Releases: InteractiveAgents/IAW
Releases · InteractiveAgents/IAW
v0.0.2 — Agent Registry, Orchestration Redesign & Generative UI
What's Changed
This release is a ground-up redesign of the agent framework's runtime architecture, replacing four legacy subsystems with a cohesive, production-grade agent platform built on Orleans grains, durable jobs, and structured UI composition.
Agent Identity & Lifecycle
- Dynamic grain IDs via
client.Get<IGit>(taskId)— each task/context gets isolated agent instances with isolated state (replaces singleton"git","shell"grain IDs) Agent<TInterface>generic base class binds agents to their contract interface at compile time- Partial-file architecture:
Agent.cs,Tools.cs,Scheduling.cs,State.cs,Events.cs,Streams.cs,Lifecycle.cs,Observers.cs
Agent Registry & Discovery
- Self-healing in-memory registry via
DiscoverAndBuildRecords()— re-runs on grain reactivation, no external persistence dependency - Agent interfaces declare static metadata via
[AgentDescription]and[AgentCapability]attributes - Auto-tool discovery: interface methods with
[Description]auto-register as AI tools
Orchestration Pipeline Redesign
DelegateAsyncnow executes work inline — MCP and DevUI get real results instead of"Task submitted"placeholdersCodeOrchestratorAgentgenerates standalone C# console apps connecting as Orleans clientsAgentSelectorwith LLM-powered reasoning for intelligent agent routing- Structured
OrchestrationResultwith success/summary/artifacts/metrics/error detail
Durable Job Scheduling (Critical Fixes)
- Fixed: Job duplication on grain reactivation — now cancels old
DurableJobviaTryCancelDurableJobAsyncbefore rescheduling ScheduleJob/ScheduleRecurringJobwith proper lifecycle management- One-shot jobs auto-remove; recurring jobs self-reschedule
Generative UI
UIParttype hierarchy:TextPart,OptionsPart,CardPart,MediaPart,FormPart,ProgressPart,SuggestionPart- Channel-agnostic UI composition —
TelegramUIAgentconverts raw output toRichOutputwith typed parts HandleCallbackonIAgentprovides unified callback entry point across channels
Model Tiers & LLM Architecture
- Three tiers:
fast(Haiku),balanced(Sonnet/GPT-5.4-Mini),reasoning(Opus) - New models:
Gpt54Mini,Gpt54Nano [Llm<TModel>]attribute binds agents to specific keyedIChatClient
Namespace Reorganization
IAW.Agents/ Infrastructure/ Git, Shell, FileSystem, Aspire LLM/ Opus46, Sonnet46, Gpt54Mini, Gpt54Nano, ... Memory/ CodeMemory, EpisodeMemory, PatternMemory, ... Knowledge/ KnowledgeAgent Orchestration/ Thread, AgentSelector, CodeOrchestrator, TelegramUI Dead agents removed: ReviewerAgent, SelfImprovementAgent, BuildAgent, UserAgent, Project
Bug Fixes
- Fixed JSON truncation corrupting downstream parsing — now truncates individual fields before serialization
- Fixed Telegram MarkdownV2 escaping and job result delivery
- Fixed AppHost path resolution by walking up from bin directory
- Unified file access model — workspace is context for path resolution, not a security boundary
- Agent registry self-heals after grain GC (no more empty registry)
Infrastructure
CodeValidatorintegrated into orchestrator for generated code validationIAWClusterfor cluster topology management- AspireAgent rewritten with Aspire MCP connection
- Auto-discover agents in CodeOrchestrator via
InterfaceCatalog
CI & Dependencies
- Bump
actions/checkoutfrom 4 to 6 - Bump
actions/setup-dotnetfrom 4 to 5 - Added README.md
Documentation
- Removed all legacy Core.V3 references across website
- Added real-world code examples: reactive agents, self-diagnostics, context, pub/sub
- Fixed outdated base class names and tutorial API
Full Changelog: v0.0.1...v0.0.2

