fix(agent): stop sending think=false to remote custom providers

[LeonSGP43]

Summary\n- only send Ollama's native \ flag to Ollama-style custom endpoints\n- keep existing local Ollama behavior intact while avoiding 4xx validation errors on remote OpenAI-compatible providers\n- add a regression test for remote custom providers\n\nWhy\n- \ currently injects \ for every endpoint when reasoning is disabled\n- remote custom providers like Mistral reject that Ollama-specific field with HTTP 422\n\nFixes #11237\n\nTesting\n- python3 -m pytest -o addopts='' tests/run_agent/test_run_agent.py -k "think_false or remote_custom_provider_does_not_receive_ollama_think_flag or non_custom_provider_unaffected"

[alt-glitch]

Competing fix for #11237 — at least 4 other PRs address the same issue: #11362, #11520, #12914, #12488. All gate think=false behind Ollama detection.

[teknium1]

Thanks for isolating the incompatible think field.

Problems

• Current main no longer has the inline request builder this PR edits: AIAgent._build_api_kwargs() is a forwarder at run_agent.py:5328-5331. The active provider-specific request construction calls CustomProfile.build_api_kwargs_extras() through agent/transports/chat_completions.py:562-594.
• The reported behavior is still live there: plugins/model-providers/custom/__init__.py:52-56 emits extra_body["think"] = False for every disabled-reasoning custom endpoint. Existing assertions at tests/plugins/model_providers/test_custom_profile.py:51-64 encode that broad behavior.

Suggested changes

• Port the endpoint predicate to CustomProfile.build_api_kwargs_extras() and gate the assignment at plugins/model-providers/custom/__init__.py:55-56 using its supplied base_url context.
• Add remote-custom and supported-Ollama cases at the profile/transport layer, replacing the stale run-agent test target.

Automated hermes-sweeper review.

[teknium1]

Current main no longer executes this inline custom-provider branch: AIAgent._build_api_kwargs() now forwards to agent.chat_completion_helpers, whose transport invokes CustomProfile.build_api_kwargs_extras() (agent/transports/chat_completions.py:562-594). Port this guard to plugins/model-providers/custom/__init__.py; otherwise the active request path will still emit think: false for remote custom endpoints.