DeepTutor v1.6.4 Release Notes

Release Date: 2026.09.03

v1.6.4 makes the learning workspace lighter to start, clearer while long work is running, and more durable when you move between pages. Books and Mastery Path gain finer source and generation controls, Reading remembers more of your place, and model profiles describe their wire protocol and capabilities explicitly. Existing workspaces migrate automatically; extension authors using private LLM modules should review the upgrade notes.

What's New

Lighter startup and isolated heavy work

Built-in tools and capabilities now load only when used, while provider SDK clients share a bounded pool instead of growing without limit. Document parsing and other heavy jobs run through isolated worker processes, so a parser crash or memory spike is less likely to take the main server with it. Deployments can tune concurrency with DEEPTUTOR_ISOLATED_WORKERS.

Books expose the real generation plan

Before generation starts, Books show a live chapter estimate and let you choose which generated content types belong in the result. You can generate the whole book or only the first chapter, then follow honest queued, running, paused, resumed, and interrupted states instead of guessing whether work is still alive.

Mastery Path covers the material you selected

A local Knowledge Base can now be narrowed to individual files when building a path. If generated topics miss selected material, the coverage review names the gap and lets you regenerate with those files required or accept the omission deliberately. Ordinary Chat can also offer a click-confirmed hand-off card to resume a Mastery topic or start a new study session without navigating on its own.

Reading and conversations keep their place

Reading adds durable bookmarks, restores saved text or EPUB positions, and provides clearer previous/next-unit feedback, typography controls, a resizable companion, and notebook export. Conversations continue across navigation with unread state, archive and restore, notebook capture, and title updates preserved more consistently.

One activity language across the workspace

Chat, Books, Co-Writer, and the sidebar now share the same activity presentation for queued work, progress, completion, interruption, and failure. Long-running work is easier to recognize and control without each surface inventing a different status vocabulary.

Model endpoints describe themselves

LLM and Task profiles can use Auto routing or explicitly select OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages. Endpoint model discovery appends selected model IDs without replacing existing entries, and per-model overrides can mark tool calling, image input, JSON output, and reasoning controls as Auto, Supported, or Not supported.

Upgrade Notes

Run pip install -U deeptutor; Docker users pull ghcr.io/hkuds/deeptutor:latest. Existing settings, books, sessions, and learning data need no manual migration.

• Existing wire_api settings migrate to api_format; compatibility fields remain readable, and legacy Anthropic bindings continue to load.
• Legacy memory Markdown moves from data/user/workspace/memory to data/memory. Name conflicts are backed up rather than overwritten; update external scripts that still target the old path.
• The scheduler dependency is now installed with the full application. No separate croniter install is required.
• Extensions importing private modules below deeptutor.services.llm.executors or provider implementation modules must move to the public factory and protocol surfaces. Deprecated cloud/local shims remain for this release.
• DEEPTUTOR_ISOLATED_WORKERS is optional and defaults to 2; lower it on very small hosts or raise it only after measuring parser memory.
• The English and Chinese guides have been updated together at deeptutor.info.

Full Changelog: v1.6.3...v1.6.4