AIchat_v1.3

IMPORTANT: If you were already using version 1.2, delete system_prompt.txt from the Some_Modeling_Agency\BepInEx\plugins\AIchat folder, as that file has been updated in the code. If you had previously edited system_prompt.txt, you can leave it as is; however, if the prompt causes issues, allow a new one to be created automatically.
The readme has also been updated with new usage instructions.
Ollama server URL changed from http://localhost:11434/api/generate to http://localhost:11434/api/chat so if you're going to use this version, make sure to change it.

New Files Created

[PsychologyEngine.cs] — The Brain

The central "Stream of Consciousness" engine. On character bind:

1. Reads 16 base-game personality modifiers from Personalidad
2. Performs dice rolls using personality traits as seeds (±15% variance)
3. Initializes 10 keyword categories and extended mood parameters

10 Keyword Categories (0-100 intensity, updated per message):

• Flattery, Intimidation, Seduction, Professionalism, Humor
• Empathy, Dominance, Vulnerability, Negotiation, Provocation

Personality Accelerators — permanent traits that multiply keyword effects:

• High pervertido → seduction keywords have 2x effect on atmosphere
• High sumiso → dominance keywords boost consent faster
• High timido → intimidation increases disgust faster

Extended Mood Parameters (mod-only, not in base game):

• Boredom: increases on bland messages, decreases on interesting ones
• Disgust: provocation + intimidation increase it
• Thawing: seduction + flattery warm her up
• Relief: empathy + humor bring relief

Atmosphere (0-100): Cold → Neutral → Warm → Hot → Intimate

Modeling Stages (replaces old binary PreferencesDiscussed):

• Discuss → Photos → Posing → Lingerie → Erotic
• Auto-progression based on atmosphere + thawing thresholds

---

[ConversationMemory.cs] — Long-Term Memory

• Rolling summary: injected as [MEMORY] block in system prompt
• Topic tagging: auto-detects important events (agreements, refusals, player name)
• Persistence: saves/loads per character to AIchat/memory/mem_{charId}.txt
• Max 20 tags to prevent token bloat

---

[SpamFilter.cs] — Event Deduplication

• Cooldown: minimum 1.5s between system messages
• Dedup: identical messages blocked within 10s window
• Batching: rapid-fire events grouped into ⚡ (3 events) msg1 | msg2 | msg3
• Toggle: on/off in settings panel

---

Modified Files

[ChatWindow.cs] — Major UX Overhaul

+ Enter to send messages (was already Enter, now explicit)
+ Shift+Enter for line breaks in input
+ Auto-focus on text input when panel opens
+ Focus removed from input after sending
+ Animated "⏳ Thinking..." status with provider name
+ Chat history persistence per character (AIchat/history/chat_{charId}.txt)
+ Debug panel (toggle via "Debug" button):
    - Psych state values (atmosphere, boredom, disgust, thawing, relief)
    - All 10 keyword strengths
    - Personality accelerators
    - Memory tags and turn count
+ Stage + Vibe display in chat header
+ Spam filter toggle in settings
+ "AllowOpenMouthCommand" toggle in settings
- Removed PreferencesDiscussed gate (stages replace it)

---

[PromptBuilder.cs] — Token-Optimized Rewrite

Old prompt: ~1500-2000 tokens of verbose prose.
New prompt: ~600-800 tokens of compact key:value notation.

- "You are wearing the following pieces of clothing: camisa, falda."
+ [CLOTHES] camisa, falda

- "[CORE PERSONALITY TRAITS] Perverted, Exhibitionist, Submissive, Extrovert(80), Submissive(60)..."
+ [PERSONA] Perverted,Exhibitionist,Submissive Perv:80 Sub:60 Dom:20 Shy:15 Extro:85

- "[HIDDEN TRAITS & KINKS (0-1)] gustoPorPervertidos(0.800), facilidadParaDesHielar(0.600)..."
+ [TRAITS] LikesPervs:0.80 EaseOfThawing:0.60 Patience:0.45 ...

- "[EMOTIONAL STATE (0-100%)] Arousal: 45%, Pleasure: 30%, Joy: 60%..."
+ [FEELING] She is currently aroused, content, warming up to intimacy.

+ [VIBE] Warm(55) [XMOOD] Thawing:40 [KWDS] Seduc:35 Flat:28 [STAGE] Photos
+ [MEMORY] Player complimented her several times...
+ [FACTS] She agreed to photos. Player's name: John.

Key changes:

• Human-readable trait names: gustoPorPervertidos → LikesPervs
• Top 15 traits only: ranked by distance from 0.5 midpoint (saves tokens)
• Natural-language mood: replaces numeric emotion events
• Work info injected: Hired:Y Salary:$100 Commission:30% Job:modeling
• Memory block: summary + fact tags from ConversationMemory
• Backward-compatible: still reads system_prompt.txt templates with new placeholders

---

[AIResponseProcessor.cs]

+ PsychologyEngine processes AI responses (atmosphere shifts)
+ SpamFilter gates system messages (dedup + cooldown)
+ ConversationMemory auto-tags AI responses
+ History auto-saved after each AI response

---

[DialogBehaviour.cs]

+ PsychologyEngine instance (initialized on character bind)
+ ConversationMemory instance (bound + loaded per character)
+ SpamFilter instance (flushes batched events in Update)
+ GetCharacterRoot() helper method
+ AllowOpenMouthCommand config field
+ SpamFilterEnabled config field
+ ModelingStage replaces PreferencesDiscussed

---

[ChatMessage.cs]

+ DateTime Timestamp field
+ ToLine() serialization (pipe-separated format)
+ FromLine() deserialization

---

[CommandExecutor.cs]

+ open_mouth command — opens jaw via IControladorDeJaw + ModificadorDeFloat (25° angle)
+ close_mouth command — closes jaw (0° angle)
+ AllowOpenMouthCommand toggle check
+ using Assets._ReusableScripts.CuchiCuchi._CharactersBasics

The mouth commands work by:

1. Finding IControladorDeJaw on the character root
2. Getting the jaw's x-axis OrdenesDeID with obtenerMaximoAbsoluto order type
3. Obtaining a ModificadorDeFloat and setting its value to 25° (open) or 0° (close)

---

Architecture Diagram

graph TD
    A["Player Message"] --> B["ChatWindow"]
    B --> C["PsychologyEngine.ProcessPlayerMessage"]
    C --> D["Keyword Detection"]
    C --> E["Atmosphere Shift"]
    C --> F["Mood Shifts"]
    C --> G["Stage Progression"]
    B --> H["API Client"]
    H --> I["PromptBuilder.GenerateSystemPrompt"]
    I --> J["Compact Identity + Traits"]
    I --> K["PsychEngine.GenerateContextBlock"]
    I --> L["PsychEngine.GenerateMoodDescription"]
    I --> M["ConversationMemory.GenerateMemoryBlock"]
    H --> N["AI Response"]
    N --> O["AIResponseProcessor"]
    O --> P["PsychEngine.ProcessAIResponse"]
    O --> Q["SpamFilter.ShouldShow"]
    O --> R["ConversationMemory.AutoTag"]
    O --> S["CommandExecutor"]
    S --> T["Game Commands"]
    T --> U["open_mouth / close_mouth"]
    T --> V["Existing commands"]

AIchat v1.2 little fixes

[x] Create API/AIResponseProcessor.cs
[x] Refactor GemmaClient.cs, OllamaClient.cs, GeminiClient.cs to use the shared processor
[x] Fix "standA" bug in CommandExecutor.cs
[x] Rebalance Sentiment rage math in SentimentAnalyzer.cs
[x] Implement Harmony hook for SetTactilTo in HarmonyPatches.cs

AIchat v1.2

• Redefined the external system_prompt.txt file. If you already have something written in that file, move it out of the AIchat folder so a new one containing more data is generated.
• Fixed the bug that prevented the AI from detecting whether there was a new model and caused it to continue using the previous model's history
• Fixed “move to wall” and “move to desk” (I haven’t yet implemented the ability to do specific wall or desk pose).
• Added Gemma (LM Studios); not yet tested.
• Now, if you make a positive, romantic, or insulting comment to the model, her consent, arousal, or rage will change.

For developers:

• PromptBuilder.cs format instruction — Added explicit rules: "comando" must be null during normal chat, never invent commands not in the allowed list, never use "comando" in dialogo text, plus concrete examples (null vs actual command)
• GeminiClient.cs maxOutputTokens — Changed 150 → 512 in both the main response call (line 50) and the summarization call (line 178)
• GeminiClient.cs safer comando extraction — Now uses JsonHelper.ExtraerValorDeClaveJsonDesdeElFinal("comando") instead of ExtraerValorDeClaveJson, finding the last "comando" key (the real JSON schema key) instead of the first one (which could match inside dialogo text). Added the new method to JsonHelper.cs
• PromptBuilder.cs LANGUAGE RULE — Changed from "LANGUAGE RULE: You MUST strictly..." to "CRITICAL LANGUAGE RULE: You MUST strictly..." with explicit "NEVER switch languages mid-conversation" reinforcement
• Created Game/SentimentAnalyzer.cs — static class with keyword-based sentiment detection (English + Spanish) for three categories:
  • Compliment → Joy +5, Consent +5
  • Romantic → Arousal +8
  • Insult → Rage +20, Joy −5, Consent −5
• Precedence: insult > romantic > compliment
• Null-guards all EmocionesFemeninas fields, clamps to [0, 100]
• Wired into ChatWindow.EnviarMensaje() — after adding user message to history, calls SentimentAnalyzer.ApplySentiment(), injects a system message with feedback, and shows the orange emotion feedback label
• Added AllowSentimentReactions toggle in Settings UI with tooltip
• Config save/load updated for AllowSentimentReactions key
• Bug fix — PromptBuilder isLocal:
  • Changed Provider != ProviderType.Gemini (which made Gemini=local, Gemma=non-local) to Provider == ProviderType.Ollama || Provider == ProviderType.Gemma — now Ollama and Gemma correctly get the jailbreak + [CMD:] format, while Gemini gets JSON format