🦀 Pincher — reflex engine as Quilt cells

A pinch in. A reflex out. <50ms, no LLM, zero marginal cost. Federates across cloud, workstation, ESP32.

Pincher is a reflex engine — it sits between an agent's intent and the world, intercepting patterns before they reach expensive machinery. quilt-pincher is a from-the-ground-up Quilt-synergistic version where every layer of the engine is a Quilt cell: the pinch is a formula cell, the match is a program cell, the store is a vector_store cell, the veto is a listener cell, the LLM compiler is an ai cell.

┌─ Quilt cells ──────────────────────────────────────────────────────┐ │ │ │ pinch(formula) ──▶ match(program) ──▶ score ≥ 0.80? ──▶ execute(program) │ │ │ │ │ │ │ 0.55 ≤ score < 0.80 │ │ │ ▼ │ │ │ confirm + execute (listener) │ │ │ │ │ │ │ │ score < 0.55 │ │ │ ▼ │ │ │ compile(ai) → new reflex │ │ │ │ │ │ │ ▼ │ │ └──────────────────▶ veto(listener) ──▶ store(vector_store) ──┘ │ │ └─────────────────────────────────────────────────────────────────────┘

Three platforms, one engine

☁️ Cloud

Browser, Node, Cloudflare Workers. Full features: real embedders, LLM compiler, federated storage, R2 mirror.

💻 Workstation

Node + SQLite + sqlite-vec. Persistent reflexes that survive restarts.

🔌 ESP32

no_std Rust. Subset: in-memory, no LLM, reflexes loaded from pre-compiled .nail bundle.

Pinch something —

list running containers show me containers what's running on the host check disk space how much disk free show memory usage deploy the service restart the api view logs

fast

medium

slow

reflexes

Reflex database

No reflexes yet. Pinch something new to compile one.

Execution log

Pinches will appear here. Fast = direct hit, no LLM. Medium = confirm + execute. Slow = LLM compile new reflex.

How it works: each pinch embeds the trigger into a 384-dimensional vector, queries the in-memory store for top-K reflexes, and dispatches to the appropriate tier. Fast tier (<50ms): direct hit, no LLM, no thinking. Medium tier (~1s): confirm + execute, slight overhead from confirmation. Slow tier (~5s): LLM compiles a brand-new reflex, which is then stored for next time. The whole engine is a Quilt sheet — the same code runs on cloud, workstation, and ESP32. View the source on github.com/SuperInstance/quilt-pincher.