v1.3.1 — Fix ONNX Runtime crash with q4 dtype

What's Changed

• Default dtype changed from q4f16 to q4 — fixes ONNX Runtime crash in browsers.

The Bug

q4f16 uses fp16 activations, which trigger a known graph optimization bug in ONNX Runtime WASM:

SimplifiedLayerNormFusion / InsertedPrecisionFreeCast

This happens across all fp16 models in browsers, not just Qwen2.5.

The Fix

q4 uses fp32 activations — fully compatible with ONNX Runtime WASM, no crashes.

Tradeoff

dtype	Size	Compatible
q4 (new default)	~750MB	✅ Browser + Node
q4f16 (old)	~500MB	❌ Crashes in browser
fp16	~1GB	❌ Crashes in browser

References

• transformers.js #1698
• microsoft/onnxruntime #25692

v1.2.3 — Qwen2.5-0.5B-Instruct default model

What's Changed

• Default Transformer.js model upgraded to Qwen2.5-0.5B-Instruct (from SmolLM2-135M-Instruct)
  • 500M params (3.7× larger), significantly better completions
  • IFEval ~51 vs 29.9 (SmolLM2-135M)
• Default dtype changed from q4 to q4f16 (~500MB download)
• Uses Transformers.js v3+ native chat template (messages API instead of manual <|im_start|> formatting)
• Default maxNewTokens increased from 30 to 40

Model comparison

Model	Params	IFEval	Download
SmolLM2-135M-Instruct (old)	135M	29.9	~30MB (q4)
Qwen2.5-0.5B-Instruct (new)	500M	~51	~500MB (q4f16)

v1.2.2 — Ghost text styling

Changed

• Ghost text now rendered dimmed (opacity 0.38) and italic — visually distinct from actual editor text, matching Copilot-style UX expectations
• Action buttons styled — accept/reject hint buttons now have subtle borders and hover states
• Styles injected once via <style> element in document.head

v1.2.1

Changed

• Default model upgraded: SmolLM-135M-Instruct → SmolLM2-135M-Instruct — ~2× better instruction following (IFEval 29.9 vs 17.2), trained on 2T tokens vs ~600B. Same 135M parameter count, no speed regression.

v1.2.0 — Built-in Transformers.js Browser-Side AI Provider

🤖 Transformers.js AI Provider (Browser-Side)

Run AI ghost text completions entirely in the browser — no server, no API keys, no setup. Powered by Transformers.js + SmolLM3-135M-Instruct with WebGPU acceleration.

✨ What's New

• createTransformersJsProvider() — Creates a fully-local AiCompletionProvider backed by Transformers.js
• transformersJs: true in aiCompletion config — one-line opt-in to browser-side AI
• getTransformersJsState() — query loading/ready/error state for UI
• preload() — trigger model download programmatically (e.g., when user clicks "Enable AI")
• Zero API keys — model runs on the user's machine using WebGPU or WASM fallback
• ~30MB download — SmolLM3-135M-Instruct, q4 quantized, cached in IndexedDB
• Optional peer dependency — @huggingface/transformers loaded dynamically, core library stays lean

📋 Usage

With the shorthand:

new IdaztianEditor({
  extensions: {
    aiCompletion: { transformersJs: true },
  },
})

With full configuration:

import { createTransformersJsProvider } from 'idaztian'

const provider = createTransformersJsProvider({
  modelId: 'HuggingFaceTB/SmolLM3-135M-Instruct',
  dtype: 'q4',
  maxNewTokens: 30,
  temperature: 0.3,
  onProgress: (pct, status) => console.log(`${pct}%: ${status}`),
  onReady: () => console.log('AI ready!'),
})

new IdaztianEditor({
  extensions: {
    aiCompletion: { provider },
  },
})

🚀 GitHub Pages Demo

The Idaztian web demo now includes an AI button in the header. Click it to download and load the SmolLM3 model — once cached, completions appear as inline ghost text when you type. Ctrl+Shift+I toggles AI on/off.

🔧 Technical

• Uses @huggingface/transformers v4 (WebGPU runtime, C++ ONNX backend)
• Model: HuggingFaceTB/SmolLM3-135M-Instruct (q4, ~30MB)
• Automatically falls back from WebGPU to WASM if WebGPU is unavailable
• English-only completions (configurable via systemPrompt)
• Works in Chrome, Edge, Firefox (WebGPU), and Safari (WASM fallback)

---

Full Changelog: v1.1.0...v1.2.0

v1.1.0 — AI Inline Ghost Text Completion

🆕 AI Inline Completion

Copilot-style ghost text completions, powered by any OpenAI-compatible API.

✨ Features

• Pluggable provider — AiCompletionProvider.fetchCompletion(context, signal) works with OpenAI, Ollama, Groq, LM Studio, or your own backend
• Ghost text — dimmed inline suggestion after the cursor
• Tab to accept, Escape to dismiss
• Debounced — configurable debounceMs, auto-cancels in-flight requests on new input
• Staleness-safe — completions from old document states are discarded
• Opt-in — disabled by default, enable via extensions.aiCompletion in config

📦 Usage

new IdaztianEditor({
  parent: document.getElementById('editor'),
  extensions: {
    aiCompletion: {
      provider: {
        async fetchCompletion(context, signal) {
          const res = await fetch('https://api.openai.com/v1/chat/completions', {
            method: 'POST',
            headers: {
              'Content-Type': 'application/json',
              'Authorization': `Bearer ${API_KEY}`,
            },
            body: JSON.stringify({
              model: 'gpt-4o-mini',
              messages: [
                { role: 'system', content: 'Continue the text naturally. Output ONLY the continuation.' },
                { role: 'user', content: context },
              ],
              max_tokens: 150,
            }),
            signal,
          });
          const data = await res.json();
          return data.choices?.[0]?.message?.content?.trim() || null;
        },
      },
      debounceMs: 500,
    },
  },
});

📄 Documentation

• Updated README with full AI completion example
• New CHANGELOG.md with full release history
• Updated landing page with AI Autocompletion feature card

---

Full Changelog: v1.0.4...v1.1.0

v1.0.4 — UX cursor navigation fixes

Fixes

Mouse Click Position Correction

• Added click-correction plugin that uses DOM-based position lookup (posAtDOM + character-level text node walk) instead of CM6's posAtCoords, which accumulates errors across block widgets
• Prevents clicks from landing 2-4 lines away from the intended target

ArrowUp/Down Navigation

• ArrowUp from document end (trailing empty line) correctly moves to end of previous line
• After CM6 processes arrow keys, detects incorrect line placement and corrects it (CM6's moveVertically can't navigate through block widgets)
• Table ArrowUp/Down handlers now scan up to 3 lines for adjacent table widgets (not just immediately adjacent lines)
• estimatedHeight + coordsAt on TableWidget for proper height oracle tracking
• Horizontal rules: Decoration.line + inline spacer widget (estimatedHeight: 49px) for accurate line height
• Removed line-height: 1.45 from .idz-code-line (variable line-heights caused height oracle drift)

Other

• Cleaned up unused CSS (removed .idz-hr-line, unused Vite boilerplate)
• Bump to v1.0.4
• 14 tests all passing

v1.0.3

ci: Configure GitHub Actions workflow for npm package publishing.

v1.0.2

chore: remove static NPM_TOKEN to establish full OIDC publishing prov…

v1.0.1

chore: enable npm provenance flag