JD.SemanticKernel.Extensions

An extensible toolkit for Microsoft Semantic Kernel that bridges Claude Code skills, plugins, and hooks into SK applications, and adds context management primitives (compaction, semantic memory) for building production-grade AI agents.

Features

• 📝 Skills — Parse SKILL.md files (YAML frontmatter + markdown) into KernelFunction or PromptTemplate
• 🔗 Hooks — Map Claude Code lifecycle events (PreToolUse, PostToolUse, etc.) to SK's IFunctionInvocationFilter and IPromptRenderFilter
• 📦 Plugins — Load .claude-plugin/ directories with skills, hooks, and MCP configs
• 🗜️ Compaction — Transparent context window management with configurable triggers and hierarchical summarization
• 🧠 Memory — Semantic memory with MMR reranking, temporal decay scoring, and query expansion
• 💾 Memory.Sqlite — SQLite-backed persistent memory storage
• 🎯 Fluent API — UseSkills(), UseHooks(), UsePlugins(), AddCompaction(), AddSemanticMemory() extension methods

Packages

Package	Description	NuGet
JD.SemanticKernel.Extensions.Skills	SKILL.md → KernelFunction/PromptTemplate
JD.SemanticKernel.Extensions.Hooks	Claude Code hooks → SK filters
JD.SemanticKernel.Extensions.Plugins	Plugin directory loader
JD.SemanticKernel.Extensions.Compaction	Context window compaction middleware
JD.SemanticKernel.Extensions.Memory	Semantic memory with MMR + temporal decay
JD.SemanticKernel.Extensions.Memory.Sqlite	SQLite memory backend
JD.SemanticKernel.Extensions	Meta-package (all of the above)

Quick Start

dotnet add package JD.SemanticKernel.Extensions

Load Skills

using JD.SemanticKernel.Extensions.Skills;

var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("gpt-4o", apiKey)
    .UseSkills("./skills/")   // Scans for SKILL.md files
    .Build();

A SKILL.md file follows the Claude Code / AgentSkills.io format:

---
name: code-reviewer
description: Reviews code for quality issues
allowed-tools: [Read, Grep, Glob]
---
# Code Reviewer

Review the provided code for:
1. Bug risks
2. Security vulnerabilities
3. Performance issues

Input: $ARGUMENTS

Configure Hooks

using JD.SemanticKernel.Extensions.Hooks;

var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("gpt-4o", apiKey)
    .UseHooks(hooks =>
    {
        hooks.OnFunctionInvoking("Bash|Execute", async ctx =>
        {
            Console.WriteLine($"Validating: {ctx.Function.Name}");
        });
        hooks.OnFunctionInvoked("Write|Edit", async ctx =>
        {
            Console.WriteLine($"Post-edit hook: {ctx.Function.Name}");
        });
        hooks.OnPromptRendering(async ctx =>
        {
            Console.WriteLine("Prompt is about to render...");
        });
    })
    .Build();

Load Plugins

using JD.SemanticKernel.Extensions.Plugins;

var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("gpt-4o", apiKey)
    .UsePlugins("./my-plugin/")         // Single plugin directory
    .UseAllPlugins("./plugins/")        // All plugins in directory
    .Build();

Plugin directories follow the .claude-plugin/ convention:

my-plugin/
├── .claude-plugin/
│   └── plugin.json          # Manifest
├── skills/
│   └── reviewer/SKILL.md    # Skills
├── hooks/
│   └── hooks.json           # Hooks
└── .mcp.json                # MCP servers (future)

Meta-Package (All-in-One)

using JD.SemanticKernel.Extensions;

var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("gpt-4o", apiKey)
    .AddClaudeCodeSkills("./skills/")
    .AddClaudeCodePlugin("./my-plugin/")
    .AddClaudeCodeHooks(hooks => hooks.OnFunctionInvoking(".*", _ => Task.CompletedTask))
    .Build();

Context Compaction

Automatically compress chat history when it grows too large, preserving key context while staying within token limits.

using JD.SemanticKernel.Extensions.Compaction;

var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion("gpt-4o", apiKey)
    .Build();

// Register compaction as transparent middleware
kernel.Services.AddCompaction(options =>
{
    options.TriggerMode = CompactionTriggerMode.ContextPercentage;
    options.Threshold = 0.70;                  // Compact at 70% of context window
    options.MaxContextWindowTokens = 128_000;  // Model's context limit
    options.PreserveLastMessages = 10;         // Always keep recent messages
    options.MinMessagesBeforeCompaction = 5;   // Don't compact short conversations
});

Trigger modes:

• TokenThreshold — Compact when estimated tokens exceed an absolute count
• ContextPercentage — Compact when usage exceeds a percentage of the context window

Token estimation:

var tokens = TokenEstimator.EstimateTokens("Hello world");          // ~2 tokens
var historyTokens = TokenEstimator.EstimateTokens(chatHistory);     // Includes overhead

Semantic Memory

Store, search, and retrieve context-relevant information with embedding-based similarity, MMR diversity reranking, and temporal decay scoring.

using JD.SemanticKernel.Extensions.Memory;

// Register with in-memory backend
kernel.Services.AddSemanticMemory(options =>
{
    options.DefaultSearchOptions = new MemorySearchOptions
    {
        TopK = 10,
        MinRelevanceScore = 0.7,
        UseMmrReranking = true,
        MmrLambda = 0.7,           // Balance relevance vs diversity
        UseTemporalDecay = true,
        TemporalDecayRate = 0.01,
    };
});

SQLite persistence:

using JD.SemanticKernel.Extensions.Memory.Sqlite;

kernel.Services.AddSqliteMemoryBackend("Data Source=memory.db");

Key capabilities:

• MMR reranking — Maximal Marginal Relevance for diverse search results
• Temporal decay — Recent memories rank higher with configurable decay rate
• Query expansion — Automatically generate alternative queries for broader recall
• Pluggable backends — InMemoryBackend (default), SqliteMemoryBackend, or implement IMemoryBackend

Hook Event Mapping

Related Projects

Building

dotnet restore
dotnet build
dotnet test

License

MIT © JD