Search plugin

First-party search capability with one URL pipeline and two model-facing entries:

• search — merged entry, visible when the current main model has no usable provider-native search. mode:"model" runs one isolated completion against the model referenced by defaultCapabilities.search with native search enabled; mode:"url" fetches the given URLs and summarizes them with the current main model.
• search_url — URL-only entry, visible when the main model does have usable native search. It reads the URLs the model provides; discovery stays with the built-in search.

Only one entry is visible at a time. The host capability projection selects it from getCapabilityFacts("search"); the plugin does not branch on vendor or model names.

Configuration

settings.json may point the merged entry's model-search target at any registered model that declares both text and search:

{
  "defaultCapabilities": {
    "search": "search-provider/search-model"
  }
}

The target's provider search mode determines the actual vendor protocol. If no target is configured, mode:"model" returns search_target_unavailable; URL mode still works.

Bounds

• 32 URLs per call, 2048 characters per URL, 2000 characters per query.
• 8 MiB raw response per page, 128 MiB per call, 4 MiB extracted text per page.
• 512 KiB summary input, 16 KiB summary output, 8192 output tokens.
• 8 concurrent fetches, 20 s per URL, 120 s total fetch/model deadlines.
• System-level disk page cache (2 GiB / 5,000 entries / 5 MiB each; 14 days base, extended by hits) and a 32-entry / 10-minute in-memory summary cache. Pages contain content only; no URLs, headers, cookies or validators.

There are no domain, scope, or private-network restrictions in the plugin itself. Deployments that need them should install a policy plugin before the tool call. No provider credentials, cookies, or browser profiles are sent to target URLs; pages with no readable server-rendered text fail with empty_content.

Status

mechanism-verified: deterministic unit tests cover registration, projection metadata, both modes, input validation, partial failures, invalid summaries, disk-cache persistence/retention/eviction, size/type failures, redirects, and cancellation. Real-task benchmarking (P4 in docs/design/搜索工具设计.md) remains a separate, authorized step.