dsh-crw

fastCRW-backed web_search and web_fetch for DeepSeek Harness.

The harness treats web access as a capability seam (ctx.web) with pluggable providers. This package registers two of them under the id crw, so the model keeps the exact same web_search / web_fetch tools and only the backend changes.

Why

The base composition ships web_search on, and web_fetch off. Its own config says why:

Fetch stays disabled and no fetch provider is mounted: that provider defers SSRF protection and the model would choose the request target.

That is a fair call. The in-box HTTP fetch provider runs fetch() in the harness process, and its README names the gap outright: no blocking of private, loopback, link-local or multicast destinations, no DNS-resolve-then-validate. A model that picks http://169.254.169.254/ gets your cloud metadata endpoint.

fastCRW resolves and validates the target server-side and refuses private ranges, so the harness process never opens a connection the model chose. That is what makes turning fetch on reasonable, and it is why fetch: true lives in this bundle rather than upstream.

The second reason is that plain HTTP is not how the open web reads any more:

URL	in-box HTTP provider	dsh-crw
producthunt.com	403, a Cloudflare interstitial	200, 27,559 chars of markdown
zillow.com	403, "Access to this page has been denied"	200, real page content
news.ycombinator.com	200, raw HTML for the tool to convert	200, markdown from the source

fastCRW escalates per page from plain HTTP through its browser tiers only when a page needs one, so the common case stays fast and the hard case still answers. Not every wall falls: an interactive Cloudflare challenge still wins sometimes.

Install

export CRW_API_KEY=...          # fastcrw.com, 500 free credits, no card
dsh plugin --profile default add dsh-crw
dsh --profile default

Get a key at fastcrw.com. The free tier is a one-time 500 credits and needs no card; one page is one credit.

To install straight from this repo instead, pin a commit and allow the build. A git install fetches sources, so pnpm has to run this package's prepare script to produce lib/, and pnpm 10+ requires you to say so explicitly:

# $DSH_HOME/profiles/<name>/pnpm-workspace.yaml
allowBuilds:
  dsh-crw: true

dsh plugin --profile default add github:us/dsh-crw#<sha>

That allowance runs this package's build on your machine at install time, so pin the commit rather than tracking the branch.

Verify the layer before booting:

dsh --profile default --dump-config    # shows a "# == dsh-crw" layer

What the bundle does

The shipped cordis.patch.yml applies three changes over @deepseek-ai/dsh-base:

• points web.searchProvider and web.fetchProvider at crw
• turns on tool-web's fetch and raises both timeouts, since a page behind a JS or anti-bot wall escalates through browser tiers before it answers
• mounts this plugin, reading $CRW_API_KEY

The base layer's own DeepSeek search provider stays mounted, so switching search back is one line in your profile's cordis.patch.yml:

- id: web
  config:
    searchProvider: deepseek-official
    fetchProvider: crw

Config

Key	Default	Meaning
apiKey	$CRW_API_KEY	fastCRW API key. Required against the cloud endpoint; empty is fine when self-hosting.
baseURL	https://fastcrw.com/api	Endpoint base; /v1/search and /v1/scrape are appended.
search	true	Register the search provider.
fetch	true	Register the fetch provider.
answer	false	Ask fastCRW to synthesize an answer over the results, returned as the seam's content. Adds an LLM call per search.
maxBodyChars	100000	Cap on returned page characters, matching the in-box provider.

- id: crw
  name: dsh-crw
  config:
    apiKey: !!js process.env.CRW_API_KEY
    answer: false

Self-hosting

fastCRW is AGPL-3.0 and the engine is one binary, so the whole web layer can stay on your own machine:

crw serve

- id: crw
  name: dsh-crw
  config:
    baseURL: http://127.0.0.1:3002

A self-hosted endpoint takes no credential, so apiKey may be empty. The provider only insists on a key when it is pointed at the cloud endpoint.

Mapping notes

• Search. POST /v1/search. Each result maps to a WebSearchSource: url, title, and snippet from the result's snippet or, failing that, its description. A result without a URL is dropped rather than given an invented one. With answer on, the synthesized answer becomes the seam's content. The seam owns the final maxResults truncation.
• Fetch. POST /v1/scrape with formats: ['markdown']. The body kind is text, not html: fastCRW already returns markdown, and html would send it back through the tool's turndown pass to convert converted output.
• Non-2xx targets. fastCRW reports a 404 target as success: false while still returning what it got. The seam's contract is the other way round, so an envelope carrying a status code is always a result. A WebError is reserved for failing to retrieve the resource at all: a rejected URL, a bad key, an exhausted balance.
• No renderer hint is sent. fastCRW's escalation ladder decides per page whether plain HTTP is enough.

Development

bun install
bun test
bun run build

License

MIT.