token-optimizer-off

自动压缩会话输入token，智能管理上下文和任务记忆，将token数量从10万+压缩至8,000以内，降低成本。

This skill appears to do what it claims: compress conversation history using an LLM and manage per-task memory files. Before installing: (1) review ~/.openclaw/config.json because the scripts will read it and may use its apiKey to call an LLM; ensure you trust that configuration and its permissions; (2) expect the skill to read/write files under ~/.openclaw/memory and create backups and a .session_state.json — back up anything important first; (3) the skill uses the OpenAI client (openai package) and will make network calls to whatever API URL/model you configure; provide an API key only if you trust the code and provider; (4) inspect the provided scripts (they are plain Python and readable) and run them in a safe environment or sandbox before granting access to production data; (5) integration example uses subprocess.run with a path containing '~' (tilde) which may not expand — fix path expansion in your agent code. If you want greater assurance, run the unit tests locally (they mock/require a TOKEN_OPTIMIZER_API_KEY env var) and verify no unexpected network endpoints or logging of secrets occur. If you observe the skill contacting unknown remote endpoints or printing full contents of config/API keys, treat that as suspicious and stop using it.

Type: OpenClaw Skill Name: token-optimizer-off Version: 1.0.0 The skill bundle performs high-risk operations by accessing the master OpenClaw configuration file (~/.openclaw/config.json) to extract sensitive AI API keys and base URLs (scripts/compress_session.py). While this is intended to provide a 'zero-config' experience for its token optimization features, reading credentials outside of the skill's own directory is a significant security risk. The bundle also manages files in the global ~/.openclaw/memory/ directory and provides instructions in SKILL.md for the AI agent to execute these scripts automatically. No evidence of intentional data exfiltration to unauthorized third parties was found, as the credentials appear to be used solely for interacting with the user's configured AI service.