v1 - The starting point
- Add tasks; Chrome built-in AI (Prompt API + JSON schema output) auto-categorizes them
- Duplicate detection badge, split offers with preview, pending queue, localStorage persistence
v2 - Small-model pipeline rewrite
- Per-task multi-phase analysis (category match → ambiguity verify → duplicate check) instead of one giant prompt - built for the small on-device model
- Categories tracked by id with AI-driven merging and a learning notes system prompt
- Queue with progress bar + live status; keep adding tasks while it works, processed one by one
- Paste multi-line text to add each non-empty line as its own task
- Clear All button with confirmation; strategy for re-analyzing existing todos; HTML escaping for safety
- Split suggestions gained accept / keep-as-one review buttons
v3 - Jobs, assistant & concurrency
- Queues persist and resume after reload
- Duplicates grouped into review cards with Combine into one (verbatim, newline-joined - no text lost) / keep separate / per-item "Not a duplicate"
- Per-task rewrite suggestions with accept/reject
- Review cards moved to the top, above the queue
- Assistant box: natural-language commands (add, delete with confirm, search, rename, reorganize) as stoppable jobs with live progress
- Reorganize button with optional guidance note
- Drag & drop tasks between categories
- Settings: up to 2 AI workers, run task analysis and assistant jobs in parallel (default on)
- "Show" buttons on job results that scroll to and flash the task
v4 - Review everything, automate anything
- Auto-accept setting: splits, duplicate combines, rewrites and structural changes apply themselves
- Structural proposals as review cards: move task between categories, rename category, merge categories
- Assistant scans: find duplicates, find splittable tasks, or both in one cleanup pass
- Example chips under the assistant that fill the input when clicked
- Split coverage guard - text the AI dropped from a split is re-added automatically; splits can group related lines into one task
- Per-task check buttons with tooltips: category, split, duplicates, rewrite
- Task text is selectable; dragging moved to a dedicated grip handle
- Best-results tip, credits, disclaimer and this dev log
v5 - Deep reasoning & safety
- Reorganize asks "are you sure" first - from the button and from the assistant alike
- Deep categorization: every task gets a specific theme label, is strictly compared against every category one by one, and extra reasoning steps weigh creating a specific new category vs reusing an existing one - richer categories instead of two giant buckets
- ETA shown for the pending queue and all long-running jobs
- Duplicate checks batched (10 tasks per prompt) so long categories stay reliable - every sibling task is still read by the AI, never fuzzy-matched
- All emojis replaced with inline SVG icons
v6 - The slowdown leak
- Fixed progressive slowdown: Prompt API sessions accumulate every exchange into their context, so reusing one session made each call slower than the last until a refresh. Every prompt now runs on a fresh clone of the base session and destroys it immediately - constant speed however long it runs (rotation fallback for builds without clone())
- Status-line updates coalesced to one DOM render per frame to cut CPU churn during long jobs
- Default AI workers changed to 1 to keep sustained load (and heat) down; bump to 2 in settings if your machine handles it