tdav/myharness

Purpose and solution shape

MyHarness is an autonomous ReAct coding/research agent backed by Ollama Cloud or compatible Ollama servers. Shared agent behavior lives in the net10.0 library; WinForms and terminal applications are presentation shells over it. Solution entry points are MyHarness/Program.cs and HarnessCli/Program.cs; core assembly and orchestration begin at Harness.Core/AgentHost.cs.

Agent host and model boundary

AgentHost constructs the Microsoft Agent Framework/Harness agent, configures OllamaSharp with Bearer authentication, registers tools, and owns workspace-scoped runtime dependencies. Settings resolve from secret.json and environment variables, with model/context-window changes requiring host recreation while preserving sessions.

• Harness.Core/AgentHost.cs

• Harness.Core/IAgentHost.cs

• Harness.Core/AppSettings.cs

The external model boundary is Ollama’s API: model listing/context discovery and streaming completion traffic go to Ollama Cloud or a local compatible endpoint.

ReAct execution and workspace safety

The agent iterates reasoning, tool selection, execution, and observation. It operates against a selected working directory and exposes file access/search, PowerShell execution, Hyperlight-isolated Python CodeAct, todo/mode/memory capabilities, and optional web tools. UI supplies approval decisions; read-only file operations are auto-approved, while mutations and execution require policy approval.

• Harness.Core/AgentHost.cs

• Harness.Core/Runtime/RunLoop.cs

• Harness.Core/Runtime/TaskQueue.cs

Shell and filesystem execution are a critical trust boundary: workspace scoping, command deny-listing, timeouts, sandboxing, and approval policy protect the host.

Sessions, long-running work, and recovery

Runtime services model queued work, progress, health, notifications, automated recovery, and long-running agent runs. Session persistence preserves conversation state across UI session changes and application restarts; memory services compact/prune context and retain longer-lived information.

• Harness.Core/Persistence/SessionStore.cs

• Harness.Core/Runtime/LongRunManager.cs

• Harness.Core/Runtime/AutoRecovery.cs

• Harness.Core/Memory/LongTermMemory.cs

• Harness.Core/Memory/ContextCompactor.cs

Goal planning and self-improvement

For multi-step work, goals are decomposed into a tree and prioritized for scheduling. Post-run analysis can refine execution strategy and optimize prompts, forming a feedback loop around agent outcomes rather than UI behavior.

• Harness.Core/Goals/GoalDecomposer.cs

• Harness.Core/Goals/GoalTree.cs

• Harness.Core/Goals/PriorityScheduler.cs

• Harness.Core/Evolution/RunAnalyzer.cs

• Harness.Core/Evolution/PromptOptimizer.cs

Extensibility, packaged knowledge, and workflows

Runtime plugins are C# source files compiled with Roslyn and loaded as one-shot tools or resident services. Plugins may call back into the agent through persistent plugin sessions; examples include web search and Telegram. This feature depends on JIT and is unavailable under NativeAOT. Shipped skill/persona documents and .NET scripts are runtime-discovered/copied resources, not compiled agent logic.

• Harness.Core/Plugins/PluginContracts.cs

• Harness.Core/Plugins/PluginManager.cs

• Harness.Core/Workflows/Magentic.cs

• Harness.Core/agents/

• Harness.Core/skills/

UI adapters and observability

WinForms provides WebView2 Markdown rendering and interactive tool approval; SharpConsoleUI provides the equivalent terminal workflow. Both create and drive AgentHost instances, stream transcripts, manage sessions/models/modes, and surface plugin logs. OpenTelemetry spans—including HTTP instrumentation—are file-exported; Serilog application logging is configured by the app hosts.

• MyHarness/UI/MainForm.cs

• MyHarness/UI/MarkdownViewer.cs

• HarnessCli/UI/MainWindow.cs

• Harness.Core/Tracing/HarnessTracing.cs

• Harness.Core/Tracing/FileSpanExporter.cs