An AI Writing Room That Actually Finishes the Book

What it does

ainovel-cli is a Go engine that produces full-length novels from a single-sentence prompt using a trio of LLM sub-agents: Architect plans the world and outline, Writer drafts chapters, and Editor judges quality. A Coordinator agent runs the entire show in a long loop, moving the project through premise, outline, writing, and review phases, while the Host process merely boots the system, recovers from crashes, and renders a TUI for observation.

The interesting bit

The system treats novelist amnesia as an engineering problem. Rather than plotting all 500 chapters upfront, it uses a “compass and rolling horizon” strategy that only details the first two volumes and generates later arcs on demand, referencing compressed summaries and character snapshots. When context windows fill, a four-stage compression pipeline—including a zero-LLM-cost mode that swaps stale messages for existing store summaries—kicks in to keep the Writer from forgetting its own plot.

Key highlights

• Actual agent separation: Architect, Writer, and Editor maintain independent contexts and communicate through structured artifacts in a Store, not by passing chat history around.
• Rolling planning for long fiction: The initial skeleton covers only the first two volumes; subsequent arcs and volumes are expanded dynamically by the Architect based on written content, arc summaries, and character snapshots.
• Step-level crash recovery: Every tool execution writes a checkpoint, allowing the process to resume at the exact plan, draft, check, or commit step rather than re-running entire chapters.
• Seven-dimensional editorial review: The Editor scores drafts on setting consistency, pacing, foreshadowing, and aesthetic quality—including sub-criteria like emotional impact and dialogue distinctiveness—with instructions to cite the original text as evidence.
• Real-time intervention: Users can inject feedback mid-stream without pausing; the system evaluates impact scope and triggers rewrites or polishing on affected chapters.

Verdict

Worth a look if you’re building multi-agent LLM pipelines or curious about automated long-form creative writing. Skip it if you want a lightweight prose toy—this is a heavy-duty orchestration framework with more state machines than a Kafka novel.