Savestate: Update DeepSeek v4 Flash ID to static 0731 version to resolve OpenRouter 400 error

Savestate: Set DeepInfra as the exclusive provider for DeepSeek v4 Pro due to privacy and price-cache optimum

Savestate: Revert DeepSeek v4 models to use -latest suffix and re-enable ZDR data_collection deny for cache discount security

Savestate: Update DeepSeek v4 config for caching and disable ZDR

Savestate: Add ZDR and OpenRouter specific routing for DeepSeek models

fix: pass OPENROUTER_API_KEY in start.sh

Savestate: Force WEB-IMAGE fallback if SiriHero card swallows the image tag

Savestate: Switch Web Image Search to Wikimedia API and fix plot-override bug

Savestate: Implement global Web Image Search for educational visualizations

Savestate: Optimize image intent prompt for scientific accuracy

Savestate: Fix syntax error in router.js

Savestate: Fix image intent evaluation by enforcing gemini-3.5-flash-lite and expanding regex fallback

feat: Enhance image intent classifier for inline visual explanations

Savestate: Update AI system prompt to explicitly detect empty text lines as gaps

Savestate: Implement Advanced Canvas Rendering Engine with Kalam handwriting font, multiply blend modes, word wrapping, and random rotation for ultra-realistic offline inpainting

Savestate: Integrate Gradio Client to hit FLUX.1-Fill-dev Space directly for high-quality AI inpainting

Savestate: Switch HF Inpainting model from FLUX.1 to SDXL due to provider availability restrictions on the free tier

Savestate: Update HuggingFace Inference API URL

Savestate: Add CORS Proxy fallback for HuggingFace FLUX.1-Fill-dev Inference API

Savestate: Fix intent classification bug where the user prompt was completely overshadowed by the massive Base64 string

Savestate: Force JSON output by appending strong instruction to prompt, because lite models often ignore system instructions

Savestate: Route intent classification to currently selected model if it is a Gemini model

Savestate: Add Vision API support to openAIClient.js to allow OpenRouter models to see images

Savestate: Route intent classification to DeepSeek API if configured, bypassing Gemini quota limits

Savestate: Add regex fallbacks for Intent Classifiers and better timeout/error reporting for HuggingFace/Gemini

Savestate: Implement Two-Step FLUX.1-Fill Inpainting Architecture for Worksheet Solving via Hugging Face Serverless API

Savestate: Implement model fallback mechanism for intent classifiers (3.5-flash -> 3.1-flash-lite)

Savestate: Switch intent classifiers to gemini-3.1-flash-lite to save quota and speed up background tasks

Savestate: Fix overly sensitive intent classifiers and switch them to gemini-3.5-flash

Savestate: Enforce image tag generation even if the AI model ignores the instruction

Savestate: Fix DeepSeek API rejecting model names by not stripping the deepseek- prefix in router

Savestate: Route image intent and fallback to the user-selected model instead of hardcoding gemini-2.5-flash to prevent API limit clashes

Savestate: Fix small models like Flash Lite ignoring the BILD tag due to conflicting SiriHero instructions

Savestate: Integrate Nano Banana 2 image generation via Pollinations and custom UI editor

Savestate: Added render distance optimization to detective board and repositioned auto-scroll button to bottom right

Savestate: Final fixes for detective board nodes, lines, auto-scroll centering and translations

Savestate: Update detective board to support vertical mouse wheel scrolling and fix SVG viewbox

Savestate: Add scroll wheel zooming to detective board and fix SVG path z-fighting issue

Savestate: Implement mathematically flawless 3D to 2D projection for detective board auto-scrolling and mouse panning, fix tooltip 3D rotation, and add scroll-to-latest button

Savestate: Fix detective board GPU texture crash by replacing massive 3D canvas with 0x0 pan layer and background-position animation

Savestate: Rewrite 3D board scrolling to use CSS transform translate instead of native overflow for flawless panning

Savestate: Add drag-to-pan feature and zoom out Detective Board for better overview

Savestate: Insert Detective Board open button as chat message bubble and actually fix isometric layout calculation

Savestate: Fix isometric projection calculation and add sticky re-open button for Detective Board

Savestate: Implement 3D Isometric Detective Board UI, fix mobile report formatting, increase report length constraint

Savestate: Fix article mobile formatting, Plotly wrapper overflow, research-agent infinite loop prevention, loader animation conflict

Savestate: UX polish, fix PDF fallback modal, gap analysis resilience, cinematic zoom out

Savestate: Final tweaks for Deep Research Canvas (async init, truncations)

Savestate: Integrate Canvas into Chat View

Savestate: Fixed runDeepResearch signature to prevent TypeError