Overview Glass features a powerful AI assistant system that provides intelligent code assistance, automated refactoring, and context-aware code generation. The assistant leverages multiple language models and integrates deeply with the editor for seamless coding workflows.
Agent Architecture Native Agent System // From crates/agent/src/agent.rs pub struct NativeAgent { sessions : HashMap < acp :: SessionId , Session >, thread_store : Entity < ThreadStore >, project_context : Entity < ProjectContext >, templates : Arc < Templates >, models : LanguageModels , // ... }
The AI assistant is built on a native agent architecture:
Sessions
Multiple concurrent conversation threads
Session-based context isolation
Persistent thread history
Context
Project-aware context gathering
Worktree snapshots
Rules and prompts integration
Thread Management struct Session { thread : Entity < Thread >, acp_thread : Entity < acp_thread :: AcpThread >, pending_save : Task <()>, _subscriptions : Vec < Subscription >, }
Each AI conversation is managed as a thread:
Internal Thread : Processes user messages and agent responsesACP Thread : Handles Agent Client Protocol communicationAuto-Save : Threads are automatically savedEvent Subscriptions : Real-time updates on thread changes
Language Models Model Selection pub struct LanguageModels { models : HashMap < acp :: ModelId , Arc < dyn LanguageModel >>, model_list : acp_thread :: AgentModelList , refresh_models_rx : watch :: Receiver <()>, // ... }
Glass supports multiple language model providers:
Recommended Models
All Models
The system curates a list of recommended models across providers for optimal performance on coding tasks.
Browse all available models grouped by provider:
Anthropic (Claude)
OpenAI (GPT)
Ollama (Local models)
And more…
pub struct AgentModelInfo { id : acp :: ModelId , name : SharedString , description : Option < SharedString >, icon : Option < AgentModelIcon >, is_latest : bool , cost : Option < SharedString >, }
Each model provides:
Name and Description : Clear identificationIcon : Visual representationVersion Status : Know if you’re using the latest versionCost Information : Understand pricing for commercial models
Model Authentication The system automatically authenticates with model providers:
fn authenticate_all_language_model_providers ( cx : & mut App ) -> Task <()>
Providers like Ollama and LM Studio are automatically detected on your local machine.
Edit Agent The Edit Agent specializes in code transformations:
Code Editing Capabilities
Parse Edit Request
Understands natural language descriptions of desired code changes
Analyze Context
Examines surrounding code and project structure
Generate Edit
Produces precise code modifications
Apply Changes
Integrates changes with conflict resolution
Edit Prediction Real-time code suggestions as you type:
// Edit prediction modes pub enum EditPredictionsMode { Disabled , Enabled , // ... }
Inline Predictions : See suggestions directly in the editorAccept Prediction : AcceptEditPrediction actionPartial Acceptance : AcceptNextWordEditPrediction or AcceptNextLineEditPredictionNavigation : Move between predictions with NextEditPrediction / PreviousEditPrediction
Toggle edit predictions on/off with the ToggleEditPrediction action from the editor.
Project Context Context Gathering pub struct ProjectContext { pub worktree_snapshots : Vec < TelemetryWorktreeSnapshot >, pub timestamp : DateTime < Utc >, }
The AI assistant automatically gathers project context:
File System
Directory structure
File types and organization
Git repository information
Code Context
Open files and recent edits
Language configurations
Diagnostic information
Rules and Prompts use prompt_store :: { ProjectContext , PromptStore , RULES_FILE_NAMES , RulesFileContext , UserRulesContext , WorktreeContext , };
Customize AI behavior with project-specific rules:
.rules filesCustom prompts : Create reusable prompt templatesWorktree context : Context specific to each workspaceUser rules : Personal preferences across all projects
Slash Commands Extend the assistant with slash commands:
Built-in Commands // From assistant_slash_commands - / default - Default assistant behavior - / diagnostics - Include current diagnostics - / file - Reference specific files - / fetch - Fetch external content - / delta - Show git changes
Extension Slash Commands // From extension manifest pub slash_commands : BTreeMap < Arc < str >, SlashCommandManifestEntry >,
Create custom slash commands via extensions:
Define Command
Add slash command to extension manifest
Implement Handler
Handle run_slash_command in extension code
Provide Completions
Implement complete_slash_command_argument for autocomplete
Agent Servers Connect to external agent services:
// From extension manifest pub agent_servers : BTreeMap < Arc < str >, AgentServerManifestEntry >,
Agent Server Features
Custom Agents : Integrate specialized AI agentsExternal Services : Connect to proprietary AI systemsTool Integration : Provide custom tools to the agent
Context Servers Enhance agent context with external data sources:
Context Server Protocol use context_server :: ContextServerId ; pub struct ContextServerRegistry { // Manages available context servers }
Context servers provide:
External Knowledge : Database schemas, API docs, etc.Real-time Data : Current metrics, logs, etc.Custom Context : Project-specific information sources
Configuration pub async fn context_server_configuration ( & self , context_server_id : Arc < str >, project : Arc < dyn ProjectDelegate >, ) -> Result < Option < ContextServerConfiguration >>
Usage and Limits Edit Prediction Usage pub struct EditPredictionUsage { // Track usage of edit predictions } // From cloud_llm_client EDIT_PREDICTIONS_USAGE_AMOUNT_HEADER_NAME EDIT_PREDICTIONS_USAGE_LIMIT_HEADER_NAME
Track your edit prediction usage:
Usage Monitoring : See how many predictions you’ve usedLimits : Understand your plan limitsHeaders : Usage information in API responses
Edit prediction features may be limited based on your subscription plan.
Control what the AI agent can do:
// From agent/tool_permissions pub struct ToolPermissions { // Manages agent tool access }
Permission System : Approve or deny agent tool usageSecurity : Prevent unauthorized actionsAuditing : Track what tools the agent uses
Templates Prompt Templates pub struct Templates { // Shared templates for all threads }
Reusable prompt templates for common tasks:
Code Review : Templates for reviewing codeRefactoring : Common refactoring patternsDocumentation : Generate documentation templatesTesting : Create test cases
Advanced Features Outline Generation // From agent/outline pub mod outline ;
Generate code outlines and structure summaries for better context understanding.
pub use pattern_extraction ::* ; pub use shell_command_parser :: extract_commands;
Extract patterns from code:
Identify repeated code patterns
Extract shell commands from conversations
Recognize code templates
Best Practices Effective AI Assistance
Provide Context
Include relevant files and diagnostics in your requests
Use Slash Commands
Leverage slash commands for specific tasks like /file or /diagnostics
Review Changes
Always review AI-generated code before accepting
Customize Rules
Create .rules files for project-specific AI behavior
Use smaller models for simple tasks to reduce latency
Enable edit predictions for real-time assistance
Configure context servers for frequently needed information
Create templates for repetitive tasks
For the best results, provide clear, specific requests and include relevant code context.