v2.0.0 — ChatHub 2.0

ChatHub 2.0 collects everything since v1.0.0 (May 2026): twenty-eight 1.x GAs, then this tag. Same upgrade as a 1.x patch — pull, restart, let migrations run. Not an API-break bump.

There is no v1.0.29. Next canaries are v2.0.1-canary.N.

Highlights

• Leave is not Stop — the server keeps generating after you switch topic, tap Back, or hide the tab
• Knowledge Base uses dedicated embeddings (OpenAI, Cohere, or Voyage) + pgvector — never the chat API key
• Code Interpreter runs on langgenius/dify-sandbox:0.2.15, not Pyodide in the image
• In-chat Image and Artifacts survive the chat bubble; SVG / HTML / Mermaid render inline
• Skills (SKILL.md, lazy load_skill), two-tier assistant memory, watermark compaction
• First-class OpenAI-compatible and Anthropic-compatible gateways, plus a prompt-cache matrix
• Server + PostgreSQL only — no Electron, no browser-local database
• Huge pastes collapse to PASTED chips (new in this tag)

Upgrade

• Image: docker.io/lqdflying/chathub:2.0.0 or :latest
• Database migrations run on container start
• Optional sidecars unchanged: Code Interpreter (dify-sandbox:0.2.15), MarkItDown
• Set RAG_EMBEDDING_* (or Settings → RAG Provider) before indexing Knowledge Base
• Docker Deployment and Upgrades

Note

Knowledge Base needs its own embedding credentials. Chat API keys are never an implicit fallback.

Added since v1.0.0

Chat

• Background generation (Graphile in PostgreSQL). Stop, retry, rewind, clear, account switch, and topic delete still cancel
• Context Export of the next real model request
• Chat Instruction as a user setting (separate from the assistant role)
• PASTED chips in the composer and topic; Shift-paste via keydown

Knowledge

• Dedicated RAG provider + pgvector HNSW
• Optional MarkItDown sidecar (PDF / Office / HTML / images → Markdown)
• Paginated Markdown chunk drawer; chat attachments stay out of Knowledge

Tools

• Tools Hub: Picbed (S3 image/video), API Tester, related utilities
• MCP OAuth 2.1 (auto-discovery, server-side tokens, refresh on 401) + Settings library
• Agent Skills: global SKILL.md bundles, lazy load_skill
• Code Interpreter on DifySandbox; Graphile can run Python after you close the tab

Media

• Durable in-chat Image (including OpenAI-compatible gpt-image-2)
• /artifacts gallery (native-ratio cards, pager, current-page batch delete)
• Inline SVG, full HTML, and Mermaid with a preview/code switch

Providers

• GPT-5.x (including Sol + reasoning floors), Kimi K2.x/K3, MiniMax, DeepSeek, Zhipu GLM
• OpenAI-compatible / Anthropic-compatible gateways (fixed model lists, opt-in native search)
• Per-provider gear panels; MiniMax history trim before overflow
• Prompt-cache matrix (stable prefixes, Completions vs Responses presets, Anthropic thinking replay)
• Cache hit rate on the token popover

Memory and account

• Topic compaction waits for a high watermark and targets a low one
• Two-tier assistant memory (fixed cards + dynamic rollups); master switch does not disable compaction
• Account isolation across RPC, WebAPI, hydration, session, and async work
• JSON backup/export under Settings → Storage (not object storage)

Changed

• Desktop chat full-width by default; simpler header
• Topic activity shows exact timestamps
• Mobile More is a fifth tab (Knowledge, Tools)
• PWA any vs maskable icons (transparent corners vs full-bleed)
• CHATHUB_DEBUG=1 redacted Pino switch; dedicated generation / tools / knowledge / compaction / cache diagnostics
• Docker startup idempotent schema repair for journal drift
• OpenAI JS SDK v6; Azure AI chat on the Node runtime in the image

Removed

• Electron / desktop and browser-local database editions
• In-image / Pyodide Code Interpreter

This tag (after v1.0.28)

• PASTED chips: composer + topic; Shift-paste; composer-scoped pending dumps; whitespace preserved; visible on focused V1Mobile input
• Mobile action-overlay 16px gutters

Docs

• Wiki Home
• Background generation
• Knowledge Base
• Code Interpreter
• Pasted text

---

Full diff: v1.0.0...v2.0.0