v2.4.2 — Performance & UX fixes

What's Changed

Performance: eliminate browser slowdown on large chats

The main chat list was re-rendering every message bubble on every streaming token (~50/sec × N messages). Fixed with:

• React.memo on MessageBubble — only the actively-streaming bubble re-renders; all completed messages are skipped. Possible because updateMessage preserves object identity for unchanged messages.
• Lifted store subscriptions — activeModels, availableModels, fontSize, moderatorId moved from each bubble to MessageList (4×N Zustand subscribers → 4 total).
• Memoized ReactMarkdown components — useMemo((): Components => ...) keyed on uniqueModels gives ReactMarkdown a stable reference across all streaming tokens.
• useMemo for model lookups — allModels, uniqueModels, model computed once per models change, not per token.
• content-visibility: auto on message wrappers — browser skips layout/paint for off-screen messages.

Bug fix: chat input vertical alignment

Text in the chat input was top-aligned relative to the action buttons. Fixed by changing the flex container from items-end to items-center.

Full Changelog

v2.4.1...v2.4.2

v2.4.1 — Token Cost Optimization

Token Cost Optimization

Reduces API token consumption by ~15-20% per debate round without any quality degradation.

Changes

• Prompt caching support: Extracted non-moderator debate rules into a stable prefix constant (PARTICIPANT_RULES_PREFIX) placed at the start of system prompts. All non-moderator models now share an identical ~270-token prefix, enabling provider-side prompt caching (Anthropic, OpenAI, Google)
• Fixed caching blocker: Previous prompts included current time (02:45 PM EET) which changed every minute, preventing any prompt caching. Now uses date-only format (2026-03-16), stable for 24h
• Earlier progressive summarization: Lowered summarization threshold from 20→14 messages and re-summarization trigger from 10→7 new messages. Context is compressed sooner in multi-round debates, reducing input tokens for subsequent rounds
• Condensed prompt formatting: Shortened date format, language rule, and merged stop rules — all without removing any behavioral instructions

What's NOT changed

All debate behavioral instructions preserved intact: thinking style diversity, @mention mechanics, stop rules, quality standards, moderator role structure (OPENING/DISCUSSION/SUMMARY), rebuttal phase addendum.

v2.4.0 — Thinking Styles & Rebuttal Phase

What's New

Cognitive Thinking Styles

Each model is automatically assigned one of 5 thinking styles that shape how it approaches arguments:

• Skeptic — questions assumptions, demands evidence
• Pragmatist — focuses on real-world feasibility and trade-offs
• Visionary — explores unconventional possibilities
• Analyst — breaks down arguments structurally with logic
• Devil's Advocate — deliberately takes opposing positions

Styles are visible as badges on each message bubble.

Structured Rebuttal Phase

A new debate phase between opening arguments and free discussion:

1. All models present their opening arguments
2. Rebuttal round — each model challenges the weakest points of others' arguments using a structured format (Challenge → Assumption → Position)
3. Free discussion continues via @mentions
4. Moderator summarizes, acknowledging genuine disagreements

Prompt Overhaul

System prompts rewritten to encourage productive skepticism:

• "Before building on any argument, identify its weakest assumption"
• "Unexplained agreement is not allowed"
• Models must justify agreement, not just echo others
• Moderator acknowledges disagreements honestly rather than forcing false consensus

Technical Details

• New round structure: Opening (80/75) → Rebuttal (65) → Discussion (70) → Summary (50)
• @mention-triggered discussion suppressed during rebuttal phase to keep it structured
• Thinking styles assigned from unused pool when models are activated, ensuring diversity
• Rebuttal responses count toward discussion budget (1 of 5 per-model, 1 of 8 total)

Files Changed

• src/types/chat.ts — Added ThinkingStyle type
• src/lib/conversationEngine.ts — Thinking styles, rebuttal phase, prompt rewrites
• src/store/chatStore.ts — Auto-assign styles on activation/load/rehydration
• src/components/ChatContainer.tsx — Rebuttal orchestration and prompt injection
• src/components/MessageBubble.tsx — Thinking style badge display
• src/lib/models.ts — Updated model definitions
• src/components/ModelDiscoveryModal.tsx — Minor adjustments

Full Changelog: v2.3.0...v2.4.0

v2.3.0 — Token Budget & Progressive Summarization

What's New

Token-Based Context Budgeting

• Replaced message-count-based context window (slice last 20) with a token-budget system that respects each model's context limit
• Automatic message compression for older messages to maximize context utilization
• Conservative token estimation (3.0 chars/token) optimized for mixed Latin/Cyrillic text

Progressive LLM Summarization

• After 20+ messages, the moderator (or a random active model) generates a semantic summary of earlier discussion
• Summary preserves verbatim: numbers, statistics, code snippets, URLs, and key parameters
• Re-triggers every 10 new messages after initial summary
• Summary replaces old messages in API context only — full messages remain visible in UI

API Cost Control

• Added explicit max_tokens to all API calls (was missing — OpenRouter defaulted to 65536, causing 402 errors)
• Server-side validation with hard cap at 8192 tokens per response
• Summarization disabled in public mode without user's own API key

Bug Fixes

• Fixed OpenRouter 402 error: "You requested up to 65536 tokens, but can only afford 58990"
• Fixed potential empty context when newest message exceeds budget
• Added context_length to all default models for accurate budget calculation

Files Changed

• src/lib/tokenBudget.ts — new token estimation & budget utilities
• src/lib/conversationEngine.ts — rewritten context window builder + summarization engine
• src/app/api/chat/route.ts — max_tokens with server-side validation
• src/lib/streamHandler.ts — maxTokens passthrough
• src/components/ChatContainer.tsx — summarization trigger & lifecycle management
• src/types/chat.ts — ContextSummary type
• src/store/chatStore.ts — contextSummary persistence
• src/lib/models.ts — context_length for default models

v2.2.0 — Onboarding, SEO & Favicons

What's new in v2.2.0

Onboarding

• New how-it-works process cards (2×2 grid) in the empty state: Pick Agents → Pose a Question → Watch the Debate → Steer Consensus
• Service title and description now visible before any agent is activated
• Audience pills: Researchers / Developers / Students / Curious minds
• Fixed content clipping on short screens (min-h-full layout fix)

SEO

• Added keywords, robots: index/follow, metadataBase
• og:image and twitter:card with icon image
• All descriptions synced to "2–5 models"

Favicons

• favicon.ico (16/32/48px) — now works in Safari and all legacy browsers
• apple-touch-icon (180×180) — iOS home screen support
• PWA manifest with icon-192.png and icon-512.png for Android / splash screens
• Theme color: #6366f1 (indigo)

v2.1.0 — Non-blocking Background Retry

What's New

Non-blocking Background Retry

Previously, when a model hit a rate limit or API error, the entire debate stalled — the system waited up to 50 seconds (5s + 15s + 30s) before other models could continue. This release fixes that.

How it works now:

• When a model fails, its concurrency slot is immediately released
• Other models continue the debate without waiting
• The failed model retries in the background on progressive delays (5s → 15s → 30s)
• When the retry timer fires, the model re-enters the priority queue and tries again
• Status messages in chat show retry progress (Model: error, retrying 1/3...)
• If all 3 retries fail, the model is marked as failed and the debate continues without it

Technical details:

• New engine methods: releaseSlot(), scheduleRetry(), clearRetry()
• hasPendingWork includes retrying models to prevent premature moderator/summarizer triggers
• All cleanup paths (stop, new round, round complete) properly clear retry timers
• Guard against double-queuing of retrying models in both analyzeForResponse and queueResponse

Full Changelog: v2.0.0...v2.1.0

v2.0.0 — 3-Phase Debate Engine, @ALL Trigger, Free Models

What's New

3-Phase Debate Engine

Complete rewrite of the conversation engine with structured phases:

• Opening — All models share independent perspectives. Moderator speaks last (priority 75 vs 80).
• Discussion — @mention-driven exchanges bounded by per-model (5) and total (8) caps. No more infinite loops.
• Summary — Moderator provides final synthesis with justified conclusions when discussion settles. Up to 3 moderation cycles.

Moderator as Full Participant

The AI moderator is no longer just a summarizer — it actively participates:

• Shares its own opinion in the opening phase
• Evaluates and challenges other participants' arguments
• Uses @mentions and @ALL to direct discussion
• Provides final summary with conclusions

@ALL Trigger

New @ALL mention addresses all participants at once. Any model that has already spoken will respond. Works for both users and AI participants. Highlighted in the UI.

Free Default Models

Default lineup is now 100% free — works without any API key:

• Nemotron 3 Super 120B (NVIDIA)
• Llama 3.3 70B (Meta)
• Mistral Small 3.1 (Mistral AI)
• Gemma 3 27B (Google)

Automatic Retry Logic

When a model returns an error, the system retries automatically:

• 3 retries with progressive delays (5s, 15s, 30s)
• Status messages in chat for each retry attempt
• Automatic moderator reassignment on permanent failure

Proactive API Key Button

Even in public mode with free models, users can add their OpenRouter API key at any time via the sidebar — unlocking paid models, web search, and bypassing rate limits.

NPX Quick Start

npx degit Lexus2016/Agent-Debate-Consensus my-debate
cd my-debate
npm install
npm run dev

UI/UX Improvements

• Onboarding empty state — Animated arrow + pulsing dots guide new users to activate agents
• "Click to activate" hint — Visible in sidebar when no agents are active
• Export header — Markdown exports now include participant list and moderator info
• File attachments — Attach text files (code, CSV, JSON, up to 100KB) to messages
• Collapsible debate history — Session list in sidebar with one-click restore

Bug Fixes

• Fixed moderator re-trigger loop (moderator no longer speaks 3 times from its own completion)
• Fixed startNewRound not clearing pending state from previous round
• Fixed empty responses not triggering moderator/summarizer
• Separate settle counter for moderator — discussion @mentions no longer consume summary budget
• Removed dead typing indicator code

---

Full Changelog: 355f438...v2.0.0

v1.0.0 — Multi-Agent Debate Platform

Agent Debate Consensus v1.0.0

A multi-agent debate platform where multiple LLMs engage in structured debates on any topic while you moderate.

Features

Core Debate Engine

• Multiple AI models debate simultaneously with priority-based response queue
• Configurable moderator role (auto-assigned or user-controlled)
• Round-based debate flow with automatic moderator summaries
• 10s cooldown between responses (bypassed by @mentions)
• User can send messages at any time during active debate

Model Management

• 4 default models: Kimi K2, Gemini 3 Pro, Claude Haiku 4.5, Grok 4.1 Fast
• Model Discovery modal — browse and add any model from OpenRouter catalog
• Configurable max active models via MAX_ACTIVE_MODELS env var
• Per-model failure tracking with automatic recovery

Chat Interface

• @mention autocomplete with model color highlighting
• File attachments — text/code files up to 100KB
• Web search toggle (OpenRouter :online suffix)
• Collapsible debate history with session management
• Temperature presets: Creative / Balanced / Precise
• Adjustable font size
• Copy individual messages as Markdown
• Export full debate as .md file
• Dark / Light theme toggle
• Responsive design with collapsible sidebar

Streaming & Performance

• Server-Sent Events (SSE) streaming via OpenAI SDK
• Per-model AbortController for individual stream cancellation
• Simulated reading delay for natural conversation flow
• Context window management (last N messages per model)
• Extended thinking / reasoning support with collapsible display

Deployment

• Supports both private (self-hosted) and public modes
• Public mode: users provide their own OpenRouter API key
• Welcome/onboarding screen for API key entry
• Zustand state persisted to localStorage

Tech Stack

• Next.js 16 (App Router)
• TypeScript
• Tailwind CSS v4
• Zustand (state management)
• OpenRouter API via OpenAI SDK

Getting Started

git clone https://github.com/Lexus2016/Agent-Debate-Consensus.git
cd Agent-Debate-Consensus
npm install
cp .env.example .env  # Add your OPENROUTER_API_KEY
npm run dev