InkOS

InkOS is a fully autonomous AI Agent system for long-form novel creation, written in TypeScript (Node.js 22+, pnpm monorepo). Current version v1.3.6, published as @actalk/inkos on npm.

Core Architecture: Multi-Agent Pipeline#

The system comprises 10 specialized agents forming a complete writing pipeline:

• Radar: Scans platform trends and reader preferences to guide story direction (pluggable)
• Planner: Reads author intent + current focus + memory retrieval results to produce chapter intent
• Composer: Selects context from all truth files by relevance, compiles rule stack and runtime artifacts
• Architect: Plans chapter structure—outline, scene beats, pacing control
• Writer: Generates prose based on curated context (word count governance + dialogue guidance)
• Observer: Extracts 9 types of facts from prose (characters, locations, resources, relationships, emotions, information, foreshadowing, time, physical states)
• Reflector: Outputs JSON delta, validated by Zod schema then immutably written to state
• Normalizer: Single-pass compress/expand to bring chapter word count into allowed range
• Auditor: Performs 33-dimension checks against 7 truth files
• Reviser: Fixes issues found by auditor; critical issues auto-fixed, others flagged for human review

When auditing fails, the pipeline automatically enters a "revise → re-audit" loop until critical issues are cleared; non-critical issues pause for human review.

Long-Term Memory: 7 Truth Files#

• current_state.md: World state—character positions, relationship networks, known information, emotional arcs
• particle_ledger.md: Resource ledger—items, money, supplies with quantity tracking and decay
• pending_hooks.md: Unclosed foreshadowing—setup, promises to readers, unresolved conflicts
• chapter_summaries.md: Chapter summaries—characters present, key events, state changes, hook dynamics
• subplot_board.md: Subplot progress board—A/B/C line statuses, stagnation detection
• emotional_arcs.md: Emotional arcs—per-character emotion tracking and growth
• character_matrix.md: Character interaction matrix—encounter records, information boundaries

Since v0.6.0, the authoritative data source has migrated to story/state/*.json (Zod schema strict validation), with Markdown retained as human-readable projections. Node 22+ automatically enables a SQLite temporal memory database (story/memory.db) supporting relevance-based retrieval.

Quality Control & Anti-Detection#

• 33-dimension continuity auditing: Comprehensive validation against 7 truth files
• Anti-AI traces: Built-in detection for high-frequency words, monotonous sentence patterns, over-summarization; revise --mode anti-detect for dedicated rewriting
• Word count governance: Separate Chinese/English counting logic (zh_chars / en_words), conservative management, max 1 correction pass
• Reliability: State snapshots and rollback, file locking for concurrent write prevention, cross-chapter duplicate detection, 11 hard rules with automatic spot-fix

Creative Assistance#

• Style mimicry: inkos style analyze extracts statistical fingerprints + LLM style guides; inkos style import injects specified book styles
• Continuing existing works: inkos import chapters imports existing text, automatically reverse-engineering 7 truth files
• Fan fiction: inkos fanfic init supports canon / au / ooc / cp modes with dedicated audit dimensions
• Creative briefs: --brief my-ideas.md passes world-building/character documents
• Input governance: author_intent.md (long-term direction) + current_focus.md (near-term focus) + per-chapter runtime artifacts

Interaction Modes#

1. Full pipeline (one-click): inkos write next
2. Atomic commands (composable): plan chapter / compose chapter / draft / audit / revise
3. Natural language Agent mode: inkos agent "help me write an urban cultivation novel"
4. TUI dashboard: inkos tui (full-screen interactive, slash command autocomplete)
5. Web Studio: inkos or inkos studio (Vite + React + Hono, book management, chapter review, progress monitoring, data analysis)

LLM Configuration & Model Support#

Built-in Provider Bank: Google Gemini, Moonshot, MiniMax, Zhipu, Bailian, DeepSeek, SiliconFlow, PPIO, OpenRouter, Ollama, CodingPlan, etc., compatible with any OpenAI API specification. Supports per-agent model routing (inkos config set-model), with fallback parsers and automatic stream interruption recovery.

Deployment & Notifications#

• Supports inkos up background daemon for continuous auto-writing
• Push notifications: Telegram / Feishu / WeCom / Webhook (HMAC-SHA256)
• Export formats: txt / md / epub
• Multi-language writing: Chinese (default) and English (--lang en)
• Published as OpenClaw Skill (clawhub install inkos)

Note: The project's self-reported benchmark data (31 chapters / 452,191 words / ~14,500 words per chapter average) has not been independently verified by third parties.