Summary

• stop streamNormalInput from overriding chat models that explicitly set completionOptions.reasoning = false
• keep the existing session-level reasoning toggle behavior for models that do allow reasoning overrides
• add a GUI regression test that verifies chat requests do not send a forced reasoning flag for an Ollama model with reasoning disabled

Why

streamNormalInput always wrote completionOptions.reasoning from the session toggle whenever hasReasoningEnabled was set. That worked for most providers, but it broke Ollama chat models that explicitly disable reasoning in config, because the GUI would still send reasoning: true and Ollama would reject the request with errors like does not support thinking.

Validation

• ran ./node_modules/.bin/vitest run src/redux/thunks/streamResponse.test.ts in gui
• ran git diff --check
• ran ./node_modules/.bin/tsc -p ./ --noEmit in gui and only hit the existing unrelated OPENROUTER_HEADERS export mismatch from ../core/llm/llms/OpenRouter.ts

Closes #11265

Summary by cubic

Respect model-level reasoning disablement in chat requests. Ollama models with completionOptions.reasoning: false no longer receive a forced reasoning flag; the session toggle still applies to models that allow overrides.

• Bug Fixes
  • Short-circuit reasoning options when the model sets completionOptions.reasoning === false to avoid Ollama “does not support thinking” errors.
  • Add a GUI regression test to verify chat requests omit reasoning for disabled Ollama models.

^{Written for commit 85f7d1f. Summary will update on new commits.}