convez.ai
AgenteDesenvolvimento WebIntermediário
alxstx

Token Minimizer

alxstx/tokenminimizer

Harness markdown que reduz tokens em agentes de IA mantendo índice em contexto

Ver no GitHub
InicialPluginScore · de 0 a 100

Um harness markdown agnóstico de ferramentas para agentes de IA em codificação (Copilot, Cursor, Claude Code) que minimiza consumo de tokens através de índices em contexto e carregamento sob demanda. Mantém um sumário sempre carregado (AGENTS.md + MEMORY.md) e puxa detalhes apenas quando necessário, permitindo que agentes trabalhem em repositórios sem re-ler todo o contexto a cada sessão.

Instalação

$npx convez-ai alxstx/tokenminimizer

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Reduzir consumo de tokens em projetos grandes com agentes Copilot/Cursor/Claude
  • Implementar fluxo de plan→implement→verify com invocação sob demanda
  • Manter memória durável e organizada em arquivos markdown entre sessões
  • Escalar trabalho com agentes sem explosão exponencial de tokens

Veredito

Solução técnica genuinamente pensada para o problema real de consumo de tokens em fluxos de agentes IA. Abordagem pragmática com filosofia clara (progressive disclosure + invocation over always-on) que escala melhor que frameworks de multi-agent convencionais.

README

A tradução para o português deste README ainda está na fila.

token-harness

A token-minimizing, tool-agnostic markdown harness for doing implementation work with AI coding agents (GitHub Copilot, Cursor, Claude Code, Codex…).

It keeps a tiny always-loaded brief plus a live memory index in context and pushes everything else to files loaded on demand — so an agent can work in your repo without re-reading the whole thing every session.

Core rule: index in context, detail on disk. Only AGENTS.md and memory/MEMORY.md load by default (a few hundred tokens). Everything else is pulled in just-in-time.

How it works

always loaded, every session ─────▶  AGENTS.md  +  memory/MEMORY.md           (~1–2K tokens)
loaded only when the task needs ──▶  memory/{architecture,decisions,tasks}.md · source files
invoked on demand (not always-on) ▶  /bootstrap-fill /bootstrap-verify /plan /implement /verify-change

Quickstart

  1. Copy AGENTS.md, memory/, harness/, and .github/ into your repo.
  2. Run /bootstrap-fill (Copilot) or paste harness/prompts/1-bootstrap-fill.md — an agent fills AGENTS.md and seeds memory/ from your repo.
  3. Run /bootstrap-verify — a second agent checks it against the real repo and flags/fixes errors.
  4. Skim the report; fix anything flagged. (~5 min)
  5. Per task: /plan → approve → /implement → /verify-change.

What's inside

PathLoadedPurpose
AGENTS.mdalwaysProject brief + operating contract (the template). < 200 lines.
memory/MEMORY.mdalwaysLive state + index. < 60 lines.
memory/{architecture,decisions,tasks}.mdon demandMap / why-log / active plan.
.github/prompts/*.prompt.mdinvokedCopilot slash commands: /plan, /implement, /verify-change, /bootstrap-fill, /bootstrap-verify.
.github/instructions/*.instructions.mdpath-scopedRules that load only for files matching applyTo.
.github/copilot-instructions.mdalways (Copilot)Forces "read memory first" behavior.
harness/referenceCanonical prompts, templates, and full docs (harness/README.md).

Principles

  • Progressive disclosure — an index in context, detail on disk; on-demand files cost nothing until opened.
  • Externalized memory — durable facts live in memory/ and are read back later, not held in a context window that resets each session.
  • Invocation over always-on — the "agents" are prompts you fire, not definitions loaded on every message.

Full design notes, token math, customization, and FAQ: harness/README.md.

Why so few agents

Existing agent frameworks tend to add agents — and tokens; some ship hundreds of thousands of tokens of always-on agent definitions, and multi-agent setups routinely burn 4–15× the tokens. This harness does the opposite: prompts you invoke instead of always-on agents, and only three roles that pay for themselves — onboard, plan, verify — plus the main loop. The token win is progressive disclosure + invocation isolation, not more agents.


A starting point — fork it, tune the prompts, make it yours.

Ver o original no GitHub

Bom para

Relacionados