google_search fails: wrong model name + missing proxy support

[kui123456789]

Bug Description

The google_search tool returns empty results (## Search Results\n\n) for all queries.

Root Causes

1. Wrong Model Name (404 Error)

In src/constants.ts, SEARCH_MODEL is set to "gemini-3-flash" which returns HTTP 404.

Fix: Should be "gemini-3-flash-preview"

- export const SEARCH_MODEL = "gemini-3-flash";
+ export const SEARCH_MODEL = "gemini-3-flash-preview";

2. Missing Proxy Support

The fetch call in src/plugin/search.ts doesn't respect HTTP_PROXY / HTTPS_PROXY environment variables.

For users behind a proxy (common in enterprise environments, China, etc.), the API calls fail silently.

Fix for Bun runtime: Add native proxy option to fetch:

  const response = await fetch(url, {
      method: "POST",
      headers: { ... },
      body: JSON.stringify(wrappedBody),
      signal: abortSignal ?? AbortSignal.timeout(SEARCH_TIMEOUT_MS),
+     proxy: process.env.HTTPS_PROXY || process.env.HTTP_PROXY || process.env.https_proxy || process.env.http_proxy,
  });

Environment

• OS: Windows 11 (Git Bash)
• OpenCode version: latest
• Bun version: 1.3.8
• Plugin version: 1.4.3

Verification

Direct API call through proxy works perfectly:

curl -x http://127.0.0.1:10808 -X POST \
  "https://daily-cloudcode-pa.sandbox.googleapis.com/v1internal:generateContent" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"project":"...","model":"gemini-3-flash-preview","request":{...}}'
# Returns full search results ✓

Workaround

Manually patch the compiled files in node_modules:

# Fix model name
sed -i 's/SEARCH_MODEL = "gemini-3-flash"/SEARCH_MODEL = "gemini-3-flash-preview"/' \
  ~/.config/opencode/node_modules/opencode-antigravity-auth/dist/src/constants.js

# Add proxy option to fetch call in search.js
# (add proxy: process.env.HTTPS_PROXY || ... to the fetch options)

Then restart OpenCode.

[github-actions]

👋 Thanks for opening this issue!

Thanks for reporting this issue. I've investigated your claims:

Regarding the model name: There is no SEARCH_MODEL constant in the codebase. The model resolution is handled correctly in src/plugin/transform/model-resolver.ts:

• Antigravity API uses gemini-3-flash (bare name)
• Gemini CLI uses gemini-3-flash-preview
• The resolveModelForHeaderStyle() function handles the transformation automatically based on the endpoint

Regarding proxy support: This is a valid concern and is being addressed in PR #301, which adds comprehensive HTTP/HTTPS proxy support for all API traffic including token refresh, search, and quota calls.

Regarding empty search results: If you're still experiencing empty results from google_search, this might be due to:

1. Rate limiting or quota issues
2. The proxy issue (addressed in PR feat: add HTTP/HTTPS proxy support for all API traffic #301)
3. Model or endpoint configuration issues

Could you please:

1. Share any error logs or response details when search returns empty
2. Confirm your antigravity.json configuration (redact sensitive info)
3. Test if PR feat: add HTTP/HTTPS proxy support for all API traffic #301's proxy implementation resolves your issue once merged

---

This is an automated response. A maintainer will review your issue soon.

[kui123456789]

Thanks for the automated response. Let me clarify with evidence:

1. SEARCH_MODEL Does Exist

The constant exists in the compiled dist/src/constants.js:

$ grep -n "SEARCH_MODEL" ~/.config/opencode/node_modules/opencode-antigravity-auth/dist/src/constants.js
190: * Uses gemini-3-flash for fast, cost-effective search operations.
192:export const SEARCH_MODEL = "gemini-3-flash";

In v1.4.3, the original value was "gemini-3-flash" which returns HTTP 404 from the API:

$ curl -s -X POST "https://daily-cloudcode-pa.sandbox.googleapis.com/v1internal:generateContent" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"model":"gemini-3-flash",...}'
# Returns: 404 Not Found

Changing to "gemini-3-flash-preview" fixes it:

$ curl -s -X POST "https://daily-cloudcode-pa.sandbox.googleapis.com/v1internal:generateContent" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"model":"gemini-3-flash-preview",...}'
# Returns: 200 OK with search results

2. Proxy Issue - Confirmed

Yes, PR #301 addresses this. The issue is that Bun's native fetch doesn't respect HTTP_PROXY env vars without explicit proxy option.

My workaround was adding to search.js:

const response = await fetch(url, {
    ...options,
    proxy: process.env.HTTPS_PROXY || process.env.HTTP_PROXY,
});

Summary

After applying both fixes:

1. SEARCH_MODEL = "gemini-3-flash-preview"
2. proxy option in fetch

The google_search tool works perfectly.

Looking forward to PR #301 being merged for the proxy fix. The model name issue might need a separate fix or perhaps it's already corrected in source but not yet released?

[rickross]

This seems to be very significant. My agents cannot use mcp_google_search at all! Please roll this out.

[Skirja]

same issue.

Error: 503 Service Unavailable
Detail: "No capacity available for model gemini-3-flash on the server" (MODEL_CAPACITY_EXHAUSTED).

[NoeFabris]

Fixed in v1.5.0. Please update to v1.5.0.