fix(thinking): disable thinking for unsupported Ollama models

[chioarub]

Fixes #1371

Overview

This pull request addresses the invalid_request_error: "llama3.1:8b" does not support thinking API failure encountered when using Ollama models that lack reasoning capabilities.

By centralizing the thinking gate and evaluating it against the model actually used for the request attempt, OpenClaude avoids constructing Anthropic-side thinking parameters for unsupported Ollama routes, including retry and fallback paths.

Changes

• Capability Gating: Introduced shouldUseThinkingForModel to consistently combine app-level thinking settings, the global thinking disable flag, and model capability support.
• Ollama Default Handling: Unknown or catalog-missing Ollama models now safely default to false for thinking support.
• Retry Resilience: API parameter construction evaluates thinking support against retryContext.model instead of the stale base request model.
• Regression Coverage: Moved the Ollama regression test to the thinking-gate boundary so it fails if claude.ts starts constructing unsupported thinking params again.

Verification

• bun test src/utils/thinking.test.ts src/services/api/openaiShim.test.ts
• git diff --check

[jatmn]

Findings

• [P2] Replace the Ollama shim test with coverage for the actual thinking gate
  src/services/api/openaiShim.test.ts:4955
  This test asserts that the OpenAI-compatible request body does not contain thinking, but the shim already only writes body.thinking for routes whose thinkingRequestFormat is deepseek-compatible; Ollama does not set that option. That means this new test can pass even if the claude.ts thinking gate regresses and still constructs Anthropic-side thinking params for llama3.1:8b, so it does not protect the #1371 failure path. Please move the regression coverage to the boundary that changed here, such as queryModel/the params passed into createOpenAIShimClient, or directly cover shouldUseThinkingForModel with the Ollama route so the test fails without the new gate.

[chioarub]

Addressed the requested test change in 0732cc58.

• Removed the downstream Ollama openaiShim request-body test; it did not protect the When using OLLAMA Model llama3 getting getting does not support thinking #1371 path because the shim already suppresses body.thinking unless the route is deepseek-compatible.
• Moved regression coverage to shouldUseThinkingForModel under the Ollama route, including llama3.1:8b while app-level thinking is enabled.
• Added coverage for a catalog-missing local model name so Ollama cannot fall through the Claude 4 heuristic path.

Verification:

• bun test src/utils/thinking.test.ts src/services/api/openaiShim.test.ts — 108 pass, 0 fail
• git diff --check

I also sanity-checked the regression shape by temporarily removing the model capability check from shouldUseThinkingForModel; the new test failed before restoring the fix.

[jatmn]

Thanks for the update. I rechecked the previously discussed paths and do not see any remaining actionable issues from my side.