feat: web search for Claude models via GPT /responses backend#273
Conversation
…kend GitHub Copilot rejects Anthropic's native server-side web_search tool on both its /v1/messages and /chat/completions endpoints, and Claude models cannot use the /responses endpoint at all — so Copilot offers no native web search path for Claude. This fulfills the web_search tool inside the proxy: when a Claude (Messages API) request carries an Anthropic web_search server tool, it is swapped for an equivalent function tool and run through an agentic loop against Copilot's /v1/messages. Each search the model requests is answered by Copilot's own GPT /responses web_search tool, and the results are reconstructed into native Anthropic server_tool_use + web_search_tool_result blocks (plus a final cited answer). Both streaming and non-streaming are supported (streaming replays the reconstructed response as a synthetic SSE sequence). Gated by the new useMessagesApiWebSearch config flag (default on); the backend GPT model is configurable via webSearchBackendModel (default gpt-5-mini). When disabled, the server tool is stripped so requests no longer 400 upstream. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Install mcp fetch or other search tools. @mahoshojoHCG |
|
@mahoshojoHCG In Claude Code, when using the WebSearch Tool to invoke a sub-model to initiate a web search, can tools that only use the websearch type be switched to the responses API or message API? Some providers' message APIs also support websearch. There are no plans to support mixing websearch-type tools with other tools, Can you adjust the code? For the switch, configuring a messageApiWebSearchModel should be sufficient. |
Per review feedback, replace the Claude agentic loop with a model switch. A Claude /v1/messages request whose only tool is web_search is now switched to a configured Responses-capable GPT model (messageApiWebSearchModel, default gpt-5-mini), which runs Copilot's native /responses web_search in a single call. The result is reconstructed into native Anthropic server_tool_use + web_search_tool_result + answer blocks (streaming replays them as a synthetic SSE sequence). Mixing web_search with other tools is intentionally unsupported: such requests have the web_search tool stripped and proceed normally. The feature is driven entirely by messageApiWebSearchModel (unset = disabled); the previous useMessagesApiWebSearch / webSearchBackendModel flags are removed. This avoids multiple Claude round-trips (cheaper, simpler) while keeping the client-facing behavior identical. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
messageApiWebSearchModel now also accepts a `provider/model` alias. When it points to a provider whose message API supports web search natively (e.g. an `anthropic`-type provider), the web_search tool is passed straight through to that provider — no translation, native results. A plain Copilot GPT model still goes through the /responses web_search path. Both return native Anthropic web-search blocks, so behavior is consistent across targets. The routing decision is extracted into a pure, unit-tested resolveWebSearchRoute helper (provider / responses / strip). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
@caozhiyuan Done — pushed two follow-up commits. 1. Both switch targets supported, via the one
The routing decision is a small pure helper ( 2. Result fidelity — kept the structured blocks. Both paths return native Mixing stays unsupported: a web_search request with any other tool just has typecheck / lint / full suite (287) all pass. |
2 participants
Problem
GitHub Copilot offers no native web search path for Claude models:
/v1/messagesendpoint rejects Anthropic's server-side web search tool:400 — "The use of the web search tool is not supported." (unsupported_value)— for everyanthropic-betavariant and bothweb_search_20250305/web_search_20260209tool types./responsesendpoint at all (supported_endpointsis only["/v1/messages","/chat/completions"]; a direct call returnsunsupported_api_for_model)./responses, and only for GPT models (confirmed working).So a
/v1/messagesrequest that carries aweb_searchserver tool currently 400s upstream.Solution
Fulfill the web search inside the proxy. When a Claude (Messages API) request carries an Anthropic
web_searchserver tool:/v1/messages; eachweb_searchcall the model makes is answered by Copilot's own GPT/responsesweb_search tool (no external API key, uses existing Copilot quota).server_tool_use+web_search_tool_resultblocks plus the final cited text, so clients render it exactly like real Anthropic web search.Both streaming and non-streaming are supported (streaming replays the reconstructed response as a synthetic Anthropic SSE sequence).
Config
useMessagesApiWebSearch(defaulttrue) — enable the feature. When off, the server tool is stripped so requests no longer 400.webSearchBackendModel(default"gpt-5-mini") — GPT model used as the search backend.Notes / limitations
web_searchcall alongside a real client tool in the same response) terminate the loop and hand control back to the client.Testing
tests/web-search-fulfill.test.ts(detection/stripping, fulfillment loop, graceful backend-error block, synthetic stream event sequence).bun run typecheck,bun run lint, and fullbun test(281 pass) all green.claude-sonnet-4.5invokedweb_search3×, each fulfilled viagpt-5-mini/responses, producing nativeserver_tool_use/web_search_tool_resultblocks, a cited answer,stop_reason: end_turn, andusage.server_tool_use.web_search_requests: 3.🤖 Generated with Claude Code