feat: voice-to-text reliability — eager init, health checks, ORT feed#14
Merged
Conversation
Add dedicated IAWSystem agent (Opus 4.6) that autonomously diagnoses, fixes, builds, tests, and deploys changes to the IAW system. Uses SendToAgent to delegate to Aspire, DotNet, FileSystem, Roslyn, Git. Fix Aspire DeployAsync: build solution via DotNet agent before restart (Aspire runs --no-build on start). Schedule deploy-verify health check. Update Thread routing to include IAWSystem for self-improvement requests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Orleans default 5-minute ResponseTimeout kills IAWSystem requests before the multi-step orchestration (LLM calls + sub-agent delegation) can complete. Override GetResponse with 30-minute timeout. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Downgrade from Opus 4.6 to Sonnet 4.6 (3-5x faster LLM responses) - Rewrite instructions: simple edits = FileSystem + DotNet + Git only - Explicitly ban Shell/Roslyn for file edits (was causing misrouting) - Minimize tool call chain for common operations Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…by a personalized message
Private [Description] methods in Agent.Scheduling.cs were never discovered by tool registration. Added RegisterSchedulingTools() that scans for non-public [Description] methods on the Agent base class. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Thread generates a 2-5 word title from the first user+assistant exchange via a lightweight LLM call. Title is cached in durable state. Excluded from AI tool discovery to prevent LLM self-invocation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Creates a new Telegram forum topic, registers it via SetTopicId, and auto-renames the topic after the first response using thread.GetTitle(). Uses TryAutoRenameTopicAsync helper called from both exit paths of StreamResponseAsync. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Lists all custom topics with message counts and inline Delete buttons. Delete callback closes the Telegram forum topic, clears thread history, and removes the project from UserProfile. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…firmation topic TryAutoRenameTopicAsync now tracks renamed topics in-memory to avoid calling EditForumTopicAsync on every message in chat- topics. HandleCleanupDeleteAsync sends confirmation to the invoking topic. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CodeValidator.Sanitize used TrimStart() which left \r on Windows,
causing EndsWith(';') to always fail — usings were never removed.
Fixed by using Trim() instead.
Added 2-minute per-build timeout to CodeOrchestratorAgent.TryBuild
to prevent test hangs when dotnet build takes too long.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ad safety CodeValidator.ValidIawNamespaces was missing IAW.Agents.Fun, causing Sanitize() to strip usings for the new Fun agents. Replaced HashSet<string> with ConcurrentDictionary for _renamedTopics since TelegramBotService is a singleton handling concurrent updates. Also fixed naming to follow _camelCase convention. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Update dependency versions and add planning/spec files for the /newchat scheduling fix. Changes: bumped multiple Aspire packages (Hosting.Orleans, Hosting.JavaScript, Hosting.Qdrant, Qdrant.Client, Azure.Storage.Blobs, Hosting.Azure.CosmosDB, Hosting.Azure.Storage, etc.) to 13.2.0 in Directory.Packages.props; upgraded Aspire.AppHost SDK to 13.2.0 in src/IAW.AppHost/Aspire.csproj; added planning and design docs for the newchat/cleanup + scheduling-tool fix (docs/superpowers/plans/... and docs/superpowers/specs/...); and added an mcp entry to .claude/settings.local.json. These changes prepare the codebase for the new Telegram topic management and scheduling tool registration work described in the docs.
mcp-call.sh and mcp-client.sh are manual curl wrappers that duplicate what the native IAW MCP server (.mcp.json) already provides. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix .mcp.json: add type:http to iaw, fix aspire args, add microsoft-learn MCP - Add Orleans gateway connection retry filter to IAWClientExtensions - Remove manual delay in Telegram StreamSubscriber (handled by retry filter) - Clean up settings.local.json permissions, enable all project MCP servers - Add Aspire skill definitions Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Eugene <59283295+ScientistFromMars@users.noreply.github.com>
After Whisper transcribes a voice message, reply with the transcribed text so the user can see what was understood before the LLM responds. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…d agents - Add Qwen25_7B/14B model classes with explicit Ollama tags for GPU sizing - Add WithEmbedding<T>() builder chain mirroring WithLLM<T>() pattern - Add NoOpEmbeddingGenerator fallback (replaces crash on missing cloud keys) - Rewrite AddEmbeddingProvider() to support Ollama/OpenAI/GitHub/NoOp - Fix Ollama connection string resolution (strip tag before sanitizing) - Switch 6 agents from concrete models to tiers (Fast/Balanced/Reasoning) - Fix /cleanup command: DeleteForumTopicAsync instead of CloseForumTopicAsync - Fix CodeOrchestrator: discover files instead of hardcoding Program.cs - Remove dead ScriptExecutor and tests Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…reporting Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
DownloadAsync/LoadAsync don't accept CancellationToken, so use Task.WhenAny with Task.Delay to enforce the 5min/2min timeouts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…tation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move all LLM and embedding models into provider sub-namespaces under Core.AI.Models (OpenAI, Anthropic, GitHub, Ollama, XAI, Google). AppHost now has 4 complete provider blocks — uncomment one to switch: OpenAI, Anthropic, GitHub Models, or Ollama local. - Add new models: Gpt54 (OpenAI reasoning), Gpt41Nano, Gpt41Mini, O4Mini (GitHub with tool calling), GitHub TextEmbedding3Small - Fix deprecated GitHub Models endpoint (models.github.ai/inference) - Update GitHub model IDs to use openai/ prefix (new API requirement) - Add / normalization to ServiceKey for prefixed model IDs - Move embedding models to provider namespaces (MxbaiEmbedLarge → Ollama, TextEmbedding3Small → OpenAI) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ext-hosted-service
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
nuget.configwith ORT Azure DevOps feed soMicrosoft.ML.OnnxRuntime.Foundrynative binaries restore on any machine (fixes silent failure on fresh PCs)SemaphoreSliminit with eagerIHostedService— Whisper model downloads/loads at startup, failures surface in Aspire Dashboard immediately instead of hanging silently on first voice messageWhisperHealthCheckreporting Healthy/Degraded/Unhealthy in Aspire Dashboard based onIWhisperReadinessstateTask.WhenAnysince Foundry Local SDK doesn't acceptCancellationTokenAddWhisperProvider<T>to register as singleton + interface alias + hosted service + health check (single shared instance)Core.AI.Models.OpenAI,.Anthropic,.GitHub,.Ollama,.XAI,.Googlewith 4 complete provider blocks in AppHost for easy switchingGpt54(OpenAI reasoning),Gpt41Nano/Gpt41Mini/O4Mini(GitHub with tool calling), GitHubTextEmbedding3Smallmodels.github.ai/inferencewithopenai/prefixed model IDsProblem
Voice-to-text via Foundry Local silently failed on second PC — no logs, no errors, no response after sending voice messages. Root cause: lazy initialization held a
SemaphoreSlimforever when model download/load stalled, and the fire-and-forget webhook handler swallowed the timeout.Additionally, all LLM models lived in a flat
Core.AI.Modelsnamespace making it impossible to switch between providers (OpenAI, Anthropic, GitHub Models, Ollama) without hunting through individual model files. The GitHub Models endpoint was deprecated (Oct 2025).Test plan
dotnet restore IAW.slnxsucceeds on fresh machine (ORT feed resolves native binaries)dotnet build— zero C# compilation errors/healthreports whisper status (Healthy or Unhealthy with error)🤖 Generated with Claude Code