Add .NET performance analysis skill and optimizing agent

[artl93]

.NET Performance Analysis Skill & Agent

Adds the analyzing-dotnet-performance skill and optimizing-dotnet-performance agent — a distilled, actionable set of ~50 .NET performance patterns designed to detect and fix anti-patterns in real codebases.

Where this comes from

Compiled from Stephen Toub's .NET performance blog series spanning .NET Core 2.0 through .NET 10. The full reference was shrunk to fit a skill context window, organized into 7 topic-specific reference files loaded on-demand based on code signals.

What's in the skill

Reference Guide	Patterns	Covers
critical-patterns.md	17	Deadlocks, crashes, ReDoS, >10x regressions — always loaded
async-patterns.md	5	ValueTask, Channels, false sharing
memory-and-strings.md	9	Span<T>, stackalloc, UTF8 literals, string allocation chains
collections-and-linq.md	9	FrozenDictionary, CollectionsMarshal, params span
io-and-serialization.md	6	JSON source generators, async FileStream, static readonly
regex-patterns.md	4	GeneratedRegex, NonBacktracking, EnumerateMatches
structural-patterns.md	1	Sealed classes (absence-based detection)

Agent: Two-Pass Analysis

1. Pass 1 (LLM-Only): Fast, intuition-driven review using the agent's own .NET knowledge — no skills loaded.
2. Pass 2 (Skill-Based Deep Scan): Loads the skill for systematic pattern matching. Deduplicates against Pass 1 and reports only new discoveries.

Key features

• Progressive disclosure: Critical patterns always loaded; topic files loaded based on code signals
• Scan recipes: grep-based detection with mandatory execution checklist
• Verify-the-Inverse: Absence patterns report ratios ("N of M classes are sealed")
• Cross-file consistency: Flags sibling files using un-optimized equivalents
• Scale-based severity: Same anti-pattern at 50+ sites escalates to codebase-wide issue
• Common Pitfalls: Guards against over-optimization (e.g., unsafe for micro-opts, LINQ removal outside hot paths)

Real-world validation

Validated against Humanizer: +8.9% average speed, +5.1% memory reduction across 81 BenchmarkDotNet benchmarks with zero functional regressions. PR example.

Comparison across runs — Run "B" is Claude Opus with no skills, Run "H" is this agent + skill.

Files

agents/optimizing-dotnet-performance.agent.md
skills/analyzing-dotnet-performance/
├── SKILL.md
└── references/
    ├── critical-patterns.md
    ├── async-patterns.md
    ├── collections-and-linq.md
    ├── io-and-serialization.md
    ├── memory-and-strings.md
    ├── regex-patterns.md
    └── structural-patterns.md

[artl93]

Toub is reviewing results.

[artl93]

found a critical bug in the agent.

[artl93]

Prompt:

look at src. Looking only at the relevant shipping code and using your agent and skills, recommend performance optimizations for this library. This library can be used in tight loops and needs optimization for things like memory and CPU. Make suggestions, rank them, note them by categories and their potential for impact. Make suggestions - do not edit. We will consider edits later.

My test repo is Humanizr/Humanizer.

[artl93]

Updated https://gist.github.com/artl93/5efc3dd3a6386bc388890b1aa37898f4 that at least hits the top items of both skills (skilled / agent is "A", llm only ls "B", 2-pass combined is "C")

[Copilot]

Pull request overview

This PR introduces a comprehensive .NET performance analysis capability with two components: the analyzing-dotnet-performance skill for systematic anti-pattern detection and the optimizing-dotnet-performance agent for guided performance optimization. The skill implements a progressive disclosure pattern with 7 reference files covering ~50 performance patterns distilled from Stephen Toub's .NET performance blog series, organized by severity (Critical/Moderate/Info) with concrete code fixes.

Changes:

• Adds analyzing-dotnet-performance skill with progressive reference loading based on code signals
• Adds optimizing-dotnet-performance agent implementing mandatory two-pass analysis (LLM intuition + systematic scanning)
• Includes 4 test fixtures demonstrating critical anti-patterns and eval scenarios for validation

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/dotnet/agents/optimizing-dotnet-performance.agent.md	Defines agent with two-pass workflow: direct analysis followed by skill-based deep scan with deduplication
src/dotnet/skills/analyzing-dotnet-performance/SKILL.md	Main skill definition with progressive disclosure workflow, severity classification, and validation checklist
src/dotnet/skills/analyzing-dotnet-performance/references/critical-patterns.md	17 critical patterns causing deadlocks, regressions, or excessive allocations with grep detection recipes
src/dotnet/skills/analyzing-dotnet-performance/references/async-patterns.md	5 async/concurrency patterns including ValueTask usage, Channels, and false sharing mitigation
src/dotnet/skills/analyzing-dotnet-performance/references/collections-and-linq.md	9 collection patterns covering FrozenDictionary, CollectionsMarshal, params span optimizations
src/dotnet/skills/analyzing-dotnet-performance/references/io-and-serialization.md	6 I/O patterns for streaming, async FileStream, Memory overloads, and static readonly devirtualization
src/dotnet/skills/analyzing-dotnet-performance/references/memory-and-strings.md	9 memory patterns including Span, stackalloc, UTF8 literals, and string allocation chains
src/dotnet/skills/analyzing-dotnet-performance/references/regex-patterns.md	4 regex patterns for GeneratedRegex, NonBacktracking, EnumerateMatches, and span-based APIs
src/dotnet/skills/analyzing-dotnet-performance/references/structural-patterns.md	Absence-based detection for unsealed classes with scale-based severity thresholds
src/dotnet/tests/analyzing-dotnet-performance/fixtures/critical-antipatterns.cs	Test fixture demonstrating sync-over-async, ReDoS vulnerability, and string concatenation in loops
src/dotnet/tests/analyzing-dotnet-performance/fixtures/collections-and-allocations.cs	Test fixture showing LINQ on hot paths, ContainsKey+indexer double-lookup, and O(n²) duplicate detection
src/dotnet/tests/analyzing-dotnet-performance/fixtures/io-and-async-issues.cs	Test fixture demonstrating HttpClient per-instance, sequential async, and file reading anti-patterns
src/dotnet/tests/analyzing-dotnet-performance/fixtures/struct-missing-iequatable.cs	Test fixture showing struct without IEquatable causing boxing in comparisons
src/dotnet/tests/analyzing-dotnet-performance/eval.yaml	4 test scenarios validating skill detection of critical patterns, collections issues, I/O problems, and struct performance concerns

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jeffschwMSFT]

lgtm. I would like to see the output from the skill validator

[jeffschwMSFT]

lgtm. I think once we have this interacting with other skills we may have more feedback. given the size we would want this to only apply when doing performance investigations. I wonder if the simd skill is better suited as part of this suite (we can explore once merged).

[danmoseley]

I didn't read the whole thing line by line as it was already. I jumped around and also used AI. I think it's good

[Copilot]

Pull request overview

Copilot reviewed 21 out of 22 changed files in this pull request and generated 6 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.