feat(web-search): intercept web search tool calls and execute via SearXNG

[jsboige]

Problem

When providers (Z.AI, GLM, OpenRouter models) emit WebSearch/WebFetch tool calls, they fail because the proxy doesn't implement server-side web search. GLM models also emit <searchWeb>query</searchWeb> tags in their text output, which go unhandled. Additionally, Claude Code's built-in WebSearch tool fires sub-agent requests that get forwarded to the model provider instead of being executed locally.

Changes

New file: packages/cli/src/handlers/shared/web-search-executor.ts

A stream-safe SearXNG client that:

• Queries a local SearXNG instance (SEARXNG_URL env var, defaults to http://search.myia.io)
• Races SearXNG against a configurable deadline (default 3s)
• Returns formatted results as text, never throws
• Extracts search queries from tool call arguments JSON

packages/cli/src/proxy-server.ts

Add interceptWebTools() function that intercepts sub-agent web search/fetch requests BEFORE handler selection. Detects the standard Claude Code WebSearch sub-agent pattern ("Perform a web search for the query: X") and returns SearXNG results as a streaming or non-streaming Anthropic-format response.

packages/cli/src/handlers/shared/stream-parsers/openai-sse.ts

Three enhancements:

1. WebSearch tool suppression -- When a structured WebSearch/WebFetch tool call is detected, set suppressed = true and buffer arguments. At finalization, execute SearXNG and inject results as a text block.

2. GLM <searchWeb> tag detection -- Detect <searchWeb>query</searchWeb> tags in accumulated text and execute SearXNG for each query.

3. Empty response handling -- When no text or tool blocks were started, inject a synthetic error message so downstream doesn't hang.

Providers affected

• Z.AI (GLM models)
• GLM direct
• OpenRouter models that emit web search tool calls
• Any provider where Claude Code's WebSearch sub-agent is routed through claudish

How to test

1. Sub-agent interception:

   claudish --model zai@glm-4.7 "search the web for recent AI news"
   # Verify: SearXNG results appear in response instead of tool call error

2. GLM tags:

   claudish --model glm-4.7 --debug "search for..."
   # Check debug log for [Streaming] Found N <searchWeb> tag(s) in text

3. SearXNG URL configuration:

   SEARXNG_URL=http://localhost:8080 claudish --model glm-4.7 "search for..."

Notes

• SearXNG is expected to be running and accessible. If SearXNG is unavailable, a timeout/failure message is injected instead of crashing.
• The SEARXNG_URL env var should be documented in the main README as an optional configuration.