Session 03 extends the Session 02 finance advisor with scalable, inspectable orchestration:

• discoverable skills
• approval-gated local shell execution
• CodeAct-style calculation boundaries
• background research tickets

The package follows the MafClaw session layout:

• code/ - complete offline finance advisor, including Session 02 file access, approvals, memory, and the new Session 03 concepts
• samples/ - numbered plain C# concepts followed by MAF bridge samples
• docs/ - setup, architecture, teaching, and troubleshooting guidance

The official Microsoft Agent Framework post is:

https://devblogs.microsoft.com/agent-framework/agent-harness-scaling-the-claw-or-harness-capabilities/

All data is mock and educational. The complete code sample is offline and does not call a model or access the network.

All 22 projects target .NET 10. The live bridges and orchestration support use MAF/Harness 1.21.0, Foundry/Shell/Hyperlight 1.21.0-preview.260911.1, Azure.AI.Projects 3.0.0-beta.2, Azure.Identity 1.21.0, configuration 10.0.12, and the Hyperlight Python guest 0.7.0 where applicable. Preview packages are limited to existing preview integrations. The session-local NuGet.Config selects nuget.org.

verification-manifest.json defines deterministic app/fixture checks, the six executable regression runners, and separately opted-in live checks. Run these regression projects with dotnet run, not dotnet test: they are executable assertion suites. Fixture success proves host mechanics, not live Foundry, hosted-search, or Hyperlight service/platform availability. No live failure automatically switches to fixtures.

dotnet run --project .\code\MafClaw.Session03.csproj

See samples/README.md for the complete numbered ladder.

The plain-C# Sample 20 makes command policy visible before the MAF shell bridge: dotnet --version is allowed, while dotnet --info is denied before launching a child process. Its five-second timeout terminates the owned process rather than merely ending the wait. This sample demonstrates an allowlist, not an OS sandbox or a human approval gate.

The live Sample 21 then lets the model propose PowerShell commands in a fresh mock workspace. MAF supplies shell environment context and approval handling; the console displays real tool results and the host verifies final filenames and unchanged content hashes. Previous demo runs are preserved.

Sample 31 currently uses automatic sandbox execution (NeverRequire), not a per-execution human approval prompt. Its host file-tool scope is unchanged. See the CodeAct policy.

• docs/README.md
• docs/setup.md
• docs/architecture.md
• docs/troubleshooting.md

Every C# file in this package starts with an objective and A/B/C step header. Short comments before major blocks identify the boundary being demonstrated, so presenters can explain the code while sharing the screen without adding implementation noise.

Samples 42-47 extend the background-agent ladder without changing 40/41: specialist teams, selective delegation, research/write/review, responsive background jobs, partial results with deadlines, and a human-approved report.

Each is a runnable MAF project with --mode live and an explicitly scripted --mode fixture. Start with the orchestration guide for all six commands, boundaries, and offline regression checks. These extensions do not change the existing offline code advisor or add real trading.