feat(conversation): introduce ConversationSession; refactor AICall, streaming, validation, and model selection #315
Conversation
…ts with observers and multi-turn flows and tool passes
…/tools - Remove legacy AICall types under Infrastructure/AICall/* - Update components, GH tools, and providers to new Core namespaces - Tweak ConversationSession and interfaces accordingly - Refresh docs/Tools/index.md and CHANGELOG entries
… EnableStreaming flag and update docs/tests
…chema via request policy
…d simplify AIToolCall
…ified replaced and invalid
…een generation and backoffice calls which validate differently
|
🏷️ This PR has been automatically assigned to milestone 0.6.0-alpha based on the version in |
There was a problem hiding this comment.
Pull Request Overview
This PR introduces a major architectural refactor focused on conversation orchestration, streaming capabilities, and centralized validation. The changes implement ConversationSession as the primary orchestrator for multi-turn conversations with tool execution, while maintaining backward compatibility through existing AIRequestCall for single-turn operations.
Key changes include:
- ConversationSession architecture: New session-based orchestration with observer patterns for multi-turn conversations and tool execution
- Unified streaming framework: Provider-agnostic streaming adapters with centralized capability checks and consistent delta emission
- Centralized validation and policy pipeline: Request/response policies for schema validation, tool validation, and response normalization
Reviewed Changes
Copilot reviewed 164 out of 164 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| Provider settings files | Added EnableStreaming setting across OpenAI, MistralAI, and DeepSeek providers |
| Provider model files | Updated model capabilities and added RetrieveApiModels() method implementations |
| Provider main files | Added streaming adapters and centralized JSON schema handling |
| JSON schema adapters | New provider-specific schema wrapping/unwrapping logic |
| Assistant settings | Added canvas button toggle and reorganized settings structure |
| Streaming infrastructure | New IStreamingAdapter interface and base classes for provider streaming |
| Validation framework | Comprehensive tool and schema validators with structured error reporting |
| Session management | ConversationSession implementation for multi-turn orchestration |
| Policy pipeline | Request/response policies for centralized validation and normalization |
| Metrics framework | Enhanced metrics tracking with correlation and domain-specific records |
Comments suppressed due to low confidence (1)
src/SmartHopper.Infrastructure/AICall/Sessions/ConversationSession.cs:1
- The timeout logic creates a potential memory leak as the original execTask continues running even after timeout. Consider implementing proper cancellation token propagation to AIToolManager.ExecuteTool() to actually cancel the underlying work.
/*
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
| }; | ||
|
|
||
| // Rebuild body immutably with context as first item and filter-out any pre-existing context interactions | ||
| var rest = body.Interactions?.Where(i => i != null && i.Agent != AIAgent.Context) ?? Enumerable.Empty<IAIInteraction>(); |
There was a problem hiding this comment.
The LINQ Where clause filters out existing context interactions, but this creates a new enumerable each time. Consider materializing this to a list if the collection will be enumerated multiple times in AddRange().
| var rest = body.Interactions?.Where(i => i != null && i.Agent != AIAgent.Context) ?? Enumerable.Empty<IAIInteraction>(); | |
| var rest = body.Interactions?.Where(i => i != null && i.Agent != AIAgent.Context).ToList() ?? new List<IAIInteraction>(); |
| if (interaction is AIInteractionToolCall || interaction is AIInteractionToolResult) | ||
| { | ||
| this.Request.Body = this.Request.Body.WithAppended(interaction); | ||
| } |
There was a problem hiding this comment.
During streaming, the request body is being rebuilt repeatedly with each interaction. This creates multiple immutable copies. Consider batching these updates or using a more efficient accumulation strategy.
| /// </summary> | ||
| public static class JsonSchemaAdapterRegistry | ||
| { | ||
| private static readonly ConcurrentDictionary<string, IJsonSchemaAdapter> _adapters = new ConcurrentDictionary<string, IJsonSchemaAdapter>(StringComparer.OrdinalIgnoreCase); |
There was a problem hiding this comment.
The static ConcurrentDictionary for adapter registration could lead to issues in testing scenarios where adapters need to be reset between tests. Consider providing a Reset() method for test scenarios or making the registry injectable.
Description
This PR delivers a major architectural refactor and feature set focused on conversation orchestration, centralized validation, streaming, and model management. Changes are summarized and grouped below based on commit messages and the [Unreleased] section in CHANGELOG.md.
IConversationSession,IConversationObserver,SessionOptions; multi‑turn and tool orchestration with observer callbacks; single source of truth for long‑running interactions.AIBodyBuilder;AIRequestKindto distinguish generation vs backoffice flows.AIRequestCall.Exec()clarified as single‑turn; multi‑turn flows handled via ConversationSession.AIProviderStreamingAdapter; OpenAI streaming adapter now derives from the shared base.ModelManager.ModelSupportsStreaming(...).ConversationSession.Stream()clearly gates unsupported models and emits actionable errors.AIMessageCodesurfaced throughAIRuntimeMessageand consumed by components (e.g., badges).IAIProvider.SelectModel(...)delegates to centralizedModelManager.SelectBestModelwhile honoring provider defaults/settings.IAIProviderModels.RetrieveModels()and registered centrally.EnableStreaming.list_generateresult output and image output propagation toImageViewer.Breaking Changes
AIRequestCall.Exec()is now explicitly single‑turn; multi‑turn/tool passes requireConversationSession.RetrieveAvailable,RetrieveCapabilities,RetrieveDefaultin favor ofRetrieveModels()and centralizedModelManager.TemplateProviderandAIToolCall.ReplaceReuseCount()(unified metrics handling).Testing Done
ImageViewer.Checklist