DeepTutor v1.6.1 Release Notes

Release Date: 2026.08.30

v1.6.0 connected the learning workspace. v1.6.1 is about the parts you touch every day: model configuration splits into a credential you enter once and a per-purpose assignment, Settings becomes a navigator instead of a wall of cards, the sidebar files study conversations under their topic and lets you arrange it yourself, and LightRAG runs as a first-party engine rather than through a wrapper. Drop-in — no migrations, no re-index.

What's New

One key, entered once

Model configuration was one layer replicated seven times: every service kept its own profile with its own credentials, so a single OpenRouter key had to be pasted into five separate pages. A connection now holds one vendor credential — API key, version, extra headers, an optional endpoint override — and links it down to every service that vendor can serve. Rotating the key in one place reaches every linked profile; deleting a connection unlinks rather than wipes. Want a different key per service? Simply never make a connection.

Task models

The two calls nobody asks for — naming a conversation after its first exchange, and writing the starting points under the composer — no longer have to run on the model you picked for actual reasoning. The catalog carries a task service shaped exactly like llm, so configuring it is the same act, and a provider can be brought over from the LLM service rather than typed again. Left empty, which is the default, both calls resolve exactly as before.

Settings is a navigator, not a hub

The card grid is gone. A persistent, searchable navigator reaches any page in one click, Models / Chat / Agents categories scroll as one page, and drafts survive navigation — Save draft now actually saves. Opening a connection shows the services it feeds on their own profiles.

Mastery Path files its own conversations

Study conversations were already ordinary sessions carrying a mastery_path_id; the sidebar had everything it needed and used none of it. They now sit under their topic and open in the tutor — with its outline, waypoint header and composer — instead of landing on /home and dropping all three. The composer placeholder is written by the task model against the waypoint in view, and Tab accepts it.

A sidebar you arrange

Feature pages fold into a More group, and both the nav and the chat history reorder by dragging — a press becomes a drag only after it travels a few pixels, so rows stay clickable. The arrangement is a view preference and lives on the machine you arranged it on.

The reading companion is the chat surface

The panel beside an open document is now assembled from /home's own components — the same message list, composer, activity drawer and outline — rather than a 380 px reimplementation of them. What stays local to reading is what actually is: the open material, the selected passage, the linked conversations, and a placeholder written against the page in view. Reading sessions also stop being excluded from the sidebar, and group under their collection.

LightRAG, first-party

The LightRAG engine talks to the SDK directly instead of going through RAG-Anything. pip install 'deeptutor[rag-lightrag]' now pulls only the supported LightRAG release and no longer drags MinerU in transitively — MinerU stays an independently selected parser, and the pdfplumber pin that conflict forced is lifted. Partially accepted ingress and uncertain initial candidates are preserved rather than discarded on append.

Fixes

DeepSeek Pro native web search is enabled and flash thinking is off on OpenAI bindings; Anthropic SDK 1.x takes temperature through extra_body again; Jina embedding retrieval roles are honored. Scanned PDFs explain why they failed to parse instead of reporting nothing found. Auth secret and token files are created owner-only. ClawHub installs are scoped by publisher when several share a slug, newly sent chat branches are selected, and ask_user payloads no longer leak Unicode escapes.

Upgrade Notes

pip install -U deeptutor; Docker users pull ghcr.io/hkuds/deeptutor:latest. No manual migration or re-index is required — existing model profiles keep working untouched, and connections and task models are both opt-in.

• Reinstall the LightRAG extra if you use that engine. pip install -U 'deeptutor[rag-lightrag]' swaps raganything for lightrag-hku. Existing indexes are read as-is. If you relied on MinerU arriving with this extra, install it deliberately now — select it under Settings → Document Parsing and either configure its cloud mode or install its local CLI.
• A connection is optional. Existing per-service credentials are untouched and keep resolving the way they always did. Make a connection only when you want one key to serve several services; linking mirrors the credential down on save, so every consumer keeps reading self-contained profiles.
• The English and Chinese guides have been updated together at deeptutor.info.

Full Changelog: v1.6.0...v1.6.1