NNyquest docs

What's New

September 5, 2026

  • Retired image models: removed unavailable direct Imagen 4 choices after Google retired them. Existing images remain accessible.

  • Media at provider cost: images, speech, music, and video have no Nyquest surcharge. Missing provider costs remain pending reconciliation instead of being estimated. Direct Google rates and video reservations use the published billable units.

  • Docs and pricing corrections: clarified hosted privacy, PAT statelessness, current authentication limits, media pricing at provider cost, and recovery from pending or unknown provider outcomes.

  • Agent accounting: new planner calls now reserve funds before dispatch, settle through the wallet ledger, and record real dollar costs. The budget check includes final synthesis. Historical runs are not repriced.

  • Documentation reader: searchable excerpts, working inline code and heading links, mobile navigation, and pages available as static HTML. The docs assistant handles interrupted streams and validates its citations.

A curated changelog of shipped features, newest first.

September 3, 2026 — Split the Diff

  • Eligible managed-chat compression savings are split 50/50 — compression already shrank your prompts before they reached the model; now half of what those tokens would have cost is credited back to your wallet on the same request, as a Split the Diff transaction. The other half is Nyquest’s managed-chat fee, subject to the provider-cost floor. Media is billed at provider cost. See Plans and Pricing.
  • Savings pill and ledger — a green ⚡ pill in the top bar keeps your running credited total, and the new Savings page shows tokens saved, the diff, your share, a per-model breakdown, and recent requests.

August 7–10, 2026 — Billing counts the whole prompt

  • Exact prompt counting — billed input tokens now come from a model-aware counter that reads both plain-text and array-form (multimodal / structured-block) message content, replacing an approximate estimate. Prompts that were previously undercounted — image and multi-part messages especially — now bill closer to their true size. See Wallet and Funds.
  • Account deletion removes your generated media — deleting an account now also deletes the files it generated (images, audio, music, video), not just database records. See Memory Privacy and Deletion.
  • No more blank replies from the auto-router — if an auto-routed model returns no content, the request retries once on that tier's deterministic default. If you pinned the model, your pick is respected and its reply stands. See Auto-Router: Tiers, Domains, and Providers.

August 2–5, 2026 — Counting with the right tokenizer

  • TokenLens counts with the destination model's tokenizer — and names which tokenizer it used, so the number matches the model you are actually sending to.
  • Target-model compiler — model and workload act as two axes whose intersection governs how a prompt may be rewritten. The decision is exposed over HTTP and surfaced in the UI, including when it declines to compress and why.
  • Compression scope tightened — a per-model backoff table stops compression on models where evaluation showed it did harm, and compression is now English-scoped by default.

July 31, 2026 — The platform markup is gone

  • No markup on platform-hosted chat — managed chat is billed at the provider's own catalog rate. The per-model rates published on the pricing page are now exactly what you pay; the previous 1.2× platform markup has been removed. BYOK was, and remains, unbilled by us. Media Studio is billed at provider cost under the current pricing policy. See Plans and Pricing.

July 27–30, 2026 — Grounding you control

  • Web grounding is caller-controlled — send x-nyquest-web: on | off | auto (or a web field in the body; the header wins) to force or suppress grounding per request. Useful for deterministic evaluations.
  • Grounding sits below your system prompt — the injected web block is inserted after your own system messages, so retrieved pages can no longer outrank your instructions.
  • Stateless observability — grounded responses report which mode ran and which sources were used, so a stateless caller can tell that live web content was injected.
  • Pinned retired models fail cleanly — pinning a model a provider has withdrawn now returns a clear client error before dispatch, naming a live alternative, instead of leaking an upstream error body. See Models, Providers, and BYOK Issues.

July 20–21, 2026 — Memory grows up

  • "What Nyquest remembers" panel (new) — Settings now lists the facts Nyquest has stored about you. Each fact can be pinned to keep it in recall, or forgotten to remove it permanently. See Memory Privacy and Deletion.
  • Per-fact control — pinning and forgetting replaces the old all-or-nothing "delete all memory" choice. See What Gets Remembered.
  • Confidence feedback — re-affirming a fact raises its confidence; contradicting one lowers it, and confidence feeds recall ranking.
  • Daily consolidation — a background pass merges duplicate facts and retires transient ones. Pinned facts are exempt.
  • Recall gate — recall fires only when the turn calls for it, so irrelevant context stops being injected. See Recall and Surfacing.

July 18–19, 2026 — TokenLens: see what your prompt costs before you send it

  • Live token count and cost estimate — every composer shows a running token count, a context-window gauge, and an estimated cost for the selected model, updating as you type.
  • Two-stage prompt optimizer — Optimize rewrites your prompt for fewer tokens, then runs an integrity-repair pass against the protected items found in your original, so names, paths, numbers, URLs and quoted text survive. You review before/after and choose whether to accept.
  • Optimization modes — Balanced (default), Code Optimize (deterministic only), or a custom instruction.
  • Voice input — a mic button on every composer. Uses your browser's built-in speech recognition where available. Audio is held in memory for the request only.

July 15–20, 2026 — A new layout and a 12th chassis

  • New navigation — a floating dock on desktop and a bottom tab bar on phones, shared by every chassis. The legacy per-chassis left icon rails have been retired. See The Dock.
  • Dedicated phone layout — phones get a purpose-built shell instead of a squeezed desktop one, with the composer always reachable.
  • Prism chassis (new) — rainbow LED filament on deep violet-black. The catalog is now 12 chassis. See Picking a Chassis.
  • Chassis deep links — ?chassis=<id> in the URL wins over your saved pick. See Switching Chassis.

July 6, 2026 — Media Studio: music & video arrive

  • Music generation (new) — the Studio now has a Music tab. Describe a mood or style and Google Lyria composes an instrumental clip you can play and download. See Media Studio.
  • Video generation (new) — a Video tab generates short clips from a prompt with Google Veo 3.1. It runs as a background job (renders in ~1–2 min), holds the cost up front, and refunds automatically if a render fails.
  • Imagen 4 images — the Vision studio adds Google Imagen 4 in Fast / Standard / Ultra tiers.
  • 30+ Gemini voices — the Voice studio adds Google's native Gemini TTS voices alongside the existing set.
  • New API endpoints — /v1/music/generate, /v1/video/generate + /v1/video/jobs/{id}, plus /v1/music/models and /v1/video/models. See the Endpoints Reference.
  • These features are powered by Nyquest's own direct model access — no BYOK required.

July 1, 2026 — The router gets domain smarts

  • Per-domain model quality — the auto-router now ranks models by how well they've actually performed in your task's domain (code, creative, reasoning, vision, general), with recent results weighted over old ones and a capped exploration budget for under-tested models. See Auto-Router: Tiers, Domains, and Providers.
  • Sharper intent detection — image- and audio-generation intent now requires strong, unambiguous phrasing. "Add a column," "explain The Picture of Dorian Gray," and "say hello" no longer misroute to media models.
  • Nyquest brand chassis — existing users were migrated once to the new default chassis; your next pick sticks permanently.

June 30, 2026 — T0 streams + a self-healing catalog

  • T0 local tier now streams — simple auto-routed tasks are served from Nyquest's own hardware at $0 with full streaming, falling back cleanly to hosted models if the local tier is unavailable.
  • Self-healing free chain — every free-mode model is probed every 6 hours with a real chat-sized request and must return actual content to stay in the chain; broken models drop out and recovered ones return automatically. See Free Mode.
  • Catalog hygiene — models that upstream providers list but can't actually serve are now detected and hidden from the picker automatically.
  • Per-key spend attribution — usage is now attributed to the exact API key that generated it, and keys can be revoked without losing their history.

June 9–12, 2026 — Provider layer + the T0 local tier

  • Free-first provider routing — the same model is often served by multiple upstream providers; Nyquest now routes to free capacity first (NVIDIA Build), with adaptive circuit breakers and graceful failover. Savings are attributed per-request.
  • Benchmarks page — Settings → Benchmarks shows live provider comparisons, and a new Preferred Provider setting lets you steer routing.
  • T0 local tier launched — a model running on Nyquest's own hardware serves simple tasks at $0.
  • Anonymous BYOK "Test connection" — probe your key against your provider before chatting, no account needed. See Try Nyquest Without an Account.
  • Transparent managed pricing — wallet-funded chat is billed from live resolved-model pricing with a flat 1.2× platform markup, enforced as a tested invariant. BYOK remains markup-free.
  • Accessibility pass — accent colors across all 11 chassis and ~60 UI surfaces now meet WCAG AA contrast as text and as button fills.

June 3–8, 2026 — The Multi-Model Splicer

  • Splicer, Phase 1 — send one prompt to up to 4 models (Pro; 6 on Teams) in parallel and get an agreement score, a synthesized answer, and the exact points of divergence. Any model in the catalog. Live on all 11 chassis. See Multi-Model Splicer.
  • Compress-before-fan-out — your prompt is compressed once, then fanned out, so token savings multiply across the cohort (~40% input savings on long prompts, shown on the consensus card).
  • Auto-router overhaul — tiers now derive from live catalog prices (not model names), routing is domain-aware, and a quality feedback loop re-ranks models within a tier from continuous background evaluations.

May 15, 2026 — Anonymous BYOK

  • Chat with your own key, no account — pass provider keys per-request; they're used in-flight and never persisted, logged, or recoverable. See Try Nyquest Without an Account.

April 30, 2026 — The Help Section Day

The single biggest content shipment in platform history. The help section grew from 7 articles to 67 articles in one day across 8 commits, with the home page going from 1 active card + 4 placeholders to 11 active cards + 0 placeholders.

Help Section — Phase 5 (this commit)

  • FAQ — 50+ scannable questions across 11 topic areas, every Q cross-linked to a deeper article
  • Glossary — 70+ terms alphabetical, definitions for every Nyquest-specific concept (chassis, BYOK, recall chunk, PAT, etc.)
  • What's New — this changelog you're reading

Help Section — Phase 4 ([PHASE 4 COMPLETE])

  • Troubleshooting (8 articles) — sign-in, billing, models, conversations, Agent Mode, error messages, browser/UI, escalation
  • Memory (4 articles) — three-layer model, what gets remembered, recall mechanics, privacy + deletion
  • Artifacts (6 articles) — local-first IndexedDB vault, generated vs saved, in-chat usage, search + tags, export/import
  • Workspace (4 articles) — chassis switching, per-chassis terminology, visual effects catalog, customization scope
  • Developer API (7 articles) — overview, auth, PATs, OpenAI compatibility, full endpoint reference, rate limits, integrations

Help Section — Phases 2 + 3b

  • Getting Started (5 articles) and Chat Basics (6 articles)
  • Projects (5 articles) and Account & Billing (5 articles)
  • Welcome intros for all 10 chassis with chassis-aware CTA
  • Empty-state CTA in the chat input area
  • Per-panel ? button in AdminConsole, ApiKeyManager, ProviderManager, ArtifactVault
  • Mobile drawer for the help nav sidebar on narrow screens

April 29, 2026

Live Support — Phase 4.5 ([COMPLETE])

  • Phase 4.5a — Discord bot connection + migration 013 (database schema for sessions, messages, operator events)
  • Phase 4.5b — session relay end-to-end (user→Nyquest→Discord and back via SSE)
  • Phase 4.5c — admin SupportTab + idle session sweep + emoji-close + status events + status banner

The result: Pro users can open /help/live-support and chat with a real human via a Discord-bridged session. Operators see all active sessions in their Discord channel, reply naturally, react with ✅ to close. Admin console has full session drilldown. Idle sessions auto-mark + auto-close after 24h.

Help Section — Phases 1 + 3

  • Phase 1 — HelpPanel component + Agent Mode articles (7) + /help slash command
  • Phase 3 — HelpButton across all 10 chassis topbars + 4-step welcome tour (first-sign-in modal)

Agent Mode — Steps 12-17

  • Step 12 — Chat/Agent toggle in icon nav rail (Ridgeline + propagated to other 9 chassis)
  • Step 13 — onRunComplete wiring + Agent Mode badge in chat history
  • Step 14-15 — Agent Mode propagated to all 9 non-Ridgeline chassis
  • Step 16 — Agents admin endpoints + admin console drilldown
  • Step 17 — User-facing AGENTS.md guide + STATUS.md MVP-complete

The 9-chassis Agent Mode propagation was the biggest agent-related cosmetic shipment — every chassis now has the Chat/Agent toggle in its icon nav.

Documentation

  • Comprehensive ARCH.md / STATUS.md / AGENTS.md updates capturing the entire arc
  • End-of-day daily 2026-04-29 doc (430 lines)
  • Plan docs locked decisions for help section + Discord live support

April 28, 2026

Agent Mode — Steps 1-11 (the foundational shipment)

A multi-day shipment crystallizing into Agent Mode MVP:

  • Migration 012 — agent_runs + agent_steps tables + projects columns
  • Tool registry types — Tool trait, Registry, PermLevel
  • First 3 tools — project_context, quote_lookup, web_search
  • Step 5 tools — memory_recall, artifact_read, web_fetch (now 6 total)
  • Planner adapter — tier-aware model selection + tool-use loop
  • Runtime loop — end-to-end agent execution
  • Step 8 — billing aggregation + memory recall + 30-day cost ceiling
  • Step 9 — HTTP surface (/v1/agents/run) + bootstrap + smoke-tested live runs
  • Step 10 — run replay endpoints (GET /v1/agents/runs/{id}) + soft-stop endpoint
  • Step 11 — UI integration in Ridgeline (Chat/Agent toggle, run timeline panel)

The first complete agent run with all 6 tools landed on April 28. The frontend integration was Ridgeline-only at that point; chassis propagation came the next day.

Earlier

April 27, 2026 and earlier

The platform's core (auth, chat completions, BYOK providers, billing, memory, projects, artifacts, the 10 chassis, OpenAI compatibility) shipped before this changelog started being curated. See the GitHub commit log for full history.

Notable earlier milestones (approximate dates):

  • April 28 — v2 frontend launched at app.nyquest.ai, replacing v1
  • April 26-27 — final 5 chassis (Quant, Apsis, Garden, Studio, Atelier) shipped
  • April 24-25 — first 5 chassis (Ridgeline, Paper, Matrix, Atlas, Courtroom) shipped
  • April 20-23 — memory three-layer system (facts + summaries + recall chunks)
  • April 15-19 — artifact vault with local-first IndexedDB storage, embedding-based search
  • April 10-14 — BYOK provider management, encrypted-at-rest keys, per-provider model lists
  • April 5-9 — wallet billing via Stripe, Pro tier auto-upgrade on first funding
  • April 1-4 — initial v2 chassis architecture, ChassisRouter + Shell pattern
  • March 2026 — v1 → v2 migration planning, Rust/Axum backend stabilization

What's NOT in this changelog

This is a shipped-features log, not a roadmap. Things deliberately out of scope:

  • Items in flight that haven't merged
  • Documentation-only changes (unless user-facing, like help articles)
  • Internal refactors with no user-facing change
  • Backend-only fixes that didn't change behavior
  • Bug fixes (those go to TROUBLESHOOTING.md)

For roadmap items, see issues on GitHub or talk to a real human.

Subscribing to updates

There's no email subscription yet. To keep up:

  • This page — bookmark /help/whats-new/changelog and check periodically
  • In-app — significant releases trigger a one-time toast notification on next login (rolled out selectively)

A roadmap doc + email subscription are on the wishlist. Tell us if it matters.

Where to next