Token Minimizer
alxstx/tokenminimizer
Harness markdown que reduz tokens em agentes de IA mantendo índice em contexto
Ver no GitHubUm 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/tokenminimizer1 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
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
- Copy
AGENTS.md,memory/,harness/, and.github/into your repo. - Run
/bootstrap-fill(Copilot) or pasteharness/prompts/1-bootstrap-fill.md— an agent fillsAGENTS.mdand seedsmemory/from your repo. - Run
/bootstrap-verify— a second agent checks it against the real repo and flags/fixes errors. - Skim the report; fix anything flagged. (~5 min)
- Per task:
/plan→ approve →/implement→/verify-change.
What's inside
| Path | Loaded | Purpose |
|---|---|---|
AGENTS.md | always | Project brief + operating contract (the template). < 200 lines. |
memory/MEMORY.md | always | Live state + index. < 60 lines. |
memory/{architecture,decisions,tasks}.md | on demand | Map / why-log / active plan. |
.github/prompts/*.prompt.md | invoked | Copilot slash commands: /plan, /implement, /verify-change, /bootstrap-fill, /bootstrap-verify. |
.github/instructions/*.instructions.md | path-scoped | Rules that load only for files matching applyTo. |
.github/copilot-instructions.md | always (Copilot) | Forces "read memory first" behavior. |
harness/ | reference | Canonical 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.