Context Stabilizer Plugin

A LangBot plugin that validates and manages conversation context using an audit model.

Features

• Context Auditing: Uses a secondary LLM model (Model B) to audit conversation context
• Steganography Detection: Detects hidden characters and zero-width characters that may indicate injection attacks
• Context Compression: Automatically compresses long context while preserving key information
• Prompt Injection Protection: Injects original system prompt reminder after compression
• Adaptive Audit Frequency: Automatically adjusts audit frequency based on audit results
• Configurable Frequency: Set how often to audit (every N conversation rounds)
• Timeout Protection: Prevents auditing from blocking the main conversation flow

Installation

1. Download the plugin
2. Place it in the LangBot plugins directory
3. Configure the plugin in the LangBot web interface

Configuration

Audit Model Settings

Config	Description	Default
audit_model_uuid	The LLM model used for auditing	Required
audit_system_prompt	System prompt for the audit model	Built-in

Frequency Settings

Config	Description	Default
audit_frequency	Rounds between audits	3
enable_adaptive_frequency	Auto-adjust frequency based on results	false
frequency_increase_step	Frequency increase step on failure	1
frequency_recovery_threshold	Passes needed before decreasing frequency	3
min_audit_frequency	Minimum interval (highest frequency)	1

Timeout Settings

Config	Description	Default
audit_timeout_seconds	Audit timeout in seconds	10
timeout_action	Action on timeout (remove_chunk/compress_all)	remove_chunk

Compression Settings

Config	Description	Default
max_context_length	Max messages before compression	50
compress_target_length	Target messages after compression	10
compression_model_uuid	Model for compression (empty = use audit model)	-
compression_prompt	Prompt for context summarization	Built-in
enable_prompt_injection	Inject original prompt after compression	true

Steganography Detection

Config	Description	Default
enable_steganography_detection	Enable hidden character detection	true
steganography_patterns	Regex patterns for detection	Built-in

Advanced Settings

Config	Description	Default
chunk_size	Messages per audit chunk	5
enable_logging	Enable detailed logging	false

Commands

Command	Description
!ctxstab status	View current session audit status
!ctxstab audit	Force audit on next message
!ctxstab compress	Force compression on next message
!ctxstab reset	Reset audit counter
!ctxstab config	View current configuration

How It Works

1. Event Listening: Listens to PromptPreProcessing event to get conversation context
2. Frequency Check: Decides whether to audit based on configured frequency
3. Steganography Detection: Detects zero-width and hidden characters in context
4. Context Splitting: Splits context into chunks for auditing
5. Model Auditing: Uses audit model to check if chunks comply with original settings
6. Result Processing: Compresses or removes problematic context based on audit results
7. Adaptive Frequency: Adjusts audit frequency based on pass/fail results (if enabled)

License

MIT

LangContextStabilizer