Magic System Workshop, scene-level contradiction checks, three multi-agent council surfaces, and a fully-local engine that ships smaller, faster, and tells you which handler answered each chat reply.
Writing & Editing
- Theme detection across the manuscript scans for recurring motifs and themes and reports them with cleaner labels (no more "morning", "took", or "stood" cluttering the topic list).
- Theme and motif drop detection uses real chapter-gap math instead of estimated shares, so quiet setup gaps no longer trigger false alarms.
- Editorial report composer lays out motifs, themes, and voice drift findings in their own sections. The voice drift block only shows up when there is something to flag.
- Lorekeeper popover in the editor: hover a flagged finding to see what triggered it without leaving the page.
- Find and Replace built into the editor.
- Breadcrumbs across the workspace so it is always clear where you are in deep nav.
- Source badges on every consistency finding name which check produced it (scene contradictions, ProseGuard, Lore ML, Lorekeeper).
- Unified strictness slider (off, low, medium, high) across every consistency panel. Replaces the older 4-step depth dropdown in Lorekeeper.
Worldbuilding
- Magic System Workshop. Three new tools for building the mechanical content of your world: Ability Workshop (templates, name patterns, random generator), Stat Forge (archetype templates and balance heuristics), and Sheet Studio (template-driven character sheet editor with a bundled template library, replacing the old Sheet Designer).
- Legendry duplicate merge: collapse two duplicate entries into one. Aliases, tags, content, and relationships all union into the surviving entry.
- Entry type prediction: the engine reads the content you have started and suggests the right Legendry entry type (character, location, faction, etc.) when you create a new entry.
Story Structure
- Scene-level contradiction checks catch knowledge violations (a character knowing something they should not), point-of-view violations (slipping into a non-POV character's head), timeline ordering issues, missing-presence and unannounced-presence problems, and dead character appearances.
- Pronoun chains feed presence checks. If a character is referenced as "he" or "she" through a scene without being named, presence detection follows the chain and counts them as present.
- Manuscript scan: theme detection runs across the whole book and reports real progress as it goes.
Councils and Studio
- Three creative surfaces are now structured multi-agent debates. Each one runs a fixed roster of agents through several handler rounds, then a deterministic composer turns the accumulated work into clean drafts plus a full report. Nothing leaves the machine.
- Worldbuilding Council. Archivist runs consistency and contradiction sweeps, Worldsmith runs name generation, Scholar pulls real-world research and synthesis. The composer produces 3 to 8 fully formed lore-entry drafts plus a council report.
- Character Council. Psychologist runs entity extraction and consistency, Dramatist runs the insight scan, Relationship Analyst runs relationship and plot analysis. The composer produces 1 to 3 character drafts plus a character council report.
- Creative Studio. Storyteller seeds the copy, Strategist pulls market and audience research, Analyst runs the marketing copy balance check at high strictness. The composer produces four marketing assets (book blurb, short blurb, tagline, elevator pitch) plus a Studio report.
Chat
- Every chat message now routes through the intent classifier and dispatches to the right specialist handler. The bubble caption tells you which agent answered and which handler did the work, so it is always obvious where the reply came from.
- Suggestion chips above the input let you pick the routing target before sending. Click a chip and the message goes straight to that handler, bypassing the classifier.
- Confidence floor. When the classifier is not sure enough, the engine asks a clarifying question instead of guessing. Tunable in Settings.
- Persona and stance pickers are gone. Voice now comes from the project-genre overlay, not from a per-message dropdown.
- Smart bookmarks and saved-page analysis run on Divinity entity extraction now. Deterministic summaries and tag sets land instantly.
The Divinity Engine
- Fully local. Every external LLM provider is gone. There are no API keys to set, no cloud calls, no per-message costs, and no internet dependency for the engine itself. Everything that used to dispatch to Anthropic, OpenRouter, or Ollama now runs through the bundled handlers on your machine.
- Authored Library expanded by roughly 50 new bundled situation packs across the three council surfaces, marketing copy feedback, chat-dispatch routing replies, and a per-handler chat-summary pack for every routable Divinity handler. Every chat reply now names the routed handler and tells you which surface to run it from.
- Voice-lint passes clean across all 220 bundled situations.
- Genre overlays: 12 genres with multiple variants across 130+ situations. The voice the engine writes in matches your project's genre.
- Intent classifier: chat suggestion chips guess what you are asking for and route to the right agent.
- Visibility learning: rules you engage with stay surfaced. Rules you ignore drift down. Per-rule overrides if you want to force a rule on or off regardless of behavior.
- Pack registry browser installs Authored Library expansion packs from inside Settings, Divinity Engine, Pack Manager.
- Variant review queue: review and approve new Authored Library variants in one place.
- Marketing copy balance check is a new Divinity handler that scores a piece of marketing prose for readability, density, hook strength, length fit, and pain-point coverage. Used automatically by the Creative Studio Analyst, and available in Etherforce for ad-hoc copy passes.
- Smaller, faster Deep Analysis Pack. Coreference resolution swapped from a 1.93 GB Longformer model to a 362 MB distilled one. Roughly 16x faster on CPU, with about 1 F1 cost on OntoNotes for the size and speed gain.
- Embedding and contradiction models swapped to multilingual e5 and mDeBERTa v3 for richer semantic similarity and better cross-language consistency catches.
- Orchestration tunables surfaced in Settings, Performance: utility scorer thresholds and behavior-tree weights. Changes take effect without a restart.
Desktop Experience
- Boot diagnostics. If the engine fails to start, the splash names the failing step and shows the underlying error instead of timing out silently.
- Real error messages everywhere. No more silent fails. No more "something went wrong." The actual cause surfaces wherever you can act on it.
- Background loads (panel mounts, doc switches, tab opens) surface inline if they hiccup, instead of popping the global error modal.
- Tray menu no longer crashes when the main window closes before the tray.
- Studio modal stays mounted when you open it from anywhere in the app. The section editor never strands you on a placeholder button.
- Etherforce gets a per-handler benchmark runner. Pick a handler, choose the fixtures you want to run, capture a baseline, and watch live progress as each fixture runs against the new code path. Diffs and pass / improved / regressed status surface inline so you can spot drift before it ships.