fix: Streaming broken with Prism 0.99+ — API mismatch and state leak

fix:Return structured output - This pull request introduces support for structured output in agents, allowing them to return structured data when a schema is defined and the response includes structured content. If structured data is not available, agents will fall back to returning plain text, preserving backward compatibility. Comprehensive unit tests have been added to verify these behaviors.

Integration of new documentation

New Features

Media Generation Agents

• New ImageAgent and AudioAgent for generating images and audio directly from your agents
• Fluent builder API: ImageAgent::run('prompt')->quality('hd')->go()
• Built-in storage methods and queued job support for async generation
• DelegateToMediaAgentTool allows LLM agents to delegate media tasks
• Chat UI now renders media responses inline

Toolbox System

• Group related tools into reusable toolboxes with ToolboxInterface
• Laravel Gate/Policy authorization at both toolbox and per-tool level
• Conditional tool inclusion via shouldIncludeTool()
• New artisan command: php artisan vizra:make:toolbox

Tool Chaining

• Compose sequential tool pipelines where output flows between tools
• Fluent API with pipe(), transform(), when(), and tap() methods
• ChainableToolInterface for tools with built-in chain support
• Detailed execution results with ToolChainResult
• Full tracing support for tool chain steps

Laravel Boost V2 Support

• Auto-discovered guidelines via resources/boost/guidelines/core.blade.php
• Opt-in Skills for advanced features:
  • vizra-agent-creation - Agent patterns
  • vizra-tool-creation - Tool development
  • vizra-workflows - Workflow orchestration
  • vizra-memory - Memory and RAG
  • vizra-evaluation - Testing framework

Add LLM usage aggregation and adjust agent output Normalize LLM usage reporting and add aggregation APIs. BaseLlmAgent now extracts promptTokens/completionTokens into input_tokens/output_tokens and computes total_tokens before ending spans. Tracer gained getUsageForSession, getUsageForTrace and getUsageDetails plus helper methods (aggregateUsage, emptyUsageResponse) to query spans, decode outputs and aggregate token counts, with safe handling when tracing is disabled or DB errors occur. Tests updated/added to cover session/trace aggregation, detailed breakdowns, disabled tracing and spans without usage data.

Bump prism to ^0.99.0 and update deps Increase prism-php/prism requirement to ^0.99.0 in composer.json and regenerate composer.lock (dependency bumps and metadata updated). Apply small changes to BaseLlmAgent and AgentManager to remain compatible with the updated dependency behavior/API. This brings bug fixes and compatibility improvements from the newer prism and related packages.

Update Livewire dependency to v4.0 and adjust related documentation

chore: release v0.0.39

Added

• Dynamic Agent Name Resolution: Agents can now be registered and executed with runtime-determined identities for multi-tenant applications (e.g., team_agent_24, team_agent_42)
• Prism v0.92.0 Streaming Support: Full support for Prism's new streaming event system with type-safe StreamEvent objects
• Streaming UI in ChatInterface: Real-time streaming responses in Vizra ADK chat UI using Livewire 3's wire:stream pattern
• Lightweight User Context: New withUserContext(array) method for passing minimal user data without serializing entire User models
• Multi-tenant MCP Config Overrides: Per-tenant MCP server configuration via AgentContext with deep-merge support
• Comprehensive Documentation: Added docs for lightweight user context and multi-tenant MCP configuration

Fixed

• Image/Document Persistence: Prism Image and Document objects now properly filtered before database save to prevent serialization errors
• Session Association: Fixed user_id extraction to check both 'user_id' and 'id' keys in context
• Artisan Command: Corrected artisan command reference in eval runner blade template

Changed

• Agent Name Resolution Strategy: Implemented three-tier resolution system (explicit context → registry → fallback)
• User Context Priority System: Three-tier priority for user data (explicit withUserContext() → toAgentContext() → toArray())
• MCPClientManager Binding: Changed from singleton to non-singleton for proper multi-tenant isolation
• Message Persistence Architecture: AgentChat handles UI display while BaseLlmAgent handles all persistence

Improved

• Streaming Event Handling: Rewrote buffer accumulation to handle both snake_case and kebab-case event types
• Template Variable Clarity: Prefixed template variables with '@' in agent, evaluation, and tool creation templates
• Test Coverage: Added tests for multi-tenant MCP behavior, config overrides, and image/document persistence

Full Changelog: https://github.com/vizra-ai/vizra-adk/compare/v0.0.32...v0.0.33

Added support for pseudo-terminal (PTY) mode and improved environment handling in MCPClient. Updated configuration to allow custom npx path and app directory access. Improved client connection management in MCPClientManager to handle stale clients and added use_pty option to server configs.

Enhances agent tracing to support sub-agent delegation with parent trace context preservation and restoration. Updates the chat interface UI to display sub-agent info, improves typing indicator and send button logic, and adds polling for running traces. Refactors prompt versioning to support Blade templates, and adds feature/browser tests for chat interface behaviors.

Introduces JudgeBuilder to enable fluent, agent-based assertion syntax in evaluations. Updates BaseEvaluation to expose judge() and make recordAssertion public. Adds comprehensive unit tests for JudgeBuilder and its integration with BaseEvaluation. Removes obsolete MCPMakeAgentCommand.

Corrects the logic for determining the 'days' value in AgentTraceCleanupCommand. Now, the command only uses the config default if the 'days' option is not provided, ensuring explicit zero values are respected.

Qdrant and in-memory vector drivers have been removed from configuration, validation, and setup logic. Only 'pgvector' and 'meilisearch' are now supported for vector storage. Also, AgentTraceCommand now includes 'input', 'output', and 'metadata' in span output.

Introduces AgentVectorProxy to simplify agent vector memory operations, making vector and RAG methods public and context-aware. Adds GeminiEmbeddingProvider for Google Gemini embeddings. Updates VectorMemoryManager and related tools/tests to use agent class instead of agent name, streamlining method signatures and usage. Enhances Meilisearch driver with fallback similarity calculation. Updates configuration for vector memory and RAG features.