@creait/dsh-web-fetch

The backend DeepSeek Harness's web_fetch tool is missing.

Why this exists

@deepseek-ai/dsh-tool-web already implements the entire model-facing web_fetch tool: argument schema, validation, timeout budget, HTML→markdown rendering, output cap with a truncation footer, and the fetch card the web UI draws. What dsh does not ship is a provider for the fetch half of the ctx.web seam, so dsh-base mounts the tool with fetch: false and the capability is inert.

The bundle 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 statement about ownership, not a verdict that fetching is unsafe. The provider is where the guards belong, and dsh's own error taxonomy names the set it expects one to carry — "invalid or blocked URLs, redirects, size and timeout limits, and unsupported content types." This package is that provider.

Install

▼
bash
复制代码
dsh plugin --profile web add @creait/dsh-web-fetch

That mounts the provider and registers it with ctx.web. It does not turn on web_fetch: the tool belongs to tool-web, which dsh-base mounts with fetch: false because no provider shipped. Turn it on in your profile patch:

▼
yaml
复制代码
- id: tool-web
  config:
    fetch: true

Restart dsh — the boot manifest is assembled at startup.

Nothing has to name the provider. dsh ships none of its own, so local is the only fetch provider registered and the seam selects it on that basis. Name it explicitly only if something else registers a second one, which is when the seam refuses to guess:

▼
yaml
复制代码
- id: web
  config:
    fetchProvider: local

Under dsh web, fetch: true on that row is not enough on its own: dsh-web-app disables the host-plane tool-web, and each agent preset mounts its own copy with fetch: false. Revive the host row instead of forking every preset, and give it only the half the presets do not own — the tool registry throws on a duplicate tool name:

▼
yaml
复制代码
- id: tool-web
  disabled: false
  config:
    search: false
    fetch: true

If you installed this before it shipped a bundle patch, your profile patch inserts the row by hand. Drop that - insert: block: insert appends unconditionally, and the second row would register the same provider id twice, which ctx.web rejects with WEB_DUPLICATE_PROVIDER.

Guards

Guard	Behaviour
Scheme	http: and https: only
Credentials	URLs carrying user:pass@ are refused outright
Address	Every resolved address is checked against loopback, RFC 1918, CGNAT, link-local (incl. 169.254.169.254), multicast, reserved and IPv6 ULA/link-local ranges
IPv4-in-IPv6	::ffff: and 64:ff9b:: addresses are unwrapped and judged as the IPv4 they carry
Multi-homed names	A hostname resolving to any blocked address is refused — which address fetch would pick is not ours to decide
Redirects	Followed by hand, capped at maxRedirects, and every hop is re-resolved and re-checked
Size	Body is streamed and cut at maxBytes; the result reports truncated
Content type	HTML and text-ish types only; anything else is an error rather than mojibake
Cancellation	The seam's AbortSignal is honoured before and during the request

A non-2xx response is a result, not an error — the status code is part of the resource's state, and the seam's contract says so.

Config

Key	Default	Meaning
maxBytes	5242880	Ceiling on the buffered body
maxRedirects	5	Hops followed before refusing
allowHosts	[]	Exact hostnames allowed to resolve into blocked ranges
blockHosts	[]	Exact hostnames always refused
userAgent	deepseek-harness/0.0.1 (web fetch)	Sent on every request

allowHosts is the deliberate escape hatch for reaching something on your own network — a LAN wiki, an internal dashboard:

▼
yaml
复制代码
- id: web-fetch
  config:
    allowHosts: ['wiki.lan', 'dashboard.lan']

Known limitation: DNS rebinding

The window between resolving a hostname and the connection being made is not closed. A name whose DNS flips to a private address inside that window would be checked as public and connected as private, because global fetch re-resolves on its own and offers no pinned-address dispatcher.

Closing it needs a custom undici dispatcher whose connect re-validates the resolved peer, which is a real dependency this package does not currently take. On a trusted workstation network the residual risk is small — but it is not zero, and allowHosts/blockHosts are the levers if your threat model needs them tighter.

Tests

▼
bash
复制代码
node --test test/*.test.js

Covers the address classifier range by range (including the IPv4-in-IPv6 unwrap paths, where a subtle parser bug would have let 64:ff9b::169.254.169.254 through) and drives the provider against a stubbed fetch for redirect re-validation, hop caps, allow/block lists, content-type handling and cancellation.

Licence

MIT