Examples

Runnable, self-contained examples live in the repository's examples/ directory. Most run fully offline (mock summarizers, no API key); the ones that call a real model tell you which env var to set. Run any of them with npx tsx examples/<name>.ts.

Core

Example	What it demonstrates
basic-chat.ts	ContextChef + Janitor for automatic history compression, reportTokenUsage-style token tracking, compiling to OpenAI format. Needs OPENAI_API_KEY.
compression-v2.ts	The v4 janitor pipeline: triggerRatio pre-rot trigger, pinned messages surviving compression, archive: 'vfs' + chef.resolveRecall, compressionGuidelines, minShrinkRatio + validateCompression quality gates, granular compression events, and compressionMode: 'incremental-anchored'. Fully offline.
durable-compaction.ts	planCompaction turn-safe splits, compactHistory one-shot compaction, the result === history no-op persist pattern, and the contrast with in-flight compression. Fully offline.
guardrail.ts	withGuardrails enforce-XML + prefill, provider-aware prefill degradation, withGuardrails(null) clearing, and v4 order-independence. Fully offline.
memory.ts	Memory with InMemoryStore, turn-based and wall-clock TTL, both memoryPlacement modes and why the tail placement is KV-cache friendly, injection selectors, and the compiled <memory> block + memory tools. Fully offline.
multi-provider.ts	The same prompt compiling to OpenAI, Anthropic, and Gemini formats — cache breakpoints for Anthropic, prefill degradation across providers. Fully offline.
skills.ts	Skill objects and SKILL.md loading, registerSkills + activateSkill, renderSkill placeholder substitution, formatSkillListing, and meta.activeSkillName. Fully offline.
snapshot-restore.ts	chef.snapshot(label) / chef.restore(snap) rollback, guardrailOptions surviving the round-trip, Janitor state (including the anchored-mode anchor doc) in snapshots. Fully offline.
tool-pruning.ts	Flat mode tag-based pruning by task, plus the two-layer architecture: stable namespaces + on-demand toolkits. Fully offline.
vfs-lifecycle.ts	offloadAsync head/tail markers + URIs, the offload:created event, the getRecallToolDefinition + chef.resolveRecall round-trip, and cleanupAsync with maxFiles / maxBytes LRU eviction. Fully offline.

Middleware

Example	What it demonstrates
ai-sdk-middleware.ts	withContextChef compress + truncate + compact in one wrapper, per-conversation sessionId isolation, the anti-double-compression guard for Anthropic server-side context management, and durable compaction with compactModelMessages. Constructs offline; the model call needs OPENAI_API_KEY.
tanstack-ai.ts	contextChefMiddleware with compress + truncate + compact on @tanstack/ai 0.44, threadId keying, and durable compaction with compactTanStackMessages. Constructs offline; the chat() call needs OPENAI_API_KEY.

Running the examples

git clone https://github.com/MyPrototypeWhat/context-chef.git
cd context-chef
pnpm install
pnpm build

npx tsx examples/compression-v2.ts   # offline
export OPENAI_API_KEY=your-key
npx tsx examples/basic-chat.ts       # calls the real API