dsh-deepseek-search-account

English | 中文

A tiny DeepSeek Harness (DSH) plugin that keeps the built-in web_search tool working when your chat model is served by an external, non-official subscription — for example OpenCode Go (opencode.ai/zen/go), another Anthropic-compatible gateway, or a pi-ai provider — while DeepSeek Harness is signed in to a DeepSeek account.

Searches then run on the official DeepSeek Search API using that DeepSeek account, so web_search keeps working without a DEEPSEEK_API_KEY.

When to use it / 适用情况

Use this plugin when all of the following are true:

• Your DSH chat model is served by a third-party subscription or gateway (OpenCode Go / Zen, another Anthropic-compatible endpoint, pi-ai, etc.), not by a DeepSeek Account model.
• You are signed in to a DeepSeek account in DSH (Settings → Account), and that account may be used for search.
• The built-in web search fails with DeepSeek search has no API key for "DEEPSEEK_API_KEY", and you do not have (or do not want to use) a DeepSeek API key.

In that situation the harness keeps working as before, but web_search returns an error while web_fetch is fine. This plugin makes web_search use the official Search API with the signed-in account, so search still runs on the DeepSeek account's official quota and the chat subscription is untouched.

Do not install it when:

• Your chat model is a DeepSeek Account model: the shipped provider already searches with the account sign-in.
• You have a DEEPSEEK_API_KEY stored: the shipped provider already works (and the Search API can be enabled in Plugins → Built-in plugin settings → Web search).
• You already use another web-search provider plugin (Exa, Tavily, SearXNG, Zhipu, etc.) that suits you.

Why it is needed

The shipped search providers (dsh-web-search-deepseek, dsh-web-deepseek-api) only use the DeepSeek account sign-in when the initiating session's model route is deepseek-account:

ctx.get("agents")?.currentInitiator()?.session.requestContext()?.provider === "deepseek-account"

With OpenCode Go as the chat provider, that check fails, the provider falls back to the DEEPSEEK_API_KEY credential, and — with no key stored — web_search fails. This plugin resolves the same account token (ctx.deepseekAccount.resolveToken(endpoint)) on every search, without looking at the conversation's model route, and falls back to DEEPSEEK_API_KEY (Bearer) only when account sign-in is unavailable.

What it does

• Registers a search provider named deepseek-account-search on the ctx.web seam.
• Calls the official DeepSeek Search API (POST https://api.deepseek.com/v1/tools/search) by default — one HTTP request per search, about 0.5–1 second, returning title, url, snippet and date; no auxiliary model turn, no extra model tokens.
• Keeps a legacy mode: messages backend (Anthropic-compatible Messages endpoint + the native web_search_20250305 server tool) as an explicit fallback.
• Ships a bundle patch (cordis.patch.yml) that registers the provider and pins web.searchProvider to it while keeping web.fetchProvider: http.

Install

From GitHub into a profile

dsh plugin --profile desktop add github:Chip-G0202/dsh-deepseek-search-account
# for a plain `dsh web` server: --profile web

The package declares dsh.bundle, so dsh plugin add records it in the profile's bundles and applies cordis.patch.yml at boot — no hand-written row is needed. Installing from this check-out works the same way:

dsh plugin --profile desktop add /path/to/dsh-deepseek-search-account

Or attach the prebuilt tarball from the latest GitHub release:

dsh plugin --profile desktop add https://github.com/Chip-G0202/dsh-deepseek-search-account/releases/latest/download/dsh-deepseek-search-account.tgz

Then restart DSH (a profile bundle list change is applied on boot). For the Desktop app use the desktop profile as above.

Manual install (file-path row)

If you prefer not to install a package, copy this folder somewhere stable and add two rows to the profile's cordis.patch.yml (~/.dsh/profiles/<profile>/cordis.patch.yml):

- insert:
    - id: web-search-deepseek-account
      name: '/path/to/dsh-deepseek-search-account/provider.js'
      config:
        mode: search-api
        count: 10
- id: web
  name: '@deepseek-ai/dsh-web'
  config:
    searchProvider: deepseek-account-search
    fetchProvider: http

Note: DSH reloads profile configuration live, but it does not re-import a plugin module when only the module path of an existing row changes. After editing provider.js, restart DSH — or toggle the row disabled: true / back — to load the new code.

Configuration

Installed as a bundle, the row lives in this package's cordis.patch.yml. Override it from the profile patch with an id-targeted entry that repeats the row name:

- id: web-search-deepseek-account
  name: dsh-deepseek-search-account
  config:
    mode: search-api
    count: 8

Key	Default	Meaning
mode	search-api	search-api = official Search API; messages = Anthropic Messages + web_search_20250305 server tool
baseURL	https://api.deepseek.com/v1/tools	Search API base URL (/search is appended)
count	10	Results per Search API request, 1–50; the tool's searchMaxResults wins when smaller
sendOriginalQuestion	true	Send ori_question (the session's latest user prompt, ≤128 chars)
apiKey / apiKeyEnv	unset / DEEPSEEK_API_KEY	API-key fallback when account sign-in is unavailable
messagesBaseURL	https://api.deepseek.com/anthropic/v1	Only mode: messages (/messages is appended)
model	deepseek-flash	Only mode: messages
apiVersion	2023-06-01	Only mode: messages (ignored by DeepSeek, per its Anthropic-compatibility docs)
maxTokens	4096	Only mode: messages
maxUses	5	Only mode: messages, server-side searches per request

Billing and quotas

• The default search-api backend bills per successful request: 5 CNY / 1000 successful requests (about 0.005 CNY per search), charged to the DeepSeek account since 2026-11-01 00:00 Beijing time. Failed requests, empty results and timeouts are not billed. See the official Search & Fetch guide.
• A search does not consume chat-model tokens and does not touch the subscription used for chatting (e.g. OpenCode Go). Only the DeepSeek account's search usage is involved.
• Latency is about 0.5–1 s per query.
• mode: messages is different: one search costs a full auxiliary model turn (thousands of input tokens, 4–6 s), and server-tool searches are also billed per call as of the same date. Use it only as a fallback.

Security

• The DeepSeek account token is read from the DSH credential service per search (ctx.deepseekAccount.resolveToken) and sent only to the api.deepseek.com origin; the account service validates the origin.
• The plugin never writes the token to disk and never exposes it to the browser.
• No runtime dependencies, no install scripts: plain ESM using node: built-ins and the global fetch.

Requirements

• DeepSeek Harness 0.2.1-rc.5 or later (the ctx.web seam and deepseekAccount service it relies on).
• Signed in to DeepSeek in DSH, with network access to api.deepseek.com.
• Node.js 22+ (the DSH desktop app bundles its own runtime).

Verifying and troubleshooting

scripts/verify.mjs runs the provider exactly as the harness does, using the real account token from ~/.dsh/.credentials.yaml (the token never leaves the process):

NODE="/Applications/DeepSeek Harness.app/Contents/Resources/runtime/primary-runtime/dependencies/node/bin/node"
"$NODE" scripts/verify.mjs "query"              # official Search API, ~0.4–0.6 s, snippets included
"$NODE" scripts/verify.mjs "query" --messages   # legacy Messages backend, ~4–6 s

If search fails, the structured error names the cause: WEB_PROVIDER_CREDENTIAL_MISSING (not signed in and no API key), WEB_PROVIDER_ERROR (HTTP/network, with the endpoint that was used), or WEB_ABORTED (cancelled). If the account sign-in is rejected, sign in to DeepSeek again — the endpoint does not need changing.

Uninstall

dsh plugin --profile desktop remove dsh-deepseek-search-account

This removes the package and its bundle layer; the profile's own search configuration is not touched. For the manual install, delete the two rows described above.

License

MIT