Inside Dynamic Juno

Dynamic Juno is a writing environment powered by cooperative AI specialists that collaborate to help authors structure, research, draft, edit, and critique both creative writing and academic papers. Instead of operating as a single generic chat assistant, the platform uses a Coordinator-Specialist model that routes inputs to highly specific agents, runs background tasks, reads and writes files directly to the project folder, and simulates critique reviews.

System Architecture

When you type a message or select a passage, your prompt is routed dynamically based on a pipeline of analysis, tool invocations, and filesystem writes. Specialist agents invoke dedicated tools to access and cross-reference various context documents (such as outlines, world building lore, and voice profiles) to maintain narrative coherence across drafts.

User Workflow & How to Work

Dynamic Juno is designed to assist you from the initial spark of an idea to a finalized, polished manuscript. By collaborating with specialized AI agents, you can maintain full creative control while leveraging automated planning, drafting, and analysis tools.

The Basics of Working in Dynamic Juno

Understanding the core interface interactions will help you write more efficiently:

Application Modes

Dynamic Juno adapts to three primary project types. Within each, genre-specific leans fine-tune the agent register and craft priorities:

Interactive Workflow Stepper

Explore the recommended stages of writing and editing. Select a phase below, and click individual steps to view their detailed roles, participating agents, and output files.

The Cooperative Agent Network

Dynamic Juno assigns specialized tasks to different specialized agents. When the Coordinator receives a query, it selects the correct agent using intent routing or respects direct user overrides (e.g., typing @writer or @research).

Interactive Flow Simulator

Select a common user action below to trace how a request coordinates between agents, databases, and files on disk. Click Simulate Flow to animate the step-by-step pipeline.

Targeted Passage Editing

When a writer highlights a specific sentence or paragraph in the Reader Panel, enters a prompt (e.g., "Make this sound more foreboding"), and clicks edit, Dynamic Juno triggers a Passage Edit workflow:

By isolating edits to the highlighted selection, Dynamic Juno prevents the LLM from rewriting portions of the chapter that the author is already happy with. This ensures the formatting of the surrounding chapter remains completely untouched.

AI Models & Presets

Dynamic Juno's 14-agent team can run on two model families. A per-agent tier system routes each agent to the right-sized model for its task — agents that perform deep multi-step reasoning get the Pro-tier model; agents with shorter, focused outputs run on Flash. Users choose a preset in Settings (per-user default) or in the composer toolbar (per-session override).

Tier Assignment

Agents that perform complex, multi-step reasoning — Coordinator, Writer, Editor, Research, Peer Review, and Citation Auditor — run on the Pro-tier model. Agents with shorter, focused outputs (Idea, Outline, Debater, Reader, Voice, World Builder) run on Flash 3.5. Lightweight conversational agents (Persona Builder, Reviewer Builder) run on Flash 3. Operator-level AGENT_MODEL_* env vars can override any assignment per-agent.

Context Caching & Session Compaction

Large AI contexts — project files, outlines, drafts, and conversation history — get expensive and slow at scale. Dynamic Juno uses provider-native prompt caching on both Gemini and Claude, plus an active compaction cycle, to keep each turn fast and affordable.

Gemini Context Caching

When a session starts, all background knowledge (ideas, outlines, world details, and previous chapter drafts) is loaded into a single large system context block. This context is cached by the Gemini API for up to 60 minutes. Subsequent requests re-use the cache, avoiding the need to process the entire history each turn.

Anthropic Prompt Caching

When running on the Premium (Claude) preset, a CachingAnthropicLlm wrapper injects Anthropic cache_control breakpoints at two tiers: a 1-hour cache on the tools definition and system instruction prefix (stable across turns), and a 5-minute cache on recent message history (refreshed as the conversation grows). This dramatically reduces response latency on long sessions where the system prefix and conversation history are large.

Active Compaction

If the session history grows too large, the Coordinator runs a compaction routine. It summarizes older conversation turns, storing the raw detail in a markdown log on disk (under projects/<project-name>/chat_history/) and replacing the active session history with the summary. This maintains maximum reasoning capability without hitting LLM context limits.

File System Indexing & Synchronization

Dynamic Juno operates as a local-first file manager. The workspace files on disk are the source of truth:

projects/
└── The Untold/
    ├── ideas.md                 # Concept brainstorming
    ├── outline.md               # Story beats and structure
    ├── metadata.json            # Word count, chapter count, genre, status
    ├── chapters/
    │   ├── chapter-01.md        # Compiled draft
    │   └── chapter-02.md
    └── World/
        ├── settings.md          # Lore and location profiles
        └── characters.md        # Persona descriptions

When you click ↻ Sync Files:

1. The server scans the projects/ directory, auto-registering any new folders as projects.
2. It dynamically scrapes the chapter files (chapter-*.md) and re-calculates chapter and word counts.
3. It saves these metrics to the project's local metadata.json and compiles everything into a central index file (.projects-index.json).
4. The UI refreshes the project list and stats dynamically without requiring a page reload.