- Add stream first byte timeout (TTFB) detection to trigger failover
when provider responds too slowly (configurable via STREAM_FIRST_BYTE_TIMEOUT)
- Add rate limit fail-open/fail-close strategy configuration
- Improve exception handling in stream prefetch with proper error classification
- Refactor UsageService with shared _prepare_usage_record method
- Add batch deletion for old usage records to avoid long transaction locks
- Update CLI adapters to use proper User-Agent headers for each CLI client
- Add composite indexes migration for usage table query optimization
- Fix streaming status display in frontend to show TTFB during streaming
- Remove sensitive JWT secret logging in auth service
- Move stream smoothing parameters (chunk_size, delay_ms) to database config
- Remove hardcoded stream smoothing constants from StreamProcessor
- Simplify dynamic delay calculation by using config values directly
- Add invalidate_models_list_cache() function to clear /v1/models endpoint cache
- Call cache invalidation on model create, update, delete, and bulk operations
- Update admin UI to allow runtime configuration of smoothing parameters
- Improve model listing freshness when models are modified
- Increase chunk size from 5 to 20 characters for fewer delays
- Reduce min delay from 15ms to 8ms for faster playback
- Reduce max delay from 24ms to 15ms for better responsiveness
- Adjust text thresholds to better differentiate content types
- Apply parameter tuning to both StreamProcessor and _LightweightSmoother
- Remove complex conditional logic for short/medium/long text differentiation
- Unify text splitting to always use consistent CHUNK_SIZE-based splitting
- Rely on dynamic delay calculation for output speed adjustment
- Reduce code complexity in both main smoother and lightweight smoother
- Move StreamSmoother functionality directly into StreamProcessor for better integration
- Create ContentExtractor strategy pattern for format-agnostic content extraction
- Implement intelligent dynamic delay calculation based on text length
- Support three text length tiers: short (char-by-char), medium (chunked), long (chunked)
- Remove manual chunk_size and delay_ms configuration - now auto-calculated
- Simplify admin UI to single toggle switch with auto timing adjustment
- Extract format detection logic to reusable content_extractors module
- Improve code maintainability with cleaner architecture
- Implement StreamSmoother class to split large content chunks into smaller pieces with delay
- Support OpenAI, Claude, and Gemini API response formats for smooth playback
- Add stream smoothing configuration to system settings (enable, chunk size, delay)
- Create streamlined API for stream smoothing with StreamSmoothingConfig dataclass
- Add admin UI controls for configuring stream smoothing parameters
- Use batch configuration loading to minimize database queries
- Enable typing effect simulation for better user experience in streaming responses
- Replace time-based throttling with background task for disconnect checks
- Remove time.monotonic() and related throttling logic
- Prevent blocking of stream transmission during connection checks
- Properly clean up background task with try/finally block
- Improve throughput and responsiveness of stream processing
- Add detection of finish_reason in OpenAI API responses to mark stream completion
- Ensures OpenAI API streams are properly marked as complete even without explicit completion events
- Complements existing completion event detection for other API formats
- Check has_completion flag before marking client disconnection as failure
- Allow graceful termination if response already completed when client disconnects
- Change logging level to info for post-completion disconnections
- Prevent false error reporting when client closes connection after receiving full response
