JD.SemanticKernel.Extensions.Compaction 0.1.100

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
• 🔌 MCP — Discover MCP servers from Claude Code, Claude Desktop, VS Code, Codex, Copilot, and JD canonical config with a PatternKit-backed provider chain
• 🗜️ 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