Changelog

v0.2.0 (Latest)

Middleware Hooks: Introduced the TekimaxPlugin interface exposing lifecycle hooks (beforeRequest, afterResponse, onStreamChunk, beforeToolExecute, afterToolExecute).
Security (PIIFilterPlugin): Built-in plugin to redact sensitive regex patterns (emails, SSNs) before they leave the client.
Scalability (MaxContextOverflowPlugin): Built-in plugin to safely truncate context windows on long-running loops while safely preserving system prompts.
Telemetry (LoggerPlugin): Standardized console logging for requests, tool executions, and tokens.

Real-Time Streaming: Native support for Server-Sent Events (SSE) via Javascript Async Iterables (client.text.generateStream()).
Embeddings API: Unified interface for text embeddings (client.text.embed()).
React Hooks: Exported useChat hook from tekimax-ts/react supporting chunked streaming directly in the frontend.

Models.dev Introspection: The SDK now optionally fuses metadata from the models.dev catalog into standard ModelDefinition objects.
Strict Capabilities: getModels() exposes an authoritative map of model limits, costs, supported modalities, and reasoning (think) capabilities.

OpenAIProvider: Added generateImage, analyzeImage, generateSpeech, and transcribeAudio.
GeminiProvider: Added analyzeImage and analyzeVideo (Video-to-Text).
AnthropicProvider: Added analyzeImage (Vision via Claude 3.5 Sonnet).
Core: Validated unified multi-modal interfaces across all modalities.

GeminiProvider: Added full support for function calling (tools and tool_choice). Gemini uses functionDeclarations internally, but the SDK normalizes this to the standard tools array format.
AnthropicProvider: Implemented response parsing for tool_use blocks, mapping Anthropic's content-block format to the unified toolCalls array.
Parity: Achieved tool calling parity across OpenAI, Ollama, Gemini, and Anthropic.

OllamaProvider: Switched import from ollama to ollama/browser to avoid node:fs externalization errors in Vite/Next.js client-side builds. The /browser sub-export strips Node-only file-system code.
Utilities: Added browser-safe fileToBase64 helper using FileReader instead of Buffer.

Reasoning: Added parseThinking utility to handle <think> blocks from models like DeepSeek-R1. Handles three cases: no think tag, streaming (tag opened but not closed), and complete blocks.
Exports: Exposed ToolDefinition and ToolCall types from root.

TekimaxProvider: Added support for Stateless Mode (full history mapping).
OllamaProvider: Added support for Cloud Authentication (apiKey option). Uses a custom fetch wrapper to inject Authorization: Bearer headers, because the Ollama JS SDK doesn't natively support auth tokens.
Verified: Consistency checks passed for Anthropic, Gemini, OpenAI, and internal adapters.

Updated core documentation to reflect stateless capabilities.
Added usage examples for authenticated Ollama connections.
Removed stale "Supported Models" lists in favor of linking to official provider documentation.

openai: 4.0.0 → ^6.17.0 (major upgrade — v6 changed response types and requires max_completion_tokens instead of max_tokens).
@anthropic-ai/sdk: 0.33.1 → ^0.72.1.
@google/generative-ai: 0.21.0 → ^0.24.1.
ollama: 0.5.11 → ^0.6.3.

Monorepo Structure: Migrated to Turborepo for unified builds across SDK, docs, and UI packages.
Unified Client: Tekimax client with AIProvider interface.
Adapters: Initial support for Anthropic, Gemini, Grok, Ollama, OpenAI, and OpenRouter.
Docs: Launched documentation site on Cloudflare Pages.