Skip to content

Agent Platform Model Selection

Mohamed Hamdarequested to merge
556185-agent_platform_model_selection into master

Agent Platform Model Selection

What This MR Does

Implements dynamic model configuration for GitLab Duo Agent Platform, enabling instance-level AI model selection with self-hosted model support.

Key Changes

  • Model Metadata Service: New ModelMetadata service retrieves and formats model configuration from instance-level feature settings
  • Feature Flag Integration: Added agent_platform_model_selection development flag with global enablement
  • Instance Feature Settings: Extended Ai::FeatureSetting to support duo_agent_platform feature
  • Agent Platform Integration: Injects model metadata via x-gitlab-self-hosted-models-metadata header to agent platform endpoints

Model Support

  • GitLab Models: Instance-level selection (e.g., claude-3-7-sonnet-20250219)
  • Self-Hosted Models: Custom OpenAI-compatible endpoints with API key protection
  • Amazon Q: Global role ARN configuration support

Instance-Level Architecture

Instance-level feature settings provide centralized model configuration for all agent platform operations. This architecture allows administrators to:

  • Configure self-hosted models globally
  • Set default model providers for all users
  • Manage API keys and endpoints centrally
  • Future Enhancement: Add user-level overrides when needed

Technical Flow

graph TD
    A[Agent Platform Request] --> B[Feature Flag Check]
    B --> C{Flag Enabled?}
    C -->|No| D[Return nil]
    C -->|Yes| E[Instance Feature Setting Lookup]
    E --> F[duo_agent_platform Setting]
    F --> G[Model Metadata Generation]
    G --> H{Metadata Present?}
    H -->|No| I[Return nil]
    H -->|Yes| J[Header Injection]
    J --> K[Agent Platform Service]
    I --> L[Standard Operation]
    K --> M[Model-Aware Operation]
  1. Global feature flag validation (agent_platform_model_selection)
  2. Instance-level duo_agent_platform feature setting lookup
  3. Model metadata extraction from feature setting
  4. JSON header injection to Agent Platform Service

Feature Configuration

  • Feature: duo_agent_platform (ID: 15) in STABLE_FEATURES
  • Flag: agent_platform_model_selection - Development flag (default disabled)
  • Scope: Instance-level configuration applies to all agent platform operations

Self-Hosted Models

Self-hosted models are configured at the instance level through Ai::FeatureSetting records. Administrators can:

  • Set custom OpenAI-compatible endpoints
  • Configure API keys securely
  • Define model identifiers
  • Enable/disable specific features

Related to: #556159 (closed) and #556185 (closed)

Edited by Mohamed Hamda

Merge request reports