of the three backends
node cli.mjs search "query"   # run one real search through the full chain
node cli.mjs status           # which provider the profile points at

probe and search talk to the network directly and do not need dsh running, which makes them the quickest way to tell whether an endpoint is down.

Limitations

Be aware of these before relying on it:

• All three backends are anonymous tiers. No quota is guaranteed. They rate-limit, they change response shapes, and they can start refusing you without notice.
• No time filtering. timeRange-style parameters are ignored.
• No relevance re-ranking. Results are deduplicated by URL and truncated; order comes from the backend.
• Bing results are scraped from HTML. A layout change breaks the parser.
• Bing collapses multi-token queries to their first token. Ask it deepseek harness plugin development and you get DeepSeek's homepage plus look-alike domains; a long Chinese query collapses to a single character. That is why it only leads for one-token queries. DeepSeek's own search API behaves the same way — the two return nearly identical result sets.
• Exa's anonymous quota is unreliable. It is a last resort, not a primary.
• Per-backend deadline: Bing 8 s, Firecrawl 9 s, Exa 6 s. Worst case with everything failing is about 23 s.
• Direct outbound access is assumed. Behind a proxy, set HTTPS_PROXY and NODE_USE_ENV_PROXY=1 in the process that runs dsh — Node's fetch does not read the system proxy.

Project layout

index.js           the provider: chain, language routing, cooldowns
cordis.patch.yml   what the bundle inserts into the profile
cli.mjs            probe / search / status
locale/            plugin-page title and description (en, zh)
docs/DETAIL.md     the page shown in Plugins
tests/smoke.mjs    offline-ish smoke test (still hits the network)
icon.svg           plugin icon

Contributing

Issues and pull requests are welcome. If you add a backend, include the evidence that it actually returns relevant results for both a Chinese and a Latin query — an HTTP 200 means nothing on its own. One backend in an earlier draft answered 200 OK and returned Olympic trapeze harnesses for "deepseek harness".

License

MIT