Guide agents to declare shell working scope

[Aaronontheweb]

Summary

• expose useful Command, WorkingDirectory, and project-root descriptions in generated tool schemas
• guide agents to use persistent project scope for multi-command work and typed WorkingDirectory for one-off commands
• preserve inline cd only when directory mutation is the behavior under test, and retry rejected project roots without falling back to stale scope
• add sanitized, structured behavioral eval cases whose assertions consume real toolCalls JSON rather than model-printable transcript markers

Validation

• dotnet build -c Release --no-restore — 0 warnings, 0 errors
• dotnet test -c Release --no-build --no-restore — all local projects pass; expected platform and integration skips only
• bash -n evals/run-evals.sh
• shellcheck -e SC2015,SC2155 evals/run-evals.sh
• pwsh ./scripts/Add-FileHeaders.ps1 -Verify
• git diff --check
• adversarial review: PASS, no warnings
• added-line PII scan: no findings

Provider-backed model eval execution remains explicitly pending because the local provider type, endpoint, and model ID are unset. The eval definitions, structured provenance checks, and harness validation are included here.