Testing

Select the phase explicitly

Use explicit --path App --workflow Orders, otherwise a named/active local context, otherwise current directory. Explicit --path bypasses an implicit active context; an explicitly named cloud context is refused even with a path. No Azure login or running host is required for local planning. Input/target failures before reporting are not successful plans.

Scaffold, then author real assertions

Inspect existing tests and output paths first. Default scaffold output is Tests/Orders under the selected host project; a dedicated test directory within the project is supported. External dedicated tests are also valid when explicitly selected.

Creation refuses an existing destination rather than adopting or overwriting it. Multiple triggers require explicit --trigger. A relative --output-path uses the selected project. The scaffold includes native project/test source, trigger fixture, SDK selection, NuGet configuration, and Git exclusions. It does not copy app credentials/settings or SDK binaries.

The starter is ignored and inconclusive, not a passing business test. Before removing both the Ignore attribute and inconclusive guard:

1. Populate synthetic trigger data for the actual workflow/trigger.
2. Mock external operations using supported SDK mocks, including required failure paths.
3. Write meaningful business assertions against required behavior/action outputs.
4. Review project imports, build targets, package sources, and test code for permitted effects.

The scaffold links the current selected workflow at build time. Do not remove valid workflow outputs, alter expected business values, or weaken assertions simply to obtain a pass. Workflow Succeeded alone does not establish business coverage.

Plan without execution

Planning inspects local authored inputs and literal references; an optional/default scaffold .csproj is XML metadata, not evaluated MSBuild or executable test discovery. It does not install, restore, generate fixtures, launch a host, or contact a cloud target.

Inspect notExecuted markers and findings. A successful plan does not accept SDK terms or authorize execution. Preserve supported historical reports’ recorded exit policies. Reports are stdout-only by default. Optional --report-path needs an existing parent; overwriting a plan report requires deliberate --force, which authorizes no SDK activity.

Review dependencies, SDK terms, and project trust

Current qualified contract: Tests SDK 1.0.3, engine 1.165.50, MSTest 4.0.2, target net8.0, selected .NET 8 SDK. The scaffold selects 8.0.426 with latest-patch roll-forward. A target framework does not select the SDK in an existing project.

.NET 8 support ends 2026-11-10. Check lifecycle and publisher compatibility at execution time. Do not silently upgrade the pinned SDK/engine/runner combination, remove selection guards, or claim a later toolchain is qualified. If no supported approved combination is available, report the native test execution blocker and keep offline inspection available.

Review the publisher’s SDK 1.0.3 license (opens in a new tab), including use/redistribution/commercial-use restrictions, data collection, and updates. The CLI’s MIT license does not license the SDK. Obtain publisher/legal clarification where needed; do not redistribute SDK binaries with this skill or a generated project.

For each invocation, obtain authorization before adding:

• --accept-sdk-terms: caller acknowledgement of applicable SDK terms, not a saved-report grant.
• --trust-project: acknowledgement that project/imports/targets/test code can execute arbitrary code and contact services.
• --restore: separate permission to resolve packages through reviewed sources.

Do not infer these permissions from activating the skill, generating a plan, a prior run, saved flags, a context, or --force. Caches and mocks are not a security sandbox or denied egress. Use an appropriately isolated environment for untrusted native code.

The scaffold NuGet configuration clears inherited sources and selects the official feed. Before restore, adapt/review it for the customer’s approved feed/proxy policy; do not prescribe a company-specific proxy to all users. Use explicit --nuget-config CONFIG_PATH if appropriate. Never bypass TLS or silently add another source after restore failure. Inherited user feed/credential/proxy environment configuration is not automatically passed to the private runner; a mirror requiring those must be configured through a supported, reviewed mechanism rather than assumed to work.

Existing lockfiles use locked restore; a new restore produces a lockfile for review. Commit appropriate packages.lock.json, not package binaries. Without --restore, build/test use --no-restore; missing dependency evidence is an error, not permission to download. The CLI does not install a global SDK to satisfy this task.

Execute only reviewed tests

Only after the above approvals, choose the required invocation:

For a reviewed previously restored project, omit --restore. From the directory containing App, after creating App\reports, save a new sanitized report with:

Use the host OS’s path separator when invoking examples. Select an existing custom project with --test-project; there is no need to regenerate tests just to run them. --timeout bounds native steps (1-3600 seconds); identity-checked cleanup can take longer.

Use absolute paths when these different bases would be ambiguous. Report parents must exist; native report destinations must be new files. For example, after creating an App\reports directory, --path App --report-path reports/native.json writes App\reports\native.json. Using --report-path native.json instead writes App\native.json, not a file in the reports directory. Results inspection must name the same saved file relative to the shell’s working directory, not the selected project.

Interpret results without overstating coverage

Native counts do not attest that arbitrary C# invoked the SDK, executed the selected workflow, or covered useful behavior. Preserve notAttested workflow/business coverage and notProvided isolation markers; complement counts with reviewed assertions.

Native stdout/stderr is not raw-echoed by the CLI. Raw TRX stays under the dedicated test directory’s .logicapp-test-state run storage and can contain fixtures, outputs, paths, and failure details. Protect it; do not share it as a sanitized report. CLI telemetry opt-out does not establish that SDK telemetry is disabled.

From the same directory containing App, inspect that saved sanitized report and the selected workflow:

Select the actual files according to the path table. Results inspection accepts closed supported preflight/native JSON, not raw SDK/TRX. It does not follow report-embedded paths, rerun tests, or authenticate authorship. Inspection exit 0 means read success; recorded outcomes/counts are historical. Optional workflow-byte comparison detects drift, not a new execution or attested coverage.

Failure and cleanup boundaries

SDK 1.0.3/engine 1.165.50 has a documented InvalidCastException for some valid nonempty typed top-level workflow outputs. Do not delete outputs, patch binaries, or change workflow semantics to hide it. Report the actual failing test/SDK limitation and seek a separately qualified publisher fix. Action-output assertions are not a universal workaround.

If cleanup cannot safely complete, retain the process receipt and ownership evidence. A later run refuses new work while such a receipt remains, including malformed receipts. Do not delete receipts or kill processes by name to unblock execution. Inspect/reconcile ownership explicitly; report unresolved cleanup. Old completed reports alone do not grant permission to stop anything.

Keep native tests/state in dedicated directories. Do not colocate them with host/workflow runtime content. Packaging excludes verified dedicated native test content and raw TRX, but arbitrary trusted code may write elsewhere: inspect package contents independently.

Gate: Required native assertions actually executed, counts/outcomes/limitations reported, approved source/terms/trust boundaries preserved, sensitive artifacts excluded, and no hidden cleanup or dependency blocker.