[BUG] Web Search fails

[Justxd22]

Bug Description

Do you have any idea why search always fail?? always 0 results

Expected Behavior
Search works

Actual Behavior
0 results

Environment

• Claude Code version: latest
• antigravity-claude-proxy version: latest
• Operating System: Arch linux
• Node.js version: v24.10.0

Checklist
Please confirm:

• I've updated antigravity-claude-proxy to the latest version
• The bug exists in the latest version
• I'm using this proxy for personal development only
• I have an active Google Cloud project with Antigravity enabled
• This issue is not related to attempting commercial use or TOS violations
• I've reviewed the README troubleshooting section

[coderabbitai]

⚠️ Possible Duplicate Issue(s)

• WebSearch tool - 0 results #27

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[Justxd22]

@badrisnarayanan i see it's a known issue... what about we fix it? Digging into Gemini-CLI code base look what i found:

Based on the code in packages/core/src/tools/web-search.ts, the Gemini CLI performs web searches through the Gemini API, which has a built-in Google Search grounding capability. It does not use an internal MCP server or a separate search API (like the custom search JSON API) for the core functionality.

Here is the breakdown of how it works:

1. Tool Execution: The WebSearchTool executes a method that calls geminiClient.generateContent
2. Special Model: It targets a specific model behavior configured as web-search
3. Grounding: The response from the Gemini API includes groundingMetadata. This metadata contains:
   - groundingChunks: The actual search result snippets (titles, URIs).
   - groundingSupports: Information on which parts of the generated text rely on which search result.
4. Processing: The tool manually processes this metadata to:
   - Insert citation markers (e.g., [1]) into the text at the correct positions.
   - Append a "Sources" list at the end of the response.

So effectively, it asks the Gemini model "Please answer this query using Google Search," and the model itself performs the retrieval-augmented generation (RAG) on Google's side, returning both the answer and the source data.

Relevant Code Snippet

  const response = await geminiClient.generateContent(
    { model: 'web-search' },
    [{ role: 'user', parts: [{ text: this.params.query }] }],
    signal,
  );

If we implemented this it would be 10x powerful than default Claude tool lol, it comes free with the pro plan so why not

[Justxd22]

i got it working, if anyone needs the code, or if u want a PR im down ... no 3rd party apis, completely free using gemini 2.5 flash with google search flag wrapped in mcp that calls the proxy (since the proxy already have oath access )

[badrisnarayanan]

Send me the pr I can take a look