[FEATURE]: Optional hosted web search in the one-shot inference API #58174

Feature hasn't been suggested before.

• I searched existing issues for one-shot generation, hosted web search, and the generate endpoint. Related requests are linked below; I did not find this exact local HTTP API proposal.

Describe the enhancement you want to request

Add optional hosted web search to the one-shot inference API, so an external application can submit one inference request, let the model search when necessary, and receive the final answer with source citations.

For example, an application should be able to ask “What is the weather in Chicago today?” without implementing its own search → inject results → generate loop, or creating, polling, and deleting an OpenCode session.

Current behavior

Tested with OpenCode Desktop/CLI 2.1.3:

• POST /api/experimental/generate works for text inference using a connected OpenAI/ChatGPT account and openai/gpt-6-luna.
• Its request schema accepts only prompt and model, with no tool configuration.
• POST /api/websearch is a separate search operation. An explicit exa search returned results in this setup.
• OpenCode Console documents a hosted search provider, but connecting a search backend does not expose search in the current one-shot generation contract.

The gap is coordinating inference and search behind a single request, rather than adding another standalone search endpoint.

Proposed experience

Illustrative proposal, not currently supported syntax; extending the existing endpoint or adding a separate one-shot tool-enabled endpoint would both work:

{
  "prompt": "What is the weather in Chicago today? Include sources and the observation time.",
  "model": {
    "providerID": "openai",
    "id": "gpt-6-luna"
  },
  "tools": [{ "type": "web_search" }],
  "toolChoice": "auto"
}

The server would use the configured search provider, including OpenCode Console hosted search where enabled. It could execute the search loop itself or use an upstream model's hosted search capability. The caller would receive the completed answer and citations without managing intermediate tool calls.

Why this is useful

This would make OpenCode's existing provider connections useful to external apps, browser integrations such as Inference Bridge, and small utilities that need occasional current information without embedding a full coding-agent session.

Acceptance criteria

• Search is explicit and optional; existing text-only requests retain their behavior.
• The model can choose whether to search, with bounded search calls and existing provider permissions/budget enforcement.
• The caller receives the final answer plus structured source URLs/citations and an indication of searches performed.
• Search failures, missing authorization, unsupported models, and spending limits are surfaced clearly.
• Cancellation stops the associated generation/search work.
• No persistent user-facing session or access to coding tools is required.

Comparisons and references

• Ollama hosted search (requester’s working comparison): With an Ollama API key, including on the free tier, hosted search works without the application implementing a client-side search loop. The requested OpenCode experience is the same: enable search, submit an inference request, and receive the grounded answer. See also Ollama’s hosted-search documentation.
• OpenAI's single-request interface: Responses API web search lets a caller enable tools: [{ "type": "web_search" }] and receive a searched answer when the model uses it.
• OpenCode Console hosted web search
• Related: [FEATURE]: Ephemeral, tool-capable sessions for the V2 plugin API #58060 requests ephemeral tool-capable sessions and mentions a one-shot helper. This issue focuses on the external HTTP inference contract and optional web search.
• Related: Go /v1/responses: the oa-compat path turns tool_search / web_search into nameless function tools, so deferred MCP tools and web search are unreachable #57397 reports route-dependent handling of web_search in the hosted Go Responses gateway; it is distinct from exposing search in the local one-shot generate API.