MaskClaw: Evolutionary Personalized Privacy Guard

MaskClaw - 端侧隐私保护 Skill 套件，提供智能打码、行为监控与规则自进化能力

Key things to consider before installing or running this skill: 1) Incomplete bundle: SKILL.md and code reference a local model server (model_server/minicpm_api.py) and platform modules (memory.chroma_manager, sandbox.*, skill_registry) that are not present in the package. Do not run the code until you confirm those components exist and are the exact implementations you expect. 2) Self-evolution persistence: The evolution engine is designed to write new skills/rules under user_skills/ and to update a skill registry DB. This is powerful and can persist generated code/artifacts. If you plan to use this, run it in an isolated environment (VM/container) and audit any generated files before allowing them to be loaded or executed. 3) Review network and file paths: The package talks to a local LLM at 127.0.0.1 and writes logs to memory/ and user_skills/. Ensure the model server is actually local and not configured to proxy to an external host. Inspect what is stored in memory/logs and user_skills for sensitive data and manage file permissions accordingly. 4) Confirm masking guarantees: The architecture claims raw screenshots are never uploaded. Validate in practice by tracing all code paths that forward data (search for any external URLs or non-local hosts) and ensure only masked images (not originals) are forwarded to any remote or cloud agent. 5) Dependency and import issues: The evolution module swallows ImportError in a try/except but then proceeds to call components (ChromaManager(), SkillDB()) without guarding for their absence — this will crash or behave unpredictably. Ensure all required internal modules or replacement stubs are present before running. 6) Test in a sandbox: If you want to evaluate functionality, run demos in a controlled environment with no network egress, minimal test data, and review all files produced. Require and manually run the 'sandbox/regression_test.py' and any validators the package claims to use before enabling auto-publishing. If you want, I can (a) list the exact lines that attempt to write to user_skills / skill_registry, (b) search the code for any non-local network endpoints, or (c) suggest a minimal safe test plan/commands to run this package in an isolated environment.

Type: OpenClaw Skill Name: maskclaw Version: 1.0.0 The MaskClaw skill bundle is a privacy-preserving framework designed to mask sensitive information (PII) and refine privacy rules through user feedback. It utilizes RapidOCR and OpenCV for local image de-identification (smart_masker.py), logs agent/user interactions to identify privacy 'corrections' (behavior_monitor.py), and employs a structured 'hill-climbing' evolution mechanic to optimize Standard Operating Procedures (SOPs) using a local LLM API (evolution_mechanic.py). The code is well-documented, lacks any indicators of data exfiltration or unauthorized execution, and its complex 'self-evolution' features are strictly constrained to refining privacy-related instructions within a local environment.