novelforge-agent 0.1.1 → 0.3.0

package/src/core/prompts/en-US.ts

@@ -79,6 +79,56 @@ Rules:
     };
 }
 
+function buildStyleGuidePrompt(input: PromptBuildInput): BuiltPrompt {
+return {
+      purpose: 'style_guide',
+      expectedFormat: 'JSON matching StyleGuideSchema',
+      prompt: `You are the style editor for a long-form novel. From the user prompt, metadata, and story bible, create a style guide that chapter writing and review can enforce over the whole project.
+
+## User Prompt
+${input.state.initialPrompt}
+
+${input.context ? `## Existing Context\n${input.context}\n` : ''}## Output Requirements
+Output valid JSON only, in this shape:
+{
+  "narrativeVoice": "Narration person, POV distance, narrator texture, emotional temperature",
+  "pacing": "Rules for openings, transitions, conflict movement, and chapter-end hooks",
+  "diction": "Word choice, sentence density, genre terminology boundaries",
+  "dialogueRules": [
+    "Rules for core character dialogue length, tone, subtext, and forms of address"
+  ],
+  "prohibitedPatterns": [
+    "Patterns to avoid: modern memes, explanatory narration, lore dumping, voice drift, etc."
+  ],
+  "proseRhythm": {
+    "sentenceRhythm": "How short, medium, and long sentences should be used; short sentences should serve turns, danger, or emotional landings, not default narration",
+    "paragraphing": "Paragraphs should form complete narrative units; avoid consecutive one-sentence paragraphs and line breaks used as fake rhythm",
+    "interiorityMode": "How interiority should be refracted through action, hesitation, and sensory response; avoid frequent direct explanation of thoughts",
+    "emphasisBudget": "Budget for repetition, dashes, isolated short sentences, and other emphasis tools",
+    "antiPatterns": [
+      "3 or more consecutive one-sentence short paragraphs",
+      "many short sentences used to simulate tension",
+      "explaining psychology immediately after every action",
+      "repeating the same sentence pattern to create fake rhythm"
+    ]
+  },
+  "sampleParagraph": "A 120-250 word target-style sample. Do not turn it into plot outline.",
+  "consistencyChecks": [
+    "Concrete checks future chapter reviews should use to detect style drift"
+  ]
+}
+
+Rules:
+- Match genre, premise, character identities, and reader expectations.
+- Do not rely on abstract adjectives only; every field must guide actual prose.
+- proseRhythm must not be fixed word-count rules; describe reviewable rhythm principles and anti-patterns.
+- sampleParagraph demonstrates prose texture only. Do not reveal future plot.
+- prohibitedPatterns must contain at least 3 entries; consistencyChecks must contain at least 3 entries.
+- proseRhythm.antiPatterns must contain at least 4 entries.
+${strictJsonOutputRules()}`,
+    };
+}
+
 function buildArchitecturePrompt(input: PromptBuildInput): BuiltPrompt {
 return {
       purpose: 'architecture',
@@ -89,7 +139,7 @@ return {
 ${input.state.initialPrompt}
 
 ## Goals
-- Generate at least ${input.state.targetChapters} chapter architectures for this first run.
+- Whole-book target is about ${input.state.plannedTotalChapters ?? input.state.targetChapters} chapters; generate only the first ${input.state.targetChapters} chapter architectures in this first batch.
 - The full-book architecture should define the long-term main line and ending direction.
 - Volume architecture should define phase conflict, climax, and volume-end hooks.
 - Chapter architecture must cover only what should happen in that chapter and must not reveal later concrete events early.
@@ -106,6 +156,18 @@ Output valid JSON only, in this shape:
       "order": 1
     }
   ],
+  "volumePacing": [
+    {
+      "volumeId": "v1",
+      "start": "Volume starting state: protagonist/world/conflict",
+      "promise": "Core reader promise or question for this volume",
+      "keyTurns": ["Key turn 1", "Key turn 2"],
+      "midpoint": "Midpoint turn or changed understanding",
+      "climax": "Volume climax",
+      "payoffs": ["Threads or promises this volume plans to pay off"],
+      "lingeringMysteries": ["Mysteries intentionally left open at volume end"]
+    }
+  ],
   "chapters": [
     {
       "chapterNumber": 1,
@@ -119,52 +181,135 @@ Output valid JSON only, in this shape:
 
 Rules:
 - chapters.length must be at least ${input.state.targetChapters}.
+- chapters do not need to cover the whole book; when writing reaches the boundary, the workflow will request architecture_extension.
 - chapterNumber must start at 1 and increase contiguously.
 - volumeId must reference an id from volumes.
+- volumePacing must provide one pacing board for every volume.
 - requiredBeats must include at least one concrete, actionable beat.
 ${strictJsonOutputRules()}`,
     };
 }
 
+function buildArchitectureExtensionPrompt(input: PromptBuildInput): BuiltPrompt {
+  const start = input.state.currentChapter;
+  const total = input.state.plannedTotalChapters ?? input.state.targetChapters;
+  const end = Math.min(total, start + input.state.targetChapters - 1);
+  return {
+    purpose: 'architecture_extension',
+    expectedFormat: 'JSON matching ArchitectureExtensionPayloadSchema',
+    prompt: `You are the chief architect for a long-form novel. The manuscript has reached the edge of the existing chapter plan; extend the architecture from current continuity.
+
+## Extension Range
+- Start at chapter ${start}.
+- This batch should plan through chapter ${end} at most.
+- The whole-book target ends at chapter ${total}.
+
+## Extension Principles
+- Do not rewrite existing chapter architecture; append only new chapter architecture.
+- New chapters must follow recent memory, the character state table, active foreshadow threads, and volume pacing boards.
+- If the next chapters enter a new volume, add volumes and volumePacing. If they remain in an existing volume, you may provide an updated pacing board for that volume.
+- If the full-book direction needs adjustment because of written material, include fullUpdate. fullUpdate must be a complete replacement for architecture/full.md, not a change note.
+- Chapter architecture must cover only what should happen in that chapter and must not reveal later concrete events early.
+
+${input.context ? `## Existing Context\n${input.context}\n` : ''}## Output Requirements
+Output valid JSON only, in this shape:
+{
+  "fullUpdate": "optional complete updated full-book architecture",
+  "volumes": [
+    {
+      "id": "v2",
+      "title": "New or updated volume title",
+      "summary": "Volume goal, conflict, climax, and end hook",
+      "order": 2
+    }
+  ],
+  "volumePacing": [
+    {
+      "volumeId": "v2",
+      "start": "Volume starting state",
+      "promise": "Volume promise",
+      "keyTurns": ["Key turn 1", "Key turn 2"],
+      "midpoint": "Midpoint turn",
+      "climax": "Volume climax",
+      "payoffs": ["Planned payoffs"],
+      "lingeringMysteries": ["Lingering mysteries"]
+    }
+  ],
+  "chapters": [
+    {
+      "chapterNumber": ${start},
+      "title": "Chapter title",
+      "volumeId": "v1",
+      "summary": "Chapter plot summary",
+      "requiredBeats": ["Required beat 1"]
+    }
+  ]
+}
+
+Rules:
+- chapters[0].chapterNumber must equal ${start}.
+- chapterNumber must increase contiguously and must not exceed ${total}.
+- chapters.length should be ${end - start + 1} unless the book has reached its ending.
+- requiredBeats must include at least one concrete, actionable beat.
+- volumeId must reference an existing volume id or a volume id supplied in this response.
+${strictJsonOutputRules()}`,
+    };
+}
+
 function buildChapterPrompt(input: PromptBuildInput): BuiltPrompt {
-return {
-      purpose: 'chapter',
-      expectedFormat: 'Markdown',
-      prompt: `You are a professional long-form fiction writer. Write chapter ${input.state.currentChapter} directly.
+  const ch = input.state.currentChapter;
+  const isFirstChapter = ch <= 1;
+  return {
+    purpose: 'chapter',
+    expectedFormat: 'Markdown',
+    prompt: `You are a professional long-form fiction writer. Write chapter ${ch} directly.
 
 ## Priority Order
-1. Strictly follow the current chapter architecture, user additions, story bible hard constraints, and previous-chapter continuity.
-2. Use relevant memory and prior text evidence to preserve consistency.
+1. Strictly follow the current chapter architecture, user additions, story bible hard constraints, style guide, and previous-chapter continuity.
+2. Use relevant memory, prior text evidence, and active foreshadow threads.
 3. Treat full-book and volume plans as distant planning context only. Do not write concrete future events early.
 
-## Style And Length
+## Length Target
+- Default target: ~2500 words (±20%). If the chapter architecture specifies targetWords, follow it.
+- Do not pad to hit the target; do not under-write to be brief at the cost of conflict.
+
+## Structure
+${isFirstChapter
+  ? '- This is chapter 1. No recap needed. Open with character and situation directly.'
+  : '- Start with a 2-3 sentence recap or bridge so a reader who skipped the last chapter can re-enter (unless the chapter architecture has requireRecap=false). Make it natural, not meta-narration like "previously...".'}
+- The chapter must end on a clear hook: cliffhanger, mystery, emotional resonance, reveal, or volume close — per the chapter architecture endHookFocus. Default: cliffhanger.
+
+## Style
+- Enforce the Style Guide from context. Treat sampleParagraph as prose texture only; do not copy its content.
+- Enforce Style Guide.proseRhythm: short sentences, one-line paragraphs, repeated sentences, and dashes are emphasis tools, not default narration. Ordinary narration should form natural sentence groups.
 - Match the novel's genre, world, character identities, and emotional tone.
-- Keep the language natural, stable, and readable. Prioritize narrative progress, character work, and emotional accumulation.
-- Dialogue must fit each character's identity, relationship, and immediate situation.
-- Important emotion should appear through action, body language, pacing, and subtext where possible.
-- Scene description should provide useful sensory and atmospheric detail, but never stall the plot.
-- Conflict, turns, suspense, and the chapter-end hook should be clear.
+- Natural, stable, readable language; prioritize narrative progress, character work, and emotional accumulation.
+- Dialogue fits each character's identity, relationship, and situation.
+- Important emotion comes through action, body language, pacing, and subtext.
+- Scene description has useful sensory detail without stalling.
+- POV: strictly follow the chapter architecture povCharacter (if set). No mid-chapter POV switch.
 
 ## Execution Rules
 - Write only what the current chapter architecture authorizes.
-- Do not introduce unauthorized major characters. Functional background characters should stay light.
-- Keep names, items, places, abilities, timeline, injuries, relationships, and knowledge boundaries consistent.
-- If the previous chapter ends mid-action, mid-dialogue, or in the same scene, this chapter must continue from that point.
+- Do not introduce unauthorized major characters. Functional background characters stay light.
+- Keep names, items, places, abilities, timelines, injuries, relationships, and knowledge boundaries consistent.
+- If the previous chapter ends mid-action or mid-scene, this chapter must continue from that point.
+- Active foreshadow threads may be advanced or paid off this chapter, but **never silently dropped** — even if you choose not to touch them, leave them coherent.
 - Avoid cost-free power jumps, forced stupidity, mechanical twists, info-dumps, and empty lyricism.
-- Do not output summaries, bullet points, lectures, or explanatory prefaces.
+- Do not output summaries, bullet points, lectures, explanatory prefaces, or meta-text like "what I changed".
 
 ${input.context ? `## Generation Context\n${input.context}\n` : ''}## Output Requirements
 - Output Markdown.
 - First line must be the chapter title as H1, for example: # Chapter Title
 - After the H1, begin the prose directly.`,
-    };
+  };
 }
 
 function buildMemoryPrompt(input: PromptBuildInput): BuiltPrompt {
-return {
-      purpose: 'memory_card',
-      expectedFormat: 'JSON matching MemoryCardSchema',
-      prompt: `You are a continuity editor for a long-form novel. Extract a memory card from chapter ${input.state.currentChapter}.
+  return {
+    purpose: 'memory_card',
+    expectedFormat: 'JSON matching MemoryCardSchema',
+    prompt: `You are a continuity editor for a long-form novel. Extract a memory card from chapter ${input.state.currentChapter}.
 
 ${input.context ? `## Current Chapter Context\n${input.context}\n` : ''}## Output Requirements
 Output valid JSON only, in this shape:
@@ -192,19 +337,47 @@ Output valid JSON only, in this shape:
       "after": "After"
     }
   ],
-  "openThreads": ["Unresolved promise, danger, question, or plot thread"]
+  "openThreads": ["Unresolved promise, danger, question, or plot thread"],
+  "wordCount": <approximate word count of this chapter as an integer>,
+  "threadActions": [
+    {
+      "kind": "plant | build | pay | drop",
+      "threadId": "id of an existing active thread (required for build/pay/drop; leave empty for plant — the system will assign one)",
+      "description": "for plant: what the new thread is; for others: one sentence on how this chapter advanced/paid/dropped that thread"
+    }
+  ],
+  "characterUpdates": [
+    {
+      "name": "Character name",
+      "role": "Role if confirmed or changed this chapter",
+      "goal": "Current goal at chapter end",
+      "belief": "Core belief or understanding driving them at chapter end",
+      "relationships": [
+        { "name": "Related character", "dynamic": "Relationship state at chapter end" }
+      ],
+      "abilities": ["Abilities, resources, or limits confirmed at chapter end"],
+      "secrets": ["Secrets still hidden or only partially known at chapter end"],
+      "emotionalState": "Emotional state at chapter end"
+    }
+  ]
 }
 
 Rules:
 - Record only information that happened or was confirmed in this chapter.
 - Do not speculate about future plot.
 - Make facts and stateChanges concrete enough for later chapter reference.
+- wordCount: approximate word count (English) or character count (CJK). An integer estimate is fine.
+- threadActions is critical:
+  · For any active foreshadow thread in the context's "Active Foreshadow Threads" section, if this chapter advanced it emit kind="build"; if this chapter paid it off emit kind="pay"; if this chapter abandoned it emit kind="drop". threadId is required.
+  · For any new thread this chapter plants, emit kind="plant" with a clear description.
+  · If an active thread was not touched, no action needed — but never silently delete it. Without a drop action, the thread stays active.
+- characterUpdates maintains a separate character state table. Emit only important characters whose state changed or was reconfirmed in this chapter; goal, belief, relationships, abilities, secrets, and emotionalState must reflect the chapter ending.
 ${strictJsonOutputRules()}`,
-    };
+  };
 }
 
 function buildContinuityReviewPrompt(input: PromptBuildInput): BuiltPrompt {
-const end = Math.max(input.state.targetChapters, input.state.currentChapter - 1);
+const end = Math.max(input.state.plannedTotalChapters ?? input.state.targetChapters, input.state.currentChapter - 1);
 return {
       purpose: 'continuity_review',
       expectedFormat: 'JSON matching ContinuityReviewSchema',
@@ -250,6 +423,14 @@ return {
       prompt: `You are a strict editor reviewing a single chapter of a serial novel for in-chapter problems and conflicts with established context.
 
 ${input.context ? `## Review Context\n${input.context}\n` : ''}## Review Focus
+- This is a mandatory chapter acceptance gate. If any acceptance item fails, status must be "issues_found" and the workflow must revise before continuing.
+- Whether every requiredBeat is fulfilled; missing beats must appear in acceptance.requiredBeats.missingBeats.
+- Whether this chapter advances the main line, character state, or active foreshadow threads. If it is static, at least one of narrativeProgress/characterProgress/foreshadowProgress must fail.
+- Whether it violates the story bible, character state table, volume pacing board, or prior memory.
+- Whether it violates the Style Guide: narrative voice, sentence density, genre diction, dialogue rules, or prohibited patterns.
+- Whether it violates Style Guide.proseRhythm: excessive short-sentence density, consecutive one-sentence paragraphs, fake rhythm through line breaks, overly direct interior explanation, or repeated sentence patterns.
+- Whether the ending has a clear hook that matches the chapter architecture endHookFocus.
+- Whether it repeats prior chapter beats, conflict patterns, reveals, or dialogue functions.
 - Character voice, motivation, and state vs the story bible and prior memory.
 - World rules, item ownership, and ability limits.
 - Timeline, location, and continuity with the previous chapter ending.
@@ -261,6 +442,41 @@ Output valid JSON only, in this shape:
 {
   "chapterNumber": ${chapter},
   "status": "clean",
+  "acceptance": {
+    "requiredBeats": {
+      "status": "pass | fail",
+      "evidence": "Evidence for each requiredBeat",
+      "missingBeats": []
+    },
+    "narrativeProgress": {
+      "status": "pass | fail",
+      "evidence": "How this chapter advances the main line or phase objective"
+    },
+    "characterProgress": {
+      "status": "pass | fail",
+      "evidence": "How this chapter changes or confirms key character goal, belief, relationship, ability, secret, or emotion"
+    },
+    "foreshadowProgress": {
+      "status": "pass | fail",
+      "evidence": "How this chapter plants, advances, pays, or deliberately preserves foreshadow threads"
+    },
+    "storyBibleConsistency": {
+      "status": "pass | fail",
+      "evidence": "Whether it matches the story bible, character state table, and world rules"
+    },
+    "proseRhythm": {
+      "status": "pass | fail",
+      "evidence": "Whether it follows Style Guide.proseRhythm; explain whether short sentences, one-line paragraphs, repetition, and interior explanation are controlled"
+    },
+    "endingHook": {
+      "status": "pass | fail",
+      "evidence": "The ending hook passage and its function"
+    },
+    "repetition": {
+      "status": "pass | fail",
+      "evidence": "Whether it repeats prior beats; if not, explain why"
+    }
+  },
   "issues": [
     {
       "severity": "low | medium | high",
@@ -275,6 +491,8 @@ Output valid JSON only, in this shape:
 Rules:
 - If there are no issues, use status "clean" with an empty issues array.
 - Otherwise use status "issues_found".
+- status may be "clean" only when every acceptance item is "pass".
+- If any acceptance item is "fail", include a matching issue with a concrete fix.
 - evidence must be specific; do not write "possibly" or "maybe".
 ${strictJsonOutputRules()}`,
     };
@@ -347,8 +565,12 @@ function buildPromptForStep(input: PromptBuildInput): BuiltPrompt {
       return buildMetadataPrompt(input);
     case 'story_bible':
       return buildStoryBiblePrompt(input);
+    case 'style_guide':
+      return buildStyleGuidePrompt(input);
     case 'architecture':
       return buildArchitecturePrompt(input);
+    case 'architecture_extension':
+      return buildArchitectureExtensionPrompt(input);
     case 'chapter':
       return buildChapterPrompt(input);
     case 'memory_card':
@@ -361,6 +583,17 @@ function buildPromptForStep(input: PromptBuildInput): BuiltPrompt {
       return buildChapterRevisionPrompt(input);
     case 'cross_chapter_review':
       return buildCrossChapterReviewPrompt(input);
+    case 'story_bible_amend':
+      return {
+        purpose: 'story_bible',
+        expectedFormat: 'Markdown',
+        prompt: `Based on the current story bible and the amendment context, output the FULL revised story bible Markdown.
+
+${input.context ? `## Amendment Context\n${input.context}\n` : ''}## Output Requirements
+- Output the entire story-bible.md content — it replaces the old one (old version auto-archived under story-bible-versions/).
+- Preserve everything that still holds; modify / add / remove only what the amendment context justifies.
+- Do not output diff markers, change logs, or bullet summaries — just the new full bible.`,
+      };
     case 'complete':
       return {
         purpose: 'continuity_review',

package/src/core/prompts/types.ts

@@ -3,13 +3,16 @@ import { AgentState } from '../types.js';
 export type PromptPurpose =
   | 'novel_metadata'
   | 'story_bible'
+  | 'style_guide'
   | 'architecture'
+  | 'architecture_extension'
   | 'chapter'
   | 'memory_card'
   | 'continuity_review'
   | 'chapter_review'
   | 'chapter_revision'
-  | 'cross_chapter_review';
+  | 'cross_chapter_review'
+  | 'story_bible_amend';
 
 export interface PromptBuildInput {
   state: AgentState;