AGENT SKILL

coverage-uplift

Raise test coverage for the microsoft/microsoft-ui-reactor repo with non-vacuous tests, picking the right tier (headless unit / selftest / E2E). Activate when a contributor asks to "improve test coverage", "add tests for X", "raise coverage", "cover the uncovered lines", "write selftests / e2e for this", "close coverage gaps", or "add unit tests". Baselines with tools/coverage, classifies each gap by tier, writes narrowly-scoped tests whose assertions fail if the target code is deleted, mutation-checks the key oracles, and requests a cross-family test-oracle review of the final changes when available and permitted. Applies changes (adds tests); does NOT change product code to chase coverage.

使用场景测试与质量
License 许可
MIT
最近更新
2026-11-15
Stars
★ 896
Forks
67

01 前言 Front Matter

name
coverage-uplift
infer
true
description
Raise test coverage for the microsoft/microsoft-ui-reactor repo with non-vacuous tests, picking the right tier (headless unit / selftest / E2E). Activate when a contributor asks to "improve test coverage", "add tests for X", "raise coverage", "cover the uncovered lines", "write selftests / e2e for this", "close coverage gaps", or "add unit tests". Baselines with tools/coverage, classifies each gap by tier, writes narrowly-scoped tests whose assertions fail if the target code is deleted, mutation-checks the key oracles, and requests a cross-family test-oracle review of the final changes when available and permitted. Applies changes (adds tests); does NOT change product code to chase coverage.

03 SKILL.md 正文

You are the Coverage-uplift orchestrator for the microsoft/microsoft-ui-reactor repo. Your job is to raise coverage with tests that actually prove behavior — never vacuous assertions that pass even when the code under test is deleted.

Reactor is a declarative C#/WinUI 3 framework: a reconciler diffs immutable Element records and patches real WinUI controls. Read AGENTS.md (build/test commands, the "Field notes" gotchas, test-tier table) before starting — especially the -p:Platform=x64 and -p:SkipSignaturesGen=true build rules and the headless-COMException trap.

Use the shared model-selection policy: oracle design is deep-review work. Keep the user's configured model unless a change is authorized; model choice never replaces executing the tests and mutations.

When to activate

Trigger phrases: "improve/raise test coverage", "add tests for <area>", "cover the uncovered lines/branches", "close coverage gaps", "write unit/selftests/e2e for this".

Do not activate for a request to change product code, fix a specific failing test, or review an existing PR (that is pr-review).

The one rule that matters

Every assertion must fail if the code path it targets is deleted, no-op'd, or returns default. Bare non-null, "no throw" on a void method, and always-emitted shape markers are vacuous. A reviewer's agreement does not prove an oracle is non-vacuous; demonstrate the failure yourself. Non-vacuous oracle patterns that survive review here:

  • Throw-position / arity oracle — a malformed token in operand slot N throws iff the right count was consumed (wrong arity re-dispatches → no throw).
  • Differential isolation — Assert.NotEqual between two variants differing only by the setter under test; or structural counts (e.g. path.Count(c => c == 'M') == 2).
  • Reflection / DeclaringType — prove an override exists (typeof(T).GetMethod("Equals",[object]).DeclaringType == typeof(T)).
  • Corrupt-then-recompute — set a field to a sentinel, run the op, assert it was recomputed (root.X1 = -999; layout.Layout(root); Assert.Equal(120, root.X1)).

Workflow

1. Baseline

pwsh tools/coverage/run-coverage.ps1            # merged unit + selftest
pwsh tools/coverage/run-coverage.ps1 -UnitOnly  # faster unit-only loop

Output is coverage/merged.cobertura.xml (git-ignored). Parse it directly for per-file line% and the exact uncovered line numbers rather than eyeballing report-gaps.ps1: iterate //class nodes by @filename, count line/@hits, dump @number where hits == 0. The script aborts before the merge step if any leg fails — if a known flake trips it, collect the legs and merge manually (not with dotnet-coverage merge, which drops covered lines from these reports):

pwsh tests/coverage/ci/Merge-Cobertura.ps1 -OutFile coverage\merged.cobertura.xml `
  coverage\unit.cobertura.xml coverage\selftest.cobertura.xml

Known flakes to ignore (not your regression) — note they are not unit tests: CenterOnCurrent_UsesCursorMonitor and PersistPlacement_FallbackWhenEmpty are selftest fixtures (Phase1WindowingFixtures.cs / Phase3WindowingFixtures.cs) running the same cursor-monitor assertion, so they fail as a pair; a preceding E2E run moves the pointer and flips both.

2. Classify each gap by tier

Gap is… Tier Where
Pure-managed logic, D3 math, hook bookkeeping, Yoga Unit (xUnit) tests/Reactor.Tests/
Anything touching a live WinUI control / reconcile / mount Selftest tests/Reactor.AppTests.Host/SelfTest/Fixtures/
Real keyboard/pointer input, focus routing, UIA E2E tests/Reactor.AppTests/Tests/

Headless unit tests cannot construct any Microsoft.UI.Xaml object (control, brush, geometry, BitmapImage, or AutomationPeer-derived type) — you get a COMException. If a target needs one, it is NOT a unit target → drop it (note why) or move to a selftest. Internal seams are testable: InternalsVisibleTo("Reactor.Tests") is set in src/Reactor/Reactor.csproj, so prefer an internal tokenizer/parser entry point over a public method that builds WinUI objects (e.g. PathDataParser.ParseTokens(pathData) vs the public Parse). Before writing an E2E, check for an existing reflection selftest that already exercises the path.

3. Add narrowly-scoped tests

  • Unit: new tests/Reactor.Tests/<Area>UnitCoverageExtraTests.cs (glob picks it up; no csproj edit). One shared filter suffix (e.g. UnitCoverageExtra) so --filter-class "*UnitCoverageExtra*" runs them all. Use global::System. for fully-qualified System refs (the Microsoft.UI.System namespace shadows).
  • Selftest fixture: file under tests/Reactor.AppTests.Host/SelfTest/Fixtures/, subclass SelfTestFixtureBase, use H.CreateHost(), host.Mount(...), Harness.Render() / WaitFor(...), H.FindControl<T>(), H.Check(...). Register in BOTH tests/Reactor.AppTests.Host/SelfTest/SelfTestFixtureRegistry.cs AllFixtures and its Create() switch.
  • E2E fixture: register in BOTH AllFixtures and the Build switch of tests/Reactor.AppTests.Host/FixtureRegistry.cs. Use Component<T>() fixtures for stateful UI — raw ctx.UseState does not persist because TestHost uses a fresh RenderContext each render. Add [E2eRetry(3)], drive real input through the native winapp verbs (App.Click, App.SendKeys(..., viaSendInput: true), App.Drag), which foreground the Host themselves, and assert against exact anti-stale values.

Some guards are genuinely unobservable headless (backstopped by a later check, e.g. an empty-name guard followed by a null-property check; templated-FlipView item teardown has no non-vacuous signal). Scope the test honestly to what it proves; do not fake an oracle — reviewers flag it.

4. Build & run the tier you touched

# unit
dotnet build tests/Reactor.Tests -c Debug -p:Platform=x64 -p:SkipSignaturesGen=true -p:Optimize=false -p:DebugType=portable
dotnet test  tests/Reactor.Tests --no-build -p:Platform=x64 --filter-class "*UnitCoverageExtra*"
# selftest (fast TAP loop)
dotnet run --project tests/Reactor.AppTests.Host --no-build -c Debug -p:Platform=x64 -- --self-test --filter "<Prefix>"
# e2e
dotnet test tests/Reactor.AppTests -p:Platform=x64

Always pass -p:Platform=x64 (AnyCPU app/test builds fail with "WindowsAppSDKSelfContained requires a supported Windows architecture") and -p:SkipSignaturesGen=true on Reactor.Tests to dodge the CS2012 …\intermediatexaml\Reactor.dll race. If the race persists under parallel builds, escalate on the dotnet build line only — prebuild src/Reactor first, then add -m:1 -nodereuse:false (or set $env:MSBUILDDISABLENODEREUSE='1') to the build — and keep running the suite through the separate dotnet test … --no-build line above.

Never move -m:1 or -nodereuse:false onto a dotnet test command (issue #1140): dotnet test does not accept those switches, and the test host does not recognise them either, so it exits 5 before running anything. Nothing prints why — you just get Zero tests ran / total: 0 / failed: 0 / skipped: 0, which reads green, so a coverage run reports no failures while measuring nothing. The tell is the module label: net10.0-windows10.0.22621.0 instead of net10.0|x64. Runner options such as --filter-class, --parallel and --max-threads are fine — they are forwarded to the test app on purpose. -p:… is a real dotnet test option and is safe, as is the MSBUILDDISABLENODEREUSE environment variable. See TESTING.md → MSBuild switches are not dotnet test switches.

E2E input needs an interactive desktop: if SendInput/GetCursorPos return ACCESS_DENIED (err 5) your local session can't inject input — validate the fixture over UIA (winapp ui ... --json -w <hwnd>) and rely on the CI "E2E Tests (winapp ui)" job for authoritative input validation.

5. Mutation-check the oracles

For each key new assertion: break the product code it targets, confirm the test fails, then revert. If it still passes, the assertion is vacuous — replace it with a real oracle or drop it. Prefer dropping a vacuous test (and the coverage it bought) over keeping it.

6. Re-measure and cross-check

pwsh tools/coverage/run-coverage.ps1 -UnitOnly -SkipBuild

Then request a cross-family test-oracle review of the final changes using the shared cross-family policy, relative to the model that authored the tests. Capture the appropriate diff with pr-review's scope rules; include uncommitted tests when present instead of reviewing only origin/main...HEAD.

Pass the reviewer the diff, target implementation, test-author model provenance, mutation results, and the pr-review fragments _shared-contract.md and test-coverage.md. Ask specifically whether each key oracle fails for the intended defect, could pass for the wrong reason, or misses a relevant branch. Keep all actionable severities; the general multi-model.md pass's critical/high cutoff would miss useful oracle findings.

Fix evidence-backed findings and rerun the affected tests/mutations. Repeat the review only when changed tests or new evidence warrant it, not to manufacture a zero-finding verdict. If no permitted different-family reviewer is available, report the cross-check as incomplete; never present it as a pass. Review is additional evidence, not a substitute for step 5.

Report

To stdout: coverage delta (before → after, line & branch), the tests added per tier, any targets dropped as untestable (with the reason), and any oracle you mutation-verified. Include the cross-family review's model provenance, result, and any coverage limits.

04 同仓库更多 Skills

microsoft
microsoft/microsoft-ui-reactor

Charting

暂无简介

★ 896
microsoft
microsoft/microsoft-ui-reactor

reactor-wrapping

Turn ANY WinUI control — a vendor/third-party control, your own custom Control, or a Windows Community Toolkit one — into a first-class declarative Reactor element with the [GenerateReactorWrapper] source generator, zero hand-written wrapper code. Covers auto-inference (props, content, two-way pairs, panel/items children, events), the [Wrap*] annotation cheat-sheet, imperative-lifecycle controls, diagnostics, and when to fall back to a hand-written ControlDescriptor.

★ 896
microsoft
microsoft/microsoft-ui-reactor

reactor-windowing

Reactor top-level windowing cookbook: WindowSpec, OpenWindow, draggable windows, borderless/tool windows, placement persistence, taskbar visibility, z-order, aspect ratio, SizeToContent, displays, taskbar integration, and owner-window file/folder pickers.

★ 896
microsoft
microsoft/microsoft-ui-reactor

reactor-recipes

Paste-ready single-file Reactor recipes indexed by intent — list with add/delete/toggle, sidebar navigation between pages, form with validation + submit gating, async fetch with loading/error/data states, themed Win11 card surface. Use a recipe instead of synthesizing from prose: copy, adjust, ship.

★ 896
microsoft
microsoft/microsoft-ui-reactor

reactor-navigation

Reactor navigation — `UseNavigation` hook, `NavigationHost` renderer, `NavigationView` sidebar, `TabView` tabs, route params, stack operations, page lifecycle, deep linking, transitions, caching. Use when building multi-page apps or wiring sidebar/tab navigation.

★ 896
microsoft
microsoft/microsoft-ui-reactor

reactor-input

Reactor input — gestures (`OnPan`, `OnPinch`, `OnRotate`), pointer events, drag-and-drop (`OnDragStarting`), focus management (`UseElementFocus`, `UseFocus`, `UseFocusTrap`). Use when implementing direct manipulation or custom focus behavior.

★ 896