refactor: migrate NativeOllamaHandler to AI SDK

[daniel-lxs]

Summary

Replace the direct ollama npm package usage in NativeOllamaHandler with the ollama-ai-provider-v2 AI SDK community provider, following the same pattern as already-migrated providers like DeepSeek, LM Studio, etc.

Changes

Provider (src/api/providers/native-ollama.ts)

• Use createOllama from ollama-ai-provider-v2 instead of the Ollama class from the ollama package
• Replace client.chat({ stream: true }) with streamText() + processAiSdkStreamPart()
• Replace client.chat({ stream: false }) with generateText()
• Use convertToAiSdkMessages instead of the custom convertToOllamaMessages function
• Use AI SDK's built-in tool support via convertToolsForAiSdk / mapToolChoice
• Pass num_ctx via providerOptions.ollama
• Preserve Ollama-specific error handling (ECONNREFUSED → "service not running", 404 → "model not found")
• Override isAiSdkProvider() to return true
• Remove the custom convertToOllamaMessages function (no longer needed)

Dependencies (src/package.json)

• Add ollama-ai-provider-v2 package
• Remove ollama npm package (no longer used by any file)

Tests (src/api/providers/__tests__/native-ollama.spec.ts)

• Rewrite all tests to mock the AI SDK (streamText, generateText) instead of the ollama package
• All 16 tests pass
• Full test suite (5381 tests) passes with no regressions

Testing

• cd src && npx vitest run api/providers/__tests__/native-ollama.spec.ts — 16/16 pass
• Full pre-push hook passes (lint, type-check, all tests)

---

Important

Refactor NativeOllamaHandler to use ollama-ai-provider-v2 AI SDK, updating provider logic, tests, and dependencies.

• Provider (native-ollama.ts):
  • Replaces Ollama class with createOllama from ollama-ai-provider-v2.
  • Uses streamText and generateText for text generation.
  • Implements convertToAiSdkMessages and convertToolsForAiSdk for message and tool conversion.
  • Preserves error handling for specific Ollama errors (ECONNREFUSED, 404).
  • Overrides isAiSdkProvider() to return true.
• Dependencies (package.json):
  • Adds ollama-ai-provider-v2.
  • Removes ollama package.
• Tests (native-ollama.spec.ts):
  • Mocks AI SDK functions (streamText, generateText) instead of ollama.
  • Updates tests to reflect new SDK usage and error handling.
  • All tests pass, including full suite (5381 tests).

^{This description was created by for 6d36235. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task

Found 2 issues related to how providerOptions.ollama is structured for ollama-ai-provider-v2.

• num_ctx is at the wrong nesting level -- should be providerOptions.ollama.options.num_ctx, not providerOptions.ollama.num_ctx (affects both createMessage and completePrompt)
• Missing think: true in providerOptions.ollama for DeepSeek R1 models -- without it, the TagMatcher removal causes reasoning to appear as raw <think> tags in text output

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

[roomote]

num_ctx is at the wrong nesting level. The ollama-ai-provider-v2 schema (ollamaProviderOptions) expects num_ctx inside a nested options object, not directly on the ollama object. The schema is: { think?: boolean, options?: { num_ctx?: number, ... } }. As written, num_ctx will be stripped by Zod validation (or cause a validation error in Zod v4 strict mode), so the context window override is silently lost. The same issue applies in completePrompt on line 126.

Suggested change

	providerOptions: { ollama: { num_ctx: this.options.ollamaNumCtx } } as any,
	providerOptions: { ollama: { options: { num_ctx: this.options.ollamaNumCtx } } } as any,

_{Fix it with Roo Code or mention @roomote and request a fix.}

[roomote]

useR1Format is computed here but only used for temperature. The old code used TagMatcher to parse <think>...</think> tags from the response content, which worked regardless of Ollama API settings. Since TagMatcher is removed, the provider now needs to emit reasoning-delta events -- but ollama-ai-provider-v2 defaults think to false and only returns structured thinking data when think: true is explicitly passed via providerOptions.ollama.think. Without it, DeepSeek R1 reasoning will appear as raw <think> tags in plain text output instead of being separated into reasoning chunks. This also affects completePrompt. The fix is to include think: true in providerOptions.ollama when useR1Format is true.

_{Fix it with Roo Code or mention @roomote and request a fix.}