Repository navigation
[awf] api-proxy/gemini: key not injected — Gemini CLI bypasses api-proxy, hits Google directly #2807
Description
Activity
👋 Hello! I noticed this issue might be related to several existing issues:
- [awf] api-proxy: Gemini API key not injected — agent calls googleapis directly #2726 (closed) — [awf] api-proxy: Gemini API key not injected — agent calls googleapis directly
- [awf] api-proxy: Gemini API key not injected — agent bypasses proxy on port 10003 #2704 (closed) — [awf] api-proxy: Gemini API key not injected — agent bypasses proxy on port 10003
- [awf] api-proxy: Gemini requests bypass api-proxy (port 10003 intercept broken) — API_KEY_INVALID errors #2666 (closed) — [awf] api-proxy: Gemini requests bypass api-proxy (port 10003 intercept broken) — API_KEY_INVALID errors
- [awf] api-proxy: Gemini engine bypasses proxy — GEMINI_API_KEY not injected, requests reach Google directly #2343 (closed) — [awf] api-proxy: Gemini engine bypasses proxy — GEMINI_API_KEY not injected, requests reach Google directly
- [awf] api-proxy: Gemini API key forwarding fails — key rejected despite being valid #2291 (closed) — [awf] api-proxy: Gemini API key forwarding fails — key rejected despite being valid
- [awf] API Proxy: Gemini key injection failing — API_KEY_INVALID despite valid key #2173 (closed) — [awf] API Proxy: Gemini key injection failing — API_KEY_INVALID despite valid key
All of these describe the same core issue: the Gemini CLI bypassing the api-proxy and sending requests directly to Google, resulting in
API_KEY_INVALIDerrors. This new issue adds the specific hypothesis thathost.docker.internalmay not resolve inside the agent container.If any of these closed issues addressed your concern or if the root cause/fix is the same, please consider closing this as a duplicate. Otherwise, feel free to clarify how your issue differs (e.g., a regression or a new environment where the fix doesn't apply)!
This is an automated message from the issue duplication detector.
Generated by Issue Duplication Detector for issue #2807 · ● 217.9K · ◷
Resolved. Gemini CLI now routes through the api-proxy at
http://172.30.0.30:10003usingGEMINI_API_BASE_URLandGOOGLE_GEMINI_BASE_URLenvironment variables (seesrc/services/api-proxy-service.ts:275-287). Key injection is handled bycontainers/api-proxy/providers/gemini.js.
Problem
Gemini engine runs fail with
400 API_KEY_INVALIDeven when using a valid Google AI Studio API key. Squid access logs confirm requests originate from the agent container (172.30.0.20) directly, not from the api-proxy (172.30.0.30). The api-proxy intercept on port 10003 is not functioning — `GEMINI_API_BASE_URL=(host.docker.internal/redacted) is set in the agent environment but the Gemini CLI bypasses it.Context
Original issue: github/gh-aw#29417
Root Cause
host.docker.internalmay not resolve inside the agent container in all environments, causing the Gemini CLI to fall back to direct requests togenerativelanguage.googleapis.comwithout the api-proxy injecting the key. Alternatively, the Gemini CLI ignores theGEMINI_API_BASE_URLenv var in certain versions.Proposed Solution
In
src/services/api-proxy-service.tsandsrc/docker-manager.ts, setGEMINI_API_BASE_URLto the api-proxy container's fixed IP ((172.30.0.30/redacted) instead ofhost.docker.internal. Add an integration test verifying Gemini traffic routes through the api-proxy (source IP is172.30.0.30in Squid logs, not172.30.0.20`).