amosblomqvist/learnharness v0.4.1 on GitHub

Notes, folders, and chains

Choose Files or Folder from Import notes to import multiple files through a continuous queue, with per-file results and support for adding more files while imports are running.
Import PDFs with selectable text alongside Markdown and text files, up to 25 MB per file. PDF text conversion also gives the tutor's PDF reader more readable words, lines, and prose-column ordering.
Organize sources in persistent folders: rename sources and folders, and drag them between folders.
Navigate folders, sources, chains, and propositions with the keyboard, and search the chains view with scoped Find.
Delete selected sources' graph contributions with confirmation. Session conversations are retained unless you explicitly choose to delete them too; original imported files are not deleted.
Notes extraction guidance now emphasizes substantive knowledge and reasoning supported by the source. Auditor review starts with ranked existing-proposition matches, with targeted follow-up searches. The teaching-session extractor prompt is unchanged.

More responsive chat typing and reliable draft handling, including successive dictation segments.
Improved modified-letter shortcuts on non-Latin keyboard layouts and shortcut rebinding.
Per-model-slot cache-retention preferences: Pi default, don't request caching, short, or long. Support and duration depend on the provider and model; these preferences do not guarantee caching.
Expanded session statistics show an observed, token-weighted cache-hit percentage once positive cache reads are recorded—not a cache-warmth indicator or TTL countdown.
Explicit recovery for saved malformed questions, without automatically restarting recovery when a session is reopened.
Claude Code fixes for unanswered-question idle timeouts and recovery from bare-named harness tool calls.
Home-relative ~/… paths in file tools and PDF reading.
A dismissible update banner for the current version in the current window, plus selection-pill layout polish.
Opt-in background-worker diagnostic capture and refreshed documentation. Diagnostic recordings can contain private learning content and should not be shared without review.

PDF note imports require selectable text; scanned-PDF OCR and image imports are not included.
Extraction quality and latency remain model-dependent; evaluation results were mixed, not a general quality or speed guarantee. The reported NVIDIA import timeout and audit-lock/recovery issues are not claimed fixed.
Malformed-question recovery does not address the separate Gemini thought_signature error when switching models mid-session.
An existing reasoning-chain limitation remains: after equivalent propositions are merged, a necessary premise can be dropped and the stored inference can become incorrectly strengthened. This issue is deferred, not fixed in this release.

Thanks to the contributors who supplied code, reports, and testing, including the work in #198, #261, #270, #301, #302, #324, #326, #347, and #359.