Native search provider audit
Checked 2026-09-24 against the configured endpoint presets and public vendor documentation. No paid inference or live search request was made. This is an implementation audit, not a claim that every model offered by a vendor supports every tool. An enabled checkbox still requires a model accepting that tool.
| Connection | Finding | Application behavior in this change |
|---|---|---|
| DeepSeek official Chat Completions | Official Harness uses the same key and /anthropic/v1/messages with web_search_20250305. | One existing entry; search opt-in routes this model's request internally. Search off keeps Completions. |
| xAI official Chat Completions | Official search documentation specifies Responses and web_search. | Same opt-in routing to same-origin Responses; no second preset. |
| OpenAI official legacy Chat Completions | The existing application supports search through Responses; the current OpenAI preset already uses it. | Legacy Completions entries can opt in without replacing their saved configuration. Current Responses and Codex accounts stay unchanged. |
| Anthropic / existing Responses connections | Existing adapters attach tools and preserve search blocks and citations. | Existing behavior retained; protocol compatibility alone does not prove model or gateway support. |
| Google Gemini | Google documents a Google Search tool and grounding/search-suggestion results. The current Google adapter has no webSearch request/response handling. | Still not integrated for this format; the hint says so rather than claiming the vendor lacks search. Requires a distinct adapter contract. |
| Qwen / DashScope | Completions uses enable_search; newer models also have Responses or Anthropic paths, with model and workspace constraints. | No blind route switch for all configured models; those formats require model-specific capability and result handling. |
| Zhipu / Z.AI | Zhipu documents its own web_search object and a separate search API. | No substitution of OpenAI/Anthropic tool schemas; format-specific integration remains outstanding. |
| Moonshot / Kimi | $web_search is a builtin_function with a tool-call/result loop. | Not treated as a provider-hosted Anthropic/Responses block; needs its own tool lifecycle integration. |
| Groq | Compound models were deprecated on September 21; current docs point to browser search for particular GPT OSS models. | No blanket capability inferred from the Groq preset or old Compound documentation. |
| MiniMax / Kimi Coding Anthropic routes | The preset already selects Anthropic Messages. Support for the exact hosted tool is not established by the protocol alone. | Existing manual opt-in retained; not newly certified by this audit. |
| Mistral, Xiaomi, Together, Fireworks, SiliconFlow, Volcengine | This audit did not establish a drop-in compatible hosted-search contract for the configured endpoints. Mistral's attempted page returned 404; Xiaomi docs transport failed. | No inferred support or automatic credential rerouting. Further vendor-specific integration remains open. |
Sources
- DeepSeek official Harness search provider
- xAI web search
- Google Search grounding
- Model Studio web search
- Zhipu web search
- Kimi built-in web search
- Groq Compound status
- Existing application contracts:
packages/shared/src/native-web-search.ts,packages/agent-runtime/src/provider-binding.ts, and the pinned pi-ai adapters.
Exact first-party origins and paths are required for request routing. Vendor names and model labels are not authority to send a relay key to an official service. Unknown connections retain their configured destination. No new discovery probes, hidden inference charges, or persistent transport migrations are introduced. Full protocol probing from issue #907 remains separate work.