paleo

agent
Guvenlik Denetimi
Uyari
Health Uyari
  • License รขโ‚ฌโ€ License: MIT
  • Description รขโ‚ฌโ€ Repository has a description
  • Active repo รขโ‚ฌโ€ Last push 0 days ago
  • Low visibility รขโ‚ฌโ€ Only 7 GitHub stars
Code Gecti
  • Code scan รขโ‚ฌโ€ Scanned 8 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions รขโ‚ฌโ€ No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

๐Ÿฆด paleo โ€” token-saving skills for LLM agents: compress output, trim context, cap budget (Claude Code / Codex / Gemini / Hermes)

README.md

๐Ÿฆด paleo

Benchmark
npm
license

Token-saving skills for LLM agents. Cut output & context tokens ~50-70% (median ~54%) without losing code/technical accuracy.

7 composable skills: terse mode, context trim, budget cap, conversation compress, summary, JSON compact, and auto-detect.


Skills

# Skill Trigger What it does
1 paleo paleo / terse Terse output โ€” no preambles, no "I'll do X now", code-first
2 paleo-trim-context long session, big context Proactively trim stale info from context window
3 paleo-budget budget 2000 / token limit Hard token cap per task, summarize if exceeded
4 paleo-converse compress conversation Condense old turns, merge duplicates, keep last N verbatim
5 paleo-summary tldr / bulky output Shrink tool results, logs, diffs to concise intisari
6 paleo-json compact json / minify Minify structured output, collapse arrays, keep parseable
7 paleo-auto paleo-auto / auto paleo ๐Ÿ†• Auto-detect session state and enable right skills automatically

Full skill reference โ†’


Quick Start

npx skills add mocasus/paleo

Or clone + copy:

git clone https://github.com/mocasus/paleo.git

One-shot: activate all 6

In your agent chat, say paleo once โ€” the agent loads the base skill which chains the others. For automatic mode: paleo-auto.


Installation

๐Ÿš Claude Code
# Install via skills CLI
npx skills add mocasus/paleo -a claude-code

# Or manual: copy to Claude skills directory
cp -r skills/paleo* ~/.claude/skills/

Plugin format available at .claude-plugin/plugin.json.

โšก Codex (OpenAI)
# Install via skills CLI
npx skills add mocasus/paleo -a codex

# Or manual
cp -r skills/paleo* ~/.codex/skills/
๐Ÿ”ท Gemini CLI
# Install via skills CLI
npx skills add mocasus/paleo -a gemini

# Or manual: copy to Gemini skills path
cp -r skills/paleo*/ ./gemini/skills/

See gemini-extension.json for Gemini extension config.

โš™๏ธ Hermes Agent
# Install via skills CLI
npx skills add mocasus/paleo -a hermes

# Or: Hermes built-in skills install
hermes skills install mocasus/paleo

# Or manual
cp -r skills/paleo* ~/.hermes/skills/

See Hermes Integration below for detailed setup.

๐Ÿ–ฑ๏ธ Cursor
npx skills add mocasus/paleo -a cursor
๐Ÿ‘จโ€๐Ÿ’ป GitHub Copilot
npx skills add mocasus/paleo -a copilot
๐ŸŒŠ Windsurf
npx skills add mocasus/paleo -a windsurf

Hermes Integration

I use Hermes myself โ€” paleo is battle-tested here first.

Install

hermes skills install mocasus/paleo

Manual trigger

Just type in your Hermes chat (Telegram, WhatsApp, etc.):

paleo

Or for auto-mode:

paleo-auto

Example session

> [20 turns in, context filling up]
> paleo-auto
๐Ÿฆด paleo-auto: enabled paleo + trim-context + converse (session length 22 turns)

> build a REST API with FastAPI
[terse, code-first response โ€” no preamble]

> budget 2000
๐Ÿฆด paleo-budget: 2000 output tokens, hard mode

Tips

  • paleo gives immediate token savings โ€” start there
  • paleo-auto watches your session and enables the right skills (best for >15 turns)
  • Combine paleo + budget for expensive models (GPT-4, Claude Opus)
  • Use paleo-converse when context hits 60%+ capacity

Benchmarks

Benchmark

Automated benchmark runs weekly via GitHub Actions. Each task is run with and without paleo, measuring delivery tokens.

Latest: median ~54% savings across 12 task types.

Task Baseline paleo Savings
Code review 1,847 891 51.8%
Bug fix 2,234 987 55.8%
Architecture plan 2,891 1,312 54.6%
API design 1,956 923 52.8%
Config setup 1,124 498 55.7%
Documentation 2,103 1,089 48.2%
Debug log analysis 2,645 1,156 56.3%
Data pipeline 2,478 1,203 51.5%
Test generation 1,892 812 57.1%
Refactoring 2,156 1,034 52.0%
Shell scripting 1,467 612 58.3%
SQL queries 1,345 598 55.5%

Full benchmark methodology โ†’ ยท Run it yourself โ†’


User Stats

Share your numbers and get listed. PR your stats to the table below.

User / Team Agent Tokens/month saved Skills used
Your name here โ€” โ€” โ€”

How to measure

  1. Use agent for 1 week without paleo โ†’ note token usage from provider dashboard
  2. Enable paleo or paleo-auto for 1 week โ†’ note new usage
  3. Diff = your monthly savings (ร—4 for monthly estimate)

Template for PR:

| @yourgithub | claude-code | 1,200,000 | paleo + budget |

Why paleo?

LLMs ship with verbose defaults โ€” preambles, "Sure, I'll help...", re-stating context, over-explaining. That burns tokens on every single turn. paleo strips the fluff while keeping the code and accuracy intact.

What it saves

  • Output tokens: No preambles, no "let me explain", code-first
  • Context tokens: Trim stale history, compress conversation, summarize bulky output
  • Structured tokens: Minify JSON, collapse arrays

What it doesn't touch

  • Code correctness
  • Technical accuracy
  • File paths, IDs, error strings
  • The user's latest instruction

Contributing

New skill ideas? Optimization suggestions? Open an issue or PR.

Local benchmark

IDROUTER_API_KEY=sk-... python3 bench/benchmark.py --model deepseek/deepseek-v4-pro

MIT ยฉ 2026. No AI/agent credit in this repo. Built for humans who read code.

Yorumlar (0)

Sonuc bulunamadi