Pular para o conteúdo
TakoAPI
Voltar às skills

context-gatekeeper

por davienzomq

AI & LLMs
1.2kDownloads informados pelo clawskills.sh
3Estrelas
0Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

Keeps the conversation token-friendly by summarizing recent exchanges, surfacing pending actions, and delivering.

Descrição

O que esta skill faz

Manages conversation token usage by summarizing history, extracting pending actions, and writing a compact briefing file before each model call. A Python script reads a plain-text exchange log and outputs a replacement context document. Run it before each prompt instead of resending the full thread.

Rather than trimming the raw conversation ad hoc, it produces a structured, size-bounded briefing that separates decisions, pending tasks, and recent turns into predictable sections.

Quando usá-la

  • Pruning a bloated multi-turn agent thread before the next API call
  • Tracking open TODOs and follow-ups across a long session
  • Reducing token costs on repetitive agentic workflows
  • Injecting a one-page decision summary into a fresh prompt
  • Auditing what the model knew at each turn via versioned summary files

Exemplo de fluxo de trabalho

ENTRADA

User asks: Summarize our Q2 planning session and keep the next prompt lean

AGENTE
  1. 1

    Append the latest user and assistant messages to context/history.txt in ROLE: text format

  2. 2

    Run context_gatekeeper.py with --history and --summary flags

  3. 3

    Script produces context/current-summary.md with up to 6 summary sentences, open action items, and the last 4 turns

  4. 4

    Inject the summary file content at the top of the next prompt

  5. 5

    Proceed with the API call using only the briefing plus 2-3 recent messages if needed

SAÍDA

A context/current-summary.md file containing a condensed briefing, pending task list, and recent exchange snippet ready for injection

Instalação

clawhub install davienzomq/context-gatekeeper

Execute o comando da CLI ou cole o prompt no Claude Code, Cursor, Codex ou no agente de código que você usa.

Links