* fix: add cache reporting support for OpenAI-Native provider
- Add normalizeUsage method to properly extract cache tokens from Responses API
- Support both detailed token shapes (input_tokens_details) and legacy fields
- Calculate cache read/write tokens with proper fallbacks
- Include reasoning tokens when available in output_tokens_details
- Ensure accurate cost calculation using uncached input tokens
This fixes the issue where caching information was not being reported
when using the OpenAI-Native provider with the Responses API.
* fix: improve cache token normalization and add comprehensive tests
- Add fallback to derive total input tokens from details when totals are missing
- Remove unused convertToOpenAiMessages import
- Add comment explaining cost calculation alignment with Gemini provider
- Add comprehensive test coverage for normalizeUsage method covering:
- Detailed token shapes with cached/miss tokens
- Legacy field names and SSE-only events
- Edge cases including missing totals with details-only
- Cost calculation with uncached input tokens
* fix: address PR review comments
- Remove incorrect fallback to missFromDetails for cache write tokens
- Fix cost calculation to pass total input tokens (calculateApiCostOpenAI handles subtraction)
- Improve readability by extracting cache detail checks to intermediate variables
- Remove redundant ?? undefined
- Update tests to reflect correct behavior (miss tokens are not cache writes)
- Add clarifying comments about cache miss vs cache write tokens
* feat: add Ollama API key support for Turbo mode
- Add ollamaApiKey field to ProviderSettings schema
- Add ollamaApiKey to SECRET_STATE_KEYS for secure storage
- Update Ollama and NativeOllama providers to use API key for authentication
- Add UI field for Ollama API key (shown when custom base URL is provided)
- Add test coverage for API key functionality
This enables users to use Ollama Turbo with datacenter-grade hardware by providing an API key for authenticated Ollama instances or cloud services.
* fix: use VSCodeTextField for Ollama API key field
Remove non-existent ApiKeyField import and use standard VSCodeTextField with password type, matching other provider implementations
* Add missing translation keys for Ollama API key support
- Add providers.ollama.apiKey and providers.ollama.apiKeyHelp to all 18 language files
- Support for authenticated Ollama instances and cloud services
- Relates to PR #7425
* refactor: improve type safety for Ollama client configuration
- Replace 'any' type with proper OllamaOptions (Config) type
- Import Config type from ollama package for better type checking
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: Daniel Riccio <ricciodaniel98@gmail.com>
* fix: handle Mistral thinking content as reasoning chunks
- Add TypeScript interfaces for Mistral content types (text and thinking)
- Update createMessage to yield reasoning chunks for thinking content
- Update completePrompt to filter out thinking content in non-streaming mode
- Add comprehensive tests for reasoning content handling
- Follow the pattern used by other providers (Anthropic, OpenAI, Gemini, etc.)
Fixes#6842
* fix: resolve TypeScript type issue in completePrompt method
* fix: handle Mistral thinking content chunks in streaming responses
- Added ContentChunkWithThinking type helper to handle thinking chunks
- Properly converts thinking content to reasoning chunks in streaming
- Filters out thinking content in non-streaming completePrompt responses
- Confirmed that Mistral API does send thinking chunks with type 'thinking'
- Works with Mistral SDK v1.9.18
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: daniel-lxs <ricciodaniel98@gmail.com>
* fix: update DeepSeek models context window to 128k
- Updated deepseek-chat and deepseek-reasoner models from 64k to 128k context window
- Updated corresponding test expectations
- Aligns with DeepSeek API documentation at https://api-docs.deepseek.com/quick_start/pricing/Fixes#7268
* feat: update deepseek-reasoner maxTokens to 64K based on official documentation
* fix: use default maxTokens values instead of maximum for DeepSeek models
- deepseek-chat: 4096 (4K default) instead of 8192 (8K max)
- deepseek-reasoner: 32768 (32K default) instead of 65536 (64K max)
- Updated tests to match new default values
- Updated description to clarify default vs max output tokens
* fix: use maximum output tokens for both DeepSeek models
- deepseek-chat: 8192 (8K max)
- deepseek-reasoner: 65536 (64K max)
- Updated tests to match maximum values
- Updated description to reflect 64K max output
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: daniel-lxs <ricciodaniel98@gmail.com>
* fix: omit temperature parameter when not explicitly set for OpenAI Compatible providers
- Modified OpenAiHandler to only include temperature when modelTemperature is defined
- Modified BaseOpenAiCompatibleProvider to only include temperature when modelTemperature is defined
- Added tests to verify temperature is omitted when undefined
- Updated existing tests to explicitly set temperature where needed
This allows backend services (LiteLLM, vLLM) to use their configured default temperatures
instead of being forced to use temperature=0 when "Use custom temperature" is unchecked.
Fixes#7187
* test: update tests to match new temperature handling behavior
- Remove temperature parameter expectations from provider tests
- Tests now expect temperature to be omitted when not explicitly set
- Aligns with PR #7188 changes to fix OpenAI Compatible provider behavior
---------
Co-authored-by: Roo Code <roomote@roocode.com>
Co-authored-by: daniel-lxs <ricciodaniel98@gmail.com>
* fix: prevent duplicate LM Studio models with case-insensitive deduplication
- Keep both listDownloadedModels and listLoaded APIs to support JIT loading
- Implement case-insensitive deduplication to prevent duplicates
- When duplicates are found, prefer loaded model data for accurate runtime info
- Add test coverage for deduplication logic
- Addresses feedback about LM Studio's JIT Model Loading feature (v0.3.5+)
Fixes#6954
* fix: correct deduplication logic to prefer loaded models
- When a loaded model ID is found in any downloaded model key (case-insensitive)
- Remove the downloaded model and replace with the loaded model
- This ensures loaded models with runtime info take precedence
- Updated tests to verify the correct deduplication behavior
* fix: improve deduplication logic and add comprehensive test coverage
- Enhanced deduplication to use path segment matching instead of simple substring
- Prevents false positives like 'llama' matching 'codellama'
- Added comprehensive test cases for edge cases and multiple scenarios
- Maintains support for JIT Model Loading feature
