[Ecosystem] Cross-language context engineering: Meet ContextChef in TypeScript

[MyPrototypeWhat]

Hi @rhettg and the contextwindow team,

I'm the author of ContextChef, a context compiler for TypeScript/JavaScript AI agents. I recently came across contextwindow and noticed we're solving a closely related problem — managing LLM conversation context with compression and tool handling — but targeting different language ecosystems.

As of April 2026, ContextChef is at v3.0.3 (core) / v1.1.0 (AI SDK middleware). Its current scope includes:

• History compression: tokenizer-based or reportTokenUsage trigger, summarization via configurable small model, mechanical compaction for zero-cost pruning
• Tool pruning: flat tag-based filtering or two-layer namespace architecture (stable core tools + on-demand toolkit loading) to prevent tool hallucinations
• Multi-provider compilation: same prompt structure compiles to OpenAI / Anthropic / Gemini with automatic cache breakpoints, prefill, and tool call format adaptation
• VFS offloading: auto-truncate large tool outputs, store originals behind context:// URIs for on-demand retrieval
• Cross-session memory: key-value store with TTL, injected as XML on compile
• Snapshot & Restore: full internal state capture for branching and rollback
• Vercel AI SDK middleware: drop-in integration with zero code changes

I think our two projects are complementary rather than competing:

	contextwindow (Go)	ContextChef (TypeScript)
Language	Go	TypeScript/JavaScript
Compression	Summary model	Summary model + mechanical compaction + circuit breaker
Tool management	Tool calling support	Dynamic pruning + namespace lazy-loading
Persistence	SQLite (built-in)	In-memory (pluggable storage adapters)
Provider support	OpenAI, Gemini, Claude	OpenAI, Anthropic, Gemini (format adapters)
State injection	—	Zod schema → XML injection
Positioning	Low-level Go library	Context compiler with modular pipeline

Go and TypeScript are probably the two most popular languages for building AI agents today. I believe cross-linking our projects would help developers discover the right tool for their stack.

Would you be open to:

1. Cross-linking our projects in each other's READMEs so users can find the implementation that matches their language
2. Sharing design notes on compression strategies — your SQLite persistence model is something we've been considering, and our namespace-based tool pruning might be interesting for contextwindow's roadmap

Happy to point to specific implementation areas if any of this is of interest.

Best,
MyPrototypeWhat