[jacek4yang]

Fixes two Grok Build session breakers in the Responses frontend (/v1/responses). Closes #43.

Bug 1: hosted web_search declaration was dropped

Grok Build (backend search enabled, the default) declares {"type":"web_search"}; the proxy dropped it, so the model never had a search tool and Grok Build's server-side search was dead at this gateway.

Fix: the declaration converts to the same upstream web_search function the Anthropic frontend offers (websearch::upstream_function). When the model calls it, the gateway executes Cline /search/websearch with the active credential (search_web, same client/route/never-rotates-pool rules as the Anthropic server-tool loop) and streams:

• response.web_search_call.in_progress per call;
• response.output_item.done carrying the full web_search_call item (id, status, action.search with query + source URLs) — also recorded for terminal reconstruction, so Grok Build can replay it next turn;
• then a continuation generation runs with the grounding appended to the upstream conversation (assistant tool_calls + role=tool results). Pure server-tool rounds continue the same downstream stream; mixed rounds (server + client tools) end the stream after executing the searches so the client can run its own tools.

A per-request budget (from the declaration's max_uses, clamped to MAX_WEB_SEARCH_USES_PER_REQUEST, default 5) bounds searches; exceeded calls surface as max_uses_exceeded failed results, never silently truncated. Continuations reuse send_chat, so effective-429 failover and no-replay-after-commit apply per round; a failed continuation is terminal (semantic output already streamed).

Bug 2: replayed web_search_call input items hard-failed the session

Grok Build replays response.completed.output items as the next turn's input. A web_search_call item hit the unsupported item type 400 and permanently broke the session (every later request failed).

Fix: web_search_call items rebuild the assistant(tool_calls)/role=tool message pair from the item's own action.search.query and action.sources URLs — nothing invented (a failed call replays as its failure; its sources list is empty). The item's own id is reused, so the tool_calls ↔ tool_call_id chain stays valid.

Unchanged

• Other hosted backend tools (x_search, code_interpreter, MCP, …) remain dropped, never forwarded, never faked.
• Reasoning exposure gating, GLM policy, shadow-store epochs, prefix telemetry: untouched paths.
• No prompt/query/result content in logs — search counts only (web_searches in the request summary).

Tests

New coverage: hosted web_search conversion (+ max_uses clamping), web_search_call replay rebuild (unit + integration), hosted-tool drop for non-web_search types, and a full streaming server-round integration test (model calls web_search → gateway executes → continuation generation runs with the grounding → client sees web_search_call items and the grounded answer, never a function_call for the server tool). Existing tool/replay/429 tests unchanged and green.

Gates: fmt OK, check OK, clippy -D warnings OK, 309 tests debug OK, 309 tests release OK.

Docs

README (web_search support in the Grok Build section) and docs/DEVELOPMENT_STATE.md updated.