Zanges
diff --git a/‎.claude/skills/review-pr/SKILL.md‎
Lines changed: 178 additions & 0 deletions b/‎.claude/skills/review-pr/SKILL.md‎
Lines changed: 178 additions & 0 deletions
diff --git a/‎.editorconfig‎
Lines changed: 90 additions & 0 deletions b/‎.editorconfig‎
Lines changed: 90 additions & 0 deletions
@@ -0,0 +1,178 @@
+---
+name: review-pr
+description: Run the Mouse2Joy multi-angle review locally on a PR or the current branch. Spawns 8 review subagents (security, correctness, regression, ux, stability, convention-compliance, performance, code-reuse) using the same reviewer prompts as the CI pipeline. Reports findings in the conversation instead of posting to GitHub. Use when the user wants to review a PR before opening it, double-check a branch, or get a second opinion on changes without waiting for CI.
+---
+
+# review-pr
+
+This is the **interactive / local** counterpart to the CI workflow at
+[`.github/workflows/claude-review.yml`](../../../.github/workflows/claude-review.yml).
+Same 8 reviewer angles, same prompts, but it runs from a Claude Code session
+and reports findings as text in the conversation instead of posting a GitHub
+review.
+
+## When to use
+
+Trigger this skill when the user asks for things like:
+
+- "Review this PR" / "review PR #12" / "give me a review of branch X"
+- "Check my branch before I open the PR"
+- "What would the CI reviewer flag on this?"
+- "Run the multi-agent review locally"
+
+Do **not** trigger if the user just wants a quick code-quality scan of one
+file or a specific question about one change — those don't need 8 angles.
+
+## What it does
+
+1. Resolves the target: PR number (via `gh pr view <n>`), explicit branch
+   name, or "the current branch" by default.
+2. Loads the 8 reviewer system prompts from
+   [`.github/claude-review/reviewers/`](../../../.github/claude-review/reviewers/).
+   These files are the **single source of truth** — they are also used by
+   the CI workflow.
+3. Lists the changed files in the diff.
+4. Runs each reviewer as a `Task` subagent in turn (sequentially, not in
+   parallel — the local skill optimizes for transparency in the conversation
+   over throughput).
+5. Aggregates findings into a single response in the conversation, with one
+   section per angle.
+
+## How to run it
+
+Follow these steps in order. The user is in this Claude Code session; you
+are the orchestrator.
+
+### Step 1: Resolve the target
+
+Look at the user's request:
+
+- If they named a PR number (e.g. "review PR #8"): use that.
+- If they named a branch: use `git diff <base>...<branch>` against
+  `origin/main`.
+- Otherwise: default to the current branch vs `origin/main`. Run
+  `git rev-parse --abbrev-ref HEAD` and check that it isn't `main`. If it
+  is, ask the user which branch / PR they meant.
+
+### Step 2: Identify changed files
+
+Run one of:
+
+- For a PR: `gh pr view <n> --json files --jq '.files[].path'`
+- For a branch: `git diff --name-only origin/main...<branch>`
+
+Hold this list — you'll pass it to every reviewer.
+
+If the list is empty, say so and stop.
+
+If the list is *only* docs (`ai-docs/**`, `**/*.md`, `LICENSE`) or *only*
+dependency files (`Directory.Packages.props`, `global.json`, csproj package
+references), tell the user "this matches the CI gate's skip rules and
+wouldn't be reviewed automatically — running anyway since you asked
+explicitly" and continue.
+
+### Step 3: Load reviewer prompts
+
+Read all 8 files from `.github/claude-review/reviewers/`:
+
+- `security.md`
+- `correctness.md`
+- `regression.md`
+- `ux.md`
+- `stability.md`
+- `convention-compliance.md`
+- `performance.md`
+- `code-reuse.md`
+
+You'll inject each one as the `Task` subagent's instructions.
+
+### Step 4: Run reviewers sequentially
+
+For each of the 8 reviewer prompts, invoke the `Task` tool with:
+
+- `subagent_type: "general-purpose"`
+- `description`: short (e.g. "Security review of PR #8")
+- `prompt`: the full text of the reviewer prompt file, followed by:
+
+  ```
+  ---
+
+  ## This review
+
+  Target: <PR #N | branch <name>>
+  Base: origin/main
+  Head: <head ref>
+
+  Changed files:
+  <bulleted list of paths>
+
+  Use Read / Grep / Glob / Bash (for `git diff` and `gh pr diff`) to
+  investigate. Don't dump everything — pull only what you actually need
+  for your angle. Output the structured JSON your instructions require.
+  ```
+
+After each subagent returns, summarize its findings into the conversation
+under a short heading like `## Security (N findings)` so the user sees
+progress, then move on. Don't try to render the JSON literally — turn it
+into readable markdown:
+
+- Summary as a sentence.
+- Critical findings as bullets with `path:L<line>` references.
+- Suggestions / nits collapsed under a "more" or omitted if low value (the
+  user is reading in real-time; respect their attention).
+
+### Step 5: Final wrap-up
+
+After all 8 finish, post one short consolidated summary:
+
+- "**Critical**: N findings" with a one-line description each.
+- "**Suggestions worth considering**: N" — list 3–5 of the most useful.
+- Token usage if you tracked it (optional).
+
+Offer to dive deeper into any specific finding the user wants to discuss.
+
+## Important constraints
+
+- **Do NOT post to GitHub.** This skill is local-only. The CI workflow
+  handles GitHub reviews. If the user wants the same review posted as a
+  GitHub review, point them at the CI pipeline (push the branch, open
+  the PR, CI runs the review automatically).
+- **Sequential, not parallel.** The CI script fan-outs 8 agents in
+  parallel; this skill runs them one at a time so the user sees progress
+  and can interrupt. If they want speed, the CI workflow exists.
+- **Reviewer prompts are the contract.** Don't paraphrase or summarize
+  them when injecting into a subagent — pass the file content verbatim.
+  This is what keeps CI review and local review in sync.
+- **Conventions referenced by the prompts** (`CLAUDE.md`, the
+  "Things that are easy to get wrong" section, `ai-docs/implementations/`)
+  are read by the subagents themselves via tool calls. You don't need to
+  pre-load them; the agent will reach for them when relevant.
+
+## What this skill does NOT do
+
+- Open PRs / push branches / merge anything.
+- Modify files (it's a review, not a fix).
+- Run the .NET build or tests (CI does that; this skill is post-CI in
+  spirit, even if there's no formal gate).
+- Replace [`/review`](https://docs.claude.com/claude-code) — that's the
+  generic Claude Code review command. This skill is the Mouse2Joy-specific
+  8-angle methodology.
+
+## Example invocations
+
+User: "Review my current branch."
+
+You:
+1. `git rev-parse --abbrev-ref HEAD` → `feature/widget-rotation`.
+2. `git diff --name-only origin/main...feature/widget-rotation` → 4 files.
+3. Read the 8 reviewer prompts.
+4. Run each via `Task` in turn, posting per-angle summaries as you go.
+5. Final consolidated summary.
+
+---
+
+User: "Run the multi-agent review on PR #12."
+
+You:
+1. `gh pr view 12 --json files,headRefName --jq ...`
+2. Same as above, but with PR-specific context.
@@ -0,0 +1,90 @@
+# EditorConfig: https://editorconfig.org
+#
+# Pairs with <AnalysisLevel>latest-recommended</AnalysisLevel> +
+# <EnforceCodeStyleInBuild>true</EnforceCodeStyleInBuild> in Directory.Build.props.
+#
+# Philosophy: keep correctness/perf rules on, silence stylistic rules that
+# would generate noise on a working WPF + native-interop codebase without
+# improving quality. Severities here override analyzer defaults at build
+# time. With TreatWarningsAsErrors=true, anything left at "warning" becomes
+# a build failure -- choose between "warning" (must fix) and "suggestion"
+# (IDE-only hint) carefully.
+
+root = true
+
+[*]
+charset = utf-8
+end_of_line = crlf
+insert_final_newline = true
+trim_trailing_whitespace = true
+indent_style = space
+indent_size = 4
+
+[*.{yml,yaml,json,md}]
+indent_size = 2
+
+[*.{csproj,props,targets,xaml}]
+indent_size = 2
+
+[*.cs]
+# ---- C# language conventions ----
+# Use the .NET defaults; keep these as suggestions so they appear in the IDE
+# but never break the build.
+dotnet_style_qualification_for_field = false:suggestion
+dotnet_style_qualification_for_property = false:suggestion
+dotnet_style_qualification_for_method = false:suggestion
+dotnet_style_qualification_for_event = false:suggestion
+csharp_style_var_when_type_is_apparent = true:suggestion
+csharp_style_var_elsewhere = false:suggestion
+
+# ---- IDE rule severities (style) ----
+# Things we genuinely want enforced.
+dotnet_diagnostic.IDE0005.severity = warning        # remove unnecessary usings
+dotnet_diagnostic.IDE0011.severity = warning        # add braces (we already use them)
+dotnet_diagnostic.IDE0044.severity = warning        # add readonly modifier
+dotnet_diagnostic.IDE0051.severity = warning        # remove unused private members
+dotnet_diagnostic.IDE0052.severity = warning        # remove unread private members
+
+# Things we explicitly do NOT want as errors -- stylistic or churny.
+dotnet_diagnostic.IDE0008.severity = silent         # use explicit type (taste)
+dotnet_diagnostic.IDE0058.severity = silent         # expression value never used (fluent assertions)
+dotnet_diagnostic.IDE0066.severity = suggestion     # use switch expression
+dotnet_diagnostic.IDE0072.severity = suggestion     # add missing cases to switch expression
+dotnet_diagnostic.IDE0290.severity = silent         # use primary constructor (churny for existing code)
+
+# ---- Code-quality (CA) rule severities ----
+# Keep on as warnings (will be errors due to TreatWarningsAsErrors=true).
+# Explicitly silence the ones that don't fit a desktop / native-interop app.
+
+dotnet_diagnostic.CA1062.severity = silent          # validate args (we use nullable refs at boundaries)
+dotnet_diagnostic.CA1805.severity = silent          # member initialized to default value (we use it to document persisted defaults on records)
+dotnet_diagnostic.CA1303.severity = silent          # do not pass literals as localized parameters (no l10n)
+dotnet_diagnostic.CA1304.severity = silent          # specify CultureInfo (desktop app, system culture is fine)
+dotnet_diagnostic.CA1305.severity = silent          # specify IFormatProvider (ditto)
+dotnet_diagnostic.CA1307.severity = silent          # specify StringComparison for clarity (ordinal everywhere by default)
+dotnet_diagnostic.CA1310.severity = warning         # specify StringComparison for correctness
+dotnet_diagnostic.CA1311.severity = silent          # specify culture for ToUpper/ToLower
+dotnet_diagnostic.CA1716.severity = silent          # identifiers should not match keywords
+dotnet_diagnostic.CA1724.severity = silent          # type names should not match namespaces
+dotnet_diagnostic.CA1812.severity = silent          # internal class never instantiated (DI containers / WPF)
+dotnet_diagnostic.CA1822.severity = suggestion      # member can be static (often false-positive for view-models)
+dotnet_diagnostic.CA1848.severity = silent          # use LoggerMessage delegates (Serilog templates already cached)
+dotnet_diagnostic.CA1851.severity = warning         # multiple enumerations of IEnumerable
+dotnet_diagnostic.CA1860.severity = warning         # prefer Count == 0 over !Any()
+dotnet_diagnostic.CA2007.severity = silent          # ConfigureAwait(false) (WPF app context)
+dotnet_diagnostic.CA1051.severity = silent          # public readonly fields on hot-path structs (Signal, etc.) are deliberate
+
+# WPF + Win32 P/Invoke realities.
+dotnet_diagnostic.CA1031.severity = suggestion      # catch general exception types (host code legitimately does this)
+dotnet_diagnostic.CA1416.severity = silent          # platform compatibility (whole app is win-only)
+dotnet_diagnostic.CA1707.severity = silent          # underscore in identifiers (Win32 constants / test names)
+dotnet_diagnostic.CA1711.severity = silent          # identifier suffix rules
+dotnet_diagnostic.CA2101.severity = silent          # specify marshaling for P/Invoke (LPWStr default is fine)
+
+# Tests: relax a few more things that don't add value in a test project.
+[**/*.Tests/**.cs]
+dotnet_diagnostic.CA1707.severity = silent          # MethodName_With_Underscores is the test convention
+dotnet_diagnostic.CA1861.severity = silent          # avoid constant arrays as args (test data)
+dotnet_diagnostic.CA2007.severity = silent
+dotnet_diagnostic.CA1859.severity = silent          # return concrete type for perf (tests intentionally use abstractions)
+dotnet_diagnostic.IDE0051.severity = silent         # unused private member (test helpers paired with peer fixtures)