Imported from xqy272/VoiceNovel (
vn_core/planner/AGENTS.md). Install upstream withnpx skills add xqy272/VoiceNovel --skill planner. Copyright stays with the author.
Reading Planner — AI Agent Guide
Purpose
Per-segment speaker attribution, reading style inference, and voice constraint generation. Uses rule-based heuristics by default, with optional LLM-based attribution for dialogue segments.
Key Concepts
- Dual attribution: Rule-based fast path for all segments; LLM path for dialogue when real LLM backend is configured
- Carry-forward: Low-confidence speakers inherit from the previous dialogue speaker
- Voice constraints: Gender, tone tags, and age traits are inferred per speaker for downstream voice casting
- Narrator fallback:
char_narratoris the universal fallback speaker ID
Module: vn_core/planner/
ReadingPlanner(llm, book_model)
plan_chapter(segments, chapter_id, scene_context) -> list[ReadingPlanEntry]
Main entry point. Processes all segments sequentially:
- For each segment, extract speaker candidate via regex
- If dialogue + real LLM: use
_llm_attribution - Otherwise: use
_rule_attribution - Run
_carry_forward_speakerspass over the plan
Speaker Extraction Patterns
"XXX说/喊/叫/道..."— speaker tag before dialogueXXX说/喊道： "..."— speaker tag with colon before quoteXXX说/问/笑道：— speaker tag with colon (no quote required)
Rule Attribution Logic
- Non-dialogue → narrator with confidence 1.0
- Dialogue with matched character in BookModel → confidence 0.85
- Dialogue with unmatched name → confidence 0.6
- Dialogue with no speaker candidate → confidence 0.3-0.4 (carry-forward)
Reading Style Inference
！→ excited, intensity 0.7？→ curious, intensity 0.3……→ hesitant, long_pause- Dialogue adds +0.2 intensity
Dependencies
vn_core.contracts.reading_plan— ReadingPlanEntry, ReadingStyle, Enhancements, VoiceConstraintsvn_core.contracts.segment— Segmentvn_core.llm_gateway— LLMGateway, LLMMessage, LLMRequestvn_core.book_model— BookModel (optional, for character lookup)
