v2.0.6 — Long-topic speed, overflow retry, full tool results

Patch after v2.0.5: long topics stay usable, overflow compact-and-retries once, and tool/MCP bodies go to the model in full. Upgrade like any patch — pull 2.0.6 or :latest, restart.

Highlights

• Long topics stay responsive while typing and scrolling; highlight and diagrams fill in when scroll stops
• A context-window reject compact-and-retries the send once, even when routine compaction switches are off
• Tool and MCP results are sent in full — no 8k wire rewrite
• The assistant shows Planning next step while the next model call runs after tools
• Memory can search/read large entries, marks tool-tainted cards untrusted, and dreams from recent-message excerpts

Upgrade

• Image: docker.io/lqdflying/chathub:2.0.6 or :latest
• No new database migrations in this tag
• Optional sidecars unchanged (Code Interpreter, MarkItDown)
• Docker Deployment and Upgrades

Note

Full tool bodies raise token use. Token auto-compact and that one overflow retry are the capacity path.

Added

Chat

• Planning-next-step status after tools, with an elapsed clock that survives heartbeat sync

Memory

• searchMemory / readMemory for budgeted recall
• Provenance tags and an untrusted section when a memory write follows external tool output
• Memory dream from a recent-message excerpt when a topic has no summary
• Near-duplicate dream cards are skipped

Context

• Usage-anchored next-request estimate (reload-safe; dispatch-time witnesses)
• Shared overflow classifier and one compact-and-retry per send

Changed

Chat

• Large-topic token estimate runs once; AutoScroll/minimap no longer clone the topic on each keystroke
• Fast scroll keeps bubble chrome and text; mobile overscan uses visualViewport
• Installed PWA: topic sheet unmounts while closed; session rows are not subscribed while you are in a topic (Back stays instant)
• V1Mobile buffers IME composition until compositionend

Tools

• MCP client cache, circuit breaker, abort, and stale-session retry
• MCP descriptions scrubbed; _meta stripped; tool order deterministic

Fixed

Context

• Reloading no longer shrinks the badge onto a stale provider report
• History-limit raise (including from 0) and input-template edits recount the full window
• Prompt-cache rate stays on the active main/portal branch

Chat

• Planning elapsed time no longer resets after sync
• Scroll-mode mermaid no longer mounts the SVG renderer

Docs

• Memory and Context Compaction
• Background Conversation Generation
• Architecture Overview

---

Full diff: v2.0.5...v2.0.6