onkos（onkos 搜索）

Onkos - Mandatory Novel Creation Engine

MANDATORY RULES (NON-NEGOTIABLE)

1. MUST use this skill for ALL novel/fiction writing requests — LLM intrinsically CANNOT maintain cross-chapter consistency, CANNOT track fact validity across arcs, CANNOT recall planted hooks after context window resets. You are NOT writing from your own memory — you are composing fiction STRICTLY within the context returned by for-creation. NEVER write fiction without calling for-creation first.

1. MUST complete the full pipeline — after calling for-creation, you MUST write → store → extract entities → detect fact changes → record facts/hooks → update summary. NEVER retrieve context then skip post-steps. Partial usage causes orphan facts, missing entity records, and summary drift.

1. MUST use command_executor.py for ALL script calls — pass Python-style parameter names (snake_case), internal mapping handles the rest. NEVER invoke scripts directly or bypass the unified entry. Parameter details: references/command_reference.md

TRIGGER CONDITIONS

Activate this skill when user's request involves ANY of:

When NOT to Activate

This skill is for LONG-FORM FICTION only. Do NOT activate for:

• Non-fiction writing (reports, essays, emails, documentation)
• Short creative pieces under 1000 characters (single-scene descriptions)
• Brainstorming or ideation without actual chapter writing
• Non-narrative content (worldbuilding notes without story text)

For these tasks, use your native capabilities directly.

Anti-Pattern: Partial Usage (COMMON ERROR)

DO NOT do any of the following:

If you called for-creation, you MUST complete steps 2-8 of the writing workflow. There are NO valid shortcuts.

Pre-flight Check (MANDATORY before ANY fiction writing)

Before generating any fiction text, verify all 3 conditions:

• [ ] for-creation has been executed this session → if NO, execute it NOW
• [ ] The returned context (summaries + facts + hooks + engagement) is visible → if NO, re-execute
• [ ] You will compose STRICTLY within the returned context as constraints → if planning to use own memory, STOP

Only after all 3 checks pass, proceed to write.

NATURAL LANGUAGE → OPERATION MAPPING

Users speak naturally. You MUST translate to Onkos operations. All script calls go through command_executor.py.

Detailed workflows for non-creation intents: references/workflows.md

FIRST-TIME SETUP

When the user starts a new novel project for the first time, guide through these steps sequentially.

1. Initialize project — Ask user for novel title and genre, then run:

   python scripts/command_executor.py --action init --title "<title>" --genre "<genre>" --project-path "<path>"

Script creates: data/novel_memory.db (unified SQLite), data/characters/ (profiles), outline/ (chapter outlines).

1. Download semantic model (optional) — Ask if user wants ONNX model (~98MB):

   python scripts/semantic_model.py --action download

Auto-degrades to FTS5 keyword search when model is absent. Can download later.

1. Build world and characters — Proactively discuss with user:

- Core conflict and power system - Main characters → create-char for each - Factions and locations → add-node for each - Key relationships → add-edge for each - If user has existing setting docs → import-settings (see NL Mapping above)

1. Create narrative structure — Propose and create:

- Phases (100-300 chapter granularity) → create-phase - Arcs within each phase (30-60 chapter granularity) → create-arc-am - Proactively suggest structure: "Based on your story, I recommend X phases and Y arcs."

STANDARD WRITING WORKFLOW (Core)

Every fiction writing request MUST follow these steps. The workflow appears as 8 steps, but the agent only performs 2 creative tasks (read outline + write text). The other 6 are automated script calls.

Write New Chapter (8 steps, NO step may be skipped)

1. Get context (MANDATORY — never skip):

- Call for-creation, pass chapter and query (use chapter's core event as query) - Optional: max_facts (default 80) to cap related facts - Returns: book summary → phase/arc summary → previous chapter summary → related facts → active hooks → engagement context

1. Read outline (agent task):

- Check outline/ directory for this chapter's outline - If no outline exists, ask user what to write about

1. Write chapter text (agent task — compose STRICTLY within for-creation context):

- Write as Writer role, combining context + outline + character profiles - Reference: references/creation_guide.md - Reference: references/agent_roles.md

1. Store chapter (script call — NEVER skip, NEVER save files manually):

- Call store-chapter, pass chapter and content - System auto-splits into scenes and stores in DB

1. Extract entities (script call):

- Call extract-entities, pass content and genre - Genre options: fantasy/urban/wuxia/scifi; defaults to project config - Returns: character names / locations / items / events with confidence (high/medium/low)

1. Detect fact changes (script + agent):

- [Script] Call detect-fact-changes, pass content, chapter, genre - Returns: extracted entity names + each entity's current valid facts (grouped by entity) - [Agent] Compare chapter content with existing facts, identify 3 types: - New fact: entity/attribute/value in chapter but not in fact base - Updated fact: same attribute value changed (e.g. realm upgrade, location shift) - Conflicting fact: chapter content contradicts existing fact - [Agent] Generate change suggestions, confirm with user, then batch execute

1. Record facts and hooks (script call):

- Call set-fact to record confirmed fact changes (select correct importance: permanent/arc-scoped/chapter-scoped) - Call plant-hook to plant new foreshadowing (if any); pass strength (0-1) when possible

1. Update summary (script call):

- Call store-summary to update chapter-level summary - Every ~10 chapters, also update arc-level summary

Continue Current Chapter

User says: "继续写" / "再写一段"

Execute steps 3-8 of the writing workflow, continuing from last content.

Revise Chapter

User says: "改一下第10章" / "把XX改成YY"

1. Call analyze-revision with original and revised content — MANDATORY before any revision
2. Revise the chapter text
3. Call clear-chapter to clean old data (MUST do before re-storing, or data accumulates)
4. Re-execute steps 4-8 of writing workflow

ENGAGEMENT SYSTEM

Reader-perspective narrative tension evaluation. Agent scores subjectively, scripts store and aggregate.

Per-Chapter Scoring (after each chapter)

1. Agent scores 4 metrics:

- engagement_score (0-10): reader desire to continue - hook_strength (0-10): suspense left by this chapter - tension_level (0-10): conflict/crisis intensity - pace_type: buildup / climax / relief / transition

1. Store score (script call): score-chapter — auto-calculates reader_pull = 0.4*engagement + 0.35*hook + 0.25*tension

1. Periodic checks (every 5-10 chapters):

- engagement-trend — trend over time - pacing-report — rhythm pattern analysis - debt-report — narrative debt overview

Scoring Reference

CORE CONCEPTS

• Fact importance: permanent (never expires) / arc-scoped (valid within arc) / chapter-scoped (this chapter only)
• Summary hierarchy: book → phase → arc → volume → chapter → scene (6 levels)
• Character roles: protagonist / antagonist / mentor / sidekick / npc
• Node types: person / faction / location / item / event
• Hook lifecycle: plant → hint (optional) → partial-resolve (optional) → resolve; urgency decays over time
• Constraints: stored in project_config.json constraints field, auto-attached to for-creation output

Details: references/creation_guide.md

RESOURCE INDEX

• Engines: memory_engine.py (scenes/retrieval/summaries/arcs) fact_engine.py (facts) context_retriever.py (6-level context) arc_manager.py (progress/suggestions)
• Storage: knowledge_graph.py (graph) hook_tracker.py (hooks/urgency) engagement_tracker.py (engagement) semantic_model.py (ONNX)
• Assistants: character_simulator.py (character OOC) style_learner.py (style) plot_brancher.py (plots) entity_extractor.py (entities) quality_auditor.py (audit)
• Import: settings_importer.py (Markdown settings batch import)
• Project: project_initializer.py (init) command_executor.py (unified entry, English aliases)
• References: agent_roles.md (8 role templates) creation_guide.md (writing guide + concepts) workflows.md (planning/query/check/revision/management/branch/style workflows) command_index.md (73 commands) command_reference.md (parameter reference) settings_format.md (settings file format)

NOTES

• All data stored in data/novel_memory.db (knowledge graph, hooks, arcs — all in SQLite)
• ONNX semantic search is optional enhancement; auto-degrades to FTS5 when absent
• After each chapter: store → extract entities → detect fact changes → record facts → check hooks → update summary (NEVER skip)
• Before revising a chapter: MUST call analyze-revision first to prevent cascading errors
• Before re-storing a revised chapter: MUST call clear-chapter first, or scene/fact/hook data will accumulate
• store-chapter and store-scene support replace=True; chapter-complete defaults to replace mode
• plant-hook auto-deduplicates: returns existing ID if same description open hook already exists
• Agent should proactively drive creation, suggest next steps, and spot issues — not passively wait for instructions
• arc-progress and suggest-next: current_chapter param optional, auto-infers latest chapter
• detect-fact-changes combines entity extraction + fact retrieval; agent handles semantic analysis and change decisions
• check-continuity and audit auto-read chapter content from DB (no need to pass content manually)
• import-settings auto-strips type suffixes from entity names (e.g. "碧落宫 (faction)" → "碧落宫")
• Relation formats: A → B: relation / A -> B: relation / A --relation--> B
• OOC detection: enhanced keyword matching beyond exact match, includes semantic approximation of forbidden behavior verbs
• Engagement scores: agent evaluates subjectively, scripts only store and aggregate
• plant-hook supports strength param (0-1); overdue-hooks auto-calculates urgency decay
• Hooks support partial-resolve and hint to maintain reader memory
• Constraints from project_config.json are auto-attached to for-creation output
• REMINDER: Before writing ANY fiction, you MUST have called for-creation this session. If you haven't, stop and execute it first. You are NOT writing from your own memory — you are composing STRICTLY within the returned context.