fix(agent): restrict think=False to Ollama/local endpoints only

[yeyitech]

Summary

• Fixes [Bug]: think=False incorrectly sent to all provider=custom endpoints, not just Ollama #11237
• The think=False parameter in extra_body is Ollama-specific. The previous guard if self.provider == "custom" matched all custom providers (Mistral, Fireworks, Together.ai, vLLM remote, etc.), causing HTTP 422 errors from cloud APIs that do not recognise the think field.
• Replace the guard with an explicit Ollama/local endpoint check that mirrors the pattern already used elsewhere in run_agent.py (_ollama_num_ctx, _is_glm_zai_direct).

Changes

In _build_api_kwargs(), replaced:

if self.provider == "custom" and self.reasoning_config and ...:

with:

_is_ollama_endpoint = (
    "ollama" in self._base_url_lower
    or ":11434" in self._base_url_lower
    or is_local_endpoint(self.base_url or "")
)
if _is_ollama_endpoint and self.reasoning_config and ...:

is_local_endpoint is already imported at the top of run_agent.py and self._base_url_lower is a cached lowercased copy of the base URL set in the base_url property setter.

Test plan

• Ollama endpoint (http://localhost:11434/v1): think=False still sent when reasoning is disabled
• Ollama via URL keyword (http://my-ollama-server/v1): think=False still sent
• Cloud custom provider (e.g. Fireworks, Together.ai): think=False no longer sent; no HTTP 422
• Mistral, vLLM remote: think=False no longer sent; no HTTP 422

🤖 Generated with Claude Code

[tong-hao]

I want this fix.

[teknium1]

Thanks for tracing the incorrect custom-provider guard. The underlying issue is still present on current main, but this patch needs to be salvaged onto the current provider-profile path.

Problems

• run_agent.py:_build_api_kwargs is now only a forwarder (run_agent.py:5328-5331); provider=custom kwargs are assembled through the registered CustomProfile (agent/chat_completion_helpers.py:849-888, plugins/model-providers/custom/__init__.py:22-60).
• The proposed is_local_endpoint() test is not Ollama-specific: it treats Docker service names and private/link-local/Tailscale addresses as local (agent/model_metadata.py:618-647). The custom profile also covers vLLM and llama.cpp (plugins/model-providers/custom/__init__.py:3-10), so local non-Ollama endpoints could still receive think=False.
• Please add regression coverage for both cloud and local non-Ollama custom endpoints; current tests still assert generic custom providers emit the field (tests/plugins/model_providers/test_custom_profile.py:51-65).

Suggested changes

• Move the endpoint-specific decision into CustomProfile.build_api_kwargs_extras and gate it on positive Ollama identification.

Automated hermes-sweeper review.