config: strip whitespace from settings, fixing the prod image search

health: report whether image search actually works; don't eat $ as math

citations: space before the link, and a hover preview card

images: rounded uncropped tiles, and a fallback search query

web: make 【domain】 citations real links

pipeline: give the hedge the web, instead of skipping it

images: search the answer, not the question; larger uncropped tiles

pipeline: Granite web-search gate, racing the hedge

evals: tunable reasoning budget, so reasoning models score fairly

evals: benchmark candidate models for the t=0 web-search gate

images on web-search answers, via Tavily

rich teaching answers, math rendering fix, emoji hard cap

web-need detection at t=0, shared answer formatting, friendly copy

answer formatting + UI: rich-answer format prompt, gpt-oss labels, input polish

router: GPT-OSS 120B generalist + Llama 4 Scout verifier; greeting fix; input UI

needs_web: flag specific numbered/named items (e.g. LeetCode 3499)

web search: tell verifier answer is web-grounded; add 'searching' stage

inject current date so models flag needs_web for recent events

bid_timeout 20->30s; narrow AI bubbles to 68%; round the input box

tier-1 lineup: Gemma 4 (general), Qwen3 Coder (coding), DeepSeek V4 Flash (logic/math)

qwen: force low bids outside coding

fix math rendering: strict LaTeX delimiters in frontier prompt + forgiving KaTeX

skip verification entirely when the creative model wins

frontier: GPT-5 -> GPT-5.6 Terra

tier-1: DeepSeek V4 Flash (general), Qwen (coding), Skyfall 36B V2 (creative)

general slot: Llama 3.1 70B -> Llama 4 Maverick

general slot: Hermes 4 70B -> Llama 3.1 70B Instruct

web search: freshness-gated OpenRouter web plugin for the agents

bidder answers: go long when detail is asked; raise token caps

simplify BID_SYSTEM per user's version (typos fixed)

fix literal 'Subheading' echo in prompts; bold back to white

general slot: Gemini 2.5 Flash Lite -> Hermes 4 70B

tier-1: Gemini 2.5 Flash Lite (general), DeepSeek V4 Flash (math), Qwen3 Coder (coding)

bidders: allow slightly more depth on explanatory answers

answers: point-per-line structure guidance, moderate spacing, lighter body text

prompt + spacing: airy answers, no wall-of-bold formatting

revert verifier: Tencent Hy3 -> GPT-OSS 120B

coding slot: Qwen3 Coder Flash (paid) -> Qwen3 Coder (free-first)

tier-1 reshuffle: V4 Flash (general), Qwen3 Coder Flash (coding), Nemotron (math)

tier-1 lineup: DeepSeek V4 Pro (general), V4 Flash (math), Nemotron 3 Ultra (coding)

swap math bidder: DeepSeek chat -> DeepSeek V4 Flash

swap models: Nemotron 3 Ultra as general bidder, Tencent Hy3 as verifier

deploy hardening: access gate, rate limiting, spend guard, env CORS

eval harness + env-swappable frontier model

instruct GPT-5 to format answers in Markdown

hard gate: GPT-5 reserved exclusively for hard STEM/reasoning queries

speculative bid answers follow the same typo/ambiguity rules as drafts

treat ambiguity as normal, not escalation-worthy

treat obvious typos as the intended question across bid/answer/verify

verifier: grade only the current question, not stale context