AI Review

AI Code Review and Project Review (aiUnit)

ViceSharp ships AI-assisted Code Review and Project Review as integration tests built on SharpNinja.aiUnit. They live in tests/ViceSharp.AiReview.Tests and run a configured frontier-model agent at xUnit discovery, writing the model's findings to a log. They are report-only: a missing, unauthenticated, or failing agent yields a normalized status:"error" findings JSON and the test still passes. Per the operator contract, an AI review never fails the build, it writes its output to a log.

Why a separate test project

aiUnit targets xUnit v2 (xunit.extensibility.core 2.9.x; its review attributes derive from the v2 Xunit.Sdk.DataAttribute). The main ViceSharp.TestHarness targets xUnit v3. The two majors define the same Xunit.* types in the same namespaces, so a single project cannot reference both (it fails to compile with CS0433). The AI reviews therefore live in their own xUnit-v2 project, which also keeps the slow, non-deterministic, host-bound review calls out of the focused unit suite.

The project is in ViceSharp.slnx (so it builds and is discoverable) but is tagged [Trait("Category", "AiReview")] and excluded from the default Nuke Test gate (Category!=Determinism&Category!=AiReview&Category!=ParityPending&Category!=ParityLegacy). Run it on demand.

What it reviews

Test	Attribute	Scope
AiCodeReview_WritesFindingsToLog_NeverFails	[AiCodeReview]	The media-capture surface (ViceSharp.Core.Media, CaptureServiceHost, capture wiring) for correctness, concurrency, lifetime, and security.
AiProjectReview_WritesFindingsToLog_NeverFails	[AiProjectReview]	Project health: requirements coverage (against the MCP Server), test/Byrd readiness, layering/boundaries, packaging.

Both prompts are prefixed with a requirements-governance preamble that points the reviewer at AGENTS-README-FIRST.yaml and the MCP Server as the single source of truth for requirements.

Each review is saved as a timestamped markdown file (the prompt plus the model's response) under docs/reviews/, named {kind}-review-{yyyyMMddTHHmmssfffZ}.md. aiUnit also writes its own per-review run logs under tests/ViceSharp.AiReview.Tests/aiunit-results/ (git-ignored).

Running with the Grok Build CLI (default strategy)

aiUnit's cli adapter speaks the Claude CLI stdin protocol. To drive Grok Build instead, this repo ships a tiny Claude-protocol shim (tools/aiunit-grok-claude-shim, an executable named claude) that forwards aiUnit's prompt to grok in headless prompt-file mode. The active grok strategy in appsettings.aiunit.json is therefore Kind: cli, Command: claude, Model: grok-build, resolved through the shim on PATH.

# 1. Publish the shim (produces claude.exe).
dotnet publish tools/aiunit-grok-claude-shim/ClaudeShim.csproj -c Release `
    -o tools/aiunit-grok-claude-shim/publish

# 2. Prepend the publish folder to PATH so 'claude' resolves to the shim
#    (which calls grok), ahead of any real Claude CLI.
$env:PATH = "F:\GitHub\vice-sharp\tools\aiunit-grok-claude-shim\publish;$env:PATH"

# 3. Make sure the Grok Build CLI is installed and logged in (grok.exe on PATH).
#    Optionally point the shim at the Grok plugin root and a smaller review root:
# $env:GROK_PLUGIN_ROOT       = "F:\GitHub\mcpserver-grok-plugin"
# $env:AIUNIT_GROK_REVIEW_ROOT = "F:\GitHub\vice-sharp"

# 4. Run only the AI reviews.
dotnet test tests/ViceSharp.AiReview.Tests/ViceSharp.AiReview.Tests.csproj `
    --filter "Category=AiReview"

# 5. Read the saved reviews (prompt + response as markdown).
#    docs/reviews/  (code-review-*.md, project-review-*.md)

Other strategies

appsettings.aiunit.json also defines claude (Claude Code CLI), claude-api (Anthropic HTTP), and grok-api (xAI OpenAI-compatible HTTP). Override the active strategy at run time without editing the file:

$env:AIUNIT_STRATEGY = "claude"      # use the Claude CLI directly (no shim needed)
$env:AIUNIT_STRATEGY = "grok-api"    # xAI HTTP API; set $env:XAI_API_KEY first

If no strategy resolves (no CLI/shim on PATH, no API key), the review attributes auto-skip or return a status:"error" payload; either way the suite stays green.