mirror of
https://github.com/dotnet/skills.git
synced 2026-09-20 09:49:54 +08:00
Art Leonard
0cee0843ca
* Add .NET deep performance fixer agent and performance patterns skill Introduce a dotnet-performance-optimizer agent and the dotnet-performance-patterns skill with reference guides covering: - Async patterns - Collections and LINQ - Critical performance patterns - I/O and serialization - Memory and strings - Regex patterns * Align agent and skill performance reporting workflow - Add two-pass analysis methodology (Pass 1: LLM-only, Pass 2: skill-based) - Consolidate skill references to match actual reference files - Remove references to non-existent skills (dotnet-jit-optimization, dotnet-async-patterns, dotnet-sync-primitives) - Update pattern count to 85+ to reflect current reference corpus * Update performance patterns skill with expanded reference coverage - Add structural-patterns.md reference file - Expand collections-and-linq, memory-and-strings, regex patterns references - Update critical-patterns with refined rules - Expand SKILL.md with additional workflow guidance * Add missing structural-patterns.md reference to agent skill list * Add non-deterministic results disclaimer to agent and skill output * Trim agent per skill authoring best practices Apply recommendations from platform.claude.com best practices: - Remove sections Claude already knows (Core Expertise, Edge Cases, Key Principles, When to Ask, Tone) - these waste context tokens - Remove skill internal reference file listing (skill handles its own progressive disclosure) - Condense 138 lines to 55 lines (-60%) - Keep only what Claude needs: two-pass workflow, output format, skill reference, escalation guidance * Trim SKILL.md per skill authoring best practices Apply recommendations from platform.claude.com best practices: - Remove Purpose section (redundant with description frontmatter) - Remove When to Use (redundant with description triggers) - Consolidate signal table (11 rows → 5, merging same-reference rows) - Remove 50-line checklist example template (keep the rule, not the example) - Remove 25-line Recipe Completeness Verification (redundant with gate) - Remove 40-line Output Format example (Claude knows markdown) - Trim Validation from 8 items + gate to 5 essential items - Trim Common Pitfalls from 3-column to 2-column - Condense Steps 5+6 into single Step 5 - 311 lines → 134 lines (-57%) * Trim reference files per Claude skill authoring best practices Apply 'Claude is already smart' principle to reference files: - Remove 6 patterns Claude already knows (Interlocked, SemaphoreSlim dup, Array.Empty, Any() vs Count(), Enum.HasFlag, Random.Shared) - Remove IEquatable<T> duplicate (kept in structural-patterns.md) - Remove 'Why a Separate File' meta-commentary from structural-patterns.md - Remove all horizontal rules between patterns - Remove explanatory prose and obvious inline comments - Preserve all Detection sections, scan recipes, and Impact statements Total: 1,819 → 1,372 lines (-25%) * Rename skill and agent to gerund form per naming conventions - dotnet-performance-patterns → analyzing-dotnet-performance - dotnet-performance-optimizer → optimizing-dotnet-performance - Update all cross-references in frontmatter and agent skill list * Rewrite descriptions per effective description guidelines - Agent: Replace second-person trigger phrases and verbose examples with concise third-person description stating what it does and when to use it - Skill: Add specificity (85+ patterns, tiered severity) and clearer triggers - Both now follow: third person, specific key terms, what + when format * Refine performance report structure and findings format for clarity and conciseness * Remove outdated reference link from SKILL.md for clarity * Remove further reading section from SKILL.md for conciseness * Add compound allocation check to performance analysis workflow * Add performance analysis scenarios for critical patterns, collections, I/O, and struct issues * Address PR #6 review feedback from Stephen Toub Remove 15 patterns that are correctness-only, redundant, or not actionable: - async void, lock+async, Async All the Way Down, NonBacktracking, Regex timeouts, Cache Regex, Budget Compiled Regex, Seal Classes stub, ConfigureAwait(false), Parallel.ForEachAsync, async lambda Action, C#10 Interpolation Handlers, Span-Based Comparison, Utf8JsonWriter, JsonSerializerContext for AOT Clarify 3 patterns where rationale was unclear: - IEquatable<T>: explain boxing via EqualityComparer<T>.Default - IsMatch guard: explain Replace always allocates even on no-match - Chained strings: split into 3 distinct anti-pattern/fix pairs Rework 3 patterns per feedback: - CollectionsMarshal.AsSpan: qualified to ultra-hot-path-only - LINQ guardrail: moved to Common Pitfalls table - params overloads: added params ReadOnlySpan<T> as .NET 9+ alternative Fix nits: - Title: arrays → substring for slicing pattern - Remove try/finally from ArrayPool example - Qualify impact statement on boxing - Remove numbered headings for consistency across reference files - Add unsafe code pitfall to Common Pitfalls - Update pattern count 85→70, critical count 24→17 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Move skill, agent, and tests to src/dotnet/ component layout Restructure to match upstream directory conventions: - skills/ → src/dotnet/skills/ - agents/ → src/dotnet/agents/ - skill tests → src/dotnet/tests/ - Update eval.yaml fixture source paths Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Fix eval.yaml source paths to resolve relative to skill directory The validator resolves source paths via join(skillPath, source), so fixture references need ../../tests/ prefix from the skill dir. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Remove validator results from tracking and add to .gitignore Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Replace synthetic eval fixtures with Humanizer-derived needle-in-haystack test suite Replace 4 synthetic Contoso fixtures with 11 realistic test fixtures modeled on patterns from the Humanizer library performance analysis (Runs A-J). 10 eval scenarios with 51 rubric items (up from 4 scenarios / 19 rubric items): - Scenarios 1-9: patterns the skill outperformed naked model on (J vs B) - Scenario 10: baseline patterns J missed (Span inconsistencies, unsealed classes) Each fixture is 240-380 lines with ~80% clean haystack code surrounding the anti-pattern needles, including decoys, positive findings, and correct patterns as contrast. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Split scenario 10 into two focused scenarios for truncation and unsealed classes The combined scenario was timing out at 120s with 2 files and 6 rubric items. Split into: - Scenario 10: Span inconsistencies and compound method chains (Truncation.cs, 3 rubric items) - Scenario 11: Unsealed leaf classes and locale hierarchy (Locales.cs, 3 rubric items) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Increase timeout to 180s for scenarios 7 and 10 These scenarios consistently timed out at 120s before the agent could synthesize findings, causing weak/negative improvement scores despite the skill detecting correct patterns. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Increase timeout to 180s for scenarios 4, 8, 9 All three timed out at 120s — agent spent time loading skill references and running grep recipes, never reaching synthesis before cutoff. Scenarios 7 and 10 already at 180s from prior fix. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Tune SKILL.md for sandboxed eval and add results report SKILL.md changes: - Add fallback for missing reference files (skip filesystem search) - Inline core scan recipes so agent works without reference files - Add 'read small files first' guidance to reduce grep overhead - Update validation checklist for reference-optional workflow Results: 73.7% overall improvement, 11/11 scenarios strong (≥60%) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Addressing @danmoseley feedback * Update src/dotnet/skills/analyzing-dotnet-performance/references/critical-patterns.md Co-authored-by: Dan Moseley <danmose@microsoft.com> * Address PR #6 review feedback: skill docs and fixture fixes Skill/reference changes: - Remove CollectionsMarshal.AsSpan section (too much churn per stephentoub) - Remove startup/initialization as examples of non-perf-sensitive code - Split Contains into own grep with false-positive note - Clarify 'exact counts' means grep match counts - GeneratedRegex: ALWAYS for static patterns, always beneficial - Rewrite regex engine section: static->GeneratedRegex, never remove NonBacktracking - Sealed class grep catches protected and file classes - StringComparison grep uses proper ERE, catches variable args Test fixture fixes: - Fix invalid regex character class in inflector-and-regex.cs - Fix Hungarian converter cache key (use original number) - Add truncation guard for truncationString >= length - Use explicit ordered array instead of dictionary keys for Farsi/Kurdish - Preserve negative sign in ToMetric null-symbol fallback - Remove stray double semicolon Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Co-authored-by: Dan Moseley <danmose@microsoft.com>
7.2 KiB
7.2 KiB