convez.ai
SkillAutomaçãoIntermediário
Adityapal67

Context Graph Compressor

Adityapal67/context-graph-compressor

Comprima conversas em grafos JSON portáteis para handoffs entre chats e modelos

Ver no GitHub
BomPluginScore · de 0 a 100

Skill que comprime conversas longas com IA em grafos de estado portáteis e estruturados em JSON, otimizando handoffs entre sessões e modelos diferentes. Reduz até 97% do consumo de tokens ao transferir contexto, com suporte a múltiplas LLMs como GPT, Gemini e Mistral.

Instalação

$npx convez-ai Adityapal67/context-graph-compressor

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
37
Forks
9
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Transferir contexto de projeto entre sessões de chat sem reenviar histórico completo
  • Compartilhar estado de conversa entre diferentes modelos de IA (Claude, GPT, Gemini)
  • Arquivar e reutilizar contexto estruturado com tags de importância (high/medium/low)
  • Otimizar custo de tokens em fluxos de trabalho com múltiplas conversas

Veredito

Solução bem pensada para um problema real: compressão de contexto estruturada e portável, com metadados (importância, status, dependências) que outras ferramentas não oferecem. Formato JSON máquina-legível é a principal vantagem sobre markdown/bullets, e o suporte cross-LLM amplia seu alcance.

README

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

context-graph-compressor

A Claude skill that compresses any chat conversation into a minimal, portable JSON context graph — so you can continue the same conversation in a new chat or a different LLM at a fraction of the token cost.


The Problem

Every message in a long chat re-sends the entire history:

Turn 50:  [turn1 + turn2 + ... + turn49 + your message]  →  ~15,000 tokens

The Solution

Compress the chat once, paste the output into a new session:

New chat turn 1:  [compressed graph ~400 tokens + your message]  →  done

Same context. 97% fewer tokens.


Why This Is Different

Several context-handling solutions already exist. Here's how this one compares:

This skillClaude's built-in compactionHandoff skills (Smithery / ClaudSkills)
Works across sessions✅ Export & paste anywhere❌ Same chat only✅
Works with other LLMs✅ GPT, Gemini, Mistral, etc.❌ Claude only❌ Claude Code only
Output formatJSON graphInternal (not exportable)Plain text / markdown bullets
Machine-readable✅ Structured, parseable❌❌
Importance tagging✅ h/m/l per node❌❌
Status tracking✅ open/resolved/deferred/blocked❌❌
Graph relationships✅ depends_on, resolves, supersedes...❌❌
Assumptions explicit✅ A node type with confidence❌❌
Downloadable file✅ .json❌❌
Token target~400–800 tokensVaries, not optimizedVaries

In plain terms

Claude's built-in compaction kicks in automatically when a chat gets too long — but it works silently inside the same conversation. You can't export it, restart clean, or take it to another LLM. It's a life raft, not a handoff.

Existing handoff skills produce a markdown bullet list — readable by humans, but not structured enough for a machine. There's no importance ranking, no node graph, no way to programmatically filter what a new LLM should prioritize. And they're all built for Claude Code, not Claude.ai.

This skill treats context as data, not prose. The JSON graph is structured so any LLM can parse it, importance tags tell the receiving model what it must know vs. what's optional, and compact mode is explicitly optimized to hit the lowest possible token count for the handoff — not just "shorter than the original."


What It Outputs

A structured JSON graph with two modes:

compact (default) — for new chat handoff

~300–500 tokens. Abbreviated keys, telegraphic summaries, no fluff.

{
  "v": 1,
  "mode": "compact",
  "desc": "Building a JWT-secured Spring Boot API with MySQL.",
  "n": [
    {
      "id": "n1",
      "t": "F",
      "i": "h",
      "s": "Spring Boot 3.2, Java 21, MySQL. Maven. Constructor injection throughout.",
      "c": [
        { "id": "n1.1", "t": "F", "i": "m", "s": "DB: app_db. ddl-auto=update. Config in application.yml." }
      ]
    },
    {
      "id": "n2",
      "t": "P",
      "i": "h",
      "s": "403 on all endpoints fixed: added .requestMatchers('/auth/**').permitAll() before anyRequest().authenticated()."
    }
  ],
  "handoff": "Continue a Spring Boot JWT API project. Stack: Java 21, MySQL, JPA. Auth working (403 fixed). N+1 query fixed with @EntityGraph. Refresh tokens deferred."
}

readable — for humans / archiving

Full field names, prose summaries, complete handoff prompt. Good for sharing with teammates or saving a project's history.


How to Use

Step 1 — Compress your chat

In any Claude chat, say:

compress this chat

or

save context / handoff to new chat / too many tokens, start fresh

The skill outputs the compact JSON and saves a downloadable .json file.

Step 2 — Paste into new chat

Start a new chat and paste:

Context from previous session:
<paste JSON here>

Continue from this state.

That's it. The new LLM picks up exactly where you left off.


Node Types

ShortFullMeaning
FfactTechnical facts, stack, config, versions
DdecisionChoices made, things ruled out
PproblemBugs, errors, and their fixes
GgoalWhat the user is building / trying to achieve
CcodeSnippets, method names, annotations
XcontextOpen/unresolved threads

Importance Levels

ShortMeaning
hHigh — losing this breaks continuity
mMedium — useful but not critical
lLow — dropped in compact mode

Installation

  1. Download context-graph-compressor.skill
  2. Go to Claude.ai → Settings → Skills
  3. Upload the .skill file
  4. Done — Claude will use it automatically when you ask to compress a chat

Works With

  • Claude (claude.ai)
  • Any LLM that accepts a system/user prompt — GPT-4, Gemini, Mistral, etc.
  • Just paste the JSON + "Continue from this state." as your first message

Files

context-graph-compressor/
├── SKILL.md       — Core skill instructions for Claude
└── README.md      — This file

License

MIT — free to use, modify, and share.

Ver o original no GitHub

Bom para

Relacionados