Changelog

v0.4.1 (Latest)

OMAT — Open Multimodal Assessment Toolkit

The full OMAT assessment layer is now built into tekimax-omat. Designed for any organization measuring learning outcomes, skill development, or program effectiveness with AI-generated formative feedback.

New: AssessmentPipeline

Accepts student/participant responses in any modality: text, speech, drawing, handwriting
Automatically transcribes speech (TranscriptionCapability) and describes images (VisionCapability) before scoring
Generates structured FormativeFeedback via JSON mode: overall summary, per-criterion scores with evidence, strengths, next steps, encouragement, and normalized score
Configurable feedback language (BCP-47), temperature, and plugin hooks
assess(), assessBatch(), and assessStream() methods

New: FairnessAuditPlugin

Collects demographic performance data across assessments
Produces FairnessReport with DemographicGroup[] and DisparityFlag[]
Configurable warning (default: 10%) and critical (default: 20%) disparity thresholds
Minimum group size protection (default: 5) to prevent small-n privacy exposure
FAIR metadata attached to reports for open data publication

New: RubricValidatorPlugin

Validates AI feedback against your rubric: criterion coverage, score ranges, level label consistency, evidence presence, suggestion presence, normalized score bounds
Strict mode: throws on validation errors (for test/benchmark pipelines)
Non-strict mode: logs warnings and errors, never blocks delivery

New: LearningProgressionPlugin

Maps criterion scores to positions on developmental learning progressions
Annotates feedback with current stage, next milestone, and observable indicators
Works for any domain: writing, coding, clinical skills, workforce readiness, etc.

New: BenchmarkRunner

Runs standardized evaluations across a BenchmarkSuite
Four metrics: accuracy (Cohen's kappa vs human scores), fairness (max demographic gap), actionability (% with concrete suggestions), alignment (criterion coverage + evidence grounding)
Results include FAIR metadata for DOI assignment and archival

New: useAssessment React hook

Wraps AssessmentPipeline with React state management
Supports streaming and non-streaming modes
Abort handling, history tracking, reset

v0.4.0

AI Action Tagging

New: AIActionTagPlugin

Attaches aiTag to every ChatResult and GenerateTextResult
Infers CRUD operation from tool call names (high confidence) or response text keywords (low confidence)
onTag callback for audit logging
aiTag available on all results even without the plugin

New: aiTag on ChatResult and GenerateTextResult

Code

interface AIActionTag {
  source: 'ai';
  operation: 'create' | 'read' | 'update' | 'delete' | 'none';
  confidence: 'high' | 'low' | 'none';
  model: string;
  timestamp: number;
  toolName?: string;
}

v0.3.0

Dynamic Context Window Management

New: TokenAwareContextPlugin (replaces MaxContextOverflowPlugin)

Token-aware trimming using chars/4 heuristic
Built-in registry of 50+ models: GPT-4o (128K), Claude (200K), Gemini 1.5 Pro (1M), etc.
Live OpenRouter API fallback for unlisted models with in-process caching
Three truncation strategies: auto, last_messages, disabled
Optional Stripe Billing Meter integration (off by default) — meter token usage per customer

New: model-context.ts registry

getModelContextInfo(modelId) — instant lookup with prefix matching
fetchModelContextWindow(modelId, apiKey?) — live fetch with cache
registerModelContext() — register private or fine-tuned models

New: Conversation class

Stateful multi-turn chat: auto-appends messages, full history, plugin support
send(), stream(), inject(), clear(), restore(), export()

v0.2.0

Plugin Architecture

TekimaxPlugin interface — lifecycle hooks: beforeRequest, afterResponse, onStreamChunk, beforeToolExecute, afterToolExecute
PIIFilterPlugin — redacts emails, SSNs, phone numbers, credit card numbers from both string and multi-part ContentPart[] messages
LoggerPlugin — request/response telemetry with token reporting
ProvisionPlugin — API gateway proxy with rate limiting and SSRF protection

Core Features

Real-time streaming via AsyncIterable<StreamChunk>
Embeddings API (client.text.embed())
useChat React hook with tool loops, streaming, abort support

v0.1.9

OpenAIProvider: generateImage, analyzeImage, generateSpeech, transcribeAudio
GeminiProvider: analyzeImage, analyzeVideo
AnthropicProvider: analyzeImage (Claude 3.x vision)
OpenRouterProvider: analyzeImage (passes through to vision-capable models)

v0.1.8

Tool Calling Parity

GeminiProvider: Full tool/function calling (functionDeclarations normalized to tools array)
AnthropicProvider: tool_use content blocks normalized to toolCalls[]
Parity across OpenAI, Ollama, Gemini, Anthropic

Reasoning

parseThinking utility for <think> blocks (DeepSeek R1 and compatible models)
think: true flag in ChatOptions

v0.1.1

TekimaxProvider: Stateless mode with full history mapping
OllamaProvider: Cloud auth (apiKey) via custom fetch wrapper; browser-safe via ollama/browser

v0.1.0 — Initial Release

Monorepo structure (Turborepo)
Tekimax unified client with AIProvider interface
Adapters: Anthropic, Gemini, Grok, Ollama, OpenAI, OpenRouter
Docs site on Cloudflare Pages