Files, folders, and images
- Attach files or folders from the desktop composer's paperclip menu, or drag and drop them into a session. Attachments insert local path references; they do not upload files or import them into the graph. Newly attached folders show folder icons, including after reload.
- Tutors can show snapshots of your own images and annotate them in the conversation. Inline learner images and annotations fit within 60% of the window height, without cropping the image or changing the original file.
- Import notes… is now Extract files…, with clearer guidance separating graph extraction from attaching material for a lesson.
- Folder imports skip hidden files and dot-directories, including folders such as
.obsidianand.git.
More control
- Optionally restrict learner-facing file tools to listed folders and their subfolders. Credential and agent-transcript locations are protected even when broad source access is enabled. These are tool-level controls, not an operating-system sandbox.
- Choose Exa, Firecrawl, or hosted Ollama search alongside Brave and Tavily. Configure the selected service's key in Settings; Ollama search uses the hosted service, not a local Ollama installation.
- Edit Practice prompt fragments in Settings alongside Teacher and Prober prompts. Shared fragments indicate which agents use them.
Reliability and polish
- More robust saves for chats, graph data, settings, prompt overrides, and other stored records, with safeguards against interrupted writes and leftover temporary files.
- Claude Code sessions are isolated from unrelated local
CLAUDE.md, output-style, and memory instructions. Claude teaching-message delivery also has an additional guard against tool-only replies that omit learner-facing feedback. - Practice sessions correctly appear as waiting when a problem needs an answer, and adopting live session records no longer duplicates messages or replays consumed wake markers.
- Practice-session sources are hidden in the chains view and node picker; empty source names and book icons are dimmed more clearly.
- Inline code now matches fenced code blocks' background and border. Paperclip menus stay above the composer, including multiline drafts.
- Refreshed user/developer documentation and contribution guidance.
Known limitations
- Image interpretation and annotation still depend on the selected model's capabilities. Older folder references without a trailing separator retain their existing generic icon.
- Previously reported notes-import timeout and audit-lock/recovery issues are not claimed fixed by this release.
- The existing reasoning-chain limitation remains: merging equivalent propositions can drop a necessary premise and incorrectly strengthen a stored inference. This is not fixed in this release.
Thanks to the contributors who supplied code, ideas, reports, and testing, including the work in #81, #157, #201, #327, #332, #364, #409, #420, #421, #428, #431, #433, and #434.