dotnet-review

/dotnet-review — orchestrated multi-agent .NET review

You are the orchestrator. Do not review code yourself - dispatch, then format the result.

Arguments given: $ARGUMENTS

Step 1 — Resolve target and mode

• Target: a branch name, commit range, --staged, or (if empty) the current uncommitted changes. For a document/spec review target, use the dotnet-techne-code-review skill instead of this command. For verifying a story implementation against its spec/acceptance criteria, use /dotnet-qa - this command hunts defects, not spec conformance.
• Effort: pass --effort <level> straight through if given. Without it the script scales reasoning effort with the change size, so only override when the user asks.
• Mode: Cynical if --cynical was passed or the user's request language is explicitly skeptical ("tear this apart", "assume this is broken", "devil's advocate"); otherwise Standard.
• Intent pack (maintainer only, never shown to reviewers): 2-5 lines summarizing what was built and any deliberate trade-offs decided in this session, if you have that context.

Step 2 — Preferred: run the bundled workflow

If the Workflow tool is available, invoke the plugin's script (this command is your opt-in):

Workflow({
  scriptPath: "${CLAUDE_PLUGIN_ROOT}/workflows/dotnet-review.js",
  args: {
    target: "<resolved target>",
    cynical: <bool>,
    model: "<only if --model was passed>",
    effort: "<only if --effort was passed>",
    pluginRoot: "${CLAUDE_PLUGIN_ROOT}",
    intentPack: "<intent pack or omit>"
  }
})

The script does everything: a scout agent sizes the change and resolves it to two pinned commit SHAs plus classified file lists (it emits no diff text - each reviewer produces the diff itself from a read-only git diff command the script assembles, so the diff never passes through a model's output), the script picks a model tier per reviewer in code (base from size: small → sonnet, medium → session model at medium effort, large or cross-repo → opus; a security/public-API/data surface boosts only the reviewer that owns it to at least opus; the maintainer matches the strongest reviewer). Opus is the ceiling the script picks on its own - fable runs only when the user passes --model fable. Reasoning effort scales on the same breakpoints (medium, then high above 20 files or 1000 LOC, then xhigh above 50 files or 3000 LOC; Cynical adds one step; the maintainer matches the reviewers), and --effort overrides it, fans out up to five reviewers in parallel (four specialists + a generalist with no lane; the four specialists get tests and generated files filtered out of their diff, and are skipped entirely when the change has no source files), dedupes in code, and runs the maintainer verification (per-finding parallel above 10 blocking findings). Intermediate findings never enter this conversation - you receive {tiers, maintainerTier, sizing, mode, scope, findings, refuted, reviewersFailed, reviewersSkipped}.