feat(download_models): update CPU dropdown model size limit and add inclusion flags for models

style(app): enhance model selector UI with improved layout and visibility

feat(download_models): add Gemma 4 E2B instruct model to model repository

fix(search): pin ddgs news/text to duckduckgo backend (avoid Bing decode errors)

fix(agent): wire Wikipedia pattern routing (explain/what-is); use wikipedia_search_term

perf(app): fast market-quote synthesis — short system prompt, fewer tokens

refactor(config): rename financial_quote_hints + provider_symbol; neutral logs

refactor: tool routing + Yahoo hints live in agent_config (not new top-level config)

fix(agent): parse index names in financial_data; skip DDG for quote-only market Qs

perf: tune for 2-4 vCPU — smaller n_batch, skip router LLM, trim synthesis context

fix: always LLM-synthesize tool results (assistant voice, not raw search dump)

feat: broader capable-assistant system prompt, shared core prompt in agent LLM, smarter tool routing

feat: add meetara Qwen3 1.7B GGUF to model list (meetara-lab/meetara-qwen3-1.7b-gguf)

feat: Yahoo Finance real-time market data tool for prices + % changes

fix: news-first search for market queries, filter landing pages, add pub_date, append month to query

fix: increase snippet from 200->400 chars, log full web data for accuracy audit

feat: direct tool display path - show web/wiki/calc results instantly without LLM pass

perf: raise gen threads for small models, cut max_tokens 450->150 for tool queries

perf: n_threads_batch for faster prompt encode, RAM logging, switch tool LLM to generate_stream_chat

feat: word+equation image support via agent-based VL post-processing

feat: add Wikipedia/DateTime/UnitConverter tools + LLM context routing + VL-calc pipeline

fix: strip VL blank lines, raise min_tokens to 1024, warn on large VL model CPU

fix: use create_chat_completion for explain path, clean system prompt config

fix: remove duplicate _detect_tool_needs call, add explain-path logging

feat: lazy VL load (text-only default, VL only on image), simple system prompt

fix: buffer-based stream to stop chat-template token leakage + KV cache clear

refactor: remove programmatic BODMAS stepper, route explain to LLM with anchored answer

Replace LLM math-explain with programmatic BODMAS explainer (no LLM hallucinations)

Add describe/show/walk as explain aliases for math step-by-step routing

Fix: add explain keyword, route explain queries through LLM with step-by-step prompt

Fix: solve-pattern now matches 'solve and explain: expr' (keyword + extra words before colon)

Fix: extract full calc expression even without trailing = ? (e.g. Calculation: 9-3/(1/3)+1)

Fix: full-expr extraction for solve: X=? and skip LLM for pure-calc responses

Fix: remove VL text path from generate_stream to prevent CLIP warmup on text queries

VL: fix blurry OCR - keep RGB, add explicit read-image prompt prefix

VL: fix ! OCR artifact + strip think blocks from VL stream

VL: fix empty output — ensure ≥ 512 visual tokens for math/OCR

core: fix two-model usage gaps

fix: remove stale is_math_task / is_text_task log reference (NameError)

VL: strip all prompt engineering — bare 'You are a helpful assistant.'

VL: fix empty response caused by complex system prompt

VL math/OCR: contrast+sharpness enhancement + forced transcription step

VL: fix math/OCR failures — task-aware prompt + 768px resolution

Fix VL temp-file race, hide copy buttons, clean loading dots

UI: replace gr.Image overlay with gr.UploadButton + thumbnail chip

VL: fix image description accuracy

UI: replace vision accordion with 📎 icon in text row (ChatGPT style)

UI: compact vision strip + tighter heavy-models control

Agent: DuckDuckGo-only web search (remove Google from hot path)

UI: fix vision upload overlay, chat scroll layout, mobile height