feat: support custom base_url for Gemini native adapter

[li0near]

Summary

When provider: "gemini" is explicitly configured, use the native Gemini REST adapter (models/{model}:generateContent) for any base_url that doesn't end in /openai — matching how other providers work unconditionally in _create_openai_client().

Gemini was the only provider with a domain-based gate:

Provider	Client creation	Gated?
copilot-acp	CopilotACPClient	No
google-gemini-cli	GeminiCloudCodeClient	No
gemini	GeminiNativeClient	Yes (relaxed by this PR)
openai/custom	OpenAI()	No
anthropic	build_anthropic_client()	No

Problem

Previously, is_native_gemini_base_url() checked for generativelanguage.googleapis.com in the URL before using the native adapter. With a custom base_url (e.g. a local proxy that speaks Gemini's native REST protocol), the check would fail and silently fall through to the OpenAI-compatible client — sending /v1/chat/completions format instead of /models/{model}:generateContent.

Fix

Replaced the domain-based is_native_gemini_base_url() check with a simpler /openai suffix check:

# Before:
if is_native_gemini_base_url(base_url):  # requires googleapis.com domain

# After:
if not base_url.rstrip('/').lower().endswith('/openai'):  # only skip for OpenAI-compat endpoints

• Any base_url (including custom proxies) → native Gemini adapter
• URLs ending in /openai (e.g. generativelanguage.googleapis.com/v1beta/openai) → OpenAI client (these speak OpenAI format)
• Empty base_url → native adapter (defaults to Google's endpoint internally)

Example config (cli-config.yaml)

model:
  default: "gemini-2.5-flash"
  provider: "gemini"
  base_url: "http://localhost:8900/gemini"
  api_key: "your-proxy-api-key"

Test plan

• Updated test_gemini_custom_base_url to assert native client is used
• test_gemini_openai_compat_base_url_keeps_openai_client unchanged (verifies /openai suffix fallback)
• Verify provider: "gemini" with no base_url still defaults to Google's endpoint
• Verify provider: "gemini" with custom base_url uses native adapter
• Verify other providers are unaffected

[teknium1]

Thanks for documenting the proxy use case. Current main has an explicit contrary routing contract, so this needs a maintainer design decision before salvage.

Problems

• Commit d393104bad deliberately limited the native adapter to the canonical Gemini endpoint and kept custom base URLs on the OpenAI-compatible path. That contract remains in agent/gemini_native_adapter.py:55-62 and is used by the live client factory at agent/agent_runtime_helpers.py:1671-1693.
• The PR changes the former implementation in run_agent.py; current run_agent.py:3987-3990 is only a forwarder after c42fa94af extracted the implementation.
• The predicate also controls auxiliary routing (agent/auxiliary_client.py:1855-1859, 1895-1899, 4904-4911), streaming (agent/chat_completion_helpers.py:2105-2112), and native request filtering (agent/transports/chat_completions.py:621-630). A one-site port would not establish a coherent endpoint contract.

Suggested changes

• First resolve whether the intentional custom-endpoint policy from d393104bad should change. If approved, update the shared classification contract and all consumers, with an end-to-end proxy-path test.

Automated hermes-sweeper review.

[teknium1]

Current main has extracted this implementation: run_agent.py:3987-3990 now forwards to agent/agent_runtime_helpers.py:1643. Port the change to the live helper only after resolving the explicit custom-endpoint routing policy in d393104bad; changing this former implementation cannot be salvaged as-is.