ContextChef

① Selection — what goes into the window

Overflow strategies bound the conversation (summarize, anchored, server, reset, chain, background); the Pruner bounds the tool list; a selector bounds memory.

🧭

② Placement — where in the window it goes

A cache-stable prefix and a volatile tail. Memory, dynamic state, skills and announcements each pick a side, and cacheAudit tells you when the choice was wrong.

💾

③ Persistence — where it lives outside the window

One context store behind memory/, notes/, vfs/ and archive/ — one StorageBackend, one set of context:// addresses.

🔎

④ Retrieval — how the model reaches back

One context tool with seven commands, plus new_context, dispatched through chef.ownsTool and chef.handleTool.

🔀

⑤ Adaptation — the provider wire format

The same compiled state becomes an OpenAI / Anthropic / Gemini payload, with prefill, cache breakpoints and tool-call shapes adapted for you.

📌

Constraint pinning

pinned: true messages survive every overflow strategy verbatim and are never cleared by compact() — policy text stays intact.

♻️

Reversible archive + recall

Whatever a strategy evicts is stored and cited by URI, so exact details stay retrievable instead of guessed at.

🤝

Handoff budget

Tokens reserved above the trigger buy the model one turn to write what matters into notes/ or memory/ before the window is cut.

☁️

Server-side context management

The server() strategy delegates LLM compression to Anthropic's compaction, with a client-side fallback that keeps one config portable.

🧩

Slots

Six composition points in the compile pipeline — veto overflow, inject RAG results, rewrite the assembled messages, observe the payload.

💿

Durable compaction

planCompaction / compactHistory (plus AI SDK and TanStack ports) compact your own message store once and persist the result.

🚦

Runtime tool gate

Pruner blocklist + checkToolCall for permission, environment safety, and rate limits — KV-cache preserving by default.

🎯

Dynamic state injection

Zod schema-based state injection forces the model to stay aligned with the current task on every call.

🛡️

Output guardrails

withGuardrails enforces an XML output contract and sets an assistant prefill, auto-degraded on providers without native prefill.

⏪

Snapshot & restore

Capture and roll back full context state — history, memory, strategy state, window lineage — for branching and error recovery.

📊

Observability

A unified event system — chef.on('compress', ...) — plus optional pipeline invariants that report dropped pins and split tool pairs.

🎭

Skills

The Skill primitive bundles instructions and tool annotations per phase, loadable from SKILL.md files (Claude Code / Mastra / OpenCode compatible).

ContextChefContext compiler for TypeScript/JavaScript AI agents