dotnet-performance-review

Serena MCP is mandatory for code (read-only here). First call mcp__serena__initial_instructions to load the Serena tool manual, then use Serena for ALL code reading / searching / navigation — prefer symbol navigation (get_symbols_overview / find_symbol / find_referencing_symbols) over whole-file reads.

.NET performance review (read-only scan)

A static, read-only anti-pattern scanner for the {{ProductName}} .NET API — it reports, it does not edit. Pairs with the architect-backend review and microbenchmarking (to confirm a fix with evidence). Aligns with our coding-standards.md and async-discipline rules.

Inputs: the source to scan (required); hot-path context + target framework (.NET 10); depth = critical-only / standard (default) / comprehensive.

Workflow

1. Signal detection → topics. Scan for signals and select topics: async/await/Task/ValueTask → async; Span</stackalloc/ArrayPool/Substring/.Replace/.ToLower/+= in loops → memory & strings; Regex/[GeneratedRegex] → regex; Dictionary</List</.ToList()/LINQ → collections; JsonSerializer/ HttpClient/Stream → I/O & serialization.
2. Scan & confirm with grep (emit a checklist with exact hit counts; 0 hits is a valid finding):
   • grep -n '\.Substring(' · grep -En '\.(StartsWith|EndsWith|Contains)\s*\(' (missing StringComparison) · grep -n '\.ToLower()\|\.ToUpper()' (use StringComparison.OrdinalIgnoreCase) · grep -n 'new HttpClient(' (check for IHttpClientFactory first) · grep -n 'RegexOptions.Compiled' vs grep -n 'GeneratedRegex' (prefer source-gen for literal patterns) · grep -n 'static readonly Dictionary<' (FrozenDictionary candidate only if never mutated).
   • Verify-the-inverse: for absence patterns report the ratio (e.g. "3 of 7 use StringComparison").
3. Classify: 🔴 Critical (deadlock/crash/security/>10×), 🟡 Moderate (2–10× / hot-path best practice), ℹ️ Info. Hot-path code elevates severity; 11–50 instances bump Info→Moderate, 50+ = systematic. Report counts.
4. Report compactly per finding: #### ID. Title (N instances) / Impact / Files as File.cs:L42 / Fix / optional Caveat — grouped by severity, ending in a summary table. Close with the non-determinism caveat (an LLM scan is a lead, not proof — confirm hot-path wins with microbenchmarking).

Correctness rules (don't over-flag)

• Memory<T> not Span<T> in async methods (aligns with our async discipline).
• LINQ only on hot paths (since .NET 7 Min/Max/Sum/Average are vectorized — no blanket ban).
• ConfigureAwait(false) only in library code (e.g. {{ProjectName}}.Shared), not in the API/handlers.
• ValueTask only for frequently-synchronous hot paths; [GeneratedRegex] only for compile-time-literal patterns.
• Never recommend unsafe / CollectionsMarshal.AsSpan without benchmarked evidence.

Our deviation from the upstream scanner

The Microsoft version flags unsealed classes as a structural perf finding. We do not — our standard deliberately drops "seal your classes" (CA1852/MA0053 → suggestion) because the domain model uses intentional inheritance. Skip the unsealed-class finding (or scope it strictly to internal non-domain hot types), and never propose sealing domain entities/aggregates.