Changelog

2026-02-24

New ClaudeTask model with status lifecycle: pending → running → completed/failed
4 API endpoints under /api/claude/tasks/ (create, pending, claim, complete) secured with bearer token auth
"Send to Claude Code" toggle added to the dictation UI — transcribed speech becomes a Claude instruction
transcribe_note management command accepts --claude-task flag to create a task after transcription
thoth_watcher.py: Mac-side polling daemon (stdlib only) that claims pending tasks, runs claude -p, and posts responses back as Thoth notes tagged "Claude Code"
Response notes include the original instruction as a blockquote followed by Claude's response
ClaudeTaskAdmin registered with status filter, instruction excerpt, and raw_id note links
Full flow tested: API → watcher → claude -p → response note created

Full production deployment on Raspberry Pi 5 (8 GB RAM, Debian Bookworm aarch64)
Gunicorn (3 workers) + nginx reverse proxy on port 80
Accessible via Tailscale at http://100.99.187.54/ and local network at http://192.168.86.53/
PostgreSQL 15 database with all 4,344 notes, 40 categories, 1,107 author tags
Systemd service (thoth.service) auto-starts on boot
All data backed up: code on GitHub (mshamma/thoth-project), data rsynced to Pi (~11.4 GB total)
Static files served directly by nginx, media (notes/resources) served from /media/

Django 6 app with Postgres backend serving the archive at http://localhost:8000/
4,344 notes imported from en_backup.db (pickled Evernote objects, XZ-compressed)
16,543 resources linked, 437 audio transcriptions connected
TF-IDF emergent categorization: 40 categories generated from content analysis (scikit-learn), with 3-layer noise filtering (StopTerm table, regex patterns, max_df suppression)
Notebook names preserved as tags only — zero influence on categorization
Yahoo late-1990s directory-style index: grid of category cards with top notes
Noguchi self-organization: viewing a note updates last_accessed, bubbling it to the top of its category
Full-text search via Postgres SearchVector/SearchRank
Note content served from pre-converted static HTML files with resource paths rewritten to /media/
Admin interface for managing stop terms, notes, categories, and tags
Management commands: import_notes, generate_categories

Tech stack page updated with full Django + Postgres design: models, category generation (TF-IDF), Noguchi self-organization, Yahoo directory-style index, noise filtering (StopTerm table + pattern filter + TF-IDF)
Key design decision: notebook names become tags with no organizational weight; categories are emergent from content analysis only

Every note now has a "← Back to index" link at the top of the page
Links use correct relative paths for both standalone notebooks (1 level) and stack notebooks (2 levels)

Added guide.html with full project documentation, usage instructions, and technical details
Added changelog.html (this file) for version history
Both linked from the index header and from each other
Neither file is overwritten by Phase 3 — maintained separately

All 443 audio files transcribed using Whisper base model (437 with content, 6 empty/silent)
269,295 total words across all transcriptions
Each transcription saved as .txt alongside the audio file
Collapsible transcription blocks injected into parent HTML notes

Fixed path encoding: switched from html.escape() to urllib.parse.quote() for src/href attributes, fixing broken images in notes with apostrophes, emoji, or special characters in titles