Yichen Jiang
103dac1b3c
The Web search switch selects the DeepSeek Search and Fetch APIs together, and both reuse the DeepSeek chat key by default. Neither provider answered readiness, so a step without a usable key was offered web_search and web_fetch whose every call failed, and a conversation on another provider's model or behind a gateway sent the chat key to DeepSeek. dsh-web gives WebFetchProvider an optional ready(llmProvider) and adds ctx.web.fetchReady(), which asks only the provider fetch() selects, as searchReady() does for search. tool-web registers web_fetch with a step activation condition on fetchReady(), so a step that cannot fetch leaves the tool out of declarations, guidance, and PTC bindings and refuses a call to it as UNKNOWN_TOOL; without web_fetch the search guidance uses its search-only text. The shipped HTTP fetch provider has no ready() and stays admitted. web-deepseek-api applies the API-key rules of Messages search to both APIs: without a literal apiKey, an apiKeyEnv other than DEEPSEEK_API_KEY, or a baseURL other than the default, the chat key goes only to the DeepSeek routes and only to an API base on the official chat route's origin. The refusal sits in the key's only resolver, and both providers' ready() answers presence in the order an operation resolves credentials. The transport repeats these checks so the package still does not depend on dsh-web-search-deepseek.
12 KiB
description, kind
| description | kind |
|---|---|
| The web access service (ctx.web): how deployments and plugin authors search the web and fetch URLs through interchangeable providers, with one selection policy and error vocabulary. | package-reference |
@deepseek-ai/dsh-web
English | 中文
Summary
Use dsh-web to search the web or fetch a URL without tying callers to a specific vendor. It selects a usable backend for each operation and gives callers consistent cancellation, errors, and result limits. Choose it for plugins or tools that call ctx.web.search() or ctx.web.fetch(); the shipped dsh-tool-web tools load it for you. A search or fetch requires a configured, usable provider because this package does not make network requests on its own.
Table of Contents
- Use this package
- Understand the implementation
- Further Exploration
- Model Experience
- Known Limitations and Deferred Work
- Dev Note
Use this package
A composition that needs web access loads the dsh-web service and mounts at least one backend — a search provider and/or a fetch provider — and plugin or tool authors then call ctx.web.search() and ctx.web.fetch() directly. The service resolves the backend for each call, so callers never see provider ids unless they configured one.
When to choose it
Choose the service when a plugin or tool must search or fetch without hard-coding a vendor; a deployment that only uses the shipped web_search/web_fetch tools gets it for free through dsh-tool-web. You do not need it when the composition never reaches the web. The service adds no network access of its own: without at least one usable provider, every call fails with a structured WebError.
Minimal configuration
Load the service and let a single mounted backend auto-select, or pin a provider id with searchProvider/fetchProvider. The environment variables $DSH_WEB_SEARCH_PROVIDER and $DSH_WEB_FETCH_PROVIDER feed the same fields and are not a separate priority chain.
- name: '@deepseek-ai/dsh-web'
- name: '@deepseek-ai/dsh-web-search-exa'
- name: '@deepseek-ai/dsh-web-fetch-http'
| Field | Default | Meaning |
|---|---|---|
searchProvider |
(unset) | Pinned search provider id; unset auto-selects when exactly one is usable |
fetchProvider |
(unset) | Pinned fetch provider id; unset auto-selects when exactly one is usable |
The generated configuration catalog is the exhaustive source for every accepted field and its JSDoc.
Searching and fetching
search() runs one query and returns an optional provider answer plus a list of citeable sources; the service enforces request.maxResults by truncating sources[] and setting truncated. fetch() retrieves one URL and returns its final URL, status code, decoded body, and a truncation flag; a non-2xx response is a result, not an error.
// Search the web; sources[] is capped to maxResults:
const result = await ctx.web.search({ query: 'deepseek harness', maxResults: 8 })
// Fetch one URL; a non-2xx response is a result, not an error:
const page = await ctx.web.fetch({ url: 'https://example.com' })
Both calls accept an optional AbortSignal that is forwarded to the provider for cancellation. The normalized request and result shapes are the contract callers build on; the vocabulary section of the web subsystem reference describes them exhaustively.
Provider selection
Each call resolves its provider at execution time, and registration or load order never matters. A configured provider id wins when it is registered and usable; without a configured id, the service runs the single usable provider or fails clearly:
| Situation | Outcome |
|---|---|
| configured id registered and usable | runs that provider |
| configured id not registered | WEB_PROVIDER_CONFIGURED_MISSING |
| configured id registered but unavailable | WEB_PROVIDER_CONFIGURED_UNAVAILABLE |
| no id, exactly one registered usable provider | runs it |
| no id, no usable provider | WEB_PROVIDER_UNAVAILABLE |
| no id, multiple usable providers | WEB_PROVIDER_AMBIGUOUS |
A provider's availability is a cheap local check — for example whether its API key is present — and never makes network calls, so selection stays fast and deterministic.
Search and fetch readiness
searchReady(llmProvider?) and fetchReady(llmProvider?) ask the provider that search() or fetch() would select whether an operation made for a model request on that chat route has a credential to send; dsh-tool-web uses the answers to decide whether each model step admits web_search and web_fetch. A provider answers through its optional ready(llmProvider), a configuration-only check that must not make network calls; a provider without it counts as ready. When selection fails, the readiness method answers true, because the operation reports that failure as a structured WebError. Readiness never changes which provider is selected.
// false only when the selected provider cannot authenticate this route's operations:
const searchOffered = await ctx.web.searchReady('deepseek-official')
const fetchOffered = await ctx.web.fetchReady('deepseek-official')
Failures and recovery
Failures throw WebError with a stable, machine-routable code; the message adds detail such as the missing provider id or the ambiguous candidate set. Callers route on the code and decide how to degrade. To change which backend a call uses, reconfigure the pinned id, mount or unmount providers, or fix the provider's configuration so its availability check passes.
Understand the implementation
Implementation internals — click to expand
This section explains the design decisions behind the service; the observable behavior is fully covered in Use this package.
Design philosophy
The package is built on one deliberate separation:
- One seam, two independent operations. Search and fetch share no request schema and no business logic, but they share one service so provider selection, cancellation, errors, and product configuration have a single owner. The parallel
Search/Fetchmethod pairs are intentional. - Selection is never order-dependent. A capability either pins a provider id or auto-selects when exactly one usable provider is registered;
search()/fetch()resolve the provider at execution time. - The service owns the result bound.
maxResultsis enforced by the seam after the provider returns, so an over-returning provider can never leak more sources than the caller asked for.
Source map
| File | Role |
|---|---|
src/index.ts |
Plugin entry: the WebRuntime service, both provider registries, execution-time selection, and per-route readiness |
src/types.ts |
Vocabulary: request/result types, the closed WebFetchBody union, and the WebError taxonomy |
Data model
The request and result types define the normalized vocabulary callers build on — one Search pair and one Fetch pair — and the exhaustive fields and JSDoc live in src/types.ts and the web subsystem reference. Two deliberate choices shape them: WebFetchBody is a closed union (html | text) owned here, so adding a kind breaks compilation until every consumer handles it; WebError extends HarnessError with an open-string code, so consumers must tolerate provider-specific values. Source fields stay optional because not every provider returns all of them.
Selection flow
At call time the service resolves the provider — configured id first, then the unique usable provider — and throws the matching WebError when no clear winner exists. A search result then passes through capSources, which truncates sources[] to maxResults and flags truncated. searchReady() and fetchReady() resolve the same provider as the matching operation, answer true when resolution throws, and otherwise return the provider's ready() answer, or true when it has none. Registration is effect-based: providers register with the calling fiber and unregister when it disposes, and a duplicate id within a capability kind is rejected at registration.
Further Exploration
Read these pages when the package-level contract is not enough. They move from the shared vocabulary to the shipped backends, the model-facing tools, and the design rationale.
- Web subsystem — the exhaustive search/fetch requests and results, provider availability, and error codes.
- Web package map — the six-package family and each role.
- dsh-tool-web — the model-facing
web_searchandweb_fetchtools over this service. - dsh-web-fetch-http — the shipped anonymous HTTP(S) fetch backend.
- Generated configuration catalog — every accepted config field and its source declaration.
- Web capability seam decision — why search and fetch share one provider-selection service.
- Search readiness decision — why the seam answers per-route search and fetch readiness for
dsh-tool-web.
Model Experience
Indirectly, through dsh-tool-web, which renders the seam's normalized search results and fetch bodies to the model and does not admit web_search or web_fetch to model steps for which searchReady() or fetchReady() answers false, while this service contributes no prompt or schema.
KV Cache effect
No direct invalidation; the named consumer owns any request-prefix changes.
Known Limitations and Deferred Work
These limits define when the service is incomplete on its own. They are current package constraints.
- No observation surface — there is no provider-change event and no capability-status query;
searchReady()andfetchReady()report only the selected provider's credential readiness and answer true when selection fails. Availability is observable only by running a search or fetch and routing the thrown code, and the no-provider failure is the genericWEB_PROVIDER_UNAVAILABLEwith no per-provider reason enumeration (Agent Note). - Search requests carry only
queryandmaxResults— provider-neutral controls (recency, domain filters, regional hints, search depth) are deferred until the backends can honor them (seam Agent Note). WebFetchBodyhas nopdfarm — text-extractable PDF support is named deferred work; the closed union makes adding it a compile-enforced change across the web packages.- Provider-backed page extraction is out of scope of
fetch()— a Firecrawl/Tavily-styleweb_extractcapability is deferred rather than widening the fetch operation.
Dev Note
Working context for maintainers — click to expand
This Dev Note is working context for maintainers: open questions and undecided directions. It is explicitly non-authoritative — shipped behavior, limits, and rationale live in the sections above and the linked Agent Notes.
Future: observing provider state
No provider-change event or capability-status query exists; consumers observe availability only by executing a call and routing the thrown code. Restoring a small observation surface is possible if a consumer needs per-provider reasons, but the archived simplification note records why the earlier one was dropped.