wukong

# wukong

> **Memory Efficient Coding Agent (MECA)** — reference implementation.

---

## MECA

MECA (Memory Efficient Coding Agent) is a class of coding agent defined by the following properties:

1. **VRAM budget is a first-class constraint** — the agent's tool schema, context management, and task decomposition are designed around a defined VRAM ceiling, not retrofitted to one
2. **Tool schema surface is minimized for reliable small model parsing** — flat, typed, no nested objects
3. **Context is actively managed per step** — token spend is tracked and pruned explicitly, not truncated as a fallback
4. **Task decomposition happens before model invocation** — tasks are broken into subtasks bounded by small model working memory limits before the model sees anything
5. **Reliability is instrumented and published** — benchmark scores per task class are part of the artifact, not anecdotal

An agent that supports local models but was not designed around these constraints is not MECA-class — it is a large-model agent with local inference shimmed in.

---

## What Is It ?

wukong is the first MECA-class implementation. Primary target: 4–8GB VRAM via ollama.

Aider, opencode, goose — all built for large models. Run them on a 7b model and they degrade silently: context assumptions are wrong, tool schemas are too complex for reliable small model parsing, no budget awareness.

wukong is built the other way. The 7b model is the design target. Everything follows from that.

---

## Architecture - Justification

| Problem | What existing agents do | What wukong does |
|---|---|---|
| 7b models hallucinate complex tool call formats | Simplify nothing, degrade silently | Flat tool schema with minimal surface area |
| Context bloat tanks small model output quality | Truncate or ignore | Active token budget tracking with explicit pruning per step |
| Full tasks exceed small model working memory | Pass full task, let model figure it out | Rule-based decomposition into bounded subtasks