Three Agents. One Conversation.
The place to ask your world a question. Multi-thread chat routes messages through the intent classifier into local Divinity Engine handlers, with @-mention support for pulling any lore entry straight into the conversation. Writers' Room runs Ishvana, Hawken, and Lagan in parallel when you want three angles at once.
Multi-Thread Conversations
Organize conversations by topic. Each thread persists independently with its own history, agent selection, and context. Switch between a worldbuilding thread, an editing thread, and a research thread without losing your place in any of them.
- Create, Switch, Delete
- Spin up new threads for different topics. Switch between them instantly. Delete old threads when they've served their purpose. Each thread holds onto its full history until you say otherwise.
- Auto-Title
- Threads get titled automatically based on the first message, and the title updates to reflect where the conversation actually ended up. You don't have to stop and name things mid-thought.
- Resizable Sidebar
- The thread list lives in a resizable sidebar (180–360px). Collapse it when you need more room for the conversation. Expand it when you're digging through your history looking for the thread where you worked out the faction hierarchy.
- Semantic Search
- Search across all threads using semantic similarity. You can find that conversation about "the trade route politics" even if you don't remember which thread it was in or which agent you asked about it.
Selectable Specialist Agents
Each agent owns different local handlers. Choose who to talk to based on what you need, use a suggestion chip to force a handler, or use Writers' Room mode to hear from three routes at once.
Ishvana
The orchestrator. Your main assistant for big-picture questions, lore search, cross-module coordination, and delegating a deeper question to one of the specialists when she's not the right agent for it.
Hawken
The writing coach. Style analysis, line-level edit suggestions, craft questions, register calibration. Hawken reads your Legendry before it says anything, and it doesn't draft your book — that part stays yours.
Lagan
The researcher. Web monitoring, scraping, API integration, pattern analysis. Lagan's job is bringing outside information into your world so the creative agents have something specific to work with.
WorldKnowledge
The fact checker. Wikipedia lookup, world-rules queries, and real-world plausibility checks for prose that makes factual claims.
Intent Suggestions & Handler Attribution
Chat now routes by intent instead of persona prompts. Suggestion chips let you choose the handler before sending; otherwise the classifier picks the best local route and asks for clarification when confidence is too low.
- Suggestion Chips
- Chips above the composer expose likely routes such as Hawken scene scan, Lagan research synthesis, or WorldKnowledge query. Click one to bypass classification.
- Confidence Floor
- When the classifier is not confident enough, the engine asks a clarifying question instead of sending the wrong handler.
- Visible Handler Caption
- Assistant bubbles show which agent answered and which
divinity/<handler>produced the response. - Genre Overlay Voice
- Voice comes from the Authored Library and the project's genre overlay, giving consistent local output without per-message persona or stance controls.
@-Mentions & Context Panel
Reference lore entries directly in the message you're typing. Pin context that should stick around for the whole thread. See exactly what the agent is working with before it answers.
- Fuzzy Search Autocomplete
- Type @ and start typing a lore entry name. Fuzzy search matches against the whole Legendry with up to eight suggestions. Pick an entry and its full content gets injected into the conversation context, so the agent has the actual source material in front of it.
- Pinned Tags
- Pin entries you reference a lot, so their content goes with every message in the thread. Useful for keeping the protagonist's character sheet in context across a long brainstorming session without having to @-mention it every single time.
- Context Panel
- A resizable right panel (220–450px) showing sources and message metadata: agent, handler id, classifier confidence, dispatch id, and context quality.
- Message Metadata
- Every response shows which agent answered, which handler ran, the confidence score, runtime, and which sources got consulted. If an answer looks weak, you can see exactly why.
Writers' Room Mode
Ask a question once, hear from three agents at the same time. They run in parallel through a single SSE stream, each bringing their own local handler route to the same prompt. Useful for the kinds of questions where a single answer would be the wrong shape.
Simultaneous Responses
All three agents respond to the same message at the same time. No waiting for Ishvana to finish before Hawken starts. You see three takes side by side, and you pick what's useful.
Color-Coded Bubbles
Each agent gets its own color-coded bubble with an identifying icon. You can tell who said what at a glance, even when the conversation runs long.
Layout Options
Three display modes. Sequential stacks the responses vertically. Columns puts them side by side when you want to compare directly. Threaded turns the responses into a hierarchical tree when the agents start responding to each other.
Single SSE Stream
All three responses arrive through one server-sent events stream, multiplexed by agent ID. No polling, no separate connections, no race conditions to worry about.
Streaming & RAG Context
You can see what the agent is doing. Pipeline stages are visible, lore retrieval is transparent, and the error messages tell you what actually went wrong. No black boxes.
- Thinking Steps
- The pipeline shows in real time. Checkmarks for completed stages, millisecond durations on each step. You can see exactly what's happening between your prompt and the response, which is useful when you're trying to figure out why an answer came back a particular way.
- Loading States
- While the agent works, you see a relevant lore snippet from your world or a thinking animation. The loading state itself tells you something, which beats staring at a blank bubble.
- Grounded Context
- ChromaDB semantic search retrieves the relevant Legendry entries and attaches them to the handler context. The agent has your world data before it answers, so the response is grounded rather than invented.
- Error Handling
- Inline red banners with dismiss buttons. Handler errors include the actual failing route and message, so you know what broke and where to look.