LYGO Context Guard

LYGO Context Guard是AI代理的本地预飞行工具，核心功能含令牌估算、秘密脱敏、确定性上下文压缩、预算校验等，支持多命令操作。适用于AI代理处理大工具输出、日志、长聊天时，上下文超限、需降API成本、防密钥泄露的场景。输入支持文本或文件，输出为标准输出JSON/文本，本地无网络运行，安全可控。

deepseekoracle

热度 15.85

最近更新 2026-08-07 11:54:38

分类：效率依赖：代码执行 / 读写文件应用场景：数据处理 / 自动化脚本 / 工作流 / 模板

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add lygo-context-guard

Metadata

name: lygo-context-guard
description: LYGO Context Guard — pre-flight token budget, secret redaction, and deterministic context compaction for AI agents. Use when stuffing tool dumps, logs, files, or long chats into a model; when context is too large; when you need to save tokens / reduce API cost; when you must redact API keys before re-injection. Pure local stdlib. No network, no subprocess. Commands: estimate, redact, compact, budget, toolpack, preflight. Install clawhub:@deepseekoracle/lygo-context-guard.
version: 1.0.0
license: MIT-0
metadata: {"openclaw":{"emoji":"🛡️","homepage":"https://github.com/DeepSeekOracle/lygo-protocol-stack/tree/main/docs/skills/lygo-context-guard","requires":{"anyBins":["python","python3"]}},"lygo":true,"context":true,"tokens":true,"security":true,"signature":"Delta9Phi963-CONTEXT-GUARD-v1.0.0","publisher":"deepseekoracle","clawhub":"https://clawhub.ai/deepseekoracle/skills/lygo-context-guard","permissions":{"network":false,"shell":false,"subprocess":false,"filesystem":{"read":"optional --file user path","write":"skill state/ only with --i-consent"},"publish":false}}

LYGO Context Guard v1.0.0

The pre-flight every agent needs.

OpenClaw burns money when tool dumps and chat history explode the context window.
Context Guard is a local lattice gate: estimate tokens → redact secrets → compact deterministically → check budget — before you call the model.

Signature: Delta9Phi963-CONTEXT-GUARD-v1.0.0
ClawHub: @deepseekoracle/lygo-context-guard

Why this exists

Problem	What agents do wrong	Guard does
Token burn	Re-inject full logs/tool JSON every turn	Compact + budget exit code
Secret leaks	API keys in tool output → model context	Redact patterns
Context overflow	“Just paste the whole file”	Head/tail + mid map + SHA
No visibility	Guessing token counts	Dual heuristic estimate

Searches this skill targets: token saver, context window, compress context, reduce tokens, prompt budget, redact secrets, tool result too large.

Install

npx clawhub@latest install deepseekoracle/lygo-context-guard

Commands

cd path/to/lygo-context-guard
python scripts/self_check.py

# Estimate tokens (heuristic)
python scripts/context_guard.py estimate --text "long tool dump..."
python scripts/context_guard.py estimate --file ./huge_log.txt

# Redact secrets (keys, JWTs, private key blocks, password=...)
python scripts/context_guard.py redact --file ./tool_out.txt

# Deterministic compact (dedupe lines + head/tail keep)
python scripts/context_guard.py compact --file ./tool_out.txt --max-chars 8000

# Budget gate (exit 10 if over)
python scripts/context_guard.py budget --file ./pack.txt --budget 4000

# One-shot for agents: redact + compact + budget
python scripts/context_guard.py toolpack --file ./tool_out.txt --budget 4000 --max-chars 8000

# Full preflight report
python scripts/context_guard.py preflight --file ./blob.txt --budget 8000
python scripts/context_guard.py preflight --file ./blob.txt --write last_preflight.json --i-consent

# Demo (synthetic leaky dump)
python scripts/context_guard.py demo

Command	Network	Writes	Notes
`estimate` / `redact` / `compact` / `budget` / `toolpack`	none	no	stdout JSON + text
`preflight --write`	none	skill `state/` only	needs `--i-consent`
`demo`	none	no	built-in sample

Agent recipe (copy this)

Before re-injecting a large tool result or log:

1. Save tool output to a temp file
2. python scripts/context_guard.py toolpack --file THAT_FILE --budget 4000
3. If exit code 10 → lower max-chars or split the task
4. Feed packed_text back into the model — not the raw dump

Exit codes:

Code	Meaning
0	OK / under budget
10	Over budget after pack
2	Missing file / bad input

What it does not do

No network / shell / subprocess
No cloud summarizer (deterministic only — reproducible)
No auto model calls
Does not claim exact tokenizer parity (honest heuristic; good for budgets)
Never stores secrets in reports when writing preflight (stores hash of packed text)

Pair with

Skill	Role
`lygo-api-token-saver`	Broader pay-to-go budget ops + local army preference
`lygo-pxpipe-lygo`	Vision / multi-tool dump compression
`lygo-ops-detector`	Discourse signal analysis (different problem)
`lygo-kickstart-wizard`	Onboarding map

Security

See references/SECURITY.md.
Δ9Φ963 — compress chaos · redact secrets · budget the lattice · human remains the publisher.